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>
87 #include <rte_memory.h>
88 #include <rte_memzone.h>
89 #include <rte_launch.h>
91 #include <rte_eal_memconfig.h>
92 #include <rte_per_lcore.h>
93 #include <rte_lcore.h>
94 #include <rte_common.h>
95 #include <rte_string_fns.h>
97 #include "eal_private.h"
98 #include "eal_internal_cfg.h"
99 #include "eal_filesystem.h"
100 #include "eal_hugepages.h"
102 #define PFN_MASK_SIZE 8
104 #ifdef RTE_LIBRTE_XEN_DOM0
105 int rte_xen_dom0_supported(void)
107 return internal_config.xen_dom0_support;
113 * Huge page mapping under linux
115 * To reserve a big contiguous amount of memory, we use the hugepage
116 * feature of linux. For that, we need to have hugetlbfs mounted. This
117 * code will create many files in this directory (one per page) and
118 * map them in virtual memory. For each page, we will retrieve its
119 * physical address and remap it in order to have a virtual contiguous
120 * zone as well as a physical contiguous zone.
123 static uint64_t baseaddr_offset;
125 static unsigned proc_pagemap_readable;
127 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
130 test_proc_pagemap_readable(void)
132 int fd = open("/proc/self/pagemap", O_RDONLY);
136 "Cannot open /proc/self/pagemap: %s. "
137 "virt2phys address translation will not work\n",
144 proc_pagemap_readable = 1;
147 /* Lock page in physical memory and prevent from swapping. */
149 rte_mem_lock_page(const void *virt)
151 unsigned long virtual = (unsigned long)virt;
152 int page_size = getpagesize();
153 unsigned long aligned = (virtual & ~ (page_size - 1));
154 return mlock((void*)aligned, page_size);
158 * Get physical address of any mapped virtual address in the current process.
161 rte_mem_virt2phy(const void *virtaddr)
164 uint64_t page, physaddr;
165 unsigned long virt_pfn;
169 /* when using dom0, /proc/self/pagemap always returns 0, check in
170 * dpdk memory by browsing the memsegs */
171 if (rte_xen_dom0_supported()) {
172 struct rte_mem_config *mcfg;
173 struct rte_memseg *memseg;
176 mcfg = rte_eal_get_configuration()->mem_config;
177 for (i = 0; i < RTE_MAX_MEMSEG; i++) {
178 memseg = &mcfg->memseg[i];
179 if (memseg->addr == NULL)
181 if (virtaddr > memseg->addr &&
182 virtaddr < RTE_PTR_ADD(memseg->addr,
184 return memseg->phys_addr +
185 RTE_PTR_DIFF(virtaddr, memseg->addr);
189 return RTE_BAD_PHYS_ADDR;
192 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
193 if (!proc_pagemap_readable)
194 return RTE_BAD_PHYS_ADDR;
196 /* standard page size */
197 page_size = getpagesize();
199 fd = open("/proc/self/pagemap", O_RDONLY);
201 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
202 __func__, strerror(errno));
203 return RTE_BAD_PHYS_ADDR;
206 virt_pfn = (unsigned long)virtaddr / page_size;
207 offset = sizeof(uint64_t) * virt_pfn;
208 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
209 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
210 __func__, strerror(errno));
212 return RTE_BAD_PHYS_ADDR;
215 retval = read(fd, &page, PFN_MASK_SIZE);
218 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
219 __func__, strerror(errno));
220 return RTE_BAD_PHYS_ADDR;
221 } else if (retval != PFN_MASK_SIZE) {
222 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
223 "but expected %d:\n",
224 __func__, retval, PFN_MASK_SIZE);
225 return RTE_BAD_PHYS_ADDR;
229 * the pfn (page frame number) are bits 0-54 (see
230 * pagemap.txt in linux Documentation)
232 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
233 + ((unsigned long)virtaddr % page_size);
239 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
240 * it by browsing the /proc/self/pagemap special file.
