On Mon, Sep 12, 2016 at 10:56:04AM -0400, Josef Bacik wrote: > I think that looping through all the sb's in the system would be > kinda shitty for this tho, we want the "get number of dirty pages" > part to be relatively fast. What if I do something like the > shrinker_control only for dirty objects. So the fs registers some > dirty_objects_control, we call into each of those and get the counts > from that. Does that sound less crappy? Thanks, Hmmm - just an off-the-wall thought on this.... If you're going to do that, then why wouldn't you simply use a "shrinker" to do the metadata writeback rather than having a hook to count dirty objects to pass to some other writeback code that calls a hook to write the metadata? That way filesystems can also implement dirty accounting and "writers" for each cache of objects they currently implement shrinkers for. i.e. just expanding shrinkers to be able to "count dirty objects" and "write dirty objects" so that we can tell filesystems to write back all their different metadata caches proportionally to the size of the page cache and it's dirty state. The existing file data and inode writeback could then just be new generic "superblock shrinker" operations, and the fs could have it's own private metadata writeback similar to the private sb shrinker callout we currently have... And, in doing so, we might be able to completely hide memcg from the writeback implementations similar to the way memcg is completely hidden from the shrinker reclaim implementations... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>