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>
59 #include <rte_memory.h>
60 #include <rte_memzone.h>
61 #include <rte_launch.h>
63 #include <rte_eal_memconfig.h>
64 #include <rte_per_lcore.h>
65 #include <rte_lcore.h>
66 #include <rte_common.h>
67 #include <rte_string_fns.h>
69 #include "eal_private.h"
70 #include "eal_internal_cfg.h"
71 #include "eal_filesystem.h"
72 #include "eal_hugepages.h"
74 #define PFN_MASK_SIZE 8
76 #ifdef RTE_LIBRTE_XEN_DOM0
77 int rte_xen_dom0_supported(void)
79 return internal_config.xen_dom0_support;
85 * Huge page mapping under linux
87 * To reserve a big contiguous amount of memory, we use the hugepage
88 * feature of linux. For that, we need to have hugetlbfs mounted. This
89 * code will create many files in this directory (one per page) and
90 * map them in virtual memory. For each page, we will retrieve its
91 * physical address and remap it in order to have a virtual contiguous
92 * zone as well as a physical contiguous zone.
95 static uint64_t baseaddr_offset;
97 static bool phys_addrs_available = true;
99 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
102 test_phys_addrs_available(void)
105 phys_addr_t physaddr;
107 /* For dom0, phys addresses can always be available */
108 if (rte_xen_dom0_supported())
111 physaddr = rte_mem_virt2phy(&tmp);
112 if (physaddr == RTE_BAD_PHYS_ADDR) {
114 "Cannot obtain physical addresses: %s. "
115 "Only vfio will function.\n",
117 phys_addrs_available = false;
122 * Get physical address of any mapped virtual address in the current process.
125 rte_mem_virt2phy(const void *virtaddr)
128 uint64_t page, physaddr;
129 unsigned long virt_pfn;
133 /* when using dom0, /proc/self/pagemap always returns 0, check in
134 * dpdk memory by browsing the memsegs */
135 if (rte_xen_dom0_supported()) {
136 struct rte_mem_config *mcfg;
137 struct rte_memseg *memseg;
140 mcfg = rte_eal_get_configuration()->mem_config;
141 for (i = 0; i < RTE_MAX_MEMSEG; i++) {
142 memseg = &mcfg->memseg[i];
143 if (memseg->addr == NULL)
145 if (virtaddr > memseg->addr &&
146 virtaddr < RTE_PTR_ADD(memseg->addr,
148 return memseg->phys_addr +
149 RTE_PTR_DIFF(virtaddr, memseg->addr);
153 return RTE_BAD_PHYS_ADDR;
156 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
157 if (!phys_addrs_available)
158 return RTE_BAD_PHYS_ADDR;
160 /* standard page size */
161 page_size = getpagesize();
163 fd = open("/proc/self/pagemap", O_RDONLY);
165 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
166 __func__, strerror(errno));
167 return RTE_BAD_PHYS_ADDR;
170 virt_pfn = (unsigned long)virtaddr / page_size;
171 offset = sizeof(uint64_t) * virt_pfn;
172 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
173 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
174 __func__, strerror(errno));
176 return RTE_BAD_PHYS_ADDR;
179 retval = read(fd, &page, PFN_MASK_SIZE);
182 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
183 __func__, strerror(errno));
184 return RTE_BAD_PHYS_ADDR;
185 } else if (retval != PFN_MASK_SIZE) {
186 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
187 "but expected %d:\n",
188 __func__, retval, PFN_MASK_SIZE);
189 return RTE_BAD_PHYS_ADDR;
193 * the pfn (page frame number) are bits 0-54 (see
194 * pagemap.txt in linux Documentation)
196 if ((page & 0x7fffffffffffffULL) == 0)
197 return RTE_BAD_PHYS_ADDR;
199 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
200 + ((unsigned long)virtaddr % page_size);
206 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
207 * it by browsing the /proc/self/pagemap special file.
