1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
7 #include <sys/sysctl.h>
14 #include <rte_errno.h>
16 #include <rte_string_fns.h>
17 #include "eal_private.h"
18 #include "eal_internal_cfg.h"
19 #include "eal_filesystem.h"
20 #include "eal_memcfg.h"
22 #define EAL_PAGE_SIZE (sysconf(_SC_PAGESIZE))
25 * Get physical address of any mapped virtual address in the current process.
28 rte_mem_virt2phy(const void *virtaddr)
30 /* XXX not implemented. This function is only used by
31 * rte_mempool_virt2iova() when hugepages are disabled. */
36 rte_mem_virt2iova(const void *virtaddr)
38 return rte_mem_virt2phy(virtaddr);
42 rte_eal_hugepage_init(void)
44 struct rte_mem_config *mcfg;
45 uint64_t total_mem = 0;
47 unsigned int i, j, seg_idx = 0;
49 /* get pointer to global configuration */
50 mcfg = rte_eal_get_configuration()->mem_config;
52 /* for debug purposes, hugetlbfs can be disabled */
53 if (internal_config.no_hugetlbfs) {
54 struct rte_memseg_list *msl;
55 struct rte_fbarray *arr;
56 struct rte_memseg *ms;
60 /* create a memseg list */
61 msl = &mcfg->memsegs[0];
63 page_sz = RTE_PGSIZE_4K;
64 n_segs = internal_config.memory / page_sz;
66 if (rte_fbarray_init(&msl->memseg_arr, "nohugemem", n_segs,
67 sizeof(struct rte_memseg))) {
68 RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
72 addr = mmap(NULL, internal_config.memory,
73 PROT_READ | PROT_WRITE,
74 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
75 if (addr == MAP_FAILED) {
76 RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
81 msl->page_sz = page_sz;
82 msl->len = internal_config.memory;
85 /* populate memsegs. each memseg is 1 page long */
86 for (cur_seg = 0; cur_seg < n_segs; cur_seg++) {
87 arr = &msl->memseg_arr;
89 ms = rte_fbarray_get(arr, cur_seg);
90 if (rte_eal_iova_mode() == RTE_IOVA_VA)
91 ms->iova = (uintptr_t)addr;
93 ms->iova = RTE_BAD_IOVA;
95 ms->hugepage_sz = page_sz;
99 rte_fbarray_set_used(arr, cur_seg);
101 addr = RTE_PTR_ADD(addr, page_sz);
106 /* map all hugepages and sort them */
107 for (i = 0; i < internal_config.num_hugepage_sizes; i ++){
108 struct hugepage_info *hpi;
109 rte_iova_t prev_end = 0;
110 int prev_ms_idx = -1;
111 uint64_t page_sz, mem_needed;
112 unsigned int n_pages, max_pages;
114 hpi = &internal_config.hugepage_info[i];
115 page_sz = hpi->hugepage_sz;
116 max_pages = hpi->num_pages[0];
117 mem_needed = RTE_ALIGN_CEIL(internal_config.memory - total_mem,
120 n_pages = RTE_MIN(mem_needed / page_sz, max_pages);
122 for (j = 0; j < n_pages; j++) {
123 struct rte_memseg_list *msl;
124 struct rte_fbarray *arr;
125 struct rte_memseg *seg;
129 size_t sysctl_size = sizeof(physaddr);
130 char physaddr_str[64];
133 /* first, check if this segment is IOVA-adjacent to
136 snprintf(physaddr_str, sizeof(physaddr_str),
137 "hw.contigmem.physaddr.%d", j);
138 error = sysctlbyname(physaddr_str, &physaddr,
139 &sysctl_size, NULL, 0);
141 RTE_LOG(ERR, EAL, "Failed to get physical addr for buffer %u "
142 "from %s\n", j, hpi->hugedir);
146 is_adjacent = prev_end != 0 && physaddr == prev_end;
147 prev_end = physaddr + hpi->hugepage_sz;
149 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS;
151 bool empty, need_hole;
152 msl = &mcfg->memsegs[msl_idx];
153 arr = &msl->memseg_arr;
155 if (msl->page_sz != page_sz)
158 empty = arr->count == 0;
160 /* we need a hole if this isn't an empty memseg
161 * list, and if previous segment was not
162 * adjacent to current one.
