Re: [patch 5/5] zfcp: Issue FCP command without holding SCSI host_lock

[Boaz Harrosh <bharrosh@panasas.com>]

On 11/17/2010 04:10 PM, Christof Schmitt wrote:
> On Wed, Nov 17, 2010 at 03:43:34PM +0200, Boaz Harrosh wrote:
>> On 11/17/2010 03:23 PM, Christof Schmitt wrote:
>>> From: Christof Schmitt <christof.schmitt@de.ibm.com>
>>>
>>> Interrupting the connection to the FCP channel while I/O requests are
>>> being issues can lead to this deadlock. scsi_dispatch_cmd already
>>> holds the host_lock while the recovery trigger tries to acquire the
>>> host_lock again when iterating through the scsi_devices.
>>>
>>>  INFO: lockdep is turned off.
>>>  BUG: spinlock lockup on CPU#1, blast/9660, 0000000078f38878
>>>  CPU: 1 Not tainted 2.6.35.7SWEN2 #2
>>>  Process blast (pid: 9660, task: 0000000071f75940, ksp: 0000000074393ac0)
>>>         0000000074393640 00000000743935c0 0000000000000002 0000000000000000
>>>         0000000074393660 00000000743935d8 00000000743935d8 00000000005590c2
>>>         0000000000000000 0000000078f38878 0000000026ede800 0000000078f38878
>>>         000000000000000d 040000000000000c 0000000074393628 0000000000000000
>>>         0000000000000000 0000000000100b2a 00000000743935c0 0000000074393600
>>>  Call Trace:
>>>  ([<0000000000100a32>] show_trace+0xee/0x144)
>>>   [<00000000003be202>] do_raw_spin_lock+0x112/0x178
>>>   [<000000000055d408>] _raw_spin_lock_irqsave+0x90/0xb0
>>>   [<00000000003f1514>] __scsi_iterate_devices+0x38/0xbc
>>>   [<00000000004849b0>] zfcp_erp_clear_adapter_status+0xd0/0x16c
>>>   [<000000000048587a>] zfcp_erp_adapter_reopen+0x3a/0xb4
>>>   [<0000000000489812>] zfcp_fsf_req_send+0x166/0x180
>>>   [<000000000048c8d6>] zfcp_fsf_fcp_cmnd+0x272/0x408
>>>   [<000000000048f864>] zfcp_scsi_queuecommand+0x11c/0x1e0
>>>   [<00000000003f1f2a>] scsi_dispatch_cmd+0x1d6/0x324
>>>   [<00000000003f9910>] scsi_request_fn+0x42c/0x56c
>>>   [<00000000003828ae>] __blk_run_queue+0x86/0x140
>>>   [<000000000037f742>] elv_insert+0x11a/0x208
>>>   [<000000000038104c>] blk_insert_cloned_request+0x84/0xe4
>>>   [<000003c0032b7c64>] dm_dispatch_request+0x6c/0x94 [dm_mod]
>>>   [<000003c0032b7d5c>] map_request+0xd0/0x100 [dm_mod]
>>>   [<000003c0032b9a78>] dm_request_fn+0xec/0x1bc [dm_mod]
>>>   [<0000000000382c0e>] generic_unplug_device+0x5a/0x6c
>>>   [<000003c0032b7f98>] dm_unplug_all+0x74/0x9c [dm_mod]
>>>   [<00000000001d1272>] sync_page+0x76/0x9c
>>>   [<00000000001d12ba>] sync_page_killable+0x22/0x60
>>>   [<000000000055a768>] __wait_on_bit_lock+0xc0/0x124
>>>   [<00000000001d1140>] __lock_page_killable+0x78/0x84
>>>   [<00000000001d351c>] generic_file_aio_read+0x5a4/0x7e8
>>>   [<0000000000228ec0>] do_sync_read+0xc8/0x12c
>>>   [<0000000000229edc>] vfs_read+0xac/0x1ac
>>>   [<000000000022a0d8>] SyS_read+0x58/0xa8
>>>   [<00000000001146de>] sysc_noemu+0x10/0x16
>>>   [<00000200000493c4>] 0x200000493c4
>>>  INFO: lockdep is turned off.
>>>
>>> Call zfcp_fsf_fcp_cmnd without the host_lock and disable the
>>> interrupts when acquiring the req_q_lock.
>>>
>>> Reviewed-by: Swen Schillig <swen@vnet.ibm.com>
>>> Signed-off-by: Christof Schmitt <christof.schmitt@de.ibm.com>
>>> ---
>>>
>>>  drivers/s390/scsi/zfcp_fsf.c  |    5 +++--
>>>  drivers/s390/scsi/zfcp_scsi.c |    3 +++
>>>  2 files changed, 6 insertions(+), 2 deletions(-)
