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
75 #include <sys/types.h>
77 #include <sys/queue.h>
82 #include <sys/ioctl.h>
88 #include <rte_memory.h>
89 #include <rte_memzone.h>
90 #include <rte_launch.h>
92 #include <rte_eal_memconfig.h>
93 #include <rte_per_lcore.h>
94 #include <rte_lcore.h>
95 #include <rte_common.h>
96 #include <rte_string_fns.h>
98 #include "eal_private.h"
99 #include "eal_internal_cfg.h"
100 #include "eal_filesystem.h"
101 #include "eal_hugepages.h"
103 #define PFN_MASK_SIZE 8
105 #ifdef RTE_LIBRTE_XEN_DOM0
106 int rte_xen_dom0_supported(void)
108 return internal_config.xen_dom0_support;
114 * Huge page mapping under linux
116 * To reserve a big contiguous amount of memory, we use the hugepage
117 * feature of linux. For that, we need to have hugetlbfs mounted. This
118 * code will create many files in this directory (one per page) and
119 * map them in virtual memory. For each page, we will retrieve its
120 * physical address and remap it in order to have a virtual contiguous
121 * zone as well as a physical contiguous zone.
124 static uint64_t baseaddr_offset;
126 static bool phys_addrs_available = true;
128 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
131 test_phys_addrs_available(void)
134 phys_addr_t physaddr;
136 physaddr = rte_mem_virt2phy(&tmp);
137 if (physaddr == RTE_BAD_PHYS_ADDR) {
139 "Cannot obtain physical addresses: %s. "
140 "Only vfio will function.\n",
142 phys_addrs_available = false;
146 /* Lock page in physical memory and prevent from swapping. */
148 rte_mem_lock_page(const void *virt)
150 unsigned long virtual = (unsigned long)virt;
151 int page_size = getpagesize();
152 unsigned long aligned = (virtual & ~ (page_size - 1));
153 return mlock((void*)aligned, page_size);
157 * Get physical address of any mapped virtual address in the current process.
160 rte_mem_virt2phy(const void *virtaddr)
163 uint64_t page, physaddr;
164 unsigned long virt_pfn;
168 /* when using dom0, /proc/self/pagemap always returns 0, check in
169 * dpdk memory by browsing the memsegs */
170 if (rte_xen_dom0_supported()) {
171 struct rte_mem_config *mcfg;
172 struct rte_memseg *memseg;
175 mcfg = rte_eal_get_configuration()->mem_config;
176 for (i = 0; i < RTE_MAX_MEMSEG; i++) {
177 memseg = &mcfg->memseg[i];
178 if (memseg->addr == NULL)
180 if (virtaddr > memseg->addr &&
181 virtaddr < RTE_PTR_ADD(memseg->addr,
183 return memseg->phys_addr +
184 RTE_PTR_DIFF(virtaddr, memseg->addr);
188 return RTE_BAD_PHYS_ADDR;
191 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
192 if (!phys_addrs_available)
193 return RTE_BAD_PHYS_ADDR;
195 /* standard page size */
196 page_size = getpagesize();
198 fd = open("/proc/self/pagemap", O_RDONLY);
200 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
201 __func__, strerror(errno));
202 return RTE_BAD_PHYS_ADDR;
205 virt_pfn = (unsigned long)virtaddr / page_size;
206 offset = sizeof(uint64_t) * virt_pfn;
207 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
208 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
209 __func__, strerror(errno));
211 return RTE_BAD_PHYS_ADDR;
214 retval = read(fd, &page, PFN_MASK_SIZE);
217 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
218 __func__, strerror(errno));
219 return RTE_BAD_PHYS_ADDR;
220 } else if (retval != PFN_MASK_SIZE) {
221 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
222 "but expected %d:\n",
223 __func__, retval, PFN_MASK_SIZE);
224 return RTE_BAD_PHYS_ADDR;
228 * the pfn (page frame number) are bits 0-54 (see
229 * pagemap.txt in linux Documentation)
231 if ((page & 0x7fffffffffffffULL) == 0)
232 return RTE_BAD_PHYS_ADDR;
234 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
235 + ((unsigned long)virtaddr % page_size);
241 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
242 * it by browsing the /proc/self/pagemap special file.
