1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation.
3 * Copyright(c) 2013 6WIND S.A.
16 #include <sys/types.h>
18 #include <sys/queue.h>
20 #include <sys/resource.h>
23 #include <sys/ioctl.h>
27 #ifdef F_ADD_SEALS /* if file sealing is supported, so is memfd */
28 #include <linux/memfd.h>
29 #define MEMFD_SUPPORTED
31 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
36 #include <rte_errno.h>
38 #include <rte_memory.h>
39 #include <rte_launch.h>
41 #include <rte_per_lcore.h>
42 #include <rte_lcore.h>
43 #include <rte_common.h>
44 #include <rte_string_fns.h>
46 #include "eal_private.h"
47 #include "eal_memalloc.h"
48 #include "eal_memcfg.h"
49 #include "eal_internal_cfg.h"
50 #include "eal_filesystem.h"
51 #include "eal_hugepages.h"
52 #include "eal_options.h"
54 #define PFN_MASK_SIZE 8
58 * Huge page mapping under linux
60 * To reserve a big contiguous amount of memory, we use the hugepage
61 * feature of linux. For that, we need to have hugetlbfs mounted. This
62 * code will create many files in this directory (one per page) and
63 * map them in virtual memory. For each page, we will retrieve its
64 * physical address and remap it in order to have a virtual contiguous
65 * zone as well as a physical contiguous zone.
68 static int phys_addrs_available = -1;
70 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
72 uint64_t eal_get_baseaddr(void)
75 * Linux kernel uses a really high address as starting address for
76 * serving mmaps calls. If there exists addressing limitations and IOVA
77 * mode is VA, this starting address is likely too high for those
78 * devices. However, it is possible to use a lower address in the
79 * process virtual address space as with 64 bits there is a lot of
82 * Current known limitations are 39 or 40 bits. Setting the starting
83 * address at 4GB implies there are 508GB or 1020GB for mapping the
84 * available hugepages. This is likely enough for most systems, although
85 * a device with addressing limitations should call
86 * rte_mem_check_dma_mask for ensuring all memory is within supported
89 return 0x100000000ULL;
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 if (phys_addrs_available == 0)
107 /* standard page size */
108 page_size = getpagesize();
110 fd = open("/proc/self/pagemap", O_RDONLY);
112 RTE_LOG(INFO, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
113 __func__, strerror(errno));
117 virt_pfn = (unsigned long)virtaddr / page_size;
118 offset = sizeof(uint64_t) * virt_pfn;
119 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
120 RTE_LOG(INFO, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
121 __func__, strerror(errno));
126 retval = read(fd, &page, PFN_MASK_SIZE);
129 RTE_LOG(INFO, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
130 __func__, strerror(errno));
132 } else if (retval != PFN_MASK_SIZE) {
133 RTE_LOG(INFO, EAL, "%s(): read %d bytes from /proc/self/pagemap "
134 "but expected %d:\n",
135 __func__, retval, PFN_MASK_SIZE);
140 * the pfn (page frame number) are bits 0-54 (see
141 * pagemap.txt in linux Documentation)
143 if ((page & 0x7fffffffffffffULL) == 0)
146 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
147 + ((unsigned long)virtaddr % page_size);
153 rte_mem_virt2iova(const void *virtaddr)
155 if (rte_eal_iova_mode() == RTE_IOVA_VA)
156 return (uintptr_t)virtaddr;
157 return rte_mem_virt2phy(virtaddr);
161 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
162 * it by browsing the /proc/self/pagemap special file.
165 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
170 for (i = 0; i < hpi->num_pages[0]; i++) {
171 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
172 if (addr == RTE_BAD_PHYS_ADDR)
174 hugepg_tbl[i].physaddr = addr;
180 * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
183 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
186 static phys_addr_t addr;
188 for (i = 0; i < hpi->num_pages[0]; i++) {
189 hugepg_tbl[i].physaddr = addr;
190 addr += hugepg_tbl[i].size;
196 * Check whether address-space layout randomization is enabled in
197 * the kernel. This is important for multi-process as it can prevent
198 * two processes mapping data to the same virtual address
200 * 0 - address space randomization disabled
201 * 1/2 - address space randomization enabled
202 * negative error code on error
208 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
211 retval = read(fd, &c, 1);
221 default: return -EINVAL;
225 static sigjmp_buf huge_jmpenv;
227 static void huge_sigbus_handler(int signo __rte_unused)
229 siglongjmp(huge_jmpenv, 1);
232 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
233 * non-static local variable in the stack frame calling sigsetjmp might be
234 * clobbered by a call to longjmp.
236 static int huge_wrap_sigsetjmp(void)
238 return sigsetjmp(huge_jmpenv, 1);
241 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
242 /* Callback for numa library. */
243 void numa_error(char *where)
245 RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
250 * Mmap all hugepages of hugepage table: it first open a file in
251 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
252 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
253 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
254 * map contiguous physical blocks in contiguous virtual blocks.
