include/exec/ram_addr.h

   1 /*
   2  * Declarations for cpu physical memory functions
   3  *
   4  * Copyright 2011 Red Hat, Inc. and/or its affiliates
   5  *
   6  * Authors:
   7  *  Avi Kivity <avi@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2 or
  10  * later.  See the COPYING file in the top-level directory.
  11  *
  12  */
  13
  14 /*
  15  * This header is for use by exec.c and memory.c ONLY.  Do not include it.
  16  * The functions declared here will be removed soon.
  17  */
  18
  19 #ifndef RAM_ADDR_H
  20 #define RAM_ADDR_H
  21
  22 #ifndef CONFIG_USER_ONLY
  23 #include "cpu.h"
  24 #include "hw/xen/xen.h"
  25 #include "sysemu/tcg.h"
  26 #include "exec/ramlist.h"
  27
  28 struct RAMBlock {
  29     struct rcu_head rcu;
  30     struct MemoryRegion *mr;
  31     uint8_t *host;
  32     uint8_t *colo_cache; /* For colo, VM's ram cache */
  33     ram_addr_t offset;
  34     ram_addr_t used_length;
  35     ram_addr_t max_length;
  36     void (*resized)(const char*, uint64_t length, void *host);
  37     uint32_t flags;
  38     /* Protected by iothread lock.  */
  39     char idstr[256];
  40     /* RCU-enabled, writes protected by the ramlist lock */
  41     QLIST_ENTRY(RAMBlock) next;
  42     QLIST_HEAD(, RAMBlockNotifier) ramblock_notifiers;
  43     int fd;
  44     size_t page_size;
  45     /* dirty bitmap used during migration */
  46     unsigned long *bmap;
  47     /* bitmap of already received pages in postcopy */
  48     unsigned long *receivedmap;
  49
  50     /*
  51      * bitmap to track already cleared dirty bitmap.  When the bit is
  52      * set, it means the corresponding memory chunk needs a log-clear.
  53      * Set this up to non-NULL to enable the capability to postpone
  54      * and split clearing of dirty bitmap on the remote node (e.g.,
  55      * KVM).  The bitmap will be set only when doing global sync.
  56      *
  57      * NOTE: this bitmap is different comparing to the other bitmaps
  58      * in that one bit can represent multiple guest pages (which is
  59      * decided by the `clear_bmap_shift' variable below).  On
  60      * destination side, this should always be NULL, and the variable
  61      * `clear_bmap_shift' is meaningless.
  62      */
  63     unsigned long *clear_bmap;
  64     uint8_t clear_bmap_shift;
  65 };
  66
  67 /**
  68  * clear_bmap_size: calculate clear bitmap size
  69  *
  70  * @pages: number of guest pages
  71  * @shift: guest page number shift
  72  *
  73  * Returns: number of bits for the clear bitmap
  74  */
  75 static inline long clear_bmap_size(uint64_t pages, uint8_t shift)
  76 {
  77     return DIV_ROUND_UP(pages, 1UL << shift);
  78 }
  79
  80 /**
  81  * clear_bmap_set: set clear bitmap for the page range
  82  *
  83  * @rb: the ramblock to operate on
  84  * @start: the start page number
  85  * @size: number of pages to set in the bitmap
  86  *
  87  * Returns: None
  88  */
  89 static inline void clear_bmap_set(RAMBlock *rb, uint64_t start,
  90                                   uint64_t npages)
  91 {
  92     uint8_t shift = rb->clear_bmap_shift;
  93
  94     bitmap_set_atomic(rb->clear_bmap, start >> shift,
  95                       clear_bmap_size(npages, shift));
  96 }
  97
  98 /**
  99  * clear_bmap_test_and_clear: test clear bitmap for the page, clear if set
 100  *
 101  * @rb: the ramblock to operate on
 102  * @page: the page number to check
 103  *
 104  * Returns: true if the bit was set, false otherwise
 105  */
 106 static inline bool clear_bmap_test_and_clear(RAMBlock *rb, uint64_t page)
 107 {
