This does not do what I wanted as all dirty buffers must be flushed
by the call to ffs_sync and any remaining dependency work would mean
that this failed.
Pointed out by: tegge
This does not do what I wanted as all dirty buffers must be flushed
by the call to ffs_sync and any remaining dependency work would mean
that this failed.
Pointed out by: tegge
Fix detection of active unlinked files by checking VI_OWEINACT and
VI_DOINGINACT in addition to v_usecount.
Defer inactive handling for unlinked files if the file system is mostly
suspended (secondary writes being blocked).
Perform deferred inactive handling after the file system is resumed.
replacement for vn_write_suspend_wait() to better account for secondary write
processing.
Close race where secondary writes could be started after ffs_sync() returned
but before the file system was marked as suspended.
Detect if secondary writes or softdep processing occurred during vnode sync
loop in ffs_sync() and retry the loop if needed.
be called without any vnode locks held. Remove calls to vn_start_write() and
vn_finished_write() in vnode_pager_putpages() and add these calls before the
vnode lock is obtained to most of the callers that don't already have them.
has many positive effects including improved smp locking, reducing
interdependencies between mounts that can lead to deadlocks, etc.
- Add the softdep worklist and various counters to the ufsmnt structure.
- Add a mount pointer to the workitem and remove mount pointers from the
various structures derived from the workitem as they are now redundant.
- Remove the poor-man's semaphore protecting softdep_process_worklist and
softdep_flushworklist. Several threads may now process the list
simultaneously.
- Add softdep_waitidle() to block the thread until all pending
dependencies being operated on by other threads have been flushed.
- Use softdep_waitidle() in unmount and snapshots to block either
operation until the fs is stable.
- Remove softdep worklist processing from the syncer and move it into the
softdep_flush() thread. This thread processes all softdep mounts
once each second and when it is called via the new softdep_speedup()
when there is a resource shortage. This removes the softdep hook
from the kernel and various hacks in header files to support it.
Reviewed by/Discussed with: tegge, truckman, mckusick
Tested by: kris
directory. vrele() may lock the passed vnode, which in these cases would
give an invalid lock order of child -> parent. These situations are
deadlock prone although do not typically deadlock because the vrele
is typically not releasing the last reference to the vnode. Users of
vrele must consider it as a call to vn_lock() and order it appropriately.
MFC After: 1 week
Sponsored by: Isilon Systems, Inc.
Tested by: kkenn
that NetBSD implemented it independently of them (don't know which one
was actually first). This saves about 24k for those times you don't
need snapshot support (like when running off a ram disk, or in an
embedded environment where size matters).
for export structure and pass that to vfs_export().
Currently in userland mount(8), an export structure is unconditionally
passed in, only for UFS. This is an attempt to move that UFS-specific
behavior out of mount(8) and into the UFS filesystem code.
callpath via vfs_getopt(), and set the appropriate MNT_* flag:
-> acls, async, force, multilabel, noasync, noatime,
-> noclusterr, noclusterw, snapshot, update
- Allow errmsg as a valid mount option via vfs_getopt(),
so we can later add a hook to propagate mount errors back
to userspace via vfs_mount_error().
- Prefer '_' to ' ', as it results in more easily parsed results in
memory monitoring tools such as vmstat.
- Remove punctuation that is incompatible with using memory type names
as file names, such as '/' characters.
- Disambiguate some collisions by adding subsystem prefixes to some
memory types.
- Generally prefer lower case to upper case.
- If the same type is defined in multiple architecture directories,
attempt to use the same name in additional cases.
Not all instances were caught in this change, so more work is required to
finish this conversion. Similar changes are required for UMA zone names.
to (max block - 1) * bsize. For DEV_BSIZE, this doubles the limit from
0.5 TB to 1 TB. For the old 4.4 FFS case, decrease the limit from 0.5 TB
to 2 GB - 1. Older systems had a 32 bit off_t so they couldn't access the
larger files anyway.
Collaboration with: bde
case by saving the value of dp->i_ino before unlocking the vnode
for the current directory and passing the saved value to VFS_VGET().
