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>
88 #include <rte_tailq.h>
90 #include <rte_eal_memconfig.h>
91 #include <rte_per_lcore.h>
92 #include <rte_lcore.h>
93 #include <rte_common.h>
94 #include <rte_string_fns.h>
96 #include "eal_private.h"
97 #include "eal_internal_cfg.h"
98 #include "eal_filesystem.h"
99 #include "eal_hugepages.h"
103 * Huge page mapping under linux
105 * To reserve a big contiguous amount of memory, we use the hugepage
106 * feature of linux. For that, we need to have hugetlbfs mounted. This
107 * code will create many files in this directory (one per page) and
108 * map them in virtual memory. For each page, we will retrieve its
109 * physical address and remap it in order to have a virtual contiguous
110 * zone as well as a physical contiguous zone.
113 static uint64_t baseaddr_offset;
115 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
117 /* Lock page in physical memory and prevent from swapping. */
119 rte_mem_lock_page(const void *virt)
121 unsigned long virtual = (unsigned long)virt;
122 int page_size = getpagesize();
123 unsigned long aligned = (virtual & ~ (page_size - 1));
124 return mlock((void*)aligned, page_size);
128 * Get physical address of any mapped virtual address in the current process.
131 rte_mem_virt2phy(const void *virtaddr)
134 uint64_t page, physaddr;
135 unsigned long virt_pfn;
139 /* standard page size */
140 page_size = getpagesize();
142 fd = open("/proc/self/pagemap", O_RDONLY);
144 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
145 __func__, strerror(errno));
146 return RTE_BAD_PHYS_ADDR;
149 virt_pfn = (unsigned long)virtaddr / page_size;
150 offset = sizeof(uint64_t) * virt_pfn;
151 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
152 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
153 __func__, strerror(errno));
155 return RTE_BAD_PHYS_ADDR;
157 if (read(fd, &page, sizeof(uint64_t)) < 0) {
158 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
159 __func__, strerror(errno));
161 return RTE_BAD_PHYS_ADDR;
165 * the pfn (page frame number) are bits 0-54 (see
166 * pagemap.txt in linux Documentation)
168 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
169 + ((unsigned long)virtaddr % page_size);
175 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
176 * it by browsing the /proc/self/pagemap special file.
179 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
184 for (i = 0; i < hpi->num_pages[0]; i++) {
185 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
186 if (addr == RTE_BAD_PHYS_ADDR)
188 hugepg_tbl[i].physaddr = addr;
194 * Check whether address-space layout randomization is enabled in
195 * the kernel. This is important for multi-process as it can prevent
196 * two processes mapping data to the same virtual address
198 * 0 - address space randomization disabled
199 * 1/2 - address space randomization enabled
200 * negative error code on error
206 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
209 retval = read(fd, &c, 1);
219 default: return -EINVAL;
224 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
225 * pointer to the mmap'd area and keep *size unmodified. Else, retry
226 * with a smaller zone: decrease *size by hugepage_sz until it reaches
227 * 0. In this case, return NULL. Note: this function returns an address
228 * which is a multiple of hugepage size.
231 get_virtual_area(size_t *size, size_t hugepage_sz)
237 if (internal_config.base_virtaddr != 0) {
238 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
243 RTE_LOG(INFO, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
245 fd = open("/dev/zero", O_RDONLY);
247 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
252 (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
253 if (addr == MAP_FAILED)
254 *size -= hugepage_sz;
255 } while (addr == MAP_FAILED && *size > 0);
257 if (addr == MAP_FAILED) {
259 RTE_LOG(INFO, EAL, "Cannot get a virtual area\n");
263 munmap(addr, (*size) + hugepage_sz);
266 /* align addr to a huge page size boundary */
267 aligned_addr = (long)addr;
268 aligned_addr += (hugepage_sz - 1);
269 aligned_addr &= (~(hugepage_sz - 1));
270 addr = (void *)(aligned_addr);
272 RTE_LOG(INFO, EAL, "Virtual area found at %p (size = 0x%zx)\n",
275 /* increment offset */
276 baseaddr_offset += *size;
282 * Mmap all hugepages of hugepage table: it first open a file in
283 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
284 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
285 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
286 * map continguous physical blocks in contiguous virtual blocks.
