forked from ~ljy/RK356X_SDK_RELEASE
blame | history | patch | commit | commitdiff
hc
2023-12-09 b22da3d8526a935aa31e086e63f60ff3246cb61c 
 kernel/arch/s390/kernel/perf_cpum_sf.c
....@@ -156,8 +156,8 @@
156156 		}
157157 	}
158158 
159
-	debug_sprintf_event(sfdbg, 5,
160
-			    "free_sampling_buffer: freed sdbt=%p\n", sfb->sdbt);
159
+	debug_sprintf_event(sfdbg, 5, "%s: freed sdbt %#lx\n", __func__,
160
+			    (unsigned long)sfb->sdbt);
161161 	memset(sfb, 0, sizeof(*sfb));
162162 }
163163 
....@@ -212,10 +212,11 @@
212212 	 * the sampling buffer origin.
213213 	 */
214214 	if (sfb->sdbt != get_next_sdbt(tail)) {
215
-		debug_sprintf_event(sfdbg, 3, "realloc_sampling_buffer: "
216
-				    "sampling buffer is not linked: origin=%p"
217
-				    "tail=%p\n",
218
-				    (void *) sfb->sdbt, (void *) tail);
215
+		debug_sprintf_event(sfdbg, 3, "%s: "
216
+				    "sampling buffer is not linked: origin %#lx"
217
+				    " tail %#lx\n", __func__,
218
+				    (unsigned long)sfb->sdbt,
219
+				    (unsigned long)tail);
219220 		return -EINVAL;
220221 	}
221222 
....@@ -264,8 +265,8 @@
264265 	*tail = (unsigned long) sfb->sdbt + 1;
265266 	sfb->tail = tail;
266267 
267
-	debug_sprintf_event(sfdbg, 4, "realloc_sampling_buffer: new buffer"
268
-			    " settings: sdbt=%lu sdb=%lu\n",
268
+	debug_sprintf_event(sfdbg, 4, "%s: new buffer"
269
+			    " settings: sdbt %lu sdb %lu\n", __func__,
269270 			    sfb->num_sdbt, sfb->num_sdb);
270271 	return rc;
271272 }
....@@ -305,12 +306,13 @@
305306 	rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL);
306307 	if (rc) {
307308 		free_sampling_buffer(sfb);
308
-		debug_sprintf_event(sfdbg, 4, "alloc_sampling_buffer: "
309
-			"realloc_sampling_buffer failed with rc=%i\n", rc);
309
+		debug_sprintf_event(sfdbg, 4, "%s: "
310
+			"realloc_sampling_buffer failed with rc %i\n",
311
+			__func__, rc);
310312 	} else
311313 		debug_sprintf_event(sfdbg, 4,
312
-			"alloc_sampling_buffer: tear=%p dear=%p\n",
313
-			sfb->sdbt, (void *) *sfb->sdbt);
314
+			"%s: tear %#lx dear %#lx\n", __func__,
315
+			(unsigned long)sfb->sdbt, (unsigned long)*sfb->sdbt);
314316 	return rc;
315317 }
316318 
....@@ -370,28 +372,33 @@
370372 
371373 static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc)
372374 {
373
-	unsigned long n_sdb, freq, factor;
375
+	unsigned long n_sdb, freq;
374376 	size_t sample_size;
375377 
376378 	/* Calculate sampling buffers using 4K pages
377379 	 *
378
-	 *    1. Determine the sample data size which depends on the used
379
-	 *	 sampling functions, for example, basic-sampling or
380
-	 *	 basic-sampling with diagnostic-sampling.
380
+	 *    1. The sampling size is 32 bytes for basic sampling. This size
381
+	 *	 is the same for all machine types. Diagnostic
382
+	 *	 sampling uses auxlilary data buffer setup which provides the
383
+	 *	 memory for SDBs using linux common code auxiliary trace
384
+	 *	 setup.
381385 	 *
382
-	 *    2. Use the sampling frequency as input.  The sampling buffer is
383
-	 *	 designed for almost one second.  This can be adjusted through
384
-	 *	 the "factor" variable.
385
-	 *	 In any case, alloc_sampling_buffer() sets the Alert Request
386
+	 *    2. Function alloc_sampling_buffer() sets the Alert Request
386387 	 *	 Control indicator to trigger a measurement-alert to harvest
387
-	 *	 sample-data-blocks (sdb).
388
+	 *	 sample-data-blocks (SDB). This is done per SDB. This
389
+	 *	 measurement alert interrupt fires quick enough to handle
390
+	 *	 one SDB, on very high frequency and work loads there might
391
+	 *	 be 2 to 3 SBDs available for sample processing.
392
+	 *	 Currently there is no need for setup alert request on every
393
+	 *	 n-th page. This is counterproductive as one IRQ triggers
394
+	 *	 a very high number of samples to be processed at one IRQ.
