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 #include <sys/mman.h>
  21
  22 #include "sysemu/xen-mapcache.h"
  23 #include "trace.h"
  24
  25
  26 //#define MAPCACHE_DEBUG
  27
  28 #ifdef MAPCACHE_DEBUG
  29 #  define DPRINTF(fmt, ...) do { \
  30     fprintf(stderr, "xen_mapcache: " fmt, ## __VA_ARGS__); \
  31 } while (0)
  32 #else
  33 #  define DPRINTF(fmt, ...) do { } while (0)
  34 #endif
  35
  36 #if HOST_LONG_BITS == 32
  37 #  define MCACHE_BUCKET_SHIFT 16
  38 #  define MCACHE_MAX_SIZE     (1UL<<31) /* 2GB Cap */
  39 #else
  40 #  define MCACHE_BUCKET_SHIFT 20
  41 #  define MCACHE_MAX_SIZE     (1UL<<35) /* 32GB Cap */
  42 #endif
  43 #define MCACHE_BUCKET_SIZE (1UL << MCACHE_BUCKET_SHIFT)
  44
  45 /* This is the size of the virtual address space reserve to QEMU that will not
  46  * be use by MapCache.
  47  * From empirical tests I observed that qemu use 75MB more than the
  48  * max_mcache_size.
  49  */
  50 #define NON_MCACHE_MEMORY_SIZE (80 * 1024 * 1024)
  51
  52 typedef struct MapCacheEntry {
  53     hwaddr paddr_index;
  54     uint8_t *vaddr_base;
  55     unsigned long *valid_mapping;
  56     uint8_t lock;
  57     hwaddr size;
  58     struct MapCacheEntry *next;
  59 } MapCacheEntry;
  60
  61 typedef struct MapCacheRev {
  62     uint8_t *vaddr_req;
  63     hwaddr paddr_index;
  64     hwaddr size;
  65     QTAILQ_ENTRY(MapCacheRev) next;
  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         if (munmap(entry->vaddr_base, entry->size) != 0) {
 168             perror("unmap fails");
 169             exit(-1);
 170         }
 171     }
 172     g_free(entry->valid_mapping);
 173     entry->valid_mapping = NULL;
 174
 175     for (i = 0; i < nb_pfn; i++) {
 176         pfns[i] = (address_index << (MCACHE_BUCKET_SHIFT-XC_PAGE_SHIFT)) + i;
 177     }
 178
 179     vaddr_base = xenforeignmemory_map(xen_fmem, xen_domid, PROT_READ|PROT_WRITE,
 180                                       nb_pfn, pfns, err);
 181     if (vaddr_base == NULL) {
 182         perror("xenforeignmemory_map");
 183         exit(-1);
 184     }
 185
 186     entry->vaddr_base = vaddr_base;
 187     entry->paddr_index = address_index;
 188     entry->size = size;
 189     entry->valid_mapping = (unsigned long *) g_malloc0(sizeof(unsigned long) *
 190             BITS_TO_LONGS(size >> XC_PAGE_SHIFT));
 191
 192     bitmap_zero(entry->valid_mapping, nb_pfn);
 193     for (i = 0; i < nb_pfn; i++) {
 194         if (!err[i]) {
 195             bitmap_set(entry->valid_mapping, i, 1);
 196         }
 197     }
 198
 199     g_free(pfns);
 200     g_free(err);
 201 }
 202
 203 static uint8_t *xen_map_cache_unlocked(hwaddr phys_addr, hwaddr size,
 204                                        uint8_t lock)
 205 {
 206     MapCacheEntry *entry, *pentry = NULL;
 207     hwaddr address_index;
 208     hwaddr address_offset;
 209     hwaddr cache_size = size;
 210     hwaddr test_bit_size;
 211     bool translated = false;
 212
 213 tryagain:
 214     address_index  = phys_addr >> MCACHE_BUCKET_SHIFT;
 215     address_offset = phys_addr & (MCACHE_BUCKET_SIZE - 1);
 216
 217     trace_xen_map_cache(phys_addr);
 218
 219     /* test_bit_size is always a multiple of XC_PAGE_SIZE */
 220     if (size) {
 221         test_bit_size = size + (phys_addr & (XC_PAGE_SIZE - 1));
 222
 223         if (test_bit_size % XC_PAGE_SIZE) {
 224             test_bit_size += XC_PAGE_SIZE - (test_bit_size % XC_PAGE_SIZE);
 225         }
 226     } else {
 227         test_bit_size = XC_PAGE_SIZE;
 228     }
 229
 230     if (mapcache->last_entry != NULL &&
 231         mapcache->last_entry->paddr_index == address_index &&
 232         !lock && !size &&
 233         test_bits(address_offset >> XC_PAGE_SHIFT,
 234                   test_bit_size >> XC_PAGE_SHIFT,
 235                   mapcache->last_entry->valid_mapping)) {
