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/cpu.h>
41#include <linux/io.h>
42#include <linux/hardirq.h>
43#include <linux/atomic.h>
44#include <linux/iommu.h>
45
46#include <asm/stacktrace.h>
47#include <asm/processor.h>
48#include <asm/debugreg.h>
49#include <asm/realmode.h>
50#include <asm/text-patching.h>
51#include <asm/ftrace.h>
52#include <asm/traps.h>
53#include <asm/desc.h>
54#include <asm/fred.h>
55#include <asm/fpu/api.h>
56#include <asm/cpu.h>
57#include <asm/cpu_entry_area.h>
58#include <asm/mce.h>
59#include <asm/fixmap.h>
60#include <asm/mach_traps.h>
61#include <asm/alternative.h>
62#include <asm/fpu/xstate.h>
63#include <asm/vm86.h>
64#include <asm/umip.h>
65#include <asm/insn.h>
66#include <asm/insn-eval.h>
67#include <asm/vdso.h>
68#include <asm/tdx.h>
69#include <asm/cfi.h>
70
71#ifdef CONFIG_X86_64
72#include <asm/x86_init.h>
73#else
74#include <asm/processor-flags.h>
75#include <asm/setup.h>
76#endif
77
78#include <asm/proto.h>
79
80DECLARE_BITMAP(system_vectors, NR_VECTORS);
81
82__always_inline int is_valid_bugaddr(unsigned long addr)
83{
84 if (addr < TASK_SIZE_MAX)
85 return 0;
86
87 /*
88 * We got #UD, if the text isn't readable we'd have gotten
89 * a different exception.
90 */
91 return *(unsigned short *)addr == INSN_UD2;
92}
93
94static nokprobe_inline int
95do_trap_no_signal(struct task_struct *tsk, int trapnr, const char *str,
96 struct pt_regs *regs, long error_code)
97{
98 if (v8086_mode(regs)) {
99 /*
100 * Traps 0, 1, 3, 4, and 5 should be forwarded to vm86.
101 * On nmi (interrupt 2), do_trap should not be called.
102 */
103 if (trapnr < X86_TRAP_UD) {
104 if (!handle_vm86_trap(a: (struct kernel_vm86_regs *) regs,
105 b: error_code, c: trapnr))
106 return 0;
107 }
108 } else if (!user_mode(regs)) {
109 if (fixup_exception(regs, trapnr, error_code, fault_addr: 0))
110 return 0;
111
112 tsk->thread.error_code = error_code;
113 tsk->thread.trap_nr = trapnr;
114 die(str, regs, error_code);
115 } else {
116 if (fixup_vdso_exception(regs, trapnr, error_code, fault_addr: 0))
117 return 0;
118 }
119
120 /*
121 * We want error_code and trap_nr set for userspace faults and
122 * kernelspace faults which result in die(), but not
123 * kernelspace faults which are fixed up. die() gives the
124 * process no chance to handle the signal and notice the
125 * kernel fault information, so that won't result in polluting
126 * the information about previously queued, but not yet
127 * delivered, faults. See also exc_general_protection below.
128 */
129 tsk->thread.error_code = error_code;
130 tsk->thread.trap_nr = trapnr;
131
132 return -1;
133}
134
135static void show_signal(struct task_struct *tsk, int signr,
136 const char *type, const char *desc,
137 struct pt_regs *regs, long error_code)
138{
139 if (show_unhandled_signals && unhandled_signal(tsk, sig: signr) &&
140 printk_ratelimit()) {
141 pr_info("%s[%d] %s%s ip:%lx sp:%lx error:%lx",
142 tsk->comm, task_pid_nr(tsk), type, desc,
143 regs->ip, regs->sp, error_code);
144 print_vma_addr(KERN_CONT " in ", rip: regs->ip);
145 pr_cont("\n");
146 }
147}
148
149static void
150do_trap(int trapnr, int signr, char *str, struct pt_regs *regs,
151 long error_code, int sicode, void __user *addr)
152{
153 struct task_struct *tsk = current;
154
155 if (!do_trap_no_signal(tsk, trapnr, str, regs, error_code))
156 return;
157
158 show_signal(tsk, signr, type: "trap ", desc: str, regs, error_code);
159
160 if (!sicode)
161 force_sig(signr);
162 else
163 force_sig_fault(sig: signr, code: sicode, addr);
164}
165NOKPROBE_SYMBOL(do_trap);
166
167static void do_error_trap(struct pt_regs *regs, long error_code, char *str,
168 unsigned long trapnr, int signr, int sicode, void __user *addr)
169{
170 RCU_LOCKDEP_WARN(!rcu_is_watching(), "entry code didn't wake RCU");
171
172 if (notify_die(val: DIE_TRAP, str, regs, err: error_code, trap: trapnr, sig: signr) !=
173 NOTIFY_STOP) {
174 cond_local_irq_enable(regs);
175 do_trap(trapnr, signr, str, regs, error_code, sicode, addr);
176 cond_local_irq_disable(regs);
177 }
178}
179
180/*
181 * Posix requires to provide the address of the faulting instruction for
182 * SIGILL (#UD) and SIGFPE (#DE) in the si_addr member of siginfo_t.
