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 <pthread.h>
  13 #include <semaphore.h>
  14 #include <sys/types.h>
  15 #include <signal.h>
  16 #include <errno.h>
  17 #include <linux/bitmap.h>
  18 #include <linux/bitops.h>
  19 #include <linux/atomic.h>
  20
  21 #include "kvm_util.h"
  22 #include "test_util.h"
  23 #include "guest_modes.h"
  24 #include "processor.h"
  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 #define TEST_DIRTY_RING_COUNT           65536
  63
  64 #define SIG_IPI SIGUSR1
  65
  66 /*
  67  * Guest/Host shared variables. Ensure addr_gva2hva() and/or
  68  * sync_global_to/from_guest() are used when accessing from
  69  * the host. READ/WRITE_ONCE() should also be used with anything
  70  * that may change.
  71  */
  72 static uint64_t host_page_size;
  73 static uint64_t guest_page_size;
  74 static uint64_t guest_num_pages;
  75 static uint64_t random_array[TEST_PAGES_PER_LOOP];
  76 static uint64_t iteration;
  77
  78 /*
  79  * Guest physical memory offset of the testing memory slot.
  80  * This will be set to the topmost valid physical address minus
  81  * the test memory size.
  82  */
  83 static uint64_t guest_test_phys_mem;
  84
  85 /*
  86  * Guest virtual memory offset of the testing memory slot.
  87  * Must not conflict with identity mapped test code.
  88  */
  89 static uint64_t guest_test_virt_mem = DEFAULT_GUEST_TEST_MEM;
  90
  91 /*
  92  * Continuously write to the first 8 bytes of a random pages within
  93  * the testing memory region.
  94  */
  95 static void guest_code(void)
  96 {
  97         uint64_t addr;
  98         int i;
  99
 100         /*
 101          * On s390x, all pages of a 1M segment are initially marked as dirty
 102          * when a page of the segment is written to for the very first time.
 103          * To compensate this specialty in this test, we need to touch all
 104          * pages during the first iteration.
 105          */
 106         for (i = 0; i < guest_num_pages; i++) {
 107                 addr = guest_test_virt_mem + i * guest_page_size;
 108                 *(uint64_t *)addr = READ_ONCE(iteration);
 109         }
 110
 111         while (true) {
 112                 for (i = 0; i < TEST_PAGES_PER_LOOP; i++) {
 113                         addr = guest_test_virt_mem;
 114                         addr += (READ_ONCE(random_array[i]) % guest_num_pages)
 115                                 * guest_page_size;
 116                         addr = align_down(addr, host_page_size);
 117                         *(uint64_t *)addr = READ_ONCE(iteration);
 118                 }
 119
 120                 /* Tell the host that we need more random numbers */
 121                 GUEST_SYNC(1);
 122         }
 123 }
 124
 125 /* Host variables */
 126 static bool host_quit;
 127
 128 /* Points to the test VM memory region on which we track dirty logs */
 129 static void *host_test_mem;
 130 static uint64_t host_num_pages;
 131
 132 /* For statistics only */
 133 static uint64_t host_dirty_count;
 134 static uint64_t host_clear_count;
 135 static uint64_t host_track_next_count;
 136
 137 /* Whether dirty ring reset is requested, or finished */
 138 static sem_t sem_vcpu_stop;
 139 static sem_t sem_vcpu_cont;
 140 /*
 141  * This is only set by main thread, and only cleared by vcpu thread.  It is
 142  * used to request vcpu thread to stop at the next GUEST_SYNC, since GUEST_SYNC
 143  * is the only place that we'll guarantee both "dirty bit" and "dirty data"
 144  * will match.  E.g., SIG_IPI won't guarantee that if the vcpu is interrupted
 145  * after setting dirty bit but before the data is written.
 146  */
 147 static atomic_t vcpu_sync_stop_requested;
 148 /*
 149  * This is updated by the vcpu thread to tell the host whether it's a
 150  * ring-full event.  It should only be read until a sem_wait() of
 151  * sem_vcpu_stop and before vcpu continues to run.
 152  */
 153 static bool dirty_ring_vcpu_ring_full;
 154 /*
 155  * This is only used for verifying the dirty pages.  Dirty ring has a very
 156  * tricky case when the ring just got full, kvm will do userspace exit due to
 157  * ring full.  When that happens, the very last PFN is set but actually the
 158  * data is not changed (the guest WRITE is not really applied yet), because
 159  * we found that the dirty ring is full, refused to continue the vcpu, and
 160  * recorded the dirty gfn with the old contents.
