1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
15 #include <sys/queue.h>
17 #include <rte_fbarray.h>
18 #include <rte_memory.h>
20 #include <rte_eal_memconfig.h>
21 #include <rte_errno.h>
24 #include "eal_memalloc.h"
25 #include "eal_private.h"
26 #include "eal_internal_cfg.h"
27 #include "eal_memcfg.h"
28 #include "malloc_heap.h"
31 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
32 * pointer to the mmap'd area and keep *size unmodified. Else, retry
33 * with a smaller zone: decrease *size by hugepage_sz until it reaches
34 * 0. In this case, return NULL. Note: this function returns an address
35 * which is a multiple of hugepage size.
38 #define MEMSEG_LIST_FMT "memseg-%" PRIu64 "k-%i-%i"
40 static void *next_baseaddr;
41 static uint64_t system_page_sz;
43 #ifdef RTE_EXEC_ENV_LINUX
44 #define RTE_DONTDUMP MADV_DONTDUMP
45 #elif defined RTE_EXEC_ENV_FREEBSD
46 #define RTE_DONTDUMP MADV_NOCORE
48 #error "madvise doesn't support this OS"
51 #define MAX_MMAP_WITH_DEFINED_ADDR_TRIES 5
53 eal_get_virtual_area(void *requested_addr, size_t *size,
54 size_t page_sz, int flags, int mmap_flags)
56 bool addr_is_hint, allow_shrink, unmap, no_align;
58 void *mapped_addr, *aligned_addr;
61 if (system_page_sz == 0)
62 system_page_sz = sysconf(_SC_PAGESIZE);
64 mmap_flags |= MAP_PRIVATE | MAP_ANONYMOUS;
66 RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
68 addr_is_hint = (flags & EAL_VIRTUAL_AREA_ADDR_IS_HINT) > 0;
69 allow_shrink = (flags & EAL_VIRTUAL_AREA_ALLOW_SHRINK) > 0;
70 unmap = (flags & EAL_VIRTUAL_AREA_UNMAP) > 0;
72 if (next_baseaddr == NULL && internal_config.base_virtaddr != 0 &&
73 rte_eal_process_type() == RTE_PROC_PRIMARY)
74 next_baseaddr = (void *) internal_config.base_virtaddr;
77 if (next_baseaddr == NULL && internal_config.base_virtaddr == 0 &&
78 rte_eal_process_type() == RTE_PROC_PRIMARY)
79 next_baseaddr = (void *) eal_get_baseaddr();
81 if (requested_addr == NULL && next_baseaddr != NULL) {
82 requested_addr = next_baseaddr;
83 requested_addr = RTE_PTR_ALIGN(requested_addr, page_sz);
87 /* we don't need alignment of resulting pointer in the following cases:
89 * 1. page size is equal to system size
90 * 2. we have a requested address, and it is page-aligned, and we will
91 * be discarding the address if we get a different one.
93 * for all other cases, alignment is potentially necessary.
95 no_align = (requested_addr != NULL &&
96 requested_addr == RTE_PTR_ALIGN(requested_addr, page_sz) &&
98 page_sz == system_page_sz;
101 map_sz = no_align ? *size : *size + page_sz;
102 if (map_sz > SIZE_MAX) {
103 RTE_LOG(ERR, EAL, "Map size too big\n");
108 mapped_addr = mmap(requested_addr, (size_t)map_sz, PROT_NONE,
110 if (mapped_addr == MAP_FAILED && allow_shrink)
113 if (mapped_addr != MAP_FAILED && addr_is_hint &&
114 mapped_addr != requested_addr) {
116 next_baseaddr = RTE_PTR_ADD(next_baseaddr, page_sz);
117 if (try <= MAX_MMAP_WITH_DEFINED_ADDR_TRIES) {
118 /* hint was not used. Try with another offset */
119 munmap(mapped_addr, map_sz);
120 mapped_addr = MAP_FAILED;
121 requested_addr = next_baseaddr;
124 } while ((allow_shrink || addr_is_hint) &&
125 mapped_addr == MAP_FAILED && *size > 0);
127 /* align resulting address - if map failed, we will ignore the value
128 * anyway, so no need to add additional checks.
