Ensure that softdep_unmount() and softdep_setup_sbupdate()
only get called for filesystems running with soft dependencies.
No functional change.
Tested by: Peter Holm and Scott Long
Sponsored by: Netflix
Convert three functions exported from ffs_softdep.c to static
functions as they are not used outside of ffs_softdep.c.
No functional change.
Tested by: Peter Holm and Scott Long
Sponsored by: Netflix
in the future in a backward compatible (API and ABI) way.
The cap_rights_t represents capability rights. We used to use one bit to
represent one right, but we are running out of spare bits. Currently the new
structure provides place for 114 rights (so 50 more than the previous
cap_rights_t), but it is possible to grow the structure to hold at least 285
rights, although we can make it even larger if 285 rights won't be enough.
The structure definition looks like this:
struct cap_rights {
uint64_t cr_rights[CAP_RIGHTS_VERSION + 2];
};
The initial CAP_RIGHTS_VERSION is 0.
The top two bits in the first element of the cr_rights[] array contain total
number of elements in the array - 2. This means if those two bits are equal to
0, we have 2 array elements.
The top two bits in all remaining array elements should be 0.
The next five bits in all array elements contain array index. Only one bit is
used and bit position in this five-bits range defines array index. This means
there can be at most five array elements in the future.
To define new right the CAPRIGHT() macro must be used. The macro takes two
arguments - an array index and a bit to set, eg.
#define CAP_PDKILL CAPRIGHT(1, 0x0000000000000800ULL)
We still support aliases that combine few rights, but the rights have to belong
to the same array element, eg:
#define CAP_LOOKUP CAPRIGHT(0, 0x0000000000000400ULL)
#define CAP_FCHMOD CAPRIGHT(0, 0x0000000000002000ULL)
#define CAP_FCHMODAT (CAP_FCHMOD | CAP_LOOKUP)
There is new API to manage the new cap_rights_t structure:
cap_rights_t *cap_rights_init(cap_rights_t *rights, ...);
void cap_rights_set(cap_rights_t *rights, ...);
void cap_rights_clear(cap_rights_t *rights, ...);
bool cap_rights_is_set(const cap_rights_t *rights, ...);
bool cap_rights_is_valid(const cap_rights_t *rights);
void cap_rights_merge(cap_rights_t *dst, const cap_rights_t *src);
void cap_rights_remove(cap_rights_t *dst, const cap_rights_t *src);
bool cap_rights_contains(const cap_rights_t *big, const cap_rights_t *little);
Capability rights to the cap_rights_init(), cap_rights_set(),
cap_rights_clear() and cap_rights_is_set() functions are provided by
separating them with commas, eg:
cap_rights_t rights;
cap_rights_init(&rights, CAP_READ, CAP_WRITE, CAP_FSTAT);
There is no need to terminate the list of rights, as those functions are
actually macros that take care of the termination, eg:
#define cap_rights_set(rights, ...) \
__cap_rights_set((rights), __VA_ARGS__, 0ULL)
void __cap_rights_set(cap_rights_t *rights, ...);
Thanks to using one bit as an array index we can assert in those functions that
there are no two rights belonging to different array elements provided
together. For example this is illegal and will be detected, because CAP_LOOKUP
belongs to element 0 and CAP_PDKILL to element 1:
cap_rights_init(&rights, CAP_LOOKUP | CAP_PDKILL);
Providing several rights that belongs to the same array's element this way is
correct, but is not advised. It should only be used for aliases definition.
This commit also breaks compatibility with some existing Capsicum system calls,
but I see no other way to do that. This should be fine as Capsicum is still
experimental and this change is not going to 9.x.
Sponsored by: The FreeBSD Foundation
allocations under low free-space conditions (-r254995), determine
that old block-preference search order used before -r249782 worked
a bit better. This change reverts to that block-preference search order.
MFC after: 2 weeks
I have 25TB Dell PERC 6 RAID5 array. When it becomes almost
full (10-20GB free), processes which write data to it start
eating 100% CPU and write speed drops below 1MB/sec (normally
to gives 400MB/sec). The revision at which it first became
apparent was http://svnweb.freebsd.org/changeset/base/249782.
The offending change reserved an area in each cylinder group to
store metadata. The new algorithm attempts to save this area for
metadata and allows its use for non-metadata only after all the
data areas have been exhausted. The size of the reserved area
defaults to half of minfree, so the filesystem reports full before
the data area can completely fill. However, in this report, the
filesystem has had minfree reduced to 1% thus forcing the metadata
area to be used for data. As the filesystem approached full, it
had only metadata areas left to allocate. The result was that
every block allocation had to scan summary data for 30,000 cylinder
groups before falling back to searching up to 30,000 metadata areas.
The fix is to give up on saving the metadata areas once the free
space reserve drops below 2%. The effect of this change is to use
the old algorithm of just accepting the first available block that
we find. Since most filesystems use the default 5% minfree, this
will have no effect on their operation. For those that want to push
to the limit, they will get their crappy block placements quickly.
Submitted by: Dmitry Sivachenko
Fix Tested by: Dmitry Sivachenko
PR: kern/181226
MFC after: 2 weeks
persist much longer than previously. Historically we had at most 100
entries; now the count may reach a million. With the increased count
we spent far too much time looking them up in the grossly undersized
newblk hash table. Configure the newblk hash table to accurately reflect
the number of entries that it must index.
Reviewed by: kib
Tested by: Peter Holm
MFC after: 2 weeks
we need to collect the highest level of allocation for each of the
different soft update dependency structures. This change collects these
statistics and makes them available using `sysctl debug.softdep.highuse'.
