4 * Copyright(c) 2010-2014 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.
35 * Copyright(c) 2013 6WIND.
37 * Redistribution and use in source and binary forms, with or without
38 * modification, are permitted provided that the following conditions
41 * * Redistributions of source code must retain the above copyright
42 * notice, this list of conditions and the following disclaimer.
43 * * Redistributions in binary form must reproduce the above copyright
44 * notice, this list of conditions and the following disclaimer in
45 * the documentation and/or other materials provided with the
47 * * Neither the name of 6WIND S.A. nor the names of its
48 * contributors may be used to endorse or promote products derived
49 * from this software without specific prior written permission.
51 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
52 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
53 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
54 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
55 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
56 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
57 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
58 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
59 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
60 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
61 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
64 #define _FILE_OFFSET_BITS 64
74 #include <sys/types.h>
76 #include <sys/queue.h>
81 #include <sys/ioctl.h>
85 #include <rte_memory.h>
86 #include <rte_memzone.h>
87 #include <rte_launch.h>
89 #include <rte_eal_memconfig.h>
90 #include <rte_per_lcore.h>
91 #include <rte_lcore.h>
92 #include <rte_common.h>
93 #include <rte_string_fns.h>
95 #include "eal_private.h"
96 #include "eal_internal_cfg.h"
97 #include "eal_filesystem.h"
98 #include "eal_hugepages.h"
102 * Huge page mapping under linux
104 * To reserve a big contiguous amount of memory, we use the hugepage
105 * feature of linux. For that, we need to have hugetlbfs mounted. This
106 * code will create many files in this directory (one per page) and
107 * map them in virtual memory. For each page, we will retrieve its
108 * physical address and remap it in order to have a virtual contiguous
109 * zone as well as a physical contiguous zone.
112 static uint64_t baseaddr_offset;
114 static unsigned proc_pagemap_readable;
116 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
118 /* Lock page in physical memory and prevent from swapping. */
120 rte_mem_lock_page(const void *virt)
122 unsigned long virtual = (unsigned long)virt;
123 int page_size = getpagesize();
124 unsigned long aligned = (virtual & ~ (page_size - 1));
125 return mlock((void*)aligned, page_size);
129 * Get physical address of any mapped virtual address in the current process.
132 rte_mem_virt2phy(const void *virtaddr)
135 uint64_t page, physaddr;
136 unsigned long virt_pfn;
140 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
141 if (!proc_pagemap_readable)
142 return RTE_BAD_PHYS_ADDR;
144 /* standard page size */
145 page_size = getpagesize();
147 fd = open("/proc/self/pagemap", O_RDONLY);
149 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
150 __func__, strerror(errno));
151 return RTE_BAD_PHYS_ADDR;
154 virt_pfn = (unsigned long)virtaddr / page_size;
155 offset = sizeof(uint64_t) * virt_pfn;
156 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
157 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
158 __func__, strerror(errno));
160 return RTE_BAD_PHYS_ADDR;
162 if (read(fd, &page, sizeof(uint64_t)) < 0) {
163 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
164 __func__, strerror(errno));
166 return RTE_BAD_PHYS_ADDR;
170 * the pfn (page frame number) are bits 0-54 (see
171 * pagemap.txt in linux Documentation)
173 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
174 + ((unsigned long)virtaddr % page_size);
180 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
181 * it by browsing the /proc/self/pagemap special file.
184 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
189 for (i = 0; i < hpi->num_pages[0]; i++) {
190 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
191 if (addr == RTE_BAD_PHYS_ADDR)
193 hugepg_tbl[i].physaddr = addr;
199 * Check whether address-space layout randomization is enabled in
200 * the kernel. This is important for multi-process as it can prevent
201 * two processes mapping data to the same virtual address
203 * 0 - address space randomization disabled
204 * 1/2 - address space randomization enabled
205 * negative error code on error
211 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
214 retval = read(fd, &c, 1);
224 default: return -EINVAL;
229 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
230 * pointer to the mmap'd area and keep *size unmodified. Else, retry
231 * with a smaller zone: decrease *size by hugepage_sz until it reaches
232 * 0. In this case, return NULL. Note: this function returns an address
233 * which is a multiple of hugepage size.
