1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation.
3 * Copyright(c) 2013 6WIND S.A.
4 */
5
6 #include <errno.h>
7 #include <fcntl.h>
8 #include <stdbool.h>
9 #include <stdlib.h>
10 #include <stdio.h>
11 #include <stdint.h>
12 #include <inttypes.h>
13 #include <string.h>
14 #include <sys/mman.h>
15 #include <sys/stat.h>
16 #include <sys/file.h>
17 #include <sys/resource.h>
18 #include <unistd.h>
19 #include <limits.h>
20 #include <signal.h>
21 #include <setjmp.h>
22 #ifdef F_ADD_SEALS /* if file sealing is supported, so is memfd */
23 #define MEMFD_SUPPORTED
24 #endif
25 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
26 #include <numa.h>
27 #include <numaif.h>
28 #endif
29
30 #include <rte_errno.h>
31 #include <rte_log.h>
32 #include <rte_memory.h>
33 #include <rte_eal.h>
34 #include <rte_lcore.h>
35 #include <rte_common.h>
36
37 #include "eal_private.h"
38 #include "eal_memalloc.h"
39 #include "eal_memcfg.h"
40 #include "eal_internal_cfg.h"
41 #include "eal_filesystem.h"
42 #include "eal_hugepages.h"
43 #include "eal_options.h"
44
45 #define PFN_MASK_SIZE 8
46
47 /**
48 * @file
49 * Huge page mapping under linux
50 *
51 * To reserve a big contiguous amount of memory, we use the hugepage
52 * feature of linux. For that, we need to have hugetlbfs mounted. This
53 * code will create many files in this directory (one per page) and
54 * map them in virtual memory. For each page, we will retrieve its
55 * physical address and remap it in order to have a virtual contiguous
56 * zone as well as a physical contiguous zone.
57 */
58
59 static int phys_addrs_available = -1;
60
61 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
62
eal_get_baseaddr(void)63 uint64_t eal_get_baseaddr(void)
64 {
65 /*
66 * Linux kernel uses a really high address as starting address for
67 * serving mmaps calls. If there exists addressing limitations and IOVA
68 * mode is VA, this starting address is likely too high for those
69 * devices. However, it is possible to use a lower address in the
70 * process virtual address space as with 64 bits there is a lot of
71 * available space.
72 *
73 * Current known limitations are 39 or 40 bits. Setting the starting
74 * address at 4GB implies there are 508GB or 1020GB for mapping the
75 * available hugepages. This is likely enough for most systems, although
76 * a device with addressing limitations should call
77 * rte_mem_check_dma_mask for ensuring all memory is within supported
78 * range.
79 */
80 return 0x100000000ULL;
81 }
82
83 /*
84 * Get physical address of any mapped virtual address in the current process.
85 */
86 phys_addr_t
rte_mem_virt2phy(const void * virtaddr)87 rte_mem_virt2phy(const void *virtaddr)
88 {
89 int fd, retval;
90 uint64_t page, physaddr;
91 unsigned long virt_pfn;
92 int page_size;
93 off_t offset;
94
95 if (phys_addrs_available == 0)
96 return RTE_BAD_IOVA;
97
98 /* standard page size */
99 page_size = getpagesize();
100
101 fd = open("/proc/self/pagemap", O_RDONLY);
102 if (fd < 0) {
103 RTE_LOG(INFO, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
104 __func__, strerror(errno));
105 return RTE_BAD_IOVA;
106 }
107
108 virt_pfn = (unsigned long)virtaddr / page_size;
109 offset = sizeof(uint64_t) * virt_pfn;
110 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
111 RTE_LOG(INFO, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
112 __func__, strerror(errno));
113 close(fd);
114 return RTE_BAD_IOVA;
115 }
116
117 retval = read(fd, &page, PFN_MASK_SIZE);
118 close(fd);
119 if (retval < 0) {
120 RTE_LOG(INFO, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
121 __func__, strerror(errno));
122 return RTE_BAD_IOVA;
123 } else if (retval != PFN_MASK_SIZE) {
124 RTE_LOG(INFO, EAL, "%s(): read %d bytes from /proc/self/pagemap "
125 "but expected %d:\n",
126 __func__, retval, PFN_MASK_SIZE);
127 return RTE_BAD_IOVA;
128 }
129
130 /*
131 * the pfn (page frame number) are bits 0-54 (see
132 * pagemap.txt in linux Documentation)
133 */
134 if ((page & 0x7fffffffffffffULL) == 0)
135 return RTE_BAD_IOVA;
136
137 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
138 + ((unsigned long)virtaddr % page_size);
139
140 return physaddr;
141 }
142
143 rte_iova_t
rte_mem_virt2iova(const void * virtaddr)144 rte_mem_virt2iova(const void *virtaddr)
145 {
146 if (rte_eal_iova_mode() == RTE_IOVA_VA)
147 return (uintptr_t)virtaddr;
148 return rte_mem_virt2phy(virtaddr);
149 }
150
151 /*
152 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
153 * it by browsing the /proc/self/pagemap special file.
154 */
155 static int
find_physaddrs(struct hugepage_file * hugepg_tbl,struct hugepage_info * hpi)156 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
157 {
158 unsigned int i;
159 phys_addr_t addr;
160
161 for (i = 0; i < hpi->num_pages[0]; i++) {
162 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
163 if (addr == RTE_BAD_PHYS_ADDR)
164 return -1;
165 hugepg_tbl[i].physaddr = addr;
166 }
167 return 0;
168 }
169
170 /*
171 * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
172 */
173 static int
set_physaddrs(struct hugepage_file * hugepg_tbl,struct hugepage_info * hpi)174 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
175 {
176 unsigned int i;
177 static phys_addr_t addr;
178
179 for (i = 0; i < hpi->num_pages[0]; i++) {
180 hugepg_tbl[i].physaddr = addr;
181 addr += hugepg_tbl[i].size;
182 }
183 return 0;
184 }
185
186 /*
187 * Check whether address-space layout randomization is enabled in
188 * the kernel. This is important for multi-process as it can prevent
189 * two processes mapping data to the same virtual address
190 * Returns:
191 * 0 - address space randomization disabled
192 * 1/2 - address space randomization enabled
193 * negative error code on error
194 */
195 static int
aslr_enabled(void)196 aslr_enabled(void)
197 {
198 char c;
199 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
200 if (fd < 0)
201 return -errno;
202 retval = read(fd, &c, 1);
203 close(fd);
204 if (retval < 0)
205 return -errno;
206 if (retval == 0)
207 return -EIO;
208 switch (c) {
209 case '0' : return 0;
210 case '1' : return 1;
211 case '2' : return 2;
212 default: return -EINVAL;
213 }
214 }
215
216 static sigjmp_buf huge_jmpenv;
217
huge_sigbus_handler(int signo __rte_unused)218 static void huge_sigbus_handler(int signo __rte_unused)
219 {
220 siglongjmp(huge_jmpenv, 1);
221 }
222
223 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
224 * non-static local variable in the stack frame calling sigsetjmp might be
225 * clobbered by a call to longjmp.
226 */
huge_wrap_sigsetjmp(void)227 static int huge_wrap_sigsetjmp(void)
228 {
229 return sigsetjmp(huge_jmpenv, 1);
230 }
231
232 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
233 /* Callback for numa library. */
numa_error(char * where)234 void numa_error(char *where)
235 {
236 RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
237 }
238 #endif
239
240 /*
241 * Mmap all hugepages of hugepage table: it first open a file in
242 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
243 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
244 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
245 * map contiguous physical blocks in contiguous virtual blocks.
