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 <asm/barrier.h>
  20 #include <linux/atomic.h>
  21
  22 #include "kvm_util.h"
  23 #include "test_util.h"
  24 #include "guest_modes.h"
  25 #include "processor.h"
  26
  27 #define VCPU_ID                         1
  28
  29 /* The memory slot index to track dirty pages */
  30 #define TEST_MEM_SLOT_INDEX             1
  31
  32 /* Default guest test virtual memory offset */
  33 #define DEFAULT_GUEST_TEST_MEM          0xc0000000
  34
  35 /* How many pages to dirty for each guest loop */
  36 #define TEST_PAGES_PER_LOOP             1024
  37
  38 /* How many host loops to run (one KVM_GET_DIRTY_LOG for each loop) */
  39 #define TEST_HOST_LOOP_N                32UL
  40
  41 /* Interval for each host loop (ms) */
  42 #define TEST_HOST_LOOP_INTERVAL         10UL
  43
  44 /* Dirty bitmaps are always little endian, so we need to swap on big endian */
  45 #if defined(__s390x__)
  46 # define BITOP_LE_SWIZZLE       ((BITS_PER_LONG-1) & ~0x7)
  47 # define test_bit_le(nr, addr) \
  48         test_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  49 # define set_bit_le(nr, addr) \
  50         set_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  51 # define clear_bit_le(nr, addr) \
  52         clear_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  53 # define test_and_set_bit_le(nr, addr) \
  54         test_and_set_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  55 # define test_and_clear_bit_le(nr, addr) \
  56         test_and_clear_bit((nr) ^ BITOP_LE_SWIZZLE, addr)
  57 #else
  58 # define test_bit_le            test_bit
  59 # define set_bit_le             set_bit
  60 # define clear_bit_le           clear_bit
  61 # define test_and_set_bit_le    test_and_set_bit
  62 # define test_and_clear_bit_le  test_and_clear_bit
  63 #endif
  64
  65 #define TEST_DIRTY_RING_COUNT           65536
  66
  67 #define SIG_IPI SIGUSR1
  68
  69 /*
  70  * Guest/Host shared variables. Ensure addr_gva2hva() and/or
  71  * sync_global_to/from_guest() are used when accessing from
  72  * the host. READ/WRITE_ONCE() should also be used with anything
  73  * that may change.
  74  */
  75 static uint64_t host_page_size;
  76 static uint64_t guest_page_size;
  77 static uint64_t guest_num_pages;
  78 static uint64_t random_array[TEST_PAGES_PER_LOOP];
  79 static uint64_t iteration;
  80
  81 /*
  82  * Guest physical memory offset of the testing memory slot.
  83  * This will be set to the topmost valid physical address minus
  84  * the test memory size.
  85  */
  86 static uint64_t guest_test_phys_mem;
  87
  88 /*
  89  * Guest virtual memory offset of the testing memory slot.
  90  * Must not conflict with identity mapped test code.
  91  */
  92 static uint64_t guest_test_virt_mem = DEFAULT_GUEST_TEST_MEM;
  93
  94 /*
  95  * Continuously write to the first 8 bytes of a random pages within
  96  * the testing memory region.
  97  */
  98 static void guest_code(void)
  99 {
 100         uint64_t addr;
 101         int i;
 102
 103         /*
 104          * On s390x, all pages of a 1M segment are initially marked as dirty
 105          * when a page of the segment is written to for the very first time.
 106          * To compensate this specialty in this test, we need to touch all
 107          * pages during the first iteration.
 108          */
 109         for (i = 0; i < guest_num_pages; i++) {
 110                 addr = guest_test_virt_mem + i * guest_page_size;
 111                 *(uint64_t *)addr = READ_ONCE(iteration);
 112         }
 113
 114         while (true) {
 115                 for (i = 0; i < TEST_PAGES_PER_LOOP; i++) {
 116                         addr = guest_test_virt_mem;
 117                         addr += (READ_ONCE(random_array[i]) % guest_num_pages)
 118                                 * guest_page_size;
 119                         addr &= ~(host_page_size - 1);
 120                         *(uint64_t *)addr = READ_ONCE(iteration);
 121                 }
 122
 123                 /* Tell the host that we need more random numbers */
 124                 GUEST_SYNC(1);
 125         }
 126 }
 127
 128 /* Host variables */
 129 static bool host_quit;
 130
 131 /* Points to the test VM memory region on which we track dirty logs */
 132 static void *host_test_mem;
 133 static uint64_t host_num_pages;
 134
 135 /* For statistics only */
 136 static uint64_t host_dirty_count;
 137 static uint64_t host_clear_count;
 138 static uint64_t host_track_next_count;
 139
 140 /* Whether dirty ring reset is requested, or finished */
 141 static sem_t sem_vcpu_stop;
 142 static sem_t sem_vcpu_cont;
 143 /*
 144  * This is only set by main thread, and only cleared by vcpu thread.  It is
 145  * used to request vcpu thread to stop at the next GUEST_SYNC, since GUEST_SYNC
 146  * is the only place that we'll guarantee both "dirty bit" and "dirty data"
 147  * will match.  E.g., SIG_IPI won't guarantee that if the vcpu is interrupted
 148  * after setting dirty bit but before the data is written.