245 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
250 for (i = 0; i < hpi->num_pages[0]; i++) {
251 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
252 if (addr == RTE_BAD_PHYS_ADDR)
254 hugepg_tbl[i].physaddr = addr;
260 * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
263 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
266 static phys_addr_t addr;
268 for (i = 0; i < hpi->num_pages[0]; i++) {
269 hugepg_tbl[i].physaddr = addr;
270 addr += hugepg_tbl[i].size;
276 * Check whether address-space layout randomization is enabled in
277 * the kernel. This is important for multi-process as it can prevent
278 * two processes mapping data to the same virtual address
280 * 0 - address space randomization disabled
281 * 1/2 - address space randomization enabled
282 * negative error code on error
288 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
291 retval = read(fd, &c, 1);
301 default: return -EINVAL;
306 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
307 * pointer to the mmap'd area and keep *size unmodified. Else, retry
308 * with a smaller zone: decrease *size by hugepage_sz until it reaches
309 * 0. In this case, return NULL. Note: this function returns an address
310 * which is a multiple of hugepage size.
313 get_virtual_area(size_t *size, size_t hugepage_sz)
319 if (internal_config.base_virtaddr != 0) {
320 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
325 RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
327 fd = open("/dev/zero", O_RDONLY);
329 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
334 (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
335 if (addr == MAP_FAILED)
336 *size -= hugepage_sz;
337 } while (addr == MAP_FAILED && *size > 0);
339 if (addr == MAP_FAILED) {
341 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
346 munmap(addr, (*size) + hugepage_sz);
349 /* align addr to a huge page size boundary */
350 aligned_addr = (long)addr;
351 aligned_addr += (hugepage_sz - 1);
352 aligned_addr &= (~(hugepage_sz - 1));
353 addr = (void *)(aligned_addr);
355 RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
358 /* increment offset */
359 baseaddr_offset += *size;
364 static sigjmp_buf huge_jmpenv;
366 static void huge_sigbus_handler(int signo __rte_unused)
368 siglongjmp(huge_jmpenv, 1);
371 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
372 * non-static local variable in the stack frame calling sigsetjmp might be
373 * clobbered by a call to longjmp.
375 static int huge_wrap_sigsetjmp(void)
377 return sigsetjmp(huge_jmpenv, 1);
381 * Mmap all hugepages of hugepage table: it first open a file in
382 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
383 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
384 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
385 * map continguous physical blocks in contiguous virtual blocks.
388 map_all_hugepages(struct hugepage_file *hugepg_tbl,
389 struct hugepage_info *hpi, int orig)
394 void *vma_addr = NULL;
397 for (i = 0; i < hpi->num_pages[0]; i++) {
398 uint64_t hugepage_sz = hpi->hugepage_sz;
401 hugepg_tbl[i].file_id = i;
402 hugepg_tbl[i].size = hugepage_sz;
403 eal_get_hugefile_path(hugepg_tbl[i].filepath,
404 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
405 hugepg_tbl[i].file_id);
406 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
409 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
410 * original map address as final map address.
412 else if ((hugepage_sz == RTE_PGSIZE_1G)
413 || (hugepage_sz == RTE_PGSIZE_16G)) {
414 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
415 hugepg_tbl[i].orig_va = NULL;
419 else if (vma_len == 0) {
420 unsigned j, num_pages;
422 /* reserve a virtual area for next contiguous
423 * physical block: count the number of
424 * contiguous physical pages. */
425 for (j = i+1; j < hpi->num_pages[0] ; j++) {
426 #ifdef RTE_ARCH_PPC_64
427 /* The physical addresses are sorted in
428 * descending order on PPC64 */
429 if (hugepg_tbl[j].physaddr !=
430 hugepg_tbl[j-1].physaddr - hugepage_sz)
433 if (hugepg_tbl[j].physaddr !=
434 hugepg_tbl[j-1].physaddr + hugepage_sz)
439 vma_len = num_pages * hugepage_sz;
441 /* get the biggest virtual memory area up to
442 * vma_len. If it fails, vma_addr is NULL, so
443 * let the kernel provide the address. */
444 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
445 if (vma_addr == NULL)
446 vma_len = hugepage_sz;
449 /* try to create hugepage file */
450 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
452 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
457 /* map the segment, and populate page tables,
458 * the kernel fills this segment with zeros */
459 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
