1 | /* |
2 | * hardware.c |
3 | * |
4 | * Copyright (C) 2016-2017 Aerospike, Inc. |
5 | * |
6 | * Portions may be licensed to Aerospike, Inc. under one or more contributor |
7 | * license agreements. |
8 | * |
9 | * This program is free software: you can redistribute it and/or modify it under |
10 | * the terms of the GNU Affero General Public License as published by the Free |
11 | * Software Foundation, either version 3 of the License, or (at your option) any |
12 | * later version. |
13 | * |
14 | * This program is distributed in the hope that it will be useful, but WITHOUT |
15 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS |
16 | * FOR A PARTICULAR PURPOSE. See the GNU Affero General Public License for more |
17 | * details. |
18 | * |
19 | * You should have received a copy of the GNU Affero General Public License |
20 | * along with this program. If not, see http://www.gnu.org/licenses/ |
21 | */ |
22 | |
23 | #include "hardware.h" |
24 | |
25 | #include <ctype.h> |
26 | #include <dirent.h> |
27 | #include <errno.h> |
28 | #include <fcntl.h> |
29 | #include <inttypes.h> |
30 | #include <libgen.h> |
31 | #include <limits.h> |
32 | #include <mntent.h> |
33 | #include <regex.h> |
34 | #include <sched.h> |
35 | #include <stdbool.h> |
36 | #include <stddef.h> |
37 | #include <stdint.h> |
38 | #include <stdio.h> |
39 | #include <stdlib.h> |
40 | #include <string.h> |
41 | #include <syscall.h> |
42 | #include <unistd.h> |
43 | |
44 | #include <sys/ioctl.h> |
45 | #include <sys/socket.h> |
46 | #include <sys/stat.h> |
47 | #include <sys/statvfs.h> |
48 | #include <sys/sysmacros.h> |
49 | #include <sys/types.h> |
50 | #include <sys/vfs.h> |
51 | |
52 | #include <linux/capability.h> |
53 | #include <linux/ethtool.h> |
54 | #include <linux/if.h> |
55 | #include <linux/limits.h> |
56 | #include <linux/mempolicy.h> |
57 | #include <linux/sockios.h> |
58 | |
59 | #include "cf_mutex.h" |
60 | #include "daemon.h" |
61 | #include "fault.h" |
62 | #include "shash.h" |
63 | #include "socket.h" |
64 | |
65 | #include "citrusleaf/alloc.h" |
66 | #include "citrusleaf/cf_clock.h" |
67 | |
68 | #include "warnings.h" |
69 | |
70 | // Only available in Linux kernel version 3.19 and later; but we'd like to |
71 | // allow compilation with older kernel headers. |
72 | #if !defined SO_INCOMING_CPU |
73 | #define SO_INCOMING_CPU 49 |
74 | #endif |
75 | |
76 | // Only available in Linux kernel version 4.12 and later; but we'd like to |
77 | // allow compilation with older kernel headers. |
78 | #if !defined SO_INCOMING_NAPI_ID |
79 | #define SO_INCOMING_NAPI_ID 56 |
80 | #endif |
81 | |
82 | // The linux/nvme_ioctl.h kernel header came in Linux 4.4, but we'd like to |
83 | // allow compilation with older kernel headers. |
84 | // |
85 | // Also, we need to be prepared for this IOCTL to fail with EINVAL, when we |
86 | // run on older kernels that don't support it. |
87 | |
88 | #define NVME_IOCTL_ADMIN_CMD _IOWR('N', 0x41, struct nvme_admin_cmd) |
89 | #define NVME_SC_INVALID_LOG_PAGE 0x109 |
90 | |
91 | struct nvme_admin_cmd { |
92 | uint8_t opcode; |
93 | uint8_t flags; |
94 | uint16_t rsvd1; |
95 | uint32_t nsid; |
96 | uint32_t cdw2; |
97 | uint32_t cdw3; |
98 | uint64_t metadata; |
99 | uint64_t addr; |
100 | uint32_t metadata_len; |
101 | uint32_t data_len; |
102 | uint32_t cdw10; |
103 | uint32_t cdw11; |
104 | uint32_t cdw12; |
105 | uint32_t cdw13; |
106 | uint32_t cdw14; |
107 | uint32_t cdw15; |
108 | uint32_t timeout_ms; |
109 | uint32_t result; |
110 | }; |
111 | |
112 | #define INVALID_INDEX ((uint16_t)-1) |
113 | #define POLICY_SCRIPT "/etc/aerospike/irqbalance-ban.sh" |
114 | |
115 | #define MEM_PAGE_SIZE (4096L) |
116 | |
117 | typedef enum { |
118 | FILE_RES_OK, |
119 | FILE_RES_NOT_FOUND, |
120 | FILE_RES_ERROR |
121 | } file_res; |
122 | |
123 | typedef enum { |
124 | CHECK_PROC_PRESENT, |
125 | CHECK_PROC_PRESENT_NO_ARG, |
126 | CHECK_PROC_ABSENT |
127 | } check_proc_res; |
128 | |
129 | typedef uint16_t os_numa_node_index; |
130 | typedef uint16_t os_package_index; |
131 | typedef uint16_t os_core_index; |
132 | |
133 | typedef uint16_t irq_number; |
134 | |
135 | typedef struct { |
136 | uint16_t n_irqs; |
137 | irq_number irqs[CPU_SETSIZE]; |
138 | uint16_t per_cpu; |
139 | } irq_list; |
140 | |
141 | static cpu_set_t g_os_cpus_online; |
142 | static cpu_set_t g_numa_node_os_cpus_online[CPU_SETSIZE]; |
143 | |
144 | static uint16_t g_n_numa_nodes; |
145 | static uint16_t g_n_cores; |
146 | static uint16_t g_n_os_cpus; |
147 | static uint16_t g_n_cpus; |
148 | static uint16_t g_n_irq_cpus; |
149 | |
150 | static os_numa_node_index g_numa_node_index_to_os_numa_node_index[CPU_SETSIZE]; |
151 | static cf_topo_os_cpu_index g_core_index_to_os_cpu_index[CPU_SETSIZE]; |
152 | static cf_topo_os_cpu_index g_cpu_index_to_os_cpu_index[CPU_SETSIZE]; |
153 | static cf_topo_cpu_index g_os_cpu_index_to_cpu_index[CPU_SETSIZE]; |
154 | |
155 | static cf_topo_numa_node_index g_i_numa_node; |
156 | |
157 | #define DEVICE_PATH_SIZE 1024 |
158 | #define DEVICE_NAME_SIZE 256 |
159 | |
160 | #define MAX_DEVICE_CHILDREN 100 |
161 | #define MAX_DEVICE_SCHEDULERS 100 |
162 | |
163 | typedef struct dev_key_s { |
164 | uint32_t major; |
165 | uint32_t minor; |
166 | } dev_key_t; |
167 | |
168 | typedef struct dev_node_s { |
169 | uint32_t n_children; |
170 | struct dev_node_s *children[MAX_DEVICE_CHILDREN]; |
171 | |
172 | char name[DEVICE_NAME_SIZE]; |
173 | char dev_path[DEVICE_PATH_SIZE]; |
174 | |
175 | char sys_home[DEVICE_PATH_SIZE]; |
176 | char sys_sched[DEVICE_PATH_SIZE]; |
177 | } dev_node_t; |
178 | |
179 | typedef struct path_data_s { |
180 | cf_storage_device_info info; |
181 | |
182 | uint32_t n_sys_scheds; |
183 | const char *sys_scheds[MAX_DEVICE_SCHEDULERS]; |
184 | |
185 | cf_clock mod_time; |
186 | } path_data_t; |
187 | |
188 | static cf_shash *g_dev_graph; |
189 | |
190 | static cf_mutex g_path_data_lock = CF_MUTEX_INIT; |
191 | static cf_shash *g_path_data; |
192 | |
193 | static file_res |
194 | read_file(const char *path, void *buff, size_t *limit) |
195 | { |
196 | cf_detail(CF_HARDWARE, "reading file %s with buffer size %zu" , path, *limit); |
197 | int32_t fd = open(path, O_RDONLY); |
198 | |
199 | if (fd < 0) { |
200 | if (errno == ENOENT) { |
201 | cf_detail(CF_HARDWARE, "file %s not found" , path); |
202 | return FILE_RES_NOT_FOUND; |
203 | } |
204 | |
205 | cf_warning(CF_HARDWARE, "error while opening file %s for reading: %d (%s)" , |
206 | path, errno, cf_strerror(errno)); |
207 | return FILE_RES_ERROR; |
208 | } |
209 | |
210 | size_t total = 0; |
211 | |
212 | while (total < *limit) { |
213 | cf_detail(CF_HARDWARE, "reading %zd byte(s) at offset %zu" , *limit - total, total); |
214 | ssize_t len = read(fd, (uint8_t *)buff + total, *limit - total); |
215 | CF_NEVER_FAILS(len); |
216 | |
217 | if (len == 0) { |
218 | cf_detail(CF_HARDWARE, "EOF" ); |
219 | break; |
220 | } |
221 | |
222 | total += (size_t)len; |
223 | } |
224 | |
225 | cf_detail(CF_HARDWARE, "read %zu byte(s) from file %s" , total, path); |
226 | file_res res; |
227 | |
228 | if (total == *limit) { |
229 | cf_warning(CF_HARDWARE, "read buffer too small for file %s" , path); |
230 | res = FILE_RES_ERROR; |
231 | } |
232 | else { |
233 | res = FILE_RES_OK; |
234 | *limit = total; |
235 | } |
236 | |
237 | CF_NEVER_FAILS(close(fd)); |
238 | return res; |
239 | } |
240 | |
241 | static file_res |
242 | write_file(const char *path, const void *buff, size_t limit) |
243 | { |
244 | cf_detail(CF_HARDWARE, "writing file %s with buffer size %zu" , path, limit); |
245 | int32_t fd = open(path, O_WRONLY | O_CREAT | O_TRUNC, 0600); |
246 | |
247 | if (fd < 0) { |
248 | if (errno == ENOENT) { |
249 | cf_detail(CF_HARDWARE, "file %s not found" , path); |
250 | return FILE_RES_NOT_FOUND; |
251 | } |
252 | |
253 | cf_warning(CF_HARDWARE, "error while opening file %s for writing: %d (%s)" , |
254 | path, errno, cf_strerror(errno)); |
255 | return FILE_RES_ERROR; |
256 | } |
257 | |
258 | size_t total = 0; |
259 | |
260 | while (total < limit) { |
261 | cf_detail(CF_HARDWARE, "writing %zd byte(s) at offset %zu" , limit - total, total); |
262 | ssize_t len = write(fd, (uint8_t *)buff + total, limit - total); |
263 | |
264 | if (len < 0) { |
265 | cf_warning(CF_HARDWARE, "error while writing to file %s: %d (%s)" , |
266 | path, errno, cf_strerror(errno)); |
267 | CF_NEVER_FAILS(close(fd)); |
268 | return FILE_RES_ERROR; |
269 | } |
270 | |
271 | total += (size_t)len; |
272 | } |
273 | |
274 | cf_detail(CF_HARDWARE, "done writing" ); |
275 | CF_NEVER_FAILS(close(fd)); |
276 | return FILE_RES_OK; |
277 | } |
278 | |
279 | static void |
280 | write_file_safe(const char *path, const void *buff, size_t limit) |
281 | { |
282 | if (write_file(path, buff, limit) != FILE_RES_OK) { |
283 | cf_crash(CF_HARDWARE, "write failed unexpectedly" ); |
284 | } |
285 | } |
286 | |
287 | static DIR * |
288 | opendir_safe(const char *path) |
289 | { |
290 | DIR *dir = opendir(path); |
291 | |
292 | if (dir == NULL) { |
293 | cf_crash(CF_HARDWARE, "error while opening directory %s: %d (%s)" , |
294 | path, errno, cf_strerror(errno)); |
295 | } |
296 | |
297 | return dir; |
298 | } |
299 | |
300 | static int32_t |
301 | readdir_safe(DIR *dir, struct dirent *ent) |
302 | { |
303 | while (true) { |
304 | errno = 0; |
305 | struct dirent *tmp = readdir(dir); |
306 | |
307 | if (tmp == NULL) { |
308 | if (errno != 0) { |
309 | cf_crash(CF_HARDWARE, "error while reading directory: %d (%s)" , |
310 | errno, cf_strerror(errno)); |
311 | } |
312 | |
313 | return -1; |
314 | } |
315 | |
316 | if (strcmp(tmp->d_name, "." ) == 0 || strcmp(tmp->d_name, ".." ) == 0) { |
317 | continue; |
318 | } |
319 | |
320 | memcpy(ent, tmp, sizeof(struct dirent)); |
321 | return 0; |
322 | } |
323 | } |
324 | |
325 | static void |
326 | closedir_safe(DIR *dir) |
327 | { |
328 | if (closedir(dir) < 0) { |
329 | cf_crash(CF_HARDWARE, "error while closing PCI device directory: %d (%s)" , |
330 | errno, cf_strerror(errno)); |
331 | } |
332 | } |
333 | |
334 | static bool |
335 | path_exists(const char *path) |
336 | { |
337 | struct stat st; |
338 | |
339 | if (stat(path, &st) < 0) { |
340 | if (errno == ENOENT) { |
341 | cf_detail(CF_HARDWARE, "path %s does not exist" , path); |
342 | return false; |
343 | } |
344 | |
345 | cf_crash(CF_HARDWARE, "error while checking for path %s: %d (%s)" , |
346 | path, errno, cf_strerror(errno)); |
347 | } |
348 | |
349 | cf_detail(CF_HARDWARE, "path %s exists" , path); |
350 | return true; |
351 | } |
352 | |
353 | static bool |
354 | path_is_dir(const char *path) |
355 | { |
356 | struct stat st; |
357 | |
358 | if (stat(path, &st) < 0) { |
359 | cf_crash(CF_HARDWARE, "error while checking path %s: %d (%s)" , |
360 | path, errno, cf_strerror(errno)); |
361 | } |
362 | |
363 | bool is_dir = S_ISDIR(st.st_mode); |
364 | |
365 | cf_detail(CF_HARDWARE, "path %s is %s directory" , path, is_dir ? |
366 | "a" : "not a" ); |
367 | |
368 | return is_dir; |
369 | } |
370 | |
371 | static bool |
372 | path_works(const char *path) |
373 | { |
374 | int32_t fd = open(path, O_RDONLY); |
375 | |
376 | if (fd < 0) { |
377 | if (errno == ENOENT || errno == EINVAL) { |
378 | cf_detail(CF_HARDWARE, "path %s does not work (open): %d (%s)" , |
379 | path, errno, cf_strerror(errno)); |
380 | return false; |