257 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
258 uint64_t *essential_memory __rte_unused)
263 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
265 int essential_prev = 0;
267 struct bitmask *oldmask = NULL;
268 bool have_numa = true;
269 unsigned long maxnode = 0;
271 /* Check if kernel supports NUMA. */
272 if (numa_available() != 0) {
273 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
278 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
279 oldmask = numa_allocate_nodemask();
280 if (get_mempolicy(&oldpolicy, oldmask->maskp,
281 oldmask->size + 1, 0, 0) < 0) {
283 "Failed to get current mempolicy: %s. "
284 "Assuming MPOL_DEFAULT.\n", strerror(errno));
285 oldpolicy = MPOL_DEFAULT;
287 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
288 if (internal_config.socket_mem[i])
293 for (i = 0; i < hpi->num_pages[0]; i++) {
294 struct hugepage_file *hf = &hugepg_tbl[i];
295 uint64_t hugepage_sz = hpi->hugepage_sz;
297 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
301 for (j = 0; j < maxnode; j++)
302 if (essential_memory[j])
306 node_id = (node_id + 1) % maxnode;
307 while (!internal_config.socket_mem[node_id]) {
314 essential_prev = essential_memory[j];
316 if (essential_memory[j] < hugepage_sz)
317 essential_memory[j] = 0;
319 essential_memory[j] -= hugepage_sz;
323 "Setting policy MPOL_PREFERRED for socket %d\n",
325 numa_set_preferred(node_id);
330 hf->size = hugepage_sz;
331 eal_get_hugefile_path(hf->filepath, sizeof(hf->filepath),
332 hpi->hugedir, hf->file_id);
333 hf->filepath[sizeof(hf->filepath) - 1] = '\0';
335 /* try to create hugepage file */
336 fd = open(hf->filepath, O_CREAT | O_RDWR, 0600);
338 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
343 /* map the segment, and populate page tables,
344 * the kernel fills this segment with zeros. we don't care where
345 * this gets mapped - we already have contiguous memory areas
346 * ready for us to map into.
348 virtaddr = mmap(NULL, hugepage_sz, PROT_READ | PROT_WRITE,
349 MAP_SHARED | MAP_POPULATE, fd, 0);
350 if (virtaddr == MAP_FAILED) {
351 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
357 hf->orig_va = virtaddr;
359 /* In linux, hugetlb limitations, like cgroup, are
360 * enforced at fault time instead of mmap(), even
361 * with the option of MAP_POPULATE. Kernel will send
362 * a SIGBUS signal. To avoid to be killed, save stack
363 * environment here, if SIGBUS happens, we can jump
366 if (huge_wrap_sigsetjmp()) {
367 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
368 "hugepages of size %u MB\n",
369 (unsigned int)(hugepage_sz / 0x100000));
370 munmap(virtaddr, hugepage_sz);
372 unlink(hugepg_tbl[i].filepath);
373 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
375 essential_memory[node_id] =
380 *(int *)virtaddr = 0;
382 /* set shared lock on the file. */
383 if (flock(fd, LOCK_SH) < 0) {
384 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
385 __func__, strerror(errno));
394 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
397 "Restoring previous memory policy: %d\n", oldpolicy);
398 if (oldpolicy == MPOL_DEFAULT) {
399 numa_set_localalloc();
400 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
401 oldmask->size + 1) < 0) {
402 RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
404 numa_set_localalloc();
408 numa_free_cpumask(oldmask);
414 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
418 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
422 unsigned i, hp_count = 0;
425 char hugedir_str[PATH_MAX];
428 f = fopen("/proc/self/numa_maps", "r");
430 RTE_LOG(NOTICE, EAL, "NUMA support not available"
431 " consider that all memory is in socket_id 0\n");
435 snprintf(hugedir_str, sizeof(hugedir_str),
436 "%s/%s", hpi->hugedir, eal_get_hugefile_prefix());
439 while (fgets(buf, sizeof(buf), f) != NULL) {
441 /* ignore non huge page */
442 if (strstr(buf, " huge ") == NULL &&
443 strstr(buf, hugedir_str) == NULL)
447 virt_addr = strtoull(buf, &end, 16);
448 if (virt_addr == 0 || end == buf) {
449 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
453 /* get node id (socket id) */
454 nodestr = strstr(buf, " N");
455 if (nodestr == NULL) {
456 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
460 end = strstr(nodestr, "=");
462 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
468 socket_id = strtoul(nodestr, &end, 0);
469 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
470 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
474 /* if we find this page in our mappings, set socket_id */
475 for (i = 0; i < hpi->num_pages[0]; i++) {
476 void *va = (void *)(unsigned long)virt_addr;
477 if (hugepg_tbl[i].orig_va == va) {
478 hugepg_tbl[i].socket_id = socket_id;
480 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
482 "Hugepage %s is on socket %d\n",
483 hugepg_tbl[i].filepath, socket_id);
489 if (hp_count < hpi->num_pages[0])
501 cmp_physaddr(const void *a, const void *b)
503 #ifndef RTE_ARCH_PPC_64
504 const struct hugepage_file *p1 = a;
505 const struct hugepage_file *p2 = b;
507 /* PowerPC needs memory sorted in reverse order from x86 */
508 const struct hugepage_file *p1 = b;
509 const struct hugepage_file *p2 = a;
511 if (p1->physaddr < p2->physaddr)
513 else if (p1->physaddr > p2->physaddr)
520 * Uses mmap to create a shared memory area for storage of data
521 * Used in this file to store the hugepage file map on disk
524 create_shared_memory(const char *filename, const size_t mem_size)
529 /* if no shared files mode is used, create anonymous memory instead */
530 if (internal_config.no_shconf) {
531 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE,
532 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
533 if (retval == MAP_FAILED)
538 fd = open(filename, O_CREAT | O_RDWR, 0600);
541 if (ftruncate(fd, mem_size) < 0) {
545 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
547 if (retval == MAP_FAILED)
553 * this copies *active* hugepages from one hugepage table to another.
554 * destination is typically the shared memory.
557 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
558 const struct hugepage_file * src, int src_size)
560 int src_pos, dst_pos = 0;
562 for (src_pos = 0; src_pos < src_size; src_pos++) {
563 if (src[src_pos].orig_va != NULL) {
564 /* error on overflow attempt */
565 if (dst_pos == dest_size)
567 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
575 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
576 unsigned num_hp_info)
578 unsigned socket, size;
579 int page, nrpages = 0;
581 /* get total number of hugepages */
582 for (size = 0; size < num_hp_info; size++)
583 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
585 internal_config.hugepage_info[size].num_pages[socket];
587 for (page = 0; page < nrpages; page++) {
588 struct hugepage_file *hp = &hugepg_tbl[page];
590 if (hp->orig_va != NULL && unlink(hp->filepath)) {
591 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
592 __func__, hp->filepath, strerror(errno));
599 * unmaps hugepages that are not going to be used. since we originally allocate
600 * ALL hugepages (not just those we need), additional unmapping needs to be done.