460 MAP_SHARED | MAP_POPULATE, fd, 0);
461 if (virtaddr == MAP_FAILED) {
462 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
469 hugepg_tbl[i].orig_va = virtaddr;
472 hugepg_tbl[i].final_va = virtaddr;
476 /* In linux, hugetlb limitations, like cgroup, are
477 * enforced at fault time instead of mmap(), even
478 * with the option of MAP_POPULATE. Kernel will send
479 * a SIGBUS signal. To avoid to be killed, save stack
480 * environment here, if SIGBUS happens, we can jump
483 if (huge_wrap_sigsetjmp()) {
484 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
485 "hugepages of size %u MB\n",
486 (unsigned)(hugepage_sz / 0x100000));
487 munmap(virtaddr, hugepage_sz);
489 unlink(hugepg_tbl[i].filepath);
492 *(int *)virtaddr = 0;
496 /* set shared flock on the file. */
497 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
498 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
499 __func__, strerror(errno));
506 vma_addr = (char *)vma_addr + hugepage_sz;
507 vma_len -= hugepage_sz;
513 /* Unmap all hugepages from original mapping */
515 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
518 for (i = 0; i < hpi->num_pages[0]; i++) {
519 if (hugepg_tbl[i].orig_va) {
520 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
521 hugepg_tbl[i].orig_va = NULL;
528 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
532 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
536 unsigned i, hp_count = 0;
539 char hugedir_str[PATH_MAX];
542 f = fopen("/proc/self/numa_maps", "r");
544 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
545 " consider that all memory is in socket_id 0\n");
549 snprintf(hugedir_str, sizeof(hugedir_str),
550 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
553 while (fgets(buf, sizeof(buf), f) != NULL) {
555 /* ignore non huge page */
556 if (strstr(buf, " huge ") == NULL &&
557 strstr(buf, hugedir_str) == NULL)
561 virt_addr = strtoull(buf, &end, 16);
562 if (virt_addr == 0 || end == buf) {
563 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
567 /* get node id (socket id) */
568 nodestr = strstr(buf, " N");
569 if (nodestr == NULL) {
570 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
574 end = strstr(nodestr, "=");
576 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
582 socket_id = strtoul(nodestr, &end, 0);
583 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
584 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
588 /* if we find this page in our mappings, set socket_id */
589 for (i = 0; i < hpi->num_pages[0]; i++) {
590 void *va = (void *)(unsigned long)virt_addr;
591 if (hugepg_tbl[i].orig_va == va) {
592 hugepg_tbl[i].socket_id = socket_id;
598 if (hp_count < hpi->num_pages[0])
610 cmp_physaddr(const void *a, const void *b)
612 #ifndef RTE_ARCH_PPC_64
613 const struct hugepage_file *p1 = a;
614 const struct hugepage_file *p2 = b;
616 /* PowerPC needs memory sorted in reverse order from x86 */
617 const struct hugepage_file *p1 = b;
618 const struct hugepage_file *p2 = a;
620 if (p1->physaddr < p2->physaddr)
622 else if (p1->physaddr > p2->physaddr)
629 * Uses mmap to create a shared memory area for storage of data
630 * Used in this file to store the hugepage file map on disk
633 create_shared_memory(const char *filename, const size_t mem_size)
636 int fd = open(filename, O_CREAT | O_RDWR, 0666);
639 if (ftruncate(fd, mem_size) < 0) {
643 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
649 * this copies *active* hugepages from one hugepage table to another.
650 * destination is typically the shared memory.
653 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
654 const struct hugepage_file * src, int src_size)
656 int src_pos, dst_pos = 0;
658 for (src_pos = 0; src_pos < src_size; src_pos++) {
659 if (src[src_pos].final_va != NULL) {
660 /* error on overflow attempt */
661 if (dst_pos == dest_size)
663 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
671 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
672 unsigned num_hp_info)
674 unsigned socket, size;
675 int page, nrpages = 0;
677 /* get total number of hugepages */
678 for (size = 0; size < num_hp_info; size++)
679 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
681 internal_config.hugepage_info[size].num_pages[socket];
683 for (page = 0; page < nrpages; page++) {
684 struct hugepage_file *hp = &hugepg_tbl[page];
686 if (hp->final_va != NULL && unlink(hp->filepath)) {
687 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
688 __func__, hp->filepath, strerror(errno));
695 * unmaps hugepages that are not going to be used. since we originally allocate
696 * ALL hugepages (not just those we need), additional unmapping needs to be done.
