1/* SPDX-License-Identifier: GPL-2.0 */ 2/* 3 * linux/arch/x86_64/entry.S 4 * 5 * Copyright (C) 1991, 1992 Linus Torvalds 6 * Copyright (C) 2000, 2001, 2002 Andi Kleen SuSE Labs 7 * Copyright (C) 2000 Pavel Machek <pavel@suse.cz> 8 * 9 * entry.S contains the system-call and fault low-level handling routines. 10 * 11 * Some of this is documented in Documentation/x86/entry_64.rst 12 * 13 * A note on terminology: 14 * - iret frame: Architecture defined interrupt frame from SS to RIP 15 * at the top of the kernel process stack. 16 * 17 * Some macro usage: 18 * - ENTRY/END: Define functions in the symbol table. 19 * - TRACE_IRQ_*: Trace hardirq state for lock debugging. 20 * - idtentry: Define exception entry points. 21 */ 22#include <linux/linkage.h> 23#include <asm/segment.h> 24#include <asm/cache.h> 25#include <asm/errno.h> 26#include <asm/asm-offsets.h> 27#include <asm/msr.h> 28#include <asm/unistd.h> 29#include <asm/thread_info.h> 30#include <asm/hw_irq.h> 31#include <asm/page_types.h> 32#include <asm/irqflags.h> 33#include <asm/paravirt.h> 34#include <asm/percpu.h> 35#include <asm/asm.h> 36#include <asm/smap.h> 37#include <asm/pgtable_types.h> 38#include <asm/export.h> 39#include <asm/frame.h> 40#include <asm/nospec-branch.h> 41#include <linux/err.h> 42 43#include "calling.h" 44 45.code64 46.section .entry.text, "ax" 47 48#ifdef CONFIG_PARAVIRT 49ENTRY(native_usergs_sysret64) 50 UNWIND_HINT_EMPTY 51 swapgs 52 sysretq 53END(native_usergs_sysret64) 54#endif /* CONFIG_PARAVIRT */ 55 56.macro TRACE_IRQS_FLAGS flags:req 57#ifdef CONFIG_TRACE_IRQFLAGS 58 btl $9, \flags /* interrupts off? */ 59 jnc 1f 60 TRACE_IRQS_ON 611: 62#endif 63.endm 64 65.macro TRACE_IRQS_IRETQ 66 TRACE_IRQS_FLAGS EFLAGS(%rsp) 67.endm 68 69/* 70 * When dynamic function tracer is enabled it will add a breakpoint 71 * to all locations that it is about to modify, sync CPUs, update 72 * all the code, sync CPUs, then remove the breakpoints. In this time 73 * if lockdep is enabled, it might jump back into the debug handler 74 * outside the updating of the IST protection. (TRACE_IRQS_ON/OFF). 75 * 76 * We need to change the IDT table before calling TRACE_IRQS_ON/OFF to 77 * make sure the stack pointer does not get reset back to the top 78 * of the debug stack, and instead just reuses the current stack. 79 */ 80#if defined(CONFIG_DYNAMIC_FTRACE) && defined(CONFIG_TRACE_IRQFLAGS) 81 82.macro TRACE_IRQS_OFF_DEBUG 83 call debug_stack_set_zero 84 TRACE_IRQS_OFF 85 call debug_stack_reset 86.endm 87 88.macro TRACE_IRQS_ON_DEBUG 89 call debug_stack_set_zero 90 TRACE_IRQS_ON 91 call debug_stack_reset 92.endm 93 94.macro TRACE_IRQS_IRETQ_DEBUG 95 btl $9, EFLAGS(%rsp) /* interrupts off? */ 96 jnc 1f 97 TRACE_IRQS_ON_DEBUG 981: 99.endm 100 101#else 102# define TRACE_IRQS_OFF_DEBUG TRACE_IRQS_OFF 103# define TRACE_IRQS_ON_DEBUG TRACE_IRQS_ON 104# define TRACE_IRQS_IRETQ_DEBUG TRACE_IRQS_IRETQ 105#endif 106 107/* 108 * 64-bit SYSCALL instruction entry. Up to 6 arguments in registers. 109 * 110 * This is the only entry point used for 64-bit system calls. The 111 * hardware interface is reasonably well designed and the register to 112 * argument mapping Linux uses fits well with the registers that are 113 * available when SYSCALL is used. 114 * 115 * SYSCALL instructions can be found inlined in libc implementations as 116 * well as some other programs and libraries. There are also a handful 117 * of SYSCALL instructions in the vDSO used, for example, as a 118 * clock_gettimeofday fallback. 119 * 120 * 64-bit SYSCALL saves rip to rcx, clears rflags.RF, then saves rflags to r11, 121 * then loads new ss, cs, and rip from previously programmed MSRs. 122 * rflags gets masked by a value from another MSR (so CLD and CLAC 123 * are not needed). SYSCALL does not save anything on the stack 124 * and does not change rsp. 125 * 126 * Registers on entry: 127 * rax system call number 128 * rcx return address 129 * r11 saved rflags (note: r11 is callee-clobbered register in C ABI) 130 * rdi arg0 131 * rsi arg1 132 * rdx arg2 133 * r10 arg3 (needs to be moved to rcx to conform to C ABI) 134 * r8 arg4 135 * r9 arg5 136 * (note: r12-r15, rbp, rbx are callee-preserved in C ABI) 137 * 138 * Only called from user space. 139 * 140 * When user can change pt_regs->foo always force IRET. That is because 141 * it deals with uncanonical addresses better. SYSRET has trouble 142 * with them due to bugs in both AMD and Intel CPUs. 143 */ 144 145ENTRY(entry_SYSCALL_64) 146 UNWIND_HINT_EMPTY 147 /* 148 * Interrupts are off on entry. 149 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON, 150 * it is too small to ever cause noticeable irq latency. 151 */ 152 153 swapgs 154 /* tss.sp2 is scratch space. */ 155 movq %rsp, PER_CPU_VAR(cpu_tss_rw + TSS_sp2) 156 SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp 157 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 158 159 /* Construct struct pt_regs on stack */ 160 pushq $__USER_DS /* pt_regs->ss */ 161 pushq PER_CPU_VAR(cpu_tss_rw + TSS_sp2) /* pt_regs->sp */ 162 pushq %r11 /* pt_regs->flags */ 163 pushq $__USER_CS /* pt_regs->cs */ 164 pushq %rcx /* pt_regs->ip */ 165GLOBAL(entry_SYSCALL_64_after_hwframe) 166 pushq %rax /* pt_regs->orig_ax */ 167 168 PUSH_AND_CLEAR_REGS rax=$-ENOSYS 169 170 TRACE_IRQS_OFF 171 172 /* IRQs are off. */ 173 movq %rax, %rdi 174 movq %rsp, %rsi 175 call do_syscall_64 /* returns with IRQs disabled */ 176 177 TRACE_IRQS_IRETQ /* we're about to change IF */ 178 179 /* 180 * Try to use SYSRET instead of IRET if we're returning to 181 * a completely clean 64-bit userspace context. If we're not, 182 * go to the slow exit path. 183 */ 184 movq RCX(%rsp), %rcx 185 movq RIP(%rsp), %r11 186 187 cmpq %rcx, %r11 /* SYSRET requires RCX == RIP */ 188 jne swapgs_restore_regs_and_return_to_usermode 189 190 /* 191 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP 192 * in kernel space. This essentially lets the user take over 193 * the kernel, since userspace controls RSP. 194 * 195 * If width of "canonical tail" ever becomes variable, this will need 196 * to be updated to remain correct on both old and new CPUs. 197 * 198 * Change top bits to match most significant bit (47th or 56th bit 199 * depending on paging mode) in the address. 200 */ 201#ifdef CONFIG_X86_5LEVEL 202 ALTERNATIVE "shl $(64 - 48), %rcx; sar $(64 - 48), %rcx", \ 203 "shl $(64 - 57), %rcx; sar $(64 - 57), %rcx", X86_FEATURE_LA57 204#else 205 shl $(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx 206 sar $(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx 207#endif 208 209 /* If this changed %rcx, it was not canonical */ 210 cmpq %rcx, %r11 211 jne swapgs_restore_regs_and_return_to_usermode 212 213 cmpq $__USER_CS, CS(%rsp) /* CS must match SYSRET */ 214 jne swapgs_restore_regs_and_return_to_usermode 215 216 movq R11(%rsp), %r11 217 cmpq %r11, EFLAGS(%rsp) /* R11 == RFLAGS */ 218 jne swapgs_restore_regs_and_return_to_usermode 219 220 /* 221 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot 222 * restore RF properly. If the slowpath sets it for whatever reason, we 223 * need to restore it correctly. 