Message ID | 20200203201745.29986-3-aarcange@redhat.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | arm64: tlb: skip tlbi broadcast for single threaded TLB flushes | expand |
Hi Andrea, Thanks for having a go at this. On Mon, Feb 03, 2020 at 03:17:45PM -0500, Andrea Arcangeli wrote: > With multiple NUMA nodes and multiple sockets, the tlbi broadcast > shall be delivered through the interconnects in turn increasing the > interconnect traffic and the latency of the tlbi broadcast instruction. There are better ways to handle this in hardware (snoop filters), so hopefully those vendors will eventually learn. > After the local TLB flush this however means the ASID context goes out > of sync in all CPUs except the local one. This can be tracked in the > mm_cpumask(mm): if the bit is set it means the asid context is stale > for that CPU. This results in an extra local ASID TLB flush only if a > single threaded process is migrated to a different CPU and only after a > TLB flush. No extra local TLB flush is needed for the common case of > single threaded processes context scheduling within the same CPU and for > multithreaded processes. Relying om mm_users is not sufficient AFAICT. Let's say on CPU0 you have a kernel thread running with the previous user pgd and ASID set in ttbr0_el1. The mm_users would still be 1 since only mm_count is incremented in context_switch(). If the user thread now runs on CPU1, a local tlbi would only invalidate the TLBs on CPU1. However, CPU0 may still walk (speculatively) the user page tables. An example where this matters is a group of small pages converted to a huge page. If CPU0 already has some TLB entries for small pages in the group but, not being aware of a TLBI for the ptes in the range, may read a block pmd entry (huge page) and we end up with a TLB conflict on CPU0 (CPU1 is fine since you do the local tlbi). There are other examples where this could go wrong as the hardware may keep intermediate pgtable entries in a walk cache. In the arm64 kernel we rely on something the architecture calls break-before-make for any page table updates and these need to be broadcast to other CPUs that may potentially have an entry in their TLB. It may be better if you used mm_cpumask to mark wherever an mm ever ran than relying on mm_users. > Skipping the tlbi instruction broadcasting is already implemented in > local_flush_tlb_all(), this patch only extends it to flush_tlb_mm(), > flush_tlb_range() and flush_tlb_page() too. > > Here's the result of 32 CPUs (ARMv8 Ampere) running mprotect at the same > time from 32 single threaded processes before the patch: > > Performance counter stats for './loop' (3 runs): > > 0 dummy > > 2.121353 +- 0.000387 seconds time elapsed ( +- 0.02% ) > > and with the patch applied: > > Performance counter stats for './loop' (3 runs): > > 0 dummy > > 0.1197750 +- 0.0000827 seconds time elapsed ( +- 0.07% ) That's a pretty artificial test and it is indeed improved by this patch. However, it would be nice to have some real-world scenarios where this matters.
Hello Catalin, On Mon, Feb 10, 2020 at 05:51:06PM +0000, Catalin Marinas wrote: > Relying om mm_users is not sufficient AFAICT. Let's say on CPU0 you have > a kernel thread running with the previous user pgd and ASID set in > ttbr0_el1. The mm_users would still be 1 since only mm_count is > incremented in context_switch(). If the user thread now runs on CPU1, a > local tlbi would only invalidate the TLBs on CPU1. However, CPU0 may > still walk (speculatively) the user page tables. > > An example where this matters is a group of small pages converted to a > huge page. If CPU0 already has some TLB entries for small pages in the > group but, not being aware of a TLBI for the ptes in the range, may read > a block pmd entry (huge page) and we end up with a TLB conflict on CPU0 > (CPU1 is fine since you do the local tlbi). > > There are other examples where this could go wrong as the hardware may > keep intermediate