Pull scheduler fixes from Ingo Molnar:
"Four misc fixes: each was deemed serious enough to warrant v3.15
inclusion"
* 'sched-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/fair: Fix tg_set_cfs_bandwidth() deadlock on rq->lock
sched/dl: Fix race in dl_task_timer()
sched: Fix sched_policy < 0 comparison
sched/numa: Fix use of spin_{un}lock_irq() when interrupts are disabled
Yoshihiro Yunomae reported that the ring buffer data for a trace
instance does not get properly cleaned up when it fails. He proposed
a patch that manually cleaned the data up and addad a bunch of labels.
The labels are not needed because all trace array is allocated with
a kzalloc which initializes it to 0 and all kfree()s can take a NULL
pointer and will ignore it.
Adding a new helper function free_trace_buffers() that can also take
null buffers to free the buffers that were allocated by
allocate_trace_buffers().
Link: http://lkml.kernel.org/r/20140605223522.32311.31664.stgit@yunodevel
Reported-by: Yoshihiro YUNOMAE <yoshihiro.yunomae.ez@hitachi.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
ftrace_trace_arrays links global_trace.list. However, global_trace
is not added to ftrace_trace_arrays if trace_alloc_buffers() failed.
As the result, ftrace_trace_arrays becomes an empty list. If
ftrace_trace_arrays is an empty list, current top_trace_array() returns
an invalid pointer. As the result, the kernel can induce memory corruption
or panic.
Current implementation does not check whether ftrace_trace_arrays is empty
list or not. So, in this patch, if ftrace_trace_arrays is empty list,
top_trace_array() returns NULL. Moreover, this patch makes all functions
calling top_trace_array() handle it appropriately.
Link: http://lkml.kernel.org/p/20140605223517.32311.99233.stgit@yunodevel
Signed-off-by: Yoshihiro YUNOMAE <yoshihiro.yunomae.ez@hitachi.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
This rwlock uses the arch_spin_lock_t as a waitqueue, and assuming the
arch_spin_lock_t is a fair lock (ticket,mcs etc..) the resulting
rwlock is a fair lock.
It fits in the same 8 bytes as the regular rwlock_t by folding the
reader and writer count into a single integer, using the remaining 4
bytes for the arch_spinlock_t.
Architectures that can single-copy adress bytes can optimize
queue_write_unlock() with a 0 write to the LSB (the write count).
Performance as measured by Davidlohr Bueso (rwlock_t -> qrwlock_t):
+--------------+-------------+---------------+
| Workload | #users | delta |
+--------------+-------------+---------------+
| alltests | > 1400 | -4.83% |
| custom | 0-100,> 100 | +1.43%,-1.57% |
| high_systime | > 1000 | -2.61 |
| shared | all | +0.32 |
+--------------+-------------+---------------+
http://www.stgolabs.net/qrwlock-stuff/aim7-results-vs-rwsem_optsin/
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
[peterz: near complete rewrite]
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: "Paul E.McKenney" <paulmck@linux.vnet.ibm.com>
Cc: linux-arch@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Link: http://lkml.kernel.org/n/tip-gac1nnl3wvs2ij87zv2xkdzq@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
perf tools like 'perf report' can aggregate samples by comm strings,
which generally works. However, there are other potential use-cases.
For example, to pair up 'calls' with 'returns' accurately (from branch
events like Intel BTS) it is necessary to identify whether the process
has exec'd. Although a comm event is generated when an 'exec' happens
it is also generated whenever the comm string is changed on a whim
(e.g. by prctl PR_SET_NAME). This patch adds a flag to the comm event
to differentiate one case from the other.
