4 * Copyright(c) 2010-2013 Intel Corporation. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of Intel Corporation nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36 * Copyright(c) 2013 6WIND.
38 * Redistribution and use in source and binary forms, with or without
39 * modification, are permitted provided that the following conditions
42 * * Redistributions of source code must retain the above copyright
43 * notice, this list of conditions and the following disclaimer.
44 * * Redistributions in binary form must reproduce the above copyright
45 * notice, this list of conditions and the following disclaimer in
46 * the documentation and/or other materials provided with the
48 * * Neither the name of 6WIND S.A. nor the names of its
49 * contributors may be used to endorse or promote products derived
50 * from this software without specific prior written permission.
52 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
53 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
54 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
55 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
56 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
57 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
58 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
59 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
60 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
61 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
62 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
74 #include <sys/types.h>
76 #include <sys/queue.h>
81 #include <sys/ioctl.h>
83 #include <sys/resource.h>
86 #include <rte_memory.h>
87 #include <rte_memzone.h>
88 #include <rte_launch.h>
89 #include <rte_tailq.h>
91 #include <rte_eal_memconfig.h>
92 #include <rte_per_lcore.h>
93 #include <rte_lcore.h>
94 #include <rte_common.h>
95 #include <rte_string_fns.h>
97 #include "eal_private.h"
98 #include "eal_internal_cfg.h"
99 #include "eal_filesystem.h"
100 #include "eal_hugepages.h"
104 * Huge page mapping under linux
106 * To reserve a big contiguous amount of memory, we use the hugepage
107 * feature of linux. For that, we need to have hugetlbfs mounted. This
108 * code will create many files in this directory (one per page) and
109 * map them in virtual memory. For each page, we will retrieve its
110 * physical address and remap it in order to have a virtual contiguous
111 * zone as well as a physical contiguous zone.
115 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
118 * Check whether address-space layout randomization is enabled in
119 * the kernel. This is important for multi-process as it can prevent
120 * two processes mapping data to the same virtual address
122 * 0 - address space randomization disabled
123 * 1/2 - address space randomization enabled
124 * negative error code on error
130 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
133 retval = read(fd, &c, 1);
143 default: return -EINVAL;
148 * Increase limit for open files for current process
151 increase_open_file_limit(void)
155 /* read current limits */
156 if (getrlimit(RLIMIT_NOFILE, &limit) != 0) {
157 RTE_LOG(ERR, EAL, "Error reading resource limit: %s\n",
162 /* check if current soft limit matches the hard limit */
163 if (limit.rlim_cur < limit.rlim_max) {
164 /* set soft limit to match hard limit */
165 limit.rlim_cur = limit.rlim_max;
168 /* we can't increase the soft limit so now we try to increase
169 * soft and hard limit. this might fail when run as non-root.
175 /* set current resource limit */
176 if (setrlimit(RLIMIT_NOFILE, &limit) != 0) {
177 RTE_LOG(ERR, EAL, "Error increasing open files limit: %s\n",
186 * Try to mmap *size bytes in /dev/zero. If it is succesful, return the
187 * pointer to the mmap'd area and keep *size unmodified. Else, retry
188 * with a smaller zone: decrease *size by hugepage_sz until it reaches
189 * 0. In this case, return NULL. Note: this function returns an address
190 * which is a multiple of hugepage size.
193 get_virtual_area(size_t *size, size_t hugepage_sz)
199 RTE_LOG(INFO, EAL, "Ask a virtual area of 0x%zu bytes\n", *size);
201 fd = open("/dev/zero", O_RDONLY);
203 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
207 addr = mmap(NULL, (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
208 if (addr == MAP_FAILED)
209 *size -= hugepage_sz;
210 } while (addr == MAP_FAILED && *size > 0);
212 if (addr == MAP_FAILED) {
214 RTE_LOG(INFO, EAL, "Cannot get a virtual area\n");
218 munmap(addr, (*size) + hugepage_sz);
221 /* align addr to a huge page size boundary */
222 aligned_addr = (long)addr;
223 aligned_addr += (hugepage_sz - 1);
224 aligned_addr &= (~(hugepage_sz - 1));
225 addr = (void *)(aligned_addr);
227 RTE_LOG(INFO, EAL, "Virtual area found at %p (size = 0x%zx)\n",
234 * Mmap all hugepages of hugepage table: it first open a file in
235 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
236 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
237 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
238 * map continguous physical blocks in contiguous virtual blocks.
