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
45 #include <sys/types.h>
47 #include <sys/queue.h>
51 #include <sys/ioctl.h>
55 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
61 #include <rte_memory.h>
62 #include <rte_memzone.h>
63 #include <rte_launch.h>
65 #include <rte_eal_memconfig.h>
66 #include <rte_per_lcore.h>
67 #include <rte_lcore.h>
68 #include <rte_common.h>
69 #include <rte_string_fns.h>
71 #include "eal_private.h"
72 #include "eal_internal_cfg.h"
73 #include "eal_filesystem.h"
74 #include "eal_hugepages.h"
76 #define PFN_MASK_SIZE 8
80 * Huge page mapping under linux
82 * To reserve a big contiguous amount of memory, we use the hugepage
83 * feature of linux. For that, we need to have hugetlbfs mounted. This
84 * code will create many files in this directory (one per page) and
85 * map them in virtual memory. For each page, we will retrieve its
86 * physical address and remap it in order to have a virtual contiguous
87 * zone as well as a physical contiguous zone.
90 static uint64_t baseaddr_offset;
92 static bool phys_addrs_available = true;
94 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
97 test_phys_addrs_available(void)
100 phys_addr_t physaddr;
102 if (!rte_eal_has_hugepages()) {
104 "Started without hugepages support, physical addresses not available\n");
105 phys_addrs_available = false;
109 physaddr = rte_mem_virt2phy(&tmp);
110 if (physaddr == RTE_BAD_PHYS_ADDR) {
112 "Cannot obtain physical addresses: %s. "
113 "Only vfio will function.\n",
115 phys_addrs_available = false;
120 * Get physical address of any mapped virtual address in the current process.
123 rte_mem_virt2phy(const void *virtaddr)
126 uint64_t page, physaddr;
127 unsigned long virt_pfn;
131 if (rte_eal_iova_mode() == RTE_IOVA_VA)
132 return (uintptr_t)virtaddr;
134 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
135 if (!phys_addrs_available)
136 return RTE_BAD_PHYS_ADDR;
138 /* standard page size */
139 page_size = getpagesize();
141 fd = open("/proc/self/pagemap", O_RDONLY);
143 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
144 __func__, strerror(errno));
145 return RTE_BAD_PHYS_ADDR;
148 virt_pfn = (unsigned long)virtaddr / page_size;
149 offset = sizeof(uint64_t) * virt_pfn;
150 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
151 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
152 __func__, strerror(errno));
154 return RTE_BAD_PHYS_ADDR;
157 retval = read(fd, &page, PFN_MASK_SIZE);
160 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
161 __func__, strerror(errno));
162 return RTE_BAD_PHYS_ADDR;
163 } else if (retval != PFN_MASK_SIZE) {
164 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
165 "but expected %d:\n",
166 __func__, retval, PFN_MASK_SIZE);
167 return RTE_BAD_PHYS_ADDR;
171 * the pfn (page frame number) are bits 0-54 (see
172 * pagemap.txt in linux Documentation)
174 if ((page & 0x7fffffffffffffULL) == 0)
175 return RTE_BAD_PHYS_ADDR;
177 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
178 + ((unsigned long)virtaddr % page_size);
184 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
185 * it by browsing the /proc/self/pagemap special file.
188 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
193 for (i = 0; i < hpi->num_pages[0]; i++) {
194 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
195 if (addr == RTE_BAD_PHYS_ADDR)
197 hugepg_tbl[i].physaddr = addr;
203 * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
206 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
209 static phys_addr_t addr;
211 for (i = 0; i < hpi->num_pages[0]; i++) {
212 hugepg_tbl[i].physaddr = addr;
213 addr += hugepg_tbl[i].size;
219 * Check whether address-space layout randomization is enabled in
220 * the kernel. This is important for multi-process as it can prevent
221 * two processes mapping data to the same virtual address
223 * 0 - address space randomization disabled
224 * 1/2 - address space randomization enabled
225 * negative error code on error
231 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
234 retval = read(fd, &c, 1);
244 default: return -EINVAL;
249 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
250 * pointer to the mmap'd area and keep *size unmodified. Else, retry
251 * with a smaller zone: decrease *size by hugepage_sz until it reaches
252 * 0. In this case, return NULL. Note: this function returns an address
253 * which is a multiple of hugepage size.
