1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation.
3 * Copyright(c) 2013 6WIND S.A.
6 #define _FILE_OFFSET_BITS 64
16 #include <sys/types.h>
18 #include <sys/queue.h>
22 #include <sys/ioctl.h>
26 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
31 #include <rte_errno.h>
33 #include <rte_memory.h>
34 #include <rte_launch.h>
36 #include <rte_eal_memconfig.h>
37 #include <rte_per_lcore.h>
38 #include <rte_lcore.h>
39 #include <rte_common.h>
40 #include <rte_string_fns.h>
42 #include "eal_private.h"
43 #include "eal_memalloc.h"
44 #include "eal_internal_cfg.h"
45 #include "eal_filesystem.h"
46 #include "eal_hugepages.h"
48 #define PFN_MASK_SIZE 8
52 * Huge page mapping under linux
54 * To reserve a big contiguous amount of memory, we use the hugepage
55 * feature of linux. For that, we need to have hugetlbfs mounted. This
56 * code will create many files in this directory (one per page) and
57 * map them in virtual memory. For each page, we will retrieve its
58 * physical address and remap it in order to have a virtual contiguous
59 * zone as well as a physical contiguous zone.
62 static bool phys_addrs_available = true;
64 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
67 test_phys_addrs_available(void)
72 if (!rte_eal_has_hugepages()) {
74 "Started without hugepages support, physical addresses not available\n");
75 phys_addrs_available = false;
79 physaddr = rte_mem_virt2phy(&tmp);
80 if (physaddr == RTE_BAD_PHYS_ADDR) {
81 if (rte_eal_iova_mode() == RTE_IOVA_PA)
83 "Cannot obtain physical addresses: %s. "
84 "Only vfio will function.\n",
86 phys_addrs_available = false;
91 * Get physical address of any mapped virtual address in the current process.
94 rte_mem_virt2phy(const void *virtaddr)
97 uint64_t page, physaddr;
98 unsigned long virt_pfn;
102 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
103 if (!phys_addrs_available)
106 /* standard page size */
107 page_size = getpagesize();
109 fd = open("/proc/self/pagemap", O_RDONLY);
111 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
112 __func__, strerror(errno));
116 virt_pfn = (unsigned long)virtaddr / page_size;
117 offset = sizeof(uint64_t) * virt_pfn;
118 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
119 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
120 __func__, strerror(errno));
125 retval = read(fd, &page, PFN_MASK_SIZE);
128 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
129 __func__, strerror(errno));
131 } else if (retval != PFN_MASK_SIZE) {
132 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
133 "but expected %d:\n",
134 __func__, retval, PFN_MASK_SIZE);
139 * the pfn (page frame number) are bits 0-54 (see
140 * pagemap.txt in linux Documentation)
142 if ((page & 0x7fffffffffffffULL) == 0)
145 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
146 + ((unsigned long)virtaddr % page_size);
152 rte_mem_virt2iova(const void *virtaddr)
154 if (rte_eal_iova_mode() == RTE_IOVA_VA)
155 return (uintptr_t)virtaddr;
156 return rte_mem_virt2phy(virtaddr);
160 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
161 * it by browsing the /proc/self/pagemap special file.
164 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
169 for (i = 0; i < hpi->num_pages[0]; i++) {
170 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
171 if (addr == RTE_BAD_PHYS_ADDR)
173 hugepg_tbl[i].physaddr = addr;
179 * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
182 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
185 static phys_addr_t addr;
187 for (i = 0; i < hpi->num_pages[0]; i++) {
188 hugepg_tbl[i].physaddr = addr;
189 addr += hugepg_tbl[i].size;
195 * Check whether address-space layout randomization is enabled in
196 * the kernel. This is important for multi-process as it can prevent
197 * two processes mapping data to the same virtual address
199 * 0 - address space randomization disabled
200 * 1/2 - address space randomization enabled
201 * negative error code on error
207 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
210 retval = read(fd, &c, 1);
220 default: return -EINVAL;
224 static sigjmp_buf huge_jmpenv;
226 static void huge_sigbus_handler(int signo __rte_unused)
228 siglongjmp(huge_jmpenv, 1);
231 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
232 * non-static local variable in the stack frame calling sigsetjmp might be
233 * clobbered by a call to longjmp.
235 static int huge_wrap_sigsetjmp(void)
237 return sigsetjmp(huge_jmpenv, 1);
240 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
241 /* Callback for numa library. */
242 void numa_error(char *where)
244 RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
249 * Mmap all hugepages of hugepage table: it first open a file in
250 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
251 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
252 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
253 * map contiguous physical blocks in contiguous virtual blocks.
256 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
257 uint64_t *essential_memory __rte_unused)
262 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
264 int essential_prev = 0;
266 struct bitmask *oldmask = numa_allocate_nodemask();
267 bool have_numa = true;
268 unsigned long maxnode = 0;
270 /* Check if kernel supports NUMA. */
271 if (numa_available() != 0) {
272 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
277 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
278 if (get_mempolicy(&oldpolicy, oldmask->maskp,
279 oldmask->size + 1, 0, 0) < 0) {
281 "Failed to get current mempolicy: %s. "
282 "Assuming MPOL_DEFAULT.\n", strerror(errno));
283 oldpolicy = MPOL_DEFAULT;
285 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
286 if (internal_config.socket_mem[i])
291 for (i = 0; i < hpi->num_pages[0]; i++) {
292 struct hugepage_file *hf = &hugepg_tbl[i];
293 uint64_t hugepage_sz = hpi->hugepage_sz;
295 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
299 for (j = 0; j < maxnode; j++)
300 if (essential_memory[j])
304 node_id = (node_id + 1) % maxnode;
305 while (!internal_config.socket_mem[node_id]) {
312 essential_prev = essential_memory[j];
314 if (essential_memory[j] < hugepage_sz)
315 essential_memory[j] = 0;
317 essential_memory[j] -= hugepage_sz;
321 "Setting policy MPOL_PREFERRED for socket %d\n",
323 numa_set_preferred(node_id);
328 hf->size = hugepage_sz;
329 eal_get_hugefile_path(hf->filepath, sizeof(hf->filepath),
330 hpi->hugedir, hf->file_id);
331 hf->filepath[sizeof(hf->filepath) - 1] = '\0';
333 /* try to create hugepage file */
334 fd = open(hf->filepath, O_CREAT | O_RDWR, 0600);
336 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
341 /* map the segment, and populate page tables,
342 * the kernel fills this segment with zeros. we don't care where
343 * this gets mapped - we already have contiguous memory areas
344 * ready for us to map into.
