Re: [Xenomai-core] [COW-BUG] __alloc_pages called from atomic context

[Jan Kiszka <jan.kiszka@domain.hid>]

[-- Attachment #1: Type: text/plain, Size: 2257 bytes --]

Gilles Chanteperdrix wrote:
> Jan Kiszka wrote:
>> Hi all,
>>
>> we are getting a lot of
>>
>> BUG: sleeping function called from invalid context at mm/page_alloc.c:1225
>> in_atomic():1, irqs_disabled():0
>>  [<c010305d>] show_trace_log_lvl+0x1a/0x2f
>>  [<c0103156>] show_trace+0x12/0x14
>>  [<c0103915>] dump_stack+0x16/0x18
>>  [<c010c4ab>] __might_sleep+0xcd/0xd3
>>  [<c0149488>] __alloc_pages+0x32/0x281
>>  [<c014fdd2>] copy_page_range+0x221/0x41e
>>  [<c010ec18>] copy_process+0x9e1/0xfe2
>>  [<c010f415>] do_fork+0x99/0x176
>>  [<c0100e75>] sys_clone+0x33/0x39
>>  [<c0102aaf>] syscall_call+0x7/0xb
>>  =======================
>>
>> here due to a Xenomai program issuing system() calls.
>>
>> After once again dissecting the "nice" mm code (sigh...), the reason
>> turned out to be plain simple:
>>
>> copy_pte_range(...);
>>   spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
>>   copy_one_pte(...);
>>     if (is_cow_mapping(vm_flags))
>>       alloc_page_vma(GFP_HIGHUSER, ...);
>>         __alloc_pages(...)
>> 	  might_sleep_if(gfp_mask & __GFP_WAIT);
>>
>> And this is true due to #define GFP_HIGHUSER (__GFP_WAIT | ...
>>
>> So the bad news is that the COW code in likely all i-pipe versions is
>> broken. But the good new is that this might be easily fixable by
>> providing the right gfp_mask. GFP_ATOMIC?
> 
> It does not look like a good solution, you are going to empty the
> GFP_ATOMIC pools. The proper solution would rather be to look at the
> real mm code (I mean not the one I wrote) and see how they cope with
> this issue.

Mmpf. What are the chances for a quick fix within the next days? We have
to consider alternatives right now here because the whole system is
meant for production purpose next week (C-ELROB '07).

OK, I'm already finding myself inside the code :-/. What about this
approach: We try to alloc with GFP_ATOMIC. Once this fails, we break
out, drop all locks (just like it happens in case of need_resched()),
try to fill up the pool, and restart then. What would reliably make
Linux refill its atomic pool?

Alternative approach: preallocate the required pages before entering the
loop in copy_pte_range. But that may require more code changes.

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 250 bytes --]