Enable unprivileged user namespace mounts of overlayfs. Overlayfs's
permission model (*) ensures that the mounter itself cannot gain additional
privileges by the act of creating an overlayfs mount.
This feature request is coming from the "rootless" container crowd.
(*) Documentation/filesystems/overlayfs.txt#Permission model
Signed-off-by: Miklos Szeredi <[email protected]>
When looking up an inode on the lower layer for which the mounter lacks
read permisison the metacopy check will fail. This causes the lookup to
fail as well, even though the directory is readable.
So ignore EACCES for the "userxattr" case and assume no metacopy for the
unreadable file.
Signed-off-by: Miklos Szeredi <[email protected]>
In case the file cannot be opened with O_NOATIME because of lack of
capabilities, then clear O_NOATIME instead of failing.
Remove WARN_ON(), since it would now trigger if O_NOATIME was cleared.
Noticed by Amir Goldstein.
Signed-off-by: Miklos Szeredi <[email protected]>
Comment above call already says this, but only EOPNOTSUPP is ignored, other
failures are not.
For example setting "user.*" will fail with EPERM on symlink/special.
Ignore this error as well.
Signed-off-by: Miklos Szeredi <[email protected]>
Optionally allow using "user.overlay." namespace instead of
"trusted.overlay."
This is necessary for overlayfs to be able to be mounted in an unprivileged
namepsace.
Make the option explicit, since it makes the filesystem format be
incompatible.
Disable redirect_dir and metacopy options, because these would allow
privilege escalation through direct manipulation of the
"user.overlay.redirect" or "user.overlay.metacopy" xattrs.
Signed-off-by: Miklos Szeredi <[email protected]>
Reviewed-by: Amir Goldstein <[email protected]>
generic_file_splice_read() and iter_file_splice_write() will call back into
f_op->iter_read() and f_op->iter_write() respectively. These already do
the real file lookup and cred override. So the code in ovl_splice_read()
and ovl_splice_write() is redundant.
In addition the ovl_file_accessed() call in ovl_splice_write() is
incorrect, though probably harmless.
Fix by calling generic_file_splice_read() and iter_file_splice_write()
directly.
Signed-off-by: Miklos Szeredi <[email protected]>
ovl_ioctl_set_flags() does a capability check using flags, but then the
real ioctl double-fetches flags and uses potentially different value.
The "Check the capability before cred override" comment misleading: user
can skip this check by presenting benign flags first and then overwriting
them to non-benign flags.
Just remove the cred override for now, hoping this doesn't cause a
regression.
The proper solution is to create a new setxflags i_op (patches are in the
works).
Xfstests don't show a regression.
Reported-by: Dmitry Vyukov <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
Reviewed-by: Amir Goldstein <[email protected]>
Fixes: dab5ca8fd9 ("ovl: add lsattr/chattr support")
Cc: <[email protected]> # v4.19
CAP_DAC_READ_SEARCH is required by open_by_handle_at(2) so check it in
ovl_decode_real_fh() as well to prevent privilege escalation for
unprivileged overlay mounts.
[Amir] If the mounter is not capable in init ns, ovl_check_origin() and
ovl_verify_index() will not function as expected and this will break index
and nfs export features. So check capability in ovl_can_decode_fh(), to
auto disable those features.
Signed-off-by: Miklos Szeredi <[email protected]>
Call remap_verify_area() on the source file as well as the destination.
When called from vfs_dedupe_file_range() the check as already been
performed, but not so if called from layered fs (overlayfs, etc...)
Could ommit the redundant check in vfs_dedupe_file_range(), but leave for
now to get error early (for fear of breaking backward compatibility).
This call shouldn't be performance sensitive.
Signed-off-by: Miklos Szeredi <[email protected]>
cap_convert_nscap() does permission checking as well as conversion of the
xattr value conditionally based on fs's user-ns.
This is needed by overlayfs and probably other layered fs (ecryptfs) and is
what vfs_foo() is supposed to do anyway.
Signed-off-by: Miklos Szeredi <[email protected]>
Acked-by: James Morris <[email protected]>
Jan Kara's analysis of the syzbot report (edited):
The reproducer opens a directory on FUSE filesystem, it then attaches
dnotify mark to the open directory. After that a fuse_do_getattr() call
finds that attributes returned by the server are inconsistent, and calls
make_bad_inode() which, among other things does:
inode->i_mode = S_IFREG;
This then confuses dnotify which doesn't tear down its structures
properly and eventually crashes.
Avoid calling make_bad_inode() on a live inode: switch to a private flag on
the fuse inode. Also add the test to ops which the bad_inode_ops would
have caught.
