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.txt 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 bt $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 */ 144 145 .pushsection .entry_trampoline, "ax" 146 147/* 148 * The code in here gets remapped into cpu_entry_area's trampoline. This means 149 * that the assembler and linker have the wrong idea as to where this code 150 * lives (and, in fact, it's mapped more than once, so it's not even at a 151 * fixed address). So we can't reference any symbols outside the entry 152 * trampoline and expect it to work. 153 * 154 * Instead, we carefully abuse %rip-relative addressing. 155 * _entry_trampoline(%rip) refers to the start of the remapped) entry 156 * trampoline. We can thus find cpu_entry_area with this macro: 157 */ 158 159#define CPU_ENTRY_AREA \ 160 _entry_trampoline - CPU_ENTRY_AREA_entry_trampoline(%rip) 161 162/* The top word of the SYSENTER stack is hot and is usable as scratch space. */ 163#define RSP_SCRATCH CPU_ENTRY_AREA_entry_stack + \ 164 SIZEOF_entry_stack - 8 + CPU_ENTRY_AREA 165 166ENTRY(entry_SYSCALL_64_trampoline) 167 UNWIND_HINT_EMPTY 168 swapgs 169 170 /* Stash the user RSP. */ 171 movq %rsp, RSP_SCRATCH 172 173 /* Note: using %rsp as a scratch reg. */ 174 SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp 175 176 /* Load the top of the task stack into RSP */ 177 movq CPU_ENTRY_AREA_tss + TSS_sp1 + CPU_ENTRY_AREA, %rsp 178 179 /* Start building the simulated IRET frame. */ 180 pushq $__USER_DS /* pt_regs->ss */ 181 pushq RSP_SCRATCH /* pt_regs->sp */ 182 pushq %r11 /* pt_regs->flags */ 183 pushq $__USER_CS /* pt_regs->cs */ 184 pushq %rcx /* pt_regs->ip */ 185 186 /* 187 * x86 lacks a near absolute jump, and we can't jump to the real 188 * entry text with a relative jump. We could push the target 189 * address and then use retq, but this destroys the pipeline on 190 * many CPUs (wasting over 20 cycles on Sandy Bridge). Instead, 191 * spill RDI and restore it in a second-stage trampoline. 192 */ 193 pushq %rdi 194 movq $entry_SYSCALL_64_stage2, %rdi 195 JMP_NOSPEC %rdi 196END(entry_SYSCALL_64_trampoline) 197 198 .popsection 199 200ENTRY(entry_SYSCALL_64_stage2) 201 UNWIND_HINT_EMPTY 202 popq %rdi 203 jmp entry_SYSCALL_64_after_hwframe 204END(entry_SYSCALL_64_stage2) 205 206ENTRY(entry_SYSCALL_64) 207 UNWIND_HINT_EMPTY 208 /* 209 * Interrupts are off on entry. 210 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON, 211 * it is too small to ever cause noticeable irq latency. 212 */ 213 214 swapgs 215 /* 216 * This path is only taken when PAGE_TABLE_ISOLATION is disabled so it 217 * is not required to switch CR3. 218 */ 219 movq %rsp, PER_CPU_VAR(rsp_scratch) 220 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 221 222 /* Construct struct pt_regs on stack */ 223 pushq $__USER_DS /* pt_regs->ss */ 224 pushq PER_CPU_VAR(rsp_scratch) /* pt_regs->sp */ 225 pushq %r11 /* pt_regs->flags */ 226 pushq $__USER_CS /* pt_regs->cs */ 227 pushq %rcx /* pt_regs->ip */ 228GLOBAL(entry_SYSCALL_64_after_hwframe) 229 pushq %rax /* pt_regs->orig_ax */ 230 231 PUSH_AND_CLEAR_REGS rax=$-ENOSYS 232 233 TRACE_IRQS_OFF 234 235 /* IRQs are off. */ 236 movq %rax, %rdi 237 movq %rsp, %rsi 238 call do_syscall_64 /* returns with IRQs disabled */ 239 240 TRACE_IRQS_IRETQ /* we're about to change IF */ 241 242 /* 243 * Try to use SYSRET instead of IRET if we're returning to 244 * a completely clean 64-bit userspace context. If we're not, 245 * go to the slow exit path. 246 */ 247 movq RCX(%rsp), %rcx 248 movq RIP(%rsp), %r11 249 250 cmpq %rcx, %r11 /* SYSRET requires RCX == RIP */ 251 jne swapgs_restore_regs_and_return_to_usermode 252 253 /* 254 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP 255 * in kernel space. This essentially lets the user take over 256 * the kernel, since userspace controls RSP. 257 * 258 * If width of "canonical tail" ever becomes variable, this will need 259 * to be updated to remain correct on both old and new CPUs. 260 * 261 * Change top bits to match most significant bit (47th or 56th bit 262 * depending on paging mode) in the address. 263 */ 264#ifdef CONFIG_X86_5LEVEL 265 ALTERNATIVE "shl $(64 - 48), %rcx; sar $(64 - 48), %rcx", \ 266 "shl $(64 - 57), %rcx; sar $(64 - 57), %rcx", X86_FEATURE_LA57 267#else 268 shl $(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx 269 sar $(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx 270#endif 271 272 /* If this changed %rcx, it was not canonical */ 273 cmpq %rcx, %r11 274 jne swapgs_restore_regs_and_return_to_usermode 275 276 cmpq $__USER_CS, CS(%rsp) /* CS must match SYSRET */ 277 jne swapgs_restore_regs_and_return_to_usermode 278 279 movq R11(%rsp), %r11 280 cmpq %r11, EFLAGS(%rsp) /* R11 == RFLAGS */ 281 jne swapgs_restore_regs_and_return_to_usermode 282 283 /* 284 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot 285 * restore RF properly. If the slowpath sets it for whatever reason, we 286 * need to restore it correctly. 287 * 288 * SYSRET can restore TF, but unlike IRET, restoring TF results in a 289 * trap from userspace immediately after SYSRET. This would cause an 290 * infinite loop whenever #DB happens with register state that satisfies 291 * the opportunistic SYSRET conditions. For example, single-stepping 292 * this user code: 293 * 294 * movq $stuck_here, %rcx 295 * pushfq 296 * popq %r11 297 * stuck_here: 298 * 299 * would never get past 'stuck_here'. 300 */ 301 testq $(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11 302 jnz swapgs_restore_regs_and_return_to_usermode 303 304 /* nothing to check for RSP */ 305 306 cmpq $__USER_DS, SS(%rsp) /* SS must match SYSRET */ 307 jne swapgs_restore_regs_and_return_to_usermode 308 309 /* 310 * We win! This label is here just for ease of understanding 311 * perf profiles. Nothing jumps here. 312 */ 313syscall_return_via_sysret: 314 /* rcx and r11 are already restored (see code above) */ 315 UNWIND_HINT_EMPTY 316 POP_REGS pop_rdi=0 skip_r11rcx=1 317 318 /* 319 * Now all regs are restored except RSP and RDI. 320 * Save old stack pointer and switch to trampoline stack. 