diff options
Diffstat (limited to 'pkg/sentry')
27 files changed, 1984 insertions, 50 deletions
diff --git a/pkg/sentry/fs/g3doc/.gitignore b/pkg/sentry/fs/g3doc/.gitignore new file mode 100644 index 000000000..2d19fc766 --- /dev/null +++ b/pkg/sentry/fs/g3doc/.gitignore @@ -0,0 +1 @@ +*.html diff --git a/pkg/sentry/fs/g3doc/fuse.md b/pkg/sentry/fs/g3doc/fuse.md new file mode 100644 index 000000000..b43c082a7 --- /dev/null +++ b/pkg/sentry/fs/g3doc/fuse.md @@ -0,0 +1,262 @@ +# Foreword + +This document describes an on-going project to support FUSE filesystems within +the sentry. This is intended to become the final documentation for this +subsystem, and is therefore written in the past tense. However FUSE support is +currently incomplete and the document will be updated as things progress. + +# FUSE: Filesystem in Userspace + +The sentry supports dispatching filesystem operations to a FUSE server, allowing +FUSE filesystem to be used with a sandbox. + +## Overview + +FUSE has two main components: + +1. A client kernel driver (canonically `fuse.ko` in Linux), which forwards + filesystem operations (usually initiated by syscalls) to the server. + +2. A server, which is a userspace daemon that implements the actual filesystem. + +The sentry implements the client component, which allows a server daemon running +within the sandbox to implement a filesystem within the sandbox. + +A FUSE filesystem is initialized with `mount(2)`, typically with the help of a +utility like `fusermount(1)`. Various mount options exist for establishing +ownership and access permissions on the filesystem, but the most important mount +option is a file descriptor used to establish communication between the client +and server. + +The FUSE device FD is obtained by opening `/dev/fuse`. During regular operation, +the client and server use the FUSE protocol described in `fuse(4)` to service +filesystem operations. See the "Protocol" section below for more information +about this protocol. The core of the sentry support for FUSE is the client-side +implementation of this protocol. + +## FUSE in the Sentry + +The sentry's FUSE client targets VFS2 and has the following components: + +- An implementation of `/dev/fuse`. + +- A VFS2 filesystem for mapping syscalls to FUSE ops. Since we're targeting + VFS2, one point of contention may be the lack of inodes in VFS2. We can + tentatively implement a kernfs-based filesystem to bridge the gap in APIs. + The kernfs base functionality can serve the role of the Linux inode cache + and, the filesystem can map VFS2 syscalls to kernfs inode operations; see + the `kernfs.Inode` interface. + +The FUSE protocol lends itself well to marshaling with `go_marshal`. The various +request and response packets can be defined in the ABI package and converted to +and from the wire format using `go_marshal`. + +### Design Goals + +- While filesystem performance is always important, the sentry's FUSE support + is primarily concerned with compatibility, with performance as a secondary + concern. + +- Avoiding deadlocks from a hung server daemon. + +- Consider the potential for denial of service from a malicious server daemon. + Protecting itself from userspace is already a design goal for the sentry, + but needs additional consideration for FUSE. Normally, an operating system + doesn't rely on userspace to make progress with filesystem operations. Since + this changes with FUSE, it opens up the possibility of creating a chain of + dependencies controlled by userspace, which could affect an entire sandbox. + For example: a FUSE op can block a syscall, which could be holding a + subsystem lock, which can then block another task goroutine. + +### Milestones + +Below are some broad goals to aim for while implementing FUSE in the sentry. +Many FUSE ops can be grouped into broad categories of functionality, and most +ops can be implemented in parallel. + +#### Minimal client that can mount a trivial FUSE filesystem. + +- Implement `/dev/fuse`. + +- Implement basic FUSE ops like `FUSE_INIT`, `FUSE_DESTROY`. + +#### Read-only mount with basic file operations + +- Implement the majority of file, directory and file descriptor FUSE ops. For + this milestone, we can skip uncommon or complex operations like mmap, mknod, + file locking, poll, and extended attributes. We can stub these out along + with any ops that modify the filesystem. The exact list of required ops are + to be determined, but the goal is to mount a real filesystem as read-only, + and be able to read contents from the filesystem in the sentry. + +#### Full read-write support + +- Implement the remaining FUSE ops and decide if we can omit rarely used + operations like ioctl. + +# Appendix + +## FUSE Protocol + +The FUSE protocol is a request-response protocol. All requests are initiated by +the client. The wire-format for the protocol is raw C structs serialized to +memory. + +All FUSE requests begin with the following request header: + +```c +struct fuse_in_header { + uint32_t len; // Length of the request, including this header. + uint32_t opcode; // Requested operation. + uint64_t unique; // A unique identifier for this request. + uint64_t nodeid; // ID of the filesystem object being operated on. + uint32_t uid; // UID of the requesting process. + uint32_t gid; // GID of the requesting process. + uint32_t pid; // PID of the requesting process. + uint32_t padding; +}; +``` + +The request is then followed by a payload specific to the `opcode`. + +All responses begin with this response header: + +```c +struct fuse_out_header { + uint32_t len; // Length of the response, including this header. + int32_t error; // Status of the request, 0 if success. + uint64_t unique; // The unique identifier from the corresponding request. +}; +``` + +The response payload also depends on the request `opcode`. If `error != 0`, the +response payload must be empty. + +### Operations + +The following is a list of all FUSE operations used in `fuse_in_header.opcode` +as of Linux v4.4, and a brief description of their purpose. These are defined in +`uapi/linux/fuse.h`. Many of these have a corresponding request and response +payload struct; `fuse(4)` has details for some of these. We also note how these +operations map to the sentry virtual filesystem. + +#### FUSE meta-operations + +These operations are specific to FUSE and don't have a corresponding action in a +generic filesystem. + +- `FUSE_INIT`: This operation initializes a new FUSE filesystem, and is the + first message sent by the client after mount. This is used for version and + feature negotiation. This is related to `mount(2)`. +- `FUSE_DESTROY`: Teardown a FUSE filesystem, related to `unmount(2)`. +- `FUSE_INTERRUPT`: Interrupts an in-flight operation, specified by the + `fuse_in_header.unique` value provided in the corresponding request header. + The client can send at most one of these per request, and will enter an + uninterruptible wait for a reply. The server is expected to reply promptly. +- `FUSE_FORGET`: A hint to the server that server should evict the indicate + node from any caches. This is wired up to `(struct + super_operations).evict_inode` in Linux, which is in turned hooked as the + inode cache shrinker which is typically triggered by system memory pressure. +- `FUSE_BATCH_FORGET`: Batch version of `FUSE_FORGET`. + +#### Filesystem Syscalls + +These FUSE ops map directly to an equivalent filesystem syscall, or family of +syscalls. The relevant syscalls have a similar name to the operation, unless +otherwise noted. + +Node creation: + +- `FUSE_MKNOD` +- `FUSE_MKDIR` +- `FUSE_CREATE`: This is equivalent to `open(2)` and `creat(2)`, which + atomically creates and opens a node. + +Node attributes and extended attributes: + +- `FUSE_GETATTR` +- `FUSE_SETATTR` +- `FUSE_SETXATTR` +- `FUSE_GETXATTR` +- `FUSE_LISTXATTR` +- `FUSE_REMOVEXATTR` + +Node link manipulation: + +- `FUSE_READLINK` +- `FUSE_LINK` +- `FUSE_SYMLINK` +- `FUSE_UNLINK` + +Directory operations: + +- `FUSE_RMDIR` +- `FUSE_RENAME` +- `FUSE_RENAME2` +- `FUSE_OPENDIR`: `open(2)` for directories. +- `FUSE_RELEASEDIR`: `close(2)` for directories. +- `FUSE_READDIR` +- `FUSE_READDIRPLUS` +- `FUSE_FSYNCDIR`: `fsync(2)` for directories. +- `FUSE_LOOKUP`: Establishes a unique identifier for a FS node. This is + reminiscent of `VirtualFilesystem.GetDentryAt` in that it resolves a path + component to a node. However the returned identifier is opaque to the + client. The server must remember this mapping, as this is how the client + will reference the node in the future. + +File operations: + +- `FUSE_OPEN`: `open(2)` for files. +- `FUSE_RELEASE`: `close(2)` for files. +- `FUSE_FSYNC` +- `FUSE_FALLOCATE` +- `FUSE_SETUPMAPPING`: Creates a memory map on a file for `mmap(2)`. +- `FUSE_REMOVEMAPPING`: Removes a memory map for `munmap(2)`. + +File locking: + +- `FUSE_GETLK` +- `FUSE_SETLK` +- `FUSE_SETLKW` +- `FUSE_COPY_FILE_RANGE` + +File descriptor operations: + +- `FUSE_IOCTL` +- `FUSE_POLL` +- `FUSE_LSEEK` + +Filesystem operations: + +- `FUSE_STATFS` + +#### Permissions + +- `FUSE_ACCESS` is used to check if a node is accessible, as part of many + syscall implementations. Maps to `vfs.FilesystemImpl.AccessAt` in the + sentry. + +#### I/O Operations + +These ops are used to read and write file pages. They're used to implement both +I/O syscalls like `read(2)`, `write(2)` and `mmap(2)`. + +- `FUSE_READ` +- `FUSE_WRITE` + +#### Miscellaneous + +- `FUSE_FLUSH`: Used by the client to indicate when a file descriptor is + closed. Distinct from `FUSE_FSYNC`, which corresponds to an `fsync(2)` + syscall from the user. Maps to `vfs.FileDescriptorImpl.Release` in the + sentry. +- `FUSE_BMAP`: Old address space API for block defrag. Probably not needed. +- `FUSE_NOTIFY_REPLY`: [TODO: what does this do?] + +# References + +- [fuse(4) Linux manual page](https://www.man7.org/linux/man-pages/man4/fuse.4.html) +- [Linux kernel FUSE documentation](https://www.kernel.org/doc/html/latest/filesystems/fuse.html) +- [The reference implementation of the Linux FUSE (Filesystem in Userspace) + interface](https://github.com/libfuse/libfuse) +- [The kernel interface of FUSE](https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/uapi/linux/fuse.h) diff --git a/pkg/sentry/fsimpl/ext/dentry.go b/pkg/sentry/fsimpl/ext/dentry.go index bfbd7c3d4..6bd1a9fc6 100644 --- a/pkg/sentry/fsimpl/ext/dentry.go +++ b/pkg/sentry/fsimpl/ext/dentry.go @@ -60,3 +60,15 @@ func (d *dentry) DecRef() { // inode.decRef(). d.inode.decRef() } + +// InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *dentry) InotifyWithParent(events uint32, cookie uint32, et vfs.EventType) {} + +// Watches implements vfs.DentryImpl.Watches. