1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation.
3 * Copyright(c) 2013 6WIND S.A.
6 #define _FILE_OFFSET_BITS 64
16 #include <sys/types.h>
18 #include <sys/queue.h>
22 #include <sys/ioctl.h>
26 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
32 #include <rte_memory.h>
33 #include <rte_launch.h>
35 #include <rte_eal_memconfig.h>
36 #include <rte_per_lcore.h>
37 #include <rte_lcore.h>
38 #include <rte_common.h>
39 #include <rte_string_fns.h>
41 #include "eal_private.h"
42 #include "eal_internal_cfg.h"
43 #include "eal_filesystem.h"
44 #include "eal_hugepages.h"
46 #define PFN_MASK_SIZE 8
50 * Huge page mapping under linux
52 * To reserve a big contiguous amount of memory, we use the hugepage
53 * feature of linux. For that, we need to have hugetlbfs mounted. This
54 * code will create many files in this directory (one per page) and
55 * map them in virtual memory. For each page, we will retrieve its
56 * physical address and remap it in order to have a virtual contiguous
57 * zone as well as a physical contiguous zone.
60 static uint64_t baseaddr_offset;
62 static bool phys_addrs_available = true;
64 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
67 test_phys_addrs_available(void)
72 if (!rte_eal_has_hugepages()) {
74 "Started without hugepages support, physical addresses not available\n");
75 phys_addrs_available = false;
79 physaddr = rte_mem_virt2phy(&tmp);
80 if (physaddr == RTE_BAD_PHYS_ADDR) {
81 if (rte_eal_iova_mode() == RTE_IOVA_PA)
83 "Cannot obtain physical addresses: %s. "
84 "Only vfio will function.\n",
86 phys_addrs_available = false;
91 * Get physical address of any mapped virtual address in the current process.
94 rte_mem_virt2phy(const void *virtaddr)
97 uint64_t page, physaddr;
98 unsigned long virt_pfn;
102 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
103 if (!phys_addrs_available)
106 /* standard page size */
107 page_size = getpagesize();
109 fd = open("/proc/self/pagemap", O_RDONLY);
111 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
112 __func__, strerror(errno));
116 virt_pfn = (unsigned long)virtaddr / page_size;
117 offset = sizeof(uint64_t) * virt_pfn;
118 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
119 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
120 __func__, strerror(errno));
125 retval = read(fd, &page, PFN_MASK_SIZE);
128 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
129 __func__, strerror(errno));
131 } else if (retval != PFN_MASK_SIZE) {
132 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
133 "but expected %d:\n",
134 __func__, retval, PFN_MASK_SIZE);
139 * the pfn (page frame number) are bits 0-54 (see
140 * pagemap.txt in linux Documentation)
142 if ((page & 0x7fffffffffffffULL) == 0)
145 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
146 + ((unsigned long)virtaddr % page_size);
152 rte_mem_virt2iova(const void *virtaddr)
154 if (rte_eal_iova_mode() == RTE_IOVA_VA)
155 return (uintptr_t)virtaddr;
156 return rte_mem_virt2phy(virtaddr);
160 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
161 * it by browsing the /proc/self/pagemap special file.
164 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
169 for (i = 0; i < hpi->num_pages[0]; i++) {
170 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
171 if (addr == RTE_BAD_PHYS_ADDR)
173 hugepg_tbl[i].physaddr = addr;
179 * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
182 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
185 static phys_addr_t addr;
187 for (i = 0; i < hpi->num_pages[0]; i++) {
188 hugepg_tbl[i].physaddr = addr;
189 addr += hugepg_tbl[i].size;
195 * Check whether address-space layout randomization is enabled in
196 * the kernel. This is important for multi-process as it can prevent
197 * two processes mapping data to the same virtual address
199 * 0 - address space randomization disabled
200 * 1/2 - address space randomization enabled
201 * negative error code on error
207 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
210 retval = read(fd, &c, 1);
220 default: return -EINVAL;
225 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
226 * pointer to the mmap'd area and keep *size unmodified. Else, retry
227 * with a smaller zone: decrease *size by hugepage_sz until it reaches
228 * 0. In this case, return NULL. Note: this function returns an address
229 * which is a multiple of hugepage size.
