On Wed, Apr 29, 2020 at 04:58:19PM -0700, Darrick J. Wong wrote: > On Wed, Apr 29, 2020 at 07:55:57AM -0700, Christoph Hellwig wrote: > > On Wed, Apr 29, 2020 at 07:28:07AM -0700, Darrick J. Wong wrote: > > > Hmm. Actually now that I think harder about it, the bmap item is > > > completely incore and fields are selectively copied to the log item. > > > This means that regular IO could set bi_owner = <some inode number> and > > > bi_ip = <the incore inode>. Recovery IO can set bi_owner but leave > > > bi_ip NULL, and then the bmap item replay can iget as needed. Now we > > > don't need this freeze/thaw thing at all. > > > > Yes, that sounds pretty reasonable. > > OTOH I was talking to Dave and we decided to try the simpler solution of > retaining the inode reference unless someone can show that the more > complex machinery is required. Sounds good. And we only need to do it when in recovery, this should save a few atomic ops during normal operations.