210 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
215 for (i = 0; i < hpi->num_pages[0]; i++) {
216 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
217 if (addr == RTE_BAD_PHYS_ADDR)
219 hugepg_tbl[i].physaddr = addr;
225 * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
228 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
231 static phys_addr_t addr;
233 for (i = 0; i < hpi->num_pages[0]; i++) {
234 hugepg_tbl[i].physaddr = addr;
235 addr += hugepg_tbl[i].size;
241 * Check whether address-space layout randomization is enabled in
242 * the kernel. This is important for multi-process as it can prevent
243 * two processes mapping data to the same virtual address
245 * 0 - address space randomization disabled
246 * 1/2 - address space randomization enabled
247 * negative error code on error
253 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
256 retval = read(fd, &c, 1);
266 default: return -EINVAL;
271 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
272 * pointer to the mmap'd area and keep *size unmodified. Else, retry
273 * with a smaller zone: decrease *size by hugepage_sz until it reaches
274 * 0. In this case, return NULL. Note: this function returns an address
275 * which is a multiple of hugepage size.
278 get_virtual_area(size_t *size, size_t hugepage_sz)
284 if (internal_config.base_virtaddr != 0) {
285 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
290 RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
292 fd = open("/dev/zero", O_RDONLY);
294 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
299 (*size) + hugepage_sz, PROT_READ,
300 #ifdef RTE_ARCH_PPC_64
301 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
306 if (addr == MAP_FAILED)
307 *size -= hugepage_sz;
308 } while (addr == MAP_FAILED && *size > 0);
310 if (addr == MAP_FAILED) {
312 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
317 munmap(addr, (*size) + hugepage_sz);
320 /* align addr to a huge page size boundary */
321 aligned_addr = (long)addr;
322 aligned_addr += (hugepage_sz - 1);
323 aligned_addr &= (~(hugepage_sz - 1));
324 addr = (void *)(aligned_addr);
326 RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
329 /* increment offset */
330 baseaddr_offset += *size;
335 static sigjmp_buf huge_jmpenv;
337 static void huge_sigbus_handler(int signo __rte_unused)
339 siglongjmp(huge_jmpenv, 1);
342 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
343 * non-static local variable in the stack frame calling sigsetjmp might be
344 * clobbered by a call to longjmp.
346 static int huge_wrap_sigsetjmp(void)
348 return sigsetjmp(huge_jmpenv, 1);
352 * Mmap all hugepages of hugepage table: it first open a file in
353 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
354 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
355 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
356 * map continguous physical blocks in contiguous virtual blocks.
359 map_all_hugepages(struct hugepage_file *hugepg_tbl,
360 struct hugepage_info *hpi, int orig)
365 void *vma_addr = NULL;
368 for (i = 0; i < hpi->num_pages[0]; i++) {
369 uint64_t hugepage_sz = hpi->hugepage_sz;
372 hugepg_tbl[i].file_id = i;
373 hugepg_tbl[i].size = hugepage_sz;
374 eal_get_hugefile_path(hugepg_tbl[i].filepath,
375 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
376 hugepg_tbl[i].file_id);
377 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
380 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
381 * original map address as final map address.
383 else if ((hugepage_sz == RTE_PGSIZE_1G)
384 || (hugepage_sz == RTE_PGSIZE_16G)) {
385 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
386 hugepg_tbl[i].orig_va = NULL;
390 else if (vma_len == 0) {
391 unsigned j, num_pages;
393 /* reserve a virtual area for next contiguous
394 * physical block: count the number of
395 * contiguous physical pages. */
396 for (j = i+1; j < hpi->num_pages[0] ; j++) {
397 #ifdef RTE_ARCH_PPC_64
398 /* The physical addresses are sorted in
399 * descending order on PPC64 */
400 if (hugepg_tbl[j].physaddr !=
401 hugepg_tbl[j-1].physaddr - hugepage_sz)
404 if (hugepg_tbl[j].physaddr !=
405 hugepg_tbl[j-1].physaddr + hugepage_sz)
410 vma_len = num_pages * hugepage_sz;
412 /* get the biggest virtual memory area up to
413 * vma_len. If it fails, vma_addr is NULL, so
414 * let the kernel provide the address. */
415 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
416 if (vma_addr == NULL)
417 vma_len = hugepage_sz;
420 /* try to create hugepage file */
421 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
423 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
428 /* map the segment, and populate page tables,
429 * the kernel fills this segment with zeros */
