mm: per-thread vma caching
authorDavidlohr Bueso <davidlohr@hp.com>
Mon, 7 Apr 2014 22:37:25 +0000 (15:37 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Mon, 7 Apr 2014 23:35:53 +0000 (16:35 -0700)
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>
12 files changed:
arch/unicore32/include/asm/mmu_context.h
fs/exec.c
fs/proc/task_mmu.c
include/linux/mm_types.h
include/linux/sched.h
include/linux/vmacache.h [new file with mode: 0644]
kernel/debug/debug_core.c
kernel/fork.c
mm/Makefile
mm/mmap.c
mm/nommu.c
mm/vmacache.c [new file with mode: 0644]

index fb5e4c6..ef470a7 100644 (file)
@@ -14,6 +14,8 @@
 
 #include <linux/compiler.h>
 #include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/vmacache.h>
 #include <linux/io.h>
 
 #include <asm/cacheflush.h>
@@ -73,7 +75,7 @@ do { \
                else \
                        mm->mmap = NULL; \
                rb_erase(&high_vma->vm_rb, &mm->mm_rb); \
-               mm->mmap_cache = NULL; \
+               vmacache_invalidate(mm); \
                mm->map_count--; \
                remove_vma(high_vma); \
        } \
index 25dfeba..b60ccf9 100644 (file)
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -26,6 +26,7 @@
 #include <linux/file.h>
 #include <linux/fdtable.h>
 #include <linux/mm.h>
+#include <linux/vmacache.h>
 #include <linux/stat.h>
 #include <linux/fcntl.h>
 #include <linux/swap.h>
@@ -822,7 +823,7 @@ EXPORT_SYMBOL(read_code);
 static int exec_mmap(struct mm_struct *mm)
 {
        struct task_struct *tsk;
-       struct mm_struct * old_mm, *active_mm;
+       struct mm_struct *old_mm, *active_mm;
 
        /* Notify parent that we're no longer interested in the old VM */
        tsk = current;
@@ -848,6 +849,8 @@ static int exec_mmap(struct mm_struct *mm)
        tsk->mm = mm;
        tsk->active_mm = mm;
        activate_mm(active_mm, mm);
+       tsk->mm->vmacache_seqnum = 0;
+       vmacache_flush(tsk);
        task_unlock(tsk);
        if (old_mm) {
                up_read(&old_mm->mmap_sem);
index fb52b54..442177b 100644 (file)
@@ -1,4 +1,5 @@
 #include <linux/mm.h>
+#include <linux/vmacache.h>
 #include <linux/hugetlb.h>
 #include <linux/huge_mm.h>
 #include <linux/mount.h>
@@ -152,7 +153,7 @@ static void *m_start(struct seq_file *m, loff_t *pos)
 
        /*
         * We remember last_addr rather than next_addr to hit with
-        * mmap_cache most of the time. We have zero last_addr at
+        * vmacache most of the time. We have zero last_addr at
         * the beginning and also after lseek. We will have -1 last_addr
         * after the end of the vmas.
         */
index 290901a..2b58d19 100644 (file)
@@ -342,9 +342,9 @@ struct mm_rss_stat {
 
 struct kioctx_table;
 struct mm_struct {
-       struct vm_area_struct * mmap;           /* list of VMAs */
+       struct vm_area_struct *mmap;            /* list of VMAs */
        struct rb_root mm_rb;
-       struct vm_area_struct * mmap_cache;     /* last find_vma result */
+       u32 vmacache_seqnum;                   /* per-thread vmacache */
 #ifdef CONFIG_MMU
        unsigned long (*get_unmapped_area) (struct file *filp,
                                unsigned long addr, unsigned long len,
index 7cb07fd..642477d 100644 (file)
@@ -132,6 +132,10 @@ struct perf_event_context;
 struct blk_plug;
 struct filename;
 
