diff options
| -rw-r--r-- | pkg/abi/linux/linux_abi_autogen_unsafe.go | 6 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/special_file.go | 5 | ||||
| -rw-r--r-- | pkg/sentry/syscalls/linux/vfs2/sync.go | 42 |
3 files changed, 39 insertions, 14 deletions
diff --git a/pkg/abi/linux/linux_abi_autogen_unsafe.go b/pkg/abi/linux/linux_abi_autogen_unsafe.go index f18f8a0e2..230e983b7 100644 --- a/pkg/abi/linux/linux_abi_autogen_unsafe.go +++ b/pkg/abi/linux/linux_abi_autogen_unsafe.go @@ -124,12 +124,12 @@ func (s *Statx) UnmarshalBytes(src []byte) { // Packed implements marshal.Marshallable.Packed. //go:nosplit func (s *Statx) Packed() bool { - return s.Atime.Packed() && s.Btime.Packed() && s.Ctime.Packed() && s.Mtime.Packed() + return s.Mtime.Packed() && s.Atime.Packed() && s.Btime.Packed() && s.Ctime.Packed() } // MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe. func (s *Statx) MarshalUnsafe(dst []byte) { - if s.Btime.Packed() && s.Ctime.Packed() && s.Mtime.Packed() && s.Atime.Packed() { + if s.Mtime.Packed() && s.Atime.Packed() && s.Btime.Packed() && s.Ctime.Packed() { safecopy.CopyIn(dst, unsafe.Pointer(s)) } else { s.MarshalBytes(dst) @@ -148,7 +148,7 @@ func (s *Statx) UnmarshalUnsafe(src []byte) { // CopyOutN implements marshal.Marshallable.CopyOutN. //go:nosplit func (s *Statx) CopyOutN(task marshal.Task, addr usermem.Addr, limit int) (int, error) { - if !s.Atime.Packed() && s.Btime.Packed() && s.Ctime.Packed() && s.Mtime.Packed() { + if !s.Mtime.Packed() && s.Atime.Packed() && s.Btime.Packed() && s.Ctime.Packed() { // Type Statx doesn't have a packed layout in memory, fall back to MarshalBytes. buf := task.CopyScratchBuffer(s.SizeBytes()) // escapes: okay. s.MarshalBytes(buf) // escapes: fallback. diff --git a/pkg/sentry/fsimpl/gofer/special_file.go b/pkg/sentry/fsimpl/gofer/special_file.go index a016cbae1..2b381af05 100644 --- a/pkg/sentry/fsimpl/gofer/special_file.go +++ b/pkg/sentry/fsimpl/gofer/special_file.go @@ -235,8 +235,5 @@ func (fd *specialFileFD) Seek(ctx context.Context, offset int64, whence int32) ( // Sync implements vfs.FileDescriptionImpl.Sync. func (fd *specialFileFD) Sync(ctx context.Context) error { - if !fd.vfsfd.IsWritable() { - return nil - } - return fd.handle.sync(ctx) + return fd.dentry().syncSharedHandle(ctx) } diff --git a/pkg/sentry/syscalls/linux/vfs2/sync.go b/pkg/sentry/syscalls/linux/vfs2/sync.go index 365250b0b..0d0ebf46a 100644 --- a/pkg/sentry/syscalls/linux/vfs2/sync.go +++ b/pkg/sentry/syscalls/linux/vfs2/sync.go @@ -65,10 +65,8 @@ func SyncFileRange(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel nbytes := args[2].Int64() flags := args[3].Uint() - if offset < 0 { - return 0, nil, syserror.EINVAL - } - if nbytes < 0 { + // Check for negative values and overflow. + if offset < 0 || offset+nbytes < 0 { return 0, nil, syserror.EINVAL } if flags&^(linux.SYNC_FILE_RANGE_WAIT_BEFORE|linux.SYNC_FILE_RANGE_WRITE|linux.SYNC_FILE_RANGE_WAIT_AFTER) != 0 { @@ -81,7 +79,37 @@ func SyncFileRange(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel } defer file.DecRef() - // TODO(gvisor.dev/issue/1897): Avoid writeback of data ranges outside of - // [offset, offset+nbytes). - return 0, nil, file.Sync(t) + // TODO(gvisor.dev/issue/1897): Currently, the only file syncing we support + // is a full-file sync, i.e. fsync(2). As a result, there are severe + // limitations on how much we support sync_file_range: + // - In Linux, sync_file_range(2) doesn't write out the file's metadata, even + // if the file size is changed. We do. + // - We always sync the entire file instead of [offset, offset+nbytes). + // - We do not support the use of WAIT_BEFORE without WAIT_AFTER. For + // correctness, we would have to perform a write-out every time WAIT_BEFORE + // was used, but this would be much more expensive than expected if there + // were no write-out operations in progress. + // - Whenever WAIT_AFTER is used, we sync the file. + // - Ignore WRITE. If this flag is used with WAIT_AFTER, then the file will + // be synced anyway. If this flag is used without WAIT_AFTER, then it is + // safe (and less expensive) to do nothing, because the syscall will not + // wait for the write-out to complete--we only need to make sure that the + // next time WAIT_BEFORE or WAIT_AFTER are used, the write-out completes. + // - According to fs/sync.c, WAIT_BEFORE|WAIT_AFTER "will detect any I/O + // errors or ENOSPC conditions and will return those to the caller, after + // clearing the EIO and ENOSPC flags in the address_space." We don't do + // this. + + if flags&linux.SYNC_FILE_RANGE_WAIT_BEFORE != 0 && + flags&linux.SYNC_FILE_RANGE_WAIT_AFTER == 0 { + t.Kernel().EmitUnimplementedEvent(t) + return 0, nil, syserror.ENOSYS + } + + if flags&linux.SYNC_FILE_RANGE_WAIT_AFTER != 0 { + if err := file.Sync(t); err != nil { + return 0, nil, syserror.ConvertIntr(err, kernel.ERESTARTSYS) + } + } + return 0, nil, nil } |