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-root.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     hwaddr size;
  57     struct MapCacheEntry *next;
  58 } MapCacheEntry;
  59
  60 typedef struct MapCacheRev {
  61     uint8_t *vaddr_req;
  62     hwaddr paddr_index;
  63     hwaddr size;
  64     QTAILQ_ENTRY(MapCacheRev) next;
  65     bool dma;
  66 } MapCacheRev;
  67
  68 typedef struct MapCache {
  69     MapCacheEntry *entry;
  70     unsigned long nr_buckets;
  71     QTAILQ_HEAD(map_cache_head, MapCacheRev) locked_entries;
  72
  73     /* For most cases (>99.9%), the page address is the same. */
  74     MapCacheEntry *last_entry;
  75     unsigned long max_mcache_size;
  76     unsigned int mcache_bucket_shift;
  77
  78     phys_offset_to_gaddr_t phys_offset_to_gaddr;
  79     QemuMutex lock;
  80     void *opaque;
  81 } MapCache;
  82
  83 static MapCache *mapcache;
  84
  85 static inline void mapcache_lock(void)
  86 {
  87     qemu_mutex_lock(&mapcache->lock);
  88 }
  89
  90 static inline void mapcache_unlock(void)
  91 {
  92     qemu_mutex_unlock(&mapcache->lock);
  93 }
  94
  95 static inline int test_bits(int nr, int size, const unsigned long *addr)
  96 {
  97     unsigned long res = find_next_zero_bit(addr, size + nr, nr);
  98     if (res >= nr + size)
  99         return 1;
 100     else
 101         return 0;
 102 }
 103
 104 void xen_map_cache_init(phys_offset_to_gaddr_t f, void *opaque)
 105 {
 106     unsigned long size;
 107     struct rlimit rlimit_as;
 108
 109     mapcache = g_malloc0(sizeof (MapCache));
 110
 111     mapcache->phys_offset_to_gaddr = f;
 112     mapcache->opaque = opaque;
 113     qemu_mutex_init(&mapcache->lock);
 114
 115     QTAILQ_INIT(&mapcache->locked_entries);
 116
 117     if (geteuid() == 0) {
 118         rlimit_as.rlim_cur = RLIM_INFINITY;
 119         rlimit_as.rlim_max = RLIM_INFINITY;
 120         mapcache->max_mcache_size = MCACHE_MAX_SIZE;
 121     } else {
 122         getrlimit(RLIMIT_AS, &rlimit_as);
 123         rlimit_as.rlim_cur = rlimit_as.rlim_max;
 124
 125         if (rlimit_as.rlim_max != RLIM_INFINITY) {
 126             fprintf(stderr, "Warning: QEMU's maximum size of virtual"
 127                     " memory is not infinity.\n");
 128         }
 129         if (rlimit_as.rlim_max < MCACHE_MAX_SIZE + NON_MCACHE_MEMORY_SIZE) {
 130             mapcache->max_mcache_size = rlimit_as.rlim_max -
 131                 NON_MCACHE_MEMORY_SIZE;
 132         } else {
 133             mapcache->max_mcache_size = MCACHE_MAX_SIZE;
 134         }
 135     }
 136
 137     setrlimit(RLIMIT_AS, &rlimit_as);
 138
 139     mapcache->nr_buckets =
 140         (((mapcache->max_mcache_size >> XC_PAGE_SHIFT) +
 141           (1UL << (MCACHE_BUCKET_SHIFT - XC_PAGE_SHIFT)) - 1) >>
 142          (MCACHE_BUCKET_SHIFT - XC_PAGE_SHIFT));
 143
 144     size = mapcache->nr_buckets * sizeof (MapCacheEntry);
 145     size = (size + XC_PAGE_SIZE - 1) & ~(XC_PAGE_SIZE - 1);
 146     DPRINTF("%s, nr_buckets = %lx size %lu\n", __func__,
 147             mapcache->nr_buckets, size);
 148     mapcache->entry = g_malloc0(size);
 149 }
 150
 151 static void xen_remap_bucket(MapCacheEntry *entry,
 152                              hwaddr size,
 153                              hwaddr address_index)
 154 {
 155     uint8_t *vaddr_base;
 156     xen_pfn_t *pfns;
 157     int *err;
 158     unsigned int i;
 159     hwaddr nb_pfn = size >> XC_PAGE_SHIFT;
 160
 161     trace_xen_remap_bucket(address_index);
 162
 163     pfns = g_malloc0(nb_pfn * sizeof (xen_pfn_t));
