Re: [5.0-rc5 regression] "scsi: kill off the legacy IO path" causes 5 minute delay during boot on Sun Blade 2500

[Jens Axboe]

On 2/10/19 9:25 AM, James Bottomley wrote:
> On Sun, 2019-02-10 at 09:05 -0700, Jens Axboe wrote:
>> On 2/10/19 8:44 AM, James Bottomley wrote:
>>> On Sun, 2019-02-10 at 10:17 +0100, Mikael Pettersson wrote:
>>>> On Sat, Feb 9, 2019 at 7:19 PM James Bottomley
>>>> <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> wrote:
>>>
>>> [...]
>>>>> I think the reason for this is that the block mq path doesn't
>>>>> feed
>>>>> the kernel entropy pool correctly, hence the need to install an
>>>>> entropy gatherer for systems that don't have other good random
>>>>> number sources.
>>>>
>>>> That does sound plausible, I admit I didn't even consider the
>>>> possibility that the old block I/O path also was an entropy
>>>> source.
>>>
>>> In theory, the new one should be as well since the rotational
>>> entropy
>>> collector is on the SCSI completion path.   I'd seen the same
>>> problem
>>> but had assumed it was something someone had done to our internal
>>> entropy pool and thus hadn't bisected it.
>>
>> The difference is that the old stack included ADD_RANDOM by default,
>> so this check:
>>
>> 	if (blk_queue_add_random(q))
>> 		add_disk_randomness(req->rq_disk);
>>
>> in scsi_end_request() would be true, and we'd add the randomness. For
>> sd, it seems to set it just fine for non-rotational drives. Could
>> this be because other devices don't? Maybe the below makes a
>> difference.
> 
> No, in both we set it per the rotational parameters of the disk in 
> 
> sd.c:sd_read_block_characteristics()
> 
> 	rot = get_unaligned_be16(&buffer[4]);
> 
> 	if (rot == 1) {
> 	
> 	blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
> 	
> 	blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, q);
> 	} else {
> 	
> 	blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
> 	
> 	blk_queue_flag_set(QUEUE_FLAG_ADD_RANDOM, q);
> 	}
> 
> 
> That check wasn't changed by the code removal.

As I said above, for sd. This isn't true for non-disks.

> Although I suspect it should be unconditional: even SSDs have what
> would appear as seek latencies at least during writes depending on the
> time taken to find an erased block or even trigger garbage collection. 
> The entropy collector is good at taking something completely regular
> and spotting the inconsistencies, so it won't matter that loads of
> "seeks" are deterministic.

The reason it isn't is that it's of limited use for SSDs where it's a
lot more predictable. And they are also a lot faster, which means the
adding randomness is more problematic from an efficiency pov.

-- 
Jens Axboe