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 /* A primary process is the only one creating these
333 * files. If there is a leftover that was not cleaned
334 * by clear_hugedir(), we must *now* make sure to drop
335 * the file or we will remap old stuff while the rest
336 * of the code is built on the assumption that a new
339 if (rte_eal_process_type() == RTE_PROC_PRIMARY &&
340 unlink(path) == -1 &&
342 RTE_LOG(DEBUG, EAL, "%s(): could not remove '%s': %s\n",
343 __func__, path, strerror(errno));
347 fd = open(path, O_CREAT | O_RDWR, 0600);
349 RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n",
350 __func__, strerror(errno));
353 /* take out a read lock */
354 if (lock(fd, LOCK_SH) < 0) {
355 RTE_LOG(ERR, EAL, "%s(): lock failed: %s\n",
356 __func__, strerror(errno));
360 fd_list[list_idx].fds[seg_idx] = fd;
367 resize_hugefile_in_memory(int fd, uint64_t fa_offset,
368 uint64_t page_sz, bool grow)
370 int flags = grow ? 0 : FALLOC_FL_PUNCH_HOLE |
374 /* grow or shrink the file */
375 ret = fallocate(fd, flags, fa_offset, page_sz);
378 RTE_LOG(DEBUG, EAL, "%s(): fallocate() failed: %s\n",
387 resize_hugefile_in_filesystem(int fd, uint64_t fa_offset, uint64_t page_sz,
393 if (fallocate_supported == 0) {
394 /* we cannot deallocate memory if fallocate() is not
395 * supported, and hugepage file is already locked at
396 * creation, so no further synchronization needed.
400 RTE_LOG(DEBUG, EAL, "%s(): fallocate not supported, not freeing page back to the system\n",
404 uint64_t new_size = fa_offset + page_sz;
405 uint64_t cur_size = get_file_size(fd);
407 /* fallocate isn't supported, fall back to ftruncate */
408 if (new_size > cur_size &&
409 ftruncate(fd, new_size) < 0) {
410 RTE_LOG(DEBUG, EAL, "%s(): ftruncate() failed: %s\n",
411 __func__, strerror(errno));
415 int flags = grow ? 0 : FALLOC_FL_PUNCH_HOLE |
420 * technically, it is perfectly safe for both primary
421 * and secondary to grow and shrink the page files:
422 * growing the file repeatedly has no effect because
423 * a page can only be allocated once, while mmap ensures
424 * that secondaries hold on to the page even after the
425 * page itself is removed from the filesystem.
427 * however, leaving growing/shrinking to the primary
428 * tends to expose bugs in fdlist page count handling,
429 * so leave this here just in case.
431 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
434 /* grow or shrink the file */
435 ret = fallocate(fd, flags, fa_offset, page_sz);
438 if (fallocate_supported == -1 &&
440 RTE_LOG(ERR, EAL, "%s(): fallocate() not supported, hugepage deallocation will be disabled\n",
443 fallocate_supported = 0;
445 RTE_LOG(DEBUG, EAL, "%s(): fallocate() failed: %s\n",
451 fallocate_supported = 1;
459 close_hugefile(int fd, char *path, int list_idx)
461 const struct internal_config *internal_conf =
462 eal_get_internal_configuration();
464 * primary process must unlink the file, but only when not in in-memory
465 * mode (as in that case there is no file to unlink).
467 if (!internal_conf->in_memory &&
468 rte_eal_process_type() == RTE_PROC_PRIMARY &&
470 RTE_LOG(ERR, EAL, "%s(): unlinking '%s' failed: %s\n",
471 __func__, path, strerror(errno));
474 fd_list[list_idx].memseg_list_fd = -1;
478 resize_hugefile(int fd, uint64_t fa_offset, uint64_t page_sz, bool grow)
480 /* in-memory mode is a special case, because we can be sure that
481 * fallocate() is supported.