232 get_virtual_area(size_t *size, size_t hugepage_sz)
239 if (internal_config.base_virtaddr != 0) {
240 int page_size = sysconf(_SC_PAGE_SIZE);
241 addr_hint = (void *) (uintptr_t)
242 (internal_config.base_virtaddr + baseaddr_offset);
243 addr_hint = RTE_PTR_ALIGN_FLOOR(addr_hint, page_size);
248 RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
251 fd = open("/dev/zero", O_RDONLY);
253 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
257 addr = mmap(addr_hint, (*size) + hugepage_sz, PROT_READ,
258 #ifdef RTE_ARCH_PPC_64
259 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
264 if (addr == MAP_FAILED) {
265 *size -= hugepage_sz;
266 } else if (addr_hint != NULL && addr != addr_hint) {
267 RTE_LOG(WARNING, EAL, "WARNING! Base virtual address "
268 "hint (%p != %p) not respected!\n",
270 RTE_LOG(WARNING, EAL, " This may cause issues with "
271 "mapping memory into secondary processes\n");
273 } while (addr == MAP_FAILED && *size > 0);
275 if (addr == MAP_FAILED) {
277 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
282 munmap(addr, (*size) + hugepage_sz);
285 /* align addr to a huge page size boundary */
286 aligned_addr = (long)addr;
287 aligned_addr += (hugepage_sz - 1);
288 aligned_addr &= (~(hugepage_sz - 1));
289 addr = (void *)(aligned_addr);
291 RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
294 /* increment offset */
295 baseaddr_offset += *size;
300 static sigjmp_buf huge_jmpenv;
302 static void huge_sigbus_handler(int signo __rte_unused)
304 siglongjmp(huge_jmpenv, 1);
307 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
308 * non-static local variable in the stack frame calling sigsetjmp might be
309 * clobbered by a call to longjmp.
311 static int huge_wrap_sigsetjmp(void)
313 return sigsetjmp(huge_jmpenv, 1);
316 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
317 /* Callback for numa library. */
318 void numa_error(char *where)
320 RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
325 * Mmap all hugepages of hugepage table: it first open a file in
326 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
327 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
328 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
329 * map contiguous physical blocks in contiguous virtual blocks.
332 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
333 uint64_t *essential_memory __rte_unused, int orig)
338 void *vma_addr = NULL;
340 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
342 int essential_prev = 0;
344 struct bitmask *oldmask = numa_allocate_nodemask();
345 bool have_numa = true;
346 unsigned long maxnode = 0;
348 /* Check if kernel supports NUMA. */
349 if (numa_available() != 0) {
350 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
354 if (orig && have_numa) {
355 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
356 if (get_mempolicy(&oldpolicy, oldmask->maskp,
357 oldmask->size + 1, 0, 0) < 0) {
359 "Failed to get current mempolicy: %s. "
360 "Assuming MPOL_DEFAULT.\n", strerror(errno));
361 oldpolicy = MPOL_DEFAULT;
363 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
364 if (internal_config.socket_mem[i])
369 for (i = 0; i < hpi->num_pages[0]; i++) {
370 uint64_t hugepage_sz = hpi->hugepage_sz;
372 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
376 for (j = 0; j < maxnode; j++)
377 if (essential_memory[j])
381 node_id = (node_id + 1) % maxnode;
382 while (!internal_config.socket_mem[node_id]) {
389 essential_prev = essential_memory[j];
391 if (essential_memory[j] < hugepage_sz)
392 essential_memory[j] = 0;
394 essential_memory[j] -= hugepage_sz;
398 "Setting policy MPOL_PREFERRED for socket %d\n",
400 numa_set_preferred(node_id);
405 hugepg_tbl[i].file_id = i;
406 hugepg_tbl[i].size = hugepage_sz;
407 eal_get_hugefile_path(hugepg_tbl[i].filepath,
408 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
409 hugepg_tbl[i].file_id);
410 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
413 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
414 * original map address as final map address.
