1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2017-2018 Intel Corporation
5 #define _FILE_OFFSET_BITS 64
15 #include <sys/types.h>
17 #include <sys/queue.h>
22 #include <sys/ioctl.h>
26 #ifdef F_ADD_SEALS /* if file sealing is supported, so is memfd */
27 #include <linux/memfd.h>
28 #define MEMFD_SUPPORTED
30 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
34 #include <linux/falloc.h>
35 #include <linux/mman.h> /* for hugetlb-related mmap flags */
37 #include <rte_common.h>
39 #include <rte_eal_memconfig.h>
41 #include <rte_errno.h>
42 #include <rte_memory.h>
43 #include <rte_spinlock.h>
45 #include "eal_filesystem.h"
46 #include "eal_internal_cfg.h"
47 #include "eal_memalloc.h"
48 #include "eal_private.h"
50 const int anonymous_hugepages_supported =
53 #define RTE_MAP_HUGE_SHIFT MAP_HUGE_SHIFT
56 #define RTE_MAP_HUGE_SHIFT 26
60 * we've already checked memfd support at compile-time, but we also need to
61 * check if we can create hugepage files with memfd.
63 * also, this is not a constant, because while we may be *compiled* with memfd
64 * hugetlbfs support, we might not be *running* on a system that supports memfd
65 * and/or memfd with hugetlbfs, so we need to be able to adjust this flag at
66 * runtime, and fall back to anonymous memory.
68 static int memfd_create_supported =
71 #define RTE_MFD_HUGETLB MFD_HUGETLB
74 #define RTE_MFD_HUGETLB 4U
78 * not all kernel version support fallocate on hugetlbfs, so fall back to
79 * ftruncate and disallow deallocation if fallocate is not supported.
81 static int fallocate_supported = -1; /* unknown */
84 * we have two modes - single file segments, and file-per-page mode.
86 * for single-file segments, we need some kind of mechanism to keep track of
87 * which hugepages can be freed back to the system, and which cannot. we cannot
88 * use flock() because they don't allow locking parts of a file, and we cannot
89 * use fcntl() due to issues with their semantics, so we will have to rely on a
90 * bunch of lockfiles for each page. so, we will use 'fds' array to keep track
91 * of per-page lockfiles. we will store the actual segment list fd in the
92 * 'memseg_list_fd' field.
94 * for file-per-page mode, each page will have its own fd, so 'memseg_list_fd'
95 * will be invalid (set to -1), and we'll use 'fds' to keep track of page fd's.
97 * we cannot know how many pages a system will have in advance, but we do know
98 * that they come in lists, and we know lengths of these lists. so, simply store
99 * a malloc'd array of fd's indexed by list and segment index.
101 * they will be initialized at startup, and filled as we allocate/deallocate
105 int *fds; /**< dynamically allocated array of segment lock fd's */
106 int memseg_list_fd; /**< memseg list fd */
107 int len; /**< total length of the array */
108 int count; /**< entries used in an array */
109 } fd_list[RTE_MAX_MEMSEG_LISTS];
111 /** local copy of a memory map, used to synchronize memory hotplug in MP */
112 static struct rte_memseg_list local_memsegs[RTE_MAX_MEMSEG_LISTS];
114 static sigjmp_buf huge_jmpenv;
116 static void __rte_unused huge_sigbus_handler(int signo __rte_unused)
118 siglongjmp(huge_jmpenv, 1);
121 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
122 * non-static local variable in the stack frame calling sigsetjmp might be
123 * clobbered by a call to longjmp.
125 static int __rte_unused huge_wrap_sigsetjmp(void)
127 return sigsetjmp(huge_jmpenv, 1);
130 static struct sigaction huge_action_old;
131 static int huge_need_recover;
133 static void __rte_unused
134 huge_register_sigbus(void)
137 struct sigaction action;
140 sigaddset(&mask, SIGBUS);
142 action.sa_mask = mask;
143 action.sa_handler = huge_sigbus_handler;
145 huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
148 static void __rte_unused
149 huge_recover_sigbus(void)
151 if (huge_need_recover) {
152 sigaction(SIGBUS, &huge_action_old, NULL);
153 huge_need_recover = 0;
157 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
162 /* Check if kernel supports NUMA. */
163 if (numa_available() != 0) {
164 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
171 prepare_numa(int *oldpolicy, struct bitmask *oldmask, int socket_id)
173 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
174 if (get_mempolicy(oldpolicy, oldmask->maskp,
175 oldmask->size + 1, 0, 0) < 0) {
177 "Failed to get current mempolicy: %s. "
178 "Assuming MPOL_DEFAULT.\n", strerror(errno));
179 *oldpolicy = MPOL_DEFAULT;
182 "Setting policy MPOL_PREFERRED for socket %d\n",
184 numa_set_preferred(socket_id);
188 restore_numa(int *oldpolicy, struct bitmask *oldmask)
191 "Restoring previous memory policy: %d\n", *oldpolicy);
192 if (*oldpolicy == MPOL_DEFAULT) {
193 numa_set_localalloc();
194 } else if (set_mempolicy(*oldpolicy, oldmask->maskp,
195 oldmask->size + 1) < 0) {
196 RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
198 numa_set_localalloc();
