From a77eaf2a9e0539326defc766d748b8eff1d89b0b Mon Sep 17 00:00:00 2001
From: Jamie Liu <jamieliu@google.com>
Date: Thu, 19 Aug 2021 17:08:02 -0700
Subject: Use MM-mapped I/O instead of buffered copies in gofer.specialFileFD.

The rationale given for using buffered copies is still valid, but it's unclear
whether holding MM locks or allocating buffers is better in practice, and the
former is at least consistent with gofer.regularFileFD (and VFS1), making
performance easier to reason about.

PiperOrigin-RevId: 391877913
---
 pkg/sentry/fsimpl/gofer/handle.go       | 41 +++++++++++++++++++++++++++++++++
 pkg/sentry/fsimpl/gofer/special_file.go | 37 ++++++++---------------------
 2 files changed, 50 insertions(+), 28 deletions(-)

(limited to 'pkg')

diff --git a/pkg/sentry/fsimpl/gofer/handle.go b/pkg/sentry/fsimpl/gofer/handle.go
index 5c57f6fea..02540a754 100644
--- a/pkg/sentry/fsimpl/gofer/handle.go
+++ b/pkg/sentry/fsimpl/gofer/handle.go
@@ -20,6 +20,7 @@ import (
 	"gvisor.dev/gvisor/pkg/p9"
 	"gvisor.dev/gvisor/pkg/safemem"
 	"gvisor.dev/gvisor/pkg/sentry/hostfd"
+	"gvisor.dev/gvisor/pkg/sync"
 )
 
 // handle represents a remote "open file descriptor", consisting of an opened
@@ -130,3 +131,43 @@ func (h *handle) writeFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, o
 	}
 	return uint64(n), cperr
 }
+
+type handleReadWriter struct {
+	ctx context.Context
+	h   *handle
+	off uint64
+}
+
+var handleReadWriterPool = sync.Pool{
+	New: func() interface{} {
+		return &handleReadWriter{}
+	},
+}
+
+func getHandleReadWriter(ctx context.Context, h *handle, offset int64) *handleReadWriter {
+	rw := handleReadWriterPool.Get().(*handleReadWriter)
+	rw.ctx = ctx
+	rw.h = h
+	rw.off = uint64(offset)
+	return rw
+}
+
+func putHandleReadWriter(rw *handleReadWriter) {
+	rw.ctx = nil
+	rw.h = nil
+	handleReadWriterPool.Put(rw)
+}
+
+// ReadToBlocks implements safemem.Reader.ReadToBlocks.
+func (rw *handleReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) {
+	n, err := rw.h.readToBlocksAt(rw.ctx, dsts, rw.off)
+	rw.off += n
+	return n, err
+}
+
+// WriteFromBlocks implements safemem.Writer.WriteFromBlocks.
+func (rw *handleReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) {
+	n, err := rw.h.writeFromBlocksAt(rw.ctx, srcs, rw.off)
+	rw.off += n
+	return n, err
+}
diff --git a/pkg/sentry/fsimpl/gofer/special_file.go b/pkg/sentry/fsimpl/gofer/special_file.go
index 144a1045e..8e3c41808 100644
--- a/pkg/sentry/fsimpl/gofer/special_file.go
+++ b/pkg/sentry/fsimpl/gofer/special_file.go
@@ -24,7 +24,6 @@ import (
 	"gvisor.dev/gvisor/pkg/fdnotifier"
 	"gvisor.dev/gvisor/pkg/metric"
 	"gvisor.dev/gvisor/pkg/p9"
-	"gvisor.dev/gvisor/pkg/safemem"
 	"gvisor.dev/gvisor/pkg/sentry/fsmetric"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
 	"gvisor.dev/gvisor/pkg/sync"
@@ -229,23 +228,13 @@ func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs
 		}
 	}
 
-	// Going through dst.CopyOutFrom() would hold MM locks around file
-	// operations of unknown duration. For regularFileFD, doing so is necessary
-	// to support mmap due to lock ordering; MM locks precede dentry.dataMu.
-	// That doesn't hold here since specialFileFD doesn't client-cache data.
-	// Just buffer the read instead.
-	buf := make([]byte, dst.NumBytes())
-	n, err := fd.handle.readToBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset))
+	rw := getHandleReadWriter(ctx, &fd.handle, offset)
+	n, err := dst.CopyOutFrom(ctx, rw)
+	putHandleReadWriter(rw)
 	if linuxerr.Equals(linuxerr.EAGAIN, err) {
 		err = linuxerr.ErrWouldBlock
 	}
-	if n == 0 {
-		return bufN, err
-	}
-	if cp, cperr := dst.CopyOut(ctx, buf[:n]); cperr != nil {
-		return bufN + int64(cp), cperr
-	}
-	return bufN + int64(n), err
+	return bufN + n, err
 }
 
 // Read implements vfs.FileDescriptionImpl.Read.
@@ -316,20 +305,15 @@ func (fd *specialFileFD) pwrite(ctx context.Context, src usermem.IOSequence, off
 		}
 	}
 
-	// Do a buffered write. See rationale in PRead.
-	buf := make([]byte, src.NumBytes())
-	copied, copyErr := src.CopyIn(ctx, buf)
-	if copied == 0 && copyErr != nil {
-		// Only return the error if we didn't get any data.
-		return 0, offset, copyErr
-	}
-	n, err := fd.handle.writeFromBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf[:copied])), uint64(offset))
+	rw := getHandleReadWriter(ctx, &fd.handle, offset)
+	n, err := src.CopyInTo(ctx, rw)
+	putHandleReadWriter(rw)
 	if linuxerr.Equals(linuxerr.EAGAIN, err) {
 		err = linuxerr.ErrWouldBlock
 	}
 	// Update offset if the offset is valid.
 	if offset >= 0 {
-		offset += int64(n)
+		offset += n
 	}
 	// Update file size for regular files.
 	if fd.isRegularFile {
@@ -340,10 +324,7 @@ func (fd *specialFileFD) pwrite(ctx context.Context, src usermem.IOSequence, off
 			atomic.StoreUint64(&d.size, uint64(offset))
 		}
 	}
-	if err != nil {
-		return int64(n), offset, err
-	}
-	return int64(n), offset, copyErr
+	return int64(n), offset, err
 }
 
 // Write implements vfs.FileDescriptionImpl.Write.
-- 
cgit v1.2.3