1 /* 2 * Copyright (C) 1991, 1992 Linus Torvalds 3 * Copyright (C) 2000, 2001, 2002 Andi Kleen, SuSE Labs 4 * 5 * Pentium III FXSR, SSE support 6 * Gareth Hughes <gareth@valinux.com>, May 2000 7 */ 8 9 /* 10 * Handle hardware traps and faults. 11 */ 12 13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 14 15 #include <linux/context_tracking.h> 16 #include <linux/interrupt.h> 17 #include <linux/kallsyms.h> 18 #include <linux/kmsan.h> 19 #include <linux/spinlock.h> 20 #include <linux/kprobes.h> 21 #include <linux/uaccess.h> 22 #include <linux/kdebug.h> 23 #include <linux/kgdb.h> 24 #include <linux/kernel.h> 25 #include <linux/export.h> 26 #include <linux/ptrace.h> 27 #include <linux/uprobes.h> 28 #include <linux/string.h> 29 #include <linux/delay.h> 30 #include <linux/errno.h> 31 #include <linux/kexec.h> 32 #include <linux/sched.h> 33 #include <linux/sched/task_stack.h> 34 #include <linux/timer.h> 35 #include <linux/init.h> 36 #include <linux/bug.h> 37 #include <linux/nmi.h> 38 #include <linux/mm.h> 39 #include <linux/smp.h> 40 #include <linux/io.h> 41 #include <linux/hardirq.h> 42 #include <linux/atomic.h> 43 #include <linux/iommu.h> 44 45 #include <asm/stacktrace.h> 46 #include <asm/processor.h> 47 #include <asm/debugreg.h> 48 #include <asm/realmode.h> 49 #include <asm/text-patching.h> 50 #include <asm/ftrace.h> 51 #include <asm/traps.h> 52 #include <asm/desc.h> 53 #include <asm/fpu/api.h> 54 #include <asm/cpu.h> 55 #include <asm/cpu_entry_area.h> 56 #include <asm/mce.h> 57 #include <asm/fixmap.h> 58 #include <asm/mach_traps.h> 59 #include <asm/alternative.h> 60 #include <asm/fpu/xstate.h> 61 #include <asm/vm86.h> 62 #include <asm/umip.h> 63 #include <asm/insn.h> 64 #include <asm/insn-eval.h> 65 #include <asm/vdso.h> 66 #include <asm/tdx.h> 67 #include <asm/cfi.h> 68 69 #ifdef CONFIG_X86_64 70 #include <asm/x86_init.h> 71 #else 72 #include <asm/processor-flags.h> 73 #include <asm/setup.h> 74 #endif 75 76 #include <asm/proto.h> 77 78 DECLARE_BITMAP(system_vectors, NR_VECTORS); 79 80 static inline void cond_local_irq_enable(struct pt_regs *regs) 81 { 82 if (regs->flags & X86_EFLAGS_IF) 83 local_irq_enable(); 84 } 85 86 static inline void cond_local_irq_disable(struct pt_regs *regs) 87 { 88 if (regs->flags & X86_EFLAGS_IF) 89 local_irq_disable(); 90 } 91 92 __always_inline int is_valid_bugaddr(unsigned long addr) 93 { 94 if (addr < TASK_SIZE_MAX) 95 return 0; 96 97 /* 98 * We got #UD, if the text isn't readable we'd have gotten 99 * a different exception. 100 */ 101 return *(unsigned short *)addr == INSN_UD2; 102 } 103 104 static nokprobe_inline int 105 do_trap_no_signal(struct task_struct *tsk, int trapnr, const char *str, 106 struct pt_regs *regs, long error_code) 107 { 108 if (v8086_mode(regs)) { 109 /* 110 * Traps 0, 1, 3, 4, and 5 should be forwarded to vm86. 111 * On nmi (interrupt 2), do_trap should not be called. 112 */ 113 if (trapnr < X86_TRAP_UD) { 114 if (!handle_vm86_trap((struct kernel_vm86_regs *) regs, 115 error_code, trapnr)) 116 return 0; 117 } 118 } else if (!user_mode(regs)) { 119 if (fixup_exception(regs, trapnr, error_code, 0)) 120 return 0; 121 122 tsk->thread.error_code = error_code; 123 tsk->thread.trap_nr = trapnr; 124 die(str, regs, error_code); 125 } else { 126 if (fixup_vdso_exception(regs, trapnr, error_code, 0)) 127 return 0; 128 } 129 130 /* 131 * We want error_code and trap_nr set for userspace faults and 132 * kernelspace faults which result in die(), but not 133 * kernelspace faults which are fixed up. die() gives the 134 * process no chance to handle the signal and notice the 135 * kernel fault information, so that won't result in polluting 136 * the information about previously queued, but not yet 137 * delivered, faults. See also exc_general_protection below. 138 */ 139 tsk->thread.error_code = error_code; 140 tsk->thread.trap_nr = trapnr; 141 142 return -1; 143 } 144 145 static void show_signal(struct task_struct *tsk, int signr, 146 const char *type, const char *desc, 147 struct pt_regs *regs, long error_code) 148 { 149 if (show_unhandled_signals && unhandled_signal(tsk, signr) && 150 printk_ratelimit()) { 151 pr_info("%s[%d] %s%s ip:%lx sp:%lx error:%lx", 152 tsk->comm, task_pid_nr(tsk), type, desc, 153 regs->ip, regs->sp, error_code); 154 print_vma_addr(KERN_CONT " in ", regs->ip); 155 pr_cont("\n"); 156 } 157 } 158 159 static void 160 do_trap(int trapnr, int signr, char *str, struct pt_regs *regs, 161 long error_code, int sicode, void __user *addr) 162 { 163 struct task_struct *tsk = current; 164 165 if (!do_trap_no_signal(tsk, trapnr, str, regs, error_code)) 166 return; 167 168 show_signal(tsk, signr, "trap ", str, regs, error_code); 169 170 if (!sicode) 171 force_sig(signr); 172 else 173 force_sig_fault(signr, sicode, addr); 174 } 175 NOKPROBE_SYMBOL(do_trap); 176 177 static void do_error_trap(struct pt_regs *regs, long error_code, char *str, 178 unsigned long trapnr, int signr, int sicode, void __user *addr) 179 { 180 RCU_LOCKDEP_WARN(!rcu_is_watching(), "entry code didn't wake RCU"); 181 182 if (notify_die(DIE_TRAP, str, regs, error_code, trapnr, signr) != 183 NOTIFY_STOP) { 184 cond_local_irq_enable(regs); 185 do_trap(trapnr, signr, str, regs, error_code, sicode, addr); 186 cond_local_irq_disable(regs); 187 } 188 } 189 190 /* 191 * Posix requires to provide the address of the faulting instruction for 192 * SIGILL (#UD) and SIGFPE (#DE) in the si_addr member of siginfo_t. 193 * 194 * This address is usually regs->ip, but when an uprobe moved the code out 195 * of line then regs->ip points to the XOL code which would confuse 196 * anything which analyzes the fault address vs. the unmodified binary. If 197 * a trap happened in XOL code then uprobe maps regs->ip back to the 198 * original instruction address. 