483 const struct internal_config *internal_conf =
484 eal_get_internal_configuration();
486 if (internal_conf->in_memory)
487 return resize_hugefile_in_memory(fd, fa_offset,
490 return resize_hugefile_in_filesystem(fd, fa_offset, page_sz,
495 alloc_seg(struct rte_memseg *ms, void *addr, int socket_id,
496 struct hugepage_info *hi, unsigned int list_idx,
497 unsigned int seg_idx)
499 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
500 int cur_socket_id = 0;
511 const struct internal_config *internal_conf =
512 eal_get_internal_configuration();
514 alloc_sz = hi->hugepage_sz;
516 /* these are checked at init, but code analyzers don't know that */
517 if (internal_conf->in_memory && !anonymous_hugepages_supported) {
518 RTE_LOG(ERR, EAL, "Anonymous hugepages not supported, in-memory mode cannot allocate memory\n");
521 if (internal_conf->in_memory && !memfd_create_supported &&
522 internal_conf->single_file_segments) {
523 RTE_LOG(ERR, EAL, "Single-file segments are not supported without memfd support\n");
527 /* in-memory without memfd is a special case */
530 if (internal_conf->in_memory && !memfd_create_supported) {
531 const int in_memory_flags = MAP_HUGETLB | MAP_FIXED |
532 MAP_PRIVATE | MAP_ANONYMOUS;
535 pagesz_flag = pagesz_flags(alloc_sz);
537 mmap_flags = in_memory_flags | pagesz_flag;
539 /* single-file segments codepath will never be active
540 * here because in-memory mode is incompatible with the
541 * fallback path, and it's stopped at EAL initialization
546 /* takes out a read lock on segment or segment list */
547 fd = get_seg_fd(path, sizeof(path), hi, list_idx, seg_idx);
549 RTE_LOG(ERR, EAL, "Couldn't get fd on hugepage file\n");
553 if (internal_conf->single_file_segments) {
554 map_offset = seg_idx * alloc_sz;
555 ret = resize_hugefile(fd, map_offset, alloc_sz, true);
559 fd_list[list_idx].count++;
562 if (ftruncate(fd, alloc_sz) < 0) {
563 RTE_LOG(DEBUG, EAL, "%s(): ftruncate() failed: %s\n",
564 __func__, strerror(errno));
567 if (internal_conf->hugepage_unlink &&
568 !internal_conf->in_memory) {
570 RTE_LOG(DEBUG, EAL, "%s(): unlink() failed: %s\n",
571 __func__, strerror(errno));
576 mmap_flags = MAP_SHARED | MAP_POPULATE | MAP_FIXED;
580 * map the segment, and populate page tables, the kernel fills
581 * this segment with zeros if it's a new page.
583 va = mmap(addr, alloc_sz, PROT_READ | PROT_WRITE, mmap_flags, fd,
586 if (va == MAP_FAILED) {
587 RTE_LOG(DEBUG, EAL, "%s(): mmap() failed: %s\n", __func__,
589 /* mmap failed, but the previous region might have been
590 * unmapped anyway. try to remap it
595 RTE_LOG(DEBUG, EAL, "%s(): wrong mmap() address\n", __func__);
596 munmap(va, alloc_sz);
600 /* In linux, hugetlb limitations, like cgroup, are
601 * enforced at fault time instead of mmap(), even
602 * with the option of MAP_POPULATE. Kernel will send
603 * a SIGBUS signal. To avoid to be killed, save stack
604 * environment here, if SIGBUS happens, we can jump
607 if (huge_wrap_sigsetjmp()) {
608 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more hugepages of size %uMB\n",
609 (unsigned int)(alloc_sz >> 20));
613 /* we need to trigger a write to the page to enforce page fault and
614 * ensure that page is accessible to us, but we can't overwrite value
615 * that is already there, so read the old value, and write itback.
616 * kernel populates the page with zeroes initially.
618 *(volatile int *)addr = *(volatile int *)addr;
620 iova = rte_mem_virt2iova(addr);
621 if (iova == RTE_BAD_PHYS_ADDR) {
622 RTE_LOG(DEBUG, EAL, "%s(): can't get IOVA addr\n",
627 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
629 * If the kernel has been built without NUMA support, get_mempolicy()
630 * will return an error. If check_numa() returns false, memory
631 * allocation is not NUMA aware and the socket_id should not be
635 ret = get_mempolicy(&cur_socket_id, NULL, 0, addr,
636 MPOL_F_NODE | MPOL_F_ADDR);
638 RTE_LOG(DEBUG, EAL, "%s(): get_mempolicy: %s\n",
639 __func__, strerror(errno));
641 } else if (cur_socket_id != socket_id) {
643 "%s(): allocation happened on wrong socket (wanted %d, got %d)\n",
644 __func__, socket_id, cur_socket_id);
649 if (rte_socket_count() > 1)
650 RTE_LOG(DEBUG, EAL, "%s(): not checking hugepage NUMA node.\n",
655 ms->hugepage_sz = alloc_sz;
657 ms->nchannel = rte_memory_get_nchannel();
658 ms->nrank = rte_memory_get_nrank();
660 ms->socket_id = socket_id;
665 munmap(addr, alloc_sz);
667 flags = EAL_RESERVE_FORCE_ADDRESS;
668 new_addr = eal_get_virtual_area(addr, &alloc_sz, alloc_sz, 0, flags);
669 if (new_addr != addr) {
670 if (new_addr != NULL)
671 munmap(new_addr, alloc_sz);
672 /* we're leaving a hole in our virtual address space. if
673 * somebody else maps this hole now, we could accidentally
674 * override it in the future.