 161  *
 162  * For this specific case, it's safe to skip checking this pfn for this
 163  * bit, because it's a redundant bit, and when the write happens later the bit
 164  * will be set again.  We use this variable to always keep track of the latest
 165  * dirty gfn we've collected, so that if a mismatch of data found later in the
 166  * verifying process, we let it pass.
 167  */
 168 static uint64_t dirty_ring_last_page;
 169
 170 enum log_mode_t {
 171         /* Only use KVM_GET_DIRTY_LOG for logging */
 172         LOG_MODE_DIRTY_LOG = 0,
 173
 174         /* Use both KVM_[GET|CLEAR]_DIRTY_LOG for logging */
 175         LOG_MODE_CLEAR_LOG = 1,
 176
 177         /* Use dirty ring for logging */
 178         LOG_MODE_DIRTY_RING = 2,
 179
 180         LOG_MODE_NUM,
 181
 182         /* Run all supported modes */
 183         LOG_MODE_ALL = LOG_MODE_NUM,
 184 };
 185
 186 /* Mode of logging to test.  Default is to run all supported modes */
 187 static enum log_mode_t host_log_mode_option = LOG_MODE_ALL;
 188 /* Logging mode for current run */
 189 static enum log_mode_t host_log_mode;
 190 static pthread_t vcpu_thread;
 191 static uint32_t test_dirty_ring_count = TEST_DIRTY_RING_COUNT;
 192
 193 static void vcpu_kick(void)
 194 {
 195         pthread_kill(vcpu_thread, SIG_IPI);
 196 }
 197
 198 /*
 199  * In our test we do signal tricks, let's use a better version of
 200  * sem_wait to avoid signal interrupts
 201  */
 202 static void sem_wait_until(sem_t *sem)
 203 {
 204         int ret;
 205
 206         do
 207                 ret = sem_wait(sem);
 208         while (ret == -1 && errno == EINTR);
 209 }
 210
 211 static bool clear_log_supported(void)
 212 {
 213         return kvm_has_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
 214 }
 215
 216 static void clear_log_create_vm_done(struct kvm_vm *vm)
 217 {
 218         u64 manual_caps;
 219
 220         manual_caps = kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
 221         TEST_ASSERT(manual_caps, "MANUAL_CAPS is zero!");
 222         manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
 223                         KVM_DIRTY_LOG_INITIALLY_SET);
 224         vm_enable_cap(vm, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, manual_caps);
 225 }
 226
 227 static void dirty_log_collect_dirty_pages(struct kvm_vcpu *vcpu, int slot,
 228                                           void *bitmap, uint32_t num_pages)
 229 {
 230         kvm_vm_get_dirty_log(vcpu->vm, slot, bitmap);
 231 }
 232
 233 static void clear_log_collect_dirty_pages(struct kvm_vcpu *vcpu, int slot,
 234                                           void *bitmap, uint32_t num_pages)
 235 {
 236         kvm_vm_get_dirty_log(vcpu->vm, slot, bitmap);
 237         kvm_vm_clear_dirty_log(vcpu->vm, slot, bitmap, 0, num_pages);
 238 }
 239
 240 /* Should only be called after a GUEST_SYNC */
 241 static void vcpu_handle_sync_stop(void)
 242 {
 243         if (atomic_read(&vcpu_sync_stop_requested)) {
 244                 /* It means main thread is sleeping waiting */
 245                 atomic_set(&vcpu_sync_stop_requested, false);
 246                 sem_post(&sem_vcpu_stop);
 247                 sem_wait_until(&sem_vcpu_cont);
 248         }
 249 }
 250
 251 static void default_after_vcpu_run(struct kvm_vcpu *vcpu, int ret, int err)
 252 {
 253         struct kvm_run *run = vcpu->run;
 254
 255         TEST_ASSERT(ret == 0 || (ret == -1 && err == EINTR),
 256                     "vcpu run failed: errno=%d", err);
 257
 258         TEST_ASSERT(get_ucall(vcpu, NULL) == UCALL_SYNC,
 259                     "Invalid guest sync status: exit_reason=%s\n",
 260                     exit_reason_str(run->exit_reason));
 261
 262         vcpu_handle_sync_stop();
 263 }
 264
 265 static bool dirty_ring_supported(void)
 266 {
 267         return kvm_has_cap(KVM_CAP_DIRTY_LOG_RING);
 268 }
 269
 270 static void dirty_ring_create_vm_done(struct kvm_vm *vm)
 271 {
 272         /*
 273          * Switch to dirty ring mode after VM creation but before any
 274          * of the vcpu creation.