430 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
431 MAP_SHARED | MAP_POPULATE, fd, 0);
432 if (virtaddr == MAP_FAILED) {
433 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
440 hugepg_tbl[i].orig_va = virtaddr;
443 hugepg_tbl[i].final_va = virtaddr;
447 /* In linux, hugetlb limitations, like cgroup, are
448 * enforced at fault time instead of mmap(), even
449 * with the option of MAP_POPULATE. Kernel will send
450 * a SIGBUS signal. To avoid to be killed, save stack
451 * environment here, if SIGBUS happens, we can jump
454 if (huge_wrap_sigsetjmp()) {
455 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
456 "hugepages of size %u MB\n",
457 (unsigned)(hugepage_sz / 0x100000));
458 munmap(virtaddr, hugepage_sz);
460 unlink(hugepg_tbl[i].filepath);
463 *(int *)virtaddr = 0;
467 /* set shared flock on the file. */
468 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
469 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
470 __func__, strerror(errno));
477 vma_addr = (char *)vma_addr + hugepage_sz;
478 vma_len -= hugepage_sz;
484 /* Unmap all hugepages from original mapping */
486 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
489 for (i = 0; i < hpi->num_pages[0]; i++) {
490 if (hugepg_tbl[i].orig_va) {
491 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
492 hugepg_tbl[i].orig_va = NULL;
499 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
503 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
507 unsigned i, hp_count = 0;
510 char hugedir_str[PATH_MAX];
513 f = fopen("/proc/self/numa_maps", "r");
515 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
516 " consider that all memory is in socket_id 0\n");
520 snprintf(hugedir_str, sizeof(hugedir_str),
521 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
524 while (fgets(buf, sizeof(buf), f) != NULL) {
526 /* ignore non huge page */
527 if (strstr(buf, " huge ") == NULL &&
528 strstr(buf, hugedir_str) == NULL)
532 virt_addr = strtoull(buf, &end, 16);
533 if (virt_addr == 0 || end == buf) {
534 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
538 /* get node id (socket id) */
539 nodestr = strstr(buf, " N");
540 if (nodestr == NULL) {
541 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
545 end = strstr(nodestr, "=");
547 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
553 socket_id = strtoul(nodestr, &end, 0);
554 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
555 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
559 /* if we find this page in our mappings, set socket_id */
560 for (i = 0; i < hpi->num_pages[0]; i++) {
561 void *va = (void *)(unsigned long)virt_addr;
562 if (hugepg_tbl[i].orig_va == va) {
563 hugepg_tbl[i].socket_id = socket_id;
569 if (hp_count < hpi->num_pages[0])
581 cmp_physaddr(const void *a, const void *b)
583 #ifndef RTE_ARCH_PPC_64
584 const struct hugepage_file *p1 = a;
585 const struct hugepage_file *p2 = b;
587 /* PowerPC needs memory sorted in reverse order from x86 */
588 const struct hugepage_file *p1 = b;
589 const struct hugepage_file *p2 = a;
591 if (p1->physaddr < p2->physaddr)
593 else if (p1->physaddr > p2->physaddr)
600 * Uses mmap to create a shared memory area for storage of data
601 * Used in this file to store the hugepage file map on disk
604 create_shared_memory(const char *filename, const size_t mem_size)
607 int fd = open(filename, O_CREAT | O_RDWR, 0666);
610 if (ftruncate(fd, mem_size) < 0) {
614 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
620 * this copies *active* hugepages from one hugepage table to another.
621 * destination is typically the shared memory.
624 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
625 const struct hugepage_file * src, int src_size)
627 int src_pos, dst_pos = 0;
629 for (src_pos = 0; src_pos < src_size; src_pos++) {
630 if (src[src_pos].final_va != NULL) {
631 /* error on overflow attempt */
632 if (dst_pos == dest_size)
634 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
642 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
643 unsigned num_hp_info)
645 unsigned socket, size;
646 int page, nrpages = 0;
648 /* get total number of hugepages */
649 for (size = 0; size < num_hp_info; size++)
650 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
652 internal_config.hugepage_info[size].num_pages[socket];
654 for (page = 0; page < nrpages; page++) {
655 struct hugepage_file *hp = &hugepg_tbl[page];
657 if (hp->final_va != NULL && unlink(hp->filepath)) {
658 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
659 __func__, hp->filepath, strerror(errno));
666 * unmaps hugepages that are not going to be used. since we originally allocate
667 * ALL hugepages (not just those we need), additional unmapping needs to be done.
