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 #define DIRTY_MEM_BITS 30 /* 1G */
 267 #define PAGE_SHIFT_4K  12
 268
 269 static void run_test(enum vm_guest_mode mode, unsigned long iterations,
 270                      unsigned long interval, uint64_t phys_offset)
 271 {
 272         unsigned int guest_pa_bits, guest_page_shift;
 273         pthread_t vcpu_thread;
 274         struct kvm_vm *vm;
 275         uint64_t max_gfn;
 276         unsigned long *bmap;
 277
 278         /*
 279          * We reserve page table for 2 times of extra dirty mem which
 280          * will definitely cover the original (1G+) test range.  Here
 281          * we do the calculation with 4K page size which is the
 282          * smallest so the page number will be enough for all archs
 283          * (e.g., 64K page size guest will need even less memory for
 284          * page tables).
 285          */
 286         vm = create_vm(mode, VCPU_ID,
 287                        2ul << (DIRTY_MEM_BITS - PAGE_SHIFT_4K),
 288                        guest_code);
 289
 290         switch (mode) {
 291         case VM_MODE_P52V48_4K:
 292                 guest_pa_bits = 52;
 293                 guest_page_shift = 12;
 294                 break;
 295         case VM_MODE_P52V48_64K:
 296                 guest_pa_bits = 52;
 297                 guest_page_shift = 16;
 298                 break;
 299         case VM_MODE_P48V48_4K:
 300                 guest_pa_bits = 48;
 301                 guest_page_shift = 12;
 302                 break;
 303         case VM_MODE_P48V48_64K:
 304                 guest_pa_bits = 48;
 305                 guest_page_shift = 16;
 306                 break;
 307         case VM_MODE_P40V48_4K:
 308                 guest_pa_bits = 40;
 309                 guest_page_shift = 12;
 310                 break;
 311         case VM_MODE_P40V48_64K:
 312                 guest_pa_bits = 40;
 313                 guest_page_shift = 16;
 314                 break;
 315         default:
 316                 TEST_ASSERT(false, "Unknown guest mode, mode: 0x%x", mode);
 317         }
 318
 319         DEBUG("Testing guest mode: %s\n", vm_guest_mode_string(mode));
 320
 321 #ifdef __x86_64__
 322         /*
 323          * FIXME
 324          * The x86_64 kvm selftests framework currently only supports a
 325          * single PML4 which restricts the number of physical address
 326          * bits we can change to 39.
 327          */
 328         guest_pa_bits = 39;
 329 #endif
 330         max_gfn = (1ul << (guest_pa_bits - guest_page_shift)) - 1;
 331         guest_page_size = (1ul << guest_page_shift);
 332         /*
 333          * A little more than 1G of guest page sized pages.  Cover the
 334          * case where the size is not aligned to 64 pages.
 335          */
 336         guest_num_pages = (1ul << (DIRTY_MEM_BITS - guest_page_shift)) + 16;
 337 #ifdef __s390x__
 338         /* Round up to multiple of 1M (segment size) */
 339         guest_num_pages = (guest_num_pages + 0xff) & ~0xffUL;
 340 #endif
 341         host_page_size = getpagesize();
 342         host_num_pages = (guest_num_pages * guest_page_size) / host_page_size +
 343                          !!((guest_num_pages * guest_page_size) % host_page_size);
 344
 345         if (!phys_offset) {
 346                 guest_test_phys_mem = (max_gfn - guest_num_pages) * guest_page_size;
 347                 guest_test_phys_mem &= ~(host_page_size - 1);
 348         } else {
 349                 guest_test_phys_mem = phys_offset;
 350         }
 351
 352 #ifdef __s390x__
 353         /* Align to 1M (segment size) */
 354         guest_test_phys_mem &= ~((1 << 20) - 1);
 355 #endif
 356
 357         DEBUG("guest physical test memory offset: 0x%lx\n", guest_test_phys_mem);
 358
 359         bmap = bitmap_alloc(host_num_pages);
 360         host_bmap_track = bitmap_alloc(host_num_pages);
 361
 362 #ifdef USE_CLEAR_DIRTY_LOG
 363         struct kvm_enable_cap cap = {};
