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