200 numa_free_cpumask(oldmask);
205 * uses fstat to report the size of a file on disk
208 get_file_size(int fd)
211 if (fstat(fd, &st) < 0)
217 pagesz_flags(uint64_t page_sz)
219 /* as per mmap() manpage, all page sizes are log2 of page size
220 * shifted by MAP_HUGE_SHIFT
222 int log2 = rte_log2_u64(page_sz);
223 return log2 << RTE_MAP_HUGE_SHIFT;
226 /* returns 1 on successful lock, 0 on unsuccessful lock, -1 on error */
227 static int lock(int fd, int type)
231 /* flock may be interrupted */
233 ret = flock(fd, type | LOCK_NB);
234 } while (ret && errno == EINTR);
236 if (ret && errno == EWOULDBLOCK) {
240 RTE_LOG(ERR, EAL, "%s(): error calling flock(): %s\n",
241 __func__, strerror(errno));
244 /* lock was successful */
248 static int get_segment_lock_fd(int list_idx, int seg_idx)
250 char path[PATH_MAX] = {0};
253 if (list_idx < 0 || list_idx >= (int)RTE_DIM(fd_list))
255 if (seg_idx < 0 || seg_idx >= fd_list[list_idx].len)
258 fd = fd_list[list_idx].fds[seg_idx];
259 /* does this lock already exist? */
263 eal_get_hugefile_lock_path(path, sizeof(path),
264 list_idx * RTE_MAX_MEMSEG_PER_LIST + seg_idx);
266 fd = open(path, O_CREAT | O_RDWR, 0660);
268 RTE_LOG(ERR, EAL, "%s(): error creating lockfile '%s': %s\n",
269 __func__, path, strerror(errno));
272 /* take out a read lock */
273 if (lock(fd, LOCK_SH) != 1) {
274 RTE_LOG(ERR, EAL, "%s(): failed to take out a readlock on '%s': %s\n",
275 __func__, path, strerror(errno));
279 /* store it for future reference */
280 fd_list[list_idx].fds[seg_idx] = fd;
281 fd_list[list_idx].count++;
285 static int unlock_segment(int list_idx, int seg_idx)
289 if (list_idx < 0 || list_idx >= (int)RTE_DIM(fd_list))
291 if (seg_idx < 0 || seg_idx >= fd_list[list_idx].len)
294 fd = fd_list[list_idx].fds[seg_idx];
296 /* upgrade lock to exclusive to see if we can remove the lockfile */
297 ret = lock(fd, LOCK_EX);
299 /* we've succeeded in taking exclusive lock, this lockfile may
302 char path[PATH_MAX] = {0};
303 eal_get_hugefile_lock_path(path, sizeof(path),
304 list_idx * RTE_MAX_MEMSEG_PER_LIST + seg_idx);
306 RTE_LOG(ERR, EAL, "%s(): error removing lockfile '%s': %s\n",
307 __func__, path, strerror(errno));
310 /* we don't want to leak the fd, so even if we fail to lock, close fd
311 * and remove it from list anyway.
314 fd_list[list_idx].fds[seg_idx] = -1;
315 fd_list[list_idx].count--;
323 get_seg_memfd(struct hugepage_info *hi __rte_unused,
324 unsigned int list_idx __rte_unused,
325 unsigned int seg_idx __rte_unused)
327 #ifdef MEMFD_SUPPORTED
329 char segname[250]; /* as per manpage, limit is 249 bytes plus null */
331 int flags = RTE_MFD_HUGETLB | pagesz_flags(hi->hugepage_sz);
333 if (internal_config.single_file_segments) {
334 fd = fd_list[list_idx].memseg_list_fd;
337 snprintf(segname, sizeof(segname), "seg_%i", list_idx);
338 fd = memfd_create(segname, flags);
340 RTE_LOG(DEBUG, EAL, "%s(): memfd create failed: %s\n",
341 __func__, strerror(errno));
344 fd_list[list_idx].memseg_list_fd = fd;
347 fd = fd_list[list_idx].fds[seg_idx];
350 snprintf(segname, sizeof(segname), "seg_%i-%i",
352 fd = memfd_create(segname, flags);
354 RTE_LOG(DEBUG, EAL, "%s(): memfd create failed: %s\n",
355 __func__, strerror(errno));
358 fd_list[list_idx].fds[seg_idx] = fd;
367 get_seg_fd(char *path, int buflen, struct hugepage_info *hi,
368 unsigned int list_idx, unsigned int seg_idx)
372 /* for in-memory mode, we only make it here when we're sure we support
373 * memfd, and this is a special case.
375 if (internal_config.in_memory)
376 return get_seg_memfd(hi, list_idx, seg_idx);
378 if (internal_config.single_file_segments) {
379 /* create a hugepage file path */
380 eal_get_hugefile_path(path, buflen, hi->hugedir, list_idx);
382 fd = fd_list[list_idx].memseg_list_fd;
385 fd = open(path, O_CREAT | O_RDWR, 0600);
387 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n",
388 __func__, strerror(errno));
391 /* take out a read lock and keep it indefinitely */
392 if (lock(fd, LOCK_SH) < 0) {
393 RTE_LOG(ERR, EAL, "%s(): lock failed: %s\n",
394 __func__, strerror(errno));
398 fd_list[list_idx].memseg_list_fd = fd;
401 /* create a hugepage file path */
402 eal_get_hugefile_path(path, buflen, hi->hugedir,
403 list_idx * RTE_MAX_MEMSEG_PER_LIST + seg_idx);
405 fd = fd_list[list_idx].fds[seg_idx];
408 fd = open(path, O_CREAT | O_RDWR, 0600);
410 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n",
411 __func__, strerror(errno));
414 /* take out a read lock */
415 if (lock(fd, LOCK_SH) < 0) {
416 RTE_LOG(ERR, EAL, "%s(): lock failed: %s\n",
417 __func__, strerror(errno));
421 fd_list[list_idx].fds[seg_idx] = fd;
428 resize_hugefile(int fd, char *path, int list_idx, int seg_idx,
429 uint64_t fa_offset, uint64_t page_sz, bool grow)
433 /* in-memory mode is a special case, because we don't need to perform
434 * any locking, and we can be sure that fallocate() is supported.
