Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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/* How many host loops to run by default (one KVM_GET_DIRTY_LOG for each loop)*/
24#define TEST_HOST_LOOP_N 2UL
25
26static int nr_vcpus = 1;
27static u64 guest_percpu_mem_size = DEFAULT_PER_VCPU_MEM_SIZE;
28static bool run_vcpus_while_disabling_dirty_logging;
29
30/* Host variables */
31static u64 dirty_log_manual_caps;
32static bool host_quit;
33static int iteration;
34static int vcpu_last_completed_iteration[KVM_MAX_VCPUS];
35
36static void vcpu_worker(struct memstress_vcpu_args *vcpu_args)
37{
38 struct kvm_vcpu *vcpu = vcpu_args->vcpu;
39 int vcpu_idx = vcpu_args->vcpu_idx;
40 u64 pages_count = 0;
41 struct kvm_run *run;
42 struct timespec start;
43 struct timespec ts_diff;
44 struct timespec total = (struct timespec){0};
45 struct timespec avg;
46 int ret;
47
48 run = vcpu->run;
49
50 while (!READ_ONCE(host_quit)) {
51 int current_iteration = READ_ONCE(iteration);
52
53 clock_gettime(CLOCK_MONOTONIC, &start);
54 ret = _vcpu_run(vcpu);
55 ts_diff = timespec_elapsed(start);
56
57 TEST_ASSERT(ret == 0, "vcpu_run failed: %d", ret);
58 TEST_ASSERT(get_ucall(vcpu, NULL) == UCALL_SYNC,
59 "Invalid guest sync status: exit_reason=%s",
60 exit_reason_str(run->exit_reason));
61
62 pr_debug("Got sync event from vCPU %d\n", vcpu_idx);
63 vcpu_last_completed_iteration[vcpu_idx] = current_iteration;
64 pr_debug("vCPU %d updated last completed iteration to %d\n",
65 vcpu_idx, vcpu_last_completed_iteration[vcpu_idx]);
66
67 if (current_iteration) {
68 pages_count += vcpu_args->pages;
69 total = timespec_add(total, ts_diff);
70 pr_debug("vCPU %d iteration %d dirty memory time: %ld.%.9lds\n",
71 vcpu_idx, current_iteration, ts_diff.tv_sec,
72 ts_diff.tv_nsec);
73 } else {
74 pr_debug("vCPU %d iteration %d populate memory time: %ld.%.9lds\n",
75 vcpu_idx, current_iteration, ts_diff.tv_sec,
76 ts_diff.tv_nsec);
77 }
78
79 /*
80 * Keep running the guest while dirty logging is being disabled
81 * (iteration is negative) so that vCPUs are accessing memory
82 * for the entire duration of zapping collapsible SPTEs.
83 */
84 while (current_iteration == READ_ONCE(iteration) &&
85 READ_ONCE(iteration) >= 0 && !READ_ONCE(host_quit)) {}
86 }
87
88 avg = timespec_div(total, vcpu_last_completed_iteration[vcpu_idx]);
89 pr_debug("\nvCPU %d dirtied 0x%lx pages over %d iterations in %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
90 vcpu_idx, pages_count, vcpu_last_completed_iteration[vcpu_idx],
91 total.tv_sec, total.tv_nsec, avg.tv_sec, avg.tv_nsec);
92}
93
94struct test_params {
95 unsigned long iterations;
96 u64 phys_offset;
97 bool partition_vcpu_memory_access;
98 enum vm_mem_backing_src_type backing_src;
99 int slots;
100 u32 write_percent;
101 bool random_access;
102};
103
104static void run_test(enum vm_guest_mode mode, void *arg)
105{
106 struct test_params *p = arg;
107 struct kvm_vm *vm;
108 unsigned long **bitmaps;
109 u64 guest_num_pages;
110 u64 host_num_pages;
111 u64 pages_per_slot;
112 struct timespec start;
113 struct timespec ts_diff;
114 struct timespec get_dirty_log_total = (struct timespec){0};
115 struct timespec vcpu_dirty_total = (struct timespec){0};
116 struct timespec avg;
117 struct timespec clear_dirty_log_total = (struct timespec){0};
118 int i;
119
120 vm = memstress_create_vm(mode, nr_vcpus, guest_percpu_mem_size,
121 p->slots, p->backing_src,
122 p->partition_vcpu_memory_access);
123
124 memstress_set_write_percent(vm, p->write_percent);
125
126 guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm->page_shift;
127 guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
128 host_num_pages = vm_num_host_pages(mode, guest_num_pages);
129 pages_per_slot = host_num_pages / p->slots;
130
131 bitmaps = memstress_alloc_bitmaps(p->slots, pages_per_slot);
132
133 if (dirty_log_manual_caps)
134 vm_enable_cap(vm, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2,
135 dirty_log_manual_caps);
136
137 /* Start the iterations */
138 iteration = 0;
139 host_quit = false;
140
141 clock_gettime(CLOCK_MONOTONIC, &start);
142 for (i = 0; i < nr_vcpus; i++)
143 vcpu_last_completed_iteration[i] = -1;
144
145 /*
146 * Use 100% writes during the population phase to ensure all
147 * memory is actually populated and not just mapped to the zero
148 * page. The prevents expensive copy-on-write faults from
149 * occurring during the dirty memory iterations below, which
150 * would pollute the performance results.
