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 "memstress.h"
20 #include "guest_modes.h"
21 #include "ucall_common.h"
22
23 #ifdef __aarch64__
24 #include "aarch64/vgic.h"
25
26 static int gic_fd;
27
28 static void arch_setup_vm(struct kvm_vm *vm, unsigned int nr_vcpus)
29 {
30 /*
31 * The test can still run even if hardware does not support GICv3, as it
32 * is only an optimization to reduce guest exits.
33 */
34 gic_fd = vgic_v3_setup(vm, nr_vcpus, 64);
35 }
36
37 static void arch_cleanup_vm(struct kvm_vm *vm)
38 {
39 if (gic_fd > 0)
40 close(gic_fd);
41 }
42
43 #else /* __aarch64__ */
44
45 static void arch_setup_vm(struct kvm_vm *vm, unsigned int nr_vcpus)
46 {
47 }
48
49 static void arch_cleanup_vm(struct kvm_vm *vm)
50 {
51 }
52
53 #endif
54
55 /* How many host loops to run by default (one KVM_GET_DIRTY_LOG for each loop)*/
56 #define TEST_HOST_LOOP_N 2UL
57
58 static int nr_vcpus = 1;
59 static uint64_t guest_percpu_mem_size = DEFAULT_PER_VCPU_MEM_SIZE;
60 static bool run_vcpus_while_disabling_dirty_logging;
61
62 /* Host variables */
63 static u64 dirty_log_manual_caps;
64 static bool host_quit;
65 static int iteration;
66 static int vcpu_last_completed_iteration[KVM_MAX_VCPUS];
67
68 static void vcpu_worker(struct memstress_vcpu_args *vcpu_args)
69 {
70 struct kvm_vcpu *vcpu = vcpu_args->vcpu;
71 int vcpu_idx = vcpu_args->vcpu_idx;
72 uint64_t pages_count = 0;
73 struct kvm_run *run;
74 struct timespec start;
75 struct timespec ts_diff;
76 struct timespec total = (struct timespec){0};
77 struct timespec avg;
78 int ret;
79
80 run = vcpu->run;
81
82 while (!READ_ONCE(host_quit)) {
83 int current_iteration = READ_ONCE(iteration);
84
85 clock_gettime(CLOCK_MONOTONIC, &start);
86 ret = _vcpu_run(vcpu);
87 ts_diff = timespec_elapsed(start);
88
89 TEST_ASSERT(ret == 0, "vcpu_run failed: %d", ret);
90 TEST_ASSERT(get_ucall(vcpu, NULL) == UCALL_SYNC,
91 "Invalid guest sync status: exit_reason=%s",
92 exit_reason_str(run->exit_reason));
93
94 pr_debug("Got sync event from vCPU %d\n", vcpu_idx);
95 vcpu_last_completed_iteration[vcpu_idx] = current_iteration;
96 pr_debug("vCPU %d updated last completed iteration to %d\n",
97 vcpu_idx, vcpu_last_completed_iteration[vcpu_idx]);
98
99 if (current_iteration) {
100 pages_count += vcpu_args->pages;
101 total = timespec_add(total, ts_diff);
102 pr_debug("vCPU %d iteration %d dirty memory time: %ld.%.9lds\n",
103 vcpu_idx, current_iteration, ts_diff.tv_sec,
104 ts_diff.tv_nsec);
105 } else {
106 pr_debug("vCPU %d iteration %d populate memory time: %ld.%.9lds\n",
107 vcpu_idx, current_iteration, ts_diff.tv_sec,
108 ts_diff.tv_nsec);
109 }
110
111 /*
112 * Keep running the guest while dirty logging is being disabled
113 * (iteration is negative) so that vCPUs are accessing memory
114 * for the entire duration of zapping collapsible SPTEs.
