[PATCH v13 39/39] drm: Mark __drm_debug as __read_mostly

[Jim Cromie via B4 Relay <devnull+jim.cromie.gmail.com@kernel.org>]

From: Jim Cromie <jim.cromie@gmail.com>

Dyndbg classmaps converts ~5,700 DRM debug callsites into static keys,
yielding 0-off-cost NOOPs that never touch memory. However, several
call patterns do not fit the direct drm_dbg() macro model; notably
open-coded drm_debug_enabled() guard blocks, drm_dbg_printer(), and
ratelimited helpers.

In upstream, the linker places __drm_debug in the same 64-byte
cacheline as hot writers: drm_lease_idr_object, drm_global_mutex.
Writes to those locks issue Request-For-Ownership (RFO) invalidations
across cores, leaving those residual bit-tests to suffer 7
false-sharing hazards and continuous cache bouncing during active
rendering (e.g. 12 concurrent vkcube clients). In contrast, writes to
__drm_debug occur almost never, only when userspace updates
"/sys/module/drm/parameters/debug".

So mark __drm_debug with __read_mostly.  This isolates it from the 7
writable symbols it currently shares a cache-line with, eliminating
those invalidations.

Signed-off-by: Jim Cromie <jim.cromie@gmail.com>
---
 drivers/gpu/drm/drm_print.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/drm_print.c b/drivers/gpu/drm/drm_print.c
index 794417c62097..1ec139c8f06f 100644
--- a/drivers/gpu/drm/drm_print.c
+++ b/drivers/gpu/drm/drm_print.c
@@ -40,7 +40,7 @@
  * __drm_debug: Enable debug output.
  * Bitmask of DRM_UT_x. See include/drm/drm_print.h for details.
  */
-u32 __drm_debug;
+u32 __drm_debug __read_mostly;
 EXPORT_SYMBOL(__drm_debug);
 
 MODULE_PARM_DESC(debug, "Enable debug output, where each bit enables a debug category.\n"

-- 
2.56.0