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)
524 int fd = open(filename, O_CREAT | O_RDWR, 0666);
527 if (ftruncate(fd, mem_size) < 0) {
531 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
533 if (retval == MAP_FAILED)
539 * this copies *active* hugepages from one hugepage table to another.
540 * destination is typically the shared memory.
543 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
544 const struct hugepage_file * src, int src_size)
546 int src_pos, dst_pos = 0;
548 for (src_pos = 0; src_pos < src_size; src_pos++) {
549 if (src[src_pos].orig_va != NULL) {
550 /* error on overflow attempt */
551 if (dst_pos == dest_size)
553 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
561 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
562 unsigned num_hp_info)
564 unsigned socket, size;
565 int page, nrpages = 0;
567 /* get total number of hugepages */
568 for (size = 0; size < num_hp_info; size++)
569 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
571 internal_config.hugepage_info[size].num_pages[socket];
573 for (page = 0; page < nrpages; page++) {
574 struct hugepage_file *hp = &hugepg_tbl[page];
576 if (hp->final_va != NULL && unlink(hp->filepath)) {
577 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
578 __func__, hp->filepath, strerror(errno));
585 * unmaps hugepages that are not going to be used. since we originally allocate
586 * ALL hugepages (not just those we need), additional unmapping needs to be done.
589 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
590 struct hugepage_info *hpi,
591 unsigned num_hp_info)
593 unsigned socket, size;
594 int page, nrpages = 0;
596 /* get total number of hugepages */
597 for (size = 0; size < num_hp_info; size++)
598 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
599 nrpages += internal_config.hugepage_info[size].num_pages[socket];
601 for (size = 0; size < num_hp_info; size++) {
602 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
603 unsigned pages_found = 0;
605 /* traverse until we have unmapped all the unused pages */
606 for (page = 0; page < nrpages; page++) {
607 struct hugepage_file *hp = &hugepg_tbl[page];
609 /* find a page that matches the criteria */
610 if ((hp->size == hpi[size].hugepage_sz) &&
611 (hp->socket_id == (int) socket)) {
613 /* if we skipped enough pages, unmap the rest */
614 if (pages_found == hpi[size].num_pages[socket]) {
617 unmap_len = hp->size;
619 /* get start addr and len of the remaining segment */
624 if (unlink(hp->filepath) == -1) {
625 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
626 __func__, hp->filepath, strerror(errno));
630 /* lock the page and skip */
636 } /* foreach socket */
637 } /* foreach pagesize */
643 remap_segment(struct hugepage_file *hugepages, int seg_start, int seg_end)
645 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
646 struct rte_memseg_list *msl;
647 struct rte_fbarray *arr;
648 int cur_page, seg_len;
649 unsigned int msl_idx;
655 page_sz = hugepages[seg_start].size;
656 socket_id = hugepages[seg_start].socket_id;
657 seg_len = seg_end - seg_start;
659 RTE_LOG(DEBUG, EAL, "Attempting to map %" PRIu64 "M on socket %i\n",
660 (seg_len * page_sz) >> 20ULL, socket_id);
662 /* find free space in memseg lists */
663 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
665 msl = &mcfg->memsegs[msl_idx];
666 arr = &msl->memseg_arr;
668 if (msl->page_sz != page_sz)
670 if (msl->socket_id != socket_id)
673 /* leave space for a hole if array is not empty */
674 empty = arr->count == 0;
675 ms_idx = rte_fbarray_find_next_n_free(arr, 0,
676 seg_len + (empty ? 0 : 1));
678 /* memseg list is full? */
682 /* leave some space between memsegs, they are not IOVA
683 * contiguous, so they shouldn't be VA contiguous either.
689 if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
690 RTE_LOG(ERR, EAL, "Could not find space for memseg. Please increase %s and/or %s in configuration.\n",
691 RTE_STR(CONFIG_RTE_MAX_MEMSEG_PER_TYPE),
692 RTE_STR(CONFIG_RTE_MAX_MEM_PER_TYPE));
696 #ifdef RTE_ARCH_PPC64
697 /* for PPC64 we go through the list backwards */
698 for (cur_page = seg_end - 1; cur_page >= seg_start;
699 cur_page--, ms_idx++) {
701 for (cur_page = seg_start; cur_page < seg_end; cur_page++, ms_idx++) {
703 struct hugepage_file *hfile = &hugepages[cur_page];
704 struct rte_memseg *ms = rte_fbarray_get(arr, ms_idx);
708 fd = open(hfile->filepath, O_RDWR);
710 RTE_LOG(ERR, EAL, "Could not open '%s': %s\n",
711 hfile->filepath, strerror(errno));
714 /* set shared lock on the file. */
715 if (flock(fd, LOCK_SH) < 0) {
716 RTE_LOG(DEBUG, EAL, "Could not lock '%s': %s\n",
717 hfile->filepath, strerror(errno));
721 memseg_len = (size_t)page_sz;
722 addr = RTE_PTR_ADD(msl->base_va, ms_idx * memseg_len);
