[PATCH] blk: don't account discard request size

[PATCH] blk: don't account discard request size

[Shaohua Li]

In a workload with discard request, the IO throughput is generally much
higher than expected. This is quite confusing checking iostat. Discard
request doesn't really write data to drive, so don't account it.

Signed-off-by: Shaohua Li <shli@fb.com>
---
 block/blk-core.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index fd154b9..0128d18 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -2138,7 +2138,11 @@ EXPORT_SYMBOL_GPL(blk_rq_err_bytes);
 
 void blk_account_io_completion(struct request *req, unsigned int bytes)
 {
-	if (blk_do_io_stat(req)) {
+	/*
+	 * discard request doesn't really write @bytes to drive,
+	 * doesn't account it
+	 **/
+	if (blk_do_io_stat(req) && !(req->cmd_flags & REQ_DISCARD)) {
 		const int rw = rq_data_dir(req);
 		struct hd_struct *part;
 		int cpu;
-- 
1.8.1

Re: [PATCH] blk: don't account discard request size

[Jeff Moyer]

Shaohua Li <shli@fb.com> writes:

> In a workload with discard request, the IO throughput is generally much
> higher than expected. This is quite confusing checking iostat. Discard
> request doesn't really write data to drive, so don't account it.
>
> Signed-off-by: Shaohua Li <shli@fb.com>
> ---
>  block/blk-core.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/block/blk-core.c b/block/blk-core.c
> index fd154b9..0128d18 100644
> --- a/block/blk-core.c
> +++ b/block/blk-core.c
> @@ -2138,7 +2138,11 @@ EXPORT_SYMBOL_GPL(blk_rq_err_bytes);
>  
>  void blk_account_io_completion(struct request *req, unsigned int bytes)
>  {
> -	if (blk_do_io_stat(req)) {
> +	/*
> +	 * discard request doesn't really write @bytes to drive,
> +	 * doesn't account it
> +	 **/
> +	if (blk_do_io_stat(req) && !(req->cmd_flags & REQ_DISCARD)) {
>  		const int rw = rq_data_dir(req);
>  		struct hd_struct *part;
>  		int cpu;

I think you want to modify __get_request to not set REQ_IO_STAT for
discard requests.  This patch will still account the start of I/O, which
means in_flight will be off.

Cheers,
Jeff

Re: [PATCH] blk: don't account discard request size

[Jens Axboe]

On 05/13/2015 09:10 AM, Jeff Moyer wrote:
> Shaohua Li <shli@fb.com> writes:
>
>> In a workload with discard request, the IO throughput is generally much
>> higher than expected. This is quite confusing checking iostat. Discard
>> request doesn't really write data to drive, so don't account it.
>>
>> Signed-off-by: Shaohua Li <shli@fb.com>
>> ---
>>   block/blk-core.c | 6 +++++-
>>   1 file changed, 5 insertions(+), 1 deletion(-)
>>
>> diff --git a/block/blk-core.c b/block/blk-core.c
>> index fd154b9..0128d18 100644
>> --- a/block/blk-core.c
>> +++ b/block/blk-core.c
>> @@ -2138,7 +2138,11 @@ EXPORT_SYMBOL_GPL(blk_rq_err_bytes);
>>
>>   void blk_account_io_completion(struct request *req, unsigned int bytes)
>>   {
>> -	if (blk_do_io_stat(req)) {
>> +	/*
>> +	 * discard request doesn't really write @bytes to drive,
>> +	 * doesn't account it
>> +	 **/
>> +	if (blk_do_io_stat(req) && !(req->cmd_flags & REQ_DISCARD)) {
>>   		const int rw = rq_data_dir(req);
>>   		struct hd_struct *part;
>>   		int cpu;
>
> I think you want to modify __get_request to not set REQ_IO_STAT for
> discard requests.  This patch will still account the start of I/O, which
> means in_flight will be off.

That would be better. But I'm still not sure we want to turn off 
accounting for discards. For the mixed write/discard cases it's 
definitely confusing. The better option would be to account it as a 
discard and not a write. Preferably in a way that would not break 
existing tools, but so that they could get updated to support it.


-- 
Jens Axboe

Re: [PATCH] blk: don't account discard request size

[Jeff Moyer]

Jens Axboe <axboe@fb.com> writes:

> That would be better. But I'm still not sure we want to turn off
> accounting for discards. For the mixed write/discard cases it's
> definitely confusing. The better option would be to account it as a
> discard and not a write. Preferably in a way that would not break
> existing tools, but so that they could get updated to support it.

Are you suggesting adding a few fields to the end of diskstats or adding
a new proc file altogether?  (or something else?)

Cheers,
Jeff

Re: [PATCH] blk: don't account discard request size

[Shaohua Li]

