Re: [RFC][PATCH] convert block trace points to TRACE_EVENT()

[Li Zefan <lizf@cn.fujitsu.com>]

Christoph Hellwig wrote:
> On Tue, May 19, 2009 at 08:59:29AM -0400, Jeff Moyer wrote:
>> I'm not sure I'm comfortable with this.  Having had to diagnose
>> performance problems using blktrace, I find the dev_t information for
>> plugs and unplugs especially interesting.  Maybe we can still figure
>> this out given the context in the trace, though.  Can you comment on
>> that?
> 
> The problem is that right now we still allow request queues shared by
> multiple gendisks, so we can't get at the gendisk.  The blktrace ioctl
> interface gets around that by doing the setup with an ioctl on the
> block device node and storing the dev_t at that point.
> 
> But I don't think we actually have any drivers left sharing a queue
> that way, and I'd love to get rid of that for various reasons.
> 

It will be great if this can be done.

>> from the single letter output to the verbose strings?  Am I just being a
>> curmudgeon?
>>
>> I have one further question, which will expose my ignorance on this
>> implementation.  Would the proposed patch change the information
>> available to the blktrace utility,
> 
> That's the output from /sys/kernel/debug/tracing/trace when the
> block events are enabled.  The blktrace utility should continue to
> work without changes, although it could be ported to use the ftrace
> binary ring buffer directly in the future.
> 

Yes, this patch doesn't affect ioctl-based blktrace or ftrace-plugin
blktrace at all.

To use blktracer:
  # echo 1 > /sys/block/sda/trace/enable
  # echo blk > /debug/tracing/current_tracer
  # cat /debug/tracing/trace_pipe

To use block trace events:
  # echo 'block:*' > /debug/tracing/set_event
  # cat /trace_pipe

You can mixed-use blktracer and trace events, to see how they differ
in the output.