346 virtaddr = mmap(NULL, hugepage_sz, PROT_READ | PROT_WRITE,
347 MAP_SHARED | MAP_POPULATE, fd, 0);
348 if (virtaddr == MAP_FAILED) {
349 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
355 hf->orig_va = virtaddr;
357 /* In linux, hugetlb limitations, like cgroup, are
358 * enforced at fault time instead of mmap(), even
359 * with the option of MAP_POPULATE. Kernel will send
360 * a SIGBUS signal. To avoid to be killed, save stack
361 * environment here, if SIGBUS happens, we can jump
364 if (huge_wrap_sigsetjmp()) {
365 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
366 "hugepages of size %u MB\n",
367 (unsigned int)(hugepage_sz / 0x100000));
368 munmap(virtaddr, hugepage_sz);
370 unlink(hugepg_tbl[i].filepath);
371 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
373 essential_memory[node_id] =
378 *(int *)virtaddr = 0;
380 /* set shared lock on the file. */
381 if (flock(fd, LOCK_SH) < 0) {
382 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
383 __func__, strerror(errno));
392 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
395 "Restoring previous memory policy: %d\n", oldpolicy);
396 if (oldpolicy == MPOL_DEFAULT) {
397 numa_set_localalloc();
398 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
399 oldmask->size + 1) < 0) {
400 RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
402 numa_set_localalloc();
405 numa_free_cpumask(oldmask);
411 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
415 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
419 unsigned i, hp_count = 0;
422 char hugedir_str[PATH_MAX];
425 f = fopen("/proc/self/numa_maps", "r");
427 RTE_LOG(NOTICE, EAL, "NUMA support not available"
428 " consider that all memory is in socket_id 0\n");
432 snprintf(hugedir_str, sizeof(hugedir_str),
433 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
436 while (fgets(buf, sizeof(buf), f) != NULL) {
438 /* ignore non huge page */
439 if (strstr(buf, " huge ") == NULL &&
440 strstr(buf, hugedir_str) == NULL)
444 virt_addr = strtoull(buf, &end, 16);
445 if (virt_addr == 0 || end == buf) {
446 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
450 /* get node id (socket id) */
451 nodestr = strstr(buf, " N");
452 if (nodestr == NULL) {
453 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
457 end = strstr(nodestr, "=");
459 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
465 socket_id = strtoul(nodestr, &end, 0);
466 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
467 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
471 /* if we find this page in our mappings, set socket_id */
472 for (i = 0; i < hpi->num_pages[0]; i++) {
473 void *va = (void *)(unsigned long)virt_addr;
474 if (hugepg_tbl[i].orig_va == va) {
475 hugepg_tbl[i].socket_id = socket_id;
477 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
479 "Hugepage %s is on socket %d\n",
480 hugepg_tbl[i].filepath, socket_id);
486 if (hp_count < hpi->num_pages[0])
498 cmp_physaddr(const void *a, const void *b)
500 #ifndef RTE_ARCH_PPC_64
501 const struct hugepage_file *p1 = a;
502 const struct hugepage_file *p2 = b;
504 /* PowerPC needs memory sorted in reverse order from x86 */
505 const struct hugepage_file *p1 = b;
506 const struct hugepage_file *p2 = a;
508 if (p1->physaddr < p2->physaddr)
510 else if (p1->physaddr > p2->physaddr)
517 * Uses mmap to create a shared memory area for storage of data
518 * Used in this file to store the hugepage file map on disk
521 create_shared_memory(const char *filename, const size_t mem_size)
526 /* if no shared files mode is used, create anonymous memory instead */
527 if (internal_config.no_shconf) {
528 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE,
529 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
530 if (retval == MAP_FAILED)
535 fd = open(filename, O_CREAT | O_RDWR, 0666);
538 if (ftruncate(fd, mem_size) < 0) {
542 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
544 if (retval == MAP_FAILED)
550 * this copies *active* hugepages from one hugepage table to another.
551 * destination is typically the shared memory.
554 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
555 const struct hugepage_file * src, int src_size)
557 int src_pos, dst_pos = 0;
559 for (src_pos = 0; src_pos < src_size; src_pos++) {
560 if (src[src_pos].orig_va != NULL) {
561 /* error on overflow attempt */
562 if (dst_pos == dest_size)
564 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
572 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
573 unsigned num_hp_info)
575 unsigned socket, size;
576 int page, nrpages = 0;
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++)
582 internal_config.hugepage_info[size].num_pages[socket];
584 for (page = 0; page < nrpages; page++) {
585 struct hugepage_file *hp = &hugepg_tbl[page];
587 if (hp->final_va != NULL && unlink(hp->filepath)) {
588 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
589 __func__, hp->filepath, strerror(errno));
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.
600 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
601 struct hugepage_info *hpi,
602 unsigned num_hp_info)
604 unsigned socket, size;
605 int page, nrpages = 0;
607 /* get total number of hugepages */
608 for (size = 0; size < num_hp_info; size++)
609 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
610 nrpages += internal_config.hugepage_info[size].num_pages[socket];
612 for (size = 0; size < num_hp_info; size++) {
613 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
614 unsigned pages_found = 0;
616 /* traverse until we have unmapped all the unused pages */
617 for (page = 0; page < nrpages; page++) {
618 struct hugepage_file *hp = &hugepg_tbl[page];
620 /* find a page that matches the criteria */
621 if ((hp->size == hpi[size].hugepage_sz) &&
622 (hp->socket_id == (int) socket)) {
624 /* if we skipped enough pages, unmap the rest */
625 if (pages_found == hpi[size].num_pages[socket]) {
628 unmap_len = hp->size;
630 /* get start addr and len of the remaining segment */
635 if (unlink(hp->filepath) == -1) {
636 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
637 __func__, hp->filepath, strerror(errno));
641 /* lock the page and skip */
647 } /* foreach socket */
648 } /* foreach pagesize */
654 remap_segment(struct hugepage_file *hugepages, int seg_start, int seg_end)
656 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
657 struct rte_memseg_list *msl;
658 struct rte_fbarray *arr;
659 int cur_page, seg_len;
660 unsigned int msl_idx;
666 page_sz = hugepages[seg_start].size;
667 socket_id = hugepages[seg_start].socket_id;
668 seg_len = seg_end - seg_start;
670 RTE_LOG(DEBUG, EAL, "Attempting to map %" PRIu64 "M on socket %i\n",
671 (seg_len * page_sz) >> 20ULL, socket_id);
673 /* find free space in memseg lists */
674 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
676 msl = &mcfg->memsegs[msl_idx];
677 arr = &msl->memseg_arr;
679 if (msl->page_sz != page_sz)
681 if (msl->socket_id != socket_id)
684 /* leave space for a hole if array is not empty */
685 empty = arr->count == 0;
686 ms_idx = rte_fbarray_find_next_n_free(arr, 0,
687 seg_len + (empty ? 0 : 1));
689 /* memseg list is full? */
693 /* leave some space between memsegs, they are not IOVA
694 * contiguous, so they shouldn't be VA contiguous either.
