1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2017-2018 Intel Corporation
14 #include <sys/types.h>
16 #include <sys/queue.h>
21 #include <sys/ioctl.h>
25 #ifdef F_ADD_SEALS /* if file sealing is supported, so is memfd */
26 #include <linux/memfd.h>
27 #define MEMFD_SUPPORTED
29 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
33 #include <linux/falloc.h>
34 #include <linux/mman.h> /* for hugetlb-related mmap flags */
36 #include <rte_common.h>
39 #include <rte_errno.h>
40 #include <rte_memory.h>
41 #include <rte_spinlock.h>
43 #include "eal_filesystem.h"
44 #include "eal_internal_cfg.h"
45 #include "eal_memalloc.h"
46 #include "eal_memcfg.h"
47 #include "eal_private.h"
49 const int anonymous_hugepages_supported =
52 #define RTE_MAP_HUGE_SHIFT MAP_HUGE_SHIFT
55 #define RTE_MAP_HUGE_SHIFT 26
59 * we've already checked memfd support at compile-time, but we also need to
60 * check if we can create hugepage files with memfd.
62 * also, this is not a constant, because while we may be *compiled* with memfd
63 * hugetlbfs support, we might not be *running* on a system that supports memfd
64 * and/or memfd with hugetlbfs, so we need to be able to adjust this flag at
65 * runtime, and fall back to anonymous memory.
67 static int memfd_create_supported =
70 #define RTE_MFD_HUGETLB MFD_HUGETLB
73 #define RTE_MFD_HUGETLB 4U
77 * not all kernel version support fallocate on hugetlbfs, so fall back to
78 * ftruncate and disallow deallocation if fallocate is not supported.
80 static int fallocate_supported = -1; /* unknown */
83 * we have two modes - single file segments, and file-per-page mode.
85 * for single-file segments, we use memseg_list_fd to store the segment fd,
86 * while the fds[] will not be allocated, and len will be set to 0.
88 * for file-per-page mode, each page will have its own fd, so 'memseg_list_fd'
89 * will be invalid (set to -1), and we'll use 'fds' to keep track of page fd's.
91 * we cannot know how many pages a system will have in advance, but we do know
92 * that they come in lists, and we know lengths of these lists. so, simply store
93 * a malloc'd array of fd's indexed by list and segment index.
95 * they will be initialized at startup, and filled as we allocate/deallocate
99 int *fds; /**< dynamically allocated array of segment lock fd's */
100 int memseg_list_fd; /**< memseg list fd */
101 int len; /**< total length of the array */
102 int count; /**< entries used in an array */
103 } fd_list[RTE_MAX_MEMSEG_LISTS];
105 /** local copy of a memory map, used to synchronize memory hotplug in MP */
106 static struct rte_memseg_list local_memsegs[RTE_MAX_MEMSEG_LISTS];
108 static sigjmp_buf huge_jmpenv;
110 static void __rte_unused huge_sigbus_handler(int signo __rte_unused)
112 siglongjmp(huge_jmpenv, 1);
115 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
116 * non-static local variable in the stack frame calling sigsetjmp might be
117 * clobbered by a call to longjmp.
119 static int __rte_unused huge_wrap_sigsetjmp(void)
121 return sigsetjmp(huge_jmpenv, 1);
124 static struct sigaction huge_action_old;
125 static int huge_need_recover;
127 static void __rte_unused
128 huge_register_sigbus(void)
131 struct sigaction action;
134 sigaddset(&mask, SIGBUS);
136 action.sa_mask = mask;
137 action.sa_handler = huge_sigbus_handler;
139 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
142 static void __rte_unused
143 huge_recover_sigbus(void)
145 if (huge_need_recover) {
146 sigaction(SIGBUS, &huge_action_old, NULL);
147 huge_need_recover = 0;
151 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
156 /* Check if kernel supports NUMA. */
157 if (numa_available() != 0) {
158 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
165 prepare_numa(int *oldpolicy, struct bitmask *oldmask, int socket_id)
167 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
168 if (get_mempolicy(oldpolicy, oldmask->maskp,
169 oldmask->size + 1, 0, 0) < 0) {
171 "Failed to get current mempolicy: %s. "
172 "Assuming MPOL_DEFAULT.\n", strerror(errno));
173 *oldpolicy = MPOL_DEFAULT;
176 "Setting policy MPOL_PREFERRED for socket %d\n",
178 numa_set_preferred(socket_id);
182 restore_numa(int *oldpolicy, struct bitmask *oldmask)
185 "Restoring previous memory policy: %d\n", *oldpolicy);
186 if (*oldpolicy == MPOL_DEFAULT) {
187 numa_set_localalloc();
188 } else if (set_mempolicy(*oldpolicy, oldmask->maskp,
189 oldmask->size + 1) < 0) {
190 RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
192 numa_set_localalloc();
194 numa_free_cpumask(oldmask);
199 * uses fstat to report the size of a file on disk
202 get_file_size(int fd)
205 if (fstat(fd, &st) < 0)
211 pagesz_flags(uint64_t page_sz)
213 /* as per mmap() manpage, all page sizes are log2 of page size
214 * shifted by MAP_HUGE_SHIFT
216 int log2 = rte_log2_u64(page_sz);
217 return log2 << RTE_MAP_HUGE_SHIFT;
220 /* returns 1 on successful lock, 0 on unsuccessful lock, -1 on error */