436 if (internal_config.in_memory) {
437 int flags = grow ? 0 : FALLOC_FL_PUNCH_HOLE |
441 /* grow or shrink the file */
442 ret = fallocate(fd, flags, fa_offset, page_sz);
445 RTE_LOG(DEBUG, EAL, "%s(): fallocate() failed: %s\n",
450 /* increase/decrease total segment count */
451 fd_list[list_idx].count += (grow ? 1 : -1);
452 if (!grow && fd_list[list_idx].count == 0) {
453 close(fd_list[list_idx].memseg_list_fd);
454 fd_list[list_idx].memseg_list_fd = -1;
460 if (fallocate_supported == 0) {
461 /* we cannot deallocate memory if fallocate() is not
462 * supported, and hugepage file is already locked at
463 * creation, so no further synchronization needed.
467 RTE_LOG(DEBUG, EAL, "%s(): fallocate not supported, not freeing page back to the system\n",
471 uint64_t new_size = fa_offset + page_sz;
472 uint64_t cur_size = get_file_size(fd);
474 /* fallocate isn't supported, fall back to ftruncate */
475 if (new_size > cur_size &&
476 ftruncate(fd, new_size) < 0) {
477 RTE_LOG(DEBUG, EAL, "%s(): ftruncate() failed: %s\n",
478 __func__, strerror(errno));
482 int flags = grow ? 0 : FALLOC_FL_PUNCH_HOLE |
486 /* if fallocate() is supported, we need to take out a
487 * read lock on allocate (to prevent other processes
488 * from deallocating this page), and take out a write
489 * lock on deallocate (to ensure nobody else is using
492 * read locks on page itself are already taken out at
493 * file creation, in get_seg_fd().
495 * we cannot rely on simple use of flock() call, because
496 * we need to be able to lock a section of the file,
497 * and we cannot use fcntl() locks, because of numerous
498 * problems with their semantics, so we will use
499 * deterministically named lock files for each section
502 * if we're shrinking the file, we want to upgrade our
503 * lock from shared to exclusive.
505 * lock_fd is an fd for a lockfile, not for the segment
508 lock_fd = get_segment_lock_fd(list_idx, seg_idx);
511 /* we are using this lockfile to determine
512 * whether this particular page is locked, as we
513 * are in single file segments mode and thus
514 * cannot use regular flock() to get this info.
516 * we want to try and take out an exclusive lock
517 * on the lock file to determine if we're the
518 * last ones using this page, and if not, we
519 * won't be shrinking it, and will instead exit
522 ret = lock(lock_fd, LOCK_EX);
524 /* drop the lock on the lockfile, so that even
525 * if we couldn't shrink the file ourselves, we
526 * are signalling to other processes that we're
527 * no longer using this page.
529 if (unlock_segment(list_idx, seg_idx))
530 RTE_LOG(ERR, EAL, "Could not unlock segment\n");
532 /* additionally, if this was the last lock on
533 * this segment list, we can safely close the
534 * page file fd, so that one of the processes
535 * could then delete the file after shrinking.
537 if (ret < 1 && fd_list[list_idx].count == 0) {
539 fd_list[list_idx].memseg_list_fd = -1;
543 RTE_LOG(ERR, EAL, "Could not lock segment\n");
547 /* failed to lock, not an error. */
551 /* grow or shrink the file */
552 ret = fallocate(fd, flags, fa_offset, page_sz);
555 if (fallocate_supported == -1 &&
557 RTE_LOG(ERR, EAL, "%s(): fallocate() not supported, hugepage deallocation will be disabled\n",
560 fallocate_supported = 0;
562 RTE_LOG(DEBUG, EAL, "%s(): fallocate() failed: %s\n",
568 fallocate_supported = 1;
570 /* we've grew/shrunk the file, and we hold an
571 * exclusive lock now. check if there are no
572 * more segments active in this segment list,
573 * and remove the file if there aren't.
575 if (fd_list[list_idx].count == 0) {
577 RTE_LOG(ERR, EAL, "%s(): unlinking '%s' failed: %s\n",
581 fd_list[list_idx].memseg_list_fd = -1;
590 alloc_seg(struct rte_memseg *ms, void *addr, int socket_id,
591 struct hugepage_info *hi, unsigned int list_idx,
592 unsigned int seg_idx)
594 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
595 int cur_socket_id = 0;
607 alloc_sz = hi->hugepage_sz;
609 /* these are checked at init, but code analyzers don't know that */
610 if (internal_config.in_memory && !anonymous_hugepages_supported) {
611 RTE_LOG(ERR, EAL, "Anonymous hugepages not supported, in-memory mode cannot allocate memory\n");
614 if (internal_config.in_memory && !memfd_create_supported &&
615 internal_config.single_file_segments) {
616 RTE_LOG(ERR, EAL, "Single-file segments are not supported without memfd support\n");
620 /* in-memory without memfd is a special case */
623 if (internal_config.in_memory && !memfd_create_supported) {
624 const int in_memory_flags = MAP_HUGETLB | MAP_FIXED |
625 MAP_PRIVATE | MAP_ANONYMOUS;
628 pagesz_flag = pagesz_flags(alloc_sz);
630 mmap_flags = in_memory_flags | pagesz_flag;
632 /* single-file segments codepath will never be active
633 * here because in-memory mode is incompatible with the
634 * fallback path, and it's stopped at EAL initialization
639 /* takes out a read lock on segment or segment list */
640 fd = get_seg_fd(path, sizeof(path), hi, list_idx, seg_idx);
642 RTE_LOG(ERR, EAL, "Couldn't get fd on hugepage file\n");
646 if (internal_config.single_file_segments) {
647 map_offset = seg_idx * alloc_sz;
648 ret = resize_hugefile(fd, path, list_idx, seg_idx,
649 map_offset, alloc_sz, true);
654 if (ftruncate(fd, alloc_sz) < 0) {
655 RTE_LOG(DEBUG, EAL, "%s(): ftruncate() failed: %s\n",
656 __func__, strerror(errno));
659 if (internal_config.hugepage_unlink &&
660 !internal_config.in_memory) {
662 RTE_LOG(DEBUG, EAL, "%s(): unlink() failed: %s\n",
663 __func__, strerror(errno));
668 mmap_flags = MAP_SHARED | MAP_POPULATE | MAP_FIXED;
672 * map the segment, and populate page tables, the kernel fills
673 * this segment with zeros if it's a new page.
