From 61c77326d1df079f202fa79403c3ccd8c5966a81 Mon Sep 17 00:00:00 2001 From: Shaohua Li Date: Mon, 16 Aug 2010 09:16:55 +0800 Subject: x86, mm: Avoid unnecessary TLB flush In x86, access and dirty bits are set automatically by CPU when CPU accesses memory. When we go into the code path of below flush_tlb_fix_spurious_fault(), we already set dirty bit for pte and don't need flush tlb. This might mean tlb entry in some CPUs hasn't dirty bit set, but this doesn't matter. When the CPUs do page write, they will automatically check the bit and no software involved. On the other hand, flush tlb in below position is harmful. Test creates CPU number of threads, each thread writes to a same but random address in same vma range and we measure the total time. Under a 4 socket system, original time is 1.96s, while with the patch, the time is 0.8s. Under a 2 socket system, there is 20% time cut too. perf shows a lot of time are taking to send ipi/handle ipi for tlb flush. Signed-off-by: Shaohua Li LKML-Reference: <20100816011655.GA362@sli10-desk.sh.intel.com> Acked-by: Suresh Siddha Cc: Andrea Archangeli Signed-off-by: H. Peter Anvin --- mm/memory.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'mm') diff --git a/mm/memory.c b/mm/memory.c index 2ed2267439df..a40da6983961 100644 --- a/mm/memory.c +++ b/mm/memory.c @@ -3147,7 +3147,7 @@ static inline int handle_pte_fault(struct mm_struct *mm, * with threads. */ if (flags & FAULT_FLAG_WRITE) - flush_tlb_page(vma, address); + flush_tlb_fix_spurious_fault(vma, address); } unlock: pte_unmap_unlock(pte, ptl); -- cgit v1.2.3 From 3ee48b6af49cf534ca2f481ecc484b156a41451d Mon Sep 17 00:00:00 2001 From: Cliff Wickman Date: Thu, 16 Sep 2010 11:44:02 -0500 Subject: mm, x86: Saving vmcore with non-lazy freeing of vmas During the reading of /proc/vmcore the kernel is doing ioremap()/iounmap() repeatedly. And the buildup of un-flushed vm_area_struct's is causing a great deal of overhead. (rb_next() is chewing up most of that time). This solution is to provide function set_iounmap_nonlazy(). It causes a subsequent call to iounmap() to immediately purge the vma area (with try_purge_vmap_area_lazy()). With this patch we have seen the time for writing a 250MB compressed dump drop from 71 seconds to 44 seconds. Signed-off-by: Cliff Wickman Cc: Andrew Morton Cc: kexec@lists.infradead.org Cc: LKML-Reference: Signed-off-by: Ingo Molnar --- arch/x86/include/asm/io.h | 1 + arch/x86/kernel/crash_dump_64.c | 1 + mm/vmalloc.c | 9 +++++++++ 3 files changed, 11 insertions(+) (limited to 'mm') diff --git a/arch/x86/include/asm/io.h b/arch/x86/include/asm/io.h index 30a3e9776123..6a45ec41ec26 100644 --- a/arch/x86/include/asm/io.h +++ b/arch/x86/include/asm/io.h @@ -206,6 +206,7 @@ static inline void __iomem *ioremap(resource_size_t offset, unsigned long size) extern void iounmap(volatile void __iomem *addr); +extern void set_iounmap_nonlazy(void); #ifdef __KERNEL__ diff --git a/arch/x86/kernel/crash_dump_64.c b/arch/x86/kernel/crash_dump_64.c index bf43188ca654..994828899e09 100644 --- a/arch/x86/kernel/crash_dump_64.c +++ b/arch/x86/kernel/crash_dump_64.c @@ -46,6 +46,7 @@ ssize_t copy_oldmem_page(unsigned long pfn, char *buf, } else memcpy(buf, vaddr + offset, csize); + set_iounmap_nonlazy(); iounmap(vaddr); return csize; } diff --git a/mm/vmalloc.c b/mm/vmalloc.c index 6b8889da69a6..d8087f0db507 100644 --- a/mm/vmalloc.c +++ b/mm/vmalloc.c @@ -516,6 +516,15 @@ static atomic_t vmap_lazy_nr = ATOMIC_INIT(0); /* for per-CPU blocks */ static void purge_fragmented_blocks_allcpus(void); +/* + * called before a call to iounmap() if the caller wants vm_area_struct's + * immediately freed. + */ +void set_iounmap_nonlazy(void) +{ + atomic_set(&vmap_lazy_nr, lazy_max_pages()+1); +} + /* * Purges all lazily-freed vmap areas. * -- cgit v1.2.3