Re: [PATCH] dcache: add fs.dentry-limit sysctl with negative-first reaper
From: Ian Kent
Date: Sun May 17 2026 - 22:56:03 EST
Hi Neil, ;)
I just happen to have been caught by this problem too!
See below.
On 18/5/26 07:55, NeilBrown wrote:
On Fri, 15 May 2026, Horst Birthelmer wrote:
From: Horst Birthelmer <hbirthelmer@xxxxxxx>Is a system-wide cap really a suitable tool? What guidance would you
The dcache only shrinks under memory pressure, which is rarely reached
on machines with ample RAM, so cached negative dentries can accumulate
without bound. Give administrators a soft cap they can set,
and a background worker that prefers negative dentries when reclaiming.
Two new sysctls under /proc/sys/fs/:
dentry-limit -- soft cap on nr_dentry. 0 (default)
disables the feature; behaviour is then
identical to before.
give to sysadmins who are considering setting a number?
Is there a better approach?
That's a good question.
In my RFC (with a different, almost trivial, approach) to get a feel for
what people thought of the idea of limiting dentry going to the LRU a
number of similar comments came up. The commit e6957c99dca5 (("vfs: Delete
the associated dentry when deleting a file") is something like what we
want but also doesn't use a limit and so lacks a calculation to work
that out. It's probably a little bit "too" aggressive.
Worse still my claim in the RFC that many entries in the dcache can lead
to a performance problem with long hash chains didn't hold true when I
looked closer at it. For example (and TBH I've forgotten the actual numbers
now but these should be close-ish), with the dcache at 12 Million entries
with around 8 million negative, AFAICT, leads to an average hash chain
length of around 12 which is a little too long but not terrible by any
means. In my experience an average chain length of 8 or less performs
really well so this can't realistically be used and by this time the
problem is already evident.
Leaving me at a loss for a reasonable way to calculate this.
According to the email you linked, a problem arises when a directory has
a great many negative children. Code which walks the list of children
(such as fsnotify) while holding a lock can suffer unpredictable delays
and result in long lock-hold times. So maybe a limit on negative
dentries for any parent is what we really want. That would be clumsy to
implement I imagine.
But the notion of dropping the dentry in ->d_delete() on last dput() is
simple enough but did see regressions (the only other place in the VFS besides
dentry_kill() that the inode is unlinked from the dentry on dput()). I wonder
if the regression wwas related to the test itself deliberately recreating
deleted files and if that really is normal behaviour. By itself that should
prevent almost all negative dentries being retained. Although file systems
could do this as well (think XFS inode recycling) it should be reasonable to
require it be left to the VFS.
But even that's not enough given that, in my case, there would still be around
4 million dentries in the LRU cache and in fsnotify there are directory child
traversals holding the parent i_lock "spinlock" that are going to cause problems.
That's all that much more puzzling when I see things like commit 172e422ffea2
("fsnotify: clear PARENT_WATCHED flags lazily") which looks like it implies
the child flag depends entirely on the parent state (what am I missing Amir?)
so why is this traversal even retained in fsnotify?
But what if we move dentries to the end of the list when they become
negative, and to the start of the list when they become positive? Then
code which walks the child list could simply abort on the first
negative.
I doubt that would be quite as easy as it sounds, but it would at least
be more focused on the observed symptom rather than some whole-system
number which only vaguely correlates with the observed symptom.
Maybe a completely different approach: change children-walking code to
drop and retake the lock (with appropriate validation) periodically.
What too would address the specific symptom.
Another good question.
I have assumed that dropping and re-taking the lock cannot be done but
this is a question I would like answered as well. Dropping and re-taking
lock would require, as Miklos pointed out to me off-list, recording the
list position with say a cursor, introducing unwanted complexity when it
would be better to accept the cost of a single extra access to the parent
flags (which I assume is one reason to set the flag in the child).
Thanks for attempting to resolve this issue, but I'm not convinced that
you have found a good solution yet.
