[PATCH 00/11] fix hungtask due to repeated traversal of inodes list

[Mateusz Guzik]

On Mon, Nov 18, 2024 at 07:44:57PM +0800, Ye Bin wrote:
> From: Ye Bin <yebin10 at huawei.com>
> 
> As commit 04646aebd30b ("fs: avoid softlockups in s_inodes iterators")
> introduces the retry logic. In the problem environment, the 'i_count'
> of millions of files is not zero. As a result, the time slice for each
> traversal to the matching inode process is almost used up, and then the
> traversal is started from scratch. The worst-case scenario is that only
> one inode can be processed after each wakeup. Because this process holds
> a lock, other processes will be stuck for a long time, causing a series
> of problems.
> To solve the problem of repeated traversal from the beginning, each time
> the CPU needs to be freed, a cursor is inserted into the linked list, and
> the traversal continues from the cursor next time.
> 

I'm surprised this was not sorted out eons ago.

My non-maintainer $0,03 is as follows:

Insertion of a marker is indeed the idiomatic way to fix this, but I
disagree with the specific implementation.

All cond_resched() handling should be moved into the magic iteration
machinery, instead of random consumers open-coding it.