243 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
248 for (i = 0; i < hpi->num_pages[0]; i++) {
249 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
250 if (addr == RTE_BAD_PHYS_ADDR)
252 hugepg_tbl[i].physaddr = addr;
258 * Check whether address-space layout randomization is enabled in
259 * the kernel. This is important for multi-process as it can prevent
260 * two processes mapping data to the same virtual address
262 * 0 - address space randomization disabled
263 * 1/2 - address space randomization enabled
264 * negative error code on error
270 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
273 retval = read(fd, &c, 1);
283 default: return -EINVAL;
288 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
289 * pointer to the mmap'd area and keep *size unmodified. Else, retry
290 * with a smaller zone: decrease *size by hugepage_sz until it reaches
291 * 0. In this case, return NULL. Note: this function returns an address
292 * which is a multiple of hugepage size.
295 get_virtual_area(size_t *size, size_t hugepage_sz)
301 if (internal_config.base_virtaddr != 0) {
302 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
307 RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
309 fd = open("/dev/zero", O_RDONLY);
311 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
316 (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
317 if (addr == MAP_FAILED)
318 *size -= hugepage_sz;
319 } while (addr == MAP_FAILED && *size > 0);
321 if (addr == MAP_FAILED) {
323 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
328 munmap(addr, (*size) + hugepage_sz);
331 /* align addr to a huge page size boundary */
332 aligned_addr = (long)addr;
333 aligned_addr += (hugepage_sz - 1);
334 aligned_addr &= (~(hugepage_sz - 1));
335 addr = (void *)(aligned_addr);
337 RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
340 /* increment offset */
341 baseaddr_offset += *size;
346 static sigjmp_buf huge_jmpenv;
348 static void huge_sigbus_handler(int signo __rte_unused)
350 siglongjmp(huge_jmpenv, 1);
353 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
354 * non-static local variable in the stack frame calling sigsetjmp might be
355 * clobbered by a call to longjmp.
357 static int huge_wrap_sigsetjmp(void)
359 return sigsetjmp(huge_jmpenv, 1);
363 * Mmap all hugepages of hugepage table: it first open a file in
364 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
365 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
366 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
367 * map continguous physical blocks in contiguous virtual blocks.
370 map_all_hugepages(struct hugepage_file *hugepg_tbl,
371 struct hugepage_info *hpi, int orig)
376 void *vma_addr = NULL;
379 for (i = 0; i < hpi->num_pages[0]; i++) {
380 uint64_t hugepage_sz = hpi->hugepage_sz;
383 hugepg_tbl[i].file_id = i;
384 hugepg_tbl[i].size = hugepage_sz;
385 eal_get_hugefile_path(hugepg_tbl[i].filepath,
386 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
387 hugepg_tbl[i].file_id);
388 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
391 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
392 * original map address as final map address.
394 else if ((hugepage_sz == RTE_PGSIZE_1G)
395 || (hugepage_sz == RTE_PGSIZE_16G)) {
396 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
397 hugepg_tbl[i].orig_va = NULL;
401 else if (vma_len == 0) {
402 unsigned j, num_pages;
404 /* reserve a virtual area for next contiguous
405 * physical block: count the number of
406 * contiguous physical pages. */
407 for (j = i+1; j < hpi->num_pages[0] ; j++) {
408 #ifdef RTE_ARCH_PPC_64
409 /* The physical addresses are sorted in
410 * descending order on PPC64 */
411 if (hugepg_tbl[j].physaddr !=
412 hugepg_tbl[j-1].physaddr - hugepage_sz)
415 if (hugepg_tbl[j].physaddr !=
416 hugepg_tbl[j-1].physaddr + hugepage_sz)
421 vma_len = num_pages * hugepage_sz;
423 /* get the biggest virtual memory area up to
424 * vma_len. If it fails, vma_addr is NULL, so
425 * let the kernel provide the address. */
426 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
427 if (vma_addr == NULL)
428 vma_len = hugepage_sz;
431 /* try to create hugepage file */
432 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
434 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
439 /* map the segment, and populate page tables,
440 * the kernel fills this segment with zeros */
441 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
442 MAP_SHARED | MAP_POPULATE, fd, 0);
443 if (virtaddr == MAP_FAILED) {
444 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
451 hugepg_tbl[i].orig_va = virtaddr;
