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 #define _GNU_SOURCE /* for program_invocation_name */
12 
13 #include <stdio.h>
14 #include <stdlib.h>
15 #include <unistd.h>
16 #include <time.h>
17 #include <pthread.h>
18 #include <linux/bitmap.h>
19 #include <linux/bitops.h>
20 
21 #include "kvm_util.h"
22 #include "perf_test_util.h"
23 #include "processor.h"
24 #include "test_util.h"
25 
26 /* How many host loops to run by default (one KVM_GET_DIRTY_LOG for each loop)*/
27 #define TEST_HOST_LOOP_N		2UL
28 
29 /* Host variables */
30 static u64 dirty_log_manual_caps;
31 static bool host_quit;
32 static uint64_t iteration;
33 static uint64_t vcpu_last_completed_iteration[MAX_VCPUS];
34 
35 static void *vcpu_worker(void *data)
36 {
37 	int ret;
38 	struct kvm_vm *vm = perf_test_args.vm;
39 	uint64_t pages_count = 0;
40 	struct kvm_run *run;
41 	struct timespec start;
42 	struct timespec ts_diff;
43 	struct timespec total = (struct timespec){0};
44 	struct timespec avg;
45 	struct vcpu_args *vcpu_args = (struct vcpu_args *)data;
46 	int vcpu_id = vcpu_args->vcpu_id;
47 
48 	vcpu_args_set(vm, vcpu_id, 1, vcpu_id);
49 	run = vcpu_state(vm, vcpu_id);
50 
51 	while (!READ_ONCE(host_quit)) {
52 		uint64_t current_iteration = READ_ONCE(iteration);
53 
54 		clock_gettime(CLOCK_MONOTONIC, &start);
55 		ret = _vcpu_run(vm, vcpu_id);
56 		ts_diff = timespec_diff_now(start);
57 
58 		TEST_ASSERT(ret == 0, "vcpu_run failed: %d\n", ret);
59 		TEST_ASSERT(get_ucall(vm, vcpu_id, NULL) == UCALL_SYNC,
60 			    "Invalid guest sync status: exit_reason=%s\n",
61 			    exit_reason_str(run->exit_reason));
62 
63 		pr_debug("Got sync event from vCPU %d\n", vcpu_id);
64 		vcpu_last_completed_iteration[vcpu_id] = current_iteration;
65 		pr_debug("vCPU %d updated last completed iteration to %lu\n",
66 			 vcpu_id, vcpu_last_completed_iteration[vcpu_id]);
67 
68 		if (current_iteration) {
69 			pages_count += vcpu_args->pages;
70 			total = timespec_add(total, ts_diff);
71 			pr_debug("vCPU %d iteration %lu dirty memory time: %ld.%.9lds\n",
72 				vcpu_id, current_iteration, ts_diff.tv_sec,
73 				ts_diff.tv_nsec);
74 		} else {
75 			pr_debug("vCPU %d iteration %lu populate memory time: %ld.%.9lds\n",
76 				vcpu_id, current_iteration, ts_diff.tv_sec,
77 				ts_diff.tv_nsec);
78 		}
79 
80 		while (current_iteration == READ_ONCE(iteration) &&
81 		       !READ_ONCE(host_quit)) {}
82 	}
83 
84 	avg = timespec_div(total, vcpu_last_completed_iteration[vcpu_id]);
85 	pr_debug("\nvCPU %d dirtied 0x%lx pages over %lu iterations in %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
86 		vcpu_id, pages_count, vcpu_last_completed_iteration[vcpu_id],
87 		total.tv_sec, total.tv_nsec, avg.tv_sec, avg.tv_nsec);
88 
89 	return NULL;
90 }
91 
92 static void run_test(enum vm_guest_mode mode, unsigned long iterations,
93 		     uint64_t phys_offset, int wr_fract)
94 {
95 	pthread_t *vcpu_threads;
96 	struct kvm_vm *vm;
97 	unsigned long *bmap;
98 	uint64_t guest_num_pages;
99 	uint64_t host_num_pages;
100 	int vcpu_id;
101 	struct timespec start;
102 	struct timespec ts_diff;
103 	struct timespec get_dirty_log_total = (struct timespec){0};
104 	struct timespec vcpu_dirty_total = (struct timespec){0};
105 	struct timespec avg;
106 	struct kvm_enable_cap cap = {};
107 	struct timespec clear_dirty_log_total = (struct timespec){0};
108 
109 	vm = create_vm(mode, nr_vcpus, guest_percpu_mem_size);
110 
111 	perf_test_args.wr_fract = wr_fract;
112 
113 	guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm_get_page_shift(vm);
114 	guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
115 	host_num_pages = vm_num_host_pages(mode, guest_num_pages);
116 	bmap = bitmap_alloc(host_num_pages);
117 
118 	if (dirty_log_manual_caps) {
119 		cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
120 		cap.args[0] = dirty_log_manual_caps;
121 		vm_enable_cap(vm, &cap);
122 	}
123 
124 	vcpu_threads = malloc(nr_vcpus * sizeof(*vcpu_threads));
