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 | |
22 | #ifdef __aarch64__ |
23 | #include "aarch64/vgic.h" |
24 | |
25 | #define GICD_BASE_GPA 0x8000000ULL |
26 | #define GICR_BASE_GPA 0x80A0000ULL |
27 | |
28 | static int gic_fd; |
29 | |
30 | static void arch_setup_vm(struct kvm_vm *vm, unsigned int nr_vcpus) |
31 | { |
32 | /* |
33 | * The test can still run even if hardware does not support GICv3, as it |
34 | * is only an optimization to reduce guest exits. |
35 | */ |
36 | gic_fd = vgic_v3_setup(vm, nr_vcpus, 64, GICD_BASE_GPA, GICR_BASE_GPA); |
37 | } |
38 | |
39 | static void arch_cleanup_vm(struct kvm_vm *vm) |
40 | { |
41 | if (gic_fd > 0) |
42 | close(gic_fd); |
43 | } |
44 | |
45 | #else /* __aarch64__ */ |
46 | |
47 | static void arch_setup_vm(struct kvm_vm *vm, unsigned int nr_vcpus) |
48 | { |
49 | } |
50 | |
51 | static void arch_cleanup_vm(struct kvm_vm *vm) |
52 | { |
53 | } |
54 | |
55 | #endif |
56 | |
57 | /* How many host loops to run by default (one KVM_GET_DIRTY_LOG for each loop)*/ |
58 | #define TEST_HOST_LOOP_N 2UL |
59 | |
60 | static int nr_vcpus = 1; |
61 | static uint64_t guest_percpu_mem_size = DEFAULT_PER_VCPU_MEM_SIZE; |
62 | static bool run_vcpus_while_disabling_dirty_logging; |
63 | |
64 | /* Host variables */ |
65 | static u64 dirty_log_manual_caps; |
66 | static bool host_quit; |
67 | static int iteration; |
68 | static int vcpu_last_completed_iteration[KVM_MAX_VCPUS]; |
69 | |
70 | static void vcpu_worker(struct memstress_vcpu_args *vcpu_args) |
71 | { |
72 | struct kvm_vcpu *vcpu = vcpu_args->vcpu; |
73 | int vcpu_idx = vcpu_args->vcpu_idx; |
74 | uint64_t pages_count = 0; |
75 | struct kvm_run *run; |
76 | struct timespec start; |
77 | struct timespec ts_diff; |
78 | struct timespec total = (struct timespec){0}; |
79 | struct timespec avg; |
80 | int ret; |
81 | |
82 | run = vcpu->run; |
83 | |
84 | while (!READ_ONCE(host_quit)) { |
85 | int current_iteration = READ_ONCE(iteration); |
86 | |
87 | clock_gettime(CLOCK_MONOTONIC, &start); |
88 | ret = _vcpu_run(vcpu); |
89 | ts_diff = timespec_elapsed(start); |
90 | |
91 | TEST_ASSERT(ret == 0, "vcpu_run failed: %d" , ret); |
92 | TEST_ASSERT(get_ucall(vcpu, NULL) == UCALL_SYNC, |
93 | "Invalid guest sync status: exit_reason=%s" , |
94 | exit_reason_str(run->exit_reason)); |
95 | |
96 | pr_debug("Got sync event from vCPU %d\n" , vcpu_idx); |
97 | vcpu_last_completed_iteration[vcpu_idx] = current_iteration; |
98 | pr_debug("vCPU %d updated last completed iteration to %d\n" , |
99 | vcpu_idx, vcpu_last_completed_iteration[vcpu_idx]); |
100 | |
101 | if (current_iteration) { |
102 | pages_count += vcpu_args->pages; |
103 | total = timespec_add(total, ts_diff); |
104 | pr_debug("vCPU %d iteration %d dirty memory time: %ld.%.9lds\n" , |
105 | vcpu_idx, current_iteration, ts_diff.tv_sec, |
106 | ts_diff.tv_nsec); |
107 | } else { |
108 | pr_debug("vCPU %d iteration %d populate memory time: %ld.%.9lds\n" , |
109 | vcpu_idx, current_iteration, ts_diff.tv_sec, |
110 | ts_diff.tv_nsec); |
111 | } |
112 | |
113 | /* |
114 | * Keep running the guest while dirty logging is being disabled |
115 | * (iteration is negative) so that vCPUs are accessing memory |
116 | * for the entire duration of zapping collapsible SPTEs. |
117 | */ |
118 | while (current_iteration == READ_ONCE(iteration) && |
119 | READ_ONCE(iteration) >= 0 && !READ_ONCE(host_quit)) {} |
120 | } |
121 | |
122 | avg = timespec_div(total, vcpu_last_completed_iteration[vcpu_idx]); |
123 | pr_debug("\nvCPU %d dirtied 0x%lx pages over %d iterations in %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n" , |
124 | vcpu_idx, pages_count, vcpu_last_completed_iteration[vcpu_idx], |
125 | total.tv_sec, total.tv_nsec, avg.tv_sec, avg.tv_nsec); |
126 | } |
127 | |
128 | struct test_params { |
129 | unsigned long iterations; |
130 | uint64_t phys_offset; |
