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(DEBUG, 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(ERR, EAL, "Cannot get a virtual area: %s\n",
269 munmap(addr, (*size) + hugepage_sz);
272 /* align addr to a huge page size boundary */
273 aligned_addr = (long)addr;
274 aligned_addr += (hugepage_sz - 1);
275 aligned_addr &= (~(hugepage_sz - 1));
276 addr = (void *)(aligned_addr);
278 RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
281 /* increment offset */
282 baseaddr_offset += *size;
288 * Mmap all hugepages of hugepage table: it first open a file in
289 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
290 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
291 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
292 * map continguous physical blocks in contiguous virtual blocks.
295 map_all_hugepages(struct hugepage_file *hugepg_tbl,
296 struct hugepage_info *hpi, int orig)
301 void *vma_addr = NULL;
304 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
305 RTE_SET_USED(vma_len);
308 for (i = 0; i < hpi->num_pages[0]; i++) {
309 uint64_t hugepage_sz = hpi->hugepage_sz;
312 hugepg_tbl[i].file_id = i;
313 hugepg_tbl[i].size = hugepage_sz;
314 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
315 eal_get_hugefile_temp_path(hugepg_tbl[i].filepath,
316 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
317 hugepg_tbl[i].file_id);
319 eal_get_hugefile_path(hugepg_tbl[i].filepath,
320 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
321 hugepg_tbl[i].file_id);
323 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
326 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
327 * original map address as final map address.
329 else if ((hugepage_sz == RTE_PGSIZE_1G)
330 || (hugepage_sz == RTE_PGSIZE_16G)) {
331 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
332 hugepg_tbl[i].orig_va = NULL;
337 #ifndef RTE_EAL_SINGLE_FILE_SEGMENTS
338 else if (vma_len == 0) {
339 unsigned j, num_pages;
341 /* reserve a virtual area for next contiguous
342 * physical block: count the number of
343 * contiguous physical pages. */
344 for (j = i+1; j < hpi->num_pages[0] ; j++) {
345 #ifdef RTE_ARCH_PPC_64
346 /* The physical addresses are sorted in
347 * descending order on PPC64 */
348 if (hugepg_tbl[j].physaddr !=
349 hugepg_tbl[j-1].physaddr - hugepage_sz)
352 if (hugepg_tbl[j].physaddr !=
353 hugepg_tbl[j-1].physaddr + hugepage_sz)
358 vma_len = num_pages * hugepage_sz;
360 /* get the biggest virtual memory area up to
361 * vma_len. If it fails, vma_addr is NULL, so
362 * let the kernel provide the address. */
363 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
364 if (vma_addr == NULL)
365 vma_len = hugepage_sz;
369 /* try to create hugepage file */
370 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
372 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
377 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
379 if (virtaddr == MAP_FAILED) {
380 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
387 hugepg_tbl[i].orig_va = virtaddr;
388 memset(virtaddr, 0, hugepage_sz);
391 hugepg_tbl[i].final_va = virtaddr;
394 /* set shared flock on the file. */
395 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
396 RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
397 __func__, strerror(errno));
404 vma_addr = (char *)vma_addr + hugepage_sz;
405 vma_len -= hugepage_sz;
410 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
413 * Remaps all hugepages into single file segments
416 remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
419 unsigned i = 0, j, num_pages, page_idx = 0;
420 void *vma_addr = NULL, *old_addr = NULL, *page_addr = NULL;
422 size_t hugepage_sz = hpi->hugepage_sz;
423 size_t total_size, offset;
424 char filepath[MAX_HUGEPAGE_PATH];
425 phys_addr_t physaddr;
428 while (i < hpi->num_pages[0]) {
431 /* for 32-bit systems, don't remap 1G pages and 16G pages,
432 * just reuse original map address as final map address.
