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"
119 test_proc_pagemap_readable(void)
121 int fd = open("/proc/self/pagemap", O_RDONLY);
125 "Cannot open /proc/self/pagemap: %s. "
126 "virt2phys address translation will not work\n",
133 proc_pagemap_readable = 1;
136 /* Lock page in physical memory and prevent from swapping. */
138 rte_mem_lock_page(const void *virt)
140 unsigned long virtual = (unsigned long)virt;
141 int page_size = getpagesize();
142 unsigned long aligned = (virtual & ~ (page_size - 1));
143 return mlock((void*)aligned, page_size);
147 * Get physical address of any mapped virtual address in the current process.
150 rte_mem_virt2phy(const void *virtaddr)
153 uint64_t page, physaddr;
154 unsigned long virt_pfn;
158 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
159 if (!proc_pagemap_readable)
160 return RTE_BAD_PHYS_ADDR;
162 /* standard page size */
163 page_size = getpagesize();
165 fd = open("/proc/self/pagemap", O_RDONLY);
167 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
168 __func__, strerror(errno));
169 return RTE_BAD_PHYS_ADDR;
172 virt_pfn = (unsigned long)virtaddr / page_size;
173 offset = sizeof(uint64_t) * virt_pfn;
174 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
175 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
176 __func__, strerror(errno));
178 return RTE_BAD_PHYS_ADDR;
180 if (read(fd, &page, sizeof(uint64_t)) < 0) {
181 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
182 __func__, strerror(errno));
184 return RTE_BAD_PHYS_ADDR;
188 * the pfn (page frame number) are bits 0-54 (see
189 * pagemap.txt in linux Documentation)
191 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
192 + ((unsigned long)virtaddr % page_size);
198 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
199 * it by browsing the /proc/self/pagemap special file.
202 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
207 for (i = 0; i < hpi->num_pages[0]; i++) {
208 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
209 if (addr == RTE_BAD_PHYS_ADDR)
211 hugepg_tbl[i].physaddr = addr;
217 * Check whether address-space layout randomization is enabled in
218 * the kernel. This is important for multi-process as it can prevent
219 * two processes mapping data to the same virtual address
221 * 0 - address space randomization disabled
222 * 1/2 - address space randomization enabled
223 * negative error code on error
229 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
232 retval = read(fd, &c, 1);
242 default: return -EINVAL;
247 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
248 * pointer to the mmap'd area and keep *size unmodified. Else, retry
249 * with a smaller zone: decrease *size by hugepage_sz until it reaches
250 * 0. In this case, return NULL. Note: this function returns an address
251 * which is a multiple of hugepage size.
254 get_virtual_area(size_t *size, size_t hugepage_sz)
260 if (internal_config.base_virtaddr != 0) {
261 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
266 RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
268 fd = open("/dev/zero", O_RDONLY);
270 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
275 (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
276 if (addr == MAP_FAILED)
277 *size -= hugepage_sz;
278 } while (addr == MAP_FAILED && *size > 0);
280 if (addr == MAP_FAILED) {
282 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
287 munmap(addr, (*size) + hugepage_sz);
290 /* align addr to a huge page size boundary */
291 aligned_addr = (long)addr;
292 aligned_addr += (hugepage_sz - 1);
293 aligned_addr &= (~(hugepage_sz - 1));
294 addr = (void *)(aligned_addr);
296 RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
299 /* increment offset */
300 baseaddr_offset += *size;
306 * Mmap all hugepages of hugepage table: it first open a file in
307 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
308 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
309 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
310 * map continguous physical blocks in contiguous virtual blocks.
313 map_all_hugepages(struct hugepage_file *hugepg_tbl,
314 struct hugepage_info *hpi, int orig)
319 void *vma_addr = NULL;
322 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
323 RTE_SET_USED(vma_len);
326 for (i = 0; i < hpi->num_pages[0]; i++) {
327 uint64_t hugepage_sz = hpi->hugepage_sz;
330 hugepg_tbl[i].file_id = i;
331 hugepg_tbl[i].size = hugepage_sz;
332 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
333 eal_get_hugefile_temp_path(hugepg_tbl[i].filepath,
334 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
335 hugepg_tbl[i].file_id);
337 eal_get_hugefile_path(hugepg_tbl[i].filepath,
338 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
339 hugepg_tbl[i].file_id);
341 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
344 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
345 * original map address as final map address.
