1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2010-2014 Intel Corporation 3 */ 4 #include <sys/mman.h> 5 #include <unistd.h> 6 #include <sys/types.h> 7 #include <sys/sysctl.h> 8 #include <inttypes.h> 9 #include <errno.h> 10 #include <string.h> 11 #include <fcntl.h> 12 13 #include <rte_eal.h> 14 #include <rte_errno.h> 15 #include <rte_log.h> 16 #include <rte_string_fns.h> 17 18 #include "eal_private.h" 19 #include "eal_internal_cfg.h" 20 #include "eal_filesystem.h" 21 #include "eal_memcfg.h" 22 #include "eal_options.h" 23 24 #define EAL_PAGE_SIZE (sysconf(_SC_PAGESIZE)) 25 26 uint64_t eal_get_baseaddr(void) 27 { 28 /* 29 * FreeBSD may allocate something in the space we will be mapping things 30 * before we get a chance to do that, so use a base address that's far 31 * away from where malloc() et al usually map things. 32 */ 33 return 0x1000000000ULL; 34 } 35 36 /* 37 * Get physical address of any mapped virtual address in the current process. 38 */ 39 phys_addr_t 40 rte_mem_virt2phy(const void *virtaddr) 41 { 42 /* XXX not implemented. This function is only used by 43 * rte_mempool_virt2iova() when hugepages are disabled. */ 44 (void)virtaddr; 45 return RTE_BAD_IOVA; 46 } 47 rte_iova_t 48 rte_mem_virt2iova(const void *virtaddr) 49 { 50 return rte_mem_virt2phy(virtaddr); 51 } 52 53 int 54 rte_eal_hugepage_init(void) 55 { 56 struct rte_mem_config *mcfg; 57 uint64_t total_mem = 0; 58 void *addr; 59 unsigned int i, j, seg_idx = 0; 60 struct internal_config *internal_conf = 61 eal_get_internal_configuration(); 62 63 /* get pointer to global configuration */ 64 mcfg = rte_eal_get_configuration()->mem_config; 65 66 /* for debug purposes, hugetlbfs can be disabled */ 67 if (internal_conf->no_hugetlbfs) { 68 struct rte_memseg_list *msl; 69 uint64_t mem_sz, page_sz; 70 int n_segs; 71 72 /* create a memseg list */ 73 msl = &mcfg->memsegs[0]; 74 75 mem_sz = internal_conf->memory; 76 page_sz = RTE_PGSIZE_4K; 77 n_segs = mem_sz / page_sz; 78 79 if (eal_memseg_list_init_named( 80 msl, "nohugemem", page_sz, n_segs, 0, true)) { 81 return -1; 82 } 83 84 addr = mmap(NULL, mem_sz, PROT_READ | PROT_WRITE, 85 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); 86 if (addr == MAP_FAILED) { 87 EAL_LOG(ERR, "%s: mmap() failed: %s", __func__, 88 strerror(errno)); 89 return -1; 90 } 91 92 msl->base_va = addr; 93 msl->len = mem_sz; 94 95 eal_memseg_list_populate(msl, addr, n_segs); 96 97 return 0; 98 } 99 100 /* map all hugepages and sort them */ 101 for (i = 0; i < internal_conf->num_hugepage_sizes; i++) { 102 struct hugepage_info *hpi; 103 rte_iova_t prev_end = 0; 104 int prev_ms_idx = -1; 105 uint64_t page_sz, mem_needed; 106 unsigned int n_pages, max_pages; 107 108 hpi = &internal_conf->hugepage_info[i]; 109 page_sz = hpi->hugepage_sz; 110 max_pages = hpi->num_pages[0]; 111 mem_needed = RTE_ALIGN_CEIL(internal_conf->memory - total_mem, 112 page_sz); 113 114 n_pages = RTE_MIN(mem_needed / page_sz, max_pages); 115 116 for (j = 0; j < n_pages; j++) { 117 struct rte_memseg_list *msl; 118 struct rte_fbarray *arr; 119 struct rte_memseg *seg; 120 int msl_idx, ms_idx; 121 rte_iova_t physaddr; 122 int error; 123 size_t sysctl_size = sizeof(physaddr); 124 char physaddr_str[64]; 125 bool is_adjacent; 126 127 /* first, check if this segment is IOVA-adjacent to 128 * the previous one. 129 */ 130 snprintf(physaddr_str, sizeof(physaddr_str), 131 "hw.contigmem.physaddr.