700 if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
701 RTE_LOG(ERR, EAL, "Could not find space for memseg. Please increase %s and/or %s in configuration.\n",
702 RTE_STR(CONFIG_RTE_MAX_MEMSEG_PER_TYPE),
703 RTE_STR(CONFIG_RTE_MAX_MEM_PER_TYPE));
707 #ifdef RTE_ARCH_PPC64
708 /* for PPC64 we go through the list backwards */
709 for (cur_page = seg_end - 1; cur_page >= seg_start;
710 cur_page--, ms_idx++) {
712 for (cur_page = seg_start; cur_page < seg_end; cur_page++, ms_idx++) {
714 struct hugepage_file *hfile = &hugepages[cur_page];
715 struct rte_memseg *ms = rte_fbarray_get(arr, ms_idx);
719 fd = open(hfile->filepath, O_RDWR);
721 RTE_LOG(ERR, EAL, "Could not open '%s': %s\n",
722 hfile->filepath, strerror(errno));
725 /* set shared lock on the file. */
726 if (flock(fd, LOCK_SH) < 0) {
727 RTE_LOG(DEBUG, EAL, "Could not lock '%s': %s\n",
728 hfile->filepath, strerror(errno));
732 memseg_len = (size_t)page_sz;
733 addr = RTE_PTR_ADD(msl->base_va, ms_idx * memseg_len);
735 /* we know this address is already mmapped by memseg list, so
736 * using MAP_FIXED here is safe
738 addr = mmap(addr, page_sz, PROT_READ | PROT_WRITE,
739 MAP_SHARED | MAP_POPULATE | MAP_FIXED, fd, 0);
740 if (addr == MAP_FAILED) {
741 RTE_LOG(ERR, EAL, "Couldn't remap '%s': %s\n",
742 hfile->filepath, strerror(errno));
747 /* we have a new address, so unmap previous one */
749 /* in 32-bit legacy mode, we have already unmapped the page */
750 if (!internal_config.legacy_mem)
751 munmap(hfile->orig_va, page_sz);
753 munmap(hfile->orig_va, page_sz);
756 hfile->orig_va = NULL;
757 hfile->final_va = addr;
759 /* rewrite physical addresses in IOVA as VA mode */
760 if (rte_eal_iova_mode() == RTE_IOVA_VA)
761 hfile->physaddr = (uintptr_t)addr;
763 /* set up memseg data */
765 ms->hugepage_sz = page_sz;
766 ms->len = memseg_len;
767 ms->iova = hfile->physaddr;
768 ms->socket_id = hfile->socket_id;
769 ms->nchannel = rte_memory_get_nchannel();
770 ms->nrank = rte_memory_get_nrank();
772 rte_fbarray_set_used(arr, ms_idx);
776 RTE_LOG(DEBUG, EAL, "Allocated %" PRIu64 "M on socket %i\n",
777 (seg_len * page_sz) >> 20, socket_id);
782 get_mem_amount(uint64_t page_sz, uint64_t max_mem)
784 uint64_t area_sz, max_pages;
786 /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */
787 max_pages = RTE_MAX_MEMSEG_PER_LIST;
788 max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem);
790 area_sz = RTE_MIN(page_sz * max_pages, max_mem);
792 /* make sure the list isn't smaller than the page size */
793 area_sz = RTE_MAX(area_sz, page_sz);
795 return RTE_ALIGN(area_sz, page_sz);
799 free_memseg_list(struct rte_memseg_list *msl)
801 if (rte_fbarray_destroy(&msl->memseg_arr)) {
802 RTE_LOG(ERR, EAL, "Cannot destroy memseg list\n");
805 memset(msl, 0, sizeof(*msl));
809 #define MEMSEG_LIST_FMT "memseg-%" PRIu64 "k-%i-%i"
811 alloc_memseg_list(struct rte_memseg_list *msl, uint64_t page_sz,
812 int n_segs, int socket_id, int type_msl_idx)
814 char name[RTE_FBARRAY_NAME_LEN];
816 snprintf(name, sizeof(name), MEMSEG_LIST_FMT, page_sz >> 10, socket_id,
818 if (rte_fbarray_init(&msl->memseg_arr, name, n_segs,
819 sizeof(struct rte_memseg))) {
820 RTE_LOG(ERR, EAL, "Cannot allocate memseg list: %s\n",
821 rte_strerror(rte_errno));
825 msl->page_sz = page_sz;
826 msl->socket_id = socket_id;
829 RTE_LOG(DEBUG, EAL, "Memseg list allocated: 0x%zxkB at socket %i\n",
830 (size_t)page_sz >> 10, socket_id);
836 alloc_va_space(struct rte_memseg_list *msl)
843 #ifdef RTE_ARCH_PPC_64
844 flags |= MAP_HUGETLB;
847 page_sz = msl->page_sz;
848 mem_sz = page_sz * msl->memseg_arr.len;
850 addr = eal_get_virtual_area(msl->base_va, &mem_sz, page_sz, 0, flags);
852 if (rte_errno == EADDRNOTAVAIL)
853 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes at [%p] - please use '--base-virtaddr' option\n",
854 (unsigned long long)mem_sz, msl->base_va);
856 RTE_LOG(ERR, EAL, "Cannot reserve memory\n");
865 * Our VA space is not preallocated yet, so preallocate it here. We need to know
866 * how many segments there are in order to map all pages into one address space,
867 * and leave appropriate holes between segments so that rte_malloc does not
868 * concatenate them into one big segment.
870 * we also need to unmap original pages to free up address space.
872 static int __rte_unused
873 prealloc_segments(struct hugepage_file *hugepages, int n_pages)
875 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
876 int cur_page, seg_start_page, end_seg, new_memseg;
877 unsigned int hpi_idx, socket, i;
878 int n_contig_segs, n_segs;
881 /* before we preallocate segments, we need to free up our VA space.
882 * we're not removing files, and we already have information about
883 * PA-contiguousness, so it is safe to unmap everything.
