arch/x86/kernel/entry_64.S

   1 /*
   2  *  linux/arch/x86_64/entry.S
   3  *
   4  *  Copyright (C) 1991, 1992  Linus Torvalds
   5  *  Copyright (C) 2000, 2001, 2002  Andi Kleen SuSE Labs
   6  *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
   7  */
   8
   9 /*
  10  * entry.S contains the system-call and fault low-level handling routines.
  11  *
  12  * NOTE: This code handles signal-recognition, which happens every time
  13  * after an interrupt and after each system call.
  14  *
  15  * Normal syscalls and interrupts don't save a full stack frame, this is
  16  * only done for syscall tracing, signals or fork/exec et.al.
  17  *
  18  * A note on terminology:
  19  * - top of stack: Architecture defined interrupt frame from SS to RIP
  20  * at the top of the kernel process stack.
  21  * - partial stack frame: partially saved registers upto R11.
  22  * - full stack frame: Like partial stack frame, but all register saved.
  23  *
  24  * Some macro usage:
  25  * - CFI macros are used to generate dwarf2 unwind information for better
  26  * backtraces. They don't change any code.
  27  * - SAVE_ALL/RESTORE_ALL - Save/restore all registers
  28  * - SAVE_ARGS/RESTORE_ARGS - Save/restore registers that C functions modify.
  29  * There are unfortunately lots of special cases where some registers
  30  * not touched. The macro is a big mess that should be cleaned up.
  31  * - SAVE_REST/RESTORE_REST - Handle the registers not saved by SAVE_ARGS.
  32  * Gives a full stack frame.
  33  * - ENTRY/END Define functions in the symbol table.
  34  * - FIXUP_TOP_OF_STACK/RESTORE_TOP_OF_STACK - Fix up the hardware stack
  35  * frame that is otherwise undefined after a SYSCALL
  36  * - TRACE_IRQ_* - Trace hard interrupt state for lock debugging.
  37  * - errorentry/paranoidentry/zeroentry - Define exception entry points.
  38  */
  39
  40 #include <linux/linkage.h>
  41 #include <asm/segment.h>
  42 #include <asm/cache.h>
  43 #include <asm/errno.h>
  44 #include <asm/dwarf2.h>
  45 #include <asm/calling.h>
  46 #include <asm/asm-offsets.h>
  47 #include <asm/msr.h>
  48 #include <asm/unistd.h>
  49 #include <asm/thread_info.h>
  50 #include <asm/hw_irq.h>
  51 #include <asm/page.h>
  52 #include <asm/irqflags.h>
  53
  54         .code64
  55
  56 #ifndef CONFIG_PREEMPT
  57 #define retint_kernel retint_restore_args
  58 #endif
  59
  60
  61 .macro TRACE_IRQS_IRETQ offset=ARGOFFSET
  62 #ifdef CONFIG_TRACE_IRQFLAGS
  63         bt   $9,EFLAGS-\offset(%rsp)    /* interrupts off? */
  64         jnc  1f
  65         TRACE_IRQS_ON
  66 1:
  67 #endif
  68 .endm
  69
  70 /*
  71  * C code is not supposed to know about undefined top of stack. Every time
  72  * a C function with an pt_regs argument is called from the SYSCALL based
  73  * fast path FIXUP_TOP_OF_STACK is needed.
  74  * RESTORE_TOP_OF_STACK syncs the syscall state after any possible ptregs
  75  * manipulation.
  76  */
  77
  78         /* %rsp:at FRAMEEND */
  79         .macro FIXUP_TOP_OF_STACK tmp
  80         movq    %gs:pda_oldrsp,\tmp
  81         movq    \tmp,RSP(%rsp)
  82         movq    $__USER_DS,SS(%rsp)
  83         movq    $__USER_CS,CS(%rsp)
  84         movq    $-1,RCX(%rsp)
  85         movq    R11(%rsp),\tmp  /* get eflags */
  86         movq    \tmp,EFLAGS(%rsp)
  87         .endm
  88
  89         .macro RESTORE_TOP_OF_STACK tmp,offset=0
  90         movq   RSP-\offset(%rsp),\tmp
  91         movq   \tmp,%gs:pda_oldrsp
  92         movq   EFLAGS-\offset(%rsp),\tmp
  93         movq   \tmp,R11-\offset(%rsp)
  94         .endm
  95
  96         .macro FAKE_STACK_FRAME child_rip
  97         /* push in order ss, rsp, eflags, cs, rip */
  98         xorl %eax, %eax
  99         pushq %rax /* ss */
 100         CFI_ADJUST_CFA_OFFSET   8
 101         /*CFI_REL_OFFSET        ss,0*/
 102         pushq %rax /* rsp */
 103         CFI_ADJUST_CFA_OFFSET   8
 104         CFI_REL_OFFSET  rsp,0
 105         pushq $(1<<9) /* eflags - interrupts on */
 106         CFI_ADJUST_CFA_OFFSET   8
 107         /*CFI_REL_OFFSET        rflags,0*/
 108         pushq $__KERNEL_CS /* cs */
 109         CFI_ADJUST_CFA_OFFSET   8
 110         /*CFI_REL_OFFSET        cs,0*/
 111         pushq \child_rip /* rip */
 112         CFI_ADJUST_CFA_OFFSET   8
 113         CFI_REL_OFFSET  rip,0
 114         pushq   %rax /* orig rax */
 115         CFI_ADJUST_CFA_OFFSET   8
 116         .endm
 117
 118         .macro UNFAKE_STACK_FRAME
 119         addq $8*6, %rsp
 120         CFI_ADJUST_CFA_OFFSET   -(6*8)
 121         .endm
 122
 123         .macro  CFI_DEFAULT_STACK start=1
 124         .if \start
 125         CFI_STARTPROC   simple
 126         CFI_SIGNAL_FRAME
 127         CFI_DEF_CFA     rsp,SS+8
 128         .else
 129         CFI_DEF_CFA_OFFSET SS+8
 130         .endif
 131         CFI_REL_OFFSET  r15,R15
 132         CFI_REL_OFFSET  r14,R14
 133         CFI_REL_OFFSET  r13,R13
 134         CFI_REL_OFFSET  r12,R12
 135         CFI_REL_OFFSET  rbp,RBP
 136         CFI_REL_OFFSET  rbx,RBX
 137         CFI_REL_OFFSET  r11,R11
 138         CFI_REL_OFFSET  r10,R10
 139         CFI_REL_OFFSET  r9,R9
 140         CFI_REL_OFFSET  r8,R8
 141         CFI_REL_OFFSET  rax,RAX
 142         CFI_REL_OFFSET  rcx,RCX
 143         CFI_REL_OFFSET  rdx,RDX
 144         CFI_REL_OFFSET  rsi,RSI
 145         CFI_REL_OFFSET  rdi,RDI
 146         CFI_REL_OFFSET  rip,RIP
 147         /*CFI_REL_OFFSET        cs,CS*/
 148         /*CFI_REL_OFFSET        rflags,EFLAGS*/
 149         CFI_REL_OFFSET  rsp,RSP
 150         /*CFI_REL_OFFSET        ss,SS*/
 151         .endm
 152 /*
 153  * A newly forked process directly context switches into this.
