Christoph Hellwig [off-list ref] 于2020年3月31日周二 下午5:16写道：

On Tue, Mar 31, 2020 at 04:45:33PM +0800, Weiping Zhang wrote:

quoted

Christoph Hellwig [off-list ref] 于2020年3月31日周二 下午4:25写道：

quoted

On Fri, Mar 27, 2020 at 02:28:59PM +0800, Weiping Zhang wrote:

quoted

Change-Id: Ibb9caf20616f83e111113ab5c824c05930c0e523
Signed-off-by: Weiping Zhang <redacted>

This needs a commit description and loose the weird change id.

OK, I rewirte commit description, it record the timestamp of issue bio
to the disk driver,
then we can get the delta time in rq_qos_done_bio. It's same as the D2C time
of blktrace.

quoted

I also think oyu need to fins a way to not bloat the bio even more,
cgroup is a really bad offender for bio size.

struct request {
    u64 io_start_time_ns;
also record this timestamp, I'll check if we can use it.

But except for a few exceptions bios are never issued directly to the
driver, requests are.  And the few exception (rsxx, umem) probably should
be rewritten to use requests.  And with generic_{start,end}_io_acct we
already have helpers to track bio based stats, which we should not
duplicate just for cgroups.

generic_{start,end}_io_acct and blk_account_io_done,
these two method use the a timeline (part->stamp), but cgroup doesn't have,
so cgroup cann't use these general helper to counting the total io ticks
for read,write and others. Block cgroup use delta = now -
bio->bi_issue[issue_time]
to counting total io ticks.

How about move it into the blk-iotrack code, rq_qos_issue will call the
rq_qos_ops.issue,  then if user doesn't enable blk-iotrack, these code
will not be executed.

Thanks