hw/i386/xen/xen-mapcache.c

   1 /*
   2  * Copyright (C) 2011       Citrix Ltd.
   3  *
   4  * This work is licensed under the terms of the GNU GPL, version 2.  See
   5  * the COPYING file in the top-level directory.
   6  *
   7  * Contributions after 2012-01-13 are licensed under the terms of the
   8  * GNU GPL, version 2 or (at your option) any later version.
   9  */
  10
  11 #include "qemu/osdep.h"
  12 #include "qemu/units.h"
  13 #include "qemu/error-report.h"
  14
  15 #include <sys/resource.h>
  16
  17 #include "hw/xen/xen-legacy-backend.h"
  18 #include "qemu/bitmap.h"
  19
  20 #include "sysemu/runstate.h"
  21 #include "sysemu/xen-mapcache.h"
  22 #include "trace.h"
  23
  24
  25 //#define MAPCACHE_DEBUG
  26
  27 #ifdef MAPCACHE_DEBUG
  28 #  define DPRINTF(fmt, ...) do { \
  29     fprintf(stderr, "xen_mapcache: " fmt, ## __VA_ARGS__); \
  30 } while (0)
  31 #else
  32 #  define DPRINTF(fmt, ...) do { } while (0)
  33 #endif
  34
  35 #if HOST_LONG_BITS == 32
  36 #  define MCACHE_BUCKET_SHIFT 16
  37 #  define MCACHE_MAX_SIZE     (1UL<<31) /* 2GB Cap */
  38 #else
  39 #  define MCACHE_BUCKET_SHIFT 20
  40 #  define MCACHE_MAX_SIZE     (1UL<<35) /* 32GB Cap */
  41 #endif
  42 #define MCACHE_BUCKET_SIZE (1UL << MCACHE_BUCKET_SHIFT)
  43
  44 /* This is the size of the virtual address space reserve to QEMU that will not
  45  * be use by MapCache.
  46  * From empirical tests I observed that qemu use 75MB more than the
  47  * max_mcache_size.
  48  */
  49 #define NON_MCACHE_MEMORY_SIZE (80 * MiB)
  50
  51 typedef struct MapCacheEntry {
  52     hwaddr paddr_index;
  53     uint8_t *vaddr_base;
  54     unsigned long *valid_mapping;
  55     uint8_t lock;
  56 #define XEN_MAPCACHE_ENTRY_DUMMY (1 << 0)
  57     uint8_t flags;
  58     hwaddr size;
  59     struct MapCacheEntry *next;
  60 } MapCacheEntry;
  61
  62 typedef struct MapCacheRev {
  63     uint8_t *vaddr_req;
  64     hwaddr paddr_index;
  65     hwaddr size;
  66     QTAILQ_ENTRY(MapCacheRev) next;
  67     bool dma;
  68 } MapCacheRev;
  69
  70 typedef struct MapCache {
  71     MapCacheEntry *entry;
  72     unsigned long nr_buckets;
  73     QTAILQ_HEAD(, MapCacheRev) locked_entries;
  74
  75     /* For most cases (>99.9%), the page address is the same. */
  76     MapCacheEntry *last_entry;
  77     unsigned long max_mcache_size;
  78     unsigned int mcache_bucket_shift;
  79
  80     phys_offset_to_gaddr_t phys_offset_to_gaddr;
  81     QemuMutex lock;
  82     void *opaque;
  83 } MapCache;
  84
  85 static MapCache *mapcache;
  86
  87 static inline void mapcache_lock(void)
  88 {
  89     qemu_mutex_lock(&mapcache->lock);
  90 }
  91
  92 static inline void mapcache_unlock(void)
  93 {
  94     qemu_mutex_unlock(&mapcache->lock);
  95 }
  96
  97 static inline int test_bits(int nr, int size, const unsigned long *addr)
  98 {
  99     unsigned long res = find_next_zero_bit(addr, size + nr, nr);
 100     if (res >= nr + size)
 101         return 1;
 102     else
 103         return 0;
 104 }