724 /* we know this address is already mmapped by memseg list, so
725 * using MAP_FIXED here is safe
727 addr = mmap(addr, page_sz, PROT_READ | PROT_WRITE,
728 MAP_SHARED | MAP_POPULATE | MAP_FIXED, fd, 0);
729 if (addr == MAP_FAILED) {
730 RTE_LOG(ERR, EAL, "Couldn't remap '%s': %s\n",
731 hfile->filepath, strerror(errno));
736 /* we have a new address, so unmap previous one */
738 /* in 32-bit legacy mode, we have already unmapped the page */
739 if (!internal_config.legacy_mem)
740 munmap(hfile->orig_va, page_sz);
742 munmap(hfile->orig_va, page_sz);
745 hfile->orig_va = NULL;
746 hfile->final_va = addr;
748 /* rewrite physical addresses in IOVA as VA mode */
749 if (rte_eal_iova_mode() == RTE_IOVA_VA)
750 hfile->physaddr = (uintptr_t)addr;
752 /* set up memseg data */
754 ms->hugepage_sz = page_sz;
755 ms->len = memseg_len;
756 ms->iova = hfile->physaddr;
757 ms->socket_id = hfile->socket_id;
758 ms->nchannel = rte_memory_get_nchannel();
759 ms->nrank = rte_memory_get_nrank();
761 rte_fbarray_set_used(arr, ms_idx);
765 RTE_LOG(DEBUG, EAL, "Allocated %" PRIu64 "M on socket %i\n",
766 (seg_len * page_sz) >> 20, socket_id);
771 get_mem_amount(uint64_t page_sz, uint64_t max_mem)
773 uint64_t area_sz, max_pages;
775 /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */
776 max_pages = RTE_MAX_MEMSEG_PER_LIST;
777 max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem);
779 area_sz = RTE_MIN(page_sz * max_pages, max_mem);
781 /* make sure the list isn't smaller than the page size */
782 area_sz = RTE_MAX(area_sz, page_sz);
784 return RTE_ALIGN(area_sz, page_sz);
788 free_memseg_list(struct rte_memseg_list *msl)
790 if (rte_fbarray_destroy(&msl->memseg_arr)) {
791 RTE_LOG(ERR, EAL, "Cannot destroy memseg list\n");
794 memset(msl, 0, sizeof(*msl));
798 #define MEMSEG_LIST_FMT "memseg-%" PRIu64 "k-%i-%i"
800 alloc_memseg_list(struct rte_memseg_list *msl, uint64_t page_sz,
801 int n_segs, int socket_id, int type_msl_idx)
803 char name[RTE_FBARRAY_NAME_LEN];
805 snprintf(name, sizeof(name), MEMSEG_LIST_FMT, page_sz >> 10, socket_id,
807 if (rte_fbarray_init(&msl->memseg_arr, name, n_segs,
808 sizeof(struct rte_memseg))) {
809 RTE_LOG(ERR, EAL, "Cannot allocate memseg list: %s\n",
810 rte_strerror(rte_errno));
814 msl->page_sz = page_sz;
815 msl->socket_id = socket_id;
818 RTE_LOG(DEBUG, EAL, "Memseg list allocated: 0x%zxkB at socket %i\n",
819 (size_t)page_sz >> 10, socket_id);
825 alloc_va_space(struct rte_memseg_list *msl)
832 #ifdef RTE_ARCH_PPC_64
833 flags |= MAP_HUGETLB;
836 page_sz = msl->page_sz;
837 mem_sz = page_sz * msl->memseg_arr.len;
839 addr = eal_get_virtual_area(msl->base_va, &mem_sz, page_sz, 0, flags);
841 if (rte_errno == EADDRNOTAVAIL)
842 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes at [%p] - please use '--base-virtaddr' option\n",
843 (unsigned long long)mem_sz, msl->base_va);
845 RTE_LOG(ERR, EAL, "Cannot reserve memory\n");
854 * Our VA space is not preallocated yet, so preallocate it here. We need to know
855 * how many segments there are in order to map all pages into one address space,
856 * and leave appropriate holes between segments so that rte_malloc does not
857 * concatenate them into one big segment.
859 * we also need to unmap original pages to free up address space.
861 static int __rte_unused
862 prealloc_segments(struct hugepage_file *hugepages, int n_pages)
864 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
865 int cur_page, seg_start_page, end_seg, new_memseg;
866 unsigned int hpi_idx, socket, i;
867 int n_contig_segs, n_segs;
870 /* before we preallocate segments, we need to free up our VA space.
871 * we're not removing files, and we already have information about
872 * PA-contiguousness, so it is safe to unmap everything.
874 for (cur_page = 0; cur_page < n_pages; cur_page++) {
875 struct hugepage_file *hpi = &hugepages[cur_page];
876 munmap(hpi->orig_va, hpi->size);
880 /* we cannot know how many page sizes and sockets we have discovered, so
881 * loop over all of them
883 for (hpi_idx = 0; hpi_idx < internal_config.num_hugepage_sizes;
886 internal_config.hugepage_info[hpi_idx].hugepage_sz;
888 for (i = 0; i < rte_socket_count(); i++) {
889 struct rte_memseg_list *msl;
891 socket = rte_socket_id_by_idx(i);
896 for (cur_page = 0; cur_page < n_pages; cur_page++) {
897 struct hugepage_file *prev, *cur;
898 int prev_seg_start_page = -1;
900 cur = &hugepages[cur_page];
901 prev = cur_page == 0 ? NULL :
902 &hugepages[cur_page - 1];
909 else if (cur->socket_id != (int) socket)
911 else if (cur->size != page_sz)
913 else if (cur_page == 0)
915 #ifdef RTE_ARCH_PPC_64
916 /* On PPC64 architecture, the mmap always start
917 * from higher address to lower address. Here,
918 * physical addresses are in descending order.
