BugLink: https://bugs.launchpad.net/bugs/1898786
[Impact]
Systems that utilise bcache can experience extremely high IO wait times when under constant IO pressure. The IO wait times seem to stay at a consistent 1 second, and never drop as long as the bcache shrinker is enabled.
If you disable the shrinker, then IO wait drops significantly to normal levels.
We did some perf analysis, and it seems we spend a huge amount of time in bch_mca_scan(), likely waiting for the mutex "&c->bucket_lock".
Looking at the recent commits in Bionic, we found the following commit merged in upstream 5.1-rc1 and backported to 4.15.0-87-generic through upstream stable:
commit 9fcc34b1a6dd4b8e5337e2b6ef45e428897eca6b
Author: Coly Li <email address hidden>
Date: Wed Nov 13 16:03:24 2019 +0800
Subject: bcache: at least try to shrink 1 node in bch_mca_scan()
Link: https://github.com/torvalds/linux/commit/9fcc34b1a6dd4b8e5337e2b6ef45e428897eca6b
It mentions in the description that:
> If sc->nr_to_scan is smaller than c->btree_pages, after the above
> calculation, variable 'nr' will be 0 and nothing will be shrunk. It is
> frequeently observed that only 1 or 2 is set to sc->nr_to_scan and make
> nr to be zero. Then bch_mca_scan() will do nothing more then acquiring
> and releasing mutex c->bucket_lock.
This seems to be what is going on here, but the above commit only addresses when nr is 0.
From what I can see, the problems we are experiencing are when nr is 1 or 2, and again, we just waste time in bch_mca_scan() waiting on c->bucket_lock, only to release c->bucket_lock since the shrinker loop never executes since there is no work to do.
[Fix]
The following commits fix the problem, and all landed in 5.6-rc1:
commit 125d98edd11464c8e0ec9eaaba7d682d0f832686
Author: Coly Li <email address hidden>
Date: Fri Jan 24 01:01:40 2020 +0800
Subject: bcache: remove member accessed from struct btree
Link: https://github.com/torvalds/linux/commit/125d98edd11464c8e0ec9eaaba7d682d0f832686
commit d5c9c470b01177e4d90cdbf178b8c7f37f5b8795
Author: Coly Li <email address hidden>
Date: Fri Jan 24 01:01:41 2020 +0800
Subject: bcache: reap c->btree_cache_freeable from the tail in bch_mca_scan()
Link: https://github.com/torvalds/linux/commit/d5c9c470b01177e4d90cdbf178b8c7f37f5b8795
commit e3de04469a49ee09c89e80bf821508df458ccee6
Author: Coly Li <email address hidden>
Date: Fri Jan 24 01:01:42 2020 +0800
Subject: bcache: reap from tail of c->btree_cache in bch_mca_scan()
Link: https://github.com/torvalds/linux/commit/e3de04469a49ee09c89e80bf821508df458ccee6
The first commit is a dependency of the other two. The first commit removes a "recently accessed" marker, used to indicate if a particular cache has been used recently, and if it has been, not consider it for cache eviction. The commit mentions that under heavy IO, all caches will end up being recently accessed, and nothing will ever be shrunk.
The second commit changes a previous design decision of skipping the first 3 caches to shrink, since it is a common case to call bch_mca_scan() with nr being 1, or 2, just like 0 was common in the very first commit I mentioned. This time, if we land on 1 or 2, the loop exits and nothing happens, and we waste time waiting on locks, just like the very first commit I mentioned. The fix is to try shrink caches from the tail of the list, and not the beginning.
The third commit fixes a minor issue where we don't want to re-arrange the linked list c->btree_cache, which is what the second commit ended up doing, and instead, just shrink the cache at the end of the linked list, and not change the order.
One minor backport / context adjustment was required in the first commit for Bionic, and the rest are all clean cherry picks to Bionic and Focal.
[Testcase]
This is kind of hard to test, since the problem shows up in production environments that are under constant IO pressure, with many different items entering and leaving the cache.
The Launchpad git server is currently suffering this issue, and has been sitting at a constant IO wait of 1 second / slightly over 1 second which was causing slow response times, which was leading to build failures when git clones ended up timing out.
We installed a test kernel, which is available in the following PPA:
https://launchpad.net/~mruffell/+archive/ubuntu/sf294907-test
Once the test kernel had been installed, IO wait times with the shrinker enabled dropped to normal levels, and the git server became responsive again. We have been monitoring the performance of the git server and watching IO wait times in grafana over the past week, and everything is looking good, and indicate that these patches solve the issue.
[Regression Potential]
If a regression were to occur, it would only affect users of bcache who have the shrinker enabled, which I believe is the default setting in bcache deployments.
In that case, users could disable the shrinker as a workaround until a fix is available. You can disable the shrinker with:
"sudo bash -c "echo 1 > /sys/fs/bcache/<uuid>/internal/btree_shrinker_disabled""
The patches remove the "recently accessed" marker, which changes the behaviour slightly when it comes to deciding what to evict from the cache. Some workloads that aren't under IO pressure may see a difference in what items are evicted from the cache, which may slightly change bcache's performance profile.
Since the patches prune the cache from the tail of the cache linked list, we will likely end up pruning more nodes than previously, since it was common to land on the first 3 caches, which were hard coded to be ignored and not pruned, which may also slightly change the bcache's performance profile.
This change was made by a bot.