This same sort of issue comes up again and again and I have thought about
it many times without actually useful ideas and it seems like I'm not
alone, ;)
Ian
NeilBrown
dentry-limit-interval-ms -- pacing for the worker while still over
the cap. Default 1000, minimum 1.
When the cap is exceeded, a delayed_work runs in two phases:
1. iterate_supers() draining only negative dentries from every LRU.
Positive entries are rotated past so the walk makes progress.
DCACHE_REFERENCED is ignored here on purpose -- an admin-imposed
cap should evict even hot negatives before any positive entry.
2. If still over the cap, iterate_supers() again with the same
isolate callback the memory-pressure shrinker uses.
Signed-off-by: Horst Birthelmer <hbirthelmer@xxxxxxx>
---
There was a discussion at LSFMM about servers with too many cached
negative dentries.
That gave me the idea to keep the dentries in general limited
if the system administrator needs it to.
This is somewhat related to [1] where it would address the same
symptoms but in a more unobtrusive way, by just garbage collecting
the negative and then the unused cache entries.
The other effect I have seen regarding this is that FUSE
will not forget inodes (no FORGET call to the FUSE server)
even after the latest reference has been closed until much later.
In a FUSE server that mirrors the kernel cached inodes in user space
because it has to keep a lot of private data for every node
this puts an unnecessarry memory strain on that userspace entity
especially if the memory is limited for its cgroup.
[1]: https://lore.kernel.org/linux-fsdevel/20260331012925.74840-1-raven@xxxxxxxxxx/
---
Documentation/admin-guide/sysctl/fs.rst | 28 +++++
fs/dcache.c | 197 ++++++++++++++++++++++++++++++++
2 files changed, 225 insertions(+)
diff --git a/Documentation/admin-guide/sysctl/fs.rst b/Documentation/admin-guide/sysctl/fs.rst
index 9b7f65c3efd8..0229aea45d85 100644
--- a/Documentation/admin-guide/sysctl/fs.rst
+++ b/Documentation/admin-guide/sysctl/fs.rst
@@ -38,6 +38,34 @@ requests. ``aio-max-nr`` allows you to change the maximum value
``aio-max-nr`` does not result in the
pre-allocation or re-sizing of any kernel data structures.
+dentry-limit
+------------
+
+Soft cap on the total number of dentries allocated system-wide (i.e. on
+``nr_dentry`` from ``dentry-state``). A value of ``0`` (the default)
+disables the feature and the dcache grows or shrinks only under memory
+pressure as before.
+
+When set to a non-zero value, a background worker is woken whenever
+the live dentry count exceeds the limit. The worker walks every
+superblock's LRU and prefers to evict negative dentries first; if it
+cannot get back under the limit using negative entries alone it falls
+back to the same LRU policy used by the memory-pressure shrinker.
+
+The limit is *soft*: allocations never fail because of it, and brief
+overshoots while the worker catches up are expected. Set the cap a
+comfortable margin above your steady-state working set.
+
+dentry-limit-interval-ms
+------------------------
+
+How often, in milliseconds, the ``dentry-limit`` worker re-runs while
+``nr_dentry`` is still above the cap. Defaults to ``1000`` (one
+second); the minimum accepted value is ``1``. Smaller values trim the
+cache more aggressively at the cost of more CPU spent walking LRUs;
+larger values let temporary spikes ride out before any work is done.
+Has no effect when ``dentry-limit`` is ``0``.
+
dentry-negative
----------------------------
diff --git a/fs/dcache.c b/fs/dcache.c
index 2c61aeea41f4..4959d2c011c0 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -144,6 +144,19 @@ static DEFINE_PER_CPU(long, nr_dentry_unused);
static DEFINE_PER_CPU(long, nr_dentry_negative);
static int dentry_negative_policy;
+/*
+ * Soft cap on the total number of dentries. When non-zero and exceeded,
+ * a background worker prunes unused dentries (preferring negative ones)
+ * until we are back under the limit. Zero (the default) disables the
+ * feature entirely; the fast path in __d_alloc() only pays the cost of
+ * a READ_ONCE and a branch in that case.