199 */ 200 static __always_inline void __user *error_get_trap_addr(struct pt_regs *regs) 201 { 202 return (void __user *)uprobe_get_trap_addr(regs); 203 } 204 205 DEFINE_IDTENTRY(exc_divide_error) 206 { 207 do_error_trap(regs, 0, "divide error", X86_TRAP_DE, SIGFPE, 208 FPE_INTDIV, error_get_trap_addr(regs)); 209 } 210 211 DEFINE_IDTENTRY(exc_overflow) 212 { 213 do_error_trap(regs, 0, "overflow", X86_TRAP_OF, SIGSEGV, 0, NULL); 214 } 215 216 #ifdef CONFIG_X86_KERNEL_IBT 217 218 static __ro_after_init bool ibt_fatal = true; 219 220 extern void ibt_selftest_ip(void); /* code label defined in asm below */ 221 222 enum cp_error_code { 223 CP_EC = (1 << 15) - 1, 224 225 CP_RET = 1, 226 CP_IRET = 2, 227 CP_ENDBR = 3, 228 CP_RSTRORSSP = 4, 229 CP_SETSSBSY = 5, 230 231 CP_ENCL = 1 << 15, 232 }; 233 234 DEFINE_IDTENTRY_ERRORCODE(exc_control_protection) 235 { 236 if (!cpu_feature_enabled(X86_FEATURE_IBT)) { 237 pr_err("Unexpected #CP\n"); 238 BUG(); 239 } 240 241 if (WARN_ON_ONCE(user_mode(regs) || (error_code & CP_EC) != CP_ENDBR)) 242 return; 243 244 if (unlikely(regs->ip == (unsigned long)&ibt_selftest_ip)) { 245 regs->ax = 0; 246 return; 247 } 248 249 pr_err("Missing ENDBR: %pS\n", (void *)instruction_pointer(regs)); 250 if (!ibt_fatal) { 251 printk(KERN_DEFAULT CUT_HERE); 252 __warn(__FILE__, __LINE__, (void *)regs->ip, TAINT_WARN, regs, NULL); 253 return; 254 } 255 BUG(); 256 } 257 258 /* Must be noinline to ensure uniqueness of ibt_selftest_ip. */ 259 noinline bool ibt_selftest(void) 260 { 261 unsigned long ret; 262 263 asm (" lea ibt_selftest_ip(%%rip), %%rax\n\t" 264 ANNOTATE_RETPOLINE_SAFE 265 " jmp *%%rax\n\t" 266 "ibt_selftest_ip:\n\t" 267 UNWIND_HINT_FUNC 268 ANNOTATE_NOENDBR 269 " nop\n\t" 270 271 : "=a" (ret) : : "memory"); 272 273 return !ret; 274 } 275 276 static int __init ibt_setup(char *str) 277 { 278 if (!strcmp(str, "off")) 279 setup_clear_cpu_cap(X86_FEATURE_IBT); 280 281 if (!strcmp(str, "warn")) 282 ibt_fatal = false; 283 284 return 1; 285 } 286 287 __setup("ibt=", ibt_setup); 288 289 #endif /* CONFIG_X86_KERNEL_IBT */ 290 291 #ifdef CONFIG_X86_F00F_BUG 292 void handle_invalid_op(struct pt_regs *regs) 293 #else 294 static inline void handle_invalid_op(struct pt_regs *regs) 295 #endif 296 { 297 do_error_trap(regs, 0, "invalid opcode", X86_TRAP_UD, SIGILL, 298 ILL_ILLOPN, error_get_trap_addr(regs)); 299 } 300 301 static noinstr bool handle_bug(struct pt_regs *regs) 302 { 303 bool handled = false; 304 305 /* 306 * Normally @regs are unpoisoned by irqentry_enter(), but handle_bug() 307 * is a rare case that uses @regs without passing them to 308 * irqentry_enter(). 309 */ 310 kmsan_unpoison_entry_regs(regs); 311 if (!is_valid_bugaddr(regs->ip)) 312 return handled; 313 314 /* 315 * All lies, just get the WARN/BUG out. 316 */ 317 instrumentation_begin(); 318 /* 319 * Since we're emulating a CALL with exceptions, restore the interrupt 320 * state to what it was at the exception site. 321 */ 322 if (regs->flags & X86_EFLAGS_IF) 323 raw_local_irq_enable(); 324 if (report_bug(regs->ip, regs) == BUG_TRAP_TYPE_WARN || 325 handle_cfi_failure(regs) == BUG_TRAP_TYPE_WARN) { 326 regs->ip += LEN_UD2; 327 handled = true; 328 } 329 if (regs->flags & X86_EFLAGS_IF) 330 raw_local_irq_disable(); 331 instrumentation_end(); 332 333 return handled; 334 } 335 336 DEFINE_IDTENTRY_RAW(exc_invalid_op) 337 { 338 irqentry_state_t state; 339 340 /* 341 * We use UD2 as a short encoding for 'CALL __WARN', as such 342 * handle it before exception entry to avoid recursive WARN 343 * in case exception entry is the one triggering WARNs. 344 */ 345 if (!user_mode(regs) && handle_bug(regs)) 346 return; 347 348 state = irqentry_enter(regs); 349 instrumentation_begin(); 350 handle_invalid_op(regs); 351 instrumentation_end(); 352 irqentry_exit(regs, state); 353 } 354 355 DEFINE_IDTENTRY(exc_coproc_segment_overrun) 356 { 357 do_error_trap(regs, 0, "coprocessor segment overrun", 358 X86_TRAP_OLD_MF, SIGFPE, 0, NULL); 359 } 360 361 DEFINE_IDTENTRY_ERRORCODE(exc_invalid_tss) 362 { 363 do_error_trap(regs, error_code, "invalid TSS", X86_TRAP_TS, SIGSEGV, 364 0, NULL); 365 } 366 367 DEFINE_IDTENTRY_ERRORCODE(exc_segment_not_present) 368 { 369 do_error_trap(regs, error_code, "segment not present", X86_TRAP_NP, 370 SIGBUS, 0, NULL); 371 } 372 373 DEFINE_IDTENTRY_ERRORCODE(exc_stack_segment) 374 { 375 do_error_trap(regs, error_code, "stack segment", X86_TRAP_SS, SIGBUS, 376 0, NULL); 377 } 378 379 DEFINE_IDTENTRY_ERRORCODE(exc_alignment_check) 380 { 381 char *str = "alignment check"; 382 383 if (notify_die(DIE_TRAP, str, regs, error_code, X86_TRAP_AC, SIGBUS) == NOTIFY_STOP) 384 return; 385 386 if (!user_mode(regs)) 387 die("Split lock detected\n", regs, error_code); 388 389 local_irq_enable(); 390 391 if (handle_user_split_lock(regs, error_code)) 392 goto out; 393 394 do_trap(X86_TRAP_AC, SIGBUS, "alignment check", regs, 395 error_code, BUS_ADRALN, NULL); 396 397 out: 398 local_irq_disable(); 399 } 400 401 #ifdef CONFIG_VMAP_STACK 402 __visible void __noreturn handle_stack_overflow(struct pt_regs *regs, 403 unsigned long fault_address, 404 struct stack_info *info) 405 { 406 const char *name = stack_type_name(info->type); 407 408 printk(KERN_EMERG "BUG: %s stack guard page was hit at %p (stack is %p..