131 | bool partition_vcpu_memory_access; |
132 | enum vm_mem_backing_src_type backing_src; |
133 | int slots; |
134 | uint32_t write_percent; |
135 | uint32_t random_seed; |
136 | bool random_access; |
137 | }; |
138 | |
139 | static void run_test(enum vm_guest_mode mode, void *arg) |
140 | { |
141 | struct test_params *p = arg; |
142 | struct kvm_vm *vm; |
143 | unsigned long **bitmaps; |
144 | uint64_t guest_num_pages; |
145 | uint64_t host_num_pages; |
146 | uint64_t pages_per_slot; |
147 | struct timespec start; |
148 | struct timespec ts_diff; |
149 | struct timespec get_dirty_log_total = (struct timespec){0}; |
150 | struct timespec vcpu_dirty_total = (struct timespec){0}; |
151 | struct timespec avg; |
152 | struct timespec clear_dirty_log_total = (struct timespec){0}; |
153 | int i; |
154 | |
155 | vm = memstress_create_vm(mode, nr_vcpus, guest_percpu_mem_size, |
156 | p->slots, p->backing_src, |
157 | p->partition_vcpu_memory_access); |
158 | |
159 | pr_info("Random seed: %u\n" , p->random_seed); |
160 | memstress_set_random_seed(vm, p->random_seed); |
161 | memstress_set_write_percent(vm, p->write_percent); |
162 | |
163 | guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm->page_shift; |
164 | guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages); |
165 | host_num_pages = vm_num_host_pages(mode, guest_num_pages); |
166 | pages_per_slot = host_num_pages / p->slots; |
167 | |
168 | bitmaps = memstress_alloc_bitmaps(p->slots, pages_per_slot); |
169 | |
170 | if (dirty_log_manual_caps) |
171 | vm_enable_cap(vm, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, |
172 | dirty_log_manual_caps); |
173 | |
174 | arch_setup_vm(vm, nr_vcpus); |
175 | |
176 | /* Start the iterations */ |
177 | iteration = 0; |
178 | host_quit = false; |
179 | |
180 | clock_gettime(CLOCK_MONOTONIC, &start); |
181 | for (i = 0; i < nr_vcpus; i++) |
182 | vcpu_last_completed_iteration[i] = -1; |
183 | |
184 | /* |
185 | * Use 100% writes during the population phase to ensure all |
186 | * memory is actually populated and not just mapped to the zero |
187 | * page. The prevents expensive copy-on-write faults from |
188 | * occurring during the dirty memory iterations below, which |
189 | * would pollute the performance results. |
190 | */ |
191 | memstress_set_write_percent(vm, 100); |
192 | memstress_set_random_access(vm, false); |
193 | memstress_start_vcpu_threads(nr_vcpus, vcpu_worker); |
194 | |
195 | /* Allow the vCPUs to populate memory */ |
196 | pr_debug("Starting iteration %d - Populating\n" , iteration); |
197 | for (i = 0; i < nr_vcpus; i++) { |
198 | while (READ_ONCE(vcpu_last_completed_iteration[i]) != |
199 | iteration) |
200 | ; |
201 | } |
202 | |
203 | ts_diff = timespec_elapsed(start); |
204 | pr_info("Populate memory time: %ld.%.9lds\n" , |
205 | ts_diff.tv_sec, ts_diff.tv_nsec); |
206 | |
207 | /* Enable dirty logging */ |
208 | clock_gettime(CLOCK_MONOTONIC, &start); |
209 | memstress_enable_dirty_logging(vm, p->slots); |
210 | ts_diff = timespec_elapsed(start); |
211 | pr_info("Enabling dirty logging time: %ld.%.9lds\n\n" , |
212 | ts_diff.tv_sec, ts_diff.tv_nsec); |
213 | |
214 | memstress_set_write_percent(vm, p->write_percent); |
215 | memstress_set_random_access(vm, p->random_access); |
216 | |
217 | while (iteration < p->iterations) { |
218 | /* |
219 | * Incrementing the iteration number will start the vCPUs |
220 | * dirtying memory again. |
221 | */ |
222 | clock_gettime(CLOCK_MONOTONIC, &start); |
223 | iteration++; |
224 | |
225 | pr_debug("Starting iteration %d\n" , iteration); |
226 | for (i = 0; i < nr_vcpus; i++) { |
227 | while (READ_ONCE(vcpu_last_completed_iteration[i]) |
228 | != iteration) |
229 | ; |
230 | } |
231 | |
232 | ts_diff = timespec_elapsed(start); |
233 | vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff); |
234 | pr_info("Iteration %d dirty memory time: %ld.%.9lds\n" , |