321 */ 322 movq %rsp, %rdi 323 movq PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp 324 325 pushq RSP-RDI(%rdi) /* RSP */ 326 pushq (%rdi) /* RDI */ 327 328 /* 329 * We are on the trampoline stack. All regs except RDI are live. 330 * We can do future final exit work right here. 331 */ 332 SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi 333 334 popq %rdi 335 popq %rsp 336 USERGS_SYSRET64 337END(entry_SYSCALL_64) 338 339/* 340 * %rdi: prev task 341 * %rsi: next task 342 */ 343ENTRY(__switch_to_asm) 344 UNWIND_HINT_FUNC 345 /* 346 * Save callee-saved registers 347 * This must match the order in inactive_task_frame 348 */ 349 pushq %rbp 350 pushq %rbx 351 pushq %r12 352 pushq %r13 353 pushq %r14 354 pushq %r15 355 356 /* switch stack */ 357 movq %rsp, TASK_threadsp(%rdi) 358 movq TASK_threadsp(%rsi), %rsp 359 360#ifdef CONFIG_STACKPROTECTOR 361 movq TASK_stack_canary(%rsi), %rbx 362 movq %rbx, PER_CPU_VAR(irq_stack_union)+stack_canary_offset 363#endif 364 365#ifdef CONFIG_RETPOLINE 366 /* 367 * When switching from a shallower to a deeper call stack 368 * the RSB may either underflow or use entries populated 369 * with userspace addresses. On CPUs where those concerns 370 * exist, overwrite the RSB with entries which capture 371 * speculative execution to prevent attack. 372 */ 373 FILL_RETURN_BUFFER %r12, RSB_CLEAR_LOOPS, X86_FEATURE_RSB_CTXSW 374#endif 375 376 /* restore callee-saved registers */ 377 popq %r15 378 popq %r14 379 popq %r13 380 popq %r12 381 popq %rbx 382 popq %rbp 383 384 jmp __switch_to 385END(__switch_to_asm) 386 387/* 388 * A newly forked process directly context switches into this address. 389 * 390 * rax: prev task we switched from 391 * rbx: kernel thread func (NULL for user thread) 392 * r12: kernel thread arg 393 */ 394ENTRY(ret_from_fork) 395 UNWIND_HINT_EMPTY 396 movq %rax, %rdi 397 call schedule_tail /* rdi: 'prev' task parameter */ 398 399 testq %rbx, %rbx /* from kernel_thread? */ 400 jnz 1f /* kernel threads are uncommon */ 401 4022: 403 UNWIND_HINT_REGS 404 movq %rsp, %rdi 405 call syscall_return_slowpath /* returns with IRQs disabled */ 406 TRACE_IRQS_ON /* user mode is traced as IRQS on */ 407 jmp swapgs_restore_regs_and_return_to_usermode 408 4091: 410 /* kernel thread */ 411 movq %r12, %rdi 412 CALL_NOSPEC %rbx 413 /* 414 * A kernel thread is allowed to return here after successfully 415 * calling do_execve(). Exit to userspace to complete the execve() 416 * syscall. 417 */ 418 movq $0, RAX(%rsp) 419 jmp 2b 420END(ret_from_fork) 421 422/* 423 * Build the entry stubs with some assembler magic. 424 * We pack 1 stub into every 8-byte block. 425 */ 426 .align 8 427ENTRY(irq_entries_start) 428 vector=FIRST_EXTERNAL_VECTOR 429 .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR) 430 UNWIND_HINT_IRET_REGS 431 pushq $(~vector+0x80) /* Note: always in signed byte range */ 432 jmp common_interrupt 433 .align 8 434 vector=vector+1 435 .endr 436END(irq_entries_start) 437 438.macro DEBUG_ENTRY_ASSERT_IRQS_OFF 439#ifdef CONFIG_DEBUG_ENTRY 440 pushq %rax 441 SAVE_FLAGS(CLBR_RAX) 442 testl $X86_EFLAGS_IF, %eax 443 jz .Lokay_\@ 444 ud2 445.Lokay_\@: 446 popq %rax 447#endif 448.endm 449 450/* 451 * Enters the IRQ stack if we're not already using it. NMI-safe. Clobbers 452 * flags and puts old RSP into old_rsp, and leaves all other GPRs alone. 453 * Requires kernel GSBASE. 454 * 455 * The invariant is that, if irq_count != -1, then the IRQ stack is in use. 456 */ 457.macro ENTER_IRQ_STACK regs=1 old_rsp save_ret=0 458 DEBUG_ENTRY_ASSERT_IRQS_OFF 459 460 .if \save_ret 461 /* 462 * If save_ret is set, the original stack contains one additional 463 * entry -- the return address. Therefore, move the address one 464 * entry below %rsp to \old_rsp. 465 */ 466 leaq 8(%rsp), \old_rsp 467 .else 468 movq %rsp, \old_rsp 469 .endif 470 471 .if \regs 472 UNWIND_HINT_REGS base=\old_rsp 473 .endif 474 475 incl PER_CPU_VAR(irq_count) 476 jnz .Lirq_stack_push_old_rsp_\@ 477 478 /* 479 * Right now, if we just incremented irq_count to zero, we've 480 * claimed the IRQ stack but we haven't switched to it yet. 481 * 482 * If anything is added that can interrupt us here without using IST, 483 * it must be *extremely* careful to limit its stack usage. This 484 * could include kprobes and a hypothetical future IST-less #DB 485 * handler. 486 * 487 * The OOPS unwinder relies on the word at the top of the IRQ 488 * stack linking back to the previous RSP for the entire time we're 489 * on the IRQ stack. For this to work reliably, we need to write 490 * it before we actually move ourselves to the IRQ stack. 491 */ 492 493 movq \old_rsp, PER_CPU_VAR(irq_stack_union + IRQ_STACK_SIZE - 8) 494 movq PER_CPU_VAR(irq_stack_ptr), %rsp 495 496#ifdef CONFIG_DEBUG_ENTRY 497 /* 498 * If the first movq above becomes wrong due to IRQ stack layout 499 * changes, the only way we'll notice is if we try to unwind right 500 * here. Assert that we set up the stack right to catch this type 501 * of bug quickly. 502 */ 503 cmpq -8(%rsp), \old_rsp 504 je .Lirq_stack_okay\@ 505 ud2 506 .Lirq_stack_okay\@: 507#endif 508 509.Lirq_stack_push_old_rsp_\@: 510 pushq \old_rsp 511 512 .if \regs 513 UNWIND_HINT_REGS indirect=1 514 .endif 515 516 .if \save_ret 517 /* 518 * Push the return address to the stack. This return address can 519 * be found at the "real" original RSP, which was offset by 8 at 520 * the beginning of this macro. 521 */ 522 pushq -8(\old_rsp) 523 .endif 524.endm 525 526/* 527 * Undoes ENTER_IRQ_STACK. 528 */ 529.macro LEAVE_IRQ_STACK regs=1 530 DEBUG_ENTRY_ASSERT_IRQS_OFF 531 /* We need to be off the IRQ stack before decrementing irq_count. */ 532 popq %rsp 533 534 .if \regs 535 UNWIND_HINT_REGS 536 .endif 537 538 /* 539 * As in ENTER_IRQ_STACK, irq_count == 0, we are still claiming 540 * the irq stack but we're not on it. 541 */ 542 543 decl PER_CPU_VAR(irq_count) 544.endm 545 546/* 547 * Interrupt entry helper function. 