221 static int lock(int fd, int type)
225 /* flock may be interrupted */
227 ret = flock(fd, type | LOCK_NB);
228 } while (ret && errno == EINTR);
230 if (ret && errno == EWOULDBLOCK) {
234 RTE_LOG(ERR, EAL, "%s(): error calling flock(): %s\n",
235 __func__, strerror(errno));
238 /* lock was successful */
243 get_seg_memfd(struct hugepage_info *hi __rte_unused,
244 unsigned int list_idx __rte_unused,
245 unsigned int seg_idx __rte_unused)
247 #ifdef MEMFD_SUPPORTED
249 char segname[250]; /* as per manpage, limit is 249 bytes plus null */
251 int flags = RTE_MFD_HUGETLB | pagesz_flags(hi->hugepage_sz);
252 const struct internal_config *internal_conf =
253 eal_get_internal_configuration();
255 if (internal_conf->single_file_segments) {
256 fd = fd_list[list_idx].memseg_list_fd;
259 snprintf(segname, sizeof(segname), "seg_%i", list_idx);
260 fd = memfd_create(segname, flags);
262 RTE_LOG(DEBUG, EAL, "%s(): memfd create failed: %s\n",
263 __func__, strerror(errno));
266 fd_list[list_idx].memseg_list_fd = fd;
269 fd = fd_list[list_idx].fds[seg_idx];
272 snprintf(segname, sizeof(segname), "seg_%i-%i",
274 fd = memfd_create(segname, flags);
276 RTE_LOG(DEBUG, EAL, "%s(): memfd create failed: %s\n",
277 __func__, strerror(errno));
280 fd_list[list_idx].fds[seg_idx] = fd;
289 get_seg_fd(char *path, int buflen, struct hugepage_info *hi,
290 unsigned int list_idx, unsigned int seg_idx)
293 const struct internal_config *internal_conf =
294 eal_get_internal_configuration();
296 /* for in-memory mode, we only make it here when we're sure we support
297 * memfd, and this is a special case.
299 if (internal_conf->in_memory)
300 return get_seg_memfd(hi, list_idx, seg_idx);
302 if (internal_conf->single_file_segments) {
303 /* create a hugepage file path */
304 eal_get_hugefile_path(path, buflen, hi->hugedir, list_idx);
306 fd = fd_list[list_idx].memseg_list_fd;
309 fd = open(path, O_CREAT | O_RDWR, 0600);
311 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n",
312 __func__, strerror(errno));
315 /* take out a read lock and keep it indefinitely */
316 if (lock(fd, LOCK_SH) < 0) {
317 RTE_LOG(ERR, EAL, "%s(): lock failed: %s\n",
318 __func__, strerror(errno));
322 fd_list[list_idx].memseg_list_fd = fd;
325 /* create a hugepage file path */
326 eal_get_hugefile_path(path, buflen, hi->hugedir,
327 list_idx * RTE_MAX_MEMSEG_PER_LIST + seg_idx);
329 fd = fd_list[list_idx].fds[seg_idx];
332 fd = open(path, O_CREAT | O_RDWR, 0600);
334 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n",
335 __func__, strerror(errno));
338 /* take out a read lock */
339 if (lock(fd, LOCK_SH) < 0) {
340 RTE_LOG(ERR, EAL, "%s(): lock failed: %s\n",
341 __func__, strerror(errno));
345 fd_list[list_idx].fds[seg_idx] = fd;
352 resize_hugefile_in_memory(int fd, uint64_t fa_offset,
353 uint64_t page_sz, bool grow)
355 int flags = grow ? 0 : FALLOC_FL_PUNCH_HOLE |
359 /* grow or shrink the file */
360 ret = fallocate(fd, flags, fa_offset, page_sz);
363 RTE_LOG(DEBUG, EAL, "%s(): fallocate() failed: %s\n",
372 resize_hugefile_in_filesystem(int fd, uint64_t fa_offset, uint64_t page_sz,
378 if (fallocate_supported == 0) {
379 /* we cannot deallocate memory if fallocate() is not
380 * supported, and hugepage file is already locked at
381 * creation, so no further synchronization needed.
385 RTE_LOG(DEBUG, EAL, "%s(): fallocate not supported, not freeing page back to the system\n",
389 uint64_t new_size = fa_offset + page_sz;
390 uint64_t cur_size = get_file_size(fd);
392 /* fallocate isn't supported, fall back to ftruncate */
393 if (new_size > cur_size &&
394 ftruncate(fd, new_size) < 0) {
395 RTE_LOG(DEBUG, EAL, "%s(): ftruncate() failed: %s\n",
396 __func__, strerror(errno));
400 int flags = grow ? 0 : FALLOC_FL_PUNCH_HOLE |
405 * technically, it is perfectly safe for both primary
406 * and secondary to grow and shrink the page files:
407 * growing the file repeatedly has no effect because
408 * a page can only be allocated once, while mmap ensures
409 * that secondaries hold on to the page even after the
410 * page itself is removed from the filesystem.
412 * however, leaving growing/shrinking to the primary
413 * tends to expose bugs in fdlist page count handling,
414 * so leave this here just in case.
416 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
419 /* grow or shrink the file */
420 ret = fallocate(fd, flags, fa_offset, page_sz);
423 if (fallocate_supported == -1 &&
425 RTE_LOG(ERR, EAL, "%s(): fallocate() not supported, hugepage deallocation will be disabled\n",
428 fallocate_supported = 0;
430 RTE_LOG(DEBUG, EAL, "%s(): fallocate() failed: %s\n",
436 fallocate_supported = 1;
444 close_hugefile(int fd, char *path, int list_idx)
446 const struct internal_config *internal_conf =
447 eal_get_internal_configuration();
449 * primary process must unlink the file, but only when not in in-memory
450 * mode (as in that case there is no file to unlink).