699 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
700 struct hugepage_info *hpi,
701 unsigned num_hp_info)
703 unsigned socket, size;
704 int page, nrpages = 0;
706 /* get total number of hugepages */
707 for (size = 0; size < num_hp_info; size++)
708 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
709 nrpages += internal_config.hugepage_info[size].num_pages[socket];
711 for (size = 0; size < num_hp_info; size++) {
712 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
713 unsigned pages_found = 0;
715 /* traverse until we have unmapped all the unused pages */
716 for (page = 0; page < nrpages; page++) {
717 struct hugepage_file *hp = &hugepg_tbl[page];
719 /* find a page that matches the criteria */
720 if ((hp->size == hpi[size].hugepage_sz) &&
721 (hp->socket_id == (int) socket)) {
723 /* if we skipped enough pages, unmap the rest */
724 if (pages_found == hpi[size].num_pages[socket]) {
727 unmap_len = hp->size;
729 /* get start addr and len of the remaining segment */
730 munmap(hp->final_va, (size_t) unmap_len);
733 if (unlink(hp->filepath) == -1) {
734 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
735 __func__, hp->filepath, strerror(errno));
739 /* lock the page and skip */
745 } /* foreach socket */
746 } /* foreach pagesize */
751 static inline uint64_t
752 get_socket_mem_size(int socket)
757 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
758 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
759 if (hpi->hugedir != NULL)
760 size += hpi->hugepage_sz * hpi->num_pages[socket];
767 * This function is a NUMA-aware equivalent of calc_num_pages.
768 * It takes in the list of hugepage sizes and the
769 * number of pages thereof, and calculates the best number of
770 * pages of each size to fulfill the request for <memory> ram
773 calc_num_pages_per_socket(uint64_t * memory,
774 struct hugepage_info *hp_info,
775 struct hugepage_info *hp_used,
776 unsigned num_hp_info)
778 unsigned socket, j, i = 0;
779 unsigned requested, available;
780 int total_num_pages = 0;
781 uint64_t remaining_mem, cur_mem;
782 uint64_t total_mem = internal_config.memory;
784 if (num_hp_info == 0)
787 /* if specific memory amounts per socket weren't requested */
788 if (internal_config.force_sockets == 0) {
789 int cpu_per_socket[RTE_MAX_NUMA_NODES];
790 size_t default_size, total_size;
793 /* Compute number of cores per socket */
794 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
795 RTE_LCORE_FOREACH(lcore_id) {
796 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
800 * Automatically spread requested memory amongst detected sockets according
801 * to number of cores from cpu mask present on each socket
803 total_size = internal_config.memory;
804 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
806 /* Set memory amount per socket */
807 default_size = (internal_config.memory * cpu_per_socket[socket])
810 /* Limit to maximum available memory on socket */
811 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
814 memory[socket] = default_size;
815 total_size -= default_size;
819 * If some memory is remaining, try to allocate it by getting all
820 * available memory from sockets, one after the other
822 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
823 /* take whatever is available */
824 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
828 memory[socket] += default_size;
829 total_size -= default_size;
833 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
834 /* skips if the memory on specific socket wasn't requested */
835 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
836 hp_used[i].hugedir = hp_info[i].hugedir;
837 hp_used[i].num_pages[socket] = RTE_MIN(
838 memory[socket] / hp_info[i].hugepage_sz,
839 hp_info[i].num_pages[socket]);
841 cur_mem = hp_used[i].num_pages[socket] *
842 hp_used[i].hugepage_sz;
844 memory[socket] -= cur_mem;
845 total_mem -= cur_mem;
847 total_num_pages += hp_used[i].num_pages[socket];
849 /* check if we have met all memory requests */
850 if (memory[socket] == 0)
853 /* check if we have any more pages left at this size, if so
854 * move on to next size */
855 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
857 /* At this point we know that there are more pages available that are
858 * bigger than the memory we want, so lets see if we can get enough
859 * from other page sizes.