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *dentry) Watches() *vfs.Watches { + return nil +} diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go index 6295f6b54..3f3bd56f0 100644 --- a/pkg/sentry/fsimpl/gofer/gofer.go +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -84,12 +84,6 @@ type filesystem struct { // devMinor is the filesystem's minor device number. devMinor is immutable. devMinor uint32 - // uid and gid are the effective KUID and KGID of the filesystem's creator, - // and are used as the owner and group for files that don't specify one. - // uid and gid are immutable. - uid auth.KUID - gid auth.KGID - // renameMu serves two purposes: // // - It synchronizes path resolution with renaming initiated by this @@ -122,6 +116,8 @@ type filesystemOptions struct { fd int aname string interop InteropMode // derived from the "cache" mount option + dfltuid auth.KUID + dfltgid auth.KGID msize uint32 version string @@ -230,6 +226,15 @@ type InternalFilesystemOptions struct { OpenSocketsByConnecting bool } +// _V9FS_DEFUID and _V9FS_DEFGID (from Linux's fs/9p/v9fs.h) are the default +// UIDs and GIDs used for files that do not provide a specific owner or group +// respectively. +const ( + // uint32(-2) doesn't work in Go. + _V9FS_DEFUID = auth.KUID(4294967294) + _V9FS_DEFGID = auth.KGID(4294967294) +) + // Name implements vfs.FilesystemType.Name. func (FilesystemType) Name() string { return Name @@ -315,6 +320,31 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt } } + // Parse the default UID and GID. + fsopts.dfltuid = _V9FS_DEFUID + if dfltuidstr, ok := mopts["dfltuid"]; ok { + delete(mopts, "dfltuid") + dfltuid, err := strconv.ParseUint(dfltuidstr, 10, 32) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid default UID: dfltuid=%s", dfltuidstr) + return nil, nil, syserror.EINVAL + } + // In Linux, dfltuid is interpreted as a UID and is converted to a KUID + // in the caller's user namespace, but goferfs isn't + // application-mountable. + fsopts.dfltuid = auth.KUID(dfltuid) + } + fsopts.dfltgid = _V9FS_DEFGID + if dfltgidstr, ok := mopts["dfltgid"]; ok { + delete(mopts, "dfltgid") + dfltgid, err := strconv.ParseUint(dfltgidstr, 10, 32) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid default UID: dfltgid=%s", dfltgidstr) + return nil, nil, syserror.EINVAL + } + fsopts.dfltgid = auth.KGID(dfltgid) + } + // Parse the 9P message size. fsopts.msize = 1024 * 1024 // 1M, tested to give good enough performance up to 64M if msizestr, ok := mopts["msize"]; ok { @@ -422,8 +452,6 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt client: client, clock: ktime.RealtimeClockFromContext(ctx), devMinor: devMinor, - uid: creds.EffectiveKUID, - gid: creds.EffectiveKGID, syncableDentries: make(map[*dentry]struct{}), specialFileFDs: make(map[*specialFileFD]struct{}), } @@ -672,8 +700,8 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma file: file, ino: qid.Path, mode: uint32(attr.Mode), - uid: uint32(fs.uid), - gid: uint32(fs.gid), + uid: uint32(fs.opts.dfltuid), + gid: uint32(fs.opts.dfltgid), blockSize: usermem.PageSize, handle: handle{ fd: -1, @@ -1011,6 +1039,18 @@ func (d *dentry) decRefLocked() { } } +// InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *dentry) InotifyWithParent(events uint32, cookie uint32, et vfs.EventType) {} + +// Watches implements vfs.DentryImpl.Watches. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *dentry) Watches() *vfs.Watches { + return nil +} + // checkCachingLocked should be called after d's reference count becomes 0 or it // becomes disowned. // diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go index a83151ad3..bbee8ccda 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -225,9 +225,21 @@ func (d *Dentry) destroy() { } } +// InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *Dentry) InotifyWithParent(events uint32, cookie uint32, et vfs.EventType) {} + +// Watches implements vfs.DentryImpl.Watches. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *Dentry) Watches() *vfs.Watches { + return nil +} + // InsertChild inserts child into the vfs dentry cache with the given name under // this dentry. This does not update the directory inode, so calling this on -// it's own isn't sufficient to insert a child into a directory. InsertChild +// its own isn't sufficient to insert a child into a directory. InsertChild // updates the link count on d if required. // // Precondition: d must represent a directory inode. diff --git a/pkg/sentry/fsimpl/tmpfs/BUILD b/pkg/sentry/fsimpl/tmpfs/BUILD index 007be1572..062321cbc 100644 --- a/pkg/sentry/fsimpl/tmpfs/BUILD +++ b/pkg/sentry/fsimpl/tmpfs/BUILD @@ -59,6 +59,7 @@ go_library( "//pkg/sentry/pgalloc", "//pkg/sentry/platform", "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/uniqueid", "//pkg/sentry/usage", "//pkg/sentry/vfs", "//pkg/sentry/vfs/lock", diff --git a/pkg/sentry/fsimpl/tmpfs/directory.go b/pkg/sentry/fsimpl/tmpfs/directory.go index f2399981b..70387cb9c 100644 --- a/pkg/sentry/fsimpl/tmpfs/directory.go +++ b/pkg/sentry/fsimpl/tmpfs/directory.go @@ -79,6 +79,7 @@ func (dir *directory) removeChildLocked(child *dentry) { dir.iterMu.Lock() dir.childList.Remove(child) dir.iterMu.Unlock() + child.unlinked = true } type directoryFD struct { @@ -112,6 +113,7 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba dir.iterMu.Lock() defer dir.iterMu.Unlock() + fd.dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) fd.inode().touchAtime(fd.vfsfd.Mount()) if fd.off == 0 { diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go index 80fa7b29d..183eb975c 100644 --- a/pkg/sentry/fsimpl/tmpfs/filesystem.go +++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go @@ -177,6 +177,12 @@ func (fs *filesystem) doCreateAt(rp *vfs.ResolvingPath, dir bool, create func(pa if err := create(parentDir, name); err != nil { return err } + + ev := linux.IN_CREATE + if dir { + ev |= linux.IN_ISDIR + } + parentDir.inode.watches.Notify(name, uint32(ev), 0, vfs.InodeEvent) parentDir.inode.touchCMtime() return nil } @@ -241,6 +247,7 @@ func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. return syserror.EMLINK } d.inode.incLinksLocked() + d.inode.watches.Notify("", linux.IN_ATTRIB, 0, vfs.InodeEvent) parentDir.insertChildLocked(fs.newDentry(d.inode), name) return nil }) @@ -354,6 +361,7 @@ afterTrailingSymlink: if err != nil { return nil, err } + parentDir.inode.watches.Notify(name, linux.IN_CREATE, 0, vfs.PathEvent) parentDir.inode.touchCMtime() return fd, nil } @@ -559,6 +567,8 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa newParentDir.inode.touchCMtime() } renamed.inode.touchCtime() + + vfs.InotifyRename(ctx, &renamed.inode.watches, &oldParentDir.inode.watches, &newParentDir.inode.watches, oldName, newName, renamed.inode.isDir()) return nil } @@ -603,8 +613,11 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error return err } parentDir.removeChildLocked(child) - parentDir.inode.decLinksLocked() // from child's ".." + parentDir.inode.watches.Notify(name, linux.IN_DELETE|linux.IN_ISDIR, 0, vfs.InodeEvent) + // Remove links for child, child/., and child/.. child.inode.decLinksLocked() + child.inode.decLinksLocked() + parentDir.inode.decLinksLocked() vfsObj.CommitDeleteDentry(&child.vfsd) parentDir.inode.touchCMtime() return nil @@ -618,7 +631,14 @@ func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts if err != nil { return err } - return d.inode.setStat(ctx, rp.Credentials(), &opts.Stat) + if err := d.inode.setStat(ctx, rp.Credentials(), &opts.Stat); err != nil { + return err + } + + if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { + d.InotifyWithParent(ev, 0, vfs.InodeEvent) + } + return nil } // StatAt implements vfs.FilesystemImpl.StatAt. @@ -698,6 +718,12 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil { return err } + + // Generate inotify events. Note that this must take place before the link + // count of the child is decremented, or else the watches may be dropped + // before these events are added. + vfs.InotifyRemoveChild(&child.inode.watches, &parentDir.inode.watches, name) + parentDir.removeChildLocked(child) child.inode.decLinksLocked() vfsObj.CommitDeleteDentry(&child.vfsd) @@ -754,7 +780,12 @@ func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt if err != nil { return err } - return d.inode.setxattr(rp.Credentials(), &opts) + if err := d.inode.setxattr(rp.Credentials(), &opts); err != nil { + return err + } + + d.InotifyWithParent(linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil } // RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. @@ -765,7 +796,12 @@ func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, if err != nil { return err } - return d.inode.removexattr(rp.Credentials(), name) + if err := d.inode.removexattr(rp.Credentials(), name); err != nil { + return err + } + + d.InotifyWithParent(linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil } // PrependPath implements vfs.FilesystemImpl.PrependPath. diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file.go b/pkg/sentry/fsimpl/tmpfs/regular_file.go index 3f433d666..fee174375 100644 --- a/pkg/sentry/fsimpl/tmpfs/regular_file.go +++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go @@ -312,7 +312,7 @@ func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, off f := fd.inode().impl.