151 */
152 memstress_set_write_percent(vm, 100);
153 memstress_set_random_access(vm, false);
154 memstress_start_vcpu_threads(nr_vcpus, vcpu_worker);
155
156 /* Allow the vCPUs to populate memory */
157 pr_debug("Starting iteration %d - Populating\n", iteration);
158 for (i = 0; i < nr_vcpus; i++) {
159 while (READ_ONCE(vcpu_last_completed_iteration[i]) !=
160 iteration)
161 ;
162 }
163
164 ts_diff = timespec_elapsed(start);
165 pr_info("Populate memory time: %ld.%.9lds\n",
166 ts_diff.tv_sec, ts_diff.tv_nsec);
167
168 /* Enable dirty logging */
169 clock_gettime(CLOCK_MONOTONIC, &start);
170 memstress_enable_dirty_logging(vm, p->slots);
171 ts_diff = timespec_elapsed(start);
172 pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
173 ts_diff.tv_sec, ts_diff.tv_nsec);
174
175 memstress_set_write_percent(vm, p->write_percent);
176 memstress_set_random_access(vm, p->random_access);
177
178 while (iteration < p->iterations) {
179 /*
180 * Incrementing the iteration number will start the vCPUs
181 * dirtying memory again.
182 */
183 clock_gettime(CLOCK_MONOTONIC, &start);
184 iteration++;
185
186 pr_debug("Starting iteration %d\n", iteration);
187 for (i = 0; i < nr_vcpus; i++) {
188 while (READ_ONCE(vcpu_last_completed_iteration[i])
189 != iteration)
190 ;
191 }
192
193 ts_diff = timespec_elapsed(start);
194 vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
195 pr_info("Iteration %d dirty memory time: %ld.%.9lds\n",
196 iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
197
198 clock_gettime(CLOCK_MONOTONIC, &start);
199 memstress_get_dirty_log(vm, bitmaps, p->slots);
200 ts_diff = timespec_elapsed(start);
201 get_dirty_log_total = timespec_add(get_dirty_log_total,
202 ts_diff);
203 pr_info("Iteration %d get dirty log time: %ld.%.9lds\n",
204 iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
205
206 if (dirty_log_manual_caps) {
207 clock_gettime(CLOCK_MONOTONIC, &start);
208 memstress_clear_dirty_log(vm, bitmaps, p->slots,
209 pages_per_slot);
210 ts_diff = timespec_elapsed(start);
211 clear_dirty_log_total = timespec_add(clear_dirty_log_total,
212 ts_diff);
213 pr_info("Iteration %d clear dirty log time: %ld.%.9lds\n",
214 iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
215 }
216 }
217
218 /*
219 * Run vCPUs while dirty logging is being disabled to stress disabling
220 * in terms of both performance and correctness. Opt-in via command
221 * line as this significantly increases time to disable dirty logging.
222 */
223 if (run_vcpus_while_disabling_dirty_logging)
224 WRITE_ONCE(iteration, -1);
225
226 /* Disable dirty logging */
227 clock_gettime(CLOCK_MONOTONIC, &start);
228 memstress_disable_dirty_logging(vm, p->slots);
229 ts_diff = timespec_elapsed(start);
230 pr_info("Disabling dirty logging time: %ld.%.9lds\n",
231 ts_diff.tv_sec, ts_diff.tv_nsec);
232
233 /*
234 * Tell the vCPU threads to quit. No need to manually check that vCPUs
235 * have stopped running after disabling dirty logging, the join will
236 * wait for them to exit.
