1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Memory bandwidth monitoring and allocation library
4  *
5  * Copyright (C) 2018 Intel Corporation
6  *
7  * Authors:
8  *    Sai Praneeth Prakhya <sai.praneeth.prakhya@intel.com>,
9  *    Fenghua Yu <fenghua.yu@intel.com>
10  */
11 #include "resctrl.h"
12 
13 #define UNCORE_IMC		"uncore_imc"
14 #define READ_FILE_NAME		"events/cas_count_read"
15 #define WRITE_FILE_NAME		"events/cas_count_write"
16 #define DYN_PMU_PATH		"/sys/bus/event_source/devices"
17 #define SCALE			0.00006103515625
18 #define MAX_IMCS		20
19 #define MAX_TOKENS		5
20 #define READ			0
21 #define WRITE			1
22 #define CON_MON_MBM_LOCAL_BYTES_PATH				\
23 	"%s/%s/mon_groups/%s/mon_data/mon_L3_%02d/mbm_local_bytes"
24 
25 #define CON_MBM_LOCAL_BYTES_PATH		\
26 	"%s/%s/mon_data/mon_L3_%02d/mbm_local_bytes"
27 
28 #define MON_MBM_LOCAL_BYTES_PATH		\
29 	"%s/mon_groups/%s/mon_data/mon_L3_%02d/mbm_local_bytes"
30 
31 #define MBM_LOCAL_BYTES_PATH			\
32 	"%s/mon_data/mon_L3_%02d/mbm_local_bytes"
33 
34 #define CON_MON_LCC_OCCUP_PATH		\
35 	"%s/%s/mon_groups/%s/mon_data/mon_L3_%02d/llc_occupancy"
36 
37 #define CON_LCC_OCCUP_PATH		\
38 	"%s/%s/mon_data/mon_L3_%02d/llc_occupancy"
39 
40 #define MON_LCC_OCCUP_PATH		\
41 	"%s/mon_groups/%s/mon_data/mon_L3_%02d/llc_occupancy"
42 
43 #define LCC_OCCUP_PATH			\
44 	"%s/mon_data/mon_L3_%02d/llc_occupancy"
45 
46 struct membw_read_format {
47 	__u64 value;         /* The value of the event */
48 	__u64 time_enabled;  /* if PERF_FORMAT_TOTAL_TIME_ENABLED */
49 	__u64 time_running;  /* if PERF_FORMAT_TOTAL_TIME_RUNNING */
50 	__u64 id;            /* if PERF_FORMAT_ID */
51 };
52 
53 struct imc_counter_config {
54 	__u32 type;
55 	__u64 event;
56 	__u64 umask;
57 	struct perf_event_attr pe;
58 	struct membw_read_format return_value;
59 	int fd;
60 };
61 
62 static char mbm_total_path[1024];
63 static int imcs;
64 static struct imc_counter_config imc_counters_config[MAX_IMCS][2];
65 
66 void membw_initialize_perf_event_attr(int i, int j)
67 {
68 	memset(&imc_counters_config[i][j].pe, 0,
69 	       sizeof(struct perf_event_attr));
70 	imc_counters_config[i][j].pe.type = imc_counters_config[i][j].type;
71 	imc_counters_config[i][j].pe.size = sizeof(struct perf_event_attr);
72 	imc_counters_config[i][j].pe.disabled = 1;
73 	imc_counters_config[i][j].pe.inherit = 1;
74 	imc_counters_config[i][j].pe.exclude_guest = 0;
75 	imc_counters_config[i][j].pe.config =
76 		imc_counters_config[i][j].umask << 8 |
77 		imc_counters_config[i][j].event;
78 	imc_counters_config[i][j].pe.sample_type = PERF_SAMPLE_IDENTIFIER;
79 	imc_counters_config[i][j].pe.read_format =
80 		PERF_FORMAT_TOTAL_TIME_ENABLED | PERF_FORMAT_TOTAL_TIME_RUNNING;
81 }
82 
83 void membw_ioctl_perf_event_ioc_reset_enable(int i, int j)
84 {
85 	ioctl(imc_counters_config[i][j].fd, PERF_EVENT_IOC_RESET, 0);
86 	ioctl(imc_counters_config[i][j].fd, PERF_EVENT_IOC_ENABLE, 0);
87 }
88 
89 void membw_ioctl_perf_event_ioc_disable(int i, int j)
90 {
91 	ioctl(imc_counters_config[i][j].fd, PERF_EVENT_IOC_DISABLE, 0);
92 }
93 
94 /*
95  * get_event_and_umask:	Parse config into event and umask
96  * @cas_count_cfg:	Config
97  * @count:		iMC number
