4 * Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
5 * Copyright(c) 2013 6WIND.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * * Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * * Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * * Neither the name of Intel Corporation nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
25 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
26 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
27 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
28 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
32 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35 #define _FILE_OFFSET_BITS 64
46 #include <sys/types.h>
48 #include <sys/queue.h>
53 #include <sys/ioctl.h>
57 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
63 #include <rte_memory.h>
64 #include <rte_memzone.h>
65 #include <rte_launch.h>
67 #include <rte_eal_memconfig.h>
68 #include <rte_per_lcore.h>
69 #include <rte_lcore.h>
70 #include <rte_common.h>
71 #include <rte_string_fns.h>
73 #include "eal_private.h"
74 #include "eal_internal_cfg.h"
75 #include "eal_filesystem.h"
76 #include "eal_hugepages.h"
78 #define PFN_MASK_SIZE 8
80 #ifdef RTE_LIBRTE_XEN_DOM0
81 int rte_xen_dom0_supported(void)
83 return internal_config.xen_dom0_support;
89 * Huge page mapping under linux
91 * To reserve a big contiguous amount of memory, we use the hugepage
92 * feature of linux. For that, we need to have hugetlbfs mounted. This
93 * code will create many files in this directory (one per page) and
94 * map them in virtual memory. For each page, we will retrieve its
95 * physical address and remap it in order to have a virtual contiguous
96 * zone as well as a physical contiguous zone.
99 static uint64_t baseaddr_offset;
101 static bool phys_addrs_available = true;
103 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
106 test_phys_addrs_available(void)
109 phys_addr_t physaddr;
111 /* For dom0, phys addresses can always be available */
112 if (rte_xen_dom0_supported())
115 if (!rte_eal_has_hugepages()) {
117 "Started without hugepages support, physical addresses not available\n");
118 phys_addrs_available = false;
122 physaddr = rte_mem_virt2phy(&tmp);
123 if (physaddr == RTE_BAD_PHYS_ADDR) {
125 "Cannot obtain physical addresses: %s. "
126 "Only vfio will function.\n",
128 phys_addrs_available = false;
133 * Get physical address of any mapped virtual address in the current process.
136 rte_mem_virt2phy(const void *virtaddr)
139 uint64_t page, physaddr;
140 unsigned long virt_pfn;
144 /* when using dom0, /proc/self/pagemap always returns 0, check in
145 * dpdk memory by browsing the memsegs */
146 if (rte_xen_dom0_supported()) {
147 struct rte_mem_config *mcfg;
148 struct rte_memseg *memseg;
151 mcfg = rte_eal_get_configuration()->mem_config;
152 for (i = 0; i < RTE_MAX_MEMSEG; i++) {
153 memseg = &mcfg->memseg[i];
154 if (memseg->addr == NULL)
156 if (virtaddr > memseg->addr &&
157 virtaddr < RTE_PTR_ADD(memseg->addr,
159 return memseg->phys_addr +
160 RTE_PTR_DIFF(virtaddr, memseg->addr);
164 return RTE_BAD_PHYS_ADDR;
167 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
168 if (!phys_addrs_available)
169 return RTE_BAD_PHYS_ADDR;
171 /* standard page size */
172 page_size = getpagesize();
174 fd = open("/proc/self/pagemap", O_RDONLY);
176 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
177 __func__, strerror(errno));
178 return RTE_BAD_PHYS_ADDR;
181 virt_pfn = (unsigned long)virtaddr / page_size;
182 offset = sizeof(uint64_t) * virt_pfn;
183 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
184 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
185 __func__, strerror(errno));
187 return RTE_BAD_PHYS_ADDR;
190 retval = read(fd, &page, PFN_MASK_SIZE);
193 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
194 __func__, strerror(errno));
195 return RTE_BAD_PHYS_ADDR;
196 } else if (retval != PFN_MASK_SIZE) {
197 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
198 "but expected %d:\n",
199 __func__, retval, PFN_MASK_SIZE);
200 return RTE_BAD_PHYS_ADDR;
204 * the pfn (page frame number) are bits 0-54 (see
205 * pagemap.txt in linux Documentation)
207 if ((page & 0x7fffffffffffffULL) == 0)
208 return RTE_BAD_PHYS_ADDR;
210 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
211 + ((unsigned long)virtaddr % page_size);
217 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
218 * it by browsing the /proc/self/pagemap special file.