434 if ((hugepage_sz == RTE_PGSIZE_1G)
435 || (hugepage_sz == RTE_PGSIZE_16G)) {
436 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
437 hugepg_tbl[i].orig_va = NULL;
443 /* reserve a virtual area for next contiguous
444 * physical block: count the number of
445 * contiguous physical pages. */
446 for (j = i+1; j < hpi->num_pages[0] ; j++) {
447 #ifdef RTE_ARCH_PPC_64
448 /* The physical addresses are sorted in descending
450 if (hugepg_tbl[j].physaddr !=
451 hugepg_tbl[j-1].physaddr - hugepage_sz)
454 if (hugepg_tbl[j].physaddr !=
455 hugepg_tbl[j-1].physaddr + hugepage_sz)
460 vma_len = num_pages * hugepage_sz;
462 socket = hugepg_tbl[i].socket_id;
464 /* get the biggest virtual memory area up to
465 * vma_len. If it fails, vma_addr is NULL, so
466 * let the kernel provide the address. */
467 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
469 /* If we can't find a big enough virtual area, work out how many pages
470 * we are going to get */
471 if (vma_addr == NULL)
473 else if (vma_len != num_pages * hugepage_sz) {
474 num_pages = vma_len / hugepage_sz;
479 hugepg_tbl[page_idx].file_id = page_idx;
480 eal_get_hugefile_path(filepath,
483 hugepg_tbl[page_idx].file_id);
485 /* try to create hugepage file */
486 fd = open(filepath, O_CREAT | O_RDWR, 0755);
488 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__, strerror(errno));
495 /* unmap current segment */
497 munmap(vma_addr, total_size);
499 /* unmap original page */
500 munmap(hugepg_tbl[i].orig_va, hugepage_sz);
501 unlink(hugepg_tbl[i].filepath);
503 total_size += hugepage_sz;
507 /* map new, bigger segment */
508 vma_addr = mmap(vma_addr, total_size,
509 PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
511 if (vma_addr == MAP_FAILED || vma_addr != old_addr) {
512 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__, strerror(errno));
517 /* touch the page. this is needed because kernel postpones mapping
518 * creation until the first page fault. with this, we pin down
519 * the page and it is marked as used and gets into process' pagemap.
521 for (offset = 0; offset < total_size; offset += hugepage_sz)
522 *((volatile uint8_t*) RTE_PTR_ADD(vma_addr, offset));
525 /* set shared flock on the file. */
526 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
527 RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
528 __func__, strerror(errno));
533 snprintf(hugepg_tbl[page_idx].filepath, MAX_HUGEPAGE_PATH, "%s",
536 physaddr = rte_mem_virt2phy(vma_addr);
538 if (physaddr == RTE_BAD_PHYS_ADDR)
541 hugepg_tbl[page_idx].final_va = vma_addr;
543 hugepg_tbl[page_idx].physaddr = physaddr;
545 hugepg_tbl[page_idx].repeated = num_pages;
547 hugepg_tbl[page_idx].socket_id = socket;
551 /* verify the memory segment - that is, check that every VA corresponds
552 * to the physical address we expect to see
554 for (offset = 0; offset < vma_len; offset += hugepage_sz) {
555 uint64_t expected_physaddr;
557 expected_physaddr = hugepg_tbl[page_idx].physaddr + offset;
558 page_addr = RTE_PTR_ADD(vma_addr, offset);
559 physaddr = rte_mem_virt2phy(page_addr);
561 if (physaddr != expected_physaddr) {
562 RTE_LOG(ERR, EAL, "Segment sanity check failed: wrong physaddr "
563 "at %p (offset 0x%" PRIx64 ": 0x%" PRIx64
564 " (expected 0x%" PRIx64 ")\n",
565 page_addr, offset, physaddr, expected_physaddr);
570 /* zero out the whole segment */
571 memset(hugepg_tbl[page_idx].final_va, 0, total_size);
576 /* zero out the rest */
577 memset(&hugepg_tbl[page_idx], 0, (hpi->num_pages[0] - page_idx) * sizeof(struct hugepage_file));
580 #else/* RTE_EAL_SINGLE_FILE_SEGMENTS=n */
582 /* Unmap all hugepages from original mapping */
584 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
587 for (i = 0; i < hpi->num_pages[0]; i++) {
588 if (hugepg_tbl[i].orig_va) {
589 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
590 hugepg_tbl[i].orig_va = NULL;
595 #endif /* RTE_EAL_SINGLE_FILE_SEGMENTS */
598 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
602 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
606 unsigned i, hp_count = 0;
609 char hugedir_str[PATH_MAX];
612 f = fopen("/proc/self/numa_maps", "r");
614 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
615 " consider that all memory is in socket_id 0\n");
619 snprintf(hugedir_str, sizeof(hugedir_str),
620 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
623 while (fgets(buf, sizeof(buf), f) != NULL) {
625 /* ignore non huge page */
626 if (strstr(buf, " huge ") == NULL &&
627 strstr(buf, hugedir_str) == NULL)
631 virt_addr = strtoull(buf, &end, 16);
632 if (virt_addr == 0 || end == buf) {
633 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
637 /* get node id (socket id) */
638 nodestr = strstr(buf, " N");
639 if (nodestr == NULL) {
640 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
644 end = strstr(nodestr, "=");
646 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
652 socket_id = strtoul(nodestr, &end, 0);
653 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
654 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
658 /* if we find this page in our mappings, set socket_id */
659 for (i = 0; i < hpi->num_pages[0]; i++) {
660 void *va = (void *)(unsigned long)virt_addr;
661 if (hugepg_tbl[i].orig_va == va) {
662 hugepg_tbl[i].socket_id = socket_id;
668 if (hp_count < hpi->num_pages[0])
680 * Sort the hugepg_tbl by physical address (lower addresses first on x86,
681 * higher address first on powerpc). We use a slow algorithm, but we won't
682 * have millions of pages, and this is only done at init time.
