lib/librte_ivshmem/rte_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 #include <fcntl.h>
  34 #include <limits.h>
  35 #include <unistd.h>
  36 #include <sys/mman.h>
  37 #include <string.h>
  38 #include <stdio.h>
  39
  40 #include <rte_eal_memconfig.h>
  41 #include <rte_memory.h>
  42 #include <rte_ivshmem.h>
  43 #include <rte_string_fns.h>
  44 #include <rte_common.h>
  45 #include <rte_log.h>
  46 #include <rte_debug.h>
  47 #include <rte_spinlock.h>
  48 #include <rte_common.h>
  49 #include <rte_malloc.h>
  50
  51 #include "rte_ivshmem.h"
  52
  53 #define IVSHMEM_CONFIG_FILE_FMT "/var/run/.dpdk_ivshmem_metadata_%s"
  54 #define IVSHMEM_QEMU_CMD_LINE_HEADER_FMT "-device ivshmem,size=%" PRIu64 "M,shm=fd%s"
  55 #define IVSHMEM_QEMU_CMD_FD_FMT ":%s:0x%" PRIx64 ":0x%" PRIx64
  56 #define IVSHMEM_QEMU_CMDLINE_BUFSIZE 1024
  57 #define IVSHMEM_MAX_PAGES (1 << 12)
  58 #define adjacent(x,y) (((x).phys_addr+(x).len)==(y).phys_addr)
  59 #define METADATA_SIZE_ALIGNED \
  60         (RTE_ALIGN_CEIL(sizeof(struct rte_ivshmem_metadata),pagesz))
  61
  62 #define GET_PAGEMAP_ADDR(in,addr,dlm,err)    \
  63 {                                      \
  64         char *end;                         \
  65         errno = 0;                         \
  66         addr = strtoull((in), &end, 16);   \
  67         if (errno != 0 || *end != (dlm)) { \
  68                 RTE_LOG(ERR, EAL, err);        \
  69                 goto error;                    \
  70         }                                  \
  71         (in) = end + 1;                    \
  72 }
  73
  74 static int pagesz;
  75
  76 struct memseg_cache_entry {
  77         char filepath[PATH_MAX];
  78         uint64_t offset;
  79         uint64_t len;
  80 };
  81
  82 struct ivshmem_config {
  83         struct rte_ivshmem_metadata * metadata;
  84         struct memseg_cache_entry memseg_cache[IVSHMEM_MAX_PAGES];
  85                 /**< account for multiple files per segment case */
  86         struct flock lock;
  87         rte_spinlock_t sl;
  88 };
  89
  90 static struct ivshmem_config
  91 ivshmem_global_config[RTE_LIBRTE_IVSHMEM_MAX_METADATA_FILES];
  92
  93 static rte_spinlock_t global_cfg_sl;
  94
  95 static struct ivshmem_config *
  96 get_config_by_name(const char * name)
  97 {
  98         struct rte_ivshmem_metadata * config;
  99         unsigned i;
 100
 101         for (i = 0; i < RTE_DIM(ivshmem_global_config); i++) {
 102                 config = ivshmem_global_config[i].metadata;
 103                 if (config == NULL)
 104                         return NULL;
 105                 if (strncmp(name, config->name, IVSHMEM_NAME_LEN) == 0)
 106                         return &ivshmem_global_config[i];
 107         }
 108
 109         return NULL;
 110 }
 111
 112 static int
 113 overlap(const struct rte_memzone * s1, const struct rte_memzone * s2)
 114 {
 115         uint64_t start1, end1, start2, end2;
 116
 117         start1 = s1->addr_64;
 118         end1 = s1->addr_64 + s1->len;
 119         start2 = s2->addr_64;
 120         end2 = s2->addr_64 + s2->len;
 121
 122         if (start1 >= start2 && start1 < end2)
 123                 return 1;
 124         if (start2 >= start1 && start2 < end1)
 125                 return 1;
 126
 127         return 0;
 128 }
 129
 130 static struct rte_memzone *
 131 get_memzone_by_addr(const void * addr)
 132 {
 133         struct rte_memzone * tmp, * mz;
 134         struct rte_mem_config * mcfg;
 135         int i;
 136
 137         mcfg = rte_eal_get_configuration()->mem_config;
 138         mz = NULL;
 139
 140         /* find memzone for the ring */
 141         for (i = 0; i < RTE_MAX_MEMZONE; i++) {