%p)\n", 409 name, (void *)fault_address, info->begin, info->end); 410 411 die("stack guard page", regs, 0); 412 413 /* Be absolutely certain we don't return. */ 414 panic("%s stack guard hit", name); 415 } 416 #endif 417 418 /* 419 * Runs on an IST stack for x86_64 and on a special task stack for x86_32. 420 * 421 * On x86_64, this is more or less a normal kernel entry. Notwithstanding the 422 * SDM's warnings about double faults being unrecoverable, returning works as 423 * expected. Presumably what the SDM actually means is that the CPU may get 424 * the register state wrong on entry, so returning could be a bad idea. 425 * 426 * Various CPU engineers have promised that double faults due to an IRET fault 427 * while the stack is read-only are, in fact, recoverable. 428 * 429 * On x86_32, this is entered through a task gate, and regs are synthesized 430 * from the TSS. Returning is, in principle, okay, but changes to regs will 431 * be lost. If, for some reason, we need to return to a context with modified 432 * regs, the shim code could be adjusted to synchronize the registers. 433 * 434 * The 32bit #DF shim provides CR2 already as an argument. On 64bit it needs 435 * to be read before doing anything else. 436 */ 437 DEFINE_IDTENTRY_DF(exc_double_fault) 438 { 439 static const char str[] = "double fault"; 440 struct task_struct *tsk = current; 441 442 #ifdef CONFIG_VMAP_STACK 443 unsigned long address = read_cr2(); 444 struct stack_info info; 445 #endif 446 447 #ifdef CONFIG_X86_ESPFIX64 448 extern unsigned char native_irq_return_iret[]; 449 450 /* 451 * If IRET takes a non-IST fault on the espfix64 stack, then we 452 * end up promoting it to a doublefault. In that case, take 453 * advantage of the fact that we're not using the normal (TSS.sp0) 454 * stack right now. We can write a fake #GP(0) frame at TSS.sp0 455 * and then modify our own IRET frame so that, when we return, 456 * we land directly at the #GP(0) vector with the stack already 457 * set up according to its expectations. 458 * 459 * The net result is that our #GP handler will think that we 460 * entered from usermode with the bad user context. 461 * 462 * No need for nmi_enter() here because we don't use RCU. 463 */ 464 if (((long)regs->sp >> P4D_SHIFT) == ESPFIX_PGD_ENTRY && 465 regs->cs == __KERNEL_CS && 466 regs->ip == (unsigned long)native_irq_return_iret) 467 { 468 struct pt_regs *gpregs = (struct pt_regs *)this_cpu_read(cpu_tss_rw.x86_tss.sp0) - 1; 469 unsigned long *p = (unsigned long *)regs->sp; 470 471 /* 472 * regs->sp points to the failing IRET frame on the 473 * ESPFIX64 stack. Copy it to the entry stack. This fills 474 * in gpregs->ss through gpregs->ip. 475 * 476 */ 477 gpregs->ip = p[0]; 478 gpregs->cs = p[1]; 479 gpregs->flags = p[2]; 480 gpregs->sp = p[3]; 481 gpregs->ss = p[4]; 482 gpregs->orig_ax = 0; /* Missing (lost) #GP error code */ 483 484 /* 485 * Adjust our frame so that we return straight to the #GP 486 * vector with the expected RSP value. This is safe because 487 * we won't enable interrupts or schedule before we invoke 488 * general_protection, so nothing will clobber the stack 489 * frame we just set up. 490 * 491 * We will enter general_protection with kernel GSBASE, 492 * which is what the stub expects, given that the faulting 493 * RIP will be the IRET instruction. 494 */ 495 regs->ip = (unsigned long)asm_exc_general_protection; 496 regs->sp = (unsigned long)&gpregs->orig_ax; 497 498 return; 499 } 500 #endif 501 502 irqentry_nmi_enter(regs); 503 instrumentation_begin(); 504 notify_die(DIE_TRAP, str, regs, error_code, X86_TRAP_DF, SIGSEGV); 505 506 tsk->thread.error_code = error_code; 507 tsk->thread.trap_nr = X86_TRAP_DF; 508 509 #ifdef CONFIG_VMAP_STACK 510 /* 511 * If we overflow the stack into a guard page, the CPU will fail 512 * to deliver #PF and will send #DF instead. Similarly, if we 513 * take any non-IST exception while too close to the bottom of 514 * the stack, the processor will get a page fault while 515 * delivering the exception and will generate a double fault. 516 * 517 * According to the SDM (footnote in 6.15 under "Interrupt 14 - 518 * Page-Fault Exception (#PF): 519 * 520 * Processors update CR2 whenever a page fault is detected. If a 521 * second page fault occurs while an earlier page fault is being 522 * delivered, the faulting linear address of the second fault will 523 * overwrite the contents of CR2 (replacing the previous 524 * address). These updates to CR2 occur even if the page fault 525 * results in a double fault or occurs during the delivery of a 526 * double fault. 527 * 528 * The logic below has a small possibility of incorrectly diagnosing 529 * some errors as stack overflows. For example, if the IDT or GDT 530 * gets corrupted such that #GP delivery fails due to a bad descriptor 531 * causing #GP and we hit this condition while CR2 coincidentally 532 * points to the stack guard page, we'll think we overflowed the 533 * stack. Given that we're going to panic one way or another 534 * if this happens, this isn't necessarily worth fixing. 