Without this change, another thread can overwrite dp->i_ino after
the current directory is unlocked, causing ufs_lookup() to lock
and return the wrong vnode in place of the vnode for its parent
directory. A deadlock can occur if dp->i_ino was changed to a
subdirectory of the current directory because the root to leaf vnode
lock ordering will be violated. A vnode lock can be leaked if
dp->i_ino was changed to point to the current directory, which
causes the current vnode lock for the current directory to be
recursed, which confuses lookup() into calling vrele() when it
should be calling vput().
The probability of this bug being triggered seems to be quite low
unless the sysctl variable debug.vfscache is set to 0.
Reviewed by: jhb
MFC after: 2 weeks
auto-start, set cnp.cn_lkflags to LK_EXCLUSIVE. This flag must now
be set so that lockmgr knows what kind of lock to acquire, and it
will panic if not specified. This resulted in a panic when using
extended attributes on UFS1 as of locking work present in the 6.x
branch.
This is a RELENG_6_0 merge candidate.
Reported by: lofi
MFC after: 3 days
updated by a process holding the snapshot lock. Another process updating a
different inode in the same inodeblock will do copy on write checks and lock in
the opposite direction.
The snapshot code force a copy on write of these blocks manually (cf. start of
expunge_ufs[12]) and these inode blocks are later put on snapblklist.
This partial fix is to 'drain' the relevant ffs_copyonwrite() operation after
installing new snapblklist. This is not a 100% solution since a failed block
allocation can cause implicit fsync() which might deadlock before the new
snapblklist has been installed.
file is flushed by a process not holding snaplk (e.g. bufdaemon). Another
process might hold snaplk and try to access the block due to ffs_copyonwrite
processing.
the cg map buffer being held when writing indirect blocks. The process ends up
in ffs_copyonwrite(), attempting to get snaplk while holding the cg map buffer
lock.
Another process might be in ffs_copyonwrite(), trying to allocate a new block
for a copy. It would hold snaplk while trying to get the cg map buffer lock.
Release the cg map buffer early and use the copy for most of the cgaccount
processing to avoid this deadlock.
skipping the call from ffs_snapremove() if the block number is zero.
Simplify snapshot locking in ffs_copyonwrite() and ffs_snapblkfree() by using
the same locking protocol for low block numbers as for larger block numbers.
This removes a lock leak that could happen if vn_lock() succeeded after
lockmgr() failed in ffs_snapblkfree().
Check if snapshot is gone before retrying a lock in ffs_copyonwrite().
reclamation. If the vnode previously was a fifo then v_op would point to
ffs_fifoops[12] instead of the expected ffs_vnodeops[12], causing a panic at
the end of ffsext_strategy.
stale flag bits left over from before the inode was recycled.
Without this change, a leftover IN_SPACECOUNTED flag could prevent
softdep_freefile() and softdep_releasefile() from incrementing
fs_pendinginodes. Because handle_workitem_freefile() unconditionally
decrements fs_pendinginodes, a negative value could be reported at
file system unmount time with a message like:
unmount pending error: blocks 0 files -3
The pending block count in fs_pendingblocks could also be negative
for similar reasons. These errors can cause the data returned by
statfs() to be slightly incorrect. Some other cleanup code in
softdep_releasefile() could also be incorrectly bypassed.
MFC after: 3 days
Add a new private thread flag to indicate that the thread should
not sleep if runningbufspace is too large.
Set this flag on the bufdaemon and syncer threads so that they skip
the waitrunningbufspace() call in bufwrite() rather than than
checking the proc pointer vs. the known proc pointers for these two
threads. A way of preventing these threads from being starved for
I/O but still placing limits on their outstanding I/O would be
desirable.
Set this flag in ffs_copyonwrite() to prevent bufwrite() calls from
blocking on the runningbufspace check while holding snaplk. This
prevents snaplk from being held for an arbitrarily long period of
time if runningbufspace is high and greatly reduces the contention
for snaplk. The disadvantage is that ffs_copyonwrite() can start
a large amount of I/O if there are a large number of snapshots,
which could cause a deadlock in other parts of the code.
