Re: [RESEND PATCH v9 2/2] arm64: support batched/deferred tlb shootdown during page reclamation/migration
On 2023/7/5 16:43, Barry Song wrote: > On Tue, Jul 4, 2023 at 10:36 PM Yicong Yang wrote: >> >> On 2023/6/30 1:26, Catalin Marinas wrote: >>> On Thu, Jun 29, 2023 at 05:31:36PM +0100, Catalin Marinas wrote: On Thu, May 18, 2023 at 02:59:34PM +0800, Yicong Yang wrote: > From: Barry Song > > on x86, batched and deferred tlb shootdown has lead to 90% > performance increase on tlb shootdown. on arm64, HW can do > tlb shootdown without software IPI. But sync tlbi is still > quite expensive. [...] > .../features/vm/TLB/arch-support.txt | 2 +- > arch/arm64/Kconfig| 1 + > arch/arm64/include/asm/tlbbatch.h | 12 > arch/arm64/include/asm/tlbflush.h | 33 - > arch/arm64/mm/flush.c | 69 +++ > arch/x86/include/asm/tlbflush.h | 5 +- > include/linux/mm_types_task.h | 4 +- > mm/rmap.c | 12 ++-- First of all, this patch needs to be split in some preparatory patches introducing/renaming functions with no functional change for x86. Once done, you can add the arm64-only changes. >> >> got it. will try to split this patch as suggested. >> Now, on the implementation, I had some comments on v7 but we didn't get to a conclusion and the thread eventually died: https://lore.kernel.org/linux-mm/y7ctoj5mwd1zb...@arm.com/ I know I said a command line argument is better than Kconfig or some random number of CPUs heuristics but it would be even better if we don't bother with any, just make this always on. >> >> ok, will make this always on. >> Barry had some comments around mprotect() being racy and that's why we have flush_tlb_batched_pending() but I don't think it's needed (or, for arm64, it can be a DSB since this patch issues the TLBIs but without the DVM Sync). So we need to clarify this (see Barry's last email on the above thread) and before attempting new versions of this patchset. With flush_tlb_batched_pending() removed (or DSB), I have a suspicion such implementation would be faster on any SoC irrespective of the number of CPUs. >>> >>> I think I got the need for flush_tlb_batched_pending(). If >>> try_to_unmap() marks the pte !present and we have a pending TLBI, >>> change_pte_range() will skip the TLB maintenance altogether since it did >>> not change the pte. So we could be left with stale TLB entries after >>> mprotect() before TTU does the batch flushing. >>> > > Good catch. > This could be also true for MADV_DONTNEED. after try_to_unmap, we run > MADV_DONTNEED on this area, as pte is not present, we don't do anything > on this PTE in zap_pte_range afterwards. > >>> We can have an arch-specific flush_tlb_batched_pending() that can be a >>> DSB only on arm64 and a full mm flush on x86. >>> >> >> We need to do a flush/dsb in flush_tlb_batched_pending() only in a race >> condition so we first check whether there's a pended batched flush and >> if so do the tlb flush. The pending checking is common and the differences >> among the archs is how to flush the TLB here within the >> flush_tlb_batched_pending(), >> on arm64 it should only be a dsb. >> >> As we only needs to maintain the TLBs already pended in batched flush, >> does it make sense to only handle those TLBs in flush_tlb_batched_pending()? >> Then we can use the arch_tlbbatch_flush() rather than flush_tlb_mm() in >> flush_tlb_batched_pending() and no arch specific function needed. > > as we have issued no-sync tlbi on those pending addresses , that means > our hardware > has already "recorded" what should be flushed in the specific mm. so > DSB only will flush > them correctly. right? > yes it's right. I was just thought something like below. arch_tlbbatch_flush() will only be a dsb on arm64 so this will match what Catalin wants. But as you told that this maybe incorrect on x86 so we'd better have arch specific implementation for flush_tlb_batched_pending() as suggested. diff --git a/mm/rmap.c b/mm/rmap.c index 9699c6011b0e..afa3571503a0 100644 --- a/mm/rmap.c +++ b/mm/rmap.c @@ -717,7 +717,7 @@ void flush_tlb_batched_pending(struct mm_struct *mm) int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT; if (pending != flushed) { - flush_tlb_mm(mm); + arch_tlbbatch_flush(¤t->tlb_ubc.arch); /* * If the new TLB flushing is pending during flushing, leave * mm->tlb_flush_batched as is, to avoid losing flushing.