246 */
247 static unsigned
map_all_hugepages(struct hugepage_file * hugepg_tbl,struct hugepage_info * hpi,uint64_t * essential_memory __rte_unused)248 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
249 uint64_t *essential_memory __rte_unused)
250 {
251 int fd;
252 unsigned i;
253 void *virtaddr;
254 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
255 int node_id = -1;
256 int essential_prev = 0;
257 int oldpolicy;
258 struct bitmask *oldmask = NULL;
259 bool have_numa = true;
260 unsigned long maxnode = 0;
261 const struct internal_config *internal_conf =
262 eal_get_internal_configuration();
263
264 /* Check if kernel supports NUMA. */
265 if (numa_available() != 0) {
266 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
267 have_numa = false;
268 }
269
270 if (have_numa) {
271 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
272 oldmask = numa_allocate_nodemask();
273 if (get_mempolicy(&oldpolicy, oldmask->maskp,
274 oldmask->size + 1, 0, 0) < 0) {
275 RTE_LOG(ERR, EAL,
276 "Failed to get current mempolicy: %s. "
277 "Assuming MPOL_DEFAULT.\n", strerror(errno));
278 oldpolicy = MPOL_DEFAULT;
279 }
280 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
281 if (internal_conf->socket_mem[i])
282 maxnode = i + 1;
283 }
284 #endif
285
286 for (i = 0; i < hpi->num_pages[0]; i++) {
287 struct hugepage_file *hf = &hugepg_tbl[i];
288 uint64_t hugepage_sz = hpi->hugepage_sz;
289
290 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
291 if (maxnode) {
292 unsigned int j;
293
294 for (j = 0; j < maxnode; j++)
295 if (essential_memory[j])
296 break;
297
298 if (j == maxnode) {
299 node_id = (node_id + 1) % maxnode;
300 while (!internal_conf->socket_mem[node_id]) {
301 node_id++;
302 node_id %= maxnode;
303 }
304 essential_prev = 0;
305 } else {
306 node_id = j;
307 essential_prev = essential_memory[j];
308
309 if (essential_memory[j] < hugepage_sz)
310 essential_memory[j] = 0;
311 else
312 essential_memory[j] -= hugepage_sz;
313 }
314
315 RTE_LOG(DEBUG, EAL,
316 "Setting policy MPOL_PREFERRED for socket %d\n",
317 node_id);
318 numa_set_preferred(node_id);
319 }
320 #endif
321
322 hf->file_id = i;
323 hf->size = hugepage_sz;
324 eal_get_hugefile_path(hf->filepath, sizeof(hf->filepath),
325 hpi->hugedir, hf->file_id);
326 hf->filepath[sizeof(hf->filepath) - 1] = '\0';
327
328 /* try to create hugepage file */
329 fd = open(hf->filepath, O_CREAT | O_RDWR, 0600);
330 if (fd < 0) {
331 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
332 strerror(errno));
333 goto out;
334 }
335
336 /* map the segment, and populate page tables,
337 * the kernel fills this segment with zeros. we don't care where
338 * this gets mapped - we already have contiguous memory areas
339 * ready for us to map into.
340 */
341 virtaddr = mmap(NULL, hugepage_sz, PROT_READ | PROT_WRITE,
342 MAP_SHARED | MAP_POPULATE, fd, 0);
343 if (virtaddr == MAP_FAILED) {
344 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
345 strerror(errno));
346 close(fd);
347 goto out;
348 }
349
350 hf->orig_va = virtaddr;
351
352 /* In linux, hugetlb limitations, like cgroup, are
353 * enforced at fault time instead of mmap(), even
354 * with the option of MAP_POPULATE. Kernel will send
355 * a SIGBUS signal. To avoid to be killed, save stack
356 * environment here, if SIGBUS happens, we can jump
357 * back here.
358 */
359 if (huge_wrap_sigsetjmp()) {
360 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
361 "hugepages of size %u MB\n",
362 (unsigned int)(hugepage_sz / 0x100000));
363 munmap(virtaddr, hugepage_sz);
364 close(fd);
365 unlink(hugepg_tbl[i].filepath);
366 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
367 if (maxnode)
368 essential_memory[node_id] =
369 essential_prev;
370 #endif
371 goto out;
372 }
373 *(int *)virtaddr = 0;
374
375 /* set shared lock on the file. */
376 if (flock(fd, LOCK_SH) < 0) {
377 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
378 __func__, strerror(errno));
379 close(fd);
380 goto out;
381 }
382
383 close(fd);
384 }
385
386 out:
387 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
388 if (maxnode) {
389 RTE_LOG(DEBUG, EAL,
390 "Restoring previous memory policy: %d\n", oldpolicy);
391 if (oldpolicy == MPOL_DEFAULT) {
392 numa_set_localalloc();
393 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
394 oldmask->size + 1) < 0) {
395 RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
396 strerror(errno));
397 numa_set_localalloc();
398 }
399 }
400 if (oldmask != NULL)
401 numa_free_cpumask(oldmask);
402 #endif
403 return i;
404 }
405
406 /*
407 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
408 * page.
409 */
410 static int
find_numasocket(struct hugepage_file * hugepg_tbl,struct hugepage_info * hpi)411 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
412 {
413 int socket_id;
414 char *end, *nodestr;
415 unsigned i, hp_count = 0;
416 uint64_t virt_addr;
417 char buf[BUFSIZ];
418 char hugedir_str[PATH_MAX];
419 FILE *f;
420
421 f = fopen("/proc/self/numa_maps", "r");
422 if (f == NULL) {
423 RTE_LOG(NOTICE, EAL, "NUMA support not available"
424 " consider that all memory is in socket_id 0\n");
425 return 0;
426 }
427
428 snprintf(hugedir_str, sizeof(hugedir_str),
429 "%s/%s", hpi->hugedir, eal_get_hugefile_prefix());
430
431 /* parse numa map */
432 while (fgets(buf, sizeof(buf), f) != NULL) {
433
434 /* ignore non huge page */
435 if (strstr(buf, " huge ") == NULL &&
436 strstr(buf, hugedir_str) == NULL)
437 continue;
438
439 /* get zone addr */
440 virt_addr = strtoull(buf, &end, 16);
441 if (virt_addr == 0 || end == buf) {
442 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
443 goto error;
444 }
445
446 /* get node id (socket id) */
447 nodestr = strstr(buf, " N");
448 if (nodestr == NULL) {
449 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
450 goto error;
451 }
452 nodestr += 2;
453 end = strstr(nodestr, "=");
454 if (end == NULL) {
455 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
456 goto error;
457 }
458 end[0] = '\0';
459 end = NULL;
460
461 socket_id = strtoul(nodestr, &end, 0);
462 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
463 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
464 goto error;
465 }
466
467 /* if we find this page in our mappings, set socket_id */
468 for (i = 0; i < hpi->num_pages[0]; i++) {
469 void *va = (void *)(unsigned long)virt_addr;
470 if (hugepg_tbl[i].orig_va == va) {
471 hugepg_tbl[i].socket_id = socket_id;
472 hp_count++;
473 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
474 RTE_LOG(DEBUG, EAL,
475 "Hugepage %s is on socket %d\n",
476 hugepg_tbl[i].filepath, socket_id);
477 #endif
478 }
479 }
480 }
481
482 if (hp_count < hpi->num_pages[0])
483 goto error;
484
485 fclose(f);
486 return 0;
487
488 error:
489 fclose(f);
490 return -1;
491 }
492
493 static int
cmp_physaddr(const void * a,const void * b)494 cmp_physaddr(const void *a, const void *b)
495 {
496 #ifndef RTE_ARCH_PPC_64
497 const struct hugepage_file *p1 = a;
498 const struct hugepage_file *p2 = b;
499 #else
500 /* PowerPC needs memory sorted in reverse order from x86 */
501 const struct hugepage_file *p1 = b;
502 const struct hugepage_file *p2 = a;
503 #endif
504 if (p1->physaddr < p2->physaddr)
505 return -1;
506 else if (p1->physaddr > p2->physaddr)
507 return 1;
508 else
509 return 0;
510 }
511
512 /*
513 * Uses mmap to create a shared memory area for storage of data
514 * Used in this file to store the hugepage file map on disk
515 */
516 static void *
create_shared_memory(const char * filename,const size_t mem_size)517 create_shared_memory(const char *filename, const size_t mem_size)
518 {
519 void *retval;
520 int fd;
521 const struct internal_config *internal_conf =
522 eal_get_internal_configuration();
523
524 /* if no shared files mode is used, create anonymous memory instead */
525 if (internal_conf->no_shconf) {
526 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE,
527 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
528 if (retval == MAP_FAILED)
529 return NULL;
530 return retval;
531 }
532
533 fd = open(filename, O_CREAT | O_RDWR, 0600);
534 if (fd < 0)
535 return NULL;
536 if (ftruncate(fd, mem_size) < 0) {
537 close(fd);
538 return NULL;
539 }
540 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
541 close(fd);
542 if (retval == MAP_FAILED)
543 return NULL;
544 return retval;
545 }
546
547 /*
548 * this copies *active* hugepages from one hugepage table to another.