130 aligned_addr = no_align ? mapped_addr :
131 RTE_PTR_ALIGN(mapped_addr, page_sz);
134 RTE_LOG(ERR, EAL, "Cannot get a virtual area of any size: %s\n",
138 } else if (mapped_addr == MAP_FAILED) {
139 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
141 /* pass errno up the call chain */
144 } else if (requested_addr != NULL && !addr_is_hint &&
145 aligned_addr != requested_addr) {
146 RTE_LOG(ERR, EAL, "Cannot get a virtual area at requested address: %p (got %p)\n",
147 requested_addr, aligned_addr);
148 munmap(mapped_addr, map_sz);
149 rte_errno = EADDRNOTAVAIL;
151 } else if (requested_addr != NULL && addr_is_hint &&
152 aligned_addr != requested_addr) {
153 RTE_LOG(WARNING, EAL, "WARNING! Base virtual address hint (%p != %p) not respected!\n",
154 requested_addr, aligned_addr);
155 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory into secondary processes\n");
156 } else if (next_baseaddr != NULL) {
157 next_baseaddr = RTE_PTR_ADD(aligned_addr, *size);
160 RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
161 aligned_addr, *size);
164 munmap(mapped_addr, map_sz);
165 } else if (!no_align) {
166 void *map_end, *aligned_end;
167 size_t before_len, after_len;
169 /* when we reserve space with alignment, we add alignment to
170 * mapping size. On 32-bit, if 1GB alignment was requested, this
171 * would waste 1GB of address space, which is a luxury we cannot
172 * afford. so, if alignment was performed, check if any unneeded
173 * address space can be unmapped back.
176 map_end = RTE_PTR_ADD(mapped_addr, (size_t)map_sz);
177 aligned_end = RTE_PTR_ADD(aligned_addr, *size);
179 /* unmap space before aligned mmap address */
180 before_len = RTE_PTR_DIFF(aligned_addr, mapped_addr);
182 munmap(mapped_addr, before_len);
184 /* unmap space after aligned end mmap address */
185 after_len = RTE_PTR_DIFF(map_end, aligned_end);
187 munmap(aligned_end, after_len);
191 /* Exclude these pages from a core dump. */
192 if (madvise(aligned_addr, *size, RTE_DONTDUMP) != 0)
193 RTE_LOG(DEBUG, EAL, "madvise failed: %s\n",
200 static struct rte_memseg *
201 virt2memseg(const void *addr, const struct rte_memseg_list *msl)
203 const struct rte_fbarray *arr;
210 /* a memseg list was specified, check if it's the right one */
211 start = msl->base_va;
212 end = RTE_PTR_ADD(start, msl->len);
214 if (addr < start || addr >= end)
217 /* now, calculate index */
218 arr = &msl->memseg_arr;
219 ms_idx = RTE_PTR_DIFF(addr, msl->base_va) / msl->page_sz;
220 return rte_fbarray_get(arr, ms_idx);
223 static struct rte_memseg_list *
224 virt2memseg_list(const void *addr)
226 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
227 struct rte_memseg_list *msl;
230 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
232 msl = &mcfg->memsegs[msl_idx];
234 start = msl->base_va;
235 end = RTE_PTR_ADD(start, msl->len);
236 if (addr >= start && addr < end)
239 /* if we didn't find our memseg list */
240 if (msl_idx == RTE_MAX_MEMSEG_LISTS)
245 struct rte_memseg_list *
246 rte_mem_virt2memseg_list(const void *addr)
248 return virt2memseg_list(addr);
256 find_virt(const struct rte_memseg_list *msl __rte_unused,
257 const struct rte_memseg *ms, void *arg)
259 struct virtiova *vi = arg;
260 if (vi->iova >= ms->iova && vi->iova < (ms->iova + ms->len)) {
261 size_t offset = vi->iova - ms->iova;
262 vi->virt = RTE_PTR_ADD(ms->addr, offset);
269 find_virt_legacy(const struct rte_memseg_list *msl __rte_unused,
270 const struct rte_memseg *ms, size_t len, void *arg)
272 struct virtiova *vi = arg;
273 if (vi->iova >= ms->iova && vi->iova < (ms->iova + len)) {
274 size_t offset = vi->iova - ms->iova;
275 vi->virt = RTE_PTR_ADD(ms->addr, offset);
283 rte_mem_iova2virt(rte_iova_t iova)
287 memset(&vi, 0, sizeof(vi));
290 /* for legacy mem, we can get away with scanning VA-contiguous segments,
291 * as we know they are PA-contiguous as well
293 if (internal_config.legacy_mem)
294 rte_memseg_contig_walk(find_virt_legacy, &vi);
296 rte_memseg_walk(find_virt, &vi);
302 rte_mem_virt2memseg(const void *addr, const struct rte_memseg_list *msl)
304 return virt2memseg(addr, msl != NULL ? msl :
305 rte_mem_virt2memseg_list(addr));
309 physmem_size(const struct rte_memseg_list *msl, void *arg)
311 uint64_t *total_len = arg;
316 *total_len += msl->memseg_arr.count * msl->page_sz;
321 /* get the total size of memory */
323 rte_eal_get_physmem_size(void)
325 uint64_t total_len = 0;
327 rte_memseg_list_walk(physmem_size, &total_len);
333 dump_memseg(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
336 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
337 int msl_idx, ms_idx, fd;
340 msl_idx = msl - mcfg->memsegs;
341 if (msl_idx < 0 || msl_idx >= RTE_MAX_MEMSEG_LISTS)
344 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
348 fd = eal_memalloc_get_seg_fd(msl_idx, ms_idx);
349 fprintf(f, "Segment %i-%i: IOVA:0x%"PRIx64", len:%zu, "
350 "virt:%p, socket_id:%"PRId32", "
351 "hugepage_sz:%"PRIu64", nchannel:%"PRIx32", "
352 "nrank:%"PRIx32" fd:%i\n",
367 * Defining here because declared in rte_memory.h, but the actual implementation
368 * is in eal_common_memalloc.c, like all other memalloc internals.