236 get_virtual_area(size_t *size, size_t hugepage_sz)
242 if (internal_config.base_virtaddr != 0) {
243 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
248 RTE_LOG(INFO, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
250 fd = open("/dev/zero", O_RDONLY);
252 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
257 (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
258 if (addr == MAP_FAILED)
259 *size -= hugepage_sz;
260 } while (addr == MAP_FAILED && *size > 0);
262 if (addr == MAP_FAILED) {
264 RTE_LOG(INFO, EAL, "Cannot get a virtual area\n");
268 munmap(addr, (*size) + hugepage_sz);
271 /* align addr to a huge page size boundary */
272 aligned_addr = (long)addr;
273 aligned_addr += (hugepage_sz - 1);
274 aligned_addr &= (~(hugepage_sz - 1));
275 addr = (void *)(aligned_addr);
277 RTE_LOG(INFO, EAL, "Virtual area found at %p (size = 0x%zx)\n",
280 /* increment offset */
281 baseaddr_offset += *size;
287 * Mmap all hugepages of hugepage table: it first open a file in
288 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
289 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
290 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
291 * map continguous physical blocks in contiguous virtual blocks.
294 map_all_hugepages(struct hugepage_file *hugepg_tbl,
295 struct hugepage_info *hpi, int orig)
300 void *vma_addr = NULL;
303 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
304 RTE_SET_USED(vma_len);
307 for (i = 0; i < hpi->num_pages[0]; i++) {
308 uint64_t hugepage_sz = hpi->hugepage_sz;
311 hugepg_tbl[i].file_id = i;
312 hugepg_tbl[i].size = hugepage_sz;
313 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
314 eal_get_hugefile_temp_path(hugepg_tbl[i].filepath,
315 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
316 hugepg_tbl[i].file_id);
318 eal_get_hugefile_path(hugepg_tbl[i].filepath,
319 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
320 hugepg_tbl[i].file_id);
322 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
325 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
326 * original map address as final map address.
328 else if ((hugepage_sz == RTE_PGSIZE_1G)
329 || (hugepage_sz == RTE_PGSIZE_16G)) {
330 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
331 hugepg_tbl[i].orig_va = NULL;
336 #ifndef RTE_EAL_SINGLE_FILE_SEGMENTS
337 else if (vma_len == 0) {
338 unsigned j, num_pages;
340 /* reserve a virtual area for next contiguous
341 * physical block: count the number of
342 * contiguous physical pages. */
343 for (j = i+1; j < hpi->num_pages[0] ; j++) {
344 #ifdef RTE_ARCH_PPC_64
345 /* The physical addresses are sorted in
346 * descending order on PPC64 */
347 if (hugepg_tbl[j].physaddr !=
348 hugepg_tbl[j-1].physaddr - hugepage_sz)
351 if (hugepg_tbl[j].physaddr !=
352 hugepg_tbl[j-1].physaddr + hugepage_sz)
357 vma_len = num_pages * hugepage_sz;
359 /* get the biggest virtual memory area up to
360 * vma_len. If it fails, vma_addr is NULL, so
361 * let the kernel provide the address. */
362 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
363 if (vma_addr == NULL)
364 vma_len = hugepage_sz;
368 /* try to create hugepage file */
369 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
371 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
376 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
378 if (virtaddr == MAP_FAILED) {
379 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
386 hugepg_tbl[i].orig_va = virtaddr;
387 memset(virtaddr, 0, hugepage_sz);
390 hugepg_tbl[i].final_va = virtaddr;
393 /* set shared flock on the file. */
394 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
395 RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
396 __func__, strerror(errno));
403 vma_addr = (char *)vma_addr + hugepage_sz;
404 vma_len -= hugepage_sz;
409 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
412 * Remaps all hugepages into single file segments
415 remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
418 unsigned i = 0, j, num_pages, page_idx = 0;
419 void *vma_addr = NULL, *old_addr = NULL, *page_addr = NULL;
421 size_t hugepage_sz = hpi->hugepage_sz;
422 size_t total_size, offset;
423 char filepath[MAX_HUGEPAGE_PATH];
424 phys_addr_t physaddr;
427 while (i < hpi->num_pages[0]) {
430 /* for 32-bit systems, don't remap 1G pages and 16G pages,
431 * just reuse original map address as final map address.