Reviewed by: kib
Tested by: Peter Holm
MFC after: 2 weeks
block copy, when copying the superblock into the snapshot. UFS1 does
not align superblock on the block boundary, and bcopy runs off the end
of the buffer.
Reported by: Andre Albsmeier <Andre.Albsmeier@siemens.com>
Reviewed by: mckusick
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
vfs_busy(mp);
vfs_write_suspend(mp);
which are problematic if other thread starts unmount between two
calls. The unmount starts a write, while vfs_write_suspend() drain
writers. On the other hand, unmount drains busy references, causing
the deadlock.
Add a flag argument to vfs_write_suspend and require the callers of it
to specify VS_SKIP_UNMOUNT flag, when the call is performed not in the
mount path, i.e. the covered vnode is not locked. The suspension is
not attempted if VS_SKIP_UNMOUNT is specified and unmount is in
progress.
Reported and tested by: Andreas Longwitz <longwitz@incore.de>
Sponsored by: The FreeBSD Foundation
MFC after: 3 weeks
Revert the simplification of the i_gen calculation.
It is still a good idea to avoid zero values and for the case
of old filesystems there is probably no advantage in using
the complete 32 bits anyways.
Discussed with: bde
MFC after: 4 weeks
Further simplify the i_gen calculation for older disks.
Having a zero here is not really a problem and this is more
similar to what is done in newfs_random().
Reported by: Xin Li
MFC after: 4 weeks
In UFS, i_gen is a random generated value and there is not way for
it to be negative. Actually, the value of i_gen is just used to
match bit patterns and it is of not consequence if the values are
signed or not.
Following other filesystems, set it to unsigned and use it as such,
Discussed by: mckusick
Reviewed by: mckusick (previous version)
MFC after: 4 weeks
- Use a shared bufobj lock in getblk() and inmem().
- Convert softdep's lk to rwlock to match the bufobj lock.
- Move INFREECNT to b_flags and protect it with the buf lock.
- Remove unnecessary locking around bremfree() and BKGRDINPROG.
Sponsored by: EMC / Isilon Storage Division
Discussed with: mckusick, kib, mdf
an error. One could argue that returning a buffer even when it is
not valid is incorrect, but bread has always returned a buffer
valid or not.
Reviewed by: kib
MFC after: 2 weeks
the return value is NULL. Based on the returned flags, the
return value should never be inspected in the case where NULL
is returned, but it is good coding practice not to return a
pointer to freed memory.
Found by: Coverity Scan, CID 1006096
Reviewed by: kib
MFC after: 2 weeks
- Don't insert BKGRDMARKER bufs into the splay or dirty/clean buf lists.
No consumers need to find them there and it complicates the tree.
These flags are all FFS specific and could be moved out of the buf
cache.
- Use pbgetvp() and pbrelvp() to associate the background and journal
bufs with the vp. Not only is this much cheaper it makes more sense
for these transient bufs.
- Fix the assertions in pbget* and pbrel*. It's not safe to check list
pointers which were never initialized. Use the BX flags instead. We
also check B_PAGING in reassignbuf() so this should cover all cases.
Discussed with: kib, mckusick, attilio
Sponsored by: EMC / Isilon Storage Division
in the pagedep and inodedep hash tables. An entry in the table is
skipped because 'pagedep_hash' and 'inodedep_hash' hold the size
of the hash tables - 1.
The chance that this would have any operational failure is extremely
unlikely. These funtions only need to find a single entry and are
only called when there are too many entries. The chance that they
would fail because all the entries are on the single skipped hash
chain are remote.
Submitted by: Pedro Martelletto
Reviewed by: kib
MFC after: 2 weeks
running time for a full fsck. It also reduces the random access time
for large files and speeds the traversal time for directory tree walks.
The key idea is to reserve a small area in each cylinder group
immediately following the inode blocks for the use of metadata,
specifically indirect blocks and directory contents. The new policy
is to preferentially place metadata in the metadata area and
everything else in the blocks that follow the metadata area.
The size of this area can be set when creating a filesystem using
newfs(8) or changed in an existing filesystem using tunefs(8).
Both utilities use the `-k held-for-metadata-blocks' option to
specify the amount of space to be held for metadata blocks in each
cylinder group. By default, newfs(8) sets this area to half of
minfree (typically 4% of the data area).
This work was inspired by a paper presented at Usenix's FAST '13:
www.usenix.org/conference/fast13/ffsck-fast-file-system-checker
Details of this implementation appears in the April 2013 of ;login:
www.usenix.org/publications/login/april-2013-volume-38-number-2.
A copy of the April 2013 ;login: paper can also be downloaded
from: www.mckusick.com/publications/faster_fsck.pdf.
Reviewed by: kib
Tested by: Peter Holm
MFC after: 4 weeks
cluster_write() and cluster_wbuild() functions. The flags to be
allowed are a subset of the GB_* flags for getblk().
Sponsored by: The FreeBSD Foundation
Tested by: pho
future further optimizations where the vm_object lock will be held
in read mode most of the time the page cache resident pool of pages
are accessed for reading purposes.
