tools/testing/selftests/kvm/dirty_log_perf_test.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * KVM dirty page logging performance test
   4  *
   5  * Based on dirty_log_test.c
   6  *
   7  * Copyright (C) 2018, Red Hat, Inc.
   8  * Copyright (C) 2020, Google, Inc.
   9  */
  10
  11 #define _GNU_SOURCE /* for program_invocation_name */
  12
  13 #include <stdio.h>
  14 #include <stdlib.h>
  15 #include <time.h>
  16 #include <pthread.h>
  17 #include <linux/bitmap.h>
  18
  19 #include "kvm_util.h"
  20 #include "test_util.h"
  21 #include "perf_test_util.h"
  22 #include "guest_modes.h"
  23
  24 /* How many host loops to run by default (one KVM_GET_DIRTY_LOG for each loop)*/
  25 #define TEST_HOST_LOOP_N                2UL
  26
  27 /* Host variables */
  28 static u64 dirty_log_manual_caps;
  29 static bool host_quit;
  30 static uint64_t iteration;
  31 static uint64_t vcpu_last_completed_iteration[MAX_VCPUS];
  32
  33 static void *vcpu_worker(void *data)
  34 {
  35         int ret;
  36         struct kvm_vm *vm = perf_test_args.vm;
  37         uint64_t pages_count = 0;
  38         struct kvm_run *run;
  39         struct timespec start;
  40         struct timespec ts_diff;
  41         struct timespec total = (struct timespec){0};
  42         struct timespec avg;
  43         struct vcpu_args *vcpu_args = (struct vcpu_args *)data;
  44         int vcpu_id = vcpu_args->vcpu_id;
  45
  46         vcpu_args_set(vm, vcpu_id, 1, vcpu_id);
  47         run = vcpu_state(vm, vcpu_id);
  48
  49         while (!READ_ONCE(host_quit)) {
  50                 uint64_t current_iteration = READ_ONCE(iteration);
  51
  52                 clock_gettime(CLOCK_MONOTONIC, &start);
  53                 ret = _vcpu_run(vm, vcpu_id);
  54                 ts_diff = timespec_diff_now(start);
  55
  56                 TEST_ASSERT(ret == 0, "vcpu_run failed: %d\n", ret);
  57                 TEST_ASSERT(get_ucall(vm, vcpu_id, NULL) == UCALL_SYNC,
  58                             "Invalid guest sync status: exit_reason=%s\n",
  59                             exit_reason_str(run->exit_reason));
  60
  61                 pr_debug("Got sync event from vCPU %d\n", vcpu_id);
  62                 vcpu_last_completed_iteration[vcpu_id] = current_iteration;
  63                 pr_debug("vCPU %d updated last completed iteration to %lu\n",
  64                          vcpu_id, vcpu_last_completed_iteration[vcpu_id]);
  65
  66                 if (current_iteration) {
  67                         pages_count += vcpu_args->pages;
  68                         total = timespec_add(total, ts_diff);
  69                         pr_debug("vCPU %d iteration %lu dirty memory time: %ld.%.9lds\n",
  70                                 vcpu_id, current_iteration, ts_diff.tv_sec,
  71                                 ts_diff.tv_nsec);
  72                 } else {
  73                         pr_debug("vCPU %d iteration %lu populate memory time: %ld.%.9lds\n",
  74                                 vcpu_id, current_iteration, ts_diff.tv_sec,
  75                                 ts_diff.tv_nsec);
  76                 }
  77
  78                 while (current_iteration == READ_ONCE(iteration) &&
  79                        !READ_ONCE(host_quit)) {}
  80         }
  81
  82         avg = timespec_div(total, vcpu_last_completed_iteration[vcpu_id]);
  83         pr_debug("\nvCPU %d dirtied 0x%lx pages over %lu iterations in %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
  84                 vcpu_id, pages_count, vcpu_last_completed_iteration[vcpu_id],
  85                 total.tv_sec, total.tv_nsec, avg.tv_sec, avg.tv_nsec);
  86
  87         return NULL;
  88 }
  89
  90 struct test_params {
  91         unsigned long iterations;
  92         uint64_t phys_offset;
  93         int wr_fract;
  94 };
  95
  96 static void run_test(enum vm_guest_mode mode, void *arg)
  97 {
  98         struct test_params *p = arg;
  99         pthread_t *vcpu_threads;
 100         struct kvm_vm *vm;
 101         unsigned long *bmap;
 102         uint64_t guest_num_pages;
 103         uint64_t host_num_pages;
 104         int vcpu_id;
 105         struct timespec start;
 106         struct timespec ts_diff;
 107         struct timespec get_dirty_log_total = (struct timespec){0};
 108         struct timespec vcpu_dirty_total = (struct timespec){0};
 109         struct timespec avg;
 110         struct kvm_enable_cap cap = {};
 111         struct timespec clear_dirty_log_total = (struct timespec){0};
 112
 113         vm = create_vm(mode, nr_vcpus, guest_percpu_mem_size);