433 if ((hugepage_sz == RTE_PGSIZE_1G)
434 || (hugepage_sz == RTE_PGSIZE_16G)) {
435 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
436 hugepg_tbl[i].orig_va = NULL;
442 /* reserve a virtual area for next contiguous
443 * physical block: count the number of
444 * contiguous physical pages. */
445 for (j = i+1; j < hpi->num_pages[0] ; j++) {
446 #ifdef RTE_ARCH_PPC_64
447 /* The physical addresses are sorted in descending
449 if (hugepg_tbl[j].physaddr !=
450 hugepg_tbl[j-1].physaddr - hugepage_sz)
453 if (hugepg_tbl[j].physaddr !=
454 hugepg_tbl[j-1].physaddr + hugepage_sz)
459 vma_len = num_pages * hugepage_sz;
461 socket = hugepg_tbl[i].socket_id;
463 /* get the biggest virtual memory area up to
464 * vma_len. If it fails, vma_addr is NULL, so
465 * let the kernel provide the address. */
466 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
468 /* If we can't find a big enough virtual area, work out how many pages
469 * we are going to get */
470 if (vma_addr == NULL)
472 else if (vma_len != num_pages * hugepage_sz) {
473 num_pages = vma_len / hugepage_sz;
478 hugepg_tbl[page_idx].file_id = page_idx;
479 eal_get_hugefile_path(filepath,
482 hugepg_tbl[page_idx].file_id);
484 /* try to create hugepage file */
485 fd = open(filepath, O_CREAT | O_RDWR, 0755);
487 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__, strerror(errno));
494 /* unmap current segment */
496 munmap(vma_addr, total_size);
498 /* unmap original page */
499 munmap(hugepg_tbl[i].orig_va, hugepage_sz);
500 unlink(hugepg_tbl[i].filepath);
502 total_size += hugepage_sz;
506 /* map new, bigger segment */
507 vma_addr = mmap(vma_addr, total_size,
508 PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
510 if (vma_addr == MAP_FAILED || vma_addr != old_addr) {
511 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__, strerror(errno));
516 /* touch the page. this is needed because kernel postpones mapping
517 * creation until the first page fault. with this, we pin down
518 * the page and it is marked as used and gets into process' pagemap.
520 for (offset = 0; offset < total_size; offset += hugepage_sz)
521 *((volatile uint8_t*) RTE_PTR_ADD(vma_addr, offset));
524 /* set shared flock on the file. */
525 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
526 RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
527 __func__, strerror(errno));
532 snprintf(hugepg_tbl[page_idx].filepath, MAX_HUGEPAGE_PATH, "%s",
535 physaddr = rte_mem_virt2phy(vma_addr);
537 if (physaddr == RTE_BAD_PHYS_ADDR)
540 hugepg_tbl[page_idx].final_va = vma_addr;
542 hugepg_tbl[page_idx].physaddr = physaddr;
544 hugepg_tbl[page_idx].repeated = num_pages;
546 hugepg_tbl[page_idx].socket_id = socket;
550 /* verify the memory segment - that is, check that every VA corresponds
551 * to the physical address we expect to see
553 for (offset = 0; offset < vma_len; offset += hugepage_sz) {
554 uint64_t expected_physaddr;
556 expected_physaddr = hugepg_tbl[page_idx].physaddr + offset;
557 page_addr = RTE_PTR_ADD(vma_addr, offset);
558 physaddr = rte_mem_virt2phy(page_addr);
560 if (physaddr != expected_physaddr) {
561 RTE_LOG(ERR, EAL, "Segment sanity check failed: wrong physaddr "
562 "at %p (offset 0x%" PRIx64 ": 0x%" PRIx64
563 " (expected 0x%" PRIx64 ")\n",
564 page_addr, offset, physaddr, expected_physaddr);
569 /* zero out the whole segment */
570 memset(hugepg_tbl[page_idx].final_va, 0, total_size);
575 /* zero out the rest */
576 memset(&hugepg_tbl[page_idx], 0, (hpi->num_pages[0] - page_idx) * sizeof(struct hugepage_file));
579 #else/* RTE_EAL_SINGLE_FILE_SEGMENTS=n */
581 /* Unmap all hugepages from original mapping */
583 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
586 for (i = 0; i < hpi->num_pages[0]; i++) {
587 if (hugepg_tbl[i].orig_va) {
588 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
589 hugepg_tbl[i].orig_va = NULL;
594 #endif /* RTE_EAL_SINGLE_FILE_SEGMENTS */
597 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
601 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
605 unsigned i, hp_count = 0;
608 char hugedir_str[PATH_MAX];
611 f = fopen("/proc/self/numa_maps", "r");
613 RTE_LOG(INFO, EAL, "cannot open /proc/self/numa_maps,"
614 " consider that all memory is in socket_id 0\n");
618 snprintf(hugedir_str, sizeof(hugedir_str),
619 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
622 while (fgets(buf, sizeof(buf), f) != NULL) {
624 /* ignore non huge page */
625 if (strstr(buf, " huge ") == NULL &&
626 strstr(buf, hugedir_str) == NULL)
630 virt_addr = strtoull(buf, &end, 16);
631 if (virt_addr == 0 || end == buf) {
632 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
636 /* get node id (socket id) */
637 nodestr = strstr(buf, " N");
638 if (nodestr == NULL) {
639 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
643 end = strstr(nodestr, "=");
645 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
651 socket_id = strtoul(nodestr, &end, 0);
652 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
653 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
657 /* if we find this page in our mappings, set socket_id */
658 for (i = 0; i < hpi->num_pages[0]; i++) {
659 void *va = (void *)(unsigned long)virt_addr;
660 if (hugepg_tbl[i].orig_va == va) {
661 hugepg_tbl[i].socket_id = socket_id;
667 if (hp_count < hpi->num_pages[0])
679 * Sort the hugepg_tbl by physical address (lower addresses first on x86,
680 * higher address first on powerpc). We use a slow algorithm, but we won't
681 * have millions of pages, and this is only done at init time.
