2 * Compatibility mode system call entry point for x86-64.
4 * Copyright 2000-2002 Andi Kleen, SuSE Labs.
7 #include <asm/asm-offsets.h>
8 #include <asm/current.h>
10 #include <asm/ia32_unistd.h>
11 #include <asm/thread_info.h>
12 #include <asm/segment.h>
13 #include <asm/irqflags.h>
16 #include <linux/linkage.h>
17 #include <linux/err.h>
19 .section .entry.text, "ax"
22 * 32-bit SYSENTER entry.
24 * 32-bit system calls through the vDSO's __kernel_vsyscall enter here
25 * on 64-bit kernels running on Intel CPUs.
27 * The SYSENTER instruction, in principle, should *only* occur in the
28 * vDSO. In practice, a small number of Android devices were shipped
29 * with a copy of Bionic that inlined a SYSENTER instruction. This
30 * never happened in any of Google's Bionic versions -- it only happened
31 * in a narrow range of Intel-provided versions.
33 * SYSENTER loads SS, RSP, CS, and RIP from previously programmed MSRs.
34 * IF and VM in RFLAGS are cleared (IOW: interrupts are off).
35 * SYSENTER does not save anything on the stack,
36 * and does not save old RIP (!!!), RSP, or RFLAGS.
39 * eax system call number
48 ENTRY(entry_SYSENTER_compat)
49 /* Interrupts are off on entry. */
51 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp
54 * User tracing code (ptrace or signal handlers) might assume that
55 * the saved RAX contains a 32-bit number when we're invoking a 32-bit
56 * syscall. Just in case the high bits are nonzero, zero-extend
57 * the syscall number. (This could almost certainly be deleted
58 * with no ill effects.)
62 /* Construct struct pt_regs on stack */
63 pushq $__USER32_DS /* pt_regs->ss */
64 pushq %rbp /* pt_regs->sp (stashed in bp) */
67 * Push flags. This is nasty. First, interrupts are currently
68 * off, but we need pt_regs->flags to have IF set. Second, even
69 * if TF was set when SYSENTER started, it's clear by now. We fix
70 * that later using TIF_SINGLESTEP.
72 pushfq /* pt_regs->flags (except IF = 0) */
73 orl $X86_EFLAGS_IF, (%rsp) /* Fix saved flags */
74 pushq $__USER32_CS /* pt_regs->cs */
75 pushq $0 /* pt_regs->ip = 0 (placeholder) */
76 pushq %rax /* pt_regs->orig_ax */
77 pushq %rdi /* pt_regs->di */
78 pushq %rsi /* pt_regs->si */
79 pushq %rdx /* pt_regs->dx */
80 pushq %rcx /* pt_regs->cx */
81 pushq $-ENOSYS /* pt_regs->ax */
82 pushq $0 /* pt_regs->r8 = 0 */
83 pushq $0 /* pt_regs->r9 = 0 */
84 pushq $0 /* pt_regs->r10 = 0 */
85 pushq $0 /* pt_regs->r11 = 0 */
86 pushq %rbx /* pt_regs->rbx */
87 pushq %rbp /* pt_regs->rbp (will be overwritten) */
88 pushq $0 /* pt_regs->r12 = 0 */
89 pushq $0 /* pt_regs->r13 = 0 */
90 pushq $0 /* pt_regs->r14 = 0 */
91 pushq $0 /* pt_regs->r15 = 0 */
95 * SYSENTER doesn't filter flags, so we need to clear NT and AC
96 * ourselves. To save a few cycles, we can check whether
97 * either was set instead of doing an unconditional popfq.
98 * This needs to happen before enabling interrupts so that
99 * we don't get preempted with NT set.
101 * If TF is set, we will single-step all the way to here -- do_debug
102 * will ignore all the traps. (Yes, this is slow, but so is
103 * single-stepping in general. This allows us to avoid having
104 * a more complicated code to handle the case where a user program
105 * forces us to single-step through the SYSENTER entry code.)
107 * NB.: .Lsysenter_fix_flags is a label with the code under it moved
108 * out-of-line as an optimization: NT is unlikely to be set in the
109 * majority of the cases and instead of polluting the I$ unnecessarily,
110 * we're keeping that code behind a branch which will predict as
111 * not-taken and therefore its instructions won't be fetched.
113 testl $X86_EFLAGS_NT|X86_EFLAGS_AC|X86_EFLAGS_TF, EFLAGS(%rsp)
114 jnz .Lsysenter_fix_flags
115 .Lsysenter_flags_fixed:
118 * User mode is traced as though IRQs are on, and SYSENTER
124 call do_fast_syscall_32
125 /* XEN PV guests always use IRET path */
126 ALTERNATIVE "testl %eax, %eax; jz .Lsyscall_32_done", \
127 "jmp .Lsyscall_32_done", X86_FEATURE_XENPV
128 jmp sysret32_from_system_call
130 .Lsysenter_fix_flags:
131 pushq $X86_EFLAGS_FIXED
133 jmp .Lsysenter_flags_fixed
134 GLOBAL(__end_entry_SYSENTER_compat)
135 ENDPROC(entry_SYSENTER_compat)
138 * 32-bit SYSCALL entry.