388395 	 *
389
-	 *    3. Compute the number of sample-data-blocks and ensure a minimum
390
-	 *	 of CPUM_SF_MIN_SDB.  Also ensure the upper limit does not
391
-	 *	 exceed a "calculated" maximum.  The symbolic maximum is
392
-	 *	 designed for basic-sampling only and needs to be increased if
393
-	 *	 diagnostic-sampling is active.
394
-	 *	 See also the remarks for these symbolic constants.
396
+	 *    3. Use the sampling frequency as input.
397
+	 *	 Compute the number of SDBs and ensure a minimum
398
+	 *	 of CPUM_SF_MIN_SDB.  Depending on frequency add some more
399
+	 *	 SDBs to handle a higher sampling rate.
400
+	 *	 Use a minimum of CPUM_SF_MIN_SDB and allow for 100 samples
401
+	 *	 (one SDB) for every 10000 HZ frequency increment.
395402 	 *
396403 	 *    4. Compute the number of sample-data-block-tables (SDBT) and
397404 	 *	 ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up
....@@ -399,10 +406,7 @@
399406 	 */
400407 	sample_size = sizeof(struct hws_basic_entry);
401408 	freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc));
402
-	factor = 1;
403
-	n_sdb = DIV_ROUND_UP(freq, factor * ((PAGE_SIZE-64) / sample_size));
404
-	if (n_sdb < CPUM_SF_MIN_SDB)
405
-		n_sdb = CPUM_SF_MIN_SDB;
409
+	n_sdb = CPUM_SF_MIN_SDB + DIV_ROUND_UP(freq, 10000);
406410 
407411 	/* If there is already a sampling buffer allocated, it is very likely
408412 	 * that the sampling facility is enabled too.  If the event to be
....@@ -417,8 +421,8 @@
417421 		return 0;
418422 
419423 	debug_sprintf_event(sfdbg, 3,
420
-			    "allocate_buffers: rate=%lu f=%lu sdb=%lu/%lu"
421
-			    " sample_size=%lu cpuhw=%p\n",
424
+			    "%s: rate %lu f %lu sdb %lu/%lu"
425
+			    " sample_size %lu cpuhw %p\n", __func__,
422426 			    SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc),
423427 			    sample_size, cpuhw);
424428 
....@@ -478,8 +482,8 @@
478482 	if (num)
479483 		sfb_account_allocs(num, hwc);
480484 
481
-	debug_sprintf_event(sfdbg, 5, "sfb: overflow: overflow=%llu ratio=%lu"
482
-			    " num=%lu\n", OVERFLOW_REG(hwc), ratio, num);
485
+	debug_sprintf_event(sfdbg, 5, "%s: overflow %llu ratio %lu num %lu\n",
486
+			    __func__, OVERFLOW_REG(hwc), ratio, num);
483487 	OVERFLOW_REG(hwc) = 0;
484488 }
485489 
....@@ -517,16 +521,15 @@
517521 	 */
518522 	rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC);
519523 	if (rc)
520
-		debug_sprintf_event(sfdbg, 5, "sfb: extend: realloc "
521
-				    "failed with rc=%i\n", rc);
524
+		debug_sprintf_event(sfdbg, 5, "%s: realloc failed with rc %i\n",
525
+				    __func__, rc);
522526 
523527 	if (sfb_has_pending_allocs(sfb, hwc))
524
-		debug_sprintf_event(sfdbg, 5, "sfb: extend: "
525
-				    "req=%lu alloc=%lu remaining=%lu\n",
526
-				    num, sfb->num_sdb - num_old,
528
+		debug_sprintf_event(sfdbg, 5, "%s: "
529
+				    "req %lu alloc %lu remaining %lu\n",
530
+				    __func__, num, sfb->num_sdb - num_old,
527531 				    sfb_pending_allocs(sfb, hwc));
528532 }
529
-
530533 
531534 /* Number of perf events counting hardware events */
532535 static atomic_t num_events;
....@@ -552,20 +555,22 @@
552555 		err = sf_disable();
553556 		if (err)
554557 			pr_err("Switching off the sampling facility failed "
555
-			       "with rc=%i\n", err);
558
+			       "with rc %i\n", err);
556559 		debug_sprintf_event(sfdbg, 5,
557
-				    "setup_pmc_cpu: initialized: cpuhw=%p\n", cpusf);
560
+				    "%s: initialized: cpuhw %p\n", __func__,
561
+				    cpusf);
558562 		break;
559563 	case PMC_RELEASE:
560564 		cpusf->flags &= ~PMU_F_RESERVED;
561565 		err = sf_disable();
562566 		if (err) {
563567 			pr_err("Switching off the sampling facility failed "
564
-			       "with rc=%i\n", err);
568
+			       "with rc %i\n", err);
565569 		} else
566570 			deallocate_buffers(cpusf);
567571 		debug_sprintf_event(sfdbg, 5,
568
-				    "setup_pmc_cpu: released: cpuhw=%p\n", cpusf);
572
+				    "%s: released: cpuhw %p\n", __func__,
573
+				    cpusf);
569574 		break;
570575 	}
571576 	if (err)
....@@ -610,13 +615,6 @@
610615 	hwc->sample_period = period;
611616 	hwc->last_period = hwc->sample_period;
612617 	local64_set(&hwc->period_left, hwc->sample_period);
613
-}
614
-
615
-static void hw_reset_registers(struct hw_perf_event *hwc,
616
-			       unsigned long *sdbt_origin)
617
-{
618
-	/* (Re)set to first sample-data-block-table */
619
-	TEAR_REG(hwc) = (unsigned long) sdbt_origin;
620618 }
621619 
622620 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