 364
 365         cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
 366         cap.args[0] = 1;
 367         vm_enable_cap(vm, &cap);
 368 #endif
 369
 370         /* Add an extra memory slot for testing dirty logging */
 371         vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS,
 372                                     guest_test_phys_mem,
 373                                     TEST_MEM_SLOT_INDEX,
 374                                     guest_num_pages,
 375                                     KVM_MEM_LOG_DIRTY_PAGES);
 376
 377         /* Do mapping for the dirty track memory slot */
 378         virt_map(vm, guest_test_virt_mem, guest_test_phys_mem,
 379                  guest_num_pages * guest_page_size, 0);
 380
 381         /* Cache the HVA pointer of the region */
 382         host_test_mem = addr_gpa2hva(vm, (vm_paddr_t)guest_test_phys_mem);
 383
 384 #ifdef __x86_64__
 385         vcpu_set_cpuid(vm, VCPU_ID, kvm_get_supported_cpuid());
 386 #endif
 387 #ifdef __aarch64__
 388         ucall_init(vm, NULL);
 389 #endif
 390
 391         /* Export the shared variables to the guest */
 392         sync_global_to_guest(vm, host_page_size);
 393         sync_global_to_guest(vm, guest_page_size);
 394         sync_global_to_guest(vm, guest_test_virt_mem);
 395         sync_global_to_guest(vm, guest_num_pages);
 396
 397         /* Start the iterations */
 398         iteration = 1;
 399         sync_global_to_guest(vm, iteration);
 400         host_quit = false;
 401         host_dirty_count = 0;
 402         host_clear_count = 0;
 403         host_track_next_count = 0;
 404
 405         pthread_create(&vcpu_thread, NULL, vcpu_worker, vm);
 406
 407         while (iteration < iterations) {
 408                 /* Give the vcpu thread some time to dirty some pages */
 409                 usleep(interval * 1000);
 410                 kvm_vm_get_dirty_log(vm, TEST_MEM_SLOT_INDEX, bmap);
 411 #ifdef USE_CLEAR_DIRTY_LOG
 412                 kvm_vm_clear_dirty_log(vm, TEST_MEM_SLOT_INDEX, bmap, 0,
 413                                        host_num_pages);
 414 #endif
 415                 vm_dirty_log_verify(bmap);
 416                 iteration++;
 417                 sync_global_to_guest(vm, iteration);
 418         }
 419
 420         /* Tell the vcpu thread to quit */
 421         host_quit = true;
 422         pthread_join(vcpu_thread, NULL);
 423
 424         DEBUG("Total bits checked: dirty (%"PRIu64"), clear (%"PRIu64"), "
 425               "track_next (%"PRIu64")\n", host_dirty_count, host_clear_count,
 426               host_track_next_count);
 427
 428         free(bmap);
 429         free(host_bmap_track);
 430         ucall_uninit(vm);
 431         kvm_vm_free(vm);
 432 }
 433
 434 struct vm_guest_mode_params {
 435         bool supported;
 436         bool enabled;
 437 };
 438 struct vm_guest_mode_params vm_guest_mode_params[NUM_VM_MODES];
 439
 440 #define vm_guest_mode_params_init(mode, supported, enabled)                                     \
 441 ({                                                                                              \
 442         vm_guest_mode_params[mode] = (struct vm_guest_mode_params){ supported, enabled };       \
 443 })
 444
 445 static void help(char *name)
 446 {
 447         int i;
 448
 449         puts("");
 450         printf("usage: %s [-h] [-i iterations] [-I interval] "
 451                "[-p offset] [-m mode]\n", name);
 452         puts("");
 453         printf(" -i: specify iteration counts (default: %"PRIu64")\n",
 454                TEST_HOST_LOOP_N);
 455         printf(" -I: specify interval in ms (default: %"PRIu64" ms)\n",
 456                TEST_HOST_LOOP_INTERVAL);
 457         printf(" -p: specify guest physical test memory offset\n"
 458                "     Warning: a low offset can conflict with the loaded test code.\n");
 459         printf(" -m: specify the guest mode ID to test "
 460                "(default: test all supported modes)\n"
