Line data Source code
1 : /* SPDX-License-Identifier: BSD-3-Clause
2 : * Copyright (C) 2017 Intel Corporation.
3 : * All rights reserved.
4 : */
5 :
6 : #include "spdk/stdinc.h"
7 :
8 : #include "env_internal.h"
9 :
10 : #include "spdk/version.h"
11 : #include "spdk/env_dpdk.h"
12 : #include "spdk/log.h"
13 : #include "spdk/config.h"
14 :
15 : #include <openssl/ssl.h>
16 : #include <openssl/err.h>
17 :
18 : #include <rte_config.h>
19 : #include <rte_eal.h>
20 : #include <rte_errno.h>
21 : #include <rte_vfio.h>
22 :
23 : #define SPDK_ENV_DPDK_DEFAULT_NAME "spdk"
24 : #define SPDK_ENV_DPDK_DEFAULT_SHM_ID -1
25 : #define SPDK_ENV_DPDK_DEFAULT_MEM_SIZE -1
26 : #define SPDK_ENV_DPDK_DEFAULT_MAIN_CORE -1
27 : #define SPDK_ENV_DPDK_DEFAULT_MEM_CHANNEL -1
28 : #define SPDK_ENV_DPDK_DEFAULT_CORE_MASK "0x1"
29 : #define SPDK_ENV_DPDK_DEFAULT_BASE_VIRTADDR 0x200000000000
30 :
31 : #define DPDK_ALLOW_PARAM "--allow"
32 : #define DPDK_BLOCK_PARAM "--block"
33 : #define DPDK_MAIN_CORE_PARAM "--main-lcore"
34 :
35 : static char **g_eal_cmdline;
36 : static int g_eal_cmdline_argcount;
37 : static bool g_external_init = true;
38 :
39 : static char *
40 0 : _sprintf_alloc(const char *format, ...)
41 : {
42 0 : va_list args;
43 0 : va_list args_copy;
44 : char *buf;
45 : size_t bufsize;
46 : int rc;
47 :
48 0 : va_start(args, format);
49 :
50 : /* Try with a small buffer first. */
51 0 : bufsize = 32;
52 :
53 : /* Limit maximum buffer size to something reasonable so we don't loop forever. */
54 0 : while (bufsize <= 1024 * 1024) {
55 0 : buf = malloc(bufsize);
56 0 : if (buf == NULL) {
57 0 : va_end(args);
58 0 : return NULL;
59 : }
60 :
61 0 : va_copy(args_copy, args);
62 0 : rc = vsnprintf(buf, bufsize, format, args_copy);
63 0 : va_end(args_copy);
64 :
65 : /*
66 : * If vsnprintf() returned a count within our current buffer size, we are done.
67 : * The count does not include the \0 terminator, so rc == bufsize is not OK.
68 : */
69 0 : if (rc >= 0 && (size_t)rc < bufsize) {
70 0 : va_end(args);
71 0 : return buf;
72 : }
73 :
74 : /*
75 : * vsnprintf() should return the required space, but some libc versions do not
76 : * implement this correctly, so just double the buffer size and try again.
77 : *
78 : * We don't need the data in buf, so rather than realloc(), use free() and malloc()
79 : * again to avoid a copy.
80 : */
81 0 : free(buf);
82 0 : bufsize *= 2;
83 : }
84 :
85 0 : va_end(args);
86 0 : return NULL;
87 : }
88 :
89 : void
90 0 : spdk_env_opts_init_ext(struct spdk_env_opts *opts, size_t opts_size)
91 : {
92 0 : if (!opts) {
93 0 : return;
94 : }
95 :
96 0 : memset(opts, 0, opts_size);
97 0 : opts->opts_size = opts_size;
98 :
99 : /* These fields were all valid before this structure was ABI
100 : * versioned, so we just set them without checking opts_size.
