On Fri, Mar 26, 2021 at 11:39:38AM -0400, Brian Foster wrote: > 1. Optimize current append transaction processing with an inode field as > noted above. > > 2. Replace the submission side append transaction entirely with a flag > or some such on the ioend that allocates the transaction at completion > time, but otherwise preserves batching behavior instituted in patch 1. I'm pretty sure I had a patch to kill off the transaction reservation a while ago and Dave objected, mostly in performance grounds in that we might have tons of reservations coming from the I/O completion workqueue that would all be blocked on the transaction reservations. This would probably be much better with the ioend merging, which should significantly reduce the amount of transactions reservations - to probably not more than common workloads using unwritten extents. I'm all for killing the transaction pre-reservations as they have created a lot of pain for us.
