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 	vcpu_args_set(vm, vcpu_id, 1, vcpu_id);
48 	run = vcpu_state(vm, vcpu_id);
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(vm, vcpu_id);
55 		ts_diff = timespec_elapsed(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 %d\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 %d 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 %d 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 %d 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 	bool partition_vcpu_memory_access;
96 	enum vm_mem_backing_src_type backing_src;
97 };
98 
99 static void run_test(enum vm_guest_mode mode, void *arg)
100 {
101 	struct test_params *p = arg;
102 	pthread_t *vcpu_threads;
103 	struct kvm_vm *vm;
104 	unsigned long *bmap;
105 	uint64_t guest_num_pages;
106 	uint64_t host_num_pages;
107 	int vcpu_id;
108 	struct timespec start;
109 	struct timespec ts_diff;
110 	struct timespec get_dirty_log_total = (struct timespec){0};
111 	struct timespec vcpu_dirty_total = (struct timespec){0};
112 	struct timespec avg;
113 	struct kvm_enable_cap cap = {};
114 	struct timespec clear_dirty_log_total = (struct timespec){0};
115 
116 	vm = perf_test_create_vm(mode, nr_vcpus, guest_percpu_mem_size,
117 				 p->backing_src);
118 
119 	perf_test_args.wr_fract = p->wr_fract;
120 
121 	guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm_get_page_shift(vm);
122 	guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
123 	host_num_pages = vm_num_host_pages(mode, guest_num_pages);
124 	bmap = bitmap_alloc(host_num_pages);
125 
126 	if (dirty_log_manual_caps) {
127 		cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
128 		cap.args[0] = dirty_log_manual_caps;
129 		vm_enable_cap(vm, &cap);
130 	}
131 
132 	vcpu_threads = malloc(nr_vcpus * sizeof(*vcpu_threads));
133 	TEST_ASSERT(vcpu_threads, "Memory allocation failed");
134 
135 	perf_test_setup_vcpus(vm, nr_vcpus, guest_percpu_mem_size,
136 			      p->partition_vcpu_memory_access);
137 
138 	sync_global_to_guest(vm, perf_test_args);
139 
140 	/* Start the iterations */
141 	iteration = 0;
142 	host_quit = false;
143 
144 	clock_gettime(CLOCK_MONOTONIC, &start);
145 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
146 		vcpu_last_completed_iteration[vcpu_id] = -1;
147 
148 		pthread_create(&vcpu_threads[vcpu_id], NULL, vcpu_worker,
149 			       &perf_test_args.vcpu_args[vcpu_id]);
150 	}
151 
152 	/* Allow the vCPUs to populate memory */
153 	pr_debug("Starting iteration %d - Populating\n", iteration);
154 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
155 		while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) !=
156 		       iteration)
157 			;
158 	}
159 
160 	ts_diff = timespec_elapsed(start);
161 	pr_info("Populate memory time: %ld.%.9lds\n",
162 		ts_diff.tv_sec, ts_diff.tv_nsec);
163 
164 	/* Enable dirty logging */
165 	clock_gettime(CLOCK_MONOTONIC, &start);
166 	vm_mem_region_set_flags(vm, PERF_TEST_MEM_SLOT_INDEX,
167 				KVM_MEM_LOG_DIRTY_PAGES);
168 	ts_diff = timespec_elapsed(start);
169 	pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
170 		ts_diff.tv_sec, ts_diff.tv_nsec);
171 
172 	while (iteration < p->iterations) {
173 		/*
174 		 * Incrementing the iteration number will start the vCPUs
175 		 * dirtying memory again.
