4 * Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
5 * Copyright(c) 2013 6WIND.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * * Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * * Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * * Neither the name of Intel Corporation nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
25 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
26 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
27 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
28 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
32 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35 #define _FILE_OFFSET_BITS 64
45 #include <sys/types.h>
47 #include <sys/queue.h>
51 #include <sys/ioctl.h>
55 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
61 #include <rte_memory.h>
62 #include <rte_launch.h>
64 #include <rte_eal_memconfig.h>
65 #include <rte_per_lcore.h>
66 #include <rte_lcore.h>
67 #include <rte_common.h>
68 #include <rte_string_fns.h>
70 #include "eal_private.h"
71 #include "eal_internal_cfg.h"
72 #include "eal_filesystem.h"
73 #include "eal_hugepages.h"
75 #define PFN_MASK_SIZE 8
79 * Huge page mapping under linux
81 * To reserve a big contiguous amount of memory, we use the hugepage
82 * feature of linux. For that, we need to have hugetlbfs mounted. This
83 * code will create many files in this directory (one per page) and
84 * map them in virtual memory. For each page, we will retrieve its
85 * physical address and remap it in order to have a virtual contiguous
86 * zone as well as a physical contiguous zone.
89 static uint64_t baseaddr_offset;
91 static bool phys_addrs_available = true;
93 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
96 test_phys_addrs_available(void)
101 if (!rte_eal_has_hugepages()) {
103 "Started without hugepages support, physical addresses not available\n");
104 phys_addrs_available = false;
108 physaddr = rte_mem_virt2phy(&tmp);
109 if (physaddr == RTE_BAD_PHYS_ADDR) {
111 "Cannot obtain physical addresses: %s. "
112 "Only vfio will function.\n",
114 phys_addrs_available = false;
119 * Get physical address of any mapped virtual address in the current process.
122 rte_mem_virt2phy(const void *virtaddr)
125 uint64_t page, physaddr;
126 unsigned long virt_pfn;
130 if (rte_eal_iova_mode() == RTE_IOVA_VA)
131 return (uintptr_t)virtaddr;
133 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
134 if (!phys_addrs_available)
135 return RTE_BAD_PHYS_ADDR;
137 /* standard page size */
138 page_size = getpagesize();
140 fd = open("/proc/self/pagemap", O_RDONLY);
142 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
143 __func__, strerror(errno));
144 return RTE_BAD_PHYS_ADDR;
147 virt_pfn = (unsigned long)virtaddr / page_size;
148 offset = sizeof(uint64_t) * virt_pfn;
149 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
150 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
151 __func__, strerror(errno));
153 return RTE_BAD_PHYS_ADDR;
156 retval = read(fd, &page, PFN_MASK_SIZE);
159 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
160 __func__, strerror(errno));
161 return RTE_BAD_PHYS_ADDR;
162 } else if (retval != PFN_MASK_SIZE) {
163 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
164 "but expected %d:\n",
165 __func__, retval, PFN_MASK_SIZE);
166 return RTE_BAD_PHYS_ADDR;
170 * the pfn (page frame number) are bits 0-54 (see
171 * pagemap.txt in linux Documentation)
173 if ((page & 0x7fffffffffffffULL) == 0)
174 return RTE_BAD_PHYS_ADDR;
176 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
177 + ((unsigned long)virtaddr % page_size);
183 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
184 * it by browsing the /proc/self/pagemap special file.
187 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
192 for (i = 0; i < hpi->num_pages[0]; i++) {
193 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
194 if (addr == RTE_BAD_PHYS_ADDR)
196 hugepg_tbl[i].physaddr = addr;
202 * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
205 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
208 static phys_addr_t addr;
210 for (i = 0; i < hpi->num_pages[0]; i++) {
211 hugepg_tbl[i].physaddr = addr;
212 addr += hugepg_tbl[i].size;
218 * Check whether address-space layout randomization is enabled in
219 * the kernel. This is important for multi-process as it can prevent
220 * two processes mapping data to the same virtual address
222 * 0 - address space randomization disabled
223 * 1/2 - address space randomization enabled
224 * negative error code on error
230 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
233 retval = read(fd, &c, 1);
243 default: return -EINVAL;
248 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
249 * pointer to the mmap'd area and keep *size unmodified. Else, retry
250 * with a smaller zone: decrease *size by hugepage_sz until it reaches
251 * 0. In this case, return NULL. Note: this function returns an address
252 * which is a multiple of hugepage size.