183 *
184 * This address is usually regs->ip, but when an uprobe moved the code out
185 * of line then regs->ip points to the XOL code which would confuse
186 * anything which analyzes the fault address vs. the unmodified binary. If
187 * a trap happened in XOL code then uprobe maps regs->ip back to the
188 * original instruction address.
189 */
190static __always_inline void __user *error_get_trap_addr(struct pt_regs *regs)
191{
192 return (void __user *)uprobe_get_trap_addr(regs);
193}
194
195DEFINE_IDTENTRY(exc_divide_error)
196{
197 do_error_trap(regs, error_code: 0, str: "divide error", X86_TRAP_DE, SIGFPE,
198 FPE_INTDIV, addr: error_get_trap_addr(regs));
199}
200
201DEFINE_IDTENTRY(exc_overflow)
202{
203 do_error_trap(regs, error_code: 0, str: "overflow", X86_TRAP_OF, SIGSEGV, sicode: 0, NULL);
204}
205
206#ifdef CONFIG_X86_F00F_BUG
207void handle_invalid_op(struct pt_regs *regs)
208#else
209static inline void handle_invalid_op(struct pt_regs *regs)
210#endif
211{
212 do_error_trap(regs, error_code: 0, str: "invalid opcode", X86_TRAP_UD, SIGILL,
213 ILL_ILLOPN, addr: error_get_trap_addr(regs));
214}
215
216static noinstr bool handle_bug(struct pt_regs *regs)
217{
218 bool handled = false;
219
220 /*
221 * Normally @regs are unpoisoned by irqentry_enter(), but handle_bug()
222 * is a rare case that uses @regs without passing them to
223 * irqentry_enter().
224 */
225 kmsan_unpoison_entry_regs(regs);
226 if (!is_valid_bugaddr(addr: regs->ip))
227 return handled;
228
229 /*
230 * All lies, just get the WARN/BUG out.
231 */
232 instrumentation_begin();
233 /*
234 * Since we're emulating a CALL with exceptions, restore the interrupt
235 * state to what it was at the exception site.
236 */
237 if (regs->flags & X86_EFLAGS_IF)
238 raw_local_irq_enable();
239 if (report_bug(bug_addr: regs->ip, regs) == BUG_TRAP_TYPE_WARN ||
240 handle_cfi_failure(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
251DEFINE_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
270DEFINE_IDTENTRY(exc_coproc_segment_overrun)
271{
272 do_error_trap(regs, error_code: 0, str: "coprocessor segment overrun",
273 X86_TRAP_OLD_MF, SIGFPE, sicode: 0, NULL);
274}
275
276DEFINE_IDTENTRY_ERRORCODE(exc_invalid_tss)
277{
278 do_error_trap(regs, error_code, str: "invalid TSS", X86_TRAP_TS, SIGSEGV,
279 sicode: 0, NULL);
280}
281
282DEFINE_IDTENTRY_ERRORCODE(exc_segment_not_present)
283{
284 do_error_trap(regs, error_code, str: "segment not present", X86_TRAP_NP,
285 SIGBUS, sicode: 0, NULL);
286}
287
288DEFINE_IDTENTRY_ERRORCODE(exc_stack_segment)
289{
290 do_error_trap(regs, error_code, str: "stack segment", X86_TRAP_SS, SIGBUS,
291 sicode: 0, NULL);
292}
293
294DEFINE_IDTENTRY_ERRORCODE(exc_alignment_check)
295{
296 char *str = "alignment check";
297
298 if (notify_die(val: DIE_TRAP, str, regs, err: 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, str: "alignment check", regs,
310 error_code, BUS_ADRALN, NULL);
311
312out:
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(type: 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(fmt: "%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 */
352DEFINE_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(val: DIE_TRAP, str, regs, err: 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(stack: (void *)address, info: &info))
463 handle_stack_overflow(regs, fault_address: address, info: &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(fmt: "Machine halted.");
469 instrumentation_end();
470}
471
472DEFINE_IDTENTRY(exc_bounds)
473{
474 if (notify_die(val: DIE_TRAP, str: "bounds", regs, err: 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, str: "bounds", regs, error_code: 0, sicode: 0, NULL);
483
484 cond_local_irq_disable(regs);
485}
486
487enum 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 */
498static 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(dst: insn_buf, src: (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: &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
533static bool fixup_iopl_exception(struct pt_regs *regs)
534{
535 struct thread_struct *t = &current->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: &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 ", rip: 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 */
568static bool try_fixup_enqcmd_gp(void)
569{
570#ifdef CONFIG_ARCH_HAS_CPU_PASID
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 /*
590 * If the mm has not been allocated a
591 * PASID, the #GP can not be fixed up.
592 */
593 if (!mm_valid_pasid(current->mm))
594 return false;
595
596 pasid = mm_get_enqcmd_pasid(current->mm);
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, val: pasid | MSR_IA32_PASID_VALID);
606 current->pasid_activated = 1;
607
608 return true;
609#else
610 return false;
611#endif
612}
613
614static bool gp_try_fixup_and_notify(struct pt_regs *regs, int trapnr,
615 unsigned long error_code, const char *str,
616 unsigned long address)
617{
618 if (fixup_exception(regs, trapnr, error_code, fault_addr: address))
619 return true;
620
621 current->thread.error_code = error_code;
622 current->thread.trap_nr = trapnr;
623
624 /*
625 * To be potentially processing a kprobe fault and to trust the result
626 * from kprobe_running(), we have to be non-preemptible.
