lib/librte_eal/linuxapp/eal/eal_ivshmem.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #ifdef RTE_LIBRTE_IVSHMEM /* hide it from coverage */
  35
  36 #include <stdint.h>
  37 #include <unistd.h>
  38 #include <inttypes.h>
  39 #include <sys/mman.h>
  40 #include <sys/file.h>
  41 #include <string.h>
  42 #include <sys/queue.h>
  43
  44 #include <rte_log.h>
  45 #include <rte_pci.h>
  46 #include <rte_memory.h>
  47 #include <rte_eal.h>
  48 #include <rte_eal_memconfig.h>
  49 #include <rte_string_fns.h>
  50 #include <rte_errno.h>
  51 #include <rte_ring.h>
  52 #include <rte_mempool.h>
  53 #include <rte_malloc.h>
  54 #include <rte_common.h>
  55 #include <rte_ivshmem.h>
  56
  57 #include "eal_internal_cfg.h"
  58 #include "eal_private.h"
  59
  60 #define PCI_VENDOR_ID_IVSHMEM 0x1Af4
  61 #define PCI_DEVICE_ID_IVSHMEM 0x1110
  62
  63 #define IVSHMEM_MAGIC 0x0BADC0DE
  64
  65 #define IVSHMEM_RESOURCE_PATH "/sys/bus/pci/devices/%04x:%02x:%02x.%x/resource2"
  66 #define IVSHMEM_CONFIG_PATH "/var/run/.%s_ivshmem_config"
  67
  68 #define PHYS 0x1
  69 #define VIRT 0x2
  70 #define IOREMAP 0x4
  71 #define FULL (PHYS|VIRT|IOREMAP)
  72
  73 #define METADATA_SIZE_ALIGNED \
  74         (RTE_ALIGN_CEIL(sizeof(struct rte_ivshmem_metadata),pagesz))
  75
  76 #define CONTAINS(x,y)\
  77         (((y).addr_64 >= (x).addr_64) && ((y).addr_64 < (x).addr_64 + (x).len))
  78
  79 #define DIM(x) (sizeof(x)/sizeof(x[0]))
  80
  81 struct ivshmem_pci_device {
  82         char path[PATH_MAX];
  83         phys_addr_t ioremap_addr;
  84 };
  85
  86 /* data type to store in config */
  87 struct ivshmem_segment {
  88         struct rte_ivshmem_metadata_entry entry;
  89         uint64_t align;
  90         char path[PATH_MAX];
  91 };
  92 struct ivshmem_shared_config {
  93         struct ivshmem_segment segment[RTE_MAX_MEMSEG];
  94         uint32_t segment_idx;
  95         struct ivshmem_pci_device pci_devs[RTE_LIBRTE_IVSHMEM_MAX_PCI_DEVS];
  96         uint32_t pci_devs_idx;
  97 };
  98 static struct ivshmem_shared_config * ivshmem_config;
  99 static int memseg_idx;
 100 static int pagesz;
 101
 102 /* Tailq heads to add rings to */
 103 TAILQ_HEAD(rte_ring_list, rte_tailq_entry);
 104
 105 /*
 106  * Utility functions
 107  */
 108
 109 static int
 110 is_ivshmem_device(struct rte_pci_device * dev)
 111 {
 112         return dev->id.vendor_id == PCI_VENDOR_ID_IVSHMEM
 113                         && dev->id.device_id == PCI_DEVICE_ID_IVSHMEM;
 114 }
 115
 116 static void *
 117 map_metadata(int fd, uint64_t len)
 118 {
 119         size_t metadata_len = sizeof(struct rte_ivshmem_metadata);
 120         size_t aligned_len = METADATA_SIZE_ALIGNED;
 121
 122         return mmap(NULL, metadata_len, PROT_READ | PROT_WRITE,
 123                         MAP_SHARED, fd, len - aligned_len);
 124 }
 125
 126 static void
 127 unmap_metadata(void * ptr)
 128 {
 129         munmap(ptr, sizeof(struct rte_ivshmem_metadata));
 130 }
 131
 132 static int
 133 has_ivshmem_metadata(int fd, uint64_t len)
 134 {
 135         struct rte_ivshmem_metadata metadata;
 136         void * ptr;
 137
 138         ptr = map_metadata(fd, len);
 139
 140         if (ptr == MAP_FAILED)
 141                 return -1;
 142
 143         metadata = *(struct rte_ivshmem_metadata*) (ptr);
 144
 145         unmap_metadata(ptr);
 146
 147         return metadata.magic_number == IVSHMEM_MAGIC;
 148 }
 149
 150 static void
 151 remove_segment(struct ivshmem_segment * ms, int len, int idx)
 152 {
 153         int i;
 154
 155         for (i = idx; i < len - 1; i++)
 156                 memcpy(&ms[i], &ms[i+1], sizeof(struct ivshmem_segment));
 157         memset(&ms[len-1], 0, sizeof(struct ivshmem_segment));
 158 }
 159
 160 static int
 161 overlap(const struct rte_memzone * mz1, const struct rte_memzone * mz2)
 162 {
 163         uint64_t start1, end1, start2, end2;