454 hugepg_tbl[i].final_va = virtaddr;
458 /* In linux, hugetlb limitations, like cgroup, are
459 * enforced at fault time instead of mmap(), even
460 * with the option of MAP_POPULATE. Kernel will send
461 * a SIGBUS signal. To avoid to be killed, save stack
462 * environment here, if SIGBUS happens, we can jump
465 if (huge_wrap_sigsetjmp()) {
466 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
467 "hugepages of size %u MB\n",
468 (unsigned)(hugepage_sz / 0x100000));
469 munmap(virtaddr, hugepage_sz);
471 unlink(hugepg_tbl[i].filepath);
474 *(int *)virtaddr = 0;
478 /* set shared flock on the file. */
479 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
480 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
481 __func__, strerror(errno));
488 vma_addr = (char *)vma_addr + hugepage_sz;
489 vma_len -= hugepage_sz;
495 /* Unmap all hugepages from original mapping */
497 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
500 for (i = 0; i < hpi->num_pages[0]; i++) {
501 if (hugepg_tbl[i].orig_va) {
502 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
503 hugepg_tbl[i].orig_va = NULL;
510 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
514 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
518 unsigned i, hp_count = 0;
521 char hugedir_str[PATH_MAX];
524 f = fopen("/proc/self/numa_maps", "r");
526 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
527 " consider that all memory is in socket_id 0\n");
531 snprintf(hugedir_str, sizeof(hugedir_str),
532 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
535 while (fgets(buf, sizeof(buf), f) != NULL) {
537 /* ignore non huge page */
538 if (strstr(buf, " huge ") == NULL &&
539 strstr(buf, hugedir_str) == NULL)
543 virt_addr = strtoull(buf, &end, 16);
544 if (virt_addr == 0 || end == buf) {
545 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
549 /* get node id (socket id) */
550 nodestr = strstr(buf, " N");
551 if (nodestr == NULL) {
552 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
556 end = strstr(nodestr, "=");
558 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
564 socket_id = strtoul(nodestr, &end, 0);
565 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
566 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
570 /* if we find this page in our mappings, set socket_id */
571 for (i = 0; i < hpi->num_pages[0]; i++) {
572 void *va = (void *)(unsigned long)virt_addr;
573 if (hugepg_tbl[i].orig_va == va) {
574 hugepg_tbl[i].socket_id = socket_id;
580 if (hp_count < hpi->num_pages[0])
592 cmp_physaddr(const void *a, const void *b)
594 #ifndef RTE_ARCH_PPC_64
595 const struct hugepage_file *p1 = (const struct hugepage_file *)a;
596 const struct hugepage_file *p2 = (const struct hugepage_file *)b;
598 /* PowerPC needs memory sorted in reverse order from x86 */
599 const struct hugepage_file *p1 = (const struct hugepage_file *)b;
600 const struct hugepage_file *p2 = (const struct hugepage_file *)a;
602 if (p1->physaddr < p2->physaddr)
604 else if (p1->physaddr > p2->physaddr)
611 * Uses mmap to create a shared memory area for storage of data
612 * Used in this file to store the hugepage file map on disk
615 create_shared_memory(const char *filename, const size_t mem_size)
618 int fd = open(filename, O_CREAT | O_RDWR, 0666);
621 if (ftruncate(fd, mem_size) < 0) {
625 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
631 * this copies *active* hugepages from one hugepage table to another.
632 * destination is typically the shared memory.
635 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
636 const struct hugepage_file * src, int src_size)
638 int src_pos, dst_pos = 0;
640 for (src_pos = 0; src_pos < src_size; src_pos++) {
641 if (src[src_pos].final_va != NULL) {
642 /* error on overflow attempt */
643 if (dst_pos == dest_size)
645 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
653 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
654 unsigned num_hp_info)
656 unsigned socket, size;
657 int page, nrpages = 0;
659 /* get total number of hugepages */
660 for (size = 0; size < num_hp_info; size++)
661 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
663 internal_config.hugepage_info[size].num_pages[socket];
665 for (page = 0; page < nrpages; page++) {
666 struct hugepage_file *hp = &hugepg_tbl[page];
668 if (hp->final_va != NULL && unlink(hp->filepath)) {
669 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
670 __func__, hp->filepath, strerror(errno));
677 * unmaps hugepages that are not going to be used. since we originally allocate
678 * ALL hugepages (not just those we need), additional unmapping needs to be done.
