Re: [PATCH] scsi: sd_zbc: update write pointer offset cache

[Johannes Thumshirn <Johannes.Thumshirn@wdc.com>]

On 12/03/2021 11:05, Shinichiro Kawasaki wrote:
> On Mar 12, 2021 / 08:58, Johannes Thumshirn wrote:
>> On 12/03/2021 09:20, Damien Le Moal wrote:
>>> On 2021/03/12 16:59, Johannes Thumshirn wrote:
>>>> On 12/03/2021 08:27, Damien Le Moal wrote:
>>>>> On 2021/03/12 13:38, Shinichiro Kawasaki wrote:
>>>>>> On Mar 11, 2021 / 15:54, Johannes Thumshirn wrote:
>>>>>>> On 11/03/2021 16:48, Bart Van Assche wrote:
>>>>>>>> On 3/11/21 7:18 AM, Johannes Thumshirn wrote:
>>>>>>>>> On 11/03/2021 16:13, Bart Van Assche wrote:
>>>>>>>>>> On 3/10/21 1:48 AM, Johannes Thumshirn wrote:
>>>>>>>>>>> Recent changes [ ... ]
>>>>>>>>>>
>>>>>>>>>> Please add Fixes: and/or Cc: stable tags as appropriate.
>>>>>>>>>
>>>>>>>>> I couldn't pin down the offending commit and I can't reproduce it locally
>>>>>>>>> as well, so I opted out of this. But it must be something between v5.11 and v5.12-rc2.
>>>>>>>>
>>>>>>>> That's weird. Did Shinichiro use a HBA? Could this be the result of a
>>>>>>>> behavior change in the HBA driver?
>>>>>>>
>>>>>>> Yes I've looked at the commits in mpt3sas, but can't really pinpoint the 
>>>>>>> offending commit TBH. 664f0dce2058 ("scsi: mpt3sas: Add support for shared 
>>>>>>> host tagset for CPU hotplug") is the only one that /looks/ as if it could
>>>>>>> be causing it, but I don't know mpt3sas well enough.
>>>>>>>
>>>>>>> FWIW added Sreekanth
>>>>>>
>>>>>> The WARNING was found in kernel v5.12-rc2 test with a SAS SMR drive and HBA
>>>>>> Broadcom 9400. It can be recreated by running blktests block/004 on the drive
>>>>>> (after reboot). It is also recreated with SATA SMR drive with the HBA, but not
>>>>>> observed with SATA drives connected to AHCI.
>>>>>>
>>>>>> I reverted the commit 664f0dce2058, then the WARNING disappeared. I suppose
>>>>>> it indicates that the commit changed HBA driver behavior.
>>>>>
>>>>> Can you send the warning splat with backtrace ?
>>>>>
>>>>
>>>> The warning splat is in the commit message:
>>>> CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.12.0-rc2+ #2
>>>>  Hardware name: Supermicro Super Server/X10SRL-F, BIOS 2.0 12/17/2015
>>>>  RIP: 0010:__local_bh_disable_ip+0x3f/0x50
>>>>  RSP: 0018:ffff8883e1409ba8 EFLAGS: 00010006
>>>>  RAX: 0000000080010001 RBX: 0000000000000001 RCX: 0000000000000013
>>>>  RDX: ffff888129e4d200 RSI: 0000000000000201 RDI: ffffffff915b9dbd
>>>>  RBP: ffff888113e9a540 R08: ffff888113e9a540 R09: 00000000000077f0
>>>>  R10: 0000000000080000 R11: 0000000000000001 R12: ffff888129e4d200
>>>>  R13: 0000000000001000 R14: 00000000000077f0 R15: ffff888129e4d218
>>>>  FS:  0000000000000000(0000) GS:ffff8883e1400000(0000) knlGS:0000000000000000
>>>>  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>>>  CR2: 00007f2f8418ebc0 CR3: 000000021202a006 CR4: 00000000001706f0
>>>>  Call Trace:
>>>>   <IRQ>
>>>>   _raw_spin_lock_bh+0x18/0x40
>>>>   sd_zbc_complete+0x43d/0x1150
>>>>   sd_done+0x631/0x1040
>>>>   ? mark_lock+0xe4/0x2fd0
>>>>   ? provisioning_mode_store+0x3f0/0x3f0
>>>>   scsi_finish_command+0x31b/0x5c0
>>>>   _scsih_io_done+0x960/0x29e0 [mpt3sas]
>>>>   ? mpt3sas_scsih_scsi_lookup_get+0x1c7/0x340 [mpt3sas]
