]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
s390/pai_ext: Enable concurrent system-wide counting/sampling
authorThomas Richter <tmricht@linux.ibm.com>
Mon, 15 Apr 2024 12:40:03 +0000 (14:40 +0200)
committerAlexander Gordeev <agordeev@linux.ibm.com>
Fri, 7 Jun 2024 14:49:07 +0000 (16:49 +0200)
The PMU for PAI NNPA counters enforces the following restriction:

- No system wide counting while system wide sampling is active.

This restriction is removed. One or more system wide counting events
can now be active at the same time while at most one system wide
sampling event is active.

Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
arch/s390/kernel/perf_pai_ext.c

index a6da7e0cc7a66dac02e9524feb802b0bfee8e0e8..5d00f77a64a5fffaf6a3745b03a532c36ba74ab1 100644 (file)
@@ -120,8 +120,10 @@ static void paiext_event_destroy(struct perf_event *event)
        struct paiext_mapptr *mp = per_cpu_ptr(paiext_root.mapptr, event->cpu);
        struct paiext_map *cpump = mp->mapptr;
 
-       free_page(PAI_SAVE_AREA(event));
        mutex_lock(&paiext_reserve_mutex);
+       if (event->attr.sample_period)
+               cpump->mode &= ~PAI_MODE_SAMPLING;
+       free_page(PAI_SAVE_AREA(event));
        if (refcount_dec_and_test(&cpump->refcnt))      /* Last reference gone */
                paiext_free(mp);
        paiext_root_free();
@@ -186,21 +188,19 @@ static int paiext_alloc(struct perf_event_attr *a, struct perf_event *event)
                        goto undo;
                }
                refcount_set(&cpump->refcnt, 1);
-               cpump->mode = a->sample_period ? PAI_MODE_SAMPLING
-                                              : PAI_MODE_COUNTING;
        } else {
                /* Multiple invocation, check what is active.
-                * Supported are multiple counter events or only one sampling
+                * Supported are multiple counter events and only one sampling
                 * event concurrently at any one time.
                 */
-               if (cpump->mode == PAI_MODE_SAMPLING ||
-                   (cpump->mode == PAI_MODE_COUNTING && a->sample_period)) {
+               if (a->sample_period && (cpump->mode & PAI_MODE_SAMPLING)) {
                        rc = -EBUSY;
                        goto undo;
                }
                refcount_inc(&cpump->refcnt);
        }
-
+       if (a->sample_period)
+               cpump->mode |= PAI_MODE_SAMPLING;
        rc = 0;
 
 undo:
@@ -335,6 +335,8 @@ static void paiext_start(struct perf_event *event, int flags)
                local64_set(&event->hw.prev_count, sum);
        } else {                                /* Sampling */
                cpump->event = event;
+               memcpy((void *)PAI_SAVE_AREA(event), cpump->area,
+                      PAIE1_CTRBLOCK_SZ);
                perf_sched_cb_inc(event->pmu);
        }
 }
@@ -493,7 +495,7 @@ static int paiext_have_sample(void)
 static void paiext_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in)
 {
        /* We started with a clean page on event installation. So read out
-        * results on schedule_out and if page was dirty, clear values.
+        * results on schedule_out and if page was dirty, save old values.
         */
        if (!sched_in)
                paiext_have_sample();