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 #ifdef __aarch64__
23 #include "aarch64/vgic.h"
24 
25 #define GICD_BASE_GPA			0x8000000ULL
26 #define GICR_BASE_GPA			0x80A0000ULL
27 
28 static int gic_fd;
29 
30 static void arch_setup_vm(struct kvm_vm *vm, unsigned int nr_vcpus)
31 {
32 	/*
33 	 * The test can still run even if hardware does not support GICv3, as it
34 	 * is only an optimization to reduce guest exits.
35 	 */
36 	gic_fd = vgic_v3_setup(vm, nr_vcpus, 64, GICD_BASE_GPA, GICR_BASE_GPA);
37 }
38 
39 static void arch_cleanup_vm(struct kvm_vm *vm)
40 {
41 	if (gic_fd > 0)
42 		close(gic_fd);
43 }
44 
45 #else /* __aarch64__ */
46 
47 static void arch_setup_vm(struct kvm_vm *vm, unsigned int nr_vcpus)
48 {
49 }
50 
51 static void arch_cleanup_vm(struct kvm_vm *vm)
52 {
53 }
54 
55 #endif
56 
57 /* How many host loops to run by default (one KVM_GET_DIRTY_LOG for each loop)*/
58 #define TEST_HOST_LOOP_N		2UL
59 
60 static int nr_vcpus = 1;
61 static uint64_t guest_percpu_mem_size = DEFAULT_PER_VCPU_MEM_SIZE;
62 static bool run_vcpus_while_disabling_dirty_logging;
63 
64 /* Host variables */
65 static u64 dirty_log_manual_caps;
66 static bool host_quit;
67 static int iteration;
68 static int vcpu_last_completed_iteration[KVM_MAX_VCPUS];
69 
70 static void vcpu_worker(struct perf_test_vcpu_args *vcpu_args)
71 {
72 	struct kvm_vcpu *vcpu = vcpu_args->vcpu;
73 	int vcpu_idx = vcpu_args->vcpu_idx;
74 	uint64_t pages_count = 0;
75 	struct kvm_run *run;
76 	struct timespec start;
77 	struct timespec ts_diff;
78 	struct timespec total = (struct timespec){0};
79 	struct timespec avg;
80 	int ret;
81 
82 	run = vcpu->run;
83 
84 	while (!READ_ONCE(host_quit)) {
85 		int current_iteration = READ_ONCE(iteration);
86 
87 		clock_gettime(CLOCK_MONOTONIC, &start);
88 		ret = _vcpu_run(vcpu);
89 		ts_diff = timespec_elapsed(start);
90 
91 		TEST_ASSERT(ret == 0, "vcpu_run failed: %d\n", ret);
92 		TEST_ASSERT(get_ucall(vcpu, NULL) == UCALL_SYNC,
93 			    "Invalid guest sync status: exit_reason=%s\n",
94 			    exit_reason_str(run->exit_reason));
95 
96 		pr_debug("Got sync event from vCPU %d\n", vcpu_idx);
97 		vcpu_last_completed_iteration[vcpu_idx] = current_iteration;
98 		pr_debug("vCPU %d updated last completed iteration to %d\n",
99 			 vcpu_idx, vcpu_last_completed_iteration[vcpu_idx]);
100 
101 		if (current_iteration) {
102 			pages_count += vcpu_args->pages;
103 			total = timespec_add(total, ts_diff);
104 			pr_debug("vCPU %d iteration %d dirty memory time: %ld.%.9lds\n",
105 				vcpu_idx, current_iteration, ts_diff.tv_sec,
106 				ts_diff.tv_nsec);
107 		} else {
108 			pr_debug("vCPU %d iteration %d populate memory time: %ld.%.9lds\n",
109 				vcpu_idx, current_iteration, ts_diff.tv_sec,
110 				ts_diff.tv_nsec);
111 		}
112 
113 		/*
114 		 * Keep running the guest while dirty logging is being disabled
115 		 * (iteration is negative) so that vCPUs are accessing memory
116 		 * for the entire duration of zapping collapsible SPTEs.
