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