224 * 225 * SYSRET can restore TF, but unlike IRET, restoring TF results in a 226 * trap from userspace immediately after SYSRET. This would cause an 227 * infinite loop whenever #DB happens with register state that satisfies 228 * the opportunistic SYSRET conditions. For example, single-stepping 229 * this user code: 230 * 231 * movq $stuck_here, %rcx 232 * pushfq 233 * popq %r11 234 * stuck_here: 235 * 236 * would never get past 'stuck_here'. 237 */ 238 testq $(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11 239 jnz swapgs_restore_regs_and_return_to_usermode 240 241 /* nothing to check for RSP */ 242 243 cmpq $__USER_DS, SS(%rsp) /* SS must match SYSRET */ 244 jne swapgs_restore_regs_and_return_to_usermode 245 246 /* 247 * We win! This label is here just for ease of understanding 248 * perf profiles. Nothing jumps here. 249 */ 250syscall_return_via_sysret: 251 /* rcx and r11 are already restored (see code above) */ 252 UNWIND_HINT_EMPTY 253 POP_REGS pop_rdi=0 skip_r11rcx=1 254 255 /* 256 * Now all regs are restored except RSP and RDI. 257 * Save old stack pointer and switch to trampoline stack. 258 */ 259 movq %rsp, %rdi 260 movq PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp 261 262 pushq RSP-RDI(%rdi) /* RSP */ 263 pushq (%rdi) /* RDI */ 264 265 /* 266 * We are on the trampoline stack. All regs except RDI are live. 267 * We can do future final exit work right here. 268 */ 269 STACKLEAK_ERASE_NOCLOBBER 270 271 SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi 272 273 popq %rdi 274 popq %rsp 275 USERGS_SYSRET64 276END(entry_SYSCALL_64) 277 278/* 279 * %rdi: prev task 280 * %rsi: next task 281 */ 282ENTRY(__switch_to_asm) 283 UNWIND_HINT_FUNC 284 /* 285 * Save callee-saved registers 286 * This must match the order in inactive_task_frame 287 */ 288 pushq %rbp 289 pushq %rbx 290 pushq %r12 291 pushq %r13 292 pushq %r14 293 pushq %r15 294 295 /* switch stack */ 296 movq %rsp, TASK_threadsp(%rdi) 297 movq TASK_threadsp(%rsi), %rsp 298 299#ifdef CONFIG_STACKPROTECTOR 300 movq TASK_stack_canary(%rsi), %rbx 301 movq %rbx, PER_CPU_VAR(fixed_percpu_data) + stack_canary_offset 302#endif 303 304#ifdef CONFIG_RETPOLINE 305 /* 306 * When switching from a shallower to a deeper call stack 307 * the RSB may either underflow or use entries populated 308 * with userspace addresses. On CPUs where those concerns 309 * exist, overwrite the RSB with entries which capture 310 * speculative execution to prevent attack. 311 */ 312 FILL_RETURN_BUFFER %r12, RSB_CLEAR_LOOPS, X86_FEATURE_RSB_CTXSW 313#endif 314 315 /* restore callee-saved registers */ 316 popq %r15 317 popq %r14 318 popq %r13 319 popq %r12 320 popq %rbx 321 popq %rbp 322 323 jmp __switch_to 324END(__switch_to_asm) 325 326/* 327 * A newly forked process directly context switches into this address. 328 * 329 * rax: prev task we switched from 330 * rbx: kernel thread func (NULL for user thread) 331 * r12: kernel thread arg 332 */ 333ENTRY(ret_from_fork) 334 UNWIND_HINT_EMPTY 335 movq %rax, %rdi 336 call schedule_tail /* rdi: 'prev' task parameter */ 337 338 testq %rbx, %rbx /* from kernel_thread? */ 339 jnz 1f /* kernel threads are uncommon */ 340 3412: 342 UNWIND_HINT_REGS 343 movq %rsp, %rdi 344 call syscall_return_slowpath /* returns with IRQs disabled */ 345 TRACE_IRQS_ON /* user mode is traced as IRQS on */ 346 jmp swapgs_restore_regs_and_return_to_usermode 347 3481: 349 /* kernel thread */ 350 UNWIND_HINT_EMPTY 351 movq %r12, %rdi 352 CALL_NOSPEC %rbx 353 /* 354 * A kernel thread is allowed to return here after successfully 355 * calling do_execve(). Exit to userspace to complete the execve() 356 * syscall. 357 */ 358 movq $0, RAX(%rsp) 359 jmp 2b 360END(ret_from_fork) 361 362/* 363 * Build the entry stubs with some assembler magic. 364 * We pack 1 stub into every 8-byte block. 365 */ 366 .align 8 367ENTRY(irq_entries_start) 368 vector=FIRST_EXTERNAL_VECTOR 369 .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR) 370 UNWIND_HINT_IRET_REGS 371 pushq $(~vector+0x80) /* Note: always in signed byte range */ 372 jmp common_interrupt 373 .align 8 374 vector=vector+1 375 .endr 376END(irq_entries_start) 377 378 .align 8 379ENTRY(spurious_entries_start) 380 vector=FIRST_SYSTEM_VECTOR 381 .rept (NR_VECTORS - FIRST_SYSTEM_VECTOR) 382 UNWIND_HINT_IRET_REGS 383 pushq $(~vector+0x80) /* Note: always in signed byte range */ 384 jmp common_spurious 385 .align 8 386 vector=vector+1 387 .endr 388END(spurious_entries_start) 389 390.macro DEBUG_ENTRY_ASSERT_IRQS_OFF 391#ifdef CONFIG_DEBUG_ENTRY 392 pushq %rax 393 SAVE_FLAGS(CLBR_RAX) 394 testl $X86_EFLAGS_IF, %eax 395 jz .Lokay_\@ 396 ud2 397.Lokay_\@: 398 popq %rax 399#endif 400.endm 401 402/* 403 * Enters the IRQ stack if we're not already using it. NMI-safe. Clobbers 404 * flags and puts old RSP into old_rsp, and leaves all other GPRs alone. 405 * Requires kernel GSBASE. 406 * 407 * The invariant is that, if irq_count != -1, then the IRQ stack is in use. 408 */ 409.macro ENTER_IRQ_STACK regs=1 old_rsp save_ret=0 410 DEBUG_ENTRY_ASSERT_IRQS_OFF 411 412 .if \save_ret 413 /* 414 * If save_ret is set, the original stack contains one additional 415 * entry -- the return address. Therefore, move the address one 416 * entry below %rsp to \old_rsp. 417 */ 418 leaq 8(%rsp), \old_rsp 419 .else 420 movq %rsp, \old_rsp 421 .endif 422 423 .if \regs 424 UNWIND_HINT_REGS base=\old_rsp 425 .endif 426 427 incl PER_CPU_VAR(irq_count) 428 jnz .Lirq_stack_push_old_rsp_\@ 429 430 /* 431 * Right now, if we just incremented irq_count to zero, we've 432 * claimed the IRQ stack but we haven't switched to it yet. 433 * 434 * If anything is added that can interrupt us here without using IST, 435 * it must be *extremely* careful to limit its stack usage. This 436 * could include kprobes and a hypothetical future IST-less #DB 437 * handler. 438 * 439 * The OOPS unwinder relies on the word at the top of the IRQ 440 * stack linking back to the previous RSP for the entire time we're 441 * on the IRQ stack. For this to work reliably, we need to write 442 * it before we actually move ourselves to the IRQ stack. 443 */ 444 445 movq \old_rsp, PER_CPU_VAR(irq_stack_backing_store + IRQ_STACK_SIZE - 8) 446 movq PER_CPU_VAR(hardirq_stack_ptr), %rsp 447 448#ifdef CONFIG_DEBUG_ENTRY 449 /* 450 * If the first movq above becomes wrong due to IRQ stack layout 451 * changes, the only way we'll notice is if we try to unwind right 452 * here. Assert that we set up the stack right to catch this type 453 * of bug quickly. 