On 8/31/18 2:22 PM, Dennis Zhou wrote:

Hi everyone,

This is a split of an earlier series I sent out [1] containing the first
3 patches with fixes from feedback. This series tackles the first
problem where blkcgs were not being destroyed.

There is a regression in blkcg destruction where references weren't
properly put causing blkcgs to never be destroyed. Previously, blkgs
were destroyed during offlining of the blkcg. This puts back the blkcg
reference a blkg holds allowing blkcg ref to reach zero. Then,
blkcg_css_free() is called as part of the final cleanup.

To address the problem, 0001 reverts the broken commit, 0002 delays
blkg destruction until writeback has finished, and 0003 closes the
window on a race condition between a css migration and dying, and
blkg association. This should fix the issue where blkg_get() was getting
called when a blkcg had already begun exiting. If a bio finds itself
here, it will just fall back to root. Oddly enough at one point,
blk-throttle was using policy data from and associating with potentially
different blkgs, thus how this was exposed.

[1] https://lore.kernel.org/lkml/20180831015356.69796-1-dennisszhou@gmail.com/T (local)

This patchset contains the following 3 patches:
  0001-Revert-blk-throttle-fix-race-between-blkcg_bio_issue.patch
  0002-blkcg-delay-blkg-destruction-until-after-writeback-h.patch
  0003-blkcg-use-tryget-logic-when-associating-a-blkg-with-.patch

0001 reverts the broken commit.
0002 delays blkg destruction until after writeback.
0003 fixes a race condition for ongoing IO and blkcg destruction.

Applied for 4.19, thanks Dennis.

-- 
Jens Axboe