289 map_all_hugepages(struct hugepage_file *hugepg_tbl,
290 struct hugepage_info *hpi, int orig)
295 void *vma_addr = NULL;
298 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
299 RTE_SET_USED(vma_len);
302 for (i = 0; i < hpi->num_pages[0]; i++) {
303 uint64_t hugepage_sz = hpi->hugepage_sz;
306 hugepg_tbl[i].file_id = i;
307 hugepg_tbl[i].size = hugepage_sz;
308 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
309 eal_get_hugefile_temp_path(hugepg_tbl[i].filepath,
310 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
311 hugepg_tbl[i].file_id);
313 eal_get_hugefile_path(hugepg_tbl[i].filepath,
314 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
315 hugepg_tbl[i].file_id);
317 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
320 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
321 * original map address as final map address.
323 else if ((hugepage_sz == RTE_PGSIZE_1G)
324 || (hugepage_sz == RTE_PGSIZE_16G)) {
325 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
326 hugepg_tbl[i].orig_va = NULL;
331 #ifndef RTE_EAL_SINGLE_FILE_SEGMENTS
332 else if (vma_len == 0) {
333 unsigned j, num_pages;
335 /* reserve a virtual area for next contiguous
336 * physical block: count the number of
337 * contiguous physical pages. */
338 for (j = i+1; j < hpi->num_pages[0] ; j++) {
339 #ifdef RTE_ARCH_PPC_64
340 /* The physical addresses are sorted in
341 * descending order on PPC64 */
342 if (hugepg_tbl[j].physaddr !=
343 hugepg_tbl[j-1].physaddr - hugepage_sz)
346 if (hugepg_tbl[j].physaddr !=
347 hugepg_tbl[j-1].physaddr + hugepage_sz)
352 vma_len = num_pages * hugepage_sz;
354 /* get the biggest virtual memory area up to
355 * vma_len. If it fails, vma_addr is NULL, so
356 * let the kernel provide the address. */
357 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
358 if (vma_addr == NULL)
359 vma_len = hugepage_sz;
363 /* try to create hugepage file */
364 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
366 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
371 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
373 if (virtaddr == MAP_FAILED) {
374 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
381 hugepg_tbl[i].orig_va = virtaddr;
382 memset(virtaddr, 0, hugepage_sz);
385 hugepg_tbl[i].final_va = virtaddr;
388 /* set shared flock on the file. */
389 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
390 RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
391 __func__, strerror(errno));
398 vma_addr = (char *)vma_addr + hugepage_sz;
399 vma_len -= hugepage_sz;
404 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
407 * Remaps all hugepages into single file segments
410 remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
413 unsigned i = 0, j, num_pages, page_idx = 0;
414 void *vma_addr = NULL, *old_addr = NULL, *page_addr = NULL;
416 size_t hugepage_sz = hpi->hugepage_sz;
417 size_t total_size, offset;
418 char filepath[MAX_HUGEPAGE_PATH];
419 phys_addr_t physaddr;
422 while (i < hpi->num_pages[0]) {
425 /* for 32-bit systems, don't remap 1G pages and 16G pages,
426 * just reuse original map address as final map address.
428 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;
437 /* reserve a virtual area for next contiguous
438 * physical block: count the number of
439 * contiguous physical pages. */
440 for (j = i+1; j < hpi->num_pages[0] ; j++) {
441 #ifdef RTE_ARCH_PPC_64
442 /* The physical addresses are sorted in descending
444 if (hugepg_tbl[j].physaddr !=
445 hugepg_tbl[j-1].physaddr - hugepage_sz)
448 if (hugepg_tbl[j].physaddr !=
449 hugepg_tbl[j-1].physaddr + hugepage_sz)
454 vma_len = num_pages * hugepage_sz;
456 socket = hugepg_tbl[i].socket_id;
458 /* get the biggest virtual memory area up to
459 * vma_len. If it fails, vma_addr is NULL, so
460 * let the kernel provide the address. */
461 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
463 /* If we can't find a big enough virtual area, work out how many pages
464 * we are going to get */
465 if (vma_addr == NULL)
467 else if (vma_len != num_pages * hugepage_sz) {
468 num_pages = vma_len / hugepage_sz;
473 hugepg_tbl[page_idx].file_id = page_idx;
474 eal_get_hugefile_path(filepath,
477 hugepg_tbl[page_idx].file_id);
479 /* try to create hugepage file */
480 fd = open(filepath, O_CREAT | O_RDWR, 0755);
482 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__, strerror(errno));
489 /* unmap current segment */
491 munmap(vma_addr, total_size);
493 /* unmap original page */
494 munmap(hugepg_tbl[i].orig_va, hugepage_sz);
495 unlink(hugepg_tbl[i].filepath);
497 total_size += hugepage_sz;
501 /* map new, bigger segment */
502 vma_addr = mmap(vma_addr, total_size,
503 PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
505 if (vma_addr == MAP_FAILED || vma_addr != old_addr) {
506 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__, strerror(errno));
511 /* touch the page. this is needed because kernel postpones mapping
512 * creation until the first page fault. with this, we pin down
513 * the page and it is marked as used and gets into process' pagemap.
