arch/x86/entry/entry_64_compat.S

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 /*
   3  * Compatibility mode system call entry point for x86-64.
   4  *
   5  * Copyright 2000-2002 Andi Kleen, SuSE Labs.
   6  */
   7 #include "calling.h"
   8 #include <asm/asm-offsets.h>
   9 #include <asm/current.h>
  10 #include <asm/errno.h>
  11 #include <asm/ia32_unistd.h>
  12 #include <asm/thread_info.h>
  13 #include <asm/segment.h>
  14 #include <asm/irqflags.h>
  15 #include <asm/asm.h>
  16 #include <asm/smap.h>
  17 #include <linux/linkage.h>
  18 #include <linux/err.h>
  19
  20         .section .entry.text, "ax"
  21
  22 /*
  23  * 32-bit SYSENTER entry.
  24  *
  25  * 32-bit system calls through the vDSO's __kernel_vsyscall enter here
  26  * on 64-bit kernels running on Intel CPUs.
  27  *
  28  * The SYSENTER instruction, in principle, should *only* occur in the
  29  * vDSO.  In practice, a small number of Android devices were shipped
  30  * with a copy of Bionic that inlined a SYSENTER instruction.  This
  31  * never happened in any of Google's Bionic versions -- it only happened
  32  * in a narrow range of Intel-provided versions.
  33  *
  34  * SYSENTER loads SS, RSP, CS, and RIP from previously programmed MSRs.
  35  * IF and VM in RFLAGS are cleared (IOW: interrupts are off).
  36  * SYSENTER does not save anything on the stack,
  37  * and does not save old RIP (!!!), RSP, or RFLAGS.
  38  *
  39  * Arguments:
  40  * eax  system call number
  41  * ebx  arg1
  42  * ecx  arg2
  43  * edx  arg3
  44  * esi  arg4
  45  * edi  arg5
  46  * ebp  user stack
  47  * 0(%ebp) arg6
  48  */
  49 ENTRY(entry_SYSENTER_compat)
  50         /* Interrupts are off on entry. */
  51         SWAPGS
  52         movq    PER_CPU_VAR(cpu_current_top_of_stack), %rsp
  53
  54         /*
  55          * User tracing code (ptrace or signal handlers) might assume that
  56          * the saved RAX contains a 32-bit number when we're invoking a 32-bit
  57          * syscall.  Just in case the high bits are nonzero, zero-extend
  58          * the syscall number.  (This could almost certainly be deleted
  59          * with no ill effects.)
  60          */
  61         movl    %eax, %eax
  62
  63         /* Construct struct pt_regs on stack */
  64         pushq   $__USER32_DS            /* pt_regs->ss */
  65         pushq   %rbp                    /* pt_regs->sp (stashed in bp) */
  66
  67         /*
  68          * Push flags.  This is nasty.  First, interrupts are currently
  69          * off, but we need pt_regs->flags to have IF set.  Second, even
  70          * if TF was set when SYSENTER started, it's clear by now.  We fix
  71          * that later using TIF_SINGLESTEP.
  72          */
  73         pushfq                          /* pt_regs->flags (except IF = 0) */
  74         orl     $X86_EFLAGS_IF, (%rsp)  /* Fix saved flags */
  75         pushq   $__USER32_CS            /* pt_regs->cs */
  76         pushq   $0                      /* pt_regs->ip = 0 (placeholder) */
  77         pushq   %rax                    /* pt_regs->orig_ax */
  78         pushq   %rdi                    /* pt_regs->di */
  79         pushq   %rsi                    /* pt_regs->si */
  80         pushq   %rdx                    /* pt_regs->dx */
  81         pushq   %rcx                    /* pt_regs->cx */
  82         pushq   $-ENOSYS                /* pt_regs->ax */
  83         pushq   $0                      /* pt_regs->r8  = 0 */
  84         pushq   $0                      /* pt_regs->r9  = 0 */
  85         pushq   $0                      /* pt_regs->r10 = 0 */
  86         pushq   $0                      /* pt_regs->r11 = 0 */
  87         pushq   %rbx                    /* pt_regs->rbx */
  88         pushq   %rbp                    /* pt_regs->rbp (will be overwritten) */
  89         pushq   $0                      /* pt_regs->r12 = 0 */
  90         pushq   $0                      /* pt_regs->r13 = 0 */
  91         pushq   $0                      /* pt_regs->r14 = 0 */
  92         pushq   $0                      /* pt_regs->r15 = 0 */
  93         cld
  94
  95         /*
  96          * SYSENTER doesn't filter flags, so we need to clear NT and AC
  97          * ourselves.  To save a few cycles, we can check whether
  98          * either was set instead of doing an unconditional popfq.