241 map_all_hugepages(struct hugepage *hugepg_tbl,
242 struct hugepage_info *hpi, int orig)
247 void *vma_addr = NULL;
250 for (i = 0; i < hpi->num_pages[0]; i++) {
251 size_t hugepage_sz = hpi->hugepage_sz;
254 hugepg_tbl[i].file_id = i;
255 hugepg_tbl[i].size = hugepage_sz;
256 eal_get_hugefile_path(hugepg_tbl[i].filepath,
257 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
258 hugepg_tbl[i].file_id);
259 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
261 #ifndef RTE_ARCH_X86_64
262 /* for 32-bit systems, don't remap 1G pages, just reuse original
263 * map address as final map address.
265 else if (hugepage_sz == RTE_PGSIZE_1G){
266 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
267 hugepg_tbl[i].orig_va = NULL;
271 else if (vma_len == 0) {
272 unsigned j, num_pages;
274 /* reserve a virtual area for next contiguous
275 * physical block: count the number of
276 * contiguous physical pages. */
277 for (j = i+1; j < hpi->num_pages[0] ; j++) {
278 if (hugepg_tbl[j].physaddr !=
279 hugepg_tbl[j-1].physaddr + hugepage_sz)
283 vma_len = num_pages * hugepage_sz;
285 /* get the biggest virtual memory area up to
286 * vma_len. If it fails, vma_addr is NULL, so
287 * let the kernel provide the address. */
288 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
289 if (vma_addr == NULL)
290 vma_len = hugepage_sz;
293 /* try to create hugepage file */
294 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
296 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
301 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
303 if (virtaddr == MAP_FAILED) {
304 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
311 hugepg_tbl[i].orig_va = virtaddr;
312 memset(virtaddr, 0, hugepage_sz);
315 hugepg_tbl[i].final_va = virtaddr;
318 /* close the file descriptor, files will be locked later */
321 vma_addr = (char *)vma_addr + hugepage_sz;
322 vma_len -= hugepage_sz;
327 /* Unmap all hugepages from original mapping. */
329 unmap_all_hugepages_orig(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
332 for (i = 0; i < hpi->num_pages[0]; i++) {
333 if (hugepg_tbl[i].orig_va) {
334 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
335 hugepg_tbl[i].orig_va = NULL;
342 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
343 * it by browsing the /proc/self/pagemap special file.
346 find_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
351 unsigned long virt_pfn;
354 /* standard page size */
355 page_size = getpagesize();
357 fd = open("/proc/self/pagemap", O_RDONLY);
359 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
360 __func__, strerror(errno));
364 for (i = 0; i < hpi->num_pages[0]; i++) {
366 virt_pfn = (unsigned long)hugepg_tbl[i].orig_va /
368 offset = sizeof(uint64_t) * virt_pfn;
369 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
370 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
371 __func__, strerror(errno));
375 if (read(fd, &page, sizeof(uint64_t)) < 0) {
376 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
377 __func__, strerror(errno));
383 * the pfn (page frame number) are bits 0-54 (see
384 * pagemap.txt in linux Documentation)
386 hugepg_tbl[i].physaddr = ((page & 0x7fffffffffffffULL) * page_size);
393 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
397 find_numasocket(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
401 unsigned i, hp_count = 0;
404 char hugedir_str[PATH_MAX];
407 f = fopen("/proc/self/numa_maps", "r");
409 RTE_LOG(INFO, EAL, "cannot open /proc/self/numa_maps,"
410 " consider that all memory is in socket_id 0\n");
414 rte_snprintf(hugedir_str, sizeof(hugedir_str),
415 "%s/", hpi->hugedir);
418 while (fgets(buf, sizeof(buf), f) != NULL) {
420 /* ignore non huge page */
421 if (strstr(buf, " huge ") == NULL &&
422 strstr(buf, hugedir_str) == NULL)
426 virt_addr = strtoull(buf, &end, 16);
427 if (virt_addr == 0 || end == buf) {
428 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
432 /* get node id (socket id) */
433 nodestr = strstr(buf, " N");
434 if (nodestr == NULL) {
435 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
439 end = strstr(nodestr, "=");
441 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
447 socket_id = strtoul(nodestr, &end, 0);
448 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
449 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
453 /* if we find this page in our mappings, set socket_id */
454 for (i = 0; i < hpi->num_pages[0]; i++) {
455 void *va = (void *)(unsigned long)virt_addr;
456 if (hugepg_tbl[i].orig_va == va) {
457 hugepg_tbl[i].socket_id = socket_id;
463 if (hp_count < hpi->num_pages[0])
475 * Sort the hugepg_tbl by physical address (lower addresses first). We
476 * use a slow algorithm, but we won't have millions of pages, and this
477 * is only done at init time.
