xref: /linux/tools/testing/selftests/resctrl/resctrl_val.c (revision c5288cda69ee2d8607f5026bd599a5cebf0ee783)
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 static const struct resctrl_test *current_test;
66 
67 void membw_initialize_perf_event_attr(int i, int j)
68 {
69 	memset(&imc_counters_config[i][j].pe, 0,
70 	       sizeof(struct perf_event_attr));
71 	imc_counters_config[i][j].pe.type = imc_counters_config[i][j].type;
72 	imc_counters_config[i][j].pe.size = sizeof(struct perf_event_attr);
73 	imc_counters_config[i][j].pe.disabled = 1;
74 	imc_counters_config[i][j].pe.inherit = 1;
75 	imc_counters_config[i][j].pe.exclude_guest = 0;
76 	imc_counters_config[i][j].pe.config =
77 		imc_counters_config[i][j].umask << 8 |
78 		imc_counters_config[i][j].event;
79 	imc_counters_config[i][j].pe.sample_type = PERF_SAMPLE_IDENTIFIER;
80 	imc_counters_config[i][j].pe.read_format =
81 		PERF_FORMAT_TOTAL_TIME_ENABLED | PERF_FORMAT_TOTAL_TIME_RUNNING;
82 }
83 
84 void membw_ioctl_perf_event_ioc_reset_enable(int i, int j)
85 {
86 	ioctl(imc_counters_config[i][j].fd, PERF_EVENT_IOC_RESET, 0);
87 	ioctl(imc_counters_config[i][j].fd, PERF_EVENT_IOC_ENABLE, 0);
88 }
89 
90 void membw_ioctl_perf_event_ioc_disable(int i, int j)
91 {
92 	ioctl(imc_counters_config[i][j].fd, PERF_EVENT_IOC_DISABLE, 0);
93 }
94 
95 /*
96  * get_event_and_umask:	Parse config into event and umask
97  * @cas_count_cfg:	Config
98  * @count:		iMC number
99  * @op:			Operation (read/write)
100  */
101 void get_event_and_umask(char *cas_count_cfg, int count, bool op)
102 {
103 	char *token[MAX_TOKENS];
104 	int i = 0;
105 
106 	strcat(cas_count_cfg, ",");
107 	token[0] = strtok(cas_count_cfg, "=,");
108 
109 	for (i = 1; i < MAX_TOKENS; i++)
110 		token[i] = strtok(NULL, "=,");
111 
112 	for (i = 0; i < MAX_TOKENS; i++) {
113 		if (!token[i])
114 			break;
115 		if (strcmp(token[i], "event") == 0) {
116 			if (op == READ)
117 				imc_counters_config[count][READ].event =
118 				strtol(token[i + 1], NULL, 16);
119 			else
120 				imc_counters_config[count][WRITE].event =
121 				strtol(token[i + 1], NULL, 16);
122 		}
123 		if (strcmp(token[i], "umask") == 0) {
124 			if (op == READ)
125 				imc_counters_config[count][READ].umask =
126 				strtol(token[i + 1], NULL, 16);
127 			else
128 				imc_counters_config[count][WRITE].umask =
129 				strtol(token[i + 1], NULL, 16);
130 		}
131 	}
132 }
133 
134 static int open_perf_event(int i, int cpu_no, int j)
135 {
136 	imc_counters_config[i][j].fd =
137 		perf_event_open(&imc_counters_config[i][j].pe, -1, cpu_no, -1,
138 				PERF_FLAG_FD_CLOEXEC);
139 
140 	if (imc_counters_config[i][j].fd == -1) {
141 		fprintf(stderr, "Error opening leader %llx\n",
142 			imc_counters_config[i][j].pe.config);
143 
144 		return -1;
145 	}
146 
147 	return 0;
148 }
149 
150 /* Get type and config (read and write) of an iMC counter */