The change is mostly mechanical but few notes are reported:
* The KPI changes as follow:
- VM_OBJECT_LOCK() -> VM_OBJECT_WLOCK()
- VM_OBJECT_TRYLOCK() -> VM_OBJECT_TRYWLOCK()
- VM_OBJECT_UNLOCK() -> VM_OBJECT_WUNLOCK()
- VM_OBJECT_LOCK_ASSERT(MA_OWNED) -> VM_OBJECT_ASSERT_WLOCKED()
(in order to avoid visibility of implementation details)
- The read-mode operations are added:
VM_OBJECT_RLOCK(), VM_OBJECT_TRYRLOCK(), VM_OBJECT_RUNLOCK(),
VM_OBJECT_ASSERT_RLOCKED(), VM_OBJECT_ASSERT_LOCKED()
* The vm/vm_pager.h namespace pollution avoidance (forcing requiring
sys/mutex.h in consumers directly to cater its inlining functions
using VM_OBJECT_LOCK()) imposes that all the vm/vm_pager.h
consumers now must include also sys/rwlock.h.
* zfs requires a quite convoluted fix to include FreeBSD rwlocks into
the compat layer because the name clash between FreeBSD and solaris
versions must be avoided.
At this purpose zfs redefines the vm_object locking functions
directly, isolating the FreeBSD components in specific compat stubs.
The KPI results heavilly broken by this commit. Thirdy part ports must
be updated accordingly (I can think off-hand of VirtualBox, for example).
Sponsored by: EMC / Isilon storage division
Reviewed by: jeff
Reviewed by: pjd (ZFS specific review)
Discussed with: alc
Tested by: pho
Current dqflush() panics when a dquot with with non-zero refcount is
encountered. The situation is possible, because quotas are turned off
before softdep workitem queue if flushed, due to the quota file writes
might create softdep workitems.
Make the encountering an active dquot in dqflush() not fatal, return
the error from quotaoff() instead. Ignore the quotaoff() failures
when ffs_flushfiles() is called in the course of softdep_flushfiles()
loop, until the last iteration. At the last loop, the quotas must be
closed, and because SU workitems should be already flushed, the
references to dquot are gone.
Sponsored by: The FreeBSD Foundation
Reported and tested by: pho
Reviewed by: mckusick
MFC after: 2 weeks
The order is inode buffer lock -> snaplk -> cg buffer lock, reversing
the order causes deadlocks.
Inode block must not be written while cg block buffer is owned. The
FFS copy on write needs to allocate a block to copy the content of the
inode block, and the cylinder group selected for the allocation might
be the same as the owned cg block. The reserved block detection code
in the ffs_copyonwrite() and ffs_bp_snapblk() is unable to detect the
situation, because the locked cg buffer is not exposed to it.
In order to maintain the dependency between initialized inode block
and the cg_initediblk pointer, look up the inode buffer in
non-blocking mode. If succeeded, brelse cg block, initialize the inode
block and write it. After the write is finished, reread cg block and
update the cg_initediblk.
If inode block is already locked by another thread, let the another
thread initialize it. If another thread raced with us after we
started writing inode block, the situation is detected by an update of
cg_initediblk. Note that double-initialization of the inode block is
harmless, the block cannot be used until cg_initediblk is incremented.
Sponsored by: The FreeBSD Foundation
In collaboration with: pho
Reviewed by: mckusick
MFC after: 1 month
X-MFC-note: after r246877
When a cylinder group runs short of inodes, a new block for inodes is
allocated, zero'ed, and written to the disk. The zero'ed inodes must
be on the disk before the cylinder group can be updated to claim them.
If the cylinder group claiming the new inodes were written before the
zero'ed block of inodes, the system could crash with the filesystem in
an unrecoverable state.
Rather than adding a soft updates dependency to ensure that the new
inode block is written before it is claimed by the cylinder group
map, we just do a barrier write of the zero'ed inode block to ensure
that it will get written before the updated cylinder group map can
be written. This change should only slow down bulk loading of newly
created filesystems since that is the primary time that new inode
blocks need to be created.
Reported by: Robert Watson
Reviewed by: kib
Tested by: Peter Holm
function, implementing the sysctl vfs.ffs.set_bufoutput (not used in
the tree yet).
- The current directory vnode dereference is unsafe since fd_cdir
could be changed and unreferenced, lock the filedesc around and vref
the fd_cdir.
- The VTOI() conversion of the fd_cdir is unsafe without first
checking that the vnode is indeed from an FFS mount, otherwise
the code dereferences a random memory.
- The cdir could be reclaimed from under us, lock it around the
checks.
- The type of the fp vnode might be not a disk, or it might have
changed while the thread was in flight, check the type.
Reviewed and tested by: mckusick
MFC after: 2 weeks
has gone below zero after the blocks in its inode are freed is a
no-op which the compiler fails to warn about because of the use of
the DIP macro. Change the sanity check to compare the number of
blocks being freed against the value i_blocks. If the number of
blocks being freed exceeds i_blocks, just set i_blocks to zero.
Reported by: Pedro Giffuni (pfg@)
MFC after: 2 weeks
mount, which means that is must not be called while the snaplock is
owned. The vfs_write_resume(9) does call the function as the
VFS_SUSP_CLEAN() method, which is too early and falls into the region
still protected by snaplock.
Add yet another flag for the vfs_write_resume_flags() to avoid calling
suspension cleanup handler after the suspend is lifted, and use it in
the ffs_snapshot() call to vfs_write_resume.
Reported and tested by: pho
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks
the write start, by adding a variation of the vfs_write_resume(9)
which accepts flags.
Use the new function to prevent a deadlock between parallel suspension
and snapshotting a UFS mount. The ffs_snapshot() code performed
vfs_write_resume() followed by vn_start_write() while owning the
snaplock. If the suspension intervene between resume and
vn_start_write(), the deadlock occured after the suspending thread
tried to lock the snaplock, most typically during the write in the
ffs_copyonwrite().