452 if (!internal_conf->in_memory &&
453 rte_eal_process_type() == RTE_PROC_PRIMARY &&
455 RTE_LOG(ERR, EAL, "%s(): unlinking '%s' failed: %s\n",
456 __func__, path, strerror(errno));
459 fd_list[list_idx].memseg_list_fd = -1;
463 resize_hugefile(int fd, uint64_t fa_offset, uint64_t page_sz, bool grow)
465 /* in-memory mode is a special case, because we can be sure that
466 * fallocate() is supported.
468 const struct internal_config *internal_conf =
469 eal_get_internal_configuration();
471 if (internal_conf->in_memory)
472 return resize_hugefile_in_memory(fd, fa_offset,
475 return resize_hugefile_in_filesystem(fd, fa_offset, page_sz,
480 alloc_seg(struct rte_memseg *ms, void *addr, int socket_id,
481 struct hugepage_info *hi, unsigned int list_idx,
482 unsigned int seg_idx)
484 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
485 int cur_socket_id = 0;
496 const struct internal_config *internal_conf =
497 eal_get_internal_configuration();
499 alloc_sz = hi->hugepage_sz;
501 /* these are checked at init, but code analyzers don't know that */
502 if (internal_conf->in_memory && !anonymous_hugepages_supported) {
503 RTE_LOG(ERR, EAL, "Anonymous hugepages not supported, in-memory mode cannot allocate memory\n");
506 if (internal_conf->in_memory && !memfd_create_supported &&
507 internal_conf->single_file_segments) {
508 RTE_LOG(ERR, EAL, "Single-file segments are not supported without memfd support\n");
512 /* in-memory without memfd is a special case */
515 if (internal_conf->in_memory && !memfd_create_supported) {
516 const int in_memory_flags = MAP_HUGETLB | MAP_FIXED |
517 MAP_PRIVATE | MAP_ANONYMOUS;
520 pagesz_flag = pagesz_flags(alloc_sz);
522 mmap_flags = in_memory_flags | pagesz_flag;
524 /* single-file segments codepath will never be active
525 * here because in-memory mode is incompatible with the
526 * fallback path, and it's stopped at EAL initialization
531 /* takes out a read lock on segment or segment list */
532 fd = get_seg_fd(path, sizeof(path), hi, list_idx, seg_idx);
534 RTE_LOG(ERR, EAL, "Couldn't get fd on hugepage file\n");
538 if (internal_conf->single_file_segments) {
539 map_offset = seg_idx * alloc_sz;
540 ret = resize_hugefile(fd, map_offset, alloc_sz, true);
544 fd_list[list_idx].count++;
547 if (ftruncate(fd, alloc_sz) < 0) {
548 RTE_LOG(DEBUG, EAL, "%s(): ftruncate() failed: %s\n",
549 __func__, strerror(errno));
552 if (internal_conf->hugepage_unlink &&
553 !internal_conf->in_memory) {
555 RTE_LOG(DEBUG, EAL, "%s(): unlink() failed: %s\n",
556 __func__, strerror(errno));
561 mmap_flags = MAP_SHARED | MAP_POPULATE | MAP_FIXED;
565 * map the segment, and populate page tables, the kernel fills
566 * this segment with zeros if it's a new page.
568 va = mmap(addr, alloc_sz, PROT_READ | PROT_WRITE, mmap_flags, fd,
571 if (va == MAP_FAILED) {
572 RTE_LOG(DEBUG, EAL, "%s(): mmap() failed: %s\n", __func__,
574 /* mmap failed, but the previous region might have been
575 * unmapped anyway. try to remap it
580 RTE_LOG(DEBUG, EAL, "%s(): wrong mmap() address\n", __func__);
581 munmap(va, alloc_sz);
585 /* In linux, hugetlb limitations, like cgroup, are
586 * enforced at fault time instead of mmap(), even
587 * with the option of MAP_POPULATE. Kernel will send
588 * a SIGBUS signal. To avoid to be killed, save stack
589 * environment here, if SIGBUS happens, we can jump
592 if (huge_wrap_sigsetjmp()) {
593 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more hugepages of size %uMB\n",
594 (unsigned int)(alloc_sz >> 20));
598 /* we need to trigger a write to the page to enforce page fault and
599 * ensure that page is accessible to us, but we can't overwrite value
600 * that is already there, so read the old value, and write itback.
601 * kernel populates the page with zeroes initially.
603 *(volatile int *)addr = *(volatile int *)addr;
605 iova = rte_mem_virt2iova(addr);
606 if (iova == RTE_BAD_PHYS_ADDR) {
607 RTE_LOG(DEBUG, EAL, "%s(): can't get IOVA addr\n",
612 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
613 ret = get_mempolicy(&cur_socket_id, NULL, 0, addr,
614 MPOL_F_NODE | MPOL_F_ADDR);
616 RTE_LOG(DEBUG, EAL, "%s(): get_mempolicy: %s\n",
617 __func__, strerror(errno));
619 } else if (cur_socket_id != socket_id) {
621 "%s(): allocation happened on wrong socket (wanted %d, got %d)\n",
622 __func__, socket_id, cur_socket_id);
626 if (rte_socket_count() > 1)
627 RTE_LOG(DEBUG, EAL, "%s(): not checking hugepage NUMA node.\n",
632 ms->hugepage_sz = alloc_sz;
634 ms->nchannel = rte_memory_get_nchannel();
635 ms->nrank = rte_memory_get_nrank();
637 ms->socket_id = socket_id;
642 munmap(addr, alloc_sz);
644 flags = EAL_RESERVE_FORCE_ADDRESS;
645 new_addr = eal_get_virtual_area(addr, &alloc_sz, alloc_sz, 0, flags);
646 if (new_addr != addr) {
647 if (new_addr != NULL)
648 munmap(new_addr, alloc_sz);
649 /* we're leaving a hole in our virtual address space. if
650 * somebody else maps this hole now, we could accidentally
651 * override it in the future.