549 * destination is typically the shared memory.
550 */
551 static int
copy_hugepages_to_shared_mem(struct hugepage_file * dst,int dest_size,const struct hugepage_file * src,int src_size)552 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
553 const struct hugepage_file * src, int src_size)
554 {
555 int src_pos, dst_pos = 0;
556
557 for (src_pos = 0; src_pos < src_size; src_pos++) {
558 if (src[src_pos].orig_va != NULL) {
559 /* error on overflow attempt */
560 if (dst_pos == dest_size)
561 return -1;
562 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
563 dst_pos++;
564 }
565 }
566 return 0;
567 }
568
569 static int
unlink_hugepage_files(struct hugepage_file * hugepg_tbl,unsigned num_hp_info)570 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
571 unsigned num_hp_info)
572 {
573 unsigned socket, size;
574 int page, nrpages = 0;
575 const struct internal_config *internal_conf =
576 eal_get_internal_configuration();
577
578 /* get total number of hugepages */
579 for (size = 0; size < num_hp_info; size++)
580 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
581 nrpages +=
582 internal_conf->hugepage_info[size].num_pages[socket];
583
584 for (page = 0; page < nrpages; page++) {
585 struct hugepage_file *hp = &hugepg_tbl[page];
586
587 if (hp->orig_va != NULL && unlink(hp->filepath)) {
588 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
589 __func__, hp->filepath, strerror(errno));
590 }
591 }
592 return 0;
593 }
594
595 /*
596 * unmaps hugepages that are not going to be used. since we originally allocate
597 * ALL hugepages (not just those we need), additional unmapping needs to be done.
598 */
599 static int
unmap_unneeded_hugepages(struct hugepage_file * hugepg_tbl,struct hugepage_info * hpi,unsigned num_hp_info)600 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
601 struct hugepage_info *hpi,
602 unsigned num_hp_info)
603 {
604 unsigned socket, size;
605 int page, nrpages = 0;
606 const struct internal_config *internal_conf =
607 eal_get_internal_configuration();
608
609 /* get total number of hugepages */
610 for (size = 0; size < num_hp_info; size++)
611 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
612 nrpages += internal_conf->hugepage_info[size].num_pages[socket];
613
614 for (size = 0; size < num_hp_info; size++) {
615 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
616 unsigned pages_found = 0;
617
618 /* traverse until we have unmapped all the unused pages */
619 for (page = 0; page < nrpages; page++) {
620 struct hugepage_file *hp = &hugepg_tbl[page];
621
622 /* find a page that matches the criteria */
623 if ((hp->size == hpi[size].hugepage_sz) &&
624 (hp->socket_id == (int) socket)) {
625
626 /* if we skipped enough pages, unmap the rest */
627 if (pages_found == hpi[size].num_pages[socket]) {
628 uint64_t unmap_len;
629
630 unmap_len = hp->size;
631
632 /* get start addr and len of the remaining segment */
633 munmap(hp->orig_va,
634 (size_t)unmap_len);
635
636 hp->orig_va = NULL;
637 if (unlink(hp->filepath) == -1) {
638 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
639 __func__, hp->filepath, strerror(errno));
640 return -1;
641 }
642 } else {
643 /* lock the page and skip */
644 pages_found++;
645 }
646
647 } /* match page */
648 } /* foreach page */
649 } /* foreach socket */
650 } /* foreach pagesize */
651
652 return 0;
653 }
654
655 static int
remap_segment(struct hugepage_file * hugepages,int seg_start,int seg_end)656 remap_segment(struct hugepage_file *hugepages, int seg_start, int seg_end)
657 {
658 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
659 struct rte_memseg_list *msl;
660 struct rte_fbarray *arr;
661 int cur_page, seg_len;
662 unsigned int msl_idx;
663 int ms_idx;
664 uint64_t page_sz;
665 size_t memseg_len;
666 int socket_id;
667 #ifndef RTE_ARCH_64
668 const struct internal_config *internal_conf =
669 eal_get_internal_configuration();
670 #endif
671 page_sz = hugepages[seg_start].size;
672 socket_id = hugepages[seg_start].socket_id;
673 seg_len = seg_end - seg_start;
674
675 RTE_LOG(DEBUG, EAL, "Attempting to map %" PRIu64 "M on socket %i\n",
676 (seg_len * page_sz) >> 20ULL, socket_id);
677
678 /* find free space in memseg lists */
679 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
680 bool empty;
681 msl = &mcfg->memsegs[msl_idx];
682 arr = &msl->memseg_arr;
683
684 if (msl->page_sz != page_sz)
685 continue;
686 if (msl->socket_id != socket_id)
687 continue;
688
689 /* leave space for a hole if array is not empty */
690 empty = arr->count == 0;
691 ms_idx = rte_fbarray_find_next_n_free(arr, 0,
692 seg_len + (empty ? 0 : 1));
693
694 /* memseg list is full? */
695 if (ms_idx < 0)
696 continue;
697
698 /* leave some space between memsegs, they are not IOVA
699 * contiguous, so they shouldn't be VA contiguous either.
700 */
701 if (!empty)
702 ms_idx++;
703 break;
704 }
705 if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
706 RTE_LOG(ERR, EAL, "Could not find space for memseg. Please increase %s and/or %s in configuration.\n",
707 RTE_STR(RTE_MAX_MEMSEG_PER_TYPE),
708 RTE_STR(RTE_MAX_MEM_MB_PER_TYPE));
709 return -1;
710 }
711
712 #ifdef RTE_ARCH_PPC_64
713 /* for PPC64 we go through the list backwards */
714 for (cur_page = seg_end - 1; cur_page >= seg_start;
715 cur_page--, ms_idx++) {
716 #else
717 for (cur_page = seg_start; cur_page < seg_end; cur_page++, ms_idx++) {
718 #endif
719 struct hugepage_file *hfile = &hugepages[cur_page];
720 struct rte_memseg *ms = rte_fbarray_get(arr, ms_idx);
721 void *addr;
722 int fd;
723
724 fd = open(hfile->filepath, O_RDWR);
725 if (fd < 0) {
726 RTE_LOG(ERR, EAL, "Could not open '%s': %s\n",
727 hfile->filepath, strerror(errno));
728 return -1;
729 }
730 /* set shared lock on the file. */
731 if (flock(fd, LOCK_SH) < 0) {
732 RTE_LOG(DEBUG, EAL, "Could not lock '%s': %s\n",
733 hfile->filepath, strerror(errno));
734 close(fd);
735 return -1;
736 }
737 memseg_len = (size_t)page_sz;
738 addr = RTE_PTR_ADD(msl->base_va, ms_idx * memseg_len);
739
740 /* we know this address is already mmapped by memseg list, so
741 * using MAP_FIXED here is safe
742 */
743 addr = mmap(addr, page_sz, PROT_READ | PROT_WRITE,
744 MAP_SHARED | MAP_POPULATE | MAP_FIXED, fd, 0);
745 if (addr == MAP_FAILED) {
746 RTE_LOG(ERR, EAL, "Couldn't remap '%s': %s\n",
747 hfile->filepath, strerror(errno));
748 close(fd);
749 return -1;
750 }
751
752 /* we have a new address, so unmap previous one */
753 #ifndef RTE_ARCH_64
754 /* in 32-bit legacy mode, we have already unmapped the page */
755 if (!internal_conf->legacy_mem)
756 munmap(hfile->orig_va, page_sz);
757 #else
758 munmap(hfile->orig_va, page_sz);
759 #endif
760
761 hfile->orig_va = NULL;
762 hfile->final_va = addr;
763
764 /* rewrite physical addresses in IOVA as VA mode */
765 if (rte_eal_iova_mode() == RTE_IOVA_VA)
766 hfile->physaddr = (uintptr_t)addr;
767
768 /* set up memseg data */
769 ms->addr = addr;
770 ms->hugepage_sz = page_sz;
771 ms->len = memseg_len;
772 ms->iova = hfile->physaddr;
773 ms->socket_id = hfile->socket_id;
774 ms->nchannel = rte_memory_get_nchannel();
775 ms->nrank = rte_memory_get_nrank();
776
777 rte_fbarray_set_used(arr, ms_idx);
778
779 /* store segment fd internally */
780 if (eal_memalloc_set_seg_fd(msl_idx, ms_idx, fd) < 0)
781 RTE_LOG(ERR, EAL, "Could not store segment fd: %s\n",
782 rte_strerror(rte_errno));
783 }
784 RTE_LOG(DEBUG, EAL, "Allocated %" PRIu64 "M on socket %i\n",
785 (seg_len * page_sz) >> 20, socket_id);
786 return 0;
787 }
788
789 static uint64_t
790 get_mem_amount(uint64_t page_sz, uint64_t max_mem)
791 {
792 uint64_t area_sz, max_pages;
793
794 /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */
795 max_pages = RTE_MAX_MEMSEG_PER_LIST;
796 max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem);
797
798 area_sz = RTE_MIN(page_sz * max_pages, max_mem);
799
800 /* make sure the list isn't smaller than the page size */
801 area_sz = RTE_MAX(area_sz, page_sz);
802
803 return RTE_ALIGN(area_sz, page_sz);
804 }
805
806 static int
807 memseg_list_free(struct rte_memseg_list *msl)
808 {
809 if (rte_fbarray_destroy(&msl->memseg_arr)) {
810 RTE_LOG(ERR, EAL, "Cannot destroy memseg list\n");
811 return -1;
812 }
813 memset(msl, 0, sizeof(*msl));
814 return 0;
815 }
816
817 /*
818 * Our VA space is not preallocated yet, so preallocate it here. We need to know
819 * how many segments there are in order to map all pages into one address space,
820 * and leave appropriate holes between segments so that rte_malloc does not
821 * concatenate them into one big segment.