221 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
226 for (i = 0; i < hpi->num_pages[0]; i++) {
227 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
228 if (addr == RTE_BAD_PHYS_ADDR)
230 hugepg_tbl[i].physaddr = addr;
236 * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
239 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
242 static phys_addr_t addr;
244 for (i = 0; i < hpi->num_pages[0]; i++) {
245 hugepg_tbl[i].physaddr = addr;
246 addr += hugepg_tbl[i].size;
252 * Check whether address-space layout randomization is enabled in
253 * the kernel. This is important for multi-process as it can prevent
254 * two processes mapping data to the same virtual address
256 * 0 - address space randomization disabled
257 * 1/2 - address space randomization enabled
258 * negative error code on error
264 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
267 retval = read(fd, &c, 1);
277 default: return -EINVAL;
282 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
283 * pointer to the mmap'd area and keep *size unmodified. Else, retry
284 * with a smaller zone: decrease *size by hugepage_sz until it reaches
285 * 0. In this case, return NULL. Note: this function returns an address
286 * which is a multiple of hugepage size.
289 get_virtual_area(size_t *size, size_t hugepage_sz)
295 if (internal_config.base_virtaddr != 0) {
296 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
301 RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
303 fd = open("/dev/zero", O_RDONLY);
305 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
310 (*size) + hugepage_sz, PROT_READ,
311 #ifdef RTE_ARCH_PPC_64
312 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
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);
362 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
363 /* Callback for numa library. */
364 void numa_error(char *where)
366 RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
371 * Mmap all hugepages of hugepage table: it first open a file in
372 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
373 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
374 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
375 * map continguous physical blocks in contiguous virtual blocks.
378 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
379 uint64_t *essential_memory __rte_unused, int orig)
384 void *vma_addr = NULL;
386 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
388 int essential_prev = 0;
390 struct bitmask *oldmask = numa_allocate_nodemask();
391 bool have_numa = true;
392 unsigned long maxnode = 0;
394 /* Check if kernel supports NUMA. */
395 if (numa_available() != 0) {
396 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
400 if (orig && have_numa) {
401 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
402 if (get_mempolicy(&oldpolicy, oldmask->maskp,
403 oldmask->size + 1, 0, 0) < 0) {
405 "Failed to get current mempolicy: %s. "
406 "Assuming MPOL_DEFAULT.\n", strerror(errno));
407 oldpolicy = MPOL_DEFAULT;
409 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
410 if (internal_config.socket_mem[i])
415 for (i = 0; i < hpi->num_pages[0]; i++) {
416 uint64_t hugepage_sz = hpi->hugepage_sz;
418 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
422 for (j = 0; j < maxnode; j++)
423 if (essential_memory[j])
427 node_id = (node_id + 1) % maxnode;
428 while (!internal_config.socket_mem[node_id]) {
435 essential_prev = essential_memory[j];
437 if (essential_memory[j] < hugepage_sz)
438 essential_memory[j] = 0;
440 essential_memory[j] -= hugepage_sz;
444 "Setting policy MPOL_PREFERRED for socket %d\n",
446 numa_set_preferred(node_id);
451 hugepg_tbl[i].file_id = i;
452 hugepg_tbl[i].size = hugepage_sz;
453 eal_get_hugefile_path(hugepg_tbl[i].filepath,
454 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
455 hugepg_tbl[i].file_id);
456 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
459 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
460 * original map address as final map address.