 149  */
 150 static atomic_t vcpu_sync_stop_requested;
 151 /*
 152  * This is updated by the vcpu thread to tell the host whether it's a
 153  * ring-full event.  It should only be read until a sem_wait() of
 154  * sem_vcpu_stop and before vcpu continues to run.
 155  */
 156 static bool dirty_ring_vcpu_ring_full;
 157 /*
 158  * This is only used for verifying the dirty pages.  Dirty ring has a very
 159  * tricky case when the ring just got full, kvm will do userspace exit due to
 160  * ring full.  When that happens, the very last PFN is set but actually the
 161  * data is not changed (the guest WRITE is not really applied yet), because
 162  * we found that the dirty ring is full, refused to continue the vcpu, and
 163  * recorded the dirty gfn with the old contents.
 164  *
 165  * For this specific case, it's safe to skip checking this pfn for this
 166  * bit, because it's a redundant bit, and when the write happens later the bit
 167  * will be set again.  We use this variable to always keep track of the latest
 168  * dirty gfn we've collected, so that if a mismatch of data found later in the
 169  * verifying process, we let it pass.
 170  */
 171 static uint64_t dirty_ring_last_page;
 172
 173 enum log_mode_t {
 174         /* Only use KVM_GET_DIRTY_LOG for logging */
 175         LOG_MODE_DIRTY_LOG = 0,
 176
 177         /* Use both KVM_[GET|CLEAR]_DIRTY_LOG for logging */
 178         LOG_MODE_CLEAR_LOG = 1,
 179
 180         /* Use dirty ring for logging */
 181         LOG_MODE_DIRTY_RING = 2,
 182
 183         LOG_MODE_NUM,
 184
 185         /* Run all supported modes */
 186         LOG_MODE_ALL = LOG_MODE_NUM,
 187 };
 188
 189 /* Mode of logging to test.  Default is to run all supported modes */
 190 static enum log_mode_t host_log_mode_option = LOG_MODE_ALL;
 191 /* Logging mode for current run */
 192 static enum log_mode_t host_log_mode;
 193 static pthread_t vcpu_thread;
 194 static uint32_t test_dirty_ring_count = TEST_DIRTY_RING_COUNT;
 195
 196 static void vcpu_kick(void)
 197 {
 198         pthread_kill(vcpu_thread, SIG_IPI);
 199 }
 200
 201 /*
 202  * In our test we do signal tricks, let's use a better version of
 203  * sem_wait to avoid signal interrupts
 204  */
 205 static void sem_wait_until(sem_t *sem)
 206 {
 207         int ret;
 208
 209         do
 210                 ret = sem_wait(sem);
 211         while (ret == -1 && errno == EINTR);
 212 }
 213
 214 static bool clear_log_supported(void)
 215 {
 216         return kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
 217 }
 218
 219 static void clear_log_create_vm_done(struct kvm_vm *vm)
 220 {
 221         struct kvm_enable_cap cap = {};
 222         u64 manual_caps;
 223
 224         manual_caps = kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
 225         TEST_ASSERT(manual_caps, "MANUAL_CAPS is zero!");
 226         manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
 227                         KVM_DIRTY_LOG_INITIALLY_SET);
 228         cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
 229         cap.args[0] = manual_caps;
 230         vm_enable_cap(vm, &cap);
 231 }
 232
 233 static void dirty_log_collect_dirty_pages(struct kvm_vm *vm, int slot,
 234                                           void *bitmap, uint32_t num_pages)
 235 {
 236         kvm_vm_get_dirty_log(vm, slot, bitmap);
 237 }
 238
 239 static void clear_log_collect_dirty_pages(struct kvm_vm *vm, int slot,
 240                                           void *bitmap, uint32_t num_pages)
 241 {
 242         kvm_vm_get_dirty_log(vm, slot, bitmap);
 243         kvm_vm_clear_dirty_log(vm, slot, bitmap, 0, num_pages);
 244 }
 245
 246 /* Should only be called after a GUEST_SYNC */
 247 static void vcpu_handle_sync_stop(void)
 248 {
 249         if (atomic_read(&vcpu_sync_stop_requested)) {
 250                 /* It means main thread is sleeping waiting */
 251                 atomic_set(&vcpu_sync_stop_requested, false);
 252                 sem_post(&sem_vcpu_stop);
 253                 sem_wait_until(&sem_vcpu_cont);
 254         }
 255 }
 256
 257 static void default_after_vcpu_run(struct kvm_vm *vm, int ret, int err)
 258 {
 259         struct kvm_run *run = vcpu_state(vm, VCPU_ID);
 260
 261         TEST_ASSERT(ret == 0 || (ret == -1 && err == EINTR),
 262                     "vcpu run failed: errno=%d", err);
 263
 264         TEST_ASSERT(get_ucall(vm, VCPU_ID, NULL) == UCALL_SYNC,
 265                     "Invalid guest sync status: exit_reason=%s\n",
 266                     exit_reason_str(run->exit_reason));
 267
 268         vcpu_handle_sync_stop();
 269 }
 270
 271 static bool dirty_ring_supported(void)
 272 {
 273         return kvm_check_cap(KVM_CAP_DIRTY_LOG_RING);
 274 }
 275
 276 static void dirty_ring_create_vm_done(struct kvm_vm *vm)
 277 {
 278         /*
 279          * Switch to dirty ring mode after VM creation but before any
 280          * of the vcpu creation.