pgtable entries in a walk cache. In the arm64 kernel > we rely on something the architecture calls break-before-make for any > page table updates and these need to be broadcast to other CPUs that may > potentially have an entry in their TLB. > > It may be better if you used mm_cpumask to mark wherever an mm ever ran > than relying on mm_users. Agreed. If we can use mm_cpumask to track where the mm ever run, then if I'm not mistaken we could optimize also multithreaded processes in the same way: if only one thread is running frequently and the others are frequently sleeping, we could issue a single tlbi broadcast (modulo invalidates of small virtual ranges). In the meantime the below should be enough to address the concern you raised of the proof of concept RFC patch. I already experimented with mm_users == 1 earlier and it doesn't change the benchmark results for the "best case" below. (untested) diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h index 772bbc45b867..a2d53b301f22 100644 --- a/arch/arm64/include/asm/tlbflush.h +++ b/arch/arm64/include/asm/tlbflush.h @@ -169,7 +169,8 @@ static inline void flush_tlb_mm(struct mm_struct *mm) unsigned long asid = __TLBI_VADDR(0, ASID(mm)); /* avoid TLB-i broadcast to remote NUMA nodes if it's a local flush */ - if (current->mm == mm && atomic_read(&mm->mm_users) <= 1) { + if (current->mm == mm && atomic_read(&mm->mm_users) <= 1 && + (system_uses_ttbr0_pan() || atomic_read(&mm->mm_count) == 1)) { int cpu = get_cpu(); cpumask_setall(mm_cpumask(mm)); @@ -177,7 +178,9 @@ static inline void flush_tlb_mm(struct mm_struct *mm) smp_mb(); - if (atomic_read(&mm->mm_users) <= 1) { + if (atomic_read(&mm->mm_users) <= 1 && + (system_uses_ttbr0_pan() || + atomic_read(&mm->mm_count) == 1)) { dsb(nshst); __tlbi(aside1, asid); __tlbi_user(aside1, asid); @@ -212,7 +215,8 @@ static inline void flush_tlb_page(struct vm_area_struct *vma, unsigned long addr = __TLBI_VADDR(uaddr, ASID(mm)); /* avoid TLB-i broadcast to remote NUMA nodes if it's a local flush */ - if (current->mm == mm && atomic_read(&mm->mm_users) <= 1) { + if (current->mm == mm && atomic_read(&mm->mm_users) <= 1 && + (system_uses_ttbr0_pan() || atomic_read(&mm->mm_count) == 1)) { int cpu = get_cpu(); cpumask_setall(mm_cpumask(mm)); @@ -220,7 +224,9 @@ static inline void flush_tlb_page(struct vm_area_struct *vma, smp_mb(); - if (atomic_read(&mm->mm_users) <= 1) { + if (atomic_read(&mm->mm_users) <= 1 && + (system_uses_ttbr0_pan() || + atomic_read(&mm->mm_count) == 1)) { dsb(nshst); __tlbi(vale1, addr); __tlbi_user(vale1, addr); @@ -264,7 +270,8 @@ static inline void __flush_tlb_range(struct vm_area_struct *vma, end = __TLBI_VADDR(end, asid); /* avoid TLB-i broadcast to remote NUMA nodes if it's a local flush */ - if (current->mm == mm && atomic_read(&mm->mm_users) <= 1) { + if (current->mm == mm && atomic_read(&mm->mm_users) <= 1 && + (system_uses_ttbr0_pan() || atomic_read(&mm->mm_count) == 1)) { int cpu = get_cpu(); cpumask_setall(mm_cpumask(mm)); @@ -272,7 +279,9 @@ static inline void __flush_tlb_range(struct vm_area_struct *vma, smp_mb(); - if (atomic_read(&mm->mm_users) <= 1) { + if (atomic_read(&mm->mm_users) <= 1 && + (system_uses_ttbr0_pan() || + atomic_read(&mm->mm_count) == 1)) { dsb(nshst); for (addr = start; addr < end; addr += stride) { if (last_level) { > That's a pretty artificial test and it is indeed improved by this patch. > However, it would be nice to have some real-world scenarios where this > matters. I don't know exactly how much we should rely on the hardware to snoop the asid on NUMA. The hardware to fully optimize would need to implement a replicated mm_cpumask bitflag for each asid and every CPU would need to tell every other CPU which asid it is loading every time it is loading it. Exactly what x86 does with mm_cpumask in software. That is ideal, but is it an arch requirement to add the above in all implementations? The case I measured has a single socket so it's even simpler