Use try_cmpxchg instead of cmpxchg (*ptr, old, new) != old in
set_nr_{and_not,if}_polling. x86 cmpxchg returns success in ZF flag,
so this change saves a compare after cmpxchg.
The definition of cmpxchg based fetch_or was changed in the
same way as atomic_fetch_##op definitions were changed
in e6790e4b5d.
Also declare these two functions as inline to ensure inlining. In the
case of set_nr_and_not_polling, the compiler (gcc) tries to outsmart
itself by constructing the boolean return value with logic operations
on the fetched value, and these extra operations enlarge the function
over the inlining threshold value.
Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20220629151552.6015-1-ubizjak@gmail.com
4feee7d126 previously added per-task forced idle accounting. This patch
extends this to also include cgroups.
rstat is used for cgroup accounting, except for the root, which uses
kcpustat in order to bypass the need for doing an rstat flush when
reading root stats.
Only cgroup v2 is supported. Similar to the task accounting, the cgroup
accounting requires that schedstats is enabled.
Signed-off-by: Josh Don <joshdon@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Tejun Heo <tj@kernel.org>
Link: https://lkml.kernel.org/r/20220629211426.3329954-1-joshdon@google.com
find_energy_efficient_cpu() integrates a margin to protect tasks from
bouncing back and forth from a CPU to another. This margin is set as being
6% of the total current energy estimated on the system. This however does
not work for two reasons:
1. The energy estimation is not a good absolute value:
compute_energy() used in feec() is a good estimation for task placement as
it allows to compare the energy with and without a task. The computed
delta will give a good overview of the cost for a certain task placement.
It, however, doesn't work as an absolute estimation for the total energy
of the system. First it adds the contribution to idle CPUs into the
energy, second it mixes util_avg with util_est values. util_avg contains
the near history for a CPU usage, it doesn't tell at all what the current
utilization is. A system that has been quite busy in the near past will
hold a very high energy and then a high margin preventing any task
migration to a lower capacity CPU, wasting energy. It even creates a
negative feedback loop: by holding the tasks on a less efficient CPU, the
margin contributes in keeping the energy high.
2. The margin handicaps small tasks:
On a system where the workload is composed mostly of small tasks (which is
often the case on Android), the overall energy will be high enough to
create a margin none of those tasks can cross. On a Pixel4, a small
utilization of 5% on all the CPUs creates a global estimated energy of 140
joules, as per the Energy Model declaration of that same device. This
means, after applying the 6% margin that any migration must save more than
8 joules to happen. No task with a utilization lower than 40 would then be
able to migrate away from the biggest CPU of the system.
The 6% of the overall system energy was brought by the following patch:
(eb92692b25 sched/fair: Speed-up energy-aware wake-ups)
It was previously 6% of the prev_cpu energy. Also, the following one
made this margin value conditional on the clusters where the task fits:
(8d4c97c105 sched/fair: Only compute base_energy_pd if necessary)
We could simply revert that margin change to what it was, but the original
version didn't have strong grounds neither and as demonstrated in (1.) the
estimated energy isn't a good absolute value. Instead, removing it
completely. It is indeed, made possible by recent changes that improved
energy estimation comparison fairness (sched/fair: Remove task_util from
effective utilization in feec()) (PM: EM: Increase energy calculation
precision) and task utilization stabilization (sched/fair: Decay task
util_avg during migration)
Without a margin, we could have feared bouncing between CPUs. But running
LISA's eas_behaviour test coverage on three different platforms (Hikey960,
RB-5 and DB-845) showed no issue.
Removing the energy margin enables more energy-optimized placements for a
more energy efficient system.
Signed-off-by: Vincent Donnefort <vincent.donnefort@arm.com>
Signed-off-by: Vincent Donnefort <vdonnefort@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Tested-by: Lukasz Luba <lukasz.luba@arm.com>
Link: https://lkml.kernel.org/r/20220621090414.433602-8-vdonnefort@google.com
The energy estimation in find_energy_efficient_cpu() (feec()) relies on
the computation of the effective utilization for each CPU of a perf domain
(PD). This effective utilization is then used as an estimation of the busy
time for this pd. The function effective_cpu_util() which gives this value,
scales the utilization relative to IRQ pressure on the CPU to take into
account that the IRQ time is hidden from the task clock. The IRQ scaling is
as follow:
effective_cpu_util = irq + (cpu_cap - irq)/cpu_cap * util
Where util is the sum of CFS/RT/DL utilization, cpu_cap the capacity of
the CPU and irq the IRQ avg time.
If now we take as an example a task placement which doesn't raise the OPP
on the candidate CPU, we can write the energy delta as:
delta = OPPcost/cpu_cap * (effective_cpu_util(cpu_util + task_util) -
effective_cpu_util(cpu_util))
= OPPcost/cpu_cap * (cpu_cap - irq)/cpu_cap * task_util
We end-up with an energy delta depending on the IRQ avg time, which is a
problem: first the time spent on IRQs by a CPU has no effect on the
additional energy that would be consumed by a task. Second, we don't want
to favour a CPU with a higher IRQ avg time value.
Nonetheless, we need to take the IRQ avg time into account. If a task
placement raises the PD's frequency, it will increase the energy cost for
the entire time where the CPU is busy. A solution is to only use
effective_cpu_util() with the CPU contribution part. The task contribution
is added separately and scaled according to prev_cpu's IRQ time.
No change for the FREQUENCY_UTIL component of the energy estimation. We
still want to get the actual frequency that would be selected after the
task placement.
Signed-off-by: Vincent Donnefort <vincent.donnefort@arm.com>
Signed-off-by: Vincent Donnefort <vdonnefort@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Tested-by: Lukasz Luba <lukasz.luba@arm.com>
Link: https://lkml.kernel.org/r/20220621090414.433602-7-vdonnefort@google.com
The Perf Domain (PD) cpumask (struct em_perf_domain.cpus) stays
invariant after Energy Model creation, i.e. it is not updated after
CPU hotplug operations.
That's why the PD mask is used in conjunction with the cpu_online_mask
(or Sched Domain cpumask). Thereby the cpu_online_mask is fetched
multiple times (in compute_energy()) during a run-queue selection
for a task.
cpu_online_mask may change during this time which can lead to wrong
energy calculations.
To be able to avoid this, use the select_rq_mask per-cpu cpumask to
create a cpumask out of PD cpumask and cpu_online_mask and pass it
through the function calls of the EAS run-queue selection path.
The PD cpumask for max_spare_cap_cpu/compute_prev_delta selection
(find_energy_efficient_cpu()) is now ANDed not only with the SD mask
but also with the cpu_online_mask. This is fine since this cpumask
has to be in syc with the one used for energy computation
(compute_energy()).
An exclusive cpuset setup with at least one asymmetric CPU capacity
island (hence the additional AND with the SD cpumask) is the obvious
exception here.
Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Lukasz Luba <lukasz.luba@arm.com>
Link: https://lkml.kernel.org/r/20220621090414.433602-6-vdonnefort@google.com
On 21/06/2022 11:04, Vincent Donnefort wrote:
> From: Dietmar Eggemann <dietmar.eggemann@arm.com>
https://lkml.kernel.org/r/202206221253.ZVyGQvPX-lkp@intel.com discovered
that this patch doesn't build anymore (on tip sched/core or linux-next)
because of commit f5b2eeb499 ("sched/fair: Consider CPU affinity when
allowing NUMA imbalance in find_idlest_group()").
New version of [PATCH v11 4/7] sched/fair: Rename select_idle_mask to
select_rq_mask below.
-- >8 --
Decouple the name of the per-cpu cpumask select_idle_mask from its usage
in select_idle_[cpu/capacity]() of the CFS run-queue selection
(select_task_rq_fair()).
This is to support the reuse of this cpumask in the Energy Aware
Scheduling (EAS) path (find_energy_efficient_cpu()) of the CFS run-queue
selection.
Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Lukasz Luba <lukasz.luba@arm.com>
Link: https://lkml.kernel.org/r/250691c7-0e2b-05ab-bedf-b245c11d9400@arm.com
effective_cpu_util() already has a `int cpu' parameter which allows to
retrieve the CPU capacity scale factor (or maximum CPU capacity) inside
this function via an arch_scale_cpu_capacity(cpu).
A lot of code calling effective_cpu_util() (or the shim
sched_cpu_util()) needs the maximum CPU capacity, i.e. it will call
arch_scale_cpu_capacity() already.
But not having to pass it into effective_cpu_util() will make the EAS
wake-up code easier, especially when the maximum CPU capacity reduced
by the thermal pressure is passed through the EAS wake-up functions.
