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/error-report.h"
  13
  14 #include <sys/resource.h>
  15
  16 #include "hw/xen/xen_backend.h"
  17 #include "qemu/bitmap.h"
  18
  19 #include <xen/hvm/params.h>
  20
  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 * 1024 * 1024)
  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(map_cache_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     if (!dummy) {
 187         vaddr_base = xenforeignmemory_map2(xen_fmem, xen_domid, vaddr,
 188                                            PROT_READ | PROT_WRITE, 0,
 189                                            nb_pfn, pfns, err);
 190         if (vaddr_base == NULL) {
 191             perror("xenforeignmemory_map2");
 192             exit(-1);
 193         }
 194     } else {
 195         /*
 196          * We create dummy mappings where we are unable to create a foreign
 197          * mapping immediately due to certain circumstances (i.e. on resume now)
 198          */
 199         vaddr_base = mmap(vaddr, size, PROT_READ | PROT_WRITE,
 200                           MAP_ANON | MAP_SHARED, -1, 0);
 201         if (vaddr_base == MAP_FAILED) {
 202             perror("mmap");
 203             exit(-1);
 204         }
 205     }
 206
 207     if (!(entry->flags & XEN_MAPCACHE_ENTRY_DUMMY)) {
 208         ram_block_notify_add(vaddr_base, size);
 209     }
 210
 211     entry->vaddr_base = vaddr_base;
 212     entry->paddr_index = address_index;
 213     entry->size = size;
 214     entry->valid_mapping = (unsigned long *) g_malloc0(sizeof(unsigned long) *
 215             BITS_TO_LONGS(size >> XC_PAGE_SHIFT));
 216
 217     if (dummy) {
 218         entry->flags |= XEN_MAPCACHE_ENTRY_DUMMY;
 219     } else {
 220         entry->flags &= ~(XEN_MAPCACHE_ENTRY_DUMMY);
 221     }
 222
 223     bitmap_zero(entry->valid_mapping, nb_pfn);
 224     for (i = 0; i < nb_pfn; i++) {
 225         if (!err[i]) {
 226             bitmap_set(entry->valid_mapping, i, 1);
 227         }
 228     }
 229
 230     g_free(pfns);
 231     g_free(err);
 232 }
 233
 234 static uint8_t *xen_map_cache_unlocked(hwaddr phys_addr, hwaddr size,
 235                                        uint8_t lock, bool dma)
 236 {
 237     MapCacheEntry *entry, *pentry = NULL,
 238                   *free_entry = NULL, *free_pentry = NULL;
 239     hwaddr address_index;
 240     hwaddr address_offset;
 241     hwaddr cache_size = size;
 242     hwaddr test_bit_size;
 243     bool translated G_GNUC_UNUSED = false;
 244     bool dummy = false;
 245
 246 tryagain:
 247     address_index  = phys_addr >> MCACHE_BUCKET_SHIFT;
 248     address_offset = phys_addr & (MCACHE_BUCKET_SIZE - 1);
 249
 250     trace_xen_map_cache(phys_addr);
 251
 252     /* test_bit_size is always a multiple of XC_PAGE_SIZE */
 253     if (size) {
 254         test_bit_size = size + (phys_addr & (XC_PAGE_SIZE - 1));
 255
 256         if (test_bit_size % XC_PAGE_SIZE) {
 257             test_bit_size += XC_PAGE_SIZE - (test_bit_size % XC_PAGE_SIZE);
 258         }
 259     } else {
 260         test_bit_size = XC_PAGE_SIZE;
 261     }
 262
 263     if (mapcache->last_entry != NULL &&
 264         mapcache->last_entry->paddr_index == address_index &&
 265         !lock && !size &&
 266         test_bits(address_offset >> XC_PAGE_SHIFT,
 267                   test_bit_size >> XC_PAGE_SHIFT,
 268                   mapcache->last_entry->valid_mapping)) {
 269         trace_xen_map_cache_return(mapcache->last_entry->vaddr_base + address_offset);
 270         return mapcache->last_entry->vaddr_base + address_offset;
 271     }
 272
 273     /* size is always a multiple of MCACHE_BUCKET_SIZE */