This bug goes back to the initial merge of fuse in 2.6.14...
Reported-by: [email protected]
Signed-off-by: Miklos Szeredi <[email protected]>
Tested-by: Jan Kara <[email protected]>
Cc: <[email protected]>
Documentation says "The lower filesystem can be any filesystem supported by
Linux". However, this is not the case, as Linux supports vfat and vfat
doesn't work as a lower filesystem
Reported-by: nerdopolis <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
Kernel has:
ATTR_KILL_PRIV -> clear "security.capability"
ATTR_KILL_SUID -> clear S_ISUID
ATTR_KILL_SGID -> clear S_ISGID if executable
Fuse has:
FUSE_WRITE_KILL_PRIV -> clear S_ISUID and S_ISGID if executable
So FUSE_WRITE_KILL_PRIV implies the complement of ATTR_KILL_PRIV, which is
somewhat confusing. Also PRIV implies all privileges, including
"security.capability".
Change the name to FUSE_WRITE_KILL_SUIDGID and make FUSE_WRITE_KILL_PRIV an
alias to perserve API compatibility
Signed-off-by: Miklos Szeredi <[email protected]>
Failure to do so may result in EEXIST even if the file only exists in the
cache and not in the filesystem.
The atomic nature of O_EXCL mandates that the cached state should be
ignored and existence verified anew.
Reported-by: Ken Schalk <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
Fuse mount now only ever has a refcount of one (before being freed) so the
count field is unnecessary.
Remove the refcounting and fold fuse_mount_put() into callers. The only
caller of fuse_mount_put() where fm->fc was NULL is fuse_dentry_automount()
and here the fuse_conn_put() can simply be omitted.
Signed-off-by: Miklos Szeredi <[email protected]>
Currently when acquiring an sb for virtiofs fuse_mount_get() is being
called from virtio_fs_set_super() if a new sb is being filled and
fuse_mount_put() is called unconditionally after sget_fc() returns.
The exact same result can be obtained by checking whether
fs_contex->s_fs_info was set to NULL (ref trasferred to sb->s_fs_info) and
only calling fuse_mount_put() if the ref wasn't transferred (error or
matching sb found).
This allows getting rid of virtio_fs_set_super() and fuse_mount_get().
Signed-off-by: Miklos Szeredi <[email protected]>
This can be triggered for example by adding the "-omand" mount option,
which will be rejected and virtio_fs_fill_super() will return an error.
In such a case the allocations for fuse_conn and fuse_mount will leak due
to s_root not yet being set and so ->put_super() not being called.
Fixes: a62a8ef9d9 ("virtio-fs: add virtiofs filesystem")
Signed-off-by: Miklos Szeredi <[email protected]>
Qian Cai reports that the WARNING in tree_insert() can be triggered by a
fuzzer with the following call chain:
invalidate_inode_pages2_range()
fuse_launder_page()
fuse_writepage_locked()
tree_insert()
The reason is that another write for the same page is already queued.
The simplest fix is to wait until the pending write is completed and only
after that queue the new write.
Since this case is very rare, the additional wait should not be a problem.
Reported-by: Qian Cai <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
Re-add lost removal of fc from fuse_conn_list and the control filesystem.
Reported-by: kernel test robot <[email protected]>
Fixes: fcee216beb ("fuse: split fuse_mount off of fuse_conn")
Signed-off-by: Miklos Szeredi <[email protected]>
After unlock_request() pages from the ap->pages[] array may be put (e.g. by
aborting the connection) and the pages can be freed.
Prevent use after free by grabbing a reference to the page before calling
unlock_request().
The original patch was created by Pradeep P V K.
Reported-by: Pradeep P V K <[email protected]>
Cc: <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
ovl_can_list() should return false for overlay private xattrs. Since
currently these use the "trusted.overlay." prefix, they will always match
the "trusted." prefix as well, hence the test for being non-trusted will
not trigger.
Prepare for using the "user.overlay." namespace by moving the test for
private xattr before the test for non-trusted.
This patch doesn't change behavior.
Signed-off-by: Miklos Szeredi <[email protected]>
Instead of passing the xattr name down to the ovl_do_*xattr() accessor
functions, pass an enumerated value. The enum can use the same names as
the the previous #define for each xattr name.
Signed-off-by: Miklos Szeredi <[email protected]>
Call ovl_do_*xattr() when accessing an overlay private xattr, vfs_*xattr()
otherwise.
This has an effect on debug output, which is made more consistent by this
patch.
Signed-off-by: Miklos Szeredi <[email protected]>
Use the convention of calling ovl_do_foo() for operations which are overlay
specific.
This patch is a no-op, and will have significance for supporting
"user.overlay." xattr namespace.