462 else if ((hugepage_sz == RTE_PGSIZE_1G)
463 || (hugepage_sz == RTE_PGSIZE_16G)) {
464 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
465 hugepg_tbl[i].orig_va = NULL;
469 else if (vma_len == 0) {
470 unsigned j, num_pages;
472 /* reserve a virtual area for next contiguous
473 * physical block: count the number of
474 * contiguous physical pages. */
475 for (j = i+1; j < hpi->num_pages[0] ; j++) {
476 #ifdef RTE_ARCH_PPC_64
477 /* The physical addresses are sorted in
478 * descending order on PPC64 */
479 if (hugepg_tbl[j].physaddr !=
480 hugepg_tbl[j-1].physaddr - hugepage_sz)
483 if (hugepg_tbl[j].physaddr !=
484 hugepg_tbl[j-1].physaddr + hugepage_sz)
489 vma_len = num_pages * hugepage_sz;
491 /* get the biggest virtual memory area up to
492 * vma_len. If it fails, vma_addr is NULL, so
493 * let the kernel provide the address. */
494 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
495 if (vma_addr == NULL)
496 vma_len = hugepage_sz;
499 /* try to create hugepage file */
500 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
502 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
507 /* map the segment, and populate page tables,
508 * the kernel fills this segment with zeros */
509 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
510 MAP_SHARED | MAP_POPULATE, fd, 0);
511 if (virtaddr == MAP_FAILED) {
512 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
519 hugepg_tbl[i].orig_va = virtaddr;
522 hugepg_tbl[i].final_va = virtaddr;
526 /* In linux, hugetlb limitations, like cgroup, are
527 * enforced at fault time instead of mmap(), even
528 * with the option of MAP_POPULATE. Kernel will send
529 * a SIGBUS signal. To avoid to be killed, save stack
530 * environment here, if SIGBUS happens, we can jump
533 if (huge_wrap_sigsetjmp()) {
534 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
535 "hugepages of size %u MB\n",
536 (unsigned)(hugepage_sz / 0x100000));
537 munmap(virtaddr, hugepage_sz);
539 unlink(hugepg_tbl[i].filepath);
540 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
542 essential_memory[node_id] =
547 *(int *)virtaddr = 0;
551 /* set shared flock on the file. */
552 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
553 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
554 __func__, strerror(errno));
561 vma_addr = (char *)vma_addr + hugepage_sz;
562 vma_len -= hugepage_sz;
566 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
569 "Restoring previous memory policy: %d\n", oldpolicy);
570 if (oldpolicy == MPOL_DEFAULT) {
571 numa_set_localalloc();
572 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
573 oldmask->size + 1) < 0) {
574 RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
576 numa_set_localalloc();
579 numa_free_cpumask(oldmask);
584 /* Unmap all hugepages from original mapping */
586 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
589 for (i = 0; i < hpi->num_pages[0]; i++) {
590 if (hugepg_tbl[i].orig_va) {
591 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
592 hugepg_tbl[i].orig_va = NULL;
599 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
603 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
607 unsigned i, hp_count = 0;
610 char hugedir_str[PATH_MAX];
613 f = fopen("/proc/self/numa_maps", "r");
615 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
616 " consider that all memory is in socket_id 0\n");
620 snprintf(hugedir_str, sizeof(hugedir_str),
621 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
624 while (fgets(buf, sizeof(buf), f) != NULL) {
626 /* ignore non huge page */
627 if (strstr(buf, " huge ") == NULL &&
628 strstr(buf, hugedir_str) == NULL)
632 virt_addr = strtoull(buf, &end, 16);
633 if (virt_addr == 0 || end == buf) {
634 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
638 /* get node id (socket id) */
639 nodestr = strstr(buf, " N");
640 if (nodestr == NULL) {
641 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
645 end = strstr(nodestr, "=");
647 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
653 socket_id = strtoul(nodestr, &end, 0);
654 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
655 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
659 /* if we find this page in our mappings, set socket_id */
660 for (i = 0; i < hpi->num_pages[0]; i++) {
661 void *va = (void *)(unsigned long)virt_addr;
662 if (hugepg_tbl[i].orig_va == va) {
663 hugepg_tbl[i].socket_id = socket_id;
665 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
667 "Hugepage %s is on socket %d\n",
668 hugepg_tbl[i].filepath, socket_id);
674 if (hp_count < hpi->num_pages[0])
686 cmp_physaddr(const void *a, const void *b)
688 #ifndef RTE_ARCH_PPC_64
689 const struct hugepage_file *p1 = a;
690 const struct hugepage_file *p2 = b;
692 /* PowerPC needs memory sorted in reverse order from x86 */
693 const struct hugepage_file *p1 = b;
694 const struct hugepage_file *p2 = a;
696 if (p1->physaddr < p2->physaddr)
698 else if (p1->physaddr > p2->physaddr)
705 * Uses mmap to create a shared memory area for storage of data
706 * Used in this file to store the hugepage file map on disk
709 create_shared_memory(const char *filename, const size_t mem_size)
712 int fd = open(filename, O_CREAT | O_RDWR, 0666);
715 if (ftruncate(fd, mem_size) < 0) {
719 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
725 * this copies *active* hugepages from one hugepage table to another.
