Disable the normal reclaim path for the txg_sync thread. This
ensures the thread will never enter dmu_tx_assign() which can
otherwise occur due to direct reclaim. If this is allowed to
happen the system can deadlock. Direct reclaim call path:
->shrink_icache_memory->prune_icache->dispose_list->
clear_inode->zpl_clear_inode->zfs_inactive->dmu_tx_assign
Under OpenSolaris all memory reclaim is done asyncronously. Under
Linux memory reclaim is done asynchronously _and_ synchronously.
When a process allocates memory with GFP_KERNEL it explicitly allows
the kernel to do reclaim on its behalf to satify the allocation.
If that GFP_KERNEL allocation fails the kernel may take more drastic
measures to reclaim the memory such as killing user space processes.
This was observed to happen with ZFS because the ARC could consume
a large fraction of the system memory but no synchronous reclaim
could be performed on it. The result was GFP_KERNEL allocations
could fail resulting in OOM events, and only moments latter the
arc_reclaim thread would free unused memory from the ARC.
This change leaves the arc_thread in place to manage the fundamental
ARC behavior. But it adds a synchronous (direct) reclaim path for
the ARC which can be called when memory is badly needed. It also
adds an asynchronous (indirect) reclaim path which is called
much more frequently to prune the ARC slab caches.
The following useful values were missing the arcstats. This change
adds them in to provide greater visibility in to the arcs behavior.
arc_no_grow 4 0
arc_tempreserve 4 0
arc_loaned_bytes 4 0
arc_meta_used 4 624774592
arc_meta_limit 4 400785408
arc_meta_max 4 625594176
Under Linux a dentry referencing an inode must be instantiated before
the inode is unlocked. To accomplish this without overly modifing
the core ZFS code the dentry it passed via the vattr_t. There are
cases such as replay when a dentry is not available. In which case
it is obviously not initialized at inode creation time, if a dentry
is needed it will be spliced as when required via d_lookup().
Provide the dnlc_reduce_cache() function which attempts to prune
cached entries from the dcache and icache. After the entries are
pruned any slabs which they may have been using are reaped.
Note the API takes a reclaim percentage but we don't have easy
access to the total number of cache entries to calculate the
reclaim count. However, in practice this doesn't need to be
exactly correct. We simply need to reclaim some useful fraction
(but not all) of the cache. The caller can determine if more
needs to be done.
By decreasing the number of target objects per slab we increase
the likelyhood that a slab can be freed. This reduces the level
of fragmentation in the slab which has been observed to be a
problem for certain workloads. The penalty for this is that we
also decrease the speed which need objects can be allocated.
One of the most common things you want to know when looking at
the slab is how much memory is being used. This information was
available in /proc/spl/kmem/slab but only on a per-slab basis.
This commit adds the following /proc/sys/kernel/spl/kmem/slab*
entries to make total slab usage easily available at a glance.
slab_kmem_total - Total kmem slab size
slab_kmem_avail - Alloc'd kmem slab size
slab_kmem_max - Max observed kmem slab size
slab_vmem_total - Total vmem slab size
slab_vmem_avail - Alloc'd vmem slab size
slab_vmem_max - Max observed vmem slab size
NOTE: The slab_*_max values are expected to over report because
they show maximum values since boot, not current values.
The 'slab_fail', 'slab_create', and 'slab_destroy' columns in the slab
output have been removed because they are virtually always zero and
not very useful.
The much more useful 'size' and 'alloc' columns have been added which
show the total slab size and how much of the total size has been
allocated to objects.
Finally, the formatting has been updated to be much more human
readable while still being friendly for tool like awk to parse.
The Linux shrinker has gone through three API changes since 2.6.22.
Rather than force every caller to understand all three APIs this
change consolidates the compatibility code in to the mm-compat.h
header. The caller then can then use a single spl provided
shrinker API which does the right thing for your kernel.
SPL_SHRINKER_CALLBACK_PROTO(shrinker_callback, cb, nr_to_scan, gfp_mask);
SPL_SHRINKER_DECLARE(shrinker_struct, shrinker_callback, seeks);
spl_register_shrinker(&shrinker_struct);
spl_unregister_shrinker(&&shrinker_struct);
spl_exec_shrinker(&shrinker_struct, nr_to_scan, gfp_mask);
While this extra structure memory does not exist under Solaris
it is needed under Linux to pass the dentry. This allows the
dentry to be easily instantiated before the inode is unlocked.