 142                 tmp = &mcfg->memzone[i];
 143
 144                 if (tmp->addr_64 == (uint64_t) addr) {
 145                         mz = tmp;
 146                         break;
 147                 }
 148         }
 149
 150         return mz;
 151 }
 152
 153 static int
 154 entry_compare(const void * a, const void * b)
 155 {
 156         const struct rte_ivshmem_metadata_entry * e1 =
 157                         (const struct rte_ivshmem_metadata_entry*) a;
 158         const struct rte_ivshmem_metadata_entry * e2 =
 159                         (const struct rte_ivshmem_metadata_entry*) b;
 160
 161         /* move unallocated zones to the end */
 162         if (e1->mz.addr == NULL && e2->mz.addr == NULL)
 163                 return 0;
 164         if (e1->mz.addr == 0)
 165                 return 1;
 166         if (e2->mz.addr == 0)
 167                 return -1;
 168
 169         return e1->mz.phys_addr > e2->mz.phys_addr;
 170 }
 171
 172 /* fills hugepage cache entry for a given start virt_addr */
 173 static int
 174 get_hugefile_by_virt_addr(uint64_t virt_addr, struct memseg_cache_entry * e)
 175 {
 176         uint64_t start_addr, end_addr;
 177         char *start,*path_end;
 178         char buf[PATH_MAX*2];
 179         FILE *f;
 180
 181         start = NULL;
 182         path_end = NULL;
 183         start_addr = 0;
 184
 185         memset(e->filepath, 0, sizeof(e->filepath));
 186
 187         /* open /proc/self/maps */
 188         f = fopen("/proc/self/maps", "r");
 189         if (f == NULL) {
 190                 RTE_LOG(ERR, EAL, "cannot open /proc/self/maps!\n");
 191                 return -1;
 192         }
 193
 194         /* parse maps */
 195         while (fgets(buf, sizeof(buf), f) != NULL) {
 196
 197                 /* get endptr to end of start addr */
 198                 start = buf;
 199
 200                 GET_PAGEMAP_ADDR(start,start_addr,'-',
 201                                 "Cannot find start address in maps!\n");
 202
 203                 /* if start address is bigger than our address, skip */
 204                 if (start_addr > virt_addr)
 205                         continue;
 206
 207                 GET_PAGEMAP_ADDR(start,end_addr,' ',
 208                                 "Cannot find end address in maps!\n");
 209
 210                 /* if end address is less than our address, skip */
 211                 if (end_addr <= virt_addr)
 212                         continue;
 213
 214                 /* find where the path starts */
 215                 start = strstr(start, "/");
 216
 217                 if (start == NULL)
 218                         continue;
 219
 220                 /* at this point, we know that this is our map.
 221                  * now let's find the file */
 222                 path_end = strstr(start, "\n");
 223                 break;
 224         }
 225
 226         if (path_end == NULL) {
 227                 RTE_LOG(ERR, EAL, "Hugefile path not found!\n");
 228                 goto error;
 229         }
 230
 231         /* calculate offset and copy the file path */
 232         snprintf(e->filepath, RTE_PTR_DIFF(path_end, start) + 1, "%s", start);
 233
 234         e->offset = virt_addr - start_addr;
 235
 236         fclose(f);
 237
 238         return 0;
 239 error:
 240         fclose(f);
 241         return -1;
 242 }
 243
 244 /*
 245  * This is a complex function. What it does is the following:
 246  *  1. Goes through metadata and gets list of hugepages involved
 247  *  2. Sorts the hugepages by size (1G first)
 248  *  3. Goes through metadata again and writes correct offsets
 249  *  4. Goes through pages and finds out their filenames, offsets etc.