822 *
823 * we also need to unmap original pages to free up address space.
824 */
825 static int __rte_unused
826 prealloc_segments(struct hugepage_file *hugepages, int n_pages)
827 {
828 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
829 int cur_page, seg_start_page, end_seg, new_memseg;
830 unsigned int hpi_idx, socket, i;
831 int n_contig_segs, n_segs;
832 int msl_idx;
833 const struct internal_config *internal_conf =
834 eal_get_internal_configuration();
835
836 /* before we preallocate segments, we need to free up our VA space.
837 * we're not removing files, and we already have information about
838 * PA-contiguousness, so it is safe to unmap everything.
839 */
840 for (cur_page = 0; cur_page < n_pages; cur_page++) {
841 struct hugepage_file *hpi = &hugepages[cur_page];
842 munmap(hpi->orig_va, hpi->size);
843 hpi->orig_va = NULL;
844 }
845
846 /* we cannot know how many page sizes and sockets we have discovered, so
847 * loop over all of them
848 */
849 for (hpi_idx = 0; hpi_idx < internal_conf->num_hugepage_sizes;
850 hpi_idx++) {
851 uint64_t page_sz =
852 internal_conf->hugepage_info[hpi_idx].hugepage_sz;
853
854 for (i = 0; i < rte_socket_count(); i++) {
855 struct rte_memseg_list *msl;
856
857 socket = rte_socket_id_by_idx(i);
858 n_contig_segs = 0;
859 n_segs = 0;
860 seg_start_page = -1;
861
862 for (cur_page = 0; cur_page < n_pages; cur_page++) {
863 struct hugepage_file *prev, *cur;
864 int prev_seg_start_page = -1;
865
866 cur = &hugepages[cur_page];
867 prev = cur_page == 0 ? NULL :
868 &hugepages[cur_page - 1];
869
870 new_memseg = 0;
871 end_seg = 0;
872
873 if (cur->size == 0)
874 end_seg = 1;
875 else if (cur->socket_id != (int) socket)
876 end_seg = 1;
877 else if (cur->size != page_sz)
878 end_seg = 1;
879 else if (cur_page == 0)
880 new_memseg = 1;
881 #ifdef RTE_ARCH_PPC_64
882 /* On PPC64 architecture, the mmap always start
883 * from higher address to lower address. Here,
884 * physical addresses are in descending order.
885 */
886 else if ((prev->physaddr - cur->physaddr) !=
887 cur->size)
888 new_memseg = 1;
889 #else
890 else if ((cur->physaddr - prev->physaddr) !=
891 cur->size)
892 new_memseg = 1;
893 #endif
894 if (new_memseg) {
895 /* if we're already inside a segment,
896 * new segment means end of current one
897 */
898 if (seg_start_page != -1) {
899 end_seg = 1;
900 prev_seg_start_page =
901 seg_start_page;
902 }
903 seg_start_page = cur_page;
904 }
905
906 if (end_seg) {
907 if (prev_seg_start_page != -1) {
908 /* we've found a new segment */
909 n_contig_segs++;
910 n_segs += cur_page -
911 prev_seg_start_page;
912 } else if (seg_start_page != -1) {
913 /* we didn't find new segment,
914 * but did end current one
915 */
916 n_contig_segs++;
917 n_segs += cur_page -
918 seg_start_page;
919 seg_start_page = -1;
920 continue;
921 } else {
922 /* we're skipping this page */
923 continue;
924 }
925 }
926 /* segment continues */
927 }
928 /* check if we missed last segment */
929 if (seg_start_page != -1) {
930 n_contig_segs++;
931 n_segs += cur_page - seg_start_page;
932 }
933
934 /* if no segments were found, do not preallocate */
935 if (n_segs == 0)
936 continue;
937
938 /* we now have total number of pages that we will
939 * allocate for this segment list. add separator pages
940 * to the total count, and preallocate VA space.
941 */
942 n_segs += n_contig_segs - 1;
943
944 /* now, preallocate VA space for these segments */
945
946 /* first, find suitable memseg list for this */
947 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS;
948 msl_idx++) {
949 msl = &mcfg->memsegs[msl_idx];
950
951 if (msl->base_va != NULL)
952 continue;
953 break;
954 }
955 if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
956 RTE_LOG(ERR, EAL, "Not enough space in memseg lists, please increase %s\n",
957 RTE_STR(RTE_MAX_MEMSEG_LISTS));
958 return -1;
959 }
960
961 /* now, allocate fbarray itself */
962 if (eal_memseg_list_init(msl, page_sz, n_segs,
963 socket, msl_idx, true) < 0)
964 return -1;
965
966 /* finally, allocate VA space */
967 if (eal_memseg_list_alloc(msl, 0) < 0) {
968 RTE_LOG(ERR, EAL, "Cannot preallocate 0x%"PRIx64"kB hugepages\n",
969 page_sz >> 10);
970 return -1;
971 }
972 }
973 }
974 return 0;
975 }
976
977 /*
978 * We cannot reallocate memseg lists on the fly because PPC64 stores pages
979 * backwards, therefore we have to process the entire memseg first before
980 * remapping it into memseg list VA space.
981 */
982 static int
983 remap_needed_hugepages(struct hugepage_file *hugepages, int n_pages)
984 {
985 int cur_page, seg_start_page, new_memseg, ret;
986
987 seg_start_page = 0;
988 for (cur_page = 0; cur_page < n_pages; cur_page++) {
989 struct hugepage_file *prev, *cur;
990
991 new_memseg = 0;
992
993 cur = &hugepages[cur_page];
994 prev = cur_page == 0 ? NULL : &hugepages[cur_page - 1];
995
996 /* if size is zero, no more pages left */
997 if (cur->size == 0)
998 break;
999
1000 if (cur_page == 0)
1001 new_memseg = 1;
1002 else if (cur->socket_id != prev->socket_id)
1003 new_memseg = 1;
1004 else if (cur->size != prev->size)
1005 new_memseg = 1;
1006 #ifdef RTE_ARCH_PPC_64
1007 /* On PPC64 architecture, the mmap always start from higher
1008 * address to lower address. Here, physical addresses are in
1009 * descending order.