255 get_virtual_area(size_t *size, size_t hugepage_sz)
261 if (internal_config.base_virtaddr != 0) {
262 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
267 RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
269 fd = open("/dev/zero", O_RDONLY);
271 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
276 (*size) + hugepage_sz, PROT_READ,
277 #ifdef RTE_ARCH_PPC_64
278 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
283 if (addr == MAP_FAILED)
284 *size -= hugepage_sz;
285 } while (addr == MAP_FAILED && *size > 0);
287 if (addr == MAP_FAILED) {
289 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
294 munmap(addr, (*size) + hugepage_sz);
297 /* align addr to a huge page size boundary */
298 aligned_addr = (long)addr;
299 aligned_addr += (hugepage_sz - 1);
300 aligned_addr &= (~(hugepage_sz - 1));
301 addr = (void *)(aligned_addr);
303 RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
306 /* increment offset */
307 baseaddr_offset += *size;
312 static sigjmp_buf huge_jmpenv;
314 static void huge_sigbus_handler(int signo __rte_unused)
316 siglongjmp(huge_jmpenv, 1);
319 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
320 * non-static local variable in the stack frame calling sigsetjmp might be
321 * clobbered by a call to longjmp.
323 static int huge_wrap_sigsetjmp(void)
325 return sigsetjmp(huge_jmpenv, 1);
328 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
329 /* Callback for numa library. */
330 void numa_error(char *where)
332 RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
337 * Mmap all hugepages of hugepage table: it first open a file in
338 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
339 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
340 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
341 * map continguous physical blocks in contiguous virtual blocks.
344 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
345 uint64_t *essential_memory __rte_unused, int orig)
350 void *vma_addr = NULL;
352 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
354 int essential_prev = 0;
356 struct bitmask *oldmask = numa_allocate_nodemask();
357 bool have_numa = true;
358 unsigned long maxnode = 0;
360 /* Check if kernel supports NUMA. */
361 if (numa_available() != 0) {
362 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
366 if (orig && have_numa) {
367 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
368 if (get_mempolicy(&oldpolicy, oldmask->maskp,
369 oldmask->size + 1, 0, 0) < 0) {
371 "Failed to get current mempolicy: %s. "
372 "Assuming MPOL_DEFAULT.\n", strerror(errno));
373 oldpolicy = MPOL_DEFAULT;
375 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
376 if (internal_config.socket_mem[i])
381 for (i = 0; i < hpi->num_pages[0]; i++) {
382 uint64_t hugepage_sz = hpi->hugepage_sz;
384 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
388 for (j = 0; j < maxnode; j++)
389 if (essential_memory[j])
393 node_id = (node_id + 1) % maxnode;
394 while (!internal_config.socket_mem[node_id]) {
401 essential_prev = essential_memory[j];
403 if (essential_memory[j] < hugepage_sz)
404 essential_memory[j] = 0;
406 essential_memory[j] -= hugepage_sz;
410 "Setting policy MPOL_PREFERRED for socket %d\n",
412 numa_set_preferred(node_id);
417 hugepg_tbl[i].file_id = i;
418 hugepg_tbl[i].size = hugepage_sz;
419 eal_get_hugefile_path(hugepg_tbl[i].filepath,
420 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
421 hugepg_tbl[i].file_id);
422 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
425 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
426 * original map address as final map address.