256 get_virtual_area(size_t *size, size_t hugepage_sz)
262 if (internal_config.base_virtaddr != 0) {
263 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
268 RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
270 fd = open("/dev/zero", O_RDONLY);
272 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
277 (*size) + hugepage_sz, PROT_READ,
278 #ifdef RTE_ARCH_PPC_64
279 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
284 if (addr == MAP_FAILED)
285 *size -= hugepage_sz;
286 } while (addr == MAP_FAILED && *size > 0);
288 if (addr == MAP_FAILED) {
290 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
295 munmap(addr, (*size) + hugepage_sz);
298 /* align addr to a huge page size boundary */
299 aligned_addr = (long)addr;
300 aligned_addr += (hugepage_sz - 1);
301 aligned_addr &= (~(hugepage_sz - 1));
302 addr = (void *)(aligned_addr);
304 RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
307 /* increment offset */
308 baseaddr_offset += *size;
313 static sigjmp_buf huge_jmpenv;
315 static void huge_sigbus_handler(int signo __rte_unused)
317 siglongjmp(huge_jmpenv, 1);
320 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
321 * non-static local variable in the stack frame calling sigsetjmp might be
322 * clobbered by a call to longjmp.
324 static int huge_wrap_sigsetjmp(void)
326 return sigsetjmp(huge_jmpenv, 1);
329 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
330 /* Callback for numa library. */
331 void numa_error(char *where)
333 RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
338 * Mmap all hugepages of hugepage table: it first open a file in
339 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
340 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
341 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
342 * map continguous physical blocks in contiguous virtual blocks.
345 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
346 uint64_t *essential_memory __rte_unused, int orig)
351 void *vma_addr = NULL;
353 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
355 int essential_prev = 0;
357 struct bitmask *oldmask = numa_allocate_nodemask();
358 bool have_numa = true;
359 unsigned long maxnode = 0;
361 /* Check if kernel supports NUMA. */
362 if (numa_available() != 0) {
363 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
367 if (orig && have_numa) {
368 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
369 if (get_mempolicy(&oldpolicy, oldmask->maskp,
370 oldmask->size + 1, 0, 0) < 0) {
372 "Failed to get current mempolicy: %s. "
373 "Assuming MPOL_DEFAULT.\n", strerror(errno));
374 oldpolicy = MPOL_DEFAULT;
376 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
377 if (internal_config.socket_mem[i])
382 for (i = 0; i < hpi->num_pages[0]; i++) {
383 uint64_t hugepage_sz = hpi->hugepage_sz;
385 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
389 for (j = 0; j < maxnode; j++)
390 if (essential_memory[j])
394 node_id = (node_id + 1) % maxnode;
395 while (!internal_config.socket_mem[node_id]) {
402 essential_prev = essential_memory[j];
404 if (essential_memory[j] < hugepage_sz)
405 essential_memory[j] = 0;
407 essential_memory[j] -= hugepage_sz;
411 "Setting policy MPOL_PREFERRED for socket %d\n",
413 numa_set_preferred(node_id);
418 hugepg_tbl[i].file_id = i;
419 hugepg_tbl[i].size = hugepage_sz;
420 eal_get_hugefile_path(hugepg_tbl[i].filepath,
421 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
422 hugepg_tbl[i].file_id);
423 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
426 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
427 * original map address as final map address.