548 * 549 * Entry runs with interrupts off. Stack layout at entry: 550 * +----------------------------------------------------+ 551 * | regs->ss | 552 * | regs->rsp | 553 * | regs->eflags | 554 * | regs->cs | 555 * | regs->ip | 556 * +----------------------------------------------------+ 557 * | regs->orig_ax = ~(interrupt number) | 558 * +----------------------------------------------------+ 559 * | return address | 560 * +----------------------------------------------------+ 561 */ 562ENTRY(interrupt_entry) 563 UNWIND_HINT_FUNC 564 ASM_CLAC 565 cld 566 567 testb $3, CS-ORIG_RAX+8(%rsp) 568 jz 1f 569 SWAPGS 570 571 /* 572 * Switch to the thread stack. The IRET frame and orig_ax are 573 * on the stack, as well as the return address. RDI..R12 are 574 * not (yet) on the stack and space has not (yet) been 575 * allocated for them. 576 */ 577 pushq %rdi 578 579 /* Need to switch before accessing the thread stack. */ 580 SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi 581 movq %rsp, %rdi 582 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 583 584 /* 585 * We have RDI, return address, and orig_ax on the stack on 586 * top of the IRET frame. That means offset=24 587 */ 588 UNWIND_HINT_IRET_REGS base=%rdi offset=24 589 590 pushq 7*8(%rdi) /* regs->ss */ 591 pushq 6*8(%rdi) /* regs->rsp */ 592 pushq 5*8(%rdi) /* regs->eflags */ 593 pushq 4*8(%rdi) /* regs->cs */ 594 pushq 3*8(%rdi) /* regs->ip */ 595 pushq 2*8(%rdi) /* regs->orig_ax */ 596 pushq 8(%rdi) /* return address */ 597 UNWIND_HINT_FUNC 598 599 movq (%rdi), %rdi 6001: 601 602 PUSH_AND_CLEAR_REGS save_ret=1 603 ENCODE_FRAME_POINTER 8 604 605 testb $3, CS+8(%rsp) 606 jz 1f 607 608 /* 609 * IRQ from user mode. 610 * 611 * We need to tell lockdep that IRQs are off. We can't do this until 612 * we fix gsbase, and we should do it before enter_from_user_mode 613 * (which can take locks). Since TRACE_IRQS_OFF is idempotent, 614 * the simplest way to handle it is to just call it twice if 615 * we enter from user mode. There's no reason to optimize this since 616 * TRACE_IRQS_OFF is a no-op if lockdep is off. 617 */ 618 TRACE_IRQS_OFF 619 620 CALL_enter_from_user_mode 621 6221: 623 ENTER_IRQ_STACK old_rsp=%rdi save_ret=1 624 /* We entered an interrupt context - irqs are off: */ 625 TRACE_IRQS_OFF 626 627 ret 628END(interrupt_entry) 629 630 631/* Interrupt entry/exit. */ 632 633 /* 634 * The interrupt stubs push (~vector+0x80) onto the stack and 635 * then jump to common_interrupt. 636 */ 637 .p2align CONFIG_X86_L1_CACHE_SHIFT 638common_interrupt: 639 addq $-0x80, (%rsp) /* Adjust vector to [-256, -1] range */ 640 call interrupt_entry 641 UNWIND_HINT_REGS indirect=1 642 call do_IRQ /* rdi points to pt_regs */ 643 /* 0(%rsp): old RSP */ 644ret_from_intr: 645 DISABLE_INTERRUPTS(CLBR_ANY) 646 TRACE_IRQS_OFF 647 648 LEAVE_IRQ_STACK 649 650 testb $3, CS(%rsp) 651 jz retint_kernel 652 653 /* Interrupt came from user space */ 654GLOBAL(retint_user) 655 mov %rsp,%rdi 656 call prepare_exit_to_usermode 657 TRACE_IRQS_IRETQ 658 659GLOBAL(swapgs_restore_regs_and_return_to_usermode) 660#ifdef CONFIG_DEBUG_ENTRY 661 /* Assert that pt_regs indicates user mode. */ 662 testb $3, CS(%rsp) 663 jnz 1f 664 ud2 6651: 666#endif 667 POP_REGS pop_rdi=0 668 669 /* 670 * The stack is now user RDI, orig_ax, RIP, CS, EFLAGS, RSP, SS. 671 * Save old stack pointer and switch to trampoline stack. 672 */ 673 movq %rsp, %rdi 674 movq PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp 675 676 /* Copy the IRET frame to the trampoline stack. */ 677 pushq 6*8(%rdi) /* SS */ 678 pushq 5*8(%rdi) /* RSP */ 679 pushq 4*8(%rdi) /* EFLAGS */ 680 pushq 3*8(%rdi) /* CS */ 681 pushq 2*8(%rdi) /* RIP */ 682 683 /* Push user RDI on the trampoline stack. */ 684 pushq (%rdi) 685 686 /* 687 * We are on the trampoline stack. All regs except RDI are live. 688 * We can do future final exit work right here. 689 */ 690 691 SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi 692 693 /* Restore RDI. */ 694 popq %rdi 695 SWAPGS 696 INTERRUPT_RETURN 697 698 699/* Returning to kernel space */ 700retint_kernel: 701#ifdef CONFIG_PREEMPT 702 /* Interrupts are off */ 703 /* Check if we need preemption */ 704 bt $9, EFLAGS(%rsp) /* were interrupts off? */ 705 jnc 1f 7060: cmpl $0, PER_CPU_VAR(__preempt_count) 707 jnz 1f 708 call preempt_schedule_irq 709 jmp 0b 7101: 711#endif 712 /* 713 * The iretq could re-enable interrupts: 714 */ 715 TRACE_IRQS_IRETQ 716 717GLOBAL(restore_regs_and_return_to_kernel) 718#ifdef CONFIG_DEBUG_ENTRY 719 /* Assert that pt_regs indicates kernel mode. */ 720 testb $3, CS(%rsp) 721 jz 1f 722 ud2 7231: 724#endif 725 POP_REGS 726 addq $8, %rsp /* skip regs->orig_ax */ 727 /* 728 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization 729 * when returning from IPI handler. 730 */ 731 INTERRUPT_RETURN 732 733ENTRY(native_iret) 734 UNWIND_HINT_IRET_REGS 735 /* 736 * Are we returning to a stack segment from the LDT? Note: in 737 * 64-bit mode SS:RSP on the exception stack is always valid. 738 */ 739#ifdef CONFIG_X86_ESPFIX64 740 testb $4, (SS-RIP)(%rsp) 741 jnz native_irq_return_ldt 742#endif 743 744.global native_irq_return_iret 745native_irq_return_iret: 746 /* 747 * This may fault. Non-paranoid faults on return to userspace are 748 * handled by fixup_bad_iret. These include #SS, #GP, and #NP. 749 * Double-faults due to espfix64 are handled in do_double_fault. 750 * Other faults here are fatal. 751 */ 752 iretq 753 754#ifdef CONFIG_X86_ESPFIX64 755native_irq_return_ldt: 756 /* 757 * We are running with user GSBASE. All GPRs contain their user 758 * values. We have a percpu ESPFIX stack that is eight slots 759 * long (see ESPFIX_STACK_SIZE). espfix_waddr points to the bottom 760 * of the ESPFIX stack. 