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
17 #include <sys/types.h>
19 #include <sys/queue.h>
21 #include <sys/resource.h>
24 #include <sys/ioctl.h>
28 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
33 #include <rte_errno.h>
35 #include <rte_memory.h>
36 #include <rte_launch.h>
38 #include <rte_eal_memconfig.h>
39 #include <rte_per_lcore.h>
40 #include <rte_lcore.h>
41 #include <rte_common.h>
42 #include <rte_string_fns.h>
44 #include "eal_private.h"
45 #include "eal_memalloc.h"
46 #include "eal_internal_cfg.h"
47 #include "eal_filesystem.h"
48 #include "eal_hugepages.h"
50 #define PFN_MASK_SIZE 8
54 * Huge page mapping under linux
56 * To reserve a big contiguous amount of memory, we use the hugepage
57 * feature of linux. For that, we need to have hugetlbfs mounted. This
58 * code will create many files in this directory (one per page) and
59 * map them in virtual memory. For each page, we will retrieve its
60 * physical address and remap it in order to have a virtual contiguous
61 * zone as well as a physical contiguous zone.
64 static bool phys_addrs_available = true;
66 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
69 test_phys_addrs_available(void)
74 if (!rte_eal_has_hugepages()) {
76 "Started without hugepages support, physical addresses not available\n");
77 phys_addrs_available = false;
81 physaddr = rte_mem_virt2phy(&tmp);
82 if (physaddr == RTE_BAD_PHYS_ADDR) {
83 if (rte_eal_iova_mode() == RTE_IOVA_PA)
85 "Cannot obtain physical addresses: %s. "
86 "Only vfio will function.\n",
88 phys_addrs_available = false;
93 * Get physical address of any mapped virtual address in the current process.
96 rte_mem_virt2phy(const void *virtaddr)
99 uint64_t page, physaddr;
100 unsigned long virt_pfn;
104 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
105 if (!phys_addrs_available)
108 /* standard page size */
109 page_size = getpagesize();
111 fd = open("/proc/self/pagemap", O_RDONLY);
113 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
114 __func__, strerror(errno));
118 virt_pfn = (unsigned long)virtaddr / page_size;
119 offset = sizeof(uint64_t) * virt_pfn;
120 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
121 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
122 __func__, strerror(errno));
127 retval = read(fd, &page, PFN_MASK_SIZE);
130 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
131 __func__, strerror(errno));
133 } else if (retval != PFN_MASK_SIZE) {
134 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
135 "but expected %d:\n",
136 __func__, retval, PFN_MASK_SIZE);
141 * the pfn (page frame number) are bits 0-54 (see
142 * pagemap.txt in linux Documentation)
144 if ((page & 0x7fffffffffffffULL) == 0)
147 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
148 + ((unsigned long)virtaddr % page_size);
154 rte_mem_virt2iova(const void *virtaddr)
156 if (rte_eal_iova_mode() == RTE_IOVA_VA)
157 return (uintptr_t)virtaddr;
158 return rte_mem_virt2phy(virtaddr);
162 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
163 * it by browsing the /proc/self/pagemap special file.
166 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
171 for (i = 0; i < hpi->num_pages[0]; i++) {
172 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
173 if (addr == RTE_BAD_PHYS_ADDR)
175 hugepg_tbl[i].physaddr = addr;
181 * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
184 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
187 static phys_addr_t addr;
189 for (i = 0; i < hpi->num_pages[0]; i++) {
190 hugepg_tbl[i].physaddr = addr;
191 addr += hugepg_tbl[i].size;
197 * Check whether address-space layout randomization is enabled in
198 * the kernel. This is important for multi-process as it can prevent
199 * two processes mapping data to the same virtual address
201 * 0 - address space randomization disabled
202 * 1/2 - address space randomization enabled
203 * negative error code on error
209 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
212 retval = read(fd, &c, 1);
222 default: return -EINVAL;
226 static sigjmp_buf huge_jmpenv;
228 static void huge_sigbus_handler(int signo __rte_unused)
230 siglongjmp(huge_jmpenv, 1);
233 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
234 * non-static local variable in the stack frame calling sigsetjmp might be
235 * clobbered by a call to longjmp.
237 static int huge_wrap_sigsetjmp(void)
239 return sigsetjmp(huge_jmpenv, 1);
242 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
243 /* Callback for numa library. */
244 void numa_error(char *where)
246 RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
251 * Mmap all hugepages of hugepage table: it first open a file in
252 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
253 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
254 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
255 * map contiguous physical blocks in contiguous virtual blocks.
258 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
259 uint64_t *essential_memory __rte_unused)
264 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
266 int essential_prev = 0;
268 struct bitmask *oldmask = NULL;
269 bool have_numa = true;
270 unsigned long maxnode = 0;
272 /* Check if kernel supports NUMA. */
273 if (numa_available() != 0) {
274 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
279 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
280 oldmask = numa_allocate_nodemask();
281 if (get_mempolicy(&oldpolicy, oldmask->maskp,
282 oldmask->size + 1, 0, 0) < 0) {
284 "Failed to get current mempolicy: %s. "
285 "Assuming MPOL_DEFAULT.\n", strerror(errno));
286 oldpolicy = MPOL_DEFAULT;
288 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
289 if (internal_config.socket_mem[i])
294 for (i = 0; i < hpi->num_pages[0]; i++) {
295 struct hugepage_file *hf = &hugepg_tbl[i];
296 uint64_t hugepage_sz = hpi->hugepage_sz;
298 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
302 for (j = 0; j < maxnode; j++)
303 if (essential_memory[j])
307 node_id = (node_id + 1) % maxnode;
308 while (!internal_config.socket_mem[node_id]) {
315 essential_prev = essential_memory[j];
317 if (essential_memory[j] < hugepage_sz)
318 essential_memory[j] = 0;
320 essential_memory[j] -= hugepage_sz;
324 "Setting policy MPOL_PREFERRED for socket %d\n",
326 numa_set_preferred(node_id);
331 hf->size = hugepage_sz;
332 eal_get_hugefile_path(hf->filepath, sizeof(hf->filepath),
333 hpi->hugedir, hf->file_id);
334 hf->filepath[sizeof(hf->filepath) - 1] = '\0';
336 /* try to create hugepage file */
337 fd = open(hf->filepath, O_CREAT | O_RDWR, 0600);
339 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
344 /* map the segment, and populate page tables,
345 * the kernel fills this segment with zeros. we don't care where
346 * this gets mapped - we already have contiguous memory areas
347 * ready for us to map into.
