Hi Frederic,
Gave this series a spin on the same system as v1.
On 2/6/26 7:52 PM, Frederic Weisbecker wrote:
Hi,
After the issue reported here:
https://lore.kernel.org/all/[email protected]/
It occurs that the idle cputime accounting is a big mess that
accumulates within two concurrent statistics, each having their own
shortcomings:
* The accounting for online CPUs which is based on the delta between
tick_nohz_start_idle() and tick_nohz_stop_idle().
Pros:
- Works when the tick is off
- Has nsecs granularity
Cons:
- Account idle steal time but doesn't substract it from idle
cputime.
- Assumes CONFIG_IRQ_TIME_ACCOUNTING by not accounting IRQs but
the IRQ time is simply ignored when
CONFIG_IRQ_TIME_ACCOUNTING=n
- The windows between 1) idle task scheduling and the first call
to tick_nohz_start_idle() and 2) idle task between the last
tick_nohz_stop_idle() and the rest of the idle time are
blindspots wrt. cputime accounting (though mostly insignificant
amount)
- Relies on private fields outside of kernel stats, with specific
accessors.
* The accounting for offline CPUs which is based on ticks and the
jiffies delta during which the tick was stopped.
Pros:
- Handles steal time correctly
- Handle CONFIG_IRQ_TIME_ACCOUNTING=y and
CONFIG_IRQ_TIME_ACCOUNTING=n correctly.
- Handles the whole idle task
- Accounts directly to kernel stats, without midlayer accumulator.
Cons:
- Doesn't elapse when the tick is off, which doesn't make it
suitable for online CPUs.
- Has TICK_NSEC granularity (jiffies)
- Needs to track the dyntick-idle ticks that were accounted and
substract them from the total jiffies time spent while the tick
was stopped. This is an ugly workaround.
Having two different accounting for a single context is not the only
problem: since those accountings are of different natures, it is
possible to observe the global idle time going backward after a CPU goes
offline, as reported by Xin Zhao.
Clean up the situation with introducing a hybrid approach that stays
coherent, fixes the backward jumps and works for both online and offline
CPUs:
* Tick based or native vtime accounting operate before the tick is
stopped and resumes once the tick is restarted.
* When the idle loop starts, switch to dynticks-idle accounting as is
done currently, except that the statistics accumulate directly to the
relevant kernel stat fields.
* Private dyntick cputime accounting fields are removed.
* Works on both online and offline case.
* Move most of the relevant code to the common sched/cputime subsystem
* Handle CONFIG_IRQ_TIME_ACCOUNTING=n correctly such that the
dynticks-idle accounting still elapses while on IRQs.
* Correctly substract idle steal cputime from idle time
Changes since v1:
- Fix deadlock involving double seq count lock on idle
- Fix build breakage on powerpc
- Fix build breakage on s390 (Heiko)
- Fix broken sysfs s390 idle time file (Heiko)
- Convert most ktime usage here into u64 (Peterz)
- Add missing (or too implicit) <linux/sched/clock.h> (Peterz)
- Fix whole idle time acccounting breakage due to missing TS_FLAG_ set
on idle entry (Shrikanth Hegde)
git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks.git
timers/core-v2
HEAD: 21458b98c80a0567d48131240317b7b73ba34c3c
Thanks,
Frederic
idle and runtime utilization with mpstat while running stress-ng looks
correct now.
However, when running hackbench I am noticing the below data. hackbench shows
severe regressions.
base: tip/master at 9c61ebbdb587a3950072700ab74a9310afe3ad73.
(nit: patch 7 is already part of tip. so skipped applying it)
+-----------------------------------------------+-------+---------+-----------+
| Test | base | +series | % Diff |
+-----------------------------------------------+-------+---------+-----------+
| HackBench Process 10 groups | 2.23 | 3.05 | -36.77% |
| HackBench Process 20 groups | 4.17 | 5.82 | -39.57% |
| HackBench Process 30 groups | 6.04 | 8.49 | -40.56% |
| HackBench Process 40 groups | 7.90 | 11.10 | -40.51% |
| HackBench thread 10 | 2.44 | 3.36 | -37.70% |
| HackBench thread 20 | 4.57 | 6.35 | -38.95% |
| HackBench Process(Pipe) 10 | 1.76 | 2.29 | -30.11% |
| HackBench Process(Pipe) 20 | 3.49 | 4.76 | -36.39% |
| HackBench Process(Pipe) 30 | 5.21 | 7.13 | -36.85% |
| HackBench Process(Pipe) 40 | 6.89 | 9.31 | -35.12% |
| HackBench thread(Pipe) 10 | 1.91 | 2.50 | -30.89% |
| HackBench thread(Pipe) 20 | 3.74 | 5.16 | -37.97% |
+-----------------------------------------------+-------+---------+-----------+
I have these in .config and I don't have nohz_full or isolated cpus.