515 for (offset = 0; offset < total_size; offset += hugepage_sz)
516 *((volatile uint8_t*) RTE_PTR_ADD(vma_addr, offset));
519 /* set shared flock on the file. */
520 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
521 RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
522 __func__, strerror(errno));
527 snprintf(hugepg_tbl[page_idx].filepath, MAX_HUGEPAGE_PATH, "%s",
530 physaddr = rte_mem_virt2phy(vma_addr);
532 if (physaddr == RTE_BAD_PHYS_ADDR)
535 hugepg_tbl[page_idx].final_va = vma_addr;
537 hugepg_tbl[page_idx].physaddr = physaddr;
539 hugepg_tbl[page_idx].repeated = num_pages;
541 hugepg_tbl[page_idx].socket_id = socket;
545 /* verify the memory segment - that is, check that every VA corresponds
546 * to the physical address we expect to see
548 for (offset = 0; offset < vma_len; offset += hugepage_sz) {
549 uint64_t expected_physaddr;
551 expected_physaddr = hugepg_tbl[page_idx].physaddr + offset;
552 page_addr = RTE_PTR_ADD(vma_addr, offset);
553 physaddr = rte_mem_virt2phy(page_addr);
555 if (physaddr != expected_physaddr) {
556 RTE_LOG(ERR, EAL, "Segment sanity check failed: wrong physaddr "
557 "at %p (offset 0x%" PRIx64 ": 0x%" PRIx64
558 " (expected 0x%" PRIx64 ")\n",
559 page_addr, offset, physaddr, expected_physaddr);
564 /* zero out the whole segment */
565 memset(hugepg_tbl[page_idx].final_va, 0, total_size);
570 /* zero out the rest */
571 memset(&hugepg_tbl[page_idx], 0, (hpi->num_pages[0] - page_idx) * sizeof(struct hugepage_file));
574 #else/* RTE_EAL_SINGLE_FILE_SEGMENTS=n */
576 /* Unmap all hugepages from original mapping */
578 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
581 for (i = 0; i < hpi->num_pages[0]; i++) {
582 if (hugepg_tbl[i].orig_va) {
583 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
584 hugepg_tbl[i].orig_va = NULL;
589 #endif /* RTE_EAL_SINGLE_FILE_SEGMENTS */
592 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
596 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
600 unsigned i, hp_count = 0;
603 char hugedir_str[PATH_MAX];
606 f = fopen("/proc/self/numa_maps", "r");
608 RTE_LOG(INFO, EAL, "cannot open /proc/self/numa_maps,"
609 " consider that all memory is in socket_id 0\n");
613 snprintf(hugedir_str, sizeof(hugedir_str),
614 "%s/", hpi->hugedir);
617 while (fgets(buf, sizeof(buf), f) != NULL) {
619 /* ignore non huge page */
620 if (strstr(buf, " huge ") == NULL &&
621 strstr(buf, hugedir_str) == NULL)
625 virt_addr = strtoull(buf, &end, 16);
626 if (virt_addr == 0 || end == buf) {
627 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
631 /* get node id (socket id) */
632 nodestr = strstr(buf, " N");
633 if (nodestr == NULL) {
634 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
638 end = strstr(nodestr, "=");
640 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
646 socket_id = strtoul(nodestr, &end, 0);
647 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
648 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
652 /* if we find this page in our mappings, set socket_id */
653 for (i = 0; i < hpi->num_pages[0]; i++) {
654 void *va = (void *)(unsigned long)virt_addr;
655 if (hugepg_tbl[i].orig_va == va) {
656 hugepg_tbl[i].socket_id = socket_id;
662 if (hp_count < hpi->num_pages[0])
674 * Sort the hugepg_tbl by physical address (lower addresses first on x86,
675 * higher address first on powerpc). We use a slow algorithm, but we won't
676 * have millions of pages, and this is only done at init time.