Re: [RESEND PATCH v9 2/2] arm64: support batched/deferred tlb shootdown during page reclamation/migration
On Tue, Jul 4, 2023 at 10:36 PM Yicong Yang wrote: > > On 2023/6/30 1:26, Catalin Marinas wrote: > > On Thu, Jun 29, 2023 at 05:31:36PM +0100, Catalin Marinas wrote: > >> On Thu, May 18, 2023 at 02:59:34PM +0800, Yicong Yang wrote: > >>> From: Barry Song > >>> > >>> on x86, batched and deferred tlb shootdown has lead to 90% > >>> performance increase on tlb shootdown. on arm64, HW can do > >>> tlb shootdown without software IPI. But sync tlbi is still > >>> quite expensive. > >> [...] > >>> .../features/vm/TLB/arch-support.txt | 2 +- > >>> arch/arm64/Kconfig| 1 + > >>> arch/arm64/include/asm/tlbbatch.h | 12 > >>> arch/arm64/include/asm/tlbflush.h | 33 - > >>> arch/arm64/mm/flush.c | 69 +++ > >>> arch/x86/include/asm/tlbflush.h | 5 +- > >>> include/linux/mm_types_task.h | 4 +- > >>> mm/rmap.c | 12 ++-- > >> > >> First of all, this patch needs to be split in some preparatory patches > >> introducing/renaming functions with no functional change for x86. Once > >> done, you can add the arm64-only changes. > >> > > got it. will try to split this patch as suggested. > > >> Now, on the implementation, I had some comments on v7 but we didn't get > >> to a conclusion and the thread eventually died: > >> > >> https://lore.kernel.org/linux-mm/y7ctoj5mwd1zb...@arm.com/ > >> > >> I know I said a command line argument is better than Kconfig or some > >> random number of CPUs heuristics but it would be even better if we don't > >> bother with any, just make this always on. > > ok, will make this always on. > > >> Barry had some comments > >> around mprotect() being racy and that's why we have > >> flush_tlb_batched_pending() but I don't think it's needed (or, for > >> arm64, it can be a DSB since this patch issues the TLBIs but without the > >> DVM Sync). So we need to clarify this (see Barry's last email on the > >> above thread) and before attempting new versions of this patchset. With > >> flush_tlb_batched_pending() removed (or DSB), I have a suspicion such > >> implementation would be faster on any SoC irrespective of the number of > >> CPUs. > > > > I think I got the need for flush_tlb_batched_pending(). If > > try_to_unmap() marks the pte !present and we have a pending TLBI, > > change_pte_range() will skip the TLB maintenance altogether since it did > > not change the pte. So we could be left with stale TLB entries after > > mprotect() before TTU does the batch flushing. > > Good catch. This could be also true for MADV_DONTNEED. after try_to_unmap, we run MADV_DONTNEED on this area, as pte is not present, we don't do anything on this PTE in zap_pte_range afterwards. > > We can have an arch-specific flush_tlb_batched_pending() that can be a > > DSB only on arm64 and a full mm flush on x86. > > > > We need to do a flush/dsb in flush_tlb_batched_pending() only in a race > condition so we first check whether there's a pended batched flush and > if so do the tlb flush. The pending checking is common and the differences > among the archs is how to flush the TLB here within the > flush_tlb_batched_pending(), > on arm64 it should only be a dsb. > > As we only needs to maintain the TLBs already pended in batched flush, > does it make sense to only handle those TLBs in flush_tlb_batched_pending()? > Then we can use the arch_tlbbatch_flush() rather than flush_tlb_mm() in > flush_tlb_batched_pending() and no arch specific function needed. as we have issued no-sync tlbi on those pending addresses , that means our hardware has already "recorded" what should be flushed in the specific mm. so DSB only will flush them correctly. right? > > Thanks. > Barry
