On Thu, Mar 11, 2010 at 01:04:33PM +0800, Miao Xie wrote: > on 2010-3-10 3:42, Paul Menage wrote: > > On Sat, Mar 6, 2010 at 6:33 PM, Miao Xie <miaox@xxxxxxxxxxxxxx> wrote: > >> > >> Before applying this patch, cpuset updates task->mems_allowed just like > >> what you said. But the allocator is still likely to see an empty nodemask. > >> This problem have been pointed out by Nick Piggin. > >> > >> The problem is following: > >> The size of nodemask_t is greater than the size of long integer, so loading > >> and storing of nodemask_t are not atomic operations. If task->mems_allowed > >> don't intersect with new_mask, such as the first word of the mask is empty > >> and only the first word of new_mask is not empty. When the allocator > >> loads a word of the mask before > >> > >> current->mems_allowed |= new_mask; > >> > >> and then loads another word of the mask after > >> > >> current->mems_allowed = new_mask; > >> > >> the allocator gets an empty nodemask. > > > > Couldn't that be solved by having the reader read the nodemask twice > > and compare them? In the normal case there's no race, so the second > > read is straight from L1 cache and is very cheap. In the unlikely case > > of a race, the reader would keep trying until it got two consistent > > values in a row. > > I think this method can't fix the problem because we can guarantee the second > read is after the update of mask completes. Any problem with using a seqlock? The other thing you could do is store a pointer to the nodemask, and allocate a new nodemask when changing it, issue a smp_wmb(), and then store the new pointer. Read side only needs a smp_read_barrier_depends() -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>