681 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
682 struct hugepage_info *hpi,
683 unsigned num_hp_info)
685 unsigned socket, size;
686 int page, nrpages = 0;
688 /* get total number of hugepages */
689 for (size = 0; size < num_hp_info; size++)
690 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
691 nrpages += internal_config.hugepage_info[size].num_pages[socket];
693 for (size = 0; size < num_hp_info; size++) {
694 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
695 unsigned pages_found = 0;
697 /* traverse until we have unmapped all the unused pages */
698 for (page = 0; page < nrpages; page++) {
699 struct hugepage_file *hp = &hugepg_tbl[page];
701 /* find a page that matches the criteria */
702 if ((hp->size == hpi[size].hugepage_sz) &&
703 (hp->socket_id == (int) socket)) {
705 /* if we skipped enough pages, unmap the rest */
706 if (pages_found == hpi[size].num_pages[socket]) {
709 unmap_len = hp->size;
711 /* get start addr and len of the remaining segment */
712 munmap(hp->final_va, (size_t) unmap_len);
715 if (unlink(hp->filepath) == -1) {
716 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
717 __func__, hp->filepath, strerror(errno));
721 /* lock the page and skip */
727 } /* foreach socket */
728 } /* foreach pagesize */
733 static inline uint64_t
734 get_socket_mem_size(int socket)
739 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
740 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
741 if (hpi->hugedir != NULL)
742 size += hpi->hugepage_sz * hpi->num_pages[socket];
749 * This function is a NUMA-aware equivalent of calc_num_pages.
750 * It takes in the list of hugepage sizes and the
751 * number of pages thereof, and calculates the best number of
752 * pages of each size to fulfill the request for <memory> ram
755 calc_num_pages_per_socket(uint64_t * memory,
756 struct hugepage_info *hp_info,
757 struct hugepage_info *hp_used,
758 unsigned num_hp_info)
760 unsigned socket, j, i = 0;
761 unsigned requested, available;
762 int total_num_pages = 0;
763 uint64_t remaining_mem, cur_mem;
764 uint64_t total_mem = internal_config.memory;
766 if (num_hp_info == 0)
769 /* if specific memory amounts per socket weren't requested */
770 if (internal_config.force_sockets == 0) {
771 int cpu_per_socket[RTE_MAX_NUMA_NODES];
772 size_t default_size, total_size;
775 /* Compute number of cores per socket */
776 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
777 RTE_LCORE_FOREACH(lcore_id) {
778 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
782 * Automatically spread requested memory amongst detected sockets according
783 * to number of cores from cpu mask present on each socket
785 total_size = internal_config.memory;
786 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
788 /* Set memory amount per socket */
789 default_size = (internal_config.memory * cpu_per_socket[socket])
792 /* Limit to maximum available memory on socket */
793 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
796 memory[socket] = default_size;
797 total_size -= default_size;
801 * If some memory is remaining, try to allocate it by getting all
802 * available memory from sockets, one after the other
804 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
805 /* take whatever is available */
806 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
810 memory[socket] += default_size;
811 total_size -= default_size;
815 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
816 /* skips if the memory on specific socket wasn't requested */
817 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
818 hp_used[i].hugedir = hp_info[i].hugedir;
819 hp_used[i].num_pages[socket] = RTE_MIN(
820 memory[socket] / hp_info[i].hugepage_sz,
821 hp_info[i].num_pages[socket]);
823 cur_mem = hp_used[i].num_pages[socket] *
824 hp_used[i].hugepage_sz;
826 memory[socket] -= cur_mem;
827 total_mem -= cur_mem;
829 total_num_pages += hp_used[i].num_pages[socket];
831 /* check if we have met all memory requests */
832 if (memory[socket] == 0)
835 /* check if we have any more pages left at this size, if so
836 * move on to next size */
837 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
839 /* At this point we know that there are more pages available that are
840 * bigger than the memory we want, so lets see if we can get enough
841 * from other page sizes.