885 for (cur_page = 0; cur_page < n_pages; cur_page++) {
886 struct hugepage_file *hpi = &hugepages[cur_page];
887 munmap(hpi->orig_va, hpi->size);
891 /* we cannot know how many page sizes and sockets we have discovered, so
892 * loop over all of them
894 for (hpi_idx = 0; hpi_idx < internal_config.num_hugepage_sizes;
897 internal_config.hugepage_info[hpi_idx].hugepage_sz;
899 for (i = 0; i < rte_socket_count(); i++) {
900 struct rte_memseg_list *msl;
902 socket = rte_socket_id_by_idx(i);
907 for (cur_page = 0; cur_page < n_pages; cur_page++) {
908 struct hugepage_file *prev, *cur;
909 int prev_seg_start_page = -1;
911 cur = &hugepages[cur_page];
912 prev = cur_page == 0 ? NULL :
913 &hugepages[cur_page - 1];
920 else if (cur->socket_id != (int) socket)
922 else if (cur->size != page_sz)
924 else if (cur_page == 0)
926 #ifdef RTE_ARCH_PPC_64
927 /* On PPC64 architecture, the mmap always start
928 * from higher address to lower address. Here,
929 * physical addresses are in descending order.
931 else if ((prev->physaddr - cur->physaddr) !=
935 else if ((cur->physaddr - prev->physaddr) !=
940 /* if we're already inside a segment,
941 * new segment means end of current one
943 if (seg_start_page != -1) {
945 prev_seg_start_page =
948 seg_start_page = cur_page;
952 if (prev_seg_start_page != -1) {
953 /* we've found a new segment */
957 } else if (seg_start_page != -1) {
958 /* we didn't find new segment,
959 * but did end current one
967 /* we're skipping this page */
971 /* segment continues */
973 /* check if we missed last segment */
974 if (seg_start_page != -1) {
976 n_segs += cur_page - seg_start_page;
979 /* if no segments were found, do not preallocate */
983 /* we now have total number of pages that we will
984 * allocate for this segment list. add separator pages
985 * to the total count, and preallocate VA space.
987 n_segs += n_contig_segs - 1;
989 /* now, preallocate VA space for these segments */
991 /* first, find suitable memseg list for this */
992 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS;
994 msl = &mcfg->memsegs[msl_idx];
996 if (msl->base_va != NULL)
1000 if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
1001 RTE_LOG(ERR, EAL, "Not enough space in memseg lists, please increase %s\n",
1002 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
1006 /* now, allocate fbarray itself */
1007 if (alloc_memseg_list(msl, page_sz, n_segs, socket,
1011 /* finally, allocate VA space */
1012 if (alloc_va_space(msl) < 0)
1020 * We cannot reallocate memseg lists on the fly because PPC64 stores pages
1021 * backwards, therefore we have to process the entire memseg first before
1022 * remapping it into memseg list VA space.
1025 remap_needed_hugepages(struct hugepage_file *hugepages, int n_pages)
1027 int cur_page, seg_start_page, new_memseg, ret;
1030 for (cur_page = 0; cur_page < n_pages; cur_page++) {
1031 struct hugepage_file *prev, *cur;
1035 cur = &hugepages[cur_page];
1036 prev = cur_page == 0 ? NULL : &hugepages[cur_page - 1];
1038 /* if size is zero, no more pages left */
1044 else if (cur->socket_id != prev->socket_id)
1046 else if (cur->size != prev->size)
1048 #ifdef RTE_ARCH_PPC_64
1049 /* On PPC64 architecture, the mmap always start from higher
1050 * address to lower address. Here, physical addresses are in
1053 else if ((prev->physaddr - cur->physaddr) != cur->size)
1056 else if ((cur->physaddr - prev->physaddr) != cur->size)
1061 /* if this isn't the first time, remap segment */
1062 if (cur_page != 0) {
1063 ret = remap_segment(hugepages, seg_start_page,
1068 /* remember where we started */
1069 seg_start_page = cur_page;
1071 /* continuation of previous memseg */
1073 /* we were stopped, but we didn't remap the last segment, do it now */
1074 if (cur_page != 0) {
1075 ret = remap_segment(hugepages, seg_start_page,
1083 static inline uint64_t
1084 get_socket_mem_size(int socket)
1089 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
1090 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1091 size += hpi->hugepage_sz * hpi->num_pages[socket];
1098 * This function is a NUMA-aware equivalent of calc_num_pages.
1099 * It takes in the list of hugepage sizes and the
1100 * number of pages thereof, and calculates the best number of
1101 * pages of each size to fulfill the request for <memory> ram
1104 calc_num_pages_per_socket(uint64_t * memory,
1105 struct hugepage_info *hp_info,
1106 struct hugepage_info *hp_used,
1107 unsigned num_hp_info)
1109 unsigned socket, j, i = 0;
1110 unsigned requested, available;
1111 int total_num_pages = 0;
1112 uint64_t remaining_mem, cur_mem;
1113 uint64_t total_mem = internal_config.memory;
1115 if (num_hp_info == 0)
1118 /* if specific memory amounts per socket weren't requested */
1119 if (internal_config.force_sockets == 0) {
1122 int cpu_per_socket[RTE_MAX_NUMA_NODES];
1123 size_t default_size;
1126 /* Compute number of cores per socket */
1127 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
1128 RTE_LCORE_FOREACH(lcore_id) {
1129 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
1133 * Automatically spread requested memory amongst detected sockets according
1134 * to number of cores from cpu mask present on each socket
1136 total_size = internal_config.memory;
1137 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1139 /* Set memory amount per socket */
1140 default_size = (internal_config.memory * cpu_per_socket[socket])
1141 / rte_lcore_count();
1143 /* Limit to maximum available memory on socket */
1144 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
1147 memory[socket] = default_size;
1148 total_size -= default_size;
1152 * If some memory is remaining, try to allocate it by getting all
1153 * available memory from sockets, one after the other
1155 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1156 /* take whatever is available */
1157 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
1161 memory[socket] += default_size;
1162 total_size -= default_size;
1165 /* in 32-bit mode, allocate all of the memory only on master
1168 total_size = internal_config.memory;
1169 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0;
1171 struct rte_config *cfg = rte_eal_get_configuration();
1172 unsigned int master_lcore_socket;
1174 master_lcore_socket =
1175 rte_lcore_to_socket_id(cfg->master_lcore);
1177 if (master_lcore_socket != socket)
1181 memory[socket] = total_size;
1187 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
1188 /* skips if the memory on specific socket wasn't requested */
1189 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
1190 strlcpy(hp_used[i].hugedir, hp_info[i].hugedir,
1191 sizeof(hp_used[i].hugedir));
1192 hp_used[i].num_pages[socket] = RTE_MIN(
1193 memory[socket] / hp_info[i].hugepage_sz,
1194 hp_info[i].num_pages[socket]);
1196 cur_mem = hp_used[i].num_pages[socket] *
1197 hp_used[i].hugepage_sz;
1199 memory[socket] -= cur_mem;
1200 total_mem -= cur_mem;
1202 total_num_pages += hp_used[i].num_pages[socket];
1204 /* check if we have met all memory requests */
1205 if (memory[socket] == 0)
1208 /* check if we have any more pages left at this size, if so
1209 * move on to next size */
1210 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
1212 /* At this point we know that there are more pages available that are
1213 * bigger than the memory we want, so lets see if we can get enough
1214 * from other page sizes.