679 sort_by_physaddr(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
683 uint64_t compare_addr;
684 struct hugepage_file tmp;
686 for (i = 0; i < hpi->num_pages[0]; i++) {
691 * browse all entries starting at 'i', and find the
692 * entry with the smallest addr
694 for (j=i; j< hpi->num_pages[0]; j++) {
696 if (compare_addr == 0 ||
697 #ifdef RTE_ARCH_PPC_64
698 hugepg_tbl[j].physaddr > compare_addr) {
700 hugepg_tbl[j].physaddr < compare_addr) {
702 compare_addr = hugepg_tbl[j].physaddr;
707 /* should not happen */
708 if (compare_idx == -1) {
709 RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
713 /* swap the 2 entries in the table */
714 memcpy(&tmp, &hugepg_tbl[compare_idx],
715 sizeof(struct hugepage_file));
716 memcpy(&hugepg_tbl[compare_idx], &hugepg_tbl[i],
717 sizeof(struct hugepage_file));
718 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage_file));
724 * Uses mmap to create a shared memory area for storage of data
725 * Used in this file to store the hugepage file map on disk
728 create_shared_memory(const char *filename, const size_t mem_size)
731 int fd = open(filename, O_CREAT | O_RDWR, 0666);
734 if (ftruncate(fd, mem_size) < 0) {
738 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
744 * this copies *active* hugepages from one hugepage table to another.
745 * destination is typically the shared memory.
748 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
749 const struct hugepage_file * src, int src_size)
751 int src_pos, dst_pos = 0;
753 for (src_pos = 0; src_pos < src_size; src_pos++) {
754 if (src[src_pos].final_va != NULL) {
755 /* error on overflow attempt */
756 if (dst_pos == dest_size)
758 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
766 * unmaps hugepages that are not going to be used. since we originally allocate
767 * ALL hugepages (not just those we need), additional unmapping needs to be done.
770 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
771 struct hugepage_info *hpi,
772 unsigned num_hp_info)
774 unsigned socket, size;
775 int page, nrpages = 0;
777 /* get total number of hugepages */
778 for (size = 0; size < num_hp_info; size++)
779 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
780 nrpages += internal_config.hugepage_info[size].num_pages[socket];
782 for (size = 0; size < num_hp_info; size++) {
783 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
784 unsigned pages_found = 0;
786 /* traverse until we have unmapped all the unused pages */
787 for (page = 0; page < nrpages; page++) {
788 struct hugepage_file *hp = &hugepg_tbl[page];
790 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
791 /* if this page was already cleared */
792 if (hp->final_va == NULL)
796 /* find a page that matches the criteria */
797 if ((hp->size == hpi[size].hugepage_sz) &&
798 (hp->socket_id == (int) socket)) {
800 /* if we skipped enough pages, unmap the rest */
801 if (pages_found == hpi[size].num_pages[socket]) {
804 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
805 unmap_len = hp->size * hp->repeated;
807 unmap_len = hp->size;
810 /* get start addr and len of the remaining segment */
811 munmap(hp->final_va, (size_t) unmap_len);
814 if (unlink(hp->filepath) == -1) {
815 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
816 __func__, hp->filepath, strerror(errno));
820 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
821 /* else, check how much do we need to map */
824 hpi[size].num_pages[socket] - pages_found;
826 /* if we need enough memory to fit into the segment */
827 if (hp->repeated <= nr_pg_left) {
828 pages_found += hp->repeated;
830 /* truncate the segment */
832 uint64_t final_size = nr_pg_left * hp->size;
833 uint64_t seg_size = hp->repeated * hp->size;
835 void * unmap_va = RTE_PTR_ADD(hp->final_va,
839 munmap(unmap_va, seg_size - final_size);
841 fd = open(hp->filepath, O_RDWR);
843 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n",
844 hp->filepath, strerror(errno));
847 if (ftruncate(fd, final_size) < 0) {
848 RTE_LOG(ERR, EAL, "Cannot truncate %s: %s\n",
849 hp->filepath, strerror(errno));
854 pages_found += nr_pg_left;
855 hp->repeated = nr_pg_left;
859 /* else, lock the page and skip */
866 } /* foreach socket */
867 } /* foreach pagesize */
872 static inline uint64_t
873 get_socket_mem_size(int socket)
878 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
879 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
880 if (hpi->hugedir != NULL)
881 size += hpi->hugepage_sz * hpi->num_pages[socket];
888 * This function is a NUMA-aware equivalent of calc_num_pages.