 250  */
 251 static int
 252 build_config(struct rte_ivshmem_metadata * metadata)
 253 {
 254         struct rte_ivshmem_metadata_entry * e_local;
 255         struct memseg_cache_entry * ms_local;
 256         struct rte_memseg pages[IVSHMEM_MAX_PAGES];
 257         struct rte_ivshmem_metadata_entry *entry;
 258         struct memseg_cache_entry * c_entry, * prev_entry;
 259         struct ivshmem_config * config;
 260         unsigned i, j, mz_iter, ms_iter;
 261         uint64_t biggest_len;
 262         int biggest_idx;
 263
 264         /* return error if we try to use an unknown config file */
 265         config = get_config_by_name(metadata->name);
 266         if (config == NULL) {
 267                 RTE_LOG(ERR, EAL, "Cannot find IVSHMEM config %s!\n", metadata->name);
 268                 goto fail_e;
 269         }
 270
 271         memset(pages, 0, sizeof(pages));
 272
 273         e_local = malloc(sizeof(config->metadata->entry));
 274         if (e_local == NULL)
 275                 goto fail_e;
 276         ms_local = malloc(sizeof(config->memseg_cache));
 277         if (ms_local == NULL)
 278                 goto fail_ms;
 279
 280
 281         /* make local copies before doing anything */
 282         memcpy(e_local, config->metadata->entry, sizeof(config->metadata->entry));
 283         memcpy(ms_local, config->memseg_cache, sizeof(config->memseg_cache));
 284
 285         qsort(e_local, RTE_DIM(config->metadata->entry), sizeof(struct rte_ivshmem_metadata_entry),
 286                         entry_compare);
 287
 288         /* first pass - collect all huge pages */
 289         for (mz_iter = 0; mz_iter < RTE_DIM(config->metadata->entry); mz_iter++) {
 290
 291                 entry = &e_local[mz_iter];
 292
 293                 uint64_t start_addr = RTE_ALIGN_FLOOR(entry->mz.addr_64,
 294                                 entry->mz.hugepage_sz);
 295                 uint64_t offset = entry->mz.addr_64 - start_addr;
 296                 uint64_t len = RTE_ALIGN_CEIL(entry->mz.len + offset,
 297                                 entry->mz.hugepage_sz);
 298
 299                 if (entry->mz.addr_64 == 0 || start_addr == 0 || len == 0)
 300                         continue;
 301
 302                 int start_page;
 303
 304                 /* find first unused page - mz are phys_addr sorted so we don't have to
 305                  * look out for holes */
 306                 for (i = 0; i < RTE_DIM(pages); i++) {
 307
 308                         /* skip if we already have this page */
 309                         if (pages[i].addr_64 == start_addr) {
 310                                 start_addr += entry->mz.hugepage_sz;
 311                                 len -= entry->mz.hugepage_sz;
 312                                 continue;
 313                         }
 314                         /* we found a new page */
 315                         else if (pages[i].addr_64 == 0) {
 316                                 start_page = i;
 317                                 break;
 318                         }
 319                 }
 320                 if (i == RTE_DIM(pages)) {
 321                         RTE_LOG(ERR, EAL, "Cannot find unused page!\n");
 322                         goto fail;
 323                 }
 324
 325                 /* populate however many pages the memzone has */
 326                 for (i = start_page; i < RTE_DIM(pages) && len != 0; i++) {
 327
 328                         pages[i].addr_64 = start_addr;
 329                         pages[i].len = entry->mz.hugepage_sz;
 330                         start_addr += entry->mz.hugepage_sz;
 331                         len -= entry->mz.hugepage_sz;
 332                 }
 333                 /* if there's still length left */
 334                 if (len != 0) {
 335                         RTE_LOG(ERR, EAL, "Not enough space for pages!\n");
 336                         goto fail;
 337                 }
 338         }
 339
 340         /* second pass - sort pages by size */
 341         for (i = 0; i < RTE_DIM(pages); i++) {
 342
 343                 if (pages[i].addr == NULL)
