On Wed, 2009-10-14 at 14:33 -0700, Frans Pop wrote: > On Wednesday 14 October 2009, reinette chatre wrote: > > We do queue the GFP_KERNEL allocations when there are only a few buffers > > remaining in the queue (8 right now) ... > > Are you sure of this? I have zero messages in my logs about allocation > failures with GFP_KERNEL, but I do have plenty with "Only 0 free buffers > remaining" with GFP_ATOMIC. That does make sense to me. We do not expect allocations with GFP_KERNEL to fail. Considering how I understand how things work I am considering the following scenario: * start with system low on available memory * now introduce incoming traffic (causing the RX code to run) * upon receipt of frame we attempt an allocation (to reclaim the buffer) with GFP_ATOMIC (state: num RX buffer free > watermark) * this fails since memory is not available * num RX buffer free reduces * does _not_ queue replenishment of buffers with GFP_KERNEL * repeat above until we hit the watermark (currently 8) * upon receipt of frame we attempt an allocation (to reclaim the buffer) with GFP_ATOMIC (state: num RX buffer free <= watermark) * this fails (now user sees big warning) * queue replenishment of buffers with GFP_KERNEL Essentially what I suspect could happen is that we do attempt to replenish the buffers with GFP_KERNEL after several failures with GFP_ATOMIC, but at that point we have already run out completely. One way to test this theory is to queue the GFP_KERNEL allocation earlier (when we still have a significant number of RX buffers available), 8 may turn out to be too small. > Does that indicate a bug or could they fall under the ratelimit somehow? In your kernel log I do see that the driver's error messages related to GFP_ATOMIC are rate limited (we see many more "order-2 allocation failure" messages than the "Failed to allocate" messages). All of these allocation failures are from the "replenish_now" code though, which is GFP_ATOMIC. So even though we do not see the "Failed to allocate" errors (which are rate limited) it seems that all allocation failures are from that (the GFP_ATOMIC) code. Reinette -- To unsubscribe from this list: send the line "unsubscribe kernel-testers" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