Call runningbufwakeup() in ffs_copyonwrite() to decrement runningbufspace
before attempting to grab snaplk so that I/O requests waiting on
snaplk are not counted in runningbufspace as being in-progress.
Increment runningbufspace again before actually launching the
original I/O request.
Prior to the above two changes, the system could deadlock if enough
I/O requests were blocked by snaplk to prevent runningbufspace from
falling below lorunningspace and one of the bawrite() calls in
ffs_copyonwrite() blocked in waitrunningbufspace() while holding
snaplk.
See <http://www.holm.cc/stress/log/cons143.html>
the directory's inode after queuing the dirrem that will decrement
the parent directory's link count. This will force the update of
the parent directory's actual link to actually be scheduled. Without
this change the parent directory's actual link count would not be
updated until ufs_inactive() cleared the inode of the newly removed
directory, which might be deferred indefinitely. ufs_inactive()
will not be called as long as any process holds a reference to the
removed directory, and ufs_inactive() will not clear the inode if
the link count is non-zero, which could be the result of an earlier
system crash.
If a background fsck is run before the update of the parent directory's
actual link count has been performed, or at least scheduled by
putting the dirrem on the leaf directory's inodedep id_bufwait list,
fsck will corrupt the file system by decrementing the parent
directory's effective link count, which was previously correct
because it already took the removal of the leaf directory into
account, and setting the actual link count to the same value as the
effective link count after the dangling, removed, leaf directory
has been removed. This happens because fsck acts based on the
actual link count, which will be too high when fsck creates the
file system snapshot that it references.
This change has the fortunate side effect of more quickly cleaning
up the large number dirrem structures that linger for an extended
time after the removal of a large directory tree. It also fixes a
potential problem with the shutdown of the syncer thread timing out
if the system is rebooted immediately after removing a large directory
tree.
Submitted by: tegge
MFC after: 3 days
osf1_signal.c:1.41, amd64/amd64/trap.c:1.291, linux_socket.c:1.60,
svr4_fcntl.c:1.36, svr4_ioctl.c:1.23, svr4_ipc.c:1.18, svr4_misc.c:1.81,
svr4_signal.c:1.34, svr4_stat.c:1.21, svr4_stream.c:1.55,
svr4_termios.c:1.13, svr4_ttold.c:1.15, svr4_util.h:1.10,
ext2_alloc.c:1.43, i386/i386/trap.c:1.279, vm86.c:1.58,
unaligned.c:1.12, imgact_elf.c:1.164, ffs_alloc.c:1.133:
Now that Giant is acquired in uprintf() and tprintf(), the caller no
longer leads to acquire Giant unless it also holds another mutex that
would generate a lock order reversal when calling into these functions.
Specifically not backed out is the acquisition of Giant in nfs_socket.c
and rpcclnt.c, where local mutexes are held and would otherwise violate
the lock order with Giant.
This aligns this code more with the eventual locking of ttys.
Suggested by: bde
in rev. 1.40 of ufs_inode.c, which allows an inode being truncated
even when the filesystem itself is marked RDONLY. A subsequent
call of UFS_TRUNCATE (ffs_truncate) would panic the system as it
asserts that it can only be called when the filesystem is mounted
read-write (same changeset, rev. 1.74 of sys/ufs/ffs/ffs_inode.c).
Because ffs_mount() already takes care of sync'ing the filesystem
to disk before being downgraded to readonly, it appears to be more
desirable that we should not permit this sort of writes to disk.
This change would fix a panic that occours when read-only mounted
a corrupted filesystem and doing some file operations.
MT6/5/4 candidate
Reviewed by: mckusick
as they both interact with the tty code (!MPSAFE) and may sleep if the
tty buffer is full (per comment).
Modify all consumers of uprintf() and tprintf() to hold Giant around
calls into these functions. In most cases, this means adding an
acquisition of Giant immediately around the function. In some cases
(nfs_timer()), it means acquiring Giant higher up in the callout.
With these changes, UFS no longer panics on SMP when either blocks are
exhausted or inodes are exhausted under load due to races in the tty
code when running without Giant.
NB: Some reduction in calls to uprintf() in the svr4 code is probably
desirable.