101 : */
102 0 : opts->name = SPDK_ENV_DPDK_DEFAULT_NAME;
103 0 : opts->core_mask = SPDK_ENV_DPDK_DEFAULT_CORE_MASK;
104 0 : opts->shm_id = SPDK_ENV_DPDK_DEFAULT_SHM_ID;
105 0 : opts->mem_size = SPDK_ENV_DPDK_DEFAULT_MEM_SIZE;
106 0 : opts->main_core = SPDK_ENV_DPDK_DEFAULT_MAIN_CORE;
107 0 : opts->mem_channel = SPDK_ENV_DPDK_DEFAULT_MEM_CHANNEL;
108 0 : opts->base_virtaddr = SPDK_ENV_DPDK_DEFAULT_BASE_VIRTADDR;
109 :
110 : #define SET_FIELD(field, value) \
111 : if (offsetof(struct spdk_env_opts, field) + sizeof(opts->field) <= opts_size) { \
112 : opts->field = value; \
113 : }
114 :
115 0 : SET_FIELD(enforce_numa, false);
116 :
117 : #undef SET_FIELD
118 : }
119 :
120 0 : SPDK_LOG_DEPRECATION_REGISTER(spdk_env_opts_init, "spdk_env_opts_init()", "v25.05", 0);
121 :
122 : void
123 0 : spdk_env_opts_init(struct spdk_env_opts *opts)
124 : {
125 0 : struct spdk_env_opts _opts = {};
126 :
127 0 : SPDK_LOG_DEPRECATED(spdk_env_opts_init);
128 :
129 : /* This function predates the ABI versioning of spdk_env_opts, so
130 : * we just copy over the defaults for the size of the structure
131 : * when the ABI versioning was first introduced.
132 : */
133 0 : spdk_env_opts_init_ext(&_opts, sizeof(_opts));
134 0 : memcpy(opts, &_opts, offsetof(struct spdk_env_opts, opts_size));
135 0 : }
136 :
137 : static void
138 0 : free_args(char **args, int argcount)
139 : {
140 : int i;
141 :
142 0 : if (args == NULL) {
143 0 : return;
144 : }
145 :
146 0 : for (i = 0; i < argcount; i++) {
147 0 : free(args[i]);
148 : }
149 :
150 0 : if (argcount) {
151 0 : free(args);
152 : }
153 : }
154 :
155 : static char **
156 0 : push_arg(char *args[], int *argcount, char *arg)
157 : {
158 : char **tmp;
159 :
160 0 : if (arg == NULL) {
161 0 : SPDK_ERRLOG("%s: NULL arg supplied\n", __func__);
162 0 : free_args(args, *argcount);
163 0 : return NULL;
164 : }
165 :
166 0 : tmp = realloc(args, sizeof(char *) * (*argcount + 1));
167 0 : if (tmp == NULL) {
168 0 : free(arg);
169 0 : free_args(args, *argcount);
170 0 : return NULL;
171 : }
172 :
173 0 : tmp[*argcount] = arg;
174 0 : (*argcount)++;
175 :
176 0 : return tmp;
177 : }
178 :
179 : #if defined(__linux__) && defined(__x86_64__)
180 :
181 : /* TODO: Can likely get this value from rlimits in the future */
182 : #define SPDK_IOMMU_VA_REQUIRED_WIDTH 48
183 : #define VTD_CAP_MGAW_SHIFT 16
184 : #define VTD_CAP_MGAW_MASK (0x3F << VTD_CAP_MGAW_SHIFT)
185 : #define RD_AMD_CAP_VASIZE_SHIFT 15
186 : #define RD_AMD_CAP_VASIZE_MASK (0x7F << RD_AMD_CAP_VASIZE_SHIFT)
187 :
188 : static int
189 0 : get_iommu_width(void)
190 : {
191 0 : int width = 0;
192 0 : glob_t glob_results = {};
193 :
194 : /* Break * and / into separate strings to appease check_format.sh comment style check. */
195 0 : glob("/sys/devices/virtual/iommu/dmar*" "/intel-iommu/cap", 0, NULL, &glob_results);