889 * It takes in the list of hugepage sizes and the
890 * number of pages thereof, and calculates the best number of
891 * pages of each size to fulfill the request for <memory> ram
894 calc_num_pages_per_socket(uint64_t * memory,
895 struct hugepage_info *hp_info,
896 struct hugepage_info *hp_used,
897 unsigned num_hp_info)
899 unsigned socket, j, i = 0;
900 unsigned requested, available;
901 int total_num_pages = 0;
902 uint64_t remaining_mem, cur_mem;
903 uint64_t total_mem = internal_config.memory;
905 if (num_hp_info == 0)
908 /* if specific memory amounts per socket weren't requested */
909 if (internal_config.force_sockets == 0) {
910 int cpu_per_socket[RTE_MAX_NUMA_NODES];
911 size_t default_size, total_size;
914 /* Compute number of cores per socket */
915 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
916 RTE_LCORE_FOREACH(lcore_id) {
917 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
921 * Automatically spread requested memory amongst detected sockets according
922 * to number of cores from cpu mask present on each socket
924 total_size = internal_config.memory;
925 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
927 /* Set memory amount per socket */
928 default_size = (internal_config.memory * cpu_per_socket[socket])
931 /* Limit to maximum available memory on socket */
932 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
935 memory[socket] = default_size;
936 total_size -= default_size;
940 * If some memory is remaining, try to allocate it by getting all
941 * available memory from sockets, one after the other
943 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
944 /* take whatever is available */
945 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
949 memory[socket] += default_size;
950 total_size -= default_size;
954 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
955 /* skips if the memory on specific socket wasn't requested */
956 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
957 hp_used[i].hugedir = hp_info[i].hugedir;
958 hp_used[i].num_pages[socket] = RTE_MIN(
959 memory[socket] / hp_info[i].hugepage_sz,
960 hp_info[i].num_pages[socket]);
962 cur_mem = hp_used[i].num_pages[socket] *
963 hp_used[i].hugepage_sz;
965 memory[socket] -= cur_mem;
966 total_mem -= cur_mem;
968 total_num_pages += hp_used[i].num_pages[socket];
970 /* check if we have met all memory requests */
971 if (memory[socket] == 0)
974 /* check if we have any more pages left at this size, if so
975 * move on to next size */
976 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
978 /* At this point we know that there are more pages available that are
979 * bigger than the memory we want, so lets see if we can get enough
980 * from other page sizes.
983 for (j = i+1; j < num_hp_info; j++)
984 remaining_mem += hp_info[j].hugepage_sz *
985 hp_info[j].num_pages[socket];
987 /* is there enough other memory, if not allocate another page and quit */
988 if (remaining_mem < memory[socket]){
989 cur_mem = RTE_MIN(memory[socket],
990 hp_info[i].hugepage_sz);
991 memory[socket] -= cur_mem;
992 total_mem -= cur_mem;
993 hp_used[i].num_pages[socket]++;
995 break; /* we are done with this socket*/
998 /* if we didn't satisfy all memory requirements per socket */
999 if (memory[socket] > 0) {
1000 /* to prevent icc errors */
1001 requested = (unsigned) (internal_config.socket_mem[socket] /
1003 available = requested -
1004 ((unsigned) (memory[socket] / 0x100000));
1005 RTE_LOG(INFO, EAL, "Not enough memory available on socket %u! "
1006 "Requested: %uMB, available: %uMB\n", socket,
1007 requested, available);
1012 /* if we didn't satisfy total memory requirements */
1013 if (total_mem > 0) {
1014 requested = (unsigned) (internal_config.memory / 0x100000);
1015 available = requested - (unsigned) (total_mem / 0x100000);
1016 RTE_LOG(INFO, EAL, "Not enough memory available! Requested: %uMB,"
1017 " available: %uMB\n", requested, available);
1020 return total_num_pages;
1024 * Prepare physical memory mapping: fill configuration structure with
1025 * these infos, return 0 on success.
