Message ID | 20200506133933.4773-1-jack@suse.cz (mailing list archive) |
---|---|
Headers | show |
Series | Fix blkparse and iowatcher for kernels >= 4.14 | expand |
Does this fix make process names (with PIDS) be shown too in bfq log messages? > Il giorno 6 mag 2020, alle ore 15:39, Jan Kara <jack@suse.cz> ha scritto: > > I was investigating a performance issue with BFQ IO scheduler and I was > pondering why I'm not seeing informational messages from BFQ. After quite > some debugging I have found out that commit 35fe6d763229 "block: use > standard blktrace API to output cgroup info for debug notes" broke standard > blktrace API - namely the informational messages logged by bfq_log_bfqq() > are no longer displayed by blkparse(8) tool. This is because these messages > have now __BLK_TA_CGROUP bit set and that breaks flags checking in > blkparse(1) and iowatcher(1). This series fixes both tools to be able to > cope with events with __BLK_TA_CGROUP flag set. > > Honza
On Wed 06-05-20 15:50:30, Paolo Valente wrote: > Does this fix make process names (with PIDS) be shown too in bfq log messages? No, I didn't change how that's handled. Also AFAICS the kernel does not add PID information to BLK_TN_MESSAGE events (which is how bfq log messages are passed) so blkparse has nothing to display even if it wanted. So poor man's solution is to derive pid from the BFQ queue name and then lookup process name from some other event. Nicer solution would be to modify blktrace handling in the kernel to add PID information to such events as well so we get similar info as when using tracing infrastructure directly. Honza > > Il giorno 6 mag 2020, alle ore 15:39, Jan Kara <jack@suse.cz> ha scritto: > > > > I was investigating a performance issue with BFQ IO scheduler and I was > > pondering why I'm not seeing informational messages from BFQ. After quite > > some debugging I have found out that commit 35fe6d763229 "block: use > > standard blktrace API to output cgroup info for debug notes" broke standard > > blktrace API - namely the informational messages logged by bfq_log_bfqq() > > are no longer displayed by blkparse(8) tool. This is because these messages > > have now __BLK_TA_CGROUP bit set and that breaks flags checking in > > blkparse(1) and iowatcher(1). This series fixes both tools to be able to > > cope with events with __BLK_TA_CGROUP flag set. > > > > Honza >
On Wed, May 06, 2020 at 03:39:30PM +0200, Jan Kara wrote: > I was investigating a performance issue with BFQ IO scheduler and I was > pondering why I'm not seeing informational messages from BFQ. After quite > some debugging I have found out that commit 35fe6d763229 "block: use > standard blktrace API to output cgroup info for debug notes" broke standard > blktrace API - namely the informational messages logged by bfq_log_bfqq() > are no longer displayed by blkparse(8) tool. This is because these messages > have now __BLK_TA_CGROUP bit set and that breaks flags checking in > blkparse(1) and iowatcher(1). This series fixes both tools to be able to > cope with events with __BLK_TA_CGROUP flag set. I'd much rather revert a kernel change that breaks frequently used userspace.
On Wed 06-05-20 07:42:51, Christoph Hellwig wrote: > On Wed, May 06, 2020 at 03:39:30PM +0200, Jan Kara wrote: > > I was investigating a performance issue with BFQ IO scheduler and I was > > pondering why I'm not seeing informational messages from BFQ. After quite > > some debugging I have found out that commit 35fe6d763229 "block: use > > standard blktrace API to output cgroup info for debug notes" broke standard > > blktrace API - namely the informational messages logged by bfq_log_bfqq() > > are no longer displayed by blkparse(8) tool. This is because these messages > > have now __BLK_TA_CGROUP bit set and that breaks flags checking in > > blkparse(1) and iowatcher(1). This series fixes both tools to be able to > > cope with events with __BLK_TA_CGROUP flag set. > > I'd much rather revert a kernel change that breaks frequently used > userspace. We've discussed it [1]. The commit was merged in 4.14 which shows that the breakage is not that severe as nobody noticed until now. Actually I did some more digging in history now and until commit 743210386c "cgroup: use cgrp->kn->id as the cgroup ID" from last November the breakage was only visible if you had CONFIG_BLK_CGROUP enabled and had bio with bi_blkcg set or trace note messages for non-trivial cgroups (effectively only BFQ informational messages). After that commit trace note messages are broken whenever you have CONFIG_BLK_CGROUP. So the reason why nobody noticed is probably because until 5.5 the breakage was actually difficult to hit. So I'm undecided whether at this point it's better to revert the original commit (as likely there is other tooling that depends on this info), just fix the fact that since commit 743210386c all trace note messages will have non-empty cgroup info, or just fixup blkparse and move on... Honza [1] https://lore.kernel.org/r/20200430114711.GA6576@quack2.suse.cz
On 5/6/20 7:39 AM, Jan Kara wrote: > I was investigating a performance issue with BFQ IO scheduler and I was > pondering why I'm not seeing informational messages from BFQ. After quite > some debugging I have found out that commit 35fe6d763229 "block: use > standard blktrace API to output cgroup info for debug notes" broke standard > blktrace API - namely the informational messages logged by bfq_log_bfqq() > are no longer displayed by blkparse(8) tool. This is because these messages > have now __BLK_TA_CGROUP bit set and that breaks flags checking in > blkparse(1) and iowatcher(1). This series fixes both tools to be able to > cope with events with __BLK_TA_CGROUP flag set. Applied, thanks.