 274     if (size) {
 275         cache_size = size + address_offset;
 276         if (cache_size % MCACHE_BUCKET_SIZE) {
 277             cache_size += MCACHE_BUCKET_SIZE - (cache_size % MCACHE_BUCKET_SIZE);
 278         }
 279     } else {
 280         cache_size = MCACHE_BUCKET_SIZE;
 281     }
 282
 283     entry = &mapcache->entry[address_index % mapcache->nr_buckets];
 284
 285     while (entry && (lock || entry->lock) && entry->vaddr_base &&
 286             (entry->paddr_index != address_index || entry->size != cache_size ||
 287              !test_bits(address_offset >> XC_PAGE_SHIFT,
 288                  test_bit_size >> XC_PAGE_SHIFT,
 289                  entry->valid_mapping))) {
 290         if (!free_entry && !entry->lock) {
 291             free_entry = entry;
 292             free_pentry = pentry;
 293         }
 294         pentry = entry;
 295         entry = entry->next;
 296     }
 297     if (!entry && free_entry) {
 298         entry = free_entry;
 299         pentry = free_pentry;
 300     }
 301     if (!entry) {
 302         entry = g_malloc0(sizeof (MapCacheEntry));
 303         pentry->next = entry;
 304         xen_remap_bucket(entry, NULL, cache_size, address_index, dummy);
 305     } else if (!entry->lock) {
 306         if (!entry->vaddr_base || entry->paddr_index != address_index ||
 307                 entry->size != cache_size ||
 308                 !test_bits(address_offset >> XC_PAGE_SHIFT,
 309                     test_bit_size >> XC_PAGE_SHIFT,
 310                     entry->valid_mapping)) {
 311             xen_remap_bucket(entry, NULL, cache_size, address_index, dummy);
 312         }
 313     }
 314
 315     if(!test_bits(address_offset >> XC_PAGE_SHIFT,
 316                 test_bit_size >> XC_PAGE_SHIFT,
 317                 entry->valid_mapping)) {
 318         mapcache->last_entry = NULL;
 319 #ifdef XEN_COMPAT_PHYSMAP
 320         if (!translated && mapcache->phys_offset_to_gaddr) {
 321             phys_addr = mapcache->phys_offset_to_gaddr(phys_addr, size);
 322             translated = true;
 323             goto tryagain;
 324         }
 325 #endif
 326         if (!dummy && runstate_check(RUN_STATE_INMIGRATE)) {
 327             dummy = true;
 328             goto tryagain;
 329         }
 330         trace_xen_map_cache_return(NULL);
 331         return NULL;
 332     }
 333
 334     mapcache->last_entry = entry;
 335     if (lock) {
 336         MapCacheRev *reventry = g_malloc0(sizeof(MapCacheRev));
 337         entry->lock++;
 338         reventry->dma = dma;
 339         reventry->vaddr_req = mapcache->last_entry->vaddr_base + address_offset;
 340         reventry->paddr_index = mapcache->last_entry->paddr_index;
 341         reventry->size = entry->size;
 342         QTAILQ_INSERT_HEAD(&mapcache->locked_entries, reventry, next);
 343     }
 344
 345     trace_xen_map_cache_return(mapcache->last_entry->vaddr_base + address_offset);
 346     return mapcache->last_entry->vaddr_base + address_offset;
 347 }
 348
 349 uint8_t *xen_map_cache(hwaddr phys_addr, hwaddr size,
 350                        uint8_t lock, bool dma)
 351 {
 352     uint8_t *p;
 353
 354     mapcache_lock();
 355     p = xen_map_cache_unlocked(phys_addr, size, lock, dma);
 356     mapcache_unlock();
 357     return p;
 358 }
 359
 360 ram_addr_t xen_ram_addr_from_mapcache(void *ptr)
 361 {
 362     MapCacheEntry *entry = NULL;
 363     MapCacheRev *reventry;
 364     hwaddr paddr_index;
 365     hwaddr size;
 366     ram_addr_t raddr;
 367     int found = 0;
 368
 369     mapcache_lock();
 370     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 371         if (reventry->vaddr_req == ptr) {