381 | } |
382 | |
383 | cf_crash(CF_HARDWARE, "error while verifying path %s (open): %d (%s)" , |
384 | path, errno, cf_strerror(errno)); |
385 | } |
386 | |
387 | uint8_t buff[1000]; |
388 | |
389 | if (read(fd, buff, sizeof(buff)) < 0) { |
390 | if (errno == EINVAL) { |
391 | cf_detail(CF_HARDWARE, "path %s does not work (read): %d (%s)" , |
392 | path, errno, cf_strerror(errno)); |
393 | CF_NEVER_FAILS(close(fd)); |
394 | return false; |
395 | } |
396 | |
397 | cf_crash(CF_HARDWARE, "error while verifying path %s (read): %d (%s)" , |
398 | path, errno, cf_strerror(errno)); |
399 | } |
400 | |
401 | cf_detail(CF_HARDWARE, "path %s works" , path); |
402 | CF_NEVER_FAILS(close(fd)); |
403 | return true; |
404 | } |
405 | |
406 | static void |
407 | set_mempolicy_safe(uint32_t mode, uint64_t *node_mask, size_t max_node) |
408 | { |
409 | if (syscall(__NR_set_mempolicy, mode, node_mask, max_node) < 0) { |
410 | cf_crash(CF_HARDWARE, "set_mempolicy() system call failed: %d (%s)" , |
411 | errno, cf_strerror(errno)); |
412 | } |
413 | } |
414 | |
415 | static void |
416 | migrate_pages_safe(pid_t pid, size_t max_node, uint64_t *from_mask, uint64_t *to_mask) |
417 | { |
418 | int64_t res = syscall(__NR_migrate_pages, pid, max_node, from_mask, to_mask); |
419 | |
420 | if (res < 0) { |
421 | cf_crash(CF_HARDWARE, "migrate_pages() syscall failed: %d (%s)" , |
422 | errno, cf_strerror(errno)); |
423 | } |
424 | |
425 | if (res > 0) { |
426 | cf_warning(CF_HARDWARE, "could not NUMA-migrate %" PRId64 " page(s)" , res); |
427 | } |
428 | } |
429 | |
430 | static void |
431 | mask_to_string(cpu_set_t *mask, char *buff, size_t limit) |
432 | { |
433 | cf_topo_os_cpu_index max; |
434 | |
435 | for (max = CPU_SETSIZE - 1; max > 0; --max) { |
436 | if (CPU_ISSET(max, mask)) { |
437 | break; |
438 | } |
439 | } |
440 | |
441 | int32_t words = max / 32 + 1; |
442 | size_t size = (size_t)words * 9; |
443 | |
444 | if (size > limit) { |
445 | cf_crash(CF_HARDWARE, "CPU mask buffer overflow: %zu vs. %zu" , size, limit); |
446 | } |
447 | |
448 | for (int32_t i = words - 1; i >= 0; --i) { |
449 | uint32_t val = 0; |
450 | |
451 | for (int32_t k = 0; k < 32; ++k) { |
452 | if (CPU_ISSET((size_t)(i * 32 + k), mask)) { |
453 | val |= 1u << k; |
454 | } |
455 | } |
456 | |
457 | snprintf(buff, limit, "%08x" , val); |
458 | |
459 | if (i > 0) { |
460 | buff[8] = ','; |
461 | } |
462 | |
463 | buff += 9; |
464 | limit -= 9; |
465 | } |
466 | } |
467 | |
468 | static file_res |
469 | read_value(const char *path, int64_t *val) |
470 | { |
471 | cf_detail(CF_HARDWARE, "reading value from file %s" , path); |
472 | |
473 | char buff[100]; |
474 | size_t limit = sizeof(buff); |
475 | file_res res = read_file(path, buff, &limit); |
476 | |
477 | if (res != FILE_RES_OK) { |
478 | return res; |
479 | } |
480 | |
481 | buff[limit - 1] = '\0'; |
482 | |
483 | cf_detail(CF_HARDWARE, "parsing value \"%s\"" , buff); |
484 | |
485 | char *end; |
486 | int64_t x = strtol(buff, &end, 10); |
487 | |
488 | if (*end != '\0' || x >= CPU_SETSIZE) { |
489 | cf_warning(CF_HARDWARE, "invalid value \"%s\" in %s" , buff, path); |
490 | return FILE_RES_ERROR; |
491 | } |
492 | |
493 | *val = x; |
494 | return FILE_RES_OK; |
495 | } |
496 | |
497 | static file_res |
498 | read_index(const char *path, uint16_t *val) |
499 | { |
500 | int64_t x; |
501 | file_res res = read_value(path, &x); |
502 | |
503 | if (res != FILE_RES_OK) { |
504 | return res; |
505 | } |
506 | |
507 | if (x < 0) { |
508 | cf_warning(CF_HARDWARE, "invalid index in %s" , path); |
509 | return FILE_RES_ERROR; |
510 | } |
511 | |
512 | *val = (uint16_t)x; |
513 | return FILE_RES_OK; |
514 | } |
515 | |
516 | static file_res |
517 | read_numa_node(const char *path, cf_topo_numa_node_index *i_numa_node) |
518 | { |
519 | int64_t x; |
520 | file_res res = read_value(path, &x); |
521 | |
522 | if (res != FILE_RES_OK) { |
523 | return res; |
524 | } |
525 | |
526 | if (x < 0) { |
527 | cf_detail(CF_HARDWARE, "no NUMA node in %s" , path); |
528 | return FILE_RES_ERROR; |
529 | } |
530 | |
531 | *i_numa_node = (cf_topo_numa_node_index)x; |
532 | return FILE_RES_OK; |
533 | } |
534 | |
535 | static file_res |
536 | read_device_numbers(const char *path, uint32_t *major, uint32_t *minor) |
537 | { |
538 | cf_detail(CF_HARDWARE, "reading device numbers from file %s" , path); |
539 | |
540 | char buff[100]; |
541 | size_t limit = sizeof(buff); |
542 | file_res res = read_file(path, buff, &limit); |
543 | |
544 | if (res != FILE_RES_OK) { |
545 | return res; |
546 | } |
547 | |
548 | buff[limit - 1] = '\0'; |
549 | |
550 | cf_detail(CF_HARDWARE, "parsing device numbers \"%s\"" , buff); |
551 | |
552 | if (sscanf(buff, "%u:%u\n" , major, minor) != 2) { |
553 | cf_warning(CF_HARDWARE, "invalid device numbers \"%s\" in %s" , buff, |
554 | path); |
555 | return FILE_RES_ERROR; |
556 | } |
557 | |
558 | return FILE_RES_OK; |
559 | } |
560 | |
561 | static file_res |
562 | read_list(const char *path, cpu_set_t *mask) |
563 | { |
564 | cf_detail(CF_HARDWARE, "reading list from file %s" , path); |
565 | char buff[1000]; |
566 | size_t limit = sizeof(buff); |
567 | file_res res = read_file(path, buff, &limit); |
568 | |
569 | if (res != FILE_RES_OK) { |
570 | return res; |
571 | } |
572 | |
573 | buff[limit - 1] = '\0'; |
574 | cf_detail(CF_HARDWARE, "parsing list \"%s\"" , buff); |
575 | |
576 | CPU_ZERO(mask); |
577 | char *walker = buff; |
578 | |
579 | while (true) { |
580 | char *delim; |
581 | uint64_t from = strtoul(walker, &delim, 10); |
582 | uint64_t thru; |
583 | |
584 | if (*delim == ',' || *delim == '\0'){ |
585 | thru = from; |
586 | } |
587 | else if (*delim == '-') { |
588 | walker = delim + 1; |
589 | thru = strtoul(walker, &delim, 10); |
590 | } |
591 | else { |
592 | cf_warning(CF_HARDWARE, "invalid list \"%s\" in %s" , buff, path); |
593 | return FILE_RES_ERROR; |
594 | } |
595 | |
596 | if (from >= CPU_SETSIZE || thru >= CPU_SETSIZE || from > thru) { |
597 | cf_warning(CF_HARDWARE, "invalid list \"%s\" in %s" , buff, path); |
598 | return FILE_RES_ERROR; |
599 | } |
600 | |
601 | cf_detail(CF_HARDWARE, "marking %d through %d" , (int32_t)from, (int32_t)thru); |
602 | |
603 | for (size_t i = from; i <= thru; ++i) { |
604 | CPU_SET(i, mask); |
605 | } |
606 | |
607 | if (*delim == '\0') { |
608 | break; |
609 | } |
610 | |
611 | walker = delim + 1; |
612 | } |
613 | |
614 | char buff2[1000]; |
615 | mask_to_string(mask, buff2, sizeof(buff2)); |
616 | cf_detail(CF_HARDWARE, "list \"%s\" -> mask %s" , buff, buff2); |
617 | |
618 | return FILE_RES_OK; |
619 | } |
620 | |
621 | static void |
622 | detect(cf_topo_numa_node_index a_numa_node) |
623 | { |
624 | if (a_numa_node == INVALID_INDEX) { |
625 | cf_detail(CF_HARDWARE, "detecting online CPUs" ); |
626 | } |
627 | else { |
628 | cf_detail(CF_HARDWARE, "detecting online CPUs on NUMA node %hu" , a_numa_node); |
629 | } |
630 | |
631 | if (read_list("/sys/devices/system/cpu/online" , &g_os_cpus_online) != FILE_RES_OK) { |
632 | cf_crash(CF_HARDWARE, "error while reading list of online CPUs" ); |
633 | } |
634 | |
635 | cf_detail(CF_HARDWARE, "learning CPU topology" ); |
636 | |
637 | cf_topo_numa_node_index os_numa_node_index_to_numa_node_index[CPU_SETSIZE]; |
638 | |
639 | for (int32_t i = 0; i < CPU_SETSIZE; ++i) { |
640 | CPU_ZERO(&g_numa_node_os_cpus_online[i]); |
641 | |
642 | g_core_index_to_os_cpu_index[i] = INVALID_INDEX; |
643 | g_cpu_index_to_os_cpu_index[i] = INVALID_INDEX; |
644 | g_os_cpu_index_to_cpu_index[i] = INVALID_INDEX; |
645 | |
646 | os_numa_node_index_to_numa_node_index[i] = INVALID_INDEX; |
647 | g_numa_node_index_to_os_numa_node_index[i] = INVALID_INDEX; |
648 | } |
649 | |
650 | cpu_set_t covered_numa_nodes; |
651 | cpu_set_t covered_cores[CPU_SETSIZE]; // One mask per package. |
652 | |
653 | CPU_ZERO(&covered_numa_nodes); |
654 | |
655 | for (int32_t i = 0; i < CPU_SETSIZE; ++i) { |
656 | CPU_ZERO(&covered_cores[i]); |
657 | } |
658 | |
659 | g_n_numa_nodes = 0; |
660 | g_n_cores = 0; |
661 | g_n_os_cpus = 0; |
662 | g_n_cpus = 0; |
663 | char path[1000]; |
664 | bool no_numa = false; |
665 | |
666 | // Loop through all CPUs in the system by looping through OS CPU indexes. |
667 | |
668 | for (g_n_os_cpus = 0; g_n_os_cpus < CPU_SETSIZE; ++g_n_os_cpus) { |
669 | cf_detail(CF_HARDWARE, "querying OS CPU index %hu" , g_n_os_cpus); |
670 | |
671 | // Let's look at the CPU's package. |
672 | |
673 | snprintf(path, sizeof(path), |
674 | "/sys/devices/system/cpu/cpu%hu/topology/physical_package_id" , |
675 | g_n_os_cpus); |
676 | os_package_index i_os_package; |
677 | file_res res = read_index(path, &i_os_package); |
678 | |
679 | // The entry doesn't exist. We've processed all available CPUs. Stop |
680 | // looping through the CPUs. |
681 | |
682 | if (res == FILE_RES_NOT_FOUND) { |
683 | break; |
684 | } |
685 | |
686 | if (res != FILE_RES_OK) { |
687 | cf_crash(CF_HARDWARE, "error while reading OS package index from %s" , path); |
688 | break; |
689 | } |
690 | |
691 | cf_detail(CF_HARDWARE, "OS package index is %hu" , i_os_package); |
692 | |
693 | // Only consider CPUs that are actually in use. |
694 | |
695 | if (!CPU_ISSET(g_n_os_cpus, &g_os_cpus_online)) { |
696 | cf_detail(CF_HARDWARE, "OS CPU index %hu is offline" , g_n_os_cpus); |
697 | continue; |
698 | } |
699 | |
700 | // Let's look at the CPU's underlying core. In Hyper Threading systems, |
701 | // two (logical) CPUs share one (physical) core. |
702 | |
703 | snprintf(path, sizeof(path), |
704 | "/sys/devices/system/cpu/cpu%hu/topology/core_id" , |
705 | g_n_os_cpus); |
706 | os_core_index i_os_core; |
707 | res = read_index(path, &i_os_core); |
708 | |
709 | if (res != FILE_RES_OK) { |
710 | cf_crash(CF_HARDWARE, "error while reading OS core index from %s" , path); |
711 | break; |
712 | } |
713 | |
714 | cf_detail(CF_HARDWARE, "OS core index is %hu" , i_os_core); |
715 | |
716 | // Consider a core when we see it for the first time. In other words, we |
717 | // consider the first Hyper Threading peer of each core to be that core. |
718 | |
719 | bool new_core; |
720 | |
721 | if (CPU_ISSET(i_os_core, &covered_cores[i_os_package])) { |
722 | cf_detail(CF_HARDWARE, "core (%hu, %hu) already covered" , i_os_core, i_os_package); |
723 | new_core = false; |
724 | } |
725 | else { |
726 | cf_detail(CF_HARDWARE, "core (%hu, %hu) is new" , i_os_core, i_os_package); |
727 | new_core = true; |
728 | CPU_SET(i_os_core, &covered_cores[i_os_package]); |
729 | } |
730 | |
731 | // Identify the NUMA node of the current CPU. We simply look for the |
732 | // current CPU's topology info subtree in each NUMA node's subtree. |
733 | // Specifically, we look for the current CPU's "core_id" entry. |
734 | |
735 | os_numa_node_index i_os_numa_node; |
736 | |
737 | for (i_os_numa_node = 0; i_os_numa_node < CPU_SETSIZE; ++i_os_numa_node) { |
738 | snprintf(path, sizeof(path), |
739 | "/sys/devices/system/cpu/cpu%hu/node%hu/cpu%hu/topology/core_id" , |
740 | g_n_os_cpus, i_os_numa_node, g_n_os_cpus); |
741 | uint16_t dummy; |
742 | res = read_index(path, &dummy); |
743 | |
744 | // We found the NUMA node that has the current CPU in its subtree. |
745 | |
746 | if (res == FILE_RES_OK) { |
747 | break; |
748 | } |
749 | |
750 | if (res != FILE_RES_NOT_FOUND) { |
751 | cf_crash(CF_HARDWARE, "error while reading core number from %s" , path); |
