tools/testing/selftests/kvm/dirty_log_test.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * KVM dirty page logging test
   4  *
   5  * Copyright (C) 2018, Red Hat, Inc.
   6  */
   7
   8 #define _GNU_SOURCE /* for program_invocation_name */
   9
  10 #include <stdio.h>
  11 #include <stdlib.h>
  12 #include <unistd.h>
  13 #include <time.h>
  14 #include <pthread.h>
  15 #include <linux/bitmap.h>
  16 #include <linux/bitops.h>
  17
  18 #include "test_util.h"
  19 #include "kvm_util.h"
  20 #include "processor.h"
  21
  22 #define DEBUG printf
  23
  24 #define VCPU_ID                         1
  25
  26 /* The memory slot index to track dirty pages */
  27 #define TEST_MEM_SLOT_INDEX             1
  28
  29 /* Default guest test virtual memory offset */
  30 #define DEFAULT_GUEST_TEST_MEM          0xc0000000
  31
  32 /* How many pages to dirty for each guest loop */
  33 #define TEST_PAGES_PER_LOOP             1024
  34
  35 /* How many host loops to run (one KVM_GET_DIRTY_LOG for each loop) */
  36 #define TEST_HOST_LOOP_N                32UL
  37
  38 /* Interval for each host loop (ms) */
  39 #define TEST_HOST_LOOP_INTERVAL         10UL
  40
  41 /* Dirty bitmaps are always little endian, so we need to swap on big endian */
  42 #if defined(__s390x__)
  43 # define BITOP_LE_SWIZZLE       ((BITS_PER_LONG-1) & ~0x7)
  44 # define test_bit_le(nr, addr) \
  45         test_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  46 # define set_bit_le(nr, addr) \
  47         set_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  48 # define clear_bit_le(nr, addr) \
  49         clear_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  50 # define test_and_set_bit_le(nr, addr) \
  51         test_and_set_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  52 # define test_and_clear_bit_le(nr, addr) \
  53         test_and_clear_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  54 #else
  55 # define test_bit_le            test_bit
  56 # define set_bit_le             set_bit
  57 # define clear_bit_le           clear_bit
  58 # define test_and_set_bit_le    test_and_set_bit
  59 # define test_and_clear_bit_le  test_and_clear_bit
  60 #endif
  61
  62 /*
  63  * Guest/Host shared variables. Ensure addr_gva2hva() and/or
  64  * sync_global_to/from_guest() are used when accessing from
  65  * the host. READ/WRITE_ONCE() should also be used with anything
  66  * that may change.
  67  */
  68 static uint64_t host_page_size;
  69 static uint64_t guest_page_size;
  70 static uint64_t guest_num_pages;
  71 static uint64_t random_array[TEST_PAGES_PER_LOOP];
  72 static uint64_t iteration;
  73
  74 /*
  75  * Guest physical memory offset of the testing memory slot.
  76  * This will be set to the topmost valid physical address minus
  77  * the test memory size.
  78  */
  79 static uint64_t guest_test_phys_mem;
  80
  81 /*
  82  * Guest virtual memory offset of the testing memory slot.
  83  * Must not conflict with identity mapped test code.
  84  */
  85 static uint64_t guest_test_virt_mem = DEFAULT_GUEST_TEST_MEM;
  86
  87 /*
  88  * Continuously write to the first 8 bytes of a random pages within
  89  * the testing memory region.
  90  */
  91 static void guest_code(void)
  92 {
  93         uint64_t addr;
  94         int i;
  95
  96         /*
  97          * On s390x, all pages of a 1M segment are initially marked as dirty
  98          * when a page of the segment is written to for the very first time.
  99          * To compensate this specialty in this test, we need to touch all
 100          * pages during the first iteration.
