on 2010-3-11 13:30, Nick Piggin wrote: >>>> 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() Comparing with my second version patch, I think both of these methods will cause worse performance and the changing of code is more. Thanks Miao -- 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>