Get rid of init_MUTEX[_LOCKED]() and use sema_init() instead.
(Ported to current XFS code by <aelder@sgi.com>.)
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
This patch reverts the driver to enabling/disabling the NFC interrupt
mask rather than enabling/disabling the system interrupt. This cleans
up the driver so that it doesn't rely on interrupts being disabled
within the interrupt handler.
For i.MX21 we keep the current behaviour, that is calling
enable_irq/disable_irq_nosync to enable/disable interrupts. This patch
is based on earlier work by John Ogness.
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
Acked-by: John Ogness <john.ogness@linutronix.de>
Tested-by: John Ogness <john.ogness@linutronix.de>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patch adds support for 32bit project quota identifiers.
On disk format is backward compatible with 16bit projid numbers. projid
on disk is now kept in two 16bit values - di_projid_lo (which holds the
same position as old 16bit projid value) and new di_projid_hi (takes
existing padding) and converts from/to 32bit value on the fly.
xfs_admin (for existing fs), mkfs.xfs (for new fs) needs to be used
to enable PROJID32BIT support.
Signed-off-by: Arkadiusz Miśkiewicz <arekm@maven.pl>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Stop having two different names for many buffer functions and use
the more descriptive xfs_buf_* names directly.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
We're not actually passing around credentials inside XFS for a while
now, so remove all xfs_cred.h with it's cred_t typedef and all
instances of it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
This header only provides one extern that isn't actually declared
anywhere, and shadowed by a macro.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
It used to have a place when it contained an automatically generated
CVS version, but these days it's entirely superflous.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Use the correct prototype for xfs_trans_committed instead of casting it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
These days inode64 should only control which AGs we allocate new
inodes from, while we still try to support reading all existing
inodes. To make this actually work the check ontop of xfs_iget
needs to be relaxed to allow inodes in all allocation groups instead
of just those that we allow allocating inodes from. Note that we
can't simply remove the check - it prevents us from accessing
invalid data when fed invalid inode numbers from NFS or bulkstat.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Update the per-cpu counters manually in xfs_trans_unreserve_and_mod_sb
and remove support for per-cpu counters from xfs_mod_incore_sb_batch
to simplify it. And added benefit is that we don't have to take
m_sb_lock for transactions that only modify per-cpu counters.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Export xfs_icsb_modify_counters and always use it for modifying
the per-cpu counters. Remove support for per-cpu counters from
xfs_mod_incore_sb to simplify it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Fail the mount if we can't allocate memory for the per-CPU counters.
This is consistent with how we handle everything else in the mount
path and makes the superblock counter modification a lot simpler.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
pahole reports the struct xfs_buf has quite a few holes in it, so
packing the structure better will reduce the size of it by 16 bytes.
Also, move all the fields used in cache lookups into the first
cacheline.
Before on x86_64:
/* size: 320, cachelines: 5 */
/* sum members: 298, holes: 6, sum holes: 22 */
After on x86_64:
/* size: 304, cachelines: 5 */
/* padding: 6 */
/* last cacheline: 48 bytes */
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
The buffer cache hash is showing typical hash scalability problems.
In large scale testing the number of cached items growing far larger
than the hash can efficiently handle. Hence we need to move to a
self-scaling cache indexing mechanism.
I have selected rbtrees for indexing becuse they can have O(log n)
search scalability, and insert and remove cost is not excessive,
even on large trees. Hence we should be able to cache large numbers
of buffers without incurring the excessive cache miss search
penalties that the hash is imposing on us.
To ensure we still have parallel access to the cache, we need
multiple trees. Rather than hashing the buffers by disk address to
select a tree, it seems more sensible to separate trees by typical
access patterns. Most operations use buffers from within a single AG
at a time, so rather than searching lots of different lists,
separate the buffer indexes out into per-AG rbtrees. This means that
searches during metadata operation have a much higher chance of
hitting cache resident nodes, and that updates of the tree are less
likely to disturb trees being accessed on other CPUs doing
independent operations.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
Memory reclaim via shrinkers has a terrible habit of having N+M
concurrent shrinker executions (N = num CPUs, M = num kswapds) all
trying to shrink the same cache. When the cache they are all working
on is protected by a single spinlock, massive contention an
slowdowns occur.
Wrap the per-ag inode caches with a reclaim mutex to serialise
reclaim access to the AG. This will block concurrent reclaim in each
AG but still allow reclaim to scan multiple AGs concurrently. Allow
shrinkers to move on to the next AG if it can't get the lock, and if
we can't get any AG, then start blocking on locks.