 108     uint8_t shift = rb->clear_bmap_shift;
 109
 110     return bitmap_test_and_clear_atomic(rb->clear_bmap, page >> shift, 1);
 111 }
 112
 113 static inline bool offset_in_ramblock(RAMBlock *b, ram_addr_t offset)
 114 {
 115     return (b && b->host && offset < b->used_length) ? true : false;
 116 }
 117
 118 static inline void *ramblock_ptr(RAMBlock *block, ram_addr_t offset)
 119 {
 120     assert(offset_in_ramblock(block, offset));
 121     return (char *)block->host + offset;
 122 }
 123
 124 static inline unsigned long int ramblock_recv_bitmap_offset(void *host_addr,
 125                                                             RAMBlock *rb)
 126 {
 127     uint64_t host_addr_offset =
 128             (uint64_t)(uintptr_t)(host_addr - (void *)rb->host);
 129     return host_addr_offset >> TARGET_PAGE_BITS;
 130 }
 131
 132 bool ramblock_is_pmem(RAMBlock *rb);
 133
 134 long qemu_minrampagesize(void);
 135 long qemu_maxrampagesize(void);
 136
 137 /**
 138  * qemu_ram_alloc_from_file,
 139  * qemu_ram_alloc_from_fd:  Allocate a ram block from the specified backing
 140  *                          file or device
 141  *
 142  * Parameters:
 143  *  @size: the size in bytes of the ram block
 144  *  @mr: the memory region where the ram block is
 145  *  @ram_flags: specify the properties of the ram block, which can be one
 146  *              or bit-or of following values
 147  *              - RAM_SHARED: mmap the backing file or device with MAP_SHARED
 148  *              - RAM_PMEM: the backend @mem_path or @fd is persistent memory
 149  *              Other bits are ignored.
 150  *  @mem_path or @fd: specify the backing file or device
 151  *  @errp: pointer to Error*, to store an error if it happens
 152  *
 153  * Return:
 154  *  On success, return a pointer to the ram block.
 155  *  On failure, return NULL.
 156  */
 157 RAMBlock *qemu_ram_alloc_from_file(ram_addr_t size, MemoryRegion *mr,
 158                                    uint32_t ram_flags, const char *mem_path,
 159                                    Error **errp);
 160 RAMBlock *qemu_ram_alloc_from_fd(ram_addr_t size, MemoryRegion *mr,
 161                                  uint32_t ram_flags, int fd,
 162                                  Error **errp);
 163
 164 RAMBlock *qemu_ram_alloc_from_ptr(ram_addr_t size, void *host,
 165                                   MemoryRegion *mr, Error **errp);
 166 RAMBlock *qemu_ram_alloc(ram_addr_t size, bool share, MemoryRegion *mr,
 167                          Error **errp);
 168 RAMBlock *qemu_ram_alloc_resizeable(ram_addr_t size, ram_addr_t max_size,
 169                                     void (*resized)(const char*,
 170                                                     uint64_t length,
 171                                                     void *host),
 172                                     MemoryRegion *mr, Error **errp);
 173 void qemu_ram_free(RAMBlock *block);
 174
 175 int qemu_ram_resize(RAMBlock *block, ram_addr_t newsize, Error **errp);
 176
 177 void qemu_ram_writeback(RAMBlock *block, ram_addr_t start, ram_addr_t length);
 178
 179 /* Clear whole block of mem */
 180 static inline void qemu_ram_block_writeback(RAMBlock *block)
 181 {
 182     qemu_ram_writeback(block, 0, block->used_length);
 183 }
 184
 185 #define DIRTY_CLIENTS_ALL     ((1 << DIRTY_MEMORY_NUM) - 1)
 186 #define DIRTY_CLIENTS_NOCODE  (DIRTY_CLIENTS_ALL & ~(1 << DIRTY_MEMORY_CODE))
 187
 188 void tb_invalidate_phys_range(ram_addr_t start, ram_addr_t end);
 189
 190 static inline bool cpu_physical_memory_get_dirty(ram_addr_t start,
