| Message ID | 98134cb73e911b2f0b59ffb76243a7777963d218.1550088114.git.khalid.aziz@oracle.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Add support for eXclusive Page Frame Ownership | expand |
> #endif > + > + /* If there is a pending TLB flush for this CPU due to XPFO > + * flush, do it now. > + */ Don't forget CodingStyle in all this, please. > + if (cpumask_test_and_clear_cpu(cpu, &pending_xpfo_flush)) { > + count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED); > + __flush_tlb_all(); > + } This seems to exist in parallel with all of the cpu_tlbstate infrastructure. Shouldn't it go in there? Also, if we're doing full flushes like this, it seems a bit wasteful to then go and do later things like invalidate_user_asid() when we *know* that the asid would have been flushed by this operation. I'm pretty sure this isn't the only __flush_tlb_all() callsite that does this, so it's not really criticism of this patch specifically. It's more of a structural issue. > +void xpfo_flush_tlb_kernel_range(unsigned long start, unsigned long end) > +{ This is a bit lightly commented. Please give this some good descriptions about the logic behind the implementation and the tradeoffs that are in play. This is doing a local flush, but deferring the flushes on all other processors, right? Can you explain the logic behind that in a comment here, please? This also has to be called with preemption disabled, right? > + struct cpumask tmp_mask; > + > + /* Balance as user space task's flush, a bit conservative */ > + if (end == TLB_FLUSH_ALL || > + (end - start) > tlb_single_page_flush_ceiling << PAGE_SHIFT) { > + do_flush_tlb_all(NULL); > + } else { > + struct flush_tlb_info info; > + > + info.start = start; > + info.end = end; > + do_kernel_range_flush(&info); > + } > + cpumask_setall(&tmp_mask); > + cpumask_clear_cpu(smp_processor_id(), &tmp_mask); > + cpumask_or(&pending_xpfo_flush, &pending_xpfo_flush, &tmp_mask); > +} Fun. cpumask_setall() is non-atomic while cpumask_clear_cpu() and cpumask_or() *are* atomic. The cpumask_clear_cpu() is operating on thread-local storage and doesn't need to be atomic. Please make it __cpumask_clear_cpu().
On 2/14/19 10:42 AM, Dave Hansen wrote: >> #endif >> + >> + /* If there is a pending TLB flush for this CPU due to XPFO >> + * flush, do it now. >> + */ > > Don't forget CodingStyle in all this, please. Of course. I will fix that. > >> + if (cpumask_test_and_clear_cpu(cpu, &pending_xpfo_flush)) { >> + count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED); >> + __flush_tlb_all(); >> + } > > This seems to exist in parallel with all of the cpu_tlbstate > infrastructure. Shouldn't it go in there? That sounds like a good idea. On the other hand, pending flush needs to be kept track of entirely within arch/x86/mm/tlb.c and using a local variable with scope limited to just that file feels like a lighter weight implementation. I could go either way. > > Also, if we're doing full flushes like this, it seems a bit wasteful to > then go and do later things like invalidate_user_asid() when we *know* > that the asid would have been flushed by this operation. I'm pretty > sure this isn't the only __flush_tlb_all() callsite that does this, so > it's not really criticism of this patch specifically. It's more of a > structural issue. > > That is a good point. It is not just wasteful, it is bound to have performance impact even if slight. >> +void xpfo_flush_tlb_kernel_range(unsigned long start, unsigned long end) >> +{ > > This is a bit lightly commented. Please give this some good > descriptions about the logic behind the implementation and the tradeoffs > that are in play. > > This is doing a local flush, but deferring the flushes on all other > processors, right? Can you explain the logic behind that in a comment > here, please? This also has to be called with preemption disabled, right? > >> + struct cpumask tmp_mask; >> + >> + /* Balance as user space task's flush, a bit conservative */ >> + if (end == TLB_FLUSH_ALL || >> + (end - start) > tlb_single_page_flush_ceiling << PAGE_SHIFT) { >> + do_flush_tlb_all(NULL); >> + } else { >> + struct flush_tlb_info info; >> + >> + info.start = start; >> + info.end = end; >> + do_kernel_range_flush(&info); >> + } >> + cpumask_setall(&tmp_mask); >> + cpumask_clear_cpu(smp_processor_id(), &tmp_mask); >> + cpumask_or(&pending_xpfo_flush, &pending_xpfo_flush, &tmp_mask); >> +} > > Fun. cpumask_setall() is non-atomic while cpumask_clear_cpu() and > cpumask_or() *are* atomic. The cpumask_clear_cpu() is operating on > thread-local storage and doesn't need to be atomic. Please make it > __cpumask_clear_cpu(). > I will fix that. Thanks! -- Khalid
diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h index f4204bf377fc..92d23629d01d 100644 --- a/arch/x86/include/asm/tlbflush.h +++ b/arch/x86/include/asm/tlbflush.h @@ -561,6 +561,7 @@ extern void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start, unsigned long end, unsigned int stride_shift, bool freed_tables); extern void flush_tlb_kernel_range(unsigned long start, unsigned long end); +extern void xpfo_flush_tlb_kernel_range(unsigned long start, unsigned long end); static inline void flush_tlb_page(struct vm_area_struct *vma, unsigned long a) { diff --git a/arch/x86/mm/tlb.c b/arch/x86/mm/tlb.c index 03b6b4c2238d..c907b643eecb 100644 --- a/arch/x86/mm/tlb.c +++ b/arch/x86/mm/tlb.c @@ -35,6 +35,15 @@ */ #define LAST_USER_MM_IBPB 0x1UL +/* + * When a full TLB flush is needed to flush stale TLB entries + * for pages that have been mapped into userspace and unmapped + * from kernel space, this TLB flush will be delayed until the + * task is scheduled on that CPU. Keep track of CPUs with + * pending full TLB flush forced by xpfo. + */ +static cpumask_t pending_xpfo_flush; + /* * We get here when we do something requiring a TLB invalidation * but could not go invalidate all of the contexts. We do the @@ -319,6 +328,15 @@ void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next, __flush_tlb_all(); } #endif + + /* If there is a pending TLB flush for this CPU due to XPFO + * flush, do it now. + */ + if (cpumask_test_and_clear_cpu(cpu, &pending_xpfo_flush)) { + count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED); + __flush_tlb_all(); + } + this_cpu_write(cpu_tlbstate.is_lazy, false); /* @@ -801,6 +819,26 @@ void flush_tlb_kernel_range(unsigned long start, unsigned long end) } } +void xpfo_flush_tlb_kernel_range(unsigned long start, unsigned long end) +{ + struct cpumask tmp_mask; + + /* Balance as user space task's flush, a bit conservative */ + if (end == TLB_FLUSH_ALL || + (end - start) > tlb_single_page_flush_ceiling << PAGE_SHIFT) { + do_flush_tlb_all(NULL); + } else { + struct flush_tlb_info info; + + info.start = start; + info.end = end; + do_kernel_range_flush(&info); + } + cpumask_setall(&tmp_mask); + cpumask_clear_cpu(smp_processor_id(), &tmp_mask); + cpumask_or(&pending_xpfo_flush, &pending_xpfo_flush, &tmp_mask); +} + void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch) { struct flush_tlb_info info = { diff --git a/arch/x86/mm/xpfo.c b/arch/x86/mm/xpfo.c index e13b99019c47..d3833532bfdc 100644 --- a/arch/x86/mm/xpfo.c +++ b/arch/x86/mm/xpfo.c @@ -115,7 +115,7 @@ inline void xpfo_flush_kernel_tlb(struct page *page, int order) return; } - flush_tlb_kernel_range(kaddr, kaddr + (1 << order) * size); + xpfo_flush_tlb_kernel_range(kaddr, kaddr + (1 << order) * size); } /* Convert a user space virtual address to a physical address.
XPFO flushes kernel space TLB entries for pages that are now mapped in userspace on not only the current CPU but also all other CPUs synchronously. Processes on each core allocating pages causes a flood of IPI messages to all other cores to flush TLB entries. Many of these messages are to flush the entire TLB on the core if the number of entries being flushed from local core exceeds tlb_single_page_flush_ceiling. The cost of TLB flush caused by unmapping pages from physmap goes up dramatically on machines with high core count. This patch flushes relevant TLB entries for current process or entire TLB depending upon number of entries for the current CPU and posts a pending TLB flush on all other CPUs when a page is unmapped from kernel space and mapped in userspace. Each core checks the pending TLB flush flag for itself on every context switch, flushes its TLB if the flag is set and clears it. This patch potentially aggregates multiple TLB flushes into one. This has very significant impact especially on machines with large core counts. To illustrate this, kernel was compiled with -j on two classes of machines - a server with high core count and large amount of memory, and a desktop class machine with more modest specs. System time from "make -j" from vanilla 4.20 kernel, 4.20 with XPFO patches before applying this patch and after applying this patch are below: Hardware: 96-core Intel Xeon Platinum 8160 CPU @ 2.10GHz, 768 GB RAM make -j60 all 4.20 950.966s 4.20+XPFO 25073.169s 26.366x 4.20+XPFO+Deferred flush 1372.874s 1.44x Hardware: 4-core Intel Core i5-3550 CPU @ 3.30GHz, 8G RAM make -j4 all 4.20 607.671s 4.20+XPFO 1588.646s 2.614x 4.20+XPFO+Deferred flush 803.989s 1.32x This patch could use more optimization. Batching more TLB entry flushes, as was suggested for earlier version of these patches, can help reduce these cases. This same code should be implemented for other architectures as well once finalized. Signed-off-by: Khalid Aziz <khalid.aziz@oracle.com> --- arch/x86/include/asm/tlbflush.h | 1 + arch/x86/mm/tlb.c | 38 +++++++++++++++++++++++++++++++++ arch/x86/mm/xpfo.c | 2 +- 3 files changed, 40 insertions(+), 1 deletion(-)