%d", j); 132 error = sysctlbyname(physaddr_str, &physaddr, 133 &sysctl_size, NULL, 0); 134 if (error < 0) { 135 EAL_LOG(ERR, "Failed to get physical addr for buffer %u " 136 "from %s", j, hpi->hugedir); 137 return -1; 138 } 139 140 is_adjacent = prev_end != 0 && physaddr == prev_end; 141 prev_end = physaddr + hpi->hugepage_sz; 142 143 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; 144 msl_idx++) { 145 bool empty, need_hole; 146 msl = &mcfg->memsegs[msl_idx]; 147 arr = &msl->memseg_arr; 148 149 if (msl->page_sz != page_sz) 150 continue; 151 152 empty = arr->count == 0; 153 154 /* we need a hole if this isn't an empty memseg 155 * list, and if previous segment was not 156 * adjacent to current one. 157 */ 158 need_hole = !empty && !is_adjacent; 159 160 /* we need 1, plus hole if not adjacent */ 161 ms_idx = rte_fbarray_find_next_n_free(arr, 162 0, 1 + (need_hole ? 1 : 0)); 163 164 /* memseg list is full? */ 165 if (ms_idx < 0) 166 continue; 167 168 if (need_hole && prev_ms_idx == ms_idx - 1) 169 ms_idx++; 170 prev_ms_idx = ms_idx; 171 172 break; 173 } 174 if (msl_idx == RTE_MAX_MEMSEG_LISTS) { 175 EAL_LOG(ERR, "Could not find space for memseg. Please increase RTE_MAX_MEMSEG_PER_LIST " 176 "RTE_MAX_MEMSEG_PER_TYPE and/or RTE_MAX_MEM_MB_PER_TYPE in configuration."); 177 return -1; 178 } 179 arr = &msl->memseg_arr; 180 seg = rte_fbarray_get(arr, ms_idx); 181 182 addr = RTE_PTR_ADD(msl->base_va, 183 (size_t)msl->page_sz * ms_idx); 184 185 /* address is already mapped in memseg list, so using 186 * MAP_FIXED here is safe. 187 */ 188 addr = mmap(addr, page_sz, PROT_READ|PROT_WRITE, 189 MAP_SHARED | MAP_FIXED, 190 hpi->lock_descriptor, 191 j * EAL_PAGE_SIZE); 192 if (addr == MAP_FAILED) { 193 EAL_LOG(ERR, "Failed to mmap buffer %u from %s", 194 j, hpi->hugedir); 195 return -1; 196 } 197 198 seg->addr = addr; 199 seg->iova = physaddr; 200 seg->hugepage_sz = page_sz; 201 seg->len = page_sz; 202 seg->nchannel = mcfg->nchannel; 203 seg->nrank = mcfg->nrank; 204 seg->socket_id = 0; 205 206 rte_fbarray_set_used(arr, ms_idx); 207 208 EAL_LOG(INFO, "Mapped memory segment %u @ %p: physaddr:0x%" 209 PRIx64", len %zu", 210 seg_idx++, addr, physaddr, page_sz); 211 212 total_mem += seg->len; 213 } 214 if (total_mem >= internal_conf->memory) 215 break; 216 } 217 if (total_mem < internal_conf->memory) { 218 EAL_LOG(ERR, "Couldn't reserve requested memory, " 219 "requested: %" PRIu64 "M " 220 "available: %" PRIu64 "M", 221 internal_conf->memory >> 20, total_mem >> 20); 222 return -1; 223 } 224 return 0; 225 } 226 227 struct attach_walk_args { 228 int fd_hugepage; 229 int seg_idx; 230 }; 231 static int 232 attach_segment(const struct rte_memseg_list *msl, const struct rte_memseg *ms, 233 void *arg) 234 { 235 struct attach_walk_args *wa = arg; 236 void *addr; 237 238 if (msl->external) 239 return 0; 240 241 addr = mmap(ms->addr, ms->len, PROT_READ | PROT_WRITE, 242 MAP_SHARED | MAP_FIXED, wa->fd_hugepage, 