535 * 536 * If necessary, we could improve the test by only diagnosing 537 * a stack overflow if the saved RSP points within 47 bytes of 538 * the bottom of the stack: if RSP == tsk_stack + 48 and we 539 * take an exception, the stack is already aligned and there 540 * will be enough room SS, RSP, RFLAGS, CS, RIP, and a 541 * possible error code, so a stack overflow would *not* double 542 * fault. With any less space left, exception delivery could 543 * fail, and, as a practical matter, we've overflowed the 544 * stack even if the actual trigger for the double fault was 545 * something else. 546 */ 547 if (get_stack_guard_info((void *)address, &info)) 548 handle_stack_overflow(regs, address, &info); 549 #endif 550 551 pr_emerg("PANIC: double fault, error_code: 0x%lx\n", error_code); 552 die("double fault", regs, error_code); 553 panic("Machine halted."); 554 instrumentation_end(); 555 } 556 557 DEFINE_IDTENTRY(exc_bounds) 558 { 559 if (notify_die(DIE_TRAP, "bounds", regs, 0, 560 X86_TRAP_BR, SIGSEGV) == NOTIFY_STOP) 561 return; 562 cond_local_irq_enable(regs); 563 564 if (!user_mode(regs)) 565 die("bounds", regs, 0); 566 567 do_trap(X86_TRAP_BR, SIGSEGV, "bounds", regs, 0, 0, NULL); 568 569 cond_local_irq_disable(regs); 570 } 571 572 enum kernel_gp_hint { 573 GP_NO_HINT, 574 GP_NON_CANONICAL, 575 GP_CANONICAL 576 }; 577 578 /* 579 * When an uncaught #GP occurs, try to determine the memory address accessed by 580 * the instruction and return that address to the caller. Also, try to figure 581 * out whether any part of the access to that address was non-canonical. 582 */ 583 static enum kernel_gp_hint get_kernel_gp_address(struct pt_regs *regs, 584 unsigned long *addr) 585 { 586 u8 insn_buf[MAX_INSN_SIZE]; 587 struct insn insn; 588 int ret; 589 590 if (copy_from_kernel_nofault(insn_buf, (void *)regs->ip, 591 MAX_INSN_SIZE)) 592 return GP_NO_HINT; 593 594 ret = insn_decode_kernel(&insn, insn_buf); 595 if (ret < 0) 596 return GP_NO_HINT; 597 598 *addr = (unsigned long)insn_get_addr_ref(&insn, regs); 599 if (*addr == -1UL) 600 return GP_NO_HINT; 601 602 #ifdef CONFIG_X86_64 603 /* 604 * Check that: 605 * - the operand is not in the kernel half 606 * - the last byte of the operand is not in the user canonical half 607 */ 608 if (*addr < ~__VIRTUAL_MASK && 609 *addr + insn.opnd_bytes - 1 > __VIRTUAL_MASK) 610 return GP_NON_CANONICAL; 611 #endif 612 613 return GP_CANONICAL; 614 } 615 616 #define GPFSTR "general protection fault" 617 618 static bool fixup_iopl_exception(struct pt_regs *regs) 619 { 620 struct thread_struct *t = ¤t->thread; 621 unsigned char byte; 622 unsigned long ip; 623 624 if (!IS_ENABLED(CONFIG_X86_IOPL_IOPERM) || t->iopl_emul != 3) 625 return false; 626 627 if (insn_get_effective_ip(regs, &ip)) 628 return false; 629 630 if (get_user(byte, (const char __user *)ip)) 631 return false; 632 633 if (byte != 0xfa && byte != 0xfb) 634 return false; 635 636 if (!t->iopl_warn && printk_ratelimit()) { 637 pr_err("%s[%d] attempts to use CLI/STI, pretending it's a NOP, ip:%lx", 638 current->comm, task_pid_nr(current), ip); 639 print_vma_addr(KERN_CONT " in ", ip); 640 pr_cont("\n"); 641 t->iopl_warn = 1; 642 } 643 644 regs->ip += 1; 645 return true; 646 } 647 648 /* 649 * The unprivileged ENQCMD instruction generates #GPs if the 650 * IA32_PASID MSR has not been populated. If possible, populate 651 * the MSR from a PASID previously allocated to the mm. 652 */ 653 static bool try_fixup_enqcmd_gp(void) 654 { 655 #ifdef CONFIG_IOMMU_SVA 656 u32 pasid; 657 658 /* 659 * MSR_IA32_PASID is managed using XSAVE. Directly 660 * writing to the MSR is only possible when fpregs 661 * are valid and the fpstate is not. This is 662 * guaranteed when handling a userspace exception 663 * in *before* interrupts are re-enabled. 664 */ 665 lockdep_assert_irqs_disabled(); 666 667 /* 668 * Hardware without ENQCMD will not generate 669 * #GPs that can be fixed up here. 670 */ 671 if (!cpu_feature_enabled(X86_FEATURE_ENQCMD)) 672 return false; 673 674 /* 675 * If the mm has not been allocated a 676 * PASID, the #GP can not be fixed up. 677 */ 678 if (!mm_valid_pasid(current->mm)) 679 return false; 680 681 pasid = current->mm->pasid; 682 683 /* 684 * Did this thread already have its PASID activated? 685 * If so, the #GP must be from something else. 686 */ 687 if (current->pasid_activated) 688 return false; 689 690 wrmsrl(MSR_IA32_PASID, pasid | MSR_IA32_PASID_VALID); 691 current->pasid_activated = 1; 692 693 return true; 694 #else 695 return false; 696 #endif 697 } 698 699 static bool gp_try_fixup_and_notify(struct pt_regs *regs, int trapnr, 700 unsigned long error_code, const char *str) 701 { 702 if (fixup_exception(regs, trapnr, error_code, 0)) 703 return true; 704 705 current->thread.error_code = error_code; 706 current->thread.trap_nr = trapnr; 707 708 /* 709 * To be potentially processing a kprobe fault and to trust the result 710 * from kprobe_running(), we have to be non-preemptible. 711 */ 712 if (!preemptible() && kprobe_running() && 713 kprobe_fault_handler(regs, trapnr)) 714 return true; 715 716 return notify_die(DIE_GPF, str, regs, error_code, trapnr, SIGSEGV) == NOTIFY_STOP; 717 } 718 719 static void gp_user_force_sig_segv(struct pt_regs *regs, int trapnr, 720 unsigned long error_code, const char *str) 721 { 722 current->thread.error_code = error_code; 723 current->thread.trap_nr = trapnr; 724 show_signal(current, SIGSEGV, "", str, regs, error_code); 725 force_sig(SIGSEGV); 726 } 727 728 DEFINE_IDTENTRY_ERRORCODE(exc_general_protection) 729 { 730 char desc[sizeof(GPFSTR) + 50 + 2*sizeof(unsigned long) + 1] = GPFSTR; 731 enum kernel_gp_hint hint = GP_NO_HINT; 732 unsigned long gp_addr; 733 734 if (user_mode(regs) && try_fixup_enqcmd_gp()) 735 return; 736 737 cond_local_irq_enable(regs); 738 739 if (static_cpu_has(X86_FEATURE_UMIP)) { 740 if (user_mode(regs) && fixup_umip_exception(regs)) 741 goto exit; 742 } 743 744 if (v8086_mode(regs)) { 745 local_irq_enable(); 746 handle_vm86_fault((struct kernel_vm86_regs *) regs, error_code); 747 local_irq_disable(); 748 return; 749 } 750 751 if (user_mode(regs)) { 752 if (fixup_iopl_exception(regs)) 753 goto exit; 754 755 if (fixup_vdso_exception(regs, X86_TRAP_GP, error_code, 0)) 756 goto exit; 757 758 gp_user_force_sig_segv(regs, X86_TRAP_GP, error_code, desc); 759 goto exit; 760 } 761 762 if (gp_try_fixup_and_notify(regs, X86_TRAP_GP, error_code, desc)) 763 goto exit; 764 765 if (error_code) 766 snprintf(desc, sizeof(desc), "segment-related " GPFSTR); 767 else 768 hint = get_kernel_gp_address(regs, &gp_addr); 769 770 if (hint != GP_NO_HINT) 771 snprintf(desc, sizeof(desc), GPFSTR ", %s 0x%lx", 772 (hint == GP_NON_CANONICAL) ? "probably for non-canonical address" 773 : "maybe for address", 774 gp_addr); 775 776 /* 777 * KASAN is interested only in the non-canonical case, clear it 778 * otherwise. 