Due to the asymmetric CPU capacity support of arm/arm64 architectures,
arch_scale_cpu_capacity(int cpu) is a per-CPU variable read access via
per_cpu(cpu_scale, cpu) on such a system.
On all other architectures it is a a compile-time constant
(SCHED_CAPACITY_SCALE).
Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Lukasz Luba <lukasz.luba@arm.com>
Link: https://lkml.kernel.org/r/20220621090414.433602-4-vdonnefort@google.com
Before being migrated to a new CPU, a task sees its PELT values
synchronized with rq last_update_time. Once done, that same task will also
have its sched_avg last_update_time reset. This means the time between
the migration and the last clock update will not be accounted for in
util_avg and a discontinuity will appear. This issue is amplified by the
PELT clock scaling. It takes currently one tick after the CPU being idle
to let clock_pelt catching up clock_task.
This is especially problematic for asymmetric CPU capacity systems which
need stable util_avg signals for task placement and energy estimation.
Ideally, this problem would be solved by updating the runqueue clocks
before the migration. But that would require taking the runqueue lock
which is quite expensive [1]. Instead estimate the missing time and update
the task util_avg with that value.
To that end, we need sched_clock_cpu() but it is a costly function. Limit
the usage to the case where the source CPU is idle as we know this is when
the clock is having the biggest risk of being outdated.
See comment in migrate_se_pelt_lag() for more details about how the PELT
value is estimated. Notice though this estimation doesn't take into account
IRQ and Paravirt time.
[1] https://lkml.kernel.org/r/20190709115759.10451-1-chris.redpath@arm.com
Signed-off-by: Vincent Donnefort <vincent.donnefort@arm.com>
Signed-off-by: Vincent Donnefort <vdonnefort@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Tested-by: Lukasz Luba <lukasz.luba@arm.com>
Link: https://lkml.kernel.org/r/20220621090414.433602-3-vdonnefort@google.com
Introducing macro helpers u64_u32_{store,load}() to factorize lockless
accesses to u64 variables for 32-bits architectures.
Users are for now cfs_rq.min_vruntime and sched_avg.last_update_time. To
accommodate the later where the copy lies outside of the structure
(cfs_rq.last_udpate_time_copy instead of sched_avg.last_update_time_copy),
use the _copy() version of those helpers.
Those new helpers encapsulate smp_rmb() and smp_wmb() synchronization and
therefore, have a small penalty for 32-bits machines in set_task_rq_fair()
and init_cfs_rq().
Signed-off-by: Vincent Donnefort <vincent.donnefort@arm.com>
Signed-off-by: Vincent Donnefort <vdonnefort@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Tested-by: Lukasz Luba <lukasz.luba@arm.com>
Link: https://lkml.kernel.org/r/20220621090414.433602-2-vdonnefort@google.com
[Problem Statement]
select_idle_cpu() might spend too much time searching for an idle CPU,
when the system is overloaded.
The following histogram is the time spent in select_idle_cpu(),
when running 224 instances of netperf on a system with 112 CPUs
per LLC domain:
@usecs:
[0] 533 | |
[1] 5495 | |
[2, 4) 12008 | |
[4, 8) 239252 | |
[8, 16) 4041924 |@@@@@@@@@@@@@@ |
[16, 32) 12357398 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[32, 64) 14820255 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[64, 128) 13047682 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[128, 256) 8235013 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[256, 512) 4507667 |@@@@@@@@@@@@@@@ |
[512, 1K) 2600472 |@@@@@@@@@ |
[1K, 2K) 927912 |@@@ |
[2K, 4K) 218720 | |
[4K, 8K) 98161 | |
[8K, 16K) 37722 | |
[16K, 32K) 6715 | |
[32K, 64K) 477 | |
[64K, 128K) 7 | |
netperf latency usecs:
=======
case load Lat_99th std%
TCP_RR thread-224 257.39 ( 0.21)
The time spent in select_idle_cpu() is visible to netperf and might have a negative
impact.
[Symptom analysis]
The patch [1] from Mel Gorman has been applied to track the efficiency
of select_idle_sibling. Copy the indicators here:
SIS Search Efficiency(se_eff%):
A ratio expressed as a percentage of runqueues scanned versus
idle CPUs found. A 100% efficiency indicates that the target,
prev or recent CPU of a task was idle at wakeup. The lower the
efficiency, the more runqueues were scanned before an idle CPU
was found.
SIS Domain Search Efficiency(dom_eff%):
Similar, except only for the slower SIS
patch.
SIS Fast Success Rate(fast_rate%):
Percentage of SIS that used target, prev or
recent CPUs.
SIS Success rate(success_rate%):
Percentage of scans that found an idle CPU.
The test is based on Aubrey's schedtests tool, including netperf, hackbench,
schbench and tbench.
Test on vanilla kernel:
schedstat_parse.py -f netperf_vanilla.log
case load se_eff% dom_eff% fast_rate% success_rate%
TCP_RR 28 threads 99.978 18.535 99.995 100.000
TCP_RR 56 threads 99.397 5.671 99.964 100.000
TCP_RR 84 threads 21.721 6.818 73.632 100.000
TCP_RR 112 threads 12.500 5.533 59.000 100.000
TCP_RR 140 threads 8.524 4.535 49.020 100.000
TCP_RR 168 threads 6.438 3.945 40.309 99.999
TCP_RR 196 threads 5.397 3.718 32.320 99.982
TCP_RR 224 threads 4.874 3.661 25.775 99.767
UDP_RR 28 threads 99.988 17.704 99.997 100.000
UDP_RR 56 threads 99.528 5.977 99.970 100.000
UDP_RR 84 threads 24.219 6.992 76.479 100.000
UDP_RR 112 threads 13.907 5.706 62.538 100.000
UDP_RR 140 threads 9.408 4.699 52.519 100.000
UDP_RR 168 threads 7.095 4.077 44.352 100.000
UDP_RR 196 threads 5.757 3.775 35.764 99.991
UDP_RR 224 threads 5.124 3.704 28.748 99.860
schedstat_parse.py -f schbench_vanilla.log
(each group has 28 tasks)
case load se_eff% dom_eff% fast_rate% success_rate%
normal 1 mthread 99.152 6.400 99.941 100.000
normal 2 mthreads 97.844 4.003 99.908 100.000
normal 3 mthreads 96.395 2.118 99.917 99.998
normal 4 mthreads 55.288 1.451 98.615 99.804
normal 5 mthreads 7.004 1.870 45.597 61.036
normal 6 mthreads 3.354 1.346 20.777 34.230
normal 7 mthreads 2.183 1.028 11.257 21.055
normal 8 mthreads 1.653 0.825 7.849 15.549
schedstat_parse.py -f hackbench_vanilla.log
(each group has 28 tasks)
case load se_eff% dom_eff% fast_rate% success_rate%
process-pipe 1 group 99.991 7.692 99.999 100.000
process-pipe 2 groups 99.934 4.615 99.997 100.000
process-pipe 3 groups 99.597 3.198 99.987 100.000
process-pipe 4 groups 98.378 2.464 99.958 100.000
process-pipe 5 groups 27.474 3.653 89.811 99.800
process-pipe 6 groups 20.201 4.098 82.763 99.570
process-pipe 7 groups 16.423 4.156 77.398 99.316
process-pipe 8 groups 13.165 3.920 72.232 98.828
process-sockets 1 group 99.977 5.882 99.999 100.000
process-sockets 2 groups 99.927 5.505 99.996 100.000
process-sockets 3 groups 99.397 3.250 99.980 100.000
process-sockets 4 groups 79.680 4.258 98.864 99.998
process-sockets 5 groups 7.673 2.503 63.659 92.115
process-sockets 6 groups 4.642 1.584 58.946 88.048
process-sockets 7 groups 3.493 1.379 49.816 81.164
process-sockets 8 groups 3.015 1.407 40.845 75.500
threads-pipe 1 group 99.997 0.000 100.000 100.000
threads-pipe 2 groups 99.894 2.932 99.997 100.000
threads-pipe 3 groups 99.611 4.117 99.983 100.000
threads-pipe 4 groups 97.703 2.624 99.937 100.000
threads-pipe 5 groups 22.919 3.623 87.150 99.764
threads-pipe 6 groups 18.016 4.038 80.491 99.557
threads-pipe 7 groups 14.663 3.991 75.239 99.247
threads-pipe 8 groups 12.242 3.808 70.651 98.644
threads-sockets 1 group 99.990 6.667 99.999 100.000
threads-sockets 2 groups 99.940 5.114 99.997 100.000
threads-sockets 3 groups 99.469 4.115 99.977 100.000
threads-sockets 4 groups 87.528 4.038 99.400 100.000
threads-sockets 5 groups 6.942 2.398 59.244 88.337
threads-sockets 6 groups 4.359 1.954 49.448 87.860
threads-sockets 7 groups 2.845 1.345 41.198 77.102
threads-sockets 8 groups 2.871 1.404 38.512 74.312
schedstat_parse.py -f tbench_vanilla.log
case load se_eff% dom_eff% fast_rate% success_rate%
loopback 28 threads 99.976 18.369 99.995 100.000
loopback 56 threads 99.222 7.799 99.934 100.000
loopback 84 threads 19.723 6.819 70.215 100.000
loopback 112 threads 11.283 5.371 55.371 99.999
loopback 140 threads 0.000 0.000 0.000 0.000
loopback 168 threads 0.000 0.000 0.000 0.000
loopback 196 threads 0.000 0.000 0.000 0.000
loopback 224 threads 0.000 0.000 0.000 0.000
According to the test above, if the system becomes busy, the
SIS Search Efficiency(se_eff%) drops significantly. Although some
benchmarks would finally find an idle CPU(success_rate% = 100%), it is
doubtful whether it is worth it to search the whole LLC domain.
