[lkml]   [2019]   [Feb]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [5.0-rc5 regression] "scsi: kill off the legacy IO path" causes 5 minute delay during boot on Sun Blade 2500
On 2/11/19 8:25 AM, James Bottomley wrote:
> On Sun, 2019-02-10 at 09:35 -0700, Jens Axboe wrote:
>> 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
>>>>>> <> 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.
> Yes, but the behaviour above doesn't change across a switch to MQ, so I
> don't quite understand how it bisects back to that change. If we're
> not gathering entropy for the device now, we wouldn't have been before
> the switch, so the entropy characteristics shouldn't have changed.

But it does, as I also wrote in that first email. The legacy queue
flags had QUEUE_FLAG_ADD_RANDOM set by default, the MQ ones do not.
Hence any non-sd device would previously ALWAYS have ADD_RANDOM
set, now none of them do. Also see the patch I sent.

>>> 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.
> But that's my point: our entropy extractor is good at weeding out
> predictable signals. Fine, it won't extract any entropy if the disk
> seek time is entirely regular, but it won't contaminate the entropy
> pool. The computational delay, I grant ... it takes a while to
> determine if any entropy is present in the signal.

But you are missing my point - if we're mostly weeding out predictable
signals, then it's pointless to take the overhead of the randomness.
This is why the MQ flag don't include it by default.

> What about feeding it with something like discard timings, which should
> be much less predictable.

That's not true, lots of devices have VERY predictable discard timings.
Most of them will have a fixed discards-per-second rate, even.

Jens Axboe

 \ /
  Last update: 2019-02-11 16:31    [W:0.044 / U:4.056 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site