235 | iteration, ts_diff.tv_sec, ts_diff.tv_nsec); |
236 | |
237 | clock_gettime(CLOCK_MONOTONIC, &start); |
238 | memstress_get_dirty_log(vm, bitmaps, p->slots); |
239 | ts_diff = timespec_elapsed(start); |
240 | get_dirty_log_total = timespec_add(get_dirty_log_total, |
241 | ts_diff); |
242 | pr_info("Iteration %d get dirty log time: %ld.%.9lds\n" , |
243 | iteration, ts_diff.tv_sec, ts_diff.tv_nsec); |
244 | |
245 | if (dirty_log_manual_caps) { |
246 | clock_gettime(CLOCK_MONOTONIC, &start); |
247 | memstress_clear_dirty_log(vm, bitmaps, p->slots, |
248 | pages_per_slot); |
249 | ts_diff = timespec_elapsed(start); |
250 | clear_dirty_log_total = timespec_add(clear_dirty_log_total, |
251 | ts_diff); |
252 | pr_info("Iteration %d clear dirty log time: %ld.%.9lds\n" , |
253 | iteration, ts_diff.tv_sec, ts_diff.tv_nsec); |
254 | } |
255 | } |
256 | |
257 | /* |
258 | * Run vCPUs while dirty logging is being disabled to stress disabling |
259 | * in terms of both performance and correctness. Opt-in via command |
260 | * line as this significantly increases time to disable dirty logging. |
261 | */ |
262 | if (run_vcpus_while_disabling_dirty_logging) |
263 | WRITE_ONCE(iteration, -1); |
264 | |
265 | /* Disable dirty logging */ |
266 | clock_gettime(CLOCK_MONOTONIC, &start); |
267 | memstress_disable_dirty_logging(vm, p->slots); |
268 | ts_diff = timespec_elapsed(start); |
269 | pr_info("Disabling dirty logging time: %ld.%.9lds\n" , |
270 | ts_diff.tv_sec, ts_diff.tv_nsec); |
271 | |
272 | /* |
273 | * Tell the vCPU threads to quit. No need to manually check that vCPUs |
274 | * have stopped running after disabling dirty logging, the join will |
275 | * wait for them to exit. |
276 | */ |
277 | host_quit = true; |
278 | memstress_join_vcpu_threads(nr_vcpus); |
279 | |
280 | avg = timespec_div(get_dirty_log_total, p->iterations); |
281 | pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n" , |
282 | p->iterations, get_dirty_log_total.tv_sec, |
283 | get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec); |
284 | |
285 | if (dirty_log_manual_caps) { |
286 | avg = timespec_div(clear_dirty_log_total, p->iterations); |
287 | pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n" , |
288 | p->iterations, clear_dirty_log_total.tv_sec, |
289 | clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec); |
290 | } |
291 | |
292 | memstress_free_bitmaps(bitmaps, p->slots); |
293 | arch_cleanup_vm(vm); |
294 | memstress_destroy_vm(vm); |
295 | } |
296 | |
297 | static void help(char *name) |
298 | { |
299 | puts("" ); |
300 | printf("usage: %s [-h] [-a] [-i iterations] [-p offset] [-g] " |
301 | "[-m mode] [-n] [-b vcpu bytes] [-v vcpus] [-o] [-r random seed ] [-s mem type]" |
302 | "[-x memslots] [-w percentage] [-c physical cpus to run test on]\n" , name); |
303 | puts("" ); |
304 | printf(" -a: access memory randomly rather than in order.\n" ); |
305 | printf(" -i: specify iteration counts (default: %" PRIu64")\n" , |
306 | TEST_HOST_LOOP_N); |
307 | printf(" -g: Do not enable KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2. This\n" |
308 | " makes KVM_GET_DIRTY_LOG clear the dirty log (i.e.\n" |
309 | " KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE is not enabled)\n" |
310 | " and writes will be tracked as soon as dirty logging is\n" |
311 | " enabled on the memslot (i.e. KVM_DIRTY_LOG_INITIALLY_SET\n" |
312 | " is not enabled).\n" ); |
313 | printf(" -p: specify guest physical test memory offset\n" |
314 | " Warning: a low offset can conflict with the loaded test code.\n" ); |
315 | guest_modes_help(); |
316 | printf(" -n: Run the vCPUs in nested mode (L2)\n" ); |
317 | printf(" -e: Run vCPUs while dirty logging is being disabled. This\n" |
318 | " can significantly increase runtime, especially if there\n" |
319 | " isn't a dedicated pCPU for the main thread.\n" ); |