347 else if ((hugepage_sz == RTE_PGSIZE_1G)
348 || (hugepage_sz == RTE_PGSIZE_16G)) {
349 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
350 hugepg_tbl[i].orig_va = NULL;
355 #ifndef RTE_EAL_SINGLE_FILE_SEGMENTS
356 else if (vma_len == 0) {
357 unsigned j, num_pages;
359 /* reserve a virtual area for next contiguous
360 * physical block: count the number of
361 * contiguous physical pages. */
362 for (j = i+1; j < hpi->num_pages[0] ; j++) {
363 #ifdef RTE_ARCH_PPC_64
364 /* The physical addresses are sorted in
365 * descending order on PPC64 */
366 if (hugepg_tbl[j].physaddr !=
367 hugepg_tbl[j-1].physaddr - hugepage_sz)
370 if (hugepg_tbl[j].physaddr !=
371 hugepg_tbl[j-1].physaddr + hugepage_sz)
376 vma_len = num_pages * hugepage_sz;
378 /* get the biggest virtual memory area up to
379 * vma_len. If it fails, vma_addr is NULL, so
380 * let the kernel provide the address. */
381 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
382 if (vma_addr == NULL)
383 vma_len = hugepage_sz;
387 /* try to create hugepage file */
388 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
390 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
395 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
397 if (virtaddr == MAP_FAILED) {
398 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
405 hugepg_tbl[i].orig_va = virtaddr;
406 memset(virtaddr, 0, hugepage_sz);
409 hugepg_tbl[i].final_va = virtaddr;
412 /* set shared flock on the file. */
413 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
414 RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
415 __func__, strerror(errno));
422 vma_addr = (char *)vma_addr + hugepage_sz;
423 vma_len -= hugepage_sz;
428 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
431 * Remaps all hugepages into single file segments
434 remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
437 unsigned i = 0, j, num_pages, page_idx = 0;
438 void *vma_addr = NULL, *old_addr = NULL, *page_addr = NULL;
440 size_t hugepage_sz = hpi->hugepage_sz;
441 size_t total_size, offset;
442 char filepath[MAX_HUGEPAGE_PATH];
443 phys_addr_t physaddr;
446 while (i < hpi->num_pages[0]) {
449 /* for 32-bit systems, don't remap 1G pages and 16G pages,
450 * just reuse original map address as final map address.
452 if ((hugepage_sz == RTE_PGSIZE_1G)
453 || (hugepage_sz == RTE_PGSIZE_16G)) {
454 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
455 hugepg_tbl[i].orig_va = NULL;
461 /* reserve a virtual area for next contiguous
462 * physical block: count the number of
463 * contiguous physical pages. */
464 for (j = i+1; j < hpi->num_pages[0] ; j++) {
465 #ifdef RTE_ARCH_PPC_64
466 /* The physical addresses are sorted in descending
468 if (hugepg_tbl[j].physaddr !=
469 hugepg_tbl[j-1].physaddr - hugepage_sz)
472 if (hugepg_tbl[j].physaddr !=
473 hugepg_tbl[j-1].physaddr + hugepage_sz)
478 vma_len = num_pages * hugepage_sz;
480 socket = hugepg_tbl[i].socket_id;
482 /* get the biggest virtual memory area up to
483 * vma_len. If it fails, vma_addr is NULL, so
484 * let the kernel provide the address. */
485 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
487 /* If we can't find a big enough virtual area, work out how many pages
488 * we are going to get */
489 if (vma_addr == NULL)
491 else if (vma_len != num_pages * hugepage_sz) {
492 num_pages = vma_len / hugepage_sz;
497 hugepg_tbl[page_idx].file_id = page_idx;
498 eal_get_hugefile_path(filepath,
501 hugepg_tbl[page_idx].file_id);
503 /* try to create hugepage file */
504 fd = open(filepath, O_CREAT | O_RDWR, 0755);
506 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__, strerror(errno));
513 /* unmap current segment */
515 munmap(vma_addr, total_size);
517 /* unmap original page */
518 munmap(hugepg_tbl[i].orig_va, hugepage_sz);
519 unlink(hugepg_tbl[i].filepath);
521 total_size += hugepage_sz;
525 /* map new, bigger segment */
526 vma_addr = mmap(vma_addr, total_size,
527 PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
529 if (vma_addr == MAP_FAILED || vma_addr != old_addr) {
530 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__, strerror(errno));
535 /* touch the page. this is needed because kernel postpones mapping
536 * creation until the first page fault. with this, we pin down
537 * the page and it is marked as used and gets into process' pagemap.