+ */
+static unsigned long sysctl_dentry_limit __read_mostly;
+static unsigned int sysctl_dentry_limit_interval_ms __read_mostly = 1000;
+static unsigned long dentry_limit_last_kick;
+
+static void dentry_limit_kick(void);
+
#if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
/* Statistics gathering. */
static struct dentry_stat_t dentry_stat = {
@@ -199,6 +212,20 @@ static int proc_nr_dentry(const struct ctl_table *table, int write, void *buffer
return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
}
+/*
+ * Writing fs.dentry-limit should give prompt feedback to admins
+ * lowering the cap, so kick the worker on every successful write.
+ */
+static int proc_dentry_limit(const struct ctl_table *table, int write,
+ void *buffer, size_t *lenp, loff_t *ppos)
+{
+ int ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
+
+ if (write && !ret)
+ dentry_limit_kick();
+ return ret;
+}
+
static const struct ctl_table fs_dcache_sysctls[] = {
{
.procname = "dentry-state",
@@ -207,6 +234,21 @@ static const struct ctl_table fs_dcache_sysctls[] = {
.mode = 0444,
.proc_handler = proc_nr_dentry,
},
+ {
+ .procname = "dentry-limit",
+ .data = &sysctl_dentry_limit,
+ .maxlen = sizeof(sysctl_dentry_limit),
+ .mode = 0644,
+ .proc_handler = proc_dentry_limit,
+ },
+ {
+ .procname = "dentry-limit-interval-ms",
+ .data = &sysctl_dentry_limit_interval_ms,
+ .maxlen = sizeof(sysctl_dentry_limit_interval_ms),
+ .mode = 0644,
+ .proc_handler = proc_douintvec_minmax,
+ .extra1 = SYSCTL_ONE,
+ },
{
.procname = "dentry-negative",
.data = &dentry_negative_policy,
@@ -1325,6 +1367,160 @@ static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
return LRU_REMOVED;
}
+#define DENTRY_LIMIT_BATCH 1024UL
+
+static void dentry_limit_worker_fn(struct work_struct *work);
+static DECLARE_DELAYED_WORK(dentry_limit_work, dentry_limit_worker_fn);
+
+/*
+ * Variant of dentry_lru_isolate() that only frees negative dentries.
+ * DCACHE_REFERENCED is intentionally not honoured here: the whole point
+ * of an admin-imposed cap on negatives is that even frequently-looked-up
+ * negative entries should be evicted before any positive dentry.
+ * Positive entries are rotated to the tail so the walk continues to
+ * make progress without disturbing their LRU position.
+ */
+static enum lru_status dentry_lru_isolate_negative(struct list_head *item,
+ struct list_lru_one *lru, void *arg)
+{
+ struct list_head *freeable = arg;
+ struct dentry *dentry = container_of(item, struct dentry, d_lru);
+
+ if (!spin_trylock(&dentry->d_lock))
+ return LRU_SKIP;
+
+ /* Same handling as dentry_lru_isolate() for in-use entries. */
+ if (dentry->d_lockref.count) {
+ d_lru_isolate(lru, dentry);
+ spin_unlock(&dentry->d_lock);
+ return LRU_REMOVED;
+ }
+
+ if (!d_is_negative(dentry)) {
+ spin_unlock(&dentry->d_lock);
+ return LRU_ROTATE;
+ }
+
+ d_lru_shrink_move(lru, dentry, freeable);
+ spin_unlock(&dentry->d_lock);
+ return LRU_REMOVED;
+}
+
+struct dentry_limit_ctx {
+ long over; /* remaining dentries to evict */
+ list_lru_walk_cb isolate;
+};
+
+static void dentry_limit_prune_sb(struct super_block *sb, void *arg)
+{
+ struct dentry_limit_ctx *ctx = arg;
+ unsigned long walked = 0;
+ unsigned long budget;
+
+ if (ctx->over <= 0)
+ return;
+
+ /*
+ * Walk up to one full pass of this superblock's LRU, in
+ * DENTRY_LIMIT_BATCH-sized chunks. The loop matters mainly for
+ * phase 1: dentry_lru_isolate_negative() returns LRU_ROTATE for
+ * positive dentries, which still counts against list_lru_walk()'s
+ * nr_to_walk. A single batch can therefore finish having freed
+ * nothing when positives crowd the head of the LRU, and without
+ * the inner loop the worker would have to wait a full
+ * dentry-limit-interval-ms before retrying never reaching the
+ * negatives buried behind a long run of positives.
