On Thu, 2025-04-03 at 18:12 +0100, Matthew Wilcox wrote: [...] > Ideas still on the table: > > - Convert all filesystems to use the XFS inode management scheme. > Nobody is thrilled by this large amount of work. > - Find a simpler version of the XFS scheme to implement for other > filesystems. What's wrong with a simpler fix: if we're in PF_MEMALLOC when we try to run inode.c:evict(), send it through a workqueue? It will require some preallocation (say using a superblock based work entry ... or simply reuse the destroy_work) but it should be doable. The analysis says that evicting from reclaim is very rare because it's a deletion race, so it shouldn't matter that it's firing once per inode with this condition. Regards, James