Signed-off-by: Miklos Szeredi <[email protected]>
This is a partial revert (with some cleanups) of commit 993a0b2aec ("ovl:
Do not lose security.capability xattr over metadata file copy-up"), which
introduced ovl_getxattr() in the first place.
Signed-off-by: Miklos Szeredi <[email protected]>
Lose the padding and the failure message (in line with other parts of the
copy up process). Return zero for both nonexistent or empty xattr.
Signed-off-by: Miklos Szeredi <[email protected]>
ovl_getattr() returns the value of an xattr in a kmalloced buffer. There
are two callers:
ovl_copy_up_meta_inode_data() (copy_up.c)
ovl_get_redirect_xattr() (util.c)
This patch just copies ovl_getxattr() to copy_up.c, the following patches
will deal with the differences in idividual callers.
Signed-off-by: Miklos Szeredi <[email protected]>
fuse_writepages_fill uses following construction:
if (wpa && ap->num_pages &&
(A || B || C)) {
action;
} else if (wpa && D) {
if (E) {
the same action;
}
}
- ap->num_pages check is always true and can be removed
- "if" and "else if" calls the same action and can be merged.
Move checking A, B, C, D, E conditions to a helper, add comments.
Original-patch-by: Vasily Averin <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
Previous patch changed handling of remount/reconfigure to ignore all
options, including those that are unknown to the fuse kernel fs. This was
done for backward compatibility, but this likely only affects the old
mount(2) API.
The new fsconfig(2) based reconfiguration could possibly be improved. This
would make the new API less of a drop in replacement for the old, OTOH this
is a good chance to get rid of some weirdnesses in the old API.
Several other behaviors might make sense:
1) unknown options are rejected, known options are ignored
2) unknown options are rejected, known options are rejected if the value
is changed, allowed otherwise
3) all options are rejected
Prior to the backward compatibility fix to ignore all options all known
options were accepted (1), even if they change the value of a mount
parameter; fuse_reconfigure() does not look at the config values set by
fuse_parse_param().
To fix that we'd need to verify that the value provided is the same as set
in the initial configuration (2). The major drawback is that this is much
more complex than just rejecting all attempts at changing options (3);
i.e. all options signify initial configuration values and don't make sense
on reconfigure.
This patch opts for (3) with the rationale that no mount options are
reconfigurable in fuse.
Signed-off-by: Miklos Szeredi <[email protected]>
The command
mount -o remount -o unknownoption /mnt/fuse
succeeds on kernel versions prior to v5.4 and fails on kernel version at or
after. This is because fuse_parse_param() rejects any unrecognised options
in case of FS_CONTEXT_FOR_RECONFIGURE, just as for FS_CONTEXT_FOR_MOUNT.
This causes a regression in case the fuse filesystem is in fstab, since
remount sends all options found there to the kernel; even ones that are
meant for the initial mount and are consumed by the userspace fuse server.
Fix this by ignoring mount options, just as fuse_remount_fs() did prior to
the conversion to the new API.
Reported-by: Stefan Priebe <[email protected]>
Fixes: c30da2e981 ("fuse: convert to use the new mount API")
Cc: <[email protected]> # v5.4
Signed-off-by: Miklos Szeredi <[email protected]>
s_op->remount_fs() is only called from legacy_reconfigure(), which is not
used after being converted to the new API.
Convert to using ->reconfigure(). This restores the previous behavior of
syncing the filesystem and rejecting MS_MANDLOCK on remount.
Fixes: c30da2e981 ("fuse: convert to use the new mount API")
Cc: <[email protected]> # v5.4
Signed-off-by: Miklos Szeredi <[email protected]>
fuse_writepages_fill() calls tree_insert() with ap->num_pages = 0 which
triggers the following warning:
WARNING: CPU: 1 PID: 17211 at fs/fuse/file.c:1728 tree_insert+0xab/0xc0 [fuse]
RIP: 0010:tree_insert+0xab/0xc0 [fuse]
Call Trace:
fuse_writepages_fill+0x5da/0x6a0 [fuse]
write_cache_pages+0x171/0x470
fuse_writepages+0x8a/0x100 [fuse]
do_writepages+0x43/0xe0
Fix up the warning and clean up the code around rb-tree insertion:
- Rename tree_insert() to fuse_insert_writeback() and make it return the
conflicting entry in case of failure
- Re-add tree_insert() as a wrapper around fuse_insert_writeback()
- Rename fuse_writepage_in_flight() to fuse_writepage_add() and reverse
the meaning of the return value to mean
+ "true" in case the writepage entry was successfully added
+ "false" in case it was in-fligt queued on an existing writepage
entry's auxiliary list or the existing writepage entry's temporary
page updated
Switch from fuse_find_writeback() + tree_insert() to
fuse_insert_writeback()
- Move setting orig_pages to before inserting/updating the entry; this may
result in the orig_pages value being discarded later in case of an
in-flight request
- In case of a new writepage entry use fuse_writepage_add()
unconditionally, only set data->wpa if the entry was added.