Re: [RESEND PATCH v9 2/2] arm64: support batched/deferred tlb shootdown during page reclamation/migration
On 2023/6/30 1:26, Catalin Marinas wrote: > On Thu, Jun 29, 2023 at 05:31:36PM +0100, Catalin Marinas wrote: >> On Thu, May 18, 2023 at 02:59:34PM +0800, Yicong Yang wrote: >>> From: Barry Song >>> >>> on x86, batched and deferred tlb shootdown has lead to 90% >>> performance increase on tlb shootdown. on arm64, HW can do >>> tlb shootdown without software IPI. But sync tlbi is still >>> quite expensive. >> [...] >>> .../features/vm/TLB/arch-support.txt | 2 +- >>> arch/arm64/Kconfig| 1 + >>> arch/arm64/include/asm/tlbbatch.h | 12 >>> arch/arm64/include/asm/tlbflush.h | 33 - >>> arch/arm64/mm/flush.c | 69 +++ >>> arch/x86/include/asm/tlbflush.h | 5 +- >>> include/linux/mm_types_task.h | 4 +- >>> mm/rmap.c | 12 ++-- >> >> First of all, this patch needs to be split in some preparatory patches >> introducing/renaming functions with no functional change for x86. Once >> done, you can add the arm64-only changes. >> got it. will try to split this patch as suggested. >> Now, on the implementation, I had some comments on v7 but we didn't get >> to a conclusion and the thread eventually died: >> >> https://lore.kernel.org/linux-mm/y7ctoj5mwd1zb...@arm.com/ >> >> I know I said a command line argument is better than Kconfig or some >> random number of CPUs heuristics but it would be even better if we don't >> bother with any, just make this always on. ok, will make this always on. >> Barry had some comments >> around mprotect() being racy and that's why we have >> flush_tlb_batched_pending() but I don't think it's needed (or, for >> arm64, it can be a DSB since this patch issues the TLBIs but without the >> DVM Sync). So we need to clarify this (see Barry's last email on the >> above thread) and before attempting new versions of this patchset. With >> flush_tlb_batched_pending() removed (or DSB), I have a suspicion such >> implementation would be faster on any SoC irrespective of the number of >> CPUs. > > I think I got the need for flush_tlb_batched_pending(). If > try_to_unmap() marks the pte !present and we have a pending TLBI, > change_pte_range() will skip the TLB maintenance altogether since it did > not change the pte. So we could be left with stale TLB entries after > mprotect() before TTU does the batch flushing. > > We can have an arch-specific flush_tlb_batched_pending() that can be a > DSB only on arm64 and a full mm flush on x86. > We need to do a flush/dsb in flush_tlb_batched_pending() only in a race condition so we first check whether there's a pended batched flush and if so do the tlb flush. The pending checking is common and the differences among the archs is how to flush the TLB here within the flush_tlb_batched_pending(), on arm64 it should only be a dsb. As we only needs to maintain the TLBs already pended in batched flush, does it make sense to only handle those TLBs in flush_tlb_batched_pending()? Then we can use the arch_tlbbatch_flush() rather than flush_tlb_mm() in flush_tlb_batched_pending() and no arch specific function needed. Thanks.