349 virtaddr = mmap(NULL, hugepage_sz, PROT_READ | PROT_WRITE,
350 MAP_SHARED | MAP_POPULATE, fd, 0);
351 if (virtaddr == MAP_FAILED) {
352 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
358 hf->orig_va = virtaddr;
360 /* In linux, hugetlb limitations, like cgroup, are
361 * enforced at fault time instead of mmap(), even
362 * with the option of MAP_POPULATE. Kernel will send
363 * a SIGBUS signal. To avoid to be killed, save stack
364 * environment here, if SIGBUS happens, we can jump
367 if (huge_wrap_sigsetjmp()) {
368 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
369 "hugepages of size %u MB\n",
370 (unsigned int)(hugepage_sz / 0x100000));
371 munmap(virtaddr, hugepage_sz);
373 unlink(hugepg_tbl[i].filepath);
374 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
376 essential_memory[node_id] =
381 *(int *)virtaddr = 0;
383 /* set shared lock on the file. */
384 if (flock(fd, LOCK_SH) < 0) {
385 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
386 __func__, strerror(errno));
395 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
398 "Restoring previous memory policy: %d\n", oldpolicy);
399 if (oldpolicy == MPOL_DEFAULT) {
400 numa_set_localalloc();
401 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
402 oldmask->size + 1) < 0) {
403 RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
405 numa_set_localalloc();
409 numa_free_cpumask(oldmask);
415 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
419 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
423 unsigned i, hp_count = 0;
426 char hugedir_str[PATH_MAX];
429 f = fopen("/proc/self/numa_maps", "r");
431 RTE_LOG(NOTICE, EAL, "NUMA support not available"
432 " consider that all memory is in socket_id 0\n");
436 snprintf(hugedir_str, sizeof(hugedir_str),
437 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
440 while (fgets(buf, sizeof(buf), f) != NULL) {
442 /* ignore non huge page */
443 if (strstr(buf, " huge ") == NULL &&
444 strstr(buf, hugedir_str) == NULL)
448 virt_addr = strtoull(buf, &end, 16);
449 if (virt_addr == 0 || end == buf) {
450 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
454 /* get node id (socket id) */
455 nodestr = strstr(buf, " N");
456 if (nodestr == NULL) {
457 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
461 end = strstr(nodestr, "=");
463 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
469 socket_id = strtoul(nodestr, &end, 0);
470 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
471 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
475 /* if we find this page in our mappings, set socket_id */
476 for (i = 0; i < hpi->num_pages[0]; i++) {
477 void *va = (void *)(unsigned long)virt_addr;
478 if (hugepg_tbl[i].orig_va == va) {
479 hugepg_tbl[i].socket_id = socket_id;
481 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
483 "Hugepage %s is on socket %d\n",
484 hugepg_tbl[i].filepath, socket_id);
490 if (hp_count < hpi->num_pages[0])
502 cmp_physaddr(const void *a, const void *b)
504 #ifndef RTE_ARCH_PPC_64
505 const struct hugepage_file *p1 = a;
506 const struct hugepage_file *p2 = b;
508 /* PowerPC needs memory sorted in reverse order from x86 */
509 const struct hugepage_file *p1 = b;
510 const struct hugepage_file *p2 = a;
512 if (p1->physaddr < p2->physaddr)
514 else if (p1->physaddr > p2->physaddr)
521 * Uses mmap to create a shared memory area for storage of data
522 * Used in this file to store the hugepage file map on disk
525 create_shared_memory(const char *filename, const size_t mem_size)
530 /* if no shared files mode is used, create anonymous memory instead */
531 if (internal_config.no_shconf) {
532 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE,
533 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
534 if (retval == MAP_FAILED)
539 fd = open(filename, O_CREAT | O_RDWR, 0666);
542 if (ftruncate(fd, mem_size) < 0) {
546 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
548 if (retval == MAP_FAILED)
554 * this copies *active* hugepages from one hugepage table to another.
555 * destination is typically the shared memory.
558 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
559 const struct hugepage_file * src, int src_size)
561 int src_pos, dst_pos = 0;
563 for (src_pos = 0; src_pos < src_size; src_pos++) {
564 if (src[src_pos].orig_va != NULL) {
565 /* error on overflow attempt */
566 if (dst_pos == dest_size)
568 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
576 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
577 unsigned num_hp_info)
579 unsigned socket, size;
580 int page, nrpages = 0;
582 /* get total number of hugepages */
583 for (size = 0; size < num_hp_info; size++)
584 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
586 internal_config.hugepage_info[size].num_pages[socket];
588 for (page = 0; page < nrpages; page++) {
589 struct hugepage_file *hp = &hugepg_tbl[page];
591 if (hp->orig_va != NULL && unlink(hp->filepath)) {
592 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
593 __func__, hp->filepath, strerror(errno));
600 * unmaps hugepages that are not going to be used. since we originally allocate
601 * ALL hugepages (not just those we need), additional unmapping needs to be done.
604 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
605 struct hugepage_info *hpi,
606 unsigned num_hp_info)
608 unsigned socket, size;
609 int page, nrpages = 0;
611 /* get total number of hugepages */
612 for (size = 0; size < num_hp_info; size++)
613 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
614 nrpages += internal_config.hugepage_info[size].num_pages[socket];
616 for (size = 0; size < num_hp_info; size++) {
617 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
618 unsigned pages_found = 0;
620 /* traverse until we have unmapped all the unused pages */
621 for (page = 0; page < nrpages; page++) {
622 struct hugepage_file *hp = &hugepg_tbl[page];
624 /* find a page that matches the criteria */
625 if ((hp->size == hpi[size].hugepage_sz) &&
626 (hp->socket_id == (int) socket)) {
628 /* if we skipped enough pages, unmap the rest */
629 if (pages_found == hpi[size].num_pages[socket]) {
632 unmap_len = hp->size;
634 /* get start addr and len of the remaining segment */
639 if (unlink(hp->filepath) == -1) {
640 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
641 __func__, hp->filepath, strerror(errno));
645 /* lock the page and skip */
651 } /* foreach socket */
652 } /* foreach pagesize */
658 remap_segment(struct hugepage_file *hugepages, int seg_start, int seg_end)
660 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
661 struct rte_memseg_list *msl;
662 struct rte_fbarray *arr;
663 int cur_page, seg_len;
664 unsigned int msl_idx;
670 page_sz = hugepages[seg_start].size;
671 socket_id = hugepages[seg_start].socket_id;
672 seg_len = seg_end - seg_start;
674 RTE_LOG(DEBUG, EAL, "Attempting to map %" PRIu64 "M on socket %i\n",
675 (seg_len * page_sz) >> 20ULL, socket_id);
677 /* find free space in memseg lists */
678 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
680 msl = &mcfg->memsegs[msl_idx];
681 arr = &msl->memseg_arr;
683 if (msl->page_sz != page_sz)
685 if (msl->socket_id != socket_id)
688 /* leave space for a hole if array is not empty */
689 empty = arr->count == 0;
690 ms_idx = rte_fbarray_find_next_n_free(arr, 0,
691 seg_len + (empty ? 0 : 1));
693 /* memseg list is full? */
697 /* leave some space between memsegs, they are not IOVA
698 * contiguous, so they shouldn't be VA contiguous either.
