diff options
author | Davidlohr Bueso <davidlohr@hp.com> | 2014-04-07 18:37:25 -0400 |
---|---|---|
committer | Linus Torvalds <torvalds@linux-foundation.org> | 2014-04-07 19:35:53 -0400 |
commit | 615d6e8756c87149f2d4c1b93d471bca002bd849 (patch) | |
tree | 45b039ccafb606a30e53c1012775efe848e789ed /mm | |
parent | d7c1755179b82d954f593ca5285b9360f2f62e9c (diff) |
mm: per-thread vma caching
This patch is a continuation of efforts trying to optimize find_vma(),
avoiding potentially expensive rbtree walks to locate a vma upon faults.
The original approach (https://lkml.org/lkml/2013/11/1/410), where the
largest vma was also cached, ended up being too specific and random,
thus further comparison with other approaches were needed. There are
two things to consider when dealing with this, the cache hit rate and
the latency of find_vma(). Improving the hit-rate does not necessarily
translate in finding the vma any faster, as the overhead of any fancy
caching schemes can be too high to consider.
We currently cache the last used vma for the whole address space, which
provides a nice optimization, reducing the total cycles in find_vma() by
up to 250%, for workloads with good locality. On the other hand, this
simple scheme is pretty much useless for workloads with poor locality.
Analyzing ebizzy runs shows that, no matter how many threads are
running, the mmap_cache hit rate is less than 2%, and in many situations
below 1%.
The proposed approach is to replace this scheme with a small per-thread
cache, maximizing hit rates at a very low maintenance cost.
Invalidations are performed by simply bumping up a 32-bit sequence
number. The only expensive operation is in the rare case of a seq
number overflow, where all caches that share the same address space are
flushed. Upon a miss, the proposed replacement policy is based on the
page number that contains the virtual address in question. Concretely,
the following results are seen on an 80 core, 8 socket x86-64 box:
1) System bootup: Most programs are single threaded, so the per-thread
scheme does improve ~50% hit rate by just adding a few more slots to
the cache.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 50.61% | 19.90 |
| patched | 73.45% | 13.58 |
+----------------+----------+------------------+
2) Kernel build: This one is already pretty good with the current
approach as we're dealing with good locality.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 75.28% | 11.03 |
| patched | 88.09% | 9.31 |
+----------------+----------+------------------+
3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 70.66% | 17.14 |
| patched | 91.15% | 12.57 |
+----------------+----------+------------------+
4) Ebizzy: There's a fair amount of variation from run to run, but this
approach always shows nearly perfect hit rates, while baseline is just
about non-existent. The amounts of cycles can fluctuate between
anywhere from ~60 to ~116 for the baseline scheme, but this approach
reduces it considerably. For instance, with 80 threads:
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 1.06% | 91.54 |
| patched | 99.97% | 14.18 |
+----------------+----------+------------------+
[akpm@linux-foundation.org: fix nommu build, per Davidlohr]
[akpm@linux-foundation.org: document vmacache_valid() logic]
[akpm@linux-foundation.org: attempt to untangle header files]
[akpm@linux-foundation.org: add vmacache_find() BUG_ON]
[hughd@google.com: add vmacache_valid_mm() (from Oleg)]
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: adjust and enhance comments]
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Michel Lespinasse <walken@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Tested-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'mm')
-rw-r--r-- | mm/Makefile | 2 | ||||
-rw-r--r-- | mm/mmap.c | 55 | ||||
-rw-r--r-- | mm/nommu.c | 24 | ||||
-rw-r--r-- | mm/vmacache.c | 112 |