862 for (j = i+1; j < num_hp_info; j++)
863 remaining_mem += hp_info[j].hugepage_sz *
864 hp_info[j].num_pages[socket];
866 /* is there enough other memory, if not allocate another page and quit */
867 if (remaining_mem < memory[socket]){
868 cur_mem = RTE_MIN(memory[socket],
869 hp_info[i].hugepage_sz);
870 memory[socket] -= cur_mem;
871 total_mem -= cur_mem;
872 hp_used[i].num_pages[socket]++;
874 break; /* we are done with this socket*/
877 /* if we didn't satisfy all memory requirements per socket */
878 if (memory[socket] > 0) {
879 /* to prevent icc errors */
880 requested = (unsigned) (internal_config.socket_mem[socket] /
882 available = requested -
883 ((unsigned) (memory[socket] / 0x100000));
884 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
885 "Requested: %uMB, available: %uMB\n", socket,
886 requested, available);
891 /* if we didn't satisfy total memory requirements */
893 requested = (unsigned) (internal_config.memory / 0x100000);
894 available = requested - (unsigned) (total_mem / 0x100000);
895 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
896 " available: %uMB\n", requested, available);
899 return total_num_pages;
903 eal_get_hugepage_mem_size(void)
908 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
909 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
910 if (hpi->hugedir != NULL) {
911 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
912 size += hpi->hugepage_sz * hpi->num_pages[j];
917 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
920 static struct sigaction huge_action_old;
921 static int huge_need_recover;
924 huge_register_sigbus(void)
927 struct sigaction action;
930 sigaddset(&mask, SIGBUS);
932 action.sa_mask = mask;
933 action.sa_handler = huge_sigbus_handler;
935 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
939 huge_recover_sigbus(void)
941 if (huge_need_recover) {
942 sigaction(SIGBUS, &huge_action_old, NULL);
943 huge_need_recover = 0;
948 * Prepare physical memory mapping: fill configuration structure with
949 * these infos, return 0 on success.
950 * 1. map N huge pages in separate files in hugetlbfs
951 * 2. find associated physical addr
952 * 3. find associated NUMA socket ID
953 * 4. sort all huge pages by physical address
954 * 5. remap these N huge pages in the correct order
955 * 6. unmap the first mapping
956 * 7. fill memsegs in configuration with contiguous zones
959 rte_eal_hugepage_init(void)
961 struct rte_mem_config *mcfg;
962 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
963 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
965 uint64_t memory[RTE_MAX_NUMA_NODES];
968 int i, j, new_memseg;
969 int nr_hugefiles, nr_hugepages = 0;
972 test_phys_addrs_available();
974 memset(used_hp, 0, sizeof(used_hp));
976 /* get pointer to global configuration */
977 mcfg = rte_eal_get_configuration()->mem_config;
979 /* hugetlbfs can be disabled */
980 if (internal_config.no_hugetlbfs) {
981 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
982 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
983 if (addr == MAP_FAILED) {
984 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
988 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
989 mcfg->memseg[0].addr = addr;
990 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
991 mcfg->memseg[0].len = internal_config.memory;
992 mcfg->memseg[0].socket_id = 0;
996 /* check if app runs on Xen Dom0 */
997 if (internal_config.xen_dom0_support) {
998 #ifdef RTE_LIBRTE_XEN_DOM0
999 /* use dom0_mm kernel driver to init memory */
1000 if (rte_xen_dom0_memory_init() < 0)
1007 /* calculate total number of hugepages available. at this point we haven't
1008 * yet started sorting them so they all are on socket 0 */
1009 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1010 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1011 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1013 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1017 * allocate a memory area for hugepage table.
1018 * this isn't shared memory yet. due to the fact that we need some
1019 * processing done on these pages, shared memory will be created
1022 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1026 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1028 hp_offset = 0; /* where we start the current page size entries */
1030 huge_register_sigbus();
1032 /* map all hugepages and sort them */
1033 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1034 unsigned pages_old, pages_new;
1035 struct hugepage_info *hpi;
1038 * we don't yet mark hugepages as used at this stage, so
1039 * we just map all hugepages available to the system
1040 * all hugepages are still located on socket 0
1042 hpi = &internal_config.hugepage_info[i];
1044 if (hpi->num_pages[0] == 0)
1047 /* map all hugepages available */
1048 pages_old = hpi->num_pages[0];
1049 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, 1);
1050 if (pages_new < pages_old) {
1052 "%d not %d hugepages of size %u MB allocated\n",
1053 pages_new, pages_old,
1054 (unsigned)(hpi->hugepage_sz / 0x100000));
1056 int pages = pages_old - pages_new;
1058 nr_hugepages -= pages;
1059 hpi->num_pages[0] = pages_new;
1064 if (phys_addrs_available) {
