debug.hashstat.rawnchash sysctl in particular as taking 7 milliseconds on
a 3GHz Intel Xeon (4x2) running 7.1. It accounted for almost a quarter of
the total runtime of 'sysctl -a'. It also performs lots of copyout's while
holding the namecache lock (this does not attempt to fix that).
MFC after: 2 weeks
stale entries, we save a copy of the directory's modification time when
the first negative cache entry was added in the directory's NFS node.
When a negative cache entry is hit during a pathname lookup, the parent
directory's modification time is checked. If it has changed, all of the
negative cache entries for that parent are purged and the lookup falls
back to using the RPC. This required adding a new cache_purge_negative()
method to the name cache to purge only negative cache entries for a given
directory.
Submitted by: mohans, Rick Macklem, Ricardo Labiaga @ NetApp
Reviewed by: mohans
mutex to a reader/writer lock. Lookup operations first grab a read lock and
perform the lookup. If the operation results in a need to modify the cache,
then it tries to do an upgrade. If that fails, it drops the read lock,
obtains a write lock, and redoes the lookup.
inside the SYSCTL() macros and thus does not need to be done for
all of the nodes scattered across the source tree.
- Mark the name-cache related sysctl's (including debug.hashstat.*) MPSAFE.
- Mark vm.loadavg MPSAFE.
- Remove GIANT_REQUIRED from vmtotal() (everything in this routine already
has sufficient locking) and mark vm.vmtotal MPSAFE.
- Mark the vm.stats.(sys|vm).* sysctls MPSAFE.
In normal operation, the number of cache entries is roughly equal to the
number of active vnodes. However, when most of the recently accessed
vnodes have many hard links, the number of cache entries can be 32000
times as large, exhausting kernel memory and provoking a panic in
kmem_malloc().
MFC after: 2 weeks
did not compared nc_dvp with supplied parent directory vnode pointer.
Add the check and note that now branches for vp != NULL and vp == NULL
are the same, thus can be merged.
Reported and reviewed by: kan
Tested by: pho
MFC after: 2 weeks
looked up would have v_dd set to a non-NULL value. This fixes a panic
seen when running installworld on a diskless system with a separate /usr
file system.
Submitted by: cracauer
Approved by: kib
on a best-effort basis. Teach vn_fullpath to use this new VOP if a
regular VFS cache lookup fails. This VOP is designed to supplement the
VFS cache to provide a better chance that a vnode-to-name lookup will
succeed.
Currently, an implementation for devfs is being committed. The default
implementation is to return ENOENT.
A big thanks to kib for the mentorship on this, and to pho for running it
through his stress test suite.
Reviewed by: arch
Approved by: kib
entries for one name. Then, creating inode with that name would remove
one entry, leaving others dormant. Reclaiming the vnode would uncover
negative entries, causing false return of ENOENT from the calls like
stat, that do not create inode.
Prevent creation of the duplicated negative entries.
Reported and debugged with: pho
Reviewed by: jhb
X-MFC: after shared lookup changes
advance of teaching vn_fullpath1() how to query file systems for
vnode-to-name mappings when cache lookups fail.
Thanks to kib for guidance and patience on this process.
Reviewed by: kib
Approved by: kib
unmounts. When we upgrade a vnode lock from shared to exclusive during
a name cache lookup, fail the lookup with EBADF if the vnode is invalidated
while we are waiting for the exclusive lock.
Also, for correctness (though I'm not sure it can occur in practice),
downgrade an exclusively locked vnode if it should be share locked.
Tested by: pho
not in the namecache when shared lookups are enabled (vfs.lookup_shared=1,
it is currently off by default) and the filesystem supports shared lookups
(e.g. NFS client). Specifically, if multiple concurrent LOOKUPs both miss
in the name cache in parallel, each of the lookups may each end up adding an
entry to the namecache resulting in duplicate entries in the namecache
for the same pathname. A subsequent removal of the mapping of that
pathname to that vnode (via remove or rename) would only evict one of the
entries from the name cache. As a result, subseqent lookups for that
pathname would still return the old vnode.
This race was observed with shared lookups over NFS where a file was updated
by writing a new file out to a temporary file name and then renaming that
temporary file to the "real" file to effect atomic updates of a file. Other
processes on the same client that were periodically reading the file would
occasionally receive an ESTALE error from open(2) because the VOP_GETATTR()
in nfs_open() would receive that error when given the stale vnode.
The fix here is to check for duplicates in cache_enter() and just return
if an entry for this same directory and leaf file name for this vnode is
already in the cache. The check for duplicates is done by walking the
per-vnode list of name cache entries. It is expected that this list should
be very small in the common case (usually 0 or 1 entries during a
cache_enter() since most files only have 1 "leaf" name).
Reviewed by: ups, scottl
MFC after: 2 months
processes are not producing absolute pathname tokens. It is required
that audited pathnames are generated relative to the global root mount
point. This modification changes our implementation of audit_canon_path(9)
and introduces a new function: vn_fullpath_global(9) which performs a
vnode -> pathname translation relative to the global mount point based
on the contents of the name cache. Much like vn_fullpath,
vn_fullpath_global is a wrapper function which called vn_fullpath1.
Further, the string parsing routines have been converted to use the
sbuf(9) framework. This change also removes the conditional acquisition
of Giant, since the vn_fullpath1 method will not dip into file system
dependent code.
