diff options
Diffstat (limited to 'pkg')
119 files changed, 3987 insertions, 869 deletions
diff --git a/pkg/abi/linux/elf.go b/pkg/abi/linux/elf.go index 40f0459a0..7c9a02f20 100644 --- a/pkg/abi/linux/elf.go +++ b/pkg/abi/linux/elf.go @@ -102,4 +102,7 @@ const ( // NT_X86_XSTATE is for x86 extended state using xsave. NT_X86_XSTATE = 0x202 + + // NT_ARM_TLS is for ARM TLS register. + NT_ARM_TLS = 0x401 ) diff --git a/pkg/abi/linux/netfilter.go b/pkg/abi/linux/netfilter.go index 314c318b6..e20c625f7 100644 --- a/pkg/abi/linux/netfilter.go +++ b/pkg/abi/linux/netfilter.go @@ -268,7 +268,7 @@ const ( NF_NAT_RANGE_PERSISTENT | NF_NAT_RANGE_PROTO_RANDOM_FULLY) ) -// NfNATIPV4Range. It corresponds to struct nf_nat_ipv4_range +// NfNATIPV4Range corresponds to struct nf_nat_ipv4_range // in include/uapi/linux/netfilter/nf_nat.h. The fields are in // network byte order. type NfNATIPV4Range struct { @@ -279,7 +279,7 @@ type NfNATIPV4Range struct { MaxPort uint16 } -// NfNATIPV4MultiRangeCompat. It corresponds to struct +// NfNATIPV4MultiRangeCompat corresponds to struct // nf_nat_ipv4_multi_range_compat in include/uapi/linux/netfilter/nf_nat.h. type NfNATIPV4MultiRangeCompat struct { RangeSize uint32 diff --git a/pkg/buffer/BUILD b/pkg/buffer/BUILD new file mode 100644 index 000000000..a77a3beea --- /dev/null +++ b/pkg/buffer/BUILD @@ -0,0 +1,39 @@ +load("//tools:defs.bzl", "go_library", "go_test") +load("//tools/go_generics:defs.bzl", "go_template_instance") + +package(licenses = ["notice"]) + +go_template_instance( + name = "buffer_list", + out = "buffer_list.go", + package = "buffer", + prefix = "buffer", + template = "//pkg/ilist:generic_list", + types = { + "Element": "*Buffer", + "Linker": "*Buffer", + }, +) + +go_library( + name = "buffer", + srcs = [ + "buffer.go", + "buffer_list.go", + "safemem.go", + "view.go", + "view_unsafe.go", + ], + visibility = ["//visibility:public"], + deps = [ + "//pkg/log", + "//pkg/safemem", + ], +) + +go_test( + name = "buffer_test", + size = "small", + srcs = ["view_test.go"], + library = ":buffer", +) diff --git a/pkg/buffer/buffer.go b/pkg/buffer/buffer.go new file mode 100644 index 000000000..d5f64609b --- /dev/null +++ b/pkg/buffer/buffer.go @@ -0,0 +1,67 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package buffer provides the implementation of a buffer view. +package buffer + +import ( + "sync" +) + +const bufferSize = 8144 // See below. + +// Buffer encapsulates a queueable byte buffer. +// +// Note that the total size is slightly less than two pages. This is done +// intentionally to ensure that the buffer object aligns with runtime +// internals. We have no hard size or alignment requirements. This two page +// size will effectively minimize internal fragmentation, but still have a +// large enough chunk to limit excessive segmentation. +// +// +stateify savable +type Buffer struct { + data [bufferSize]byte + read int + write int + bufferEntry +} + +// Reset resets internal data. +// +// This must be called before use. +func (b *Buffer) Reset() { + b.read = 0 + b.write = 0 +} + +// Empty indicates the buffer is empty. +// +// This indicates there is no data left to read. +func (b *Buffer) Empty() bool { + return b.read == b.write +} + +// Full indicates the buffer is full. +// +// This indicates there is no capacity left to write. +func (b *Buffer) Full() bool { + return b.write == len(b.data) +} + +// bufferPool is a pool for buffers. +var bufferPool = sync.Pool{ + New: func() interface{} { + return new(Buffer) + }, +} diff --git a/pkg/buffer/safemem.go b/pkg/buffer/safemem.go new file mode 100644 index 000000000..071aaa488 --- /dev/null +++ b/pkg/buffer/safemem.go @@ -0,0 +1,131 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package buffer + +import ( + "io" + + "gvisor.dev/gvisor/pkg/safemem" +) + +// WriteBlock returns this buffer as a write Block. +func (b *Buffer) WriteBlock() safemem.Block { + return safemem.BlockFromSafeSlice(b.data[b.write:]) +} + +// ReadBlock returns this buffer as a read Block. +func (b *Buffer) ReadBlock() safemem.Block { + return safemem.BlockFromSafeSlice(b.data[b.read:b.write]) +} + +// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. +// +// This will advance the write index. +func (v *View) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { + need := int(srcs.NumBytes()) + if need == 0 { + return 0, nil + } + + var ( + dst safemem.BlockSeq + blocks []safemem.Block + ) + + // Need at least one buffer. + firstBuf := v.data.Back() + if firstBuf == nil { + firstBuf = bufferPool.Get().(*Buffer) + v.data.PushBack(firstBuf) + } + + // Does the last block have sufficient capacity alone? + if l := len(firstBuf.data) - firstBuf.write; l >= need { + dst = safemem.BlockSeqOf(firstBuf.WriteBlock()) + } else { + // Append blocks until sufficient. + need -= l + blocks = append(blocks, firstBuf.WriteBlock()) + for need > 0 { + emptyBuf := bufferPool.Get().(*Buffer) + v.data.PushBack(emptyBuf) + need -= len(emptyBuf.data) // Full block. + blocks = append(blocks, emptyBuf.WriteBlock()) + } + dst = safemem.BlockSeqFromSlice(blocks) + } + + // Perform the copy. + n, err := safemem.CopySeq(dst, srcs) + v.size += int64(n) + + // Update all indices. + for left := int(n); left > 0; firstBuf = firstBuf.Next() { + if l := len(firstBuf.data) - firstBuf.write; left >= l { + firstBuf.write += l // Whole block. + left -= l + } else { + firstBuf.write += left // Partial block. + left = 0 + } + } + + return n, err +} + +// ReadToBlocks implements safemem.Reader.ReadToBlocks. +// +// This will not advance the read index; the caller should follow +// this call with a call to TrimFront in order to remove the read +// data from the buffer. This is done to support pipe sematics. +func (v *View) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { + need := int(dsts.NumBytes()) + if need == 0 { + return 0, nil + } + + var ( + src safemem.BlockSeq + blocks []safemem.Block + ) + + firstBuf := v.data.Front() + if firstBuf == nil { + return 0, io.EOF + } + + // Is all the data in a single block? + if l := firstBuf.write - firstBuf.read; l >= need { + src = safemem.BlockSeqOf(firstBuf.ReadBlock()) + } else { + // Build a list of all the buffers. + need -= l + blocks = append(blocks, firstBuf.ReadBlock()) + for buf := firstBuf.Next(); buf != nil && need > 0; buf = buf.Next() { + need -= buf.write - buf.read + blocks = append(blocks, buf.ReadBlock()) + } + src = safemem.BlockSeqFromSlice(blocks) + } + + // Perform the copy. + n, err := safemem.CopySeq(dsts, src) + + // See above: we would normally advance the read index here, but we + // don't do that in order to support pipe semantics. We rely on a + // separate call to TrimFront() in this case. + + return n, err +} diff --git a/pkg/buffer/view.go b/pkg/buffer/view.go new file mode 100644 index 000000000..00fc11e9c --- /dev/null +++ b/pkg/buffer/view.go @@ -0,0 +1,382 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package buffer + +import ( + "fmt" + "io" +) + +// View is a non-linear buffer. +// +// All methods are thread compatible. +// +// +stateify savable +type View struct { + data bufferList + size int64 +} + +// TrimFront removes the first count bytes from the buffer. +func (v *View) TrimFront(count int64) { + if count >= v.size { + v.advanceRead(v.size) + } else { + v.advanceRead(count) + } +} + +// Read implements io.Reader.Read. +// +// Note that reading does not advance the read index. This must be done +// manually using TrimFront or other methods. +func (v *View) Read(p []byte) (int, error) { + return v.ReadAt(p, 0) +} + +// ReadAt implements io.ReaderAt.ReadAt. +func (v *View) ReadAt(p []byte, offset int64) (int, error) { + var ( + skipped int64 + done int64 + ) + for buf := v.data.Front(); buf != nil && done < int64(len(p)); buf = buf.Next() { + needToSkip := int(offset - skipped) + if l := buf.write - buf.read; l <= needToSkip { + skipped += int64(l) + continue + } + + // Actually read data. + n := copy(p[done:], buf.data[buf.read+needToSkip:buf.write]) + skipped += int64(needToSkip) + done += int64(n) + } + if int(done) < len(p) { + return int(done), io.EOF + } + return int(done), nil +} + +// Write implements io.Writer.Write. +func (v *View) Write(p []byte) (int, error) { + v.Append(p) // Does not fail. + return len(p), nil +} + +// advanceRead advances the view's read index. +// +// Precondition: there must be sufficient bytes in the buffer. +func (v *View) advanceRead(count int64) { + for buf := v.data.Front(); buf != nil && count > 0; { + l := int64(buf.write - buf.read) + if l > count { + // There is still data for reading. + buf.read += int(count) + v.size -= count + count = 0 + break + } + + // Read from this buffer. + buf.read += int(l) + count -= l + v.size -= l + + // When all data has been read from a buffer, we push + // it into the empty buffer pool for reuse. + oldBuf := buf + buf = buf.Next() // Iterate. + v.data.Remove(oldBuf) + oldBuf.Reset() + bufferPool.Put(oldBuf) + } + if count > 0 { + panic(fmt.Sprintf("advanceRead still has %d bytes remaining", count)) + } +} + +// Truncate truncates the view to the given bytes. +func (v *View) Truncate(length int64) { + if length < 0 || length >= v.size { + return // Nothing to do. + } + for buf := v.data.Back(); buf != nil && v.size > length; buf = v.data.Back() { + l := int64(buf.write - buf.read) // Local bytes. + switch { + case v.size-l >= length: + // Drop the buffer completely; see above. + v.data.Remove(buf) + v.size -= l + buf.Reset() + bufferPool.Put(buf) + + case v.size > length && v.size-l < length: + // Just truncate the buffer locally. + delta := (length - (v.size - l)) + buf.write = buf.read + int(delta) + v.size = length + + default: + // Should never happen. + panic("invalid buffer during truncation") + } + } + v.size = length // Save the new size. +} + +// Grow grows the given view to the number of bytes. If zero +// is true, all these bytes will be zero. If zero is false, +// then this is the caller's responsibility. +// +// Precondition: length must be >= 0. +func (v *View) Grow(length int64, zero bool) { + if length < 0 { + panic("negative length provided") + } + for v.size < length { + buf := v.data.Back() + + // Is there at least one buffer? + if buf == nil || buf.Full() { + buf = bufferPool.Get().(*Buffer) + v.data.PushBack(buf) + } + + // Write up to length bytes. + l := len(buf.data) - buf.write + if int64(l) > length-v.size { + l = int(length - v.size) + } + + // Zero the written section; note that this pattern is + // specifically recognized and optimized by the compiler. + if zero { + for i := buf.write; i < buf.write+l; i++ { + buf.data[i] = 0 + } + } + + // Advance the index. + buf.write += l + v.size += int64(l) + } +} + +// Prepend prepends the given data. +func (v *View) Prepend(data []byte) { + // Is there any space in the first buffer? + if buf := v.data.Front(); buf != nil && buf.read > 0 { + // Fill up before the first write. + avail := buf.read + copy(buf.data[0:], data[len(data)-avail:]) + data = data[:len(data)-avail] + v.size += int64(avail) + } + + for len(data) > 0 { + // Do we need an empty buffer? + buf := bufferPool.Get().(*Buffer) + v.data.PushFront(buf) + + // The buffer is empty; copy last chunk. + start := len(data) - len(buf.data) + if start < 0 { + start = 0 // Everything. + } + + // We have to put the data at the end of the current + // buffer in order to ensure that the next prepend will + // correctly fill up the beginning of this buffer. + bStart := len(buf.data) - len(data[start:]) + n := copy(buf.data[bStart:], data[start:]) + buf.read = bStart + buf.write = len(buf.data) + data = data[:start] + v.size += int64(n) + } +} + +// Append appends the given data. +func (v *View) Append(data []byte) { + for done := 0; done < len(data); { + buf := v.data.Back() + + // Find the first empty buffer. + if buf == nil || buf.Full() { + buf = bufferPool.Get().(*Buffer) + v.data.PushBack(buf) + } + + // Copy in to the given buffer. + n := copy(buf.data[buf.write:], data[done:]) + done += n + buf.write += n + v.size += int64(n) + } +} + +// Flatten returns a flattened copy of this data. +// +// This method should not be used in any performance-sensitive paths. It may +// allocate a fresh byte slice sufficiently large to contain all the data in +// the buffer. +// +// N.B. Tee data still belongs to this view, as if there is a single buffer +// present, then it will be returned directly. This should be used for +// temporary use only, and a reference to the given slice should not be held. +func (v *View) Flatten() []byte { + if buf := v.data.Front(); buf.Next() == nil { + return buf.data[buf.read:buf.write] // Only one buffer. + } + data := make([]byte, 0, v.size) // Need to flatten. + for buf := v.data.Front(); buf != nil; buf = buf.Next() { + // Copy to the allocated slice. + data = append(data, buf.data[buf.read:buf.write]...) + } + return data +} + +// Size indicates the total amount of data available in this view. +func (v *View) Size() (sz int64) { + sz = v.size // Pre-calculated. + return sz +} + +// Copy makes a strict copy of this view. +func (v *View) Copy() (other View) { + for buf := v.data.Front(); buf != nil; buf = buf.Next() { + other.Append(buf.data[buf.read:buf.write]) + } + return other +} + +// Apply applies the given function across all valid data. +func (v *View) Apply(fn func([]byte)) { + for buf := v.data.Front(); buf != nil; buf = buf.Next() { + if l := int64(buf.write - buf.read); l > 0 { + fn(buf.data[buf.read:buf.write]) + } + } +} + +// Merge merges the provided View with this one. +// +// The other view will be empty after this operation. +func (v *View) Merge(other *View) { + // Copy over all buffers. + for buf := other.data.Front(); buf != nil && !buf.Empty(); buf = other.data.Front() { + other.data.Remove(buf) + v.data.PushBack(buf) + } + + // Adjust sizes. + v.size += other.size + other.size = 0 +} + +// WriteFromReader writes to the buffer from an io.Reader. +func (v *View) WriteFromReader(r io.Reader, count int64) (int64, error) { + var ( + done int64 + n int + err error + ) + for done < count { + buf := v.data.Back() + + // Find the first empty buffer. + if buf == nil || buf.Full() { + buf = bufferPool.Get().(*Buffer) + v.data.PushBack(buf) + } + + // Is this less than the minimum batch? + if len(buf.data[buf.write:]) < minBatch && (count-done) >= int64(minBatch) { + tmp := make([]byte, minBatch) + n, err = r.Read(tmp) + v.Write(tmp[:n]) + done += int64(n) + if err != nil { + break + } + continue + } + + // Limit the read, if necessary. + end := len(buf.data) + if int64(end-buf.write) > (count - done) { + end = buf.write + int(count-done) + } + + // Pass the relevant portion of the buffer. + n, err = r.Read(buf.data[buf.write:end]) + buf.write += n + done += int64(n) + v.size += int64(n) + if err == io.EOF { + err = nil // Short write allowed. + break + } else if err != nil { + break + } + } + return done, err +} + +// ReadToWriter reads from the buffer into an io.Writer. +// +// N.B. This does not consume the bytes read. TrimFront should +// be called appropriately after this call in order to do so. +func (v *View) ReadToWriter(w io.Writer, count int64) (int64, error) { + var ( + done int64 + n int + err error + ) + offset := 0 // Spill-over for batching. + for buf := v.data.Front(); buf != nil && done < count; buf = buf.Next() { + l := buf.write - buf.read - offset + + // Is this less than the minimum batch? + if l < minBatch && (count-done) >= int64(minBatch) && (v.size-done) >= int64(minBatch) { + tmp := make([]byte, minBatch) + n, err = v.ReadAt(tmp, done) + w.Write(tmp[:n]) + done += int64(n) + offset = n - l // Reset below. + if err != nil { + break + } + continue + } + + // Limit the write if necessary. + if int64(l) >= (count - done) { + l = int(count - done) + } + + // Perform the actual write. + n, err = w.Write(buf.data[buf.read+offset : buf.read+offset+l]) + done += int64(n) + if err != nil { + break + } + + // Reset spill-over. + offset = 0 + } + return done, err +} diff --git a/pkg/buffer/view_test.go b/pkg/buffer/view_test.go new file mode 100644 index 000000000..37e652f16 --- /dev/null +++ b/pkg/buffer/view_test.go @@ -0,0 +1,233 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package buffer + +import ( + "bytes" + "strings" + "testing" +) + +func TestView(t *testing.T) { + testCases := []struct { + name string + input string + output string + ops []func(*View) + }{ + // Prepend. + { + name: "prepend", + input: "world", + ops: []func(*View){ + func(v *View) { + v.Prepend([]byte("hello ")) + }, + }, + output: "hello world", + }, + { + name: "prepend fill", + input: strings.Repeat("1", bufferSize-1), + ops: []func(*View){ + func(v *View) { + v.Prepend([]byte("0")) + }, + }, + output: "0" + strings.Repeat("1", bufferSize-1), + }, + { + name: "prepend overflow", + input: strings.Repeat("1", bufferSize), + ops: []func(*View){ + func(v *View) { + v.Prepend([]byte("0")) + }, + }, + output: "0" + strings.Repeat("1", bufferSize), + }, + { + name: "prepend multiple buffers", + input: strings.Repeat("1", bufferSize-1), + ops: []func(*View){ + func(v *View) { + v.Prepend([]byte(strings.Repeat("0", bufferSize*3))) + }, + }, + output: strings.Repeat("0", bufferSize*3) + strings.Repeat("1", bufferSize-1), + }, + + // Append. + { + name: "append", + input: "hello", + ops: []func(*View){ + func(v *View) { + v.Append([]byte(" world")) + }, + }, + output: "hello world", + }, + { + name: "append fill", + input: strings.Repeat("1", bufferSize-1), + ops: []func(*View){ + func(v *View) { + v.Append([]byte("0")) + }, + }, + output: strings.Repeat("1", bufferSize-1) + "0", + }, + { + name: "append overflow", + input: strings.Repeat("1", bufferSize), + ops: []func(*View){ + func(v *View) { + v.Append([]byte("0")) + }, + }, + output: strings.Repeat("1", bufferSize) + "0", + }, + { + name: "append multiple buffers", + input: strings.Repeat("1", bufferSize-1), + ops: []func(*View){ + func(v *View) { + v.Append([]byte(strings.Repeat("0", bufferSize*3))) + }, + }, + output: strings.Repeat("1", bufferSize-1) + strings.Repeat("0", bufferSize*3), + }, + + // Truncate. + { + name: "truncate", + input: "hello world", + ops: []func(*View){ + func(v *View) { + v.Truncate(5) + }, + }, + output: "hello", + }, + { + name: "truncate multiple buffers", + input: strings.Repeat("1", bufferSize*2), + ops: []func(*View){ + func(v *View) { + v.Truncate(bufferSize*2 - 1) + }, + }, + output: strings.Repeat("1", bufferSize*2-1), + }, + { + name: "truncate multiple buffers to one buffer", + input: strings.Repeat("1", bufferSize*2), + ops: []func(*View){ + func(v *View) { + v.Truncate(5) + }, + }, + output: "11111", + }, + + // TrimFront. + { + name: "trim", + input: "hello world", + ops: []func(*View){ + func(v *View) { + v.TrimFront(6) + }, + }, + output: "world", + }, + { + name: "trim multiple buffers", + input: strings.Repeat("1", bufferSize*2), + ops: []func(*View){ + func(v *View) { + v.TrimFront(1) + }, + }, + output: strings.Repeat("1", bufferSize*2-1), + }, + { + name: "trim multiple buffers to one buffer", + input: strings.Repeat("1", bufferSize*2), + ops: []func(*View){ + func(v *View) { + v.TrimFront(bufferSize*2 - 1) + }, + }, + output: "1", + }, + + // Grow. + { + name: "grow", + input: "hello world", + ops: []func(*View){ + func(v *View) { + v.Grow(1, true) + }, + }, + output: "hello world", + }, + { + name: "grow from zero", + ops: []func(*View){ + func(v *View) { + v.Grow(1024, true) + }, + }, + output: strings.Repeat("\x00", 1024), + }, + { + name: "grow from non-zero", + input: strings.Repeat("1", bufferSize), + ops: []func(*View){ + func(v *View) { + v.Grow(bufferSize*2, true) + }, + }, + output: strings.Repeat("1", bufferSize) + strings.Repeat("\x00", bufferSize), + }, + } + + for _, tc := range testCases { + t.Run(tc.name, func(t *testing.T) { + // Construct the new view. + var view View + view.Append([]byte(tc.input)) + + // Run all operations. + for _, op := range tc.ops { + op(&view) + } + + // Flatten and validate. + out := view.Flatten() + if !bytes.Equal([]byte(tc.output), out) { + t.Errorf("expected %q, got %q", tc.output, string(out)) + } + + // Ensure the size is correct. + if len(out) != int(view.Size()) { + t.Errorf("size is wrong: expected %d, got %d", len(out), view.Size()) + } + }) + } +} diff --git a/pkg/sentry/kernel/pipe/buffer_test.go b/pkg/buffer/view_unsafe.go index 4d54b8b8f..d1ef39b26 100644 --- a/pkg/sentry/kernel/pipe/buffer_test.go +++ b/pkg/buffer/view_unsafe.go @@ -1,4 +1,4 @@ -// Copyright 2019 The gVisor Authors. +// Copyright 2020 The gVisor Authors. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -12,21 +12,14 @@ // See the License for the specific language governing permissions and // limitations under the License. -package pipe +package buffer import ( - "testing" "unsafe" - - "gvisor.dev/gvisor/pkg/usermem" ) -func TestBufferSize(t *testing.T) { - bufferSize := unsafe.Sizeof(buffer{}) - if bufferSize < usermem.PageSize { - t.Errorf("buffer is less than a page") - } - if bufferSize > (2 * usermem.PageSize) { - t.Errorf("buffer is greater than two pages") - } -} +// minBatch is the smallest Read or Write operation that the +// WriteFromReader and ReadToWriter functions will use. +// +// This is defined as the size of a native pointer. +const minBatch = int(unsafe.Sizeof(uintptr(0))) diff --git a/pkg/ilist/list.go b/pkg/ilist/list.go index 019caadca..8f93e4d6d 100644 --- a/pkg/ilist/list.go +++ b/pkg/ilist/list.go @@ -88,8 +88,9 @@ func (l *List) Back() Element { // PushFront inserts the element e at the front of list l. func (l *List) PushFront(e Element) { - ElementMapper{}.linkerFor(e).SetNext(l.head) - ElementMapper{}.linkerFor(e).SetPrev(nil) + linker := ElementMapper{}.linkerFor(e) + linker.SetNext(l.head) + linker.SetPrev(nil) if l.head != nil { ElementMapper{}.linkerFor(l.head).SetPrev(e) @@ -102,8 +103,9 @@ func (l *List) PushFront(e Element) { // PushBack inserts the element e at the back of list l. func (l *List) PushBack(e Element) { - ElementMapper{}.linkerFor(e).SetNext(nil) - ElementMapper{}.linkerFor(e).SetPrev(l.tail) + linker := ElementMapper{}.linkerFor(e) + linker.SetNext(nil) + linker.SetPrev(l.tail) if l.tail != nil { ElementMapper{}.linkerFor(l.tail).SetNext(e) @@ -132,10 +134,14 @@ func (l *List) PushBackList(m *List) { // InsertAfter inserts e after b. func (l *List) InsertAfter(b, e Element) { - a := ElementMapper{}.linkerFor(b).Next() - ElementMapper{}.linkerFor(e).SetNext(a) - ElementMapper{}.linkerFor(e).SetPrev(b) - ElementMapper{}.linkerFor(b).SetNext(e) + bLinker := ElementMapper{}.linkerFor(b) + eLinker := ElementMapper{}.linkerFor(e) + + a := bLinker.Next() + + eLinker.SetNext(a) + eLinker.SetPrev(b) + bLinker.SetNext(e) if a != nil { ElementMapper{}.linkerFor(a).SetPrev(e) @@ -146,10 +152,13 @@ func (l *List) InsertAfter(b, e Element) { // InsertBefore inserts e before a. func (l *List) InsertBefore(a, e Element) { - b := ElementMapper{}.linkerFor(a).Prev() - ElementMapper{}.linkerFor(e).SetNext(a) - ElementMapper{}.linkerFor(e).SetPrev(b) - ElementMapper{}.linkerFor(a).SetPrev(e) + aLinker := ElementMapper{}.linkerFor(a) + eLinker := ElementMapper{}.linkerFor(e) + + b := aLinker.Prev() + eLinker.SetNext(a) + eLinker.SetPrev(b) + aLinker.SetPrev(e) if b != nil { ElementMapper{}.linkerFor(b).SetNext(e) @@ -160,8 +169,9 @@ func (l *List) InsertBefore(a, e Element) { // Remove removes e from l. func (l *List) Remove(e Element) { - prev := ElementMapper{}.linkerFor(e).Prev() - next := ElementMapper{}.linkerFor(e).Next() + linker := ElementMapper{}.linkerFor(e) + prev := linker.Prev() + next := linker.Next() if prev != nil { ElementMapper{}.linkerFor(prev).SetNext(next) @@ -174,6 +184,9 @@ func (l *List) Remove(e Element) { } else { l.tail = prev } + + linker.SetNext(nil) + linker.SetPrev(nil) } // Entry is a default implementation of Linker. Users can add anonymous fields diff --git a/pkg/safemem/seq_test.go b/pkg/safemem/seq_test.go index eba4bb535..de34005e9 100644 --- a/pkg/safemem/seq_test.go +++ b/pkg/safemem/seq_test.go @@ -20,6 +20,27 @@ import ( "testing" ) +func TestBlockSeqOfEmptyBlock(t *testing.T) { + bs := BlockSeqOf(Block{}) + if !bs.IsEmpty() { + t.Errorf("BlockSeqOf(Block{}).IsEmpty(): got false, wanted true; BlockSeq is %v", bs) + } +} + +func TestBlockSeqOfNonemptyBlock(t *testing.T) { + b := BlockFromSafeSlice(make([]byte, 1)) + bs := BlockSeqOf(b) + if bs.IsEmpty() { + t.Fatalf("BlockSeqOf(non-empty Block).IsEmpty(): got true, wanted false; BlockSeq is %v", bs) + } + if head := bs.Head(); head != b { + t.Fatalf("BlockSeqOf(non-empty Block).Head(): got %v, wanted %v", head, b) + } + if tail := bs.Tail(); !tail.IsEmpty() { + t.Fatalf("BlockSeqOf(non-empty Block).Tail().IsEmpty(): got false, wanted true: tail is %v", tail) + } +} + type blockSeqTest struct { desc string diff --git a/pkg/safemem/seq_unsafe.go b/pkg/safemem/seq_unsafe.go index dcdfc9600..f5f0574f8 100644 --- a/pkg/safemem/seq_unsafe.go +++ b/pkg/safemem/seq_unsafe.go @@ -56,6 +56,9 @@ type BlockSeq struct { // BlockSeqOf returns a BlockSeq representing the single Block b. func BlockSeqOf(b Block) BlockSeq { + if b.length == 0 { + return BlockSeq{} + } bs := BlockSeq{ data: b.start, length: -1, diff --git a/pkg/seccomp/seccomp_test.go b/pkg/seccomp/seccomp_test.go index da5a5e4b2..88766f33b 100644 --- a/pkg/seccomp/seccomp_test.go +++ b/pkg/seccomp/seccomp_test.go @@ -451,7 +451,7 @@ func TestRandom(t *testing.T) { } } - fmt.Printf("Testing filters: %v", syscallRules) + t.Logf("Testing filters: %v", syscallRules) instrs, err := BuildProgram([]RuleSet{ RuleSet{ Rules: syscallRules, diff --git a/pkg/sentry/arch/arch_aarch64.go b/pkg/sentry/arch/arch_aarch64.go index d7794db63..b998f84fc 100644 --- a/pkg/sentry/arch/arch_aarch64.go +++ b/pkg/sentry/arch/arch_aarch64.go @@ -32,6 +32,12 @@ import ( const ( // SyscallWidth is the width of insturctions. SyscallWidth = 4 + + // fpsimdMagic is the magic number which is used in fpsimd_context. + fpsimdMagic = 0x46508001 + + // fpsimdContextSize is the size of fpsimd_context. + fpsimdContextSize = 0x210 ) // ARMTrapFlag is the mask for the trap flag. @@ -40,24 +46,24 @@ const ARMTrapFlag = uint64(1) << 21 // aarch64FPState is aarch64 floating point state. type aarch64FPState []byte -// initAarch64FPState (defined in asm files) sets up initial state. -func initAarch64FPState(data *FloatingPointData) { - // TODO(gvisor.dev/issue/1238): floating-point is not supported. +// initAarch64FPState sets up initial state. +func initAarch64FPState(data aarch64FPState) { + binary.LittleEndian.PutUint32(data, fpsimdMagic) + binary.LittleEndian.PutUint32(data[4:], fpsimdContextSize) } func newAarch64FPStateSlice() []byte { - return alignedBytes(4096, 32)[:4096] + return alignedBytes(4096, 16)[:fpsimdContextSize] } // newAarch64FPState returns an initialized floating point state. // // The returned state is large enough to store all floating point state // supported by host, even if the app won't use much of it due to a restricted -// FeatureSet. Since they may still be able to see state not advertised by -// CPUID we must ensure it does not contain any sentry state. +// FeatureSet. func newAarch64FPState() aarch64FPState { f := aarch64FPState(newAarch64FPStateSlice()) - initAarch64FPState(f.FloatingPointData()) + initAarch64FPState(f) return f } @@ -89,6 +95,9 @@ type State struct { // Our floating point state. aarch64FPState `state:"wait"` + // TLS pointer + TPValue uint64 + // FeatureSet is a pointer to the currently active feature set. FeatureSet *cpuid.FeatureSet } @@ -136,10 +145,11 @@ func (s State) Proto() *rpb.Registers { // Fork creates and returns an identical copy of the state. func (s *State) Fork() State { - // TODO(gvisor.dev/issue/1238): floating-point is not supported. return State{ - Regs: s.Regs, - FeatureSet: s.FeatureSet, + Regs: s.Regs, + aarch64FPState: s.aarch64FPState.fork(), + TPValue: s.TPValue, + FeatureSet: s.FeatureSet, } } @@ -249,6 +259,7 @@ func (s *State) PtraceSetFPRegs(src io.Reader) (int, error) { const ( _NT_PRSTATUS = 1 _NT_PRFPREG = 2 + _NT_ARM_TLS = 0x401 ) // PtraceGetRegSet implements Context.PtraceGetRegSet. @@ -288,8 +299,10 @@ func New(arch Arch, fs *cpuid.FeatureSet) Context { case ARM64: return &context64{ State{ - FeatureSet: fs, + aarch64FPState: newAarch64FPState(), + FeatureSet: fs, }, + []aarch64FPState(nil), } } panic(fmt.Sprintf("unknown architecture %v", arch)) diff --git a/pkg/sentry/arch/arch_arm64.go b/pkg/sentry/arch/arch_arm64.go index ac98897b5..db99c5acb 100644 --- a/pkg/sentry/arch/arch_arm64.go +++ b/pkg/sentry/arch/arch_arm64.go @@ -53,6 +53,11 @@ const ( preferredPIELoadAddr usermem.Addr = maxAddr64 / 6 * 5 ) +var ( + // CPUIDInstruction doesn't exist on ARM64. + CPUIDInstruction = []byte{} +) + // These constants are selected as heuristics to help make the Platform's // potentially limited address space conform as closely to Linux as possible. const ( @@ -68,6 +73,7 @@ const ( // context64 represents an ARM64 context. type context64 struct { State + sigFPState []aarch64FPState // fpstate to be restored on sigreturn. } // Arch implements Context.Arch. @@ -75,10 +81,19 @@ func (c *context64) Arch() Arch { return ARM64 } +func (c *context64) copySigFPState() []aarch64FPState { + var sigfps []aarch64FPState + for _, s := range c.sigFPState { + sigfps = append(sigfps, s.fork()) + } + return sigfps +} + // Fork returns an exact copy of this context. func (c *context64) Fork() Context { return &context64{ - State: c.State.Fork(), + State: c.State.Fork(), + sigFPState: c.copySigFPState(), } } @@ -125,16 +140,17 @@ func (c *context64) SetStack(value uintptr) { // TLS returns the current TLS pointer. func (c *context64) TLS() uintptr { - // TODO(gvisor.dev/issue/1238): TLS is not supported. - // MRS_TPIDR_EL0 - return 0 + return uintptr(c.TPValue) } // SetTLS sets the current TLS pointer. Returns false if value is invalid. func (c *context64) SetTLS(value uintptr) bool { - // TODO(gvisor.dev/issue/1238): TLS is not supported. - // MSR_TPIDR_EL0 - return false + if value >= uintptr(maxAddr64) { + return false + } + + c.TPValue = uint64(value) + return true } // SetOldRSeqInterruptedIP implements Context.SetOldRSeqInterruptedIP. diff --git a/pkg/sentry/arch/signal_arm64.go b/pkg/sentry/arch/signal_arm64.go index 4f4cc46a8..0c1db4b13 100644 --- a/pkg/sentry/arch/signal_arm64.go +++ b/pkg/sentry/arch/signal_arm64.go @@ -30,14 +30,29 @@ type SignalContext64 struct { Sp uint64 Pc uint64 Pstate uint64 - _pad [8]byte // __attribute__((__aligned__(16))) - Reserved [4096]uint8 + _pad [8]byte // __attribute__((__aligned__(16))) + Fpsimd64 FpsimdContext // size = 528 + Reserved [3568]uint8 +} + +type aarch64Ctx struct { + Magic uint32 + Size uint32 +} + +// FpsimdContext is equivalent to struct fpsimd_context on arm64 +// (arch/arm64/include/uapi/asm/sigcontext.h). +type FpsimdContext struct { + Head aarch64Ctx + Fpsr uint32 + Fpcr uint32 + Vregs [64]uint64 // actually [32]uint128 } // UContext64 is equivalent to ucontext on arm64(arch/arm64/include/uapi/asm/ucontext.h). type UContext64 struct { Flags uint64 - Link *UContext64 + Link uint64 Stack SignalStack Sigset linux.SignalSet // glibc uses a 1024-bit sigset_t diff --git a/pkg/sentry/control/pprof.go b/pkg/sentry/control/pprof.go index 151808911..663e51989 100644 --- a/pkg/sentry/control/pprof.go +++ b/pkg/sentry/control/pprof.go @@ -117,9 +117,9 @@ func (p *Profile) HeapProfile(o *ProfileOpts, _ *struct{}) error { return nil } -// Goroutine is an RPC stub which dumps out the stack trace for all running -// goroutines. -func (p *Profile) Goroutine(o *ProfileOpts, _ *struct{}) error { +// GoroutineProfile is an RPC stub which dumps out the stack trace for all +// running goroutines. +func (p *Profile) GoroutineProfile(o *ProfileOpts, _ *struct{}) error { if len(o.FilePayload.Files) < 1 { return errNoOutput } @@ -131,6 +131,34 @@ func (p *Profile) Goroutine(o *ProfileOpts, _ *struct{}) error { return nil } +// BlockProfile is an RPC stub which dumps out the stack trace that led to +// blocking on synchronization primitives. +func (p *Profile) BlockProfile(o *ProfileOpts, _ *struct{}) error { + if len(o.FilePayload.Files) < 1 { + return errNoOutput + } + output := o.FilePayload.Files[0] + defer output.Close() + if err := pprof.Lookup("block").WriteTo(output, 0); err != nil { + return err + } + return nil +} + +// MutexProfile is an RPC stub which dumps out the stack trace of holders of +// contended mutexes. +func (p *Profile) MutexProfile(o *ProfileOpts, _ *struct{}) error { + if len(o.FilePayload.Files) < 1 { + return errNoOutput + } + output := o.FilePayload.Files[0] + defer output.Close() + if err := pprof.Lookup("mutex").WriteTo(output, 0); err != nil { + return err + } + return nil +} + // StartTrace is an RPC stub which starts collection of an execution trace. func (p *Profile) StartTrace(o *ProfileOpts, _ *struct{}) error { if len(o.FilePayload.Files) < 1 { diff --git a/pkg/sentry/fs/dev/BUILD b/pkg/sentry/fs/dev/BUILD index 9b6bb26d0..9379a4d7b 100644 --- a/pkg/sentry/fs/dev/BUILD +++ b/pkg/sentry/fs/dev/BUILD @@ -26,6 +26,7 @@ go_library( "//pkg/sentry/fs/fsutil", "//pkg/sentry/fs/ramfs", "//pkg/sentry/fs/tmpfs", + "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/memmap", "//pkg/sentry/mm", diff --git a/pkg/sentry/fs/dev/dev.go b/pkg/sentry/fs/dev/dev.go index 7e66c29b0..acbd401a0 100644 --- a/pkg/sentry/fs/dev/dev.go +++ b/pkg/sentry/fs/dev/dev.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/ramfs" "gvisor.dev/gvisor/pkg/sentry/fs/tmpfs" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/usermem" ) @@ -124,10 +125,12 @@ func New(ctx context.Context, msrc *fs.MountSource) *fs.Inode { "ptmx": newSymlink(ctx, "pts/ptmx", msrc), "tty": newCharacterDevice(ctx, newTTYDevice(ctx, fs.RootOwner, 0666), msrc, ttyDevMajor, ttyDevMinor), + } - "net": newDirectory(ctx, map[string]*fs.Inode{ + if isNetTunSupported(inet.StackFromContext(ctx)) { + contents["net"] = newDirectory(ctx, map[string]*fs.Inode{ "tun": newCharacterDevice(ctx, newNetTunDevice(ctx, fs.RootOwner, 0666), msrc, netTunDevMajor, netTunDevMinor), - }, msrc), + }, msrc) } iops := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) diff --git a/pkg/sentry/fs/dev/net_tun.go b/pkg/sentry/fs/dev/net_tun.go index 755644488..dc7ad075a 100644 --- a/pkg/sentry/fs/dev/net_tun.go +++ b/pkg/sentry/fs/dev/net_tun.go @@ -20,6 +20,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/socket/netstack" "gvisor.dev/gvisor/pkg/syserror" @@ -168,3 +169,9 @@ func (fops *netTunFileOperations) EventRegister(e *waiter.Entry, mask waiter.Eve func (fops *netTunFileOperations) EventUnregister(e *waiter.Entry) { fops.device.EventUnregister(e) } + +// isNetTunSupported returns whether /dev/net/tun device is supported for s. +func isNetTunSupported(s inet.Stack) bool { + _, ok := s.(*netstack.Stack) + return ok +} diff --git a/pkg/sentry/fs/dirent.go b/pkg/sentry/fs/dirent.go index acab0411a..e0b32e1c1 100644 --- a/pkg/sentry/fs/dirent.go +++ b/pkg/sentry/fs/dirent.go @@ -1438,8 +1438,8 @@ func lockForRename(oldParent *Dirent, oldName string, newParent *Dirent, newName }, nil } -func checkSticky(ctx context.Context, dir *Dirent, victim *Dirent) error { - uattr, err := dir.Inode.UnstableAttr(ctx) +func (d *Dirent) checkSticky(ctx context.Context, victim *Dirent) error { + uattr, err := d.Inode.UnstableAttr(ctx) if err != nil { return syserror.EPERM } @@ -1465,30 +1465,33 @@ func checkSticky(ctx context.Context, dir *Dirent, victim *Dirent) error { return syserror.EPERM } -// MayDelete determines whether `name`, a child of `dir`, can be deleted or +// MayDelete determines whether `name`, a child of `d`, can be deleted or // renamed by `ctx`. // // Compare Linux kernel fs/namei.c:may_delete. -func MayDelete(ctx context.Context, root, dir *Dirent, name string) error { - if err := dir.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil { +func (d *Dirent) MayDelete(ctx context.Context, root *Dirent, name string) error { + if err := d.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil { return err } - victim, err := dir.Walk(ctx, root, name) + unlock := d.lockDirectory() + defer unlock() + + victim, err := d.walk(ctx, root, name, true /* may unlock */) if err != nil { return err } defer victim.DecRef() - return mayDelete(ctx, dir, victim) + return d.mayDelete(ctx, victim) } // mayDelete determines whether `victim`, a child of `dir`, can be deleted or // renamed by `ctx`. // // Preconditions: `dir` is writable and executable by `ctx`. -func mayDelete(ctx context.Context, dir, victim *Dirent) error { - if err := checkSticky(ctx, dir, victim); err != nil { +func (d *Dirent) mayDelete(ctx context.Context, victim *Dirent) error { + if err := d.checkSticky(ctx, victim); err != nil { return err } @@ -1542,7 +1545,7 @@ func Rename(ctx context.Context, root *Dirent, oldParent *Dirent, oldName string defer renamed.DecRef() // Check that the renamed dirent is deletable. - if err := mayDelete(ctx, oldParent, renamed); err != nil { + if err := oldParent.mayDelete(ctx, renamed); err != nil { return err } @@ -1580,7 +1583,7 @@ func Rename(ctx context.Context, root *Dirent, oldParent *Dirent, oldName string // across the Rename, so must call DecRef manually (no defer). // Check that we can delete replaced. - if err := mayDelete(ctx, newParent, replaced); err != nil { + if err := newParent.mayDelete(ctx, replaced); err != nil { replaced.DecRef() return err } diff --git a/pkg/sentry/fs/dirent_cache.go b/pkg/sentry/fs/dirent_cache.go index 25514ace4..33de32c69 100644 --- a/pkg/sentry/fs/dirent_cache.go +++ b/pkg/sentry/fs/dirent_cache.go @@ -101,8 +101,6 @@ func (c *DirentCache) remove(d *Dirent) { panic(fmt.Sprintf("trying to remove %v, which is not in the dirent cache", d)) } c.list.Remove(d) - d.SetPrev(nil) - d.SetNext(nil) d.DecRef() c.currentSize-- if c.limit != nil { diff --git a/pkg/sentry/fs/fsutil/inode.go b/pkg/sentry/fs/fsutil/inode.go index daecc4ffe..1922ff08c 100644 --- a/pkg/sentry/fs/fsutil/inode.go +++ b/pkg/sentry/fs/fsutil/inode.go @@ -259,8 +259,8 @@ func (i *InodeSimpleExtendedAttributes) ListXattr(context.Context, *fs.Inode, ui // RemoveXattr implements fs.InodeOperations.RemoveXattr. func (i *InodeSimpleExtendedAttributes) RemoveXattr(_ context.Context, _ *fs.Inode, name string) error { - i.mu.RLock() - defer i.mu.RUnlock() + i.mu.Lock() + defer i.mu.Unlock() if _, ok := i.xattrs[name]; ok { delete(i.xattrs, name) return nil diff --git a/pkg/sentry/fs/host/util.go b/pkg/sentry/fs/host/util.go index e37e687c6..7c60dc1db 100644 --- a/pkg/sentry/fs/host/util.go +++ b/pkg/sentry/fs/host/util.go @@ -24,7 +24,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/device" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/syserror" ) @@ -152,9 +152,9 @@ func unstableAttr(mo *superOperations, s *syscall.Stat_t) fs.UnstableAttr { Usage: s.Blocks * 512, Perms: fs.FilePermsFromMode(linux.FileMode(s.Mode)), Owner: owner(mo, s), - AccessTime: ktime.FromUnix(s.Atim.Sec, s.Atim.Nsec), - ModificationTime: ktime.FromUnix(s.Mtim.Sec, s.Mtim.Nsec), - StatusChangeTime: ktime.FromUnix(s.Ctim.Sec, s.Ctim.Nsec), + AccessTime: time.FromUnix(s.Atim.Sec, s.Atim.Nsec), + ModificationTime: time.FromUnix(s.Mtim.Sec, s.Mtim.Nsec), + StatusChangeTime: time.FromUnix(s.Ctim.Sec, s.Ctim.Nsec), Links: uint64(s.Nlink), } } @@ -165,6 +165,8 @@ type dirInfo struct { bufp int // location of next record in buf. } +// LINT.IfChange + // isBlockError unwraps os errors and checks if they are caused by EAGAIN or // EWOULDBLOCK. This is so they can be transformed into syserror.ErrWouldBlock. func isBlockError(err error) bool { @@ -177,6 +179,8 @@ func isBlockError(err error) bool { return false } +// LINT.ThenChange(../../fsimpl/host/util.go) + func hostEffectiveKIDs() (uint32, []uint32, error) { gids, err := os.Getgroups() if err != nil { diff --git a/pkg/sentry/fs/inotify.go b/pkg/sentry/fs/inotify.go index 928c90aa0..e3a715c1f 100644 --- a/pkg/sentry/fs/inotify.go +++ b/pkg/sentry/fs/inotify.go @@ -143,7 +143,10 @@ func (i *Inotify) Read(ctx context.Context, _ *File, dst usermem.IOSequence, _ i } var writeLen int64 - for event := i.events.Front(); event != nil; event = event.Next() { + for it := i.events.Front(); it != nil; { + event := it + it = it.Next() + // Does the buffer have enough remaining space to hold the event we're // about to write out? if dst.NumBytes() < int64(event.sizeOf()) { diff --git a/pkg/sentry/fs/proc/net.go b/pkg/sentry/fs/proc/net.go index 95d5817ff..bd18177d4 100644 --- a/pkg/sentry/fs/proc/net.go +++ b/pkg/sentry/fs/proc/net.go @@ -40,47 +40,48 @@ import ( // LINT.IfChange -// newNet creates a new proc net entry. -func (p *proc) newNetDir(ctx context.Context, k *kernel.Kernel, msrc *fs.MountSource) *fs.Inode { +// newNetDir creates a new proc net entry. +func newNetDir(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + k := t.Kernel() + var contents map[string]*fs.Inode - // TODO(gvisor.dev/issue/1833): Support for using the network stack in the - // network namespace of the calling process. We should make this per-process, - // a.k.a. /proc/PID/net, and make /proc/net a symlink to /proc/self/net. - if s := p.k.RootNetworkNamespace().Stack(); s != nil { + if s := t.NetworkNamespace().Stack(); s != nil { + // TODO(gvisor.dev/issue/1833): Make sure file contents reflect the task + // network namespace. contents = map[string]*fs.Inode{ - "dev": seqfile.NewSeqFileInode(ctx, &netDev{s: s}, msrc), - "snmp": seqfile.NewSeqFileInode(ctx, &netSnmp{s: s}, msrc), + "dev": seqfile.NewSeqFileInode(t, &netDev{s: s}, msrc), + "snmp": seqfile.NewSeqFileInode(t, &netSnmp{s: s}, msrc), // The following files are simple stubs until they are // implemented in netstack, if the file contains a // header the stub is just the header otherwise it is // an empty file. - "arp": newStaticProcInode(ctx, msrc, []byte("IP address HW type Flags HW address Mask Device\n")), + "arp": newStaticProcInode(t, msrc, []byte("IP address HW type Flags HW address Mask Device\n")), - "netlink": newStaticProcInode(ctx, msrc, []byte("sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n")), - "netstat": newStaticProcInode(ctx, msrc, []byte("TcpExt: SyncookiesSent SyncookiesRecv SyncookiesFailed EmbryonicRsts PruneCalled RcvPruned OfoPruned OutOfWindowIcmps LockDroppedIcmps ArpFilter TW TWRecycled TWKilled PAWSPassive PAWSActive PAWSEstab DelayedACKs DelayedACKLocked DelayedACKLost ListenOverflows ListenDrops TCPPrequeued TCPDirectCopyFromBacklog TCPDirectCopyFromPrequeue TCPPrequeueDropped TCPHPHits TCPHPHitsToUser TCPPureAcks TCPHPAcks TCPRenoRecovery TCPSackRecovery TCPSACKReneging TCPFACKReorder TCPSACKReorder TCPRenoReorder TCPTSReorder TCPFullUndo TCPPartialUndo TCPDSACKUndo TCPLossUndo TCPLostRetransmit TCPRenoFailures TCPSackFailures TCPLossFailures TCPFastRetrans TCPForwardRetrans TCPSlowStartRetrans TCPTimeouts TCPLossProbes TCPLossProbeRecovery TCPRenoRecoveryFail TCPSackRecoveryFail TCPSchedulerFailed TCPRcvCollapsed TCPDSACKOldSent TCPDSACKOfoSent TCPDSACKRecv TCPDSACKOfoRecv TCPAbortOnData TCPAbortOnClose TCPAbortOnMemory TCPAbortOnTimeout TCPAbortOnLinger TCPAbortFailed TCPMemoryPressures TCPSACKDiscard TCPDSACKIgnoredOld TCPDSACKIgnoredNoUndo TCPSpuriousRTOs TCPMD5NotFound TCPMD5Unexpected TCPMD5Failure TCPSackShifted TCPSackMerged TCPSackShiftFallback TCPBacklogDrop TCPMinTTLDrop TCPDeferAcceptDrop IPReversePathFilter TCPTimeWaitOverflow TCPReqQFullDoCookies TCPReqQFullDrop TCPRetransFail TCPRcvCoalesce TCPOFOQueue TCPOFODrop TCPOFOMerge TCPChallengeACK TCPSYNChallenge TCPFastOpenActive TCPFastOpenActiveFail TCPFastOpenPassive TCPFastOpenPassiveFail TCPFastOpenListenOverflow TCPFastOpenCookieReqd TCPSpuriousRtxHostQueues BusyPollRxPackets TCPAutoCorking TCPFromZeroWindowAdv TCPToZeroWindowAdv TCPWantZeroWindowAdv TCPSynRetrans TCPOrigDataSent TCPHystartTrainDetect TCPHystartTrainCwnd TCPHystartDelayDetect TCPHystartDelayCwnd TCPACKSkippedSynRecv TCPACKSkippedPAWS TCPACKSkippedSeq TCPACKSkippedFinWait2 TCPACKSkippedTimeWait TCPACKSkippedChallenge TCPWinProbe TCPKeepAlive TCPMTUPFail TCPMTUPSuccess\n")), - "packet": newStaticProcInode(ctx, msrc, []byte("sk RefCnt Type Proto Iface R Rmem User Inode\n")), - "protocols": newStaticProcInode(ctx, msrc, []byte("protocol size sockets memory press maxhdr slab module cl co di ac io in de sh ss gs se re sp bi br ha uh gp em\n")), + "netlink": newStaticProcInode(t, msrc, []byte("sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n")), + "netstat": newStaticProcInode(t, msrc, []byte("TcpExt: SyncookiesSent SyncookiesRecv SyncookiesFailed EmbryonicRsts PruneCalled RcvPruned OfoPruned OutOfWindowIcmps LockDroppedIcmps ArpFilter TW TWRecycled TWKilled PAWSPassive PAWSActive PAWSEstab DelayedACKs DelayedACKLocked DelayedACKLost ListenOverflows ListenDrops TCPPrequeued TCPDirectCopyFromBacklog TCPDirectCopyFromPrequeue TCPPrequeueDropped TCPHPHits TCPHPHitsToUser TCPPureAcks TCPHPAcks TCPRenoRecovery TCPSackRecovery TCPSACKReneging TCPFACKReorder TCPSACKReorder TCPRenoReorder TCPTSReorder TCPFullUndo TCPPartialUndo TCPDSACKUndo TCPLossUndo TCPLostRetransmit TCPRenoFailures TCPSackFailures TCPLossFailures TCPFastRetrans TCPForwardRetrans TCPSlowStartRetrans TCPTimeouts TCPLossProbes TCPLossProbeRecovery TCPRenoRecoveryFail TCPSackRecoveryFail TCPSchedulerFailed TCPRcvCollapsed TCPDSACKOldSent TCPDSACKOfoSent TCPDSACKRecv TCPDSACKOfoRecv TCPAbortOnData TCPAbortOnClose TCPAbortOnMemory TCPAbortOnTimeout TCPAbortOnLinger TCPAbortFailed TCPMemoryPressures TCPSACKDiscard TCPDSACKIgnoredOld TCPDSACKIgnoredNoUndo TCPSpuriousRTOs TCPMD5NotFound TCPMD5Unexpected TCPMD5Failure TCPSackShifted TCPSackMerged TCPSackShiftFallback TCPBacklogDrop TCPMinTTLDrop TCPDeferAcceptDrop IPReversePathFilter TCPTimeWaitOverflow TCPReqQFullDoCookies TCPReqQFullDrop TCPRetransFail TCPRcvCoalesce TCPOFOQueue TCPOFODrop TCPOFOMerge TCPChallengeACK TCPSYNChallenge TCPFastOpenActive TCPFastOpenActiveFail TCPFastOpenPassive TCPFastOpenPassiveFail TCPFastOpenListenOverflow TCPFastOpenCookieReqd TCPSpuriousRtxHostQueues BusyPollRxPackets TCPAutoCorking TCPFromZeroWindowAdv TCPToZeroWindowAdv TCPWantZeroWindowAdv TCPSynRetrans TCPOrigDataSent TCPHystartTrainDetect TCPHystartTrainCwnd TCPHystartDelayDetect TCPHystartDelayCwnd TCPACKSkippedSynRecv TCPACKSkippedPAWS TCPACKSkippedSeq TCPACKSkippedFinWait2 TCPACKSkippedTimeWait TCPACKSkippedChallenge TCPWinProbe TCPKeepAlive TCPMTUPFail TCPMTUPSuccess\n")), + "packet": newStaticProcInode(t, msrc, []byte("sk RefCnt Type Proto Iface R Rmem User Inode\n")), + "protocols": newStaticProcInode(t, msrc, []byte("protocol size sockets memory press maxhdr slab module cl co di ac io in de sh ss gs se re sp bi br ha uh gp em\n")), // Linux sets psched values to: nsec per usec, psched // tick in ns, 1000000, high res timer ticks per sec // (ClockGetres returns 1ns resolution). - "psched": newStaticProcInode(ctx, msrc, []byte(fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)))), - "ptype": newStaticProcInode(ctx, msrc, []byte("Type Device Function\n")), - "route": seqfile.NewSeqFileInode(ctx, &netRoute{s: s}, msrc), - "tcp": seqfile.NewSeqFileInode(ctx, &netTCP{k: k}, msrc), - "udp": seqfile.NewSeqFileInode(ctx, &netUDP{k: k}, msrc), - "unix": seqfile.NewSeqFileInode(ctx, &netUnix{k: k}, msrc), + "psched": newStaticProcInode(t, msrc, []byte(fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)))), + "ptype": newStaticProcInode(t, msrc, []byte("Type Device Function\n")), + "route": seqfile.NewSeqFileInode(t, &netRoute{s: s}, msrc), + "tcp": seqfile.NewSeqFileInode(t, &netTCP{k: k}, msrc), + "udp": seqfile.NewSeqFileInode(t, &netUDP{k: k}, msrc), + "unix": seqfile.NewSeqFileInode(t, &netUnix{k: k}, msrc), } if s.SupportsIPv6() { - contents["if_inet6"] = seqfile.NewSeqFileInode(ctx, &ifinet6{s: s}, msrc) - contents["ipv6_route"] = newStaticProcInode(ctx, msrc, []byte("")) - contents["tcp6"] = seqfile.NewSeqFileInode(ctx, &netTCP6{k: k}, msrc) - contents["udp6"] = newStaticProcInode(ctx, msrc, []byte(" sl local_address remote_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode\n")) + contents["if_inet6"] = seqfile.NewSeqFileInode(t, &ifinet6{s: s}, msrc) + contents["ipv6_route"] = newStaticProcInode(t, msrc, []byte("")) + contents["tcp6"] = seqfile.NewSeqFileInode(t, &netTCP6{k: k}, msrc) + contents["udp6"] = newStaticProcInode(t, msrc, []byte(" sl local_address remote_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode\n")) } } - d := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) - return newProcInode(ctx, d, msrc, fs.SpecialDirectory, nil) + d := ramfs.NewDir(t, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) + return newProcInode(t, d, msrc, fs.SpecialDirectory, t) } // ifinet6 implements seqfile.SeqSource for /proc/net/if_inet6. @@ -837,4 +838,4 @@ func (n *netUDP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se return data, 0 } -// LINT.ThenChange(../../fsimpl/proc/tasks_net.go) +// LINT.ThenChange(../../fsimpl/proc/task_net.go) diff --git a/pkg/sentry/fs/proc/proc.go b/pkg/sentry/fs/proc/proc.go index c8abb5052..c659224a7 100644 --- a/pkg/sentry/fs/proc/proc.go +++ b/pkg/sentry/fs/proc/proc.go @@ -70,6 +70,7 @@ func New(ctx context.Context, msrc *fs.MountSource, cgroupControllers map[string "loadavg": seqfile.NewSeqFileInode(ctx, &loadavgData{}, msrc), "meminfo": seqfile.NewSeqFileInode(ctx, &meminfoData{k}, msrc), "mounts": newProcInode(ctx, ramfs.NewSymlink(ctx, fs.RootOwner, "self/mounts"), msrc, fs.Symlink, nil), + "net": newProcInode(ctx, ramfs.NewSymlink(ctx, fs.RootOwner, "self/net"), msrc, fs.Symlink, nil), "self": newSelf(ctx, pidns, msrc), "stat": seqfile.NewSeqFileInode(ctx, &statData{k}, msrc), "thread-self": newThreadSelf(ctx, pidns, msrc), @@ -86,7 +87,6 @@ func New(ctx context.Context, msrc *fs.MountSource, cgroupControllers map[string } // Add more contents that need proc to be initialized. - p.AddChild(ctx, "net", p.newNetDir(ctx, k, msrc)) p.AddChild(ctx, "sys", p.newSysDir(ctx, msrc)) return newProcInode(ctx, p, msrc, fs.SpecialDirectory, nil), nil diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go index 8ab8d8a02..03cc788c8 100644 --- a/pkg/sentry/fs/proc/task.go +++ b/pkg/sentry/fs/proc/task.go @@ -72,24 +72,27 @@ var _ fs.InodeOperations = (*taskDir)(nil) // newTaskDir creates a new proc task entry. func (p *proc) newTaskDir(t *kernel.Task, msrc *fs.MountSource, isThreadGroup bool) *fs.Inode { contents := map[string]*fs.Inode{ - "auxv": newAuxvec(t, msrc), - "cmdline": newExecArgInode(t, msrc, cmdlineExecArg), - "comm": newComm(t, msrc), - "environ": newExecArgInode(t, msrc, environExecArg), - "exe": newExe(t, msrc), - "fd": newFdDir(t, msrc), - "fdinfo": newFdInfoDir(t, msrc), - "gid_map": newGIDMap(t, msrc), - "io": newIO(t, msrc, isThreadGroup), - "maps": newMaps(t, msrc), - "mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), - "mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), - "ns": newNamespaceDir(t, msrc), - "smaps": newSmaps(t, msrc), - "stat": newTaskStat(t, msrc, isThreadGroup, p.pidns), - "statm": newStatm(t, msrc), - "status": newStatus(t, msrc, p.pidns), - "uid_map": newUIDMap(t, msrc), + "auxv": newAuxvec(t, msrc), + "cmdline": newExecArgInode(t, msrc, cmdlineExecArg), + "comm": newComm(t, msrc), + "environ": newExecArgInode(t, msrc, environExecArg), + "exe": newExe(t, msrc), + "fd": newFdDir(t, msrc), + "fdinfo": newFdInfoDir(t, msrc), + "gid_map": newGIDMap(t, msrc), + "io": newIO(t, msrc, isThreadGroup), + "maps": newMaps(t, msrc), + "mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), + "mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), + "net": newNetDir(t, msrc), + "ns": newNamespaceDir(t, msrc), + "oom_score": newOOMScore(t, msrc), + "oom_score_adj": newOOMScoreAdj(t, msrc), + "smaps": newSmaps(t, msrc), + "stat": newTaskStat(t, msrc, isThreadGroup, p.pidns), + "statm": newStatm(t, msrc), + "status": newStatus(t, msrc, p.pidns), + "uid_map": newUIDMap(t, msrc), } if isThreadGroup { contents["task"] = p.newSubtasks(t, msrc) @@ -796,4 +799,92 @@ func (f *auxvecFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequenc return int64(n), err } +// newOOMScore returns a oom_score file. It is a stub that always returns 0. +// TODO(gvisor.dev/issue/1967) +func newOOMScore(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + return newStaticProcInode(t, msrc, []byte("0\n")) +} + +// oomScoreAdj is a file containing the oom_score adjustment for a task. +// +// +stateify savable +type oomScoreAdj struct { + fsutil.SimpleFileInode + + t *kernel.Task +} + +// +stateify savable +type oomScoreAdjFile struct { + fsutil.FileGenericSeek `state:"nosave"` + fsutil.FileNoIoctl `state:"nosave"` + fsutil.FileNoMMap `state:"nosave"` + fsutil.FileNoSplice `state:"nosave"` + fsutil.FileNoopFlush `state:"nosave"` + fsutil.FileNoopFsync `state:"nosave"` + fsutil.FileNoopRelease `state:"nosave"` + fsutil.FileNotDirReaddir `state:"nosave"` + fsutil.FileUseInodeUnstableAttr `state:"nosave"` + waiter.AlwaysReady `state:"nosave"` + + t *kernel.Task +} + +// newOOMScoreAdj returns a oom_score_adj file. +func newOOMScoreAdj(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + i := &oomScoreAdj{ + SimpleFileInode: *fsutil.NewSimpleFileInode(t, fs.RootOwner, fs.FilePermsFromMode(0644), linux.PROC_SUPER_MAGIC), + t: t, + } + return newProcInode(t, i, msrc, fs.SpecialFile, t) +} + +// Truncate implements fs.InodeOperations.Truncate. Truncate is called when +// O_TRUNC is specified for any kind of existing Dirent but is not called via +// (f)truncate for proc files. +func (*oomScoreAdj) Truncate(context.Context, *fs.Inode, int64) error { + return nil +} + +// GetFile implements fs.InodeOperations.GetFile. +func (o *oomScoreAdj) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { + return fs.NewFile(ctx, dirent, flags, &oomScoreAdjFile{t: o.t}), nil +} + +// Read implements fs.FileOperations.Read. +func (f *oomScoreAdjFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { + if offset != 0 { + return 0, io.EOF + } + adj, err := f.t.OOMScoreAdj() + if err != nil { + return 0, err + } + adjBytes := []byte(strconv.FormatInt(int64(adj), 10) + "\n") + n, err := dst.CopyOut(ctx, adjBytes) + return int64(n), err +} + +// Write implements fs.FileOperations.Write. +func (f *oomScoreAdjFile) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Limit input size so as not to impact performance if input size is large. + src = src.TakeFirst(usermem.PageSize - 1) + + var v int32 + n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts) + if err != nil { + return 0, err + } + + if err := f.t.SetOOMScoreAdj(v); err != nil { + return 0, err + } + + return n, nil +} + // LINT.ThenChange(../../fsimpl/proc/task.go|../../fsimpl/proc/task_files.go) diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go index d00850e25..c4a8f0b38 100644 --- a/pkg/sentry/fsimpl/gofer/gofer.go +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -1045,13 +1045,13 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool // using the old file descriptor, preventing us from safely // closing it. We could handle this by invalidating existing // memmap.Translations, but this is expensive. Instead, use - // dup2() to make the old file descriptor refer to the new file + // dup3 to make the old file descriptor refer to the new file // description, then close the new file descriptor (which is no // longer needed). Racing