On Wed, May 13, 2020 at 02:15:19PM -0700, Andrew Morton wrote: > On Tue, 12 May 2020 17:29:36 -0400 Johannes Weiner <hannes@xxxxxxxxxxx> wrote: > > > > > ... > > > > Solution > > > > This patch fixes the aging inversion described above on > > !CONFIG_HIGHMEM systems, without reintroducing the problems associated > > with excessive shrinker LRU rotations, by keeping populated inodes off > > the shrinker LRUs entirely. > > > > Currently, inodes are kept off the shrinker LRU as long as they have > > an elevated i_count, indicating an active user. Unfortunately, the > > page cache cannot simply hold an i_count reference, because unlink() > > *should* result in the inode being dropped and its cache invalidated. > > > > Instead, this patch makes iput_final() consult the state of the page > > cache and punt the LRU linking to the VM if the inode is still > > populated; the VM in turn checks the inode state when it depopulates > > the page cache, and adds the inode to the LRU if necessary. > > > > This is not unlike what we do for dirty inodes, which are moved off > > the LRU permanently until writeback completion puts them back on (iff > > still unused). We can reuse the same code -- inode_add_lru() - here. > > > > This is also not unlike page reclaim, where the lower VM layer has to > > negotiate state with the higher VFS layer. Follow existing precedence > > and handle the inversion as much as possible on the VM side: > > > > - introduce an I_PAGES flag that the VM maintains under the i_lock, so > > that any inode code holding that lock can check the page cache state > > without having to lock and inspect the struct address_space > > Maintaining the same info in two places is a hassle. Is this > optimization worthwhile? Hm, maybe not. I'll try to get rid of it and test cache / LRU state directly. > > - introduce inode_pages_set() and inode_pages_clear() to maintain the > > inode LRU state from the VM side, then update all cache mutators to > > use them when populating the first cache entry or clearing the last > > > > With this, the concept of "inodesteal" - where the inode shrinker > > drops page cache - is relegated to CONFIG_HIGHMEM systems only. The VM > > is in charge of the cache, the shrinker in charge of struct inode. > > How tested is this on highmem machines? I don't have a highmem machine, but my code is ifdeffed out on CONFIG_HIGHMEM so the behavior shouldn't have changed there.