....@@ -674,7 +672,7 @@
674672 	rcu_read_lock();
675673 
676674 	perf_prepare_sample(&header, data, event, regs);
677
-	if (perf_output_begin(&handle, event, header.size))
675
+	if (perf_output_begin(&handle, data, event, header.size))
678676 		goto out;
679677 
680678 	/* Update the process ID (see also kernel/events/core.c) */
....@@ -687,13 +685,88 @@
687685 	rcu_read_unlock();
688686 }
689687 
688
+static unsigned long getrate(bool freq, unsigned long sample,
689
+			     struct hws_qsi_info_block *si)
690
+{
691
+	unsigned long rate;
692
+
693
+	if (freq) {
694
+		rate = freq_to_sample_rate(si, sample);
695
+		rate = hw_limit_rate(si, rate);
696
+	} else {
697
+		/* The min/max sampling rates specifies the valid range
698
+		 * of sample periods.  If the specified sample period is
699
+		 * out of range, limit the period to the range boundary.
700
+		 */
701
+		rate = hw_limit_rate(si, sample);
702
+
703
+		/* The perf core maintains a maximum sample rate that is
704
+		 * configurable through the sysctl interface.  Ensure the
705
+		 * sampling rate does not exceed this value.  This also helps
706
+		 * to avoid throttling when pushing samples with
707
+		 * perf_event_overflow().
708
+		 */
709
+		if (sample_rate_to_freq(si, rate) >
710
+		    sysctl_perf_event_sample_rate) {
711
+			debug_sprintf_event(sfdbg, 1, "%s: "
712
+					    "Sampling rate exceeds maximum "
713
+					    "perf sample rate\n", __func__);
714
+			rate = 0;
715
+		}
716
+	}
717
+	return rate;
718
+}
719
+
720
+/* The sampling information (si) contains information about the
721
+ * min/max sampling intervals and the CPU speed.  So calculate the
722
+ * correct sampling interval and avoid the whole period adjust
723
+ * feedback loop.
724
+ *
725
+ * Since the CPU Measurement sampling facility can not handle frequency
726
+ * calculate the sampling interval when frequency is specified using
727
+ * this formula:
728
+ *	interval := cpu_speed * 1000000 / sample_freq
729
+ *
730
+ * Returns errno on bad input and zero on success with parameter interval
731
+ * set to the correct sampling rate.
732
+ *
733
+ * Note: This function turns off freq bit to avoid calling function
734
+ * perf_adjust_period(). This causes frequency adjustment in the common
735
+ * code part which causes tremendous variations in the counter values.
736
+ */
737
+static int __hw_perf_event_init_rate(struct perf_event *event,
738
+				     struct hws_qsi_info_block *si)
739
+{
740
+	struct perf_event_attr *attr = &event->attr;
741
+	struct hw_perf_event *hwc = &event->hw;
742
+	unsigned long rate;
743
+
744
+	if (attr->freq) {
745
+		if (!attr->sample_freq)
746
+			return -EINVAL;
747
+		rate = getrate(attr->freq, attr->sample_freq, si);
748
+		attr->freq = 0;		/* Don't call  perf_adjust_period() */
749
+		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FREQ_MODE;
750
+	} else {
751
+		rate = getrate(attr->freq, attr->sample_period, si);
752
+		if (!rate)
753
+			return -EINVAL;
754
+	}
755
+	attr->sample_period = rate;
756
+	SAMPL_RATE(hwc) = rate;
757
+	hw_init_period(hwc, SAMPL_RATE(hwc));
758
+	debug_sprintf_event(sfdbg, 4, "%s: cpu %d period %#llx freq %d,%#lx\n",
759
+			    __func__, event->cpu, event->attr.sample_period,
760
+			    event->attr.freq, SAMPLE_FREQ_MODE(hwc));
761
+	return 0;
762
+}
763
+
690764 static int __hw_perf_event_init(struct perf_event *event)
691765 {
692766 	struct cpu_hw_sf *cpuhw;
693767 	struct hws_qsi_info_block si;
694768 	struct perf_event_attr *attr = &event->attr;
695769 	struct hw_perf_event *hwc = &event->hw;
696
-	unsigned long rate;
697770 	int cpu, err;
698771 
699772 	/* Reserve CPU-measurement sampling facility */
....@@ -741,6 +814,12 @@
741814 		goto out;
742815 	}
743816 
817
+	if (si.ribm & CPU_MF_SF_RIBM_NOTAV) {
818
+		pr_warn("CPU Measurement Facility sampling is temporarily not available\n");
819
+		err = -EBUSY;
820
+		goto out;
821
+	}
822
+
744823 	/* Always enable basic sampling */
745824 	SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;
746825 
....@@ -759,43 +838,9 @@
759838 	if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS)
760839 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS;
761840 
762
-	/* The sampling information (si) contains information about the
763
-	 * min/max sampling intervals and the CPU speed.  So calculate the
764
-	 * correct sampling interval and avoid the whole period adjust
765
-	 * feedback loop.