Making distclean in module
make[1]: Entering directory `/zfs/module'
make -C SUBDIRS=`pwd` clean
make: Entering an unknown directory
make: *** SUBDIRS=/zfs/module: No such file or directory. Stop.
When using --with-config=user the 'distclean' target would fail
because it assumes the kernel configuration infrastrure is set up.
This is not the case, nor does it need to be, because the
'--with-config=user' option will prune the entire ./module subtree
from SUBDIRS. This prevents most build rules from operating in the
./module directory.
However, the 'dist*' rules will still traverse this directory
because it is listed in DIST_SUBDIRS. This is correct because we
need to ensure the dist rules package the directory contents
regardless of the configuration for the 'dist' rule. The correct
way to handle this is to only invoke the kernel build system as
part of the 'clean' rule when CONFIG_KERNEL_TRUE is set.
Initial fix provided by Darik Horn <dajhorn@vanadac.com>.
This commit is a slightly refined form of the original.
Some udev hooks are not designed to be idempotent, so calling udevadm
trigger outside of the distribution's initialization scripts can have
unexpected (and potentially dangerous) side effects. For example, the
system time may change or devices may appear multiple times. See Ubuntu
launchpad bug 320200 and this mailing list post for more details:
https://lists.ubuntu.com/archives/ubuntu-devel/2009-January/027260.html
To avoid these problems we call udevadm trigger with --action=change
--subsystem-match=block. The first argument tells udev just to refresh
devices, and make sure everything's as it should be. The second
argument limits the scope to block devices, so devices belonging to
other subsystems cannot be affected.
This doesn't fix the problem on older udev implementations that don't
provide udevadm but instead have udevtrigger as a standalone program.
In this case the above options aren't available so there's no way to
call call udevtrigger safely. But we can live with that since this
issue only exists in optional test and helper scripts, and most
zfs-on-linux users are running newer systems anyways.
The CHAOS 5 kernels are now packaged identially to the RHEL6 kernels.
Therefore we can simply use the RHEL6 rules in the spec file when
building packages.
The CHAOS 5 kernels are now packaged identially to the RHEL6 kernels.
Therefore we can simply use the RHEL6 rules in the spec file when
building packages.
This build failure was accidentally introduced by previous commit
bfd214a which fixed the load average. Unfortunately, the wrapper
for cv_wait_interruptible was not available in the zfs_context.h
user compatibility code. I failed to notice this because I didn't
rebuild everything cleanly before committing.
undefined reference to `cv_wait_interruptible'
collect2: ld returned 1 exit status
Closes#181
Kernel threads which sleep uninterruptibly on Linux are marked in the (D)
state. These threads are usually in the process of performing IO and are
thus counted against the load average. The txg_quiesce and txg_sync threads
were always sleeping uninterruptibly and thus inflating the load average.
This change makes them sleep interruptibly. Some care is required however
because these threads may now be woken early by signals. In this case the
callers are all careful to check that the required conditions are met after
waking up. If we're woken early due to a signal they will simply go back
to sleep. In this case these changes are safe.
Closes#175
The dracut change caused an error during "make rpm". The cause
is simple, RHEL5 does not recognize the %{datarootdir} macro in
zfs.spec. It was changed to %{datadir} which fixes the build.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
The udev/rules.d scripts must use absolute paths to their support
binaries. However, where those binaries get installed depends
on what --prefix was set to when the package was configured.
This change makes the udev/rules.d helpers to *.in files which
are processed by configure. This allows them to be dynamically
updated to include the specified --prefix.
Additionally, this change updates 60-zvol.rules to handle both
the 'add' and 'change' actions. This ensures that that all
valid zvol devices are correctly linked.
This commit fixes issue on
https://github.com/behlendorf/zfs/issues/#issue/172
Changes:
- update BLKZNAME to use _IOR instead of _IO. Kernel 2.6.32 allows
read parameters (copy_to_user) with _IO, while newer kernels (tested
Archlinux's 2.6.37 kernel) enforces _IOR (which is correct)
- fix return code and message on error
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Solaris credentials don't have an fsuid/fsguid field but Linux
credentials do. To handle this case the Solaris API is being
modestly extended to include the crgetfsuid()/crgetfsgid()
helper functions.