To prevent reclaimers from continually scanning the same inodes in
each AG, add a cursor that tracks where the last reclaim got up to
and start from that point on the next reclaim. This should avoid
only ever scanning a small number of inodes at the satart of each AG
and not making progress. If we have a non-shrinker based reclaim
pass, ignore the cursor and reset it to zero once we are done.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Batch and optimise the per-ag inode lookup for reclaim to minimise
scanning overhead. This involves gang lookups on the radix trees to
get multiple inodes during each tree walk, and tighter validation of
what inodes can be reclaimed without blocking befor we take any
locks.
This is based on ideas suggested in a proof-of-concept patch
posted by Nick Piggin.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
With the reclaim code separated from the generic walking code, it is
simple to implement batched lookups for the generic walk code.
Separate out the inode validation from the execute operations and
modify the tree lookups to get a batch of inodes at a time.
Reclaim operations will be optimised separately.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
When doing read side inode cache walks, the code to validate and
grab an inode is common to all callers. Split it out of the execute
callbacks in preparation for batching lookups. Similarly, split out
the inode reference dropping from the execute callbacks into the
main lookup look to be symmetric with the grab.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
The reclaim walk requires different locking and has a slightly
different walk algorithm, so separate it out so that it can be
optimised separately.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
For RT and external log devices, we never use hashed buffers on them
now. Remove the buftarg hash tables that are set up for them.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
When we are checking we can access the last block of each device, we
do not need to use cached buffers as they will be tossed away
immediately. Use uncached buffers for size checks so that all IO
prior to full in-memory structure initialisation does not use the
buffer cache.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
Filesystem level managed buffers are buffers that have their
lifecycle controlled by the filesystem layer, not the buffer cache.
We currently cache these buffers, which makes cleanup and cache
walking somewhat troublesome. Convert the fs managed buffers to
uncached buffers obtained by via xfs_buf_get_uncached(), and remove
the XBF_FS_MANAGED special cases from the buffer cache.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
Each buffer contains both a buftarg pointer and a mount pointer. If
we add a mount pointer into the buftarg, we can avoid needing the
b_mount field in every buffer and grab it from the buftarg when
needed instead. This shrinks the xfs_buf by 8 bytes.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
To avoid the need to use cached buffers for single-shot or buffers
cached at the filesystem level, introduce a new buffer read
primitive that bypasses the cache an reads directly from disk.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
xfs_buf_get_nodaddr() is really used to allocate a buffer that is
uncached. While it is not directly assigned a disk address, the fact
that they are not cached is a more important distinction. With the
upcoming uncached buffer read primitive, we should be consistent
with this disctinction.
While there, make page allocation in xfs_buf_get_nodaddr() safe
against memory reclaim re-entrancy into the filesystem by allowing
a flags parameter to be passed.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
Under heavy multi-way parallel create workloads, the VFS struggles
to write back all the inodes that have been changed in age order.
The bdi flusher thread becomes CPU bound, spending 85% of it's time
in the VFS code, mostly traversing the superblock dirty inode list
to separate dirty inodes old enough to flush.
We already keep an index of all metadata changes in age order - in
the AIL - and continued log pressure will do age ordered writeback
without any extra overhead at all. If there is no pressure on the
log, the xfssyncd will periodically write back metadata in ascending
disk address offset order so will be very efficient.
Hence we can stop marking VFS inodes dirty during transaction commit
or when changing timestamps during transactions. This will keep the
inodes in the superblock dirty list to those containing data or
unlogged metadata changes.
However, the timstamp changes are slightly more complex than this -
there are a couple of places that do unlogged updates of the
timestamps, and the VFS need to be informed of these. Hence add a
new function xfs_trans_ichgtime() for transactional changes,
and leave xfs_ichgtime() for the non-transactional changes.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
When we start taking a reference to the per-ag for every cached
buffer in the system, kernel lockstat profiling on an 8-way create
workload shows the mp->m_perag_lock has higher acquisition rates
than the inode lock and has significantly more contention. That is,
it becomes the highest contended lock in the system.