761 * 762 * We clobber RAX and RDI in this code. We stash RDI on the 763 * normal stack and RAX on the ESPFIX stack. 764 * 765 * The ESPFIX stack layout we set up looks like this: 766 * 767 * --- top of ESPFIX stack --- 768 * SS 769 * RSP 770 * RFLAGS 771 * CS 772 * RIP <-- RSP points here when we're done 773 * RAX <-- espfix_waddr points here 774 * --- bottom of ESPFIX stack --- 775 */ 776 777 pushq %rdi /* Stash user RDI */ 778 SWAPGS /* to kernel GS */ 779 SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi /* to kernel CR3 */ 780 781 movq PER_CPU_VAR(espfix_waddr), %rdi 782 movq %rax, (0*8)(%rdi) /* user RAX */ 783 movq (1*8)(%rsp), %rax /* user RIP */ 784 movq %rax, (1*8)(%rdi) 785 movq (2*8)(%rsp), %rax /* user CS */ 786 movq %rax, (2*8)(%rdi) 787 movq (3*8)(%rsp), %rax /* user RFLAGS */ 788 movq %rax, (3*8)(%rdi) 789 movq (5*8)(%rsp), %rax /* user SS */ 790 movq %rax, (5*8)(%rdi) 791 movq (4*8)(%rsp), %rax /* user RSP */ 792 movq %rax, (4*8)(%rdi) 793 /* Now RAX == RSP. */ 794 795 andl $0xffff0000, %eax /* RAX = (RSP & 0xffff0000) */ 796 797 /* 798 * espfix_stack[31:16] == 0. The page tables are set up such that 799 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of 800 * espfix_waddr for any X. That is, there are 65536 RO aliases of 801 * the same page. Set up RSP so that RSP[31:16] contains the 802 * respective 16 bits of the /userspace/ RSP and RSP nonetheless 803 * still points to an RO alias of the ESPFIX stack. 804 */ 805 orq PER_CPU_VAR(espfix_stack), %rax 806 807 SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi 808 SWAPGS /* to user GS */ 809 popq %rdi /* Restore user RDI */ 810 811 movq %rax, %rsp 812 UNWIND_HINT_IRET_REGS offset=8 813 814 /* 815 * At this point, we cannot write to the stack any more, but we can 816 * still read. 817 */ 818 popq %rax /* Restore user RAX */ 819 820 /* 821 * RSP now points to an ordinary IRET frame, except that the page 822 * is read-only and RSP[31:16] are preloaded with the userspace 823 * values. We can now IRET back to userspace. 824 */ 825 jmp native_irq_return_iret 826#endif 827END(common_interrupt) 828 829/* 830 * APIC interrupts. 831 */ 832.macro apicinterrupt3 num sym do_sym 833ENTRY(\sym) 834 UNWIND_HINT_IRET_REGS 835 pushq $~(\num) 836.Lcommon_\sym: 837 call interrupt_entry 838 UNWIND_HINT_REGS indirect=1 839 call \do_sym /* rdi points to pt_regs */ 840 jmp ret_from_intr 841END(\sym) 842.endm 843 844/* Make sure APIC interrupt handlers end up in the irqentry section: */ 845#define PUSH_SECTION_IRQENTRY .pushsection .irqentry.text, "ax" 846#define POP_SECTION_IRQENTRY .popsection 847 848.macro apicinterrupt num sym do_sym 849PUSH_SECTION_IRQENTRY 850apicinterrupt3 \num \sym \do_sym 851POP_SECTION_IRQENTRY 852.endm 853 854#ifdef CONFIG_SMP 855apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR irq_move_cleanup_interrupt smp_irq_move_cleanup_interrupt 856apicinterrupt3 REBOOT_VECTOR reboot_interrupt smp_reboot_interrupt 857#endif 858 859#ifdef CONFIG_X86_UV 860apicinterrupt3 UV_BAU_MESSAGE uv_bau_message_intr1 uv_bau_message_interrupt 861#endif 862 863apicinterrupt LOCAL_TIMER_VECTOR apic_timer_interrupt smp_apic_timer_interrupt 864apicinterrupt X86_PLATFORM_IPI_VECTOR x86_platform_ipi smp_x86_platform_ipi 865 866#ifdef CONFIG_HAVE_KVM 867apicinterrupt3 POSTED_INTR_VECTOR kvm_posted_intr_ipi smp_kvm_posted_intr_ipi 868apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR kvm_posted_intr_wakeup_ipi smp_kvm_posted_intr_wakeup_ipi 869apicinterrupt3 POSTED_INTR_NESTED_VECTOR kvm_posted_intr_nested_ipi smp_kvm_posted_intr_nested_ipi 870#endif 871 872#ifdef CONFIG_X86_MCE_THRESHOLD 873apicinterrupt THRESHOLD_APIC_VECTOR threshold_interrupt smp_threshold_interrupt 874#endif 875 876#ifdef CONFIG_X86_MCE_AMD 877apicinterrupt DEFERRED_ERROR_VECTOR deferred_error_interrupt smp_deferred_error_interrupt 878#endif 879 880#ifdef CONFIG_X86_THERMAL_VECTOR 881apicinterrupt THERMAL_APIC_VECTOR thermal_interrupt smp_thermal_interrupt 882#endif 883 884#ifdef CONFIG_SMP 885apicinterrupt CALL_FUNCTION_SINGLE_VECTOR call_function_single_interrupt smp_call_function_single_interrupt 886apicinterrupt CALL_FUNCTION_VECTOR call_function_interrupt smp_call_function_interrupt 887apicinterrupt RESCHEDULE_VECTOR reschedule_interrupt smp_reschedule_interrupt 888#endif 889 890apicinterrupt ERROR_APIC_VECTOR error_interrupt smp_error_interrupt 891apicinterrupt SPURIOUS_APIC_VECTOR spurious_interrupt smp_spurious_interrupt 892 893#ifdef CONFIG_IRQ_WORK 894apicinterrupt IRQ_WORK_VECTOR irq_work_interrupt smp_irq_work_interrupt 895#endif 896 897/* 898 * Exception entry points. 899 */ 900#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss_rw) + (TSS_ist + ((x) - 1) * 8) 901 902.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1 903ENTRY(\sym) 904 UNWIND_HINT_IRET_REGS offset=\has_error_code*8 905 906 /* Sanity check */ 907 .if \shift_ist != -1 && \paranoid == 0 908 .error "using shift_ist requires paranoid=1" 909 .endif 910 911 ASM_CLAC 912 913 .if \has_error_code == 0 914 pushq $-1 /* ORIG_RAX: no syscall to restart */ 915 .endif 916 917 .if \paranoid == 1 918 testb $3, CS-ORIG_RAX(%rsp) /* If coming from userspace, switch stacks */ 919 jnz .Lfrom_usermode_switch_stack_\@ 920 .endif 921 922 .if \paranoid 923 call paranoid_entry 924 .else 925 call error_entry 926 .endif 927 UNWIND_HINT_REGS 928 /* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */ 929 930 .if \paranoid 931 .if \shift_ist != -1 932 TRACE_IRQS_OFF_DEBUG /* reload IDT in case of recursion */ 933 .else 934 TRACE_IRQS_OFF 935 .endif 936 .endif 937 938 movq %rsp, %rdi /* pt_regs pointer */ 939 940 .if \has_error_code 941 movq ORIG_RAX(%rsp), %rsi /* get error code */ 942 movq $-1, ORIG_RAX(%rsp) /* no syscall to restart */ 943 .else 944 xorl %esi, %esi /* no error code */ 945 .endif 946 947 .if \shift_ist != -1 948 subq $EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist) 949 .endif 950 951 call \do_sym 952 953 .if \shift_ist != -1 954 addq $EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist) 955 .endif 956 957 /* these procedures expect "no swapgs" flag in ebx */ 958 .if \paranoid 959 jmp paranoid_exit 960 .else 961 jmp error_exit 962 .endif 963 964 .if \paranoid == 1 965 /* 966 * Entry from userspace. Switch stacks and treat it 967 * as a normal entry. This means that paranoid handlers 968 * run in real process context if user_mode(regs). 