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 static int nr_vcpus = 1;
26 static uint64_t guest_percpu_mem_size = DEFAULT_PER_VCPU_MEM_SIZE;
27 
28 /* Host variables */
29 static u64 dirty_log_manual_caps;
30 static bool host_quit;
31 static int iteration;
32 static int vcpu_last_completed_iteration[KVM_MAX_VCPUS];
33 
34 static void *vcpu_worker(void *data)
35 {
36 	int ret;
37 	struct kvm_vm *vm = perf_test_args.vm;
38 	uint64_t pages_count = 0;
39 	struct kvm_run *run;
40 	struct timespec start;
41 	struct timespec ts_diff;
42 	struct timespec total = (struct timespec){0};
43 	struct timespec avg;
44 	struct perf_test_vcpu_args *vcpu_args = (struct perf_test_vcpu_args *)data;
45 	int vcpu_id = vcpu_args->vcpu_id;
46 
47 	run = vcpu_state(vm, vcpu_id);
48 
49 	while (!READ_ONCE(host_quit)) {
50 		int current_iteration = READ_ONCE(iteration);
51 
52 		clock_gettime(CLOCK_MONOTONIC, &start);
53 		ret = _vcpu_run(vm, vcpu_id);
54 		ts_diff = timespec_elapsed(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 %d\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 %d 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 %d 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 %d 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 	bool partition_vcpu_memory_access;
95 	enum vm_mem_backing_src_type backing_src;
96 	int slots;
97 };
98 
99 static void toggle_dirty_logging(struct kvm_vm *vm, int slots, bool enable)
100 {
101 	int i;
102 
103 	for (i = 0; i < slots; i++) {
104 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
105 		int flags = enable ? KVM_MEM_LOG_DIRTY_PAGES : 0;
106 
107 		vm_mem_region_set_flags(vm, slot, flags);
108 	}
109 }
110 
111 static inline void enable_dirty_logging(struct kvm_vm *vm, int slots)
112 {
113 	toggle_dirty_logging(vm, slots, true);
114 }
115 
116 static inline void disable_dirty_logging(struct kvm_vm *vm, int slots)
117 {
118 	toggle_dirty_logging(vm, slots, false);
119 }
120 
121 static void get_dirty_log(struct kvm_vm *vm, unsigned long *bitmaps[], int slots)
122 {
123 	int i;
124 
125 	for (i = 0; i < slots; i++) {
126 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
127 
128 		kvm_vm_get_dirty_log(vm, slot, bitmaps[i]);
129 	}
130 }
131 
132 static void clear_dirty_log(struct kvm_vm *vm, unsigned long *bitmaps[],
133 			    int slots, uint64_t pages_per_slot)
134 {
135 	int i;
136 
137 	for (i = 0; i < slots; i++) {
138 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
139 
140 		kvm_vm_clear_dirty_log(vm, slot, bitmaps[i], 0, pages_per_slot);
141 	}
142 }
143 
144 static unsigned long **alloc_bitmaps(int slots, uint64_t pages_per_slot)
145 {
146 	unsigned long **bitmaps;
147 	int i;
148 
149 	bitmaps = malloc(slots * sizeof(bitmaps[0]));
150 	TEST_ASSERT(bitmaps, "Failed to allocate bitmaps array.");
151 
152 	for (i = 0; i < slots; i++) {
153 		bitmaps[i] = bitmap_zalloc(pages_per_slot);
154 		TEST_ASSERT(bitmaps[i], "Failed to allocate slot bitmap.");
155 	}
156 
157 	return bitmaps;
158 }
159 
160 static void free_bitmaps(unsigned long *bitmaps[], int slots)
161 {
162 	int i;
163 
164 	for (i = 0; i < slots; i++)
165 		free(bitmaps[i]);
166 
167 	free(bitmaps);
168 }
169 
170 static void run_test(enum vm_guest_mode mode, void *arg)
171 {
172 	struct test_params *p = arg;
173 	pthread_t *vcpu_threads;
174 	struct kvm_vm *vm;
175 	unsigned long **bitmaps;
176 	uint64_t guest_num_pages;
177 	uint64_t host_num_pages;
178 	uint64_t pages_per_slot;
179 	int vcpu_id;
180 	struct timespec start;
181 	struct timespec ts_diff;
182 	struct timespec get_dirty_log_total = (struct timespec){0};
183 	struct timespec vcpu_dirty_total = (struct timespec){0};
184 	struct timespec avg;
185 	struct kvm_enable_cap cap = {};