In order to determine whether the kernel supports the new flag, a
selection bit named 'exec' is added to struct perf_event_attr. The
bit does nothing but will cause perf_event_open() to fail if the bit
is set on kernels that do not have it defined.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/537D9EBE.7030806@intel.com
Cc: Paul Mackerras <paulus@samba.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: linux-fsdevel@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When calculating the average and standard deviation, it is required that
the count be less than UINT_MAX, otherwise the do_div() will get
undefined results. After 2^32 counts of data, the average and standard
deviation should pretty much be set anyway.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
I've been told that do_div() expects an unsigned 64 bit number, and
is undefined if a signed is used. This gave a warning on the MIPS
build. I'm not sure if a signed 64 bit dividend is really an issue
or not, but the calculation this is used for is standard deviation,
and that isn't going to be negative. We can just convert it to
unsigned and be safe.
Reported-by: David Daney <ddaney.cavm@gmail.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Pull ARM updates from Russell King:
- Major clean-up of the L2 cache support code. The existing mess was
becoming rather unmaintainable through all the additions that others
have done over time. This turns it into a much nicer structure, and
implements a few performance improvements as well.
- Clean up some of the CP15 control register tweaks for alignment
support, moving some code and data into alignment.c
- DMA properties for ARM, from Santosh and reviewed by DT people. This
adds DT properties to specify bus translations we can't discover
automatically, and to indicate whether devices are coherent.
- Hibernation support for ARM
- Make ftrace work with read-only text in modules
- add suspend support for PJ4B CPUs
- rework interrupt masking for undefined instruction handling, which
allows us to enable interrupts earlier in the handling of these
exceptions.
- support for big endian page tables
- fix stacktrace support to exclude stacktrace functions from the
trace, and add save_stack_trace_regs() implementation so that kprobes
can record stack traces.
- Add support for the Cortex-A17 CPU.
- Remove last vestiges of ARM710 support.
- Removal of ARM "meminfo" structure, finally converting us solely to
memblock to handle the early memory initialisation.
* 'for-linus' of git://ftp.arm.linux.org.uk/~rmk/linux-arm: (142 commits)
ARM: ensure C page table setup code follows assembly code (part II)
ARM: ensure C page table setup code follows assembly code
ARM: consolidate last remaining open-coded alignment trap enable
ARM: remove global cr_no_alignment
ARM: remove CPU_CP15 conditional from alignment.c
ARM: remove unused adjust_cr() function
ARM: move "noalign" command line option to alignment.c
ARM: provide common method to clear bits in CPU control register
ARM: 8025/1: Get rid of meminfo
ARM: 8060/1: mm: allow sub-architectures to override PCI I/O memory type
ARM: 8066/1: correction for ARM patch 8031/2
ARM: 8049/1: ftrace/add save_stack_trace_regs() implementation
ARM: 8065/1: remove last use of CONFIG_CPU_ARM710
ARM: 8062/1: Modify ldrt fixup handler to re-execute the userspace instruction
ARM: 8047/1: rwsem: use asm-generic rwsem implementation
ARM: l2c: trial at enabling some Cortex-A9 optimisations
ARM: l2c: add warnings for stuff modifying aux_ctrl register values
ARM: l2c: print a warning with L2C-310 caches if the cache size is modified
ARM: l2c: remove old .set_debug method
ARM: l2c: kill L2X0_AUX_CTRL_MASK before anyone else makes use of this
...
Merge futex fixes from Thomas Gleixner:
"So with more awake and less futex wreckaged brain, I went through my
list of points again and came up with the following 4 patches.
1) Prevent pi requeueing on the same futex
I kept Kees check for uaddr1 == uaddr2 as a early check for private
futexes and added a key comparison to both futex_requeue and
futex_wait_requeue_pi.
Sebastian, sorry for the confusion yesterday night. I really
misunderstood your question.
You are right the check is pointless for shared futexes where the
same physical address is mapped to two different virtual addresses.
2) Sanity check atomic acquisiton in futex_lock_pi_atomic
That's basically what Darren suggested.
I just simplified it to use futex_top_waiter() to find kernel
internal state. If state is found return -EINVAL and do not bother
to fix up the user space variable. It's corrupted already.
3) Ensure state consistency in futex_unlock_pi
The code is silly versus the owner died bit. There is no point to
preserve it on unlock when the user space thread owns the futex.
What's worse is that it does not update the user space value when
the owner died bit is set. So the kernel itself creates observable
inconsistency.