176 		 */
177 		clock_gettime(CLOCK_MONOTONIC, &start);
178 		iteration++;
179 
180 		pr_debug("Starting iteration %d\n", iteration);
181 		for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
182 			while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id])
183 			       != iteration)
184 				;
185 		}
186 
187 		ts_diff = timespec_elapsed(start);
188 		vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
189 		pr_info("Iteration %d dirty memory time: %ld.%.9lds\n",
190 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
191 
192 		clock_gettime(CLOCK_MONOTONIC, &start);
193 		kvm_vm_get_dirty_log(vm, PERF_TEST_MEM_SLOT_INDEX, bmap);
194 
195 		ts_diff = timespec_elapsed(start);
196 		get_dirty_log_total = timespec_add(get_dirty_log_total,
197 						   ts_diff);
198 		pr_info("Iteration %d get dirty log time: %ld.%.9lds\n",
199 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
200 
201 		if (dirty_log_manual_caps) {
202 			clock_gettime(CLOCK_MONOTONIC, &start);
203 			kvm_vm_clear_dirty_log(vm, PERF_TEST_MEM_SLOT_INDEX, bmap, 0,
204 					       host_num_pages);
205 
206 			ts_diff = timespec_elapsed(start);
207 			clear_dirty_log_total = timespec_add(clear_dirty_log_total,
208 							     ts_diff);
209 			pr_info("Iteration %d clear dirty log time: %ld.%.9lds\n",
210 				iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
211 		}
212 	}
213 
214 	/* Disable dirty logging */
215 	clock_gettime(CLOCK_MONOTONIC, &start);
216 	vm_mem_region_set_flags(vm, PERF_TEST_MEM_SLOT_INDEX, 0);
217 	ts_diff = timespec_elapsed(start);
218 	pr_info("Disabling dirty logging time: %ld.%.9lds\n",
219 		ts_diff.tv_sec, ts_diff.tv_nsec);
220 
221 	/* Tell the vcpu thread to quit */
222 	host_quit = true;
223 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++)
224 		pthread_join(vcpu_threads[vcpu_id], NULL);
225 
226 	avg = timespec_div(get_dirty_log_total, p->iterations);
227 	pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
228 		p->iterations, get_dirty_log_total.tv_sec,
229 		get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
230 
231 	if (dirty_log_manual_caps) {
232 		avg = timespec_div(clear_dirty_log_total, p->iterations);
233 		pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
234 			p->iterations, clear_dirty_log_total.tv_sec,
235 			clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
236 	}
237 
238 	free(bmap);
239 	free(vcpu_threads);
240 	perf_test_destroy_vm(vm);
241 }
242 
243 static void help(char *name)
244 {
245 	puts("");
246 	printf("usage: %s [-h] [-i iterations] [-p offset] "
247 	       "[-m mode] [-b vcpu bytes] [-v vcpus] [-o] [-s mem type]\n", name);
248 	puts("");
249 	printf(" -i: specify iteration counts (default: %"PRIu64")\n",
250 	       TEST_HOST_LOOP_N);
251 	printf(" -p: specify guest physical test memory offset\n"
252 	       "     Warning: a low offset can conflict with the loaded test code.\n");
253 	guest_modes_help();
254 	printf(" -b: specify the size of the memory region which should be\n"
255 	       "     dirtied by each vCPU. e.g. 10M or 3G.\n"
256 	       "     (default: 1G)\n");
257 	printf(" -f: specify the fraction of pages which should be written to\n"
258 	       "     as opposed to simply read, in the form\n"
259 	       "     1/<fraction of pages to write>.\n"
260 	       "     (default: 1 i.e. all pages are written to.)\n");
261 	printf(" -v: specify the number of vCPUs to run.\n");
262 	printf(" -o: Overlap guest memory accesses instead of partitioning\n"
263 	       "     them into a separate region of memory for each vCPU.\n");
264 	printf(" -s: specify the type of memory that should be used to\n"
265 	       "     back the guest data region.\n\n");
266 	backing_src_help();
267 	puts("");
268 	exit(0);
269 }
270 
271 int main(int argc, char *argv[])
272 {
273 	int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS);
274 	struct test_params p = {
275 		.iterations = TEST_HOST_LOOP_N,
276 		.wr_fract = 1,
277 		.partition_vcpu_memory_access = true,
278 		.backing_src = VM_MEM_SRC_ANONYMOUS,
279 	};
280 	int opt;
281 
282 	dirty_log_manual_caps =
283 		kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
284 	dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
285 				  KVM_DIRTY_LOG_INITIALLY_SET);
286 
287 	guest_modes_append_default();
288 
289 	while ((opt = getopt(argc, argv, "hi:p:m:b:f:v:os:")) != -1) {
290 		switch (opt) {
291 		case 'i':
292 			p.iterations = atoi(optarg);
293 			break;
294 		case 'p':
295 			p.phys_offset = strtoull(optarg, NULL, 0);
296 			break;
297 		case 'm':
298 			guest_modes_cmdline(optarg);
299 			break;
300 		case 'b':
301 			guest_percpu_mem_size = parse_size(optarg);
302 			break;
303 		case 'f':
304 			p.wr_fract = atoi(optarg);
305 			TEST_ASSERT(p.wr_fract >= 1,
306 				    "Write fraction cannot be less than one");
307 			break;
308 		case 'v':
309 			nr_vcpus = atoi(optarg);
310 			TEST_ASSERT(nr_vcpus > 0 && nr_vcpus <= max_vcpus,
311 				    "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
312 			break;
313 		case 'o':
314 			p.partition_vcpu_memory_access = false;
315 		case 's':
316 			p.backing_src = parse_backing_src_type(optarg);
317 			break;
318 		case 'h':
319 		default:
320 			help(argv[0]);
321 			break;
322 		}
323 	}
324 
325 	TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations");
326 
327 	pr_info("Test iterations: %"PRIu64"\n",	p.iterations);
328 
329 	for_each_guest_mode(run_test, &p);
330 
331 	return 0;
332 }
333