1 // SPDX-License-Identifier: GPL-2.0
2 #define _GNU_SOURCE /* for program_invocation_short_name */
3 #include <fcntl.h>
4 #include <pthread.h>
5 #include <sched.h>
6 #include <semaphore.h>
7 #include <signal.h>
8 #include <stdio.h>
9 #include <stdlib.h>
10 #include <string.h>
11 #include <sys/ioctl.h>
12 #include <sys/mman.h>
13 
14 #include <linux/compiler.h>
15 
16 #include <test_util.h>
17 #include <kvm_util.h>
18 #include <processor.h>
19 
20 #define VCPU_ID 0
21 
22 /*
23  * s390x needs at least 1MB alignment, and the x86_64 MOVE/DELETE tests need a
24  * 2MB sized and aligned region so that the initial region corresponds to
25  * exactly one large page.
26  */
27 #define MEM_REGION_SIZE		0x200000
28 
29 #ifdef __x86_64__
30 /*
31  * Somewhat arbitrary location and slot, intended to not overlap anything.
32  */
33 #define MEM_REGION_GPA		0xc0000000
34 #define MEM_REGION_SLOT		10
35 
36 static const uint64_t MMIO_VAL = 0xbeefull;
37 
38 extern const uint64_t final_rip_start;
39 extern const uint64_t final_rip_end;
40 
41 static sem_t vcpu_ready;
42 
43 static inline uint64_t guest_spin_on_val(uint64_t spin_val)
44 {
45 	uint64_t val;
46 
47 	do {
48 		val = READ_ONCE(*((uint64_t *)MEM_REGION_GPA));
49 	} while (val == spin_val);
50 
51 	GUEST_SYNC(0);
52 	return val;
53 }
54 
55 static void *vcpu_worker(void *data)
56 {
57 	struct kvm_vm *vm = data;
58 	struct kvm_run *run;
59 	struct ucall uc;
60 	uint64_t cmd;
61 
62 	/*
63 	 * Loop until the guest is done.  Re-enter the guest on all MMIO exits,
64 	 * which will occur if the guest attempts to access a memslot after it
65 	 * has been deleted or while it is being moved .
66 	 */
67 	run = vcpu_state(vm, VCPU_ID);
68 
69 	while (1) {
70 		vcpu_run(vm, VCPU_ID);
71 
72 		if (run->exit_reason == KVM_EXIT_IO) {
73 			cmd = get_ucall(vm, VCPU_ID, &uc);
74 			if (cmd != UCALL_SYNC)
75 				break;
76 
77 			sem_post(&vcpu_ready);
78 			continue;
79 		}
80 
81 		if (run->exit_reason != KVM_EXIT_MMIO)
82 			break;
83 
84 		TEST_ASSERT(!run->mmio.is_write, "Unexpected exit mmio write");
85 		TEST_ASSERT(run->mmio.len == 8,
86 			    "Unexpected exit mmio size = %u", run->mmio.len);
87 
88 		TEST_ASSERT(run->mmio.phys_addr == MEM_REGION_GPA,
89 			    "Unexpected exit mmio address = 0x%llx",
90 			    run->mmio.phys_addr);
91 		memcpy(run->mmio.data, &MMIO_VAL, 8);
92 	}
93 
94 	if (run->exit_reason == KVM_EXIT_IO && cmd == UCALL_ABORT)
95 		TEST_FAIL("%s at %s:%ld, val = %lu", (const char *)uc.args[0],
96 			  __FILE__, uc.args[1], uc.args[2]);
97 
98 	return NULL;
99 }
100 
101 static void wait_for_vcpu(void)
102 {
103 	struct timespec ts;
104 
105 	TEST_ASSERT(!clock_gettime(CLOCK_REALTIME, &ts),
106 		    "clock_gettime() failed: %d\n", errno);
107 
108 	ts.tv_sec += 2;
109 	TEST_ASSERT(!sem_timedwait(&vcpu_ready, &ts),
110 		    "sem_timedwait() failed: %d\n", errno);
111 
112 	/* Wait for the vCPU thread to reenter the guest. */
113 	usleep(100000);
114 }
115 
116 static struct kvm_vm *spawn_vm(pthread_t *vcpu_thread, void *guest_code)
117 {
118 	struct kvm_vm *vm;
119 	uint64_t *hva;
120 	uint64_t gpa;
121 
122 	vm = vm_create_default(VCPU_ID, 0, guest_code);
123 
124 	vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS_THP,
125 				    MEM_REGION_GPA, MEM_REGION_SLOT,
126 				    MEM_REGION_SIZE / getpagesize(), 0);
127 
128 	/*
129 	 * Allocate and map two pages so that the GPA accessed by guest_code()
130 	 * stays valid across the memslot move.
