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 <semaphore.h>
  16 #include <sys/types.h>
  17 #include <signal.h>
  18 #include <errno.h>
  19 #include <linux/bitmap.h>
  20 #include <linux/bitops.h>
  21 #include <asm/barrier.h>
  22
  23 #include "test_util.h"
  24 #include "kvm_util.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 dirty_ring_vcpu_stop;
 142 static sem_t dirty_ring_vcpu_cont;
 143 /*
 144  * This is updated by the vcpu thread to tell the host whether it's a
 145  * ring-full event.  It should only be read until a sem_wait() of
 146  * dirty_ring_vcpu_stop and before vcpu continues to run.
 147  */
 148 static bool dirty_ring_vcpu_ring_full;
 149 /*
 150  * This is only used for verifying the dirty pages.  Dirty ring has a very
 151  * tricky case when the ring just got full, kvm will do userspace exit due to
 152  * ring full.  When that happens, the very last PFN is set but actually the
 153  * data is not changed (the guest WRITE is not really applied yet), because
 154  * we found that the dirty ring is full, refused to continue the vcpu, and
 155  * recorded the dirty gfn with the old contents.
 156  *
 157  * For this specific case, it's safe to skip checking this pfn for this
 158  * bit, because it's a redundant bit, and when the write happens later the bit
 159  * will be set again.  We use this variable to always keep track of the latest
 160  * dirty gfn we've collected, so that if a mismatch of data found later in the
 161  * verifying process, we let it pass.
 162  */
 163 static uint64_t dirty_ring_last_page;
 164
 165 enum log_mode_t {
 166         /* Only use KVM_GET_DIRTY_LOG for logging */
 167         LOG_MODE_DIRTY_LOG = 0,
 168
 169         /* Use both KVM_[GET|CLEAR]_DIRTY_LOG for logging */
 170         LOG_MODE_CLEAR_LOG = 1,
 171
 172         /* Use dirty ring for logging */
 173         LOG_MODE_DIRTY_RING = 2,
 174
 175         LOG_MODE_NUM,
 176
 177         /* Run all supported modes */
 178         LOG_MODE_ALL = LOG_MODE_NUM,
 179 };
 180
 181 /* Mode of logging to test.  Default is to run all supported modes */
 182 static enum log_mode_t host_log_mode_option = LOG_MODE_ALL;
 183 /* Logging mode for current run */
 184 static enum log_mode_t host_log_mode;
 185 static pthread_t vcpu_thread;
 186
 187 static void vcpu_kick(void)
 188 {
 189         pthread_kill(vcpu_thread, SIG_IPI);
 190 }
 191
 192 /*
 193  * In our test we do signal tricks, let's use a better version of
 194  * sem_wait to avoid signal interrupts
 195  */
 196 static void sem_wait_until(sem_t *sem)
 197 {
 198         int ret;
 199
 200         do
 201                 ret = sem_wait(sem);
 202         while (ret == -1 && errno == EINTR);
 203 }
 204
 205 static bool clear_log_supported(void)
 206 {
 207         return kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
 208 }
 209
 210 static void clear_log_create_vm_done(struct kvm_vm *vm)
 211 {
 212         struct kvm_enable_cap cap = {};
 213         u64 manual_caps;
 214
 215         manual_caps = kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
 216         TEST_ASSERT(manual_caps, "MANUAL_CAPS is zero!");
 217         manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
 218                         KVM_DIRTY_LOG_INITIALLY_SET);
 219         cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
 220         cap.args[0] = manual_caps;
 221         vm_enable_cap(vm, &cap);
 222 }
 223
 224 static void dirty_log_collect_dirty_pages(struct kvm_vm *vm, int slot,
 225                                           void *bitmap, uint32_t num_pages)
 226 {
 227         kvm_vm_get_dirty_log(vm, slot, bitmap);
 228 }
 229
 230 static void clear_log_collect_dirty_pages(struct kvm_vm *vm, int slot,
 231                                           void *bitmap, uint32_t num_pages)
 232 {
 233         kvm_vm_get_dirty_log(vm, slot, bitmap);
 234         kvm_vm_clear_dirty_log(vm, slot, bitmap, 0, num_pages);
 235 }
 236
 237 static void default_after_vcpu_run(struct kvm_vm *vm, int ret, int err)
 238 {
 239         struct kvm_run *run = vcpu_state(vm, VCPU_ID);
 240
 241         TEST_ASSERT(ret == 0 || (ret == -1 && err == EINTR),
 242                     "vcpu run failed: errno=%d", err);
 243
 244         TEST_ASSERT(get_ucall(vm, VCPU_ID, NULL) == UCALL_SYNC,
 245                     "Invalid guest sync status: exit_reason=%s\n",
 246                     exit_reason_str(run->exit_reason));
 247 }
 248
 249 static bool dirty_ring_supported(void)
 250 {
 251         return kvm_check_cap(KVM_CAP_DIRTY_LOG_RING);
 252 }
 253
 254 static void dirty_ring_create_vm_done(struct kvm_vm *vm)
 255 {
 256         /*
 257          * Switch to dirty ring mode after VM creation but before any
 258          * of the vcpu creation.