because it could be optimized all in-core. Even with a single socket I'm not sure what's going wrong in the chip: it felt like it's the engine that does the broadcast that runs serially system wide and then all CPUs have to wait on it. Still your question if it'll make a difference in practice is a good one and I don't have a sure answer yet. I suppose before doing more benchmarking it's better to make a new version of this that uses mm_cpumask to track where the asid was ever loaded as you suggested, so that it will also optimize away tlbi broadcaasts from multithreaded processes where only one thread is running frequently? Thanks! Andrea
Hi Andrea, On Mon, Feb 10, 2020 at 03:14:11PM -0500, Andrea Arcangeli wrote: > On Mon, Feb 10, 2020 at 05:51:06PM +0000, Catalin Marinas wrote: > > It may be better if you used mm_cpumask to mark wherever an mm ever ran > > than relying on mm_users. > > Agreed. > > If we can use mm_cpumask to track where the mm ever run, then if I'm > not mistaken we could optimize also multithreaded processes in the > same way: if only one thread is running frequently and the others are > frequently sleeping, we could issue a single tlbi broadcast (modulo > invalidates of small virtual ranges). Possibly, though not sure how you'd detect such scenario. > In the meantime the below should be enough to address the concern you > raised of the proof of concept RFC patch. > > I already experimented with mm_users == 1 earlier and it doesn't > change the benchmark results for the "best case" below. > > (untested) > > diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h > index 772bbc45b867..a2d53b301f22 100644 > --- a/arch/arm64/include/asm/tlbflush.h > +++ b/arch/arm64/include/asm/tlbflush.h [...] > @@ -212,7 +215,8 @@ static inline void flush_tlb_page(struct vm_area_struct *vma, > unsigned long addr = __TLBI_VADDR(uaddr, ASID(mm)); > > /* avoid TLB-i broadcast to remote NUMA nodes if it's a local flush */ > - if (current->mm == mm && atomic_read(&mm->mm_users) <= 1) { > + if (current->mm == mm && atomic_read(&mm->mm_users) <= 1 && > + (system_uses_ttbr0_pan() || atomic_read(&mm->mm_count) == 1)) { > int cpu = get_cpu(); > > cpumask_setall(mm_cpumask(mm)); I think there is another race here. IIUC, the assumption you make is that when mm_users <= 1 && mm_count == 1, the only active user of this pgd/ASID is on the CPU doing the TLBI. This is not the case for try_to_unmap() where the above condition may be true but the active thread on a different CPU won't notice the local TLBI. > > That's a pretty artificial test and it is indeed improved by this patch. > > However, it would be nice to have some real-world scenarios where this > > matters. [...] > Still your question if it'll make a difference in practice is a good > one and I don't have a sure answer yet. I suppose before doing more > benchmarking it's better to make a new version of this that uses > mm_cpumask to track where the asid was ever loaded as you suggested, > so that it will also optimize away tlbi broadcasts from multithreaded > processes where only one thread is running frequently? I was actually curious what triggered this patch series, whether you've seen a real use-case where the TLBI was a bottleneck.
Hello, On Tue, Feb 11, 2020 at 02:00:25PM +0000, Catalin Marinas wrote: > I think there is another race here. IIUC, the assumption you make is > that when mm_users <= 1 && mm_count == 1, the only active user of this > pgd/ASID is on the CPU doing the TLBI. This is not the case for > try_to_unmap() where the above condition may be true but the active > thread on a different CPU won't notice the local TLBI. The "current->mm == mm" check is what shall prevent the above. Thanks, Andrea