 164     err = g_malloc0(nb_pfn * sizeof (int));
 165
 166     if (entry->vaddr_base != NULL) {
 167         ram_block_notify_remove(entry->vaddr_base, entry->size);
 168         if (munmap(entry->vaddr_base, entry->size) != 0) {
 169             perror("unmap fails");
 170             exit(-1);
 171         }
 172     }
 173     g_free(entry->valid_mapping);
 174     entry->valid_mapping = NULL;
 175
 176     for (i = 0; i < nb_pfn; i++) {
 177         pfns[i] = (address_index << (MCACHE_BUCKET_SHIFT-XC_PAGE_SHIFT)) + i;
 178     }
 179
 180     vaddr_base = xenforeignmemory_map(xen_fmem, xen_domid, PROT_READ|PROT_WRITE,
 181                                       nb_pfn, pfns, err);
 182     if (vaddr_base == NULL) {
 183         perror("xenforeignmemory_map");
 184         exit(-1);
 185     }
 186
 187     entry->vaddr_base = vaddr_base;
 188     entry->paddr_index = address_index;
 189     entry->size = size;
 190     entry->valid_mapping = (unsigned long *) g_malloc0(sizeof(unsigned long) *
 191             BITS_TO_LONGS(size >> XC_PAGE_SHIFT));
 192
 193     ram_block_notify_add(entry->vaddr_base, entry->size);
 194     bitmap_zero(entry->valid_mapping, nb_pfn);
 195     for (i = 0; i < nb_pfn; i++) {
 196         if (!err[i]) {
 197             bitmap_set(entry->valid_mapping, i, 1);
 198         }
 199     }
 200
 201     g_free(pfns);
 202     g_free(err);
 203 }
 204
 205 static uint8_t *xen_map_cache_unlocked(hwaddr phys_addr, hwaddr size,
 206                                        uint8_t lock, bool dma)
 207 {
 208     MapCacheEntry *entry, *pentry = NULL;
 209     hwaddr address_index;
 210     hwaddr address_offset;
 211     hwaddr cache_size = size;
 212     hwaddr test_bit_size;
 213     bool translated = false;
 214
 215 tryagain:
 216     address_index  = phys_addr >> MCACHE_BUCKET_SHIFT;
 217     address_offset = phys_addr & (MCACHE_BUCKET_SIZE - 1);
 218
 219     trace_xen_map_cache(phys_addr);
 220
 221     /* test_bit_size is always a multiple of XC_PAGE_SIZE */
 222     if (size) {
 223         test_bit_size = size + (phys_addr & (XC_PAGE_SIZE - 1));
 224
 225         if (test_bit_size % XC_PAGE_SIZE) {
 226             test_bit_size += XC_PAGE_SIZE - (test_bit_size % XC_PAGE_SIZE);
 227         }
 228     } else {
 229         test_bit_size = XC_PAGE_SIZE;
 230     }
 231
 232     if (mapcache->last_entry != NULL &&
 233         mapcache->last_entry->paddr_index == address_index &&
 234         !lock && !size &&
 235         test_bits(address_offset >> XC_PAGE_SHIFT,
 236                   test_bit_size >> XC_PAGE_SHIFT,
 237                   mapcache->last_entry->valid_mapping)) {
 238         trace_xen_map_cache_return(mapcache->last_entry->vaddr_base + address_offset);
 239         return mapcache->last_entry->vaddr_base + address_offset;
 240     }
 241
 242     /* size is always a multiple of MCACHE_BUCKET_SIZE */
 243     if (size) {
 244         cache_size = size + address_offset;
 245         if (cache_size % MCACHE_BUCKET_SIZE) {
 246             cache_size += MCACHE_BUCKET_SIZE - (cache_size % MCACHE_BUCKET_SIZE);
 247         }
 248     } else {
 249         cache_size = MCACHE_BUCKET_SIZE;
 250     }
 251
 252     entry = &mapcache->entry[address_index % mapcache->nr_buckets];
 253
 254     while (entry && entry->lock && entry->vaddr_base &&
 255             (entry->paddr_index != address_index || entry->size != cache_size ||
 256              !test_bits(address_offset >> XC_PAGE_SHIFT,
 257                  test_bit_size >> XC_PAGE_SHIFT,