 259          */
 260         vm_enable_dirty_ring(vm, TEST_DIRTY_RING_COUNT *
 261                              sizeof(struct kvm_dirty_gfn));
 262 }
 263
 264 static inline bool dirty_gfn_is_dirtied(struct kvm_dirty_gfn *gfn)
 265 {
 266         return gfn->flags == KVM_DIRTY_GFN_F_DIRTY;
 267 }
 268
 269 static inline void dirty_gfn_set_collected(struct kvm_dirty_gfn *gfn)
 270 {
 271         gfn->flags = KVM_DIRTY_GFN_F_RESET;
 272 }
 273
 274 static uint32_t dirty_ring_collect_one(struct kvm_dirty_gfn *dirty_gfns,
 275                                        int slot, void *bitmap,
 276                                        uint32_t num_pages, uint32_t *fetch_index)
 277 {
 278         struct kvm_dirty_gfn *cur;
 279         uint32_t count = 0;
 280
 281         while (true) {
 282                 cur = &dirty_gfns[*fetch_index % TEST_DIRTY_RING_COUNT];
 283                 if (!dirty_gfn_is_dirtied(cur))
 284                         break;
 285                 TEST_ASSERT(cur->slot == slot, "Slot number didn't match: "
 286                             "%u != %u", cur->slot, slot);
 287                 TEST_ASSERT(cur->offset < num_pages, "Offset overflow: "
 288                             "0x%llx >= 0x%x", cur->offset, num_pages);
 289                 //pr_info("fetch 0x%x page %llu\n", *fetch_index, cur->offset);
 290                 set_bit_le(cur->offset, bitmap);
 291                 dirty_ring_last_page = cur->offset;
 292                 dirty_gfn_set_collected(cur);
 293                 (*fetch_index)++;
 294                 count++;
 295         }
 296
 297         return count;
 298 }
 299
 300 static void dirty_ring_wait_vcpu(void)
 301 {
 302         /* This makes sure that hardware PML cache flushed */
 303         vcpu_kick();
 304         sem_wait_until(&dirty_ring_vcpu_stop);
 305 }
 306
 307 static void dirty_ring_continue_vcpu(void)
 308 {
 309         pr_info("Notifying vcpu to continue\n");
 310         sem_post(&dirty_ring_vcpu_cont);
 311 }
 312
 313 static void dirty_ring_collect_dirty_pages(struct kvm_vm *vm, int slot,
 314                                            void *bitmap, uint32_t num_pages)
 315 {
 316         /* We only have one vcpu */
 317         static uint32_t fetch_index = 0;
 318         uint32_t count = 0, cleared;
 319         bool continued_vcpu = false;
 320
 321         dirty_ring_wait_vcpu();
 322
 323         if (!dirty_ring_vcpu_ring_full) {
 324                 /*
 325                  * This is not a ring-full event, it's safe to allow
 326                  * vcpu to continue
 327                  */
 328                 dirty_ring_continue_vcpu();
 329                 continued_vcpu = true;
 330         }
 331
 332         /* Only have one vcpu */
 333         count = dirty_ring_collect_one(vcpu_map_dirty_ring(vm, VCPU_ID),
 334                                        slot, bitmap, num_pages, &fetch_index);
 335
 336         cleared = kvm_vm_reset_dirty_ring(vm);
 337
 338         /* Cleared pages should be the same as collected */
 339         TEST_ASSERT(cleared == count, "Reset dirty pages (%u) mismatch "
 340                     "with collected (%u)", cleared, count);
 341
 342         if (!continued_vcpu) {
 343                 TEST_ASSERT(dirty_ring_vcpu_ring_full,
 344                             "Didn't continue vcpu even without ring full");
 345                 dirty_ring_continue_vcpu();
 346         }
 347
 348         pr_info("Iteration %ld collected %u pages\n", iteration, count);
 349 }
 350