480 sort_by_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
484 uint64_t smallest_addr;
487 for (i = 0; i < hpi->num_pages[0]; i++) {
492 * browse all entries starting at 'i', and find the
493 * entry with the smallest addr
495 for (j=i; j< hpi->num_pages[0]; j++) {
497 if (smallest_addr == 0 ||
498 hugepg_tbl[j].physaddr < smallest_addr) {
499 smallest_addr = hugepg_tbl[j].physaddr;
504 /* should not happen */
505 if (smallest_idx == -1) {
506 RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
510 /* swap the 2 entries in the table */
511 memcpy(&tmp, &hugepg_tbl[smallest_idx], sizeof(struct hugepage));
512 memcpy(&hugepg_tbl[smallest_idx], &hugepg_tbl[i],
513 sizeof(struct hugepage));
514 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage));
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)
527 int fd = open(filename, O_CREAT | O_RDWR, 0666);
530 if (ftruncate(fd, mem_size) < 0) {
534 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
540 * this copies *active* hugepages from one hugepage table to another.
541 * destination is typically the shared memory.
544 copy_hugepages_to_shared_mem(struct hugepage * dst, int dest_size,
545 const struct hugepage * src, int src_size)
547 int src_pos, dst_pos = 0;
549 for (src_pos = 0; src_pos < src_size; src_pos++) {
550 if (src[src_pos].final_va != NULL) {
551 /* error on overflow attempt */
552 if (dst_pos == dest_size)
554 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage));
562 * unmaps hugepages that are not going to be used. since we originally allocate
563 * ALL hugepages (not just those we need), additional unmapping needs to be done.
566 unmap_unneeded_hugepages(struct hugepage *hugepg_tbl,
567 struct hugepage_info *hpi,
568 unsigned num_hp_info)
570 unsigned socket, size;
571 int page, nrpages = 0;
574 /* get total number of hugepages */
575 for (size = 0; size < num_hp_info; size++)
576 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
577 nrpages += internal_config.hugepage_info[size].num_pages[socket];
579 for (size = 0; size < num_hp_info; size++) {
580 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
581 unsigned pages_found = 0;
582 /* traverse until we have unmapped all the unused pages */
583 for (page = 0; page < nrpages; page++) {
584 struct hugepage *hp = &hugepg_tbl[page];
586 /* find a page that matches the criteria */
587 if ((hp->size == hpi[size].hugepage_sz) &&
588 (hp->socket_id == (int) socket)) {
590 /* if we skipped enough pages, unmap the rest */
591 if (pages_found == hpi[size].num_pages[socket]) {
592 munmap(hp->final_va, hp->size);
595 /* lock the page and skip */
597 /* try and open the hugepage file */
598 while ((fd = open(hp->filepath, O_CREAT | O_RDWR, 0755)) < 0) {
599 /* if we can't open due to resource limits */
600 if (errno == EMFILE) {
601 RTE_LOG(INFO, EAL, "Increasing open file limit\n");
603 /* if we manage to increase resource limit, try again */
604 if (increase_open_file_limit() == 0)
608 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
612 /* try and lock the hugepage */
613 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
614 RTE_LOG(ERR, EAL, "Locking hugepage file failed!\n");
623 } /* foreach socket */
624 } /* foreach pagesize */
629 static inline uint64_t
630 get_socket_mem_size(int socket)
635 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
636 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
637 if (hpi->hugedir != NULL)
638 size += hpi->hugepage_sz * hpi->num_pages[socket];
645 * This function is a NUMA-aware equivalent of calc_num_pages.