callers may use the old or new file // description, but this doesn't matter since they refer to the // same file (unless d.fs.opts.overlayfsStaleRead is true, // which we handle separately). - if err := syscall.Dup2(int(h.fd), int(d.handle.fd)); err != nil { + if err := syscall.Dup3(int(h.fd), int(d.handle.fd), 0); err != nil { d.handleMu.Unlock() ctx.Warningf("gofer.dentry.ensureSharedHandle: failed to dup fd %d to fd %d: %v", h.fd, d.handle.fd, err) h.close(ctx) diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go index 54c1031a7..e95209661 100644 --- a/pkg/sentry/fsimpl/gofer/regular_file.go +++ b/pkg/sentry/fsimpl/gofer/regular_file.go @@ -361,8 +361,15 @@ func (rw *dentryReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, erro rw.d.handleMu.RLock() if (rw.d.handle.fd >= 0 && !rw.d.fs.opts.forcePageCache) || rw.d.fs.opts.interop == InteropModeShared || rw.direct { n, err := rw.d.handle.writeFromBlocksAt(rw.ctx, srcs, rw.off) - rw.d.handleMu.RUnlock() rw.off += n + rw.d.dataMu.Lock() + if rw.off > rw.d.size { + atomic.StoreUint64(&rw.d.size, rw.off) + // The remote file's size will implicitly be extended to the correct + // value when we write back to it. + } + rw.d.dataMu.Unlock() + rw.d.handleMu.RUnlock() return n, err } diff --git a/pkg/sentry/fsimpl/host/BUILD b/pkg/sentry/fsimpl/host/BUILD new file mode 100644 index 000000000..731f192b3 --- /dev/null +++ b/pkg/sentry/fsimpl/host/BUILD @@ -0,0 +1,27 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "host", + srcs = [ + "default_file.go", + "host.go", + "util.go", + ], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/log", + "//pkg/refs", + "//pkg/safemem", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/memmap", + "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserror", + "//pkg/usermem", + "@org_golang_x_sys//unix:go_default_library", + ], +) diff --git a/pkg/sentry/fsimpl/host/default_file.go b/pkg/sentry/fsimpl/host/default_file.go new file mode 100644 index 000000000..172cdb161 --- /dev/null +++ b/pkg/sentry/fsimpl/host/default_file.go @@ -0,0 +1,233 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "math" + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// defaultFileFD implements FileDescriptionImpl for non-socket, non-TTY files. +type defaultFileFD struct { + fileDescription + + // canMap specifies whether we allow the file to be memory mapped. + canMap bool + + // mu protects the fields below. + mu sync.Mutex + + // offset specifies the current file offset. + offset int64 +} + +// TODO(gvisor.dev/issue/1672): Implement Waitable interface. + +// PRead implements FileDescriptionImpl. +func (f *defaultFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + return 0, syserror.ESPIPE + } + + return readFromHostFD(ctx, f.inode.hostFD, dst, offset, int(opts.Flags)) +} + +// Read implements FileDescriptionImpl. +func (f *defaultFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + // These files can't be memory mapped, assert this. + if f.canMap { + panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped") + } + + f.mu.Lock() + n, err := readFromHostFD(ctx, f.inode.hostFD, dst, -1, int(opts.Flags)) + f.mu.Unlock() + if isBlockError(err) { + // If we got any data at all, return it as a "completed" partial read + // rather than retrying until complete. + if n != 0 { + err = nil + } else { + err = syserror.ErrWouldBlock + } + } + return n, err + } + // TODO(gvisor.dev/issue/1672): Cache pages, when forced to do so. + f.mu.Lock() + n, err := readFromHostFD(ctx, f.inode.hostFD, dst, f.offset, int(opts.Flags)) + f.offset += n + f.mu.Unlock() + return n, err +} + +func readFromHostFD(ctx context.Context, fd int, dst usermem.IOSequence, offset int64, flags int) (int64, error) { + if flags&^(linux.RWF_VALID) != 0 { + return 0, syserror.EOPNOTSUPP + } + + reader := safemem.FromVecReaderFunc{ + func(srcs [][]byte) (int64, error) { + n, err := unix.Preadv2(fd, srcs, offset, flags) + return int64(n), err + }, + } + n, err := dst.CopyOutFrom(ctx, reader) + return int64(n), err +} + +// PWrite implements FileDescriptionImpl. +func (f *defaultFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + return 0, syserror.ESPIPE + } + + return writeToHostFD(ctx, f.inode.hostFD, src, offset, int(opts.Flags)) +} + +// Write implements FileDescriptionImpl. +func (f *defaultFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + // These files can't be memory mapped, assert this. + if f.canMap { + panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped") + } + + f.mu.Lock() + n, err := writeToHostFD(ctx, f.inode.hostFD, src, -1, int(opts.Flags)) + f.mu.Unlock() + if isBlockError(err) { + err = syserror.ErrWouldBlock + } + return n, err + } + // TODO(gvisor.dev/issue/1672): Cache pages, when forced to do so. + // TODO(gvisor.dev/issue/1672): Write to end of file and update offset if O_APPEND is set on this file. + f.mu.Lock() + n, err := writeToHostFD(ctx, f.inode.hostFD, src, f.offset, int(opts.Flags)) + f.offset += n + f.mu.Unlock() + return n, err +} + +func writeToHostFD(ctx context.Context, fd int, src usermem.IOSequence, offset int64, flags int) (int64, error) { + if flags&^(linux.RWF_VALID) != 0 { + return 0, syserror.EOPNOTSUPP + } + + writer := safemem.FromVecWriterFunc{ + func(srcs [][]byte) (int64, error) { + n, err := unix.Pwritev2(fd, srcs, offset, flags) + return int64(n), err + }, + } + n, err := src.CopyInTo(ctx, writer) + return int64(n), err +} + +// Seek implements FileDescriptionImpl. +// +// Note that we do not support seeking on directories, since we do not even +// allow directory fds to be imported at all. +func (f *defaultFileFD) Seek(_ context.Context, offset int64, whence int32) (int64, error) { + // TODO(b/34716638): Some char devices do support seeking, e.g. /dev/null. + if f.inode.isStream { + return 0, syserror.ESPIPE + } + + f.mu.Lock() + defer f.mu.Unlock() + + switch whence { + case linux.SEEK_SET: + if offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset = offset + + case linux.SEEK_CUR: + // Check for overflow. Note that underflow cannot occur, since f.offset >= 0. + if offset > math.MaxInt64-f.offset { + return f.offset, syserror.EOVERFLOW + } + if f.offset+offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset += offset + + case linux.SEEK_END: + var s syscall.Stat_t + if err := syscall.Fstat(f.inode.hostFD, &s); err != nil { + return f.offset, err + } + size := s.Size + + // Check for overflow. Note that underflow cannot occur, since size >= 0. + if offset > math.MaxInt64-size { + return f.offset, syserror.EOVERFLOW + } + if size+offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset = size + offset + + case linux.SEEK_DATA, linux.SEEK_HOLE: + // Modifying the offset in the host file table should not matter, since + // this is the only place where we use it. + // + // For reading and writing, we always rely on our internal offset. + n, err := unix.Seek(f.inode.hostFD, offset, int(whence)) + if err != nil { + return f.offset, err + } + f.offset = n + + default: + // Invalid whence. + return f.offset, syserror.EINVAL + } + + return f.offset, nil +} + +// Sync implements FileDescriptionImpl. +func (f *defaultFileFD) Sync(context.Context) error { + // TODO(gvisor.dev/issue/1672): Currently we do not support the SyncData optimization, so we always sync everything. + return unix.Fsync(f.inode.hostFD) +} + +// ConfigureMMap implements FileDescriptionImpl. +func (f *defaultFileFD) ConfigureMMap(_ context.Context, opts *memmap.MMapOpts) error { + if !f.canMap { + return syserror.ENODEV + } + // TODO(gvisor.dev/issue/1672): Implement ConfigureMMap and Mappable interface. + return syserror.ENODEV +} diff --git a/pkg/sentry/fsimpl/host/host.go b/pkg/sentry/fsimpl/host/host.go new file mode 100644 index 000000000..c205e6a0b --- /dev/null +++ b/pkg/sentry/fsimpl/host/host.go @@ -0,0 +1,286 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package host provides a filesystem implementation for host files imported as +// file descriptors. +package host + +import ( + "errors" + "fmt" + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" +) + +// filesystem implements vfs.FilesystemImpl. +type filesystem struct { + kernfs.Filesystem +} + +// ImportFD sets up and returns a vfs.FileDescription from a donated fd. +func ImportFD(mnt *vfs.Mount, hostFD int, ownerUID auth.KUID, ownerGID auth.KGID, isTTY bool) (*vfs.FileDescription, error) { + // Must be importing to a mount of host.filesystem. + fs, ok := mnt.Filesystem().Impl().(*filesystem) + if !ok { + return nil, fmt.Errorf("can't import host FDs into filesystems of type %T", mnt.Filesystem().Impl()) + } + + // Retrieve metadata. + var s syscall.Stat_t + if err := syscall.Fstat(hostFD, &s); err != nil { + return nil, err + } + + fileMode := linux.FileMode(s.Mode) + fileType := fileMode.FileType() + // Pipes, character devices, and sockets can return EWOULDBLOCK for + // operations that would block. + isStream := fileType == syscall.S_IFIFO || fileType == syscall.S_IFCHR || fileType == syscall.S_IFSOCK + + i := &inode{ + hostFD: hostFD, + isStream: isStream, + isTTY: isTTY, + ino: fs.NextIno(), + mode: fileMode, + uid: ownerUID, + gid: ownerGID, + } + + d := &kernfs.Dentry{} + d.Init(i) + // i.open will take a reference on d. + defer d.DecRef() + + return i.open(d.VFSDentry(), mnt) +} + +// inode implements kernfs.Inode. +type inode struct { + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + + // When the reference count reaches zero, the host fd is closed. + refs.AtomicRefCount + + // hostFD contains the host fd that this file was originally created from, + // which must be available at time of restore. + // + // This field is initialized at creation time and is immutable. + hostFD int + + // isStream is true if the host fd points to a file representing a stream, + // e.g. a socket or a pipe. Such files are not seekable and can return + // EWOULDBLOCK for I/O operations. + // + // This field is initialized at creation time and is immutable. + isStream bool + + // isTTY is true if this file represents a TTY. + // + // This field is initialized at creation time and is immutable. + isTTY bool + + // ino is an inode number unique within this filesystem. + ino uint64 + + // mu protects the inode metadata below. + mu sync.Mutex + + // mode is the file mode of this inode. Note that this value may become out + // of date if the mode is changed on the host, e.g. with chmod. + mode linux.FileMode + + // uid and gid of the file owner. Note that these refer to the owner of the + // file created on import, not the fd on the host. + uid auth.KUID + gid auth.KGID +} + +// Note that these flags may become out of date, since they can be modified +// on the host, e.g. with fcntl. +func fileFlagsFromHostFD(fd int) (int, error) { + flags, err := unix.FcntlInt(uintptr(fd), syscall.F_GETFL, 0) + if err != nil { + log.Warningf("Failed to get file flags for donated FD %d: %v", fd, err) + return 0, err + } + // TODO(gvisor.dev/issue/1672): implement behavior corresponding to these allowed flags. + flags &= syscall.O_ACCMODE | syscall.O_DIRECT | syscall.O_NONBLOCK | syscall.O_DSYNC | syscall.O_SYNC | syscall.O_APPEND + return flags, nil +} + +// CheckPermissions implements kernfs.Inode. +func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, atx vfs.AccessTypes) error { + return vfs.GenericCheckPermissions(creds, atx, false /* isDir */, uint16(i.mode), i.uid, i.gid) +} + +// Mode implements kernfs.Inode. +func (i *inode) Mode() linux.FileMode { + return i.mode +} + +// Stat implements kernfs.Inode. +func (i *inode) Stat(_ *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + var s unix.Statx_t + if err := unix.Statx(i.hostFD, "", int(unix.AT_EMPTY_PATH|opts.Sync), int(opts.Mask), &s); err != nil { + return linux.Statx{}, err + } + ls := unixToLinuxStatx(s) + + // Use our own internal inode number and file owner. + // + // TODO(gvisor.dev/issue/1672): Use a kernfs-specific device number as well. + // If we use the device number from the host, it may collide with another + // sentry-internal device number. We handle device/inode numbers without + // relying on the host to prevent collisions. + ls.Ino = i.ino + ls.UID = uint32(i.uid) + ls.GID = uint32(i.gid) + + // Update file mode from the host. + i.mode = linux.FileMode(ls.Mode) + + return ls, nil +} + +// SetStat implements kernfs.Inode. +func (i *inode) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { + s := opts.Stat + + m := s.Mask + if m == 0 { + return nil + } + if m&(linux.STATX_UID|linux.STATX_GID) != 0 { + return syserror.EPERM + } + if m&linux.STATX_MODE != 0 { + if err := syscall.Fchmod(i.hostFD, uint32(s.Mode)); err != nil { + return err + } + i.mode = linux.FileMode(s.Mode) + } + if m&linux.STATX_SIZE != 0 { + if err := syscall.Ftruncate(i.hostFD, int64(s.Size)); err != nil { + return err + } + } + if m&(linux.STATX_ATIME|linux.STATX_MTIME) != 0 { + timestamps := []unix.Timespec{ + toTimespec(s.Atime, m&linux.STATX_ATIME == 0), + toTimespec(s.Mtime, m&linux.STATX_MTIME == 0), + } + if err := unix.UtimesNanoAt(i.hostFD, "", timestamps, unix.AT_EMPTY_PATH); err != nil { + return err + } + } + return nil +} + +// DecRef implements kernfs.Inode. +func (i *inode) DecRef() { + i.AtomicRefCount.DecRefWithDestructor(i.Destroy) +} + +// Destroy implements kernfs.Inode. +func (i *inode) Destroy() { + if err := unix.Close(i.hostFD); err != nil { + log.Warningf("failed to close host fd %d: %v", i.hostFD, err) + } +} + +// Open implements kernfs.Inode. +func (i *inode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + return i.open(vfsd, rp.Mount()) +} + +func (i *inode) open(d *vfs.Dentry, mnt *vfs.Mount) (*vfs.FileDescription, error) { + + fileType := i.mode.FileType() + if fileType == syscall.S_IFSOCK { + if i.isTTY { + return nil, errors.New("cannot use host socket as TTY") + } + // TODO(gvisor.dev/issue/1672): support importing sockets. + return nil, errors.New("importing host sockets not supported") + } + + if i.isTTY { + // TODO(gvisor.dev/issue/1672): support importing host fd as TTY. + return nil, errors.New("importing host fd as TTY not supported") + } + + // For simplicity, set offset to 0. Technically, we should + // only set to 0 on files that are not seekable (sockets, pipes, etc.), + // and use the offset from the host fd otherwise. + fd := &defaultFileFD{ + fileDescription: fileDescription{ + inode: i, + }, + canMap: canMap(uint32(fileType)), + mu: sync.Mutex{}, + offset: 0, + } + + vfsfd := &fd.vfsfd + flags, err := fileFlagsFromHostFD(i.hostFD) + if err != nil { + return nil, err + } + + if err := vfsfd.Init(fd, uint32(flags), mnt, d, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return vfsfd, nil +} + +// fileDescription is embedded by host fd implementations of FileDescriptionImpl. +type fileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + + // inode is vfsfd.Dentry().Impl().(*kernfs.Dentry).Inode().(*inode), but + // cached to reduce indirections and casting. fileDescription does not hold + // a reference on the inode through the inode field (since one is already + // held via the Dentry). + // + // inode is immutable after fileDescription creation. + inode *inode +} + +// SetStat implements vfs.FileDescriptionImpl. +func (f *fileDescription) SetStat(_ context.Context, opts vfs.SetStatOptions) error { + return f.inode.SetStat(nil, opts) +} + +// Stat implements vfs.FileDescriptionImpl. +func (f *fileDescription) Stat(_ context.Context, opts vfs.StatOptions) (linux.Statx, error) { + return f.inode.Stat(nil, opts) +} + +// Release implements vfs.FileDescriptionImpl. +func (f *fileDescription) Release() { + // noop +} diff --git a/pkg/sentry/fsimpl/host/util.go b/pkg/sentry/fsimpl/host/util.go new file mode 100644 index 000000000..e1ccacb4d --- /dev/null +++ b/pkg/sentry/fsimpl/host/util.go @@ -0,0 +1,86 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/syserror" +) + +func toTimespec(ts linux.StatxTimestamp, omit bool) unix.Timespec { + if omit { + return unix.Timespec{ + Sec: 0, + Nsec: unix.UTIME_OMIT, + } + } + return unix.Timespec{ + Sec: int64(ts.Sec), + Nsec: int64(ts.Nsec), + } +} + +func unixToLinuxStatx(s unix.Statx_t) linux.Statx { + return linux.Statx{ + Mask: s.Mask, + Blksize: s.Blksize, + Attributes: s.Attributes, + Nlink: s.Nlink, + UID: s.Uid, + GID: s.Gid, + Mode: s.Mode, + Ino: s.Ino, + Size: s.Size, + Blocks: s.Blocks, + AttributesMask: s.Attributes_mask, + Atime: unixToLinuxStatxTimestamp(s.Atime), + Btime: unixToLinuxStatxTimestamp(s.Btime), + Ctime: unixToLinuxStatxTimestamp(s.Ctime), + Mtime: unixToLinuxStatxTimestamp(s.Mtime), + RdevMajor: s.Rdev_major, + RdevMinor: s.Rdev_minor, + DevMajor: s.Dev_major, + DevMinor: s.Dev_minor, + } +} + +func unixToLinuxStatxTimestamp(ts unix.StatxTimestamp) linux.StatxTimestamp { + return linux.StatxTimestamp{Sec: ts.Sec, Nsec: ts.Nsec} +} + +// wouldBlock returns true for file types that can return EWOULDBLOCK +// for blocking operations, e.g. pipes, character devices, and sockets. +func wouldBlock(fileType uint32) bool { + return fileType == syscall.S_IFIFO || fileType == syscall.S_IFCHR || fileType == syscall.S_IFSOCK +} + +// canMap returns true if a file with fileType is allowed to be memory mapped. +// This is ported over from VFS1, but it's probably not the best way for us +// to check if a file can be memory mapped. +func canMap(fileType uint32) bool { + // TODO(gvisor.dev/issue/1672): Also allow "special files" to be mapped (see fs/host:canMap()). + // + // TODO(b/38213152): Some obscure character devices can be mapped. + return fileType == syscall.S_IFREG +} + +// isBlockError checks if an error is EAGAIN or EWOULDBLOCK. +// If so, they can be transformed into syserror.ErrWouldBlock. +func isBlockError(err error) bool { + return err == syserror.EAGAIN || err == syserror.EWOULDBLOCK +} diff --git a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go index d092ccb2a..1c026f4d8 100644 --- a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go +++ b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go @@ -122,7 +122,7 @@ func (fd *DynamicBytesFD) Release() {} // Stat implements vfs.FileDescriptionImpl.Stat. func (fd *DynamicBytesFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { fs := fd.vfsfd.VirtualDentry().Mount().Filesystem() - return fd.inode.Stat(fs), nil + return fd.inode.Stat(fs, opts) } // SetStat implements vfs.FileDescriptionImpl.SetStat. diff --git a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go index 5650512e0..da821d524 100644 --- a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go @@ -107,9 +107,13 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent fs.mu.Lock() defer fs.mu.Unlock() + opts := vfs.StatOptions{Mask: linux.STATX_INO} // Handle ".". if fd.off == 0 { - stat := fd.inode().Stat(vfsFS) + stat, err := fd.inode().Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: ".", Type: linux.DT_DIR, @@ -125,7 +129,10 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent // Handle "..". if fd.off == 1 { parentInode := vfsd.ParentOrSelf().Impl().(*Dentry).inode - stat := parentInode.Stat(vfsFS) + stat, err := parentInode.Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: "..", Type: linux.FileMode(stat.Mode).DirentType(), @@ -146,7 +153,10 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent childIdx := fd.off - 2 for it := fd.children.nthLocked(childIdx); it != nil; it = it.Next() { inode := it.Dentry.Impl().(*Dentry).inode - stat := inode.Stat(vfsFS) + stat, err := inode.Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: it.Name, Type: linux.FileMode(stat.Mode).DirentType(), @@ -190,7 +200,7 @@ func (fd *GenericDirectoryFD) Seek(ctx context.Context, offset int64, whence int func (fd *GenericDirectoryFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { fs := fd.filesystem() inode := fd.inode() - return inode.Stat(fs), nil + return inode.Stat(fs, opts) } // SetStat implements vfs.FileDescriptionImpl.SetStat. diff --git a/pkg/sentry/fsimpl/kernfs/filesystem.go b/pkg/sentry/fsimpl/kernfs/filesystem.go index 292f58afd..1d7e04ad4 100644 --- a/pkg/sentry/fsimpl/kernfs/filesystem.go +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -12,10 +12,10 @@ // See the License for the specific language governing permissions and // limitations under the License. -// This file implements vfs.FilesystemImpl for kernfs. - package kernfs +// This file implements vfs.FilesystemImpl for kernfs. + import ( "fmt" @@ -634,7 +634,7 @@ func (fs *Filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if err != nil { return linux.Statx{}, err } - return inode.Stat(fs.VFSFilesystem()), nil + return inode.Stat(fs.VFSFilesystem(), opts) } // StatFSAt implements vfs.FilesystemImpl.StatFSAt. diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go index 099d70a16..d50018b18 100644 --- a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go @@ -36,20 +36,20 @@ type InodeNoopRefCount struct { } // IncRef implements Inode.IncRef. -func (n *InodeNoopRefCount) IncRef() { +func (InodeNoopRefCount) IncRef() { } // DecRef implements Inode.DecRef. -func (n *InodeNoopRefCount) DecRef() { +func (InodeNoopRefCount) DecRef() { } // TryIncRef implements Inode.TryIncRef. -func (n *InodeNoopRefCount) TryIncRef() bool { +func (InodeNoopRefCount) TryIncRef() bool { return true } // Destroy implements Inode.Destroy. -func (n *InodeNoopRefCount) Destroy() { +func (InodeNoopRefCount) Destroy() { } // InodeDirectoryNoNewChildren partially implements the Inode interface. @@ -58,27 +58,27 @@ func (n *InodeNoopRefCount) Destroy() { type InodeDirectoryNoNewChildren struct{} // NewFile implements Inode.NewFile. -func (*InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewDir implements Inode.NewDir. -func (*InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewLink implements Inode.NewLink. -func (*InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewSymlink implements Inode.NewSymlink. -func (*InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewNode implements Inode.NewNode. -func (*InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } @@ -90,62 +90,62 @@ type InodeNotDirectory struct { } // HasChildren implements Inode.HasChildren. -func (*InodeNotDirectory) HasChildren() bool { +func (InodeNotDirectory) HasChildren() bool { return false } // NewFile implements Inode.NewFile. -func (*InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { panic("NewFile called on non-directory inode") } // NewDir implements Inode.NewDir. -func (*InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { panic("NewDir called on non-directory inode") } // NewLink implements Inode.NewLinkink. -func (*InodeNotDirectory) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { panic("NewLink called on non-directory inode") } // NewSymlink implements Inode.NewSymlink. -func (*InodeNotDirectory) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { panic("NewSymlink called on non-directory inode") } // NewNode implements Inode.NewNode. -func (*InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { panic("NewNode called on non-directory inode") } // Unlink implements Inode.Unlink. -func (*InodeNotDirectory) Unlink(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) Unlink(context.Context, string, *vfs.Dentry) error { panic("Unlink called on non-directory inode") } // RmDir implements Inode.RmDir. -func (*InodeNotDirectory) RmDir(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) RmDir(context.Context, string, *vfs.Dentry) error { panic("RmDir called on non-directory inode") } // Rename implements Inode.Rename. -func (*InodeNotDirectory) Rename(context.Context, string, string, *vfs.Dentry, *vfs.Dentry) (*vfs.Dentry, error) { +func (InodeNotDirectory) Rename(context.Context, string, string, *vfs.Dentry, *vfs.Dentry) (*vfs.Dentry, error) { panic("Rename called on non-directory inode") } // Lookup implements Inode.Lookup. -func (*InodeNotDirectory) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNotDirectory) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { panic("Lookup called on non-directory inode") } // IterDirents implements Inode.IterDirents. -func (*InodeNotDirectory) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { +func (InodeNotDirectory) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { panic("IterDirents called on non-directory inode") } // Valid implements Inode.Valid. -func (*InodeNotDirectory) Valid(context.Context) bool { +func (InodeNotDirectory) Valid(context.Context) bool { return true } @@ -157,17 +157,17 @@ func (*InodeNotDirectory) Valid(context.Context) bool { type InodeNoDynamicLookup struct{} // Lookup implements Inode.Lookup. -func (*InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { return nil, syserror.ENOENT } // IterDirents implements Inode.IterDirents. -func (*InodeNoDynamicLookup) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { +func (InodeNoDynamicLookup) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { return offset, nil } // Valid implements Inode.Valid. -func (*InodeNoDynamicLookup) Valid(ctx context.Context) bool { +func (InodeNoDynamicLookup) Valid(ctx context.Context) bool { return true } @@ -177,7 +177,7 @@ func (*InodeNoDynamicLookup) Valid(ctx context.Context) bool { type InodeNotSymlink struct{} // Readlink implements Inode.Readlink. -func (*InodeNotSymlink) Readlink(context.Context) (string, error) { +func (InodeNotSymlink) Readlink(context.Context) (string, error) { return "", syserror.EINVAL } @@ -219,7 +219,7 @@ func (a *InodeAttrs) Mode() linux.FileMode { // Stat partially implements Inode.Stat. Note that this function doesn't provide // all the stat fields, and the embedder should consider extending the result // with filesystem-specific fields. -func (a *InodeAttrs) Stat(*vfs.Filesystem) linux.Statx { +func (a *InodeAttrs) Stat(*vfs.Filesystem, vfs.StatOptions) (linux.Statx, error) { var stat linux.Statx stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_NLINK stat.Ino = atomic.LoadUint64(&a.ino) @@ -230,7 +230,7 @@ func (a *InodeAttrs) Stat(*vfs.Filesystem) linux.Statx { // TODO: Implement other stat fields like timestamps. - return stat + return stat, nil } // SetStat implements Inode.SetStat. diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go index c74fa999b..a8ab2a2ba 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -176,8 +176,6 @@ type Dentry struct { vfsd vfs.Dentry inode Inode - refs uint64 - // flags caches useful information about the dentry from the inode. See the // dflags* consts above. Must be accessed by atomic ops. flags uint32 @@ -302,7 +300,8 @@ type Inode interface { // this inode. The returned file description should hold a reference on the // inode for its lifetime. // - // Precondition: !rp.Done(). vfsd.Impl() must be a kernfs Dentry. + // Precondition: rp.Done(). vfsd.Impl() must be the kernfs Dentry containing + // the inode on which Open() is being called. Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) } @@ -328,7 +327,7 @@ type inodeMetadata interface { // Stat returns the metadata for this inode. This corresponds to // vfs.FilesystemImpl.StatAt. - Stat(fs *vfs.Filesystem) linux.Statx + Stat(fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) // SetStat updates the metadata for this inode. This corresponds to // vfs.FilesystemImpl.SetStatAt. diff --git a/pkg/sentry/fsimpl/proc/BUILD b/pkg/sentry/fsimpl/proc/BUILD index a83245866..bb609a305 100644 --- a/pkg/sentry/fsimpl/proc/BUILD +++ b/pkg/sentry/fsimpl/proc/BUILD @@ -9,9 +9,9 @@ go_library( "subtasks.go", "task.go", "task_files.go", + "task_net.go", "tasks.go", "tasks_files.go", - "tasks_net.go", "tasks_sys.go", ], visibility = ["//pkg/sentry:internal"], diff --git a/pkg/sentry/fsimpl/proc/subtasks.go b/pkg/sentry/fsimpl/proc/subtasks.go index f3f4e49b4..611645f3f 100644 --- a/pkg/sentry/fsimpl/proc/subtasks.go +++ b/pkg/sentry/fsimpl/proc/subtasks.go @@ -121,8 +121,13 @@ func (i *subtasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.O } // Stat implements kernfs.Inode. -func (i *subtasksInode) Stat(vsfs *vfs.Filesystem) linux.Statx { - stat := i.InodeAttrs.Stat(vsfs) - stat.Nlink += uint32(i.task.ThreadGroup().Count()) - return stat +func (i *subtasksInode) Stat(vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.InodeAttrs.Stat(vsfs, opts) + if err != nil { + return linux.Statx{}, err + } + if opts.Mask&linux.STATX_NLINK != 0 { + stat.Nlink += uint32(i.task.ThreadGroup().Count()) + } + return stat, nil } diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go index 2d814668a..493acbd1b 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -57,16 +57,19 @@ func newTaskInode(inoGen InoGenerator, task *kernel.Task, pidns *kernel.PIDNames "maps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &mapsData{task: task}), //"mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), //"mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), + "net": newTaskNetDir(task, inoGen), "ns": newTaskOwnedDir(task, inoGen.NextIno(), 0511, map[string]*kernfs.Dentry{ "net": newNamespaceSymlink(task, inoGen.NextIno(), "net"), "pid": newNamespaceSymlink(task, inoGen.NextIno(), "pid"), "user": newNamespaceSymlink(task, inoGen.NextIno(), "user"), }), - "smaps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &smapsData{task: task}), - "stat": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), - "statm": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statmData{task: task}), - "status": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statusData{task: task, pidns: pidns}), - "uid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: false}), + "oom_score": newTaskOwnedFile(task, inoGen.NextIno(), 0444, newStaticFile("0\n")), + "oom_score_adj": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &oomScoreAdj{task: task}), + "smaps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &smapsData{task: task}), + "stat": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), + "statm": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statmData{task: task}), + "status": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statusData{task: task, pidns: pidns}), + "uid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: false}), } if isThreadGroup { contents["task"] = newSubtasks(task, pidns, inoGen, cgroupControllers) @@ -152,12 +155,21 @@ func newTaskOwnedDir(task *kernel.Task, ino uint64, perm linux.FileMode, childre } // Stat implements kernfs.Inode. -func (i *taskOwnedInode) Stat(fs *vfs.Filesystem) linux.Statx { - stat := i.Inode.Stat(fs) - uid, gid := i.getOwner(linux.FileMode(stat.Mode)) - stat.UID = uint32(uid) - stat.GID = uint32(gid) - return stat +func (i *taskOwnedInode) Stat(fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.Inode.Stat(fs, opts) + if err != nil { + return linux.Statx{}, err + } + if opts.Mask&(linux.STATX_UID|linux.STATX_GID) != 0 { + uid, gid := i.getOwner(linux.FileMode(stat.Mode)) + if opts.Mask&linux.STATX_UID != 0 { + stat.UID = uint32(uid) + } + if opts.Mask&linux.STATX_GID != 0 { + stat.GID = uint32(gid) + } + } + return stat, nil } // CheckPermissions implements kernfs.Inode. @@ -234,7 +246,7 @@ func newNamespaceSymlink(task *kernel.Task, ino uint64, ns string) *kernfs.Dentr // member, there is one entry containing three colon-separated fields: // hierarchy-ID:controller-list:cgroup-path" func newCgroupData(controllers map[string]string) dynamicInode { - buf := bytes.Buffer{} + var buf bytes.Buffer // The hierarchy ids must be positive integers (for cgroup v1), but the // exact number does not matter, so long as they are unique. We can diff --git a/pkg/sentry/fsimpl/proc/task_files.go b/pkg/sentry/fsimpl/proc/task_files.go index efd3b3453..5a231ac86 100644 --- a/pkg/sentry/fsimpl/proc/task_files.go +++ b/pkg/sentry/fsimpl/proc/task_files.go @@ -525,3 +525,46 @@ func (i *ioData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "cancelled_write_bytes: %d\n", io.BytesWriteCancelled) return nil } + +// oomScoreAdj is a stub of the /proc/<pid>/oom_score_adj file. +// +// +stateify savable +type oomScoreAdj struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ vfs.WritableDynamicBytesSource = (*oomScoreAdj)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (o *oomScoreAdj) Generate(ctx context.Context, buf *bytes.Buffer) error { + adj, err := o.task.OOMScoreAdj() + if err != nil { + return err + } + fmt.Fprintf(buf, "%d\n", adj) + return nil +} + +// Write implements vfs.WritableDynamicBytesSource.Write. +func (o *oomScoreAdj) Write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Limit input size so as not to impact performance if input size is large. + src = src.TakeFirst(usermem.PageSize - 1) + + var v int32 + n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts) + if err != nil { + return 0, err + } + + if err := o.task.SetOOMScoreAdj(v); err != nil { + return 0, err + } + + return n, nil +} diff --git a/pkg/sentry/fsimpl/proc/tasks_net.go b/pkg/sentry/fsimpl/proc/task_net.go index d4e1812d8..373a7b17d 100644 --- a/pkg/sentry/fsimpl/proc/tasks_net.go +++ b/pkg/sentry/fsimpl/proc/task_net.go @@ -37,12 +37,13 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *kernfs.Dentry { +func newTaskNetDir(task *kernel.Task, inoGen InoGenerator) *kernfs.Dentry { + k := task.Kernel() + pidns := task.PIDNamespace() + root := auth.NewRootCredentials(pidns.UserNamespace()) + var contents map[string]*kernfs.Dentry - // TODO(gvisor.dev/issue/1833): Support for using the network stack in the - // network namespace of the calling process. We should make this per-process, - // a.k.a. /proc/PID/net, and make /proc/net a symlink to /proc/self/net. - if stack := k.RootNetworkNamespace().Stack(); stack != nil { + if stack := task.NetworkNamespace().Stack(); stack != nil { const ( arp = "IP address HW type Flags HW address Mask Device\n" netlink = "sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n" @@ -53,6 +54,8 @@ func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *k ) psched := fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)) + // TODO(gvisor.dev/issue/1833): Make sure file contents reflect the task + // network namespace. contents = map[string]*kernfs.Dentry{ "dev": newDentry(root, inoGen.NextIno(), 0444, &netDevData{stack: stack}), "snmp": newDentry(root, inoGen.NextIno(), 0444, &netSnmpData{stack: stack}), @@ -84,7 +87,7 @@ func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *k } } - return kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, contents) + return newTaskOwnedDir(task, inoGen.NextIno(), 0555, contents) } // ifinet6 implements vfs.DynamicBytesSource for /proc/net/if_inet6. diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go index 10c08fa90..d203cebd4 100644 --- a/pkg/sentry/fsimpl/proc/tasks.go +++ b/pkg/sentry/fsimpl/proc/tasks.go @@ -72,7 +72,7 @@ func newTasksInode(inoGen InoGenerator, k *kernel.Kernel, pidns *kernel.PIDNames "sys": newSysDir(root, inoGen, k), "meminfo": newDentry(root, inoGen.NextIno(), 0444, &meminfoData{}), "mounts": kernfs.NewStaticSymlink(root, inoGen.NextIno(), "self/mounts"), - "net": newNetDir(root, inoGen, k), + "net": kernfs.NewStaticSymlink(root, inoGen.NextIno(), "self/net"), "stat": newDentry(root, inoGen.NextIno(), 0444, &statData{k: k}), "uptime": newDentry(root, inoGen.NextIno(), 0444, &uptimeData{}), "version": newDentry(root, inoGen.NextIno(), 0444, &versionData{k: k}), @@ -211,17 +211,22 @@ func (i *tasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.Open return fd.VFSFileDescription(), nil } -func (i *tasksInode) Stat(vsfs *vfs.Filesystem) linux.Statx { - stat := i.InodeAttrs.Stat(vsfs) +func (i *tasksInode) Stat(vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.InodeAttrs.Stat(vsfs, opts) + if err != nil { + return linux.Statx{}, err + } - // Add dynamic children to link count. - for _, tg := range i.pidns.ThreadGroups() { - if leader := tg.Leader(); leader != nil { - stat.Nlink++ + if opts.Mask&linux.STATX_NLINK != 0 { + // Add dynamic children to link count. + for _, tg := range i.pidns.ThreadGroups() { + if leader := tg.Leader(); leader != nil { + stat.Nlink++ + } } } - return stat + return stat, nil } func cpuInfoData(k *kernel.Kernel) string { diff --git a/pkg/sentry/fsimpl/proc/tasks_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go index c5d531fe0..1bb9430c0 100644 --- a/pkg/sentry/fsimpl/proc/tasks_test.go +++ b/pkg/sentry/fsimpl/proc/tasks_test.go @@ -50,7 +50,7 @@ var ( "loadavg": linux.DT_REG, "meminfo": linux.DT_REG, "mounts": linux.DT_LNK, - "net": linux.DT_DIR, + "net": linux.DT_LNK, "self": linux.DT_LNK, "stat": linux.DT_REG, "sys": linux.DT_DIR, @@ -63,21 +63,24 @@ var ( "thread-self": threadSelfLink.NextOff, } taskStaticFiles = map[string]testutil.DirentType{ - "auxv": linux.DT_REG, - "cgroup": linux.DT_REG, - "cmdline": linux.DT_REG, - "comm": linux.DT_REG, - "environ": linux.DT_REG, - "gid_map": linux.DT_REG, - "io": linux.DT_REG, - "maps": linux.DT_REG, - "ns": linux.DT_DIR, - "smaps": linux.DT_REG, - "stat": linux.DT_REG, - "statm": linux.DT_REG, - "status": linux.DT_REG, - "task": linux.DT_DIR, - "uid_map": linux.DT_REG, + "auxv": linux.DT_REG, + "cgroup": linux.DT_REG, + "cmdline": linux.DT_REG, + "comm": linux.DT_REG, + "environ": linux.DT_REG, + "gid_map": linux.DT_REG, + "io": linux.DT_REG, + "maps": linux.DT_REG, + "net": linux.DT_DIR, + "ns": linux.DT_DIR, + "oom_score": linux.DT_REG, + "oom_score_adj": linux.DT_REG, + "smaps": linux.DT_REG, + "stat": linux.DT_REG, + "statm": linux.DT_REG, + "status": linux.DT_REG, + "task": linux.DT_DIR, + "uid_map": linux.DT_REG, } ) diff --git a/pkg/sentry/inet/namespace.go b/pkg/sentry/inet/namespace.go index c16667e7f..029af3025 100644 --- a/pkg/sentry/inet/namespace.go +++ b/pkg/sentry/inet/namespace.go @@ -23,7 +23,10 @@ type Namespace struct { // creator allows kernel to create new network stack for network namespaces. // If nil, no networking will function if network is namespaced. - creator NetworkStackCreator + // + // At afterLoad(), creator will be used to create network stack. Stateify + // needs to wait for this field to be loaded before calling afterLoad(). + creator NetworkStackCreator `state:"wait"` // isRoot indicates whether this is the root network namespace. isRoot bool diff --git a/pkg/sentry/kernel/epoll/epoll_state.go b/pkg/sentry/kernel/epoll/epoll_state.go index a0d35d350..8e9f200d0 100644 --- a/pkg/sentry/kernel/epoll/epoll_state.go +++ b/pkg/sentry/kernel/epoll/epoll_state.go @@ -38,11 +38,14 @@ func (e *EventPoll) afterLoad() { } } - for it := e.waitingList.Front(); it != nil; it = it.Next() { - if it.id.File.Readiness(it.mask) != 0 { - e.waitingList.Remove(it) - e.readyList.PushBack(it) - it.curList = &e.readyList + for it := e.waitingList.Front(); it != nil; { + entry := it + it = it.Next() + + if entry.id.File.Readiness(entry.mask) != 0 { + e.waitingList.Remove(entry) + e.readyList.PushBack(entry) + entry.curList = &e.readyList e.Notify(waiter.EventIn) } } diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 8b76750e9..1d627564f 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -755,6 +755,8 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} { return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: @@ -1481,6 +1483,8 @@ func (ctx supervisorContext) Value(key interface{}) interface{} { return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: diff --git a/pkg/sentry/kernel/pipe/BUILD b/pkg/sentry/kernel/pipe/BUILD index 4c049d5b4..f29dc0472 100644 --- a/pkg/sentry/kernel/pipe/BUILD +++ b/pkg/sentry/kernel/pipe/BUILD @@ -1,25 +1,10 @@ load("//tools:defs.bzl", "go_library", "go_test") -load("//tools/go_generics:defs.bzl", "go_template_instance") package(licenses = ["notice"]) -go_template_instance( - name = "buffer_list", - out = "buffer_list.go", - package = "pipe", - prefix = "buffer", - template = "//pkg/ilist:generic_list", - types = { - "Element": "*buffer", - "Linker": "*buffer", - }, -) - go_library( name = "pipe", srcs = [ - "buffer.go", - "buffer_list.go", "device.go", "node.go", "pipe.go", @@ -33,8 +18,8 @@ go_library( deps = [ "//pkg/abi/linux", "//pkg/amutex", + "//pkg/buffer", "//pkg/context", - "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/device", "//pkg/sentry/fs", @@ -51,7 +36,6 @@ go_test( name = "pipe_test", size = "small", srcs = [ - "buffer_test.go", "node_test.go", "pipe_test.go", ], diff --git a/pkg/sentry/kernel/pipe/buffer.go b/pkg/sentry/kernel/pipe/buffer.go deleted file mode 100644 index fe3be5dbd..000000000 --- a/pkg/sentry/kernel/pipe/buffer.go +++ /dev/null @@ -1,115 +0,0 @@ -// Copyright 2018 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package pipe - -import ( - "io" - - "gvisor.dev/gvisor/pkg/safemem" - "gvisor.dev/gvisor/pkg/sync" -) - -// buffer encapsulates a queueable byte buffer. -// -// Note that the total size is slightly less than two pages. This -// is done intentionally to ensure that the buffer object aligns -// with runtime internals. We have no hard size or alignment -// requirements. This two page size will effectively minimize -// internal fragmentation, but still have a large enough chunk -// to limit excessive segmentation. -// -// +stateify savable -type buffer struct { - data [8144]byte - read int - write int - bufferEntry -} - -// Reset resets internal data. -// -// This must be called before use. -func (b *buffer) Reset() { - b.read = 0 - b.write = 0 -} - -// Empty indicates the buffer is empty. -// -// This indicates there is no data left to read. -func (b *buffer) Empty() bool { - return b.read == b.write -} - -// Full indicates the buffer is full. -// -// This indicates there is no capacity left to write. -func (b *buffer) Full() bool { - return b.write == len(b.data) -} - -// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. -func (b *buffer) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { - dst := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(b.data[b.write:])) - n, err := safemem.CopySeq(dst, srcs) - b.write += int(n) - return n, err -} - -// WriteFromReader writes to the buffer from an io.Reader. -func (b *buffer) WriteFromReader(r io.Reader, count int64) (int64, error) { - dst := b.data[b.write:] - if count < int64(len(dst)) { - dst = b.data[b.write:][:count] - } - n, err := r.Read(dst) - b.write += n - return int64(n), err -} - -// ReadToBlocks implements safemem.Reader.ReadToBlocks. -func (b *buffer) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { - src := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(b.data[b.read:b.write])) - n, err := safemem.CopySeq(dsts, src) - b.read += int(n) - return n, err -} - -// ReadToWriter reads from the buffer into an io.Writer. -func (b *buffer) ReadToWriter(w io.Writer, count int64, dup bool) (int64, error) { - src := b.data[b.read:b.write] - if count < int64(len(src)) { - src = b.data[b.read:][:count] - } - n, err := w.Write(src) - if !dup { - b.read += n - } - return int64(n), err -} - -// bufferPool is a pool for buffers. -var bufferPool = sync.Pool{ - New: func() interface{} { - return new(buffer) - }, -} - -// newBuffer grabs a new buffer from the pool. -func newBuffer() *buffer { - b := bufferPool.Get().(*buffer) - b.Reset() - return b -} diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go index 08410283f..725e9db7d 100644 --- a/pkg/sentry/kernel/pipe/pipe.go +++ b/pkg/sentry/kernel/pipe/pipe.go @@ -20,6 +20,7 @@ import ( "sync/atomic" "syscall" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sync" @@ -70,10 +71,10 @@ type Pipe struct { // mu protects all pipe internal state below. mu sync.Mutex `state:"nosave"` - // data is the buffer queue of pipe contents. + // view is the underlying set of buffers. // // This is protected by mu. - data bufferList + view buffer.View // max is the maximum size of the pipe in bytes. When this max has been // reached, writers will get EWOULDBLOCK. @@ -81,11 +82,6 @@ type Pipe struct { // This is protected by mu. max int64 - // size is the current size of the pipe in bytes. - // - // This is protected by mu. - size int64 - // hadWriter indicates if this pipe ever had a writer. Note that this // does not necessarily indicate there is *currently* a writer, just // that there has been a writer at some point since the pipe was @@ -196,7 +192,7 @@ type readOps struct { limit func(int64) // read performs the actual read operation. - read func(*buffer) (int64, error) + read func(*buffer.View) (int64, error) } // read reads data from the pipe into dst and returns the number of bytes @@ -213,7 +209,7 @@ func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) { defer p.mu.Unlock() // Is the pipe empty? - if p.size == 0 { + if p.view.Size() == 0 { if !p.HasWriters() { // There are no writers, return EOF. return 0, nil @@ -222,71 +218,13 @@ func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) { } // Limit how much we consume. - if ops.left() > p.size { - ops.limit(p.size) + if ops.left() > p.view.Size() { + ops.limit(p.view.Size()) } - done := int64(0) - for ops.left() > 0 { - // Pop the first buffer. - first := p.data.Front() - if first == nil { - break - } - - // Copy user data. - n, err := ops.read(first) - done += int64(n) - p.size -= n - - // Empty buffer? - if first.Empty() { - // Push to the free list. - p.data.Remove(first) - bufferPool.Put(first) - } - - // Handle errors. - if err != nil { - return done, err - } - } - - return done, nil -} - -// dup duplicates all data from this pipe into the given writer. -// -// There is no blocking behavior implemented here. The writer may propagate -// some blocking error. All the writes must be complete writes. -func (p *Pipe) dup(ctx context.Context, ops readOps) (int64, error) { - p.mu.Lock() - defer p.mu.Unlock() - - // Is the pipe empty? - if p.size == 0 { - if !p.HasWriters() { - // See above. - return 0, nil - } - return 0, syserror.ErrWouldBlock - } - - // Limit how much we consume. - if ops.left() > p.size { - ops.limit(p.size) - } - - done := int64(0) - for buf := p.data.Front(); buf != nil; buf = buf.Next() { - n, err := ops.read(buf) - done += n - if err != nil { - return done, err - } - } - - return done, nil + // Copy user data; the read op is responsible for trimming. + done, err := ops.read(&p.view) + return done, err } type writeOps struct { @@ -297,7 +235,7 @@ type writeOps struct { limit func(int64) // write should write to the provided buffer. - write func(*buffer) (int64, error) + write func(*buffer.View) (int64, error) } // write writes data from sv into the pipe and returns the number of bytes @@ -317,33 +255,19 @@ func (p *Pipe) write(ctx context.Context, ops writeOps) (int64, error) { // POSIX requires that a write smaller than atomicIOBytes (PIPE_BUF) be // atomic, but requires no atomicity for writes larger than this. wanted := ops.left() - if avail := p.max - p.size; wanted > avail { + if avail := p.max - p.view.Size(); wanted > avail { if wanted <= p.atomicIOBytes { return 0, syserror.ErrWouldBlock } ops.limit(avail) } - done := int64(0) - for ops.left() > 0 { - // Need a new buffer? - last := p.data.Back() - if last == nil || last.Full() { - // Add a new buffer to the data list. - last = newBuffer() - p.data.PushBack(last) - } - - // Copy user data. - n, err := ops.write(last) - done += int64(n) - p.size += n - - // Handle errors. - if err != nil { - return done, err - } + // Copy user data. + done, err := ops.write(&p.view) + if err != nil { + return done, err } + if wanted > done { // Partial write due to full pipe. return done, syserror.ErrWouldBlock @@ -396,7 +320,7 @@ func (p *Pipe) HasWriters() bool { // Precondition: mu must be held. func (p *Pipe) rReadinessLocked() waiter.EventMask { ready := waiter.EventMask(0) - if p.HasReaders() && p.data.Front() != nil { + if p.HasReaders() && p.view.Size() != 0 { ready |= waiter.EventIn } if !p.HasWriters() && p.hadWriter { @@ -422,7 +346,7 @@ func (p *Pipe) rReadiness() waiter.EventMask { // Precondition: mu must be held. func (p *Pipe) wReadinessLocked() waiter.EventMask { ready := waiter.EventMask(0) - if p.HasWriters() && p.size < p.max { + if p.HasWriters() && p.view.Size() < p.max { ready |= waiter.EventOut } if !p.HasReaders() { @@ -451,7 +375,7 @@ func (p *Pipe) rwReadiness() waiter.EventMask { func (p *Pipe) queued() int64 { p.mu.Lock() defer p.mu.Unlock() - return p.size + return p.view.Size() } // FifoSize implements fs.FifoSizer.FifoSize. @@ -474,7 +398,7 @@ func (p *Pipe) SetFifoSize(size int64) (int64, error) { } p.mu.Lock() defer p.mu.Unlock() - if size < p.size { + if size < p.view.Size() { return 0, syserror.EBUSY } p.max = size diff --git a/pkg/sentry/kernel/pipe/pipe_util.go b/pkg/sentry/kernel/pipe/pipe_util.go index 80158239e..5a1d4fd57 100644 --- a/pkg/sentry/kernel/pipe/pipe_util.go +++ b/pkg/sentry/kernel/pipe/pipe_util.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/amutex" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sync" @@ -49,9 +50,10 @@ func (p *Pipe) Read(ctx context.Context, dst usermem.IOSequence) (int64, error) limit: func(l int64) { dst = dst.TakeFirst64(l) }, - read: func(buf *buffer) (int64, error) { - n, err := dst.CopyOutFrom(ctx, buf) + read: func(view *buffer.View) (int64, error) { + n, err := dst.CopyOutFrom(ctx, view) dst = dst.DropFirst64(n) + view.TrimFront(n) return n, err }, }) @@ -70,16 +72,15 @@ func (p *Pipe) WriteTo(ctx context.Context, w io.Writer, count int64, dup bool) limit: func(l int64) { count = l }, - read: func(buf *buffer) (int64, error) { - n, err := buf.ReadToWriter(w, count, dup) + read: func(view *buffer.View) (int64, error) { + n, err := view.ReadToWriter(w, count) + if !dup { + view.TrimFront(n) + } count -= n return n, err }, } - if dup { - // There is no notification for dup operations. - return p.dup(ctx, ops) - } n, err := p.read(ctx, ops) if n > 0 { p.Notify(waiter.EventOut) @@ -96,8 +97,8 @@ func (p *Pipe) Write(ctx context.Context, src usermem.IOSequence) (int64, error) limit: func(l int64) { src = src.TakeFirst64(l) }, - write: func(buf *buffer) (int64, error) { - n, err := src.CopyInTo(ctx, buf) + write: func(view *buffer.View) (int64, error) { + n, err := src.CopyInTo(ctx, view) src = src.DropFirst64(n) return n, err }, @@ -117,8 +118,8 @@ func (p *Pipe) ReadFrom(ctx context.Context, r io.Reader, count int64) (int64, e limit: func(l int64) { count = l }, - write: func(buf *buffer) (int64, error) { - n, err := buf.WriteFromReader(r, count) + write: func(view *buffer.View) (int64, error) { + n, err := view.WriteFromReader(r, count) count -= n return n, err }, diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index 2cee2e6ed..c0dbbe890 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -37,6 +37,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/usage" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -554,6 +555,13 @@ type Task struct { // // startTime is protected by mu. startTime ktime.Time + + // oomScoreAdj is the task's OOM score adjustment. This is currently not + // used but is maintained for consistency. + // TODO(gvisor.dev/issue/1967) + // + // oomScoreAdj is protected by mu, and is owned by the task goroutine. + oomScoreAdj int32 } func (t *Task) savePtraceTracer() *Task { @@ -847,3 +855,28 @@ func (t *Task) AbstractSockets() *AbstractSocketNamespace { func (t *Task) ContainerID() string { return t.containerID } + +// OOMScoreAdj gets the task's OOM score adjustment. +func (t *Task) OOMScoreAdj() (int32, error) { + t.mu.Lock() + defer t.mu.Unlock() + if t.ExitState() == TaskExitDead { + return 0, syserror.ESRCH + } + return t.oomScoreAdj, nil +} + +// SetOOMScoreAdj sets the task's OOM score adjustment. The value should be +// between -1000 and 1000 inclusive. +func (t *Task) SetOOMScoreAdj(adj int32) error { + t.mu.Lock() + defer t.mu.Unlock() + if t.ExitState() == TaskExitDead { + return syserror.ESRCH + } + if adj > 1000 || adj < -1000 { + return syserror.EINVAL + } + t.oomScoreAdj = adj + return nil +} diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index 78866f280..dda502bb8 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -264,6 +264,11 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { rseqSignature = t.rseqSignature } + adj, err := t.OOMScoreAdj() + if err != nil { + return 0, nil, err + } + cfg := &TaskConfig{ Kernel: t.k, ThreadGroup: tg, @@ -282,6 +287,7 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { RSeqAddr: rseqAddr, RSeqSignature: rseqSignature, ContainerID: t.ContainerID(), + OOMScoreAdj: adj, } if opts.NewThreadGroup { cfg.Parent = t diff --git a/pkg/sentry/kernel/task_run.go b/pkg/sentry/kernel/task_run.go index 5568c91bc..799cbcd93 100644 --- a/pkg/sentry/kernel/task_run.go +++ b/pkg/sentry/kernel/task_run.go @@ -126,13 +126,39 @@ func (t *Task) doStop() { } } +func (*runApp) handleCPUIDInstruction(t *Task) error { + if len(arch.CPUIDInstruction) == 0 { + // CPUID emulation isn't supported, but this code can be + // executed, because the ptrace platform returns + // ErrContextSignalCPUID on page faults too. Look at + // pkg/sentry/platform/ptrace/ptrace.go:context.Switch for more + // details. + return platform.ErrContextSignal + } + // Is this a CPUID instruction? + region := trace.StartRegion(t.traceContext, cpuidRegion) + expected := arch.CPUIDInstruction[:] + found := make([]byte, len(expected)) + _, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found) + if err == nil && bytes.Equal(expected, found) { + // Skip the cpuid instruction. + t.Arch().CPUIDEmulate(t) + t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected))) + region.End() + + return nil + } + region.End() // Not an actual CPUID, but required copy-in. + return platform.ErrContextSignal +} + // The runApp state checks for interrupts before executing untrusted // application code. // // +stateify savable type runApp struct{} -func (*runApp) execute(t *Task) taskRunState { +func (app *runApp) execute(t *Task) taskRunState { if t.interrupted() { // Checkpointing instructs tasks to stop by sending an interrupt, so we // must check for stops before entering runInterrupt (instead of @@ -237,21 +263,10 @@ func (*runApp) execute(t *Task) taskRunState { return (*runApp)(nil) case platform.ErrContextSignalCPUID: - // Is this a CPUID instruction? - region := trace.StartRegion(t.traceContext, cpuidRegion) - expected := arch.CPUIDInstruction[:] - found := make([]byte, len(expected)) - _, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found) - if err == nil && bytes.Equal(expected, found) { - // Skip the cpuid instruction. - t.Arch().CPUIDEmulate(t) - t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected))) - region.End() - + if err := app.handleCPUIDInstruction(t); err == nil { // Resume execution. return (*runApp)(nil) } - region.End() // Not an actual CPUID, but required copy-in. // The instruction at the given RIP was not a CPUID, and we // fallthrough to the default signal deliver behavior below. diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index a5035bb7f..2bbf48bb8 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -93,6 +93,9 @@ type TaskConfig struct { // ContainerID is the container the new task belongs to. ContainerID string + + // oomScoreAdj is the task's OOM score adjustment. + OOMScoreAdj int32 } // NewTask creates a new task defined by cfg. @@ -143,6 +146,7 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { rseqSignature: cfg.RSeqSignature, futexWaiter: futex.NewWaiter(), containerID: cfg.ContainerID, + oomScoreAdj: cfg.OOMScoreAdj, } t.creds.Store(cfg.Credentials) t.endStopCond.L = &t.tg.signalHandlers.mu