675 va = mmap(addr, alloc_sz, PROT_READ | PROT_WRITE, mmap_flags, fd,
678 if (va == MAP_FAILED) {
679 RTE_LOG(DEBUG, EAL, "%s(): mmap() failed: %s\n", __func__,
681 /* mmap failed, but the previous region might have been
682 * unmapped anyway. try to remap it
687 RTE_LOG(DEBUG, EAL, "%s(): wrong mmap() address\n", __func__);
688 munmap(va, alloc_sz);
692 /* In linux, hugetlb limitations, like cgroup, are
693 * enforced at fault time instead of mmap(), even
694 * with the option of MAP_POPULATE. Kernel will send
695 * a SIGBUS signal. To avoid to be killed, save stack
696 * environment here, if SIGBUS happens, we can jump
699 if (huge_wrap_sigsetjmp()) {
700 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more hugepages of size %uMB\n",
701 (unsigned int)(alloc_sz >> 20));
705 /* we need to trigger a write to the page to enforce page fault and
706 * ensure that page is accessible to us, but we can't overwrite value
707 * that is already there, so read the old value, and write itback.
708 * kernel populates the page with zeroes initially.
710 *(volatile int *)addr = *(volatile int *)addr;
712 iova = rte_mem_virt2iova(addr);
713 if (iova == RTE_BAD_PHYS_ADDR) {
714 RTE_LOG(DEBUG, EAL, "%s(): can't get IOVA addr\n",
719 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
720 move_pages(getpid(), 1, &addr, NULL, &cur_socket_id, 0);
722 if (cur_socket_id != socket_id) {
724 "%s(): allocation happened on wrong socket (wanted %d, got %d)\n",
725 __func__, socket_id, cur_socket_id);
731 ms->hugepage_sz = alloc_sz;
733 ms->nchannel = rte_memory_get_nchannel();
734 ms->nrank = rte_memory_get_nrank();
736 ms->socket_id = socket_id;
741 munmap(addr, alloc_sz);
744 new_addr = eal_get_virtual_area(addr, &alloc_sz, alloc_sz, 0, flags);
745 if (new_addr != addr) {
746 if (new_addr != NULL)
747 munmap(new_addr, alloc_sz);
748 /* we're leaving a hole in our virtual address space. if
749 * somebody else maps this hole now, we could accidentally
750 * override it in the future.
752 RTE_LOG(CRIT, EAL, "Can't mmap holes in our virtual address space\n");
755 /* some codepaths will return negative fd, so exit early */
759 if (internal_config.single_file_segments) {
760 resize_hugefile(fd, path, list_idx, seg_idx, map_offset,
762 /* ignore failure, can't make it any worse */
764 /* only remove file if we can take out a write lock */
765 if (internal_config.hugepage_unlink == 0 &&
766 internal_config.in_memory == 0 &&
767 lock(fd, LOCK_EX) == 1)
770 fd_list[list_idx].fds[seg_idx] = -1;
776 free_seg(struct rte_memseg *ms, struct hugepage_info *hi,
777 unsigned int list_idx, unsigned int seg_idx)
784 /* erase page data */
785 memset(ms->addr, 0, ms->len);
787 if (mmap(ms->addr, ms->len, PROT_READ,
788 MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, 0) ==
790 RTE_LOG(DEBUG, EAL, "couldn't unmap page\n");
796 /* if we're using anonymous hugepages, nothing to be done */
797 if (internal_config.in_memory && !memfd_create_supported)
800 /* if we've already unlinked the page, nothing needs to be done */
801 if (!internal_config.in_memory && internal_config.hugepage_unlink)
805 memset(ms, 0, sizeof(*ms));
809 /* if we are not in single file segments mode, we're going to unmap the
810 * segment and thus drop the lock on original fd, but hugepage dir is
811 * now locked so we can take out another one without races.
813 fd = get_seg_fd(path, sizeof(path), hi, list_idx, seg_idx);
817 if (internal_config.single_file_segments) {
818 map_offset = seg_idx * ms->len;
819 if (resize_hugefile(fd, path, list_idx, seg_idx, map_offset,
824 /* if we're able to take out a write lock, we're the last one
825 * holding onto this page.