752 | } |
753 | } |
754 | |
755 | // Some Docker installations seem to not have any NUMA information |
756 | // in /sys. In this case, assume a system with a single NUMA node. |
757 | |
758 | if (i_os_numa_node == CPU_SETSIZE) { |
759 | cf_detail(CF_HARDWARE, "OS CPU index %hu does not have a NUMA node" , g_n_os_cpus); |
760 | no_numa = true; |
761 | i_os_numa_node = 0; |
762 | } |
763 | |
764 | cf_detail(CF_HARDWARE, "OS NUMA node index is %hu" , i_os_numa_node); |
765 | |
766 | // Again, just like with cores, we consider a NUMA node when we encounter |
767 | // it for the first time. |
768 | |
769 | bool new_numa_node; |
770 | |
771 | if (CPU_ISSET(i_os_numa_node, &covered_numa_nodes)) { |
772 | cf_detail(CF_HARDWARE, "OS NUMA node index %hu already covered" , i_os_numa_node); |
773 | new_numa_node = false; |
774 | } |
775 | else { |
776 | cf_detail(CF_HARDWARE, "OS NUMA node index %hu is new" , i_os_numa_node); |
777 | new_numa_node = true; |
778 | CPU_SET(i_os_numa_node, &covered_numa_nodes); |
779 | |
780 | // For now, we only support a 64-bit bitmask (= one uint64_t). |
781 | |
782 | if (i_os_numa_node >= 64) { |
783 | cf_crash(CF_HARDWARE, "OS NUMA node index %hu too high" , i_os_numa_node); |
784 | } |
785 | } |
786 | |
787 | // Now we know that the CPU is online and we know, whether it is in a newly |
788 | // seen core (new_core) and/or a newly seen NUMA node (new_numa_node). |
789 | |
790 | cf_topo_numa_node_index i_numa_node; |
791 | |
792 | if (new_numa_node) { |
793 | i_numa_node = g_n_numa_nodes; |
794 | ++g_n_numa_nodes; |
795 | os_numa_node_index_to_numa_node_index[i_os_numa_node] = i_numa_node; |
796 | g_numa_node_index_to_os_numa_node_index[i_numa_node] = i_os_numa_node; |
797 | cf_detail(CF_HARDWARE, "OS NUMA node index %hu -> new NUMA node index %hu" , |
798 | i_os_numa_node, i_numa_node); |
799 | } |
800 | else { |
801 | i_numa_node = os_numa_node_index_to_numa_node_index[i_os_numa_node]; |
802 | cf_detail(CF_HARDWARE, "OS NUMA node index %hu -> existing NUMA node index %hu" , |
803 | i_os_numa_node, i_numa_node); |
804 | } |
805 | |
806 | cf_detail(CF_HARDWARE, "OS CPU index %hu on NUMA node index %hu" , g_n_os_cpus, i_numa_node); |
807 | CPU_SET(g_n_os_cpus, &g_numa_node_os_cpus_online[i_numa_node]); |
808 | |
809 | // If we're in NUMA mode and the CPU isn't on the NUMA mode that we're |
810 | // running on, then ignore the CPU. |
811 | |
812 | if (a_numa_node != INVALID_INDEX && a_numa_node != i_numa_node) { |
813 | cf_detail(CF_HARDWARE, "skipping unwanted NUMA node index %hu" , i_numa_node); |
814 | continue; |
815 | } |
816 | |
817 | // If the CPU is a new core, then map a new core index to the OS CPU index. |
818 | |
819 | if (new_core) { |
820 | g_core_index_to_os_cpu_index[g_n_cores] = g_n_os_cpus; |
821 | cf_detail(CF_HARDWARE, "core index %hu -> OS CPU index %hu" , g_n_cores, g_n_os_cpus); |
822 | ++g_n_cores; |
823 | } |
824 | |
825 | // Map the OS CPU index to a new CPU index and vice versa. |
826 | |
827 | g_os_cpu_index_to_cpu_index[g_n_os_cpus] = g_n_cpus; |
828 | g_cpu_index_to_os_cpu_index[g_n_cpus] = g_n_os_cpus; |
829 | |
830 | cf_detail(CF_HARDWARE, "OS CPU index %hu <-> CPU index %hu" , g_n_os_cpus, g_n_cpus); |
831 | ++g_n_cpus; |
832 | } |
833 | |
834 | if (g_n_os_cpus == CPU_SETSIZE) { |
835 | cf_crash(CF_HARDWARE, "too many CPUs" ); |
836 | } |
837 | |
838 | if (a_numa_node != INVALID_INDEX && no_numa) { |
839 | cf_warning(CF_HARDWARE, "no NUMA information found in /sys" ); |
840 | } |
841 | |
842 | g_i_numa_node = a_numa_node; |
843 | } |
844 | |
845 | static void |
846 | pin_to_numa_node(cf_topo_numa_node_index a_numa_node) |
847 | { |
848 | cf_info(CF_HARDWARE, "pinning to NUMA node %hu" , a_numa_node); |
849 | |
850 | // Move the current thread (and all of its future descendants) to the CPUs |
851 | // on the selected NUMA node. |
852 | |
853 | cpu_set_t cpu_set; |
854 | CPU_ZERO(&cpu_set); |
855 | |
856 | for (cf_topo_cpu_index i_cpu = 0; i_cpu < g_n_cpus; ++i_cpu) { |
857 | cf_topo_os_cpu_index i_os_cpu = g_cpu_index_to_os_cpu_index[i_cpu]; |
858 | CPU_SET(i_os_cpu, &cpu_set); |
859 | } |
860 | |
861 | char buff[1000]; |
862 | mask_to_string(&cpu_set, buff, sizeof(buff)); |
863 | cf_detail(CF_HARDWARE, "NUMA node %hu CPU mask: %s" , a_numa_node, buff); |
864 | |
865 | if (sched_setaffinity(0, sizeof(cpu_set), &cpu_set) < 0) { |
866 | cf_crash(CF_HARDWARE, "error while pinning thread to NUMA node %hu: %d (%s)" , |
867 | a_numa_node, errno, cf_strerror(errno)); |
868 | } |
869 | |
870 | // Force future memory allocations to the selected NUMA node. |
871 | |
872 | os_numa_node_index i_os_numa_node = g_numa_node_index_to_os_numa_node_index[a_numa_node]; |
873 | uint64_t to_mask = 1UL << i_os_numa_node; |
874 | cf_detail(CF_HARDWARE, "NUMA node mask (to): %016" PRIx64, to_mask); |
875 | |
876 | // Unlike select(), we have to pass "number of valid bits + 1". |
877 | set_mempolicy_safe(MPOL_BIND, &to_mask, 65); |
878 | |
879 | // Make sure we can migrate shared memory that we later attach and map. |
880 | cf_process_add_startup_cap(CAP_SYS_NICE); |
881 | } |
882 | |
883 | static uint32_t |
884 | pick_random(uint32_t limit) |
885 | { |
886 | static __thread uint64_t state = 0; |
887 | |
888 | if (state == 0) { |
889 | state = (uint64_t)syscall(SYS_gettid); |
890 | } |
891 | |
892 | state = state * 6364136223846793005 + 1; |
893 | |
894 | if (state == 0) { |
895 | state = 1; |
896 | } |
897 | |
898 | return (uint32_t)((state >> 32) % limit); |
899 | } |
900 | |
901 | uint16_t |
902 | cf_topo_count_cores(void) |
903 | { |
904 | return g_n_cores; |
905 | } |
906 | |
907 | uint16_t |
908 | cf_topo_count_cpus(void) |
909 | { |
910 | return g_n_cpus; |
911 | } |
912 | |
913 | static cf_topo_cpu_index |
914 | os_cpu_index_to_cpu_index(cf_topo_os_cpu_index i_os_cpu) |
915 | { |
916 | cf_detail(CF_HARDWARE, "translating OS CPU index %hu" , i_os_cpu); |
917 | |
918 | if (i_os_cpu >= g_n_os_cpus) { |
919 | cf_crash(CF_HARDWARE, "invalid OS CPU index %hu" , i_os_cpu); |
920 | } |
921 | |
922 | cf_topo_cpu_index i_cpu = g_os_cpu_index_to_cpu_index[i_os_cpu]; |
923 | |
924 | if (i_cpu == INVALID_INDEX) { |
925 | cf_detail(CF_HARDWARE, "foreign OS CPU index %hu" , i_os_cpu); |
926 | } |
927 | else { |
928 | cf_detail(CF_HARDWARE, "CPU index is %hu" , i_cpu); |
929 | } |
930 | |
931 | return i_cpu; |
932 | } |
933 | |
934 | cf_topo_cpu_index |
935 | cf_topo_current_cpu(void) |
936 | { |
937 | cf_detail(CF_HARDWARE, "getting current OS CPU index" ); |
938 | int32_t os = sched_getcpu(); |
939 | |
940 | if (os < 0) { |
941 | cf_crash(CF_HARDWARE, "error while getting OS CPU index: %d (%s)" , |
942 | errno, cf_strerror(errno)); |
943 | } |
944 | |
945 | return os_cpu_index_to_cpu_index((cf_topo_os_cpu_index)os); |
946 | } |
947 | |
948 | cf_topo_cpu_index |
949 | cf_topo_socket_cpu(const cf_socket *sock) |
950 | { |
951 | cf_detail(CF_HARDWARE, "determining CPU index for socket FD %d" , CSFD(sock)); |
952 | |
953 | int32_t os; |
954 | socklen_t len = sizeof(os); |
955 | |
956 | if (getsockopt(sock->fd, SOL_SOCKET, SO_INCOMING_CPU, &os, &len) < 0) { |
957 | cf_crash(CF_HARDWARE, "error while determining incoming OS CPU index: %d (%s)" , |
958 | errno, cf_strerror(errno)); |
959 | } |
960 | |
961 | cf_detail(CF_HARDWARE, "OS CPU index is %d" , os); |
962 | cf_topo_cpu_index i_cpu = os_cpu_index_to_cpu_index((cf_topo_os_cpu_index)os); |
963 | |
964 | // 1. The incoming connection was handled on the wrong NUMA node. In this case, |
965 | // pick a random CPU on the correct NUMA node. |
966 | |
967 | if (i_cpu == INVALID_INDEX) { |
968 | i_cpu = (cf_topo_cpu_index)pick_random(g_n_cpus); |
969 | cf_detail(CF_HARDWARE, "picking random CPU index %hu" , i_cpu); |
970 | return i_cpu; |
971 | } |
972 | |
973 | // 2. The incoming connection was handled on a CPU that doesn't get any NIC |
974 | // interrupts. This should not happen for connections from other machines, but |
975 | // it does happen for connections from the local machine, because they don't |
976 | // go through the NIC hardware. In this case, pick a random CPU. |
977 | |
978 | if (i_cpu >= g_n_irq_cpus) { |
979 | i_cpu = (cf_topo_cpu_index)pick_random(g_n_cpus); |
980 | cf_detail(CF_HARDWARE, "randomizing unexpected CPU index >%hu to %hu" , |
981 | g_n_irq_cpus - 1, i_cpu); |
982 | return i_cpu; |
983 | } |
984 | |
985 | // 3. Otherwise, redistribute. The first g_n_irq_cpus CPUs out of a total of |
986 | // g_n_cpus CPUs get NIC interrupts. Suppose we have 2 NIC queues and 8 CPUs, |
987 | // i.e., that g_n_irq_cpus == 2 and g_n_cpus == 8. We want to redistribute |
988 | // evenly across the 8 CPUs, i.e., each CPU should be picked with a probability |
989 | // of 0.125. |
990 | |
991 | // We're currently running on one of the 2 CPUs that get NIC interrupts, on |
992 | // either with a probability of p1 = 0.5. We want to stay on the current CPU |
993 | // with a probability of p2 = g_n_irq_cpus / g_n_cpus == 2 / 8 == 0.25, which |
994 | // yields the desired total probability of p1 * p2 = 0.5 * 0.25 = 0.125. |
995 | |
996 | if (pick_random(100000) < g_n_irq_cpus * (uint32_t)100000 / g_n_cpus) { |
997 | cf_detail(CF_HARDWARE, "staying on CPU index %hu" , i_cpu); |
998 | return i_cpu; |
999 | } |
1000 | |
1001 | // 4. Otherwise, if we switch CPUs, then we jump to a CPU that doesn't receive |
1002 | // NIC interrupts, i.e., one of the remaining 6 CPUs [2 .. 8] in our example. |
1003 | // This reaches each CPU with a probability of (1 - p2) / 6 = 0.125. |
1004 | |
1005 | i_cpu = (cf_topo_cpu_index)(g_n_irq_cpus + |
1006 | pick_random((uint32_t)g_n_cpus - (uint32_t)g_n_irq_cpus)); |
1007 | cf_detail(CF_HARDWARE, "redirecting to CPU index %hu" , i_cpu); |
1008 | return i_cpu; |
1009 | } |
1010 | |
1011 | cf_topo_napi_id |
1012 | cf_topo_socket_napi_id(const cf_socket *sock) |
1013 | { |
1014 | cf_topo_napi_id id; |
1015 | socklen_t len = sizeof(id); |
1016 | |
1017 | if (getsockopt(sock->fd, SOL_SOCKET, SO_INCOMING_NAPI_ID, &id, &len) < 0) { |
1018 | cf_crash(CF_HARDWARE, "SO_INCOMING_NAPI_ID failed: %d (%s)" , errno, |
1019 | cf_strerror(errno)); |
1020 | } |
1021 | |
1022 | cf_detail(CF_HARDWARE, "incoming connection with NAPI-id %d" , id); |
1023 | return id; |
1024 | } |
1025 | |
1026 | static void |
1027 | pin_to_os_cpu(cf_topo_os_cpu_index i_os_cpu) |
1028 | { |
1029 | cf_detail(CF_HARDWARE, "pinning to OS CPU index %hu" , i_os_cpu); |
1030 | |
1031 | cpu_set_t cpu_set; |
1032 | CPU_ZERO(&cpu_set); |
1033 | CPU_SET(i_os_cpu, &cpu_set); |
1034 | |
1035 | if (sched_setaffinity(0, sizeof(cpu_set), &cpu_set) < 0) { |
1036 | cf_crash(CF_HARDWARE, "error while pinning thread to OS CPU %hu: %d (%s)" , |
1037 | i_os_cpu, errno, cf_strerror(errno)); |
1038 | } |
1039 | } |
1040 | |
1041 | void |
1042 | cf_topo_pin_to_core(cf_topo_core_index i_core) |
1043 | { |
1044 | cf_detail(CF_HARDWARE, "pinning to core index %hu" , i_core); |
1045 | |
1046 | if (i_core >= g_n_cores) { |
1047 | cf_crash(CF_HARDWARE, "invalid core index %hu" , i_core); |
1048 | } |
1049 | |
1050 | pin_to_os_cpu(g_core_index_to_os_cpu_index[i_core]); |
1051 | } |
1052 | |
1053 | void |
1054 | cf_topo_pin_to_cpu(cf_topo_cpu_index i_cpu) |
1055 | { |
1056 | cf_detail(CF_HARDWARE, "pinning to CPU index %hu" , i_cpu); |
1057 | |
1058 | if (i_cpu >= g_n_cpus) { |