676 RTE_LOG(CRIT, EAL, "Can't mmap holes in our virtual address space\n");
678 /* roll back the ref count */
679 if (internal_conf->single_file_segments)
680 fd_list[list_idx].count--;
682 /* some codepaths will return negative fd, so exit early */
686 if (internal_conf->single_file_segments) {
687 resize_hugefile(fd, map_offset, alloc_sz, false);
688 /* ignore failure, can't make it any worse */
690 /* if refcount is at zero, close the file */
691 if (fd_list[list_idx].count == 0)
692 close_hugefile(fd, path, list_idx);
694 /* only remove file if we can take out a write lock */
695 if (internal_conf->hugepage_unlink == 0 &&
696 internal_conf->in_memory == 0 &&
697 lock(fd, LOCK_EX) == 1)
700 fd_list[list_idx].fds[seg_idx] = -1;
706 free_seg(struct rte_memseg *ms, struct hugepage_info *hi,
707 unsigned int list_idx, unsigned int seg_idx)
712 const struct internal_config *internal_conf =
713 eal_get_internal_configuration();
715 /* erase page data */
716 memset(ms->addr, 0, ms->len);
718 if (mmap(ms->addr, ms->len, PROT_NONE,
719 MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, 0) ==
721 RTE_LOG(DEBUG, EAL, "couldn't unmap page\n");
725 eal_mem_set_dump(ms->addr, ms->len, false);
727 /* if we're using anonymous hugepages, nothing to be done */
728 if (internal_conf->in_memory && !memfd_create_supported) {
729 memset(ms, 0, sizeof(*ms));
733 /* if we are not in single file segments mode, we're going to unmap the
734 * segment and thus drop the lock on original fd, but hugepage dir is
735 * now locked so we can take out another one without races.
737 fd = get_seg_fd(path, sizeof(path), hi, list_idx, seg_idx);
741 if (internal_conf->single_file_segments) {
742 map_offset = seg_idx * ms->len;
743 if (resize_hugefile(fd, map_offset, ms->len, false))
746 if (--(fd_list[list_idx].count) == 0)
747 close_hugefile(fd, path, list_idx);
751 /* if we're able to take out a write lock, we're the last one
752 * holding onto this page.
754 if (!internal_conf->in_memory && !internal_conf->hugepage_unlink) {
755 ret = lock(fd, LOCK_EX);
757 /* no one else is using this page */
762 /* closing fd will drop the lock */
764 fd_list[list_idx].fds[seg_idx] = -1;
767 memset(ms, 0, sizeof(*ms));
769 return ret < 0 ? -1 : 0;
772 struct alloc_walk_param {
773 struct hugepage_info *hi;
774 struct rte_memseg **ms;
776 unsigned int segs_allocated;
782 alloc_seg_walk(const struct rte_memseg_list *msl, void *arg)
784 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
785 struct alloc_walk_param *wa = arg;
786 struct rte_memseg_list *cur_msl;
788 int cur_idx, start_idx, j, dir_fd = -1;
789 unsigned int msl_idx, need, i;
790 const struct internal_config *internal_conf =
791 eal_get_internal_configuration();
793 if (msl->page_sz != wa->page_sz)
795 if (msl->socket_id != wa->socket)
798 page_sz = (size_t)msl->page_sz;
800 msl_idx = msl - mcfg->memsegs;
801 cur_msl = &mcfg->memsegs[msl_idx];
805 /* try finding space in memseg list */
807 /* if we require exact number of pages in a list, find them */
808 cur_idx = rte_fbarray_find_next_n_free(&cur_msl->memseg_arr, 0,
816 /* we don't require exact number of pages, so we're going to go
817 * for best-effort allocation. that means finding the biggest
818 * unused block, and going with that.
820 cur_idx = rte_fbarray_find_biggest_free(&cur_msl->memseg_arr,
825 /* adjust the size to possibly be smaller than original
826 * request, but do not allow it to be bigger.
828 cur_len = rte_fbarray_find_contig_free(&cur_msl->memseg_arr,
830 need = RTE_MIN(need, (unsigned int)cur_len);
833 /* do not allow any page allocations during the time we're allocating,
834 * because file creation and locking operations are not atomic,
835 * and we might be the first or the last ones to use a particular page,
836 * so we need to ensure atomicity of every operation.