The perag lookup is trivial to convert to lock-less RCU lookups
because perag structures never go away. Hence the only thing we need
to protect against is tree structure changes during a grow. This can
be done simply by replacing the locking in xfs_perag_get() with RCU
read locking. This removes the mp->m_perag_lock completely from this
path.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
When we start taking references per cached buffer to the the perag
it is cached on, it will blow the current debug maximum reference
count assert out of the water. The assert has never caught a bug,
and we have tracing to track changes if there ever is a problem,
so just remove it.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
When commiting a transaction, we do a lock CIL state lock round trip
on every single log vector we insert into the CIL. This is resulting
in the lock being as hot as the inode and dcache locks on 8-way
create workloads. Rework the insertion loops to bring the number
of lock round trips to one per transaction for log vectors, and one
more do the busy extents.
Also change the allocation of the log vector buffer not to zero it
as we copy over the entire allocated buffer anyway.
This patch also includes a structural cleanup to the CIL item
insertion provided by Christoph Hellwig.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
Ionut Gabriel Popescu <poyo_vl@yahoo.com> submitted a simple change
to eliminate some "may be used uninitialized" warnings when building
XFS. The reported condition seems to be something that GCC did not
used to recognize or report. The warnings were produced by:
gcc version 4.5.0 20100604
[gcc-4_5-branch revision 160292] (SUSE Linux)
Signed-off-by: Ionut Gabriel Popescu <poyo_vl@yahoo.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
The implementation os ->kill_root only differ by either simply
zeroing out the now unused buffer in the btree cursor in the inode
allocation btree or using xfs_btree_setbuf in the allocation btree.
Initially both of them used xfs_btree_setbuf, but the use in the
ialloc btree was removed early on because it interacted badly with
xfs_trans_binval.
In addition to zeroing out the buffer in the cursor xfs_btree_setbuf
updates the bc_ra array in the btree cursor, and calls
xfs_trans_brelse on the buffer previous occupying the slot.
The bc_ra update should be done for the alloc btree updated too,
although the lack of it does not cause serious problems. The
xfs_trans_brelse call on the other hand is effectively a no-op in
the end - it keeps decrementing the bli_recur refcount until it hits
zero, and then just skips out because the buffer will always be
dirty at this point. So removing it for the allocation btree is
just fine.
So unify the code and move it to xfs_btree.c. While we're at it
also replace the call to xfs_btree_setbuf with a NULL bp argument in
xfs_btree_del_cursor with a direct call to xfs_trans_brelse given
that the cursor is beeing freed just after this and the state
updates are superflous. After this xfs_btree_setbuf is only used
with a non-NULL bp argument and can thus be simplified.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
In xfs_qm_dqflush we know that q_blkno must be initialized already from a
previous xfs_qm_dqread. So instead of calling xfs_qm_dqtobp we can
simply read the quota buffer directly. This also saves us from a duplicate
xfs_qm_dqcheck call check and allows xfs_qm_dqtobp to be simplified now
that it is always called for a newly initialized inode. In addition to
that properly unwind all locks in xfs_qm_dqflush when xfs_qm_dqcheck
fails.
This mirrors a similar cleanup in the inode lookup done earlier.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
There is no need to have the users and group/project quota locked at the
same time. Get rid of xfs_qm_dqget_noattach and just do a xfs_qm_dqget
inside xfs_qm_quotacheck_dqadjust for the quota we are operating on
right now. The new version of xfs_qm_quotacheck_dqadjust holds the
inode lock over it's operations, which is not a problem as it simply
increments counters and there is no concern about log contention
during mount time.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
XFS_IOC_ZERO_RANGE is the equivalent of an atomic XFS_IOC_UNRESVSP/
XFS_IOC_RESVSP call pair. It enabled ranges of written data to be
turned into zeroes without requiring IO or having to free and
reallocate the extents in the range given as would occur if we had
to punch and then preallocate them separately. This enables
applications to zero parts of files very quickly without changing
the layout of the files in any way.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
While XFS passes ranges to operate on from the core code, the
functions being called ignore the either the entire range or the end
of the range. This is historical because when the function were
written linux didn't have the necessary range operations. Update the
functions to use the correct operations.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
* 'for-linus/i2c/2636-rc8' of git://git.fluff.org/bjdooks/linux:
i2c-imx: do not allow interruptions when waiting for I2C to complete
i2c-davinci: Fix TX setup for more SoCs
When CPU is idle and on first interrupt, irq_enter calls tick_check_idle()
to notify interruption from idle. But, there is a problem if this call
is done after __irq_enter, as all routines in __irq_enter may find
stale time due to yet to be done tick_check_idle.