 461                "     This option may be used multiple times.\n"
 462                "     Guest mode IDs:\n");
 463         for (i = 0; i < NUM_VM_MODES; ++i) {
 464                 printf("         %d:    %s%s\n", i, vm_guest_mode_string(i),
 465                        vm_guest_mode_params[i].supported ? " (supported)" : "");
 466         }
 467         puts("");
 468         exit(0);
 469 }
 470
 471 int main(int argc, char *argv[])
 472 {
 473         unsigned long iterations = TEST_HOST_LOOP_N;
 474         unsigned long interval = TEST_HOST_LOOP_INTERVAL;
 475         bool mode_selected = false;
 476         uint64_t phys_offset = 0;
 477         unsigned int mode;
 478         int opt, i;
 479 #ifdef __aarch64__
 480         unsigned int host_ipa_limit;
 481 #endif
 482
 483 #ifdef USE_CLEAR_DIRTY_LOG
 484         if (!kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2)) {
 485                 fprintf(stderr, "KVM_CLEAR_DIRTY_LOG not available, skipping tests\n");
 486                 exit(KSFT_SKIP);
 487         }
 488 #endif
 489
 490 #ifdef __x86_64__
 491         vm_guest_mode_params_init(VM_MODE_P52V48_4K, true, true);
 492 #endif
 493 #ifdef __aarch64__
 494         vm_guest_mode_params_init(VM_MODE_P40V48_4K, true, true);
 495         vm_guest_mode_params_init(VM_MODE_P40V48_64K, true, true);
 496
 497         host_ipa_limit = kvm_check_cap(KVM_CAP_ARM_VM_IPA_SIZE);
 498         if (host_ipa_limit >= 52)
 499                 vm_guest_mode_params_init(VM_MODE_P52V48_64K, true, true);
 500         if (host_ipa_limit >= 48) {
 501                 vm_guest_mode_params_init(VM_MODE_P48V48_4K, true, true);
 502                 vm_guest_mode_params_init(VM_MODE_P48V48_64K, true, true);
 503         }
 504 #endif
 505 #ifdef __s390x__
 506         vm_guest_mode_params_init(VM_MODE_P40V48_4K, true, true);
 507 #endif
 508
 509         while ((opt = getopt(argc, argv, "hi:I:p:m:")) != -1) {
 510                 switch (opt) {
 511                 case 'i':
 512                         iterations = strtol(optarg, NULL, 10);
 513                         break;
 514                 case 'I':
 515                         interval = strtol(optarg, NULL, 10);
 516                         break;
 517                 case 'p':
 518                         phys_offset = strtoull(optarg, NULL, 0);
 519                         break;
 520                 case 'm':
 521                         if (!mode_selected) {
 522                                 for (i = 0; i < NUM_VM_MODES; ++i)
 523                                         vm_guest_mode_params[i].enabled = false;
 524                                 mode_selected = true;
 525                         }
 526                         mode = strtoul(optarg, NULL, 10);
 527                         TEST_ASSERT(mode < NUM_VM_MODES,
 528                                     "Guest mode ID %d too big", mode);
 529                         vm_guest_mode_params[mode].enabled = true;
 530                         break;
 531                 case 'h':
 532                 default:
 533                         help(argv[0]);
 534                         break;
 535                 }
 536         }
 537
 538         TEST_ASSERT(iterations > 2, "Iterations must be greater than two");
 539         TEST_ASSERT(interval > 0, "Interval must be greater than zero");
 540
 541         DEBUG("Test iterations: %"PRIu64", interval: %"PRIu64" (ms)\n",
 542               iterations, interval);
 543
 544         srandom(time(0));
 545
 546         for (i = 0; i < NUM_VM_MODES; ++i) {
 547                 if (!vm_guest_mode_params[i].enabled)
 548                         continue;
 549                 TEST_ASSERT(vm_guest_mode_params[i].supported,
 550                             "Guest mode ID %d (%s) not supported.",
 551                             i, vm_guest_mode_string(i));
 552                 run_test(i, iterations, interval, phys_offset);
 553         }
 554
 555         return 0;
 556 }