704 if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
705 RTE_LOG(ERR, EAL, "Could not find space for memseg. Please increase %s and/or %s in configuration.\n",
706 RTE_STR(CONFIG_RTE_MAX_MEMSEG_PER_TYPE),
707 RTE_STR(CONFIG_RTE_MAX_MEM_PER_TYPE));
711 #ifdef RTE_ARCH_PPC64
712 /* for PPC64 we go through the list backwards */
713 for (cur_page = seg_end - 1; cur_page >= seg_start;
714 cur_page--, ms_idx++) {
716 for (cur_page = seg_start; cur_page < seg_end; cur_page++, ms_idx++) {
718 struct hugepage_file *hfile = &hugepages[cur_page];
719 struct rte_memseg *ms = rte_fbarray_get(arr, ms_idx);
723 fd = open(hfile->filepath, O_RDWR);
725 RTE_LOG(ERR, EAL, "Could not open '%s': %s\n",
726 hfile->filepath, strerror(errno));
729 /* set shared lock on the file. */
730 if (flock(fd, LOCK_SH) < 0) {
731 RTE_LOG(DEBUG, EAL, "Could not lock '%s': %s\n",
732 hfile->filepath, strerror(errno));
736 memseg_len = (size_t)page_sz;
737 addr = RTE_PTR_ADD(msl->base_va, ms_idx * memseg_len);
739 /* we know this address is already mmapped by memseg list, so
740 * using MAP_FIXED here is safe
742 addr = mmap(addr, page_sz, PROT_READ | PROT_WRITE,
743 MAP_SHARED | MAP_POPULATE | MAP_FIXED, fd, 0);
744 if (addr == MAP_FAILED) {
745 RTE_LOG(ERR, EAL, "Couldn't remap '%s': %s\n",
746 hfile->filepath, strerror(errno));
751 /* we have a new address, so unmap previous one */
753 /* in 32-bit legacy mode, we have already unmapped the page */
754 if (!internal_config.legacy_mem)
755 munmap(hfile->orig_va, page_sz);
757 munmap(hfile->orig_va, page_sz);
760 hfile->orig_va = NULL;
761 hfile->final_va = addr;
763 /* rewrite physical addresses in IOVA as VA mode */
764 if (rte_eal_iova_mode() == RTE_IOVA_VA)
765 hfile->physaddr = (uintptr_t)addr;
767 /* set up memseg data */
769 ms->hugepage_sz = page_sz;
770 ms->len = memseg_len;
771 ms->iova = hfile->physaddr;
772 ms->socket_id = hfile->socket_id;
773 ms->nchannel = rte_memory_get_nchannel();
774 ms->nrank = rte_memory_get_nrank();
776 rte_fbarray_set_used(arr, ms_idx);
778 /* store segment fd internally */
779 if (eal_memalloc_set_seg_fd(msl_idx, ms_idx, fd) < 0)
780 RTE_LOG(ERR, EAL, "Could not store segment fd: %s\n",
781 rte_strerror(rte_errno));
783 RTE_LOG(DEBUG, EAL, "Allocated %" PRIu64 "M on socket %i\n",
784 (seg_len * page_sz) >> 20, socket_id);
789 get_mem_amount(uint64_t page_sz, uint64_t max_mem)
791 uint64_t area_sz, max_pages;
793 /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */
794 max_pages = RTE_MAX_MEMSEG_PER_LIST;
795 max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem);
797 area_sz = RTE_MIN(page_sz * max_pages, max_mem);
799 /* make sure the list isn't smaller than the page size */
800 area_sz = RTE_MAX(area_sz, page_sz);
802 return RTE_ALIGN(area_sz, page_sz);
806 free_memseg_list(struct rte_memseg_list *msl)
808 if (rte_fbarray_destroy(&msl->memseg_arr)) {
809 RTE_LOG(ERR, EAL, "Cannot destroy memseg list\n");
812 memset(msl, 0, sizeof(*msl));
816 #define MEMSEG_LIST_FMT "memseg-%" PRIu64 "k-%i-%i"
818 alloc_memseg_list(struct rte_memseg_list *msl, uint64_t page_sz,
819 int n_segs, int socket_id, int type_msl_idx)
821 char name[RTE_FBARRAY_NAME_LEN];
823 snprintf(name, sizeof(name), MEMSEG_LIST_FMT, page_sz >> 10, socket_id,
825 if (rte_fbarray_init(&msl->memseg_arr, name, n_segs,
826 sizeof(struct rte_memseg))) {
827 RTE_LOG(ERR, EAL, "Cannot allocate memseg list: %s\n",
828 rte_strerror(rte_errno));
832 msl->page_sz = page_sz;
833 msl->socket_id = socket_id;
836 RTE_LOG(DEBUG, EAL, "Memseg list allocated: 0x%zxkB at socket %i\n",
837 (size_t)page_sz >> 10, socket_id);
843 alloc_va_space(struct rte_memseg_list *msl)
850 #ifdef RTE_ARCH_PPC_64
851 flags |= MAP_HUGETLB;
854 page_sz = msl->page_sz;
855 mem_sz = page_sz * msl->memseg_arr.len;
857 addr = eal_get_virtual_area(msl->base_va, &mem_sz, page_sz, 0, flags);
859 if (rte_errno == EADDRNOTAVAIL)
860 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes at [%p] - please use '--base-virtaddr' option\n",
861 (unsigned long long)mem_sz, msl->base_va);
863 RTE_LOG(ERR, EAL, "Cannot reserve memory\n");
873 * Our VA space is not preallocated yet, so preallocate it here. We need to know
874 * how many segments there are in order to map all pages into one address space,
875 * and leave appropriate holes between segments so that rte_malloc does not
876 * concatenate them into one big segment.
878 * we also need to unmap original pages to free up address space.
880 static int __rte_unused
881 prealloc_segments(struct hugepage_file *hugepages, int n_pages)
883 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
884 int cur_page, seg_start_page, end_seg, new_memseg;
885 unsigned int hpi_idx, socket, i;
886 int n_contig_segs, n_segs;
889 /* before we preallocate segments, we need to free up our VA space.
890 * we're not removing files, and we already have information about
891 * PA-contiguousness, so it is safe to unmap everything.
893 for (cur_page = 0; cur_page < n_pages; cur_page++) {
894 struct hugepage_file *hpi = &hugepages[cur_page];
895 munmap(hpi->orig_va, hpi->size);
899 /* we cannot know how many page sizes and sockets we have discovered, so
900 * loop over all of them
902 for (hpi_idx = 0; hpi_idx < internal_config.num_hugepage_sizes;
905 internal_config.hugepage_info[hpi_idx].hugepage_sz;
907 for (i = 0; i < rte_socket_count(); i++) {
908 struct rte_memseg_list *msl;
910 socket = rte_socket_id_by_idx(i);
915 for (cur_page = 0; cur_page < n_pages; cur_page++) {
916 struct hugepage_file *prev, *cur;
917 int prev_seg_start_page = -1;
919 cur = &hugepages[cur_page];
920 prev = cur_page == 0 ? NULL :
921 &hugepages[cur_page - 1];
928 else if (cur->socket_id != (int) socket)
930 else if (cur->size != page_sz)
932 else if (cur_page == 0)
934 #ifdef RTE_ARCH_PPC_64
935 /* On PPC64 architecture, the mmap always start
936 * from higher address to lower address. Here,
937 * physical addresses are in descending order.
939 else if ((prev->physaddr - cur->physaddr) !=
943 else if ((cur->physaddr - prev->physaddr) !=
948 /* if we're already inside a segment,
949 * new segment means end of current one
951 if (seg_start_page != -1) {
953 prev_seg_start_page =
956 seg_start_page = cur_page;
960 if (prev_seg_start_page != -1) {
961 /* we've found a new segment */
965 } else if (seg_start_page != -1) {
966 /* we didn't find new segment,
967 * but did end current one
975 /* we're skipping this page */
979 /* segment continues */
981 /* check if we missed last segment */
982 if (seg_start_page != -1) {
984 n_segs += cur_page - seg_start_page;
987 /* if no segments were found, do not preallocate */
991 /* we now have total number of pages that we will
992 * allocate for this segment list. add separator pages
993 * to the total count, and preallocate VA space.
995 n_segs += n_contig_segs - 1;
997 /* now, preallocate VA space for these segments */
999 /* first, find suitable memseg list for this */
1000 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS;
1002 msl = &mcfg->memsegs[msl_idx];
1004 if (msl->base_va != NULL)
1008 if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
1009 RTE_LOG(ERR, EAL, "Not enough space in memseg lists, please increase %s\n",
1010 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
1014 /* now, allocate fbarray itself */
1015 if (alloc_memseg_list(msl, page_sz, n_segs, socket,
1019 /* finally, allocate VA space */
1020 if (alloc_va_space(msl) < 0)
1028 * We cannot reallocate memseg lists on the fly because PPC64 stores pages
1029 * backwards, therefore we have to process the entire memseg first before
1030 * remapping it into memseg list VA space.