827 if (!internal_config.in_memory) {
828 ret = lock(fd, LOCK_EX);
830 /* no one else is using this page */
835 /* closing fd will drop the lock */
837 fd_list[list_idx].fds[seg_idx] = -1;
840 memset(ms, 0, sizeof(*ms));
842 return ret < 0 ? -1 : 0;
845 struct alloc_walk_param {
846 struct hugepage_info *hi;
847 struct rte_memseg **ms;
849 unsigned int segs_allocated;
855 alloc_seg_walk(const struct rte_memseg_list *msl, void *arg)
857 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
858 struct alloc_walk_param *wa = arg;
859 struct rte_memseg_list *cur_msl;
861 int cur_idx, start_idx, j, dir_fd = -1;
862 unsigned int msl_idx, need, i;
864 if (msl->page_sz != wa->page_sz)
866 if (msl->socket_id != wa->socket)
869 page_sz = (size_t)msl->page_sz;
871 msl_idx = msl - mcfg->memsegs;
872 cur_msl = &mcfg->memsegs[msl_idx];
876 /* try finding space in memseg list */
877 cur_idx = rte_fbarray_find_next_n_free(&cur_msl->memseg_arr, 0, need);
882 /* do not allow any page allocations during the time we're allocating,
883 * because file creation and locking operations are not atomic,
884 * and we might be the first or the last ones to use a particular page,
885 * so we need to ensure atomicity of every operation.
887 * during init, we already hold a write lock, so don't try to take out
890 if (wa->hi->lock_descriptor == -1 && !internal_config.in_memory) {
891 dir_fd = open(wa->hi->hugedir, O_RDONLY);
893 RTE_LOG(ERR, EAL, "%s(): Cannot open '%s': %s\n",
894 __func__, wa->hi->hugedir, strerror(errno));
897 /* blocking writelock */
898 if (flock(dir_fd, LOCK_EX)) {
899 RTE_LOG(ERR, EAL, "%s(): Cannot lock '%s': %s\n",
900 __func__, wa->hi->hugedir, strerror(errno));
906 for (i = 0; i < need; i++, cur_idx++) {
907 struct rte_memseg *cur;
910 cur = rte_fbarray_get(&cur_msl->memseg_arr, cur_idx);
911 map_addr = RTE_PTR_ADD(cur_msl->base_va,
914 if (alloc_seg(cur, map_addr, wa->socket, wa->hi,
916 RTE_LOG(DEBUG, EAL, "attempted to allocate %i segments, but only %i were allocated\n",
919 /* if exact number wasn't requested, stop */
924 for (j = start_idx; j < cur_idx; j++) {
925 struct rte_memseg *tmp;
926 struct rte_fbarray *arr =
927 &cur_msl->memseg_arr;
929 tmp = rte_fbarray_get(arr, j);
930 rte_fbarray_set_free(arr, j);
932 /* free_seg may attempt to create a file, which
935 if (free_seg(tmp, wa->hi, msl_idx, j))
936 RTE_LOG(DEBUG, EAL, "Cannot free page\n");
940 memset(wa->ms, 0, sizeof(*wa->ms) * wa->n_segs);
949 rte_fbarray_set_used(&cur_msl->memseg_arr, cur_idx);
952 wa->segs_allocated = i;
960 struct free_walk_param {
961 struct hugepage_info *hi;
962 struct rte_memseg *ms;
965 free_seg_walk(const struct rte_memseg_list *msl, void *arg)
967 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
968 struct rte_memseg_list *found_msl;
969 struct free_walk_param *wa = arg;
970 uintptr_t start_addr, end_addr;
971 int msl_idx, seg_idx, ret, dir_fd = -1;
973 start_addr = (uintptr_t) msl->base_va;
974 end_addr = start_addr + msl->len;
976 if ((uintptr_t)wa->ms->addr < start_addr ||
977 (uintptr_t)wa->ms->addr >= end_addr)
980 msl_idx = msl - mcfg->memsegs;
981 seg_idx = RTE_PTR_DIFF(wa->ms->addr, start_addr) / msl->page_sz;
984 found_msl = &mcfg->memsegs[msl_idx];
986 /* do not allow any page allocations during the time we're freeing,
987 * because file creation and locking operations are not atomic,
988 * and we might be the first or the last ones to use a particular page,
989 * so we need to ensure atomicity of every operation.