429 else if ((hugepage_sz == RTE_PGSIZE_1G)
430 || (hugepage_sz == RTE_PGSIZE_16G)) {
431 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
432 hugepg_tbl[i].orig_va = NULL;
436 else if (vma_len == 0) {
437 unsigned j, num_pages;
439 /* reserve a virtual area for next contiguous
440 * physical block: count the number of
441 * contiguous physical pages. */
442 for (j = i+1; j < hpi->num_pages[0] ; j++) {
443 #ifdef RTE_ARCH_PPC_64
444 /* The physical addresses are sorted in
445 * descending order on PPC64 */
446 if (hugepg_tbl[j].physaddr !=
447 hugepg_tbl[j-1].physaddr - hugepage_sz)
450 if (hugepg_tbl[j].physaddr !=
451 hugepg_tbl[j-1].physaddr + hugepage_sz)
456 vma_len = num_pages * hugepage_sz;
458 /* get the biggest virtual memory area up to
459 * vma_len. If it fails, vma_addr is NULL, so
460 * let the kernel provide the address. */
461 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
462 if (vma_addr == NULL)
463 vma_len = hugepage_sz;
466 /* try to create hugepage file */
467 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
469 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
474 /* map the segment, and populate page tables,
475 * the kernel fills this segment with zeros */
476 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
477 MAP_SHARED | MAP_POPULATE, fd, 0);
478 if (virtaddr == MAP_FAILED) {
479 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
486 hugepg_tbl[i].orig_va = virtaddr;
489 hugepg_tbl[i].final_va = virtaddr;
493 /* In linux, hugetlb limitations, like cgroup, are
494 * enforced at fault time instead of mmap(), even
495 * with the option of MAP_POPULATE. Kernel will send
496 * a SIGBUS signal. To avoid to be killed, save stack
497 * environment here, if SIGBUS happens, we can jump
500 if (huge_wrap_sigsetjmp()) {
501 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
502 "hugepages of size %u MB\n",
503 (unsigned)(hugepage_sz / 0x100000));
504 munmap(virtaddr, hugepage_sz);
506 unlink(hugepg_tbl[i].filepath);
507 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
509 essential_memory[node_id] =
514 *(int *)virtaddr = 0;
518 /* set shared flock on the file. */
519 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
520 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
521 __func__, strerror(errno));
528 vma_addr = (char *)vma_addr + hugepage_sz;
529 vma_len -= hugepage_sz;
533 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
536 "Restoring previous memory policy: %d\n", oldpolicy);
537 if (oldpolicy == MPOL_DEFAULT) {
538 numa_set_localalloc();
539 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
540 oldmask->size + 1) < 0) {
541 RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
543 numa_set_localalloc();
546 numa_free_cpumask(oldmask);
551 /* Unmap all hugepages from original mapping */
553 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
556 for (i = 0; i < hpi->num_pages[0]; i++) {
557 if (hugepg_tbl[i].orig_va) {
558 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
559 hugepg_tbl[i].orig_va = NULL;
566 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
570 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
574 unsigned i, hp_count = 0;
577 char hugedir_str[PATH_MAX];
580 f = fopen("/proc/self/numa_maps", "r");
582 RTE_LOG(NOTICE, EAL, "NUMA support not available"
583 " consider that all memory is in socket_id 0\n");
587 snprintf(hugedir_str, sizeof(hugedir_str),
588 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
591 while (fgets(buf, sizeof(buf), f) != NULL) {
593 /* ignore non huge page */
594 if (strstr(buf, " huge ") == NULL &&
595 strstr(buf, hugedir_str) == NULL)
599 virt_addr = strtoull(buf, &end, 16);
600 if (virt_addr == 0 || end == buf) {
601 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
605 /* get node id (socket id) */
606 nodestr = strstr(buf, " N");
607 if (nodestr == NULL) {
608 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
612 end = strstr(nodestr, "=");
614 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
620 socket_id = strtoul(nodestr, &end, 0);
621 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
622 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
626 /* if we find this page in our mappings, set socket_id */
627 for (i = 0; i < hpi->num_pages[0]; i++) {
628 void *va = (void *)(unsigned long)virt_addr;
629 if (hugepg_tbl[i].orig_va == va) {
630 hugepg_tbl[i].socket_id = socket_id;
632 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
634 "Hugepage %s is on socket %d\n",
635 hugepg_tbl[i].filepath, socket_id);
641 if (hp_count < hpi->num_pages[0])
653 cmp_physaddr(const void *a, const void *b)
655 #ifndef RTE_ARCH_PPC_64
656 const struct hugepage_file *p1 = a;
657 const struct hugepage_file *p2 = b;
659 /* PowerPC needs memory sorted in reverse order from x86 */
660 const struct hugepage_file *p1 = b;
661 const struct hugepage_file *p2 = a;
663 if (p1->physaddr < p2->physaddr)
665 else if (p1->physaddr > p2->physaddr)
672 * Uses mmap to create a shared memory area for storage of data
673 * Used in this file to store the hugepage file map on disk
676 create_shared_memory(const char *filename, const size_t mem_size)
679 int fd = open(filename, O_CREAT | O_RDWR, 0666);
682 if (ftruncate(fd, mem_size) < 0) {
686 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
688 if (retval == MAP_FAILED)
694 * this copies *active* hugepages from one hugepage table to another.