920 else if ((prev->physaddr - cur->physaddr) !=
924 else if ((cur->physaddr - prev->physaddr) !=
929 /* if we're already inside a segment,
930 * new segment means end of current one
932 if (seg_start_page != -1) {
934 prev_seg_start_page =
937 seg_start_page = cur_page;
941 if (prev_seg_start_page != -1) {
942 /* we've found a new segment */
946 } else if (seg_start_page != -1) {
947 /* we didn't find new segment,
948 * but did end current one
956 /* we're skipping this page */
960 /* segment continues */
962 /* check if we missed last segment */
963 if (seg_start_page != -1) {
965 n_segs += cur_page - seg_start_page;
968 /* if no segments were found, do not preallocate */
972 /* we now have total number of pages that we will
973 * allocate for this segment list. add separator pages
974 * to the total count, and preallocate VA space.
976 n_segs += n_contig_segs - 1;
978 /* now, preallocate VA space for these segments */
980 /* first, find suitable memseg list for this */
981 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS;
983 msl = &mcfg->memsegs[msl_idx];
985 if (msl->base_va != NULL)
989 if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
990 RTE_LOG(ERR, EAL, "Not enough space in memseg lists, please increase %s\n",
991 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
995 /* now, allocate fbarray itself */
996 if (alloc_memseg_list(msl, page_sz, n_segs, socket,
1000 /* finally, allocate VA space */
1001 if (alloc_va_space(msl) < 0)
1009 * We cannot reallocate memseg lists on the fly because PPC64 stores pages
1010 * backwards, therefore we have to process the entire memseg first before
1011 * remapping it into memseg list VA space.
1014 remap_needed_hugepages(struct hugepage_file *hugepages, int n_pages)
1016 int cur_page, seg_start_page, new_memseg, ret;
1019 for (cur_page = 0; cur_page < n_pages; cur_page++) {
1020 struct hugepage_file *prev, *cur;
1024 cur = &hugepages[cur_page];
1025 prev = cur_page == 0 ? NULL : &hugepages[cur_page - 1];
1027 /* if size is zero, no more pages left */
1033 else if (cur->socket_id != prev->socket_id)
1035 else if (cur->size != prev->size)
1037 #ifdef RTE_ARCH_PPC_64
1038 /* On PPC64 architecture, the mmap always start from higher
1039 * address to lower address. Here, physical addresses are in
1042 else if ((prev->physaddr - cur->physaddr) != cur->size)
1045 else if ((cur->physaddr - prev->physaddr) != cur->size)
1050 /* if this isn't the first time, remap segment */
1051 if (cur_page != 0) {
1052 ret = remap_segment(hugepages, seg_start_page,
1057 /* remember where we started */
1058 seg_start_page = cur_page;
1060 /* continuation of previous memseg */
1062 /* we were stopped, but we didn't remap the last segment, do it now */
1063 if (cur_page != 0) {
1064 ret = remap_segment(hugepages, seg_start_page,
1072 static inline uint64_t
1073 get_socket_mem_size(int socket)
1078 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
1079 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1080 if (strnlen(hpi->hugedir, sizeof(hpi->hugedir)) != 0)
1081 size += hpi->hugepage_sz * hpi->num_pages[socket];
1088 * This function is a NUMA-aware equivalent of calc_num_pages.
1089 * It takes in the list of hugepage sizes and the
1090 * number of pages thereof, and calculates the best number of
1091 * pages of each size to fulfill the request for <memory> ram
1094 calc_num_pages_per_socket(uint64_t * memory,
1095 struct hugepage_info *hp_info,
1096 struct hugepage_info *hp_used,
1097 unsigned num_hp_info)
1099 unsigned socket, j, i = 0;
1100 unsigned requested, available;
1101 int total_num_pages = 0;
1102 uint64_t remaining_mem, cur_mem;
1103 uint64_t total_mem = internal_config.memory;
1105 if (num_hp_info == 0)
1108 /* if specific memory amounts per socket weren't requested */
1109 if (internal_config.force_sockets == 0) {
1112 int cpu_per_socket[RTE_MAX_NUMA_NODES];
1113 size_t default_size;
1116 /* Compute number of cores per socket */
1117 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
1118 RTE_LCORE_FOREACH(lcore_id) {
1119 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
1123 * Automatically spread requested memory amongst detected sockets according
1124 * to number of cores from cpu mask present on each socket
1126 total_size = internal_config.memory;
1127 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1129 /* Set memory amount per socket */
1130 default_size = (internal_config.memory * cpu_per_socket[socket])
1131 / rte_lcore_count();
1133 /* Limit to maximum available memory on socket */
1134 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
1137 memory[socket] = default_size;
1138 total_size -= default_size;
1142 * If some memory is remaining, try to allocate it by getting all
1143 * available memory from sockets, one after the other
1145 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1146 /* take whatever is available */
1147 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
1151 memory[socket] += default_size;
1152 total_size -= default_size;
1155 /* in 32-bit mode, allocate all of the memory only on master
1158 total_size = internal_config.memory;
1159 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0;
1161 struct rte_config *cfg = rte_eal_get_configuration();
1162 unsigned int master_lcore_socket;
1164 master_lcore_socket =
1165 rte_lcore_to_socket_id(cfg->master_lcore);
1167 if (master_lcore_socket != socket)
1171 memory[socket] = total_size;
1177 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
1178 /* skips if the memory on specific socket wasn't requested */
1179 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
1180 strlcpy(hp_used[i].hugedir, hp_info[i].hugedir,
1181 sizeof(hp_used[i].hugedir));
1182 hp_used[i].num_pages[socket] = RTE_MIN(
1183 memory[socket] / hp_info[i].hugepage_sz,
1184 hp_info[i].num_pages[socket]);
1186 cur_mem = hp_used[i].num_pages[socket] *
1187 hp_used[i].hugepage_sz;
1189 memory[socket] -= cur_mem;
1190 total_mem -= cur_mem;
1192 total_num_pages += hp_used[i].num_pages[socket];
1194 /* check if we have met all memory requests */
1195 if (memory[socket] == 0)
1198 /* check if we have any more pages left at this size, if so
1199 * move on to next size */
1200 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
1202 /* At this point we know that there are more pages available that are
1203 * bigger than the memory we want, so lets see if we can get enough
1204 * from other page sizes.