[Proposal]
It would be ideal to have a crystal ball to answer this question:
How many CPUs must a wakeup path walk down, before it can find an idle
CPU? Many potential metrics could be used to predict the number.
One candidate is the sum of util_avg in this LLC domain. The benefit
of choosing util_avg is that it is a metric of accumulated historic
activity, which seems to be smoother than instantaneous metrics
(such as rq->nr_running). Besides, choosing the sum of util_avg
would help predict the load of the LLC domain more precisely, because
SIS_PROP uses one CPU's idle time to estimate the total LLC domain idle
time.
In summary, the lower the util_avg is, the more select_idle_cpu()
should scan for idle CPU, and vice versa. When the sum of util_avg
in this LLC domain hits 85% or above, the scan stops. The reason to
choose 85% as the threshold is that this is the imbalance_pct(117)
when a LLC sched group is overloaded.
Introduce the quadratic function:
y = SCHED_CAPACITY_SCALE - p * x^2
and y'= y / SCHED_CAPACITY_SCALE
x is the ratio of sum_util compared to the CPU capacity:
x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
y' is the ratio of CPUs to be scanned in the LLC domain,
and the number of CPUs to scan is calculated by:
nr_scan = llc_weight * y'
Choosing quadratic function is because:
[1] Compared to the linear function, it scans more aggressively when the
sum_util is low.
[2] Compared to the exponential function, it is easier to calculate.
[3] It seems that there is no accurate mapping between the sum of util_avg
and the number of CPUs to be scanned. Use heuristic scan for now.
For a platform with 112 CPUs per LLC, the number of CPUs to scan is:
sum_util% 0 5 15 25 35 45 55 65 75 85 86 ...
scan_nr 112 111 108 102 93 81 65 47 25 1 0 ...
For a platform with 16 CPUs per LLC, the number of CPUs to scan is:
sum_util% 0 5 15 25 35 45 55 65 75 85 86 ...
scan_nr 16 15 15 14 13 11 9 6 3 0 0 ...
Furthermore, to minimize the overhead of calculating the metrics in
select_idle_cpu(), borrow the statistics from periodic load balance.
As mentioned by Abel, on a platform with 112 CPUs per LLC, the
sum_util calculated by periodic load balance after 112 ms would
decay to about 0.5 * 0.5 * 0.5 * 0.7 = 8.75%, thus bringing a delay
in reflecting the latest utilization. But it is a trade-off.
Checking the util_avg in newidle load balance would be more frequent,
but it brings overhead - multiple CPUs write/read the per-LLC shared
variable and introduces cache contention. Tim also mentioned that,
it is allowed to be non-optimal in terms of scheduling for the
short-term variations, but if there is a long-term trend in the load
behavior, the scheduler can adjust for that.
When SIS_UTIL is enabled, the select_idle_cpu() uses the nr_scan
calculated by SIS_UTIL instead of the one from SIS_PROP. As Peter and
Mel suggested, SIS_UTIL should be enabled by default.
This patch is based on the util_avg, which is very sensitive to the
CPU frequency invariance. There is an issue that, when the max frequency
has been clamp, the util_avg would decay insanely fast when
the CPU is idle. Commit addca28512 ("cpufreq: intel_pstate: Handle no_turbo
in frequency invariance") could be used to mitigate this symptom, by adjusting
the arch_max_freq_ratio when turbo is disabled. But this issue is still
not thoroughly fixed, because the current code is unaware of the user-specified
max CPU frequency.
[Test result]
netperf and tbench were launched with 25% 50% 75% 100% 125% 150%
175% 200% of CPU number respectively. Hackbench and schbench were launched
by 1, 2 ,4, 8 groups. Each test lasts for 100 seconds and repeats 3 times.
The following is the benchmark result comparison between
baseline:vanilla v5.19-rc1 and compare:patched kernel. Positive compare%
indicates better performance.
Each netperf test is a:
netperf -4 -H 127.0.1 -t TCP/UDP_RR -c -C -l 100
netperf.throughput
=======
case load baseline(std%) compare%( std%)
TCP_RR 28 threads 1.00 ( 0.34) -0.16 ( 0.40)
TCP_RR 56 threads 1.00 ( 0.19) -0.02 ( 0.20)
TCP_RR 84 threads 1.00 ( 0.39) -0.47 ( 0.40)
TCP_RR 112 threads 1.00 ( 0.21) -0.66 ( 0.22)
TCP_RR 140 threads 1.00 ( 0.19) -0.69 ( 0.19)
TCP_RR 168 threads 1.00 ( 0.18) -0.48 ( 0.18)
TCP_RR 196 threads 1.00 ( 0.16) +194.70 ( 16.43)
TCP_RR 224 threads 1.00 ( 0.16) +197.30 ( 7.85)
UDP_RR 28 threads 1.00 ( 0.37) +0.35 ( 0.33)
UDP_RR 56 threads 1.00 ( 11.18) -0.32 ( 0.21)
UDP_RR 84 threads 1.00 ( 1.46) -0.98 ( 0.32)
UDP_RR 112 threads 1.00 ( 28.85) -2.48 ( 19.61)
UDP_RR 140 threads 1.00 ( 0.70) -0.71 ( 14.04)
UDP_RR 168 threads 1.00 ( 14.33) -0.26 ( 11.16)
UDP_RR 196 threads 1.00 ( 12.92) +186.92 ( 20.93)
UDP_RR 224 threads 1.00 ( 11.74) +196.79 ( 18.62)
Take the 224 threads as an example, the SIS search metrics changes are
illustrated below:
vanilla patched
4544492 +237.5% 15338634 sched_debug.cpu.sis_domain_search.avg
38539 +39686.8% 15333634 sched_debug.cpu.sis_failed.avg
128300000 -87.9% 15551326 sched_debug.cpu.sis_scanned.avg
5842896 +162.7% 15347978 sched_debug.cpu.sis_search.avg
There is -87.9% less CPU scans after patched, which indicates lower overhead.
Besides, with this patch applied, there is -13% less rq lock contention
in perf-profile.calltrace.cycles-pp._raw_spin_lock.raw_spin_rq_lock_nested
.try_to_wake_up.default_wake_function.woken_wake_function.
This might help explain the performance improvement - Because this patch allows
the waking task to remain on the previous CPU, rather than grabbing other CPUs'
lock.