766
-	 */
767
-	rate = 0;
768
-	if (attr->freq) {
769
-		if (!attr->sample_freq) {
770
-			err = -EINVAL;
771
-			goto out;
772
-		}
773
-		rate = freq_to_sample_rate(&si, attr->sample_freq);
774
-		rate = hw_limit_rate(&si, rate);
775
-		attr->freq = 0;
776
-		attr->sample_period = rate;
777
-	} else {
778
-		/* The min/max sampling rates specifies the valid range
779
-		 * of sample periods.  If the specified sample period is
780
-		 * out of range, limit the period to the range boundary.
781
-		 */
782
-		rate = hw_limit_rate(&si, hwc->sample_period);
783
-
784
-		/* The perf core maintains a maximum sample rate that is
785
-		 * configurable through the sysctl interface.  Ensure the
786
-		 * sampling rate does not exceed this value.  This also helps
787
-		 * to avoid throttling when pushing samples with
788
-		 * perf_event_overflow().
789
-		 */
790
-		if (sample_rate_to_freq(&si, rate) >
791
-		      sysctl_perf_event_sample_rate) {
792
-			err = -EINVAL;
793
-			debug_sprintf_event(sfdbg, 1, "Sampling rate exceeds maximum perf sample rate\n");
794
-			goto out;
795
-		}
796
-	}
797
-	SAMPL_RATE(hwc) = rate;
798
-	hw_init_period(hwc, SAMPL_RATE(hwc));
841
+	err =  __hw_perf_event_init_rate(event, &si);
842
+	if (err)
843
+		goto out;
799844 
800845 	/* Initialize sample data overflow accounting */
801846 	hwc->extra_reg.reg = REG_OVERFLOW;
....@@ -836,12 +881,21 @@
836881 	return err;
837882 }
838883 
884
+static bool is_callchain_event(struct perf_event *event)
885
+{
886
+	u64 sample_type = event->attr.sample_type;
887
+
888
+	return sample_type & (PERF_SAMPLE_CALLCHAIN | PERF_SAMPLE_REGS_USER |
889
+			      PERF_SAMPLE_STACK_USER);
890
+}
891
+
839892 static int cpumsf_pmu_event_init(struct perf_event *event)
840893 {
841894 	int err;
842895 
843896 	/* No support for taken branch sampling */
844
-	if (has_branch_stack(event))
897
+	/* No support for callchain, stacks and registers */
898
+	if (has_branch_stack(event) || is_callchain_event(event))
845899 		return -EOPNOTSUPP;
846900 
847901 	switch (event->attr.type) {
....@@ -867,7 +921,7 @@
867921 
868922 	/* Check online status of the CPU to which the event is pinned */
869923 	if (event->cpu >= 0 && !cpu_online(event->cpu))
870
-			return -ENODEV;
924
+		return -ENODEV;
871925 
872926 	/* Force reset of idle/hv excludes regardless of what the
873927 	 * user requested.