1207 for (j = i+1; j < num_hp_info; j++)
1208 remaining_mem += hp_info[j].hugepage_sz *
1209 hp_info[j].num_pages[socket];
1211 /* is there enough other memory, if not allocate another page and quit */
1212 if (remaining_mem < memory[socket]){
1213 cur_mem = RTE_MIN(memory[socket],
1214 hp_info[i].hugepage_sz);
1215 memory[socket] -= cur_mem;
1216 total_mem -= cur_mem;
1217 hp_used[i].num_pages[socket]++;
1219 break; /* we are done with this socket*/
1222 /* if we didn't satisfy all memory requirements per socket */
1223 if (memory[socket] > 0 &&
1224 internal_config.socket_mem[socket] != 0) {
1225 /* to prevent icc errors */
1226 requested = (unsigned) (internal_config.socket_mem[socket] /
1228 available = requested -
1229 ((unsigned) (memory[socket] / 0x100000));
1230 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
1231 "Requested: %uMB, available: %uMB\n", socket,
1232 requested, available);
1237 /* if we didn't satisfy total memory requirements */
1238 if (total_mem > 0) {
1239 requested = (unsigned) (internal_config.memory / 0x100000);
1240 available = requested - (unsigned) (total_mem / 0x100000);
1241 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
1242 " available: %uMB\n", requested, available);
1245 return total_num_pages;
1248 static inline size_t
1249 eal_get_hugepage_mem_size(void)
1254 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
1255 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1256 if (strnlen(hpi->hugedir, sizeof(hpi->hugedir)) != 0) {
1257 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1258 size += hpi->hugepage_sz * hpi->num_pages[j];
1263 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
1266 static struct sigaction huge_action_old;
1267 static int huge_need_recover;
1270 huge_register_sigbus(void)
1273 struct sigaction action;
1276 sigaddset(&mask, SIGBUS);
1277 action.sa_flags = 0;
1278 action.sa_mask = mask;
1279 action.sa_handler = huge_sigbus_handler;
1281 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1285 huge_recover_sigbus(void)
1287 if (huge_need_recover) {
1288 sigaction(SIGBUS, &huge_action_old, NULL);
1289 huge_need_recover = 0;
1294 * Prepare physical memory mapping: fill configuration structure with
1295 * these infos, return 0 on success.
1296 * 1. map N huge pages in separate files in hugetlbfs
1297 * 2. find associated physical addr
1298 * 3. find associated NUMA socket ID
1299 * 4. sort all huge pages by physical address
1300 * 5. remap these N huge pages in the correct order
1301 * 6. unmap the first mapping
1302 * 7. fill memsegs in configuration with contiguous zones
1305 eal_legacy_hugepage_init(void)
1307 struct rte_mem_config *mcfg;
1308 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1309 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1310 struct rte_fbarray *arr;
1311 struct rte_memseg *ms;
1313 uint64_t memory[RTE_MAX_NUMA_NODES];
1317 int nr_hugefiles, nr_hugepages = 0;
1320 test_phys_addrs_available();
1322 memset(used_hp, 0, sizeof(used_hp));
1324 /* get pointer to global configuration */
1325 mcfg = rte_eal_get_configuration()->mem_config;
1327 /* hugetlbfs can be disabled */
1328 if (internal_config.no_hugetlbfs) {
1329 struct rte_memseg_list *msl;
1331 int n_segs, cur_seg;
1333 /* nohuge mode is legacy mode */
1334 internal_config.legacy_mem = 1;
1336 /* create a memseg list */
1337 msl = &mcfg->memsegs[0];
1339 page_sz = RTE_PGSIZE_4K;
1340 n_segs = internal_config.memory / page_sz;
1342 if (rte_fbarray_init(&msl->memseg_arr, "nohugemem", n_segs,
1343 sizeof(struct rte_memseg))) {
1344 RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
1348 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1349 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
1350 if (addr == MAP_FAILED) {
1351 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1355 msl->base_va = addr;
1356 msl->page_sz = page_sz;
1359 /* populate memsegs. each memseg is one page long */
1360 for (cur_seg = 0; cur_seg < n_segs; cur_seg++) {
1361 arr = &msl->memseg_arr;
1363 ms = rte_fbarray_get(arr, cur_seg);
1364 if (rte_eal_iova_mode() == RTE_IOVA_VA)
1365 ms->iova = (uintptr_t)addr;
1367 ms->iova = RTE_BAD_IOVA;
1369 ms->hugepage_sz = page_sz;
1373 rte_fbarray_set_used(arr, cur_seg);
1375 addr = RTE_PTR_ADD(addr, (size_t)page_sz);
1380 /* calculate total number of hugepages available. at this point we haven't
1381 * yet started sorting them so they all are on socket 0 */
1382 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1383 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1384 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1386 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1390 * allocate a memory area for hugepage table.