186 	struct timespec clear_dirty_log_total = (struct timespec){0};
187 
188 	vm = perf_test_create_vm(mode, nr_vcpus, guest_percpu_mem_size,
189 				 p->slots, p->backing_src,
190 				 p->partition_vcpu_memory_access);
191 
192 	perf_test_args.wr_fract = p->wr_fract;
193 
194 	guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm_get_page_shift(vm);
195 	guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
196 	host_num_pages = vm_num_host_pages(mode, guest_num_pages);
197 	pages_per_slot = host_num_pages / p->slots;
198 
199 	bitmaps = alloc_bitmaps(p->slots, pages_per_slot);
200 
201 	if (dirty_log_manual_caps) {
202 		cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
203 		cap.args[0] = dirty_log_manual_caps;
204 		vm_enable_cap(vm, &cap);
205 	}
206 
207 	vcpu_threads = malloc(nr_vcpus * sizeof(*vcpu_threads));
208 	TEST_ASSERT(vcpu_threads, "Memory allocation failed");
209 
210 	sync_global_to_guest(vm, perf_test_args);
211 
212 	/* Start the iterations */
213 	iteration = 0;
214 	host_quit = false;
215 
216 	clock_gettime(CLOCK_MONOTONIC, &start);
217 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
218 		vcpu_last_completed_iteration[vcpu_id] = -1;
219 
220 		pthread_create(&vcpu_threads[vcpu_id], NULL, vcpu_worker,
221 			       &perf_test_args.vcpu_args[vcpu_id]);
222 	}
223 
224 	/* Allow the vCPUs to populate memory */
225 	pr_debug("Starting iteration %d - Populating\n", iteration);
226 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
227 		while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) !=
228 		       iteration)
229 			;
230 	}
231 
232 	ts_diff = timespec_elapsed(start);
233 	pr_info("Populate memory time: %ld.%.9lds\n",
234 		ts_diff.tv_sec, ts_diff.tv_nsec);
235 
236 	/* Enable dirty logging */
237 	clock_gettime(CLOCK_MONOTONIC, &start);
238 	enable_dirty_logging(vm, p->slots);
239 	ts_diff = timespec_elapsed(start);
240 	pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
241 		ts_diff.tv_sec, ts_diff.tv_nsec);
242 
243 	while (iteration < p->iterations) {
244 		/*
245 		 * Incrementing the iteration number will start the vCPUs
246 		 * dirtying memory again.
247 		 */
248 		clock_gettime(CLOCK_MONOTONIC, &start);
249 		iteration++;
250 
251 		pr_debug("Starting iteration %d\n", iteration);
252 		for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
253 			while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id])
254 			       != iteration)
255 				;
256 		}
257 
258 		ts_diff = timespec_elapsed(start);
259 		vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
260 		pr_info("Iteration %d dirty memory time: %ld.%.9lds\n",
261 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
262 
263 		clock_gettime(CLOCK_MONOTONIC, &start);
264 		get_dirty_log(vm, bitmaps, p->slots);
265 		ts_diff = timespec_elapsed(start);
266 		get_dirty_log_total = timespec_add(get_dirty_log_total,
267 						   ts_diff);
268 		pr_info("Iteration %d get dirty log time: %ld.%.9lds\n",
269 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
270 
271 		if (dirty_log_manual_caps) {
272 			clock_gettime(CLOCK_MONOTONIC, &start);
273 			clear_dirty_log(vm, bitmaps, p->slots, pages_per_slot);
274 			ts_diff = timespec_elapsed(start);
275 			clear_dirty_log_total = timespec_add(clear_dirty_log_total,
276 							     ts_diff);
277 			pr_info("Iteration %d clear dirty log time: %ld.%.9lds\n",
278 				iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
279 		}
280 	}
281 
282 	/* Disable dirty logging */
283 	clock_gettime(CLOCK_MONOTONIC, &start);
284 	disable_dirty_logging(vm, p->slots);
285 	ts_diff = timespec_elapsed(start);
286 	pr_info("Disabling dirty logging time: %ld.%.9lds\n",
287 		ts_diff.tv_sec, ts_diff.tv_nsec);
288 
289 	/* Tell the vcpu thread to quit */
290 	host_quit = true;
