mm: fix Committed_AS underflow on large NR_CPUS environment
authorKOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Thu, 30 Apr 2009 22:08:51 +0000 (15:08 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 2 May 2009 22:36:10 +0000 (15:36 -0700)
The Committed_AS field can underflow in certain situations:

>         # while true; do cat /proc/meminfo  | grep _AS; sleep 1; done | uniq -c
>               1 Committed_AS: 18446744073709323392 kB
>              11 Committed_AS: 18446744073709455488 kB
>               6 Committed_AS:    35136 kB
>               5 Committed_AS: 18446744073709454400 kB
>               7 Committed_AS:    35904 kB
>               3 Committed_AS: 18446744073709453248 kB
>               2 Committed_AS:    34752 kB
>               9 Committed_AS: 18446744073709453248 kB
>               8 Committed_AS:    34752 kB
>               3 Committed_AS: 18446744073709320960 kB
>               7 Committed_AS: 18446744073709454080 kB
>               3 Committed_AS: 18446744073709320960 kB
>               5 Committed_AS: 18446744073709454080 kB
>               6 Committed_AS: 18446744073709320960 kB

Because NR_CPUS can be greater than 1000 and meminfo_proc_show() does
not check for underflow.

But NR_CPUS proportional isn't good calculation.  In general,
possibility of lock contention is proportional to the number of online
cpus, not theorical maximum cpus (NR_CPUS).

The current kernel has generic percpu-counter stuff.  using it is right
way.  it makes code simplify and percpu_counter_read_positive() don't
make underflow issue.

Reported-by: Dave Hansen <dave@linux.vnet.ibm.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Eric B Munson <ebmunson@us.ibm.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: <stable@kernel.org> [All kernel versions]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/proc/meminfo.c
include/linux/mman.h
mm/mmap.c
mm/nommu.c
mm/swap.c

index 74ea974f5ca6d86bd249ee920527be9c95157522..c6b0302af4c40268e9511e34d16f3385f817e1d6 100644 (file)
@@ -35,7 +35,7 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
 #define K(x) ((x) << (PAGE_SHIFT - 10))
        si_meminfo(&i);
        si_swapinfo(&i);
-       committed = atomic_long_read(&vm_committed_space);
+       committed = percpu_counter_read_positive(&vm_committed_as);
        allowed = ((totalram_pages - hugetlb_total_pages())
                * sysctl_overcommit_ratio / 100) + total_swap_pages;
 
index 30d1073bac3ba63459123db829283fa9ec7eb8e5..9872d6ca58ae9a7b88dcc97387f86a434a633db5 100644 (file)
 
 #ifdef __KERNEL__
 #include <linux/mm.h>
+#include <linux/percpu_counter.h>
 
 #include <asm/atomic.h>
 
 extern int sysctl_overcommit_memory;
 extern int sysctl_overcommit_ratio;
-extern atomic_long_t vm_committed_space;
+extern struct percpu_counter vm_committed_as;
 
-#ifdef CONFIG_SMP
-extern void vm_acct_memory(long pages);
-#else
 static inline void vm_acct_memory(long pages)
 {
-       atomic_long_add(pages, &vm_committed_space);
+       percpu_counter_add(&vm_committed_as, pages);
 }
-#endif
 
 static inline void vm_unacct_memory(long pages)
 {
index 3303d1ba8e87aca1ba34ca74f92d0cd9038de396..6b7b1a95944bf267bd21cd981703280fe25a39e9 100644 (file)
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -85,7 +85,7 @@ EXPORT_SYMBOL(vm_get_page_prot);
 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
 int sysctl_overcommit_ratio = 50;      /* default is 50% */
 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
-atomic_long_t vm_committed_space = ATOMIC_LONG_INIT(0);
+struct percpu_counter vm_committed_as;
 
 /*
  * Check that a process has enough memory to allocate a new virtual
@@ -179,11 +179,7 @@ int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
        if (mm)
                allowed -= mm->total_vm / 32;
 