 191                                                  ram_addr_t length,
 192                                                  unsigned client)
 193 {
 194     DirtyMemoryBlocks *blocks;
 195     unsigned long end, page;
 196     unsigned long idx, offset, base;
 197     bool dirty = false;
 198
 199     assert(client < DIRTY_MEMORY_NUM);
 200
 201     end = TARGET_PAGE_ALIGN(start + length) >> TARGET_PAGE_BITS;
 202     page = start >> TARGET_PAGE_BITS;
 203
 204     WITH_RCU_READ_LOCK_GUARD() {
 205         blocks = atomic_rcu_read(&ram_list.dirty_memory[client]);
 206
 207         idx = page / DIRTY_MEMORY_BLOCK_SIZE;
 208         offset = page % DIRTY_MEMORY_BLOCK_SIZE;
 209         base = page - offset;
 210         while (page < end) {
 211             unsigned long next = MIN(end, base + DIRTY_MEMORY_BLOCK_SIZE);
 212             unsigned long num = next - base;
 213             unsigned long found = find_next_bit(blocks->blocks[idx],
 214                                                 num, offset);
 215             if (found < num) {
 216                 dirty = true;
 217                 break;
 218             }
 219
 220             page = next;
 221             idx++;
 222             offset = 0;
 223             base += DIRTY_MEMORY_BLOCK_SIZE;
 224         }
 225     }
 226
 227     return dirty;
 228 }
 229
 230 static inline bool cpu_physical_memory_all_dirty(ram_addr_t start,
 231                                                  ram_addr_t length,
 232                                                  unsigned client)
 233 {
 234     DirtyMemoryBlocks *blocks;
 235     unsigned long end, page;
 236     unsigned long idx, offset, base;
 237     bool dirty = true;
 238
 239     assert(client < DIRTY_MEMORY_NUM);
 240
 241     end = TARGET_PAGE_ALIGN(start + length) >> TARGET_PAGE_BITS;
 242     page = start >> TARGET_PAGE_BITS;
 243
 244     RCU_READ_LOCK_GUARD();
 245
 246     blocks = atomic_rcu_read(&ram_list.dirty_memory[client]);
 247
 248     idx = page / DIRTY_MEMORY_BLOCK_SIZE;
 249     offset = page % DIRTY_MEMORY_BLOCK_SIZE;
 250     base = page - offset;
 251     while (page < end) {
 252         unsigned long next = MIN(end, base + DIRTY_MEMORY_BLOCK_SIZE);
 253         unsigned long num = next - base;
 254         unsigned long found = find_next_zero_bit(blocks->blocks[idx], num, offset);
 255         if (found < num) {
 256             dirty = false;
 257             break;
 258         }
 259
 260         page = next;
 261         idx++;
 262         offset = 0;
 263         base += DIRTY_MEMORY_BLOCK_SIZE;
 264     }
 265
 266     return dirty;
 267 }
 268
 269 static inline bool cpu_physical_memory_get_dirty_flag(ram_addr_t addr,
 270                                                       unsigned client)
 271 {
 272     return cpu_physical_memory_get_dirty(addr, 1, client);
 273 }
 274
 275 static inline bool cpu_physical_memory_is_clean(ram_addr_t addr)
 276 {
 277     bool vga = cpu_physical_memory_get_dirty_flag(addr, DIRTY_MEMORY_VGA);
 278     bool code = cpu_physical_memory_get_dirty_flag(addr, DIRTY_MEMORY_CODE);
 279     bool migration =
 280         cpu_physical_memory_get_dirty_flag(addr, DIRTY_MEMORY_MIGRATION);
 281     return !(vga && code && migration);
 282 }
 283
 284 static inline uint8_t cpu_physical_memory_range_includes_clean(ram_addr_t start,
 285                                                                ram_addr_t length,
 286                                                                uint8_t mask)
 287 {
 288     uint8_t ret = 0;
 289
 290     if (mask & (1 << DIRTY_MEMORY_VGA) &&