539 for (offset = 0; offset < total_size; offset += hugepage_sz)
540 *((volatile uint8_t*) RTE_PTR_ADD(vma_addr, offset));
543 /* set shared flock on the file. */
544 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
545 RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
546 __func__, strerror(errno));
551 snprintf(hugepg_tbl[page_idx].filepath, MAX_HUGEPAGE_PATH, "%s",
554 physaddr = rte_mem_virt2phy(vma_addr);
556 if (physaddr == RTE_BAD_PHYS_ADDR)
559 hugepg_tbl[page_idx].final_va = vma_addr;
561 hugepg_tbl[page_idx].physaddr = physaddr;
563 hugepg_tbl[page_idx].repeated = num_pages;
565 hugepg_tbl[page_idx].socket_id = socket;
569 /* verify the memory segment - that is, check that every VA corresponds
570 * to the physical address we expect to see
572 for (offset = 0; offset < vma_len; offset += hugepage_sz) {
573 uint64_t expected_physaddr;
575 expected_physaddr = hugepg_tbl[page_idx].physaddr + offset;
576 page_addr = RTE_PTR_ADD(vma_addr, offset);
577 physaddr = rte_mem_virt2phy(page_addr);
579 if (physaddr != expected_physaddr) {
580 RTE_LOG(ERR, EAL, "Segment sanity check failed: wrong physaddr "
581 "at %p (offset 0x%" PRIx64 ": 0x%" PRIx64
582 " (expected 0x%" PRIx64 ")\n",
583 page_addr, offset, physaddr, expected_physaddr);
588 /* zero out the whole segment */
589 memset(hugepg_tbl[page_idx].final_va, 0, total_size);
594 /* zero out the rest */
595 memset(&hugepg_tbl[page_idx], 0, (hpi->num_pages[0] - page_idx) * sizeof(struct hugepage_file));
598 #else/* RTE_EAL_SINGLE_FILE_SEGMENTS=n */
600 /* Unmap all hugepages from original mapping */
602 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
605 for (i = 0; i < hpi->num_pages[0]; i++) {
606 if (hugepg_tbl[i].orig_va) {
607 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
608 hugepg_tbl[i].orig_va = NULL;
613 #endif /* RTE_EAL_SINGLE_FILE_SEGMENTS */
616 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
620 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
624 unsigned i, hp_count = 0;
627 char hugedir_str[PATH_MAX];
630 f = fopen("/proc/self/numa_maps", "r");
632 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
633 " consider that all memory is in socket_id 0\n");
637 snprintf(hugedir_str, sizeof(hugedir_str),
638 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
641 while (fgets(buf, sizeof(buf), f) != NULL) {
643 /* ignore non huge page */
644 if (strstr(buf, " huge ") == NULL &&
645 strstr(buf, hugedir_str) == NULL)
649 virt_addr = strtoull(buf, &end, 16);
650 if (virt_addr == 0 || end == buf) {
651 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
655 /* get node id (socket id) */
656 nodestr = strstr(buf, " N");
657 if (nodestr == NULL) {
658 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
662 end = strstr(nodestr, "=");
664 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
670 socket_id = strtoul(nodestr, &end, 0);
671 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
672 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
676 /* if we find this page in our mappings, set socket_id */
677 for (i = 0; i < hpi->num_pages[0]; i++) {
678 void *va = (void *)(unsigned long)virt_addr;
679 if (hugepg_tbl[i].orig_va == va) {
680 hugepg_tbl[i].socket_id = socket_id;
686 if (hp_count < hpi->num_pages[0])
698 * Sort the hugepg_tbl by physical address (lower addresses first on x86,
699 * higher address first on powerpc). We use a slow algorithm, but we won't
700 * have millions of pages, and this is only done at init time.
