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 "malloc_heap.h"
30 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
31 * pointer to the mmap'd area and keep *size unmodified. Else, retry
32 * with a smaller zone: decrease *size by hugepage_sz until it reaches
33 * 0. In this case, return NULL. Note: this function returns an address
34 * which is a multiple of hugepage size.
37 #define MEMSEG_LIST_FMT "memseg-%" PRIu64 "k-%i-%i"
39 static void *next_baseaddr;
40 static uint64_t system_page_sz;
44 * Linux kernel uses a really high address as starting address for serving
45 * mmaps calls. If there exists addressing limitations and IOVA mode is VA,
46 * this starting address is likely too high for those devices. However, it
47 * is possible to use a lower address in the process virtual address space
48 * as with 64 bits there is a lot of available space.
50 * Current known limitations are 39 or 40 bits. Setting the starting address
51 * at 4GB implies there are 508GB or 1020GB for mapping the available
52 * hugepages. This is likely enough for most systems, although a device with
53 * addressing limitations should call rte_mem_check_dma_mask for ensuring all
54 * memory is within supported range.
56 static uint64_t baseaddr = 0x100000000;
59 #define MAX_MMAP_WITH_DEFINED_ADDR_TRIES 5
61 eal_get_virtual_area(void *requested_addr, size_t *size,
62 size_t page_sz, int flags, int mmap_flags)
64 bool addr_is_hint, allow_shrink, unmap, no_align;
66 void *mapped_addr, *aligned_addr;
69 if (system_page_sz == 0)
70 system_page_sz = sysconf(_SC_PAGESIZE);
72 mmap_flags |= MAP_PRIVATE | MAP_ANONYMOUS;
74 RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
76 addr_is_hint = (flags & EAL_VIRTUAL_AREA_ADDR_IS_HINT) > 0;
77 allow_shrink = (flags & EAL_VIRTUAL_AREA_ALLOW_SHRINK) > 0;
78 unmap = (flags & EAL_VIRTUAL_AREA_UNMAP) > 0;
80 if (next_baseaddr == NULL && internal_config.base_virtaddr != 0 &&
81 rte_eal_process_type() == RTE_PROC_PRIMARY)
82 next_baseaddr = (void *) internal_config.base_virtaddr;
85 if (next_baseaddr == NULL && internal_config.base_virtaddr == 0 &&
86 rte_eal_process_type() == RTE_PROC_PRIMARY)
87 next_baseaddr = (void *) baseaddr;
89 if (requested_addr == NULL && next_baseaddr != NULL) {
90 requested_addr = next_baseaddr;
91 requested_addr = RTE_PTR_ALIGN(requested_addr, page_sz);
95 /* we don't need alignment of resulting pointer in the following cases:
97 * 1. page size is equal to system size
98 * 2. we have a requested address, and it is page-aligned, and we will
99 * be discarding the address if we get a different one.
101 * for all other cases, alignment is potentially necessary.
103 no_align = (requested_addr != NULL &&
104 requested_addr == RTE_PTR_ALIGN(requested_addr, page_sz) &&
106 page_sz == system_page_sz;
109 map_sz = no_align ? *size : *size + page_sz;
110 if (map_sz > SIZE_MAX) {
111 RTE_LOG(ERR, EAL, "Map size too big\n");
116 mapped_addr = mmap(requested_addr, (size_t)map_sz, PROT_READ,
118 if (mapped_addr == MAP_FAILED && allow_shrink)
121 if (mapped_addr != MAP_FAILED && addr_is_hint &&
122 mapped_addr != requested_addr) {
124 next_baseaddr = RTE_PTR_ADD(next_baseaddr, page_sz);
125 if (try <= MAX_MMAP_WITH_DEFINED_ADDR_TRIES) {
126 /* hint was not used. Try with another offset */
127 munmap(mapped_addr, map_sz);
128 mapped_addr = MAP_FAILED;
129 requested_addr = next_baseaddr;
132 } while ((allow_shrink || addr_is_hint) &&
133 mapped_addr == MAP_FAILED && *size > 0);
135 /* align resulting address - if map failed, we will ignore the value
136 * anyway, so no need to add additional checks.