1065 /* find physical addresses for each hugepage */
1066 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1067 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1068 "for %u MB pages\n",
1069 (unsigned int)(hpi->hugepage_sz / 0x100000));
1073 /* set physical addresses for each hugepage */
1074 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1075 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1076 "for %u MB pages\n",
1077 (unsigned int)(hpi->hugepage_sz / 0x100000));
1082 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1083 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1084 (unsigned)(hpi->hugepage_sz / 0x100000));
1088 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1089 sizeof(struct hugepage_file), cmp_physaddr);
1091 /* remap all hugepages */
1092 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) !=
1093 hpi->num_pages[0]) {
1094 RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1095 (unsigned)(hpi->hugepage_sz / 0x100000));
1099 /* unmap original mappings */
1100 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1103 /* we have processed a num of hugepages of this size, so inc offset */
1104 hp_offset += hpi->num_pages[0];
1107 huge_recover_sigbus();
1109 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1110 internal_config.memory = eal_get_hugepage_mem_size();
1112 nr_hugefiles = nr_hugepages;
1115 /* clean out the numbers of pages */
1116 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1117 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1118 internal_config.hugepage_info[i].num_pages[j] = 0;
1120 /* get hugepages for each socket */
1121 for (i = 0; i < nr_hugefiles; i++) {
1122 int socket = tmp_hp[i].socket_id;
1124 /* find a hugepage info with right size and increment num_pages */
1125 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1126 (int)internal_config.num_hugepage_sizes);
1127 for (j = 0; j < nb_hpsizes; j++) {
1128 if (tmp_hp[i].size ==
1129 internal_config.hugepage_info[j].hugepage_sz) {
1130 internal_config.hugepage_info[j].num_pages[socket]++;
1135 /* make a copy of socket_mem, needed for number of pages calculation */
1136 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1137 memory[i] = internal_config.socket_mem[i];
1139 /* calculate final number of pages */
1140 nr_hugepages = calc_num_pages_per_socket(memory,
1141 internal_config.hugepage_info, used_hp,
1142 internal_config.num_hugepage_sizes);
1144 /* error if not enough memory available */
1145 if (nr_hugepages < 0)
1149 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1150 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1151 if (used_hp[i].num_pages[j] > 0) {
1153 "Requesting %u pages of size %uMB"
1154 " from socket %i\n",
1155 used_hp[i].num_pages[j],
1157 (used_hp[i].hugepage_sz / 0x100000),
1163 /* create shared memory */
1164 hugepage = create_shared_memory(eal_hugepage_info_path(),
1165 nr_hugefiles * sizeof(struct hugepage_file));
1167 if (hugepage == NULL) {
1168 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1171 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1174 * unmap pages that we won't need (looks at used_hp).
1175 * also, sets final_va to NULL on pages that were unmapped.
1177 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1178 internal_config.num_hugepage_sizes) < 0) {
1179 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1184 * copy stuff from malloc'd hugepage* to the actual shared memory.
1185 * this procedure only copies those hugepages that have final_va
1186 * not NULL. has overflow protection.
1188 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1189 tmp_hp, nr_hugefiles) < 0) {
1190 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1194 /* free the hugepage backing files */
1195 if (internal_config.hugepage_unlink &&
1196 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1197 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1201 /* free the temporary hugepage table */
1205 /* first memseg index shall be 0 after incrementing it below */
1207 for (i = 0; i < nr_hugefiles; i++) {
1210 /* if this is a new section, create a new memseg */
1213 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1215 else if (hugepage[i].size != hugepage[i-1].size)
1218 #ifdef RTE_ARCH_PPC_64
1219 /* On PPC64 architecture, the mmap always start from higher
1220 * virtual address to lower address. Here, both the physical
1221 * address and virtual address are in descending order */
1222 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1225 else if (((unsigned long)hugepage[i-1].final_va -
1226 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1229 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1232 else if (((unsigned long)hugepage[i].final_va -
1233 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1239 if (j == RTE_MAX_MEMSEG)
1242 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1243 mcfg->memseg[j].addr = hugepage[i].final_va;
1244 mcfg->memseg[j].len = hugepage[i].size;
1245 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1246 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1248 /* continuation of previous memseg */
1250 #ifdef RTE_ARCH_PPC_64
1251 /* Use the phy and virt address of the last page as segment