969 */ 970.Lfrom_usermode_switch_stack_\@: 971 call error_entry 972 973 movq %rsp, %rdi /* pt_regs pointer */ 974 975 .if \has_error_code 976 movq ORIG_RAX(%rsp), %rsi /* get error code */ 977 movq $-1, ORIG_RAX(%rsp) /* no syscall to restart */ 978 .else 979 xorl %esi, %esi /* no error code */ 980 .endif 981 982 call \do_sym 983 984 jmp error_exit 985 .endif 986END(\sym) 987.endm 988 989idtentry divide_error do_divide_error has_error_code=0 990idtentry overflow do_overflow has_error_code=0 991idtentry bounds do_bounds has_error_code=0 992idtentry invalid_op do_invalid_op has_error_code=0 993idtentry device_not_available do_device_not_available has_error_code=0 994idtentry double_fault do_double_fault has_error_code=1 paranoid=2 995idtentry coprocessor_segment_overrun do_coprocessor_segment_overrun has_error_code=0 996idtentry invalid_TSS do_invalid_TSS has_error_code=1 997idtentry segment_not_present do_segment_not_present has_error_code=1 998idtentry spurious_interrupt_bug do_spurious_interrupt_bug has_error_code=0 999idtentry coprocessor_error do_coprocessor_error has_error_code=0 1000idtentry alignment_check do_alignment_check has_error_code=1 1001idtentry simd_coprocessor_error do_simd_coprocessor_error has_error_code=0 1002 1003 1004 /* 1005 * Reload gs selector with exception handling 1006 * edi: new selector 1007 */ 1008ENTRY(native_load_gs_index) 1009 FRAME_BEGIN 1010 pushfq 1011 DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI) 1012 TRACE_IRQS_OFF 1013 SWAPGS 1014.Lgs_change: 1015 movl %edi, %gs 10162: ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE 1017 SWAPGS 1018 TRACE_IRQS_FLAGS (%rsp) 1019 popfq 1020 FRAME_END 1021 ret 1022ENDPROC(native_load_gs_index) 1023EXPORT_SYMBOL(native_load_gs_index) 1024 1025 _ASM_EXTABLE(.Lgs_change, bad_gs) 1026 .section .fixup, "ax" 1027 /* running with kernelgs */ 1028bad_gs: 1029 SWAPGS /* switch back to user gs */ 1030.macro ZAP_GS 1031 /* This can't be a string because the preprocessor needs to see it. */ 1032 movl $__USER_DS, %eax 1033 movl %eax, %gs 1034.endm 1035 ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG 1036 xorl %eax, %eax 1037 movl %eax, %gs 1038 jmp 2b 1039 .previous 1040 1041/* Call softirq on interrupt stack. Interrupts are off. */ 1042ENTRY(do_softirq_own_stack) 1043 pushq %rbp 1044 mov %rsp, %rbp 1045 ENTER_IRQ_STACK regs=0 old_rsp=%r11 1046 call __do_softirq 1047 LEAVE_IRQ_STACK regs=0 1048 leaveq 1049 ret 1050ENDPROC(do_softirq_own_stack) 1051 1052#ifdef CONFIG_XEN 1053idtentry hypervisor_callback xen_do_hypervisor_callback has_error_code=0 1054 1055/* 1056 * A note on the "critical region" in our callback handler. 1057 * We want to avoid stacking callback handlers due to events occurring 1058 * during handling of the last event. To do this, we keep events disabled 1059 * until we've done all processing. HOWEVER, we must enable events before 1060 * popping the stack frame (can't be done atomically) and so it would still 1061 * be possible to get enough handler activations to overflow the stack. 1062 * Although unlikely, bugs of that kind are hard to track down, so we'd 1063 * like to avoid the possibility. 1064 * So, on entry to the handler we detect whether we interrupted an 1065 * existing activation in its critical region -- if so, we pop the current 1066 * activation and restart the handler using the previous one. 1067 */ 1068ENTRY(xen_do_hypervisor_callback) /* do_hypervisor_callback(struct *pt_regs) */ 1069 1070/* 1071 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will 1072 * see the correct pointer to the pt_regs 1073 */ 1074 UNWIND_HINT_FUNC 1075 movq %rdi, %rsp /* we don't return, adjust the stack frame */ 1076 UNWIND_HINT_REGS 1077 1078 ENTER_IRQ_STACK old_rsp=%r10 1079 call xen_evtchn_do_upcall 1080 LEAVE_IRQ_STACK 1081 1082#ifndef CONFIG_PREEMPT 1083 call xen_maybe_preempt_hcall 1084#endif 1085 jmp error_exit 1086END(xen_do_hypervisor_callback) 1087 1088/* 1089 * Hypervisor uses this for application faults while it executes. 1090 * We get here for two reasons: 1091 * 1. Fault while reloading DS, ES, FS or GS 1092 * 2. Fault while executing IRET 1093 * Category 1 we do not need to fix up as Xen has already reloaded all segment 1094 * registers that could be reloaded and zeroed the others. 1095 * Category 2 we fix up by killing the current process. We cannot use the 1096 * normal Linux return path in this case because if we use the IRET hypercall 1097 * to pop the stack frame we end up in an infinite loop of failsafe callbacks. 1098 * We distinguish between categories by comparing each saved segment register 1099 * with its current contents: any discrepancy means we in category 1. 