diff --git a/pkg/sentry/platform/kvm/bluepill.go b/pkg/sentry/platform/kvm/bluepill.go index 35cd55fef..4b23f7803 100644 --- a/pkg/sentry/platform/kvm/bluepill.go +++ b/pkg/sentry/platform/kvm/bluepill.go @@ -81,12 +81,6 @@ func (c *vCPU) die(context *arch.SignalContext64, msg string) { // Save the death message, which will be thrown. c.dieState.message = msg - // Reload all registers to have an accurate stack trace when we return - // to host mode. This means that the stack should be unwound correctly. - if errno := c.getUserRegisters(&c.dieState.guestRegs); errno != 0 { - throw(msg) - } - // Setup the trampoline. dieArchSetup(c, context, &c.dieState.guestRegs) } diff --git a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go index a63a6a071..99cac665d 100644 --- a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go @@ -31,6 +31,12 @@ import ( // //go:nosplit func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) { + // Reload all registers to have an accurate stack trace when we return + // to host mode. This means that the stack should be unwound correctly. + if errno := c.getUserRegisters(&c.dieState.guestRegs); errno != 0 { + throw(c.dieState.message) + } + // If the vCPU is in user mode, we set the stack to the stored stack // value in the vCPU itself. We don't want to unwind the user stack. if guestRegs.RFLAGS&ring0.UserFlagsSet == ring0.UserFlagsSet { diff --git a/pkg/sentry/platform/kvm/bluepill_arm64.s b/pkg/sentry/platform/kvm/bluepill_arm64.s index c61700892..04efa0147 100644 --- a/pkg/sentry/platform/kvm/bluepill_arm64.s +++ b/pkg/sentry/platform/kvm/bluepill_arm64.s @@ -82,6 +82,8 @@ fallback: // dieTrampoline: see bluepill.go, bluepill_arm64_unsafe.go for documentation. TEXT ·dieTrampoline(SB),NOSPLIT,$0 - // TODO(gvisor.dev/issue/1249): dieTrampoline supporting for Arm64. - MOVD R9, 8(RSP) - BL ·dieHandler(SB) + // R0: Fake the old PC as caller + // R1: First argument (vCPU) + MOVD.P R1, 8(RSP) // R1: First argument (vCPU) + MOVD.P R0, 8(RSP) // R0: Fake the old PC as caller + B ·dieHandler(SB) diff --git a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go index 2f02c03cf..195331383 100644 --- a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go @@ -18,9 +18,30 @@ package kvm import ( "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) +// dieArchSetup initialies the state for dieTrampoline. +// +// The arm64 dieTrampoline requires the vCPU to be set in R1, and the last PC +// to be in R0. The trampoline then simulates a call to dieHandler from the +// provided PC. +// //go:nosplit func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) { - // TODO(gvisor.dev/issue/1249): dieTrampoline supporting for Arm64. + // If the vCPU is in user mode, we set the stack to the stored stack + // value in the vCPU itself. We don't want to unwind the user stack. + if guestRegs.Regs.Pstate&ring0.PSR_MODE_MASK == ring0.PSR_MODE_EL0t { + regs := c.CPU.Registers() + context.Regs[0] = regs.Regs[0] + context.Sp = regs.Sp + context.Regs[29] = regs.Regs[29] // stack base address + } else { + context.Regs[0] = guestRegs.Regs.Pc + context.Sp = guestRegs.Regs.Sp + context.Regs[29] = guestRegs.Regs.Regs[29] + context.Pstate = guestRegs.Regs.Pstate + } + context.Regs[1] = uint64(uintptr(unsafe.Pointer(c))) + context.Pc = uint64(dieTrampolineAddr) } diff --git a/pkg/sentry/platform/kvm/kvm.go b/pkg/sentry/platform/kvm/kvm.go index 972ba85c3..a9b4af43e 100644 --- a/pkg/sentry/platform/kvm/kvm.go +++ b/pkg/sentry/platform/kvm/kvm.go @@ -27,6 +27,38 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// userMemoryRegion is a region of physical memory. +// +// This mirrors kvm_memory_region. +type userMemoryRegion struct { + slot uint32 + flags uint32 + guestPhysAddr uint64 + memorySize uint64 + userspaceAddr uint64 +} + +// runData is the run structure. This may be mapped for synchronous register +// access (although that doesn't appear to be supported by my kernel at least). +// +// This mirrors kvm_run. +type runData struct { + requestInterruptWindow uint8 + _ [7]uint8 + + exitReason uint32 + readyForInterruptInjection uint8 + ifFlag uint8 + _ [2]uint8 + + cr8 uint64 + apicBase uint64 + + // This is the union data for exits. Interpretation depends entirely on + // the exitReason above (see vCPU code for more information). + data [32]uint64 +} + // KVM represents a lightweight VM context. type KVM struct { platform.NoCPUPreemptionDetection diff --git a/pkg/sentry/platform/kvm/kvm_amd64.go b/pkg/sentry/platform/kvm/kvm_amd64.go index c5a6f9c7d..093497bc4 100644 --- a/pkg/sentry/platform/kvm/kvm_amd64.go +++ b/pkg/sentry/platform/kvm/kvm_amd64.go @@ -21,17 +21,6 @@ import ( "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) -// userMemoryRegion is a region of physical memory. -// -// This mirrors kvm_memory_region. -type userMemoryRegion struct { - slot uint32 - flags uint32 - guestPhysAddr uint64 - memorySize uint64 - userspaceAddr uint64 -} - // userRegs represents KVM user registers. // // This mirrors kvm_regs. @@ -169,27 +158,6 @@ type modelControlRegisters struct { entries [16]modelControlRegister } -// runData is the run structure. This may be mapped for synchronous register -// access (although that doesn't appear to be supported by my kernel at least). -// -// This mirrors kvm_run. -type runData struct { - requestInterruptWindow uint8 - _ [7]uint8 - - exitReason uint32 - readyForInterruptInjection uint8 - ifFlag uint8 - _ [2]uint8 - - cr8 uint64 - apicBase uint64 - - // This is the union data for exits. Interpretation depends entirely on - // the exitReason above (see vCPU code for more information). - data [32]uint64 -} - // cpuidEntry is a single CPUID entry. // // This mirrors kvm_cpuid_entry2. diff --git a/pkg/sentry/platform/kvm/kvm_arm64.go b/pkg/sentry/platform/kvm/kvm_arm64.go index 2319c86d3..79045651e 100644 --- a/pkg/sentry/platform/kvm/kvm_arm64.go +++ b/pkg/sentry/platform/kvm/kvm_arm64.go @@ -20,17 +20,6 @@ import ( "syscall" ) -// userMemoryRegion is a region of physical memory. -// -// This mirrors kvm_memory_region. -type userMemoryRegion struct { - slot uint32 - flags uint32 - guestPhysAddr uint64 - memorySize uint64 - userspaceAddr uint64 -} - type kvmOneReg struct { id uint64 addr uint64 @@ -53,27 +42,6 @@ type userRegs struct { fpRegs userFpsimdState } -// runData is the run structure. This may be mapped for synchronous register -// access (although that doesn't appear to be supported by my kernel at least). -// -// This mirrors kvm_run. -type runData struct { - requestInterruptWindow uint8 - _ [7]uint8 - - exitReason uint32 - readyForInterruptInjection uint8 - ifFlag uint8 - _ [2]uint8 - - cr8 uint64 - apicBase uint64 - - // This is the union data for exits. Interpretation depends entirely on - // the exitReason above (see vCPU code for more information). - data [32]uint64 -} - // updateGlobalOnce does global initialization. It has to be called only once. func updateGlobalOnce(fd int) error { physicalInit() diff --git a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go index 1c8384e6b..b531f2f85 100644 --- a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go @@ -29,30 +29,6 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -// setMemoryRegion initializes a region. -// -// This may be called from bluepillHandler, and therefore returns an errno -// directly (instead of wrapping in an error) to avoid allocations. -// -//go:nosplit -func (m *machine) setMemoryRegion(slot int, physical, length, virtual uintptr) syscall.Errno { - userRegion := userMemoryRegion{ - slot: uint32(slot), - flags: 0, - guestPhysAddr: uint64(physical), - memorySize: uint64(length), - userspaceAddr: uint64(virtual), - } - - // Set the region. - _, _, errno := syscall.RawSyscall( - syscall.SYS_IOCTL, - uintptr(m.fd), - _KVM_SET_USER_MEMORY_REGION, - uintptr(unsafe.Pointer(&userRegion))) - return errno -} - type kvmVcpuInit struct { target uint32 features [7]uint32 diff --git a/pkg/sentry/platform/ptrace/BUILD b/pkg/sentry/platform/ptrace/BUILD index 95abd321e..30402c2df 100644 --- a/pkg/sentry/platform/ptrace/BUILD +++ b/pkg/sentry/platform/ptrace/BUILD @@ -9,6 +9,7 @@ go_library( "ptrace.go", "ptrace_amd64.go", "ptrace_arm64.go", + "ptrace_arm64_unsafe.go", "ptrace_unsafe.go", "stub_amd64.s", "stub_arm64.s", diff --git a/pkg/sentry/platform/ptrace/ptrace_amd64.go b/pkg/sentry/platform/ptrace/ptrace_amd64.go index db0212538..24fc5dc62 100644 --- a/pkg/sentry/platform/ptrace/ptrace_amd64.go +++ b/pkg/sentry/platform/ptrace/ptrace_amd64.go @@ -31,3 +31,17 @@ func fpRegSet(useXsave bool) uintptr { func stackPointer(r *syscall.PtraceRegs) uintptr { return uintptr(r.Rsp) } + +// x86 use the fs_base register to store the TLS pointer which can be +// get/set in "func (t *thread) get/setRegs(regs *syscall.PtraceRegs)". +// So both of the get/setTLS() operations are noop here. + +// getTLS gets the thread local storage register. +func (t *thread) getTLS(tls *uint64) error { + return nil +} + +// setTLS sets the thread local storage register. +func (t *thread) setTLS(tls *uint64) error { + return nil +} diff --git a/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go b/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go new file mode 100644 index 000000000..32b8a6be9 --- /dev/null +++ b/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go @@ -0,0 +1,62 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package ptrace + +import ( + "syscall" + "unsafe" + + "gvisor.dev/gvisor/pkg/abi/linux" +) + +// getTLS gets the thread local storage register. +func (t *thread) getTLS(tls *uint64) error { + iovec := syscall.Iovec{ + Base: (*byte)(unsafe.Pointer(tls)), + Len: uint64(unsafe.Sizeof(*tls)), + } + _, _, errno := syscall.RawSyscall6( + syscall.SYS_PTRACE, + syscall.PTRACE_GETREGSET, + uintptr(t.tid), + linux.NT_ARM_TLS, + uintptr(unsafe.Pointer(&iovec)), + 0, 0) + if errno != 0 { + return errno + } + return nil +} + +// setTLS sets the thread local storage register. +func (t *thread) setTLS(tls *uint64) error { + iovec := syscall.Iovec{ + Base: (*byte)(unsafe.Pointer(tls)), + Len: uint64(unsafe.Sizeof(*tls)), + } + _, _, errno := syscall.RawSyscall6( + syscall.SYS_PTRACE, + syscall.PTRACE_SETREGSET, + uintptr(t.tid), + linux.NT_ARM_TLS, + uintptr(unsafe.Pointer(&iovec)), + 0, 0) + if errno != 0 { + return errno + } + return nil +} diff --git a/pkg/sentry/platform/ptrace/subprocess.go b/pkg/sentry/platform/ptrace/subprocess.go index 31b7cec53..a644609ef 100644 --- a/pkg/sentry/platform/ptrace/subprocess.go +++ b/pkg/sentry/platform/ptrace/subprocess.go @@ -506,6 +506,9 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { regs := &ac.StateData().Regs t.resetSysemuRegs(regs) + // Extract TLS register + tls := uint64(ac.TLS()) + // Check for interrupts, and ensure that future interrupts will signal t. if !c.interrupt.Enable(t) { // Pending interrupt; simulate. @@ -526,6 +529,9 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { if err := t.setFPRegs(fpState, uint64(fpLen), useXsave); err != nil { panic(fmt.Sprintf("ptrace set fpregs (%+v) failed: %v", fpState, err)) } + if err := t.setTLS(&tls); err != nil { + panic(fmt.Sprintf("ptrace set tls (%+v) failed: %v", tls, err)) + } for { // Start running until the next system call. @@ -555,6 +561,12 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { if err := t.getFPRegs(fpState, uint64(fpLen), useXsave); err != nil { panic(fmt.Sprintf("ptrace get fpregs failed: %v", err)) } + if err := t.getTLS(&tls); err != nil { + panic(fmt.Sprintf("ptrace get tls failed: %v", err)) + } + if !ac.SetTLS(uintptr(tls)) { + panic(fmt.Sprintf("tls value %v is invalid", tls)) + } // Is it a system call? if sig == (syscallEvent | syscall.SIGTRAP) { diff --git a/pkg/sentry/platform/ptrace/subprocess_amd64.go b/pkg/sentry/platform/ptrace/subprocess_amd64.go index e99798c56..cd74945e7 100644 --- a/pkg/sentry/platform/ptrace/subprocess_amd64.go +++ b/pkg/sentry/platform/ptrace/subprocess_amd64.go @@ -21,6 +21,7 @@ import ( "strings" "syscall" + "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/seccomp" "gvisor.dev/gvisor/pkg/sentry/arch" @@ -183,13 +184,76 @@ func enableCpuidFault() { // appendArchSeccompRules append architecture specific seccomp rules when creating BPF program. // Ref attachedThread() for more detail. -func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet { - return append(rules, seccomp.RuleSet{ - Rules: seccomp.SyscallRules{ - syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ - {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, +func appendArchSeccompRules(rules []seccomp.RuleSet, defaultAction linux.BPFAction) []seccomp.RuleSet { + rules = append(rules, + // Rules for trapping vsyscall access. + seccomp.RuleSet{ + Rules: seccomp.SyscallRules{ + syscall.SYS_GETTIMEOFDAY: {}, + syscall.SYS_TIME: {}, + unix.SYS_GETCPU: {}, // SYS_GETCPU was not defined in package syscall on amd64. }, - }, - Action: linux.SECCOMP_RET_ALLOW, - }) + Action: linux.SECCOMP_RET_TRAP, + Vsyscall: true, + }) + if defaultAction != linux.SECCOMP_RET_ALLOW { + rules = append(rules, + seccomp.RuleSet{ + Rules: seccomp.SyscallRules{ + syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ + {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, + }, + }, + Action: linux.SECCOMP_RET_ALLOW, + }) + } + return rules +} + +// probeSeccomp returns true iff seccomp is run after ptrace notifications, +// which is generally the case for kernel version >= 4.8. This check is dynamic +// because kernels have be backported behavior. +// +// See createStub for more information. +// +// Precondition: the runtime OS thread must be locked. +func probeSeccomp() bool { + // Create a completely new, destroyable process. + t, err := attachedThread(0, linux.SECCOMP_RET_ERRNO) + if err != nil { + panic(fmt.Sprintf("seccomp probe failed: %v", err)) + } + defer t.destroy() + + // Set registers to the yield system call. This call is not allowed + // by the filters specified in the attachThread function. + regs := createSyscallRegs(&t.initRegs, syscall.SYS_SCHED_YIELD) + if err := t.setRegs(®s); err != nil { + panic(fmt.Sprintf("ptrace set regs failed: %v", err)) + } + + for { + // Attempt an emulation. + if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, unix.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { + panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) + } + + sig := t.wait(stopped) + if sig == (syscallEvent | syscall.SIGTRAP) { + // Did the seccomp errno hook already run? This would + // indicate that seccomp is first in line and we're + // less than 4.8. + if err := t.getRegs(®s); err != nil { + panic(fmt.Sprintf("ptrace get-regs failed: %v", err)) + } + if _, err := syscallReturnValue(®s); err == nil { + // The seccomp errno mode ran first, and reset + // the error in the registers. + return false + } + // The seccomp hook did not run yet, and therefore it + // is safe to use RET_KILL mode for dispatched calls. + return true + } + } } diff --git a/pkg/sentry/platform/ptrace/subprocess_arm64.go b/pkg/sentry/platform/ptrace/subprocess_arm64.go index 7b975137f..7f5c393f0 100644 --- a/pkg/sentry/platform/ptrace/subprocess_arm64.go +++ b/pkg/sentry/platform/ptrace/subprocess_arm64.go @@ -160,6 +160,15 @@ func enableCpuidFault() { // appendArchSeccompRules append architecture specific seccomp rules when creating BPF program. // Ref attachedThread() for more detail. -func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet { +func appendArchSeccompRules(rules []seccomp.RuleSet, defaultAction linux.BPFAction) []seccomp.RuleSet { return rules } + +// probeSeccomp returns true if seccomp is run after ptrace notifications, +// which is generally the case for kernel version >= 4.8. +// +// On arm64, the support of PTRACE_SYSEMU was added in the 5.3 kernel, so +// probeSeccomp can always return true. +func probeSeccomp() bool { + return true +} diff --git a/pkg/sentry/platform/ptrace/subprocess_linux.go b/pkg/sentry/platform/ptrace/subprocess_linux.go index 74968dfdf..2ce528601 100644 --- a/pkg/sentry/platform/ptrace/subprocess_linux.go +++ b/pkg/sentry/platform/ptrace/subprocess_linux.go @@ -20,7 +20,6 @@ import ( "fmt" "syscall" - "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/procid" @@ -30,54 +29,6 @@ import ( const syscallEvent syscall.Signal = 0x80 -// probeSeccomp returns true iff seccomp is run after ptrace notifications, -// which is generally the case for kernel version >= 4.8. This check is dynamic -// because kernels have be backported behavior. -// -// See createStub for more information. -// -// Precondition: the runtime OS thread must be locked. -func probeSeccomp() bool { - // Create a completely new, destroyable process. - t, err := attachedThread(0, linux.SECCOMP_RET_ERRNO) - if err != nil { - panic(fmt.Sprintf("seccomp probe failed: %v", err)) - } - defer t.destroy() - - // Set registers to the yield system call. This call is not allowed - // by the filters specified in the attachThread function. - regs := createSyscallRegs(&t.initRegs, syscall.SYS_SCHED_YIELD) - if err := t.setRegs(®s); err != nil { - panic(fmt.Sprintf("ptrace set regs failed: %v", err)) - } - - for { - // Attempt an emulation. - if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, unix.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { - panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) - } - - sig := t.wait(stopped) - if sig == (syscallEvent | syscall.SIGTRAP) { - // Did the seccomp errno hook already run? This would - // indicate that seccomp is first in line and we're - // less than 4.8. - if err := t.getRegs(®s); err != nil { - panic(fmt.Sprintf("ptrace get-regs failed: %v", err)) - } - if _, err := syscallReturnValue(®s); err == nil { - // The seccomp errno mode ran first, and reset - // the error in the registers. - return false - } - // The seccomp hook did not run yet, and therefore it - // is safe to use RET_KILL mode for dispatched calls. - return true - } - } -} - // createStub creates a fresh stub processes. // // Precondition: the runtime OS thread must be locked. @@ -123,18 +74,7 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro // stub and all its children. This is used to create child stubs // (below), so we must include the ability to fork, but otherwise lock // down available calls only to what is needed. - rules := []seccomp.RuleSet{ - // Rules for trapping vsyscall access. - { - Rules: seccomp.SyscallRules{ - syscall.SYS_GETTIMEOFDAY: {}, - syscall.SYS_TIME: {}, - unix.SYS_GETCPU: {}, // SYS_GETCPU was not defined in package syscall on amd64. - }, - Action: linux.SECCOMP_RET_TRAP, - Vsyscall: true, - }, - } + rules := []seccomp.RuleSet{} if defaultAction != linux.SECCOMP_RET_ALLOW { rules = append(rules, seccomp.RuleSet{ Rules: seccomp.SyscallRules{ @@ -173,9 +113,8 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro }, Action: linux.SECCOMP_RET_ALLOW, }) - - rules = appendArchSeccompRules(rules) } + rules = appendArchSeccompRules(rules, defaultAction) instrs, err := seccomp.BuildProgram(rules, defaultAction) if err != nil { return nil, err diff --git a/pkg/sentry/platform/ring0/aarch64.go b/pkg/sentry/platform/ring0/aarch64.go index f6da41c27..8122ac6e2 100644 --- a/pkg/sentry/platform/ring0/aarch64.go +++ b/pkg/sentry/platform/ring0/aarch64.go @@ -27,26 +27,27 @@ const ( _PTE_PGT_BASE = 0x7000 _PTE_PGT_SIZE = 0x1000 - _PSR_MODE_EL0t = 0x0 - _PSR_MODE_EL1t = 0x4 - _PSR_MODE_EL1h = 0x5 - _PSR_EL_MASK = 0xf - - _PSR_D_BIT = 0x200 - _PSR_A_BIT = 0x100 - _PSR_I_BIT = 0x80 - _PSR_F_BIT = 0x40 + _PSR_D_BIT = 0x00000200 + _PSR_A_BIT = 0x00000100 + _PSR_I_BIT = 0x00000080 + _PSR_F_BIT = 0x00000040 ) const ( + // PSR bits + PSR_MODE_EL0t = 0x00000000 + PSR_MODE_EL1t = 0x00000004 + PSR_MODE_EL1h = 0x00000005 + PSR_MODE_MASK = 0x0000000f + // KernelFlagsSet should always be set in the kernel. - KernelFlagsSet = _PSR_MODE_EL1h + KernelFlagsSet = PSR_MODE_EL1h // UserFlagsSet are always set in userspace. - UserFlagsSet = _PSR_MODE_EL0t + UserFlagsSet = PSR_MODE_EL0t - KernelFlagsClear = _PSR_EL_MASK - UserFlagsClear = _PSR_EL_MASK + KernelFlagsClear = PSR_MODE_MASK + UserFlagsClear = PSR_MODE_MASK PsrDefaultSet = _PSR_D_BIT | _PSR_A_BIT | _PSR_I_BIT | _PSR_F_BIT ) diff --git a/pkg/sentry/platform/ring0/entry_arm64.s b/pkg/sentry/platform/ring0/entry_arm64.s index baa6c4910..d42eda37b 100644 --- a/pkg/sentry/platform/ring0/entry_arm64.s +++ b/pkg/sentry/platform/ring0/entry_arm64.s @@ -25,10 +25,14 @@ // not available for calls. // +// ERET returns using the ELR and SPSR for the current exception level. #define ERET() \ WORD $0xd69f03e0 +// RSV_REG is a register that holds el1 information temporarily. #define RSV_REG R18_PLATFORM + +// RSV_REG_APP is a register that holds el0 information temporarily. #define RSV_REG_APP R9 #define FPEN_NOTRAP 0x3 @@ -36,6 +40,12 @@ #define FPEN_ENABLE (FPEN_NOTRAP << FPEN_SHIFT) +// Saves a register set. +// +// This is a macro because it may need to executed in contents where a stack is +// not available for calls. +// +// The following registers are not saved: R9, R18. #define REGISTERS_SAVE(reg, offset) \ MOVD R0, offset+PTRACE_R0(reg); \ MOVD R1, offset+PTRACE_R1(reg); \ @@ -67,6 +77,12 @@ MOVD R29, offset+PTRACE_R29(reg); \ MOVD R30, offset+PTRACE_R30(reg); +// Loads a register set. +// +// This is a macro because it may need to executed in contents where a stack is +// not available for calls. +// +// The following registers are not loaded: R9, R18. #define REGISTERS_LOAD(reg, offset) \ MOVD offset+PTRACE_R0(reg), R0; \ MOVD offset+PTRACE_R1(reg), R1; \ @@ -98,7 +114,7 @@ MOVD offset+PTRACE_R29(reg), R29; \ MOVD offset+PTRACE_R30(reg), R30; -//NOP +// NOP-s #define nop31Instructions() \ WORD $0xd503201f; \ WORD $0xd503201f; \ @@ -254,6 +270,7 @@ #define ESR_ELx_WFx_ISS_WFE (UL(1) << 0) #define ESR_ELx_xVC_IMM_MASK ((1UL << 16) - 1) +// LOAD_KERNEL_ADDRESS loads a kernel address. #define LOAD_KERNEL_ADDRESS(from, to) \ MOVD from, to; \ ORR $0xffff000000000000, to, to; @@ -263,15 +280,18 @@ LOAD_KERNEL_ADDRESS(CPU_SELF(from), RSV_REG); \ MOVD $CPU_STACK_TOP(RSV_REG), RSV_REG; \ MOVD RSV_REG, RSP; \ + WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 ISB $15; \ DSB $15; +// SWITCH_TO_APP_PAGETABLE sets a new pagetable for a container application. #define SWITCH_TO_APP_PAGETABLE(from) \ MOVD CPU_TTBR0_APP(from), RSV_REG; \ WORD $0xd5182012; \ // MSR R18, TTBR0_EL1 ISB $15; \ DSB $15; +// SWITCH_TO_KVM_PAGETABLE sets the kvm pagetable. #define SWITCH_TO_KVM_PAGETABLE(from) \ MOVD CPU_TTBR0_KVM(from), RSV_REG; \ WORD $0xd5182012; \ // MSR R18, TTBR0_EL1 @@ -294,6 +314,7 @@ WORD $0xd5181040; \ //MSR R0, CPACR_EL1 ISB $15; +// KERNEL_ENTRY_FROM_EL0 is the entry code of the vcpu from el0 to el1. #define KERNEL_ENTRY_FROM_EL0 \ SUB $16, RSP, RSP; \ // step1, save r18, r9 into kernel temporary stack. STP (RSV_REG, RSV_REG_APP), 16*0(RSP); \ @@ -315,19 +336,22 @@ WORD $0xd5384103; \ // MRS SP_EL0, R3 MOVD R3, PTRACE_SP(RSV_REG_APP); +// KERNEL_ENTRY_FROM_EL1 is the entry code of the vcpu from el1 to el1. #define KERNEL_ENTRY_FROM_EL1 \ WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 - REGISTERS_SAVE(RSV_REG, CPU_REGISTERS); \ // save sentry context + REGISTERS_SAVE(RSV_REG, CPU_REGISTERS); \ // Save sentry context. MOVD RSV_REG_APP, CPU_REGISTERS+PTRACE_R9(RSV_REG); \ WORD $0xd5384004; \ // MRS SPSR_EL1, R4 MOVD R4, CPU_REGISTERS+PTRACE_PSTATE(RSV_REG); \ MRS ELR_EL1, R4; \ MOVD R4, CPU_REGISTERS+PTRACE_PC(RSV_REG); \ MOVD RSP, R4; \ - MOVD R4, CPU_REGISTERS+PTRACE_SP(RSV_REG); + MOVD R4, CPU_REGISTERS+PTRACE_SP(RSV_REG); \ + LOAD_KERNEL_STACK(RSV_REG); // Load the temporary stack. +// Halt halts execution. TEXT ·Halt(SB),NOSPLIT,$0 - // clear bluepill. + // Clear bluepill. WORD $0xd538d092 //MRS TPIDR_EL1, R18 CMP RSV_REG, R9 BNE mmio_exit @@ -341,8 +365,22 @@ mmio_exit: // MMIO_EXIT. MOVD $0, R9 MOVD R0, 0xffff000000001000(R9) - B ·kernelExitToEl1(SB) + RET + +// HaltAndResume halts execution and point the pointer to the resume function. +TEXT ·HaltAndResume(SB),NOSPLIT,$0 + BL ·Halt(SB) + B ·kernelExitToEl1(SB) // Resume. +// HaltEl1SvcAndResume calls Hooks.KernelSyscall and resume. +TEXT ·HaltEl1SvcAndResume(SB),NOSPLIT,$0 + WORD $0xd538d092 // MRS TPIDR_EL1, R18 + MOVD CPU_SELF(RSV_REG), R3 // Load vCPU. + MOVD R3, 8(RSP) // First argument (vCPU). + CALL ·kernelSyscall(SB) // Call the trampoline. + B ·kernelExitToEl1(SB) // Resume. + +// Shutdown stops the guest. TEXT ·Shutdown(SB),NOSPLIT,$0 // PSCI EVENT. MOVD $0x84000009, R0 @@ -429,6 +467,7 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 TEXT ·kernelExitToEl1(SB),NOSPLIT,$0 ERET() +// Start is the CPU entrypoint. TEXT ·Start(SB),NOSPLIT,$0 IRQ_DISABLE MOVD R8, RSV_REG @@ -437,18 +476,23 @@ TEXT ·Start(SB),NOSPLIT,$0 B ·kernelExitToEl1(SB) +// El1_sync_invalid is the handler for an invalid EL1_sync. TEXT ·El1_sync_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_irq_invalid is the handler for an invalid El1_irq. TEXT ·El1_irq_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_fiq_invalid is the handler for an invalid El1_fiq. TEXT ·El1_fiq_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_error_invalid is the handler for an invalid El1_error. TEXT ·El1_error_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_sync is the handler for El1_sync. TEXT ·El1_sync(SB),NOSPLIT,$0 KERNEL_ENTRY_FROM_EL1 WORD $0xd5385219 // MRS ESR_EL1, R25 @@ -484,10 +528,10 @@ el1_da: MOVD $PageFault, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) el1_ia: - B ·Halt(SB) + B ·HaltAndResume(SB) el1_sp_pc: B ·Shutdown(SB) @@ -496,7 +540,9 @@ el1_undef: B ·Shutdown(SB) el1_svc: - B ·Halt(SB) + MOVD $0, CPU_ERROR_CODE(RSV_REG) + MOVD $0, CPU_ERROR_TYPE(RSV_REG) + B ·HaltEl1SvcAndResume(SB) el1_dbg: B ·Shutdown(SB) @@ -508,15 +554,19 @@ el1_fpsimd_acc: el1_invalid: B ·Shutdown(SB) +// El1_irq is the handler for El1_irq. TEXT ·El1_irq(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_fiq is the handler for El1_fiq. TEXT ·El1_fiq(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_error is the handler for El1_error. TEXT ·El1_error(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El0_sync is the handler for El0_sync. TEXT ·El0_sync(SB),NOSPLIT,$0 KERNEL_ENTRY_FROM_EL0 WORD $0xd5385219 // MRS ESR_EL1, R25 @@ -554,7 +604,7 @@ el0_svc: MOVD $Syscall, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) el0_da: WORD $0xd538d092 //MRS TPIDR_EL1, R18 @@ -568,7 +618,7 @@ el0_da: MOVD $PageFault, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) el0_ia: B ·Shutdown(SB) @@ -613,7 +663,7 @@ TEXT ·El0_error(SB),NOSPLIT,$0 MOVD $VirtualizationException, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) TEXT ·El0_sync_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) @@ -627,6 +677,7 @@ TEXT ·El0_fiq_invalid(SB),NOSPLIT,$0 TEXT ·El0_error_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// Vectors implements exception vector table. TEXT ·Vectors(SB),NOSPLIT,$0 B ·El1_sync_invalid(SB) nop31Instructions() diff --git a/pkg/sentry/platform/ring0/kernel_arm64.go b/pkg/sentry/platform/ring0/kernel_arm64.go index c3d341998..ccacaea6b 100644 --- a/pkg/sentry/platform/ring0/kernel_arm64.go +++ b/pkg/sentry/platform/ring0/kernel_arm64.go @@ -16,6 +16,14 @@ package ring0 +// HaltAndResume halts execution and point the pointer to the resume function. +//go:nosplit +func HaltAndResume() + +// HaltEl1SvcAndResume calls Hooks.KernelSyscall and resume. +//go:nosplit +func HaltEl1SvcAndResume() + // init initializes architecture-specific state. func (k *Kernel) init(opts KernelOpts) { // Save the root page tables. diff --git a/pkg/sentry/platform/ring0/pagetables/BUILD b/pkg/sentry/platform/ring0/pagetables/BUILD index 4f2406ce3..581841555 