On Fri, Dec 23, 2022 at 09:52:07PM +0800, Yin Fengwei wrote: > Kernel build regression with LLVM was reported here: > https://lore.kernel.org/all/Y1GCYXGtEVZbcv%2F5@dev-arch.thelio-3990X/ > with commit f35b5d7d676e ("mm: align larger anonymous mappings on THP > boundaries"). And the commit f35b5d7d676e was reverted. > > It turned out the regression is related with madvise(MADV_DONTNEED) > was used by ld.lld. But with none PMD_SIZE aligned parameter len. > trace-bpfcc captured: > 531607 531732 ld.lld do_madvise.part.0 start: 0x7feca9000000, len: 0x7fb000, behavior: 0x4 > 531607 531793 ld.lld do_madvise.part.0 start: 0x7fec86a00000, len: 0x7fb000, behavior: 0x4 > > If the underneath physical page is THP, the madvise(MADV_DONTNNED) can > trigger split_queue_lock contention raised significantly. perf showed > following data: > 14.85% 0.00% ld.lld [kernel.kallsyms] [k] > entry_SYSCALL_64_after_hwframe > 11.52% > entry_SYSCALL_64_after_hwframe > do_syscall_64 > __x64_sys_madvise > do_madvise.part.0 > zap_page_range > unmap_single_vma > unmap_page_range > page_remove_rmap > deferred_split_huge_page > __lock_text_start > native_queued_spin_lock_slowpath > > If THP can't be removed from rmap as whole THP, partial THP will be > removed from rmap by removing sub-pages from rmap. Even the THP > head page is added to deferred queue already, the split_queue_lock > will be acquired and check whether the THP head page is in the queue > already. Thus, the contention of split_queue_lock is raised. > > Before acquire split_queue_lock, check and bail out early if the THP > head page is in the queue already. The checking without holding > split_queue_lock could race with deferred_split_scan, but it doesn't > impact the correctness here. > > Test result of building kernel with ld.lld: > commit 7b5a0b664ebe (parent commit of f35b5d7d676e): > time -f "\t%E real,\t%U user,\t%S sys" make LD=ld.lld -skj96 allmodconfig all > 6:07.99 real, 26367.77 user, 5063.35 sys > > commit f35b5d7d676e: > time -f "\t%E real,\t%U user,\t%S sys" make LD=ld.lld -skj96 allmodconfig all > 7:22.15 real, 26235.03 user, 12504.55 sys > > commit f35b5d7d676e with the fixing patch: > time -f "\t%E real,\t%U user,\t%S sys" make LD=ld.lld -skj96 allmodconfig all > 6:08.49 real, 26520.15 user, 5047.91 sys > > Signed-off-by: Yin Fengwei <fengwei.yin@xxxxxxxxx> I cannot say whether or not this is a good idea or not but it does resolve the regression I reported: Benchmark 1: x86_64 allmodconfig (GCC + ld.lld) @ 1b929c02afd3 ("Linux 6.2-rc1") on 6.0.0-rc3-debug-00016-g7b5a0b664ebe Time (mean ± σ): 383.003 s ± 0.680 s [User: 34737.850 s, System: 7287.079 s] Range (min … max): 382.218 s … 383.413 s 3 runs Benchmark 1: x86_64 allmodconfig (GCC + ld.lld) @ 1b929c02afd3 ("Linux 6.2-rc1") on 6.0.0-rc3-debug-00017-gf35b5d7d676e Time (mean ± σ): 437.886 s ± 1.030 s [User: 35888.658 s, System: 14048.871 s] Range (min … max): 436.865 s … 438.924 s 3 runs Benchmark 1: x86_64 allmodconfig (GCC + ld.lld) @ 1b929c02afd3 ("Linux 6.2-rc1") on 6.0.0-rc3-debug-00017-gf35b5d7d676e-dirty Time (mean ± σ): 384.371 s ± 1.004 s [User: 35402.880 s, System: 6401.691 s] Range (min … max): 383.547 s … 385.489 s 3 runs Tested-by: Nathan Chancellor <nathan@xxxxxxxxxx> > --- > My first thought was to change the per node deferred queue to per cpu. > It's complicated and may be overkill. > > For the race without lock acquired, I didn't see obvious issue here. But I > could miss something here. Let me know if I did. Thanks. > > > mm/huge_memory.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/mm/huge_memory.c b/mm/huge_memory.c > index abe6cfd92ffa..7cde9f702e63 100644 > --- a/mm/huge_memory.c > +++ b/mm/huge_memory.c > @@ -2837,6 +2837,9 @@ void deferred_split_huge_page(struct page *page) > if (PageSwapCache(page)) > return; > > + if (!list_empty(page_deferred_list(page))) > + return; > + > spin_lock_irqsave(&ds_queue->split_queue_lock, flags); > if (list_empty(page_deferred_list(page))) { > count_vm_event(THP_DEFERRED_SPLIT_PAGE); > > base-commit: 8395ae05cb5a2e31d36106e8c85efa11cda849be > -- > 2.34.1 > > >