653 RTE_LOG(CRIT, EAL, "Can't mmap holes in our virtual address space\n");
655 /* roll back the ref count */
656 if (internal_conf->single_file_segments)
657 fd_list[list_idx].count--;
659 /* some codepaths will return negative fd, so exit early */
663 if (internal_conf->single_file_segments) {
664 resize_hugefile(fd, map_offset, alloc_sz, false);
665 /* ignore failure, can't make it any worse */
667 /* if refcount is at zero, close the file */
668 if (fd_list[list_idx].count == 0)
669 close_hugefile(fd, path, list_idx);
671 /* only remove file if we can take out a write lock */
672 if (internal_conf->hugepage_unlink == 0 &&
673 internal_conf->in_memory == 0 &&
674 lock(fd, LOCK_EX) == 1)
677 fd_list[list_idx].fds[seg_idx] = -1;
683 free_seg(struct rte_memseg *ms, struct hugepage_info *hi,
684 unsigned int list_idx, unsigned int seg_idx)
690 const struct internal_config *internal_conf =
691 eal_get_internal_configuration();
693 /* erase page data */
694 memset(ms->addr, 0, ms->len);
696 if (mmap(ms->addr, ms->len, PROT_NONE,
697 MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, 0) ==
699 RTE_LOG(DEBUG, EAL, "couldn't unmap page\n");
703 eal_mem_set_dump(ms->addr, ms->len, false);
707 /* if we're using anonymous hugepages, nothing to be done */
708 if (internal_conf->in_memory && !memfd_create_supported)
711 /* if we've already unlinked the page, nothing needs to be done */
712 if (!internal_conf->in_memory && internal_conf->hugepage_unlink)
716 memset(ms, 0, sizeof(*ms));
720 /* if we are not in single file segments mode, we're going to unmap the
721 * segment and thus drop the lock on original fd, but hugepage dir is
722 * now locked so we can take out another one without races.
724 fd = get_seg_fd(path, sizeof(path), hi, list_idx, seg_idx);
728 if (internal_conf->single_file_segments) {
729 map_offset = seg_idx * ms->len;
730 if (resize_hugefile(fd, map_offset, ms->len, false))
733 if (--(fd_list[list_idx].count) == 0)
734 close_hugefile(fd, path, list_idx);
738 /* if we're able to take out a write lock, we're the last one
739 * holding onto this page.
741 if (!internal_conf->in_memory) {
742 ret = lock(fd, LOCK_EX);
744 /* no one else is using this page */
749 /* closing fd will drop the lock */
751 fd_list[list_idx].fds[seg_idx] = -1;
754 memset(ms, 0, sizeof(*ms));
756 return ret < 0 ? -1 : 0;
759 struct alloc_walk_param {
760 struct hugepage_info *hi;
761 struct rte_memseg **ms;
763 unsigned int segs_allocated;
769 alloc_seg_walk(const struct rte_memseg_list *msl, void *arg)
771 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
772 struct alloc_walk_param *wa = arg;
773 struct rte_memseg_list *cur_msl;
775 int cur_idx, start_idx, j, dir_fd = -1;
776 unsigned int msl_idx, need, i;
777 const struct internal_config *internal_conf =
778 eal_get_internal_configuration();
780 if (msl->page_sz != wa->page_sz)
782 if (msl->socket_id != wa->socket)
785 page_sz = (size_t)msl->page_sz;
787 msl_idx = msl - mcfg->memsegs;
788 cur_msl = &mcfg->memsegs[msl_idx];
792 /* try finding space in memseg list */
794 /* if we require exact number of pages in a list, find them */
795 cur_idx = rte_fbarray_find_next_n_free(&cur_msl->memseg_arr, 0,
803 /* we don't require exact number of pages, so we're going to go
804 * for best-effort allocation. that means finding the biggest
805 * unused block, and going with that.
807 cur_idx = rte_fbarray_find_biggest_free(&cur_msl->memseg_arr,
812 /* adjust the size to possibly be smaller than original
813 * request, but do not allow it to be bigger.
815 cur_len = rte_fbarray_find_contig_free(&cur_msl->memseg_arr,
817 need = RTE_MIN(need, (unsigned int)cur_len);
820 /* do not allow any page allocations during the time we're allocating,
821 * because file creation and locking operations are not atomic,
822 * and we might be the first or the last ones to use a particular page,
823 * so we need to ensure atomicity of every operation.