684 sort_by_physaddr(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
688 uint64_t compare_addr;
689 struct hugepage_file tmp;
691 for (i = 0; i < hpi->num_pages[0]; i++) {
696 * browse all entries starting at 'i', and find the
697 * entry with the smallest addr
699 for (j=i; j< hpi->num_pages[0]; j++) {
701 if (compare_addr == 0 ||
702 #ifdef RTE_ARCH_PPC_64
703 hugepg_tbl[j].physaddr > compare_addr) {
705 hugepg_tbl[j].physaddr < compare_addr) {
707 compare_addr = hugepg_tbl[j].physaddr;
712 /* should not happen */
713 if (compare_idx == -1) {
714 RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
718 /* swap the 2 entries in the table */
719 memcpy(&tmp, &hugepg_tbl[compare_idx],
720 sizeof(struct hugepage_file));
721 memcpy(&hugepg_tbl[compare_idx], &hugepg_tbl[i],
722 sizeof(struct hugepage_file));
723 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage_file));
729 * Uses mmap to create a shared memory area for storage of data
730 * Used in this file to store the hugepage file map on disk
733 create_shared_memory(const char *filename, const size_t mem_size)
736 int fd = open(filename, O_CREAT | O_RDWR, 0666);
739 if (ftruncate(fd, mem_size) < 0) {
743 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
749 * this copies *active* hugepages from one hugepage table to another.
750 * destination is typically the shared memory.
753 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
754 const struct hugepage_file * src, int src_size)
756 int src_pos, dst_pos = 0;
758 for (src_pos = 0; src_pos < src_size; src_pos++) {
759 if (src[src_pos].final_va != NULL) {
760 /* error on overflow attempt */
761 if (dst_pos == dest_size)
763 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
771 * unmaps hugepages that are not going to be used. since we originally allocate
772 * ALL hugepages (not just those we need), additional unmapping needs to be done.
775 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
776 struct hugepage_info *hpi,
777 unsigned num_hp_info)
779 unsigned socket, size;
780 int page, nrpages = 0;
782 /* get total number of hugepages */
783 for (size = 0; size < num_hp_info; size++)
784 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
785 nrpages += internal_config.hugepage_info[size].num_pages[socket];
787 for (size = 0; size < num_hp_info; size++) {
788 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
789 unsigned pages_found = 0;
791 /* traverse until we have unmapped all the unused pages */
792 for (page = 0; page < nrpages; page++) {
793 struct hugepage_file *hp = &hugepg_tbl[page];
795 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
796 /* if this page was already cleared */
797 if (hp->final_va == NULL)
801 /* find a page that matches the criteria */
802 if ((hp->size == hpi[size].hugepage_sz) &&
803 (hp->socket_id == (int) socket)) {
805 /* if we skipped enough pages, unmap the rest */
806 if (pages_found == hpi[size].num_pages[socket]) {
809 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
810 unmap_len = hp->size * hp->repeated;
812 unmap_len = hp->size;
815 /* get start addr and len of the remaining segment */
816 munmap(hp->final_va, (size_t) unmap_len);
819 if (unlink(hp->filepath) == -1) {
820 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
821 __func__, hp->filepath, strerror(errno));
825 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
826 /* else, check how much do we need to map */
829 hpi[size].num_pages[socket] - pages_found;
831 /* if we need enough memory to fit into the segment */
832 if (hp->repeated <= nr_pg_left) {
833 pages_found += hp->repeated;
835 /* truncate the segment */
837 uint64_t final_size = nr_pg_left * hp->size;
838 uint64_t seg_size = hp->repeated * hp->size;
840 void * unmap_va = RTE_PTR_ADD(hp->final_va,
844 munmap(unmap_va, seg_size - final_size);
846 fd = open(hp->filepath, O_RDWR);
848 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n",
849 hp->filepath, strerror(errno));
852 if (ftruncate(fd, final_size) < 0) {
853 RTE_LOG(ERR, EAL, "Cannot truncate %s: %s\n",
854 hp->filepath, strerror(errno));
859 pages_found += nr_pg_left;
860 hp->repeated = nr_pg_left;
864 /* else, lock the page and skip */
871 } /* foreach socket */
872 } /* foreach pagesize */
877 static inline uint64_t
878 get_socket_mem_size(int socket)
883 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
884 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
885 if (hpi->hugedir != NULL)
886 size += hpi->hugepage_sz * hpi->num_pages[socket];
893 * This function is a NUMA-aware equivalent of calc_num_pages.