703 sort_by_physaddr(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
707 uint64_t compare_addr;
708 struct hugepage_file tmp;
710 for (i = 0; i < hpi->num_pages[0]; i++) {
715 * browse all entries starting at 'i', and find the
716 * entry with the smallest addr
718 for (j=i; j< hpi->num_pages[0]; j++) {
720 if (compare_addr == 0 ||
721 #ifdef RTE_ARCH_PPC_64
722 hugepg_tbl[j].physaddr > compare_addr) {
724 hugepg_tbl[j].physaddr < compare_addr) {
726 compare_addr = hugepg_tbl[j].physaddr;
731 /* should not happen */
732 if (compare_idx == -1) {
733 RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
737 /* swap the 2 entries in the table */
738 memcpy(&tmp, &hugepg_tbl[compare_idx],
739 sizeof(struct hugepage_file));
740 memcpy(&hugepg_tbl[compare_idx], &hugepg_tbl[i],
741 sizeof(struct hugepage_file));
742 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage_file));
748 * Uses mmap to create a shared memory area for storage of data
749 * Used in this file to store the hugepage file map on disk
752 create_shared_memory(const char *filename, const size_t mem_size)
755 int fd = open(filename, O_CREAT | O_RDWR, 0666);
758 if (ftruncate(fd, mem_size) < 0) {
762 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
768 * this copies *active* hugepages from one hugepage table to another.
769 * destination is typically the shared memory.
772 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
773 const struct hugepage_file * src, int src_size)
775 int src_pos, dst_pos = 0;
777 for (src_pos = 0; src_pos < src_size; src_pos++) {
778 if (src[src_pos].final_va != NULL) {
779 /* error on overflow attempt */
780 if (dst_pos == dest_size)
782 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
790 * unmaps hugepages that are not going to be used. since we originally allocate
791 * ALL hugepages (not just those we need), additional unmapping needs to be done.
794 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
795 struct hugepage_info *hpi,
796 unsigned num_hp_info)
798 unsigned socket, size;
799 int page, nrpages = 0;
801 /* get total number of hugepages */
802 for (size = 0; size < num_hp_info; size++)
803 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
804 nrpages += internal_config.hugepage_info[size].num_pages[socket];
806 for (size = 0; size < num_hp_info; size++) {
807 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
808 unsigned pages_found = 0;
810 /* traverse until we have unmapped all the unused pages */
811 for (page = 0; page < nrpages; page++) {
812 struct hugepage_file *hp = &hugepg_tbl[page];
814 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
815 /* if this page was already cleared */
816 if (hp->final_va == NULL)
820 /* find a page that matches the criteria */
821 if ((hp->size == hpi[size].hugepage_sz) &&
822 (hp->socket_id == (int) socket)) {
824 /* if we skipped enough pages, unmap the rest */
825 if (pages_found == hpi[size].num_pages[socket]) {
828 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
829 unmap_len = hp->size * hp->repeated;
831 unmap_len = hp->size;
834 /* get start addr and len of the remaining segment */
835 munmap(hp->final_va, (size_t) unmap_len);
838 if (unlink(hp->filepath) == -1) {
839 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
840 __func__, hp->filepath, strerror(errno));
844 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
845 /* else, check how much do we need to map */
848 hpi[size].num_pages[socket] - pages_found;
850 /* if we need enough memory to fit into the segment */
851 if (hp->repeated <= nr_pg_left) {
852 pages_found += hp->repeated;
854 /* truncate the segment */
856 uint64_t final_size = nr_pg_left * hp->size;
857 uint64_t seg_size = hp->repeated * hp->size;
859 void * unmap_va = RTE_PTR_ADD(hp->final_va,
863 munmap(unmap_va, seg_size - final_size);
865 fd = open(hp->filepath, O_RDWR);
867 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n",
868 hp->filepath, strerror(errno));
871 if (ftruncate(fd, final_size) < 0) {
872 RTE_LOG(ERR, EAL, "Cannot truncate %s: %s\n",
873 hp->filepath, strerror(errno));
878 pages_found += nr_pg_left;
879 hp->repeated = nr_pg_left;
883 /* else, lock the page and skip */
890 } /* foreach socket */
891 } /* foreach pagesize */
896 static inline uint64_t
897 get_socket_mem_size(int socket)
902 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
903 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
904 if (hpi->hugedir != NULL)
905 size += hpi->hugepage_sz * hpi->num_pages[socket];
912 * This function is a NUMA-aware equivalent of calc_num_pages.
