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, int slots, unsigned long *bitmap,
122 			  uint64_t nr_pages)
123 {
124 	uint64_t slot_pages = nr_pages / slots;
125 	int i;
126 
127 	for (i = 0; i < slots; i++) {
128 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
129 		unsigned long *slot_bitmap = bitmap + i * slot_pages;
130 
131 		kvm_vm_get_dirty_log(vm, slot, slot_bitmap);
132 	}
133 }
134 
135 static void clear_dirty_log(struct kvm_vm *vm, int slots, unsigned long *bitmap,
136 			    uint64_t nr_pages)
137 {
138 	uint64_t slot_pages = nr_pages / slots;
139 	int i;
140 
141 	for (i = 0; i < slots; i++) {
142 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
143 		unsigned long *slot_bitmap = bitmap + i * slot_pages;
144 
145 		kvm_vm_clear_dirty_log(vm, slot, slot_bitmap, 0, slot_pages);
146 	}
147 }
148 
149 static void run_test(enum vm_guest_mode mode, void *arg)
150 {
151 	struct test_params *p = arg;
152 	pthread_t *vcpu_threads;
153 	struct kvm_vm *vm;
154 	unsigned long *bmap;
155 	uint64_t guest_num_pages;
156 	uint64_t host_num_pages;
157 	int vcpu_id;
158 	struct timespec start;
159 	struct timespec ts_diff;
160 	struct timespec get_dirty_log_total = (struct timespec){0};
161 	struct timespec vcpu_dirty_total = (struct timespec){0};
162 	struct timespec avg;
163 	struct kvm_enable_cap cap = {};
164 	struct timespec clear_dirty_log_total = (struct timespec){0};
165 
166 	vm = perf_test_create_vm(mode, nr_vcpus, guest_percpu_mem_size,
167 				 p->slots, p->backing_src);
168 
169 	perf_test_args.wr_fract = p->wr_fract;
170 
171 	guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm_get_page_shift(vm);
172 	guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
173 	host_num_pages = vm_num_host_pages(mode, guest_num_pages);
174 	bmap = bitmap_zalloc(host_num_pages);
175 
176 	if (dirty_log_manual_caps) {
177 		cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
178 		cap.args[0] = dirty_log_manual_caps;
179 		vm_enable_cap(vm, &cap);
180 	}
181 
182 	vcpu_threads = malloc(nr_vcpus * sizeof(*vcpu_threads));
183 	TEST_ASSERT(vcpu_threads, "Memory allocation failed");
184 
185 	perf_test_setup_vcpus(vm, nr_vcpus, guest_percpu_mem_size,
186 			      p->partition_vcpu_memory_access);
187 
188 	sync_global_to_guest(vm, perf_test_args);
189 
190 	/* Start the iterations */
191 	iteration = 0;
192 	host_quit = false;
193 
194 	clock_gettime(CLOCK_MONOTONIC, &start);
195 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
196 		vcpu_last_completed_iteration[vcpu_id] = -1;
197 
198 		pthread_create(&vcpu_threads[vcpu_id], NULL, vcpu_worker,
199 			       &perf_test_args.vcpu_args[vcpu_id]);
200 	}
201 
202 	/* Allow the vCPUs to populate memory */
203 	pr_debug("Starting iteration %d - Populating\n", iteration);
204 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
205 		while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) !=
206 		       iteration)
207 			;
208 	}
209 
210 	ts_diff = timespec_elapsed(start);
211 	pr_info("Populate memory time: %ld.%.9lds\n",
212 		ts_diff.tv_sec, ts_diff.tv_nsec);
213 
214 	/* Enable dirty logging */
215 	clock_gettime(CLOCK_MONOTONIC, &start);
216 	enable_dirty_logging(vm, p->slots);
217 	ts_diff = timespec_elapsed(start);
218 	pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
219 		ts_diff.tv_sec, ts_diff.tv_nsec);
220 
221 	while (iteration < p->iterations) {
222 		/*
223 		 * Incrementing the iteration number will start the vCPUs
224 		 * dirtying memory again.