 372             paddr_index = reventry->paddr_index;
 373             size = reventry->size;
 374             found = 1;
 375             break;
 376         }
 377     }
 378     if (!found) {
 379         fprintf(stderr, "%s, could not find %p\n", __func__, ptr);
 380         QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 381             DPRINTF("   "TARGET_FMT_plx" -> %p is present\n", reventry->paddr_index,
 382                     reventry->vaddr_req);
 383         }
 384         abort();
 385         return 0;
 386     }
 387
 388     entry = &mapcache->entry[paddr_index % mapcache->nr_buckets];
 389     while (entry && (entry->paddr_index != paddr_index || entry->size != size)) {
 390         entry = entry->next;
 391     }
 392     if (!entry) {
 393         DPRINTF("Trying to find address %p that is not in the mapcache!\n", ptr);
 394         raddr = 0;
 395     } else {
 396         raddr = (reventry->paddr_index << MCACHE_BUCKET_SHIFT) +
 397              ((unsigned long) ptr - (unsigned long) entry->vaddr_base);
 398     }
 399     mapcache_unlock();
 400     return raddr;
 401 }
 402
 403 static void xen_invalidate_map_cache_entry_unlocked(uint8_t *buffer)
 404 {
 405     MapCacheEntry *entry = NULL, *pentry = NULL;
 406     MapCacheRev *reventry;
 407     hwaddr paddr_index;
 408     hwaddr size;
 409     int found = 0;
 410
 411     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 412         if (reventry->vaddr_req == buffer) {
 413             paddr_index = reventry->paddr_index;
 414             size = reventry->size;
 415             found = 1;
 416             break;
 417         }
 418     }
 419     if (!found) {
 420         DPRINTF("%s, could not find %p\n", __func__, buffer);
 421         QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 422             DPRINTF("   "TARGET_FMT_plx" -> %p is present\n", reventry->paddr_index, reventry->vaddr_req);
 423         }
 424         return;
 425     }
 426     QTAILQ_REMOVE(&mapcache->locked_entries, reventry, next);
 427     g_free(reventry);
 428
 429     if (mapcache->last_entry != NULL &&
 430         mapcache->last_entry->paddr_index == paddr_index) {
 431         mapcache->last_entry = NULL;
 432     }
 433
 434     entry = &mapcache->entry[paddr_index % mapcache->nr_buckets];
 435     while (entry && (entry->paddr_index != paddr_index || entry->size != size)) {
 436         pentry = entry;
 437         entry = entry->next;
 438     }
 439     if (!entry) {
 440         DPRINTF("Trying to unmap address %p that is not in the mapcache!\n", buffer);
 441         return;
 442     }
 443     entry->lock--;
 444     if (entry->lock > 0 || pentry == NULL) {
 445         return;
 446     }
 447
 448     pentry->next = entry->next;
 449     ram_block_notify_remove(entry->vaddr_base, entry->size);
 450     if (munmap(entry->vaddr_base, entry->size) != 0) {
 451         perror("unmap fails");
 452         exit(-1);
 453     }
 454     g_free(entry->valid_mapping);
 455     g_free(entry);
 456 }
 457
 458 void xen_invalidate_map_cache_entry(uint8_t *buffer)
 459 {
 460     mapcache_lock();
 461     xen_invalidate_map_cache_entry_unlocked(buffer);
 462     mapcache_unlock();
 463 }
 464
 465 void xen_invalidate_map_cache(void)
 466 {
 467     unsigned long i;
 468     MapCacheRev *reventry;
 469
 470     /* Flush pending AIO before destroying the mapcache */
 471     bdrv_drain_all();
 472
 473     mapcache_lock();
 474
 475     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 476         if (!reventry->dma) {
 477             continue;
 478         }
 479         fprintf(stderr, "Locked DMA mapping while invalidating mapcache!"