Each hackbench test is a:
hackbench -g $job --process/threads --pipe/sockets -l 1000000 -s 100
hackbench.throughput
=========
case load baseline(std%) compare%( std%)
process-pipe 1 group 1.00 ( 1.29) +0.57 ( 0.47)
process-pipe 2 groups 1.00 ( 0.27) +0.77 ( 0.81)
process-pipe 4 groups 1.00 ( 0.26) +1.17 ( 0.02)
process-pipe 8 groups 1.00 ( 0.15) -4.79 ( 0.02)
process-sockets 1 group 1.00 ( 0.63) -0.92 ( 0.13)
process-sockets 2 groups 1.00 ( 0.03) -0.83 ( 0.14)
process-sockets 4 groups 1.00 ( 0.40) +5.20 ( 0.26)
process-sockets 8 groups 1.00 ( 0.04) +3.52 ( 0.03)
threads-pipe 1 group 1.00 ( 1.28) +0.07 ( 0.14)
threads-pipe 2 groups 1.00 ( 0.22) -0.49 ( 0.74)
threads-pipe 4 groups 1.00 ( 0.05) +1.88 ( 0.13)
threads-pipe 8 groups 1.00 ( 0.09) -4.90 ( 0.06)
threads-sockets 1 group 1.00 ( 0.25) -0.70 ( 0.53)
threads-sockets 2 groups 1.00 ( 0.10) -0.63 ( 0.26)
threads-sockets 4 groups 1.00 ( 0.19) +11.92 ( 0.24)
threads-sockets 8 groups 1.00 ( 0.08) +4.31 ( 0.11)
Each tbench test is a:
tbench -t 100 $job 127.0.0.1
tbench.throughput
======
case load baseline(std%) compare%( std%)
loopback 28 threads 1.00 ( 0.06) -0.14 ( 0.09)
loopback 56 threads 1.00 ( 0.03) -0.04 ( 0.17)
loopback 84 threads 1.00 ( 0.05) +0.36 ( 0.13)
loopback 112 threads 1.00 ( 0.03) +0.51 ( 0.03)
loopback 140 threads 1.00 ( 0.02) -1.67 ( 0.19)
loopback 168 threads 1.00 ( 0.38) +1.27 ( 0.27)
loopback 196 threads 1.00 ( 0.11) +1.34 ( 0.17)
loopback 224 threads 1.00 ( 0.11) +1.67 ( 0.22)
Each schbench test is a:
schbench -m $job -t 28 -r 100 -s 30000 -c 30000
schbench.latency_90%_us
========
case load baseline(std%) compare%( std%)
normal 1 mthread 1.00 ( 31.22) -7.36 ( 20.25)*
normal 2 mthreads 1.00 ( 2.45) -0.48 ( 1.79)
normal 4 mthreads 1.00 ( 1.69) +0.45 ( 0.64)
normal 8 mthreads 1.00 ( 5.47) +9.81 ( 14.28)
*Consider the Standard Deviation, this -7.36% regression might not be valid.
Also, a OLTP workload with a commercial RDBMS has been tested, and there
is no significant change.
There were concerns that unbalanced tasks among CPUs would cause problems.
For example, suppose the LLC domain is composed of 8 CPUs, and 7 tasks are
bound to CPU0~CPU6, while CPU7 is idle:
CPU0 CPU1 CPU2 CPU3 CPU4 CPU5 CPU6 CPU7
util_avg 1024 1024 1024 1024 1024 1024 1024 0
Since the util_avg ratio is 87.5%( = 7/8 ), which is higher than 85%,
select_idle_cpu() will not scan, thus CPU7 is undetected during scan.
But according to Mel, it is unlikely the CPU7 will be idle all the time
because CPU7 could pull some tasks via CPU_NEWLY_IDLE.
lkp(kernel test robot) has reported a regression on stress-ng.sock on a
very busy system. According to the sched_debug statistics, it might be caused
by SIS_UTIL terminates the scan and chooses a previous CPU earlier, and this
might introduce more context switch, especially involuntary preemption, which
impacts a busy stress-ng. This regression has shown that, not all benchmarks
in every scenario benefit from idle CPU scan limit, and it needs further
investigation.
Besides, there is slight regression in hackbench's 16 groups case when the
LLC domain has 16 CPUs. Prateek mentioned that we should scan aggressively
in an LLC domain with 16 CPUs. Because the cost to search for an idle one
among 16 CPUs is negligible. The current patch aims to propose a generic
solution and only considers the util_avg. Something like the below could
be applied on top of the current patch to fulfill the requirement:
if (llc_weight <= 16)
nr_scan = nr_scan * 32 / llc_weight;
For LLC domain with 16 CPUs, the nr_scan will be expanded to 2 times large.
The smaller the CPU number this LLC domain has, the larger nr_scan will be
expanded. This needs further investigation.
There is also ongoing work[2] from Abel to filter out the busy CPUs during
wakeup, to further speed up the idle CPU scan. And it could be a following-up
optimization on top of this change.
Suggested-by: Tim Chen <tim.c.chen@intel.com>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Yicong Yang <yangyicong@hisilicon.com>
Tested-by: Mohini Narkhede <mohini.narkhede@intel.com>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://lore.kernel.org/r/20220612163428.849378-1-yu.c.chen@intel.com
sched_setattr(2) issues via kernel/sched/core.c:__sched_setscheduler()
a CAP_SYS_NICE audit event unconditionally, even when the requested
operation does not require that capability / is unprivileged, i.e. for
reducing niceness.
This is relevant in connection with SELinux, where a capability check
results in a policy decision and by default a denial message on
insufficient permission is issued.
It can lead to three undesired cases:
1. A denial message is generated, even in case the operation was an
unprivileged one and thus the syscall succeeded, creating noise.
2. To avoid the noise from 1. the policy writer adds a rule to ignore
those denial messages, hiding future syscalls, where the task
performs an actual privileged operation, leading to hidden limited
functionality of that task.
3. To avoid the noise from 1. the policy writer adds a rule to allow
the task the capability CAP_SYS_NICE, while it does not need it,
violating the principle of least privilege.
Conduct privilged/unprivileged categorization first and perform a
capable test (and at most once) only if needed.
Signed-off-by: Christian Göttsche <cgzones@googlemail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20220615152505.310488-1-cgzones@googlemail.com
Wakelist can help avoid cache bouncing and offload the overhead of waker
cpu. So far, using wakelist within the same llc only happens on
WF_ON_CPU, and this limitation could be removed to further improve
wakeup performance.
The commit 518cd62341 ("sched: Only queue remote wakeups when
crossing cache boundaries") disabled queuing tasks on wakelist when
the cpus share llc. This is because, at that time, the scheduler must
send IPIs to do ttwu_queue_wakelist. Nowadays, ttwu_queue_wakelist also
supports TIF_POLLING, so this is not a problem now when the wakee cpu is
in idle polling.
Benefits:
Queuing the task on idle cpu can help improving performance on waker cpu
and utilization on wakee cpu, and further improve locality because
the wakee cpu can handle its own rq. This patch helps improving rt on
our real java workloads where wakeup happens frequently.
Consider the normal condition (CPU0 and CPU1 share same llc)
Before this patch:
CPU0 CPU1
select_task_rq() idle
rq_lock(CPU1->rq)
enqueue_task(CPU1->rq)
notify CPU1 (by sending IPI or CPU1 polling)
resched()
After this patch:
CPU0 CPU1
select_task_rq() idle
add to wakelist of CPU1
notify CPU1 (by sending IPI or CPU1 polling)
rq_lock(CPU1->rq)
enqueue_task(CPU1->rq)
resched()
We see CPU0 can finish its work earlier. It only needs to put task to
wakelist and return.
While CPU1 is idle, so let itself handle its own runqueue data.
This patch brings no difference about IPI.
This patch only takes effect when the wakee cpu is:
1) idle polling
2) idle not polling
For 1), there will be no IPI with or without this patch.
For 2), there will always be an IPI before or after this patch.
Before this patch: waker cpu will enqueue task and check preempt. Since
"idle" will be sure to be preempted, waker cpu must send a resched IPI.
After this patch: waker cpu will put the task to the wakelist of wakee
cpu, and send an IPI.
Benchmark:
We've tested schbench, unixbench, and hachbench on both x86 and arm64.