 344                         break;
 345
 346                 biggest_len = 0;
 347                 biggest_idx = -1;
 348
 349                 /*
 350                  * browse all entries starting at 'i', and find the
 351                  * entry with the smallest addr
 352                  */
 353                 for (j=i; j< RTE_DIM(pages); j++) {
 354                         if (pages[j].addr == NULL)
 355                                         break;
 356                         if (biggest_len == 0 ||
 357                                 pages[j].len > biggest_len) {
 358                                 biggest_len = pages[j].len;
 359                                 biggest_idx = j;
 360                         }
 361                 }
 362
 363                 /* should not happen */
 364                 if (biggest_idx == -1) {
 365                         RTE_LOG(ERR, EAL, "Error sorting by size!\n");
 366                         goto fail;
 367                 }
 368                 if (i != (unsigned) biggest_idx) {
 369                         struct rte_memseg tmp;
 370
 371                         memcpy(&tmp, &pages[biggest_idx], sizeof(struct rte_memseg));
 372
 373                         /* we don't want to break contiguousness, so instead of just
 374                          * swapping segments, we move all the preceding segments to the
 375                          * right and then put the old segment @ biggest_idx in place of
 376                          * segment @ i */
 377                         for (j = biggest_idx - 1; j >= i; j--) {
 378                                 memcpy(&pages[j+1], &pages[j], sizeof(struct rte_memseg));
 379                                 memset(&pages[j], 0, sizeof(struct rte_memseg));
 380                                 if (j == 0)
 381                                         break;
 382                         }
 383
 384                         /* put old biggest segment to its new place */
 385                         memcpy(&pages[i], &tmp, sizeof(struct rte_memseg));
 386                 }
 387         }
 388
 389         /* third pass - write correct offsets */
 390         for (mz_iter = 0; mz_iter < RTE_DIM(config->metadata->entry); mz_iter++) {
 391
 392                 uint64_t offset = 0;
 393
 394                 entry = &e_local[mz_iter];
 395
 396                 if (entry->mz.addr_64 == 0)
 397                         break;
 398
 399                 /* find page for current memzone */
 400                 for (i = 0; i < RTE_DIM(pages); i++) {
 401                         /* we found our page */
 402                         if (entry->mz.addr_64 >= pages[i].addr_64 &&
 403                                         entry->mz.addr_64 < pages[i].addr_64 + pages[i].len) {
 404                                 entry->offset = (entry->mz.addr_64 - pages[i].addr_64) +
 405                                                 offset;
 406                                 break;
 407                         }
 408                         offset += pages[i].len;
 409                 }
 410                 if (i == RTE_DIM(pages)) {
 411                         RTE_LOG(ERR, EAL, "Page not found!\n");
 412                         goto fail;
 413                 }
 414         }
 415
 416         ms_iter = 0;
 417         prev_entry = NULL;
 418
 419         /* fourth pass - create proper memseg cache */
 420         for (i = 0; i < RTE_DIM(pages) &&
 421                         ms_iter <= RTE_DIM(config->memseg_cache); i++) {
 422                 if (pages[i].addr_64 == 0)
 423                         break;
 424
 425
 426                 if (ms_iter == RTE_DIM(pages)) {
 427                         RTE_LOG(ERR, EAL, "The universe has collapsed!\n");
 428                         goto fail;
 429                 }
 430
 431                 c_entry = &ms_local[ms_iter];
 432                 c_entry->len = pages[i].len;
 433
 434                 if (get_hugefile_by_virt_addr(pages[i].addr_64, c_entry) < 0)
 435                         goto fail;
 436
 437                 /* if previous entry has the same filename and is contiguous,
