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 VCPU_ID                         1
  23
  24 /* The memory slot index to track dirty pages */
  25 #define TEST_MEM_SLOT_INDEX             1
  26
  27 /* Default guest test virtual memory offset */
  28 #define DEFAULT_GUEST_TEST_MEM          0xc0000000
  29
  30 /* How many pages to dirty for each guest loop */
  31 #define TEST_PAGES_PER_LOOP             1024
  32
  33 /* How many host loops to run (one KVM_GET_DIRTY_LOG for each loop) */
  34 #define TEST_HOST_LOOP_N                32UL
  35
  36 /* Interval for each host loop (ms) */
  37 #define TEST_HOST_LOOP_INTERVAL         10UL
  38
  39 /* Dirty bitmaps are always little endian, so we need to swap on big endian */
  40 #if defined(__s390x__)
  41 # define BITOP_LE_SWIZZLE       ((BITS_PER_LONG-1) & ~0x7)
  42 # define test_bit_le(nr, addr) \
  43         test_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  44 # define set_bit_le(nr, addr) \
  45         set_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  46 # define clear_bit_le(nr, addr) \
  47         clear_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  48 # define test_and_set_bit_le(nr, addr) \
  49         test_and_set_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  50 # define test_and_clear_bit_le(nr, addr) \
  51         test_and_clear_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  52 #else
  53 # define test_bit_le            test_bit
  54 # define set_bit_le             set_bit
  55 # define clear_bit_le           clear_bit
  56 # define test_and_set_bit_le    test_and_set_bit
  57 # define test_and_clear_bit_le  test_and_clear_bit
  58 #endif
  59
  60 /*
  61  * Guest/Host shared variables. Ensure addr_gva2hva() and/or
  62  * sync_global_to/from_guest() are used when accessing from
  63  * the host. READ/WRITE_ONCE() should also be used with anything
  64  * that may change.
  65  */
  66 static uint64_t host_page_size;
  67 static uint64_t guest_page_size;
  68 static uint64_t guest_num_pages;
  69 static uint64_t random_array[TEST_PAGES_PER_LOOP];
  70 static uint64_t iteration;
  71
  72 /*
  73  * Guest physical memory offset of the testing memory slot.
  74  * This will be set to the topmost valid physical address minus
  75  * the test memory size.
  76  */
  77 static uint64_t guest_test_phys_mem;
  78
  79 /*
  80  * Guest virtual memory offset of the testing memory slot.
  81  * Must not conflict with identity mapped test code.
  82  */
  83 static uint64_t guest_test_virt_mem = DEFAULT_GUEST_TEST_MEM;
  84
  85 /*
  86  * Continuously write to the first 8 bytes of a random pages within
  87  * the testing memory region.
  88  */
  89 static void guest_code(void)
  90 {
  91         uint64_t addr;
  92         int i;
  93
  94         /*
  95          * On s390x, all pages of a 1M segment are initially marked as dirty
  96          * when a page of the segment is written to for the very first time.
  97          * To compensate this specialty in this test, we need to touch all
  98          * pages during the first iteration.
  99          */
 100         for (i = 0; i < guest_num_pages; i++) {
 101                 addr = guest_test_virt_mem + i * guest_page_size;
 102                 *(uint64_t *)addr = READ_ONCE(iteration);
 103         }
 104
 105         while (true) {
 106                 for (i = 0; i < TEST_PAGES_PER_LOOP; i++) {
 107                         addr = guest_test_virt_mem;
 108                         addr += (READ_ONCE(random_array[i]) % guest_num_pages)
 109                                 * guest_page_size;
 110                         addr &= ~(host_page_size - 1);
 111                         *(uint64_t *)addr = READ_ONCE(iteration);
 112                 }
 113
 114                 /* Tell the host that we need more random numbers */
 115                 GUEST_SYNC(1);
 116         }
 117 }
 118
 119 /* Host variables */
 120 static bool host_quit;
 121
 122 /* Points to the test VM memory region on which we track dirty logs */
 123 static void *host_test_mem;
 124 static uint64_t host_num_pages;
 125
 126 /* For statistics only */
 127 static uint64_t host_dirty_count;
 128 static uint64_t host_clear_count;
 129 static uint64_t host_track_next_count;
 130
 131 /*
 132  * We use this bitmap to track some pages that should have its dirty
 133  * bit set in the _next_ iteration.  For example, if we detected the
 134  * page value changed to current iteration but at the same time the
 135  * page bit is cleared in the latest bitmap, then the system must
 136  * report that write in the next get dirty log call.