98  * @op:			Operation (read/write)
99  */
100 void get_event_and_umask(char *cas_count_cfg, int count, bool op)
101 {
102 	char *token[MAX_TOKENS];
103 	int i = 0;
104 
105 	strcat(cas_count_cfg, ",");
106 	token[0] = strtok(cas_count_cfg, "=,");
107 
108 	for (i = 1; i < MAX_TOKENS; i++)
109 		token[i] = strtok(NULL, "=,");
110 
111 	for (i = 0; i < MAX_TOKENS; i++) {
112 		if (!token[i])
113 			break;
114 		if (strcmp(token[i], "event") == 0) {
115 			if (op == READ)
116 				imc_counters_config[count][READ].event =
117 				strtol(token[i + 1], NULL, 16);
118 			else
119 				imc_counters_config[count][WRITE].event =
120 				strtol(token[i + 1], NULL, 16);
121 		}
122 		if (strcmp(token[i], "umask") == 0) {
123 			if (op == READ)
124 				imc_counters_config[count][READ].umask =
125 				strtol(token[i + 1], NULL, 16);
126 			else
127 				imc_counters_config[count][WRITE].umask =
128 				strtol(token[i + 1], NULL, 16);
129 		}
130 	}
131 }
132 
133 static int open_perf_event(int i, int cpu_no, int j)
134 {
135 	imc_counters_config[i][j].fd =
136 		perf_event_open(&imc_counters_config[i][j].pe, -1, cpu_no, -1,
137 				PERF_FLAG_FD_CLOEXEC);
138 
139 	if (imc_counters_config[i][j].fd == -1) {
140 		fprintf(stderr, "Error opening leader %llx\n",
141 			imc_counters_config[i][j].pe.config);
142 
143 		return -1;
144 	}
145 
146 	return 0;
147 }
148 
149 /* Get type and config (read and write) of an iMC counter */
150 static int read_from_imc_dir(char *imc_dir, int count)
151 {
152 	char cas_count_cfg[1024], imc_counter_cfg[1024], imc_counter_type[1024];
153 	FILE *fp;
154 
155 	/* Get type of iMC counter */
156 	sprintf(imc_counter_type, "%s%s", imc_dir, "type");
157 	fp = fopen(imc_counter_type, "r");
158 	if (!fp) {
159 		perror("Failed to open imc counter type file");
160 
161 		return -1;
162 	}
163 	if (fscanf(fp, "%u", &imc_counters_config[count][READ].type) <= 0) {
164 		perror("Could not get imc type");
165 		fclose(fp);
166 
167 		return -1;
168 	}
169 	fclose(fp);
170 
171 	imc_counters_config[count][WRITE].type =
172 				imc_counters_config[count][READ].type;
173 
174 	/* Get read config */
175 	sprintf(imc_counter_cfg, "%s%s", imc_dir, READ_FILE_NAME);
176 	fp = fopen(imc_counter_cfg, "r");
177 	if (!fp) {
178 		perror("Failed to open imc config file");
179 
180 		return -1;
181 	}
182 	if (fscanf(fp, "%s", cas_count_cfg) <= 0) {
183 		perror("Could not get imc cas count read");
184 		fclose(fp);
185 
186 		return -1;
187 	}
188 	fclose(fp);
189 
190 	get_event_and_umask(cas_count_cfg, count, READ);
191 
192 	/* Get write config */
193 	sprintf(imc_counter_cfg, "%s%s", imc_dir, WRITE_FILE_NAME);
194 	fp = fopen(imc_counter_cfg, "r");
195 	if (!fp) {
196 		perror("Failed to open imc config file");
197 
198 		return -1;
199 	}
200 	if  (fscanf(fp, "%s", cas_count_cfg) <= 0) {
201 		perror("Could not get imc cas count write");
202 		fclose(fp);
203 
204 		return -1;
205 	}
206 	fclose(fp);
207 
208 	get_event_and_umask(cas_count_cfg, count, WRITE);
209 
210 	return 0;
211 }
212 
213 /*
214  * A system can have 'n' number of iMC (Integrated Memory Controller)
215  * counters, get that 'n'. For each iMC counter get it's type and config.
