On 13/08/2026 19:22, Iván Ezequiel Rodriguez wrote:
Replace the O(N) scan of watched names on each directory inotify event
with a Hash_table keyed by (parent_wd, basename), addressing the
long-standing FIXME in tail_forever_inotify.
Keep ownership consistent with wd_to_name: parent_by_name is created
only for Follow_name and returned to the caller, so fallback paths can
return without local hash cleanup.
Matching cost for N directory events that each compare against N names
(illustrative host timings of the scan vs hash lookup alone):
N=1000: ~0.018s -> ~0.000s
N=5000: ~0.39s -> ~0.000s
N=10000: ~1.5s -> ~0.000s
End-to-end recreate storms remain largely file-system bound on this
host; the hash removes the quadratic name-matching CPU from that path.
+/* Hash Follow_name watches by parent directory wd + basename so directory
+ inotify events can be matched in expected O(1) instead of scanning all
+ files. */
+static size_t
+parent_name_hasher (const void *entry, size_t tabsize)
+{
+ const struct File_spec *spec = entry;
+ char const *name = spec->name + spec->basename_start;
+ size_t value = spec->parent_wd;
+ for (unsigned char c; (c = *name); name++)
+ value = value * 31 + c;
+ return value % tabsize;
As discussed recently, we could use hash_string() for that.
In general the patch looks sensible.
I'll give it a full review tomorrow hopefully.
thanks,
Padraig
