brintos

brintos / linux-shallow public Read only

0
0
Text · 16.1 KiB · 2709421 Raw
743 lines · c
1// SPDX-License-Identifier: GPL-2.02#include <api/fs/fs.h>3#include "cpumap.h"4#include "debug.h"5#include "event.h"6#include <assert.h>7#include <dirent.h>8#include <stdio.h>9#include <stdlib.h>10#include <linux/bitmap.h>11#include "asm/bug.h"12 13#include <linux/ctype.h>14#include <linux/zalloc.h>15#include <internal/cpumap.h>16 17static struct perf_cpu max_cpu_num;18static struct perf_cpu max_present_cpu_num;19static int max_node_num;20/**21 * The numa node X as read from /sys/devices/system/node/nodeX indexed by the22 * CPU number.23 */24static int *cpunode_map;25 26bool perf_record_cpu_map_data__test_bit(int i,27					const struct perf_record_cpu_map_data *data)28{29	int bit_word32 = i / 32;30	__u32 bit_mask32 = 1U << (i & 31);31	int bit_word64 = i / 64;32	__u64 bit_mask64 = ((__u64)1) << (i & 63);33 34	return (data->mask32_data.long_size == 4)35		? (bit_word32 < data->mask32_data.nr) &&36		(data->mask32_data.mask[bit_word32] & bit_mask32) != 037		: (bit_word64 < data->mask64_data.nr) &&38		(data->mask64_data.mask[bit_word64] & bit_mask64) != 0;39}40 41/* Read ith mask value from data into the given 64-bit sized bitmap */42static void perf_record_cpu_map_data__read_one_mask(const struct perf_record_cpu_map_data *data,43						    int i, unsigned long *bitmap)44{45#if __SIZEOF_LONG__ == 846	if (data->mask32_data.long_size == 4)47		bitmap[0] = data->mask32_data.mask[i];48	else49		bitmap[0] = data->mask64_data.mask[i];50#else51	if (data->mask32_data.long_size == 4) {52		bitmap[0] = data->mask32_data.mask[i];53		bitmap[1] = 0;54	} else {55#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__56		bitmap[0] = (unsigned long)(data->mask64_data.mask[i] >> 32);57		bitmap[1] = (unsigned long)data->mask64_data.mask[i];58#else59		bitmap[0] = (unsigned long)data->mask64_data.mask[i];60		bitmap[1] = (unsigned long)(data->mask64_data.mask[i] >> 32);61#endif62	}63#endif64}65static struct perf_cpu_map *cpu_map__from_entries(const struct perf_record_cpu_map_data *data)66{67	struct perf_cpu_map *map;68 69	map = perf_cpu_map__empty_new(data->cpus_data.nr);70	if (map) {71		unsigned i;72 73		for (i = 0; i < data->cpus_data.nr; i++) {74			/*75			 * Special treatment for -1, which is not real cpu number,76			 * and we need to use (int) -1 to initialize map[i],77			 * otherwise it would become 65535.78			 */79			if (data->cpus_data.cpu[i] == (u16) -1)80				RC_CHK_ACCESS(map)->map[i].cpu = -1;81			else82				RC_CHK_ACCESS(map)->map[i].cpu = (int) data->cpus_data.cpu[i];83		}84	}85 86	return map;87}88 89static struct perf_cpu_map *cpu_map__from_mask(const struct perf_record_cpu_map_data *data)90{91	DECLARE_BITMAP(local_copy, 64);92	int weight = 0, mask_nr = data->mask32_data.nr;93	struct perf_cpu_map *map;94 95	for (int i = 0; i < mask_nr; i++) {96		perf_record_cpu_map_data__read_one_mask(data, i, local_copy);97		weight += bitmap_weight(local_copy, 64);98	}99 100	map = perf_cpu_map__empty_new(weight);101	if (!map)102		return NULL;103 104	for (int i = 0, j = 0; i < mask_nr; i++) {105		int cpus_per_i = (i * data->mask32_data.long_size  * BITS_PER_BYTE);106		int cpu;107 108		perf_record_cpu_map_data__read_one_mask(data, i, local_copy);109		for_each_set_bit(cpu, local_copy, 64)110			RC_CHK_ACCESS(map)->map[j++].cpu = cpu + cpus_per_i;111	}112	return map;113 114}115 116static struct perf_cpu_map *cpu_map__from_range(const struct perf_record_cpu_map_data *data)117{118	struct perf_cpu_map *map;119	unsigned int i = 0;120 121	map = perf_cpu_map__empty_new(data->range_cpu_data.end_cpu -122				data->range_cpu_data.start_cpu + 1 + data->range_cpu_data.any_cpu);123	if (!map)124		return NULL;125 126	if (data->range_cpu_data.any_cpu)127		RC_CHK_ACCESS(map)->map[i++].cpu = -1;128 129	for (int cpu = data->range_cpu_data.start_cpu; cpu <= data->range_cpu_data.end_cpu;130	     i++, cpu++)131		RC_CHK_ACCESS(map)->map[i].cpu = cpu;132 133	return map;134}135 136struct perf_cpu_map *cpu_map__new_data(const struct perf_record_cpu_map_data *data)137{138	switch (data->type) {139	case PERF_CPU_MAP__CPUS:140		return cpu_map__from_entries(data);141	case PERF_CPU_MAP__MASK:142		return cpu_map__from_mask(data);143	case PERF_CPU_MAP__RANGE_CPUS:144		return cpu_map__from_range(data);145	default:146		pr_err("cpu_map__new_data unknown type %d\n", data->type);147		return NULL;148	}149}150 151size_t cpu_map__fprintf(struct perf_cpu_map *map, FILE *fp)152{153#define BUFSIZE 1024154	char buf[BUFSIZE];155 156	cpu_map__snprint(map, buf, sizeof(buf));157	return fprintf(fp, "%s\n", buf);158#undef BUFSIZE159}160 161struct perf_cpu_map *perf_cpu_map__empty_new(int nr)162{163	struct perf_cpu_map *cpus = perf_cpu_map__alloc(nr);164 165	if (cpus != NULL) {166		for (int i = 0; i < nr; i++)167			RC_CHK_ACCESS(cpus)->map[i].cpu = -1;168	}169 170	return cpus;171}172 173struct cpu_aggr_map *cpu_aggr_map__empty_new(int nr)174{175	struct cpu_aggr_map *cpus = malloc(sizeof(*cpus) + sizeof(struct aggr_cpu_id) * nr);176 177	if (cpus != NULL) {178		int i;179 180		cpus->nr = nr;181		for (i = 0; i < nr; i++)182			cpus->map[i] = aggr_cpu_id__empty();183	}184 185	return cpus;186}187 188static int cpu__get_topology_int(int cpu, const char *name, int *value)189{190	char path[PATH_MAX];191 192	snprintf(path, PATH_MAX,193		"devices/system/cpu/cpu%d/topology/%s", cpu, name);194 195	return sysfs__read_int(path, value);196}197 198int cpu__get_socket_id(struct perf_cpu cpu)199{200	int value, ret = cpu__get_topology_int(cpu.cpu, "physical_package_id", &value);201	return ret ?: value;202}203 204struct aggr_cpu_id aggr_cpu_id__socket(struct perf_cpu cpu, void *data __maybe_unused)205{206	struct aggr_cpu_id id = aggr_cpu_id__empty();207 208	id.socket = cpu__get_socket_id(cpu);209	return id;210}211 212static int aggr_cpu_id__cmp(const void *a_pointer, const void *b_pointer)213{214	struct aggr_cpu_id *a = (struct aggr_cpu_id *)a_pointer;215	struct aggr_cpu_id *b = (struct aggr_cpu_id *)b_pointer;216 217	if (a->node != b->node)218		return a->node - b->node;219	else if (a->socket != b->socket)220		return a->socket - b->socket;221	else if (a->die != b->die)222		return a->die - b->die;223	else if (a->cluster != b->cluster)224		return a->cluster - b->cluster;225	else if (a->cache_lvl != b->cache_lvl)226		return a->cache_lvl - b->cache_lvl;227	else if (a->cache != b->cache)228		return a->cache - b->cache;229	else if (a->core != b->core)230		return a->core - b->core;231	else232		return a->thread_idx - b->thread_idx;233}234 235struct cpu_aggr_map *cpu_aggr_map__new(const struct perf_cpu_map *cpus,236				       aggr_cpu_id_get_t get_id,237				       void *data, bool needs_sort)238{239	int idx;240	struct perf_cpu cpu;241	struct cpu_aggr_map *c = cpu_aggr_map__empty_new(perf_cpu_map__nr(cpus));242 243	if (!c)244		return NULL;245 246	/* Reset size as it may only be partially filled */247	c->nr = 0;248 249	perf_cpu_map__for_each_cpu(cpu, idx, cpus) {250		bool duplicate = false;251		struct aggr_cpu_id cpu_id = get_id(cpu, data);252 253		for (int j = 0; j < c->nr; j++) {254			if (aggr_cpu_id__equal(&cpu_id, &c->map[j])) {255				duplicate = true;256				break;257			}258		}259		if (!duplicate) {260			