991 * during init, we already hold a write lock, so don't try to take out
994 if (wa->hi->lock_descriptor == -1 && !internal_config.in_memory) {
995 dir_fd = open(wa->hi->hugedir, O_RDONLY);
997 RTE_LOG(ERR, EAL, "%s(): Cannot open '%s': %s\n",
998 __func__, wa->hi->hugedir, strerror(errno));
1001 /* blocking writelock */
1002 if (flock(dir_fd, LOCK_EX)) {
1003 RTE_LOG(ERR, EAL, "%s(): Cannot lock '%s': %s\n",
1004 __func__, wa->hi->hugedir, strerror(errno));
1010 found_msl->version++;
1012 rte_fbarray_set_free(&found_msl->memseg_arr, seg_idx);
1014 ret = free_seg(wa->ms, wa->hi, msl_idx, seg_idx);
1026 eal_memalloc_alloc_seg_bulk(struct rte_memseg **ms, int n_segs, size_t page_sz,
1027 int socket, bool exact)
1030 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
1031 bool have_numa = false;
1033 struct bitmask *oldmask;
1035 struct alloc_walk_param wa;
1036 struct hugepage_info *hi = NULL;
1038 memset(&wa, 0, sizeof(wa));
1040 /* dynamic allocation not supported in legacy mode */
1041 if (internal_config.legacy_mem)
1044 for (i = 0; i < (int) RTE_DIM(internal_config.hugepage_info); i++) {
1046 internal_config.hugepage_info[i].hugepage_sz) {
1047 hi = &internal_config.hugepage_info[i];
1052 RTE_LOG(ERR, EAL, "%s(): can't find relevant hugepage_info entry\n",
1057 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
1059 oldmask = numa_allocate_nodemask();
1060 prepare_numa(&oldpolicy, oldmask, socket);
1069 wa.page_sz = page_sz;
1071 wa.segs_allocated = 0;
1073 /* memalloc is locked, so it's safe to use thread-unsafe version */
1074 ret = rte_memseg_list_walk_thread_unsafe(alloc_seg_walk, &wa);
1076 RTE_LOG(ERR, EAL, "%s(): couldn't find suitable memseg_list\n",
1079 } else if (ret > 0) {
1080 ret = (int)wa.segs_allocated;
1083 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
1085 restore_numa(&oldpolicy, oldmask);
1091 eal_memalloc_alloc_seg(size_t page_sz, int socket)
1093 struct rte_memseg *ms;
1094 if (eal_memalloc_alloc_seg_bulk(&ms, 1, page_sz, socket, true) < 0)
1096 /* return pointer to newly allocated memseg */
1101 eal_memalloc_free_seg_bulk(struct rte_memseg **ms, int n_segs)
1105 /* dynamic free not supported in legacy mode */
1106 if (internal_config.legacy_mem)
1109 for (seg = 0; seg < n_segs; seg++) {
1110 struct rte_memseg *cur = ms[seg];
1111 struct hugepage_info *hi = NULL;
1112 struct free_walk_param wa;
1115 /* if this page is marked as unfreeable, fail */
1116 if (cur->flags & RTE_MEMSEG_FLAG_DO_NOT_FREE) {
1117 RTE_LOG(DEBUG, EAL, "Page is not allowed to be freed\n");
1122 memset(&wa, 0, sizeof(wa));
1124 for (i = 0; i < (int)RTE_DIM(internal_config.hugepage_info);
1126 hi = &internal_config.hugepage_info[i];
1127 if (cur->hugepage_sz == hi->hugepage_sz)
1130 if (i == (int)RTE_DIM(internal_config.hugepage_info)) {
1131 RTE_LOG(ERR, EAL, "Can't find relevant hugepage_info entry\n");
1139 /* memalloc is locked, so it's safe to use thread-unsafe version
1141 walk_res = rte_memseg_list_walk_thread_unsafe(free_seg_walk,
1146 RTE_LOG(ERR, EAL, "Couldn't find memseg list\n");
1153 eal_memalloc_free_seg(struct rte_memseg *ms)
1155 /* dynamic free not supported in legacy mode */
1156 if (internal_config.legacy_mem)
1159 return eal_memalloc_free_seg_bulk(&ms, 1);
1163 sync_chunk(struct rte_memseg_list *primary_msl,
1164 struct rte_memseg_list *local_msl, struct hugepage_info *hi,
1165 unsigned int msl_idx, bool used, int start, int end)
1167 struct rte_fbarray *l_arr, *p_arr;
1168 int i, ret, chunk_len, diff_len;
1170 l_arr = &local_msl->memseg_arr;
1171 p_arr = &primary_msl->memseg_arr;
1173 /* we need to aggregate allocations/deallocations into bigger chunks,
1174 * as we don't want to spam the user with per-page callbacks.
1176 * to avoid any potential issues, we also want to trigger
1177 * deallocation callbacks *before* we actually deallocate
1178 * memory, so that the user application could wrap up its use
1179 * before it goes away.
1182 chunk_len = end - start;
1184 /* find how many contiguous pages we can map/unmap for this chunk */
1186 rte_fbarray_find_contig_free(l_arr, start) :
1187 rte_fbarray_find_contig_used(l_arr, start);
1189 /* has to be at least one page */
1193 diff_len = RTE_MIN(chunk_len, diff_len);
1195 /* if we are freeing memory, notify the application */
1197 struct rte_memseg *ms;
1199 size_t len, page_sz;
1201 ms = rte_fbarray_get(l_arr, start);
1202 start_va = ms->addr;
1203 page_sz = (size_t)primary_msl->page_sz;
1204 len = page_sz * diff_len;
1206 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
1210 for (i = 0; i < diff_len; i++) {
1211 struct rte_memseg *p_ms, *l_ms;
1212 int seg_idx = start + i;
1214 l_ms = rte_fbarray_get(l_arr, seg_idx);
1215 p_ms = rte_fbarray_get(p_arr, seg_idx);
1217 if (l_ms == NULL || p_ms == NULL)
1221 ret = alloc_seg(l_ms, p_ms->addr,
1222 p_ms->socket_id, hi,
1226 rte_fbarray_set_used(l_arr, seg_idx);
1228 ret = free_seg(l_ms, hi, msl_idx, seg_idx);
1229 rte_fbarray_set_free(l_arr, seg_idx);
1235 /* if we just allocated memory, notify the application */
1237 struct rte_memseg *ms;
1239 size_t len, page_sz;
1241 ms = rte_fbarray_get(l_arr, start);
1242 start_va = ms->addr;
1243 page_sz = (size_t)primary_msl->page_sz;
1244 len = page_sz * diff_len;
1246 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC,
1250 /* calculate how much we can advance until next chunk */
1252 rte_fbarray_find_contig_used(l_arr, start) :
1253 rte_fbarray_find_contig_free(l_arr, start);
1254 ret = RTE_MIN(chunk_len, diff_len);
1260 sync_status(struct rte_memseg_list *primary_msl,
1261 struct rte_memseg_list *local_msl, struct hugepage_info *hi,
1262 unsigned int msl_idx, bool used)
1264 struct rte_fbarray *l_arr, *p_arr;
1265 int p_idx, l_chunk_len, p_chunk_len, ret;
1268 /* this is a little bit tricky, but the basic idea is - walk both lists
1269 * and spot any places where there are discrepancies. walking both lists
1270 * and noting discrepancies in a single go is a hard problem, so we do
1271 * it in two passes - first we spot any places where allocated segments
1272 * mismatch (i.e. ensure that everything that's allocated in the primary
1273 * is also allocated in the secondary), and then we do it by looking at
1274 * free segments instead.