1217 for (j = i+1; j < num_hp_info; j++)
1218 remaining_mem += hp_info[j].hugepage_sz *
1219 hp_info[j].num_pages[socket];
1221 /* is there enough other memory, if not allocate another page and quit */
1222 if (remaining_mem < memory[socket]){
1223 cur_mem = RTE_MIN(memory[socket],
1224 hp_info[i].hugepage_sz);
1225 memory[socket] -= cur_mem;
1226 total_mem -= cur_mem;
1227 hp_used[i].num_pages[socket]++;
1229 break; /* we are done with this socket*/
1232 /* if we didn't satisfy all memory requirements per socket */
1233 if (memory[socket] > 0 &&
1234 internal_config.socket_mem[socket] != 0) {
1235 /* to prevent icc errors */
1236 requested = (unsigned) (internal_config.socket_mem[socket] /
1238 available = requested -
1239 ((unsigned) (memory[socket] / 0x100000));
1240 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
1241 "Requested: %uMB, available: %uMB\n", socket,
1242 requested, available);
1247 /* if we didn't satisfy total memory requirements */
1248 if (total_mem > 0) {
1249 requested = (unsigned) (internal_config.memory / 0x100000);
1250 available = requested - (unsigned) (total_mem / 0x100000);
1251 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
1252 " available: %uMB\n", requested, available);
1255 return total_num_pages;
1258 static inline size_t
1259 eal_get_hugepage_mem_size(void)
1264 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
1265 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1266 if (strnlen(hpi->hugedir, sizeof(hpi->hugedir)) != 0) {
1267 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1268 size += hpi->hugepage_sz * hpi->num_pages[j];
1273 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
1276 static struct sigaction huge_action_old;
1277 static int huge_need_recover;
1280 huge_register_sigbus(void)
1283 struct sigaction action;
1286 sigaddset(&mask, SIGBUS);
1287 action.sa_flags = 0;
1288 action.sa_mask = mask;
1289 action.sa_handler = huge_sigbus_handler;
1291 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1295 huge_recover_sigbus(void)
1297 if (huge_need_recover) {
1298 sigaction(SIGBUS, &huge_action_old, NULL);
1299 huge_need_recover = 0;
1304 * Prepare physical memory mapping: fill configuration structure with
1305 * these infos, return 0 on success.
1306 * 1. map N huge pages in separate files in hugetlbfs
1307 * 2. find associated physical addr
1308 * 3. find associated NUMA socket ID
1309 * 4. sort all huge pages by physical address
1310 * 5. remap these N huge pages in the correct order
1311 * 6. unmap the first mapping
1312 * 7. fill memsegs in configuration with contiguous zones
1315 eal_legacy_hugepage_init(void)
1317 struct rte_mem_config *mcfg;
1318 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1319 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1320 struct rte_fbarray *arr;
1321 struct rte_memseg *ms;
1323 uint64_t memory[RTE_MAX_NUMA_NODES];
1327 int nr_hugefiles, nr_hugepages = 0;
1330 test_phys_addrs_available();
1332 memset(used_hp, 0, sizeof(used_hp));
1334 /* get pointer to global configuration */
1335 mcfg = rte_eal_get_configuration()->mem_config;
1337 /* hugetlbfs can be disabled */
1338 if (internal_config.no_hugetlbfs) {
1339 struct rte_memseg_list *msl;
1341 int n_segs, cur_seg;
1343 /* nohuge mode is legacy mode */
1344 internal_config.legacy_mem = 1;
1346 /* create a memseg list */
1347 msl = &mcfg->memsegs[0];
1349 page_sz = RTE_PGSIZE_4K;
1350 n_segs = internal_config.memory / page_sz;
1352 if (rte_fbarray_init(&msl->memseg_arr, "nohugemem", n_segs,
1353 sizeof(struct rte_memseg))) {
1354 RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
1358 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1359 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
1360 if (addr == MAP_FAILED) {
1361 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1365 msl->base_va = addr;
1366 msl->page_sz = page_sz;
1369 /* populate memsegs. each memseg is one page long */
1370 for (cur_seg = 0; cur_seg < n_segs; cur_seg++) {
1371 arr = &msl->memseg_arr;
1373 ms = rte_fbarray_get(arr, cur_seg);
1374 if (rte_eal_iova_mode() == RTE_IOVA_VA)
1375 ms->iova = (uintptr_t)addr;
1377 ms->iova = RTE_BAD_IOVA;
1379 ms->hugepage_sz = page_sz;
1383 rte_fbarray_set_used(arr, cur_seg);
1385 addr = RTE_PTR_ADD(addr, (size_t)page_sz);
1390 /* calculate total number of hugepages available. at this point we haven't
1391 * yet started sorting them so they all are on socket 0 */
1392 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1393 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1394 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1396 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1400 * allocate a memory area for hugepage table.
