1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Test for x86 KVM_CAP_SYNC_REGS 4 * 5 * Copyright (C) 2018, Google LLC. 6 * 7 * Verifies expected behavior of x86 KVM_CAP_SYNC_REGS functionality, 8 * including requesting an invalid register set, updates to/from values 9 * in kvm_run.s.regs when kvm_valid_regs and kvm_dirty_regs are toggled. 10 */ 11 12 #define _GNU_SOURCE /* for program_invocation_short_name */ 13 #include <fcntl.h> 14 #include <stdio.h> 15 #include <stdlib.h> 16 #include <string.h> 17 #include <sys/ioctl.h> 18 #include <pthread.h> 19 20 #include "kvm_test_harness.h" 21 #include "test_util.h" 22 #include "kvm_util.h" 23 #include "processor.h" 24 25 #define UCALL_PIO_PORT ((uint16_t)0x1000) 26 27 struct ucall uc_none = { 28 .cmd = UCALL_NONE, 29 }; 30 31 /* 32 * ucall is embedded here to protect against compiler reshuffling registers 33 * before calling a function. In this test we only need to get KVM_EXIT_IO 34 * vmexit and preserve RBX, no additional information is needed. 35 */ 36 void guest_code(void) 37 { 38 asm volatile("1: in %[port], %%al\n" 39 "add $0x1, %%rbx\n" 40 "jmp 1b" 41 : : [port] "d" (UCALL_PIO_PORT), "D" (&uc_none) 42 : "rax", "rbx"); 43 } 44 45 KVM_ONE_VCPU_TEST_SUITE(sync_regs_test); 46 47 static void compare_regs(struct kvm_regs *left, struct kvm_regs *right) 48 { 49 #define REG_COMPARE(reg) \ 50 TEST_ASSERT(left->reg == right->reg, \ 51 "Register " #reg \ 52 " values did not match: 0x%llx, 0x%llx", \ 53 left->reg, right->reg) 54 REG_COMPARE(rax); 55 REG_COMPARE(rbx); 56 REG_COMPARE(rcx); 57 REG_COMPARE(rdx); 58 REG_COMPARE(rsi); 59 REG_COMPARE(rdi); 60 REG_COMPARE(rsp); 61 REG_COMPARE(rbp); 62 REG_COMPARE(r8); 63 REG_COMPARE(r9); 64 REG_COMPARE(r10); 65 REG_COMPARE(r11); 66 REG_COMPARE(r12); 67 REG_COMPARE(r13); 68 REG_COMPARE(r14); 69 REG_COMPARE(r15); 70 REG_COMPARE(rip); 71 REG_COMPARE(rflags); 72 #undef REG_COMPARE 73 } 74 75 static void compare_sregs(struct kvm_sregs *left, struct kvm_sregs *right) 76 { 77 } 78 79 static void compare_vcpu_events(struct kvm_vcpu_events *left, 80 struct kvm_vcpu_events *right) 81 { 82 } 83 84 #define TEST_SYNC_FIELDS (KVM_SYNC_X86_REGS|KVM_SYNC_X86_SREGS|KVM_SYNC_X86_EVENTS) 85 #define INVALID_SYNC_FIELD 0x80000000 86 87 /* 88 * Set an exception as pending *and* injected while KVM is processing events. 89 * KVM is supposed to ignore/drop pending exceptions if userspace is also 90 * requesting that an exception be injected. 91 */ 92 static void *race_events_inj_pen(void *arg) 93 { 94 struct kvm_run *run = (struct kvm_run *)arg; 95 struct kvm_vcpu_events *events = &run->s.regs.events; 96 97 WRITE_ONCE(events->exception.nr, UD_VECTOR); 98 99 for (;;) { 100 WRITE_ONCE(run->kvm_dirty_regs, KVM_SYNC_X86_EVENTS); 101 WRITE_ONCE(events->flags, 0); 102 WRITE_ONCE(events->exception.injected, 1); 103 WRITE_ONCE(events->exception.pending, 1); 104 105 pthread_testcancel(); 106 } 107 108 return NULL; 109 } 110 111 /* 112 * Set an invalid exception vector while KVM is processing events. KVM is 113 * supposed to reject any vector >= 32, as well as NMIs (vector 2). 