Specifically, trace calls in __irq_enter when they use global clock and also
account_system_vtime change in this patch as it wants to use sched_clock_cpu()
to do proper irq timing.
But, tick_check_idle was moved after __irq_enter intentionally to
prevent problem of unneeded ksoftirqd wakeups by the commit ee5f80a:
irq: call __irq_enter() before calling the tick_idle_check
Impact: avoid spurious ksoftirqd wakeups
Moving tick_check_idle() before __irq_enter and wrapping it with
local_bh_enable/disable would solve both the problems.
Fixed-by: Yong Zhang <yong.zhang0@gmail.com>
Signed-off-by: Venkatesh Pallipadi <venki@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1286237003-12406-9-git-send-email-venki@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The idea was suggested by Peter Zijlstra here:
http://marc.info/?l=linux-kernel&m=127476934517534&w=2
irq time is technically not available to the tasks running on the CPU.
This patch removes irq time from CPU power piggybacking on
sched_rt_avg_update().
Tested this by keeping CPU X busy with a network intensive task having 75%
oa a single CPU irq processing (hard+soft) on a 4-way system. And start seven
cycle soakers on the system. Without this change, there will be two tasks on
each CPU. With this change, there is a single task on irq busy CPU X and
remaining 7 tasks are spread around among other 3 CPUs.
Signed-off-by: Venkatesh Pallipadi <venki@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1286237003-12406-8-git-send-email-venki@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Scheduler accounts both softirq and interrupt processing times to the
currently running task. This means, if the interrupt processing was
for some other task in the system, then the current task ends up being
penalized as it gets shorter runtime than otherwise.
Change sched task accounting to acoount only actual task time from
currently running task. Now update_curr(), modifies the delta_exec to
depend on rq->clock_task.
Note that this change only handles CONFIG_IRQ_TIME_ACCOUNTING case. We can
extend this to CONFIG_VIRT_CPU_ACCOUNTING with minimal effort. But, thats
for later.
This change will impact scheduling behavior in interrupt heavy conditions.
Tested on a 4-way system with eth0 handled by CPU 2 and a network heavy
task (nc) running on CPU 3 (and no RSS/RFS). With that I have CPU 2
spending 75%+ of its time in irq processing. CPU 3 spending around 35%
time running nc task.
Now, if I run another CPU intensive task on CPU 2, without this change
/proc/<pid>/schedstat shows 100% of time accounted to this task. With this
change, it rightly shows less than 25% accounted to this task as remaining
time is actually spent on irq processing.
Signed-off-by: Venkatesh Pallipadi <venki@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1286237003-12406-7-git-send-email-venki@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch adds IRQ_TIME_ACCOUNTING option on x86 and runtime enables it
when TSC is enabled.
This change just enables fine grained irq time accounting, isn't used yet.
Following patches use it for different purposes.
Signed-off-by: Venkatesh Pallipadi <venki@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1286237003-12406-6-git-send-email-venki@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
s390/powerpc/ia64 have support for CONFIG_VIRT_CPU_ACCOUNTING which does
the fine granularity accounting of user, system, hardirq, softirq times.
Adding that option on archs like x86 will be challenging however, given the
state of TSC reliability on various platforms and also the overhead it will
add in syscall entry exit.
Instead, add a lighter variant that only does finer accounting of
hardirq and softirq times, providing precise irq times (instead of timer tick
based samples). This accounting is added with a new config option
CONFIG_IRQ_TIME_ACCOUNTING so that there won't be any overhead for users not
interested in paying the perf penalty.
This accounting is based on sched_clock, with the code being generic.
So, other archs may find it useful as well.
This patch just adds the core logic and does not enable this logic yet.
Signed-off-by: Venkatesh Pallipadi <venki@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1286237003-12406-5-git-send-email-venki@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
To account softirq time cleanly in scheduler, we need to identify whether
softirq is invoked in ksoftirqd context or softirq at hardirq tail context.
Add PF_KSOFTIRQD for that purpose.
As all PF flag bits are currently taken, create space by moving one of the
infrequently used bits (PF_THREAD_BOUND) down in task_struct to be along
with some other state fields.
Signed-off-by: Venkatesh Pallipadi <venki@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1286237003-12406-4-git-send-email-venki@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Just a minor cleanup patch that makes things easier to the following patches.
No functionality change in this patch.
