Until now, x86 can queue jump_label text-pokes as long as the
poke-addr is monotonically increasing, but flushes the queue when a
new poke-addr is less than the previous.

Dynamic-debug now uses the queued static-key API, but the advantage is
limited; we see a ~2x reduction in IPIs.  Although the pr_debug
descriptors are ordered, the patch-address in them are not; about 1/2
violate the ordering constraint.

So this patch drops that requirement, and sorts the text-pokes by
their address before applying them.  Doing so lets us fill the queue
before sorting then flushing, giving a dramatic ~125x reduction in
IPIs over the traditional single IPI per pr_debug.

Other arches don't need a queue, and so have nothing to sort.

   #> dd_ipis
   [   23.100381] dyndbg: query 0: "module !virtio* +p " mod:*
   [   23.103432] dyndbg: query 1: "-p" mod:*
   Delta-CAL (IPI): 242

Signed-off-by: Jim Cromie <jim.cromie@gmail.com>
---
 arch/x86/kernel/alternative.c | 42 ++++++++++++++++-------------------
 lib/dynamic_debug.c           |  1 +
 2 files changed, 20 insertions(+), 23 deletions(-)

diff --git a/arch/x86/kernel/alternative.c b/arch/x86/kernel/alternative.c
index e87da25d1236..92987954b8aa 100644
--- a/arch/x86/kernel/alternative.c
+++ b/arch/x86/kernel/alternative.c
@@ -2,6 +2,7 @@
 #define pr_fmt(fmt) "SMP alternatives: " fmt
 
 #include <linux/mmu_context.h>
+#include <linux/sort.h>
 #include <linux/perf_event.h>
 #include <linux/vmalloc.h>
 #include <linux/memory.h>
@@ -2823,6 +2824,18 @@ static __always_inline int patch_cmp(const void *tpl_a, const void *tpl_b)
 	return 0;
 }
 
+static int text_poke_loc_cmp(const void *a, const void *b)
+{
+	const struct smp_text_poke_loc *tpl_a = a;
+	const struct smp_text_poke_loc *tpl_b = b;
+
+	if (tpl_a->rel_addr < tpl_b->rel_addr)
+		return -1;
+	if (tpl_a->rel_addr > tpl_b->rel_addr)
+		return 1;
+	return 0;
+}
+
 noinstr int smp_text_poke_int3_handler(struct pt_regs *regs)
 {
 	struct smp_text_poke_loc *tpl;
@@ -2935,6 +2948,10 @@ void smp_text_poke_batch_finish(void)
 	if (!text_poke_array.nr_entries)
 		return;
 
+	if (text_poke_array.nr_entries > 1)
+		sort(text_poke_array.vec, text_poke_array.nr_entries,
+		     sizeof(struct smp_text_poke_loc), text_poke_loc_cmp, NULL);
+
 	lockdep_assert_held(&text_mutex);
 
 	/*
@@ -3151,28 +3168,6 @@ static void __smp_text_poke_batch_add(void *addr, const void *opcode, size_t len
 	}
 }
 
-/*
- * We hard rely on the text_poke_array.vec being ordered; ensure this is so by flushing
- * early if needed.
- */
-static bool text_poke_addr_ordered(void *addr)
-{
-	WARN_ON_ONCE(!addr);
-
-	if (!text_poke_array.nr_entries)
-		return true;
-
-	/*
-	 * If the last current entry's address is higher than the
-	 * new entry's address we'd like to add, then ordering
-	 * is violated and we must first flush all pending patching
-	 * requests:
-	 */
-	if (text_poke_addr(text_poke_array.vec + text_poke_array.nr_entries-1) > addr)
-		return false;
-
-	return true;
-}
 
 /**
  * smp_text_poke_batch_add() -- update instruction on live kernel on SMP, batched
@@ -3189,8 +3184,9 @@ static bool text_poke_addr_ordered(void *addr)
  */
 void __ref smp_text_poke_batch_add(void *addr, const void *opcode, size_t len, const void *emulate)
 {
-	if (text_poke_array.nr_entries == TEXT_POKE_ARRAY_MAX || !text_poke_addr_ordered(addr))
+	if (text_poke_array.nr_entries == TEXT_POKE_ARRAY_MAX)
 		smp_text_poke_batch_finish();
+
 	__smp_text_poke_batch_add(addr, opcode, len, emulate);
 }
 
diff --git a/lib/dynamic_debug.c b/lib/dynamic_debug.c
index 18a71a9108d3..b5060749464e 100644
--- a/lib/dynamic_debug.c
+++ b/lib/dynamic_debug.c
@@ -264,6 +264,7 @@ static int ddebug_change(const struct ddebug_query *query,
 		}
 	}
 	mutex_unlock(&ddebug_lock);
+	v2pr_info("applied %d queued updates to sites in total\n", nfound);
 
 	if (!nfound && verbose)
 		pr_info("no matches for query\n");
-- 
2.53.0