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 /*
21  * s390x needs at least 1MB alignment, and the x86_64 MOVE/DELETE tests need a
22  * 2MB sized and aligned region so that the initial region corresponds to
23  * exactly one large page.
24  */
25 #define MEM_REGION_SIZE		0x200000
26 
27 #ifdef __x86_64__
28 /*
29  * Somewhat arbitrary location and slot, intended to not overlap anything.
30  */
31 #define MEM_REGION_GPA		0xc0000000
32 #define MEM_REGION_SLOT		10
33 
34 static const uint64_t MMIO_VAL = 0xbeefull;
35 
36 extern const uint64_t final_rip_start;
37 extern const uint64_t final_rip_end;
38 
39 static sem_t vcpu_ready;
40 
41 static inline uint64_t guest_spin_on_val(uint64_t spin_val)
42 {
43 	uint64_t val;
44 
45 	do {
46 		val = READ_ONCE(*((uint64_t *)MEM_REGION_GPA));
47 	} while (val == spin_val);
48 
49 	GUEST_SYNC(0);
50 	return val;
51 }
52 
53 static void *vcpu_worker(void *data)
54 {
55 	struct kvm_vcpu *vcpu = data;
56 	struct kvm_run *run = vcpu->run;
57 	struct ucall uc;
58 	uint64_t cmd;
59 
60 	/*
61 	 * Loop until the guest is done.  Re-enter the guest on all MMIO exits,
62 	 * which will occur if the guest attempts to access a memslot after it
63 	 * has been deleted or while it is being moved .
64 	 */
65 	while (1) {
66 		vcpu_run(vcpu);
67 
68 		if (run->exit_reason == KVM_EXIT_IO) {
69 			cmd = get_ucall(vcpu, &uc);
70 			if (cmd != UCALL_SYNC)
71 				break;
72 
73 			sem_post(&vcpu_ready);
74 			continue;
75 		}
76 
77 		if (run->exit_reason != KVM_EXIT_MMIO)
78 			break;
79 
80 		TEST_ASSERT(!run->mmio.is_write, "Unexpected exit mmio write");
81 		TEST_ASSERT(run->mmio.len == 8,
82 			    "Unexpected exit mmio size = %u", run->mmio.len);
83 
84 		TEST_ASSERT(run->mmio.phys_addr == MEM_REGION_GPA,
85 			    "Unexpected exit mmio address = 0x%llx",
86 			    run->mmio.phys_addr);
87 		memcpy(run->mmio.data, &MMIO_VAL, 8);
88 	}
89 
90 	if (run->exit_reason == KVM_EXIT_IO && cmd == UCALL_ABORT)
91 		REPORT_GUEST_ASSERT_1(uc, "val = %lu");
92 
93 	return NULL;
94 }
95 
96 static void wait_for_vcpu(void)
97 {
98 	struct timespec ts;
99 
100 	TEST_ASSERT(!clock_gettime(CLOCK_REALTIME, &ts),
101 		    "clock_gettime() failed: %d\n", errno);
102 
103 	ts.tv_sec += 2;
104 	TEST_ASSERT(!sem_timedwait(&vcpu_ready, &ts),
105 		    "sem_timedwait() failed: %d\n", errno);
106 
107 	/* Wait for the vCPU thread to reenter the guest. */
108 	usleep(100000);
109 }
110 
111 static struct kvm_vm *spawn_vm(struct kvm_vcpu **vcpu, pthread_t *vcpu_thread,
112 			       void *guest_code)
113 {
114 	struct kvm_vm *vm;
115 	uint64_t *hva;
116 	uint64_t gpa;
117 
118 	vm = vm_create_with_one_vcpu(vcpu, guest_code);
119 
120 	vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS_THP,
121 				    MEM_REGION_GPA, MEM_REGION_SLOT,
122 				    MEM_REGION_SIZE / getpagesize(), 0);
123 
124 	/*
125 	 * Allocate and map two pages so that the GPA accessed by guest_code()
126 	 * stays valid across the memslot move.