216  * Also, each counter has two configs, one for read and the other for write.
217  * A config again has two parts, event and umask.
218  * Enumerate all these details into an array of structures.
219  *
220  * Return: >= 0 on success. < 0 on failure.
221  */
222 static int num_of_imcs(void)
223 {
224 	char imc_dir[512], *temp;
225 	unsigned int count = 0;
226 	struct dirent *ep;
227 	int ret;
228 	DIR *dp;
229 
230 	dp = opendir(DYN_PMU_PATH);
231 	if (dp) {
232 		while ((ep = readdir(dp))) {
233 			temp = strstr(ep->d_name, UNCORE_IMC);
234 			if (!temp)
235 				continue;
236 
237 			/*
238 			 * imc counters are named as "uncore_imc_<n>", hence
239 			 * increment the pointer to point to <n>. Note that
240 			 * sizeof(UNCORE_IMC) would count for null character as
241 			 * well and hence the last underscore character in
242 			 * uncore_imc'_' need not be counted.
243 			 */
244 			temp = temp + sizeof(UNCORE_IMC);
245 
246 			/*
247 			 * Some directories under "DYN_PMU_PATH" could have
248 			 * names like "uncore_imc_free_running", hence, check if
249 			 * first character is a numerical digit or not.
250 			 */
251 			if (temp[0] >= '0' && temp[0] <= '9') {
252 				sprintf(imc_dir, "%s/%s/", DYN_PMU_PATH,
253 					ep->d_name);
254 				ret = read_from_imc_dir(imc_dir, count);
255 				if (ret) {
256 					closedir(dp);
257 
258 					return ret;
259 				}
260 				count++;
261 			}
262 		}
263 		closedir(dp);
264 		if (count == 0) {
265 			perror("Unable find iMC counters!\n");
266 
267 			return -1;
268 		}
269 	} else {
270 		perror("Unable to open PMU directory!\n");
271 
272 		return -1;
273 	}
274 
275 	return count;
276 }
277 
278 static int initialize_mem_bw_imc(void)
279 {
280 	int imc, j;
281 
282 	imcs = num_of_imcs();
283 	if (imcs <= 0)
284 		return imcs;
285 
286 	/* Initialize perf_event_attr structures for all iMC's */
287 	for (imc = 0; imc < imcs; imc++) {
288 		for (j = 0; j < 2; j++)
289 			membw_initialize_perf_event_attr(imc, j);
290 	}
291 
292 	return 0;
293 }
294 
295 /*
296  * get_mem_bw_imc:	Memory band width as reported by iMC counters
297  * @cpu_no:		CPU number that the benchmark PID is binded to
298  * @bw_report:		Bandwidth report type (reads, writes)
299  *
300  * Memory B/W utilized by a process on a socket can be calculated using
301  * iMC counters. Perf events are used to read these counters.