 164         uint64_t p_start1, p_end1, p_start2, p_end2;
 165         uint64_t i_start1, i_end1, i_start2, i_end2;
 166         int result = 0;
 167
 168         /* gather virtual addresses */
 169         start1 = mz1->addr_64;
 170         end1 = mz1->addr_64 + mz1->len;
 171         start2 = mz2->addr_64;
 172         end2 = mz2->addr_64 + mz2->len;
 173
 174         /* gather physical addresses */
 175         p_start1 = mz1->phys_addr;
 176         p_end1 = mz1->phys_addr + mz1->len;
 177         p_start2 = mz2->phys_addr;
 178         p_end2 = mz2->phys_addr + mz2->len;
 179
 180         /* gather ioremap addresses */
 181         i_start1 = mz1->ioremap_addr;
 182         i_end1 = mz1->ioremap_addr + mz1->len;
 183         i_start2 = mz2->ioremap_addr;
 184         i_end2 = mz2->ioremap_addr + mz2->len;
 185
 186         /* check for overlap in virtual addresses */
 187         if (start1 >= start2 && start1 < end2)
 188                 result |= VIRT;
 189         if (start2 >= start1 && start2 < end1)
 190                 result |= VIRT;
 191
 192         /* check for overlap in physical addresses */
 193         if (p_start1 >= p_start2 && p_start1 < p_end2)
 194                 result |= PHYS;
 195         if (p_start2 >= p_start1 && p_start2 < p_end1)
 196                 result |= PHYS;
 197
 198         /* check for overlap in ioremap addresses */
 199         if (i_start1 >= i_start2 && i_start1 < i_end2)
 200                 result |= IOREMAP;
 201         if (i_start2 >= i_start1 && i_start2 < i_end1)
 202                 result |= IOREMAP;
 203
 204         return result;
 205 }
 206
 207 static int
 208 adjacent(const struct rte_memzone * mz1, const struct rte_memzone * mz2)
 209 {
 210         uint64_t start1, end1, start2, end2;
 211         uint64_t p_start1, p_end1, p_start2, p_end2;
 212         uint64_t i_start1, i_end1, i_start2, i_end2;
 213         int result = 0;
 214
 215         /* gather virtual addresses */
 216         start1 = mz1->addr_64;
 217         end1 = mz1->addr_64 + mz1->len;
 218         start2 = mz2->addr_64;
 219         end2 = mz2->addr_64 + mz2->len;
 220
 221         /* gather physical addresses */
 222         p_start1 = mz1->phys_addr;
 223         p_end1 = mz1->phys_addr + mz1->len;
 224         p_start2 = mz2->phys_addr;
 225         p_end2 = mz2->phys_addr + mz2->len;
 226
 227         /* gather ioremap addresses */
 228         i_start1 = mz1->ioremap_addr;
 229         i_end1 = mz1->ioremap_addr + mz1->len;
 230         i_start2 = mz2->ioremap_addr;
 231         i_end2 = mz2->ioremap_addr + mz2->len;
 232
 233         /* check if segments are virtually adjacent */
 234         if (start1 == end2)
 235                 result |= VIRT;
 236         if (start2 == end1)
 237                 result |= VIRT;
 238
 239         /* check if segments are physically adjacent */
 240         if (p_start1 == p_end2)
 241                 result |= PHYS;
 242         if (p_start2 == p_end1)
 243                 result |= PHYS;
 244
 245         /* check if segments are ioremap-adjacent */
 246         if (i_start1 == i_end2)
 247                 result |= IOREMAP;
 248         if (i_start2 == i_end1)
 249                 result |= IOREMAP;
 250
 251         return result;
 252 }
 253
 254 static int
 255 has_adjacent_segments(struct ivshmem_segment * ms, int len)
 256 {
 257         int i, j, a;
 258
 259         for (i = 0; i < len; i++)
 260                 for (j = i + 1; j < len; j++) {
 261                         a = adjacent(&ms[i].entry.mz, &ms[j].entry.mz);
 262
 263                         /* check if segments are adjacent virtually and/or physically but
 264                          * not ioremap (since that would indicate that they are from
 265                          * different PCI devices and thus don't need to be concatenated.