 351 static void dirty_ring_after_vcpu_run(struct kvm_vm *vm, int ret, int err)
 352 {
 353         struct kvm_run *run = vcpu_state(vm, VCPU_ID);
 354
 355         /* A ucall-sync or ring-full event is allowed */
 356         if (get_ucall(vm, VCPU_ID, NULL) == UCALL_SYNC) {
 357                 /* We should allow this to continue */
 358                 ;
 359         } else if (run->exit_reason == KVM_EXIT_DIRTY_RING_FULL ||
 360                    (ret == -1 && err == EINTR)) {
 361                 /* Update the flag first before pause */
 362                 WRITE_ONCE(dirty_ring_vcpu_ring_full,
 363                            run->exit_reason == KVM_EXIT_DIRTY_RING_FULL);
 364                 sem_post(&dirty_ring_vcpu_stop);
 365                 pr_info("vcpu stops because %s...\n",
 366                         dirty_ring_vcpu_ring_full ?
 367                         "dirty ring is full" : "vcpu is kicked out");
 368                 sem_wait_until(&dirty_ring_vcpu_cont);
 369                 pr_info("vcpu continues now.\n");
 370         } else {
 371                 TEST_ASSERT(false, "Invalid guest sync status: "
 372                             "exit_reason=%s\n",
 373                             exit_reason_str(run->exit_reason));
 374         }
 375 }
 376
 377 static void dirty_ring_before_vcpu_join(void)
 378 {
 379         /* Kick another round of vcpu just to make sure it will quit */
 380         sem_post(&dirty_ring_vcpu_cont);
 381 }
 382
 383 struct log_mode {
 384         const char *name;
 385         /* Return true if this mode is supported, otherwise false */
 386         bool (*supported)(void);
 387         /* Hook when the vm creation is done (before vcpu creation) */
 388         void (*create_vm_done)(struct kvm_vm *vm);
 389         /* Hook to collect the dirty pages into the bitmap provided */
 390         void (*collect_dirty_pages) (struct kvm_vm *vm, int slot,
 391                                      void *bitmap, uint32_t num_pages);
 392         /* Hook to call when after each vcpu run */
 393         void (*after_vcpu_run)(struct kvm_vm *vm, int ret, int err);
 394         void (*before_vcpu_join) (void);
 395 } log_modes[LOG_MODE_NUM] = {
 396         {
 397                 .name = "dirty-log",
 398                 .collect_dirty_pages = dirty_log_collect_dirty_pages,
 399                 .after_vcpu_run = default_after_vcpu_run,
 400         },
 401         {
 402                 .name = "clear-log",
 403                 .supported = clear_log_supported,
 404                 .create_vm_done = clear_log_create_vm_done,
 405                 .collect_dirty_pages = clear_log_collect_dirty_pages,
 406                 .after_vcpu_run = default_after_vcpu_run,
 407         },
 408         {
 409                 .name = "dirty-ring",
 410                 .supported = dirty_ring_supported,
 411                 .create_vm_done = dirty_ring_create_vm_done,
 412                 .collect_dirty_pages = dirty_ring_collect_dirty_pages,
 413                 .before_vcpu_join = dirty_ring_before_vcpu_join,
 414                 .after_vcpu_run = dirty_ring_after_vcpu_run,
 415         },
 416 };
 417
 418 /*
 419  * We use this bitmap to track some pages that should have its dirty
 420  * bit set in the _next_ iteration.  For example, if we detected the
 421  * page value changed to current iteration but at the same time the
 422  * page bit is cleared in the latest bitmap, then the system must
 423  * report that write in the next get dirty log call.