237 */
238 host_quit = true;
239 memstress_join_vcpu_threads(nr_vcpus);
240
241 avg = timespec_div(get_dirty_log_total, p->iterations);
242 pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
243 p->iterations, get_dirty_log_total.tv_sec,
244 get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
245
246 if (dirty_log_manual_caps) {
247 avg = timespec_div(clear_dirty_log_total, p->iterations);
248 pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
249 p->iterations, clear_dirty_log_total.tv_sec,
250 clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
251 }
252
253 memstress_free_bitmaps(bitmaps, p->slots);
254 memstress_destroy_vm(vm);
255}
256
257static void help(char *name)
258{
259 puts("");
260 printf("usage: %s [-h] [-a] [-i iterations] [-p offset] [-g] "
261 "[-m mode] [-n] [-b vcpu bytes] [-v vcpus] [-o] [-r random seed ] [-s mem type]"
262 "[-x memslots] [-w percentage] [-c physical cpus to run test on]\n", name);
263 puts("");
264 printf(" -a: access memory randomly rather than in order.\n");
265 printf(" -i: specify iteration counts (default: %"PRIu64")\n",
266 TEST_HOST_LOOP_N);
267 printf(" -g: Do not enable KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2. This\n"
268 " makes KVM_GET_DIRTY_LOG clear the dirty log (i.e.\n"
269 " KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE is not enabled)\n"
270 " and writes will be tracked as soon as dirty logging is\n"
271 " enabled on the memslot (i.e. KVM_DIRTY_LOG_INITIALLY_SET\n"
272 " is not enabled).\n");
273 printf(" -p: specify guest physical test memory offset\n"
274 " Warning: a low offset can conflict with the loaded test code.\n");
275 guest_modes_help();
276 printf(" -n: Run the vCPUs in nested mode (L2)\n");
277 printf(" -e: Run vCPUs while dirty logging is being disabled. This\n"
278 " can significantly increase runtime, especially if there\n"
279 " isn't a dedicated pCPU for the main thread.\n");
280 printf(" -b: specify the size of the memory region which should be\n"
281 " dirtied by each vCPU. e.g. 10M or 3G.\n"
282 " (default: 1G)\n");
283 printf(" -v: specify the number of vCPUs to run.\n");
284 printf(" -o: Overlap guest memory accesses instead of partitioning\n"
285 " them into a separate region of memory for each vCPU.\n");
286 printf(" -r: specify the starting random seed.\n");
287 backing_src_help("-s");
288 printf(" -x: Split the memory region into this number of memslots.\n"
289 " (default: 1)\n");
290 printf(" -w: specify the percentage of pages which should be written to\n"
291 " as an integer from 0-100 inclusive. This is probabilistic,\n"
292 " so -w X means each page has an X%% chance of writing\n"
293 " and a (100-X)%% chance of reading.\n"
294 " (default: 100 i.e. all pages are written to.)\n");
295 kvm_print_vcpu_pinning_help();
296 puts("");
297 exit(0);
298}
299
300int main(int argc, char *argv[])
301{
302 int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS);
303 const char *pcpu_list = NULL;
304 struct test_params p = {
305 .iterations = TEST_HOST_LOOP_N,
306 .partition_vcpu_memory_access = true,
307 .backing_src = DEFAULT_VM_MEM_SRC,
308 .slots = 1,
309 .write_percent = 100,
310 };
311 int opt;
312
313 /* Override the seed to be deterministic by default. */
314 guest_random_seed = 1;
315
316 dirty_log_manual_caps =
317 kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
318 dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
319 KVM_DIRTY_LOG_INITIALLY_SET);
320
321 guest_modes_append_default();
322
323 while ((opt = getopt(argc, argv, "ab:c:eghi:m:nop:r:s:v:x:w:")) != -1) {
324 switch (opt) {
325 case 'a':
326 p.random_access = true;
327 break;
328 case 'b':
329 guest_percpu_mem_size = parse_size(optarg);
330 break;
331 case 'c':
332 pcpu_list = optarg;
333 break;
334 case 'e':
335 /* 'e' is for evil. */
336 run_vcpus_while_disabling_dirty_logging = true;
337 break;
338 case 'g':
339 dirty_log_manual_caps = 0;
340 break;
341 case 'h':
342 help(argv[0]);
343 break;
344 case 'i':
345 p.iterations = atoi_positive("Number of iterations", optarg);
346 break;
347 case 'm':
348 guest_modes_cmdline(optarg);
349 break;
350 case 'n':
351 memstress_args.nested = true;
352 break;
353 case 'o':
354 p.partition_vcpu_memory_access = false;
355 break;
356 case 'p':
357 p.phys_offset = strtoull(optarg, NULL, 0);
358 break;
359 case 'r':
360 guest_random_seed = atoi_positive("Random seed", optarg);
361 break;
362 case 's':
363 p.backing_src = parse_backing_src_type(optarg);
364 break;
365 case 'v':
366 nr_vcpus = atoi_positive("Number of vCPUs", optarg);
367 TEST_ASSERT(nr_vcpus <= max_vcpus,
368 "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
369 break;
370 case 'w':
371 p.write_percent = atoi_non_negative("Write percentage", optarg);
372 TEST_ASSERT(p.write_percent <= 100,
373 "Write percentage must be between 0 and 100");
374 break;
375 case 'x':
376 p.slots = atoi_positive("Number of slots", optarg);
377 break;
378 default:
379 help(argv[0]);
380 break;
381 }
382 }
383
384 if (pcpu_list) {
385 kvm_parse_vcpu_pinning(pcpu_list, memstress_args.vcpu_to_pcpu,
386 nr_vcpus);
387 memstress_args.pin_vcpus = true;
388 }
389
390 TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations");
391
392 pr_info("Test iterations: %"PRIu64"\n", p.iterations);
393
394 for_each_guest_mode(run_test, &p);
395
396 return 0;
397}