 480                 " "TARGET_FMT_plx" -> %p is present\n",
 481                 reventry->paddr_index, reventry->vaddr_req);
 482     }
 483
 484     for (i = 0; i < mapcache->nr_buckets; i++) {
 485         MapCacheEntry *entry = &mapcache->entry[i];
 486
 487         if (entry->vaddr_base == NULL) {
 488             continue;
 489         }
 490         if (entry->lock > 0) {
 491             continue;
 492         }
 493
 494         if (munmap(entry->vaddr_base, entry->size) != 0) {
 495             perror("unmap fails");
 496             exit(-1);
 497         }
 498
 499         entry->paddr_index = 0;
 500         entry->vaddr_base = NULL;
 501         entry->size = 0;
 502         g_free(entry->valid_mapping);
 503         entry->valid_mapping = NULL;
 504     }
 505
 506     mapcache->last_entry = NULL;
 507
 508     mapcache_unlock();
 509 }
 510
 511 static uint8_t *xen_replace_cache_entry_unlocked(hwaddr old_phys_addr,
 512                                                  hwaddr new_phys_addr,
 513                                                  hwaddr size)
 514 {
 515     MapCacheEntry *entry;
 516     hwaddr address_index, address_offset;
 517     hwaddr test_bit_size, cache_size = size;
 518
 519     address_index  = old_phys_addr >> MCACHE_BUCKET_SHIFT;
 520     address_offset = old_phys_addr & (MCACHE_BUCKET_SIZE - 1);
 521
 522     assert(size);
 523     /* test_bit_size is always a multiple of XC_PAGE_SIZE */
 524     test_bit_size = size + (old_phys_addr & (XC_PAGE_SIZE - 1));
 525     if (test_bit_size % XC_PAGE_SIZE) {
 526         test_bit_size += XC_PAGE_SIZE - (test_bit_size % XC_PAGE_SIZE);
 527     }
 528     cache_size = size + address_offset;
 529     if (cache_size % MCACHE_BUCKET_SIZE) {
 530         cache_size += MCACHE_BUCKET_SIZE - (cache_size % MCACHE_BUCKET_SIZE);
 531     }
 532
 533     entry = &mapcache->entry[address_index % mapcache->nr_buckets];
 534     while (entry && !(entry->paddr_index == address_index &&
 535                       entry->size == cache_size)) {
 536         entry = entry->next;
 537     }
 538     if (!entry) {
 539         DPRINTF("Trying to update an entry for "TARGET_FMT_plx \
 540                 "that is not in the mapcache!\n", old_phys_addr);
 541         return NULL;
 542     }
 543
 544     address_index  = new_phys_addr >> MCACHE_BUCKET_SHIFT;
 545     address_offset = new_phys_addr & (MCACHE_BUCKET_SIZE - 1);
 546
 547     fprintf(stderr, "Replacing a dummy mapcache entry for "TARGET_FMT_plx \
 548             " with "TARGET_FMT_plx"\n", old_phys_addr, new_phys_addr);
 549
 550     xen_remap_bucket(entry, entry->vaddr_base,
 551                      cache_size, address_index, false);
 552     if (!test_bits(address_offset >> XC_PAGE_SHIFT,
 553                 test_bit_size >> XC_PAGE_SHIFT,
 554                 entry->valid_mapping)) {
 555         DPRINTF("Unable to update a mapcache entry for "TARGET_FMT_plx"!\n",
 556                 old_phys_addr);
 557         return NULL;
 558     }
 559
 560     return entry->vaddr_base + address_offset;
 561 }
 562
 563 uint8_t *xen_replace_cache_entry(hwaddr old_phys_addr,
 564                                  hwaddr new_phys_addr,
 565                                  hwaddr size)
 566 {
 567     uint8_t *p;
 568
 569     mapcache_lock();
 570     p = xen_replace_cache_entry_unlocked(old_phys_addr, new_phys_addr, size);
 571     mapcache_unlock();
 572     return p;
 573 }