| Message ID | 20220208100750.1189808-1-nsaenzju@redhat.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | mm/page_alloc: Remote per-cpu lists drain support | expand |
Hi Nicolas，
When I applied the patchset on the following commit and tested on QEMU, I came
accross the following CallTrace.
commit dd81e1c7d5fb126e5fbc5c9e334d7b3ec29a16a0
I wrote a userspace application to consume the memory. When the memory is used
out, the OOM killer is triggered and the following Calltrace is printed. I am
not sure if it is related to this patchset. But when I reverted this patchset,
the 'NULL pointer' Calltrace didn't show.
root@syzkaller:~# [ 148.041840][ T476] Unable to handle kernel NULL pointer
dereference at virtual address 0000000000000058
[ 148.042838][ T476] Mem abort info:
[ 148.043188][ T476] ESR = 0x96000004
[ 148.043540][ T476] EC = 0x25: DABT (current EL), IL = 32 bits
[ 148.044102][ T476] SET = 0, FnV = 0
[ 148.044451][ T476] EA = 0, S1PTW = 0
[ 148.044812][ T476] FSC = 0x04: level 0 translation fault
[ 148.045333][ T476] Data abort info:
[ 148.045666][ T476] ISV = 0, ISS = 0x00000004
[ 148.046103][ T476] CM = 0, WnR = 0
[ 148.046444][ T476] user pgtable: 4k pages, 48-bit VAs, pgdp=00000000abcec000
[ 148.047187][ T476] [0000000000000058] pgd=0000000000000000, p4d=0000000000000000
[ 148.047915][ T476] Internal error: Oops: 96000004 [#1] SMP
[ 148.048450][ T476] Dumping ftrace buffer:
[ 148.048843][ T476] (ftrace buffer empty)
[ 148.049473][ T476] Modules linked in:
[ 148.049834][ T476] CPU: 1 PID: 476 Comm: test Not tainted 5.17.0-rc1+ #19
[ 148.050485][ T476] Hardware name: QEMU KVM Virtual Machine, BIOS 0.0.0
02/06/2015
[ 148.051242][ T476] pstate: 80400005 (Nzcv daif +PAN -UAO -TCO -DIT -SSBS
BTYPE=--)
[ 148.052019][ T476] pc : __drain_all_pages+0x1c4/0x318
[ 148.052548][ T476] lr : __drain_all_pages+0x1a0/0x318
[ 148.053070][ T476] sp : ffff80000ce9ba20
[ 148.053477][ T476] x29: ffff80000ce9ba20 x28: 0000000000000000 x27:
0000000000000000
[ 148.054268][ T476] x26: 0000000000000000 x25: ffff800009799000 x24:
0000000000000000
[ 148.055067][ T476] x23: ffff80000979f460 x22: ffff800009bac000 x21:
ffff800009ed7110
[ 148.055997][ T476] x20: ffff80000979a000 x19: 0000000000000000 x18:
0000000000000000
[ 148.056784][ T476] x17: 0000000000000000 x16: 0000000000000000 x15:
0000000000000000
[ 148.058493][ T476] x14: 0000000000000000 x13: 0000000000000000 x12:
ffff80000ce9b558
[ 148.059649][ T476] x11: 0000000000000000 x10: ffff000bfffe6240 x9 :
0000000000000002
[ 148.060775][ T476] x8 : 0000000000000000 x7 : ffffffffffffffff x6 :
0000000000000000
[ 148.061609][ T476] x5 : 0000000000000000 x4 : ffff80000979bb70 x3 :
ffff800bf6a18000
[ 148.062435][ T476] x2 : 0000000000000080 x1 : 00000000000c0000 x0 :
ffff000bfffe6240
[ 148.063267][ T476] Call trace:
[ 148.063602][ T476] __drain_all_pages+0x1c4/0x318
[ 148.064110][ T476] __alloc_pages_slowpath.constprop.126+0x328/0xac8
[ 148.064790][ T476] __alloc_pages+0x270/0x330
[ 148.065260][ T476] alloc_pages+0xb0/0x160
[ 148.065703][ T476] __pte_alloc+0x44/0x180
[ 148.066147][ T476] __handle_mm_fault+0x59c/0xb30
[ 148.066656][ T476] handle_mm_fault+0xc4/0x210
[ 148.067139][ T476] do_page_fault+0x1cc/0x480
[ 148.067607][ T476] do_translation_fault+0x80/0x98
[ 148.068265][ T476] do_mem_abort+0x50/0xa0
[ 148.068724][ T476] el0_da+0x40/0x118
[ 148.069132][ T476] el0t_64_sync_handler+0x60/0xb0
[ 148.069659][ T476] el0t_64_sync+0x16c/0x170
[ 148.070127][ T476] Code: b4fff7e0 f9404401 b4ffffa1 f87c7ae3 (f9402f01)
[ 148.070874][ T476] ---[ end trace 0000000000000000 ]---
[ 149.059678][ T229] systemd-journal invoked oom-killer:
gfp_mask=0x1140cca(GFP_HIGHUSER_MOVABLE|__GFP_COMP), order=0, oom_score_adj=0
[ 149.064280][ T229] CPU: 0 PID: 229 Comm: systemd-journal Tainted: G D
5.17.0-rc1+ #19
[ 149.069110][ T229] Hardware name: QEMU KVM Virtual Machine, BIOS 0.0.0
02/06/2015
[ 149.073070][ T229] Call trace:
[ 149.074392][ T229] dump_backtrace+0x114/0x120
[ 149.077045][ T229] show_stack+0x2c/0x78
[ 149.079680][ T229] dump_stack_lvl+0x64/0x7c
[ 149.081741][ T229] dump_stack+0x14/0x2c
[ 149.083760][ T229] dump_header+0x48/0x200
[ 149.085772][ T229] oom_kill_process+0x230/0x2e0
[ 149.087877][ T229] out_of_memory+0xf0/0x5a8
[ 149.090776][ T229] __alloc_pages_slowpath.constprop.126+0x830/0xac8
[ 149.103397][ T229] __alloc_pages+0x270/0x330
[ 149.112662][ T229] alloc_pages+0xb0/0x160
[ 149.113088][ T229] folio_alloc+0x28/0x50
[ 149.113504][ T229] filemap_alloc_folio+0x98/0xc8
[ 149.113990][ T229] __filemap_get_folio+0x194/0x2f0
[ 149.114492][ T229] filemap_fault+0x434/0x708
[ 149.115799][ T229] __do_fault+0x3c/0x118
[ 149.117842][ T229] do_fault+0x160/0x428
[ 149.119893][ T229] __handle_mm_fault+0x498/0xb30
[ 149.123249][ T229] handle_mm_fault+0xc4/0x210
[ 149.125558][ T229] do_page_fault+0x1cc/0x480
[ 149.128347][ T229] do_translation_fault+0x80/0x98
[ 149.130444][ T229] do_mem_abort+0x50/0xa0
[ 149.132513][ T229] el0_da+0x40/0x118
[ 149.135152][ T229] el0t_64_sync_handler+0x60/0xb0
[ 149.137802][ T229] el0t_64_sync+0x16c/0x170
[ 149.140088][ T229] Mem-Info:
[ 149.141197][ T229] active_anon:108 inactive_anon:12226341 isolated_anon:0
[ 149.141197][ T229] active_file:68 inactive_file:6 isolated_file:0
[ 149.141197][ T229] unevictable:0 dirty:0 writeback:0
[ 149.141197][ T229] slab_reclaimable:5964 slab_unreclaimable:14242
[ 149.141197][ T229] mapped:41 shmem:995 pagetables:24564 bounce:0
[ 149.141197][ T229] kernel_misc_reclaimable:0
[ 149.141197][ T229] free:52077 free_pcp:1845 free_cma:13312
Thanks,
Xiongfeng
On 2022/2/8 18:07, Nicolas Saenz Julienne wrote:
> This series replaces mm/page_alloc's per-cpu page lists drain mechanism with
> one that allows accessing the lists remotely. Currently, only a local CPU is
> permitted to change its per-cpu lists, and it's expected to do so, on-demand,
> whenever a process demands it by means of queueing a drain task on the local
> CPU. This causes problems for NOHZ_FULL CPUs and real-time systems that can't
> take any sort of interruption and to some lesser extent inconveniences idle and
> virtualised systems.
>
> The new algorithm will atomically switch the pointer to the per-cpu page lists
> and use RCU to make sure it's not being concurrently used before draining the
> lists. And its main benefit of is that it fixes the issue for good, avoiding
> the need for configuration based heuristics or having to modify applications
> (i.e. using the isolation prctrl being worked by Marcello Tosatti ATM).
>
> All this with minimal performance implications: a page allocation
> microbenchmark was run on multiple systems and architectures generally showing
> no performance differences, only the more extreme cases showed a 1-3%
> degradation. See data below. Needless to say that I'd appreciate if someone
> could validate my values independently.
>
> The approach has been stress-tested: I forced 100 drains/s while running
> mmtests' pft in a loop for a full day on multiple machines and archs (arm64,
> x86_64, ppc64le).
>
> Note that this is not the first attempt at fixing this per-cpu page lists:
> - The first attempt[1] tried to conditionally change the pagesets locking
> scheme based the NOHZ_FULL config. It was deemed hard to maintain as the
> NOHZ_FULL code path would be rarely tested. Also, this only solves the issue
> for NOHZ_FULL setups, which isn't ideal.
> - The second[2] unanimously switched the local_locks to per-cpu spinlocks. The
> performance degradation was too big.
>
> Previous RFC: https://lkml.org/lkml/2021/10/8/793
>
> Thanks!
>
> [1] https://lkml.org/lkml/2021/9/21/599
> [2] https://lkml.org/lkml/2021/11/3/644
>
> ---
>
> Changes since RFC:
> - Get rid of aesthetic changes that affected performance
> - Add more documentation
> - Add better commit messages
> - Pass sparse tests
> - Verify this_cpu_*() usage
> - Performance measurements
>
> Nicolas Saenz Julienne (2):
> mm/page_alloc: Access lists in 'struct per_cpu_pages' indirectly
> mm/page_alloc: Add remote draining support to per-cpu lists
>
> include/linux/mmzone.h | 28 +++++-
> mm/page_alloc.c | 212 ++++++++++++++++++++++++++---------------
> mm/vmstat.c | 6 +-
> 3 files changed, 162 insertions(+), 84 deletions(-)
>
>
> -------------------------Performance results-----------------------------
>
> I'm focusing on mmtests' Page Fault Test (pft), as it's page allocator
> intensive.
>
> - AMD Daytona Reference System, 2 sockets, AMD EPYC 7742, Zen 2, 64-Core,
> 4 NUMA nodes, x86_64
>
> pft timings:
> vanilla rcu
> Amean system-1 58.52 ( 0.00%) 58.92 * -0.68%*
> Amean system-4 61.00 ( 0.00%) 61.41 * -0.67%*
> Amean system-7 61.55 ( 0.00%) 61.74 * -0.30%*
> Amean system-12 64.91 ( 0.00%) 64.94 * -0.05%*
> Amean system-21 98.80 ( 0.00%) 99.92 * -1.13%*
> Amean system-30 147.68 ( 0.00%) 145.83 * 1.25%*
> Amean system-48 237.04 ( 0.00%) 241.29 * -1.79%*
> Amean system-79 286.61 ( 0.00%) 283.72 * 1.01%*
> Amean system-110 303.40 ( 0.00%) 299.91 * 1.15%*
> Amean system-128 345.07 ( 0.00%) 342.10 * 0.86%*
> Amean elapsed-1 61.21 ( 0.00%) 61.65 * -0.71%*
> Amean elapsed-4 15.94 ( 0.00%) 16.05 * -0.69%*
> Amean elapsed-7 9.24 ( 0.00%) 9.28 * -0.47%*
> Amean elapsed-12 5.70 ( 0.00%) 5.70 * -0.07%*
> Amean elapsed-21 5.11 ( 0.00%) 5.06 * 1.13%*
> Amean elapsed-30 5.28 ( 0.00%) 5.14 * 2.73%*
> Amean elapsed-48 5.28 ( 0.00%) 5.24 * 0.74%*
> Amean elapsed-79 4.41 ( 0.00%) 4.31 * 2.17%*
> Amean elapsed-110 3.45 ( 0.00%) 3.44 * 0.40%*
> Amean elapsed-128 2.75 ( 0.00%) 2.75 * -0.28%*
>
> - AMD Speedway Reference System, 2 sockets, AMD EPYC 7601, Zen 1, 64-core, 8
> NUMA nodes, x86_64. Lots of variance between tests on this platform. It'll
> easily swing -+5% on each result.