695 * destination is typically the shared memory.
698 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
699 const struct hugepage_file * src, int src_size)
701 int src_pos, dst_pos = 0;
703 for (src_pos = 0; src_pos < src_size; src_pos++) {
704 if (src[src_pos].final_va != NULL) {
705 /* error on overflow attempt */
706 if (dst_pos == dest_size)
708 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
716 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
717 unsigned num_hp_info)
719 unsigned socket, size;
720 int page, nrpages = 0;
722 /* get total number of hugepages */
723 for (size = 0; size < num_hp_info; size++)
724 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
726 internal_config.hugepage_info[size].num_pages[socket];
728 for (page = 0; page < nrpages; page++) {
729 struct hugepage_file *hp = &hugepg_tbl[page];
731 if (hp->final_va != NULL && unlink(hp->filepath)) {
732 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
733 __func__, hp->filepath, strerror(errno));
740 * unmaps hugepages that are not going to be used. since we originally allocate
741 * ALL hugepages (not just those we need), additional unmapping needs to be done.
744 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
745 struct hugepage_info *hpi,
746 unsigned num_hp_info)
748 unsigned socket, size;
749 int page, nrpages = 0;
751 /* get total number of hugepages */
752 for (size = 0; size < num_hp_info; size++)
753 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
754 nrpages += internal_config.hugepage_info[size].num_pages[socket];
756 for (size = 0; size < num_hp_info; size++) {
757 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
758 unsigned pages_found = 0;
760 /* traverse until we have unmapped all the unused pages */
761 for (page = 0; page < nrpages; page++) {
762 struct hugepage_file *hp = &hugepg_tbl[page];
764 /* find a page that matches the criteria */
765 if ((hp->size == hpi[size].hugepage_sz) &&
766 (hp->socket_id == (int) socket)) {
768 /* if we skipped enough pages, unmap the rest */
769 if (pages_found == hpi[size].num_pages[socket]) {
772 unmap_len = hp->size;
774 /* get start addr and len of the remaining segment */
775 munmap(hp->final_va, (size_t) unmap_len);
778 if (unlink(hp->filepath) == -1) {
779 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
780 __func__, hp->filepath, strerror(errno));
784 /* lock the page and skip */
790 } /* foreach socket */
791 } /* foreach pagesize */
796 static inline uint64_t
797 get_socket_mem_size(int socket)
802 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
803 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
804 if (hpi->hugedir != NULL)
805 size += hpi->hugepage_sz * hpi->num_pages[socket];
812 * This function is a NUMA-aware equivalent of calc_num_pages.
