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 #ifdef __aarch64__
22 #include "aarch64/vgic.h"
23 
24 #define GICD_BASE_GPA			0x8000000ULL
25 #define GICR_BASE_GPA			0x80A0000ULL
26 #endif
27 
28 /* How many host loops to run by default (one KVM_GET_DIRTY_LOG for each loop)*/
29 #define TEST_HOST_LOOP_N		2UL
30 
31 static int nr_vcpus = 1;
32 static uint64_t guest_percpu_mem_size = DEFAULT_PER_VCPU_MEM_SIZE;
33 
34 /* Host variables */
35 static u64 dirty_log_manual_caps;
36 static bool host_quit;
37 static int iteration;
38 static int vcpu_last_completed_iteration[KVM_MAX_VCPUS];
39 
40 static void vcpu_worker(struct perf_test_vcpu_args *vcpu_args)
41 {
42 	int ret;
43 	struct kvm_vm *vm = perf_test_args.vm;
44 	uint64_t pages_count = 0;
45 	struct kvm_run *run;
46 	struct timespec start;
47 	struct timespec ts_diff;
48 	struct timespec total = (struct timespec){0};
49 	struct timespec avg;
50 	int vcpu_id = vcpu_args->vcpu_id;
51 
52 	run = vcpu_state(vm, vcpu_id);
53 
54 	while (!READ_ONCE(host_quit)) {
55 		int current_iteration = READ_ONCE(iteration);
56 
57 		clock_gettime(CLOCK_MONOTONIC, &start);
58 		ret = _vcpu_run(vm, vcpu_id);
59 		ts_diff = timespec_elapsed(start);
60 
61 		TEST_ASSERT(ret == 0, "vcpu_run failed: %d\n", ret);
62 		TEST_ASSERT(get_ucall(vm, vcpu_id, NULL) == UCALL_SYNC,
63 			    "Invalid guest sync status: exit_reason=%s\n",
64 			    exit_reason_str(run->exit_reason));
65 
66 		pr_debug("Got sync event from vCPU %d\n", vcpu_id);
67 		vcpu_last_completed_iteration[vcpu_id] = current_iteration;
68 		pr_debug("vCPU %d updated last completed iteration to %d\n",
69 			 vcpu_id, vcpu_last_completed_iteration[vcpu_id]);
70 
71 		if (current_iteration) {
72 			pages_count += vcpu_args->pages;
73 			total = timespec_add(total, ts_diff);
74 			pr_debug("vCPU %d iteration %d dirty memory time: %ld.%.9lds\n",
75 				vcpu_id, current_iteration, ts_diff.tv_sec,
76 				ts_diff.tv_nsec);
77 		} else {
78 			pr_debug("vCPU %d iteration %d populate memory time: %ld.%.9lds\n",
79 				vcpu_id, current_iteration, ts_diff.tv_sec,
80 				ts_diff.tv_nsec);
81 		}
82 
83 		while (current_iteration == READ_ONCE(iteration) &&
84 		       !READ_ONCE(host_quit)) {}
85 	}
86 
87 	avg = timespec_div(total, vcpu_last_completed_iteration[vcpu_id]);
88 	pr_debug("\nvCPU %d dirtied 0x%lx pages over %d iterations in %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
89 		vcpu_id, pages_count, vcpu_last_completed_iteration[vcpu_id],
90 		total.tv_sec, total.tv_nsec, avg.tv_sec, avg.tv_nsec);
91 }
92 
93 struct test_params {
94 	unsigned long iterations;
95 	uint64_t phys_offset;
96 	int wr_fract;
97 	bool partition_vcpu_memory_access;
98 	enum vm_mem_backing_src_type backing_src;
99 	int slots;
100 };
101 
102 static void toggle_dirty_logging(struct kvm_vm *vm, int slots, bool enable)
103 {
104 	int i;
105 
106 	for (i = 0; i < slots; i++) {
107 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
108 		int flags = enable ? KVM_MEM_LOG_DIRTY_PAGES : 0;
109 
110 		vm_mem_region_set_flags(vm, slot, flags);
111 	}
112 }
113 
114 static inline void enable_dirty_logging(struct kvm_vm *vm, int slots)
115 {
116 	toggle_dirty_logging(vm, slots, true);
117 }
118 
119 static inline void disable_dirty_logging(struct kvm_vm *vm, int slots)