 105
 106 void xen_map_cache_init(phys_offset_to_gaddr_t f, void *opaque)
 107 {
 108     unsigned long size;
 109     struct rlimit rlimit_as;
 110
 111     mapcache = g_malloc0(sizeof (MapCache));
 112
 113     mapcache->phys_offset_to_gaddr = f;
 114     mapcache->opaque = opaque;
 115     qemu_mutex_init(&mapcache->lock);
 116
 117     QTAILQ_INIT(&mapcache->locked_entries);
 118
 119     if (geteuid() == 0) {
 120         rlimit_as.rlim_cur = RLIM_INFINITY;
 121         rlimit_as.rlim_max = RLIM_INFINITY;
 122         mapcache->max_mcache_size = MCACHE_MAX_SIZE;
 123     } else {
 124         getrlimit(RLIMIT_AS, &rlimit_as);
 125         rlimit_as.rlim_cur = rlimit_as.rlim_max;
 126
 127         if (rlimit_as.rlim_max != RLIM_INFINITY) {
 128             warn_report("QEMU's maximum size of virtual"
 129                         " memory is not infinity");
 130         }
 131         if (rlimit_as.rlim_max < MCACHE_MAX_SIZE + NON_MCACHE_MEMORY_SIZE) {
 132             mapcache->max_mcache_size = rlimit_as.rlim_max -
 133                 NON_MCACHE_MEMORY_SIZE;
 134         } else {
 135             mapcache->max_mcache_size = MCACHE_MAX_SIZE;
 136         }
 137     }
 138
 139     setrlimit(RLIMIT_AS, &rlimit_as);
 140
 141     mapcache->nr_buckets =
 142         (((mapcache->max_mcache_size >> XC_PAGE_SHIFT) +
 143           (1UL << (MCACHE_BUCKET_SHIFT - XC_PAGE_SHIFT)) - 1) >>
 144          (MCACHE_BUCKET_SHIFT - XC_PAGE_SHIFT));
 145
 146     size = mapcache->nr_buckets * sizeof (MapCacheEntry);
 147     size = (size + XC_PAGE_SIZE - 1) & ~(XC_PAGE_SIZE - 1);
 148     DPRINTF("%s, nr_buckets = %lx size %lu\n", __func__,
 149             mapcache->nr_buckets, size);
 150     mapcache->entry = g_malloc0(size);
 151 }
 152
 153 static void xen_remap_bucket(MapCacheEntry *entry,
 154                              void *vaddr,
 155                              hwaddr size,
 156                              hwaddr address_index,
 157                              bool dummy)
 158 {
 159     uint8_t *vaddr_base;
 160     xen_pfn_t *pfns;
 161     int *err;
 162     unsigned int i;
 163     hwaddr nb_pfn = size >> XC_PAGE_SHIFT;
 164
 165     trace_xen_remap_bucket(address_index);
 166
 167     pfns = g_malloc0(nb_pfn * sizeof (xen_pfn_t));
 168     err = g_malloc0(nb_pfn * sizeof (int));
 169
 170     if (entry->vaddr_base != NULL) {
 171         if (!(entry->flags & XEN_MAPCACHE_ENTRY_DUMMY)) {
 172             ram_block_notify_remove(entry->vaddr_base, entry->size);
 173         }
 174         if (munmap(entry->vaddr_base, entry->size) != 0) {
 175             perror("unmap fails");
 176             exit(-1);
 177         }
 178     }
 179     g_free(entry->valid_mapping);
 180     entry->valid_mapping = NULL;
 181
 182     for (i = 0; i < nb_pfn; i++) {
 183         pfns[i] = (address_index << (MCACHE_BUCKET_SHIFT-XC_PAGE_SHIFT)) + i;
 184     }
 185
 186     /*
 187      * If the caller has requested the mapping at a specific address use
 188      * MAP_FIXED to make sure it's honored.