838 * during init, we already hold a write lock, so don't try to take out
841 if (wa->hi->lock_descriptor == -1 && !internal_conf->in_memory) {
842 dir_fd = open(wa->hi->hugedir, O_RDONLY);
844 RTE_LOG(ERR, EAL, "%s(): Cannot open '%s': %s\n",
845 __func__, wa->hi->hugedir, strerror(errno));
848 /* blocking writelock */
849 if (flock(dir_fd, LOCK_EX)) {
850 RTE_LOG(ERR, EAL, "%s(): Cannot lock '%s': %s\n",
851 __func__, wa->hi->hugedir, strerror(errno));
857 for (i = 0; i < need; i++, cur_idx++) {
858 struct rte_memseg *cur;
861 cur = rte_fbarray_get(&cur_msl->memseg_arr, cur_idx);
862 map_addr = RTE_PTR_ADD(cur_msl->base_va,
865 if (alloc_seg(cur, map_addr, wa->socket, wa->hi,
867 RTE_LOG(DEBUG, EAL, "attempted to allocate %i segments, but only %i were allocated\n",
870 /* if exact number wasn't requested, stop */
875 for (j = start_idx; j < cur_idx; j++) {
876 struct rte_memseg *tmp;
877 struct rte_fbarray *arr =
878 &cur_msl->memseg_arr;
880 tmp = rte_fbarray_get(arr, j);
881 rte_fbarray_set_free(arr, j);
883 /* free_seg may attempt to create a file, which
886 if (free_seg(tmp, wa->hi, msl_idx, j))
887 RTE_LOG(DEBUG, EAL, "Cannot free page\n");
891 memset(wa->ms, 0, sizeof(*wa->ms) * wa->n_segs);
900 rte_fbarray_set_used(&cur_msl->memseg_arr, cur_idx);
903 wa->segs_allocated = i;
908 /* if we didn't allocate any segments, move on to the next list */
912 struct free_walk_param {
913 struct hugepage_info *hi;
914 struct rte_memseg *ms;
917 free_seg_walk(const struct rte_memseg_list *msl, void *arg)
919 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
920 struct rte_memseg_list *found_msl;
921 struct free_walk_param *wa = arg;
922 uintptr_t start_addr, end_addr;
923 int msl_idx, seg_idx, ret, dir_fd = -1;
924 const struct internal_config *internal_conf =
925 eal_get_internal_configuration();
927 start_addr = (uintptr_t) msl->base_va;
928 end_addr = start_addr + msl->len;
930 if ((uintptr_t)wa->ms->addr < start_addr ||
931 (uintptr_t)wa->ms->addr >= end_addr)
934 msl_idx = msl - mcfg->memsegs;
935 seg_idx = RTE_PTR_DIFF(wa->ms->addr, start_addr) / msl->page_sz;
938 found_msl = &mcfg->memsegs[msl_idx];
940 /* do not allow any page allocations during the time we're freeing,
941 * because file creation and locking operations are not atomic,
942 * and we might be the first or the last ones to use a particular page,
943 * so we need to ensure atomicity of every operation.
945 * during init, we already hold a write lock, so don't try to take out
948 if (wa->hi->lock_descriptor == -1 && !internal_conf->in_memory) {
949 dir_fd = open(wa->hi->hugedir, O_RDONLY);
951 RTE_LOG(ERR, EAL, "%s(): Cannot open '%s': %s\n",
952 __func__, wa->hi->hugedir, strerror(errno));
955 /* blocking writelock */
956 if (flock(dir_fd, LOCK_EX)) {
957 RTE_LOG(ERR, EAL, "%s(): Cannot lock '%s': %s\n",
958 __func__, wa->hi->hugedir, strerror(errno));
964 found_msl->version++;
966 rte_fbarray_set_free(&found_msl->memseg_arr, seg_idx);
968 ret = free_seg(wa->ms, wa->hi, msl_idx, seg_idx);
980 eal_memalloc_alloc_seg_bulk(struct rte_memseg **ms, int n_segs, size_t page_sz,
981 int socket, bool exact)
984 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
985 bool have_numa = false;
987 struct bitmask *oldmask;
989 struct alloc_walk_param wa;
990 struct hugepage_info *hi = NULL;
991 struct internal_config *internal_conf =
992 eal_get_internal_configuration();
994 memset(&wa, 0, sizeof(wa));
996 /* dynamic allocation not supported in legacy mode */
997 if (internal_conf->legacy_mem)
1000 for (i = 0; i < (int) RTE_DIM(internal_conf->hugepage_info); i++) {
1002 internal_conf->hugepage_info[i].hugepage_sz) {
1003 hi = &internal_conf->hugepage_info[i];
1008 RTE_LOG(ERR, EAL, "%s(): can't find relevant hugepage_info entry\n",
1013 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
1015 oldmask = numa_allocate_nodemask();
1016 prepare_numa(&oldpolicy, oldmask, socket);
1025 wa.page_sz = page_sz;
1027 wa.segs_allocated = 0;
1029 /* memalloc is locked, so it's safe to use thread-unsafe version */
1030 ret = rte_memseg_list_walk_thread_unsafe(alloc_seg_walk, &wa);
1032 RTE_LOG(ERR, EAL, "%s(): couldn't find suitable memseg_list\n",
1035 } else if (ret > 0) {
1036 ret = (int)wa.segs_allocated;
1039 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
1041 restore_numa(&oldpolicy, oldmask);
1047 eal_memalloc_alloc_seg(size_t page_sz, int socket)
1049 struct rte_memseg *ms;