1059 | cf_crash(CF_HARDWARE, "invalid CPU index %hu" , i_cpu); |
1060 | } |
1061 | |
1062 | pin_to_os_cpu(g_cpu_index_to_os_cpu_index[i_cpu]); |
1063 | } |
1064 | |
1065 | static check_proc_res |
1066 | check_proc(const char *name, int32_t argc, const char *argv[]) |
1067 | { |
1068 | cf_detail(CF_HARDWARE, "looking for process %s" , name); |
1069 | |
1070 | for (int32_t i = 0; i < argc; ++i) { |
1071 | cf_detail(CF_HARDWARE, "argv[%d]: %s" , i, argv[i]); |
1072 | } |
1073 | |
1074 | DIR *dir = opendir_safe("/proc" ); |
1075 | struct dirent ent; |
1076 | char cmd[10000]; |
1077 | size_t limit; |
1078 | bool found = false; |
1079 | |
1080 | while (readdir_safe(dir, &ent) >= 0) { |
1081 | bool numeric = true; |
1082 | |
1083 | for (int32_t i = 0; ent.d_name[i] != 0; ++i) { |
1084 | if (!isascii(ent.d_name[i]) || !isdigit(ent.d_name[i])) { |
1085 | numeric = false; |
1086 | break; |
1087 | } |
1088 | } |
1089 | |
1090 | if (!numeric) { |
1091 | continue; |
1092 | } |
1093 | |
1094 | char path[500]; |
1095 | snprintf(path, sizeof(path), "/proc/%s/cmdline" , ent.d_name); |
1096 | |
1097 | limit = sizeof(cmd) - 1; |
1098 | file_res rfr = read_file(path, cmd, &limit); |
1099 | |
1100 | // Can legitimately happen, if the process has exited in the meantime. |
1101 | if (rfr == FILE_RES_NOT_FOUND) { |
1102 | continue; |
1103 | } |
1104 | |
1105 | if (rfr == FILE_RES_ERROR) { |
1106 | cf_crash(CF_HARDWARE, "error while reading file %s" , path); |
1107 | } |
1108 | |
1109 | if (limit > 0 && cmd[limit - 1] != 0) { |
1110 | cmd[limit] = 0; |
1111 | } |
1112 | |
1113 | const char *name2 = strrchr(cmd, '/'); |
1114 | |
1115 | if (name2 != NULL) { |
1116 | ++name2; |
1117 | } |
1118 | else { |
1119 | name2 = cmd; |
1120 | } |
1121 | |
1122 | if (strcmp(name2, name) == 0) { |
1123 | found = true; |
1124 | break; |
1125 | } |
1126 | } |
1127 | |
1128 | closedir_safe(dir); |
1129 | |
1130 | if (!found) { |
1131 | cf_detail(CF_HARDWARE, "process %s absent" , name); |
1132 | return CHECK_PROC_ABSENT; |
1133 | } |
1134 | |
1135 | cf_detail(CF_HARDWARE, "process %s is %s" , name, cmd); |
1136 | |
1137 | if (argc > 0) { |
1138 | int32_t i_arg = 0; |
1139 | |
1140 | for (size_t off = strlen(cmd) + 1; off < limit; off += strlen(cmd + off) + 1) { |
1141 | cf_detail(CF_HARDWARE, "checking argument %s against %s" , cmd + off, argv[i_arg]); |
1142 | |
1143 | if (strcmp(cmd + off, argv[i_arg]) == 0) { |
1144 | ++i_arg; |
1145 | |
1146 | if (i_arg >= argc) { |
1147 | break; |
1148 | } |
1149 | } |
1150 | else { |
1151 | i_arg = 0; |
1152 | } |
1153 | } |
1154 | |
1155 | if (i_arg >= argc) { |
1156 | cf_detail(CF_HARDWARE, "process %s present with argument" , name); |
1157 | return CHECK_PROC_PRESENT; |
1158 | } |
1159 | } |
1160 | |
1161 | cf_detail(CF_HARDWARE, "process %s present" , name); |
1162 | return CHECK_PROC_PRESENT_NO_ARG; |
1163 | } |
1164 | |
1165 | static uint16_t |
1166 | interface_queues(const char *if_name, const char *format) |
1167 | { |
1168 | uint16_t n_queues = 0; |
1169 | |
1170 | while (true) { |
1171 | char path[1000]; |
1172 | snprintf(path, sizeof(path), format, if_name, n_queues); |
1173 | cf_detail(CF_HARDWARE, "checking for working path %s" , path); |
1174 | |
1175 | if (!path_works(path)) { |
1176 | cf_detail(CF_HARDWARE, "path does not work" ); |
1177 | break; |
1178 | } |
1179 | |
1180 | ++n_queues; |
1181 | } |
1182 | |
1183 | cf_assert(n_queues != 0, CF_HARDWARE, "interface %s has no queues" , if_name); |
1184 | |
1185 | return n_queues; |
1186 | } |
1187 | |
1188 | static uint16_t |
1189 | interface_rx_queues(const char *if_name) |
1190 | { |
1191 | cf_detail(CF_HARDWARE, "getting receive queues for interface %s" , if_name); |
1192 | return interface_queues(if_name, "/sys/class/net/%s/queues/rx-%hu/rps_cpus" ); |
1193 | } |
1194 | |
1195 | static uint16_t |
1196 | interface_tx_queues(const char *if_name) |
1197 | { |
1198 | cf_detail(CF_HARDWARE, "getting transmit queues for interface %s" , if_name); |
1199 | return interface_queues(if_name, "/sys/class/net/%s/queues/tx-%hu/xps_cpus" ); |
1200 | } |
1201 | |
1202 | static int |
1203 | comp_irq_number(const void *lhs, const void *rhs) |
1204 | { |
1205 | return *(irq_number *)lhs - *(irq_number *)rhs; |
1206 | } |
1207 | |
1208 | static void |
1209 | interface_irqs(const char *if_name, irq_list *irqs) |
1210 | { |
1211 | cf_detail(CF_HARDWARE, "getting IRQs for interface %s" , if_name); |
1212 | |
1213 | DIR *dir = opendir_safe("/sys/bus/pci/devices" ); |
1214 | struct dirent ent; |
1215 | char path[PATH_MAX]; |
1216 | bool found = false; |
1217 | |
1218 | while (readdir_safe(dir, &ent) >= 0) { |
1219 | snprintf(path, sizeof(path), "/sys/bus/pci/devices/%s/net/%s/ifindex" , |
1220 | ent.d_name, if_name); |
1221 | bool exists = path_exists(path); |
1222 | |
1223 | if (!exists) { |
1224 | for (int32_t i = 0; i < 100; ++i) { |
1225 | snprintf(path, sizeof(path), "/sys/bus/pci/devices/%s/virtio%d/net/%s/ifindex" , |
1226 | ent.d_name, i, if_name); |
1227 | exists = path_exists(path); |
1228 | |
1229 | if (exists) { |
1230 | break; |
1231 | } |
1232 | } |
1233 | } |
1234 | |
1235 | if (!exists) { |
1236 | continue; |
1237 | } |
1238 | |
1239 | snprintf(path, sizeof(path), "/sys/bus/pci/devices/%s/msi_irqs" , ent.d_name); |
1240 | |
1241 | if (!path_exists(path)) { |
1242 | cf_crash(CF_HARDWARE, "interface %s does not support MSIs" , if_name); |
1243 | } |
1244 | |
1245 | cf_detail(CF_HARDWARE, "interface %s is %s" , if_name, ent.d_name); |
1246 | found = true; |
1247 | break; |
1248 | } |
1249 | |
1250 | closedir_safe(dir); |
1251 | |
1252 | if (!found) { |
1253 | cf_crash(CF_HARDWARE, "interface %s does not have a PCI device entry" , if_name); |
1254 | } |
1255 | |
1256 | dir = opendir_safe(path); |
1257 | int32_t count = 0; |
1258 | irq_number irq_nums[CPU_SETSIZE]; |
1259 | |
1260 | while (readdir_safe(dir, &ent) >= 0) { |
1261 | char *end; |
1262 | uint64_t tmp = strtoul(ent.d_name, &end, 10); |
1263 | |
1264 | if (*end != 0 || tmp > 65535) { |
1265 | cf_crash(CF_HARDWARE, "invalid IRQ number %s in %s" , ent.d_name, path); |
1266 | } |
1267 | |
1268 | if (count >= CPU_SETSIZE) { |
1269 | cf_crash(CF_HARDWARE, "too many IRQs in %s" , path); |
1270 | } |
1271 | |
1272 | cf_detail(CF_HARDWARE, "interface %s has IRQ %hu" , if_name, (irq_number)tmp); |
1273 | irq_nums[count] = (irq_number)tmp; |
1274 | ++count; |
1275 | } |
1276 | |
1277 | closedir_safe(dir); |
1278 | |
1279 | // Sort IRQ numbers, so that RX and TX interrupts pair up nicely when |
1280 | // populating irqs->irqs[]. |
1281 | qsort(irq_nums, (size_t)count, sizeof(irq_number), comp_irq_number); |
1282 | |
1283 | char actions[count][100]; |
1284 | memset(actions, 0, sizeof(actions)); |
1285 | |
1286 | FILE *fh = fopen("/proc/interrupts" , "r" ); |
1287 | |
1288 | if (fh == NULL) { |
1289 | cf_crash(CF_HARDWARE, "error while opening /proc/interrupts" ); |
1290 | } |
1291 | |
1292 | int32_t line_no = 0; |
1293 | char line[25000]; |
1294 | |
1295 | while (fgets(line, sizeof(line), fh) != NULL) { |
1296 | ++line_no; |
1297 | |
1298 | if (line_no == 1) { |
1299 | continue; |
1300 | } |
1301 | |
1302 | int32_t i = 0; |
1303 | |
1304 | while (line[i] == ' ') { |
1305 | ++i; |
1306 | } |
1307 | |
1308 | irq_number irq_num = 0; |
1309 | |
1310 | while (line[i] >= '0' && line[i] <= '9') { |
1311 | irq_num = (irq_number)(irq_num * 10 + line[i] - '0'); |
1312 | ++i; |
1313 | } |
1314 | |
1315 | if (line[i] != ':') { |
1316 | continue; |
1317 | } |
1318 | |
1319 | while (line[i] != 0 && line[i] != '\n') { |
1320 | ++i; |
1321 | } |
1322 | |
1323 | line[i] = 0; |
1324 | |
1325 | while (i >= 0 && line[i] != ' ') { |
1326 | --i; |
1327 | } |
1328 | |
1329 | char *action = line + i + 1; |
1330 | |
1331 | if (strlen(action) >= sizeof(actions[0])) { |
1332 | cf_crash(CF_HARDWARE, "oversize action in line %d in /proc/interrupts: %s" , |
1333 | line_no, action); |
1334 | } |
1335 | |
1336 | cf_detail(CF_HARDWARE, "IRQ %hu has action %s" , irq_num, action); |
1337 | |
1338 | for (i = 0; i < count; ++i) { |
1339 | if (irq_nums[i] == irq_num) { |
1340 | int32_t m = 0; |
1341 | |
1342 | // Remove any digits, so that the queue index goes away and all queues |
1343 | // look alike. Also, normalize to lower case. For example: |
1344 | // |
1345 | // "i40e-em1-TxRx-0" -> "ie-em-txrx-" |
1346 | // "i40e-em1-TxRx-1" -> "ie-em-txrx-" |
1347 | // ... |
1348 | |
1349 | for (int32_t k = 0; action[k] != 0; ++k) { |
1350 | if (action[k] < '0' || action[k] > '9') { |
1351 | actions[i][m] = (char)tolower((uint8_t)action[k]); |
1352 | ++m; |
1353 | } |
1354 | } |
1355 | |
1356 | actions[i][m] = 0; |
1357 | cf_detail(CF_HARDWARE, "action pattern is %s" , actions[i]); |
1358 | break; |
1359 | } |
1360 | } |
1361 | } |
1362 | |
1363 | fclose(fh); |
1364 | |
1365 | int32_t n_groups = 0; |
1366 | int32_t group_sizes[count]; |
1367 | int32_t [count]; |
1368 | int32_t action_groups[count]; |
1369 | int32_t inactive_group = -1; |
1370 | |
1371 | for (int32_t i = 0; i < count; ++i) { |
1372 | group_sizes[i] = 0; |
1373 | group_extra[i] = 0; |
1374 | action_groups[i] = -1; |
1375 | } |
1376 | |
1377 | // Group by action pattern. |
1378 | |
1379 | for (int32_t i = 0; i < count; ++i) { |
1380 | if (action_groups[i] >= 0) { |
1381 | continue; |
1382 | } |
1383 | |
1384 | action_groups[i] = n_groups; |
1385 | ++group_sizes[n_groups]; |
1386 | |
1387 | if (actions[i][0] == 0) { |
1388 | inactive_group = n_groups; |
1389 | cf_detail(CF_HARDWARE, "inactive IRQs in new group %d" , n_groups); |
1390 | } |
1391 | else { |
1392 | cf_detail(CF_HARDWARE, "new group %d: %s" , n_groups, actions[i]); |
1393 | } |
1394 | |
1395 | for (int32_t k = i + 1; k < count; ++k) { |
1396 | if (strcmp(actions[i], actions[k]) == 0) { |
1397 | action_groups[k] = n_groups; |
1398 | ++group_sizes[n_groups]; |
1399 | } |
1400 | } |
1401 | |
1402 | cf_detail(CF_HARDWARE, "group %d has %d member(s)" , n_groups, group_sizes[n_groups]); |
1403 | |
1404 | // Prefer groups whose action patterns have "rx", "tx", "input", or "output" in them. |
1405 | |
1406 | if (strstr(actions[i], "rx" ) != NULL || strstr(actions[i], "tx" ) != NULL || |
1407 | strstr(actions[i], "input" ) != NULL || strstr(actions[i], "output" ) != NULL) { |
1408 | cf_detail(CF_HARDWARE, "preferring group %d" , n_groups); |
1409 | group_extra[n_groups] = 1; |
1410 | } |
1411 | |
1412 | ++n_groups; |
1413 | } |
1414 | |
1415 | // Find the two largest groups. |
1416 | |
1417 | int32_t a = -1; |
1418 | int32_t b = -1; |
1419 | |
1420 | for (int32_t i = 0; i < n_groups; ++i) { |
1421 | if (i != inactive_group && |
1422 | (a < 0 || group_sizes[i] + group_extra[i] > group_sizes[a] + group_extra[a])) { |
1423 | a = i; |
1424 | } |
1425 | } |
1426 | |
1427 | if (a < 0) { |
1428 | cf_crash(CF_HARDWARE, "no active interrupts for interface %s" , if_name); |
1429 | } |
1430 | |
1431 | for (int32_t i = 0; i < n_groups; ++i) { |
1432 | if (i != inactive_group && i != a && |
1433 | (b < 0 || group_sizes[i] + group_extra[i] > group_sizes[b] + group_extra[b])) { |
1434 | b = i; |
1435 | } |
1436 | } |
1437 | |
1438 | cf_detail(CF_HARDWARE, "largest groups: %d, %d" , a, b); |
1439 | |
1440 | // If the two largest groups have an equal number of members, then we assume |
1441 | // that it's a NIC with separate RX and TX queue IRQs. |
1442 | |