Another "optimization" is to retry an atomic unlock. That's
pointless as in a sane environment user space would not call into
that code if it could have unlocked it atomically. So we always
check whether there is kernel state around and only if there is
none, we do the unlock by setting the user space value to 0.
4) Sanitize lookup_pi_state
lookup_pi_state is ambigous about TID == 0 in the user space value.
This can be a valid state even if there is kernel state on this
uaddr, but we miss a few corner case checks.
I tried to come up with a smaller solution hacking the checks into
the current cruft, but it turned out to be ugly as hell and I got
more confused than I was before. So I rewrote the sanity checks
along the state documentation with awful lots of commentry"
* emailed patches from Thomas Gleixner <tglx@linutronix.de>:
futex: Make lookup_pi_state more robust
futex: Always cleanup owner tid in unlock_pi
futex: Validate atomic acquisition in futex_lock_pi_atomic()
futex-prevent-requeue-pi-on-same-futex.patch futex: Forbid uaddr == uaddr2 in futex_requeue(..., requeue_pi=1)
The current implementation of lookup_pi_state has ambigous handling of
the TID value 0 in the user space futex. We can get into the kernel
even if the TID value is 0, because either there is a stale waiters bit
or the owner died bit is set or we are called from the requeue_pi path
or from user space just for fun.
The current code avoids an explicit sanity check for pid = 0 in case
that kernel internal state (waiters) are found for the user space
address. This can lead to state leakage and worse under some
circumstances.
Handle the cases explicit:
Waiter | pi_state | pi->owner | uTID | uODIED | ?
[1] NULL | --- | --- | 0 | 0/1 | Valid
[2] NULL | --- | --- | >0 | 0/1 | Valid
[3] Found | NULL | -- | Any | 0/1 | Invalid
[4] Found | Found | NULL | 0 | 1 | Valid
[5] Found | Found | NULL | >0 | 1 | Invalid
[6] Found | Found | task | 0 | 1 | Valid
[7] Found | Found | NULL | Any | 0 | Invalid
[8] Found | Found | task | ==taskTID | 0/1 | Valid
[9] Found | Found | task | 0 | 0 | Invalid
[10] Found | Found | task | !=taskTID | 0/1 | Invalid
[1] Indicates that the kernel can acquire the futex atomically. We
came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
[2] Valid, if TID does not belong to a kernel thread. If no matching
thread is found then it indicates that the owner TID has died.
[3] Invalid. The waiter is queued on a non PI futex
[4] Valid state after exit_robust_list(), which sets the user space
value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
[5] The user space value got manipulated between exit_robust_list()
and exit_pi_state_list()
[6] Valid state after exit_pi_state_list() which sets the new owner in
the pi_state but cannot access the user space value.
[7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
[8] Owner and user space value match
[9] There is no transient state which sets the user space TID to 0
except exit_robust_list(), but this is indicated by the
FUTEX_OWNER_DIED bit. See [4]
[10] There is no transient state which leaves owner and user space
TID out of sync.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Kees Cook <keescook@chromium.org>
Cc: Will Drewry <wad@chromium.org>
Cc: Darren Hart <dvhart@linux.intel.com>
Cc: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If the owner died bit is set at futex_unlock_pi, we currently do not
cleanup the user space futex. So the owner TID of the current owner
(the unlocker) persists. That's observable inconsistant state,
especially when the ownership of the pi state got transferred.
Clean it up unconditionally.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Kees Cook <keescook@chromium.org>
Cc: Will Drewry <wad@chromium.org>
Cc: Darren Hart <dvhart@linux.intel.com>
Cc: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We need to protect the atomic acquisition in the kernel against rogue
user space which sets the user space futex to 0, so the kernel side
acquisition succeeds while there is existing state in the kernel
associated to the real owner.
Verify whether the futex has waiters associated with kernel state. If
it has, return -EINVAL. The state is corrupted already, so no point in
cleaning it up. Subsequent calls will fail as well. Not our problem.