>>>>   ? __lock_acquire+0x166b/0x58b0
>>>>   ? _get_st_from_smid+0x4a/0x80 [mpt3sas]
>>>>   _base_process_reply_queue+0x23f/0x26e0 [mpt3sas]
>>>>   ? lock_is_held_type+0x98/0x110
>>>>   ? find_held_lock+0x2c/0x110
>>>>   ? mpt3sas_base_sync_reply_irqs+0x360/0x360 [mpt3sas]
>>>>   _base_interrupt+0x8d/0xd0 [mpt3sas]
>>>>   ? rcu_read_lock_sched_held+0x3f/0x70
>>>>   __handle_irq_event_percpu+0x24d/0x600
>>>>   handle_irq_event+0xef/0x240
>>>>   ? handle_irq_event_percpu+0x110/0x110
>>>>   handle_edge_irq+0x1f6/0xb60
>>>>   __common_interrupt+0x75/0x160
>>>>   common_interrupt+0x7b/0xa0
>>>>   </IRQ>
>>>>   asm_common_interrupt+0x1e/0x40
>>>>
>>>
>>> Looking at patch 664f0dce2058, all that seems to be done is to enable
>>> nr_hw_queue > 1. I do not see any change of locking context or irq handling.
>>> From the backtrace, it does not look like scsi_finish_command() is called from
>>> softirq... Probably a change in that area is responsible ?
>>>
>>
>>
>> In scsi_lib.c we only have these two patches in that area:
>>
>> 684da7628d93 ("block: remove unnecessary argument from blk_execute_rq")
>> 962c8dcdd5fa ("scsi: core: Add a new error code DID_TRANSPORT_MARGINAL in scsi.h")
>>
>> and none of them can cause the failure either. In block we have:
>>
>> 0a2efafbb1c7 ("blk-mq: Always complete remote completions requests in softirq")
>>
>> but this doesn't look guilty as well, all it does is raising a softirq for all
>> block completions local and remote.
> 
> In blk_mq_complete_request_remote(), I found the following code.
> 
> 	if (rq->q->nr_hw_queues == 1) {
> 		blk_mq_raise_softirq(rq);
> 		return true;
> 	}
> 	return false;
> 
> My mere guess is that the commit 664f0dce2058 changed the shost->nr_hw_queue
> from zero to a value larger than 1 (with my test system, it is 8), it is
> propagated to rq->q->nr_hw_queues, then blk_mq_raise_softirq() is no longer
> called.
> 
> The call stack I assume is as follows: without calling blk_mq_raise_softirq(),
> there are all executed in IRQ context, probably.
> 
>   _scsih_io_done()
>     scmd->scsi_done() = scsi_mq_done()
>       blk_mq_complete_request()
>         blk_mq_complete_request_remote() ... did not call blk_mq_raise_softirq()
>         rq->q->mq_ops->complete() = scsi_soft_irq_done()
> 	  scsi_finish_command()
> 	    drv->done() = sd_done()
> 
> Will confirm this guess further.
> 

But commit 0a2efafbb1c7 ("blk-mq: Always complete remote completions requests 
in softirq") changed it to:

 
-       /*
-        * For most of single queue controllers, there is only one irq vector
-        * for handling I/O completion, and the only irq's affinity is set
-        * to all possible CPUs.  On most of ARCHs, this affinity means the irq
-        * is handled on one specific CPU.
-        *
-        * So complete I/O requests in softirq context in case of single queue
-        * devices to avoid degrading I/O performance due to irqsoff latency.
-        */
-       if (rq->q->nr_hw_queues == 1)
-               blk_mq_trigger_softirq(rq);
-       else
-               rq->q->mq_ops->complete(rq);
+       blk_mq_trigger_softirq(rq);
 }

So to my understanding, we will always complete in a softirq.