1050 if (eal_memalloc_alloc_seg_bulk(&ms, 1, page_sz, socket, true) < 0)
1052 /* return pointer to newly allocated memseg */
1057 eal_memalloc_free_seg_bulk(struct rte_memseg **ms, int n_segs)
1060 struct internal_config *internal_conf =
1061 eal_get_internal_configuration();
1063 /* dynamic free not supported in legacy mode */
1064 if (internal_conf->legacy_mem)
1067 for (seg = 0; seg < n_segs; seg++) {
1068 struct rte_memseg *cur = ms[seg];
1069 struct hugepage_info *hi = NULL;
1070 struct free_walk_param wa;
1073 /* if this page is marked as unfreeable, fail */
1074 if (cur->flags & RTE_MEMSEG_FLAG_DO_NOT_FREE) {
1075 RTE_LOG(DEBUG, EAL, "Page is not allowed to be freed\n");
1080 memset(&wa, 0, sizeof(wa));
1082 for (i = 0; i < (int)RTE_DIM(internal_conf->hugepage_info);
1084 hi = &internal_conf->hugepage_info[i];
1085 if (cur->hugepage_sz == hi->hugepage_sz)
1088 if (i == (int)RTE_DIM(internal_conf->hugepage_info)) {
1089 RTE_LOG(ERR, EAL, "Can't find relevant hugepage_info entry\n");
1097 /* memalloc is locked, so it's safe to use thread-unsafe version
1099 walk_res = rte_memseg_list_walk_thread_unsafe(free_seg_walk,
1104 RTE_LOG(ERR, EAL, "Couldn't find memseg list\n");
1111 eal_memalloc_free_seg(struct rte_memseg *ms)
1113 const struct internal_config *internal_conf =
1114 eal_get_internal_configuration();
1116 /* dynamic free not supported in legacy mode */
1117 if (internal_conf->legacy_mem)
1120 return eal_memalloc_free_seg_bulk(&ms, 1);
1124 sync_chunk(struct rte_memseg_list *primary_msl,
1125 struct rte_memseg_list *local_msl, struct hugepage_info *hi,
1126 unsigned int msl_idx, bool used, int start, int end)
1128 struct rte_fbarray *l_arr, *p_arr;
1129 int i, ret, chunk_len, diff_len;
1131 l_arr = &local_msl->memseg_arr;
1132 p_arr = &primary_msl->memseg_arr;
1134 /* we need to aggregate allocations/deallocations into bigger chunks,
1135 * as we don't want to spam the user with per-page callbacks.
1137 * to avoid any potential issues, we also want to trigger
1138 * deallocation callbacks *before* we actually deallocate
1139 * memory, so that the user application could wrap up its use
1140 * before it goes away.
1143 chunk_len = end - start;
1145 /* find how many contiguous pages we can map/unmap for this chunk */
1147 rte_fbarray_find_contig_free(l_arr, start) :
1148 rte_fbarray_find_contig_used(l_arr, start);
1150 /* has to be at least one page */
1154 diff_len = RTE_MIN(chunk_len, diff_len);
1156 /* if we are freeing memory, notify the application */
1158 struct rte_memseg *ms;
1160 size_t len, page_sz;
1162 ms = rte_fbarray_get(l_arr, start);
1163 start_va = ms->addr;
1164 page_sz = (size_t)primary_msl->page_sz;
1165 len = page_sz * diff_len;
1167 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
1171 for (i = 0; i < diff_len; i++) {
1172 struct rte_memseg *p_ms, *l_ms;
1173 int seg_idx = start + i;
1175 l_ms = rte_fbarray_get(l_arr, seg_idx);
1176 p_ms = rte_fbarray_get(p_arr, seg_idx);
1178 if (l_ms == NULL || p_ms == NULL)
1182 ret = alloc_seg(l_ms, p_ms->addr,
1183 p_ms->socket_id, hi,
1187 rte_fbarray_set_used(l_arr, seg_idx);
1189 ret = free_seg(l_ms, hi, msl_idx, seg_idx);
1190 rte_fbarray_set_free(l_arr, seg_idx);
1196 /* if we just allocated memory, notify the application */
1198 struct rte_memseg *ms;
1200 size_t len, page_sz;
1202 ms = rte_fbarray_get(l_arr, start);
1203 start_va = ms->addr;
1204 page_sz = (size_t)primary_msl->page_sz;
1205 len = page_sz * diff_len;
1207 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC,
1211 /* calculate how much we can advance until next chunk */
1213 rte_fbarray_find_contig_used(l_arr, start) :
1214 rte_fbarray_find_contig_free(l_arr, start);
1215 ret = RTE_MIN(chunk_len, diff_len);
1221 sync_status(struct rte_memseg_list *primary_msl,
1222 struct rte_memseg_list *local_msl, struct hugepage_info *hi,
1223 unsigned int msl_idx, bool used)
1225 struct rte_fbarray *l_arr, *p_arr;
1226 int p_idx, l_chunk_len, p_chunk_len, ret;
1229 /* this is a little bit tricky, but the basic idea is - walk both lists
1230 * and spot any places where there are discrepancies. walking both lists
1231 * and noting discrepancies in a single go is a hard problem, so we do
1232 * it in two passes - first we spot any places where allocated segments
1233 * mismatch (i.e. ensure that everything that's allocated in the primary
1234 * is also allocated in the secondary), and then we do it by looking at
1235 * free segments instead.