 281          */
 282         vm_enable_dirty_ring(vm, test_dirty_ring_count *
 283                              sizeof(struct kvm_dirty_gfn));
 284 }
 285
 286 static inline bool dirty_gfn_is_dirtied(struct kvm_dirty_gfn *gfn)
 287 {
 288         return gfn->flags == KVM_DIRTY_GFN_F_DIRTY;
 289 }
 290
 291 static inline void dirty_gfn_set_collected(struct kvm_dirty_gfn *gfn)
 292 {
 293         gfn->flags = KVM_DIRTY_GFN_F_RESET;
 294 }
 295
 296 static uint32_t dirty_ring_collect_one(struct kvm_dirty_gfn *dirty_gfns,
 297                                        int slot, void *bitmap,
 298                                        uint32_t num_pages, uint32_t *fetch_index)
 299 {
 300         struct kvm_dirty_gfn *cur;
 301         uint32_t count = 0;
 302
 303         while (true) {
 304                 cur = &dirty_gfns[*fetch_index % test_dirty_ring_count];
 305                 if (!dirty_gfn_is_dirtied(cur))
 306                         break;
 307                 TEST_ASSERT(cur->slot == slot, "Slot number didn't match: "
 308                             "%u != %u", cur->slot, slot);
 309                 TEST_ASSERT(cur->offset < num_pages, "Offset overflow: "
 310                             "0x%llx >= 0x%x", cur->offset, num_pages);
 311                 //pr_info("fetch 0x%x page %llu\n", *fetch_index, cur->offset);
 312                 set_bit_le(cur->offset, bitmap);
 313                 dirty_ring_last_page = cur->offset;
 314                 dirty_gfn_set_collected(cur);
 315                 (*fetch_index)++;
 316                 count++;
 317         }
 318
 319         return count;
 320 }
 321
 322 static void dirty_ring_wait_vcpu(void)
 323 {
 324         /* This makes sure that hardware PML cache flushed */
 325         vcpu_kick();
 326         sem_wait_until(&sem_vcpu_stop);
 327 }
 328
 329 static void dirty_ring_continue_vcpu(void)
 330 {
 331         pr_info("Notifying vcpu to continue\n");
 332         sem_post(&sem_vcpu_cont);
 333 }
 334
 335 static void dirty_ring_collect_dirty_pages(struct kvm_vm *vm, int slot,
 336                                            void *bitmap, uint32_t num_pages)
 337 {
 338         /* We only have one vcpu */
 339         static uint32_t fetch_index = 0;
 340         uint32_t count = 0, cleared;
 341         bool continued_vcpu = false;
 342
 343         dirty_ring_wait_vcpu();
 344
 345         if (!dirty_ring_vcpu_ring_full) {
 346                 /*
 347                  * This is not a ring-full event, it's safe to allow
 348                  * vcpu to continue
 349                  */
 350                 dirty_ring_continue_vcpu();
 351                 continued_vcpu = true;
 352         }
 353
 354         /* Only have one vcpu */
 355         count = dirty_ring_collect_one(vcpu_map_dirty_ring(vm, VCPU_ID),
 356                                        slot, bitmap, num_pages, &fetch_index);
 357
 358         cleared = kvm_vm_reset_dirty_ring(vm);
 359
 360         /* Cleared pages should be the same as collected */
 361         TEST_ASSERT(cleared == count, "Reset dirty pages (%u) mismatch "
 362                     "with collected (%u)", cleared, count);
 363
 364         if (!continued_vcpu) {
 365                 TEST_ASSERT(dirty_ring_vcpu_ring_full,
 366                             "Didn't continue vcpu even without ring full");
 367                 dirty_ring_continue_vcpu();
 368         }
 369
 370         pr_info("Iteration %ld collected %u pages\n", iteration, count);
 371 }
 372
 373 static void dirty_ring_after_vcpu_run(struct kvm_vm *vm, int ret, int err)
 374 {
 375         struct kvm_run *run = vcpu_state(vm, VCPU_ID);
 376
 377         /* A ucall-sync or ring-full event is allowed */
 378         if (get_ucall(vm, VCPU_ID, NULL) == UCALL_SYNC) {
 379                 /* We should allow this to continue */
 380                 ;
 381         } else if (run->exit_reason == KVM_EXIT_DIRTY_RING_FULL ||
 382                    (ret == -1 && err == EINTR)) {
 383                 /* Update the flag first before pause */
 384                 WRITE_ONCE(dirty_ring_vcpu_ring_full,
 385                            run->exit_reason == KVM_EXIT_DIRTY_RING_FULL);
 386                 sem_post(&sem_vcpu_stop);
 387                 pr_info("vcpu stops because %s...\n",
 388                         dirty_ring_vcpu_ring_full ?