1026 * 1. map N huge pages in separate files in hugetlbfs
1027 * 2. find associated physical addr
1028 * 3. find associated NUMA socket ID
1029 * 4. sort all huge pages by physical address
1030 * 5. remap these N huge pages in the correct order
1031 * 6. unmap the first mapping
1032 * 7. fill memsegs in configuration with contiguous zones
1035 rte_eal_hugepage_init(void)
1037 struct rte_mem_config *mcfg;
1038 struct hugepage_file *hugepage, *tmp_hp = NULL;
1039 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1041 uint64_t memory[RTE_MAX_NUMA_NODES];
1044 int i, j, new_memseg;
1045 int nr_hugefiles, nr_hugepages = 0;
1047 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1048 int new_pages_count[MAX_HUGEPAGE_SIZES];
1051 memset(used_hp, 0, sizeof(used_hp));
1053 /* get pointer to global configuration */
1054 mcfg = rte_eal_get_configuration()->mem_config;
1056 /* hugetlbfs can be disabled */
1057 if (internal_config.no_hugetlbfs) {
1058 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1059 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1060 if (addr == MAP_FAILED) {
1061 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1065 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
1066 mcfg->memseg[0].addr = addr;
1067 mcfg->memseg[0].len = internal_config.memory;
1068 mcfg->memseg[0].socket_id = SOCKET_ID_ANY;
1072 /* check if app runs on Xen Dom0 */
1073 if (internal_config.xen_dom0_support) {
1074 #ifdef RTE_LIBRTE_XEN_DOM0
1075 /* use dom0_mm kernel driver to init memory */
1076 if (rte_xen_dom0_memory_init() < 0)
1084 /* calculate total number of hugepages available. at this point we haven't
1085 * yet started sorting them so they all are on socket 0 */
1086 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1087 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1088 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1090 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1094 * allocate a memory area for hugepage table.
1095 * this isn't shared memory yet. due to the fact that we need some
1096 * processing done on these pages, shared memory will be created
1099 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1103 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1105 hp_offset = 0; /* where we start the current page size entries */
1107 /* map all hugepages and sort them */
1108 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1109 struct hugepage_info *hpi;
1112 * we don't yet mark hugepages as used at this stage, so
1113 * we just map all hugepages available to the system
1114 * all hugepages are still located on socket 0
1116 hpi = &internal_config.hugepage_info[i];
1118 if (hpi->num_pages[0] == 0)
1121 /* map all hugepages available */
1122 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
1123 RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
1124 (unsigned)(hpi->hugepage_sz / 0x100000));
1128 /* find physical addresses and sockets for each hugepage */
1129 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0){
1130 RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
1131 (unsigned)(hpi->hugepage_sz / 0x100000));
1135 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1136 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1137 (unsigned)(hpi->hugepage_sz / 0x100000));
1141 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
1144 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1145 /* remap all hugepages into single file segments */
1146 new_pages_count[i] = remap_all_hugepages(&tmp_hp[hp_offset], hpi);
1147 if (new_pages_count[i] < 0){
1148 RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1149 (unsigned)(hpi->hugepage_sz / 0x100000));
1153 /* we have processed a num of hugepages of this size, so inc offset */
1154 hp_offset += new_pages_count[i];
1156 /* remap all hugepages */
1157 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
1158 RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1159 (unsigned)(hpi->hugepage_sz / 0x100000));
1163 /* unmap original mappings */
1164 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1167 /* we have processed a num of hugepages of this size, so inc offset */
1168 hp_offset += hpi->num_pages[0];
1172 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1174 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1175 nr_hugefiles += new_pages_count[i];
1178 nr_hugefiles = nr_hugepages;
1182 /* clean out the numbers of pages */
1183 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1184 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1185 internal_config.hugepage_info[i].num_pages[j] = 0;
1187 /* get hugepages for each socket */
1188 for (i = 0; i < nr_hugefiles; i++) {
1189 int socket = tmp_hp[i].socket_id;
1191 /* find a hugepage info with right size and increment num_pages */
1192 for (j = 0; j < (int) internal_config.num_hugepage_sizes; j++) {
1193 if (tmp_hp[i].size ==
1194 internal_config.hugepage_info[j].hugepage_sz) {
1195 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1196 internal_config.hugepage_info[j].num_pages[socket] +=
1199 internal_config.hugepage_info[j].num_pages[socket]++;
1205 /* make a copy of socket_mem, needed for number of pages calculation */