151 static int read_from_imc_dir(char *imc_dir, int count)
152 {
153 	char cas_count_cfg[1024], imc_counter_cfg[1024], imc_counter_type[1024];
154 	FILE *fp;
155 
156 	/* Get type of iMC counter */
157 	sprintf(imc_counter_type, "%s%s", imc_dir, "type");
158 	fp = fopen(imc_counter_type, "r");
159 	if (!fp) {
160 		ksft_perror("Failed to open iMC counter type file");
161 
162 		return -1;
163 	}
164 	if (fscanf(fp, "%u", &imc_counters_config[count][READ].type) <= 0) {
165 		ksft_perror("Could not get iMC type");
166 		fclose(fp);
167 
168 		return -1;
169 	}
170 	fclose(fp);
171 
172 	imc_counters_config[count][WRITE].type =
173 				imc_counters_config[count][READ].type;
174 
175 	/* Get read config */
176 	sprintf(imc_counter_cfg, "%s%s", imc_dir, READ_FILE_NAME);
177 	fp = fopen(imc_counter_cfg, "r");
178 	if (!fp) {
179 		ksft_perror("Failed to open iMC config file");
180 
181 		return -1;
182 	}
183 	if (fscanf(fp, "%s", cas_count_cfg) <= 0) {
184 		ksft_perror("Could not get iMC cas count read");
185 		fclose(fp);
186 
187 		return -1;
188 	}
189 	fclose(fp);
190 
191 	get_event_and_umask(cas_count_cfg, count, READ);
192 
193 	/* Get write config */
194 	sprintf(imc_counter_cfg, "%s%s", imc_dir, WRITE_FILE_NAME);
195 	fp = fopen(imc_counter_cfg, "r");
196 	if (!fp) {
197 		ksft_perror("Failed to open iMC config file");
198 
199 		return -1;
200 	}
201 	if  (fscanf(fp, "%s", cas_count_cfg) <= 0) {
202 		ksft_perror("Could not get iMC cas count write");
203 		fclose(fp);
204 
205 		return -1;
206 	}
207 	fclose(fp);
208 
209 	get_event_and_umask(cas_count_cfg, count, WRITE);
210 
211 	return 0;
212 }
213 
214 /*
215  * A system can have 'n' number of iMC (Integrated Memory Controller)
216  * counters, get that 'n'. For each iMC counter get it's type and config.
217  * Also, each counter has two configs, one for read and the other for write.
218  * A config again has two parts, event and umask.
219  * Enumerate all these details into an array of structures.
220  *
221  * Return: >= 0 on success. < 0 on failure.
222  */
223 static int num_of_imcs(void)
224 {
225 	char imc_dir[512], *temp;
226 	unsigned int count = 0;
227 	struct dirent *ep;
228 	int ret;
229 	DIR *dp;
230 
231 	dp = opendir(DYN_PMU_PATH);
232 	if (dp) {
233 		while ((ep = readdir(dp))) {
234 			temp = strstr(ep->d_name, UNCORE_IMC);
235 			if (!temp)
236 				continue;
237 
238 			/*
239 			 * imc counters are named as "uncore_imc_<n>", hence
240 			 * increment the pointer to point to <n>. Note that
241 			 * sizeof(UNCORE_IMC) would count for null character as
242 			 * well and hence the last underscore character in
243 			 * uncore_imc'_' need not be counted.
244 			 */
245 			temp = temp + sizeof(UNCORE_IMC);
246 
247 			/*
248 			 * Some directories under "DYN_PMU_PATH" could have
249 			 * names like "uncore_imc_free_running", hence, check if
250 			 * first character is a numerical digit or not.
