> On Mon, 17 Apr 2023 20:17:45 +0200 Lorenzo Bianconi wrote: > > > I do not see why this would be different than having buffers sitting > > > in some tcp receive > > > (or out or order) queues for a few minutes ? > > > > The main issue in my tests (and even in mt76 I think) is the pages are not returned > > to the pool for a very long time (even hours) and doing so the pool is like in a > > 'limbo' state where it is not actually deallocated and page_pool_release_retry > > continues complaining about it. I think this is because we do not have more tcp > > traffic to deallocate them, but I am not so familiar with tcp codebase :) > > I've seen the page leaks too in my recent testing but I just assumed > I fumbled the quick bnxt conversion to page pool. Could it be something > with page frags? It happened a lot if I used page frags, IIRC mt76 is > using page frags, too. my device is allocating a full order 0 page from the pool so it is not related to fragmented pages as fixed for mt76 by Alex and Felix. > > Is drgn available for your target? You could try to scan the pages on > the system and see if you can find what's still pointing to the page > pool (assuming they are indeed leaked and not returned to the page > allocator without releasing :() I will test it but since setting sysctl_skb_defer_max to 0 fixes the issue, I think the pages are still properly linked to the pool, they are just not returned to it. I proved it using the other patch I posted [0] where I can see the counter of returned pages incrementing from time to time (in a very long time slot..). Unrelated to this issue, but debugging it I think a found a page_pool leak in skb_condense() [1] where we can reallocate the skb data using kmalloc for a page_pool recycled skb. Regards, Lorenzo [0] https://lore.kernel.org/netdev/20230417111204.08f19827@xxxxxxxxxx/T/#t [1] https://github.com/torvalds/linux/blob/master/net/core/skbuff.c#L6602
Attachment:
signature.asc
Description: PGP signature