302  *
303  * Return: = 0 on success. < 0 on failure.
304  */
305 static int get_mem_bw_imc(int cpu_no, char *bw_report, float *bw_imc)
306 {
307 	float reads, writes, of_mul_read, of_mul_write;
308 	int imc, j, ret;
309 
310 	/* Start all iMC counters to log values (both read and write) */
311 	reads = 0, writes = 0, of_mul_read = 1, of_mul_write = 1;
312 	for (imc = 0; imc < imcs; imc++) {
313 		for (j = 0; j < 2; j++) {
314 			ret = open_perf_event(imc, cpu_no, j);
315 			if (ret)
316 				return -1;
317 		}
318 		for (j = 0; j < 2; j++)
319 			membw_ioctl_perf_event_ioc_reset_enable(imc, j);
320 	}
321 
322 	sleep(1);
323 
324 	/* Stop counters after a second to get results (both read and write) */
325 	for (imc = 0; imc < imcs; imc++) {
326 		for (j = 0; j < 2; j++)
327 			membw_ioctl_perf_event_ioc_disable(imc, j);
328 	}
329 
330 	/*
331 	 * Get results which are stored in struct type imc_counter_config
332 	 * Take over flow into consideration before calculating total b/w
333 	 */
334 	for (imc = 0; imc < imcs; imc++) {
335 		struct imc_counter_config *r =
336 			&imc_counters_config[imc][READ];
337 		struct imc_counter_config *w =
338 			&imc_counters_config[imc][WRITE];
339 
340 		if (read(r->fd, &r->return_value,
341 			 sizeof(struct membw_read_format)) == -1) {
342 			perror("Couldn't get read b/w through iMC");
343 
344 			return -1;
345 		}
346 
347 		if (read(w->fd, &w->return_value,
348 			 sizeof(struct membw_read_format)) == -1) {
349 			perror("Couldn't get write bw through iMC");
350 
351 			return -1;
352 		}
353 
354 		__u64 r_time_enabled = r->return_value.time_enabled;
355 		__u64 r_time_running = r->return_value.time_running;
356 
357 		if (r_time_enabled != r_time_running)
358 			of_mul_read = (float)r_time_enabled /
359 					(float)r_time_running;
360 
361 		__u64 w_time_enabled = w->return_value.time_enabled;
362 		__u64 w_time_running = w->return_value.time_running;
363 
364 		if (w_time_enabled != w_time_running)
365 			of_mul_write = (float)w_time_enabled /
366 					(float)w_time_running;
367 		reads += r->return_value.value * of_mul_read * SCALE;
368 		writes += w->return_value.value * of_mul_write * SCALE;
369 	}
370 
371 	for (imc = 0; imc < imcs; imc++) {
372 		close(imc_counters_config[imc][READ].fd);
373 		close(imc_counters_config[imc][WRITE].fd);
374 	}
375 
376 	if (strcmp(bw_report, "reads") == 0) {
377 		*bw_imc = reads;
378 		return 0;
379 	}
380 
381 	if (strcmp(bw_report, "writes") == 0) {
382 		*bw_imc = writes;
383 		return 0;
384 	}
385 
386 	*bw_imc = reads + writes;
387 	return 0;
388 }
389 
390 void set_mbm_path(const char *ctrlgrp, const char *mongrp, int resource_id)
391 {
392 	if (ctrlgrp && mongrp)
393 		sprintf(mbm_total_path, CON_MON_MBM_LOCAL_BYTES_PATH,
394 			RESCTRL_PATH, ctrlgrp, mongrp, resource_id);
395 	else if (!ctrlgrp && mongrp)
396 		sprintf(mbm_total_path, MON_MBM_LOCAL_BYTES_PATH, RESCTRL_PATH,
397 			mongrp, resource_id);
398 	else if (ctrlgrp && !mongrp)