1443 | if (b >= 0 && group_sizes[a] == group_sizes[b]) { |
1444 | cf_detail(CF_HARDWARE, "assuming %d separate RX and TX queue IRQ(s)" , |
1445 | group_sizes[a] + group_sizes[b]); |
1446 | int32_t ia = 0; |
1447 | int32_t ib = 0; |
1448 | |
1449 | // Make RX and TX queue IRQs take turns in the IRQ list. |
1450 | |
1451 | for (int32_t k = 0; k < count; ++k) { |
1452 | if (action_groups[k] == a) { |
1453 | irqs->irqs[ia * 2] = irq_nums[k]; |
1454 | cf_detail(CF_HARDWARE, "irqs[%d] = %hu" , ia * 2, irq_nums[k]); |
1455 | ++ia; |
1456 | } |
1457 | else if (action_groups[k] == b) { |
1458 | irqs->irqs[ib * 2 + 1] = irq_nums[k]; |
1459 | cf_detail(CF_HARDWARE, "irqs[%d] = %hu" , ib * 2 + 1, irq_nums[k]); |
1460 | ++ib; |
1461 | } |
1462 | } |
1463 | |
1464 | irqs->n_irqs = (uint16_t)(group_sizes[a] + group_sizes[b]); |
1465 | |
1466 | // Send pairs of two consecutive IRQs in the IRQ list (= the RX and the |
1467 | // TX queue IRQ of a given NIC queue pair) to the same CPU. |
1468 | |
1469 | irqs->per_cpu = 2; |
1470 | return; |
1471 | } |
1472 | |
1473 | // Otherwise, we assume that it's a NIC with combined RX and TX queue IRQs |
1474 | // and that the largest group contains these IRQs. |
1475 | |
1476 | cf_detail(CF_HARDWARE, "assuming %d combined RX and TX queue IRQ(s)" , group_sizes[a]); |
1477 | int32_t ia = 0; |
1478 | |
1479 | for (int32_t k = 0; k < count; ++k) { |
1480 | if (action_groups[k] == a) { |
1481 | irqs->irqs[ia] = irq_nums[k]; |
1482 | cf_detail(CF_HARDWARE, "irqs[%d] = %hu" , ia, irq_nums[k]); |
1483 | ++ia; |
1484 | } |
1485 | } |
1486 | |
1487 | irqs->n_irqs = (uint16_t)group_sizes[a]; |
1488 | |
1489 | // Send each IRQ in the IRQ list to a different CPU. |
1490 | |
1491 | irqs->per_cpu = 1; |
1492 | } |
1493 | |
1494 | static void |
1495 | pin_irq(irq_number i_irq, cf_topo_os_cpu_index i_os_cpu) |
1496 | { |
1497 | cf_detail(CF_HARDWARE, "pinning IRQ number %hu to OS CPU index %hu" , i_irq, i_os_cpu); |
1498 | |
1499 | cpu_set_t mask; |
1500 | CPU_ZERO(&mask); |
1501 | CPU_SET(i_os_cpu, &mask); |
1502 | |
1503 | char mask_str[200]; |
1504 | mask_to_string(&mask, mask_str, sizeof(mask_str)); |
1505 | cf_detail(CF_HARDWARE, "CPU mask is %s" , mask_str); |
1506 | |
1507 | char path[1000]; |
1508 | snprintf(path, sizeof(path), "/proc/irq/%hu/smp_affinity" , i_irq); |
1509 | |
1510 | if (write_file(path, mask_str, strlen(mask_str)) != FILE_RES_OK) { |
1511 | cf_crash(CF_HARDWARE, "error while pinning IRQ, path %s" , path); |
1512 | } |
1513 | } |
1514 | |
1515 | static cf_topo_os_cpu_index |
1516 | fix_os_cpu_index(cf_topo_os_cpu_index i_os_cpu, const cpu_set_t *online) |
1517 | { |
1518 | while (true) { |
1519 | if (i_os_cpu >= g_n_os_cpus) { |
1520 | i_os_cpu = 0; |
1521 | } |
1522 | |
1523 | if (CPU_ISSET(i_os_cpu, online)) { |
1524 | return i_os_cpu; |
1525 | } |
1526 | |
1527 | ++i_os_cpu; |
1528 | } |
1529 | } |
1530 | |
1531 | static void |
1532 | config_steering(const char *format, const char *if_name, uint16_t n_queues, bool enable) |
1533 | { |
1534 | uint16_t i_queue; |
1535 | cpu_set_t masks[n_queues]; |
1536 | |
1537 | for (i_queue = 0; i_queue < n_queues; ++i_queue) { |
1538 | CPU_ZERO(&masks[i_queue]); |
1539 | } |
1540 | |
1541 | if (enable) { |
1542 | i_queue = 0; |
1543 | |
1544 | for (cf_topo_os_cpu_index i_os_cpu = 0; i_os_cpu < g_n_os_cpus; ++i_os_cpu) { |
1545 | if (CPU_ISSET(i_os_cpu, &g_os_cpus_online)) { |
1546 | CPU_SET(i_os_cpu, &masks[i_queue % n_queues]); |
1547 | ++i_queue; |
1548 | } |
1549 | } |
1550 | } |
1551 | |
1552 | for (i_queue = 0; i_queue < n_queues; ++i_queue) { |
1553 | char path[1000]; |
1554 | snprintf(path, sizeof(path), format, if_name, i_queue); |
1555 | cf_detail(CF_HARDWARE, "path is %s" , path); |
1556 | |
1557 | char mask_str[200]; |
1558 | mask_to_string(&masks[i_queue], mask_str, sizeof(mask_str)); |
1559 | cf_detail(CF_HARDWARE, "CPU mask is %s" , mask_str); |
1560 | |
1561 | write_file_safe(path, mask_str, strlen(mask_str)); |
1562 | } |
1563 | } |
1564 | |
1565 | static void |
1566 | enable_xps(const char *if_name) |
1567 | { |
1568 | cf_detail(CF_HARDWARE, "enabling XPS for interface %s" , if_name); |
1569 | uint16_t n_queues = interface_tx_queues(if_name); |
1570 | config_steering("/sys/class/net/%s/queues/tx-%hu/xps_cpus" , if_name, n_queues, true); |
1571 | } |
1572 | |
1573 | static void |
1574 | disable_rps(const char *if_name) |
1575 | { |
1576 | cf_detail(CF_HARDWARE, "disabling RPS for interface %s" , if_name); |
1577 | uint16_t n_queues = interface_rx_queues(if_name); |
1578 | config_steering("/sys/class/net/%s/queues/rx-%hu/rps_cpus" , if_name, n_queues, false); |
1579 | } |
1580 | |
1581 | static void |
1582 | config_rfs(const char *if_name, bool enable) |
1583 | { |
1584 | cf_detail(CF_HARDWARE, "%s RFS for interface %s" , enable ? "enabling" : "disabling" , if_name); |
1585 | |
1586 | uint16_t n_queues = interface_rx_queues(if_name); |
1587 | uint32_t sz_glob = enable ? 1000000 : 0; |
1588 | uint32_t sz_queue = sz_glob / n_queues; |
1589 | |
1590 | cf_detail(CF_HARDWARE, "global size is %u, per-queue size is %u" , sz_glob, sz_queue); |
1591 | |
1592 | char string[50]; |
1593 | snprintf(string, sizeof(string), "%u" , sz_glob); |
1594 | write_file_safe("/proc/sys/net/core/rps_sock_flow_entries" , string, strlen(string)); |
1595 | |
1596 | snprintf(string, sizeof(string), "%u" , sz_queue); |
1597 | |
1598 | for (uint16_t i_queue = 0; i_queue < n_queues; ++i_queue) { |
1599 | char path[1000]; |
1600 | snprintf(path, sizeof(path), "/sys/class/net/%s/queues/rx-%hu/rps_flow_cnt" , |
1601 | if_name, i_queue); |
1602 | write_file_safe(path, string, strlen(string)); |
1603 | } |
1604 | } |
1605 | |
1606 | static void |
1607 | enable_coalescing(const char *if_name) |
1608 | { |
1609 | cf_detail(CF_HARDWARE, "enabling interrupt coalescing for interface %s" , if_name); |
1610 | int32_t sock = socket(AF_INET, SOCK_DGRAM, 0); |
1611 | |
1612 | if (sock < 0) { |
1613 | cf_crash(CF_HARDWARE, "error while create ethtool socket: %d (%s)" , errno, cf_strerror(errno)); |
1614 | } |
1615 | |
1616 | struct ifreq req; |
1617 | memset(&req, 0, sizeof(req)); |
1618 | |
1619 | if (strlen(if_name) > IFNAMSIZ - 1) { |
1620 | cf_crash(CF_HARDWARE, "invalid interface name %s" , if_name); |
1621 | } |
1622 | |
1623 | strcpy(req.ifr_name, if_name); |
1624 | struct ethtool_coalesce coal = { .cmd = ETHTOOL_GCOALESCE }; |
1625 | req.ifr_data = &coal; |
1626 | |
1627 | if (ioctl(sock, SIOCETHTOOL, &req) < 0) { |
1628 | if (errno == EOPNOTSUPP) { |
1629 | cf_detail(CF_HARDWARE, "interface %s does not support ETHTOOL_GCOALESCE" , if_name); |
1630 | goto cleanup1; |
1631 | } |
1632 | |
1633 | cf_crash(CF_HARDWARE, "error while getting interface settings: %d (%s)" , |
1634 | errno, cf_strerror(errno)); |
1635 | } |
1636 | |
1637 | cf_detail(CF_HARDWARE, "current interface settings: adaptive = %u, usecs = %u" , |
1638 | coal.use_adaptive_rx_coalesce, coal.rx_coalesce_usecs); |
1639 | |
1640 | if (coal.use_adaptive_rx_coalesce != 0 || coal.rx_coalesce_usecs >= 100) { |
1641 | cf_detail(CF_HARDWARE, "leaving interface settings untouched" ); |
1642 | goto cleanup1; |
1643 | } |
1644 | |
1645 | cf_detail(CF_HARDWARE, "adjusting interface settings" ); |
1646 | coal = (struct ethtool_coalesce){ |
1647 | .cmd = ETHTOOL_SCOALESCE, |
1648 | .rx_coalesce_usecs = 100 // .1 ms for now, which adds .05 ms to a request on average. |
1649 | }; |
1650 | |
1651 | if (ioctl(sock, SIOCETHTOOL, &req) < 0) { |
1652 | if (errno == EOPNOTSUPP) { |
1653 | cf_detail(CF_HARDWARE, "interface %s does not support ETHTOOL_SCOALESCE" , if_name); |
1654 | goto cleanup1; |
1655 | } |
1656 | |
1657 | cf_crash(CF_HARDWARE, "error while adjusting interface settings: %d (%s)" , |
1658 | errno, cf_strerror(errno)); |
1659 | } |
1660 | |
1661 | cleanup1: |
1662 | CF_NEVER_FAILS(close(sock)); |
1663 | } |
1664 | |
1665 | static void |
1666 | check_irqbalance(void) |
1667 | { |
1668 | cf_detail(CF_HARDWARE, "checking irqbalance" ); |
1669 | |
1670 | check_proc_res res = check_proc("irqbalance" , 1, (const char *[]){ |
1671 | "--policyscript=" POLICY_SCRIPT |
1672 | }); |
1673 | |
1674 | if (res == CHECK_PROC_PRESENT_NO_ARG) { |
1675 | res = check_proc("irqbalance" , 2, (const char *[]){ |
1676 | "--policyscript" , |
1677 | POLICY_SCRIPT |
1678 | }); |
1679 | } |
1680 | |
1681 | if (res == CHECK_PROC_PRESENT_NO_ARG) { |
1682 | res = check_proc("irqbalance" , 1, (const char *[]){ |
1683 | "-l" POLICY_SCRIPT |
1684 | }); |
1685 | } |
1686 | |
1687 | if (res == CHECK_PROC_PRESENT_NO_ARG) { |
1688 | res = check_proc("irqbalance" , 2, (const char *[]){ |
1689 | "-l" , |
1690 | POLICY_SCRIPT |
1691 | }); |
1692 | } |
1693 | |
1694 | if (res == CHECK_PROC_PRESENT_NO_ARG) { |
1695 | cf_crash_nostack(CF_HARDWARE, "please disable irqbalance or run it with the Aerospike policy script, /etc/aerospike/irqbalance-ban.sh" ); |
1696 | } |
1697 | } |
1698 | |
1699 | static void |
1700 | config_interface(const char *if_name, bool rfs, irq_list *irqs) |
1701 | { |
1702 | uint16_t n_irq_cpus = 0; |
1703 | cf_topo_os_cpu_index i_os_cpu = fix_os_cpu_index(0, &g_os_cpus_online); |
1704 | |
1705 | for (uint16_t i = 0; i < irqs->n_irqs; ++i) { |
1706 | pin_irq(irqs->irqs[i], i_os_cpu); |
1707 | |
1708 | if (i % irqs->per_cpu == irqs->per_cpu - 1) { |
1709 | ++n_irq_cpus; |
1710 | i_os_cpu = fix_os_cpu_index((cf_topo_os_cpu_index)(i_os_cpu + 1), &g_os_cpus_online); |
1711 | } |
1712 | } |
1713 | |
1714 | cf_detail(CF_HARDWARE, "interface %s with %hu RX interrupt(s)" , if_name, n_irq_cpus); |
1715 | |
1716 | if (g_n_irq_cpus == 0) { |
1717 | g_n_irq_cpus = n_irq_cpus; |
1718 | } |
1719 | else if (n_irq_cpus != g_n_irq_cpus) { |
1720 | cf_crash(CF_HARDWARE, "interface %s with inconsistent number of RX interrupts: %hu vs. %hu" , |
1721 | if_name, n_irq_cpus, g_n_irq_cpus); |
1722 | } |
1723 | |
1724 | disable_rps(if_name); |
1725 | config_rfs(if_name, rfs); |
1726 | enable_xps(if_name); |
1727 | |
1728 | // Redistributing packets with RFS causes inter-CPU interrupts, which increases |
1729 | // the interrupt load on the machine. For low-end systems, make sure that |
1730 | // interrupt coalescing is enabled. |
1731 | // |
1732 | // We consider a machine low-end, if we handle interrupts on 25% or less of the |
1733 | // available CPUs (i.e., if the number of NIC queues is 25% or less of the number |
1734 | // of available CPUs) and it has fewer than 4 NIC queues. |
1735 | // |
1736 | // Better (i.e., NUMA) machines typically come with adaptive interrupt coalescing |
1737 | // enabled by default. That's why we only do this here and not in the NUMA case. |
1738 | |
1739 | if (rfs && n_irq_cpus <= g_n_cpus / 4 && n_irq_cpus < 4) { |
1740 | enable_coalescing(if_name); |
1741 | } |
1742 | } |
1743 | |
1744 | static void |
1745 | config_interface_numa(const char *if_name, irq_list *irqs) |
1746 | { |
1747 | uint16_t n_irq_cpus = 0; |
1748 | cf_topo_os_cpu_index i_os_cpu[g_n_numa_nodes]; |
1749 | uint16_t i_numa_node; |
1750 | |
1751 | for (i_numa_node = 0; i_numa_node < g_n_numa_nodes; ++i_numa_node) { |
1752 | i_os_cpu[i_numa_node] = fix_os_cpu_index(0, &g_numa_node_os_cpus_online[i_numa_node]); |
1753 | } |
1754 | |
1755 | i_numa_node = 0; |
1756 | |
1757 | // This configures the IRQs for all NUMA nodes. If multiple asd processes are |
1758 | // running, each process does this, but each does it identically. Hence there |