825 * during init, we already hold a write lock, so don't try to take out
828 if (wa->hi->lock_descriptor == -1 && !internal_conf->in_memory) {
829 dir_fd = open(wa->hi->hugedir, O_RDONLY);
831 RTE_LOG(ERR, EAL, "%s(): Cannot open '%s': %s\n",
832 __func__, wa->hi->hugedir, strerror(errno));
835 /* blocking writelock */
836 if (flock(dir_fd, LOCK_EX)) {
837 RTE_LOG(ERR, EAL, "%s(): Cannot lock '%s': %s\n",
838 __func__, wa->hi->hugedir, strerror(errno));
844 for (i = 0; i < need; i++, cur_idx++) {
845 struct rte_memseg *cur;
848 cur = rte_fbarray_get(&cur_msl->memseg_arr, cur_idx);
849 map_addr = RTE_PTR_ADD(cur_msl->base_va,
852 if (alloc_seg(cur, map_addr, wa->socket, wa->hi,
854 RTE_LOG(DEBUG, EAL, "attempted to allocate %i segments, but only %i were allocated\n",
857 /* if exact number wasn't requested, stop */
862 for (j = start_idx; j < cur_idx; j++) {
863 struct rte_memseg *tmp;
864 struct rte_fbarray *arr =
865 &cur_msl->memseg_arr;
867 tmp = rte_fbarray_get(arr, j);
868 rte_fbarray_set_free(arr, j);
870 /* free_seg may attempt to create a file, which
873 if (free_seg(tmp, wa->hi, msl_idx, j))
874 RTE_LOG(DEBUG, EAL, "Cannot free page\n");
878 memset(wa->ms, 0, sizeof(*wa->ms) * wa->n_segs);
887 rte_fbarray_set_used(&cur_msl->memseg_arr, cur_idx);
890 wa->segs_allocated = i;
895 /* if we didn't allocate any segments, move on to the next list */
899 struct free_walk_param {
900 struct hugepage_info *hi;
901 struct rte_memseg *ms;
904 free_seg_walk(const struct rte_memseg_list *msl, void *arg)
906 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
907 struct rte_memseg_list *found_msl;
908 struct free_walk_param *wa = arg;
909 uintptr_t start_addr, end_addr;
910 int msl_idx, seg_idx, ret, dir_fd = -1;
911 const struct internal_config *internal_conf =
912 eal_get_internal_configuration();
914 start_addr = (uintptr_t) msl->base_va;
915 end_addr = start_addr + msl->len;
917 if ((uintptr_t)wa->ms->addr < start_addr ||
918 (uintptr_t)wa->ms->addr >= end_addr)
921 msl_idx = msl - mcfg->memsegs;
922 seg_idx = RTE_PTR_DIFF(wa->ms->addr, start_addr) / msl->page_sz;
925 found_msl = &mcfg->memsegs[msl_idx];
927 /* do not allow any page allocations during the time we're freeing,
928 * because file creation and locking operations are not atomic,
929 * and we might be the first or the last ones to use a particular page,
930 * so we need to ensure atomicity of every operation.
932 * during init, we already hold a write lock, so don't try to take out
935 if (wa->hi->lock_descriptor == -1 && !internal_conf->in_memory) {
936 dir_fd = open(wa->hi->hugedir, O_RDONLY);
938 RTE_LOG(ERR, EAL, "%s(): Cannot open '%s': %s\n",
939 __func__, wa->hi->hugedir, strerror(errno));
942 /* blocking writelock */
943 if (flock(dir_fd, LOCK_EX)) {
944 RTE_LOG(ERR, EAL, "%s(): Cannot lock '%s': %s\n",
945 __func__, wa->hi->hugedir, strerror(errno));
951 found_msl->version++;
953 rte_fbarray_set_free(&found_msl->memseg_arr, seg_idx);
955 ret = free_seg(wa->ms, wa->hi, msl_idx, seg_idx);
967 eal_memalloc_alloc_seg_bulk(struct rte_memseg **ms, int n_segs, size_t page_sz,
968 int socket, bool exact)
971 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
972 bool have_numa = false;
974 struct bitmask *oldmask;
976 struct alloc_walk_param wa;
977 struct hugepage_info *hi = NULL;
978 struct internal_config *internal_conf =
979 eal_get_internal_configuration();
981 memset(&wa, 0, sizeof(wa));
983 /* dynamic allocation not supported in legacy mode */
984 if (internal_conf->legacy_mem)
987 for (i = 0; i < (int) RTE_DIM(internal_conf->hugepage_info); i++) {
989 internal_conf->hugepage_info[i].hugepage_sz) {
990 hi = &internal_conf->hugepage_info[i];
995 RTE_LOG(ERR, EAL, "%s(): can't find relevant hugepage_info entry\n",
1000 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
1002 oldmask = numa_allocate_nodemask();
1003 prepare_numa(&oldpolicy, oldmask, socket);
1012 wa.page_sz = page_sz;
1014 wa.segs_allocated = 0;
1016 /* memalloc is locked, so it's safe to use thread-unsafe version */
1017 ret = rte_memseg_list_walk_thread_unsafe(alloc_seg_walk, &wa);
1019 RTE_LOG(ERR, EAL, "%s(): couldn't find suitable memseg_list\n",
1022 } else if (ret > 0) {
1023 ret = (int)wa.segs_allocated;
1026 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
1028 restore_numa(&oldpolicy, oldmask);
1034 eal_memalloc_alloc_seg(size_t page_sz, int socket)
1036 struct rte_memseg *ms;
1037 if (eal_memalloc_alloc_seg_bulk(&ms, 1, page_sz, socket, true) < 0)