627 */
628 if (!preemptible() && kprobe_running() &&
629 kprobe_fault_handler(regs, trapnr))
630 return true;
631
632 return notify_die(val: DIE_GPF, str, regs, err: error_code, trap: trapnr, SIGSEGV) == NOTIFY_STOP;
633}
634
635static void gp_user_force_sig_segv(struct pt_regs *regs, int trapnr,
636 unsigned long error_code, const char *str)
637{
638 current->thread.error_code = error_code;
639 current->thread.trap_nr = trapnr;
640 show_signal(current, SIGSEGV, type: "", desc: str, regs, error_code);
641 force_sig(SIGSEGV);
642}
643
644DEFINE_IDTENTRY_ERRORCODE(exc_general_protection)
645{
646 char desc[sizeof(GPFSTR) + 50 + 2*sizeof(unsigned long) + 1] = GPFSTR;
647 enum kernel_gp_hint hint = GP_NO_HINT;
648 unsigned long gp_addr;
649
650 if (user_mode(regs) && try_fixup_enqcmd_gp())
651 return;
652
653 cond_local_irq_enable(regs);
654
655 if (static_cpu_has(X86_FEATURE_UMIP)) {
656 if (user_mode(regs) && fixup_umip_exception(regs))
657 goto exit;
658 }
659
660 if (v8086_mode(regs)) {
661 local_irq_enable();
662 handle_vm86_fault((struct kernel_vm86_regs *) regs, error_code);
663 local_irq_disable();
664 return;
665 }
666
667 if (user_mode(regs)) {
668 if (fixup_iopl_exception(regs))
669 goto exit;
670
671 if (fixup_vdso_exception(regs, X86_TRAP_GP, error_code, fault_addr: 0))
672 goto exit;
673
674 gp_user_force_sig_segv(regs, X86_TRAP_GP, error_code, str: desc);
675 goto exit;
676 }
677
678 if (gp_try_fixup_and_notify(regs, X86_TRAP_GP, error_code, str: desc, address: 0))
679 goto exit;
680
681 if (error_code)
682 snprintf(buf: desc, size: sizeof(desc), fmt: "segment-related " GPFSTR);
683 else
684 hint = get_kernel_gp_address(regs, addr: &gp_addr);
685
686 if (hint != GP_NO_HINT)
687 snprintf(buf: desc, size: sizeof(desc), GPFSTR ", %s 0x%lx",
688 (hint == GP_NON_CANONICAL) ? "probably for non-canonical address"
689 : "maybe for address",
690 gp_addr);
691
692 /*
693 * KASAN is interested only in the non-canonical case, clear it
694 * otherwise.
695 */
696 if (hint != GP_NON_CANONICAL)
697 gp_addr = 0;
698
699 die_addr(str: desc, regs, err: error_code, gp_addr);
700
701exit:
702 cond_local_irq_disable(regs);
703}
704
705static bool do_int3(struct pt_regs *regs)
706{
707 int res;
708
709#ifdef CONFIG_KGDB_LOW_LEVEL_TRAP
710 if (kgdb_ll_trap(cmd: DIE_INT3, str: "int3", regs, err: 0, X86_TRAP_BP,
711 SIGTRAP) == NOTIFY_STOP)
712 return true;
713#endif /* CONFIG_KGDB_LOW_LEVEL_TRAP */
714
715#ifdef CONFIG_KPROBES
716 if (kprobe_int3_handler(regs))
717 return true;
718#endif
719 res = notify_die(val: DIE_INT3, str: "int3", regs, err: 0, X86_TRAP_BP, SIGTRAP);
720
721 return res == NOTIFY_STOP;
722}
723NOKPROBE_SYMBOL(do_int3);
724
725static void do_int3_user(struct pt_regs *regs)
726{
727 if (do_int3(regs))
728 return;
729
730 cond_local_irq_enable(regs);
731 do_trap(X86_TRAP_BP, SIGTRAP, str: "int3", regs, error_code: 0, sicode: 0, NULL);
732 cond_local_irq_disable(regs);
733}
734
735DEFINE_IDTENTRY_RAW(exc_int3)
736{
737 /*
738 * poke_int3_handler() is completely self contained code; it does (and
739 * must) *NOT* call out to anything, lest it hits upon yet another
740 * INT3.
741 */
742 if (poke_int3_handler(regs))
743 return;
744
745 /*
746 * irqentry_enter_from_user_mode() uses static_branch_{,un}likely()
747 * and therefore can trigger INT3, hence poke_int3_handler() must
748 * be done before. If the entry came from kernel mode, then use
749 * nmi_enter() because the INT3 could have been hit in any context
750 * including NMI.
