Re: BLK_TN_PROCESS events not delivered for all devices

[Jens Axboe <axboe@kernel.dk>]

On 09/17/2013 11:10 AM, Jan Kara wrote:
> On Tue 17-09-13 08:29:07, Jens Axboe wrote:
>> On 09/16/2013 03:23 AM, Jan Kara wrote:
>>>   Hi,
>>>
>>>   I've been looking into a problem where BLK_TN_PROCESS events are not
>>> delivered to all devices which are being traced. This results in process
>>> name being (null) when trace for a single device is parsed.
>>>
>>> The reason for this problem is that trace_note_tsk() is called only if
>>> tsk->btrace_seq != blktrace_seq and it updates tsk->btrace_seq to
>>> blktrace_seq. Thus after a trace for another device is started
>>> BLK_TN_PROCESS event is sent only on behalf of the first device with which
>>> the task interacts. That isn't necessarily the new device thus traces for
>>> some devices accumulate several BLK_TN_PROCESS events for one task while
>>> other have none. Is this a known problem and is this intended to work
>>> better?
>>>
>>> I was thinking how to fix that for a while and it doesn't seem to be
>>> possible without tracking with each block trace which tasks it has been
>>> notified about. And that is relatively expensive...
>>
>> It is unfortunately a known issue... I have not come up with a good way
>> to fix it either, while keeping it cheap. So if you think of something,
>> do let me know.
>   Hum... How about linking all running block traces (struct blk_trace) in a
> linked list and sending BLK_TN_PROCESS to all the traces? Sure we will be
> spamming with BLK_TN_PROCESS a bit but starting a trace isn't such a common
> thing so it shouldn't be too bad. What do you think?

That might be good enough. I'm not worried about start/stop type
expenses, those things generally don't matter. And the list wont add any
fast path overhead when tracing.

-- 
Jens Axboe