Fixes: 6b2fb79963 ("fuse: optimize writepages search")
Reported-by: kernel test robot <[email protected]>
Original-path-by: Vasily Averin <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
In fuse_writepage_end() the old writepages entry needs to be removed from
the rbtree before inserting the new one, otherwise tree_insert() would
fail. This is a very rare codepath and no reproducer exists.
Signed-off-by: Miklos Szeredi <[email protected]>
Overlayfs is using clone_private_mount() to create internal mounts for
underlying layers. These are used for operations requiring a path, such as
dentry_open().
Since these private mounts are not in any namespace they are treated as
short term, "detached" mounts and mntput() involves taking the global
mount_lock, which can result in serious cacheline pingpong.
Make these private mounts longterm instead, which trade the penalty on
mntput() for a slightly longer shutdown time due to an added RCU grace
period when putting these mounts.
Introduce a new helper kern_unmount_many() that can take care of multiple
longterm mounts with a single RCU grace period.
Cc: Al Viro <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
ofs->upper_mnt is copied to ->layers[0].mnt and ->layers[0].trap could be
used instead of a separate ->upperdir_trap.
Split the lowerdir option early to get the number of layers, then allocate
the ->layers array, and finally fill the upper and lower layers, as before.
Get rid of path_put_init() in ovl_lower_dir(), since the only caller will
take care of that.
[Colin Ian King] Fix null pointer dereference on null stack pointer on
error return found by Coverity.
Signed-off-by: Miklos Szeredi <[email protected]>
Call inode_permission() on real inode before opening regular file on one of
the underlying layers.
In some cases ovl_permission() already checks access to an underlying file,
but it misses the metacopy case, and possibly other ones as well.
Removing the redundant permission check from ovl_permission() should be
considered later.
Signed-off-by: Miklos Szeredi <[email protected]>
Verify LSM permissions for underlying file, since vfs_ioctl() doesn't do
it.
[Stephen Rothwell] export security_file_ioctl
Signed-off-by: Miklos Szeredi <[email protected]>
Check permission before opening a real file.
ovl_path_open() is used by readdir and copy-up routines.
ovl_permission() theoretically already checked copy up permissions, but it
doesn't hurt to re-do these checks during the actual copy-up.
For directory reading ovl_permission() only checks access to topmost
underlying layer. Readdir on a merged directory accesses layers below the
topmost one as well. Permission wasn't checked for these layers.
Note: modifying ovl_permission() to perform this check would be far more
complex and hence more bug prone. The result is less precise permissions
returned in access(2). If this turns out to be an issue, we can revisit
this bug.
Signed-off-by: Miklos Szeredi <[email protected]>
In preparation for more permission checking, override credentials for
directory operations on the underlying filesystems.
Signed-off-by: Miklos Szeredi <[email protected]>
The three instances of ovl_path_open() in overlayfs/readdir.c do three
different things:
- pass f_flags from overlay file
- pass O_RDONLY | O_DIRECTORY
- pass just O_RDONLY
The value of f_flags can be (other than O_RDONLY):
O_WRONLY - not possible for a directory
O_RDWR - not possible for a directory
O_CREAT - masked out by dentry_open()
O_EXCL - masked out by dentry_open()
O_NOCTTY - masked out by dentry_open()
O_TRUNC - masked out by dentry_open()
O_APPEND - no effect on directory ops
O_NDELAY - no effect on directory ops
O_NONBLOCK - no effect on directory ops
__O_SYNC - no effect on directory ops
O_DSYNC - no effect on directory ops
FASYNC - no effect on directory ops
O_DIRECT - no effect on directory ops
O_LARGEFILE - ?
O_DIRECTORY - only affects lookup
O_NOFOLLOW - only affects lookup
O_NOATIME - overlay sets this unconditionally in ovl_path_open()
O_CLOEXEC - only affects fd allocation
O_PATH - no effect on directory ops
__O_TMPFILE - not possible for a directory
Fon non-merge directories we use the underlying filesystem's iterate; in
this case honor O_LARGEFILE from the original file to make sure that open
doesn't get rejected.
For merge directories it's safe to pass O_LARGEFILE unconditionally since
userspace will only see the artificial offsets created by overlayfs.
