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(struct perf_test_vcpu_args *vcpu_args)
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 	int vcpu_id = vcpu_args->vcpu_id;
45 
46 	run = vcpu_state(vm, vcpu_id);
47 
48 	while (!READ_ONCE(host_quit)) {
49 		int current_iteration = READ_ONCE(iteration);
50 
51 		clock_gettime(CLOCK_MONOTONIC, &start);
52 		ret = _vcpu_run(vm, vcpu_id);
53 		ts_diff = timespec_elapsed(start);
54 
55 		TEST_ASSERT(ret == 0, "vcpu_run failed: %d\n", ret);
56 		TEST_ASSERT(get_ucall(vm, vcpu_id, NULL) == UCALL_SYNC,
57 			    "Invalid guest sync status: exit_reason=%s\n",
58 			    exit_reason_str(run->exit_reason));
59 
60 		pr_debug("Got sync event from vCPU %d\n", vcpu_id);
61 		vcpu_last_completed_iteration[vcpu_id] = current_iteration;
62 		pr_debug("vCPU %d updated last completed iteration to %d\n",
63 			 vcpu_id, vcpu_last_completed_iteration[vcpu_id]);
64 
65 		if (current_iteration) {
66 			pages_count += vcpu_args->pages;
67 			total = timespec_add(total, ts_diff);
68 			pr_debug("vCPU %d iteration %d dirty memory time: %ld.%.9lds\n",
69 				vcpu_id, current_iteration, ts_diff.tv_sec,
70 				ts_diff.tv_nsec);
71 		} else {
72 			pr_debug("vCPU %d iteration %d populate memory time: %ld.%.9lds\n",
73 				vcpu_id, current_iteration, ts_diff.tv_sec,
74 				ts_diff.tv_nsec);
75 		}
76 
77 		while (current_iteration == READ_ONCE(iteration) &&
78 		       !READ_ONCE(host_quit)) {}
79 	}
80 
81 	avg = timespec_div(total, vcpu_last_completed_iteration[vcpu_id]);
82 	pr_debug("\nvCPU %d dirtied 0x%lx pages over %d iterations in %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
83 		vcpu_id, pages_count, vcpu_last_completed_iteration[vcpu_id],
84 		total.tv_sec, total.tv_nsec, avg.tv_sec, avg.tv_nsec);
85 }
86 
87 struct test_params {
88 	unsigned long iterations;
89 	uint64_t phys_offset;
90 	int wr_fract;
91 	bool partition_vcpu_memory_access;
92 	enum vm_mem_backing_src_type backing_src;
93 	int slots;
94 };
95 
96 static void toggle_dirty_logging(struct kvm_vm *vm, int slots, bool enable)
97 {
98 	int i;
99 
100 	for (i = 0; i < slots; i++) {
101 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
102 		int flags = enable ? KVM_MEM_LOG_DIRTY_PAGES : 0;
103 
104 		vm_mem_region_set_flags(vm, slot, flags);
105 	}
106 }
107 
108 static inline void enable_dirty_logging(struct kvm_vm *vm, int slots)
109 {
110 	toggle_dirty_logging(vm, slots, true);
111 }
112 
113 static inline void disable_dirty_logging(struct kvm_vm *vm, int slots)
114 {
115 	toggle_dirty_logging(vm, slots, false);
116 }
117 
118 static void get_dirty_log(struct kvm_vm *vm, unsigned long *bitmaps[], int slots)
119 {
120 	int i;
121 
122 	for (i = 0; i < slots; i++) {
123 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
124 
125 		kvm_vm_get_dirty_log(vm, slot, bitmaps[i]);
126 	}
127 }
128 
129 static void clear_dirty_log(struct kvm_vm *vm, unsigned long *bitmaps[],
130 			    int slots, uint64_t pages_per_slot)
131 {
132 	int i;
133 
134 	for (i = 0; i < slots; i++) {
135 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
136 
137 		kvm_vm_clear_dirty_log(vm, slot, bitmaps[i], 0, pages_per_slot);
138 	}
139 }
140 
141 static unsigned long **alloc_bitmaps(int slots, uint64_t pages_per_slot)
142 {
143 	unsigned long **bitmaps;
144 	int i;
145 
146 	bitmaps = malloc(slots * sizeof(bitmaps[0]));
147 	TEST_ASSERT(bitmaps, "Failed to allocate bitmaps array.");
148 
149 	for (i = 0; i < slots; i++) {
150 		bitmaps[i] = bitmap_zalloc(pages_per_slot);
151 		TEST_ASSERT(bitmaps[i], "Failed to allocate slot bitmap.");
152 	}
153 
154 	return bitmaps;
155 }
156 
157 static void free_bitmaps(unsigned long *bitmaps[], int slots)
158 {
159 	int i;
160 
161 	for (i = 0; i < slots; i++)
162 		free(bitmaps[i]);
163 
164 	free(bitmaps);
165 }
166 
167 static void run_test(enum vm_guest_mode mode, void *arg)
168 {