1759 | // isn't any conflict. |
1760 | |
1761 | for (uint16_t i = 0; i < irqs->n_irqs; ++i) { |
1762 | char mask_str[200]; |
1763 | mask_to_string(&g_numa_node_os_cpus_online[i_numa_node], mask_str, sizeof(mask_str)); |
1764 | cf_detail(CF_HARDWARE, "NUMA node index %hu CPU mask is %s" , i_numa_node, mask_str); |
1765 | |
1766 | pin_irq(irqs->irqs[i], i_os_cpu[i_numa_node]); |
1767 | |
1768 | if (i % irqs->per_cpu == irqs->per_cpu - 1) { |
1769 | // Only count CPUs on our NUMA node. |
1770 | |
1771 | if (i_numa_node == g_i_numa_node) { |
1772 | ++n_irq_cpus; |
1773 | } |
1774 | |
1775 | i_os_cpu[i_numa_node] = |
1776 | fix_os_cpu_index((cf_topo_os_cpu_index)(i_os_cpu[i_numa_node] + 1), |
1777 | &g_numa_node_os_cpus_online[i_numa_node]); |
1778 | i_numa_node = (uint16_t)((i_numa_node + 1) % g_n_numa_nodes); |
1779 | } |
1780 | } |
1781 | |
1782 | cf_detail(CF_HARDWARE, "interface %s with %hu RX interrupt(s) on NUMA node %hu" , |
1783 | if_name, n_irq_cpus, g_i_numa_node); |
1784 | |
1785 | if (g_n_irq_cpus == 0) { |
1786 | g_n_irq_cpus = n_irq_cpus; |
1787 | } |
1788 | else if (n_irq_cpus != g_n_irq_cpus) { |
1789 | cf_crash(CF_HARDWARE, "interface %s with inconsistent number of RX interrupts: %hu vs. %hu" , |
1790 | if_name, n_irq_cpus, g_n_irq_cpus); |
1791 | } |
1792 | |
1793 | disable_rps(if_name); |
1794 | config_rfs(if_name, true); |
1795 | enable_xps(if_name); |
1796 | } |
1797 | |
1798 | static void |
1799 | optimize_interface(const char *if_name) |
1800 | { |
1801 | cf_detail(CF_HARDWARE, "optimizing interface %s" , if_name); |
1802 | uint16_t n_queues = interface_rx_queues(if_name); |
1803 | irq_list irqs; |
1804 | interface_irqs(if_name, &irqs); |
1805 | |
1806 | cf_info(CF_HARDWARE, "detected %hu NIC receive queue(s), %hu interrupt(s) for %s" , |
1807 | n_queues, irqs.n_irqs, if_name); |
1808 | |
1809 | // We either expect one interrupt per RX queue (shared with TX) or two |
1810 | // interrupts per RX queue (one RX, one TX). |
1811 | |
1812 | uint16_t n_irq_cpus = irqs.n_irqs / irqs.per_cpu; |
1813 | |
1814 | if (n_irq_cpus != n_queues) { |
1815 | cf_crash(CF_HARDWARE, "suspicious NIC interrupt count %hu with %hu NIC receive queue(s)" , |
1816 | irqs.n_irqs, n_queues); |
1817 | } |
1818 | |
1819 | if (n_irq_cpus == g_n_cpus) { |
1820 | if (g_i_numa_node != INVALID_INDEX) { |
1821 | cf_detail(CF_HARDWARE, "setting up for a fancy interface with NUMA" ); |
1822 | config_interface_numa(if_name, &irqs); |
1823 | } |
1824 | else { |
1825 | cf_detail(CF_HARDWARE, "setting up for a fancy interface, no NUMA" ); |
1826 | config_interface(if_name, false, &irqs); |
1827 | } |
1828 | } |
1829 | else { |
1830 | if (n_irq_cpus <= g_n_cpus / 4) { |
1831 | cf_warning(CF_HARDWARE, "%s has very few NIC queues; only %hu out of %hu CPUs handle(s) NIC interrupts" , |
1832 | if_name, n_irq_cpus, g_n_cpus); |
1833 | } |
1834 | |
1835 | if (g_i_numa_node != INVALID_INDEX) { |
1836 | cf_detail(CF_HARDWARE, "setting up for a lame interface with NUMA" ); |
1837 | config_interface_numa(if_name, &irqs); |
1838 | } |
1839 | else { |
1840 | cf_detail(CF_HARDWARE, "setting up for a lame interface, no NUMA" ); |
1841 | config_interface(if_name, true, &irqs); |
1842 | } |
1843 | } |
1844 | } |
1845 | |
1846 | // Make sure that we are running on appropriate kernel. |
1847 | static void |
1848 | check_socket_option(int optname, const char *tag) |
1849 | { |
1850 | int32_t fd = socket(AF_INET, SOCK_DGRAM, IPPROTO_UDP); |
1851 | |
1852 | if (fd < 0) { |
1853 | cf_crash(CF_HARDWARE, "error while creating UDP test socket: %d (%s)" , |
1854 | errno, cf_strerror(errno)); |
1855 | } |
1856 | |
1857 | int32_t val; |
1858 | socklen_t val_len = sizeof(val); |
1859 | |
1860 | if (getsockopt(fd, SOL_SOCKET, optname, &val, &val_len) < 0) { |
1861 | if (errno == ENOPROTOOPT) { |
1862 | cf_crash_nostack(CF_HARDWARE, "auto-pin requires %s or later" , tag); |
1863 | } |
1864 | |
1865 | cf_crash(CF_HARDWARE, "error while testing for socket option: %d (%s)" , |
1866 | errno, cf_strerror(errno)); |
1867 | } |
1868 | |
1869 | CF_NEVER_FAILS(close(fd)); |
1870 | } |
1871 | |
1872 | // Reconfigure NIC queues and interrupts. |
1873 | static void |
1874 | optimize_interfaces(const cf_addr_list *addrs) |
1875 | { |
1876 | if (addrs->n_addrs == 0) { |
1877 | cf_crash_nostack(CF_HARDWARE, "auto-pinning requires binding the service to one or more network interfaces" ); |
1878 | } |
1879 | |
1880 | for (uint32_t i = 0; i < addrs->n_addrs; ++i) { |
1881 | const char *if_name = addrs->addrs[i]; |
1882 | |
1883 | if (!cf_inter_is_inter_name(if_name)) { |
1884 | cf_crash_nostack(CF_HARDWARE, "auto-pinning requires binding the service to network interfaces; \"%s\" isn't a network interface" , |
1885 | if_name); |
1886 | } |
1887 | |
1888 | char phys_name[50]; |
1889 | CF_NEVER_FAILS(cf_inter_get_physical(if_name, phys_name, sizeof(phys_name))); |
1890 | |
1891 | char *exp_names[100]; |
1892 | uint32_t n_exp = sizeof(exp_names) / sizeof(exp_names[0]); |
1893 | cf_inter_expand_bond(phys_name, exp_names, &n_exp); |
1894 | |
1895 | for (uint32_t k = 0; k < n_exp; ++k) { |
1896 | optimize_interface(exp_names[k]); |
1897 | cf_free(exp_names[k]); |
1898 | } |
1899 | } |
1900 | } |
1901 | |
1902 | void |
1903 | cf_topo_config(cf_topo_auto_pin auto_pin, cf_topo_numa_node_index a_numa_node, |
1904 | const cf_addr_list *addrs) |
1905 | { |
1906 | // Detect the NUMA topology. |
1907 | |
1908 | switch (auto_pin) { |
1909 | case CF_TOPO_AUTO_PIN_NONE: |
1910 | case CF_TOPO_AUTO_PIN_CPU: |
1911 | detect(INVALID_INDEX); |
1912 | break; |
1913 | |
1914 | case CF_TOPO_AUTO_PIN_NUMA: |
1915 | case CF_TOPO_AUTO_PIN_ADQ: |
1916 | detect(a_numa_node); |
1917 | |
1918 | // Clamp the given NUMA node index to the valid range. We can only do this |
1919 | // after we know what g_n_numa_nodes is, which is initialized by the above |
1920 | // call to detect(). |
1921 | |
1922 | if (a_numa_node >= g_n_numa_nodes) { |
1923 | cf_topo_numa_node_index orig = a_numa_node; |
1924 | a_numa_node = (cf_topo_numa_node_index)(a_numa_node % g_n_numa_nodes); |
1925 | cf_detail(CF_HARDWARE, "invalid NUMA node index: %hu, clamping to %hu" , orig, a_numa_node); |
1926 | detect(a_numa_node); |
1927 | } |
1928 | |
1929 | break; |
1930 | |
1931 | default: |
1932 | cf_crash(CF_HARDWARE, "bad auto-pin value %d" , auto_pin); |
1933 | break; |
1934 | } |
1935 | |
1936 | // If we don't do any pinning, then we're done after NUMA topology detection. |
1937 | if (auto_pin == CF_TOPO_AUTO_PIN_NONE) { |
1938 | return; |
1939 | } |
1940 | |
1941 | check_irqbalance(); // ensure irqbalance is disabled |
1942 | |
1943 | switch (auto_pin) { |
1944 | case CF_TOPO_AUTO_PIN_CPU: |
1945 | check_socket_option(SO_INCOMING_CPU, "Linux kernel 3.19" ); |
1946 | optimize_interfaces(addrs); |
1947 | break; |
1948 | case CF_TOPO_AUTO_PIN_NUMA: |
1949 | check_socket_option(SO_INCOMING_CPU, "Linux kernel 3.19" ); |
1950 | optimize_interfaces(addrs); |
1951 | pin_to_numa_node(a_numa_node); |
1952 | break; |
1953 | case CF_TOPO_AUTO_PIN_ADQ: |
1954 | check_socket_option(SO_INCOMING_NAPI_ID, "Linux kernel 4.12" ); |
1955 | pin_to_numa_node(a_numa_node); |
1956 | break; |
1957 | default: |
1958 | cf_crash(CF_HARDWARE, "bad auto-pin value %d" , auto_pin); |
1959 | break; |
1960 | } |
1961 | } |
1962 | |
1963 | void |
1964 | cf_topo_force_map_memory(const uint8_t *from, size_t size) |
1965 | { |
1966 | if (g_i_numa_node == INVALID_INDEX || size == 0) { |
1967 | return; |
1968 | } |
1969 | |
1970 | cf_assert(from, CF_HARDWARE, "invalid cf_topo_force_map_memory() call" ); |
1971 | |
1972 | // Read one byte per memory page to force otherwise lazy mapping. |
1973 | |
1974 | const uint8_t *start = (const uint8_t *) |
1975 | (((int64_t)from + (MEM_PAGE_SIZE - 1)) & -MEM_PAGE_SIZE); |
1976 | const uint8_t *end = from + size; |
1977 | const volatile uint8_t *p_byte; |
1978 | |
1979 | // In case 'from' was not page-aligned, take care of the partial page. |
1980 | if (start > from) { |
1981 | p_byte = from; |
1982 | p_byte[0]; |
1983 | } |
1984 | |
1985 | for (p_byte = start; p_byte < end; p_byte += MEM_PAGE_SIZE) { |
1986 | p_byte[0]; |
1987 | } |
1988 | } |
1989 | |
1990 | void |
1991 | cf_topo_migrate_memory(void) |
1992 | { |
1993 | if (g_i_numa_node == INVALID_INDEX) { |
1994 | return; |
1995 | } |
1996 | |
1997 | // Migrate existing memory allocations to the selected NUMA node. |
1998 | |
1999 | os_numa_node_index i_os_numa_node = g_numa_node_index_to_os_numa_node_index[g_i_numa_node]; |
2000 | uint64_t to_mask = 1UL << i_os_numa_node; |
2001 | cf_detail(CF_HARDWARE, "NUMA node mask (to): %016" PRIx64, to_mask); |
2002 | |
2003 | uint64_t from_mask = 0; |
2004 | |
2005 | for (cf_topo_numa_node_index i_numa_node = 0; i_numa_node < g_n_numa_nodes; ++i_numa_node) { |
2006 | i_os_numa_node = g_numa_node_index_to_os_numa_node_index[i_numa_node]; |
2007 | from_mask |= 1u << i_os_numa_node; |
2008 | } |
2009 | |
2010 | from_mask &= ~to_mask; |
2011 | cf_detail(CF_HARDWARE, "NUMA node mask (from): %016" PRIx64, from_mask); |
2012 | |
2013 | if (from_mask != 0) { |
2014 | cf_info(CF_HARDWARE, "migrating shared memory to local NUMA node - this may take a bit" ); |
2015 | // Unlike select(), we have to pass "number of valid bits + 1". |
2016 | migrate_pages_safe(0, 65, &from_mask, &to_mask); |
2017 | } |
2018 | } |
2019 | |
2020 | void |
2021 | cf_topo_info(void) |
2022 | { |
2023 | if (g_i_numa_node == INVALID_INDEX) { |
2024 | cf_info(CF_HARDWARE, "detected %hu CPU(s), %hu core(s), %hu NUMA node(s)" , |
2025 | g_n_cpus, g_n_cores, g_n_numa_nodes); |
2026 | } |
2027 | else { |
2028 | cf_info(CF_HARDWARE, "detected %hu CPU(s), %hu core(s) on NUMA node %hu of %hu" , |
2029 | g_n_cpus, g_n_cores, g_i_numa_node, g_n_numa_nodes); |
2030 | } |
2031 | } |
2032 | |
2033 | static uint32_t |
2034 | dev_key_hash(const void *k) |
2035 | { |
2036 | const dev_key_t *key = k; |
2037 | return (1 + key->major) * (1 + key->minor); |
2038 | } |
2039 | |
2040 | static void |
2041 | add_child(const dev_key_t *key, dev_node_t *node, const dev_key_t *child_key, |
2042 | dev_node_t *child_node) |
2043 | { |
2044 | cf_detail(CF_HARDWARE, "parent %u:%u -> child %u:%u" , |
2045 | key->major, key->minor, child_key->major, child_key->minor); |
2046 | |
2047 | node->children[node->n_children] = child_node; |
2048 | ++node->n_children; |
2049 | } |
2050 | |
2051 | static void |
2052 | collect_edges(const char *sys_dir, const char *prefix, bool flip, |
2053 | const dev_key_t *key, dev_node_t *node) |
2054 | { |
2055 | cf_detail(CF_HARDWARE, "collecting devices in %s" , sys_dir); |
2056 | |
2057 | if (!path_exists(sys_dir)) { |
2058 | return; |
2059 | } |
2060 | |
2061 | size_t prefix_len = strlen(prefix); |
2062 | |
2063 | DIR *dir = opendir_safe(sys_dir); |
2064 | struct dirent ent; |
2065 | |
2066 | while (readdir_safe(dir, &ent) >= 0) { |
2067 | cf_detail(CF_HARDWARE, "considering %s" , ent.d_name); |
2068 | |
2069 | if (prefix_len > 0 && strncmp(ent.d_name, prefix, prefix_len) != 0) { |
2070 | cf_detail(CF_HARDWARE, "prefix mismatch" ); |
2071 | continue; |
2072 | } |
2073 | |
2074 | char sys_path[DEVICE_PATH_SIZE]; |
2075 | snprintf(sys_path, DEVICE_PATH_SIZE, "%s/%s" , sys_dir, ent.d_name); |
2076 | |
2077 | if (!path_is_dir(sys_path)) { |
2078 | cf_detail(CF_HARDWARE, "not a directory" ); |