399 		sprintf(mbm_total_path, CON_MBM_LOCAL_BYTES_PATH, RESCTRL_PATH,
400 			ctrlgrp, resource_id);
401 	else if (!ctrlgrp && !mongrp)
402 		sprintf(mbm_total_path, MBM_LOCAL_BYTES_PATH, RESCTRL_PATH,
403 			resource_id);
404 }
405 
406 /*
407  * initialize_mem_bw_resctrl:	Appropriately populate "mbm_total_path"
408  * @ctrlgrp:			Name of the control monitor group (con_mon grp)
409  * @mongrp:			Name of the monitor group (mon grp)
410  * @cpu_no:			CPU number that the benchmark PID is binded to
411  * @resctrl_val:		Resctrl feature (Eg: mbm, mba.. etc)
412  */
413 static void initialize_mem_bw_resctrl(const char *ctrlgrp, const char *mongrp,
414 				      int cpu_no, char *resctrl_val)
415 {
416 	int resource_id;
417 
418 	if (get_resource_id(cpu_no, &resource_id) < 0) {
419 		perror("Could not get resource_id");
420 		return;
421 	}
422 
423 	if (!strncmp(resctrl_val, MBM_STR, sizeof(MBM_STR)))
424 		set_mbm_path(ctrlgrp, mongrp, resource_id);
425 
426 	if (!strncmp(resctrl_val, MBA_STR, sizeof(MBA_STR))) {
427 		if (ctrlgrp)
428 			sprintf(mbm_total_path, CON_MBM_LOCAL_BYTES_PATH,
429 				RESCTRL_PATH, ctrlgrp, resource_id);
430 		else
431 			sprintf(mbm_total_path, MBM_LOCAL_BYTES_PATH,
432 				RESCTRL_PATH, resource_id);
433 	}
434 }
435 
436 /*
437  * Get MBM Local bytes as reported by resctrl FS
438  * For MBM,
439  * 1. If con_mon grp and mon grp are given, then read from con_mon grp's mon grp
440  * 2. If only con_mon grp is given, then read from con_mon grp
441  * 3. If both are not given, then read from root con_mon grp
442  * For MBA,
443  * 1. If con_mon grp is given, then read from it
444  * 2. If con_mon grp is not given, then read from root con_mon grp
445  */
446 static int get_mem_bw_resctrl(unsigned long *mbm_total)
447 {
448 	FILE *fp;
449 
450 	fp = fopen(mbm_total_path, "r");
451 	if (!fp) {
452 		perror("Failed to open total bw file");
453 
454 		return -1;
455 	}
456 	if (fscanf(fp, "%lu", mbm_total) <= 0) {
457 		perror("Could not get mbm local bytes");
458 		fclose(fp);
459 
460 		return -1;
461 	}
462 	fclose(fp);
463 
464 	return 0;
465 }
466 
467 pid_t bm_pid, ppid;
468 
469 void ctrlc_handler(int signum, siginfo_t *info, void *ptr)
470 {
471 	/* Only kill child after bm_pid is set after fork() */
472 	if (bm_pid)
473 		kill(bm_pid, SIGKILL);
474 	umount_resctrlfs();
475 	tests_cleanup();
476 	ksft_print_msg("Ending\n\n");
477 
478 	exit(EXIT_SUCCESS);
479 }
480 
481 /*
482  * Register CTRL-C handler for parent, as it has to kill
483  * child process before exiting.
484  */
485 int signal_handler_register(void)
486 {
487 	struct sigaction sigact = {};
488 	int ret = 0;
489 
490 	bm_pid = 0;
491 
492 	sigact.sa_sigaction = ctrlc_handler;
493 	sigemptyset(&sigact.sa_mask);
494 	sigact.sa_flags = SA_SIGINFO;
495 	if (sigaction(SIGINT, &sigact, NULL) ||
496 	    sigaction(SIGTERM, &sigact, NULL) ||
497 	    sigaction(SIGHUP, &sigact, NULL)) {
498 		perror("# sigaction");
499 		ret = -1;
500 	}
501 	return ret;
502 }
503 
504 /*
505  * Reset signal handler to SIG_DFL.