NB: In the case of nfs_timer(), calling uprintf() while holding a mutex,
or even in a callout at all, is a bad idea, and will generate warnings
and potential upset. This needs to be fixed, but was a problem before
this change.
NB: uprintf()/tprintf() sleeping is generally a bad ideas, as is having
non-MPSAFE tty code.
MFC after: 1 week
Also introduce an aclinit function which will be used to create the UMA zone
for use by file systems at system start up.
MFC after: 1 month
Discussed with: rwatson
Don't free a struct inodedep if another process is allocating saved inode
memory for the same struct inodedep in initiate_write_inodeblock_ufs[12]().
Handle disappearing dependencies in softdep_disk_io_initiation().
Reviewed by: mckusick
or unreadable blocks, make sure to destroy the mutex we created.
Also fix an unrelated typo in a comment.
Found by: Peter Holm's stress tests
Reviewed by: dwmalone
MFC after: 3 days
make the b_iodone callback responsible for setting it if it is needed.
Previously, it was set unconditionally by bufdone() without holding
whichever lock is shared by the b_iodone callback and the corresponding
top-half function. Consequently, in a race, the top-half function could
conclude that operation was done before the b_iodone callback finished.
See, for example, aio_physwakeup() and aio_fphysio().
Note: I don't believe that the other, more widely-used b_iodone callbacks
are affected.
Discussed with: jeff
Reviewed by: phk
MFC after: 2 weeks
UFS by:
- Making the pre and post hooks for the VOP functions work even when
DEBUG_VFS_LOCKS is not defined.
- Moving the KNOTE activations into the corresponding VOP hooks.
- Creating a MNTK_NOKNOTE flag for the mnt_kern_flag field of struct
mount that permits filesystems to disable the new behavior.
- Creating a default VOP_KQFILTER function: vfs_kqfilter()
My benchmarks have not revealed any performance degradation.
Reviewed by: jeff, bde
Approved by: rwatson, jmg (kqueue changes), grehan (mentor)
file's access time should be updated when it gets executed. A while
ago the mechanism used to exec was changed to use a more mmap based
mechanism and this behavior was broken as a side-effect of that.
A new vnode flag is added that gets set when the file gets executed,
and the VOP_SETATTR() vnode operation gets called. The underlying
filesystem is expected to handle it based on its own semantics, some
filesystems don't support access time at all. Those that do should
handle it in a way that does not block, does not generate I/O if possible,
etc. In particular vn_start_write() has not been called. The UFS code
handles it the same way as it would normally handle the access time if
a file was read - the IN_ACCESS flag gets set in the inode but no other
action happens at this point. The actual time update will happen later
during a sync (which handles all the necessary locking).
Got me into this: cperciva
Discussed with: a lot with bde, a little with kan
Showed patches to: phk, jeffr, standards@, arch@
Minor discussion on: arch@
are subtle differences in the read and write completion path. Instead,
grab an extra write ref so the write path can drop it when we recursively
call bufdone(). I believe this may be the source of the wrong bufobj
panics.
Reported by: pho, kkenn
occur on a filesystem running with soft updates after a crash and
before a background fsck has been run. To prevent discrepancies
from arising in a background fsck that may already be running,
the directory is removed but its inode is not freed and is left
with the residual reference count. When encountered by the
background fsck it will be reclaimed.
export. This was happening anyway since this file manually sets DEBUG.
- Add a sysctl for the number of items on the worklist.
- Use a more canonical loop restart in softdep_fsync_mountdev, it saves
some code at the expense of a goto and makes me worry less about
modifying a variable that should be private to the TAILQ_FOREACH_SAFE
macro.
- Don't intermingle direct calls to lockmgr and indirect calls through
VOPs. This will be important in the future.
- Dont lock the devvp's interlock just to release it on the next line by
passing LK_INTERLOCK to lockmgr.
- Restructure ffs_snapshot_unmount so we don't call free() with the
devvp's interlock locked.
because it may change identities while we're sleeping on the lock.
Otherwise we may bail out of ffs_sync() early due to an error from
deadfs.
- Collapse a VOP_UNLOCK, vrele into a single vput().
two bugs.