115 */
116 while (current_iteration == READ_ONCE(iteration) &&
117 READ_ONCE(iteration) >= 0 && !READ_ONCE(host_quit)) {}
118 }
119
120 avg = timespec_div(total, vcpu_last_completed_iteration[vcpu_idx]);
121 pr_debug("\nvCPU %d dirtied 0x%lx pages over %d iterations in %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
122 vcpu_idx, pages_count, vcpu_last_completed_iteration[vcpu_idx],
123 total.tv_sec, total.tv_nsec, avg.tv_sec, avg.tv_nsec);
124 }
125
126 struct test_params {
127 unsigned long iterations;
128 uint64_t phys_offset;
129 bool partition_vcpu_memory_access;
130 enum vm_mem_backing_src_type backing_src;
131 int slots;
132 uint32_t write_percent;
133 bool random_access;
134 };
135
136 static void run_test(enum vm_guest_mode mode, void *arg)
137 {
138 struct test_params *p = arg;
139 struct kvm_vm *vm;
140 unsigned long **bitmaps;
141 uint64_t guest_num_pages;
142 uint64_t host_num_pages;
143 uint64_t pages_per_slot;
144 struct timespec start;
145 struct timespec ts_diff;
146 struct timespec get_dirty_log_total = (struct timespec){0};
147 struct timespec vcpu_dirty_total = (struct timespec){0};
148 struct timespec avg;
149 struct timespec clear_dirty_log_total = (struct timespec){0};
150 int i;
151
152 vm = memstress_create_vm(mode, nr_vcpus, guest_percpu_mem_size,
153 p->slots, p->backing_src,
154 p->partition_vcpu_memory_access);
155
156 memstress_set_write_percent(vm, p->write_percent);
157
158 guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm->page_shift;
159 guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
160 host_num_pages = vm_num_host_pages(mode, guest_num_pages);
161 pages_per_slot = host_num_pages / p->slots;
162
163 bitmaps = memstress_alloc_bitmaps(p->slots, pages_per_slot);
164
165 if (dirty_log_manual_caps)
166 vm_enable_cap(vm, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2,
167 dirty_log_manual_caps);
168
169 arch_setup_vm(vm, nr_vcpus);
170
171 /* Start the iterations */
172 iteration = 0;
173 host_quit = false;
174
175 clock_gettime(CLOCK_MONOTONIC, &start);
176 for (i = 0; i < nr_vcpus; i++)
177 vcpu_last_completed_iteration[i] = -1;
178
179 /*
180 * Use 100% writes during the population phase to ensure all
181 * memory is actually populated and not just mapped to the zero
182 * page. The prevents expensive copy-on-write faults from
183 * occurring during the dirty memory iterations below, which
184 * would pollute the performance results.
185 */
186 memstress_set_write_percent(vm, 100);
187 memstress_set_random_access(vm, false);
188 memstress_start_vcpu_threads(nr_vcpus, vcpu_worker);
189
190 /* Allow the vCPUs to populate memory */
191 pr_debug("Starting iteration %d - Populating\n", iteration);
192 for (i = 0; i < nr_vcpus; i++) {
193 while (READ_ONCE(vcpu_last_completed_iteration[i]) !=
194 iteration)
195 ;
196 }
197
198 ts_diff = timespec_elapsed(start);
199 pr_info("Populate memory time: %ld.%.9lds\n",
200 ts_diff.tv_sec, ts_diff.tv_nsec);
201
202 /* Enable dirty logging */
203 clock_gettime(CLOCK_MONOTONIC, &start);
204 memstress_enable_dirty_logging(vm, p->slots);
205 ts_diff = timespec_elapsed(start);
206 pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
207 ts_diff.tv_sec, ts_diff.tv_nsec);
208
209 memstress_set_write_percent(vm, p->write_percent);
210 memstress_set_random_access(vm, p->random_access);
211
212 while (iteration < p->iterations) {
213 /*
214 * Incrementing the iteration number will start the vCPUs
215 * dirtying memory again.