251 			 */
252 			if (temp[0] >= '0' && temp[0] <= '9') {
253 				sprintf(imc_dir, "%s/%s/", DYN_PMU_PATH,
254 					ep->d_name);
255 				ret = read_from_imc_dir(imc_dir, count);
256 				if (ret) {
257 					closedir(dp);
258 
259 					return ret;
260 				}
261 				count++;
262 			}
263 		}
264 		closedir(dp);
265 		if (count == 0) {
266 			ksft_print_msg("Unable to find iMC counters\n");
267 
268 			return -1;
269 		}
270 	} else {
271 		ksft_perror("Unable to open PMU directory");
272 
273 		return -1;
274 	}
275 
276 	return count;
277 }
278 
279 static int initialize_mem_bw_imc(void)
280 {
281 	int imc, j;
282 
283 	imcs = num_of_imcs();
284 	if (imcs <= 0)
285 		return imcs;
286 
287 	/* Initialize perf_event_attr structures for all iMC's */
288 	for (imc = 0; imc < imcs; imc++) {
289 		for (j = 0; j < 2; j++)
290 			membw_initialize_perf_event_attr(imc, j);
291 	}
292 
293 	return 0;
294 }
295 
296 /*
297  * get_mem_bw_imc:	Memory band width as reported by iMC counters
298  * @cpu_no:		CPU number that the benchmark PID is binded to
299  * @bw_report:		Bandwidth report type (reads, writes)
300  *
301  * Memory B/W utilized by a process on a socket can be calculated using
302  * iMC counters. Perf events are used to read these counters.
303  *
304  * Return: = 0 on success. < 0 on failure.
305  */
306 static int get_mem_bw_imc(int cpu_no, char *bw_report, float *bw_imc)
307 {
308 	float reads, writes, of_mul_read, of_mul_write;
309 	int imc, j, ret;
310 
311 	/* Start all iMC counters to log values (both read and write) */
312 	reads = 0, writes = 0, of_mul_read = 1, of_mul_write = 1;
313 	for (imc = 0; imc < imcs; imc++) {
314 		for (j = 0; j < 2; j++) {
315 			ret = open_perf_event(imc, cpu_no, j);
316 			if (ret)
317 				return -1;
318 		}
319 		for (j = 0; j < 2; j++)
320 			membw_ioctl_perf_event_ioc_reset_enable(imc, j);
321 	}
322 
323 	sleep(1);
324 
325 	/* Stop counters after a second to get results (both read and write) */
326 	for (imc = 0; imc < imcs; imc++) {
327 		for (j = 0; j < 2; j++)
328 			membw_ioctl_perf_event_ioc_disable(imc, j);
329 	}
330 
331 	/*
332 	 * Get results which are stored in struct type imc_counter_config
333 	 * Take over flow into consideration before calculating total b/w
334 	 */
335 	for (imc = 0; imc < imcs; imc++) {
336 		struct imc_counter_config *r =
337 			&imc_counters_config[imc][READ];
338 		struct imc_counter_config *w =
339 			&imc_counters_config[imc][WRITE];
340 
341 		if (read(r->fd, &r->return_value,
342 			 sizeof(struct membw_read_format)) == -1) {
343 			ksft_perror("Couldn't get read b/w through iMC");
344 
345 			return -1;
346 		}
347 
348 		if (read(w->fd, &w->return_value,
349 			 sizeof(struct membw_read_format)) == -1) {
350 			ksft_perror("Couldn't get write bw through iMC");
351 
352 			return -1;
353 		}
354 
355 		__u64 r_time_enabled = r->return_value.time_enabled;
356 		__u64 r_time_running = r->return_value.time_running;
357 
358 		if (r_time_enabled != r_time_running)
359 			of_mul_read = (float)r_time_enabled /
360 					(float)r_time_running;
361 
362 		__u64 w_time_enabled = w->return_value.time_enabled;
363 		__u64 w_time_running = w->return_value.time_running;
364 
365 		if (w_time_enabled != w_time_running)
366 			of_mul_write = (float)w_time_enabled /
367 					(float)w_time_running;
368 		reads += r->return_value.value * of_mul_read * SCALE;
369 		writes += w->return_value.value * of_mul_write * SCALE;