Signed-off-by: Miklos Szeredi <[email protected]>
After the copy operation completes the cache is not up-to-date. Truncate
all pages in the interval that has successfully been copied.
Truncating completely copied dirty pages is okay, since the data has been
overwritten anyway. Truncating partially copied dirty pages is not okay;
add a comment for now.
Fixes: 88bc7d5097 ("fuse: add support for copy_file_range()")
Signed-off-by: Miklos Szeredi <[email protected]>
a) Dirty cache needs to be written back not just in the writeback_cache
case, since the dirty pages may come from memory maps.
b) The fuse_writeback_range() helper takes an inclusive interval, so the
end position needs to be pos+len-1 instead of pos+len.
Fixes: 88bc7d5097 ("fuse: add support for copy_file_range()")
Signed-off-by: Miklos Szeredi <[email protected]>
A GETATTR request can race with FUSE_NOTIFY_INVAL_INODE, resulting in the
attribute cache being updated with stale information after the
invalidation.
Fix this by bumping the attribute version in fuse_reverse_inval_inode().
Reported-by: Krzysztof Rusek <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
page_count() is unstable. Unless there has been an RCU grace period
between when the page was removed from the page cache and now, a
speculative reference may exist from the page cache.
Reported-by: Matthew Wilcox <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
Instead of custom page dumping, use the standard helper.
Reported-by: Matthew Wilcox <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
Fuse mounts without "allow_other" are off-limits to all non-owners. Yet it
makes sense to allow querying st_dev on the root, since this value is
provided by the kernel, not the userspace filesystem.
Allow statx(2) with a zero request mask to succeed on a fuse mounts for all
users.
Reported-by: Nikolaus Rath <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
We want cached data to synced with the userspace filesystem on close(), for
example to allow getting correct st_blocks value. Do this regardless of
whether the userspace filesystem implements a FLUSH method or not.
Signed-off-by: Miklos Szeredi <[email protected]>
POSIX defines faccessat() as having a fourth "flags" argument, while the
linux syscall doesn't have it. Glibc tries to emulate AT_EACCESS and
AT_SYMLINK_NOFOLLOW, but AT_EACCESS emulation is broken.
Add a new faccessat(2) syscall with the added flags argument and implement
both flags.
The value of AT_EACCESS is defined in glibc headers to be the same as
AT_REMOVEDIR. Use this value for the kernel interface as well, together
with the explanatory comment.
Also add AT_EMPTY_PATH support, which is not documented by POSIX, but can
be useful and is trivial to implement.
Signed-off-by: Miklos Szeredi <[email protected]>
Parsing "silent" and clearing SB_SILENT makes zero sense.
Parsing "silent" and setting SB_SILENT would make a bit more sense, but
apparently nobody cares.
Signed-off-by: Miklos Szeredi <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Unlike the others, this is _not_ a standard option accepted by mount(8).
In fact SB_POSIXACL is an internal flag, and accepting MS_POSIXACL on the
mount(2) interface is possibly a bug.
The only filesystem that apparently wants to handle the "posixacl" option
is 9p, but it has special handling of that option besides setting
SB_POSIXACL.
Signed-off-by: Miklos Szeredi <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Makes little sense to keep this blacklist synced with what mount(8) parses
and what it doesn't. E.g. it has various forms of "*atime" options, but
not "atime"...
Signed-off-by: Miklos Szeredi <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Determining whether a path or file descriptor refers to a mountpoint (or
more precisely a mount root) is not trivial using current tools.
Add a flag to statx that indicates whether the path or fd refers to the
root of a mount or not.
Cc: [email protected]
Cc: [email protected]
Reported-by: Lennart Poettering <[email protected]>
Reported-by: J. Bruce Fields <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
IS_NOATIME(inode) is defined as __IS_FLG(inode, SB_RDONLY|SB_NOATIME), so
generic_fillattr() will clear STATX_ATIME from the result_mask if the super
block is marked read only.
This was probably not the intention, so fix to only clear STATX_ATIME if
the fs doesn't support atime at all.
Signed-off-by: Miklos Szeredi <[email protected]>
Acked-by: David Howells <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Constants of the *_ALL type can be actively harmful due to the fact that
developers will usually fail to consider the possible effects of future
changes to the definition.
Deprecate STATX_ALL in the uapi, while no damage has been done yet.
We could keep something like this around in the kernel, but there's
actually no point, since all filesystems should be explicitly checking
flags that they support and not rely on the VFS masking unknown ones out: a
flag could be known to the VFS, yet not known to the filesystem.
Cc: David Howells <[email protected]>
Cc: [email protected]
Cc: [email protected]
Signed-off-by: Miklos Szeredi <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
This makes it possible to use utimensat on an O_PATH file (including
symlinks).