1033 remap_needed_hugepages(struct hugepage_file *hugepages, int n_pages)
1035 int cur_page, seg_start_page, new_memseg, ret;
1038 for (cur_page = 0; cur_page < n_pages; cur_page++) {
1039 struct hugepage_file *prev, *cur;
1043 cur = &hugepages[cur_page];
1044 prev = cur_page == 0 ? NULL : &hugepages[cur_page - 1];
1046 /* if size is zero, no more pages left */
1052 else if (cur->socket_id != prev->socket_id)
1054 else if (cur->size != prev->size)
1056 #ifdef RTE_ARCH_PPC_64
1057 /* On PPC64 architecture, the mmap always start from higher
1058 * address to lower address. Here, physical addresses are in
1061 else if ((prev->physaddr - cur->physaddr) != cur->size)
1064 else if ((cur->physaddr - prev->physaddr) != cur->size)
1069 /* if this isn't the first time, remap segment */
1070 if (cur_page != 0) {
1071 ret = remap_segment(hugepages, seg_start_page,
1076 /* remember where we started */
1077 seg_start_page = cur_page;
1079 /* continuation of previous memseg */
1081 /* we were stopped, but we didn't remap the last segment, do it now */
1082 if (cur_page != 0) {
1083 ret = remap_segment(hugepages, seg_start_page,
1091 static inline uint64_t
1092 get_socket_mem_size(int socket)
1097 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
1098 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1099 size += hpi->hugepage_sz * hpi->num_pages[socket];
1106 * This function is a NUMA-aware equivalent of calc_num_pages.
1107 * It takes in the list of hugepage sizes and the
1108 * number of pages thereof, and calculates the best number of
1109 * pages of each size to fulfill the request for <memory> ram
1112 calc_num_pages_per_socket(uint64_t * memory,
1113 struct hugepage_info *hp_info,
1114 struct hugepage_info *hp_used,
1115 unsigned num_hp_info)
1117 unsigned socket, j, i = 0;
1118 unsigned requested, available;
1119 int total_num_pages = 0;
1120 uint64_t remaining_mem, cur_mem;
1121 uint64_t total_mem = internal_config.memory;
1123 if (num_hp_info == 0)
1126 /* if specific memory amounts per socket weren't requested */
1127 if (internal_config.force_sockets == 0) {
1130 int cpu_per_socket[RTE_MAX_NUMA_NODES];
1131 size_t default_size;
1134 /* Compute number of cores per socket */
1135 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
1136 RTE_LCORE_FOREACH(lcore_id) {
1137 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
1141 * Automatically spread requested memory amongst detected sockets according
1142 * to number of cores from cpu mask present on each socket
1144 total_size = internal_config.memory;
1145 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1147 /* Set memory amount per socket */
1148 default_size = (internal_config.memory * cpu_per_socket[socket])
1149 / rte_lcore_count();
1151 /* Limit to maximum available memory on socket */
1152 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
1155 memory[socket] = default_size;
1156 total_size -= default_size;
1160 * If some memory is remaining, try to allocate it by getting all
1161 * available memory from sockets, one after the other
1163 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1164 /* take whatever is available */
1165 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
1169 memory[socket] += default_size;
1170 total_size -= default_size;
1173 /* in 32-bit mode, allocate all of the memory only on master
1176 total_size = internal_config.memory;
1177 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0;
1179 struct rte_config *cfg = rte_eal_get_configuration();
1180 unsigned int master_lcore_socket;
1182 master_lcore_socket =
1183 rte_lcore_to_socket_id(cfg->master_lcore);
1185 if (master_lcore_socket != socket)
1189 memory[socket] = total_size;
1195 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
1196 /* skips if the memory on specific socket wasn't requested */
1197 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
1198 strlcpy(hp_used[i].hugedir, hp_info[i].hugedir,
1199 sizeof(hp_used[i].hugedir));
1200 hp_used[i].num_pages[socket] = RTE_MIN(
1201 memory[socket] / hp_info[i].hugepage_sz,
1202 hp_info[i].num_pages[socket]);
1204 cur_mem = hp_used[i].num_pages[socket] *
1205 hp_used[i].hugepage_sz;
1207 memory[socket] -= cur_mem;
1208 total_mem -= cur_mem;
1210 total_num_pages += hp_used[i].num_pages[socket];
1212 /* check if we have met all memory requests */
1213 if (memory[socket] == 0)
1216 /* check if we have any more pages left at this size, if so
1217 * move on to next size */
1218 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
1220 /* At this point we know that there are more pages available that are
1221 * bigger than the memory we want, so lets see if we can get enough
1222 * from other page sizes.
1225 for (j = i+1; j < num_hp_info; j++)
1226 remaining_mem += hp_info[j].hugepage_sz *
1227 hp_info[j].num_pages[socket];
1229 /* is there enough other memory, if not allocate another page and quit */
1230 if (remaining_mem < memory[socket]){
1231 cur_mem = RTE_MIN(memory[socket],
1232 hp_info[i].hugepage_sz);
1233 memory[socket] -= cur_mem;
1234 total_mem -= cur_mem;
1235 hp_used[i].num_pages[socket]++;
1237 break; /* we are done with this socket*/
1240 /* if we didn't satisfy all memory requirements per socket */
1241 if (memory[socket] > 0 &&
1242 internal_config.socket_mem[socket] != 0) {
1243 /* to prevent icc errors */
1244 requested = (unsigned) (internal_config.socket_mem[socket] /
1246 available = requested -
1247 ((unsigned) (memory[socket] / 0x100000));
1248 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
1249 "Requested: %uMB, available: %uMB\n", socket,
1250 requested, available);
1255 /* if we didn't satisfy total memory requirements */
1256 if (total_mem > 0) {
1257 requested = (unsigned) (internal_config.memory / 0x100000);
1258 available = requested - (unsigned) (total_mem / 0x100000);
1259 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
1260 " available: %uMB\n", requested, available);
1263 return total_num_pages;
1266 static inline size_t
1267 eal_get_hugepage_mem_size(void)
1272 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
1273 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1274 if (strnlen(hpi->hugedir, sizeof(hpi->hugedir)) != 0) {
1275 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1276 size += hpi->hugepage_sz * hpi->num_pages[j];
1281 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
1284 static struct sigaction huge_action_old;
1285 static int huge_need_recover;
1288 huge_register_sigbus(void)
1291 struct sigaction action;
1294 sigaddset(&mask, SIGBUS);
1295 action.sa_flags = 0;
1296 action.sa_mask = mask;
1297 action.sa_handler = huge_sigbus_handler;
1299 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1303 huge_recover_sigbus(void)
1305 if (huge_need_recover) {
1306 sigaction(SIGBUS, &huge_action_old, NULL);
1307 huge_need_recover = 0;
1312 * Prepare physical memory mapping: fill configuration structure with
1313 * these infos, return 0 on success.