 137  */
 138 static unsigned long *host_bmap_track;
 139
 140 static void generate_random_array(uint64_t *guest_array, uint64_t size)
 141 {
 142         uint64_t i;
 143
 144         for (i = 0; i < size; i++)
 145                 guest_array[i] = random();
 146 }
 147
 148 static void *vcpu_worker(void *data)
 149 {
 150         int ret;
 151         struct kvm_vm *vm = data;
 152         uint64_t *guest_array;
 153         uint64_t pages_count = 0;
 154         struct kvm_run *run;
 155
 156         run = vcpu_state(vm, VCPU_ID);
 157
 158         guest_array = addr_gva2hva(vm, (vm_vaddr_t)random_array);
 159         generate_random_array(guest_array, TEST_PAGES_PER_LOOP);
 160
 161         while (!READ_ONCE(host_quit)) {
 162                 /* Let the guest dirty the random pages */
 163                 ret = _vcpu_run(vm, VCPU_ID);
 164                 TEST_ASSERT(ret == 0, "vcpu_run failed: %d\n", ret);
 165                 if (get_ucall(vm, VCPU_ID, NULL) == UCALL_SYNC) {
 166                         pages_count += TEST_PAGES_PER_LOOP;
 167                         generate_random_array(guest_array, TEST_PAGES_PER_LOOP);
 168                 } else {
 169                         TEST_ASSERT(false,
 170                                     "Invalid guest sync status: "
 171                                     "exit_reason=%s\n",
 172                                     exit_reason_str(run->exit_reason));
 173                 }
 174         }
 175
 176         pr_info("Dirtied %"PRIu64" pages\n", pages_count);
 177
 178         return NULL;
 179 }
 180
 181 static void vm_dirty_log_verify(enum vm_guest_mode mode, unsigned long *bmap)
 182 {
 183         uint64_t step = vm_num_host_pages(mode, 1);
 184         uint64_t page;
 185         uint64_t *value_ptr;
 186
 187         for (page = 0; page < host_num_pages; page += step) {
 188                 value_ptr = host_test_mem + page * host_page_size;
 189
 190                 /* If this is a special page that we were tracking... */
 191                 if (test_and_clear_bit_le(page, host_bmap_track)) {
 192                         host_track_next_count++;
 193                         TEST_ASSERT(test_bit_le(page, bmap),
 194                                     "Page %"PRIu64" should have its dirty bit "
 195                                     "set in this iteration but it is missing",
 196                                     page);
 197                 }
 198
 199                 if (test_bit_le(page, bmap)) {
 200                         host_dirty_count++;
 201                         /*
 202                          * If the bit is set, the value written onto
 203                          * the corresponding page should be either the
 204                          * previous iteration number or the current one.
 205                          */
 206                         TEST_ASSERT(*value_ptr == iteration ||
 207                                     *value_ptr == iteration - 1,
 208                                     "Set page %"PRIu64" value %"PRIu64
 209                                     " incorrect (iteration=%"PRIu64")",
 210                                     page, *value_ptr, iteration);
 211                 } else {
 212                         host_clear_count++;
 213                         /*
 214                          * If cleared, the value written can be any
 215                          * value smaller or equals to the iteration
 216                          * number.  Note that the value can be exactly
 217                          * (iteration-1) if that write can happen
 218                          * like this:
 219                          *
 220                          * (1) increase loop count to "iteration-1"
 221                          * (2) write to page P happens (with value
 222                          *     "iteration-1")
 223                          * (3) get dirty log for "iteration-1"; we'll
 224                          *     see that page P bit is set (dirtied),
 225                          *     and not set the bit in host_bmap_track
 226                          * (4) increase loop count to "iteration"
 227                          *     (which is current iteration)
 228                          * (5) get dirty log for current iteration,
 229                          *     we'll see that page P is cleared, with
 230                          *     value "iteration-1".