 154  */
 155 /* rdi: prev */
 156 ENTRY(ret_from_fork)
 157         CFI_DEFAULT_STACK
 158         push kernel_eflags(%rip)
 159         CFI_ADJUST_CFA_OFFSET 4
 160         popf                            # reset kernel eflags
 161         CFI_ADJUST_CFA_OFFSET -4
 162         call schedule_tail
 163         GET_THREAD_INFO(%rcx)
 164         testl $(_TIF_SYSCALL_TRACE|_TIF_SYSCALL_AUDIT),threadinfo_flags(%rcx)
 165         jnz rff_trace
 166 rff_action:
 167         RESTORE_REST
 168         testl $3,CS-ARGOFFSET(%rsp)     # from kernel_thread?
 169         je   int_ret_from_sys_call
 170         testl $_TIF_IA32,threadinfo_flags(%rcx)
 171         jnz  int_ret_from_sys_call
 172         RESTORE_TOP_OF_STACK %rdi,ARGOFFSET
 173         jmp ret_from_sys_call
 174 rff_trace:
 175         movq %rsp,%rdi
 176         call syscall_trace_leave
 177         GET_THREAD_INFO(%rcx)
 178         jmp rff_action
 179         CFI_ENDPROC
 180 END(ret_from_fork)
 181
 182 /*
 183  * System call entry. Upto 6 arguments in registers are supported.
 184  *
 185  * SYSCALL does not save anything on the stack and does not change the
 186  * stack pointer.
 187  */
 188
 189 /*
 190  * Register setup:
 191  * rax  system call number
 192  * rdi  arg0
 193  * rcx  return address for syscall/sysret, C arg3
 194  * rsi  arg1
 195  * rdx  arg2
 196  * r10  arg3    (--> moved to rcx for C)
 197  * r8   arg4
 198  * r9   arg5
 199  * r11  eflags for syscall/sysret, temporary for C
 200  * r12-r15,rbp,rbx saved by C code, not touched.
 201  *
 202  * Interrupts are off on entry.
 203  * Only called from user space.
 204  *
 205  * XXX  if we had a free scratch register we could save the RSP into the stack frame
 206  *      and report it properly in ps. Unfortunately we haven't.
 207  *
 208  * When user can change the frames always force IRET. That is because
 209  * it deals with uncanonical addresses better. SYSRET has trouble
 210  * with them due to bugs in both AMD and Intel CPUs.
 211  */
 212
 213 ENTRY(system_call)
 214         CFI_STARTPROC   simple
 215         CFI_SIGNAL_FRAME
 216         CFI_DEF_CFA     rsp,PDA_STACKOFFSET
 217         CFI_REGISTER    rip,rcx
 218         /*CFI_REGISTER  rflags,r11*/
 219         swapgs
 220         movq    %rsp,%gs:pda_oldrsp
 221         movq    %gs:pda_kernelstack,%rsp
 222         /*
 223          * No need to follow this irqs off/on section - it's straight
 224          * and short:
 225          */
 226         sti
 227         SAVE_ARGS 8,1
 228         movq  %rax,ORIG_RAX-ARGOFFSET(%rsp)
 229         movq  %rcx,RIP-ARGOFFSET(%rsp)
 230         CFI_REL_OFFSET rip,RIP-ARGOFFSET
 231         GET_THREAD_INFO(%rcx)
 232         testl $(_TIF_SYSCALL_TRACE|_TIF_SYSCALL_AUDIT|_TIF_SECCOMP),threadinfo_flags(%rcx)
 233         jnz tracesys
 234         cmpq $__NR_syscall_max,%rax
 235         ja badsys
 236         movq %r10,%rcx
 237         call *sys_call_table(,%rax,8)  # XXX:    rip relative
 238         movq %rax,RAX-ARGOFFSET(%rsp)
 239 /*
 240  * Syscall return path ending with SYSRET (fast path)
 241  * Has incomplete stack frame and undefined top of stack.
