Dave Chinner reported the following on https://lkml.org/lkml/2015/3/1/226
Across the board the 4.0-rc1 numbers are much slower, and the degradation
is far worse when using the large memory footprint configs. Perf points
straight at the cause - this is from 4.0-rc1 on the "-o bhash=101073" config:
- 56.07% 56.07% [kernel] [k] default_send_IPI_mask_sequence_phys
- default_send_IPI_mask_sequence_phys
- 99.99% physflat_send_IPI_mask
- 99.37% native_send_call_func_ipi
smp_call_function_many
- native_flush_tlb_others
- 99.85% flush_tlb_page
ptep_clear_flush
try_to_unmap_one
rmap_walk
try_to_unmap
migrate_pages
migrate_misplaced_page
- handle_mm_fault
- 99.73% __do_page_fault
trace_do_page_fault
do_async_page_fault
+ async_page_fault
0.63% native_send_call_func_single_ipi
generic_exec_single
smp_call_function_single
This is showing excessive migration activity even though excessive
migrations are meant to get throttled. Normally, the scan rate is tuned
on a per-task basis depending on the locality of faults. However, if
migrations fail for any reason then the PTE scanner may scan faster if
the faults continue to be remote. This means there is higher system CPU
overhead and fault trapping at exactly the time we know that migrations
cannot happen. This patch tracks when migration failures occur and
slows the PTE scanner.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reported-by: Dave Chinner <david@fromorbit.com>
Tested-by: Dave Chinner <david@fromorbit.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
At least some gcc versions - validly afaict - warn about potentially
using max_group uninitialized: There's no way the compiler can prove
that the body of the conditional where it and max_faults get set/
updated gets executed; in fact, without knowing all the details of
other scheduler code, I can't prove this either.
Generally the necessary change would appear to be to clear max_group
prior to entering the inner loop, and break out of the outer loop when
it ends up being all clear after the inner one. This, however, seems
inefficient, and afaict the same effect can be achieved by exiting the
outer loop when max_faults is still zero after the inner loop.
[ mingo: changed the solution to zero initialization: uninitialized_var()
needs to die, as it's an actively dangerous construct: if in the future
a known-proven-good piece of code is changed to have a true, buggy
uninitialized variable, the compiler warning is then supressed...
The better long term solution is to clean up the code flow, so that
even simple minded compilers (and humans!) are able to read it without
getting a headache. ]
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/54C2139202000078000588F7@mail.emea.novell.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The original purpose of rq::skip_clock_update was to avoid 'costly' clock
updates for back to back wakeup-preempt pairs. The big problem with it
has always been that the rq variable is unaware of the context and
causes indiscrimiate clock skips.
Rework the entire thing and create a sense of context by only allowing
schedule() to skip clock updates. (XXX can we measure the cost of the
added store?)
By ensuring only schedule can ever skip an update, we guarantee we're
never more than 1 tick behind on the update.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: umgwanakikbuti@gmail.com
Link: http://lkml.kernel.org/r/20150105103554.432381549@infradead.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Child has the same decay_count as parent. If it's not zero,
we add it to parent's cfs_rq->removed_load:
wake_up_new_task()->set_task_cpu()->migrate_task_rq_fair().
Child's load is a just garbade after copying of parent,
it hasn't been on cfs_rq yet, and it must not be added to
cfs_rq::removed_load in migrate_task_rq_fair().
The patch moves sched_entity::avg::decay_count intialization
in sched_fork(). So, migrate_task_rq_fair() does not change
removed_load.
Signed-off-by: Kirill Tkhai <ktkhai@parallels.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Ben Segall <bsegall@google.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/1418644618.6074.13.camel@tkhai
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In __synchronize_entity_decay(), if "decays" happens to be zero,
se->avg.decay_count will not be zeroed, holding the positive value
assigned when dequeued last time.
This is problematic in the following case:
If this runnable task is CFS-balanced to other CPUs soon afterwards,
migrate_task_rq_fair() will treat it as a blocked task due to its
non-zero decay_count, thereby adding its load to cfs_rq->removed_load
wrongly.
Thus, we must zero se->avg.decay_count in this case as well.
Signed-off-by: Xunlei Pang <pang.xunlei@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Ben Segall <bsegall@google.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/1418745509-2609-1-git-send-email-pang.xunlei@linaro.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Commit caeb178c60 ("sched/fair: Make update_sd_pick_busiest() return
'true' on a busier sd") changes groups to be ranked in the order of
overloaded > imbalance > other, and busiest group is picked according
to this order.
sgs->group_capacity_factor is used to check if the group is overloaded.
When the child domain prefers tasks to go to siblings first, the
sgs->group_capacity_factor will be set lower than one in order to
move all the excess tasks away.