844 for (j = i+1; j < num_hp_info; j++)
845 remaining_mem += hp_info[j].hugepage_sz *
846 hp_info[j].num_pages[socket];
848 /* is there enough other memory, if not allocate another page and quit */
849 if (remaining_mem < memory[socket]){
850 cur_mem = RTE_MIN(memory[socket],
851 hp_info[i].hugepage_sz);
852 memory[socket] -= cur_mem;
853 total_mem -= cur_mem;
854 hp_used[i].num_pages[socket]++;
856 break; /* we are done with this socket*/
859 /* if we didn't satisfy all memory requirements per socket */
860 if (memory[socket] > 0) {
861 /* to prevent icc errors */
862 requested = (unsigned) (internal_config.socket_mem[socket] /
864 available = requested -
865 ((unsigned) (memory[socket] / 0x100000));
866 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
867 "Requested: %uMB, available: %uMB\n", socket,
868 requested, available);
873 /* if we didn't satisfy total memory requirements */
875 requested = (unsigned) (internal_config.memory / 0x100000);
876 available = requested - (unsigned) (total_mem / 0x100000);
877 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
878 " available: %uMB\n", requested, available);
881 return total_num_pages;
885 eal_get_hugepage_mem_size(void)
890 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
891 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
892 if (hpi->hugedir != NULL) {
893 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
894 size += hpi->hugepage_sz * hpi->num_pages[j];
899 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
902 static struct sigaction huge_action_old;
903 static int huge_need_recover;
906 huge_register_sigbus(void)
909 struct sigaction action;
912 sigaddset(&mask, SIGBUS);
914 action.sa_mask = mask;
915 action.sa_handler = huge_sigbus_handler;
917 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
921 huge_recover_sigbus(void)
923 if (huge_need_recover) {
924 sigaction(SIGBUS, &huge_action_old, NULL);
925 huge_need_recover = 0;
930 * Prepare physical memory mapping: fill configuration structure with
931 * these infos, return 0 on success.
932 * 1. map N huge pages in separate files in hugetlbfs
933 * 2. find associated physical addr
934 * 3. find associated NUMA socket ID
935 * 4. sort all huge pages by physical address
936 * 5. remap these N huge pages in the correct order
937 * 6. unmap the first mapping
938 * 7. fill memsegs in configuration with contiguous zones
941 rte_eal_hugepage_init(void)
943 struct rte_mem_config *mcfg;
944 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
945 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
947 uint64_t memory[RTE_MAX_NUMA_NODES];
950 int i, j, new_memseg;
951 int nr_hugefiles, nr_hugepages = 0;
954 test_proc_pagemap_readable();
956 memset(used_hp, 0, sizeof(used_hp));
958 /* get pointer to global configuration */
959 mcfg = rte_eal_get_configuration()->mem_config;
961 /* hugetlbfs can be disabled */
962 if (internal_config.no_hugetlbfs) {
963 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
964 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
965 if (addr == MAP_FAILED) {
966 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
970 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
971 mcfg->memseg[0].addr = addr;
972 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
973 mcfg->memseg[0].len = internal_config.memory;
974 mcfg->memseg[0].socket_id = 0;
978 /* check if app runs on Xen Dom0 */
979 if (internal_config.xen_dom0_support) {
980 #ifdef RTE_LIBRTE_XEN_DOM0
981 /* use dom0_mm kernel driver to init memory */
982 if (rte_xen_dom0_memory_init() < 0)
989 /* calculate total number of hugepages available. at this point we haven't
990 * yet started sorting them so they all are on socket 0 */
991 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
992 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
993 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
995 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
999 * allocate a memory area for hugepage table.