+#define VMACACHE_BITS 2
+#define VMACACHE_SIZE (1U << VMACACHE_BITS)
+#define VMACACHE_MASK (VMACACHE_SIZE - 1)
+
 /*
  * List of flags we want to share for kernel threads,
  * if only because they are not used by them anyway.
@@ -1235,6 +1239,9 @@ struct task_struct {
 #ifdef CONFIG_COMPAT_BRK
        unsigned brk_randomized:1;
 #endif
+       /* per-thread vma caching */
+       u32 vmacache_seqnum;
+       struct vm_area_struct *vmacache[VMACACHE_SIZE];
 #if defined(SPLIT_RSS_COUNTING)
        struct task_rss_stat    rss_stat;
 #endif
diff --git a/include/linux/vmacache.h b/include/linux/vmacache.h
new file mode 100644 (file)
index 0000000..c3fa0fd
--- /dev/null
@@ -0,0 +1,38 @@
+#ifndef __LINUX_VMACACHE_H
+#define __LINUX_VMACACHE_H
+
+#include <linux/sched.h>
+#include <linux/mm.h>
+
+/*
+ * Hash based on the page number. Provides a good hit rate for
+ * workloads with good locality and those with random accesses as well.
+ */
+#define VMACACHE_HASH(addr) ((addr >> PAGE_SHIFT) & VMACACHE_MASK)
+
+static inline void vmacache_flush(struct task_struct *tsk)
+{
+       memset(tsk->vmacache, 0, sizeof(tsk->vmacache));
+}
+
+extern void vmacache_flush_all(struct mm_struct *mm);
+extern void vmacache_update(unsigned long addr, struct vm_area_struct *newvma);
+extern struct vm_area_struct *vmacache_find(struct mm_struct *mm,
+                                                   unsigned long addr);
+
+#ifndef CONFIG_MMU
+extern struct vm_area_struct *vmacache_find_exact(struct mm_struct *mm,
+                                                 unsigned long start,
+                                                 unsigned long end);
+#endif
+
+static inline void vmacache_invalidate(struct mm_struct *mm)
+{
+       mm->vmacache_seqnum++;
+
+       /* deal with overflows */
+       if (unlikely(mm->vmacache_seqnum == 0))
+               vmacache_flush_all(mm);
+}
+
+#endif /* __LINUX_VMACACHE_H */
index 99982a7..2956c8d 100644 (file)
@@ -49,6 +49,7 @@
 #include <linux/pid.h>
 #include <linux/smp.h>
 #include <linux/mm.h>
+#include <linux/vmacache.h>
 #include <linux/rcupdate.h>
 
 #include <asm/cacheflush.h>
@@ -224,10 +225,17 @@ static void kgdb_flush_swbreak_addr(unsigned long addr)
        if (!CACHE_FLUSH_IS_SAFE)
                return;
 
-       if (current->mm && current->mm->mmap_cache) {
-               flush_cache_range(current->mm->mmap_cache,
-                                 addr, addr + BREAK_INSTR_SIZE);
+       if (current->mm) {
+               int i;
+
+               for (i = 0; i < VMACACHE_SIZE; i++) {
+                       if (!current->vmacache[i])
+                               continue;
+                       flush_cache_range(current->vmacache[i],
+                                         addr, addr + BREAK_INSTR_SIZE);
+               }
        }
+
        /* Force flush instruction cache if it was outside the mm */
        flush_icache_range(addr, addr + BREAK_INSTR_SIZE);
 }
index e40c0a0..bc0e96b 100644 (file)
@@ -28,6 +28,8 @@
 #include <linux/mman.h>
 #include <linux/mmu_notifier.h>
 #include <linux/fs.h>
+#include <linux/mm.h>
+#include <linux/vmacache.h>
 #include <linux/nsproxy.h>
 #include <linux/capability.h>
 #include <linux/cpu.h>
@@ -364,7 +366,7 @@ static int dup_mmap(struct mm_struct *mm, struct mm_struct *oldmm)
 
        mm->locked_vm = 0;
        mm->mmap = NULL;
-       mm->mmap_cache = NULL;
+       mm->vmacache_seqnum = 0;
        mm->map_count = 0;
        cpumask_clear(mm_cpumask(mm));
        mm->mm_rb = RB_ROOT;
@@ -882,6 +884,9 @@ static int copy_mm(unsigned long clone_flags, struct task_struct *tsk)
        if (!oldmm)
                return 0;
 