 389                         "dirty ring is full" : "vcpu is kicked out");
 390                 sem_wait_until(&sem_vcpu_cont);
 391                 pr_info("vcpu continues now.\n");
 392         } else {
 393                 TEST_ASSERT(false, "Invalid guest sync status: "
 394                             "exit_reason=%s\n",
 395                             exit_reason_str(run->exit_reason));
 396         }
 397 }
 398
 399 static void dirty_ring_before_vcpu_join(void)
 400 {
 401         /* Kick another round of vcpu just to make sure it will quit */
 402         sem_post(&sem_vcpu_cont);
 403 }
 404
 405 struct log_mode {
 406         const char *name;
 407         /* Return true if this mode is supported, otherwise false */
 408         bool (*supported)(void);
 409         /* Hook when the vm creation is done (before vcpu creation) */
 410         void (*create_vm_done)(struct kvm_vm *vm);
 411         /* Hook to collect the dirty pages into the bitmap provided */
 412         void (*collect_dirty_pages) (struct kvm_vm *vm, int slot,
 413                                      void *bitmap, uint32_t num_pages);
 414         /* Hook to call when after each vcpu run */
 415         void (*after_vcpu_run)(struct kvm_vm *vm, int ret, int err);
 416         void (*before_vcpu_join) (void);
 417 } log_modes[LOG_MODE_NUM] = {
 418         {
 419                 .name = "dirty-log",
 420                 .collect_dirty_pages = dirty_log_collect_dirty_pages,
 421                 .after_vcpu_run = default_after_vcpu_run,
 422         },
 423         {
 424                 .name = "clear-log",
 425                 .supported = clear_log_supported,
 426                 .create_vm_done = clear_log_create_vm_done,
 427                 .collect_dirty_pages = clear_log_collect_dirty_pages,
 428                 .after_vcpu_run = default_after_vcpu_run,
 429         },
 430         {
 431                 .name = "dirty-ring",
 432                 .supported = dirty_ring_supported,
 433                 .create_vm_done = dirty_ring_create_vm_done,
 434                 .collect_dirty_pages = dirty_ring_collect_dirty_pages,
 435                 .before_vcpu_join = dirty_ring_before_vcpu_join,
 436                 .after_vcpu_run = dirty_ring_after_vcpu_run,
 437         },
 438 };
 439
 440 /*
 441  * We use this bitmap to track some pages that should have its dirty
 442  * bit set in the _next_ iteration.  For example, if we detected the
 443  * page value changed to current iteration but at the same time the
 444  * page bit is cleared in the latest bitmap, then the system must
 445  * report that write in the next get dirty log call.