1100 */ 1101ENTRY(xen_failsafe_callback) 1102 UNWIND_HINT_EMPTY 1103 movl %ds, %ecx 1104 cmpw %cx, 0x10(%rsp) 1105 jne 1f 1106 movl %es, %ecx 1107 cmpw %cx, 0x18(%rsp) 1108 jne 1f 1109 movl %fs, %ecx 1110 cmpw %cx, 0x20(%rsp) 1111 jne 1f 1112 movl %gs, %ecx 1113 cmpw %cx, 0x28(%rsp) 1114 jne 1f 1115 /* All segments match their saved values => Category 2 (Bad IRET). */ 1116 movq (%rsp), %rcx 1117 movq 8(%rsp), %r11 1118 addq $0x30, %rsp 1119 pushq $0 /* RIP */ 1120 UNWIND_HINT_IRET_REGS offset=8 1121 jmp general_protection 11221: /* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */ 1123 movq (%rsp), %rcx 1124 movq 8(%rsp), %r11 1125 addq $0x30, %rsp 1126 UNWIND_HINT_IRET_REGS 1127 pushq $-1 /* orig_ax = -1 => not a system call */ 1128 PUSH_AND_CLEAR_REGS 1129 ENCODE_FRAME_POINTER 1130 jmp error_exit 1131END(xen_failsafe_callback) 1132 1133apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \ 1134 xen_hvm_callback_vector xen_evtchn_do_upcall 1135 1136#endif /* CONFIG_XEN */ 1137 1138#if IS_ENABLED(CONFIG_HYPERV) 1139apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \ 1140 hyperv_callback_vector hyperv_vector_handler 1141 1142apicinterrupt3 HYPERV_REENLIGHTENMENT_VECTOR \ 1143 hyperv_reenlightenment_vector hyperv_reenlightenment_intr 1144 1145apicinterrupt3 HYPERV_STIMER0_VECTOR \ 1146 hv_stimer0_callback_vector hv_stimer0_vector_handler 1147#endif /* CONFIG_HYPERV */ 1148 1149idtentry debug do_debug has_error_code=0 paranoid=1 shift_ist=DEBUG_STACK 1150idtentry int3 do_int3 has_error_code=0 1151idtentry stack_segment do_stack_segment has_error_code=1 1152 1153#ifdef CONFIG_XEN 1154idtentry xennmi do_nmi has_error_code=0 1155idtentry xendebug do_debug has_error_code=0 1156idtentry xenint3 do_int3 has_error_code=0 1157#endif 1158 1159idtentry general_protection do_general_protection has_error_code=1 1160idtentry page_fault do_page_fault has_error_code=1 1161 1162#ifdef CONFIG_KVM_GUEST 1163idtentry async_page_fault do_async_page_fault has_error_code=1 1164#endif 1165 1166#ifdef CONFIG_X86_MCE 1167idtentry machine_check do_mce has_error_code=0 paranoid=1 1168#endif 1169 1170/* 1171 * Save all registers in pt_regs, and switch gs if needed. 1172 * Use slow, but surefire "are we in kernel?" check. 1173 * Return: ebx=0: need swapgs on exit, ebx=1: otherwise 1174 */ 1175ENTRY(paranoid_entry) 1176 UNWIND_HINT_FUNC 1177 cld 1178 PUSH_AND_CLEAR_REGS save_ret=1 1179 ENCODE_FRAME_POINTER 8 1180 movl $1, %ebx 1181 movl $MSR_GS_BASE, %ecx 1182 rdmsr 1183 testl %edx, %edx 1184 js 1f /* negative -> in kernel */ 1185 SWAPGS 1186 xorl %ebx, %ebx 1187 11881: 1189 SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14 1190 1191 ret 1192END(paranoid_entry) 1193 1194/* 1195 * "Paranoid" exit path from exception stack. This is invoked 1196 * only on return from non-NMI IST interrupts that came 1197 * from kernel space. 1198 * 1199 * We may be returning to very strange contexts (e.g. very early 1200 * in syscall entry), so checking for preemption here would 1201 * be complicated. Fortunately, we there's no good reason 1202 * to try to handle preemption here. 1203 * 1204 * On entry, ebx is "no swapgs" flag (1: don't need swapgs, 0: need it) 1205 */ 1206ENTRY(paranoid_exit) 1207 UNWIND_HINT_REGS 1208 DISABLE_INTERRUPTS(CLBR_ANY) 1209 TRACE_IRQS_OFF_DEBUG 1210 testl %ebx, %ebx /* swapgs needed? */ 1211 jnz .Lparanoid_exit_no_swapgs 1212 TRACE_IRQS_IRETQ 1213 RESTORE_CR3 scratch_reg=%rbx save_reg=%r14 1214 SWAPGS_UNSAFE_STACK 1215 jmp .Lparanoid_exit_restore 1216.Lparanoid_exit_no_swapgs: 1217 TRACE_IRQS_IRETQ_DEBUG 1218 RESTORE_CR3 scratch_reg=%rbx save_reg=%r14 1219.Lparanoid_exit_restore: 1220 jmp restore_regs_and_return_to_kernel 1221END(paranoid_exit) 1222 1223/* 1224 * Save all registers in pt_regs, and switch GS if needed. 1225 */ 1226ENTRY(error_entry) 1227 UNWIND_HINT_FUNC 1228 cld 1229 PUSH_AND_CLEAR_REGS save_ret=1 1230 ENCODE_FRAME_POINTER 8 1231 testb $3, CS+8(%rsp) 1232 jz .Lerror_kernelspace 1233 1234 /* 1235 * We entered from user mode or we're pretending to have entered 1236 * from user mode due to an IRET fault. 1237 */ 1238 SWAPGS 1239 /* We have user CR3. Change to kernel CR3. */ 1240 SWITCH_TO_KERNEL_CR3 scratch_reg=%rax 1241 1242.Lerror_entry_from_usermode_after_swapgs: 1243 /* Put us onto the real thread stack. */ 1244 popq %r12 /* save return addr in %12 */ 1245 movq %rsp, %rdi /* arg0 = pt_regs pointer */ 1246 call sync_regs 1247 movq %rax, %rsp /* switch stack */ 1248 ENCODE_FRAME_POINTER 1249 pushq %r12 1250 1251 /* 1252 * We need to tell lockdep that IRQs are off. We can't do this until 1253 * we fix gsbase, and we should do it before enter_from_user_mode 1254 * (which can take locks). 1255 */ 1256 TRACE_IRQS_OFF 1257 CALL_enter_from_user_mode 1258 ret 1259 1260.Lerror_entry_done: 1261 TRACE_IRQS_OFF 1262 ret 1263 1264 /* 1265 * There are two places in the kernel that can potentially fault with 1266 * usergs. Handle them here. B stepping K8s sometimes report a 1267 * truncated RIP for IRET exceptions returning to compat mode. Check 1268 * for these here too. 1269 */ 1270.Lerror_kernelspace: 1271 leaq native_irq_return_iret(%rip), %rcx 1272 cmpq %rcx, RIP+8(%rsp) 1273 je .Lerror_bad_iret 1274 movl %ecx, %eax /* zero extend */ 1275 cmpq %rax, RIP+8(%rsp) 1276 je .Lbstep_iret 1277 cmpq $.Lgs_change, RIP+8(%rsp) 1278 jne .Lerror_entry_done 1279 1280 /* 1281 * hack: .Lgs_change can fail with user gsbase. If this happens, fix up 1282 * gsbase and proceed. We'll fix up the exception and land in 1283 * .Lgs_change's error handler with kernel gsbase. 1284 */ 1285 SWAPGS 1286 SWITCH_TO_KERNEL_CR3 scratch_reg=%rax 1287 jmp .Lerror_entry_done 1288 1289.Lbstep_iret: 1290 /* Fix truncated RIP */ 1291 movq %rcx, RIP+8(%rsp) 1292 /* fall through */ 1293 1294.Lerror_bad_iret: 1295 /* 1296 * We came from an IRET to user mode, so we have user 1297 * gsbase and CR3. Switch to kernel gsbase and CR3: 1298 */ 1299 SWAPGS 1300 SWITCH_TO_KERNEL_CR3 scratch_reg=%rax 1301 1302 /* 1303 * Pretend that the exception came from user mode: set up pt_regs 1304 * as if we faulted immediately after IRET. 1305 */ 1306 mov %rsp, %rdi 1307 call fixup_bad_iret 1308 mov %rax, %rsp 1309 jmp .Lerror_entry_from_usermode_after_swapgs 1310END(error_entry) 1311 1312ENTRY(error_exit) 1313 UNWIND_HINT_REGS 1314 DISABLE_INTERRUPTS(CLBR_ANY) 1315 TRACE_IRQS_OFF 1316 testb $3, CS(%rsp) 1317 jz retint_kernel 1318 jmp retint_user 1319END(error_exit) 1320 1321/* 1322 * Runs on exception stack. Xen PV does not go through this path at all, 1323 * so we can use real assembly here. 1324 * 1325 * Registers: 1326 * %r14: Used to save/restore the CR3 of the interrupted context 1327 * when PAGE_TABLE_ISOLATION is in use. Do not clobber. 