1 // SPDX-License-Identifier: GPL-2.0
2 #define _GNU_SOURCE
3 
4 #include <stdio.h>
5 #include <stdlib.h>
6 #include <pthread.h>
7 #include <semaphore.h>
8 #include <sys/types.h>
9 #include <signal.h>
10 #include <errno.h>
11 #include <linux/bitmap.h>
12 #include <linux/bitops.h>
13 #include <linux/atomic.h>
14 #include <linux/sizes.h>
15 
16 #include "kvm_util.h"
17 #include "test_util.h"
18 #include "guest_modes.h"
19 #include "processor.h"
20 
guest_code(uint64_t start_gpa,uint64_t end_gpa,uint64_t stride)21 static void guest_code(uint64_t start_gpa, uint64_t end_gpa, uint64_t stride)
22 {
23 	uint64_t gpa;
24 
25 	for (gpa = start_gpa; gpa < end_gpa; gpa += stride)
26 		*((volatile uint64_t *)gpa) = gpa;
27 
28 	GUEST_DONE();
29 }
30 
31 struct vcpu_info {
32 	struct kvm_vcpu *vcpu;
33 	uint64_t start_gpa;
34 	uint64_t end_gpa;
35 };
36 
37 static int nr_vcpus;
38 static atomic_t rendezvous;
39 
rendezvous_with_boss(void)40 static void rendezvous_with_boss(void)
41 {
42 	int orig = atomic_read(&rendezvous);
43 
44 	if (orig > 0) {
45 		atomic_dec_and_test(&rendezvous);
46 		while (atomic_read(&rendezvous) > 0)
47 			cpu_relax();
48 	} else {
49 		atomic_inc(&rendezvous);
50 		while (atomic_read(&rendezvous) < 0)
51 			cpu_relax();
52 	}
53 }
54 
run_vcpu(struct kvm_vcpu * vcpu)55 static void run_vcpu(struct kvm_vcpu *vcpu)
56 {
57 	vcpu_run(vcpu);
58 	TEST_ASSERT_EQ(get_ucall(vcpu, NULL), UCALL_DONE);
59 }
60 
vcpu_worker(void * data)61 static void *vcpu_worker(void *data)
62 {
63 	struct vcpu_info *info = data;
64 	struct kvm_vcpu *vcpu = info->vcpu;
65 	struct kvm_vm *vm = vcpu->vm;
66 	struct kvm_sregs sregs;
67 	struct kvm_regs regs;
68 
69 	vcpu_args_set(vcpu, 3, info->start_gpa, info->end_gpa, vm->page_size);
70 
71 	/* Snapshot regs before the first run. */
72 	vcpu_regs_get(vcpu, &regs);
73 	rendezvous_with_boss();
74 
75 	run_vcpu(vcpu);
76 	rendezvous_with_boss();
77 	vcpu_regs_set(vcpu, &regs);
78 	vcpu_sregs_get(vcpu, &sregs);
79 #ifdef __x86_64__
80 	/* Toggle CR0.WP to trigger a MMU context reset. */
81 	sregs.cr0 ^= X86_CR0_WP;
82 #endif
83 	vcpu_sregs_set(vcpu, &sregs);
84 	rendezvous_with_boss();
85 
86 	run_vcpu(vcpu);
87 	rendezvous_with_boss();
88 
89 	return NULL;
90 }
91 
spawn_workers(struct kvm_vm * vm,struct kvm_vcpu ** vcpus,uint64_t start_gpa,uint64_t end_gpa)92 static pthread_t *spawn_workers(struct kvm_vm *vm, struct kvm_vcpu **vcpus,
93 				uint64_t start_gpa, uint64_t end_gpa)
94 {
95 	struct vcpu_info *info;
96 	uint64_t gpa, nr_bytes;
97 	pthread_t *threads;
98 	int i;
99 
100 	threads = malloc(nr_vcpus * sizeof(*threads));
101 	TEST_ASSERT(threads, "Failed to allocate vCPU threads");
102 
103 	info = malloc(nr_vcpus * sizeof(*info));
104 	TEST_ASSERT(info, "Failed to allocate vCPU gpa ranges");