  99          * This needs to happen before enabling interrupts so that
 100          * we don't get preempted with NT set.
 101          *
 102          * If TF is set, we will single-step all the way to here -- do_debug
 103          * will ignore all the traps.  (Yes, this is slow, but so is
 104          * single-stepping in general.  This allows us to avoid having
 105          * a more complicated code to handle the case where a user program
 106          * forces us to single-step through the SYSENTER entry code.)
 107          *
 108          * NB.: .Lsysenter_fix_flags is a label with the code under it moved
 109          * out-of-line as an optimization: NT is unlikely to be set in the
 110          * majority of the cases and instead of polluting the I$ unnecessarily,
 111          * we're keeping that code behind a branch which will predict as
 112          * not-taken and therefore its instructions won't be fetched.
 113          */
 114         testl   $X86_EFLAGS_NT|X86_EFLAGS_AC|X86_EFLAGS_TF, EFLAGS(%rsp)
 115         jnz     .Lsysenter_fix_flags
 116 .Lsysenter_flags_fixed:
 117
 118         /*
 119          * User mode is traced as though IRQs are on, and SYSENTER
 120          * turned them off.
 121          */
 122         TRACE_IRQS_OFF
 123
 124         movq    %rsp, %rdi
 125         call    do_fast_syscall_32
 126         /* XEN PV guests always use IRET path */
 127         ALTERNATIVE "testl %eax, %eax; jz .Lsyscall_32_done", \
 128                     "jmp .Lsyscall_32_done", X86_FEATURE_XENPV
 129         jmp     sysret32_from_system_call
 130
 131 .Lsysenter_fix_flags:
 132         pushq   $X86_EFLAGS_FIXED
 133         popfq
 134         jmp     .Lsysenter_flags_fixed
 135 GLOBAL(__end_entry_SYSENTER_compat)
 136 ENDPROC(entry_SYSENTER_compat)
 137
 138 /*
 139  * 32-bit SYSCALL entry.
 140  *
 141  * 32-bit system calls through the vDSO's __kernel_vsyscall enter here
 142  * on 64-bit kernels running on AMD CPUs.
 143  *
 144  * The SYSCALL instruction, in principle, should *only* occur in the
 145  * vDSO.  In practice, it appears that this really is the case.
 146  * As evidence:
 147  *
 148  *  - The calling convention for SYSCALL has changed several times without
 149  *    anyone noticing.
 150  *
 151  *  - Prior to the in-kernel X86_BUG_SYSRET_SS_ATTRS fixup, anything
 152  *    user task that did SYSCALL without immediately reloading SS
 153  *    would randomly crash.
 154  *
 155  *  - Most programmers do not directly target AMD CPUs, and the 32-bit
 156  *    SYSCALL instruction does not exist on Intel CPUs.  Even on AMD
 157  *    CPUs, Linux disables the SYSCALL instruction on 32-bit kernels
 158  *    because the SYSCALL instruction in legacy/native 32-bit mode (as
 159  *    opposed to compat mode) is sufficiently poorly designed as to be
 160  *    essentially unusable.
 161  *
 162  * 32-bit SYSCALL saves RIP to RCX, clears RFLAGS.RF, then saves
 163  * RFLAGS to R11, then loads new SS, CS, and RIP from previously
 164  * programmed MSRs.  RFLAGS gets masked by a value from another MSR
 165  * (so CLD and CLAC are not needed).  SYSCALL does not save anything on
 166  * the stack and does not change RSP.
 167  *
 168  * Note: RFLAGS saving+masking-with-MSR happens only in Long mode
 169  * (in legacy 32-bit mode, IF, RF and VM bits are cleared and that's it).
 170  * Don't get confused: RFLAGS saving+masking depends on Long Mode Active bit
 171  * (EFER.LMA=1), NOT on bitness of userspace where SYSCALL executes
 172  * or target CS descriptor's L bit (SYSCALL does not read segment descriptors).
