From: Patrick Farrell Date: Tue, 15 Jun 2021 14:23:04 +0000 (-0400) Subject: LU-13799 osc: Improve osc_queue_sync_pages X-Git-Tag: 2.14.53~14 X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=commitdiff_plain;h=87c4535f7a5d239aad4e936545a72d0199ccd9ba LU-13799 osc: Improve osc_queue_sync_pages This patch was split and partially done in: https://review.whamcloud.com/38214 So the text below refers to the combination of this patch and that one. This patch now just improves a looped atomic add by replacing with a single one. The rest of the grant calcuation change is in https://review.whamcloud.com/38214 (I am retaining the text below to show the performance improvement) ---------- osc_queue_sync_pages now has a grant calculation component, this has a pretty painful impact on the new faster DIO performance. Specifically, per page ktime_get() and the per-page atomic_add cost close to 10% of total CPU time in the DIO path. We can make this per batch of pages rather than for each page, which reduces this cost from 10% of CPU to almost nothing. This improves write performance by about 10% (but has no effect on reads, since they don't use grant). This patch reduces i/o time in ms/GiB by: Write: 10 ms/GiB Read: 0 ms/GiB Totals: Write: 158 ms/GiB Read: 161 ms/GiB mpirun -np 1 $IOR -w -t 1G -b 64G -o $FILE --posix.odirect Before patch: write 6071 After patch: write 6470 (Read is similar.) This also fixes a mistake in c24c25dc1b / LU-13419 where it removed the shrink interval update entirely from the direct i/o path. Fixes: c24c25dc1b ("LU-13419 osc: Move shrink update to per-write") Signed-off-by: Patrick Farrell Change-Id: Ic606e03be58239c291ec0382fa89eba64560da53 Reviewed-on: https://review.whamcloud.com/39482 Reviewed-by: Andreas Dilger Reviewed-by: Wang Shilong Tested-by: jenkins Tested-by: Maloo Reviewed-by: Oleg Drokin --- diff --git a/lustre/osc/osc_cache.c b/lustre/osc/osc_cache.c index 658e8f4..c33cfae 100644 --- a/lustre/osc/osc_cache.c +++ b/lustre/osc/osc_cache.c @@ -2634,8 +2634,8 @@ int osc_queue_sync_pages(const struct lu_env *env, struct cl_io *io, list_for_each_entry(oap, list, oap_pending_item) { osc_consume_write_grant(cli, &oap->oap_brw_page); - atomic_long_inc(&obd_dirty_pages); } + atomic_long_add(page_count, &obd_dirty_pages); osc_unreserve_grant_nolock(cli, grants, 0); ext->oe_grants = grants; } else { @@ -2649,6 +2649,7 @@ int osc_queue_sync_pages(const struct lu_env *env, struct cl_io *io, "not enough grant available, switching to sync for this i/o\n"); } spin_unlock(&cli->cl_loi_list_lock); + osc_update_next_shrink(cli); } ext->oe_is_rdma_only = !!(brw_flags & OBD_BRW_RDMA_ONLY);