The vnode locking was modified to use vhold()/vdrop() instead the vref()
and vrele(). This will modify the hold count instead of modifying the
user count. This makes more sense since it's the kernel that requires
the reference to the vnode. This also makes sure that the vnode does not
get recycled we hold the reference to it. [1]
Discussed with: rwatson
Reviewed by: kib [1]
MFC after: 2 weeks
no longer needed, but for now we still want to be consistent with other
similar checks in the tree.
- Call ASSERT_VOP_ELOCKED() only when vget() returns 0.
Reviewed by: jeff
after each SYSINIT() macro invocation. This makes a number of
lightweight C parsers much happier with the FreeBSD kernel
source, including cflow's prcc and lxr.
MFC after: 1 month
Discussed with: imp, rink
always curthread.
As KPI gets broken by this patch, manpages and __FreeBSD_version will be
updated by further commits.
Tested by: Andrea Barberio <insomniac at slackware dot it>
conjuction with 'thread' argument passing which is always curthread.
Remove the unuseful extra-argument and pass explicitly curthread to lower
layer functions, when necessary.
KPI results broken by this change, which should affect several ports, so
version bumping and manpage update will be further committed.
Tested by: kris, pho, Diego Sardina <siarodx at gmail dot com>
Remove this argument and pass curthread directly to underlying
VOP_LOCK1() VFS method. This modify makes the code cleaner and in
particular remove an annoying dependence helping next lockmgr() cleanup.
KPI results, obviously, changed.
Manpage and FreeBSD_version will be updated through further commits.
As a side note, would be valuable to say that next commits will address
a similar cleanup about VFS methods, in particular vop_lock1 and
vop_unlock.
Tested by: Diego Sardina <siarodx at gmail dot com>,
Andrea Di Pasquale <whyx dot it at gmail dot com>
parent vnode and relock it after locking child vnode. The problem was that
we always relock it exclusively, even when it was share-locked.
Discussed with: jeff
and flags with an sxlock. This leads to a significant and measurable
performance improvement as a result of access to shared locking for
frequent lookup operations, reduced general overhead, and reduced overhead
in the event of contention. All of these are imported for threaded
applications where simultaneous access to a shared file descriptor array
occurs frequently. Kris has reported 2x-4x transaction rate improvements
on 8-core MySQL benchmarks; smaller improvements can be expected for many
workloads as a result of reduced overhead.
- Generally eliminate the distinction between "fast" and regular
acquisisition of the filedesc lock; the plan is that they will now all
be fast. Change all locking instances to either shared or exclusive
locks.
- Correct a bug (pointed out by kib) in fdfree() where previously msleep()
was called without the mutex held; sx_sleep() is now always called with
the sxlock held exclusively.
- Universally hold the struct file lock over changes to struct file,
rather than the filedesc lock or no lock. Always update the f_ops
field last. A further memory barrier is required here in the future
(discussed with jhb).
- Improve locking and reference management in linux_at(), which fails to
properly acquire vnode references before using vnode pointers. Annotate
improper use of vn_fullpath(), which will be replaced at a future date.
In fcntl(), we conservatively acquire an exclusive lock, even though in
some cases a shared lock may be sufficient, which should be revisited.
The dropping of the filedesc lock in fdgrowtable() is no longer required
as the sxlock can be held over the sleep operation; we should consider
removing that (pointed out by attilio).
Tested by: kris
Discussed with: jhb, kris, attilio, jeff
- Remove also "MP SAFE" after prior "MPSAFE" pass. (suggested by bde)
- Remove extra blank lines in some cases.
- Add extra blank lines in some cases.
- Remove no-op comments consisting solely of the function name, the word
"syscall", or the system call name.
- Add punctuation.
- Re-wrap some comments.
filesystem agnostic. We are not touching any file system specific functions
in this code path. Since we have a cache lock, there is really no need to
keep Giant around here.
This eliminates Giant acquisitions for any syscall which is auditing pathnames.
Discussed with: jeff
cache_zap() to clear the v_dd pointers when a directory vnode is forcibly
discarded. For this to work, all vnodes with v_dd pointers to a directory
must also have name cache entries linked via v_cache_dst to that dvp
otherwise we could not find them at cache_purge() time. The following
code snipit could break this guarantee by unlinking a directory before
fetching it's dotdot. The dotdot lookup would initialize the v_dd field
of the unlinked directory which could never be cleared. To fix this
we don't initialize v_dd for orphaned vnodes.
printf("rmdir: %d\n", rmdir("../foo")); /* foo is cwd */
printf("chdir: %d\n", chdir(".."));
printf("%s\n", getwd(NULL));
Sponsored by: Isilon Systems, Inc.
Discovered by: kkenn
Approved by: re (blanket vfs)
vnodes whose names it caches, so we no longer need a `generation
number' to tell us if a referenced vnode is invalid. Replace the use
of the parent's v_id in the hash function with the address of the
parent vnode.
Tested by: Peter Holm
Glanced at by: jeff, phk
except for places where people forget to update one of them. We now
collect only one set of stats for both of these routines. Other
changes in this commit include:
- Start acquiring Giant again in vn_fullpath(), since it is required
when crossing a mount point.
- Expand the scope of the cache lock to avoid dropping it and
picking it up again for every pathname component. This also
makes it trivial to avoid races in stats collection.
- Assert that nc_dvp == v_dd for directories instead of returning
an error to userland when this is not true. AFAIK, it should
always be true when v_dd is non-null.
- For vn_fullpath(), handle the first (non-directory) vnode
separately.
Glanced at by: jeff, phk
to cache_lookup(). This allows us to acquire the vnode interlock before
dropping the cache lock. This protects the vnodes identity until we
have locked it.
Sponsored by: Isilon Systems, Inc.