120 {
121 	toggle_dirty_logging(vm, slots, false);
122 }
123 
124 static void get_dirty_log(struct kvm_vm *vm, unsigned long *bitmaps[], int slots)
125 {
126 	int i;
127 
128 	for (i = 0; i < slots; i++) {
129 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
130 
131 		kvm_vm_get_dirty_log(vm, slot, bitmaps[i]);
132 	}
133 }
134 
135 static void clear_dirty_log(struct kvm_vm *vm, unsigned long *bitmaps[],
136 			    int slots, uint64_t pages_per_slot)
137 {
138 	int i;
139 
140 	for (i = 0; i < slots; i++) {
141 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
142 
143 		kvm_vm_clear_dirty_log(vm, slot, bitmaps[i], 0, pages_per_slot);
144 	}
145 }
146 
147 static unsigned long **alloc_bitmaps(int slots, uint64_t pages_per_slot)
148 {
149 	unsigned long **bitmaps;
150 	int i;
151 
152 	bitmaps = malloc(slots * sizeof(bitmaps[0]));
153 	TEST_ASSERT(bitmaps, "Failed to allocate bitmaps array.");
154 
155 	for (i = 0; i < slots; i++) {
156 		bitmaps[i] = bitmap_zalloc(pages_per_slot);
157 		TEST_ASSERT(bitmaps[i], "Failed to allocate slot bitmap.");
158 	}
159 
160 	return bitmaps;
161 }
162 
163 static void free_bitmaps(unsigned long *bitmaps[], int slots)
164 {
165 	int i;
166 
167 	for (i = 0; i < slots; i++)
168 		free(bitmaps[i]);
169 
170 	free(bitmaps);
171 }
172 
173 static void run_test(enum vm_guest_mode mode, void *arg)
174 {
175 	struct test_params *p = arg;
176 	struct kvm_vm *vm;
177 	unsigned long **bitmaps;
178 	uint64_t guest_num_pages;
179 	uint64_t host_num_pages;
180 	uint64_t pages_per_slot;
181 	int vcpu_id;
182 	struct timespec start;
183 	struct timespec ts_diff;
184 	struct timespec get_dirty_log_total = (struct timespec){0};
185 	struct timespec vcpu_dirty_total = (struct timespec){0};
186 	struct timespec avg;
187 	struct kvm_enable_cap cap = {};
188 	struct timespec clear_dirty_log_total = (struct timespec){0};
189 
190 	vm = perf_test_create_vm(mode, nr_vcpus, guest_percpu_mem_size,
191 				 p->slots, p->backing_src,
192 				 p->partition_vcpu_memory_access);
193 
194 	perf_test_set_wr_fract(vm, p->wr_fract);
195 
196 	guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm_get_page_shift(vm);
197 	guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
198 	host_num_pages = vm_num_host_pages(mode, guest_num_pages);
199 	pages_per_slot = host_num_pages / p->slots;
200 
201 	bitmaps = alloc_bitmaps(p->slots, pages_per_slot);
202 
203 	if (dirty_log_manual_caps) {
204 		cap.cap = KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2;
205 		cap.args[0] = dirty_log_manual_caps;
206 		vm_enable_cap(vm, &cap);
207 	}
208 
209 #ifdef __aarch64__
210 	vgic_v3_setup(vm, nr_vcpus, 64, GICD_BASE_GPA, GICR_BASE_GPA);
211 #endif
212 
213 	/* Start the iterations */
214 	iteration = 0;
215 	host_quit = false;
216 
217 	clock_gettime(CLOCK_MONOTONIC, &start);
218 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++)
219 		vcpu_last_completed_iteration[vcpu_id] = -1;
220 
221 	perf_test_start_vcpu_threads(nr_vcpus, vcpu_worker);
222 
223 	/* Allow the vCPUs to populate memory */
224 	pr_debug("Starting iteration %d - Populating\n", iteration);
225 	for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
226 		while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) !=