416 else if ((hugepage_sz == RTE_PGSIZE_1G)
417 || (hugepage_sz == RTE_PGSIZE_16G)) {
418 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
419 hugepg_tbl[i].orig_va = NULL;
423 else if (vma_len == 0) {
424 unsigned j, num_pages;
426 /* reserve a virtual area for next contiguous
427 * physical block: count the number of
428 * contiguous physical pages. */
429 for (j = i+1; j < hpi->num_pages[0] ; j++) {
430 #ifdef RTE_ARCH_PPC_64
431 /* The physical addresses are sorted in
432 * descending order on PPC64 */
433 if (hugepg_tbl[j].physaddr !=
434 hugepg_tbl[j-1].physaddr - hugepage_sz)
437 if (hugepg_tbl[j].physaddr !=
438 hugepg_tbl[j-1].physaddr + hugepage_sz)
443 vma_len = num_pages * hugepage_sz;
445 /* get the biggest virtual memory area up to
446 * vma_len. If it fails, vma_addr is NULL, so
447 * let the kernel provide the address. */
448 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
449 if (vma_addr == NULL)
450 vma_len = hugepage_sz;
453 /* try to create hugepage file */
454 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
456 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
461 /* map the segment, and populate page tables,
462 * the kernel fills this segment with zeros */
463 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
464 MAP_SHARED | MAP_POPULATE, fd, 0);
465 if (virtaddr == MAP_FAILED) {
466 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
473 hugepg_tbl[i].orig_va = virtaddr;
476 hugepg_tbl[i].final_va = virtaddr;
480 /* In linux, hugetlb limitations, like cgroup, are
481 * enforced at fault time instead of mmap(), even
482 * with the option of MAP_POPULATE. Kernel will send
483 * a SIGBUS signal. To avoid to be killed, save stack
484 * environment here, if SIGBUS happens, we can jump
487 if (huge_wrap_sigsetjmp()) {
488 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
489 "hugepages of size %u MB\n",
490 (unsigned)(hugepage_sz / 0x100000));
491 munmap(virtaddr, hugepage_sz);
493 unlink(hugepg_tbl[i].filepath);
494 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
496 essential_memory[node_id] =
501 *(int *)virtaddr = 0;
505 /* set shared flock on the file. */
506 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
507 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
508 __func__, strerror(errno));
515 vma_addr = (char *)vma_addr + hugepage_sz;
516 vma_len -= hugepage_sz;
520 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
523 "Restoring previous memory policy: %d\n", oldpolicy);
524 if (oldpolicy == MPOL_DEFAULT) {
525 numa_set_localalloc();
526 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
527 oldmask->size + 1) < 0) {
528 RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
530 numa_set_localalloc();
533 numa_free_cpumask(oldmask);
538 /* Unmap all hugepages from original mapping */
540 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
543 for (i = 0; i < hpi->num_pages[0]; i++) {
544 if (hugepg_tbl[i].orig_va) {
545 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
546 hugepg_tbl[i].orig_va = NULL;
553 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
557 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
561 unsigned i, hp_count = 0;
564 char hugedir_str[PATH_MAX];
567 f = fopen("/proc/self/numa_maps", "r");
569 RTE_LOG(NOTICE, EAL, "NUMA support not available"
570 " consider that all memory is in socket_id 0\n");
574 snprintf(hugedir_str, sizeof(hugedir_str),
575 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
578 while (fgets(buf, sizeof(buf), f) != NULL) {
580 /* ignore non huge page */
581 if (strstr(buf, " huge ") == NULL &&
582 strstr(buf, hugedir_str) == NULL)
586 virt_addr = strtoull(buf, &end, 16);
587 if (virt_addr == 0 || end == buf) {
588 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
592 /* get node id (socket id) */
593 nodestr = strstr(buf, " N");
594 if (nodestr == NULL) {
595 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
599 end = strstr(nodestr, "=");
601 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
607 socket_id = strtoul(nodestr, &end, 0);
608 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