On 2/4/20 5:17 AM, Andrea Arcangeli wrote: > With multiple NUMA nodes and multiple sockets, the tlbi broadcast > shall be delivered through the interconnects in turn increasing the > interconnect traffic and the latency of the tlbi broadcast instruction. > > Even within a single NUMA node the latency of the tlbi broadcast > instruction increases almost linearly with the number of CPUs trying to > send tlbi broadcasts at the same time. > > When the process is single threaded however we can achieve full SMP > scalability by skipping the tlbi broadcasting. Other arches already > deploy this optimization. > > After the local TLB flush this however means the ASID context goes out > of sync in all CPUs except the local one. This can be tracked in the > mm_cpumask(mm): if the bit is set it means the asid context is stale > for that CPU. This results in an extra local ASID TLB flush only if a > single threaded process is migrated to a different CPU and only after a > TLB flush. No extra local TLB flush is needed for the common case of > single threaded processes context scheduling within the same CPU and for > multithreaded processes. > > Skipping the tlbi instruction broadcasting is already implemented in > local_flush_tlb_all(), this patch only extends it to flush_tlb_mm(), > flush_tlb_range() and flush_tlb_page() too. > > Here's the result of 32 CPUs (ARMv8 Ampere) running mprotect at the same > time from 32 single threaded processes before the patch: > > Performance counter stats for './loop' (3 runs): > > 0 dummy > > 2.121353 +- 0.000387 seconds time elapsed ( +- 0.02% ) > > and with the patch applied: > > Performance counter stats for './loop' (3 runs): > > 0 dummy > > 0.1197750 +- 0.0000827 seconds time elapsed ( +- 0.07% ) Hi, I have tested this patch on thunderX2 with Himeno benchmark[1] with LARGE calculation size. Here are the results. w/o patch: MFLOPS : 1149.480174 w/ patch: MFLOPS : 1110.653003 In order to validate the effectivness of the patch, I ran a single-threded program, which calls mprotect() in a loop to issue the tlbi broadcast instruction on a CPU core. At the same time, I ran Himeno benchmark on another CPU core. The results are: w/o patch: MFLOPS : 860.238792 w/ patch: MFLOPS : 1110.449666 Though Himeno benchmark is a microbenchmark, I hope it helps. [1] http://accc.riken.jp/en/supercom/documents/himenobmt/ Thanks, QI Fuli > > Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> > --- > arch/arm64/include/asm/efi.h | 2 +- > arch/arm64/include/asm/mmu.h | 3 +- > arch/arm64/include/asm/mmu_context.h | 10 +-- > arch/arm64/include/asm/tlbflush.h | 91 +++++++++++++++++++++++++++- > arch/arm64/mm/context.c | 13 +++- > 5 files changed, 109 insertions(+), 10 deletions(-) > > diff --git a/arch/arm64/include/asm/efi.h b/arch/arm64/include/asm/efi.h > index 44531a69d32b..5d9a1433d918 100644 > --- a/arch/arm64/include/asm/efi.h > +++ b/arch/arm64/include/asm/efi.h > @@ -131,7 +131,7 @@ static inline void efifb_setup_from_dmi(struct screen_info *si, const char *opt) > > static inline void efi_set_pgd(struct mm_struct *mm) > { > - __switch_mm(mm); > + __switch_mm(mm, smp_processor_id()); > > if (system_uses_ttbr0_pan()) { > if (mm != current->active_mm) { > diff --git a/arch/arm64/include/asm/mmu.h b/arch/arm64/include/asm/mmu.h > index e4d862420bb4..1f84289d3e44 100644 > --- a/arch/arm64/include/asm/mmu.h > +++ b/arch/arm64/include/asm/mmu.h > @@ -27,7 +27,8 @@ typedef struct { > * ASID change and therefore doesn't need to reload the counter using > * atomic64_read. > */ > -#define ASID(mm) ((mm)->context.id.counter & 0xffff) > +#define __ASID(asid) ((asid) & 0xffff) > +#define ASID(mm) __ASID((mm)->context.id.counter) > > extern bool arm64_use_ng_mappings; > > diff --git a/arch/arm64/include/asm/mmu_context.h b/arch/arm64/include/asm/mmu_context.h > index 3827ff4040a3..5ec264297968 100644 > --- a/arch/arm64/include/asm/mmu_context.h > +++ b/arch/arm64/include/asm/mmu_context.h > @@ -210,10 +210,8 @@ enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk) > update_saved_ttbr0(tsk, &init_mm); > } > > -static inline void __switch_mm(struct mm_struct *next) > +static inline void __switch_mm(struct mm_struct *next, unsigned int cpu) > { > - unsigned