1000 * this isn't shared memory yet. due to the fact that we need some
1001 * processing done on these pages, shared memory will be created
1004 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1008 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1010 hp_offset = 0; /* where we start the current page size entries */
1012 huge_register_sigbus();
1014 /* map all hugepages and sort them */
1015 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1016 unsigned pages_old, pages_new;
1017 struct hugepage_info *hpi;
1020 * we don't yet mark hugepages as used at this stage, so
1021 * we just map all hugepages available to the system
1022 * all hugepages are still located on socket 0
1024 hpi = &internal_config.hugepage_info[i];
1026 if (hpi->num_pages[0] == 0)
1029 /* map all hugepages available */
1030 pages_old = hpi->num_pages[0];
1031 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, 1);
1032 if (pages_new < pages_old) {
1034 "%d not %d hugepages of size %u MB allocated\n",
1035 pages_new, pages_old,
1036 (unsigned)(hpi->hugepage_sz / 0x100000));
1038 int pages = pages_old - pages_new;
1040 nr_hugepages -= pages;
1041 hpi->num_pages[0] = pages_new;
1046 /* find physical addresses and sockets for each hugepage */
1047 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0){
1048 RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
1049 (unsigned)(hpi->hugepage_sz / 0x100000));
1053 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1054 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1055 (unsigned)(hpi->hugepage_sz / 0x100000));
1059 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1060 sizeof(struct hugepage_file), cmp_physaddr);
1062 /* remap all hugepages */
1063 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) !=
1064 hpi->num_pages[0]) {
1065 RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1066 (unsigned)(hpi->hugepage_sz / 0x100000));
1070 /* unmap original mappings */
1071 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1074 /* we have processed a num of hugepages of this size, so inc offset */
1075 hp_offset += hpi->num_pages[0];
1078 huge_recover_sigbus();
1080 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1081 internal_config.memory = eal_get_hugepage_mem_size();
1083 nr_hugefiles = nr_hugepages;
1086 /* clean out the numbers of pages */
1087 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1088 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1089 internal_config.hugepage_info[i].num_pages[j] = 0;
1091 /* get hugepages for each socket */
1092 for (i = 0; i < nr_hugefiles; i++) {
1093 int socket = tmp_hp[i].socket_id;
1095 /* find a hugepage info with right size and increment num_pages */
1096 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1097 (int)internal_config.num_hugepage_sizes);
1098 for (j = 0; j < nb_hpsizes; j++) {
1099 if (tmp_hp[i].size ==
1100 internal_config.hugepage_info[j].hugepage_sz) {
1101 internal_config.hugepage_info[j].num_pages[socket]++;
1106 /* make a copy of socket_mem, needed for number of pages calculation */
1107 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1108 memory[i] = internal_config.socket_mem[i];
1110 /* calculate final number of pages */
1111 nr_hugepages = calc_num_pages_per_socket(memory,
1112 internal_config.hugepage_info, used_hp,
1113 internal_config.num_hugepage_sizes);
1115 /* error if not enough memory available */
1116 if (nr_hugepages < 0)
1120 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1121 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1122 if (used_hp[i].num_pages[j] > 0) {
1124 "Requesting %u pages of size %uMB"
1125 " from socket %i\n",
1126 used_hp[i].num_pages[j],
1128 (used_hp[i].hugepage_sz / 0x100000),
1134 /* create shared memory */
1135 hugepage = create_shared_memory(eal_hugepage_info_path(),
1136 nr_hugefiles * sizeof(struct hugepage_file));
1138 if (hugepage == NULL) {
1139 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1142 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1145 * unmap pages that we won't need (looks at used_hp).
1146 * also, sets final_va to NULL on pages that were unmapped.
1148 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1149 internal_config.num_hugepage_sizes) < 0) {
1150 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1155 * copy stuff from malloc'd hugepage* to the actual shared memory.
1156 * this procedure only copies those hugepages that have final_va
1157 * not NULL. has overflow protection.
1159 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1160 tmp_hp, nr_hugefiles) < 0) {
1161 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1165 /* free the hugepage backing files */
1166 if (internal_config.hugepage_unlink &&
1167 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1168 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1172 /* free the temporary hugepage table */
1176 /* first memseg index shall be 0 after incrementing it below */
1178 for (i = 0; i < nr_hugefiles; i++) {
1181 /* if this is a new section, create a new memseg */
1184 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1186 else if (hugepage[i].size != hugepage[i-1].size)
1189 #ifdef RTE_ARCH_PPC_64
1190 /* On PPC64 architecture, the mmap always start from higher
1191 * virtual address to lower address. Here, both the physical
1192 * address and virtual address are in descending order */
1193 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1196 else if (((unsigned long)hugepage[i-1].final_va -