CONFIG_TICK_ONESHOT=y
CONFIG_NO_HZ_COMMON=y
# CONFIG_HZ_PERIODIC is not set
# CONFIG_NO_HZ_IDLE is not set
CONFIG_NO_HZ_FULL=y
# CPU/Task time and stats accounting
#
CONFIG_VIRT_CPU_ACCOUNTING=y
CONFIG_VIRT_CPU_ACCOUNTING_GEN=y
CONFIG_IRQ_TIME_ACCOUNTING=y
CONFIG_HAVE_SCHED_AVG_IRQ=y
I did a git bisect and below is what it says.
git bisect start
# status: waiting for both good and bad commits
# bad: [6821315886a3b5267ea31d29dba26fd34647fbbc] sched/cputime: Handle
dyntick-idle steal time correctly
git bisect bad 6821315886a3b5267ea31d29dba26fd34647fbbc
# status: waiting for good commit(s), bad commit known
# good: [9c61ebbdb587a3950072700ab74a9310afe3ad73] Merge branch into
tip/master: 'x86/sev'
git bisect good 9c61ebbdb587a3950072700ab74a9310afe3ad73
# good: [dc8bb3c84d162f7d9aa6becf9f8392474f92655a] tick/sched: Remove nohz
disabled special case in cputime fetch
git bisect good dc8bb3c84d162f7d9aa6becf9f8392474f92655a
# good: [5070a778a581cd668f5d717f85fb22b078d8c20c] tick/sched: Account tickless
idle cputime only when tick is stopped
git bisect good 5070a778a581cd668f5d717f85fb22b078d8c20c
# bad: [1e0ccc25a9a74b188b239c4de716fde279adbf8e] sched/cputime: Provide
get_cpu_[idle|iowait]_time_us() off-case
git bisect bad 1e0ccc25a9a74b188b239c4de716fde279adbf8e
# bad: [ee7c735b76071000d401869fc2883c451ee3fa61] tick/sched: Consolidate idle
time fetching APIs
git bisect bad ee7c735b76071000d401869fc2883c451ee3fa61
# first bad commit: [ee7c735b76071000d401869fc2883c451ee3fa61] tick/sched:
Consolidate idle time fetching APIs
I did a perf diff between the two (collected perf record -a for hackbench 60
process 10000 loops)
perf diff base series:
# Baseline Delta Abs Shared Object Symbol
# ........ ......... ...........................
................................................
#
+5.43% [kernel.kallsyms] [k] __update_freelist_slow
0.00% +4.55% [kernel.kallsyms] [k] _raw_spin_lock
+3.35% [kernel.kallsyms] [k] __memcg_slab_free_hook
0.55% +2.58% [kernel.kallsyms] [k] sock_wfree
+2.51% [kernel.kallsyms] [k] __account_obj_stock
2.29% -2.29% [kernel.kallsyms] [k] _raw_write_lock_irq
+2.25% [kernel.kallsyms] [k] _copy_from_iter
+1.96% [kernel.kallsyms] [k] fdget_pos
+1.87% [kernel.kallsyms] [k] _copy_to_iter
+1.69% [kernel.kallsyms] [k] sock_def_readable
2.32% -1.68% [kernel.kallsyms] [k] mod_memcg_lruvec_state
0.82% +1.67% [kernel.kallsyms] [k] skb_set_owner_w
0.08% +1.65% [kernel.kallsyms] [k] vfs_read
0.42% +1.57% [kernel.kallsyms] [k]
kmem_cache_alloc_node_noprof
1.53% -1.53% [kernel.kallsyms] [k]
kmem_cache_alloc_lru_noprof
1.56% -1.41% [kernel.kallsyms] [k] simple_copy_to_iter
0.27% +1.32% [kernel.kallsyms] [k] kfree
0.01% +1.25% [kernel.kallsyms] [k] __slab_free
0.19% +1.24% [kernel.kallsyms] [k] kmem_cache_free
1.23% -1.23% [kernel.kallsyms] [k] __pcs_replace_full_main
0.35% +1.21% [kernel.kallsyms] [k] __skb_datagram_iter
0.21% +1.13% [kernel.kallsyms] [k] sock_alloc_send_pskb
+1.09% [kernel.kallsyms] [k] mutex_lock
+0.98% [kernel.kallsyms].head.text [k] 0x0000000000013004
I haven't gone through the series yet. trying to go through meanwhile.
maybe different allocation scheme or more allocation/free everytime instead of
pre-allocated percpu variables?
First thought of reporting it. Let me know if you need any additional data.