506  * Non-Value return because the caller should keep
507  * the error code of other path even if sigaction fails.
508  */
509 void signal_handler_unregister(void)
510 {
511 	struct sigaction sigact = {};
512 
513 	sigact.sa_handler = SIG_DFL;
514 	sigemptyset(&sigact.sa_mask);
515 	if (sigaction(SIGINT, &sigact, NULL) ||
516 	    sigaction(SIGTERM, &sigact, NULL) ||
517 	    sigaction(SIGHUP, &sigact, NULL)) {
518 		perror("# sigaction");
519 	}
520 }
521 
522 /*
523  * print_results_bw:	the memory bandwidth results are stored in a file
524  * @filename:		file that stores the results
525  * @bm_pid:		child pid that runs benchmark
526  * @bw_imc:		perf imc counter value
527  * @bw_resc:		memory bandwidth value
528  *
529  * Return:		0 on success. non-zero on failure.
530  */
531 static int print_results_bw(char *filename,  int bm_pid, float bw_imc,
532 			    unsigned long bw_resc)
533 {
534 	unsigned long diff = fabs(bw_imc - bw_resc);
535 	FILE *fp;
536 
537 	if (strcmp(filename, "stdio") == 0 || strcmp(filename, "stderr") == 0) {
538 		printf("Pid: %d \t Mem_BW_iMC: %f \t ", bm_pid, bw_imc);
539 		printf("Mem_BW_resc: %lu \t Difference: %lu\n", bw_resc, diff);
540 	} else {
541 		fp = fopen(filename, "a");
542 		if (!fp) {
543 			perror("Cannot open results file");
544 
545 			return errno;
546 		}
547 		if (fprintf(fp, "Pid: %d \t Mem_BW_iMC: %f \t Mem_BW_resc: %lu \t Difference: %lu\n",
548 			    bm_pid, bw_imc, bw_resc, diff) <= 0) {
549 			fclose(fp);
550 			perror("Could not log results.");
551 
552 			return errno;
553 		}
554 		fclose(fp);
555 	}
556 
557 	return 0;
558 }
559 
560 static void set_cmt_path(const char *ctrlgrp, const char *mongrp, char sock_num)
561 {
562 	if (strlen(ctrlgrp) && strlen(mongrp))
563 		sprintf(llc_occup_path,	CON_MON_LCC_OCCUP_PATH,	RESCTRL_PATH,
564 			ctrlgrp, mongrp, sock_num);
565 	else if (!strlen(ctrlgrp) && strlen(mongrp))
566 		sprintf(llc_occup_path,	MON_LCC_OCCUP_PATH, RESCTRL_PATH,
567 			mongrp, sock_num);
568 	else if (strlen(ctrlgrp) && !strlen(mongrp))
569 		sprintf(llc_occup_path,	CON_LCC_OCCUP_PATH, RESCTRL_PATH,
570 			ctrlgrp, sock_num);
571 	else if (!strlen(ctrlgrp) && !strlen(mongrp))
572 		sprintf(llc_occup_path, LCC_OCCUP_PATH,	RESCTRL_PATH, sock_num);
573 }
574 
575 /*
576  * initialize_llc_occu_resctrl:	Appropriately populate "llc_occup_path"
577  * @ctrlgrp:			Name of the control monitor group (con_mon grp)
578  * @mongrp:			Name of the monitor group (mon grp)
579  * @cpu_no:			CPU number that the benchmark PID is binded to
580  * @resctrl_val:		Resctrl feature (Eg: cat, cmt.. etc)
581  */
582 static void initialize_llc_occu_resctrl(const char *ctrlgrp, const char *mongrp,
583 					int cpu_no, char *resctrl_val)
584 {
585 	int resource_id;
586 
587 	if (get_resource_id(cpu_no, &resource_id) < 0) {
588 		perror("# Unable to resource_id");
589 		return;
590 	}
591 
592 	if (!strncmp(resctrl_val, CMT_STR, sizeof(CMT_STR)))
593 		set_cmt_path(ctrlgrp, mongrp, resource_id);
594 }
595 
596 static int
597 measure_vals(struct resctrl_val_param *param, unsigned long *bw_resc_start)
598 {
599 	unsigned long bw_resc, bw_resc_end;
600 	float bw_imc;
601 	int ret;
602 
603 	/*
604 	 * Measure memory bandwidth from resctrl and from
605 	 * another source which is perf imc value or could
606 	 * be something else if perf imc event is not available.