454 */ 455 cmpq -8(%rsp), \old_rsp 456 je .Lirq_stack_okay\@ 457 ud2 458 .Lirq_stack_okay\@: 459#endif 460 461.Lirq_stack_push_old_rsp_\@: 462 pushq \old_rsp 463 464 .if \regs 465 UNWIND_HINT_REGS indirect=1 466 .endif 467 468 .if \save_ret 469 /* 470 * Push the return address to the stack. This return address can 471 * be found at the "real" original RSP, which was offset by 8 at 472 * the beginning of this macro. 473 */ 474 pushq -8(\old_rsp) 475 .endif 476.endm 477 478/* 479 * Undoes ENTER_IRQ_STACK. 480 */ 481.macro LEAVE_IRQ_STACK regs=1 482 DEBUG_ENTRY_ASSERT_IRQS_OFF 483 /* We need to be off the IRQ stack before decrementing irq_count. */ 484 popq %rsp 485 486 .if \regs 487 UNWIND_HINT_REGS 488 .endif 489 490 /* 491 * As in ENTER_IRQ_STACK, irq_count == 0, we are still claiming 492 * the irq stack but we're not on it. 493 */ 494 495 decl PER_CPU_VAR(irq_count) 496.endm 497 498/* 499 * Interrupt entry helper function. 500 * 501 * Entry runs with interrupts off. Stack layout at entry: 502 * +----------------------------------------------------+ 503 * | regs->ss | 504 * | regs->rsp | 505 * | regs->eflags | 506 * | regs->cs | 507 * | regs->ip | 508 * +----------------------------------------------------+ 509 * | regs->orig_ax = ~(interrupt number) | 510 * +----------------------------------------------------+ 511 * | return address | 512 * +----------------------------------------------------+ 513 */ 514ENTRY(interrupt_entry) 515 UNWIND_HINT_FUNC 516 ASM_CLAC 517 cld 518 519 testb $3, CS-ORIG_RAX+8(%rsp) 520 jz 1f 521 SWAPGS 522 523 /* 524 * Switch to the thread stack. The IRET frame and orig_ax are 525 * on the stack, as well as the return address. RDI..R12 are 526 * not (yet) on the stack and space has not (yet) been 527 * allocated for them. 528 */ 529 pushq %rdi 530 531 /* Need to switch before accessing the thread stack. */ 532 SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi 533 movq %rsp, %rdi 534 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 535 536 /* 537 * We have RDI, return address, and orig_ax on the stack on 538 * top of the IRET frame. That means offset=24 539 */ 540 UNWIND_HINT_IRET_REGS base=%rdi offset=24 541 542 pushq 7*8(%rdi) /* regs->ss */ 543 pushq 6*8(%rdi) /* regs->rsp */ 544 pushq 5*8(%rdi) /* regs->eflags */ 545 pushq 4*8(%rdi) /* regs->cs */ 546 pushq 3*8(%rdi) /* regs->ip */ 547 pushq 2*8(%rdi) /* regs->orig_ax */ 548 pushq 8(%rdi) /* return address */ 549 UNWIND_HINT_FUNC 550 551 movq (%rdi), %rdi 5521: 553 554 PUSH_AND_CLEAR_REGS save_ret=1 555 ENCODE_FRAME_POINTER 8 556 557 testb $3, CS+8(%rsp) 558 jz 1f 559 560 /* 561 * IRQ from user mode. 562 * 563 * We need to tell lockdep that IRQs are off. We can't do this until 564 * we fix gsbase, and we should do it before enter_from_user_mode 565 * (which can take locks). Since TRACE_IRQS_OFF is idempotent, 566 * the simplest way to handle it is to just call it twice if 567 * we enter from user mode. There's no reason to optimize this since 568 * TRACE_IRQS_OFF is a no-op if lockdep is off. 569 */ 570 TRACE_IRQS_OFF 571 572 CALL_enter_from_user_mode 573 5741: 575 ENTER_IRQ_STACK old_rsp=%rdi save_ret=1 576 /* We entered an interrupt context - irqs are off: */ 577 TRACE_IRQS_OFF 578 579 ret 580END(interrupt_entry) 581_ASM_NOKPROBE(interrupt_entry) 582 583 584/* Interrupt entry/exit. */ 585 586/* 587 * The interrupt stubs push (~vector+0x80) onto the stack and 588 * then jump to common_spurious/interrupt. 589 */ 590common_spurious: 591 addq $-0x80, (%rsp) /* Adjust vector to [-256, -1] range */ 592 call interrupt_entry 593 UNWIND_HINT_REGS indirect=1 594 call smp_spurious_interrupt /* rdi points to pt_regs */ 595 jmp ret_from_intr 596END(common_spurious) 597_ASM_NOKPROBE(common_spurious) 598 599/* common_interrupt is a hotpath. Align it */ 600 .p2align CONFIG_X86_L1_CACHE_SHIFT 601common_interrupt: 602 addq $-0x80, (%rsp) /* Adjust vector to [-256, -1] range */ 603 call interrupt_entry 604 UNWIND_HINT_REGS indirect=1 605 call do_IRQ /* rdi points to pt_regs */ 606 /* 0(%rsp): old RSP */ 607ret_from_intr: 608 DISABLE_INTERRUPTS(CLBR_ANY) 609 TRACE_IRQS_OFF 610 611 LEAVE_IRQ_STACK 612 613 testb $3, CS(%rsp) 614 jz retint_kernel 615 616 /* Interrupt came from user space */ 617GLOBAL(retint_user) 618 mov %rsp,%rdi 619 call prepare_exit_to_usermode 620 TRACE_IRQS_IRETQ 621 622GLOBAL(swapgs_restore_regs_and_return_to_usermode) 623#ifdef CONFIG_DEBUG_ENTRY 624 /* Assert that pt_regs indicates user mode. */ 625 testb $3, CS(%rsp) 626 jnz 1f 627 ud2 6281: 629#endif 630 POP_REGS pop_rdi=0 631 632 /* 633 * The stack is now user RDI, orig_ax, RIP, CS, EFLAGS, RSP, SS. 634 * Save old stack pointer and switch to trampoline stack. 635 */ 636 movq %rsp, %rdi 637 movq PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp 638 639 /* Copy the IRET frame to the trampoline stack. */ 640 pushq 6*8(%rdi) /* SS */ 641 pushq 5*8(%rdi) /* RSP */ 642 pushq 4*8(%rdi) /* EFLAGS */ 643 pushq 3*8(%rdi) /* CS */ 644 pushq 2*8(%rdi) /* RIP */ 645 646 /* Push user RDI on the trampoline stack. */ 647 pushq (%rdi) 648 649 /* 650 * We are on the trampoline stack. All regs except RDI are live. 651 * We can do future final exit work right here. 652 */ 653 STACKLEAK_ERASE_NOCLOBBER 654 655 SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi 656 657 /* Restore RDI. */ 658 popq %rdi 659 SWAPGS 660 INTERRUPT_RETURN 661 662 663/* Returning to kernel space */ 664retint_kernel: 665#ifdef CONFIG_PREEMPT 666 /* Interrupts are off */ 667 /* Check if we need preemption */ 668 btl $9, EFLAGS(%rsp) /* were interrupts off? */ 669 jnc 1f 670 cmpl $0, PER_CPU_VAR(__preempt_count) 671 jnz 1f 672 call preempt_schedule_irq 6731: 674#endif 675 /* 676 * The iretq could re-enable interrupts: 677 */ 678 TRACE_IRQS_IRETQ 679 680GLOBAL(restore_regs_and_return_to_kernel) 681#ifdef CONFIG_DEBUG_ENTRY 682 /* Assert that pt_regs indicates kernel mode. */ 683 testb $3, CS(%rsp) 684 jz 1f 685 ud2 6861: 687#endif 688 POP_REGS 689 addq $8, %rsp /* skip regs->orig_ax */ 690 /* 691 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization 692 * when returning from IPI handler. 693 */ 694 INTERRUPT_RETURN 695 696ENTRY(native_iret) 697 UNWIND_HINT_IRET_REGS 698 /* 699 * Are we returning to a stack segment from the LDT? Note: in 700 * 64-bit mode SS:RSP on the exception stack is always valid. 701 */ 702#ifdef CONFIG_X86_ESPFIX64 703 testb $4, (SS-RIP)(%rsp) 704 jnz native_irq_return_ldt 705#endif 706 707.global native_irq_return_iret 708native_irq_return_iret: 709 /* 710 * This may fault. Non-paranoid faults on return to userspace are 711 * handled by fixup_bad_iret. These include #SS, #GP, and #NP. 712 * Double-faults due to espfix64 are handled in do_double_fault. 