It supersedes the nonstandard utimensat(fd, NULL, ...) form.
Cc: [email protected]
Cc: [email protected]
Signed-off-by: Miklos Szeredi <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Split out a helper that overrides the credentials in preparation for
actually doing the access check.
This prepares for the next patch that optionally disables the creds
override.
Suggested-by: Christoph Hellwig <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
If mounts are deleted after a read(2) call on /proc/self/mounts (or its
kin), the subsequent read(2) could miss a mount that comes after the
deleted one in the list. This is because the file position is interpreted
as the number mount entries from the start of the list.
E.g. first read gets entries #0 to #9; the seq file index will be 10. Then
entry #5 is deleted, resulting in #10 becoming #9 and #11 becoming #10,
etc... The next read will continue from entry #10, and #9 is missed.
Solve this by adding a cursor entry for each open instance. Taking the
global namespace_sem for write seems excessive, since we are only dealing
with a per-namespace list. Instead add a per-namespace spinlock and use
that together with namespace_sem taken for read to protect against
concurrent modification of the mount list. This may reduce parallelism of
is_local_mountpoint(), but it's hardly a big contention point. We could
also use RCU freeing of cursors to make traversal not need additional
locks, if that turns out to be neceesary.
Only move the cursor once for each read (cursor is not added on open) to
minimize cacheline invalidation. When EOF is reached, the cursor is taken
off the list, in order to prevent an excessive number of cursors due to
inactive open file descriptors.
Reported-by: Karel Zak <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
Avi Kivity reports that on fuse filesystems running in a user namespace
asyncronous fsync fails with EOVERFLOW.
The reason is that f_ops->fsync() is called with the creds of the kthread
performing aio work instead of the creds of the process originally
submitting IOCB_CMD_FSYNC.
Fuse sends the creds of the caller in the request header and it needs to
translate the uid and gid into the server's user namespace. Since the
kthread is running in init_user_ns, the translation will fail and the
operation returns an error.
It can be argued that fsync doesn't actually need any creds, but just
zeroing out those fields in the header (as with requests that currently
don't take creds) is a backward compatibility risk.
Instead of working around this issue in fuse, solve the core of the problem
by calling the filesystem with the proper creds.
Reported-by: Avi Kivity <[email protected]>
Tested-by: Giuseppe Scrivano <[email protected]>
Fixes: c9582eb0ff ("fuse: Fail all requests with invalid uids or gids")
Cc: [email protected] # 4.18+
Signed-off-by: Miklos Szeredi <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Whiteouts, unlike real device node should not require privileges to create.
The general concern with device nodes is that opening them can have side
effects. The kernel already avoids zero major (see
Documentation/admin-guide/devices.txt). To be on the safe side the patch
explicitly forbids registering a char device with 0/0 number (see
cdev_add()).
This guarantees that a non-O_PATH open on a whiteout will fail with ENODEV;
i.e. it won't have any side effect.
Signed-off-by: Miklos Szeredi <[email protected]>
Changes to underlying layers should not cause WARN_ON(), but this repro
does:
mkdir w l u mnt
sudo mount -t overlay -o workdir=w,lowerdir=l,upperdir=u overlay mnt
touch mnt/h
ln u/h u/k
rm -rf mnt/k
rm -rf mnt/h
dmesg
------------[ cut here ]------------
WARNING: CPU: 1 PID: 116244 at fs/inode.c:302 drop_nlink+0x28/0x40
After upper hardlinks were added while overlay is mounted, unlinking all
overlay hardlinks drops overlay nlink to zero before all upper inodes
are unlinked.
After unlink/rename prevent i_nlink from going to zero if there are still
hashed aliases (i.e. cached hard links to the victim) remaining.
Reported-by: Phasip <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
No reason to prevent upper layer being a remote filesystem. Do the
revalidation in that case, just as we already do for lower layers.
This lets virtiofs be used as upper layer, which appears to be a real use
case.
Signed-off-by: Miklos Szeredi <[email protected]>
Allow completely skipping ->revalidate() on a per-dentry basis, in case the
underlying layers used for a dentry do not themselves have ->revalidate().
E.g. negative overlay dentry has no underlying layers, hence revalidate is
unnecessary. Or if lower layer is remote but overlay dentry is pure-upper,
then can skip revalidate.
The following places need to update whether the dentry needs revalidate or
not:
- fill-super (root dentry)
- lookup
- create
- fh_to_dentry
Signed-off-by: Miklos Szeredi <[email protected]>
Following patch will allow remote as upper layer, but not overlay stacked
on upper layer. Separate the two concepts.
This patch is doesn't change behavior.