 231                          */
 232                         TEST_ASSERT(*value_ptr <= iteration,
 233                                     "Clear page %"PRIu64" value %"PRIu64
 234                                     " incorrect (iteration=%"PRIu64")",
 235                                     page, *value_ptr, iteration);
 236                         if (*value_ptr == iteration) {
 237                                 /*
 238                                  * This page is _just_ modified; it
 239                                  * should report its dirtyness in the
 240                                  * next run
 241                                  */
 242                                 set_bit_le(page, host_bmap_track);
 243                         }
 244                 }
 245         }
 246 }
 247
 248 static struct kvm_vm *create_vm(enum vm_guest_mode mode, uint32_t vcpuid,
 249                                 uint64_t extra_mem_pages, void *guest_code)
 250 {
 251         struct kvm_vm *vm;
 252         uint64_t extra_pg_pages = extra_mem_pages / 512 * 2;
 253
 254         pr_info("Testing guest mode: %s\n", vm_guest_mode_string(mode));
 255
 256         vm = _vm_create(mode, DEFAULT_GUEST_PHY_PAGES + extra_pg_pages, O_RDWR);
 257         kvm_vm_elf_load(vm, program_invocation_name, 0, 0);
 258 #ifdef __x86_64__
 259         vm_create_irqchip(vm);
 260 #endif
 261         vm_vcpu_add_default(vm, vcpuid, guest_code);
 262         return vm;
 263 }
 264
 265 #define DIRTY_MEM_BITS 30 /* 1G */
 266 #define PAGE_SHIFT_4K  12
 267
 268 #ifdef USE_CLEAR_DIRTY_LOG
 269 static u64 dirty_log_manual_caps;
 270 #endif
 271
 272 static void run_test(enum vm_guest_mode mode, unsigned long iterations,
 273                      unsigned long interval, uint64_t phys_offset)
 274 {
 275         pthread_t vcpu_thread;
 276         struct kvm_vm *vm;
 277         unsigned long *bmap;
 278
 279         /*
 280          * We reserve page table for 2 times of extra dirty mem which
 281          * will definitely cover the original (1G+) test range.  Here
 282          * we do the calculation with 4K page size which is the
 283          * smallest so the page number will be enough for all archs
 284          * (e.g., 64K page size guest will need even less memory for
 285          * page tables).
 286          */
 287         vm = create_vm(mode, VCPU_ID,
 288                        2ul << (DIRTY_MEM_BITS - PAGE_SHIFT_4K),
 289                        guest_code);
 290
 291         guest_page_size = vm_get_page_size(vm);
 292         /*
 293          * A little more than 1G of guest page sized pages.  Cover the
 294          * case where the size is not aligned to 64 pages.