 446  */
 447 static unsigned long *host_bmap_track;
 448
 449 static void log_modes_dump(void)
 450 {
 451         int i;
 452
 453         printf("all");
 454         for (i = 0; i < LOG_MODE_NUM; i++)
 455                 printf(", %s", log_modes[i].name);
 456         printf("\n");
 457 }
 458
 459 static bool log_mode_supported(void)
 460 {
 461         struct log_mode *mode = &log_modes[host_log_mode];
 462
 463         if (mode->supported)
 464                 return mode->supported();
 465
 466         return true;
 467 }
 468
 469 static void log_mode_create_vm_done(struct kvm_vm *vm)
 470 {
 471         struct log_mode *mode = &log_modes[host_log_mode];
 472
 473         if (mode->create_vm_done)
 474                 mode->create_vm_done(vm);
 475 }
 476
 477 static void log_mode_collect_dirty_pages(struct kvm_vm *vm, int slot,
 478                                          void *bitmap, uint32_t num_pages)
 479 {
 480         struct log_mode *mode = &log_modes[host_log_mode];
 481
 482         TEST_ASSERT(mode->collect_dirty_pages != NULL,
 483                     "collect_dirty_pages() is required for any log mode!");
 484         mode->collect_dirty_pages(vm, slot, bitmap, num_pages);
 485 }
 486
 487 static void log_mode_after_vcpu_run(struct kvm_vm *vm, int ret, int err)
 488 {
 489         struct log_mode *mode = &log_modes[host_log_mode];
 490
 491         if (mode->after_vcpu_run)
 492                 mode->after_vcpu_run(vm, ret, err);
 493 }
 494
 495 static void log_mode_before_vcpu_join(void)
 496 {
 497         struct log_mode *mode = &log_modes[host_log_mode];
 498
 499         if (mode->before_vcpu_join)
 500                 mode->before_vcpu_join();
 501 }
 502
 503 static void generate_random_array(uint64_t *guest_array, uint64_t size)
 504 {
 505         uint64_t i;
 506
 507         for (i = 0; i < size; i++)
 508                 guest_array[i] = random();
 509 }
 510
 511 static void *vcpu_worker(void *data)
 512 {
 513         int ret, vcpu_fd;
 514         struct kvm_vm *vm = data;
 515         uint64_t *guest_array;
 516         uint64_t pages_count = 0;
 517         struct kvm_signal_mask *sigmask = alloca(offsetof(struct kvm_signal_mask, sigset)
 518                                                  + sizeof(sigset_t));
 519         sigset_t *sigset = (sigset_t *) &sigmask->sigset;
 520
 521         vcpu_fd = vcpu_get_fd(vm, VCPU_ID);
 522
 523         /*
 524          * SIG_IPI is unblocked atomically while in KVM_RUN.  It causes the
 525          * ioctl to return with -EINTR, but it is still pending and we need
 526          * to accept it with the sigwait.
 527          */
 528         sigmask->len = 8;
 529         pthread_sigmask(0, NULL, sigset);
 530         sigdelset(sigset, SIG_IPI);
 531         vcpu_ioctl(vm, VCPU_ID, KVM_SET_SIGNAL_MASK, sigmask);
 532
 533         sigemptyset(sigset);
 534         sigaddset(sigset, SIG_IPI);
 535
 536         guest_array = addr_gva2hva(vm, (vm_vaddr_t)random_array);
 537
 538         while (!READ_ONCE(host_quit)) {
 539                 /* Clear any existing kick signals */
 540                 generate_random_array(guest_array, TEST_PAGES_PER_LOOP);
 541                 pages_count += TEST_PAGES_PER_LOOP;
 542                 /* Let the guest dirty the random pages */
 543                 ret = ioctl(vcpu_fd, KVM_RUN, NULL);
 544                 if (ret == -1 && errno == EINTR) {
 545                         int sig = -1;
 546                         sigwait(sigset, &sig);
 547                         assert(sig == SIG_IPI);
 548                 }
 549                 log_mode_after_vcpu_run(vm, ret, errno);
 550         }
 551
 552         pr_info("Dirtied %"PRIu64" pages\n", pages_count);
 553
 554         return NULL;
 555 }
 556
 557 static void vm_dirty_log_verify(enum vm_guest_mode mode, unsigned long *bmap)
 558 {
 559         uint64_t step = vm_num_host_pages(mode, 1);
 560         uint64_t page;
 561         uint64_t *value_ptr;
 562         uint64_t min_iter = 0;
 563
 564         for (page = 0; page < host_num_pages; page += step) {
 565                 value_ptr = host_test_mem + page * host_page_size;
 566
 567                 /* If this is a special page that we were tracking... */
 568                 if (test_and_clear_bit_le(page, host_bmap_track)) {
 569                         host_track_next_count++;
 570                         TEST_ASSERT(test_bit_le(page, bmap),
 571                                     "Page %"PRIu64" should have its dirty bit "
 572                                     "set in this iteration but it is missing",
 573                                     page);
 574                 }
 575
 576                 if (test_and_clear_bit_le(page, bmap)) {
 577                         bool matched;
 578
 579                         host_dirty_count++;
 580
 581                         /*
 582                          * If the bit is set, the value written onto
 583                          * the corresponding page should be either the
 584                          * previous iteration number or the current one.
 585                          */
 586                         matched = (*value_ptr == iteration ||
 587                                    *value_ptr == iteration - 1);
 588
 589                         if (host_log_mode == LOG_MODE_DIRTY_RING && !matched) {
 590                                 if (*value_ptr == iteration - 2 && min_iter <= iteration - 2) {
 591                                         /*
 592                                          * Short answer: this case is special
 593                                          * only for dirty ring test where the
 594                                          * page is the last page before a kvm
 595                                          * dirty ring full in iteration N-2.