However, group overloaded status is not updated when
sgs->group_capacity_factor is set to lower than one, which leads to us
missing to find the busiest group.
This patch fixes it by updating group overloaded status when sg capacity
factor is set to one, in order to find the busiest group accurately.
Signed-off-by: Wanpeng Li <wanpeng.li@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Kirill Tkhai <ktkhai@parallels.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/1415144690-25196-1-git-send-email-wanpeng.li@linux.intel.com
[ Fixed the changelog. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Commit d670ec1317 "posix-cpu-timers: Cure SMP wobbles" fixes one glibc
test case in cost of breaking another one. After that commit, calling
clock_nanosleep(TIMER_ABSTIME, X) and then clock_gettime(&Y) can result
of Y time being smaller than X time.
Reproducer/tester can be found further below, it can be compiled and ran by:
gcc -o tst-cpuclock2 tst-cpuclock2.c -pthread
while ./tst-cpuclock2 ; do : ; done
This reproducer, when running on a buggy kernel, will complain
about "clock_gettime difference too small".
Issue happens because on start in thread_group_cputimer() we initialize
sum_exec_runtime of cputimer with threads runtime not yet accounted and
then add the threads runtime to running cputimer again on scheduler
tick, making it's sum_exec_runtime bigger than actual threads runtime.
KOSAKI Motohiro posted a fix for this problem, but that patch was never
applied: https://lkml.org/lkml/2013/5/26/191 .
This patch takes different approach to cure the problem. It calls
update_curr() when cputimer starts, that assure we will have updated
stats of running threads and on the next schedule tick we will account
only the runtime that elapsed from cputimer start. That also assure we
have consistent state between cpu times of individual threads and cpu
time of the process consisted by those threads.
Full reproducer (tst-cpuclock2.c):
#define _GNU_SOURCE
#include <unistd.h>
#include <sys/syscall.h>
#include <stdio.h>
#include <time.h>
#include <pthread.h>
#include <stdint.h>
#include <inttypes.h>
/* Parameters for the Linux kernel ABI for CPU clocks. */
#define CPUCLOCK_SCHED 2
#define MAKE_PROCESS_CPUCLOCK(pid, clock) \
((~(clockid_t) (pid) << 3) | (clockid_t) (clock))
static pthread_barrier_t barrier;
/* Help advance the clock. */
static void *chew_cpu(void *arg)
{
pthread_barrier_wait(&barrier);
while (1) ;
return NULL;
}
/* Don't use the glibc wrapper. */
static int do_nanosleep(int flags, const struct timespec *req)
{
clockid_t clock_id = MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED);
return syscall(SYS_clock_nanosleep, clock_id, flags, req, NULL);
}
static int64_t tsdiff(const struct timespec *before, const struct timespec *after)
{
int64_t before_i = before->tv_sec * 1000000000ULL + before->tv_nsec;
int64_t after_i = after->tv_sec * 1000000000ULL + after->tv_nsec;
return after_i - before_i;
}
int main(void)
{
int result = 0;
pthread_t th;
pthread_barrier_init(&barrier, NULL, 2);
if (pthread_create(&th, NULL, chew_cpu, NULL) != 0) {
perror("pthread_create");
return 1;
}
pthread_barrier_wait(&barrier);
/* The test. */
struct timespec before, after, sleeptimeabs;
int64_t sleepdiff, diffabs;
const struct timespec sleeptime = {.tv_sec = 0,.tv_nsec = 100000000 };
/* The relative nanosleep. Not sure why this is needed, but its presence
seems to make it easier to reproduce the problem. */
if (do_nanosleep(0, &sleeptime) != 0) {
perror("clock_nanosleep");
return 1;
}
/* Get the current time. */
if (clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &before) < 0) {
perror("clock_gettime[2]");
return 1;
}
/* Compute the absolute sleep time based on the current time. */
uint64_t nsec = before.tv_nsec + sleeptime.tv_nsec;
sleeptimeabs.tv_sec = before.tv_sec + nsec / 1000000000;
sleeptimeabs.tv_nsec = nsec % 1000000000;
/* Sleep for the computed time. */
if (do_nanosleep(TIMER_ABSTIME, &sleeptimeabs) != 0) {
perror("absolute clock_nanosleep");
return 1;
}
/* Get the time after the sleep. */
if (clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &after) < 0) {
perror("clock_gettime[3]");
return 1;
}
/* The time after sleep should always be equal to or after the absolute sleep
time passed to clock_nanosleep. */
sleepdiff = tsdiff(&sleeptimeabs, &after);
if (sleepdiff < 0) {
printf("absolute clock_nanosleep woke too early: %" PRId64 "\n", sleepdiff);
result = 1;
printf("Before %llu.%09llu\n", before.tv_sec, before.tv_nsec);
printf("After %llu.%09llu\n", after.tv_sec, after.tv_nsec);
printf("Sleep %llu.%09llu\n", sleeptimeabs.tv_sec, sleeptimeabs.tv_nsec);
}
/* The difference between the timestamps taken before and after the
clock_nanosleep call should be equal to or more than the duration of the
sleep. */
diffabs = tsdiff(&before, &after);
if (diffabs < sleeptime.tv_nsec) {
printf("clock_gettime difference too small: %" PRId64 "\n", diffabs);
result = 1;
}
pthread_cancel(th);
return result;
}
Signed-off-by: Stanislaw Gruszka <sgruszka@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/20141112155843.GA24803@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch simplifies task_struct by removing the four numa_* pointers
in the same array and replacing them with the array pointer. By doing this,
on x86_64, the size of task_struct is reduced by 3 ulong pointers (24 bytes on
x86_64).