On x86 (Intel Xeon Platinum 8269CY):
schbench -m 2 -t 8
Latency percentiles (usec) before after
50.0000th: 8 6
75.0000th: 10 7
90.0000th: 11 8
95.0000th: 12 8
*99.0000th: 13 10
99.5000th: 15 11
99.9000th: 18 14
Unixbench with full threads (104)
before after
Dhrystone 2 using register variables 3011862938 3009935994 -0.06%
Double-Precision Whetstone 617119.3 617298.5 0.03%
Execl Throughput 27667.3 27627.3 -0.14%
File Copy 1024 bufsize 2000 maxblocks 785871.4 784906.2 -0.12%
File Copy 256 bufsize 500 maxblocks 210113.6 212635.4 1.20%
File Copy 4096 bufsize 8000 maxblocks 2328862.2 2320529.1 -0.36%
Pipe Throughput 145535622.8 145323033.2 -0.15%
Pipe-based Context Switching 3221686.4 3583975.4 11.25%
Process Creation 101347.1 103345.4 1.97%
Shell Scripts (1 concurrent) 120193.5 123977.8 3.15%
Shell Scripts (8 concurrent) 17233.4 17138.4 -0.55%
System Call Overhead 5300604.8 5312213.6 0.22%
hackbench -g 1 -l 100000
before after
Time 3.246 2.251
On arm64 (Ampere Altra):
schbench -m 2 -t 8
Latency percentiles (usec) before after
50.0000th: 14 10
75.0000th: 19 14
90.0000th: 22 16
95.0000th: 23 16
*99.0000th: 24 17
99.5000th: 24 17
99.9000th: 28 25
Unixbench with full threads (80)
before after
Dhrystone 2 using register variables 3536194249 3537019613 0.02%
Double-Precision Whetstone 629383.6 629431.6 0.01%
Execl Throughput 65920.5 65846.2 -0.11%
File Copy 1024 bufsize 2000 maxblocks 1063722.8 1064026.8 0.03%
File Copy 256 bufsize 500 maxblocks 322684.5 318724.5 -1.23%
File Copy 4096 bufsize 8000 maxblocks 2348285.3 2328804.8 -0.83%
Pipe Throughput 133542875.3 131619389.8 -1.44%
Pipe-based Context Switching 3215356.1 3576945.1 11.25%
Process Creation 108520.5 120184.6 10.75%
Shell Scripts (1 concurrent) 122636.3 121888 -0.61%
Shell Scripts (8 concurrent) 17462.1 17381.4 -0.46%
System Call Overhead 4429998.9 4435006.7 0.11%
hackbench -g 1 -l 100000
before after
Time 4.217 2.916
Our patch has improvement on schbench, hackbench
and Pipe-based Context Switching of unixbench
when there exists idle cpus,
and no obvious regression on other tests of unixbench.
This can help improve rt in scenes where wakeup happens frequently.
Signed-off-by: Tianchen Ding <dtcccc@linux.alibaba.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <vschneid@redhat.com>
Link: https://lore.kernel.org/r/20220608233412.327341-3-dtcccc@linux.alibaba.com
While doing newidle load balancing, it is possible for new tasks to
arrive, such as with pending wakeups. newidle_balance() already accounts
for this by exiting the sched_domain load_balance() iteration if it
detects these cases. This is very important for minimizing wakeup
latency.
However, if we are already in load_balance(), we may stay there for a
while before returning back to newidle_balance(). This is most
exacerbated if we enter a 'goto redo' loop in the LBF_ALL_PINNED case. A
very straightforward workaround to this is to adjust should_we_balance()
to bail out if we're doing a CPU_NEWLY_IDLE balance and new tasks are
detected.
This was tested with the following reproduction:
- two threads that take turns sleeping and waking each other up are
affined to two cores
- a large number of threads with 100% utilization are pinned to all
other cores
Without this patch, wakeup latency was ~120us for the pair of threads,
almost entirely spent in load_balance(). With this patch, wakeup latency
is ~6us.
Signed-off-by: Josh Don <joshdon@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20220609025515.2086253-1-joshdon@google.com
sysctl_sched_dl_period_max and sysctl_sched_dl_period_min are unsigned
integer, but proc_dointvec() wouldn't return error even if we set a
negative number.
Use proc_douintvec_minmax() instead of proc_dointvec(). Add extra1 for
sysctl_sched_dl_period_max and extra2 for sysctl_sched_dl_period_min.
It's just an optimization for match data and proc_handler in struct
ctl_table. The 'if (period < min || period > max)' in __checkparam_dl()
will work fine even if there hasn't this patch.
Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Daniel Bristot de Oliveira <bristot@kernel.org>
Link: https://lore.kernel.org/r/20220607101807.249965-1-yajun.deng@linux.dev
We notice the rq leaf_cfs_rq_list has two problems when do bugfix
backports and some test profiling.
1. cfs_rqs under throttled subtree could be added to the list, and
make their fully decayed ancestors on the list, even though not needed.
2. #1 also make the leaf_cfs_rq_list management complex and error prone,
this is the list of related bugfix so far:
commit 31bc6aeaab ("sched/fair: Optimize update_blocked_averages()")
commit fe61468b2c ("sched/fair: Fix enqueue_task_fair warning")
commit b34cb07dde ("sched/fair: Fix enqueue_task_fair() warning some more")
commit 39f23ce07b ("sched/fair: Fix unthrottle_cfs_rq() for leaf_cfs_rq list")
commit 0258bdfaff ("sched/fair: Fix unfairness caused by missing load decay")
commit a7b359fc6a ("sched/fair: Correctly insert cfs_rq's to list on unthrottle")
commit fdaba61ef8 ("sched/fair: Ensure that the CFS parent is added after unthrottling")
commit 2630cde267 ("sched/fair: Add ancestors of unthrottled undecayed cfs_rq")
commit 31bc6aeaab ("sched/fair: Optimize update_blocked_averages()")
delete every cfs_rq under throttled subtree from rq->leaf_cfs_rq_list,
and delete the throttled_hierarchy() test in update_blocked_averages(),
which optimized update_blocked_averages().
But those later bugfix add cfs_rqs under throttled subtree back to
rq->leaf_cfs_rq_list again, with their fully decayed ancestors, for
the integrity of rq->leaf_cfs_rq_list.
This patch takes another method, skip all cfs_rqs under throttled
hierarchy when list_add_leaf_cfs_rq(), to completely make cfs_rqs
under throttled subtree off the leaf_cfs_rq_list.
So we don't need to consider throttled related things in
enqueue_entity(), unthrottle_cfs_rq() and enqueue_task_fair(),
which simplify the code a lot. Also optimize update_blocked_averages()
since cfs_rqs under throttled hierarchy and their ancestors
won't be on the leaf_cfs_rq_list.
Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20220601021848.76943-1-zhouchengming@bytedance.com
In the case of systems containing multiple LLCs per socket, like
AMD Zen systems, users want to spread bandwidth hungry applications
across multiple LLCs. Stream is one such representative workload where
the best performance is obtained by limiting one stream thread per LLC.
To ensure this, users are known to pin the tasks to a specify a subset
of the CPUs consisting of one CPU per LLC while running such bandwidth
hungry tasks.
Suppose we kickstart a multi-threaded task like stream with 8 threads
using taskset or numactl to run on a subset of CPUs on a 2 socket Zen3
server where each socket contains 128 CPUs
(0-63,128-191 in one socket, 64-127,192-255 in another socket)
Eg: numactl -C 0,16,32,48,64,80,96,112 ./stream8
Here each CPU in the list is from a different LLC and 4 of those LLCs
are on one socket, while the other 4 are on another socket.
Ideally we would prefer that each stream thread runs on a different
CPU from the allowed list of CPUs. However, the current heuristics in
find_idlest_group() do not allow this during the initial placement.
Suppose the first socket (0-63,128-191) is our local group from which
we are kickstarting the stream tasks. The first four stream threads
will be placed in this socket. When it comes to placing the 5th
thread, all the allowed CPUs are from the local group (0,16,32,48)
would have been taken.
However, the current scheduler code simply checks if the number of
tasks in the local group is fewer than the allowed numa-imbalance
threshold. This threshold was previously 25% of the NUMA domain span
(in this case threshold = 32) but after the v6 of Mel's patchset
"Adjust NUMA imbalance for multiple LLCs", got merged in sched-tip,
Commit: e496132ebe ("sched/fair: Adjust the allowed NUMA imbalance
when SD_NUMA spans multiple LLCs") it is now equal to number of LLCs
in the NUMA domain, for processors with multiple LLCs.
(in this case threshold = 8).
For this example, the number of tasks will always be within threshold
and thus all the 8 stream threads will be woken up on the first socket
thereby resulting in sub-optimal performance.
The following sched_wakeup_new tracepoint output shows the initial
placement of tasks in the current tip/sched/core on the Zen3 machine:
stream-5313 [016] d..2. 627.005036: sched_wakeup_new: comm=stream pid=5315 prio=120 target_cpu=032
stream-5313 [016] d..2. 627.005086: sched_wakeup_new: comm=stream pid=5316 prio=120 target_cpu=048
stream-5313 [016] d..2. 627.005141: sched_wakeup_new: comm=stream pid=5317 prio=120 target_cpu=000
stream-5313 [016] d..2. 627.005183: sched_wakeup_new: comm=stream pid=5318 prio=120 target_cpu=016
stream-5313 [016] d..2. 627.005218: sched_wakeup_new: comm=stream pid=5319 prio=120 target_cpu=016
stream-5313 [016] d..2. 627.005256: sched_wakeup_new: comm=stream pid=5320 prio=120 target_cpu=016
stream-5313 [016] d..2. 627.005295: sched_wakeup_new: comm=stream pid=5321 prio=120 target_cpu=016
Once the first four threads are distributed among the allowed CPUs of
socket one, the rest of the treads start piling on these same CPUs
when clearly there are CPUs on the second socket that can be used.