127 	 */
128 	gpa = vm_phy_pages_alloc(vm, 2, MEM_REGION_GPA, MEM_REGION_SLOT);
129 	TEST_ASSERT(gpa == MEM_REGION_GPA, "Failed vm_phy_pages_alloc\n");
130 
131 	virt_map(vm, MEM_REGION_GPA, MEM_REGION_GPA, 2);
132 
133 	/* Ditto for the host mapping so that both pages can be zeroed. */
134 	hva = addr_gpa2hva(vm, MEM_REGION_GPA);
135 	memset(hva, 0, 2 * 4096);
136 
137 	pthread_create(vcpu_thread, NULL, vcpu_worker, *vcpu);
138 
139 	/* Ensure the guest thread is spun up. */
140 	wait_for_vcpu();
141 
142 	return vm;
143 }
144 
145 
146 static void guest_code_move_memory_region(void)
147 {
148 	uint64_t val;
149 
150 	GUEST_SYNC(0);
151 
152 	/*
153 	 * Spin until the memory region starts getting moved to a
154 	 * misaligned address.
155 	 * Every region move may or may not trigger MMIO, as the
156 	 * window where the memslot is invalid is usually quite small.
157 	 */
158 	val = guest_spin_on_val(0);
159 	GUEST_ASSERT_1(val == 1 || val == MMIO_VAL, val);
160 
161 	/* Spin until the misaligning memory region move completes. */
162 	val = guest_spin_on_val(MMIO_VAL);
163 	GUEST_ASSERT_1(val == 1 || val == 0, val);
164 
165 	/* Spin until the memory region starts to get re-aligned. */
166 	val = guest_spin_on_val(0);
167 	GUEST_ASSERT_1(val == 1 || val == MMIO_VAL, val);
168 
169 	/* Spin until the re-aligning memory region move completes. */
170 	val = guest_spin_on_val(MMIO_VAL);
171 	GUEST_ASSERT_1(val == 1, val);
172 
173 	GUEST_DONE();
174 }
175 
176 static void test_move_memory_region(void)
177 {
178 	pthread_t vcpu_thread;
179 	struct kvm_vcpu *vcpu;
180 	struct kvm_vm *vm;
181 	uint64_t *hva;
182 
183 	vm = spawn_vm(&vcpu, &vcpu_thread, guest_code_move_memory_region);
184 
185 	hva = addr_gpa2hva(vm, MEM_REGION_GPA);
186 
187 	/*
188 	 * Shift the region's base GPA.  The guest should not see "2" as the
189 	 * hva->gpa translation is misaligned, i.e. the guest is accessing a
190 	 * different host pfn.
191 	 */
192 	vm_mem_region_move(vm, MEM_REGION_SLOT, MEM_REGION_GPA - 4096);
193 	WRITE_ONCE(*hva, 2);
194 
195 	/*
196 	 * The guest _might_ see an invalid memslot and trigger MMIO, but it's
197 	 * a tiny window.  Spin and defer the sync until the memslot is
198 	 * restored and guest behavior is once again deterministic.
199 	 */
200 	usleep(100000);
201 
202 	/*
203 	 * Note, value in memory needs to be changed *before* restoring the
204 	 * memslot, else the guest could race the update and see "2".
