Quelle  dirty_log_perf_test.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/*
 * KVM dirty page logging performance test
 *
 * Based on dirty_log_test.c
 *
 * Copyright (C) 2018, Red Hat, Inc.
 * Copyright (C) 2020, Google, Inc.
 */

#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <pthread.h>
#include <linux/bitmap.h>

#include "kvm_util.h"
#include "test_util.h"
#include "memstress.h"
#include "guest_modes.h"
#include "ucall_common.h"

#ifdef __aarch64__
#include "arm64/vgic.h"

static int gic_fd;

static void arch_setup_vm(struct kvm_vm *vm, unsigned int nr_vcpus)
{
 /*
 * The test can still run even if hardware does not support GICv3, as it
 * is only an optimization to reduce guest exits.
 */
 gic_fd = vgic_v3_setup(vm, nr_vcpus, 64);
}

static void arch_cleanup_vm(struct kvm_vm *vm)
{
 if (gic_fd > 0)
  close(gic_fd);
}

#else /* __aarch64__ */

static void arch_setup_vm(struct kvm_vm *vm, unsigned int nr_vcpus)
{
}

static void arch_cleanup_vm(struct kvm_vm *vm)
{
}

#endif

/* How many host loops to run by default (one KVM_GET_DIRTY_LOG for each loop)*/
#define TEST_HOST_LOOP_N  2UL

static int nr_vcpus = 1;
static uint64_t guest_percpu_mem_size = DEFAULT_PER_VCPU_MEM_SIZE;
static bool run_vcpus_while_disabling_dirty_logging;

/* Host variables */
static u64 dirty_log_manual_caps;
static bool host_quit;
static int iteration;
static int vcpu_last_completed_iteration[KVM_MAX_VCPUS];

static void vcpu_worker(struct memstress_vcpu_args *vcpu_args)
{
 struct kvm_vcpu *vcpu = vcpu_args->vcpu;
 int vcpu_idx = vcpu_args->vcpu_idx;
 uint64_t pages_count = 0;
 struct kvm_run *run;
 struct timespec start;
 struct timespec ts_diff;
 struct timespec total = (struct timespec){0};
 struct timespec avg;
 int ret;

 run = vcpu->run;

 while (!READ_ONCE(host_quit)) {
  int current_iteration = READ_ONCE(iteration);

  clock_gettime(CLOCK_MONOTONIC, &start);
  ret = _vcpu_run(vcpu);
  ts_diff = timespec_elapsed(start);

  TEST_ASSERT(ret == 0, "vcpu_run failed: %d", ret);
  TEST_ASSERT(get_ucall(vcpu, NULL) == UCALL_SYNC,
       "Invalid guest sync status: exit_reason=%s",
       exit_reason_str(run->exit_reason));

  pr_debug("Got sync event from vCPU %d\n", vcpu_idx);
  vcpu_last_completed_iteration[vcpu_idx] = current_iteration;
  pr_debug("vCPU %d updated last completed iteration to %d\n",
    vcpu_idx, vcpu_last_completed_iteration[vcpu_idx]);

  if (current_iteration) {
   pages_count += vcpu_args->pages;
   total = timespec_add(total, ts_diff);
   pr_debug("vCPU %d iteration %d dirty memory time: %ld.%.9lds\n",
    vcpu_idx, current_iteration, ts_diff.tv_sec,
    ts_diff.tv_nsec);
  } else {
   pr_debug("vCPU %d iteration %d populate memory time: %ld.%.9lds\n",
    vcpu_idx, current_iteration, ts_diff.tv_sec,
    ts_diff.tv_nsec);
  }

  /*
 * Keep running the guest while dirty logging is being disabled
 * (iteration is negative) so that vCPUs are accessing memory
 * for the entire duration of zapping collapsible SPTEs.
 */
  while (current_iteration == READ_ONCE(iteration) &&
         READ_ONCE(iteration) >= 0 && !READ_ONCE(host_quit)) {}
 }

 avg = timespec_div(total, vcpu_last_completed_iteration[vcpu_idx]);
 pr_debug("\nvCPU %d dirtied 0x%lx pages over %d iterations in %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
  vcpu_idx, pages_count, vcpu_last_completed_iteration[vcpu_idx],
  total.tv_sec, total.tv_nsec, avg.tv_sec, avg.tv_nsec);
}

struct test_params {
 unsigned long iterations;
 uint64_t phys_offset;
 bool partition_vcpu_memory_access;
 enum vm_mem_backing_src_type backing_src;
 int slots;
 uint32_t write_percent;
 bool random_access;
};

static void run_test(enum vm_guest_mode mode, void *arg)
{
 struct test_params *p = arg;
 struct kvm_vm *vm;
 unsigned long **bitmaps;
 uint64_t guest_num_pages;
 uint64_t host_num_pages;
 uint64_t pages_per_slot;
 struct timespec start;
 struct timespec ts_diff;
 struct timespec get_dirty_log_total = (struct timespec){0};
 struct timespec vcpu_dirty_total = (struct timespec){0};
 struct timespec avg;
 struct timespec clear_dirty_log_total = (struct timespec){0};
 int i;

 vm = memstress_create_vm(mode, nr_vcpus, guest_percpu_mem_size,
     p->slots, p->backing_src,
     p->partition_vcpu_memory_access);

 memstress_set_write_percent(vm, p->write_percent);

 guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm->page_shift;
 guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
 host_num_pages = vm_num_host_pages(mode, guest_num_pages);
 pages_per_slot = host_num_pages / p->slots;

 bitmaps = memstress_alloc_bitmaps(p->slots, pages_per_slot);

 if (dirty_log_manual_caps)
  vm_enable_cap(vm, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2,
         dirty_log_manual_caps);

 arch_setup_vm(vm, nr_vcpus);