428 else if ((hugepage_sz == RTE_PGSIZE_1G)
429 || (hugepage_sz == RTE_PGSIZE_16G)) {
430 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
431 hugepg_tbl[i].orig_va = NULL;
435 else if (vma_len == 0) {
436 unsigned j, num_pages;
438 /* reserve a virtual area for next contiguous
439 * physical block: count the number of
440 * contiguous physical pages. */
441 for (j = i+1; j < hpi->num_pages[0] ; j++) {
442 #ifdef RTE_ARCH_PPC_64
443 /* The physical addresses are sorted in
444 * descending order on PPC64 */
445 if (hugepg_tbl[j].physaddr !=
446 hugepg_tbl[j-1].physaddr - hugepage_sz)
449 if (hugepg_tbl[j].physaddr !=
450 hugepg_tbl[j-1].physaddr + hugepage_sz)
455 vma_len = num_pages * hugepage_sz;
457 /* get the biggest virtual memory area up to
458 * vma_len. If it fails, vma_addr is NULL, so
459 * let the kernel provide the address. */
460 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
461 if (vma_addr == NULL)
462 vma_len = hugepage_sz;
465 /* try to create hugepage file */
466 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
468 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
473 /* map the segment, and populate page tables,
474 * the kernel fills this segment with zeros */
475 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
476 MAP_SHARED | MAP_POPULATE, fd, 0);
477 if (virtaddr == MAP_FAILED) {
478 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
485 hugepg_tbl[i].orig_va = virtaddr;
488 hugepg_tbl[i].final_va = virtaddr;
492 /* In linux, hugetlb limitations, like cgroup, are
493 * enforced at fault time instead of mmap(), even
494 * with the option of MAP_POPULATE. Kernel will send
495 * a SIGBUS signal. To avoid to be killed, save stack
496 * environment here, if SIGBUS happens, we can jump
499 if (huge_wrap_sigsetjmp()) {
500 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
501 "hugepages of size %u MB\n",
502 (unsigned)(hugepage_sz / 0x100000));
503 munmap(virtaddr, hugepage_sz);
505 unlink(hugepg_tbl[i].filepath);
506 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
508 essential_memory[node_id] =
513 *(int *)virtaddr = 0;
517 /* set shared flock on the file. */
518 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
519 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
520 __func__, strerror(errno));
527 vma_addr = (char *)vma_addr + hugepage_sz;
528 vma_len -= hugepage_sz;
532 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
535 "Restoring previous memory policy: %d\n", oldpolicy);
536 if (oldpolicy == MPOL_DEFAULT) {
537 numa_set_localalloc();
538 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
539 oldmask->size + 1) < 0) {
540 RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
542 numa_set_localalloc();
545 numa_free_cpumask(oldmask);
550 /* Unmap all hugepages from original mapping */
552 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
555 for (i = 0; i < hpi->num_pages[0]; i++) {
556 if (hugepg_tbl[i].orig_va) {
557 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
558 hugepg_tbl[i].orig_va = NULL;
565 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
569 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
573 unsigned i, hp_count = 0;
576 char hugedir_str[PATH_MAX];
579 f = fopen("/proc/self/numa_maps", "r");
581 RTE_LOG(NOTICE, EAL, "NUMA support not available"
582 " consider that all memory is in socket_id 0\n");
586 snprintf(hugedir_str, sizeof(hugedir_str),
587 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
590 while (fgets(buf, sizeof(buf), f) != NULL) {
592 /* ignore non huge page */
593 if (strstr(buf, " huge ") == NULL &&
594 strstr(buf, hugedir_str) == NULL)
598 virt_addr = strtoull(buf, &end, 16);
599 if (virt_addr == 0 || end == buf) {
600 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
604 /* get node id (socket id) */
605 nodestr = strstr(buf, " N");
606 if (nodestr == NULL) {
607 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
611 end = strstr(nodestr, "=");
613 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
619 socket_id = strtoul(nodestr, &end, 0);
620 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