>
> pft timings:
> vanilla rcu
> Amean system-1 69.20 ( 0.00%) 66.21 * 4.32%*
> Amean system-4 70.79 ( 0.00%) 69.01 * 2.52%*
> Amean system-7 71.34 ( 0.00%) 69.16 * 3.05%*
> Amean system-12 74.00 ( 0.00%) 72.74 * 1.70%*
> Amean system-21 86.01 ( 0.00%) 85.70 * 0.36%*
> Amean system-30 89.21 ( 0.00%) 89.93 * -0.80%*
> Amean system-48 92.39 ( 0.00%) 92.43 * -0.04%*
> Amean system-79 120.19 ( 0.00%) 121.30 * -0.92%*
> Amean system-110 172.79 ( 0.00%) 179.37 * -3.81%*
> Amean system-128 201.70 ( 0.00%) 212.57 * -5.39%*
> Amean elapsed-1 72.23 ( 0.00%) 69.29 * 4.08%*
> Amean elapsed-4 18.69 ( 0.00%) 18.28 * 2.20%*
> Amean elapsed-7 10.80 ( 0.00%) 10.54 * 2.41%*
> Amean elapsed-12 6.62 ( 0.00%) 6.53 * 1.30%*
> Amean elapsed-21 4.68 ( 0.00%) 4.69 * -0.14%*
> Amean elapsed-30 3.44 ( 0.00%) 3.50 * -1.66%*
> Amean elapsed-48 2.40 ( 0.00%) 2.42 * -1.00%*
> Amean elapsed-79 2.05 ( 0.00%) 2.09 * -1.90%*
> Amean elapsed-110 1.83 ( 0.00%) 1.91 * -4.60%*
> Amean elapsed-128 1.75 ( 0.00%) 1.85 * -5.99%*
>
> - IBM 9006-22C system, 2 sockets, POWER9, 64-Core, 1 NUMA node per cpu,
> pppc64le.
>
> pft timings:
> vanilla rcu
> Amean system-1 1.82 ( 0.00%) 1.85 * -1.43%*
> Amean system-4 2.18 ( 0.00%) 2.22 * -2.02%*
> Amean system-7 3.27 ( 0.00%) 3.28 * -0.15%*
> Amean system-12 5.22 ( 0.00%) 5.20 * 0.26%*
> Amean system-21 10.10 ( 0.00%) 10.20 * -1.00%*
> Amean system-30 15.00 ( 0.00%) 14.52 * 3.20%*
> Amean system-48 26.41 ( 0.00%) 25.96 * 1.71%*
> Amean system-79 29.35 ( 0.00%) 29.70 * -1.21%*
> Amean system-110 24.01 ( 0.00%) 23.40 * 2.54%*
> Amean system-128 24.57 ( 0.00%) 25.32 * -3.06%*
> Amean elapsed-1 1.85 ( 0.00%) 1.87 * -1.28%*
> Amean elapsed-4 0.56 ( 0.00%) 0.57 * -1.72%*
> Amean elapsed-7 0.51 ( 0.00%) 0.50 * 0.07%*
> Amean elapsed-12 0.51 ( 0.00%) 0.51 * 0.06%*
> Amean elapsed-21 0.54 ( 0.00%) 0.54 * 0.06%*
> Amean elapsed-30 0.54 ( 0.00%) 0.53 * 2.22%*
> Amean elapsed-48 0.58 ( 0.00%) 0.57 * 1.73%*
> Amean elapsed-79 0.49 ( 0.00%) 0.48 * 0.89%*
> Amean elapsed-110 0.37 ( 0.00%) 0.37 * -1.08%*
> Amean elapsed-128 0.33 ( 0.00%) 0.33 * 0.00%*
>
> - Ampere MtSnow, 1 socket, Neoverse-N1, 80-Cores, 1 NUMA node, arm64.
>
> pft timings:
> vanilla rcu
> Amean system-1 11.92 ( 0.00%) 11.99 * -0.61%*
> Amean system-4 13.13 ( 0.00%) 13.09 * 0.31%*
> Amean system-7 13.91 ( 0.00%) 13.94 * -0.20%*
> Amean system-12 15.77 ( 0.00%) 15.69 * 0.48%*
> Amean system-21 21.32 ( 0.00%) 21.42 * -0.46%*
> Amean system-30 28.58 ( 0.00%) 29.12 * -1.90%*
> Amean system-48 47.41 ( 0.00%) 46.91 * 1.04%*
> Amean system-79 76.76 ( 0.00%) 77.16 * -0.52%*
> Amean system-80 77.98 ( 0.00%) 78.23 * -0.32%*
> Amean elapsed-1 12.46 ( 0.00%) 12.53 * -0.58%*
> Amean elapsed-4 3.47 ( 0.00%) 3.46 * 0.34%*
> Amean elapsed-7 2.18 ( 0.00%) 2.21 * -1.58%*
> Amean elapsed-12 1.41 ( 0.00%) 1.42 * -0.80%*
> Amean elapsed-21 1.09 ( 0.00%) 1.12 * -2.60%*
> Amean elapsed-30 0.98 ( 0.00%) 1.01 * -3.08%*
> Amean elapsed-48 1.08 ( 0.00%) 1.10 * -1.78%*
> Amean elapsed-79 1.32 ( 0.00%) 1.28 * 2.71%*
> Amean elapsed-80 1.32 ( 0.00%) 1.28 * 3.23%*
>
> - Dell R430, 2 sockets, Intel Xeon E5-2640 v3, Sandy Bridge, 16-Cores, 2 NUMA
> nodes, x86_64.
>
> pft timings:
> vanilla rcu
> Amean system-1 11.10 ( 0.00%) 11.07 * 0.24%*
> Amean system-3 11.14 ( 0.00%) 11.10 * 0.34%*
> Amean system-5 11.18 ( 0.00%) 11.13 * 0.47%*
> Amean system-7 11.21 ( 0.00%) 11.17 * 0.38%*
> Amean system-12 11.28 ( 0.00%) 11.28 ( -0.03%)
> Amean system-18 13.24 ( 0.00%) 13.25 * -0.11%*
> Amean system-24 17.12 ( 0.00%) 17.14 ( -0.13%)
> Amean system-30 21.10 ( 0.00%) 21.23 * -0.60%*
> Amean system-32 22.31 ( 0.00%) 22.47 * -0.71%*
> Amean elapsed-1 11.76 ( 0.00%) 11.73 * 0.29%*
> Amean elapsed-3 3.93 ( 0.00%) 3.93 * 0.17%*
> Amean elapsed-5 2.39 ( 0.00%) 2.37 * 0.74%*
> Amean elapsed-7 1.72 ( 0.00%) 1.71 * 0.81%*
> Amean elapsed-12 1.02 ( 0.00%) 1.03 ( -0.42%)
> Amean elapsed-18 1.13 ( 0.00%) 1.14 ( -0.18%)
> Amean elapsed-24 0.87 ( 0.00%) 0.88 * -0.65%*
> Amean elapsed-30 0.77 ( 0.00%) 0.78 * -0.86%*
> Amean elapsed-32 0.74 ( 0.00%) 0.74 ( 0.00%)
>
> - HPE Apollo 70, 2 sockets, Cavium ThunderX2, 128-Cores, 2 NUMA nodes, arm64.
> NOTE: The test here only goes up to 128 for some reason, although there are
> 256 CPUs. Maybe a mmtests issue? I didn't investigate.
>
> pft timings:
> vanilla rcu
> Amean system-1 4.42 ( 0.00%) 4.36 * 1.29%*
> Amean system-4 4.56 ( 0.00%) 4.51 * 1.05%*
> Amean system-7 4.63 ( 0.00%) 4.65 * -0.42%*
> Amean system-12 5.96 ( 0.00%) 6.02 * -1.00%*
> Amean system-21 10.97 ( 0.00%) 11.01 * -0.32%*
> Amean system-30 16.01 ( 0.00%) 16.04 * -0.19%*
> Amean system-48 26.81 ( 0.00%) 26.78 * 0.09%*
> Amean system-79 30.80 ( 0.00%) 30.85 * -0.16%*
> Amean system-110 31.87 ( 0.00%) 31.93 * -0.19%*
> Amean system-128 36.27 ( 0.00%) 36.31 * -0.10%*
> Amean elapsed-1 4.88 ( 0.00%) 4.85 * 0.60%*
> Amean elapsed-4 1.27 ( 0.00%) 1.26 * 1.00%*
> Amean elapsed-7 0.73 ( 0.00%) 0.74 * -0.46%*
> Amean elapsed-12 0.55 ( 0.00%) 0.55 * 1.09%*
> Amean elapsed-21 0.59 ( 0.00%) 0.60 * -0.96%*
> Amean elapsed-30 0.60 ( 0.00%) 0.60 * 0.28%*
> Amean elapsed-48 0.60 ( 0.00%) 0.60 * 0.44%*
> Amean elapsed-79 0.49 ( 0.00%) 0.49 * -0.07%*
> Amean elapsed-110 0.36 ( 0.00%) 0.36 * 0.28%*
> Amean elapsed-128 0.31 ( 0.00%) 0.31 * -0.43%*
>
> - Raspberry Pi 4, 1 socket, bcm2711, Cortex-A72, 4-Cores, 1 NUMA node, arm64.
>
> pft timings:
> vanilla rcu
> Amean system-1 0.67 ( 0.00%) 0.67 * -1.25%*
> Amean system-3 1.30 ( 0.00%) 1.29 * 0.62%*
> Amean system-4 1.61 ( 0.00%) 1.59 * 0.95%*
> Amean elapsed-1 0.71 ( 0.00%) 0.72 * -1.17%*
> Amean elapsed-3 0.45 ( 0.00%) 0.45 * 0.88%*
> Amean elapsed-4 0.42 ( 0.00%) 0.42 * 1.19%*
>
>
Hi Xiongfeng, thanks for taking the time to look at this. On Wed, 2022-02-09 at 16:55 +0800, Xiongfeng Wang wrote: > Hi Nicolas， > > When I applied the patchset on the following commit and tested on QEMU, I came > accross the following CallTrace. > commit dd81e1c7d5fb126e5fbc5c9e334d7b3ec29a16a0 > > I wrote a userspace application to consume the memory. When the memory is used > out, the OOM killer is triggered and the following Calltrace is printed. I am > not sure if it is related to this patchset. But when I reverted this patchset, > the 'NULL pointer' Calltrace didn't show. It's a silly mistake on my part, while cleaning up the code I messed up one of the 'struct per_cpu_pages' accessors. This should fix it: ------------------------->8------------------------- diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 0caa7155ca34..e65b991c3dc8 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -3279,7 +3279,7 @@ static void __drain_all_pages(struct zone *zone, bool force_all_cpus) has_pcps = true; } else { for_each_populated_zone(z) { - pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); + pcp = per_cpu_ptr(z->per_cpu_pageset, cpu); lp = rcu_dereference_protected(pcp->lp, mutex_is_locked(&pcpu_drain_mutex)); if (lp->count) { ------------------------->8------------------------- Thanks!
Hi, On 2022/2/9 17:45, Nicolas Saenz Julienne wrote: > Hi Xiongfeng, thanks for taking the time to look at this. > > On Wed, 2022-02-09 at 16:55 +0800, Xiongfeng Wang wrote: >> Hi Nicolas， >> >> When I applied the patchset on the following commit and tested on QEMU, I came >> accross the following CallTrace. >> commit dd81e1c7d5fb126e5fbc5c9e334d7b3ec29a16a0 >> >> I wrote a userspace application to consume the memory. When the memory is used >> out, the OOM killer is triggered and the following Calltrace is printed. I am >> not sure if it is related to this patchset. But when I reverted this patchset, >> the 'NULL pointer' Calltrace didn't show. > > It's a silly mistake on my part, while cleaning up the code I messed up one of > the 'struct per_cpu_pages' accessors. This should fix it: > > ------------------------->8------------------------- > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > index 0caa7155ca34..e65b991c3dc8 100644 > --- a/mm/page_alloc.c > +++ b/mm/page_alloc.c > @@ -3279,7 +3279,7 @@ static void __drain_all_pages(struct zone *zone, bool force_all_cpus) > has_pcps = true; > } else { > for_each_populated_zone(z) { > - pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); > + pcp = per_cpu_ptr(z->per_cpu_pageset, cpu); > lp = rcu_dereference_protected(pcp->lp, > mutex_is_locked(&pcpu_drain_mutex)); > if (lp->count) { I have tested it. It works well. No more 'NULL pointer' Calltrace. Thanks, Xiongfeng > ------------------------->8------------------------- > > Thanks! >
On Wed, 2022-02-09 at 19:26 +0800, Xiongfeng Wang wrote: > Hi, > > On 2022/2/9 17:45, Nicolas Saenz Julienne wrote: > > Hi Xiongfeng, thanks for taking the time to look at this. > > > > On Wed, 2022-02-09 at 16:55 +0800, Xiongfeng Wang wrote: > > > Hi Nicolas， > > > > > > When I applied the patchset on the following commit and tested on QEMU, I came > > > accross the following CallTrace. > > > commit dd81e1c7d5fb126e5fbc5c9e334d7b3ec29a16a0 > > > > > > I wrote a userspace application to consume the memory. When the memory is used > > > out, the OOM killer is triggered and the following Calltrace is printed. I am > > > not sure if it is related to this patchset. But when I reverted this patchset, > > > the 'NULL pointer' Calltrace didn't show. > > > > It's a silly mistake on my part, while cleaning up the code I messed up one of > > the 'struct per_cpu_pages' accessors. This should fix it: > > > > ------------------------->8------------------------- > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > > index 0caa7155ca34..e65b991c3dc8 100644 > > --- a/mm/page_alloc.c > > +++ b/mm/page_alloc.c > > @@ -3279,7 +3279,7 @@ static void __drain_all_pages(struct zone *zone, bool force_all_cpus) > > has_pcps = true; > > } else { > > for_each_populated_zone(z) { > > - pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); > > + pcp = per_cpu_ptr(z->per_cpu_pageset, cpu); > > lp = rcu_dereference_protected(pcp->lp, > > mutex_is_locked(&pcpu_drain_mutex)); > > if (lp->count) { > > I have tested it. It works well. No more 'NULL pointer' Calltrace. Thanks!