1314 * 1. map N huge pages in separate files in hugetlbfs
1315 * 2. find associated physical addr
1316 * 3. find associated NUMA socket ID
1317 * 4. sort all huge pages by physical address
1318 * 5. remap these N huge pages in the correct order
1319 * 6. unmap the first mapping
1320 * 7. fill memsegs in configuration with contiguous zones
1323 eal_legacy_hugepage_init(void)
1325 struct rte_mem_config *mcfg;
1326 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1327 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1328 struct rte_fbarray *arr;
1329 struct rte_memseg *ms;
1331 uint64_t memory[RTE_MAX_NUMA_NODES];
1335 int nr_hugefiles, nr_hugepages = 0;
1338 test_phys_addrs_available();
1340 memset(used_hp, 0, sizeof(used_hp));
1342 /* get pointer to global configuration */
1343 mcfg = rte_eal_get_configuration()->mem_config;
1345 /* hugetlbfs can be disabled */
1346 if (internal_config.no_hugetlbfs) {
1347 struct rte_memseg_list *msl;
1349 int n_segs, cur_seg;
1351 /* nohuge mode is legacy mode */
1352 internal_config.legacy_mem = 1;
1354 /* create a memseg list */
1355 msl = &mcfg->memsegs[0];
1357 page_sz = RTE_PGSIZE_4K;
1358 n_segs = internal_config.memory / page_sz;
1360 if (rte_fbarray_init(&msl->memseg_arr, "nohugemem", n_segs,
1361 sizeof(struct rte_memseg))) {
1362 RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
1366 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1367 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
1368 if (addr == MAP_FAILED) {
1369 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1373 msl->base_va = addr;
1374 msl->page_sz = page_sz;
1376 msl->len = internal_config.memory;
1378 /* populate memsegs. each memseg is one page long */
1379 for (cur_seg = 0; cur_seg < n_segs; cur_seg++) {
1380 arr = &msl->memseg_arr;
1382 ms = rte_fbarray_get(arr, cur_seg);
1383 if (rte_eal_iova_mode() == RTE_IOVA_VA)
1384 ms->iova = (uintptr_t)addr;
1386 ms->iova = RTE_BAD_IOVA;
1388 ms->hugepage_sz = page_sz;
1392 rte_fbarray_set_used(arr, cur_seg);
1394 addr = RTE_PTR_ADD(addr, (size_t)page_sz);
1399 /* calculate total number of hugepages available. at this point we haven't
1400 * yet started sorting them so they all are on socket 0 */
1401 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1402 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1403 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1405 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1409 * allocate a memory area for hugepage table.
1410 * this isn't shared memory yet. due to the fact that we need some
1411 * processing done on these pages, shared memory will be created
1414 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1418 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1420 hp_offset = 0; /* where we start the current page size entries */
1422 huge_register_sigbus();
1424 /* make a copy of socket_mem, needed for balanced allocation. */
1425 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1426 memory[i] = internal_config.socket_mem[i];
1428 /* map all hugepages and sort them */
1429 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1430 unsigned pages_old, pages_new;
1431 struct hugepage_info *hpi;
1434 * we don't yet mark hugepages as used at this stage, so
1435 * we just map all hugepages available to the system
1436 * all hugepages are still located on socket 0
1438 hpi = &internal_config.hugepage_info[i];
1440 if (hpi->num_pages[0] == 0)
1443 /* map all hugepages available */
1444 pages_old = hpi->num_pages[0];
1445 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, memory);
1446 if (pages_new < pages_old) {
1448 "%d not %d hugepages of size %u MB allocated\n",
1449 pages_new, pages_old,
1450 (unsigned)(hpi->hugepage_sz / 0x100000));
1452 int pages = pages_old - pages_new;
1454 nr_hugepages -= pages;
1455 hpi->num_pages[0] = pages_new;
1460 if (phys_addrs_available &&
1461 rte_eal_iova_mode() != RTE_IOVA_VA) {
1462 /* find physical addresses for each hugepage */
1463 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1464 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1465 "for %u MB pages\n",
1466 (unsigned int)(hpi->hugepage_sz / 0x100000));
1470 /* set physical addresses for each hugepage */
1471 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1472 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1473 "for %u MB pages\n",
1474 (unsigned int)(hpi->hugepage_sz / 0x100000));
1479 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1480 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1481 (unsigned)(hpi->hugepage_sz / 0x100000));
1485 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1486 sizeof(struct hugepage_file), cmp_physaddr);
1488 /* we have processed a num of hugepages of this size, so inc offset */
1489 hp_offset += hpi->num_pages[0];
1492 huge_recover_sigbus();
1494 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1495 internal_config.memory = eal_get_hugepage_mem_size();
1497 nr_hugefiles = nr_hugepages;
1500 /* clean out the numbers of pages */
1501 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1502 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1503 internal_config.hugepage_info[i].num_pages[j] = 0;
1505 /* get hugepages for each socket */
1506 for (i = 0; i < nr_hugefiles; i++) {
1507 int socket = tmp_hp[i].socket_id;
1509 /* find a hugepage info with right size and increment num_pages */
1510 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1511 (int)internal_config.num_hugepage_sizes);
1512 for (j = 0; j < nb_hpsizes; j++) {
1513 if (tmp_hp[i].size ==
1514 internal_config.hugepage_info[j].hugepage_sz) {
1515 internal_config.hugepage_info[j].num_pages[socket]++;
1520 /* make a copy of socket_mem, needed for number of pages calculation */
1521 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1522 memory[i] = internal_config.socket_mem[i];
1524 /* calculate final number of pages */
1525 nr_hugepages = calc_num_pages_per_socket(memory,
1526 internal_config.hugepage_info, used_hp,
1527 internal_config.num_hugepage_sizes);
1529 /* error if not enough memory available */
1530 if (nr_hugepages < 0)
1534 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1535 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1536 if (used_hp[i].num_pages[j] > 0) {
1538 "Requesting %u pages of size %uMB"
1539 " from socket %i\n",
1540 used_hp[i].num_pages[j],
1542 (used_hp[i].hugepage_sz / 0x100000),
1548 /* create shared memory */
1549 hugepage = create_shared_memory(eal_hugepage_data_path(),
1550 nr_hugefiles * sizeof(struct hugepage_file));
1552 if (hugepage == NULL) {
1553 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1556 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1559 * unmap pages that we won't need (looks at used_hp).
1560 * also, sets final_va to NULL on pages that were unmapped.
1562 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1563 internal_config.num_hugepage_sizes) < 0) {
1564 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1569 * copy stuff from malloc'd hugepage* to the actual shared memory.
1570 * this procedure only copies those hugepages that have orig_va
1571 * not NULL. has overflow protection.