370 	}
371 
372 	for (imc = 0; imc < imcs; imc++) {
373 		close(imc_counters_config[imc][READ].fd);
374 		close(imc_counters_config[imc][WRITE].fd);
375 	}
376 
377 	if (strcmp(bw_report, "reads") == 0) {
378 		*bw_imc = reads;
379 		return 0;
380 	}
381 
382 	if (strcmp(bw_report, "writes") == 0) {
383 		*bw_imc = writes;
384 		return 0;
385 	}
386 
387 	*bw_imc = reads + writes;
388 	return 0;
389 }
390 
391 void set_mbm_path(const char *ctrlgrp, const char *mongrp, int domain_id)
392 {
393 	if (ctrlgrp && mongrp)
394 		sprintf(mbm_total_path, CON_MON_MBM_LOCAL_BYTES_PATH,
395 			RESCTRL_PATH, ctrlgrp, mongrp, domain_id);
396 	else if (!ctrlgrp && mongrp)
397 		sprintf(mbm_total_path, MON_MBM_LOCAL_BYTES_PATH, RESCTRL_PATH,
398 			mongrp, domain_id);
399 	else if (ctrlgrp && !mongrp)
400 		sprintf(mbm_total_path, CON_MBM_LOCAL_BYTES_PATH, RESCTRL_PATH,
401 			ctrlgrp, domain_id);
402 	else if (!ctrlgrp && !mongrp)
403 		sprintf(mbm_total_path, MBM_LOCAL_BYTES_PATH, RESCTRL_PATH,
404 			domain_id);
405 }
406 
407 /*
408  * initialize_mem_bw_resctrl:	Appropriately populate "mbm_total_path"
409  * @ctrlgrp:			Name of the control monitor group (con_mon grp)
410  * @mongrp:			Name of the monitor group (mon grp)
411  * @cpu_no:			CPU number that the benchmark PID is binded to
412  * @resctrl_val:		Resctrl feature (Eg: mbm, mba.. etc)
413  */
414 static void initialize_mem_bw_resctrl(const char *ctrlgrp, const char *mongrp,
415 				      int cpu_no, char *resctrl_val)
416 {
417 	int domain_id;
418 
419 	if (get_domain_id("MB", cpu_no, &domain_id) < 0) {
420 		ksft_print_msg("Could not get domain ID\n");
421 		return;
422 	}
423 
424 	if (!strncmp(resctrl_val, MBM_STR, sizeof(MBM_STR)))
425 		set_mbm_path(ctrlgrp, mongrp, domain_id);
426 
427 	if (!strncmp(resctrl_val, MBA_STR, sizeof(MBA_STR))) {
428 		if (ctrlgrp)
429 			sprintf(mbm_total_path, CON_MBM_LOCAL_BYTES_PATH,
430 				RESCTRL_PATH, ctrlgrp, domain_id);
431 		else
432 			sprintf(mbm_total_path, MBM_LOCAL_BYTES_PATH,
433 				RESCTRL_PATH, domain_id);
434 	}
435 }
436 
437 /*
438  * Get MBM Local bytes as reported by resctrl FS
439  * For MBM,
440  * 1. If con_mon grp and mon grp are given, then read from con_mon grp's mon grp
441  * 2. If only con_mon grp is given, then read from con_mon grp
442  * 3. If both are not given, then read from root con_mon grp
443  * For MBA,
444  * 1. If con_mon grp is given, then read from it
445  * 2. If con_mon grp is not given, then read from root con_mon grp
446  */
447 static int get_mem_bw_resctrl(unsigned long *mbm_total)
448 {
449 	FILE *fp;
450 
451 	fp = fopen(mbm_total_path, "r");
452 	if (!fp) {
453 		ksft_perror("Failed to open total bw file");
454 
455 		return -1;
456 	}
457 	if (fscanf(fp, "%lu", mbm_total) <= 0) {
458 		ksft_perror("Could not get mbm local bytes");
459 		fclose(fp);
460 
461 		return -1;
462 	}
463 	fclose(fp);
464 
465 	return 0;
466 }
467 
468 pid_t bm_pid, ppid;
469 
470 void ctrlc_handler(int signum, siginfo_t *info, void *ptr)
471 {
472 	/* Only kill child after bm_pid is set after fork() */
473 	if (bm_pid)
474 		kill(bm_pid, SIGKILL);
475 	umount_resctrlfs();
476 	if (current_test && current_test->cleanup)
477 		current_test->cleanup();
478 	ksft_print_msg("Ending\n\n");
479 
480 	exit(EXIT_SUCCESS);
481 }
482 
483 /*
484  * Register CTRL-C handler for parent, as it has to kill
485  * child process before exiting.