 173  *
 174  * Arguments:
 175  * eax  system call number
 176  * ecx  return address
 177  * ebx  arg1
 178  * ebp  arg2    (note: not saved in the stack frame, should not be touched)
 179  * edx  arg3
 180  * esi  arg4
 181  * edi  arg5
 182  * esp  user stack
 183  * 0(%esp) arg6
 184  */
 185 ENTRY(entry_SYSCALL_compat)
 186         /* Interrupts are off on entry. */
 187         swapgs
 188
 189         /* Stash user ESP and switch to the kernel stack. */
 190         movl    %esp, %r8d
 191         movq    PER_CPU_VAR(cpu_current_top_of_stack), %rsp
 192
 193         /* Construct struct pt_regs on stack */
 194         pushq   $__USER32_DS            /* pt_regs->ss */
 195         pushq   %r8                     /* pt_regs->sp */
 196         pushq   %r11                    /* pt_regs->flags */
 197         pushq   $__USER32_CS            /* pt_regs->cs */
 198         pushq   %rcx                    /* pt_regs->ip */
 199 GLOBAL(entry_SYSCALL_compat_after_hwframe)
 200         movl    %eax, %eax              /* discard orig_ax high bits */
 201         pushq   %rax                    /* pt_regs->orig_ax */
 202         pushq   %rdi                    /* pt_regs->di */
 203         pushq   %rsi                    /* pt_regs->si */
 204         pushq   %rdx                    /* pt_regs->dx */
 205         pushq   %rbp                    /* pt_regs->cx (stashed in bp) */
 206         pushq   $-ENOSYS                /* pt_regs->ax */
 207         pushq   $0                      /* pt_regs->r8  = 0 */
 208         pushq   $0                      /* pt_regs->r9  = 0 */
 209         pushq   $0                      /* pt_regs->r10 = 0 */
 210         pushq   $0                      /* pt_regs->r11 = 0 */
 211         pushq   %rbx                    /* pt_regs->rbx */
 212         pushq   %rbp                    /* pt_regs->rbp (will be overwritten) */
 213         pushq   $0                      /* pt_regs->r12 = 0 */
 214         pushq   $0                      /* pt_regs->r13 = 0 */
 215         pushq   $0                      /* pt_regs->r14 = 0 */
 216         pushq   $0                      /* pt_regs->r15 = 0 */
 217
 218         /*
 219          * User mode is traced as though IRQs are on, and SYSENTER
 220          * turned them off.
 221          */
 222         TRACE_IRQS_OFF
 223
 224         movq    %rsp, %rdi
 225         call    do_fast_syscall_32
 226         /* XEN PV guests always use IRET path */
 227         ALTERNATIVE "testl %eax, %eax; jz .Lsyscall_32_done", \
 228                     "jmp .Lsyscall_32_done", X86_FEATURE_XENPV
 229
 230         /* Opportunistic SYSRET */
 231 sysret32_from_system_call:
 232         TRACE_IRQS_ON                   /* User mode traces as IRQs on. */
 233         movq    RBX(%rsp), %rbx         /* pt_regs->rbx */
 234         movq    RBP(%rsp), %rbp         /* pt_regs->rbp */
 235         movq    EFLAGS(%rsp), %r11      /* pt_regs->flags (in r11) */
 236         movq    RIP(%rsp), %rcx         /* pt_regs->ip (in rcx) */
 237         addq    $RAX, %rsp              /* Skip r8-r15 */
 238         popq    %rax                    /* pt_regs->rax */
 239         popq    %rdx                    /* Skip pt_regs->cx */
 240         popq    %rdx                    /* pt_regs->dx */
 241         popq    %rsi                    /* pt_regs->si */
 242         popq    %rdi                    /* pt_regs->di */
 243
 244         /*
 245          * USERGS_SYSRET32 does:
 246          *  GSBASE = user's GS base
 247          *  EIP = ECX
 248          *  RFLAGS = R11
 249          *  CS = __USER32_CS
 250          *  SS = __USER_DS
 251          *
 252          * ECX will not match pt_regs->cx, but we're returning to a vDSO
 253          * trampoline that will fix up RCX, so this is okay.
