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;
121 /* Lock page in physical memory and prevent from swapping. */
123 rte_mem_lock_page(const void *virt)
125 unsigned long virtual = (unsigned long)virt;
126 int page_size = getpagesize();
127 unsigned long aligned = (virtual & ~ (page_size - 1));
128 return mlock((void*)aligned, page_size);
132 * Get physical address of any mapped virtual address in the current process.
135 rte_mem_virt2phy(const void *virtaddr)
138 uint64_t page, physaddr;
139 unsigned long virt_pfn;
143 /* when using dom0, /proc/self/pagemap always returns 0, check in
144 * dpdk memory by browsing the memsegs */
145 if (rte_xen_dom0_supported()) {
146 struct rte_mem_config *mcfg;
147 struct rte_memseg *memseg;
150 mcfg = rte_eal_get_configuration()->mem_config;
151 for (i = 0; i < RTE_MAX_MEMSEG; i++) {
152 memseg = &mcfg->memseg[i];
153 if (memseg->addr == NULL)
155 if (virtaddr > memseg->addr &&
156 virtaddr < RTE_PTR_ADD(memseg->addr,
158 return memseg->phys_addr +
159 RTE_PTR_DIFF(virtaddr, memseg->addr);
163 return RTE_BAD_PHYS_ADDR;
166 /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
167 if (!phys_addrs_available)
168 return RTE_BAD_PHYS_ADDR;
170 /* standard page size */
171 page_size = getpagesize();
173 fd = open("/proc/self/pagemap", O_RDONLY);
175 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
176 __func__, strerror(errno));
177 return RTE_BAD_PHYS_ADDR;
180 virt_pfn = (unsigned long)virtaddr / page_size;
181 offset = sizeof(uint64_t) * virt_pfn;
182 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
183 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
184 __func__, strerror(errno));
186 return RTE_BAD_PHYS_ADDR;
189 retval = read(fd, &page, PFN_MASK_SIZE);
192 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
193 __func__, strerror(errno));
194 return RTE_BAD_PHYS_ADDR;
195 } else if (retval != PFN_MASK_SIZE) {
196 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
197 "but expected %d:\n",
198 __func__, retval, PFN_MASK_SIZE);
199 return RTE_BAD_PHYS_ADDR;
203 * the pfn (page frame number) are bits 0-54 (see
204 * pagemap.txt in linux Documentation)
206 if ((page & 0x7fffffffffffffULL) == 0)
207 return RTE_BAD_PHYS_ADDR;
209 physaddr = ((page & 0x7fffffffffffffULL) * page_size)
210 + ((unsigned long)virtaddr % page_size);
216 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
217 * it by browsing the /proc/self/pagemap special file.
220 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
225 for (i = 0; i < hpi->num_pages[0]; i++) {
226 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
227 if (addr == RTE_BAD_PHYS_ADDR)
229 hugepg_tbl[i].physaddr = addr;
235 * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
238 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
241 static phys_addr_t addr;
243 for (i = 0; i < hpi->num_pages[0]; i++) {
244 hugepg_tbl[i].physaddr = addr;
245 addr += hugepg_tbl[i].size;
251 * Check whether address-space layout randomization is enabled in
252 * the kernel. This is important for multi-process as it can prevent
253 * two processes mapping data to the same virtual address
255 * 0 - address space randomization disabled
256 * 1/2 - address space randomization enabled
257 * negative error code on error
263 int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
266 retval = read(fd, &c, 1);
276 default: return -EINVAL;
281 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
282 * pointer to the mmap'd area and keep *size unmodified. Else, retry
283 * with a smaller zone: decrease *size by hugepage_sz until it reaches
284 * 0. In this case, return NULL. Note: this function returns an address
285 * which is a multiple of hugepage size.