713 * Other faults here are fatal. 714 */ 715 iretq 716 717#ifdef CONFIG_X86_ESPFIX64 718native_irq_return_ldt: 719 /* 720 * We are running with user GSBASE. All GPRs contain their user 721 * values. We have a percpu ESPFIX stack that is eight slots 722 * long (see ESPFIX_STACK_SIZE). espfix_waddr points to the bottom 723 * of the ESPFIX stack. 724 * 725 * We clobber RAX and RDI in this code. We stash RDI on the 726 * normal stack and RAX on the ESPFIX stack. 727 * 728 * The ESPFIX stack layout we set up looks like this: 729 * 730 * --- top of ESPFIX stack --- 731 * SS 732 * RSP 733 * RFLAGS 734 * CS 735 * RIP <-- RSP points here when we're done 736 * RAX <-- espfix_waddr points here 737 * --- bottom of ESPFIX stack --- 738 */ 739 740 pushq %rdi /* Stash user RDI */ 741 SWAPGS /* to kernel GS */ 742 SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi /* to kernel CR3 */ 743 744 movq PER_CPU_VAR(espfix_waddr), %rdi 745 movq %rax, (0*8)(%rdi) /* user RAX */ 746 movq (1*8)(%rsp), %rax /* user RIP */ 747 movq %rax, (1*8)(%rdi) 748 movq (2*8)(%rsp), %rax /* user CS */ 749 movq %rax, (2*8)(%rdi) 750 movq (3*8)(%rsp), %rax /* user RFLAGS */ 751 movq %rax, (3*8)(%rdi) 752 movq (5*8)(%rsp), %rax /* user SS */ 753 movq %rax, (5*8)(%rdi) 754 movq (4*8)(%rsp), %rax /* user RSP */ 755 movq %rax, (4*8)(%rdi) 756 /* Now RAX == RSP. */ 757 758 andl $0xffff0000, %eax /* RAX = (RSP & 0xffff0000) */ 759 760 /* 761 * espfix_stack[31:16] == 0. The page tables are set up such that 762 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of 763 * espfix_waddr for any X. That is, there are 65536 RO aliases of 764 * the same page. Set up RSP so that RSP[31:16] contains the 765 * respective 16 bits of the /userspace/ RSP and RSP nonetheless 766 * still points to an RO alias of the ESPFIX stack. 767 */ 768 orq PER_CPU_VAR(espfix_stack), %rax 769 770 SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi 771 SWAPGS /* to user GS */ 772 popq %rdi /* Restore user RDI */ 773 774 movq %rax, %rsp 775 UNWIND_HINT_IRET_REGS offset=8 776 777 /* 778 * At this point, we cannot write to the stack any more, but we can 779 * still read. 780 */ 781 popq %rax /* Restore user RAX */ 782 783 /* 784 * RSP now points to an ordinary IRET frame, except that the page 785 * is read-only and RSP[31:16] are preloaded with the userspace 786 * values. We can now IRET back to userspace. 787 */ 788 jmp native_irq_return_iret 789#endif 790END(common_interrupt) 791_ASM_NOKPROBE(common_interrupt) 792 793/* 794 * APIC interrupts. 795 */ 796.macro apicinterrupt3 num sym do_sym 797ENTRY(\sym) 798 UNWIND_HINT_IRET_REGS 799 pushq $~(\num) 800.Lcommon_\sym: 801 call interrupt_entry 802 UNWIND_HINT_REGS indirect=1 803 call \do_sym /* rdi points to pt_regs */ 804 jmp ret_from_intr 805END(\sym) 806_ASM_NOKPROBE(\sym) 807.endm 808 809/* Make sure APIC interrupt handlers end up in the irqentry section: */ 810#define PUSH_SECTION_IRQENTRY .pushsection .irqentry.text, "ax" 811#define POP_SECTION_IRQENTRY .popsection 812 813.macro apicinterrupt num sym do_sym 814PUSH_SECTION_IRQENTRY 815apicinterrupt3 \num \sym \do_sym 816POP_SECTION_IRQENTRY 817.endm 818 819#ifdef CONFIG_SMP 820apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR irq_move_cleanup_interrupt smp_irq_move_cleanup_interrupt 821apicinterrupt3 REBOOT_VECTOR reboot_interrupt smp_reboot_interrupt 822#endif 823 824#ifdef CONFIG_X86_UV 825apicinterrupt3 UV_BAU_MESSAGE uv_bau_message_intr1 uv_bau_message_interrupt 826#endif 827 828apicinterrupt LOCAL_TIMER_VECTOR apic_timer_interrupt smp_apic_timer_interrupt 829apicinterrupt X86_PLATFORM_IPI_VECTOR x86_platform_ipi smp_x86_platform_ipi 830 831#ifdef CONFIG_HAVE_KVM 832apicinterrupt3 POSTED_INTR_VECTOR kvm_posted_intr_ipi smp_kvm_posted_intr_ipi 833apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR kvm_posted_intr_wakeup_ipi smp_kvm_posted_intr_wakeup_ipi 834apicinterrupt3 POSTED_INTR_NESTED_VECTOR kvm_posted_intr_nested_ipi smp_kvm_posted_intr_nested_ipi 835#endif 836 837#ifdef CONFIG_X86_MCE_THRESHOLD 838apicinterrupt THRESHOLD_APIC_VECTOR threshold_interrupt smp_threshold_interrupt 839#endif 840 841#ifdef CONFIG_X86_MCE_AMD 842apicinterrupt DEFERRED_ERROR_VECTOR deferred_error_interrupt smp_deferred_error_interrupt 843#endif 844 845#ifdef CONFIG_X86_THERMAL_VECTOR 846apicinterrupt THERMAL_APIC_VECTOR thermal_interrupt smp_thermal_interrupt 847#endif 848 849#ifdef CONFIG_SMP 850apicinterrupt CALL_FUNCTION_SINGLE_VECTOR call_function_single_interrupt smp_call_function_single_interrupt 851apicinterrupt CALL_FUNCTION_VECTOR call_function_interrupt smp_call_function_interrupt 852apicinterrupt RESCHEDULE_VECTOR reschedule_interrupt smp_reschedule_interrupt 853#endif 854 855apicinterrupt ERROR_APIC_VECTOR error_interrupt smp_error_interrupt 856apicinterrupt SPURIOUS_APIC_VECTOR spurious_interrupt smp_spurious_interrupt 857 858#ifdef CONFIG_IRQ_WORK 859apicinterrupt IRQ_WORK_VECTOR irq_work_interrupt smp_irq_work_interrupt 860#endif 861 862/* 863 * Exception entry points. 864 */ 865#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss_rw) + (TSS_ist + (x) * 8) 866 867/** 868 * idtentry - Generate an IDT entry stub 869 * @sym: Name of the generated entry point 870 * @do_sym: C function to be called 871 * @has_error_code: True if this IDT vector has an error code on the stack 872 * @paranoid: non-zero means that this vector may be invoked from 873 * kernel mode with user GSBASE and/or user CR3. 874 * 2 is special -- see below. 875 * @shift_ist: Set to an IST index if entries from kernel mode should 876 * decrement the IST stack so that nested entries get a 877 * fresh stack. (This is for #DB, which has a nasty habit 878 * of recursing.) 879 * 880 * idtentry generates an IDT stub that sets up a usable kernel context, 881 * creates struct pt_regs, and calls @do_sym. The stub has the following 882 * special behaviors: 883 * 884 * On an entry from user mode, the stub switches from the trampoline or 885 * IST stack to the normal thread stack. On an exit to user mode, the 886 * normal exit-to-usermode path is invoked. 887 * 888 * On an exit to kernel mode, if @paranoid == 0, we check for preemption, 889 * whereas we omit the preemption check if @paranoid != 0. This is purely 890 * because the implementation is simpler this way. The kernel only needs 891 * to check for asynchronous kernel preemption when IRQ handlers return. 