Should io_limits/queue depth flushed persist while using verify_state?

Should io_limits/queue depth flushed persist while using verify_state?

[Sitsofe Wheeler]

While using fio-2.2.4 the following snippet produces a verification
error for me on the second fio invocation:

fio --bs=4k --ioengine=libaio --direct=1 --verify=meta --iodepth=128
--io_limit=300k --verify_state_save=1 --rw=randwrite --name=write3
--filename=/dev/dm-7 && fio --bs=4k --ioengine=libaio --direct=1
--verify=meta --iodepth=128 --verify_state_load=1 --rw=randread
--name=write3 --filename=/dev/dm-7

However if I put the iodepth of the second fio command down to 64 or
set io_limit the error goes away. I have also sometimes seen an error
that looks similar if the first fio is killed via ctrl-C. Is this to
be expected?

-- 
Sitsofe | http://sucs.org/~sits/

Re: Should io_limits/queue depth flushed persist while using verify_state?

[Jens Axboe]

On 01/15/2015 04:01 AM, Sitsofe Wheeler wrote:
> While using fio-2.2.4 the following snippet produces a verification
> error for me on the second fio invocation:
>
> fio --bs=4k --ioengine=libaio --direct=1 --verify=meta --iodepth=128
> --io_limit=300k --verify_state_save=1 --rw=randwrite --name=write3
> --filename=/dev/dm-7 && fio --bs=4k --ioengine=libaio --direct=1
> --verify=meta --iodepth=128 --verify_state_load=1 --rw=randread
> --name=write3 --filename=/dev/dm-7
>
> However if I put the iodepth of the second fio command down to 64 or
> set io_limit the error goes away. I have also sometimes seen an error
> that looks similar if the first fio is killed via ctrl-C. Is this to
> be expected?

Doesn't reproduce here, but both cases would seem to imply that there's 
something not quite right with the completed ios in the verify state 
output. I'll take a look.

-- 
Jens Axboe

Re: Should io_limits/queue depth flushed persist while using verify_state?

[Jens Axboe]

[-- Attachment #1: Type: text/plain, Size: 1105 bytes --]

On 01/15/2015 03:45 PM, Jens Axboe wrote:
> On 01/15/2015 04:01 AM, Sitsofe Wheeler wrote:
>> While using fio-2.2.4 the following snippet produces a verification
>> error for me on the second fio invocation:
>>
>> fio --bs=4k --ioengine=libaio --direct=1 --verify=meta --iodepth=128
>> --io_limit=300k --verify_state_save=1 --rw=randwrite --name=write3
>> --filename=/dev/dm-7 && fio --bs=4k --ioengine=libaio --direct=1
>> --verify=meta --iodepth=128 --verify_state_load=1 --rw=randread
>> --name=write3 --filename=/dev/dm-7
>>
>> However if I put the iodepth of the second fio command down to 64 or
>> set io_limit the error goes away. I have also sometimes seen an error
>> that looks similar if the first fio is killed via ctrl-C. Is this to
>> be expected?
>
> Doesn't reproduce here, but both cases would seem to imply that there's
> something not quite right with the completed ios in the verify state
> output. I'll take a look.

I wonder if this will do the trick, can you check? If not, I'll need the 
full output of the two runs, and the state files saved from the first run.


-- 
Jens Axboe


[-- Attachment #2: depth-ver.patch --]
[-- Type: text/x-patch, Size: 663 bytes --]

diff --git a/verify.c b/verify.c
index 205f01a2fc30..260fa2ae2fad 100644
--- a/verify.c
+++ b/verify.c
@@ -1540,10 +1540,12 @@ int verify_state_should_stop(struct thread_data *td, struct io_u *io_u)
 		return 0;
 
 	/*
-	 * If we're not into the window of issues - depth yet, continue
+	 * If we're not into the window of issues - depth yet, continue. If
+	 * issue is shorter than depth, do check.
 	 */
-	if (td->io_blocks[DDIR_READ] < s->depth ||
-	    s->numberio - td->io_blocks[DDIR_READ] > s->depth)
+	if ((td->io_blocks[DDIR_READ] < s->depth ||
+	    s->numberio - td->io_blocks[DDIR_READ] > s->depth) &&
+	    s->numberio > s->depth)
 		return 0;
 
 	/*

Re: Should io_limits/queue depth flushed persist while using verify_state?