288 get_virtual_area(size_t *size, size_t hugepage_sz)
294 if (internal_config.base_virtaddr != 0) {
295 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
300 RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
302 fd = open("/dev/zero", O_RDONLY);
304 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
309 (*size) + hugepage_sz, PROT_READ,
310 #ifdef RTE_ARCH_PPC_64
311 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
316 if (addr == MAP_FAILED)
317 *size -= hugepage_sz;
318 } while (addr == MAP_FAILED && *size > 0);
320 if (addr == MAP_FAILED) {
322 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
327 munmap(addr, (*size) + hugepage_sz);
330 /* align addr to a huge page size boundary */
331 aligned_addr = (long)addr;
332 aligned_addr += (hugepage_sz - 1);
333 aligned_addr &= (~(hugepage_sz - 1));
334 addr = (void *)(aligned_addr);
336 RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
339 /* increment offset */
340 baseaddr_offset += *size;
345 static sigjmp_buf huge_jmpenv;
347 static void huge_sigbus_handler(int signo __rte_unused)
349 siglongjmp(huge_jmpenv, 1);
352 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
353 * non-static local variable in the stack frame calling sigsetjmp might be
354 * clobbered by a call to longjmp.
356 static int huge_wrap_sigsetjmp(void)
358 return sigsetjmp(huge_jmpenv, 1);
362 * Mmap all hugepages of hugepage table: it first open a file in
363 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
364 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
365 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
366 * map continguous physical blocks in contiguous virtual blocks.
369 map_all_hugepages(struct hugepage_file *hugepg_tbl,
370 struct hugepage_info *hpi, int orig)
375 void *vma_addr = NULL;
378 for (i = 0; i < hpi->num_pages[0]; i++) {
379 uint64_t hugepage_sz = hpi->hugepage_sz;
382 hugepg_tbl[i].file_id = i;
383 hugepg_tbl[i].size = hugepage_sz;
384 eal_get_hugefile_path(hugepg_tbl[i].filepath,
385 sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
386 hugepg_tbl[i].file_id);
387 hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
390 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
391 * original map address as final map address.
393 else if ((hugepage_sz == RTE_PGSIZE_1G)
394 || (hugepage_sz == RTE_PGSIZE_16G)) {
395 hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
396 hugepg_tbl[i].orig_va = NULL;
400 else if (vma_len == 0) {
401 unsigned j, num_pages;
403 /* reserve a virtual area for next contiguous
404 * physical block: count the number of
405 * contiguous physical pages. */
406 for (j = i+1; j < hpi->num_pages[0] ; j++) {
407 #ifdef RTE_ARCH_PPC_64
408 /* The physical addresses are sorted in
409 * descending order on PPC64 */
410 if (hugepg_tbl[j].physaddr !=
411 hugepg_tbl[j-1].physaddr - hugepage_sz)
414 if (hugepg_tbl[j].physaddr !=
415 hugepg_tbl[j-1].physaddr + hugepage_sz)
420 vma_len = num_pages * hugepage_sz;
422 /* get the biggest virtual memory area up to
423 * vma_len. If it fails, vma_addr is NULL, so
424 * let the kernel provide the address. */
425 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
426 if (vma_addr == NULL)
427 vma_len = hugepage_sz;
430 /* try to create hugepage file */
431 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
433 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
438 /* map the segment, and populate page tables,
439 * the kernel fills this segment with zeros */
440 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
441 MAP_SHARED | MAP_POPULATE, fd, 0);
442 if (virtaddr == MAP_FAILED) {
443 RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
450 hugepg_tbl[i].orig_va = virtaddr;
453 hugepg_tbl[i].final_va = virtaddr;
457 /* In linux, hugetlb limitations, like cgroup, are
458 * enforced at fault time instead of mmap(), even
459 * with the option of MAP_POPULATE. Kernel will send
460 * a SIGBUS signal. To avoid to be killed, save stack
461 * environment here, if SIGBUS happens, we can jump
464 if (huge_wrap_sigsetjmp()) {
465 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
466 "hugepages of size %u MB\n",
467 (unsigned)(hugepage_sz / 0x100000));
468 munmap(virtaddr, hugepage_sz);
470 unlink(hugepg_tbl[i].filepath);
473 *(int *)virtaddr = 0;
477 /* set shared flock on the file. */
478 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
479 RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
480 __func__, strerror(errno));
487 vma_addr = (char *)vma_addr + hugepage_sz;
488 vma_len -= hugepage_sz;
494 /* Unmap all hugepages from original mapping */
496 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
499 for (i = 0; i < hpi->num_pages[0]; i++) {
500 if (hugepg_tbl[i].orig_va) {
501 munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
502 hugepg_tbl[i].orig_va = NULL;
509 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
513 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