 295          */
 296         guest_num_pages = (1ul << (DIRTY_MEM_BITS -
 297                                    vm_get_page_shift(vm))) + 3;
 298         guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
 299
 300         host_page_size = getpagesize();
 301         host_num_pages = vm_num_host_pages(mode, guest_num_pages);
 302
 303         if (!phys_offset) {
 304                 guest_test_phys_mem = (vm_get_max_gfn(vm) -
 305                                        guest_num_pages) * guest_page_size;
 306                 guest_test_phys_mem &= ~(host_page_size - 1);
 307         } else {
 308                 guest_test_phys_mem = phys_offset;
 309         }
 310
 311 #ifdef __s390x__
 312         /* Align to 1M (segment size) */
 313         guest_test_phys_mem &= ~((1 << 20) - 1);
 314 #endif
 315
 316         pr_info("guest physical test memory offset: 0x%lx\n", guest_test_phys_mem);
 317
 318         bmap = bitmap_alloc(host_num_pages);
 319         host_bmap_track = bitmap_alloc(host_num_pages);
 320
 321 #ifdef USE_CLEAR_DIRTY_LOG
 322         struct kvm_enable_cap cap = {};
 323
 324         cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
 325         cap.args[0] = dirty_log_manual_caps;
 326         vm_enable_cap(vm, &cap);
 327 #endif
 328
 329         /* Add an extra memory slot for testing dirty logging */
 330         vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS,
 331                                     guest_test_phys_mem,
 332                                     TEST_MEM_SLOT_INDEX,
 333                                     guest_num_pages,
 334                                     KVM_MEM_LOG_DIRTY_PAGES);
 335
 336         /* Do mapping for the dirty track memory slot */
 337         virt_map(vm, guest_test_virt_mem, guest_test_phys_mem,
 338                  guest_num_pages * guest_page_size, 0);
 339
 340         /* Cache the HVA pointer of the region */
 341         host_test_mem = addr_gpa2hva(vm, (vm_paddr_t)guest_test_phys_mem);
 342
 343 #ifdef __x86_64__
 344         vcpu_set_cpuid(vm, VCPU_ID, kvm_get_supported_cpuid());
 345 #endif
 346         ucall_init(vm, NULL);
 347
 348         /* Export the shared variables to the guest */
 349         sync_global_to_guest(vm, host_page_size);
 350         sync_global_to_guest(vm, guest_page_size);
 351         sync_global_to_guest(vm, guest_test_virt_mem);
 352         sync_global_to_guest(vm, guest_num_pages);
 353
 354         /* Start the iterations */
 355         iteration = 1;
 356         sync_global_to_guest(vm, iteration);
 357         host_quit = false;
 358         host_dirty_count = 0;
 359         host_clear_count = 0;
 360         host_track_next_count = 0;
 361
 362         pthread_create(&vcpu_thread, NULL, vcpu_worker, vm);
 363
 364         while (iteration < iterations) {
 365                 /* Give the vcpu thread some time to dirty some pages */
 366                 usleep(interval * 1000);
 367                 kvm_vm_get_dirty_log(vm, TEST_MEM_SLOT_INDEX, bmap);
 368 #ifdef USE_CLEAR_DIRTY_LOG
 369                 kvm_vm_clear_dirty_log(vm, TEST_MEM_SLOT_INDEX, bmap, 0,
 370                                        host_num_pages);
 371 #endif
 372                 vm_dirty_log_verify(mode, bmap);
 373                 iteration++;
 374                 sync_global_to_guest(vm, iteration);
 375         }
 376
 377         /* Tell the vcpu thread to quit */
 378         host_quit = true;
 379         pthread_join(vcpu_thread, NULL);
 380
 381         pr_info("Total bits checked: dirty (%"PRIu64"), clear (%"PRIu64"), "
 382                 "track_next (%"PRIu64")\n", host_dirty_count, host_clear_count,
 383                 host_track_next_count);
 384
 385         free(bmap);
 386         free(host_bmap_track);
 387         ucall_uninit(vm);
 388         kvm_vm_free(vm);
 389 }
 390
 391 struct guest_mode {
 392         bool supported;
 393         bool enabled;
 394 };
 395 static struct guest_mode guest_modes[NUM_VM_MODES];
 396
 397 #define guest_mode_init(mode, supported, enabled) ({ \
 398         guest_modes[mode] = (struct guest_mode){ supported, enabled }; \
 399 })
 400
 401 static void help(char *name)
 402 {
 403         int i;
 404
 405         puts("");
 406         printf("usage: %s [-h] [-i iterations] [-I interval] "
 407                "[-p offset] [-m mode]\n", name);
 408         puts("");
 409         printf(" -i: specify iteration counts (default: %"PRIu64")\n",
 410                TEST_HOST_LOOP_N);