 189      */
 190     if (!dummy) {
 191         vaddr_base = xenforeignmemory_map2(xen_fmem, xen_domid, vaddr,
 192                                            PROT_READ | PROT_WRITE,
 193                                            vaddr ? MAP_FIXED : 0,
 194                                            nb_pfn, pfns, err);
 195         if (vaddr_base == NULL) {
 196             perror("xenforeignmemory_map2");
 197             exit(-1);
 198         }
 199     } else {
 200         /*
 201          * We create dummy mappings where we are unable to create a foreign
 202          * mapping immediately due to certain circumstances (i.e. on resume now)
 203          */
 204         vaddr_base = mmap(vaddr, size, PROT_READ | PROT_WRITE,
 205                           MAP_ANON | MAP_SHARED | (vaddr ? MAP_FIXED : 0),
 206                           -1, 0);
 207         if (vaddr_base == MAP_FAILED) {
 208             perror("mmap");
 209             exit(-1);
 210         }
 211     }
 212
 213     if (!(entry->flags & XEN_MAPCACHE_ENTRY_DUMMY)) {
 214         ram_block_notify_add(vaddr_base, size);
 215     }
 216
 217     entry->vaddr_base = vaddr_base;
 218     entry->paddr_index = address_index;
 219     entry->size = size;
 220     entry->valid_mapping = (unsigned long *) g_malloc0(sizeof(unsigned long) *
 221             BITS_TO_LONGS(size >> XC_PAGE_SHIFT));
 222
 223     if (dummy) {
 224         entry->flags |= XEN_MAPCACHE_ENTRY_DUMMY;
 225     } else {
 226         entry->flags &= ~(XEN_MAPCACHE_ENTRY_DUMMY);
 227     }
 228
 229     bitmap_zero(entry->valid_mapping, nb_pfn);
 230     for (i = 0; i < nb_pfn; i++) {
 231         if (!err[i]) {
 232             bitmap_set(entry->valid_mapping, i, 1);
 233         }
 234     }
 235
 236     g_free(pfns);
 237     g_free(err);
 238 }
 239
 240 static uint8_t *xen_map_cache_unlocked(hwaddr phys_addr, hwaddr size,
 241                                        uint8_t lock, bool dma)
 242 {
 243     MapCacheEntry *entry, *pentry = NULL,
 244                   *free_entry = NULL, *free_pentry = NULL;
 245     hwaddr address_index;
 246     hwaddr address_offset;
 247     hwaddr cache_size = size;
 248     hwaddr test_bit_size;
 249     bool translated G_GNUC_UNUSED = false;
 250     bool dummy = false;
 251
 252 tryagain:
 253     address_index  = phys_addr >> MCACHE_BUCKET_SHIFT;
 254     address_offset = phys_addr & (MCACHE_BUCKET_SIZE - 1);
 255
 256     trace_xen_map_cache(phys_addr);
 257
 258     /* test_bit_size is always a multiple of XC_PAGE_SIZE */
 259     if (size) {
 260         test_bit_size = size + (phys_addr & (XC_PAGE_SIZE - 1));
 261
 262         if (test_bit_size % XC_PAGE_SIZE) {
 263             test_bit_size += XC_PAGE_SIZE - (test_bit_size % XC_PAGE_SIZE);
 264         }
 265     } else {
 266         test_bit_size = XC_PAGE_SIZE;
 267     }
 268
 269     if (mapcache->last_entry != NULL &&
 270         mapcache->last_entry->paddr_index == address_index &&
 271         !lock && !size &&
 272         test_bits(address_offset >> XC_PAGE_SHIFT,
 273                   test_bit_size >> XC_PAGE_SHIFT,
 274                   mapcache->last_entry->valid_mapping)) {
 275         trace_xen_map_cache_return(mapcache->last_entry->vaddr_base + address_offset);
 276         return mapcache->last_entry->vaddr_base + address_offset;
 277     }
 278
 279     /* size is always a multiple of MCACHE_BUCKET_SIZE */
 280     if (size) {
 281         cache_size = size + address_offset;
 282         if (cache_size % MCACHE_BUCKET_SIZE) {
 283             cache_size += MCACHE_BUCKET_SIZE - (cache_size % MCACHE_BUCKET_SIZE);
