Messages in this thread | | | Date | Wed, 30 Apr 2008 14:09:36 -0700 | From | Andrew Morton <> | Subject | Re: [PATCH 1/2] mm: Fix overcommit overflow |
| |
On Wed, 30 Apr 2008 21:42:41 +0100 Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:
> Sami Farin reported an overflow in the overcommit handling on 64bit boxes. > > We use atomic_t for page counting which is fine on 32bit but on 64bit > overflows. We can use atomic64_t but there is a problem - most heavy > users of zero overcommit are embedded people whose 64bit atomics are > really slow and expensive operations. Thus we use a few defines to flip > 32 or 64bit according to the size of a long > > (Split into two diffs to keep ownership correct) > > Signed-off-by: Alan Cox <alan@redhat.com> > > diff -u --new-file --recursive --exclude-from /usr/src/exclude linux.vanilla-2.6.25-mm1/include/linux/mman.h linux-2.6.25-mm1/include/linux/mman.h > --- linux.vanilla-2.6.25-mm1/include/linux/mman.h 2008-04-28 11:35:21.000000000 +0100 > +++ linux-2.6.25-mm1/include/linux/mman.h 2008-04-30 11:21:10.000000000 +0100 > @@ -15,16 +15,31 @@ > > #include <asm/atomic.h> > > +/* 32bit platforms have a virtual address space in pages we > + can fit into an atomic_t. We want to avoid atomic64_t on > + such boxes as it is often expensive and most strict overcommit > + users turn out to be embedded low power processors */ > + > +#if (BITS_PER_LONG == 32) > +#define vm_atomic_t atomic_t > +#define vm_atomic_read atomic_read > +#define vm_atomic_add atomic_add > +#else > +#define vm_atomic_t atomic64_t > +#define vm_atomic_read atomic64_read > +#define vm_atomic_add atomic64_add > +#endif
ew.
> extern int sysctl_overcommit_memory; > extern int sysctl_overcommit_ratio; > -extern atomic_t vm_committed_space; > +extern vm_atomic_t vm_committed_space; > > #ifdef CONFIG_SMP > extern void vm_acct_memory(long pages); > #else > static inline void vm_acct_memory(long pages) > { > - atomic_add(pages, &vm_committed_space); > + vm_atomic_add(pages, &vm_committed_space); > } > #endif > > diff -u --new-file --recursive --exclude-from /usr/src/exclude linux.vanilla-2.6.25-mm1/mm/mmap.c linux-2.6.25-mm1/mm/mmap.c > --- linux.vanilla-2.6.25-mm1/mm/mmap.c 2008-04-28 11:36:52.000000000 +0100 > +++ linux-2.6.25-mm1/mm/mmap.c 2008-04-30 11:17:03.000000000 +0100 > @@ -80,7 +80,7 @@ > 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_t vm_committed_space = ATOMIC_INIT(0); > +vm_atomic_t vm_committed_space = ATOMIC_INIT(0);
That'll need to be ATOMIC64_INIT on 64-bit.
> > /* > * Check that a process has enough memory to allocate a new virtual > @@ -177,7 +177,7 @@ > * cast `allowed' as a signed long because vm_committed_space > * sometimes has a negative value > */ > - if (atomic_read(&vm_committed_space) < (long)allowed) > + if (vm_atomic_read(&vm_committed_space) < (long)allowed) > return 0; > error: > vm_unacct_memory(pages); > diff -u --new-file --recursive --exclude-from /usr/src/exclude linux.vanilla-2.6.25-mm1/mm/swap.c linux-2.6.25-mm1/mm/swap.c > --- linux.vanilla-2.6.25-mm1/mm/swap.c 2008-04-28 11:36:52.000000000 +0100 > +++ linux-2.6.25-mm1/mm/swap.c 2008-04-30 11:18:05.000000000 +0100 > @@ -503,7 +503,7 @@ > local = &__get_cpu_var(committed_space); > *local += pages; > if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) { > - atomic_add(*local, &vm_committed_space); > + vm_atomic_add(*local, &vm_committed_space); > *local = 0; > } > preempt_enable();
But afacit the existing atomic_long_t does exactly what you want?
| |