On 2019/09/19 10:56, Ming Lei wrote:
> On Thu, Sep 19, 2019 at 08:26:32AM +0000, Damien Le Moal wrote:
>> On 2019/09/18 18:30, Alan Stern wrote:
>>> On Wed, 18 Sep 2019, Andrea Vai wrote:
>>>
>>>>> Also, I wonder if the changing the size of the data transfers would
>>>>> make any difference. This is easy to try; just write "64" to
>>>>> /sys/block/sd?/queue/max_sectors_kb (where the ? is the appropriate
>>>>> drive letter) after the drive is plugged in but before the test
>>>>> starts.
>>>>
>>>> ok, so I duplicated the tests above for the "64" case (it was
>>>> initially set as "120", if it is relevant to know), leading to 40 tests named as
>>>>
>>>> bad.mon.out_50000000_64_TIMESTAMP
>>>> bad.mon.out_50000000_non64_TIMESTAMP
>>>> good.mon.out_50000000_64_TIMESTAMP
>>>> good.mon.out_50000000_non64_TIMESTAMP
>>>>
>>>> where "64" denotes the ones done with that value in max_sectors_kb,
>>>> and "not64" the ones without it (as far as I can tell, it has been
>>>> always "120").
>>>>
>>>> So, we have 40 traces total. Each set of 10 trials is identified by
>>>> a text file, which contains the output log of the test script (and the
>>>> timestamps), also available in the download zipfile.
>>>>
>>>> Just to summarize here the times, they are respectively (number
>>>> expressed in seconds):
>>>>
>>>> BAD:
>>>> Logs: log_10trials_50MB_BAD_NonCanc_64.txt,
>>>> log_10trials_50MB_BAD_NonCanc_non64.txt
>>>> 64: 34, 34, 35, 39, 37, 32, 42, 44, 43, 40
>>>> not64: 61, 71, 59, 71, 62, 75, 62, 70, 62, 68
>>>> GOOD:
>>>> Logs: log_10trials_50MB_GOOD_NonCanc_64.txt,
>>>> log_10trials_50MB_GOOD_NonCanc_non64.txt
>>>> 64: 34, 32, 35, 34, 35, 33, 34, 33, 33, 33
>>>> not64: 32, 30, 32, 31, 31, 30, 32, 30, 32, 31
>>>
>>> The improvement from using "64" with the bad kernel is quite large.
>>> That alone would be a big help for you.
>>>
>>> However, I did see what appears to be a very significant difference
>>> between the bad and good kernel traces. It has to do with the order in
>>> which the blocks are accessed.
>>>
>>> Here is an extract from one of the bad traces. I have erased all the
>>> information except for the columns containing the block numbers to be
>>> written:
>>>
>>> 00019628 00
>>> 00019667 00
>>> 00019628 80
>>> 00019667 80
>>> 00019629 00
>>> 00019668 00
>>> 00019629 80
>>> 00019668 80
>>>
>>> Here is the equivalent portion from one of the good traces:
>>>
>>> 00019628 00
>>> 00019628 80
>>> 00019629 00
>>> 00019629 80
>>> 0001962a 00
>>> 0001962a 80
>>> 0001962b 00
>>> 0001962b 80
>>>
>>> Notice that under the good kernel, the block numbers increase
>>> monotonically in a single sequence. But under the bad kernel, the
>>> block numbers are not monotonic -- it looks like there are two separate
>>> threads each with its own strictly increasing sequence.
>>>
>>> This is exactly the sort of difference one might expect to see from
>>> the commit f664a3cc17b7 ("scsi: kill off the legacy IO path") you
>>> identified as the cause of the problem. With multiqueue I/O, it's not
>>> surprising to see multiple sequences of block numbers.
>>>
>>> Add it's not at all surprising that a consumer-grade USB storage device
>>> might do a much worse job of handling non-sequential writes than
>>> sequential ones.
>>>
>>> Which leads to a simple question for the SCSI or block-layer
>>> maintainers: Is there a sysfs setting Andrea can tweak which will
>>> effectively restrict a particular disk device down to a single I/O
>>> queue, forcing sequential access?
>>
>> The scheduling inefficiency you are seeing may be coming from the fact that the
>> block layer does a direct issue of requests, bypassing the elevator, under some
>> conditions. One of these is sync requests on a multiqueue device. We hit this
>> problem on Zoned disks which can easily return an error for write requests
>> without the elevator throttling writes per zones (zone write locking). This
>> problem was discovered by Hans (on CC).
>>
>> I discussed this with Hannes yesterday and we think we have a fix, but we'll
>> need to do a lot of testing as all block devices are potentially impacted by the
>> change, including stacked drivers (DM). Performance regression is scary with any
>> change in that area (see blk_mq_make_request() and use of
>> blk_mq_try_issue_directly() vs blk_mq_sched_insert_request()).
>
> Not sure this one is same with yours, for USB, mq-deadline is used at
> default, and direct issue won't be possible. Direct issue is only used
> in case of none or underlying queues of DM multipath.
For a multi-queue zoned disk, mq-deadline is also set, but we have observed
unaligned write IO errors for sync writes because of mq-deadline being bypassed
and as a result zones not being write-locked.
In blk_mq_make_request(), at the end, you have:
} else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
!data.hctx->dispatch_busy)) {
blk_mq_try_issue_directly(data.hctx, rq, &cookie);
} else {
blk_mq_sched_insert_request(rq, false, true, true);
}
Which I read as "for a sync req on a multi-queue device, direct issue",
regardless of the elevator being none or something else.
The correct test should probably be:
} else if (!q->elevator &&
((q->nr_hw_queues > 1 && is_sync) ||
!data.hctx->dispatch_busy))) {
blk_mq_try_issue_directly(data.hctx, rq, &cookie);
} else {
blk_mq_sched_insert_request(rq, false, true, true);
}
That is, never bypass the elevator if one is set. Thoughts ?
--
Damien Le Moal
Western Digital Research