- ffs_disk_prewrite was pulling the vp from the buf and checking for
COPYONWRITE, when really it wanted the vp from the bufobj that we're
writing to, which is the devvp. This lead to us skipping the copy on
write to all file data, which significantly broke snapshots for the
last few months.
- When the SOFTUPDATES option was not included in the kernel config we
would also skip the copy on write check, which would effectively disable
snapshots.
- Remove an invalid mp_fixme().
Debugging tips from: mckusick
Reported by: iedowse, others
Discussed with: phk
rely on ufs to always leave the parent locked except in the ISDOTDOT
case. Adjust asserts to deal with these changes.
Sponsored by: Isilon Systems, Inc.
add more work are forced to process two worklist items first.
However, processing an item may generate additional work, causing the
unlucky thread to recursively process the worklist. Add a per-thread
flag to detect this situation and avoid the recursion. This should
fix the stack overflows that could occur while removing large
directory trees.
Tested by: kris
Reviewed by: mckusick
the filesystem. Check that rather than VI_XLOCK.
- Shorten ffs_reload by one step. The old check for an inactive vnode
was slightly racey, and the code which deals with still active vnodes
is not much more expensive.
Sponsored by: Isilon Systems, Inc.
required.
- In ufs_close(), don't do the EAGAIN vrele hack, the top layer now calls
vn_start_write before the lock is acquired as it should.
Sponsored by: Isilon Systems, Inc.
- Also in ufs_inactive, don't acquire the vnode interlock where it isn't
strictly needed. Also owning the vnode interlock while calling vprint()
will cause locking assertions to trip.
Sponsored by: Isilon Systems, Inc.
very slow process, especially for large file systems that is just
recovered from a crash.
Since the summary is already re-sync'ed every 30 second, we will
not lag behind too much after a crash. With this consideration
in mind, it is more reasonable to transfer the responsibility to
background fsck, to reduce the delay after a crash.
Add a new sysctl variable, vfs.ffs.compute_summary_at_mount, to
control this behavior. When set to nonzero, we will get the
"old" behavior, that the summary is computed immediately at mount
time.
Add five new sysctl variables to adjust ndir, nbfree, nifree,
nffree and numclusters respectively. Teach fsck_ffs about these
API, however, intentionally not to check the existence, since
kernels without these sysctls must have recomputed the summary
and hence no adjustments are necessary.
This change has eliminated the usual tens of minutes of delay of
mounting large dirty volumes.
Reviewed by: mckusick
MFC After: 1 week
patch from kan@).
Pull bufobj_invalbuf() out of vinvalbuf() and make g_vfs call it on
close. This is not yet a generally safe function, but for this very
specific use it is safe. This solves the problem with buffers not
being flushed by unmount or after failed mount attempts.
invalidate pending io and dependencies. However, vinvalbuf() rightfully
does not call vnode_pager_setsize() for us. We must do this here. This
could potentially have caused numerous kinds of bugs, but it was
specifically causing msync() deadlocks because msync() was writing
flushing pages that should not have been valid.
Sponsored by: Isilon Systems, Inc.
Reported by: kkenn
behaviour of chflags within a jail. If set to 0 (the default), then a
jailed root user is treated as an unprivileged user; if set to 1, then
a jailed root user is treated the same as an unjailed root user.
This is necessary to allow "make installworld" to work inside a jail,
since it attempts to manipulate the system immutable flag on certain
files.
Discussed with: csjp, rwatson
MFC after: 2 weeks
Give FFS vnodes a specific bufwrite method which contains all the
background write stuff and then calls into the default bufwrite()
for the rest of the job.
Remove all the background write related stuff from the normal bufwrite.
This drags the softdep_move_dependencies() back into FFS.
Long term, it is worth looking at simply copying the data into
allocated memory and issuing the bio directly and not create the
"shadow buf" in the first place (just like copy-on-write is done
in snapshots for instance). I don't think we really gain anything
but complexity from doing this with a buf.
The "business class upgrade" was implemented in UFS's VOP_LOCK
implementation ufs_lock() which is the wrong layer, so move it to
ffs_lock().