(*regularFile) if end := offset + srclen; end < offset { // Overflow. - return 0, syserror.EFBIG + return 0, syserror.EINVAL } var err error diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go index 1e781aecd..f0e098702 100644 --- a/pkg/sentry/fsimpl/tmpfs/tmpfs.go +++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go @@ -163,6 +163,11 @@ type dentry struct { // filesystem.mu. name string + // unlinked indicates whether this dentry has been unlinked from its parent. + // It is only set to true on an unlink operation, and never set from true to + // false. unlinked is protected by filesystem.mu. + unlinked bool + // dentryEntry (ugh) links dentries into their parent directory.childList. dentryEntry @@ -201,6 +206,26 @@ func (d *dentry) DecRef() { d.inode.decRef() } +// InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. +func (d *dentry) InotifyWithParent(events uint32, cookie uint32, et vfs.EventType) { + if d.inode.isDir() { + events |= linux.IN_ISDIR + } + + // The ordering below is important, Linux always notifies the parent first. + if d.parent != nil { + // Note that d.parent or d.name may be stale if there is a concurrent + // rename operation. Inotify does not provide consistency guarantees. + d.parent.inode.watches.NotifyWithExclusions(d.name, events, cookie, et, d.unlinked) + } + d.inode.watches.Notify("", events, cookie, et) +} + +// Watches implements vfs.DentryImpl.Watches. +func (d *dentry) Watches() *vfs.Watches { + return &d.inode.watches +} + // inode represents a filesystem object. type inode struct { // fs is the owning filesystem. fs is immutable. @@ -209,11 +234,9 @@ type inode struct { // refs is a reference count. refs is accessed using atomic memory // operations. // - // A reference is held on all inodes that are reachable in the filesystem - // tree. For non-directories (which may have multiple hard links), this - // means that a reference is dropped when nlink reaches 0. For directories, - // nlink never reaches 0 due to the "." entry; instead, - // filesystem.RmdirAt() drops the reference. + // A reference is held on all inodes as long as they are reachable in the + // filesystem tree, i.e. nlink is nonzero. This reference is dropped when + // nlink reaches 0. refs int64 // xattrs implements extended attributes. @@ -238,6 +261,9 @@ type inode struct { // Advisory file locks, which lock at the inode level. locks lock.FileLocks + // Inotify watches for this inode. + watches vfs.Watches + impl interface{} // immutable } @@ -259,6 +285,7 @@ func (i *inode) init(impl interface{}, fs *filesystem, creds *auth.Credentials, i.ctime = now i.mtime = now // i.nlink initialized by caller + i.watches = vfs.Watches{} i.impl = impl } @@ -276,14 +303,17 @@ func (i *inode) incLinksLocked() { atomic.AddUint32(&i.nlink, 1) } -// decLinksLocked decrements i's link count. +// decLinksLocked decrements i's link count. If the link count reaches 0, we +// remove a reference on i as well. // // Preconditions: filesystem.mu must be locked for writing. i.nlink != 0. func (i *inode) decLinksLocked() { if i.nlink == 0 { panic("tmpfs.inode.decLinksLocked() called with no existing links") } - atomic.AddUint32(&i.nlink, ^uint32(0)) + if atomic.AddUint32(&i.nlink, ^uint32(0)) == 0 { + i.decRef() + } } func (i *inode) incRef() { @@ -306,6 +336,7 @@ func (i *inode) tryIncRef() bool { func (i *inode) decRef() { if refs := atomic.AddInt64(&i.refs, -1); refs == 0 { + i.watches.HandleDeletion() if regFile, ok := i.impl.(*regularFile); ok { // Release memory used by regFile to store data. Since regFile is // no longer usable, we don't need to grab any locks or update any @@ -627,8 +658,12 @@ func (fd *fileDescription) filesystem() *filesystem { return fd.vfsfd.Mount().Filesystem().Impl().(*filesystem) } +func (fd *fileDescription) dentry() *dentry { + return fd.vfsfd.Dentry().Impl().(*dentry) +} + func (fd *fileDescription) inode() *inode { - return fd.vfsfd.Dentry().Impl().(*dentry).inode + return fd.dentry().inode } // Stat implements vfs.FileDescriptionImpl.Stat. @@ -641,7 +676,15 @@ func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linu // SetStat implements vfs.FileDescriptionImpl.SetStat. func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { creds := auth.CredentialsFromContext(ctx) - return fd.inode().setStat(ctx, creds, &opts.Stat) + d := fd.dentry() + if err := d.inode.setStat(ctx, creds, &opts.Stat); err != nil { + return err + } + + if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { + d.InotifyWithParent(ev, 0, vfs.InodeEvent) + } + return nil } // Listxattr implements vfs.FileDescriptionImpl.Listxattr. @@ -656,12 +699,26 @@ func (fd *fileDescription) Getxattr(ctx context.Context, opts vfs.GetxattrOption // Setxattr implements vfs.FileDescriptionImpl.Setxattr. func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error { - return fd.inode().setxattr(auth.CredentialsFromContext(ctx), &opts) + d := fd.dentry() + if err := d.inode.setxattr(auth.CredentialsFromContext(ctx), &opts); err != nil { + return err + } + + // Generate inotify events. + d.InotifyWithParent(linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil } // Removexattr implements vfs.FileDescriptionImpl.Removexattr. func (fd *fileDescription) Removexattr(ctx context.Context, name string) error { - return fd.inode().removexattr(auth.CredentialsFromContext(ctx), name) + d := fd.dentry() + if err := d.inode.removexattr(auth.CredentialsFromContext(ctx), name); err != nil { + return err + } + + // Generate inotify events. + d.InotifyWithParent(linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil } // NewMemfd creates a new tmpfs regular file and file description that can back diff --git a/pkg/sentry/kernel/fd_table.go b/pkg/sentry/kernel/fd_table.go index ed40b5303..dbfcef0fa 100644 --- a/pkg/sentry/kernel/fd_table.go +++ b/pkg/sentry/kernel/fd_table.go @@ -152,7 +152,13 @@ func (f *FDTable) drop(file *fs.File) { // dropVFS2 drops the table reference. func (f *FDTable) dropVFS2(file *vfs.FileDescription) { // TODO(gvisor.dev/issue/1480): Release locks. - // TODO(gvisor.dev/issue/1479): Send inotify events. + + // Generate inotify events. + ev := uint32(linux.IN_CLOSE_NOWRITE) + if file.IsWritable() { + ev = linux.IN_CLOSE_WRITE + } + file.Dentry().InotifyWithParent(ev, 0, vfs.PathEvent) // Drop the table reference. file.DecRef() diff --git a/pkg/sentry/kernel/pipe/BUILD b/pkg/sentry/kernel/pipe/BUILD index f29dc0472..7bfa9075a 100644 --- a/pkg/sentry/kernel/pipe/BUILD +++ b/pkg/sentry/kernel/pipe/BUILD @@ -8,6 +8,7 @@ go_library( "device.go", "node.go", "pipe.go", + "pipe_unsafe.go", "pipe_util.go", "reader.go", "reader_writer.go", @@ -20,6 +21,7 @@ go_library( "//pkg/amutex", "//pkg/buffer", "//pkg/context", + "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/device", "//pkg/sentry/fs", diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go index 62c8691f1..79645d7d2 100644 --- a/pkg/sentry/kernel/pipe/pipe.go +++ b/pkg/sentry/kernel/pipe/pipe.go @@ -207,7 +207,10 @@ func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) { p.mu.Lock() defer p.mu.Unlock() + return p.readLocked(ctx, ops) +} +func (p *Pipe) readLocked(ctx context.Context, ops readOps) (int64, error) { // Is the pipe empty? if p.view.Size() == 0 { if !p.HasWriters() { @@ -246,7 +249,10 @@ type writeOps struct { func (p *Pipe) write(ctx context.Context, ops writeOps) (int64, error) { p.mu.Lock() defer p.mu.Unlock() + return p.writeLocked(ctx, ops) +} +func (p *Pipe) writeLocked(ctx context.Context, ops writeOps) (int64, error) { // Can't write to a pipe with no readers. if !p.HasReaders() { return 0, syscall.EPIPE diff --git a/pkg/sentry/kernel/pipe/pipe_unsafe.go b/pkg/sentry/kernel/pipe/pipe_unsafe.go new file mode 100644 index 000000000..dd60cba24 --- /dev/null +++ b/pkg/sentry/kernel/pipe/pipe_unsafe.go @@ -0,0 +1,35 @@ +// Copyright 2019 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 ( + "unsafe" +) + +// lockTwoPipes locks both x.mu and y.mu in an order that is guaranteed to be +// consistent for both lockTwoPipes(x, y) and lockTwoPipes(y, x), such that +// concurrent calls cannot deadlock. +// +// Preconditions: x != y. +func lockTwoPipes(x, y *Pipe) { + // Lock the two pipes in order of increasing address. + if uintptr(unsafe.Pointer(x)) < uintptr(unsafe.Pointer(y)) { + x.mu.Lock() + y.mu.Lock() + } else { + y.mu.Lock() + x.mu.Lock() + } +} diff --git a/pkg/sentry/kernel/pipe/vfs.go b/pkg/sentry/kernel/pipe/vfs.go index b54f08a30..2602bed72 100644 --- a/pkg/sentry/kernel/pipe/vfs.go +++ b/pkg/sentry/kernel/pipe/vfs.go @@ -16,7 +16,9 @@ package pipe import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" @@ -150,7 +152,9 @@ func (vp *VFSPipe) newFD(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32) * return &fd.vfsfd } -// VFSPipeFD implements vfs.FileDescriptionImpl for pipes. +// VFSPipeFD implements vfs.FileDescriptionImpl for pipes. It also implements +// non-atomic usermem.IO methods, allowing it to be passed as usermem.IO to +// other FileDescriptions for splice(2) and tee(2). type VFSPipeFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -229,3 +233,216 @@ func (fd *VFSPipeFD) PipeSize() int64 { func (fd *VFSPipeFD) SetPipeSize(size int64) (int64, error) { return fd.pipe.SetFifoSize(size) } + +// IOSequence returns a useremm.IOSequence that reads up to count bytes from, +// or writes up to count bytes to, fd. +func (fd *VFSPipeFD) IOSequence(count int64) usermem.IOSequence { + return usermem.IOSequence{ + IO: fd, + Addrs: usermem.AddrRangeSeqOf(usermem.AddrRange{0, usermem.Addr(count)}), + } +} + +// CopyIn implements usermem.IO.CopyIn. +func (fd *VFSPipeFD) CopyIn(ctx context.Context, addr usermem.Addr, dst []byte, opts usermem.IOOpts) (int, error) { + origCount := int64(len(dst)) + n, err := fd.pipe.read(ctx, readOps{ + left: func() int64 { + return int64(len(dst)) + }, + limit: func(l int64) { + dst = dst[:l] + }, + read: func(view *buffer.View) (int64, error) { + n, err := view.ReadAt(dst, 0) + view.TrimFront(int64(n)) + return int64(n), err + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventOut) + } + if err == nil && n != origCount { + return int(n), syserror.ErrWouldBlock + } + return int(n), err +} + +// CopyOut implements usermem.IO.CopyOut. +func (fd *VFSPipeFD) CopyOut(ctx context.Context, addr usermem.Addr, src []byte, opts usermem.IOOpts) (int, error) { + origCount := int64(len(src)) + n, err := fd.pipe.write(ctx, writeOps{ + left: func() int64 { + return int64(len(src)) + }, + limit: func(l int64) { + src = src[:l] + }, + write: func(view *buffer.View) (int64, error) { + view.Append(src) + return int64(len(src)), nil + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventIn) + } + if err == nil && n != origCount { + return int(n), syserror.ErrWouldBlock + } + return int(n), err +} + +// ZeroOut implements usermem.IO.ZeroOut. +func (fd *VFSPipeFD) ZeroOut(ctx context.Context, addr usermem.Addr, toZero int64, opts usermem.IOOpts) (int64, error) { + origCount := toZero + n, err := fd.pipe.write(ctx, writeOps{ + left: func() int64 { + return toZero + }, + limit: func(l int64) { + toZero = l + }, + write: func(view *buffer.View) (int64, error) { + view.Grow(view.Size()+toZero, true /* zero */) + return toZero, nil + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventIn) + } + if err == nil && n != origCount { + return n, syserror.ErrWouldBlock + } + return n, err +} + +// CopyInTo implements usermem.IO.CopyInTo. +func (fd *VFSPipeFD) CopyInTo(ctx context.Context, ars usermem.AddrRangeSeq, dst safemem.Writer, opts usermem.IOOpts) (int64, error) { + count := ars.NumBytes() + if count == 0 { + return 0, nil + } + origCount := count + n, err := fd.pipe.read(ctx, readOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + read: func(view *buffer.View) (int64, error) { + n, err := view.ReadToSafememWriter(dst, uint64(count)) + view.TrimFront(int64(n)) + return int64(n), err + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventOut) + } + if err == nil && n != origCount { + return n, syserror.ErrWouldBlock + } + return n, err +} + +// CopyOutFrom implements usermem.IO.CopyOutFrom. +func (fd *VFSPipeFD) CopyOutFrom(ctx context.Context, ars usermem.AddrRangeSeq, src safemem.Reader, opts usermem.IOOpts) (int64, error) { + count := ars.NumBytes() + if count == 0 { + return 0, nil + } + origCount := count + n, err := fd.pipe.write(ctx, writeOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + write: func(view *buffer.View) (int64, error) { + n, err := view.WriteFromSafememReader(src, uint64(count)) + return int64(n), err + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventIn) + } + if err == nil && n != origCount { + return n, syserror.ErrWouldBlock + } + return n, err +} + +// SwapUint32 implements usermem.IO.SwapUint32. +func (fd *VFSPipeFD) SwapUint32(ctx context.Context, addr usermem.Addr, new uint32, opts usermem.IOOpts) (uint32, error) { + // How did a pipe get passed as the virtual address space to futex(2)? + panic("VFSPipeFD.SwapUint32 called unexpectedly") +} + +// CompareAndSwapUint32 implements usermem.IO.CompareAndSwapUint32. +func (fd *VFSPipeFD) CompareAndSwapUint32(ctx context.Context, addr usermem.Addr, old, new uint32, opts usermem.IOOpts) (uint32, error) { + panic("VFSPipeFD.CompareAndSwapUint32 called unexpectedly") +} + +// LoadUint32 implements usermem.IO.LoadUint32. +func (fd *VFSPipeFD) LoadUint32(ctx context.Context, addr usermem.Addr, opts usermem.IOOpts) (uint32, error) { + panic("VFSPipeFD.LoadUint32 called unexpectedly") +} + +// Splice reads up to count bytes from src and writes them to dst. It returns +// the number of bytes moved. +// +// Preconditions: count > 0. +func Splice(ctx context.Context, dst, src *VFSPipeFD, count int64) (int64, error) { + return spliceOrTee(ctx, dst, src, count, true /* removeFromSrc */) +} + +// Tee reads up to count bytes from src and writes them to dst, without +// removing the read bytes from src. It returns the number of bytes copied. +// +// Preconditions: count > 0. +func Tee(ctx context.Context, dst, src *VFSPipeFD, count int64) (int64, error) { + return spliceOrTee(ctx, dst, src, count, false /* removeFromSrc */) +} + +// Preconditions: count > 0. +func spliceOrTee(ctx context.Context, dst, src *VFSPipeFD, count int64, removeFromSrc bool) (int64, error) { + if dst.pipe == src.pipe { + return 0, syserror.EINVAL + } + + lockTwoPipes(dst.pipe, src.pipe) + defer dst.pipe.mu.Unlock() + defer src.pipe.mu.Unlock() + + n, err := dst.pipe.writeLocked(ctx, writeOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + write: func(dstView *buffer.View) (int64, error) { + return src.pipe.readLocked(ctx, readOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + read: func(srcView *buffer.View) (int64, error) { + n, err := srcView.ReadToSafememWriter(dstView, uint64(count)) + if n > 0 && removeFromSrc { + srcView.TrimFront(int64(n)) + } + return int64(n), err + }, + }) + }, + }) + if n > 0 { + dst.pipe.Notify(waiter.EventIn) + src.pipe.Notify(waiter.EventOut) + } + return n, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/BUILD b/pkg/sentry/syscalls/linux/vfs2/BUILD index f882ef840..9c8b44f64 100644 --- a/pkg/sentry/syscalls/linux/vfs2/BUILD +++ b/pkg/sentry/syscalls/linux/vfs2/BUILD @@ -12,6 +12,7 @@ go_library( "filesystem.go", "fscontext.go", "getdents.go", + "inotify.go", "ioctl.go", "memfd.go", "mmap.go", @@ -22,6 +23,7 @@ go_library( "setstat.go", "signal.go", "socket.go", + "splice.go", "stat.go", "stat_amd64.go", "stat_arm64.go", diff --git a/pkg/sentry/syscalls/linux/vfs2/inotify.go b/pkg/sentry/syscalls/linux/vfs2/inotify.go new file mode 100644 index 000000000..7d50b6a16 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/inotify.go @@ -0,0 +1,134 @@ +// 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 vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +const allFlags = linux.IN_NONBLOCK | linux.IN_CLOEXEC + +// InotifyInit1 implements the inotify_init1() syscalls. +func InotifyInit1(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + flags := args[0].Int() + if flags&^allFlags != 0 { + return 0, nil, syserror.EINVAL + } + + ino, err := vfs.NewInotifyFD(t, t.Kernel().VFS(), uint32(flags)) + if err != nil { + return 0, nil, err + } + defer ino.DecRef() + + fd, err := t.NewFDFromVFS2(0, ino, kernel.FDFlags{ + CloseOnExec: flags&linux.IN_CLOEXEC != 0, + }) + + if err != nil { + return 0, nil, err + } + + return uintptr(fd), nil, nil +} + +// InotifyInit implements the inotify_init() syscalls. +func InotifyInit(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + args[0].Value = 0 + return InotifyInit1(t, args) +} + +// fdToInotify resolves an fd to an inotify object. If successful, the file will +// have an extra ref and the caller is responsible for releasing the ref. +func fdToInotify(t *kernel.Task, fd int32) (*vfs.Inotify, *vfs.FileDescription, error) { + f := t.GetFileVFS2(fd) + if f == nil { + // Invalid fd. + return nil, nil, syserror.EBADF + } + + ino, ok := f.Impl().(*vfs.Inotify) + if !ok { + // Not an inotify fd. + f.DecRef() + return nil, nil, syserror.EINVAL + } + + return ino, f, nil +} + +// InotifyAddWatch implements the inotify_add_watch() syscall. +func InotifyAddWatch(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + mask := args[2].Uint() + + // "EINVAL: The given event mask contains no valid events." + // -- inotify_add_watch(2) + if validBits := mask & linux.ALL_INOTIFY_BITS; validBits == 0 { + return 0, nil, syserror.EINVAL + } + + // "IN_DONT_FOLLOW: Don't dereference pathname if it is a symbolic link." + // -- inotify(7) + follow := followFinalSymlink + if mask&linux.IN_DONT_FOLLOW == 0 { + follow = nofollowFinalSymlink + } + + ino, f, err := fdToInotify(t, fd) + if err != nil { + return 0, nil, err + } + defer f.DecRef() + + path, err := copyInPath(t, addr) + if err != nil { + return 0, nil, err + } + if mask&linux.IN_ONLYDIR != 0 { + path.Dir = true + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, follow) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + d, err := t.Kernel().VFS().GetDentryAt(t, t.Credentials(), &tpop.pop, &vfs.GetDentryOptions{}) + if err != nil { + return 0, nil, err + } + defer d.DecRef() + + fd = ino.AddWatch(d.Dentry(), mask) + return uintptr(fd), nil, err +} + +// InotifyRmWatch implements the inotify_rm_watch() syscall. +func InotifyRmWatch(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + wd := args[1].Int() + + ino, f, err := fdToInotify(t, fd) + if err != nil { + return 0, nil, err + } + defer f.DecRef() + return 0, nil, ino.RmWatch(wd) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/read_write.go b/pkg/sentry/syscalls/linux/vfs2/read_write.go index 3a7ef24f5..7f9debd4a 100644 --- a/pkg/sentry/syscalls/linux/vfs2/read_write.go +++ b/pkg/sentry/syscalls/linux/vfs2/read_write.go @@ -93,11 +93,17 @@ func Readv(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall func read(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { n, err := file.Read(t, dst, opts) if err != syserror.ErrWouldBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } return n, err } allowBlock, deadline, hasDeadline := blockPolicy(t, file) if !allowBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } return n, err } @@ -128,6 +134,9 @@ func read(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, opt } file.EventUnregister(&w) + if total > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } return total, err } @@ -248,11 +257,17 @@ func Preadv2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca func pread(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { n, err := file.PRead(t, dst, offset, opts) if err != syserror.ErrWouldBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } return n, err } allowBlock, deadline, hasDeadline := blockPolicy(t, file) if !allowBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } return n, err } @@ -283,6 +298,9 @@ func pread(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, of } file.EventUnregister(&w) + if total > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } return total, err } @@ -345,11 +363,17 @@ func Writev(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal func write(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { n, err := file.Write(t, src, opts) if err != syserror.ErrWouldBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_MODIFY, 0, vfs.PathEvent) + } return n, err } allowBlock, deadline, hasDeadline := blockPolicy(t, file) if !allowBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_MODIFY, 0, vfs.PathEvent) + } return n, err } @@ -380,6 +404,9 @@ func write(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, op } file.EventUnregister(&w) + if total > 0 { + file.Dentry().InotifyWithParent(linux.IN_MODIFY, 0, vfs.PathEvent) + } return total, err } @@ -500,11 +527,17 @@ func Pwritev2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc func pwrite(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { n, err := file.PWrite(t, src, offset, opts) if err != syserror.ErrWouldBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_MODIFY, 0, vfs.PathEvent) + } return n, err } allowBlock, deadline, hasDeadline := blockPolicy(t, file) if !allowBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } return n, err } @@ -535,6 +568,9 @@ func pwrite(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, o } file.EventUnregister(&w) + if total > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } return total, err } diff --git a/pkg/sentry/syscalls/linux/vfs2/splice.go b/pkg/sentry/syscalls/linux/vfs2/splice.go new file mode 100644 index 000000000..8f3c22a02 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/splice.