 236         trace_xen_map_cache_return(mapcache->last_entry->vaddr_base + address_offset);
 237         return mapcache->last_entry->vaddr_base + address_offset;
 238     }
 239
 240     /* size is always a multiple of MCACHE_BUCKET_SIZE */
 241     if (size) {
 242         cache_size = size + address_offset;
 243         if (cache_size % MCACHE_BUCKET_SIZE) {
 244             cache_size += MCACHE_BUCKET_SIZE - (cache_size % MCACHE_BUCKET_SIZE);
 245         }
 246     } else {
 247         cache_size = MCACHE_BUCKET_SIZE;
 248     }
 249
 250     entry = &mapcache->entry[address_index % mapcache->nr_buckets];
 251
 252     while (entry && entry->lock && entry->vaddr_base &&
 253             (entry->paddr_index != address_index || entry->size != cache_size ||
 254              !test_bits(address_offset >> XC_PAGE_SHIFT,
 255                  test_bit_size >> XC_PAGE_SHIFT,
 256                  entry->valid_mapping))) {
 257         pentry = entry;
 258         entry = entry->next;
 259     }
 260     if (!entry) {
 261         entry = g_malloc0(sizeof (MapCacheEntry));
 262         pentry->next = entry;
 263         xen_remap_bucket(entry, cache_size, address_index);
 264     } else if (!entry->lock) {
 265         if (!entry->vaddr_base || entry->paddr_index != address_index ||
 266                 entry->size != cache_size ||
 267                 !test_bits(address_offset >> XC_PAGE_SHIFT,
 268                     test_bit_size >> XC_PAGE_SHIFT,
 269                     entry->valid_mapping)) {
 270             xen_remap_bucket(entry, cache_size, address_index);
 271         }
 272     }
 273
 274     if(!test_bits(address_offset >> XC_PAGE_SHIFT,
 275                 test_bit_size >> XC_PAGE_SHIFT,
 276                 entry->valid_mapping)) {
 277         mapcache->last_entry = NULL;
 278         if (!translated && mapcache->phys_offset_to_gaddr) {
 279             phys_addr = mapcache->phys_offset_to_gaddr(phys_addr, size, mapcache->opaque);
 280             translated = true;
 281             goto tryagain;
 282         }
 283         trace_xen_map_cache_return(NULL);
 284         return NULL;
 285     }
 286
 287     mapcache->last_entry = entry;
 288     if (lock) {
 289         MapCacheRev *reventry = g_malloc0(sizeof(MapCacheRev));
 290         entry->lock++;
 291         reventry->vaddr_req = mapcache->last_entry->vaddr_base + address_offset;
 292         reventry->paddr_index = mapcache->last_entry->paddr_index;
 293         reventry->size = entry->size;
 294         QTAILQ_INSERT_HEAD(&mapcache->locked_entries, reventry, next);
 295     }
 296
 297     trace_xen_map_cache_return(mapcache->last_entry->vaddr_base + address_offset);
 298     return mapcache->last_entry->vaddr_base + address_offset;
 299 }
 300
 301 uint8_t *xen_map_cache(hwaddr phys_addr, hwaddr size,
 302                        uint8_t lock)
 303 {
 304     uint8_t *p;
 305
 306     mapcache_lock();
 307     p = xen_map_cache_unlocked(phys_addr, size, lock);
 308     mapcache_unlock();
 309     return p;
 310 }
 311
 312 ram_addr_t xen_ram_addr_from_mapcache(void *ptr)
 313 {
 314     MapCacheEntry *entry = NULL;
 315     MapCacheRev *reventry;
 316     hwaddr paddr_index;
 317     hwaddr size;
 318     ram_addr_t raddr;
 319     int found = 0;
 320
 321     mapcache_lock();
 322     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 323         if (reventry->vaddr_req == ptr) {
 324             paddr_index = reventry->paddr_index;
 325             size = reventry->size;
 326             found = 1;
 327             break;
 328         }
 329     }
 330     if (!found) {
 331         fprintf(stderr, "%s, could not find %p\n", __func__, ptr);
 332         QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 333             DPRINTF("   "TARGET_FMT_plx" -> %p is present\n", reventry->paddr_index,
 334                     reventry->vaddr_req);
 335         }
 336         abort();