894 * It takes in the list of hugepage sizes and the
895 * number of pages thereof, and calculates the best number of
896 * pages of each size to fulfill the request for <memory> ram
899 calc_num_pages_per_socket(uint64_t * memory,
900 struct hugepage_info *hp_info,
901 struct hugepage_info *hp_used,
902 unsigned num_hp_info)
904 unsigned socket, j, i = 0;
905 unsigned requested, available;
906 int total_num_pages = 0;
907 uint64_t remaining_mem, cur_mem;
908 uint64_t total_mem = internal_config.memory;
910 if (num_hp_info == 0)
913 /* if specific memory amounts per socket weren't requested */
914 if (internal_config.force_sockets == 0) {
915 int cpu_per_socket[RTE_MAX_NUMA_NODES];
916 size_t default_size, total_size;
919 /* Compute number of cores per socket */
920 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
921 RTE_LCORE_FOREACH(lcore_id) {
922 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
926 * Automatically spread requested memory amongst detected sockets according
927 * to number of cores from cpu mask present on each socket
929 total_size = internal_config.memory;
930 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
932 /* Set memory amount per socket */
933 default_size = (internal_config.memory * cpu_per_socket[socket])
936 /* Limit to maximum available memory on socket */
937 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
940 memory[socket] = default_size;
941 total_size -= default_size;
945 * If some memory is remaining, try to allocate it by getting all
946 * available memory from sockets, one after the other
948 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
949 /* take whatever is available */
950 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
954 memory[socket] += default_size;
955 total_size -= default_size;
959 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
960 /* skips if the memory on specific socket wasn't requested */
961 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
962 hp_used[i].hugedir = hp_info[i].hugedir;
963 hp_used[i].num_pages[socket] = RTE_MIN(
964 memory[socket] / hp_info[i].hugepage_sz,
965 hp_info[i].num_pages[socket]);
967 cur_mem = hp_used[i].num_pages[socket] *
968 hp_used[i].hugepage_sz;
970 memory[socket] -= cur_mem;
971 total_mem -= cur_mem;
973 total_num_pages += hp_used[i].num_pages[socket];
975 /* check if we have met all memory requests */
976 if (memory[socket] == 0)
979 /* check if we have any more pages left at this size, if so
980 * move on to next size */
981 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
983 /* At this point we know that there are more pages available that are
984 * bigger than the memory we want, so lets see if we can get enough
985 * from other page sizes.
988 for (j = i+1; j < num_hp_info; j++)
989 remaining_mem += hp_info[j].hugepage_sz *
990 hp_info[j].num_pages[socket];
992 /* is there enough other memory, if not allocate another page and quit */
993 if (remaining_mem < memory[socket]){
994 cur_mem = RTE_MIN(memory[socket],
995 hp_info[i].hugepage_sz);
996 memory[socket] -= cur_mem;
997 total_mem -= cur_mem;
998 hp_used[i].num_pages[socket]++;
1000 break; /* we are done with this socket*/
1003 /* if we didn't satisfy all memory requirements per socket */
1004 if (memory[socket] > 0) {
1005 /* to prevent icc errors */
1006 requested = (unsigned) (internal_config.socket_mem[socket] /
1008 available = requested -
1009 ((unsigned) (memory[socket] / 0x100000));
1010 RTE_LOG(INFO, EAL, "Not enough memory available on socket %u! "
1011 "Requested: %uMB, available: %uMB\n", socket,
1012 requested, available);
1017 /* if we didn't satisfy total memory requirements */
1018 if (total_mem > 0) {
1019 requested = (unsigned) (internal_config.memory / 0x100000);
1020 available = requested - (unsigned) (total_mem / 0x100000);
1021 RTE_LOG(INFO, EAL, "Not enough memory available! Requested: %uMB,"
1022 " available: %uMB\n", requested, available);
1025 return total_num_pages;
1029 * Prepare physical memory mapping: fill configuration structure with
1030 * these infos, return 0 on success.