196 0 : glob("/sys/class/iommu/ivhd*" "/amd-iommu/cap", GLOB_APPEND, NULL, &glob_results);
197 :
198 0 : for (size_t i = 0; i < glob_results.gl_pathc; i++) {
199 0 : const char *filename = glob_results.gl_pathv[0];
200 0 : FILE *file = fopen(filename, "r");
201 0 : uint64_t cap_reg = 0;
202 :
203 0 : if (file == NULL) {
204 0 : continue;
205 : }
206 :
207 0 : if (fscanf(file, "%" PRIx64, &cap_reg) == 1) {
208 0 : if (strstr(filename, "intel-iommu") != NULL) {
209 : /* We have an Intel IOMMU */
210 0 : int mgaw = ((cap_reg & VTD_CAP_MGAW_MASK) >> VTD_CAP_MGAW_SHIFT) + 1;
211 :
212 0 : if (width == 0 || (mgaw > 0 && mgaw < width)) {
213 0 : width = mgaw;
214 : }
215 0 : } else if (strstr(filename, "amd-iommu") != NULL) {
216 : /* We have an AMD IOMMU */
217 0 : int mgaw = ((cap_reg & RD_AMD_CAP_VASIZE_MASK) >> RD_AMD_CAP_VASIZE_SHIFT) + 1;
218 :
219 0 : if (width == 0 || (mgaw > 0 && mgaw < width)) {
220 0 : width = mgaw;
221 : }
222 : }
223 : }
224 :
225 0 : fclose(file);
226 : }
227 :
228 0 : globfree(&glob_results);
229 0 : return width;
230 : }
231 :
232 : #endif
233 :
234 : static int
235 0 : build_eal_cmdline(const struct spdk_env_opts *opts)
236 : {
237 0 : int argcount = 0;
238 : char **args;
239 : bool no_huge;
240 :
241 0 : args = NULL;
242 0 : no_huge = opts->no_huge || (opts->env_context && strstr(opts->env_context, "--no-huge") != NULL);
243 :
244 : /* set the program name */
245 0 : args = push_arg(args, &argcount, _sprintf_alloc("%s", opts->name));
246 0 : if (args == NULL) {
247 0 : return -1;
248 : }
249 :
250 : /* disable shared configuration files when in single process mode. This allows for cleaner shutdown */
251 0 : if (opts->shm_id < 0) {
252 0 : args = push_arg(args, &argcount, _sprintf_alloc("%s", "--no-shconf"));
253 0 : if (args == NULL) {
254 0 : return -1;
255 : }
256 : }
257 :
258 : /* Either lcore_map or core_mask must be set. If both, or none specified, fail */
259 0 : if ((opts->core_mask == NULL) == (opts->lcore_map == NULL)) {
260 0 : if (opts->core_mask && opts->lcore_map) {
261 0 : fprintf(stderr,
262 : "Both, lcore map and core mask are provided, while only one can be set\n");
263 : } else {
264 0 : fprintf(stderr, "Core mask or lcore map must be specified\n");
265 : }
266 0 : free_args(args, argcount);
267 0 : return -1;
268 : }
269 :
270 0 : if (opts->lcore_map) {
271 : /* If lcore list is set, generate --lcores parameter */
272 0 : args = push_arg(args, &argcount, _sprintf_alloc("--lcores=%s", opts->lcore_map));
273 0 : } else if (opts->core_mask[0] == '-') {
274 : /*
275 : * Set the coremask:
276 : *
277 : * - if it starts with '-', we presume it's literal EAL arguments such
278 : * as --lcores.
279 : *
280 : * - if it starts with '[', we presume it's a core list to use with the
281 : * -l option.
282 : *
283 : * - otherwise, it's a CPU mask of the form "0xff.." as expected by the
284 : * -c option.