117 		 */
118 		while (current_iteration == READ_ONCE(iteration) &&
119 		       READ_ONCE(iteration) >= 0 && !READ_ONCE(host_quit)) {}
120 	}
121 
122 	avg = timespec_div(total, vcpu_last_completed_iteration[vcpu_idx]);
123 	pr_debug("\nvCPU %d dirtied 0x%lx pages over %d iterations in %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
124 		vcpu_idx, pages_count, vcpu_last_completed_iteration[vcpu_idx],
125 		total.tv_sec, total.tv_nsec, avg.tv_sec, avg.tv_nsec);
126 }
127 
128 struct test_params {
129 	unsigned long iterations;
130 	uint64_t phys_offset;
131 	int wr_fract;
132 	bool partition_vcpu_memory_access;
133 	enum vm_mem_backing_src_type backing_src;
134 	int slots;
135 };
136 
137 static void toggle_dirty_logging(struct kvm_vm *vm, int slots, bool enable)
138 {
139 	int i;
140 
141 	for (i = 0; i < slots; i++) {
142 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
143 		int flags = enable ? KVM_MEM_LOG_DIRTY_PAGES : 0;
144 
145 		vm_mem_region_set_flags(vm, slot, flags);
146 	}
147 }
148 
149 static inline void enable_dirty_logging(struct kvm_vm *vm, int slots)
150 {
151 	toggle_dirty_logging(vm, slots, true);
152 }
153 
154 static inline void disable_dirty_logging(struct kvm_vm *vm, int slots)
155 {
156 	toggle_dirty_logging(vm, slots, false);
157 }
158 
159 static void get_dirty_log(struct kvm_vm *vm, unsigned long *bitmaps[], int slots)
160 {
161 	int i;
162 
163 	for (i = 0; i < slots; i++) {
164 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
165 
166 		kvm_vm_get_dirty_log(vm, slot, bitmaps[i]);
167 	}
168 }
169 
170 static void clear_dirty_log(struct kvm_vm *vm, unsigned long *bitmaps[],
171 			    int slots, uint64_t pages_per_slot)
172 {
173 	int i;
174 
175 	for (i = 0; i < slots; i++) {
176 		int slot = PERF_TEST_MEM_SLOT_INDEX + i;
177 
178 		kvm_vm_clear_dirty_log(vm, slot, bitmaps[i], 0, pages_per_slot);
179 	}
180 }
181 
182 static unsigned long **alloc_bitmaps(int slots, uint64_t pages_per_slot)
183 {
184 	unsigned long **bitmaps;
185 	int i;
186 
187 	bitmaps = malloc(slots * sizeof(bitmaps[0]));
188 	TEST_ASSERT(bitmaps, "Failed to allocate bitmaps array.");
189 
190 	for (i = 0; i < slots; i++) {
191 		bitmaps[i] = bitmap_zalloc(pages_per_slot);
192 		TEST_ASSERT(bitmaps[i], "Failed to allocate slot bitmap.");
193 	}
194 
195 	return bitmaps;
196 }
197 
198 static void free_bitmaps(unsigned long *bitmaps[], int slots)
199 {
200 	int i;
201 
202 	for (i = 0; i < slots; i++)
203 		free(bitmaps[i]);
204 
205 	free(bitmaps);
206 }
207 
208 static void run_test(enum vm_guest_mode mode, void *arg)
209 {
210 	struct test_params *p = arg;
211 	struct kvm_vm *vm;
212 	unsigned long **bitmaps;
213 	uint64_t guest_num_pages;
214 	uint64_t host_num_pages;
215 	uint64_t pages_per_slot;
216 	struct timespec start;
217 	struct timespec ts_diff;
218 	struct timespec get_dirty_log_total = (struct timespec){0};
219 	struct timespec vcpu_dirty_total = (struct timespec){0};
220 	struct timespec avg;
221 	struct timespec clear_dirty_log_total = (struct timespec){0};
222 	int i;
223 
224 	vm = perf_test_create_vm(mode, nr_vcpus, guest_percpu_mem_size,
225 				 p->slots, p->backing_src,
226 				 p->partition_vcpu_memory_access);
227 
228 	perf_test_set_wr_fract(vm, p->wr_fract);
229 
230 	guest_num_pages = (nr_vcpus * guest_percpu_mem_size) >> vm->page_shift;
231 	guest_num_pages = vm_adjust_num_guest_pages(mode, guest_num_pages);
232 	host_num_pages = vm_num_host_pages(mode, guest_num_pages);
233 	pages_per_slot = host_num_pages / p->slots;
234 
235 	bitmaps = alloc_bitmaps(p->slots, pages_per_slot);
236 
237 	if (dirty_log_manual_caps)
238 		vm_enable_cap(vm, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2,
239 			      dirty_log_manual_caps);
240 
241 	arch_setup_vm(vm, nr_vcpus);
242 
243 	/* Start the iterations */
244 	iteration = 0;
245 	host_quit = false;
246 
247 	clock_gettime(CLOCK_MONOTONIC, &start);
248 	for (i = 0; i < nr_vcpus; i++)
249 		vcpu_last_completed_iteration[i] = -1;
250 
251 	perf_test_start_vcpu_threads(nr_vcpus, vcpu_worker);
252 
253 	/* Allow the vCPUs to populate memory */
254 	pr_debug("Starting iteration %d - Populating\n", iteration);
255 	for (i = 0; i < nr_vcpus; i++) {
256 		while (READ_ONCE(vcpu_last_completed_iteration[i]) !=
257 		       iteration)
258 			;
259 	}
260 
261 	ts_diff = timespec_elapsed(start);
262 	pr_info("Populate memory time: %ld.%.9lds\n",
263 		ts_diff.tv_sec, ts_diff.tv_nsec);
264 
265 	/* Enable dirty logging */
266 	clock_gettime(CLOCK_MONOTONIC, &start);
267 	enable_dirty_logging(vm, p->slots);
268 	ts_diff = timespec_elapsed(start);
269 	pr_info("Enabling dirty logging time: %ld.%.9lds\n\n",
270 		ts_diff.tv_sec, ts_diff.tv_nsec);
271 
272 	while (iteration < p->iterations) {
273 		/*
274 		 * Incrementing the iteration number will start the vCPUs
275 		 * dirtying memory again.