....@@ -915,9 +969,10 @@
915969 			 * buffer extents
916970 			 */
917971 			sfb_account_overflows(cpuhw, hwc);
918
-			if (sfb_has_pending_allocs(&cpuhw->sfb, hwc))
919
-				extend_sampling_buffer(&cpuhw->sfb, hwc);
972
+			extend_sampling_buffer(&cpuhw->sfb, hwc);
920973 		}
974
+		/* Rate may be adjusted with ioctl() */
975
+		cpuhw->lsctl.interval = SAMPL_RATE(&cpuhw->event->hw);
921976 	}
922977 
923978 	/* (Re)enable the PMU and sampling facility */
....@@ -927,7 +982,7 @@
927982 	err = lsctl(&cpuhw->lsctl);
928983 	if (err) {
929984 		cpuhw->flags &= ~PMU_F_ENABLED;
930
-		pr_err("Loading sampling controls failed: op=%i err=%i\n",
985
+		pr_err("Loading sampling controls failed: op %i err %i\n",
931986 			1, err);
932987 		return;
933988 	}
....@@ -935,10 +990,11 @@
935990 	/* Load current program parameter */
936991 	lpp(&S390_lowcore.lpp);
937992 
938
-	debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i "
939
-			    "tear=%p dear=%p\n", cpuhw->lsctl.es, cpuhw->lsctl.cs,
940
-			    cpuhw->lsctl.ed, cpuhw->lsctl.cd,
941
-			    (void *) cpuhw->lsctl.tear, (void *) cpuhw->lsctl.dear);
993
+	debug_sprintf_event(sfdbg, 6, "%s: es %i cs %i ed %i cd %i "
994
+			    "interval %#lx tear %#lx dear %#lx\n", __func__,
995
+			    cpuhw->lsctl.es, cpuhw->lsctl.cs, cpuhw->lsctl.ed,
996
+			    cpuhw->lsctl.cd, cpuhw->lsctl.interval,
997
+			    cpuhw->lsctl.tear, cpuhw->lsctl.dear);
942998 }
943999 
9441000 static void cpumsf_pmu_disable(struct pmu *pmu)
....@@ -961,13 +1017,14 @@
9611017 
9621018 	err = lsctl(&inactive);
9631019 	if (err) {
964
-		pr_err("Loading sampling controls failed: op=%i err=%i\n",
1020
+		pr_err("Loading sampling controls failed: op %i err %i\n",
9651021 			2, err);
9661022 		return;
9671023 	}
9681024 
9691025 	/* Save state of TEAR and DEAR register contents */
970
-	if (!qsi(&si)) {
1026
+	err = qsi(&si);
1027
+	if (!err) {
9711028 		/* TEAR/DEAR values are valid only if the sampling facility is
9721029 		 * enabled.  Note that cpumsf_pmu_disable() might be called even
9731030 		 * for a disabled sampling facility because cpumsf_pmu_enable()
....@@ -978,8 +1035,8 @@
9781035 			cpuhw->lsctl.dear = si.dear;
9791036 		}
9801037 	} else
981
-		debug_sprintf_event(sfdbg, 3, "cpumsf_pmu_disable: "
982
-				    "qsi() failed with err=%i\n", err);
1038
+		debug_sprintf_event(sfdbg, 3, "%s: qsi() failed with err %i\n",
1039
+				    __func__, err);
9831040 
9841041 	cpuhw->flags &= ~PMU_F_ENABLED;
9851042 }
....@@ -1092,14 +1149,6 @@
10921149 	local64_add(count, &event->count);
10931150 }
10941151 
1095
-static void debug_sample_entry(struct hws_basic_entry *sample,
1096
-			       struct hws_trailer_entry *te)
1097
-{
1098
-	debug_sprintf_event(sfdbg, 4, "hw_collect_samples: Found unknown "
1099
-			    "sampling data entry: te->f=%i basic.def=%04x (%p)\n",
1100
-			    te->f, sample->def, sample);
1101
-}
1102
-
11031152 /* hw_collect_samples() - Walk through a sample-data-block and collect samples
11041153  * @event:	The perf event
11051154  * @sdbt:	Sample-data-block table
....@@ -1153,7 +1202,11 @@
11531202 				/* Count discarded samples */
11541203 				*overflow += 1;
11551204 		} else {
1156
-			debug_sample_entry(sample, te);
1205
+			debug_sprintf_event(sfdbg, 4,
1206
+					    "%s: Found unknown"
1207
+					    " sampling data entry: te->f %i"
1208
+					    " basic.def %#4x (%p)\n", __func__,
1209
+					    te->f, sample->def, sample);
11571210 			/* Sample slot is not yet written or other record.
11581211 			 *
11591212 			 * This condition can occur if the buffer was reused
....@@ -1228,9 +1281,9 @@
12281281 			sampl_overflow += te->overflow;
12291282 
12301283 		/* Timestamps are valid for full sample-data-blocks only */
1231
-		debug_sprintf_event(sfdbg, 6, "hw_perf_event_update: sdbt=%p "
1232
-				    "overflow=%llu timestamp=0x%llx\n",
1233
-				    sdbt, te->overflow,
1284
+		debug_sprintf_event(sfdbg, 6, "%s: sdbt %#lx "
1285
+				    "overflow %llu timestamp %#llx\n",
1286
+				    __func__, (unsigned long)sdbt, te->overflow,
12341287 				    (te->f) ? trailer_timestamp(te) : 0ULL);
12351288 
12361289 		/* Collect all samples from a single sample-data-block and
....@@ -1284,9 +1337,11 @@
12841337 	}
12851338 
12861339 	if (sampl_overflow || event_overflow)
1287
-		debug_sprintf_event(sfdbg, 4, "hw_perf_event_update: "
1288
-				    "overflow stats: sample=%llu event=%llu\n",
1289
-				    sampl_overflow, event_overflow);
1340
+		debug_sprintf_event(sfdbg, 4, "%s: "
1341
+				    "overflows: sample %llu event %llu"
1342
+				    " total %llu num_sdb %llu\n",
1343
+				    __func__, sampl_overflow, event_overflow,
1344
+				    OVERFLOW_REG(hwc), num_sdb);
12901345 }
12911346 
12921347 #define AUX_SDB_INDEX(aux, i) ((i) % aux->sfb.num_sdb)
....@@ -1339,7 +1394,8 @@
13391394 	te = aux_sdb_trailer(aux, aux->alert_mark);
13401395 	te->flags &= ~SDB_TE_ALERT_REQ_MASK;
13411396 
1342
-	debug_sprintf_event(sfdbg, 6, "aux_output_end: collect %lx SDBs\n", i);
1397
+	debug_sprintf_event(sfdbg, 6, "%s: SDBs %ld range %ld head %ld\n",
1398
+			    __func__, i, range_scan, aux->head);
13431399 }
13441400 
13451401 /*
....@@ -1372,6 +1428,10 @@
13721428 	 * SDBs between aux->head and aux->empty_mark are already ready
13731429 	 * for new data. range_scan is num of SDBs not within them.
