On Wed, May 26, 2021 at 11:19:47PM -0700, Darrick J. Wong wrote: > On Thu, May 27, 2021 at 02:52:02PM +1000, Dave Chinner wrote: > > From: Dave Chinner <dchinner@xxxxxxxxxx> > > > > Ever since we moved to deferred freeing of extents, we only every > > free one extent per transaction. We separated the bulk unmapping of > > extents from the submission of EFI/free/EFD transactions, and hence > > while we unmap extents in bulk, we only every free one per > > transaction. > > > > Our transaction reservations still live in the era from before > > deferred freeing of extents, so still refer to "xfs_bmap_finish" > > and it needing to free multiple extents per transaction. These > > freeing reservations can now all be reduced to a single extent to > > reflect how we currently free extents. > > > > This significantly reduces the reservation sizes for operations like > > truncate and directory operations where they currently reserve space > > for freeing up to 4 extents per transaction. > > > > For a 4kB block size filesytsem with reflink=1,rmapbt=1, the > > reservation sizes change like this: > > > > Reservation Before After > > (index) logres logcount logres logcount > > 0 write 314104 8 314104 8 > > 1 itruncate 579456 8 148608 8 > > 2 rename 435840 2 307936 2 > > 3 link 191600 2 191600 2 > > 4 remove 312960 2 174328 2 > > 5 symlink 470656 3 470656 3 > > 6 create 469504 2 469504 2 > > 7 create_tmpfile 490240 2 490240 2 > > 8 mkdir 469504 3 469504 3 > > 9 ifree 508664 2 508664 2 > > 10 ichange 5752 0 5752 0 > > 11 growdata 147840 2 147840 2 > > 12 addafork 178936 2 178936 2 > > 13 writeid 760 0 760 0 > > 14 attrinval 578688 1 147840 1 > > 15 attrsetm 26872 3 26872 3 > > 16 attrsetrt 16896 0 16896 0 > > 17 attrrm 292224 3 148608 3 > > 18 clearagi 4224 0 4224 0 > > 19 growrtalloc 173944 2 173944 2 > > 20 growrtzero 4224 0 4224 0 > > 21 growrtfree 10096 0 10096 0 > > 22 qm_setqlim 232 1 232 1 > > 23 qm_dqalloc 318327 8 318327 8 > > 24 qm_quotaoff 4544 1 4544 1 > > 25 qm_equotaoff 320 1 320 1 > > 26 sb 4224 1 4224 1 > > 27 fsyncts 760 0 760 0 > > MAX 579456 8 318327 8 > > > > So we can see that many of the reservations have gone substantially > > down in size. itruncate, rename, remove, attrinval and attrrm are > > much smaller now. The maximum reservation size has gone from being > > attrinval at 579456*8 bytes to dqalloc at 318327*8 bytes. This is a > > substantial improvement for common operations. > > If you're going to play around with log reservations, can you have a > quick look at the branch I made to fix all the oversized reservations > that we make for rmap and reflink? > > https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=reflink-speedups > > That's what's next after deferred inode inactivation lands. They all look reasonable at a first pass. I'd need to do more than read the patches to say they are definitely correct, but they certainly don't seem unreasonable to me. I'd also have to run numbers like the table above so I can quantify the reductions, but nothing shouted "don't do this!" to me.... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