285 : */
286 0 : args = push_arg(args, &argcount, _sprintf_alloc("%s", opts->core_mask));
287 0 : } else if (opts->core_mask[0] == '[') {
288 0 : char *l_arg = _sprintf_alloc("-l %s", opts->core_mask + 1);
289 :
290 0 : if (l_arg != NULL) {
291 0 : int len = strlen(l_arg);
292 :
293 0 : if (l_arg[len - 1] == ']') {
294 0 : l_arg[len - 1] = '\0';
295 : }
296 : }
297 0 : args = push_arg(args, &argcount, l_arg);
298 : } else {
299 0 : args = push_arg(args, &argcount, _sprintf_alloc("-c %s", opts->core_mask));
300 : }
301 :
302 0 : if (args == NULL) {
303 0 : return -1;
304 : }
305 :
306 : /* set the memory channel number */
307 0 : if (opts->mem_channel > 0) {
308 0 : args = push_arg(args, &argcount, _sprintf_alloc("-n %d", opts->mem_channel));
309 0 : if (args == NULL) {
310 0 : return -1;
311 : }
312 : }
313 :
314 : /* set the memory size */
315 0 : if (opts->mem_size >= 0) {
316 0 : args = push_arg(args, &argcount, _sprintf_alloc("-m %d", opts->mem_size));
317 0 : if (args == NULL) {
318 0 : return -1;
319 : }
320 : }
321 :
322 : /* set no huge pages */
323 0 : if (opts->no_huge) {
324 0 : mem_disable_huge_pages();
325 : }
326 :
327 0 : if (opts->enforce_numa) {
328 0 : mem_enforce_numa();
329 : }
330 :
331 : /* set the main core */
332 0 : if (opts->main_core > 0) {
333 0 : args = push_arg(args, &argcount, _sprintf_alloc("%s=%d",
334 : DPDK_MAIN_CORE_PARAM, opts->main_core));
335 0 : if (args == NULL) {
336 0 : return -1;
337 : }
338 : }
339 :
340 : /* set no pci if enabled */
341 0 : if (opts->no_pci) {
342 0 : args = push_arg(args, &argcount, _sprintf_alloc("--no-pci"));
343 0 : if (args == NULL) {
344 0 : return -1;
345 : }
346 : }
347 :
348 0 : if (no_huge) {
349 0 : if (opts->hugepage_single_segments || opts->unlink_hugepage || opts->hugedir) {
350 0 : fprintf(stderr, "--no-huge invalid with other hugepage options\n");
351 0 : free_args(args, argcount);
352 0 : return -1;
353 : }
354 :
355 0 : if (opts->mem_size < 0) {
356 0 : fprintf(stderr,
357 : "Disabling hugepages requires specifying how much memory "
358 : "will be allocated using -s parameter\n");
359 0 : free_args(args, argcount);
360 0 : return -1;
361 : }
362 :
363 : /* iova-mode=pa is incompatible with no_huge */
364 0 : if (opts->iova_mode &&
365 0 : (strcmp(opts->iova_mode, "pa") == 0)) {
366 0 : fprintf(stderr, "iova-mode=pa is incompatible with specified "
367 : "no-huge parameter\n");
368 0 : free_args(args, argcount);
369 0 : return -1;
370 : }
371 :
372 0 : args = push_arg(args, &argcount, _sprintf_alloc("--no-huge"));
373 0 : args = push_arg(args, &argcount, _sprintf_alloc("--iova-mode=va"));
374 :
375 : } else {
376 : /* create just one hugetlbfs file */
377 0 : if (opts->hugepage_single_segments) {
378 0 : args = push_arg(args, &argcount, _sprintf_alloc("--single-file-segments"));
379 0 : if (args == NULL) {
380 0 : return -1;
381 : }
382 : }
383 :
384 : /* unlink hugepages after initialization */
385 : /* Note: Automatically unlink hugepage when shm_id < 0, since it means we're not using
386 : * multi-process so we don't need the hugepage links anymore. But we need to make sure
387 : * we don't specify --huge-unlink implicitly if --single-file-segments was specified since
388 : * DPDK doesn't support that.
389 : */
390 0 : if (opts->unlink_hugepage ||
391 0 : (opts->shm_id < 0 && !opts->hugepage_single_segments)) {
392 0 : args = push_arg(args, &argcount, _sprintf_alloc("--huge-unlink"));
393 0 : if (args == NULL) {
394 0 : return -1;
395 : }
396 : }
397 :
398 : /* use a specific hugetlbfs mount */
399 0 : if (opts->hugedir) {
400 0 : args = push_arg(args, &argcount, _sprintf_alloc("--huge-dir=%s", opts->hugedir));
401 0 : if (args == NULL) {
402 0 : return -1;
403 : }
404 : }
405 : }
406 :
407 0 : if (opts->num_pci_addr) {
408 : size_t i;
409 0 : char bdf[32];
410 0 : struct spdk_pci_addr *pci_addr =
411 0 : opts->pci_blocked ? opts->pci_blocked : opts->pci_allowed;
412 :
413 0 : for (i = 0; i < opts->num_pci_addr; i++) {
414 0 : spdk_pci_addr_fmt(bdf, 32, &pci_addr[i]);
415 0 : args = push_arg(args, &argcount, _sprintf_alloc("%s=%s",
416 0 : (opts->pci_blocked ? DPDK_BLOCK_PARAM : DPDK_ALLOW_PARAM),
417 : bdf));
418 0 : if (args == NULL) {
419 0 : return -1;
420 : }
421 : }
422 : }
423 :
424 : /* Disable DPDK telemetry information by default, can be modified with env_context.