 275          */
 276         vm_enable_dirty_ring(vm, test_dirty_ring_count *
 277                              sizeof(struct kvm_dirty_gfn));
 278 }
 279
 280 static inline bool dirty_gfn_is_dirtied(struct kvm_dirty_gfn *gfn)
 281 {
 282         return gfn->flags == KVM_DIRTY_GFN_F_DIRTY;
 283 }
 284
 285 static inline void dirty_gfn_set_collected(struct kvm_dirty_gfn *gfn)
 286 {
 287         gfn->flags = KVM_DIRTY_GFN_F_RESET;
 288 }
 289
 290 static uint32_t dirty_ring_collect_one(struct kvm_dirty_gfn *dirty_gfns,
 291                                        int slot, void *bitmap,
 292                                        uint32_t num_pages, uint32_t *fetch_index)
 293 {
 294         struct kvm_dirty_gfn *cur;
 295         uint32_t count = 0;
 296
 297         while (true) {
 298                 cur = &dirty_gfns[*fetch_index % test_dirty_ring_count];
 299                 if (!dirty_gfn_is_dirtied(cur))
 300                         break;
 301                 TEST_ASSERT(cur->slot == slot, "Slot number didn't match: "
 302                             "%u != %u", cur->slot, slot);
 303                 TEST_ASSERT(cur->offset < num_pages, "Offset overflow: "
 304                             "0x%llx >= 0x%x", cur->offset, num_pages);
 305                 //pr_info("fetch 0x%x page %llu\n", *fetch_index, cur->offset);
 306                 set_bit_le(cur->offset, bitmap);
 307                 dirty_ring_last_page = cur->offset;
 308                 dirty_gfn_set_collected(cur);
 309                 (*fetch_index)++;
 310                 count++;
 311         }
 312
 313         return count;
 314 }
 315
 316 static void dirty_ring_wait_vcpu(void)
 317 {
 318         /* This makes sure that hardware PML cache flushed */
 319         vcpu_kick();
 320         sem_wait_until(&sem_vcpu_stop);
 321 }
 322
 323 static void dirty_ring_continue_vcpu(void)
 324 {
 325         pr_info("Notifying vcpu to continue\n");
 326         sem_post(&sem_vcpu_cont);
 327 }
 328
 329 static void dirty_ring_collect_dirty_pages(struct kvm_vcpu *vcpu, int slot,
 330                                            void *bitmap, uint32_t num_pages)
 331 {
 332         /* We only have one vcpu */
 333         static uint32_t fetch_index = 0;
 334         uint32_t count = 0, cleared;
 335         bool continued_vcpu = false;
 336
 337         dirty_ring_wait_vcpu();
 338
 339         if (!dirty_ring_vcpu_ring_full) {
 340                 /*
 341                  * This is not a ring-full event, it's safe to allow
 342                  * vcpu to continue
 343                  */
 344                 dirty_ring_continue_vcpu();
 345                 continued_vcpu = true;
 346         }
 347
 348         /* Only have one vcpu */
 349         count = dirty_ring_collect_one(vcpu_map_dirty_ring(vcpu),
 350                                        slot, bitmap, num_pages, &fetch_index);
 351
 352         cleared = kvm_vm_reset_dirty_ring(vcpu->vm);
 353
 354         /* Cleared pages should be the same as collected */
 355         TEST_ASSERT(cleared == count, "Reset dirty pages (%u) mismatch "
 356                     "with collected (%u)", cleared, count);
 357
 358         if (!continued_vcpu) {
 359                 TEST_ASSERT(dirty_ring_vcpu_ring_full,
 360                             "Didn't continue vcpu even without ring full");
 361                 dirty_ring_continue_vcpu();
 362         }
 363
 364         pr_info("Iteration %ld collected %u pages\n", iteration, count);
 365 }
 366
 367 static void dirty_ring_after_vcpu_run(struct kvm_vcpu *vcpu, int ret, int err)
 368 {
 369         struct kvm_run *run = vcpu->run;
 370
 371         /* A ucall-sync or ring-full event is allowed */
 372         if (get_ucall(vcpu, NULL) == UCALL_SYNC) {
 373                 /* We should allow this to continue */
 374                 ;
 375         } else if (run->exit_reason == KVM_EXIT_DIRTY_RING_FULL ||
 376                    (ret == -1 && err == EINTR)) {
 377                 /* Update the flag first before pause */
 378                 WRITE_ONCE(dirty_ring_vcpu_ring_full,
 379                            run->exit_reason == KVM_EXIT_DIRTY_RING_FULL);
 380                 sem_post(&sem_vcpu_stop);
 381                 pr_info("vcpu stops because %s...\n",
 382                         dirty_ring_vcpu_ring_full ?