 266                          */
 267                         if ((a & (VIRT|PHYS)) > 0 && (a & IOREMAP) == 0)
 268                                 return 1;
 269                 }
 270         return 0;
 271 }
 272
 273 static int
 274 has_overlapping_segments(struct ivshmem_segment * ms, int len)
 275 {
 276         int i, j;
 277
 278         for (i = 0; i < len; i++)
 279                 for (j = i + 1; j < len; j++)
 280                         if (overlap(&ms[i].entry.mz, &ms[j].entry.mz))
 281                                 return 1;
 282         return 0;
 283 }
 284
 285 static int
 286 seg_compare(const void * a, const void * b)
 287 {
 288         const struct ivshmem_segment * s1 = (const struct ivshmem_segment*) a;
 289         const struct ivshmem_segment * s2 = (const struct ivshmem_segment*) b;
 290
 291         /* move unallocated zones to the end */
 292         if (s1->entry.mz.addr == NULL && s2->entry.mz.addr == NULL)
 293                 return 0;
 294         if (s1->entry.mz.addr == 0)
 295                 return 1;
 296         if (s2->entry.mz.addr == 0)
 297                 return -1;
 298
 299         return s1->entry.mz.phys_addr > s2->entry.mz.phys_addr;
 300 }
 301
 302 #ifdef RTE_LIBRTE_IVSHMEM_DEBUG
 303 static void
 304 entry_dump(struct rte_ivshmem_metadata_entry *e)
 305 {
 306         RTE_LOG(DEBUG, EAL, "\tvirt: %p-%p\n", e->mz.addr,
 307                         RTE_PTR_ADD(e->mz.addr, e->mz.len));
 308         RTE_LOG(DEBUG, EAL, "\tphys: 0x%" PRIx64 "-0x%" PRIx64 "\n",
 309                         e->mz.phys_addr,
 310                         e->mz.phys_addr + e->mz.len);
 311         RTE_LOG(DEBUG, EAL, "\tio: 0x%" PRIx64 "-0x%" PRIx64 "\n",
 312                         e->mz.ioremap_addr,
 313                         e->mz.ioremap_addr + e->mz.len);
 314         RTE_LOG(DEBUG, EAL, "\tlen: 0x%" PRIx64 "\n", e->mz.len);
 315         RTE_LOG(DEBUG, EAL, "\toff: 0x%" PRIx64 "\n", e->offset);
 316 }
 317 #endif
 318
 319
 320
 321 /*
 322  * Actual useful code
 323  */
 324
 325 /* read through metadata mapped from the IVSHMEM device */
 326 static int
 327 read_metadata(char * path, int path_len, int fd, uint64_t flen)
 328 {
 329         struct rte_ivshmem_metadata metadata;
 330         struct rte_ivshmem_metadata_entry * entry;
 331         int idx, i;
 332         void * ptr;
 333
 334         ptr = map_metadata(fd, flen);
 335
 336         if (ptr == MAP_FAILED)
 337                 return -1;
 338
 339         metadata = *(struct rte_ivshmem_metadata*) (ptr);
 340
 341         unmap_metadata(ptr);
 342
 343         RTE_LOG(DEBUG, EAL, "Parsing metadata for \"%s\"\n", metadata.name);
 344
 345         idx = ivshmem_config->segment_idx;
 346
 347         for (i = 0; i < RTE_LIBRTE_IVSHMEM_MAX_ENTRIES &&
 348                 idx <= RTE_MAX_MEMSEG; i++) {
 349
 350                 if (idx == RTE_MAX_MEMSEG) {
 351                         RTE_LOG(ERR, EAL, "Not enough memory segments!\n");
 352                         return -1;
 353                 }
 354
 355                 entry = &metadata.entry[i];
 356
 357                 /* stop on uninitialized memzone */
 358                 if (entry->mz.len == 0)
 359                         break;
 360
 361                 /* copy metadata entry */
 362                 memcpy(&ivshmem_config->segment[idx].entry, entry,
 363                                 sizeof(struct rte_ivshmem_metadata_entry));
 364
 365                 /* copy path */
 366                 snprintf(ivshmem_config->segment[idx].path, path_len, "%s", path);
 367
 368                 idx++;
 369         }
 370         ivshmem_config->segment_idx = idx;
 371
 372         return 0;
 373 }
 374
 375 /* check through each segment and look for adjacent or overlapping ones. */
 376 static int
 377 cleanup_segments(struct ivshmem_segment * ms, int tbl_len)
 378 {
 379         struct ivshmem_segment * s, * tmp;
 380         int i, j, concat, seg_adjacent, seg_overlapping;
 381         uint64_t start1, start2, end1, end2, p_start1, p_start2, i_start1, i_start2;
 382
 383         qsort(ms, tbl_len, sizeof(struct ivshmem_segment),
 384                                 seg_compare);
 385
 386         while (has_overlapping_segments(ms, tbl_len) ||
 387                         has_adjacent_segments(ms, tbl_len)) {
 388
 389                 for (i = 0; i < tbl_len; i++) {
 390                         s = &ms[i];
 391
 392                         concat = 0;
 393
 394                         for (j = i + 1; j < tbl_len; j++) {
 395                                 tmp = &ms[j];
 396
 397                                 /* check if this segment is overlapping with existing segment,
 398                                  * or is adjacent to existing segment */
 399                                 seg_overlapping = overlap(&s->entry.mz, &tmp->entry.mz);
 400                                 seg_adjacent = adjacent(&s->entry.mz, &tmp->entry.mz);
 401
 402                                 /* check if segments fully overlap or are fully adjacent */
 403                                 if ((seg_adjacent == FULL) || (seg_overlapping == FULL)) {
 404