 438                  * clear current entry and increase previous entry's length
 439                  */
 440                 if (prev_entry != NULL &&
 441                                 strncmp(c_entry->filepath, prev_entry->filepath,
 442                                 sizeof(c_entry->filepath)) == 0 &&
 443                                 prev_entry->offset + prev_entry->len == c_entry->offset) {
 444                         prev_entry->len += pages[i].len;
 445                         memset(c_entry, 0, sizeof(struct memseg_cache_entry));
 446                 }
 447                 else {
 448                         prev_entry = c_entry;
 449                         ms_iter++;
 450                 }
 451         }
 452
 453         /* update current configuration with new valid data */
 454         memcpy(config->metadata->entry, e_local, sizeof(config->metadata->entry));
 455         memcpy(config->memseg_cache, ms_local, sizeof(config->memseg_cache));
 456
 457         free(ms_local);
 458         free(e_local);
 459
 460         return 0;
 461 fail:
 462         free(ms_local);
 463 fail_ms:
 464         free(e_local);
 465 fail_e:
 466         return -1;
 467 }
 468
 469 static int
 470 add_memzone_to_metadata(const struct rte_memzone * mz,
 471                 struct ivshmem_config * config)
 472 {
 473         struct rte_ivshmem_metadata_entry * entry;
 474         unsigned i;
 475
 476         rte_spinlock_lock(&config->sl);
 477
 478         /* find free slot in this config */
 479         for (i = 0; i < RTE_DIM(config->metadata->entry); i++) {
 480                 entry = &config->metadata->entry[i];
 481
 482                 if (&entry->mz.addr_64 != 0 && overlap(mz, &entry->mz)) {
 483                         RTE_LOG(ERR, EAL, "Overlapping memzones!\n");
 484                         goto fail;
 485                 }
 486
 487                 /* if addr is zero, the memzone is probably free */
 488                 if (entry->mz.addr_64 == 0) {
 489                         RTE_LOG(DEBUG, EAL, "Adding memzone '%s' at %p to metadata %s\n",
 490                                         mz->name, mz->addr, config->metadata->name);
 491                         memcpy(&entry->mz, mz, sizeof(struct rte_memzone));
 492
 493                         /* run config file parser */
 494                         if (build_config(config->metadata) < 0)
 495                                 goto fail;
 496
 497                         break;
 498                 }
 499         }
 500
 501         /* if we reached the maximum, that means we have no place in config */
 502         if (i == RTE_DIM(config->metadata->entry)) {
 503                 RTE_LOG(ERR, EAL, "No space left in IVSHMEM metadata %s!\n",
 504                                 config->metadata->name);
 505                 goto fail;
 506         }
 507 #ifdef RTE_LIBRTE_IVSHMEM
 508         struct rte_mem_config *mcfg;
 509         unsigned int idx;
 510
 511         mcfg = rte_eal_get_configuration()->mem_config;
 512
 513         rte_rwlock_write_lock(&mcfg->mlock);
 514
 515         idx = ((uintptr_t)mz - (uintptr_t)mcfg->memzone);
 516         idx = idx / sizeof(struct rte_memzone);
 517
 518         /* mark the memzone not freeable */
 519         mcfg->memzone[idx].ioremap_addr = mz->phys_addr;
 520
 521         rte_rwlock_write_unlock(&mcfg->mlock);
 522 #endif
 523         rte_spinlock_unlock(&config->sl);
 524         return 0;
 525 fail:
 526         rte_spinlock_unlock(&config->sl);
 527         return -1;
 528 }
 529
 530 static int
 531 add_ring_to_metadata(const struct rte_ring * r,
 532                 struct ivshmem_config * config)
 533 {
 534         struct rte_memzone * mz;
 535
 536         mz = get_memzone_by_addr(r);
 537
 538         if (!mz) {
 539                 RTE_LOG(ERR, EAL, "Cannot find memzone for ring!\n");
 540                 return -1;
 541         }
 542
 543         return add_memzone_to_metadata(mz, config);
 544 }
 545
 546 static int
 547 add_mempool_to_metadata(const struct rte_mempool * mp,
 548                 struct ivshmem_config * config)
 549 {
 550         struct rte_memzone * mz;