913 * It takes in the list of hugepage sizes and the
914 * number of pages thereof, and calculates the best number of
915 * pages of each size to fulfill the request for <memory> ram
918 calc_num_pages_per_socket(uint64_t * memory,
919 struct hugepage_info *hp_info,
920 struct hugepage_info *hp_used,
921 unsigned num_hp_info)
923 unsigned socket, j, i = 0;
924 unsigned requested, available;
925 int total_num_pages = 0;
926 uint64_t remaining_mem, cur_mem;
927 uint64_t total_mem = internal_config.memory;
929 if (num_hp_info == 0)
932 /* if specific memory amounts per socket weren't requested */
933 if (internal_config.force_sockets == 0) {
934 int cpu_per_socket[RTE_MAX_NUMA_NODES];
935 size_t default_size, total_size;
938 /* Compute number of cores per socket */
939 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
940 RTE_LCORE_FOREACH(lcore_id) {
941 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
945 * Automatically spread requested memory amongst detected sockets according
946 * to number of cores from cpu mask present on each socket
948 total_size = internal_config.memory;
949 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
951 /* Set memory amount per socket */
952 default_size = (internal_config.memory * cpu_per_socket[socket])
955 /* Limit to maximum available memory on socket */
956 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
959 memory[socket] = default_size;
960 total_size -= default_size;
964 * If some memory is remaining, try to allocate it by getting all
965 * available memory from sockets, one after the other
967 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
968 /* take whatever is available */
969 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
973 memory[socket] += default_size;
974 total_size -= default_size;
978 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
979 /* skips if the memory on specific socket wasn't requested */
980 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
981 hp_used[i].hugedir = hp_info[i].hugedir;
982 hp_used[i].num_pages[socket] = RTE_MIN(
983 memory[socket] / hp_info[i].hugepage_sz,
984 hp_info[i].num_pages[socket]);
986 cur_mem = hp_used[i].num_pages[socket] *
987 hp_used[i].hugepage_sz;
989 memory[socket] -= cur_mem;
990 total_mem -= cur_mem;
992 total_num_pages += hp_used[i].num_pages[socket];
994 /* check if we have met all memory requests */
995 if (memory[socket] == 0)
998 /* check if we have any more pages left at this size, if so
999 * move on to next size */
1000 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
1002 /* At this point we know that there are more pages available that are
1003 * bigger than the memory we want, so lets see if we can get enough
1004 * from other page sizes.
1007 for (j = i+1; j < num_hp_info; j++)
1008 remaining_mem += hp_info[j].hugepage_sz *
1009 hp_info[j].num_pages[socket];
1011 /* is there enough other memory, if not allocate another page and quit */
1012 if (remaining_mem < memory[socket]){
1013 cur_mem = RTE_MIN(memory[socket],
1014 hp_info[i].hugepage_sz);
1015 memory[socket] -= cur_mem;
1016 total_mem -= cur_mem;
1017 hp_used[i].num_pages[socket]++;
1019 break; /* we are done with this socket*/
1022 /* if we didn't satisfy all memory requirements per socket */
1023 if (memory[socket] > 0) {
1024 /* to prevent icc errors */
1025 requested = (unsigned) (internal_config.socket_mem[socket] /
1027 available = requested -
1028 ((unsigned) (memory[socket] / 0x100000));
1029 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
1030 "Requested: %uMB, available: %uMB\n", socket,
1031 requested, available);
1036 /* if we didn't satisfy total memory requirements */
1037 if (total_mem > 0) {
1038 requested = (unsigned) (internal_config.memory / 0x100000);
1039 available = requested - (unsigned) (total_mem / 0x100000);
1040 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
1041 " available: %uMB\n", requested, available);
1044 return total_num_pages;
1048 * Prepare physical memory mapping: fill configuration structure with
1049 * these infos, return 0 on success.