1206 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1207 memory[i] = internal_config.socket_mem[i];
1209 /* calculate final number of pages */
1210 nr_hugepages = calc_num_pages_per_socket(memory,
1211 internal_config.hugepage_info, used_hp,
1212 internal_config.num_hugepage_sizes);
1214 /* error if not enough memory available */
1215 if (nr_hugepages < 0)
1219 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1220 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1221 if (used_hp[i].num_pages[j] > 0) {
1223 "Requesting %u pages of size %uMB"
1224 " from socket %i\n",
1225 used_hp[i].num_pages[j],
1227 (used_hp[i].hugepage_sz / 0x100000),
1233 /* create shared memory */
1234 hugepage = create_shared_memory(eal_hugepage_info_path(),
1235 nr_hugefiles * sizeof(struct hugepage_file));
1237 if (hugepage == NULL) {
1238 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1241 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1244 * unmap pages that we won't need (looks at used_hp).
1245 * also, sets final_va to NULL on pages that were unmapped.
1247 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1248 internal_config.num_hugepage_sizes) < 0) {
1249 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1254 * copy stuff from malloc'd hugepage* to the actual shared memory.
1255 * this procedure only copies those hugepages that have final_va
1256 * not NULL. has overflow protection.
1258 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1259 tmp_hp, nr_hugefiles) < 0) {
1260 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1264 /* free the temporary hugepage table */
1268 /* find earliest free memseg - this is needed because in case of IVSHMEM,
1269 * segments might have already been initialized */
1270 for (j = 0; j < RTE_MAX_MEMSEG; j++)
1271 if (mcfg->memseg[j].addr == NULL) {
1272 /* move to previous segment and exit loop */
1277 for (i = 0; i < nr_hugefiles; i++) {
1280 /* if this is a new section, create a new memseg */
1283 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1285 else if (hugepage[i].size != hugepage[i-1].size)
1288 #ifdef RTE_ARCH_PPC_64
1289 /* On PPC64 architecture, the mmap always start from higher
1290 * virtual address to lower address. Here, both the physical
1291 * address and virtual address are in descending order */
1292 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1295 else if (((unsigned long)hugepage[i-1].final_va -
1296 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1299 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1302 else if (((unsigned long)hugepage[i].final_va -
1303 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1309 if (j == RTE_MAX_MEMSEG)
1312 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1313 mcfg->memseg[j].addr = hugepage[i].final_va;
1314 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1315 mcfg->memseg[j].len = hugepage[i].size * hugepage[i].repeated;
1317 mcfg->memseg[j].len = hugepage[i].size;
1319 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1320 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1322 /* continuation of previous memseg */
1324 #ifdef RTE_ARCH_PPC_64
1325 /* Use the phy and virt address of the last page as segment
1326 * address for IBM Power architecture */
1327 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1328 mcfg->memseg[j].addr = hugepage[i].final_va;
1330 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1332 hugepage[i].memseg_id = j;
1335 if (i < nr_hugefiles) {
1336 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1337 "from %d requested\n"
1338 "Current %s=%d is not enough\n"
1339 "Please either increase it or request less amount "
1341 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1355 * uses fstat to report the size of a file on disk
1361 if (fstat(fd, &st) < 0)
1367 * This creates the memory mappings in the secondary process to match that of
1368 * the server process. It goes through each memory segment in the DPDK runtime
1369 * configuration and finds the hugepages which form that segment, mapping them
1370 * in order to form a contiguous block in the virtual memory space
1373 rte_eal_hugepage_attach(void)
1375 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1376 const struct hugepage_file *hp = NULL;
1377 unsigned num_hp = 0;
1378 unsigned i, s = 0; /* s used to track the segment number */
1380 int fd, fd_zero = -1, fd_hugepage = -1;
1382 if (aslr_enabled() > 0) {
1383 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1384 "(ASLR) is enabled in the kernel.\n");
1385 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1386 "into secondary processes\n");
1389 if (internal_config.xen_dom0_support) {
1390 #ifdef RTE_LIBRTE_XEN_DOM0
1391 if (rte_xen_dom0_memory_attach() < 0) {
1392 RTE_LOG(ERR, EAL,"Failed to attach memory setments of primay "
1400 fd_zero = open("/dev/zero", O_RDONLY);
1402 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1405 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1406 if (fd_hugepage < 0) {
1407 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1411 /* map all segments into memory to make sure we get the addrs */
1412 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1416 * the first memory segment with len==0 is the one that
1417 * follows the last valid segment.