670 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
671 struct hugepage_info *hpi,
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++)
680 nrpages += internal_config.hugepage_info[size].num_pages[socket];
682 for (size = 0; size < num_hp_info; size++) {
683 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
684 unsigned pages_found = 0;
686 /* traverse until we have unmapped all the unused pages */
687 for (page = 0; page < nrpages; page++) {
688 struct hugepage_file *hp = &hugepg_tbl[page];
690 /* find a page that matches the criteria */
691 if ((hp->size == hpi[size].hugepage_sz) &&
692 (hp->socket_id == (int) socket)) {
694 /* if we skipped enough pages, unmap the rest */
695 if (pages_found == hpi[size].num_pages[socket]) {
698 unmap_len = hp->size;
700 /* get start addr and len of the remaining segment */
701 munmap(hp->final_va, (size_t) unmap_len);
704 if (unlink(hp->filepath) == -1) {
705 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
706 __func__, hp->filepath, strerror(errno));
710 /* lock the page and skip */
716 } /* foreach socket */
717 } /* foreach pagesize */
722 static inline uint64_t
723 get_socket_mem_size(int socket)
728 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
729 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
730 if (hpi->hugedir != NULL)
731 size += hpi->hugepage_sz * hpi->num_pages[socket];
738 * This function is a NUMA-aware equivalent of calc_num_pages.
739 * It takes in the list of hugepage sizes and the
740 * number of pages thereof, and calculates the best number of
741 * pages of each size to fulfill the request for <memory> ram
744 calc_num_pages_per_socket(uint64_t * memory,
745 struct hugepage_info *hp_info,
746 struct hugepage_info *hp_used,
747 unsigned num_hp_info)
749 unsigned socket, j, i = 0;
750 unsigned requested, available;
751 int total_num_pages = 0;
752 uint64_t remaining_mem, cur_mem;
753 uint64_t total_mem = internal_config.memory;
755 if (num_hp_info == 0)
758 /* if specific memory amounts per socket weren't requested */
759 if (internal_config.force_sockets == 0) {
760 int cpu_per_socket[RTE_MAX_NUMA_NODES];
761 size_t default_size, total_size;
764 /* Compute number of cores per socket */
765 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
766 RTE_LCORE_FOREACH(lcore_id) {
767 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
771 * Automatically spread requested memory amongst detected sockets according
772 * to number of cores from cpu mask present on each socket
774 total_size = internal_config.memory;
775 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
777 /* Set memory amount per socket */
778 default_size = (internal_config.memory * cpu_per_socket[socket])
781 /* Limit to maximum available memory on socket */
782 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
785 memory[socket] = default_size;
786 total_size -= default_size;
790 * If some memory is remaining, try to allocate it by getting all
791 * available memory from sockets, one after the other
793 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
794 /* take whatever is available */
795 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
799 memory[socket] += default_size;
800 total_size -= default_size;
804 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
805 /* skips if the memory on specific socket wasn't requested */
806 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
807 hp_used[i].hugedir = hp_info[i].hugedir;
808 hp_used[i].num_pages[socket] = RTE_MIN(
809 memory[socket] / hp_info[i].hugepage_sz,
810 hp_info[i].num_pages[socket]);
812 cur_mem = hp_used[i].num_pages[socket] *
813 hp_used[i].hugepage_sz;
815 memory[socket] -= cur_mem;
816 total_mem -= cur_mem;
818 total_num_pages += hp_used[i].num_pages[socket];
820 /* check if we have met all memory requests */
821 if (memory[socket] == 0)
824 /* check if we have any more pages left at this size, if so
825 * move on to next size */
826 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
828 /* At this point we know that there are more pages available that are
829 * bigger than the memory we want, so lets see if we can get enough
830 * from other page sizes.