685 sort_by_physaddr(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
689 uint64_t compare_addr;
690 struct hugepage_file tmp;
692 for (i = 0; i < hpi->num_pages[0]; i++) {
697 * browse all entries starting at 'i', and find the
698 * entry with the smallest addr
700 for (j=i; j< hpi->num_pages[0]; j++) {
702 if (compare_addr == 0 ||
703 #ifdef RTE_ARCH_PPC_64
704 hugepg_tbl[j].physaddr > compare_addr) {
706 hugepg_tbl[j].physaddr < compare_addr) {
708 compare_addr = hugepg_tbl[j].physaddr;
713 /* should not happen */
714 if (compare_idx == -1) {
715 RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
719 /* swap the 2 entries in the table */
720 memcpy(&tmp, &hugepg_tbl[compare_idx],
721 sizeof(struct hugepage_file));
722 memcpy(&hugepg_tbl[compare_idx], &hugepg_tbl[i],
723 sizeof(struct hugepage_file));
724 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage_file));
730 * Uses mmap to create a shared memory area for storage of data
731 * Used in this file to store the hugepage file map on disk
734 create_shared_memory(const char *filename, const size_t mem_size)
737 int fd = open(filename, O_CREAT | O_RDWR, 0666);
740 if (ftruncate(fd, mem_size) < 0) {
744 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
750 * this copies *active* hugepages from one hugepage table to another.
751 * destination is typically the shared memory.
754 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
755 const struct hugepage_file * src, int src_size)
757 int src_pos, dst_pos = 0;
759 for (src_pos = 0; src_pos < src_size; src_pos++) {
760 if (src[src_pos].final_va != NULL) {
761 /* error on overflow attempt */
762 if (dst_pos == dest_size)
764 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
772 * unmaps hugepages that are not going to be used. since we originally allocate
773 * ALL hugepages (not just those we need), additional unmapping needs to be done.
776 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
777 struct hugepage_info *hpi,
778 unsigned num_hp_info)
780 unsigned socket, size;
781 int page, nrpages = 0;
783 /* get total number of hugepages */
784 for (size = 0; size < num_hp_info; size++)
785 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
786 nrpages += internal_config.hugepage_info[size].num_pages[socket];
788 for (size = 0; size < num_hp_info; size++) {
789 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
790 unsigned pages_found = 0;
792 /* traverse until we have unmapped all the unused pages */
793 for (page = 0; page < nrpages; page++) {
794 struct hugepage_file *hp = &hugepg_tbl[page];
796 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
797 /* if this page was already cleared */
798 if (hp->final_va == NULL)
802 /* find a page that matches the criteria */
803 if ((hp->size == hpi[size].hugepage_sz) &&
804 (hp->socket_id == (int) socket)) {
806 /* if we skipped enough pages, unmap the rest */
807 if (pages_found == hpi[size].num_pages[socket]) {
810 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
811 unmap_len = hp->size * hp->repeated;
813 unmap_len = hp->size;
816 /* get start addr and len of the remaining segment */
817 munmap(hp->final_va, (size_t) unmap_len);
820 if (unlink(hp->filepath) == -1) {
821 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
822 __func__, hp->filepath, strerror(errno));
826 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
827 /* else, check how much do we need to map */
830 hpi[size].num_pages[socket] - pages_found;
832 /* if we need enough memory to fit into the segment */
833 if (hp->repeated <= nr_pg_left) {
834 pages_found += hp->repeated;
836 /* truncate the segment */
838 uint64_t final_size = nr_pg_left * hp->size;
839 uint64_t seg_size = hp->repeated * hp->size;
841 void * unmap_va = RTE_PTR_ADD(hp->final_va,
845 munmap(unmap_va, seg_size - final_size);
847 fd = open(hp->filepath, O_RDWR);
849 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n",
850 hp->filepath, strerror(errno));
853 if (ftruncate(fd, final_size) < 0) {
854 RTE_LOG(ERR, EAL, "Cannot truncate %s: %s\n",
855 hp->filepath, strerror(errno));
860 pages_found += nr_pg_left;
861 hp->repeated = nr_pg_left;
865 /* else, lock the page and skip */
872 } /* foreach socket */
873 } /* foreach pagesize */
878 static inline uint64_t
879 get_socket_mem_size(int socket)
884 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
885 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
886 if (hpi->hugedir != NULL)
887 size += hpi->hugepage_sz * hpi->num_pages[socket];
894 * This function is a NUMA-aware equivalent of calc_num_pages.