751 */
752 if (user_mode(regs)) {
753 irqentry_enter_from_user_mode(regs);
754 instrumentation_begin();
755 do_int3_user(regs);
756 instrumentation_end();
757 irqentry_exit_to_user_mode(regs);
758 } else {
759 irqentry_state_t irq_state = irqentry_nmi_enter(regs);
760
761 instrumentation_begin();
762 if (!do_int3(regs))
763 die("int3", regs, 0);
764 instrumentation_end();
765 irqentry_nmi_exit(regs, irq_state);
766 }
767}
768
769#ifdef CONFIG_X86_64
770/*
771 * Help handler running on a per-cpu (IST or entry trampoline) stack
772 * to switch to the normal thread stack if the interrupted code was in
773 * user mode. The actual stack switch is done in entry_64.S
774 */
775asmlinkage __visible noinstr struct pt_regs *sync_regs(struct pt_regs *eregs)
776{
777 struct pt_regs *regs = (struct pt_regs *)current_top_of_stack() - 1;
778 if (regs != eregs)
779 *regs = *eregs;
780 return regs;
781}
782
783#ifdef CONFIG_AMD_MEM_ENCRYPT
784asmlinkage __visible noinstr struct pt_regs *vc_switch_off_ist(struct pt_regs *regs)
785{
786 unsigned long sp, *stack;
787 struct stack_info info;
788 struct pt_regs *regs_ret;
789
790 /*
791 * In the SYSCALL entry path the RSP value comes from user-space - don't
792 * trust it and switch to the current kernel stack
793 */
794 if (ip_within_syscall_gap(regs)) {
795 sp = current_top_of_stack();
796 goto sync;
797 }
798
799 /*
800 * From here on the RSP value is trusted. Now check whether entry
801 * happened from a safe stack. Not safe are the entry or unknown stacks,
802 * use the fall-back stack instead in this case.
803 */
804 sp = regs->sp;
805 stack = (unsigned long *)sp;
806
807 if (!get_stack_info_noinstr(stack, current, info: &info) || info.type == STACK_TYPE_ENTRY ||
808 info.type > STACK_TYPE_EXCEPTION_LAST)
809 sp = __this_cpu_ist_top_va(VC2);
810
811sync:
812 /*
813 * Found a safe stack - switch to it as if the entry didn't happen via
814 * IST stack. The code below only copies pt_regs, the real switch happens
815 * in assembly code.
816 */
817 sp = ALIGN_DOWN(sp, 8) - sizeof(*regs_ret);
818
819 regs_ret = (struct pt_regs *)sp;
820 *regs_ret = *regs;
821
822 return regs_ret;
823}
824#endif
825
826asmlinkage __visible noinstr struct pt_regs *fixup_bad_iret(struct pt_regs *bad_regs)
827{
828 struct pt_regs tmp, *new_stack;
829
830 /*
831 * This is called from entry_64.S early in handling a fault
832 * caused by a bad iret to user mode. To handle the fault
833 * correctly, we want to move our stack frame to where it would
834 * be had we entered directly on the entry stack (rather than
835 * just below the IRET frame) and we want to pretend that the
836 * exception came from the IRET target.
837 */
838 new_stack = (struct pt_regs *)__this_cpu_read(cpu_tss_rw.x86_tss.sp0) - 1;
839
840 /* Copy the IRET target to the temporary storage. */
841 __memcpy(to: &tmp.ip, from: (void *)bad_regs->sp, len: 5*8);
842
843 /* Copy the remainder of the stack from the current stack. */
844 __memcpy(to: &tmp, from: bad_regs, offsetof(struct pt_regs, ip));
845
846 /* Update the entry stack */
847 __memcpy(to: new_stack, from: &tmp, len: sizeof(tmp));
848
849 BUG_ON(!user_mode(new_stack));
850 return new_stack;
851}
852#endif
853
854static bool is_sysenter_singlestep(struct pt_regs *regs)
855{
856 /*
857 * We don't try for precision here. If we're anywhere in the region of
858 * code that can be single-stepped in the SYSENTER entry path, then
859 * assume that this is a useless single-step trap due to SYSENTER
860 * being invoked with TF set. (We don't know in advance exactly
861 * which instructions will be hit because BTF could plausibly
862 * be set.)
863 */
864#ifdef CONFIG_X86_32
865 return (regs->ip - (unsigned long)__begin_SYSENTER_singlestep_region) <
866 (unsigned long)__end_SYSENTER_singlestep_region -
867 (unsigned long)__begin_SYSENTER_singlestep_region;
868#elif defined(CONFIG_IA32_EMULATION)
869 return (regs->ip - (unsigned long)entry_SYSENTER_compat) <
870 (unsigned long)__end_entry_SYSENTER_compat -
871 (unsigned long)entry_SYSENTER_compat;
872#else
873 return false;
874#endif
875}
876
877static __always_inline unsigned long debug_read_clear_dr6(void)
878{
879 unsigned long dr6;
880
881 /*
882 * The Intel SDM says:
883 *
884 * Certain debug exceptions may clear bits 0-3. The remaining
885 * contents of the DR6 register are never cleared by the
886 * processor. To avoid confusion in identifying debug
887 * exceptions, debug handlers should clear the register before
888 * returning to the interrupted task.
889 *
890 * Keep it simple: clear DR6 immediately.
891 */
892 get_debugreg(dr6, 6);
893 set_debugreg(DR6_RESERVED, reg: 6);
894 dr6 ^= DR6_RESERVED; /* Flip to positive polarity */
895
896 return dr6;
897}
898
899/*
900 * Our handling of the processor debug registers is non-trivial.