int cpu = smp_processor_id(); > - > /* > * init_mm.pgd does not contain any user mappings and it is always > * active for kernel addresses in TTBR1. Just set the reserved TTBR0. > @@ -230,8 +228,12 @@ static inline void > switch_mm(struct mm_struct *prev, struct mm_struct *next, > struct task_struct *tsk) > { > + unsigned int cpu = smp_processor_id(); > + > if (prev != next) > - __switch_mm(next); > + __switch_mm(next, cpu); > + else if (cpumask_test_and_clear_cpu(cpu, mm_cpumask(next))) > + local_flush_tlb_asid(atomic64_read(&next->context.id)); > > /* > * Update the saved TTBR0_EL1 of the scheduled-in task as the previous > diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h > index bc3949064725..283f97af3fc5 100644 > --- a/arch/arm64/include/asm/tlbflush.h > +++ b/arch/arm64/include/asm/tlbflush.h > @@ -136,6 +136,15 @@ static inline void local_flush_tlb_all(void) > isb(); > } > > +static inline void local_flush_tlb_asid(unsigned long asid) > +{ > + asid = __TLBI_VADDR(0, __ASID(asid)); > + dsb(nshst); > + __tlbi(aside1, asid); > + __tlbi_user(aside1, asid); > + dsb(nsh); > +} > + > static inline void flush_tlb_all(void) > { > dsb(ishst); > @@ -148,6 +157,27 @@ static inline void flush_tlb_mm(struct mm_struct *mm) > { > unsigned long asid = __TLBI_VADDR(0, ASID(mm)); > > + /* avoid TLB-i broadcast to remote NUMA nodes if it's a local flush */ > + if (current->mm == mm && atomic_read(&mm->mm_users) <= 1) { > + int cpu = get_cpu(); > + > + cpumask_setall(mm_cpumask(mm)); > + cpumask_clear_cpu(cpu, mm_cpumask(mm)); > + > + smp_mb(); > + > + if (atomic_read(&mm->mm_users) <= 1) { > + dsb(nshst); > + __tlbi(aside1, asid); > + __tlbi_user(aside1, asid); > + dsb(nsh); > + > + put_cpu(); > + return; > + } > + put_cpu(); > + } > + > dsb(ishst); > __tlbi(aside1is, asid); > __tlbi_user(aside1is, asid); > @@ -167,7 +197,33 @@ static inline void flush_tlb_page_nosync(struct vm_area_struct *vma, > static inline void flush_tlb_page(struct vm_area_struct *vma, > unsigned long uaddr) > { > - flush_tlb_page_nosync(vma, uaddr); > + struct mm_struct *mm = vma->vm_mm; > + unsigned long addr = __TLBI_VADDR(uaddr, ASID(mm)); > + > + /* avoid TLB-i broadcast to remote NUMA nodes if it's a local flush */ > + if (current->mm == mm && atomic_read(&mm->mm_users) <= 1) { > + int cpu = get_cpu(); > + > + cpumask_setall(mm_cpumask(mm)); > + cpumask_clear_cpu(cpu, mm_cpumask(mm)); > + > + smp_mb(); > + > + if (atomic_read(&mm->mm_users) <= 1) { > + dsb(nshst); > + __tlbi(vale1, addr); > + __tlbi_user(vale1, addr); > + dsb(nsh); > + > + put_cpu(); > + return; > + } > + put_cpu(); > + } > + > + dsb(ishst); > + __tlbi(vale1is, addr); > + __tlbi_user(vale1is, addr); > dsb(ish); > } > > @@ -181,14 +237,15 @@ static inline void __flush_tlb_range(struct vm_area_struct *vma, > unsigned long start, unsigned long end, > unsigned long stride, bool last_level) > { > - unsigned long asid = ASID(vma->vm_mm); > + struct mm_struct *mm = vma->vm_mm; > + unsigned long asid = ASID(mm); > unsigned long addr; > > start = round_down(start, stride); > end = round_up(end, stride); > > if ((end - start) >= (MAX_TLBI_OPS * stride)) { > - flush_tlb_mm(vma->vm_mm); > + flush_tlb_mm(mm); > return; > } > > @@ -198,6 +255,34 @@ static inline void __flush_tlb_range(struct vm_area_struct *vma, > start = __TLBI_VADDR(start, asid); > end = __TLBI_VADDR(end, asid); > > + /* avoid TLB-i broadcast to remote NUMA nodes if it's a local flush */ > + if (current->mm == mm && atomic_read(&mm->mm_users) <= 1) { > + int cpu = get_cpu(); > + > + cpumask_setall(mm_cpumask(mm)); > + cpumask_clear_cpu(cpu, mm_cpumask(mm)); > + > + smp_mb(); > + > + if (atomic_read(&mm->mm_users) <= 1) { > + dsb(nshst); > + for (addr = start; addr < end; addr += stride) { > + if (last_level) { > + __tlbi(vale1, addr); > + __tlbi_user(vale1, addr); > + } else { > + __tlbi(vae1, addr); > + __tlbi_user(vae1, addr); > + } > + } > + dsb(nsh); > + > + put_cpu(); > + return; > + } > + put_cpu(); > + } > + > dsb(ishst); > for (addr = start; addr < end; addr += stride) { > if (last_level) { > diff --git a/arch/arm64/mm/context.c b/arch/arm64/mm/context.c > index 8ef73e89d514..b44459d64dff 100644 > --- a/arch/arm64/mm/context.c > +++ b/arch/arm64/mm/context.c > @@ -198,6 +198,7 @@ void check_and_switch_context(struct mm_struct *mm, unsigned int cpu) > { > unsigned long flags; > u64 asid, old_active_asid; > + int need_flush = 0; > > if (system_supports_cnp()) > cpu_set_reserved_ttbr0(); > @@ -233,7 +234,8 @@ void check_and_switch_context(struct mm_struct *mm, unsigned int cpu) > atomic64_set(&mm->context.id, asid); > } > > - if (cpumask_test_and_clear_cpu(cpu, &tlb_flush_pending)) > + need_flush = cpumask_test_and_clear_cpu(cpu, &tlb_flush_pending); > + if (need_flush) > local_flush_tlb_all(); > > atomic64_set(&per_cpu(active_asids, cpu), asid); > @@ -241,6 +243,15 @@ void check_and_switch_context(struct mm_struct *mm, unsigned int cpu) > > switch_mm_fastpath: > > + /* > + * Enforce CPU ordering between the mmget() in use_mm() and > + * the below cpumask_test_and_clear_cpu(). > + */ > + smp_mb(); > + if (cpumask_test_and_clear_cpu(cpu, mm_cpumask(mm)) && > + likely(!need_flush)) > + local_flush_tlb_asid(asid); > + > arm64_apply_bp_hardening(); > > /* > > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel >
On Wed, Feb 12, 2020 at 02:13:56PM +0000, qi.fuli@fujitsu.com wrote: > On 2/4/20 5:17 AM, Andrea Arcangeli wrote: > > With multiple NUMA nodes and multiple sockets, the tlbi broadcast > > shall be delivered through the interconnects in turn increasing the > > interconnect traffic and the latency of the tlbi broadcast instruction. > > > > Even within a single NUMA node the latency of the tlbi broadcast > > instruction increases almost linearly with the number of CPUs trying to > > send tlbi broadcasts at the same time. > > > > When the process is single threaded however we can achieve full SMP > > scalability by skipping the tlbi broadcasting. Other arches already > > deploy this optimization. > > > > After the local TLB flush this however means the ASID context goes out > > of sync in all CPUs except the local one. This can be tracked in the > > mm_cpumask(mm): if the bit is set it means the asid context is stale > > for that CPU. This results in an extra local ASID TLB flush only if a > > single threaded process is migrated to a different CPU and only after a > > TLB flush. No extra local TLB flush is needed for the common case of > > single threaded processes context scheduling within the same CPU and for > > multithreaded processes. > > > > Skipping the tlbi instruction broadcasting is already implemented in > > local_flush_tlb_all(), this patch only extends it to flush_tlb_mm(), > > flush_tlb_range() and flush_tlb_page() too. > > > > Here's the result of 32 CPUs (ARMv8 Ampere) running mprotect at the same > > time from 32 single threaded processes before the patch: > > > > Performance counter stats for './loop' (3 runs): > > > > 0 dummy > > > > 2.121353 +- 0.000387 seconds time elapsed ( +- 0.02% ) > > > > and with the patch applied: > > > > Performance counter stats for './loop' (3 runs): > > > > 0 dummy > > > > 0.1197750 +- 0.0000827 seconds time elapsed ( +- 0.07% ) > > I have tested this patch on thunderX2 with Himeno benchmark[1] with > LARGE calculation size. Here are the results. > > w/o patch: MFLOPS : 1149.480174 > w/ patch: MFLOPS : 1110.653003 > > In order to validate the effectivness of the patch, I ran a > single-threded program, which calls mprotect() in a loop to issue the > tlbi broadcast instruction on a CPU core. At the same time, I ran Himeno > benchmark on another CPU core. The results are: > > w/o patch: MFLOPS : 860.238792 > w/ patch: MFLOPS : 1110.449666 > > Though Himeno benchmark is a microbenchmark, I hope it helps. It doesn't really help. What if you have a two-thread program calling mprotect() in a loop? IOW, how is this relevant to real-world scenarios? Thanks.
