b18e99
From 8b9a0af8ca012217bf90d1dc0694f85b49ae09da Mon Sep 17 00:00:00 2001
b18e99
From: Noah Goldstein <goldstein.w.n@gmail.com>
b18e99
Date: Tue, 18 Jul 2023 10:27:59 -0500
b18e99
Subject: [PATCH] [PATCH v1] x86: Use `3/4*sizeof(per-thread-L3)` as low bound
b18e99
 for NT threshold.
b18e99
Content-type: text/plain; charset=UTF-8
b18e99
b18e99
On some machines we end up with incomplete cache information. This can
b18e99
make the new calculation of `sizeof(total-L3)/custom-divisor` end up
b18e99
lower than intended (and lower than the prior value). So reintroduce
b18e99
the old bound as a lower bound to avoid potentially regressing code
b18e99
where we don't have complete information to make the decision.
b18e99
Reviewed-by: DJ Delorie <dj@redhat.com>
b18e99
---
b18e99
 sysdeps/x86/dl-cacheinfo.h | 15 ++++++++++++---
b18e99
 1 file changed, 12 insertions(+), 3 deletions(-)
b18e99
b18e99
[DJ - ported to C8S]
b18e99
b18e99
diff -rup b2/sysdeps/x86/cacheinfo.h b3/sysdeps/x86/cacheinfo.h
b18e99
--- b2/sysdeps/x86/cacheinfo.h	2023-08-08 13:55:16.474680016 -0400
b18e99
+++ b3/sysdeps/x86/cacheinfo.h	2023-08-08 13:59:14.507988958 -0400
b18e99
@@ -401,12 +401,20 @@ init_cacheinfo (void)
b18e99
      provides proper LRU hints so that the maximum thrashing
b18e99
      capped at 1/associativity. */
b18e99
   unsigned long int non_temporal_threshold = shared / 4;
b18e99
+  /* If the computed non_temporal_threshold <= 3/4 * per-thread L3, we most
b18e99
+     likely have incorrect/incomplete cache info in which case, default to
b18e99
+     3/4 * per-thread L3 to avoid regressions.  */
b18e99
+  unsigned long int non_temporal_threshold_lowbound
b18e99
+      = shared_per_thread * 3 / 4;
b18e99
+  if (non_temporal_threshold < non_temporal_threshold_lowbound)
b18e99
+    non_temporal_threshold = non_temporal_threshold_lowbound;
b18e99
+
b18e99
   /* If no ERMS, we use the per-thread L3 chunking. Normal cacheable stores run
b18e99
      a higher risk of actually thrashing the cache as they don't have a HW LRU
b18e99
      hint. As well, their performance in highly parallel situations is
b18e99
      noticeably worse.  */
b18e99
   if (!CPU_FEATURE_USABLE_P (cpu_features, ERMS))
b18e99
-    non_temporal_threshold = shared_per_thread * 3 / 4;
b18e99
+    non_temporal_threshold = non_temporal_threshold_lowbound;
b18e99
 
b18e99
   __x86_shared_non_temporal_threshold
b18e99
     = (cpu_features->non_temporal_threshold != 0