607 	 * Compare the two values to validate resctrl value.
608 	 * It takes 1sec to measure the data.
609 	 */
610 	ret = get_mem_bw_imc(param->cpu_no, param->bw_report, &bw_imc);
611 	if (ret < 0)
612 		return ret;
613 
614 	ret = get_mem_bw_resctrl(&bw_resc_end);
615 	if (ret < 0)
616 		return ret;
617 
618 	bw_resc = (bw_resc_end - *bw_resc_start) / MB;
619 	ret = print_results_bw(param->filename, bm_pid, bw_imc, bw_resc);
620 	if (ret)
621 		return ret;
622 
623 	*bw_resc_start = bw_resc_end;
624 
625 	return 0;
626 }
627 
628 /*
629  * resctrl_val:	execute benchmark and measure memory bandwidth on
630  *			the benchmark
631  * @benchmark_cmd:	benchmark command and its arguments
632  * @param:		parameters passed to resctrl_val()
633  *
634  * Return:		0 on success. non-zero on failure.
635  */
636 int resctrl_val(const char * const *benchmark_cmd, struct resctrl_val_param *param)
637 {
638 	char *resctrl_val = param->resctrl_val;
639 	unsigned long bw_resc_start = 0;
640 	struct sigaction sigact;
641 	int ret = 0, pipefd[2];
642 	char pipe_message = 0;
643 	union sigval value;
644 
645 	if (strcmp(param->filename, "") == 0)
646 		sprintf(param->filename, "stdio");
647 
648 	if (!strncmp(resctrl_val, MBA_STR, sizeof(MBA_STR)) ||
649 	    !strncmp(resctrl_val, MBM_STR, sizeof(MBM_STR))) {
650 		ret = validate_bw_report_request(param->bw_report);
651 		if (ret)
652 			return ret;
653 	}
654 
655 	/*
656 	 * If benchmark wasn't successfully started by child, then child should
657 	 * kill parent, so save parent's pid
658 	 */
659 	ppid = getpid();
660 
661 	if (pipe(pipefd)) {
662 		perror("# Unable to create pipe");
663 
664 		return -1;
665 	}
666 
667 	/*
668 	 * Fork to start benchmark, save child's pid so that it can be killed
669 	 * when needed
670 	 */
671 	fflush(stdout);
672 	bm_pid = fork();
673 	if (bm_pid == -1) {
674 		perror("# Unable to fork");
675 
676 		return -1;
677 	}
678 
679 	if (bm_pid == 0) {
680 		/*
681 		 * Mask all signals except SIGUSR1, parent uses SIGUSR1 to
682 		 * start benchmark
683 		 */
684 		sigfillset(&sigact.sa_mask);
685 		sigdelset(&sigact.sa_mask, SIGUSR1);
686 
687 		sigact.sa_sigaction = run_benchmark;
688 		sigact.sa_flags = SA_SIGINFO;
689 
690 		/* Register for "SIGUSR1" signal from parent */
691 		if (sigaction(SIGUSR1, &sigact, NULL))
692 			PARENT_EXIT("Can't register child for signal");
693 
694 		/* Tell parent that child is ready */
695 		close(pipefd[0]);
696 		pipe_message = 1;
697 		if (write(pipefd[1], &pipe_message, sizeof(pipe_message)) <
698 		    sizeof(pipe_message)) {
699 			perror("# failed signaling parent process");
700 			close(pipefd[1]);
701 			return -1;
702 		}
703 		close(pipefd[1]);
704 
705 		/* Suspend child until delivery of "SIGUSR1" from parent */
706 		sigsuspend(&sigact.sa_mask);
707 
708 		PARENT_EXIT("Child is done");
709 	}
710 
711 	ksft_print_msg("Benchmark PID: %d\n", bm_pid);
712 
713 	/*
714 	 * The cast removes constness but nothing mutates benchmark_cmd within
715 	 * the context of this process. At the receiving process, it becomes
716 	 * argv, which is mutable, on exec() but that's after fork() so it
717 	 * doesn't matter for the process running the tests.