726 * destination is typically the shared memory.
729 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
730 const struct hugepage_file * src, int src_size)
732 int src_pos, dst_pos = 0;
734 for (src_pos = 0; src_pos < src_size; src_pos++) {
735 if (src[src_pos].final_va != NULL) {
736 /* error on overflow attempt */
737 if (dst_pos == dest_size)
739 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
747 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
748 unsigned num_hp_info)
750 unsigned socket, size;
751 int page, nrpages = 0;
753 /* get total number of hugepages */
754 for (size = 0; size < num_hp_info; size++)
755 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
757 internal_config.hugepage_info[size].num_pages[socket];
759 for (page = 0; page < nrpages; page++) {
760 struct hugepage_file *hp = &hugepg_tbl[page];
762 if (hp->final_va != NULL && unlink(hp->filepath)) {
763 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
764 __func__, hp->filepath, strerror(errno));
771 * unmaps hugepages that are not going to be used. since we originally allocate
772 * ALL hugepages (not just those we need), additional unmapping needs to be done.
775 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
776 struct hugepage_info *hpi,
777 unsigned num_hp_info)
779 unsigned socket, size;
780 int page, nrpages = 0;
782 /* get total number of hugepages */
783 for (size = 0; size < num_hp_info; size++)
784 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
785 nrpages += internal_config.hugepage_info[size].num_pages[socket];
787 for (size = 0; size < num_hp_info; size++) {
788 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
789 unsigned pages_found = 0;
791 /* traverse until we have unmapped all the unused pages */
792 for (page = 0; page < nrpages; page++) {
793 struct hugepage_file *hp = &hugepg_tbl[page];
795 /* find a page that matches the criteria */
796 if ((hp->size == hpi[size].hugepage_sz) &&
797 (hp->socket_id == (int) socket)) {
799 /* if we skipped enough pages, unmap the rest */
800 if (pages_found == hpi[size].num_pages[socket]) {
803 unmap_len = hp->size;
805 /* get start addr and len of the remaining segment */
806 munmap(hp->final_va, (size_t) unmap_len);
809 if (unlink(hp->filepath) == -1) {
810 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
811 __func__, hp->filepath, strerror(errno));
815 /* lock the page and skip */
821 } /* foreach socket */
822 } /* foreach pagesize */
827 static inline uint64_t
828 get_socket_mem_size(int socket)
833 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
834 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
835 if (hpi->hugedir != NULL)
836 size += hpi->hugepage_sz * hpi->num_pages[socket];
843 * This function is a NUMA-aware equivalent of calc_num_pages.