1031 * 1. map N huge pages in separate files in hugetlbfs
1032 * 2. find associated physical addr
1033 * 3. find associated NUMA socket ID
1034 * 4. sort all huge pages by physical address
1035 * 5. remap these N huge pages in the correct order
1036 * 6. unmap the first mapping
1037 * 7. fill memsegs in configuration with contiguous zones
1040 rte_eal_hugepage_init(void)
1042 struct rte_mem_config *mcfg;
1043 struct hugepage_file *hugepage, *tmp_hp = NULL;
1044 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1046 uint64_t memory[RTE_MAX_NUMA_NODES];
1049 int i, j, new_memseg;
1050 int nr_hugefiles, nr_hugepages = 0;
1052 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1053 int new_pages_count[MAX_HUGEPAGE_SIZES];
1056 memset(used_hp, 0, sizeof(used_hp));
1058 /* get pointer to global configuration */
1059 mcfg = rte_eal_get_configuration()->mem_config;
1061 /* hugetlbfs can be disabled */
1062 if (internal_config.no_hugetlbfs) {
1063 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1064 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1065 if (addr == MAP_FAILED) {
1066 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1070 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
1071 mcfg->memseg[0].addr = addr;
1072 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1073 mcfg->memseg[0].len = internal_config.memory;
1074 mcfg->memseg[0].socket_id = SOCKET_ID_ANY;
1078 /* check if app runs on Xen Dom0 */
1079 if (internal_config.xen_dom0_support) {
1080 #ifdef RTE_LIBRTE_XEN_DOM0
1081 /* use dom0_mm kernel driver to init memory */
1082 if (rte_xen_dom0_memory_init() < 0)
1090 /* calculate total number of hugepages available. at this point we haven't
1091 * yet started sorting them so they all are on socket 0 */
1092 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1093 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1094 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1096 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1100 * allocate a memory area for hugepage table.
1101 * this isn't shared memory yet. due to the fact that we need some
1102 * processing done on these pages, shared memory will be created
1105 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1109 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1111 hp_offset = 0; /* where we start the current page size entries */
1113 /* map all hugepages and sort them */
1114 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1115 struct hugepage_info *hpi;
1118 * we don't yet mark hugepages as used at this stage, so
1119 * we just map all hugepages available to the system
1120 * all hugepages are still located on socket 0
1122 hpi = &internal_config.hugepage_info[i];
1124 if (hpi->num_pages[0] == 0)
1127 /* map all hugepages available */
1128 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
1129 RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
1130 (unsigned)(hpi->hugepage_sz / 0x100000));
1134 /* find physical addresses and sockets for each hugepage */
1135 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0){
1136 RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
1137 (unsigned)(hpi->hugepage_sz / 0x100000));
1141 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1142 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1143 (unsigned)(hpi->hugepage_sz / 0x100000));
1147 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
1150 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1151 /* remap all hugepages into single file segments */
1152 new_pages_count[i] = remap_all_hugepages(&tmp_hp[hp_offset], hpi);
1153 if (new_pages_count[i] < 0){
1154 RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1155 (unsigned)(hpi->hugepage_sz / 0x100000));
1159 /* we have processed a num of hugepages of this size, so inc offset */
1160 hp_offset += new_pages_count[i];
1162 /* remap all hugepages */
1163 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
1164 RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1165 (unsigned)(hpi->hugepage_sz / 0x100000));
1169 /* unmap original mappings */
1170 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1173 /* we have processed a num of hugepages of this size, so inc offset */
1174 hp_offset += hpi->num_pages[0];
1178 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1180 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1181 nr_hugefiles += new_pages_count[i];
1184 nr_hugefiles = nr_hugepages;
1188 /* clean out the numbers of pages */
1189 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1190 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1191 internal_config.hugepage_info[i].num_pages[j] = 0;
1193 /* get hugepages for each socket */
1194 for (i = 0; i < nr_hugefiles; i++) {
1195 int socket = tmp_hp[i].socket_id;
1197 /* find a hugepage info with right size and increment num_pages */
1198 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1199 (int)internal_config.num_hugepage_sizes);
1200 for (j = 0; j < nb_hpsizes; j++) {
1201 if (tmp_hp[i].size ==
1202 internal_config.hugepage_info[j].hugepage_sz) {
1203 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1204 internal_config.hugepage_info[j].num_pages[socket] +=
1207 internal_config.hugepage_info[j].num_pages[socket]++;
1213 /* make a copy of socket_mem, needed for number of pages calculation */