833 for (j = i+1; j < num_hp_info; j++)
834 remaining_mem += hp_info[j].hugepage_sz *
835 hp_info[j].num_pages[socket];
837 /* is there enough other memory, if not allocate another page and quit */
838 if (remaining_mem < memory[socket]){
839 cur_mem = RTE_MIN(memory[socket],
840 hp_info[i].hugepage_sz);
841 memory[socket] -= cur_mem;
842 total_mem -= cur_mem;
843 hp_used[i].num_pages[socket]++;
845 break; /* we are done with this socket*/
848 /* if we didn't satisfy all memory requirements per socket */
849 if (memory[socket] > 0) {
850 /* to prevent icc errors */
851 requested = (unsigned) (internal_config.socket_mem[socket] /
853 available = requested -
854 ((unsigned) (memory[socket] / 0x100000));
855 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
856 "Requested: %uMB, available: %uMB\n", socket,
857 requested, available);
862 /* if we didn't satisfy total memory requirements */
864 requested = (unsigned) (internal_config.memory / 0x100000);
865 available = requested - (unsigned) (total_mem / 0x100000);
866 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
867 " available: %uMB\n", requested, available);
870 return total_num_pages;
874 eal_get_hugepage_mem_size(void)
879 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
880 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
881 if (hpi->hugedir != NULL) {
882 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
883 size += hpi->hugepage_sz * hpi->num_pages[j];
888 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
891 static struct sigaction huge_action_old;
892 static int huge_need_recover;
895 huge_register_sigbus(void)
898 struct sigaction action;
901 sigaddset(&mask, SIGBUS);
903 action.sa_mask = mask;
904 action.sa_handler = huge_sigbus_handler;
906 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
910 huge_recover_sigbus(void)
912 if (huge_need_recover) {
913 sigaction(SIGBUS, &huge_action_old, NULL);
914 huge_need_recover = 0;
919 * Prepare physical memory mapping: fill configuration structure with
920 * these infos, return 0 on success.
921 * 1. map N huge pages in separate files in hugetlbfs
922 * 2. find associated physical addr
923 * 3. find associated NUMA socket ID
924 * 4. sort all huge pages by physical address
925 * 5. remap these N huge pages in the correct order
926 * 6. unmap the first mapping
927 * 7. fill memsegs in configuration with contiguous zones
930 rte_eal_hugepage_init(void)
932 struct rte_mem_config *mcfg;
933 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
934 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
936 uint64_t memory[RTE_MAX_NUMA_NODES];
939 int i, j, new_memseg;
940 int nr_hugefiles, nr_hugepages = 0;
943 test_phys_addrs_available();
945 memset(used_hp, 0, sizeof(used_hp));
947 /* get pointer to global configuration */
948 mcfg = rte_eal_get_configuration()->mem_config;
950 /* hugetlbfs can be disabled */
951 if (internal_config.no_hugetlbfs) {
952 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
953 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
954 if (addr == MAP_FAILED) {
955 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
959 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
960 mcfg->memseg[0].addr = addr;
961 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
962 mcfg->memseg[0].len = internal_config.memory;
963 mcfg->memseg[0].socket_id = 0;
967 /* check if app runs on Xen Dom0 */
968 if (internal_config.xen_dom0_support) {
969 #ifdef RTE_LIBRTE_XEN_DOM0
970 /* use dom0_mm kernel driver to init memory */
971 if (rte_xen_dom0_memory_init() < 0)
978 /* calculate total number of hugepages available. at this point we haven't
979 * yet started sorting them so they all are on socket 0 */
980 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
981 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
982 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
984 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
988 * allocate a memory area for hugepage table.
989 * this isn't shared memory yet. due to the fact that we need some
990 * processing done on these pages, shared memory will be created
993 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
997 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
999 hp_offset = 0; /* where we start the current page size entries */
1001 huge_register_sigbus();
1003 /* map all hugepages and sort them */
1004 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1005 unsigned pages_old, pages_new;
1006 struct hugepage_info *hpi;
1009 * we don't yet mark hugepages as used at this stage, so
1010 * we just map all hugepages available to the system
1011 * all hugepages are still located on socket 0
1013 hpi = &internal_config.hugepage_info[i];
1015 if (hpi->num_pages[0] == 0)
1018 /* map all hugepages available */
1019 pages_old = hpi->num_pages[0];
1020 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, 1);
1021 if (pages_new < pages_old) {
1023 "%d not %d hugepages of size %u MB allocated\n",
1024 pages_new, pages_old,
1025 (unsigned)(hpi->hugepage_sz / 0x100000));
1027 int pages = pages_old - pages_new;
1029 nr_hugepages -= pages;
1030 hpi->num_pages[0] = pages_new;
1035 if (phys_addrs_available) {