895 * It takes in the list of hugepage sizes and the
896 * number of pages thereof, and calculates the best number of
897 * pages of each size to fulfill the request for <memory> ram
900 calc_num_pages_per_socket(uint64_t * memory,
901 struct hugepage_info *hp_info,
902 struct hugepage_info *hp_used,
903 unsigned num_hp_info)
905 unsigned socket, j, i = 0;
906 unsigned requested, available;
907 int total_num_pages = 0;
908 uint64_t remaining_mem, cur_mem;
909 uint64_t total_mem = internal_config.memory;
911 if (num_hp_info == 0)
914 /* if specific memory amounts per socket weren't requested */
915 if (internal_config.force_sockets == 0) {
916 int cpu_per_socket[RTE_MAX_NUMA_NODES];
917 size_t default_size, total_size;
920 /* Compute number of cores per socket */
921 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
922 RTE_LCORE_FOREACH(lcore_id) {
923 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
927 * Automatically spread requested memory amongst detected sockets according
928 * to number of cores from cpu mask present on each socket
930 total_size = internal_config.memory;
931 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
933 /* Set memory amount per socket */
934 default_size = (internal_config.memory * cpu_per_socket[socket])
937 /* Limit to maximum available memory on socket */
938 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
941 memory[socket] = default_size;
942 total_size -= default_size;
946 * If some memory is remaining, try to allocate it by getting all
947 * available memory from sockets, one after the other
949 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
950 /* take whatever is available */
951 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
955 memory[socket] += default_size;
956 total_size -= default_size;
960 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
961 /* skips if the memory on specific socket wasn't requested */
962 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
963 hp_used[i].hugedir = hp_info[i].hugedir;
964 hp_used[i].num_pages[socket] = RTE_MIN(
965 memory[socket] / hp_info[i].hugepage_sz,
966 hp_info[i].num_pages[socket]);
968 cur_mem = hp_used[i].num_pages[socket] *
969 hp_used[i].hugepage_sz;
971 memory[socket] -= cur_mem;
972 total_mem -= cur_mem;
974 total_num_pages += hp_used[i].num_pages[socket];
976 /* check if we have met all memory requests */
977 if (memory[socket] == 0)
980 /* check if we have any more pages left at this size, if so
981 * move on to next size */
982 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
984 /* At this point we know that there are more pages available that are
985 * bigger than the memory we want, so lets see if we can get enough
986 * from other page sizes.
989 for (j = i+1; j < num_hp_info; j++)
990 remaining_mem += hp_info[j].hugepage_sz *
991 hp_info[j].num_pages[socket];
993 /* is there enough other memory, if not allocate another page and quit */
994 if (remaining_mem < memory[socket]){
995 cur_mem = RTE_MIN(memory[socket],
996 hp_info[i].hugepage_sz);
997 memory[socket] -= cur_mem;
998 total_mem -= cur_mem;
999 hp_used[i].num_pages[socket]++;
1001 break; /* we are done with this socket*/
1004 /* if we didn't satisfy all memory requirements per socket */
1005 if (memory[socket] > 0) {
1006 /* to prevent icc errors */
1007 requested = (unsigned) (internal_config.socket_mem[socket] /
1009 available = requested -
1010 ((unsigned) (memory[socket] / 0x100000));
1011 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
1012 "Requested: %uMB, available: %uMB\n", socket,
1013 requested, available);
1018 /* if we didn't satisfy total memory requirements */
1019 if (total_mem > 0) {
1020 requested = (unsigned) (internal_config.memory / 0x100000);
1021 available = requested - (unsigned) (total_mem / 0x100000);
1022 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
1023 " available: %uMB\n", requested, available);
1026 return total_num_pages;
1030 * Prepare physical memory mapping: fill configuration structure with
1031 * these infos, return 0 on success.