603 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
604 struct hugepage_info *hpi,
605 unsigned num_hp_info)
607 unsigned socket, size;
608 int page, nrpages = 0;
610 /* get total number of hugepages */
611 for (size = 0; size < num_hp_info; size++)
612 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
613 nrpages += internal_config.hugepage_info[size].num_pages[socket];
615 for (size = 0; size < num_hp_info; size++) {
616 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
617 unsigned pages_found = 0;
619 /* traverse until we have unmapped all the unused pages */
620 for (page = 0; page < nrpages; page++) {
621 struct hugepage_file *hp = &hugepg_tbl[page];
623 /* find a page that matches the criteria */
624 if ((hp->size == hpi[size].hugepage_sz) &&
625 (hp->socket_id == (int) socket)) {
627 /* if we skipped enough pages, unmap the rest */
628 if (pages_found == hpi[size].num_pages[socket]) {
631 unmap_len = hp->size;
633 /* get start addr and len of the remaining segment */
638 if (unlink(hp->filepath) == -1) {
639 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
640 __func__, hp->filepath, strerror(errno));
644 /* lock the page and skip */
650 } /* foreach socket */
651 } /* foreach pagesize */
657 remap_segment(struct hugepage_file *hugepages, int seg_start, int seg_end)
659 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
660 struct rte_memseg_list *msl;
661 struct rte_fbarray *arr;
662 int cur_page, seg_len;
663 unsigned int msl_idx;
669 page_sz = hugepages[seg_start].size;
670 socket_id = hugepages[seg_start].socket_id;
671 seg_len = seg_end - seg_start;
673 RTE_LOG(DEBUG, EAL, "Attempting to map %" PRIu64 "M on socket %i\n",
674 (seg_len * page_sz) >> 20ULL, socket_id);
676 /* find free space in memseg lists */
677 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
679 msl = &mcfg->memsegs[msl_idx];
680 arr = &msl->memseg_arr;
682 if (msl->page_sz != page_sz)
684 if (msl->socket_id != socket_id)
687 /* leave space for a hole if array is not empty */
688 empty = arr->count == 0;
689 ms_idx = rte_fbarray_find_next_n_free(arr, 0,
690 seg_len + (empty ? 0 : 1));
692 /* memseg list is full? */
696 /* leave some space between memsegs, they are not IOVA
697 * contiguous, so they shouldn't be VA contiguous either.
703 if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
704 RTE_LOG(ERR, EAL, "Could not find space for memseg. Please increase %s and/or %s in configuration.\n",
705 RTE_STR(CONFIG_RTE_MAX_MEMSEG_PER_TYPE),
706 RTE_STR(CONFIG_RTE_MAX_MEM_PER_TYPE));
710 #ifdef RTE_ARCH_PPC_64
711 /* for PPC64 we go through the list backwards */
712 for (cur_page = seg_end - 1; cur_page >= seg_start;
713 cur_page--, ms_idx++) {
715 for (cur_page = seg_start; cur_page < seg_end; cur_page++, ms_idx++) {
717 struct hugepage_file *hfile = &hugepages[cur_page];
718 struct rte_memseg *ms = rte_fbarray_get(arr, ms_idx);
722 fd = open(hfile->filepath, O_RDWR);
724 RTE_LOG(ERR, EAL, "Could not open '%s': %s\n",
725 hfile->filepath, strerror(errno));
728 /* set shared lock on the file. */
729 if (flock(fd, LOCK_SH) < 0) {
730 RTE_LOG(DEBUG, EAL, "Could not lock '%s': %s\n",
731 hfile->filepath, strerror(errno));
735 memseg_len = (size_t)page_sz;
736 addr = RTE_PTR_ADD(msl->base_va, ms_idx * memseg_len);
738 /* we know this address is already mmapped by memseg list, so
739 * using MAP_FIXED here is safe
741 addr = mmap(addr, page_sz, PROT_READ | PROT_WRITE,
742 MAP_SHARED | MAP_POPULATE | MAP_FIXED, fd, 0);
743 if (addr == MAP_FAILED) {
744 RTE_LOG(ERR, EAL, "Couldn't remap '%s': %s\n",
745 hfile->filepath, strerror(errno));
750 /* we have a new address, so unmap previous one */
752 /* in 32-bit legacy mode, we have already unmapped the page */
753 if (!internal_config.legacy_mem)
754 munmap(hfile->orig_va, page_sz);
756 munmap(hfile->orig_va, page_sz);
759 hfile->orig_va = NULL;
760 hfile->final_va = addr;
762 /* rewrite physical addresses in IOVA as VA mode */
763 if (rte_eal_iova_mode() == RTE_IOVA_VA)
764 hfile->physaddr = (uintptr_t)addr;
766 /* set up memseg data */
768 ms->hugepage_sz = page_sz;
769 ms->len = memseg_len;
770 ms->iova = hfile->physaddr;
771 ms->socket_id = hfile->socket_id;
772 ms->nchannel = rte_memory_get_nchannel();
773 ms->nrank = rte_memory_get_nrank();
775 rte_fbarray_set_used(arr, ms_idx);
777 /* store segment fd internally */
778 if (eal_memalloc_set_seg_fd(msl_idx, ms_idx, fd) < 0)
779 RTE_LOG(ERR, EAL, "Could not store segment fd: %s\n",
780 rte_strerror(rte_errno));
782 RTE_LOG(DEBUG, EAL, "Allocated %" PRIu64 "M on socket %i\n",
783 (seg_len * page_sz) >> 20, socket_id);
788 get_mem_amount(uint64_t page_sz, uint64_t max_mem)
790 uint64_t area_sz, max_pages;
792 /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */
793 max_pages = RTE_MAX_MEMSEG_PER_LIST;
794 max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem);
796 area_sz = RTE_MIN(page_sz * max_pages, max_mem);
798 /* make sure the list isn't smaller than the page size */
799 area_sz = RTE_MAX(area_sz, page_sz);
801 return RTE_ALIGN(area_sz, page_sz);
805 memseg_list_free(struct rte_memseg_list *msl)
807 if (rte_fbarray_destroy(&msl->memseg_arr)) {
808 RTE_LOG(ERR, EAL, "Cannot destroy memseg list\n");
811 memset(msl, 0, sizeof(*msl));
816 * Our VA space is not preallocated yet, so preallocate it here. We need to know
817 * how many segments there are in order to map all pages into one address space,
818 * and leave appropriate holes between segments so that rte_malloc does not
819 * concatenate them into one big segment.