243 wa->seg_idx * EAL_PAGE_SIZE); 244 if (addr == MAP_FAILED || addr != ms->addr) 245 return -1; 246 wa->seg_idx++; 247 248 return 0; 249 } 250 251 int 252 rte_eal_hugepage_attach(void) 253 { 254 struct hugepage_info *hpi; 255 int fd_hugepage = -1; 256 unsigned int i; 257 struct internal_config *internal_conf = 258 eal_get_internal_configuration(); 259 260 hpi = &internal_conf->hugepage_info[0]; 261 262 for (i = 0; i < internal_conf->num_hugepage_sizes; i++) { 263 const struct hugepage_info *cur_hpi = &hpi[i]; 264 struct attach_walk_args wa; 265 266 memset(&wa, 0, sizeof(wa)); 267 268 /* Obtain a file descriptor for contiguous memory */ 269 fd_hugepage = open(cur_hpi->hugedir, O_RDWR); 270 if (fd_hugepage < 0) { 271 EAL_LOG(ERR, "Could not open %s", 272 cur_hpi->hugedir); 273 goto error; 274 } 275 wa.fd_hugepage = fd_hugepage; 276 wa.seg_idx = 0; 277 278 /* Map the contiguous memory into each memory segment */ 279 if (rte_memseg_walk(attach_segment, &wa) < 0) { 280 EAL_LOG(ERR, "Failed to mmap buffer %u from %s", 281 wa.seg_idx, cur_hpi->hugedir); 282 goto error; 283 } 284 285 close(fd_hugepage); 286 fd_hugepage = -1; 287 } 288 289 /* hugepage_info is no longer required */ 290 return 0; 291 292 error: 293 if (fd_hugepage >= 0) 294 close(fd_hugepage); 295 return -1; 296 } 297 298 int 299 rte_eal_using_phys_addrs(void) 300 { 301 return 0; 302 } 303 304 static uint64_t 305 get_mem_amount(uint64_t page_sz, uint64_t max_mem) 306 { 307 uint64_t area_sz, max_pages; 308 309 /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */ 310 max_pages = RTE_MAX_MEMSEG_PER_LIST; 311 max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem); 312 313 area_sz = RTE_MIN(page_sz * max_pages, max_mem); 314 315 /* make sure the list isn't smaller than the page size */ 316 area_sz = RTE_MAX(area_sz, page_sz); 317 318 return RTE_ALIGN(area_sz, page_sz); 319 } 320 321 static int 322 memseg_list_alloc(struct rte_memseg_list *msl) 323 { 324 int flags = 0; 325 326 #ifdef RTE_ARCH_PPC_64 327 flags |= EAL_RESERVE_HUGEPAGES; 328 #endif 329 return eal_memseg_list_alloc(msl, flags); 330 } 331 332 static int 333 memseg_primary_init(void) 334 { 335 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config; 336 int hpi_idx, msl_idx = 0; 337 struct rte_memseg_list *msl; 338 uint64_t max_mem, total_mem; 339 struct internal_config *internal_conf = 340 eal_get_internal_configuration(); 341 342 /* no-huge does not need this at all */ 343 if (internal_conf->no_hugetlbfs) 344 return 0; 345 346 /* FreeBSD has an issue where core dump will dump the entire memory 347 * contents, including anonymous zero-page memory. Therefore, while we 348 * will be limiting total amount of memory to RTE_MAX_MEM_MB, we will 349 * also be further limiting total memory amount to whatever memory is 350 * available to us through contigmem driver (plus spacing blocks). 351 * 352 * so, at each stage, we will be checking how much memory we are 353 * preallocating, and adjust all the values accordingly. 