Avoid overloading the CPU scheduler with enormous load averages
when handling high call-rate NFS loads. When the knfsd bottom half
is made aware of an incoming call by the socket layer, it tries to
choose an nfsd thread and wake it up. As long as there are idle
threads, one will be woken up.
If there are lot of nfsd threads (a sensible configuration when
the server is disk-bound or is running an HSM), there will be many
more nfsd threads than CPUs to run them. Under a high call-rate
low service-time workload, the result is that almost every nfsd is
runnable, but only a handful are actually able to run. This situation
causes two significant problems:
1. The CPU scheduler takes over 10% of each CPU, which is robbing
the nfsd threads of valuable CPU time.
2. At a high enough load, the nfsd threads starve userspace threads
of CPU time, to the point where daemons like portmap and rpc.mountd
do not schedule for tens of seconds at a time. Clients attempting
to mount an NFS filesystem timeout at the very first step (opening
a TCP connection to portmap) because portmap cannot wake up from
select() and call accept() in time.
Disclaimer: these effects were observed on a SLES9 kernel, modern
kernels' schedulers may behave more gracefully.
The solution is simple: keep in each svc_pool a counter of the number
of threads which have been woken but have not yet run, and do not wake
any more if that count reaches an arbitrary small threshold.
Testing was on a 4 CPU 4 NIC Altix using 4 IRIX clients, each with 16
synthetic client threads simulating an rsync (i.e. recursive directory
listing) workload reading from an i386 RH9 install image (161480
regular files in 10841 directories) on the server. That tree is small
enough to fill in the server's RAM so no disk traffic was involved.
This setup gives a sustained call rate in excess of 60000 calls/sec
before being CPU-bound on the server. The server was running 128 nfsds.
Profiling showed schedule() taking 6.7% of every CPU, and __wake_up()
taking 5.2%. This patch drops those contributions to 3.0% and 2.2%.
Load average was over 120 before the patch, and 20.9 after.
This patch is a forward-ported version of knfsd-avoid-nfsd-overload
which has been shipping in the SGI "Enhanced NFS" product since 2006.
It has been posted before:
http://article.gmane.org/gmane.linux.nfs/10374
Signed-off-by: Greg Banks <gnb@sgi.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Stop gathering the data that feeds the 'th' line in /proc/net/rpc/nfsd
because the questionable data provided is not worth the scalability
impact of calculating it. Instead, always report zeroes. The current
approach suffers from three major issues:
1. update_thread_usage() increments buckets by call service
time or call arrival time...in jiffies. On lightly loaded
machines, call service times are usually < 1 jiffy; on
heavily loaded machines call arrival times will be << 1 jiffy.
So a large portion of the updates to the buckets are rounded
down to zero, and the histogram is undercounting.
2. As seen previously on the nfs mailing list, the format in which
the histogram is presented is cryptic, difficult to explain,
and difficult to use.
3. Updating the histogram requires taking a global spinlock and
dirtying the global variables nfsd_last_call, nfsd_busy, and
nfsdstats *twice* on every RPC call, which is a significant
scaling limitation.
Testing on a 4 CPU 4 NIC Altix using 4 IRIX clients each doing
1K streaming reads at full line rate, shows the stats update code
(inlined into nfsd()) takes about 1.7% of each CPU. This patch drops
the contribution from nfsd() into the profile noise.
This patch is a forward-ported version of knfsd-remove-nfsd-threadstats
which has been shipping in the SGI "Enhanced NFS" product since 2006.
In that time, exactly one customer has noticed that the threadstats
were missing. It has been previously posted:
http://article.gmane.org/gmane.linux.nfs/10376
and more recently requested to be posted again.
Signed-off-by: Greg Banks <gnb@sgi.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
The main nfsd code was recently modified to no longer do lookups from
withing the readdir callback, to avoid locking problems on certain
filesystems.
This (rather hacky, and overdue for replacement) NFSv4 recovery code has
the same problem. Fix it to build up a list of names (instead of
dentries) and do the lookups afterwards.
Reported symptoms were a deadlock in the xfs code (called from
nfsd4_recdir_load), with /var/lib/nfs on xfs.