164 need_hole = !empty && !is_adjacent;
166 /* we need 1, plus hole if not adjacent */
167 ms_idx = rte_fbarray_find_next_n_free(arr,
168 0, 1 + (need_hole ? 1 : 0));
170 /* memseg list is full? */
174 if (need_hole && prev_ms_idx == ms_idx - 1)
176 prev_ms_idx = ms_idx;
180 if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
181 RTE_LOG(ERR, EAL, "Could not find space for memseg. Please increase %s and/or %s in configuration.\n",
182 RTE_STR(CONFIG_RTE_MAX_MEMSEG_PER_TYPE),
183 RTE_STR(CONFIG_RTE_MAX_MEM_PER_TYPE));
186 arr = &msl->memseg_arr;
187 seg = rte_fbarray_get(arr, ms_idx);
189 addr = RTE_PTR_ADD(msl->base_va,
190 (size_t)msl->page_sz * ms_idx);
192 /* address is already mapped in memseg list, so using
193 * MAP_FIXED here is safe.
195 addr = mmap(addr, page_sz, PROT_READ|PROT_WRITE,
196 MAP_SHARED | MAP_FIXED,
197 hpi->lock_descriptor,
199 if (addr == MAP_FAILED) {
200 RTE_LOG(ERR, EAL, "Failed to mmap buffer %u from %s\n",
206 seg->iova = physaddr;
207 seg->hugepage_sz = page_sz;
209 seg->nchannel = mcfg->nchannel;
210 seg->nrank = mcfg->nrank;
213 rte_fbarray_set_used(arr, ms_idx);
215 RTE_LOG(INFO, EAL, "Mapped memory segment %u @ %p: physaddr:0x%"
217 seg_idx++, addr, physaddr, page_sz);
219 total_mem += seg->len;
221 if (total_mem >= internal_config.memory)
224 if (total_mem < internal_config.memory) {
225 RTE_LOG(ERR, EAL, "Couldn't reserve requested memory, "
226 "requested: %" PRIu64 "M "
227 "available: %" PRIu64 "M\n",
228 internal_config.memory >> 20, total_mem >> 20);
234 struct attach_walk_args {
239 attach_segment(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
242 struct attach_walk_args *wa = arg;
248 addr = mmap(ms->addr, ms->len, PROT_READ | PROT_WRITE,
249 MAP_SHARED | MAP_FIXED, wa->fd_hugepage,
250 wa->seg_idx * EAL_PAGE_SIZE);
251 if (addr == MAP_FAILED || addr != ms->addr)
259 rte_eal_hugepage_attach(void)
261 const struct hugepage_info *hpi;
262 int fd_hugepage = -1;
265 hpi = &internal_config.hugepage_info[0];
267 for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
268 const struct hugepage_info *cur_hpi = &hpi[i];
269 struct attach_walk_args wa;
271 memset(&wa, 0, sizeof(wa));
273 /* Obtain a file descriptor for contiguous memory */
274 fd_hugepage = open(cur_hpi->hugedir, O_RDWR);
275 if (fd_hugepage < 0) {
276 RTE_LOG(ERR, EAL, "Could not open %s\n",
280 wa.fd_hugepage = fd_hugepage;
283 /* Map the contiguous memory into each memory segment */
284 if (rte_memseg_walk(attach_segment, &wa) < 0) {
285 RTE_LOG(ERR, EAL, "Failed to mmap buffer %u from %s\n",
286 wa.seg_idx, cur_hpi->hugedir);
294 /* hugepage_info is no longer required */
298 if (fd_hugepage >= 0)
304 rte_eal_using_phys_addrs(void)
310 get_mem_amount(uint64_t page_sz, uint64_t max_mem)
312 uint64_t area_sz, max_pages;
314 /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */
315 max_pages = RTE_MAX_MEMSEG_PER_LIST;
316 max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem);
318 area_sz = RTE_MIN(page_sz * max_pages, max_mem);
320 /* make sure the list isn't smaller than the page size */
321 area_sz = RTE_MAX(area_sz, page_sz);
323 return RTE_ALIGN(area_sz, page_sz);
326 #define MEMSEG_LIST_FMT "memseg-%" PRIu64 "k-%i-%i"
328 alloc_memseg_list(struct rte_memseg_list *msl, uint64_t page_sz,
329 int n_segs, int socket_id, int type_msl_idx)
331 char name[RTE_FBARRAY_NAME_LEN];
333 snprintf(name, sizeof(name), MEMSEG_LIST_FMT, page_sz >> 10, socket_id,
335 if (rte_fbarray_init(&msl->memseg_arr, name, n_segs,
336 sizeof(struct rte_memseg))) {
337 RTE_LOG(ERR, EAL, "Cannot allocate memseg list: %s\n",
338 rte_strerror(rte_errno));
342 msl->page_sz = page_sz;
343 msl->socket_id = socket_id;
346 RTE_LOG(DEBUG, EAL, "Memseg list allocated: 0x%zxkB at socket %i\n",
347 (size_t)page_sz >> 10, socket_id);
353 alloc_va_space(struct rte_memseg_list *msl)
360 #ifdef RTE_ARCH_PPC_64
361 flags |= MAP_HUGETLB;
364 page_sz = msl->page_sz;
