2012-10-19 17:43:07

by Chaoxing Lin

Subject: bug in mesh_sync_adjust_tbtt()

I have checked a few kernel versions between 3.5 and 3.6.2, even the latest version 3.6.2 has the same problem.


In file net/mac80211/mesh_sync.c

Function: coid ?mesh_sync_adjust_tbtt()
{
??? spin_lock_bh(&ifmsh->sync_offset_lock);
..
??? tsf = drv_get_tsf(local, sdata);
??? if (tsf != -1ULL)
??????? drv_set_tsf(local, sdata, tsf + tsfdelta);

??? spin_unlock_bh(&ifmsh->sync_offset_lock);
}


In file net/mac80211/driver-ops.h

static inline u64 drv_get_tsf(struct ieee80211_local *local,
????????????????? struct ieee80211_sub_if_data *sdata)
{??
.
??? might_sleep();
.
}

The problem is drv_get_tsf() can sleep, but it's called in sleep-prohibit context (enclosed by?spin_lock_bh() and??? spin_unlock_bh())

As a result, when running mesh traffic for a while, kernel keeps complaining like below.

One question is, can I get rid of might_sleep() in drv_get/set_tsf()? What would be the impact?


Jan 10 01:46:36 2005 default user.err kernel: BUG: scheduling while atomic: kworker/u:1/37/0x10000200
Jan 10 01:46:36 2005 default user.warn kernel: Modules linked in: ip6table_filter ip6_tables iptable_filter ip_tables x_tables staNat(PO) arc4 ath9k(O
Jan 10 01:46:36 2005 default user.warn kernel: Call Trace:
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbd20] [c00072b0] show_stack+0x58/0x148 (unreliable)
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbd60] [c00413dc] __schedule_bug+0x48/0x60
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbd70] [c035273c] __schedule+0x68/0x46c
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbe60] [c0041418] __cond_resched+0x24/0x38
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbe80] [c0352bd4] _cond_resched+0x2c/0x48
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbe90] [d53af420] mesh_sync_adjust_tbtt+0x8c/0x160 [mac80211]
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbec0] [d53aa53c] ieee80211_mesh_work+0x17c/0x184 [mac80211]
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbee0] [d538e384] ieee80211_iface_work+0x2a8/0x2b0 [mac80211]
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbf20] [c00321d8] process_one_work+0x1e0/0x2dc
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbf50] [c0033b84] worker_thread+0x24c/0x348
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbfa0] [c003804c] kthread+0xa4/0xa8
Jan 10 01:46:36 2005 default user.warn kernel: [ce9bbff0] [c000ccac] kernel_thread+0x4c/0x68

2012-10-19 18:34:57

by Thomas Pedersen

Subject: Re: bug in mesh_sync_adjust_tbtt()

Hi Lin,

On Fri, Oct 19, 2012 at 10:43 AM, Chaoxing Lin
<[email protected]> wrote:
> I have checked a few kernel versions between 3.5 and 3.6.2, even the latest version 3.6.2 has the same problem.

Yes, this has been fixed in
https://patchwork.kernel.org/patch/1556391/. Unfortunately I forgot to
CC [email protected], so this fix won't make it in until 3.7.

Submitted to [email protected] and lesson (hopefully) learned.
This should be fixed in 3.6.3

Thanks,
Thomas