 114
 115         perf_test_args.wr_fract = p->wr_fract;
 116
 117         guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm_get_page_shift(vm);
 118         guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
 119         host_num_pages = vm_num_host_pages(mode, guest_num_pages);
 120         bmap = bitmap_alloc(host_num_pages);
 121
 122         if (dirty_log_manual_caps) {
 123                 cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
 124                 cap.args[0] = dirty_log_manual_caps;
 125                 vm_enable_cap(vm, &cap);
 126         }
 127
 128         vcpu_threads = malloc(nr_vcpus * sizeof(*vcpu_threads));
 129         TEST_ASSERT(vcpu_threads, "Memory allocation failed");
 130
 131         add_vcpus(vm, nr_vcpus, guest_percpu_mem_size);
 132
 133         sync_global_to_guest(vm, perf_test_args);
 134
 135         /* Start the iterations */
 136         iteration = 0;
 137         host_quit = false;
 138
 139         clock_gettime(CLOCK_MONOTONIC, &start);
 140         for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
 141                 pthread_create(&vcpu_threads[vcpu_id], NULL, vcpu_worker,
 142                                &perf_test_args.vcpu_args[vcpu_id]);
 143         }
 144
 145         /* Allow the vCPU to populate memory */
 146         pr_debug("Starting iteration %lu - Populating\n", iteration);
 147         while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) != iteration)
 148                 pr_debug("Waiting for vcpu_last_completed_iteration == %lu\n",
 149                         iteration);
 150
 151         ts_diff = timespec_diff_now(start);
 152         pr_info("Populate memory time: %ld.%.9lds\n",
 153                 ts_diff.tv_sec, ts_diff.tv_nsec);
 154
 155         /* Enable dirty logging */
 156         clock_gettime(CLOCK_MONOTONIC, &start);
 157         vm_mem_region_set_flags(vm, TEST_MEM_SLOT_INDEX,
 158                                 KVM_MEM_LOG_DIRTY_PAGES);
 159         ts_diff = timespec_diff_now(start);
 160         pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
 161                 ts_diff.tv_sec, ts_diff.tv_nsec);
 162
 163         while (iteration < p->iterations) {
 164                 /*
 165                  * Incrementing the iteration number will start the vCPUs
 166                  * dirtying memory again.
 167                  */
 168                 clock_gettime(CLOCK_MONOTONIC, &start);
 169                 iteration++;
 170
 171                 pr_debug("Starting iteration %lu\n", iteration);
 172                 for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
 173                         while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) != iteration)
 174                                 pr_debug("Waiting for vCPU %d vcpu_last_completed_iteration == %lu\n",
 175                                          vcpu_id, iteration);
 176                 }
 177
 178                 ts_diff = timespec_diff_now(start);
 179                 vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
 180                 pr_info("Iteration %lu dirty memory time: %ld.%.9lds\n",
 181                         iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
 182
 183                 clock_gettime(CLOCK_MONOTONIC, &start);
 184                 kvm_vm_get_dirty_log(vm, TEST_MEM_SLOT_INDEX, bmap);
 185
 186                 ts_diff = timespec_diff_now(start);
 187                 get_dirty_log_total = timespec_add(get_dirty_log_total,
 188                                                    ts_diff);
 189                 pr_info("Iteration %lu get dirty log time: %ld.%.9lds\n",
 190                         iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
 191
 192                 if (dirty_log_manual_caps) {
 193                         clock_gettime(CLOCK_MONOTONIC, &start);
 194                         kvm_vm_clear_dirty_log(vm, TEST_MEM_SLOT_INDEX, bmap, 0,
 195                                                host_num_pages);
 196
 197                         ts_diff = timespec_diff_now(start);
 198                         clear_dirty_log_total = timespec_add(clear_dirty_log_total,
 199                                                              ts_diff);
 200                         pr_info("Iteration %lu clear dirty log time: %ld.%.9lds\n",
 201                                 iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
 202                 }
 203         }
 204
 205         /* Tell the vcpu thread to quit */
 206         host_quit = true;