354 */ 355 356 max_mem = (uint64_t)RTE_MAX_MEM_MB << 20; 357 total_mem = 0; 358 359 /* create memseg lists */ 360 for (hpi_idx = 0; hpi_idx < (int) internal_conf->num_hugepage_sizes; 361 hpi_idx++) { 362 uint64_t max_type_mem, total_type_mem = 0; 363 uint64_t avail_mem; 364 int type_msl_idx, max_segs, avail_segs, total_segs = 0; 365 struct hugepage_info *hpi; 366 uint64_t hugepage_sz; 367 368 hpi = &internal_conf->hugepage_info[hpi_idx]; 369 hugepage_sz = hpi->hugepage_sz; 370 371 /* no NUMA support on FreeBSD */ 372 373 /* check if we've already exceeded total memory amount */ 374 if (total_mem >= max_mem) 375 break; 376 377 /* first, calculate theoretical limits according to config */ 378 max_type_mem = RTE_MIN(max_mem - total_mem, 379 (uint64_t)RTE_MAX_MEM_MB_PER_TYPE << 20); 380 max_segs = RTE_MAX_MEMSEG_PER_TYPE; 381 382 /* now, limit all of that to whatever will actually be 383 * available to us, because without dynamic allocation support, 384 * all of that extra memory will be sitting there being useless 385 * and slowing down core dumps in case of a crash. 386 * 387 * we need (N*2)-1 segments because we cannot guarantee that 388 * each segment will be IOVA-contiguous with the previous one, 389 * so we will allocate more and put spaces between segments 390 * that are non-contiguous. 391 */ 392 avail_segs = (hpi->num_pages[0] * 2) - 1; 393 avail_mem = avail_segs * hugepage_sz; 394 395 max_type_mem = RTE_MIN(avail_mem, max_type_mem); 396 max_segs = RTE_MIN(avail_segs, max_segs); 397 398 type_msl_idx = 0; 399 while (total_type_mem < max_type_mem && 400 total_segs < max_segs) { 401 uint64_t cur_max_mem, cur_mem; 402 unsigned int n_segs; 403 404 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) { 405 EAL_LOG(ERR, 406 "No more space in memseg lists, please increase RTE_MAX_MEMSEG_LISTS"); 407 return -1; 408 } 409 410 msl = &mcfg->memsegs[msl_idx++]; 411 412 cur_max_mem = max_type_mem - total_type_mem; 413 414 cur_mem = get_mem_amount(hugepage_sz, 415 cur_max_mem); 416 n_segs = cur_mem / hugepage_sz; 417 418 if (eal_memseg_list_init(msl, hugepage_sz, n_segs, 419 0, type_msl_idx, false)) 420 return -1; 421 422 total_segs += msl->memseg_arr.len; 423 total_type_mem = total_segs * hugepage_sz; 424 type_msl_idx++; 425 426 if (memseg_list_alloc(msl)) { 427 EAL_LOG(ERR, "Cannot allocate VA space for memseg list"); 428 return -1; 429 } 430 } 431 total_mem += total_type_mem; 432 } 433 return 0; 434 } 435 436 static int 437 memseg_secondary_init(void) 438 { 439 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config; 440 int msl_idx = 0; 441 struct rte_memseg_list *msl; 442 443 for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) { 444 445 msl = &mcfg->memsegs[msl_idx]; 446 447 /* skip empty and external memseg lists */ 448 if (msl->memseg_arr.len == 0 || msl->external) 449 continue; 450 451 if (rte_fbarray_attach(&msl->memseg_arr)) { 452 EAL_LOG(ERR, "Cannot attach to primary process memseg lists"); 453 return -1; 454 } 455 456 /* preallocate VA space */ 457 if (memseg_list_alloc(msl)) { 458 EAL_LOG(ERR, "Cannot preallocate VA space for hugepage memory"); 459 return -1; 460 } 461 } 462 463 return 0; 464 } 465 466 int 467 rte_eal_memseg_init(void) 468 { 469 return rte_eal_process_type() == RTE_PROC_PRIMARY ? 470 memseg_primary_init() : 471 memseg_secondary_init(); 472 } 473