125 	TEST_ASSERT(vcpu_threads, "Memory allocation failed");
126 
127 	add_vcpus(vm, nr_vcpus, guest_percpu_mem_size);
128 
129 	sync_global_to_guest(vm, perf_test_args);
130 
131 	/* Start the iterations */
132 	iteration = 0;
133 	host_quit = false;
134 
135 	clock_gettime(CLOCK_MONOTONIC, &start);
136 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
137 		pthread_create(&vcpu_threads[vcpu_id], NULL, vcpu_worker,
138 			       &perf_test_args.vcpu_args[vcpu_id]);
139 	}
140 
141 	/* Allow the vCPU to populate memory */
142 	pr_debug("Starting iteration %lu - Populating\n", iteration);
143 	while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) != iteration)
144 		pr_debug("Waiting for vcpu_last_completed_iteration == %lu\n",
145 			iteration);
146 
147 	ts_diff = timespec_diff_now(start);
148 	pr_info("Populate memory time: %ld.%.9lds\n",
149 		ts_diff.tv_sec, ts_diff.tv_nsec);
150 
151 	/* Enable dirty logging */
152 	clock_gettime(CLOCK_MONOTONIC, &start);
153 	vm_mem_region_set_flags(vm, TEST_MEM_SLOT_INDEX,
154 				KVM_MEM_LOG_DIRTY_PAGES);
155 	ts_diff = timespec_diff_now(start);
156 	pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
157 		ts_diff.tv_sec, ts_diff.tv_nsec);
158 
159 	while (iteration < iterations) {
160 		/*
161 		 * Incrementing the iteration number will start the vCPUs
162 		 * dirtying memory again.
163 		 */
164 		clock_gettime(CLOCK_MONOTONIC, &start);
165 		iteration++;
166 
167 		pr_debug("Starting iteration %lu\n", iteration);
168 		for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
169 			while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) != iteration)
170 				pr_debug("Waiting for vCPU %d vcpu_last_completed_iteration == %lu\n",
171 					 vcpu_id, iteration);
172 		}
173 
174 		ts_diff = timespec_diff_now(start);
175 		vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
176 		pr_info("Iteration %lu dirty memory time: %ld.%.9lds\n",
177 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
178 
179 		clock_gettime(CLOCK_MONOTONIC, &start);
180 		kvm_vm_get_dirty_log(vm, TEST_MEM_SLOT_INDEX, bmap);
181 
182 		ts_diff = timespec_diff_now(start);
183 		get_dirty_log_total = timespec_add(get_dirty_log_total,
184 						   ts_diff);
185 		pr_info("Iteration %lu get dirty log time: %ld.%.9lds\n",
186 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
187 
188 		if (dirty_log_manual_caps) {
189 			clock_gettime(CLOCK_MONOTONIC, &start);
190 			kvm_vm_clear_dirty_log(vm, TEST_MEM_SLOT_INDEX, bmap, 0,
191 					       host_num_pages);
192 
193 			ts_diff = timespec_diff_now(start);
194 			clear_dirty_log_total = timespec_add(clear_dirty_log_total,
195 							     ts_diff);
196 			pr_info("Iteration %lu clear dirty log time: %ld.%.9lds\n",
197 				iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
198 		}
199 	}
200 
201 	/* Tell the vcpu thread to quit */
202 	host_quit = true;
203 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++)
204 		pthread_join(vcpu_threads[vcpu_id], NULL);
205 
206 	/* Disable dirty logging */
207 	clock_gettime(CLOCK_MONOTONIC, &start);
208 	vm_mem_region_set_flags(vm, TEST_MEM_SLOT_INDEX, 0);
209 	ts_diff = timespec_diff_now(start);
210 	pr_info("Disabling dirty logging time: %ld.%.9lds\n",
211 		ts_diff.tv_sec, ts_diff.tv_nsec);
212 
213 	avg = timespec_div(get_dirty_log_total, iterations);
214 	pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
215 		iterations, get_dirty_log_total.tv_sec,
216 		get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
217 
218 	if (dirty_log_manual_caps) {
219 		avg = timespec_div(clear_dirty_log_total, iterations);
220 		pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
221 			iterations, clear_dirty_log_total.tv_sec,
222 			clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
223 	}
224 
225 	free(bmap);
226 	free(vcpu_threads);
227 	ucall_uninit(vm);
228 	kvm_vm_free(vm);
229 }
230 
231 struct guest_mode {
232 	bool supported;