1276 * we also need to aggregate changes into chunks, as we have to call
1277 * callbacks per allocation, not per page.
1279 l_arr = &local_msl->memseg_arr;
1280 p_arr = &primary_msl->memseg_arr;
1283 p_idx = rte_fbarray_find_next_used(p_arr, 0);
1285 p_idx = rte_fbarray_find_next_free(p_arr, 0);
1287 while (p_idx >= 0) {
1288 int next_chunk_search_idx;
1291 p_chunk_len = rte_fbarray_find_contig_used(p_arr,
1293 l_chunk_len = rte_fbarray_find_contig_used(l_arr,
1296 p_chunk_len = rte_fbarray_find_contig_free(p_arr,
1298 l_chunk_len = rte_fbarray_find_contig_free(l_arr,
1301 /* best case scenario - no differences (or bigger, which will be
1302 * fixed during next iteration), look for next chunk
1304 if (l_chunk_len >= p_chunk_len) {
1305 next_chunk_search_idx = p_idx + p_chunk_len;
1309 /* if both chunks start at the same point, skip parts we know
1310 * are identical, and sync the rest. each call to sync_chunk
1311 * will only sync contiguous segments, so we need to call this
1312 * until we are sure there are no more differences in this
1315 start = p_idx + l_chunk_len;
1316 end = p_idx + p_chunk_len;
1318 ret = sync_chunk(primary_msl, local_msl, hi, msl_idx,
1321 } while (start < end && ret >= 0);
1322 /* if ret is negative, something went wrong */
1326 next_chunk_search_idx = p_idx + p_chunk_len;
1328 /* skip to end of this chunk */
1330 p_idx = rte_fbarray_find_next_used(p_arr,
1331 next_chunk_search_idx);
1333 p_idx = rte_fbarray_find_next_free(p_arr,
1334 next_chunk_search_idx);
1341 sync_existing(struct rte_memseg_list *primary_msl,
1342 struct rte_memseg_list *local_msl, struct hugepage_info *hi,
1343 unsigned int msl_idx)
1347 /* do not allow any page allocations during the time we're allocating,
1348 * because file creation and locking operations are not atomic,
1349 * and we might be the first or the last ones to use a particular page,
1350 * so we need to ensure atomicity of every operation.
1352 dir_fd = open(hi->hugedir, O_RDONLY);
1354 RTE_LOG(ERR, EAL, "%s(): Cannot open '%s': %s\n", __func__,
1355 hi->hugedir, strerror(errno));
1358 /* blocking writelock */
1359 if (flock(dir_fd, LOCK_EX)) {
1360 RTE_LOG(ERR, EAL, "%s(): Cannot lock '%s': %s\n", __func__,
1361 hi->hugedir, strerror(errno));
1366 /* ensure all allocated space is the same in both lists */
1367 ret = sync_status(primary_msl, local_msl, hi, msl_idx, true);
1371 /* ensure all unallocated space is the same in both lists */
1372 ret = sync_status(primary_msl, local_msl, hi, msl_idx, false);
1376 /* update version number */
1377 local_msl->version = primary_msl->version;
1388 sync_walk(const struct rte_memseg_list *msl, void *arg __rte_unused)
1390 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1391 struct rte_memseg_list *primary_msl, *local_msl;
1392 struct hugepage_info *hi = NULL;
1399 msl_idx = msl - mcfg->memsegs;
1400 primary_msl = &mcfg->memsegs[msl_idx];
1401 local_msl = &local_memsegs[msl_idx];
1403 for (i = 0; i < RTE_DIM(internal_config.hugepage_info); i++) {
1405 internal_config.hugepage_info[i].hugepage_sz;
1406 uint64_t msl_sz = primary_msl->page_sz;
1407 if (msl_sz == cur_sz) {
1408 hi = &internal_config.hugepage_info[i];
1413 RTE_LOG(ERR, EAL, "Can't find relevant hugepage_info entry\n");
1417 /* if versions don't match, synchronize everything */
1418 if (local_msl->version != primary_msl->version &&
1419 sync_existing(primary_msl, local_msl, hi, msl_idx))
1426 eal_memalloc_sync_with_primary(void)
1428 /* nothing to be done in primary */
1429 if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1432 /* memalloc is locked, so it's safe to call thread-unsafe version */
1433 if (rte_memseg_list_walk_thread_unsafe(sync_walk, NULL))
1439 secondary_msl_create_walk(const struct rte_memseg_list *msl,
1440 void *arg __rte_unused)
1442 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1443 struct rte_memseg_list *primary_msl, *local_msl;
1444 char name[PATH_MAX];
1450 msl_idx = msl - mcfg->memsegs;
1451 primary_msl = &mcfg->memsegs[msl_idx];
1452 local_msl = &local_memsegs[msl_idx];
1454 /* create distinct fbarrays for each secondary */
1455 snprintf(name, RTE_FBARRAY_NAME_LEN, "%s_%i",
1456 primary_msl->memseg_arr.name, getpid());
1458 ret = rte_fbarray_init(&local_msl->memseg_arr, name,
1459 primary_msl->memseg_arr.len,
1460 primary_msl->memseg_arr.elt_sz);
1462 RTE_LOG(ERR, EAL, "Cannot initialize local memory map\n");
1465 local_msl->base_va = primary_msl->base_va;
1466 local_msl->len = primary_msl->len;
1472 alloc_list(int list_idx, int len)
1477 /* ensure we have space to store fd per each possible segment */