Re: [RESEND PATCH v9 2/2] arm64: support batched/deferred tlb shootdown during page reclamation/migration
On Thu, Jun 29, 2023 at 05:31:36PM +0100, Catalin Marinas wrote: > On Thu, May 18, 2023 at 02:59:34PM +0800, Yicong Yang wrote: > > From: Barry Song > > > > on x86, batched and deferred tlb shootdown has lead to 90% > > performance increase on tlb shootdown. on arm64, HW can do > > tlb shootdown without software IPI. But sync tlbi is still > > quite expensive. > [...] > > .../features/vm/TLB/arch-support.txt | 2 +- > > arch/arm64/Kconfig| 1 + > > arch/arm64/include/asm/tlbbatch.h | 12 > > arch/arm64/include/asm/tlbflush.h | 33 - > > arch/arm64/mm/flush.c | 69 +++ > > arch/x86/include/asm/tlbflush.h | 5 +- > > include/linux/mm_types_task.h | 4 +- > > mm/rmap.c | 12 ++-- > > First of all, this patch needs to be split in some preparatory patches > introducing/renaming functions with no functional change for x86. Once > done, you can add the arm64-only changes. > > Now, on the implementation, I had some comments on v7 but we didn't get > to a conclusion and the thread eventually died: > > https://lore.kernel.org/linux-mm/y7ctoj5mwd1zb...@arm.com/ > > I know I said a command line argument is better than Kconfig or some > random number of CPUs heuristics but it would be even better if we don't > bother with any, just make this always on. Barry had some comments > around mprotect() being racy and that's why we have > flush_tlb_batched_pending() but I don't think it's needed (or, for > arm64, it can be a DSB since this patch issues the TLBIs but without the > DVM Sync). So we need to clarify this (see Barry's last email on the > above thread) and before attempting new versions of this patchset. With > flush_tlb_batched_pending() removed (or DSB), I have a suspicion such > implementation would be faster on any SoC irrespective of the number of > CPUs. I think I got the need for flush_tlb_batched_pending(). If try_to_unmap() marks the pte !present and we have a pending TLBI, change_pte_range() will skip the TLB maintenance altogether since it did not change the pte. So we could be left with stale TLB entries after mprotect() before TTU does the batch flushing. We can have an arch-specific flush_tlb_batched_pending() that can be a DSB only on arm64 and a full mm flush on x86. -- Catalin
Re: [RESEND PATCH v9 2/2] arm64: support batched/deferred tlb shootdown during page reclamation/migration
On Thu, May 18, 2023 at 02:59:34PM +0800, Yicong Yang wrote: > From: Barry Song > > on x86, batched and deferred tlb shootdown has lead to 90% > performance increase on tlb shootdown. on arm64, HW can do > tlb shootdown without software IPI. But sync tlbi is still > quite expensive. [...] > .../features/vm/TLB/arch-support.txt | 2 +- > arch/arm64/Kconfig| 1 + > arch/arm64/include/asm/tlbbatch.h | 12 > arch/arm64/include/asm/tlbflush.h | 33 - > arch/arm64/mm/flush.c | 69 +++ > arch/x86/include/asm/tlbflush.h | 5 +- > include/linux/mm_types_task.h | 4 +- > mm/rmap.c | 12 ++-- First of all, this patch needs to be split in some preparatory patches introducing/renaming functions with no functional change for x86. Once done, you can add the arm64-only changes. Now, on the implementation, I had some comments on v7 but we didn't get to a conclusion and the thread eventually died: https://lore.kernel.org/linux-mm/y7ctoj5mwd1zb...@arm.com/ I know I said a command line argument is better than Kconfig or some random number of CPUs heuristics but it would be even better if we don't bother with any, just make this always on. Barry had some comments around mprotect() being racy and that's why we have flush_tlb_batched_pending() but I don't think it's needed (or, for arm64, it can be a DSB since this patch issues the TLBIs but without the DVM Sync). So we need to clarify this (see Barry's last email on the above thread) and before attempting new versions of this patchset. With flush_tlb_batched_pending() removed (or DSB), I have a suspicion such implementation would be faster on any SoC irrespective of the number of CPUs. -- Catalin
[RESEND PATCH v9 2/2] arm64: support batched/deferred tlb shootdown during page reclamation/migration