 405 #ifdef RTE_LIBRTE_IVSHMEM_DEBUG
 406                                         RTE_LOG(DEBUG, EAL, "Concatenating segments\n");
 407                                         RTE_LOG(DEBUG, EAL, "Segment %i:\n", i);
 408                                         entry_dump(&s->entry);
 409                                         RTE_LOG(DEBUG, EAL, "Segment %i:\n", j);
 410                                         entry_dump(&tmp->entry);
 411 #endif
 412
 413                                         start1 = s->entry.mz.addr_64;
 414                                         start2 = tmp->entry.mz.addr_64;
 415                                         p_start1 = s->entry.mz.phys_addr;
 416                                         p_start2 = tmp->entry.mz.phys_addr;
 417                                         i_start1 = s->entry.mz.ioremap_addr;
 418                                         i_start2 = tmp->entry.mz.ioremap_addr;
 419                                         end1 = s->entry.mz.addr_64 + s->entry.mz.len;
 420                                         end2 = tmp->entry.mz.addr_64 + tmp->entry.mz.len;
 421
 422                                         /* settle for minimum start address and maximum length */
 423                                         s->entry.mz.addr_64 = RTE_MIN(start1, start2);
 424                                         s->entry.mz.phys_addr = RTE_MIN(p_start1, p_start2);
 425                                         s->entry.mz.ioremap_addr = RTE_MIN(i_start1, i_start2);
 426                                         s->entry.offset = RTE_MIN(s->entry.offset, tmp->entry.offset);
 427                                         s->entry.mz.len = RTE_MAX(end1, end2) - s->entry.mz.addr_64;
 428                                         concat = 1;
 429
 430 #ifdef RTE_LIBRTE_IVSHMEM_DEBUG
 431                                         RTE_LOG(DEBUG, EAL, "Resulting segment:\n");
 432                                         entry_dump(&s->entry);
 433
 434 #endif
 435                                 }
 436                                 /* if segments not fully overlap, we have an error condition.
 437                                  * adjacent segments can coexist.
 438                                  */
 439                                 else if (seg_overlapping > 0) {
 440                                         RTE_LOG(ERR, EAL, "Segments %i and %i overlap!\n", i, j);
 441 #ifdef RTE_LIBRTE_IVSHMEM_DEBUG
 442                                         RTE_LOG(DEBUG, EAL, "Segment %i:\n", i);
 443                                         entry_dump(&s->entry);
 444                                         RTE_LOG(DEBUG, EAL, "Segment %i:\n", j);
 445                                         entry_dump(&tmp->entry);
 446 #endif
 447                                         return -1;
 448                                 }
 449                                 if (concat)
 450                                         break;
 451                         }
 452                         /* if we concatenated, remove segment at j */
 453                         if (concat) {
 454                                 remove_segment(ms, tbl_len, j);
 455                                 tbl_len--;
 456                                 break;
 457                         }
 458                 }
 459         }
 460
 461         return tbl_len;
 462 }
 463
 464 static int
 465 create_shared_config(void)
 466 {
 467         char path[PATH_MAX];
 468         int fd;
 469
 470         /* build ivshmem config file path */
 471         snprintf(path, sizeof(path), IVSHMEM_CONFIG_PATH,
 472                         internal_config.hugefile_prefix);
 473
 474         fd = open(path, O_CREAT | O_RDWR, 0600);
 475
 476         if (fd < 0) {
 477                 RTE_LOG(ERR, EAL, "Could not open %s: %s\n", path, strerror(errno));
 478                 return -1;
 479         }
 480
 481         /* try ex-locking first - if the file is locked, we have a problem */
 482         if (flock(fd, LOCK_EX | LOCK_NB) == -1) {
 483                 RTE_LOG(ERR, EAL, "Locking %s failed: %s\n", path, strerror(errno));
 484                 close(fd);
 485                 return -1;
 486         }
 487
 488         if (ftruncate(fd, sizeof(struct ivshmem_shared_config)) < 0) {
 489                 RTE_LOG(ERR, EAL, "ftruncate failed: %s\n", strerror(errno));
 490                 return -1;
 491         }
 492
 493         ivshmem_config = mmap(NULL, sizeof(struct ivshmem_shared_config),
 494                         PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 495
 496         if (ivshmem_config == MAP_FAILED)
 497                 return -1;
 498
 499         memset(ivshmem_config, 0, sizeof(struct ivshmem_shared_config));
 500
 501         /* change the exclusive lock we got earlier to a shared lock */
 502         if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 503                 RTE_LOG(ERR, EAL, "Locking %s failed: %s \n", path, strerror(errno));
 504                 return -1;
 505         }
 506
 507         close(fd);
 508
 509         return 0;
 510 }
 511
 512 /* open shared config file and, if present, map the config.
