On Mon, Nov 09, 2026 at 08:49:04PM +0100, Jiri Olsa wrote:
quoted hunk ↗ jump to hunk
--- a/kernel/static_call_inline.c
+++ b/kernel/static_call_inline.c
@@ -15,6 +15,18 @@ extern struct static_call_site __start_static_call_sites[],
extern struct static_call_tramp_key __start_static_call_tramp_key[],
__stop_static_call_tramp_key[];
+struct static_call_site_index {
+ struct static_call_site *base;
+ unsigned int nr;
+ u32 entries[];
+};
+
+/*
+ * Index static call sites by code address for text reservation lookups.
+ * The site table itself is sorted by static-call-key address.
+ */
+static struct static_call_site_index *static_call_site_index;
+
int static_call_initialized;
/*@@ -41,7 +53,7 @@ static void static_call_unlock(void)
mutex_unlock(&static_call_mutex);
}
-static inline void *static_call_addr(struct static_call_site *site)
+static inline void *static_call_addr(const struct static_call_site *site)
{
return (void *)((long)site->addr + (long)&site->addr);
}@@ -301,6 +313,79 @@ static int addr_conflict(struct static_call_site *site, void *start, void *end)
return 0;
}
+static int static_call_site_index_cmp(const void *a, const void *b, const void *priv)
+{
+ const struct static_call_site *base = priv;
+ unsigned int ia = *(u32 *)a;
+ unsigned int ib = *(u32 *)b;
+ unsigned long ca = (unsigned long)static_call_addr(base + ia);
+ unsigned long cb = (unsigned long)static_call_addr(base + ib);
+
+ return cmp_int(ca, cb);
+}
+
+static struct static_call_site_index *
+static_call_site_index_alloc(struct static_call_site *start,
+ struct static_call_site *stop)
+{
+ struct static_call_site_index *index;
+ unsigned int nr = stop - start;
+ unsigned int i;
+
+ if (!nr)
+ return NULL;
+
+ index = kvmalloc_flex(*index, entries, nr, GFP_KERNEL);
+ if (!index)
+ return NULL;
+
+ index->nr = nr;
+ index->base = start;
+ for (i = 0; i < nr; i++)
+ index->entries[i] = i;
+
+ sort_r(index->entries, nr, sizeof(index->entries[0]),
+ static_call_site_index_cmp, NULL, start);
+
+ return index;
+}
+
+static int
+static_call_site_index_text_reserved(struct static_call_site_index *index,
+ void *start, void *end, bool init)
+{
+ unsigned int low = 0, high = index->nr, i;
+
+ /* Find the first site after @start. */
+ while (low < high) {
+ unsigned int mid = low + (high - low) / 2;
+ void *addr = static_call_addr(index->base + index->entries[mid]);
+
+ if (addr <= start)
+ low = mid + 1;
+ else
+ high = mid;
+ }
+
+ /* The preceding site may overlap @start. */
+ if (low)
+ low--;
+
+ for (i = low; i < index->nr; i++) {
+ struct static_call_site *site = index->base + index->entries[i];
+ void *addr = static_call_addr(site);
+
+ /* Later sites cannot overlap once this address is past @end. */
+ if (addr > end)
+ break;
+ if ((init || !static_call_is_init(site)) &&
+ addr_conflict(site, start, end))
+ return 1;
+ }
+
+ return 0;
+}
Looking at this in a little more detail; I'm confused.
You're literally adding an index, so that you can find the
static_call_size, while all you really need is static_call_addr().
That is, you could have just created an array of relocated site->addr
and sorted that, that would safe you a ton of indirection all around.
That is:
index->entries[i] = (long)static_call_addr(start + i) - (long)&index->entries[i];
That will make your lookup take at least half the amount of loads.
Also, you really should support modules, there is no point being fast
until you load a module.