821 * we also need to unmap original pages to free up address space.
823 static int __rte_unused
824 prealloc_segments(struct hugepage_file *hugepages, int n_pages)
826 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
827 int cur_page, seg_start_page, end_seg, new_memseg;
828 unsigned int hpi_idx, socket, i;
829 int n_contig_segs, n_segs;
832 /* before we preallocate segments, we need to free up our VA space.
833 * we're not removing files, and we already have information about
834 * PA-contiguousness, so it is safe to unmap everything.
836 for (cur_page = 0; cur_page < n_pages; cur_page++) {
837 struct hugepage_file *hpi = &hugepages[cur_page];
838 munmap(hpi->orig_va, hpi->size);
842 /* we cannot know how many page sizes and sockets we have discovered, so
843 * loop over all of them
845 for (hpi_idx = 0; hpi_idx < internal_config.num_hugepage_sizes;
848 internal_config.hugepage_info[hpi_idx].hugepage_sz;
850 for (i = 0; i < rte_socket_count(); i++) {
851 struct rte_memseg_list *msl;
853 socket = rte_socket_id_by_idx(i);
858 for (cur_page = 0; cur_page < n_pages; cur_page++) {
859 struct hugepage_file *prev, *cur;
860 int prev_seg_start_page = -1;
862 cur = &hugepages[cur_page];
863 prev = cur_page == 0 ? NULL :
864 &hugepages[cur_page - 1];
871 else if (cur->socket_id != (int) socket)
873 else if (cur->size != page_sz)
875 else if (cur_page == 0)
877 #ifdef RTE_ARCH_PPC_64
878 /* On PPC64 architecture, the mmap always start
879 * from higher address to lower address. Here,
880 * physical addresses are in descending order.
882 else if ((prev->physaddr - cur->physaddr) !=
886 else if ((cur->physaddr - prev->physaddr) !=
891 /* if we're already inside a segment,
892 * new segment means end of current one
894 if (seg_start_page != -1) {
896 prev_seg_start_page =
899 seg_start_page = cur_page;
903 if (prev_seg_start_page != -1) {
904 /* we've found a new segment */
908 } else if (seg_start_page != -1) {
909 /* we didn't find new segment,
910 * but did end current one
918 /* we're skipping this page */
922 /* segment continues */
924 /* check if we missed last segment */
925 if (seg_start_page != -1) {
927 n_segs += cur_page - seg_start_page;
930 /* if no segments were found, do not preallocate */
934 /* we now have total number of pages that we will
935 * allocate for this segment list. add separator pages
936 * to the total count, and preallocate VA space.
938 n_segs += n_contig_segs - 1;
940 /* now, preallocate VA space for these segments */
942 /* first, find suitable memseg list for this */
943 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS;
945 msl = &mcfg->memsegs[msl_idx];
947 if (msl->base_va != NULL)
951 if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
952 RTE_LOG(ERR, EAL, "Not enough space in memseg lists, please increase %s\n",
953 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
957 /* now, allocate fbarray itself */
958 if (eal_memseg_list_init(msl, page_sz, n_segs,
959 socket, msl_idx, true) < 0)
962 /* finally, allocate VA space */
963 if (eal_memseg_list_alloc(msl, 0) < 0) {
964 RTE_LOG(ERR, EAL, "Cannot preallocate 0x%"PRIx64"kB hugepages\n",
974 * We cannot reallocate memseg lists on the fly because PPC64 stores pages
975 * backwards, therefore we have to process the entire memseg first before
976 * remapping it into memseg list VA space.
979 remap_needed_hugepages(struct hugepage_file *hugepages, int n_pages)
981 int cur_page, seg_start_page, new_memseg, ret;
984 for (cur_page = 0; cur_page < n_pages; cur_page++) {
985 struct hugepage_file *prev, *cur;
989 cur = &hugepages[cur_page];
990 prev = cur_page == 0 ? NULL : &hugepages[cur_page - 1];
992 /* if size is zero, no more pages left */
998 else if (cur->socket_id != prev->socket_id)
1000 else if (cur->size != prev->size)
1002 #ifdef RTE_ARCH_PPC_64
1003 /* On PPC64 architecture, the mmap always start from higher
1004 * address to lower address. Here, physical addresses are in
1007 else if ((prev->physaddr - cur->physaddr) != cur->size)
1010 else if ((cur->physaddr - prev->physaddr) != cur->size)
1015 /* if this isn't the first time, remap segment */
1016 if (cur_page != 0) {
1017 ret = remap_segment(hugepages, seg_start_page,
1022 /* remember where we started */
1023 seg_start_page = cur_page;
1025 /* continuation of previous memseg */
1027 /* we were stopped, but we didn't remap the last segment, do it now */
1028 if (cur_page != 0) {
1029 ret = remap_segment(hugepages, seg_start_page,
1037 static inline size_t
1038 eal_get_hugepage_mem_size(void)
1043 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
1044 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1045 if (strnlen(hpi->hugedir, sizeof(hpi->hugedir)) != 0) {
1046 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1047 size += hpi->hugepage_sz * hpi->num_pages[j];
1052 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
1055 static struct sigaction huge_action_old;
1056 static int huge_need_recover;
1059 huge_register_sigbus(void)
1062 struct sigaction action;
1065 sigaddset(&mask, SIGBUS);
1066 action.sa_flags = 0;
1067 action.sa_mask = mask;
1068 action.sa_handler = huge_sigbus_handler;
1070 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1074 huge_recover_sigbus(void)
1076 if (huge_need_recover) {
1077 sigaction(SIGBUS, &huge_action_old, NULL);
1078 huge_need_recover = 0;
1083 * Prepare physical memory mapping: fill configuration structure with
1084 * these infos, return 0 on success.