621 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
625 /* if we find this page in our mappings, set socket_id */
626 for (i = 0; i < hpi->num_pages[0]; i++) {
627 void *va = (void *)(unsigned long)virt_addr;
628 if (hugepg_tbl[i].orig_va == va) {
629 hugepg_tbl[i].socket_id = socket_id;
631 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
633 "Hugepage %s is on socket %d\n",
634 hugepg_tbl[i].filepath, socket_id);
640 if (hp_count < hpi->num_pages[0])
652 cmp_physaddr(const void *a, const void *b)
654 #ifndef RTE_ARCH_PPC_64
655 const struct hugepage_file *p1 = a;
656 const struct hugepage_file *p2 = b;
658 /* PowerPC needs memory sorted in reverse order from x86 */
659 const struct hugepage_file *p1 = b;
660 const struct hugepage_file *p2 = a;
662 if (p1->physaddr < p2->physaddr)
664 else if (p1->physaddr > p2->physaddr)
671 * Uses mmap to create a shared memory area for storage of data
672 * Used in this file to store the hugepage file map on disk
675 create_shared_memory(const char *filename, const size_t mem_size)
678 int fd = open(filename, O_CREAT | O_RDWR, 0666);
681 if (ftruncate(fd, mem_size) < 0) {
685 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
687 if (retval == MAP_FAILED)
693 * this copies *active* hugepages from one hugepage table to another.
694 * destination is typically the shared memory.
697 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
698 const struct hugepage_file * src, int src_size)
700 int src_pos, dst_pos = 0;
702 for (src_pos = 0; src_pos < src_size; src_pos++) {
703 if (src[src_pos].final_va != NULL) {
704 /* error on overflow attempt */
705 if (dst_pos == dest_size)
707 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
715 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
716 unsigned num_hp_info)
718 unsigned socket, size;
719 int page, nrpages = 0;
721 /* get total number of hugepages */
722 for (size = 0; size < num_hp_info; size++)
723 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
725 internal_config.hugepage_info[size].num_pages[socket];
727 for (page = 0; page < nrpages; page++) {
728 struct hugepage_file *hp = &hugepg_tbl[page];
730 if (hp->final_va != NULL && unlink(hp->filepath)) {
731 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
732 __func__, hp->filepath, strerror(errno));
739 * unmaps hugepages that are not going to be used. since we originally allocate
740 * ALL hugepages (not just those we need), additional unmapping needs to be done.
743 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
744 struct hugepage_info *hpi,
745 unsigned num_hp_info)
747 unsigned socket, size;
748 int page, nrpages = 0;
750 /* get total number of hugepages */
751 for (size = 0; size < num_hp_info; size++)
752 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
753 nrpages += internal_config.hugepage_info[size].num_pages[socket];
755 for (size = 0; size < num_hp_info; size++) {
756 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
757 unsigned pages_found = 0;
759 /* traverse until we have unmapped all the unused pages */
760 for (page = 0; page < nrpages; page++) {
761 struct hugepage_file *hp = &hugepg_tbl[page];
763 /* find a page that matches the criteria */
764 if ((hp->size == hpi[size].hugepage_sz) &&
765 (hp->socket_id == (int) socket)) {
767 /* if we skipped enough pages, unmap the rest */
768 if (pages_found == hpi[size].num_pages[socket]) {
771 unmap_len = hp->size;
773 /* get start addr and len of the remaining segment */
774 munmap(hp->final_va, (size_t) unmap_len);
777 if (unlink(hp->filepath) == -1) {
778 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
779 __func__, hp->filepath, strerror(errno));
783 /* lock the page and skip */
789 } /* foreach socket */
790 } /* foreach pagesize */
795 static inline uint64_t
796 get_socket_mem_size(int socket)
801 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
802 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
803 if (hpi->hugedir != NULL)
804 size += hpi->hugepage_sz * hpi->num_pages[socket];
811 * This function is a NUMA-aware equivalent of calc_num_pages.