 284         }
 285     } else {
 286         cache_size = MCACHE_BUCKET_SIZE;
 287     }
 288
 289     entry = &mapcache->entry[address_index % mapcache->nr_buckets];
 290
 291     while (entry && (lock || entry->lock) && entry->vaddr_base &&
 292             (entry->paddr_index != address_index || entry->size != cache_size ||
 293              !test_bits(address_offset >> XC_PAGE_SHIFT,
 294                  test_bit_size >> XC_PAGE_SHIFT,
 295                  entry->valid_mapping))) {
 296         if (!free_entry && !entry->lock) {
 297             free_entry = entry;
 298             free_pentry = pentry;
 299         }
 300         pentry = entry;
 301         entry = entry->next;
 302     }
 303     if (!entry && free_entry) {
 304         entry = free_entry;
 305         pentry = free_pentry;
 306     }
 307     if (!entry) {
 308         entry = g_malloc0(sizeof (MapCacheEntry));
 309         pentry->next = entry;
 310         xen_remap_bucket(entry, NULL, cache_size, address_index, dummy);
 311     } else if (!entry->lock) {
 312         if (!entry->vaddr_base || entry->paddr_index != address_index ||
 313                 entry->size != cache_size ||
 314                 !test_bits(address_offset >> XC_PAGE_SHIFT,
 315                     test_bit_size >> XC_PAGE_SHIFT,
 316                     entry->valid_mapping)) {
 317             xen_remap_bucket(entry, NULL, cache_size, address_index, dummy);
 318         }
 319     }
 320
 321     if(!test_bits(address_offset >> XC_PAGE_SHIFT,
 322                 test_bit_size >> XC_PAGE_SHIFT,
 323                 entry->valid_mapping)) {
 324         mapcache->last_entry = NULL;
 325 #ifdef XEN_COMPAT_PHYSMAP
 326         if (!translated && mapcache->phys_offset_to_gaddr) {
 327             phys_addr = mapcache->phys_offset_to_gaddr(phys_addr, size);
 328             translated = true;
 329             goto tryagain;
 330         }
 331 #endif
 332         if (!dummy && runstate_check(RUN_STATE_INMIGRATE)) {
 333             dummy = true;
 334             goto tryagain;
 335         }
 336         trace_xen_map_cache_return(NULL);
 337         return NULL;
 338     }
 339
 340     mapcache->last_entry = entry;
 341     if (lock) {
 342         MapCacheRev *reventry = g_malloc0(sizeof(MapCacheRev));
 343         entry->lock++;
 344         reventry->dma = dma;
 345         reventry->vaddr_req = mapcache->last_entry->vaddr_base + address_offset;
 346         reventry->paddr_index = mapcache->last_entry->paddr_index;
 347         reventry->size = entry->size;
 348         QTAILQ_INSERT_HEAD(&mapcache->locked_entries, reventry, next);
 349     }
 350
 351     trace_xen_map_cache_return(mapcache->last_entry->vaddr_base + address_offset);
 352     return mapcache->last_entry->vaddr_base + address_offset;
 353 }
 354
 355 uint8_t *xen_map_cache(hwaddr phys_addr, hwaddr size,
 356                        uint8_t lock, bool dma)
 357 {
 358     uint8_t *p;
 359
 360     mapcache_lock();
 361     p = xen_map_cache_unlocked(phys_addr, size, lock, dma);
 362     mapcache_unlock();
 363     return p;
 364 }
 365
 366 ram_addr_t xen_ram_addr_from_mapcache(void *ptr)
 367 {
 368     MapCacheEntry *entry = NULL;
 369     MapCacheRev *reventry;
 370     hwaddr paddr_index;
 371     hwaddr size;
 372     ram_addr_t raddr;
 373     int found = 0;
 374
 375     mapcache_lock();
 376     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 377         if (reventry->vaddr_req == ptr) {
 378             paddr_index = reventry->paddr_index;
 379             size = reventry->size;