1328 */ 1329ENTRY(nmi) 1330 UNWIND_HINT_IRET_REGS 1331 1332 /* 1333 * We allow breakpoints in NMIs. If a breakpoint occurs, then 1334 * the iretq it performs will take us out of NMI context. 1335 * This means that we can have nested NMIs where the next 1336 * NMI is using the top of the stack of the previous NMI. We 1337 * can't let it execute because the nested NMI will corrupt the 1338 * stack of the previous NMI. NMI handlers are not re-entrant 1339 * anyway. 1340 * 1341 * To handle this case we do the following: 1342 * Check the a special location on the stack that contains 1343 * a variable that is set when NMIs are executing. 1344 * The interrupted task's stack is also checked to see if it 1345 * is an NMI stack. 1346 * If the variable is not set and the stack is not the NMI 1347 * stack then: 1348 * o Set the special variable on the stack 1349 * o Copy the interrupt frame into an "outermost" location on the 1350 * stack 1351 * o Copy the interrupt frame into an "iret" location on the stack 1352 * o Continue processing the NMI 1353 * If the variable is set or the previous stack is the NMI stack: 1354 * o Modify the "iret" location to jump to the repeat_nmi 1355 * o return back to the first NMI 1356 * 1357 * Now on exit of the first NMI, we first clear the stack variable 1358 * The NMI stack will tell any nested NMIs at that point that it is 1359 * nested. Then we pop the stack normally with iret, and if there was 1360 * a nested NMI that updated the copy interrupt stack frame, a 1361 * jump will be made to the repeat_nmi code that will handle the second 1362 * NMI. 1363 * 1364 * However, espfix prevents us from directly returning to userspace 1365 * with a single IRET instruction. Similarly, IRET to user mode 1366 * can fault. We therefore handle NMIs from user space like 1367 * other IST entries. 1368 */ 1369 1370 ASM_CLAC 1371 1372 /* Use %rdx as our temp variable throughout */ 1373 pushq %rdx 1374 1375 testb $3, CS-RIP+8(%rsp) 1376 jz .Lnmi_from_kernel 1377 1378 /* 1379 * NMI from user mode. We need to run on the thread stack, but we 1380 * can't go through the normal entry paths: NMIs are masked, and 1381 * we don't want to enable interrupts, because then we'll end 1382 * up in an awkward situation in which IRQs are on but NMIs 1383 * are off. 1384 * 1385 * We also must not push anything to the stack before switching 1386 * stacks lest we corrupt the "NMI executing" variable. 1387 */ 1388 1389 swapgs 1390 cld 1391 SWITCH_TO_KERNEL_CR3 scratch_reg=%rdx 1392 movq %rsp, %rdx 1393 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 1394 UNWIND_HINT_IRET_REGS base=%rdx offset=8 1395 pushq 5*8(%rdx) /* pt_regs->ss */ 1396 pushq 4*8(%rdx) /* pt_regs->rsp */ 1397 pushq 3*8(%rdx) /* pt_regs->flags */ 1398 pushq 2*8(%rdx) /* pt_regs->cs */ 1399 pushq 1*8(%rdx) /* pt_regs->rip */ 1400 UNWIND_HINT_IRET_REGS 1401 pushq $-1 /* pt_regs->orig_ax */ 1402 PUSH_AND_CLEAR_REGS rdx=(%rdx) 1403 ENCODE_FRAME_POINTER 1404 1405 /* 1406 * At this point we no longer need to worry about stack damage 1407 * due to nesting -- we're on the normal thread stack and we're 1408 * done with the NMI stack. 1409 */ 1410 1411 movq %rsp, %rdi 1412 movq $-1, %rsi 1413 call do_nmi 1414 1415 /* 1416 * Return back to user mode. We must *not* do the normal exit 1417 * work, because we don't want to enable interrupts. 1418 */ 1419 jmp swapgs_restore_regs_and_return_to_usermode 1420 1421.Lnmi_from_kernel: 1422 /* 1423 * Here's what our stack frame will look like: 1424 * +---------------------------------------------------------+ 1425 * | original SS | 1426 * | original Return RSP | 1427 * | original RFLAGS | 1428 * | original CS | 1429 * | original RIP | 1430 * +---------------------------------------------------------+ 1431 * | temp storage for rdx | 1432 * +---------------------------------------------------------+ 1433 * | "NMI executing" variable | 1434 * +---------------------------------------------------------+ 1435 * | iret SS } Copied from "outermost" frame | 1436 * | iret Return RSP } on each loop iteration; overwritten | 1437 * | iret RFLAGS } by a nested NMI to force another | 1438 * | iret CS } iteration if needed. | 1439 * | iret RIP } | 1440 * +---------------------------------------------------------+ 1441 * | outermost SS } initialized in first_nmi; | 1442 * | outermost Return RSP } will not be changed before | 1443 * | outermost RFLAGS } NMI processing is done. | 1444 * | outermost CS } Copied to "iret" frame on each | 1445 * | outermost RIP } iteration. | 1446 * +---------------------------------------------------------+ 1447 * | pt_regs | 1448 * +---------------------------------------------------------+ 1449 * 1450 * The "original" frame is used by hardware. Before re-enabling 1451 * NMIs, we need to be done with it, and we need to leave enough 1452 * space for the asm code here. 1453 * 1454 * We return by executing IRET while RSP points to the "iret" frame. 1455 * That will either return for real or it will loop back into NMI 1456 * processing. 1457 * 1458 * The "outermost" frame is copied to the "iret" frame on each 1459 * iteration of the loop, so each iteration starts with the "iret" 1460 * frame pointing to the final return target. 1461 */ 1462 1463 /* 1464 * Determine whether we're a nested NMI. 1465 * 1466 * If we interrupted kernel code between repeat_nmi and 1467 * end_repeat_nmi, then we are a nested NMI. We must not 1468 * modify the "iret" frame because it's being written by 1469 * the outer NMI. That's okay; the outer NMI handler is 1470 * about to about to call do_nmi anyway, so we can just 1471 * resume the outer NMI. 1472 */ 1473 1474 movq $repeat_nmi, %rdx 1475 cmpq 8(%rsp), %rdx 1476 ja 1f 1477 movq $end_repeat_nmi, %rdx 1478 cmpq 8(%rsp), %rdx 1479 ja nested_nmi_out 14801: 1481 1482 /* 1483 * Now check "NMI executing". If it's set, then we're nested. 