1039 /* return pointer to newly allocated memseg */
1044 eal_memalloc_free_seg_bulk(struct rte_memseg **ms, int n_segs)
1047 struct internal_config *internal_conf =
1048 eal_get_internal_configuration();
1050 /* dynamic free not supported in legacy mode */
1051 if (internal_conf->legacy_mem)
1054 for (seg = 0; seg < n_segs; seg++) {
1055 struct rte_memseg *cur = ms[seg];
1056 struct hugepage_info *hi = NULL;
1057 struct free_walk_param wa;
1060 /* if this page is marked as unfreeable, fail */
1061 if (cur->flags & RTE_MEMSEG_FLAG_DO_NOT_FREE) {
1062 RTE_LOG(DEBUG, EAL, "Page is not allowed to be freed\n");
1067 memset(&wa, 0, sizeof(wa));
1069 for (i = 0; i < (int)RTE_DIM(internal_conf->hugepage_info);
1071 hi = &internal_conf->hugepage_info[i];
1072 if (cur->hugepage_sz == hi->hugepage_sz)
1075 if (i == (int)RTE_DIM(internal_conf->hugepage_info)) {
1076 RTE_LOG(ERR, EAL, "Can't find relevant hugepage_info entry\n");
1084 /* memalloc is locked, so it's safe to use thread-unsafe version
1086 walk_res = rte_memseg_list_walk_thread_unsafe(free_seg_walk,
1091 RTE_LOG(ERR, EAL, "Couldn't find memseg list\n");
1098 eal_memalloc_free_seg(struct rte_memseg *ms)
1100 const struct internal_config *internal_conf =
1101 eal_get_internal_configuration();
1103 /* dynamic free not supported in legacy mode */
1104 if (internal_conf->legacy_mem)
1107 return eal_memalloc_free_seg_bulk(&ms, 1);
1111 sync_chunk(struct rte_memseg_list *primary_msl,
1112 struct rte_memseg_list *local_msl, struct hugepage_info *hi,
1113 unsigned int msl_idx, bool used, int start, int end)
1115 struct rte_fbarray *l_arr, *p_arr;
1116 int i, ret, chunk_len, diff_len;
1118 l_arr = &local_msl->memseg_arr;
1119 p_arr = &primary_msl->memseg_arr;
1121 /* we need to aggregate allocations/deallocations into bigger chunks,
1122 * as we don't want to spam the user with per-page callbacks.
1124 * to avoid any potential issues, we also want to trigger
1125 * deallocation callbacks *before* we actually deallocate
1126 * memory, so that the user application could wrap up its use
1127 * before it goes away.
1130 chunk_len = end - start;
1132 /* find how many contiguous pages we can map/unmap for this chunk */
1134 rte_fbarray_find_contig_free(l_arr, start) :
1135 rte_fbarray_find_contig_used(l_arr, start);
1137 /* has to be at least one page */
1141 diff_len = RTE_MIN(chunk_len, diff_len);
1143 /* if we are freeing memory, notify the application */
1145 struct rte_memseg *ms;
1147 size_t len, page_sz;
1149 ms = rte_fbarray_get(l_arr, start);
1150 start_va = ms->addr;
1151 page_sz = (size_t)primary_msl->page_sz;
1152 len = page_sz * diff_len;
1154 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
1158 for (i = 0; i < diff_len; i++) {
1159 struct rte_memseg *p_ms, *l_ms;
1160 int seg_idx = start + i;
1162 l_ms = rte_fbarray_get(l_arr, seg_idx);
1163 p_ms = rte_fbarray_get(p_arr, seg_idx);
1165 if (l_ms == NULL || p_ms == NULL)
1169 ret = alloc_seg(l_ms, p_ms->addr,
1170 p_ms->socket_id, hi,
1174 rte_fbarray_set_used(l_arr, seg_idx);
1176 ret = free_seg(l_ms, hi, msl_idx, seg_idx);
1177 rte_fbarray_set_free(l_arr, seg_idx);
1183 /* if we just allocated memory, notify the application */
1185 struct rte_memseg *ms;
1187 size_t len, page_sz;
1189 ms = rte_fbarray_get(l_arr, start);
1190 start_va = ms->addr;
1191 page_sz = (size_t)primary_msl->page_sz;
1192 len = page_sz * diff_len;
1194 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC,
1198 /* calculate how much we can advance until next chunk */
1200 rte_fbarray_find_contig_used(l_arr, start) :
1201 rte_fbarray_find_contig_free(l_arr, start);
1202 ret = RTE_MIN(chunk_len, diff_len);
1208 sync_status(struct rte_memseg_list *primary_msl,
1209 struct rte_memseg_list *local_msl, struct hugepage_info *hi,
1210 unsigned int msl_idx, bool used)
1212 struct rte_fbarray *l_arr, *p_arr;
1213 int p_idx, l_chunk_len, p_chunk_len, ret;
1216 /* this is a little bit tricky, but the basic idea is - walk both lists
1217 * and spot any places where there are discrepancies. walking both lists
1218 * and noting discrepancies in a single go is a hard problem, so we do
1219 * it in two passes - first we spot any places where allocated segments
1220 * mismatch (i.e. ensure that everything that's allocated in the primary
1221 * is also allocated in the secondary), and then we do it by looking at
1222 * free segments instead.
1224 * we also need to aggregate changes into chunks, as we have to call
1225 * callbacks per allocation, not per page.