105 
106 	nr_bytes = ((end_gpa - start_gpa) / nr_vcpus) &
107 			~((uint64_t)vm->page_size - 1);
108 	TEST_ASSERT(nr_bytes, "C'mon, no way you have %d CPUs", nr_vcpus);
109 
110 	for (i = 0, gpa = start_gpa; i < nr_vcpus; i++, gpa += nr_bytes) {
111 		info[i].vcpu = vcpus[i];
112 		info[i].start_gpa = gpa;
113 		info[i].end_gpa = gpa + nr_bytes;
114 		pthread_create(&threads[i], NULL, vcpu_worker, &info[i]);
115 	}
116 	return threads;
117 }
118 
rendezvous_with_vcpus(struct timespec * time,const char * name)119 static void rendezvous_with_vcpus(struct timespec *time, const char *name)
120 {
121 	int i, rendezvoused;
122 
123 	pr_info("Waiting for vCPUs to finish %s...\n", name);
124 
125 	rendezvoused = atomic_read(&rendezvous);
126 	for (i = 0; abs(rendezvoused) != 1; i++) {
127 		usleep(100);
128 		if (!(i & 0x3f))
129 			pr_info("\r%d vCPUs haven't rendezvoused...",
130 				abs(rendezvoused) - 1);
131 		rendezvoused = atomic_read(&rendezvous);
132 	}
133 
134 	clock_gettime(CLOCK_MONOTONIC, time);
135 
136 	/* Release the vCPUs after getting the time of the previous action. */
137 	pr_info("\rAll vCPUs finished %s, releasing...\n", name);
138 	if (rendezvoused > 0)
139 		atomic_set(&rendezvous, -nr_vcpus - 1);
140 	else
141 		atomic_set(&rendezvous, nr_vcpus + 1);
142 }
143 
calc_default_nr_vcpus(void)144 static void calc_default_nr_vcpus(void)
145 {
146 	cpu_set_t possible_mask;
147 	int r;
148 
149 	r = sched_getaffinity(0, sizeof(possible_mask), &possible_mask);
150 	TEST_ASSERT(!r, "sched_getaffinity failed, errno = %d (%s)",
151 		    errno, strerror(errno));
152 
153 	nr_vcpus = CPU_COUNT(&possible_mask) * 3/4;
154 	TEST_ASSERT(nr_vcpus > 0, "Uh, no CPUs?");
155 }
156 
main(int argc,char * argv[])157 int main(int argc, char *argv[])
158 {
159 	/*
160 	 * Skip the first 4gb and slot0.  slot0 maps <1gb and is used to back
161 	 * the guest's code, stack, and page tables.  Because selftests creates
162 	 * an IRQCHIP, a.k.a. a local APIC, KVM creates an internal memslot
163 	 * just below the 4gb boundary.  This test could create memory at
164 	 * 1gb-3gb,but it's simpler to skip straight to 4gb.
165 	 */
166 	const uint64_t start_gpa = SZ_4G;
167 	const int first_slot = 1;
168 
169 	struct timespec time_start, time_run1, time_reset, time_run2;
170 	uint64_t max_gpa, gpa, slot_size, max_mem, i;
171 	int max_slots, slot, opt, fd;
172 	bool hugepages = false;
173 	struct kvm_vcpu **vcpus;
174 	pthread_t *threads;
175 	struct kvm_vm *vm;
176 	void *mem;
177 
178 	/*
179 	 * Default to 2gb so that maxing out systems with MAXPHADDR=46, which
180 	 * are quite common for x86, requires changing only max_mem (KVM allows
181 	 * 32k memslots, 32k * 2gb == ~64tb of guest memory).