4 files changed, 158 insertions, 35 deletions
diff --git a/mm/Makefile b/mm/Makefile index cdd741519ee0..23a6f7e23019 100644 --- a/mm/Makefile +++ b/mm/Makefile | |||
@@ -16,7 +16,7 @@ obj-y := filemap.o mempool.o oom_kill.o fadvise.o \ | |||
16 | readahead.o swap.o truncate.o vmscan.o shmem.o \ | 16 | readahead.o swap.o truncate.o vmscan.o shmem.o \ |
17 | util.o mmzone.o vmstat.o backing-dev.o \ | 17 | util.o mmzone.o vmstat.o backing-dev.o \ |
18 | mm_init.o mmu_context.o percpu.o slab_common.o \ | 18 | mm_init.o mmu_context.o percpu.o slab_common.o \ |
19 | compaction.o balloon_compaction.o \ | 19 | compaction.o balloon_compaction.o vmacache.o \ |
20 | interval_tree.o list_lru.o workingset.o $(mmu-y) | 20 | interval_tree.o list_lru.o workingset.o $(mmu-y) |
21 | 21 | ||
22 | obj-y += init-mm.o | 22 | obj-y += init-mm.o |
@@ -10,6 +10,7 @@ | |||
10 | #include <linux/slab.h> | 10 | #include <linux/slab.h> |
11 | #include <linux/backing-dev.h> | 11 | #include <linux/backing-dev.h> |
12 | #include <linux/mm.h> | 12 | #include <linux/mm.h> |
13 | #include <linux/vmacache.h> | ||
13 | #include <linux/shm.h> | 14 | #include <linux/shm.h> |
14 | #include <linux/mman.h> | 15 | #include <linux/mman.h> |
15 | #include <linux/pagemap.h> | 16 | #include <linux/pagemap.h> |
@@ -681,8 +682,9 @@ __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma, | |||
681 | prev->vm_next = next = vma->vm_next; | 682 | prev->vm_next = next = vma->vm_next; |
682 | if (next) | 683 | if (next) |
683 | next->vm_prev = prev; | 684 | next->vm_prev = prev; |
684 | if (mm->mmap_cache == vma) | 685 | |
685 | mm->mmap_cache = prev; | 686 | /* Kill the cache */ |
687 | vmacache_invalidate(mm); | ||
686 | } | 688 | } |
687 | 689 | ||
688 | /* | 690 | /* |
@@ -1989,34 +1991,33 @@ EXPORT_SYMBOL(get_unmapped_area); | |||
1989 | /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ | 1991 | /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ |
1990 | struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) | 1992 | struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) |
1991 | { | 1993 | { |
1992 | struct vm_area_struct *vma = NULL; | 1994 | struct rb_node *rb_node; |
1995 | struct vm_area_struct *vma; | ||
1993 | 1996 | ||
1994 | /* Check the cache first. */ | 1997 | /* Check the cache first. */ |
1995 | /* (Cache hit rate is typically around 35%.) */ | 1998 | vma = vmacache_find(mm, addr); |
1996 | vma = ACCESS_ONCE(mm->mmap_cache); | 1999 | if (likely(vma)) |
1997 | if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) { | 2000 | return vma; |
1998 | struct rb_node *rb_node; | ||
1999 | 2001 | ||
2000 | rb_node = mm->mm_rb.rb_node; | 2002 | rb_node = mm->mm_rb.rb_node; |
2001 | vma = NULL; | 2003 | vma = NULL; |
2002 | 2004 | ||
2003 | while (rb_node) { | 2005 | while (rb_node) { |
2004 | struct vm_area_struct *vma_tmp; | 2006 | struct vm_area_struct *tmp; |
2005 | 2007 | ||
2006 | vma_tmp = rb_entry(rb_node, | 2008 | tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); |
2007 | struct vm_area_struct, vm_rb); | 2009 | |
2008 | 2010 | if (tmp->vm_end > addr) { | |
2009 | if (vma_tmp->vm_end > addr) { | 2011 | vma = tmp; |
2010 | vma = vma_tmp; | 2012 | if (tmp->vm_start <= addr) |
2011 | if (vma_tmp->vm_start <= addr) | 2013 | break; |
2012 | break; | 2014 | rb_node = rb_node->rb_left; |
2013 | rb_node = rb_node->rb_left; | 2015 | } else |
2014 | } else | 2016 | rb_node = rb_node->rb_right; |
2015 | rb_node = rb_node->rb_right; | ||
2016 | } | ||
2017 | if (vma) | ||
2018 | mm->mmap_cache = vma; | ||
2019 | } | 2017 | } |
2018 | |||
2019 | if (vma) | ||
2020 | vmacache_update(addr, vma); | ||
2020 | return vma; | 2021 | return vma; |
2021 | } | 2022 | } |
2022 | 2023 | ||
@@ -2388,7 +2389,9 @@ detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, | |||
2388 | } else | 2389 | } else |