 383                         "dirty ring is full" : "vcpu is kicked out");
 384                 sem_wait_until(&sem_vcpu_cont);
 385                 pr_info("vcpu continues now.\n");
 386         } else {
 387                 TEST_ASSERT(false, "Invalid guest sync status: "
 388                             "exit_reason=%s\n",
 389                             exit_reason_str(run->exit_reason));
 390         }
 391 }
 392
 393 static void dirty_ring_before_vcpu_join(void)
 394 {
 395         /* Kick another round of vcpu just to make sure it will quit */
 396         sem_post(&sem_vcpu_cont);
 397 }
 398
 399 struct log_mode {
 400         const char *name;
 401         /* Return true if this mode is supported, otherwise false */
 402         bool (*supported)(void);
 403         /* Hook when the vm creation is done (before vcpu creation) */
 404         void (*create_vm_done)(struct kvm_vm *vm);
 405         /* Hook to collect the dirty pages into the bitmap provided */
 406         void (*collect_dirty_pages) (struct kvm_vcpu *vcpu, int slot,
 407                                      void *bitmap, uint32_t num_pages);
 408         /* Hook to call when after each vcpu run */
 409         void (*after_vcpu_run)(struct kvm_vcpu *vcpu, int ret, int err);
 410         void (*before_vcpu_join) (void);
 411 } log_modes[LOG_MODE_NUM] = {
 412         {
 413                 .name = "dirty-log",
 414                 .collect_dirty_pages = dirty_log_collect_dirty_pages,
 415                 .after_vcpu_run = default_after_vcpu_run,
 416         },
 417         {
 418                 .name = "clear-log",
 419                 .supported = clear_log_supported,
 420                 .create_vm_done = clear_log_create_vm_done,
 421                 .collect_dirty_pages = clear_log_collect_dirty_pages,
 422                 .after_vcpu_run = default_after_vcpu_run,
 423         },
 424         {
 425                 .name = "dirty-ring",
 426                 .supported = dirty_ring_supported,
 427                 .create_vm_done = dirty_ring_create_vm_done,
 428                 .collect_dirty_pages = dirty_ring_collect_dirty_pages,
 429                 .before_vcpu_join = dirty_ring_before_vcpu_join,
 430                 .after_vcpu_run = dirty_ring_after_vcpu_run,
 431         },
 432 };
 433
 434 /*
 435  * We use this bitmap to track some pages that should have its dirty
 436  * bit set in the _next_ iteration.  For example, if we detected the
 437  * page value changed to current iteration but at the same time the
 438  * page bit is cleared in the latest bitmap, then the system must
 439  * report that write in the next get dirty log call.
 440  */
 441 static unsigned long *host_bmap_track;
 442
 443 static void log_modes_dump(void)
 444 {
 445         int i;
 446
 447         printf("all");
 448         for (i = 0; i < LOG_MODE_NUM; i++)
 449                 printf(", %s", log_modes[i].name);
 450         printf("\n");
 451 }
 452
 453 static bool log_mode_supported(void)
 454 {
 455         struct log_mode *mode = &log_modes[host_log_mode];
 456
 457         if (mode->supported)
 458                 return mode->supported();
 459
 460         return true;
 461 }
 462
 463 static void log_mode_create_vm_done(struct kvm_vm *vm)
 464 {
 465         struct log_mode *mode = &log_modes[host_log_mode];
 466
 467         if (mode->create_vm_done)
 468                 mode->create_vm_done(vm);
 469 }
 470
 471 static void log_mode_collect_dirty_pages(struct kvm_vcpu *vcpu, int slot,
 472                                          void *bitmap, uint32_t num_pages)
 473 {
 474         struct log_mode *mode = &log_modes[host_log_mode];
 475
 476         TEST_ASSERT(mode->collect_dirty_pages != NULL,
 477                     "collect_dirty_pages() is required for any log mode!");
 478         mode->collect_dirty_pages(vcpu, slot, bitmap, num_pages);
 479 }
 480
 481 static void log_mode_after_vcpu_run(struct kvm_vcpu *vcpu, int ret, int err)
 482 {
 483         struct log_mode *mode = &log_modes[host_log_mode];
 484
 485         if (mode->after_vcpu_run)
 486                 mode->after_vcpu_run(vcpu, ret, err);
 487 }
 488
 489 static void log_mode_before_vcpu_join(void)
 490 {