371 rte_mem_event_callback_register(const char *name, rte_mem_event_callback_t clb,
374 /* FreeBSD boots with legacy mem enabled by default */
375 if (internal_config.legacy_mem) {
376 RTE_LOG(DEBUG, EAL, "Registering mem event callbacks not supported\n");
380 return eal_memalloc_mem_event_callback_register(name, clb, arg);
384 rte_mem_event_callback_unregister(const char *name, void *arg)
386 /* FreeBSD boots with legacy mem enabled by default */
387 if (internal_config.legacy_mem) {
388 RTE_LOG(DEBUG, EAL, "Registering mem event callbacks not supported\n");
392 return eal_memalloc_mem_event_callback_unregister(name, arg);
396 rte_mem_alloc_validator_register(const char *name,
397 rte_mem_alloc_validator_t clb, int socket_id, size_t limit)
399 /* FreeBSD boots with legacy mem enabled by default */
400 if (internal_config.legacy_mem) {
401 RTE_LOG(DEBUG, EAL, "Registering mem alloc validators not supported\n");
405 return eal_memalloc_mem_alloc_validator_register(name, clb, socket_id,
410 rte_mem_alloc_validator_unregister(const char *name, int socket_id)
412 /* FreeBSD boots with legacy mem enabled by default */
413 if (internal_config.legacy_mem) {
414 RTE_LOG(DEBUG, EAL, "Registering mem alloc validators not supported\n");
418 return eal_memalloc_mem_alloc_validator_unregister(name, socket_id);
421 /* Dump the physical memory layout on console */
423 rte_dump_physmem_layout(FILE *f)
425 rte_memseg_walk(dump_memseg, f);
429 check_iova(const struct rte_memseg_list *msl __rte_unused,
430 const struct rte_memseg *ms, void *arg)
432 uint64_t *mask = arg;
435 /* higher address within segment */
436 iova = (ms->iova + ms->len) - 1;
440 RTE_LOG(DEBUG, EAL, "memseg iova %"PRIx64", len %zx, out of range\n",
443 RTE_LOG(DEBUG, EAL, "\tusing dma mask %"PRIx64"\n", *mask);
447 #define MAX_DMA_MASK_BITS 63
449 /* check memseg iovas are within the required range based on dma mask */
451 check_dma_mask(uint8_t maskbits, bool thread_unsafe)
453 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
457 /* Sanity check. We only check width can be managed with 64 bits
458 * variables. Indeed any higher value is likely wrong. */
459 if (maskbits > MAX_DMA_MASK_BITS) {
460 RTE_LOG(ERR, EAL, "wrong dma mask size %u (Max: %u)\n",
461 maskbits, MAX_DMA_MASK_BITS);
465 /* create dma mask */
466 mask = ~((1ULL << maskbits) - 1);
469 ret = rte_memseg_walk_thread_unsafe(check_iova, &mask);
471 ret = rte_memseg_walk(check_iova, &mask);
475 * Dma mask precludes hugepage usage.
476 * This device can not be used and we do not need to keep
482 * we need to keep the more restricted maskbit for checking
483 * potential dynamic memory allocation in the future.
485 mcfg->dma_maskbits = mcfg->dma_maskbits == 0 ? maskbits :
486 RTE_MIN(mcfg->dma_maskbits, maskbits);
492 rte_mem_check_dma_mask(uint8_t maskbits)
494 return check_dma_mask(maskbits, false);
498 rte_mem_check_dma_mask_thread_unsafe(uint8_t maskbits)
500 return check_dma_mask(maskbits, true);
504 * Set dma mask to use when memory initialization is done.