140 * 32-bit system calls through the vDSO's __kernel_vsyscall enter here
141 * on 64-bit kernels running on AMD CPUs.
143 * The SYSCALL instruction, in principle, should *only* occur in the
144 * vDSO. In practice, it appears that this really is the case.
147 * - The calling convention for SYSCALL has changed several times without
150 * - Prior to the in-kernel X86_BUG_SYSRET_SS_ATTRS fixup, anything
151 * user task that did SYSCALL without immediately reloading SS
152 * would randomly crash.
154 * - Most programmers do not directly target AMD CPUs, and the 32-bit
155 * SYSCALL instruction does not exist on Intel CPUs. Even on AMD
156 * CPUs, Linux disables the SYSCALL instruction on 32-bit kernels
157 * because the SYSCALL instruction in legacy/native 32-bit mode (as
158 * opposed to compat mode) is sufficiently poorly designed as to be
159 * essentially unusable.
161 * 32-bit SYSCALL saves RIP to RCX, clears RFLAGS.RF, then saves
162 * RFLAGS to R11, then loads new SS, CS, and RIP from previously
163 * programmed MSRs. RFLAGS gets masked by a value from another MSR
164 * (so CLD and CLAC are not needed). SYSCALL does not save anything on
165 * the stack and does not change RSP.
167 * Note: RFLAGS saving+masking-with-MSR happens only in Long mode
168 * (in legacy 32-bit mode, IF, RF and VM bits are cleared and that's it).
169 * Don't get confused: RFLAGS saving+masking depends on Long Mode Active bit
170 * (EFER.LMA=1), NOT on bitness of userspace where SYSCALL executes
171 * or target CS descriptor's L bit (SYSCALL does not read segment descriptors).
174 * eax system call number
177 * ebp arg2 (note: not saved in the stack frame, should not be touched)
184 ENTRY(entry_SYSCALL_compat)
185 /* Interrupts are off on entry. */
188 /* Stash user ESP and switch to the kernel stack. */
190 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp
192 /* Zero-extending 32-bit regs, do not remove */
195 /* Construct struct pt_regs on stack */
196 pushq $__USER32_DS /* pt_regs->ss */
197 pushq %r8 /* pt_regs->sp */
198 pushq %r11 /* pt_regs->flags */
199 pushq $__USER32_CS /* pt_regs->cs */
200 pushq %rcx /* pt_regs->ip */
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 */
219 * User mode is traced as though IRQs are on, and SYSENTER
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
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 */
245 * USERGS_SYSRET32 does:
246 * GSBASE = user's GS base
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.
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.
262 movq RSP-ORIG_RAX(%rsp), %rsp
265 END(entry_SYSCALL_compat)
268 * 32-bit legacy system call entry.
270 * 32-bit x86 Linux system calls traditionally used the INT $0x80
271 * instruction. INT $0x80 lands here.
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
281 * This is considered a slow path. It is not used by most libc
282 * implementations on modern hardware except during process startup.
285 * eax system call number
293 ENTRY(entry_INT80_compat)
295 * Interrupts are off on entry.
297 PARAVIRT_ADJUST_EXCEPTION_FRAME
298 ASM_CLAC /* Do this early to minimize exposure */
302 * User tracing code (ptrace or signal handlers) might assume that
303 * the saved RAX contains a 32-bit number when we're invoking a 32-bit
304 * syscall. Just in case the high bits are nonzero, zero-extend
305 * the syscall number. (This could almost certainly be deleted
306 * with no ill effects.)
310 /* Construct struct pt_regs on stack (iret frame is already on stack) */
311 pushq %rax /* pt_regs->orig_ax */
312 pushq %rdi /* pt_regs->di */
313 pushq %rsi /* pt_regs->si */
314 pushq %rdx /* pt_regs->dx */
315 pushq %rcx /* pt_regs->cx */
316 pushq $-ENOSYS /* pt_regs->ax */
317 pushq $0 /* pt_regs->r8 = 0 */
318 pushq $0 /* pt_regs->r9 = 0 */
319 pushq $0 /* pt_regs->r10 = 0 */
320 pushq $0 /* pt_regs->r11 = 0 */
321 pushq %rbx /* pt_regs->rbx */
322 pushq %rbp /* pt_regs->rbp */
323 pushq %r12 /* pt_regs->r12 */
324 pushq %r13 /* pt_regs->r13 */
325 pushq %r14 /* pt_regs->r14 */
326 pushq %r15 /* pt_regs->r15 */
330 * User mode is traced as though IRQs are on, and the interrupt
331 * gate turned them off.
336 call do_int80_syscall_32
339 /* Go back to user mode. */
342 jmp restore_regs_and_iret
343 END(entry_INT80_compat)
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).
351 * The native 64-bit kernel's sys_clone() implements the latter,
352 * so we need to swap arguments here before calling it: