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 uint64_t iteration;
32 static uint64_t 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 	vcpu_args_set(vm, vcpu_id, 1, vcpu_id);
48 	run = vcpu_state(vm, vcpu_id);
49 
50 	while (!READ_ONCE(host_quit)) {
51 		uint64_t current_iteration = READ_ONCE(iteration);
52 
53 		clock_gettime(CLOCK_MONOTONIC, &start);
54 		ret = _vcpu_run(vm, vcpu_id);
55 		ts_diff = timespec_diff_now(start);
56 
57 		TEST_ASSERT(ret == 0, "vcpu_run failed: %d\n", ret);
58 		TEST_ASSERT(get_ucall(vm, vcpu_id, NULL) == UCALL_SYNC,
59 			    "Invalid guest sync status: exit_reason=%s\n",
60 			    exit_reason_str(run->exit_reason));
61 
62 		pr_debug("Got sync event from vCPU %d\n", vcpu_id);
63 		vcpu_last_completed_iteration[vcpu_id] = current_iteration;
64 		pr_debug("vCPU %d updated last completed iteration to %lu\n",
65 			 vcpu_id, vcpu_last_completed_iteration[vcpu_id]);
66 
67 		if (current_iteration) {
68 			pages_count += vcpu_args->pages;
69 			total = timespec_add(total, ts_diff);
70 			pr_debug("vCPU %d iteration %lu dirty memory time: %ld.%.9lds\n",
71 				vcpu_id, current_iteration, ts_diff.tv_sec,
72 				ts_diff.tv_nsec);
73 		} else {
74 			pr_debug("vCPU %d iteration %lu populate memory time: %ld.%.9lds\n",
75 				vcpu_id, current_iteration, ts_diff.tv_sec,
76 				ts_diff.tv_nsec);
77 		}
78 
79 		while (current_iteration == READ_ONCE(iteration) &&
80 		       !READ_ONCE(host_quit)) {}
81 	}
82 
83 	avg = timespec_div(total, vcpu_last_completed_iteration[vcpu_id]);
84 	pr_debug("\nvCPU %d dirtied 0x%lx pages over %lu iterations in %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
85 		vcpu_id, pages_count, vcpu_last_completed_iteration[vcpu_id],
86 		total.tv_sec, total.tv_nsec, avg.tv_sec, avg.tv_nsec);
87 
88 	return NULL;
89 }
90 
91 struct test_params {
92 	unsigned long iterations;
93 	uint64_t phys_offset;
94 	int wr_fract;
95 };
96 
97 static void run_test(enum vm_guest_mode mode, void *arg)
98 {
99 	struct test_params *p = arg;
100 	pthread_t *vcpu_threads;
101 	struct kvm_vm *vm;
102 	unsigned long *bmap;
103 	uint64_t guest_num_pages;
104 	uint64_t host_num_pages;
105 	int vcpu_id;
106 	struct timespec start;
107 	struct timespec ts_diff;
108 	struct timespec get_dirty_log_total = (struct timespec){0};
109 	struct timespec vcpu_dirty_total = (struct timespec){0};
110 	struct timespec avg;
111 	struct kvm_enable_cap cap = {};
112 	struct timespec clear_dirty_log_total = (struct timespec){0};
113 
114 	vm = perf_test_create_vm(mode, nr_vcpus, guest_percpu_mem_size);
115 
116 	perf_test_args.wr_fract = p->wr_fract;
117 
118 	guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm_get_page_shift(vm);
119 	guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
120 	host_num_pages = vm_num_host_pages(mode, guest_num_pages);
121 	bmap = bitmap_alloc(host_num_pages);
122 
123 	if (dirty_log_manual_caps) {
124 		cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
125 		cap.args[0] = dirty_log_manual_caps;
126 		vm_enable_cap(vm, &cap);
127 	}
128 
129 	vcpu_threads = malloc(nr_vcpus * sizeof(*vcpu_threads));
130 	TEST_ASSERT(vcpu_threads, "Memory allocation failed");
131 
132 	perf_test_setup_vcpus(vm, nr_vcpus, guest_percpu_mem_size);
133 
134 	sync_global_to_guest(vm, perf_test_args);
135 
136 	/* Start the iterations */
137 	iteration = 0;
138 	host_quit = false;
139 
140 	clock_gettime(CLOCK_MONOTONIC, &start);
141 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
142 		pthread_create(&vcpu_threads[vcpu_id], NULL, vcpu_worker,
143 			       &perf_test_args.vcpu_args[vcpu_id]);
144 	}
145 
146 	/* Allow the vCPU to populate memory */
147 	pr_debug("Starting iteration %lu - Populating\n", iteration);
148 	while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) != iteration)
149 		pr_debug("Waiting for vcpu_last_completed_iteration == %lu\n",
150 			iteration);
151 
152 	ts_diff = timespec_diff_now(start);
153 	pr_info("Populate memory time: %ld.%.9lds\n",
154 		ts_diff.tv_sec, ts_diff.tv_nsec);
155 
156 	/* Enable dirty logging */
157 	clock_gettime(CLOCK_MONOTONIC, &start);
158 	vm_mem_region_set_flags(vm, PERF_TEST_MEM_SLOT_INDEX,
159 				KVM_MEM_LOG_DIRTY_PAGES);
160 	ts_diff = timespec_diff_now(start);
161 	pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
162 		ts_diff.tv_sec, ts_diff.tv_nsec);
163 
164 	while (iteration < p->iterations) {
165 		/*
166 		 * Incrementing the iteration number will start the vCPUs
167 		 * dirtying memory again.