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 vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/waiter" +) + +// Splice implements Linux syscall splice(2). +func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + inFD := args[0].Int() + inOffsetPtr := args[1].Pointer() + outFD := args[2].Int() + outOffsetPtr := args[3].Pointer() + count := int64(args[4].SizeT()) + flags := args[5].Int() + + if count == 0 { + return 0, nil, nil + } + if count > int64(kernel.MAX_RW_COUNT) { + count = int64(kernel.MAX_RW_COUNT) + } + + // Check for invalid flags. + if flags&^(linux.SPLICE_F_MOVE|linux.SPLICE_F_NONBLOCK|linux.SPLICE_F_MORE|linux.SPLICE_F_GIFT) != 0 { + return 0, nil, syserror.EINVAL + } + + // Get file descriptions. + inFile := t.GetFileVFS2(inFD) + if inFile == nil { + return 0, nil, syserror.EBADF + } + defer inFile.DecRef() + outFile := t.GetFileVFS2(outFD) + if outFile == nil { + return 0, nil, syserror.EBADF + } + defer outFile.DecRef() + + // Check that both files support the required directionality. + if !inFile.IsReadable() || !outFile.IsWritable() { + return 0, nil, syserror.EBADF + } + + // The operation is non-blocking if anything is non-blocking. + // + // N.B. This is a rather simplistic heuristic that avoids some + // poor edge case behavior since the exact semantics here are + // underspecified and vary between versions of Linux itself. + nonBlock := ((inFile.StatusFlags()|outFile.StatusFlags())&linux.O_NONBLOCK != 0) || (flags&linux.SPLICE_F_NONBLOCK != 0) + + // At least one file description must represent a pipe. + inPipeFD, inIsPipe := inFile.Impl().(*pipe.VFSPipeFD) + outPipeFD, outIsPipe := outFile.Impl().(*pipe.VFSPipeFD) + if !inIsPipe && !outIsPipe { + return 0, nil, syserror.EINVAL + } + + // Copy in offsets. + inOffset := int64(-1) + if inOffsetPtr != 0 { + if inIsPipe { + return 0, nil, syserror.ESPIPE + } + if inFile.Options().DenyPRead { + return 0, nil, syserror.EINVAL + } + if _, err := t.CopyIn(inOffsetPtr, &inOffset); err != nil { + return 0, nil, err + } + if inOffset < 0 { + return 0, nil, syserror.EINVAL + } + } + outOffset := int64(-1) + if outOffsetPtr != 0 { + if outIsPipe { + return 0, nil, syserror.ESPIPE + } + if outFile.Options().DenyPWrite { + return 0, nil, syserror.EINVAL + } + if _, err := t.CopyIn(outOffsetPtr, &outOffset); err != nil { + return 0, nil, err + } + if outOffset < 0 { + return 0, nil, syserror.EINVAL + } + } + + // Move data. + var ( + n int64 + err error + inCh chan struct{} + outCh chan struct{} + ) + for { + // If both input and output are pipes, delegate to the pipe + // implementation. Otherwise, exactly one end is a pipe, which we + // ensure is consistently ordered after the non-pipe FD's locks by + // passing the pipe FD as usermem.IO to the non-pipe end. + switch { + case inIsPipe && outIsPipe: + n, err = pipe.Splice(t, outPipeFD, inPipeFD, count) + case inIsPipe: + if outOffset != -1 { + n, err = outFile.PWrite(t, inPipeFD.IOSequence(count), outOffset, vfs.WriteOptions{}) + outOffset += n + } else { + n, err = outFile.Write(t, inPipeFD.IOSequence(count), vfs.WriteOptions{}) + } + case outIsPipe: + if inOffset != -1 { + n, err = inFile.PRead(t, outPipeFD.IOSequence(count), inOffset, vfs.ReadOptions{}) + inOffset += n + } else { + n, err = inFile.Read(t, outPipeFD.IOSequence(count), vfs.ReadOptions{}) + } + } + if n != 0 || err != syserror.ErrWouldBlock || nonBlock { + break + } + + // Note that the blocking behavior here is a bit different than the + // normal pattern. Because we need to have both data to read and data + // to write simultaneously, we actually explicitly block on both of + // these cases in turn before returning to the splice operation. + if inFile.Readiness(eventMaskRead)&eventMaskRead == 0 { + if inCh == nil { + inCh = make(chan struct{}, 1) + inW, _ := waiter.NewChannelEntry(inCh) + inFile.EventRegister(&inW, eventMaskRead) + defer inFile.EventUnregister(&inW) + continue // Need to refresh readiness. + } + if err = t.Block(inCh); err != nil { + break + } + } + if outFile.Readiness(eventMaskWrite)&eventMaskWrite == 0 { + if outCh == nil { + outCh = make(chan struct{}, 1) + outW, _ := waiter.NewChannelEntry(outCh) + outFile.EventRegister(&outW, eventMaskWrite) + defer outFile.EventUnregister(&outW) + continue // Need to refresh readiness. + } + if err = t.Block(outCh); err != nil { + break + } + } + } + + // Copy updated offsets out. + if inOffsetPtr != 0 { + if _, err := t.CopyOut(inOffsetPtr, &inOffset); err != nil { + return 0, nil, err + } + } + if outOffsetPtr != 0 { + if _, err := t.CopyOut(outOffsetPtr, &outOffset); err != nil { + return 0, nil, err + } + } + + if n == 0 { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Tee implements Linux syscall tee(2). +func Tee(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + inFD := args[0].Int() + outFD := args[1].Int() + count := int64(args[2].SizeT()) + flags := args[3].Int() + + if count == 0 { + return 0, nil, nil + } + if count > int64(kernel.MAX_RW_COUNT) { + count = int64(kernel.MAX_RW_COUNT) + } + + // Check for invalid flags. + if flags&^(linux.SPLICE_F_MOVE|linux.SPLICE_F_NONBLOCK|linux.SPLICE_F_MORE|linux.SPLICE_F_GIFT) != 0 { + return 0, nil, syserror.EINVAL + } + + // Get file descriptions. + inFile := t.GetFileVFS2(inFD) + if inFile == nil { + return 0, nil, syserror.EBADF + } + defer inFile.DecRef() + outFile := t.GetFileVFS2(outFD) + if outFile == nil { + return 0, nil, syserror.EBADF + } + defer outFile.DecRef() + + // Check that both files support the required directionality. + if !inFile.IsReadable() || !outFile.IsWritable() { + return 0, nil, syserror.EBADF + } + + // The operation is non-blocking if anything is non-blocking. + // + // N.B. This is a rather simplistic heuristic that avoids some + // poor edge case behavior since the exact semantics here are + // underspecified and vary between versions of Linux itself. + nonBlock := ((inFile.StatusFlags()|outFile.StatusFlags())&linux.O_NONBLOCK != 0) || (flags&linux.SPLICE_F_NONBLOCK != 0) + + // Both file descriptions must represent pipes. + inPipeFD, inIsPipe := inFile.Impl().(*pipe.VFSPipeFD) + outPipeFD, outIsPipe := outFile.Impl().(*pipe.VFSPipeFD) + if !inIsPipe || !outIsPipe { + return 0, nil, syserror.EINVAL + } + + // Copy data. + var ( + inCh chan struct{} + outCh chan struct{} + ) + for { + n, err := pipe.Tee(t, outPipeFD, inPipeFD, count) + if n != 0 { + return uintptr(n), nil, nil + } + if err != syserror.ErrWouldBlock || nonBlock { + return 0, nil, err + } + + // Note that the blocking behavior here is a bit different than the + // normal pattern. Because we need to have both data to read and data + // to write simultaneously, we actually explicitly block on both of + // these cases in turn before returning to the tee operation. + if inFile.Readiness(eventMaskRead)&eventMaskRead == 0 { + if inCh == nil { + inCh = make(chan struct{}, 1) + inW, _ := waiter.NewChannelEntry(inCh) + inFile.EventRegister(&inW, eventMaskRead) + defer inFile.EventUnregister(&inW) + continue // Need to refresh readiness. + } + if err := t.Block(inCh); err != nil { + return 0, nil, err + } + } + if outFile.Readiness(eventMaskWrite)&eventMaskWrite == 0 { + if outCh == nil { + outCh = make(chan struct{}, 1) + outW, _ := waiter.NewChannelEntry(outCh) + outFile.EventRegister(&outW, eventMaskWrite) + defer outFile.EventUnregister(&outW) + continue // Need to refresh readiness. + } + if err := t.Block(outCh); err != nil { + return 0, nil, err + } + } + } +} diff --git a/pkg/sentry/syscalls/linux/vfs2/vfs2.go b/pkg/sentry/syscalls/linux/vfs2/vfs2.go index a332d01bd..ef8358b8a 100644 --- a/pkg/sentry/syscalls/linux/vfs2/vfs2.go +++ b/pkg/sentry/syscalls/linux/vfs2/vfs2.go @@ -116,9 +116,9 @@ func Override() { s.Table[232] = syscalls.Supported("epoll_wait", EpollWait) s.Table[233] = syscalls.Supported("epoll_ctl", EpollCtl) s.Table[235] = syscalls.Supported("utimes", Utimes) - delete(s.Table, 253) // inotify_init - delete(s.Table, 254) // inotify_add_watch - delete(s.Table, 255) // inotify_rm_watch + s.Table[253] = syscalls.PartiallySupported("inotify_init", InotifyInit, "inotify events are only available inside the sandbox.", nil) + s.Table[254] = syscalls.PartiallySupported("inotify_add_watch", InotifyAddWatch, "inotify events are only available inside the sandbox.", nil) + s.Table[255] = syscalls.PartiallySupported("inotify_rm_watch", InotifyRmWatch, "inotify events are only available inside the sandbox.", nil) s.Table[257] = syscalls.Supported("openat", Openat) s.Table[258] = syscalls.Supported("mkdirat", Mkdirat) s.Table[259] = syscalls.Supported("mknodat", Mknodat) @@ -134,8 +134,8 @@ func Override() { s.Table[269] = syscalls.Supported("faccessat", Faccessat) s.Table[270] = syscalls.Supported("pselect", Pselect) s.Table[271] = syscalls.Supported("ppoll", Ppoll) - delete(s.Table, 275) // splice - delete(s.Table, 276) // tee + s.Table[275] = syscalls.Supported("splice", Splice) + s.Table[276] = syscalls.Supported("tee", Tee) s.Table[277] = syscalls.Supported("sync_file_range", SyncFileRange) s.Table[280] = syscalls.Supported("utimensat", Utimensat) s.Table[281] = syscalls.Supported("epoll_pwait", EpollPwait) @@ -151,7 +151,7 @@ func Override() { s.Table[291] = syscalls.Supported("epoll_create1", EpollCreate1) s.Table[292] = syscalls.Supported("dup3", Dup3) s.Table[293] = syscalls.Supported("pipe2", Pipe2) - delete(s.Table, 294) // inotify_init1 + s.Table[294] = syscalls.PartiallySupported("inotify_init1", InotifyInit1, "inotify events are only available inside the sandbox.", nil) s.Table[295] = syscalls.Supported("preadv", Preadv) s.Table[296] = syscalls.Supported("pwritev", Pwritev) s.Table[299] = syscalls.Supported("recvmmsg", RecvMMsg) diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index 94d69c1cc..774cc66cc 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -15,6 +15,18 @@ go_template_instance( }, ) +go_template_instance( + name = "event_list", + out = "event_list.go", + package = "vfs", + prefix = "event", + template = "//pkg/ilist:generic_list", + types = { + "Element": "*Event", + "Linker": "*Event", + }, +) + go_library( name = "vfs", srcs = [ @@ -25,11 +37,13 @@ go_library( "device.go", "epoll.go", "epoll_interest_list.go", + "event_list.go", "file_description.go", "file_description_impl_util.go", "filesystem.go", "filesystem_impl_util.go", "filesystem_type.go", + "inotify.go", "mount.go", "mount_unsafe.go", "options.go", @@ -57,6 +71,7 @@ go_library( "//pkg/sentry/limits", "//pkg/sentry/memmap", "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/uniqueid", "//pkg/sync", "//pkg/syserror", "//pkg/usermem", diff --git a/pkg/sentry/vfs/anonfs.go b/pkg/sentry/vfs/anonfs.go index caf770fd5..b7c6b60b8 100644 --- a/pkg/sentry/vfs/anonfs.go +++ b/pkg/sentry/vfs/anonfs.go @@ -297,3 +297,15 @@ func (d *anonDentry) TryIncRef() bool { func (d *anonDentry) DecRef() { // no-op } + +// InotifyWithParent implements DentryImpl.InotifyWithParent. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *anonDentry) InotifyWithParent(events uint32, cookie uint32, et EventType) {} + +// Watches implements DentryImpl.Watches. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *anonDentry) Watches() *Watches { + return nil +} diff --git a/pkg/sentry/vfs/dentry.go b/pkg/sentry/vfs/dentry.go index 8624dbd5d..24af13eb1 100644 --- a/pkg/sentry/vfs/dentry.go +++ b/pkg/sentry/vfs/dentry.go @@ -103,6 +103,22 @@ type DentryImpl interface { // DecRef decrements the Dentry's reference count. DecRef() + + // InotifyWithParent notifies all watches on the targets represented by this + // dentry and its parent. The parent's watches are notified first, followed + // by this dentry's. + // + // InotifyWithParent automatically adds the IN_ISDIR flag for dentries + // representing directories. + // + // Note that the events may not actually propagate up to the user, depending + // on the event masks. + InotifyWithParent(events uint32, cookie uint32, et EventType) + + // Watches returns the set of inotify watches for the file corresponding to + // the Dentry. Dentries that are hard links to the same underlying file + // share the same watches. + Watches() *Watches } // IncRef increments d's reference count. @@ -133,6 +149,17 @@ func (d *Dentry) isMounted() bool { return atomic.LoadUint32(&d.mounts) != 0 } +// InotifyWithParent notifies all watches on the inodes for this dentry and +// its parent of events. +func (d *Dentry) InotifyWithParent(events uint32, cookie uint32, et EventType) { + d.impl.InotifyWithParent(events, cookie, et) +} + +// Watches returns the set of inotify watches associated with d. +func (d *Dentry) Watches() *Watches { + return d.impl.Watches() +} + // The following functions are exported so that filesystem implementations can // use them. The vfs package, and users of VFS, should not call these // functions. diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go index cfabd936c..bb294563d 100644 --- a/pkg/sentry/vfs/file_description.go +++ b/pkg/sentry/vfs/file_description.go @@ -210,6 +210,11 @@ func (fd *FileDescription) VirtualDentry() VirtualDentry { return fd.vd } +// Options returns the options passed to fd.Init(). +func (fd *FileDescription) Options() FileDescriptionOptions { + return fd.opts +} + // StatusFlags returns file description status flags, as for fcntl(F_GETFL). func (fd *FileDescription) StatusFlags() uint32 { return atomic.LoadUint32(&fd.statusFlags) diff --git a/pkg/sentry/vfs/inotify.go b/pkg/sentry/vfs/inotify.go new file mode 100644 index 000000000..05a3051a4 --- /dev/null +++ b/pkg/sentry/vfs/inotify.go @@ -0,0 +1,697 @@ +// 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 vfs + +import ( + "bytes" + "fmt" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/uniqueid" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// inotifyEventBaseSize is the base size of linux's struct inotify_event. This +// must be a power 2 for rounding below. +const inotifyEventBaseSize = 16 + +// EventType defines different kinds of inotfiy events. +// +// The way events are labelled appears somewhat arbitrary, but they must match +// Linux so that IN_EXCL_UNLINK behaves as it does in Linux. +type EventType uint8 + +// PathEvent and InodeEvent correspond to FSNOTIFY_EVENT_PATH and +// FSNOTIFY_EVENT_INODE in Linux. +const ( + PathEvent EventType = iota + InodeEvent EventType = iota +) + +// Inotify represents an inotify instance created by inotify_init(2) or +// inotify_init1(2). Inotify implements FileDescriptionImpl. +// +// Lock ordering: +// Inotify.mu -> Watches.mu -> Inotify.evMu +// +// +stateify savable +type Inotify struct { + vfsfd FileDescription + FileDescriptionDefaultImpl + DentryMetadataFileDescriptionImpl + + // Unique identifier for this inotify instance. We don't just reuse the + // inotify fd because fds can be duped. These should not be exposed to the + // user, since we may aggressively reuse an id on S/R. + id uint64 + + // queue is used to notify interested parties when the inotify instance + // becomes readable or writable. + queue waiter.Queue `state:"nosave"` + + // evMu *only* protects the events list. We need a separate lock while + // queuing events: using mu may violate lock ordering, since at that point + // the calling goroutine may already hold Watches.mu. + evMu sync.Mutex `state:"nosave"` + + // A list of pending events for this inotify instance. Protected by evMu. + events eventList + + // A scratch buffer, used to serialize inotify events. Allocate this + // ahead of time for the sake of performance. Protected by evMu. + scratch []byte + + // mu protects the fields below. + mu sync.Mutex `state:"nosave"` + + // nextWatchMinusOne is used to allocate watch descriptors on this Inotify + // instance. Note that Linux starts numbering watch descriptors from 1. + nextWatchMinusOne int32 + + // Map from watch descriptors to watch objects. + watches map[int32]*Watch +} + +var _ FileDescriptionImpl = (*Inotify)(nil) + +// NewInotifyFD constructs a new Inotify instance. +func NewInotifyFD(ctx context.Context, vfsObj *VirtualFilesystem, flags uint32) (*FileDescription, error) { + // O_CLOEXEC affects file descriptors, so it must be handled outside of vfs. + flags &^= linux.O_CLOEXEC + if flags&^linux.O_NONBLOCK != 0 { + return nil, syserror.EINVAL + } + + id := uniqueid.GlobalFromContext(ctx) + vd := vfsObj.NewAnonVirtualDentry(fmt.Sprintf("[inotifyfd:%d]", id)) + defer vd.DecRef() + fd := &Inotify{ + id: id, + scratch: make([]byte, inotifyEventBaseSize), + watches: make(map[int32]*Watch), + } + if err := fd.vfsfd.Init(fd, flags, vd.Mount(), vd.Dentry(), &FileDescriptionOptions{ + UseDentryMetadata: true, + DenyPRead: true, + DenyPWrite: true, + }); err != nil { + return nil, err + } + return &fd.vfsfd, nil +} + +// Release implements FileDescriptionImpl.Release. Release removes all +// watches and frees all resources for an inotify instance. +func (i *Inotify) Release() { + // We need to hold i.mu to avoid a race with concurrent calls to + // Inotify.handleDeletion from Watches. There's no risk of Watches + // accessing this Inotify after the destructor ends, because we remove all + // references to it below. + i.mu.Lock() + defer i.mu.Unlock() + for _, w := range i.watches { + // Remove references to the watch from the watches set on the target. We + // don't need to worry about the references from i.watches, since this + // file description is about to be destroyed. + w.set.Remove(i.id) + } +} + +// EventRegister implements waiter.Waitable. +func (i *Inotify) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + i.queue.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable. +func (i *Inotify) EventUnregister(e *waiter.Entry) { + i.queue.EventUnregister(e) +} + +// Readiness implements waiter.Waitable.Readiness. +// +// Readiness indicates whether there are pending events for an inotify instance. +func (i *Inotify) Readiness(mask waiter.EventMask) waiter.EventMask { + ready := waiter.EventMask(0) + + i.evMu.Lock() + defer i.evMu.Unlock() + + if !i.events.Empty() { + ready |= waiter.EventIn + } + + return mask & ready +} + +// PRead implements FileDescriptionImpl. +func (*Inotify) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts ReadOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// PWrite implements FileDescriptionImpl. +func (*Inotify) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts WriteOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// Write implements FileDescriptionImpl.Write. +func (*Inotify) Write(ctx context.Context, src usermem.IOSequence, opts WriteOptions) (int64, error) { + return 0, syserror.EBADF +} + +// Read implements FileDescriptionImpl.Read. +func (i *Inotify) Read(ctx context.Context, dst usermem.IOSequence, opts ReadOptions) (int64, error) { + if dst.NumBytes() < inotifyEventBaseSize { + return 0, syserror.EINVAL + } + + i.evMu.Lock() + defer i.evMu.Unlock() + + if i.events.Empty() { + // Nothing to read yet, tell caller to block. + return 0, syserror.ErrWouldBlock + } + + var writeLen int64 + for it := i.events.Front(); it != nil; { + // Advance `it` before the element is removed from the list, or else + // it.Next() will always be 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()) { + if writeLen > 0 { + // Buffer wasn't big enough for all pending events, but we did + // write some events out. + return writeLen, nil + } + return 0, syserror.EINVAL + } + + // Linux always dequeues an available event as long as there's enough + // buffer space to copy it out, even if the copy below fails. Emulate + // this behaviour. + i.events.Remove(event) + + // Buffer has enough space, copy event to the read buffer. + n, err := event.CopyTo(ctx, i.scratch, dst) + if err != nil { + return 0, err + } + + writeLen += n + dst = dst.DropFirst64(n) + } + return writeLen, nil +} + +// Ioctl