609 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
613 /* if we find this page in our mappings, set socket_id */
614 for (i = 0; i < hpi->num_pages[0]; i++) {
615 void *va = (void *)(unsigned long)virt_addr;
616 if (hugepg_tbl[i].orig_va == va) {
617 hugepg_tbl[i].socket_id = socket_id;
619 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
621 "Hugepage %s is on socket %d\n",
622 hugepg_tbl[i].filepath, socket_id);
628 if (hp_count < hpi->num_pages[0])
640 cmp_physaddr(const void *a, const void *b)
642 #ifndef RTE_ARCH_PPC_64
643 const struct hugepage_file *p1 = a;
644 const struct hugepage_file *p2 = b;
646 /* PowerPC needs memory sorted in reverse order from x86 */
647 const struct hugepage_file *p1 = b;
648 const struct hugepage_file *p2 = a;
650 if (p1->physaddr < p2->physaddr)
652 else if (p1->physaddr > p2->physaddr)
659 * Uses mmap to create a shared memory area for storage of data
660 * Used in this file to store the hugepage file map on disk
663 create_shared_memory(const char *filename, const size_t mem_size)
666 int fd = open(filename, O_CREAT | O_RDWR, 0666);
669 if (ftruncate(fd, mem_size) < 0) {
673 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
675 if (retval == MAP_FAILED)
681 * this copies *active* hugepages from one hugepage table to another.
682 * destination is typically the shared memory.
685 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
686 const struct hugepage_file * src, int src_size)
688 int src_pos, dst_pos = 0;
690 for (src_pos = 0; src_pos < src_size; src_pos++) {
691 if (src[src_pos].final_va != NULL) {
692 /* error on overflow attempt */
693 if (dst_pos == dest_size)
695 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
703 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
704 unsigned num_hp_info)
706 unsigned socket, size;
707 int page, nrpages = 0;
709 /* get total number of hugepages */
710 for (size = 0; size < num_hp_info; size++)
711 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
713 internal_config.hugepage_info[size].num_pages[socket];
715 for (page = 0; page < nrpages; page++) {
716 struct hugepage_file *hp = &hugepg_tbl[page];
718 if (hp->final_va != NULL && unlink(hp->filepath)) {
719 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
720 __func__, hp->filepath, strerror(errno));
727 * unmaps hugepages that are not going to be used. since we originally allocate
728 * ALL hugepages (not just those we need), additional unmapping needs to be done.
731 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
732 struct hugepage_info *hpi,
733 unsigned num_hp_info)
735 unsigned socket, size;
736 int page, nrpages = 0;
738 /* get total number of hugepages */
739 for (size = 0; size < num_hp_info; size++)
740 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
741 nrpages += internal_config.hugepage_info[size].num_pages[socket];
743 for (size = 0; size < num_hp_info; size++) {
744 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
745 unsigned pages_found = 0;
747 /* traverse until we have unmapped all the unused pages */
748 for (page = 0; page < nrpages; page++) {
749 struct hugepage_file *hp = &hugepg_tbl[page];
751 /* find a page that matches the criteria */
752 if ((hp->size == hpi[size].hugepage_sz) &&
753 (hp->socket_id == (int) socket)) {
755 /* if we skipped enough pages, unmap the rest */
756 if (pages_found == hpi[size].num_pages[socket]) {
759 unmap_len = hp->size;
761 /* get start addr and len of the remaining segment */
762 munmap(hp->final_va, (size_t) unmap_len);
765 if (unlink(hp->filepath) == -1) {
766 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
767 __func__, hp->filepath, strerror(errno));
771 /* lock the page and skip */
777 } /* foreach socket */
778 } /* foreach pagesize */
783 static inline uint64_t
784 get_socket_mem_size(int socket)
789 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
790 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
791 if (hpi->hugedir != NULL)
792 size += hpi->hugepage_sz * hpi->num_pages[socket];
799 * This function is a NUMA-aware equivalent of calc_num_pages.
