diff options
Diffstat (limited to 'pkg/sentry/vfs/file_description.go')
| -rw-r--r-- | pkg/sentry/vfs/file_description.go | 302 |
1 files changed, 252 insertions, 50 deletions
diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go index 6575afd16..6afe280bc 100644 --- a/pkg/sentry/vfs/file_description.go +++ b/pkg/sentry/vfs/file_description.go @@ -20,8 +20,10 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/waiter" ) @@ -38,49 +40,58 @@ type FileDescription struct { // operations. refs int64 + // statusFlags contains status flags, "initialized by open(2) and possibly + // modified by fcntl()" - fcntl(2). statusFlags is accessed using atomic + // memory operations. + statusFlags uint32 + // vd is the filesystem location at which this FileDescription was opened. // A reference is held on vd. vd is immutable. vd VirtualDentry + opts FileDescriptionOptions + // impl is the FileDescriptionImpl associated with this Filesystem. impl is // immutable. This should be the last field in FileDescription. impl FileDescriptionImpl } -// Init must be called before first use of fd. It takes ownership of references -// on mnt and d held by the caller. -func (fd *FileDescription) Init(impl FileDescriptionImpl, mnt *Mount, d *Dentry) { +// FileDescriptionOptions contains options to FileDescription.Init(). +type FileDescriptionOptions struct { + // If AllowDirectIO is true, allow O_DIRECT to be set on the file. This is + // usually only the case if O_DIRECT would actually have an effect. + AllowDirectIO bool + + // If UseDentryMetadata is true, calls to FileDescription methods that + // interact with file and filesystem metadata (Stat, SetStat, StatFS, + // Listxattr, Getxattr, Setxattr, Removexattr) are implemented by calling + // the corresponding FilesystemImpl methods instead of the corresponding + // FileDescriptionImpl methods. + // + // UseDentryMetadata is intended for file descriptions that are implemented + // outside of individual filesystems, such as pipes, sockets, and device + // special files. FileDescriptions for which UseDentryMetadata is true may + // embed DentryMetadataFileDescriptionImpl to obtain appropriate + // implementations of FileDescriptionImpl methods that should not be + // called. + UseDentryMetadata bool +} + +// Init must be called before first use of fd. It takes references on mnt and +// d. statusFlags is the initial file description status flags, which is +// usually the full set of flags passed to open(2). +func (fd *FileDescription) Init(impl FileDescriptionImpl, statusFlags uint32, mnt *Mount, d *Dentry, opts *FileDescriptionOptions) { fd.refs = 1 + fd.statusFlags = statusFlags | linux.O_LARGEFILE fd.vd = VirtualDentry{ mount: mnt, dentry: d, } + fd.vd.IncRef() + fd.opts = *opts fd.impl = impl } -// Impl returns the FileDescriptionImpl associated with fd. -func (fd *FileDescription) Impl() FileDescriptionImpl { - return fd.impl -} - -// Mount returns the mount on which fd was opened. It does not take a reference -// on the returned Mount. -func (fd *FileDescription) Mount() *Mount { - return fd.vd.mount -} - -// Dentry returns the dentry at which fd was opened. It does not take a -// reference on the returned Dentry. -func (fd *FileDescription) Dentry() *Dentry { - return fd.vd.dentry -} - -// VirtualDentry returns the location at which fd was opened. It does not take -// a reference on the returned VirtualDentry. -func (fd *FileDescription) VirtualDentry() VirtualDentry { - return fd.vd -} - // IncRef increments fd's reference count. func (fd *FileDescription) IncRef() { atomic.AddInt64(&fd.refs, 1) @@ -112,6 +123,82 @@ func (fd *FileDescription) DecRef() { } } +// Mount returns the mount on which fd was opened. It does not take a reference +// on the returned Mount. +func (fd *FileDescription) Mount() *Mount { + return fd.vd.mount +} + +// Dentry returns the dentry at which fd was opened. It does not take a +// reference on the returned Dentry. +func (fd *FileDescription) Dentry() *Dentry { + return fd.vd.dentry +} + +// VirtualDentry returns the location at which fd was opened. It does not take +// a reference on the returned VirtualDentry. +func (fd *FileDescription) VirtualDentry() VirtualDentry { + return fd.vd +} + +// StatusFlags returns file description status flags, as for fcntl(F_GETFL). +func (fd *FileDescription) StatusFlags() uint32 { + return atomic.LoadUint32(&fd.statusFlags) +} + +// SetStatusFlags sets file description status flags, as for fcntl(F_SETFL). +func (fd *FileDescription) SetStatusFlags(ctx context.Context, creds *auth.Credentials, flags uint32) error { + // Compare Linux's fs/fcntl.c:setfl(). + oldFlags := fd.StatusFlags() + // Linux documents this check as "O_APPEND cannot be cleared if the file is + // marked as append-only and the file is open for write", which would make + // sense. However, the check as actually implemented seems to be "O_APPEND + // cannot be changed if the file is marked as append-only". + if (flags^oldFlags)&linux.O_APPEND != 0 { + stat, err := fd.Stat(ctx, StatOptions{ + // There is no mask bit for stx_attributes. + Mask: 0, + // Linux just reads inode::i_flags directly. + Sync: linux.AT_STATX_DONT_SYNC, + }) + if err != nil { + return err + } + if (stat.AttributesMask&linux.STATX_ATTR_APPEND != 0) && (stat.Attributes&linux.STATX_ATTR_APPEND != 0) { + return syserror.EPERM + } + } + if (flags&linux.O_NOATIME != 0) && (oldFlags&linux.O_NOATIME == 0) { + stat, err := fd.Stat(ctx, StatOptions{ + Mask: linux.STATX_UID, + // Linux's inode_owner_or_capable() just reads inode::i_uid + // directly. + Sync: linux.AT_STATX_DONT_SYNC, + }) + if err != nil { + return err + } + if stat.Mask&linux.STATX_UID == 0 { + return syserror.EPERM + } + if !CanActAsOwner(creds, auth.KUID(stat.UID)) { + return syserror.EPERM + } + } + if flags&linux.O_DIRECT != 0 && !fd.opts.AllowDirectIO { + return syserror.EINVAL + } + // TODO(jamieliu): FileDescriptionImpl.SetOAsync()? + const settableFlags = linux.O_APPEND | linux.O_ASYNC | linux.O_DIRECT | linux.O_NOATIME | linux.O_NONBLOCK + atomic.StoreUint32(&fd.statusFlags, (oldFlags&^settableFlags)|(flags&settableFlags)) + return nil +} + +// Impl returns the FileDescriptionImpl associated with fd. +func (fd *FileDescription) Impl() FileDescriptionImpl { + return fd.impl +} + // FileDescriptionImpl contains implementation details for an FileDescription. // Implementations of FileDescriptionImpl should contain their associated // FileDescription by value as their first field. @@ -120,6 +207,8 @@ func (fd *FileDescription) DecRef() { // be interpreted as IDs in the root UserNamespace (i.e. as auth.KUID and // auth.KGID respectively). // +// All methods may return errors not specified. +// // FileDescriptionImpl is analogous to Linux's struct file_operations. type FileDescriptionImpl interface { // Release is called when the associated FileDescription reaches zero @@ -131,14 +220,6 @@ type FileDescriptionImpl interface { // prevent the file descriptor from being closed. OnClose(ctx context.Context) error - // StatusFlags returns file description status flags, as for - // fcntl(F_GETFL). - StatusFlags(ctx context.Context) (uint32, error) - - // SetStatusFlags sets file description status flags, as for - // fcntl(F_SETFL). - SetStatusFlags(ctx context.Context, flags uint32) error - // Stat returns metadata for the file represented by the FileDescription. Stat(ctx context.Context, opts StatOptions) (linux.Statx, error) @@ -156,6 +237,10 @@ type FileDescriptionImpl interface { // PRead reads from the file into dst, starting at the given