1032 * 1. map N huge pages in separate files in hugetlbfs
1033 * 2. find associated physical addr
1034 * 3. find associated NUMA socket ID
1035 * 4. sort all huge pages by physical address
1036 * 5. remap these N huge pages in the correct order
1037 * 6. unmap the first mapping
1038 * 7. fill memsegs in configuration with contiguous zones
1041 rte_eal_hugepage_init(void)
1043 struct rte_mem_config *mcfg;
1044 struct hugepage_file *hugepage, *tmp_hp = NULL;
1045 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1047 uint64_t memory[RTE_MAX_NUMA_NODES];
1050 int i, j, new_memseg;
1051 int nr_hugefiles, nr_hugepages = 0;
1053 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1054 int new_pages_count[MAX_HUGEPAGE_SIZES];
1057 memset(used_hp, 0, sizeof(used_hp));
1059 /* get pointer to global configuration */
1060 mcfg = rte_eal_get_configuration()->mem_config;
1062 /* hugetlbfs can be disabled */
1063 if (internal_config.no_hugetlbfs) {
1064 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1065 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1066 if (addr == MAP_FAILED) {
1067 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1071 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
1072 mcfg->memseg[0].addr = addr;
1073 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1074 mcfg->memseg[0].len = internal_config.memory;
1075 mcfg->memseg[0].socket_id = 0;
1079 /* check if app runs on Xen Dom0 */
1080 if (internal_config.xen_dom0_support) {
1081 #ifdef RTE_LIBRTE_XEN_DOM0
1082 /* use dom0_mm kernel driver to init memory */
1083 if (rte_xen_dom0_memory_init() < 0)
1091 /* calculate total number of hugepages available. at this point we haven't
1092 * yet started sorting them so they all are on socket 0 */
1093 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1094 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1095 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1097 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1101 * allocate a memory area for hugepage table.
1102 * this isn't shared memory yet. due to the fact that we need some
1103 * processing done on these pages, shared memory will be created
1106 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1110 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1112 hp_offset = 0; /* where we start the current page size entries */
1114 /* map all hugepages and sort them */
1115 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1116 struct hugepage_info *hpi;
1119 * we don't yet mark hugepages as used at this stage, so
1120 * we just map all hugepages available to the system
1121 * all hugepages are still located on socket 0
1123 hpi = &internal_config.hugepage_info[i];
1125 if (hpi->num_pages[0] == 0)
1128 /* map all hugepages available */
1129 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
1130 RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
1131 (unsigned)(hpi->hugepage_sz / 0x100000));
1135 /* find physical addresses and sockets for each hugepage */
1136 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0){
1137 RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
1138 (unsigned)(hpi->hugepage_sz / 0x100000));
1142 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1143 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1144 (unsigned)(hpi->hugepage_sz / 0x100000));
1148 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
1151 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1152 /* remap all hugepages into single file segments */
1153 new_pages_count[i] = remap_all_hugepages(&tmp_hp[hp_offset], hpi);
1154 if (new_pages_count[i] < 0){
1155 RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1156 (unsigned)(hpi->hugepage_sz / 0x100000));
1160 /* we have processed a num of hugepages of this size, so inc offset */
1161 hp_offset += new_pages_count[i];
1163 /* remap all hugepages */
1164 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
1165 RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1166 (unsigned)(hpi->hugepage_sz / 0x100000));
1170 /* unmap original mappings */
1171 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1174 /* we have processed a num of hugepages of this size, so inc offset */
1175 hp_offset += hpi->num_pages[0];
1179 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1181 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1182 nr_hugefiles += new_pages_count[i];
1185 nr_hugefiles = nr_hugepages;
1189 /* clean out the numbers of pages */
1190 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1191 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1192 internal_config.hugepage_info[i].num_pages[j] = 0;
1194 /* get hugepages for each socket */
1195 for (i = 0; i < nr_hugefiles; i++) {
1196 int socket = tmp_hp[i].socket_id;
1198 /* find a hugepage info with right size and increment num_pages */
1199 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1200 (int)internal_config.num_hugepage_sizes);
1201 for (j = 0; j < nb_hpsizes; j++) {
1202 if (tmp_hp[i].size ==
1203 internal_config.hugepage_info[j].hugepage_sz) {
1204 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1205 internal_config.hugepage_info[j].num_pages[socket] +=
1208 internal_config.hugepage_info[j].num_pages[socket]++;
1214 /* make a copy of socket_mem, needed for number of pages calculation */