 254          *
 255          * R12-R15 are callee-saved, so they contain whatever was in them
 256          * when the system call started, which is already known to user
 257          * code.  We zero R8-R10 to avoid info leaks.
 258          */
 259         xorq    %r8, %r8
 260         xorq    %r9, %r9
 261         xorq    %r10, %r10
 262         movq    RSP-ORIG_RAX(%rsp), %rsp
 263         swapgs
 264         sysretl
 265 END(entry_SYSCALL_compat)
 266
 267 /*
 268  * 32-bit legacy system call entry.
 269  *
 270  * 32-bit x86 Linux system calls traditionally used the INT $0x80
 271  * instruction.  INT $0x80 lands here.
 272  *
 273  * This entry point can be used by 32-bit and 64-bit programs to perform
 274  * 32-bit system calls.  Instances of INT $0x80 can be found inline in
 275  * various programs and libraries.  It is also used by the vDSO's
 276  * __kernel_vsyscall fallback for hardware that doesn't support a faster
 277  * entry method.  Restarted 32-bit system calls also fall back to INT
 278  * $0x80 regardless of what instruction was originally used to do the
 279  * system call.
 280  *
 281  * This is considered a slow path.  It is not used by most libc
 282  * implementations on modern hardware except during process startup.
 283  *
 284  * Arguments:
 285  * eax  system call number
 286  * ebx  arg1
 287  * ecx  arg2
 288  * edx  arg3
 289  * esi  arg4
 290  * edi  arg5
 291  * ebp  arg6
 292  */
 293 ENTRY(entry_INT80_compat)
 294         /*
 295          * Interrupts are off on entry.
 296          */
 297         ASM_CLAC                        /* Do this early to minimize exposure */
 298         SWAPGS
 299
 300         /*
 301          * User tracing code (ptrace or signal handlers) might assume that
 302          * the saved RAX contains a 32-bit number when we're invoking a 32-bit
 303          * syscall.  Just in case the high bits are nonzero, zero-extend
 304          * the syscall number.  (This could almost certainly be deleted
 305          * with no ill effects.)
 306          */
 307         movl    %eax, %eax
 308
 309         pushq   %rax                    /* pt_regs->orig_ax */
 310
 311         /* switch to thread stack expects orig_ax to be pushed */
 312         call    switch_to_thread_stack
 313
 314         pushq   %rdi                    /* pt_regs->di */
 315         pushq   %rsi                    /* pt_regs->si */
 316         pushq   %rdx                    /* pt_regs->dx */
 317         pushq   %rcx                    /* pt_regs->cx */
 318         pushq   $-ENOSYS                /* pt_regs->ax */
 319         pushq   $0                      /* pt_regs->r8  = 0 */
 320         pushq   $0                      /* pt_regs->r9  = 0 */
 321         pushq   $0                      /* pt_regs->r10 = 0 */
 322         pushq   $0                      /* pt_regs->r11 = 0 */
 323         pushq   %rbx                    /* pt_regs->rbx */
 324         pushq   %rbp                    /* pt_regs->rbp */
 325         pushq   %r12                    /* pt_regs->r12 */
 326         pushq   %r13                    /* pt_regs->r13 */
 327         pushq   %r14                    /* pt_regs->r14 */
 328         pushq   %r15                    /* pt_regs->r15 */
 329         cld
 330
 331         /*
 332          * User mode is traced as though IRQs are on, and the interrupt
 333          * gate turned them off.
 334          */
 335         TRACE_IRQS_OFF
 336
 337         movq    %rsp, %rdi
 338         call    do_int80_syscall_32
 339 .Lsyscall_32_done:
 340
 341         /* Go back to user mode. */
 342         TRACE_IRQS_ON
 343         jmp     swapgs_restore_regs_and_return_to_usermode
 344 END(entry_INT80_compat)
 345
 346 ENTRY(stub32_clone)
 347         /*
 348          * The 32-bit clone ABI is: clone(..., int tls_val, int *child_tidptr).
 349          * The 64-bit clone ABI is: clone(..., int *child_tidptr, int tls_val).
 350          *
 351          * The native 64-bit kernel's sys_clone() implements the latter,
 352          * so we need to swap arguments here before calling it:
 353          */
 354         xchg    %r8, %rcx
 355         jmp     sys_clone
 356 ENDPROC(stub32_clone)