On Mon, Apr 20, 2020 at 08:53:06AM +1000, Dave Chinner wrote: > On Fri, Apr 17, 2020 at 11:08:47AM -0400, Brian Foster wrote: > > Hi all, > > > > This actually started as what I intended to be a cleanup of xfsaild > > error handling and the fact that unexpected errors are kind of lost in > > the ->iop_push() handlers of flushable log items. Some discussion with > > Dave on that is available here[1]. I was thinking of genericizing the > > behavior, but I'm not so sure that is possible now given the error > > handling requirements of the associated items. > > > > While thinking through that, I ended up incorporating various cleanups > > in the somewhat confusing and erratic error handling on the periphery of > > xfsaild, such as the flush handlers. Most of these are straightforward > > cleanups except for patch 9, which I think requires careful review and > > is of debatable value. I have used patch 12 to run an hour or so of > > highly concurrent fsstress load against it and will execute a longer run > > over the weekend now that fstests has completed. > > > > Thoughts, reviews, flames appreciated. > > I'll need to do something thinking on this patchset - I have a > patchset that touches a lot of the same code I'm working on right > now to pin inode cluster buffers in memory when the inode is dirtied > so we don't get RMW cycles in AIL flushing. > > That code gets rid of xfs_iflush() completely, removes dirty inodes > from the AIL and tracks only ordered cluster buffers in the AIL for > inode writeback (i.e. reduces AIL tracked log items by up to 30x). > It also only does inode writeback from the ordered cluster buffers. > Ok. I could see that being reason enough to drop the iflush iodone patch, given that it depends on a bit of a rework/hack. A cleaner solution requires more thought and it might not be worth the time if the code is going away. Most of the rest are straightforward cleanups though so I wouldn't expect complex conflict resolution. It's hard to say for sure without seeing the code, of course.. > The idea behind this is to make inode flushing completely > non-blocking, and to simply inode cluster flushing to simply iterate > all the dirty inodes attached to the buffer. This gets rid of radix > tree lookups and races with reclaim, and gets rid of having to > special case a locked inode in the cluster iteration code. > Sounds interesting, but it's not really clear to me what the general flushing dynamic looks like in this model. I.e., you mention xfs_iflush() goes away, but cluster flushing still exists in some form, so I can't really tell if xfs_iflush() going away is tied to a functional change or primarily a refactoring/cleanup. Anyways, no need to go into the weeds if the code will eventually clarify.. > I was looking at this as the model to then apply to dquot flushing, > too, because it currently does not have cluster flushing, and hence > flushes dquots individually, even though there can be multiple dirty > dquots per buffer. Some of this patchset moves the dquot flushing a > bit closer to the inode code, so those parts are going to be useful > regardless of everything else.... > Makes sense. > Do you have a git tree I could pull this from to see how bad the > conflicts are? > I don't have a public tree. I suppose I could look into getting kernel.org access if somebody could point me in the right direction for that. :) In the meantime I could make a private tree accessible to you directly if that's helpful.. Brian > Cheers, > > Dave. > -- > Dave Chinner > david@xxxxxxxxxxxxx >