1484 * This will not detect if we interrupted an outer NMI just 1485 * before IRET. 1486 */ 1487 cmpl $1, -8(%rsp) 1488 je nested_nmi 1489 1490 /* 1491 * Now test if the previous stack was an NMI stack. This covers 1492 * the case where we interrupt an outer NMI after it clears 1493 * "NMI executing" but before IRET. We need to be careful, though: 1494 * there is one case in which RSP could point to the NMI stack 1495 * despite there being no NMI active: naughty userspace controls 1496 * RSP at the very beginning of the SYSCALL targets. We can 1497 * pull a fast one on naughty userspace, though: we program 1498 * SYSCALL to mask DF, so userspace cannot cause DF to be set 1499 * if it controls the kernel's RSP. We set DF before we clear 1500 * "NMI executing". 1501 */ 1502 lea 6*8(%rsp), %rdx 1503 /* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */ 1504 cmpq %rdx, 4*8(%rsp) 1505 /* If the stack pointer is above the NMI stack, this is a normal NMI */ 1506 ja first_nmi 1507 1508 subq $EXCEPTION_STKSZ, %rdx 1509 cmpq %rdx, 4*8(%rsp) 1510 /* If it is below the NMI stack, it is a normal NMI */ 1511 jb first_nmi 1512 1513 /* Ah, it is within the NMI stack. */ 1514 1515 testb $(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp) 1516 jz first_nmi /* RSP was user controlled. */ 1517 1518 /* This is a nested NMI. */ 1519 1520nested_nmi: 1521 /* 1522 * Modify the "iret" frame to point to repeat_nmi, forcing another 1523 * iteration of NMI handling. 1524 */ 1525 subq $8, %rsp 1526 leaq -10*8(%rsp), %rdx 1527 pushq $__KERNEL_DS 1528 pushq %rdx 1529 pushfq 1530 pushq $__KERNEL_CS 1531 pushq $repeat_nmi 1532 1533 /* Put stack back */ 1534 addq $(6*8), %rsp 1535 1536nested_nmi_out: 1537 popq %rdx 1538 1539 /* We are returning to kernel mode, so this cannot result in a fault. */ 1540 iretq 1541 1542first_nmi: 1543 /* Restore rdx. */ 1544 movq (%rsp), %rdx 1545 1546 /* Make room for "NMI executing". */ 1547 pushq $0 1548 1549 /* Leave room for the "iret" frame */ 1550 subq $(5*8), %rsp 1551 1552 /* Copy the "original" frame to the "outermost" frame */ 1553 .rept 5 1554 pushq 11*8(%rsp) 1555 .endr 1556 UNWIND_HINT_IRET_REGS 1557 1558 /* Everything up to here is safe from nested NMIs */ 1559 1560#ifdef CONFIG_DEBUG_ENTRY 1561 /* 1562 * For ease of testing, unmask NMIs right away. Disabled by 1563 * default because IRET is very expensive. 1564 */ 1565 pushq $0 /* SS */ 1566 pushq %rsp /* RSP (minus 8 because of the previous push) */ 1567 addq $8, (%rsp) /* Fix up RSP */ 1568 pushfq /* RFLAGS */ 1569 pushq $__KERNEL_CS /* CS */ 1570 pushq $1f /* RIP */ 1571 iretq /* continues at repeat_nmi below */ 1572 UNWIND_HINT_IRET_REGS 15731: 1574#endif 1575 1576repeat_nmi: 1577 /* 1578 * If there was a nested NMI, the first NMI's iret will return 1579 * here. But NMIs are still enabled and we can take another 1580 * nested NMI. The nested NMI checks the interrupted RIP to see 1581 * if it is between repeat_nmi and end_repeat_nmi, and if so 1582 * it will just return, as we are about to repeat an NMI anyway. 1583 * This makes it safe to copy to the stack frame that a nested 1584 * NMI will update. 1585 * 1586 * RSP is pointing to "outermost RIP". gsbase is unknown, but, if 1587 * we're repeating an NMI, gsbase has the same value that it had on 1588 * the first iteration. paranoid_entry will load the kernel 1589 * gsbase if needed before we call do_nmi. "NMI executing" 1590 * is zero. 1591 */ 1592 movq $1, 10*8(%rsp) /* Set "NMI executing". */ 1593 1594 /* 1595 * Copy the "outermost" frame to the "iret" frame. NMIs that nest 1596 * here must not modify the "iret" frame while we're writing to 1597 * it or it will end up containing garbage. 1598 */ 1599 addq $(10*8), %rsp 1600 .rept 5 1601 pushq -6*8(%rsp) 1602 .endr 1603 subq $(5*8), %rsp 1604end_repeat_nmi: 1605 1606 /* 1607 * Everything below this point can be preempted by a nested NMI. 1608 * If this happens, then the inner NMI will change the "iret" 1609 * frame to point back to repeat_nmi. 1610 */ 1611 pushq $-1 /* ORIG_RAX: no syscall to restart */ 1612 1613 /* 1614 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit 1615 * as we should not be calling schedule in NMI context. 1616 * Even with normal interrupts enabled. An NMI should not be 1617 * setting NEED_RESCHED or anything that normal interrupts and 1618 * exceptions might do. 1619 */ 1620 call paranoid_entry 1621 UNWIND_HINT_REGS 1622 1623 /* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */ 1624 movq %rsp, %rdi 1625 movq $-1, %rsi 1626 call do_nmi 1627 1628 RESTORE_CR3 scratch_reg=%r15 save_reg=%r14 1629 1630 testl %ebx, %ebx /* swapgs needed? */ 1631 jnz nmi_restore 1632nmi_swapgs: 1633 SWAPGS_UNSAFE_STACK 1634nmi_restore: 1635 POP_REGS 1636 1637 /* 1638 * Skip orig_ax and the "outermost" frame to point RSP at the "iret" 1639 * at the "iret" frame. 1640 */ 1641 addq $6*8, %rsp 1642 1643 /* 1644 * Clear "NMI executing". Set DF first so that we can easily 1645 * distinguish the remaining code between here and IRET from 1646 * the SYSCALL entry and exit paths. 1647 * 1648 * We arguably should just inspect RIP instead, but I (Andy) wrote 1649 * this code when I had the misapprehension that Xen PV supported 1650 * NMIs, and Xen PV would break that approach. 1651 */ 1652 std 1653 movq $0, 5*8(%rsp) /* clear "NMI executing" */ 1654 1655 /* 1656 * iretq reads the "iret" frame and exits the NMI stack in a 1657 * single instruction. We are returning to kernel mode, so this 1658 * cannot result in a fault. Similarly, we don't need to worry 1659 * about espfix64 on the way back to kernel mode. 1660 */ 1661 iretq 1662END(nmi) 1663 1664ENTRY(ignore_sysret) 1665 UNWIND_HINT_EMPTY 1666 mov $-ENOSYS, %eax 1667 sysret 1668END(ignore_sysret) 1669 1670ENTRY(rewind_stack_do_exit) 1671 UNWIND_HINT_FUNC 1672 /* Prevent any naive code from trying to unwind to our caller. */ 1673 xorl %ebp, %ebp 1674 1675 movq PER_CPU_VAR(cpu_current_top_of_stack), %rax 1676 leaq -PTREGS_SIZE(%rax), %rsp 1677 UNWIND_HINT_FUNC sp_offset=PTREGS_SIZE 1678 1679 call do_exit 1680END(rewind_stack_do_exit) 1681