812 * It takes in the list of hugepage sizes and the
813 * number of pages thereof, and calculates the best number of
814 * pages of each size to fulfill the request for <memory> ram
817 calc_num_pages_per_socket(uint64_t * memory,
818 struct hugepage_info *hp_info,
819 struct hugepage_info *hp_used,
820 unsigned num_hp_info)
822 unsigned socket, j, i = 0;
823 unsigned requested, available;
824 int total_num_pages = 0;
825 uint64_t remaining_mem, cur_mem;
826 uint64_t total_mem = internal_config.memory;
828 if (num_hp_info == 0)
831 /* if specific memory amounts per socket weren't requested */
832 if (internal_config.force_sockets == 0) {
833 int cpu_per_socket[RTE_MAX_NUMA_NODES];
834 size_t default_size, total_size;
837 /* Compute number of cores per socket */
838 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
839 RTE_LCORE_FOREACH(lcore_id) {
840 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
844 * Automatically spread requested memory amongst detected sockets according
845 * to number of cores from cpu mask present on each socket
847 total_size = internal_config.memory;
848 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
850 /* Set memory amount per socket */
851 default_size = (internal_config.memory * cpu_per_socket[socket])
854 /* Limit to maximum available memory on socket */
855 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
858 memory[socket] = default_size;
859 total_size -= default_size;
863 * If some memory is remaining, try to allocate it by getting all
864 * available memory from sockets, one after the other
866 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
867 /* take whatever is available */
868 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
872 memory[socket] += default_size;
873 total_size -= default_size;
877 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
878 /* skips if the memory on specific socket wasn't requested */
879 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
880 hp_used[i].hugedir = hp_info[i].hugedir;
881 hp_used[i].num_pages[socket] = RTE_MIN(
882 memory[socket] / hp_info[i].hugepage_sz,
883 hp_info[i].num_pages[socket]);
885 cur_mem = hp_used[i].num_pages[socket] *
886 hp_used[i].hugepage_sz;
888 memory[socket] -= cur_mem;
889 total_mem -= cur_mem;
891 total_num_pages += hp_used[i].num_pages[socket];
893 /* check if we have met all memory requests */
894 if (memory[socket] == 0)
897 /* check if we have any more pages left at this size, if so
898 * move on to next size */
899 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
901 /* At this point we know that there are more pages available that are
902 * bigger than the memory we want, so lets see if we can get enough
903 * from other page sizes.
906 for (j = i+1; j < num_hp_info; j++)
907 remaining_mem += hp_info[j].hugepage_sz *
908 hp_info[j].num_pages[socket];
910 /* is there enough other memory, if not allocate another page and quit */
911 if (remaining_mem < memory[socket]){
912 cur_mem = RTE_MIN(memory[socket],
913 hp_info[i].hugepage_sz);
914 memory[socket] -= cur_mem;
915 total_mem -= cur_mem;
916 hp_used[i].num_pages[socket]++;
918 break; /* we are done with this socket*/
921 /* if we didn't satisfy all memory requirements per socket */
922 if (memory[socket] > 0) {
923 /* to prevent icc errors */
924 requested = (unsigned) (internal_config.socket_mem[socket] /
926 available = requested -
927 ((unsigned) (memory[socket] / 0x100000));
928 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
929 "Requested: %uMB, available: %uMB\n", socket,
930 requested, available);
935 /* if we didn't satisfy total memory requirements */
937 requested = (unsigned) (internal_config.memory / 0x100000);
938 available = requested - (unsigned) (total_mem / 0x100000);
939 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
940 " available: %uMB\n", requested, available);
943 return total_num_pages;
947 eal_get_hugepage_mem_size(void)
952 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
953 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
954 if (hpi->hugedir != NULL) {
955 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
956 size += hpi->hugepage_sz * hpi->num_pages[j];
961 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
964 static struct sigaction huge_action_old;
965 static int huge_need_recover;
968 huge_register_sigbus(void)
971 struct sigaction action;
974 sigaddset(&mask, SIGBUS);
976 action.sa_mask = mask;
977 action.sa_handler = huge_sigbus_handler;
979 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
983 huge_recover_sigbus(void)
985 if (huge_need_recover) {
986 sigaction(SIGBUS, &huge_action_old, NULL);
987 huge_need_recover = 0;
992 * Prepare physical memory mapping: fill configuration structure with
993 * these infos, return 0 on success.
