Messages in this thread | | | Date | Sun, 8 Jan 2012 17:49:49 -0800 | From | Tejun Heo <> | Subject | Re: [PATCH 03/11] block: block_bio_complete tracepoint was missing |
| |
Hello,
On Mon, Jan 09, 2012 at 10:30:06AM +0900, Namhyung Kim wrote: > Just adding the TP unconditionally will produce duplicated (in some > sense) reports about the completion. For example, normal request > based IO reports whole request completion prior to its bio's, and > further
Request and bio completions are separate events. There's nothing wrong with reporting them separately. In fact, I think they should be reported separately.
>, some of nested block IO handling routines - bounced bio and > btrfs with compression, etc - call bio_endio() more than once. Also > there are cases that bio fails before it's enqueued for some reason.
They are actually separate bio's being completed. I don't think trying to put extra semantics on TP itself is a good idea. In general, TP signals that such event happened with sufficient information and it's the consumers' responsibility to make sense of what's going on. BIO_CLONED/BOUNCED are there.
> I have no idea about the ioblame can deal with all of such corner > cases. However it might confuse blktrace somewhat, I guess.
Unless someone is doing memcpy() on bio's, ioblame should be okay. It only considers bio's which went through actual submission.
> I already posted similar patch a couple of weeks ago, but didn't > receive a comment yet. [1] Please take a look this too :)
I'll reply there but don't think imposing such extra logic on TP is a good idea.
> After a quick glance, the ioblame seems to carry all IO accounting > info through the first bio in the request. If so, why don't you use > the request structure for that?
Because there are bio based drivers which don't use requests at all.
Thanks.
-- tejun
| |