216 */
217 clock_gettime(CLOCK_MONOTONIC, &start);
218 iteration++;
219
220 pr_debug("Starting iteration %d\n", iteration);
221 for (i = 0; i < nr_vcpus; i++) {
222 while (READ_ONCE(vcpu_last_completed_iteration[i])
223 != iteration)
224 ;
225 }
226
227 ts_diff = timespec_elapsed(start);
228 vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
229 pr_info("Iteration %d dirty memory time: %ld.%.9lds\n",
230 iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
231
232 clock_gettime(CLOCK_MONOTONIC, &start);
233 memstress_get_dirty_log(vm, bitmaps, p->slots);
234 ts_diff = timespec_elapsed(start);
235 get_dirty_log_total = timespec_add(get_dirty_log_total,
236 ts_diff);
237 pr_info("Iteration %d get dirty log time: %ld.%.9lds\n",
238 iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
239
240 if (dirty_log_manual_caps) {
241 clock_gettime(CLOCK_MONOTONIC, &start);
242 memstress_clear_dirty_log(vm, bitmaps, p->slots,
243 pages_per_slot);
244 ts_diff = timespec_elapsed(start);
245 clear_dirty_log_total = timespec_add(clear_dirty_log_total,
246 ts_diff);
247 pr_info("Iteration %d clear dirty log time: %ld.%.9lds\n",
248 iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
249 }
250 }
251
252 /*
253 * Run vCPUs while dirty logging is being disabled to stress disabling
254 * in terms of both performance and correctness. Opt-in via command
255 * line as this significantly increases time to disable dirty logging.
256 */
257 if (run_vcpus_while_disabling_dirty_logging)
258 WRITE_ONCE(iteration, -1);
259
260 /* Disable dirty logging */
261 clock_gettime(CLOCK_MONOTONIC, &start);
262 memstress_disable_dirty_logging(vm, p->slots);
263 ts_diff = timespec_elapsed(start);
264 pr_info("Disabling dirty logging time: %ld.%.9lds\n",
265 ts_diff.tv_sec, ts_diff.tv_nsec);
266
267 /*
268 * Tell the vCPU threads to quit. No need to manually check that vCPUs
269 * have stopped running after disabling dirty logging, the join will
270 * wait for them to exit.
271 */
272 host_quit = true;
273 memstress_join_vcpu_threads(nr_vcpus);
274
275 avg = timespec_div(get_dirty_log_total, p->iterations);
276 pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
277 p->iterations, get_dirty_log_total.tv_sec,
278 get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
279
280 if (dirty_log_manual_caps) {
281 avg = timespec_div(clear_dirty_log_total, p->iterations);
282 pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
283 p->iterations, clear_dirty_log_total.tv_sec,
284 clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
285 }
286
287 memstress_free_bitmaps(bitmaps, p->slots);
288 arch_cleanup_vm(vm);
289 memstress_destroy_vm(vm);
290 }
291
292 static void help(char *name)
293 {
294 puts("");
295 printf("usage: %s [-h] [-a] [-i iterations] [-p offset] [-g] "
296 "[-m mode] [-n] [-b vcpu bytes] [-v vcpus] [-o] [-r random seed ] [-s mem type]"
297 "[-x memslots] [-w percentage] [-c physical cpus to run test on]\n", name);
298 puts("");
299 printf(" -a: access memory randomly rather than in order.\n");
300 printf(" -i: specify iteration counts (default: %"PRIu64")\n",
301 TEST_HOST_LOOP_N);
302 printf(" -g: Do not enable KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2. This\n"
303 " makes KVM_GET_DIRTY_LOG clear the dirty log (i.e.\n"
304 " KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE is not enabled)\n"
305 " and writes will be tracked as soon as dirty logging is\n"
306 " enabled on the memslot (i.e. KVM_DIRTY_LOG_INITIALLY_SET\n"
307 " is not enabled).\n");
308 printf(" -p: specify guest physical test memory offset\n"
309 " Warning: a low offset can conflict with the loaded test code.\n");
310 guest_modes_help();