Following the initial pile up on a small number of CPUs, though the
load-balancer eventually kicks in, it takes a while to get to {4}{4}
and even {4}{4} isn't stable as we observe a bunch of ping ponging
between {4}{4} to {5}{3} and back before a stable state is reached
much later (1 Stream thread per allowed CPU) and no more migration is
required.
We can detect this piling and avoid it by checking if the number of
allowed CPUs in the local group are fewer than the number of tasks
running in the local group and use this information to spread the
5th task out into the next socket (after all, the goal in this
slowpath is to find the idlest group and the idlest CPU during the
initial placement!).
The following sched_wakeup_new tracepoint output shows the initial
placement of tasks after adding this fix on the Zen3 machine:
stream-4485 [016] d..2. 230.784046: sched_wakeup_new: comm=stream pid=4487 prio=120 target_cpu=032
stream-4485 [016] d..2. 230.784123: sched_wakeup_new: comm=stream pid=4488 prio=120 target_cpu=048
stream-4485 [016] d..2. 230.784167: sched_wakeup_new: comm=stream pid=4489 prio=120 target_cpu=000
stream-4485 [016] d..2. 230.784222: sched_wakeup_new: comm=stream pid=4490 prio=120 target_cpu=112
stream-4485 [016] d..2. 230.784271: sched_wakeup_new: comm=stream pid=4491 prio=120 target_cpu=096
stream-4485 [016] d..2. 230.784322: sched_wakeup_new: comm=stream pid=4492 prio=120 target_cpu=080
stream-4485 [016] d..2. 230.784368: sched_wakeup_new: comm=stream pid=4493 prio=120 target_cpu=064
We see that threads are using all of the allowed CPUs and there is
no pileup.
No output is generated for tracepoint sched_migrate_task with this
patch due to a perfect initial placement which removes the need
for balancing later on - both across NUMA boundaries and within
NUMA boundaries for stream.
Following are the results from running 8 Stream threads with and
without pinning on a dual socket Zen3 Machine (2 x 64C/128T):
During the testing of this patch, the tip sched/core was at
commit: 089c02ae27 "ftrace: Use preemption model accessors for trace
header printout"
Pinning is done using: numactl -C 0,16,32,48,64,80,96,112 ./stream8
5.18.0-rc1 5.18.0-rc1 5.18.0-rc1
tip sched/core tip sched/core tip sched/core
(no pinning) + pinning + this-patch
+ pinning
Copy: 109364.74 (0.00 pct) 94220.50 (-13.84 pct) 158301.28 (44.74 pct)
Scale: 109670.26 (0.00 pct) 90210.59 (-17.74 pct) 149525.64 (36.34 pct)
Add: 129029.01 (0.00 pct) 101906.00 (-21.02 pct) 186658.17 (44.66 pct)
Triad: 127260.05 (0.00 pct) 106051.36 (-16.66 pct) 184327.30 (44.84 pct)
Pinning currently hurts the performance compared to unbound case on
tip/sched/core. With the addition of this patch, we are able to
outperform tip/sched/core by a good margin with pinning.
Following are the results from running 16 Stream threads with and
without pinning on a dual socket IceLake Machine (2 x 32C/64T):
NUMA Topology of Intel Skylake machine:
Node 1: 0,2,4,6 ... 126 (Even numbers)
Node 2: 1,3,5,7 ... 127 (Odd numbers)
Pinning is done using: numactl -C 0-15 ./stream16
5.18.0-rc1 5.18.0-rc1 5.18.0-rc1
tip sched/core tip sched/core tip sched/core
(no pinning) +pinning + this-patch
+ pinning
Copy: 85815.31 (0.00 pct) 149819.21 (74.58 pct) 156807.48 (82.72 pct)
Scale: 64795.60 (0.00 pct) 97595.07 (50.61 pct) 99871.96 (54.13 pct)
Add: 71340.68 (0.00 pct) 111549.10 (56.36 pct) 114598.33 (60.63 pct)
Triad: 68890.97 (0.00 pct) 111635.16 (62.04 pct) 114589.24 (66.33 pct)
In case of Icelake machine, with single LLC per socket, pinning across
the two sockets reduces cache contention, thus showing great
improvement in pinned case which is further benefited by this patch.
Signed-off-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Link: https://lkml.kernel.org/r/20220407111222.22649-1-kprateek.nayak@amd.com
For a single LLC per node, a NUMA imbalance is allowed up until 25%
of CPUs sharing a node could be active. One intent of the cut-off is
to avoid an imbalance of memory channels but there is no topological
information based on active memory channels. Furthermore, there can
be differences between nodes depending on the number of populated
DIMMs.
A cut-off of 25% was arbitrary but generally worked. It does have a severe
corner cases though when an parallel workload is using 25% of all available
CPUs over-saturates memory channels. This can happen due to the initial
forking of tasks that get pulled more to one node after early wakeups
(e.g. a barrier synchronisation) that is not quickly corrected by the
load balancer. The LB may fail to act quickly as the parallel tasks are
considered to be poor migrate candidates due to locality or cache hotness.
On a range of modern Intel CPUs, 12.5% appears to be a better cut-off
assuming all memory channels are populated and is used as the new cut-off
point. A minimum of 1 is specified to allow a communicating pair to
remain local even for CPUs with low numbers of cores. For modern AMDs,
there are multiple LLCs and are not affected.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://lore.kernel.org/r/20220520103519.1863-5-mgorman@techsingularity.net
The imbalance limitations are applied inconsistently at fork time
and at runtime. At fork, a new task can remain local until there are
too many running tasks even if the degree of imbalance is larger than
NUMA_IMBALANCE_MIN which is different to runtime. Secondly, the imbalance
figure used during load balancing is different to the one used at NUMA
placement. Load balancing uses the number of tasks that must move to
restore imbalance where as NUMA balancing uses the total imbalance.
In combination, it is possible for a parallel workload that uses a small
number of CPUs without applying scheduler policies to have very variable
run-to-run performance.
[lkp@intel.com: Fix build breakage for arc-allyesconfig]
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://lore.kernel.org/r/20220520103519.1863-4-mgorman@techsingularity.net
Pull workqueue fixes from Tejun Heo:
"Tetsuo's patch to trigger build warnings if system-wide wq's are
flushed along with a TP type update and trivial comment update"
* tag 'wq-for-5.19-rc1-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq:
workqueue: Switch to new kerneldoc syntax for named variable macro argument
workqueue: Fix type of cpu in trace event
workqueue: Wrap flush_workqueue() using a macro
Pull power management fixes from Rafael Wysocki:
"These fix an intel_idle issue introduced during the 5.16 development
cycle and two recent regressions in the system reboot/poweroff code.
Specifics:
- Fix CPUIDLE_FLAG_IRQ_ENABLE handling in intel_idle (Peter Zijlstra)
- Allow all platforms to use the global poweroff handler and make
non-syscall poweroff code paths work again (Dmitry Osipenko)"
* tag 'pm-5.19-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
cpuidle,intel_idle: Fix CPUIDLE_FLAG_IRQ_ENABLE
kernel/reboot: Fix powering off using a non-syscall code paths
kernel/reboot: Use static handler for register_platform_power_off()
Merge fixes for regressions introduced by the recent rework of the
system reboot/poweroff code.
* pm-sysoff:
kernel/reboot: Fix powering off using a non-syscall code paths
kernel/reboot: Use static handler for register_platform_power_off()
Pull xen updates from Juergen Gross:
- a small cleanup removing "export" of an __init function
- a small series adding a new infrastructure for platform flags
- a series adding generic virtio support for Xen guests (frontend side)
* tag 'for-linus-5.19a-rc2-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/xen/tip:
xen: unexport __init-annotated xen_xlate_map_ballooned_pages()
arm/xen: Assign xen-grant DMA ops for xen-grant DMA devices
xen/grant-dma-ops: Retrieve the ID of backend's domain for DT devices
xen/grant-dma-iommu: Introduce stub IOMMU driver
dt-bindings: Add xen,grant-dma IOMMU description for xen-grant DMA ops
xen/virtio: Enable restricted memory access using Xen grant mappings
xen/grant-dma-ops: Add option to restrict memory access under Xen
xen/grants: support allocating consecutive grants
arm/xen: Introduce xen_setup_dma_ops()
virtio: replace arch_has_restricted_virtio_memory_access()
kernel: add platform_has() infrastructure
Pull KVM fixes from Paolo Bonzini:
- syzkaller NULL pointer dereference
- TDP MMU performance issue with disabling dirty logging
- 5.14 regression with SVM TSC scaling
- indefinite stall on applying live patches
- unstable selftest
- memory leak from wrong copy-and-paste
- missed PV TLB flush when racing with emulation
* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm:
KVM: x86: do not report a vCPU as preempted outside instruction boundaries
KVM: x86: do not set st->preempted when going back to user space
KVM: SVM: fix tsc scaling cache logic
KVM: selftests: Make hyperv_clock selftest more stable
KVM: x86/MMU: Zap non-leaf SPTEs when disabling dirty logging
x86: drop bogus "cc" clobber from __try_cmpxchg_user_asm()
KVM: x86/mmu: Check every prev_roots in __kvm_mmu_free_obsolete_roots()
entry/kvm: Exit to user mode when TIF_NOTIFY_SIGNAL is set
KVM: Don't null dereference ops->destroy
There are other methods of powering off machine than the reboot syscall.
