Messages in this thread | | | Date | Thu, 9 Jun 2011 12:42:26 -0700 (PDT) | From | David Rientjes <> | Subject | Re: [PATCH v2] trace: Set __GFP_NORETRY flag for ring buffer allocating process |
| |
On Thu, 9 Jun 2011, Steven Rostedt wrote:
> > Unfortunately, __GFP_NORETRY is racy and don't work as expected. > > If free memory is not enough, the thread may start to reclaim and > > another thread can steal the reclaimed memory. And thread0 don't retry. > > > > Then, thread0's alloc page may fail even though system have enough reclaimable > > memory. > > > > thread0 thread1 > > --------------------------------------------------------------- > > alloc_pages() > > get_page_from_freelist() -> fail > > try_to_free_pages() > > alloc_pages() > > get_page_from_freelist() -> success > > get_page_from_freelist() -> fail again > > > > I think this is mm issue, and afaik, Minchan and some developers are > > working on fixing it. but _now_ your patch doesn't work. > > Have you seen this fail in practice? > > I'm not too concern if it only triggers when memory is tight. But if it > is triggering on normal cases, then that worries me. >
It would only happen if there was an antagonist that stole the reclaimed pages before your __GFP_NORETRY allocation could allocate them, resulting in the system being oom again as it was before reclaim occurred. Without __GFP_NORETRY, we'd automatically retry these allocations in a loop until we found the memory since they are order-0, so the only side effect would be an increased latency in the allocation. I think if we still end up oom after reclaiming memory that was allocated by another thread that we probably don't want to be expanding the ring buffer and, thus, I see no problem with just failing.
| |