 337         return 0;
 338     }
 339
 340     entry = &mapcache->entry[paddr_index % mapcache->nr_buckets];
 341     while (entry && (entry->paddr_index != paddr_index || entry->size != size)) {
 342         entry = entry->next;
 343     }
 344     if (!entry) {
 345         DPRINTF("Trying to find address %p that is not in the mapcache!\n", ptr);
 346         raddr = 0;
 347     } else {
 348         raddr = (reventry->paddr_index << MCACHE_BUCKET_SHIFT) +
 349              ((unsigned long) ptr - (unsigned long) entry->vaddr_base);
 350     }
 351     mapcache_unlock();
 352     return raddr;
 353 }
 354
 355 static void xen_invalidate_map_cache_entry_unlocked(uint8_t *buffer)
 356 {
 357     MapCacheEntry *entry = NULL, *pentry = NULL;
 358     MapCacheRev *reventry;
 359     hwaddr paddr_index;
 360     hwaddr size;
 361     int found = 0;
 362
 363     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 364         if (reventry->vaddr_req == buffer) {
 365             paddr_index = reventry->paddr_index;
 366             size = reventry->size;
 367             found = 1;
 368             break;
 369         }
 370     }
 371     if (!found) {
 372         DPRINTF("%s, could not find %p\n", __func__, buffer);
 373         QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 374             DPRINTF("   "TARGET_FMT_plx" -> %p is present\n", reventry->paddr_index, reventry->vaddr_req);
 375         }
 376         return;
 377     }
 378     QTAILQ_REMOVE(&mapcache->locked_entries, reventry, next);
 379     g_free(reventry);
 380
 381     if (mapcache->last_entry != NULL &&
 382         mapcache->last_entry->paddr_index == paddr_index) {
 383         mapcache->last_entry = NULL;
 384     }
 385
 386     entry = &mapcache->entry[paddr_index % mapcache->nr_buckets];
 387     while (entry && (entry->paddr_index != paddr_index || entry->size != size)) {
 388         pentry = entry;
 389         entry = entry->next;
 390     }
 391     if (!entry) {
 392         DPRINTF("Trying to unmap address %p that is not in the mapcache!\n", buffer);
 393         return;
 394     }
 395     entry->lock--;
 396     if (entry->lock > 0 || pentry == NULL) {
 397         return;
 398     }
 399
 400     pentry->next = entry->next;
 401     if (munmap(entry->vaddr_base, entry->size) != 0) {
 402         perror("unmap fails");
 403         exit(-1);
 404     }
 405     g_free(entry->valid_mapping);
 406     g_free(entry);
 407 }
 408
 409 void xen_invalidate_map_cache_entry(uint8_t *buffer)
 410 {
 411     mapcache_lock();
 412     xen_invalidate_map_cache_entry_unlocked(buffer);
 413     mapcache_unlock();
 414 }
 415
 416 void xen_invalidate_map_cache(void)
 417 {
 418     unsigned long i;
 419     MapCacheRev *reventry;
 420
 421     /* Flush pending AIO before destroying the mapcache */
 422     bdrv_drain_all();
 423
 424     mapcache_lock();
 425
 426     QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
 427         DPRINTF("There should be no locked mappings at this time, "
 428                 "but "TARGET_FMT_plx" -> %p is present\n",
 429                 reventry->paddr_index, reventry->vaddr_req);
 430     }
 431
 432     for (i = 0; i < mapcache->nr_buckets; i++) {
 433         MapCacheEntry *entry = &mapcache->entry[i];
 434
 435         if (entry->vaddr_base == NULL) {
 436             continue;
 437         }
 438         if (entry->lock > 0) {
 439             continue;
 440         }
 441
 442         if (munmap(entry->vaddr_base, entry->size) != 0) {
 443             perror("unmap fails");
 444             exit(-1);
 445         }
 446
 447         entry->paddr_index = 0;
 448         entry->vaddr_base = NULL;
 449         entry->size = 0;
 450         g_free(entry->valid_mapping);
 451         entry->valid_mapping = NULL;
 452     }
 453
 454     mapcache->last_entry = NULL;
 455
 456     mapcache_unlock();
 457 }