+       /* initialize the new vmacache entries */
+       vmacache_flush(tsk);
+
        if (clone_flags & CLONE_VM) {
                atomic_inc(&oldmm->mm_users);
                mm = oldmm;
index cdd7415..23a6f7e 100644 (file)
@@ -16,7 +16,7 @@ obj-y                 := filemap.o mempool.o oom_kill.o fadvise.o \
                           readahead.o swap.o truncate.o vmscan.o shmem.o \
                           util.o mmzone.o vmstat.o backing-dev.o \
                           mm_init.o mmu_context.o percpu.o slab_common.o \
-                          compaction.o balloon_compaction.o \
+                          compaction.o balloon_compaction.o vmacache.o \
                           interval_tree.o list_lru.o workingset.o $(mmu-y)
 
 obj-y += init-mm.o
index 46433e1..b1202cf 100644 (file)
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -10,6 +10,7 @@
 #include <linux/slab.h>
 #include <linux/backing-dev.h>
 #include <linux/mm.h>
+#include <linux/vmacache.h>
 #include <linux/shm.h>
 #include <linux/mman.h>
 #include <linux/pagemap.h>
@@ -681,8 +682,9 @@ __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
        prev->vm_next = next = vma->vm_next;
        if (next)
                next->vm_prev = prev;
-       if (mm->mmap_cache == vma)
-               mm->mmap_cache = prev;
+
+       /* Kill the cache */
+       vmacache_invalidate(mm);
 }
 
 /*
@@ -1989,34 +1991,33 @@ EXPORT_SYMBOL(get_unmapped_area);
 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
 {
-       struct vm_area_struct *vma = NULL;
+       struct rb_node *rb_node;
+       struct vm_area_struct *vma;
 
        /* Check the cache first. */
-       /* (Cache hit rate is typically around 35%.) */
-       vma = ACCESS_ONCE(mm->mmap_cache);
-       if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
-               struct rb_node *rb_node;
+       vma = vmacache_find(mm, addr);
+       if (likely(vma))
+               return vma;
 
-               rb_node = mm->mm_rb.rb_node;
-               vma = NULL;
+       rb_node = mm->mm_rb.rb_node;
+       vma = NULL;
 
-               while (rb_node) {
-                       struct vm_area_struct *vma_tmp;
-
-                       vma_tmp = rb_entry(rb_node,
-                                          struct vm_area_struct, vm_rb);
-
-                       if (vma_tmp->vm_end > addr) {
-                               vma = vma_tmp;
-                               if (vma_tmp->vm_start <= addr)
-                                       break;
-                               rb_node = rb_node->rb_left;
-                       } else
-                               rb_node = rb_node->rb_right;
-               }
-               if (vma)
-                       mm->mmap_cache = vma;
+       while (rb_node) {
+               struct vm_area_struct *tmp;
+
+               tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
+
+               if (tmp->vm_end > addr) {
+                       vma = tmp;
+                       if (tmp->vm_start <= addr)
+                               break;
+                       rb_node = rb_node->rb_left;
+               } else
+                       rb_node = rb_node->rb_right;
        }
+
+       if (vma)
+               vmacache_update(addr, vma);
        return vma;
 }
 
@@ -2388,7 +2389,9 @@ detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
        } else
                mm->highest_vm_end = prev ? prev->vm_end : 0;
        tail_vma->vm_next = NULL;
-       mm->mmap_cache = NULL;          /* Kill the cache. */
+
+       /* Kill the cache */
+       vmacache_invalidate(mm);
 }
 
 /*
index e194825..5d3f352 100644 (file)
@@ -15,6 +15,7 @@
 
 #include <linux/export.h>
 #include <linux/mm.h>
+#include <linux/vmacache.h>
 #include <linux/mman.h>
 #include <linux/swap.h>
 #include <linux/file.h>
@@ -768,16 +769,23 @@ static void add_vma_to_mm(struct mm_struct *mm, struct vm_area_struct *vma)
  */
 static void delete_vma_from_mm(struct vm_area_struct *vma)
 {
+       int i;
        struct address_space *mapping;
        struct mm_struct *mm = vma->vm_mm;
+       struct task_struct *curr = current;
 
        kenter("%p", vma);
 
        protect_vma(vma, 0);
 
        mm->map_count--;
-       if (mm->mmap_cache == vma)
-               mm->mmap_cache = NULL;
+       for (i = 0; i < VMACACHE_SIZE; i++) {
+               /* if the vma is cached, invalidate the entire cache */
+               if (curr->vmacache[i] == vma) {
+                       vmacache_invalidate(curr->mm);
+                       break;
+               }
+       }
 
        /* remove the VMA from the mapping */
        if (vma->vm_file) {
@@ -825,8 +833,8 @@ struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
        struct vm_area_struct *vma;
 
        /* check the cache first */
-       vma = ACCESS_ONCE(mm->mmap_cache);
-       if (vma && vma->vm_start <= addr && vma->vm_end > addr)
+       vma = vmacache_find(mm, addr);
+       if (likely(vma))
                return vma;
 