892 * 893 * If @paranoid == 0, then the stub will handle IRET faults by pretending 894 * that the fault came from user mode. It will handle gs_change faults by 895 * pretending that the fault happened with kernel GSBASE. Since this handling 896 * is omitted for @paranoid != 0, the #GP, #SS, and #NP stubs must have 897 * @paranoid == 0. This special handling will do the wrong thing for 898 * espfix-induced #DF on IRET, so #DF must not use @paranoid == 0. 899 * 900 * @paranoid == 2 is special: the stub will never switch stacks. This is for 901 * #DF: if the thread stack is somehow unusable, we'll still get a useful OOPS. 902 */ 903.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1 ist_offset=0 create_gap=0 904ENTRY(\sym) 905 UNWIND_HINT_IRET_REGS offset=\has_error_code*8 906 907 /* Sanity check */ 908 .if \shift_ist != -1 && \paranoid == 0 909 .error "using shift_ist requires paranoid=1" 910 .endif 911 912 ASM_CLAC 913 914 .if \has_error_code == 0 915 pushq $-1 /* ORIG_RAX: no syscall to restart */ 916 .endif 917 918 .if \paranoid == 1 919 testb $3, CS-ORIG_RAX(%rsp) /* If coming from userspace, switch stacks */ 920 jnz .Lfrom_usermode_switch_stack_\@ 921 .endif 922 923 .if \create_gap == 1 924 /* 925 * If coming from kernel space, create a 6-word gap to allow the 926 * int3 handler to emulate a call instruction. 927 */ 928 testb $3, CS-ORIG_RAX(%rsp) 929 jnz .Lfrom_usermode_no_gap_\@ 930 .rept 6 931 pushq 5*8(%rsp) 932 .endr 933 UNWIND_HINT_IRET_REGS offset=8 934.Lfrom_usermode_no_gap_\@: 935 .endif 936 937 .if \paranoid 938 call paranoid_entry 939 .else 940 call error_entry 941 .endif 942 UNWIND_HINT_REGS 943 /* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */ 944 945 .if \paranoid 946 .if \shift_ist != -1 947 TRACE_IRQS_OFF_DEBUG /* reload IDT in case of recursion */ 948 .else 949 TRACE_IRQS_OFF 950 .endif 951 .endif 952 953 movq %rsp, %rdi /* pt_regs pointer */ 954 955 .if \has_error_code 956 movq ORIG_RAX(%rsp), %rsi /* get error code */ 957 movq $-1, ORIG_RAX(%rsp) /* no syscall to restart */ 958 .else 959 xorl %esi, %esi /* no error code */ 960 .endif 961 962 .if \shift_ist != -1 963 subq $\ist_offset, CPU_TSS_IST(\shift_ist) 964 .endif 965 966 call \do_sym 967 968 .if \shift_ist != -1 969 addq $\ist_offset, CPU_TSS_IST(\shift_ist) 970 .endif 971 972 /* these procedures expect "no swapgs" flag in ebx */ 973 .if \paranoid 974 jmp paranoid_exit 975 .else 976 jmp error_exit 977 .endif 978 979 .if \paranoid == 1 980 /* 981 * Entry from userspace. Switch stacks and treat it 982 * as a normal entry. This means that paranoid handlers 983 * run in real process context if user_mode(regs). 984 */ 985.Lfrom_usermode_switch_stack_\@: 986 call error_entry 987 988 movq %rsp, %rdi /* pt_regs pointer */ 989 990 .if \has_error_code 991 movq ORIG_RAX(%rsp), %rsi /* get error code */ 992 movq $-1, ORIG_RAX(%rsp) /* no syscall to restart */ 993 .else 994 xorl %esi, %esi /* no error code */ 995 .endif 996 997 call \do_sym 998 999 jmp error_exit 1000 .endif 1001_ASM_NOKPROBE(\sym) 1002END(\sym) 1003.endm 1004 1005idtentry divide_error do_divide_error has_error_code=0 1006idtentry overflow do_overflow has_error_code=0 1007idtentry bounds do_bounds has_error_code=0 1008idtentry invalid_op do_invalid_op has_error_code=0 1009idtentry device_not_available do_device_not_available has_error_code=0 1010idtentry double_fault do_double_fault has_error_code=1 paranoid=2 1011idtentry coprocessor_segment_overrun do_coprocessor_segment_overrun has_error_code=0 1012idtentry invalid_TSS do_invalid_TSS has_error_code=1 1013idtentry segment_not_present do_segment_not_present has_error_code=1 1014idtentry spurious_interrupt_bug do_spurious_interrupt_bug has_error_code=0 1015idtentry coprocessor_error do_coprocessor_error has_error_code=0 1016idtentry alignment_check do_alignment_check has_error_code=1 1017idtentry simd_coprocessor_error do_simd_coprocessor_error has_error_code=0 1018 1019 1020 /* 1021 * Reload gs selector with exception handling 1022 * edi: new selector 1023 */ 1024ENTRY(native_load_gs_index) 1025 FRAME_BEGIN 1026 pushfq 1027 DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI) 1028 TRACE_IRQS_OFF 1029 SWAPGS 1030.Lgs_change: 1031 movl %edi, %gs 10322: ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE 1033 SWAPGS 1034 TRACE_IRQS_FLAGS (%rsp) 1035 popfq 1036 FRAME_END 1037 ret 1038ENDPROC(native_load_gs_index) 1039EXPORT_SYMBOL(native_load_gs_index) 1040 1041 _ASM_EXTABLE(.Lgs_change, bad_gs) 1042 .section .fixup, "ax" 1043 /* running with kernelgs */ 1044bad_gs: 1045 SWAPGS /* switch back to user gs */ 1046.macro ZAP_GS 1047 /* This can't be a string because the preprocessor needs to see it. */ 1048 movl $__USER_DS, %eax 1049 movl %eax, %gs 1050.endm 1051 ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG 1052 xorl %eax, %eax 1053 movl %eax, %gs 1054 jmp 2b 1055 .previous 1056 1057/* Call softirq on interrupt stack. Interrupts are off. */ 1058ENTRY(do_softirq_own_stack) 1059 pushq %rbp 1060 mov %rsp, %rbp 1061 ENTER_IRQ_STACK regs=0 old_rsp=%r11 1062 call __do_softirq 1063 LEAVE_IRQ_STACK regs=0 1064 leaveq 1065 ret 1066ENDPROC(do_softirq_own_stack) 1067 1068#ifdef CONFIG_XEN_PV 1069idtentry hypervisor_callback xen_do_hypervisor_callback has_error_code=0 1070 1071/* 1072 * A note on the "critical region" in our callback handler. 1073 * We want to avoid stacking callback handlers due to events occurring 1074 * during handling of the last event. To do this, we keep events disabled 1075 * until we've done all processing. HOWEVER, we must enable events before 1076 * popping the stack frame (can't be done atomically) and so it would still 1077 * be possible to get enough handler activations to overflow the stack. 1078 * Although unlikely, bugs of that kind are hard to track down, so we'd 1079 * like to avoid the possibility. 1080 * So, on entry to the handler we detect whether we interrupted an 1081 * existing activation in its critical region -- if so, we pop the current 1082 * activation and restart the handler using the previous one. 1083 */ 1084ENTRY(xen_do_hypervisor_callback) /* do_hypervisor_callback(struct *pt_regs) */ 1085 1086/* 1087 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will 1088 * see the correct pointer to the pt_regs 1089 */ 1090 UNWIND_HINT_FUNC 1091 movq %rdi, %rsp /* we don't return, adjust the stack frame */ 1092 UNWIND_HINT_REGS 1093 1094 ENTER_IRQ_STACK old_rsp=%r10 1095 call xen_evtchn_do_upcall 1096 LEAVE_IRQ_STACK 1097 1098#ifndef CONFIG_PREEMPT 1099 call xen_maybe_preempt_hcall 1100#endif 1101 jmp error_exit 1102END(xen_do_hypervisor_callback) 1103 1104/* 1105 * Hypervisor uses this for application faults while it executes. 1106 * We get here for two reasons: 1107 * 1. Fault while reloading DS, ES, FS or GS 1108 * 2. Fault while executing IRET 1109 * Category 1 we do not need to fix up as Xen has already reloaded all segment 1110 * registers that could be reloaded and zeroed the others. 