Reported and tested by: Andreas Longwitz <longwitz@incore.de>
Reviewed by: mckusick
MFC after: 2 weeks
X-MFC-note: make the vfs_write_resume(9) function a macro after the MFC,
in HEAD
When kern_yield() was introduced with the possibility to specify
a new priority, the behaviour changed by not lowering priority at all
in the consumers, making the yielding mechanism highly ineffective for
high priority kthreads like bufdaemon, syncer, vlrudaemon, etc.
There are no evidences that consumers could bear with such change in
semantic and this situation could finally lead to bugs similar to the
ones fixed in r244240.
Re-specify userland pri for kthreads involved.
Tested by: pho
Reviewed by: kib, mdf
MFC after: 1 week
received granular locking) but the comment present in UFS has been
copied all over other filesystems code incorrectly for several times.
Removes comments that makes no sense now.
Reviewed by: kib
MFC after: 3 days
crash. When truncating a file that never made it to disk we use the
canceled allocation dependencies to hold the journal records until
the truncation completes. Previously allocdirect dependencies on
the id_bufwait list were not considered and their journal space
could expire before the bitmaps were written. Cancel them and attach
them to the freeblks as we do for other allocdirects.
- Add KTR traces that were used to debug this problem.
- When adding jsegdeps, always use jwork_insert() so we don't have more
than one segdep on a given jwork list.
Sponsored by: EMC / Isilon Storage Division
When a background copy of a cg is written we complete any work associated
with that bmsafemap. If new work has been added to the non-background
copy of the buffer it will be completed before the next write happens.
The solution is to do the rollbacks when we make the copy so only those
dependencies that were present at the time of writing will be completed
when the background write completes. This would've resulted in various
bitmap related corruptions and panics. It also would've expired journal
entries early causing journal replay to miss some records.
MFC after: 2 weeks
solve power loss problems with dishonest write caches. However, it
should improve the situation and force a full fsck when it is unable
to resolve with the journal.
- Resolve a case where the journal could wrap in an unsafe way causing
us to prematurely lose journal entries in very specific scenarios.
Discussed with: mckusick
MFC after: 1 month
(implemented by ffs_reallocblks_ufs[12]) relocates the file's blocks
so as to cluster them together into a contiguous set of blocks on
the disk.
When the cluster crosses the boundary into the first indirect block,
the first indirect block is initially allocated in a position
immediately following the last direct block. Block reallocation
would usually destroy locality by moving the indirect block out of
the way to keep the data blocks contiguous. This change compensates
for this problem by noting that the first indirect block should be
left immediately following the last direct block. It then tries
to start a new cluster of contiguous blocks (referenced by the
indirect block) immediately following the indirect block.
We should also do this for other indirect block boundaries, but it
is only important for the first one.
Suggested by: Bruce Evans
MFC: 2 weeks
the previous diradd had already finished it could have been reclaimed
already. This would only happen under heavy dependency pressure.
Reported by: Andrey Zonov <zont@FreeBSD.org>
Discussed with: mckusick
MFC after: 1 week
extended using growfs(8). The problem here is that geom_label checks if
the filesystem size recorded in UFS superblock is equal to the provider
(i.e. device) size. This check cannot be removed due to backward
compatibility. On the other hand, in most cases growfs(8) cannot set
fs_size in the superblock to match the provider size, because, differently
from newfs(8), it cannot recompute cylinder group sizes.
To fix this problem, add another superblock field, fs_providersize, used
only for this purpose. The geom_label(4) will attach if either fs_size
(filesystem created with newfs(8)) or fs_providersize (filesystem expanded
using growfs(8)) matches the device size.
PR: kern/165962
Reviewed by: mckusick
Sponsored by: FreeBSD Foundation
with softupdates went away. Note that this does not fix the problem
entirely; I'm committing it now to make it easier for someone to pick
up the work.
Reviewed by: mckusick
In particular, do not lock Giant conditionally when calling into the
filesystem module, remove the VFS_LOCK_GIANT() and related
macros. Stop handling buffers belonging to non-mpsafe filesystems.
The VFS_VERSION is bumped to indicate the interface change which does
not result in the interface signatures changes.
Conducted and reviewed by: attilio
Tested by: pho
to pull vm_param.h was removed. Other big dependency of vm_page.h on
vm_param.h are PA_LOCK* definitions, which are only needed for
in-kernel code, because modules use KBI-safe functions to lock the
pages.
Stop including vm_param.h into vm_page.h. Include vm_param.h
explicitely for the kernel code which needs it.
Suggested and reviewed by: alc
MFC after: 2 weeks
now fully encapsulates all accesses to f_offset, and extends f_offset
locking to other consumers that need it, in particular, to lseek() and
variants of getdirentries().
Ensure that on 32bit architectures f_offset, which is 64bit quantity,
always read and written under the mtxpool protection. This fixes
apparently easy to trigger race when parallel lseek()s or lseek() and
read/write could destroy file offset.
The already broken ABI emulations, including iBCS and SysV, are not
converted (yet).
Tested by: pho
No objections from: jhb
MFC after: 3 weeks
is performed before exact size of the buffer is calculated, but the
buffer cannot have size greater then the total space allocated for
extended attributes. The existing check is executing with precise
size, but it is too late, since buffer needs to be allocated in
advance.
Also, adapt to uio_resid being of ssize_t type. Use lblktosize instead of
multiplying by fs block size by hand as well.