 596                                          *
 597                                          * Long answer: Assuming ring size R,
 598                                          * one possible condition is:
 599                                          *
 600                                          *      main thr       vcpu thr
 601                                          *      --------       --------
 602                                          *    iter=1
 603                                          *                   write 1 to page 0~(R-1)
 604                                          *                   full, vmexit
 605                                          *    collect 0~(R-1)
 606                                          *    kick vcpu
 607                                          *                   write 1 to (R-1)~(2R-2)
 608                                          *                   full, vmexit
 609                                          *    iter=2
 610                                          *    collect (R-1)~(2R-2)
 611                                          *    kick vcpu
 612                                          *                   write 1 to (2R-2)
 613                                          *                   (NOTE!!! "1" cached in cpu reg)
 614                                          *                   write 2 to (2R-1)~(3R-3)
 615                                          *                   full, vmexit
 616                                          *    iter=3
 617                                          *    collect (2R-2)~(3R-3)
 618                                          *    (here if we read value on page
 619                                          *     "2R-2" is 1, while iter=3!!!)
 620                                          *
 621                                          * This however can only happen once per iteration.
 622                                          */
 623                                         min_iter = iteration - 1;
 624                                         continue;
 625                                 } else if (page == dirty_ring_last_page) {
 626                                         /*
 627                                          * Please refer to comments in
 628                                          * dirty_ring_last_page.
 629                                          */
 630                                         continue;
 631                                 }
 632                         }
 633
 634                         TEST_ASSERT(matched,
 635                                     "Set page %"PRIu64" value %"PRIu64
 636                                     " incorrect (iteration=%"PRIu64")",
 637                                     page, *value_ptr, iteration);
 638                 } else {
 639                         host_clear_count++;
 640                         /*
 641                          * If cleared, the value written can be any
 642                          * value smaller or equals to the iteration
 643                          * number.  Note that the value can be exactly
 644                          * (iteration-1) if that write can happen
 645                          * like this:
 646                          *
 647                          * (1) increase loop count to "iteration-1"
 648                          * (2) write to page P happens (with value
 649                          *     "iteration-1")
 650                          * (3) get dirty log for "iteration-1"; we'll
 651                          *     see that page P bit is set (dirtied),
 652                          *     and not set the bit in host_bmap_track
 653                          * (4) increase loop count to "iteration"
 654                          *     (which is current iteration)
 655                          * (5) get dirty log for current iteration,
 656                          *     we'll see that page P is cleared, with
 657                          *     value "iteration-1".
 658                          */
 659                         TEST_ASSERT(*value_ptr <= iteration,
 660                                     "Clear page %"PRIu64" value %"PRIu64
 661                                     " incorrect (iteration=%"PRIu64")",
 662                                     page, *value_ptr, iteration);
 663                         if (*value_ptr == iteration) {
 664                                 /*
 665                                  * This page is _just_ modified; it
 666                                  * should report its dirtyness in the
 667                                  * next run
 668                                  */
 669                                 set_bit_le(page, host_bmap_track);
 670                         }
 671                 }
 672         }
 673 }
 674
 675 static struct kvm_vm *create_vm(enum vm_guest_mode mode, uint32_t vcpuid,
 676                                 uint64_t extra_mem_pages, void *guest_code)
 677 {
 678         struct kvm_vm *vm;
 679         uint64_t extra_pg_pages = extra_mem_pages / 512 * 2;
 680
 681         pr_info("Testing guest mode: %s\n", vm_guest_mode_string(mode));
 682
 683         vm = vm_create(mode, DEFAULT_GUEST_PHY_PAGES + extra_pg_pages, O_RDWR);
 684         kvm_vm_elf_load(vm, program_invocation_name, 0, 0);
 685 #ifdef __x86_64__
 686         vm_create_irqchip(vm);
 687 #endif
 688         log_mode_create_vm_done(vm);
 689         vm_vcpu_add_default(vm, vcpuid, guest_code);
 690         return vm;
 691 }
 692
 693 #define DIRTY_MEM_BITS 30 /* 1G */
 694 #define PAGE_SHIFT_4K  12
 695
 696 struct test_params {
 697         unsigned long iterations;
 698         unsigned long interval;
 699         uint64_t phys_offset;
 700 };
 701
 702 static void run_test(enum vm_guest_mode mode, void *arg)
 703 {
 704         struct test_params *p = arg;
 705         struct kvm_vm *vm;
 706         unsigned long *bmap;
 707
 708         if (!log_mode_supported()) {
 709                 print_skip("Log mode '%s' not supported",
 710                            log_modes[host_log_mode].name);
 711                 return;
 712         }
 713
 714         /*
 715          * We reserve page table for 2 times of extra dirty mem which
 716          * will definitely cover the original (1G+) test range.  Here
 717          * we do the calculation with 4K page size which is the
 718          * smallest so the page number will be enough for all archs
 719          * (e.g., 64K page size guest will need even less memory for
 720          * page tables).