205 	 */
206 	WRITE_ONCE(*hva, 1);
207 
208 	/* Restore the original base, the guest should see "1". */
209 	vm_mem_region_move(vm, MEM_REGION_SLOT, MEM_REGION_GPA);
210 	wait_for_vcpu();
211 	/* Defered sync from when the memslot was misaligned (above). */
212 	wait_for_vcpu();
213 
214 	pthread_join(vcpu_thread, NULL);
215 
216 	kvm_vm_free(vm);
217 }
218 
219 static void guest_code_delete_memory_region(void)
220 {
221 	uint64_t val;
222 
223 	GUEST_SYNC(0);
224 
225 	/* Spin until the memory region is deleted. */
226 	val = guest_spin_on_val(0);
227 	GUEST_ASSERT_1(val == MMIO_VAL, val);
228 
229 	/* Spin until the memory region is recreated. */
230 	val = guest_spin_on_val(MMIO_VAL);
231 	GUEST_ASSERT_1(val == 0, val);
232 
233 	/* Spin until the memory region is deleted. */
234 	val = guest_spin_on_val(0);
235 	GUEST_ASSERT_1(val == MMIO_VAL, val);
236 
237 	asm("1:\n\t"
238 	    ".pushsection .rodata\n\t"
239 	    ".global final_rip_start\n\t"
240 	    "final_rip_start: .quad 1b\n\t"
241 	    ".popsection");
242 
243 	/* Spin indefinitely (until the code memslot is deleted). */
244 	guest_spin_on_val(MMIO_VAL);
245 
246 	asm("1:\n\t"
247 	    ".pushsection .rodata\n\t"
248 	    ".global final_rip_end\n\t"
249 	    "final_rip_end: .quad 1b\n\t"
250 	    ".popsection");
251 
252 	GUEST_ASSERT_1(0, 0);
253 }
254 
255 static void test_delete_memory_region(void)
256 {
257 	pthread_t vcpu_thread;
258 	struct kvm_vcpu *vcpu;
259 	struct kvm_regs regs;
260 	struct kvm_run *run;
261 	struct kvm_vm *vm;
262 
263 	vm = spawn_vm(&vcpu, &vcpu_thread, guest_code_delete_memory_region);
264 
265 	/* Delete the memory region, the guest should not die. */
266 	vm_mem_region_delete(vm, MEM_REGION_SLOT);
267 	wait_for_vcpu();
268 
269 	/* Recreate the memory region.  The guest should see "0". */
270 	vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS_THP,
271 				    MEM_REGION_GPA, MEM_REGION_SLOT,
272 				    MEM_REGION_SIZE / getpagesize(), 0);
273 	wait_for_vcpu();
274 
275 	/* Delete the region again so that there's only one memslot left. */
276 	vm_mem_region_delete(vm, MEM_REGION_SLOT);
277 	wait_for_vcpu();
278 
279 	/*
280 	 * Delete the primary memslot.  This should cause an emulation error or
281 	 * shutdown due to the page tables getting nuked.
282 	 */
283 	vm_mem_region_delete(vm, 0);
284 
285 	pthread_join(vcpu_thread, NULL);
286 
287 	run = vcpu->run;
288 
289 	TEST_ASSERT(run->exit_reason == KVM_EXIT_SHUTDOWN ||
290 		    run->exit_reason == KVM_EXIT_INTERNAL_ERROR,
291 		    "Unexpected exit reason = %d", run->exit_reason);
292 
293 	vcpu_regs_get(vcpu, &regs);
294 
295 	/*
296 	 * On AMD, after KVM_EXIT_SHUTDOWN the VMCB has been reinitialized already,
297 	 * so the instruction pointer would point to the reset vector.
298 	 */
299 	if (run->exit_reason == KVM_EXIT_INTERNAL_ERROR)
300 		TEST_ASSERT(regs.rip >= final_rip_start &&
301 			    regs.rip < final_rip_end,
302 			    "Bad rip, expected 0x%lx - 0x%lx, got 0x%llx\n",
303 			    final_rip_start, final_rip_end, regs.rip);
304 
305 	kvm_vm_free(vm);
306 }
307 
308 static void test_zero_memory_regions(void)
309 {
310 	struct kvm_vcpu *vcpu;
311 	struct kvm_run *run;
312 	struct kvm_vm *vm;
313 
314 	pr_info("Testing KVM_RUN with zero added memory regions\n");
315 
316 	vm = vm_create_barebones();
317 	vcpu = __vm_vcpu_add(vm, 0);
318 
319 	vm_ioctl(vm, KVM_SET_NR_MMU_PAGES, (void *)64ul);
320 	vcpu_run(vcpu);
321 
322 	run = vcpu->run;
323 	TEST_ASSERT(run->exit_reason == KVM_EXIT_INTERNAL_ERROR,
324 		    "Unexpected exit_reason = %u\n", run->exit_reason);
325 
326 	kvm_vm_free(vm);
327 }
328 #endif /* __x86_64__ */
329 
330 /*
331  * Test it can be added memory slots up to KVM_CAP_NR_MEMSLOTS, then any
332  * tentative to add further slots should fail.
