Re: [PATCH] scsi: storvsc: Cap scsi_driver.can_queue to fix a hang issue during boot

[John Garry <john.garry@huawei.com>]

On 06/10/2021 16:01, Michael Kelley wrote:
> From: John Garry <john.garry@huawei.com> Sent: Wednesday, October 6, 2021 1:17 AM
>>
>> On 06/10/2021 08:03, Dexuan Cui wrote:
>>> After commit ea2f0f77538c, a 416-CPU VM running on Hyper-V hangs during
>>> boot because scsi_add_host_with_dma() sets shost->cmd_per_lun to a
>>> negative number:
>>> 	'max_outstanding_req_per_channel' is 352,
>>> 	'max_sub_channels' is (416 - 1) / 4 = 103, so in storvsc_probe(),
>>> scsi_driver.can_queue = 352 * (103 + 1) * (100 - 10) / 100 = 32947, which
>>> is bigger than SHRT_MAX (i.e. 32767).
>>
>> Out of curiosity, are these values realistic? You're capping can_queue
>> just because of a data size issue, so, if these values are realistic,
>> seems a weak reason.
>>
> 
> The calculated value of can_queue is not realistic.  The blk-mq layer
> caps the number of tags at 10240, 

nit: 1024, I think

> so the excessively large value
> calculated here didn't definitively break anything, though it can be
> poor from a performance tuning standpoint. The algorithm used here
> is fairly broken, particularly in VMs with large CPU counts.  I have an
> effort underway to fix it, but its part of a bigger set of changes to also
> do a better job on the perf tuning aspects.
> 
>>>
>>> Fix the hang issue by capping scsi_driver.can_queue.
>>>
>>> Add the below Fixed tag though ea2f0f77538c itself is good.
>>>
>>> Fixes: ea2f0f77538c ("scsi: core: Cap scsi_host cmd_per_lun at can_queue")
>>> Cc: stable@vger.kernel.org
>>> Signed-off-by: Dexuan Cui <decui@microsoft.com>
>>> ---
>>>    drivers/scsi/storvsc_drv.c | 10 ++++++++++
>>>    1 file changed, 10 insertions(+)
>>>
>>> diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c
>>> index ebbbc1299c62..ba374908aec2 100644
>>> --- a/drivers/scsi/storvsc_drv.c
>>> +++ b/drivers/scsi/storvsc_drv.c
>>> @@ -1976,6 +1976,16 @@ static int storvsc_probe(struct hv_device *device,
>>>    				(max_sub_channels + 1) *
>>>    				(100 - ring_avail_percent_lowater) / 100;
>>>
>>> +	/*
>>> +	 * v5.14 (see commit ea2f0f77538c) implicitly requires that
>>> +	 * scsi_driver.can_queue should not exceed SHRT_MAX, otherwise
>>> +	 * scsi_add_host_with_dma() sets shost->cmd_per_lun to a negative
>>> +	 * number (note: the type of the "cmd_per_lun" field is "short"), and
>>> +	 * the system may hang during early boot.
>>> +	 */
>>
>> The different data sizes for cmd_per_lun and can_queue are problematic here.
>>
>> I'd be more inclined to set cmd_per_lun to the same data size as
>> can_queue. We did discuss this when ea2f0f77538c was upstreamed
>> (actually it was the other way around - setting can_queue to 16b).
> 
> I can see that making can_queue be 16 bits would make sense.
> And it also seems that both cmd_per_lun and can_queue should be
> unsigned, though I don't the implications of making such a change.
> 
> But in today's world where cmd_per_lun is "short" and can_queue
> is "int",  ea2f0f77538c seems incorrect to me.  The comparison should
> be done as "int", not "short", in order to prevent the truncation
> problem with can_queue that Dexuan's patch is trying to address.

Yeah, right. I think can_queue values > short_max was considered outside 
the realms of what is realistic then, hence my sloppy programming.

> The result will always fit in back into the "short" cmd_per_lun since
> it is calculating a "min" function.
> 
>>
>> Thanks,
>> John
>>
>>
>>> +	if (scsi_driver.can_queue > SHRT_MAX)
>>> +		scsi_driver.can_queue = SHRT_MAX;
>>> +
> 
> This fix works, but is a more of a temporary hack until I can finish
> a larger overhaul of the algorithm. 


> But for now, I think the better
> fix is for ea2f0f77538c to do the comparison as "int" instead of "short".
> 

That seems better to me. But Let's wait for other possible opinion.

Thanks,
John