844 * It takes in the list of hugepage sizes and the
845 * number of pages thereof, and calculates the best number of
846 * pages of each size to fulfill the request for <memory> ram
849 calc_num_pages_per_socket(uint64_t * memory,
850 struct hugepage_info *hp_info,
851 struct hugepage_info *hp_used,
852 unsigned num_hp_info)
854 unsigned socket, j, i = 0;
855 unsigned requested, available;
856 int total_num_pages = 0;
857 uint64_t remaining_mem, cur_mem;
858 uint64_t total_mem = internal_config.memory;
860 if (num_hp_info == 0)
863 /* if specific memory amounts per socket weren't requested */
864 if (internal_config.force_sockets == 0) {
865 int cpu_per_socket[RTE_MAX_NUMA_NODES];
866 size_t default_size, total_size;
869 /* Compute number of cores per socket */
870 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
871 RTE_LCORE_FOREACH(lcore_id) {
872 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
876 * Automatically spread requested memory amongst detected sockets according
877 * to number of cores from cpu mask present on each socket
879 total_size = internal_config.memory;
880 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
882 /* Set memory amount per socket */
883 default_size = (internal_config.memory * cpu_per_socket[socket])
886 /* Limit to maximum available memory on socket */
887 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
890 memory[socket] = default_size;
891 total_size -= default_size;
895 * If some memory is remaining, try to allocate it by getting all
896 * available memory from sockets, one after the other
898 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
899 /* take whatever is available */
900 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
904 memory[socket] += default_size;
905 total_size -= default_size;
909 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
910 /* skips if the memory on specific socket wasn't requested */
911 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
912 hp_used[i].hugedir = hp_info[i].hugedir;
913 hp_used[i].num_pages[socket] = RTE_MIN(
914 memory[socket] / hp_info[i].hugepage_sz,
915 hp_info[i].num_pages[socket]);
917 cur_mem = hp_used[i].num_pages[socket] *
918 hp_used[i].hugepage_sz;
920 memory[socket] -= cur_mem;
921 total_mem -= cur_mem;
923 total_num_pages += hp_used[i].num_pages[socket];
925 /* check if we have met all memory requests */
926 if (memory[socket] == 0)
929 /* check if we have any more pages left at this size, if so
930 * move on to next size */
931 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
933 /* At this point we know that there are more pages available that are
934 * bigger than the memory we want, so lets see if we can get enough
935 * from other page sizes.
938 for (j = i+1; j < num_hp_info; j++)
939 remaining_mem += hp_info[j].hugepage_sz *
940 hp_info[j].num_pages[socket];
942 /* is there enough other memory, if not allocate another page and quit */
943 if (remaining_mem < memory[socket]){
944 cur_mem = RTE_MIN(memory[socket],
945 hp_info[i].hugepage_sz);
946 memory[socket] -= cur_mem;
947 total_mem -= cur_mem;
948 hp_used[i].num_pages[socket]++;
950 break; /* we are done with this socket*/
953 /* if we didn't satisfy all memory requirements per socket */
954 if (memory[socket] > 0) {
955 /* to prevent icc errors */
956 requested = (unsigned) (internal_config.socket_mem[socket] /
958 available = requested -
959 ((unsigned) (memory[socket] / 0x100000));
960 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
961 "Requested: %uMB, available: %uMB\n", socket,
962 requested, available);
967 /* if we didn't satisfy total memory requirements */
969 requested = (unsigned) (internal_config.memory / 0x100000);
970 available = requested - (unsigned) (total_mem / 0x100000);
971 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
972 " available: %uMB\n", requested, available);
975 return total_num_pages;
979 eal_get_hugepage_mem_size(void)
984 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
985 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
986 if (hpi->hugedir != NULL) {
987 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
988 size += hpi->hugepage_sz * hpi->num_pages[j];
993 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
996 static struct sigaction huge_action_old;
997 static int huge_need_recover;
1000 huge_register_sigbus(void)
1003 struct sigaction action;
1006 sigaddset(&mask, SIGBUS);
1007 action.sa_flags = 0;
1008 action.sa_mask = mask;
1009 action.sa_handler = huge_sigbus_handler;
1011 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1015 huge_recover_sigbus(void)
1017 if (huge_need_recover) {
1018 sigaction(SIGBUS, &huge_action_old, NULL);
1019 huge_need_recover = 0;
1024 * Prepare physical memory mapping: fill configuration structure with
1025 * these infos, return 0 on success.