365 mem_sz = page_sz * msl->memseg_arr.len;
367 addr = eal_get_virtual_area(msl->base_va, &mem_sz, page_sz, 0, flags);
369 if (rte_errno == EADDRNOTAVAIL)
370 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes at [%p] - please use '--base-virtaddr' option\n",
371 (unsigned long long)mem_sz, msl->base_va);
373 RTE_LOG(ERR, EAL, "Cannot reserve memory\n");
384 memseg_primary_init(void)
386 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
387 int hpi_idx, msl_idx = 0;
388 struct rte_memseg_list *msl;
389 uint64_t max_mem, total_mem;
391 /* no-huge does not need this at all */
392 if (internal_config.no_hugetlbfs)
395 /* FreeBSD has an issue where core dump will dump the entire memory
396 * contents, including anonymous zero-page memory. Therefore, while we
397 * will be limiting total amount of memory to RTE_MAX_MEM_MB, we will
398 * also be further limiting total memory amount to whatever memory is
399 * available to us through contigmem driver (plus spacing blocks).
401 * so, at each stage, we will be checking how much memory we are
402 * preallocating, and adjust all the values accordingly.
405 max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
408 /* create memseg lists */
409 for (hpi_idx = 0; hpi_idx < (int) internal_config.num_hugepage_sizes;
411 uint64_t max_type_mem, total_type_mem = 0;
413 int type_msl_idx, max_segs, avail_segs, total_segs = 0;
414 struct hugepage_info *hpi;
415 uint64_t hugepage_sz;
417 hpi = &internal_config.hugepage_info[hpi_idx];
418 hugepage_sz = hpi->hugepage_sz;
420 /* no NUMA support on FreeBSD */
422 /* check if we've already exceeded total memory amount */
423 if (total_mem >= max_mem)
426 /* first, calculate theoretical limits according to config */
427 max_type_mem = RTE_MIN(max_mem - total_mem,
428 (uint64_t)RTE_MAX_MEM_MB_PER_TYPE << 20);
429 max_segs = RTE_MAX_MEMSEG_PER_TYPE;
431 /* now, limit all of that to whatever will actually be
432 * available to us, because without dynamic allocation support,
433 * all of that extra memory will be sitting there being useless
434 * and slowing down core dumps in case of a crash.
436 * we need (N*2)-1 segments because we cannot guarantee that
437 * each segment will be IOVA-contiguous with the previous one,
438 * so we will allocate more and put spaces inbetween segments
439 * that are non-contiguous.
441 avail_segs = (hpi->num_pages[0] * 2) - 1;
442 avail_mem = avail_segs * hugepage_sz;
444 max_type_mem = RTE_MIN(avail_mem, max_type_mem);
445 max_segs = RTE_MIN(avail_segs, max_segs);
448 while (total_type_mem < max_type_mem &&
449 total_segs < max_segs) {
450 uint64_t cur_max_mem, cur_mem;
453 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
455 "No more space in memseg lists, please increase %s\n",
456 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
460 msl = &mcfg->memsegs[msl_idx++];
462 cur_max_mem = max_type_mem - total_type_mem;
464 cur_mem = get_mem_amount(hugepage_sz,
466 n_segs = cur_mem / hugepage_sz;
468 if (alloc_memseg_list(msl, hugepage_sz, n_segs,
472 total_segs += msl->memseg_arr.len;
473 total_type_mem = total_segs * hugepage_sz;
476 if (alloc_va_space(msl)) {
477 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list\n");
481 total_mem += total_type_mem;
487 memseg_secondary_init(void)
489 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
491 struct rte_memseg_list *msl;
493 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
495 msl = &mcfg->memsegs[msl_idx];
497 /* skip empty memseg lists */
498 if (msl->memseg_arr.len == 0)
501 if (rte_fbarray_attach(&msl->memseg_arr)) {
502 RTE_LOG(ERR, EAL, "Cannot attach to primary process memseg lists\n");
506 /* preallocate VA space */
507 if (alloc_va_space(msl)) {
508 RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
517 rte_eal_memseg_init(void)
519 return rte_eal_process_type() == RTE_PROC_PRIMARY ?
520 memseg_primary_init() :
521 memseg_secondary_init();