On Mon, 2012-04-09 at 13:58 -0500, Larry Finger wrote:

As it happens with both iwlwifi and e1000e, it seems to be a problem further up 
the food chain.

I don't know much about iwlwifi, but loading it with the module parameter 
"amsdu_size_8K=0" seems to select 4K rather than 8K buffers. That will hurt 
performance, but it should fix the memory fragmentation. There have also been 
some problems with aggregation that are fixed by setting the option "11n_disable=3".

I think Marc posted stack traces showing problem on transmit side.

09:44:12   [<ffffffff810ba394>] ? __alloc_pages_nodemask+0x6b2/0x726
09:44:12   [<ffffffff810e9081>] ? kmem_getpages+0x4c/0xd9
09:44:12   [<ffffffff810e9081>] ? kmem_getpages+0x4c/0xd9
09:44:12   [<ffffffff810ea421>] ? fallback_alloc+0x123/0x1c2
09:44:12   [<ffffffff8127daae>] ? pskb_expand_head+0xe0/0x24a
09:44:12   [<ffffffff810ea923>] ? __kmalloc+0xba/0x112
09:44:12   [<ffffffff8127daae>] ? pskb_expand_head+0xe0/0x24a
09:44:12   [<ffffffffa07a81c0>] ? ieee80211_skb_resize+0x64/0x9d [mac80211]
09:44:12   [<ffffffffa07aa087>] ? ieee80211_subif_start_xmit+0x68e/0x80c [mac80211]
09:44:12   [<ffffffffa07922c8>] ? ieee80211_tx_status_irqsafe+0x2e/0x7f [mac80211]
09:44:12   [<ffffffff812877f3>] ? dev_hard_start_xmit+0x3fc/0x543
09:44:12   [<ffffffff81070551>] ? arch_local_irq_save+0x11/0x17
09:44:12   [<ffffffff8129da01>] ? sch_direct_xmit+0x5e/0x12f
09:44:12   [<ffffffff8129dbc9>] ? __qdisc_run+0xf7/0x10f

I dont really understand how it can happen, with MTU=1500