425 : * Prevents creation of dpdk_telemetry socket and additional pthread for it.
426 : */
427 0 : args = push_arg(args, &argcount, _sprintf_alloc("--no-telemetry"));
428 0 : if (args == NULL) {
429 0 : return -1;
430 : }
431 :
432 : /* Lower default EAL loglevel to RTE_LOG_NOTICE - normal, but significant messages.
433 : * This can be overridden by specifying the same option in opts->env_context
434 : */
435 0 : args = push_arg(args, &argcount, strdup("--log-level=lib.eal:6"));
436 0 : if (args == NULL) {
437 0 : return -1;
438 : }
439 :
440 : /* Lower default CRYPTO loglevel to RTE_LOG_WARNING to avoid a ton of init msgs.
441 : * This can be overridden by specifying the same option in opts->env_context
442 : */
443 0 : args = push_arg(args, &argcount, strdup("--log-level=lib.cryptodev:5"));
444 0 : if (args == NULL) {
445 0 : return -1;
446 : }
447 :
448 : /* Lower default POWER loglevel to RTE_LOG_WARNING to avoid a ton of init msgs.
449 : * This can be overridden by specifying the same option in opts->env_context
450 : */
451 0 : args = push_arg(args, &argcount, strdup("--log-level=lib.power:5"));
452 0 : if (args == NULL) {
453 0 : return -1;
454 : }
455 :
456 : /* `user1` log type is used by rte_vhost, which prints an INFO log for each received
457 : * vhost user message. We don't want that. The same log type is also used by a couple
458 : * of other DPDK libs, but none of which we make use right now. If necessary, this can
459 : * be overridden via opts->env_context.
460 : */
461 0 : args = push_arg(args, &argcount, strdup("--log-level=user1:6"));
462 0 : if (args == NULL) {
463 0 : return -1;
464 : }
465 :
466 : #ifdef __linux__
467 :
468 0 : if (opts->iova_mode) {
469 : /* iova-mode=pa is incompatible with no_huge */
470 0 : args = push_arg(args, &argcount, _sprintf_alloc("--iova-mode=%s", opts->iova_mode));
471 0 : if (args == NULL) {
472 0 : return -1;
473 : }
474 : } else {
475 : /* When using vfio with enable_unsafe_noiommu_mode=Y, we need iova-mode=pa,
476 : * but DPDK guesses it should be iova-mode=va. Add a check and force
477 : * iova-mode=pa here. */
478 0 : if (!no_huge && rte_vfio_noiommu_is_enabled()) {
479 0 : args = push_arg(args, &argcount, _sprintf_alloc("--iova-mode=pa"));
480 0 : if (args == NULL) {
481 0 : return -1;
482 : }
483 : }
484 :
485 : #if defined(__x86_64__)
486 : /* DPDK by default guesses that it should be using iova-mode=va so that it can
487 : * support running as an unprivileged user. However, some systems (especially
488 : * virtual machines) don't have an IOMMU capable of handling the full virtual
489 : * address space and DPDK doesn't currently catch that. Add a check in SPDK
490 : * and force iova-mode=pa here. */
491 0 : if (!no_huge && get_iommu_width() < SPDK_IOMMU_VA_REQUIRED_WIDTH) {
492 0 : args = push_arg(args, &argcount, _sprintf_alloc("--iova-mode=pa"));
493 0 : if (args == NULL) {
494 0 : return -1;
495 : }
496 : }
497 : #elif defined(__PPC64__)
498 : /* On Linux + PowerPC, DPDK doesn't support VA mode at all. Unfortunately, it doesn't correctly
499 : * auto-detect at the moment, so we'll just force it here. */
500 : args = push_arg(args, &argcount, _sprintf_alloc("--iova-mode=pa"));
501 : if (args == NULL) {
502 : return -1;
503 : }
504 : #endif
505 : }
506 :
507 :
508 : /* Set the base virtual address - it must be an address that is not in the
509 : * ASAN shadow region, otherwise ASAN-enabled builds will ignore the
510 : * mmap hint.