-       /*
-        * cast `allowed' as a signed long because vm_committed_space
-        * sometimes has a negative value
-        */
-       if (atomic_long_read(&vm_committed_space) < (long)allowed)
+       if (percpu_counter_read_positive(&vm_committed_as) < allowed)
                return 0;
 error:
        vm_unacct_memory(pages);
@@ -2481,4 +2477,8 @@ void mm_drop_all_locks(struct mm_struct *mm)
  */
 void __init mmap_init(void)
 {
+       int ret;
+
+       ret = percpu_counter_init(&vm_committed_as, 0);
+       VM_BUG_ON(ret);
 }
index 72eda4aee2cb46d72e070271d82ae13daef78b36..809998aa7b509dd072862def05f420a14e1185cf 100644 (file)
@@ -62,7 +62,7 @@ void *high_memory;
 struct page *mem_map;
 unsigned long max_mapnr;
 unsigned long num_physpages;
-atomic_long_t vm_committed_space = ATOMIC_LONG_INIT(0);
+struct percpu_counter vm_committed_as;
 int sysctl_overcommit_memory = OVERCOMMIT_GUESS; /* heuristic overcommit */
 int sysctl_overcommit_ratio = 50; /* default is 50% */
 int sysctl_max_map_count = DEFAULT_MAX_MAP_COUNT;
@@ -463,6 +463,10 @@ SYSCALL_DEFINE1(brk, unsigned long, brk)
  */
 void __init mmap_init(void)
 {
+       int ret;
+
+       ret = percpu_counter_init(&vm_committed_as, 0);
+       VM_BUG_ON(ret);
        vm_region_jar = KMEM_CACHE(vm_region, SLAB_PANIC);
 }
 
@@ -1847,12 +1851,9 @@ int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
        if (mm)
                allowed -= mm->total_vm / 32;
 
-       /*
-        * cast `allowed' as a signed long because vm_committed_space
-        * sometimes has a negative value
-        */
-       if (atomic_long_read(&vm_committed_space) < (long)allowed)
+       if (percpu_counter_read_positive(&vm_committed_as) < allowed)
                return 0;
+
 error:
        vm_unacct_memory(pages);
 
index bede23ce64ea28e52d991168f2c93cd2243c893a..cb29ae5d33abfce703991cbfebd7bc1956752942 100644 (file)
--- a/mm/swap.c
+++ b/mm/swap.c
@@ -491,49 +491,6 @@ unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 
 EXPORT_SYMBOL(pagevec_lookup_tag);
 
-#ifdef CONFIG_SMP
-/*
- * We tolerate a little inaccuracy to avoid ping-ponging the counter between
- * CPUs
- */
-#define ACCT_THRESHOLD max(16, NR_CPUS * 2)
-
-static DEFINE_PER_CPU(long, committed_space);
-
-void vm_acct_memory(long pages)
-{
-       long *local;
-
-       preempt_disable();
-       local = &__get_cpu_var(committed_space);
-       *local += pages;
-       if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
-               atomic_long_add(*local, &vm_committed_space);
-               *local = 0;
-       }
-       preempt_enable();
-}
-
-#ifdef CONFIG_HOTPLUG_CPU
-
-/* Drop the CPU's cached committed space back into the central pool. */
-static int cpu_swap_callback(struct notifier_block *nfb,
-                            unsigned long action,
-                            void *hcpu)
-{
-       long *committed;
-
-       committed = &per_cpu(committed_space, (long)hcpu);
-       if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
-               atomic_long_add(*committed, &vm_committed_space);
-               *committed = 0;
-               drain_cpu_pagevecs((long)hcpu);
-       }
-       return NOTIFY_OK;
-}
-#endif /* CONFIG_HOTPLUG_CPU */
-#endif /* CONFIG_SMP */
-
 /*
  * Perform any setup for the swap system
  */
@@ -554,7 +511,4 @@ void __init swap_setup(void)
         * Right now other parts of the system means that we
         * _really_ don't want to cluster much more
         */
-#ifdef CONFIG_HOTPLUG_CPU
-       hotcpu_notifier(cpu_swap_callback, 0);
-#endif
 }