100644 --- a/pkg/sentry/platform/ring0/pagetables/BUILD +++ b/pkg/sentry/platform/ring0/pagetables/BUILD @@ -80,7 +80,7 @@ go_library( "pagetables_amd64.go", "pagetables_arm64.go", "pagetables_x86.go", - "pcids_x86.go", + "pcids.go", "walker_amd64.go", "walker_arm64.go", "walker_empty.go", diff --git a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go b/pkg/sentry/platform/ring0/pagetables/pcids.go index e199bae18..9206030bf 100644 --- a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go +++ b/pkg/sentry/platform/ring0/pagetables/pcids.go @@ -12,8 +12,6 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build i386 amd64 - package pagetables import ( diff --git a/pkg/sentry/sighandling/sighandling.go b/pkg/sentry/sighandling/sighandling.go index ba1f9043d..959ef7217 100644 --- a/pkg/sentry/sighandling/sighandling.go +++ b/pkg/sentry/sighandling/sighandling.go @@ -83,6 +83,10 @@ func StartSignalForwarding(handler func(linux.Signal)) func() { // for their handling. var sigchans []chan os.Signal for sig := 1; sig <= numSignals+1; sig++ { + // SIGURG is used by Go's runtime scheduler. + if sig == int(linux.SIGURG) { + continue + } sigchan := make(chan os.Signal, 1) sigchans = append(sigchans, sigchan) signal.Notify(sigchan, syscall.Signal(sig)) diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index c8a9e67b8..e57ef02a1 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -673,7 +673,7 @@ func parseTarget(filter iptables.IPHeaderFilter, optVal []byte) (iptables.Target // For now, redirect target only supports destination port change. // Port range and IP range are not supported yet. if nfRange.RangeIPV4.Flags&linux.NF_NAT_RANGE_PROTO_SPECIFIED == 0 { - return nil, fmt.Errorf("netfilter.SetEntries: invalid argument.") + return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") } target.RangeProtoSpecified = true @@ -682,7 +682,7 @@ func parseTarget(filter iptables.IPHeaderFilter, optVal []byte) (iptables.Target // TODO(gvisor.dev/issue/170): Port range is not supported yet. if nfRange.RangeIPV4.MinPort != nfRange.RangeIPV4.MaxPort { - return nil, fmt.Errorf("netfilter.SetEntries: invalid argument.") + return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") } // Convert port from big endian to little endian. diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index e187276c5..13a9a60b4 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -712,14 +712,44 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo // Bind implements the linux syscall bind(2) for sockets backed by // tcpip.Endpoint. func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { - addr, family, err := AddressAndFamily(sockaddr) - if err != nil { - return err + if len(sockaddr) < 2 { + return syserr.ErrInvalidArgument } - if err := s.checkFamily(family, true /* exact */); err != nil { - return err + + family := usermem.ByteOrder.Uint16(sockaddr) + var addr tcpip.FullAddress + + // Bind for AF_PACKET requires only family, protocol and ifindex. + // In function AddressAndFamily, we check the address length which is + // not needed for AF_PACKET bind. + if family == linux.AF_PACKET { + var a linux.SockAddrLink + if len(sockaddr) < sockAddrLinkSize { + return syserr.ErrInvalidArgument + } + binary.Unmarshal(sockaddr[:sockAddrLinkSize], usermem.ByteOrder, &a) + + if a.Protocol != uint16(s.protocol) { + return syserr.ErrInvalidArgument + } + + addr = tcpip.FullAddress{ + NIC: tcpip.NICID(a.InterfaceIndex), + Addr: tcpip.Address(a.HardwareAddr[:header.EthernetAddressSize]), + } + } else { + var err *syserr.Error + addr, family, err = AddressAndFamily(sockaddr) + if err != nil { + return err + } + + if err = s.checkFamily(family, true /* exact */); err != nil { + return err + } + + addr = s.mapFamily(addr, family) } - addr = s.mapFamily(addr, family) // Issue the bind request to the endpoint. return syserr.TranslateNetstackError(s.Endpoint.Bind(addr)) @@ -2637,7 +2667,9 @@ func (s *SocketOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, } // Add bytes removed from the endpoint but not yet sent to the caller. + s.readMu.Lock() v += len(s.readView) + s.readMu.Unlock() if v > math.MaxInt32 { v = math.MaxInt32 diff --git a/pkg/sentry/strace/socket.go b/pkg/sentry/strace/socket.go index 51e6d81b2..c0512de89 100644 --- a/pkg/sentry/strace/socket.go +++ b/pkg/sentry/strace/socket.go @@ -632,4 +632,13 @@ var sockOptNames = map[uint64]abi.ValueSet{ linux.MCAST_MSFILTER: "MCAST_MSFILTER", linux.IPV6_ADDRFORM: "IPV6_ADDRFORM", }, + linux.SOL_NETLINK: { + linux.NETLINK_BROADCAST_ERROR: "NETLINK_BROADCAST_ERROR", + linux.NETLINK_CAP_ACK: "NETLINK_CAP_ACK", + linux.NETLINK_DUMP_STRICT_CHK: "NETLINK_DUMP_STRICT_CHK", + linux.NETLINK_EXT_ACK: "NETLINK_EXT_ACK", + linux.NETLINK_LIST_MEMBERSHIPS: "NETLINK_LIST_MEMBERSHIPS", + linux.NETLINK_NO_ENOBUFS: "NETLINK_NO_ENOBUFS", + linux.NETLINK_PKTINFO: "NETLINK_PKTINFO", + }, } diff --git a/pkg/sentry/syscalls/linux/BUILD b/pkg/sentry/syscalls/linux/BUILD index c7883e68e..0d24fd3c4 100644 --- a/pkg/sentry/syscalls/linux/BUILD +++ b/pkg/sentry/syscalls/linux/BUILD @@ -42,6 +42,8 @@ go_library( "sys_socket.go", "sys_splice.go", "sys_stat.go", + "sys_stat_amd64.go", + "sys_stat_arm64.go", "sys_sync.go", "sys_sysinfo.go", "sys_syslog.go", diff --git a/pkg/sentry/syscalls/linux/sys_file.go b/pkg/sentry/syscalls/linux/sys_file.go index c21f14dc0..d10a9bed8 100644 --- a/pkg/sentry/syscalls/linux/sys_file.go +++ b/pkg/sentry/syscalls/linux/sys_file.go @@ -1236,7 +1236,7 @@ func rmdirAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { return syserror.ENOTEMPTY } - if err := fs.MayDelete(t, root, d, name); err != nil { + if err := d.MayDelete(t, root, name); err != nil { return err } @@ -1517,7 +1517,7 @@ func unlinkAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { return syserror.ENOTDIR } - if err := fs.MayDelete(t, root, d, name); err != nil { + if err := d.MayDelete(t, root, name); err != nil { return err } diff --git a/pkg/sentry/syscalls/linux/sys_stat.go b/pkg/sentry/syscalls/linux/sys_stat.go index 701b27b4a..9bd2df104 100644 --- a/pkg/sentry/syscalls/linux/sys_stat.go +++ b/pkg/sentry/syscalls/linux/sys_stat.go @@ -25,24 +25,6 @@ import ( // LINT.IfChange -func statFromAttrs(t *kernel.Task, sattr fs.StableAttr, uattr fs.UnstableAttr) linux.Stat { - return linux.Stat{ - Dev: sattr.DeviceID, - Ino: sattr.InodeID, - Nlink: uattr.Links, - Mode: sattr.Type.LinuxType() | uint32(uattr.Perms.LinuxMode()), - UID: uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow()), - GID: uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow()), - Rdev: uint64(linux.MakeDeviceID(sattr.DeviceFileMajor, sattr.DeviceFileMinor)), - Size: uattr.Size, - Blksize: sattr.BlockSize, - Blocks: uattr.Usage / 512, - ATime: uattr.AccessTime.Timespec(), - MTime: uattr.ModificationTime.Timespec(), - CTime: uattr.StatusChangeTime.Timespec(), - } -} - // Stat implements linux syscall stat(2). func Stat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() diff --git a/pkg/sentry/syscalls/linux/sys_stat_amd64.go b/pkg/sentry/syscalls/linux/sys_stat_amd64.go new file mode 100644 index 000000000..0a04a6113 --- /dev/null +++ b/pkg/sentry/syscalls/linux/sys_stat_amd64.go @@ -0,0 +1,45 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build amd64 + +package linux + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/kernel" +) + +// LINT.IfChange + +func statFromAttrs(t *kernel.Task, sattr fs.StableAttr, uattr fs.UnstableAttr) linux.Stat { + return linux.Stat{ + Dev: sattr.DeviceID, + Ino: sattr.InodeID, + Nlink: uattr.Links, + Mode: sattr.Type.LinuxType() | uint32(uattr.Perms.LinuxMode()), + UID: uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow()), + GID: uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(sattr.DeviceFileMajor, sattr.DeviceFileMinor)), + Size: uattr.Size, + Blksize: sattr.BlockSize, + Blocks: uattr.Usage / 512, + ATime: uattr.AccessTime.Timespec(), + MTime: uattr.ModificationTime.Timespec(), + CTime: uattr.StatusChangeTime.Timespec(), + } +} + +// LINT.ThenChange(vfs2/stat_amd64.go) diff --git a/pkg/sentry/syscalls/linux/sys_stat_arm64.go b/pkg/sentry/syscalls/linux/sys_stat_arm64.go new file mode 100644 index 000000000..5a3b1bfad --- /dev/null +++ b/pkg/sentry/syscalls/linux/sys_stat_arm64.go @@ -0,0 +1,45 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package linux + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/kernel" +) + +// LINT.IfChange + +func statFromAttrs(t *kernel.Task, sattr fs.StableAttr, uattr fs.UnstableAttr) linux.Stat { + return linux.Stat{ + Dev: sattr.DeviceID, + Ino: sattr.InodeID, + Nlink: uint32(uattr.Links), + Mode: sattr.Type.LinuxType() | uint32(uattr.Perms.LinuxMode()), + UID: uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow()), + GID: uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(sattr.DeviceFileMajor, sattr.DeviceFileMinor)), + Size: uattr.Size, + Blksize: int32(sattr.BlockSize), + Blocks: uattr.Usage / 512, + ATime: uattr.AccessTime.Timespec(), + MTime: uattr.ModificationTime.Timespec(), + CTime: uattr.StatusChangeTime.Timespec(), + } +} + +// LINT.ThenChange(vfs2/stat_arm64.go) diff --git a/pkg/sentry/syscalls/linux/vfs2/BUILD b/pkg/sentry/syscalls/linux/vfs2/BUILD index f51761e81..e7695e995 100644 --- a/pkg/sentry/syscalls/linux/vfs2/BUILD +++ b/pkg/sentry/syscalls/linux/vfs2/BUILD @@ -22,6 +22,8 @@ go_library( "read_write.go", "setstat.go", "stat.go", + "stat_amd64.go", + "stat_arm64.go", "sync.go", "xattr.go", ], diff --git a/pkg/sentry/syscalls/linux/vfs2/stat.go b/pkg/sentry/syscalls/linux/vfs2/stat.go index dca8d7011..12c532310 100644 --- a/pkg/sentry/syscalls/linux/vfs2/stat.go +++ b/pkg/sentry/syscalls/linux/vfs2/stat.go @@ -113,29 +113,6 @@ func fstatat(t *kernel.Task, dirfd int32, pathAddr, statAddr usermem.Addr, flags return stat.CopyOut(t, statAddr) } -// This takes both input and output as pointer arguments to avoid copying large -// structs. -func convertStatxToUserStat(t *kernel.Task, statx *linux.Statx, stat *linux.Stat) { - // Linux just copies fields from struct kstat without regard to struct - // kstat::result_mask (fs/stat.c:cp_new_stat()), so we do too. - userns := t.UserNamespace() - *stat = linux.Stat{ - Dev: uint64(linux.MakeDeviceID(uint16(statx.DevMajor), statx.DevMinor)), - Ino: statx.Ino, - Nlink: uint64(statx.Nlink), - Mode: uint32(statx.Mode), - UID: uint32(auth.KUID(statx.UID).In(userns).OrOverflow()), - GID: uint32(auth.KGID(statx.GID).In(userns).OrOverflow()), - Rdev: uint64(linux.MakeDeviceID(uint16(statx.RdevMajor), statx.RdevMinor)), - Size: int64(statx.Size), - Blksize: int64(statx.Blksize), - Blocks: int64(statx.Blocks), - ATime: timespecFromStatxTimestamp(statx.Atime), - MTime: timespecFromStatxTimestamp(statx.Mtime), - CTime: timespecFromStatxTimestamp(statx.Ctime), - } -} - func timespecFromStatxTimestamp(sxts linux.StatxTimestamp) linux.Timespec { return linux.Timespec{ Sec: sxts.Sec, diff --git a/pkg/sentry/syscalls/linux/vfs2/stat_amd64.go b/pkg/sentry/syscalls/linux/vfs2/stat_amd64.go new file mode 100644 index 000000000..2da538fc6 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/stat_amd64.go @@ -0,0 +1,46 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build amd64 + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" +) + +// This takes both input and output as pointer arguments to avoid copying large +// structs. +func convertStatxToUserStat(t *kernel.Task, statx *linux.Statx, stat *linux.Stat) { + // Linux just copies fields from struct kstat without regard to struct + // kstat::result_mask (fs/stat.c:cp_new_stat()), so we do too. + userns := t.UserNamespace() + *stat = linux.Stat{ + Dev: uint64(linux.MakeDeviceID(uint16(statx.DevMajor), statx.DevMinor)), + Ino: statx.Ino, + Nlink: uint64(statx.Nlink), + Mode: uint32(statx.Mode), + UID: uint32(auth.KUID(statx.UID).In(userns).OrOverflow()), + GID: uint32(auth.KGID(statx.GID).In(userns).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(uint16(statx.RdevMajor), statx.RdevMinor)), + Size: int64(statx.Size), + Blksize: int64(statx.Blksize), + Blocks: int64(statx.Blocks), + ATime: timespecFromStatxTimestamp(statx.Atime), + MTime: timespecFromStatxTimestamp(statx.Mtime), + CTime: timespecFromStatxTimestamp(statx.Ctime), + } +} diff --git a/pkg/sentry/syscalls/linux/vfs2/stat_arm64.go b/pkg/sentry/syscalls/linux/vfs2/stat_arm64.go new file mode 100644 index 000000000..88b9c7627 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/stat_arm64.go @@ -0,0 +1,46 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" +) + +// This takes both input and output as pointer arguments to avoid copying large +// structs. +func convertStatxToUserStat(t *kernel.Task, statx *linux.Statx, stat *linux.Stat) { + // Linux just copies fields from struct kstat without regard to struct + // kstat::result_mask (fs/stat.c:cp_new_stat()), so we do too. + userns := t.UserNamespace() + *stat = linux.Stat{ + Dev: uint64(linux.MakeDeviceID(uint16(statx.DevMajor), statx.DevMinor)), + Ino: statx.Ino, + Nlink: uint32(statx.Nlink), + Mode: uint32(statx.Mode), + UID: uint32(auth.KUID(statx.UID).In(userns).OrOverflow()), + GID: uint32(auth.KGID(statx.GID).In(userns).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(uint16(statx.RdevMajor), statx.RdevMinor)), + Size: int64(statx.Size), + Blksize: int32(statx.Blksize), + Blocks: int64(statx.Blocks), + ATime: timespecFromStatxTimestamp(statx.Atime), + MTime: timespecFromStatxTimestamp(statx.Mtime), + CTime: timespecFromStatxTimestamp(statx.Ctime), + } +} diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index 07c8383e6..cb4deb068 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -42,10 +42,13 @@ go_library( deps = [ "//pkg/abi/linux", "//pkg/context", + "//pkg/fd", "//pkg/fspath", "//pkg/gohacks", "//pkg/log", + "//pkg/safemem", "//pkg/sentry/arch", + "//pkg/sentry/fs", "//pkg/sentry/fs/lock", "//pkg/sentry/kernel/auth", "//pkg/sentry/memmap", @@ -53,6 +56,7 @@ go_library( "//pkg/syserror", "//pkg/usermem", "//pkg/waiter", + "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/vfs/file_description_impl_util.go b/pkg/sentry/vfs/file_description_impl_util.go index c2a52ec1b..45191d1c3 100644 --- a/pkg/sentry/vfs/file_description_impl_util.go +++ b/pkg/sentry/vfs/file_description_impl_util.go @@ -33,8 +33,8 @@ import ( // implementations to adapt: // - Have a local fileDescription struct (containing FileDescription) which // embeds FileDescriptionDefaultImpl and overrides the default methods -// which are common to all fd implementations for that for that filesystem -// like StatusFlags, SetStatusFlags, Stat, SetStat, StatFS, etc. +// which are common to all fd implementations for that filesystem like +// StatusFlags, SetStatusFlags, Stat, SetStat, StatFS, etc. // - This should be embedded in all file description implementations as the // first field by value. // - Directory FDs would also embed DirectoryFileDescriptionDefaultImpl. diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go index 9912df799..31a4e5480 100644 --- a/pkg/sentry/vfs/mount.go +++ b/pkg/sentry/vfs/mount.go @@ -139,6 +139,23 @@ func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth return mntns, nil } +// NewDisconnectedMount returns a Mount representing fs with the given root +// (which may be nil). The new Mount is not associated with any MountNamespace +// and is not connected to any other Mounts. References are taken on fs and +// root. +func (vfs *VirtualFilesystem) NewDisconnectedMount(fs *Filesystem, root *Dentry, opts *MountOptions) (*Mount, error) { + fs.IncRef() + if root != nil { + root.IncRef() + } + return &Mount{ + vfs: vfs, + fs: fs, + root: root, + refs: 1, + }, nil +} + // MountAt creates and mounts a Filesystem configured by the given arguments. func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentials, source string, target *PathOperation, fsTypeName string, opts *MountOptions) error { rft := vfs.getFilesystemType(fsTypeName) diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go index 73f8043be..bde81e1ef 100644 --- a/pkg/sentry/vfs/vfs.go +++ b/pkg/sentry/vfs/vfs.go @@ -126,17 +126,23 @@ func (vfs *VirtualFilesystem) Init() error { // Construct vfs.anonMount. anonfsDevMinor, err := vfs.GetAnonBlockDevMinor() if err != nil { - return err + // This shouldn't be possible since anonBlockDevMinorNext was + // initialized to 1 above (no device numbers have been allocated yet). + panic(fmt.Sprintf("VirtualFilesystem.Init: device number allocation for anonfs failed: %v", err)) } anonfs := anonFilesystem{ devMinor: anonfsDevMinor, } anonfs.vfsfs.Init(vfs, &anonfs) - vfs.anonMount = &Mount{ - vfs: vfs, - fs: &anonfs.vfsfs, - refs: 1, + defer anonfs.vfsfs.DecRef() + anonMount, err := vfs.NewDisconnectedMount(&anonfs.vfsfs, nil, &MountOptions{}) + if err != nil { + // We should not be passing any MountOptions that would cause + // construction of this mount to fail. + panic(fmt.Sprintf("VirtualFilesystem.Init: anonfs mount failed: %v", err)) } + vfs.anonMount = anonMount + return nil } diff --git a/pkg/sentry/watchdog/watchdog.go b/pkg/sentry/watchdog/watchdog.go index bfb2fac26..f7d6009a0 100644 --- a/pkg/sentry/watchdog/watchdog.go +++ b/pkg/sentry/watchdog/watchdog.go @@ -221,7 +221,7 @@ func (w *Watchdog) waitForStart() { return } var buf bytes.Buffer - buf.WriteString("Watchdog.Start() not called within %s:\n") + buf.WriteString(fmt.Sprintf("Watchdog.Start() not called within %s", w.StartupTimeout)) w.doAction(w.StartupTimeoutAction, false, &buf) } @@ -325,7 +325,7 @@ func (w *Watchdog) report(offenders map[*kernel.Task]*offender, newTaskFound boo func (w *Watchdog) reportStuckWatchdog() { var buf bytes.Buffer - buf.WriteString("Watchdog goroutine is stuck:\n") + buf.WriteString("Watchdog goroutine is stuck:") w.doAction(w.TaskTimeoutAction, false, &buf) } @@ -359,7 +359,7 @@ func (w *Watchdog) doAction(action Action, skipStack bool, msg *bytes.Buffer) { case <-metricsEmitted: case <-time.After(1 * time.Second): } - panic(fmt.Sprintf("Stack for running G's are skipped while panicking.\n%s", msg.String())) + panic(fmt.Sprintf("%s\nStack for running G's are skipped while panicking.", msg.String())) default: panic(fmt.Sprintf("Unknown watchdog action %v", action)) diff --git a/pkg/tcpip/checker/checker.go b/pkg/tcpip/checker/checker.go index c6c160dfc..8dc0f7c0e 100644 --- a/pkg/tcpip/checker/checker.go +++ b/pkg/tcpip/checker/checker.go @@ -785,6 +785,52 @@ func NDPNSTargetAddress(want tcpip.Address) TransportChecker { } } +// ndpOptions checks that optsBuf only contains opts. +func ndpOptions(t *testing.T, optsBuf header.NDPOptions, opts []header.NDPOption) { + t.Helper() + + it, err := optsBuf.Iter(true) + if err != nil { + t.Errorf("optsBuf.Iter(true): %s", err) + return + } + + i := 0 + for { + opt, done, err := it.Next() + if err != nil { + // This should never happen as Iter(true) above did not return an error. + t.Fatalf("unexpected error when iterating over NDP options: %s", err) + } + if done { + break + } + + if i >= len(opts) { + t.Errorf("got unexpected option: %s", opt) + continue + } + + switch wantOpt := opts[i].(type) { + case header.NDPSourceLinkLayerAddressOption: + gotOpt, ok := opt.(header.NDPSourceLinkLayerAddressOption) + if !ok { + t.Errorf("got type = %T at index = %d; want = %T", opt, i, wantOpt) + } else if got, want := gotOpt.EthernetAddress(), wantOpt.EthernetAddress(); got != want { + t.Errorf("got EthernetAddress() = %s at index %d, want = %s", got, i, want) + } + default: + t.Fatalf("checker not implemented for expected NDP option: %T", wantOpt) + } + + i++ + } + + if missing := opts[i:]; len(missing) > 0 { + t.Errorf("missing options: %s", missing) + } +} + // NDPNSOptions creates a checker that checks that the packet contains the // provided NDP options within an NDP Neighbor Solicitation message. // @@ -796,47 +842,31 @@ func NDPNSOptions(opts []header.NDPOption) TransportChecker { icmp := h.(header.ICMPv6) ns := header.NDPNeighborSolicit(icmp.NDPPayload()) - it, err := ns.Options().Iter(true) - if err != nil { - t.Errorf("opts.Iter(true): %s", err) - return - } - - i := 0 - for { - opt, done, _ := it.Next() - if done { - break - } - - if i >= len(opts) { - t.Errorf("got unexpected option: %s", opt) - continue - } - - switch wantOpt := opts[i].(type) { - case header.NDPSourceLinkLayerAddressOption: - gotOpt, ok := opt.(header.NDPSourceLinkLayerAddressOption) - if !ok { - t.Errorf("got type = %T at index = %d; want = %T", opt, i, wantOpt) - } else if got, want := gotOpt.EthernetAddress(), wantOpt.EthernetAddress(); got != want { - t.Errorf("got EthernetAddress() = %s at index %d, want = %s", got, i, want) - } - default: - panic("not implemented") - } - - i++ - } - - if missing := opts[i:]; len(missing) > 0 { - t.Errorf("missing options: %s", missing) - } + ndpOptions(t, ns.Options(), opts) } } // NDPRS creates a checker that checks that the packet contains a valid NDP // Router Solicitation message (as per the raw wire format). -func NDPRS() NetworkChecker { - return NDP(header.ICMPv6RouterSolicit, header.NDPRSMinimumSize) +// +// checkers may assume that a valid ICMPv6 is passed to it containing a valid +// NDPRS as far as the size of the message is concerned. The values within the +// message are up to checkers to validate. +func NDPRS(checkers ...TransportChecker) NetworkChecker { + return NDP(header.ICMPv6RouterSolicit, header.NDPRSMinimumSize, checkers...) +} + +// NDPRSOptions creates a checker that checks that the packet contains the +// provided NDP options within an NDP Router Solicitation message. +// +// The returned TransportChecker assumes that a valid ICMPv6 is passed to it +// containing a valid NDPRS message as far as the size is concerned. +func NDPRSOptions(opts []header.NDPOption) TransportChecker { + return func(t *testing.T, h header.Transport) { + t.Helper() + + icmp := h.(header.ICMPv6) + rs := header.NDPRouterSolicit(icmp.NDPPayload()) + ndpOptions(t, rs.Options(), opts) + } } diff --git a/pkg/tcpip/header/ipv6.go b/pkg/tcpip/header/ipv6.go index 70e6ce095..76e88e9b3 100644 --- a/pkg/tcpip/header/ipv6.go +++ b/pkg/tcpip/header/ipv6.go @@ -115,6 +115,19 @@ const ( // for the secret key used to generate an opaque interface identifier as // outlined by RFC 7217. OpaqueIIDSecretKeyMinBytes = 16 + + // ipv6MulticastAddressScopeByteIdx is the byte where the scope (scop) field + // is located within a multicast IPv6 address, as per RFC 4291 section 2.7. + ipv6MulticastAddressScopeByteIdx = 1 + + // ipv6MulticastAddressScopeMask is the mask for the scope (scop) field, + // within the byte holding the field, as per RFC 4291 section 2.7. + ipv6MulticastAddressScopeMask = 0xF + + // ipv6LinkLocalMulticastScope is the value of the scope (scop) field within + // a multicast IPv6 address that indicates the address has link-local scope, + // as per RFC 4291 section 2.7. + ipv6LinkLocalMulticastScope = 2 ) // IPv6EmptySubnet is the empty IPv6 subnet. It may also be known as the @@ -340,6 +353,12 @@ func IsV6LinkLocalAddress(addr tcpip.Address) bool { return addr[0] == 0xfe && (addr[1]&0xc0) == 0x80 } +// IsV6LinkLocalMulticastAddress determines if the provided address is an IPv6 +// link-local multicast address. +func IsV6LinkLocalMulticastAddress(addr tcpip.Address) bool { + return IsV6MulticastAddress(addr) && addr[ipv6MulticastAddressScopeByteIdx]&ipv6MulticastAddressScopeMask == ipv6LinkLocalMulticastScope +} + // IsV6UniqueLocalAddress determines if the provided address is an IPv6 // unique-local address (within the prefix FC00::/7). func IsV6UniqueLocalAddress(addr tcpip.Address) bool { @@ -411,6 +430,9 @@ func ScopeForIPv6Address(addr tcpip.Address) (IPv6AddressScope, *tcpip.Error) { } switch { + case IsV6LinkLocalMulticastAddress(addr): + return LinkLocalScope, nil + case IsV6LinkLocalAddress(addr): return LinkLocalScope, nil diff --git a/pkg/tcpip/header/ipv6_test.go b/pkg/tcpip/header/ipv6_test.go index c3ad503aa..426a873b1 100644 --- a/pkg/tcpip/header/ipv6_test.go +++ b/pkg/tcpip/header/ipv6_test.go @@ -27,11 +27,12 @@ import ( ) const ( - linkAddr = tcpip.LinkAddress("\x02\x02\x03\x04\x05\x06") - linkLocalAddr = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") - uniqueLocalAddr1 = tcpip.Address("\xfc\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") - uniqueLocalAddr2 = tcpip.Address("\xfd\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") - globalAddr = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + linkAddr = tcpip.LinkAddress("\x02\x02\x03\x04\x05\x06") + linkLocalAddr = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + linkLocalMulticastAddr = tcpip.Address("\xff\x02\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + uniqueLocalAddr1 = tcpip.Address("\xfc\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + uniqueLocalAddr2 = tcpip.Address("\xfd\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") + globalAddr = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") ) func TestEthernetAdddressToModifiedEUI64(t *testing.T) { @@ -256,6 +257,85 @@ func TestIsV6UniqueLocalAddress(t *testing.T) { } } +func TestIsV6LinkLocalMulticastAddress(t *testing.T) { + tests := []struct { + name string + addr tcpip.Address + expected bool + }{ + { + name: "Valid Link Local Multicast", + addr: linkLocalMulticastAddr, + expected: true, + }, + { + name: "Valid Link Local Multicast with flags", + addr: "\xff\xf2\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01", + expected: true, + }, + { + name: "Link Local Unicast", + addr: linkLocalAddr, + expected: false, + }, + { + name: "IPv4 Multicast", + addr: "\xe0\x00\x00\x01", + expected: false, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + if got := header.IsV6LinkLocalMulticastAddress(test.addr); got != test.expected { + t.Errorf("got header.IsV6LinkLocalMulticastAddress(%s) = %t, want = %t", test.addr, got, test.expected) + } + }) + } +} + +func TestIsV6LinkLocalAddress(t *testing.T) { + tests := []struct { + name string + addr tcpip.Address + expected bool + }{ + { + name: "Valid Link Local Unicast", + addr: linkLocalAddr, + expected: true, + }, + { + name: "Link Local Multicast", + addr: linkLocalMulticastAddr, + expected: false, + }, + { + name: "Unique Local", + addr: uniqueLocalAddr1, + expected: false, + }, + { + name: "Global", + addr: globalAddr, + expected: false, + }, + { + name: "IPv4 Link Local", + addr: "\xa9\xfe\x00\x01", + expected: false, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + if got := header.IsV6LinkLocalAddress(test.addr); got != test.expected { + t.Errorf("got header.IsV6LinkLocalAddress(%s) = %t, want = %t", test.addr, got, test.expected) + } + }) + } +} + func TestScopeForIPv6Address(t *testing.T) { tests := []struct { name string @@ -270,12 +350,18 @@ func TestScopeForIPv6Address(t *testing.T) { err: nil, }, { - name: "Link Local", + name: "Link Local Unicast", addr: linkLocalAddr, scope: header.LinkLocalScope, err: nil, }, { + name: "Link Local Multicast", + addr: linkLocalMulticastAddr, + scope: header.LinkLocalScope, + err: nil, + }, + { name: "Global", addr: globalAddr, scope: header.GlobalScope, diff --git a/pkg/tcpip/network/fragmentation/fragmentation.go b/pkg/tcpip/network/fragmentation/fragmentation.go index 92f2aa13a..f42abc4bb 100644 --- a/pkg/tcpip/network/fragmentation/fragmentation.go +++ b/pkg/tcpip/network/fragmentation/fragmentation.go @@ -115,10 +115,12 @@ func (f *Fragmentation) Process(id uint32, first, last uint16, more bool, vv buf // Evict reassemblers if we are consuming more memory than highLimit until // we reach lowLimit. if f.size > f.highLimit { - tail := f.rList.Back() - for f.size > f.lowLimit && tail != nil { + for f.size > f.lowLimit { + tail := f.rList.Back() + if tail == nil { + break + } f.release(tail) - tail = tail.Prev() } } f.mu.Unlock() diff --git a/pkg/tcpip/stack/BUILD b/pkg/tcpip/stack/BUILD index 705cf01ee..6c029b2fb 100644 --- a/pkg/tcpip/stack/BUILD +++ b/pkg/tcpip/stack/BUILD @@ -18,6 +18,8 @@ go_template_instance( go_library( name = "stack", srcs = [ + "dhcpv6configurationfromndpra_string.go", + "forwarder.go", "icmp_rate_limit.go", "linkaddrcache.go", "linkaddrentry_list.go", @@ -79,6 +81,7 @@ go_test( name = "stack_test", size = "small", srcs = [ + "forwarder_test.go", "linkaddrcache_test.go", "nic_test.go", ], diff --git a/pkg/tcpip/stack/dhcpv6configurationfromndpra_string.go b/pkg/tcpip/stack/dhcpv6configurationfromndpra_string.go new file mode 100644 index 000000000..8b4213eec --- /dev/null +++ b/pkg/tcpip/stack/dhcpv6configurationfromndpra_string.go @@ -0,0 +1,39 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Code generated by "stringer -type=DHCPv6ConfigurationFromNDPRA"; DO NOT EDIT. + +package stack + +import "strconv" + +func _() { + // An "invalid array index" compiler error signifies that the constant values have changed. + // Re-run the stringer command to generate them again. + var x [1]struct{} + _ = x[DHCPv6NoConfiguration-0] + _ = x[DHCPv6ManagedAddress-1] + _ = x[DHCPv6OtherConfigurations-2] +} + +const _DHCPv6ConfigurationFromNDPRA_name = "DHCPv6NoConfigurationDHCPv6ManagedAddressDHCPv6OtherConfigurations" + +var _DHCPv6ConfigurationFromNDPRA_index = [...]uint8{0, 21, 41, 66} + +func (i DHCPv6ConfigurationFromNDPRA) String() string { + if i < 0 || i >= DHCPv6ConfigurationFromNDPRA(len(_DHCPv6ConfigurationFromNDPRA_index)-1) { + return "DHCPv6ConfigurationFromNDPRA(" + strconv.FormatInt(int64(i), 10) + ")" + } + return _DHCPv6ConfigurationFromNDPRA_name[_DHCPv6ConfigurationFromNDPRA_index[i]:_DHCPv6ConfigurationFromNDPRA_index[i+1]] +} diff --git a/pkg/tcpip/stack/forwarder.go b/pkg/tcpip/stack/forwarder.go new file mode 100644 index 000000000..631953935 --- /dev/null +++ b/pkg/tcpip/stack/forwarder.go @@ -0,0 +1,131 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package stack + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/tcpip" +) + +const ( + // maxPendingResolutions is the maximum number of pending link-address + // resolutions. + maxPendingResolutions = 64 + maxPendingPacketsPerResolution = 256 +) + +type pendingPacket struct { + nic *NIC + route *Route + proto tcpip.NetworkProtocolNumber + pkt tcpip.PacketBuffer +} + +type forwardQueue struct { + sync.Mutex + + // The packets to send once the resolver completes. + packets map[<-chan struct{}][]*pendingPacket + + // FIFO of channels used to cancel the oldest goroutine waiting for + // link-address resolution. + cancelChans []chan struct{} +} + +func newForwardQueue() *forwardQueue { + return &forwardQueue{packets: make(map[<-chan struct{}][]*pendingPacket)} +} + +func (f *forwardQueue) enqueue(ch <-chan struct{}, n *NIC, r *Route, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { + shouldWait := false + + f.Lock() + packets, ok := f.packets[ch] + if !ok { + shouldWait = true + } + for len(packets) == maxPendingPacketsPerResolution { + p := packets[0] + packets = packets[1:] + p.nic.stack.stats.IP.OutgoingPacketErrors.Increment() + p.route.Release() + } + if l := len(packets); l >= maxPendingPacketsPerResolution { + panic(fmt.Sprintf("max pending packets for resolution reached; got %d packets, max = %d", l, maxPendingPacketsPerResolution)) + } + f.packets[ch] = append(packets, &pendingPacket{ + nic: n, + route: r, + proto: protocol, + pkt: pkt, + }) + f.Unlock() + + if !shouldWait { + return + } + + // Wait for the link-address resolution to complete. + // Start a goroutine with a forwarding-cancel channel so that we can + // limit the maximum number of goroutines running concurrently. + cancel := f.newCancelChannel() + go func() { + cancelled := false + select { + case <-ch: + case <-cancel: + cancelled = true + } + + f.Lock() + packets := f.packets[ch] + delete(f.packets, ch) + f.Unlock() + + for _, p := range packets { + if cancelled { + p.nic.stack.stats.IP.OutgoingPacketErrors.Increment() + } else if _, err := p.route.Resolve(nil); err != nil { + p.nic.stack.stats.IP.OutgoingPacketErrors.Increment() + } else { + p.nic.forwardPacket(p.route, p.proto, p.pkt) + } + p.route.Release() + } + }() +} + +// newCancelChannel creates a channel that can cancel a pending forwarding +// activity. The oldest channel is closed if the number of open channels would +// exceed maxPendingResolutions. +func (f *forwardQueue) newCancelChannel() chan struct{} { + f.Lock() + defer f.Unlock() + + if len(f.cancelChans) == maxPendingResolutions { + ch := f.cancelChans[0] + f.cancelChans = f.cancelChans[1:] + close(ch) + } + if l := len(f.cancelChans); l >= maxPendingResolutions { + panic(fmt.Sprintf("max pending resolutions reached; got %d active resolutions, max = %d", l, maxPendingResolutions)) + } + + ch := make(chan struct{}) + f.cancelChans = append(f.cancelChans, ch) + return ch +} diff --git a/pkg/tcpip/stack/forwarder_test.go b/pkg/tcpip/stack/forwarder_test.go new file mode 100644 index 000000000..321b7524d --- /dev/null +++ b/pkg/tcpip/stack/forwarder_test.go @@ -0,0 +1,635 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package stack + +import ( + "encoding/binary" + "math" + "testing" + "time" + + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/buffer" +) + +const ( + fwdTestNetNumber tcpip.NetworkProtocolNumber = math.MaxUint32 + fwdTestNetHeaderLen = 12 + fwdTestNetDefaultPrefixLen = 8 + + // fwdTestNetDefaultMTU is the MTU, in bytes, used throughout the tests, + // except where another value is explicitly used. It is chosen to match + // the MTU of loopback interfaces on linux systems. + fwdTestNetDefaultMTU = 65536 +) + +// fwdTestNetworkEndpoint is a network-layer protocol endpoint. +// Headers of this protocol are fwdTestNetHeaderLen bytes, but we currently only +// use the first three: destination address, source address, and transport +// protocol. They're all one byte fields to simplify parsing. +type fwdTestNetworkEndpoint struct { + nicID tcpip.NICID + id NetworkEndpointID + prefixLen int + proto *fwdTestNetworkProtocol + dispatcher TransportDispatcher + ep LinkEndpoint +} + +func (f *fwdTestNetworkEndpoint) MTU() uint32 { + return f.ep.MTU() - uint32(f.MaxHeaderLength()) +} + +func (f *fwdTestNetworkEndpoint) NICID() tcpip.NICID { + return f.nicID +} + +func (f *fwdTestNetworkEndpoint) PrefixLen() int { + return f.prefixLen +} + +func (*fwdTestNetworkEndpoint) DefaultTTL() uint8 { + return 123 +} + +func (f *fwdTestNetworkEndpoint) ID() *NetworkEndpointID { + return &f.id +} + +func (f *fwdTestNetworkEndpoint) HandlePacket(r *Route, pkt tcpip.PacketBuffer) { + // Consume the network header. + b := pkt.Data.First() + pkt.Data.TrimFront(fwdTestNetHeaderLen) + + // Dispatch the packet to the transport protocol. + f.dispatcher.DeliverTransportPacket(r, tcpip.TransportProtocolNumber(b[2]), pkt) +} + +func (f *fwdTestNetworkEndpoint) MaxHeaderLength() uint16 { + return f.ep.MaxHeaderLength() + fwdTestNetHeaderLen +} + +func (f *fwdTestNetworkEndpoint) PseudoHeaderChecksum(protocol tcpip.TransportProtocolNumber, dstAddr tcpip.Address) uint16 { + return 0 +} + +func (f *fwdTestNetworkEndpoint) Capabilities() LinkEndpointCapabilities { + return f.ep.Capabilities() +} + +func (f *fwdTestNetworkEndpoint) WritePacket(r *Route, gso *GSO, params NetworkHeaderParams, pkt tcpip.PacketBuffer) *tcpip.Error { + // Add the protocol's header to the packet and send it to the link + // endpoint. + b := pkt.Header.Prepend(fwdTestNetHeaderLen) + b[0] = r.RemoteAddress[0] + b[1] = f.id.LocalAddress[0] + b[2] = byte(params.Protocol) + + return f.ep.WritePacket(r, gso, fwdTestNetNumber, pkt) +} + +// WritePackets implements LinkEndpoint.WritePackets. +func (f *fwdTestNetworkEndpoint) WritePackets(r *Route, gso *GSO, pkts []tcpip.PacketBuffer, params NetworkHeaderParams) (int, *tcpip.Error) { + panic("not implemented") +} + +func (*fwdTestNetworkEndpoint) WriteHeaderIncludedPacket(r *Route, pkt tcpip.PacketBuffer) *tcpip.Error { + return tcpip.ErrNotSupported +} + +func (*fwdTestNetworkEndpoint) Close() {} + +// fwdTestNetworkProtocol is a network-layer protocol that implements Address +// resolution. +type fwdTestNetworkProtocol struct { + addrCache *linkAddrCache + addrResolveDelay time.Duration + onLinkAddressResolved func(cache *linkAddrCache, addr tcpip.Address) + onResolveStaticAddress func(tcpip.Address) (tcpip.LinkAddress, bool) +} + +func (f *fwdTestNetworkProtocol) Number() tcpip.NetworkProtocolNumber { + return fwdTestNetNumber +} + +func (f *fwdTestNetworkProtocol) MinimumPacketSize() int { + return fwdTestNetHeaderLen +} + +func (f *fwdTestNetworkProtocol) DefaultPrefixLen() int { + return fwdTestNetDefaultPrefixLen +} + +func (*fwdTestNetworkProtocol) ParseAddresses(v buffer.View) (src, dst tcpip.Address) { + return tcpip.Address(v[1:2]), tcpip.Address(v[0:1]) +} + +func (f *fwdTestNetworkProtocol) NewEndpoint(nicID tcpip.NICID, addrWithPrefix tcpip.AddressWithPrefix, linkAddrCache LinkAddressCache, dispatcher TransportDispatcher, ep LinkEndpoint, _ *Stack) (NetworkEndpoint, *tcpip.Error) { + return &fwdTestNetworkEndpoint{ + nicID: nicID, + id: NetworkEndpointID{LocalAddress: addrWithPrefix.Address}, + prefixLen: addrWithPrefix.PrefixLen, + proto: f, + dispatcher: dispatcher, + ep: ep, + }, nil +} + +func (f *fwdTestNetworkProtocol) SetOption(option interface{}) *tcpip.Error { + return tcpip.ErrUnknownProtocolOption +} + +func (f *fwdTestNetworkProtocol) Option(option interface{}) *tcpip.Error { + return tcpip.ErrUnknownProtocolOption +} + +func (f *fwdTestNetworkProtocol) Close() {} + +func (f *fwdTestNetworkProtocol) Wait() {} + +func (f *fwdTestNetworkProtocol) LinkAddressRequest(addr, localAddr tcpip.Address, linkEP LinkEndpoint) *tcpip.Error { + if f.addrCache != nil && f.onLinkAddressResolved != nil { + time.AfterFunc(f.addrResolveDelay, func() { + f.onLinkAddressResolved(f.addrCache, addr) + }) + } + return nil +} + +func (f *fwdTestNetworkProtocol) ResolveStaticAddress(addr tcpip.Address) (tcpip.LinkAddress, bool) { + if f.onResolveStaticAddress != nil { + return f.onResolveStaticAddress(addr) + } + return "", false +} + +func (f *fwdTestNetworkProtocol) LinkAddressProtocol() tcpip.NetworkProtocolNumber { + return fwdTestNetNumber +} + +// fwdTestPacketInfo holds all the information about an outbound packet. +type fwdTestPacketInfo struct { + RemoteLinkAddress tcpip.LinkAddress + LocalLinkAddress tcpip.LinkAddress + Pkt tcpip.PacketBuffer +} + +type fwdTestLinkEndpoint struct { + dispatcher NetworkDispatcher + mtu uint32 + linkAddr tcpip.LinkAddress + + // C is where outbound packets are queued. + C chan fwdTestPacketInfo +} + +// InjectInbound injects an inbound packet. +func (e *fwdTestLinkEndpoint) InjectInbound(protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { + e.InjectLinkAddr(protocol, "", pkt) +} + +// InjectLinkAddr injects an inbound packet with a remote link address. +func (e *fwdTestLinkEndpoint) InjectLinkAddr(protocol tcpip.NetworkProtocolNumber, remote tcpip.LinkAddress, pkt tcpip.PacketBuffer) { + e.dispatcher.DeliverNetworkPacket(e, remote, "" /* local */, protocol, pkt) +} + +// Attach saves the stack network-layer dispatcher for use later when packets +// are injected. +func (e *fwdTestLinkEndpoint) Attach(dispatcher NetworkDispatcher) { + e.dispatcher = dispatcher +} + +// IsAttached implements stack.LinkEndpoint.IsAttached. +func (e *fwdTestLinkEndpoint) IsAttached() bool { + return e.dispatcher != nil +} + +// MTU implements stack.LinkEndpoint.MTU. It returns the value initialized +// during construction. +func (e *fwdTestLinkEndpoint) MTU() uint32 { + return e.mtu +} + +// Capabilities implements stack.LinkEndpoint.Capabilities. +func (e fwdTestLinkEndpoint) Capabilities() LinkEndpointCapabilities { + caps := LinkEndpointCapabilities(0) + return caps | CapabilityResolutionRequired +} + +// GSOMaxSize returns the maximum GSO packet size. +func (*fwdTestLinkEndpoint) GSOMaxSize() uint32 { + return 1 << 15 +} + +// MaxHeaderLength returns the maximum size of the link layer header. Given it +// doesn't have a header, it just returns 0. +func (*fwdTestLinkEndpoint) MaxHeaderLength() uint16 { + return 0 +} + +// LinkAddress returns the link address of this endpoint. +func (e *fwdTestLinkEndpoint) LinkAddress() tcpip.LinkAddress { + return e.linkAddr +} + +func (e fwdTestLinkEndpoint) WritePacket(r *Route, gso *GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error { + p := fwdTestPacketInfo{ + RemoteLinkAddress: r.RemoteLinkAddress, + LocalLinkAddress: r.LocalLinkAddress, + Pkt: pkt, + } + + select { + case e.C <- p: + default: + } + + return nil +} + +// WritePackets stores outbound packets into the channel. +func (e *fwdTestLinkEndpoint) WritePackets(r *Route, gso *GSO, pkts []tcpip.PacketBuffer, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { + n := 0 + for _, pkt := range pkts { + e.WritePacket(r, gso, protocol, pkt) + n++ + } + + return n, nil +} + +// WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket. +func (e *fwdTestLinkEndpoint) WriteRawPacket(vv buffer.VectorisedView) *tcpip.Error { + p := fwdTestPacketInfo{ + Pkt: tcpip.PacketBuffer{Data: vv}, + } + + select { + case e.C <- p: + default: + } + + return nil +} + +// Wait implements stack.LinkEndpoint.Wait. +func (*fwdTestLinkEndpoint) Wait() {} + +func fwdTestNetFactory(t *testing.T, proto *fwdTestNetworkProtocol) (ep1, ep2 *fwdTestLinkEndpoint) { + // Create a stack with the network protocol and two NICs. + s := New(Options{ + NetworkProtocols: []NetworkProtocol{proto}, + }) + + proto.addrCache = s.linkAddrCache + + // Enable forwarding. + s.SetForwarding(true) + + // NIC 1 has the link address "a", and added the network address 1. + ep1 = &fwdTestLinkEndpoint{ + C: make(chan fwdTestPacketInfo, 300), + mtu: fwdTestNetDefaultMTU, + linkAddr: "a", + } + if err := s.CreateNIC(1, ep1); err != nil { + t.Fatal("CreateNIC #1 failed:", err) + } + if err := s.AddAddress(1, fwdTestNetNumber, "\x01"); err != nil { + t.Fatal("AddAddress #1 failed:", err) + } + + // NIC 2 has the link address "b", and added the network address 2. + ep2 = &fwdTestLinkEndpoint{ + C: make(chan fwdTestPacketInfo, 300), + mtu: fwdTestNetDefaultMTU, + linkAddr: "b", + } + if err := s.CreateNIC(2, ep2); err != nil { + t.Fatal("CreateNIC #2 failed:", err) + } + if err := s.AddAddress(2, fwdTestNetNumber, "\x02"); err != nil { + t.Fatal("AddAddress #2 failed:", err) + } + + // Route all packets to NIC 2. + { + subnet, err := tcpip.NewSubnet("\x00", "\x00") + if err != nil { + t.Fatal(err) + } + s.SetRouteTable([]tcpip.Route{{Destination: subnet, NIC: 2}}) + } + + return ep1, ep2 +} + +func TestForwardingWithStaticResolver(t *testing.T) { + // Create a network protocol with a static resolver. + proto := &fwdTestNetworkProtocol{ + onResolveStaticAddress: + // The network address 3 is resolved to the link address "c". + func(addr tcpip.Address) (tcpip.LinkAddress, bool) { + if addr == "\x03" { + return "c", true + } + return "", false + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + // Inject an inbound packet to address 3 on NIC 1, and see if it is + // forwarded to NIC 2. + buf := buffer.NewView(30) + buf[0] = 3 + ep1.InjectInbound(fwdTestNetNumber, tcpip.PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + default: + t.Fatal("packet not forwarded") + } + + // Test that the static address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } +} + +func TestForwardingWithFakeResolver(t *testing.T) { + // Create a network protocol with a fake resolver. + proto := &fwdTestNetworkProtocol{ + addrResolveDelay: 500 * time.Millisecond, + onLinkAddressResolved: func(cache *linkAddrCache, addr tcpip.Address) { + // Any address will be resolved to the link address "c". + cache.add(tcpip.FullAddress{NIC: 2, Addr: addr}, "c") + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + // Inject an inbound packet to address 3 on NIC 1, and see if it is + // forwarded to NIC 2. + buf := buffer.NewView(30) + buf[0] = 3 + ep1.InjectInbound(fwdTestNetNumber, tcpip.PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + case <-time.After(time.Second): + t.Fatal("packet not forwarded") + } + + // Test that the address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } +} + +func TestForwardingWithNoResolver(t *testing.T) { + // Create a network protocol without a resolver. + proto := &fwdTestNetworkProtocol{} + + ep1, ep2 := fwdTestNetFactory(t, proto) + + // inject an inbound packet to address 3 on NIC 1, and see if it is + // forwarded to NIC 2. + buf := buffer.NewView(30) + buf[0] = 3 + ep1.InjectInbound(fwdTestNetNumber, tcpip.PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + + select { + case <-ep2.C: + t.Fatal("Packet should not be forwarded") + case <-time.After(time.Second): + } +} + +func TestForwardingWithFakeResolverPartialTimeout(t *testing.T) { + // Create a network protocol with a fake resolver. + proto := &fwdTestNetworkProtocol{ + addrResolveDelay: 500 * time.Millisecond, + onLinkAddressResolved: func(cache *linkAddrCache, addr tcpip.Address) { + // Only packets to address 3 will be resolved to the + // link address "c". + if addr == "\x03" { + cache.add(tcpip.FullAddress{NIC: 2, Addr: addr}, "c") + } + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + // Inject an inbound packet to address 4 on NIC 1. This packet should + // not be forwarded. + buf := buffer.NewView(30) + buf[0] = 4 + ep1.InjectInbound(fwdTestNetNumber, tcpip.PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + + // Inject an inbound packet to address 3 on NIC 1, and see if it is + // forwarded to NIC 2. + buf = buffer.NewView(30) + buf[0] = 3 + ep1.InjectInbound(fwdTestNetNumber, tcpip.PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + case <-time.After(time.Second): + t.Fatal("packet not forwarded") + } + + b := p.Pkt.Header.View() + if b[0] != 3 { + t.Fatalf("got b[0] = %d, want = 3", b[0]) + } + + // Test that the address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } +} + +func TestForwardingWithFakeResolverTwoPackets(t *testing.T) { + // Create a network protocol with a fake resolver. + proto := &fwdTestNetworkProtocol{ + addrResolveDelay: 500 * time.Millisecond, + onLinkAddressResolved: func(cache *linkAddrCache, addr tcpip.Address) { + // Any packets will be resolved to the link address "c". + cache.add(tcpip.FullAddress{NIC: 2, Addr: addr}, "c") + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + // Inject two inbound packets to address 3 on NIC 1. + for i := 0; i < 2; i++ { + buf := buffer.NewView(30) + buf[0] = 3 + ep1.InjectInbound(fwdTestNetNumber, tcpip.PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + } + + for i := 0; i < 2; i++ { + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + case <-time.After(time.Second): + t.Fatal("packet not forwarded") + } + + b := p.Pkt.Header.View() + if b[0] != 3 { + t.Fatalf("got b[0] = %d, want = 3", b[0]) + } + + // Test that the address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } + } +} + +func TestForwardingWithFakeResolverManyPackets(t *testing.T) { + // Create a network protocol with a fake resolver. + proto := &fwdTestNetworkProtocol{ + addrResolveDelay: 500 * time.Millisecond, + onLinkAddressResolved: func(cache *linkAddrCache, addr tcpip.Address) { + // Any packets will be resolved to the link address "c". + cache.add(tcpip.FullAddress{NIC: 2, Addr: addr}, "c") + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + for i := 0; i < maxPendingPacketsPerResolution+5; i++ { + // Inject inbound 'maxPendingPacketsPerResolution + 5' packets on NIC 1. + buf := buffer.NewView(30) + buf[0] = 3 + // Set the packet sequence number. + binary.BigEndian.PutUint16(buf[fwdTestNetHeaderLen:], uint16(i)) + ep1.InjectInbound(fwdTestNetNumber, tcpip.PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + } + + for i := 0; i < maxPendingPacketsPerResolution; i++ { + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + case <-time.After(time.Second): + t.Fatal("packet not forwarded") + } + + b := p.Pkt.Header.View() + if b[0] != 3 { + t.Fatalf("got b[0] = %d, want = 3", b[0]) + } + // The first 5 packets should not be forwarded so the the + // sequemnce number should start with 5. + want := uint16(i + 5) + if n := binary.BigEndian.Uint16(b[fwdTestNetHeaderLen:]); n != want { + t.Fatalf("got the packet #%d, want = #%d", n, want) + } + + // Test that the address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } + } +} + +func TestForwardingWithFakeResolverManyResolutions(t *testing.T) { + // Create a network protocol with a fake resolver. + proto := &fwdTestNetworkProtocol{ + addrResolveDelay: 500 * time.Millisecond, + onLinkAddressResolved: func(cache *linkAddrCache, addr tcpip.Address) { + // Any packets will be resolved to the link address "c". + cache.add(tcpip.FullAddress{NIC: 2, Addr: addr}, "c") + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + for i := 0; i < maxPendingResolutions+5; i++ { + // Inject inbound 'maxPendingResolutions + 5' packets on NIC 1. + // Each packet has a different destination address (3 to + // maxPendingResolutions + 7). + buf := buffer.NewView(30) + buf[0] = byte(3 + i) + ep1.InjectInbound(fwdTestNetNumber, tcpip.PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + } + + for i := 0; i < maxPendingResolutions; i++ { + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + case <-time.After(time.Second): + t.Fatal("packet not forwarded") + } + + // The first 5 packets (address 3 to 7) should not be forwarded + // because their address resolutions are interrupted. + b := p.Pkt.Header.View() + if b[0] < 8 { + t.Fatalf("got b[0] = %d, want b[0] >= 8", b[0]) + } + + // Test that the address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } + } +} diff --git a/pkg/tcpip/stack/ndp.go b/pkg/tcpip/stack/ndp.go index f651871ce..a9f4d5dad 100644 --- a/pkg/tcpip/stack/ndp.go +++ b/pkg/tcpip/stack/ndp.go @@ -1220,9 +1220,15 @@ func (ndp *ndpState) startSolicitingRouters() { } ndp.rtrSolicitTimer = time.AfterFunc(delay, func() { - // Send an RS message with the unspecified source address. - ref := ndp.nic.getRefOrCreateTemp(header.IPv6ProtocolNumber, header.IPv6Any, NeverPrimaryEndpoint, forceSpoofing) - r := makeRoute(header.IPv6ProtocolNumber, header.IPv6Any, header.IPv6AllRoutersMulticastAddress, ndp.nic.linkEP.LinkAddress(), ref, false, false) + // As per RFC 4861 section 4.1, the source of the RS is an address assigned + // to the sending interface, or the unspecified address if no address is + // assigned to the sending interface. + ref := ndp.nic.primaryIPv6Endpoint(header.IPv6AllRoutersMulticastAddress) + if ref == nil { + ref = ndp.nic.getRefOrCreateTemp(header.IPv6ProtocolNumber, header.IPv6Any, NeverPrimaryEndpoint, forceSpoofing) + } + localAddr := ref.ep.ID().LocalAddress + r := makeRoute(header.IPv6ProtocolNumber, localAddr, header.IPv6AllRoutersMulticastAddress, ndp.nic.linkEP.LinkAddress(), ref, false, false) defer r.Release() // Route should resolve immediately since @@ -1234,10 +1240,25 @@ func (ndp *ndpState) startSolicitingRouters() { log.Fatalf("ndp: route resolution not immediate for route to send NDP RS (%s -> %s on NIC(%d))", header.IPv6Any, header.IPv6AllRoutersMulticastAddress, ndp.nic.ID()) } - payloadSize := header.ICMPv6HeaderSize + header.NDPRSMinimumSize + // As per RFC 4861 section 4.1, an NDP RS SHOULD include the source + // link-layer address option if the source address of the NDP RS is + // specified. This option MUST NOT be included if the source address is + // unspecified. + // + // TODO(b/141011931): Validate a LinkEndpoint's link address (provided by + // LinkEndpoint.LinkAddress) before reaching this point. + var optsSerializer header.NDPOptionsSerializer + if localAddr != header.IPv6Any && header.IsValidUnicastEthernetAddress(r.LocalLinkAddress) { + optsSerializer = header.NDPOptionsSerializer{ + header.NDPSourceLinkLayerAddressOption(r.LocalLinkAddress), + } + } + payloadSize := header.ICMPv6HeaderSize + header.NDPRSMinimumSize + int(optsSerializer.Length()) hdr := buffer.NewPrependable(int(r.MaxHeaderLength()) + payloadSize) pkt := header.ICMPv6(hdr.Prepend(payloadSize)) pkt.SetType(header.ICMPv6RouterSolicit) + rs := header.NDPRouterSolicit(pkt.NDPPayload()) + rs.Options().Serialize(optsSerializer) pkt.SetChecksum(header.ICMPv6Checksum(pkt, r.LocalAddress, r.RemoteAddress, buffer.VectorisedView{})) sent := r.Stats().ICMP.V6PacketsSent diff --git a/pkg/tcpip/stack/ndp_test.go b/pkg/tcpip/stack/ndp_test.go index 6e9306d09..98b1c807c 100644 --- a/pkg/tcpip/stack/ndp_test.go +++ b/pkg/tcpip/stack/ndp_test.go @@ -3384,6 +3384,10 @@ func TestRouterSolicitation(t *testing.T) { tests := []struct { name string linkHeaderLen uint16 + linkAddr tcpip.LinkAddress + nicAddr tcpip.Address + expectedSrcAddr tcpip.Address + expectedNDPOpts []header.NDPOption maxRtrSolicit uint8 rtrSolicitInt time.Duration effectiveRtrSolicitInt time.Duration @@ -3392,6 +3396,7 @@ func TestRouterSolicitation(t *testing.T) { }{ { name: "Single RS with delay", + expectedSrcAddr: header.IPv6Any, maxRtrSolicit: 1, rtrSolicitInt: time.Second, effectiveRtrSolicitInt: time.Second, @@ -3401,6 +3406,8 @@ func TestRouterSolicitation(t *testing.T) { { name: "Two RS with delay", linkHeaderLen: 1, + nicAddr: llAddr1, + expectedSrcAddr: llAddr1, maxRtrSolicit: 2, rtrSolicitInt: time.Second, effectiveRtrSolicitInt: time.Second, @@ -3408,8 +3415,14 @@ func TestRouterSolicitation(t *testing.T) { effectiveMaxRtrSolicitDelay: 500 * time.Millisecond, }, { - name: "Single RS without delay", - linkHeaderLen: 2, + name: "Single RS without delay", + linkHeaderLen: 2, + linkAddr: linkAddr1, + nicAddr: llAddr1, + expectedSrcAddr: llAddr1, + expectedNDPOpts: []header.NDPOption{ + header.NDPSourceLinkLayerAddressOption(linkAddr1), + }, maxRtrSolicit: 1, rtrSolicitInt: time.Second, effectiveRtrSolicitInt: time.Second, @@ -3419,6 +3432,8 @@ func TestRouterSolicitation(t *testing.T) { { name: "Two RS without delay and invalid zero interval", linkHeaderLen: 3, + linkAddr: linkAddr1, + expectedSrcAddr: header.IPv6Any, maxRtrSolicit: 2, rtrSolicitInt: 0, effectiveRtrSolicitInt: 4 * time.Second, @@ -3427,6 +3442,8 @@ func TestRouterSolicitation(t *testing.T) { }, { name: "Three RS without delay", + linkAddr: linkAddr1, + expectedSrcAddr: header.IPv6Any, maxRtrSolicit: 3, rtrSolicitInt: 500 * time.Millisecond, effectiveRtrSolicitInt: 500 * time.Millisecond, @@ -3435,6 +3452,8 @@ func TestRouterSolicitation(t *testing.T) { }, { name: "Two RS with invalid negative delay", + linkAddr: linkAddr1, + expectedSrcAddr: header.IPv6Any, maxRtrSolicit: 2, rtrSolicitInt: time.Second, effectiveRtrSolicitInt: time.Second, @@ -3457,7 +3476,7 @@ func TestRouterSolicitation(t *testing.T) { t.Run(test.name, func(t *testing.T) { t.Parallel() e := channelLinkWithHeaderLength{ - Endpoint: channel.New(int(test.maxRtrSolicit), 1280, linkAddr1), + Endpoint: channel.New(int(test.maxRtrSolicit), 1280, test.linkAddr), headerLength: test.linkHeaderLen, } e.Endpoint.LinkEPCapabilities |= stack.CapabilityResolutionRequired @@ -3481,10 +3500,10 @@ func TestRouterSolicitation(t *testing.T) { checker.IPv6(t, p.Pkt.Header.View(), - checker.SrcAddr(header.IPv6Any), + checker.SrcAddr(test.expectedSrcAddr), checker.DstAddr(header.IPv6AllRoutersMulticastAddress), checker.TTL(header.NDPHopLimit), - checker.NDPRS(), + checker.NDPRS(checker.NDPRSOptions(test.expectedNDPOpts)), ) if l, want := p.Pkt.Header.AvailableLength(), int(test.linkHeaderLen); l != want { @@ -3510,13 +3529,19 @@ func TestRouterSolicitation(t *testing.T) { t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) } - // Make