517 unsigned i, hp_count = 0;
520 char hugedir_str[PATH_MAX];
523 f = fopen("/proc/self/numa_maps", "r");
525 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
526 " consider that all memory is in socket_id 0\n");
530 snprintf(hugedir_str, sizeof(hugedir_str),
531 "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
534 while (fgets(buf, sizeof(buf), f) != NULL) {
536 /* ignore non huge page */
537 if (strstr(buf, " huge ") == NULL &&
538 strstr(buf, hugedir_str) == NULL)
542 virt_addr = strtoull(buf, &end, 16);
543 if (virt_addr == 0 || end == buf) {
544 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
548 /* get node id (socket id) */
549 nodestr = strstr(buf, " N");
550 if (nodestr == NULL) {
551 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
555 end = strstr(nodestr, "=");
557 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
563 socket_id = strtoul(nodestr, &end, 0);
564 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
565 RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
569 /* if we find this page in our mappings, set socket_id */
570 for (i = 0; i < hpi->num_pages[0]; i++) {
571 void *va = (void *)(unsigned long)virt_addr;
572 if (hugepg_tbl[i].orig_va == va) {
573 hugepg_tbl[i].socket_id = socket_id;
579 if (hp_count < hpi->num_pages[0])
591 cmp_physaddr(const void *a, const void *b)
593 #ifndef RTE_ARCH_PPC_64
594 const struct hugepage_file *p1 = a;
595 const struct hugepage_file *p2 = b;
597 /* PowerPC needs memory sorted in reverse order from x86 */
598 const struct hugepage_file *p1 = b;
599 const struct hugepage_file *p2 = a;
601 if (p1->physaddr < p2->physaddr)
603 else if (p1->physaddr > p2->physaddr)
610 * Uses mmap to create a shared memory area for storage of data
611 * Used in this file to store the hugepage file map on disk
614 create_shared_memory(const char *filename, const size_t mem_size)
617 int fd = open(filename, O_CREAT | O_RDWR, 0666);
620 if (ftruncate(fd, mem_size) < 0) {
624 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
630 * this copies *active* hugepages from one hugepage table to another.
631 * destination is typically the shared memory.
634 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
635 const struct hugepage_file * src, int src_size)
637 int src_pos, dst_pos = 0;
639 for (src_pos = 0; src_pos < src_size; src_pos++) {
640 if (src[src_pos].final_va != NULL) {
641 /* error on overflow attempt */
642 if (dst_pos == dest_size)
644 memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
652 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
653 unsigned num_hp_info)
655 unsigned socket, size;
656 int page, nrpages = 0;
658 /* get total number of hugepages */
659 for (size = 0; size < num_hp_info; size++)
660 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
662 internal_config.hugepage_info[size].num_pages[socket];
664 for (page = 0; page < nrpages; page++) {
665 struct hugepage_file *hp = &hugepg_tbl[page];
667 if (hp->final_va != NULL && unlink(hp->filepath)) {
668 RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
669 __func__, hp->filepath, strerror(errno));
676 * unmaps hugepages that are not going to be used. since we originally allocate
677 * ALL hugepages (not just those we need), additional unmapping needs to be done.
680 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
681 struct hugepage_info *hpi,
682 unsigned num_hp_info)
684 unsigned socket, size;
685 int page, nrpages = 0;
687 /* get total number of hugepages */
688 for (size = 0; size < num_hp_info; size++)
689 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
690 nrpages += internal_config.hugepage_info[size].num_pages[socket];
692 for (size = 0; size < num_hp_info; size++) {
693 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
694 unsigned pages_found = 0;
696 /* traverse until we have unmapped all the unused pages */
697 for (page = 0; page < nrpages; page++) {
698 struct hugepage_file *hp = &hugepg_tbl[page];
700 /* find a page that matches the criteria */
701 if ((hp->size == hpi[size].hugepage_sz) &&
702 (hp->socket_id == (int) socket)) {
704 /* if we skipped enough pages, unmap the rest */
705 if (pages_found == hpi[size].num_pages[socket]) {
708 unmap_len = hp->size;
710 /* get start addr and len of the remaining segment */
711 munmap(hp->final_va, (size_t) unmap_len);
714 if (unlink(hp->filepath) == -1) {
715 RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
716 __func__, hp->filepath, strerror(errno));
720 /* lock the page and skip */
726 } /* foreach socket */
727 } /* foreach pagesize */
732 static inline uint64_t
733 get_socket_mem_size(int socket)
738 for (i = 0; i < internal_config.num_hugepage_sizes; i++){
739 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
740 if (hpi->hugedir != NULL)
741 size += hpi->hugepage_sz * hpi->num_pages[socket];
748 * This function is a NUMA-aware equivalent of calc_num_pages.