 291         !cpu_physical_memory_all_dirty(start, length, DIRTY_MEMORY_VGA)) {
 292         ret |= (1 << DIRTY_MEMORY_VGA);
 293     }
 294     if (mask & (1 << DIRTY_MEMORY_CODE) &&
 295         !cpu_physical_memory_all_dirty(start, length, DIRTY_MEMORY_CODE)) {
 296         ret |= (1 << DIRTY_MEMORY_CODE);
 297     }
 298     if (mask & (1 << DIRTY_MEMORY_MIGRATION) &&
 299         !cpu_physical_memory_all_dirty(start, length, DIRTY_MEMORY_MIGRATION)) {
 300         ret |= (1 << DIRTY_MEMORY_MIGRATION);
 301     }
 302     return ret;
 303 }
 304
 305 static inline void cpu_physical_memory_set_dirty_flag(ram_addr_t addr,
 306                                                       unsigned client)
 307 {
 308     unsigned long page, idx, offset;
 309     DirtyMemoryBlocks *blocks;
 310
 311     assert(client < DIRTY_MEMORY_NUM);
 312
 313     page = addr >> TARGET_PAGE_BITS;
 314     idx = page / DIRTY_MEMORY_BLOCK_SIZE;
 315     offset = page % DIRTY_MEMORY_BLOCK_SIZE;
 316
 317     RCU_READ_LOCK_GUARD();
 318
 319     blocks = atomic_rcu_read(&ram_list.dirty_memory[client]);
 320
 321     set_bit_atomic(offset, blocks->blocks[idx]);
 322 }
 323
 324 static inline void cpu_physical_memory_set_dirty_range(ram_addr_t start,
 325                                                        ram_addr_t length,
 326                                                        uint8_t mask)
 327 {
 328     DirtyMemoryBlocks *blocks[DIRTY_MEMORY_NUM];
 329     unsigned long end, page;
 330     unsigned long idx, offset, base;
 331     int i;
 332
 333     if (!mask && !xen_enabled()) {
 334         return;
 335     }
 336
 337     end = TARGET_PAGE_ALIGN(start + length) >> TARGET_PAGE_BITS;
 338     page = start >> TARGET_PAGE_BITS;
 339
 340     WITH_RCU_READ_LOCK_GUARD() {
 341         for (i = 0; i < DIRTY_MEMORY_NUM; i++) {
 342             blocks[i] = atomic_rcu_read(&ram_list.dirty_memory[i]);
 343         }
 344
 345         idx = page / DIRTY_MEMORY_BLOCK_SIZE;
 346         offset = page % DIRTY_MEMORY_BLOCK_SIZE;
 347         base = page - offset;
 348         while (page < end) {
 349             unsigned long next = MIN(end, base + DIRTY_MEMORY_BLOCK_SIZE);
 350
 351             if (likely(mask & (1 << DIRTY_MEMORY_MIGRATION))) {
 352                 bitmap_set_atomic(blocks[DIRTY_MEMORY_MIGRATION]->blocks[idx],
 353                                   offset, next - page);
 354             }
 355             if (unlikely(mask & (1 << DIRTY_MEMORY_VGA))) {
 356                 bitmap_set_atomic(blocks[DIRTY_MEMORY_VGA]->blocks[idx],
 357                                   offset, next - page);
 358             }
 359             if (unlikely(mask & (1 << DIRTY_MEMORY_CODE))) {
 360                 bitmap_set_atomic(blocks[DIRTY_MEMORY_CODE]->blocks[idx],
 361                                   offset, next - page);
 362             }
 363
 364             page = next;
 365             idx++;
 366             offset = 0;
 367             base += DIRTY_MEMORY_BLOCK_SIZE;
 368         }
 369     }
 370
 371     xen_hvm_modified_memory(start, length);
 372 }
 373
 374 #if !defined(_WIN32)
 375 static inline void cpu_physical_memory_set_dirty_lebitmap(unsigned long *bitmap,
 376                                                           ram_addr_t start,
 377                                                           ram_addr_t pages)
 378 {
 379     unsigned long i, j;
 380     unsigned long page_number, c;
 381     hwaddr addr;
 382     ram_addr_t ram_addr;
 383     unsigned long len = (pages + HOST_LONG_BITS - 1) / HOST_LONG_BITS;
 384     unsigned long hpratio = qemu_real_host_page_size / TARGET_PAGE_SIZE;