Hi Nicolas, I found there are several 'lru_add_drain_per_cpu()' worker thread on the NOHZ_FULL CPUs. I am just wondering do you have plan to add support for remote per-cpu LRU pagevec drain ? Thanks, Xiongfeng On 2022/2/8 18:07, Nicolas Saenz Julienne wrote: > This series replaces mm/page_alloc's per-cpu page lists drain mechanism with > one that allows accessing the lists remotely. Currently, only a local CPU is > permitted to change its per-cpu lists, and it's expected to do so, on-demand, > whenever a process demands it by means of queueing a drain task on the local > CPU. This causes problems for NOHZ_FULL CPUs and real-time systems that can't > take any sort of interruption and to some lesser extent inconveniences idle and > virtualised systems. > > The new algorithm will atomically switch the pointer to the per-cpu page lists > and use RCU to make sure it's not being concurrently used before draining the > lists. And its main benefit of is that it fixes the issue for good, avoiding > the need for configuration based heuristics or having to modify applications > (i.e. using the isolation prctrl being worked by Marcello Tosatti ATM). > > All this with minimal performance implications: a page allocation > microbenchmark was run on multiple systems and architectures generally showing > no performance differences, only the more extreme cases showed a 1-3% > degradation. See data below. Needless to say that I'd appreciate if someone > could validate my values independently. > > The approach has been stress-tested: I forced 100 drains/s while running > mmtests' pft in a loop for a full day on multiple machines and archs (arm64, > x86_64, ppc64le). > > Note that this is not the first attempt at fixing this per-cpu page lists: > - The first attempt[1] tried to conditionally change the pagesets locking > scheme based the NOHZ_FULL config. It was deemed hard to maintain as the > NOHZ_FULL code path would be rarely tested. Also, this only solves the issue > for NOHZ_FULL setups, which isn't ideal. > - The second[2] unanimously switched the local_locks to per-cpu spinlocks. The > performance degradation was too big. > > Previous RFC: https://lkml.org/lkml/2021/10/8/793 > > Thanks! > > [1] https://lkml.org/lkml/2021/9/21/599 > [2] https://lkml.org/lkml/2021/11/3/644 > > --- > > Changes since RFC: > - Get rid of aesthetic changes that affected performance > - Add more documentation > - Add better commit messages > - Pass sparse tests > - Verify this_cpu_*() usage > - Performance measurements > > Nicolas Saenz Julienne (2): > mm/page_alloc: Access lists in 'struct per_cpu_pages' indirectly > mm/page_alloc: Add remote draining support to per-cpu lists > > include/linux/mmzone.h | 28 +++++- > mm/page_alloc.c | 212 ++++++++++++++++++++++++++--------------- > mm/vmstat.c | 6 +- > 3 files changed, 162 insertions(+), 84 deletions(-) > > > -------------------------Performance results----------------------------- > > I'm focusing on mmtests' Page Fault Test (pft), as it's page allocator > intensive. > > - AMD Daytona Reference System, 2 sockets, AMD EPYC 7742, Zen 2, 64-Core, > 4 NUMA nodes, x86_64 > > pft timings: > vanilla rcu > Amean system-1 58.52 ( 0.00%) 58.92 * -0.68%* > Amean system-4 61.00 ( 0.00%) 61.41 * -0.67%* > Amean system-7 61.55 ( 0.00%) 61.74 * -0.30%* > Amean system-12 64.91 ( 0.00%) 64.94 * -0.05%* > Amean system-21 98.80 ( 0.00%) 99.92 * -1.13%* > Amean system-30 147.68 ( 0.00%) 145.83 * 1.25%* > Amean system-48 237.04 ( 0.00%) 241.29 * -1.79%* > Amean system-79 286.61 ( 0.00%) 283.72 * 1.01%* > Amean system-110 303.40 ( 0.00%) 299.91 * 1.15%* > Amean system-128 345.07 ( 0.00%) 342.10 * 0.86%* > Amean elapsed-1 61.21 ( 0.00%) 61.65 * -0.71%* > Amean elapsed-4 15.94 ( 0.00%) 16.05 * -0.69%* > Amean elapsed-7 9.24 ( 0.00%) 9.28 * -0.47%* > Amean elapsed-12 5.70 ( 0.00%) 5.70 * -0.07%* > Amean elapsed-21 5.11 ( 0.00%) 5.06 * 1.13%* > Amean elapsed-30 5.28 ( 0.00%) 5.14 * 2.73%* > Amean elapsed-48 5.28 ( 0.00%) 5.24 * 0.74%* > Amean elapsed-79 4.41 ( 0.00%) 4.31 * 2.17%* > Amean elapsed-110 3.45 ( 0.00%) 3.44 * 0.40%* > Amean elapsed-128 2.75 ( 0.00%) 2.75 * -0.28%* > > - AMD Speedway Reference System, 2 sockets, AMD EPYC 7601, Zen 1, 64-core, 8 > NUMA nodes, x86_64. Lots of variance between tests on this platform. It'll > easily swing -+5% on each result. > > pft timings: > vanilla rcu > Amean system-1 69.20 ( 0.00%) 66.21 * 4.32%* > Amean system-4 70.79 ( 0.00%) 69.01 * 2.52%* > Amean system-7 71.34 ( 0.00%) 69.16 * 3.05%* > Amean system-12 74.00 ( 0.00%) 72.74 * 1.70%* > Amean system-21 86.01 ( 0.00%) 85.70 * 0.36%* > Amean system-30 89.21 ( 0.00%) 89.93 * -0.80%* > Amean system-48 92.39 ( 0.00%) 92.43 * -0.04%* > Amean system-79 120.19 ( 0.00%) 121.30 * -0.92%* > Amean system-110 172.79 ( 0.00%) 179.37 * -3.81%* > Amean system-128 201.70 ( 0.00%) 212.57 * -5.39%* > Amean elapsed-1 72.23 ( 0.00%) 69.29 * 4.08%* > Amean elapsed-4 18.69 ( 0.00%) 18.28 * 2.20%* > Amean elapsed-7 10.80 ( 0.00%) 10.54 * 2.41%* > Amean elapsed-12 6.62 ( 0.00%) 6.53 * 1.30%* > Amean elapsed-21 4.68 ( 0.00%) 4.69 * -0.14%* > Amean elapsed-30 3.44 ( 0.00%) 3.50 * -1.66%* > Amean elapsed-48 2.40 ( 0.00%) 2.42 * -1.00%* > Amean elapsed-79 2.05 ( 0.00%) 2.09 * -1.90%* > Amean elapsed-110 1.83 ( 0.00%) 1.91 * -4.60%* > Amean elapsed-128 1.75 ( 0.00%) 1.85 * -5.99%* > > - IBM 9006-22C system, 2 sockets, POWER9, 64-Core, 1 NUMA node per cpu, > pppc64le. > > pft timings: > vanilla rcu > Amean system-1 1.82 ( 0.00%) 1.85 * -1.43%* > Amean system-4 2.18 ( 0.00%) 2.22 * -2.02%* > Amean system-7 3.27 ( 0.00%) 3.28 * -0.15%* > Amean system-12 5.22 ( 0.00%) 5.20 * 0.26%* > Amean system-21 10.10 ( 0.00%) 10.20 * -1.00%* > Amean system-30 15.00 ( 0.00%) 14.52 * 3.20%* > Amean system-48 26.41 ( 0.00%) 25.96 * 1.71%* > Amean system-79 29.35 ( 0.00%) 29.70 * -1.21%* > Amean system-110 24.01 ( 0.00%) 23.40 * 2.54%* > Amean system-128 24.57 ( 0.00%) 25.32 * -3.06%* > Amean elapsed-1 1.85 ( 0.00%) 1.87 * -1.28%* > Amean elapsed-4 0.56 ( 0.00%) 0.57 * -1.72%* > Amean elapsed-7 0.51 ( 0.00%) 0.50 * 0.07%* > Amean elapsed-12 0.51 ( 0.00%) 0.51 * 0.06%* > Amean elapsed-21 0.54 ( 0.00%) 0.54 * 0.06%* > Amean elapsed-30 0.54 ( 0.00%) 0.53 * 2.22%* > Amean elapsed-48 0.58 ( 0.00%) 0.57 * 1.73%* > Amean elapsed-79 0.49 ( 0.00%) 0.48 * 0.89%* > Amean elapsed-110 0.37 ( 0.00%) 0.37 * -1.08%* > Amean elapsed-128 0.33 ( 0.00%) 0.33 * 0.00%* > > - Ampere MtSnow, 1 socket, Neoverse-N1, 80-Cores, 1 NUMA node, arm64. > > pft timings: > vanilla rcu > Amean system-1 11.92 ( 0.00%) 11.99 * -0.61%* > Amean system-4 13.13 ( 0.00%) 13.09 * 0.31%* > Amean system-7 13.91 ( 0.00%) 13.94 * -0.20%* > Amean system-12 15.77 ( 0.00%) 15.69 * 0.48%* > Amean system-21 21.32 ( 0.00%) 21.42 * -0.46%* > Amean system-30 28.58 ( 0.00%) 29.12 * -1.90%* > Amean system-48 47.41 ( 0.00%) 46.91 * 1.04%* > Amean system-79 76.76 ( 0.00%) 77.16 * -0.52%* > Amean system-80 77.98 ( 0.00%) 78.23 * -0.32%* > Amean elapsed-1 12.46 ( 0.00%) 12.53 * -0.58%* > Amean elapsed-4 3.47 ( 0.00%) 3.46 * 0.34%* > Amean elapsed-7 2.18 ( 0.00%) 2.21 * -1.58%* > Amean elapsed-12 1.41 ( 0.00%) 1.42 * -0.80%* > Amean elapsed-21 1.09 ( 0.00%) 1.12 * -2.60%* > Amean elapsed-30 0.98 ( 0.00%) 1.01 * -3.08%* > Amean elapsed-48 1.08 ( 0.00%) 1.10 * -1.78%* > Amean elapsed-79 1.32 ( 0.00%) 1.28 * 2.71%* > Amean elapsed-80 1.32 ( 0.00%) 1.28 * 3.23%* > > - Dell R430, 2 sockets, Intel Xeon E5-2640 v3, Sandy Bridge, 16-Cores, 2 NUMA > nodes, x86_64. > > pft timings: > vanilla rcu > Amean system-1 11.10 ( 0.00%) 11.07 * 0.24%* > Amean system-3 11.14 ( 0.00%) 11.10 * 0.34%* > Amean system-5 11.18 ( 0.00%) 11.13 * 0.47%* > Amean system-7 11.21 ( 0.00%) 11.17 * 0.38%* > Amean system-12 11.28 ( 0.00%) 11.28 ( -0.03%) > Amean system-18 13.24 ( 0.00%) 13.25 * -0.11%* > Amean system-24 17.12 ( 0.00%) 17.14 ( -0.13%) > Amean system-30 21.10 ( 0.00%) 21.23 * -0.60%* > Amean system-32 22.31 ( 0.00%) 22.47 * -0.71%* > Amean elapsed-1 11.76 ( 0.00%) 11.73 * 0.29%* > Amean elapsed-3 3.93 ( 0.00%) 3.93 * 0.17%* > Amean elapsed-5 2.39 ( 0.00%) 2.37 * 0.74%* > Amean elapsed-7 1.72 ( 0.00%) 1.71 * 0.81%* > Amean elapsed-12 1.02 ( 0.00%) 1.03 ( -0.42%) > Amean elapsed-18 1.13 ( 0.00%) 1.14 ( -0.18%) > Amean elapsed-24 0.87 ( 0.00%) 0.88 * -0.65%* > Amean elapsed-30 0.77 ( 0.00%) 0.78 * -0.86%* > Amean elapsed-32 0.74 ( 0.00%) 0.74 ( 0.00%) > > - HPE Apollo 70, 2 sockets, Cavium ThunderX2, 128-Cores, 2 NUMA nodes, arm64. > NOTE: The test here only goes up to 128 for some reason, although there are > 256 CPUs. Maybe a mmtests issue? I didn't investigate. > > pft timings: > vanilla rcu > Amean system-1 4.42 ( 0.00%) 4.36 * 1.29%* > Amean system-4 4.56 ( 0.00%) 4.51 * 1.05%* > Amean system-7 4.63 ( 0.00%) 4.65 * -0.42%* > Amean system-12 5.96 ( 0.00%) 6.02 * -1.00%* > Amean system-21 10.97 ( 0.00%) 11.01 * -0.32%* > Amean system-30 16.01 ( 0.00%) 16.04 * -0.19%* > Amean system-48 26.81 ( 0.00%) 26.78 * 0.09%* > Amean system-79 30.80 ( 0.00%) 30.85 * -0.16%* > Amean system-110 31.87 ( 0.00%) 31.93 * -0.19%* > Amean system-128 36.27 ( 0.00%) 36.31 * -0.10%* > Amean elapsed-1 4.88 ( 0.00%) 4.85 * 0.60%* > Amean elapsed-4 1.27 ( 0.00%) 1.26 * 1.00%* > Amean elapsed-7 0.73 ( 0.00%) 0.74 * -0.46%* > Amean elapsed-12 0.55 ( 0.00%) 0.55 * 1.09%* > Amean elapsed-21 0.59 ( 0.00%) 0.60 * -0.96%* > Amean elapsed-30 0.60 ( 0.00%) 0.60 * 0.28%* > Amean elapsed-48 0.60 ( 0.00%) 0.60 * 0.44%* > Amean elapsed-79 0.49 ( 0.00%) 0.49 * -0.07%* > Amean elapsed-110 0.36 ( 0.00%) 0.36 * 0.28%* > Amean elapsed-128 0.31 ( 0.00%) 0.31 * -0.43%* > > - Raspberry Pi 4, 1 socket, bcm2711, Cortex-A72, 4-Cores, 1 NUMA node, arm64. > > pft timings: > vanilla rcu > Amean system-1 0.67 ( 0.00%) 0.67 * -1.25%* > Amean system-3 1.30 ( 0.00%) 1.29 * 0.62%* > Amean system-4 1.61 ( 0.00%) 1.59 * 0.95%* > Amean elapsed-1 0.71 ( 0.00%) 0.72 * -1.17%* > Amean elapsed-3 0.45 ( 0.00%) 0.45 * 0.88%* > Amean elapsed-4 0.42 ( 0.00%) 0.42 * 1.19%* > >
On Thu, 2022-02-10 at 18:59 +0800, Xiongfeng Wang wrote: > Hi Nicolas, > > I found there are several 'lru_add_drain_per_cpu()' worker thread on the > NOHZ_FULL CPUs. I am just wondering do you have plan to add support for remote > per-cpu LRU pagevec drain ? Yes, this is something we've been meaning to fix too. But haven't started looking into it yet (maybe Marcelo can contradict me here). Regards, Nicolas