994 * 1. map N huge pages in separate files in hugetlbfs
995 * 2. find associated physical addr
996 * 3. find associated NUMA socket ID
997 * 4. sort all huge pages by physical address
998 * 5. remap these N huge pages in the correct order
999 * 6. unmap the first mapping
1000 * 7. fill memsegs in configuration with contiguous zones
1003 rte_eal_hugepage_init(void)
1005 struct rte_mem_config *mcfg;
1006 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1007 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1009 uint64_t memory[RTE_MAX_NUMA_NODES];
1012 int i, j, new_memseg;
1013 int nr_hugefiles, nr_hugepages = 0;
1016 test_phys_addrs_available();
1018 memset(used_hp, 0, sizeof(used_hp));
1020 /* get pointer to global configuration */
1021 mcfg = rte_eal_get_configuration()->mem_config;
1023 /* hugetlbfs can be disabled */
1024 if (internal_config.no_hugetlbfs) {
1025 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1026 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1027 if (addr == MAP_FAILED) {
1028 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1032 if (rte_eal_iova_mode() == RTE_IOVA_VA)
1033 mcfg->memseg[0].phys_addr = (uintptr_t)addr;
1035 mcfg->memseg[0].phys_addr = RTE_BAD_PHYS_ADDR;
1036 mcfg->memseg[0].addr = addr;
1037 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1038 mcfg->memseg[0].len = internal_config.memory;
1039 mcfg->memseg[0].socket_id = 0;
1043 /* calculate total number of hugepages available. at this point we haven't
1044 * yet started sorting them so they all are on socket 0 */
1045 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1046 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1047 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1049 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1053 * allocate a memory area for hugepage table.
1054 * this isn't shared memory yet. due to the fact that we need some
1055 * processing done on these pages, shared memory will be created
1058 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1062 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1064 hp_offset = 0; /* where we start the current page size entries */
1066 huge_register_sigbus();
1068 /* make a copy of socket_mem, needed for balanced allocation. */
1069 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1070 memory[i] = internal_config.socket_mem[i];
1073 /* map all hugepages and sort them */
1074 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1075 unsigned pages_old, pages_new;
1076 struct hugepage_info *hpi;
1079 * we don't yet mark hugepages as used at this stage, so
1080 * we just map all hugepages available to the system
1081 * all hugepages are still located on socket 0
1083 hpi = &internal_config.hugepage_info[i];
1085 if (hpi->num_pages[0] == 0)
1088 /* map all hugepages available */
1089 pages_old = hpi->num_pages[0];
1090 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi,
1092 if (pages_new < pages_old) {
1094 "%d not %d hugepages of size %u MB allocated\n",
1095 pages_new, pages_old,
1096 (unsigned)(hpi->hugepage_sz / 0x100000));
1098 int pages = pages_old - pages_new;
1100 nr_hugepages -= pages;
1101 hpi->num_pages[0] = pages_new;
1106 if (phys_addrs_available) {
1107 /* find physical addresses for each hugepage */
1108 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1109 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1110 "for %u MB pages\n",
1111 (unsigned int)(hpi->hugepage_sz / 0x100000));
1115 /* set physical addresses for each hugepage */
1116 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1117 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1118 "for %u MB pages\n",
1119 (unsigned int)(hpi->hugepage_sz / 0x100000));
1124 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1125 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1126 (unsigned)(hpi->hugepage_sz / 0x100000));
1130 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1131 sizeof(struct hugepage_file), cmp_physaddr);
1133 /* remap all hugepages */
1134 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, NULL, 0) !=
1135 hpi->num_pages[0]) {
1136 RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1137 (unsigned)(hpi->hugepage_sz / 0x100000));
1141 /* unmap original mappings */
1142 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1145 /* we have processed a num of hugepages of this size, so inc offset */
1146 hp_offset += hpi->num_pages[0];
1149 huge_recover_sigbus();
1151 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1152 internal_config.memory = eal_get_hugepage_mem_size();
1154 nr_hugefiles = nr_hugepages;
1157 /* clean out the numbers of pages */
1158 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1159 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1160 internal_config.hugepage_info[i].num_pages[j] = 0;
1162 /* get hugepages for each socket */
1163 for (i = 0; i < nr_hugefiles; i++) {
1164 int socket = tmp_hp[i].socket_id;
1166 /* find a hugepage info with right size and increment num_pages */
1167 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1168 (int)internal_config.num_hugepage_sizes);
1169 for (j = 0; j < nb_hpsizes; j++) {
1170 if (tmp_hp[i].size ==
1171 internal_config.hugepage_info[j].hugepage_sz) {
1172 internal_config.hugepage_info[j].num_pages[socket]++;