486  */
487 int signal_handler_register(const struct resctrl_test *test)
488 {
489 	struct sigaction sigact = {};
490 	int ret = 0;
491 
492 	bm_pid = 0;
493 
494 	current_test = test;
495 	sigact.sa_sigaction = ctrlc_handler;
496 	sigemptyset(&sigact.sa_mask);
497 	sigact.sa_flags = SA_SIGINFO;
498 	if (sigaction(SIGINT, &sigact, NULL) ||
499 	    sigaction(SIGTERM, &sigact, NULL) ||
500 	    sigaction(SIGHUP, &sigact, NULL)) {
501 		ksft_perror("sigaction");
502 		ret = -1;
503 	}
504 	return ret;
505 }
506 
507 /*
508  * Reset signal handler to SIG_DFL.
509  * Non-Value return because the caller should keep
510  * the error code of other path even if sigaction fails.
511  */
512 void signal_handler_unregister(void)
513 {
514 	struct sigaction sigact = {};
515 
516 	current_test = NULL;
517 	sigact.sa_handler = SIG_DFL;
518 	sigemptyset(&sigact.sa_mask);
519 	if (sigaction(SIGINT, &sigact, NULL) ||
520 	    sigaction(SIGTERM, &sigact, NULL) ||
521 	    sigaction(SIGHUP, &sigact, NULL)) {
522 		ksft_perror("sigaction");
523 	}
524 }
525 
526 /*
527  * print_results_bw:	the memory bandwidth results are stored in a file
528  * @filename:		file that stores the results
529  * @bm_pid:		child pid that runs benchmark
530  * @bw_imc:		perf imc counter value
531  * @bw_resc:		memory bandwidth value
532  *
533  * Return:		0 on success, < 0 on error.
534  */
535 static int print_results_bw(char *filename,  int bm_pid, float bw_imc,
536 			    unsigned long bw_resc)
537 {
538 	unsigned long diff = fabs(bw_imc - bw_resc);
539 	FILE *fp;
540 
541 	if (strcmp(filename, "stdio") == 0 || strcmp(filename, "stderr") == 0) {
542 		printf("Pid: %d \t Mem_BW_iMC: %f \t ", bm_pid, bw_imc);
543 		printf("Mem_BW_resc: %lu \t Difference: %lu\n", bw_resc, diff);
544 	} else {
545 		fp = fopen(filename, "a");
546 		if (!fp) {
547 			ksft_perror("Cannot open results file");
548 
549 			return -1;
550 		}
551 		if (fprintf(fp, "Pid: %d \t Mem_BW_iMC: %f \t Mem_BW_resc: %lu \t Difference: %lu\n",
552 			    bm_pid, bw_imc, bw_resc, diff) <= 0) {
553 			ksft_print_msg("Could not log results\n");
554 			fclose(fp);
555 
556 			return -1;
557 		}
558 		fclose(fp);
559 	}
560 
561 	return 0;
562 }
563 
564 static void set_cmt_path(const char *ctrlgrp, const char *mongrp, char sock_num)
565 {
566 	if (strlen(ctrlgrp) && strlen(mongrp))