138 aligned_addr = no_align ? mapped_addr :
139 RTE_PTR_ALIGN(mapped_addr, page_sz);
142 RTE_LOG(ERR, EAL, "Cannot get a virtual area of any size: %s\n",
146 } else if (mapped_addr == MAP_FAILED) {
147 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
149 /* pass errno up the call chain */
152 } else if (requested_addr != NULL && !addr_is_hint &&
153 aligned_addr != requested_addr) {
154 RTE_LOG(ERR, EAL, "Cannot get a virtual area at requested address: %p (got %p)\n",
155 requested_addr, aligned_addr);
156 munmap(mapped_addr, map_sz);
157 rte_errno = EADDRNOTAVAIL;
159 } else if (requested_addr != NULL && addr_is_hint &&
160 aligned_addr != requested_addr) {
161 RTE_LOG(WARNING, EAL, "WARNING! Base virtual address hint (%p != %p) not respected!\n",
162 requested_addr, aligned_addr);
163 RTE_LOG(WARNING, EAL, " This may cause issues with mapping memory into secondary processes\n");
164 } else if (next_baseaddr != NULL) {
165 next_baseaddr = RTE_PTR_ADD(aligned_addr, *size);
168 RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
169 aligned_addr, *size);
172 munmap(mapped_addr, map_sz);
173 } else if (!no_align) {
174 void *map_end, *aligned_end;
175 size_t before_len, after_len;
177 /* when we reserve space with alignment, we add alignment to
178 * mapping size. On 32-bit, if 1GB alignment was requested, this
179 * would waste 1GB of address space, which is a luxury we cannot
180 * afford. so, if alignment was performed, check if any unneeded
181 * address space can be unmapped back.
184 map_end = RTE_PTR_ADD(mapped_addr, (size_t)map_sz);
185 aligned_end = RTE_PTR_ADD(aligned_addr, *size);
187 /* unmap space before aligned mmap address */
188 before_len = RTE_PTR_DIFF(aligned_addr, mapped_addr);
190 munmap(mapped_addr, before_len);
192 /* unmap space after aligned end mmap address */
193 after_len = RTE_PTR_DIFF(map_end, aligned_end);
195 munmap(aligned_end, after_len);
201 static struct rte_memseg *
202 virt2memseg(const void *addr, const struct rte_memseg_list *msl)
204 const struct rte_fbarray *arr;
211 /* a memseg list was specified, check if it's the right one */
212 start = msl->base_va;
213 end = RTE_PTR_ADD(start, msl->len);
215 if (addr < start || addr >= end)
218 /* now, calculate index */
219 arr = &msl->memseg_arr;
220 ms_idx = RTE_PTR_DIFF(addr, msl->base_va) / msl->page_sz;
221 return rte_fbarray_get(arr, ms_idx);
224 static struct rte_memseg_list *
225 virt2memseg_list(const void *addr)
227 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
228 struct rte_memseg_list *msl;
231 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
233 msl = &mcfg->memsegs[msl_idx];
235 start = msl->base_va;
236 end = RTE_PTR_ADD(start, msl->len);
237 if (addr >= start && addr < end)
240 /* if we didn't find our memseg list */
241 if (msl_idx == RTE_MAX_MEMSEG_LISTS)
246 struct rte_memseg_list *
247 rte_mem_virt2memseg_list(const void *addr)
249 return virt2memseg_list(addr);
257 find_virt(const struct rte_memseg_list *msl __rte_unused,
258 const struct rte_memseg *ms, void *arg)
260 struct virtiova *vi = arg;
261 if (vi->iova >= ms->iova && vi->iova < (ms->iova + ms->len)) {
262 size_t offset = vi->iova - ms->iova;
263 vi->virt = RTE_PTR_ADD(ms->addr, offset);
270 find_virt_legacy(const struct rte_memseg_list *msl __rte_unused,
271 const struct rte_memseg *ms, size_t len, void *arg)
273 struct virtiova *vi = arg;
274 if (vi->iova >= ms->iova && vi->iova < (ms->iova + len)) {
275 size_t offset = vi->iova - ms->iova;
276 vi->virt = RTE_PTR_ADD(ms->addr, offset);
284 rte_mem_iova2virt(rte_iova_t iova)
288 memset(&vi, 0, sizeof(vi));
291 /* for legacy mem, we can get away with scanning VA-contiguous segments,
292 * as we know they are PA-contiguous as well
294 if (internal_config.legacy_mem)
295 rte_memseg_contig_walk(find_virt_legacy, &vi);
297 rte_memseg_walk(find_virt, &vi);
303 rte_mem_virt2memseg(const void *addr, const struct rte_memseg_list *msl)
305 return virt2memseg(addr, msl != NULL ? msl :
306 rte_mem_virt2memseg_list(addr));
310 physmem_size(const struct rte_memseg_list *msl, void *arg)
312 uint64_t *total_len = arg;
317 *total_len += msl->memseg_arr.count * msl->page_sz;
322 /* get the total size of memory */
324 rte_eal_get_physmem_size(void)
326 uint64_t total_len = 0;
328 rte_memseg_list_walk(physmem_size, &total_len);
334 dump_memseg(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
337 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
338 int msl_idx, ms_idx, fd;
341 msl_idx = msl - mcfg->memsegs;
342 if (msl_idx < 0 || msl_idx >= RTE_MAX_MEMSEG_LISTS)
345 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
349 fd = eal_memalloc_get_seg_fd(msl_idx, ms_idx);
350 fprintf(f, "Segment %i-%i: IOVA:0x%"PRIx64", len:%zu, "
351 "virt:%p, socket_id:%"PRId32", "
352 "hugepage_sz:%"PRIu64", nchannel:%"PRIx32", "
353 "nrank:%"PRIx32" fd:%i\n",
368 * Defining here because declared in rte_memory.h, but the actual implementation
369 * is in eal_common_memalloc.c, like all other memalloc internals.