 101          */
 102         for (i = 0; i < guest_num_pages; i++) {
 103                 addr = guest_test_virt_mem + i * guest_page_size;
 104                 *(uint64_t *)addr = READ_ONCE(iteration);
 105         }
 106
 107         while (true) {
 108                 for (i = 0; i < TEST_PAGES_PER_LOOP; i++) {
 109                         addr = guest_test_virt_mem;
 110                         addr += (READ_ONCE(random_array[i]) % guest_num_pages)
 111                                 * guest_page_size;
 112                         addr &= ~(host_page_size - 1);
 113                         *(uint64_t *)addr = READ_ONCE(iteration);
 114                 }
 115
 116                 /* Tell the host that we need more random numbers */
 117                 GUEST_SYNC(1);
 118         }
 119 }
 120
 121 /* Host variables */
 122 static bool host_quit;
 123
 124 /* Points to the test VM memory region on which we track dirty logs */
 125 static void *host_test_mem;
 126 static uint64_t host_num_pages;
 127
 128 /* For statistics only */
 129 static uint64_t host_dirty_count;
 130 static uint64_t host_clear_count;
 131 static uint64_t host_track_next_count;
 132
 133 /*
 134  * We use this bitmap to track some pages that should have its dirty
 135  * bit set in the _next_ iteration.  For example, if we detected the
 136  * page value changed to current iteration but at the same time the
 137  * page bit is cleared in the latest bitmap, then the system must
 138  * report that write in the next get dirty log call.
 139  */
 140 static unsigned long *host_bmap_track;
 141
 142 static void generate_random_array(uint64_t *guest_array, uint64_t size)
 143 {
 144         uint64_t i;
 145
 146         for (i = 0; i < size; i++)
 147                 guest_array[i] = random();
 148 }
 149
 150 static void *vcpu_worker(void *data)
 151 {
 152         int ret;
 153         struct kvm_vm *vm = data;
 154         uint64_t *guest_array;
 155         uint64_t pages_count = 0;
 156         struct kvm_run *run;
 157
 158         run = vcpu_state(vm, VCPU_ID);
 159
 160         guest_array = addr_gva2hva(vm, (vm_vaddr_t)random_array);
 161         generate_random_array(guest_array, TEST_PAGES_PER_LOOP);
 162
 163         while (!READ_ONCE(host_quit)) {
 164                 /* Let the guest dirty the random pages */
 165                 ret = _vcpu_run(vm, VCPU_ID);
 166                 TEST_ASSERT(ret == 0, "vcpu_run failed: %d\n", ret);
 167                 if (get_ucall(vm, VCPU_ID, NULL) == UCALL_SYNC) {
 168                         pages_count += TEST_PAGES_PER_LOOP;
 169                         generate_random_array(guest_array, TEST_PAGES_PER_LOOP);
 170                 } else {
 171                         TEST_ASSERT(false,
 172                                     "Invalid guest sync status: "
 173                                     "exit_reason=%s\n",
 174                                     exit_reason_str(run->exit_reason));
 175                 }
 176         }
 177
 178         DEBUG("Dirtied %"PRIu64" pages\n", pages_count);
 179
 180         return NULL;
 181 }
 182
 183 static void vm_dirty_log_verify(unsigned long *bmap)
 184 {
 185         uint64_t page;
 186         uint64_t *value_ptr;
 187         uint64_t step = host_page_size >= guest_page_size ? 1 :
 188                                 guest_page_size / host_page_size;
 189
 190         for (page = 0; page < host_num_pages; page += step) {
 191                 value_ptr = host_test_mem + page * host_page_size;
 192
 193                 /* If this is a special page that we were tracking... */
 194                 if (test_and_clear_bit_le(page, host_bmap_track)) {
 195                         host_track_next_count++;
 196                         TEST_ASSERT(test_bit_le(page, bmap),
 197                                     "Page %"PRIu64" should have its dirty bit "
 198                                     "set in this iteration but it is missing",
 199                                     page);
 200                 }
 201
 202                 if (test_bit_le(page, bmap)) {
 203                         host_dirty_count++;
 204                         /*
 205                          * If the bit is set, the value written onto
 206                          * the corresponding page should be either the
 207                          * previous iteration number or the current one.
 208                          */
 209                         TEST_ASSERT(*value_ptr == iteration ||
 210                                     *value_ptr == iteration - 1,
 211                                     "Set page %"PRIu64" value %"PRIu64
 212                                     " incorrect (iteration=%"PRIu64")",
 213                                     page, *value_ptr, iteration);
 214                 } else {
 215                         host_clear_count++;
 216                         /*
 217                          * If cleared, the value written can be any
 218                          * value smaller or equals to the iteration
 219                          * number.  Note that the value can be exactly
 220                          * (iteration-1) if that write can happen
 221                          * like this:
 222                          *
 223                          * (1) increase loop count to "iteration-1"
 224                          * (2) write to page P happens (with value
 225                          *     "iteration-1")
 226                          * (3) get dirty log for "iteration-1"; we'll
 227                          *     see that page P bit is set (dirtied),
 228                          *     and not set the bit in host_bmap_track
 229                          * (4) increase loop count to "iteration"
 230                          *     (which is current iteration)
 231                          * (5) get dirty log for current iteration,
 232                          *     we'll see that page P is cleared, with
 233                          *     value "iteration-1".