511 : *
512 : * Ref: https://github.com/google/sanitizers/wiki/AddressSanitizerAlgorithm
513 : */
514 0 : args = push_arg(args, &argcount, _sprintf_alloc("--base-virtaddr=0x%" PRIx64, opts->base_virtaddr));
515 0 : if (args == NULL) {
516 0 : return -1;
517 : }
518 :
519 : /* --match-allocation prevents DPDK from merging or splitting system memory allocations under the hood.
520 : * This is critical for RDMA when attempting to use an rte_mempool based buffer pool. If DPDK merges two
521 : * physically or IOVA contiguous memory regions, then when we go to allocate a buffer pool, it can split
522 : * the memory for a buffer over two allocations meaning the buffer will be split over a memory region.
523 : */
524 :
525 : /* --no-huge is incompatible with --match-allocations
526 : * Ref: https://doc.dpdk.org/guides/prog_guide/env_abstraction_layer.html#hugepage-allocation-matching
527 : */
528 0 : if (!no_huge &&
529 0 : (!opts->env_context || strstr(opts->env_context, "--legacy-mem") == NULL)) {
530 0 : args = push_arg(args, &argcount, _sprintf_alloc("%s", "--match-allocations"));
531 0 : if (args == NULL) {
532 0 : return -1;
533 : }
534 : }
535 :
536 0 : if (opts->shm_id < 0) {
537 0 : args = push_arg(args, &argcount, _sprintf_alloc("--file-prefix=spdk_pid%d",
538 : getpid()));
539 0 : if (args == NULL) {
540 0 : return -1;
541 : }
542 : } else {
543 0 : args = push_arg(args, &argcount, _sprintf_alloc("--file-prefix=spdk%d",
544 : opts->shm_id));
545 0 : if (args == NULL) {
546 0 : return -1;
547 : }
548 :
549 : /* set the process type */
550 0 : args = push_arg(args, &argcount, _sprintf_alloc("--proc-type=auto"));
551 0 : if (args == NULL) {
552 0 : return -1;
553 : }
554 : }
555 :
556 : /* --vfio-vf-token used for VF initialized by vfio_pci driver. */
557 0 : if (opts->vf_token) {
558 0 : args = push_arg(args, &argcount, _sprintf_alloc("--vfio-vf-token=%s",
559 : opts->vf_token));
560 0 : if (args == NULL) {
561 0 : return -1;
562 : }
563 : }
564 : #endif
565 :
566 0 : if (opts->env_context) {
567 0 : char *ptr = strdup(opts->env_context);
568 0 : char *tok = strtok(ptr, " \t");
569 :
570 : /* DPDK expects each argument as a separate string in the argv
571 : * array, so we need to tokenize here in case the caller
572 : * passed multiple arguments in the env_context string.