 380             found = 1;
 381             break;
 382         }
 383     }
 384     if (!found) {
 385         fprintf(stderr, "%s, could not find %p\n", __func__, ptr);
 386         QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 387             DPRINTF("   "TARGET_FMT_plx" -> %p is present\n", reventry->paddr_index,
 388                     reventry->vaddr_req);
 389         }
 390         abort();
 391         return 0;
 392     }
 393
 394     entry = &mapcache->entry[paddr_index % mapcache->nr_buckets];
 395     while (entry && (entry->paddr_index != paddr_index || entry->size != size)) {
 396         entry = entry->next;
 397     }
 398     if (!entry) {
 399         DPRINTF("Trying to find address %p that is not in the mapcache!\n", ptr);
 400         raddr = 0;
 401     } else {
 402         raddr = (reventry->paddr_index << MCACHE_BUCKET_SHIFT) +
 403              ((unsigned long) ptr - (unsigned long) entry->vaddr_base);
 404     }
 405     mapcache_unlock();
 406     return raddr;
 407 }
 408
 409 static void xen_invalidate_map_cache_entry_unlocked(uint8_t *buffer)
 410 {
 411     MapCacheEntry *entry = NULL, *pentry = NULL;
 412     MapCacheRev *reventry;
 413     hwaddr paddr_index;
 414     hwaddr size;
 415     int found = 0;
 416
 417     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 418         if (reventry->vaddr_req == buffer) {
 419             paddr_index = reventry->paddr_index;
 420             size = reventry->size;
 421             found = 1;
 422             break;
 423         }
 424     }
 425     if (!found) {
 426         DPRINTF("%s, could not find %p\n", __func__, buffer);
 427         QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 428             DPRINTF("   "TARGET_FMT_plx" -> %p is present\n", reventry->paddr_index, reventry->vaddr_req);
 429         }
 430         return;
 431     }
 432     QTAILQ_REMOVE(&mapcache->locked_entries, reventry, next);
 433     g_free(reventry);
 434
 435     if (mapcache->last_entry != NULL &&
 436         mapcache->last_entry->paddr_index == paddr_index) {
 437         mapcache->last_entry = NULL;
 438     }
 439
 440     entry = &mapcache->entry[paddr_index % mapcache->nr_buckets];
 441     while (entry && (entry->paddr_index != paddr_index || entry->size != size)) {
 442         pentry = entry;
 443         entry = entry->next;
 444     }
 445     if (!entry) {
 446         DPRINTF("Trying to unmap address %p that is not in the mapcache!\n", buffer);
 447         return;
 448     }
 449     entry->lock--;
 450     if (entry->lock > 0 || pentry == NULL) {
 451         return;
 452     }
 453
 454     pentry->next = entry->next;
 455     ram_block_notify_remove(entry->vaddr_base, entry->size);
 456     if (munmap(entry->vaddr_base, entry->size) != 0) {
 457         perror("unmap fails");
 458         exit(-1);
 459     }
 460     g_free(entry->valid_mapping);
 461     g_free(entry);
 462 }
 463
 464 void xen_invalidate_map_cache_entry(uint8_t *buffer)
 465 {
 466     mapcache_lock();
 467     xen_invalidate_map_cache_entry_unlocked(buffer);
 468     mapcache_unlock();
 469 }
 470
 471 void xen_invalidate_map_cache(void)
 472 {
 473     unsigned long i;
 474     MapCacheRev *reventry;
 475
 476     /* Flush pending AIO before destroying the mapcache */
 477     bdrv_drain_all();
 478
 479     mapcache_lock();
 480
 481     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 482         if (!reventry->dma) {
 483             continue;
 484         }
 485         fprintf(stderr, "Locked DMA mapping while invalidating mapcache!"
