tag name | btree-dynamic-depth-5.16_2021-10-14 (e920beb5ac09bafe33e6efb138f833a272ffc615) |
tag date | 2021-10-14 13:14:35 -0700 |
tagged by | Darrick J. Wong <djwong@kernel.org> |
tagged object | commit 078e7a3a27... |
xfs: support dynamic btree cursor height
In what's left of this series, we rearrange the incore btree cursor so
that we can support btrees of any height. This will become necessary
for realtime rmap and reflink since we'd like to handle tall trees
without bloating the AG btree cursors.
Chandan Babu pointed out that his large extent counters series depends
on the ability to have btree cursors of arbitrary heights, so I've
ported this to 5.15-rc4 so his patchsets won't have to depend on
djwong-dev for submission.
Following the review discussions about the dynamic btree cursor height
patches, I've throw together another series to reduce the size of the
btree cursor, compute the absolute maximum possible btree heights for
each btree type, and now each btree cursor has its own slab cache:
$ grep xfs.*cur /proc/slabinfo
xfs_refcbt_cur 0 0 200 20 1 : tunables 0 0 0 : slabdata 4 4 0
xfs_rmapbt_cur 0 0 248 16 1 : tunables 0 0 0 : slabdata 4 4 0
xfs_bmbt_cur 0 0 248 16 1 : tunables 0 0 0 : slabdata 4 4 0
xfs_inobt_cur 0 0 216 18 1 : tunables 0 0 0 : slabdata 4 4 0
xfs_bnobt_cur 0 0 216 18 1 : tunables 0 0 0 : slabdata 4 4 0
I've also rigged up the debugger to make it easier to extract the actual
height information:
$ xfs_db /dev/sda -c 'btheight -w absmax all'
bnobt: 7
cntbt: 7
inobt: 7
finobt: 7
bmapbt: 9
refcountbt: 6
rmapbt: 9
As you can see from the slabinfo output, this no longer means that we're
allocating 224-byte cursors for all five btree types. Even with the
extra overhead of supporting dynamic cursor sizes and per-btree caches,
we still come out ahead in terms of cursor size for three of the five
btree types.
This series now also includes a couple of patches to reduce holes and
unnecessary fields in the btree cursor.
Patches 1, 5, 11, 12, 13, 16, and 17 still need review.
v2: reduce scrub btree checker memory footprint even more, put the one
fixpatch first, use struct_size, fix 80col problems, move all the
btree cache work to a separate series
v3: rebase to 5.15-rc4, fold in the per-btree cursor cache patches,
remove all the references to "zones" since they're called "caches"
in Linux
v4: improve documentation, clean up some geometry functions, simplify
the max possible height computation code
-----BEGIN PGP SIGNATURE-----
iQIzBAABCgAdFiEEUzaAxoMeQq6m2jMV+H93GTRKtOsFAmFoj6sACgkQ+H93GTRK
tOtkjg//ULsQwkslBXdf/eEEd5i+/bAUp5EFSMZpyc43XbQnzhD4FaHuv9whuai1
u0whbor/eam2FJdQkTtjCW//OO2F1n5IkOdY/euETQU1f0CrJUP8n30yW6W5z8wE
/FtbYeLjfHDLN51v3dXU4VsdedFJ0s3TaolQq7VQlRHerdx2XbmbSkHVFkzR1TAR
3mTEvfPN4RU0OkDTp5OmtTH0bJySOa5H3lV7OUA0pxBE34TX6au3pxTnjj6E/3Qu
xyG+ojgWpw4hMuOi50Q3OaAxoOwDMrcTCC8Fg7GSIO/fj5Qu/e0fkdqYNWbHvIRM
YHxCD6zM/mrg79LrumhwOQgoc9+6UTMKMWOE8LWST91tU8htsGCJgkVxXhgLbEIO
MetxaI47FiGH1DIB094hEeKFZzQu2nBRtnqg05RKAEvEuq8lFEsYUhG90uqaFTou
sQ1wooLdjCQ6WuOI98D7RXN7S2miaC+BmoKuMLH9nJ0Q3j0S9cfy3ZaLtRz51z0g
5kR0pLfey4D1dzGhtsT0nZ/Ay7UqW+lqWDO7qnF7DTSaqsZxFqsx6hoIfr1jJPL3
ede9HZ17oiAQCv9dNhu1JfJ4TpFUvgZjNrJ4zJSIqShF6TWDSQ16KS1v4xh01eB9
OGnFPSeDM6ziswWpyosCJLEGR3Z3+IlZTtTARgbl/4D5mbA/GBg=
=8kAj
-----END PGP SIGNATURE-----