diff --git a/arch/arm64/include/asm/efi.h b/arch/arm64/include/asm/efi.h index 44531a69d32b..5d9a1433d918 100644 --- a/arch/arm64/include/asm/efi.h +++ b/arch/arm64/include/asm/efi.h @@ -131,7 +131,7 @@ static inline void efifb_setup_from_dmi(struct screen_info *si, const char *opt) static inline void efi_set_pgd(struct mm_struct *mm) { - __switch_mm(mm); + __switch_mm(mm, smp_processor_id()); if (system_uses_ttbr0_pan()) { if (mm != current->active_mm) { diff --git a/arch/arm64/include/asm/mmu.h b/arch/arm64/include/asm/mmu.h index e4d862420bb4..1f84289d3e44 100644 --- a/arch/arm64/include/asm/mmu.h +++ b/arch/arm64/include/asm/mmu.h @@ -27,7 +27,8 @@ typedef struct { * ASID change and therefore doesn't need to reload the counter using * atomic64_read. */ -#define ASID(mm) ((mm)->context.id.counter & 0xffff) +#define __ASID(asid) ((asid) & 0xffff) +#define ASID(mm) __ASID((mm)->context.id.counter) extern bool arm64_use_ng_mappings; diff --git a/arch/arm64/include/asm/mmu_context.h b/arch/arm64/include/asm/mmu_context.h index 3827ff4040a3..5ec264297968 100644 --- a/arch/arm64/include/asm/mmu_context.h +++ b/arch/arm64/include/asm/mmu_context.h @@ -210,10 +210,8 @@ enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk) update_saved_ttbr0(tsk, &init_mm); } -static inline void __switch_mm(struct mm_struct *next) +static inline void __switch_mm(struct mm_struct *next, unsigned int cpu) { - unsigned int cpu = smp_processor_id(); - /* * init_mm.pgd does not contain any user mappings and it is always * active for kernel addresses in TTBR1. Just set the reserved TTBR0. @@ -230,8 +228,12 @@ static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next, struct task_struct *tsk) { + unsigned int cpu = smp_processor_id(); + if (prev != next) - __switch_mm(next); + __switch_mm(next, cpu); + else if (cpumask_test_and_clear_cpu(cpu, mm_cpumask(next))) + local_flush_tlb_asid(atomic64_read(&next->context.id)); /* * Update the saved TTBR0_EL1 of the scheduled-in task as the previous diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h index bc3949064725..283f97af3fc5 100644 --- a/arch/arm64/include/asm/tlbflush.h +++ b/arch/arm64/include/asm/tlbflush.h @@ -136,6 +136,15 @@ static inline void local_flush_tlb_all(void) isb(); } +static inline void local_flush_tlb_asid(unsigned long asid) +{ + asid = __TLBI_VADDR(0, __ASID(asid)); + dsb(nshst); + __tlbi(aside1, asid); + __tlbi_user(aside1, asid); + dsb(nsh); +} + static inline void flush_tlb_all(void) { dsb(ishst); @@ -148,6 +157,27 @@ static inline void flush_tlb_mm(struct mm_struct *mm) { unsigned long asid = __TLBI_VADDR(0, ASID(mm)); + /* avoid TLB-i broadcast to remote NUMA nodes if it's a local flush */ + if (current->mm == mm && atomic_read(&mm->mm_users) <= 1) { + int cpu = get_cpu(); + + cpumask_setall(mm_cpumask(mm)); + cpumask_clear_cpu(cpu, mm_cpumask(mm)); + + smp_mb(); + + if (atomic_read(&mm->mm_users) <= 1) { + dsb(nshst); + __tlbi(aside1, asid); + __tlbi_user(aside1, asid); + dsb(nsh); + + put_cpu(); + return; + } + put_cpu(); + } + dsb(ishst); __tlbi(aside1is, asid); __tlbi_user(aside1is, asid); @@ -167,7 +197,33 @@ static inline void flush_tlb_page_nosync(struct vm_area_struct *vma, static inline void flush_tlb_page(struct vm_area_struct *vma, unsigned long uaddr) { - flush_tlb_page_nosync(vma, uaddr); + struct mm_struct *mm = vma->vm_mm; + unsigned long addr = __TLBI_VADDR(uaddr, ASID(mm)); + + /* avoid TLB-i broadcast to remote NUMA nodes if it's a local flush */ + if (current->mm == mm && atomic_read(&mm->mm_users) <= 1) { + int cpu = get_cpu(); + + cpumask_setall(mm_cpumask(mm)); + cpumask_clear_cpu(cpu, mm_cpumask(mm)); + + smp_mb(); + + if (atomic_read(&mm->mm_users) <= 1) { + dsb(nshst); + __tlbi(vale1, addr); + __tlbi_user(vale1, addr); + dsb(nsh); + + put_cpu(); + return; + } + put_cpu(); + } + + dsb(ishst); + __tlbi(vale1is, addr); + __tlbi_user(vale1is, addr); dsb(ish); } @@ -181,14 +237,15 @@ static inline void __flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end, unsigned long stride, bool last_level) { - unsigned long asid = ASID(vma->vm_mm); + struct