On Tue, Feb 08, 2022 at 11:07:48AM +0100, Nicolas Saenz Julienne wrote: > This series replaces mm/page_alloc's per-cpu page lists drain mechanism with > one that allows accessing the lists remotely. Currently, only a local CPU is > permitted to change its per-cpu lists, and it's expected to do so, on-demand, > whenever a process demands it by means of queueing a drain task on the local > CPU. This causes problems for NOHZ_FULL CPUs and real-time systems that can't > take any sort of interruption and to some lesser extent inconveniences idle and > virtualised systems. > I know this has been sitting here for a long while. Last few weeks have not been fun. > Note that this is not the first attempt at fixing this per-cpu page lists: > - The first attempt[1] tried to conditionally change the pagesets locking > scheme based the NOHZ_FULL config. It was deemed hard to maintain as the > NOHZ_FULL code path would be rarely tested. Also, this only solves the issue > for NOHZ_FULL setups, which isn't ideal. > - The second[2] unanimously switched the local_locks to per-cpu spinlocks. The > performance degradation was too big. > For unrelated reasons I looked at using llist to avoid locks entirely. It turns out it's not possible and needs a lock. We know "local_locks to per-cpu spinlocks" took a large penalty so I considered alternatives on how a lock could be used. I found it's possible to both remote drain the lists and avoid the disable/enable of IRQs entirely as long as a preempting IRQ is willing to take the zone lock instead (should be very rare). The IRQ part is a bit hairy though as softirqs are also a problem and preempt-rt needs different rules and the llist has to sort PCP refills which might be a loss in total. However, the remote draining may still be interesting. The full series is at https://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git/ mm-pcpllist-v1r2 It's still waiting on tests to complete and not all the changelogs are complete which is why it's not posted. This is a comparison of vanilla vs "local_locks to per-cpu spinlocks" versus the git series up to "mm/page_alloc: Remotely drain per-cpu lists" for the page faulting microbench I originally complained about. The test machine is a 2-socket CascadeLake machine. pft timings 5.17.0-rc5 5.17.0-rc5 5.17.0-rc5 vanilla mm-remotedrain-v2r1 mm-pcpdrain-v1r1 Amean elapsed-1 32.54 ( 0.00%) 33.08 * -1.66%* 32.82 * -0.86%* Amean elapsed-4 8.66 ( 0.00%) 9.24 * -6.72%* 8.69 * -0.38%* Amean elapsed-7 5.02 ( 0.00%) 5.43 * -8.16%* 5.05 * -0.55%* Amean elapsed-12 3.07 ( 0.00%) 3.38 * -10.00%* 3.09 * -0.72%* Amean elapsed-21 2.36 ( 0.00%) 2.38 * -0.89%* 2.19 * 7.39%* Amean elapsed-30 1.75 ( 0.00%) 1.87 * -6.50%* 1.62 * 7.59%* Amean elapsed-48 1.71 ( 0.00%) 2.00 * -17.32%* 1.71 ( -0.08%) Amean elapsed-79 1.56 ( 0.00%) 1.62 * -3.84%* 1.56 ( -0.02%) Amean elapsed-80 1.57 ( 0.00%) 1.65 * -5.31%* 1.57 ( -0.04%) Note the local_lock conversion took 1 1-17% penalty while the git tree takes a negligile penalty while still allowing remote drains. It might have some potential while being less complex than the RCU approach.
On 2/8/22 11:07, Nicolas Saenz Julienne wrote: > This series replaces mm/page_alloc's per-cpu page lists drain mechanism with > one that allows accessing the lists remotely. Currently, only a local CPU is > permitted to change its per-cpu lists, and it's expected to do so, on-demand, > whenever a process demands it by means of queueing a drain task on the local > CPU. This causes problems for NOHZ_FULL CPUs and real-time systems that can't > take any sort of interruption and to some lesser extent inconveniences idle and > virtualised systems. > > The new algorithm will atomically switch the pointer to the per-cpu page lists > and use RCU to make sure it's not being concurrently used before draining the > lists. And its main benefit of is that it fixes the issue for good, avoiding > the need for configuration based heuristics or having to modify applications > (i.e. using the isolation prctrl being worked by Marcello Tosatti ATM). > > All this with minimal performance implications: a page allocation > microbenchmark was run on multiple systems and architectures generally showing > no performance differences, only the more extreme cases showed a 1-3% > degradation. See data below. Needless to say that I'd appreciate if someone > could validate my values independently. > > The approach has been stress-tested: I forced 100 drains/s while running > mmtests' pft in a loop for a full day on multiple machines and archs (arm64, > x86_64, ppc64le). > > Note that this is not the first attempt at fixing this per-cpu page lists: > - The first attempt[1] tried to conditionally change the pagesets locking > scheme based the NOHZ_FULL config. It was deemed hard to maintain as the > NOHZ_FULL code path would be rarely tested. Also, this only solves the issue > for NOHZ_FULL setups, which isn't ideal. > - The second[2] unanimously switched the local_locks to per-cpu spinlocks. The > performance degradation was too big. For completeness, what was the fate of the approach to have pcp->high = 0 for NOHZ cpus? [1] It would be nice to have documented why it wasn't feasible. Too much overhead for when these CPUs eventually do allocate, or some other unforeseen issue? Thanks. [1] https://lore.kernel.org/all/43462fe11258395f4e885c3d594a3ed1b604b858.camel@redhat.com/ > Previous RFC: https://lkml.org/lkml/2021/10/8/793 > > Thanks! > > [1] https://lkml.org/lkml/2021/9/21/599 > [2] https://lkml.org/lkml/2021/11/3/644 > > --- > > Changes since RFC: > - Get rid of aesthetic changes that affected performance > - Add more documentation > - Add better commit messages > - Pass sparse tests > - Verify this_cpu_*() usage > - Performance measurements > > Nicolas Saenz Julienne (2): > mm/page_alloc: Access lists in 'struct per_cpu_pages' indirectly > mm/page_alloc: Add remote draining support to per-cpu lists > > include/linux/mmzone.h | 28 +++++- > mm/page_alloc.c | 212 ++++++++++++++++++++++++++--------------- > mm/vmstat.c | 6 +- > 3 files changed, 162 insertions(+), 84 deletions(-) > > > -------------------------Performance results----------------------------- > > I'm focusing on mmtests' Page Fault Test (pft), as it's page allocator > intensive. > > - AMD Daytona Reference System, 2 sockets, AMD EPYC 7742, Zen 2, 64-Core, > 4 NUMA nodes, x86_64 > > pft timings: > vanilla rcu > Amean system-1 58.52 ( 0.00%) 58.92 * -0.68%* > Amean system-4 61.00 ( 0.00%) 61.41 * -0.67%* > Amean system-7 61.55 ( 0.00%) 61.74 * -0.30%* > Amean system-12 64.91 ( 0.00%) 64.94 * -0.05%* > Amean system-21 98.80 ( 0.00%) 99.92 * -1.13%* > Amean system-30 147.68 ( 0.00%) 145.83 * 1.25%* > Amean system-48 237.04 ( 0.00%) 241.29 * -1.79%* > Amean system-79 286.61 ( 0.00%) 283.72 * 1.01%* > Amean system-110 303.40 ( 0.00%) 299.91 * 1.15%* > Amean system-128 345.07 ( 0.00%) 342.10 * 0.86%* > Amean elapsed-1 61.21 ( 0.00%) 61.65 * -0.71%* > Amean elapsed-4 15.94 ( 0.00%) 16.05 * -0.69%* > Amean elapsed-7 9.24 ( 0.00%) 9.28 * -0.47%* > Amean elapsed-12 5.70 ( 0.00%) 5.70 * -0.07%* > Amean elapsed-21 5.11 ( 0.00%) 5.06 * 1.13%* > Amean elapsed-30 5.28 ( 0.00%) 5.14 * 2.73%* > Amean elapsed-48 5.28 ( 0.00%) 5.24 * 0.74%* > Amean elapsed-79 4.41 ( 0.00%) 4.31 * 2.17%* > Amean elapsed-110 3.45 ( 0.00%) 3.44 * 0.40%* > Amean elapsed-128 2.75 ( 0.00%) 2.75 * -0.28%* > > - AMD Speedway Reference System, 2 sockets, AMD EPYC 7601, Zen 1, 64-core, 8 > NUMA nodes, x86_64. Lots of variance between tests on this platform. It'll > easily swing -+5% on each result. > > pft timings: > vanilla rcu > Amean system-1 69.20 ( 0.00%) 66.21 * 4.32%* > Amean system-4 70.79 ( 0.00%) 69.01 * 2.52%* > Amean system-7 71.34 ( 0.00%) 69.16 * 3.05%* > Amean system-12 74.00 ( 0.00%) 72.74 * 1.70%* > Amean system-21 86.01 ( 0.00%) 85.70 * 0.36%* > Amean system-30 89.21 ( 0.00%) 89.93 * -0.80%* > Amean system-48 92.39 ( 0.00%) 92.43 * -0.04%* > Amean system-79 120.19 ( 0.00%) 121.30 * -0.92%* > Amean system-110 172.79 ( 0.00%) 179.37 * -3.81%* > Amean system-128 201.70 ( 0.00%) 212.57 * -5.39%* > Amean elapsed-1 72.23 ( 0.00%) 69.29 * 4.08%* > Amean elapsed-4 18.69 ( 0.00%) 18.28 * 2.20%* > Amean elapsed-7 10.80 ( 0.00%) 10.54 * 2.41%* > Amean elapsed-12 6.62 ( 0.00%) 6.53 * 1.30%* > Amean elapsed-21 4.68 ( 0.00%) 4.69 * -0.14%* > Amean elapsed-30 3.44 ( 0.00%) 3.50 * -1.66%* > Amean elapsed-48 2.40 ( 0.00%) 2.42 * -1.00%* > Amean elapsed-79 2.05 ( 0.00%) 2.09 * -1.90%* > Amean elapsed-110 1.83 ( 0.00%) 1.91 * -4.60%* > Amean elapsed-128 1.75 ( 0.00%) 1.85 * -5.99%* > > - IBM 9006-22C system, 2 sockets, POWER9, 64-Core, 1 NUMA node per cpu, > pppc64le. > > pft timings: > vanilla rcu > Amean system-1 1.82 ( 0.00%) 1.85 * -1.43%* > Amean system-4 2.18 ( 0.00%) 2.22 * -2.02%* > Amean system-7 3.27 ( 0.00%) 3.28 * -0.15%* > Amean system-12 5.22 ( 0.00%) 5.20 * 0.26%* > Amean system-21 10.10 ( 0.00%) 10.20 * -1.00%* > Amean system-30 15.00 ( 0.00%) 14.52 * 3.20%* > Amean system-48 26.41 ( 0.00%) 25.96 * 1.71%* > Amean system-79 29.35 ( 0.00%) 29.70 * -1.21%* > Amean system-110 24.01 ( 0.00%) 23.40 * 2.54%* > Amean system-128 24.57 ( 0.00%) 25.32 * -3.06%* > Amean elapsed-1 1.85 ( 0.00%) 1.87 * -1.28%* > Amean elapsed-4 0.56 ( 0.00%) 0.57 * -1.72%* > Amean elapsed-7 0.51 ( 0.00%) 0.50 * 0.07%* > Amean elapsed-12 0.51 ( 0.00%) 0.51 * 0.06%* > Amean elapsed-21 0.54 ( 0.00%) 0.54 * 0.06%* > Amean elapsed-30 0.54 ( 0.00%) 0.53 * 2.22%* > Amean elapsed-48 0.58 ( 0.00%) 0.57 * 1.73%* > Amean elapsed-79 0.49 ( 0.00%) 0.48 * 0.89%* > Amean elapsed-110 0.37 ( 0.00%) 0.37 * -1.08%* > Amean elapsed-128 0.33 ( 0.00%) 0.33 * 0.00%* > > - Ampere MtSnow, 1 socket, Neoverse-N1, 80-Cores, 1 NUMA node, arm64. > > pft timings: > vanilla rcu > Amean system-1 11.92 ( 0.00%) 11.99 * -0.61%* > Amean system-4 13.13 ( 0.00%) 13.09 * 0.31%* > Amean system-7 13.91 ( 0.00%) 13.94 * -0.20%* > Amean system-12 15.77 ( 0.00%) 15.69 * 0.48%* > Amean system-21 21.32 ( 0.00%) 21.42 * -0.46%* > Amean system-30 28.58 ( 0.00%) 29.12 * -1.90%* > Amean system-48 47.41 ( 0.00%) 46.91 * 1.04%* > Amean system-79 76.76 ( 0.00%) 77.16 * -0.52%* > Amean system-80 77.98 ( 0.00%) 78.23 * -0.32%* > Amean elapsed-1 12.46 ( 0.00%) 12.53 * -0.58%* > Amean elapsed-4 3.47 ( 0.00%) 3.46 * 0.34%* > Amean elapsed-7 2.18 ( 0.00%) 2.21 * -1.58%* > Amean elapsed-12 1.41 ( 0.00%) 1.42 * -0.80%* > Amean elapsed-21 1.09 ( 0.00%) 1.12 * -2.60%* > Amean elapsed-30 0.98 ( 0.00%) 1.01 * -3.08%* > Amean elapsed-48 1.08 ( 0.00%) 1.10 * -1.78%* > Amean elapsed-79 1.32 ( 0.00%) 1.28 * 2.71%* > Amean elapsed-80 1.32 ( 0.00%) 1.28 * 3.23%* > > - Dell R430, 2 sockets, Intel Xeon E5-2640 v3, Sandy Bridge, 16-Cores, 2 NUMA > nodes, x86_64. > > pft timings: > vanilla rcu > Amean system-1 11.10 ( 0.00%) 11.07 * 0.24%* > Amean system-3 11.14 ( 0.00%) 11.10 * 0.34%* > Amean system-5 11.18 ( 0.00%) 11.13 * 0.47%* > Amean system-7 11.21 ( 0.00%) 11.17 * 0.38%* > Amean system-12 11.28 ( 0.00%) 11.28 ( -0.03%) > Amean system-18 13.24 ( 0.00%) 13.25 * -0.11%* > Amean system-24 17.12 ( 0.00%) 17.14 ( -0.13%) > Amean system-30 21.10 ( 0.00%) 21.23 * -0.60%* > Amean