779 */ 780 if (hint != GP_NON_CANONICAL) 781 gp_addr = 0; 782 783 die_addr(desc, regs, error_code, gp_addr); 784 785 exit: 786 cond_local_irq_disable(regs); 787 } 788 789 static bool do_int3(struct pt_regs *regs) 790 { 791 int res; 792 793 #ifdef CONFIG_KGDB_LOW_LEVEL_TRAP 794 if (kgdb_ll_trap(DIE_INT3, "int3", regs, 0, X86_TRAP_BP, 795 SIGTRAP) == NOTIFY_STOP) 796 return true; 797 #endif /* CONFIG_KGDB_LOW_LEVEL_TRAP */ 798 799 #ifdef CONFIG_KPROBES 800 if (kprobe_int3_handler(regs)) 801 return true; 802 #endif 803 res = notify_die(DIE_INT3, "int3", regs, 0, X86_TRAP_BP, SIGTRAP); 804 805 return res == NOTIFY_STOP; 806 } 807 NOKPROBE_SYMBOL(do_int3); 808 809 static void do_int3_user(struct pt_regs *regs) 810 { 811 if (do_int3(regs)) 812 return; 813 814 cond_local_irq_enable(regs); 815 do_trap(X86_TRAP_BP, SIGTRAP, "int3", regs, 0, 0, NULL); 816 cond_local_irq_disable(regs); 817 } 818 819 DEFINE_IDTENTRY_RAW(exc_int3) 820 { 821 /* 822 * poke_int3_handler() is completely self contained code; it does (and 823 * must) *NOT* call out to anything, lest it hits upon yet another 824 * INT3. 825 */ 826 if (poke_int3_handler(regs)) 827 return; 828 829 /* 830 * irqentry_enter_from_user_mode() uses static_branch_{,un}likely() 831 * and therefore can trigger INT3, hence poke_int3_handler() must 832 * be done before. If the entry came from kernel mode, then use 833 * nmi_enter() because the INT3 could have been hit in any context 834 * including NMI. 835 */ 836 if (user_mode(regs)) { 837 irqentry_enter_from_user_mode(regs); 838 instrumentation_begin(); 839 do_int3_user(regs); 840 instrumentation_end(); 841 irqentry_exit_to_user_mode(regs); 842 } else { 843 irqentry_state_t irq_state = irqentry_nmi_enter(regs); 844 845 instrumentation_begin(); 846 if (!do_int3(regs)) 847 die("int3", regs, 0); 848 instrumentation_end(); 849 irqentry_nmi_exit(regs, irq_state); 850 } 851 } 852 853 #ifdef CONFIG_X86_64 854 /* 855 * Help handler running on a per-cpu (IST or entry trampoline) stack 856 * to switch to the normal thread stack if the interrupted code was in 857 * user mode. The actual stack switch is done in entry_64.S 858 */ 859 asmlinkage __visible noinstr struct pt_regs *sync_regs(struct pt_regs *eregs) 860 { 861 struct pt_regs *regs = (struct pt_regs *)this_cpu_read(pcpu_hot.top_of_stack) - 1; 862 if (regs != eregs) 863 *regs = *eregs; 864 return regs; 865 } 866 867 #ifdef CONFIG_AMD_MEM_ENCRYPT 868 asmlinkage __visible noinstr struct pt_regs *vc_switch_off_ist(struct pt_regs *regs) 869 { 870 unsigned long sp, *stack; 871 struct stack_info info; 872 struct pt_regs *regs_ret; 873 874 /* 875 * In the SYSCALL entry path the RSP value comes from user-space - don't 876 * trust it and switch to the current kernel stack 877 */ 878 if (ip_within_syscall_gap(regs)) { 879 sp = this_cpu_read(pcpu_hot.top_of_stack); 880 goto sync; 881 } 882 883 /* 884 * From here on the RSP value is trusted. Now check whether entry 885 * happened from a safe stack. Not safe are the entry or unknown stacks, 886 * use the fall-back stack instead in this case. 887 */ 888 sp = regs->sp; 889 stack = (unsigned long *)sp; 890 891 if (!get_stack_info_noinstr(stack, current, &info) || info.type == STACK_TYPE_ENTRY || 892 info.type > STACK_TYPE_EXCEPTION_LAST) 893 sp = __this_cpu_ist_top_va(VC2); 894 895 sync: 896 /* 897 * Found a safe stack - switch to it as if the entry didn't happen via 898 * IST stack. The code below only copies pt_regs, the real switch happens 899 * in assembly code. 900 */ 901 sp = ALIGN_DOWN(sp, 8) - sizeof(*regs_ret); 902 903 regs_ret = (struct pt_regs *)sp; 904 *regs_ret = *regs; 905 906 return regs_ret; 907 } 908 #endif 909 910 asmlinkage __visible noinstr struct pt_regs *fixup_bad_iret(struct pt_regs *bad_regs) 911 { 912 struct pt_regs tmp, *new_stack; 913 914 /* 915 * This is called from entry_64.S early in handling a fault 916 * caused by a bad iret to user mode. To handle the fault 917 * correctly, we want to move our stack frame to where it would 918 * be had we entered directly on the entry stack (rather than 919 * just below the IRET frame) and we want to pretend that the 920 * exception came from the IRET target. 