Reported and tested by: pho
MFC after: 1 week
and bmsafemap dependency structures in inodedep_lookup() and
bmsafemap_lookup() respectively. The setup of these structures must
be done while holding the soft-dependency mutex. If the inodedep is
allocated first, it may be freed in the I/O completion callback when
the mutex is released to allocate the bmsafemap. If the bmsafemap is
allocated first, it may be freed in the I/O completion callback when
the mutex is released to allocate the inodedep.
To resolve this problem, bmsafemap_lookup has had a parameter added
that allows a pre-malloc'ed bmsafemap to be passed in so that it does
not need to release the mutex to create a new bmsafemap. The
softdep_setup_inomapdep() routine pre-malloc's a bmsafemap dependency
before acquiring the mutex and starting to build the inodedep with a
call to inodedep_lookup(). The subsequent call to bmsafemap_lookup()
is passed this pre-allocated bmsafemap entry so that it need not
release the mutex if it needs to create a new one.
Reported by: Peter Holm
Tested by: Peter Holm
MFC after: 1 week
over just the active vnodes associated with a mount point to replace
MNT_VNODE_FOREACH_ALL in the vfs_msync, ffs_sync_lazy, and qsync
routines.
The vfs_msync routine is run every 30 seconds for every writably
mounted filesystem. It ensures that any files mmap'ed from the
filesystem with modified pages have those pages queued to be
written back to the file from which they are mapped.
The ffs_lazy_sync and qsync routines are run every 30 seconds for
every writably mounted UFS/FFS filesystem. The ffs_lazy_sync routine
ensures that any files that have been accessed in the previous
30 seconds have had their access times queued for updating in the
filesystem. The qsync routine ensures that any files with modified
quotas have those quotas queued to be written back to their
associated quota file.
In a system configured with 250,000 vnodes, less than 1000 are
typically active at any point in time. Prior to this change all
250,000 vnodes would be locked and inspected twice every minute
by the syncer. For UFS/FFS filesystems they would be locked and
inspected six times every minute (twice by each of these three
routines since each of these routines does its own pass over the
vnodes associated with a mount point). With this change the syncer
now locks and inspects only the tiny set of vnodes that are active.
Reviewed by: kib
Tested by: Peter Holm
MFC after: 2 weeks
The primary changes are that the user of the interface no longer
needs to manage the mount-mutex locking and that the vnode that
is returned has its mutex locked (thus avoiding the need to check
to see if its is DOOMED or other possible end of life senarios).
To minimize compatibility issues for third-party developers, the
old MNT_VNODE_FOREACH interface will remain available so that this
change can be MFC'ed to 9. Following the MFC to 9, MNT_VNODE_FOREACH
will be removed in head.
The reason for this update is to prepare for the addition of the
MNT_VNODE_FOREACH_ACTIVE interface that will loop over just the
active vnodes associated with a mount point (typically less than
1% of the vnodes associated with the mount point).
Reviewed by: kib
Tested by: Peter Holm
MFC after: 2 weeks
static and declare its prototype in sys/vnode.h) so that it can be
called from process_deferred_inactive() (in ufs/ffs/ffs_snapshot.c)
instead of the body of vinactive() being cut and pasted into
process_deferred_inactive().
Reviewed by: kib
MFC after: 2 weeks
and it is no longer referenced by a user process. The inode for a
file whose name has been removed, but is still referenced at the
time of a crash will still be allocated in the filesystem, but will
have no references (e.g., they will have no names referencing them
from any directory).
With traditional soft updates these unreferenced inodes will be
found and reclaimed when the background fsck is run. When using
journaled soft updates, the kernel must keep track of these inodes
so that it can find and reclaim them during the cleanup process.
Their existence cannot be stored in the journal as the journal only
handles short-term events, and they may persist for days. So, they
are tracked by keeping them in a linked list whose head pointer is
stored in the superblock. The journal tracks them only until their
linked list pointers have been commited to disk. Part of the cleanup
process involves traversing the list of unreferenced inodes and
reclaiming them.
This bug was triggered when confusion arose in the commit steps
of keeping the unreferenced-inode linked list coherent on disk.
Notably, a race between the link() system call adding a link-count
to a file and the unlink() system call removing a link-count to
the file. Here if the unlink() ran after link() had looked up
the file but before link() had incremented the link-count of the
file, the file's link-count would drop to zero before the link()
incremented it back up to one. If the file was referenced by a
user process, the first transition through zero made it appear
that it should be added to the unreferenced-inode list when in
fact it should not have been added. If the new name created by
link() was deleted within a few seconds (with the file still
referenced by a user process) it would legitimately be a candidate
for addition to the unreferenced-inode list. The result was that
there were two attempts to add the same inode to the unreferenced-inode
list which scrambled the unreferenced-inode list's pointers leading
to a panic. The fix is to detect and avoid the false attempt at
adding it to the unreferenced-inode list by having the link()
system call check to see if the link count is zero before it
increments it. If it is, the link() fails with ENOENT (showing that
it has failed the link()/unlink() race).
While tracking down this bug, we have added additional assertions
to detect the problem sooner and also simplified some of the code.
Reported by: Kirk Russell
Fix submitted by: Jeff Roberson
Tested by: Peter Holm
PR: kern/159971
MFC (to 9 only): 2 weeks
While we have a snapshot vnode unlocked to avoid a deadlock with another
inode in the same inode block being updated, the filesystem containing
it may be forcibly unmounted. When that happens the snapshot vnode is
revoked. We need to check for that condition and fail appropriately.
This change will be included along with 232351 when it is MFC'ed to 9.