 424  */
 425 static unsigned long *host_bmap_track;
 426
 427 static void log_modes_dump(void)
 428 {
 429         int i;
 430
 431         printf("all");
 432         for (i = 0; i < LOG_MODE_NUM; i++)
 433                 printf(", %s", log_modes[i].name);
 434         printf("\n");
 435 }
 436
 437 static bool log_mode_supported(void)
 438 {
 439         struct log_mode *mode = &log_modes[host_log_mode];
 440
 441         if (mode->supported)
 442                 return mode->supported();
 443
 444         return true;
 445 }
 446
 447 static void log_mode_create_vm_done(struct kvm_vm *vm)
 448 {
 449         struct log_mode *mode = &log_modes[host_log_mode];
 450
 451         if (mode->create_vm_done)
 452                 mode->create_vm_done(vm);
 453 }
 454
 455 static void log_mode_collect_dirty_pages(struct kvm_vm *vm, int slot,
 456                                          void *bitmap, uint32_t num_pages)
 457 {
 458         struct log_mode *mode = &log_modes[host_log_mode];
 459
 460         TEST_ASSERT(mode->collect_dirty_pages != NULL,
 461                     "collect_dirty_pages() is required for any log mode!");
 462         mode->collect_dirty_pages(vm, slot, bitmap, num_pages);
 463 }
 464
 465 static void log_mode_after_vcpu_run(struct kvm_vm *vm, int ret, int err)
 466 {
 467         struct log_mode *mode = &log_modes[host_log_mode];
 468
 469         if (mode->after_vcpu_run)
 470                 mode->after_vcpu_run(vm, ret, err);
 471 }
 472
 473 static void log_mode_before_vcpu_join(void)
 474 {
 475         struct log_mode *mode = &log_modes[host_log_mode];
 476
 477         if (mode->before_vcpu_join)
 478                 mode->before_vcpu_join();
 479 }
 480
 481 static void generate_random_array(uint64_t *guest_array, uint64_t size)
 482 {
 483         uint64_t i;
 484
 485         for (i = 0; i < size; i++)
 486                 guest_array[i] = random();
 487 }
 488
 489 static void *vcpu_worker(void *data)
 490 {
 491         int ret, vcpu_fd;
 492         struct kvm_vm *vm = data;
 493         uint64_t *guest_array;
 494         uint64_t pages_count = 0;
 495         struct kvm_signal_mask *sigmask = alloca(offsetof(struct kvm_signal_mask, sigset)
 496                                                  + sizeof(sigset_t));
 497         sigset_t *sigset = (sigset_t *) &sigmask->sigset;
 498
 499         vcpu_fd = vcpu_get_fd(vm, VCPU_ID);
 500
 501         /*
 502          * SIG_IPI is unblocked atomically while in KVM_RUN.  It causes the
 503          * ioctl to return with -EINTR, but it is still pending and we need
 504          * to accept it with the sigwait.
 505          */
 506         sigmask->len = 8;
 507         pthread_sigmask(0, NULL, sigset);
 508         vcpu_ioctl(vm, VCPU_ID, KVM_SET_SIGNAL_MASK, sigmask);
 509         sigaddset(sigset, SIG_IPI);
 510         pthread_sigmask(SIG_BLOCK, sigset, NULL);
 511
 512         sigemptyset(sigset);
 513         sigaddset(sigset, SIG_IPI);
 514
 515         guest_array = addr_gva2hva(vm, (vm_vaddr_t)random_array);
 516
 517         while (!READ_ONCE(host_quit)) {
 518                 /* Clear any existing kick signals */
 519                 generate_random_array(guest_array, TEST_PAGES_PER_LOOP);
 520                 pages_count += TEST_PAGES_PER_LOOP;
 521                 /* Let the guest dirty the random pages */
 522                 ret = ioctl(vcpu_fd, KVM_RUN, NULL);
 523                 if (ret == -1 && errno == EINTR) {
 524                         int sig = -1;
 525                         sigwait(sigset, &sig);
 526                         assert(sig == SIG_IPI);
 527                 }
 528                 log_mode_after_vcpu_run(vm, ret, errno);
 529         }
 530
 531         pr_info("Dirtied %"PRIu64" pages\n", pages_count);
 532
 533         return NULL;
 534 }
 535
 536 static void vm_dirty_log_verify(enum vm_guest_mode mode, unsigned long *bmap)
 537 {
 538         uint64_t step = vm_num_host_pages(mode, 1);
 539         uint64_t page;
 540         uint64_t *value_ptr;
 541         uint64_t min_iter = 0;
 542
 543         for (page = 0; page < host_num_pages; page += step) {
 544                 value_ptr = host_test_mem + page * host_page_size;
 545
 546                 /* If this is a special page that we were tracking... */
 547                 if (test_and_clear_bit_le(page, host_bmap_track)) {
 548                         host_track_next_count++;
 549                         TEST_ASSERT(test_bit_le(page, bmap),
 550                                     "Page %"PRIu64" should have its dirty bit "
 551                                     "set in this iteration but it is missing",
 552                                     page);
 553                 }
 554
 555                 if (test_and_clear_bit_le(page, bmap)) {
 556                         bool matched;
 557
 558                         host_dirty_count++;
 559
 560                         /*
 561                          * If the bit is set, the value written onto
 562                          * the corresponding page should be either the
 563                          * previous iteration number or the current one.