169 	struct test_params *p = arg;
170 	struct kvm_vm *vm;
171 	unsigned long **bitmaps;
172 	uint64_t guest_num_pages;
173 	uint64_t host_num_pages;
174 	uint64_t pages_per_slot;
175 	int vcpu_id;
176 	struct timespec start;
177 	struct timespec ts_diff;
178 	struct timespec get_dirty_log_total = (struct timespec){0};
179 	struct timespec vcpu_dirty_total = (struct timespec){0};
180 	struct timespec avg;
181 	struct kvm_enable_cap cap = {};
182 	struct timespec clear_dirty_log_total = (struct timespec){0};
183 
184 	vm = perf_test_create_vm(mode, nr_vcpus, guest_percpu_mem_size,
185 				 p->slots, p->backing_src,
186 				 p->partition_vcpu_memory_access);
187 
188 	perf_test_set_wr_fract(vm, p->wr_fract);
189 
190 	guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm_get_page_shift(vm);
191 	guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
192 	host_num_pages = vm_num_host_pages(mode, guest_num_pages);
193 	pages_per_slot = host_num_pages / p->slots;
194 
195 	bitmaps = alloc_bitmaps(p->slots, pages_per_slot);
196 
197 	if (dirty_log_manual_caps) {
198 		cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
199 		cap.args[0] = dirty_log_manual_caps;
200 		vm_enable_cap(vm, &cap);
201 	}
202 
203 	/* Start the iterations */
204 	iteration = 0;
205 	host_quit = false;
206 
207 	clock_gettime(CLOCK_MONOTONIC, &start);
208 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++)
209 		vcpu_last_completed_iteration[vcpu_id] = -1;
210 
211 	perf_test_start_vcpu_threads(nr_vcpus, vcpu_worker);
212 
213 	/* Allow the vCPUs to populate memory */
214 	pr_debug("Starting iteration %d - Populating\n", iteration);
215 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
216 		while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) !=
217 		       iteration)
218 			;
219 	}
220 
221 	ts_diff = timespec_elapsed(start);
222 	pr_info("Populate memory time: %ld.%.9lds\n",
223 		ts_diff.tv_sec, ts_diff.tv_nsec);
224 
225 	/* Enable dirty logging */
226 	clock_gettime(CLOCK_MONOTONIC, &start);
227 	enable_dirty_logging(vm, p->slots);
228 	ts_diff = timespec_elapsed(start);
229 	pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
230 		ts_diff.tv_sec, ts_diff.tv_nsec);
231 
232 	while (iteration < p->iterations) {
233 		/*
234 		 * Incrementing the iteration number will start the vCPUs
235 		 * dirtying memory again.
236 		 */
237 		clock_gettime(CLOCK_MONOTONIC, &start);
238 		iteration++;
239 
240 		pr_debug("Starting iteration %d\n", iteration);
241 		for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
242 			while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id])
243 			       != iteration)
244 				;
245 		}
246 
247 		ts_diff = timespec_elapsed(start);
248 		vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
249 		pr_info("Iteration %d dirty memory time: %ld.%.9lds\n",
250 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
251 
252 		clock_gettime(CLOCK_MONOTONIC, &start);
253 		get_dirty_log(vm, bitmaps, p->slots);
254 		ts_diff = timespec_elapsed(start);
255 		get_dirty_log_total = timespec_add(get_dirty_log_total,
256 						   ts_diff);
257 		pr_info("Iteration %d get dirty log time: %ld.%.9lds\n",
258 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
259 
260 		if (dirty_log_manual_caps) {
261 			clock_gettime(CLOCK_MONOTONIC, &start);
262 			clear_dirty_log(vm, bitmaps, p->slots, pages_per_slot);
263 			ts_diff = timespec_elapsed(start);
264 			clear_dirty_log_total = timespec_add(clear_dirty_log_total,
265 							     ts_diff);
266 			pr_info("Iteration %d clear dirty log time: %ld.%.9lds\n",
267 				iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
268 		}
269 	}
270 
271 	/* Disable dirty logging */
272 	clock_gettime(CLOCK_MONOTONIC, &start);
273 	disable_dirty_logging(vm, p->slots);
274 	ts_diff = timespec_elapsed(start);
275 	pr_info("Disabling dirty logging time: %ld.%.9lds\n",