646 * It takes in the list of hugepage sizes and the
647 * number of pages thereof, and calculates the best number of
648 * pages of each size to fulfill the request for <memory> ram
651 calc_num_pages_per_socket(uint64_t * memory,
652 struct hugepage_info *hp_info,
653 struct hugepage_info *hp_used,
654 unsigned num_hp_info)
656 unsigned socket, j, i = 0;
657 unsigned requested, available;
658 int total_num_pages = 0;
659 uint64_t remaining_mem, cur_mem;
660 uint64_t total_mem = internal_config.memory;
662 if (num_hp_info == 0)
665 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
666 /* if specific memory amounts per socket weren't requested */
667 if (internal_config.force_sockets == 0) {
668 /* take whatever is available */
669 memory[socket] = RTE_MIN(get_socket_mem_size(socket),
672 /* skips if the memory on specific socket wasn't requested */
673 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
674 hp_used[i].hugedir = hp_info[i].hugedir;
675 hp_used[i].num_pages[socket] = RTE_MIN(
676 memory[socket] / hp_info[i].hugepage_sz,
677 hp_info[i].num_pages[socket]);
679 cur_mem = hp_used[i].num_pages[socket] *
680 hp_used[i].hugepage_sz;
682 memory[socket] -= cur_mem;
683 total_mem -= cur_mem;
685 total_num_pages += hp_used[i].num_pages[socket];
687 /* check if we have met all memory requests */
688 if (memory[socket] == 0)
691 /* check if we have any more pages left at this size, if so
692 * move on to next size */
693 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
695 /* At this point we know that there are more pages available that are
696 * bigger than the memory we want, so lets see if we can get enough
697 * from other page sizes.
700 for (j = i+1; j < num_hp_info; j++)
701 remaining_mem += hp_info[j].hugepage_sz *
702 hp_info[j].num_pages[socket];
704 /* is there enough other memory, if not allocate another page and quit */
705 if (remaining_mem < memory[socket]){
706 cur_mem = RTE_MIN(memory[socket],
707 hp_info[i].hugepage_sz);
708 memory[socket] -= cur_mem;
709 total_mem -= cur_mem;
710 hp_used[i].num_pages[socket]++;
712 break; /* we are done with this socket*/
715 /* if we didn't satisfy all memory requirements per socket */
716 if (memory[socket] > 0) {
717 /* to prevent icc errors */
718 requested = (unsigned) (internal_config.socket_mem[socket] /
720 available = requested -
721 ((unsigned) (memory[socket] / 0x100000));
722 RTE_LOG(INFO, EAL, "Not enough memory available on socket %u! "
723 "Requested: %uMB, available: %uMB\n", socket,
724 requested, available);
729 /* if we didn't satisfy total memory requirements */
731 requested = (unsigned) (internal_config.memory / 0x100000);
732 available = requested - (unsigned) (total_mem / 0x100000);
733 RTE_LOG(INFO, EAL, "Not enough memory available! Requested: %uMB,"
734 " available: %uMB\n", requested, available);
737 return total_num_pages;
741 * Prepare physical memory mapping: fill configuration structure with
742 * these infos, return 0 on success.