Spotted by: kib
Reviewed by: kib
to enable the collection of counts of synchronous and asynchronous
reads and writes for its associated filesystem. The counts are
displayed using `mount -v'.
Ensure that buffers used for paging indicate the vnode from
which they are operating so that counts of paging I/O operations
from the filesystem are collected.
This checkin only adds the setting of the mount point for the
UFS/FFS filesystem, but it would be trivial to add the setting
and clearing of the mount point at filesystem mount/unmount
time for other filesystems too.
Reviewed by: kib
While there, make some style adjustments, like missed () around
return values.
Submitted by: bde
Reviewed by: mckusick
Tested by: pho
MFC after: 2 weeks
The bawrite() schedules the write to happen immediately, and its use
frees the current thread to do more cleanups.
Submitted by: bde
Reviewed by: mckusick
Tested by: pho
MFC after: 2 weeks
Synchronous inode block update is not needed for MNT_LAZY callers (syncer),
and since waitfor values are not zero, code did unneccessary synchronous
update.
Submitted by: bde
Reviewed by: mckusick
Tested by: pho
MFC after: 2 weeks
snapshots on UFS filesystems running with journaled soft updates.
This is the first of several bugs that need to be fixed before
removing the restriction added in -r230250 to prevent the use
of snapshots on filesystems running with journaled soft updates.
The deadlock occurs when holding the snapshot lock (snaplk)
and then trying to flush an inode via ffs_update(). We become
blocked by another process trying to flush a different inode
contained in the same inode block that we need. It holds the
inode block for which we are waiting locked. When it tries to
write the inode block, it gets blocked waiting for the our
snaplk when it calls ffs_copyonwrite() to see if the inode
block needs to be copied in our snapshot.
The most obvious place that this deadlock arises is in the
ffs_copyonwrite() routine when it updates critical metadata
in a snapshot and tries to write it out before proceeding.
The fix here is to write the data and indirect block pointer
for the snapshot, but to skip the call to ffs_update() to
write the snapshot inode. To ensure that we will never have
to update a pointer in the inode itself, the ffs_snapshot()
routine that creates the snapshot has to ensure that all the
direct blocks are allocated as part of the creation of the
snapshot.
A less obvious place that this deadlock occurs is when we hold
the snaplk because we are deleting a snapshot. In the course of
doing the deletion, we need to allocate various soft update
dependency structures and allocate some journal space. If we
hit a resource limit while doing this we decrease the resources
in use by flushing out an existing dirty file to get it to give
up the soft dependency resources that it holds. The flush can
cause an ffs_update() to be done on the inode for the file that
we have selected to flush resulting in the same deadlock as
described above when the inode that we have chosen to flush
resides in the same inode block as the snapshot inode that we hold.
The fix is to defer cleaning up any time that the inode on which
we are operating is a snapshot.
Help and review by: Jeff Roberson
Tested by: Peter Holm
MFC (to 9 only) after: 2 weeks
Add the sysctl debug.iosize_max_clamp, enabled by default. Setting the
sysctl to zero allows to perform the SSIZE_MAX-sized i/o requests from
the usermode.
Discussed with: bde, das (previous versions)
MFC after: 1 month
the kernel allocated a buffer but did not zero it as it was about
to be completely filled by a uiomove() from the user's buffer.
However, if the uiomove() failed, the old contents of the buffer
could be exposed especially if the file was being mmap'ed. The
fix was to always zero the buffer when it was allocated.
This change first attempts the uiomove() to the newly allocated
(and dirty) buffer and only zeros it if the uiomove() fails. The
effect is to eliminate the gratuitous zeroing of the buffer in
the usual case where the uiomove() successfully fills it.
Reviewed by: kib
Tested by: scottl
MFC after: 2 weeks (to 9 only)
every 30 seconds. This spike in I/O caused the system to pause every
30 seconds which was quite annoying. So, the way that sync worked
was changed so that when a vnode was first dirtied, it was put on
a 30-second cleaning queue (see the syncer_workitem_pending queues
in kern/vfs_subr.c). If the file has not been written or deleted
after 30 seconds, the syncer pushes it out. As the syncer runs once
per second, dirty files are trickled out slowly over the 30-second
period instead of all at once by a call to sync(2).
The one drawback to this is that it does not cover the filesystem
metadata. To handle the metadata, vfs_allocate_syncvnode() is called
to create a "filesystem syncer vnode" at mount time which cycles
around the cleaning queue being sync'ed every 30 seconds. In the
original design, the only things it would sync for UFS were the
filesystem metadata: inode blocks, cylinder group bitmaps, and the
superblock (e.g., by VOP_FSYNC'ing devvp, the device vnode from
which the filesystem is mounted).
Somewhere in its path to integration with FreeBSD the flushing of
the filesystem syncer vnode got changed to sync every vnode associated
with the filesystem. The result of this change is to return to the
old filesystem-wide flush every 30-seconds behavior and makes the
whole 30-second delay per vnode useless.
This change goes back to the originally intended trickle out sync
behavior. Key to ensuring that all the intended semantics are
preserved (e.g., that all inode updates get flushed within a bounded
period of time) is that all inode modifications get pushed to their
corresponding inode blocks so that the metadata flush by the
filesystem syncer vnode gets them to the disk in a timely way.
Thanks to Konstantin Belousov (kib@) for doing the audit and commit
-r231122 which ensures that all of these updates are being made.
Reviewed by: kib
Tested by: scottl
MFC after: 2 weeks
list. If softdep_sync_buf() discovers such dependency, it should do
nothing, which is safe as it is only waiting on the parent buffer to
be written, so it can be removed.