1214 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1215 memory[i] = internal_config.socket_mem[i];
1217 /* calculate final number of pages */
1218 nr_hugepages = calc_num_pages_per_socket(memory,
1219 internal_config.hugepage_info, used_hp,
1220 internal_config.num_hugepage_sizes);
1222 /* error if not enough memory available */
1223 if (nr_hugepages < 0)
1227 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1228 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1229 if (used_hp[i].num_pages[j] > 0) {
1231 "Requesting %u pages of size %uMB"
1232 " from socket %i\n",
1233 used_hp[i].num_pages[j],
1235 (used_hp[i].hugepage_sz / 0x100000),
1241 /* create shared memory */
1242 hugepage = create_shared_memory(eal_hugepage_info_path(),
1243 nr_hugefiles * sizeof(struct hugepage_file));
1245 if (hugepage == NULL) {
1246 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1249 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1252 * unmap pages that we won't need (looks at used_hp).
1253 * also, sets final_va to NULL on pages that were unmapped.
1255 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1256 internal_config.num_hugepage_sizes) < 0) {
1257 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1262 * copy stuff from malloc'd hugepage* to the actual shared memory.
1263 * this procedure only copies those hugepages that have final_va
1264 * not NULL. has overflow protection.
1266 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1267 tmp_hp, nr_hugefiles) < 0) {
1268 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1272 /* free the temporary hugepage table */
1276 /* find earliest free memseg - this is needed because in case of IVSHMEM,
1277 * segments might have already been initialized */
1278 for (j = 0; j < RTE_MAX_MEMSEG; j++)
1279 if (mcfg->memseg[j].addr == NULL) {
1280 /* move to previous segment and exit loop */
1285 for (i = 0; i < nr_hugefiles; i++) {
1288 /* if this is a new section, create a new memseg */
1291 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1293 else if (hugepage[i].size != hugepage[i-1].size)
1296 #ifdef RTE_ARCH_PPC_64
1297 /* On PPC64 architecture, the mmap always start from higher
1298 * virtual address to lower address. Here, both the physical
1299 * address and virtual address are in descending order */
1300 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1303 else if (((unsigned long)hugepage[i-1].final_va -
1304 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1307 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1310 else if (((unsigned long)hugepage[i].final_va -
1311 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1317 if (j == RTE_MAX_MEMSEG)
1320 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1321 mcfg->memseg[j].addr = hugepage[i].final_va;
1322 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1323 mcfg->memseg[j].len = hugepage[i].size * hugepage[i].repeated;
1325 mcfg->memseg[j].len = hugepage[i].size;
1327 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1328 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1330 /* continuation of previous memseg */
1332 #ifdef RTE_ARCH_PPC_64
1333 /* Use the phy and virt address of the last page as segment
1334 * address for IBM Power architecture */
1335 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1336 mcfg->memseg[j].addr = hugepage[i].final_va;
1338 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1340 hugepage[i].memseg_id = j;
1343 if (i < nr_hugefiles) {
1344 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1345 "from %d requested\n"
1346 "Current %s=%d is not enough\n"
1347 "Please either increase it or request less amount "
1349 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1363 * uses fstat to report the size of a file on disk
1369 if (fstat(fd, &st) < 0)
1375 * This creates the memory mappings in the secondary process to match that of
1376 * the server process. It goes through each memory segment in the DPDK runtime
1377 * configuration and finds the hugepages which form that segment, mapping them
1378 * in order to form a contiguous block in the virtual memory space
1381 rte_eal_hugepage_attach(void)
1383 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1384 const struct hugepage_file *hp = NULL;
1385 unsigned num_hp = 0;
1386 unsigned i, s = 0; /* s used to track the segment number */
1388 int fd, fd_zero = -1, fd_hugepage = -1;
1390 if (aslr_enabled() > 0) {
1391 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1392 "(ASLR) is enabled in the kernel.\n");
1393 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1394 "into secondary processes\n");
1397 if (internal_config.xen_dom0_support) {
1398 #ifdef RTE_LIBRTE_XEN_DOM0
1399 if (rte_xen_dom0_memory_attach() < 0) {
1400 RTE_LOG(ERR, EAL,"Failed to attach memory setments of primay "
1408 fd_zero = open("/dev/zero", O_RDONLY);
1410 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1413 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1414 if (fd_hugepage < 0) {
1415 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1419 /* map all segments into memory to make sure we get the addrs */
1420 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1424 * the first memory segment with len==0 is the one that
1425 * follows the last valid segment.