[Sitsofe Wheeler]

[-- Attachment #1: Type: text/plain, Size: 523 bytes --]

On 16 January 2015 at 17:24, Jens Axboe <axboe@kernel.dk> wrote:
>
> I wonder if this will do the trick, can you check? If not, I'll need the
> full output of the two runs, and the state files saved from the first run.

I tried revision ae703cdf31532e337cc18c259c883bf5314aa43a which seems
to already have that patch rolled into it (via commit
def0009508f0f3c763c4de5e0b62388b42544faf ) but it made no
difference...

Find the two runs with --debug=all and and the state file attached.

-- 
Sitsofe | http://sucs.org/~sits/

[-- Attachment #2: fio-depth-ver.tar.xz --]
[-- Type: application/octet-stream, Size: 14860 bytes --]

Re: Should io_limits/queue depth flushed persist while using verify_state?

[Sitsofe Wheeler]

On 17 January 2015 at 09:11, Sitsofe Wheeler <sitsofe@gmail.com> wrote:
> On 16 January 2015 at 17:24, Jens Axboe <axboe@kernel.dk> wrote:
>>
>> I wonder if this will do the trick, can you check? If not, I'll need the
>> full output of the two runs, and the state files saved from the first run.
>
> I tried revision ae703cdf31532e337cc18c259c883bf5314aa43a which seems
> to already have that patch rolled into it (via commit
> def0009508f0f3c763c4de5e0b62388b42544faf ) but it made no
> difference...
>
> Find the two runs with --debug=all and and the state file attached.

Just checking - did the above arrive correctly?

-- 
Sitsofe | http://sucs.org/~sits/

Re: Should io_limits/queue depth flushed persist while using verify_state?

[Jens Axboe]

On 01/28/2015 10:18 PM, Sitsofe Wheeler wrote:
> On 17 January 2015 at 09:11, Sitsofe Wheeler <sitsofe@gmail.com> wrote:
>> On 16 January 2015 at 17:24, Jens Axboe <axboe@kernel.dk> wrote:
>>>
>>> I wonder if this will do the trick, can you check? If not, I'll need the
>>> full output of the two runs, and the state files saved from the first run.
>>
>> I tried revision ae703cdf31532e337cc18c259c883bf5314aa43a which seems
>> to already have that patch rolled into it (via commit
>> def0009508f0f3c763c4de5e0b62388b42544faf ) but it made no
>> difference...
>>
>> Find the two runs with --debug=all and and the state file attached.
>
> Just checking - did the above arrive correctly?

I apparently missed it... Will take a look, but traveling in the next 
few days so response might be a bit slower.

-- 
Jens Axboe

Re: Should io_limits/queue depth flushed persist while using verify_state?

[Sitsofe Wheeler]

On 29 January 2015 at 16:43, Jens Axboe <axboe@kernel.dk> wrote:
> On 01/28/2015 10:18 PM, Sitsofe Wheeler wrote:
>>
>> On 17 January 2015 at 09:11, Sitsofe Wheeler <sitsofe@gmail.com> wrote:
>>>
>>> On 16 January 2015 at 17:24, Jens Axboe <axboe@kernel.dk> wrote:
>>>>
>>>> I wonder if this will do the trick, can you check? If not, I'll need the
>>>> full output of the two runs, and the state files saved from the first
>>>> run.
>>>
>>> I tried revision ae703cdf31532e337cc18c259c883bf5314aa43a which seems
>>> to already have that patch rolled into it (via commit
>>> def0009508f0f3c763c4de5e0b62388b42544faf ) but it made no
>>> difference...
>>>
>>> Find the two runs with --debug=all and and the state file attached.
>>
>> Just checking - did the above arrive correctly?
>
> I apparently missed it... Will take a look, but traveling in the next few
> days so response might be a bit slower.

Just pinging this one again... If it's too soon just let me know!

-- 
Sitsofe | http://sucs.org/~sits/