 513  * having no config file is not an error condition, as we later check if
 514  * ivshmem_config is NULL (if it is, that means nothing was mapped). */
 515 static int
 516 open_shared_config(void)
 517 {
 518         char path[PATH_MAX];
 519         int fd;
 520
 521         /* build ivshmem config file path */
 522         snprintf(path, sizeof(path), IVSHMEM_CONFIG_PATH,
 523                         internal_config.hugefile_prefix);
 524
 525         fd = open(path, O_RDONLY);
 526
 527         /* if the file doesn't exist, just return success */
 528         if (fd < 0 && errno == ENOENT)
 529                 return 0;
 530         /* else we have an error condition */
 531         else if (fd < 0) {
 532                 RTE_LOG(ERR, EAL, "Could not open %s: %s\n",
 533                                 path, strerror(errno));
 534                 return -1;
 535         }
 536
 537         /* try ex-locking first - if the lock *does* succeed, this means it's a
 538          * stray config file, so it should be deleted.
 539          */
 540         if (flock(fd, LOCK_EX | LOCK_NB) != -1) {
 541
 542                 /* if we can't remove the file, something is wrong */
 543                 if (unlink(path) < 0) {
 544                         RTE_LOG(ERR, EAL, "Could not remove %s: %s\n", path,
 545                                         strerror(errno));
 546                         return -1;
 547                 }
 548
 549                 /* release the lock */
 550                 flock(fd, LOCK_UN);
 551                 close(fd);
 552
 553                 /* return success as having a stray config file is equivalent to not
 554                  * having config file at all.
 555                  */
 556                 return 0;
 557         }
 558
 559         ivshmem_config = mmap(NULL, sizeof(struct ivshmem_shared_config),
 560                         PROT_READ, MAP_SHARED, fd, 0);
 561
 562         if (ivshmem_config == MAP_FAILED)
 563                 return -1;
 564
 565         /* place a shared lock on config file */
 566         if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 567                 RTE_LOG(ERR, EAL, "Locking %s failed: %s \n", path, strerror(errno));
 568                 return -1;
 569         }
 570
 571         close(fd);
 572
 573         return 0;
 574 }
 575
 576 /*
 577  * This function does the following:
 578  *
 579  * 1) Builds a table of ivshmem_segments with proper offset alignment
 580  * 2) Cleans up that table so that we don't have any overlapping or adjacent
 581  *    memory segments
 582  * 3) Creates memsegs from this table and maps them into memory.
 583  */
 584 static inline int
 585 map_all_segments(void)
 586 {
 587         struct ivshmem_segment ms_tbl[RTE_MAX_MEMSEG];
 588         struct ivshmem_pci_device * pci_dev;
 589         struct rte_mem_config * mcfg;
 590         struct ivshmem_segment * seg;
 591         int fd, fd_zero;
 592         unsigned i, j;
 593         struct rte_memzone mz;
 594         struct rte_memseg ms;
 595         void * base_addr;
 596         uint64_t align, len;
 597         phys_addr_t ioremap_addr;
 598
 599         ioremap_addr = 0;
 600
 601         memset(ms_tbl, 0, sizeof(ms_tbl));
 602         memset(&mz, 0, sizeof(struct rte_memzone));
 603         memset(&ms, 0, sizeof(struct rte_memseg));
 604
 605         /* first, build a table of memsegs to map, to avoid failed mmaps due to
 606          * overlaps
 607          */
 608         for (i = 0; i < ivshmem_config->segment_idx && i <= RTE_MAX_MEMSEG; i++) {
 609                 if (i == RTE_MAX_MEMSEG) {
 610                         RTE_LOG(ERR, EAL, "Too many segments requested!\n");
 611                         return -1;
 612                 }
 613
 614                 seg = &ivshmem_config->segment[i];
 615
 616                 /* copy segment to table */
 617                 memcpy(&ms_tbl[i], seg, sizeof(struct ivshmem_segment));
 618
 619                 /* find ioremap addr */
 620                 for (j = 0; j < DIM(ivshmem_config->pci_devs); j++) {
 621                         pci_dev = &ivshmem_config->pci_devs[j];
 622                         if (!strncmp(pci_dev->path, seg->path, sizeof(pci_dev->path))) {
 623                                 ioremap_addr = pci_dev->ioremap_addr;
 624                                 break;
 625                         }
 626                 }
 627                 if (ioremap_addr == 0) {
 628                         RTE_LOG(ERR, EAL, "Cannot find ioremap addr!\n");
 629                         return -1;
 630                 }
 631
 632                 /* work out alignments */
 633                 align = seg->entry.mz.addr_64 -
 634                                 RTE_ALIGN_FLOOR(seg->entry.mz.addr_64, 0x1000);
 635                 len = RTE_ALIGN_CEIL(seg->entry.mz.len + align, 0x1000);