567 		sprintf(llc_occup_path,	CON_MON_LCC_OCCUP_PATH,	RESCTRL_PATH,
568 			ctrlgrp, mongrp, sock_num);
569 	else if (!strlen(ctrlgrp) && strlen(mongrp))
570 		sprintf(llc_occup_path,	MON_LCC_OCCUP_PATH, RESCTRL_PATH,
571 			mongrp, sock_num);
572 	else if (strlen(ctrlgrp) && !strlen(mongrp))
573 		sprintf(llc_occup_path,	CON_LCC_OCCUP_PATH, RESCTRL_PATH,
574 			ctrlgrp, sock_num);
575 	else if (!strlen(ctrlgrp) && !strlen(mongrp))
576 		sprintf(llc_occup_path, LCC_OCCUP_PATH,	RESCTRL_PATH, sock_num);
577 }
578 
579 /*
580  * initialize_llc_occu_resctrl:	Appropriately populate "llc_occup_path"
581  * @ctrlgrp:			Name of the control monitor group (con_mon grp)
582  * @mongrp:			Name of the monitor group (mon grp)
583  * @cpu_no:			CPU number that the benchmark PID is binded to
584  * @resctrl_val:		Resctrl feature (Eg: cat, cmt.. etc)
585  */
586 static void initialize_llc_occu_resctrl(const char *ctrlgrp, const char *mongrp,
587 					int cpu_no, char *resctrl_val)
588 {
589 	int domain_id;
590 
591 	if (get_domain_id("L3", cpu_no, &domain_id) < 0) {
592 		ksft_print_msg("Could not get domain ID\n");
593 		return;
594 	}
595 
596 	if (!strncmp(resctrl_val, CMT_STR, sizeof(CMT_STR)))
597 		set_cmt_path(ctrlgrp, mongrp, domain_id);
598 }
599 
600 static int measure_vals(const struct user_params *uparams,
601 			struct resctrl_val_param *param,
602 			unsigned long *bw_resc_start)
603 {
604 	unsigned long bw_resc, bw_resc_end;
605 	float bw_imc;
606 	int ret;
607 
608 	/*
609 	 * Measure memory bandwidth from resctrl and from
610 	 * another source which is perf imc value or could
611 	 * be something else if perf imc event is not available.
612 	 * Compare the two values to validate resctrl value.
613 	 * It takes 1sec to measure the data.
614 	 */
615 	ret = get_mem_bw_imc(uparams->cpu, param->bw_report, &bw_imc);
616 	if (ret < 0)
617 		return ret;
618 
619 	ret = get_mem_bw_resctrl(&bw_resc_end);
620 	if (ret < 0)
621 		return ret;
622 
623 	bw_resc = (bw_resc_end - *bw_resc_start) / MB;
624 	ret = print_results_bw(param->filename, bm_pid, bw_imc, bw_resc);
625 	if (ret)
626 		return ret;
627 
628 	*bw_resc_start = bw_resc_end;
629 
630 	return 0;
631 }
632 
633 /*
634  * run_benchmark - Run a specified benchmark or fill_buf (default benchmark)
635  *		   in specified signal. Direct benchmark stdio to /dev/null.