1227 l_arr = &local_msl->memseg_arr;
1228 p_arr = &primary_msl->memseg_arr;
1231 p_idx = rte_fbarray_find_next_used(p_arr, 0);
1233 p_idx = rte_fbarray_find_next_free(p_arr, 0);
1235 while (p_idx >= 0) {
1236 int next_chunk_search_idx;
1239 p_chunk_len = rte_fbarray_find_contig_used(p_arr,
1241 l_chunk_len = rte_fbarray_find_contig_used(l_arr,
1244 p_chunk_len = rte_fbarray_find_contig_free(p_arr,
1246 l_chunk_len = rte_fbarray_find_contig_free(l_arr,
1249 /* best case scenario - no differences (or bigger, which will be
1250 * fixed during next iteration), look for next chunk
1252 if (l_chunk_len >= p_chunk_len) {
1253 next_chunk_search_idx = p_idx + p_chunk_len;
1257 /* if both chunks start at the same point, skip parts we know
1258 * are identical, and sync the rest. each call to sync_chunk
1259 * will only sync contiguous segments, so we need to call this
1260 * until we are sure there are no more differences in this
1263 start = p_idx + l_chunk_len;
1264 end = p_idx + p_chunk_len;
1266 ret = sync_chunk(primary_msl, local_msl, hi, msl_idx,
1269 } while (start < end && ret >= 0);
1270 /* if ret is negative, something went wrong */
1274 next_chunk_search_idx = p_idx + p_chunk_len;
1276 /* skip to end of this chunk */
1278 p_idx = rte_fbarray_find_next_used(p_arr,
1279 next_chunk_search_idx);
1281 p_idx = rte_fbarray_find_next_free(p_arr,
1282 next_chunk_search_idx);
1289 sync_existing(struct rte_memseg_list *primary_msl,
1290 struct rte_memseg_list *local_msl, struct hugepage_info *hi,
1291 unsigned int msl_idx)
1295 /* do not allow any page allocations during the time we're allocating,
1296 * because file creation and locking operations are not atomic,
1297 * and we might be the first or the last ones to use a particular page,
1298 * so we need to ensure atomicity of every operation.
1300 dir_fd = open(hi->hugedir, O_RDONLY);
1302 RTE_LOG(ERR, EAL, "%s(): Cannot open '%s': %s\n", __func__,
1303 hi->hugedir, strerror(errno));
1306 /* blocking writelock */
1307 if (flock(dir_fd, LOCK_EX)) {
1308 RTE_LOG(ERR, EAL, "%s(): Cannot lock '%s': %s\n", __func__,
1309 hi->hugedir, strerror(errno));
1314 /* ensure all allocated space is the same in both lists */
1315 ret = sync_status(primary_msl, local_msl, hi, msl_idx, true);
1319 /* ensure all unallocated space is the same in both lists */
1320 ret = sync_status(primary_msl, local_msl, hi, msl_idx, false);
1324 /* update version number */
1325 local_msl->version = primary_msl->version;
1336 sync_walk(const struct rte_memseg_list *msl, void *arg __rte_unused)
1338 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1339 struct rte_memseg_list *primary_msl, *local_msl;
1340 struct hugepage_info *hi = NULL;
1343 struct internal_config *internal_conf =
1344 eal_get_internal_configuration();
1349 msl_idx = msl - mcfg->memsegs;
1350 primary_msl = &mcfg->memsegs[msl_idx];
1351 local_msl = &local_memsegs[msl_idx];
1353 for (i = 0; i < RTE_DIM(internal_conf->hugepage_info); i++) {
1355 internal_conf->hugepage_info[i].hugepage_sz;
1356 uint64_t msl_sz = primary_msl->page_sz;
1357 if (msl_sz == cur_sz) {
1358 hi = &internal_conf->hugepage_info[i];
1363 RTE_LOG(ERR, EAL, "Can't find relevant hugepage_info entry\n");
1367 /* if versions don't match, synchronize everything */
1368 if (local_msl->version != primary_msl->version &&
1369 sync_existing(primary_msl, local_msl, hi, msl_idx))
1376 eal_memalloc_sync_with_primary(void)
1378 /* nothing to be done in primary */
1379 if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1382 /* memalloc is locked, so it's safe to call thread-unsafe version */
1383 if (rte_memseg_list_walk_thread_unsafe(sync_walk, NULL))
1389 secondary_msl_create_walk(const struct rte_memseg_list *msl,
1390 void *arg __rte_unused)
1392 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1393 struct rte_memseg_list *primary_msl, *local_msl;
1394 char name[PATH_MAX];
1400 msl_idx = msl - mcfg->memsegs;
1401 primary_msl = &mcfg->memsegs[msl_idx];
1402 local_msl = &local_memsegs[msl_idx];
1404 /* create distinct fbarrays for each secondary */
1405 snprintf(name, RTE_FBARRAY_NAME_LEN, "%s_%i",
1406 primary_msl->memseg_arr.name, getpid());
1408 ret = rte_fbarray_init(&local_msl->memseg_arr, name,
1409 primary_msl->memseg_arr.len,
1410 primary_msl->memseg_arr.elt_sz);
1412 RTE_LOG(ERR, EAL, "Cannot initialize local memory map\n");
1415 local_msl->base_va = primary_msl->base_va;
1416 local_msl->len = primary_msl->len;
1422 alloc_list(int list_idx, int len)
1426 const struct internal_config *internal_conf =
1427 eal_get_internal_configuration();
1429 /* single-file segments mode does not need fd list */
1430 if (!internal_conf->single_file_segments) {
1431 /* ensure we have space to store fd per each possible segment */
1432 data = malloc(sizeof(int) * len);