800 * It takes in the list of hugepage sizes and the
801 * number of pages thereof, and calculates the best number of
802 * pages of each size to fulfill the request for <memory> ram
805 calc_num_pages_per_socket(uint64_t * memory,
806 struct hugepage_info *hp_info,
807 struct hugepage_info *hp_used,
808 unsigned num_hp_info)
810 unsigned socket, j, i = 0;
811 unsigned requested, available;
812 int total_num_pages = 0;
813 uint64_t remaining_mem, cur_mem;
814 uint64_t total_mem = internal_config.memory;
816 if (num_hp_info == 0)
819 /* if specific memory amounts per socket weren't requested */
820 if (internal_config.force_sockets == 0) {
821 int cpu_per_socket[RTE_MAX_NUMA_NODES];
822 size_t default_size, total_size;
825 /* Compute number of cores per socket */
826 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
827 RTE_LCORE_FOREACH(lcore_id) {
828 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
832 * Automatically spread requested memory amongst detected sockets according
833 * to number of cores from cpu mask present on each socket
835 total_size = internal_config.memory;
836 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
838 /* Set memory amount per socket */
839 default_size = (internal_config.memory * cpu_per_socket[socket])
842 /* Limit to maximum available memory on socket */
843 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
846 memory[socket] = default_size;
847 total_size -= default_size;
851 * If some memory is remaining, try to allocate it by getting all
852 * available memory from sockets, one after the other
854 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
855 /* take whatever is available */
856 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
860 memory[socket] += default_size;
861 total_size -= default_size;
865 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
866 /* skips if the memory on specific socket wasn't requested */
867 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
868 hp_used[i].hugedir = hp_info[i].hugedir;
869 hp_used[i].num_pages[socket] = RTE_MIN(
870 memory[socket] / hp_info[i].hugepage_sz,
871 hp_info[i].num_pages[socket]);
873 cur_mem = hp_used[i].num_pages[socket] *
874 hp_used[i].hugepage_sz;
876 memory[socket] -= cur_mem;
877 total_mem -= cur_mem;
879 total_num_pages += hp_used[i].num_pages[socket];
881 /* check if we have met all memory requests */
882 if (memory[socket] == 0)
885 /* check if we have any more pages left at this size, if so
886 * move on to next size */
887 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
889 /* At this point we know that there are more pages available that are
890 * bigger than the memory we want, so lets see if we can get enough
891 * from other page sizes.
894 for (j = i+1; j < num_hp_info; j++)
895 remaining_mem += hp_info[j].hugepage_sz *
896 hp_info[j].num_pages[socket];
898 /* is there enough other memory, if not allocate another page and quit */
899 if (remaining_mem < memory[socket]){
900 cur_mem = RTE_MIN(memory[socket],
901 hp_info[i].hugepage_sz);
902 memory[socket] -= cur_mem;
903 total_mem -= cur_mem;
904 hp_used[i].num_pages[socket]++;
906 break; /* we are done with this socket*/
909 /* if we didn't satisfy all memory requirements per socket */
910 if (memory[socket] > 0) {
911 /* to prevent icc errors */
912 requested = (unsigned) (internal_config.socket_mem[socket] /
914 available = requested -
915 ((unsigned) (memory[socket] / 0x100000));
916 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
917 "Requested: %uMB, available: %uMB\n", socket,
918 requested, available);
923 /* if we didn't satisfy total memory requirements */
925 requested = (unsigned) (internal_config.memory / 0x100000);
926 available = requested - (unsigned) (total_mem / 0x100000);
927 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
928 " available: %uMB\n", requested, available);
931 return total_num_pages;
935 eal_get_hugepage_mem_size(void)
940 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
941 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
942 if (hpi->hugedir != NULL) {
943 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
944 size += hpi->hugepage_sz * hpi->num_pages[j];
949 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
952 static struct sigaction huge_action_old;
953 static int huge_need_recover;
956 huge_register_sigbus(void)
959 struct sigaction action;
962 sigaddset(&mask, SIGBUS);
964 action.sa_mask = mask;
965 action.sa_handler = huge_sigbus_handler;
967 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
971 huge_recover_sigbus(void)
973 if (huge_need_recover) {
974 sigaction(SIGBUS, &huge_action_old, NULL);
975 huge_need_recover = 0;
980 * Prepare physical memory mapping: fill configuration structure with
981 * these infos, return 0 on success.