 551         int ret;
 552
 553         mz = get_memzone_by_addr(mp);
 554         ret = 0;
 555
 556         if (!mz) {
 557                 RTE_LOG(ERR, EAL, "Cannot find memzone for mempool!\n");
 558                 return -1;
 559         }
 560
 561         /* mempool consists of memzone and ring */
 562         ret = add_memzone_to_metadata(mz, config);
 563         if (ret < 0)
 564                 return -1;
 565
 566         return add_ring_to_metadata(mp->ring, config);
 567 }
 568
 569 int
 570 rte_ivshmem_metadata_add_ring(const struct rte_ring * r, const char * name)
 571 {
 572         struct ivshmem_config * config;
 573
 574         if (name == NULL || r == NULL)
 575                 return -1;
 576
 577         config = get_config_by_name(name);
 578
 579         if (config == NULL) {
 580                 RTE_LOG(ERR, EAL, "Cannot find IVSHMEM config %s!\n", name);
 581                 return -1;
 582         }
 583
 584         return add_ring_to_metadata(r, config);
 585 }
 586
 587 int
 588 rte_ivshmem_metadata_add_memzone(const struct rte_memzone * mz, const char * name)
 589 {
 590         struct ivshmem_config * config;
 591
 592         if (name == NULL || mz == NULL)
 593                 return -1;
 594
 595         config = get_config_by_name(name);
 596
 597         if (config == NULL) {
 598                 RTE_LOG(ERR, EAL, "Cannot find IVSHMEM config %s!\n", name);
 599                 return -1;
 600         }
 601
 602         return add_memzone_to_metadata(mz, config);
 603 }
 604
 605 int
 606 rte_ivshmem_metadata_add_mempool(const struct rte_mempool * mp, const char * name)
 607 {
 608         struct ivshmem_config * config;
 609
 610         if (name == NULL || mp == NULL)
 611                 return -1;
 612
 613         config = get_config_by_name(name);
 614
 615         if (config == NULL) {
 616                 RTE_LOG(ERR, EAL, "Cannot find IVSHMEM config %s!\n", name);
 617                 return -1;
 618         }
 619
 620         return add_mempool_to_metadata(mp, config);
 621 }
 622
 623 static inline void
 624 ivshmem_config_path(char *buffer, size_t bufflen, const char *name)
 625 {
 626         snprintf(buffer, bufflen, IVSHMEM_CONFIG_FILE_FMT, name);
 627 }
 628
 629
 630
 631 static inline
 632 void *ivshmem_metadata_create(const char *name, size_t size,
 633                 struct flock *lock)
 634 {
 635         int retval, fd;
 636         void *metadata_addr;
 637         char pathname[PATH_MAX];
 638
 639         ivshmem_config_path(pathname, sizeof(pathname), name);
 640
 641         fd = open(pathname, O_RDWR | O_CREAT, 0660);
 642         if (fd < 0) {
 643                 RTE_LOG(ERR, EAL, "Cannot open '%s'\n", pathname);
 644                 return NULL;
 645         }
 646
 647         size = METADATA_SIZE_ALIGNED;
 648
 649         retval = fcntl(fd, F_SETLK, lock);
 650         if (retval < 0){
 651                 close(fd);
 652                 RTE_LOG(ERR, EAL, "Cannot create lock on '%s'. Is another "
 653                                 "process using it?\n", pathname);
 654                 return NULL;
 655         }
 656
 657         retval = ftruncate(fd, size);
 658         if (retval < 0){
 659                 close(fd);
 660                 RTE_LOG(ERR, EAL, "Cannot resize '%s'\n", pathname);
 661                 return NULL;
 662         }
 663
 664         metadata_addr = mmap(NULL, size,
 665                                 PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 666
 667         if (metadata_addr == MAP_FAILED){
 668                 RTE_LOG(ERR, EAL, "Cannot mmap memory for '%s'\n", pathname);
 669
 670                 /* we don't care if we can't unlock */
 671                 fcntl(fd, F_UNLCK, lock);
 672                 close(fd);
 673
 674                 return NULL;
 675         }
 676
 677         return metadata_addr;
 678 }
 679
 680 int rte_ivshmem_metadata_create(const char *name)
 681 {
 682         struct ivshmem_config * ivshmem_config;
 683         unsigned index;
 684
 685         if (pagesz == 0)
 686                 pagesz = getpagesize();
 687
 688         if (name == NULL)