372 rte_mem_event_callback_register(const char *name, rte_mem_event_callback_t clb,
375 /* FreeBSD boots with legacy mem enabled by default */
376 if (internal_config.legacy_mem) {
377 RTE_LOG(DEBUG, EAL, "Registering mem event callbacks not supported\n");
381 return eal_memalloc_mem_event_callback_register(name, clb, arg);
385 rte_mem_event_callback_unregister(const char *name, void *arg)
387 /* FreeBSD boots with legacy mem enabled by default */
388 if (internal_config.legacy_mem) {
389 RTE_LOG(DEBUG, EAL, "Registering mem event callbacks not supported\n");
393 return eal_memalloc_mem_event_callback_unregister(name, arg);
397 rte_mem_alloc_validator_register(const char *name,
398 rte_mem_alloc_validator_t clb, int socket_id, size_t limit)
400 /* FreeBSD boots with legacy mem enabled by default */
401 if (internal_config.legacy_mem) {
402 RTE_LOG(DEBUG, EAL, "Registering mem alloc validators not supported\n");
406 return eal_memalloc_mem_alloc_validator_register(name, clb, socket_id,
411 rte_mem_alloc_validator_unregister(const char *name, int socket_id)
413 /* FreeBSD boots with legacy mem enabled by default */
414 if (internal_config.legacy_mem) {
415 RTE_LOG(DEBUG, EAL, "Registering mem alloc validators not supported\n");
419 return eal_memalloc_mem_alloc_validator_unregister(name, socket_id);
422 /* Dump the physical memory layout on console */
424 rte_dump_physmem_layout(FILE *f)
426 rte_memseg_walk(dump_memseg, f);
430 check_iova(const struct rte_memseg_list *msl __rte_unused,
431 const struct rte_memseg *ms, void *arg)
433 uint64_t *mask = arg;
436 /* higher address within segment */
437 iova = (ms->iova + ms->len) - 1;
441 RTE_LOG(DEBUG, EAL, "memseg iova %"PRIx64", len %zx, out of range\n",
444 RTE_LOG(DEBUG, EAL, "\tusing dma mask %"PRIx64"\n", *mask);
448 #define MAX_DMA_MASK_BITS 63
450 /* check memseg iovas are within the required range based on dma mask */
452 check_dma_mask(uint8_t maskbits, bool thread_unsafe)
454 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
458 /* Sanity check. We only check width can be managed with 64 bits
459 * variables. Indeed any higher value is likely wrong. */
460 if (maskbits > MAX_DMA_MASK_BITS) {
461 RTE_LOG(ERR, EAL, "wrong dma mask size %u (Max: %u)\n",
462 maskbits, MAX_DMA_MASK_BITS);
466 /* create dma mask */
467 mask = ~((1ULL << maskbits) - 1);
470 ret = rte_memseg_walk_thread_unsafe(check_iova, &mask);
472 ret = rte_memseg_walk(check_iova, &mask);
476 * Dma mask precludes hugepage usage.
477 * This device can not be used and we do not need to keep
483 * we need to keep the more restricted maskbit for checking
484 * potential dynamic memory allocation in the future.
486 mcfg->dma_maskbits = mcfg->dma_maskbits == 0 ? maskbits :
487 RTE_MIN(mcfg->dma_maskbits, maskbits);
493 rte_mem_check_dma_mask(uint8_t maskbits)
495 return check_dma_mask(maskbits, false);
499 rte_mem_check_dma_mask_thread_unsafe(uint8_t maskbits)
501 return check_dma_mask(maskbits, true);
505 * Set dma mask to use when memory initialization is done.