1391 * this isn't shared memory yet. due to the fact that we need some
1392 * processing done on these pages, shared memory will be created
1395 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1399 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1401 hp_offset = 0; /* where we start the current page size entries */
1403 huge_register_sigbus();
1405 /* make a copy of socket_mem, needed for balanced allocation. */
1406 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1407 memory[i] = internal_config.socket_mem[i];
1409 /* map all hugepages and sort them */
1410 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1411 unsigned pages_old, pages_new;
1412 struct hugepage_info *hpi;
1415 * we don't yet mark hugepages as used at this stage, so
1416 * we just map all hugepages available to the system
1417 * all hugepages are still located on socket 0
1419 hpi = &internal_config.hugepage_info[i];
1421 if (hpi->num_pages[0] == 0)
1424 /* map all hugepages available */
1425 pages_old = hpi->num_pages[0];
1426 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, memory);
1427 if (pages_new < pages_old) {
1429 "%d not %d hugepages of size %u MB allocated\n",
1430 pages_new, pages_old,
1431 (unsigned)(hpi->hugepage_sz / 0x100000));
1433 int pages = pages_old - pages_new;
1435 nr_hugepages -= pages;
1436 hpi->num_pages[0] = pages_new;
1441 if (phys_addrs_available &&
1442 rte_eal_iova_mode() != RTE_IOVA_VA) {
1443 /* find physical addresses for each hugepage */
1444 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1445 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1446 "for %u MB pages\n",
1447 (unsigned int)(hpi->hugepage_sz / 0x100000));
1451 /* set physical addresses for each hugepage */
1452 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1453 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1454 "for %u MB pages\n",
1455 (unsigned int)(hpi->hugepage_sz / 0x100000));
1460 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1461 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1462 (unsigned)(hpi->hugepage_sz / 0x100000));
1466 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1467 sizeof(struct hugepage_file), cmp_physaddr);
1469 /* we have processed a num of hugepages of this size, so inc offset */
1470 hp_offset += hpi->num_pages[0];
1473 huge_recover_sigbus();
1475 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1476 internal_config.memory = eal_get_hugepage_mem_size();
1478 nr_hugefiles = nr_hugepages;
1481 /* clean out the numbers of pages */
1482 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1483 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1484 internal_config.hugepage_info[i].num_pages[j] = 0;
1486 /* get hugepages for each socket */
1487 for (i = 0; i < nr_hugefiles; i++) {
1488 int socket = tmp_hp[i].socket_id;
1490 /* find a hugepage info with right size and increment num_pages */
1491 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1492 (int)internal_config.num_hugepage_sizes);
1493 for (j = 0; j < nb_hpsizes; j++) {
1494 if (tmp_hp[i].size ==
1495 internal_config.hugepage_info[j].hugepage_sz) {
1496 internal_config.hugepage_info[j].num_pages[socket]++;
1501 /* make a copy of socket_mem, needed for number of pages calculation */
1502 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1503 memory[i] = internal_config.socket_mem[i];
1505 /* calculate final number of pages */
1506 nr_hugepages = calc_num_pages_per_socket(memory,
1507 internal_config.hugepage_info, used_hp,
1508 internal_config.num_hugepage_sizes);
1510 /* error if not enough memory available */
1511 if (nr_hugepages < 0)
1515 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1516 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1517 if (used_hp[i].num_pages[j] > 0) {
1519 "Requesting %u pages of size %uMB"
1520 " from socket %i\n",
1521 used_hp[i].num_pages[j],
1523 (used_hp[i].hugepage_sz / 0x100000),
1529 /* create shared memory */
1530 hugepage = create_shared_memory(eal_hugepage_data_path(),
1531 nr_hugefiles * sizeof(struct hugepage_file));
1533 if (hugepage == NULL) {
1534 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1537 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1540 * unmap pages that we won't need (looks at used_hp).
1541 * also, sets final_va to NULL on pages that were unmapped.
1543 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1544 internal_config.num_hugepage_sizes) < 0) {
1545 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1550 * copy stuff from malloc'd hugepage* to the actual shared memory.
1551 * this procedure only copies those hugepages that have orig_va
1552 * not NULL. has overflow protection.