227 		       iteration)
228 			;
229 	}
230 
231 	ts_diff = timespec_elapsed(start);
232 	pr_info("Populate memory time: %ld.%.9lds\n",
233 		ts_diff.tv_sec, ts_diff.tv_nsec);
234 
235 	/* Enable dirty logging */
236 	clock_gettime(CLOCK_MONOTONIC, &start);
237 	enable_dirty_logging(vm, p->slots);
238 	ts_diff = timespec_elapsed(start);
239 	pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
240 		ts_diff.tv_sec, ts_diff.tv_nsec);
241 
242 	while (iteration < p->iterations) {
243 		/*
244 		 * Incrementing the iteration number will start the vCPUs
245 		 * dirtying memory again.
246 		 */
247 		clock_gettime(CLOCK_MONOTONIC, &start);
248 		iteration++;
249 
250 		pr_debug("Starting iteration %d\n", iteration);
251 		for (vcpu_id = 0; vcpu_id < nr_vcpus; vcpu_id++) {
252 			while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id])
253 			       != iteration)
254 				;
255 		}
256 
257 		ts_diff = timespec_elapsed(start);
258 		vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
259 		pr_info("Iteration %d dirty memory time: %ld.%.9lds\n",
260 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
261 
262 		clock_gettime(CLOCK_MONOTONIC, &start);
263 		get_dirty_log(vm, bitmaps, p->slots);
264 		ts_diff = timespec_elapsed(start);
265 		get_dirty_log_total = timespec_add(get_dirty_log_total,
266 						   ts_diff);
267 		pr_info("Iteration %d get dirty log time: %ld.%.9lds\n",
268 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
269 
270 		if (dirty_log_manual_caps) {
271 			clock_gettime(CLOCK_MONOTONIC, &start);
272 			clear_dirty_log(vm, bitmaps, p->slots, pages_per_slot);
273 			ts_diff = timespec_elapsed(start);
274 			clear_dirty_log_total = timespec_add(clear_dirty_log_total,
275 							     ts_diff);
276 			pr_info("Iteration %d clear dirty log time: %ld.%.9lds\n",
277 				iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
278 		}
279 	}
280 
281 	/* Disable dirty logging */
282 	clock_gettime(CLOCK_MONOTONIC, &start);
283 	disable_dirty_logging(vm, p->slots);
284 	ts_diff = timespec_elapsed(start);
285 	pr_info("Disabling dirty logging time: %ld.%.9lds\n",
286 		ts_diff.tv_sec, ts_diff.tv_nsec);
287 
288 	/* Tell the vcpu thread to quit */
289 	host_quit = true;
290 	perf_test_join_vcpu_threads(nr_vcpus);
291 
292 	avg = timespec_div(get_dirty_log_total, p->iterations);
293 	pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
294 		p->iterations, get_dirty_log_total.tv_sec,
295 		get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
296 
297 	if (dirty_log_manual_caps) {
298 		avg = timespec_div(clear_dirty_log_total, p->iterations);
299 		pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
300 			p->iterations, clear_dirty_log_total.tv_sec,
301 			clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
302 	}
303 
304 	free_bitmaps(bitmaps, p->slots);
305 	perf_test_destroy_vm(vm);
306 }
307 
308 static void help(char *name)
309 {
310 	puts("");
311 	printf("usage: %s [-h] [-i iterations] [-p offset] [-g]"
312 	       "[-m mode] [-b vcpu bytes] [-v vcpus] [-o] [-s mem type]"
313 	       "[-x memslots]\n", name);
314 	puts("");
315 	printf(" -i: specify iteration counts (default: %"PRIu64")\n",
316 	       TEST_HOST_LOOP_N);