320 | printf(" -b: specify the size of the memory region which should be\n" |
321 | " dirtied by each vCPU. e.g. 10M or 3G.\n" |
322 | " (default: 1G)\n" ); |
323 | printf(" -v: specify the number of vCPUs to run.\n" ); |
324 | printf(" -o: Overlap guest memory accesses instead of partitioning\n" |
325 | " them into a separate region of memory for each vCPU.\n" ); |
326 | printf(" -r: specify the starting random seed.\n" ); |
327 | backing_src_help("-s" ); |
328 | printf(" -x: Split the memory region into this number of memslots.\n" |
329 | " (default: 1)\n" ); |
330 | printf(" -w: specify the percentage of pages which should be written to\n" |
331 | " as an integer from 0-100 inclusive. This is probabilistic,\n" |
332 | " so -w X means each page has an X%% chance of writing\n" |
333 | " and a (100-X)%% chance of reading.\n" |
334 | " (default: 100 i.e. all pages are written to.)\n" ); |
335 | kvm_print_vcpu_pinning_help(); |
336 | puts("" ); |
337 | exit(0); |
338 | } |
339 | |
340 | int main(int argc, char *argv[]) |
341 | { |
342 | int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS); |
343 | const char *pcpu_list = NULL; |
344 | struct test_params p = { |
345 | .iterations = TEST_HOST_LOOP_N, |
346 | .partition_vcpu_memory_access = true, |
347 | .backing_src = DEFAULT_VM_MEM_SRC, |
348 | .slots = 1, |
349 | .random_seed = 1, |
350 | .write_percent = 100, |
351 | }; |
352 | int opt; |
353 | |
354 | dirty_log_manual_caps = |
355 | kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2); |
356 | dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | |
357 | KVM_DIRTY_LOG_INITIALLY_SET); |
358 | |
359 | guest_modes_append_default(); |
360 | |
361 | while ((opt = getopt(argc, argv, "ab:c:eghi:m:nop:r:s:v:x:w:" )) != -1) { |
362 | switch (opt) { |
363 | case 'a': |
364 | p.random_access = true; |
365 | break; |
366 | case 'b': |
367 | guest_percpu_mem_size = parse_size(optarg); |
368 | break; |
369 | case 'c': |
370 | pcpu_list = optarg; |
371 | break; |
372 | case 'e': |
373 | /* 'e' is for evil. */ |
374 | run_vcpus_while_disabling_dirty_logging = true; |
375 | break; |
376 | case 'g': |
377 | dirty_log_manual_caps = 0; |
378 | break; |
379 | case 'h': |
380 | help(name: argv[0]); |
381 | break; |
382 | case 'i': |
383 | p.iterations = atoi_positive("Number of iterations" , optarg); |
384 | break; |
385 | case 'm': |
386 | guest_modes_cmdline(optarg); |
387 | break; |
388 | case 'n': |
389 | memstress_args.nested = true; |
390 | break; |
391 | case 'o': |
392 | p.partition_vcpu_memory_access = false; |
393 | break; |
394 | case 'p': |
395 | p.phys_offset = strtoull(optarg, NULL, 0); |
396 | break; |
397 | case 'r': |
398 | p.random_seed = atoi_positive("Random seed" , optarg); |
399 | break; |
400 | case 's': |
401 | p.backing_src = parse_backing_src_type(optarg); |
402 | break; |
403 | case 'v': |
404 | nr_vcpus = atoi_positive("Number of vCPUs" , optarg); |
405 | TEST_ASSERT(nr_vcpus <= max_vcpus, |
406 | "Invalid number of vcpus, must be between 1 and %d" , max_vcpus); |
407 | break; |
408 | case 'w': |
409 | p.write_percent = atoi_non_negative("Write percentage" , optarg); |
410 | TEST_ASSERT(p.write_percent <= 100, |
411 | "Write percentage must be between 0 and 100" ); |
412 | break; |
413 | case 'x': |
414 | p.slots = atoi_positive("Number of slots" , optarg); |
415 | break; |
416 | default: |
417 | help(name: argv[0]); |
418 | break; |
419 | } |
420 | } |
421 | |
422 | if (pcpu_list) { |
423 | kvm_parse_vcpu_pinning(pcpu_list, memstress_args.vcpu_to_pcpu, |
424 | nr_vcpus); |
425 | memstress_args.pin_vcpus = true; |
426 | } |
427 | |
428 | TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations" ); |
429 | |
430 | pr_info("Test iterations: %" PRIu64"\n" , p.iterations); |
431 | |
432 | for_each_guest_mode(run_test, &p); |
433 | |
434 | return 0; |
435 | } |
436 | |