1085 * 1. map N huge pages in separate files in hugetlbfs
1086 * 2. find associated physical addr
1087 * 3. find associated NUMA socket ID
1088 * 4. sort all huge pages by physical address
1089 * 5. remap these N huge pages in the correct order
1090 * 6. unmap the first mapping
1091 * 7. fill memsegs in configuration with contiguous zones
1094 eal_legacy_hugepage_init(void)
1096 struct rte_mem_config *mcfg;
1097 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1098 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1100 uint64_t memory[RTE_MAX_NUMA_NODES];
1104 int nr_hugefiles, nr_hugepages = 0;
1107 memset(used_hp, 0, sizeof(used_hp));
1109 /* get pointer to global configuration */
1110 mcfg = rte_eal_get_configuration()->mem_config;
1112 /* hugetlbfs can be disabled */
1113 if (internal_config.no_hugetlbfs) {
1114 void *prealloc_addr;
1116 struct rte_memseg_list *msl;
1117 int n_segs, fd, flags;
1118 #ifdef MEMFD_SUPPORTED
1123 /* nohuge mode is legacy mode */
1124 internal_config.legacy_mem = 1;
1126 /* nohuge mode is single-file segments mode */
1127 internal_config.single_file_segments = 1;
1129 /* create a memseg list */
1130 msl = &mcfg->memsegs[0];
1132 mem_sz = internal_config.memory;
1133 page_sz = RTE_PGSIZE_4K;
1134 n_segs = mem_sz / page_sz;
1136 if (eal_memseg_list_init_named(
1137 msl, "nohugemem", page_sz, n_segs, 0, true)) {
1141 /* set up parameters for anonymous mmap */
1143 flags = MAP_PRIVATE | MAP_ANONYMOUS;
1145 #ifdef MEMFD_SUPPORTED
1146 /* create a memfd and store it in the segment fd table */
1147 memfd = memfd_create("nohuge", 0);
1149 RTE_LOG(DEBUG, EAL, "Cannot create memfd: %s\n",
1151 RTE_LOG(DEBUG, EAL, "Falling back to anonymous map\n");
1153 /* we got an fd - now resize it */
1154 if (ftruncate(memfd, internal_config.memory) < 0) {
1155 RTE_LOG(ERR, EAL, "Cannot resize memfd: %s\n",
1157 RTE_LOG(ERR, EAL, "Falling back to anonymous map\n");
1160 /* creating memfd-backed file was successful.
1161 * we want changes to memfd to be visible to
1162 * other processes (such as vhost backend), so
1163 * map it as shared memory.
1165 RTE_LOG(DEBUG, EAL, "Using memfd for anonymous memory\n");
1171 /* preallocate address space for the memory, so that it can be
1172 * fit into the DMA mask.
1174 if (eal_memseg_list_alloc(msl, 0)) {
1175 RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
1179 prealloc_addr = msl->base_va;
1180 addr = mmap(prealloc_addr, mem_sz, PROT_READ | PROT_WRITE,
1181 flags | MAP_FIXED, fd, 0);
1182 if (addr == MAP_FAILED || addr != prealloc_addr) {
1183 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1185 munmap(prealloc_addr, mem_sz);
1189 /* we're in single-file segments mode, so only the segment list
1190 * fd needs to be set up.
1193 if (eal_memalloc_set_seg_list_fd(0, fd) < 0) {
1194 RTE_LOG(ERR, EAL, "Cannot set up segment list fd\n");
1195 /* not a serious error, proceed */
1199 eal_memseg_list_populate(msl, addr, n_segs);
1201 if (mcfg->dma_maskbits &&
1202 rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
1204 "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask.\n",
1206 if (rte_eal_iova_mode() == RTE_IOVA_VA &&
1207 rte_eal_using_phys_addrs())
1209 "%s(): Please try initializing EAL with --iova-mode=pa parameter.\n",
1216 /* calculate total number of hugepages available. at this point we haven't
1217 * yet started sorting them so they all are on socket 0 */
1218 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1219 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1220 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1222 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1226 * allocate a memory area for hugepage table.