1111 * Category 2 we fix up by killing the current process. We cannot use the 1112 * normal Linux return path in this case because if we use the IRET hypercall 1113 * to pop the stack frame we end up in an infinite loop of failsafe callbacks. 1114 * We distinguish between categories by comparing each saved segment register 1115 * with its current contents: any discrepancy means we in category 1. 1116 */ 1117ENTRY(xen_failsafe_callback) 1118 UNWIND_HINT_EMPTY 1119 movl %ds, %ecx 1120 cmpw %cx, 0x10(%rsp) 1121 jne 1f 1122 movl %es, %ecx 1123 cmpw %cx, 0x18(%rsp) 1124 jne 1f 1125 movl %fs, %ecx 1126 cmpw %cx, 0x20(%rsp) 1127 jne 1f 1128 movl %gs, %ecx 1129 cmpw %cx, 0x28(%rsp) 1130 jne 1f 1131 /* All segments match their saved values => Category 2 (Bad IRET). */ 1132 movq (%rsp), %rcx 1133 movq 8(%rsp), %r11 1134 addq $0x30, %rsp 1135 pushq $0 /* RIP */ 1136 UNWIND_HINT_IRET_REGS offset=8 1137 jmp general_protection 11381: /* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */ 1139 movq (%rsp), %rcx 1140 movq 8(%rsp), %r11 1141 addq $0x30, %rsp 1142 UNWIND_HINT_IRET_REGS 1143 pushq $-1 /* orig_ax = -1 => not a system call */ 1144 PUSH_AND_CLEAR_REGS 1145 ENCODE_FRAME_POINTER 1146 jmp error_exit 1147END(xen_failsafe_callback) 1148#endif /* CONFIG_XEN_PV */ 1149 1150#ifdef CONFIG_XEN_PVHVM 1151apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \ 1152 xen_hvm_callback_vector xen_evtchn_do_upcall 1153#endif 1154 1155 1156#if IS_ENABLED(CONFIG_HYPERV) 1157apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \ 1158 hyperv_callback_vector hyperv_vector_handler 1159 1160apicinterrupt3 HYPERV_REENLIGHTENMENT_VECTOR \ 1161 hyperv_reenlightenment_vector hyperv_reenlightenment_intr 1162 1163apicinterrupt3 HYPERV_STIMER0_VECTOR \ 1164 hv_stimer0_callback_vector hv_stimer0_vector_handler 1165#endif /* CONFIG_HYPERV */ 1166 1167#if IS_ENABLED(CONFIG_ACRN_GUEST) 1168apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \ 1169 acrn_hv_callback_vector acrn_hv_vector_handler 1170#endif 1171 1172idtentry debug do_debug has_error_code=0 paranoid=1 shift_ist=IST_INDEX_DB ist_offset=DB_STACK_OFFSET 1173idtentry int3 do_int3 has_error_code=0 create_gap=1 1174idtentry stack_segment do_stack_segment has_error_code=1 1175 1176#ifdef CONFIG_XEN_PV 1177idtentry xennmi do_nmi has_error_code=0 1178idtentry xendebug do_debug has_error_code=0 1179idtentry xenint3 do_int3 has_error_code=0 1180#endif 1181 1182idtentry general_protection do_general_protection has_error_code=1 1183idtentry page_fault do_page_fault has_error_code=1 1184 1185#ifdef CONFIG_KVM_GUEST 1186idtentry async_page_fault do_async_page_fault has_error_code=1 1187#endif 1188 1189#ifdef CONFIG_X86_MCE 1190idtentry machine_check do_mce has_error_code=0 paranoid=1 1191#endif 1192 1193/* 1194 * Save all registers in pt_regs, and switch gs if needed. 1195 * Use slow, but surefire "are we in kernel?" check. 1196 * Return: ebx=0: need swapgs on exit, ebx=1: otherwise 1197 */ 1198ENTRY(paranoid_entry) 1199 UNWIND_HINT_FUNC 1200 cld 1201 PUSH_AND_CLEAR_REGS save_ret=1 1202 ENCODE_FRAME_POINTER 8 1203 movl $1, %ebx 1204 movl $MSR_GS_BASE, %ecx 1205 rdmsr 1206 testl %edx, %edx 1207 js 1f /* negative -> in kernel */ 1208 SWAPGS 1209 xorl %ebx, %ebx 1210 12111: 1212 /* 1213 * Always stash CR3 in %r14. This value will be restored, 1214 * verbatim, at exit. Needed if paranoid_entry interrupted 1215 * another entry that already switched to the user CR3 value 1216 * but has not yet returned to userspace. 1217 * 1218 * This is also why CS (stashed in the "iret frame" by the 1219 * hardware at entry) can not be used: this may be a return 1220 * to kernel code, but with a user CR3 value. 1221 */ 1222 SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14 1223 1224 ret 1225END(paranoid_entry) 1226 1227/* 1228 * "Paranoid" exit path from exception stack. This is invoked 1229 * only on return from non-NMI IST interrupts that came 1230 * from kernel space. 1231 * 1232 * We may be returning to very strange contexts (e.g. very early 1233 * in syscall entry), so checking for preemption here would 1234 * be complicated. Fortunately, we there's no good reason 1235 * to try to handle preemption here. 1236 * 1237 * On entry, ebx is "no swapgs" flag (1: don't need swapgs, 0: need it) 1238 */ 1239ENTRY(paranoid_exit) 1240 UNWIND_HINT_REGS 1241 DISABLE_INTERRUPTS(CLBR_ANY) 1242 TRACE_IRQS_OFF_DEBUG 1243 testl %ebx, %ebx /* swapgs needed? */ 1244 jnz .Lparanoid_exit_no_swapgs 1245 TRACE_IRQS_IRETQ 1246 /* Always restore stashed CR3 value (see paranoid_entry) */ 1247 RESTORE_CR3 scratch_reg=%rbx save_reg=%r14 1248 SWAPGS_UNSAFE_STACK 1249 jmp .Lparanoid_exit_restore 1250.Lparanoid_exit_no_swapgs: 1251 TRACE_IRQS_IRETQ_DEBUG 1252 /* Always restore stashed CR3 value (see paranoid_entry) */ 1253 RESTORE_CR3 scratch_reg=%rbx save_reg=%r14 1254.Lparanoid_exit_restore: 1255 jmp restore_regs_and_return_to_kernel 1256END(paranoid_exit) 1257 1258/* 1259 * Save all registers in pt_regs, and switch GS if needed. 1260 */ 1261ENTRY(error_entry) 1262 UNWIND_HINT_FUNC 1263 cld 1264 PUSH_AND_CLEAR_REGS save_ret=1 1265 ENCODE_FRAME_POINTER 8 1266 testb $3, CS+8(%rsp) 1267 jz .Lerror_kernelspace 1268 1269 /* 1270 * We entered from user mode or we're pretending to have entered 1271 * from user mode due to an IRET fault. 1272 */ 1273 SWAPGS 1274 /* We have user CR3. Change to kernel CR3. */ 1275 SWITCH_TO_KERNEL_CR3 scratch_reg=%rax 1276 1277.Lerror_entry_from_usermode_after_swapgs: 1278 /* Put us onto the real thread stack. */ 1279 popq %r12 /* save return addr in %12 */ 1280 movq %rsp, %rdi /* arg0 = pt_regs pointer */ 1281 call sync_regs 1282 movq %rax, %rsp /* switch stack */ 1283 ENCODE_FRAME_POINTER 1284 pushq %r12 1285 1286 /* 1287 * We need to tell lockdep that IRQs are off. We can't do this until 1288 * we fix gsbase, and we should do it before enter_from_user_mode 1289 * (which can take locks). 1290 */ 1291 TRACE_IRQS_OFF 1292 CALL_enter_from_user_mode 1293 ret 1294 1295.Lerror_entry_done: 1296 TRACE_IRQS_OFF 1297 ret 1298 1299 /* 1300 * There are two places in the kernel that can potentially fault with 1301 * usergs. Handle them here. B stepping K8s sometimes report a 1302 * truncated RIP for IRET exceptions returning to compat mode. Check 1303 * for these here too. 1304 */ 1305.Lerror_kernelspace: 1306 leaq native_irq_return_iret(%rip), %rcx 1307 cmpq %rcx, RIP+8(%rsp) 1308 je .Lerror_bad_iret 1309 movl %ecx, %eax /* zero extend */ 1310 cmpq %rax, RIP+8(%rsp) 1311 je .Lbstep_iret 1312 cmpq $.Lgs_change, RIP+8(%rsp) 1313 jne .Lerror_entry_done 1314 1315 /* 1316 * hack: .Lgs_change can fail with user gsbase. If this happens, fix up 1317 * gsbase and proceed. We'll fix up the exception and land in 1318 * .Lgs_change's error handler with kernel gsbase. 