 242  */
 243 ret_from_sys_call:
 244         movl $_TIF_ALLWORK_MASK,%edi
 245         /* edi: flagmask */
 246 sysret_check:
 247         LOCKDEP_SYS_EXIT
 248         GET_THREAD_INFO(%rcx)
 249         cli
 250         TRACE_IRQS_OFF
 251         movl threadinfo_flags(%rcx),%edx
 252         andl %edi,%edx
 253         jnz  sysret_careful
 254         CFI_REMEMBER_STATE
 255         /*
 256          * sysretq will re-enable interrupts:
 257          */
 258         TRACE_IRQS_ON
 259         movq RIP-ARGOFFSET(%rsp),%rcx
 260         CFI_REGISTER    rip,rcx
 261         RESTORE_ARGS 0,-ARG_SKIP,1
 262         /*CFI_REGISTER  rflags,r11*/
 263         movq    %gs:pda_oldrsp,%rsp
 264         swapgs
 265         sysretq
 266
 267         CFI_RESTORE_STATE
 268         /* Handle reschedules */
 269         /* edx: work, edi: workmask */
 270 sysret_careful:
 271         bt $TIF_NEED_RESCHED,%edx
 272         jnc sysret_signal
 273         TRACE_IRQS_ON
 274         sti
 275         pushq %rdi
 276         CFI_ADJUST_CFA_OFFSET 8
 277         call schedule
 278         popq  %rdi
 279         CFI_ADJUST_CFA_OFFSET -8
 280         jmp sysret_check
 281
 282         /* Handle a signal */
 283 sysret_signal:
 284         TRACE_IRQS_ON
 285         sti
 286         testl $_TIF_DO_NOTIFY_MASK,%edx
 287         jz    1f
 288
 289         /* Really a signal */
 290         /* edx: work flags (arg3) */
 291         leaq do_notify_resume(%rip),%rax
 292         leaq -ARGOFFSET(%rsp),%rdi # &pt_regs -> arg1
 293         xorl %esi,%esi # oldset -> arg2
 294         call ptregscall_common
 295 1:      movl $_TIF_NEED_RESCHED,%edi
 296         /* Use IRET because user could have changed frame. This
 297            works because ptregscall_common has called FIXUP_TOP_OF_STACK. */
 298         cli
 299         TRACE_IRQS_OFF
 300         jmp int_with_check
 301
 302 badsys:
 303         movq $-ENOSYS,RAX-ARGOFFSET(%rsp)
 304         jmp ret_from_sys_call
 305
 306         /* Do syscall tracing */
 307 tracesys:
 308         SAVE_REST
 309         movq $-ENOSYS,RAX(%rsp)
 310         FIXUP_TOP_OF_STACK %rdi
 311         movq %rsp,%rdi
 312         call syscall_trace_enter
 313         LOAD_ARGS ARGOFFSET  /* reload args from stack in case ptrace changed it */
 314         RESTORE_REST
 315         cmpq $__NR_syscall_max,%rax
 316         movq $-ENOSYS,%rcx
 317         cmova %rcx,%rax
 318         ja  1f
 319         movq %r10,%rcx  /* fixup for C */
 320         call *sys_call_table(,%rax,8)
 321 1:      movq %rax,RAX-ARGOFFSET(%rsp)
 322         /* Use IRET because user could have changed frame */
 323
 324 /*
 325  * Syscall return path ending with IRET.
 326  * Has correct top of stack, but partial stack frame.
 327  */
 328         .globl int_ret_from_sys_call
 329 int_ret_from_sys_call:
 330         cli
 331         TRACE_IRQS_OFF
 332         testl $3,CS-ARGOFFSET(%rsp)
 333         je retint_restore_args
 334         movl $_TIF_ALLWORK_MASK,%edi
 335         /* edi: mask to check */
 336 int_with_check:
 337         LOCKDEP_SYS_EXIT_IRQ
 338         GET_THREAD_INFO(%rcx)
 339         movl threadinfo_flags(%rcx),%edx
 340         andl %edi,%edx
 341         jnz   int_careful
 342         andl    $~TS_COMPAT,threadinfo_status(%rcx)
 343         jmp   retint_swapgs
 344
 345         /* Either reschedule or signal or syscall exit tracking needed. */
 346         /* First do a reschedule test. */
 347         /* edx: work, edi: workmask */
 348 int_careful:
 349         bt $TIF_NEED_RESCHED,%edx
 350         jnc  int_very_careful
 351         TRACE_IRQS_ON
 352         sti
 353         pushq %rdi
 354         CFI_ADJUST_CFA_OFFSET 8
 355         call schedule
 356         popq %rdi
 357         CFI_ADJUST_CFA_OFFSET -8
 358         cli
 359         TRACE_IRQS_OFF
 360         jmp int_with_check
 361
 362         /* handle signals and tracing -- both require a full stack frame */
 363 int_very_careful:
 364         TRACE_IRQS_ON
 365         sti
 366         SAVE_REST
 367         /* Check for syscall exit trace */
 368         testl $(_TIF_SYSCALL_TRACE|_TIF_SYSCALL_AUDIT|_TIF_SINGLESTEP),%edx
 369         jz int_signal
 370         pushq %rdi
 371         CFI_ADJUST_CFA_OFFSET 8
 372         leaq 8(%rsp),%rdi       # &ptregs -> arg1
 373         call syscall_trace_leave
 374         popq %rdi
 375         CFI_ADJUST_CFA_OFFSET -8
 376         andl $~(_TIF_SYSCALL_TRACE|_TIF_SYSCALL_AUDIT|_TIF_SINGLESTEP),%edi
 377         jmp int_restore_rest
 378
 379 int_signal:
 380         testl $_TIF_DO_NOTIFY_MASK,%edx
 381         jz 1f
 382         movq %rsp,%rdi          # &ptregs -> arg1
 383         xorl %esi,%esi          # oldset -> arg2
 384         call do_notify_resume
 385 1:      movl $_TIF_NEED_RESCHED,%edi
 386 int_restore_rest:
 387         RESTORE_REST
 388         cli
 389         TRACE_IRQS_OFF
 390         jmp int_with_check
 391         CFI_ENDPROC
 392 END(system_call)
 393
 394 /*
 395  * Certain special system calls that need to save a complete full stack frame.