1237 * we also need to aggregate changes into chunks, as we have to call
1238 * callbacks per allocation, not per page.
1240 l_arr = &local_msl->memseg_arr;
1241 p_arr = &primary_msl->memseg_arr;
1244 p_idx = rte_fbarray_find_next_used(p_arr, 0);
1246 p_idx = rte_fbarray_find_next_free(p_arr, 0);
1248 while (p_idx >= 0) {
1249 int next_chunk_search_idx;
1252 p_chunk_len = rte_fbarray_find_contig_used(p_arr,
1254 l_chunk_len = rte_fbarray_find_contig_used(l_arr,
1257 p_chunk_len = rte_fbarray_find_contig_free(p_arr,
1259 l_chunk_len = rte_fbarray_find_contig_free(l_arr,
1262 /* best case scenario - no differences (or bigger, which will be
1263 * fixed during next iteration), look for next chunk
1265 if (l_chunk_len >= p_chunk_len) {
1266 next_chunk_search_idx = p_idx + p_chunk_len;
1270 /* if both chunks start at the same point, skip parts we know
1271 * are identical, and sync the rest. each call to sync_chunk
1272 * will only sync contiguous segments, so we need to call this
1273 * until we are sure there are no more differences in this
1276 start = p_idx + l_chunk_len;
1277 end = p_idx + p_chunk_len;
1279 ret = sync_chunk(primary_msl, local_msl, hi, msl_idx,
1282 } while (start < end && ret >= 0);
1283 /* if ret is negative, something went wrong */
1287 next_chunk_search_idx = p_idx + p_chunk_len;
1289 /* skip to end of this chunk */
1291 p_idx = rte_fbarray_find_next_used(p_arr,
1292 next_chunk_search_idx);
1294 p_idx = rte_fbarray_find_next_free(p_arr,
1295 next_chunk_search_idx);
1302 sync_existing(struct rte_memseg_list *primary_msl,
1303 struct rte_memseg_list *local_msl, struct hugepage_info *hi,
1304 unsigned int msl_idx)
1308 /* do not allow any page allocations during the time we're allocating,
1309 * because file creation and locking operations are not atomic,
1310 * and we might be the first or the last ones to use a particular page,
1311 * so we need to ensure atomicity of every operation.
1313 dir_fd = open(hi->hugedir, O_RDONLY);
1315 RTE_LOG(ERR, EAL, "%s(): Cannot open '%s': %s\n", __func__,
1316 hi->hugedir, strerror(errno));
1319 /* blocking writelock */
1320 if (flock(dir_fd, LOCK_EX)) {
1321 RTE_LOG(ERR, EAL, "%s(): Cannot lock '%s': %s\n", __func__,
1322 hi->hugedir, strerror(errno));
1327 /* ensure all allocated space is the same in both lists */
1328 ret = sync_status(primary_msl, local_msl, hi, msl_idx, true);
1332 /* ensure all unallocated space is the same in both lists */
1333 ret = sync_status(primary_msl, local_msl, hi, msl_idx, false);
1337 /* update version number */
1338 local_msl->version = primary_msl->version;
1349 sync_walk(const struct rte_memseg_list *msl, void *arg __rte_unused)
1351 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1352 struct rte_memseg_list *primary_msl, *local_msl;
1353 struct hugepage_info *hi = NULL;
1356 struct internal_config *internal_conf =
1357 eal_get_internal_configuration();
1362 msl_idx = msl - mcfg->memsegs;
1363 primary_msl = &mcfg->memsegs[msl_idx];
1364 local_msl = &local_memsegs[msl_idx];
1366 for (i = 0; i < RTE_DIM(internal_conf->hugepage_info); i++) {
1368 internal_conf->hugepage_info[i].hugepage_sz;
1369 uint64_t msl_sz = primary_msl->page_sz;
1370 if (msl_sz == cur_sz) {
1371 hi = &internal_conf->hugepage_info[i];
1376 RTE_LOG(ERR, EAL, "Can't find relevant hugepage_info entry\n");
1380 /* if versions don't match, synchronize everything */
1381 if (local_msl->version != primary_msl->version &&
1382 sync_existing(primary_msl, local_msl, hi, msl_idx))
1389 eal_memalloc_sync_with_primary(void)
1391 /* nothing to be done in primary */
1392 if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1395 /* memalloc is locked, so it's safe to call thread-unsafe version */
1396 if (rte_memseg_list_walk_thread_unsafe(sync_walk, NULL))