743 * 1. map N huge pages in separate files in hugetlbfs
744 * 2. find associated physical addr
745 * 3. find associated NUMA socket ID
746 * 4. sort all huge pages by physical address
747 * 5. remap these N huge pages in the correct order
748 * 6. unmap the first mapping
749 * 7. fill memsegs in configuration with contiguous zones
752 rte_eal_hugepage_init(void)
754 struct rte_mem_config *mcfg;
755 struct hugepage *hugepage, *tmp_hp = NULL;
756 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
758 uint64_t memory[RTE_MAX_NUMA_NODES];
761 int i, j, new_memseg;
762 int nrpages, total_pages = 0;
765 memset(used_hp, 0, sizeof(used_hp));
767 /* get pointer to global configuration */
768 mcfg = rte_eal_get_configuration()->mem_config;
770 /* for debug purposes, hugetlbfs can be disabled */
771 if (internal_config.no_hugetlbfs) {
772 addr = malloc(internal_config.memory);
773 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
774 mcfg->memseg[0].addr = addr;
775 mcfg->memseg[0].len = internal_config.memory;
776 mcfg->memseg[0].socket_id = 0;
781 /* calculate total number of hugepages available. at this point we haven't
782 * yet started sorting them so they all are on socket 0 */
783 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
784 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
785 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
787 total_pages += internal_config.hugepage_info[i].num_pages[0];
791 * allocate a memory area for hugepage table.
792 * this isn't shared memory yet. due to the fact that we need some
793 * processing done on these pages, shared memory will be created
796 tmp_hp = malloc(total_pages * sizeof(struct hugepage));
800 memset(tmp_hp, 0, total_pages * sizeof(struct hugepage));
802 hp_offset = 0; /* where we start the current page size entries */
804 /* map all hugepages and sort them */
805 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
806 struct hugepage_info *hpi;
809 * we don't yet mark hugepages as used at this stage, so
810 * we just map all hugepages available to the system
811 * all hugepages are still located on socket 0
813 hpi = &internal_config.hugepage_info[i];
815 if (hpi->num_pages == 0)
818 /* map all hugepages available */
819 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
820 RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
821 (unsigned)(hpi->hugepage_sz / 0x100000));
825 /* find physical addresses and sockets for each hugepage */
826 if (find_physaddr(&tmp_hp[hp_offset], hpi) < 0){
827 RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
828 (unsigned)(hpi->hugepage_sz / 0x100000));
832 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
833 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
834 (unsigned)(hpi->hugepage_sz / 0x100000));
838 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
841 /* remap all hugepages */
842 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
843 RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
844 (unsigned)(hpi->hugepage_sz / 0x100000));
848 /* unmap original mappings */
849 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
852 /* we have processed a num of hugepages of this size, so inc offset */
853 hp_offset += hpi->num_pages[0];
856 /* clean out the numbers of pages */
857 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
858 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
859 internal_config.hugepage_info[i].num_pages[j] = 0;
861 /* get hugepages for each socket */
862 for (i = 0; i < total_pages; i++) {
863 int socket = tmp_hp[i].socket_id;
865 /* find a hugepage info with right size and increment num_pages */
866 for (j = 0; j < (int) internal_config.num_hugepage_sizes; j++) {
867 if (tmp_hp[i].size ==
868 internal_config.hugepage_info[j].hugepage_sz) {
869 internal_config.hugepage_info[j].num_pages[socket]++;
874 /* make a copy of socket_mem, needed for number of pages calculation */
875 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
876 memory[i] = internal_config.socket_mem[i];
878 /* calculate final number of pages */
879 nrpages = calc_num_pages_per_socket(memory,
880 internal_config.hugepage_info, used_hp,
881 internal_config.num_hugepage_sizes);
883 /* error if not enough memory available */
888 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
889 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
890 if (used_hp[i].num_pages[j] > 0) {
892 "Requesting %u pages of size %uMB"
894 used_hp[i].num_pages[j],
896 (used_hp[i].hugepage_sz / 0x100000),
902 /* create shared memory */
903 hugepage = create_shared_memory(eal_hugepage_info_path(),
904 nrpages * sizeof(struct hugepage));
906 if (hugepage == NULL) {
907 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
912 * unmap pages that we won't need (looks at used_hp).
913 * also, sets final_va to NULL on pages that were unmapped.
915 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
916 internal_config.num_hugepage_sizes) < 0) {
917 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
922 * copy stuff from malloc'd hugepage* to the actual shared memory.
923 * this procedure only copies those hugepages that have final_va
924 * not NULL. has overflow protection.