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Reported-by: David Warren <warren@atmos.washington.edu>
Currently putpubfh returns NFSERR_OPNOTSUPP, which isn't actually
allowed for v4. The right error is probably NFSERR_NOTSUPP.
But let's just implement it; though rarely seen, it can be used by
Solaris (with a special mount option), is mandated by the rfc, and is
trivial for us to support.
Thanks to Yang Hongyang for pointing out the original problem, and to
Mike Eisler, Tom Talpey, Trond Myklebust, and Dave Noveck for further
argument....
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
If a filesystem being written to via NFS returns a short write count
(as opposed to an error) to nfsd, nfsd treats that as a success for
the entire write, rather than the short count that actually succeeded.
For example, given a 8192 byte write, if the underlying filesystem
only writes 4096 bytes, nfsd will ack back to the nfs client that all
8192 bytes were written. The nfs client does have retry logic for
short writes, but this is never called as the client is told the
complete write succeeded.
There are probably other ways it could happen, but in my case it
happened with a fuse (filesystem in userspace) filesystem which can
rather easily have a partial write.
Here is a patch to properly return the short write count to the
client.
Signed-off-by: David Shaw <dshaw@jabberwocky.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Thanks for Bill Baker at sun.com for catching this
at Connectathon 2009.
This bug was introduced in 2.6.27
Signed-off-by: Benny Halevy <bhalevy@panasas.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
The errors returned aren't used. Just return 0 and make them available
to a dprintk(). Also, consistently use -ERRNO errors instead of nfs
errors.
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Reviewed-by: Benny Halevy <bhalevy@panasas.com>
As part of reducing the scope of the client_mutex, and in order to
remove the need for mutexes from the callback code (so that callbacks
can be done as asynchronous rpc calls), move manipulations of the
file_hashtable under the recall_lock.
Update the relevant comments while we're here.
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Cc: Alexandros Batsakis <batsakis@netapp.com>
Reviewed-by: Benny Halevy <bhalevy@panasas.com>
Since free_client() is guaranteed to only be called once, and to only
touch the client structure itself (not any common data structures), it
has no need for the state lock.
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Cc: Alexandros Batsakis <batsakis@netapp.com>
Previous cleanup reveals an obvious (though harmless) bug: when
delegreturn gets a stateid that isn't for a delegation, it should return
an error rather than doing nothing.
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Delegreturn is enough a special case for preprocess_stateid_op to
warrant just open-coding it in delegreturn.
There should be no change in behavior here; we're just reshuffling code.
Thanks to Yang Hongyang for catching a critical typo.
Reviewed-by: Yang Hongyang <yanghy@cn.fujitsu.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
I can't recall ever seeing these printk's used to debug a problem. I'll
happily put them back if we see a case where they'd be useful. (Though
if we do that the find_XXX() errors would probably be better
reported in find_XXX() functions themselves.)
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Note that we exit this first big "if" with stp == NULL if and only if we
took the first branch; therefore, the second "if" is redundant, and we
can just combine the two, simplifying the logic.
Reviewed-by: Yang Hongyang <yanghy@cn.fujitsu.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
The base versions handle constant folding now, none of these headers
are exported to userspace, so the __ prefixed versions are not
necessary.
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Reviewed-by: NeilBrown <neilb@suse.de>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
sometimes HPUX nfs client sends a create request to linux nfs server(v2/v3).
the dump of the request is like:
obj_attributes
mode: value follows
set_it: value follows (1)
mode: 00
uid: no value
set_it: no value (0)
gid: value follows
set_it: value follows (1)
gid: 8030
size: value follows
set_it: value follows (1)
size: 0
atime: don't change
set_it: don't change (0)
mtime: don't change
set_it: don't change (0)
note that mode is 00(havs no rwx privilege even for the owner) and it requires
to set size to 0.
as current nfsd(v2/v3) implementation, the server does mainly 2 steps:
1) creates the file in mode specified by calling vfs_create().
2) sets attributes for the file by calling nfsd_setattr().
at step 2), it finally calls file system specific setattr() function which may
fail when checking permission because changing size needs WRITE privilege but
it has none since mode is 000.
for this case, a new file created, we may simply ignore the request of
setting size to 0, so that WRITE privilege is not needed and the open
succeeds.
Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
--
vfs.c | 19 +++++++++++++++++++
1 file changed, 19 insertions(+)
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
No change in behavior, just rearranging the switch so that we break out
of the switch if and only if we're in the wait case.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
not having the state locked before putting the client/delegation causes a bug.
Also removed the comment from the function header about the state being already locked
Signed-off-by: Alexandros Batsakis <batsakis@netapp.com>
Signed-off-by: Benny Halevy <bhalevy@panasas.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
The use of |= is confusing--the bitmask is always initialized to zero in
this case, so we're effectively just doing an assignment here.
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Enable NFSD only when FILE_LOCKING is enabled, since we don't want to
support NFSD without FILE_LOCKING.
Signed-off-by: Manish Katiyar <mkatiyar@gmail.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
MSDOS_SUPER_MAGIC is defined in <linux/magic.h>,
so use MSDOS_SUPER_MAGIC directly.
Signed-off-by: Qinghuang Feng <qhfeng.kernel@gmail.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
The caller always knows specifically whether it's releasing a lockowner
or an openowner, and the code is simpler if we use separate functions
(and the apparent recursion is gone).
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
The flags here attempt to make the code more general, but I find it
actually just adds confusion.
I think it's clearer to separate the logic for the open and lock cases
entirely. And eventually we may want to separate the stateowner and
stateid types as well, as many of the fields aren't shared between the
lock and open cases.
Also move to eliminate forward references.
Start with the stateid's.
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Reviewed-by: Benny Halevy <bhalevy@panasas.com>
The benet driver is now in the proper place in drivers/net/benet, so we
can remove the staging version.
Acked-by: Sathya Perla <sathyap@serverengines.com>
Cc: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound-2.6:
ALSA: Fix vunmap and free order in snd_free_sgbuf_pages()
ALSA: mixart, fix lock imbalance
ALSA: pcm_oss, fix locking typo
ALSA: oss-mixer - Fixes recording gain control
ALSA: hda - Workaround for buggy DMA position on ATI controllers
ALSA: hda - Fix DMA mask for ATI controllers
ALSA: opl3sa2 - Fix NULL dereference when suspending snd_opl3sa2
In snd_free_sgbuf_pags(), vunmap() is called after releasing the SG
pages, and it causes errors on Xen as Xen manages the pages
differently. Although no significant errors have been reported on
the actual hardware, this order should be fixed other way round,
first vunmap() then free pages.
Cc: Jan Beulich <jbeulich@novell.com>
Cc: <stable@kernel.org>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
s/mutex_lock/mutex_unlock/ on 2 fail paths in snd_pcm_oss_proc_write.
Probably a typo, lock should be unlocked when leaving the function.
Signed-off-by: Jiri Slaby <jirislaby@gmail.com>
Cc: <stable@kernel.org>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
At the time of initialization, SNDRV_MIXER_OSS_PRESENT_PVOLUME bit is not
set for MIC (slot 7).
So, the same should not be checked when an application tries to do gain
control for audio recording devices.
Just check slot->present for SNDRV_MIXER_OSS_PRESENT_CVOLUME independently.
Verified with a simple application which opens /dev/dsp for recording and
/dev/mixer for volume control.
Have tested two usb audio mic devices.
Signed-off-by: Viral Mehta <viral.mehta@einfochips.com>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
The position-buffer on ATI controllers are unreliable as well as
on VIA chips, thus the same workaround for DMA position reading as
VIA is useful for ATI.
Cc: <stable@kernel.org>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
ATI controllers (at least some SB0600 models) appear buggy to handle
64bit DMA. As a workaround, reset GCAP bit0 and let the driver to
use only 32bit DMA on these controllers.
Cc: <stable@kernel.org>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
* 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4:
ext4: fix bb_prealloc_list corruption due to wrong group locking
ext4: fix bogus BUG_ONs in in mballoc code
ext4: Print the find_group_flex() warning only once
ext4: fix header check in ext4_ext_search_right() for deep extent trees.