506 * This function should ONLY be used by code executed before the memory
507 * initialization. PMDs should use rte_mem_check_dma_mask if addressing
508 * limitations by the device.
511 rte_mem_set_dma_mask(uint8_t maskbits)
513 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
515 mcfg->dma_maskbits = mcfg->dma_maskbits == 0 ? maskbits :
516 RTE_MIN(mcfg->dma_maskbits, maskbits);
519 /* return the number of memory channels */
520 unsigned rte_memory_get_nchannel(void)
522 return rte_eal_get_configuration()->mem_config->nchannel;
525 /* return the number of memory rank */
526 unsigned rte_memory_get_nrank(void)
528 return rte_eal_get_configuration()->mem_config->nrank;
532 rte_eal_memdevice_init(void)
534 struct rte_config *config;
536 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
539 config = rte_eal_get_configuration();
540 config->mem_config->nchannel = internal_config.force_nchannel;
541 config->mem_config->nrank = internal_config.force_nrank;
546 /* Lock page in physical memory and prevent from swapping. */
548 rte_mem_lock_page(const void *virt)
550 unsigned long virtual = (unsigned long)virt;
551 int page_size = getpagesize();
552 unsigned long aligned = (virtual & ~(page_size - 1));
553 return mlock((void *)aligned, page_size);
557 rte_memseg_contig_walk_thread_unsafe(rte_memseg_contig_walk_t func, void *arg)
559 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
560 int i, ms_idx, ret = 0;
562 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
563 struct rte_memseg_list *msl = &mcfg->memsegs[i];
564 const struct rte_memseg *ms;
565 struct rte_fbarray *arr;
567 if (msl->memseg_arr.count == 0)
570 arr = &msl->memseg_arr;
572 ms_idx = rte_fbarray_find_next_used(arr, 0);
573 while (ms_idx >= 0) {
577 ms = rte_fbarray_get(arr, ms_idx);
579 /* find how many more segments there are, starting with
582 n_segs = rte_fbarray_find_contig_used(arr, ms_idx);
583 len = n_segs * msl->page_sz;
585 ret = func(msl, ms, len, arg);
588 ms_idx = rte_fbarray_find_next_used(arr,
596 rte_memseg_contig_walk(rte_memseg_contig_walk_t func, void *arg)
600 /* do not allow allocations/frees/init while we iterate */
601 rte_mcfg_mem_read_lock();
602 ret = rte_memseg_contig_walk_thread_unsafe(func, arg);
603 rte_mcfg_mem_read_unlock();
609 rte_memseg_walk_thread_unsafe(rte_memseg_walk_t func, void *arg)
611 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
612 int i, ms_idx, ret = 0;
614 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
615 struct rte_memseg_list *msl = &mcfg->memsegs[i];
616 const struct rte_memseg *ms;
617 struct rte_fbarray *arr;
619 if (msl->memseg_arr.count == 0)
622 arr = &msl->memseg_arr;
624 ms_idx = rte_fbarray_find_next_used(arr, 0);
625 while (ms_idx >= 0) {
626 ms = rte_fbarray_get(arr, ms_idx);
627 ret = func(msl, ms, arg);
630 ms_idx = rte_fbarray_find_next_used(arr, ms_idx + 1);
637 rte_memseg_walk(rte_memseg_walk_t func, void *arg)
641 /* do not allow allocations/frees/init while we iterate */
642 rte_mcfg_mem_read_lock();
643 ret = rte_memseg_walk_thread_unsafe(func, arg);
644 rte_mcfg_mem_read_unlock();
650 rte_memseg_list_walk_thread_unsafe(rte_memseg_list_walk_t func, void *arg)
652 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
655 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
656 struct rte_memseg_list *msl = &mcfg->memsegs[i];
658 if (msl->base_va == NULL)
661 ret = func(msl, arg);
669 rte_memseg_list_walk(rte_memseg_list_walk_t func, void *arg)
673 /* do not allow allocations/frees/init while we iterate */
674 rte_mcfg_mem_read_lock();
675 ret = rte_memseg_list_walk_thread_unsafe(func, arg);
676 rte_mcfg_mem_read_unlock();
682 rte_memseg_get_fd_thread_unsafe(const struct rte_memseg *ms)
684 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
685 struct rte_memseg_list *msl;