 385     unsigned long page = BIT_WORD(start >> TARGET_PAGE_BITS);
 386
 387     /* start address is aligned at the start of a word? */
 388     if ((((page * BITS_PER_LONG) << TARGET_PAGE_BITS) == start) &&
 389         (hpratio == 1)) {
 390         unsigned long **blocks[DIRTY_MEMORY_NUM];
 391         unsigned long idx;
 392         unsigned long offset;
 393         long k;
 394         long nr = BITS_TO_LONGS(pages);
 395
 396         idx = (start >> TARGET_PAGE_BITS) / DIRTY_MEMORY_BLOCK_SIZE;
 397         offset = BIT_WORD((start >> TARGET_PAGE_BITS) %
 398                           DIRTY_MEMORY_BLOCK_SIZE);
 399
 400         WITH_RCU_READ_LOCK_GUARD() {
 401             for (i = 0; i < DIRTY_MEMORY_NUM; i++) {
 402                 blocks[i] = atomic_rcu_read(&ram_list.dirty_memory[i])->blocks;
 403             }
 404
 405             for (k = 0; k < nr; k++) {
 406                 if (bitmap[k]) {
 407                     unsigned long temp = leul_to_cpu(bitmap[k]);
 408
 409                     atomic_or(&blocks[DIRTY_MEMORY_VGA][idx][offset], temp);
 410
 411                     if (global_dirty_log) {
 412                         atomic_or(&blocks[DIRTY_MEMORY_MIGRATION][idx][offset],
 413                                   temp);
 414                     }
 415
 416                     if (tcg_enabled()) {
 417                         atomic_or(&blocks[DIRTY_MEMORY_CODE][idx][offset],
 418                                   temp);
 419                     }
 420                 }
 421
 422                 if (++offset >= BITS_TO_LONGS(DIRTY_MEMORY_BLOCK_SIZE)) {
 423                     offset = 0;
 424                     idx++;
 425                 }
 426             }
 427         }
 428
 429         xen_hvm_modified_memory(start, pages << TARGET_PAGE_BITS);
 430     } else {
 431         uint8_t clients = tcg_enabled() ? DIRTY_CLIENTS_ALL : DIRTY_CLIENTS_NOCODE;
 432
 433         if (!global_dirty_log) {
 434             clients &= ~(1 << DIRTY_MEMORY_MIGRATION);
 435         }
 436
 437         /*
 438          * bitmap-traveling is faster than memory-traveling (for addr...)
 439          * especially when most of the memory is not dirty.
 440          */
 441         for (i = 0; i < len; i++) {
 442             if (bitmap[i] != 0) {
 443                 c = leul_to_cpu(bitmap[i]);
 444                 do {
 445                     j = ctzl(c);
 446                     c &= ~(1ul << j);
 447                     page_number = (i * HOST_LONG_BITS + j) * hpratio;
 448                     addr = page_number * TARGET_PAGE_SIZE;
 449                     ram_addr = start + addr;
 450                     cpu_physical_memory_set_dirty_range(ram_addr,
 451                                        TARGET_PAGE_SIZE * hpratio, clients);
 452                 } while (c != 0);
 453             }
 454         }
 455     }
 456 }
 457 #endif /* not _WIN32 */
 458
 459 bool cpu_physical_memory_test_and_clear_dirty(ram_addr_t start,
 460                                               ram_addr_t length,
 461                                               unsigned client);
 462
 463 DirtyBitmapSnapshot *cpu_physical_memory_snapshot_and_clear_dirty
 464     (MemoryRegion *mr, hwaddr offset, hwaddr length, unsigned client);
 465
 466 bool cpu_physical_memory_snapshot_get_dirty(DirtyBitmapSnapshot *snap,
 467                                             ram_addr_t start,
 468                                             ram_addr_t length);
 469
 470 static inline void cpu_physical_memory_clear_dirty_range(ram_addr_t start,
