In the Linux kernel, the following vulnerability has been resolved:
btrfs: do not strictly require dirty metadata threshold for metadata writepages
[BUG] There is an internal report that over 1000 processes are waiting at the io_schedule_timeout() of balance_dirty_pages(), causing a system hang and trigger a kernel coredump.
The kernel is v6.4 kernel based, but the root problem still applies to any upstream kernel before v6.18.
[CAUSE] From Jan Kara for his wisdom on the dirty page balance behavior first.
This cgroup dirty limit was what was actually playing the role here because the cgroup had only a small amount of memory and so the dirty limit for it was something like 16MB.
Dirty throttling is responsible for enforcing that nobody can dirty (significantly) more dirty memory than there's dirty limit. Thus when a task is dirtying pages it periodically enters into balance_dirty_pages() and we let it sleep there to slow down the dirtying.
When the system is over dirty limit already (either globally or within a cgroup of the running task), we will not let the task exit from balance_dirty_pages() until the number of dirty pages drops below the limit.
So in this particular case, as I already mentioned, there was a cgroup with relatively small amount of memory and as a result with dirty limit set at 16MB. A task from that cgroup has dirtied about 28MB worth of pages in btrfs btree inode and these were practically the only dirty pages in that cgroup.
So that means the only way to reduce the dirty pages of that cgroup is to writeback the dirty pages of btrfs btree inode, and only after that those processes can exit balance_dirty_pages().
Now back to the btrfs part, btree_writepages() is responsible for writing back dirty btree inode pages.
The problem here is, there is a btrfs internal threshold that if the btree inode's dirty bytes are below the 32M threshold, it will not do any writeback.
This behavior is to batch as much metadata as possible so we won't write back those tree blocks and then later re-COW them again for another modification.
This internal 32MiB is higher than the existing dirty page size (28MiB), meaning no writeback will happen, causing a deadlock between btrfs and cgroup:
- Btrfs doesn't want to write back btree inode until more dirty pages
- Cgroup/MM doesn't want more dirty pages for btrfs btree inode
Thanks Jan Kara a lot for the analysis of the root cause.
[ENHANCEMENT] Since kernel commit b55102826d7d ("btrfs: set AS_KERNEL_FILE on the btree_inode"), btrfs btree inode pages will only be charged to the root cgroup which should have a much larger limit than btrfs' 32MiB threshold. So it should not affect newer kernels.
But for all current LTS kernels, they are all affected by this problem, and backporting the whole AS_KERNEL_FILE may not be a good idea.
Even for newer kernels I still think it's a good idea to get rid of the internal threshold at btree_writepages(), since for most cases cgroup/MM has a better view of full system memory usage than btrfs' fixed threshold.
For internal callers using btrfs_btree_balance_dirty() since that function is already doing internal threshold check, we don't need to bother them.
But for external callers of btree_writepages(), just respect their requests and write back whatever they want, ignoring the internal btrfs threshold to avoid such deadlock on btree inode dirty page balancing.