921 */ 922 new_stack = (struct pt_regs *)__this_cpu_read(cpu_tss_rw.x86_tss.sp0) - 1; 923 924 /* Copy the IRET target to the temporary storage. */ 925 __memcpy(&tmp.ip, (void *)bad_regs->sp, 5*8); 926 927 /* Copy the remainder of the stack from the current stack. */ 928 __memcpy(&tmp, bad_regs, offsetof(struct pt_regs, ip)); 929 930 /* Update the entry stack */ 931 __memcpy(new_stack, &tmp, sizeof(tmp)); 932 933 BUG_ON(!user_mode(new_stack)); 934 return new_stack; 935 } 936 #endif 937 938 static bool is_sysenter_singlestep(struct pt_regs *regs) 939 { 940 /* 941 * We don't try for precision here. If we're anywhere in the region of 942 * code that can be single-stepped in the SYSENTER entry path, then 943 * assume that this is a useless single-step trap due to SYSENTER 944 * being invoked with TF set. (We don't know in advance exactly 945 * which instructions will be hit because BTF could plausibly 946 * be set.) 947 */ 948 #ifdef CONFIG_X86_32 949 return (regs->ip - (unsigned long)__begin_SYSENTER_singlestep_region) < 950 (unsigned long)__end_SYSENTER_singlestep_region - 951 (unsigned long)__begin_SYSENTER_singlestep_region; 952 #elif defined(CONFIG_IA32_EMULATION) 953 return (regs->ip - (unsigned long)entry_SYSENTER_compat) < 954 (unsigned long)__end_entry_SYSENTER_compat - 955 (unsigned long)entry_SYSENTER_compat; 956 #else 957 return false; 958 #endif 959 } 960 961 static __always_inline unsigned long debug_read_clear_dr6(void) 962 { 963 unsigned long dr6; 964 965 /* 966 * The Intel SDM says: 967 * 968 * Certain debug exceptions may clear bits 0-3. The remaining 969 * contents of the DR6 register are never cleared by the 970 * processor. To avoid confusion in identifying debug 971 * exceptions, debug handlers should clear the register before 972 * returning to the interrupted task. 973 * 974 * Keep it simple: clear DR6 immediately. 975 */ 976 get_debugreg(dr6, 6); 977 set_debugreg(DR6_RESERVED, 6); 978 dr6 ^= DR6_RESERVED; /* Flip to positive polarity */ 979 980 return dr6; 981 } 982 983 /* 984 * Our handling of the processor debug registers is non-trivial. 985 * We do not clear them on entry and exit from the kernel. Therefore 986 * it is possible to get a watchpoint trap here from inside the kernel. 987 * However, the code in ./ptrace.c has ensured that the user can 988 * only set watchpoints on userspace addresses. Therefore the in-kernel 989 * watchpoint trap can only occur in code which is reading/writing 990 * from user space. Such code must not hold kernel locks (since it 991 * can equally take a page fault), therefore it is safe to call 992 * force_sig_info even though that claims and releases locks. 993 * 994 * Code in ./signal.c ensures that the debug control register 995 * is restored before we deliver any signal, and therefore that 996 * user code runs with the correct debug control register even though 997 * we clear it here. 998 * 999 * Being careful here means that we don't have to be as careful in a 1000 * lot of more complicated places (task switching can be a bit lazy 1001 * about restoring all the debug state, and ptrace doesn't have to 1002 * find every occurrence of the TF bit that could be saved away even 1003 * by user code) 1004 * 1005 * May run on IST stack. 1006 */ 1007 1008 static bool notify_debug(struct pt_regs *regs, unsigned long *dr6) 1009 { 1010 /* 1011 * Notifiers will clear bits in @dr6 to indicate the event has been 1012 * consumed - hw_breakpoint_handler(), single_stop_cont(). 1013 * 1014 * Notifiers will set bits in @virtual_dr6 to indicate the desire 1015 * for signals - ptrace_triggered(), kgdb_hw_overflow_handler(). 1016 */ 1017 if (notify_die(DIE_DEBUG, "debug", regs, (long)dr6, 0, SIGTRAP) == NOTIFY_STOP) 1018 return true; 1019 1020 return false; 1021 } 1022 1023 static __always_inline void exc_debug_kernel(struct pt_regs *regs, 1024 unsigned long dr6) 1025 { 1026 /* 1027 * Disable breakpoints during exception handling; recursive exceptions 1028 * are exceedingly 'fun'. 1029 * 1030 * Since this function is NOKPROBE, and that also applies to 1031 * HW_BREAKPOINT_X, we can't hit a breakpoint before this (XXX except a 1032 * HW_BREAKPOINT_W on our stack) 1033 * 1034 * Entry text is excluded for HW_BP_X and cpu_entry_area, which 1035 * includes the entry stack is excluded for everything. 1036 */ 1037 unsigned long dr7 = local_db_save(); 1038 irqentry_state_t irq_state = irqentry_nmi_enter(regs); 1039 instrumentation_begin(); 1040 1041 /* 1042 * If something gets miswired and we end up here for a user mode 1043 * #DB, we will malfunction. 1044 */ 1045 WARN_ON_ONCE(user_mode(regs)); 1046 1047 if (test_thread_flag(TIF_BLOCKSTEP)) { 1048 /* 1049 * The SDM says "The processor clears the BTF flag when it 1050 * generates a debug exception." but PTRACE_BLOCKSTEP requested 1051 * it for userspace, but we just took a kernel #DB, so re-set 1052 * BTF. 1053 */ 1054 unsigned long debugctl; 1055 1056 rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl); 1057 debugctl |= DEBUGCTLMSR_BTF; 1058 wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl); 1059 } 1060 1061 /* 1062 * Catch SYSENTER with TF set and clear DR_STEP. If this hit a 1063 * watchpoint at the same time then that will still be handled. 1064 */ 1065 if ((dr6 & DR_STEP) && is_sysenter_singlestep(regs)) 1066 dr6 &= ~DR_STEP; 1067 1068 /* 1069 * The kernel doesn't use INT1 1070 */ 1071 if (!dr6) 1072 goto out; 1073 1074 if (notify_debug(regs, &dr6)) 1075 goto out; 1076 1077 /* 1078 * The kernel doesn't use TF single-step outside of: 1079 * 1080 * - Kprobes, consumed through kprobe_debug_handler() 1081 * - KGDB, consumed through notify_debug() 1082 * 1083 * So if we get here with DR_STEP set, something is wonky. 1084 * 1085 * A known way to trigger this is through QEMU's GDB stub, 1086 * which leaks #DB into the guest and causes IST recursion. 1087 */ 1088 if (WARN_ON_ONCE(dr6 & DR_STEP)) 1089 regs->flags &= ~X86_EFLAGS_TF; 1090 out: 1091 instrumentation_end(); 1092 irqentry_nmi_exit(regs, irq_state); 1093 1094 local_db_restore(dr7); 1095 } 1096 1097 static __always_inline void exc_debug_user(struct pt_regs *regs, 1098 unsigned long dr6) 1099 { 1100 bool icebp; 1101 1102 /* 1103 * If something gets miswired and we end up here for a kernel mode 1104 * #DB, we will malfunction. 