901 * We do not clear them on entry and exit from the kernel. Therefore
902 * it is possible to get a watchpoint trap here from inside the kernel.
903 * However, the code in ./ptrace.c has ensured that the user can
904 * only set watchpoints on userspace addresses. Therefore the in-kernel
905 * watchpoint trap can only occur in code which is reading/writing
906 * from user space. Such code must not hold kernel locks (since it
907 * can equally take a page fault), therefore it is safe to call
908 * force_sig_info even though that claims and releases locks.
909 *
910 * Code in ./signal.c ensures that the debug control register
911 * is restored before we deliver any signal, and therefore that
912 * user code runs with the correct debug control register even though
913 * we clear it here.
914 *
915 * Being careful here means that we don't have to be as careful in a
916 * lot of more complicated places (task switching can be a bit lazy
917 * about restoring all the debug state, and ptrace doesn't have to
918 * find every occurrence of the TF bit that could be saved away even
919 * by user code)
920 *
921 * May run on IST stack.
922 */
923
924static bool notify_debug(struct pt_regs *regs, unsigned long *dr6)
925{
926 /*
927 * Notifiers will clear bits in @dr6 to indicate the event has been
928 * consumed - hw_breakpoint_handler(), single_stop_cont().
929 *
930 * Notifiers will set bits in @virtual_dr6 to indicate the desire
931 * for signals - ptrace_triggered(), kgdb_hw_overflow_handler().
932 */
933 if (notify_die(val: DIE_DEBUG, str: "debug", regs, err: (long)dr6, trap: 0, SIGTRAP) == NOTIFY_STOP)
934 return true;
935
936 return false;
937}
938
939static noinstr void exc_debug_kernel(struct pt_regs *regs, 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 * For FRED, nested #DB should just work fine. But when a watchpoint or
953 * breakpoint is set in the code path which is executed by #DB handler,
954 * it results in an endless recursion and stack overflow. Thus we stay
955 * with the IDT approach, i.e., save DR7 and disable #DB.
956 */
957 unsigned long dr7 = local_db_save();
958 irqentry_state_t irq_state = irqentry_nmi_enter(regs);
959 instrumentation_begin();
960
961 /*
962 * If something gets miswired and we end up here for a user mode
963 * #DB, we will malfunction.
964 */
965 WARN_ON_ONCE(user_mode(regs));
966
967 if (test_thread_flag(TIF_BLOCKSTEP)) {
968 /*
969 * The SDM says "The processor clears the BTF flag when it
970 * generates a debug exception." but PTRACE_BLOCKSTEP requested
971 * it for userspace, but we just took a kernel #DB, so re-set
972 * BTF.
973 */
974 unsigned long debugctl;
975
976 rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
977 debugctl |= DEBUGCTLMSR_BTF;
978 wrmsrl(MSR_IA32_DEBUGCTLMSR, val: debugctl);
979 }
980
981 /*
982 * Catch SYSENTER with TF set and clear DR_STEP. If this hit a
983 * watchpoint at the same time then that will still be handled.
984 */
985 if (!cpu_feature_enabled(X86_FEATURE_FRED) &&
986 (dr6 & DR_STEP) && is_sysenter_singlestep(regs))
987 dr6 &= ~DR_STEP;
988
989 /*
990 * The kernel doesn't use INT1
991 */
992 if (!dr6)
993 goto out;
994
995 if (notify_debug(regs, dr6: &dr6))
996 goto out;
997
998 /*
999 * The kernel doesn't use TF single-step outside of:
1000 *
1001 * - Kprobes, consumed through kprobe_debug_handler()
1002 * - KGDB, consumed through notify_debug()
1003 *
1004 * So if we get here with DR_STEP set, something is wonky.
1005 *
1006 * A known way to trigger this is through QEMU's GDB stub,
1007 * which leaks #DB into the guest and causes IST recursion.
1008 */
1009 if (WARN_ON_ONCE(dr6 & DR_STEP))
1010 regs->flags &= ~X86_EFLAGS_TF;
1011out:
1012 instrumentation_end();
1013 irqentry_nmi_exit(regs, irq_state);
1014
1015 local_db_restore(dr7);
1016}
1017
1018static noinstr void exc_debug_user(struct pt_regs *regs, unsigned long dr6)
1019{
1020 bool icebp;
1021
1022 /*
1023 * If something gets miswired and we end up here for a kernel mode
1024 * #DB, we will malfunction.
1025 */
1026 WARN_ON_ONCE(!user_mode(regs));
1027
1028 /*
1029 * NB: We can't easily clear DR7 here because
1030 * irqentry_exit_to_usermode() can invoke ptrace, schedule, access
1031 * user memory, etc. This means that a recursive #DB is possible. If
1032 * this happens, that #DB will hit exc_debug_kernel() and clear DR7.
1033 * Since we're not on the IST stack right now, everything will be
1034 * fine.
1035 */
1036
1037 irqentry_enter_from_user_mode(regs);
1038 instrumentation_begin();
1039
1040 /*
1041 * Start the virtual/ptrace DR6 value with just the DR_STEP mask
1042 * of the real DR6. ptrace_triggered() will set the DR_TRAPn bits.
