DLPX-93371 test_stbtrace_io failed because it stbtrace did not produce any output

[tperkins-perforce]

Background

There'a set of bpftrace files that are failing with blk_mq_end_request for kprobe. This fix is changing the kprobe to use a higher level function using raw tracepoints.

Problem

After upgrading to 24.04, the stbtrace io command fails to generate any output.

Solution

Using raw tracepoints and changing the callback to be traced was changed from blk_mq_start_request to block_io_start, and blk_mq_end_request to block_io_done. The underlying kernel functions did not always call blk_mq_end_request, so the output for the kprobe was not always hit with a corresponding kprobe on blk_mq_start_request.

This way we can get the raw tracepoint of block_io_start/block_io_done, which is always called.

Testing Done

Running fio, iostat, and estat related tools.

Implementation

Changed kprobes to raw tracepoints.

[prakashsurya]

LGTM, thanks!

[sebroy]

This looks great, thanks for persevering through this one. I had a thought while reviewing this for us to consider (perhaps as a next step following the integration of this PR):

It looks like we may have all of the metadata and metrics needed in the disk_io_done raw tracepoint itself without the need to attach a probe to disk_io_start. In the disk_io_done tracepoint, we have the full request structure, and that structure also contains the timestamp when the I/O request started. If we have all we need in the disk_io_done probe, then we could eliminate the overhead of having to attach anything to disk_io_start...

[sebroy]

It looks like there's a typo in the PR summary, the Jira issue referenced doesn't exist. This'll have to be fixed so that the PR checks can pass.