1573 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1574 tmp_hp, nr_hugefiles) < 0) {
1575 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1580 /* for legacy 32-bit mode, we did not preallocate VA space, so do it */
1581 if (internal_config.legacy_mem &&
1582 prealloc_segments(hugepage, nr_hugefiles)) {
1583 RTE_LOG(ERR, EAL, "Could not preallocate VA space for hugepages\n");
1588 /* remap all pages we do need into memseg list VA space, so that those
1589 * pages become first-class citizens in DPDK memory subsystem
1591 if (remap_needed_hugepages(hugepage, nr_hugefiles)) {
1592 RTE_LOG(ERR, EAL, "Couldn't remap hugepage files into memseg lists\n");
1596 /* free the hugepage backing files */
1597 if (internal_config.hugepage_unlink &&
1598 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1599 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1603 /* free the temporary hugepage table */
1607 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1609 /* we're not going to allocate more pages, so release VA space for
1610 * unused memseg lists
1612 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1613 struct rte_memseg_list *msl = &mcfg->memsegs[i];
1616 /* skip inactive lists */
1617 if (msl->base_va == NULL)
1619 /* skip lists where there is at least one page allocated */
1620 if (msl->memseg_arr.count > 0)
1622 /* this is an unused list, deallocate it */
1624 munmap(msl->base_va, mem_sz);
1625 msl->base_va = NULL;
1627 /* destroy backing fbarray */
1628 rte_fbarray_destroy(&msl->memseg_arr);
1634 huge_recover_sigbus();
1636 if (hugepage != NULL)
1637 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1642 static int __rte_unused
1643 hugepage_count_walk(const struct rte_memseg_list *msl, void *arg)
1645 struct hugepage_info *hpi = arg;
1647 if (msl->page_sz != hpi->hugepage_sz)
1650 hpi->num_pages[msl->socket_id] += msl->memseg_arr.len;
1655 limits_callback(int socket_id, size_t cur_limit, size_t new_len)
1657 RTE_SET_USED(socket_id);
1658 RTE_SET_USED(cur_limit);
1659 RTE_SET_USED(new_len);
1664 eal_hugepage_init(void)
1666 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1667 uint64_t memory[RTE_MAX_NUMA_NODES];
1668 int hp_sz_idx, socket_id;
1670 test_phys_addrs_available();
1672 memset(used_hp, 0, sizeof(used_hp));
1675 hp_sz_idx < (int) internal_config.num_hugepage_sizes;
1678 struct hugepage_info dummy;
1681 /* also initialize used_hp hugepage sizes in used_hp */
1682 struct hugepage_info *hpi;
1683 hpi = &internal_config.hugepage_info[hp_sz_idx];
1684 used_hp[hp_sz_idx].hugepage_sz = hpi->hugepage_sz;
1687 /* for 32-bit, limit number of pages on socket to whatever we've
1688 * preallocated, as we cannot allocate more.
1690 memset(&dummy, 0, sizeof(dummy));
1691 dummy.hugepage_sz = hpi->hugepage_sz;
1692 if (rte_memseg_list_walk(hugepage_count_walk, &dummy) < 0)
1695 for (i = 0; i < RTE_DIM(dummy.num_pages); i++) {
1696 hpi->num_pages[i] = RTE_MIN(hpi->num_pages[i],
1697 dummy.num_pages[i]);
1702 /* make a copy of socket_mem, needed for balanced allocation. */
1703 for (hp_sz_idx = 0; hp_sz_idx < RTE_MAX_NUMA_NODES; hp_sz_idx++)
1704 memory[hp_sz_idx] = internal_config.socket_mem[hp_sz_idx];
1706 /* calculate final number of pages */
1707 if (calc_num_pages_per_socket(memory,
1708 internal_config.hugepage_info, used_hp,
1709 internal_config.num_hugepage_sizes) < 0)
1713 hp_sz_idx < (int)internal_config.num_hugepage_sizes;
1715 for (socket_id = 0; socket_id < RTE_MAX_NUMA_NODES;
1717 struct rte_memseg **pages;
1718 struct hugepage_info *hpi = &used_hp[hp_sz_idx];
1719 unsigned int num_pages = hpi->num_pages[socket_id];
1720 int num_pages_alloc, i;
1725 pages = malloc(sizeof(*pages) * num_pages);
1727 RTE_LOG(DEBUG, EAL, "Allocating %u pages of size %" PRIu64 "M on socket %i\n",
1728 num_pages, hpi->hugepage_sz >> 20, socket_id);
1730 num_pages_alloc = eal_memalloc_alloc_seg_bulk(pages,
1731 num_pages, hpi->hugepage_sz,
1733 if (num_pages_alloc < 0) {
1738 /* mark preallocated pages as unfreeable */
1739 for (i = 0; i < num_pages_alloc; i++) {
1740 struct rte_memseg *ms = pages[i];
1741 ms->flags |= RTE_MEMSEG_FLAG_DO_NOT_FREE;
1746 /* if socket limits were specified, set them */
1747 if (internal_config.force_socket_limits) {
1749 for (i = 0; i < RTE_MAX_NUMA_NODES; i++) {
1750 uint64_t limit = internal_config.socket_limit[i];
1753 if (rte_mem_alloc_validator_register("socket-limit",
1754 limits_callback, i, limit))
1755 RTE_LOG(ERR, EAL, "Failed to register socket limits validator callback\n");
1762 * uses fstat to report the size of a file on disk
1768 if (fstat(fd, &st) < 0)
1774 * This creates the memory mappings in the secondary process to match that of
1775 * the server process. It goes through each memory segment in the DPDK runtime
1776 * configuration and finds the hugepages which form that segment, mapping them
1777 * in order to form a contiguous block in the virtual memory space
1780 eal_legacy_hugepage_attach(void)
1782 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1783 struct hugepage_file *hp = NULL;
1784 unsigned int num_hp = 0;
1786 unsigned int cur_seg;
1788 int fd, fd_hugepage = -1;
1790 if (aslr_enabled() > 0) {
1791 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1792 "(ASLR) is enabled in the kernel.\n");
1793 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1794 "into secondary processes\n");
1797 test_phys_addrs_available();
1799 fd_hugepage = open(eal_hugepage_data_path(), O_RDONLY);
1800 if (fd_hugepage < 0) {
1801 RTE_LOG(ERR, EAL, "Could not open %s\n",
1802 eal_hugepage_data_path());
1806 size = getFileSize(fd_hugepage);
1807 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1808 if (hp == MAP_FAILED) {
1809 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1810 eal_hugepage_data_path());
1814 num_hp = size / sizeof(struct hugepage_file);
1815 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1817 /* map all segments into memory to make sure we get the addrs. the
1818 * segments themselves are already in memseg list (which is shared and
1819 * has its VA space already preallocated), so we just need to map
1820 * everything into correct addresses.
1822 for (i = 0; i < num_hp; i++) {
1823 struct hugepage_file *hf = &hp[i];
1824 size_t map_sz = hf->size;
1825 void *map_addr = hf->final_va;
1826 int msl_idx, ms_idx;
1827 struct rte_memseg_list *msl;
1828 struct rte_memseg *ms;
1830 /* if size is zero, no more pages left */
1834 fd = open(hf->filepath, O_RDWR);
1836 RTE_LOG(ERR, EAL, "Could not open %s: %s\n",
1837 hf->filepath, strerror(errno));
1841 map_addr = mmap(map_addr, map_sz, PROT_READ | PROT_WRITE,
1842 MAP_SHARED | MAP_FIXED, fd, 0);
1843 if (map_addr == MAP_FAILED) {
1844 RTE_LOG(ERR, EAL, "Could not map %s: %s\n",
1845 hf->filepath, strerror(errno));
1849 /* set shared lock on the file. */
1850 if (flock(fd, LOCK_SH) < 0) {
1851 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed: %s\n",
1852 __func__, strerror(errno));
1856 /* find segment data */
1857 msl = rte_mem_virt2memseg_list(map_addr);
1859 RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg list\n",
1863 ms = rte_mem_virt2memseg(map_addr, msl);
1865 RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg\n",
1870 msl_idx = msl - mcfg->memsegs;
1871 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
1873 RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg idx\n",
1878 /* store segment fd internally */
1879 if (eal_memalloc_set_seg_fd(msl_idx, ms_idx, fd) < 0)
1880 RTE_LOG(ERR, EAL, "Could not store segment fd: %s\n",
1881 rte_strerror(rte_errno));
1883 /* unmap the hugepage config file, since we are done using it */
1891 /* map all segments into memory to make sure we get the addrs */
1893 for (cur_seg = 0; cur_seg < i; cur_seg++) {
1894 struct hugepage_file *hf = &hp[i];
1895 size_t map_sz = hf->size;
1896 void *map_addr = hf->final_va;
1898 munmap(map_addr, map_sz);
1900 if (hp != NULL && hp != MAP_FAILED)
1902 if (fd_hugepage >= 0)
1908 eal_hugepage_attach(void)
1910 if (eal_memalloc_sync_with_primary()) {
1911 RTE_LOG(ERR, EAL, "Could not map memory from primary process\n");
1912 if (aslr_enabled() > 0)
1913 RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel and retry running both primary and secondary processes\n");
1920 rte_eal_hugepage_init(void)
1922 return internal_config.legacy_mem ?