1227 * this isn't shared memory yet. due to the fact that we need some
1228 * processing done on these pages, shared memory will be created
1231 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1235 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1237 hp_offset = 0; /* where we start the current page size entries */
1239 huge_register_sigbus();
1241 /* make a copy of socket_mem, needed for balanced allocation. */
1242 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1243 memory[i] = internal_config.socket_mem[i];
1245 /* map all hugepages and sort them */
1246 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1247 unsigned pages_old, pages_new;
1248 struct hugepage_info *hpi;
1251 * we don't yet mark hugepages as used at this stage, so
1252 * we just map all hugepages available to the system
1253 * all hugepages are still located on socket 0
1255 hpi = &internal_config.hugepage_info[i];
1257 if (hpi->num_pages[0] == 0)
1260 /* map all hugepages available */
1261 pages_old = hpi->num_pages[0];
1262 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, memory);
1263 if (pages_new < pages_old) {
1265 "%d not %d hugepages of size %u MB allocated\n",
1266 pages_new, pages_old,
1267 (unsigned)(hpi->hugepage_sz / 0x100000));
1269 int pages = pages_old - pages_new;
1271 nr_hugepages -= pages;
1272 hpi->num_pages[0] = pages_new;
1277 if (rte_eal_using_phys_addrs() &&
1278 rte_eal_iova_mode() != RTE_IOVA_VA) {
1279 /* find physical addresses for each hugepage */
1280 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1281 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1282 "for %u MB pages\n",
1283 (unsigned int)(hpi->hugepage_sz / 0x100000));
1287 /* set physical addresses for each hugepage */
1288 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1289 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1290 "for %u MB pages\n",
1291 (unsigned int)(hpi->hugepage_sz / 0x100000));
1296 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1297 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1298 (unsigned)(hpi->hugepage_sz / 0x100000));
1302 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1303 sizeof(struct hugepage_file), cmp_physaddr);
1305 /* we have processed a num of hugepages of this size, so inc offset */
1306 hp_offset += hpi->num_pages[0];
1309 huge_recover_sigbus();
1311 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1312 internal_config.memory = eal_get_hugepage_mem_size();
1314 nr_hugefiles = nr_hugepages;
1317 /* clean out the numbers of pages */
1318 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1319 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1320 internal_config.hugepage_info[i].num_pages[j] = 0;
1322 /* get hugepages for each socket */
1323 for (i = 0; i < nr_hugefiles; i++) {
1324 int socket = tmp_hp[i].socket_id;
1326 /* find a hugepage info with right size and increment num_pages */
1327 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1328 (int)internal_config.num_hugepage_sizes);
1329 for (j = 0; j < nb_hpsizes; j++) {
1330 if (tmp_hp[i].size ==
1331 internal_config.hugepage_info[j].hugepage_sz) {
1332 internal_config.hugepage_info[j].num_pages[socket]++;
1337 /* make a copy of socket_mem, needed for number of pages calculation */
1338 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1339 memory[i] = internal_config.socket_mem[i];
1341 /* calculate final number of pages */
1342 nr_hugepages = eal_dynmem_calc_num_pages_per_socket(memory,
1343 internal_config.hugepage_info, used_hp,
1344 internal_config.num_hugepage_sizes);
1346 /* error if not enough memory available */
1347 if (nr_hugepages < 0)
1351 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1352 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1353 if (used_hp[i].num_pages[j] > 0) {
1355 "Requesting %u pages of size %uMB"
1356 " from socket %i\n",
1357 used_hp[i].num_pages[j],
1359 (used_hp[i].hugepage_sz / 0x100000),
1365 /* create shared memory */
1366 hugepage = create_shared_memory(eal_hugepage_data_path(),
1367 nr_hugefiles * sizeof(struct hugepage_file));
1369 if (hugepage == NULL) {
1370 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1373 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1376 * unmap pages that we won't need (looks at used_hp).
1377 * also, sets final_va to NULL on pages that were unmapped.
1379 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1380 internal_config.num_hugepage_sizes) < 0) {
1381 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1386 * copy stuff from malloc'd hugepage* to the actual shared memory.
1387 * this procedure only copies those hugepages that have orig_va
1388 * not NULL. has overflow protection.
1390 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1391 tmp_hp, nr_hugefiles) < 0) {
1392 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1397 /* for legacy 32-bit mode, we did not preallocate VA space, so do it */
1398 if (internal_config.legacy_mem &&
1399 prealloc_segments(hugepage, nr_hugefiles)) {
1400 RTE_LOG(ERR, EAL, "Could not preallocate VA space for hugepages\n");
1405 /* remap all pages we do need into memseg list VA space, so that those
1406 * pages become first-class citizens in DPDK memory subsystem
1408 if (remap_needed_hugepages(hugepage, nr_hugefiles)) {
1409 RTE_LOG(ERR, EAL, "Couldn't remap hugepage files into memseg lists\n");
1413 /* free the hugepage backing files */
1414 if (internal_config.hugepage_unlink &&
1415 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1416 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1420 /* free the temporary hugepage table */
1424 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1427 /* we're not going to allocate more pages, so release VA space for
1428 * unused memseg lists
1430 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1431 struct rte_memseg_list *msl = &mcfg->memsegs[i];
1434 /* skip inactive lists */
1435 if (msl->base_va == NULL)
1437 /* skip lists where there is at least one page allocated */
1438 if (msl->memseg_arr.count > 0)
1440 /* this is an unused list, deallocate it */
1442 munmap(msl->base_va, mem_sz);
1443 msl->base_va = NULL;
1446 /* destroy backing fbarray */
1447 rte_fbarray_destroy(&msl->memseg_arr);
1450 if (mcfg->dma_maskbits &&
1451 rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
1453 "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask.\n",
1461 huge_recover_sigbus();
1463 if (hugepage != NULL)
1464 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1470 * uses fstat to report the size of a file on disk
1476 if (fstat(fd, &st) < 0)
1482 * This creates the memory mappings in the secondary process to match that of
1483 * the server process. It goes through each memory segment in the DPDK runtime
1484 * configuration and finds the hugepages which form that segment, mapping them
1485 * in order to form a contiguous block in the virtual memory space
1488 eal_legacy_hugepage_attach(void)
1490 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1491 struct hugepage_file *hp = NULL;
1492 unsigned int num_hp = 0;
1494 unsigned int cur_seg;
1496 int fd, fd_hugepage = -1;
1498 if (aslr_enabled() > 0) {
1499 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1500 "(ASLR) is enabled in the kernel.\n");
1501 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1502 "into secondary processes\n");
1505 fd_hugepage = open(eal_hugepage_data_path(), O_RDONLY);
1506 if (fd_hugepage < 0) {
1507 RTE_LOG(ERR, EAL, "Could not open %s\n",
1508 eal_hugepage_data_path());
1512 size = getFileSize(fd_hugepage);
1513 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1514 if (hp == MAP_FAILED) {
1515 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1516 eal_hugepage_data_path());
1520 num_hp = size / sizeof(struct hugepage_file);
1521 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1523 /* map all segments into memory to make sure we get the addrs. the
1524 * segments themselves are already in memseg list (which is shared and
1525 * has its VA space already preallocated), so we just need to map
1526 * everything into correct addresses.
