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