1554 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1555 tmp_hp, nr_hugefiles) < 0) {
1556 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1561 /* for legacy 32-bit mode, we did not preallocate VA space, so do it */
1562 if (internal_config.legacy_mem &&
1563 prealloc_segments(hugepage, nr_hugefiles)) {
1564 RTE_LOG(ERR, EAL, "Could not preallocate VA space for hugepages\n");
1569 /* remap all pages we do need into memseg list VA space, so that those
1570 * pages become first-class citizens in DPDK memory subsystem
1572 if (remap_needed_hugepages(hugepage, nr_hugefiles)) {
1573 RTE_LOG(ERR, EAL, "Couldn't remap hugepage files into memseg lists\n");
1577 /* free the hugepage backing files */
1578 if (internal_config.hugepage_unlink &&
1579 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1580 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1584 /* free the temporary hugepage table */
1588 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1590 /* we're not going to allocate more pages, so release VA space for
1591 * unused memseg lists
1593 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1594 struct rte_memseg_list *msl = &mcfg->memsegs[i];
1597 /* skip inactive lists */
1598 if (msl->base_va == NULL)
1600 /* skip lists where there is at least one page allocated */
1601 if (msl->memseg_arr.count > 0)
1603 /* this is an unused list, deallocate it */
1604 mem_sz = (size_t)msl->page_sz * msl->memseg_arr.len;
1605 munmap(msl->base_va, mem_sz);
1606 msl->base_va = NULL;
1608 /* destroy backing fbarray */
1609 rte_fbarray_destroy(&msl->memseg_arr);
1615 huge_recover_sigbus();
1617 if (hugepage != NULL)
1618 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1623 static int __rte_unused
1624 hugepage_count_walk(const struct rte_memseg_list *msl, void *arg)
1626 struct hugepage_info *hpi = arg;
1628 if (msl->page_sz != hpi->hugepage_sz)
1631 hpi->num_pages[msl->socket_id] += msl->memseg_arr.len;
1636 limits_callback(int socket_id, size_t cur_limit, size_t new_len)
1638 RTE_SET_USED(socket_id);
1639 RTE_SET_USED(cur_limit);
1640 RTE_SET_USED(new_len);
1645 eal_hugepage_init(void)
1647 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1648 uint64_t memory[RTE_MAX_NUMA_NODES];
1649 int hp_sz_idx, socket_id;
1651 test_phys_addrs_available();
1653 memset(used_hp, 0, sizeof(used_hp));
1656 hp_sz_idx < (int) internal_config.num_hugepage_sizes;
1659 struct hugepage_info dummy;
1662 /* also initialize used_hp hugepage sizes in used_hp */
1663 struct hugepage_info *hpi;
1664 hpi = &internal_config.hugepage_info[hp_sz_idx];
1665 used_hp[hp_sz_idx].hugepage_sz = hpi->hugepage_sz;
1668 /* for 32-bit, limit number of pages on socket to whatever we've
1669 * preallocated, as we cannot allocate more.
1671 memset(&dummy, 0, sizeof(dummy));
1672 dummy.hugepage_sz = hpi->hugepage_sz;
1673 if (rte_memseg_list_walk(hugepage_count_walk, &dummy) < 0)
1676 for (i = 0; i < RTE_DIM(dummy.num_pages); i++) {
1677 hpi->num_pages[i] = RTE_MIN(hpi->num_pages[i],
1678 dummy.num_pages[i]);
1683 /* make a copy of socket_mem, needed for balanced allocation. */
1684 for (hp_sz_idx = 0; hp_sz_idx < RTE_MAX_NUMA_NODES; hp_sz_idx++)
1685 memory[hp_sz_idx] = internal_config.socket_mem[hp_sz_idx];
1687 /* calculate final number of pages */
1688 if (calc_num_pages_per_socket(memory,
1689 internal_config.hugepage_info, used_hp,
1690 internal_config.num_hugepage_sizes) < 0)
1694 hp_sz_idx < (int)internal_config.num_hugepage_sizes;
1696 for (socket_id = 0; socket_id < RTE_MAX_NUMA_NODES;
1698 struct rte_memseg **pages;
1699 struct hugepage_info *hpi = &used_hp[hp_sz_idx];
1700 unsigned int num_pages = hpi->num_pages[socket_id];
1701 int num_pages_alloc, i;
1706 pages = malloc(sizeof(*pages) * num_pages);
1708 RTE_LOG(DEBUG, EAL, "Allocating %u pages of size %" PRIu64 "M on socket %i\n",
1709 num_pages, hpi->hugepage_sz >> 20, socket_id);
1711 num_pages_alloc = eal_memalloc_alloc_seg_bulk(pages,
1712 num_pages, hpi->hugepage_sz,
1714 if (num_pages_alloc < 0) {
1719 /* mark preallocated pages as unfreeable */
1720 for (i = 0; i < num_pages_alloc; i++) {
1721 struct rte_memseg *ms = pages[i];
1722 ms->flags |= RTE_MEMSEG_FLAG_DO_NOT_FREE;
1727 /* if socket limits were specified, set them */
1728 if (internal_config.force_socket_limits) {
1730 for (i = 0; i < RTE_MAX_NUMA_NODES; i++) {
1731 uint64_t limit = internal_config.socket_limit[i];
1734 if (rte_mem_alloc_validator_register("socket-limit",
1735 limits_callback, i, limit))
1736 RTE_LOG(ERR, EAL, "Failed to register socket limits validator callback\n");
1743 * uses fstat to report the size of a file on disk
1749 if (fstat(fd, &st) < 0)
1755 * This creates the memory mappings in the secondary process to match that of
1756 * the server process. It goes through each memory segment in the DPDK runtime
1757 * configuration and finds the hugepages which form that segment, mapping them
1758 * in order to form a contiguous block in the virtual memory space
1761 eal_legacy_hugepage_attach(void)
1763 struct hugepage_file *hp = NULL;
1764 unsigned int num_hp = 0;
1766 unsigned int cur_seg;
1768 int fd, fd_hugepage = -1;
1770 if (aslr_enabled() > 0) {
1771 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1772 "(ASLR) is enabled in the kernel.\n");
1773 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1774 "into secondary processes\n");
1777 test_phys_addrs_available();
1779 fd_hugepage = open(eal_hugepage_data_path(), O_RDONLY);
1780 if (fd_hugepage < 0) {
1781 RTE_LOG(ERR, EAL, "Could not open %s\n",
1782 eal_hugepage_data_path());
1786 size = getFileSize(fd_hugepage);
1787 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1788 if (hp == MAP_FAILED) {
1789 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1790 eal_hugepage_data_path());
1794 num_hp = size / sizeof(struct hugepage_file);
1795 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1797 /* map all segments into memory to make sure we get the addrs. the
1798 * segments themselves are already in memseg list (which is shared and
1799 * has its VA space already preallocated), so we just need to map
1800 * everything into correct addresses.
