The unlazy sequence of an rcuwalk lookup occurs a bit earlier than normal for O_CREAT lookups (i.e. in open_last_lookups()). The create logic here historically invoked complete_walk(), which clears the nd->root.mnt pointer when appropriate before the unlazy. This changed in commit 72287417abd1 ("open_last_lookups(): don't abuse complete_walk() when all we want is unlazy"), which refactored the create path to invoke unlazy_walk() and not consider nd->root.mnt. This tweak negatively impacts performance on a concurrent open(O_CREAT) workload to multiple independent mounts beneath the root directory. This attributes to increased spinlock contention on the root dentry via legitimize_root(), to the point where the spinlock becomes the primary bottleneck over the directory inode rwsem of the individual submounts. For example, the completion rate of a 32k thread aim7 create/close benchmark that repeatedly passes O_CREAT to open preexisting files drops from over 700k "jobs per minute" to 30, increasing the overall test time from a few minutes to over an hour. A similar, more simplified test to create a set of opener tasks across a set of submounts can demonstrate the problem more quickly. For example, consider sets of 100 open/close tasks each running against 64 independent filesystem mounts (i.e. 6400 tasks total), with each task completing 10k iterations before it exits. On an 80xcpu box running v5.16.0-rc2, this test completes in 50-55s. With this patch applied, the same test completes in 10-15s. This is not the most realistic workload in the world as it factors out inode allocation in the filesystem. The contention can also be avoided by more selective use of O_CREAT or via use of relative pathnames. That said, this regression appears to be an unintentional side effect of code cleanup and might be unexpected for users. Restore original behavior prior to commit 72287417abd1 by factoring the nd->root handling logic from complete_walk() into a new helper and invoke that from both places. Note that the LOOKUP_CACHE logic is not required here because it is incompatible with O_CREAT. Otherwise the tradeoff for this change is that this may impact behavior when an absolute path O_CREAT lookup lands on a symlink that contains another absolute path. The unlazy sequence of the create lookup now clears the nd->root mount pointer, which means that once we read said link via step_into(), the subsequent nd_jump_root() calls into set_root() to grab the mount pointer again (from refwalk mode). This is historical behavior for O_CREAT and less common than the current behavior of a typical create lookup unnecessarily legitimizing the root dentry. Signed-off-by: Brian Foster <bfoster@xxxxxxxxxx> --- v2: - Refactor to a more granular helper to avoid confusion with complete_walk(). - Update the commit log description with some reasoning on functionality of the helper and side effects of clearing the mnt pointer. v1: https://lore.kernel.org/linux-fsdevel/20220105180259.115760-1-bfoster@xxxxxxxxxx/ Hi all, This v2 incorporates some tweaks based on Al's feedback. Al expressed some concern over potential confusion of the complete_walk_rcu() helper as an alternative for complete_walk(), presumably because this _rcu() naming convention is used as such elsewhere. While thinking about a better name, it seemed cleaner to just reduce the scope of the helper to minimal logic. I'm still not totally sure on the name, but that can be changed (or we could always just open code it). Ian, I dropped your R-b tag from v1 since the logic did change a bit, so another look is appreciated. Thanks. Brian fs/namei.c | 23 ++++++++++++++--------- 1 file changed, 14 insertions(+), 9 deletions(-) diff --git a/fs/namei.c b/fs/namei.c index 1f9d2187c765..a72e144a0a37 100644 --- a/fs/namei.c +++ b/fs/namei.c @@ -856,6 +856,18 @@ static inline int d_revalidate(struct dentry *dentry, unsigned int flags) return 1; } +static inline bool nd_reset_root_and_unlazy(struct nameidata *nd) +{ + /* + * We don't want to zero nd->root for scoped-lookups or + * externally-managed nd->root. + */ + if (!(nd->state & ND_ROOT_PRESET)) + if (!(nd->flags & LOOKUP_IS_SCOPED)) + nd->root.mnt = NULL; + return try_to_unlazy(nd); +} + /** * complete_walk - successful completion of path walk * @nd: pointer nameidata @@ -872,15 +884,8 @@ static int complete_walk(struct nameidata *nd) int status; if (nd->flags & LOOKUP_RCU) { - /* - * We don't want to zero nd->root for scoped-lookups or - * externally-managed nd->root. - */ - if (!(nd->state & ND_ROOT_PRESET)) - if (!(nd->flags & LOOKUP_IS_SCOPED)) - nd->root.mnt = NULL; nd->flags &= ~LOOKUP_CACHED; - if (!try_to_unlazy(nd)) + if (!nd_reset_root_and_unlazy(nd)) return -ECHILD; } @@ -3326,7 +3331,7 @@ static const char *open_last_lookups(struct nameidata *nd, } else { /* create side of things */ if (nd->flags & LOOKUP_RCU) { - if (!try_to_unlazy(nd)) + if (!nd_reset_root_and_unlazy(nd)) return ERR_PTR(-ECHILD); } audit_inode(nd->name, dir, AUDIT_INODE_PARENT); -- 2.31.1