 564                          */
 565                         matched = (*value_ptr == iteration ||
 566                                    *value_ptr == iteration - 1);
 567
 568                         if (host_log_mode == LOG_MODE_DIRTY_RING && !matched) {
 569                                 if (*value_ptr == iteration - 2 && min_iter <= iteration - 2) {
 570                                         /*
 571                                          * Short answer: this case is special
 572                                          * only for dirty ring test where the
 573                                          * page is the last page before a kvm
 574                                          * dirty ring full in iteration N-2.
 575                                          *
 576                                          * Long answer: Assuming ring size R,
 577                                          * one possible condition is:
 578                                          *
 579                                          *      main thr       vcpu thr
 580                                          *      --------       --------
 581                                          *    iter=1
 582                                          *                   write 1 to page 0~(R-1)
 583                                          *                   full, vmexit
 584                                          *    collect 0~(R-1)
 585                                          *    kick vcpu
 586                                          *                   write 1 to (R-1)~(2R-2)
 587                                          *                   full, vmexit
 588                                          *    iter=2
 589                                          *    collect (R-1)~(2R-2)
 590                                          *    kick vcpu
 591                                          *                   write 1 to (2R-2)
 592                                          *                   (NOTE!!! "1" cached in cpu reg)
 593                                          *                   write 2 to (2R-1)~(3R-3)
 594                                          *                   full, vmexit
 595                                          *    iter=3
 596                                          *    collect (2R-2)~(3R-3)
 597                                          *    (here if we read value on page
 598                                          *     "2R-2" is 1, while iter=3!!!)
 599                                          *
 600                                          * This however can only happen once per iteration.
 601                                          */
 602                                         min_iter = iteration - 1;
 603                                         continue;
 604                                 } else if (page == dirty_ring_last_page) {
 605                                         /*
 606                                          * Please refer to comments in
 607                                          * dirty_ring_last_page.
 608                                          */
 609                                         continue;
 610                                 }
 611                         }
 612
 613                         TEST_ASSERT(matched,
 614                                     "Set page %"PRIu64" value %"PRIu64
 615                                     " incorrect (iteration=%"PRIu64")",
 616                                     page, *value_ptr, iteration);
 617                 } else {
 618                         host_clear_count++;
 619                         /*
 620                          * If cleared, the value written can be any
 621                          * value smaller or equals to the iteration
 622                          * number.  Note that the value can be exactly
 623                          * (iteration-1) if that write can happen
 624                          * like this:
 625                          *
 626                          * (1) increase loop count to "iteration-1"
 627                          * (2) write to page P happens (with value
 628                          *     "iteration-1")
 629                          * (3) get dirty log for "iteration-1"; we'll
 630                          *     see that page P bit is set (dirtied),
 631                          *     and not set the bit in host_bmap_track
 632                          * (4) increase loop count to "iteration"
 633                          *     (which is current iteration)
 634                          * (5) get dirty log for current iteration,
 635                          *     we'll see that page P is cleared, with
 636                          *     value "iteration-1".