114 */ 115 static void *race_events_exc(void *arg) 116 { 117 struct kvm_run *run = (struct kvm_run *)arg; 118 struct kvm_vcpu_events *events = &run->s.regs.events; 119 120 for (;;) { 121 WRITE_ONCE(run->kvm_dirty_regs, KVM_SYNC_X86_EVENTS); 122 WRITE_ONCE(events->flags, 0); 123 WRITE_ONCE(events->exception.nr, UD_VECTOR); 124 WRITE_ONCE(events->exception.pending, 1); 125 WRITE_ONCE(events->exception.nr, 255); 126 127 pthread_testcancel(); 128 } 129 130 return NULL; 131 } 132 133 /* 134 * Toggle CR4.PAE while KVM is processing SREGS, EFER.LME=1 with CR4.PAE=0 is 135 * illegal, and KVM's MMU heavily relies on vCPU state being valid. 136 */ 137 static noinline void *race_sregs_cr4(void *arg) 138 { 139 struct kvm_run *run = (struct kvm_run *)arg; 140 __u64 *cr4 = &run->s.regs.sregs.cr4; 141 __u64 pae_enabled = *cr4; 142 __u64 pae_disabled = *cr4 & ~X86_CR4_PAE; 143 144 for (;;) { 145 WRITE_ONCE(run->kvm_dirty_regs, KVM_SYNC_X86_SREGS); 146 WRITE_ONCE(*cr4, pae_enabled); 147 asm volatile(".rept 512\n\t" 148 "nop\n\t" 149 ".endr"); 150 WRITE_ONCE(*cr4, pae_disabled); 151 152 pthread_testcancel(); 153 } 154 155 return NULL; 156 } 157 158 static void race_sync_regs(struct kvm_vcpu *vcpu, void *racer) 159 { 160 const time_t TIMEOUT = 2; /* seconds, roughly */ 161 struct kvm_x86_state *state; 162 struct kvm_translation tr; 163 struct kvm_run *run; 164 pthread_t thread; 165 time_t t; 166 167 run = vcpu->run; 168 169 run->kvm_valid_regs = KVM_SYNC_X86_SREGS; 170 vcpu_run(vcpu); 171 run->kvm_valid_regs = 0; 172 173 /* Save state *before* spawning the thread that mucks with vCPU state. */ 174 state = vcpu_save_state(vcpu); 175 176 /* 177 * Selftests run 64-bit guests by default, both EFER.LME and CR4.PAE 178 * should already be set in guest state. 179 */ 180 TEST_ASSERT((run->s.regs.sregs.cr4 & X86_CR4_PAE) && 181 (run->s.regs.sregs.efer & EFER_LME), 182 "vCPU should be in long mode, CR4.PAE=%d, EFER.LME=%d", 183 !!(run->s.regs.sregs.cr4 & X86_CR4_PAE), 184 !!(run->s.regs.sregs.efer & EFER_LME)); 185 186 TEST_ASSERT_EQ(pthread_create(&thread, NULL, racer, (void *)run), 0); 187 188 for (t = time(NULL) + TIMEOUT; time(NULL) < t;) { 189 /* 190 * Reload known good state if the vCPU triple faults, e.g. due 191 * to the unhandled #GPs being injected. VMX preserves state 192 * on shutdown, but SVM synthesizes an INIT as the VMCB state 193 * is architecturally undefined on triple fault. 194 */ 195 if (!__vcpu_run(vcpu) && run->exit_reason == KVM_EXIT_SHUTDOWN) 196 vcpu_load_state(vcpu, state); 197 198 if (racer == race_sregs_cr4) { 199 tr = (struct kvm_translation) { .linear_address = 0 }; 200 __vcpu_ioctl(vcpu, KVM_TRANSLATE, &tr); 201 } 202 } 203 204 TEST_ASSERT_EQ(pthread_cancel(thread), 0); 205 TEST_ASSERT_EQ(pthread_join(thread, NULL), 0); 206 207 kvm_x86_state_cleanup(state); 208 } 209 210 KVM_ONE_VCPU_TEST(sync_regs_test, read_invalid, guest_code) 211 { 212 struct kvm_run *run = vcpu->run; 213 int rv; 214 215 /* Request reading invalid register set from VCPU. */ 216 run->kvm_valid_regs = INVALID_SYNC_FIELD; 217 rv = _vcpu_run(vcpu); 218 TEST_ASSERT(rv < 0 && errno == EINVAL, 219 "Invalid