Addititionally, because the crget*() helpers are implemented
identically regardless of HAVE_CRED_STRUCT they have been
moved outside the #ifdef to common code. This simplification
means we only have one version of the helper to keep to to date.
Certain stock kernels (Debian Lenny) are built with zlib_inflate.ko
as a kernel module. To ensure 'make check' works in-tree load this
module before loading the spl module. This is now required for the
zlib splat regression test.
The .freeze_fs/.unfreeze_fs hooks were not added until Linux 2.6.29
Since these hooks are currently unused they are being removed to
allow support of older kernels.
As of Linux 2.6.29 a clean credential API was added to the Linux kernel.
Previously the credential was embedded in the task_struct. Because the
SPL already has considerable support for handling this API change the
ZPL code has been updated to use the Solaris credential API.
Added insert_inode_locked() helper function, prior to this most callers
used insert_inode_hash(). The older method doesn't check for collisions
in the inode_hashtable but it still acceptible for use. Fallback to
using insert_inode_hash() when insert_inode_locked() is unavailable.
The blk_queue_stackable() queue flag was added in 2.6.27 to handle dm
stacking drivers. Prior to this request stacking drivers were detected
by checking (q->request_fn == NULL), for earlier kernels we revert to
this legacy behavior.
Older glibc <sys/mount.h> headers did not define all the available
umount2(2) flags. Both MNT_FORCE and MNT_DETACH are supported in the
kernel back to 2.4.11 so we define them correctly if they are missing.
Closes#95
Now that KM_SLEEP is not defined as GFP_NOFS there is the possibility
of synchronous reclaim deadlocks. These deadlocks never existed in the
original OpenSolaris code because all memory reclaim on Solaris is done
asyncronously. Linux does both synchronous (direct) and asynchronous
(indirect) reclaim.
This commit addresses a deadlock caused by inode eviction. A KM_SLEEP
allocation may trigger direct memory reclaim and shrink the inode cache.
This can occur while a mutex in the array of ZFS_OBJ_HOLD mutexes is
held. Through the ->shrink_icache_memory()->evict()->zfs_inactive()->
zfs_zinactive() call path the same mutex may be reacquired resulting
in a deadlock. To avoid this deadlock the process must not reacquire
the mutex when it is already holding it.
This is a reasonable fix for now but longer term the ZFS_OBJ_HOLD
mutex locking should be reevaluated. This infrastructure already
prevents us from ever using the Linux lock dependency analysis tools,
and it may limit scalability.
It used to be the case that all KM_SLEEP allocations were GFS_NOFS.
Unfortunately this often resulted in the kernel being unable to
reclaim the ARC, inode, and dentry caches in a timely manor.
The fix was to make KM_SLEEP a GFP_KERNEL allocation in the SPL.
However, this increases the posibility of deadlocking the system
on a zfs write thread. If a zfs write thread attempts to perform
an allocation it may trigger synchronous reclaim. This reclaim
may attempt to flush dirty data/inode to disk to free memory.
Unforunately, this write cannot finish because the write thread
which would handle it is holding the previous transaction open.
Deadlock.
To avoid this all allocations in the zfs write thread path must
use KM_PUSHPAGE which prohibits synchronous reclaim for that
thread. In this way forward progress in ensured. The risk
with this change is I missed updating an allocation for the
write threads leaving an increased posibility of deadlock. If
any deadlocks remain they will be unlikely but we'll have to
make sure they all get fixed.
As part of vmalloc() a __pte_alloc_kernel() allocation may occur. This
internal allocation does not honor the gfp flags passed to vmalloc().
This means even when vmalloc(GFP_NOFS) is called it is possible that a
synchronous reclaim will occur. This reclaim can trigger file IO which
can result in a deadlock. This issue can be avoided by explicitly
setting PF_MEMALLOC on the process to subvert synchronous reclaim when
vmalloc() is called with !__GFP_FS.
An example stack of the deadlock can be found here (1), along with the
upstream kernel bug (2), and the original bug discussion on the
linux-mm mailing list (3). This code can be properly autoconf'ed
when the upstream bug is fixed.