813 * It takes in the list of hugepage sizes and the
814 * number of pages thereof, and calculates the best number of
815 * pages of each size to fulfill the request for <memory> ram
818 calc_num_pages_per_socket(uint64_t * memory,
819 struct hugepage_info *hp_info,
820 struct hugepage_info *hp_used,
821 unsigned num_hp_info)
823 unsigned socket, j, i = 0;
824 unsigned requested, available;
825 int total_num_pages = 0;
826 uint64_t remaining_mem, cur_mem;
827 uint64_t total_mem = internal_config.memory;
829 if (num_hp_info == 0)
832 /* if specific memory amounts per socket weren't requested */
833 if (internal_config.force_sockets == 0) {
834 int cpu_per_socket[RTE_MAX_NUMA_NODES];
835 size_t default_size, total_size;
838 /* Compute number of cores per socket */
839 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
840 RTE_LCORE_FOREACH(lcore_id) {
841 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
845 * Automatically spread requested memory amongst detected sockets according
846 * to number of cores from cpu mask present on each socket
848 total_size = internal_config.memory;
849 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
851 /* Set memory amount per socket */
852 default_size = (internal_config.memory * cpu_per_socket[socket])
855 /* Limit to maximum available memory on socket */
856 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
859 memory[socket] = default_size;
860 total_size -= default_size;
864 * If some memory is remaining, try to allocate it by getting all
865 * available memory from sockets, one after the other
867 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
868 /* take whatever is available */
869 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
873 memory[socket] += default_size;
874 total_size -= default_size;
878 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
879 /* skips if the memory on specific socket wasn't requested */
880 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
881 hp_used[i].hugedir = hp_info[i].hugedir;
882 hp_used[i].num_pages[socket] = RTE_MIN(
883 memory[socket] / hp_info[i].hugepage_sz,
884 hp_info[i].num_pages[socket]);
886 cur_mem = hp_used[i].num_pages[socket] *
887 hp_used[i].hugepage_sz;
889 memory[socket] -= cur_mem;
890 total_mem -= cur_mem;
892 total_num_pages += hp_used[i].num_pages[socket];
894 /* check if we have met all memory requests */
895 if (memory[socket] == 0)
898 /* check if we have any more pages left at this size, if so
899 * move on to next size */
900 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
902 /* At this point we know that there are more pages available that are
903 * bigger than the memory we want, so lets see if we can get enough
904 * from other page sizes.
907 for (j = i+1; j < num_hp_info; j++)
908 remaining_mem += hp_info[j].hugepage_sz *
909 hp_info[j].num_pages[socket];
911 /* is there enough other memory, if not allocate another page and quit */
912 if (remaining_mem < memory[socket]){
913 cur_mem = RTE_MIN(memory[socket],
914 hp_info[i].hugepage_sz);
915 memory[socket] -= cur_mem;
916 total_mem -= cur_mem;
917 hp_used[i].num_pages[socket]++;
919 break; /* we are done with this socket*/
922 /* if we didn't satisfy all memory requirements per socket */
923 if (memory[socket] > 0) {
924 /* to prevent icc errors */
925 requested = (unsigned) (internal_config.socket_mem[socket] /
927 available = requested -
928 ((unsigned) (memory[socket] / 0x100000));
929 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
930 "Requested: %uMB, available: %uMB\n", socket,
931 requested, available);
936 /* if we didn't satisfy total memory requirements */
938 requested = (unsigned) (internal_config.memory / 0x100000);
939 available = requested - (unsigned) (total_mem / 0x100000);
940 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
941 " available: %uMB\n", requested, available);
944 return total_num_pages;
948 eal_get_hugepage_mem_size(void)
953 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
954 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
955 if (hpi->hugedir != NULL) {
956 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
957 size += hpi->hugepage_sz * hpi->num_pages[j];
962 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
965 static struct sigaction huge_action_old;
966 static int huge_need_recover;
969 huge_register_sigbus(void)
972 struct sigaction action;
975 sigaddset(&mask, SIGBUS);
977 action.sa_mask = mask;
978 action.sa_handler = huge_sigbus_handler;
980 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
984 huge_recover_sigbus(void)
986 if (huge_need_recover) {
987 sigaction(SIGBUS, &huge_action_old, NULL);
988 huge_need_recover = 0;
993 * Prepare physical memory mapping: fill configuration structure with
994 * these infos, return 0 on success.