 /* Start the iterations */
 iteration = 0;
 host_quit = false;

 clock_gettime(CLOCK_MONOTONIC, &start);
 for (i = 0; i < nr_vcpus; i++)
  vcpu_last_completed_iteration[i] = -1;

 /*
 * Use 100% writes during the population phase to ensure all
 * memory is actually populated and not just mapped to the zero
 * page. The prevents expensive copy-on-write faults from
 * occurring during the dirty memory iterations below, which
 * would pollute the performance results.
 */
 memstress_set_write_percent(vm, 100);
 memstress_set_random_access(vm, false);
 memstress_start_vcpu_threads(nr_vcpus, vcpu_worker);

 /* Allow the vCPUs to populate memory */
 pr_debug("Starting iteration %d - Populating\n", iteration);
 for (i = 0; i < nr_vcpus; i++) {
  while (READ_ONCE(vcpu_last_completed_iteration[i]) !=
         iteration)
   ;
 }

 ts_diff = timespec_elapsed(start);
 pr_info("Populate memory time: %ld.%.9lds\n",
  ts_diff.tv_sec, ts_diff.tv_nsec);

 /* Enable dirty logging */
 clock_gettime(CLOCK_MONOTONIC, &start);
 memstress_enable_dirty_logging(vm, p->slots);
 ts_diff = timespec_elapsed(start);
 pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
  ts_diff.tv_sec, ts_diff.tv_nsec);

 memstress_set_write_percent(vm, p->write_percent);
 memstress_set_random_access(vm, p->random_access);

 while (iteration < p->iterations) {
  /*
 * Incrementing the iteration number will start the vCPUs
 * dirtying memory again.
 */
  clock_gettime(CLOCK_MONOTONIC, &start);
  iteration++;

  pr_debug("Starting iteration %d\n", iteration);
  for (i = 0; i < nr_vcpus; i++) {
   while (READ_ONCE(vcpu_last_completed_iteration[i])
          != iteration)
    ;
  }

  ts_diff = timespec_elapsed(start);
  vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
  pr_info("Iteration %d dirty memory time: %ld.%.9lds\n",
   iteration, ts_diff.tv_sec, ts_diff.tv_nsec);

  clock_gettime(CLOCK_MONOTONIC, &start);
  memstress_get_dirty_log(vm, bitmaps, p->slots);
  ts_diff = timespec_elapsed(start);
  get_dirty_log_total = timespec_add(get_dirty_log_total,
         ts_diff);
  pr_info("Iteration %d get dirty log time: %ld.%.9lds\n",
   iteration, ts_diff.tv_sec, ts_diff.tv_nsec);

  if (dirty_log_manual_caps) {
   clock_gettime(CLOCK_MONOTONIC, &start);
   memstress_clear_dirty_log(vm, bitmaps, p->slots,
        pages_per_slot);
   ts_diff = timespec_elapsed(start);
   clear_dirty_log_total = timespec_add(clear_dirty_log_total,
            ts_diff);
   pr_info("Iteration %d clear dirty log time: %ld.%.9lds\n",
    iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
  }
 }

 /*
 * Run vCPUs while dirty logging is being disabled to stress disabling
 * in terms of both performance and correctness.  Opt-in via command
 * line as this significantly increases time to disable dirty logging.
 */
 if (run_vcpus_while_disabling_dirty_logging)
  WRITE_ONCE(iteration, -1);

 /* Disable dirty logging */
 clock_gettime(CLOCK_MONOTONIC, &start);
 memstress_disable_dirty_logging(vm, p->slots);
 ts_diff = timespec_elapsed(start);
 pr_info("Disabling dirty logging time: %ld.%.9lds\n",
  ts_diff.tv_sec, ts_diff.tv_nsec);