On Wed, May 13, 2015 at 10:20:12AM -0400, Jens Axboe wrote:
> On 05/13/2015 09:10 AM, Jeff Moyer wrote:
> >Shaohua Li <shli@fb.com> writes:
> >
> >>In a workload with discard request, the IO throughput is generally much
> >>higher than expected. This is quite confusing checking iostat. Discard
> >>request doesn't really write data to drive, so don't account it.
> >>
> >>Signed-off-by: Shaohua Li <shli@fb.com>
> >>---
> >>  block/blk-core.c | 6 +++++-
> >>  1 file changed, 5 insertions(+), 1 deletion(-)
> >>
> >>diff --git a/block/blk-core.c b/block/blk-core.c
> >>index fd154b9..0128d18 100644
> >>--- a/block/blk-core.c
> >>+++ b/block/blk-core.c
> >>@@ -2138,7 +2138,11 @@ EXPORT_SYMBOL_GPL(blk_rq_err_bytes);
> >>
> >>  void blk_account_io_completion(struct request *req, unsigned int bytes)
> >>  {
> >>-	if (blk_do_io_stat(req)) {
> >>+	/*
> >>+	 * discard request doesn't really write @bytes to drive,
> >>+	 * doesn't account it
> >>+	 **/
> >>+	if (blk_do_io_stat(req) && !(req->cmd_flags & REQ_DISCARD)) {
> >>  		const int rw = rq_data_dir(req);
> >>  		struct hd_struct *part;
> >>  		int cpu;
> >
> >I think you want to modify __get_request to not set REQ_IO_STAT for
> >discard requests.  This patch will still account the start of I/O, which
> >means in_flight will be off.
> 
> That would be better. But I'm still not sure we want to turn off
> accounting for discards. For the mixed write/discard cases it's
> definitely confusing. The better option would be to account it as a
> discard and not a write. Preferably in a way that would not break
> existing tools, but so that they could get updated to support it.

It's intentional discard IO start gets accounted, so tools will show
there is IO. I'm not sure if this is better though.

Adding separate columns for discard (maybe flush too) is definitely
preferred. Is breaking existing tools really ok?

Thanks,
Shaohua

Re: [PATCH] blk: don't account discard request size

[Jens Axboe]

On 05/13/2015 11:00 AM, Jeff Moyer wrote:
> Jens Axboe <axboe@fb.com> writes:
>
>> That would be better. But I'm still not sure we want to turn off
>> accounting for discards. For the mixed write/discard cases it's
>> definitely confusing. The better option would be to account it as a
>> discard and not a write. Preferably in a way that would not break
>> existing tools, but so that they could get updated to support it.
>
> Are you suggesting adding a few fields to the end of diskstats or adding
> a new proc file altogether?  (or something else?)

I didn't suggest any specific solution. Obviously it'd be nice if we 
could just extend diskstats, but that might break userland. Worth 
checking up on.

-- 
Jens Axboe

Re: [PATCH] blk: don't account discard request size

[Jens Axboe]

On 05/13/2015 11:22 AM, Shaohua Li wrote:
> On Wed, May 13, 2015 at 10:20:12AM -0400, Jens Axboe wrote:
>> On 05/13/2015 09:10 AM, Jeff Moyer wrote:
>>> Shaohua Li <shli@fb.com> writes:
>>>
>>>> In a workload with discard request, the IO throughput is generally much
>>>> higher than expected. This is quite confusing checking iostat. Discard
>>>> request doesn't really write data to drive, so don't account it.
>>>>
>>>> Signed-off-by: Shaohua Li <shli@fb.com>
>>>> ---
>>>>   block/blk-core.c | 6 +++++-
>>>>   1 file changed, 5 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/block/blk-core.c b/block/blk-core.c
>>>> index fd154b9..0128d18 100644
>>>> --- a/block/blk-core.c
>>>> +++ b/block/blk-core.c
>>>> @@ -2138,7 +2138,11 @@ EXPORT_SYMBOL_GPL(blk_rq_err_bytes);
>>>>
>>>>   void blk_account_io_completion(struct request *req, unsigned int bytes)
>>>>   {
>>>> -	if (blk_do_io_stat(req)) {
>>>> +	/*
>>>> +	 * discard request doesn't really write @bytes to drive,
>>>> +	 * doesn't account it
>>>> +	 **/
>>>> +	if (blk_do_io_stat(req) && !(req->cmd_flags & REQ_DISCARD)) {
>>>>   		const int rw = rq_data_dir(req);
>>>>   		struct hd_struct *part;
>>>>   		int cpu;
>>>
>>> I think you want to modify __get_request to not set REQ_IO_STAT for
>>> discard requests.  This patch will still account the start of I/O, which
>>> means in_flight will be off.
>>
>> That would be better. But I'm still not sure we want to turn off
>> accounting for discards. For the mixed write/discard cases it's
>> definitely confusing. The better option would be to account it as a
>> discard and not a write. Preferably in a way that would not break
>> existing tools, but so that they could get updated to support it.
>
> It's intentional discard IO start gets accounted, so tools will show
> there is IO. I'm not sure if this is better though.
>
> Adding separate columns for discard (maybe flush too) is definitely
> preferred. Is breaking existing tools really ok?

We can't break then, I was just curious if adding a field to the end of 
the diskstats would potentially not break old applications. If not, they 
could just be updated to grab the new field too.

-- 
Jens Axboe

Re: [PATCH] blk: don't account discard request size

[Jeff Moyer]

Jens Axboe <axboe@fb.com> writes:

> On 05/13/2015 11:00 AM, Jeff Moyer wrote:
>> Jens Axboe <axboe@fb.com> writes:
>>
>>> That would be better. But I'm still not sure we want to turn off
>>> accounting for discards. For the mixed write/discard cases it's
>>> definitely confusing. The better option would be to account it as a
>>> discard and not a write. Preferably in a way that would not break
>>> existing tools, but so that they could get updated to support it.
>>
>> Are you suggesting adding a few fields to the end of diskstats or adding
>> a new proc file altogether?  (or something else?)
>
> I didn't suggest any specific solution. Obviously it'd be nice if we
> could just extend diskstats, but that might break userland. Worth
> checking up on.

OK, I didn't know if you had any other tricks up your sleeve.  We've had
luck adding fields to the end of some files in the past (I'd really have
to dig to remember what files), but I'm fairly certain this would break
*something*.  It's definitely safer (albeit uglier) to just add a file.

-Jeff