749 * It takes in the list of hugepage sizes and the
750 * number of pages thereof, and calculates the best number of
751 * pages of each size to fulfill the request for <memory> ram
754 calc_num_pages_per_socket(uint64_t * memory,
755 struct hugepage_info *hp_info,
756 struct hugepage_info *hp_used,
757 unsigned num_hp_info)
759 unsigned socket, j, i = 0;
760 unsigned requested, available;
761 int total_num_pages = 0;
762 uint64_t remaining_mem, cur_mem;
763 uint64_t total_mem = internal_config.memory;
765 if (num_hp_info == 0)
768 /* if specific memory amounts per socket weren't requested */
769 if (internal_config.force_sockets == 0) {
770 int cpu_per_socket[RTE_MAX_NUMA_NODES];
771 size_t default_size, total_size;
774 /* Compute number of cores per socket */
775 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
776 RTE_LCORE_FOREACH(lcore_id) {
777 cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
781 * Automatically spread requested memory amongst detected sockets according
782 * to number of cores from cpu mask present on each socket
784 total_size = internal_config.memory;
785 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
787 /* Set memory amount per socket */
788 default_size = (internal_config.memory * cpu_per_socket[socket])
791 /* Limit to maximum available memory on socket */
792 default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
795 memory[socket] = default_size;
796 total_size -= default_size;
800 * If some memory is remaining, try to allocate it by getting all
801 * available memory from sockets, one after the other
803 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
804 /* take whatever is available */
805 default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
809 memory[socket] += default_size;
810 total_size -= default_size;
814 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
815 /* skips if the memory on specific socket wasn't requested */
816 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
817 hp_used[i].hugedir = hp_info[i].hugedir;
818 hp_used[i].num_pages[socket] = RTE_MIN(
819 memory[socket] / hp_info[i].hugepage_sz,
820 hp_info[i].num_pages[socket]);
822 cur_mem = hp_used[i].num_pages[socket] *
823 hp_used[i].hugepage_sz;
825 memory[socket] -= cur_mem;
826 total_mem -= cur_mem;
828 total_num_pages += hp_used[i].num_pages[socket];
830 /* check if we have met all memory requests */
831 if (memory[socket] == 0)
834 /* check if we have any more pages left at this size, if so
835 * move on to next size */
836 if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
838 /* At this point we know that there are more pages available that are
839 * bigger than the memory we want, so lets see if we can get enough
840 * from other page sizes.
843 for (j = i+1; j < num_hp_info; j++)
844 remaining_mem += hp_info[j].hugepage_sz *
845 hp_info[j].num_pages[socket];
847 /* is there enough other memory, if not allocate another page and quit */
848 if (remaining_mem < memory[socket]){
849 cur_mem = RTE_MIN(memory[socket],
850 hp_info[i].hugepage_sz);
851 memory[socket] -= cur_mem;
852 total_mem -= cur_mem;
853 hp_used[i].num_pages[socket]++;
855 break; /* we are done with this socket*/
858 /* if we didn't satisfy all memory requirements per socket */
859 if (memory[socket] > 0) {
860 /* to prevent icc errors */
861 requested = (unsigned) (internal_config.socket_mem[socket] /
863 available = requested -
864 ((unsigned) (memory[socket] / 0x100000));
865 RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
866 "Requested: %uMB, available: %uMB\n", socket,
867 requested, available);
872 /* if we didn't satisfy total memory requirements */
874 requested = (unsigned) (internal_config.memory / 0x100000);
875 available = requested - (unsigned) (total_mem / 0x100000);
876 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
877 " available: %uMB\n", requested, available);
880 return total_num_pages;
884 eal_get_hugepage_mem_size(void)
889 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
890 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
891 if (hpi->hugedir != NULL) {
892 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
893 size += hpi->hugepage_sz * hpi->num_pages[j];
898 return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
901 static struct sigaction huge_action_old;
902 static int huge_need_recover;
905 huge_register_sigbus(void)
908 struct sigaction action;
911 sigaddset(&mask, SIGBUS);
913 action.sa_mask = mask;
914 action.sa_handler = huge_sigbus_handler;
916 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
920 huge_recover_sigbus(void)
922 if (huge_need_recover) {
923 sigaction(SIGBUS, &huge_action_old, NULL);
924 huge_need_recover = 0;
929 * Prepare physical memory mapping: fill configuration structure with
930 * these infos, return 0 on success.