Previously we missed to cover those methods and it created power-off
regression for some machines, like the PowerPC e500.
Fix this problem by moving the legacy sys-off handler registration to
the latest phase of power-off process and making the kernel_can_power_off()
check the legacy pm_power_off presence.
Tested-by: Michael Ellerman <mpe@ellerman.id.au> # ppce500
Reported-by: Michael Ellerman <mpe@ellerman.id.au> # ppce500
Fixes: da007f171f ("kernel/reboot: Change registration order of legacy power-off handler")
Signed-off-by: Dmitry Osipenko <dmitry.osipenko@collabora.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The verifier allows programs to call global functions as long as their
argument types match, using BTF to check the function arguments. One of the
allowed argument types to such global functions is PTR_TO_CTX; however the
check for this fails on BPF_PROG_TYPE_EXT functions because the verifier
uses the wrong type to fetch the vmlinux BTF ID for the program context
type. This failure is seen when an XDP program is loaded using
libxdp (which loads it as BPF_PROG_TYPE_EXT and attaches it to a global XDP
type program).
Fix the issue by passing in the target program type instead of the
BPF_PROG_TYPE_EXT type to bpf_prog_get_ctx() when checking function
argument compatibility.
The first Fixes tag refers to the latest commit that touched the code in
question, while the second one points to the code that first introduced
the global function call verification.
v2:
- Use resolve_prog_type()
Fixes: 3363bd0cfb ("bpf: Extend kfunc with PTR_TO_CTX, PTR_TO_MEM argument support")
Fixes: 51c39bb1d5 ("bpf: Introduce function-by-function verification")
Reported-by: Simon Sundberg <simon.sundberg@kau.se>
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/r/20220606075253.28422-1-toke@redhat.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The kvmalloc_array() function is safer because it has a check for
integer overflows. These sizes come from the user and I was not
able to see any bounds checking so an integer overflow seems like a
realistic concern.
Fixes: 0dcac27254 ("bpf: Add multi kprobe link")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/Yo9VRVMeHbALyjUH@kili
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Since flush operation synchronously waits for completion, flushing
system-wide WQs (e.g. system_wq) might introduce possibility of deadlock
due to unexpected locking dependency. Tejun Heo commented at [1] that it
makes no sense at all to call flush_workqueue() on the shared WQs as the
caller has no idea what it's gonna end up waiting for.
Although there is flush_scheduled_work() which flushes system_wq WQ with
"Think twice before calling this function! It's very easy to get into
trouble if you don't take great care." warning message, syzbot found a
circular locking dependency caused by flushing system_wq WQ [2].
Therefore, let's change the direction to that developers had better use
their local WQs if flush_scheduled_work()/flush_workqueue(system_*_wq) is
inevitable.
Steps for converting system-wide WQs into local WQs are explained at [3],
and a conversion to stop flushing system-wide WQs is in progress. Now we
want some mechanism for preventing developers who are not aware of this
conversion from again start flushing system-wide WQs.
Since I found that WARN_ON() is complete but awkward approach for teaching
developers about this problem, let's use __compiletime_warning() for
incomplete but handy approach. For completeness, we will also insert
WARN_ON() into __flush_workqueue() after all in-tree users stopped calling
flush_scheduled_work().
Link: https://lore.kernel.org/all/YgnQGZWT%2Fn3VAITX@slm.duckdns.org/ [1]
Link: https://syzkaller.appspot.com/bug?extid=bde0f89deacca7c765b8 [2]
Link: https://lkml.kernel.org/r/49925af7-78a8-a3dd-bce6-cfc02e1a9236@I-love.SAKURA.ne.jp [3]
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Tejun Heo <tj@kernel.org>
A livepatch transition may stall indefinitely when a kvm vCPU is heavily
loaded. To the host, the vCPU task is a user thread which is spending a
very long time in the ioctl(KVM_RUN) syscall. During livepatch
transition, set_notify_signal() will be called on such tasks to
interrupt the syscall so that the task can be transitioned. This
interrupts guest execution, but when xfer_to_guest_mode_work() sees that
TIF_NOTIFY_SIGNAL is set but not TIF_SIGPENDING it concludes that an
exit to user mode is unnecessary, and guest execution is resumed without
transitioning the task for the livepatch.
This handling of TIF_NOTIFY_SIGNAL is incorrect, as set_notify_signal()
is expected to break tasks out of interruptible kernel loops and cause
them to return to userspace. Change xfer_to_guest_mode_work() to handle
TIF_NOTIFY_SIGNAL the same as TIF_SIGPENDING, signaling to the vCPU run
loop that an exit to userpsace is needed. Any pending task_work will be
run when get_signal() is called from exit_to_user_mode_loop(), so there
is no longer any need to run task work from xfer_to_guest_mode_work().
Suggested-by: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Petr Mladek <pmladek@suse.com>
Signed-off-by: Seth Forshee <sforshee@digitalocean.com>
Message-Id: <20220504180840.2907296-1-sforshee@digitalocean.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Pull dma-mapping fixes from Christoph Hellwig:
- fix a regressin in setting swiotlb ->force_bounce (me)
- make dma-debug less chatty (Rob Clark)
* tag 'dma-mapping-5.19-2022-06-06' of git://git.infradead.org/users/hch/dma-mapping:
swiotlb: fix setting ->force_bounce
dma-debug: make things less spammy under memory pressure
Add a simple infrastructure for setting, resetting and querying
platform feature flags.
Flags can be either global or architecture specific.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com>
Tested-by: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com> # Arm64 only
Reviewed-by: Christoph Hellwig <hch@lst.de>
Acked-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Juergen Gross <jgross@suse.com>
Pull delay-accounting update from Andrew Morton:
"A single featurette for delay accounting.
Delayed a bit because, unusually, it had dependencies on both the
mm-stable and mm-nonmm-stable queues"
* tag 'mm-nonmm-stable-2022-06-05' of git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm:
delayacct: track delays from write-protect copy
Pull scheduler fix from Thomas Gleixner:
"Fix the fallout of sysctl code move which placed the init function
wrong"
* tag 'sched-urgent-2022-06-05' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/autogroup: Fix sysctl move
Pull perf fixes from Thomas Gleixner:
- Make the ICL event constraints match reality
- Remove a unused local variable
* tag 'perf-urgent-2022-06-05' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
perf/core: Remove unused local variable
perf/x86/intel: Fix event constraints for ICL
Pull mount handling updates from Al Viro:
"Cleanups (and one fix) around struct mount handling.
The fix is usermode_driver.c one - once you've done kern_mount(), you
must kern_unmount(); simple mntput() will end up with a leak. Several
failure exits in there messed up that way... In practice you won't hit
those particular failure exits without fault injection, though"
* tag 'pull-18-rc1-work.mount' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
move mount-related externs from fs.h to mount.h
blob_to_mnt(): kern_unmount() is needed to undo kern_mount()
m->mnt_root->d_inode->i_sb is a weird way to spell m->mnt_sb...
linux/mount.h: trim includes
uninline may_mount() and don't opencode it in fspick(2)/fsopen(2)
Pull ptrace_stop cleanups from Eric Biederman:
"While looking at the ptrace problems with PREEMPT_RT and the problems
Peter Zijlstra was encountering with ptrace in his freezer rewrite I
identified some cleanups to ptrace_stop that make sense on their own
and move make resolving the other problems much simpler.
The biggest issue is the habit of the ptrace code to change
task->__state from the tracer to suppress TASK_WAKEKILL from waking up
the tracee. No other code in the kernel does that and it is straight
forward to update signal_wake_up and friends to make that unnecessary.