1036 /* find physical addresses for each hugepage */
1037 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1038 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1039 "for %u MB pages\n",
1040 (unsigned int)(hpi->hugepage_sz / 0x100000));
1044 /* set physical addresses for each hugepage */
1045 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1046 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1047 "for %u MB pages\n",
1048 (unsigned int)(hpi->hugepage_sz / 0x100000));
1053 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1054 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1055 (unsigned)(hpi->hugepage_sz / 0x100000));
1059 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1060 sizeof(struct hugepage_file), cmp_physaddr);
1062 /* remap all hugepages */
1063 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) !=
1064 hpi->num_pages[0]) {
1065 RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1066 (unsigned)(hpi->hugepage_sz / 0x100000));
1070 /* unmap original mappings */
1071 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1074 /* we have processed a num of hugepages of this size, so inc offset */
1075 hp_offset += hpi->num_pages[0];
1078 huge_recover_sigbus();
1080 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1081 internal_config.memory = eal_get_hugepage_mem_size();
1083 nr_hugefiles = nr_hugepages;
1086 /* clean out the numbers of pages */
1087 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1088 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1089 internal_config.hugepage_info[i].num_pages[j] = 0;
1091 /* get hugepages for each socket */
1092 for (i = 0; i < nr_hugefiles; i++) {
1093 int socket = tmp_hp[i].socket_id;
1095 /* find a hugepage info with right size and increment num_pages */
1096 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1097 (int)internal_config.num_hugepage_sizes);
1098 for (j = 0; j < nb_hpsizes; j++) {
1099 if (tmp_hp[i].size ==
1100 internal_config.hugepage_info[j].hugepage_sz) {
1101 internal_config.hugepage_info[j].num_pages[socket]++;
1106 /* make a copy of socket_mem, needed for number of pages calculation */
1107 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1108 memory[i] = internal_config.socket_mem[i];
1110 /* calculate final number of pages */
1111 nr_hugepages = calc_num_pages_per_socket(memory,
1112 internal_config.hugepage_info, used_hp,
1113 internal_config.num_hugepage_sizes);
1115 /* error if not enough memory available */
1116 if (nr_hugepages < 0)
1120 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1121 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1122 if (used_hp[i].num_pages[j] > 0) {
1124 "Requesting %u pages of size %uMB"
1125 " from socket %i\n",
1126 used_hp[i].num_pages[j],
1128 (used_hp[i].hugepage_sz / 0x100000),
1134 /* create shared memory */
1135 hugepage = create_shared_memory(eal_hugepage_info_path(),
1136 nr_hugefiles * sizeof(struct hugepage_file));
1138 if (hugepage == NULL) {
1139 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1142 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1145 * unmap pages that we won't need (looks at used_hp).
1146 * also, sets final_va to NULL on pages that were unmapped.
1148 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1149 internal_config.num_hugepage_sizes) < 0) {
1150 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1155 * copy stuff from malloc'd hugepage* to the actual shared memory.
1156 * this procedure only copies those hugepages that have final_va
1157 * not NULL. has overflow protection.
1159 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1160 tmp_hp, nr_hugefiles) < 0) {
1161 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1165 /* free the hugepage backing files */
1166 if (internal_config.hugepage_unlink &&
1167 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1168 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1172 /* free the temporary hugepage table */
1176 /* first memseg index shall be 0 after incrementing it below */
1178 for (i = 0; i < nr_hugefiles; i++) {
1181 /* if this is a new section, create a new memseg */
1184 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1186 else if (hugepage[i].size != hugepage[i-1].size)
1189 #ifdef RTE_ARCH_PPC_64
1190 /* On PPC64 architecture, the mmap always start from higher
1191 * virtual address to lower address. Here, both the physical
1192 * address and virtual address are in descending order */
1193 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1196 else if (((unsigned long)hugepage[i-1].final_va -
1197 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1200 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1203 else if (((unsigned long)hugepage[i].final_va -
1204 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1210 if (j == RTE_MAX_MEMSEG)
1213 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1214 mcfg->memseg[j].addr = hugepage[i].final_va;
1215 mcfg->memseg[j].len = hugepage[i].size;
1216 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1217 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1219 /* continuation of previous memseg */
1221 #ifdef RTE_ARCH_PPC_64
1222 /* Use the phy and virt address of the last page as segment