1010 */
1011 else if ((prev->physaddr - cur->physaddr) != cur->size)
1012 new_memseg = 1;
1013 #else
1014 else if ((cur->physaddr - prev->physaddr) != cur->size)
1015 new_memseg = 1;
1016 #endif
1017
1018 if (new_memseg) {
1019 /* if this isn't the first time, remap segment */
1020 if (cur_page != 0) {
1021 ret = remap_segment(hugepages, seg_start_page,
1022 cur_page);
1023 if (ret != 0)
1024 return -1;
1025 }
1026 /* remember where we started */
1027 seg_start_page = cur_page;
1028 }
1029 /* continuation of previous memseg */
1030 }
1031 /* we were stopped, but we didn't remap the last segment, do it now */
1032 if (cur_page != 0) {
1033 ret = remap_segment(hugepages, seg_start_page,
1034 cur_page);
1035 if (ret != 0)
1036 return -1;
1037 }
1038 return 0;
1039 }
1040
1041 static inline size_t
1042 eal_get_hugepage_mem_size(void)
1043 {
1044 uint64_t size = 0;
1045 unsigned i, j;
1046 struct internal_config *internal_conf =
1047 eal_get_internal_configuration();
1048
1049 for (i = 0; i < internal_conf->num_hugepage_sizes; i++) {
1050 struct hugepage_info *hpi = &internal_conf->hugepage_info[i];
1051 if (strnlen(hpi->hugedir, sizeof(hpi->hugedir)) != 0) {
1052 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1053 size += hpi->hugepage_sz * hpi->num_pages[j];
1054 }
1055 }
1056 }
1057
1058 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
1059 }
1060
1061 static struct sigaction huge_action_old;
1062 static int huge_need_recover;
1063
1064 static void
1065 huge_register_sigbus(void)
1066 {
1067 sigset_t mask;
1068 struct sigaction action;
1069
1070 sigemptyset(&mask);
1071 sigaddset(&mask, SIGBUS);
1072 action.sa_flags = 0;
1073 action.sa_mask = mask;
1074 action.sa_handler = huge_sigbus_handler;
1075
1076 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1077 }
1078
1079 static void
1080 huge_recover_sigbus(void)
1081 {
1082 if (huge_need_recover) {
1083 sigaction(SIGBUS, &huge_action_old, NULL);
1084 huge_need_recover = 0;
1085 }
1086 }
1087
1088 /*
1089 * Prepare physical memory mapping: fill configuration structure with
1090 * these infos, return 0 on success.
1091 * 1. map N huge pages in separate files in hugetlbfs
1092 * 2. find associated physical addr
1093 * 3. find associated NUMA socket ID
1094 * 4. sort all huge pages by physical address
1095 * 5. remap these N huge pages in the correct order
1096 * 6. unmap the first mapping
1097 * 7. fill memsegs in configuration with contiguous zones
1098 */
1099 static int
1100 eal_legacy_hugepage_init(void)
1101 {
1102 struct rte_mem_config *mcfg;
1103 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1104 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1105 struct internal_config *internal_conf =
1106 eal_get_internal_configuration();
1107
1108 uint64_t memory[RTE_MAX_NUMA_NODES];
1109
1110 unsigned hp_offset;
1111 int i, j;
1112 int nr_hugefiles, nr_hugepages = 0;
1113 void *addr;
1114
1115 memset(used_hp, 0, sizeof(used_hp));
1116
1117 /* get pointer to global configuration */
1118 mcfg = rte_eal_get_configuration()->mem_config;
1119
1120 /* hugetlbfs can be disabled */
1121 if (internal_conf->no_hugetlbfs) {
1122 void *prealloc_addr;
1123 size_t mem_sz;
1124 struct rte_memseg_list *msl;
1125 int n_segs, fd, flags;
1126 #ifdef MEMFD_SUPPORTED
1127 int memfd;
1128 #endif
1129 uint64_t page_sz;
1130
1131 /* nohuge mode is legacy mode */
1132 internal_conf->legacy_mem = 1;
1133
1134 /* nohuge mode is single-file segments mode */
1135 internal_conf->single_file_segments = 1;
1136
1137 /* create a memseg list */
1138 msl = &mcfg->memsegs[0];
1139
1140 mem_sz = internal_conf->memory;
1141 page_sz = RTE_PGSIZE_4K;
1142 n_segs = mem_sz / page_sz;
1143
1144 if (eal_memseg_list_init_named(
1145 msl, "nohugemem", page_sz, n_segs, 0, true)) {
1146 return -1;
1147 }
1148
1149 /* set up parameters for anonymous mmap */
1150 fd = -1;
1151 flags = MAP_PRIVATE | MAP_ANONYMOUS;
1152
1153 #ifdef MEMFD_SUPPORTED
1154 /* create a memfd and store it in the segment fd table */
1155 memfd = memfd_create("nohuge", 0);
1156 if (memfd < 0) {
1157 RTE_LOG(DEBUG, EAL, "Cannot create memfd: %s\n",
1158 strerror(errno));
1159 RTE_LOG(DEBUG, EAL, "Falling back to anonymous map\n");
1160 } else {
1161 /* we got an fd - now resize it */
1162 if (ftruncate(memfd, internal_conf->memory) < 0) {
1163 RTE_LOG(ERR, EAL, "Cannot resize memfd: %s\n",
1164 strerror(errno));
1165 RTE_LOG(ERR, EAL, "Falling back to anonymous map\n");
1166 close(memfd);
1167 } else {
1168 /* creating memfd-backed file was successful.
1169 * we want changes to memfd to be visible to
1170 * other processes (such as vhost backend), so
1171 * map it as shared memory.
1172 */
1173 RTE_LOG(DEBUG, EAL, "Using memfd for anonymous memory\n");
1174 fd = memfd;
1175 flags = MAP_SHARED;
1176 }
1177 }
1178 #endif
1179 /* preallocate address space for the memory, so that it can be
1180 * fit into the DMA mask.
1181 */
1182 if (eal_memseg_list_alloc(msl, 0)) {
1183 RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
1184 return -1;
1185 }
1186
1187 prealloc_addr = msl->base_va;
1188 addr = mmap(prealloc_addr, mem_sz, PROT_READ | PROT_WRITE,
1189 flags | MAP_FIXED, fd, 0);
1190 if (addr == MAP_FAILED || addr != prealloc_addr) {
1191 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1192 strerror(errno));
1193 munmap(prealloc_addr, mem_sz);
1194 return -1;
1195 }
1196
1197 /* we're in single-file segments mode, so only the segment list
1198 * fd needs to be set up.
1199 */
1200 if (fd != -1) {
1201 if (eal_memalloc_set_seg_list_fd(0, fd) < 0) {
1202 RTE_LOG(ERR, EAL, "Cannot set up segment list fd\n");
1203 /* not a serious error, proceed */
1204 }
1205 }
1206
1207 eal_memseg_list_populate(msl, addr, n_segs);
1208
1209 if (mcfg->dma_maskbits &&
1210 rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
1211 RTE_LOG(ERR, EAL,
1212 "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask.\n",
1213 __func__);
1214 if (rte_eal_iova_mode() == RTE_IOVA_VA &&
1215 rte_eal_using_phys_addrs())
1216 RTE_LOG(ERR, EAL,
1217 "%s(): Please try initializing EAL with --iova-mode=pa parameter.\n",
1218 __func__);
1219 goto fail;
1220 }
1221 return 0;
1222 }
1223
1224 /* calculate total number of hugepages available. at this point we haven't
1225 * yet started sorting them so they all are on socket 0 */
1226 for (i = 0; i < (int) internal_conf->num_hugepage_sizes; i++) {
1227 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1228 used_hp[i].hugepage_sz = internal_conf->hugepage_info[i].hugepage_sz;
1229
1230 nr_hugepages += internal_conf->hugepage_info[i].num_pages[0];
1231 }
1232
1233 /*
1234 * allocate a memory area for hugepage table.
1235 * this isn't shared memory yet. due to the fact that we need some
1236 * processing done on these pages, shared memory will be created
1237 * at a later stage.