1105 */ 1106 WARN_ON_ONCE(!user_mode(regs)); 1107 1108 /* 1109 * NB: We can't easily clear DR7 here because 1110 * irqentry_exit_to_usermode() can invoke ptrace, schedule, access 1111 * user memory, etc. This means that a recursive #DB is possible. If 1112 * this happens, that #DB will hit exc_debug_kernel() and clear DR7. 1113 * Since we're not on the IST stack right now, everything will be 1114 * fine. 1115 */ 1116 1117 irqentry_enter_from_user_mode(regs); 1118 instrumentation_begin(); 1119 1120 /* 1121 * Start the virtual/ptrace DR6 value with just the DR_STEP mask 1122 * of the real DR6. ptrace_triggered() will set the DR_TRAPn bits. 1123 * 1124 * Userspace expects DR_STEP to be visible in ptrace_get_debugreg(6) 1125 * even if it is not the result of PTRACE_SINGLESTEP. 1126 */ 1127 current->thread.virtual_dr6 = (dr6 & DR_STEP); 1128 1129 /* 1130 * The SDM says "The processor clears the BTF flag when it 1131 * generates a debug exception." Clear TIF_BLOCKSTEP to keep 1132 * TIF_BLOCKSTEP in sync with the hardware BTF flag. 1133 */ 1134 clear_thread_flag(TIF_BLOCKSTEP); 1135 1136 /* 1137 * If dr6 has no reason to give us about the origin of this trap, 1138 * then it's very likely the result of an icebp/int01 trap. 1139 * User wants a sigtrap for that. 1140 */ 1141 icebp = !dr6; 1142 1143 if (notify_debug(regs, &dr6)) 1144 goto out; 1145 1146 /* It's safe to allow irq's after DR6 has been saved */ 1147 local_irq_enable(); 1148 1149 if (v8086_mode(regs)) { 1150 handle_vm86_trap((struct kernel_vm86_regs *)regs, 0, X86_TRAP_DB); 1151 goto out_irq; 1152 } 1153 1154 /* #DB for bus lock can only be triggered from userspace. */ 1155 if (dr6 & DR_BUS_LOCK) 1156 handle_bus_lock(regs); 1157 1158 /* Add the virtual_dr6 bits for signals. */ 1159 dr6 |= current->thread.virtual_dr6; 1160 if (dr6 & (DR_STEP | DR_TRAP_BITS) || icebp) 1161 send_sigtrap(regs, 0, get_si_code(dr6)); 1162 1163 out_irq: 1164 local_irq_disable(); 1165 out: 1166 instrumentation_end(); 1167 irqentry_exit_to_user_mode(regs); 1168 } 1169 1170 #ifdef CONFIG_X86_64 1171 /* IST stack entry */ 1172 DEFINE_IDTENTRY_DEBUG(exc_debug) 1173 { 1174 exc_debug_kernel(regs, debug_read_clear_dr6()); 1175 } 1176 1177 /* User entry, runs on regular task stack */ 1178 DEFINE_IDTENTRY_DEBUG_USER(exc_debug) 1179 { 1180 exc_debug_user(regs, debug_read_clear_dr6()); 1181 } 1182 #else 1183 /* 32 bit does not have separate entry points. */ 1184 DEFINE_IDTENTRY_RAW(exc_debug) 1185 { 1186 unsigned long dr6 = debug_read_clear_dr6(); 1187 1188 if (user_mode(regs)) 1189 exc_debug_user(regs, dr6); 1190 else 1191 exc_debug_kernel(regs, dr6); 1192 } 1193 #endif 1194 1195 /* 1196 * Note that we play around with the 'TS' bit in an attempt to get 1197 * the correct behaviour even in the presence of the asynchronous 1198 * IRQ13 behaviour 1199 */ 1200 static void math_error(struct pt_regs *regs, int trapnr) 1201 { 1202 struct task_struct *task = current; 1203 struct fpu *fpu = &task->thread.fpu; 1204 int si_code; 1205 char *str = (trapnr == X86_TRAP_MF) ? "fpu exception" : 1206 "simd exception"; 1207 1208 cond_local_irq_enable(regs); 1209 1210 if (!user_mode(regs)) { 1211 if (fixup_exception(regs, trapnr, 0, 0)) 1212 goto exit; 1213 1214 task->thread.error_code = 0; 1215 task->thread.trap_nr = trapnr; 1216 1217 if (notify_die(DIE_TRAP, str, regs, 0, trapnr, 1218 SIGFPE) != NOTIFY_STOP) 1219 die(str, regs, 0); 1220 goto exit; 1221 } 1222 1223 /* 1224 * Synchronize the FPU register state to the memory register state 1225 * if necessary. This allows the exception handler to inspect it. 1226 */ 1227 fpu_sync_fpstate(fpu); 1228 1229 task->thread.trap_nr = trapnr; 1230 task->thread.error_code = 0; 1231 1232 si_code = fpu__exception_code(fpu, trapnr); 1233 /* Retry when we get spurious exceptions: */ 1234 if (!si_code) 1235 goto exit; 1236 1237 if (fixup_vdso_exception(regs, trapnr, 0, 0)) 1238 goto exit; 1239 1240 force_sig_fault(SIGFPE, si_code, 1241 (void __user *)uprobe_get_trap_addr(regs)); 1242 exit: 1243 cond_local_irq_disable(regs); 1244 } 1245 1246 DEFINE_IDTENTRY(exc_coprocessor_error) 1247 { 1248 math_error(regs, X86_TRAP_MF); 1249 } 1250 1251 DEFINE_IDTENTRY(exc_simd_coprocessor_error) 1252 { 1253 if (IS_ENABLED(CONFIG_X86_INVD_BUG)) { 1254 /* AMD 486 bug: INVD in CPL 0 raises #XF instead of #GP */ 1255 if (!static_cpu_has(X86_FEATURE_XMM)) { 1256 __exc_general_protection(regs, 0); 1257 return; 1258 } 1259 } 1260 math_error(regs, X86_TRAP_XF); 1261 } 1262 1263 DEFINE_IDTENTRY(exc_spurious_interrupt_bug) 1264 { 1265 /* 1266 * This addresses a Pentium Pro Erratum: 1267 * 1268 * PROBLEM: If the APIC subsystem is configured in mixed mode with 1269 * Virtual Wire mode implemented through the local APIC, an 1270 * interrupt vector of 0Fh (Intel reserved encoding) may be 1271 * generated by the local APIC (Int 15). This vector may be 1272 * generated upon receipt of a spurious interrupt (an interrupt 1273 * which is removed before the system receives the INTA sequence) 1274 * instead of the programmed 8259 spurious interrupt vector. 1275 * 1276 * IMPLICATION: The spurious interrupt vector programmed in the 1277 * 8259 is normally handled by an operating system's spurious 1278 * interrupt handler. However, a vector of 0Fh is unknown to some 1279 * operating systems, which would crash if this erratum occurred. 