 689                 return -1;
 690
 691         rte_spinlock_lock(&global_cfg_sl);
 692
 693         for (index = 0; index < RTE_DIM(ivshmem_global_config); index++) {
 694                 if (ivshmem_global_config[index].metadata == NULL) {
 695                         ivshmem_config = &ivshmem_global_config[index];
 696                         break;
 697                 }
 698         }
 699
 700         if (index == RTE_DIM(ivshmem_global_config)) {
 701                 RTE_LOG(ERR, EAL, "Cannot create more ivshmem config files. "
 702                 "Maximum has been reached\n");
 703                 rte_spinlock_unlock(&global_cfg_sl);
 704                 return -1;
 705         }
 706
 707         ivshmem_config->lock.l_type = F_WRLCK;
 708         ivshmem_config->lock.l_whence = SEEK_SET;
 709
 710         ivshmem_config->lock.l_start = 0;
 711         ivshmem_config->lock.l_len = METADATA_SIZE_ALIGNED;
 712
 713         ivshmem_global_config[index].metadata = ((struct rte_ivshmem_metadata *)
 714                         ivshmem_metadata_create(
 715                                         name,
 716                                         sizeof(struct rte_ivshmem_metadata),
 717                                         &ivshmem_config->lock));
 718
 719         if (ivshmem_global_config[index].metadata == NULL) {
 720                 rte_spinlock_unlock(&global_cfg_sl);
 721                 return -1;
 722         }
 723
 724         /* Metadata setup */
 725         memset(ivshmem_config->metadata, 0, sizeof(struct rte_ivshmem_metadata));
 726         ivshmem_config->metadata->magic_number = IVSHMEM_MAGIC;
 727         snprintf(ivshmem_config->metadata->name,
 728                         sizeof(ivshmem_config->metadata->name), "%s", name);
 729
 730         rte_spinlock_unlock(&global_cfg_sl);
 731
 732         return 0;
 733 }
 734
 735 int
 736 rte_ivshmem_metadata_cmdline_generate(char *buffer, unsigned size, const char *name)
 737 {
 738         const struct memseg_cache_entry * ms_cache, *entry;
 739         struct ivshmem_config * config;
 740         char cmdline[IVSHMEM_QEMU_CMDLINE_BUFSIZE], *cmdline_ptr;
 741         char cfg_file_path[PATH_MAX];
 742         unsigned remaining_len, tmplen, iter;
 743         uint64_t shared_mem_size, zero_size, total_size;
 744
 745         if (buffer == NULL || name == NULL)
 746                 return -1;
 747
 748         config = get_config_by_name(name);
 749
 750         if (config == NULL) {
 751                 RTE_LOG(ERR, EAL, "Config %s not found!\n", name);
 752                 return -1;
 753         }
 754
 755         rte_spinlock_lock(&config->sl);
 756
 757         /* prepare metadata file path */
 758         snprintf(cfg_file_path, sizeof(cfg_file_path), IVSHMEM_CONFIG_FILE_FMT,
 759                         config->metadata->name);
 760
 761         ms_cache = config->memseg_cache;
 762
 763         cmdline_ptr = cmdline;
 764         remaining_len = sizeof(cmdline);
 765
 766         shared_mem_size = 0;
 767         iter = 0;
 768
 769         while ((ms_cache[iter].len != 0) && (iter < RTE_DIM(config->metadata->entry))) {
 770
 771                 entry = &ms_cache[iter];
 772
 773                 /* Offset and sizes within the current pathname */
 774                 tmplen = snprintf(cmdline_ptr, remaining_len, IVSHMEM_QEMU_CMD_FD_FMT,
 775                                 entry->filepath, entry->offset, entry->len);
 776
 777                 shared_mem_size += entry->len;
 778
 779                 cmdline_ptr = RTE_PTR_ADD(cmdline_ptr, tmplen);
 780                 remaining_len -= tmplen;
 781
 782                 if (remaining_len == 0) {
 783                         RTE_LOG(ERR, EAL, "Command line too long!\n");
 784                         rte_spinlock_unlock(&config->sl);
 785                         return -1;
 786                 }
 787
 788                 iter++;
 789         }
 790
 791         total_size = rte_align64pow2(shared_mem_size + METADATA_SIZE_ALIGNED);
 792         zero_size = total_size - shared_mem_size - METADATA_SIZE_ALIGNED;
 793
 794         /* add /dev/zero to command-line to fill the space */