995 * 1. map N huge pages in separate files in hugetlbfs
996 * 2. find associated physical addr
997 * 3. find associated NUMA socket ID
998 * 4. sort all huge pages by physical address
999 * 5. remap these N huge pages in the correct order
1000 * 6. unmap the first mapping
1001 * 7. fill memsegs in configuration with contiguous zones
1004 rte_eal_hugepage_init(void)
1006 struct rte_mem_config *mcfg;
1007 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1008 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1010 uint64_t memory[RTE_MAX_NUMA_NODES];
1013 int i, j, new_memseg;
1014 int nr_hugefiles, nr_hugepages = 0;
1017 test_phys_addrs_available();
1019 memset(used_hp, 0, sizeof(used_hp));
1021 /* get pointer to global configuration */
1022 mcfg = rte_eal_get_configuration()->mem_config;
1024 /* hugetlbfs can be disabled */
1025 if (internal_config.no_hugetlbfs) {
1026 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1027 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1028 if (addr == MAP_FAILED) {
1029 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1033 if (rte_eal_iova_mode() == RTE_IOVA_VA)
1034 mcfg->memseg[0].phys_addr = (uintptr_t)addr;
1036 mcfg->memseg[0].phys_addr = RTE_BAD_PHYS_ADDR;
1037 mcfg->memseg[0].addr = addr;
1038 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1039 mcfg->memseg[0].len = internal_config.memory;
1040 mcfg->memseg[0].socket_id = 0;
1044 /* calculate total number of hugepages available. at this point we haven't
1045 * yet started sorting them so they all are on socket 0 */
1046 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1047 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1048 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1050 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1054 * allocate a memory area for hugepage table.
1055 * this isn't shared memory yet. due to the fact that we need some
1056 * processing done on these pages, shared memory will be created
1059 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1063 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1065 hp_offset = 0; /* where we start the current page size entries */
1067 huge_register_sigbus();
1069 /* make a copy of socket_mem, needed for balanced allocation. */
1070 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1071 memory[i] = internal_config.socket_mem[i];
1074 /* map all hugepages and sort them */
1075 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1076 unsigned pages_old, pages_new;
1077 struct hugepage_info *hpi;
1080 * we don't yet mark hugepages as used at this stage, so
1081 * we just map all hugepages available to the system
1082 * all hugepages are still located on socket 0
1084 hpi = &internal_config.hugepage_info[i];
1086 if (hpi->num_pages[0] == 0)
1089 /* map all hugepages available */
1090 pages_old = hpi->num_pages[0];
1091 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi,
1093 if (pages_new < pages_old) {
1095 "%d not %d hugepages of size %u MB allocated\n",
1096 pages_new, pages_old,
1097 (unsigned)(hpi->hugepage_sz / 0x100000));
1099 int pages = pages_old - pages_new;
1101 nr_hugepages -= pages;
1102 hpi->num_pages[0] = pages_new;
1107 if (phys_addrs_available) {
1108 /* find physical addresses for each hugepage */
1109 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1110 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1111 "for %u MB pages\n",
1112 (unsigned int)(hpi->hugepage_sz / 0x100000));
1116 /* set physical addresses for each hugepage */
1117 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1118 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1119 "for %u MB pages\n",
1120 (unsigned int)(hpi->hugepage_sz / 0x100000));
1125 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1126 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1127 (unsigned)(hpi->hugepage_sz / 0x100000));
1131 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1132 sizeof(struct hugepage_file), cmp_physaddr);
1134 /* remap all hugepages */
1135 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, NULL, 0) !=
1136 hpi->num_pages[0]) {
1137 RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1138 (unsigned)(hpi->hugepage_sz / 0x100000));
1142 /* unmap original mappings */
1143 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1146 /* we have processed a num of hugepages of this size, so inc offset */
1147 hp_offset += hpi->num_pages[0];
1150 huge_recover_sigbus();
1152 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1153 internal_config.memory = eal_get_hugepage_mem_size();
1155 nr_hugefiles = nr_hugepages;
1158 /* clean out the numbers of pages */
1159 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1160 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1161 internal_config.hugepage_info[i].num_pages[j] = 0;
1163 /* get hugepages for each socket */
1164 for (i = 0; i < nr_hugefiles; i++) {
1165 int socket = tmp_hp[i].socket_id;
1167 /* find a hugepage info with right size and increment num_pages */
1168 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1169 (int)internal_config.num_hugepage_sizes);
1170 for (j = 0; j < nb_hpsizes; j++) {
1171 if (tmp_hp[i].size ==
1172 internal_config.hugepage_info[j].hugepage_sz) {
1173 internal_config.hugepage_info[j].num_pages[socket]++;