 234                          */
 235                         TEST_ASSERT(*value_ptr <= iteration,
 236                                     "Clear page %"PRIu64" value %"PRIu64
 237                                     " incorrect (iteration=%"PRIu64")",
 238                                     page, *value_ptr, iteration);
 239                         if (*value_ptr == iteration) {
 240                                 /*
 241                                  * This page is _just_ modified; it
 242                                  * should report its dirtyness in the
 243                                  * next run
 244                                  */
 245                                 set_bit_le(page, host_bmap_track);
 246                         }
 247                 }
 248         }
 249 }
 250
 251 static struct kvm_vm *create_vm(enum vm_guest_mode mode, uint32_t vcpuid,
 252                                 uint64_t extra_mem_pages, void *guest_code)
 253 {
 254         struct kvm_vm *vm;
 255         uint64_t extra_pg_pages = extra_mem_pages / 512 * 2;
 256
 257         vm = _vm_create(mode, DEFAULT_GUEST_PHY_PAGES + extra_pg_pages, O_RDWR);
 258         kvm_vm_elf_load(vm, program_invocation_name, 0, 0);
 259 #ifdef __x86_64__
 260         vm_create_irqchip(vm);
 261 #endif
 262         vm_vcpu_add_default(vm, vcpuid, guest_code);
 263         return vm;
 264 }
 265
 266 static void run_test(enum vm_guest_mode mode, unsigned long iterations,
 267                      unsigned long interval, uint64_t phys_offset)
 268 {
 269         unsigned int guest_pa_bits, guest_page_shift;
 270         pthread_t vcpu_thread;
 271         struct kvm_vm *vm;
 272         uint64_t max_gfn;
 273         unsigned long *bmap;
 274
 275         switch (mode) {
 276         case VM_MODE_P52V48_4K:
 277                 guest_pa_bits = 52;
 278                 guest_page_shift = 12;
 279                 break;
 280         case VM_MODE_P52V48_64K:
 281                 guest_pa_bits = 52;
 282                 guest_page_shift = 16;
 283                 break;
 284         case VM_MODE_P48V48_4K:
 285                 guest_pa_bits = 48;
 286                 guest_page_shift = 12;
 287                 break;
 288         case VM_MODE_P48V48_64K:
 289                 guest_pa_bits = 48;
 290                 guest_page_shift = 16;
 291                 break;
 292         case VM_MODE_P40V48_4K:
 293                 guest_pa_bits = 40;
 294                 guest_page_shift = 12;
 295                 break;
 296         case VM_MODE_P40V48_64K:
 297                 guest_pa_bits = 40;
 298                 guest_page_shift = 16;
 299                 break;
 300         default:
 301                 TEST_ASSERT(false, "Unknown guest mode, mode: 0x%x", mode);
 302         }
 303
 304         DEBUG("Testing guest mode: %s\n", vm_guest_mode_string(mode));
 305
 306 #ifdef __x86_64__
 307         /*
 308          * FIXME
 309          * The x86_64 kvm selftests framework currently only supports a
 310          * single PML4 which restricts the number of physical address
 311          * bits we can change to 39.
 312          */
 313         guest_pa_bits = 39;
 314 #endif
 315         max_gfn = (1ul << (guest_pa_bits - guest_page_shift)) - 1;
 316         guest_page_size = (1ul << guest_page_shift);
 317         /*
 318          * A little more than 1G of guest page sized pages.  Cover the
 319          * case where the size is not aligned to 64 pages.