13741430 	 */
1431
+	debug_sprintf_event(sfdbg, 6,
1432
+			    "%s: range %ld head %ld alert %ld empty %ld\n",
1433
+			    __func__, range, aux->head, aux->alert_mark,
1434
+			    aux->empty_mark);
13751435 	if (range > AUX_SDB_NUM_EMPTY(aux)) {
13761436 		range_scan = range - AUX_SDB_NUM_EMPTY(aux);
13771437 		idx = aux->empty_mark + 1;
....@@ -1397,15 +1457,11 @@
13971457 	cpuhw->lsctl.tear = base + offset * sizeof(unsigned long);
13981458 	cpuhw->lsctl.dear = aux->sdb_index[head];
13991459 
1400
-	debug_sprintf_event(sfdbg, 6, "aux_output_begin: "
1401
-			    "head->alert_mark->empty_mark (num_alert, range)"
1402
-			    "[%lx -> %lx -> %lx] (%lx, %lx) "
1403
-			    "tear index %lx, tear %lx dear %lx\n",
1460
+	debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld empty %ld "
1461
+			    "index %ld tear %#lx dear %#lx\n", __func__,
14041462 			    aux->head, aux->alert_mark, aux->empty_mark,
1405
-			    AUX_SDB_NUM_ALERT(aux), range,
14061463 			    head / CPUM_SF_SDB_PER_TABLE,
1407
-			    cpuhw->lsctl.tear,
1408
-			    cpuhw->lsctl.dear);
1464
+			    cpuhw->lsctl.tear, cpuhw->lsctl.dear);
14091465 
14101466 	return 0;
14111467 }
....@@ -1467,9 +1523,12 @@
14671523 			     unsigned long long *overflow)
14681524 {
14691525 	unsigned long long orig_overflow, orig_flags, new_flags;
1470
-	unsigned long i, range_scan, idx;
1526
+	unsigned long i, range_scan, idx, idx_old;
14711527 	struct hws_trailer_entry *te;
14721528 
1529
+	debug_sprintf_event(sfdbg, 6, "%s: range %ld head %ld alert %ld "
1530
+			    "empty %ld\n", __func__, range, aux->head,
1531
+			    aux->alert_mark, aux->empty_mark);
14731532 	if (range <= AUX_SDB_NUM_EMPTY(aux))
14741533 		/*
14751534 		 * No need to scan. All SDBs in range are marked as empty.
....@@ -1492,7 +1551,7 @@
14921551 	 * indicator fall into this range, set it.
14931552 	 */
14941553 	range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1495
-	idx = aux->empty_mark + 1;
1554
+	idx_old = idx = aux->empty_mark + 1;
14961555 	for (i = 0; i < range_scan; i++, idx++) {
14971556 		te = aux_sdb_trailer(aux, idx);
14981557 		do {
....@@ -1512,6 +1571,9 @@
15121571 	/* Update empty_mark to new position */
15131572 	aux->empty_mark = aux->head + range - 1;
15141573 
1574
+	debug_sprintf_event(sfdbg, 6, "%s: range_scan %ld idx %ld..%ld "
1575
+			    "empty %ld\n", __func__, range_scan, idx_old,
1576
+			    idx - 1, aux->empty_mark);
15151577 	return true;
15161578 }
15171579 
....@@ -1533,8 +1595,9 @@
15331595 
15341596 	/* Inform user space new data arrived */
15351597 	size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1598
+	debug_sprintf_event(sfdbg, 6, "%s: #alert %ld\n", __func__,
1599
+			    size >> PAGE_SHIFT);
15361600 	perf_aux_output_end(handle, size);
1537
-	num_sdb = aux->sfb.num_sdb;
15381601 
15391602 	num_sdb = aux->sfb.num_sdb;
15401603 	while (!done) {
....@@ -1544,7 +1607,9 @@
15441607 			pr_err("The AUX buffer with %lu pages for the "
15451608 			       "diagnostic-sampling mode is full\n",
15461609 				num_sdb);
1547
-			debug_sprintf_event(sfdbg, 1, "AUX buffer used up\n");
1610
+			debug_sprintf_event(sfdbg, 1,
1611
+					    "%s: AUX buffer used up\n",
1612
+					    __func__);
15481613 			break;
15491614 		}
15501615 		if (WARN_ON_ONCE(!aux))
....@@ -1566,24 +1631,24 @@
15661631 			size = range << PAGE_SHIFT;
15671632 			perf_aux_output_end(&cpuhw->handle, size);
15681633 			pr_err("Sample data caused the AUX buffer with %lu "
1569
-			       "pages to overflow\n", num_sdb);
1570
-			debug_sprintf_event(sfdbg, 1, "head %lx range %lx "
1571
-					    "overflow %llx\n",
1634