 258                  entry->valid_mapping))) {
 259         pentry = entry;
 260         entry = entry->next;
 261     }
 262     if (!entry) {
 263         entry = g_malloc0(sizeof (MapCacheEntry));
 264         pentry->next = entry;
 265         xen_remap_bucket(entry, cache_size, address_index);
 266     } else if (!entry->lock) {
 267         if (!entry->vaddr_base || entry->paddr_index != address_index ||
 268                 entry->size != cache_size ||
 269                 !test_bits(address_offset >> XC_PAGE_SHIFT,
 270                     test_bit_size >> XC_PAGE_SHIFT,
 271                     entry->valid_mapping)) {
 272             xen_remap_bucket(entry, cache_size, address_index);
 273         }
 274     }
 275
 276     if(!test_bits(address_offset >> XC_PAGE_SHIFT,
 277                 test_bit_size >> XC_PAGE_SHIFT,
 278                 entry->valid_mapping)) {
 279         mapcache->last_entry = NULL;
 280         if (!translated && mapcache->phys_offset_to_gaddr) {
 281             phys_addr = mapcache->phys_offset_to_gaddr(phys_addr, size, mapcache->opaque);
 282             translated = true;
 283             goto tryagain;
 284         }
 285         trace_xen_map_cache_return(NULL);
 286         return NULL;
 287     }
 288
 289     mapcache->last_entry = entry;
 290     if (lock) {
 291         MapCacheRev *reventry = g_malloc0(sizeof(MapCacheRev));
 292         entry->lock++;
 293         reventry->dma = dma;
 294         reventry->vaddr_req = mapcache->last_entry->vaddr_base + address_offset;
 295         reventry->paddr_index = mapcache->last_entry->paddr_index;
 296         reventry->size = entry->size;
 297         QTAILQ_INSERT_HEAD(&mapcache->locked_entries, reventry, next);
 298     }
 299
 300     trace_xen_map_cache_return(mapcache->last_entry->vaddr_base + address_offset);
 301     return mapcache->last_entry->vaddr_base + address_offset;
 302 }
 303
 304 uint8_t *xen_map_cache(hwaddr phys_addr, hwaddr size,
 305                        uint8_t lock, bool dma)
 306 {
 307     uint8_t *p;
 308
 309     mapcache_lock();
 310     p = xen_map_cache_unlocked(phys_addr, size, lock, dma);
 311     mapcache_unlock();
 312     return p;
 313 }
 314
 315 ram_addr_t xen_ram_addr_from_mapcache(void *ptr)
 316 {
 317     MapCacheEntry *entry = NULL;
 318     MapCacheRev *reventry;
 319     hwaddr paddr_index;
 320     hwaddr size;
 321     ram_addr_t raddr;
 322     int found = 0;
 323
 324     mapcache_lock();
 325     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 326         if (reventry->vaddr_req == ptr) {
 327             paddr_index = reventry->paddr_index;
 328             size = reventry->size;
 329             found = 1;
 330             break;
 331         }
 332     }
 333     if (!found) {
 334         fprintf(stderr, "%s, could not find %p\n", __func__, ptr);
 335         QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 336             DPRINTF("   "TARGET_FMT_plx" -> %p is present\n", reventry->paddr_index,
 337                     reventry->vaddr_req);
 338         }
 339         abort();
 340         return 0;
 341     }
 342
 343     entry = &mapcache->entry[paddr_index % mapcache->nr_buckets];
 344     while (entry && (entry->paddr_index != paddr_index || entry->size != size)) {
 345         entry = entry->next;
 346     }
 347     if (!entry) {
 348         DPRINTF("Trying to find address %p that is not in the mapcache!\n", ptr);
 349         raddr = 0;
 350     } else {
 351         raddr = (reventry->paddr_index << MCACHE_BUCKET_SHIFT) +