 486                 " "TARGET_FMT_plx" -> %p is present\n",
 487                 reventry->paddr_index, reventry->vaddr_req);
 488     }
 489
 490     for (i = 0; i < mapcache->nr_buckets; i++) {
 491         MapCacheEntry *entry = &mapcache->entry[i];
 492
 493         if (entry->vaddr_base == NULL) {
 494             continue;
 495         }
 496         if (entry->lock > 0) {
 497             continue;
 498         }
 499
 500         if (munmap(entry->vaddr_base, entry->size) != 0) {
 501             perror("unmap fails");
 502             exit(-1);
 503         }
 504
 505         entry->paddr_index = 0;
 506         entry->vaddr_base = NULL;
 507         entry->size = 0;
 508         g_free(entry->valid_mapping);
 509         entry->valid_mapping = NULL;
 510     }
 511
 512     mapcache->last_entry = NULL;
 513
 514     mapcache_unlock();
 515 }
 516
 517 static uint8_t *xen_replace_cache_entry_unlocked(hwaddr old_phys_addr,
 518                                                  hwaddr new_phys_addr,
 519                                                  hwaddr size)
 520 {
 521     MapCacheEntry *entry;
 522     hwaddr address_index, address_offset;
 523     hwaddr test_bit_size, cache_size = size;
 524
 525     address_index  = old_phys_addr >> MCACHE_BUCKET_SHIFT;
 526     address_offset = old_phys_addr & (MCACHE_BUCKET_SIZE - 1);
 527
 528     assert(size);
 529     /* test_bit_size is always a multiple of XC_PAGE_SIZE */
 530     test_bit_size = size + (old_phys_addr & (XC_PAGE_SIZE - 1));
 531     if (test_bit_size % XC_PAGE_SIZE) {
 532         test_bit_size += XC_PAGE_SIZE - (test_bit_size % XC_PAGE_SIZE);
 533     }
 534     cache_size = size + address_offset;
 535     if (cache_size % MCACHE_BUCKET_SIZE) {
 536         cache_size += MCACHE_BUCKET_SIZE - (cache_size % MCACHE_BUCKET_SIZE);
 537     }
 538
 539     entry = &mapcache->entry[address_index % mapcache->nr_buckets];
 540     while (entry && !(entry->paddr_index == address_index &&
 541                       entry->size == cache_size)) {
 542         entry = entry->next;
 543     }
 544     if (!entry) {
 545         DPRINTF("Trying to update an entry for "TARGET_FMT_plx \
 546                 "that is not in the mapcache!\n", old_phys_addr);
 547         return NULL;
 548     }
 549
 550     address_index  = new_phys_addr >> MCACHE_BUCKET_SHIFT;
 551     address_offset = new_phys_addr & (MCACHE_BUCKET_SIZE - 1);
 552
 553     fprintf(stderr, "Replacing a dummy mapcache entry for "TARGET_FMT_plx \
 554             " with "TARGET_FMT_plx"\n", old_phys_addr, new_phys_addr);
 555
 556     xen_remap_bucket(entry, entry->vaddr_base,
 557                      cache_size, address_index, false);
 558     if (!test_bits(address_offset >> XC_PAGE_SHIFT,
 559                 test_bit_size >> XC_PAGE_SHIFT,
 560                 entry->valid_mapping)) {
 561         DPRINTF("Unable to update a mapcache entry for "TARGET_FMT_plx"!\n",
 562                 old_phys_addr);
 563         return NULL;
 564     }
 565
 566     return entry->vaddr_base + address_offset;
 567 }
 568
 569 uint8_t *xen_replace_cache_entry(hwaddr old_phys_addr,
 570                                  hwaddr new_phys_addr,
 571                                  hwaddr size)
 572 {
 573     uint8_t *p;
 574
 575     mapcache_lock();
 576     p = xen_replace_cache_entry_unlocked(old_phys_addr, new_phys_addr, size);
 577     mapcache_unlock();
 578     return p;
 579 }