1178 /* make a copy of socket_mem, needed for number of pages calculation */
1179 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1180 memory[i] = internal_config.socket_mem[i];
1182 /* calculate final number of pages */
1183 nr_hugepages = calc_num_pages_per_socket(memory,
1184 internal_config.hugepage_info, used_hp,
1185 internal_config.num_hugepage_sizes);
1187 /* error if not enough memory available */
1188 if (nr_hugepages < 0)
1192 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1193 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1194 if (used_hp[i].num_pages[j] > 0) {
1196 "Requesting %u pages of size %uMB"
1197 " from socket %i\n",
1198 used_hp[i].num_pages[j],
1200 (used_hp[i].hugepage_sz / 0x100000),
1206 /* create shared memory */
1207 hugepage = create_shared_memory(eal_hugepage_info_path(),
1208 nr_hugefiles * sizeof(struct hugepage_file));
1210 if (hugepage == NULL) {
1211 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1214 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1217 * unmap pages that we won't need (looks at used_hp).
1218 * also, sets final_va to NULL on pages that were unmapped.
1220 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1221 internal_config.num_hugepage_sizes) < 0) {
1222 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1227 * copy stuff from malloc'd hugepage* to the actual shared memory.
1228 * this procedure only copies those hugepages that have final_va
1229 * not NULL. has overflow protection.
1231 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1232 tmp_hp, nr_hugefiles) < 0) {
1233 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1237 /* free the hugepage backing files */
1238 if (internal_config.hugepage_unlink &&
1239 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1240 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1244 /* free the temporary hugepage table */
1248 /* first memseg index shall be 0 after incrementing it below */
1250 for (i = 0; i < nr_hugefiles; i++) {
1253 /* if this is a new section, create a new memseg */
1256 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1258 else if (hugepage[i].size != hugepage[i-1].size)
1261 #ifdef RTE_ARCH_PPC_64
1262 /* On PPC64 architecture, the mmap always start from higher
1263 * virtual address to lower address. Here, both the physical
1264 * address and virtual address are in descending order */
1265 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1268 else if (((unsigned long)hugepage[i-1].final_va -
1269 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1272 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1275 else if (((unsigned long)hugepage[i].final_va -
1276 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1282 if (j == RTE_MAX_MEMSEG)
1285 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1286 mcfg->memseg[j].addr = hugepage[i].final_va;
1287 mcfg->memseg[j].len = hugepage[i].size;
1288 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1289 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1291 /* continuation of previous memseg */
1293 #ifdef RTE_ARCH_PPC_64
1294 /* Use the phy and virt address of the last page as segment
1295 * address for IBM Power architecture */
1296 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1297 mcfg->memseg[j].addr = hugepage[i].final_va;
1299 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1301 hugepage[i].memseg_id = j;
1304 if (i < nr_hugefiles) {
1305 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1306 "from %d requested\n"
1307 "Current %s=%d is not enough\n"
1308 "Please either increase it or request less amount "
1310 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1315 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1320 huge_recover_sigbus();
1322 if (hugepage != NULL)
1323 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1329 * uses fstat to report the size of a file on disk
1335 if (fstat(fd, &st) < 0)
1341 * This creates the memory mappings in the secondary process to match that of
1342 * the server process. It goes through each memory segment in the DPDK runtime
1343 * configuration and finds the hugepages which form that segment, mapping them
1344 * in order to form a contiguous block in the virtual memory space
1347 rte_eal_hugepage_attach(void)
1349 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1350 struct hugepage_file *hp = NULL;
1351 unsigned num_hp = 0;
1352 unsigned i, s = 0; /* s used to track the segment number */
1353 unsigned max_seg = RTE_MAX_MEMSEG;
1355 int fd, fd_zero = -1, fd_hugepage = -1;
1357 if (aslr_enabled() > 0) {
1358 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1359 "(ASLR) is enabled in the kernel.\n");
1360 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1361 "into secondary processes\n");
1364 test_phys_addrs_available();
1366 fd_zero = open("/dev/zero", O_RDONLY);
1368 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1371 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1372 if (fd_hugepage < 0) {
1373 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1377 /* map all segments into memory to make sure we get the addrs */
1378 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1382 * the first memory segment with len==0 is the one that
1383 * follows the last valid segment.