offset, and // returns the number of bytes read. PRead is permitted to return partial // reads with a nil error. + // + // Errors: + // + // - If opts.Flags specifies unsupported options, PRead returns EOPNOTSUPP. PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts ReadOptions) (int64, error) // Read is similar to PRead, but does not specify an offset. @@ -165,6 +250,10 @@ type FileDescriptionImpl interface { // the number of bytes read; note that POSIX 2.9.7 "Thread Interactions // with Regular File Operations" requires that all operations that may // mutate the FileDescription offset are serialized. + // + // Errors: + // + // - If opts.Flags specifies unsupported options, Read returns EOPNOTSUPP. Read(ctx context.Context, dst usermem.IOSequence, opts ReadOptions) (int64, error) // PWrite writes src to the file, starting at the given offset, and returns @@ -174,6 +263,11 @@ type FileDescriptionImpl interface { // As in Linux (but not POSIX), if O_APPEND is in effect for the // FileDescription, PWrite should ignore the offset and append data to the // end of the file. + // + // Errors: + // + // - If opts.Flags specifies unsupported options, PWrite returns + // EOPNOTSUPP. PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts WriteOptions) (int64, error) // Write is similar to PWrite, but does not specify an offset, which is @@ -183,6 +277,10 @@ type FileDescriptionImpl interface { // PWrite that uses a FileDescription offset, to make it possible for // remote filesystems to implement O_APPEND correctly (i.e. atomically with // respect to writers outside the scope of VFS). + // + // Errors: + // + // - If opts.Flags specifies unsupported options, Write returns EOPNOTSUPP. Write(ctx context.Context, src usermem.IOSequence, opts WriteOptions) (int64, error) // IterDirents invokes cb on each entry in the directory represented by the @@ -212,7 +310,21 @@ type FileDescriptionImpl interface { // Ioctl implements the ioctl(2) syscall. Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) - // TODO: extended attributes; file locking + // Listxattr returns all extended attribute names for the file. + Listxattr(ctx context.Context) ([]string, error) + + // Getxattr returns the value associated with the given extended attribute + // for the file. + Getxattr(ctx context.Context, name string) (string, error) + + // Setxattr changes the value associated with the given extended attribute + // for the file. + Setxattr(ctx context.Context, opts SetxattrOptions) error + + // Removexattr removes the given extended attribute from the file. + Removexattr(ctx context.Context, name string) error + + // TODO: file locking } // Dirent holds the information contained in struct linux_dirent64. @@ -249,31 +361,49 @@ func (fd *FileDescription) OnClose(ctx context.Context) error { return fd.impl.OnClose(ctx) } -// StatusFlags returns file description status flags, as for fcntl(F_GETFL). -func (fd *FileDescription) StatusFlags(ctx context.Context) (uint32, error) { - flags, err := fd.impl.StatusFlags(ctx) - flags |= linux.O_LARGEFILE - return flags, err -} - -// SetStatusFlags sets file description status flags, as for fcntl(F_SETFL). -func (fd *FileDescription) SetStatusFlags(ctx context.Context, flags uint32) error { - return fd.impl.SetStatusFlags(ctx, flags) -} - // Stat returns metadata for the file represented by fd. func (fd *FileDescription) Stat(ctx context.Context, opts StatOptions) (linux.Statx, error) { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + stat, err := fd.vd.mount.fs.impl.StatAt(ctx, rp, opts) + vfsObj.putResolvingPath(rp) + return stat, err + } return fd.impl.Stat(ctx, opts) } // SetStat updates metadata for the file represented by fd. func (fd *FileDescription) SetStat(ctx context.Context, opts SetStatOptions) error { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + err := fd.vd.mount.fs.impl.SetStatAt(ctx, rp, opts) + vfsObj.putResolvingPath(rp) + return err + } return fd.impl.SetStat(ctx, opts) } // StatFS returns metadata for the filesystem containing the file represented // by fd. func (fd *FileDescription) StatFS(ctx context.Context) (linux.Statfs, error) { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + statfs, err := fd.vd.mount.fs.impl.StatFSAt(ctx, rp) + vfsObj.putResolvingPath(rp) + return statfs, err + } return fd.impl.StatFS(ctx) } @@ -329,6 +459,78 @@ func (fd *FileDescription) Ioctl(ctx context.Context, uio usermem.IO, args arch. return fd.impl.Ioctl(ctx, uio, args) } +// Listxattr returns all extended attribute names for the file represented by +// fd. +func (fd *FileDescription) Listxattr(ctx context.Context) ([]string, error) { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + names, err := fd.vd.mount.fs.impl.ListxattrAt(ctx, rp) + vfsObj.putResolvingPath(rp) + return names, err + } + names, err := fd.impl.Listxattr(ctx) + if err == syserror.ENOTSUP { + // Linux doesn't actually return ENOTSUP in this case; instead, + // fs/xattr.c:vfs_listxattr() falls back to allowing the security + // subsystem to return security extended attributes, which by default + // don't exist. + return nil, nil + } + return names, err +} + +// Getxattr returns the value associated with the given extended attribute for +// the file represented by fd. +func (fd *FileDescription) Getxattr(ctx context.Context, name string) (string, error) { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + val, err := fd.vd.mount.fs.impl.GetxattrAt(ctx, rp, name) + vfsObj.putResolvingPath(rp) + return val, err + } + return fd.impl.Getxattr(ctx, name) +} + +// Setxattr changes the value associated with the given extended attribute for +// the file represented by fd. +func (fd *FileDescription) Setxattr(ctx context.Context, opts SetxattrOptions) error { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + err := fd.vd.mount.fs.impl.SetxattrAt(ctx, rp, opts) + vfsObj.putResolvingPath(rp) + return err + } + return fd.impl.Setxattr(ctx, opts) +} + +// Removexattr removes the given extended attribute from the file represented +// by fd. +func (fd *FileDescription) Removexattr(ctx context.Context, name string) error { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + err := fd.vd.mount.fs.impl.RemovexattrAt(ctx, rp, name) + vfsObj.putResolvingPath(rp) + return err + } + return fd.impl.Removexattr(ctx, name) +} + // SyncFS instructs the filesystem containing fd to execute the semantics of // syncfs(2). func (fd *FileDescription) SyncFS(ctx context.Context) error { @@ -347,7 +549,7 @@ func (fd *FileDescription) MappedName(ctx context.Context) string { // DeviceID implements memmap.MappingIdentity.DeviceID. func (fd *FileDescription) DeviceID() uint64 { - stat, err := fd.impl.Stat(context.Background(), StatOptions{ + stat, err := fd.Stat(context.Background(), StatOptions{ // There is no STATX_DEV; we assume that Stat will return it if it's // available regardless of mask. Mask: 0, @@ -363,7 +565,7 @@ func (fd *FileDescription) DeviceID() uint64 { // InodeID implements memmap.MappingIdentity.InodeID. func (fd *FileDescription) InodeID() uint64 { - stat, err := fd.impl.Stat(context.Background(), StatOptions{ + stat, err := fd.Stat(context.Background(), StatOptions{ Mask: linux.STATX_INO, // fs/proc/task_mmu.c:show_map_vma() just reads inode::i_ino directly. Sync: linux.AT_STATX_DONT_SYNC, @@ -376,5 +578,5 @@ func (fd *FileDescription) InodeID() uint64 { // Msync implements memmap.MappingIdentity.Msync. func (fd *FileDescription) Msync(ctx context.Context, mr memmap.MappableRange) error { - return fd.impl.Sync(ctx) + return fd.Sync(ctx) } |