implements fs.FileOperations.Ioctl. +func (i *Inotify) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { + switch args[1].Int() { + case linux.FIONREAD: + i.evMu.Lock() + defer i.evMu.Unlock() + var n uint32 + for e := i.events.Front(); e != nil; e = e.Next() { + n += uint32(e.sizeOf()) + } + var buf [4]byte + usermem.ByteOrder.PutUint32(buf[:], n) + _, err := uio.CopyOut(ctx, args[2].Pointer(), buf[:], usermem.IOOpts{}) + return 0, err + + default: + return 0, syserror.ENOTTY + } +} + +func (i *Inotify) queueEvent(ev *Event) { + i.evMu.Lock() + + // Check if we should coalesce the event we're about to queue with the last + // one currently in the queue. Events are coalesced if they are identical. + if last := i.events.Back(); last != nil { + if ev.equals(last) { + // "Coalesce" the two events by simply not queuing the new one. We + // don't need to raise a waiter.EventIn notification because no new + // data is available for reading. + i.evMu.Unlock() + return + } + } + + i.events.PushBack(ev) + + // Release mutex before notifying waiters because we don't control what they + // can do. + i.evMu.Unlock() + + i.queue.Notify(waiter.EventIn) +} + +// newWatchLocked creates and adds a new watch to target. +// +// Precondition: i.mu must be locked. +func (i *Inotify) newWatchLocked(target *Dentry, mask uint32) *Watch { + targetWatches := target.Watches() + w := &Watch{ + owner: i, + wd: i.nextWatchIDLocked(), + set: targetWatches, + mask: mask, + } + + // Hold the watch in this inotify instance as well as the watch set on the + // target. + i.watches[w.wd] = w + targetWatches.Add(w) + return w +} + +// newWatchIDLocked allocates and returns a new watch descriptor. +// +// Precondition: i.mu must be locked. +func (i *Inotify) nextWatchIDLocked() int32 { + i.nextWatchMinusOne++ + return i.nextWatchMinusOne +} + +// handleDeletion handles the deletion of the target of watch w. It removes w +// from i.watches and a watch removal event is generated. +func (i *Inotify) handleDeletion(w *Watch) { + i.mu.Lock() + _, found := i.watches[w.wd] + delete(i.watches, w.wd) + i.mu.Unlock() + + if found { + i.queueEvent(newEvent(w.wd, "", linux.IN_IGNORED, 0)) + } +} + +// AddWatch constructs a new inotify watch and adds it to the target. It +// returns the watch descriptor returned by inotify_add_watch(2). +func (i *Inotify) AddWatch(target *Dentry, mask uint32) int32 { + // Note: Locking this inotify instance protects the result returned by + // Lookup() below. With the lock held, we know for sure the lookup result + // won't become stale because it's impossible for *this* instance to + // add/remove watches on target. + i.mu.Lock() + defer i.mu.Unlock() + + // Does the target already have a watch from this inotify instance? + if existing := target.Watches().Lookup(i.id); existing != nil { + newmask := mask + if mask&linux.IN_MASK_ADD != 0 { + // "Add (OR) events to watch mask for this pathname if it already + // exists (instead of replacing mask)." -- inotify(7) + newmask |= atomic.LoadUint32(&existing.mask) + } + atomic.StoreUint32(&existing.mask, newmask) + return existing.wd + } + + // No existing watch, create a new watch. + w := i.newWatchLocked(target, mask) + return w.wd +} + +// RmWatch looks up an inotify watch for the given 'wd' and configures the +// target to stop sending events to this inotify instance. +func (i *Inotify) RmWatch(wd int32) error { + i.mu.Lock() + + // Find the watch we were asked to removed. + w, ok := i.watches[wd] + if !ok { + i.mu.Unlock() + return syserror.EINVAL + } + + // Remove the watch from this instance. + delete(i.watches, wd) + + // Remove the watch from the watch target. + w.set.Remove(w.OwnerID()) + i.mu.Unlock() + + // Generate the event for the removal. + i.queueEvent(newEvent(wd, "", linux.IN_IGNORED, 0)) + + return nil +} + +// Watches is the collection of all inotify watches on a single file. +// +// +stateify savable +type Watches struct { + // mu protects the fields below. + mu sync.RWMutex `state:"nosave"` + + // ws is the map of active watches in this collection, keyed by the inotify + // instance id of the owner. + ws map[uint64]*Watch +} + +// Lookup returns the watch owned by an inotify instance with the given id. +// Returns nil if no such watch exists. +// +// Precondition: the inotify instance with the given id must be locked to +// prevent the returned watch from being concurrently modified or replaced in +// Inotify.watches. +func (w *Watches) Lookup(id uint64) *Watch { + w.mu.Lock() + defer w.mu.Unlock() + return w.ws[id] +} + +// Add adds watch into this set of watches. +// +// Precondition: the inotify instance with the given id must be locked. +func (w *Watches) Add(watch *Watch) { + w.mu.Lock() + defer w.mu.Unlock() + + owner := watch.OwnerID() + // Sanity check, we should never have two watches for one owner on the + // same target. + if _, exists := w.ws[owner]; exists { + panic(fmt.Sprintf("Watch collision with ID %+v", owner)) + } + if w.ws == nil { + w.ws = make(map[uint64]*Watch) + } + w.ws[owner] = watch +} + +// Remove removes a watch with the given id from this set of watches and +// releases it. The caller is responsible for generating any watch removal +// event, as appropriate. The provided id must match an existing watch in this +// collection. +// +// Precondition: the inotify instance with the given id must be locked. +func (w *Watches) Remove(id uint64) { + w.mu.Lock() + defer w.mu.Unlock() + + if w.ws == nil { + // This watch set is being destroyed. The thread executing the + // destructor is already in the process of deleting all our watches. We + // got here with no references on the target because we raced with the + // destructor notifying all the watch owners of destruction. See the + // comment in Watches.HandleDeletion for why this race exists. + return + } + + if _, ok := w.ws[id]; !ok { + // While there's technically no problem with silently ignoring a missing + // watch, this is almost certainly a bug. + panic(fmt.Sprintf("Attempt to remove a watch, but no watch found with provided id %+v.", id)) + } + delete(w.ws, id) +} + +// Notify queues a new event with all watches in this set. +func (w *Watches) Notify(name string, events, cookie uint32, et EventType) { + w.NotifyWithExclusions(name, events, cookie, et, false) +} + +// NotifyWithExclusions queues a new event with watches in this set. Watches +// with IN_EXCL_UNLINK are skipped if the event is coming from a child that +// has been unlinked. +func (w *Watches) NotifyWithExclusions(name string, events, cookie uint32, et EventType, unlinked bool) { + // N.B. We don't defer the unlocks because Notify is in the hot path of + // all IO operations, and the defer costs too much for small IO + // operations. + w.mu.RLock() + for _, watch := range w.ws { + if unlinked && watch.ExcludeUnlinkedChildren() && et == PathEvent { + continue + } + watch.Notify(name, events, cookie) + } + w.mu.RUnlock() +} + +// HandleDeletion is called when the watch target is destroyed to emit +// the appropriate events. +func (w *Watches) HandleDeletion() { + w.Notify("", linux.IN_DELETE_SELF, 0, InodeEvent) + + // TODO(gvisor.dev/issue/1479): This doesn't work because maps are not copied + // by value. Ideally, we wouldn't have this circular locking so we can just + // notify of IN_DELETE_SELF in the same loop below. + // + // We can't hold w.mu while calling watch.handleDeletion to preserve lock + // ordering w.r.t to the owner inotify instances. Instead, atomically move + // the watches map into a local variable so we can iterate over it safely. + // + // Because of this however, it is possible for the watches' owners to reach + // this inode while the inode has no refs. This is still safe because the + // owners can only reach the inode until this function finishes calling + // watch.handleDeletion below and the inode is guaranteed to exist in the + // meantime. But we still have to be very careful not to rely on inode state + // that may have been already destroyed. + var ws map[uint64]*Watch + w.mu.Lock() + ws = w.ws + w.ws = nil + w.mu.Unlock() + + for _, watch := range ws { + // TODO(gvisor.dev/issue/1479): consider refactoring this. + watch.handleDeletion() + } +} + +// Watch represent a particular inotify watch created by inotify_add_watch. +// +// +stateify savable +type Watch struct { + // Inotify instance which owns this watch. + owner *Inotify + + // Descriptor for this watch. This is unique across an inotify instance. + wd int32 + + // set is the watch set containing this watch. It belongs to the target file + // of this watch. + set *Watches + + // Events being monitored via this watch. Must be accessed with atomic + // memory operations. + mask uint32 +} + +// OwnerID returns the id of the inotify instance that owns this watch. +func (w *Watch) OwnerID() uint64 { + return w.owner.id +} + +// ExcludeUnlinkedChildren indicates whether the watched object should continue +// to be notified of events of its children after they have been unlinked, e.g. +// for an open file descriptor. +// +// TODO(gvisor.dev/issue/1479): Implement IN_EXCL_UNLINK. +// We can do this by keeping track of the set of unlinked children in Watches +// to skip notification. +func (w *Watch) ExcludeUnlinkedChildren() bool { + return atomic.LoadUint32(&w.mask)&linux.IN_EXCL_UNLINK != 0 +} + +// Notify queues a new event on this watch. +func (w *Watch) Notify(name string, events uint32, cookie uint32) { + mask := atomic.LoadUint32(&w.mask) + if mask&events == 0 { + // We weren't watching for this event. + return + } + + // Event mask should include bits matched from the watch plus all control + // event bits. + unmaskableBits := ^uint32(0) &^ linux.IN_ALL_EVENTS + effectiveMask := unmaskableBits | mask + matchedEvents := effectiveMask & events + w.owner.queueEvent(newEvent(w.wd, name, matchedEvents, cookie)) +} + +// handleDeletion handles the deletion of w's target. +func (w *Watch) handleDeletion() { + w.owner.handleDeletion(w) +} + +// Event represents a struct inotify_event from linux. +// +// +stateify savable +type Event struct { + eventEntry + + wd int32 + mask uint32 + cookie uint32 + + // len is computed based on the name field is set automatically by + // Event.setName. It should be 0 when no name is set; otherwise it is the + // length of the name slice. + len uint32 + + // The name field has special padding requirements and should only be set by + // calling Event.setName. + name []byte +} + +func newEvent(wd int32, name string, events, cookie uint32) *Event { + e := &Event{ + wd: wd, + mask: events, + cookie: cookie, + } + if name != "" { + e.setName(name) + } + return e +} + +// paddedBytes converts a go string to a null-terminated c-string, padded with +// null bytes to a total size of 'l'. 