 636
 637                 /* save original alignments */
 638                 ms_tbl[i].align = align;
 639
 640                 /* create a memory zone */
 641                 mz.addr_64 = seg->entry.mz.addr_64 - align;
 642                 mz.len = len;
 643                 mz.hugepage_sz = seg->entry.mz.hugepage_sz;
 644                 mz.phys_addr = seg->entry.mz.phys_addr - align;
 645
 646                 /* find true physical address */
 647                 mz.ioremap_addr = ioremap_addr + seg->entry.offset - align;
 648
 649                 ms_tbl[i].entry.offset = seg->entry.offset - align;
 650
 651                 memcpy(&ms_tbl[i].entry.mz, &mz, sizeof(struct rte_memzone));
 652         }
 653
 654         /* clean up the segments */
 655         memseg_idx = cleanup_segments(ms_tbl, ivshmem_config->segment_idx);
 656
 657         if (memseg_idx < 0)
 658                 return -1;
 659
 660         mcfg = rte_eal_get_configuration()->mem_config;
 661
 662         fd_zero = open("/dev/zero", O_RDWR);
 663
 664         if (fd_zero < 0) {
 665                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero: %s\n", strerror(errno));
 666                 return -1;
 667         }
 668
 669         /* create memsegs and put them into DPDK memory */
 670         for (i = 0; i < (unsigned) memseg_idx; i++) {
 671
 672                 seg = &ms_tbl[i];
 673
 674                 ms.addr_64 = seg->entry.mz.addr_64;
 675                 ms.hugepage_sz = seg->entry.mz.hugepage_sz;
 676                 ms.len = seg->entry.mz.len;
 677                 ms.nchannel = rte_memory_get_nchannel();
 678                 ms.nrank = rte_memory_get_nrank();
 679                 ms.phys_addr = seg->entry.mz.phys_addr;
 680                 ms.ioremap_addr = seg->entry.mz.ioremap_addr;
 681                 ms.socket_id = seg->entry.mz.socket_id;
 682
 683                 base_addr = mmap(ms.addr, ms.len,
 684                                 PROT_READ | PROT_WRITE, MAP_PRIVATE, fd_zero, 0);
 685
 686                 if (base_addr == MAP_FAILED || base_addr != ms.addr) {
 687                         RTE_LOG(ERR, EAL, "Cannot map /dev/zero!\n");
 688                         return -1;
 689                 }
 690
 691                 fd = open(seg->path, O_RDWR);
 692
 693                 if (fd < 0) {
 694                         RTE_LOG(ERR, EAL, "Cannot open %s: %s\n", seg->path,
 695                                         strerror(errno));
 696                         return -1;
 697                 }
 698
 699                 munmap(ms.addr, ms.len);
 700
 701                 base_addr = mmap(ms.addr, ms.len,
 702                                 PROT_READ | PROT_WRITE, MAP_SHARED, fd,
 703                                 seg->entry.offset);
 704
 705
 706                 if (base_addr == MAP_FAILED || base_addr != ms.addr) {
 707                         RTE_LOG(ERR, EAL, "Cannot map segment into memory: "
 708                                         "expected %p got %p (%s)\n", ms.addr, base_addr,
 709                                         strerror(errno));
 710                         return -1;
 711                 }
 712
 713                 RTE_LOG(DEBUG, EAL, "Memory segment mapped: %p (len %" PRIx64 ") at "
 714                                 "offset 0x%" PRIx64 "\n",
 715                                 ms.addr, ms.len, seg->entry.offset);
 716
 717                 /* put the pointers back into their real positions using original
 718                  * alignment */
 719                 ms.addr_64 += seg->align;
 720                 ms.phys_addr += seg->align;
 721                 ms.ioremap_addr += seg->align;
 722                 ms.len -= seg->align;
 723
 724                 /* at this point, the rest of DPDK memory is not initialized, so we
 725                  * expect memsegs to be empty */
 726                 memcpy(&mcfg->memseg[i], &ms,
 727                                 sizeof(struct rte_memseg));
 728
 729                 close(fd);
 730
 731                 RTE_LOG(DEBUG, EAL, "IVSHMEM segment found, size: 0x%lx\n",
 732                                 ms.len);
 733         }
 734
 735         return 0;
 736 }
 737
 738 /* this happens at a later stage, after general EAL memory initialization */
 739 int
 740 rte_eal_ivshmem_obj_init(void)
 741 {
 742         struct rte_ring_list* ring_list = NULL;
 743         struct rte_mem_config * mcfg;
 744         struct ivshmem_segment * seg;
 745         struct rte_memzone * mz;
 746         struct rte_ring * r;
 747         struct rte_tailq_entry *te;
 748         unsigned i, ms, idx;
 749         uint64_t offset;
 750
 751         /* secondary process would not need any object discovery - it'll all
 752          * already be in shared config */
 753         if (rte_eal_process_type() != RTE_PROC_PRIMARY || ivshmem_config == NULL)