1528 for (i = 0; i < num_hp; i++) {
1529 struct hugepage_file *hf = &hp[i];
1530 size_t map_sz = hf->size;
1531 void *map_addr = hf->final_va;
1532 int msl_idx, ms_idx;
1533 struct rte_memseg_list *msl;
1534 struct rte_memseg *ms;
1536 /* if size is zero, no more pages left */
1540 fd = open(hf->filepath, O_RDWR);
1542 RTE_LOG(ERR, EAL, "Could not open %s: %s\n",
1543 hf->filepath, strerror(errno));
1547 map_addr = mmap(map_addr, map_sz, PROT_READ | PROT_WRITE,
1548 MAP_SHARED | MAP_FIXED, fd, 0);
1549 if (map_addr == MAP_FAILED) {
1550 RTE_LOG(ERR, EAL, "Could not map %s: %s\n",
1551 hf->filepath, strerror(errno));
1555 /* set shared lock on the file. */
1556 if (flock(fd, LOCK_SH) < 0) {
1557 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed: %s\n",
1558 __func__, strerror(errno));
1562 /* find segment data */
1563 msl = rte_mem_virt2memseg_list(map_addr);
1565 RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg list\n",
1569 ms = rte_mem_virt2memseg(map_addr, msl);
1571 RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg\n",
1576 msl_idx = msl - mcfg->memsegs;
1577 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
1579 RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg idx\n",
1584 /* store segment fd internally */
1585 if (eal_memalloc_set_seg_fd(msl_idx, ms_idx, fd) < 0)
1586 RTE_LOG(ERR, EAL, "Could not store segment fd: %s\n",
1587 rte_strerror(rte_errno));
1589 /* unmap the hugepage config file, since we are done using it */
1595 munmap(hp[i].final_va, hp[i].size);
1599 /* unwind mmap's done so far */
1600 for (cur_seg = 0; cur_seg < i; cur_seg++)
1601 munmap(hp[cur_seg].final_va, hp[cur_seg].size);
1603 if (hp != NULL && hp != MAP_FAILED)
1605 if (fd_hugepage >= 0)
1611 eal_hugepage_attach(void)
1613 if (eal_memalloc_sync_with_primary()) {
1614 RTE_LOG(ERR, EAL, "Could not map memory from primary process\n");
1615 if (aslr_enabled() > 0)
1616 RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel and retry running both primary and secondary processes\n");
1623 rte_eal_hugepage_init(void)
1625 return internal_config.legacy_mem ?
1626 eal_legacy_hugepage_init() :
1627 eal_dynmem_hugepage_init();
1631 rte_eal_hugepage_attach(void)
1633 return internal_config.legacy_mem ?
1634 eal_legacy_hugepage_attach() :
1635 eal_hugepage_attach();
1639 rte_eal_using_phys_addrs(void)
1641 if (phys_addrs_available == -1) {
1644 if (rte_eal_has_hugepages() != 0 &&
1645 rte_mem_virt2phy(&tmp) != RTE_BAD_PHYS_ADDR)
1646 phys_addrs_available = 1;
1648 phys_addrs_available = 0;
1650 return phys_addrs_available;
1653 static int __rte_unused
1654 memseg_primary_init_32(void)
1656 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1657 int active_sockets, hpi_idx, msl_idx = 0;
1658 unsigned int socket_id, i;
1659 struct rte_memseg_list *msl;
1660 uint64_t extra_mem_per_socket, total_extra_mem, total_requested_mem;
1663 /* no-huge does not need this at all */
1664 if (internal_config.no_hugetlbfs)
1667 /* this is a giant hack, but desperate times call for desperate
1668 * measures. in legacy 32-bit mode, we cannot preallocate VA space,
1669 * because having upwards of 2 gigabytes of VA space already mapped will
1670 * interfere with our ability to map and sort hugepages.
1672 * therefore, in legacy 32-bit mode, we will be initializing memseg
1673 * lists much later - in eal_memory.c, right after we unmap all the
1674 * unneeded pages. this will not affect secondary processes, as those
1675 * should be able to mmap the space without (too many) problems.
1677 if (internal_config.legacy_mem)
1680 /* 32-bit mode is a very special case. we cannot know in advance where
1681 * the user will want to allocate their memory, so we have to do some
1685 total_requested_mem = 0;
1686 if (internal_config.force_sockets)
1687 for (i = 0; i < rte_socket_count(); i++) {
1690 socket_id = rte_socket_id_by_idx(i);
1691 mem = internal_config.socket_mem[socket_id];
1697 total_requested_mem += mem;
1700 total_requested_mem = internal_config.memory;
1702 max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
1703 if (total_requested_mem > max_mem) {
1704 RTE_LOG(ERR, EAL, "Invalid parameters: 32-bit process can at most use %uM of memory\n",
1705 (unsigned int)(max_mem >> 20));
1708 total_extra_mem = max_mem - total_requested_mem;
1709 extra_mem_per_socket = active_sockets == 0 ? total_extra_mem :
1710 total_extra_mem / active_sockets;
1712 /* the allocation logic is a little bit convoluted, but here's how it
1713 * works, in a nutshell:
1714 * - if user hasn't specified on which sockets to allocate memory via
1715 * --socket-mem, we allocate all of our memory on master core socket.
