Re: Regression since 2.6.30-rc1 - Page allocation failures

[Johannes Berg]

Hi Larry,

> I have been getting a number of page allocation failures of late on my x86_64
> system. The user-space app that hits the problem varies; however, there is
> always a network connection. In addition, the failure is always of order 1. For
> the instance reported below, git was updating a tree mounted on an NFS volume.
> The problem started with 2.6.30-rc1, but persists with -rc3.
> 
> I have seen the problem with the b43 and p54usb wireless drivers. The user-space
> apps include cc1, git components, and firefox; thus not all involve nfs.
> 
> I will be happy to supply any further information.
> 
> kernel: git: page allocation failure. order:1, mode:0x4020

> kernel:  [<ffffffffa0220711>] ieee80211_skb_resize+0x91/0xc7 [mac80211]
...
> kernel: 427298 pages non-shared

It's a little curious that you have order:1 allocations here. I think
below the printout you showed you would see a line like
	phy0: failed to reallocate TX buffer
Is that there? Could you track down that message in the code (around
net/mac80211/tx.c line 1368) and add to it the following:

	"%d, %d, need %d, %d", skb->data_len, skb->len, head_need, tail_need

and then try to reproduce this problem please? With both p54 and b43 you
should only have packets far smaller than 4096 bytes so no order 1
allocation should be needed, afaict.

johannes

Attachment: signature.asc
Description: This is a digitally signed message part