 320          */
 321         guest_num_pages = (1ul << (30 - guest_page_shift)) + 16;
 322 #ifdef __s390x__
 323         /* Round up to multiple of 1M (segment size) */
 324         guest_num_pages = (guest_num_pages + 0xff) & ~0xffUL;
 325 #endif
 326         host_page_size = getpagesize();
 327         host_num_pages = (guest_num_pages * guest_page_size) / host_page_size +
 328                          !!((guest_num_pages * guest_page_size) % host_page_size);
 329
 330         if (!phys_offset) {
 331                 guest_test_phys_mem = (max_gfn - guest_num_pages) * guest_page_size;
 332                 guest_test_phys_mem &= ~(host_page_size - 1);
 333         } else {
 334                 guest_test_phys_mem = phys_offset;
 335         }
 336
 337 #ifdef __s390x__
 338         /* Align to 1M (segment size) */
 339         guest_test_phys_mem &= ~((1 << 20) - 1);
 340 #endif
 341
 342         DEBUG("guest physical test memory offset: 0x%lx\n", guest_test_phys_mem);
 343
 344         bmap = bitmap_alloc(host_num_pages);
 345         host_bmap_track = bitmap_alloc(host_num_pages);
 346
 347         vm = create_vm(mode, VCPU_ID, guest_num_pages, guest_code);
 348
 349 #ifdef USE_CLEAR_DIRTY_LOG
 350         struct kvm_enable_cap cap = {};
 351
 352         cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
 353         cap.args[0] = 1;
 354         vm_enable_cap(vm, &cap);
 355 #endif
 356
 357         /* Add an extra memory slot for testing dirty logging */
 358         vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS,
 359                                     guest_test_phys_mem,
 360                                     TEST_MEM_SLOT_INDEX,
 361                                     guest_num_pages,
 362                                     KVM_MEM_LOG_DIRTY_PAGES);
 363
 364         /* Do mapping for the dirty track memory slot */
 365         virt_map(vm, guest_test_virt_mem, guest_test_phys_mem,
 366                  guest_num_pages * guest_page_size, 0);
 367
 368         /* Cache the HVA pointer of the region */
 369         host_test_mem = addr_gpa2hva(vm, (vm_paddr_t)guest_test_phys_mem);
 370
 371 #ifdef __x86_64__
 372         vcpu_set_cpuid(vm, VCPU_ID, kvm_get_supported_cpuid());
 373 #endif
 374 #ifdef __aarch64__
 375         ucall_init(vm, NULL);
 376 #endif
 377
 378         /* Export the shared variables to the guest */
 379         sync_global_to_guest(vm, host_page_size);
 380         sync_global_to_guest(vm, guest_page_size);
 381         sync_global_to_guest(vm, guest_test_virt_mem);
 382         sync_global_to_guest(vm, guest_num_pages);
 383
 384         /* Start the iterations */
 385         iteration = 1;
 386         sync_global_to_guest(vm, iteration);
 387         host_quit = false;
 388         host_dirty_count = 0;
 389         host_clear_count = 0;
 390         host_track_next_count = 0;
 391
 392         pthread_create(&vcpu_thread, NULL, vcpu_worker, vm);
 393
 394         while (iteration < iterations) {
 395                 /* Give the vcpu thread some time to dirty some pages */
 396                 usleep(interval * 1000);
 397                 kvm_vm_get_dirty_log(vm, TEST_MEM_SLOT_INDEX, bmap);
 398 #ifdef USE_CLEAR_DIRTY_LOG
 399                 kvm_vm_clear_dirty_log(vm, TEST_MEM_SLOT_INDEX, bmap, 0,
 400                                        host_num_pages);
 401 #endif
 402                 vm_dirty_log_verify(bmap);
 403                 iteration++;
 404                 sync_global_to_guest(vm, iteration);
 405         }
 406
 407         /* Tell the vcpu thread to quit */
 408         host_quit = true;
 409         pthread_join(vcpu_thread, NULL);
 410
 411         DEBUG("Total bits checked: dirty (%"PRIu64"), clear (%"PRIu64"), "
 412               "track_next (%"PRIu64")\n", host_dirty_count, host_clear_count,
 413               host_track_next_count);
 414
 415         free(bmap);
 416         free(host_bmap_track);
 417         ucall_uninit(vm);
 418         kvm_vm_free(vm);
 419 }
 420
 421 struct vm_guest_mode_params {
 422         bool supported;
 423         bool enabled;
 424 };
 425 struct vm_guest_mode_params vm_guest_mode_params[NUM_VM_MODES];
 426
 427 #define vm_guest_mode_params_init(mode, supported, enabled)                                     \
 428 ({                                                                                              \
 429         vm_guest_mode_params[mode] = (struct vm_guest_mode_params){ supported, enabled };       \
 430 })
 431
 432 static void help(char *name)
 433 {
 434         int i;
 435
 436         puts("");