1252 * address for IBM Power architecture */
1253 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1254 mcfg->memseg[j].addr = hugepage[i].final_va;
1256 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1258 hugepage[i].memseg_id = j;
1261 if (i < nr_hugefiles) {
1262 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1263 "from %d requested\n"
1264 "Current %s=%d is not enough\n"
1265 "Please either increase it or request less amount "
1267 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1272 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1277 huge_recover_sigbus();
1279 if (hugepage != NULL)
1280 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1286 * uses fstat to report the size of a file on disk
1292 if (fstat(fd, &st) < 0)
1298 * This creates the memory mappings in the secondary process to match that of
1299 * the server process. It goes through each memory segment in the DPDK runtime
1300 * configuration and finds the hugepages which form that segment, mapping them
1301 * in order to form a contiguous block in the virtual memory space
1304 rte_eal_hugepage_attach(void)
1306 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1307 struct hugepage_file *hp = NULL;
1308 unsigned num_hp = 0;
1309 unsigned i, s = 0; /* s used to track the segment number */
1310 unsigned max_seg = RTE_MAX_MEMSEG;
1312 int fd, fd_zero = -1, fd_hugepage = -1;
1314 if (aslr_enabled() > 0) {
1315 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1316 "(ASLR) is enabled in the kernel.\n");
1317 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1318 "into secondary processes\n");
1321 test_phys_addrs_available();
1323 if (internal_config.xen_dom0_support) {
1324 #ifdef RTE_LIBRTE_XEN_DOM0
1325 if (rte_xen_dom0_memory_attach() < 0) {
1326 RTE_LOG(ERR, EAL, "Failed to attach memory segments of primary "
1334 fd_zero = open("/dev/zero", O_RDONLY);
1336 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1339 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1340 if (fd_hugepage < 0) {
1341 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1345 /* map all segments into memory to make sure we get the addrs */
1346 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1350 * the first memory segment with len==0 is the one that
1351 * follows the last valid segment.
1353 if (mcfg->memseg[s].len == 0)
1357 * fdzero is mmapped to get a contiguous block of virtual
1358 * addresses of the appropriate memseg size.
1359 * use mmap to get identical addresses as the primary process.
1361 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1362 PROT_READ, MAP_PRIVATE, fd_zero, 0);
1363 if (base_addr == MAP_FAILED ||
1364 base_addr != mcfg->memseg[s].addr) {
1366 if (base_addr != MAP_FAILED) {
1367 /* errno is stale, don't use */
1368 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1369 "in /dev/zero at [%p], got [%p] - "
1370 "please use '--base-virtaddr' option\n",
1371 (unsigned long long)mcfg->memseg[s].len,
1372 mcfg->memseg[s].addr, base_addr);
1373 munmap(base_addr, mcfg->memseg[s].len);
1375 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1376 "in /dev/zero at [%p]: '%s'\n",
1377 (unsigned long long)mcfg->memseg[s].len,
1378 mcfg->memseg[s].addr, strerror(errno));
1380 if (aslr_enabled() > 0) {
1381 RTE_LOG(ERR, EAL, "It is recommended to "
1382 "disable ASLR in the kernel "
1383 "and retry running both primary "
1384 "and secondary processes\n");
1390 size = getFileSize(fd_hugepage);
1391 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1392 if (hp == MAP_FAILED) {
1393 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1397 num_hp = size / sizeof(struct hugepage_file);
1398 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1401 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1402 void *addr, *base_addr;
1403 uintptr_t offset = 0;
1404 size_t mapping_size;
1406 * free previously mapped memory so we can map the
1407 * hugepages into the space
1409 base_addr = mcfg->memseg[s].addr;
1410 munmap(base_addr, mcfg->memseg[s].len);
1412 /* find the hugepages for this segment and map them
1413 * we don't need to worry about order, as the server sorted the
1414 * entries before it did the second mmap of them */
1415 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1416 if (hp[i].memseg_id == (int)s){
1417 fd = open(hp[i].filepath, O_RDWR);
1419 RTE_LOG(ERR, EAL, "Could not open %s\n",
1423 mapping_size = hp[i].size;
1424 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1425 mapping_size, PROT_READ | PROT_WRITE,
1427 close(fd); /* close file both on success and on failure */
1428 if (addr == MAP_FAILED ||
1429 addr != RTE_PTR_ADD(base_addr, offset)) {
1430 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1434 offset+=mapping_size;
1437 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1438 (unsigned long long)mcfg->memseg[s].len);
1441 /* unmap the hugepage config file, since we are done using it */
1448 for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1449 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1450 if (hp != NULL && hp != MAP_FAILED)
1454 if (fd_hugepage >= 0)
1460 rte_eal_using_phys_addrs(void)
1462 return phys_addrs_available;