1043 *
1044 * Userspace expects DR_STEP to be visible in ptrace_get_debugreg(6)
1045 * even if it is not the result of PTRACE_SINGLESTEP.
1046 */
1047 current->thread.virtual_dr6 = (dr6 & DR_STEP);
1048
1049 /*
1050 * The SDM says "The processor clears the BTF flag when it
1051 * generates a debug exception." Clear TIF_BLOCKSTEP to keep
1052 * TIF_BLOCKSTEP in sync with the hardware BTF flag.
1053 */
1054 clear_thread_flag(TIF_BLOCKSTEP);
1055
1056 /*
1057 * If dr6 has no reason to give us about the origin of this trap,
1058 * then it's very likely the result of an icebp/int01 trap.
1059 * User wants a sigtrap for that.
1060 */
1061 icebp = !dr6;
1062
1063 if (notify_debug(regs, dr6: &dr6))
1064 goto out;
1065
1066 /* It's safe to allow irq's after DR6 has been saved */
1067 local_irq_enable();
1068
1069 if (v8086_mode(regs)) {
1070 handle_vm86_trap(a: (struct kernel_vm86_regs *)regs, b: 0, X86_TRAP_DB);
1071 goto out_irq;
1072 }
1073
1074 /* #DB for bus lock can only be triggered from userspace. */
1075 if (dr6 & DR_BUS_LOCK)
1076 handle_bus_lock(regs);
1077
1078 /* Add the virtual_dr6 bits for signals. */
1079 dr6 |= current->thread.virtual_dr6;
1080 if (dr6 & (DR_STEP | DR_TRAP_BITS) || icebp)
1081 send_sigtrap(regs, error_code: 0, si_code: get_si_code(condition: dr6));
1082
1083out_irq:
1084 local_irq_disable();
1085out:
1086 instrumentation_end();
1087 irqentry_exit_to_user_mode(regs);
1088}
1089
1090#ifdef CONFIG_X86_64
1091/* IST stack entry */
1092DEFINE_IDTENTRY_DEBUG(exc_debug)
1093{
1094 exc_debug_kernel(regs, dr6: debug_read_clear_dr6());
1095}
1096
1097/* User entry, runs on regular task stack */
1098DEFINE_IDTENTRY_DEBUG_USER(exc_debug)
1099{
1100 exc_debug_user(regs, dr6: debug_read_clear_dr6());
1101}
1102
1103#ifdef CONFIG_X86_FRED
1104/*
1105 * When occurred on different ring level, i.e., from user or kernel
1106 * context, #DB needs to be handled on different stack: User #DB on
1107 * current task stack, while kernel #DB on a dedicated stack.
1108 *
1109 * This is exactly how FRED event delivery invokes an exception
1110 * handler: ring 3 event on level 0 stack, i.e., current task stack;
1111 * ring 0 event on the #DB dedicated stack specified in the
1112 * IA32_FRED_STKLVLS MSR. So unlike IDT, the FRED debug exception
1113 * entry stub doesn't do stack switch.
1114 */
1115DEFINE_FREDENTRY_DEBUG(exc_debug)
1116{
1117 /*
1118 * FRED #DB stores DR6 on the stack in the format which
1119 * debug_read_clear_dr6() returns for the IDT entry points.
1120 */
1121 unsigned long dr6 = fred_event_data(regs);
1122
1123 if (user_mode(regs))
1124 exc_debug_user(regs, dr6);
1125 else
1126 exc_debug_kernel(regs, dr6);
1127}
1128#endif /* CONFIG_X86_FRED */
1129
1130#else
1131/* 32 bit does not have separate entry points. */
1132DEFINE_IDTENTRY_RAW(exc_debug)
1133{
1134 unsigned long dr6 = debug_read_clear_dr6();
1135
1136 if (user_mode(regs))
1137 exc_debug_user(regs, dr6);
1138 else
1139 exc_debug_kernel(regs, dr6);
1140}
1141#endif
1142
1143/*
1144 * Note that we play around with the 'TS' bit in an attempt to get
1145 * the correct behaviour even in the presence of the asynchronous
1146 * IRQ13 behaviour
1147 */
1148static void math_error(struct pt_regs *regs, int trapnr)
1149{
1150 struct task_struct *task = current;
1151 struct fpu *fpu = &task->thread.fpu;
1152 int si_code;
1153 char *str = (trapnr == X86_TRAP_MF) ? "fpu exception" :
1154 "simd exception";
1155
1156 cond_local_irq_enable(regs);
1157
1158 if (!user_mode(regs)) {
1159 if (fixup_exception(regs, trapnr, error_code: 0, fault_addr: 0))
1160 goto exit;
1161
1162 task->thread.error_code = 0;
1163 task->thread.trap_nr = trapnr;
1164
1165 if (notify_die(val: DIE_TRAP, str, regs, err: 0, trap: trapnr,
1166 SIGFPE) != NOTIFY_STOP)
1167 die(str, regs, 0);
1168 goto exit;
1169 }
1170
1171 /*
1172 * Synchronize the FPU register state to the memory register state
1173 * if necessary. This allows the exception handler to inspect it.