Peter's task freezer sets frozen tasks to a new state TASK_FROZEN and
then it stores them by calling "wake_up_state(t, TASK_FROZEN)" relying
on the fact that all stopped states except the special stop states can
tolerate spurious wake up and recover their state.
The state of stopped and traced tasked is changed to be stored in
task->jobctl as well as in task->__state. This makes it possible for
the freezer to recover tasks in these special states, as well as
serving as a general cleanup. With a little more work in that
direction I believe TASK_STOPPED can learn to tolerate spurious wake
ups and become an ordinary stop state.
The TASK_TRACED state has to remain a special state as the registers
for a process are only reliably available when the process is stopped
in the scheduler. Fundamentally ptrace needs acess to the saved
register values of a task.
There are bunch of semi-random ptrace related cleanups that were found
while looking at these issues.
One cleanup that deserves to be called out is from commit 57b6de08b5
("ptrace: Admit ptrace_stop can generate spuriuos SIGTRAPs"). This
makes a change that is technically user space visible, in the handling
of what happens to a tracee when a tracer dies unexpectedly. According
to our testing and our understanding of userspace nothing cares that
spurious SIGTRAPs can be generated in that case"
* tag 'ptrace_stop-cleanup-for-v5.19' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm/user-namespace:
sched,signal,ptrace: Rework TASK_TRACED, TASK_STOPPED state
ptrace: Always take siglock in ptrace_resume
ptrace: Don't change __state
ptrace: Admit ptrace_stop can generate spuriuos SIGTRAPs
ptrace: Document that wait_task_inactive can't fail
ptrace: Reimplement PTRACE_KILL by always sending SIGKILL
signal: Use lockdep_assert_held instead of assert_spin_locked
ptrace: Remove arch_ptrace_attach
ptrace/xtensa: Replace PT_SINGLESTEP with TIF_SINGLESTEP
ptrace/um: Replace PT_DTRACE with TIF_SINGLESTEP
signal: Replace __group_send_sig_info with send_signal_locked
signal: Rename send_signal send_signal_locked
Pull kthread updates from Eric Biederman:
"This updates init and user mode helper tasks to be ordinary user mode
tasks.
Commit 40966e316f ("kthread: Ensure struct kthread is present for
all kthreads") caused init and the user mode helper threads that call
kernel_execve to have struct kthread allocated for them. This struct
kthread going away during execve in turned made a use after free of
struct kthread possible.
Here, commit 343f4c49f2 ("kthread: Don't allocate kthread_struct for
init and umh") is enough to fix the use after free and is simple
enough to be backportable.
The rest of the changes pass struct kernel_clone_args to clean things
up and cause the code to make sense.
In making init and the user mode helpers tasks purely user mode tasks
I ran into two complications. The function task_tick_numa was
detecting tasks without an mm by testing for the presence of
PF_KTHREAD. The initramfs code in populate_initrd_image was using
flush_delayed_fput to ensuere the closing of all it's file descriptors
was complete, and flush_delayed_fput does not work in a userspace
thread.
I have looked and looked and more complications and in my code review
I have not found any, and neither has anyone else with the code
sitting in linux-next"
* tag 'kthread-cleanups-for-v5.19' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm/user-namespace:
sched: Update task_tick_numa to ignore tasks without an mm
fork: Stop allowing kthreads to call execve
fork: Explicitly set PF_KTHREAD
init: Deal with the init process being a user mode process
fork: Generalize PF_IO_WORKER handling
fork: Explicity test for idle tasks in copy_thread
fork: Pass struct kernel_clone_args into copy_thread
kthread: Don't allocate kthread_struct for init and umh
Pull arm64 fixes from Catalin Marinas:
"Most of issues addressed were introduced during this merging window.
- Initialise jump labels before setup_machine_fdt(), needed by commit
f5bda35fba ("random: use static branch for crng_ready()").
- Sparse warnings: missing prototype, incorrect __user annotation.
- Skip SVE kselftest if not sufficient vector lengths supported"
* tag 'arm64-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux:
kselftest/arm64: signal: Skip SVE signal test if not enough VLs supported
arm64: Initialize jump labels before setup_machine_fdt()
arm64: hibernate: Fix syntax errors in comments
arm64: Remove the __user annotation for the restore_za_context() argument
ftrace/fgraph: fix increased missing-prototypes warnings
Pull networking fixes from Jakub Kicinski:
"Including fixes from bpf and netfilter.
Current release - new code bugs:
- af_packet: make sure to pull the MAC header, avoid skb panic in GSO
- ptp_clockmatrix: fix inverted logic in is_single_shot()
- netfilter: flowtable: fix missing FLOWI_FLAG_ANYSRC flag
- dt-bindings: net: adin: fix adi,phy-output-clock description syntax
- wifi: iwlwifi: pcie: rename CAUSE macro, avoid MIPS build warning
Previous releases - regressions:
- Revert "net: af_key: add check for pfkey_broadcast in function
pfkey_process"
- tcp: fix tcp_mtup_probe_success vs wrong snd_cwnd
- nf_tables: disallow non-stateful expression in sets earlier
- nft_limit: clone packet limits' cost value
- nf_tables: double hook unregistration in netns path
- ping6: fix ping -6 with interface name
Previous releases - always broken:
- sched: fix memory barriers to prevent skbs from getting stuck in
lockless qdiscs
- neigh: set lower cap for neigh_managed_work rearming, avoid
constantly scheduling the probe work
- bpf: fix probe read error on big endian in ___bpf_prog_run()
- amt: memory leak and error handling fixes
Misc:
- ipv6: expand & rename accept_unsolicited_na to accept_untracked_na"
* tag 'net-5.19-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net: (80 commits)
net/af_packet: make sure to pull mac header
net: add debug info to __skb_pull()
net: CONFIG_DEBUG_NET depends on CONFIG_NET
stmmac: intel: Add RPL-P PCI ID
net: stmmac: use dev_err_probe() for reporting mdio bus registration failure
tipc: check attribute length for bearer name
ice: fix access-beyond-end in the switch code
nfp: remove padding in nfp_nfdk_tx_desc
ax25: Fix ax25 session cleanup problems
net: usb: qmi_wwan: Add support for Cinterion MV31 with new baseline
sfc/siena: fix wrong tx channel offset with efx_separate_tx_channels
sfc/siena: fix considering that all channels have TX queues
socket: Don't use u8 type in uapi socket.h
net/sched: act_api: fix error code in tcf_ct_flow_table_fill_tuple_ipv6()
net: ping6: Fix ping -6 with interface name
macsec: fix UAF bug for real_dev
octeontx2-af: fix error code in is_valid_offset()
wifi: mac80211: fix use-after-free in chanctx code
bonding: guard ns_targets by CONFIG_IPV6
tcp: tcp_rtx_synack() can be called from process context
...
The register_platform_power_off() fails on m68k platform due to the
memory allocation error that happens at a very early boot time when
memory allocator isn't available yet. Fix it by using a static sys-off
handler for the platform-level power-off handlers.
Fixes: f0f7e5265b ("m68k: Switch to new sys-off handler API")
Reported-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Dmitry Osipenko <dmitry.osipenko@collabora.com>
Reviewed-by: Geert Uytterhoeven <geert@linux-m68k.org>
Tested-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Pull livepatching cleanup from Petr Mladek:
- Remove duplicated livepatch code [Christophe]
* tag 'livepatching-for-5.19' of git://git.kernel.org/pub/scm/linux/kernel/git/livepatching/livepatching:
livepatch: Remove klp_arch_set_pc() and asm/livepatch.h
Pull printk fixup from Petr Mladek:
- Revert inappropriate use of wake_up_interruptible_all() in printk()
* tag 'printk-for-5.19-fixup' of git://git.kernel.org/pub/scm/linux/kernel/git/printk/linux:
Revert "printk: wake up all waiters"
The swiotlb_init refactor messed up assigning ->force_bounce by doing
it in different places based on what caused the setting of the flag.
Fix this by passing the SWIOTLB_* flags to swiotlb_init_io_tlb_mem
and just setting it there.
Fixes: c6af2aa9ff ("swiotlb: make the swiotlb_init interface more useful")
Reported-by: Nathan Chancellor <nathan@kernel.org>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Nathan Chancellor <nathan@kernel.org>
Limit the error msg to avoid flooding the console. If you have a lot of
threads hitting this at once, they could have already gotten passed the
dma_debug_disabled() check before they get to the point of allocation
failure, resulting in quite a lot of this error message spamming the
log. Use pr_err_once() to limit that.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Signed-off-by: Christoph Hellwig <hch@lst.de>