1401 * this isn't shared memory yet. due to the fact that we need some
1402 * processing done on these pages, shared memory will be created
1405 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1409 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1411 hp_offset = 0; /* where we start the current page size entries */
1413 huge_register_sigbus();
1415 /* make a copy of socket_mem, needed for balanced allocation. */
1416 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1417 memory[i] = internal_config.socket_mem[i];
1419 /* map all hugepages and sort them */
1420 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1421 unsigned pages_old, pages_new;
1422 struct hugepage_info *hpi;
1425 * we don't yet mark hugepages as used at this stage, so
1426 * we just map all hugepages available to the system
1427 * all hugepages are still located on socket 0
1429 hpi = &internal_config.hugepage_info[i];
1431 if (hpi->num_pages[0] == 0)
1434 /* map all hugepages available */
1435 pages_old = hpi->num_pages[0];
1436 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, memory);
1437 if (pages_new < pages_old) {
1439 "%d not %d hugepages of size %u MB allocated\n",
1440 pages_new, pages_old,
1441 (unsigned)(hpi->hugepage_sz / 0x100000));
1443 int pages = pages_old - pages_new;
1445 nr_hugepages -= pages;
1446 hpi->num_pages[0] = pages_new;
1451 if (phys_addrs_available &&
1452 rte_eal_iova_mode() != RTE_IOVA_VA) {
1453 /* find physical addresses for each hugepage */
1454 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1455 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1456 "for %u MB pages\n",
1457 (unsigned int)(hpi->hugepage_sz / 0x100000));
1461 /* set physical addresses for each hugepage */
1462 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1463 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1464 "for %u MB pages\n",
1465 (unsigned int)(hpi->hugepage_sz / 0x100000));
1470 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1471 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1472 (unsigned)(hpi->hugepage_sz / 0x100000));
1476 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1477 sizeof(struct hugepage_file), cmp_physaddr);
1479 /* we have processed a num of hugepages of this size, so inc offset */
1480 hp_offset += hpi->num_pages[0];
1483 huge_recover_sigbus();
1485 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1486 internal_config.memory = eal_get_hugepage_mem_size();
1488 nr_hugefiles = nr_hugepages;
1491 /* clean out the numbers of pages */
1492 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1493 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1494 internal_config.hugepage_info[i].num_pages[j] = 0;
1496 /* get hugepages for each socket */
1497 for (i = 0; i < nr_hugefiles; i++) {
1498 int socket = tmp_hp[i].socket_id;
1500 /* find a hugepage info with right size and increment num_pages */
1501 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1502 (int)internal_config.num_hugepage_sizes);
1503 for (j = 0; j < nb_hpsizes; j++) {
1504 if (tmp_hp[i].size ==
1505 internal_config.hugepage_info[j].hugepage_sz) {
1506 internal_config.hugepage_info[j].num_pages[socket]++;
1511 /* make a copy of socket_mem, needed for number of pages calculation */
1512 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1513 memory[i] = internal_config.socket_mem[i];
1515 /* calculate final number of pages */
1516 nr_hugepages = calc_num_pages_per_socket(memory,
1517 internal_config.hugepage_info, used_hp,
1518 internal_config.num_hugepage_sizes);
1520 /* error if not enough memory available */
1521 if (nr_hugepages < 0)
1525 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1526 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1527 if (used_hp[i].num_pages[j] > 0) {
1529 "Requesting %u pages of size %uMB"
1530 " from socket %i\n",
1531 used_hp[i].num_pages[j],
1533 (used_hp[i].hugepage_sz / 0x100000),
1539 /* create shared memory */
1540 hugepage = create_shared_memory(eal_hugepage_data_path(),
1541 nr_hugefiles * sizeof(struct hugepage_file));
1543 if (hugepage == NULL) {
1544 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1547 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1550 * unmap pages that we won't need (looks at used_hp).
1551 * also, sets final_va to NULL on pages that were unmapped.
1553 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1554 internal_config.num_hugepage_sizes) < 0) {
1555 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1560 * copy stuff from malloc'd hugepage* to the actual shared memory.
1561 * this procedure only copies those hugepages that have orig_va
1562 * not NULL. has overflow protection.
1564 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1565 tmp_hp, nr_hugefiles) < 0) {
1566 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1571 /* for legacy 32-bit mode, we did not preallocate VA space, so do it */
1572 if (internal_config.legacy_mem &&
1573 prealloc_segments(hugepage, nr_hugefiles)) {
1574 RTE_LOG(ERR, EAL, "Could not preallocate VA space for hugepages\n");
1579 /* remap all pages we do need into memseg list VA space, so that those
1580 * pages become first-class citizens in DPDK memory subsystem
1582 if (remap_needed_hugepages(hugepage, nr_hugefiles)) {
1583 RTE_LOG(ERR, EAL, "Couldn't remap hugepage files into memseg lists\n");
1587 /* free the hugepage backing files */
1588 if (internal_config.hugepage_unlink &&
1589 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1590 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1594 /* free the temporary hugepage table */
1598 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1600 /* we're not going to allocate more pages, so release VA space for
1601 * unused memseg lists
1603 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1604 struct rte_memseg_list *msl = &mcfg->memsegs[i];
1607 /* skip inactive lists */
1608 if (msl->base_va == NULL)
1610 /* skip lists where there is at least one page allocated */
1611 if (msl->memseg_arr.count > 0)
1613 /* this is an unused list, deallocate it */
1614 mem_sz = (size_t)msl->page_sz * msl->memseg_arr.len;
1615 munmap(msl->base_va, mem_sz);
1616 msl->base_va = NULL;
1618 /* destroy backing fbarray */
1619 rte_fbarray_destroy(&msl->memseg_arr);
1625 huge_recover_sigbus();
1627 if (hugepage != NULL)
1628 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1633 static int __rte_unused
1634 hugepage_count_walk(const struct rte_memseg_list *msl, void *arg)
1636 struct hugepage_info *hpi = arg;
1638 if (msl->page_sz != hpi->hugepage_sz)
1641 hpi->num_pages[msl->socket_id] += msl->memseg_arr.len;
1646 limits_callback(int socket_id, size_t cur_limit, size_t new_len)
1648 RTE_SET_USED(socket_id);
1649 RTE_SET_USED(cur_limit);
1650 RTE_SET_USED(new_len);
1655 eal_hugepage_init(void)
1657 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1658 uint64_t memory[RTE_MAX_NUMA_NODES];
1659 int hp_sz_idx, socket_id;
1661 test_phys_addrs_available();
1663 memset(used_hp, 0, sizeof(used_hp));
1666 hp_sz_idx < (int) internal_config.num_hugepage_sizes;
1669 struct hugepage_info dummy;
1672 /* also initialize used_hp hugepage sizes in used_hp */
1673 struct hugepage_info *hpi;
1674 hpi = &internal_config.hugepage_info[hp_sz_idx];
1675 used_hp[hp_sz_idx].hugepage_sz = hpi->hugepage_sz;
1678 /* for 32-bit, limit number of pages on socket to whatever we've
1679 * preallocated, as we cannot allocate more.
