summaryrefslogtreecommitdiffhomepage
path: root/pkg/sentry/kernel/shm
diff options
context:
space:
mode:
authorIan Lewis <ianmlewis@gmail.com>2020-08-17 21:44:31 -0400
committerIan Lewis <ianmlewis@gmail.com>2020-08-17 21:44:31 -0400
commitac324f646ee3cb7955b0b45a7453aeb9671cbdf1 (patch)
tree0cbc5018e8807421d701d190dc20525726c7ca76 /pkg/sentry/kernel/shm
parent352ae1022ce19de28fc72e034cc469872ad79d06 (diff)
parent6d0c5803d557d453f15ac6f683697eeb46dab680 (diff)
Merge branch 'master' into ip-forwarding
- Merges aleksej-paschenko's with HEAD - Adds vfs2 support for ip_forward
Diffstat (limited to 'pkg/sentry/kernel/shm')
-rw-r--r--pkg/sentry/kernel/shm/BUILD9
-rw-r--r--pkg/sentry/kernel/shm/shm.go106
2 files changed, 74 insertions, 41 deletions
diff --git a/pkg/sentry/kernel/shm/BUILD b/pkg/sentry/kernel/shm/BUILD
index cd48945e6..c211fc8d0 100644
--- a/pkg/sentry/kernel/shm/BUILD
+++ b/pkg/sentry/kernel/shm/BUILD
@@ -1,4 +1,4 @@
-load("//tools/go_stateify:defs.bzl", "go_library")
+load("//tools:defs.bzl", "go_library")
package(licenses = ["notice"])
@@ -8,22 +8,21 @@ go_library(
"device.go",
"shm.go",
],
- importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/shm",
visibility = ["//pkg/sentry:internal"],
deps = [
"//pkg/abi/linux",
+ "//pkg/context",
"//pkg/log",
"//pkg/refs",
- "//pkg/sentry/context",
"//pkg/sentry/device",
"//pkg/sentry/fs",
"//pkg/sentry/kernel/auth",
"//pkg/sentry/kernel/time",
"//pkg/sentry/memmap",
"//pkg/sentry/pgalloc",
- "//pkg/sentry/platform",
"//pkg/sentry/usage",
- "//pkg/sentry/usermem",
+ "//pkg/sync",
"//pkg/syserror",
+ "//pkg/usermem",
],
)
diff --git a/pkg/sentry/kernel/shm/shm.go b/pkg/sentry/kernel/shm/shm.go
index 5bd610f68..13ec7afe0 100644
--- a/pkg/sentry/kernel/shm/shm.go
+++ b/pkg/sentry/kernel/shm/shm.go
@@ -35,21 +35,20 @@ package shm
import (
"fmt"
- "sync"
"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/context"
"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/memmap"
"gvisor.dev/gvisor/pkg/sentry/pgalloc"
- "gvisor.dev/gvisor/pkg/sentry/platform"
"gvisor.dev/gvisor/pkg/sentry/usage"
- "gvisor.dev/gvisor/pkg/sentry/usermem"
+ "gvisor.dev/gvisor/pkg/sync"
"gvisor.dev/gvisor/pkg/syserror"
+ "gvisor.dev/gvisor/pkg/usermem"
)
// Key represents a shm segment key. Analogous to a file name.
@@ -71,9 +70,20 @@ type Registry struct {
mu sync.Mutex `state:"nosave"`
// shms maps segment ids to segments.
+ //
+ // shms holds all referenced segments, which are removed on the last
+ // DecRef. Thus, it cannot itself hold a reference on the Shm.
+ //
+ // Since removal only occurs after the last (unlocked) DecRef, there
+ // exists a short window during which a Shm still exists in Shm, but is
+ // unreferenced. Users must use TryIncRef to determine if the Shm is
+ // still valid.
shms map[ID]*Shm
// keysToShms maps segment keys to segments.