182 	 */
183 	slot_size = SZ_2G;
184 
185 	max_slots = kvm_check_cap(KVM_CAP_NR_MEMSLOTS);
186 	TEST_ASSERT(max_slots > first_slot, "KVM is broken");
187 
188 	/* All KVM MMUs should be able to survive a 128gb guest. */
189 	max_mem = 128ull * SZ_1G;
190 
191 	calc_default_nr_vcpus();
192 
193 	while ((opt = getopt(argc, argv, "c:h:m:s:H")) != -1) {
194 		switch (opt) {
195 		case 'c':
196 			nr_vcpus = atoi_positive("Number of vCPUs", optarg);
197 			break;
198 		case 'm':
199 			max_mem = 1ull * atoi_positive("Memory size", optarg) * SZ_1G;
200 			break;
201 		case 's':
202 			slot_size = 1ull * atoi_positive("Slot size", optarg) * SZ_1G;
203 			break;
204 		case 'H':
205 			hugepages = true;
206 			break;
207 		case 'h':
208 		default:
209 			printf("usage: %s [-c nr_vcpus] [-m max_mem_in_gb] [-s slot_size_in_gb] [-H]\n", argv[0]);
210 			exit(1);
211 		}
212 	}
213 
214 	vcpus = malloc(nr_vcpus * sizeof(*vcpus));
215 	TEST_ASSERT(vcpus, "Failed to allocate vCPU array");
216 
217 	vm = vm_create_with_vcpus(nr_vcpus, guest_code, vcpus);
218 
219 	max_gpa = vm->max_gfn << vm->page_shift;
220 	TEST_ASSERT(max_gpa > (4 * slot_size), "MAXPHYADDR <4gb ");
221 
222 	fd = kvm_memfd_alloc(slot_size, hugepages);
223 	mem = mmap(NULL, slot_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
224 	TEST_ASSERT(mem != MAP_FAILED, "mmap() failed");
225 
226 	TEST_ASSERT(!madvise(mem, slot_size, MADV_NOHUGEPAGE), "madvise() failed");
227 
228 	/* Pre-fault the memory to avoid taking mmap_sem on guest page faults. */
229 	for (i = 0; i < slot_size; i += vm->page_size)
230 		((uint8_t *)mem)[i] = 0xaa;
231 
232 	gpa = 0;
233 	for (slot = first_slot; slot < max_slots; slot++) {
234 		gpa = start_gpa + ((slot - first_slot) * slot_size);
235 		if (gpa + slot_size > max_gpa)
236 			break;
237 
238 		if ((gpa - start_gpa) >= max_mem)
239 			break;
240 
241 		vm_set_user_memory_region(vm, slot, 0, gpa, slot_size, mem);
242 
243 #ifdef __x86_64__
244 		/* Identity map memory in the guest using 1gb pages. */
245 		for (i = 0; i < slot_size; i += SZ_1G)
246 			__virt_pg_map(vm, gpa + i, gpa + i, PG_LEVEL_1G);
247 #else
248 		for (i = 0; i < slot_size; i += vm->page_size)
249 			virt_pg_map(vm, gpa + i, gpa + i);
250 #endif
251 	}
252 
253 	atomic_set(&rendezvous, nr_vcpus + 1);
254 	threads = spawn_workers(vm, vcpus, start_gpa, gpa);
255 
256 	free(vcpus);
257 	vcpus = NULL;
258 
259 	pr_info("Running with %lugb of guest memory and %u vCPUs\n",
260 		(gpa - start_gpa) / SZ_1G, nr_vcpus);
261 
262 	rendezvous_with_vcpus(&time_start, "spawning");
263 	rendezvous_with_vcpus(&time_run1, "run 1");
264 	rendezvous_with_vcpus(&time_reset, "reset");
265 	rendezvous_with_vcpus(&time_run2, "run 2");
266 
267 	time_run2  = timespec_sub(time_run2,   time_reset);
268 	time_reset = timespec_sub(time_reset, time_run1);
269 	time_run1  = timespec_sub(time_run1,   time_start);
270 
271 	pr_info("run1 = %ld.%.9lds, reset = %ld.%.9lds, run2 =  %ld.%.9lds\n",
272 		time_run1.tv_sec, time_run1.tv_nsec,
273 		time_reset.tv_sec, time_reset.tv_nsec,
274 		time_run2.tv_sec, time_run2.tv_nsec);
275 
276 	/*
277 	 * Delete even numbered slots (arbitrary) and unmap the first half of
278 	 * the backing (also arbitrary) to verify KVM correctly drops all
279 	 * references to the removed regions.
280 	 */
281 	for (slot = (slot - 1) & ~1ull; slot >= first_slot; slot -= 2)
282 		vm_set_user_memory_region(vm, slot, 0, 0, 0, NULL);
283 
284 	munmap(mem, slot_size / 2);
285 
286 	/* Sanity check that the vCPUs actually ran. */
287 	for (i = 0; i < nr_vcpus; i++)
288 		pthread_join(threads[i], NULL);
289 
290 	/*
291 	 * Deliberately exit without deleting the remaining memslots or closing
292 	 * kvm_fd to test cleanup via mmu_notifier.release.
293 	 */
294 }
295