 754                 return 0;
 755
 756         /* check that we have an initialised ring tail queue */
 757         ring_list = RTE_TAILQ_LOOKUP(RTE_TAILQ_RING_NAME, rte_ring_list);
 758         if (ring_list == NULL) {
 759                 RTE_LOG(ERR, EAL, "No rte_ring tailq found!\n");
 760                 return -1;
 761         }
 762
 763         mcfg = rte_eal_get_configuration()->mem_config;
 764
 765         /* create memzones */
 766         for (i = 0; i < ivshmem_config->segment_idx && i <= RTE_MAX_MEMZONE; i++) {
 767
 768                 seg = &ivshmem_config->segment[i];
 769
 770                 /* add memzone */
 771                 if (mcfg->memzone_cnt == RTE_MAX_MEMZONE) {
 772                         RTE_LOG(ERR, EAL, "No more memory zones available!\n");
 773                         return -1;
 774                 }
 775
 776                 idx = mcfg->memzone_cnt;
 777
 778                 RTE_LOG(DEBUG, EAL, "Found memzone: '%s' at %p (len 0x%" PRIx64 ")\n",
 779                                 seg->entry.mz.name, seg->entry.mz.addr, seg->entry.mz.len);
 780
 781                 memcpy(&mcfg->memzone[idx], &seg->entry.mz,
 782                                 sizeof(struct rte_memzone));
 783
 784                 /* find ioremap address */
 785                 for (ms = 0; ms <= RTE_MAX_MEMSEG; ms++) {
 786                         if (ms == RTE_MAX_MEMSEG) {
 787                                 RTE_LOG(ERR, EAL, "Physical address of segment not found!\n");
 788                                 return -1;
 789                         }
 790                         if (CONTAINS(mcfg->memseg[ms], mcfg->memzone[idx])) {
 791                                 offset = mcfg->memzone[idx].addr_64 -
 792                                                                 mcfg->memseg[ms].addr_64;
 793                                 mcfg->memzone[idx].ioremap_addr = mcfg->memseg[ms].ioremap_addr +
 794                                                 offset;
 795                                 break;
 796                         }
 797                 }
 798
 799                 mcfg->memzone_cnt++;
 800         }
 801
 802         rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
 803
 804         /* find rings */
 805         for (i = 0; i < mcfg->memzone_cnt; i++) {
 806                 mz = &mcfg->memzone[i];
 807
 808                 /* check if memzone has a ring prefix */
 809                 if (strncmp(mz->name, RTE_RING_MZ_PREFIX,
 810                                 sizeof(RTE_RING_MZ_PREFIX) - 1) != 0)
 811                         continue;
 812
 813                 r = (struct rte_ring*) (mz->addr_64);
 814
 815                 te = rte_zmalloc("RING_TAILQ_ENTRY", sizeof(*te), 0);
 816                 if (te == NULL) {
 817                         RTE_LOG(ERR, EAL, "Cannot allocate ring tailq entry!\n");
 818                         return -1;
 819                 }
 820
 821                 te->data = (void *) r;
 822
 823                 TAILQ_INSERT_TAIL(ring_list, te, next);
 824
 825                 RTE_LOG(DEBUG, EAL, "Found ring: '%s' at %p\n", r->name, mz->addr);
 826         }
 827         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 828
 829 #ifdef RTE_LIBRTE_IVSHMEM_DEBUG
 830         rte_memzone_dump(stdout);
 831         rte_ring_list_dump(stdout);
 832 #endif
 833
 834         return 0;
 835 }
 836
 837 /* initialize ivshmem structures */
 838 int rte_eal_ivshmem_init(void)
 839 {
 840         struct rte_pci_device * dev;
 841         struct rte_pci_resource * res;
 842         int fd, ret;
 843         char path[PATH_MAX];
 844
 845         /* initialize everything to 0 */
 846         memset(path, 0, sizeof(path));
 847         ivshmem_config = NULL;
 848
 849         pagesz = getpagesize();
 850
 851         RTE_LOG(DEBUG, EAL, "Searching for IVSHMEM devices...\n");
 852
 853         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 854
 855                 if (open_shared_config() < 0) {
 856                         RTE_LOG(ERR, EAL, "Could not open IVSHMEM config!\n");
 857                         return -1;
 858                 }
 859         }
 860         else {
 861
 862                 TAILQ_FOREACH(dev, &pci_device_list, next) {
 863
 864                         if (is_ivshmem_device(dev)) {
 865
 866                                 /* IVSHMEM memory is always on BAR2 */
 867                                 res = &dev->mem_resource[2];
 868
 869                                 /* if we don't have a BAR2 */
 870                                 if (res->len == 0)
 871                                         continue;
 872
 873                                 /* construct pci device path */
 874                                 snprintf(path, sizeof(path), IVSHMEM_RESOURCE_PATH,