A new parameter is added to the task_faults_idx function so that it can return
an index to the correct offset, corresponding with the old precalculated
pointers.
All of the code in sched/ that depended on task_faults_idx and numa_* was
changed in order to match the new logic.
Signed-off-by: Iulia Manda <iulia.manda21@gmail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: mgorman@suse.de
Cc: dave@stgolabs.net
Cc: riel@redhat.com
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/20141031001331.GA30662@winterfell
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In pseudo-interleaved numa_groups, all tasks try to relocate to
the group's preferred_nid. When a group is spread across multiple
NUMA nodes, this can lead to tasks swapping their location with
other tasks inside the same group, instead of swapping location with
tasks from other NUMA groups. This can keep NUMA groups from converging.
Examining all nodes, when dealing with a task in a pseudo-interleaved
NUMA group, avoids this problem. Note that only CPUs in nodes that
improve the task or group score are examined, so the loop isn't too
bad.
Tested-by: Vinod Chegu <chegu_vinod@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: "Vinod Chegu" <chegu_vinod@hp.com>
Cc: mgorman@suse.de
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/20141009172747.0d97c38c@annuminas.surriel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In order to do task placement on systems with complex NUMA topologies,
it is necessary to count the faults on nodes nearby the node that is
being examined for a potential move.
In case of a system with a backplane interconnect, we are dealing with
groups of NUMA nodes; each of the nodes within a group is the same number
of hops away from nodes in other groups in the system. Optimal placement
on this topology is achieved by counting all nearby nodes equally. When
comparing nodes A and B at distance N, nearby nodes are those at distances
smaller than N from nodes A or B.
Placement strategy on a system with a glueless mesh NUMA topology needs
to be different, because there are no natural groups of nodes determined
by the hardware. Instead, when dealing with two nodes A and B at distance
N, N >= 2, there will be intermediate nodes at distance < N from both nodes
A and B. Good placement can be achieved by right shifting the faults on
nearby nodes by the number of hops from the node being scored. In this
context, a nearby node is any node less than the maximum distance in the
system away from the node. Those nodes are skipped for efficiency reasons,
there is no real policy reason to do so.
Placement policy on directly connected NUMA systems is not affected.
Signed-off-by: Rik van Riel <riel@redhat.com>
Tested-by: Chegu Vinod <chegu_vinod@hp.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: mgorman@suse.de
Cc: chegu_vinod@hp.com
Link: http://lkml.kernel.org/r/1413530994-9732-5-git-send-email-riel@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
While offling node by hot removing memory, the following divide error
occurs:
divide error: 0000 [#1] SMP
[...]
Call Trace:
[...] handle_mm_fault
[...] ? try_to_wake_up
[...] ? wake_up_state
[...] __do_page_fault
[...] ? do_futex
[...] ? put_prev_entity
[...] ? __switch_to
[...] do_page_fault
[...] page_fault
[...]
RIP [<ffffffff810a7081>] task_numa_fault
RSP <ffff88084eb2bcb0>
The issue occurs as follows:
1. When page fault occurs and page is allocated from node 1,
task_struct->numa_faults_buffer_memory[] of node 1 is
incremented and p->numa_faults_locality[] is also incremented
as follows:
o numa_faults_buffer_memory[] o numa_faults_locality[]
NR_NUMA_HINT_FAULT_TYPES
| 0 | 1 |
---------------------------------- ----------------------
node 0 | 0 | 0 | remote | 0 |
node 1 | 0 | 1 | locale | 1 |
---------------------------------- ----------------------
2. node 1 is offlined by hot removing memory.
3. When page fault occurs, fault_types[] is calculated by using
p->numa_faults_buffer_memory[] of all online nodes in
task_numa_placement(). But node 1 was offline by step 2. So
the fault_types[] is calculated by using only
p->numa_faults_buffer_memory[] of node 0. So both of fault_types[]
are set to 0.
