Messages in this thread | | | Date | Tue, 25 Oct 2011 15:12:26 -0700 (PDT) | From | David Rientjes <> | Subject | Re: [PATCH] mm: avoid livelock on !__GFP_FS allocations |
| |
On Tue, 25 Oct 2011, Colin Cross wrote:
> GFP_KERNEL is __GFP_WAIT | __GFP_IO | __GFP_FS. Once driver suspend > has started, gfp_allowed_mask is ~(__GFP_IO | GFP_FS), so any call to > __alloc_pages_nodemask(GFP_KERNEL, ...) gets masked to effectively > __alloc_pages_nodemask(__GFP_WAIT, ...). >
Just passing __GFP_WAIT is the problem that you're trying to address, though. Why not include __GFP_NORETRY since you know the liklihood of allocation being successful on the second iteration is very slim since you're not in a context where you can force reclaim or oom killing?
> The loop is in __alloc_pages_slowpath, from the rebalance label to > should_alloc_retry.
The loop is by design and is activated because you're just passing __GFP_WAIT in this context for no sensible reason.
| |