1319 */ 1320 SWAPGS 1321 SWITCH_TO_KERNEL_CR3 scratch_reg=%rax 1322 jmp .Lerror_entry_done 1323 1324.Lbstep_iret: 1325 /* Fix truncated RIP */ 1326 movq %rcx, RIP+8(%rsp) 1327 /* fall through */ 1328 1329.Lerror_bad_iret: 1330 /* 1331 * We came from an IRET to user mode, so we have user 1332 * gsbase and CR3. Switch to kernel gsbase and CR3: 1333 */ 1334 SWAPGS 1335 SWITCH_TO_KERNEL_CR3 scratch_reg=%rax 1336 1337 /* 1338 * Pretend that the exception came from user mode: set up pt_regs 1339 * as if we faulted immediately after IRET. 1340 */ 1341 mov %rsp, %rdi 1342 call fixup_bad_iret 1343 mov %rax, %rsp 1344 jmp .Lerror_entry_from_usermode_after_swapgs 1345END(error_entry) 1346 1347ENTRY(error_exit) 1348 UNWIND_HINT_REGS 1349 DISABLE_INTERRUPTS(CLBR_ANY) 1350 TRACE_IRQS_OFF 1351 testb $3, CS(%rsp) 1352 jz retint_kernel 1353 jmp retint_user 1354END(error_exit) 1355 1356/* 1357 * Runs on exception stack. Xen PV does not go through this path at all, 1358 * so we can use real assembly here. 1359 * 1360 * Registers: 1361 * %r14: Used to save/restore the CR3 of the interrupted context 1362 * when PAGE_TABLE_ISOLATION is in use. Do not clobber. 1363 */ 1364ENTRY(nmi) 1365 UNWIND_HINT_IRET_REGS 1366 1367 /* 1368 * We allow breakpoints in NMIs. If a breakpoint occurs, then 1369 * the iretq it performs will take us out of NMI context. 1370 * This means that we can have nested NMIs where the next 1371 * NMI is using the top of the stack of the previous NMI. We 1372 * can't let it execute because the nested NMI will corrupt the 1373 * stack of the previous NMI. NMI handlers are not re-entrant 1374 * anyway. 1375 * 1376 * To handle this case we do the following: 1377 * Check the a special location on the stack that contains 1378 * a variable that is set when NMIs are executing. 1379 * The interrupted task's stack is also checked to see if it 1380 * is an NMI stack. 1381 * If the variable is not set and the stack is not the NMI 1382 * stack then: 1383 * o Set the special variable on the stack 1384 * o Copy the interrupt frame into an "outermost" location on the 1385 * stack 1386 * o Copy the interrupt frame into an "iret" location on the stack 1387 * o Continue processing the NMI 1388 * If the variable is set or the previous stack is the NMI stack: 1389 * o Modify the "iret" location to jump to the repeat_nmi 1390 * o return back to the first NMI 1391 * 1392 * Now on exit of the first NMI, we first clear the stack variable 1393 * The NMI stack will tell any nested NMIs at that point that it is 1394 * nested. Then we pop the stack normally with iret, and if there was 1395 * a nested NMI that updated the copy interrupt stack frame, a 1396 * jump will be made to the repeat_nmi code that will handle the second 1397 * NMI. 1398 * 1399 * However, espfix prevents us from directly returning to userspace 1400 * with a single IRET instruction. Similarly, IRET to user mode 1401 * can fault. We therefore handle NMIs from user space like 1402 * other IST entries. 1403 */ 1404 1405 ASM_CLAC 1406 1407 /* Use %rdx as our temp variable throughout */ 1408 pushq %rdx 1409 1410 testb $3, CS-RIP+8(%rsp) 1411 jz .Lnmi_from_kernel 1412 1413 /* 1414 * NMI from user mode. We need to run on the thread stack, but we 1415 * can't go through the normal entry paths: NMIs are masked, and 1416 * we don't want to enable interrupts, because then we'll end 1417 * up in an awkward situation in which IRQs are on but NMIs 1418 * are off. 1419 * 1420 * We also must not push anything to the stack before switching 1421 * stacks lest we corrupt the "NMI executing" variable. 1422 */ 1423 1424 swapgs 1425 cld 1426 SWITCH_TO_KERNEL_CR3 scratch_reg=%rdx 1427 movq %rsp, %rdx 1428 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 1429 UNWIND_HINT_IRET_REGS base=%rdx offset=8 1430 pushq 5*8(%rdx) /* pt_regs->ss */ 1431 pushq 4*8(%rdx) /* pt_regs->rsp */ 1432 pushq 3*8(%rdx) /* pt_regs->flags */ 1433 pushq 2*8(%rdx) /* pt_regs->cs */ 1434 pushq 1*8(%rdx) /* pt_regs->rip */ 1435 UNWIND_HINT_IRET_REGS 1436 pushq $-1 /* pt_regs->orig_ax */ 1437 PUSH_AND_CLEAR_REGS rdx=(%rdx) 1438 ENCODE_FRAME_POINTER 1439 1440 /* 1441 * At this point we no longer need to worry about stack damage 1442 * due to nesting -- we're on the normal thread stack and we're 1443 * done with the NMI stack. 1444 */ 1445 1446 movq %rsp, %rdi 1447 movq $-1, %rsi 1448 call do_nmi 1449 1450 /* 1451 * Return back to user mode. We must *not* do the normal exit 1452 * work, because we don't want to enable interrupts. 1453 */ 1454 jmp swapgs_restore_regs_and_return_to_usermode 1455 1456.Lnmi_from_kernel: 1457 /* 1458 * Here's what our stack frame will look like: 1459 * +---------------------------------------------------------+ 1460 * | original SS | 1461 * | original Return RSP | 1462 * | original RFLAGS | 1463 * | original CS | 1464 * | original RIP | 1465 * +---------------------------------------------------------+ 1466 * | temp storage for rdx | 1467 * +---------------------------------------------------------+ 1468 * | "NMI executing" variable | 1469 * +---------------------------------------------------------+ 1470 * | iret SS } Copied from "outermost" frame | 1471 * | iret Return RSP } on each loop iteration; overwritten | 1472 * | iret RFLAGS } by a nested NMI to force another | 1473 * | iret CS } iteration if needed. | 1474 * | iret RIP } | 1475 * +---------------------------------------------------------+ 1476 * | outermost SS } initialized in first_nmi; | 1477 * | outermost Return RSP } will not be changed before | 1478 * | outermost RFLAGS } NMI processing is done. | 1479 * | outermost CS } Copied to "iret" frame on each | 1480 * | outermost RIP } iteration. | 1481 * +---------------------------------------------------------+ 1482 * | pt_regs | 1483 * +---------------------------------------------------------+ 1484 * 1485 * The "original" frame is used by hardware. Before re-enabling 1486 * NMIs, we need to be done with it, and we need to leave enough 1487 * space for the asm code here. 1488 * 1489 * We return by executing IRET while RSP points to the "iret" frame. 1490 * That will either return for real or it will loop back into NMI 1491 * processing. 1492 * 1493 * The "outermost" frame is copied to the "iret" frame on each 1494 * iteration of the loop, so each iteration starts with the "iret" 1495 * frame pointing to the final return target. 