Re: [PATCH] scsi: mpt3sas: fix hang on ata passthru commands

[Jason Baron]

On 12/29/2016 03:02 AM, Christoph Hellwig wrote:

> On Wed, Dec 28, 2016 at 11:30:24PM -0500, Jason Baron wrote:
>> Add a new parameter to scsi_internal_device_block() to decide whether
>> or not to invoke scsi_wait_for_queuecommand().
> We'll also need to deal with the blk-mq wait path that Bart has been
> working on (I think it's already in the scsi tree, but I'd have to
> check).
Ok, I'm not sure either.

> Also adding a bool flag for the last call in a function is style that's
> a little annoying.
>
> I'd prefer to add a scsi_internal_device_block_nowait that contains
> all the code except for the waiting, and then make
> scsi_internal_device_block_nowait a wrapper around it.  Or drop the
> annoying internal for both while we're at it :)

The proposed patch brings the code in-line with what is in 4.8 stable
where scsi_internal_device_block() does not call
scsi_wait_for_queuecommand(). So I saw it as a minimal fix to make
my system boot again :)

I was wondering if the original fix is racy in that there could be multiple
threads in the queuecommand. Perhaps we should do something like:
       
    if (ata_12_16_cmd(scmd))
{                                                                                                                                   

                if (!test_and_set_bit(MPT_DEVICE_EXCLUSIVE,
&sas_device_priv_data->flags))
{                                                                         
                       
scsi_internal_device_block(scmd->device);                                                                                                    

                }
else                                                                                                                                               

                        return
SCSI_MLQUEUE_HOST_BUSY;                                                                                                               

        }        

where scsi_internal_device_block() could be taught to wait for
request_fn_active becoming 1 instead of 0.

Thanks,

-Jason