 471                                                          ram_addr_t length)
 472 {
 473     cpu_physical_memory_test_and_clear_dirty(start, length, DIRTY_MEMORY_MIGRATION);
 474     cpu_physical_memory_test_and_clear_dirty(start, length, DIRTY_MEMORY_VGA);
 475     cpu_physical_memory_test_and_clear_dirty(start, length, DIRTY_MEMORY_CODE);
 476 }
 477
 478
 479 /* Called with RCU critical section */
 480 static inline
 481 uint64_t cpu_physical_memory_sync_dirty_bitmap(RAMBlock *rb,
 482                                                ram_addr_t start,
 483                                                ram_addr_t length,
 484                                                uint64_t *real_dirty_pages)
 485 {
 486     ram_addr_t addr;
 487     unsigned long word = BIT_WORD((start + rb->offset) >> TARGET_PAGE_BITS);
 488     uint64_t num_dirty = 0;
 489     unsigned long *dest = rb->bmap;
 490
 491     /* start address and length is aligned at the start of a word? */
 492     if (((word * BITS_PER_LONG) << TARGET_PAGE_BITS) ==
 493          (start + rb->offset) &&
 494         !(length & ((BITS_PER_LONG << TARGET_PAGE_BITS) - 1))) {
 495         int k;
 496         int nr = BITS_TO_LONGS(length >> TARGET_PAGE_BITS);
 497         unsigned long * const *src;
 498         unsigned long idx = (word * BITS_PER_LONG) / DIRTY_MEMORY_BLOCK_SIZE;
 499         unsigned long offset = BIT_WORD((word * BITS_PER_LONG) %
 500                                         DIRTY_MEMORY_BLOCK_SIZE);
 501         unsigned long page = BIT_WORD(start >> TARGET_PAGE_BITS);
 502
 503         src = atomic_rcu_read(
 504                 &ram_list.dirty_memory[DIRTY_MEMORY_MIGRATION])->blocks;
 505
 506         for (k = page; k < page + nr; k++) {
 507             if (src[idx][offset]) {
 508                 unsigned long bits = atomic_xchg(&src[idx][offset], 0);
 509                 unsigned long new_dirty;
 510                 *real_dirty_pages += ctpopl(bits);
 511                 new_dirty = ~dest[k];
 512                 dest[k] |= bits;
 513                 new_dirty &= bits;
 514                 num_dirty += ctpopl(new_dirty);
 515             }
 516
 517             if (++offset >= BITS_TO_LONGS(DIRTY_MEMORY_BLOCK_SIZE)) {
 518                 offset = 0;
 519                 idx++;
 520             }
 521         }
 522
 523         if (rb->clear_bmap) {
 524             /*
 525              * Postpone the dirty bitmap clear to the point before we
 526              * really send the pages, also we will split the clear
 527              * dirty procedure into smaller chunks.
 528              */
 529             clear_bmap_set(rb, start >> TARGET_PAGE_BITS,
 530                            length >> TARGET_PAGE_BITS);
 531         } else {
 532             /* Slow path - still do that in a huge chunk */
 533             memory_region_clear_dirty_bitmap(rb->mr, start, length);
 534         }
 535     } else {
 536         ram_addr_t offset = rb->offset;
 537
 538         for (addr = 0; addr < length; addr += TARGET_PAGE_SIZE) {
 539             if (cpu_physical_memory_test_and_clear_dirty(
 540                         start + addr + offset,
 541                         TARGET_PAGE_SIZE,
 542                         DIRTY_MEMORY_MIGRATION)) {
 543                 *real_dirty_pages += 1;
 544                 long k = (start + addr) >> TARGET_PAGE_BITS;
 545                 if (!test_and_set_bit(k, dest)) {
 546                     num_dirty++;
 547                 }
 548             }
 549         }
 550     }
 551
 552     return num_dirty;
 553 }
 554 #endif
 555 #endif