931 * 1. map N huge pages in separate files in hugetlbfs
932 * 2. find associated physical addr
933 * 3. find associated NUMA socket ID
934 * 4. sort all huge pages by physical address
935 * 5. remap these N huge pages in the correct order
936 * 6. unmap the first mapping
937 * 7. fill memsegs in configuration with contiguous zones
940 rte_eal_hugepage_init(void)
942 struct rte_mem_config *mcfg;
943 struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
944 struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
946 uint64_t memory[RTE_MAX_NUMA_NODES];
949 int i, j, new_memseg;
950 int nr_hugefiles, nr_hugepages = 0;
953 test_phys_addrs_available();
955 memset(used_hp, 0, sizeof(used_hp));
957 /* get pointer to global configuration */
958 mcfg = rte_eal_get_configuration()->mem_config;
960 /* hugetlbfs can be disabled */
961 if (internal_config.no_hugetlbfs) {
962 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
963 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
964 if (addr == MAP_FAILED) {
965 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
969 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
970 mcfg->memseg[0].addr = addr;
971 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
972 mcfg->memseg[0].len = internal_config.memory;
973 mcfg->memseg[0].socket_id = 0;
977 /* check if app runs on Xen Dom0 */
978 if (internal_config.xen_dom0_support) {
979 #ifdef RTE_LIBRTE_XEN_DOM0
980 /* use dom0_mm kernel driver to init memory */
981 if (rte_xen_dom0_memory_init() < 0)
988 /* calculate total number of hugepages available. at this point we haven't
989 * yet started sorting them so they all are on socket 0 */
990 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
991 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
992 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
994 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
998 * allocate a memory area for hugepage table.
999 * this isn't shared memory yet. due to the fact that we need some
1000 * processing done on these pages, shared memory will be created
1003 tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1007 memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1009 hp_offset = 0; /* where we start the current page size entries */
1011 huge_register_sigbus();
1013 /* map all hugepages and sort them */
1014 for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1015 unsigned pages_old, pages_new;
1016 struct hugepage_info *hpi;
1019 * we don't yet mark hugepages as used at this stage, so
1020 * we just map all hugepages available to the system
1021 * all hugepages are still located on socket 0
1023 hpi = &internal_config.hugepage_info[i];
1025 if (hpi->num_pages[0] == 0)
1028 /* map all hugepages available */
1029 pages_old = hpi->num_pages[0];
1030 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, 1);
1031 if (pages_new < pages_old) {
1033 "%d not %d hugepages of size %u MB allocated\n",
1034 pages_new, pages_old,
1035 (unsigned)(hpi->hugepage_sz / 0x100000));
1037 int pages = pages_old - pages_new;
1039 nr_hugepages -= pages;
1040 hpi->num_pages[0] = pages_new;
1045 if (phys_addrs_available) {
1046 /* find physical addresses for each hugepage */
1047 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1048 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1049 "for %u MB pages\n",
1050 (unsigned int)(hpi->hugepage_sz / 0x100000));
1054 /* set physical addresses for each hugepage */
1055 if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1056 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1057 "for %u MB pages\n",
1058 (unsigned int)(hpi->hugepage_sz / 0x100000));
1063 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1064 RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1065 (unsigned)(hpi->hugepage_sz / 0x100000));
1069 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1070 sizeof(struct hugepage_file), cmp_physaddr);
1072 /* remap all hugepages */
1073 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) !=
1074 hpi->num_pages[0]) {
1075 RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1076 (unsigned)(hpi->hugepage_sz / 0x100000));
1080 /* unmap original mappings */
1081 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1084 /* we have processed a num of hugepages of this size, so inc offset */
1085 hp_offset += hpi->num_pages[0];
1088 huge_recover_sigbus();
1090 if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1091 internal_config.memory = eal_get_hugepage_mem_size();
1093 nr_hugefiles = nr_hugepages;