982 * 1. map N huge pages in separate files in hugetlbfs
983 * 2. find associated physical addr
984 * 3. find associated NUMA socket ID
985 * 4. sort all huge pages by physical address
986 * 5. remap these N huge pages in the correct order
987 * 6. unmap the first mapping
988 * 7. fill memsegs in configuration with contiguous zones
991 rte_eal_hugepage_init(void)
993 struct rte_mem_config *mcfg;
994 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
995 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
997 uint64_t memory[RTE_MAX_NUMA_NODES];
1000 int i, j, new_memseg;
1001 int nr_hugefiles, nr_hugepages = 0;
1004 test_phys_addrs_available();
1006 memset(used_hp, 0, sizeof(used_hp));
1008 /* get pointer to global configuration */
1009 mcfg = rte_eal_get_configuration()->mem_config;
1011 /* hugetlbfs can be disabled */
1012 if (internal_config.no_hugetlbfs) {
1013 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1014 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1015 if (addr == MAP_FAILED) {
1016 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1020 if (rte_eal_iova_mode() == RTE_IOVA_VA)
1021 mcfg->memseg[0].iova = (uintptr_t)addr;
1023 mcfg->memseg[0].iova = RTE_BAD_IOVA;
1024 mcfg->memseg[0].addr = addr;
1025 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1026 mcfg->memseg[0].len = internal_config.memory;
1027 mcfg->memseg[0].socket_id = 0;
1031 /* calculate total number of hugepages available. at this point we haven't
1032 * yet started sorting them so they all are on socket 0 */
1033 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1034 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1035 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1037 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1041 * allocate a memory area for hugepage table.
1042 * this isn't shared memory yet. due to the fact that we need some
1043 * processing done on these pages, shared memory will be created
1046 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1050 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1052 hp_offset = 0; /* where we start the current page size entries */
1054 huge_register_sigbus();
1056 /* make a copy of socket_mem, needed for balanced allocation. */
1057 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1058 memory[i] = internal_config.socket_mem[i];
1061 /* map all hugepages and sort them */
1062 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1063 unsigned pages_old, pages_new;
1064 struct hugepage_info *hpi;
1067 * we don't yet mark hugepages as used at this stage, so
1068 * we just map all hugepages available to the system
1069 * all hugepages are still located on socket 0
1071 hpi = &internal_config.hugepage_info[i];
1073 if (hpi->num_pages[0] == 0)
1076 /* map all hugepages available */
1077 pages_old = hpi->num_pages[0];
1078 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi,
1080 if (pages_new < pages_old) {
1082 "%d not %d hugepages of size %u MB allocated\n",
1083 pages_new, pages_old,
1084 (unsigned)(hpi->hugepage_sz / 0x100000));
1086 int pages = pages_old - pages_new;
1088 nr_hugepages -= pages;
1089 hpi->num_pages[0] = pages_new;
1094 if (phys_addrs_available) {
1095 /* find physical addresses for each hugepage */
1096 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1097 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1098 "for %u MB pages\n",
1099 (unsigned int)(hpi->hugepage_sz / 0x100000));
1103 /* set physical addresses for each hugepage */
1104 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1105 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1106 "for %u MB pages\n",
1107 (unsigned int)(hpi->hugepage_sz / 0x100000));
1112 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1113 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1114 (unsigned)(hpi->hugepage_sz / 0x100000));
1118 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1119 sizeof(struct hugepage_file), cmp_physaddr);
1121 /* remap all hugepages */
1122 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, NULL, 0) !=
1123 hpi->num_pages[0]) {
1124 RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1125 (unsigned)(hpi->hugepage_sz / 0x100000));
1129 /* unmap original mappings */
1130 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1133 /* we have processed a num of hugepages of this size, so inc offset */
1134 hp_offset += hpi->num_pages[0];
1137 huge_recover_sigbus();
1139 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1140 internal_config.memory = eal_get_hugepage_mem_size();
1142 nr_hugefiles = nr_hugepages;
1145 /* clean out the numbers of pages */
1146 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1147 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1148 internal_config.hugepage_info[i].num_pages[j] = 0;
1150 /* get hugepages for each socket */
1151 for (i = 0; i < nr_hugefiles; i++) {
1152 int socket = tmp_hp[i].socket_id;
1154 /* find a hugepage info with right size and increment num_pages */
1155 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1156 (int)internal_config.num_hugepage_sizes);
1157 for (j = 0; j < nb_hpsizes; j++) {
1158 if (tmp_hp[i].size ==
1159 internal_config.hugepage_info[j].hugepage_sz) {
1160 internal_config.hugepage_info[j].num_pages[socket]++;