Committed on behalf of: jeff
MFC after: 1 week
filesystem running with journaled soft updates. Until these problems
have been tracked down, return ENOTSUPP when an attempt is made to
take a snapshot on a filesystem running with journaled soft updates.
MFC after: 2 weeks
vfs_mount_error error message facility provided by the nmount
interface.
Clean up formatting of mount warnings which still need to use
kernel printf's since they do not return errors.
Requested by: Craig Rodrigues <rodrigc@crodrigues.org>
MFC after: 2 weeks
While there, remove a useless check from the code. memcchr() always
returns characters unequal to 0xff in this case, so inosused[i] ^ 0xff
can never be equal to zero. Also, the fact that memcchr() returns a
pointer instead of the number of bytes until the end, makes conversion
to an offset far more easy.
The SYSCTL_NODE macro defines a list that stores all child-elements of
that node. If there's no SYSCTL_DECL macro anywhere else, there's no
reason why it shouldn't be static.
whle tracking down the system hang reported in kern/160662 and
corrected in revision 225806. The LOR is not the cause of the system
hang and indeed cannot cause an actual deadlock. However, it can
be easily eliminated by defering the acquisition of a buflock until
after all the vnode locks have been acquired.
Reported by: Hans Ottevanger
PR: kern/160662
Remove mapped pages for all dataset vnodes in zfs_rezget() using
new vn_pages_remove() to fix mmapped files changed by
zfs rollback or zfs receive -F.
PR: kern/160035, kern/156933
Reviewed by: kib, pjd
Approved by: re (kib)
MFC after: 1 week
(1) opt_capsicum.h is no longer required in ffs_alloc.c, so remove the
#include.
(2) portalfs depends on opt_capsicum.h, so have the Makefile generate one
if required.
These affect only modules built without a kernel (i.e, not buildkernel,
but yes buildworld if the dubious MODULES_WITH_WORLD is used).
Approved by: re (bz)
Sponsored by: Google Inc
kernel for FreeBSD 9.0:
Add a new capability mask argument to fget(9) and friends, allowing system
call code to declare what capabilities are required when an integer file
descriptor is converted into an in-kernel struct file *. With options
CAPABILITIES compiled into the kernel, this enforces capability
protection; without, this change is effectively a no-op.
Some cases require special handling, such as mmap(2), which must preserve
information about the maximum rights at the time of mapping in the memory
map so that they can later be enforced in mprotect(2) -- this is done by
narrowing the rights in the existing max_protection field used for similar
purposes with file permissions.
In namei(9), we assert that the code is not reached from within capability
mode, as we're not yet ready to enforce namespace capabilities there.
This will follow in a later commit.
Update two capability names: CAP_EVENT and CAP_KEVENT become
CAP_POST_KEVENT and CAP_POLL_KEVENT to more accurately indicate what they
represent.
Approved by: re (bz)
Submitted by: jonathan
Sponsored by: Google Inc
so that it is visible to userland programs. This change enables
the `mount' command with no arguments to be able to show if a
filesystem is mounted using journaled soft updates as opposed
to just normal soft updates.
Approved by: re (bz)
(typically fsck_ffs) to register that it wishes to use FFS specific
sysctl's to update the filesystem. This ensures that two checkers
cannot run on a given filesystem at the same time and that no other
process accidentally or maliciously uses the filesystem updating
sysctls inappropriately. This functionality is needed by the
journaling soft-updates recovery code.
flag (FS_SUJ) when determining whether to do journaling-based
operations. The mount flag is set only when journaling is active
while the superblock flag is set to indicate that journaling is to
be used. For example, when the filesystem is mounted read-only, the
journaling may be present (FS_SUJ) but not active (MNTK_SUJ).
Inappropriate checking of the FS_SUJ flag was causing some
journaling actions to be attempted at inappropriate times.
filesystems to be opened for writing. This functionality used to
be special-cased for just the root filesystem, but with this change
is now available for all UFS filesystems. This change is needed for
journaled soft updates recovery.
Discussed with: Jeff Roberson
This will most likely cause new block allocations which can recurse
into request cleanup.
- While here optimize the ufs locking slightly. We need only acquire and
drop once.
- process_removes() and process_truncates() also is only needed once.
- Attempt to flush each item on the worklist once but do not loop forever
if some can not be completed.
Discussed with: mckusick
option to vm_object_page_remove() asserts that the specified range of pages
is not mapped, or more precisely that none of these pages have any managed
mappings. Thus, vm_object_page_remove() need not call pmap_remove_all() on
the pages.
This change not only saves time by eliminating pointless calls to
pmap_remove_all(), but it also eliminates an inconsistency in the use of
pmap_remove_all() versus related functions, like pmap_remove_write(). It
eliminates harmless but pointless calls to pmap_remove_all() that were being
performed on PG_UNMANAGED pages.
Update all of the existing assertions on pmap_remove_all() to reflect this
change.
Reviewed by: kib
and usr.sbin/makefs/ffs/ffs_subr.c as they have no need of anything in that
file. No other programs or libraries include <ufs/ffs/ffs_extern.h> (nor
should they as it is totally in-kernel interfaces). For added protection
I enclosed the entire contents of <ufs/ffs/ffs_extern.h> in ifdef _KERNEL.
Feedback from: Bruce Evans and Tai-hwa Liang
messages for a filesystem being out of space need to be moved so that
they do not print out until after a failed cleanup attempt.