From: Barry Song on x86, batched and deferred tlb shootdown has lead to 90% performance increase on tlb shootdown. on arm64, HW can do tlb shootdown without software IPI. But sync tlbi is still quite expensive. Even running a simplest program which requires swapout can prove this is true, #include #include #include #include int main() { #define SIZE (1 * 1024 * 1024) volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, -1, 0); memset(p, 0x88, SIZE); for (int k = 0; k < 1; k++) { /* swap in */ for (int i = 0; i < SIZE; i += 4096) { (void)p[i]; } /* swap out */ madvise(p, SIZE, MADV_PAGEOUT); } } Perf result on snapdragon 888 with 8 cores by using zRAM as the swap block device. ~ # perf record taskset -c 4 ./a.out [ perf record: Woken up 10 times to write data ] [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ] ~ # perf report # To display the perf.data header info, please use --header/--header-only options. # To display the perf.data header info, please use --header/--header-only options. # # # Total Lost Samples: 0 # # Samples: 60K of event 'cycles' # Event count (approx.): 35706225414 # # Overhead Command Shared Object Symbol # ... . . # 21.07% a.out[kernel.kallsyms] [k] _raw_spin_unlock_irq 8.23% a.out[kernel.kallsyms] [k] _raw_spin_unlock_irqrestore 6.67% a.out[kernel.kallsyms] [k] filemap_map_pages 6.16% a.out[kernel.kallsyms] [k] __zram_bvec_write 5.36% a.out[kernel.kallsyms] [k] ptep_clear_flush 3.71% a.out[kernel.kallsyms] [k] _raw_spin_lock 3.49% a.out[kernel.kallsyms] [k] memset64 1.63% a.out[kernel.kallsyms] [k] clear_page 1.42% a.out[kernel.kallsyms] [k] _raw_spin_unlock 1.26% a.out[kernel.kallsyms] [k] mod_zone_state.llvm.8525150236079521930 1.23% a.out[kernel.kallsyms] [k] xas_load 1.15% a.out[kernel.kallsyms] [k] zram_slot_lock ptep_clear_flush() takes 5.36% CPU in the micro-benchmark swapping in/out a page mapped by only one process. If the page is mapped by multiple processes, typically, like more than 100 on a phone, the overhead would be much higher as we have to run tlb flush 100 times for one single page. Plus, tlb flush overhead will increase with the number of CPU cores due to the bad scalability of tlb shootdown in HW, so those ARM64 servers should expect much higher overhead. Further perf annonate shows 95% cpu time of ptep_clear_flush is actually used by the final dsb() to wait for the completion of tlb flush. This provides us a very good chance to leverage the existing batched tlb in kernel. The minimum modification is that we only send async tlbi in the first stage and we send dsb while we have to sync in the second stage. With the above simplest micro benchmark, collapsed time to finish the program decreases around 5%. Typical collapsed time w/o patch: ~ # time taskset -c 4 ./a.out 0.21user 14.34system 0:14.69elapsed w/ patch: ~ # time taskset -c 4 ./a.out 0.22user 13.45system 0:13.80elapsed Also, Yicong Yang added the following observation. Tested with benchmark in the commit on Kunpeng920 arm64 server, observed an improvement around 12.5% with command `time ./swap_bench`. w/o w/ real0m13.460s 0m11.771s user0m0.248s0m0.279s sys 0m12.039s 0m11.458s Originally it's noticed a 16.99% overhead of ptep_clear_flush() which has been eliminated by this patch: [root@localhost yang]# perf record -- ./swap_bench && perf report [...] 16.99% swap_bench [kernel.kallsyms] [k] ptep_clear_flush It is tested on 4,8,128 CPU platforms and shows to be beneficial on large systems but may not have improvement on small systems like on a 4 CPU platform. So make ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH depends on CONFIG_EXPERT for this stage and add a runtime tunable to allow to disable it according to the scenario. Also this patch improve the performance of page migration. Using pmbench and tries to migrate the pages of pmbench between node 0 and node 1 for 20 times, this patch decrease the time used more than 50% and saved the time used by ptep_clear_flush(). This patch extends arch_tlbbatch_add_mm() to take an address of the target page to support the feature on arm64. Also rename it to arch_tlbbatch_add_pending() to better match its function since we don't need to handle the mm on arm64 and add_mm is not proper. add_pending will make sense to both as on x86 we're pending the TLB flush operations while on arm64 w