1681 memset(&dummy, 0, sizeof(dummy));
1682 dummy.hugepage_sz = hpi->hugepage_sz;
1683 if (rte_memseg_list_walk(hugepage_count_walk, &dummy) < 0)
1686 for (i = 0; i < RTE_DIM(dummy.num_pages); i++) {
1687 hpi->num_pages[i] = RTE_MIN(hpi->num_pages[i],
1688 dummy.num_pages[i]);
1693 /* make a copy of socket_mem, needed for balanced allocation. */
1694 for (hp_sz_idx = 0; hp_sz_idx < RTE_MAX_NUMA_NODES; hp_sz_idx++)
1695 memory[hp_sz_idx] = internal_config.socket_mem[hp_sz_idx];
1697 /* calculate final number of pages */
1698 if (calc_num_pages_per_socket(memory,
1699 internal_config.hugepage_info, used_hp,
1700 internal_config.num_hugepage_sizes) < 0)
1704 hp_sz_idx < (int)internal_config.num_hugepage_sizes;
1706 for (socket_id = 0; socket_id < RTE_MAX_NUMA_NODES;
1708 struct rte_memseg **pages;
1709 struct hugepage_info *hpi = &used_hp[hp_sz_idx];
1710 unsigned int num_pages = hpi->num_pages[socket_id];
1711 int num_pages_alloc, i;
1716 pages = malloc(sizeof(*pages) * num_pages);
1718 RTE_LOG(DEBUG, EAL, "Allocating %u pages of size %" PRIu64 "M on socket %i\n",
1719 num_pages, hpi->hugepage_sz >> 20, socket_id);
1721 num_pages_alloc = eal_memalloc_alloc_seg_bulk(pages,
1722 num_pages, hpi->hugepage_sz,
1724 if (num_pages_alloc < 0) {
1729 /* mark preallocated pages as unfreeable */
1730 for (i = 0; i < num_pages_alloc; i++) {
1731 struct rte_memseg *ms = pages[i];
1732 ms->flags |= RTE_MEMSEG_FLAG_DO_NOT_FREE;
1737 /* if socket limits were specified, set them */
1738 if (internal_config.force_socket_limits) {
1740 for (i = 0; i < RTE_MAX_NUMA_NODES; i++) {
1741 uint64_t limit = internal_config.socket_limit[i];
1744 if (rte_mem_alloc_validator_register("socket-limit",
1745 limits_callback, i, limit))
1746 RTE_LOG(ERR, EAL, "Failed to register socket limits validator callback\n");
1753 * uses fstat to report the size of a file on disk
1759 if (fstat(fd, &st) < 0)
1765 * This creates the memory mappings in the secondary process to match that of
1766 * the server process. It goes through each memory segment in the DPDK runtime
1767 * configuration and finds the hugepages which form that segment, mapping them
1768 * in order to form a contiguous block in the virtual memory space
1771 eal_legacy_hugepage_attach(void)
1773 struct hugepage_file *hp = NULL;
1774 unsigned int num_hp = 0;
1776 unsigned int cur_seg;
1778 int fd, fd_hugepage = -1;
1780 if (aslr_enabled() > 0) {
1781 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1782 "(ASLR) is enabled in the kernel.\n");
1783 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1784 "into secondary processes\n");
1787 test_phys_addrs_available();
1789 fd_hugepage = open(eal_hugepage_data_path(), O_RDONLY);
1790 if (fd_hugepage < 0) {
1791 RTE_LOG(ERR, EAL, "Could not open %s\n",
1792 eal_hugepage_data_path());
1796 size = getFileSize(fd_hugepage);
1797 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1798 if (hp == MAP_FAILED) {
1799 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1800 eal_hugepage_data_path());
1804 num_hp = size / sizeof(struct hugepage_file);
1805 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1807 /* map all segments into memory to make sure we get the addrs. the
1808 * segments themselves are already in memseg list (which is shared and
1809 * has its VA space already preallocated), so we just need to map
1810 * everything into correct addresses.
1812 for (i = 0; i < num_hp; i++) {
1813 struct hugepage_file *hf = &hp[i];
1814 size_t map_sz = hf->size;
1815 void *map_addr = hf->final_va;
1817 /* if size is zero, no more pages left */
1821 fd = open(hf->filepath, O_RDWR);
1823 RTE_LOG(ERR, EAL, "Could not open %s: %s\n",
1824 hf->filepath, strerror(errno));
1828 map_addr = mmap(map_addr, map_sz, PROT_READ | PROT_WRITE,
1829 MAP_SHARED | MAP_FIXED, fd, 0);
1830 if (map_addr == MAP_FAILED) {
1831 RTE_LOG(ERR, EAL, "Could not map %s: %s\n",
1832 hf->filepath, strerror(errno));
1837 /* set shared lock on the file. */
1838 if (flock(fd, LOCK_SH) < 0) {
1839 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed: %s\n",
1840 __func__, strerror(errno));
1847 /* unmap the hugepage config file, since we are done using it */
1853 /* map all segments into memory to make sure we get the addrs */
1855 for (cur_seg = 0; cur_seg < i; cur_seg++) {
1856 struct hugepage_file *hf = &hp[i];
1857 size_t map_sz = hf->size;
1858 void *map_addr = hf->final_va;
1860 munmap(map_addr, map_sz);
1862 if (hp != NULL && hp != MAP_FAILED)
1864 if (fd_hugepage >= 0)
1870 eal_hugepage_attach(void)
1872 if (eal_memalloc_sync_with_primary()) {
1873 RTE_LOG(ERR, EAL, "Could not map memory from primary process\n");
1874 if (aslr_enabled() > 0)
1875 RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel and retry running both primary and secondary processes\n");
1882 rte_eal_hugepage_init(void)
1884 return internal_config.legacy_mem ?
1885 eal_legacy_hugepage_init() :
1886 eal_hugepage_init();
1890 rte_eal_hugepage_attach(void)
1892 return internal_config.legacy_mem ?
1893 eal_legacy_hugepage_attach() :
1894 eal_hugepage_attach();
1898 rte_eal_using_phys_addrs(void)
1900 return phys_addrs_available;
1903 static int __rte_unused
1904 memseg_primary_init_32(void)
1906 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1907 int active_sockets, hpi_idx, msl_idx = 0;
1908 unsigned int socket_id, i;
1909 struct rte_memseg_list *msl;
1910 uint64_t extra_mem_per_socket, total_extra_mem, total_requested_mem;
1913 /* no-huge does not need this at all */
1914 if (internal_config.no_hugetlbfs)
1917 /* this is a giant hack, but desperate times call for desperate
1918 * measures. in legacy 32-bit mode, we cannot preallocate VA space,
1919 * because having upwards of 2 gigabytes of VA space already mapped will
1920 * interfere with our ability to map and sort hugepages.
1922 * therefore, in legacy 32-bit mode, we will be initializing memseg
1923 * lists much later - in eal_memory.c, right after we unmap all the
1924 * unneeded pages. this will not affect secondary processes, as those
1925 * should be able to mmap the space without (too many) problems.
1927 if (internal_config.legacy_mem)
1930 /* 32-bit mode is a very special case. we cannot know in advance where
1931 * the user will want to allocate their memory, so we have to do some
1935 total_requested_mem = 0;
1936 if (internal_config.force_sockets)
1937 for (i = 0; i < rte_socket_count(); i++) {
1940 socket_id = rte_socket_id_by_idx(i);
1941 mem = internal_config.socket_mem[socket_id];
1947 total_requested_mem += mem;
1950 total_requested_mem = internal_config.memory;
1952 max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
1953 if (total_requested_mem > max_mem) {
1954 RTE_LOG(ERR, EAL, "Invalid parameters: 32-bit process can at most use %uM of memory\n",
1955 (unsigned int)(max_mem >> 20));
1958 total_extra_mem = max_mem - total_requested_mem;
1959 extra_mem_per_socket = active_sockets == 0 ? total_extra_mem :
1960 total_extra_mem / active_sockets;
1962 /* the allocation logic is a little bit convoluted, but here's how it
1963 * works, in a nutshell:
1964 * - if user hasn't specified on which sockets to allocate memory via
1965 * --socket-mem, we allocate all of our memory on master core socket.