 637                          */
 638                         TEST_ASSERT(*value_ptr <= iteration,
 639                                     "Clear page %"PRIu64" value %"PRIu64
 640                                     " incorrect (iteration=%"PRIu64")",
 641                                     page, *value_ptr, iteration);
 642                         if (*value_ptr == iteration) {
 643                                 /*
 644                                  * This page is _just_ modified; it
 645                                  * should report its dirtyness in the
 646                                  * next run
 647                                  */
 648                                 set_bit_le(page, host_bmap_track);
 649                         }
 650                 }
 651         }
 652 }
 653
 654 static struct kvm_vm *create_vm(enum vm_guest_mode mode, uint32_t vcpuid,
 655                                 uint64_t extra_mem_pages, void *guest_code)
 656 {
 657         struct kvm_vm *vm;
 658         uint64_t extra_pg_pages = extra_mem_pages / 512 * 2;
 659
 660         pr_info("Testing guest mode: %s\n", vm_guest_mode_string(mode));
 661
 662         vm = vm_create(mode, DEFAULT_GUEST_PHY_PAGES + extra_pg_pages, O_RDWR);
 663         kvm_vm_elf_load(vm, program_invocation_name, 0, 0);
 664 #ifdef __x86_64__
 665         vm_create_irqchip(vm);
 666 #endif
 667         log_mode_create_vm_done(vm);
 668         vm_vcpu_add_default(vm, vcpuid, guest_code);
 669         return vm;
 670 }
 671
 672 #define DIRTY_MEM_BITS 30 /* 1G */
 673 #define PAGE_SHIFT_4K  12
 674
 675 static void run_test(enum vm_guest_mode mode, unsigned long iterations,
 676                      unsigned long interval, uint64_t phys_offset)
 677 {
 678         struct kvm_vm *vm;
 679         unsigned long *bmap;
 680
 681         if (!log_mode_supported()) {
 682                 print_skip("Log mode '%s' not supported",
 683                            log_modes[host_log_mode].name);
 684                 return;
 685         }
 686
 687         /*
 688          * We reserve page table for 2 times of extra dirty mem which
 689          * will definitely cover the original (1G+) test range.  Here
 690          * we do the calculation with 4K page size which is the
 691          * smallest so the page number will be enough for all archs
 692          * (e.g., 64K page size guest will need even less memory for
 693          * page tables).
 694          */
 695         vm = create_vm(mode, VCPU_ID,
 696                        2ul << (DIRTY_MEM_BITS - PAGE_SHIFT_4K),
 697                        guest_code);
 698
 699         guest_page_size = vm_get_page_size(vm);
 700         /*
 701          * A little more than 1G of guest page sized pages.  Cover the
 702          * case where the size is not aligned to 64 pages.
 703          */
 704         guest_num_pages = (1ul << (DIRTY_MEM_BITS -
 705                                    vm_get_page_shift(vm))) + 3;
 706         guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
 707
 708         host_page_size = getpagesize();
 709         host_num_pages = vm_num_host_pages(mode, guest_num_pages);
 710
 711         if (!phys_offset) {
 712                 guest_test_phys_mem = (vm_get_max_gfn(vm) -
 713                                        guest_num_pages) * guest_page_size;
 714                 guest_test_phys_mem &= ~(host_page_size - 1);
 715         } else {
 716                 guest_test_phys_mem = phys_offset;
 717         }
 718
 719 #ifdef __s390x__
 720         /* Align to 1M (segment size) */
 721         guest_test_phys_mem &= ~((1 << 20) - 1);
 722 #endif
 723
 724         pr_info("guest physical test memory offset: 0x%lx\n", guest_test_phys_mem);
 725
 726         bmap = bitmap_alloc(host_num_pages);
 727         host_bmap_track = bitmap_alloc(host_num_pages);
 728
 729         /* Add an extra memory slot for testing dirty logging */
 730         vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS,
 731                                     guest_test_phys_mem,
 732                                     TEST_MEM_SLOT_INDEX,
 733                                     guest_num_pages,
 734                                     KVM_MEM_LOG_DIRTY_PAGES);
 735
 736         /* Do mapping for the dirty track memory slot */