system-32 22.31 ( 0.00%) 22.47 * -0.71%* > Amean elapsed-1 11.76 ( 0.00%) 11.73 * 0.29%* > Amean elapsed-3 3.93 ( 0.00%) 3.93 * 0.17%* > Amean elapsed-5 2.39 ( 0.00%) 2.37 * 0.74%* > Amean elapsed-7 1.72 ( 0.00%) 1.71 * 0.81%* > Amean elapsed-12 1.02 ( 0.00%) 1.03 ( -0.42%) > Amean elapsed-18 1.13 ( 0.00%) 1.14 ( -0.18%) > Amean elapsed-24 0.87 ( 0.00%) 0.88 * -0.65%* > Amean elapsed-30 0.77 ( 0.00%) 0.78 * -0.86%* > Amean elapsed-32 0.74 ( 0.00%) 0.74 ( 0.00%) > > - HPE Apollo 70, 2 sockets, Cavium ThunderX2, 128-Cores, 2 NUMA nodes, arm64. > NOTE: The test here only goes up to 128 for some reason, although there are > 256 CPUs. Maybe a mmtests issue? I didn't investigate. > > pft timings: > vanilla rcu > Amean system-1 4.42 ( 0.00%) 4.36 * 1.29%* > Amean system-4 4.56 ( 0.00%) 4.51 * 1.05%* > Amean system-7 4.63 ( 0.00%) 4.65 * -0.42%* > Amean system-12 5.96 ( 0.00%) 6.02 * -1.00%* > Amean system-21 10.97 ( 0.00%) 11.01 * -0.32%* > Amean system-30 16.01 ( 0.00%) 16.04 * -0.19%* > Amean system-48 26.81 ( 0.00%) 26.78 * 0.09%* > Amean system-79 30.80 ( 0.00%) 30.85 * -0.16%* > Amean system-110 31.87 ( 0.00%) 31.93 * -0.19%* > Amean system-128 36.27 ( 0.00%) 36.31 * -0.10%* > Amean elapsed-1 4.88 ( 0.00%) 4.85 * 0.60%* > Amean elapsed-4 1.27 ( 0.00%) 1.26 * 1.00%* > Amean elapsed-7 0.73 ( 0.00%) 0.74 * -0.46%* > Amean elapsed-12 0.55 ( 0.00%) 0.55 * 1.09%* > Amean elapsed-21 0.59 ( 0.00%) 0.60 * -0.96%* > Amean elapsed-30 0.60 ( 0.00%) 0.60 * 0.28%* > Amean elapsed-48 0.60 ( 0.00%) 0.60 * 0.44%* > Amean elapsed-79 0.49 ( 0.00%) 0.49 * -0.07%* > Amean elapsed-110 0.36 ( 0.00%) 0.36 * 0.28%* > Amean elapsed-128 0.31 ( 0.00%) 0.31 * -0.43%* > > - Raspberry Pi 4, 1 socket, bcm2711, Cortex-A72, 4-Cores, 1 NUMA node, arm64. > > pft timings: > vanilla rcu > Amean system-1 0.67 ( 0.00%) 0.67 * -1.25%* > Amean system-3 1.30 ( 0.00%) 1.29 * 0.62%* > Amean system-4 1.61 ( 0.00%) 1.59 * 0.95%* > Amean elapsed-1 0.71 ( 0.00%) 0.72 * -1.17%* > Amean elapsed-3 0.45 ( 0.00%) 0.45 * 0.88%* > Amean elapsed-4 0.42 ( 0.00%) 0.42 * 1.19%* > >
On Thu, 2022-03-03 at 14:27 +0100, Vlastimil Babka wrote: > On 2/8/22 11:07, Nicolas Saenz Julienne wrote: > > This series replaces mm/page_alloc's per-cpu page lists drain mechanism with > > one that allows accessing the lists remotely. Currently, only a local CPU is > > permitted to change its per-cpu lists, and it's expected to do so, on-demand, > > whenever a process demands it by means of queueing a drain task on the local > > CPU. This causes problems for NOHZ_FULL CPUs and real-time systems that can't > > take any sort of interruption and to some lesser extent inconveniences idle and > > virtualised systems. > > > > The new algorithm will atomically switch the pointer to the per-cpu page lists > > and use RCU to make sure it's not being concurrently used before draining the > > lists. And its main benefit of is that it fixes the issue for good, avoiding > > the need for configuration based heuristics or having to modify applications > > (i.e. using the isolation prctrl being worked by Marcello Tosatti ATM). > > > > All this with minimal performance implications: a page allocation > > microbenchmark was run on multiple systems and architectures generally showing > > no performance differences, only the more extreme cases showed a 1-3% > > degradation. See data below. Needless to say that I'd appreciate if someone > > could validate my values independently. > > > > The approach has been stress-tested: I forced 100 drains/s while running > > mmtests' pft in a loop for a full day on multiple machines and archs (arm64, > > x86_64, ppc64le). > > > > Note that this is not the first attempt at fixing this per-cpu page lists: > > - The first attempt[1] tried to conditionally change the pagesets locking > > scheme based the NOHZ_FULL config. It was deemed hard to maintain as the > > NOHZ_FULL code path would be rarely tested. Also, this only solves the issue > > for NOHZ_FULL setups, which isn't ideal. > > - The second[2] unanimously switched the local_locks to per-cpu spinlocks. The > > performance degradation was too big. > > For completeness, what was the fate of the approach to have pcp->high = 0 > for NOHZ cpus? [1] It would be nice to have documented why it wasn't > feasible. Too much overhead for when these CPUs eventually do allocate, or > some other unforeseen issue? Thanks. Yes sorry, should've been more explicit on why I haven't gone that way yet. Some points: - As I mention above, not only CPU isolation users care for this. RT and HPC do too. This is my main motivation for focusing on this solution, or potentially Mel's. - Fully disabling pcplists on nohz_full CPUs is too drastic, as isolated CPUs might want to retain the performance edge while not running their sensitive workloads. (I remember Christoph Lamenter's commenting about this on the previous RFC). - So the idea would be to selectively disable pcplists upon entering in the really 'isolated' area. This could be achieved with Marcelo Tosatti's new WIP prctrl[1]. And if we decide the current solutions are unacceptable I'll have a go at it. Thanks! [1] https://lore.kernel.org/lkml/20220204173554.534186379@fedora.localdomain/T/
Hi Mel, Thanks for having a look at this. On Thu, 2022-03-03 at 11:45 +0000, Mel Gorman wrote: > On Tue, Feb 08, 2022 at 11:07:48AM +0100, Nicolas Saenz Julienne wrote: > > This series replaces mm/page_alloc's per-cpu page lists drain mechanism with > > one that allows accessing the lists remotely. Currently, only a local CPU is > > permitted to change its per-cpu lists, and it's expected to do so, on-demand, > > whenever a process demands it by means of queueing a drain task on the local > > CPU. This causes problems for NOHZ_FULL CPUs and real-time systems that can't > > take any sort of interruption and to some lesser extent inconveniences idle and > > virtualised systems. > > > > I know this has been sitting here for a long while. Last few weeks have > not been fun. No problem. > > Note that this is not the first attempt at fixing this per-cpu page lists: > > - The first attempt[1] tried to conditionally change the pagesets locking > > scheme based the NOHZ_FULL config. It was deemed hard to maintain as the > > NOHZ_FULL code path would be rarely tested. Also, this only solves the issue > > for NOHZ_FULL setups, which isn't ideal. > > - The second[2] unanimously switched the local_locks to per-cpu spinlocks. The > > performance degradation was too big. > > > > For unrelated reasons I looked at using llist to avoid locks entirely. It > turns out it's not possible and needs a lock. We know "local_locks to > per-cpu spinlocks" took a large penalty so I considered alternatives on > how a lock could be used. I found it's possible to both remote drain > the lists and avoid the disable/enable of IRQs entirely as long as a > preempting IRQ is willing to take the zone lock instead (should be very > rare). The IRQ part is a bit hairy though as softirqs are also a problem > and preempt-rt needs different rules and the llist has to sort PCP > refills which might be a loss in total. However, the remote draining may > still be interesting. The full series is at > https://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git/ mm-pcpllist-v1r2 I'll have a proper look at it soon. Regards,
On Thu, Mar 03, 2022 at 11:45:50AM +0000, Mel Gorman wrote: > On Tue, Feb 08, 2022 at 11:07:48AM +0100, Nicolas Saenz Julienne wrote: > > This series replaces mm/page_alloc's per-cpu page lists drain mechanism with > > one that allows accessing the lists remotely. Currently, only a local CPU is > > permitted to change its per-cpu lists, and it's expected to do so, on-demand, > > whenever a process demands it by means of queueing a drain task on the local > > CPU. This causes problems for NOHZ_FULL CPUs and real-time systems that can't > > take any sort of interruption and to some lesser extent inconveniences idle and > > virtualised systems. > > > > I know this has been sitting here for a long while. Last few weeks have > not been fun. > > > Note that this is not the first attempt at fixing this per-cpu page lists: > > - The first attempt[1] tried to conditionally change the pagesets locking > > scheme based the NOHZ_FULL config. It was deemed hard to maintain as the > > NOHZ_FULL code path would be rarely tested. Also, this only solves the issue > > for NOHZ_FULL setups, which isn't ideal. > > - The second[2] unanimously switched the local_locks to per-cpu spinlocks. The > > performance degradation was too big. > > > > For unrelated reasons I looked at using llist to avoid locks entirely. It > turns out it's not possible and needs a lock. We know "local_locks to > per-cpu spinlocks" took a large penalty so I considered alternatives on > how a lock could be used. I found it's possible to both remote drain > the lists and avoid the disable/enable of IRQs entirely as long as a > preempting IRQ is willing to take the zone lock instead (should be very > rare). The IRQ part is a bit hairy though as softirqs are also a problem > and preempt-rt needs different rules and the llist has to sort PCP > refills which might be a loss in total. However, the remote draining may > still be interesting. The full series is at > https://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git/ mm-pcpllist-v1r2 > > It's still waiting on tests to complete and not all the changelogs are > complete which is why it's not posted. > > This is a comparison of vanilla vs "local_locks to per-cpu spinlocks" > versus the git series up to "mm/page_alloc: Remotely drain per-cpu lists" > for the page faulting microbench I originally complained about. The test > machine is a 2-socket CascadeLake machine. > > pft timings > 5.17.0-rc5 5.17.0-rc5 5.17.0-rc5 > vanilla mm-remotedrain-v2r1 mm-pcpdrain-v1r1 > Amean elapsed-1 32.54 ( 0.00%) 33.08 * -1.66%* 32.82 * -0.86%* > Amean elapsed-4 8.66 ( 0.00%) 9.24 * -6.72%* 8.69 * -0.38%* > Amean elapsed-7 5.02 ( 0.00%) 5.43 * -8.16%* 5.05 * -0.55%* > Amean elapsed-12 3.07 ( 0.00%) 3.38 * -10.00%* 3.09 * -0.72%* > Amean elapsed-21 2.36 ( 0.00%) 2.38 * -0.89%* 2.19 * 7.39%* > Amean elapsed-30 1.75 ( 0.00%) 1.87 * -6.50%* 1.62 * 7.59%* > Amean elapsed-48 1.71 ( 0.00%) 2.00 * -17.32%* 1.71 ( -0.08%) > Amean elapsed-79 1.56 ( 0.00%) 1.62 * -3.84%* 1.56 ( -0.02%) > Amean elapsed-80 1.57 ( 0.00%) 1.65 * -5.31%* 1.57 ( -0.04%) > > Note the local_lock conversion took 1 1-17% penalty while the git tree > takes a negligile penalty while still allowing remote drains. It might > have some potential while being less complex than the RCU approach. Nice! Hopefully a spinlock can be added to "struct lru_pvecs" without degradation in performance, similarly to what is done here.
