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 vcpu_set_cpuid(vm, VCPU_ID, kvm_get_supported_cpuid()); 125 126 vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS_THP, 127 MEM_REGION_GPA, MEM_REGION_SLOT, 128 MEM_REGION_SIZE / getpagesize(), 0); 129 130 /* 131 * Allocate and map two pages so that the GPA accessed by guest_code() 132 * stays valid across the memslot move. 133 */ 134 gpa = vm_phy_pages_alloc(vm, 2, MEM_REGION_GPA, MEM_REGION_SLOT); 135 TEST_ASSERT(gpa == MEM_REGION_GPA, "Failed vm_phy_pages_alloc\n"); 136 137 virt_map(vm, MEM_REGION_GPA, MEM_REGION_GPA, 2, 0); 138 139 /* Ditto for the host mapping so that both pages can be zeroed. */ 140 hva = addr_gpa2hva(vm, MEM_REGION_GPA); 141 memset(hva, 0, 2 * 4096); 142 143 pthread_create(vcpu_thread, NULL, vcpu_worker, vm); 144 145 /* Ensure the guest thread is spun up. */ 146 wait_for_vcpu(); 147 148 return vm; 149 } 150 151 152 static void guest_code_move_memory_region(void) 153 { 154 uint64_t val; 155 156 GUEST_SYNC(0); 157 158 /* 159 * Spin until the memory region starts getting moved to a 160 * misaligned address. 161 * Every region move may or may not trigger MMIO, as the 162 * window where the memslot is invalid is usually quite small. 163 */ 164 val = guest_spin_on_val(0); 165 GUEST_ASSERT_1(val == 1 || val == MMIO_VAL, val); 166 167 /* Spin until the misaligning memory region move completes. */ 168 val = guest_spin_on_val(MMIO_VAL); 169 GUEST_ASSERT_1(val == 1 || val == 0, val); 170 171 /* Spin until the memory region starts to get re-aligned. */ 172 val = guest_spin_on_val(0); 173 GUEST_ASSERT_1(val == 1 || val == MMIO_VAL, val); 174 175 /* Spin until the re-aligning memory region move completes. */ 176 val = guest_spin_on_val(MMIO_VAL); 177 GUEST_ASSERT_1(val == 1, val); 178 179 GUEST_DONE(); 180 } 181 182 static void test_move_memory_region(void) 183 { 184 pthread_t vcpu_thread; 185 struct kvm_vm *vm; 186 uint64_t *hva; 187 188 vm = spawn_vm(&vcpu_thread, guest_code_move_memory_region); 189 190 hva = addr_gpa2hva(vm, MEM_REGION_GPA); 191 192 /* 193 * Shift the region's base GPA. The guest should not see "2" as the 194 * hva->gpa translation is misaligned, i.e. the guest is accessing a 195 * different host pfn. 196 */ 197 vm_mem_region_move(vm, MEM_REGION_SLOT, MEM_REGION_GPA - 4096); 198 WRITE_ONCE(*hva, 2); 199 200 /* 201 * The guest _might_ see an invalid memslot and trigger MMIO, but it's 202 * a tiny window. Spin and defer the sync until the memslot is 203 * restored and guest behavior is once again deterministic. 204 */ 205 usleep(100000); 206 207 /* 208 * Note, value in memory needs to be changed *before* restoring the 209 * memslot, else the guest could race the update and see "2". 210 */ 211 WRITE_ONCE(*hva, 1); 212 213 /* Restore the original base, the guest should see "1". */ 214 vm_mem_region_move(vm, MEM_REGION_SLOT, MEM_REGION_GPA); 215 wait_for_vcpu(); 216 /* Defered sync from when the memslot was misaligned (above). */ 217 wait_for_vcpu(); 218 219 pthread_join(vcpu_thread, NULL); 220 221 kvm_vm_free(vm); 222 } 223 224 static void guest_code_delete_memory_region(void) 225 { 226 uint64_t val; 227 228 GUEST_SYNC(0); 229 230 /* Spin until the memory region is deleted. */ 231 val = guest_spin_on_val(0); 232 GUEST_ASSERT_1(val == MMIO_VAL, val); 233 234 /* Spin until the memory region is recreated. */ 235 val = guest_spin_on_val(MMIO_VAL); 236 GUEST_ASSERT_1(val == 0, val); 237 238 /* Spin until the memory region is deleted. */ 239 val = guest_spin_on_val(0); 240 GUEST_ASSERT_1(val == MMIO_VAL, val); 241 242 asm("1:\n\t" 243 ".pushsection .rodata\n\t" 244 ".global final_rip_start\n\t" 245 "final_rip_start: .quad 1b\n\t" 246 ".popsection"); 247 248 /* Spin indefinitely (until the code memslot is deleted). */ 249 guest_spin_on_val(MMIO_VAL); 250 251 asm("1:\n\t" 252 ".pushsection .rodata\n\t" 253 ".global final_rip_end\n\t" 254 "final_rip_end: .quad 1b\n\t" 255 ".popsection"); 256 257 GUEST_ASSERT_1(0, 0); 258 } 259 260 static void test_delete_memory_region(void) 261 { 262 pthread_t vcpu_thread; 263 struct kvm_regs regs; 264 struct kvm_run *run; 265 struct kvm_vm *vm; 266 267 vm = spawn_vm(&vcpu_thread, guest_code_delete_memory_region); 268 269 /* Delete the memory region, the guest should not die. */ 270 vm_mem_region_delete(vm, MEM_REGION_SLOT); 271 wait_for_vcpu(); 272 273 /* Recreate the memory region. The guest should see "0". */ 274 vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS_THP, 275 MEM_REGION_GPA, MEM_REGION_SLOT, 276 MEM_REGION_SIZE / getpagesize(), 0); 277 wait_for_vcpu(); 278 279 /* Delete the region again so that there's only one memslot left. */ 280 vm_mem_region_delete(vm, MEM_REGION_SLOT); 281 wait_for_vcpu(); 282 283 /* 284 * Delete the primary memslot. This should cause an emulation error or 285 * shutdown due to the page tables getting nuked. 