 396  */
 397
 398         .macro PTREGSCALL label,func,arg
 399         .globl \label
 400 \label:
 401         leaq    \func(%rip),%rax
 402         leaq    -ARGOFFSET+8(%rsp),\arg /* 8 for return address */
 403         jmp     ptregscall_common
 404 END(\label)
 405         .endm
 406
 407         CFI_STARTPROC
 408
 409         PTREGSCALL stub_clone, sys_clone, %r8
 410         PTREGSCALL stub_fork, sys_fork, %rdi
 411         PTREGSCALL stub_vfork, sys_vfork, %rdi
 412         PTREGSCALL stub_rt_sigsuspend, sys_rt_sigsuspend, %rdx
 413         PTREGSCALL stub_sigaltstack, sys_sigaltstack, %rdx
 414         PTREGSCALL stub_iopl, sys_iopl, %rsi
 415
 416 ENTRY(ptregscall_common)
 417         popq %r11
 418         CFI_ADJUST_CFA_OFFSET -8
 419         CFI_REGISTER rip, r11
 420         SAVE_REST
 421         movq %r11, %r15
 422         CFI_REGISTER rip, r15
 423         FIXUP_TOP_OF_STACK %r11
 424         call *%rax
 425         RESTORE_TOP_OF_STACK %r11
 426         movq %r15, %r11
 427         CFI_REGISTER rip, r11
 428         RESTORE_REST
 429         pushq %r11
 430         CFI_ADJUST_CFA_OFFSET 8
 431         CFI_REL_OFFSET rip, 0
 432         ret
 433         CFI_ENDPROC
 434 END(ptregscall_common)
 435
 436 ENTRY(stub_execve)
 437         CFI_STARTPROC
 438         popq %r11
 439         CFI_ADJUST_CFA_OFFSET -8
 440         CFI_REGISTER rip, r11
 441         SAVE_REST
 442         FIXUP_TOP_OF_STACK %r11
 443         call sys_execve
 444         RESTORE_TOP_OF_STACK %r11
 445         movq %rax,RAX(%rsp)
 446         RESTORE_REST
 447         jmp int_ret_from_sys_call
 448         CFI_ENDPROC
 449 END(stub_execve)
 450
 451 /*
 452  * sigreturn is special because it needs to restore all registers on return.
 453  * This cannot be done with SYSRET, so use the IRET return path instead.
 454  */
 455 ENTRY(stub_rt_sigreturn)
 456         CFI_STARTPROC
 457         addq $8, %rsp
 458         CFI_ADJUST_CFA_OFFSET   -8
 459         SAVE_REST
 460         movq %rsp,%rdi
 461         FIXUP_TOP_OF_STACK %r11
 462         call sys_rt_sigreturn
 463         movq %rax,RAX(%rsp) # fixme, this could be done at the higher layer
 464         RESTORE_REST
 465         jmp int_ret_from_sys_call
 466         CFI_ENDPROC
 467 END(stub_rt_sigreturn)
 468
 469 /*
 470  * initial frame state for interrupts and exceptions
 471  */
 472         .macro _frame ref
 473         CFI_STARTPROC simple
 474         CFI_SIGNAL_FRAME
 475         CFI_DEF_CFA rsp,SS+8-\ref
 476         /*CFI_REL_OFFSET ss,SS-\ref*/
 477         CFI_REL_OFFSET rsp,RSP-\ref
 478         /*CFI_REL_OFFSET rflags,EFLAGS-\ref*/
 479         /*CFI_REL_OFFSET cs,CS-\ref*/
 480         CFI_REL_OFFSET rip,RIP-\ref
 481         .endm
 482
 483 /* initial frame state for interrupts (and exceptions without error code) */
 484 #define INTR_FRAME _frame RIP
 485 /* initial frame state for exceptions with error code (and interrupts with
 486    vector already pushed) */
 487 #define XCPT_FRAME _frame ORIG_RAX
 488
 489 /*
 490  * Interrupt entry/exit.
 491  *
 492  * Interrupt entry points save only callee clobbered registers in fast path.
 493  *
 494  * Entry runs with interrupts off.
 495  */
 496
 497 /* 0(%rsp): interrupt number */
 498         .macro interrupt func
 499         cld
 500         SAVE_ARGS
 501         leaq -ARGOFFSET(%rsp),%rdi      # arg1 for handler
 502         pushq %rbp
 503         CFI_ADJUST_CFA_OFFSET   8
 504         CFI_REL_OFFSET          rbp, 0
 505         movq %rsp,%rbp
 506         CFI_DEF_CFA_REGISTER    rbp
 507         testl $3,CS(%rdi)
 508         je 1f
 509         swapgs
 510         /* irqcount is used to check if a CPU is already on an interrupt
 511            stack or not. While this is essentially redundant with preempt_count
 512            it is a little cheaper to use a separate counter in the PDA
 513            (short of moving irq_enter into assembly, which would be too
 514             much work) */
 515 1:      incl    %gs:pda_irqcount
 516         cmoveq %gs:pda_irqstackptr,%rsp
 517         push    %rbp                    # backlink for old unwinder
 518         /*
 519          * We entered an interrupt context - irqs are off:
 520          */
 521         TRACE_IRQS_OFF
 522         call \func
 523         .endm
 524
 525 ENTRY(common_interrupt)
 526         XCPT_FRAME
 527         interrupt do_IRQ
 528         /* 0(%rsp): oldrsp-ARGOFFSET */
 529 ret_from_intr:
 530         cli
 531         TRACE_IRQS_OFF
 532         decl %gs:pda_irqcount
 533         leaveq
 534         CFI_DEF_CFA_REGISTER    rsp
 535         CFI_ADJUST_CFA_OFFSET   -8
 536 exit_intr:
 537         GET_THREAD_INFO(%rcx)
 538         testl $3,CS-ARGOFFSET(%rsp)
 539         je retint_kernel
 540
 541         /* Interrupt came from user space */
 542         /*
 543          * Has a correct top of stack, but a partial stack frame
 544          * %rcx: thread info. Interrupts off.