On Mon, Mar 07, 2022 at 02:57:47PM +0100, Nicolas Saenz Julienne wrote: > > > Note that this is not the first attempt at fixing this per-cpu page lists: > > > - The first attempt[1] tried to conditionally change the pagesets locking > > > scheme based the NOHZ_FULL config. It was deemed hard to maintain as the > > > NOHZ_FULL code path would be rarely tested. Also, this only solves the issue > > > for NOHZ_FULL setups, which isn't ideal. > > > - The second[2] unanimously switched the local_locks to per-cpu spinlocks. The > > > performance degradation was too big. > > > > > > > For unrelated reasons I looked at using llist to avoid locks entirely. It > > turns out it's not possible and needs a lock. We know "local_locks to > > per-cpu spinlocks" took a large penalty so I considered alternatives on > > how a lock could be used. I found it's possible to both remote drain > > the lists and avoid the disable/enable of IRQs entirely as long as a > > preempting IRQ is willing to take the zone lock instead (should be very > > rare). The IRQ part is a bit hairy though as softirqs are also a problem > > and preempt-rt needs different rules and the llist has to sort PCP > > refills which might be a loss in total. However, the remote draining may > > still be interesting. The full series is at > > https://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git/ mm-pcpllist-v1r2 > > I'll have a proper look at it soon. > Thanks. I'm still delayed actually finishing the series as most of my time is dedicated to a separate issue. However, there is at least one bug in there at patch "mm/page_alloc: Remotely drain per-cpu lists" that causes a lockup under severe memory pressure. The fix is diff --git a/mm/page_alloc.c b/mm/page_alloc.c index c9a6f2b5548e..11b54f383d04 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -3065,10 +3065,8 @@ static int rmqueue_bulk(struct zone *zone, unsigned int order, */ void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) { - unsigned long flags; int to_drain, batch; - pcp_local_lock(&pagesets.lock, flags); batch = READ_ONCE(pcp->batch); to_drain = min(pcp->count, batch); if (to_drain > 0) { @@ -3076,7 +3074,6 @@ void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) free_pcppages_bulk(zone, to_drain, pcp, 0); spin_unlock(&pcp->lock); } - pcp_local_unlock(&pagesets.lock, flags); } #endif @@ -3088,16 +3085,12 @@ static void drain_pages_zone(unsigned int cpu, struct zone *zone) unsigned long flags; struct per_cpu_pages *pcp; - pcp_local_lock(&pagesets.lock, flags); - pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); if (pcp->count) { spin_lock(&pcp->lock); free_pcppages_bulk(zone, pcp->count, pcp, 0); spin_unlock(&pcp->lock); } - - pcp_local_unlock(&pagesets.lock, flags); } /*
Hi Mel, On Thu, 2022-03-03 at 11:45 +0000, Mel Gorman wrote: > For unrelated reasons I looked at using llist to avoid locks entirely. It > turns out it's not possible and needs a lock. We know "local_locks to > per-cpu spinlocks" took a large penalty so I considered alternatives on > how a lock could be used. I found it's possible to both remote drain > the lists and avoid the disable/enable of IRQs entirely as long as a > preempting IRQ is willing to take the zone lock instead (should be very > rare). The IRQ part is a bit hairy though as softirqs are also a problem > and preempt-rt needs different rules and the llist has to sort PCP > refills which might be a loss in total. However, the remote draining may > still be interesting. The full series is at > https://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git/ mm-pcpllist-v1r2 > > It's still waiting on tests to complete and not all the changelogs are > complete which is why it's not posted. > > This is a comparison of vanilla vs "local_locks to per-cpu spinlocks" > versus the git series up to "mm/page_alloc: Remotely drain per-cpu lists" > for the page faulting microbench I originally complained about. The test > machine is a 2-socket CascadeLake machine. > > pft timings > 5.17.0-rc5 5.17.0-rc5 5.17.0-rc5 > vanilla mm-remotedrain-v2r1 mm-pcpdrain-v1r1 > Amean elapsed-1 32.54 ( 0.00%) 33.08 * -1.66%* 32.82 * -0.86%* > Amean elapsed-4 8.66 ( 0.00%) 9.24 * -6.72%* 8.69 * -0.38%* > Amean elapsed-7 5.02 ( 0.00%) 5.43 * -8.16%* 5.05 * -0.55%* > Amean elapsed-12 3.07 ( 0.00%) 3.38 * -10.00%* 3.09 * -0.72%* > Amean elapsed-21 2.36 ( 0.00%) 2.38 * -0.89%* 2.19 * 7.39%* > Amean elapsed-30 1.75 ( 0.00%) 1.87 * -6.50%* 1.62 * 7.59%* > Amean elapsed-48 1.71 ( 0.00%) 2.00 * -17.32%* 1.71 ( -0.08%) > Amean elapsed-79 1.56 ( 0.00%) 1.62 * -3.84%* 1.56 ( -0.02%) > Amean elapsed-80 1.57 ( 0.00%) 1.65 * -5.31%* 1.57 ( -0.04%) > > Note the local_lock conversion took 1 1-17% penalty while the git tree > takes a negligile penalty while still allowing remote drains. It might > have some potential while being less complex than the RCU approach. I finally got some time to look at this and made some progress: First, I belive your 'mm-remotedrain-v2r1' results are wrong/inflated due to a bug in my series. Essentially, all 'this_cpu_ptr()' calls should've been 'raw_cpu_ptr()' and your build, which I bet enables CONFIG_DEBUG_PREEMPT, wasted time trowing warnings about per-cpu variable usage with preemption enabled. Making the overall performance look worse than it actually is. My build didn't enable it, which made me miss this whole issue. I'm sorry for the noise and time wasted on such a silly thing. Note that the local_lock to spin_lock conversion can handle the preeemption alright, it is part of the design[1]. As for your idea of not disabling interrupts in the hot paths, it seems to close the performance gap created by the lock conversion. That said, I'm not sure I understand why you find the need to keep the local_locks around, not only it casuses problems for RT systems, but IIUC they aren't really protecting anything other than the 'this_cpu_ptr()' usage (which isn't really needed). I rewrote your patch on top of my lock conversion series and I'm in the process of testing it on multiple systems[2]. Let me know what you think. Thanks! [1] It follows this pattern: struct per_cpu_pages *pcp; pcp = raw_cpu_ptr(page_zone(page)->per_cpu_pageset); // <- Migration here is OK: spin_lock protects vs eventual pcplist // access from local CPU as long as all list access happens through the // pcp pointer. spin_lock(&pcp->lock); do_stuff_with_pcp_lists(pcp); spin_unlock(&pcp->lock); [2] See: git://git.kernel.org/pub/scm/linux/kernel/git/nsaenz/linux-rpi.git pcpdrain-sl-v3r1
On Thu, Mar 24, 2022 at 07:59:56PM +0100, Nicolas Saenz Julienne wrote: > Hi Mel, > > On Thu, 2022-03-03 at 11:45 +0000, Mel Gorman wrote: > > For unrelated reasons I looked at using llist to avoid locks entirely. It > > turns out it's not possible and needs a lock. We know "local_locks to > > per-cpu spinlocks" took a large penalty so I considered alternatives on > > how a lock could be used. I found it's possible to both remote drain > > the lists and avoid the disable/enable of IRQs entirely as long as a > > preempting IRQ is willing to take the zone lock instead (should be very > > rare). The IRQ part is a bit hairy though as softirqs are also a problem > > and preempt-rt needs different rules and the llist has to sort PCP > > refills which might be a loss in total. However, the remote draining may > > still be interesting. The full series is at > > https://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git/ mm-pcpllist-v1r2 > > > > It's still waiting on tests to complete and not all the changelogs are > > complete which is why it's not posted. > > > > This is a comparison of vanilla vs "local_locks to per-cpu spinlocks" > > versus the git series up to "mm/page_alloc: Remotely drain per-cpu lists" > > for the page faulting microbench I originally complained about. The test > > machine is a 2-socket CascadeLake machine. > > > > pft timings > > 5.17.0-rc5 5.17.0-rc5 5.17.0-rc5 > > vanilla mm-remotedrain-v2r1 mm-pcpdrain-v1r1 > > Amean elapsed-1 32.54 ( 0.00%) 33.08 * -1.66%* 32.82 * -0.86%* > > Amean elapsed-4 8.66 ( 0.00%) 9.24 * -6.72%* 8.69 * -0.38%* > > Amean elapsed-7 5.02 ( 0.00%) 5.43 * -8.16%* 5.05 * -0.55%* > > Amean elapsed-12 3.07 ( 0.00%) 3.38 * -10.00%* 3.09 * -0.72%* > > Amean elapsed-21 2.36 ( 0.00%) 2.38 * -0.89%* 2.19 * 7.39%* > > Amean elapsed-30 1.75 ( 0.00%) 1.87 * -6.50%* 1.62 * 7.59%* > > Amean elapsed-48 1.71 ( 0.00%) 2.00 * -17.32%* 1.71 ( -0.08%) > > Amean elapsed-79 1.56 ( 0.00%) 1.62 * -3.84%* 1.56 ( -0.02%) > > Amean elapsed-80 1.57 ( 0.00%) 1.65 * -5.31%* 1.57 ( -0.04%) > > > > Note the local_lock conversion took 1 1-17% penalty while the git tree > > takes a negligile penalty while still allowing remote drains. It might > > have some potential while being less complex than the RCU approach. > > I finally got some time to look at this and made some progress: > > First, I belive your 'mm-remotedrain-v2r1' results are wrong/inflated due to a > bug in my series. Essentially, all 'this_cpu_ptr()' calls should've been > 'raw_cpu_ptr()' and your build, which I bet enables CONFIG_DEBUG_PREEMPT, I no longer have the logs but it could have and I didn't check the dmesg at the time to see if there were warnings in it. It really should add something to parse that log and automatically report if there are unexpected warnings, oops or prove-locking warnings if enabled. > wasted time trowing warnings about per-cpu variable usage with preemption > enabled. Making the overall performance look worse than it actually is. My > build didn't enable it, which made me miss this whole issue. I'm sorry for the > noise and time wasted on such a silly thing. Note that the local_lock to > spin_lock conversion can handle the preeemption alright, it is part of the > design[1]. > > As for your idea of not disabling interrupts in the hot paths, it seems to > close the performance gap created by the lock conversion. That said, I'm not > sure I understand why you find the need to keep the local_locks around, not > only it casuses problems for RT systems, but IIUC they aren't really protecting > anything other than the 'this_cpu_ptr()' usage (which isn't really needed). The local lock was preserved because something has to stabilise the per-cpu pointer due to preemption and migration. On !RT, that's a preempt_disable and on RT it's a spinlock both which prevent a migration. The secondary goal of using local lock was to allow some allocations to be done without disabling IRQs at all with the penalty that an IRQ arriving in at the wrong time will have to allocate directly from the buddy lists which should be rare. > I > rewrote your patch on top of my lock conversion series and I'm in the process > of testing it on multiple systems[2]. > > Let me know what you think. > Thanks! > > [1] It follows this pattern: > > struct per_cpu_pages *pcp; > > pcp = raw_cpu_ptr(page_zone(page)->per_cpu_pageset); > // <- Migration here is OK: spin_lock protects vs eventual pcplist > // access from local CPU as long as all list access happens through the > // pcp pointer. > spin_lock(&pcp->lock); > do_stuff_with_pcp_lists(pcp); > spin_unlock(&pcp->lock); > And this was the part I am concerned with. We are accessing a PCP structure that is not necessarily the one belonging to the CPU we are currently running on. This type of pattern is warned about in Documentation/locking/locktypes.rst ---8<--- A typical scenario is protection of per-CPU variables in thread context:: struct foo *p = get_cpu_ptr(&var1); spin_lock(&p->lock); p->count += this_cpu_read(var2); This is correct code on a non-PREEMPT_RT kernel, but on a PREEMPT_RT kernel this breaks. The PREEMPT_RT-specific change of spinlock_t semantics does not allow to acquire p->lock because get_cpu_ptr() implicitly disables preemption. The following substitution works on both kernels:: ---8<--- Now we don't explicitly have this pattern because there isn't an obvious this_cpu_read() for example but it can accidentally happen for counting. __count_zid_vm_events -> __count_vm_events -> raw_cpu_add is an example although a harmless one. Any of the mod_page_state ones are more problematic though because we lock one PCP but potentially update the per-cpu pcp stats of another CPU of a different PCP that we have not locked and those counters must be accurate. It *might* still be safe but it's subtle, it could be easily accidentally broken in the future and it would be hard to detect because it would be very slow corruption of VM counters like NR_FREE_PAGES that must be accurate.