1419 if (mcfg->memseg[s].len == 0)
1422 #ifdef RTE_LIBRTE_IVSHMEM
1424 * if segment has ioremap address set, it's an IVSHMEM segment and
1425 * doesn't need mapping as it was already mapped earlier
1427 if (mcfg->memseg[s].ioremap_addr != 0)
1432 * fdzero is mmapped to get a contiguous block of virtual
1433 * addresses of the appropriate memseg size.
1434 * use mmap to get identical addresses as the primary process.
1436 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1437 PROT_READ, MAP_PRIVATE, fd_zero, 0);
1438 if (base_addr == MAP_FAILED ||
1439 base_addr != mcfg->memseg[s].addr) {
1440 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1441 "in /dev/zero to requested address [%p]: '%s'\n",
1442 (unsigned long long)mcfg->memseg[s].len,
1443 mcfg->memseg[s].addr, strerror(errno));
1444 if (aslr_enabled() > 0) {
1445 RTE_LOG(ERR, EAL, "It is recommended to "
1446 "disable ASLR in the kernel "
1447 "and retry running both primary "
1448 "and secondary processes\n");
1454 size = getFileSize(fd_hugepage);
1455 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1457 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1461 num_hp = size / sizeof(struct hugepage_file);
1462 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1465 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1466 void *addr, *base_addr;
1467 uintptr_t offset = 0;
1468 size_t mapping_size;
1469 #ifdef RTE_LIBRTE_IVSHMEM
1471 * if segment has ioremap address set, it's an IVSHMEM segment and
1472 * doesn't need mapping as it was already mapped earlier
1474 if (mcfg->memseg[s].ioremap_addr != 0) {
1480 * free previously mapped memory so we can map the
1481 * hugepages into the space
1483 base_addr = mcfg->memseg[s].addr;
1484 munmap(base_addr, mcfg->memseg[s].len);
1486 /* find the hugepages for this segment and map them
1487 * we don't need to worry about order, as the server sorted the
1488 * entries before it did the second mmap of them */
1489 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1490 if (hp[i].memseg_id == (int)s){
1491 fd = open(hp[i].filepath, O_RDWR);
1493 RTE_LOG(ERR, EAL, "Could not open %s\n",
1497 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1498 mapping_size = hp[i].size * hp[i].repeated;
1500 mapping_size = hp[i].size;
1502 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1503 mapping_size, PROT_READ | PROT_WRITE,
1505 close(fd); /* close file both on success and on failure */
1506 if (addr == MAP_FAILED ||
1507 addr != RTE_PTR_ADD(base_addr, offset)) {
1508 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1512 offset+=mapping_size;
1515 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1516 (unsigned long long)mcfg->memseg[s].len);
1519 /* unmap the hugepage config file, since we are done using it */
1520 munmap((void *)(uintptr_t)hp, size);
1528 if (fd_hugepage >= 0)
1534 rte_eal_memdevice_init(void)
1536 struct rte_config *config;
1538 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1541 config = rte_eal_get_configuration();
1542 config->mem_config->nchannel = internal_config.force_nchannel;
1543 config->mem_config->nrank = internal_config.force_nrank;
1549 /* init memory subsystem */
1551 rte_eal_memory_init(void)
1553 RTE_LOG(INFO, EAL, "Setting up memory...\n");
1554 const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
1555 rte_eal_hugepage_init() :
1556 rte_eal_hugepage_attach();
1560 if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)