1165 /* make a copy of socket_mem, needed for number of pages calculation */
1166 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1167 memory[i] = internal_config.socket_mem[i];
1169 /* calculate final number of pages */
1170 nr_hugepages = calc_num_pages_per_socket(memory,
1171 internal_config.hugepage_info, used_hp,
1172 internal_config.num_hugepage_sizes);
1174 /* error if not enough memory available */
1175 if (nr_hugepages < 0)
1179 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1180 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1181 if (used_hp[i].num_pages[j] > 0) {
1183 "Requesting %u pages of size %uMB"
1184 " from socket %i\n",
1185 used_hp[i].num_pages[j],
1187 (used_hp[i].hugepage_sz / 0x100000),
1193 /* create shared memory */
1194 hugepage = create_shared_memory(eal_hugepage_info_path(),
1195 nr_hugefiles * sizeof(struct hugepage_file));
1197 if (hugepage == NULL) {
1198 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1201 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1204 * unmap pages that we won't need (looks at used_hp).
1205 * also, sets final_va to NULL on pages that were unmapped.
1207 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1208 internal_config.num_hugepage_sizes) < 0) {
1209 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1214 * copy stuff from malloc'd hugepage* to the actual shared memory.
1215 * this procedure only copies those hugepages that have final_va
1216 * not NULL. has overflow protection.
1218 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1219 tmp_hp, nr_hugefiles) < 0) {
1220 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1224 /* free the hugepage backing files */
1225 if (internal_config.hugepage_unlink &&
1226 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1227 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1231 /* free the temporary hugepage table */
1235 /* first memseg index shall be 0 after incrementing it below */
1237 for (i = 0; i < nr_hugefiles; i++) {
1240 /* if this is a new section, create a new memseg */
1243 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1245 else if (hugepage[i].size != hugepage[i-1].size)
1248 #ifdef RTE_ARCH_PPC_64
1249 /* On PPC64 architecture, the mmap always start from higher
1250 * virtual address to lower address. Here, both the physical
1251 * address and virtual address are in descending order */
1252 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1255 else if (((unsigned long)hugepage[i-1].final_va -
1256 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1259 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1262 else if (((unsigned long)hugepage[i].final_va -
1263 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1269 if (j == RTE_MAX_MEMSEG)
1272 mcfg->memseg[j].iova = hugepage[i].physaddr;
1273 mcfg->memseg[j].addr = hugepage[i].final_va;
1274 mcfg->memseg[j].len = hugepage[i].size;
1275 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1276 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1278 /* continuation of previous memseg */
1280 #ifdef RTE_ARCH_PPC_64
1281 /* Use the phy and virt address of the last page as segment
1282 * address for IBM Power architecture */
1283 mcfg->memseg[j].iova = hugepage[i].physaddr;
1284 mcfg->memseg[j].addr = hugepage[i].final_va;
1286 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1288 hugepage[i].memseg_id = j;
1291 if (i < nr_hugefiles) {
1292 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1293 "from %d requested\n"
1294 "Current %s=%d is not enough\n"
1295 "Please either increase it or request less amount "
1297 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1302 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1307 huge_recover_sigbus();
1309 if (hugepage != NULL)
1310 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1316 * uses fstat to report the size of a file on disk
1322 if (fstat(fd, &st) < 0)
1328 * This creates the memory mappings in the secondary process to match that of
1329 * the server process. It goes through each memory segment in the DPDK runtime
1330 * configuration and finds the hugepages which form that segment, mapping them
1331 * in order to form a contiguous block in the virtual memory space
1334 rte_eal_hugepage_attach(void)
1336 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1337 struct hugepage_file *hp = NULL;
1338 unsigned num_hp = 0;
1339 unsigned i, s = 0; /* s used to track the segment number */
1340 unsigned max_seg = RTE_MAX_MEMSEG;
1342 int fd, fd_zero = -1, fd_hugepage = -1;
1344 if (aslr_enabled() > 0) {
1345 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1346 "(ASLR) is enabled in the kernel.\n");
1347 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1348 "into secondary processes\n");
1351 test_phys_addrs_available();
1353 fd_zero = open("/dev/zero", O_RDONLY);
1355 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1358 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1359 if (fd_hugepage < 0) {
1360 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1364 /* map all segments into memory to make sure we get the addrs */
1365 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1369 * the first memory segment with len==0 is the one that
1370 * follows the last valid segment.