1478 data = malloc(sizeof(int) * len);
1480 RTE_LOG(ERR, EAL, "Unable to allocate space for file descriptors\n");
1483 /* set all fd's as invalid */
1484 for (i = 0; i < len; i++)
1487 fd_list[list_idx].fds = data;
1488 fd_list[list_idx].len = len;
1489 fd_list[list_idx].count = 0;
1490 fd_list[list_idx].memseg_list_fd = -1;
1496 fd_list_create_walk(const struct rte_memseg_list *msl,
1497 void *arg __rte_unused)
1499 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1506 msl_idx = msl - mcfg->memsegs;
1507 len = msl->memseg_arr.len;
1509 return alloc_list(msl_idx, len);
1513 eal_memalloc_set_seg_fd(int list_idx, int seg_idx, int fd)
1515 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1517 /* single file segments mode doesn't support individual segment fd's */
1518 if (internal_config.single_file_segments)
1521 /* if list is not allocated, allocate it */
1522 if (fd_list[list_idx].len == 0) {
1523 int len = mcfg->memsegs[list_idx].memseg_arr.len;
1525 if (alloc_list(list_idx, len) < 0)
1528 fd_list[list_idx].fds[seg_idx] = fd;
1534 eal_memalloc_set_seg_list_fd(int list_idx, int fd)
1536 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1538 /* non-single file segment mode doesn't support segment list fd's */
1539 if (!internal_config.single_file_segments)
1542 /* if list is not allocated, allocate it */
1543 if (fd_list[list_idx].len == 0) {
1544 int len = mcfg->memsegs[list_idx].memseg_arr.len;
1546 if (alloc_list(list_idx, len) < 0)
1550 fd_list[list_idx].memseg_list_fd = fd;
1556 eal_memalloc_get_seg_fd(int list_idx, int seg_idx)
1560 if (internal_config.in_memory || internal_config.no_hugetlbfs) {
1561 #ifndef MEMFD_SUPPORTED
1562 /* in in-memory or no-huge mode, we rely on memfd support */
1565 /* memfd supported, but hugetlbfs memfd may not be */
1566 if (!internal_config.no_hugetlbfs && !memfd_create_supported)
1570 if (internal_config.single_file_segments) {
1571 fd = fd_list[list_idx].memseg_list_fd;
1572 } else if (fd_list[list_idx].len == 0) {
1573 /* list not initialized */
1576 fd = fd_list[list_idx].fds[seg_idx];
1584 test_memfd_create(void)
1586 #ifdef MEMFD_SUPPORTED
1588 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
1589 uint64_t pagesz = internal_config.hugepage_info[i].hugepage_sz;
1590 int pagesz_flag = pagesz_flags(pagesz);
1593 flags = pagesz_flag | RTE_MFD_HUGETLB;
1594 int fd = memfd_create("test", flags);
1596 /* we failed - let memalloc know this isn't working */
1597 if (errno == EINVAL) {
1598 memfd_create_supported = 0;
1599 return 0; /* not supported */
1602 /* we got other error - something's wrong */
1603 return -1; /* error */
1606 return 1; /* supported */
1609 return 0; /* not supported */
1613 eal_memalloc_get_seg_fd_offset(int list_idx, int seg_idx, size_t *offset)
1615 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1617 if (internal_config.in_memory || internal_config.no_hugetlbfs) {
1618 #ifndef MEMFD_SUPPORTED
1619 /* in in-memory or no-huge mode, we rely on memfd support */
1622 /* memfd supported, but hugetlbfs memfd may not be */
1623 if (!internal_config.no_hugetlbfs && !memfd_create_supported)
1627 /* fd_list not initialized? */
1628 if (fd_list[list_idx].len == 0)
1630 if (internal_config.single_file_segments) {
1631 size_t pgsz = mcfg->memsegs[list_idx].page_sz;
1633 /* segment not active? */
1634 if (fd_list[list_idx].memseg_list_fd < 0)
1636 *offset = pgsz * seg_idx;
1638 /* segment not active? */
1639 if (fd_list[list_idx].fds[seg_idx] < 0)
1647 eal_memalloc_init(void)
1649 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1650 if (rte_memseg_list_walk(secondary_msl_create_walk, NULL) < 0)
1652 if (rte_eal_process_type() == RTE_PROC_PRIMARY &&
1653 internal_config.in_memory) {
1654 int mfd_res = test_memfd_create();
1657 RTE_LOG(ERR, EAL, "Unable to check if memfd is supported\n");
1661 RTE_LOG(DEBUG, EAL, "Using memfd for anonymous memory\n");
1663 RTE_LOG(INFO, EAL, "Using memfd is not supported, falling back to anonymous hugepages\n");
1665 /* we only support single-file segments mode with in-memory mode
1666 * if we support hugetlbfs with memfd_create. this code will
1669 if (internal_config.single_file_segments &&
1671 RTE_LOG(ERR, EAL, "Single-file segments mode cannot be used without memfd support\n");
1674 /* this cannot ever happen but better safe than sorry */
1675 if (!anonymous_hugepages_supported) {
1676 RTE_LOG(ERR, EAL, "Using anonymous memory is not supported\n");
1681 /* initialize all of the fd lists */
1682 if (rte_memseg_list_walk(fd_list_create_walk, NULL))