1215 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1216 memory[i] = internal_config.socket_mem[i];
1218 /* calculate final number of pages */
1219 nr_hugepages = calc_num_pages_per_socket(memory,
1220 internal_config.hugepage_info, used_hp,
1221 internal_config.num_hugepage_sizes);
1223 /* error if not enough memory available */
1224 if (nr_hugepages < 0)
1228 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1229 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1230 if (used_hp[i].num_pages[j] > 0) {
1232 "Requesting %u pages of size %uMB"
1233 " from socket %i\n",
1234 used_hp[i].num_pages[j],
1236 (used_hp[i].hugepage_sz / 0x100000),
1242 /* create shared memory */
1243 hugepage = create_shared_memory(eal_hugepage_info_path(),
1244 nr_hugefiles * sizeof(struct hugepage_file));
1246 if (hugepage == NULL) {
1247 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1250 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1253 * unmap pages that we won't need (looks at used_hp).
1254 * also, sets final_va to NULL on pages that were unmapped.
1256 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1257 internal_config.num_hugepage_sizes) < 0) {
1258 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1263 * copy stuff from malloc'd hugepage* to the actual shared memory.
1264 * this procedure only copies those hugepages that have final_va
1265 * not NULL. has overflow protection.
1267 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1268 tmp_hp, nr_hugefiles) < 0) {
1269 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1273 /* free the temporary hugepage table */
1277 /* find earliest free memseg - this is needed because in case of IVSHMEM,
1278 * segments might have already been initialized */
1279 for (j = 0; j < RTE_MAX_MEMSEG; j++)
1280 if (mcfg->memseg[j].addr == NULL) {
1281 /* move to previous segment and exit loop */
1286 for (i = 0; i < nr_hugefiles; i++) {
1289 /* if this is a new section, create a new memseg */
1292 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1294 else if (hugepage[i].size != hugepage[i-1].size)
1297 #ifdef RTE_ARCH_PPC_64
1298 /* On PPC64 architecture, the mmap always start from higher
1299 * virtual address to lower address. Here, both the physical
1300 * address and virtual address are in descending order */
1301 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1304 else if (((unsigned long)hugepage[i-1].final_va -
1305 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1308 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1311 else if (((unsigned long)hugepage[i].final_va -
1312 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1318 if (j == RTE_MAX_MEMSEG)
1321 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1322 mcfg->memseg[j].addr = hugepage[i].final_va;
1323 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1324 mcfg->memseg[j].len = hugepage[i].size * hugepage[i].repeated;
1326 mcfg->memseg[j].len = hugepage[i].size;
1328 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1329 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1331 /* continuation of previous memseg */
1333 #ifdef RTE_ARCH_PPC_64
1334 /* Use the phy and virt address of the last page as segment
1335 * address for IBM Power architecture */
1336 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1337 mcfg->memseg[j].addr = hugepage[i].final_va;
1339 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1341 hugepage[i].memseg_id = j;
1344 if (i < nr_hugefiles) {
1345 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1346 "from %d requested\n"
1347 "Current %s=%d is not enough\n"
1348 "Please either increase it or request less amount "
1350 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1364 * uses fstat to report the size of a file on disk
1370 if (fstat(fd, &st) < 0)
1376 * This creates the memory mappings in the secondary process to match that of
1377 * the server process. It goes through each memory segment in the DPDK runtime
1378 * configuration and finds the hugepages which form that segment, mapping them
1379 * in order to form a contiguous block in the virtual memory space
1382 rte_eal_hugepage_attach(void)
1384 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1385 const struct hugepage_file *hp = NULL;
1386 unsigned num_hp = 0;
1387 unsigned i, s = 0; /* s used to track the segment number */
1389 int fd, fd_zero = -1, fd_hugepage = -1;
1391 if (aslr_enabled() > 0) {
1392 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1393 "(ASLR) is enabled in the kernel.\n");
1394 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1395 "into secondary processes\n");
1398 if (internal_config.xen_dom0_support) {
1399 #ifdef RTE_LIBRTE_XEN_DOM0
1400 if (rte_xen_dom0_memory_attach() < 0) {
1401 RTE_LOG(ERR, EAL,"Failed to attach memory setments of primay "
1409 fd_zero = open("/dev/zero", O_RDONLY);
1411 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1414 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1415 if (fd_hugepage < 0) {
1416 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1420 /* map all segments into memory to make sure we get the addrs */
1421 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1425 * the first memory segment with len==0 is the one that
1426 * follows the last valid segment.