686 struct rte_fbarray *arr;
687 int msl_idx, seg_idx, ret;
694 msl = rte_mem_virt2memseg_list(ms->addr);
699 arr = &msl->memseg_arr;
701 msl_idx = msl - mcfg->memsegs;
702 seg_idx = rte_fbarray_find_idx(arr, ms);
704 if (!rte_fbarray_is_used(arr, seg_idx)) {
709 /* segment fd API is not supported for external segments */
715 ret = eal_memalloc_get_seg_fd(msl_idx, seg_idx);
724 rte_memseg_get_fd(const struct rte_memseg *ms)
728 rte_mcfg_mem_read_lock();
729 ret = rte_memseg_get_fd_thread_unsafe(ms);
730 rte_mcfg_mem_read_unlock();
736 rte_memseg_get_fd_offset_thread_unsafe(const struct rte_memseg *ms,
739 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
740 struct rte_memseg_list *msl;
741 struct rte_fbarray *arr;
742 int msl_idx, seg_idx, ret;
744 if (ms == NULL || offset == NULL) {
749 msl = rte_mem_virt2memseg_list(ms->addr);
754 arr = &msl->memseg_arr;
756 msl_idx = msl - mcfg->memsegs;
757 seg_idx = rte_fbarray_find_idx(arr, ms);
759 if (!rte_fbarray_is_used(arr, seg_idx)) {
764 /* segment fd API is not supported for external segments */
770 ret = eal_memalloc_get_seg_fd_offset(msl_idx, seg_idx, offset);
779 rte_memseg_get_fd_offset(const struct rte_memseg *ms, size_t *offset)
783 rte_mcfg_mem_read_lock();
784 ret = rte_memseg_get_fd_offset_thread_unsafe(ms, offset);
785 rte_mcfg_mem_read_unlock();
791 rte_extmem_register(void *va_addr, size_t len, rte_iova_t iova_addrs[],
792 unsigned int n_pages, size_t page_sz)
794 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
795 unsigned int socket_id, n;
798 if (va_addr == NULL || page_sz == 0 || len == 0 ||
799 !rte_is_power_of_2(page_sz) ||
800 RTE_ALIGN(len, page_sz) != len ||
801 ((len / page_sz) != n_pages && iova_addrs != NULL) ||
802 !rte_is_aligned(va_addr, page_sz)) {
806 rte_mcfg_mem_write_lock();
808 /* make sure the segment doesn't already exist */
809 if (malloc_heap_find_external_seg(va_addr, len) != NULL) {
815 /* get next available socket ID */
816 socket_id = mcfg->next_socket_id;
817 if (socket_id > INT32_MAX) {
818 RTE_LOG(ERR, EAL, "Cannot assign new socket ID's\n");
824 /* we can create a new memseg */
826 if (malloc_heap_create_external_seg(va_addr, iova_addrs, n,
827 page_sz, "extmem", socket_id) == NULL) {
832 /* memseg list successfully created - increment next socket ID */
833 mcfg->next_socket_id++;
835 rte_mcfg_mem_write_unlock();
840 rte_extmem_unregister(void *va_addr, size_t len)
842 struct rte_memseg_list *msl;
845 if (va_addr == NULL || len == 0) {
849 rte_mcfg_mem_write_lock();
851 /* find our segment */
852 msl = malloc_heap_find_external_seg(va_addr, len);
859 ret = malloc_heap_destroy_external_seg(msl);
861 rte_mcfg_mem_write_unlock();
866 sync_memory(void *va_addr, size_t len, bool attach)
868 struct rte_memseg_list *msl;
871 if (va_addr == NULL || len == 0) {
875 rte_mcfg_mem_write_lock();
877 /* find our segment */
878 msl = malloc_heap_find_external_seg(va_addr, len);
885 ret = rte_fbarray_attach(&msl->memseg_arr);
887 ret = rte_fbarray_detach(&msl->memseg_arr);
890 rte_mcfg_mem_write_unlock();
895 rte_extmem_attach(void *va_addr, size_t len)
897 return sync_memory(va_addr, len, true);
901 rte_extmem_detach(void *va_addr, size_t len)
903 return sync_memory(va_addr, len, false);
906 /* init memory subsystem */
908 rte_eal_memory_init(void)
910 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
912 RTE_LOG(DEBUG, EAL, "Setting up physically contiguous memory...\n");
917 /* lock mem hotplug here, to prevent races while we init */
918 rte_mcfg_mem_read_lock();
920 if (rte_eal_memseg_init() < 0)
923 if (eal_memalloc_init() < 0)
926 retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
927 rte_eal_hugepage_init() :
928 rte_eal_hugepage_attach();
932 if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)
937 rte_mcfg_mem_read_unlock();