333  */
334 static void test_add_max_memory_regions(void)
335 {
336 	int ret;
337 	struct kvm_vm *vm;
338 	uint32_t max_mem_slots;
339 	uint32_t slot;
340 	void *mem, *mem_aligned, *mem_extra;
341 	size_t alignment;
342 
343 #ifdef __s390x__
344 	/* On s390x, the host address must be aligned to 1M (due to PGSTEs) */
345 	alignment = 0x100000;
346 #else
347 	alignment = 1;
348 #endif
349 
350 	max_mem_slots = kvm_check_cap(KVM_CAP_NR_MEMSLOTS);
351 	TEST_ASSERT(max_mem_slots > 0,
352 		    "KVM_CAP_NR_MEMSLOTS should be greater than 0");
353 	pr_info("Allowed number of memory slots: %i\n", max_mem_slots);
354 
355 	vm = vm_create_barebones();
356 
357 	/* Check it can be added memory slots up to the maximum allowed */
358 	pr_info("Adding slots 0..%i, each memory region with %dK size\n",
359 		(max_mem_slots - 1), MEM_REGION_SIZE >> 10);
360 
361 	mem = mmap(NULL, (size_t)max_mem_slots * MEM_REGION_SIZE + alignment,
362 		   PROT_READ | PROT_WRITE,
363 		   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
364 	TEST_ASSERT(mem != MAP_FAILED, "Failed to mmap() host");
365 	mem_aligned = (void *)(((size_t) mem + alignment - 1) & ~(alignment - 1));
366 
367 	for (slot = 0; slot < max_mem_slots; slot++)
368 		vm_set_user_memory_region(vm, slot, 0,
369 					  ((uint64_t)slot * MEM_REGION_SIZE),
370 					  MEM_REGION_SIZE,
371 					  mem_aligned + (uint64_t)slot * MEM_REGION_SIZE);
372 
373 	/* Check it cannot be added memory slots beyond the limit */
374 	mem_extra = mmap(NULL, MEM_REGION_SIZE, PROT_READ | PROT_WRITE,
375 			 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
376 	TEST_ASSERT(mem_extra != MAP_FAILED, "Failed to mmap() host");
377 
378 	ret = __vm_set_user_memory_region(vm, max_mem_slots, 0,
379 					  (uint64_t)max_mem_slots * MEM_REGION_SIZE,
380 					  MEM_REGION_SIZE, mem_extra);
381 	TEST_ASSERT(ret == -1 && errno == EINVAL,
382 		    "Adding one more memory slot should fail with EINVAL");
383 
384 	munmap(mem, (size_t)max_mem_slots * MEM_REGION_SIZE + alignment);
385 	munmap(mem_extra, MEM_REGION_SIZE);
386 	kvm_vm_free(vm);
387 }
388 
389 int main(int argc, char *argv[])
390 {
391 #ifdef __x86_64__
392 	int i, loops;
393 #endif
394 
395 #ifdef __x86_64__
396 	/*
397 	 * FIXME: the zero-memslot test fails on aarch64 and s390x because
398 	 * KVM_RUN fails with ENOEXEC or EFAULT.
399 	 */
400 	test_zero_memory_regions();
401 #endif
402 
403 	test_add_max_memory_regions();
404 
405 #ifdef __x86_64__
406 	if (argc > 1)
407 		loops = atoi_positive("Number of iterations", argv[1]);
408 	else
409 		loops = 10;
410 
411 	pr_info("Testing MOVE of in-use region, %d loops\n", loops);
412 	for (i = 0; i < loops; i++)
413 		test_move_memory_region();
414 
415 	pr_info("Testing DELETE of in-use region, %d loops\n", loops);
416 	for (i = 0; i < loops; i++)
417 		test_delete_memory_region();
418 #endif
419 
420 	return 0;
421 }
422