+ //
+ // Shms in keysToShms are guaranteed to be referenced, as they are
+ // removed by disassociateKey before the last DecRef.
keysToShms map[Key]*Shm
// Sum of the sizes of all existing segments rounded up to page size, in
@@ -95,10 +105,18 @@ func NewRegistry(userNS *auth.UserNamespace) *Registry {
}
// FindByID looks up a segment given an ID.
+//
+// FindByID returns a reference on Shm.
func (r *Registry) FindByID(id ID) *Shm {
r.mu.Lock()
defer r.mu.Unlock()
- return r.shms[id]
+ s := r.shms[id]
+ // Take a reference on s. If TryIncRef fails, s has reached the last
+ // DecRef, but hasn't quite been removed from r.shms yet.
+ if s != nil && s.TryIncRef() {
+ return s
+ }
+ return nil
}
// dissociateKey removes the association between a segment and its key,
@@ -119,6 +137,8 @@ func (r *Registry) dissociateKey(s *Shm) {
// FindOrCreate looks up or creates a segment in the registry. It's functionally
// analogous to open(2).
+//
+// FindOrCreate returns a reference on Shm.
func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size uint64, mode linux.FileMode, private, create, exclusive bool) (*Shm, error) {
if (create || private) && (size < linux.SHMMIN || size > linux.SHMMAX) {
// "A new segment was to be created and size is less than SHMMIN or
@@ -166,6 +186,7 @@ func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size ui
return nil, syserror.EEXIST
}
+ shm.IncRef()
return shm, nil
}
@@ -193,7 +214,14 @@ func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size ui
// Need to create a new segment.
creator := fs.FileOwnerFromContext(ctx)
perms := fs.FilePermsFromMode(mode)
- return r.newShm(ctx, pid, key, creator, perms, size)
+ s, err := r.newShm(ctx, pid, key, creator, perms, size)
+ if err != nil {
+ return nil, err
+ }
+ // The initial reference is held by s itself. Take another to return to
+ // the caller.
+ s.IncRef()
+ return s, nil
}
// newShm creates a new segment in the registry.
@@ -296,22 +324,26 @@ func (r *Registry) remove(s *Shm) {
// Shm represents a single shared memory segment.
//
-// Shm segment are backed directly by an allocation from platform
-// memory. Segments are always mapped as a whole, greatly simplifying how
-// mappings are tracked. However note that mremap and munmap calls may cause the
-// vma for a segment to become fragmented; which requires special care when
-// unmapping a segment. See mm/shm.go.
+// Shm segment are backed directly by an allocation from platform memory.
+// Segments are always mapped as a whole, greatly simplifying how mappings are
+// tracked. However note that mremap and munmap calls may cause the vma for a
+// segment to become fragmented; which requires special care when unmapping a
+// segment. See mm/shm.go.
//
// Segments persist until they are explicitly marked for destruction via
-// shmctl(SHM_RMID).
+// MarkDestroyed().
//
// Shm implements memmap.Mappable and memmap.MappingIdentity.
//
// +stateify savable
type Shm struct {
- // AtomicRefCount tracks the number of references to this segment from
- // maps. A segment always holds a reference to itself, until it's marked for
+ // AtomicRefCount tracks the number of references to this segment.
+ //
+ // A segment holds a reference to itself until it is marked for
// destruction.
+ //
+ // In addition to direct users, the MemoryManager will hold references
+ // via MappingIdentity.
refs.AtomicRefCount
mfp pgalloc.MemoryFileProvider
@@ -337,7 +369,7 @@ type Shm struct {
// fr is the offset into mfp.MemoryFile() that backs this contents of this
// segment. Immutable.
- fr platform.FileRange
+ fr memmap.FileRange
// mu protects all fields below.
mu sync.Mutex `state:"nosave"`
@@ -399,8 +431,8 @@ func (s *Shm) InodeID() uint64 {
// DecRef overrides refs.RefCount.DecRef with a destructor.
//
// Precondition: Caller must not hold s.mu.