168 		 */
169 		clock_gettime(CLOCK_MONOTONIC, &start);
170 		iteration++;
171 
172 		pr_debug("Starting iteration %lu\n", iteration);
173 		for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
174 			while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) != iteration)
175 				pr_debug("Waiting for vCPU %d vcpu_last_completed_iteration == %lu\n",
176 					 vcpu_id, iteration);
177 		}
178 
179 		ts_diff = timespec_diff_now(start);
180 		vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
181 		pr_info("Iteration %lu dirty memory time: %ld.%.9lds\n",
182 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
183 
184 		clock_gettime(CLOCK_MONOTONIC, &start);
185 		kvm_vm_get_dirty_log(vm, PERF_TEST_MEM_SLOT_INDEX, bmap);
186 
187 		ts_diff = timespec_diff_now(start);
188 		get_dirty_log_total = timespec_add(get_dirty_log_total,
189 						   ts_diff);
190 		pr_info("Iteration %lu get dirty log time: %ld.%.9lds\n",
191 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
192 
193 		if (dirty_log_manual_caps) {
194 			clock_gettime(CLOCK_MONOTONIC, &start);
195 			kvm_vm_clear_dirty_log(vm, PERF_TEST_MEM_SLOT_INDEX, bmap, 0,
196 					       host_num_pages);
197 
198 			ts_diff = timespec_diff_now(start);
199 			clear_dirty_log_total = timespec_add(clear_dirty_log_total,
200 							     ts_diff);
201 			pr_info("Iteration %lu clear dirty log time: %ld.%.9lds\n",
202 				iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
203 		}
204 	}
205 
206 	/* Tell the vcpu thread to quit */
207 	host_quit = true;
208 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++)
209 		pthread_join(vcpu_threads[vcpu_id], NULL);
210 
211 	/* Disable dirty logging */
212 	clock_gettime(CLOCK_MONOTONIC, &start);
213 	vm_mem_region_set_flags(vm, PERF_TEST_MEM_SLOT_INDEX, 0);
214 	ts_diff = timespec_diff_now(start);
215 	pr_info("Disabling dirty logging time: %ld.%.9lds\n",
216 		ts_diff.tv_sec, ts_diff.tv_nsec);
217 
218 	avg = timespec_div(get_dirty_log_total, p->iterations);
219 	pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
220 		p->iterations, get_dirty_log_total.tv_sec,
221 		get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
222 
223 	if (dirty_log_manual_caps) {
224 		avg = timespec_div(clear_dirty_log_total, p->iterations);
225 		pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
226 			p->iterations, clear_dirty_log_total.tv_sec,
227 			clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
228 	}
229 
230 	free(bmap);
231 	free(vcpu_threads);
232 	perf_test_destroy_vm(vm);
233 }
234 
235 static void help(char *name)
236 {
237 	puts("");
238 	printf("usage: %s [-h] [-i iterations] [-p offset] "
239 	       "[-m mode] [-b vcpu bytes] [-v vcpus]\n", name);
240 	puts("");
241 	printf(" -i: specify iteration counts (default: %"PRIu64")\n",
242 	       TEST_HOST_LOOP_N);
243 	printf(" -p: specify guest physical test memory offset\n"
244 	       "     Warning: a low offset can conflict with the loaded test code.\n");
245 	guest_modes_help();
246 	printf(" -b: specify the size of the memory region which should be\n"
247 	       "     dirtied by each vCPU. e.g. 10M or 3G.\n"
248 	       "     (default: 1G)\n");
249 	printf(" -f: specify the fraction of pages which should be written to\n"
250 	       "     as opposed to simply read, in the form\n"
251 	       "     1/<fraction of pages to write>.\n"
252 	       "     (default: 1 i.e. all pages are written to.)\n");
253 	printf(" -v: specify the number of vCPUs to run.\n");
254 	puts("");
255 	exit(0);
256 }
257 
258 int main(int argc, char *argv[])
259 {
260 	int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS);
261 	struct test_params p = {
262 		.iterations = TEST_HOST_LOOP_N,
263 		.wr_fract = 1,
264 	};
265 	int opt;
266 
267 	dirty_log_manual_caps =
268 		kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
269 	dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
270 				  KVM_DIRTY_LOG_INITIALLY_SET);
271 
272 	guest_modes_append_default();
273 
274 	while ((opt = getopt(argc, argv, "hi:p:m:b:f:v:")) != -1) {
275 		switch (opt) {
276 		case 'i':
277 			p.iterations = strtol(optarg, NULL, 10);
278 			break;
279 		case 'p':
280 			p.phys_offset = strtoull(optarg, NULL, 0);
281 			break;
282 		case 'm':
283 			guest_modes_cmdline(optarg);
284 			break;
285 		case 'b':
286 			guest_percpu_mem_size = parse_size(optarg);
287 			break;
288 		case 'f':
289 			p.wr_fract = atoi(optarg);
290 			TEST_ASSERT(p.wr_fract >= 1,
291 				    "Write fraction cannot be less than one");
292 			break;
293 		case 'v':
294 			nr_vcpus = atoi(optarg);
295 			TEST_ASSERT(nr_vcpus > 0 && nr_vcpus <= max_vcpus,
296 				    "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
297 			break;
298 		case 'h':
299 		default:
300 			help(argv[0]);
301 			break;
302 		}
303 	}
304 
305 	TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations");
306 
307 	pr_info("Test iterations: %"PRIu64"\n",	p.iterations);
308 
309 	for_each_guest_mode(run_test, &p);
310 
311 	return 0;
312 }
313