 491         struct log_mode *mode = &log_modes[host_log_mode];
 492
 493         if (mode->before_vcpu_join)
 494                 mode->before_vcpu_join();
 495 }
 496
 497 static void generate_random_array(uint64_t *guest_array, uint64_t size)
 498 {
 499         uint64_t i;
 500
 501         for (i = 0; i < size; i++)
 502                 guest_array[i] = random();
 503 }
 504
 505 static void *vcpu_worker(void *data)
 506 {
 507         int ret;
 508         struct kvm_vcpu *vcpu = data;
 509         struct kvm_vm *vm = vcpu->vm;
 510         uint64_t *guest_array;
 511         uint64_t pages_count = 0;
 512         struct kvm_signal_mask *sigmask = alloca(offsetof(struct kvm_signal_mask, sigset)
 513                                                  + sizeof(sigset_t));
 514         sigset_t *sigset = (sigset_t *) &sigmask->sigset;
 515
 516         /*
 517          * SIG_IPI is unblocked atomically while in KVM_RUN.  It causes the
 518          * ioctl to return with -EINTR, but it is still pending and we need
 519          * to accept it with the sigwait.
 520          */
 521         sigmask->len = 8;
 522         pthread_sigmask(0, NULL, sigset);
 523         sigdelset(sigset, SIG_IPI);
 524         vcpu_ioctl(vcpu, KVM_SET_SIGNAL_MASK, sigmask);
 525
 526         sigemptyset(sigset);
 527         sigaddset(sigset, SIG_IPI);
 528
 529         guest_array = addr_gva2hva(vm, (vm_vaddr_t)random_array);
 530
 531         while (!READ_ONCE(host_quit)) {
 532                 /* Clear any existing kick signals */
 533                 generate_random_array(guest_array, TEST_PAGES_PER_LOOP);
 534                 pages_count += TEST_PAGES_PER_LOOP;
 535                 /* Let the guest dirty the random pages */
 536                 ret = __vcpu_run(vcpu);
 537                 if (ret == -1 && errno == EINTR) {
 538                         int sig = -1;
 539                         sigwait(sigset, &sig);
 540                         assert(sig == SIG_IPI);
 541                 }
 542                 log_mode_after_vcpu_run(vcpu, ret, errno);
 543         }
 544
 545         pr_info("Dirtied %"PRIu64" pages\n", pages_count);
 546
 547         return NULL;
 548 }
 549
 550 static void vm_dirty_log_verify(enum vm_guest_mode mode, unsigned long *bmap)
 551 {
 552         uint64_t step = vm_num_host_pages(mode, 1);
 553         uint64_t page;
 554         uint64_t *value_ptr;
 555         uint64_t min_iter = 0;
 556
 557         for (page = 0; page < host_num_pages; page += step) {
 558                 value_ptr = host_test_mem + page * host_page_size;
 559
 560                 /* If this is a special page that we were tracking... */
 561                 if (test_and_clear_bit_le(page, host_bmap_track)) {
 562                         host_track_next_count++;
 563                         TEST_ASSERT(test_bit_le(page, bmap),
 564                                     "Page %"PRIu64" should have its dirty bit "
 565                                     "set in this iteration but it is missing",
 566                                     page);
 567                 }
 568
 569                 if (test_and_clear_bit_le(page, bmap)) {
 570                         bool matched;
 571
 572                         host_dirty_count++;
 573
 574                         /*
 575                          * If the bit is set, the value written onto
 576                          * the corresponding page should be either the
 577                          * previous iteration number or the current one.
 578                          */
 579                         matched = (*value_ptr == iteration ||
 580                                    *value_ptr == iteration - 1);
 581
 582                         if (host_log_mode == LOG_MODE_DIRTY_RING && !matched) {
 583                                 if (*value_ptr == iteration - 2 && min_iter <= iteration - 2) {
 584                                         /*
 585                                          * Short answer: this case is special
 586                                          * only for dirty ring test where the
 587                                          * page is the last page before a kvm
 588                                          * dirty ring full in iteration N-2.