Hi Mel, On Fri, 2022-03-25 at 10:48 +0000, Mel Gorman wrote: > > [1] It follows this pattern: > > > > struct per_cpu_pages *pcp; > > > > pcp = raw_cpu_ptr(page_zone(page)->per_cpu_pageset); > > // <- Migration here is OK: spin_lock protects vs eventual pcplist > > // access from local CPU as long as all list access happens through the > > // pcp pointer. > > spin_lock(&pcp->lock); > > do_stuff_with_pcp_lists(pcp); > > spin_unlock(&pcp->lock); > > > > And this was the part I am concerned with. We are accessing a PCP > structure that is not necessarily the one belonging to the CPU we > are currently running on. This type of pattern is warned about in > Documentation/locking/locktypes.rst > > ---8<--- > A typical scenario is protection of per-CPU variables in thread context:: > > struct foo *p = get_cpu_ptr(&var1); > > spin_lock(&p->lock); > p->count += this_cpu_read(var2); > > This is correct code on a non-PREEMPT_RT kernel, but on a PREEMPT_RT kernel > this breaks. The PREEMPT_RT-specific change of spinlock_t semantics does > not allow to acquire p->lock because get_cpu_ptr() implicitly disables > preemption. The following substitution works on both kernels:: > ---8<--- > > Now we don't explicitly have this pattern because there isn't an > obvious this_cpu_read() for example but it can accidentally happen for > counting. __count_zid_vm_events -> __count_vm_events -> raw_cpu_add is > an example although a harmless one. > > Any of the mod_page_state ones are more problematic though because we > lock one PCP but potentially update the per-cpu pcp stats of another CPU > of a different PCP that we have not locked and those counters must be > accurate. But IIUC vmstats don't track pcplist usage (i.e. adding a page into the local pcplist doesn't affect the count at all). It is only when interacting with the buddy allocator that they get updated. It makes sense for the CPU that adds/removes pages from the allocator to do the stat update, regardless of the page's journey. > It *might* still be safe but it's subtle, it could be easily accidentally > broken in the future and it would be hard to detect because it would be > very slow corruption of VM counters like NR_FREE_PAGES that must be > accurate. What does accurate mean here? vmstat consumers don't get accurate data, only snapshots. And as I comment above you can't infer information about pcplist usage from these stats. So, I see no real need for CPU locality when updating them (which we're still retaining nonetheless, as per my comment above), the only thing that is really needed is atomicity, achieved by disabling IRQs (and preemption on RT). And this, even with your solution, is achieved through the struct zone's spin_lock (plus a preempt_disable() in RT). All in all, my point is that none of the stats are affected by the change, nor have a dependency with the pcplists handling. And if we ever have the need to pin vmstat updates to pcplist usage they should share the same pcp structure. That said, I'm happy with either solution as long as we get remote pcplist draining. So if still unconvinced, let me know how can I help. I have access to all sorts of machines to validate perf results, time to review, or even to move the series forward. Thanks!
On Mon, Mar 28, 2022 at 03:51:43PM +0200, Nicolas Saenz Julienne wrote: > > Now we don't explicitly have this pattern because there isn't an > > obvious this_cpu_read() for example but it can accidentally happen for > > counting. __count_zid_vm_events -> __count_vm_events -> raw_cpu_add is > > an example although a harmless one. > > > > Any of the mod_page_state ones are more problematic though because we > > lock one PCP but potentially update the per-cpu pcp stats of another CPU > > of a different PCP that we have not locked and those counters must be > > accurate. > > But IIUC vmstats don't track pcplist usage (i.e. adding a page into the local > pcplist doesn't affect the count at all). It is only when interacting with the > buddy allocator that they get updated. It makes sense for the CPU that > adds/removes pages from the allocator to do the stat update, regardless of the > page's journey. > It probably doesn't, I didn't audit it. As I said, it's subtle which is why I'm wary of relying on accidental safety of getting a per-cpu pointer that may not be stable. Even if it was ok *now*, I would worry that it would break in the future. There already has been cases where patches tried to move vmstats outside the appropriate locking accidentally. > > It *might* still be safe but it's subtle, it could be easily accidentally > > broken in the future and it would be hard to detect because it would be > > very slow corruption of VM counters like NR_FREE_PAGES that must be > > accurate. > > What does accurate mean here? vmstat consumers don't get accurate data, only > snapshots. They are accurate in that they have "Eventual Consistency". zone_page_state_snapshot exists to get a more accurate count but there is always some drift but it still is accurate eventually. There is a clear distinction between VM counters which can be inaccurate they are just to assist debugging and vmstats like NR_FREE_PAGES that the kernel uses to make decisions. It potentially gets very problematic if a per-cpu pointer acquired from one zone gets migrated to another zone and the wrong vmstat is updated. It *might* still be ok, I haven't audited it but if there is a possible that two CPUs can be doing a RMW on one per-cpu vmstat structure, it will corrupt and it'll be difficult to detect. > And as I comment above you can't infer information about pcplist > usage from these stats. So, I see no real need for CPU locality when updating > them (which we're still retaining nonetheless, as per my comment above), the > only thing that is really needed is atomicity, achieved by disabling IRQs (and > preemption on RT). And this, even with your solution, is achieved through the > struct zone's spin_lock (plus a preempt_disable() in RT). > Yes, but under the series I had, I was using local_lock to stabilise what CPU is being used before acquiring the per-cpu pointer. Strictly speaking, it doesn't need a local_lock but the local_lock is clearer in terms of what is being protected and it works with PROVE_LOCKING which already caught a problematic softirq interaction for me when developing the series. > All in all, my point is that none of the stats are affected by the change, nor > have a dependency with the pcplists handling. And if we ever have the need to > pin vmstat updates to pcplist usage they should share the same pcp structure. > That said, I'm happy with either solution as long as we get remote pcplist > draining. So if still unconvinced, let me know how can I help. I have access to > all sorts of machines to validate perf results, time to review, or even to move > the series forward. > I also want the remote draining for PREEMPT_RT to avoid interference of isolated CPUs due to workqueue activity but whatever the solution, I would be happier if the per-cpu lock is acquired with the CPU stablised and covers the scope of any vmstat delta updates stored in the per-cpu structure. The earliest I will be rebasing my series is 5.18-rc1 as I see limited value in basing it on 5.17 aiming for a 5.19 merge window.
Hi Mel, On Thu, 2022-03-03 at 11:45 +0000, Mel Gorman wrote: > On Tue, Feb 08, 2022 at 11:07:48AM +0100, Nicolas Saenz Julienne wrote: > > This series replaces mm/page_alloc's per-cpu page lists drain mechanism with > > one that allows accessing the lists remotely. Currently, only a local CPU is > > permitted to change its per-cpu lists, and it's expected to do so, on-demand, > > whenever a process demands it by means of queueing a drain task on the local > > CPU. This causes problems for NOHZ_FULL CPUs and real-time systems that can't > > take any sort of interruption and to some lesser extent inconveniences idle and > > virtualised systems. > > > > I know this has been sitting here for a long while. Last few weeks have > not been fun. > > > Note that this is not the first attempt at fixing this per-cpu page lists: > > - The first attempt[1] tried to conditionally change the pagesets locking > > scheme based the NOHZ_FULL config. It was deemed hard to maintain as the > > NOHZ_FULL code path would be rarely tested. Also, this only solves the issue > > for NOHZ_FULL setups, which isn't ideal. > > - The second[2] unanimously switched the local_locks to per-cpu spinlocks. The > > performance degradation was too big. > > > > For unrelated reasons I looked at using llist to avoid locks entirely. It > turns out it's not possible and needs a lock. We know "local_locks to > per-cpu spinlocks" took a large penalty so I considered alternatives on > how a lock could be used. I found it's possible to both remote drain > the lists and avoid the disable/enable of IRQs entirely as long as a > preempting IRQ is willing to take the zone lock instead (should be very > rare). The IRQ part is a bit hairy though as softirqs are also a problem > and preempt-rt needs different rules and the llist has to sort PCP > refills which might be a loss in total. However, the remote draining may > still be interesting. The full series is at > https://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git/ mm-pcpllist-v1r2 > > It's still waiting on tests to complete and not all the changelogs are > complete which is why it's not posted. > > This is a comparison of vanilla vs "local_locks to per-cpu spinlocks" > versus the git series up to "mm/page_alloc: Remotely drain per-cpu lists" > for the page faulting microbench I originally complained about. The test > machine is a 2-socket CascadeLake machine. > > pft timings > 5.17.0-rc5 5.17.0-rc5 5.17.0-rc5 > vanilla mm-remotedrain-v2r1 mm-pcpdrain-v1r1 > Amean elapsed-1 32.54 ( 0.00%) 33.08 * -1.66%* 32.82 * -0.86%* > Amean elapsed-4 8.66 ( 0.00%) 9.24 * -6.72%* 8.69 * -0.38%* > Amean elapsed-7 5.02 ( 0.00%) 5.43 * -8.16%* 5.05 * -0.55%* > Amean elapsed-12 3.07 ( 0.00%) 3.38 * -10.00%* 3.09 * -0.72%* > Amean elapsed-21 2.36 ( 0.00%) 2.38 * -0.89%* 2.19 * 7.39%* > Amean elapsed-30 1.75 ( 0.00%) 1.87 * -6.50%* 1.62 * 7.59%* > Amean elapsed-48 1.71 ( 0.00%) 2.00 * -17.32%* 1.71 ( -0.08%) > Amean elapsed-79 1.56 ( 0.00%) 1.62 * -3.84%* 1.56 ( -0.02%) > Amean elapsed-80 1.57 ( 0.00%) 1.65 * -5.31%* 1.57 ( -0.04%) > > Note the local_lock conversion took 1 1-17% penalty while the git tree > takes a negligile penalty while still allowing remote drains. It might > have some potential while being less complex than the RCU approach. I've been made aware of a problem with the spin_trylock() approach. It doesn't work for UP since in that context spin_lock() is a NOOP (well, it only disables preemption). So nothing prevents a race with an IRQ. It's kinda funny, breakages generally happen the other way around. Regards,
On Wed, Mar 30, 2022 at 01:29:04PM +0200, Nicolas Saenz Julienne wrote: > Hi Mel, > > On Thu, 2022-03-03 at 11:45 +0000, Mel Gorman wrote: > > On Tue, Feb 08, 2022 at 11:07:48AM +0100, Nicolas Saenz Julienne wrote: > > > This series replaces mm/page_alloc's per-cpu page lists drain mechanism with > > > one that allows accessing the lists remotely. Currently, only a local CPU is > > > permitted to change its per-cpu lists, and it's expected to do so, on-demand, > > > whenever a process demands it by means of queueing a drain task on the local > > > CPU. This causes problems for NOHZ_FULL CPUs and real-time systems that can't > > > take any sort of interruption and to some lesser extent inconveniences idle and > > > virtualised systems. > > > > > > > I know this has been sitting here for a long while. Last few weeks have > > not been fun. > > > > > Note that this is not the first attempt at fixing this per-cpu page lists: > > > - The first attempt[1] tried to conditionally change the pagesets locking > > > scheme based the NOHZ_FULL config. It was deemed hard to maintain as the > > > NOHZ_FULL code path would be rarely tested. Also, this only solves the issue > > > for NOHZ_FULL setups, which isn't ideal. > > > - The second[2] unanimously switched the local_locks to per-cpu spinlocks. The > > > performance degradation was too big. > > > > > > > For unrelated reasons I looked at using llist to avoid locks entirely. It > > turns out it's not possible and needs a lock. We know "local_locks to > > per-cpu spinlocks" took a large penalty so I considered alternatives on > > how a lock could be used. I found it's possible to both remote drain > > the lists and avoid the disable/enable of IRQs entirely as long as a > > preempting IRQ is willing to take the zone lock instead (should be very > > rare). The IRQ part is a bit hairy though as softirqs are also a problem > > and preempt-rt needs different rules and the llist has to sort PCP > > refills which might be a loss in total. However, the remote draining may > > still be interesting. The full series is at > > https://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git/ mm-pcpllist-v1r2 > > > > It's still waiting on tests to complete and not all the changelogs are > > complete which is why it's not posted. > > > > This is a comparison of vanilla vs "local_locks to per-cpu spinlocks" > > versus the git series up to "mm/page_alloc: Remotely drain per-cpu lists" > > for the page faulting microbench I originally complained about. The test > > machine is a 2-socket CascadeLake machine. > > > > pft timings > > 5.17.0-rc5 5.17.0-rc5 5.17.0-rc5 > > vanilla mm-remotedrain-v2r1 mm-pcpdrain-v1r1 > > Amean elapsed-1 32.54 ( 0.00%) 33.08 * -1.66%* 32.82 * -0.86%* > > Amean elapsed-4 8.66 ( 0.00%) 9.24 * -6.72%* 8.69 * -0.38%* > > Amean elapsed-7 5.02 ( 0.00%) 5.43 * -8.16%* 5.05 * -0.55%* > > Amean elapsed-12 3.07 ( 0.00%) 3.38 * -10.00%* 3.09 * -0.72%* > > Amean elapsed-21 2.36 ( 0.00%) 2.38 * -0.89%* 2.19 * 7.39%* > > Amean elapsed-30 1.75 ( 0.00%) 1.87 * -6.50%* 1.62 * 7.59%* > > Amean elapsed-48 1.71 ( 0.00%) 2.00 * -17.32%* 1.71 ( -0.08%) > > Amean elapsed-79 1.56 ( 0.00%) 1.62 * -3.84%* 1.56 ( -0.02%) > > Amean elapsed-80 1.57 ( 0.00%) 1.65 * -5.31%* 1.57 ( -0.04%) > > > > Note the local_lock conversion took 1 1-17% penalty while the git tree > > takes a negligile penalty while still allowing remote drains. It might > > have some potential while being less complex than the RCU approach. > > I've been made aware of a problem with the spin_trylock() approach. It doesn't > work for UP since in that context spin_lock() is a NOOP (well, it only disables > preemption). So nothing prevents a race with an IRQ. > I didn't think of UP being a problem. I'm offline shortly until early next week but superficially the spin_[try]lock for PCP would need a pcp_lock and pcp_trylock helpers. On SMP, it would be the equivalent lock. On UP, pcp_lock would map to spin_lock but pcp_trylock would likely need to map to spin_lock_irqsave. It means that UP would always disable IRQs but that would be no worse than the current allocator.