+ *
+ * The budget is snapshot at entry so a filesystem allocating
+ * dentries faster than we drain them can't keep us spinning here
+ * forever; freshly added dentries are picked up on the next
+ * worker invocation.
+ *
+ * Phase 2 normally exits much sooner: its isolate callback frees
+ * any non-referenced dentry, so ctx->over typically hits zero
+ * inside the first batch. The worst-case over-eviction is one
+ * batch past the cap, which is within the soft semantics of
+ * fs.dentry-limit.
+ */
+ budget = list_lru_count(&sb->s_dentry_lru);
+
+ while (ctx->over > 0 && walked < budget) {
+ LIST_HEAD(dispose);
+ unsigned long nr;
+ long freed;
+
+ nr = min(DENTRY_LIMIT_BATCH, budget - walked);
+ freed = list_lru_walk(&sb->s_dentry_lru, ctx->isolate,
+ &dispose, nr);
+ shrink_dentry_list(&dispose);
+
+ ctx->over -= freed;
+ walked += nr;
+
+ cond_resched();
+ }
+}
+
+static void dentry_limit_worker_fn(struct work_struct *work)
+{
+ struct dentry_limit_ctx ctx;
+ unsigned long limit = READ_ONCE(sysctl_dentry_limit);
+ unsigned int ms;
+ long nr;
+
+ if (!limit)
+ return;
+
+ nr = get_nr_dentry();
+ if (nr <= (long)limit)
+ return;
+
+ ctx.over = nr - (long)limit;
+
+ /* Phase 1: drain negative dentries across every superblock. */
+ ctx.isolate = dentry_lru_isolate_negative;
+ iterate_supers(dentry_limit_prune_sb, &ctx);
+
+ /* Phase 2: still over? Apply the ordinary LRU policy. */
+ if (ctx.over > 0) {
+ ctx.isolate = dentry_lru_isolate;
+ iterate_supers(dentry_limit_prune_sb, &ctx);
+ }
+
+ /*
+ * Re-arm while still above the limit. Re-read the sysctls in
+ * case the admin raised the cap or disabled the feature during
+ * the walk.
+ */
+ limit = READ_ONCE(sysctl_dentry_limit);
+ if (!limit || get_nr_dentry() <= (long)limit)
+ return;
+
+ ms = READ_ONCE(sysctl_dentry_limit_interval_ms);
+ queue_delayed_work(system_unbound_wq, &dentry_limit_work,
+ msecs_to_jiffies(ms));
+}
+
+static void dentry_limit_kick(void)
+{
+ unsigned long limit = READ_ONCE(sysctl_dentry_limit);
+ unsigned long now;
+
+ if (!limit)
+ return;
+ if (delayed_work_pending(&dentry_limit_work))
+ return;
+
+ now = jiffies;
+ if (time_before(now, READ_ONCE(dentry_limit_last_kick) + HZ / 10))
+ return;
+ WRITE_ONCE(dentry_limit_last_kick, now);
+
+ if (get_nr_dentry() <= (long)limit)
+ return;
+
+ queue_delayed_work(system_unbound_wq, &dentry_limit_work, 0);
+}
/**
* shrink_dcache_sb - shrink dcache for a superblock
@@ -1868,6 +2064,7 @@ static struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
}
this_cpu_inc(nr_dentry);
+ dentry_limit_kick();
return dentry;
}
---
base-commit: 5d6919055dec134de3c40167a490f33c74c12581
change-id: 20260513-limit-dentries-cache-63685729672b
Best regards,
--
Horst Birthelmer <hbirthelmer@xxxxxxx>