131 	 */
132 	gpa = vm_phy_pages_alloc(vm, 2, MEM_REGION_GPA, MEM_REGION_SLOT);
133 	TEST_ASSERT(gpa == MEM_REGION_GPA, "Failed vm_phy_pages_alloc\n");
134 
135 	virt_map(vm, MEM_REGION_GPA, MEM_REGION_GPA, 2);
136 
137 	/* Ditto for the host mapping so that both pages can be zeroed. */
138 	hva = addr_gpa2hva(vm, MEM_REGION_GPA);
139 	memset(hva, 0, 2 * 4096);
140 
141 	pthread_create(vcpu_thread, NULL, vcpu_worker, vm);
142 
143 	/* Ensure the guest thread is spun up. */
144 	wait_for_vcpu();
145 
146 	return vm;
147 }
148 
149 
150 static void guest_code_move_memory_region(void)
151 {
152 	uint64_t val;
153 
154 	GUEST_SYNC(0);
155 
156 	/*
157 	 * Spin until the memory region starts getting moved to a
158 	 * misaligned address.
159 	 * Every region move may or may not trigger MMIO, as the
160 	 * window where the memslot is invalid is usually quite small.
161 	 */
162 	val = guest_spin_on_val(0);
163 	GUEST_ASSERT_1(val == 1 || val == MMIO_VAL, val);
164 
165 	/* Spin until the misaligning memory region move completes. */
166 	val = guest_spin_on_val(MMIO_VAL);
167 	GUEST_ASSERT_1(val == 1 || val == 0, val);
168 
169 	/* Spin until the memory region starts to get re-aligned. */
170 	val = guest_spin_on_val(0);
171 	GUEST_ASSERT_1(val == 1 || val == MMIO_VAL, val);
172 
173 	/* Spin until the re-aligning memory region move completes. */
174 	val = guest_spin_on_val(MMIO_VAL);
175 	GUEST_ASSERT_1(val == 1, val);
176 
177 	GUEST_DONE();
178 }
179 
180 static void test_move_memory_region(void)
181 {
182 	pthread_t vcpu_thread;
183 	struct kvm_vm *vm;
184 	uint64_t *hva;
185 
186 	vm = spawn_vm(&vcpu_thread, guest_code_move_memory_region);
187 
188 	hva = addr_gpa2hva(vm, MEM_REGION_GPA);
189 
190 	/*
191 	 * Shift the region's base GPA.  The guest should not see "2" as the
192 	 * hva->gpa translation is misaligned, i.e. the guest is accessing a
193 	 * different host pfn.
194 	 */
195 	vm_mem_region_move(vm, MEM_REGION_SLOT, MEM_REGION_GPA - 4096);
196 	WRITE_ONCE(*hva, 2);
197 
198 	/*
199 	 * The guest _might_ see an invalid memslot and trigger MMIO, but it's
200 	 * a tiny window.  Spin and defer the sync until the memslot is
201 	 * restored and guest behavior is once again deterministic.
202 	 */
203 	usleep(100000);
204 
205 	/*
206 	 * Note, value in memory needs to be changed *before* restoring the
207 	 * memslot, else the guest could race the update and see "2".