1197 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1200 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1203 else if (((unsigned long)hugepage[i].final_va -
1204 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1210 if (j == RTE_MAX_MEMSEG)
1213 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1214 mcfg->memseg[j].addr = hugepage[i].final_va;
1215 mcfg->memseg[j].len = hugepage[i].size;
1216 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1217 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1219 /* continuation of previous memseg */
1221 #ifdef RTE_ARCH_PPC_64
1222 /* Use the phy and virt address of the last page as segment
1223 * address for IBM Power architecture */
1224 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1225 mcfg->memseg[j].addr = hugepage[i].final_va;
1227 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1229 hugepage[i].memseg_id = j;
1232 if (i < nr_hugefiles) {
1233 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1234 "from %d requested\n"
1235 "Current %s=%d is not enough\n"
1236 "Please either increase it or request less amount "
1238 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1243 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1248 huge_recover_sigbus();
1250 if (hugepage != NULL)
1251 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1257 * uses fstat to report the size of a file on disk
1263 if (fstat(fd, &st) < 0)
1269 * This creates the memory mappings in the secondary process to match that of
1270 * the server process. It goes through each memory segment in the DPDK runtime
1271 * configuration and finds the hugepages which form that segment, mapping them
1272 * in order to form a contiguous block in the virtual memory space
1275 rte_eal_hugepage_attach(void)
1277 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1278 struct hugepage_file *hp = NULL;
1279 unsigned num_hp = 0;
1280 unsigned i, s = 0; /* s used to track the segment number */
1281 unsigned max_seg = RTE_MAX_MEMSEG;
1283 int fd, fd_zero = -1, fd_hugepage = -1;
1285 if (aslr_enabled() > 0) {
1286 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1287 "(ASLR) is enabled in the kernel.\n");
1288 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1289 "into secondary processes\n");
1292 test_proc_pagemap_readable();
1294 if (internal_config.xen_dom0_support) {
1295 #ifdef RTE_LIBRTE_XEN_DOM0
1296 if (rte_xen_dom0_memory_attach() < 0) {
1297 RTE_LOG(ERR, EAL, "Failed to attach memory segments of primary "
1305 fd_zero = open("/dev/zero", O_RDONLY);
1307 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1310 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1311 if (fd_hugepage < 0) {
1312 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1316 /* map all segments into memory to make sure we get the addrs */
1317 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1321 * the first memory segment with len==0 is the one that
1322 * follows the last valid segment.
1324 if (mcfg->memseg[s].len == 0)
1328 * fdzero is mmapped to get a contiguous block of virtual
1329 * addresses of the appropriate memseg size.
1330 * use mmap to get identical addresses as the primary process.
1332 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1333 PROT_READ, MAP_PRIVATE, fd_zero, 0);
1334 if (base_addr == MAP_FAILED ||
1335 base_addr != mcfg->memseg[s].addr) {
1336 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1337 "in /dev/zero to requested address [%p]: '%s'\n",
1338 (unsigned long long)mcfg->memseg[s].len,
1339 mcfg->memseg[s].addr, strerror(errno));
1341 if (base_addr != MAP_FAILED)
1342 munmap(base_addr, mcfg->memseg[s].len);
1343 if (aslr_enabled() > 0) {
1344 RTE_LOG(ERR, EAL, "It is recommended to "
1345 "disable ASLR in the kernel "
1346 "and retry running both primary "
1347 "and secondary processes\n");
1353 size = getFileSize(fd_hugepage);
1354 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1355 if (hp == MAP_FAILED) {
1356 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1360 num_hp = size / sizeof(struct hugepage_file);
1361 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1364 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1365 void *addr, *base_addr;
1366 uintptr_t offset = 0;
1367 size_t mapping_size;
1369 * free previously mapped memory so we can map the
1370 * hugepages into the space
1372 base_addr = mcfg->memseg[s].addr;
1373 munmap(base_addr, mcfg->memseg[s].len);
1375 /* find the hugepages for this segment and map them
1376 * we don't need to worry about order, as the server sorted the
1377 * entries before it did the second mmap of them */
1378 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1379 if (hp[i].memseg_id == (int)s){
1380 fd = open(hp[i].filepath, O_RDWR);
1382 RTE_LOG(ERR, EAL, "Could not open %s\n",
1386 mapping_size = hp[i].size;
1387 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1388 mapping_size, PROT_READ | PROT_WRITE,
1390 close(fd); /* close file both on success and on failure */
1391 if (addr == MAP_FAILED ||
1392 addr != RTE_PTR_ADD(base_addr, offset)) {
1393 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1397 offset+=mapping_size;
1400 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1401 (unsigned long long)mcfg->memseg[s].len);
1404 /* unmap the hugepage config file, since we are done using it */
1411 for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1412 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1413 if (hp != NULL && hp != MAP_FAILED)
1417 if (fd_hugepage >= 0)