311 printf(" -n: Run the vCPUs in nested mode (L2)\n");
312 printf(" -e: Run vCPUs while dirty logging is being disabled. This\n"
313 " can significantly increase runtime, especially if there\n"
314 " isn't a dedicated pCPU for the main thread.\n");
315 printf(" -b: specify the size of the memory region which should be\n"
316 " dirtied by each vCPU. e.g. 10M or 3G.\n"
317 " (default: 1G)\n");
318 printf(" -v: specify the number of vCPUs to run.\n");
319 printf(" -o: Overlap guest memory accesses instead of partitioning\n"
320 " them into a separate region of memory for each vCPU.\n");
321 printf(" -r: specify the starting random seed.\n");
322 backing_src_help("-s");
323 printf(" -x: Split the memory region into this number of memslots.\n"
324 " (default: 1)\n");
325 printf(" -w: specify the percentage of pages which should be written to\n"
326 " as an integer from 0-100 inclusive. This is probabilistic,\n"
327 " so -w X means each page has an X%% chance of writing\n"
328 " and a (100-X)%% chance of reading.\n"
329 " (default: 100 i.e. all pages are written to.)\n");
330 kvm_print_vcpu_pinning_help();
331 puts("");
332 exit(0);
333 }
334
335 int main(int argc, char *argv[])
336 {
337 int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS);
338 const char *pcpu_list = NULL;
339 struct test_params p = {
340 .iterations = TEST_HOST_LOOP_N,
341 .partition_vcpu_memory_access = true,
342 .backing_src = DEFAULT_VM_MEM_SRC,
343 .slots = 1,
344 .write_percent = 100,
345 };
346 int opt;
347
348 /* Override the seed to be deterministic by default. */
349 guest_random_seed = 1;
350
351 dirty_log_manual_caps =
352 kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
353 dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
354 KVM_DIRTY_LOG_INITIALLY_SET);
355
356 guest_modes_append_default();
357
358 while ((opt = getopt(argc, argv, "ab:c:eghi:m:nop:r:s:v:x:w:")) != -1) {
359 switch (opt) {
360 case 'a':
361 p.random_access = true;
362 break;
363 case 'b':
364 guest_percpu_mem_size = parse_size(optarg);
365 break;
366 case 'c':
367 pcpu_list = optarg;
368 break;
369 case 'e':
370 /* 'e' is for evil. */
371 run_vcpus_while_disabling_dirty_logging = true;
372 break;
373 case 'g':
374 dirty_log_manual_caps = 0;
375 break;
376 case 'h':
377 help(argv[0]);
378 break;
379 case 'i':
380 p.iterations = atoi_positive("Number of iterations", optarg);
381 break;
382 case 'm':
383 guest_modes_cmdline(optarg);
384 break;
385 case 'n':
386 memstress_args.nested = true;
387 break;
388 case 'o':
389 p.partition_vcpu_memory_access = false;
390 break;
391 case 'p':
392 p.phys_offset = strtoull(optarg, NULL, 0);
393 break;
394 case 'r':
395 guest_random_seed = atoi_positive("Random seed", optarg);
396 break;
397 case 's':
398 p.backing_src = parse_backing_src_type(optarg);
399 break;
400 case 'v':
401 nr_vcpus = atoi_positive("Number of vCPUs", optarg);
402 TEST_ASSERT(nr_vcpus <= max_vcpus,
403 "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
404 break;
405 case 'w':
406 p.write_percent = atoi_non_negative("Write percentage", optarg);
407 TEST_ASSERT(p.write_percent <= 100,
408 "Write percentage must be between 0 and 100");
409 break;
410 case 'x':
411 p.slots = atoi_positive("Number of slots", optarg);
412 break;
413 default:
414 help(argv[0]);
415 break;
416 }
417 }
418
419 if (pcpu_list) {
420 kvm_parse_vcpu_pinning(pcpu_list, memstress_args.vcpu_to_pcpu,
421 nr_vcpus);
422 memstress_args.pin_vcpus = true;
423 }
424
425 TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations");
426
427 pr_info("Test iterations: %"PRIu64"\n", p.iterations);
428
429 for_each_guest_mode(run_test, &p);
430
431 return 0;
432 }
433
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.