1402 secondary_msl_create_walk(const struct rte_memseg_list *msl,
1403 void *arg __rte_unused)
1405 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1406 struct rte_memseg_list *primary_msl, *local_msl;
1407 char name[PATH_MAX];
1413 msl_idx = msl - mcfg->memsegs;
1414 primary_msl = &mcfg->memsegs[msl_idx];
1415 local_msl = &local_memsegs[msl_idx];
1417 /* create distinct fbarrays for each secondary */
1418 snprintf(name, RTE_FBARRAY_NAME_LEN, "%s_%i",
1419 primary_msl->memseg_arr.name, getpid());
1421 ret = rte_fbarray_init(&local_msl->memseg_arr, name,
1422 primary_msl->memseg_arr.len,
1423 primary_msl->memseg_arr.elt_sz);
1425 RTE_LOG(ERR, EAL, "Cannot initialize local memory map\n");
1428 local_msl->base_va = primary_msl->base_va;
1429 local_msl->len = primary_msl->len;
1435 secondary_msl_destroy_walk(const struct rte_memseg_list *msl,
1436 void *arg __rte_unused)
1438 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1439 struct rte_memseg_list *local_msl;
1445 msl_idx = msl - mcfg->memsegs;
1446 local_msl = &local_memsegs[msl_idx];
1448 ret = rte_fbarray_destroy(&local_msl->memseg_arr);
1450 RTE_LOG(ERR, EAL, "Cannot destroy local memory map\n");
1453 local_msl->base_va = NULL;
1460 alloc_list(int list_idx, int len)
1464 const struct internal_config *internal_conf =
1465 eal_get_internal_configuration();
1467 /* single-file segments mode does not need fd list */
1468 if (!internal_conf->single_file_segments) {
1469 /* ensure we have space to store fd per each possible segment */
1470 data = malloc(sizeof(int) * len);
1472 RTE_LOG(ERR, EAL, "Unable to allocate space for file descriptors\n");
1475 /* set all fd's as invalid */
1476 for (i = 0; i < len; i++)
1478 fd_list[list_idx].fds = data;
1479 fd_list[list_idx].len = len;
1481 fd_list[list_idx].fds = NULL;
1482 fd_list[list_idx].len = 0;
1485 fd_list[list_idx].count = 0;
1486 fd_list[list_idx].memseg_list_fd = -1;
1492 destroy_list(int list_idx)
1494 const struct internal_config *internal_conf =
1495 eal_get_internal_configuration();
1497 /* single-file segments mode does not need fd list */
1498 if (!internal_conf->single_file_segments) {
1499 int *fds = fd_list[list_idx].fds;
1501 /* go through each fd and ensure it's closed */
1502 for (i = 0; i < fd_list[list_idx].len; i++) {
1509 fd_list[list_idx].fds = NULL;
1510 fd_list[list_idx].len = 0;
1511 } else if (fd_list[list_idx].memseg_list_fd >= 0) {
1512 close(fd_list[list_idx].memseg_list_fd);
1513 fd_list[list_idx].count = 0;
1514 fd_list[list_idx].memseg_list_fd = -1;
1520 fd_list_create_walk(const struct rte_memseg_list *msl,
1521 void *arg __rte_unused)
1523 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1530 msl_idx = msl - mcfg->memsegs;
1531 len = msl->memseg_arr.len;
1533 return alloc_list(msl_idx, len);
1537 fd_list_destroy_walk(const struct rte_memseg_list *msl, void *arg __rte_unused)
1539 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1545 msl_idx = msl - mcfg->memsegs;
1547 return destroy_list(msl_idx);
1551 eal_memalloc_set_seg_fd(int list_idx, int seg_idx, int fd)
1553 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1554 const struct internal_config *internal_conf =
1555 eal_get_internal_configuration();
1557 /* single file segments mode doesn't support individual segment fd's */
1558 if (internal_conf->single_file_segments)
1561 /* if list is not allocated, allocate it */
1562 if (fd_list[list_idx].len == 0) {
1563 int len = mcfg->memsegs[list_idx].memseg_arr.len;
1565 if (alloc_list(list_idx, len) < 0)
1568 fd_list[list_idx].fds[seg_idx] = fd;
1574 eal_memalloc_set_seg_list_fd(int list_idx, int fd)
1576 const struct internal_config *internal_conf =
1577 eal_get_internal_configuration();
1579 /* non-single file segment mode doesn't support segment list fd's */
1580 if (!internal_conf->single_file_segments)