317 	printf(" -g: Do not enable KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2. This\n"
318 	       "     makes KVM_GET_DIRTY_LOG clear the dirty log (i.e.\n"
319 	       "     KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE is not enabled)\n"
320 	       "     and writes will be tracked as soon as dirty logging is\n"
321 	       "     enabled on the memslot (i.e. KVM_DIRTY_LOG_INITIALLY_SET\n"
322 	       "     is not enabled).\n");
323 	printf(" -p: specify guest physical test memory offset\n"
324 	       "     Warning: a low offset can conflict with the loaded test code.\n");
325 	guest_modes_help();
326 	printf(" -b: specify the size of the memory region which should be\n"
327 	       "     dirtied by each vCPU. e.g. 10M or 3G.\n"
328 	       "     (default: 1G)\n");
329 	printf(" -f: specify the fraction of pages which should be written to\n"
330 	       "     as opposed to simply read, in the form\n"
331 	       "     1/<fraction of pages to write>.\n"
332 	       "     (default: 1 i.e. all pages are written to.)\n");
333 	printf(" -v: specify the number of vCPUs to run.\n");
334 	printf(" -o: Overlap guest memory accesses instead of partitioning\n"
335 	       "     them into a separate region of memory for each vCPU.\n");
336 	backing_src_help("-s");
337 	printf(" -x: Split the memory region into this number of memslots.\n"
338 	       "     (default: 1)\n");
339 	puts("");
340 	exit(0);
341 }
342 
343 int main(int argc, char *argv[])
344 {
345 	int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS);
346 	struct test_params p = {
347 		.iterations = TEST_HOST_LOOP_N,
348 		.wr_fract = 1,
349 		.partition_vcpu_memory_access = true,
350 		.backing_src = DEFAULT_VM_MEM_SRC,
351 		.slots = 1,
352 	};
353 	int opt;
354 
355 	dirty_log_manual_caps =
356 		kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
357 	dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
358 				  KVM_DIRTY_LOG_INITIALLY_SET);
359 
360 	guest_modes_append_default();
361 
362 	while ((opt = getopt(argc, argv, "ghi:p:m:b:f:v:os:x:")) != -1) {
363 		switch (opt) {
364 		case 'g':
365 			dirty_log_manual_caps = 0;
366 			break;
367 		case 'i':
368 			p.iterations = atoi(optarg);
369 			break;
370 		case 'p':
371 			p.phys_offset = strtoull(optarg, NULL, 0);
372 			break;
373 		case 'm':
374 			guest_modes_cmdline(optarg);
375 			break;
376 		case 'b':
377 			guest_percpu_mem_size = parse_size(optarg);
378 			break;
379 		case 'f':
380 			p.wr_fract = atoi(optarg);
381 			TEST_ASSERT(p.wr_fract >= 1,
382 				    "Write fraction cannot be less than one");
383 			break;
384 		case 'v':
385 			nr_vcpus = atoi(optarg);
386 			TEST_ASSERT(nr_vcpus > 0 && nr_vcpus <= max_vcpus,
387 				    "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
388 			break;
389 		case 'o':
390 			p.partition_vcpu_memory_access = false;
391 			break;
392 		case 's':
393 			p.backing_src = parse_backing_src_type(optarg);
394 			break;
395 		case 'x':
396 			p.slots = atoi(optarg);
397 			break;
398 		case 'h':
399 		default:
400 			help(argv[0]);
401 			break;
402 		}
403 	}
404 
405 	TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations");
406 
407 	pr_info("Test iterations: %"PRIu64"\n",	p.iterations);
408 
409 	for_each_guest_mode(run_test, &p);
410 
411 	return 0;
412 }
413