507 * This function should ONLY be used by code executed before the memory
508 * initialization. PMDs should use rte_mem_check_dma_mask if addressing
509 * limitations by the device.
512 rte_mem_set_dma_mask(uint8_t maskbits)
514 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
516 mcfg->dma_maskbits = mcfg->dma_maskbits == 0 ? maskbits :
517 RTE_MIN(mcfg->dma_maskbits, maskbits);
520 /* return the number of memory channels */
521 unsigned rte_memory_get_nchannel(void)
523 return rte_eal_get_configuration()->mem_config->nchannel;
526 /* return the number of memory rank */
527 unsigned rte_memory_get_nrank(void)
529 return rte_eal_get_configuration()->mem_config->nrank;
533 rte_eal_memdevice_init(void)
535 struct rte_config *config;
537 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
540 config = rte_eal_get_configuration();
541 config->mem_config->nchannel = internal_config.force_nchannel;
542 config->mem_config->nrank = internal_config.force_nrank;
547 /* Lock page in physical memory and prevent from swapping. */
549 rte_mem_lock_page(const void *virt)
551 unsigned long virtual = (unsigned long)virt;
552 int page_size = getpagesize();
553 unsigned long aligned = (virtual & ~(page_size - 1));
554 return mlock((void *)aligned, page_size);
558 rte_memseg_contig_walk_thread_unsafe(rte_memseg_contig_walk_t func, void *arg)
560 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
561 int i, ms_idx, ret = 0;
563 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
564 struct rte_memseg_list *msl = &mcfg->memsegs[i];
565 const struct rte_memseg *ms;
566 struct rte_fbarray *arr;
568 if (msl->memseg_arr.count == 0)
571 arr = &msl->memseg_arr;
573 ms_idx = rte_fbarray_find_next_used(arr, 0);
574 while (ms_idx >= 0) {
578 ms = rte_fbarray_get(arr, ms_idx);
580 /* find how many more segments there are, starting with
583 n_segs = rte_fbarray_find_contig_used(arr, ms_idx);
584 len = n_segs * msl->page_sz;
586 ret = func(msl, ms, len, arg);
589 ms_idx = rte_fbarray_find_next_used(arr,
597 rte_memseg_contig_walk(rte_memseg_contig_walk_t func, void *arg)
601 /* do not allow allocations/frees/init while we iterate */
602 rte_mcfg_mem_read_lock();
603 ret = rte_memseg_contig_walk_thread_unsafe(func, arg);
604 rte_mcfg_mem_read_unlock();
610 rte_memseg_walk_thread_unsafe(rte_memseg_walk_t func, void *arg)
612 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
613 int i, ms_idx, ret = 0;
615 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
616 struct rte_memseg_list *msl = &mcfg->memsegs[i];
617 const struct rte_memseg *ms;
618 struct rte_fbarray *arr;
620 if (msl->memseg_arr.count == 0)
623 arr = &msl->memseg_arr;
625 ms_idx = rte_fbarray_find_next_used(arr, 0);
626 while (ms_idx >= 0) {
627 ms = rte_fbarray_get(arr, ms_idx);
628 ret = func(msl, ms, arg);
631 ms_idx = rte_fbarray_find_next_used(arr, ms_idx + 1);
638 rte_memseg_walk(rte_memseg_walk_t func, void *arg)
642 /* do not allow allocations/frees/init while we iterate */
643 rte_mcfg_mem_read_lock();
644 ret = rte_memseg_walk_thread_unsafe(func, arg);
645 rte_mcfg_mem_read_unlock();
651 rte_memseg_list_walk_thread_unsafe(rte_memseg_list_walk_t func, void *arg)
653 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
656 for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
657 struct rte_memseg_list *msl = &mcfg->memsegs[i];
659 if (msl->base_va == NULL)
662 ret = func(msl, arg);
670 rte_memseg_list_walk(rte_memseg_list_walk_t func, void *arg)
674 /* do not allow allocations/frees/init while we iterate */
675 rte_mcfg_mem_read_lock();
676 ret = rte_memseg_list_walk_thread_unsafe(func, arg);
677 rte_mcfg_mem_read_unlock();
683 rte_memseg_get_fd_thread_unsafe(const struct rte_memseg *ms)
685 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
686 struct rte_memseg_list *msl;