233 	bool enabled;
234 };
235 static struct guest_mode guest_modes[NUM_VM_MODES];
236 
237 #define guest_mode_init(mode, supported, enabled) ({ \
238 	guest_modes[mode] = (struct guest_mode){ supported, enabled }; \
239 })
240 
241 static void help(char *name)
242 {
243 	int i;
244 
245 	puts("");
246 	printf("usage: %s [-h] [-i iterations] [-p offset] "
247 	       "[-m mode] [-b vcpu bytes] [-v vcpus]\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 	printf(" -m: specify the guest mode ID to test "
254 	       "(default: test all supported modes)\n"
255 	       "     This option may be used multiple times.\n"
256 	       "     Guest mode IDs:\n");
257 	for (i = 0; i < NUM_VM_MODES; ++i) {
258 		printf("         %d:    %s%s\n", i, vm_guest_mode_string(i),
259 		       guest_modes[i].supported ? " (supported)" : "");
260 	}
261 	printf(" -b: specify the size of the memory region which should be\n"
262 	       "     dirtied by each vCPU. e.g. 10M or 3G.\n"
263 	       "     (default: 1G)\n");
264 	printf(" -f: specify the fraction of pages which should be written to\n"
265 	       "     as opposed to simply read, in the form\n"
266 	       "     1/<fraction of pages to write>.\n"
267 	       "     (default: 1 i.e. all pages are written to.)\n");
268 	printf(" -v: specify the number of vCPUs to run.\n");
269 	puts("");
270 	exit(0);
271 }
272 
273 int main(int argc, char *argv[])
274 {
275 	unsigned long iterations = TEST_HOST_LOOP_N;
276 	bool mode_selected = false;
277 	uint64_t phys_offset = 0;
278 	unsigned int mode;
279 	int opt, i;
280 	int wr_fract = 1;
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 #ifdef __x86_64__
288 	guest_mode_init(VM_MODE_PXXV48_4K, true, true);
289 #endif
290 #ifdef __aarch64__
291 	guest_mode_init(VM_MODE_P40V48_4K, true, true);
292 	guest_mode_init(VM_MODE_P40V48_64K, true, true);
293 
294 	{
295 		unsigned int limit = kvm_check_cap(KVM_CAP_ARM_VM_IPA_SIZE);
296 
297 		if (limit >= 52)
298 			guest_mode_init(VM_MODE_P52V48_64K, true, true);
299 		if (limit >= 48) {
300 			guest_mode_init(VM_MODE_P48V48_4K, true, true);
301 			guest_mode_init(VM_MODE_P48V48_64K, true, true);
302 		}
303 	}
304 #endif
305 #ifdef __s390x__
306 	guest_mode_init(VM_MODE_P40V48_4K, true, true);
307 #endif
308 
309 	while ((opt = getopt(argc, argv, "hi:p:m:b:f:v:")) != -1) {
310 		switch (opt) {
311 		case 'i':
312 			iterations = strtol(optarg, NULL, 10);
313 			break;
314 		case 'p':
315 			phys_offset = strtoull(optarg, NULL, 0);
316 			break;
317 		case 'm':
318 			if (!mode_selected) {
319 				for (i = 0; i < NUM_VM_MODES; ++i)
320 					guest_modes[i].enabled = false;
321 				mode_selected = true;
322 			}
323 			mode = strtoul(optarg, NULL, 10);
324 			TEST_ASSERT(mode < NUM_VM_MODES,
325 				    "Guest mode ID %d too big", mode);
326 			guest_modes[mode].enabled = true;
327 			break;
328 		case 'b':
329 			guest_percpu_mem_size = parse_size(optarg);
330 			break;
331 		case 'f':
332 			wr_fract = atoi(optarg);
333 			TEST_ASSERT(wr_fract >= 1,
334 				    "Write fraction cannot be less than one");
335 			break;
336 		case 'v':
337 			nr_vcpus = atoi(optarg);
338 			TEST_ASSERT(nr_vcpus > 0,
339 				    "Must have a positive number of vCPUs");
340 			TEST_ASSERT(nr_vcpus <= MAX_VCPUS,
341 				    "This test does not currently support\n"
342 				    "more than %d vCPUs.", MAX_VCPUS);
343 			break;
344 		case 'h':
345 		default:
346 			help(argv[0]);
347 			break;
348 		}
349 	}
350 
351 	TEST_ASSERT(iterations >= 2, "The test should have at least two iterations");
352 
353 	pr_info("Test iterations: %"PRIu64"\n",	iterations);
354 
355 	for (i = 0; i < NUM_VM_MODES; ++i) {
356 		if (!guest_modes[i].enabled)
357 			continue;
358 		TEST_ASSERT(guest_modes[i].supported,
359 			    "Guest mode ID %d (%s) not supported.",
360 			    i, vm_guest_mode_string(i));
361 		run_test(i, iterations, phys_offset, wr_fract);
362 	}
363 
364 	return 0;
365 }
366