 207         for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++)
 208                 pthread_join(vcpu_threads[vcpu_id], NULL);
 209
 210         /* Disable dirty logging */
 211         clock_gettime(CLOCK_MONOTONIC, &start);
 212         vm_mem_region_set_flags(vm, TEST_MEM_SLOT_INDEX, 0);
 213         ts_diff = timespec_diff_now(start);
 214         pr_info("Disabling dirty logging time: %ld.%.9lds\n",
 215                 ts_diff.tv_sec, ts_diff.tv_nsec);
 216
 217         avg = timespec_div(get_dirty_log_total, p->iterations);
 218         pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
 219                 p->iterations, get_dirty_log_total.tv_sec,
 220                 get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
 221
 222         if (dirty_log_manual_caps) {
 223                 avg = timespec_div(clear_dirty_log_total, p->iterations);
 224                 pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
 225                         p->iterations, clear_dirty_log_total.tv_sec,
 226                         clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
 227         }
 228
 229         free(bmap);
 230         free(vcpu_threads);
 231         ucall_uninit(vm);
 232         kvm_vm_free(vm);
 233 }
 234
 235 static void help(char *name)
 236 {
 237         puts("");
 238         printf("usage: %s [-h] [-i iterations] [-p offset] "
 239                "[-m mode] [-b vcpu bytes] [-v vcpus]\n", name);
 240         puts("");
 241         printf(" -i: specify iteration counts (default: %"PRIu64")\n",
 242                TEST_HOST_LOOP_N);
 243         printf(" -p: specify guest physical test memory offset\n"
 244                "     Warning: a low offset can conflict with the loaded test code.\n");
 245         guest_modes_help();
 246         printf(" -b: specify the size of the memory region which should be\n"
 247                "     dirtied by each vCPU. e.g. 10M or 3G.\n"
 248                "     (default: 1G)\n");
 249         printf(" -f: specify the fraction of pages which should be written to\n"
 250                "     as opposed to simply read, in the form\n"
 251                "     1/<fraction of pages to write>.\n"
 252                "     (default: 1 i.e. all pages are written to.)\n");
 253         printf(" -v: specify the number of vCPUs to run.\n");
 254         puts("");
 255         exit(0);
 256 }
 257
 258 int main(int argc, char *argv[])
 259 {
 260         int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS);
 261         struct test_params p = {
 262                 .iterations = TEST_HOST_LOOP_N,
 263                 .wr_fract = 1,
 264         };
 265         int opt;
 266
 267         dirty_log_manual_caps =
 268                 kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
 269         dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
 270                                   KVM_DIRTY_LOG_INITIALLY_SET);
 271
 272         guest_modes_append_default();
 273
 274         while ((opt = getopt(argc, argv, "hi:p:m:b:f:v:")) != -1) {
 275                 switch (opt) {
 276                 case 'i':
 277                         p.iterations = strtol(optarg, NULL, 10);
 278                         break;
 279                 case 'p':
 280                         p.phys_offset = strtoull(optarg, NULL, 0);
 281                         break;
 282                 case 'm':
 283                         guest_modes_cmdline(optarg);
 284                         break;
 285                 case 'b':
 286                         guest_percpu_mem_size = parse_size(optarg);
 287                         break;
 288                 case 'f':
 289                         p.wr_fract = atoi(optarg);
 290                         TEST_ASSERT(p.wr_fract >= 1,
 291                                     "Write fraction cannot be less than one");
 292                         break;
 293                 case 'v':
 294                         nr_vcpus = atoi(optarg);
 295                         TEST_ASSERT(nr_vcpus > 0 && nr_vcpus <= max_vcpus,
 296                                     "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
 297                         break;
 298                 case 'h':
 299                 default:
 300                         help(argv[0]);
 301                         break;
 302                 }
 303         }
 304
 305         TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations");
 306
 307         pr_info("Test iterations: %"PRIu64"\n", p.iterations);
 308
 309         for_each_guest_mode(run_test, &p);
 310
 311         return 0;
 312 }