 437         printf("usage: %s [-h] [-i iterations] [-I interval] "
 438                "[-p offset] [-m mode]\n", name);
 439         puts("");
 440         printf(" -i: specify iteration counts (default: %"PRIu64")\n",
 441                TEST_HOST_LOOP_N);
 442         printf(" -I: specify interval in ms (default: %"PRIu64" ms)\n",
 443                TEST_HOST_LOOP_INTERVAL);
 444         printf(" -p: specify guest physical test memory offset\n"
 445                "     Warning: a low offset can conflict with the loaded test code.\n");
 446         printf(" -m: specify the guest mode ID to test "
 447                "(default: test all supported modes)\n"
 448                "     This option may be used multiple times.\n"
 449                "     Guest mode IDs:\n");
 450         for (i = 0; i < NUM_VM_MODES; ++i) {
 451                 printf("         %d:    %s%s\n", i, vm_guest_mode_string(i),
 452                        vm_guest_mode_params[i].supported ? " (supported)" : "");
 453         }
 454         puts("");
 455         exit(0);
 456 }
 457
 458 int main(int argc, char *argv[])
 459 {
 460         unsigned long iterations = TEST_HOST_LOOP_N;
 461         unsigned long interval = TEST_HOST_LOOP_INTERVAL;
 462         bool mode_selected = false;
 463         uint64_t phys_offset = 0;
 464         unsigned int mode;
 465         int opt, i;
 466 #ifdef __aarch64__
 467         unsigned int host_ipa_limit;
 468 #endif
 469
 470 #ifdef USE_CLEAR_DIRTY_LOG
 471         if (!kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2)) {
 472                 fprintf(stderr, "KVM_CLEAR_DIRTY_LOG not available, skipping tests\n");
 473                 exit(KSFT_SKIP);
 474         }
 475 #endif
 476
 477 #ifdef __x86_64__
 478         vm_guest_mode_params_init(VM_MODE_P52V48_4K, true, true);
 479 #endif
 480 #ifdef __aarch64__
 481         vm_guest_mode_params_init(VM_MODE_P40V48_4K, true, true);
 482         vm_guest_mode_params_init(VM_MODE_P40V48_64K, true, true);
 483
 484         host_ipa_limit = kvm_check_cap(KVM_CAP_ARM_VM_IPA_SIZE);
 485         if (host_ipa_limit >= 52)
 486                 vm_guest_mode_params_init(VM_MODE_P52V48_64K, true, true);
 487         if (host_ipa_limit >= 48) {
 488                 vm_guest_mode_params_init(VM_MODE_P48V48_4K, true, true);
 489                 vm_guest_mode_params_init(VM_MODE_P48V48_64K, true, true);
 490         }
 491 #endif
 492 #ifdef __s390x__
 493         vm_guest_mode_params_init(VM_MODE_P40V48_4K, true, true);
 494 #endif
 495
 496         while ((opt = getopt(argc, argv, "hi:I:p:m:")) != -1) {
 497                 switch (opt) {
 498                 case 'i':
 499                         iterations = strtol(optarg, NULL, 10);
 500                         break;
 501                 case 'I':
 502                         interval = strtol(optarg, NULL, 10);
 503                         break;
 504                 case 'p':
 505                         phys_offset = strtoull(optarg, NULL, 0);
 506                         break;
 507                 case 'm':
 508                         if (!mode_selected) {
 509                                 for (i = 0; i < NUM_VM_MODES; ++i)
 510                                         vm_guest_mode_params[i].enabled = false;
 511                                 mode_selected = true;
 512                         }
 513                         mode = strtoul(optarg, NULL, 10);
 514                         TEST_ASSERT(mode < NUM_VM_MODES,
 515                                     "Guest mode ID %d too big", mode);
 516                         vm_guest_mode_params[mode].enabled = true;
 517                         break;
 518                 case 'h':
 519                 default:
 520                         help(argv[0]);
 521                         break;
 522                 }
 523         }
 524
 525         TEST_ASSERT(iterations > 2, "Iterations must be greater than two");
 526         TEST_ASSERT(interval > 0, "Interval must be greater than zero");
 527
 528         DEBUG("Test iterations: %"PRIu64", interval: %"PRIu64" (ms)\n",
 529               iterations, interval);
 530
 531         srandom(time(0));
 532
 533         for (i = 0; i < NUM_VM_MODES; ++i) {
 534                 if (!vm_guest_mode_params[i].enabled)
 535                         continue;
 536                 TEST_ASSERT(vm_guest_mode_params[i].supported,
 537                             "Guest mode ID %d (%s) not supported.",
 538                             i, vm_guest_mode_string(i));
 539                 run_test(i, iterations, interval, phys_offset);
 540         }
 541
 542         return 0;
 543 }