208 	 */
209 	WRITE_ONCE(*hva, 1);
210 
211 	/* Restore the original base, the guest should see "1". */
212 	vm_mem_region_move(vm, MEM_REGION_SLOT, MEM_REGION_GPA);
213 	wait_for_vcpu();
214 	/* Defered sync from when the memslot was misaligned (above). */
215 	wait_for_vcpu();
216 
217 	pthread_join(vcpu_thread, NULL);
218 
219 	kvm_vm_free(vm);
220 }
221 
222 static void guest_code_delete_memory_region(void)
223 {
224 	uint64_t val;
225 
226 	GUEST_SYNC(0);
227 
228 	/* Spin until the memory region is deleted. */
229 	val = guest_spin_on_val(0);
230 	GUEST_ASSERT_1(val == MMIO_VAL, val);
231 
232 	/* Spin until the memory region is recreated. */
233 	val = guest_spin_on_val(MMIO_VAL);
234 	GUEST_ASSERT_1(val == 0, val);
235 
236 	/* Spin until the memory region is deleted. */
237 	val = guest_spin_on_val(0);
238 	GUEST_ASSERT_1(val == MMIO_VAL, val);
239 
240 	asm("1:\n\t"
241 	    ".pushsection .rodata\n\t"
242 	    ".global final_rip_start\n\t"
243 	    "final_rip_start: .quad 1b\n\t"
244 	    ".popsection");
245 
246 	/* Spin indefinitely (until the code memslot is deleted). */
247 	guest_spin_on_val(MMIO_VAL);
248 
249 	asm("1:\n\t"
250 	    ".pushsection .rodata\n\t"
251 	    ".global final_rip_end\n\t"
252 	    "final_rip_end: .quad 1b\n\t"
253 	    ".popsection");
254 
255 	GUEST_ASSERT_1(0, 0);
256 }
257 
258 static void test_delete_memory_region(void)
259 {
260 	pthread_t vcpu_thread;
261 	struct kvm_regs regs;
262 	struct kvm_run *run;
263 	struct kvm_vm *vm;
264 
265 	vm = spawn_vm(&vcpu_thread, guest_code_delete_memory_region);
266 
267 	/* Delete the memory region, the guest should not die. */
268 	vm_mem_region_delete(vm, MEM_REGION_SLOT);
269 	wait_for_vcpu();
270 
271 	/* Recreate the memory region.  The guest should see "0". */
272 	vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS_THP,
273 				    MEM_REGION_GPA, MEM_REGION_SLOT,
274 				    MEM_REGION_SIZE / getpagesize(), 0);
275 	wait_for_vcpu();
276 
277 	/* Delete the region again so that there's only one memslot left. */
278 	vm_mem_region_delete(vm, MEM_REGION_SLOT);
279 	wait_for_vcpu();
280 
281 	/*
282 	 * Delete the primary memslot.  This should cause an emulation error or
283 	 * shutdown due to the page tables getting nuked.
284 	 */
285 	vm_mem_region_delete(vm, 0);
286 
287 	pthread_join(vcpu_thread, NULL);
288 
289 	run = vcpu_state(vm, VCPU_ID);
290 
291 	TEST_ASSERT(run->exit_reason == KVM_EXIT_SHUTDOWN ||
292 		    run->exit_reason == KVM_EXIT_INTERNAL_ERROR,
293 		    "Unexpected exit reason = %d", run->exit_reason);
294 
295 	vcpu_regs_get(vm, VCPU_ID, &regs);
296 
297 	/*
298 	 * On AMD, after KVM_EXIT_SHUTDOWN the VMCB has been reinitialized already,
299 	 * so the instruction pointer would point to the reset vector.
300 	 */
301 	if (run->exit_reason == KVM_EXIT_INTERNAL_ERROR)
302 		TEST_ASSERT(regs.rip >= final_rip_start &&
303 			    regs.rip < final_rip_end,
304 			    "Bad rip, expected 0x%lx - 0x%lx, got 0x%llx\n",
305 			    final_rip_start, final_rip_end, regs.rip);
306 
307 	kvm_vm_free(vm);
308 }
309 
310 static void test_zero_memory_regions(void)
311 {
312 	struct kvm_run *run;
313 	struct kvm_vm *vm;
314 
315 	pr_info("Testing KVM_RUN with zero added memory regions\n");
316 
317 	vm = vm_create(VM_MODE_DEFAULT, 0, O_RDWR);
318 	vm_vcpu_add(vm, VCPU_ID);
319 
320 	TEST_ASSERT(!ioctl(vm_get_fd(vm), KVM_SET_NR_MMU_PAGES, 64),
321 		    "KVM_SET_NR_MMU_PAGES failed, errno = %d\n", errno);
322 	vcpu_run(vm, VCPU_ID);
323 
324 	run = vcpu_state(vm, VCPU_ID);
325 	TEST_ASSERT(run->exit_reason == KVM_EXIT_INTERNAL_ERROR,
326 		    "Unexpected exit_reason = %u\n", run->exit_reason);
327 
328 	kvm_vm_free(vm);
329 }
330 #endif /* __x86_64__ */
331 
332 static int test_memory_region_add(struct kvm_vm *vm, void *mem, uint32_t slot,
333 				   uint32_t size, uint64_t guest_addr)
334 {
335 	struct kvm_userspace_memory_region region;
336 	int ret;
337 
338 	region.slot = slot;
339 	region.flags = 0;
340 	region.guest_phys_addr = guest_addr;
341 	region.memory_size = size;
342 	region.userspace_addr = (uintptr_t) mem;
343 	ret = ioctl(vm_get_fd(vm), KVM_SET_USER_MEMORY_REGION, &region);
344 
345 	return ret;
346 }
347 
348 /*
349  * Test it can be added memory slots up to KVM_CAP_NR_MEMSLOTS, then any
350  * tentative to add further slots should fail.