1174 */
1175 fpu_sync_fpstate(fpu);
1176
1177 task->thread.trap_nr = trapnr;
1178 task->thread.error_code = 0;
1179
1180 si_code = fpu__exception_code(fpu, trap_nr: trapnr);
1181 /* Retry when we get spurious exceptions: */
1182 if (!si_code)
1183 goto exit;
1184
1185 if (fixup_vdso_exception(regs, trapnr, error_code: 0, fault_addr: 0))
1186 goto exit;
1187
1188 force_sig_fault(SIGFPE, code: si_code,
1189 addr: (void __user *)uprobe_get_trap_addr(regs));
1190exit:
1191 cond_local_irq_disable(regs);
1192}
1193
1194DEFINE_IDTENTRY(exc_coprocessor_error)
1195{
1196 math_error(regs, X86_TRAP_MF);
1197}
1198
1199DEFINE_IDTENTRY(exc_simd_coprocessor_error)
1200{
1201 if (IS_ENABLED(CONFIG_X86_INVD_BUG)) {
1202 /* AMD 486 bug: INVD in CPL 0 raises #XF instead of #GP */
1203 if (!static_cpu_has(X86_FEATURE_XMM)) {
1204 __exc_general_protection(regs, error_code: 0);
1205 return;
1206 }
1207 }
1208 math_error(regs, X86_TRAP_XF);
1209}
1210
1211DEFINE_IDTENTRY(exc_spurious_interrupt_bug)
1212{
1213 /*
1214 * This addresses a Pentium Pro Erratum:
1215 *
1216 * PROBLEM: If the APIC subsystem is configured in mixed mode with
1217 * Virtual Wire mode implemented through the local APIC, an
1218 * interrupt vector of 0Fh (Intel reserved encoding) may be
1219 * generated by the local APIC (Int 15). This vector may be
1220 * generated upon receipt of a spurious interrupt (an interrupt
1221 * which is removed before the system receives the INTA sequence)
1222 * instead of the programmed 8259 spurious interrupt vector.
1223 *
1224 * IMPLICATION: The spurious interrupt vector programmed in the
1225 * 8259 is normally handled by an operating system's spurious
1226 * interrupt handler. However, a vector of 0Fh is unknown to some
1227 * operating systems, which would crash if this erratum occurred.
1228 *
1229 * In theory this could be limited to 32bit, but the handler is not
1230 * hurting and who knows which other CPUs suffer from this.
1231 */
1232}
1233
1234static bool handle_xfd_event(struct pt_regs *regs)
1235{
1236 u64 xfd_err;
1237 int err;
1238
1239 if (!IS_ENABLED(CONFIG_X86_64) || !cpu_feature_enabled(X86_FEATURE_XFD))
1240 return false;
1241
1242 rdmsrl(MSR_IA32_XFD_ERR, xfd_err);
1243 if (!xfd_err)
1244 return false;
1245
1246 wrmsrl(MSR_IA32_XFD_ERR, val: 0);
1247
1248 /* Die if that happens in kernel space */
1249 if (WARN_ON(!user_mode(regs)))
1250 return false;
1251
1252 local_irq_enable();
1253
1254 err = xfd_enable_feature(xfd_err);
1255
1256 switch (err) {
1257 case -EPERM:
1258 force_sig_fault(SIGILL, ILL_ILLOPC, addr: error_get_trap_addr(regs));
1259 break;
1260 case -EFAULT:
1261 force_sig(SIGSEGV);
1262 break;
1263 }
1264
1265 local_irq_disable();
1266 return true;
1267}
1268
1269DEFINE_IDTENTRY(exc_device_not_available)
1270{
1271 unsigned long cr0 = read_cr0();
1272
1273 if (handle_xfd_event(regs))
1274 return;
1275
1276#ifdef CONFIG_MATH_EMULATION
1277 if (!boot_cpu_has(X86_FEATURE_FPU) && (cr0 & X86_CR0_EM)) {
1278 struct math_emu_info info = { };
1279
1280 cond_local_irq_enable(regs);
1281
1282 info.regs = regs;
1283 math_emulate(&info);
1284
1285 cond_local_irq_disable(regs);
1286 return;
1287 }
1288#endif
1289
1290 /* This should not happen. */
1291 if (WARN(cr0 & X86_CR0_TS, "CR0.TS was set")) {
1292 /* Try to fix it up and carry on. */
1293 write_cr0(x: cr0 & ~X86_CR0_TS);
1294 } else {
1295 /*
1296 * Something terrible happened, and we're better off trying
1297 * to kill the task than getting stuck in a never-ending
1298 * loop of #NM faults.
1299 */
1300 die("unexpected #NM exception", regs, 0);
1301 }
1302}
1303
1304#ifdef CONFIG_INTEL_TDX_GUEST
1305
1306#define VE_FAULT_STR "VE fault"
1307
1308static void ve_raise_fault(struct pt_regs *regs, long error_code,
1309 unsigned long address)
1310{
1311 if (user_mode(regs)) {
1312 gp_user_force_sig_segv(regs, X86_TRAP_VE, error_code, VE_FAULT_STR);
1313 return;
1314 }
1315
1316 if (gp_try_fixup_and_notify(regs, X86_TRAP_VE, error_code,
1317 VE_FAULT_STR, address)) {
1318 return;
1319 }
1320
1321 die_addr(VE_FAULT_STR, regs, err: error_code, gp_addr: address);
1322}
1323
1324/*
1325 * Virtualization Exceptions (#VE) are delivered to TDX guests due to
1326 * specific guest actions which may happen in either user space or the
1327 * kernel:
1328 *
1329 * * Specific instructions (WBINVD, for example)
1330 * * Specific MSR accesses
1331 * * Specific CPUID leaf accesses
1332 * * Access to specific guest physical addresses
1333 *
1334 * In the settings that Linux will run in, virtualization exceptions are
1335 * never generated on accesses to normal, TD-private memory that has been
1336 * accepted (by BIOS or with tdx_enc_status_changed()).
