On Fri, Feb 18, 2022 at 3:20 PM HAGIO KAZUHITO(萩尾 一仁) <k-hagio-ab@xxxxxxx> wrote:
Hi Lianbo,
-----Original Message-----
> -----Original Message-----
> > Admittedly, I haven't looked into the details, but those simple numbers for
> > pending IO have been very valuable for me, when analyzing dumps with
> > dm-multipath over (FC-attached) SCSI disks. Aren't those pending IO numbers
> > available elsewhere in the kernel? Maybe not in debugfs anymore, but I suppose
> > the (mq) block layer does keep track of them?
>
> AFAIK, unfortunately there is no simple counter, crash needs to parse sbitmap
> structures to count them. (please let us know if there is a simpler solution.)
> Currently Sergey has worked on adding sbitmap function [1], and we plan to make
> use of it after merging the patch. It may take some time, so Lianbo's patch is
> a temporary workaround for the error.
>
> In addition, I would like to add that the counter for the ctx->rq_completed/ctx->rq_dispatched(for 'dev
> -d') may become inaccurate due to racing per-cpu values in the kernel. I'm considering whether crash should
> print a warning or "not supported" for "dev -d", which can avoid misleading users. So far we have encountered
> two similar cases.
>
> What's your opinion?Kazu.
Thank you for raising this up.
The counters can have been inaccurate with multiple CPUs since the
introduction of them in kernel, right? No other conditions?
There seems to be no other conditions.
Hmm, in this case, personally I would suggest to print a warning on
machines with multiple cpus.
btw, is there any upstream patch or email saying about the inaccuracy?
No, I didn't see it, and only saw it in internal bugs.
It would be better to add it to a commit log or/and code comment.
That would be good if anyone knows about the related discussion.
Thanks.
Lianbo
Thanks,
Kazu
-- Crash-utility mailing list Crash-utility@xxxxxxxxxx https://listman.redhat.com/mailman/listinfo/crash-utility