2079 | continue; |
2080 | } |
2081 | |
2082 | snprintf(sys_path, DEVICE_PATH_SIZE, "%s/%s/dev" , sys_dir, ent.d_name); |
2083 | |
2084 | dev_key_t sub_key; |
2085 | |
2086 | if (read_device_numbers(sys_path, &sub_key.major, &sub_key.minor) != |
2087 | FILE_RES_OK) { |
2088 | cf_detail(CF_HARDWARE, "no device numbers" ); |
2089 | continue; |
2090 | } |
2091 | |
2092 | dev_node_t *sub_node; |
2093 | |
2094 | if (cf_shash_get(g_dev_graph, &sub_key, &sub_node) != CF_SHASH_OK) { |
2095 | cf_warning(CF_HARDWARE, "no node for sub device %s/%s (%u:%u)" , |
2096 | sys_dir, ent.d_name, sub_key.major, sub_key.minor); |
2097 | continue; |
2098 | } |
2099 | |
2100 | if (!flip) { |
2101 | add_child(&sub_key, sub_node, key, node); |
2102 | } |
2103 | else { |
2104 | add_child(key, node, &sub_key, sub_node); |
2105 | } |
2106 | } |
2107 | |
2108 | closedir_safe(dir); |
2109 | } |
2110 | |
2111 | static int32_t |
2112 | create_device_edges(const void *k, void *v, void *udata) |
2113 | { |
2114 | (void)udata; |
2115 | |
2116 | const dev_key_t *key = k; |
2117 | dev_node_t **node = v; |
2118 | |
2119 | cf_detail(CF_HARDWARE, "creating edges for %s" , (*node)->sys_home); |
2120 | |
2121 | // Collect partitions on a device. |
2122 | collect_edges((*node)->sys_home, (*node)->name, false, key, *node); |
2123 | |
2124 | char sys_slaves[DEVICE_PATH_SIZE + 7]; // +7 to silence the compiler |
2125 | snprintf(sys_slaves, DEVICE_PATH_SIZE + 7, "%s/slaves" , (*node)->sys_home); |
2126 | |
2127 | // Collect inter-device dependencies. |
2128 | collect_edges(sys_slaves, "" , true, key, *node); |
2129 | |
2130 | return CF_SHASH_OK; |
2131 | } |
2132 | |
2133 | static void |
2134 | build_device_graph(void) |
2135 | { |
2136 | // Step 1. Create a device map entry for each device. Don't yet link them |
2137 | // into a device dependency graph. |
2138 | |
2139 | static const char *sys_dirs[] = { |
2140 | "/sys/class/nvme" , |
2141 | "/sys/class/block" , |
2142 | NULL |
2143 | }; |
2144 | |
2145 | g_dev_graph = cf_shash_create(dev_key_hash, sizeof(dev_key_t), |
2146 | sizeof(dev_node_t *), 256, 0); |
2147 | |
2148 | for (int32_t i_dir = 0; sys_dirs[i_dir] != NULL; ++i_dir) { |
2149 | const char *sys_dir = sys_dirs[i_dir]; |
2150 | |
2151 | cf_detail(CF_HARDWARE, "collecting devices in %s" , sys_dir); |
2152 | |
2153 | if (!path_exists(sys_dir)) { |
2154 | cf_detail(CF_HARDWARE, "directory does not exist" ); |
2155 | continue; |
2156 | } |
2157 | |
2158 | DIR *dir = opendir_safe(sys_dir); |
2159 | struct dirent ent; |
2160 | |
2161 | while (readdir_safe(dir, &ent) >= 0) { |
2162 | cf_detail(CF_HARDWARE, "considering %s" , ent.d_name); |
2163 | |
2164 | char sys_path[DEVICE_PATH_SIZE]; |
2165 | snprintf(sys_path, DEVICE_PATH_SIZE, "%s/%s/dev" , sys_dir, |
2166 | ent.d_name); |
2167 | |
2168 | dev_key_t key; |
2169 | |
2170 | if (read_device_numbers(sys_path, &key.major, &key.minor) != |
2171 | FILE_RES_OK) { |
2172 | cf_detail(CF_HARDWARE, "no device numbers" ); |
2173 | continue; |
2174 | } |
2175 | |
2176 | dev_node_t *node = cf_malloc(sizeof(dev_node_t)); |
2177 | memset(node, 0, sizeof(dev_node_t)); |
2178 | |
2179 | snprintf(node->name, DEVICE_NAME_SIZE, "%s" , ent.d_name); |
2180 | snprintf(node->dev_path, DEVICE_PATH_SIZE, "/dev/%s" , ent.d_name); |
2181 | |
2182 | snprintf(node->sys_home, DEVICE_PATH_SIZE, "%s/%s" , sys_dir, |
2183 | ent.d_name); |
2184 | |
2185 | snprintf(sys_path, DEVICE_PATH_SIZE, "%s/%s/queue/scheduler" , |
2186 | sys_dir, ent.d_name); |
2187 | |
2188 | if (path_exists(sys_path)) { |
2189 | strcpy(node->sys_sched, sys_path); |
2190 | } |
2191 | |
2192 | cf_detail(CF_HARDWARE, "new device %s (%u:%u), home %s, " |
2193 | "scheduler %s" , node->dev_path, key.major, key.minor, |
2194 | node->sys_home, node->sys_sched[0] != 0 ? |
2195 | node->sys_sched : "-" ); |
2196 | |
2197 | if (cf_shash_put_unique(g_dev_graph, &key, &node) != CF_SHASH_OK) { |
2198 | cf_warning(CF_HARDWARE, "duplicate device %s (%u:%u)" , |
2199 | node->dev_path, key.major, key.minor); |
2200 | } |
2201 | } |
2202 | |
2203 | closedir_safe(dir); |
2204 | } |
2205 | |
2206 | // Step 2. Link the devices in the device map to create the device |
2207 | // dependency graph. Here's an example graph path for logical volume |
2208 | // lv_foo on encrypted partition sda3: |
2209 | // |
2210 | // lv_foo 253:1 -> sda3_crypt 253:0 -> sda3 8:3 -> sda 8:0 |
2211 | // |
2212 | // In short: Going from parents to children takes you closer to |
2213 | // physical devices. |
2214 | // |
2215 | // Devices can have multiple parents, e.g., sda could have sda1, sda2, |
2216 | // and sda3. |
2217 | // |
2218 | // Devices can also have multiple children, e.g., lv_bar could have |
2219 | // children sda1 and sdb1. |
2220 | |
2221 | cf_detail(CF_HARDWARE, "creating device edges" ); |
2222 | cf_shash_reduce(g_dev_graph, create_device_edges, NULL); |
2223 | } |
2224 | |
2225 | static char * |
2226 | get_mounted_device(const char *fs_path) |
2227 | { |
2228 | cf_detail(CF_HARDWARE, "mapping mount point %s" , fs_path); |
2229 | |
2230 | char *fs_real = realpath(fs_path, NULL); |
2231 | |
2232 | if (fs_real == NULL) { |
2233 | cf_warning(CF_HARDWARE, "failed to resolve path %s: %d (%s)" , |
2234 | fs_path, errno, cf_strerror(errno)); |
2235 | return NULL; |
2236 | } |
2237 | |
2238 | cf_detail(CF_HARDWARE, "resolved path %s" , fs_real); |
2239 | |
2240 | FILE *fh = setmntent("/proc/mounts" , "r" ); |
2241 | |
2242 | struct mntent mnt; |
2243 | char buff[1000]; |
2244 | |
2245 | size_t best_len = 0; |
2246 | char best_path[DEVICE_PATH_SIZE]; |
2247 | |
2248 | while (getmntent_r(fh, &mnt, buff, sizeof(buff)) != NULL) { |
2249 | cf_detail(CF_HARDWARE, "mount point %s" , mnt.mnt_dir); |
2250 | |
2251 | char *mount_real = realpath(mnt.mnt_dir, NULL); |
2252 | |
2253 | if (mount_real == NULL) { |
2254 | // Don't warn; current user may simply not be allowed access to all |
2255 | // mount points. |
2256 | cf_detail(CF_HARDWARE, |
2257 | "failed to resolve mount point %s: %d (%s)" , |
2258 | mnt.mnt_dir, errno, cf_strerror(errno)); |
2259 | continue; |
2260 | } |
2261 | |
2262 | cf_detail(CF_HARDWARE, "resolved mount point %s" , mount_real); |
2263 | |
2264 | size_t len = strlen(mount_real); |
2265 | |
2266 | if (len > best_len && strncmp(fs_real, mount_real, len) == 0) { |
2267 | strcpy(best_path, mnt.mnt_fsname); |
2268 | best_len = len; |
2269 | cf_detail(CF_HARDWARE, "new best %s with length %zu" , |
2270 | best_path, best_len); |
2271 | } |
2272 | |
2273 | free(mount_real); |
2274 | } |
2275 | |
2276 | endmntent(fh); |
2277 | free(fs_real); |
2278 | |
2279 | if (best_len == 0) { |
2280 | cf_warning(CF_HARDWARE, "no mount point found for %s" , fs_path); |
2281 | return NULL; |
2282 | } |
2283 | |
2284 | if (strncmp(best_path, "/dev" , 4) != 0) { |
2285 | // Don't warn; could be tmpfs, etc. |
2286 | cf_detail(CF_HARDWARE, "invalid device %s found for %s" , best_path, |
2287 | fs_path); |
2288 | return NULL; |
2289 | } |
2290 | |
2291 | char *best_real = realpath(best_path, NULL); |
2292 | |
2293 | if (best_real == NULL) { |
2294 | cf_warning(CF_HARDWARE, |
2295 | "failed to resolve mounted device %s: %d (%s)" , best_path, |
2296 | errno, cf_strerror(errno)); |
2297 | return NULL; |
2298 | } |
2299 | |
2300 | // Return a result allocated with the cf_*() allocation functions. |
2301 | |
2302 | char *res = cf_strdup(best_real); |
2303 | free(best_real); |
2304 | |
2305 | cf_detail(CF_HARDWARE, "mount point is %s" , res); |
2306 | return res; |
2307 | } |
2308 | |
2309 | static bool |
2310 | get_dev_key(const char *dev_path, dev_key_t *key) |
2311 | { |
2312 | cf_detail(CF_HARDWARE, "getting device key for %s" , dev_path); |
2313 | |
2314 | struct stat st; |
2315 | |
2316 | if (stat(dev_path, &st) < 0) { |
2317 | cf_warning(CF_HARDWARE, "failed to query meta data for %s: %d (%s)" , |
2318 | dev_path, errno, cf_strerror(errno)); |
2319 | return false; |
2320 | } |
2321 | |
2322 | if (!S_ISBLK(st.st_mode) && !S_ISCHR(st.st_mode)) { |
2323 | cf_warning(CF_HARDWARE, "%s is not a device" , dev_path); |
2324 | return false; |
2325 | } |
2326 | |
2327 | key->major = major(st.st_rdev); |
2328 | key->minor = minor(st.st_rdev); |
2329 | |
2330 | cf_detail(CF_HARDWARE, "device key %u:%u" , key->major, key->minor); |
2331 | return true; |
2332 | } |
2333 | |
2334 | static cf_topo_numa_node_index |
2335 | get_numa_node(const char *sys_path) |
2336 | { |
2337 | cf_detail(CF_HARDWARE, "finding NUMA node for %s" , sys_path); |
2338 | |
2339 | char *sys_real = realpath(sys_path, NULL); |
2340 | |
2341 | if (sys_real == NULL) { |
2342 | cf_warning(CF_HARDWARE, "failed to resolve path %s: %d (%s)" , |
2343 | sys_path, errno, cf_strerror(errno)); |
2344 | return INVALID_INDEX; |
2345 | } |
2346 | |
2347 | cf_topo_numa_node_index res = INVALID_INDEX; |
2348 | |
2349 | for (int32_t i = 0; i < 25; ++i) { |
2350 | cf_detail(CF_HARDWARE, "considering %s" , sys_real); |
2351 | |
2352 | char sys_numa[DEVICE_PATH_SIZE]; |
2353 | snprintf(sys_numa, DEVICE_PATH_SIZE, "%s/numa_node" , sys_real); |
2354 | |
2355 | cf_topo_numa_node_index tmp; |
2356 | |
2357 | if (read_numa_node(sys_numa, &tmp) == FILE_RES_OK) { |
2358 | cf_detail(CF_HARDWARE, "NUMA node found" ); |
2359 | res = tmp; |
2360 | break; |
2361 | } |
2362 | |
2363 | int32_t i_slash = -1; |
2364 | |
2365 | for (int32_t k = 0; sys_real[k] != 0; ++k) { |
2366 | if (sys_real[k] == '/') { |
2367 | i_slash = k; |
2368 | } |
2369 | } |
2370 | |
2371 | if (i_slash < 1) { |
2372 | break; |
2373 | } |
2374 | |
2375 | sys_real[i_slash] = 0; |
2376 | } |
2377 | |
2378 | free(sys_real); |
2379 | return res; |
2380 | } |
2381 | |
2382 | static int32_t |
2383 | get_nvme_age(const char *dev_path) |
2384 | { |
2385 | static const uint32_t SZ_BUFF = 512; |
2386 | |
2387 | cf_detail(CF_HARDWARE, "getting age for %s" , dev_path); |
2388 | |
2389 | if (!cf_process_has_cap(CAP_SYS_ADMIN)) { |
2390 | cf_detail(CF_HARDWARE, "insufficient privileges to query %s" , |
2391 | dev_path); |
2392 | return -1; |
2393 | } |
2394 | |
2395 | int32_t fd = open(dev_path, O_RDONLY); |
2396 | |
2397 | if (fd < 0) { |
2398 | if (errno == EACCES) { |
2399 | cf_detail(CF_HARDWARE, "insufficient privileges to open %s" , |
2400 | dev_path); |
2401 | } |
2402 | else { |
2403 | cf_warning(CF_HARDWARE, "failed to open %s: %d (%s)" , |
2404 | dev_path, errno, cf_strerror(errno)); |
2405 | } |
2406 | |
2407 | return -1; |
2408 | } |
2409 | |
2410 | uint8_t *buff = cf_valloc(SZ_BUFF); |
2411 | |
2412 | // Silence Valgrind, which doesn't know about this ioctl. |
2413 | |
2414 | memset(buff, 0, SZ_BUFF); |
2415 | |
2416 | // NVMe specification: https://bit.ly/2HPAS99 |
2417 | // |
2418 | // - See 4.2 for overall command format. |
2419 | // - See 5.14 for specifics of the Get Log page command. |
2420 | // |
2421 | // "0's based value" in the spec means that a value x in a data |
2422 | // structure actually means x + 1. |
2423 | |
2424 | uint32_t numdl = (SZ_BUFF / 4) - 1; // number of dwords lower (0's based) |
2425 | uint32_t lid = 2; // log page identifier: 2 (SMART log) |
2426 | |
2427 | uint32_t cdw10 = (numdl << 16) | lid; |
2428 | |
2429 | struct nvme_admin_cmd cmd = { |
2430 | .opcode = 0x02, // Get Log Page |
2431 | .nsid = 0xffffffff, // no namespace |
2432 | .addr = (uint64_t)buff, // result buffer |