1802 for (i = 0; i < num_hp; i++) {
1803 struct hugepage_file *hf = &hp[i];
1804 size_t map_sz = hf->size;
1805 void *map_addr = hf->final_va;
1807 /* if size is zero, no more pages left */
1811 fd = open(hf->filepath, O_RDWR);
1813 RTE_LOG(ERR, EAL, "Could not open %s: %s\n",
1814 hf->filepath, strerror(errno));
1818 map_addr = mmap(map_addr, map_sz, PROT_READ | PROT_WRITE,
1819 MAP_SHARED | MAP_FIXED, fd, 0);
1820 if (map_addr == MAP_FAILED) {
1821 RTE_LOG(ERR, EAL, "Could not map %s: %s\n",
1822 hf->filepath, strerror(errno));
1827 /* set shared lock on the file. */
1828 if (flock(fd, LOCK_SH) < 0) {
1829 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed: %s\n",
1830 __func__, strerror(errno));
1837 /* unmap the hugepage config file, since we are done using it */
1843 /* map all segments into memory to make sure we get the addrs */
1845 for (cur_seg = 0; cur_seg < i; cur_seg++) {
1846 struct hugepage_file *hf = &hp[i];
1847 size_t map_sz = hf->size;
1848 void *map_addr = hf->final_va;
1850 munmap(map_addr, map_sz);
1852 if (hp != NULL && hp != MAP_FAILED)
1854 if (fd_hugepage >= 0)
1860 eal_hugepage_attach(void)
1862 if (eal_memalloc_sync_with_primary()) {
1863 RTE_LOG(ERR, EAL, "Could not map memory from primary process\n");
1864 if (aslr_enabled() > 0)
1865 RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel and retry running both primary and secondary processes\n");
1872 rte_eal_hugepage_init(void)
1874 return internal_config.legacy_mem ?
1875 eal_legacy_hugepage_init() :
1876 eal_hugepage_init();
1880 rte_eal_hugepage_attach(void)
1882 return internal_config.legacy_mem ?
1883 eal_legacy_hugepage_attach() :
1884 eal_hugepage_attach();
1888 rte_eal_using_phys_addrs(void)
1890 return phys_addrs_available;
1893 static int __rte_unused
1894 memseg_primary_init_32(void)
1896 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1897 int active_sockets, hpi_idx, msl_idx = 0;
1898 unsigned int socket_id, i;
1899 struct rte_memseg_list *msl;
1900 uint64_t extra_mem_per_socket, total_extra_mem, total_requested_mem;
1903 /* no-huge does not need this at all */
1904 if (internal_config.no_hugetlbfs)
1907 /* this is a giant hack, but desperate times call for desperate
1908 * measures. in legacy 32-bit mode, we cannot preallocate VA space,
1909 * because having upwards of 2 gigabytes of VA space already mapped will
1910 * interfere with our ability to map and sort hugepages.
1912 * therefore, in legacy 32-bit mode, we will be initializing memseg
1913 * lists much later - in eal_memory.c, right after we unmap all the
1914 * unneeded pages. this will not affect secondary processes, as those
1915 * should be able to mmap the space without (too many) problems.
1917 if (internal_config.legacy_mem)
1920 /* 32-bit mode is a very special case. we cannot know in advance where
1921 * the user will want to allocate their memory, so we have to do some
1925 total_requested_mem = 0;
1926 if (internal_config.force_sockets)
1927 for (i = 0; i < rte_socket_count(); i++) {
1930 socket_id = rte_socket_id_by_idx(i);
1931 mem = internal_config.socket_mem[socket_id];
1937 total_requested_mem += mem;
1940 total_requested_mem = internal_config.memory;
1942 max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
1943 if (total_requested_mem > max_mem) {
1944 RTE_LOG(ERR, EAL, "Invalid parameters: 32-bit process can at most use %uM of memory\n",
1945 (unsigned int)(max_mem >> 20));
1948 total_extra_mem = max_mem - total_requested_mem;
1949 extra_mem_per_socket = active_sockets == 0 ? total_extra_mem :
1950 total_extra_mem / active_sockets;
1952 /* the allocation logic is a little bit convoluted, but here's how it
1953 * works, in a nutshell:
1954 * - if user hasn't specified on which sockets to allocate memory via
1955 * --socket-mem, we allocate all of our memory on master core socket.
1956 * - if user has specified sockets to allocate memory on, there may be
1957 * some "unused" memory left (e.g. if user has specified --socket-mem
1958 * such that not all memory adds up to 2 gigabytes), so add it to all
1959 * sockets that are in use equally.
1961 * page sizes are sorted by size in descending order, so we can safely
1962 * assume that we dispense with bigger page sizes first.
1965 /* create memseg lists */
1966 for (i = 0; i < rte_socket_count(); i++) {
1967 int hp_sizes = (int) internal_config.num_hugepage_sizes;
1968 uint64_t max_socket_mem, cur_socket_mem;
1969 unsigned int master_lcore_socket;
1970 struct rte_config *cfg = rte_eal_get_configuration();
1973 socket_id = rte_socket_id_by_idx(i);
1975 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
1980 /* if we didn't specifically request memory on this socket */
1981 skip = active_sockets != 0 &&
1982 internal_config.socket_mem[socket_id] == 0;
1983 /* ...or if we didn't specifically request memory on *any*
1984 * socket, and this is not master lcore
1986 master_lcore_socket = rte_lcore_to_socket_id(cfg->master_lcore);
1987 skip |= active_sockets == 0 && socket_id != master_lcore_socket;
1990 RTE_LOG(DEBUG, EAL, "Will not preallocate memory on socket %u\n",
1995 /* max amount of memory on this socket */
1996 max_socket_mem = (active_sockets != 0 ?