 737         virt_map(vm, guest_test_virt_mem, guest_test_phys_mem, guest_num_pages, 0);
 738
 739         /* Cache the HVA pointer of the region */
 740         host_test_mem = addr_gpa2hva(vm, (vm_paddr_t)guest_test_phys_mem);
 741
 742 #ifdef __x86_64__
 743         vcpu_set_cpuid(vm, VCPU_ID, kvm_get_supported_cpuid());
 744 #endif
 745         ucall_init(vm, NULL);
 746
 747         /* Export the shared variables to the guest */
 748         sync_global_to_guest(vm, host_page_size);
 749         sync_global_to_guest(vm, guest_page_size);
 750         sync_global_to_guest(vm, guest_test_virt_mem);
 751         sync_global_to_guest(vm, guest_num_pages);
 752
 753         /* Start the iterations */
 754         iteration = 1;
 755         sync_global_to_guest(vm, iteration);
 756         host_quit = false;
 757         host_dirty_count = 0;
 758         host_clear_count = 0;
 759         host_track_next_count = 0;
 760
 761         pthread_create(&vcpu_thread, NULL, vcpu_worker, vm);
 762
 763         while (iteration < iterations) {
 764                 /* Give the vcpu thread some time to dirty some pages */
 765                 usleep(interval * 1000);
 766                 log_mode_collect_dirty_pages(vm, TEST_MEM_SLOT_INDEX,
 767                                              bmap, host_num_pages);
 768                 vm_dirty_log_verify(mode, bmap);
 769                 iteration++;
 770                 sync_global_to_guest(vm, iteration);
 771         }
 772
 773         /* Tell the vcpu thread to quit */
 774         host_quit = true;
 775         log_mode_before_vcpu_join();
 776         pthread_join(vcpu_thread, NULL);
 777
 778         pr_info("Total bits checked: dirty (%"PRIu64"), clear (%"PRIu64"), "
 779                 "track_next (%"PRIu64")\n", host_dirty_count, host_clear_count,
 780                 host_track_next_count);
 781
 782         free(bmap);
 783         free(host_bmap_track);
 784         ucall_uninit(vm);
 785         kvm_vm_free(vm);
 786 }
 787
 788 struct guest_mode {
 789         bool supported;
 790         bool enabled;
 791 };
 792 static struct guest_mode guest_modes[NUM_VM_MODES];
 793
 794 #define guest_mode_init(mode, supported, enabled) ({ \
 795         guest_modes[mode] = (struct guest_mode){ supported, enabled }; \
 796 })
 797
 798 static void help(char *name)
 799 {
 800         int i;
 801
 802         puts("");
 803         printf("usage: %s [-h] [-i iterations] [-I interval] "
 804                "[-p offset] [-m mode]\n", name);
 805         puts("");
 806         printf(" -i: specify iteration counts (default: %"PRIu64")\n",
 807                TEST_HOST_LOOP_N);
 808         printf(" -I: specify interval in ms (default: %"PRIu64" ms)\n",
 809                TEST_HOST_LOOP_INTERVAL);
 810         printf(" -p: specify guest physical test memory offset\n"
 811                "     Warning: a low offset can conflict with the loaded test code.\n");
 812         printf(" -M: specify the host logging mode "
 813                "(default: run all log modes).  Supported modes: \n\t");
 814         log_modes_dump();
 815         printf(" -m: specify the guest mode ID to test "
 816                "(default: test all supported modes)\n"
 817                "     This option may be used multiple times.\n"
 818                "     Guest mode IDs:\n");
 819         for (i = 0; i < NUM_VM_MODES; ++i) {
 820                 printf("         %d:    %s%s\n", i, vm_guest_mode_string(i),
 821                        guest_modes[i].supported ? " (supported)" : "");
 822         }
 823         puts("");
 824         exit(0);
 825 }
 826
 827 int main(int argc, char *argv[])
 828 {
 829         unsigned long iterations = TEST_HOST_LOOP_N;
 830         unsigned long interval = TEST_HOST_LOOP_INTERVAL;
 831         bool mode_selected = false;
 832         uint64_t phys_offset = 0;
 833         unsigned int mode;
 834         int opt, i, j;
 835
 836         sem_init(&dirty_ring_vcpu_stop, 0, 0);
 837         sem_init(&dirty_ring_vcpu_cont, 0, 0);
 838
 839 #ifdef __x86_64__
 840         guest_mode_init(VM_MODE_PXXV48_4K, true, true);
 841 #endif
 842 #ifdef __aarch64__
 843         guest_mode_init(VM_MODE_P40V48_4K, true, true);
 844         guest_mode_init(VM_MODE_P40V48_64K, true, true);
 845
 846         {