1) http://github.com/behlendorf/zfs/issues/labels/Vmalloc#issue/133
2) http://bugzilla.kernel.org/show_bug.cgi?id=30702
3) http://marc.info/?l=linux-mm&m=128942194520631&w=4
As originally described in commit 82b8c8fa64
this was done to prevent certain deadlocks from occuring in the system.
However, as suspected the price for doing this proved to be too high.
The VM is having a hard time effectively reclaiming memory thus we are
reverting this change.
However, we still need to fundamentally handle the issue. Under
Solaris the KM_PUSHPAGE mask is used commonly in I/O paths to ensure
a memory allocations will succeed. We leverage this fact and redefine
KM_PUSHPAGE to include GFP_NOFS. This ensures that in these common
I/O path we don't trigger additional reclaim. This minimizes the
change to the Solaris code.
Compiling with 'LDFLAGS=-Wl,--as-needed' exposed the fact that
there were some library linking problems introduced by mount_zfs.
In particular, the libzfs library does use nvpair symbols, and
mount_zfs contains no dependencies on libzpool.
Closes#161Closes#162
New versions glibc declare getcwd() with the warn_unused_result attribute.
This results in a warning because the updated mount helper was not
checking this return value. This issue was fixed by checking the return
type and in the case of an error simply returning the passed dataset.
One possible, but unlikely, error would be having your cwd directory
unlinked while the mount command was running.
cmd/mount_zfs/mount_zfs.c: In function ‘parse_dataset’:
cmd/mount_zfs/mount_zfs.c:223:2: error: ignoring return value of
‘getcwd’, declared with attribute warn_unused_result
To simplify the process of using zfs as your root filesystem a
zfs-drucat sub-package has been added. This sub-package adds a zfs
dracut module which allows your initramfs to be rebuilt with zfs
support. The process for doing this is still complicated but there
is clearly interest from the community about getting this working
well and documented. This should help lay some of the groundwork.
Longer term these changes should be pushed in the upstream dracut
package. Once that occurs this subpackage will no longer be
required for new systems, however we may want to conditionally
build this package in the future for systems running older
dracut versions.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
To support automatically mounting your zfs on filesystem on boot
a basic init script is needed. Unfortunately, every distribution
has their own idea of the _right_ way to do things. Rather than
write one very complicated portable init script, which would be
invariably replaced by the distributions own anyway. I have
instead added support to provide multiple distribution specific
init scripts.
The correct init script for your distribution will be selected
by ZFS_AC_DEFAULT_PACKAGE which will set DEFAULT_INIT_SCRIPT.
During 'make install' the correct script for your system will
be installed from zfs/etc/init.d/zfs.DEFAULT_INIT_SCRIPT to the
usual /etc/init.d/zfs location.
Currently, there is zfs.fedora and a more generic zfs.lsb init
script. Hopefully, the distribution maintainers who know best
how they want their init scripts to function will feedback their
approved versions to be included in the project.
This change does not consider upstart jobs but I'm not at all
opposed to add that sort of thing.
Register the missing .remount_fs handler. This handler isn't strictly
required because the VFS does a pretty good job updating most of the
MS_* flags. However, there's no harm in using the hook to call the
registered zpl callback for various MS_* flags. Additionaly, this
allows us to lay the ground work for more complicated argument parsing
in the future.
Register the missing .sync_fs handler. This is a noop in most cases
because the usual requirement is that sync just be initiated. As part
of the DMU's normal transaction processing txgs will be frequently
synced. However, when the 'wait' flag is set the requirement is that
.sync_fs must not return until the data is safe on disk. With the
addition of the .sync_fs handler this is now properly implemented.
When updating /etc/mtab we should be careful and strip certain
options. In particular, we need to strip 'zfsutil' because if
we don't the mount utility will helpfull provide it to the
mount helper when we issue mount(8) again. This subverts the
check that the caller is zfs(8) and not mount(8).
Allow the mount(8) utility to always operate on all datasets when
remounting them read-only. This critical for rc.sysinit/umountroot
which remounts the root filesystem read-only during shutdown to
ensure everything is correctly flushed to disk.
Fix minor typo, the check to set zfsutil should use the bitwise
'&'. I must have accidentally hit the adjacent '*' and obviously
neither the compiler or my code review caught this. Fix it now.