1923 eal_legacy_hugepage_init() :
1924 eal_hugepage_init();
1928 rte_eal_hugepage_attach(void)
1930 return internal_config.legacy_mem ?
1931 eal_legacy_hugepage_attach() :
1932 eal_hugepage_attach();
1936 rte_eal_using_phys_addrs(void)
1938 return phys_addrs_available;
1941 static int __rte_unused
1942 memseg_primary_init_32(void)
1944 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1945 int active_sockets, hpi_idx, msl_idx = 0;
1946 unsigned int socket_id, i;
1947 struct rte_memseg_list *msl;
1948 uint64_t extra_mem_per_socket, total_extra_mem, total_requested_mem;
1951 /* no-huge does not need this at all */
1952 if (internal_config.no_hugetlbfs)
1955 /* this is a giant hack, but desperate times call for desperate
1956 * measures. in legacy 32-bit mode, we cannot preallocate VA space,
1957 * because having upwards of 2 gigabytes of VA space already mapped will
1958 * interfere with our ability to map and sort hugepages.
1960 * therefore, in legacy 32-bit mode, we will be initializing memseg
1961 * lists much later - in eal_memory.c, right after we unmap all the
1962 * unneeded pages. this will not affect secondary processes, as those
1963 * should be able to mmap the space without (too many) problems.
1965 if (internal_config.legacy_mem)
1968 /* 32-bit mode is a very special case. we cannot know in advance where
1969 * the user will want to allocate their memory, so we have to do some
1973 total_requested_mem = 0;
1974 if (internal_config.force_sockets)
1975 for (i = 0; i < rte_socket_count(); i++) {
1978 socket_id = rte_socket_id_by_idx(i);
1979 mem = internal_config.socket_mem[socket_id];
1985 total_requested_mem += mem;
1988 total_requested_mem = internal_config.memory;
1990 max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
1991 if (total_requested_mem > max_mem) {
1992 RTE_LOG(ERR, EAL, "Invalid parameters: 32-bit process can at most use %uM of memory\n",
1993 (unsigned int)(max_mem >> 20));
1996 total_extra_mem = max_mem - total_requested_mem;
1997 extra_mem_per_socket = active_sockets == 0 ? total_extra_mem :
1998 total_extra_mem / active_sockets;
2000 /* the allocation logic is a little bit convoluted, but here's how it
2001 * works, in a nutshell:
2002 * - if user hasn't specified on which sockets to allocate memory via
2003 * --socket-mem, we allocate all of our memory on master core socket.
2004 * - if user has specified sockets to allocate memory on, there may be
2005 * some "unused" memory left (e.g. if user has specified --socket-mem
2006 * such that not all memory adds up to 2 gigabytes), so add it to all
2007 * sockets that are in use equally.
2009 * page sizes are sorted by size in descending order, so we can safely
2010 * assume that we dispense with bigger page sizes first.
2013 /* create memseg lists */
2014 for (i = 0; i < rte_socket_count(); i++) {
2015 int hp_sizes = (int) internal_config.num_hugepage_sizes;
2016 uint64_t max_socket_mem, cur_socket_mem;
2017 unsigned int master_lcore_socket;
2018 struct rte_config *cfg = rte_eal_get_configuration();
2021 socket_id = rte_socket_id_by_idx(i);
2023 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
2028 /* if we didn't specifically request memory on this socket */
2029 skip = active_sockets != 0 &&
2030 internal_config.socket_mem[socket_id] == 0;
2031 /* ...or if we didn't specifically request memory on *any*
2032 * socket, and this is not master lcore
2034 master_lcore_socket = rte_lcore_to_socket_id(cfg->master_lcore);
2035 skip |= active_sockets == 0 && socket_id != master_lcore_socket;
2038 RTE_LOG(DEBUG, EAL, "Will not preallocate memory on socket %u\n",
2043 /* max amount of memory on this socket */
2044 max_socket_mem = (active_sockets != 0 ?
2045 internal_config.socket_mem[socket_id] :
2046 internal_config.memory) +
2047 extra_mem_per_socket;
2050 for (hpi_idx = 0; hpi_idx < hp_sizes; hpi_idx++) {
2051 uint64_t max_pagesz_mem, cur_pagesz_mem = 0;
2052 uint64_t hugepage_sz;
2053 struct hugepage_info *hpi;
2054 int type_msl_idx, max_segs, total_segs = 0;
2056 hpi = &internal_config.hugepage_info[hpi_idx];
2057 hugepage_sz = hpi->hugepage_sz;
2059 /* check if pages are actually available */
2060 if (hpi->num_pages[socket_id] == 0)
2063 max_segs = RTE_MAX_MEMSEG_PER_TYPE;
2064 max_pagesz_mem = max_socket_mem - cur_socket_mem;
2066 /* make it multiple of page size */
2067 max_pagesz_mem = RTE_ALIGN_FLOOR(max_pagesz_mem,
2070 RTE_LOG(DEBUG, EAL, "Attempting to preallocate "
2071 "%" PRIu64 "M on socket %i\n",
2072 max_pagesz_mem >> 20, socket_id);
2075 while (cur_pagesz_mem < max_pagesz_mem &&
2076 total_segs < max_segs) {
2078 unsigned int n_segs;
2080 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2082 "No more space in memseg lists, please increase %s\n",
2083 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2087 msl = &mcfg->memsegs[msl_idx];
2089 cur_mem = get_mem_amount(hugepage_sz,
2091 n_segs = cur_mem / hugepage_sz;
2093 if (alloc_memseg_list(msl, hugepage_sz, n_segs,
2094 socket_id, type_msl_idx)) {
2095 /* failing to allocate a memseg list is
2098 RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
2102 if (alloc_va_space(msl)) {
2103 /* if we couldn't allocate VA space, we
2104 * can try with smaller page sizes.
2106 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list, retrying with different page size\n");
2107 /* deallocate memseg list */
2108 if (free_memseg_list(msl))
2113 total_segs += msl->memseg_arr.len;
2114 cur_pagesz_mem = total_segs * hugepage_sz;
2118 cur_socket_mem += cur_pagesz_mem;
2120 if (cur_socket_mem == 0) {
2121 RTE_LOG(ERR, EAL, "Cannot allocate VA space on socket %u\n",
2130 static int __rte_unused
2131 memseg_primary_init(void)
2133 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2138 int i, hpi_idx, msl_idx, ret = -1; /* fail unless told to succeed */
2139 struct rte_memseg_list *msl;
2140 uint64_t max_mem, max_mem_per_type;
2141 unsigned int max_seglists_per_type;
2142 unsigned int n_memtypes, cur_type;
2144 /* no-huge does not need this at all */
2145 if (internal_config.no_hugetlbfs)
2149 * figuring out amount of memory we're going to have is a long and very
2150 * involved process. the basic element we're operating with is a memory
2151 * type, defined as a combination of NUMA node ID and page size (so that
2152 * e.g. 2 sockets with 2 page sizes yield 4 memory types in total).