 352              ((unsigned long) ptr - (unsigned long) entry->vaddr_base);
 353     }
 354     mapcache_unlock();
 355     return raddr;
 356 }
 357
 358 static void xen_invalidate_map_cache_entry_unlocked(uint8_t *buffer)
 359 {
 360     MapCacheEntry *entry = NULL, *pentry = NULL;
 361     MapCacheRev *reventry;
 362     hwaddr paddr_index;
 363     hwaddr size;
 364     int found = 0;
 365
 366     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 367         if (reventry->vaddr_req == buffer) {
 368             paddr_index = reventry->paddr_index;
 369             size = reventry->size;
 370             found = 1;
 371             break;
 372         }
 373     }
 374     if (!found) {
 375         DPRINTF("%s, could not find %p\n", __func__, buffer);
 376         QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 377             DPRINTF("   "TARGET_FMT_plx" -> %p is present\n", reventry->paddr_index, reventry->vaddr_req);
 378         }
 379         return;
 380     }
 381     QTAILQ_REMOVE(&mapcache->locked_entries, reventry, next);
 382     g_free(reventry);
 383
 384     if (mapcache->last_entry != NULL &&
 385         mapcache->last_entry->paddr_index == paddr_index) {
 386         mapcache->last_entry = NULL;
 387     }
 388
 389     entry = &mapcache->entry[paddr_index % mapcache->nr_buckets];
 390     while (entry && (entry->paddr_index != paddr_index || entry->size != size)) {
 391         pentry = entry;
 392         entry = entry->next;
 393     }
 394     if (!entry) {
 395         DPRINTF("Trying to unmap address %p that is not in the mapcache!\n", buffer);
 396         return;
 397     }
 398     entry->lock--;
 399     if (entry->lock > 0 || pentry == NULL) {
 400         return;
 401     }
 402
 403     pentry->next = entry->next;
 404     ram_block_notify_remove(entry->vaddr_base, entry->size);
 405     if (munmap(entry->vaddr_base, entry->size) != 0) {
 406         perror("unmap fails");
 407         exit(-1);
 408     }
 409     g_free(entry->valid_mapping);
 410     g_free(entry);
 411 }
 412
 413 void xen_invalidate_map_cache_entry(uint8_t *buffer)
 414 {
 415     mapcache_lock();
 416     xen_invalidate_map_cache_entry_unlocked(buffer);
 417     mapcache_unlock();
 418 }
 419
 420 void xen_invalidate_map_cache(void)
 421 {
 422     unsigned long i;
 423     MapCacheRev *reventry;
 424
 425     /* Flush pending AIO before destroying the mapcache */
 426     bdrv_drain_all();
 427
 428     mapcache_lock();
 429
 430     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 431         if (!reventry->dma) {
 432             continue;
 433         }
 434         fprintf(stderr, "Locked DMA mapping while invalidating mapcache!"
 435                 " "TARGET_FMT_plx" -> %p is present\n",
 436                 reventry->paddr_index, reventry->vaddr_req);
 437     }
 438
 439     for (i = 0; i < mapcache->nr_buckets; i++) {
 440         MapCacheEntry *entry = &mapcache->entry[i];
 441
 442         if (entry->vaddr_base == NULL) {
 443             continue;
 444         }
 445         if (entry->lock > 0) {
 446             continue;
 447         }
 448
 449         if (munmap(entry->vaddr_base, entry->size) != 0) {
 450             perror("unmap fails");
 451             exit(-1);
 452         }
 453
 454         entry->paddr_index = 0;
 455         entry->vaddr_base = NULL;
 456         entry->size = 0;
 457         g_free(entry->valid_mapping);
 458         entry->valid_mapping = NULL;
 459     }
 460
 461     mapcache->last_entry = NULL;
 462
 463     mapcache_unlock();
 464 }