1496 */ 1497 1498 /* 1499 * Determine whether we're a nested NMI. 1500 * 1501 * If we interrupted kernel code between repeat_nmi and 1502 * end_repeat_nmi, then we are a nested NMI. We must not 1503 * modify the "iret" frame because it's being written by 1504 * the outer NMI. That's okay; the outer NMI handler is 1505 * about to about to call do_nmi anyway, so we can just 1506 * resume the outer NMI. 1507 */ 1508 1509 movq $repeat_nmi, %rdx 1510 cmpq 8(%rsp), %rdx 1511 ja 1f 1512 movq $end_repeat_nmi, %rdx 1513 cmpq 8(%rsp), %rdx 1514 ja nested_nmi_out 15151: 1516 1517 /* 1518 * Now check "NMI executing". If it's set, then we're nested. 1519 * This will not detect if we interrupted an outer NMI just 1520 * before IRET. 1521 */ 1522 cmpl $1, -8(%rsp) 1523 je nested_nmi 1524 1525 /* 1526 * Now test if the previous stack was an NMI stack. This covers 1527 * the case where we interrupt an outer NMI after it clears 1528 * "NMI executing" but before IRET. We need to be careful, though: 1529 * there is one case in which RSP could point to the NMI stack 1530 * despite there being no NMI active: naughty userspace controls 1531 * RSP at the very beginning of the SYSCALL targets. We can 1532 * pull a fast one on naughty userspace, though: we program 1533 * SYSCALL to mask DF, so userspace cannot cause DF to be set 1534 * if it controls the kernel's RSP. We set DF before we clear 1535 * "NMI executing". 1536 */ 1537 lea 6*8(%rsp), %rdx 1538 /* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */ 1539 cmpq %rdx, 4*8(%rsp) 1540 /* If the stack pointer is above the NMI stack, this is a normal NMI */ 1541 ja first_nmi 1542 1543 subq $EXCEPTION_STKSZ, %rdx 1544 cmpq %rdx, 4*8(%rsp) 1545 /* If it is below the NMI stack, it is a normal NMI */ 1546 jb first_nmi 1547 1548 /* Ah, it is within the NMI stack. */ 1549 1550 testb $(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp) 1551 jz first_nmi /* RSP was user controlled. */ 1552 1553 /* This is a nested NMI. */ 1554 1555nested_nmi: 1556 /* 1557 * Modify the "iret" frame to point to repeat_nmi, forcing another 1558 * iteration of NMI handling. 1559 */ 1560 subq $8, %rsp 1561 leaq -10*8(%rsp), %rdx 1562 pushq $__KERNEL_DS 1563 pushq %rdx 1564 pushfq 1565 pushq $__KERNEL_CS 1566 pushq $repeat_nmi 1567 1568 /* Put stack back */ 1569 addq $(6*8), %rsp 1570 1571nested_nmi_out: 1572 popq %rdx 1573 1574 /* We are returning to kernel mode, so this cannot result in a fault. */ 1575 iretq 1576 1577first_nmi: 1578 /* Restore rdx. */ 1579 movq (%rsp), %rdx 1580 1581 /* Make room for "NMI executing". */ 1582 pushq $0 1583 1584 /* Leave room for the "iret" frame */ 1585 subq $(5*8), %rsp 1586 1587 /* Copy the "original" frame to the "outermost" frame */ 1588 .rept 5 1589 pushq 11*8(%rsp) 1590 .endr 1591 UNWIND_HINT_IRET_REGS 1592 1593 /* Everything up to here is safe from nested NMIs */ 1594 1595#ifdef CONFIG_DEBUG_ENTRY 1596 /* 1597 * For ease of testing, unmask NMIs right away. Disabled by 1598 * default because IRET is very expensive. 1599 */ 1600 pushq $0 /* SS */ 1601 pushq %rsp /* RSP (minus 8 because of the previous push) */ 1602 addq $8, (%rsp) /* Fix up RSP */ 1603 pushfq /* RFLAGS */ 1604 pushq $__KERNEL_CS /* CS */ 1605 pushq $1f /* RIP */ 1606 iretq /* continues at repeat_nmi below */ 1607 UNWIND_HINT_IRET_REGS 16081: 1609#endif 1610 1611repeat_nmi: 1612 /* 1613 * If there was a nested NMI, the first NMI's iret will return 1614 * here. But NMIs are still enabled and we can take another 1615 * nested NMI. The nested NMI checks the interrupted RIP to see 1616 * if it is between repeat_nmi and end_repeat_nmi, and if so 1617 * it will just return, as we are about to repeat an NMI anyway. 1618 * This makes it safe to copy to the stack frame that a nested 1619 * NMI will update. 1620 * 1621 * RSP is pointing to "outermost RIP". gsbase is unknown, but, if 1622 * we're repeating an NMI, gsbase has the same value that it had on 1623 * the first iteration. paranoid_entry will load the kernel 1624 * gsbase if needed before we call do_nmi. "NMI executing" 1625 * is zero. 1626 */ 1627 movq $1, 10*8(%rsp) /* Set "NMI executing". */ 1628 1629 /* 1630 * Copy the "outermost" frame to the "iret" frame. NMIs that nest 1631 * here must not modify the "iret" frame while we're writing to 1632 * it or it will end up containing garbage. 1633 */ 1634 addq $(10*8), %rsp 1635 .rept 5 1636 pushq -6*8(%rsp) 1637 .endr 1638 subq $(5*8), %rsp 1639end_repeat_nmi: 1640 1641 /* 1642 * Everything below this point can be preempted by a nested NMI. 1643 * If this happens, then the inner NMI will change the "iret" 1644 * frame to point back to repeat_nmi. 1645 */ 1646 pushq $-1 /* ORIG_RAX: no syscall to restart */ 1647 1648 /* 1649 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit 1650 * as we should not be calling schedule in NMI context. 1651 * Even with normal interrupts enabled. An NMI should not be 1652 * setting NEED_RESCHED or anything that normal interrupts and 1653 * exceptions might do. 1654 */ 1655 call paranoid_entry 1656 UNWIND_HINT_REGS 1657 1658 /* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */ 1659 movq %rsp, %rdi 1660 movq $-1, %rsi 1661 call do_nmi 1662 1663 /* Always restore stashed CR3 value (see paranoid_entry) */ 1664 RESTORE_CR3 scratch_reg=%r15 save_reg=%r14 1665 1666 testl %ebx, %ebx /* swapgs needed? */ 1667 jnz nmi_restore 1668nmi_swapgs: 1669 SWAPGS_UNSAFE_STACK 1670nmi_restore: 1671 POP_REGS 1672 1673 /* 1674 * Skip orig_ax and the "outermost" frame to point RSP at the "iret" 1675 * at the "iret" frame. 1676 */ 1677 addq $6*8, %rsp 1678 1679 /* 1680 * Clear "NMI executing". Set DF first so that we can easily 1681 * distinguish the remaining code between here and IRET from 1682 * the SYSCALL entry and exit paths. 1683 * 1684 * We arguably should just inspect RIP instead, but I (Andy) wrote 1685 * this code when I had the misapprehension that Xen PV supported 1686 * NMIs, and Xen PV would break that approach. 1687 */ 1688 std 1689 movq $0, 5*8(%rsp) /* clear "NMI executing" */ 1690 1691 /* 1692 * iretq reads the "iret" frame and exits the NMI stack in a 1693 * single instruction. We are returning to kernel mode, so this 1694 * cannot result in a fault. Similarly, we don't need to worry 1695 * about espfix64 on the way back to kernel mode. 1696 */ 1697 iretq 1698END(nmi) 1699 1700#ifndef CONFIG_IA32_EMULATION 1701/* 1702 * This handles SYSCALL from 32-bit code. There is no way to program 1703 * MSRs to fully disable 32-bit SYSCALL. 1704 */ 1705ENTRY(ignore_sysret) 1706 UNWIND_HINT_EMPTY 1707 mov $-ENOSYS, %eax 1708 sysret 1709END(ignore_sysret) 1710#endif 1711 1712ENTRY(rewind_stack_do_exit) 1713 UNWIND_HINT_FUNC 1714 /* Prevent any naive code from trying to unwind to our caller. */ 1715 xorl %ebp, %ebp 1716 1717 movq PER_CPU_VAR(cpu_current_top_of_stack), %rax 1718 leaq -PTREGS_SIZE(%rax), %rsp 1719 UNWIND_HINT_FUNC sp_offset=PTREGS_SIZE 1720 1721 call do_exit 1722END(rewind_stack_do_exit) 1723