573 : */
574 0 : while (tok != NULL) {
575 0 : args = push_arg(args, &argcount, strdup(tok));
576 0 : tok = strtok(NULL, " \t");
577 : }
578 :
579 0 : free(ptr);
580 : }
581 :
582 0 : g_eal_cmdline = args;
583 0 : g_eal_cmdline_argcount = argcount;
584 0 : return argcount;
585 : }
586 :
587 : int
588 0 : spdk_env_dpdk_post_init(bool legacy_mem)
589 : {
590 : int rc;
591 :
592 0 : rc = pci_env_init();
593 0 : if (rc < 0) {
594 0 : SPDK_ERRLOG("pci_env_init() failed\n");
595 0 : return rc;
596 : }
597 :
598 0 : rc = mem_map_init(legacy_mem);
599 0 : if (rc < 0) {
600 0 : SPDK_ERRLOG("Failed to allocate mem_map\n");
601 0 : return rc;
602 : }
603 :
604 0 : rc = vtophys_init();
605 0 : if (rc < 0) {
606 0 : SPDK_ERRLOG("Failed to initialize vtophys\n");
607 0 : return rc;
608 : }
609 :
610 0 : return 0;
611 : }
612 :
613 : void
614 0 : spdk_env_dpdk_post_fini(void)
615 : {
616 0 : pci_env_fini();
617 :
618 0 : free_args(g_eal_cmdline, g_eal_cmdline_argcount);
619 0 : g_eal_cmdline = NULL;
620 0 : g_eal_cmdline_argcount = 0;
621 0 : }
622 :
623 : static void
624 0 : env_copy_opts(struct spdk_env_opts *opts, const struct spdk_env_opts *opts_user, size_t opts_size)
625 : {
626 0 : spdk_env_opts_init_ext(opts, sizeof(*opts));
627 0 : memcpy(opts, opts_user, offsetof(struct spdk_env_opts, opts_size));
628 0 : opts->opts_size = opts_size;
629 :
630 : #define SET_FIELD(field) \
631 : if (offsetof(struct spdk_env_opts, field) + sizeof(opts->field) <= opts_size) { \
632 : opts->field = opts_user->field; \
633 : }
634 :
635 0 : SET_FIELD(enforce_numa);
636 :
637 : #undef SET_FIELD
638 0 : }
639 :
640 : int
641 0 : spdk_env_init_ext(const struct spdk_env_opts *opts_user)
642 : {
643 0 : struct spdk_env_opts opts_local = {};
644 0 : struct spdk_env_opts *opts = &opts_local;
645 0 : char **dpdk_args = NULL;
646 0 : char *args_print = NULL, *args_tmp = NULL;
647 : OPENSSL_INIT_SETTINGS *settings;
648 : int i, rc;
649 : int orig_optind;
650 : bool legacy_mem;
651 :
652 : /* If SPDK env has been initialized before, then only pci env requires
653 : * reinitialization.
654 : */
655 0 : if (g_external_init == false) {
656 0 : if (opts_user != NULL) {
657 0 : fprintf(stderr, "Invalid arguments to reinitialize SPDK env\n");
658 0 : return -EINVAL;
659 : }
660 :
661 0 : printf("Starting %s / %s reinitialization...\n", SPDK_VERSION_STRING, rte_version());
662 0 : pci_env_reinit();
663 :
664 0 : return 0;
665 : }
666 :
667 0 : if (opts_user == NULL) {
668 0 : fprintf(stderr, "NULL arguments to initialize DPDK\n");
669 0 : return -EINVAL;
670 : }
671 :
672 0 : if (opts_user->opts_size < offsetof(struct spdk_env_opts, opts_size) + sizeof(opts->opts_size)) {
673 0 : fprintf(stderr, "Invalid opts->opts_size\n");
674 0 : return -EINVAL;
675 : }
676 :
677 0 : env_copy_opts(opts, opts_user, opts_user->opts_size);
678 :
679 0 : settings = OPENSSL_INIT_new();
680 0 : if (!settings) {
681 0 : fprintf(stderr, "Failed to create openssl settings object\n");
682 0 : ERR_print_errors_fp(stderr);
683 0 : return -ENOMEM;
684 : }
685 :
686 : #if OPENSSL_VERSION_NUMBER >= 0x30000000 /* OPENSSL 3.0.0 */
687 : OPENSSL_INIT_set_config_file_flags(settings, 0);
688 : #endif
689 0 : rc = OPENSSL_init_ssl(OPENSSL_INIT_LOAD_CONFIG, settings);
690 0 : if (rc != 1) {
691 0 : fprintf(stderr, "Failed to initialize OpenSSL\n");
692 0 : ERR_print_errors_fp(stderr);
693 0 : return -EINVAL;
694 : }
695 0 : OPENSSL_INIT_free(settings);
696 :
697 0 : rc = build_eal_cmdline(opts);
698 0 : if (rc < 0) {
699 0 : SPDK_ERRLOG("Invalid arguments to initialize DPDK\n");
700 0 : return -EINVAL;
701 : }
702 :
703 0 : SPDK_PRINTF("Starting %s / %s initialization...\n", SPDK_VERSION_STRING, rte_version());
704 :
705 0 : args_print = _sprintf_alloc("[ DPDK EAL parameters: ");
706 0 : if (args_print == NULL) {
707 0 : return -ENOMEM;
708 : }
709 0 : for (i = 0; i < g_eal_cmdline_argcount; i++) {
710 0 : args_tmp = args_print;
711 0 : args_print = _sprintf_alloc("%s%s ", args_tmp, g_eal_cmdline[i]);
712 0 : if (args_print == NULL) {
713 0 : free(args_tmp);
714 0 : return -ENOMEM;
715 : }
716 0 : free(args_tmp);
717 : }
718 0 : SPDK_PRINTF("%s]\n", args_print);
719 0 : free(args_print);
720 :
721 : /* DPDK rearranges the array we pass to it, so make a copy
722 : * before passing so we can still free the individual strings
723 : * correctly.