1428 if (mcfg->memseg[s].len == 0)
1431 #ifdef RTE_LIBRTE_IVSHMEM
1433 * if segment has ioremap address set, it's an IVSHMEM segment and
1434 * doesn't need mapping as it was already mapped earlier
1436 if (mcfg->memseg[s].ioremap_addr != 0)
1441 * fdzero is mmapped to get a contiguous block of virtual
1442 * addresses of the appropriate memseg size.
1443 * use mmap to get identical addresses as the primary process.
1445 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1446 PROT_READ, MAP_PRIVATE, fd_zero, 0);
1447 if (base_addr == MAP_FAILED ||
1448 base_addr != mcfg->memseg[s].addr) {
1449 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1450 "in /dev/zero to requested address [%p]: '%s'\n",
1451 (unsigned long long)mcfg->memseg[s].len,
1452 mcfg->memseg[s].addr, strerror(errno));
1453 if (aslr_enabled() > 0) {
1454 RTE_LOG(ERR, EAL, "It is recommended to "
1455 "disable ASLR in the kernel "
1456 "and retry running both primary "
1457 "and secondary processes\n");
1463 size = getFileSize(fd_hugepage);
1464 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1466 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1470 num_hp = size / sizeof(struct hugepage_file);
1471 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1474 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1475 void *addr, *base_addr;
1476 uintptr_t offset = 0;
1477 size_t mapping_size;
1478 #ifdef RTE_LIBRTE_IVSHMEM
1480 * if segment has ioremap address set, it's an IVSHMEM segment and
1481 * doesn't need mapping as it was already mapped earlier
1483 if (mcfg->memseg[s].ioremap_addr != 0) {
1489 * free previously mapped memory so we can map the
1490 * hugepages into the space
1492 base_addr = mcfg->memseg[s].addr;
1493 munmap(base_addr, mcfg->memseg[s].len);
1495 /* find the hugepages for this segment and map them
1496 * we don't need to worry about order, as the server sorted the
1497 * entries before it did the second mmap of them */
1498 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1499 if (hp[i].memseg_id == (int)s){
1500 fd = open(hp[i].filepath, O_RDWR);
1502 RTE_LOG(ERR, EAL, "Could not open %s\n",
1506 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1507 mapping_size = hp[i].size * hp[i].repeated;
1509 mapping_size = hp[i].size;
1511 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1512 mapping_size, PROT_READ | PROT_WRITE,
1514 close(fd); /* close file both on success and on failure */
1515 if (addr == MAP_FAILED ||
1516 addr != RTE_PTR_ADD(base_addr, offset)) {
1517 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1521 offset+=mapping_size;
1524 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1525 (unsigned long long)mcfg->memseg[s].len);
1528 /* unmap the hugepage config file, since we are done using it */
1529 munmap((void *)(uintptr_t)hp, size);
1537 if (fd_hugepage >= 0)
1543 rte_eal_memdevice_init(void)
1545 struct rte_config *config;
1547 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1550 config = rte_eal_get_configuration();
1551 config->mem_config->nchannel = internal_config.force_nchannel;
1552 config->mem_config->nrank = internal_config.force_nrank;
1558 test_proc_pagemap_readable(void)
1560 int fd = open("/proc/self/pagemap", O_RDONLY);
1570 /* init memory subsystem */
1572 rte_eal_memory_init(void)
1574 RTE_LOG(INFO, EAL, "Setting up memory...\n");
1576 proc_pagemap_readable = test_proc_pagemap_readable();
1577 if (!proc_pagemap_readable)
1579 "Cannot open /proc/self/pagemap: %s. "
1580 "virt2phys address translation will not work\n",
1583 const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
1584 rte_eal_hugepage_init() :
1585 rte_eal_hugepage_attach();
1589 if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)