1966 * - if user has specified sockets to allocate memory on, there may be
1967 * some "unused" memory left (e.g. if user has specified --socket-mem
1968 * such that not all memory adds up to 2 gigabytes), so add it to all
1969 * sockets that are in use equally.
1971 * page sizes are sorted by size in descending order, so we can safely
1972 * assume that we dispense with bigger page sizes first.
1975 /* create memseg lists */
1976 for (i = 0; i < rte_socket_count(); i++) {
1977 int hp_sizes = (int) internal_config.num_hugepage_sizes;
1978 uint64_t max_socket_mem, cur_socket_mem;
1979 unsigned int master_lcore_socket;
1980 struct rte_config *cfg = rte_eal_get_configuration();
1983 socket_id = rte_socket_id_by_idx(i);
1985 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
1990 /* if we didn't specifically request memory on this socket */
1991 skip = active_sockets != 0 &&
1992 internal_config.socket_mem[socket_id] == 0;
1993 /* ...or if we didn't specifically request memory on *any*
1994 * socket, and this is not master lcore
1996 master_lcore_socket = rte_lcore_to_socket_id(cfg->master_lcore);
1997 skip |= active_sockets == 0 && socket_id != master_lcore_socket;
2000 RTE_LOG(DEBUG, EAL, "Will not preallocate memory on socket %u\n",
2005 /* max amount of memory on this socket */
2006 max_socket_mem = (active_sockets != 0 ?
2007 internal_config.socket_mem[socket_id] :
2008 internal_config.memory) +
2009 extra_mem_per_socket;
2012 for (hpi_idx = 0; hpi_idx < hp_sizes; hpi_idx++) {
2013 uint64_t max_pagesz_mem, cur_pagesz_mem = 0;
2014 uint64_t hugepage_sz;
2015 struct hugepage_info *hpi;
2016 int type_msl_idx, max_segs, total_segs = 0;
2018 hpi = &internal_config.hugepage_info[hpi_idx];
2019 hugepage_sz = hpi->hugepage_sz;
2021 /* check if pages are actually available */
2022 if (hpi->num_pages[socket_id] == 0)
2025 max_segs = RTE_MAX_MEMSEG_PER_TYPE;
2026 max_pagesz_mem = max_socket_mem - cur_socket_mem;
2028 /* make it multiple of page size */
2029 max_pagesz_mem = RTE_ALIGN_FLOOR(max_pagesz_mem,
2032 RTE_LOG(DEBUG, EAL, "Attempting to preallocate "
2033 "%" PRIu64 "M on socket %i\n",
2034 max_pagesz_mem >> 20, socket_id);
2037 while (cur_pagesz_mem < max_pagesz_mem &&
2038 total_segs < max_segs) {
2040 unsigned int n_segs;
2042 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2044 "No more space in memseg lists, please increase %s\n",
2045 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2049 msl = &mcfg->memsegs[msl_idx];
2051 cur_mem = get_mem_amount(hugepage_sz,
2053 n_segs = cur_mem / hugepage_sz;
2055 if (alloc_memseg_list(msl, hugepage_sz, n_segs,
2056 socket_id, type_msl_idx)) {
2057 /* failing to allocate a memseg list is
2060 RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
2064 if (alloc_va_space(msl)) {
2065 /* if we couldn't allocate VA space, we
2066 * can try with smaller page sizes.
2068 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list, retrying with different page size\n");
2069 /* deallocate memseg list */
2070 if (free_memseg_list(msl))
2075 total_segs += msl->memseg_arr.len;
2076 cur_pagesz_mem = total_segs * hugepage_sz;
2080 cur_socket_mem += cur_pagesz_mem;
2082 if (cur_socket_mem == 0) {
2083 RTE_LOG(ERR, EAL, "Cannot allocate VA space on socket %u\n",
2092 static int __rte_unused
2093 memseg_primary_init(void)
2095 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2096 int i, socket_id, hpi_idx, msl_idx = 0;
2097 struct rte_memseg_list *msl;
2098 uint64_t max_mem, total_mem;
2100 /* no-huge does not need this at all */
2101 if (internal_config.no_hugetlbfs)
2104 max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
2107 /* create memseg lists */
2108 for (hpi_idx = 0; hpi_idx < (int) internal_config.num_hugepage_sizes;
2110 struct hugepage_info *hpi;
2111 uint64_t hugepage_sz;
2113 hpi = &internal_config.hugepage_info[hpi_idx];
2114 hugepage_sz = hpi->hugepage_sz;
2116 for (i = 0; i < (int) rte_socket_count(); i++) {
2117 uint64_t max_type_mem, total_type_mem = 0;
2118 int type_msl_idx, max_segs, total_segs = 0;
2120 socket_id = rte_socket_id_by_idx(i);
2122 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
2127 if (total_mem >= max_mem)
2130 max_type_mem = RTE_MIN(max_mem - total_mem,
2131 (uint64_t)RTE_MAX_MEM_MB_PER_TYPE << 20);
2132 max_segs = RTE_MAX_MEMSEG_PER_TYPE;
2135 while (total_type_mem < max_type_mem &&
2136 total_segs < max_segs) {
2137 uint64_t cur_max_mem, cur_mem;
2138 unsigned int n_segs;
2140 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2142 "No more space in memseg lists, please increase %s\n",
2143 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2147 msl = &mcfg->memsegs[msl_idx++];
2149 cur_max_mem = max_type_mem - total_type_mem;
2151 cur_mem = get_mem_amount(hugepage_sz,
2153 n_segs = cur_mem / hugepage_sz;
2155 if (alloc_memseg_list(msl, hugepage_sz, n_segs,
2156 socket_id, type_msl_idx))
2159 total_segs += msl->memseg_arr.len;
2160 total_type_mem = total_segs * hugepage_sz;
2163 if (alloc_va_space(msl)) {
2164 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list\n");
2168 total_mem += total_type_mem;
2175 memseg_secondary_init(void)
2177 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2179 struct rte_memseg_list *msl;
2181 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
2183 msl = &mcfg->memsegs[msl_idx];
2185 /* skip empty memseg lists */
2186 if (msl->memseg_arr.len == 0)
2189 if (rte_fbarray_attach(&msl->memseg_arr)) {
2190 RTE_LOG(ERR, EAL, "Cannot attach to primary process memseg lists\n");
2194 /* preallocate VA space */
2195 if (alloc_va_space(msl)) {
2196 RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
2205 rte_eal_memseg_init(void)
2207 return rte_eal_process_type() == RTE_PROC_PRIMARY ?
2209 memseg_primary_init_32() :
2211 memseg_primary_init() :
2213 memseg_secondary_init();