Suggested by: Jeff Roberson
to resolve errors which can cause corruption on recovery with the old
synchronous mechanism.
- Append partial truncation freework structures to indirdeps while
truncation is proceeding. These prevent new block pointers from
becoming valid until truncation completes and serialize truncations.
- On completion of a partial truncate journal work waits for zeroed
pointers to hit indirects.
- softdep_journal_freeblocks() handles last frag allocation and last
block zeroing.
- vtruncbuf/ffs_page_remove moved into softdep_*_freeblocks() so it
is only implemented in one place.
- Block allocation failure handling moved up one level so it does not
proceed with buf locks held. This permits us to do more extensive
reclaims when filesystem space is exhausted.
- softdep_sync_metadata() is broken into two parts, the first executes
once at the start of ffs_syncvnode() and flushes truncations and
inode dependencies. The second is called on each locked buf. This
eliminates excessive looping and rollbacks.
- Improve the mechanism in process_worklist_item() that handles
acquiring vnode locks for handle_workitem_remove() so that it works
more generally and does not loop excessively over the same worklist
items on each call.
- Don't corrupt directories by zeroing the tail in fsck. This is only
done for regular files.
- Push a fsync complete record for files that need it so the checker
knows a truncation in the journal is no longer valid.
Discussed with: mckusick, kib (ffs_pages_remove and ffs_truncate parts)
Tested by: pho
Eliminate one (of several) possible conflicting buffer locks when
trying to reclaim blocks. Rest of fix to be incorporated as part
of SUJ update by jeff.
Pointed out by: Kostik Belousov
method, so that callers can indicate the minimum vnode
locking requirement. This will allow some file systems to choose
to return a LK_SHARED locked vnode when LK_SHARED is specified
for the flags argument. This patch only adds the flag. It
does not change any file system to use it and all callers
specify LK_EXCLUSIVE, so file system semantics are not changed.
Reviewed by: kib
should not change. Fetch the td_user_pri under the thread lock. This
is probably not necessary but a magic number also seems preferable to
knowing the implementation details here.
Requested by: Jason Behmer < jason DOT behmer AT isilon DOT com >
goes to zero. E.g., the vnode might be only shared-locked at the time of
vput() call. Such vnodes are kept in the hash, so they can be found later.
If ffs_valloc() allocated an inode that has its vnode cached in hash, and
still owing the inactivation, then vget() call from ffs_valloc() clears
VI_OWEINACT, and then the vnode is reused for the newly allocated inode.
The problem is, the vnode is not reclaimed before it is put to the new
use. ffs_valloc() recycles vnode vm object, but this is not enough.
In particular, at least v_vflag should be cleared, and several bits of
UFS state need to be removed.
It is very inconvenient to call vgone() at this point. Instead, move
some parts of ufs_reclaim() into helper function ufs_prepare_reclaim(),
and call the helper from VOP_RECLAIM and ffs_valloc().
Reviewed by: mckusick
Tested by: pho
MFC after: 3 weeks
for a new journal specific partial truncate routine.
- Use dep_current[] in place of specific dependency counts. This is
automatically maintained when workitems are allocated and has
less risk of becoming incorrect.
- Keep a hash of indirect blocks that have recently been freed and are
still referenced in the journal.
- Lookup blocks in this hash before forcing a new block write to wait on
the journal entry to hit the disk. This is only necessary to avoid
confusion between old identities as indirects and new identities as
file blocks.
- Don't free jseg structures until the journal has written a record that
invalidates it. This keeps the indirect block information around for
as long as is required to be safe.
- Force an empty journal block write when required to flush out stale
journal data that is simply waiting for the oldest valid sequence
number to advance beyond it.
will be removed. Permit the journal to proceed so that we don't leave
a rollback in a cg for a very long time as this can cause terrible perf
problems in low memory situations.
Tested by: pho
up and declaring a filesystem out of space. Especially necessary when
running on a small filesystem. With this improvement, it should be
possible to use soft updates on a small root filesystem.
Kudos to: Peter Holm
Testing by: Peter Holm
MFC: 2 weeks
- In softdep_revert_mkdir() find the dotaddref before we attempt to cancel
the jaddref so we can make assumptions about where the dotaddref is on
the list. cancel_jaddref() does not always remove items from the list
anymore.
- Always set GOINGAWAY on an inode in softdep_freefile() if DEPCOMPLETE
was never set. This ensures that dependencies will continue to be
processed on the inowait/bufwait list and is more an artifact of
the structure of the code than a pure ordering problem.
- Always set DEPCOMPLETE on canceled jaddrefs so that they can be freed
appropriately. This normally occurs when the refs are added to the
journal but if they are canceled before this point the state would
never be set and the dependency could never be freed.
Reported by: pho
Tested by: pho
journal blocks, instead of hard coding 512 byte sector size. Journal need
to atomically write the block, that can only be guaranteed at the device
sector size, not larger. Attempt to write less then sector size results in
driver errors.
Note that this is the first structure in UFS that depends on the
sector size. Other elements are written in the units of fragments.
In collaboration with: pho
Reviewed by: jeff
Tested by: bz, pho
SU+J is not included as a FEATURE macro:
- it was not in the tree during the GSoC
- I do not see an option to en-/disable it in NOTES
Two minor changes where made during the review compared to what was developed
during GSoC 2010.
No FreeBSD version bump, the userland application to query the features will
be committed last and can serve as an indication of the availablility if
needed.
Sponsored by: Google Summer of Code 2010
Submitted by: kibab
Reviewed by: kib
X-MFC after: to be determined in last commit with code from this project