 847                 unsigned int limit = kvm_check_cap(KVM_CAP_ARM_VM_IPA_SIZE);
 848
 849                 if (limit >= 52)
 850                         guest_mode_init(VM_MODE_P52V48_64K, true, true);
 851                 if (limit >= 48) {
 852                         guest_mode_init(VM_MODE_P48V48_4K, true, true);
 853                         guest_mode_init(VM_MODE_P48V48_64K, true, true);
 854                 }
 855         }
 856 #endif
 857 #ifdef __s390x__
 858         guest_mode_init(VM_MODE_P40V48_4K, true, true);
 859 #endif
 860
 861         while ((opt = getopt(argc, argv, "hi:I:p:m:M:")) != -1) {
 862                 switch (opt) {
 863                 case 'i':
 864                         iterations = strtol(optarg, NULL, 10);
 865                         break;
 866                 case 'I':
 867                         interval = strtol(optarg, NULL, 10);
 868                         break;
 869                 case 'p':
 870                         phys_offset = strtoull(optarg, NULL, 0);
 871                         break;
 872                 case 'm':
 873                         if (!mode_selected) {
 874                                 for (i = 0; i < NUM_VM_MODES; ++i)
 875                                         guest_modes[i].enabled = false;
 876                                 mode_selected = true;
 877                         }
 878                         mode = strtoul(optarg, NULL, 10);
 879                         TEST_ASSERT(mode < NUM_VM_MODES,
 880                                     "Guest mode ID %d too big", mode);
 881                         guest_modes[mode].enabled = true;
 882                         break;
 883                 case 'M':
 884                         if (!strcmp(optarg, "all")) {
 885                                 host_log_mode_option = LOG_MODE_ALL;
 886                                 break;
 887                         }
 888                         for (i = 0; i < LOG_MODE_NUM; i++) {
 889                                 if (!strcmp(optarg, log_modes[i].name)) {
 890                                         pr_info("Setting log mode to: '%s'\n",
 891                                                 optarg);
 892                                         host_log_mode_option = i;
 893                                         break;
 894                                 }
 895                         }
 896                         if (i == LOG_MODE_NUM) {
 897                                 printf("Log mode '%s' invalid. Please choose "
 898                                        "from: ", optarg);
 899                                 log_modes_dump();
 900                                 exit(1);
 901                         }
 902                         break;
 903                 case 'h':
 904                 default:
 905                         help(argv[0]);
 906                         break;
 907                 }
 908         }
 909
 910         TEST_ASSERT(iterations > 2, "Iterations must be greater than two");
 911         TEST_ASSERT(interval > 0, "Interval must be greater than zero");
 912
 913         pr_info("Test iterations: %"PRIu64", interval: %"PRIu64" (ms)\n",
 914                 iterations, interval);
 915
 916         srandom(time(0));
 917
 918         for (i = 0; i < NUM_VM_MODES; ++i) {
 919                 if (!guest_modes[i].enabled)
 920                         continue;
 921                 TEST_ASSERT(guest_modes[i].supported,
 922                             "Guest mode ID %d (%s) not supported.",
 923                             i, vm_guest_mode_string(i));
 924                 if (host_log_mode_option == LOG_MODE_ALL) {
 925                         /* Run each log mode */
 926                         for (j = 0; j < LOG_MODE_NUM; j++) {
 927                                 pr_info("Testing Log Mode '%s'\n",
 928                                         log_modes[j].name);
 929                                 host_log_mode = j;
 930                                 run_test(i, iterations, interval, phys_offset);
 931                         }
 932                 } else {
 933                         host_log_mode = host_log_mode_option;
 934                         run_test(i, iterations, interval, phys_offset);
 935                 }
 936         }
 937
 938         return 0;
 939 }