1223 * address for IBM Power architecture */
1224 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1225 mcfg->memseg[j].addr = hugepage[i].final_va;
1227 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1229 hugepage[i].memseg_id = j;
1232 if (i < nr_hugefiles) {
1233 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1234 "from %d requested\n"
1235 "Current %s=%d is not enough\n"
1236 "Please either increase it or request less amount "
1238 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1243 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1248 huge_recover_sigbus();
1250 if (hugepage != NULL)
1251 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1257 * uses fstat to report the size of a file on disk
1263 if (fstat(fd, &st) < 0)
1269 * This creates the memory mappings in the secondary process to match that of
1270 * the server process. It goes through each memory segment in the DPDK runtime
1271 * configuration and finds the hugepages which form that segment, mapping them
1272 * in order to form a contiguous block in the virtual memory space
1275 rte_eal_hugepage_attach(void)
1277 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1278 struct hugepage_file *hp = NULL;
1279 unsigned num_hp = 0;
1280 unsigned i, s = 0; /* s used to track the segment number */
1281 unsigned max_seg = RTE_MAX_MEMSEG;
1283 int fd, fd_zero = -1, fd_hugepage = -1;
1285 if (aslr_enabled() > 0) {
1286 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1287 "(ASLR) is enabled in the kernel.\n");
1288 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1289 "into secondary processes\n");
1292 test_phys_addrs_available();
1294 if (internal_config.xen_dom0_support) {
1295 #ifdef RTE_LIBRTE_XEN_DOM0
1296 if (rte_xen_dom0_memory_attach() < 0) {
1297 RTE_LOG(ERR, EAL, "Failed to attach memory segments of primary "
1305 fd_zero = open("/dev/zero", O_RDONLY);
1307 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1310 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1311 if (fd_hugepage < 0) {
1312 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1316 /* map all segments into memory to make sure we get the addrs */
1317 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1321 * the first memory segment with len==0 is the one that
1322 * follows the last valid segment.
1324 if (mcfg->memseg[s].len == 0)
1328 * fdzero is mmapped to get a contiguous block of virtual
1329 * addresses of the appropriate memseg size.
1330 * use mmap to get identical addresses as the primary process.
1332 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1334 #ifdef RTE_ARCH_PPC_64
1335 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1340 if (base_addr == MAP_FAILED ||
1341 base_addr != mcfg->memseg[s].addr) {
1343 if (base_addr != MAP_FAILED) {
1344 /* errno is stale, don't use */
1345 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1346 "in /dev/zero at [%p], got [%p] - "
1347 "please use '--base-virtaddr' option\n",
1348 (unsigned long long)mcfg->memseg[s].len,
1349 mcfg->memseg[s].addr, base_addr);
1350 munmap(base_addr, mcfg->memseg[s].len);
1352 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1353 "in /dev/zero at [%p]: '%s'\n",
1354 (unsigned long long)mcfg->memseg[s].len,
1355 mcfg->memseg[s].addr, strerror(errno));
1357 if (aslr_enabled() > 0) {
1358 RTE_LOG(ERR, EAL, "It is recommended to "
1359 "disable ASLR in the kernel "
1360 "and retry running both primary "
1361 "and secondary processes\n");
1367 size = getFileSize(fd_hugepage);
1368 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1369 if (hp == MAP_FAILED) {
1370 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1374 num_hp = size / sizeof(struct hugepage_file);
1375 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1378 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1379 void *addr, *base_addr;
1380 uintptr_t offset = 0;
1381 size_t mapping_size;
1383 * free previously mapped memory so we can map the
1384 * hugepages into the space
1386 base_addr = mcfg->memseg[s].addr;
1387 munmap(base_addr, mcfg->memseg[s].len);
1389 /* find the hugepages for this segment and map them
1390 * we don't need to worry about order, as the server sorted the
1391 * entries before it did the second mmap of them */
1392 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1393 if (hp[i].memseg_id == (int)s){
1394 fd = open(hp[i].filepath, O_RDWR);
1396 RTE_LOG(ERR, EAL, "Could not open %s\n",
1400 mapping_size = hp[i].size;
1401 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1402 mapping_size, PROT_READ | PROT_WRITE,
1404 close(fd); /* close file both on success and on failure */
1405 if (addr == MAP_FAILED ||
1406 addr != RTE_PTR_ADD(base_addr, offset)) {
1407 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1411 offset+=mapping_size;
1414 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1415 (unsigned long long)mcfg->memseg[s].len);
1418 /* unmap the hugepage config file, since we are done using it */
1425 for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1426 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1427 if (hp != NULL && hp != MAP_FAILED)
1431 if (fd_hugepage >= 0)
1437 rte_eal_using_phys_addrs(void)
1439 return phys_addrs_available;