From b13b1d2d8692b437203de7a404c6b809d2cc4d99 Mon Sep 17 00:00:00 2001 From: Shaohua Li Date: Tue, 8 Apr 2014 15:58:09 +0800 Subject: [PATCH 1/2] x86/mm: In the PTE swapout page reclaim case clear the accessed bit instead of flushing the TLB We use the accessed bit to age a page at page reclaim time, and currently we also flush the TLB when doing so. But in some workloads TLB flush overhead is very heavy. In my simple multithreaded app with a lot of swap to several pcie SSDs, removing the tlb flush gives about 20% ~ 30% swapout speedup. Fortunately just removing the TLB flush is a valid optimization: on x86 CPUs, clearing the accessed bit without a TLB flush doesn't cause data corruption. It could cause incorrect page aging and the (mistaken) reclaim of hot pages, but the chance of that should be relatively low. So as a performance optimization don't flush the TLB when clearing the accessed bit, it will eventually be flushed by a context switch or a VM operation anyway. [ In the rare event of it not getting flushed for a long time the delay shouldn't really matter because there's no real memory pressure for swapout to react to. ] Suggested-by: Linus Torvalds Signed-off-by: Shaohua Li Acked-by: Rik van Riel Acked-by: Mel Gorman Acked-by: Hugh Dickins Acked-by: Johannes Weiner Cc: linux-mm@kvack.org Cc: Peter Zijlstra Link: http://lkml.kernel.org/r/20140408075809.GA1764@kernel.org [ Rewrote the changelog and the code comments. ] Signed-off-by: Ingo Molnar --- arch/x86/mm/pgtable.c | 21 ++++++++++++++------- 1 file changed, 14 insertions(+), 7 deletions(-) diff --git a/arch/x86/mm/pgtable.c b/arch/x86/mm/pgtable.c index c96314abd144..0004ac72dbdd 100644 --- a/arch/x86/mm/pgtable.c +++ b/arch/x86/mm/pgtable.c @@ -399,13 +399,20 @@ int pmdp_test_and_clear_young(struct vm_area_struct *vma, int ptep_clear_flush_young(struct vm_area_struct *vma, unsigned long address, pte_t *ptep) { - int young; - - young = ptep_test_and_clear_young(vma, address, ptep); - if (young) - flush_tlb_page(vma, address); - - return young; + /* + * On x86 CPUs, clearing the accessed bit without a TLB flush + * doesn't cause data corruption. [ It could cause incorrect + * page aging and the (mistaken) reclaim of hot pages, but the + * chance of that should be relatively low. ] + * + * So as a performance optimization don't flush the TLB when + * clearing the accessed bit, it will eventually be flushed by + * a context switch or a VM operation anyway. [ In the rare + * event of it not getting flushed for a long time the delay + * shouldn't really matter because there's no real memory + * pressure for swapout to react to. ] + */ + return ptep_test_and_clear_young(vma, address, ptep); } #ifdef CONFIG_TRANSPARENT_HUGEPAGE From c81c8a1eeede61e92a15103748c23d100880cc8a Mon Sep 17 00:00:00 2001 From: Roland Dreier Date: Fri, 2 May 2014 11:18:41 -0700 Subject: [PATCH 2/2] x86, ioremap: Speed up check for RAM pages In __ioremap_caller() (the guts of ioremap), we loop over the range of pfns being remapped and checks each one individually with page_is_ram(). For large ioremaps, this can be very slow. For example, we have a device with a 256 GiB PCI BAR, and ioremapping this BAR can take 20+ seconds -- sometimes long enough to trigger the soft lockup detector! Internally, page_is_ram() calls walk_system_ram_range() on a single page. Instead, we can make a single call to walk_system_ram_range() from __ioremap_caller(), and do our further checks only for any RAM pages that we find. For the common case of MMIO, this saves an enormous amount of work, since the range being ioremapped doesn't intersect system RAM at all. With this change, ioremap on our 256 GiB BAR takes less than 1 second. Signed-off-by: Roland Dreier Link: http://lkml.kernel.org/r/1399054721-1331-1-git-send-email-roland@kernel.org Signed-off-by: H. Peter Anvin --- arch/x86/mm/ioremap.c | 26 +++++++++++++++++++------- 1 file changed, 19 insertions(+), 7 deletions(-) diff --git a/arch/x86/mm/ioremap.c b/arch/x86/mm/ioremap.c index 597ac155c91c..bc7527e109c8 100644 --- a/arch/x86/mm/ioremap.c +++ b/arch/x86/mm/ioremap.c @@ -50,6 +50,21 @@ int ioremap_change_attr(unsigned long vaddr, unsigned long size, return err; } +static int __ioremap_check_ram(unsigned long start_pfn, unsigned long nr_pages, + void *arg) +{ + unsigned long i; + + for (i = 0; i < nr_pages; ++i) + if (pfn_valid(start_pfn + i) && + !PageReserved(pfn_to_page(start_pfn + i))) + return 1; + + WARN_ONCE(1, "ioremap on RAM pfn 0x%lx\n", start_pfn); + + return 0; +} + /* * Remap an arbitrary physical address space into the kernel virtual * address space. Needed when the kernel wants to access high addresses @@ -93,14 +108,11 @@ static void __iomem *__ioremap_caller(resource_size_t phys_addr, /* * Don't allow anybody to remap normal RAM that we're using.. */ + pfn = phys_addr >> PAGE_SHIFT; last_pfn = last_addr >> PAGE_SHIFT; - for (pfn = phys_addr >> PAGE_SHIFT; pfn <= last_pfn; pfn++) { - int is_ram = page_is_ram(pfn); - - if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn))) - return NULL; - WARN_ON_ONCE(is_ram); - } + if (walk_system_ram_range(pfn, last_pfn - pfn + 1, NULL, + __ioremap_check_ram) == 1) + return NULL; /* * Mappings have to be page-aligned