4. The values(0) of fault_types[] pass to update_task_scan_period().
5. numa_faults_locality[1] is set to 1. So the following division is
calculated.
static void update_task_scan_period(struct task_struct *p,
unsigned long shared, unsigned long private){
...
ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
}
6. But both of private and shared are set to 0. So divide error
occurs here.
The divide error is rare case because the trigger is node offline.
This patch always increments denominator for avoiding divide error.
Signed-off-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/54475703.8000505@jp.fujitsu.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Unlocked access to dst_rq->curr in task_numa_compare() is racy.
If curr task is exiting this may be a reason of use-after-free:
task_numa_compare() do_exit()
... current->flags |= PF_EXITING;
... release_task()
... ~~delayed_put_task_struct()~~
... schedule()
rcu_read_lock() ...
cur = ACCESS_ONCE(dst_rq->curr) ...
... rq->curr = next;
... context_switch()
... finish_task_switch()
... put_task_struct()
... __put_task_struct()
... free_task_struct()
task_numa_assign() ...
get_task_struct() ...
As noted by Oleg:
<<The lockless get_task_struct(tsk) is only safe if tsk == current
and didn't pass exit_notify(), or if this tsk was found on a rcu
protected list (say, for_each_process() or find_task_by_vpid()).
IOW, it is only safe if release_task() was not called before we
take rcu_read_lock(), in this case we can rely on the fact that
delayed_put_pid() can not drop the (potentially) last reference
until rcu_read_unlock().
And as Kirill pointed out task_numa_compare()->task_numa_assign()
path does get_task_struct(dst_rq->curr) and this is not safe. The
task_struct itself can't go away, but rcu_read_lock() can't save
us from the final put_task_struct() in finish_task_switch(); this
reference goes away without rcu gp>>
The patch provides simple check of PF_EXITING flag. If it's not set,
this guarantees that call_rcu() of delayed_put_task_struct() callback
hasn't happened yet, so we can safely do get_task_struct() in
task_numa_assign().
Locked dst_rq->lock protects from concurrency with the last schedule().
Reusing or unmapping of cur's memory may happen without it.
Suggested-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Kirill Tkhai <ktkhai@parallels.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/1413962231.19914.130.camel@tkhai
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull percpu consistent-ops changes from Tejun Heo:
"Way back, before the current percpu allocator was implemented, static
and dynamic percpu memory areas were allocated and handled separately
and had their own accessors. The distinction has been gone for many
years now; however, the now duplicate two sets of accessors remained
with the pointer based ones - this_cpu_*() - evolving various other
operations over time. During the process, we also accumulated other
inconsistent operations.
This pull request contains Christoph's patches to clean up the
duplicate accessor situation. __get_cpu_var() uses are replaced with
with this_cpu_ptr() and __this_cpu_ptr() with raw_cpu_ptr().
Unfortunately, the former sometimes is tricky thanks to C being a bit
messy with the distinction between lvalues and pointers, which led to
a rather ugly solution for cpumask_var_t involving the introduction of
this_cpu_cpumask_var_ptr().
This converts most of the uses but not all. Christoph will follow up
with the remaining conversions in this merge window and hopefully
remove the obsolete accessors"
* 'for-3.18-consistent-ops' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu: (38 commits)
irqchip: Properly fetch the per cpu offset
percpu: Resolve ambiguities in __get_cpu_var/cpumask_var_t -fix
ia64: sn_nodepda cannot be assigned to after this_cpu conversion. Use __this_cpu_write.
percpu: Resolve ambiguities in __get_cpu_var/cpumask_var_t
Revert "powerpc: Replace __get_cpu_var uses"
percpu: Remove __this_cpu_ptr
clocksource: Replace __this_cpu_ptr with raw_cpu_ptr
sparc: Replace __get_cpu_var uses
avr32: Replace __get_cpu_var with __this_cpu_write
blackfin: Replace __get_cpu_var uses
tile: Use this_cpu_ptr() for hardware counters
tile: Replace __get_cpu_var uses
powerpc: Replace __get_cpu_var uses
alpha: Replace __get_cpu_var
ia64: Replace __get_cpu_var uses
s390: cio driver &__get_cpu_var replacements
s390: Replace __get_cpu_var uses
mips: Replace __get_cpu_var uses
MIPS: Replace __get_cpu_var uses in FPU emulator.
arm: Replace __this_cpu_ptr with raw_cpu_ptr
...
Mel Gorman
Ingo Molnar
Jan Beulich
Peter Zijlstra
Kirill Tkhai
Xunlei Pang
Tetsuo Handa
Yuyang Du
Wanpeng Li
Stanislaw Gruszka
Iulia Manda
Yao Dongdong
Rik van Riel
Yasuaki Ishimatsu
Linus Torvalds