        /* 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)
                if (vma->vm_start > addr)
                        return NULL;
                if (vma->vm_end > addr) {
-                       mm->mmap_cache = vma;
+                       vmacache_update(addr, vma);
                        return vma;
                }
        }
@@ -874,8 +882,8 @@ static struct vm_area_struct *find_vma_exact(struct mm_struct *mm,
        unsigned long end = addr + len;
 
        /* check the cache first */
-       vma = mm->mmap_cache;
-       if (vma && vma->vm_start == addr && vma->vm_end == end)
+       vma = vmacache_find_exact(mm, addr, end);
+       if (vma)
                return vma;
 
        /* 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,
                if (vma->vm_start > addr)
                        return NULL;
                if (vma->vm_end == end) {
-                       mm->mmap_cache = vma;
+                       vmacache_update(addr, vma);
                        return vma;
                }
        }
diff --git a/mm/vmacache.c b/mm/vmacache.c
new file mode 100644 (file)
index 0000000..d4224b3
--- /dev/null
@@ -0,0 +1,112 @@
+/*
+ * Copyright (C) 2014 Davidlohr Bueso.
+ */
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/vmacache.h>
+
+/*
+ * Flush vma caches for threads that share a given mm.
+ *
+ * The operation is safe because the caller holds the mmap_sem
+ * exclusively and other threads accessing the vma cache will
+ * have mmap_sem held at least for read, so no extra locking
+ * is required to maintain the vma cache.
+ */
+void vmacache_flush_all(struct mm_struct *mm)
+{
+       struct task_struct *g, *p;
+
+       rcu_read_lock();
+       for_each_process_thread(g, p) {
+               /*
+                * Only flush the vmacache pointers as the
+                * mm seqnum is already set and curr's will
+                * be set upon invalidation when the next
+                * lookup is done.
+                */
+               if (mm == p->mm)
+                       vmacache_flush(p);
+       }
+       rcu_read_unlock();
+}
+
+/*
+ * This task may be accessing a foreign mm via (for example)
+ * get_user_pages()->find_vma().  The vmacache is task-local and this
+ * task's vmacache pertains to a different mm (ie, its own).  There is
+ * nothing we can do here.
+ *
+ * Also handle the case where a kernel thread has adopted this mm via use_mm().
+ * That kernel thread's vmacache is not applicable to this mm.
+ */
+static bool vmacache_valid_mm(struct mm_struct *mm)
+{
+       return current->mm == mm && !(current->flags & PF_KTHREAD);
+}
+
+void vmacache_update(unsigned long addr, struct vm_area_struct *newvma)
+{
+       if (vmacache_valid_mm(newvma->vm_mm))
+               current->vmacache[VMACACHE_HASH(addr)] = newvma;
+}
+
+static bool vmacache_valid(struct mm_struct *mm)
+{
+       struct task_struct *curr;
+
+       if (!vmacache_valid_mm(mm))
+               return false;
+
+       curr = current;
+       if (mm->vmacache_seqnum != curr->vmacache_seqnum) {
+               /*
+                * First attempt will always be invalid, initialize
+                * the new cache for this task here.
+                */
+               curr->vmacache_seqnum = mm->vmacache_seqnum;
+               vmacache_flush(curr);
+               return false;
+       }
+       return true;
+}
+
+struct vm_area_struct *vmacache_find(struct mm_struct *mm, unsigned long addr)
+{
+       int i;
+
+       if (!vmacache_valid(mm))
+               return NULL;
+
+       for (i = 0; i < VMACACHE_SIZE; i++) {
+               struct vm_area_struct *vma = current->vmacache[i];
+
+               if (vma && vma->vm_start <= addr && vma->vm_end > addr) {
+                       BUG_ON(vma->vm_mm != mm);
+                       return vma;
+               }
+       }
+
+       return NULL;
+}
+
+#ifndef CONFIG_MMU
+struct vm_area_struct *vmacache_find_exact(struct mm_struct *mm,
+                                          unsigned long start,
+                                          unsigned long end)
+{
+       int i;
+
+       if (!vmacache_valid(mm))
+               return NULL;
+
+       for (i = 0; i < VMACACHE_SIZE; i++) {
+               struct vm_area_struct *vma = current->vmacache[i];
+
+               if (vma && vma->vm_start == start && vma->vm_end == end)
+                       return vma;
+       }
+
+       return NULL;
+}
+#endif