 589                                          *
 590                                          * Long answer: Assuming ring size R,
 591                                          * one possible condition is:
 592                                          *
 593                                          *      main thr       vcpu thr
 594                                          *      --------       --------
 595                                          *    iter=1
 596                                          *                   write 1 to page 0~(R-1)
 597                                          *                   full, vmexit
 598                                          *    collect 0~(R-1)
 599                                          *    kick vcpu
 600                                          *                   write 1 to (R-1)~(2R-2)
 601                                          *                   full, vmexit
 602                                          *    iter=2
 603                                          *    collect (R-1)~(2R-2)
 604                                          *    kick vcpu
 605                                          *                   write 1 to (2R-2)
 606                                          *                   (NOTE!!! "1" cached in cpu reg)
 607                                          *                   write 2 to (2R-1)~(3R-3)
 608                                          *                   full, vmexit
 609                                          *    iter=3
 610                                          *    collect (2R-2)~(3R-3)
 611                                          *    (here if we read value on page
 612                                          *     "2R-2" is 1, while iter=3!!!)
 613                                          *
 614                                          * This however can only happen once per iteration.
 615                                          */
 616                                         min_iter = iteration - 1;
 617                                         continue;
 618                                 } else if (page == dirty_ring_last_page) {
 619                                         /*
 620                                          * Please refer to comments in
 621                                          * dirty_ring_last_page.
 622                                          */
 623                                         continue;
 624                                 }
 625                         }
 626
 627                         TEST_ASSERT(matched,
 628                                     "Set page %"PRIu64" value %"PRIu64
 629                                     " incorrect (iteration=%"PRIu64")",
 630                                     page, *value_ptr, iteration);
 631                 } else {
 632                         host_clear_count++;
 633                         /*
 634                          * If cleared, the value written can be any
 635                          * value smaller or equals to the iteration
 636                          * number.  Note that the value can be exactly
 637                          * (iteration-1) if that write can happen
 638                          * like this:
 639                          *
 640                          * (1) increase loop count to "iteration-1"
 641                          * (2) write to page P happens (with value
 642                          *     "iteration-1")
 643                          * (3) get dirty log for "iteration-1"; we'll
 644                          *     see that page P bit is set (dirtied),
 645                          *     and not set the bit in host_bmap_track
 646                          * (4) increase loop count to "iteration"
 647                          *     (which is current iteration)
 648                          * (5) get dirty log for current iteration,
 649                          *     we'll see that page P is cleared, with
 650                          *     value "iteration-1".
 651                          */
 652                         TEST_ASSERT(*value_ptr <= iteration,
 653                                     "Clear page %"PRIu64" value %"PRIu64
 654                                     " incorrect (iteration=%"PRIu64")",
 655                                     page, *value_ptr, iteration);
 656                         if (*value_ptr == iteration) {
 657                                 /*
 658                                  * This page is _just_ modified; it
 659                                  * should report its dirtyness in the
 660                                  * next run
 661                                  */
 662                                 set_bit_le(page, host_bmap_track);
 663                         }
 664                 }
 665         }
 666 }
 667
 668 static struct kvm_vm *create_vm(enum vm_guest_mode mode, struct kvm_vcpu **vcpu,
 669                                 uint64_t extra_mem_pages, void *guest_code)
 670 {
 671         struct kvm_vm *vm;
 672
 673         pr_info("Testing guest mode: %s\n", vm_guest_mode_string(mode));
 674
 675         vm = __vm_create(mode, 1, extra_mem_pages);
 676
 677         log_mode_create_vm_done(vm);
 678         *vcpu = vm_vcpu_add(vm, 0, guest_code);
 679         return vm;
 680 }
 681
 682 #define DIRTY_MEM_BITS 30 /* 1G */
 683 #define PAGE_SHIFT_4K  12
 684
 685 struct test_params {
 686         unsigned long iterations;
 687         unsigned long interval;
 688         uint64_t phys_offset;
 689 };
 690
 691 static void run_test(enum vm_guest_mode mode, void *arg)
 692 {
 693         struct test_params *p = arg;
 694         struct kvm_vcpu *vcpu;
 695         struct kvm_vm *vm;
 696         unsigned long *bmap;
 697
 698         if (!log_mode_supported()) {
 699                 print_skip("Log mode '%s' not supported",
 700                            log_modes[host_log_mode].name);
 701                 return;
 702         }
 703
 704         /*
 705          * We reserve page table for 2 times of extra dirty mem which
 706          * will definitely cover the original (1G+) test range.  Here
 707          * we do the calculation with 4K page size which is the
 708          * smallest so the page number will be enough for all archs
 709          * (e.g., 64K page size guest will need even less memory for
 710          * page tables).