+			       "pages to overflow\n", aux->sfb.num_sdb);
1635
+			debug_sprintf_event(sfdbg, 1, "%s: head %ld range %ld "
1636
+					    "overflow %lld\n", __func__,
15721637 					    aux->head, range, overflow);
15731638 		} else {
15741639 			size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
15751640 			perf_aux_output_end(&cpuhw->handle, size);
1576
-			debug_sprintf_event(sfdbg, 6, "head %lx alert %lx "
1641
+			debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld "
15771642 					    "already full, try another\n",
1643
+					    __func__,
15781644 					    aux->head, aux->alert_mark);
15791645 		}
15801646 	}
15811647 
15821648 	if (done)
1583
-		debug_sprintf_event(sfdbg, 6, "aux_reset_buffer: "
1584
-				    "[%lx -> %lx -> %lx] (%lx, %lx)\n",
1585
-				    aux->head, aux->alert_mark, aux->empty_mark,
1586
-				    AUX_SDB_NUM_ALERT(aux), range);
1649
+		debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld "
1650
+				    "empty %ld\n", __func__, aux->head,
1651
+				    aux->alert_mark, aux->empty_mark);
15871652 }
15881653 
15891654 /*
....@@ -1606,8 +1671,7 @@
16061671 	kfree(aux->sdb_index);
16071672 	kfree(aux);
16081673 
1609
-	debug_sprintf_event(sfdbg, 4, "aux_buffer_free: free "
1610
-			    "%lu SDBTs\n", num_sdbt);
1674
+	debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu\n", __func__, num_sdbt);
16111675 }
16121676 
16131677 static void aux_sdb_init(unsigned long sdb)
....@@ -1665,7 +1729,7 @@
16651729 	sfb = &aux->sfb;
16661730 
16671731 	/* Allocate sdbt_index for fast reference */
1668
-	n_sdbt = (nr_pages + CPUM_SF_SDB_PER_TABLE - 1) / CPUM_SF_SDB_PER_TABLE;
1732
+	n_sdbt = DIV_ROUND_UP(nr_pages, CPUM_SF_SDB_PER_TABLE);
16691733 	aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL);
16701734 	if (!aux->sdbt_index)
16711735 		goto no_sdbt_index;
....@@ -1715,8 +1779,7 @@
17151779 	 */
17161780 	aux->empty_mark = sfb->num_sdb - 1;
17171781 
1718
-	debug_sprintf_event(sfdbg, 4, "aux_buffer_setup: setup %lu SDBTs"
1719
-			    " and %lu SDBs\n",
1782
+	debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu SDBs %lu\n", __func__,
17201783 			    sfb->num_sdbt, sfb->num_sdb);
17211784 
17221785 	return aux;
....@@ -1737,6 +1800,44 @@
17371800 static void cpumsf_pmu_read(struct perf_event *event)
17381801 {
17391802 	/* Nothing to do ... updates are interrupt-driven */
1803
+}
1804
+
1805
+/* Check if the new sampling period/freqeuncy is appropriate.
1806
+ *
1807
+ * Return non-zero on error and zero on passed checks.
1808
+ */
1809
+static int cpumsf_pmu_check_period(struct perf_event *event, u64 value)
1810
+{
1811
+	struct hws_qsi_info_block si;
1812
+	unsigned long rate;
1813
+	bool do_freq;
1814
+
1815
+	memset(&si, 0, sizeof(si));
1816
+	if (event->cpu == -1) {
1817
+		if (qsi(&si))
1818
+			return -ENODEV;
1819
+	} else {
1820
+		/* Event is pinned to a particular CPU, retrieve the per-CPU
1821
+		 * sampling structure for accessing the CPU-specific QSI.
1822
+		 */
1823
+		struct cpu_hw_sf *cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
1824
+
1825
+		si = cpuhw->qsi;
1826
+	}
1827
+
1828
+	do_freq = !!SAMPLE_FREQ_MODE(&event->hw);
1829
+	rate = getrate(do_freq, value, &si);
1830
+	if (!rate)
1831
+		return -EINVAL;
1832
+
1833
+	event->attr.sample_period = rate;
1834
+	SAMPL_RATE(&event->hw) = rate;
1835
+	hw_init_period(&event->hw, SAMPL_RATE(&event->hw));
1836
+	debug_sprintf_event(sfdbg, 4, "%s:"
1837
+			    " cpu %d value %#llx period %#llx freq %d\n",
1838
+			    __func__, event->cpu, value,
1839
+			    event->attr.sample_period, do_freq);
1840
+	return 0;
17401841 }
17411842 
17421843 /* Activate sampling control.