1096 /* clean out the numbers of pages */
1097 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1098 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1099 internal_config.hugepage_info[i].num_pages[j] = 0;
1101 /* get hugepages for each socket */
1102 for (i = 0; i < nr_hugefiles; i++) {
1103 int socket = tmp_hp[i].socket_id;
1105 /* find a hugepage info with right size and increment num_pages */
1106 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1107 (int)internal_config.num_hugepage_sizes);
1108 for (j = 0; j < nb_hpsizes; j++) {
1109 if (tmp_hp[i].size ==
1110 internal_config.hugepage_info[j].hugepage_sz) {
1111 internal_config.hugepage_info[j].num_pages[socket]++;
1116 /* make a copy of socket_mem, needed for number of pages calculation */
1117 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1118 memory[i] = internal_config.socket_mem[i];
1120 /* calculate final number of pages */
1121 nr_hugepages = calc_num_pages_per_socket(memory,
1122 internal_config.hugepage_info, used_hp,
1123 internal_config.num_hugepage_sizes);
1125 /* error if not enough memory available */
1126 if (nr_hugepages < 0)
1130 for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1131 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1132 if (used_hp[i].num_pages[j] > 0) {
1134 "Requesting %u pages of size %uMB"
1135 " from socket %i\n",
1136 used_hp[i].num_pages[j],
1138 (used_hp[i].hugepage_sz / 0x100000),
1144 /* create shared memory */
1145 hugepage = create_shared_memory(eal_hugepage_info_path(),
1146 nr_hugefiles * sizeof(struct hugepage_file));
1148 if (hugepage == NULL) {
1149 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1152 memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1155 * unmap pages that we won't need (looks at used_hp).
1156 * also, sets final_va to NULL on pages that were unmapped.
1158 if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1159 internal_config.num_hugepage_sizes) < 0) {
1160 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1165 * copy stuff from malloc'd hugepage* to the actual shared memory.
1166 * this procedure only copies those hugepages that have final_va
1167 * not NULL. has overflow protection.
1169 if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1170 tmp_hp, nr_hugefiles) < 0) {
1171 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1175 /* free the hugepage backing files */
1176 if (internal_config.hugepage_unlink &&
1177 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1178 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1182 /* free the temporary hugepage table */
1186 /* first memseg index shall be 0 after incrementing it below */
1188 for (i = 0; i < nr_hugefiles; i++) {
1191 /* if this is a new section, create a new memseg */
1194 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1196 else if (hugepage[i].size != hugepage[i-1].size)
1199 #ifdef RTE_ARCH_PPC_64
1200 /* On PPC64 architecture, the mmap always start from higher
1201 * virtual address to lower address. Here, both the physical
1202 * address and virtual address are in descending order */
1203 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1206 else if (((unsigned long)hugepage[i-1].final_va -
1207 (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1210 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1213 else if (((unsigned long)hugepage[i].final_va -
1214 (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1220 if (j == RTE_MAX_MEMSEG)
1223 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1224 mcfg->memseg[j].addr = hugepage[i].final_va;
1225 mcfg->memseg[j].len = hugepage[i].size;
1226 mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1227 mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1229 /* continuation of previous memseg */
1231 #ifdef RTE_ARCH_PPC_64
1232 /* Use the phy and virt address of the last page as segment
1233 * address for IBM Power architecture */
1234 mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1235 mcfg->memseg[j].addr = hugepage[i].final_va;
1237 mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1239 hugepage[i].memseg_id = j;
1242 if (i < nr_hugefiles) {
1243 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1244 "from %d requested\n"
1245 "Current %s=%d is not enough\n"
1246 "Please either increase it or request less amount "
1248 i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1253 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1258 huge_recover_sigbus();
1260 if (hugepage != NULL)
1261 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1267 * uses fstat to report the size of a file on disk