This series replaces mm/page_alloc's per-cpu page lists drain mechanism with one that allows accessing the lists remotely. Currently, only a local CPU is permitted to change its per-cpu lists, and it's expected to do so, on-demand, whenever a process demands it by means of queueing a drain task on the local CPU. This causes problems for NOHZ_FULL CPUs and real-time systems that can't take any sort of interruption and to some lesser extent inconveniences idle and virtualised systems. The new algorithm will atomically switch the pointer to the per-cpu page lists and use RCU to make sure it's not being concurrently used before draining the lists. And its main benefit of is that it fixes the issue for good, avoiding the need for configuration based heuristics or having to modify applications (i.e. using the isolation prctrl being worked by Marcello Tosatti ATM). All this with minimal performance implications: a page allocation microbenchmark was run on multiple systems and architectures generally showing no performance differences, only the more extreme cases showed a 1-3% degradation. See data below. Needless to say that I'd appreciate if someone could validate my values independently. The approach has been stress-tested: I forced 100 drains/s while running mmtests' pft in a loop for a full day on multiple machines and archs (arm64, x86_64, ppc64le). Note that this is not the first attempt at fixing this per-cpu page lists: - The first attempt[1] tried to conditionally change the pagesets locking scheme based the NOHZ_FULL config. It was deemed hard to maintain as the NOHZ_FULL code path would be rarely tested. Also, this only solves the issue for NOHZ_FULL setups, which isn't ideal. - The second[2] unanimously switched the local_locks to per-cpu spinlocks. The performance degradation was too big. Previous RFC: https://lkml.org/lkml/2021/10/8/793 Thanks! [1] https://lkml.org/lkml/2021/9/21/599 [2] https://lkml.org/lkml/2021/11/3/644 --- Changes since RFC: - Get rid of aesthetic changes that affected performance - Add more documentation - Add better commit messages - Pass sparse tests - Verify this_cpu_*() usage - Performance measurements Nicolas Saenz Julienne (2): mm/page_alloc: Access lists in 'struct per_cpu_pages' indirectly mm/page_alloc: Add remote draining support to per-cpu lists include/linux/mmzone.h | 28 +++++- mm/page_alloc.c | 212 ++++++++++++++++++++++++++--------------- mm/vmstat.c | 6 +- 3 files changed, 162 insertions(+), 84 deletions(-) -------------------------Performance results----------------------------- I'm focusing on mmtests' Page Fault Test (pft), as it's page allocator intensive. - AMD Daytona Reference System, 2 sockets, AMD EPYC 7742, Zen 2, 64-Core, 4 NUMA nodes, x86_64 pft timings: vanilla rcu Amean system-1 58.52 ( 0.00%) 58.92 * -0.68%* Amean system-4 61.00 ( 0.00%) 61.41 * -0.67%* Amean system-7 61.55 ( 0.00%) 61.74 * -0.30%* Amean system-12 64.91 ( 0.00%) 64.94 * -0.05%* Amean system-21 98.80 ( 0.00%) 99.92 * -1.13%* Amean system-30 147.68 ( 0.00%) 145.83 * 1.25%* Amean system-48 237.04 ( 0.00%) 241.29 * -1.79%* Amean system-79 286.61 ( 0.00%) 283.72 * 1.01%* Amean system-110 303.40 ( 0.00%) 299.91 * 1.15%* Amean system-128 345.07 ( 0.00%) 342.10 * 0.86%* Amean elapsed-1 61.21 ( 0.00%) 61.65 * -0.71%* Amean elapsed-4 15.94 ( 0.00%) 16.05 * -0.69%* Amean elapsed-7 9.24 ( 0.00%) 9.28 * -0.47%* Amean elapsed-12 5.70 ( 0.00%) 5.70 * -0.07%* Amean elapsed-21 5.11 ( 0.00%) 5.06 * 1.13%* Amean elapsed-30 5.28 ( 0.00%) 5.14 * 2.73%* Amean elapsed-48 5.28 ( 0.00%) 5.24 * 0.74%* Amean elapsed-79 4.41 ( 0.00%) 4.31 * 2.17%* Amean elapsed-110 3.45 ( 0.00%) 3.44 * 0.40%* Amean elapsed-128 2.75 ( 0.00%) 2.75 * -0.28%* - AMD Speedway Reference System, 2 sockets, AMD EPYC 7601, Zen 1, 64-core, 8 NUMA nodes, x86_64. Lots of variance between tests on this platform. It'll easily swing -+5% on each result. pft timings: vanilla rcu Amean system-1 69.20 ( 0.00%) 66.21 * 4.32%* Amean system-4 70.79 ( 0.00%) 69.01 * 2.52%* Amean system-7 71.34 ( 0.00%) 69.16 * 3.05%* Amean system-12 74.00 ( 0.00%) 72.74 * 1.70%* Amean system-21 86.01 ( 0.00%) 85.70 * 0.36%* Amean system-30 89.21 ( 0.00%) 89.93 * -0.80%* Amean system-48 92.39 ( 0.00%) 92.43 * -0.04%* Amean system-79 120.19 ( 0.00%) 121.30 * -0.92%* Amean system-110 172.79 ( 0.00%) 179.37 * -3.81%* Amean system-128 201.70 ( 0.00%) 212.57 * -5.39%* Amean elapsed-1 72.23 ( 0.00%) 69.29 * 4.08%* Amean elapsed-4 18.69 ( 0.00%) 18.28 * 2.20%* Amean elapsed-7 10.80 ( 0.00%) 10.54 * 2.41%* Amean elapsed-12 6.62 ( 0.00%) 6.53 * 1.30%* Amean elapsed-21 4.68 ( 0.00%) 4.69 * -0.14%* Amean elapsed-30 3.44 ( 0.00%) 3.50 * -1.66%* Amean elapsed-48 2.40 ( 0.00%) 2.42 * -1.00%* Amean elapsed-79 2.05 ( 0.00%) 2.09 * -1.90%* Amean elapsed-110 1.83 ( 0.00%) 1.91 * -4.60%* Amean elapsed-128 1.75 ( 0.00%) 1.85 * -5.99%* - IBM 9006-22C system, 2 sockets, POWER9, 64-Core, 1 NUMA node per cpu, pppc64le. pft timings: vanilla rcu Amean system-1 1.82 ( 0.00%) 1.85 * -1.43%* Amean system-4 2.18 ( 0.00%) 2.22 * -2.02%* Amean system-7 3.27 ( 0.00%) 3.28 * -0.15%* Amean system-12 5.22 ( 0.00%) 5.20 * 0.26%* Amean system-21 10.10 ( 0.00%) 10.20 * -1.00%* Amean system-30 15.00 ( 0.00%) 14.52 * 3.20%* Amean system-48 26.41 ( 0.00%) 25.96 * 1.71%* Amean system-79 29.35 ( 0.00%) 29.70 * -1.21%* Amean system-110 24.01 ( 0.00%) 23.40 * 2.54%* Amean system-128 24.57 ( 0.00%) 25.32 * -3.06%* Amean elapsed-1 1.85 ( 0.00%) 1.87 * -1.28%* Amean elapsed-4 0.56 ( 0.00%) 0.57 * -1.72%* Amean elapsed-7 0.51 ( 0.00%) 0.50 * 0.07%* Amean elapsed-12 0.51 ( 0.00%) 0.51 * 0.06%* Amean elapsed-21 0.54 ( 0.00%) 0.54 * 0.06%* Amean elapsed-30 0.54 ( 0.00%) 0.53 * 2.22%* Amean elapsed-48 0.58 ( 0.00%) 0.57 * 1.73%* Amean elapsed-79 0.49 ( 0.00%) 0.48 * 0.89%* Amean elapsed-110 0.37 ( 0.00%) 0.37 * -1.08%* Amean elapsed-128 0.33 ( 0.00%) 0.33 * 0.00%* - Ampere MtSnow, 1 socket, Neoverse-N1, 80-Cores, 1 NUMA node, arm64. pft timings: vanilla rcu Amean system-1 11.92 ( 0.00%) 11.99 * -0.61%* Amean system-4 13.13 ( 0.00%) 13.09 * 0.31%* Amean system-7 13.91 ( 0.00%) 13.94 * -0.20%* Amean system-12 15.77 ( 0.00%) 15.69 * 0.48%* Amean system-21 21.32 ( 0.00%) 21.42 * -0.46%* Amean system-30 28.58 ( 0.00%) 29.12 * -1.90%* Amean system-48 47.41 ( 0.00%) 46.91 * 1.04%* Amean system-79 76.76 ( 0.00%) 77.16 * -0.52%* Amean system-80 77.98 ( 0.00%) 78.23 * -0.32%* Amean elapsed-1 12.46 ( 0.00%) 12.53 * -0.58%* Amean elapsed-4 3.47 ( 0.00%) 3.46 * 0.34%* Amean elapsed-7 2.18 ( 0.00%) 2.21 * -1.58%* Amean elapsed-12 1.41 ( 0.00%) 1.42 * -0.80%* Amean elapsed-21 1.09 ( 0.00%) 1.12 * -2.60%* Amean elapsed-30 0.98 ( 0.00%) 1.01 * -3.08%* Amean elapsed-48 1.08 ( 0.00%) 1.10 * -1.78%* Amean elapsed-79 1.32 ( 0.00%) 1.28 * 2.71%* Amean elapsed-80 1.32 ( 0.00%) 1.28 * 3.23%* - Dell R430, 2 sockets, Intel Xeon E5-2640 v3, Sandy Bridge, 16-Cores, 2 NUMA nodes, x86_64. pft timings: vanilla rcu Amean system-1 11.10 ( 0.00%) 11.07 * 0.24%* Amean system-3 11.14 ( 0.00%) 11.10 * 0.34%* Amean system-5 11.18 ( 0.00%) 11.13 * 0.47%* Amean system-7 11.21 ( 0.00%) 11.17 * 0.38%* Amean system-12 11.28 ( 0.00%) 11.28 ( -0.03%) Amean system-18 13.24 ( 0.00%) 13.25 * -0.11%* Amean system-24 17.12 ( 0.00%) 17.14 ( -0.13%) Amean system-30 21.10 ( 0.00%) 21.23 * -0.60%* Amean system-32 22.31 ( 0.00%) 22.47 * -0.71%* Amean elapsed-1 11.76 ( 0.00%) 11.73 * 0.29%* Amean elapsed-3 3.93 ( 0.00%) 3.93 * 0.17%* Amean elapsed-5 2.39 ( 0.00%) 2.37 * 0.74%* Amean elapsed-7 1.72 ( 0.00%) 1.71 * 0.81%* Amean elapsed-12 1.02 ( 0.00%) 1.03 ( -0.42%) Amean elapsed-18 1.13 ( 0.00%) 1.14 ( -0.18%) Amean elapsed-24 0.87 ( 0.00%) 0.88 * -0.65%* Amean elapsed-30 0.77 ( 0.00%) 0.78 * -0.86%* Amean elapsed-32 0.74 ( 0.00%) 0.74 ( 0.00%) - HPE Apollo 70, 2 sockets, Cavium ThunderX2, 128-Cores, 2 NUMA nodes, arm64. NOTE: The test here only goes up to 128 for some reason, although there are 256 CPUs. Maybe a mmtests issue? I didn't investigate. pft timings: vanilla rcu Amean system-1 4.42 ( 0.00%) 4.36 * 1.29%* Amean system-4 4.56 ( 0.00%) 4.51 * 1.05%* Amean system-7 4.63 ( 0.00%) 4.65 * -0.42%* Amean system-12 5.96 ( 0.00%) 6.02 * -1.00%* Amean system-21 10.97 ( 0.00%) 11.01 * -0.32%* Amean system-30 16.01 ( 0.00%) 16.04 * -0.19%* Amean system-48 26.81 ( 0.00%) 26.78 * 0.09%* Amean system-79 30.80 ( 0.00%) 30.85 * -0.16%* Amean system-110 31.87 ( 0.00%) 31.93 * -0.19%* Amean system-128 36.27 ( 0.00%) 36.31 * -0.10%* Amean elapsed-1 4.88 ( 0.00%) 4.85 * 0.60%* Amean elapsed-4 1.27 ( 0.00%) 1.26 * 1.00%* Amean elapsed-7 0.73 ( 0.00%) 0.74 * -0.46%* Amean elapsed-12 0.55 ( 0.00%) 0.55 * 1.09%* Amean elapsed-21 0.59 ( 0.00%) 0.60 * -0.96%* Amean elapsed-30 0.60 ( 0.00%) 0.60 * 0.28%* Amean elapsed-48 0.60 ( 0.00%) 0.60 * 0.44%* Amean elapsed-79 0.49 ( 0.00%) 0.49 * -0.07%* Amean elapsed-110 0.36 ( 0.00%) 0.36 * 0.28%* Amean elapsed-128 0.31 ( 0.00%) 0.31 * -0.43%* - Raspberry Pi 4, 1 socket, bcm2711, Cortex-A72, 4-Cores, 1 NUMA node, arm64. pft timings: vanilla rcu Amean system-1 0.67 ( 0.00%) 0.67 * -1.25%* Amean system-3 1.30 ( 0.00%) 1.29 * 0.62%* Amean system-4 1.61 ( 0.00%) 1.59 * 0.95%* Amean elapsed-1 0.71 ( 0.00%) 0.72 * -1.17%* Amean elapsed-3 0.45 ( 0.00%) 0.45 * 0.88%* Amean elapsed-4 0.42 ( 0.00%) 0.42 * 1.19%*