 711          */
 712         vm = create_vm(mode, &vcpu,
 713                        2ul << (DIRTY_MEM_BITS - PAGE_SHIFT_4K), guest_code);
 714
 715         guest_page_size = vm->page_size;
 716         /*
 717          * A little more than 1G of guest page sized pages.  Cover the
 718          * case where the size is not aligned to 64 pages.
 719          */
 720         guest_num_pages = (1ul << (DIRTY_MEM_BITS - vm->page_shift)) + 3;
 721         guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
 722
 723         host_page_size = getpagesize();
 724         host_num_pages = vm_num_host_pages(mode, guest_num_pages);
 725
 726         if (!p->phys_offset) {
 727                 guest_test_phys_mem = (vm->max_gfn - guest_num_pages) *
 728                                       guest_page_size;
 729                 guest_test_phys_mem = align_down(guest_test_phys_mem, host_page_size);
 730         } else {
 731                 guest_test_phys_mem = p->phys_offset;
 732         }
 733
 734 #ifdef __s390x__
 735         /* Align to 1M (segment size) */
 736         guest_test_phys_mem = align_down(guest_test_phys_mem, 1 << 20);
 737 #endif
 738
 739         pr_info("guest physical test memory offset: 0x%lx\n", guest_test_phys_mem);
 740
 741         bmap = bitmap_zalloc(host_num_pages);
 742         host_bmap_track = bitmap_zalloc(host_num_pages);
 743
 744         /* Add an extra memory slot for testing dirty logging */
 745         vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS,
 746                                     guest_test_phys_mem,
 747                                     TEST_MEM_SLOT_INDEX,
 748                                     guest_num_pages,
 749                                     KVM_MEM_LOG_DIRTY_PAGES);
 750
 751         /* Do mapping for the dirty track memory slot */
 752         virt_map(vm, guest_test_virt_mem, guest_test_phys_mem, guest_num_pages);
 753
 754         /* Cache the HVA pointer of the region */
 755         host_test_mem = addr_gpa2hva(vm, (vm_paddr_t)guest_test_phys_mem);
 756
 757         ucall_init(vm, NULL);
 758
 759         /* Export the shared variables to the guest */
 760         sync_global_to_guest(vm, host_page_size);
 761         sync_global_to_guest(vm, guest_page_size);
 762         sync_global_to_guest(vm, guest_test_virt_mem);
 763         sync_global_to_guest(vm, guest_num_pages);
 764
 765         /* Start the iterations */
 766         iteration = 1;
 767         sync_global_to_guest(vm, iteration);
 768         host_quit = false;
 769         host_dirty_count = 0;
 770         host_clear_count = 0;
 771         host_track_next_count = 0;
 772
 773         pthread_create(&vcpu_thread, NULL, vcpu_worker, vcpu);
 774
 775         while (iteration < p->iterations) {
 776                 /* Give the vcpu thread some time to dirty some pages */
 777                 usleep(p->interval * 1000);
 778                 log_mode_collect_dirty_pages(vcpu, TEST_MEM_SLOT_INDEX,
 779                                              bmap, host_num_pages);
 780
 781                 /*
 782                  * See vcpu_sync_stop_requested definition for details on why
 783                  * we need to stop vcpu when verify data.
 784                  */
 785                 atomic_set(&vcpu_sync_stop_requested, true);
 786                 sem_wait_until(&sem_vcpu_stop);
 787                 /*
 788                  * NOTE: for dirty ring, it's possible that we didn't stop at
 789                  * GUEST_SYNC but instead we stopped because ring is full;
 790                  * that's okay too because ring full means we're only missing
 791                  * the flush of the last page, and since we handle the last
 792                  * page specially verification will succeed anyway.