1997 internal_config.socket_mem[socket_id] :
1998 internal_config.memory) +
1999 extra_mem_per_socket;
2002 for (hpi_idx = 0; hpi_idx < hp_sizes; hpi_idx++) {
2003 uint64_t max_pagesz_mem, cur_pagesz_mem = 0;
2004 uint64_t hugepage_sz;
2005 struct hugepage_info *hpi;
2006 int type_msl_idx, max_segs, total_segs = 0;
2008 hpi = &internal_config.hugepage_info[hpi_idx];
2009 hugepage_sz = hpi->hugepage_sz;
2011 /* check if pages are actually available */
2012 if (hpi->num_pages[socket_id] == 0)
2015 max_segs = RTE_MAX_MEMSEG_PER_TYPE;
2016 max_pagesz_mem = max_socket_mem - cur_socket_mem;
2018 /* make it multiple of page size */
2019 max_pagesz_mem = RTE_ALIGN_FLOOR(max_pagesz_mem,
2022 RTE_LOG(DEBUG, EAL, "Attempting to preallocate "
2023 "%" PRIu64 "M on socket %i\n",
2024 max_pagesz_mem >> 20, socket_id);
2027 while (cur_pagesz_mem < max_pagesz_mem &&
2028 total_segs < max_segs) {
2030 unsigned int n_segs;
2032 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2034 "No more space in memseg lists, please increase %s\n",
2035 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2039 msl = &mcfg->memsegs[msl_idx];
2041 cur_mem = get_mem_amount(hugepage_sz,
2043 n_segs = cur_mem / hugepage_sz;
2045 if (alloc_memseg_list(msl, hugepage_sz, n_segs,
2046 socket_id, type_msl_idx)) {
2047 /* failing to allocate a memseg list is
2050 RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
2054 if (alloc_va_space(msl)) {
2055 /* if we couldn't allocate VA space, we
2056 * can try with smaller page sizes.
2058 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list, retrying with different page size\n");
2059 /* deallocate memseg list */
2060 if (free_memseg_list(msl))
2065 total_segs += msl->memseg_arr.len;
2066 cur_pagesz_mem = total_segs * hugepage_sz;
2070 cur_socket_mem += cur_pagesz_mem;
2072 if (cur_socket_mem == 0) {
2073 RTE_LOG(ERR, EAL, "Cannot allocate VA space on socket %u\n",
2082 static int __rte_unused
2083 memseg_primary_init(void)
2085 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2086 int i, socket_id, hpi_idx, msl_idx = 0;
2087 struct rte_memseg_list *msl;
2088 uint64_t max_mem, total_mem;
2090 /* no-huge does not need this at all */
2091 if (internal_config.no_hugetlbfs)
2094 max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
2097 /* create memseg lists */
2098 for (hpi_idx = 0; hpi_idx < (int) internal_config.num_hugepage_sizes;
2100 struct hugepage_info *hpi;
2101 uint64_t hugepage_sz;
2103 hpi = &internal_config.hugepage_info[hpi_idx];
2104 hugepage_sz = hpi->hugepage_sz;
2106 for (i = 0; i < (int) rte_socket_count(); i++) {
2107 uint64_t max_type_mem, total_type_mem = 0;
2108 int type_msl_idx, max_segs, total_segs = 0;
2110 socket_id = rte_socket_id_by_idx(i);
2112 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
2117 if (total_mem >= max_mem)
2120 max_type_mem = RTE_MIN(max_mem - total_mem,
2121 (uint64_t)RTE_MAX_MEM_MB_PER_TYPE << 20);
2122 max_segs = RTE_MAX_MEMSEG_PER_TYPE;
2125 while (total_type_mem < max_type_mem &&
2126 total_segs < max_segs) {
2127 uint64_t cur_max_mem, cur_mem;
2128 unsigned int n_segs;
2130 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2132 "No more space in memseg lists, please increase %s\n",
2133 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2137 msl = &mcfg->memsegs[msl_idx++];
2139 cur_max_mem = max_type_mem - total_type_mem;
2141 cur_mem = get_mem_amount(hugepage_sz,
2143 n_segs = cur_mem / hugepage_sz;
2145 if (alloc_memseg_list(msl, hugepage_sz, n_segs,
2146 socket_id, type_msl_idx))
2149 total_segs += msl->memseg_arr.len;
2150 total_type_mem = total_segs * hugepage_sz;
2153 if (alloc_va_space(msl)) {
2154 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list\n");
2158 total_mem += total_type_mem;
2165 memseg_secondary_init(void)
2167 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2169 struct rte_memseg_list *msl;
2171 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
2173 msl = &mcfg->memsegs[msl_idx];
2175 /* skip empty memseg lists */
2176 if (msl->memseg_arr.len == 0)
2179 if (rte_fbarray_attach(&msl->memseg_arr)) {
2180 RTE_LOG(ERR, EAL, "Cannot attach to primary process memseg lists\n");
2184 /* preallocate VA space */
2185 if (alloc_va_space(msl)) {
2186 RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
2195 rte_eal_memseg_init(void)
2197 return rte_eal_process_type() == RTE_PROC_PRIMARY ?
2199 memseg_primary_init_32() :
2201 memseg_primary_init() :
2203 memseg_secondary_init();