[ tglx: Use futex_top_waiter() and explain why we do not need to try
restoring the already corrupted user space state. ]
Signed-off-by: Darren Hart <dvhart@linux.intel.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: Will Drewry <wad@chromium.org>
Cc: stable@vger.kernel.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If uaddr == uaddr2, then we have broken the rule of only requeueing from
a non-pi futex to a pi futex with this call. If we attempt this, then
dangling pointers may be left for rt_waiter resulting in an exploitable
condition.
This change brings futex_requeue() in line with futex_wait_requeue_pi()
which performs the same check as per commit 6f7b0a2a5c ("futex: Forbid
uaddr == uaddr2 in futex_wait_requeue_pi()")
[ tglx: Compare the resulting keys as well, as uaddrs might be
different depending on the mapping ]
Fixes CVE-2014-3153.
Reported-by: Pinkie Pie
Signed-off-by: Will Drewry <wad@chromium.org>
Signed-off-by: Kees Cook <keescook@chromium.org>
Cc: stable@vger.kernel.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Darren Hart <dvhart@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Introduce saved_cmdlines_size file for changing the number of saved pid-comms.
saved_cmdlines currently stores 128 command names using SAVED_CMDLINES, but
'no-existing processes' names are often lost in saved_cmdlines when we
read the trace data. So, by introducing saved_cmdlines_size file, we can
now change the 128 command names saved to something much larger if needed.
When we write a value to saved_cmdlines_size, the number of the value will
be stored in pid-comm list:
# echo 1024 > /sys/kernel/debug/tracing/saved_cmdlines_size
Here, 1024 command names can be stored. The default number is 128 and the maximum
number is PID_MAX_DEFAULT (=32768 if CONFIG_BASE_SMALL is not set). So, if we
want to avoid losing any command names, we need to set 32768 to
saved_cmdlines_size.
We can read the maximum number of the list:
# cat /sys/kernel/debug/tracing/saved_cmdlines_size
128
Link: http://lkml.kernel.org/p/20140605012427.22115.16173.stgit@yunodevel
Signed-off-by: Yoshihiro YUNOMAE <yoshihiro.yunomae.ez@hitachi.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Pull x86 cdso updates from Peter Anvin:
"Vdso cleanups and improvements largely from Andy Lutomirski. This
makes the vdso a lot less ''special''"
* 'x86/vdso' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/vdso, build: Make LE access macros clearer, host-safe
x86/vdso, build: Fix cross-compilation from big-endian architectures
x86/vdso, build: When vdso2c fails, unlink the output
x86, vdso: Fix an OOPS accessing the HPET mapping w/o an HPET
x86, mm: Replace arch_vma_name with vm_ops->name for vsyscalls
x86, mm: Improve _install_special_mapping and fix x86 vdso naming
mm, fs: Add vm_ops->name as an alternative to arch_vma_name
x86, vdso: Fix an OOPS accessing the HPET mapping w/o an HPET
x86, vdso: Remove vestiges of VDSO_PRELINK and some outdated comments
x86, vdso: Move the vvar and hpet mappings next to the 64-bit vDSO
x86, vdso: Move the 32-bit vdso special pages after the text
x86, vdso: Reimplement vdso.so preparation in build-time C
x86, vdso: Move syscall and sysenter setup into kernel/cpu/common.c
x86, vdso: Clean up 32-bit vs 64-bit vdso params
x86, mm: Ensure correct alignment of the fixmap
After booting with cgroup_disable=memory, I still saw memcg files
in the default hierarchy, and I can write to them, though it won't
take effect.
# dmesg
...
Disabling memory control group subsystem
...
# mount -t cgroup -o __DEVEL__sane_behavior xxx /cgroup
# ls /cgroup
...
memory.failcnt memory.move_charge_at_immigrate
memory.force_empty memory.numa_stat
memory.limit_in_bytes memory.oom_control
...
# cat /cgroup/memory.usage_in_bytes
0
tj: Minor comment update.