1238 */
1239 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1240 if (tmp_hp == NULL)
1241 goto fail;
1242
1243 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1244
1245 hp_offset = 0; /* where we start the current page size entries */
1246
1247 huge_register_sigbus();
1248
1249 /* make a copy of socket_mem, needed for balanced allocation. */
1250 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1251 memory[i] = internal_conf->socket_mem[i];
1252
1253 /* map all hugepages and sort them */
1254 for (i = 0; i < (int)internal_conf->num_hugepage_sizes; i++) {
1255 unsigned pages_old, pages_new;
1256 struct hugepage_info *hpi;
1257
1258 /*
1259 * we don't yet mark hugepages as used at this stage, so
1260 * we just map all hugepages available to the system
1261 * all hugepages are still located on socket 0
1262 */
1263 hpi = &internal_conf->hugepage_info[i];
1264
1265 if (hpi->num_pages[0] == 0)
1266 continue;
1267
1268 /* map all hugepages available */
1269 pages_old = hpi->num_pages[0];
1270 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, memory);
1271 if (pages_new < pages_old) {
1272 RTE_LOG(DEBUG, EAL,
1273 "%d not %d hugepages of size %u MB allocated\n",
1274 pages_new, pages_old,
1275 (unsigned)(hpi->hugepage_sz / 0x100000));
1276
1277 int pages = pages_old - pages_new;
1278
1279 nr_hugepages -= pages;
1280 hpi->num_pages[0] = pages_new;
1281 if (pages_new == 0)
1282 continue;
1283 }
1284
1285 if (rte_eal_using_phys_addrs() &&
1286 rte_eal_iova_mode() != RTE_IOVA_VA) {
1287 /* find physical addresses for each hugepage */
1288 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1289 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1290 "for %u MB pages\n",
1291 (unsigned int)(hpi->hugepage_sz / 0x100000));
1292 goto fail;
1293 }
1294 } else {
1295 /* set physical addresses for each hugepage */
1296 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1297 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1298 "for %u MB pages\n",
1299 (unsigned int)(hpi->hugepage_sz / 0x100000));
1300 goto fail;
1301 }
1302 }
1303
1304 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1305 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1306 (unsigned)(hpi->hugepage_sz / 0x100000));
1307 goto fail;
1308 }
1309
1310 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1311 sizeof(struct hugepage_file), cmp_physaddr);
1312
1313 /* we have processed a num of hugepages of this size, so inc offset */
1314 hp_offset += hpi->num_pages[0];
1315 }
1316
1317 huge_recover_sigbus();
1318
1319 if (internal_conf->memory == 0 && internal_conf->force_sockets == 0)
1320 internal_conf->memory = eal_get_hugepage_mem_size();
1321
1322 nr_hugefiles = nr_hugepages;
1323
1324
1325 /* clean out the numbers of pages */
1326 for (i = 0; i < (int) internal_conf->num_hugepage_sizes; i++)
1327 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1328 internal_conf->hugepage_info[i].num_pages[j] = 0;
1329
1330 /* get hugepages for each socket */
1331 for (i = 0; i < nr_hugefiles; i++) {
1332 int socket = tmp_hp[i].socket_id;
1333
1334 /* find a hugepage info with right size and increment num_pages */
1335 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1336 (int)internal_conf->num_hugepage_sizes);
1337 for (j = 0; j < nb_hpsizes; j++) {
1338 if (tmp_hp[i].size ==
1339 internal_conf->hugepage_info[j].hugepage_sz) {
1340 internal_conf->hugepage_info[j].num_pages[socket]++;
1341 }
1342 }
1343 }
1344
1345 /* make a copy of socket_mem, needed for number of pages calculation */
1346 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1347 memory[i] = internal_conf->socket_mem[i];
1348
1349 /* calculate final number of pages */
1350 nr_hugepages = eal_dynmem_calc_num_pages_per_socket(memory,
1351 internal_conf->hugepage_info, used_hp,
1352 internal_conf->num_hugepage_sizes);
1353
1354 /* error if not enough memory available */
1355 if (nr_hugepages < 0)
1356 goto fail;
1357
1358 /* reporting in! */
1359 for (i = 0; i < (int) internal_conf->num_hugepage_sizes; i++) {
1360 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1361 if (used_hp[i].num_pages[j] > 0) {
1362 RTE_LOG(DEBUG, EAL,
1363 "Requesting %u pages of size %uMB"
1364 " from socket %i\n",
1365 used_hp[i].num_pages[j],
1366 (unsigned)
1367 (used_hp[i].hugepage_sz / 0x100000),
1368 j);
1369 }
1370 }
1371 }
1372
1373 /* create shared memory */
1374 hugepage = create_shared_memory(eal_hugepage_data_path(),
1375 nr_hugefiles * sizeof(struct hugepage_file));
1376
1377 if (hugepage == NULL) {
1378 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1379 goto fail;
1380 }
1381 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1382
1383 /*
1384 * unmap pages that we won't need (looks at used_hp).
1385 * also, sets final_va to NULL on pages that were unmapped.
1386 */
1387 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1388 internal_conf->num_hugepage_sizes) < 0) {
1389 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1390 goto fail;
1391 }
1392
1393 /*
1394 * copy stuff from malloc'd hugepage* to the actual shared memory.
1395 * this procedure only copies those hugepages that have orig_va
1396 * not NULL. has overflow protection.
1397 */
1398 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1399 tmp_hp, nr_hugefiles) < 0) {
1400 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1401 goto fail;
1402 }
1403
1404 #ifndef RTE_ARCH_64
1405 /* for legacy 32-bit mode, we did not preallocate VA space, so do it */
1406 if (internal_conf->legacy_mem &&
1407 prealloc_segments(hugepage, nr_hugefiles)) {
1408 RTE_LOG(ERR, EAL, "Could not preallocate VA space for hugepages\n");
1409 goto fail;
1410 }
1411 #endif
1412
1413 /* remap all pages we do need into memseg list VA space, so that those
1414 * pages become first-class citizens in DPDK memory subsystem
1415 */
1416 if (remap_needed_hugepages(hugepage, nr_hugefiles)) {
1417 RTE_LOG(ERR, EAL, "Couldn't remap hugepage files into memseg lists\n");
1418 goto fail;
1419 }
1420
1421 /* free the hugepage backing files */
1422 if (internal_conf->hugepage_file.unlink_before_mapping &&
1423 unlink_hugepage_files(tmp_hp, internal_conf->num_hugepage_sizes) < 0) {
1424 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1425 goto fail;
1426 }
1427
1428 /* free the temporary hugepage table */
1429 free(tmp_hp);
1430 tmp_hp = NULL;
1431
1432 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1433 hugepage = NULL;
1434
1435 /* we're not going to allocate more pages, so release VA space for
1436 * unused memseg lists
1437 */
1438 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1439 struct rte_memseg_list *msl = &mcfg->memsegs[i];
1440 size_t mem_sz;
1441
1442 /* skip inactive lists */
1443 if (msl->base_va == NULL)
1444 continue;
1445 /* skip lists where there is at least one page allocated */
1446 if (msl->memseg_arr.count > 0)
1447 continue;
1448 /* this is an unused list, deallocate it */
1449 mem_sz = msl->len;
1450 munmap(msl->base_va, mem_sz);
1451 msl->base_va = NULL;
1452 msl->heap = 0;
1453
1454 /* destroy backing fbarray */
1455 rte_fbarray_destroy(&msl->memseg_arr);
1456 }
1457
1458 if (mcfg->dma_maskbits &&
1459 rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
1460 RTE_LOG(ERR, EAL,
1461 "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask.\n",
1462 __func__);
1463 goto fail;
1464 }
1465
1466 return 0;
1467
1468 fail:
1469 huge_recover_sigbus();
1470 free(tmp_hp);
1471 if (hugepage != NULL)
1472 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1473
1474 return -1;
1475 }
1476
1477 /*
1478 * uses fstat to report the size of a file on disk
1479 */
1480 static off_t
1481 getFileSize(int fd)
1482 {
1483 struct stat st;
1484 if (fstat(fd, &st) < 0)
1485 return 0;
1486 return st.st_size;
1487 }
1488
1489 /*
1490 * This creates the memory mappings in the secondary process to match that of
1491 * the server process. It goes through each memory segment in the DPDK runtime
1492 * configuration and finds the hugepages which form that segment, mapping them
1493 * in order to form a contiguous block in the virtual memory space
1494 */
1495 static int
1496 eal_legacy_hugepage_attach(void)
1497 {
1498 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1499 struct hugepage_file *hp = NULL;
1500 unsigned int num_hp = 0;
1501 unsigned int i = 0;
1502 unsigned int cur_seg;
1503 off_t size = 0;
1504 int fd, fd_hugepage = -1;
1505
1506 if (aslr_enabled() > 0) {
1507 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1508 "(ASLR) is enabled in the kernel.\n");
1509 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1510 "into secondary processes\n");
1511 }
1512
1513 fd_hugepage = open(eal_hugepage_data_path(), O_RDONLY);
1514 if (fd_hugepage < 0) {
1515 RTE_LOG(ERR, EAL, "Could not open %s\n",
1516 eal_hugepage_data_path());
1517 goto error;
1518 }
1519
1520 size = getFileSize(fd_hugepage);
1521 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1522 if (hp == MAP_FAILED) {
1523 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1524 eal_hugepage_data_path());
1525 goto error;
1526 }
1527
1528 num_hp = size / sizeof(struct hugepage_file);
1529 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1530
1531 /* map all segments into memory to make sure we get the addrs. the
1532 * segments themselves are already in memseg list (which is shared and
1533 * has its VA space already preallocated), so we just need to map
1534 * everything into correct addresses.