 545          */
 546 retint_with_reschedule:
 547         movl $_TIF_WORK_MASK,%edi
 548 retint_check:
 549         LOCKDEP_SYS_EXIT_IRQ
 550         movl threadinfo_flags(%rcx),%edx
 551         andl %edi,%edx
 552         CFI_REMEMBER_STATE
 553         jnz  retint_careful
 554
 555 retint_swapgs:          /* return to user-space */
 556         /*
 557          * The iretq could re-enable interrupts:
 558          */
 559         cli
 560         TRACE_IRQS_IRETQ
 561         swapgs
 562         jmp restore_args
 563
 564 retint_restore_args:    /* return to kernel space */
 565         cli
 566         /*
 567          * The iretq could re-enable interrupts:
 568          */
 569         TRACE_IRQS_IRETQ
 570 restore_args:
 571         RESTORE_ARGS 0,8,0
 572 iret_label:
 573         iretq
 574
 575         .section __ex_table,"a"
 576         .quad iret_label,bad_iret
 577         .previous
 578         .section .fixup,"ax"
 579         /* force a signal here? this matches i386 behaviour */
 580         /* running with kernel gs */
 581 bad_iret:
 582         movq $11,%rdi   /* SIGSEGV */
 583         TRACE_IRQS_ON
 584         sti
 585         jmp do_exit
 586         .previous
 587
 588         /* edi: workmask, edx: work */
 589 retint_careful:
 590         CFI_RESTORE_STATE
 591         bt    $TIF_NEED_RESCHED,%edx
 592         jnc   retint_signal
 593         TRACE_IRQS_ON
 594         sti
 595         pushq %rdi
 596         CFI_ADJUST_CFA_OFFSET   8
 597         call  schedule
 598         popq %rdi
 599         CFI_ADJUST_CFA_OFFSET   -8
 600         GET_THREAD_INFO(%rcx)
 601         cli
 602         TRACE_IRQS_OFF
 603         jmp retint_check
 604
 605 retint_signal:
 606         testl $_TIF_DO_NOTIFY_MASK,%edx
 607         jz    retint_swapgs
 608         TRACE_IRQS_ON
 609         sti
 610         SAVE_REST
 611         movq $-1,ORIG_RAX(%rsp)
 612         xorl %esi,%esi          # oldset
 613         movq %rsp,%rdi          # &pt_regs
 614         call do_notify_resume
 615         RESTORE_REST
 616         cli
 617         TRACE_IRQS_OFF
 618         movl $_TIF_NEED_RESCHED,%edi
 619         GET_THREAD_INFO(%rcx)
 620         jmp retint_check
 621
 622 #ifdef CONFIG_PREEMPT
 623         /* Returning to kernel space. Check if we need preemption */
 624         /* rcx:  threadinfo. interrupts off. */
 625 ENTRY(retint_kernel)
 626         cmpl $0,threadinfo_preempt_count(%rcx)
 627         jnz  retint_restore_args
 628         bt  $TIF_NEED_RESCHED,threadinfo_flags(%rcx)
 629         jnc  retint_restore_args
 630         bt   $9,EFLAGS-ARGOFFSET(%rsp)  /* interrupts off? */
 631         jnc  retint_restore_args
 632         call preempt_schedule_irq
 633         jmp exit_intr
 634 #endif
 635
 636         CFI_ENDPROC
 637 END(common_interrupt)
 638
 639 /*
 640  * APIC interrupts.
 641  */
 642         .macro apicinterrupt num,func
 643         INTR_FRAME
 644         pushq $~(\num)
 645         CFI_ADJUST_CFA_OFFSET 8
 646         interrupt \func
 647         jmp ret_from_intr
 648         CFI_ENDPROC
 649         .endm
 650
 651 ENTRY(thermal_interrupt)
 652         apicinterrupt THERMAL_APIC_VECTOR,smp_thermal_interrupt
 653 END(thermal_interrupt)
 654
 655 ENTRY(threshold_interrupt)
 656         apicinterrupt THRESHOLD_APIC_VECTOR,mce_threshold_interrupt
 657 END(threshold_interrupt)
 658
 659 #ifdef CONFIG_SMP
 660 ENTRY(reschedule_interrupt)
 661         apicinterrupt RESCHEDULE_VECTOR,smp_reschedule_interrupt
 662 END(reschedule_interrupt)
 663
 664         .macro INVALIDATE_ENTRY num
 665 ENTRY(invalidate_interrupt\num)
 666         apicinterrupt INVALIDATE_TLB_VECTOR_START+\num,smp_invalidate_interrupt
 667 END(invalidate_interrupt\num)
 668         .endm
 669
 670         INVALIDATE_ENTRY 0
 671         INVALIDATE_ENTRY 1
 672         INVALIDATE_ENTRY 2
 673         INVALIDATE_ENTRY 3
 674         INVALIDATE_ENTRY 4
 675         INVALIDATE_ENTRY 5
 676         INVALIDATE_ENTRY 6
 677         INVALIDATE_ENTRY 7
 678
 679 ENTRY(call_function_interrupt)
 680         apicinterrupt CALL_FUNCTION_VECTOR,smp_call_function_interrupt
 681 END(call_function_interrupt)
 682 ENTRY(irq_move_cleanup_interrupt)
 683         apicinterrupt IRQ_MOVE_CLEANUP_VECTOR,smp_irq_move_cleanup_interrupt
 684 END(irq_move_cleanup_interrupt)
 685 #endif
 686
 687 ENTRY(apic_timer_interrupt)
 688         apicinterrupt LOCAL_TIMER_VECTOR,smp_apic_timer_interrupt
 689 END(apic_timer_interrupt)
 690
 691 ENTRY(error_interrupt)
 692         apicinterrupt ERROR_APIC_VECTOR,smp_error_interrupt
 693 END(error_interrupt)
 694
 695 ENTRY(spurious_interrupt)
 696         apicinterrupt SPURIOUS_APIC_VECTOR,smp_spurious_interrupt
 697 END(spurious_interrupt)
 698
 699 /*
 700  * Exception entry points.