Signed-off-by: Li Zefan <lizefan@huawei.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
copy_insn() fails with -EIO if ->readpage == NULL, but this error
is not propagated unless uprobe_register() path finds ->mm which
already mmaps this file. In this case (say) "perf record" does not
actually install the probe, but the user can't know about this.
Move this check into uprobe_register() so that this problem can be
detected earlier and reported to user.
Note: this is still not perfect,
- copy_insn() and arch_uprobe_analyze_insn() should be called
by uprobe_register() but this is not simple, we need vm_file
for read_mapping_page() (although perhaps we can pass NULL),
and we need ->mm for is_64bit_mm() (although this logic is
broken anyway).
- uprobe_register() should be called by create_trace_uprobe(),
not by probe_event_enable(), so that an error can be detected
at "perf probe -x" time. This also needs more changes in the
core uprobe code, uprobe register/unregister interface was
poorly designed from the very beginning.
Reported-by: Denys Vlasenko <dvlasenk@redhat.com>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/20140519184054.GA6750@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
sched_cfs_period_timer() reads cfs_b->period without locks before calling
do_sched_cfs_period_timer(), and similarly unthrottle_offline_cfs_rqs()
would read cfs_b->period without the right lock. Thus a simultaneous
change of bandwidth could cause corruption on any platform where ktime_t
or u64 writes/reads are not atomic.
Extend cfs_b->lock from do_sched_cfs_period_timer() to include the read of
cfs_b->period to solve that issue; unthrottle_offline_cfs_rqs() can just
use 1 rather than the exact quota, much like distribute_cfs_runtime()
does.
There is also an unlocked read of cfs_b->runtime_expires, but a race
there would only delay runtime expiry by a tick. Still, the comparison
should just be != anyway, which clarifies even that problem.
Signed-off-by: Ben Segall <bsegall@google.com>
Tested-by: Roman Gushchin <klamm@yandex-team.ru>
[peterz: Fix compile warn]
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20140519224945.20303.93530.stgit@sword-of-the-dawn.mtv.corp.google.com
Cc: pjt@google.com
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
tg_set_cfs_bandwidth() sets cfs_b->timer_active to 0 to
force the period timer restart. It's not safe, because
can lead to deadlock, described in commit 927b54fccb:
"__start_cfs_bandwidth calls hrtimer_cancel while holding rq->lock,
waiting for the hrtimer to finish. However, if sched_cfs_period_timer
runs for another loop iteration, the hrtimer can attempt to take
rq->lock, resulting in deadlock."
Three CPUs must be involved:
CPU0 CPU1 CPU2
take rq->lock period timer fired
... take cfs_b lock
... ... tg_set_cfs_bandwidth()
throttle_cfs_rq() release cfs_b lock take cfs_b lock
... distribute_cfs_runtime() timer_active = 0
take cfs_b->lock wait for rq->lock ...
__start_cfs_bandwidth()
{wait for timer callback
break if timer_active == 1}
So, CPU0 and CPU1 are deadlocked.
Instead of resetting cfs_b->timer_active, tg_set_cfs_bandwidth can
wait for period timer callbacks (ignoring cfs_b->timer_active) and
restart the timer explicitly.
Signed-off-by: Roman Gushchin <klamm@yandex-team.ru>
Reviewed-by: Ben Segall <bsegall@google.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/87wqdi9g8e.wl\%klamm@yandex-team.ru
Cc: pjt@google.com
Cc: chris.j.arges@canonical.com
Cc: gregkh@linuxfoundation.org
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Throttled task is still on rq, and it may be moved to other cpu
if user is playing with sched_setaffinity(). Therefore, unlocked
task_rq() access makes the race.
Juri Lelli reports he got this race when dl_bandwidth_enabled()
was not set.
Other thing, pointed by Peter Zijlstra:
"Now I suppose the problem can still actually happen when
you change the root domain and trigger a effective affinity
change that way".
To fix that we do the same as made in __task_rq_lock(). We do not
use __task_rq_lock() itself, because it has a useful lockdep check,
which is not correct in case of dl_task_timer(). We do not need
pi_lock locked here. This case is an exception (PeterZ):
"The only reason we don't strictly need ->pi_lock now is because
we're guaranteed to have p->state == TASK_RUNNING here and are
thus free of ttwu races".