1050 * 1. map N huge pages in separate files in hugetlbfs
1051 * 2. find associated physical addr
1052 * 3. find associated NUMA socket ID
1053 * 4. sort all huge pages by physical address
1054 * 5. remap these N huge pages in the correct order
1055 * 6. unmap the first mapping
1056 * 7. fill memsegs in configuration with contiguous zones
1059 rte_eal_hugepage_init(void)
1061 struct rte_mem_config *mcfg;
1062 struct hugepage_file *hugepage, *tmp_hp = NULL;
1063 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1065 uint64_t memory[RTE_MAX_NUMA_NODES];
1068 int i, j, new_memseg;
1069 int nr_hugefiles, nr_hugepages = 0;
1071 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1072 int new_pages_count[MAX_HUGEPAGE_SIZES];
1075 test_proc_pagemap_readable();
1077 memset(used_hp, 0, sizeof(used_hp));
1079 /* get pointer to global configuration */
1080 mcfg = rte_eal_get_configuration()->mem_config;
1082 /* hugetlbfs can be disabled */
1083 if (internal_config.no_hugetlbfs) {
1084 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1085 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1086 if (addr == MAP_FAILED) {
1087 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1091 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
1092 mcfg->memseg[0].addr = addr;
1093 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1094 mcfg->memseg[0].len = internal_config.memory;
1095 mcfg->memseg[0].socket_id = 0;
1099 /* check if app runs on Xen Dom0 */
1100 if (internal_config.xen_dom0_support) {
1101 #ifdef RTE_LIBRTE_XEN_DOM0
1102 /* use dom0_mm kernel driver to init memory */
1103 if (rte_xen_dom0_memory_init() < 0)
1110 /* calculate total number of hugepages available. at this point we haven't
1111 * yet started sorting them so they all are on socket 0 */
1112 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1113 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1114 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1116 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1120 * allocate a memory area for hugepage table.
1121 * this isn't shared memory yet. due to the fact that we need some
1122 * processing done on these pages, shared memory will be created
1125 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1129 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1131 hp_offset = 0; /* where we start the current page size entries */
1133 /* map all hugepages and sort them */
1134 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1135 struct hugepage_info *hpi;
1138 * we don't yet mark hugepages as used at this stage, so
1139 * we just map all hugepages available to the system
1140 * all hugepages are still located on socket 0
1142 hpi = &internal_config.hugepage_info[i];
1144 if (hpi->num_pages[0] == 0)
1147 /* map all hugepages available */
1148 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
1149 RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
1150 (unsigned)(hpi->hugepage_sz / 0x100000));
1154 /* find physical addresses and sockets for each hugepage */
1155 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0){
1156 RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
1157 (unsigned)(hpi->hugepage_sz / 0x100000));
1161 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1162 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1163 (unsigned)(hpi->hugepage_sz / 0x100000));
1167 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
1170 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1171 /* remap all hugepages into single file segments */
1172 new_pages_count[i] = remap_all_hugepages(&tmp_hp[hp_offset], hpi);
1173 if (new_pages_count[i] < 0){
1174 RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1175 (unsigned)(hpi->hugepage_sz / 0x100000));
1179 /* we have processed a num of hugepages of this size, so inc offset */
1180 hp_offset += new_pages_count[i];
1182 /* remap all hugepages */
1183 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
1184 RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1185 (unsigned)(hpi->hugepage_sz / 0x100000));
1189 /* unmap original mappings */
1190 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1193 /* we have processed a num of hugepages of this size, so inc offset */
1194 hp_offset += hpi->num_pages[0];
1198 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1200 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1201 nr_hugefiles += new_pages_count[i];
1204 nr_hugefiles = nr_hugepages;
1208 /* clean out the numbers of pages */