'l' must be large enough for all the bytes +// in the 's' plus at least one null byte. +func paddedBytes(s string, l uint32) []byte { + if l < uint32(len(s)+1) { + panic("Converting string to byte array results in truncation, this can lead to buffer-overflow due to the missing null-byte!") + } + b := make([]byte, l) + copy(b, s) + + // b was zero-value initialized during make(), so the rest of the slice is + // already filled with null bytes. + + return b +} + +// setName sets the optional name for this event. +func (e *Event) setName(name string) { + // We need to pad the name such that the entire event length ends up a + // multiple of inotifyEventBaseSize. + unpaddedLen := len(name) + 1 + // Round up to nearest multiple of inotifyEventBaseSize. + e.len = uint32((unpaddedLen + inotifyEventBaseSize - 1) & ^(inotifyEventBaseSize - 1)) + // Make sure we haven't overflowed and wrapped around when rounding. + if unpaddedLen > int(e.len) { + panic("Overflow when rounding inotify event size, the 'name' field was too big.") + } + e.name = paddedBytes(name, e.len) +} + +func (e *Event) sizeOf() int { + s := inotifyEventBaseSize + int(e.len) + if s < inotifyEventBaseSize { + panic("overflow") + } + return s +} + +// CopyTo serializes this event to dst. buf is used as a scratch buffer to +// construct the output. We use a buffer allocated ahead of time for +// performance. buf must be at least inotifyEventBaseSize bytes. +func (e *Event) CopyTo(ctx context.Context, buf []byte, dst usermem.IOSequence) (int64, error) { + usermem.ByteOrder.PutUint32(buf[0:], uint32(e.wd)) + usermem.ByteOrder.PutUint32(buf[4:], e.mask) + usermem.ByteOrder.PutUint32(buf[8:], e.cookie) + usermem.ByteOrder.PutUint32(buf[12:], e.len) + + writeLen := 0 + + n, err := dst.CopyOut(ctx, buf) + if err != nil { + return 0, err + } + writeLen += n + dst = dst.DropFirst(n) + + if e.len > 0 { + n, err = dst.CopyOut(ctx, e.name) + if err != nil { + return 0, err + } + writeLen += n + } + + // Santiy check. + if writeLen != e.sizeOf() { + panic(fmt.Sprintf("Serialized unexpected amount of data for an event, expected %d, wrote %d.", e.sizeOf(), writeLen)) + } + + return int64(writeLen), nil +} + +func (e *Event) equals(other *Event) bool { + return e.wd == other.wd && + e.mask == other.mask && + e.cookie == other.cookie && + e.len == other.len && + bytes.Equal(e.name, other.name) +} + +// InotifyEventFromStatMask generates the appropriate events for an operation +// that set the stats specified in mask. +func InotifyEventFromStatMask(mask uint32) uint32 { + var ev uint32 + if mask&(linux.STATX_UID|linux.STATX_GID|linux.STATX_MODE) != 0 { + ev |= linux.IN_ATTRIB + } + if mask&linux.STATX_SIZE != 0 { + ev |= linux.IN_MODIFY + } + + if (mask & (linux.STATX_ATIME | linux.STATX_MTIME)) == (linux.STATX_ATIME | linux.STATX_MTIME) { + // Both times indicates a utime(s) call. + ev |= linux.IN_ATTRIB + } else if mask&linux.STATX_ATIME != 0 { + ev |= linux.IN_ACCESS + } else if mask&linux.STATX_MTIME != 0 { + mask |= linux.IN_MODIFY + } + return ev +} + +// InotifyRemoveChild sends the appriopriate notifications to the watch sets of +// the child being removed and its parent. +func InotifyRemoveChild(self, parent *Watches, name string) { + self.Notify("", linux.IN_ATTRIB, 0, InodeEvent) + parent.Notify(name, linux.IN_DELETE, 0, InodeEvent) + // TODO(gvisor.dev/issue/1479): implement IN_EXCL_UNLINK. +} + +// InotifyRename sends the appriopriate notifications to the watch sets of the +// file being renamed and its old/new parents. +func InotifyRename(ctx context.Context, renamed, oldParent, newParent *Watches, oldName, newName string, isDir bool) { + var dirEv uint32 + if isDir { + dirEv = linux.IN_ISDIR + } + cookie := uniqueid.InotifyCookie(ctx) + oldParent.Notify(oldName, dirEv|linux.IN_MOVED_FROM, cookie, InodeEvent) + newParent.Notify(newName, dirEv|linux.IN_MOVED_TO, cookie, InodeEvent) + // Somewhat surprisingly, self move events do not have a cookie. + renamed.Notify("", linux.IN_MOVE_SELF, 0, InodeEvent) +} diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go index 02850b65c..e4ac6524b 100644 --- a/pkg/sentry/vfs/mount.go +++ b/pkg/sentry/vfs/mount.go @@ -28,9 +28,6 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) -// lastMountID is used to allocate mount ids. Must be accessed atomically. -var lastMountID uint64 - // A Mount is a replacement of a Dentry (Mount.key.point) from one Filesystem // (Mount.key.parent.fs) with a Dentry (Mount.root) from another Filesystem // (Mount.fs), which applies to path resolution in the context of a particular @@ -97,7 +94,7 @@ type Mount struct { func newMount(vfs *VirtualFilesystem, fs *Filesystem, root *Dentry, mntns *MountNamespace, opts *MountOptions) *Mount { mnt := &Mount{ - ID: atomic.AddUint64(&lastMountID, 1), + ID: atomic.AddUint64(&vfs.lastMountID, 1), vfs: vfs, fs: fs, root: root, @@ -111,6 +108,16 @@ func newMount(vfs *VirtualFilesystem, fs *Filesystem, root *Dentry, mntns *Mount return mnt } +// Options returns a copy of the MountOptions currently applicable to mnt. +func (mnt *Mount) Options() MountOptions { + mnt.vfs.mountMu.Lock() + defer mnt.vfs.mountMu.Unlock() + return MountOptions{ + Flags: mnt.flags, + ReadOnly: mnt.readOnly(), + } +} + // A MountNamespace is a collection of Mounts. // // MountNamespaces are reference-counted. Unless otherwise specified, all @@ -148,7 +155,7 @@ type MountNamespace struct { func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth.Credentials, source, fsTypeName string, opts *GetFilesystemOptions) (*MountNamespace, error) { rft := vfs.getFilesystemType(fsTypeName) if rft == nil { - ctx.Warningf("Unknown filesystem: %s", fsTypeName) + ctx.Warningf("Unknown filesystem type: %s", fsTypeName) return nil, syserror.ENODEV } fs, root, err := rft.fsType.GetFilesystem(ctx, vfs, creds, source, *opts) @@ -175,26 +182,34 @@ func (vfs *VirtualFilesystem) NewDisconnectedMount(fs *Filesystem, root *Dentry, return newMount(vfs, fs, root, nil /* mntns */, opts), 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 { +// MountDisconnected creates a Filesystem configured by the given arguments, +// then returns a Mount representing it. The new Mount is not associated with +// any MountNamespace and is not connected to any other Mounts. +func (vfs *VirtualFilesystem) MountDisconnected(ctx context.Context, creds *auth.Credentials, source string, fsTypeName string, opts *MountOptions) (*Mount, error) { rft := vfs.getFilesystemType(fsTypeName) if rft == nil { - return syserror.ENODEV + return nil, syserror.ENODEV } if !opts.InternalMount && !rft.opts.AllowUserMount { - return syserror.ENODEV + return nil, syserror.ENODEV } fs, root, err := rft.fsType.GetFilesystem(ctx, vfs, creds, source, opts.GetFilesystemOptions) if err != nil { - return err + return nil, err } + defer root.DecRef() + defer fs.DecRef() + return vfs.NewDisconnectedMount(fs, root, opts) +} +// ConnectMountAt connects mnt at the path represented by target. +// +// Preconditions: mnt must be disconnected. +func (vfs *VirtualFilesystem) ConnectMountAt(ctx context.Context, creds *auth.Credentials, mnt *Mount, target *PathOperation) error { // We can't hold vfs.mountMu while calling FilesystemImpl methods due to // lock ordering. vd, err := vfs.GetDentryAt(ctx, creds, target, &GetDentryOptions{}) if err != nil { - root.DecRef() - fs.DecRef() return err } vfs.mountMu.Lock() @@ -204,8 +219,6 @@ func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentia vd.dentry.mu.Unlock() vfs.mountMu.Unlock() vd.DecRef() - root.DecRef() - fs.DecRef() return syserror.ENOENT } // vd might have been mounted over between vfs.GetDentryAt() and @@ -238,7 +251,6 @@ func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentia // point and the mount root are directories, or neither are, and returns // ENOTDIR if this is not the case. mntns := vd.mount.ns - mnt := newMount(vfs, fs, root, mntns, opts) vfs.mounts.seq.BeginWrite() vfs.connectLocked(mnt, vd, mntns) vfs.mounts.seq.EndWrite() @@ -247,6 +259,19 @@ func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentia return 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 { + mnt, err := vfs.MountDisconnected(ctx, creds, source, fsTypeName, opts) + if err != nil { + return err + } + if err := vfs.ConnectMountAt(ctx, creds, mnt, target); err != nil { + mnt.DecRef() + return err + } + return nil +} + // UmountAt removes the Mount at the given path. func (vfs *VirtualFilesystem) UmountAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *UmountOptions) error { if opts.Flags&^(linux.MNT_FORCE|linux.MNT_DETACH) != 0 { @@ -377,6 +402,7 @@ func (vfs *VirtualFilesystem) connectLocked(mnt *Mount, vd VirtualDentry, mntns } vd.mount.children[mnt] = struct{}{} atomic.AddUint32(&vd.dentry.mounts, 1) + mnt.ns = mntns mntns.mountpoints[vd.dentry]++ vfs.mounts.insertSeqed(mnt) vfsmpmounts, ok := vfs.mountpoints[vd.dentry] diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go index 8d7f8f8af..52643a7c5 100644 --- a/pkg/sentry/vfs/vfs.go +++ b/pkg/sentry/vfs/vfs.go @@ -82,6 +82,10 @@ type VirtualFilesystem struct { // mountpoints is analogous to Linux's mountpoint_hashtable. mountpoints map[*Dentry]map[*Mount]struct{} + // lastMountID is the last allocated mount ID. lastMountID is accessed + // using atomic memory operations. + lastMountID uint64 + // anonMount is a Mount, not included in mounts or mountpoints, // representing an anonFilesystem. anonMount is used to back // VirtualDentries returned by VirtualFilesystem.NewAnonVirtualDentry(). @@ -418,6 +422,7 @@ func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credential } } + fd.Dentry().InotifyWithParent(linux.IN_OPEN, 0, PathEvent) return fd, nil } if !rp.handleError(err) { |