1273 if (fstat(fd, &st) < 0)
1279 * This creates the memory mappings in the secondary process to match that of
1280 * the server process. It goes through each memory segment in the DPDK runtime
1281 * configuration and finds the hugepages which form that segment, mapping them
1282 * in order to form a contiguous block in the virtual memory space
1285 rte_eal_hugepage_attach(void)
1287 const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1288 struct hugepage_file *hp = NULL;
1289 unsigned num_hp = 0;
1290 unsigned i, s = 0; /* s used to track the segment number */
1291 unsigned max_seg = RTE_MAX_MEMSEG;
1293 int fd, fd_zero = -1, fd_hugepage = -1;
1295 if (aslr_enabled() > 0) {
1296 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1297 "(ASLR) is enabled in the kernel.\n");
1298 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory "
1299 "into secondary processes\n");
1302 test_phys_addrs_available();
1304 if (internal_config.xen_dom0_support) {
1305 #ifdef RTE_LIBRTE_XEN_DOM0
1306 if (rte_xen_dom0_memory_attach() < 0) {
1307 RTE_LOG(ERR, EAL, "Failed to attach memory segments of primary "
1315 fd_zero = open("/dev/zero", O_RDONLY);
1317 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1320 fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1321 if (fd_hugepage < 0) {
1322 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1326 /* map all segments into memory to make sure we get the addrs */
1327 for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1331 * the first memory segment with len==0 is the one that
1332 * follows the last valid segment.
1334 if (mcfg->memseg[s].len == 0)
1338 * fdzero is mmapped to get a contiguous block of virtual
1339 * addresses of the appropriate memseg size.
1340 * use mmap to get identical addresses as the primary process.
1342 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1344 #ifdef RTE_ARCH_PPC_64
1345 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1350 if (base_addr == MAP_FAILED ||
1351 base_addr != mcfg->memseg[s].addr) {
1353 if (base_addr != MAP_FAILED) {
1354 /* errno is stale, don't use */
1355 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1356 "in /dev/zero at [%p], got [%p] - "
1357 "please use '--base-virtaddr' option\n",
1358 (unsigned long long)mcfg->memseg[s].len,
1359 mcfg->memseg[s].addr, base_addr);
1360 munmap(base_addr, mcfg->memseg[s].len);
1362 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1363 "in /dev/zero at [%p]: '%s'\n",
1364 (unsigned long long)mcfg->memseg[s].len,
1365 mcfg->memseg[s].addr, strerror(errno));
1367 if (aslr_enabled() > 0) {
1368 RTE_LOG(ERR, EAL, "It is recommended to "
1369 "disable ASLR in the kernel "
1370 "and retry running both primary "
1371 "and secondary processes\n");
1377 size = getFileSize(fd_hugepage);
1378 hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1379 if (hp == MAP_FAILED) {
1380 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1384 num_hp = size / sizeof(struct hugepage_file);
1385 RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1388 while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1389 void *addr, *base_addr;
1390 uintptr_t offset = 0;
1391 size_t mapping_size;
1393 * free previously mapped memory so we can map the
1394 * hugepages into the space
1396 base_addr = mcfg->memseg[s].addr;
1397 munmap(base_addr, mcfg->memseg[s].len);
1399 /* find the hugepages for this segment and map them
1400 * we don't need to worry about order, as the server sorted the
1401 * entries before it did the second mmap of them */
1402 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1403 if (hp[i].memseg_id == (int)s){
1404 fd = open(hp[i].filepath, O_RDWR);
1406 RTE_LOG(ERR, EAL, "Could not open %s\n",
1410 mapping_size = hp[i].size;
1411 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1412 mapping_size, PROT_READ | PROT_WRITE,
1414 close(fd); /* close file both on success and on failure */
1415 if (addr == MAP_FAILED ||
1416 addr != RTE_PTR_ADD(base_addr, offset)) {
1417 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1421 offset+=mapping_size;
1424 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1425 (unsigned long long)mcfg->memseg[s].len);
1428 /* unmap the hugepage config file, since we are done using it */
1435 for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1436 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1437 if (hp != NULL && hp != MAP_FAILED)
1441 if (fd_hugepage >= 0)
1447 rte_eal_using_phys_addrs(void)
1449 return phys_addrs_available;