291 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++)
292 		pthread_join(vcpu_threads[vcpu_id], NULL);
293 
294 	avg = timespec_div(get_dirty_log_total, p->iterations);
295 	pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
296 		p->iterations, get_dirty_log_total.tv_sec,
297 		get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
298 
299 	if (dirty_log_manual_caps) {
300 		avg = timespec_div(clear_dirty_log_total, p->iterations);
301 		pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
302 			p->iterations, clear_dirty_log_total.tv_sec,
303 			clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
304 	}
305 
306 	free_bitmaps(bitmaps, p->slots);
307 	free(vcpu_threads);
308 	perf_test_destroy_vm(vm);
309 }
310 
311 static void help(char *name)
312 {
313 	puts("");
314 	printf("usage: %s [-h] [-i iterations] [-p offset] "
315 	       "[-m mode] [-b vcpu bytes] [-v vcpus] [-o] [-s mem type]"
316 	       "[-x memslots]\n", name);
317 	puts("");
318 	printf(" -i: specify iteration counts (default: %"PRIu64")\n",
319 	       TEST_HOST_LOOP_N);
320 	printf(" -p: specify guest physical test memory offset\n"
321 	       "     Warning: a low offset can conflict with the loaded test code.\n");
322 	guest_modes_help();
323 	printf(" -b: specify the size of the memory region which should be\n"
324 	       "     dirtied by each vCPU. e.g. 10M or 3G.\n"
325 	       "     (default: 1G)\n");
326 	printf(" -f: specify the fraction of pages which should be written to\n"
327 	       "     as opposed to simply read, in the form\n"
328 	       "     1/<fraction of pages to write>.\n"
329 	       "     (default: 1 i.e. all pages are written to.)\n");
330 	printf(" -v: specify the number of vCPUs to run.\n");
331 	printf(" -o: Overlap guest memory accesses instead of partitioning\n"
332 	       "     them into a separate region of memory for each vCPU.\n");
333 	backing_src_help("-s");
334 	printf(" -x: Split the memory region into this number of memslots.\n"
335 	       "     (default: 1)\n");
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 	struct test_params p = {
344 		.iterations = TEST_HOST_LOOP_N,
345 		.wr_fract = 1,
346 		.partition_vcpu_memory_access = true,
347 		.backing_src = DEFAULT_VM_MEM_SRC,
348 		.slots = 1,
349 	};
350 	int opt;
351 
352 	dirty_log_manual_caps =
353 		kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
354 	dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
355 				  KVM_DIRTY_LOG_INITIALLY_SET);
356 
357 	guest_modes_append_default();
358 
359 	while ((opt = getopt(argc, argv, "hi:p:m:b:f:v:os:x:")) != -1) {
360 		switch (opt) {
361 		case 'i':
362 			p.iterations = atoi(optarg);
363 			break;
364 		case 'p':
365 			p.phys_offset = strtoull(optarg, NULL, 0);
366 			break;
367 		case 'm':
368 			guest_modes_cmdline(optarg);
369 			break;
370 		case 'b':
371 			guest_percpu_mem_size = parse_size(optarg);
372 			break;
373 		case 'f':
374 			p.wr_fract = atoi(optarg);
375 			TEST_ASSERT(p.wr_fract >= 1,
376 				    "Write fraction cannot be less than one");
377 			break;
378 		case 'v':
379 			nr_vcpus = atoi(optarg);
380 			TEST_ASSERT(nr_vcpus > 0 && nr_vcpus <= max_vcpus,
381 				    "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
382 			break;
383 		case 'o':
384 			p.partition_vcpu_memory_access = false;
385 			break;
386 		case 's':
387 			p.backing_src = parse_backing_src_type(optarg);
388 			break;
389 		case 'x':
390 			p.slots = atoi(optarg);
391 			break;
392 		case 'h':
393 		default:
394 			help(argv[0]);
395 			break;
396 		}
397 	}
398 
399 	TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations");
400 
401 	pr_info("Test iterations: %"PRIu64"\n",	p.iterations);
402 
403 	for_each_guest_mode(run_test, &p);
404 
405 	return 0;
406 }
407