On Tue, Apr 28, 2020 at 04:45:57PM -0700, Darrick J. Wong wrote: > On Tue, Apr 28, 2020 at 03:41:32PM -0700, Darrick J. Wong wrote: > > On Sat, Apr 25, 2020 at 11:28:01AM -0700, Christoph Hellwig wrote: > > > On Tue, Apr 21, 2020 at 07:07:13PM -0700, Darrick J. Wong wrote: > > > > From: Darrick J. Wong <darrick.wong@xxxxxxxxxx> > > > > > > > > Move the extent free intent and intent-done log recovery code into the > > > > per-item source code files and use dispatch functions to call them. We > > > > do these one at a time because there's a lot of code to move. No > > > > functional changes. > > > > > > What is the reason for splitting xlog_recover_item_type vs > > > xlog_recover_intent_type? To me it would seem more logical to have > > > one operation vector, with some ops only set for intents. > > > > Partly because I started by refactoring only the intent items, and then > > decided to prepend a series to do everything; and partly to be stingy > > with bytes. :P > > > > That said, I like your suggestion of every XFS_LI_* code gets its own > > xlog_recover_item_type so I'll go do that. > > Aha, now I remember why those two are separate types -- the > process_intent and cancel_intent functions operate on the xfs_log_item > that gets created from the xlog_recover_item that we pulled out of the > log, whereas the other functions are called directly on the > xlog_recovery_item. There's no direct link between the log item and the > recovery log item, nor is there a good way to link through their > dispatch functions. Maybe those should move to xfs_item_ops as they operate on a "live" xfs_log_item? (they'd need to grow names clearly related to recovery of course). In fact except for slightly different calling convention ->cancel_intent already seems to be identical to ->abort_intent in xfs_item_ops, so that would be one off the list. Btw, it seems like we should drop the ail_lock before calling ->process_intent as all instances do that anyway, and it would keep the locking a little more centralized, and it will allow killing one pointless wrapper in each instance. Maybe we can also move the recovered flag to the generic log item flags?