1583 fd_list[list_idx].memseg_list_fd = fd;
1589 eal_memalloc_get_seg_fd(int list_idx, int seg_idx)
1592 const struct internal_config *internal_conf =
1593 eal_get_internal_configuration();
1595 if (internal_conf->in_memory || internal_conf->no_hugetlbfs) {
1596 #ifndef MEMFD_SUPPORTED
1597 /* in in-memory or no-huge mode, we rely on memfd support */
1600 /* memfd supported, but hugetlbfs memfd may not be */
1601 if (!internal_conf->no_hugetlbfs && !memfd_create_supported)
1605 if (internal_conf->single_file_segments) {
1606 fd = fd_list[list_idx].memseg_list_fd;
1607 } else if (fd_list[list_idx].len == 0) {
1608 /* list not initialized */
1611 fd = fd_list[list_idx].fds[seg_idx];
1619 test_memfd_create(void)
1621 #ifdef MEMFD_SUPPORTED
1622 const struct internal_config *internal_conf =
1623 eal_get_internal_configuration();
1625 for (i = 0; i < internal_conf->num_hugepage_sizes; i++) {
1626 uint64_t pagesz = internal_conf->hugepage_info[i].hugepage_sz;
1627 int pagesz_flag = pagesz_flags(pagesz);
1630 flags = pagesz_flag | RTE_MFD_HUGETLB;
1631 int fd = memfd_create("test", flags);
1633 /* we failed - let memalloc know this isn't working */
1634 if (errno == EINVAL) {
1635 memfd_create_supported = 0;
1636 return 0; /* not supported */
1639 /* we got other error - something's wrong */
1640 return -1; /* error */
1643 return 1; /* supported */
1646 return 0; /* not supported */
1650 eal_memalloc_get_seg_fd_offset(int list_idx, int seg_idx, size_t *offset)
1652 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1653 const struct internal_config *internal_conf =
1654 eal_get_internal_configuration();
1656 if (internal_conf->in_memory || internal_conf->no_hugetlbfs) {
1657 #ifndef MEMFD_SUPPORTED
1658 /* in in-memory or no-huge mode, we rely on memfd support */
1661 /* memfd supported, but hugetlbfs memfd may not be */
1662 if (!internal_conf->no_hugetlbfs && !memfd_create_supported)
1666 if (internal_conf->single_file_segments) {
1667 size_t pgsz = mcfg->memsegs[list_idx].page_sz;
1669 /* segment not active? */
1670 if (fd_list[list_idx].memseg_list_fd < 0)
1672 *offset = pgsz * seg_idx;
1674 /* fd_list not initialized? */
1675 if (fd_list[list_idx].len == 0)
1678 /* segment not active? */
1679 if (fd_list[list_idx].fds[seg_idx] < 0)
1687 eal_memalloc_cleanup(void)
1689 /* close all remaining fd's - these are per-process, so it's safe */
1690 if (rte_memseg_list_walk_thread_unsafe(fd_list_destroy_walk, NULL))
1693 /* destroy the shadow page table if we're a secondary process */
1694 if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1697 if (rte_memseg_list_walk_thread_unsafe(secondary_msl_destroy_walk,
1705 eal_memalloc_init(void)
1707 const struct internal_config *internal_conf =
1708 eal_get_internal_configuration();
1710 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1711 if (rte_memseg_list_walk(secondary_msl_create_walk, NULL) < 0)
1713 if (rte_eal_process_type() == RTE_PROC_PRIMARY &&
1714 internal_conf->in_memory) {
1715 int mfd_res = test_memfd_create();
1718 RTE_LOG(ERR, EAL, "Unable to check if memfd is supported\n");
1722 RTE_LOG(DEBUG, EAL, "Using memfd for anonymous memory\n");
1724 RTE_LOG(INFO, EAL, "Using memfd is not supported, falling back to anonymous hugepages\n");
1726 /* we only support single-file segments mode with in-memory mode
1727 * if we support hugetlbfs with memfd_create. this code will
1730 if (internal_conf->single_file_segments &&
1732 RTE_LOG(ERR, EAL, "Single-file segments mode cannot be used without memfd support\n");
1735 /* this cannot ever happen but better safe than sorry */
1736 if (!anonymous_hugepages_supported) {
1737 RTE_LOG(ERR, EAL, "Using anonymous memory is not supported\n");
1742 /* initialize all of the fd lists */
1743 if (rte_memseg_list_walk(fd_list_create_walk, NULL))