636  * @signum:	signal number
637  * @info:	signal info
638  * @ucontext:	user context in signal handling
639  */
640 static void run_benchmark(int signum, siginfo_t *info, void *ucontext)
641 {
642 	int operation, ret, memflush;
643 	char **benchmark_cmd;
644 	size_t span;
645 	bool once;
646 	FILE *fp;
647 
648 	benchmark_cmd = info->si_ptr;
649 
650 	/*
651 	 * Direct stdio of child to /dev/null, so that only parent writes to
652 	 * stdio (console)
653 	 */
654 	fp = freopen("/dev/null", "w", stdout);
655 	if (!fp) {
656 		ksft_perror("Unable to direct benchmark status to /dev/null");
657 		PARENT_EXIT();
658 	}
659 
660 	if (strcmp(benchmark_cmd[0], "fill_buf") == 0) {
661 		/* Execute default fill_buf benchmark */
662 		span = strtoul(benchmark_cmd[1], NULL, 10);
663 		memflush =  atoi(benchmark_cmd[2]);
664 		operation = atoi(benchmark_cmd[3]);
665 		if (!strcmp(benchmark_cmd[4], "true")) {
666 			once = true;
667 		} else if (!strcmp(benchmark_cmd[4], "false")) {
668 			once = false;
669 		} else {
670 			ksft_print_msg("Invalid once parameter\n");
671 			PARENT_EXIT();
672 		}
673 
674 		if (run_fill_buf(span, memflush, operation, once))
675 			fprintf(stderr, "Error in running fill buffer\n");
676 	} else {
677 		/* Execute specified benchmark */
678 		ret = execvp(benchmark_cmd[0], benchmark_cmd);
679 		if (ret)
680 			ksft_perror("execvp");
681 	}
682 
683 	fclose(stdout);
684 	ksft_print_msg("Unable to run specified benchmark\n");
685 	PARENT_EXIT();
686 }
687 
688 /*
689  * resctrl_val:	execute benchmark and measure memory bandwidth on
690  *			the benchmark
691  * @test:		test information structure
692  * @uparams:		user supplied parameters
693  * @benchmark_cmd:	benchmark command and its arguments
694  * @param:		parameters passed to resctrl_val()
695  *
696  * Return:		0 when the test was run, < 0 on error.
697  */
698 int resctrl_val(const struct resctrl_test *test,
699 		const struct user_params *uparams,
700 		const char * const *benchmark_cmd,
701 		struct resctrl_val_param *param)
702 {
703 	char *resctrl_val = param->resctrl_val;
704 	unsigned long bw_resc_start = 0;
705 	struct sigaction sigact;
706 	int ret = 0, pipefd[2];
707 	char pipe_message = 0;
708 	union sigval value;
709 
710 	if (strcmp(param->filename, "") == 0)
711 		sprintf(param->filename, "stdio");
712 
713 	if (!strncmp(resctrl_val, MBA_STR, sizeof(MBA_STR)) ||
714 	    !strncmp(resctrl_val, MBM_STR, sizeof(MBM_STR))) {
715 		ret = validate_bw_report_request(param->bw_report);
716 		if (ret)
717 			return ret;
718 	}
719 
720 	/*
721 	 * If benchmark wasn't successfully started by child, then child should
722 	 * kill parent, so save parent's pid
723 	 */
724 	ppid = getpid();
725 
726 	if (pipe(pipefd)) {
727 		ksft_perror("Unable to create pipe");
728 
729 		return -1;
730 	}
731 
732 	/*
733 	 * Fork to start benchmark, save child's pid so that it can be killed
734 	 * when needed
735 	 */
736 	fflush(stdout);
737 	bm_pid = fork();
738 	if (bm_pid == -1) {
739 		ksft_perror("Unable to fork");
740 
741 		return -1;
742 	}
743 
744 	if (bm_pid == 0) {
745 		/*
746 		 * Mask all signals except SIGUSR1, parent uses SIGUSR1 to
747 		 * start benchmark
748 		 */
749 		sigfillset(&sigact.sa_mask);
750 		sigdelset(&sigact.sa_mask, SIGUSR1);
751 
752 		sigact.sa_sigaction = run_benchmark;
753 		sigact.sa_flags = SA_SIGINFO;
754 
755 		/* Register for "SIGUSR1" signal from parent */
756 		if (sigaction(SIGUSR1, &sigact, NULL)) {
757 			ksft_perror("Can't register child for signal");
758 			PARENT_EXIT();
759 		}
760 
761 		/* Tell parent that child is ready */
762 		close(pipefd[0]);
763 		pipe_message = 1;
764 		if (write(pipefd[1], &pipe_message, sizeof(pipe_message)) <
765 		    sizeof(pipe_message)) {
766 			ksft_perror("Failed signaling parent process");
767 			close(pipefd[1]);
768 			return -1;
769 		}
770 		close(pipefd[1]);
771 
772 		/* Suspend child until delivery of "SIGUSR1" from parent */
773 		sigsuspend(&sigact.sa_mask);
774 
775 		ksft_perror("Child is done");
776 		PARENT_EXIT();
777 	}
778 
779 	ksft_print_msg("Benchmark PID: %d\n", bm_pid);
780 
781 	/*
782 	 * The cast removes constness but nothing mutates benchmark_cmd within
783 	 * the context of this process. At the receiving process, it becomes
784 	 * argv, which is mutable, on exec() but that's after fork() so it
785 	 * doesn't matter for the process running the tests.