1209 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1210 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1211 internal_config.hugepage_info[i].num_pages[j] = 0;
1213 /* get hugepages for each socket */
1214 for (i = 0; i < nr_hugefiles; i++) {
1215 int socket = tmp_hp[i].socket_id;
1217 /* find a hugepage info with right size and increment num_pages */
1218 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1219 (int)internal_config.num_hugepage_sizes);
1220 for (j = 0; j < nb_hpsizes; j++) {
1221 if (tmp_hp[i].size ==
1222 internal_config.hugepage_info[j].hugepage_sz) {
1223 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1224 internal_config.hugepage_info[j].num_pages[socket] +=
1227 internal_config.hugepage_info[j].num_pages[socket]++;
1233 /* make a copy of socket_mem, needed for number of pages calculation */
1234 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1235 memory[i] = internal_config.socket_mem[i];
1237 /* calculate final number of pages */
1238 nr_hugepages = calc_num_pages_per_socket(memory,
1239 internal_config.hugepage_info, used_hp,
1240 internal_config.num_hugepage_sizes);
1242 /* error if not enough memory available */
1243 if (nr_hugepages < 0)
1247 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1248 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1249 if (used_hp[i].num_pages[j] > 0) {
1251 "Requesting %u pages of size %uMB"
1252 " from socket %i\n",
1253 used_hp[i].num_pages[j],
1255 (used_hp[i].hugepage_sz / 0x100000),
1261 /* create shared memory */
1262 hugepage = create_shared_memory(eal_hugepage_info_path(),
1263 nr_hugefiles * sizeof(struct hugepage_file));
1265 if (hugepage == NULL) {
1266 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1269 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1272 * unmap pages that we won't need (looks at used_hp).
1273 * also, sets final_va to NULL on pages that were unmapped.
1275 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1276 internal_config.num_hugepage_sizes) < 0) {
1277 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1282 * copy stuff from malloc'd hugepage* to the actual shared memory.
1283 * this procedure only copies those hugepages that have final_va
1284 * not NULL. has overflow protection.
1286 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1287 tmp_hp, nr_hugefiles) < 0) {
1288 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1292 /* free the temporary hugepage table */
1296 /* find earliest free memseg - this is needed because in case of IVSHMEM,
1297 * segments might have already been initialized */
1298 for (j = 0; j < RTE_MAX_MEMSEG; j++)
1299 if (mcfg->memseg[j].addr == NULL) {
1300 /* move to previous segment and exit loop */
1305 for (i = 0; i < nr_hugefiles; i++) {
1308 /* if this is a new section, create a new memseg */
1311 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1313 else if (hugepage[i].size != hugepage[i-1].size)
1316 #ifdef RTE_ARCH_PPC_64
1317 /* On PPC64 architecture, the mmap always start from higher
1318 * virtual address to lower address. Here, both the physical
1319 * address and virtual address are in descending order */
1320 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1323 else if (((unsigned long)hugepage[i-1].final_va -
1324 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1327 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1330 else if (((unsigned long)hugepage[i].final_va -
1331 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1337 if (j == RTE_MAX_MEMSEG)
1340 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1341 mcfg->memseg[j].addr = hugepage[i].final_va;
1342 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1343 mcfg->memseg[j].len = hugepage[i].size * hugepage[i].repeated;
1345 mcfg->memseg[j].len = hugepage[i].size;
1347 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1348 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1350 /* continuation of previous memseg */
1352 #ifdef RTE_ARCH_PPC_64
1353 /* Use the phy and virt address of the last page as segment
1354 * address for IBM Power architecture */
1355 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1356 mcfg->memseg[j].addr = hugepage[i].final_va;
1358 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1360 hugepage[i].memseg_id = j;
1363 if (i < nr_hugefiles) {
1364 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1365 "from %d requested\n"
1366 "Current %s=%d is not enough\n"
1367 "Please either increase it or request less amount "