ZFS should only change the i/o scheduler for a disk when it has
ownership of the whole disk. This is basically the same logic as
adjusting the write cache behavior on a disk. This change updates
the vdev disk code to skip partitions when setting the i/o scheduler.
Closes#152
When run with a root '/' cwd the mount.zfs helper would strip not
only the '/' but also the next character from the dataset name.
For example, '/tank' was changed to 'ank' instead of just 'tank'.
Originally, this was done for the '/tmp' cwd case where we needed
to strip the '/' following the cwd. For example '/tmp/tank' needed
to remove the '/tmp' cwd plus 1 character for the '/'.
This change fixes the problem by checking the cwd and if it ends in
a '/' it does not strip and extra character. Otherwise it will strip
the next character. I believe this should only ever be true for the
root directory.
Closes#148
Because we are dependent of the system mount/umount utilities to
ensure correct mtab locking, we should not suppress their error
output. During a successful mount/umount they will be silent,
but during a failure the error message they print is the only sure
way to know why a mount failed. This is because the (u)mount(8)
return code does not contain the result of the system call issued.
The only way to clearly idenify why thing failed is to rely on
the error message printed by the tool.
Longer term once libmount is available we can issue the mount/umount
system calls within the tool and still be ensured correct mtab locking.
Closed#107
Several issues related to strange mount/umount behavior were reported
and this commit should address most of them. The original idea was
to put in place a zfs mount helper (mount.zfs). This helper is used
to enforce 'legacy' mount behavior, and perform any extra mount argument
processing (selinux, zfsutil, etc). This helper wasn't ready for the
0.6.0-rc1 release but with this change it's functional but needs to
extensively tested.
This change addresses the following open issues.
Closes#101Closes#107Closes#113Closes#115Closes#119
Due to an uninitialized variable files opened with O_APPEND may
overwrite the start of the file rather than append to it. This
was introduced accidentally when I removed the Solaris vnodes.
The zfs_range_lock_writer() function used to key off zf->z_vnode
to determine if a znode_t was for a zvol of zpl object. With
the removal of vnodes this was replaced by the flag zp->z_is_zvol.
This flag was used to control the append behavior for range locks.
Unfortunately, this value was never properly initialized after
the vnode removal. However, because most of memory is usually
zeros it happened to be set correctly most of the time making
the bug appear racy. Properly initializing zp->z_is_zvol to
zero completely resolves the problem with O_APPEND.
Closes#126
Move 'bulk' and 'xattr_bulk' from the stack to the heap to minimize
stack space usage. These two arrays consumed 448 bytes on the stack
and have been replaced by two 8 byte points for a total stack space
saving of 432 bytes. The zfs_setattr() path had been previously
observed to overrun the stack in certain circumstances.
The original range lock implementation had to be modified by commit
8926ab7 because it was unsafe on Linux. In particular, calling
cv_destroy() immediately after cv_broadcast() is dangerous because
the waiters may still be asleep. Thus the following cv_destroy()
will free memory which may still be in use.
This was fixed by updating cv_destroy() to block on waiters but
this in turn introduced a deadlock. The deadlock was resolved
with the use of a taskq to move the offending free outside the
range lock. This worked well but using the taskq for the free
resulted in a serious performace hit. This is somewhat ironic
because at the time I felt using the taskq might improve things
by making the free asynchronous.
This patch refines the original fix and moves the free from the
taskq to a private free list. Then items which must be free'd
are simply inserted in to the list. When the range lock is dropped
it's safe to free the items. The list is walked and all rl_t
entries are freed.
This change improves small cached read performance by 26x. This
was expected because for small reads the number of locking calls
goes up significantly. More surprisingly this change significantly
improves large cache read performance. This probably attributable
to better cpu/memory locality. Very likely the same processor
which allocated the memory is now freeing it.
bs ext3 zfs zfs+fix faster
----------------------------------------------
512 435 3 79 26x
1k 820 7 160 22x
2k 1536 14 305 21x
4k 2764 28 572 20x
8k 3788 50 1024 20x
16k 4300 86 1843 21x
32k 4505 138 2560 18x
64k 5324 252 3891 15x
128k 5427 276 4710 17x
256k 5427 413 5017 12x
512k 5427 497 5324 10x
1m 5427 521 5632 10x
Closes#142
Manuel Amador (Rudd-O)
Brian Behlendorf
Ned Bass
Fajar A. Nugraha