786 	 */
787 	value.sival_ptr = (void *)benchmark_cmd;
788 
789 	/* Taskset benchmark to specified cpu */
790 	ret = taskset_benchmark(bm_pid, uparams->cpu, NULL);
791 	if (ret)
792 		goto out;
793 
794 	/* Write benchmark to specified control&monitoring grp in resctrl FS */
795 	ret = write_bm_pid_to_resctrl(bm_pid, param->ctrlgrp, param->mongrp,
796 				      resctrl_val);
797 	if (ret)
798 		goto out;
799 
800 	if (!strncmp(resctrl_val, MBM_STR, sizeof(MBM_STR)) ||
801 	    !strncmp(resctrl_val, MBA_STR, sizeof(MBA_STR))) {
802 		ret = initialize_mem_bw_imc();
803 		if (ret)
804 			goto out;
805 
806 		initialize_mem_bw_resctrl(param->ctrlgrp, param->mongrp,
807 					  uparams->cpu, resctrl_val);
808 	} else if (!strncmp(resctrl_val, CMT_STR, sizeof(CMT_STR)))
809 		initialize_llc_occu_resctrl(param->ctrlgrp, param->mongrp,
810 					    uparams->cpu, resctrl_val);
811 
812 	/* Parent waits for child to be ready. */
813 	close(pipefd[1]);
814 	while (pipe_message != 1) {
815 		if (read(pipefd[0], &pipe_message, sizeof(pipe_message)) <
816 		    sizeof(pipe_message)) {
817 			ksft_perror("Failed reading message from child process");
818 			close(pipefd[0]);
819 			goto out;
820 		}
821 	}
822 	close(pipefd[0]);
823 
824 	/* Signal child to start benchmark */
825 	if (sigqueue(bm_pid, SIGUSR1, value) == -1) {
826 		ksft_perror("sigqueue SIGUSR1 to child");
827 		ret = -1;
828 		goto out;
829 	}
830 
831 	/* Give benchmark enough time to fully run */
832 	sleep(1);
833 
834 	/* Test runs until the callback setup() tells the test to stop. */
835 	while (1) {
836 		ret = param->setup(test, uparams, param);
837 		if (ret == END_OF_TESTS) {
838 			ret = 0;
839 			break;
840 		}
841 		if (ret < 0)
842 			break;
843 
844 		if (!strncmp(resctrl_val, MBM_STR, sizeof(MBM_STR)) ||
845 		    !strncmp(resctrl_val, MBA_STR, sizeof(MBA_STR))) {
846 			ret = measure_vals(uparams, param, &bw_resc_start);
847 			if (ret)
848 				break;
849 		} else if (!strncmp(resctrl_val, CMT_STR, sizeof(CMT_STR))) {
850 			sleep(1);
851 			ret = measure_llc_resctrl(param->filename, bm_pid);
852 			if (ret)
853 				break;
854 		}
855 	}
856 
857 out:
858 	kill(bm_pid, SIGKILL);
859 
860 	return ret;
861 }
862