Signed-off-by: Miklos Szeredi <[email protected]>
Use a common loop for plain and weak revalidation. This will aid doing
revalidation on upper layer.
This patch doesn't change behavior.
Signed-off-by: Miklos Szeredi <[email protected]>
This issue came up with NFSv4 as the lower layer, which generates
"system.nfs4_acl" xattrs (even for plain old unix permissions). Prior to
this patch this prevented copy-up from succeeding.
The overlayfs permission model mandates that permissions are checked
locally for the task and remotely for the mounter(*). NFS4 ACLs are not
supported by the Linux kernel currently, hence they cannot be enforced
locally. Which means it is indifferent whether this attribute is copied or
not.
Generalize this to any xattr that is not used in access checking (i.e. it's
not a POSIX ACL and not in the "security." namespace).
Incidentally, best effort copying of xattrs seems to also be the behavior
of "cp -a", which is what overlayfs tries to mimic.
(*) Documentation/filesystems/overlayfs.txt#Permission model
Signed-off-by: Miklos Szeredi <[email protected]>
Lockdep reports "WARNING: lock held when returning to user space!" due to
async write holding freeze lock over the write. Apparently aio.c already
deals with this by lying to lockdep about the state of the lock.
Do the same here. No need to check for S_IFREG() here since these file ops
are regular-only.
Reported-by: [email protected]
Fixes: 2406a307ac ("ovl: implement async IO routines")
Signed-off-by: Miklos Szeredi <[email protected]>
Normal, synchronous requests will have their args allocated on the stack.
After the FR_FINISHED bit is set by receiving the reply from the userspace
fuse server, the originating task may return and reuse the stack frame,
resulting in an Oops if the args structure is dereferenced.
Fix by setting a flag in the request itself upon initializing, indicating
whether it has an asynchronous ->end() callback.
Reported-by: Kyle Sanderson <[email protected]>
Reported-by: Michael Stapelberg <[email protected]>
Fixes: 2b319d1f6f ("fuse: don't dereference req->args on finished request")
Cc: <[email protected]> # v5.4
Tested-by: Michael Stapelberg <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
Handle the special case of fuse_readpages() wanting to read the last page
of a hugest file possible and overflowing the end offset in the process.
This is basically to unbreak xfstests:generic/525 and prevent filesystems
from doing bad things with an overflowing offset.
Reported-by: Xiao Yang <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
fuse_direct_io() can end up advancing the iterator by more than the amount
of data read or written. This case is handled by the generic code if going
through ->direct_IO(), but not in the FOPEN_DIRECT_IO case.
Fix by reverting the extra bytes from the iterator in case of error or a
short count.
To test: install lxcfs, then the following testcase
int fd = open("/var/lib/lxcfs/proc/uptime", O_RDONLY);
sendfile(1, fd, NULL, 16777216);
sendfile(1, fd, NULL, 16777216);
will spew WARN_ON() in iov_iter_pipe().
Reported-by: Peter Geis <[email protected]>
Reported-by: Al Viro <[email protected]>
Fixes: 3c3db095b6 ("fuse: use iov_iter based generic splice helpers")
Cc: <[email protected]> # v5.1
Signed-off-by: Miklos Szeredi <[email protected]>
ovl_lseek() is using ssize_t to return the value from vfs_llseek(). On a
32-bit kernel ssize_t is a 32-bit signed int, which overflows above 2 GB.
Assign the return value of vfs_llseek() to loff_t to fix this.
Reported-by: Boris Gjenero <[email protected]>
Fixes: 9e46b840c7 ("ovl: support stacked SEEK_HOLE/SEEK_DATA")
Cc: <[email protected]> # v4.19
Signed-off-by: Miklos Szeredi <[email protected]>
Buffered read in fuse normally goes via:
-> generic_file_buffered_read()
-> fuse_readpages()
-> fuse_send_readpages()
->fuse_simple_request() [called since v5.4]
In the case of a read request, fuse_simple_request() will return a
non-negative bytecount on success or a negative error value. A positive
bytecount was taken to be an error and the PG_error flag set on the page.
This resulted in generic_file_buffered_read() falling back to ->readpage(),
which would repeat the read request and succeed. Because of the repeated
read succeeding the bug was not detected with regression tests or other use
cases.
The FTP module in GVFS however fails the second read due to the
non-seekable nature of FTP downloads.
Fix by checking and ignoring positive return value from
fuse_simple_request().
Reported-by: Ondrej Holy <[email protected]>
Link: https://gitlab.gnome.org/GNOME/gvfs/issues/441
Fixes: 134831e36b ("fuse: convert readpages to simple api")
Cc: <[email protected]> # v5.4
Signed-off-by: Miklos Szeredi <[email protected]>
exit_aio() is sometimes stuck in wait_for_completion() after aio is issued
with direct IO and the task receives a signal.