kvm_valid_regs did not cause expected KVM_RUN error: %d", 220 rv); 221 run->kvm_valid_regs = 0; 222 223 run->kvm_valid_regs = INVALID_SYNC_FIELD | TEST_SYNC_FIELDS; 224 rv = _vcpu_run(vcpu); 225 TEST_ASSERT(rv < 0 && errno == EINVAL, 226 "Invalid kvm_valid_regs did not cause expected KVM_RUN error: %d", 227 rv); 228 run->kvm_valid_regs = 0; 229 } 230 231 KVM_ONE_VCPU_TEST(sync_regs_test, set_invalid, guest_code) 232 { 233 struct kvm_run *run = vcpu->run; 234 int rv; 235 236 /* Request setting invalid register set into VCPU. */ 237 run->kvm_dirty_regs = INVALID_SYNC_FIELD; 238 rv = _vcpu_run(vcpu); 239 TEST_ASSERT(rv < 0 && errno == EINVAL, 240 "Invalid kvm_dirty_regs did not cause expected KVM_RUN error: %d", 241 rv); 242 run->kvm_dirty_regs = 0; 243 244 run->kvm_dirty_regs = INVALID_SYNC_FIELD | TEST_SYNC_FIELDS; 245 rv = _vcpu_run(vcpu); 246 TEST_ASSERT(rv < 0 && errno == EINVAL, 247 "Invalid kvm_dirty_regs did not cause expected KVM_RUN error: %d", 248 rv); 249 run->kvm_dirty_regs = 0; 250 } 251 252 KVM_ONE_VCPU_TEST(sync_regs_test, req_and_verify_all_valid, guest_code) 253 { 254 struct kvm_run *run = vcpu->run; 255 struct kvm_vcpu_events events; 256 struct kvm_sregs sregs; 257 struct kvm_regs regs; 258 259 /* Request and verify all valid register sets. */ 260 /* TODO: BUILD TIME CHECK: TEST_ASSERT(KVM_SYNC_X86_NUM_FIELDS != 3); */ 261 run->kvm_valid_regs = TEST_SYNC_FIELDS; 262 vcpu_run(vcpu); 263 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO); 264 265 vcpu_regs_get(vcpu, ®s); 266 compare_regs(®s, &run->s.regs.regs); 267 268 vcpu_sregs_get(vcpu, &sregs); 269 compare_sregs(&sregs, &run->s.regs.sregs); 270 271 vcpu_events_get(vcpu, &events); 272 compare_vcpu_events(&events, &run->s.regs.events); 273 } 274 275 KVM_ONE_VCPU_TEST(sync_regs_test, set_and_verify_various, guest_code) 276 { 277 struct kvm_run *run = vcpu->run; 278 struct kvm_vcpu_events events; 279 struct kvm_sregs sregs; 280 struct kvm_regs regs; 281 282 /* Run once to get register set */ 283 run->kvm_valid_regs = TEST_SYNC_FIELDS; 284 vcpu_run(vcpu); 285 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO); 286 287 /* Set and verify various register values. */ 288 run->s.regs.regs.rbx = 0xBAD1DEA; 289 run->s.regs.sregs.apic_base = 1 << 11; 290 /* TODO run->s.regs.events.XYZ = ABC; */ 291 292 run->kvm_valid_regs = TEST_SYNC_FIELDS; 293 run->kvm_dirty_regs = KVM_SYNC_X86_REGS | KVM_SYNC_X86_SREGS; 294 vcpu_run(vcpu); 295 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO); 296 TEST_ASSERT(run->s.regs.regs.rbx == 0xBAD1DEA + 1, 297 "rbx sync regs value incorrect 0x%llx.", 298 run->s.regs.regs.rbx); 299 TEST_ASSERT(run->s.regs.sregs.apic_base == 1 << 11, 300 "apic_base sync regs value incorrect 0x%llx.", 301 run->s.regs.sregs.apic_base); 302 303 vcpu_regs_get(vcpu, ®s); 304 compare_regs(®s, &run->s.regs.regs); 305 306 vcpu_sregs_get(vcpu, &sregs); 307 compare_sregs(&sregs, &run->s.regs.sregs); 308 309 vcpu_events_get(vcpu, &events); 310 compare_vcpu_events(&events, &run->s.regs.events); 311 } 312 313 KVM_ONE_VCPU_TEST(sync_regs_test, clear_kvm_dirty_regs_bits, guest_code) 314 { 315 struct kvm_run *run = vcpu->run; 316 317 /* Clear kvm_dirty_regs bits, verify new s.regs values are 318 * overwritten with existing guest values. 