1337 *
1338 * Syscall entry code has a critical window where the kernel stack is not
1339 * yet set up. Any exception in this window leads to hard to debug issues
1340 * and can be exploited for privilege escalation. Exceptions in the NMI
1341 * entry code also cause issues. Returning from the exception handler with
1342 * IRET will re-enable NMIs and nested NMI will corrupt the NMI stack.
1343 *
1344 * For these reasons, the kernel avoids #VEs during the syscall gap and
1345 * the NMI entry code. Entry code paths do not access TD-shared memory,
1346 * MMIO regions, use #VE triggering MSRs, instructions, or CPUID leaves
1347 * that might generate #VE. VMM can remove memory from TD at any point,
1348 * but access to unaccepted (or missing) private memory leads to VM
1349 * termination, not to #VE.
1350 *
1351 * Similarly to page faults and breakpoints, #VEs are allowed in NMI
1352 * handlers once the kernel is ready to deal with nested NMIs.
1353 *
1354 * During #VE delivery, all interrupts, including NMIs, are blocked until
1355 * TDGETVEINFO is called. It prevents #VE nesting until the kernel reads
1356 * the VE info.
1357 *
1358 * If a guest kernel action which would normally cause a #VE occurs in
1359 * the interrupt-disabled region before TDGETVEINFO, a #DF (fault
1360 * exception) is delivered to the guest which will result in an oops.
1361 *
1362 * The entry code has been audited carefully for following these expectations.
1363 * Changes in the entry code have to be audited for correctness vs. this
1364 * aspect. Similarly to #PF, #VE in these places will expose kernel to
1365 * privilege escalation or may lead to random crashes.
1366 */
1367DEFINE_IDTENTRY(exc_virtualization_exception)
1368{
1369 struct ve_info ve;
1370
1371 /*
1372 * NMIs/Machine-checks/Interrupts will be in a disabled state
1373 * till TDGETVEINFO TDCALL is executed. This ensures that VE
1374 * info cannot be overwritten by a nested #VE.
1375 */
1376 tdx_get_ve_info(ve: &ve);
1377
1378 cond_local_irq_enable(regs);
1379
1380 /*
1381 * If tdx_handle_virt_exception() could not process
1382 * it successfully, treat it as #GP(0) and handle it.
1383 */
1384 if (!tdx_handle_virt_exception(regs, ve: &ve))
1385 ve_raise_fault(regs, error_code: 0, address: ve.gla);
1386
1387 cond_local_irq_disable(regs);
1388}
1389
1390#endif
1391
1392#ifdef CONFIG_X86_32
1393DEFINE_IDTENTRY_SW(iret_error)
1394{
1395 local_irq_enable();
1396 if (notify_die(DIE_TRAP, "iret exception", regs, 0,
1397 X86_TRAP_IRET, SIGILL) != NOTIFY_STOP) {
1398 do_trap(X86_TRAP_IRET, SIGILL, "iret exception", regs, 0,
1399 ILL_BADSTK, (void __user *)NULL);
1400 }
1401 local_irq_disable();
1402}
1403#endif
1404
1405/* Do not enable FRED by default yet. */
1406static bool enable_fred __ro_after_init = false;
1407
1408#ifdef CONFIG_X86_FRED
1409static int __init fred_setup(char *str)
1410{
1411 if (!str)
1412 return -EINVAL;
1413
1414 if (!cpu_feature_enabled(X86_FEATURE_FRED))
1415 return 0;
1416
1417 if (!strcmp(str, "on"))
1418 enable_fred = true;
1419 else if (!strcmp(str, "off"))
1420 enable_fred = false;
1421 else
1422 pr_warn("invalid FRED option: 'fred=%s'\n", str);
1423 return 0;
1424}
1425early_param("fred", fred_setup);
1426#endif
1427
1428void __init trap_init(void)
1429{
1430 if (cpu_feature_enabled(X86_FEATURE_FRED) && !enable_fred)
1431 setup_clear_cpu_cap(X86_FEATURE_FRED);
1432
1433 /* Init cpu_entry_area before IST entries are set up */
1434 setup_cpu_entry_areas();
1435
1436 /* Init GHCB memory pages when running as an SEV-ES guest */
1437 sev_es_init_vc_handling();
1438
1439 /* Initialize TSS before setting up traps so ISTs work */
1440 cpu_init_exception_handling();
1441
1442 /* Setup traps as cpu_init() might #GP */
1443 if (!cpu_feature_enabled(X86_FEATURE_FRED))
1444 idt_setup_traps();
1445
1446 cpu_init();
1447}
1448

source code of linux/arch/x86/kernel/traps.c