1372 if (mcfg->memseg[s].len == 0)
1376 * fdzero is mmapped to get a contiguous block of virtual
1377 * addresses of the appropriate memseg size.
1378 * use mmap to get identical addresses as the primary process.
1380 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1382 #ifdef RTE_ARCH_PPC_64
1383 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1388 if (base_addr == MAP_FAILED ||
1389 base_addr != mcfg->memseg[s].addr) {
1391 if (base_addr != MAP_FAILED) {
1392 /* errno is stale, don't use */
1393 RTE_LOG(ERR, EAL, "Could not mmap %zu bytes "
1394 "in /dev/zero at [%p], got [%p] - "
1395 "please use '--base-virtaddr' option\n",
1396 mcfg->memseg[s].len,
1397 mcfg->memseg[s].addr, base_addr);
1398 munmap(base_addr, mcfg->memseg[s].len);
1400 RTE_LOG(ERR, EAL, "Could not mmap %zu bytes "
1401 "in /dev/zero at [%p]: '%s'\n",
1402 mcfg->memseg[s].len,
1403 mcfg->memseg[s].addr, strerror(errno));
1405 if (aslr_enabled() > 0) {
1406 RTE_LOG(ERR, EAL, "It is recommended to "
1407 "disable ASLR in the kernel "
1408 "and retry running both primary "
1409 "and secondary processes\n");
1415 size = getFileSize(fd_hugepage);
1416 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1417 if (hp == MAP_FAILED) {
1418 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1422 num_hp = size / sizeof(struct hugepage_file);
1423 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1426 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1427 void *addr, *base_addr;
1428 uintptr_t offset = 0;
1429 size_t mapping_size;
1431 * free previously mapped memory so we can map the
1432 * hugepages into the space
1434 base_addr = mcfg->memseg[s].addr;
1435 munmap(base_addr, mcfg->memseg[s].len);
1437 /* find the hugepages for this segment and map them
1438 * we don't need to worry about order, as the server sorted the
1439 * entries before it did the second mmap of them */
1440 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1441 if (hp[i].memseg_id == (int)s){
1442 fd = open(hp[i].filepath, O_RDWR);
1444 RTE_LOG(ERR, EAL, "Could not open %s\n",
1448 mapping_size = hp[i].size;
1449 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1450 mapping_size, PROT_READ | PROT_WRITE,
1452 close(fd); /* close file both on success and on failure */
1453 if (addr == MAP_FAILED ||
1454 addr != RTE_PTR_ADD(base_addr, offset)) {
1455 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1459 offset+=mapping_size;
1462 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1463 (unsigned long long)mcfg->memseg[s].len);
1466 /* unmap the hugepage config file, since we are done using it */
1473 for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1474 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1475 if (hp != NULL && hp != MAP_FAILED)
1479 if (fd_hugepage >= 0)
1485 rte_eal_using_phys_addrs(void)
1487 return phys_addrs_available;