 /*
 * Tell the vCPU threads to quit.  No need to manually check that vCPUs
 * have stopped running after disabling dirty logging, the join will
 * wait for them to exit.
 */
 host_quit = true;
 memstress_join_vcpu_threads(nr_vcpus);

 avg = timespec_div(get_dirty_log_total, p->iterations);
 pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
  p->iterations, get_dirty_log_total.tv_sec,
  get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);

 if (dirty_log_manual_caps) {
  avg = timespec_div(clear_dirty_log_total, p->iterations);
  pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
   p->iterations, clear_dirty_log_total.tv_sec,
   clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
 }

 memstress_free_bitmaps(bitmaps, p->slots);
 arch_cleanup_vm(vm);
 memstress_destroy_vm(vm);
}

static void help(char *name)
{
 puts("");
 printf("usage: %s [-h] [-a] [-i iterations] [-p offset] [-g] "
        "[-m mode] [-n] [-b vcpu bytes] [-v vcpus] [-o] [-r random seed ] [-s mem type]"
        "[-x memslots] [-w percentage] [-c physical cpus to run test on]\n", name);
 puts("");
 printf(" -a: access memory randomly rather than in order.\n");
 printf(" -i: specify iteration counts (default: %"PRIu64")\n",
        TEST_HOST_LOOP_N);
 printf(" -g: Do not enable KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2. This\n"
        " makes KVM_GET_DIRTY_LOG clear the dirty log (i.e.\n"
        " KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE is not enabled)\n"
        " and writes will be tracked as soon as dirty logging is\n"
        " enabled on the memslot (i.e. KVM_DIRTY_LOG_INITIALLY_SET\n"
        " is not enabled).\n");
 printf(" -p: specify guest physical test memory offset\n"
        " Warning: a low offset can conflict with the loaded test code.\n");
 guest_modes_help();
 printf(" -n: Run the vCPUs in nested mode (L2)\n");
 printf(" -e: Run vCPUs while dirty logging is being disabled. This\n"
        " can significantly increase runtime, especially if there\n"
        " isn't a dedicated pCPU for the main thread.\n");
 printf(" -b: specify the size of the memory region which should be\n"
        " dirtied by each vCPU. e.g. 10M or 3G.\n"
        " (default: 1G)\n");
 printf(" -v: specify the number of vCPUs to run.\n");
 printf(" -o: Overlap guest memory accesses instead of partitioning\n"
        " them into a separate region of memory for each vCPU.\n");
 printf(" -r: specify the starting random seed.\n");
 backing_src_help("-s");
 printf(" -x: Split the memory region into this number of memslots.\n"
        " (default: 1)\n");
 printf(" -w: specify the percentage of pages which should be written to\n"
        " as an integer from 0-100 inclusive. This is probabilistic,\n"
        " so -w X means each page has an X%% chance of writing\n"
        " and a (100-X)%% chance of reading.\n"
        " (default: 100 i.e. all pages are written to.)\n");
 kvm_print_vcpu_pinning_help();
 puts("");
 exit(0);
}

int main(int argc, char *argv[])
{
 int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS);
 const char *pcpu_list = NULL;
 struct test_params p = {
  .iterations = TEST_HOST_LOOP_N,
  .partition_vcpu_memory_access = true,
  .backing_src = DEFAULT_VM_MEM_SRC,
  .slots = 1,
  .write_percent = 100,
 };
 int opt;

 /* Override the seed to be deterministic by default. */
 guest_random_seed = 1;

 dirty_log_manual_caps =
  kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
 dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
      KVM_DIRTY_LOG_INITIALLY_SET);

 guest_modes_append_default();

 while ((opt = getopt(argc, argv, "ab:c:eghi:m:nop:r:s:v:x:w:")) != -1) {
  switch (opt) {
  case 'a':
   p.random_access = true;
   break;
  case 'b':
   guest_percpu_mem_size = parse_size(optarg);
   break;
  case 'c':
   pcpu_list = optarg;
   break;
  case 'e':
   /* 'e' is for evil. */
   run_vcpus_while_disabling_dirty_logging = true;
   break;
  case 'g':
   dirty_log_manual_caps = 0;
   break;
  case 'h':
   help(argv[0]);
   break;
  case 'i':
   p.iterations = atoi_positive("Number of iterations", optarg);
   break;
  case 'm':
   guest_modes_cmdline(optarg);
   break;
  case 'n':
   memstress_args.nested = true;
   break;
  case 'o':
   p.partition_vcpu_memory_access = false;
   break;
  case 'p':
   p.phys_offset = strtoull(optarg, NULL, 0);
   break;
  case 'r':
   guest_random_seed = atoi_positive("Random seed", optarg);
   break;
  case 's':
   p.backing_src = parse_backing_src_type(optarg);
   break;
  case 'v':
   nr_vcpus = atoi_positive("Number of vCPUs", optarg);
   TEST_ASSERT(nr_vcpus <= max_vcpus,
        "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
   break;
  case 'w':
   p.write_percent = atoi_non_negative("Write percentage", optarg);
   TEST_ASSERT(p.write_percent <= 100,
        "Write percentage must be between 0 and 100");
   break;
  case 'x':
   p.slots = atoi_positive("Number of slots", optarg);
   break;
  default:
   help(argv[0]);
   break;
  }
 }

 if (pcpu_list) {
  kvm_parse_vcpu_pinning(pcpu_list, memstress_args.vcpu_to_pcpu,
           nr_vcpus);
  memstress_args.pin_vcpus = true;
 }

 TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations");

 pr_info("Test iterations: %"PRIu64"\n", p.iterations);

 for_each_guest_mode(run_test, &p);

 return 0;
}