276 		 */
277 		clock_gettime(CLOCK_MONOTONIC, &start);
278 		iteration++;
279 
280 		pr_debug("Starting iteration %d\n", iteration);
281 		for (i = 0; i < nr_vcpus; i++) {
282 			while (READ_ONCE(vcpu_last_completed_iteration[i])
283 			       != iteration)
284 				;
285 		}
286 
287 		ts_diff = timespec_elapsed(start);
288 		vcpu_dirty_total = timespec_add(vcpu_dirty_total, ts_diff);
289 		pr_info("Iteration %d dirty memory time: %ld.%.9lds\n",
290 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
291 
292 		clock_gettime(CLOCK_MONOTONIC, &start);
293 		get_dirty_log(vm, bitmaps, p->slots);
294 		ts_diff = timespec_elapsed(start);
295 		get_dirty_log_total = timespec_add(get_dirty_log_total,
296 						   ts_diff);
297 		pr_info("Iteration %d get dirty log time: %ld.%.9lds\n",
298 			iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
299 
300 		if (dirty_log_manual_caps) {
301 			clock_gettime(CLOCK_MONOTONIC, &start);
302 			clear_dirty_log(vm, bitmaps, p->slots, pages_per_slot);
303 			ts_diff = timespec_elapsed(start);
304 			clear_dirty_log_total = timespec_add(clear_dirty_log_total,
305 							     ts_diff);
306 			pr_info("Iteration %d clear dirty log time: %ld.%.9lds\n",
307 				iteration, ts_diff.tv_sec, ts_diff.tv_nsec);
308 		}
309 	}
310 
311 	/*
312 	 * Run vCPUs while dirty logging is being disabled to stress disabling
313 	 * in terms of both performance and correctness.  Opt-in via command
314 	 * line as this significantly increases time to disable dirty logging.
315 	 */
316 	if (run_vcpus_while_disabling_dirty_logging)
317 		WRITE_ONCE(iteration, -1);
318 
319 	/* Disable dirty logging */
320 	clock_gettime(CLOCK_MONOTONIC, &start);
321 	disable_dirty_logging(vm, p->slots);
322 	ts_diff = timespec_elapsed(start);
323 	pr_info("Disabling dirty logging time: %ld.%.9lds\n",
324 		ts_diff.tv_sec, ts_diff.tv_nsec);
325 
326 	/*
327 	 * Tell the vCPU threads to quit.  No need to manually check that vCPUs
328 	 * have stopped running after disabling dirty logging, the join will
329 	 * wait for them to exit.