c->map[c->nr] = cpu_id;261			c->nr++;262		}263	}264	/* Trim. */265	if (c->nr != perf_cpu_map__nr(cpus)) {266		struct cpu_aggr_map *trimmed_c =267			realloc(c,268				sizeof(struct cpu_aggr_map) + sizeof(struct aggr_cpu_id) * c->nr);269 270		if (trimmed_c)271			c = trimmed_c;272	}273 274	/* ensure we process id in increasing order */275	if (needs_sort)276		qsort(c->map, c->nr, sizeof(struct aggr_cpu_id), aggr_cpu_id__cmp);277 278	return c;279 280}281 282int cpu__get_die_id(struct perf_cpu cpu)283{284	int value, ret = cpu__get_topology_int(cpu.cpu, "die_id", &value);285 286	return ret ?: value;287}288 289struct aggr_cpu_id aggr_cpu_id__die(struct perf_cpu cpu, void *data)290{291	struct aggr_cpu_id id;292	int die;293 294	die = cpu__get_die_id(cpu);295	/* There is no die_id on legacy system. */296	if (die == -1)297		die = 0;298 299	/*300	 * die_id is relative to socket, so start301	 * with the socket ID and then add die to302	 * make a unique ID.303	 */304	id = aggr_cpu_id__socket(cpu, data);305	if (aggr_cpu_id__is_empty(&id))306		return id;307 308	id.die = die;309	return id;310}311 312int cpu__get_cluster_id(struct perf_cpu cpu)313{314	int value, ret = cpu__get_topology_int(cpu.cpu, "cluster_id", &value);315 316	return ret ?: value;317}318 319struct aggr_cpu_id aggr_cpu_id__cluster(struct perf_cpu cpu, void *data)320{321	int cluster = cpu__get_cluster_id(cpu);322	struct aggr_cpu_id id;323 324	/* There is no cluster_id on legacy system. */325	if (cluster == -1)326		cluster = 0;327 328	id = aggr_cpu_id__die(cpu, data);329	if (aggr_cpu_id__is_empty(&id))330		return id;331 332	id.cluster = cluster;333	return id;334}335 336int cpu__get_core_id(struct perf_cpu cpu)337{338	int value, ret = cpu__get_topology_int(cpu.cpu, "core_id", &value);339	return ret ?: value;340}341 342struct aggr_cpu_id aggr_cpu_id__core(struct perf_cpu cpu, void *data)343{344	struct aggr_cpu_id id;345	int core = cpu__get_core_id(cpu);346 347	/* aggr_cpu_id__die returns a struct with socket die, and cluster set. */348	id = aggr_cpu_id__cluster(cpu, data);349	if (aggr_cpu_id__is_empty(&id))350		return id;351 352	/*353	 * core_id is relative to socket and die, we need a global id.354	 * So we combine the result from cpu_map__get_die with the core id355	 */356	id.core = core;357	return id;358 359}360 361struct aggr_cpu_id aggr_cpu_id__cpu(struct perf_cpu cpu, void *data)362{363	struct aggr_cpu_id id;364 365	/* aggr_cpu_id__core returns a struct with socket, die and core set. */366	id = aggr_cpu_id__core(cpu, data);367	if (aggr_cpu_id__is_empty(&id))368		return id;369 370	id.cpu = cpu;371	return id;372 373}374 375struct aggr_cpu_id aggr_cpu_id__node(struct perf_cpu cpu, void *data __maybe_unused)376{377	struct aggr_cpu_id id = aggr_cpu_id__empty();378 379	id.node = cpu__get_node(cpu);380	return id;381}382 383struct aggr_cpu_id aggr_cpu_id__global(struct perf_cpu cpu, void *data __maybe_unused)384{385	struct aggr_cpu_id id = aggr_cpu_id__empty();386 387	/* it always aggregates to the cpu 0 */388	cpu.cpu = 0;389	id.cpu = cpu;390	return id;391}392 393/* setup simple routines to easily access node numbers given a cpu number */394static int get_max_num(char *path, int *max)395{396	size_t num;397	char *buf;398	int err = 0;399 400	if (filename__read_str(path, &buf, &num))401		return -1;402 403	buf[num] = '\0';404 405	/* start on the right, to find highest node num */406	while (--num) {407		if ((buf[num] == ',') || (buf[num] == '-')) {408			num++;409			break;410		}411	}412	if (sscanf(&buf[num], "%d", max) < 1) {413		err = -1;414		goto out;415	}416 417	/* convert from 0-based to 1-based */418	(*max)++;419 420out:421	