 721          */
 722         vm = create_vm(mode, VCPU_ID,
 723                        2ul << (DIRTY_MEM_BITS - PAGE_SHIFT_4K),
 724                        guest_code);
 725
 726         guest_page_size = vm_get_page_size(vm);
 727         /*
 728          * A little more than 1G of guest page sized pages.  Cover the
 729          * case where the size is not aligned to 64 pages.
 730          */
 731         guest_num_pages = (1ul << (DIRTY_MEM_BITS -
 732                                    vm_get_page_shift(vm))) + 3;
 733         guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
 734
 735         host_page_size = getpagesize();
 736         host_num_pages = vm_num_host_pages(mode, guest_num_pages);
 737
 738         if (!p->phys_offset) {
 739                 guest_test_phys_mem = (vm_get_max_gfn(vm) -
 740                                        guest_num_pages) * guest_page_size;
 741                 guest_test_phys_mem &= ~(host_page_size - 1);
 742         } else {
 743                 guest_test_phys_mem = p->phys_offset;
 744         }
 745
 746 #ifdef __s390x__
 747         /* Align to 1M (segment size) */
 748         guest_test_phys_mem &= ~((1 << 20) - 1);
 749 #endif
 750
 751         pr_info("guest physical test memory offset: 0x%lx\n", guest_test_phys_mem);
 752
 753         bmap = bitmap_alloc(host_num_pages);
 754         host_bmap_track = bitmap_alloc(host_num_pages);
 755
 756         /* Add an extra memory slot for testing dirty logging */
 757         vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS,
 758                                     guest_test_phys_mem,
 759                                     TEST_MEM_SLOT_INDEX,
 760                                     guest_num_pages,
 761                                     KVM_MEM_LOG_DIRTY_PAGES);
 762
 763         /* Do mapping for the dirty track memory slot */
 764         virt_map(vm, guest_test_virt_mem, guest_test_phys_mem, guest_num_pages, 0);
 765
 766         /* Cache the HVA pointer of the region */
 767         host_test_mem = addr_gpa2hva(vm, (vm_paddr_t)guest_test_phys_mem);
 768
 769         ucall_init(vm, NULL);
 770
 771         /* Export the shared variables to the guest */
 772         sync_global_to_guest(vm, host_page_size);
 773         sync_global_to_guest(vm, guest_page_size);
 774         sync_global_to_guest(vm, guest_test_virt_mem);
 775         sync_global_to_guest(vm, guest_num_pages);
 776
 777         /* Start the iterations */
 778         iteration = 1;
 779         sync_global_to_guest(vm, iteration);
 780         host_quit = false;
 781         host_dirty_count = 0;
 782         host_clear_count = 0;
 783         host_track_next_count = 0;
 784
 785         pthread_create(&vcpu_thread, NULL, vcpu_worker, vm);
 786
 787         while (iteration < p->iterations) {
 788                 /* Give the vcpu thread some time to dirty some pages */
 789                 usleep(p->interval * 1000);
 790                 log_mode_collect_dirty_pages(vm, TEST_MEM_SLOT_INDEX,
 791                                              bmap, host_num_pages);
 792
 793                 /*
 794                  * See vcpu_sync_stop_requested definition for details on why
 795                  * we need to stop vcpu when verify data.
 796                  */
 797                 atomic_set(&vcpu_sync_stop_requested, true);
 798                 sem_wait_until(&sem_vcpu_stop);
 799                 /*
 800                  * NOTE: for dirty ring, it's possible that we didn't stop at
 801                  * GUEST_SYNC but instead we stopped because ring is full;
 802                  * that's okay too because ring full means we're only missing
 803                  * the flush of the last page, and since we handle the last
 804                  * page specially verification will succeed anyway.