-func (s *Shm) DecRef() {
- s.DecRefWithDestructor(s.destroy)
+func (s *Shm) DecRef(ctx context.Context) {
+ s.DecRefWithDestructor(ctx, s.destroy)
}
// Msync implements memmap.MappingIdentity.Msync. Msync is a no-op for shm
@@ -428,7 +460,7 @@ func (s *Shm) AddMapping(ctx context.Context, _ memmap.MappingSpace, _ usermem.A
func (s *Shm) RemoveMapping(ctx context.Context, _ memmap.MappingSpace, _ usermem.AddrRange, _ uint64, _ bool) {
s.mu.Lock()
defer s.mu.Unlock()
- // TODO(b/38173783): RemoveMapping may be called during task exit, when ctx
+ // RemoveMapping may be called during task exit, when ctx
// is context.Background. Gracefully handle missing clocks. Failing to
// update the detach time in these cases is ok, since no one can observe the
// omission.
@@ -484,9 +516,8 @@ type AttachOpts struct {
// ConfigureAttach creates an mmap configuration for the segment with the
// requested attach options.
//
-// ConfigureAttach returns with a ref on s on success. The caller should drop
-// this once the map is installed. This reference prevents s from being
-// destroyed before the returned configuration is used.
+// Postconditions: The returned MMapOpts are valid only as long as a reference
+// continues to be held on s.
func (s *Shm) ConfigureAttach(ctx context.Context, addr usermem.Addr, opts AttachOpts) (memmap.MMapOpts, error) {
s.mu.Lock()
defer s.mu.Unlock()
@@ -504,7 +535,6 @@ func (s *Shm) ConfigureAttach(ctx context.Context, addr usermem.Addr, opts Attac
// in the user namespace that governs its IPC namespace." - man shmat(2)
return memmap.MMapOpts{}, syserror.EACCES
}
- s.IncRef()
return memmap.MMapOpts{
Length: s.size,
Offset: 0,
@@ -549,10 +579,15 @@ func (s *Shm) IPCStat(ctx context.Context) (*linux.ShmidDS, error) {
}
creds := auth.CredentialsFromContext(ctx)
- nattach := uint64(s.ReadRefs())
- // Don't report the self-reference we keep prior to being marked for
- // destruction. However, also don't report a count of -1 for segments marked
- // as destroyed, with no mappings.
+ // Use the reference count as a rudimentary count of the number of
+ // attaches. We exclude:
+ //
+ // 1. The reference the caller holds.
+ // 2. The self-reference held by s prior to destruction.
+ //
+ // Note that this may still overcount by including transient references
+ // used in concurrent calls.
+ nattach := uint64(s.ReadRefs()) - 1
if !s.pendingDestruction {
nattach--
}
@@ -607,7 +642,7 @@ func (s *Shm) Set(ctx context.Context, ds *linux.ShmidDS) error {
return nil
}
-func (s *Shm) destroy() {
+func (s *Shm) destroy(context.Context) {
s.mfp.MemoryFile().DecRef(s.fr)
s.registry.remove(s)
}
@@ -616,22 +651,21 @@ func (s *Shm) destroy() {
// destroyed once it has no references. MarkDestroyed may be called multiple
// times, and is safe to call after a segment has already been destroyed. See
// shmctl(IPC_RMID).
-func (s *Shm) MarkDestroyed() {
+func (s *Shm) MarkDestroyed(ctx context.Context) {
s.registry.dissociateKey(s)
s.mu.Lock()
- // Only drop the segment's self-reference once, when destruction is
- // requested. Otherwise, repeated calls to shmctl(IPC_RMID) would force a
- // segment to be destroyed prematurely, potentially with active maps to the
- // segment's address range. Remaining references are dropped when the
- // segment is detached or unmaped.
+ defer s.mu.Unlock()
if !s.pendingDestruction {
s.pendingDestruction = true
- s.mu.Unlock() // Must release s.mu before calling s.DecRef.
- s.DecRef()
+ // Drop the self-reference so destruction occurs when all
+ // external references are gone.
+ //
+ // N.B. This cannot be the final DecRef, as the caller also
+ // holds a reference.
+ s.DecRef(ctx)
return
}
- s.mu.Unlock()
}
// checkOwnership verifies whether a segment may be accessed by ctx as an