351  */
352 static void test_add_max_memory_regions(void)
353 {
354 	int ret;
355 	struct kvm_vm *vm;
356 	uint32_t max_mem_slots;
357 	uint32_t slot;
358 	void *mem, *mem_aligned, *mem_extra;
359 	size_t alignment;
360 
361 #ifdef __s390x__
362 	/* On s390x, the host address must be aligned to 1M (due to PGSTEs) */
363 	alignment = 0x100000;
364 #else
365 	alignment = 1;
366 #endif
367 
368 	max_mem_slots = kvm_check_cap(KVM_CAP_NR_MEMSLOTS);
369 	TEST_ASSERT(max_mem_slots > 0,
370 		    "KVM_CAP_NR_MEMSLOTS should be greater than 0");
371 	pr_info("Allowed number of memory slots: %i\n", max_mem_slots);
372 
373 	vm = vm_create(VM_MODE_DEFAULT, 0, O_RDWR);
374 
375 	/* Check it can be added memory slots up to the maximum allowed */
376 	pr_info("Adding slots 0..%i, each memory region with %dK size\n",
377 		(max_mem_slots - 1), MEM_REGION_SIZE >> 10);
378 
379 	mem = mmap(NULL, (size_t)max_mem_slots * MEM_REGION_SIZE + alignment,
380 		   PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
381 	TEST_ASSERT(mem != MAP_FAILED, "Failed to mmap() host");
382 	mem_aligned = (void *)(((size_t) mem + alignment - 1) & ~(alignment - 1));
383 
384 	for (slot = 0; slot < max_mem_slots; slot++) {
385 		ret = test_memory_region_add(vm, mem_aligned +
386 					     ((uint64_t)slot * MEM_REGION_SIZE),
387 					     slot, MEM_REGION_SIZE,
388 					     (uint64_t)slot * MEM_REGION_SIZE);
389 		TEST_ASSERT(ret == 0, "KVM_SET_USER_MEMORY_REGION IOCTL failed,\n"
390 			    "  rc: %i errno: %i slot: %i\n",
391 			    ret, errno, slot);
392 	}
393 
394 	/* Check it cannot be added memory slots beyond the limit */
395 	mem_extra = mmap(NULL, MEM_REGION_SIZE, PROT_READ | PROT_WRITE,
396 			 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
397 	TEST_ASSERT(mem_extra != MAP_FAILED, "Failed to mmap() host");
398 
399 	ret = test_memory_region_add(vm, mem_extra, max_mem_slots, MEM_REGION_SIZE,
400 				     (uint64_t)max_mem_slots * MEM_REGION_SIZE);
401 	TEST_ASSERT(ret == -1 && errno == EINVAL,
402 		    "Adding one more memory slot should fail with EINVAL");
403 
404 	munmap(mem, (size_t)max_mem_slots * MEM_REGION_SIZE + alignment);
405 	munmap(mem_extra, MEM_REGION_SIZE);
406 	kvm_vm_free(vm);
407 }
408 
409 int main(int argc, char *argv[])
410 {
411 #ifdef __x86_64__
412 	int i, loops;
413 #endif
414 
415 	/* Tell stdout not to buffer its content */
416 	setbuf(stdout, NULL);
417 
418 #ifdef __x86_64__
419 	/*
420 	 * FIXME: the zero-memslot test fails on aarch64 and s390x because
421 	 * KVM_RUN fails with ENOEXEC or EFAULT.
422 	 */
423 	test_zero_memory_regions();
424 #endif
425 
426 	test_add_max_memory_regions();
427 
428 #ifdef __x86_64__
429 	if (argc > 1)
430 		loops = atoi(argv[1]);
431 	else
432 		loops = 10;
433 
434 	pr_info("Testing MOVE of in-use region, %d loops\n", loops);
435 	for (i = 0; i < loops; i++)
436 		test_move_memory_region();
437 
438 	pr_info("Testing DELETE of in-use region, %d loops\n", loops);
439 	for (i = 0; i < loops; i++)
440 		test_delete_memory_region();
441 #endif
442 
443 	return 0;
444 }
445