330 	 */
331 	host_quit = true;
332 	perf_test_join_vcpu_threads(nr_vcpus);
333 
334 	avg = timespec_div(get_dirty_log_total, p->iterations);
335 	pr_info("Get dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
336 		p->iterations, get_dirty_log_total.tv_sec,
337 		get_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
338 
339 	if (dirty_log_manual_caps) {
340 		avg = timespec_div(clear_dirty_log_total, p->iterations);
341 		pr_info("Clear dirty log over %lu iterations took %ld.%.9lds. (Avg %ld.%.9lds/iteration)\n",
342 			p->iterations, clear_dirty_log_total.tv_sec,
343 			clear_dirty_log_total.tv_nsec, avg.tv_sec, avg.tv_nsec);
344 	}
345 
346 	free_bitmaps(bitmaps, p->slots);
347 	arch_cleanup_vm(vm);
348 	perf_test_destroy_vm(vm);
349 }
350 
351 static void help(char *name)
352 {
353 	puts("");
354 	printf("usage: %s [-h] [-i iterations] [-p offset] [-g] "
355 	       "[-m mode] [-n] [-b vcpu bytes] [-v vcpus] [-o] [-s mem type]"
356 	       "[-x memslots]\n", name);
357 	puts("");
358 	printf(" -i: specify iteration counts (default: %"PRIu64")\n",
359 	       TEST_HOST_LOOP_N);
360 	printf(" -g: Do not enable KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2. This\n"
361 	       "     makes KVM_GET_DIRTY_LOG clear the dirty log (i.e.\n"
362 	       "     KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE is not enabled)\n"
363 	       "     and writes will be tracked as soon as dirty logging is\n"
364 	       "     enabled on the memslot (i.e. KVM_DIRTY_LOG_INITIALLY_SET\n"
365 	       "     is not enabled).\n");
366 	printf(" -p: specify guest physical test memory offset\n"
367 	       "     Warning: a low offset can conflict with the loaded test code.\n");
368 	guest_modes_help();
369 	printf(" -n: Run the vCPUs in nested mode (L2)\n");
370 	printf(" -e: Run vCPUs while dirty logging is being disabled.  This\n"
371 	       "     can significantly increase runtime, especially if there\n"
372 	       "     isn't a dedicated pCPU for the main thread.\n");
373 	printf(" -b: specify the size of the memory region which should be\n"
374 	       "     dirtied by each vCPU. e.g. 10M or 3G.\n"
375 	       "     (default: 1G)\n");
376 	printf(" -f: specify the fraction of pages which should be written to\n"
377 	       "     as opposed to simply read, in the form\n"
378 	       "     1/<fraction of pages to write>.\n"
379 	       "     (default: 1 i.e. all pages are written to.)\n");
380 	printf(" -v: specify the number of vCPUs to run.\n");
381 	printf(" -o: Overlap guest memory accesses instead of partitioning\n"
382 	       "     them into a separate region of memory for each vCPU.\n");
383 	backing_src_help("-s");
384 	printf(" -x: Split the memory region into this number of memslots.\n"
385 	       "     (default: 1)\n");
386 	puts("");
387 	exit(0);
388 }
389 
390 int main(int argc, char *argv[])
391 {
392 	int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS);
393 	struct test_params p = {
394 		.iterations = TEST_HOST_LOOP_N,
395 		.wr_fract = 1,
396 		.partition_vcpu_memory_access = true,
397 		.backing_src = DEFAULT_VM_MEM_SRC,
398 		.slots = 1,
399 	};
400 	int opt;
401 
402 	dirty_log_manual_caps =
403 		kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
404 	dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
405 				  KVM_DIRTY_LOG_INITIALLY_SET);
406 
407 	guest_modes_append_default();
408 
409 	while ((opt = getopt(argc, argv, "eghi:p:m:nb:f:v:os:x:")) != -1) {
410 		switch (opt) {
411 		case 'e':
412 			/* 'e' is for evil. */
413 			run_vcpus_while_disabling_dirty_logging = true;
414 		case 'g':
415 			dirty_log_manual_caps = 0;
416 			break;
417 		case 'i':
418 			p.iterations = atoi(optarg);
419 			break;
420 		case 'p':
421 			p.phys_offset = strtoull(optarg, NULL, 0);
422 			break;
423 		case 'm':
424 			guest_modes_cmdline(optarg);
425 			break;
426 		case 'n':
427 			perf_test_args.nested = true;
428 			break;
429 		case 'b':
430 			guest_percpu_mem_size = parse_size(optarg);
431 			break;
432 		case 'f':
433 			p.wr_fract = atoi(optarg);
434 			TEST_ASSERT(p.wr_fract >= 1,
435 				    "Write fraction cannot be less than one");
436 			break;
437 		case 'v':
438 			nr_vcpus = atoi(optarg);
439 			TEST_ASSERT(nr_vcpus > 0 && nr_vcpus <= max_vcpus,
440 				    "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
441 			break;
442 		case 'o':
443 			p.partition_vcpu_memory_access = false;
444 			break;
445 		case 's':
446 			p.backing_src = parse_backing_src_type(optarg);
447 			break;
448 		case 'x':
449 			p.slots = atoi(optarg);
450 			break;
451 		case 'h':
452 		default:
453 			help(argv[0]);
454 			break;
455 		}
456 	}
457 
458 	TEST_ASSERT(p.iterations >= 2, "The test should have at least two iterations");
459 
460 	pr_info("Test iterations: %"PRIu64"\n",	p.iterations);
461 
462 	for_each_guest_mode(run_test, &p);
463 
464 	return 0;
465 }
466