276 		ts_diff.tv_sec, ts_diff.tv_nsec);
277 
278 	/* Tell the vcpu thread to quit */
279 	host_quit = true;
280 	perf_test_join_vcpu_threads(nr_vcpus);
281 
282 	avg = timespec_div(get_dirty_log_total, p->iterations);
283 	pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
284 		p->iterations, get_dirty_log_total.tv_sec,
285 		get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
286 
287 	if (dirty_log_manual_caps) {
288 		avg = timespec_div(clear_dirty_log_total, p->iterations);
289 		pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
290 			p->iterations, clear_dirty_log_total.tv_sec,
291 			clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
292 	}
293 
294 	free_bitmaps(bitmaps, p->slots);
295 	perf_test_destroy_vm(vm);
296 }
297 
298 static void help(char *name)
299 {
300 	puts("");
301 	printf("usage: %s [-h] [-i iterations] [-p offset] "
302 	       "[-m mode] [-b vcpu bytes] [-v vcpus] [-o] [-s mem type]"
303 	       "[-x memslots]\n", name);
304 	puts("");
305 	printf(" -i: specify iteration counts (default: %"PRIu64")\n",
306 	       TEST_HOST_LOOP_N);
307 	printf(" -p: specify guest physical test memory offset\n"
308 	       "     Warning: a low offset can conflict with the loaded test code.\n");
309 	guest_modes_help();
310 	printf(" -b: specify the size of the memory region which should be\n"
311 	       "     dirtied by each vCPU. e.g. 10M or 3G.\n"
312 	       "     (default: 1G)\n");
313 	printf(" -f: specify the fraction of pages which should be written to\n"
314 	       "     as opposed to simply read, in the form\n"
315 	       "     1/<fraction of pages to write>.\n"
316 	       "     (default: 1 i.e. all pages are written to.)\n");
317 	printf(" -v: specify the number of vCPUs to run.\n");
318 	printf(" -o: Overlap guest memory accesses instead of partitioning\n"
319 	       "     them into a separate region of memory for each vCPU.\n");
320 	backing_src_help("-s");
321 	printf(" -x: Split the memory region into this number of memslots.\n"
322 	       "     (default: 1)\n");
323 	puts("");
324 	exit(0);
325 }
326 
327 int main(int argc, char *argv[])
328 {
329 	int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS);
330 	struct test_params p = {
331 		.iterations = TEST_HOST_LOOP_N,
332 		.wr_fract = 1,
333 		.partition_vcpu_memory_access = true,
334 		.backing_src = DEFAULT_VM_MEM_SRC,
335 		.slots = 1,
336 	};
337 	int opt;
338 
339 	dirty_log_manual_caps =
340 		kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
341 	dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
342 				  KVM_DIRTY_LOG_INITIALLY_SET);
343 
344 	guest_modes_append_default();
345 
346 	while ((opt = getopt(argc, argv, "hi:p:m:b:f:v:os:x:")) != -1) {
347 		switch (opt) {
348 		case 'i':
349 			p.iterations = atoi(optarg);
350 			break;
351 		case 'p':
352 			p.phys_offset = strtoull(optarg, NULL, 0);
353 			break;
354 		case 'm':
355 			guest_modes_cmdline(optarg);
356 			break;
357 		case 'b':
358 			guest_percpu_mem_size = parse_size(optarg);
359 			break;
360 		case 'f':
361 			p.wr_fract = atoi(optarg);
362 			TEST_ASSERT(p.wr_fract >= 1,
363 				    "Write fraction cannot be less than one");
364 			break;
365 		case 'v':
366 			nr_vcpus = atoi(optarg);
367 			TEST_ASSERT(nr_vcpus > 0 && nr_vcpus <= max_vcpus,
368 				    "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
369 			break;
370 		case 'o':
371 			p.partition_vcpu_memory_access = false;
372 			break;
373 		case 's':
374 			p.backing_src = parse_backing_src_type(optarg);
375 			break;
376 		case 'x':
377 			p.slots = atoi(optarg);
378 			break;
379 		case 'h':
380 		default:
381 			help(argv[0]);
382 			break;
383 		}
384 	}
385 
386 	TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations");
387 
388 	pr_info("Test iterations: %"PRIu64"\n",	p.iterations);
389 
390 	for_each_guest_mode(run_test, &p);
391 
392 	return 0;
393 }
394