>>>
>>> --- a/drivers/s390/scsi/zfcp_fsf.c
>>> +++ b/drivers/s390/scsi/zfcp_fsf.c
>>> @@ -2170,12 +2170,13 @@ int zfcp_fsf_fcp_cmnd(struct scsi_cmnd *
>>>  	struct zfcp_adapter *adapter = zfcp_sdev->port->adapter;
>>>  	struct zfcp_qdio *qdio = adapter->qdio;
>>>  	struct fsf_qtcb_bottom_io *io;
>>> +	unsigned long flags;
>>>  
>>>  	if (unlikely(!(atomic_read(&zfcp_sdev->status) &
>>>  		       ZFCP_STATUS_COMMON_UNBLOCKED)))
>>>  		return -EBUSY;
>>>  
>>> -	spin_lock(&qdio->req_q_lock);
>>> +	spin_lock_irqsave(&qdio->req_q_lock, flags);
>>>  	if (atomic_read(&qdio->req_q_free) <= 0) {
>>>  		atomic_inc(&qdio->req_q_full);
>>>  		goto out;
>>> @@ -2239,7 +2240,7 @@ failed_scsi_cmnd:
>>>  	zfcp_fsf_req_free(req);
>>>  	scsi_cmnd->host_scribble = NULL;
>>>  out:
>>> -	spin_unlock(&qdio->req_q_lock);
>>> +	spin_unlock_irqrestore(&qdio->req_q_lock, flags);
>>>  	return retval;
>>>  }
>>>  
>>> --- a/drivers/s390/scsi/zfcp_scsi.c
>>> +++ b/drivers/s390/scsi/zfcp_scsi.c
>>> @@ -83,6 +83,7 @@ static int zfcp_scsi_queuecommand_lck(st
>>>  	struct zfcp_adapter *adapter = zfcp_sdev->port->adapter;
>>>  	struct fc_rport *rport = starget_to_rport(scsi_target(scpnt->device));
>>>  	int    status, scsi_result, ret;
>>> +	struct scsi_device *sdev = scpnt->device;
>>>  
>>>  	/* reset the status for this request */
>>>  	scpnt->result = 0;
>>> @@ -118,7 +119,9 @@ static int zfcp_scsi_queuecommand_lck(st
>>>  		return 0;
>>>  	}
>>>  
>>> +	spin_unlock_irq(sdev->host->host_lock);
>>>  	ret = zfcp_fsf_fcp_cmnd(scpnt);
>>> +	spin_lock_irq(sdev->host->host_lock);
>>
>> CCing Jeff
>>
>> that locks is taken in your own driver three lines below at the
>> DEF_SCSI_QCMD macro invocation
>>
>> Please do the proper host-lock-removal. The first time you are
>> touching this code. (See example patch to libata by Jeff Garzik)
> 
> With the current code, the serial_number has to be updated for each
> command since scsi_error still has the check for the serial_number in
> scsi_try_to_abort_cmd.  The change above is a bug fix for the zfcp
> changes introduced in 2.6.37-rc1, and i would like to fix this now. I
> remember from the host_lock discussion that the serial_number changes
> will happen in the 2.6.38 kernel. To me, it looks like there a two
> changes needed anyway.
> 

Sigh! yes this is a deficiency I don't like about the current
push-down. I think we could do Nic's atomic serial_number as an
intermediate solution, or the pending fix to scsi_eh, and not force
a double change to every driver like we have now.

But I still think it's better to open code DEF_SCSI_QCMD than
to have that ugly 
unlock
...
lock

That looks like an ugly typo (And is slower actually)

Jeff ?
Boaz

> Either
>  - make the above change now (as bug fix)
>  - remove the host_lock from zfcp's queuecommand function when the
>    serial_number becomes optional
> or
>  - change the queuecommand function now to include:
>    take host_lock, call scsi_cmd_get_serial, release host_lock
>  - remove the sequence again when the serial_number becomes optional
> 
> I opted for the first approach, to have a smaller patch now. If the
> second approach is preferred, i can send an updated patch.
> 
> Christof
> --
> To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>