225 		 */
226 		clock_gettime(CLOCK_MONOTONIC, &start);
227 		iteration++;
228 
229 		pr_debug("Starting iteration %d\n", iteration);
230 		for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
231 			while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id])
232 			       != iteration)
233 				;
234 		}
235 
236 		ts_diff = timespec_elapsed(start);
237 		vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
238 		pr_info("Iteration %d dirty memory time: %ld.%.9lds\n",
239 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
240 
241 		clock_gettime(CLOCK_MONOTONIC, &start);
242 		get_dirty_log(vm, p->slots, bmap, host_num_pages);
243 		ts_diff = timespec_elapsed(start);
244 		get_dirty_log_total = timespec_add(get_dirty_log_total,
245 						   ts_diff);
246 		pr_info("Iteration %d get dirty log time: %ld.%.9lds\n",
247 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
248 
249 		if (dirty_log_manual_caps) {
250 			clock_gettime(CLOCK_MONOTONIC, &start);
251 			clear_dirty_log(vm, p->slots, bmap, host_num_pages);
252 			ts_diff = timespec_elapsed(start);
253 			clear_dirty_log_total = timespec_add(clear_dirty_log_total,
254 							     ts_diff);
255 			pr_info("Iteration %d clear dirty log time: %ld.%.9lds\n",
256 				iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
257 		}
258 	}
259 
260 	/* Disable dirty logging */
261 	clock_gettime(CLOCK_MONOTONIC, &start);
262 	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 	/* Tell the vcpu thread to quit */
268 	host_quit = true;
269 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++)
270 		pthread_join(vcpu_threads[vcpu_id], NULL);
271 
272 	avg = timespec_div(get_dirty_log_total, p->iterations);
273 	pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
274 		p->iterations, get_dirty_log_total.tv_sec,
275 		get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
276 
277 	if (dirty_log_manual_caps) {
278 		avg = timespec_div(clear_dirty_log_total, p->iterations);
279 		pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
280 			p->iterations, clear_dirty_log_total.tv_sec,
281 			clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
282 	}
283 
284 	free(bmap);
285 	free(vcpu_threads);
286 	perf_test_destroy_vm(vm);
287 }
288 
289 static void help(char *name)
290 {
291 	puts("");
292 	printf("usage: %s [-h] [-i iterations] [-p offset] "
293 	       "[-m mode] [-b vcpu bytes] [-v vcpus] [-o] [-s mem type]"
294 	       "[-x memslots]\n", name);
295 	puts("");
296 	printf(" -i: specify iteration counts (default: %"PRIu64")\n",
297 	       TEST_HOST_LOOP_N);
298 	printf(" -p: specify guest physical test memory offset\n"
299 	       "     Warning: a low offset can conflict with the loaded test code.\n");
300 	guest_modes_help();
301 	printf(" -b: specify the size of the memory region which should be\n"
302 	       "     dirtied by each vCPU. e.g. 10M or 3G.\n"
303 	       "     (default: 1G)\n");
304 	printf(" -f: specify the fraction of pages which should be written to\n"
305 	       "     as opposed to simply read, in the form\n"
306 	       "     1/<fraction of pages to write>.\n"
307 	       "     (default: 1 i.e. all pages are written to.)\n");
308 	printf(" -v: specify the number of vCPUs to run.\n");
309 	printf(" -o: Overlap guest memory accesses instead of partitioning\n"
310 	       "     them into a separate region of memory for each vCPU.\n");
311 	printf(" -s: specify the type of memory that should be used to\n"
312 	       "     back the guest data region.\n\n");
313 	printf(" -x: Split the memory region into this number of memslots.\n"
314 	       "     (default: 1)");
315 	backing_src_help();
316 	puts("");
317 	exit(0);
318 }
319 
320 int main(int argc, char *argv[])
321 {
322 	int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS);
323 	struct test_params p = {
324 		.iterations = TEST_HOST_LOOP_N,
325 		.wr_fract = 1,
326 		.partition_vcpu_memory_access = true,
327 		.backing_src = VM_MEM_SRC_ANONYMOUS,
328 		.slots = 1,
329 	};
330 	int opt;
331 
332 	dirty_log_manual_caps =
333 		kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
334 	dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
335 				  KVM_DIRTY_LOG_INITIALLY_SET);
336 
337 	guest_modes_append_default();
338 
339 	while ((opt = getopt(argc, argv, "hi:p:m:b:f:v:os:x:")) != -1) {
340 		switch (opt) {
341 		case 'i':
342 			p.iterations = atoi(optarg);
343 			break;
344 		case 'p':
345 			p.phys_offset = strtoull(optarg, NULL, 0);
346 			break;
347 		case 'm':
348 			guest_modes_cmdline(optarg);
349 			break;
350 		case 'b':
351 			guest_percpu_mem_size = parse_size(optarg);
352 			break;
353 		case 'f':
354 			p.wr_fract = atoi(optarg);
355 			TEST_ASSERT(p.wr_fract >= 1,
356 				    "Write fraction cannot be less than one");
357 			break;
358 		case 'v':
359 			nr_vcpus = atoi(optarg);
360 			TEST_ASSERT(nr_vcpus > 0 && nr_vcpus <= max_vcpus,
361 				    "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
362 			break;
363 		case 'o':
364 			p.partition_vcpu_memory_access = false;
365 			break;
366 		case 's':
367 			p.backing_src = parse_backing_src_type(optarg);
368 			break;
369 		case 'x':
370 			p.slots = atoi(optarg);
371 			break;
372 		case 'h':
373 		default:
374 			help(argv[0]);
375 			break;
376 		}
377 	}
378 
379 	TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations");
380 
381 	pr_info("Test iterations: %"PRIu64"\n",	p.iterations);
382 
383 	for_each_guest_mode(run_test, &p);
384 
385 	return 0;
386 }
387