 701  */
 702         .macro zeroentry sym
 703         INTR_FRAME
 704         pushq $0        /* push error code/oldrax */
 705         CFI_ADJUST_CFA_OFFSET 8
 706         pushq %rax      /* push real oldrax to the rdi slot */
 707         CFI_ADJUST_CFA_OFFSET 8
 708         CFI_REL_OFFSET rax,0
 709         leaq  \sym(%rip),%rax
 710         jmp error_entry
 711         CFI_ENDPROC
 712         .endm
 713
 714         .macro errorentry sym
 715         XCPT_FRAME
 716         pushq %rax
 717         CFI_ADJUST_CFA_OFFSET 8
 718         CFI_REL_OFFSET rax,0
 719         leaq  \sym(%rip),%rax
 720         jmp error_entry
 721         CFI_ENDPROC
 722         .endm
 723
 724         /* error code is on the stack already */
 725         /* handle NMI like exceptions that can happen everywhere */
 726         .macro paranoidentry sym, ist=0, irqtrace=1
 727         SAVE_ALL
 728         cld
 729         movl $1,%ebx
 730         movl  $MSR_GS_BASE,%ecx
 731         rdmsr
 732         testl %edx,%edx
 733         js    1f
 734         swapgs
 735         xorl  %ebx,%ebx
 736 1:
 737         .if \ist
 738         movq    %gs:pda_data_offset, %rbp
 739         .endif
 740         movq %rsp,%rdi
 741         movq ORIG_RAX(%rsp),%rsi
 742         movq $-1,ORIG_RAX(%rsp)
 743         .if \ist
 744         subq    $EXCEPTION_STKSZ, per_cpu__init_tss + TSS_ist + (\ist - 1) * 8(%rbp)
 745         .endif
 746         call \sym
 747         .if \ist
 748         addq    $EXCEPTION_STKSZ, per_cpu__init_tss + TSS_ist + (\ist - 1) * 8(%rbp)
 749         .endif
 750         cli
 751         .if \irqtrace
 752         TRACE_IRQS_OFF
 753         .endif
 754         .endm
 755
 756         /*
 757          * "Paranoid" exit path from exception stack.
 758          * Paranoid because this is used by NMIs and cannot take
 759          * any kernel state for granted.
 760          * We don't do kernel preemption checks here, because only
 761          * NMI should be common and it does not enable IRQs and
 762          * cannot get reschedule ticks.
 763          *
 764          * "trace" is 0 for the NMI handler only, because irq-tracing
 765          * is fundamentally NMI-unsafe. (we cannot change the soft and
 766          * hard flags at once, atomically)
 767          */
 768         .macro paranoidexit trace=1
 769         /* ebx: no swapgs flag */
 770 paranoid_exit\trace:
 771         testl %ebx,%ebx                         /* swapgs needed? */
 772         jnz paranoid_restore\trace
 773         testl $3,CS(%rsp)
 774         jnz   paranoid_userspace\trace
 775 paranoid_swapgs\trace:
 776         .if \trace
 777         TRACE_IRQS_IRETQ 0
 778         .endif
 779         swapgs
 780 paranoid_restore\trace:
 781         RESTORE_ALL 8
 782         iretq
 783 paranoid_userspace\trace:
 784         GET_THREAD_INFO(%rcx)
 785         movl threadinfo_flags(%rcx),%ebx
 786         andl $_TIF_WORK_MASK,%ebx
 787         jz paranoid_swapgs\trace
 788         movq %rsp,%rdi                  /* &pt_regs */
 789         call sync_regs
 790         movq %rax,%rsp                  /* switch stack for scheduling */
 791         testl $_TIF_NEED_RESCHED,%ebx
 792         jnz paranoid_schedule\trace
 793         movl %ebx,%edx                  /* arg3: thread flags */
 794         .if \trace
 795         TRACE_IRQS_ON
 796         .endif
 797         sti
 798         xorl %esi,%esi                  /* arg2: oldset */
 799         movq %rsp,%rdi                  /* arg1: &pt_regs */
 800         call do_notify_resume
 801         cli
 802         .if \trace
 803         TRACE_IRQS_OFF
 804         .endif
 805         jmp paranoid_userspace\trace
 806 paranoid_schedule\trace:
 807         .if \trace
 808         TRACE_IRQS_ON
 809         .endif
 810         sti
 811         call schedule
 812         cli
 813         .if \trace
 814         TRACE_IRQS_OFF
 815         .endif
 816         jmp paranoid_userspace\trace
 817         CFI_ENDPROC
 818         .endm
 819
 820 /*
 821  * Exception entry point. This expects an error code/orig_rax on the stack
 822  * and the exception handler in %rax.