1026 * 1. map N huge pages in separate files in hugetlbfs
1027 * 2. find associated physical addr
1028 * 3. find associated NUMA socket ID
1029 * 4. sort all huge pages by physical address
1030 * 5. remap these N huge pages in the correct order
1031 * 6. unmap the first mapping
1032 * 7. fill memsegs in configuration with contiguous zones
1035 rte_eal_hugepage_init(void)
1037 struct rte_mem_config *mcfg;
1038 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1039 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1041 uint64_t memory[RTE_MAX_NUMA_NODES];
1044 int i, j, new_memseg;
1045 int nr_hugefiles, nr_hugepages = 0;
1048 test_phys_addrs_available();
1050 memset(used_hp, 0, sizeof(used_hp));
1052 /* get pointer to global configuration */
1053 mcfg = rte_eal_get_configuration()->mem_config;
1055 /* hugetlbfs can be disabled */
1056 if (internal_config.no_hugetlbfs) {
1057 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1058 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1059 if (addr == MAP_FAILED) {
1060 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1064 mcfg->memseg[0].phys_addr = RTE_BAD_PHYS_ADDR;
1065 mcfg->memseg[0].addr = addr;
1066 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1067 mcfg->memseg[0].len = internal_config.memory;
1068 mcfg->memseg[0].socket_id = 0;
1072 /* check if app runs on Xen Dom0 */
1073 if (internal_config.xen_dom0_support) {
1074 #ifdef RTE_LIBRTE_XEN_DOM0
1075 /* use dom0_mm kernel driver to init memory */
1076 if (rte_xen_dom0_memory_init() < 0)
1083 /* calculate total number of hugepages available. at this point we haven't
1084 * yet started sorting them so they all are on socket 0 */
1085 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1086 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1087 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1089 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1093 * allocate a memory area for hugepage table.
1094 * this isn't shared memory yet. due to the fact that we need some
1095 * processing done on these pages, shared memory will be created
1098 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1102 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1104 hp_offset = 0; /* where we start the current page size entries */
1106 huge_register_sigbus();
1108 /* make a copy of socket_mem, needed for balanced allocation. */
1109 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1110 memory[i] = internal_config.socket_mem[i];
1113 /* map all hugepages and sort them */
1114 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1115 unsigned pages_old, pages_new;
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 pages_old = hpi->num_pages[0];
1130 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi,
1132 if (pages_new < pages_old) {
1134 "%d not %d hugepages of size %u MB allocated\n",
1135 pages_new, pages_old,
1136 (unsigned)(hpi->hugepage_sz / 0x100000));
1138 int pages = pages_old - pages_new;
1140 nr_hugepages -= pages;
1141 hpi->num_pages[0] = pages_new;
1146 if (phys_addrs_available) {
1147 /* find physical addresses for each hugepage */
1148 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1149 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1150 "for %u MB pages\n",
1151 (unsigned int)(hpi->hugepage_sz / 0x100000));
1155 /* set physical addresses for each hugepage */
1156 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1157 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1158 "for %u MB pages\n",
1159 (unsigned int)(hpi->hugepage_sz / 0x100000));
1164 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1165 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1166 (unsigned)(hpi->hugepage_sz / 0x100000));
1170 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1171 sizeof(struct hugepage_file), cmp_physaddr);
1173 /* remap all hugepages */
1174 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, NULL, 0) !=
1175 hpi->num_pages[0]) {
1176 RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1177 (unsigned)(hpi->hugepage_sz / 0x100000));
1181 /* unmap original mappings */
1182 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1185 /* we have processed a num of hugepages of this size, so inc offset */
1186 hp_offset += hpi->num_pages[0];
1189 huge_recover_sigbus();
1191 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1192 internal_config.memory = eal_get_hugepage_mem_size();
1194 nr_hugefiles = nr_hugepages;
1197 /* clean out the numbers of pages */
1198 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1199 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1200 internal_config.hugepage_info[i].num_pages[j] = 0;
1202 /* get hugepages for each socket */
1203 for (i = 0; i < nr_hugefiles; i++) {
1204 int socket = tmp_hp[i].socket_id;
1206 /* find a hugepage info with right size and increment num_pages */
1207 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1208 (int)internal_config.num_hugepage_sizes);
1209 for (j = 0; j < nb_hpsizes; j++) {
1210 if (tmp_hp[i].size ==
1211 internal_config.hugepage_info[j].hugepage_sz) {
1212 internal_config.hugepage_info[j].num_pages[socket]++;
1217 /* make a copy of socket_mem, needed for number of pages calculation */
1218 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1219 memory[i] = internal_config.socket_mem[i];