free(buf);422	return err;423}424 425/* Determine highest possible cpu in the system for sparse allocation */426static void set_max_cpu_num(void)427{428	const char *mnt;429	char path[PATH_MAX];430	int ret = -1;431 432	/* set up default */433	max_cpu_num.cpu = 4096;434	max_present_cpu_num.cpu = 4096;435 436	mnt = sysfs__mountpoint();437	if (!mnt)438		goto out;439 440	/* get the highest possible cpu number for a sparse allocation */441	ret = snprintf(path, PATH_MAX, "%s/devices/system/cpu/possible", mnt);442	if (ret >= PATH_MAX) {443		pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX);444		goto out;445	}446 447	ret = get_max_num(path, &max_cpu_num.cpu);448	if (ret)449		goto out;450 451	/* get the highest present cpu number for a sparse allocation */452	ret = snprintf(path, PATH_MAX, "%s/devices/system/cpu/present", mnt);453	if (ret >= PATH_MAX) {454		pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX);455		goto out;456	}457 458	ret = get_max_num(path, &max_present_cpu_num.cpu);459 460out:461	if (ret)462		pr_err("Failed to read max cpus, using default of %d\n", max_cpu_num.cpu);463}464 465/* Determine highest possible node in the system for sparse allocation */466static void set_max_node_num(void)467{468	const char *mnt;469	char path[PATH_MAX];470	int ret = -1;471 472	/* set up default */473	max_node_num = 8;474 475	mnt = sysfs__mountpoint();476	if (!mnt)477		goto out;478 479	/* get the highest possible cpu number for a sparse allocation */480	ret = snprintf(path, PATH_MAX, "%s/devices/system/node/possible", mnt);481	if (ret >= PATH_MAX) {482		pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX);483		goto out;484	}485 486	ret = get_max_num(path, &max_node_num);487 488out:489	if (ret)490		pr_err("Failed to read max nodes, using default of %d\n", max_node_num);491}492 493int cpu__max_node(void)494{495	if (unlikely(!max_node_num))496		set_max_node_num();497 498	return max_node_num;499}500 501struct perf_cpu cpu__max_cpu(void)502{503	if (unlikely(!max_cpu_num.cpu))504		set_max_cpu_num();505 506	return max_cpu_num;507}508 509struct perf_cpu cpu__max_present_cpu(void)510{511	if (unlikely(!max_present_cpu_num.cpu))512		set_max_cpu_num();513 514	return max_present_cpu_num;515}516 517 518int cpu__get_node(struct perf_cpu cpu)519{520	if (unlikely(cpunode_map == NULL)) {521		pr_debug("cpu_map not initialized\n");522		return -1;523	}524 525	return cpunode_map[cpu.cpu];526}527 528static int init_cpunode_map(void)529{530	int i;531 532	set_max_cpu_num();533	set_max_node_num();534 535	cpunode_map = calloc(max_cpu_num.cpu, sizeof(int));536	if (!cpunode_map) {537		pr_err("%s: calloc failed\n", __func__);538		return -1;539	}540 541	for (i = 0; i < max_cpu_num.cpu; i++)542		cpunode_map[i] = -1;543 544	return 0;545}546 547int cpu__setup_cpunode_map(void)548{549	struct dirent *dent1, *dent2;550	DIR *dir1, *dir2;551	unsigned int cpu, mem;552	char buf[PATH_MAX];553	char path[PATH_MAX];554	const char *mnt;555	int n;556 557	/* initialize globals */558	if (init_cpunode_map())559		return -1;560 561	mnt = sysfs__mountpoint();562	if (!mnt)563		return 0;564 565	n = snprintf(path, PATH_MAX, "%s/devices/system/node", mnt);566	if (n >= PATH_MAX) {567		pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX);568		return -1;569	}570 571	dir1 = opendir(path);572	if (!dir1)573		return 0;574 575	/* walk tree and setup map */576	while ((dent1 = readdir(dir1)) != NULL) {577		if (dent1->d_type != DT_DIR || sscanf(dent1->d_name, "node%u", &mem) < 1)578			continue;579 580		n = snprintf(buf, PATH_MAX, "%s/%s", path, dent1->d_name);581		if (n >= PATH_MAX) {582			pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX);583			continue;584		}585 586		dir2 = opendir(buf);587		if (!dir2)588			continue;589		while ((dent2 = readdir(dir2)) != NULL) {590			if (dent2->d_type != DT_LNK || sscanf(dent2->d_name, "cpu%u", &cpu) < 1)591				continue;592			cpunode_map[cpu] = mem;593		}594		closedir(dir2);595	}596	closedir(dir1);597	return 0;598}599 600size_t cpu_map__snprint(struct perf_cpu_map *map, char *buf, size_t size)601{602	int i, start = -1;603	bool first = true;604	size_t ret = 0;605 606#define COMMA first ? "" : ","607 608	for (i = 0; i < perf_cpu_map__nr(map) + 1; i++) {609		struct perf_cpu cpu = { .cpu = INT_MAX };610		bool last = i == perf_cpu_map__nr(map);611 612		if (!last)613			cpu = perf_cpu_map__cpu(map, i);614 615		if (start == -1) {616			start = i;617			if (last) {618				ret += snprintf(buf + ret, size - ret,619						"%s%d", COMMA,620						perf_cpu_map__cpu(map, i).cpu);621			}622		} else if (((i - start) != (cpu.cpu - perf_cpu_map__cpu(map, start).cpu)) || last) {623			int end = i - 1;624 625			if (start == end) {626				ret += snprintf(buf + ret, size - ret,627						"%s%d", COMMA,628						perf_cpu_map__cpu(map, start).cpu);629			} else {630				ret += snprintf(buf + ret, size - ret,631						"%s%d-%d", COMMA,632						perf_cpu_map__cpu(map, start).cpu, perf_cpu_map__cpu(map, end).cpu);633			}634			first = false;635			start = i;636		}637	}638 639#undef COMMA640 641	pr_debug2("cpumask list: %s\n", buf);642	return ret;643}644 645static char hex_char(unsigned char val)646{647	if (val < 10)648		return val + '0';649	if (val < 16)650		return val - 10 + 'a';651	return '?';652}653 654size_t cpu_map__snprint_mask(struct perf_cpu_map *map, char *buf, size_t size)655{656	int idx;657	char *ptr = buf;658	unsigned char *bitmap;659	struct perf_cpu c, last_cpu = perf_cpu_map__max(map);660 661	if (buf == NULL)662		return 0;663 664	bitmap = zalloc(last_cpu.cpu / 8 + 1);665	if (bitmap == NULL) {666		buf[0] = '\0';667		return 0;668	}669 670	perf_cpu_map__for_each_cpu(c, idx, map)671		bitmap[c.cpu / 8] |= 1 << (c.cpu % 8);672 673	for (int cpu = last_cpu.cpu / 4 * 4; cpu >= 0; cpu -= 4) {674		unsigned char bits = bitmap[cpu / 8];675 676		if (cpu % 8)677			bits >>= 4;678		else679			bits &= 0xf;680 681		*ptr++ = hex_char(bits);682		if ((cpu % 32) == 0 && cpu > 0)683			*ptr++ = ',';684	}685	*ptr = '\0';686	free(bitmap);687 688	buf[size - 1] = '\0';689	return ptr - buf;690}691 692struct perf_cpu_map *cpu_map__online(void) /* thread unsafe */693{694	static struct perf_cpu_map *online;695 696	if (!online)697		online = perf_cpu_map__new_online_cpus(); /* from /sys/devices/system/cpu/online */698 699	return online;700}701 702bool aggr_cpu_id__equal(const struct aggr_cpu_id *a, const struct aggr_cpu_id *b)703{704	return a->thread_idx == b->thread_idx &&705		a->node == b->node &&706		a->socket == b->socket &&707		a->die == b->die &&708		a->cluster == b->cluster &&709		a->cache_lvl == b->cache_lvl &&710		a->cache == b->cache &&711		a->core == b->core &&712		a->cpu.cpu == b->cpu.cpu;713}714 715bool aggr_cpu_id__is_empty(const struct aggr_cpu_id *a)716{717	return a->thread_idx == -1 &&718		a->node == -1 &&719		a->socket == -1 &&720		a->die == -1 &&721		a->cluster == -1 &&722		a->cache_lvl == -1 &&723		a->cache == -1 &&724		a->core == -1 &&725		a->cpu.cpu == -1;726}727 728struct aggr_cpu_id aggr_cpu_id__empty(void)729{730	struct aggr_cpu_id ret = {731		.thread_idx = -1,732		.node = -1,733		.socket = -1,734		.die = -1,735		.cluster = -1,736		.cache_lvl = -1,737		.cache = -1,738		.core = -1,739		.cpu = (struct perf_cpu){ .cpu = -1 },740	};741	return ret;742}743