 793                  */
 794                 assert(host_log_mode == LOG_MODE_DIRTY_RING ||
 795                        atomic_read(&vcpu_sync_stop_requested) == false);
 796                 vm_dirty_log_verify(mode, bmap);
 797                 sem_post(&sem_vcpu_cont);
 798
 799                 iteration++;
 800                 sync_global_to_guest(vm, iteration);
 801         }
 802
 803         /* Tell the vcpu thread to quit */
 804         host_quit = true;
 805         log_mode_before_vcpu_join();
 806         pthread_join(vcpu_thread, NULL);
 807
 808         pr_info("Total bits checked: dirty (%"PRIu64"), clear (%"PRIu64"), "
 809                 "track_next (%"PRIu64")\n", host_dirty_count, host_clear_count,
 810                 host_track_next_count);
 811
 812         free(bmap);
 813         free(host_bmap_track);
 814         ucall_uninit(vm);
 815         kvm_vm_free(vm);
 816 }
 817
 818 static void help(char *name)
 819 {
 820         puts("");
 821         printf("usage: %s [-h] [-i iterations] [-I interval] "
 822                "[-p offset] [-m mode]\n", name);
 823         puts("");
 824         printf(" -c: specify dirty ring size, in number of entries\n");
 825         printf("     (only useful for dirty-ring test; default: %"PRIu32")\n",
 826                TEST_DIRTY_RING_COUNT);
 827         printf(" -i: specify iteration counts (default: %"PRIu64")\n",
 828                TEST_HOST_LOOP_N);
 829         printf(" -I: specify interval in ms (default: %"PRIu64" ms)\n",
 830                TEST_HOST_LOOP_INTERVAL);
 831         printf(" -p: specify guest physical test memory offset\n"
 832                "     Warning: a low offset can conflict with the loaded test code.\n");
 833         printf(" -M: specify the host logging mode "
 834                "(default: run all log modes).  Supported modes: \n\t");
 835         log_modes_dump();
 836         guest_modes_help();
 837         puts("");
 838         exit(0);
 839 }
 840
 841 int main(int argc, char *argv[])
 842 {
 843         struct test_params p = {
 844                 .iterations = TEST_HOST_LOOP_N,
 845                 .interval = TEST_HOST_LOOP_INTERVAL,
 846         };
 847         int opt, i;
 848         sigset_t sigset;
 849
 850         sem_init(&sem_vcpu_stop, 0, 0);
 851         sem_init(&sem_vcpu_cont, 0, 0);
 852
 853         guest_modes_append_default();
 854
 855         while ((opt = getopt(argc, argv, "c:hi:I:p:m:M:")) != -1) {
 856                 switch (opt) {
 857                 case 'c':
 858                         test_dirty_ring_count = strtol(optarg, NULL, 10);
 859                         break;
 860                 case 'i':
 861                         p.iterations = strtol(optarg, NULL, 10);
 862                         break;
 863                 case 'I':
 864                         p.interval = strtol(optarg, NULL, 10);
 865                         break;
 866                 case 'p':
 867                         p.phys_offset = strtoull(optarg, NULL, 0);
 868                         break;
 869                 case 'm':
 870                         guest_modes_cmdline(optarg);
 871                         break;
 872                 case 'M':
 873                         if (!strcmp(optarg, "all")) {
 874                                 host_log_mode_option = LOG_MODE_ALL;
 875                                 break;
 876                         }
 877                         for (i = 0; i < LOG_MODE_NUM; i++) {
 878                                 if (!strcmp(optarg, log_modes[i].name)) {
 879                                         pr_info("Setting log mode to: '%s'\n",
 880                                                 optarg);
 881                                         host_log_mode_option = i;
 882                                         break;
 883                                 }
 884                         }
 885                         if (i == LOG_MODE_NUM) {
 886                                 printf("Log mode '%s' invalid. Please choose "
 887                                        "from: ", optarg);
 888                                 log_modes_dump();
 889                                 exit(1);
 890                         }
 891                         break;
 892                 case 'h':
 893                 default:
 894                         help(argv[0]);
 895                         break;
 896                 }
 897         }
 898
 899         TEST_ASSERT(p.iterations > 2, "Iterations must be greater than two");
 900         TEST_ASSERT(p.interval > 0, "Interval must be greater than zero");
 901
 902         pr_info("Test iterations: %"PRIu64", interval: %"PRIu64" (ms)\n",
 903                 p.iterations, p.interval);
 904
 905         srandom(time(0));
 906
 907         /* Ensure that vCPU threads start with SIG_IPI blocked.  */
 908         sigemptyset(&sigset);
 909         sigaddset(&sigset, SIG_IPI);
 910         pthread_sigmask(SIG_BLOCK, &sigset, NULL);
 911
 912         if (host_log_mode_option == LOG_MODE_ALL) {
 913                 /* Run each log mode */
 914                 for (i = 0; i < LOG_MODE_NUM; i++) {
 915                         pr_info("Testing Log Mode '%s'\n", log_modes[i].name);
 916                         host_log_mode = i;
 917                         for_each_guest_mode(run_test, &p);
 918                 }
 919         } else {
 920                 host_log_mode = host_log_mode_option;
 921                 for_each_guest_mode(run_test, &p);
 922         }
 923
 924         return 0;
 925 }