1221 /* calculate final number of pages */
1222 nr_hugepages = calc_num_pages_per_socket(memory,
1223 internal_config.hugepage_info, used_hp,
1224 internal_config.num_hugepage_sizes);
1226 /* error if not enough memory available */
1227 if (nr_hugepages < 0)
1231 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1232 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1233 if (used_hp[i].num_pages[j] > 0) {
1235 "Requesting %u pages of size %uMB"
1236 " from socket %i\n",
1237 used_hp[i].num_pages[j],
1239 (used_hp[i].hugepage_sz / 0x100000),
1245 /* create shared memory */
1246 hugepage = create_shared_memory(eal_hugepage_info_path(),
1247 nr_hugefiles * sizeof(struct hugepage_file));
1249 if (hugepage == NULL) {
1250 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1253 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1256 * unmap pages that we won't need (looks at used_hp).
1257 * also, sets final_va to NULL on pages that were unmapped.
1259 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1260 internal_config.num_hugepage_sizes) < 0) {
1261 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1266 * copy stuff from malloc'd hugepage* to the actual shared memory.
1267 * this procedure only copies those hugepages that have final_va
1268 * not NULL. has overflow protection.
1270 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1271 tmp_hp, nr_hugefiles) < 0) {
1272 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1276 /* free the hugepage backing files */
1277 if (internal_config.hugepage_unlink &&
1278 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1279 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1283 /* free the temporary hugepage table */
1287 /* first memseg index shall be 0 after incrementing it below */
1289 for (i = 0; i < nr_hugefiles; i++) {
1292 /* if this is a new section, create a new memseg */
1295 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1297 else if (hugepage[i].size != hugepage[i-1].size)
1300 #ifdef RTE_ARCH_PPC_64
1301 /* On PPC64 architecture, the mmap always start from higher
1302 * virtual address to lower address. Here, both the physical
1303 * address and virtual address are in descending order */
1304 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1307 else if (((unsigned long)hugepage[i-1].final_va -
1308 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1311 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1314 else if (((unsigned long)hugepage[i].final_va -
1315 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1321 if (j == RTE_MAX_MEMSEG)
1324 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1325 mcfg->memseg[j].addr = hugepage[i].final_va;
1326 mcfg->memseg[j].len = hugepage[i].size;
1327 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1328 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1330 /* continuation of previous memseg */
1332 #ifdef RTE_ARCH_PPC_64
1333 /* Use the phy and virt address of the last page as segment
1334 * address for IBM Power architecture */
1335 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1336 mcfg->memseg[j].addr = hugepage[i].final_va;
1338 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1340 hugepage[i].memseg_id = j;
1343 if (i < nr_hugefiles) {
1344 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1345 "from %d requested\n"
1346 "Current %s=%d is not enough\n"
1347 "Please either increase it or request less amount "
1349 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1354 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1359 huge_recover_sigbus();
1361 if (hugepage != NULL)
1362 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1368 * uses fstat to report the size of a file on disk
1374 if (fstat(fd, &st) < 0)
1380 * This creates the memory mappings in the secondary process to match that of
1381 * the server process. It goes through each memory segment in the DPDK runtime
1382 * configuration and finds the hugepages which form that segment, mapping them
1383 * in order to form a contiguous block in the virtual memory space
1386 rte_eal_hugepage_attach(void)
1388 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1389 struct hugepage_file *hp = NULL;
1390 unsigned num_hp = 0;
1391 unsigned i, s = 0; /* s used to track the segment number */
1392 unsigned max_seg = RTE_MAX_MEMSEG;
1394 int fd, fd_zero = -1, fd_hugepage = -1;
1396 if (aslr_enabled() > 0) {
1397 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1398 "(ASLR) is enabled in the kernel.\n");
1399 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1400 "into secondary processes\n");
1403 test_phys_addrs_available();
1405 if (internal_config.xen_dom0_support) {
1406 #ifdef RTE_LIBRTE_XEN_DOM0
1407 if (rte_xen_dom0_memory_attach() < 0) {
1408 RTE_LOG(ERR, EAL, "Failed to attach memory segments of primary "
1416 fd_zero = open("/dev/zero", O_RDONLY);
1418 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1421 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1422 if (fd_hugepage < 0) {
1423 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1427 /* map all segments into memory to make sure we get the addrs */
1428 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1432 * the first memory segment with len==0 is the one that
1433 * follows the last valid segment.