926 if (copy_hugepages_to_shared_mem(hugepage, nrpages,
927 tmp_hp, total_pages) < 0) {
928 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
932 /* free the temporary hugepage table */
936 memset(mcfg->memseg, 0, sizeof(mcfg->memseg));
938 for (i = 0; i < nrpages; i++) {
941 /* if this is a new section, create a new memseg */
944 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
946 else if (hugepage[i].size != hugepage[i-1].size)
948 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
951 else if (((unsigned long)hugepage[i].final_va -
952 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
957 if (j == RTE_MAX_MEMSEG)
960 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
961 mcfg->memseg[j].addr = hugepage[i].final_va;
962 mcfg->memseg[j].len = hugepage[i].size;
963 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
964 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
966 /* continuation of previous memseg */
968 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
970 hugepage[i].memseg_id = j;
974 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
975 "from %d requested\n"
976 "Current %s=%d is not enough\n"
977 "Please either increase it or request less amount "
979 i, nrpages, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
995 * uses fstat to report the size of a file on disk
1001 if (fstat(fd, &st) < 0)
1007 * This creates the memory mappings in the secondary process to match that of
1008 * the server process. It goes through each memory segment in the DPDK runtime
1009 * configuration and finds the hugepages which form that segment, mapping them
1010 * in order to form a contiguous block in the virtual memory space
1013 rte_eal_hugepage_attach(void)
1015 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1016 const struct hugepage *hp = NULL;
1017 unsigned num_hp = 0;
1018 unsigned i, s = 0; /* s used to track the segment number */
1020 int fd, fd_zero = -1, fd_hugepage = -1;
1022 if (aslr_enabled() > 0) {
1023 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1024 "(ASLR) is enabled in the kernel.\n");
1025 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1026 "into secondary processes\n");
1029 fd_zero = open("/dev/zero", O_RDONLY);
1031 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1034 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1035 if (fd_hugepage < 0) {
1036 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1040 size = getFileSize(fd_hugepage);
1041 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1043 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1047 num_hp = size / sizeof(struct hugepage);
1048 RTE_LOG(DEBUG, EAL, "Analysing %u hugepages\n", num_hp);
1050 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1051 void *addr, *base_addr;
1052 uintptr_t offset = 0;
1054 /* fdzero is mmapped to get a contiguous block of virtual addresses
1055 * get a block of free memory of the appropriate size -
1056 * use mmap to attempt to get an identical address as server.
1058 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1059 PROT_READ, MAP_PRIVATE, fd_zero, 0);
1060 if (base_addr == MAP_FAILED || base_addr != mcfg->memseg[s].addr) {
1061 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1062 "in /dev/zero to requested address [%p]\n",
1063 (unsigned long long)mcfg->memseg[s].len,
1064 mcfg->memseg[s].addr);
1065 if (aslr_enabled() > 0)
1066 RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel "
1067 "and retry running both primary and secondary processes\n");
1070 /* free memory so we can map the hugepages into the space */
1071 munmap(base_addr, mcfg->memseg[s].len);
1073 /* find the hugepages for this segment and map them
1074 * we don't need to worry about order, as the server sorted the
1075 * entries before it did the second mmap of them */
1076 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1077 if (hp[i].memseg_id == (int)s){
1078 fd = open(hp[i].filepath, O_RDWR);
1080 RTE_LOG(ERR, EAL, "Could not open %s\n",
1084 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1085 hp[i].size, PROT_READ | PROT_WRITE,
1086 MAP_SHARED | MAP_FIXED, fd, 0);
1087 close(fd); /* close file both on success and on failure */
1088 if (addr == MAP_FAILED) {
1089 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1096 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1097 (unsigned long long)mcfg->memseg[s].len);
1107 if (fd_hugepage >= 0)
1113 rte_eal_memdevice_init(void)
1115 struct rte_config *config;
1117 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1120 config = rte_eal_get_configuration();
1121 config->mem_config->nchannel = internal_config.force_nchannel;
1122 config->mem_config->nrank = internal_config.force_nrank;
1128 /* init memory subsystem */
1130 rte_eal_memory_init(void)
1132 RTE_LOG(INFO, EAL, "Setting up hugepage memory...\n");
1133 const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
1134 rte_eal_hugepage_init() :
1135 rte_eal_hugepage_attach();
1139 if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)