....@@ -1810,7 +1911,7 @@
18101911 	if (!SAMPL_DIAG_MODE(&event->hw)) {
18111912 		cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt;
18121913 		cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
1813
-		hw_reset_registers(&event->hw, cpuhw->sfb.sdbt);
1914
+		TEAR_REG(&event->hw) = (unsigned long) cpuhw->sfb.sdbt;
18141915 	}
18151916 
18161917 	/* Ensure sampling functions are in the disabled state.  If disabled,
....@@ -1865,10 +1966,30 @@
18651966 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
18661967 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
18671968 
1868
-static struct attribute *cpumsf_pmu_events_attr[] = {
1869
-	CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC),
1870
-	NULL,
1871
-	NULL,
1969
+/* Attribute list for CPU_SF.
1970
+ *
1971
+ * The availablitiy depends on the CPU_MF sampling facility authorization
1972
+ * for basic + diagnositic samples. This is determined at initialization
1973
+ * time by the sampling facility device driver.
1974
+ * If the authorization for basic samples is turned off, it should be
1975
+ * also turned off for diagnostic sampling.
1976
+ *
1977
+ * During initialization of the device driver, check the authorization
1978
+ * level for diagnostic sampling and installs the attribute
1979
+ * file for diagnostic sampling if necessary.
1980
+ *
1981
+ * For now install a placeholder to reference all possible attributes:
1982
+ * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG.
1983
+ * Add another entry for the final NULL pointer.
1984
+ */
1985
+enum {
1986
+	SF_CYCLES_BASIC_ATTR_IDX = 0,
1987
+	SF_CYCLES_BASIC_DIAG_ATTR_IDX,
1988
+	SF_CYCLES_ATTR_MAX
1989
+};
1990
+
1991
+static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = {
1992
+	[SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC)
18721993 };
18731994 
18741995 PMU_FORMAT_ATTR(event, "config:0-63");
....@@ -1882,10 +2003,12 @@
18822003 	.name = "events",
18832004 	.attrs = cpumsf_pmu_events_attr,
18842005 };
2006
+
18852007 static struct attribute_group cpumsf_pmu_format_group = {
18862008 	.name = "format",
18872009 	.attrs = cpumsf_pmu_format_attr,
18882010 };
2011
+
18892012 static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
18902013 	&cpumsf_pmu_events_group,
18912014 	&cpumsf_pmu_format_group,
....@@ -1908,6 +2031,8 @@
19082031 
19092032 	.setup_aux    = aux_buffer_setup,
19102033 	.free_aux     = aux_buffer_free,
2034
+
2035
+	.check_period = cpumsf_pmu_check_period,
19112036 };
19122037 
19132038 static void cpumf_measurement_alert(struct ext_code ext_code,
....@@ -1941,7 +2066,8 @@
19412066 
19422067 	/* Report measurement alerts only for non-PRA codes */
19432068 	if (alert != CPU_MF_INT_SF_PRA)
1944
-		debug_sprintf_event(sfdbg, 6, "measurement alert: 0x%x\n", alert);
2069
+		debug_sprintf_event(sfdbg, 6, "%s: alert %#x\n", __func__,
2070
+				    alert);
19452071 
19462072 	/* Sampling authorization change request */
19472073 	if (alert & CPU_MF_INT_SF_SACA)
....@@ -1962,6 +2088,7 @@
19622088 		sf_disable();
19632089 	}
19642090 }
2091
+
19652092 static int cpusf_pmu_setup(unsigned int cpu, int flags)
19662093 {
19672094 	/* Ignore the notification if no events are scheduled on the PMU.
....@@ -2018,7 +2145,7 @@
20182145 
20192146 	sfb_set_limits(min, max);
20202147 	pr_info("The sampling buffer limits have changed to: "
2021
-		"min=%lu max=%lu (diag=x%lu)\n",
2148
+		"min %lu max %lu (diag %lu)\n",
20222149 		CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
20232150 	return 0;
20242151 }
....@@ -2036,7 +2163,7 @@
20362163 static void __init pr_cpumsf_err(unsigned int reason)
20372164 {
20382165 	pr_err("Sampling facility support for perf is not available: "
2039
-	       "reason=%04x\n", reason);
2166
+	       "reason %#x\n", reason);
20402167 }
20412168 
20422169 static int __init init_cpum_sampling_pmu(void)
....@@ -2063,7 +2190,10 @@
20632190 
20642191 	if (si.ad) {
20652192 		sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2066
-		cpumsf_pmu_events_attr[1] =
2193
+		/* Sampling of diagnostic data authorized,
2194
+		 * install event into attribute list of PMU device.
2195
+		 */
2196
+		cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] =
20672197 			CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
20682198 	}
20692199 
....@@ -2096,5 +2226,6 @@
20962226 out:
20972227 	return err;
20982228 }
2229
+
20992230 arch_initcall(init_cpum_sampling_pmu);
21002231 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0644);