 823  */
 824 KPROBE_ENTRY(error_entry)
 825         _frame RDI
 826         CFI_REL_OFFSET rax,0
 827         /* rdi slot contains rax, oldrax contains error code */
 828         cld
 829         subq  $14*8,%rsp
 830         CFI_ADJUST_CFA_OFFSET   (14*8)
 831         movq %rsi,13*8(%rsp)
 832         CFI_REL_OFFSET  rsi,RSI
 833         movq 14*8(%rsp),%rsi    /* load rax from rdi slot */
 834         CFI_REGISTER    rax,rsi
 835         movq %rdx,12*8(%rsp)
 836         CFI_REL_OFFSET  rdx,RDX
 837         movq %rcx,11*8(%rsp)
 838         CFI_REL_OFFSET  rcx,RCX
 839         movq %rsi,10*8(%rsp)    /* store rax */
 840         CFI_REL_OFFSET  rax,RAX
 841         movq %r8, 9*8(%rsp)
 842         CFI_REL_OFFSET  r8,R8
 843         movq %r9, 8*8(%rsp)
 844         CFI_REL_OFFSET  r9,R9
 845         movq %r10,7*8(%rsp)
 846         CFI_REL_OFFSET  r10,R10
 847         movq %r11,6*8(%rsp)
 848         CFI_REL_OFFSET  r11,R11
 849         movq %rbx,5*8(%rsp)
 850         CFI_REL_OFFSET  rbx,RBX
 851         movq %rbp,4*8(%rsp)
 852         CFI_REL_OFFSET  rbp,RBP
 853         movq %r12,3*8(%rsp)
 854         CFI_REL_OFFSET  r12,R12
 855         movq %r13,2*8(%rsp)
 856         CFI_REL_OFFSET  r13,R13
 857         movq %r14,1*8(%rsp)
 858         CFI_REL_OFFSET  r14,R14
 859         movq %r15,(%rsp)
 860         CFI_REL_OFFSET  r15,R15
 861         xorl %ebx,%ebx
 862         testl $3,CS(%rsp)
 863         je  error_kernelspace
 864 error_swapgs:
 865         swapgs
 866 error_sti:
 867         movq %rdi,RDI(%rsp)
 868         CFI_REL_OFFSET  rdi,RDI
 869         movq %rsp,%rdi
 870         movq ORIG_RAX(%rsp),%rsi        /* get error code */
 871         movq $-1,ORIG_RAX(%rsp)
 872         call *%rax
 873         /* ebx: no swapgs flag (1: don't need swapgs, 0: need it) */
 874 error_exit:
 875         movl %ebx,%eax
 876         RESTORE_REST
 877         cli
 878         TRACE_IRQS_OFF
 879         GET_THREAD_INFO(%rcx)
 880         testl %eax,%eax
 881         jne  retint_kernel
 882         LOCKDEP_SYS_EXIT_IRQ
 883         movl  threadinfo_flags(%rcx),%edx
 884         movl  $_TIF_WORK_MASK,%edi
 885         andl  %edi,%edx
 886         jnz  retint_careful
 887         jmp retint_swapgs
 888         CFI_ENDPROC
 889
 890 error_kernelspace:
 891         incl %ebx
 892        /* There are two places in the kernel that can potentially fault with
 893           usergs. Handle them here. The exception handlers after
 894            iret run with kernel gs again, so don't set the user space flag.
 895            B stepping K8s sometimes report an truncated RIP for IRET
 896            exceptions returning to compat mode. Check for these here too. */
 897         leaq iret_label(%rip),%rbp
 898         cmpq %rbp,RIP(%rsp)
 899         je   error_swapgs
 900         movl %ebp,%ebp  /* zero extend */
 901         cmpq %rbp,RIP(%rsp)
 902         je   error_swapgs
 903         cmpq $gs_change,RIP(%rsp)
 904         je   error_swapgs
 905         jmp  error_sti
 906 KPROBE_END(error_entry)
 907
 908        /* Reload gs selector with exception handling */
 909        /* edi:  new selector */
 910 ENTRY(load_gs_index)
 911         CFI_STARTPROC
 912         pushf
 913         CFI_ADJUST_CFA_OFFSET 8
 914         cli
 915         swapgs
 916 gs_change:
 917         movl %edi,%gs
 918 2:      mfence          /* workaround */
 919         swapgs
 920         popf
 921         CFI_ADJUST_CFA_OFFSET -8
 922         ret
 923         CFI_ENDPROC
 924 ENDPROC(load_gs_index)
 925
 926         .section __ex_table,"a"
 927         .align 8
 928         .quad gs_change,bad_gs
 929         .previous
 930         .section .fixup,"ax"
 931         /* running with kernelgs */
 932 bad_gs:
 933         swapgs                  /* switch back to user gs */
 934         xorl %eax,%eax
 935         movl %eax,%gs
 936         jmp  2b
 937         .previous
 938
 939 /*
 940  * Create a kernel thread.
 941  *
 942  * C extern interface:
 943  *      extern long kernel_thread(int (*fn)(void *), void * arg, unsigned long flags)
 944  *
 945  * asm input arguments:
 946  *      rdi: fn, rsi: arg, rdx: flags
 947  */
 948 ENTRY(kernel_thread)
 949         CFI_STARTPROC
 950         FAKE_STACK_FRAME $child_rip
 951         SAVE_ALL
 952
 953         # rdi: flags, rsi: usp, rdx: will be &pt_regs
 954         movq %rdx,%rdi
 955         orq  kernel_thread_flags(%rip),%rdi
 956         movq $-1, %rsi
 957         movq %rsp, %rdx
 958
 959         xorl %r8d,%r8d
 960         xorl %r9d,%r9d
 961
 962         # clone now
 963         call do_fork
 964         movq %rax,RAX(%rsp)
 965         xorl %edi,%edi
 966
 967         /*
 968          * It isn't worth to check for reschedule here,
 969          * so internally to the x86_64 port you can rely on kernel_thread()
 970          * not to reschedule the child before returning, this avoids the need
 971          * of hacks for example to fork off the per-CPU idle tasks.
 972          * [Hopefully no generic code relies on the reschedule -AK]
 973          */
 974         RESTORE_ALL
 975         UNFAKE_STACK_FRAME
 976         ret
 977         CFI_ENDPROC
 978 ENDPROC(kernel_thread)
 979
 980 child_rip:
 981         pushq $0                # fake return address
 982         CFI_STARTPROC
 983         /*
 984          * Here we are in the child and the registers are set as they were
 985          * at kernel_thread() invocation in the parent.
 986          */
 987         movq %rdi, %rax
 988         movq %rsi, %rdi
 989         call *%rax
 990         # exit
 991         mov %eax, %edi
 992         call do_exit
 993         CFI_ENDPROC
 994 ENDPROC(child_rip)
 995
 996 /*
 997  * execve(). This function needs to use IRET, not SYSRET, to set up all state properly.