2389 | mm->highest_vm_end = prev ? prev->vm_end : 0; | 2390 | mm->highest_vm_end = prev ? prev->vm_end : 0; |
2390 | tail_vma->vm_next = NULL; | 2391 | tail_vma->vm_next = NULL; |
2391 | mm->mmap_cache = NULL; /* Kill the cache. */ | 2392 | |
2393 | /* Kill the cache */ | ||
2394 | vmacache_invalidate(mm); | ||
2392 | } | 2395 | } |
2393 | 2396 | ||
2394 | /* | 2397 | /* |
diff --git a/mm/nommu.c b/mm/nommu.c index e19482533ce3..5d3f3524bbdc 100644 --- a/mm/nommu.c +++ b/mm/nommu.c | |||
@@ -15,6 +15,7 @@ | |||
15 | 15 | ||
16 | #include <linux/export.h> | 16 | #include <linux/export.h> |
17 | #include <linux/mm.h> | 17 | #include <linux/mm.h> |
18 | #include <linux/vmacache.h> | ||
18 | #include <linux/mman.h> | 19 | #include <linux/mman.h> |
19 | #include <linux/swap.h> | 20 | #include <linux/swap.h> |
20 | #include <linux/file.h> | 21 | #include <linux/file.h> |
@@ -768,16 +769,23 @@ static void add_vma_to_mm(struct mm_struct *mm, struct vm_area_struct *vma) | |||
768 | */ | 769 | */ |
769 | static void delete_vma_from_mm(struct vm_area_struct *vma) | 770 | static void delete_vma_from_mm(struct vm_area_struct *vma) |
770 | { | 771 | { |
772 | int i; | ||
771 | struct address_space *mapping; | 773 | struct address_space *mapping; |
772 | struct mm_struct *mm = vma->vm_mm; | 774 | struct mm_struct *mm = vma->vm_mm; |
775 | struct task_struct *curr = current; | ||
773 | 776 | ||
774 | kenter("%p", vma); | 777 | kenter("%p", vma); |
775 | 778 | ||
776 | protect_vma(vma, 0); | 779 | protect_vma(vma, 0); |
777 | 780 | ||
778 | mm->map_count--; | 781 | mm->map_count--; |
779 | if (mm->mmap_cache == vma) | 782 | for (i = 0; i < VMACACHE_SIZE; i++) { |
780 | mm->mmap_cache = NULL; | 783 | /* if the vma is cached, invalidate the entire cache */ |
784 | if (curr->vmacache[i] == vma) { | ||
785 | vmacache_invalidate(curr->mm); | ||
786 | break; | ||
787 | } | ||
788 | } | ||
781 | 789 | ||
782 | /* remove the VMA from the mapping */ | 790 | /* remove the VMA from the mapping */ |
783 | if (vma->vm_file) { | 791 | if (vma->vm_file) { |
@@ -825,8 +833,8 @@ struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) | |||
825 | struct vm_area_struct *vma; | 833 | struct vm_area_struct *vma; |
826 | 834 | ||
827 | /* check the cache first */ | 835 | /* check the cache first */ |
828 | vma = ACCESS_ONCE(mm->mmap_cache); | 836 | vma = vmacache_find(mm, addr); |
829 | if (vma && vma->vm_start <= addr && vma->vm_end > addr) | 837 | if (likely(vma)) |
830 | return vma; | 838 | return vma; |
831 | 839 | ||
832 | /* trawl the list (there may be multiple mappings in which addr | 840 | /* trawl the list (there may be multiple mappings in which addr |
@@ -835,7 +843,7 @@ struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) | |||
835 | if (vma->vm_start > addr) | 843 | if (vma->vm_start > addr) |
836 | return NULL; | 844 | return NULL; |
837 | if (vma->vm_end > addr) { | 845 | if (vma->vm_end > addr) { |
838 | mm->mmap_cache = vma; | 846 | vmacache_update(addr, vma); |
839 | return vma; | 847 | return vma; |
840 | } | 848 | } |
841 | } | 849 | } |
@@ -874,8 +882,8 @@ static struct vm_area_struct *find_vma_exact(struct mm_struct *mm, | |||
874 | unsigned long end = addr + len; | 882 | unsigned long end = addr + len; |
875 | 883 | ||
876 | /* check the cache first */ | 884 | /* check the cache first */ |
877 | vma = mm->mmap_cache; | 885 | vma = vmacache_find_exact(mm, addr, end); |
878 | if (vma && vma->vm_start == addr && vma->vm_end == end) | 886 | if (vma) |
879 | return vma; | 887 | return vma; |
880 | 888 | ||
881 | /* trawl the list (there may be multiple mappings in which addr | 889 | /* trawl the list (there may be multiple mappings in which addr |
@@ -886,7 +894,7 @@ static struct vm_area_struct *find_vma_exact(struct mm_struct *mm, | |||
886 | if (vma->vm_start > addr) | 894 | if (vma->vm_start > addr) |
887 | return NULL; | 895 | return NULL; |
888 | if (vma->vm_end == end) { | 896 | if (vma->vm_end == end) { |
