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