1177 /* make a copy of socket_mem, needed for number of pages calculation */
1178 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1179 memory[i] = internal_config.socket_mem[i];
1181 /* calculate final number of pages */
1182 nr_hugepages = calc_num_pages_per_socket(memory,
1183 internal_config.hugepage_info, used_hp,
1184 internal_config.num_hugepage_sizes);
1186 /* error if not enough memory available */
1187 if (nr_hugepages < 0)
1191 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1192 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1193 if (used_hp[i].num_pages[j] > 0) {
1195 "Requesting %u pages of size %uMB"
1196 " from socket %i\n",
1197 used_hp[i].num_pages[j],
1199 (used_hp[i].hugepage_sz / 0x100000),
1205 /* create shared memory */
1206 hugepage = create_shared_memory(eal_hugepage_info_path(),
1207 nr_hugefiles * sizeof(struct hugepage_file));
1209 if (hugepage == NULL) {
1210 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1213 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1216 * unmap pages that we won't need (looks at used_hp).
1217 * also, sets final_va to NULL on pages that were unmapped.
1219 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1220 internal_config.num_hugepage_sizes) < 0) {
1221 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1226 * copy stuff from malloc'd hugepage* to the actual shared memory.
1227 * this procedure only copies those hugepages that have final_va
1228 * not NULL. has overflow protection.
1230 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1231 tmp_hp, nr_hugefiles) < 0) {
1232 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1236 /* free the hugepage backing files */
1237 if (internal_config.hugepage_unlink &&
1238 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1239 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1243 /* free the temporary hugepage table */
1247 /* first memseg index shall be 0 after incrementing it below */
1249 for (i = 0; i < nr_hugefiles; i++) {
1252 /* if this is a new section, create a new memseg */
1255 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1257 else if (hugepage[i].size != hugepage[i-1].size)
1260 #ifdef RTE_ARCH_PPC_64
1261 /* On PPC64 architecture, the mmap always start from higher
1262 * virtual address to lower address. Here, both the physical
1263 * address and virtual address are in descending order */
1264 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1267 else if (((unsigned long)hugepage[i-1].final_va -
1268 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1271 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1274 else if (((unsigned long)hugepage[i].final_va -
1275 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1281 if (j == RTE_MAX_MEMSEG)
1284 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1285 mcfg->memseg[j].addr = hugepage[i].final_va;
1286 mcfg->memseg[j].len = hugepage[i].size;
1287 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1288 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1290 /* continuation of previous memseg */
1292 #ifdef RTE_ARCH_PPC_64
1293 /* Use the phy and virt address of the last page as segment
1294 * address for IBM Power architecture */
1295 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1296 mcfg->memseg[j].addr = hugepage[i].final_va;
1298 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1300 hugepage[i].memseg_id = j;
1303 if (i < nr_hugefiles) {
1304 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1305 "from %d requested\n"
1306 "Current %s=%d is not enough\n"
1307 "Please either increase it or request less amount "
1309 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1314 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1319 huge_recover_sigbus();
1321 if (hugepage != NULL)
1322 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1328 * uses fstat to report the size of a file on disk
1334 if (fstat(fd, &st) < 0)
1340 * This creates the memory mappings in the secondary process to match that of
1341 * the server process. It goes through each memory segment in the DPDK runtime
1342 * configuration and finds the hugepages which form that segment, mapping them
1343 * in order to form a contiguous block in the virtual memory space
1346 rte_eal_hugepage_attach(void)
1348 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1349 struct hugepage_file *hp = NULL;
1350 unsigned num_hp = 0;
1351 unsigned i, s = 0; /* s used to track the segment number */
1352 unsigned max_seg = RTE_MAX_MEMSEG;
1354 int fd, fd_zero = -1, fd_hugepage = -1;
1356 if (aslr_enabled() > 0) {
1357 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1358 "(ASLR) is enabled in the kernel.\n");
1359 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1360 "into secondary processes\n");
1363 test_phys_addrs_available();
1365 fd_zero = open("/dev/zero", O_RDONLY);
1367 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1370 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1371 if (fd_hugepage < 0) {
1372 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1376 /* map all segments into memory to make sure we get the addrs */
1377 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1381 * the first memory segment with len==0 is the one that
1382 * follows the last valid segment.