sure each RS got sent at the right - // times. + if addr := test.nicAddr; addr != "" { + if err := s.AddAddress(nicID, header.IPv6ProtocolNumber, addr); err != nil { + t.Fatalf("AddAddress(%d, %d, %s) = %s", nicID, header.IPv6ProtocolNumber, addr, err) + } + } + + // Make sure each RS is sent at the right time. remaining := test.maxRtrSolicit if remaining > 0 { waitForPkt(test.effectiveMaxRtrSolicitDelay + defaultAsyncEventTimeout) remaining-- } + for ; remaining > 0; remaining-- { waitForNothing(test.effectiveRtrSolicitInt - defaultTimeout) waitForPkt(defaultAsyncEventTimeout) diff --git a/pkg/tcpip/stack/nic.go b/pkg/tcpip/stack/nic.go index 271e55be0..3cd5fec71 100644 --- a/pkg/tcpip/stack/nic.go +++ b/pkg/tcpip/stack/nic.go @@ -452,7 +452,7 @@ func (n *NIC) primaryIPv6Endpoint(remoteAddr tcpip.Address) *referencedNetworkEn cs := make([]ipv6AddrCandidate, 0, len(primaryAddrs)) for _, r := range primaryAddrs { // If r is not valid for outgoing connections, it is not a valid endpoint. - if !r.isValidForOutgoing() { + if !r.isValidForOutgoingRLocked() { continue } @@ -1213,10 +1213,6 @@ func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.Link n.stack.stats.IP.InvalidDestinationAddressesReceived.Increment() return } - defer r.Release() - - r.LocalLinkAddress = n.linkEP.LinkAddress() - r.RemoteLinkAddress = remote // Found a NIC. n := r.ref.nic @@ -1225,24 +1221,33 @@ func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.Link ok = ok && ref.isValidForOutgoingRLocked() && ref.tryIncRef() n.mu.RUnlock() if ok { + r.LocalLinkAddress = n.linkEP.LinkAddress() + r.RemoteLinkAddress = remote r.RemoteAddress = src // TODO(b/123449044): Update the source NIC as well. ref.ep.HandlePacket(&r, pkt) ref.decRef() - } else { - // n doesn't have a destination endpoint. - // Send the packet out of n. - pkt.Header = buffer.NewPrependableFromView(pkt.Data.First()) - pkt.Data.RemoveFirst() - - // TODO(b/128629022): use route.WritePacket. - if err := n.linkEP.WritePacket(&r, nil /* gso */, protocol, pkt); err != nil { - r.Stats().IP.OutgoingPacketErrors.Increment() - } else { - n.stats.Tx.Packets.Increment() - n.stats.Tx.Bytes.IncrementBy(uint64(pkt.Header.UsedLength() + pkt.Data.Size())) + r.Release() + return + } + + // n doesn't have a destination endpoint. + // Send the packet out of n. + // TODO(b/128629022): move this logic to route.WritePacket. + if ch, err := r.Resolve(nil); err != nil { + if err == tcpip.ErrWouldBlock { + n.stack.forwarder.enqueue(ch, n, &r, protocol, pkt) + // forwarder will release route. + return } + n.stack.stats.IP.InvalidDestinationAddressesReceived.Increment() + r.Release() + return } + + // The link-address resolution finished immediately. + n.forwardPacket(&r, protocol, pkt) + r.Release() return } @@ -1252,6 +1257,20 @@ func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.Link } } +func (n *NIC) forwardPacket(r *Route, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { + // TODO(b/143425874) Decrease the TTL field in forwarded packets. + pkt.Header = buffer.NewPrependableFromView(pkt.Data.First()) + pkt.Data.RemoveFirst() + + if err := n.linkEP.WritePacket(r, nil /* gso */, protocol, pkt); err != nil { + r.Stats().IP.OutgoingPacketErrors.Increment() + return + } + + n.stats.Tx.Packets.Increment() + n.stats.Tx.Bytes.IncrementBy(uint64(pkt.Header.UsedLength() + pkt.Data.Size())) +} + // DeliverTransportPacket delivers the packets to the appropriate transport // protocol endpoint. func (n *NIC) DeliverTransportPacket(r *Route, protocol tcpip.TransportProtocolNumber, pkt tcpip.PacketBuffer) { diff --git a/pkg/tcpip/stack/stack.go b/pkg/tcpip/stack/stack.go index ebb6c5e3b..6f423874a 100644 --- a/pkg/tcpip/stack/stack.go +++ b/pkg/tcpip/stack/stack.go @@ -462,6 +462,10 @@ type Stack struct { // opaqueIIDOpts hold the options for generating opaque interface identifiers // (IIDs) as outlined by RFC 7217. opaqueIIDOpts OpaqueInterfaceIdentifierOptions + + // forwarder holds the packets that wait for their link-address resolutions + // to complete, and forwards them when each resolution is done. + forwarder *forwardQueue } // UniqueID is an abstract generator of unique identifiers. @@ -551,11 +555,13 @@ type TransportEndpointInfo struct { RegisterNICID tcpip.NICID } -// AddrNetProto unwraps the specified address if it is a V4-mapped V6 address -// and returns the network protocol number to be used to communicate with the -// specified address. It returns an error if the passed address is incompatible -// with the receiver. -func (e *TransportEndpointInfo) AddrNetProto(addr tcpip.FullAddress, v6only bool) (tcpip.FullAddress, tcpip.NetworkProtocolNumber, *tcpip.Error) { +// AddrNetProtoLocked unwraps the specified address if it is a V4-mapped V6 +// address and returns the network protocol number to be used to communicate +// with the specified address. It returns an error if the passed address is +// incompatible with the receiver. +// +// Preconditon: the parent endpoint mu must be held while calling this method. +func (e *TransportEndpointInfo) AddrNetProtoLocked(addr tcpip.FullAddress, v6only bool) (tcpip.FullAddress, tcpip.NetworkProtocolNumber, *tcpip.Error) { netProto := e.NetProto switch len(addr.Addr) { case header.IPv4AddressSize: @@ -639,6 +645,7 @@ func New(opts Options) *Stack { uniqueIDGenerator: opts.UniqueID, ndpDisp: opts.NDPDisp, opaqueIIDOpts: opts.OpaqueIIDOpts, + forwarder: newForwardQueue(), } // Add specified network protocols. diff --git a/pkg/tcpip/stack/stack_test.go b/pkg/tcpip/stack/stack_test.go index edf6bec52..e15db40fb 100644 --- a/pkg/tcpip/stack/stack_test.go +++ b/pkg/tcpip/stack/stack_test.go @@ -2790,13 +2790,14 @@ func TestNewPEBOnPromotionToPermanent(t *testing.T) { func TestIPv6SourceAddressSelectionScopeAndSameAddress(t *testing.T) { const ( - linkLocalAddr1 = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") - linkLocalAddr2 = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") - uniqueLocalAddr1 = tcpip.Address("\xfc\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") - uniqueLocalAddr2 = tcpip.Address("\xfd\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") - globalAddr1 = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") - globalAddr2 = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") - nicID = 1 + linkLocalAddr1 = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + linkLocalAddr2 = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") + linkLocalMulticastAddr = tcpip.Address("\xff\x02\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + uniqueLocalAddr1 = tcpip.Address("\xfc\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + uniqueLocalAddr2 = tcpip.Address("\xfd\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") + globalAddr1 = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + globalAddr2 = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") + nicID = 1 ) // Rule 3 is not tested here, and is instead tested by NDP's AutoGenAddr test. @@ -2870,6 +2871,18 @@ func TestIPv6SourceAddressSelectionScopeAndSameAddress(t *testing.T) { expectedLocalAddr: linkLocalAddr1, }, { + name: "Link Local most preferred for link local multicast (last address)", + nicAddrs: []tcpip.Address{globalAddr1, uniqueLocalAddr1, linkLocalAddr1}, + connectAddr: linkLocalMulticastAddr, + expectedLocalAddr: linkLocalAddr1, + }, + { + name: "Link Local most preferred for link local multicast (first address)", + nicAddrs: []tcpip.Address{linkLocalAddr1, uniqueLocalAddr1, globalAddr1}, + connectAddr: linkLocalMulticastAddr, + expectedLocalAddr: linkLocalAddr1, + }, + { name: "Unique Local most preferred (last address)", nicAddrs: []tcpip.Address{uniqueLocalAddr1, globalAddr1, linkLocalAddr1}, connectAddr: uniqueLocalAddr2, diff --git a/pkg/tcpip/transport/icmp/endpoint.go b/pkg/tcpip/transport/icmp/endpoint.go index 426da1ee6..2a396e9bc 100644 --- a/pkg/tcpip/transport/icmp/endpoint.go +++ b/pkg/tcpip/transport/icmp/endpoint.go @@ -291,15 +291,13 @@ func (e *endpoint) write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, <-c nicID = e.BindNICID } - toCopy := *to - to = &toCopy - netProto, err := e.checkV4Mapped(to) + dst, netProto, err := e.checkV4MappedLocked(*to) if err != nil { return 0, nil, err } - // Find the enpoint. - r, err := e.stack.FindRoute(nicID, e.BindAddr, to.Addr, netProto, false /* multicastLoop */) + // Find the endpoint. + r, err := e.stack.FindRoute(nicID, e.BindAddr, dst.Addr, netProto, false /* multicastLoop */) if err != nil { return 0, nil, err } @@ -480,13 +478,14 @@ func send6(r *stack.Route, ident uint16, data buffer.View, ttl uint8) *tcpip.Err }) } -func (e *endpoint) checkV4Mapped(addr *tcpip.FullAddress) (tcpip.NetworkProtocolNumber, *tcpip.Error) { - unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProto(*addr, false /* v6only */) +// checkV4MappedLocked determines the effective network protocol and converts +// addr to its canonical form. +func (e *endpoint) checkV4MappedLocked(addr tcpip.FullAddress) (tcpip.FullAddress, tcpip.NetworkProtocolNumber, *tcpip.Error) { + unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProtoLocked(addr, false /* v6only */) if err != nil { - return 0, err + return tcpip.FullAddress{}, 0, err } - *addr = unwrapped - return netProto, nil + return unwrapped, netProto, nil } // Disconnect implements tcpip.Endpoint.Disconnect. @@ -517,7 +516,7 @@ func (e *endpoint) Connect(addr tcpip.FullAddress) *tcpip.Error { return tcpip.ErrInvalidEndpointState } - netProto, err := e.checkV4Mapped(&addr) + addr, netProto, err := e.checkV4MappedLocked(addr) if err != nil { return err } @@ -630,7 +629,7 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) *tcpip.Error { return tcpip.ErrInvalidEndpointState } - netProto, err := e.checkV4Mapped(&addr) + addr, netProto, err := e.checkV4MappedLocked(addr) if err != nil { return err } diff --git a/pkg/tcpip/transport/packet/endpoint.go b/pkg/tcpip/transport/packet/endpoint.go index 5722815e9..09a1cd436 100644 --- a/pkg/tcpip/transport/packet/endpoint.go +++ b/pkg/tcpip/transport/packet/endpoint.go @@ -76,6 +76,7 @@ type endpoint struct { sndBufSize int closed bool stats tcpip.TransportEndpointStats `state:"nosave"` + bound bool } // NewEndpoint returns a new packet endpoint. @@ -125,6 +126,7 @@ func (ep *endpoint) Close() { } ep.closed = true + ep.bound = false ep.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) } @@ -216,7 +218,24 @@ func (ep *endpoint) Bind(addr tcpip.FullAddress) *tcpip.Error { // sll_family (should be AF_PACKET), sll_protocol, and sll_ifindex." // - packet(7). - return tcpip.ErrNotSupported + ep.mu.Lock() + defer ep.mu.Unlock() + + if ep.bound { + return tcpip.ErrAlreadyBound + } + + // Unregister endpoint with all the nics. + ep.stack.UnregisterPacketEndpoint(0, ep.netProto, ep) + + // Bind endpoint to receive packets from specific interface. + if err := ep.stack.RegisterPacketEndpoint(addr.NIC, ep.netProto, ep); err != nil { + return err + } + + ep.bound = true + + return nil } // GetLocalAddress implements tcpip.Endpoint.GetLocalAddress. diff --git a/pkg/tcpip/transport/tcp/BUILD b/pkg/tcpip/transport/tcp/BUILD index 272e8f570..a32f9eacf 100644 --- a/pkg/tcpip/transport/tcp/BUILD +++ b/pkg/tcpip/transport/tcp/BUILD @@ -32,6 +32,7 @@ go_library( srcs = [ "accept.go", "connect.go", + "connect_unsafe.go", "cubic.go", "cubic_state.go", "dispatcher.go", diff --git a/pkg/tcpip/transport/tcp/accept.go b/pkg/tcpip/transport/tcp/accept.go index 13e383ffc..85049e54e 100644 --- a/pkg/tcpip/transport/tcp/accept.go +++ b/pkg/tcpip/transport/tcp/accept.go @@ -236,6 +236,7 @@ func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, i n.effectiveNetProtos = []tcpip.NetworkProtocolNumber{s.route.NetProto} n.rcvBufSize = int(l.rcvWnd) n.amss = mssForRoute(&n.route) + n.setEndpointState(StateConnecting) n.maybeEnableTimestamp(rcvdSynOpts) n.maybeEnableSACKPermitted(rcvdSynOpts) diff --git a/pkg/tcpip/transport/tcp/connect.go b/pkg/tcpip/transport/tcp/connect.go index 7730e6445..c0f73ef16 100644 --- a/pkg/tcpip/transport/tcp/connect.go +++ b/pkg/tcpip/transport/tcp/connect.go @@ -295,6 +295,7 @@ func (h *handshake) synSentState(s *segment) *tcpip.Error { h.state = handshakeSynRcvd h.ep.mu.Lock() ttl := h.ep.ttl + amss := h.ep.amss h.ep.setEndpointState(StateSynRecv) h.ep.mu.Unlock() synOpts := header.TCPSynOptions{ @@ -307,7 +308,7 @@ func (h *handshake) synSentState(s *segment) *tcpip.Error { // permits SACK. This is not explicitly defined in the RFC but // this is the behaviour implemented by Linux. SACKPermitted: rcvSynOpts.SACKPermitted, - MSS: h.ep.amss, + MSS: amss, } if ttl == 0 { ttl = s.route.DefaultTTL() @@ -356,6 +357,10 @@ func (h *handshake) synRcvdState(s *segment) *tcpip.Error { return tcpip.ErrInvalidEndpointState } + h.ep.mu.RLock() + amss := h.ep.amss + h.ep.mu.RUnlock() + h.resetState() synOpts := header.TCPSynOptions{ WS: h.rcvWndScale, @@ -363,7 +368,7 @@ func (h *handshake) synRcvdState(s *segment) *tcpip.Error { TSVal: h.ep.timestamp(), TSEcr: h.ep.recentTimestamp(), SACKPermitted: h.ep.sackPermitted, - MSS: h.ep.amss, + MSS: amss, } h.ep.sendSynTCP(&s.route, h.ep.ID, h.ep.ttl, h.ep.sendTOS, h.flags, h.iss, h.ackNum, h.rcvWnd, synOpts) return nil @@ -530,6 +535,7 @@ func (h *handshake) execute() *tcpip.Error { // Send the initial SYN segment and loop until the handshake is // completed. + h.ep.mu.Lock() h.ep.amss = calculateAdvertisedMSS(h.ep.userMSS, h.ep.route) synOpts := header.TCPSynOptions{ @@ -540,6 +546,7 @@ func (h *handshake) execute() *tcpip.Error { SACKPermitted: bool(sackEnabled), MSS: h.ep.amss, } + h.ep.mu.Unlock() // Execute is also called in a listen context so we want to make sure we // only send the TS/SACK option when we received the TS/SACK in the @@ -577,7 +584,7 @@ func (h *handshake) execute() *tcpip.Error { case wakerForNotification: n := h.ep.fetchNotifications() - if n¬ifyClose != 0 { + if (n¬ifyClose)|(n¬ifyAbort) != 0 { return tcpip.ErrAborted } if n¬ifyDrain != 0 { @@ -617,17 +624,17 @@ func parseSynSegmentOptions(s *segment) header.TCPSynOptions { var optionPool = sync.Pool{ New: func() interface{} { - return make([]byte, maxOptionSize) + return &[maxOptionSize]byte{} }, } func getOptions() []byte { - return optionPool.Get().([]byte) + return (*optionPool.Get().(*[maxOptionSize]byte))[:] } func putOptions(options []byte) { // Reslice to full capacity. - optionPool.Put(options[0:cap(options)]) + optionPool.Put(optionsToArray(options)) } func makeSynOptions(opts header.TCPSynOptions) []byte { diff --git a/pkg/tcpip/transport/tcp/connect_unsafe.go b/pkg/tcpip/transport/tcp/connect_unsafe.go new file mode 100644 index 000000000..cfc304616 --- /dev/null +++ b/pkg/tcpip/transport/tcp/connect_unsafe.go @@ -0,0 +1,30 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package tcp + +import ( + "reflect" + "unsafe" +) + +// optionsToArray converts a slice of capacity >-= maxOptionSize to an array. +// +// optionsToArray panics if the capacity of options is smaller than +// maxOptionSize. +func optionsToArray(options []byte) *[maxOptionSize]byte { + // Reslice to full capacity. + options = options[0:maxOptionSize] + return (*[maxOptionSize]byte)(unsafe.Pointer((*reflect.SliceHeader)(unsafe.Pointer(&options)).Data)) +} diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go index f1ad19dac..dc9c18b6f 100644 --- a/pkg/tcpip/transport/tcp/endpoint.go +++ b/pkg/tcpip/transport/tcp/endpoint.go @@ -798,7 +798,21 @@ func (e *endpoint) Abort() { // If the endpoint disconnected after the check, nothing needs to be // done, so sending a notification which will potentially be ignored is // fine. - if e.EndpointState().connected() { + // + // If the endpoint connecting finishes after the check, the endpoint + // is either in a connected state (where we would notifyAbort anyway), + // SYN-RECV (where we would also notifyAbort anyway), or in an error + // state where nothing is required and the notification can be safely + // ignored. + // + // Endpoints where a Close during connecting or SYN-RECV state would be + // problematic are set to state connecting before being registered (and + // thus possible to be Aborted). They are never available in initial + // state. + // + // Endpoints transitioning from initial to connecting state may be + // safely either closed or sent notifyAbort. + if s := e.EndpointState(); s == StateConnecting || s == StateSynRecv || s.connected() { e.notifyProtocolGoroutine(notifyAbort) return } @@ -945,15 +959,18 @@ func (e *endpoint) initialReceiveWindow() int { // ModerateRecvBuf adjusts the receive buffer and the advertised window // based on the number of bytes copied to user space. func (e *endpoint) ModerateRecvBuf(copied int) { + e.mu.RLock() e.rcvListMu.Lock() if e.rcvAutoParams.disabled { e.rcvListMu.Unlock() + e.mu.RUnlock() return } now := time.Now() if rtt := e.rcvAutoParams.rtt; rtt == 0 || now.Sub(e.rcvAutoParams.measureTime) < rtt { e.rcvAutoParams.copied += copied e.rcvListMu.Unlock() + e.mu.RUnlock() return } prevRTTCopied := e.rcvAutoParams.copied + copied @@ -994,7 +1011,7 @@ func (e *endpoint) ModerateRecvBuf(copied int) { e.rcvBufSize = rcvWnd availAfter := e.receiveBufferAvailableLocked() mask := uint32(notifyReceiveWindowChanged) - if crossed, above := e.windowCrossedACKThreshold(availAfter - availBefore); crossed && above { + if crossed, above := e.windowCrossedACKThresholdLocked(availAfter - availBefore); crossed && above { mask |= notifyNonZeroReceiveWindow } e.notifyProtocolGoroutine(mask) @@ -1009,6 +1026,7 @@ func (e *endpoint) ModerateRecvBuf(copied int) { e.rcvAutoParams.measureTime = now e.rcvAutoParams.copied = 0 e.rcvListMu.Unlock() + e.mu.RUnlock() } // IPTables implements tcpip.Endpoint.IPTables. @@ -1038,7 +1056,6 @@ func (e *endpoint) Read(*tcpip.FullAddress) (buffer.View, tcpip.ControlMessages, v, err := e.readLocked() e.rcvListMu.Unlock() - e.mu.RUnlock() if err == tcpip.ErrClosedForReceive { @@ -1071,7 +1088,7 @@ func (e *endpoint) readLocked() (buffer.View, *tcpip.Error) { // enough buffer space, to either fit an aMSS or half a receive buffer // (whichever smaller), then notify the protocol goroutine to send a // window update. - if crossed, above := e.windowCrossedACKThreshold(len(v)); crossed && above { + if crossed, above := e.windowCrossedACKThresholdLocked(len(v)); crossed && above { e.notifyProtocolGoroutine(notifyNonZeroReceiveWindow) } @@ -1289,9 +1306,9 @@ func (e *endpoint) Peek(vec [][]byte) (int64, tcpip.ControlMessages, *tcpip.Erro return num, tcpip.ControlMessages{}, nil } -// windowCrossedACKThreshold checks if the receive window to be announced now -// would be under aMSS or under half receive buffer, whichever smaller. This is -// useful as a receive side silly window syndrome prevention mechanism. If +// windowCrossedACKThresholdLocked checks if the receive window to be announced +// now would be under aMSS or under half receive buffer, whichever smaller. This +// is useful as a receive side silly window syndrome prevention mechanism. If // window grows to reasonable value, we should send ACK to the sender to inform // the rx space is now large. We also want ensure a series of small read()'s // won't trigger a flood of spurious tiny ACK's. @@ -1302,7 +1319,9 @@ func (e *endpoint) Peek(vec [][]byte) (int64, tcpip.ControlMessages, *tcpip.Erro // crossed will be true if the window size crossed the ACK threshold. // above will be true if the new window is >= ACK threshold and false // otherwise. -func (e *endpoint) windowCrossedACKThreshold(deltaBefore int) (crossed bool, above bool) { +// +// Precondition: e.mu and e.rcvListMu must be held. +func (e *endpoint) windowCrossedACKThresholdLocked(deltaBefore int) (crossed bool, above bool) { newAvail := e.receiveBufferAvailableLocked() oldAvail := newAvail - deltaBefore if oldAvail < 0 { @@ -1365,6 +1384,7 @@ func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { mask := uint32(notifyReceiveWindowChanged) + e.mu.RLock() e.rcvListMu.Lock() // Make sure the receive buffer size allows us to send a @@ -1391,11 +1411,11 @@ func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { // Immediately send an ACK to uncork the sender silly window // syndrome prevetion, when our available space grows above aMSS // or half receive buffer, whichever smaller. - if crossed, above := e.windowCrossedACKThreshold(availAfter - availBefore); crossed && above { + if crossed, above := e.windowCrossedACKThresholdLocked(availAfter - availBefore); crossed && above { mask |= notifyNonZeroReceiveWindow } e.rcvListMu.Unlock() - + e.mu.RUnlock() e.notifyProtocolGoroutine(mask) return nil @@ -1854,13 +1874,14 @@ func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error { } } -func (e *endpoint) checkV4Mapped(addr *tcpip.FullAddress) (tcpip.NetworkProtocolNumber, *tcpip.Error) { - unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProto(*addr, e.v6only) +// checkV4MappedLocked determines the effective network protocol and converts +// addr to its canonical form. +func (e *endpoint) checkV4MappedLocked(addr tcpip.FullAddress) (tcpip.FullAddress, tcpip.NetworkProtocolNumber, *tcpip.Error) { + unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProtoLocked(addr, e.v6only) if err != nil { - return 0, err + return tcpip.FullAddress{}, 0, err } - *addr = unwrapped - return netProto, nil + return unwrapped, netProto, nil } // Disconnect implements tcpip.Endpoint.Disconnect. @@ -1890,7 +1911,7 @@ func (e *endpoint) connect(addr tcpip.FullAddress, handshake bool, run bool) *tc connectingAddr := addr.Addr - netProto, err := e.checkV4Mapped(&addr) + addr, netProto, err := e.checkV4MappedLocked(addr) if err != nil { return err } @@ -2096,10 +2117,13 @@ func (e *endpoint) Shutdown(flags tcpip.ShutdownFlags) *tcpip.Error { // Close for write. if (e.shutdownFlags & tcpip.ShutdownWrite) != 0 { e.sndBufMu.Lock() - if e.sndClosed { // Already closed. e.sndBufMu.Unlock() + if e.EndpointState() == StateTimeWait { + e.mu.Unlock() + return tcpip.ErrNotConnected + } break } @@ -2256,7 +2280,7 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) (err *tcpip.Error) { } e.BindAddr = addr.Addr - netProto, err := e.checkV4Mapped(&addr) + addr, netProto, err := e.checkV4MappedLocked(addr) if err != nil { return err } @@ -2400,13 +2424,14 @@ func (e *endpoint) updateSndBufferUsage(v int) { // to be read, or when the connection is closed for receiving (in which case // s will be nil). func (e *endpoint) readyToRead(s *segment) { + e.mu.RLock() e.rcvListMu.Lock() if s != nil { s.incRef() e.rcvBufUsed += s.data.Size() // Increase counter if the receive window falls down below MSS // or half receive buffer size, whichever smaller. - if crossed, above := e.windowCrossedACKThreshold(-s.data.Size()); crossed && !above { + if crossed, above := e.windowCrossedACKThresholdLocked(-s.data.Size()); crossed && !above { e.stats.ReceiveErrors.ZeroRcvWindowState.Increment() } e.rcvList.PushBack(s) @@ -2414,7 +2439,7 @@ func (e *endpoint) readyToRead(s *segment) { e.rcvClosed = true } e.rcvListMu.Unlock() - + e.mu.RUnlock() e.waiterQueue.Notify(waiter.EventIn) } diff --git a/pkg/tcpip/transport/tcp/rcv.go b/pkg/tcpip/transport/tcp/rcv.go index 958f03ac1..d80aff1b6 100644 --- a/pkg/tcpip/transport/tcp/rcv.go +++ b/pkg/tcpip/transport/tcp/rcv.go @@ -195,6 +195,10 @@ func (r *receiver) consumeSegment(s *segment, segSeq seqnum.Value, segLen seqnum for i := first; i < len(r.pendingRcvdSegments); i++ { r.pendingRcvdSegments[i].decRef() + // Note that slice truncation does not allow garbage collection of + // truncated items, thus truncated items must be set to nil to avoid + // memory leaks. + r.pendingRcvdSegments[i] = nil } r.pendingRcvdSegments = r.pendingRcvdSegments[:first] diff --git a/pkg/tcpip/transport/tcp/segment_heap.go b/pkg/tcpip/transport/tcp/segment_heap.go index 9fd061d7d..e28f213ba 100644 --- a/pkg/tcpip/transport/tcp/segment_heap.go +++ b/pkg/tcpip/transport/tcp/segment_heap.go @@ -41,6 +41,7 @@ func (h *segmentHeap) Pop() interface{} { old := *h n := len(old) x := old[n-1] + old[n-1] = nil *h = old[:n-1] return x } diff --git a/pkg/tcpip/transport/tcp/testing/context/context.go b/pkg/tcpip/transport/tcp/testing/context/context.go index 1e9a0dea3..8cea20fb5 100644 --- a/pkg/tcpip/transport/tcp/testing/context/context.go +++ b/pkg/tcpip/transport/tcp/testing/context/context.go @@ -204,6 +204,7 @@ func (c *Context) Cleanup() { if c.EP != nil { c.EP.Close() } + c.Stack().Close() } // Stack returns a reference to the stack in the Context. diff --git a/pkg/tcpip/transport/udp/endpoint.go b/pkg/tcpip/transport/udp/endpoint.go index 1c6a600b8..0af4514e1 100644 --- a/pkg/tcpip/transport/udp/endpoint.go +++ b/pkg/tcpip/transport/udp/endpoint.go @@ -443,19 +443,19 @@ func (e *endpoint) write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, <-c return 0, nil, tcpip.ErrBroadcastDisabled } - netProto, err := e.checkV4Mapped(to) + dst, netProto, err := e.checkV4MappedLocked(*to) if err != nil { return 0, nil, err } - r, _, err := e.connectRoute(nicID, *to, netProto) + r, _, err := e.connectRoute(nicID, dst, netProto) if err != nil { return 0, nil, err } defer r.Release() route = &r - dstPort = to.Port + dstPort = dst.Port } if route.IsResolutionRequired() { @@ -566,7 +566,7 @@ func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { defer e.mu.Unlock() fa := tcpip.FullAddress{Addr: v.InterfaceAddr} - netProto, err := e.checkV4Mapped(&fa) + fa, netProto, err := e.checkV4MappedLocked(fa) if err != nil { return err } @@ -927,13 +927,14 @@ func sendUDP(r *stack.Route, data buffer.VectorisedView, localPort, remotePort u return nil } -func (e *endpoint) checkV4Mapped(addr *tcpip.FullAddress) (tcpip.NetworkProtocolNumber, *tcpip.Error) { - unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProto(*addr, e.v6only) +// checkV4MappedLocked determines the effective network protocol and converts +// addr to its canonical form. +func (e *endpoint) checkV4MappedLocked(addr tcpip.FullAddress) (tcpip.FullAddress, tcpip.NetworkProtocolNumber, *tcpip.Error) { + unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProtoLocked(addr, e.v6only) if err != nil { - return 0, err + return tcpip.FullAddress{}, 0, err } - *addr = unwrapped - return netProto, nil + return unwrapped, netProto, nil } // Disconnect implements tcpip.Endpoint.Disconnect. @@ -981,10 +982,6 @@ func (e *endpoint) Disconnect() *tcpip.Error { // Connect connects the endpoint to its peer. Specifying a NIC is optional. func (e *endpoint) Connect(addr tcpip.FullAddress) *tcpip.Error { - netProto, err := e.checkV4Mapped(&addr) - if err != nil { - return err - } if addr.Port == 0 { // We don't support connecting to port zero. return tcpip.ErrInvalidEndpointState @@ -1012,6 +1009,11 @@ func (e *endpoint) Connect(addr tcpip.FullAddress) *tcpip.Error { return tcpip.ErrInvalidEndpointState } + addr, netProto, err := e.checkV4MappedLocked(addr) + if err != nil { + return err + } + r, nicID, err := e.connectRoute(nicID, addr, netProto) if err != nil { return err @@ -1139,7 +1141,7 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) *tcpip.Error { return tcpip.ErrInvalidEndpointState } - netProto, err := e.checkV4Mapped(&addr) + addr, netProto, err := e.checkV4MappedLocked(addr) if err != nil { return err } diff --git a/pkg/tcpip/transport/udp/endpoint_state.go b/pkg/tcpip/transport/udp/endpoint_state.go index 43fb047ed..466bd9381 100644 --- a/pkg/tcpip/transport/udp/endpoint_state.go +++ b/pkg/tcpip/transport/udp/endpoint_state.go @@ -69,6 +69,9 @@ func (e *endpoint) afterLoad() { // Resume implements tcpip.ResumableEndpoint.Resume. func (e *endpoint) Resume(s *stack.Stack) { + e.mu.Lock() + defer e.mu.Unlock() + e.stack = s for _, m := range e.multicastMemberships { |