1434 RTE_LOG(ERR, EAL, "Unable to allocate space for file descriptors\n");
1437 /* set all fd's as invalid */
1438 for (i = 0; i < len; i++)
1440 fd_list[list_idx].fds = data;
1441 fd_list[list_idx].len = len;
1443 fd_list[list_idx].fds = NULL;
1444 fd_list[list_idx].len = 0;
1447 fd_list[list_idx].count = 0;
1448 fd_list[list_idx].memseg_list_fd = -1;
1454 fd_list_create_walk(const struct rte_memseg_list *msl,
1455 void *arg __rte_unused)
1457 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1464 msl_idx = msl - mcfg->memsegs;
1465 len = msl->memseg_arr.len;
1467 return alloc_list(msl_idx, len);
1471 eal_memalloc_set_seg_fd(int list_idx, int seg_idx, int fd)
1473 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1474 const struct internal_config *internal_conf =
1475 eal_get_internal_configuration();
1477 /* single file segments mode doesn't support individual segment fd's */
1478 if (internal_conf->single_file_segments)
1481 /* if list is not allocated, allocate it */
1482 if (fd_list[list_idx].len == 0) {
1483 int len = mcfg->memsegs[list_idx].memseg_arr.len;
1485 if (alloc_list(list_idx, len) < 0)
1488 fd_list[list_idx].fds[seg_idx] = fd;
1494 eal_memalloc_set_seg_list_fd(int list_idx, int fd)
1496 const struct internal_config *internal_conf =
1497 eal_get_internal_configuration();
1499 /* non-single file segment mode doesn't support segment list fd's */
1500 if (!internal_conf->single_file_segments)
1503 fd_list[list_idx].memseg_list_fd = fd;
1509 eal_memalloc_get_seg_fd(int list_idx, int seg_idx)
1512 const struct internal_config *internal_conf =
1513 eal_get_internal_configuration();
1515 if (internal_conf->in_memory || internal_conf->no_hugetlbfs) {
1516 #ifndef MEMFD_SUPPORTED
1517 /* in in-memory or no-huge mode, we rely on memfd support */
1520 /* memfd supported, but hugetlbfs memfd may not be */
1521 if (!internal_conf->no_hugetlbfs && !memfd_create_supported)
1525 if (internal_conf->single_file_segments) {
1526 fd = fd_list[list_idx].memseg_list_fd;
1527 } else if (fd_list[list_idx].len == 0) {
1528 /* list not initialized */
1531 fd = fd_list[list_idx].fds[seg_idx];
1539 test_memfd_create(void)
1541 #ifdef MEMFD_SUPPORTED
1542 const struct internal_config *internal_conf =
1543 eal_get_internal_configuration();
1545 for (i = 0; i < internal_conf->num_hugepage_sizes; i++) {
1546 uint64_t pagesz = internal_conf->hugepage_info[i].hugepage_sz;
1547 int pagesz_flag = pagesz_flags(pagesz);
1550 flags = pagesz_flag | RTE_MFD_HUGETLB;
1551 int fd = memfd_create("test", flags);
1553 /* we failed - let memalloc know this isn't working */
1554 if (errno == EINVAL) {
1555 memfd_create_supported = 0;
1556 return 0; /* not supported */
1559 /* we got other error - something's wrong */
1560 return -1; /* error */
1563 return 1; /* supported */
1566 return 0; /* not supported */
1570 eal_memalloc_get_seg_fd_offset(int list_idx, int seg_idx, size_t *offset)
1572 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1573 const struct internal_config *internal_conf =
1574 eal_get_internal_configuration();
1576 if (internal_conf->in_memory || internal_conf->no_hugetlbfs) {
1577 #ifndef MEMFD_SUPPORTED
1578 /* in in-memory or no-huge mode, we rely on memfd support */
1581 /* memfd supported, but hugetlbfs memfd may not be */
1582 if (!internal_conf->no_hugetlbfs && !memfd_create_supported)
1586 if (internal_conf->single_file_segments) {
1587 size_t pgsz = mcfg->memsegs[list_idx].page_sz;
1589 /* segment not active? */
1590 if (fd_list[list_idx].memseg_list_fd < 0)
1592 *offset = pgsz * seg_idx;
1594 /* fd_list not initialized? */
1595 if (fd_list[list_idx].len == 0)
1598 /* segment not active? */
1599 if (fd_list[list_idx].fds[seg_idx] < 0)
1607 eal_memalloc_init(void)
1609 const struct internal_config *internal_conf =
1610 eal_get_internal_configuration();
1612 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1613 if (rte_memseg_list_walk(secondary_msl_create_walk, NULL) < 0)
1615 if (rte_eal_process_type() == RTE_PROC_PRIMARY &&
1616 internal_conf->in_memory) {
1617 int mfd_res = test_memfd_create();
1620 RTE_LOG(ERR, EAL, "Unable to check if memfd is supported\n");
1624 RTE_LOG(DEBUG, EAL, "Using memfd for anonymous memory\n");
1626 RTE_LOG(INFO, EAL, "Using memfd is not supported, falling back to anonymous hugepages\n");
1628 /* we only support single-file segments mode with in-memory mode
1629 * if we support hugetlbfs with memfd_create. this code will
1632 if (internal_conf->single_file_segments &&
1634 RTE_LOG(ERR, EAL, "Single-file segments mode cannot be used without memfd support\n");
1637 /* this cannot ever happen but better safe than sorry */
1638 if (!anonymous_hugepages_supported) {
1639 RTE_LOG(ERR, EAL, "Using anonymous memory is not supported\n");
1644 /* initialize all of the fd lists */
1645 if (rte_memseg_list_walk(fd_list_create_walk, NULL))