1280 * 1281 * In theory this could be limited to 32bit, but the handler is not 1282 * hurting and who knows which other CPUs suffer from this. 1283 */ 1284 } 1285 1286 static bool handle_xfd_event(struct pt_regs *regs) 1287 { 1288 u64 xfd_err; 1289 int err; 1290 1291 if (!IS_ENABLED(CONFIG_X86_64) || !cpu_feature_enabled(X86_FEATURE_XFD)) 1292 return false; 1293 1294 rdmsrl(MSR_IA32_XFD_ERR, xfd_err); 1295 if (!xfd_err) 1296 return false; 1297 1298 wrmsrl(MSR_IA32_XFD_ERR, 0); 1299 1300 /* Die if that happens in kernel space */ 1301 if (WARN_ON(!user_mode(regs))) 1302 return false; 1303 1304 local_irq_enable(); 1305 1306 err = xfd_enable_feature(xfd_err); 1307 1308 switch (err) { 1309 case -EPERM: 1310 force_sig_fault(SIGILL, ILL_ILLOPC, error_get_trap_addr(regs)); 1311 break; 1312 case -EFAULT: 1313 force_sig(SIGSEGV); 1314 break; 1315 } 1316 1317 local_irq_disable(); 1318 return true; 1319 } 1320 1321 DEFINE_IDTENTRY(exc_device_not_available) 1322 { 1323 unsigned long cr0 = read_cr0(); 1324 1325 if (handle_xfd_event(regs)) 1326 return; 1327 1328 #ifdef CONFIG_MATH_EMULATION 1329 if (!boot_cpu_has(X86_FEATURE_FPU) && (cr0 & X86_CR0_EM)) { 1330 struct math_emu_info info = { }; 1331 1332 cond_local_irq_enable(regs); 1333 1334 info.regs = regs; 1335 math_emulate(&info); 1336 1337 cond_local_irq_disable(regs); 1338 return; 1339 } 1340 #endif 1341 1342 /* This should not happen. */ 1343 if (WARN(cr0 & X86_CR0_TS, "CR0.TS was set")) { 1344 /* Try to fix it up and carry on. */ 1345 write_cr0(cr0 & ~X86_CR0_TS); 1346 } else { 1347 /* 1348 * Something terrible happened, and we're better off trying 1349 * to kill the task than getting stuck in a never-ending 1350 * loop of #NM faults. 1351 */ 1352 die("unexpected #NM exception", regs, 0); 1353 } 1354 } 1355 1356 #ifdef CONFIG_INTEL_TDX_GUEST 1357 1358 #define VE_FAULT_STR "VE fault" 1359 1360 static void ve_raise_fault(struct pt_regs *regs, long error_code) 1361 { 1362 if (user_mode(regs)) { 1363 gp_user_force_sig_segv(regs, X86_TRAP_VE, error_code, VE_FAULT_STR); 1364 return; 1365 } 1366 1367 if (gp_try_fixup_and_notify(regs, X86_TRAP_VE, error_code, VE_FAULT_STR)) 1368 return; 1369 1370 die_addr(VE_FAULT_STR, regs, error_code, 0); 1371 } 1372 1373 /* 1374 * Virtualization Exceptions (#VE) are delivered to TDX guests due to 1375 * specific guest actions which may happen in either user space or the 1376 * kernel: 1377 * 1378 * * Specific instructions (WBINVD, for example) 1379 * * Specific MSR accesses 1380 * * Specific CPUID leaf accesses 1381 * * Access to specific guest physical addresses 1382 * 1383 * In the settings that Linux will run in, virtualization exceptions are 1384 * never generated on accesses to normal, TD-private memory that has been 1385 * accepted (by BIOS or with tdx_enc_status_changed()). 1386 * 1387 * Syscall entry code has a critical window where the kernel stack is not 1388 * yet set up. Any exception in this window leads to hard to debug issues 1389 * and can be exploited for privilege escalation. Exceptions in the NMI 1390 * entry code also cause issues. Returning from the exception handler with 1391 * IRET will re-enable NMIs and nested NMI will corrupt the NMI stack. 1392 * 1393 * For these reasons, the kernel avoids #VEs during the syscall gap and 1394 * the NMI entry code. Entry code paths do not access TD-shared memory, 1395 * MMIO regions, use #VE triggering MSRs, instructions, or CPUID leaves 1396 * that might generate #VE. VMM can remove memory from TD at any point, 1397 * but access to unaccepted (or missing) private memory leads to VM 1398 * termination, not to #VE. 1399 * 1400 * Similarly to page faults and breakpoints, #VEs are allowed in NMI 1401 * handlers once the kernel is ready to deal with nested NMIs. 1402 * 1403 * During #VE delivery, all interrupts, including NMIs, are blocked until 1404 * TDGETVEINFO is called. It prevents #VE nesting until the kernel reads 1405 * the VE info. 1406 * 1407 * If a guest kernel action which would normally cause a #VE occurs in 1408 * the interrupt-disabled region before TDGETVEINFO, a #DF (fault 1409 * exception) is delivered to the guest which will result in an oops. 1410 * 1411 * The entry code has been audited carefully for following these expectations. 1412 * Changes in the entry code have to be audited for correctness vs. this 1413 * aspect. Similarly to #PF, #VE in these places will expose kernel to 1414 * privilege escalation or may lead to random crashes. 1415 */ 1416 DEFINE_IDTENTRY(exc_virtualization_exception) 1417 { 1418 struct ve_info ve; 1419 1420 /* 1421 * NMIs/Machine-checks/Interrupts will be in a disabled state 1422 * till TDGETVEINFO TDCALL is executed. This ensures that VE 1423 * info cannot be overwritten by a nested #VE. 1424 */ 1425 tdx_get_ve_info(&ve); 1426 1427 cond_local_irq_enable(regs); 1428 1429 /* 1430 * If tdx_handle_virt_exception() could not process 1431 * it successfully, treat it as #GP(0) and handle it. 1432 */ 1433 if (!tdx_handle_virt_exception(regs, &ve)) 1434 ve_raise_fault(regs, 0); 1435 1436 cond_local_irq_disable(regs); 1437 } 1438 1439 #endif 1440 1441 #ifdef CONFIG_X86_32 1442 DEFINE_IDTENTRY_SW(iret_error) 1443 { 1444 local_irq_enable(); 1445 if (notify_die(DIE_TRAP, "iret exception", regs, 0, 1446 X86_TRAP_IRET, SIGILL) != NOTIFY_STOP) { 1447 do_trap(X86_TRAP_IRET, SIGILL, "iret exception", regs, 0, 1448 ILL_BADSTK, (void __user *)NULL); 1449 } 1450 local_irq_disable(); 1451 } 1452 #endif 1453 1454 void __init trap_init(void) 1455 { 1456 /* Init cpu_entry_area before IST entries are set up */ 1457 setup_cpu_entry_areas(); 1458 1459 /* Init GHCB memory pages when running as an SEV-ES guest */ 1460 sev_es_init_vc_handling(); 1461 1462 /* Initialize TSS before setting up traps so ISTs work */ 1463 cpu_init_exception_handling(); 1464 /* Setup traps as cpu_init() might #GP */ 1465 idt_setup_traps(); 1466 cpu_init(); 1467 } 1468