mm_struct *mm = vma->vm_mm; + unsigned long asid = ASID(mm); unsigned long addr; start = round_down(start, stride); end = round_up(end, stride); if ((end - start) >= (MAX_TLBI_OPS * stride)) { - flush_tlb_mm(vma->vm_mm); + flush_tlb_mm(mm); return; } @@ -198,6 +255,34 @@ static inline void __flush_tlb_range(struct vm_area_struct *vma, start = __TLBI_VADDR(start, asid); end = __TLBI_VADDR(end, asid); + /* avoid TLB-i broadcast to remote NUMA nodes if it's a local flush */ + if (current->mm == mm && atomic_read(&mm->mm_users) <= 1) { + int cpu = get_cpu(); + + cpumask_setall(mm_cpumask(mm)); + cpumask_clear_cpu(cpu, mm_cpumask(mm)); + + smp_mb(); + + if (atomic_read(&mm->mm_users) <= 1) { + dsb(nshst); + for (addr = start; addr < end; addr += stride) { + if (last_level) { + __tlbi(vale1, addr); + __tlbi_user(vale1, addr); + } else { + __tlbi(vae1, addr); + __tlbi_user(vae1, addr); + } + } + dsb(nsh); + + put_cpu(); + return; + } + put_cpu(); + } + dsb(ishst); for (addr = start; addr < end; addr += stride) { if (last_level) { diff --git a/arch/arm64/mm/context.c b/arch/arm64/mm/context.c index 8ef73e89d514..b44459d64dff 100644 --- a/arch/arm64/mm/context.c +++ b/arch/arm64/mm/context.c @@ -198,6 +198,7 @@ void check_and_switch_context(struct mm_struct *mm, unsigned int cpu) { unsigned long flags; u64 asid, old_active_asid; + int need_flush = 0; if (system_supports_cnp()) cpu_set_reserved_ttbr0(); @@ -233,7 +234,8 @@ void check_and_switch_context(struct mm_struct *mm, unsigned int cpu) atomic64_set(&mm->context.id, asid); } - if (cpumask_test_and_clear_cpu(cpu, &tlb_flush_pending)) + need_flush = cpumask_test_and_clear_cpu(cpu, &tlb_flush_pending); + if (need_flush) local_flush_tlb_all(); atomic64_set(&per_cpu(active_asids, cpu), asid); @@ -241,6 +243,15 @@ void check_and_switch_context(struct mm_struct *mm, unsigned int cpu) switch_mm_fastpath: + /* + * Enforce CPU ordering between the mmget() in use_mm() and + * the below cpumask_test_and_clear_cpu(). + */ + smp_mb(); + if (cpumask_test_and_clear_cpu(cpu, mm_cpumask(mm)) && + likely(!need_flush)) + local_flush_tlb_asid(asid); + arm64_apply_bp_hardening(); /*
With multiple NUMA nodes and multiple sockets, the tlbi broadcast shall be delivered through the interconnects in turn increasing the interconnect traffic and the latency of the tlbi broadcast instruction. Even within a single NUMA node the latency of the tlbi broadcast instruction increases almost linearly with the number of CPUs trying to send tlbi broadcasts at the same time. When the process is single threaded however we can achieve full SMP scalability by skipping the tlbi broadcasting. Other arches already deploy this optimization. After the local TLB flush this however means the ASID context goes out of sync in all CPUs except the local one. This can be tracked in the mm_cpumask(mm): if the bit is set it means the asid context is stale for that CPU. This results in an extra local ASID TLB flush only if a single threaded process is migrated to a different CPU and only after a TLB flush. No extra local TLB flush is needed for the common case of single threaded processes context scheduling within the same CPU and for multithreaded processes. Skipping the tlbi instruction broadcasting is already implemented in local_flush_tlb_all(), this patch only extends it to flush_tlb_mm(), flush_tlb_range() and flush_tlb_page() too. Here's the result of 32 CPUs (ARMv8 Ampere) running mprotect at the same time from 32 single threaded processes before the patch: Performance counter stats for './loop' (3 runs): 0 dummy 2.121353 +- 0.000387 seconds time elapsed ( +- 0.02% ) and with the patch applied: Performance counter stats for './loop' (3 runs): 0 dummy 0.1197750 +- 0.0000827 seconds time elapsed ( +- 0.07% ) Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> --- arch/arm64/include/asm/efi.h | 2 +- arch/arm64/include/asm/mmu.h | 3 +- arch/arm64/include/asm/mmu_context.h | 10 +-- arch/arm64/include/asm/tlbflush.h | 91 +++++++++++++++++++++++++++- arch/arm64/mm/context.c | 13 +++- 5 files changed, 109 insertions(+), 10 deletions(-)