1427 if (mcfg->memseg[s].len == 0)
1430 #ifdef RTE_LIBRTE_IVSHMEM
1432 * if segment has ioremap address set, it's an IVSHMEM segment and
1433 * doesn't need mapping as it was already mapped earlier
1435 if (mcfg->memseg[s].ioremap_addr != 0)
1440 * fdzero is mmapped to get a contiguous block of virtual
1441 * addresses of the appropriate memseg size.
1442 * use mmap to get identical addresses as the primary process.
1444 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1445 PROT_READ, MAP_PRIVATE, fd_zero, 0);
1446 if (base_addr == MAP_FAILED ||
1447 base_addr != mcfg->memseg[s].addr) {
1448 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1449 "in /dev/zero to requested address [%p]: '%s'\n",
1450 (unsigned long long)mcfg->memseg[s].len,
1451 mcfg->memseg[s].addr, strerror(errno));
1452 if (aslr_enabled() > 0) {
1453 RTE_LOG(ERR, EAL, "It is recommended to "
1454 "disable ASLR in the kernel "
1455 "and retry running both primary "
1456 "and secondary processes\n");
1462 size = getFileSize(fd_hugepage);
1463 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1465 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1469 num_hp = size / sizeof(struct hugepage_file);
1470 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1473 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1474 void *addr, *base_addr;
1475 uintptr_t offset = 0;
1476 size_t mapping_size;
1477 #ifdef RTE_LIBRTE_IVSHMEM
1479 * if segment has ioremap address set, it's an IVSHMEM segment and
1480 * doesn't need mapping as it was already mapped earlier
1482 if (mcfg->memseg[s].ioremap_addr != 0) {
1488 * free previously mapped memory so we can map the
1489 * hugepages into the space
1491 base_addr = mcfg->memseg[s].addr;
1492 munmap(base_addr, mcfg->memseg[s].len);
1494 /* find the hugepages for this segment and map them
1495 * we don't need to worry about order, as the server sorted the
1496 * entries before it did the second mmap of them */
1497 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1498 if (hp[i].memseg_id == (int)s){
1499 fd = open(hp[i].filepath, O_RDWR);
1501 RTE_LOG(ERR, EAL, "Could not open %s\n",
1505 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1506 mapping_size = hp[i].size * hp[i].repeated;
1508 mapping_size = hp[i].size;
1510 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1511 mapping_size, PROT_READ | PROT_WRITE,
1513 close(fd); /* close file both on success and on failure */
1514 if (addr == MAP_FAILED ||
1515 addr != RTE_PTR_ADD(base_addr, offset)) {
1516 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1520 offset+=mapping_size;
1523 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1524 (unsigned long long)mcfg->memseg[s].len);
1527 /* unmap the hugepage config file, since we are done using it */
1528 munmap((void *)(uintptr_t)hp, size);
1536 if (fd_hugepage >= 0)
1542 rte_eal_memdevice_init(void)
1544 struct rte_config *config;
1546 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1549 config = rte_eal_get_configuration();
1550 config->mem_config->nchannel = internal_config.force_nchannel;
1551 config->mem_config->nrank = internal_config.force_nrank;
1557 test_proc_pagemap_readable(void)
1559 int fd = open("/proc/self/pagemap", O_RDONLY);
1569 /* init memory subsystem */
1571 rte_eal_memory_init(void)
1573 RTE_LOG(INFO, EAL, "Setting up memory...\n");
1575 proc_pagemap_readable = test_proc_pagemap_readable();
1576 if (!proc_pagemap_readable)
1578 "Cannot open /proc/self/pagemap: %s. "
1579 "virt2phys address translation will not work\n",
1582 const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
1583 rte_eal_hugepage_init() :
1584 rte_eal_hugepage_attach();
1588 if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)