 875                                                 dev->addr.domain, dev->addr.bus, dev->addr.devid,
 876                                                 dev->addr.function);
 877
 878                                 /* try to find memseg */
 879                                 fd = open(path, O_RDWR);
 880                                 if (fd < 0) {
 881                                         RTE_LOG(ERR, EAL, "Could not open %s\n", path);
 882                                         return -1;
 883                                 }
 884
 885                                 /* check if it's a DPDK IVSHMEM device */
 886                                 ret = has_ivshmem_metadata(fd, res->len);
 887
 888                                 /* is DPDK device */
 889                                 if (ret == 1) {
 890
 891                                         /* config file creation is deferred until the first
 892                                          * DPDK device is found. then, it has to be created
 893                                          * only once. */
 894                                         if (ivshmem_config == NULL &&
 895                                                         create_shared_config() < 0) {
 896                                                 RTE_LOG(ERR, EAL, "Could not create IVSHMEM config!\n");
 897                                                 close(fd);
 898                                                 return -1;
 899                                         }
 900
 901                                         if (read_metadata(path, sizeof(path), fd, res->len) < 0) {
 902                                                 RTE_LOG(ERR, EAL, "Could not read metadata from"
 903                                                                 " device %02x:%02x.%x!\n", dev->addr.bus,
 904                                                                 dev->addr.devid, dev->addr.function);
 905                                                 close(fd);
 906                                                 return -1;
 907                                         }
 908
 909                                         if (ivshmem_config->pci_devs_idx == RTE_LIBRTE_IVSHMEM_MAX_PCI_DEVS) {
 910                                                 RTE_LOG(WARNING, EAL,
 911                                                                 "IVSHMEM PCI device limit exceeded. Increase "
 912                                                                 "CONFIG_RTE_LIBRTE_IVSHMEM_MAX_PCI_DEVS  in "
 913                                                                 "your config file.\n");
 914                                                 break;
 915                                         }
 916
 917                                         RTE_LOG(INFO, EAL, "Found IVSHMEM device %02x:%02x.%x\n",
 918                                                         dev->addr.bus, dev->addr.devid, dev->addr.function);
 919
 920                                         ivshmem_config->pci_devs[ivshmem_config->pci_devs_idx].ioremap_addr = res->phys_addr;
 921                                         snprintf(ivshmem_config->pci_devs[ivshmem_config->pci_devs_idx].path,
 922                                                         sizeof(ivshmem_config->pci_devs[ivshmem_config->pci_devs_idx].path),
 923                                                         "%s", path);
 924
 925                                         ivshmem_config->pci_devs_idx++;
 926                                 }
 927                                 /* failed to read */
 928                                 else if (ret < 0) {
 929                                         RTE_LOG(ERR, EAL, "Could not read IVSHMEM device: %s\n",
 930                                                         strerror(errno));
 931                                         close(fd);
 932                                         return -1;
 933                                 }
 934                                 /* not a DPDK device */
 935                                 else
 936                                         RTE_LOG(DEBUG, EAL, "Skipping non-DPDK IVSHMEM device\n");
 937
 938                                 /* close the BAR fd */
 939                                 close(fd);
 940                         }
 941                 }
 942         }
 943
 944         /* ivshmem_config is not NULL only if config was created and/or mapped */
 945         if (ivshmem_config) {
 946                 if (map_all_segments() < 0) {
 947                         RTE_LOG(ERR, EAL, "Mapping IVSHMEM segments failed!\n");
 948                         return -1;
 949                 }
 950         }
 951         else {
 952                 RTE_LOG(DEBUG, EAL, "No IVSHMEM configuration found! \n");
 953         }
 954
 955         return 0;
 956 }
 957
 958 #endif