1385 if (mcfg->memseg[s].len == 0)
1389 * fdzero is mmapped to get a contiguous block of virtual
1390 * addresses of the appropriate memseg size.
1391 * use mmap to get identical addresses as the primary process.
1393 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1395 #ifdef RTE_ARCH_PPC_64
1396 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1401 if (base_addr == MAP_FAILED ||
1402 base_addr != mcfg->memseg[s].addr) {
1404 if (base_addr != MAP_FAILED) {
1405 /* errno is stale, don't use */
1406 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1407 "in /dev/zero at [%p], got [%p] - "
1408 "please use '--base-virtaddr' option\n",
1409 (unsigned long long)mcfg->memseg[s].len,
1410 mcfg->memseg[s].addr, base_addr);
1411 munmap(base_addr, mcfg->memseg[s].len);
1413 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1414 "in /dev/zero at [%p]: '%s'\n",
1415 (unsigned long long)mcfg->memseg[s].len,
1416 mcfg->memseg[s].addr, strerror(errno));
1418 if (aslr_enabled() > 0) {
1419 RTE_LOG(ERR, EAL, "It is recommended to "
1420 "disable ASLR in the kernel "
1421 "and retry running both primary "
1422 "and secondary processes\n");
1428 size = getFileSize(fd_hugepage);
1429 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1430 if (hp == MAP_FAILED) {
1431 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1435 num_hp = size / sizeof(struct hugepage_file);
1436 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1439 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1440 void *addr, *base_addr;
1441 uintptr_t offset = 0;
1442 size_t mapping_size;
1444 * free previously mapped memory so we can map the
1445 * hugepages into the space
1447 base_addr = mcfg->memseg[s].addr;
1448 munmap(base_addr, mcfg->memseg[s].len);
1450 /* find the hugepages for this segment and map them
1451 * we don't need to worry about order, as the server sorted the
1452 * entries before it did the second mmap of them */
1453 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1454 if (hp[i].memseg_id == (int)s){
1455 fd = open(hp[i].filepath, O_RDWR);
1457 RTE_LOG(ERR, EAL, "Could not open %s\n",
1461 mapping_size = hp[i].size;
1462 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1463 mapping_size, PROT_READ | PROT_WRITE,
1465 close(fd); /* close file both on success and on failure */
1466 if (addr == MAP_FAILED ||
1467 addr != RTE_PTR_ADD(base_addr, offset)) {
1468 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1472 offset+=mapping_size;
1475 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1476 (unsigned long long)mcfg->memseg[s].len);
1479 /* unmap the hugepage config file, since we are done using it */
1486 for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1487 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1488 if (hp != NULL && hp != MAP_FAILED)
1492 if (fd_hugepage >= 0)
1498 rte_eal_using_phys_addrs(void)
1500 return phys_addrs_available;