687 struct rte_fbarray *arr;
688 int msl_idx, seg_idx, ret;
695 msl = rte_mem_virt2memseg_list(ms->addr);
700 arr = &msl->memseg_arr;
702 msl_idx = msl - mcfg->memsegs;
703 seg_idx = rte_fbarray_find_idx(arr, ms);
705 if (!rte_fbarray_is_used(arr, seg_idx)) {
710 /* segment fd API is not supported for external segments */
716 ret = eal_memalloc_get_seg_fd(msl_idx, seg_idx);
725 rte_memseg_get_fd(const struct rte_memseg *ms)
729 rte_mcfg_mem_read_lock();
730 ret = rte_memseg_get_fd_thread_unsafe(ms);
731 rte_mcfg_mem_read_unlock();
737 rte_memseg_get_fd_offset_thread_unsafe(const struct rte_memseg *ms,
740 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
741 struct rte_memseg_list *msl;
742 struct rte_fbarray *arr;
743 int msl_idx, seg_idx, ret;
745 if (ms == NULL || offset == NULL) {
750 msl = rte_mem_virt2memseg_list(ms->addr);
755 arr = &msl->memseg_arr;
757 msl_idx = msl - mcfg->memsegs;
758 seg_idx = rte_fbarray_find_idx(arr, ms);
760 if (!rte_fbarray_is_used(arr, seg_idx)) {
765 /* segment fd API is not supported for external segments */
771 ret = eal_memalloc_get_seg_fd_offset(msl_idx, seg_idx, offset);
780 rte_memseg_get_fd_offset(const struct rte_memseg *ms, size_t *offset)
784 rte_mcfg_mem_read_lock();
785 ret = rte_memseg_get_fd_offset_thread_unsafe(ms, offset);
786 rte_mcfg_mem_read_unlock();
792 rte_extmem_register(void *va_addr, size_t len, rte_iova_t iova_addrs[],
793 unsigned int n_pages, size_t page_sz)
795 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
796 unsigned int socket_id, n;
799 if (va_addr == NULL || page_sz == 0 || len == 0 ||
800 !rte_is_power_of_2(page_sz) ||
801 RTE_ALIGN(len, page_sz) != len ||
802 ((len / page_sz) != n_pages && iova_addrs != NULL) ||
803 !rte_is_aligned(va_addr, page_sz)) {
807 rte_mcfg_mem_write_lock();
809 /* make sure the segment doesn't already exist */
810 if (malloc_heap_find_external_seg(va_addr, len) != NULL) {
816 /* get next available socket ID */
817 socket_id = mcfg->next_socket_id;
818 if (socket_id > INT32_MAX) {
819 RTE_LOG(ERR, EAL, "Cannot assign new socket ID's\n");
825 /* we can create a new memseg */
827 if (malloc_heap_create_external_seg(va_addr, iova_addrs, n,
828 page_sz, "extmem", socket_id) == NULL) {
833 /* memseg list successfully created - increment next socket ID */
834 mcfg->next_socket_id++;
836 rte_mcfg_mem_write_unlock();
841 rte_extmem_unregister(void *va_addr, size_t len)
843 struct rte_memseg_list *msl;
846 if (va_addr == NULL || len == 0) {
850 rte_mcfg_mem_write_lock();
852 /* find our segment */
853 msl = malloc_heap_find_external_seg(va_addr, len);
860 ret = malloc_heap_destroy_external_seg(msl);
862 rte_mcfg_mem_write_unlock();
867 sync_memory(void *va_addr, size_t len, bool attach)
869 struct rte_memseg_list *msl;
872 if (va_addr == NULL || len == 0) {
876 rte_mcfg_mem_write_lock();
878 /* find our segment */
879 msl = malloc_heap_find_external_seg(va_addr, len);
886 ret = rte_fbarray_attach(&msl->memseg_arr);
888 ret = rte_fbarray_detach(&msl->memseg_arr);
891 rte_mcfg_mem_write_unlock();
896 rte_extmem_attach(void *va_addr, size_t len)
898 return sync_memory(va_addr, len, true);
902 rte_extmem_detach(void *va_addr, size_t len)
904 return sync_memory(va_addr, len, false);
907 /* init memory subsystem */
909 rte_eal_memory_init(void)
911 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
913 RTE_LOG(DEBUG, EAL, "Setting up physically contiguous memory...\n");
918 /* lock mem hotplug here, to prevent races while we init */
919 rte_mcfg_mem_read_lock();
921 if (rte_eal_memseg_init() < 0)
924 if (eal_memalloc_init() < 0)
927 retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
928 rte_eal_hugepage_init() :
929 rte_eal_hugepage_attach();
933 if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)
938 rte_mcfg_mem_read_unlock();