319 */ 320 run->kvm_valid_regs = TEST_SYNC_FIELDS; 321 run->kvm_dirty_regs = 0; 322 run->s.regs.regs.rbx = 0xDEADBEEF; 323 vcpu_run(vcpu); 324 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO); 325 TEST_ASSERT(run->s.regs.regs.rbx != 0xDEADBEEF, 326 "rbx sync regs value incorrect 0x%llx.", 327 run->s.regs.regs.rbx); 328 } 329 330 KVM_ONE_VCPU_TEST(sync_regs_test, clear_kvm_valid_and_dirty_regs, guest_code) 331 { 332 struct kvm_run *run = vcpu->run; 333 struct kvm_regs regs; 334 335 /* Run once to get register set */ 336 run->kvm_valid_regs = TEST_SYNC_FIELDS; 337 vcpu_run(vcpu); 338 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO); 339 340 /* Clear kvm_valid_regs bits and kvm_dirty_bits. 341 * Verify s.regs values are not overwritten with existing guest values 342 * and that guest values are not overwritten with kvm_sync_regs values. 343 */ 344 run->kvm_valid_regs = 0; 345 run->kvm_dirty_regs = 0; 346 run->s.regs.regs.rbx = 0xAAAA; 347 vcpu_regs_get(vcpu, ®s); 348 regs.rbx = 0xBAC0; 349 vcpu_regs_set(vcpu, ®s); 350 vcpu_run(vcpu); 351 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO); 352 TEST_ASSERT(run->s.regs.regs.rbx == 0xAAAA, 353 "rbx sync regs value incorrect 0x%llx.", 354 run->s.regs.regs.rbx); 355 vcpu_regs_get(vcpu, ®s); 356 TEST_ASSERT(regs.rbx == 0xBAC0 + 1, 357 "rbx guest value incorrect 0x%llx.", 358 regs.rbx); 359 } 360 361 KVM_ONE_VCPU_TEST(sync_regs_test, clear_kvm_valid_regs_bits, guest_code) 362 { 363 struct kvm_run *run = vcpu->run; 364 struct kvm_regs regs; 365 366 /* Run once to get register set */ 367 run->kvm_valid_regs = TEST_SYNC_FIELDS; 368 vcpu_run(vcpu); 369 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO); 370 371 /* Clear kvm_valid_regs bits. Verify s.regs values are not overwritten 372 * with existing guest values but that guest values are overwritten 373 * with kvm_sync_regs values. 374 */ 375 run->kvm_valid_regs = 0; 376 run->kvm_dirty_regs = TEST_SYNC_FIELDS; 377 run->s.regs.regs.rbx = 0xBBBB; 378 vcpu_run(vcpu); 379 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO); 380 TEST_ASSERT(run->s.regs.regs.rbx == 0xBBBB, 381 "rbx sync regs value incorrect 0x%llx.", 382 run->s.regs.regs.rbx); 383 vcpu_regs_get(vcpu, ®s); 384 TEST_ASSERT(regs.rbx == 0xBBBB + 1, 385 "rbx guest value incorrect 0x%llx.", 386 regs.rbx); 387 } 388 389 KVM_ONE_VCPU_TEST(sync_regs_test, race_cr4, guest_code) 390 { 391 race_sync_regs(vcpu, race_sregs_cr4); 392 } 393 394 KVM_ONE_VCPU_TEST(sync_regs_test, race_exc, guest_code) 395 { 396 race_sync_regs(vcpu, race_events_exc); 397 } 398 399 KVM_ONE_VCPU_TEST(sync_regs_test, race_inj_pen, guest_code) 400 { 401 race_sync_regs(vcpu, race_events_inj_pen); 402 } 403 404 int main(int argc, char *argv[]) 405 { 406 int cap; 407 408 cap = kvm_check_cap(KVM_CAP_SYNC_REGS); 409 TEST_REQUIRE((cap & TEST_SYNC_FIELDS) == TEST_SYNC_FIELDS); 410 TEST_REQUIRE(!(cap & INVALID_SYNC_FIELD)); 411 412 return test_harness_run(argc, argv); 413 } 414