 998  *
 999  * C extern interface:
1000  *       extern long execve(char *name, char **argv, char **envp)
1001  *
1002  * asm input arguments:
1003  *      rdi: name, rsi: argv, rdx: envp
1004  *
1005  * We want to fallback into:
1006  *      extern long sys_execve(char *name, char **argv,char **envp, struct pt_regs regs)
1007  *
1008  * do_sys_execve asm fallback arguments:
1009  *      rdi: name, rsi: argv, rdx: envp, fake frame on the stack
1010  */
1011 ENTRY(kernel_execve)
1012         CFI_STARTPROC
1013         FAKE_STACK_FRAME $0
1014         SAVE_ALL
1015         call sys_execve
1016         movq %rax, RAX(%rsp)
1017         RESTORE_REST
1018         testq %rax,%rax
1019         je int_ret_from_sys_call
1020         RESTORE_ARGS
1021         UNFAKE_STACK_FRAME
1022         ret
1023         CFI_ENDPROC
1024 ENDPROC(kernel_execve)
1025
1026 KPROBE_ENTRY(page_fault)
1027         errorentry do_page_fault
1028 KPROBE_END(page_fault)
1029
1030 ENTRY(coprocessor_error)
1031         zeroentry do_coprocessor_error
1032 END(coprocessor_error)
1033
1034 ENTRY(simd_coprocessor_error)
1035         zeroentry do_simd_coprocessor_error
1036 END(simd_coprocessor_error)
1037
1038 ENTRY(device_not_available)
1039         zeroentry math_state_restore
1040 END(device_not_available)
1041
1042         /* runs on exception stack */
1043 KPROBE_ENTRY(debug)
1044         INTR_FRAME
1045         pushq $0
1046         CFI_ADJUST_CFA_OFFSET 8
1047         paranoidentry do_debug, DEBUG_STACK
1048         paranoidexit
1049 KPROBE_END(debug)
1050
1051         /* runs on exception stack */
1052 KPROBE_ENTRY(nmi)
1053         INTR_FRAME
1054         pushq $-1
1055         CFI_ADJUST_CFA_OFFSET 8
1056         paranoidentry do_nmi, 0, 0
1057 #ifdef CONFIG_TRACE_IRQFLAGS
1058         paranoidexit 0
1059 #else
1060         jmp paranoid_exit1
1061         CFI_ENDPROC
1062 #endif
1063 KPROBE_END(nmi)
1064
1065 KPROBE_ENTRY(int3)
1066         INTR_FRAME
1067         pushq $0
1068         CFI_ADJUST_CFA_OFFSET 8
1069         paranoidentry do_int3, DEBUG_STACK
1070         jmp paranoid_exit1
1071         CFI_ENDPROC
1072 KPROBE_END(int3)
1073
1074 ENTRY(overflow)
1075         zeroentry do_overflow
1076 END(overflow)
1077
1078 ENTRY(bounds)
1079         zeroentry do_bounds
1080 END(bounds)
1081
1082 ENTRY(invalid_op)
1083         zeroentry do_invalid_op
1084 END(invalid_op)
1085
1086 ENTRY(coprocessor_segment_overrun)
1087         zeroentry do_coprocessor_segment_overrun
1088 END(coprocessor_segment_overrun)
1089
1090 ENTRY(reserved)
1091         zeroentry do_reserved
1092 END(reserved)
1093
1094         /* runs on exception stack */
1095 ENTRY(double_fault)
1096         XCPT_FRAME
1097         paranoidentry do_double_fault
1098         jmp paranoid_exit1
1099         CFI_ENDPROC
1100 END(double_fault)
1101
1102 ENTRY(invalid_TSS)
1103         errorentry do_invalid_TSS
1104 END(invalid_TSS)
1105
1106 ENTRY(segment_not_present)
1107         errorentry do_segment_not_present
1108 END(segment_not_present)
1109
1110         /* runs on exception stack */
1111 ENTRY(stack_segment)
1112         XCPT_FRAME
1113         paranoidentry do_stack_segment
1114         jmp paranoid_exit1
1115         CFI_ENDPROC
1116 END(stack_segment)
1117
1118 KPROBE_ENTRY(general_protection)
1119         errorentry do_general_protection
1120 KPROBE_END(general_protection)
1121
1122 ENTRY(alignment_check)
1123         errorentry do_alignment_check
1124 END(alignment_check)
1125
1126 ENTRY(divide_error)
1127         zeroentry do_divide_error
1128 END(divide_error)
1129
1130 ENTRY(spurious_interrupt_bug)
1131         zeroentry do_spurious_interrupt_bug
1132 END(spurious_interrupt_bug)
1133
1134 #ifdef CONFIG_X86_MCE
1135         /* runs on exception stack */
1136 ENTRY(machine_check)
1137         INTR_FRAME
1138         pushq $0
1139         CFI_ADJUST_CFA_OFFSET 8
1140         paranoidentry do_machine_check
1141         jmp paranoid_exit1
1142         CFI_ENDPROC
1143 END(machine_check)
1144 #endif
1145
1146 /* Call softirq on interrupt stack. Interrupts are off. */
1147 ENTRY(call_softirq)
1148         CFI_STARTPROC
1149         push %rbp
1150         CFI_ADJUST_CFA_OFFSET   8
1151         CFI_REL_OFFSET rbp,0
1152         mov  %rsp,%rbp
1153         CFI_DEF_CFA_REGISTER rbp
1154         incl %gs:pda_irqcount
1155         cmove %gs:pda_irqstackptr,%rsp
1156         push  %rbp                      # backlink for old unwinder
1157         call __do_softirq
1158         leaveq
1159         CFI_DEF_CFA_REGISTER    rsp
1160         CFI_ADJUST_CFA_OFFSET   -8
1161         decl %gs:pda_irqcount
1162         ret
1163         CFI_ENDPROC
1164 ENDPROC(call_softirq)
1165
1166 KPROBE_ENTRY(ignore_sysret)
1167         CFI_STARTPROC
1168         mov $-ENOSYS,%eax
1169         sysret
1170         CFI_ENDPROC
1171 ENDPROC(ignore_sysret)