Re: [PATCH][noop-iosched] don't reuse a freed request

[Tejun Heo <htejun@gmail.com>]

Jens Axboe wrote:
> On Mon, Oct 31 2005, Tejun Heo wrote:
> 
>>Hi, guys.
>>
>>Jens Axboe wrote:
>>
>>>On Mon, Oct 31 2005, Arnaldo Carvalho de Melo wrote:
>>>
>>>
>>>>Hi,
>>>>
>>>>	I'm getting the oops below when trying to use qemu with a kernel
>>>>built with just the noop iosched, I'm never had looked at this code 
>>>>before,
>>>>so I did a quick hack that seems enough for my case.
>>>>
>>>>	Ah, this is with a fairly recent git tree (today), haven't checked
>>>>if it is present in 2.6.14.
>>>>
>>>>Best Regards,
>>>>
>>>>- Arnaldo
>>>>
>>>>Unable to handle kernel paging request at virtual address c5f20f60
>>>>printing eip:
>>>>c01b0ecd
>>>>*pde = 00017067
>>>>*pte = 05f20000
>>>>Oops: 0000 [#1]
>>>>DEBUG_PAGEALLOC
>>>>Modules linked in:
>>>>CPU:    0
>>>>EIP:    0060:[<c01b0ecd>]    Not tainted VLI
>>>>EFLAGS: 00000046   (2.6.14acme)
>>>>EIP is at elv_rq_merge_ok+0x15/0x7b
>>>>eax: 00000014   ebx: c5f20f58   ecx: 000003f8   edx: 00000046
>>>>esi: c12a5a90   edi: c5f20f58   ebp: c11658d0   esp: c11658c4
>>>>ds: 007b   es: 007b   ss: 0068
>>>>Process swapper (pid: 1, threadinfo=c1165000 task=c1164af0)
>>>>Stack: c0251883 c5ecfe4c c5d688c0 c1165904 c01b0f48 c5f20f58 c12a5a90 
>>>>00000000
>>>>     c5874000 c018c5e1 c5f15f24 0000002b 00000000 c5ecfe4c c5d688c0 
>>>>     c12a5a90
>>>>     c1165920 c01b128d c5f20f58 c12a5a90 000a568a 00000000 00000002 
>>>>     c1165960
>>>>Call Trace:
>>>>[<c0102a63>] show_stack+0x78/0x83
>>>>[<c0102b88>] show_registers+0x100/0x167
>>>>[<c0102d35>] die+0xcb/0x140
>>>>[<c0234308>] do_page_fault+0x393/0x53a
>>>>[<c0102777>] error_code+0x4f/0x54
>>>>[<c01b0f48>] elv_try_merge+0x15/0x84
>>>>[<c01b128d>] elv_merge+0x1d/0x4f
>>>>[<c01b41d9>] __make_request+0xb2/0x425
>>>>[<c01b46f9>] generic_make_request+0x125/0x137
>>>
>>>
>>>Hrmpf, this looks really bad. Tejun, clearly there are still paths where
>>>->last_rq isn't being cleared.
>>>
>>
>>I'm currently debugging this.  The problem is that we are using generic 
>>dispatch queue directly in the noop and merging is NOT allowed on 
>>dispatch queues but generic handling of last_merge tries to merge 
>>requests.  I'm still trying to verify this, so I'll be back with results 
>>soon.
>>
>>
>>>>--- a/drivers/block/ll_rw_blk.c
>>>>+++ b/drivers/block/ll_rw_blk.c
>>>>@@ -1787,6 +1787,9 @@ static inline void blk_free_request(requ
>>>>	if (rq->flags & REQ_ELVPRIV)
>>>>		elv_put_request(q, rq);
>>>>	mempool_free(rq, q->rq.rq_pool);
>>>>+
>>>>+	if (rq == q->last_merge)
>>>>+		q->last_merge = NULL;
>>>>}
>>>>
>>>>static inline struct request *
>>>
>>>
>>>It's most likely a bug getting this far in the first place, but does it
>>>fix things for you? I'll get on this asap.
>>>
>>
>>If the bug is where I think it is, I think the proper thing to do is to 
>>use separate list_head in noop instead of using generic dispatch queue 
>>directly thus making noop consistent with other ioscheds.
>>
>>I'm more worried about oops w/ cfq Arnaldo reported in this thread. 
>>I'll track that down as soon as I'm done with this one.
> 
> 
> So either we disable merging for noop by setting REQ_NOMERGE in
> elevator_noop_add_request(), or we add a noop_list and do the
> dispatching like in the other io schedulers. I'd prefer the latter,
> merging is still beneficial for noop (and it has always done it).

Just verified.  It happens when elv_merge_requests() happens.  The 
merged request should be unlinked from list but noop does not have any 
merge handling callbacks ATM.

Sorry about the hassle. :-(

> 
> For now, we should add the former.

Yeap, also verified oops doesn't happen with the following patch.

I'll soon post a patch to convert noop such that it does proper 
dispatching.  BTW, while I was looking at the code, I found something 
else, in elv_former/latter_request functions, if the iosched doesn't 
supply the callbacks, it uses rq->queue_list.prev/next implicitly 
(without this, this noop bug wouldn't have been triggered).  I think 
this code is not necessary anymore.  What do you think?

Thanks.

-- 
tejun