286 */ 287 vm_mem_region_delete(vm, 0); 288 289 pthread_join(vcpu_thread, NULL); 290 291 run = vcpu_state(vm, VCPU_ID); 292 293 TEST_ASSERT(run->exit_reason == KVM_EXIT_SHUTDOWN || 294 run->exit_reason == KVM_EXIT_INTERNAL_ERROR, 295 "Unexpected exit reason = %d", run->exit_reason); 296 297 vcpu_regs_get(vm, VCPU_ID, ®s); 298 299 /* 300 * On AMD, after KVM_EXIT_SHUTDOWN the VMCB has been reinitialized already, 301 * so the instruction pointer would point to the reset vector. 302 */ 303 if (run->exit_reason == KVM_EXIT_INTERNAL_ERROR) 304 TEST_ASSERT(regs.rip >= final_rip_start && 305 regs.rip < final_rip_end, 306 "Bad rip, expected 0x%lx - 0x%lx, got 0x%llx\n", 307 final_rip_start, final_rip_end, regs.rip); 308 309 kvm_vm_free(vm); 310 } 311 312 static void test_zero_memory_regions(void) 313 { 314 struct kvm_run *run; 315 struct kvm_vm *vm; 316 317 pr_info("Testing KVM_RUN with zero added memory regions\n"); 318 319 vm = vm_create(VM_MODE_DEFAULT, 0, O_RDWR); 320 vm_vcpu_add(vm, VCPU_ID); 321 322 TEST_ASSERT(!ioctl(vm_get_fd(vm), KVM_SET_NR_MMU_PAGES, 64), 323 "KVM_SET_NR_MMU_PAGES failed, errno = %d\n", errno); 324 vcpu_run(vm, VCPU_ID); 325 326 run = vcpu_state(vm, VCPU_ID); 327 TEST_ASSERT(run->exit_reason == KVM_EXIT_INTERNAL_ERROR, 328 "Unexpected exit_reason = %u\n", run->exit_reason); 329 330 kvm_vm_free(vm); 331 } 332 #endif /* __x86_64__ */ 333 334 /* 335 * Test it can be added memory slots up to KVM_CAP_NR_MEMSLOTS, then any 336 * tentative to add further slots should fail. 337 */ 338 static void test_add_max_memory_regions(void) 339 { 340 int ret; 341 struct kvm_vm *vm; 342 uint32_t max_mem_slots; 343 uint32_t slot; 344 uint64_t guest_addr = 0x0; 345 uint64_t mem_reg_npages; 346 void *mem; 347 348 max_mem_slots = kvm_check_cap(KVM_CAP_NR_MEMSLOTS); 349 TEST_ASSERT(max_mem_slots > 0, 350 "KVM_CAP_NR_MEMSLOTS should be greater than 0"); 351 pr_info("Allowed number of memory slots: %i\n", max_mem_slots); 352 353 vm = vm_create(VM_MODE_DEFAULT, 0, O_RDWR); 354 355 mem_reg_npages = vm_calc_num_guest_pages(VM_MODE_DEFAULT, MEM_REGION_SIZE); 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 for (slot = 0; slot < max_mem_slots; slot++) { 361 vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS, 362 guest_addr, slot, mem_reg_npages, 363 0); 364 guest_addr += MEM_REGION_SIZE; 365 } 366 367 /* Check it cannot be added memory slots beyond the limit */ 368 mem = mmap(NULL, MEM_REGION_SIZE, PROT_READ | PROT_WRITE, 369 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); 370 TEST_ASSERT(mem != MAP_FAILED, "Failed to mmap() host"); 371 372 ret = ioctl(vm_get_fd(vm), KVM_SET_USER_MEMORY_REGION, 373 &(struct kvm_userspace_memory_region) {slot, 0, guest_addr, 374 MEM_REGION_SIZE, (uint64_t) mem}); 375 TEST_ASSERT(ret == -1 && errno == EINVAL, 376 "Adding one more memory slot should fail with EINVAL"); 377 378 munmap(mem, MEM_REGION_SIZE); 379 kvm_vm_free(vm); 380 } 381 382 int main(int argc, char *argv[]) 383 { 384 #ifdef __x86_64__ 385 int i, loops; 386 #endif 387 388 /* Tell stdout not to buffer its content */ 389 setbuf(stdout, NULL); 390 391 #ifdef __x86_64__ 392 /* 393 * FIXME: the zero-memslot test fails on aarch64 and s390x because 394 * KVM_RUN fails with ENOEXEC or EFAULT. 395 */ 396 test_zero_memory_regions(); 397 #endif 398 399 test_add_max_memory_regions(); 400 401 #ifdef __x86_64__ 402 if (argc > 1) 403 loops = atoi(argv[1]); 404 else 405 loops = 10; 406 407 pr_info("Testing MOVE of in-use region, %d loops\n", loops); 408 for (i = 0; i < loops; i++) 409 test_move_memory_region(); 410 411 pr_info("Testing DELETE of in-use region, %d loops\n", loops); 412 for (i = 0; i < loops; i++) 413 test_delete_memory_region(); 414 #endif 415 416 return 0; 417 } 418