 795         tmplen = snprintf(cmdline_ptr, remaining_len, IVSHMEM_QEMU_CMD_FD_FMT,
 796                         "/dev/zero",
 797                         (uint64_t)0x0,
 798                         zero_size);
 799
 800         cmdline_ptr = RTE_PTR_ADD(cmdline_ptr, tmplen);
 801         remaining_len -= tmplen;
 802
 803         if (remaining_len == 0) {
 804                 RTE_LOG(ERR, EAL, "Command line too long!\n");
 805                 rte_spinlock_unlock(&config->sl);
 806                 return -1;
 807         }
 808
 809         /* add metadata file to the end of command-line */
 810         tmplen = snprintf(cmdline_ptr, remaining_len, IVSHMEM_QEMU_CMD_FD_FMT,
 811                         cfg_file_path,
 812                         (uint64_t)0x0,
 813                         METADATA_SIZE_ALIGNED);
 814
 815         cmdline_ptr = RTE_PTR_ADD(cmdline_ptr, tmplen);
 816         remaining_len -= tmplen;
 817
 818         if (remaining_len == 0) {
 819                 RTE_LOG(ERR, EAL, "Command line too long!\n");
 820                 rte_spinlock_unlock(&config->sl);
 821                 return -1;
 822         }
 823
 824         /* if current length of the command line is bigger than the buffer supplied
 825          * by the user, or if command-line is bigger than what IVSHMEM accepts */
 826         if ((sizeof(cmdline) - remaining_len) > size) {
 827                 RTE_LOG(ERR, EAL, "Buffer is too short!\n");
 828                 rte_spinlock_unlock(&config->sl);
 829                 return -1;
 830         }
 831         /* complete the command-line */
 832         snprintf(buffer, size,
 833                         IVSHMEM_QEMU_CMD_LINE_HEADER_FMT,
 834                         total_size >> 20,
 835                         cmdline);
 836
 837         rte_spinlock_unlock(&config->sl);
 838
 839         return 0;
 840 }
 841
 842 void
 843 rte_ivshmem_metadata_dump(FILE *f, const char *name)
 844 {
 845         unsigned i = 0;
 846         struct ivshmem_config * config;
 847         struct rte_ivshmem_metadata_entry *entry;
 848 #ifdef RTE_LIBRTE_IVSHMEM_DEBUG
 849         uint64_t addr;
 850         uint64_t end, hugepage_sz;
 851         struct memseg_cache_entry e;
 852 #endif
 853
 854         if (name == NULL)
 855                 return;
 856
 857         /* return error if we try to use an unknown config file */
 858         config = get_config_by_name(name);
 859         if (config == NULL) {
 860                 RTE_LOG(ERR, EAL, "Cannot find IVSHMEM config %s!\n", name);
 861                 return;
 862         }
 863
 864         rte_spinlock_lock(&config->sl);
 865
 866         entry = &config->metadata->entry[0];
 867
 868         while (entry->mz.addr != NULL && i < RTE_DIM(config->metadata->entry)) {
 869
 870                 fprintf(f, "Entry %u: name:<%-20s>, phys:0x%-15lx, len:0x%-15lx, "
 871                         "virt:%-15p, off:0x%-15lx\n",
 872                         i,
 873                         entry->mz.name,
 874                         entry->mz.phys_addr,
 875                         entry->mz.len,
 876                         entry->mz.addr,
 877                         entry->offset);
 878                 i++;
 879
 880 #ifdef RTE_LIBRTE_IVSHMEM_DEBUG
 881                 fprintf(f, "\tHugepage files:\n");
 882
 883                 hugepage_sz = entry->mz.hugepage_sz;
 884                 addr = RTE_ALIGN_FLOOR(entry->mz.addr_64, hugepage_sz);
 885                 end = addr + RTE_ALIGN_CEIL(entry->mz.len + (entry->mz.addr_64 - addr),
 886                                 hugepage_sz);
 887
 888                 for (; addr < end; addr += hugepage_sz) {
 889                         memset(&e, 0, sizeof(e));
 890
 891                         get_hugefile_by_virt_addr(addr, &e);
 892
 893                         fprintf(f, "\t0x%"PRIx64 "-0x%" PRIx64 " offset: 0x%" PRIx64 " %s\n",
 894                                         addr, addr + hugepage_sz, e.offset, e.filepath);
 895                 }
 896 #endif
 897                 entry++;
 898         }
 899
 900         rte_spinlock_unlock(&config->sl);
 901 }