889 | mm->mmap_cache = vma; | 897 | vmacache_update(addr, vma); |
890 | return vma; | 898 | return vma; |
891 | } | 899 | } |
892 | } | 900 | } |
diff --git a/mm/vmacache.c b/mm/vmacache.c new file mode 100644 index 000000000000..d4224b397c0e --- /dev/null +++ b/mm/vmacache.c | |||
@@ -0,0 +1,112 @@ | |||
1 | /* | ||
2 | * Copyright (C) 2014 Davidlohr Bueso. | ||
3 | */ | ||
4 | #include <linux/sched.h> | ||
5 | #include <linux/mm.h> | ||
6 | #include <linux/vmacache.h> | ||
7 | |||
8 | /* | ||
9 | * Flush vma caches for threads that share a given mm. | ||
10 | * | ||
11 | * The operation is safe because the caller holds the mmap_sem | ||
12 | * exclusively and other threads accessing the vma cache will | ||
13 | * have mmap_sem held at least for read, so no extra locking | ||
14 | * is required to maintain the vma cache. | ||
15 | */ | ||
16 | void vmacache_flush_all(struct mm_struct *mm) | ||
17 | { | ||
18 | struct task_struct *g, *p; | ||
19 | |||
20 | rcu_read_lock(); | ||
21 | for_each_process_thread(g, p) { | ||
22 | /* | ||
23 | * Only flush the vmacache pointers as the | ||
24 | * mm seqnum is already set and curr's will | ||
25 | * be set upon invalidation when the next | ||
26 | * lookup is done. | ||
27 | */ | ||
28 | if (mm == p->mm) | ||
29 | vmacache_flush(p); | ||
30 | } | ||
31 | rcu_read_unlock(); | ||
32 | } | ||
33 | |||
34 | /* | ||
35 | * This task may be accessing a foreign mm via (for example) | ||
36 | * get_user_pages()->find_vma(). The vmacache is task-local and this | ||
37 | * task's vmacache pertains to a different mm (ie, its own). There is | ||
38 | * nothing we can do here. | ||
39 | * | ||
40 | * Also handle the case where a kernel thread has adopted this mm via use_mm(). | ||
41 | * That kernel thread's vmacache is not applicable to this mm. | ||
42 | */ | ||
43 | static bool vmacache_valid_mm(struct mm_struct *mm) | ||
44 | { | ||
45 | return current->mm == mm && !(current->flags & PF_KTHREAD); | ||
46 | } | ||
47 | |||
48 | void vmacache_update(unsigned long addr, struct vm_area_struct *newvma) | ||
49 | { | ||
50 | if (vmacache_valid_mm(newvma->vm_mm)) | ||
51 | current->vmacache[VMACACHE_HASH(addr)] = newvma; | ||
52 | } | ||
53 | |||
54 | static bool vmacache_valid(struct mm_struct *mm) | ||
55 | { | ||
56 | struct task_struct *curr; | ||
57 | |||
58 | if (!vmacache_valid_mm(mm)) | ||
59 | return false; | ||
60 | |||
61 | curr = current; | ||
62 | if (mm->vmacache_seqnum != curr->vmacache_seqnum) { | ||
63 | /* | ||
64 | * First attempt will always be invalid, initialize | ||
65 | * the new cache for this task here. | ||
66 | */ | ||
67 | curr->vmacache_seqnum = mm->vmacache_seqnum; | ||
68 | vmacache_flush(curr); | ||
69 | return false; | ||
70 | } | ||
71 | return true; | ||
72 | } | ||
73 | |||
74 | struct vm_area_struct *vmacache_find(struct mm_struct *mm, unsigned long addr) | ||
75 | { | ||
76 | int i; | ||
77 | |||
78 | if (!vmacache_valid(mm)) | ||
79 | return NULL; | ||
80 | |||
81 | for (i = 0; i < VMACACHE_SIZE; i++) { | ||
82 | struct vm_area_struct *vma = current->vmacache[i]; | ||
83 | |||
84 | if (vma && vma->vm_start <= addr && vma->vm_end > addr) { | ||
85 | BUG_ON(vma->vm_mm != mm); | ||
86 | return vma; | ||
87 | } | ||
88 | } | ||
89 | |||
90 | return NULL; | ||
91 | } | ||
92 | |||
93 | #ifndef CONFIG_MMU | ||
94 | struct vm_area_struct *vmacache_find_exact(struct mm_struct *mm, | ||
95 | unsigned long start, | ||
96 | unsigned long end) | ||
97 | { | ||
98 | int i; | ||
99 | |||
100 | if (!vmacache_valid(mm)) | ||
101 | return NULL; | ||
102 | |||
103 | for (i = 0; i < VMACACHE_SIZE; i++) { | ||
104 | struct vm_area_struct *vma = current->vmacache[i]; | ||
105 | |||
106 | if (vma && vma->vm_start == start && vma->vm_end == end) | ||
107 | return vma; | ||
108 | } | ||
109 | |||
110 | return NULL; | ||
111 | } | ||
112 | #endif | ||