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