1535 */
1536 for (i = 0; i < num_hp; i++) {
1537 struct hugepage_file *hf = &hp[i];
1538 size_t map_sz = hf->size;
1539 void *map_addr = hf->final_va;
1540 int msl_idx, ms_idx;
1541 struct rte_memseg_list *msl;
1542 struct rte_memseg *ms;
1543
1544 /* if size is zero, no more pages left */
1545 if (map_sz == 0)
1546 break;
1547
1548 fd = open(hf->filepath, O_RDWR);
1549 if (fd < 0) {
1550 RTE_LOG(ERR, EAL, "Could not open %s: %s\n",
1551 hf->filepath, strerror(errno));
1552 goto error;
1553 }
1554
1555 map_addr = mmap(map_addr, map_sz, PROT_READ | PROT_WRITE,
1556 MAP_SHARED | MAP_FIXED, fd, 0);
1557 if (map_addr == MAP_FAILED) {
1558 RTE_LOG(ERR, EAL, "Could not map %s: %s\n",
1559 hf->filepath, strerror(errno));
1560 goto fd_error;
1561 }
1562
1563 /* set shared lock on the file. */
1564 if (flock(fd, LOCK_SH) < 0) {
1565 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed: %s\n",
1566 __func__, strerror(errno));
1567 goto mmap_error;
1568 }
1569
1570 /* find segment data */
1571 msl = rte_mem_virt2memseg_list(map_addr);
1572 if (msl == NULL) {
1573 RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg list\n",
1574 __func__);
1575 goto mmap_error;
1576 }
1577 ms = rte_mem_virt2memseg(map_addr, msl);
1578 if (ms == NULL) {
1579 RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg\n",
1580 __func__);
1581 goto mmap_error;
1582 }
1583
1584 msl_idx = msl - mcfg->memsegs;
1585 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
1586 if (ms_idx < 0) {
1587 RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg idx\n",
1588 __func__);
1589 goto mmap_error;
1590 }
1591
1592 /* store segment fd internally */
1593 if (eal_memalloc_set_seg_fd(msl_idx, ms_idx, fd) < 0)
1594 RTE_LOG(ERR, EAL, "Could not store segment fd: %s\n",
1595 rte_strerror(rte_errno));
1596 }
1597 /* unmap the hugepage config file, since we are done using it */
1598 munmap(hp, size);
1599 close(fd_hugepage);
1600 return 0;
1601
1602 mmap_error:
1603 munmap(hp[i].final_va, hp[i].size);
1604 fd_error:
1605 close(fd);
1606 error:
1607 /* unwind mmap's done so far */
1608 for (cur_seg = 0; cur_seg < i; cur_seg++)
1609 munmap(hp[cur_seg].final_va, hp[cur_seg].size);
1610
1611 if (hp != NULL && hp != MAP_FAILED)
1612 munmap(hp, size);
1613 if (fd_hugepage >= 0)
1614 close(fd_hugepage);
1615 return -1;
1616 }
1617
1618 static int
1619 eal_hugepage_attach(void)
1620 {
1621 if (eal_memalloc_sync_with_primary()) {
1622 RTE_LOG(ERR, EAL, "Could not map memory from primary process\n");
1623 if (aslr_enabled() > 0)
1624 RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel and retry running both primary and secondary processes\n");
1625 return -1;
1626 }
1627 return 0;
1628 }
1629
1630 int
1631 rte_eal_hugepage_init(void)
1632 {
1633 const struct internal_config *internal_conf =
1634 eal_get_internal_configuration();
1635
1636 return internal_conf->legacy_mem ?
1637 eal_legacy_hugepage_init() :
1638 eal_dynmem_hugepage_init();
1639 }
1640
1641 int
1642 rte_eal_hugepage_attach(void)
1643 {
1644 const struct internal_config *internal_conf =
1645 eal_get_internal_configuration();
1646
1647 return internal_conf->legacy_mem ?
1648 eal_legacy_hugepage_attach() :
1649 eal_hugepage_attach();
1650 }
1651
1652 int
1653 rte_eal_using_phys_addrs(void)
1654 {
1655 if (phys_addrs_available == -1) {
1656 uint64_t tmp = 0;
1657
1658 if (rte_eal_has_hugepages() != 0 &&
1659 rte_mem_virt2phy(&tmp) != RTE_BAD_PHYS_ADDR)
1660 phys_addrs_available = 1;
1661 else
1662 phys_addrs_available = 0;
1663 }
1664 return phys_addrs_available;
1665 }
1666
1667 static int __rte_unused
1668 memseg_primary_init_32(void)
1669 {
1670 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1671 int active_sockets, hpi_idx, msl_idx = 0;
1672 unsigned int socket_id, i;
1673 struct rte_memseg_list *msl;
1674 uint64_t extra_mem_per_socket, total_extra_mem, total_requested_mem;
1675 uint64_t max_mem;
1676 struct internal_config *internal_conf =
1677 eal_get_internal_configuration();
1678
1679 /* no-huge does not need this at all */
1680 if (internal_conf->no_hugetlbfs)
1681 return 0;
1682
1683 /* this is a giant hack, but desperate times call for desperate
1684 * measures. in legacy 32-bit mode, we cannot preallocate VA space,
1685 * because having upwards of 2 gigabytes of VA space already mapped will
1686 * interfere with our ability to map and sort hugepages.
1687 *
1688 * therefore, in legacy 32-bit mode, we will be initializing memseg
1689 * lists much later - in eal_memory.c, right after we unmap all the
1690 * unneeded pages. this will not affect secondary processes, as those
1691 * should be able to mmap the space without (too many) problems.
1692 */
1693 if (internal_conf->legacy_mem)
1694 return 0;
1695
1696 /* 32-bit mode is a very special case. we cannot know in advance where
1697 * the user will want to allocate their memory, so we have to do some
1698 * heuristics.
1699 */
1700 active_sockets = 0;
1701 total_requested_mem = 0;
1702 if (internal_conf->force_sockets)
1703 for (i = 0; i < rte_socket_count(); i++) {
1704 uint64_t mem;
1705
1706 socket_id = rte_socket_id_by_idx(i);
1707 mem = internal_conf->socket_mem[socket_id];
1708
1709 if (mem == 0)
1710 continue;
1711
1712 active_sockets++;
1713 total_requested_mem += mem;
1714 }
1715 else
1716 total_requested_mem = internal_conf->memory;
1717
1718 max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
1719 if (total_requested_mem > max_mem) {
1720 RTE_LOG(ERR, EAL, "Invalid parameters: 32-bit process can at most use %uM of memory\n",
1721 (unsigned int)(max_mem >> 20));
1722 return -1;
1723 }
1724 total_extra_mem = max_mem - total_requested_mem;
1725 extra_mem_per_socket = active_sockets == 0 ? total_extra_mem :
1726 total_extra_mem / active_sockets;
1727
1728 /* the allocation logic is a little bit convoluted, but here's how it
1729 * works, in a nutshell:
1730 * - if user hasn't specified on which sockets to allocate memory via
1731 * --socket-mem, we allocate all of our memory on main core socket.
