Hello All, Please find the RFCv3 patchset which adds support for iomap subpage dirty state tracking which improves write performance and should reduce the write amplification problem on platforms with smaller filesystem blocksize compared to pagesize. E.g. On Power with 64k default pagesize and with 4k filesystem blocksize. RFCv2 -> RFCv3 =============== 1. Addressed review comments on adding accessor APIs for both uptodate and dirty iop bitmap. (todo-1 of rfcv2). Addressed few other review comments from Christoph & Matthew. 2. Performance testing of these patches reveal the same performance improvement i.e. the given fio workload shows 16x perf improvement on nvme drive. (completed todo-3 of rfcv2) 3. Addressed todo-4 of rfcv2 Few TODOs =========== 1. Test gfs2 and zonefs with these changes (todo-2 of rfcv2) 2. Look into todo-5 of rfcv2 xfstests testing with default options and 1k blocksize on x86 reveals no new issues. Also didn't observe any surprises on Power with 4k blocksize. (Please do suggest if there are any specific xfstests config options (for xfs) which are good to get it tested for this patch series?) Copy-Paste Cover letter of RFCv2 ================================ RFC -> RFCv2 ============= 1. One of the key fix in v2 is that earlier when the folio gets marked as dirty, we were never marking the bits dirty in iop bitmap. This patch adds support for iomap_dirty_folio() as new ->dirty_folio() aops callback, which sets the dirty bitmap in iop and later call filemap_dirty_folio(). This was one of the review comment that was discussed in RFC. 2. One of the other key fix identified in testing was that iop structure could get allocated at the time of the writeback if the folio is uptodate. (since it can get freed during memory pressure or during truncate_inode_partial_folio() in case of large folio). This could then cause nothing to get written if we have not marked the necessary bits as dirty in iop->state[]. Patch-1 & Patch-3 takes care of that. TODOs ====== 1. I still need to work on macros which we could declare and use for easy reference to uptodate/dirty bits in iop->state[] bitmap (based on previous review comments). 2. Test xfstests on other filesystems which are using the iomap buffered write path (gfs2, zonefs). 3. Latest performance testing with this patch series (I am not expecting any surprises here. The perf improvements should be more or less similar to rfc). 4. To address one of the todo in Patch-3. I think I missed to address it and noticed it only now before sending. But it should be easily addressable. I can address it in the next revision along with others. 5. To address one of the other review comments like what happens with a large folio. Can we limit the size of bitmaps if the folio is too large e.g. > 2MB. [RH] - I can start looking into this area too, if we think these patches are looking good. My preference would be to work on todos 1-4 as part of this patch series and take up bitmap optimization as a follow-up work for next part. Please do let me know your thoughts and suggestions on this. Note: I have done a 4k bs test with auto group on Power with 64k pagesize and I haven't found any surprises. I am also running a full bench of all tests with x86 and 1k blocksize, but it still hasn't completed. I can update the results once it completes. Also as we discussed, all the dirty and uptodate bitmap tracking code for iomap_page's state[] bitmap, is still contained within iomap buffered-io.c file. I would appreciate any review comments/feedback and help on this work i.e. adding subpage size dirty tracking to reduce write amplification problem and improve buffered write performance. Kindly note that w/o these patches, below type of workload gets severly impacted. Performance Results from RFC [1]: ================================= 1. Performance testing of below fio workload reveals ~16x performance improvement on nvme with XFS (4k blocksize) on Power (64K pagesize) FIO reported write bw scores, improved from ~28 MBps to ~452 MBps. <test_randwrite.fio> [global] ioengine=psync rw=randwrite overwrite=1 pre_read=1 direct=0 bs=4k size=1G dir=./ numjobs=8 fdatasync=1 runtime=60 iodepth=64 group_reporting=1 [fio-run] 2. Also our internal performance team reported that this patch improves there database workload performance by around ~83% (with XFS on Power) [1]: https://lore.kernel.org/linux-xfs/cover.1666928993.git.ritesh.list@xxxxxxxxx/ Ritesh Harjani (IBM) (3): iomap: Allocate iop in ->write_begin() early iomap: Change uptodate variable name to state iomap: Support subpage size dirty tracking to improve write performance fs/gfs2/aops.c | 2 +- fs/iomap/buffered-io.c | 166 ++++++++++++++++++++++++++++++++++++----- fs/xfs/xfs_aops.c | 2 +- fs/zonefs/super.c | 2 +- include/linux/iomap.h | 1 + 5 files changed, 150 insertions(+), 23 deletions(-) -- 2.39.2
