Re: Bug: ACPI, scheduling while atomic (was Re: [PATCH 0/4] sched: Make sleep inside atomic detection work on !PREEMPT)

[Randy Dunlap <rdunlap@xenotime.net>]

On Thu, 9 Jun 2011 00:49:41 +0200 Frederic Weisbecker wrote:

> On Wed, Jun 08, 2011 at 07:48:31PM +0200, Frederic Weisbecker wrote:
> > Aside it may mostly avoid the need for a specific PROVE_RCU
> > check when we sleep inside an rcu read side critical section.
> > 
> > Better make sleeping inside atomic sections work everywhere.
> 
> BTW, it has led to detect a bug in the ACPI code. It happens in
> !CONFIG_PREEMPT:
> 
> [    0.160187] BUG: scheduling while atomic: swapper/0/0x10000002
> [    0.166016] no locks held by swapper/0.
> [    0.170014] Modules linked in:
> [    0.173107] Pid: 0, comm: swapper Not tainted 2.6.39+ #124
> [    0.180014] Call Trace:
> [    0.182481]  [<ffffffff81048685>] __schedule_bug+0x85/0x90
> [    0.187967]  [<ffffffff817da98c>] schedule+0x75c/0xa40
> [    0.190022]  [<ffffffff8109a1fd>] ? trace_hardirqs_on+0xd/0x10
> [    0.200023]  [<ffffffff813879c0>] ? acpi_ps_free_op+0x22/0x24
> [    0.205776]  [<ffffffff810554a5>] __cond_resched+0x25/0x40
> [    0.210022]  [<ffffffff817daf3b>] _cond_resched+0x2b/0x40
> [    0.215420]  [<ffffffff81386cbe>] acpi_ps_complete_op+0x262/0x278
> [    0.220023]  [<ffffffff813874df>] acpi_ps_parse_loop+0x80b/0x960
> [    0.230023]  [<ffffffff81386607>] acpi_ps_parse_aml+0x98/0x274
> [    0.235859]  [<ffffffff81384cbb>] acpi_ns_one_complete_parse+0x103/0x120
> [    0.240021]  [<ffffffff810886da>] ? up+0x2a/0x50
> [    0.244641]  [<ffffffff81384cf3>] acpi_ns_parse_table+0x1b/0x34
> [    0.250022]  [<ffffffff8138242a>] acpi_ns_load_table+0x4a/0x8c
> [    0.260023]  [<ffffffff8138947c>] acpi_load_tables+0x9c/0x15d
> [    0.265774]  [<ffffffff81d0b0f8>] acpi_early_init+0x6c/0xf7
> [    0.270022]  [<ffffffff81cd8d31>] start_kernel+0x400/0x415
> [    0.275508]  [<ffffffff81cd8346>] x86_64_start_reservations+0x131/0x135
> [    0.280022]  [<ffffffff81cd844d>] x86_64_start_kernel+0x103/0x112
> 
> ACPI_PREEMPTION_POINT() is called from acpi_ps_complete_op() and schedules
> if !PREEMPT. But preemption is disabled as we are in early bootup.

This still happens in 3.1-rc[123].
Was there a patch for it?

---
~Randy
*** Remember to use Documentation/SubmitChecklist when testing your code ***