"Close NETIF_F_LLTX race conditions" problem

"Close NETIF_F_LLTX race conditions" problem

[OGAWA Hirofumi]

Hi,

I got the following warning. The cause of this warning is the context
of ->hard_start_xmit() was changed. Before that patch the IRQ was
enableing, but now it was disabled.

With this change, the assumption of some drivers was changed, so now
some drivers need to change the ->hard_start_xmit().
(e.g, dev_kfree_skb() -> dev_kfree_skb_irq(), spin_lock_irq() -> spin_lock()).

Umm.. Should we fix the all drivers? Any idea?

Thanks.

Badness in local_bh_enable at kernel/softirq.c:140
 [<c0123139>] local_bh_enable+0x79/0x80
 [<e8a7872a>] svc_write_space+0x2a/0xa0 [sunrpc]
 [<c024e38a>] sock_wfree+0x4a/0x50
 [<c024f913>] __kfree_skb+0x63/0x140
 [<e899841c>] rtl8139_start_xmit+0x7c/0x160 [8139too]
 [<c0262c2d>] qdisc_restart+0x14d/0x1c0
 [<c02556ff>] dev_queue_xmit+0x1cf/0x270
 [<c026dffd>] ip_finish_output+0xcd/0x210
 [<c02950dd>] inet_sendmsg+0x4d/0x60
 [<c02701e4>] ip_push_pending_frames+0x2c4/0x460
 [<c028cd6c>] udp_push_pending_frames+0x16c/0x2b0
 [<c028d6ad>] udp_sendpage+0xdd/0x150
 [<c012d2a8>] set_current_groups+0x78/0x80
 [<c029518e>] inet_sendpage+0x9e/0xd0
 [<e8a78329>] svc_sendto+0xa9/0x2a0 [sunrpc]
 [<c024f8a4>] kfree_skbmem+0x24/0x30
 [<e8a78a5e>] svc_udp_sendto+0x1e/0x40 [sunrpc]
 [<e8a79d1b>] svc_send+0xcb/0x110 [sunrpc]
 [<e8a778e0>] svc_process+0x220/0x6b0 [sunrpc]
 [<c01193e0>] default_wake_function+0x0/0x20
 [<e8b4941b>] nfsd+0x1cb/0x350 [nfsd]
 [<e8b49250>] nfsd+0x0/0x350 [nfsd]
 [<c0101375>] kernel_thread_helper+0x5/0x10
-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>

Re: "Close NETIF_F_LLTX race conditions" problem

[Eric Lemoine]

On Sun, 23 Jan 2005 21:19:04 +0900, OGAWA Hirofumi
<hirofumi@mail.parknet.co.jp> wrote:
> Hi,
> 
> I got the following warning. The cause of this warning is the context
> of ->hard_start_xmit() was changed. Before that patch the IRQ was
> enableing, but now it was disabled.
> 
> With this change, the assumption of some drivers was changed, so now
> some drivers need to change the ->hard_start_xmit().
> (e.g, dev_kfree_skb() -> dev_kfree_skb_irq(), spin_lock_irq() -> spin_lock()).

This is a known issue (in 2.6.11-rc2). Linus' tree should certainly
include the revert patch at this time.

--
Eric

Re: "Close NETIF_F_LLTX race conditions" problem

[Francois Romieu]

OGAWA Hirofumi <hirofumi@mail.parknet.co.jp> :
[context if hard_start_xmit changed]
> Umm.. Should we fix the all drivers? Any idea?

The bk-commit mailing-list shows that the culprit has been reverted.

Actually it would probably make sense to audit/change the drivers to
fulfill the requirements for Davem's patch before he resurrects it.

--
Ueimor