1369 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1383 * uses fstat to report the size of a file on disk
1389 if (fstat(fd, &st) < 0)
1395 * This creates the memory mappings in the secondary process to match that of
1396 * the server process. It goes through each memory segment in the DPDK runtime
1397 * configuration and finds the hugepages which form that segment, mapping them
1398 * in order to form a contiguous block in the virtual memory space
1401 rte_eal_hugepage_attach(void)
1403 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1404 const struct hugepage_file *hp = NULL;
1405 unsigned num_hp = 0;
1406 unsigned i, s = 0; /* s used to track the segment number */
1408 int fd, fd_zero = -1, fd_hugepage = -1;
1410 if (aslr_enabled() > 0) {
1411 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1412 "(ASLR) is enabled in the kernel.\n");
1413 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1414 "into secondary processes\n");
1417 test_proc_pagemap_readable();
1419 if (internal_config.xen_dom0_support) {
1420 #ifdef RTE_LIBRTE_XEN_DOM0
1421 if (rte_xen_dom0_memory_attach() < 0) {
1422 RTE_LOG(ERR, EAL,"Failed to attach memory setments of primay "
1430 fd_zero = open("/dev/zero", O_RDONLY);
1432 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1435 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1436 if (fd_hugepage < 0) {
1437 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1441 /* map all segments into memory to make sure we get the addrs */
1442 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1446 * the first memory segment with len==0 is the one that
1447 * follows the last valid segment.
1449 if (mcfg->memseg[s].len == 0)
1452 #ifdef RTE_LIBRTE_IVSHMEM
1454 * if segment has ioremap address set, it's an IVSHMEM segment and
1455 * doesn't need mapping as it was already mapped earlier
1457 if (mcfg->memseg[s].ioremap_addr != 0)
1462 * fdzero is mmapped to get a contiguous block of virtual
1463 * addresses of the appropriate memseg size.
1464 * use mmap to get identical addresses as the primary process.
1466 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1467 PROT_READ, MAP_PRIVATE, fd_zero, 0);
1468 if (base_addr == MAP_FAILED ||
1469 base_addr != mcfg->memseg[s].addr) {
1470 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1471 "in /dev/zero to requested address [%p]: '%s'\n",
1472 (unsigned long long)mcfg->memseg[s].len,
1473 mcfg->memseg[s].addr, strerror(errno));
1474 if (aslr_enabled() > 0) {
1475 RTE_LOG(ERR, EAL, "It is recommended to "
1476 "disable ASLR in the kernel "
1477 "and retry running both primary "
1478 "and secondary processes\n");
1484 size = getFileSize(fd_hugepage);
1485 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1487 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1491 num_hp = size / sizeof(struct hugepage_file);
1492 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1495 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1496 void *addr, *base_addr;
1497 uintptr_t offset = 0;
1498 size_t mapping_size;
1499 #ifdef RTE_LIBRTE_IVSHMEM
1501 * if segment has ioremap address set, it's an IVSHMEM segment and
1502 * doesn't need mapping as it was already mapped earlier
1504 if (mcfg->memseg[s].ioremap_addr != 0) {
1510 * free previously mapped memory so we can map the
1511 * hugepages into the space
1513 base_addr = mcfg->memseg[s].addr;
1514 munmap(base_addr, mcfg->memseg[s].len);
1516 /* find the hugepages for this segment and map them
1517 * we don't need to worry about order, as the server sorted the
1518 * entries before it did the second mmap of them */
1519 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1520 if (hp[i].memseg_id == (int)s){
1521 fd = open(hp[i].filepath, O_RDWR);
1523 RTE_LOG(ERR, EAL, "Could not open %s\n",
1527 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1528 mapping_size = hp[i].size * hp[i].repeated;
1530 mapping_size = hp[i].size;
1532 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1533 mapping_size, PROT_READ | PROT_WRITE,
1535 close(fd); /* close file both on success and on failure */
1536 if (addr == MAP_FAILED ||
1537 addr != RTE_PTR_ADD(base_addr, offset)) {
1538 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1542 offset+=mapping_size;
1545 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1546 (unsigned long long)mcfg->memseg[s].len);
1549 /* unmap the hugepage config file, since we are done using it */
1550 munmap((void *)(uintptr_t)hp, size);
1558 if (fd_hugepage >= 0)