1716 * - if user has specified sockets to allocate memory on, there may be
1717 * some "unused" memory left (e.g. if user has specified --socket-mem
1718 * such that not all memory adds up to 2 gigabytes), so add it to all
1719 * sockets that are in use equally.
1721 * page sizes are sorted by size in descending order, so we can safely
1722 * assume that we dispense with bigger page sizes first.
1725 /* create memseg lists */
1726 for (i = 0; i < rte_socket_count(); i++) {
1727 int hp_sizes = (int) internal_config.num_hugepage_sizes;
1728 uint64_t max_socket_mem, cur_socket_mem;
1729 unsigned int master_lcore_socket;
1730 struct rte_config *cfg = rte_eal_get_configuration();
1733 socket_id = rte_socket_id_by_idx(i);
1735 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
1736 /* we can still sort pages by socket in legacy mode */
1737 if (!internal_config.legacy_mem && socket_id > 0)
1741 /* if we didn't specifically request memory on this socket */
1742 skip = active_sockets != 0 &&
1743 internal_config.socket_mem[socket_id] == 0;
1744 /* ...or if we didn't specifically request memory on *any*
1745 * socket, and this is not master lcore
1747 master_lcore_socket = rte_lcore_to_socket_id(cfg->master_lcore);
1748 skip |= active_sockets == 0 && socket_id != master_lcore_socket;
1751 RTE_LOG(DEBUG, EAL, "Will not preallocate memory on socket %u\n",
1756 /* max amount of memory on this socket */
1757 max_socket_mem = (active_sockets != 0 ?
1758 internal_config.socket_mem[socket_id] :
1759 internal_config.memory) +
1760 extra_mem_per_socket;
1763 for (hpi_idx = 0; hpi_idx < hp_sizes; hpi_idx++) {
1764 uint64_t max_pagesz_mem, cur_pagesz_mem = 0;
1765 uint64_t hugepage_sz;
1766 struct hugepage_info *hpi;
1767 int type_msl_idx, max_segs, total_segs = 0;
1769 hpi = &internal_config.hugepage_info[hpi_idx];
1770 hugepage_sz = hpi->hugepage_sz;
1772 /* check if pages are actually available */
1773 if (hpi->num_pages[socket_id] == 0)
1776 max_segs = RTE_MAX_MEMSEG_PER_TYPE;
1777 max_pagesz_mem = max_socket_mem - cur_socket_mem;
1779 /* make it multiple of page size */
1780 max_pagesz_mem = RTE_ALIGN_FLOOR(max_pagesz_mem,
1783 RTE_LOG(DEBUG, EAL, "Attempting to preallocate "
1784 "%" PRIu64 "M on socket %i\n",
1785 max_pagesz_mem >> 20, socket_id);
1788 while (cur_pagesz_mem < max_pagesz_mem &&
1789 total_segs < max_segs) {
1791 unsigned int n_segs;
1793 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
1795 "No more space in memseg lists, please increase %s\n",
1796 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
1800 msl = &mcfg->memsegs[msl_idx];
1802 cur_mem = get_mem_amount(hugepage_sz,
1804 n_segs = cur_mem / hugepage_sz;
1806 if (eal_memseg_list_init(msl, hugepage_sz,
1807 n_segs, socket_id, type_msl_idx,
1809 /* failing to allocate a memseg list is
1812 RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
1816 if (eal_memseg_list_alloc(msl, 0)) {
1817 /* if we couldn't allocate VA space, we
1818 * can try with smaller page sizes.
1820 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list, retrying with different page size\n");
1821 /* deallocate memseg list */
1822 if (memseg_list_free(msl))
1827 total_segs += msl->memseg_arr.len;
1828 cur_pagesz_mem = total_segs * hugepage_sz;
1832 cur_socket_mem += cur_pagesz_mem;
1834 if (cur_socket_mem == 0) {
1835 RTE_LOG(ERR, EAL, "Cannot allocate VA space on socket %u\n",
1844 static int __rte_unused
1845 memseg_primary_init(void)
1847 return eal_dynmem_memseg_lists_init();
1851 memseg_secondary_init(void)
1853 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1855 struct rte_memseg_list *msl;
1857 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
1859 msl = &mcfg->memsegs[msl_idx];
1861 /* skip empty memseg lists */
1862 if (msl->memseg_arr.len == 0)
1865 if (rte_fbarray_attach(&msl->memseg_arr)) {
1866 RTE_LOG(ERR, EAL, "Cannot attach to primary process memseg lists\n");
1870 /* preallocate VA space */
1871 if (eal_memseg_list_alloc(msl, 0)) {
1872 RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
1881 rte_eal_memseg_init(void)
1883 /* increase rlimit to maximum */
1886 if (getrlimit(RLIMIT_NOFILE, &lim) == 0) {
1887 /* set limit to maximum */
1888 lim.rlim_cur = lim.rlim_max;
1890 if (setrlimit(RLIMIT_NOFILE, &lim) < 0) {
1891 RTE_LOG(DEBUG, EAL, "Setting maximum number of open files failed: %s\n",
1894 RTE_LOG(DEBUG, EAL, "Setting maximum number of open files to %"
1896 (uint64_t)lim.rlim_cur);
1899 RTE_LOG(ERR, EAL, "Cannot get current resource limits\n");
1901 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
1902 if (!internal_config.legacy_mem && rte_socket_count() > 1) {
1903 RTE_LOG(WARNING, EAL, "DPDK is running on a NUMA system, but is compiled without NUMA support.\n");
1904 RTE_LOG(WARNING, EAL, "This will have adverse consequences for performance and usability.\n");
1905 RTE_LOG(WARNING, EAL, "Please use --"OPT_LEGACY_MEM" option, or recompile with NUMA support.\n");
1909 return rte_eal_process_type() == RTE_PROC_PRIMARY ?
1911 memseg_primary_init_32() :
1913 memseg_primary_init() :
1915 memseg_secondary_init();