Also, as long as we have not abandonned advanced vfs-stacking we
should not preclude it from happening: instead of implementing a
copy locally, use the VOP_LOCK_APV(&ufs) to correctly arrive at
vop_stdlock() at the bottom.
The "business class upgrade" was implemented in UFS's VOP_LOCK
implementation ufs_lock() which is the wrong layer, so move it to
ffs_lock().
Also, as long as we have not abandonned advanced vfs-stacking we
should not preclude it from happening: instead of implementing a
copy locally, use the VOP_LOCK_APV(&ufs) to correctly arrive at
vop_stdlock() at the bottom.
executed. This appears to violate most of the UNIX-ish standards.
One example quote from:
http://www.opengroup.org/onlinepubs/009695399/functions/exec.html
Upon successful completion, the exec functions shall mark for update
the st_atime field of the file. If an exec function failed but was
able to locate the process image file, whether the st_atime field is
marked for update is unspecified. Should the exec function succeed,
the process image file shall be considered to have been opened with
open().
This appears to take care of it for ufs filesystems, doing the necessary
sanity checks (read-only filesystem, etc) without violating any other
standards (setting atime for any open appears to be allowed in any standards
I could find).
Noticed by: cperciva
Reviewed by: kan, rwatson
with NFS.
We are moving responsibility for creating the vnode_pager object into
the filesystems which own the vnode, and this is one of the places
we have to cover.
We call vnode_create_vobject() directly because we own the vnode.
If we can get the size easily, pass it as an argument to save the
call to VOP_GETATTR() in vnode_create_vobject()
- Expand the scope of lk to cover not only interrupt races, but also
top-half races, which includes many new uses over global top-half
only data.
- Get rid of interlocked_sleep() and use msleep or BUF_LOCK where
appropriate.
- Use the lk mutex in place of the various hand rolled semaphores.
- Stop dropping the lk lock before we panic.
- Fix getdirtybuf() callers so that they reacquire access to whatever
softdep datastructure they were inxpecting in the failure/retry
case. Previously, sleeps in getdirtybuf() could leave us with
pointers to bad memory.
- Update handling of ffs to be compatible with ffs locking changes.
Sponsored By: Isilon Systems, Inc.
- Use the buffer lock on the superblock buf to serialize calls to
sbupdate.
- Set the MNTK_MPSAFE flag when QUOTA is not defined in the kernel.
Sponsored By: Isilon Systems, Inc.
it is now quite naturally protected by the ufsmount mutex.
- Use the ufs lock to protect various fields in struct fs, primarily the
cg summary needs protection to avoid allocation races. Several
functions have been slightly re-arranged to reduce the number of
lock operations.
- Adjust several functions (blkfree, freefile, etc.) to accept a
ufsmount as an argument so that we may access the ufs lock.
Sponsored By: Isilon Systems, Inc.
any per-instance global data that is not already protected by a
buf or vnode lock. Presently, only fields in ffs's struct fs utilize
this lock.
- Sort some ufsmount members so that fields used for quotas are grouped
together. This is in anticipation of quota locking.
Sponsored By: Isilon Systems, Inc.
Without this fix, when ACLs are set via tunefs(8) on the root file system,
they are removed on boot when 'mount -a' is called, because mount(8)
called for the root file system always add MNT_UPDATE flag and MNT_UPDATE
flag isn't perfect.
Now, one cannot remove ACLs stored in superblock (configured with tunefs(8))
via 'mount -a' nor 'mount -u -o noacls <file system>', but it is still
possible to mount file system which doesn't have ACLs in superblock via
'mount -o acls <file system>' or /etc/fstab's 'acls' option.
Reported by: Lech Lorens/pl.comp.os.bsd
Discussed with: phk, rwatson
Reviewed by: rwatson
MFC after: 2 weeks
I'm not sure why a credential was added to these in the first place, it is
not used anywhere and it doesn't make much sense:
The credentials for syncing a file (ability to write to the
file) should be checked at the system call level.
Credentials for syncing one or more filesystems ("none")
should be checked at the system call level as well.
If the filesystem implementation needs a particular credential
to carry out the syncing it would logically have to the
cached mount credential, or a credential cached along with
any delayed write data.
Discussed with: rwatson