724 : */
725 0 : dpdk_args = calloc(g_eal_cmdline_argcount, sizeof(char *));
726 0 : if (dpdk_args == NULL) {
727 0 : SPDK_ERRLOG("Failed to allocate dpdk_args\n");
728 0 : return -ENOMEM;
729 : }
730 0 : memcpy(dpdk_args, g_eal_cmdline, sizeof(char *) * g_eal_cmdline_argcount);
731 :
732 0 : fflush(stdout);
733 0 : orig_optind = optind;
734 0 : optind = 1;
735 0 : rc = rte_eal_init(g_eal_cmdline_argcount, dpdk_args);
736 0 : optind = orig_optind;
737 :
738 0 : free(dpdk_args);
739 :
740 0 : if (rc < 0) {
741 0 : if (rte_errno == EALREADY) {
742 0 : SPDK_ERRLOG("DPDK already initialized\n");
743 : } else {
744 0 : SPDK_ERRLOG("Failed to initialize DPDK\n");
745 : }
746 0 : return -rte_errno;
747 : }
748 :
749 0 : legacy_mem = false;
750 0 : if (opts->env_context && strstr(opts->env_context, "--legacy-mem") != NULL) {
751 0 : legacy_mem = true;
752 : }
753 :
754 0 : rc = spdk_env_dpdk_post_init(legacy_mem);
755 0 : if (rc == 0) {
756 0 : g_external_init = false;
757 : }
758 :
759 0 : return rc;
760 : }
761 :
762 0 : SPDK_LOG_DEPRECATION_REGISTER(spdk_env_init, "spdk_env_init()", "v25.05", 0);
763 :
764 : int
765 0 : spdk_env_init(const struct spdk_env_opts *opts_user)
766 : {
767 0 : SPDK_LOG_DEPRECATED(spdk_env_init);
768 :
769 : /* This is an old API, prior to having a size-versions spdk_env_opts
770 : * struct. This means we do not know the size of the _opts parameter,
771 : * we must assume it has size matching the structure when the
772 : * versioning was first introduced. So allocate latest version of
773 : * this struct on the stack and use env_copy_opts to copy over the
774 : * bytes we know must be valid.
775 : */
776 0 : struct spdk_env_opts opts = {};
777 :
778 0 : env_copy_opts(&opts, opts_user, offsetof(struct spdk_env_opts, opts_size) + sizeof(opts.opts_size));
779 0 : return spdk_env_init_ext(&opts);
780 : }
781 :
782 : /* We use priority 101 which is the highest priority level available
783 : * to applications (the toolchains reserve 1 to 100 for internal usage).
784 : * This ensures this destructor runs last, after any other destructors
785 : * that might still need the environment up and running.
786 : */
787 : __attribute__((destructor(101))) static void
788 0 : dpdk_cleanup(void)
789 : {
790 : /* Only call rte_eal_cleanup if the SPDK env library called rte_eal_init. */
791 0 : if (!g_external_init) {
792 0 : rte_eal_cleanup();
793 : }
794 0 : }
795 :
796 : void
797 0 : spdk_env_fini(void)
798 : {
799 0 : spdk_env_dpdk_post_fini();
800 0 : }
801 :
802 : bool
803 0 : spdk_env_dpdk_external_init(void)
804 : {
805 0 : return g_external_init;
806 : }
|