Date: Fri, 13 Oct 2023 23:53:55 GMT From: Mateusz Guzik <mjg@FreeBSD.org> To: src-committers@FreeBSD.org, dev-commits-src-all@FreeBSD.org, dev-commits-src-branches@FreeBSD.org Subject: git: 12e81504ef7a - stable/13 - vfs: prefix regular vnlru with a special case for free vnodes Message-ID: <202310132353.39DNrtJU032274@gitrepo.freebsd.org>
next in thread | raw e-mail | index | archive | help
The branch stable/13 has been updated by mjg: URL: https://cgit.FreeBSD.org/src/commit/?id=12e81504ef7aa1802c8ede384f07ae65fec48bdc commit 12e81504ef7aa1802c8ede384f07ae65fec48bdc Author: Mateusz Guzik <mjg@FreeBSD.org> AuthorDate: 2023-09-14 19:08:40 +0000 Commit: Mateusz Guzik <mjg@FreeBSD.org> CommitDate: 2023-10-13 23:48:13 +0000 vfs: prefix regular vnlru with a special case for free vnodes Works around severe performance problems in certain corner cases, see the commentary added. Modifying vnlru logic has proven rather error prone in the past and a release is near, thus take the easy way out and fix it without having to dig into the current machinery. (cherry picked from commit 90a008e94bb205e5b8f3c41d57e155b59a6be95d) --- sys/kern/vfs_subr.c | 117 ++++++++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 113 insertions(+), 4 deletions(-) diff --git a/sys/kern/vfs_subr.c b/sys/kern/vfs_subr.c index 5ddbff77a221..e4b9ff8e1ba6 100644 --- a/sys/kern/vfs_subr.c +++ b/sys/kern/vfs_subr.c @@ -1417,6 +1417,14 @@ vnlru_free_locked(int count) return (ret); } +static int +vnlru_free(int count) +{ + + mtx_lock(&vnode_list_mtx); + return (vnlru_free_locked(count)); +} + void vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp) { @@ -1611,6 +1619,106 @@ vnlru_kick_cond(void) mtx_unlock(&vnode_list_mtx); } +static void +vnlru_proc_sleep(void) +{ + + if (vnlruproc_sig) { + vnlruproc_sig = 0; + wakeup(&vnlruproc_sig); + } + msleep(vnlruproc, &vnode_list_mtx, PVFS|PDROP, "vlruwt", hz); +} + +/* + * A lighter version of the machinery below. + * + * Tries to reach goals only by recycling free vnodes and does not invoke + * uma_reclaim(UMA_RECLAIM_DRAIN). + * + * This works around pathological behavior in vnlru in presence of tons of free + * vnodes, but without having to rewrite the machinery at this time. Said + * behavior boils down to continuously trying to reclaim all kinds of vnodes + * (cycling through all levels of "force") when the count is transiently above + * limit. This happens a lot when all vnodes are used up and vn_alloc + * speculatively increments the counter. + * + * Sample testcase: vnode limit 8388608, 20 separate directory trees each with + * 1 million files in total and 20 find(1) processes stating them in parallel + * (one per each tree). + * + * On a kernel with only stock machinery this needs anywhere between 60 and 120 + * seconds to execute (time varies *wildly* between runs). With the workaround + * it consistently stays around 20 seconds. + * + * That is to say the entire thing needs a fundamental redesign (most notably + * to accommodate faster recycling), the above only tries to get it ouf the way. + * + * Return values are: + * -1 -- fallback to regular vnlru loop + * 0 -- do nothing, go to sleep + * >0 -- recycle this many vnodes + */ +static long +vnlru_proc_light_pick(void) +{ + u_long rnumvnodes, rfreevnodes; + + if (vstir || vnlruproc_sig == 1) + return (-1); + + rnumvnodes = atomic_load_long(&numvnodes); + rfreevnodes = vnlru_read_freevnodes(); + + /* + * vnode limit might have changed and now we may be at a significant + * excess. Bail if we can't sort it out with free vnodes. + */ + if (rnumvnodes > desiredvnodes) { + if (rnumvnodes - rfreevnodes >= desiredvnodes || + rfreevnodes <= wantfreevnodes) { + return (-1); + } + + return (rnumvnodes - desiredvnodes); + } + + /* + * Don't try to reach wantfreevnodes target if there are too few vnodes + * to begin with. + */ + if (rnumvnodes < wantfreevnodes) { + return (0); + } + + if (rfreevnodes < wantfreevnodes) { + return (-1); + } + + return (0); +} + +static bool +vnlru_proc_light(void) +{ + long freecount; + + mtx_assert(&vnode_list_mtx, MA_NOTOWNED); + + freecount = vnlru_proc_light_pick(); + if (freecount == -1) + return (false); + + if (freecount != 0) { + vnlru_free(freecount); + } + + mtx_lock(&vnode_list_mtx); + vnlru_proc_sleep(); + mtx_assert(&vnode_list_mtx, MA_NOTOWNED); + return (true); +} + static void vnlru_proc(void) { @@ -1626,6 +1734,10 @@ vnlru_proc(void) want_reread = false; for (;;) { kproc_suspend_check(vnlruproc); + + if (force == 0 && vnlru_proc_light()) + continue; + mtx_lock(&vnode_list_mtx); rnumvnodes = atomic_load_long(&numvnodes); @@ -1656,10 +1768,7 @@ vnlru_proc(void) vstir = false; } if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) { - vnlruproc_sig = 0; - wakeup(&vnlruproc_sig); - msleep(vnlruproc, &vnode_list_mtx, - PVFS|PDROP, "vlruwt", hz); + vnlru_proc_sleep(); continue; } rfreevnodes = vnlru_read_freevnodes();
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?202310132353.39DNrtJU032274>