 411         printf(" -I: specify interval in ms (default: %"PRIu64" ms)\n",
 412                TEST_HOST_LOOP_INTERVAL);
 413         printf(" -p: specify guest physical test memory offset\n"
 414                "     Warning: a low offset can conflict with the loaded test code.\n");
 415         printf(" -m: specify the guest mode ID to test "
 416                "(default: test all supported modes)\n"
 417                "     This option may be used multiple times.\n"
 418                "     Guest mode IDs:\n");
 419         for (i = 0; i < NUM_VM_MODES; ++i) {
 420                 printf("         %d:    %s%s\n", i, vm_guest_mode_string(i),
 421                        guest_modes[i].supported ? " (supported)" : "");
 422         }
 423         puts("");
 424         exit(0);
 425 }
 426
 427 int main(int argc, char *argv[])
 428 {
 429         unsigned long iterations = TEST_HOST_LOOP_N;
 430         unsigned long interval = TEST_HOST_LOOP_INTERVAL;
 431         bool mode_selected = false;
 432         uint64_t phys_offset = 0;
 433         unsigned int mode;
 434         int opt, i;
 435
 436 #ifdef USE_CLEAR_DIRTY_LOG
 437         dirty_log_manual_caps =
 438                 kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
 439         if (!dirty_log_manual_caps) {
 440                 print_skip("KVM_CLEAR_DIRTY_LOG not available");
 441                 exit(KSFT_SKIP);
 442         }
 443         dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
 444                                   KVM_DIRTY_LOG_INITIALLY_SET);
 445 #endif
 446
 447 #ifdef __x86_64__
 448         guest_mode_init(VM_MODE_PXXV48_4K, true, true);
 449 #endif
 450 #ifdef __aarch64__
 451         guest_mode_init(VM_MODE_P40V48_4K, true, true);
 452         guest_mode_init(VM_MODE_P40V48_64K, true, true);
 453
 454         {
 455                 unsigned int limit = kvm_check_cap(KVM_CAP_ARM_VM_IPA_SIZE);
 456
 457                 if (limit >= 52)
 458                         guest_mode_init(VM_MODE_P52V48_64K, true, true);
 459                 if (limit >= 48) {
 460                         guest_mode_init(VM_MODE_P48V48_4K, true, true);
 461                         guest_mode_init(VM_MODE_P48V48_64K, true, true);
 462                 }
 463         }
 464 #endif
 465 #ifdef __s390x__
 466         guest_mode_init(VM_MODE_P40V48_4K, true, true);
 467 #endif
 468
 469         while ((opt = getopt(argc, argv, "hi:I:p:m:")) != -1) {
 470                 switch (opt) {
 471                 case 'i':
 472                         iterations = strtol(optarg, NULL, 10);
 473                         break;
 474                 case 'I':
 475                         interval = strtol(optarg, NULL, 10);
 476                         break;
 477                 case 'p':
 478                         phys_offset = strtoull(optarg, NULL, 0);
 479                         break;
 480                 case 'm':
 481                         if (!mode_selected) {
 482                                 for (i = 0; i < NUM_VM_MODES; ++i)
 483                                         guest_modes[i].enabled = false;
 484                                 mode_selected = true;
 485                         }
 486                         mode = strtoul(optarg, NULL, 10);
 487                         TEST_ASSERT(mode < NUM_VM_MODES,
 488                                     "Guest mode ID %d too big", mode);
 489                         guest_modes[mode].enabled = true;
 490                         break;
 491                 case 'h':
 492                 default:
 493                         help(argv[0]);
 494                         break;
 495                 }
 496         }
 497
 498         TEST_ASSERT(iterations > 2, "Iterations must be greater than two");
 499         TEST_ASSERT(interval > 0, "Interval must be greater than zero");
 500
 501         pr_info("Test iterations: %"PRIu64", interval: %"PRIu64" (ms)\n",
 502                 iterations, interval);
 503
 504         srandom(time(0));
 505
 506         for (i = 0; i < NUM_VM_MODES; ++i) {
 507                 if (!guest_modes[i].enabled)
 508                         continue;
 509                 TEST_ASSERT(guest_modes[i].supported,
 510                             "Guest mode ID %d (%s) not supported.",
 511                             i, vm_guest_mode_string(i));
 512                 run_test(i, iterations, interval, phys_offset);
 513         }
 514
 515         return 0;
 516 }