1384 if (mcfg->memseg[s].len == 0)
1388 * fdzero is mmapped to get a contiguous block of virtual
1389 * addresses of the appropriate memseg size.
1390 * use mmap to get identical addresses as the primary process.
1392 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1394 #ifdef RTE_ARCH_PPC_64
1395 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1400 if (base_addr == MAP_FAILED ||
1401 base_addr != mcfg->memseg[s].addr) {
1403 if (base_addr != MAP_FAILED) {
1404 /* errno is stale, don't use */
1405 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1406 "in /dev/zero at [%p], got [%p] - "
1407 "please use '--base-virtaddr' option\n",
1408 (unsigned long long)mcfg->memseg[s].len,
1409 mcfg->memseg[s].addr, base_addr);
1410 munmap(base_addr, mcfg->memseg[s].len);
1412 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1413 "in /dev/zero at [%p]: '%s'\n",
1414 (unsigned long long)mcfg->memseg[s].len,
1415 mcfg->memseg[s].addr, strerror(errno));
1417 if (aslr_enabled() > 0) {
1418 RTE_LOG(ERR, EAL, "It is recommended to "
1419 "disable ASLR in the kernel "
1420 "and retry running both primary "
1421 "and secondary processes\n");
1427 size = getFileSize(fd_hugepage);
1428 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1429 if (hp == MAP_FAILED) {
1430 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1434 num_hp = size / sizeof(struct hugepage_file);
1435 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1438 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1439 void *addr, *base_addr;
1440 uintptr_t offset = 0;
1441 size_t mapping_size;
1443 * free previously mapped memory so we can map the
1444 * hugepages into the space
1446 base_addr = mcfg->memseg[s].addr;
1447 munmap(base_addr, mcfg->memseg[s].len);
1449 /* find the hugepages for this segment and map them
1450 * we don't need to worry about order, as the server sorted the
1451 * entries before it did the second mmap of them */
1452 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1453 if (hp[i].memseg_id == (int)s){
1454 fd = open(hp[i].filepath, O_RDWR);
1456 RTE_LOG(ERR, EAL, "Could not open %s\n",
1460 mapping_size = hp[i].size;
1461 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1462 mapping_size, PROT_READ | PROT_WRITE,
1464 close(fd); /* close file both on success and on failure */
1465 if (addr == MAP_FAILED ||
1466 addr != RTE_PTR_ADD(base_addr, offset)) {
1467 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1471 offset+=mapping_size;
1474 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1475 (unsigned long long)mcfg->memseg[s].len);
1478 /* unmap the hugepage config file, since we are done using it */
1485 for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1486 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1487 if (hp != NULL && hp != MAP_FAILED)
1491 if (fd_hugepage >= 0)
1497 rte_eal_using_phys_addrs(void)
1499 return phys_addrs_available;