 805                  */
 806                 assert(host_log_mode == LOG_MODE_DIRTY_RING ||
 807                        atomic_read(&vcpu_sync_stop_requested) == false);
 808                 vm_dirty_log_verify(mode, bmap);
 809                 sem_post(&sem_vcpu_cont);
 810
 811                 iteration++;
 812                 sync_global_to_guest(vm, iteration);
 813         }
 814
 815         /* Tell the vcpu thread to quit */
 816         host_quit = true;
 817         log_mode_before_vcpu_join();
 818         pthread_join(vcpu_thread, NULL);
 819
 820         pr_info("Total bits checked: dirty (%"PRIu64"), clear (%"PRIu64"), "
 821                 "track_next (%"PRIu64")\n", host_dirty_count, host_clear_count,
 822                 host_track_next_count);
 823
 824         free(bmap);
 825         free(host_bmap_track);
 826         ucall_uninit(vm);
 827         kvm_vm_free(vm);
 828 }
 829
 830 static void help(char *name)
 831 {
 832         puts("");
 833         printf("usage: %s [-h] [-i iterations] [-I interval] "
 834                "[-p offset] [-m mode]\n", name);
 835         puts("");
 836         printf(" -c: specify dirty ring size, in number of entries\n");
 837         printf("     (only useful for dirty-ring test; default: %"PRIu32")\n",
 838                TEST_DIRTY_RING_COUNT);
 839         printf(" -i: specify iteration counts (default: %"PRIu64")\n",
 840                TEST_HOST_LOOP_N);
 841         printf(" -I: specify interval in ms (default: %"PRIu64" ms)\n",
 842                TEST_HOST_LOOP_INTERVAL);
 843         printf(" -p: specify guest physical test memory offset\n"
 844                "     Warning: a low offset can conflict with the loaded test code.\n");
 845         printf(" -M: specify the host logging mode "
 846                "(default: run all log modes).  Supported modes: \n\t");
 847         log_modes_dump();
 848         guest_modes_help();
 849         puts("");
 850         exit(0);
 851 }
 852
 853 int main(int argc, char *argv[])
 854 {
 855         struct test_params p = {
 856                 .iterations = TEST_HOST_LOOP_N,
 857                 .interval = TEST_HOST_LOOP_INTERVAL,
 858         };
 859         int opt, i;
 860         sigset_t sigset;
 861
 862         sem_init(&sem_vcpu_stop, 0, 0);
 863         sem_init(&sem_vcpu_cont, 0, 0);
 864
 865         guest_modes_append_default();
 866
 867         while ((opt = getopt(argc, argv, "c:hi:I:p:m:M:")) != -1) {
 868                 switch (opt) {
 869                 case 'c':
 870                         test_dirty_ring_count = strtol(optarg, NULL, 10);
 871                         break;
 872                 case 'i':
 873                         p.iterations = strtol(optarg, NULL, 10);
 874                         break;
 875                 case 'I':
 876                         p.interval = strtol(optarg, NULL, 10);
 877                         break;
 878                 case 'p':
 879                         p.phys_offset = strtoull(optarg, NULL, 0);
 880                         break;
 881                 case 'm':
 882                         guest_modes_cmdline(optarg);
 883                         break;
 884                 case 'M':
 885                         if (!strcmp(optarg, "all")) {
 886                                 host_log_mode_option = LOG_MODE_ALL;
 887                                 break;
 888                         }
 889                         for (i = 0; i < LOG_MODE_NUM; i++) {
 890                                 if (!strcmp(optarg, log_modes[i].name)) {
 891                                         pr_info("Setting log mode to: '%s'\n",
 892                                                 optarg);
 893                                         host_log_mode_option = i;
 894                                         break;
 895                                 }
 896                         }
 897                         if (i == LOG_MODE_NUM) {
 898                                 printf("Log mode '%s' invalid. Please choose "
 899                                        "from: ", optarg);
 900                                 log_modes_dump();
 901                                 exit(1);
 902                         }
 903                         break;
 904                 case 'h':
 905                 default:
 906                         help(argv[0]);
 907                         break;
 908                 }
 909         }
 910
 911         TEST_ASSERT(p.iterations > 2, "Iterations must be greater than two");
 912         TEST_ASSERT(p.interval > 0, "Interval must be greater than zero");
 913
 914         pr_info("Test iterations: %"PRIu64", interval: %"PRIu64" (ms)\n",
 915                 p.iterations, p.interval);
 916
 917         srandom(time(0));
 918
 919         /* Ensure that vCPU threads start with SIG_IPI blocked.  */
 920         sigemptyset(&sigset);
 921         sigaddset(&sigset, SIG_IPI);
 922         pthread_sigmask(SIG_BLOCK, &sigset, NULL);
 923
 924         if (host_log_mode_option == LOG_MODE_ALL) {
 925                 /* Run each log mode */
 926                 for (i = 0; i < LOG_MODE_NUM; i++) {
 927                         pr_info("Testing Log Mode '%s'\n", log_modes[i].name);
 928                         host_log_mode = i;
 929                         for_each_guest_mode(run_test, &p);
 930                 }
 931         } else {
 932                 host_log_mode = host_log_mode_option;
 933                 for_each_guest_mode(run_test, &p);
 934         }
 935
 936         return 0;
 937 }