2154 * deciding amount of memory going towards each memory type is a
2155 * balancing act between maximum segments per type, maximum memory per
2156 * type, and number of detected NUMA nodes. the goal is to make sure
2157 * each memory type gets at least one memseg list.
2159 * the total amount of memory is limited by RTE_MAX_MEM_MB value.
2161 * the total amount of memory per type is limited by either
2162 * RTE_MAX_MEM_MB_PER_TYPE, or by RTE_MAX_MEM_MB divided by the number
2163 * of detected NUMA nodes. additionally, maximum number of segments per
2164 * type is also limited by RTE_MAX_MEMSEG_PER_TYPE. this is because for
2165 * smaller page sizes, it can take hundreds of thousands of segments to
2166 * reach the above specified per-type memory limits.
2168 * additionally, each type may have multiple memseg lists associated
2169 * with it, each limited by either RTE_MAX_MEM_MB_PER_LIST for bigger
2170 * page sizes, or RTE_MAX_MEMSEG_PER_LIST segments for smaller ones.
2172 * the number of memseg lists per type is decided based on the above
2173 * limits, and also taking number of detected NUMA nodes, to make sure
2174 * that we don't run out of memseg lists before we populate all NUMA
2175 * nodes with memory.
2177 * we do this in three stages. first, we collect the number of types.
2178 * then, we figure out memory constraints and populate the list of
2179 * would-be memseg lists. then, we go ahead and allocate the memseg
2183 /* create space for mem types */
2184 n_memtypes = internal_config.num_hugepage_sizes * rte_socket_count();
2185 memtypes = calloc(n_memtypes, sizeof(*memtypes));
2186 if (memtypes == NULL) {
2187 RTE_LOG(ERR, EAL, "Cannot allocate space for memory types\n");
2191 /* populate mem types */
2193 for (hpi_idx = 0; hpi_idx < (int) internal_config.num_hugepage_sizes;
2195 struct hugepage_info *hpi;
2196 uint64_t hugepage_sz;
2198 hpi = &internal_config.hugepage_info[hpi_idx];
2199 hugepage_sz = hpi->hugepage_sz;
2201 for (i = 0; i < (int) rte_socket_count(); i++, cur_type++) {
2202 int socket_id = rte_socket_id_by_idx(i);
2204 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
2208 memtypes[cur_type].page_sz = hugepage_sz;
2209 memtypes[cur_type].socket_id = socket_id;
2211 RTE_LOG(DEBUG, EAL, "Detected memory type: "
2212 "socket_id:%u hugepage_sz:%" PRIu64 "\n",
2213 socket_id, hugepage_sz);
2217 /* set up limits for types */
2218 max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
2219 max_mem_per_type = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_TYPE << 20,
2220 max_mem / n_memtypes);
2222 * limit maximum number of segment lists per type to ensure there's
2223 * space for memseg lists for all NUMA nodes with all page sizes
2225 max_seglists_per_type = RTE_MAX_MEMSEG_LISTS / n_memtypes;
2227 if (max_seglists_per_type == 0) {
2228 RTE_LOG(ERR, EAL, "Cannot accommodate all memory types, please increase %s\n",
2229 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2233 /* go through all mem types and create segment lists */
2235 for (cur_type = 0; cur_type < n_memtypes; cur_type++) {
2236 unsigned int cur_seglist, n_seglists, n_segs;
2237 unsigned int max_segs_per_type, max_segs_per_list;
2238 struct memtype *type = &memtypes[cur_type];
2239 uint64_t max_mem_per_list, pagesz;
2242 pagesz = type->page_sz;
2243 socket_id = type->socket_id;
2246 * we need to create segment lists for this type. we must take
2247 * into account the following things:
2249 * 1. total amount of memory we can use for this memory type
2250 * 2. total amount of memory per memseg list allowed
2251 * 3. number of segments needed to fit the amount of memory
2252 * 4. number of segments allowed per type
2253 * 5. number of segments allowed per memseg list
2254 * 6. number of memseg lists we are allowed to take up
2257 /* calculate how much segments we will need in total */
2258 max_segs_per_type = max_mem_per_type / pagesz;
2259 /* limit number of segments to maximum allowed per type */
2260 max_segs_per_type = RTE_MIN(max_segs_per_type,
2261 (unsigned int)RTE_MAX_MEMSEG_PER_TYPE);
2262 /* limit number of segments to maximum allowed per list */
2263 max_segs_per_list = RTE_MIN(max_segs_per_type,
2264 (unsigned int)RTE_MAX_MEMSEG_PER_LIST);
2266 /* calculate how much memory we can have per segment list */
2267 max_mem_per_list = RTE_MIN(max_segs_per_list * pagesz,
2268 (uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20);
2270 /* calculate how many segments each segment list will have */
2271 n_segs = RTE_MIN(max_segs_per_list, max_mem_per_list / pagesz);
2273 /* calculate how many segment lists we can have */
2274 n_seglists = RTE_MIN(max_segs_per_type / n_segs,
2275 max_mem_per_type / max_mem_per_list);
2277 /* limit number of segment lists according to our maximum */
2278 n_seglists = RTE_MIN(n_seglists, max_seglists_per_type);
2280 RTE_LOG(DEBUG, EAL, "Creating %i segment lists: "
2281 "n_segs:%i socket_id:%i hugepage_sz:%" PRIu64 "\n",
2282 n_seglists, n_segs, socket_id, pagesz);
2284 /* create all segment lists */
2285 for (cur_seglist = 0; cur_seglist < n_seglists; cur_seglist++) {
2286 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2288 "No more space in memseg lists, please increase %s\n",
2289 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2292 msl = &mcfg->memsegs[msl_idx++];
2294 if (alloc_memseg_list(msl, pagesz, n_segs,
2295 socket_id, cur_seglist))
2298 if (alloc_va_space(msl)) {
2299 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list\n");
2304 /* we're successful */
2312 memseg_secondary_init(void)
2314 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2316 struct rte_memseg_list *msl;
2318 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
2320 msl = &mcfg->memsegs[msl_idx];
2322 /* skip empty memseg lists */
2323 if (msl->memseg_arr.len == 0)
2326 if (rte_fbarray_attach(&msl->memseg_arr)) {
2327 RTE_LOG(ERR, EAL, "Cannot attach to primary process memseg lists\n");
2331 /* preallocate VA space */
2332 if (alloc_va_space(msl)) {
2333 RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
2342 rte_eal_memseg_init(void)
2344 /* increase rlimit to maximum */
2347 if (getrlimit(RLIMIT_NOFILE, &lim) == 0) {
2348 /* set limit to maximum */
2349 lim.rlim_cur = lim.rlim_max;
2351 if (setrlimit(RLIMIT_NOFILE, &lim) < 0) {
2352 RTE_LOG(DEBUG, EAL, "Setting maximum number of open files failed: %s\n",
2355 RTE_LOG(DEBUG, EAL, "Setting maximum number of open files to %"
2357 (uint64_t)lim.rlim_cur);
2360 RTE_LOG(ERR, EAL, "Cannot get current resource limits\n");
2363 return rte_eal_process_type() == RTE_PROC_PRIMARY ?
2365 memseg_primary_init_32() :
2367 memseg_primary_init() :
2369 memseg_secondary_init();