Signed-off-by: Kirill Tkhai <tkhai@yandex.ru>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: <stable@vger.kernel.org> # v3.14+
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/3056991400578422@web14g.yandex.ru
Signed-off-by: Ingo Molnar <mingo@kernel.org>
As Peter Zijlstra told me, we have the following path:
do_exit()
exit_itimers()
itimer_delete()
spin_lock_irqsave(&timer->it_lock, &flags);
timer_delete_hook(timer);
kc->timer_del(timer) := posix_cpu_timer_del()
put_task_struct()
__put_task_struct()
task_numa_free()
spin_lock(&grp->lock);
Which means that task_numa_free() can be called with interrupts
disabled, which means that we should not be using spin_lock_irq() but
spin_lock_irqsave() instead. Otherwise we are enabling interrupts while
holding an interrupt unsafe lock!
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner<tglx@linutronix.de>
Cc: Mike Galbraith <umgwanakikbuti@gmail.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/20140527182541.GH11096@twins.programming.kicks-ass.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We have reached the point where our mutexes are quite fine tuned
for a number of situations. This includes the use of heuristics
and optimistic spinning, based on MCS locking techniques.
Exclusive ownership of read-write semaphores are, conceptually,
just about the same as mutexes, making them close cousins. To
this end we need to make them both perform similarly, and
right now, rwsems are simply not up to it. This was discovered
by both reverting commit 4fc3f1d6 (mm/rmap, migration: Make
rmap_walk_anon() and try_to_unmap_anon() more scalable) and
similarly, converting some other mutexes (ie: i_mmap_mutex) to
rwsems. This creates a situation where users have to choose
between a rwsem and mutex taking into account this important
performance difference. Specifically, biggest difference between
both locks is when we fail to acquire a mutex in the fastpath,
optimistic spinning comes in to play and we can avoid a large
amount of unnecessary sleeping and overhead of moving tasks in
and out of wait queue. Rwsems do not have such logic.
This patch, based on the work from Tim Chen and I, adds support
for write-side optimistic spinning when the lock is contended.
It also includes support for the recently added cancelable MCS
locking for adaptive spinning. Note that is is only applicable
to the xadd method, and the spinlock rwsem variant remains intact.
Allowing optimistic spinning before putting the writer on the wait
queue reduces wait queue contention and provided greater chance
for the rwsem to get acquired. With these changes, rwsem is on par
with mutex. The performance benefits can be seen on a number of
workloads. For instance, on a 8 socket, 80 core 64bit Westmere box,
aim7 shows the following improvements in throughput:
+--------------+---------------------+-----------------+
| Workload | throughput-increase | number of users |
+--------------+---------------------+-----------------+
| alltests | 20% | >1000 |
| custom | 27%, 60% | 10-100, >1000 |
| high_systime | 36%, 30% | >100, >1000 |
| shared | 58%, 29% | 10-100, >1000 |
+--------------+---------------------+-----------------+
There was also improvement on smaller systems, such as a quad-core
x86-64 laptop running a 30Gb PostgreSQL (pgbench) workload for up
to +60% in throughput for over 50 clients. Additionally, benefits
were also noticed in exim (mail server) workloads. Furthermore, no
performance regression have been seen at all.
Based-on-work-from: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
[peterz: rej fixup due to comment patches, sched/rt.h header]
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Alex Shi <alex.shi@linaro.org>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Hurley <peter@hurleysoftware.com>
Cc: "Paul E.McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Jason Low <jason.low2@hp.com>
Cc: Aswin Chandramouleeswaran <aswin@hp.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: "Scott J Norton" <scott.norton@hp.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Chris Mason <clm@fb.com>
Cc: Josef Bacik <jbacik@fusionio.com>
Link: http://lkml.kernel.org/r/1399055055.6275.15.camel@buesod1.americas.hpqcorp.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