2433 | .data_len = SZ_BUFF, // size of result buffer |
2434 | .cdw10 = cdw10 // command arguments |
2435 | }; |
2436 | |
2437 | cf_process_enable_cap(CAP_SYS_ADMIN); |
2438 | |
2439 | cf_detail(CF_HARDWARE, "querying %s" , dev_path); |
2440 | int32_t res = ioctl(fd, NVME_IOCTL_ADMIN_CMD, &cmd); |
2441 | |
2442 | cf_process_disable_cap(CAP_SYS_ADMIN); |
2443 | |
2444 | if (res < 0) { |
2445 | // Older kernels that don't support the IOCTL return EINVAL. |
2446 | // Submitting to non-NVMe devices causes ENOTTY. |
2447 | if (errno != EINVAL && errno != ENOTTY) { |
2448 | cf_warning(CF_HARDWARE, "failed to submit command to %s: %d (%s)" , |
2449 | dev_path, errno, cf_strerror(errno)); |
2450 | } |
2451 | |
2452 | cf_free(buff); |
2453 | close(fd); |
2454 | return -1; |
2455 | } |
2456 | |
2457 | if (res > 0){ |
2458 | // Some virtualized environments don't provide a SMART log page. |
2459 | if (res != NVME_SC_INVALID_LOG_PAGE) { |
2460 | cf_warning(CF_HARDWARE, "failed to submit command to %s: 0x%x" , |
2461 | dev_path, res); |
2462 | } |
2463 | |
2464 | cf_free(buff); |
2465 | close(fd); |
2466 | return -1; |
2467 | } |
2468 | |
2469 | // 0 <= age <= 255 - reported percentage used may exceed 100, when a drive |
2470 | // lives longer than predicted by its vendor. |
2471 | |
2472 | int32_t age = buff[5]; |
2473 | cf_detail(CF_HARDWARE, "percentage lived %d" , age); |
2474 | |
2475 | cf_free(buff); |
2476 | close(fd); |
2477 | |
2478 | return age; |
2479 | } |
2480 | |
2481 | static void |
2482 | update_path_data(path_data_t *data) |
2483 | { |
2484 | cf_storage_device_info *info = &data->info; |
2485 | |
2486 | cf_detail(CF_HARDWARE, "updating path data for %s" , info->dev_path); |
2487 | |
2488 | for (uint32_t i = 0; i < info->n_phys; ++i) { |
2489 | cf_detail(CF_HARDWARE, "updating %s" , info->phys[i].dev_path); |
2490 | info->phys[i].nvme_age = get_nvme_age(info->phys[i].dev_path); |
2491 | } |
2492 | |
2493 | data->mod_time = cf_get_seconds(); |
2494 | } |
2495 | |
2496 | static void |
2497 | visit_children(path_data_t *data, dev_node_t *node) |
2498 | { |
2499 | cf_storage_device_info *info = &data->info; |
2500 | |
2501 | cf_detail(CF_HARDWARE, "considering %s for %s" , node->dev_path, |
2502 | info->dev_path); |
2503 | |
2504 | if (node->sys_sched[0] != 0) { |
2505 | cf_detail(CF_HARDWARE, "found scheduler %s" , node->sys_sched); |
2506 | |
2507 | uint32_t n_sys_scheds = data->n_sys_scheds; |
2508 | |
2509 | if (n_sys_scheds >= CF_STORAGE_MAX_PHYS) { |
2510 | cf_warning(CF_HARDWARE, "too many schedulers for %s" , |
2511 | info->dev_path); |
2512 | return; |
2513 | } |
2514 | |
2515 | data->sys_scheds[n_sys_scheds] = node->sys_sched; |
2516 | ++data->n_sys_scheds; |
2517 | } |
2518 | |
2519 | if (node->n_children == 0) { |
2520 | cf_detail(CF_HARDWARE, "found physical device" ); |
2521 | |
2522 | uint32_t n_phys = info->n_phys; |
2523 | |
2524 | if (n_phys >= CF_STORAGE_MAX_PHYS) { |
2525 | cf_warning(CF_HARDWARE, "too many physical devices for %s" , |
2526 | info->dev_path); |
2527 | return; |
2528 | } |
2529 | |
2530 | info->phys[n_phys].dev_path = node->dev_path; |
2531 | info->phys[n_phys].numa_node = get_numa_node(node->sys_home); |
2532 | info->phys[n_phys].nvme_age = -1; |
2533 | |
2534 | ++info->n_phys; |
2535 | return; |
2536 | } |
2537 | |
2538 | cf_detail(CF_HARDWARE, "examining children" ); |
2539 | |
2540 | for (uint32_t i = 0; i < node->n_children; ++i) { |
2541 | visit_children(data, node->children[i]); |
2542 | } |
2543 | } |
2544 | |
2545 | static path_data_t * |
2546 | new_path_data(const char *any_path) |
2547 | { |
2548 | cf_detail(CF_HARDWARE, "creating path data for %s" , any_path); |
2549 | |
2550 | path_data_t *data = cf_malloc(sizeof(path_data_t)); |
2551 | struct stat st; |
2552 | |
2553 | if (stat(any_path, &st) < 0) { |
2554 | cf_warning(CF_HARDWARE, "failed to query meta data for %s: %d (%s)" , |
2555 | any_path, errno, cf_strerror(errno)); |
2556 | cf_free(data); |
2557 | return NULL; |
2558 | } |
2559 | |
2560 | cf_storage_device_info *info = &data->info; |
2561 | |
2562 | if (S_ISREG(st.st_mode) || S_ISDIR(st.st_mode)) { |
2563 | cf_detail(CF_HARDWARE, "%s is a file or directory" , any_path); |
2564 | info->dev_path = get_mounted_device(any_path); |
2565 | |
2566 | if (info->dev_path == NULL) { |
2567 | cf_free(data); |
2568 | return NULL; |
2569 | } |
2570 | } |
2571 | else if (S_ISBLK(st.st_mode) || S_ISCHR(st.st_mode)) { |
2572 | cf_detail(CF_HARDWARE, "%s is a device" , any_path); |
2573 | info->dev_path = cf_strdup(any_path); |
2574 | } |
2575 | else { |
2576 | cf_warning(CF_HARDWARE, "%s with unknown type 0x%x" , any_path, |
2577 | st.st_mode & S_IFMT); |
2578 | cf_free(data); |
2579 | return NULL; |
2580 | } |
2581 | |
2582 | cf_detail(CF_HARDWARE, "mapping device %s" , info->dev_path); |
2583 | |
2584 | dev_key_t key; |
2585 | |
2586 | if (!get_dev_key(info->dev_path, &key)) { |
2587 | cf_free(info->dev_path); |
2588 | cf_free(data); |
2589 | return NULL; |
2590 | } |
2591 | |
2592 | dev_node_t *node; |
2593 | |
2594 | if (cf_shash_get(g_dev_graph, &key, &node) != CF_SHASH_OK) { |
2595 | cf_warning(CF_HARDWARE, "no node for device key %u:%u" , key.major, |
2596 | key.minor); |
2597 | cf_free(info->dev_path); |
2598 | cf_free(data); |
2599 | return NULL; |
2600 | } |
2601 | |
2602 | cf_detail(CF_HARDWARE, "collecting dependency info" ); |
2603 | |
2604 | data->n_sys_scheds = 0; |
2605 | info->n_phys = 0; |
2606 | |
2607 | visit_children(data, node); |
2608 | |
2609 | cf_detail(CF_HARDWARE, "populating NVMe age" ); |
2610 | update_path_data(data); |
2611 | |
2612 | return data; |
2613 | } |
2614 | |
2615 | static path_data_t * |
2616 | get_path_data(const char *any_path) |
2617 | { |
2618 | cf_detail(CF_HARDWARE, "getting path data for %s" , any_path); |
2619 | |
2620 | cf_mutex_lock(&g_path_data_lock); |
2621 | |
2622 | if (g_dev_graph == NULL) { |
2623 | build_device_graph(); |
2624 | } |
2625 | |
2626 | if (g_path_data == NULL) { |
2627 | g_path_data = cf_shash_create(cf_shash_fn_zstr, |
2628 | DEVICE_PATH_SIZE, sizeof(path_data_t *), 256, 0); |
2629 | } |
2630 | |
2631 | size_t len = strlen(any_path); |
2632 | |
2633 | if (len >= DEVICE_PATH_SIZE) { |
2634 | cf_warning(CF_HARDWARE, "device path %s is too long" , any_path); |
2635 | cf_mutex_unlock(&g_path_data_lock); |
2636 | return NULL; |
2637 | } |
2638 | |
2639 | char key[DEVICE_PATH_SIZE]; |
2640 | |
2641 | memcpy(key, any_path, len); |
2642 | memset(key + len, 0, DEVICE_PATH_SIZE - len); |
2643 | |
2644 | path_data_t *data; |
2645 | |
2646 | if (cf_shash_get(g_path_data, key, &data) != CF_SHASH_OK) { |
2647 | cf_detail(CF_HARDWARE, "no path data" ); |
2648 | |
2649 | data = new_path_data(any_path); |
2650 | |
2651 | if (data == NULL) { |
2652 | cf_mutex_unlock(&g_path_data_lock); |
2653 | return NULL; |
2654 | } |
2655 | |
2656 | cf_shash_put_unique(g_path_data, key, &data); |
2657 | } |
2658 | else { |
2659 | cf_detail(CF_HARDWARE, "existing path data" ); |
2660 | } |
2661 | |
2662 | cf_clock now = cf_get_seconds(); |
2663 | |
2664 | if (now > data->mod_time + 86400) { |
2665 | update_path_data(data); |
2666 | } |
2667 | |
2668 | cf_mutex_unlock(&g_path_data_lock); |
2669 | return data; |
2670 | } |
2671 | |
2672 | cf_storage_device_info * |
2673 | cf_storage_get_device_info(const char *path) |
2674 | { |
2675 | cf_detail(CF_HARDWARE, "getting device info for %s" , path); |
2676 | |
2677 | path_data_t *data = get_path_data(path); |
2678 | |
2679 | if (data == NULL) { |
2680 | return NULL; |
2681 | } |
2682 | |
2683 | return &data->info; |
2684 | } |
2685 | |
2686 | void |
2687 | cf_storage_set_scheduler(const char *path, const char *sched) |
2688 | { |
2689 | cf_detail(CF_HARDWARE, "setting scheduler for %s to %s" , path, sched); |
2690 | |
2691 | path_data_t *data = get_path_data(path); |
2692 | |
2693 | if (data == NULL) { |
2694 | cf_warning(CF_HARDWARE, "couldn't find path data for %s" , path); |
2695 | return; |
2696 | } |
2697 | |
2698 | bool failed = false; |
2699 | |
2700 | for (uint32_t i = 0; i < data->n_sys_scheds; ++i) { |
2701 | if (write_file(data->sys_scheds[i], sched, strlen(sched)) != |
2702 | FILE_RES_OK) { |
2703 | failed = true; |
2704 | } |
2705 | } |
2706 | |
2707 | if (failed) { |
2708 | cf_warning(CF_HARDWARE, "couldn't set scheduler for %s to %s" , path, |
2709 | sched); |
2710 | } |
2711 | else { |
2712 | cf_info(CF_HARDWARE, "set scheduler for %s to %s" , path, sched); |
2713 | } |
2714 | } |
2715 | |
2716 | int64_t |
2717 | cf_storage_file_system_size(const char *path) |
2718 | { |
2719 | struct stat file; |
2720 | |
2721 | if (stat(path, &file) < 0) { |
2722 | switch (errno) { |
2723 | case ENOENT: |
2724 | cf_warning(CF_HARDWARE, "mount point %s does not exist" , path); |
2725 | break; |
2726 | |
2727 | case EACCES: |
2728 | cf_warning(CF_HARDWARE, "access to mount point %s denied" , path); |
2729 | break; |
2730 | |
2731 | default: |
2732 | cf_warning(CF_HARDWARE, |
2733 | "error while querying mount point %s: %d (%s)" , path, |
2734 | errno, cf_strerror(errno)); |
2735 | break; |
2736 | } |
2737 | |
2738 | return -1; |
2739 | } |
2740 | |
2741 | if (!S_ISDIR(file.st_mode)) { |
2742 | cf_warning(CF_HARDWARE, "mount point %s is not a directory" , path); |
2743 | return -1; |
2744 | } |
2745 | |
2746 | struct statfs fs; |
2747 | |
2748 | if (statfs(path, &fs) < 0) { |
2749 | cf_warning(CF_HARDWARE, |
2750 | "error while querying mount point %s: %d (%s)" , path, |
2751 | errno, cf_strerror(errno)); |
2752 | return -1; |
2753 | } |
2754 | |
2755 | int64_t sz = (int64_t)fs.f_bsize * (int64_t)fs.f_blocks; |
2756 | |
2757 | cf_detail(CF_HARDWARE, "file system size of %s is %ld" , path, sz); |
2758 | return sz; |
2759 | } |
2760 | |
2761 | bool |
2762 | cf_storage_is_root_fs(const char *path) |
2763 | { |
2764 | struct statvfs vfs; |
2765 | |
2766 | if (statvfs("/" , &vfs) < 0) { |
2767 | cf_crash(CF_HARDWARE, "cannot stat root directory" ); |
2768 | } |
2769 | |
2770 | uint64_t root_id = vfs.f_fsid; |
2771 | |
2772 | if (statvfs(path, &vfs) < 0) { |
2773 | cf_warning(CF_HARDWARE, "cannot stat %s: %d (%s)" , path, errno, |
2774 | cf_strerror(errno)); |
2775 | return false; |
2776 | } |
2777 | |
2778 | return vfs.f_fsid == root_id; |
2779 | } |
2780 | |
2781 | void |
2782 | cf_page_cache_dirty_limits(void) |
2783 | { |
2784 | write_file_safe("/proc/sys/vm/dirty_bytes" , "16777216" , 8); |
2785 | write_file_safe("/proc/sys/vm/dirty_background_bytes" , "1" , 1); |
2786 | write_file_safe("/proc/sys/vm/dirty_expire_centisecs" , "1" , 1); |
2787 | write_file_safe("/proc/sys/vm/dirty_writeback_centisecs" , "10" , 2); |
2788 | } |
2789 | |
2790 | bool |
2791 | cf_mount_is_local(const char *path) |
2792 | { |
2793 | if (g_i_numa_node == INVALID_INDEX) { |
2794 | cf_detail(CF_HARDWARE, "not NUMA pinned" ); |
2795 | return true; |
2796 | } |
2797 | |
2798 | cf_storage_device_info *info = cf_storage_get_device_info(path); |
2799 | cf_topo_numa_node_index numa_node = info->phys[0].numa_node; |
2800 | |
2801 | for (uint32_t i = 1; i < info->n_phys; i++) { |
2802 | if (info->phys[i].numa_node != numa_node) { |
2803 | cf_crash_nostack(CF_HARDWARE, "can't numa pin %s (%s,%s)" , path, |
2804 | info->phys[0].dev_path, info->phys[i].dev_path); |
2805 | } |
2806 | } |
2807 | |
2808 | return numa_node == g_i_numa_node; |
2809 | } |
2810 | |