718 	 */
719 	value.sival_ptr = (void *)benchmark_cmd;
720 
721 	/* Taskset benchmark to specified cpu */
722 	ret = taskset_benchmark(bm_pid, param->cpu_no);
723 	if (ret)
724 		goto out;
725 
726 	/* Write benchmark to specified control&monitoring grp in resctrl FS */
727 	ret = write_bm_pid_to_resctrl(bm_pid, param->ctrlgrp, param->mongrp,
728 				      resctrl_val);
729 	if (ret)
730 		goto out;
731 
732 	if (!strncmp(resctrl_val, MBM_STR, sizeof(MBM_STR)) ||
733 	    !strncmp(resctrl_val, MBA_STR, sizeof(MBA_STR))) {
734 		ret = initialize_mem_bw_imc();
735 		if (ret)
736 			goto out;
737 
738 		initialize_mem_bw_resctrl(param->ctrlgrp, param->mongrp,
739 					  param->cpu_no, resctrl_val);
740 	} else if (!strncmp(resctrl_val, CMT_STR, sizeof(CMT_STR)))
741 		initialize_llc_occu_resctrl(param->ctrlgrp, param->mongrp,
742 					    param->cpu_no, resctrl_val);
743 
744 	/* Parent waits for child to be ready. */
745 	close(pipefd[1]);
746 	while (pipe_message != 1) {
747 		if (read(pipefd[0], &pipe_message, sizeof(pipe_message)) <
748 		    sizeof(pipe_message)) {
749 			perror("# failed reading message from child process");
750 			close(pipefd[0]);
751 			goto out;
752 		}
753 	}
754 	close(pipefd[0]);
755 
756 	/* Signal child to start benchmark */
757 	if (sigqueue(bm_pid, SIGUSR1, value) == -1) {
758 		perror("# sigqueue SIGUSR1 to child");
759 		ret = errno;
760 		goto out;
761 	}
762 
763 	/* Give benchmark enough time to fully run */
764 	sleep(1);
765 
766 	/* Test runs until the callback setup() tells the test to stop. */
767 	while (1) {
768 		ret = param->setup(param);
769 		if (ret == END_OF_TESTS) {
770 			ret = 0;
771 			break;
772 		}
773 		if (ret < 0)
774 			break;
775 
776 		if (!strncmp(resctrl_val, MBM_STR, sizeof(MBM_STR)) ||
777 		    !strncmp(resctrl_val, MBA_STR, sizeof(MBA_STR))) {
778 			ret = measure_vals(param, &bw_resc_start);
779 			if (ret)
780 				break;
781 		} else if (!strncmp(resctrl_val, CMT_STR, sizeof(CMT_STR))) {
782 			sleep(1);
783 			ret = measure_cache_vals(param, bm_pid);
784 			if (ret)
785 				break;
786 		}
787 	}
788 
789 out:
790 	kill(bm_pid, SIGKILL);
791 
792 	return ret;
793 }
794