On 6/23/20 6:48 PM, Artur Paszkiewicz wrote:
On 6/23/20 4:21 PM, Guoqing Jiang wrote:
Hi Artur,
On 6/8/20 9:13 AM, Artur Paszkiewicz wrote:
On 6/5/20 10:19 PM, jeffm@xxxxxxxx wrote:
The i/o accounting published in /proc/diskstats for mdraid is currently
broken. md_make_request does the accounting for every bio passed but
when a bio needs to be split, all the split bios are also submitted
through md_make_request, resulting in multiple accounting.
Hi Jeff,
I sent a patch a few days ago which should fix this issue. Can you check
it out?
https://marc.info/?l=linux-raid&m=159102814820539
I need to account some extra statistics for bio such as latency and size,
so it is kind of relies on your patch, then I read the code again.
And besides my previous comment. I think you don't need clone bio for all
personalities. For md-multipath, raid1 and raid10, you can track the start
time by add it to those structures (multipath_bh, r1bio and r10bio), then
one extra copy could be avoided.
What do you think?
You're right, cloning can be avoided for those personalities. I wanted
to keep it clean and centralized in the generic md code. I think we
should have a common completion callback and some common request
structure for all md personalities, like struct md_io which I'm using in
my patch.How about we add it to the existing stucts like r1bio etc. and
use that for io accounting, and clone the bio with the struct otherwise?
Or maybe remove the cloning from the personalities and instead make the
bio cloned in md_make_request available to them? Does this make sense?
Centralization definitely has it's advantage, but given the difference
among personalities, not sure it is possible or make sense to clone bio
in common md layer.
For raid1/10, we at least need to care about barrier and read balance
before clone bio, and I would prefer to not change current mechanism
since it is really a big modification from my understanding.
Thanks,
Guoqing