1732 * - if user has specified sockets to allocate memory on, there may be
1733 * some "unused" memory left (e.g. if user has specified --socket-mem
1734 * such that not all memory adds up to 2 gigabytes), so add it to all
1735 * sockets that are in use equally.
1736 *
1737 * page sizes are sorted by size in descending order, so we can safely
1738 * assume that we dispense with bigger page sizes first.
1739 */
1740
1741 /* create memseg lists */
1742 for (i = 0; i < rte_socket_count(); i++) {
1743 int hp_sizes = (int) internal_conf->num_hugepage_sizes;
1744 uint64_t max_socket_mem, cur_socket_mem;
1745 unsigned int main_lcore_socket;
1746 struct rte_config *cfg = rte_eal_get_configuration();
1747 bool skip;
1748
1749 socket_id = rte_socket_id_by_idx(i);
1750
1751 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
1752 /* we can still sort pages by socket in legacy mode */
1753 if (!internal_conf->legacy_mem && socket_id > 0)
1754 break;
1755 #endif
1756
1757 /* if we didn't specifically request memory on this socket */
1758 skip = active_sockets != 0 &&
1759 internal_conf->socket_mem[socket_id] == 0;
1760 /* ...or if we didn't specifically request memory on *any*
1761 * socket, and this is not main lcore
1762 */
1763 main_lcore_socket = rte_lcore_to_socket_id(cfg->main_lcore);
1764 skip |= active_sockets == 0 && socket_id != main_lcore_socket;
1765
1766 if (skip) {
1767 RTE_LOG(DEBUG, EAL, "Will not preallocate memory on socket %u\n",
1768 socket_id);
1769 continue;
1770 }
1771
1772 /* max amount of memory on this socket */
1773 max_socket_mem = (active_sockets != 0 ?
1774 internal_conf->socket_mem[socket_id] :
1775 internal_conf->memory) +
1776 extra_mem_per_socket;
1777 cur_socket_mem = 0;
1778
1779 for (hpi_idx = 0; hpi_idx < hp_sizes; hpi_idx++) {
1780 uint64_t max_pagesz_mem, cur_pagesz_mem = 0;
1781 uint64_t hugepage_sz;
1782 struct hugepage_info *hpi;
1783 int type_msl_idx, max_segs, total_segs = 0;
1784
1785 hpi = &internal_conf->hugepage_info[hpi_idx];
1786 hugepage_sz = hpi->hugepage_sz;
1787
1788 /* check if pages are actually available */
1789 if (hpi->num_pages[socket_id] == 0)
1790 continue;
1791
1792 max_segs = RTE_MAX_MEMSEG_PER_TYPE;
1793 max_pagesz_mem = max_socket_mem - cur_socket_mem;
1794
1795 /* make it multiple of page size */
1796 max_pagesz_mem = RTE_ALIGN_FLOOR(max_pagesz_mem,
1797 hugepage_sz);
1798
1799 RTE_LOG(DEBUG, EAL, "Attempting to preallocate "
1800 "%" PRIu64 "M on socket %i\n",
1801 max_pagesz_mem >> 20, socket_id);
1802
1803 type_msl_idx = 0;
1804 while (cur_pagesz_mem < max_pagesz_mem &&
1805 total_segs < max_segs) {
1806 uint64_t cur_mem;
1807 unsigned int n_segs;
1808
1809 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
1810 RTE_LOG(ERR, EAL,
1811 "No more space in memseg lists, please increase %s\n",
1812 RTE_STR(RTE_MAX_MEMSEG_LISTS));
1813 return -1;
1814 }
1815
1816 msl = &mcfg->memsegs[msl_idx];
1817
1818 cur_mem = get_mem_amount(hugepage_sz,
1819 max_pagesz_mem);
1820 n_segs = cur_mem / hugepage_sz;
1821
1822 if (eal_memseg_list_init(msl, hugepage_sz,
1823 n_segs, socket_id, type_msl_idx,
1824 true)) {
1825 /* failing to allocate a memseg list is
1826 * a serious error.
1827 */
1828 RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
1829 return -1;
1830 }
1831
1832 if (eal_memseg_list_alloc(msl, 0)) {
1833 /* if we couldn't allocate VA space, we
1834 * can try with smaller page sizes.
1835 */
1836 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list, retrying with different page size\n");
1837 /* deallocate memseg list */
1838 if (memseg_list_free(msl))
1839 return -1;
1840 break;
1841 }
1842
1843 total_segs += msl->memseg_arr.len;
1844 cur_pagesz_mem = total_segs * hugepage_sz;
1845 type_msl_idx++;
1846 msl_idx++;
1847 }
1848 cur_socket_mem += cur_pagesz_mem;
1849 }
1850 if (cur_socket_mem == 0) {
1851 RTE_LOG(ERR, EAL, "Cannot allocate VA space on socket %u\n",
1852 socket_id);
1853 return -1;
1854 }
1855 }
1856
1857 return 0;
1858 }
1859
1860 static int __rte_unused
1861 memseg_primary_init(void)
1862 {
1863 return eal_dynmem_memseg_lists_init();
1864 }
1865
1866 static int
1867 memseg_secondary_init(void)
1868 {
1869 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1870 int msl_idx = 0;
1871 struct rte_memseg_list *msl;
1872
1873 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
1874
1875 msl = &mcfg->memsegs[msl_idx];
1876
1877 /* skip empty and external memseg lists */
1878 if (msl->memseg_arr.len == 0 || msl->external)
1879 continue;
1880
1881 if (rte_fbarray_attach(&msl->memseg_arr)) {
1882 RTE_LOG(ERR, EAL, "Cannot attach to primary process memseg lists\n");
1883 return -1;
1884 }
1885
1886 /* preallocate VA space */
1887 if (eal_memseg_list_alloc(msl, 0)) {
1888 RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
1889 return -1;
1890 }
1891 }
1892
1893 return 0;
1894 }
1895
1896 int
1897 rte_eal_memseg_init(void)
1898 {
1899 /* increase rlimit to maximum */
1900 struct rlimit lim;
1901
1902 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
1903 const struct internal_config *internal_conf =
1904 eal_get_internal_configuration();
1905 #endif
1906 if (getrlimit(RLIMIT_NOFILE, &lim) == 0) {
1907 /* set limit to maximum */
1908 lim.rlim_cur = lim.rlim_max;
1909
1910 if (setrlimit(RLIMIT_NOFILE, &lim) < 0) {
1911 RTE_LOG(DEBUG, EAL, "Setting maximum number of open files failed: %s\n",
1912 strerror(errno));
1913 } else {
1914 RTE_LOG(DEBUG, EAL, "Setting maximum number of open files to %"
1915 PRIu64 "\n",
1916 (uint64_t)lim.rlim_cur);
1917 }
1918 } else {
1919 RTE_LOG(ERR, EAL, "Cannot get current resource limits\n");
1920 }
1921 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
1922 if (!internal_conf->legacy_mem && rte_socket_count() > 1) {
1923 RTE_LOG(WARNING, EAL, "DPDK is running on a NUMA system, but is compiled without NUMA support.\n");
1924 RTE_LOG(WARNING, EAL, "This will have adverse consequences for performance and usability.\n");
1925 RTE_LOG(WARNING, EAL, "Please use --"OPT_LEGACY_MEM" option, or recompile with NUMA support.\n");
1926 }
1927 #endif
1928
1929 return rte_eal_process_type() == RTE_PROC_PRIMARY ?
1930 #ifndef RTE_ARCH_64
1931 memseg_primary_init_32() :
1932 #else
1933 memseg_primary_init() :
1934 #endif
1935 memseg_secondary_init();
1936 }
1937