Signed-off-by: Venkatesh Pallipadi <venki@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1286237003-12406-3-git-send-email-venki@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Peter Zijlstra found a bug in the way softirq time is accounted in
VIRT_CPU_ACCOUNTING on this thread:
http://lkml.indiana.edu/hypermail//linux/kernel/1009.2/01366.html
The problem is, softirq processing uses local_bh_disable internally. There
is no way, later in the flow, to differentiate between whether softirq is
being processed or is it just that bh has been disabled. So, a hardirq when bh
is disabled results in time being wrongly accounted as softirq.
Looking at the code a bit more, the problem exists in !VIRT_CPU_ACCOUNTING
as well. As account_system_time() in normal tick based accouting also uses
softirq_count, which will be set even when not in softirq with bh disabled.
Peter also suggested solution of using 2*SOFTIRQ_OFFSET as irq count
for local_bh_{disable,enable} and using just SOFTIRQ_OFFSET while softirq
processing. The patch below does that and adds API in_serving_softirq() which
returns whether we are currently processing softirq or not.
Also changes one of the usages of softirq_count in net/sched/cls_cgroup.c
to in_serving_softirq.
Looks like many usages of in_softirq really want in_serving_softirq. Those
changes can be made individually on a case by case basis.
Signed-off-by: Venkatesh Pallipadi <venki@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1286237003-12406-2-git-send-email-venki@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
When SD_PREFER_SIBLING is set on a sched domain, drop group_capacity to 1
only if the local group has extra capacity. The extra check prevents the case
where you always pull from the heaviest group when it is already under-utilized
(possible with a large weight task outweighs the tasks on the system).
For example, consider a 16-cpu quad-core quad-socket machine with MC and NUMA
scheduling domains. Let's say we spawn 15 nice0 tasks and one nice-15 task,
and each task is running on one core. In this case, we observe the following
events when balancing at the NUMA domain:
- find_busiest_group() will always pick the sched group containing the niced
task to be the busiest group.
- find_busiest_queue() will then always pick one of the cpus running the
nice0 task (never picks the cpu with the nice -15 task since
weighted_cpuload > imbalance).
- The load balancer fails to migrate the task since it is the running task
and increments sd->nr_balance_failed.
- It repeats the above steps a few more times until sd->nr_balance_failed > 5,
at which point it kicks off the active load balancer, wakes up the migration
thread and kicks the nice 0 task off the cpu.
The load balancer doesn't stop until we kick out all nice 0 tasks from
the sched group, leaving you with 3 idle cpus and one cpu running the
nice -15 task.
When balancing at the NUMA domain, we drop sgs.group_capacity to 1 if the child
domain (in this case MC) has SD_PREFER_SIBLING set. Subsequent load checks are
not relevant because the niced task has a very large weight.
In this patch, we add an extra condition to the "if(prefer_sibling)" check in
update_sd_lb_stats(). We drop the capacity of a group only if the local group
has extra capacity, ie. nr_running < group_capacity. This patch preserves the
original intent of the prefer_siblings check (to spread tasks across the system
in low utilization scenarios) and fixes the case above.
It helps in the following ways:
- In low utilization cases (where nr_tasks << nr_cpus), we still drop
group_capacity down to 1 if we prefer siblings.
- On very busy systems (where nr_tasks >> nr_cpus), sgs.nr_running will most
likely be > sgs.group_capacity.
- When balancing large weight tasks, if the local group does not have extra
capacity, we do not pick the group with the niced task as the busiest group.
This prevents failed balances, active migration and the under-utilization
described above.
Signed-off-by: Nikhil Rao <ncrao@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1287173550-30365-5-git-send-email-ncrao@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch forces a load balance on a newly idle cpu when the local group has
extra capacity and the busiest group does not have any. It improves system
utilization when balancing tasks with a large weight differential.
Under certain situations, such as a niced down task (i.e. nice = -15) in the
presence of nr_cpus NICE0 tasks, the niced task lands on a sched group and
kicks away other tasks because of its large weight. This leads to sub-optimal
utilization of the machine. Even though the sched group has capacity, it does
not pull tasks because sds.this_load >> sds.max_load, and f_b_g() returns NULL.
With this patch, if the local group has extra capacity, we shortcut the checks
in f_b_g() and try to pull a task over. A sched group has extra capacity if the
group capacity is greater than the number of running tasks in that group.
Thanks to Mike Galbraith for discussions leading to this patch and for the
insight to reuse SD_NEWIDLE_BALANCE.
Signed-off-by: Nikhil Rao <ncrao@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1287173550-30365-4-git-send-email-ncrao@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