The reason is failure to call ->ki_complete() due to a leaked reference to
fuse_io_priv. This happens in fuse_async_req_send() if
fuse_simple_background() returns an error (e.g. -EINTR).
In this case the error value is propagated via io->err, so return success
to not confuse callers.
This issue is tracked as a virtio-fs issue:
https://gitlab.com/virtio-fs/qemu/issues/14
Reported-by: Masayoshi Mizuma <[email protected]>
Fixes: 45ac96ed7c ("fuse: convert direct_io to simple api")
Cc: <[email protected]> # v5.4
Signed-off-by: Miklos Szeredi <[email protected]>
When adding a new hard link, make sure that i_nlink doesn't overflow.
Fixes: ac45d61357 ("fuse: fix nlink after unlink")
Cc: <[email protected]> # v3.4
Signed-off-by: Miklos Szeredi <[email protected]>
Make sure filesystem is not returning a bogus number of bytes written.
Fixes: ea9b9907b8 ("fuse: implement perform_write")
Cc: <[email protected]> # v2.6.26
Signed-off-by: Miklos Szeredi <[email protected]>
If a filesystem returns negative inode sizes, future reads on the file were
causing the cpu to spin on truncate_pagecache.
Create a helper to validate the attributes. This now does two things:
- check the file mode
- check if the file size fits in i_size without overflowing
Reported-by: Arijit Banerjee <[email protected]>
Fixes: d8a5ba4545 ("[PATCH] FUSE - core")
Cc: <[email protected]> # v2.6.14
Signed-off-by: Miklos Szeredi <[email protected]>
If writeback cache is enabled, then writes might get reordered with
chmod/chown/utimes. The problem with this is that performing the write in
the fuse daemon might itself change some of these attributes. In such case
the following sequence of operations will result in file ending up with the
wrong mode, for example:
int fd = open ("suid", O_WRONLY|O_CREAT|O_EXCL);
write (fd, "1", 1);
fchown (fd, 0, 0);
fchmod (fd, 04755);
close (fd);
This patch fixes this by flushing pending writes before performing
chown/chmod/utimes.
Reported-by: Giuseppe Scrivano <[email protected]>
Tested-by: Giuseppe Scrivano <[email protected]>
Fixes: 4d99ff8f12 ("fuse: Turn writeback cache on")
Cc: <[email protected]> # v3.15+
Signed-off-by: Miklos Szeredi <[email protected]>
If the FUSE_READDIRPLUS_AUTO feature is enabled, then lookups on a
directory before/during readdir are used as an indication that READDIRPLUS
should be used instead of READDIR. However if the lookup turns out to be
negative, then selecting READDIRPLUS makes no sense.
Signed-off-by: Miklos Szeredi <[email protected]>
Move the check for async request after check for the request being already
finished and done with.
Reported-by: [email protected]
Fixes: d49937749f ("fuse: stop copying args to fuse_req")
Signed-off-by: Miklos Szeredi <[email protected]>
Virtio-fs does not accept any mount options, so it's confusing and wrong to
show any in /proc/mounts.
Reported-by: Stefan Hajnoczi <[email protected]>
Signed-off-by: Miklos Szeredi <[email protected]>
Don't hold onto dentry in lru list if need to re-lookup it anyway at next
access. Only do this if explicitly enabled, otherwise it could result in
performance regression.
More advanced version of this patch would periodically flush out dentries
from the lru which have gone stale.
Signed-off-by: Miklos Szeredi <[email protected]>
The size of struct fuse_req was reduced from 392B to 144B on a non-debug
config, thus the sanitize_global_limit() helper was setting a larger
default limit. This doesn't really reflect reduction in the memory used by
requests, since the fields removed from fuse_req were added to fuse_args
derived structs; e.g. sizeof(struct fuse_writepages_args) is 248B, thus
resulting in slightly more memory being used for writepage requests
overalll (due to using 256B slabs).
Make the calculatation ignore the size of fuse_req and use the old 392B
value.
Signed-off-by: Miklos Szeredi <[email protected]>
The page array pointers are also duplicated across fuse_args_pages and
fuse_req. Get rid of the fuse_req ones.
Signed-off-by: Miklos Szeredi <[email protected]>
No need to duplicate the argument arrays in fuse_req, so just dereference
req->args instead of copying to the fuse_req internal ones.
This allows further cleanup of the fuse_req structure.
Signed-off-by: Miklos Szeredi <[email protected]>
Miklos Szeredi