1435 if (mcfg->memseg[s].len == 0)
1439 * fdzero is mmapped to get a contiguous block of virtual
1440 * addresses of the appropriate memseg size.
1441 * use mmap to get identical addresses as the primary process.
1443 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1445 #ifdef RTE_ARCH_PPC_64
1446 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1451 if (base_addr == MAP_FAILED ||
1452 base_addr != mcfg->memseg[s].addr) {
1454 if (base_addr != MAP_FAILED) {
1455 /* errno is stale, don't use */
1456 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1457 "in /dev/zero at [%p], got [%p] - "
1458 "please use '--base-virtaddr' option\n",
1459 (unsigned long long)mcfg->memseg[s].len,
1460 mcfg->memseg[s].addr, base_addr);
1461 munmap(base_addr, mcfg->memseg[s].len);
1463 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1464 "in /dev/zero at [%p]: '%s'\n",
1465 (unsigned long long)mcfg->memseg[s].len,
1466 mcfg->memseg[s].addr, strerror(errno));
1468 if (aslr_enabled() > 0) {
1469 RTE_LOG(ERR, EAL, "It is recommended to "
1470 "disable ASLR in the kernel "
1471 "and retry running both primary "
1472 "and secondary processes\n");
1478 size = getFileSize(fd_hugepage);
1479 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1480 if (hp == MAP_FAILED) {
1481 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1485 num_hp = size / sizeof(struct hugepage_file);
1486 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1489 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1490 void *addr, *base_addr;
1491 uintptr_t offset = 0;
1492 size_t mapping_size;
1494 * free previously mapped memory so we can map the
1495 * hugepages into the space
1497 base_addr = mcfg->memseg[s].addr;
1498 munmap(base_addr, mcfg->memseg[s].len);
1500 /* find the hugepages for this segment and map them
1501 * we don't need to worry about order, as the server sorted the
1502 * entries before it did the second mmap of them */
1503 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1504 if (hp[i].memseg_id == (int)s){
1505 fd = open(hp[i].filepath, O_RDWR);
1507 RTE_LOG(ERR, EAL, "Could not open %s\n",
1511 mapping_size = hp[i].size;
1512 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1513 mapping_size, PROT_READ | PROT_WRITE,
1515 close(fd); /* close file both on success and on failure */
1516 if (addr == MAP_FAILED ||
1517 addr != RTE_PTR_ADD(base_addr, offset)) {
1518 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1522 offset+=mapping_size;
1525 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1526 (unsigned long long)mcfg->memseg[s].len);
1529 /* unmap the hugepage config file, since we are done using it */
1536 for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1537 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1538 if (hp != NULL && hp != MAP_FAILED)
1542 if (fd_hugepage >= 0)
1548 rte_eal_using_phys_addrs(void)
1550 return phys_addrs_available;