1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3 * Compatibility mode system call entry point
4 *
5 * Copyright 2000-2002 Andi Kleen, SuSE Labs.
6 */
7 #include <asm/asm-offsets.h>
8 #include <asm/current.h>
9 #include <asm/errno.h>
10 #include <asm/thread_info.h>
11 #include <asm/segment.h>
12 #include <asm/irqflags.h>
13 #include <asm/asm.h>
14 #include <asm/smap.h>
15 #include <asm/nospec-branch.h>
16 #include <linux/linkage.h>
17 #include <linux/err.h>
18
19 #include "calling.h"
20
21 .section .entry.text, "ax"
22
23 /*
24 * 32-bit SYSENTER entry.
25 *
26 * 32-bit system calls through the vDSO's __ke
27 * on 64-bit kernels running on Intel CPUs.
28 *
29 * The SYSENTER instruction, in principle, sho
30 * vDSO. In practice, a small number of Andro
31 * with a copy of Bionic that inlined a SYSENT
32 * never happened in any of Google's Bionic ve
33 * in a narrow range of Intel-provided version
34 *
35 * SYSENTER loads SS, RSP, CS, and RIP from pr
36 * IF and VM in RFLAGS are cleared (IOW: inter
37 * SYSENTER does not save anything on the stac
38 * and does not save old RIP (!!!), RSP, or RF
39 *
40 * Arguments:
41 * eax system call number
42 * ebx arg1
43 * ecx arg2
44 * edx arg3
45 * esi arg4
46 * edi arg5
47 * ebp user stack
48 * 0(%ebp) arg6
49 */
50 SYM_CODE_START(entry_SYSENTER_compat)
51 UNWIND_HINT_ENTRY
52 ENDBR
53 /* Interrupts are off on entry. */
54 swapgs
55
56 pushq %rax
57 SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
58 popq %rax
59
60 movq PER_CPU_VAR(pcpu_hot + X86_top
61
62 /* Construct struct pt_regs on stack *
63 pushq $__USER_DS /* pt_
64 pushq $0 /* pt_
65
66 /*
67 * Push flags. This is nasty. First,
68 * off, but we need pt_regs->flags to
69 * was set in usermode, it's still set
70 * through this code. do_SYSENTER_32(
71 */
72 pushfq /* pt_
73 pushq $__USER32_CS /* pt_
74 pushq $0 /* pt_
75 SYM_INNER_LABEL(entry_SYSENTER_compat_after_hw
76
77 /*
78 * User tracing code (ptrace or signal
79 * the saved RAX contains a 32-bit num
80 * syscall. Just in case the high bit
81 * the syscall number. (This could al
82 * with no ill effects.)
83 */
84 movl %eax, %eax
85
86 pushq %rax /* pt_
87 PUSH_AND_CLEAR_REGS rax=$-ENOSYS
88 UNWIND_HINT_REGS
89
90 cld
91
92 /*
93 * SYSENTER doesn't filter flags, so w
94 * ourselves. To save a few cycles, w
95 * either was set instead of doing an
96 * This needs to happen before enablin
97 * we don't get preempted with NT set.
98 *
99 * If TF is set, we will single-step a
100 * will ignore all the traps. (Yes, t
101 * single-stepping in general. This a
102 * a more complicated code to handle t
103 * forces us to single-step through th
104 *
105 * NB.: .Lsysenter_fix_flags is a labe
106 * out-of-line as an optimization: NT
107 * majority of the cases and instead o
108 * we're keeping that code behind a br
109 * not-taken and therefore its instruc
110 */
111 testl $X86_EFLAGS_NT|X86_EFLAGS_AC|X
112 jnz .Lsysenter_fix_flags
113 .Lsysenter_flags_fixed:
114
115 /*
116 * CPU bugs mitigations mechanisms can
117 * should be invoked after making sure
118 * single-step is ignored only for ins
119 * entry_SYSENTER_compat function.
120 */
121 IBRS_ENTER
122 UNTRAIN_RET
123 CLEAR_BRANCH_HISTORY
124
125 movq %rsp, %rdi
126 call do_SYSENTER_32
127 jmp sysret32_from_system_call
128
129 .Lsysenter_fix_flags:
130 pushq $X86_EFLAGS_FIXED
131 popfq
132 jmp .Lsysenter_flags_fixed
133 SYM_INNER_LABEL(__end_entry_SYSENTER_compat, S
134 SYM_CODE_END(entry_SYSENTER_compat)
135
136 /*
137 * 32-bit SYSCALL entry.
138 *
139 * 32-bit system calls through the vDSO's __ke
140 * on 64-bit kernels running on AMD CPUs.
141 *
142 * The SYSCALL instruction, in principle, shou
143 * vDSO. In practice, it appears that this re
144 * As evidence:
145 *
146 * - The calling convention for SYSCALL has c
147 * anyone noticing.
148 *
149 * - Prior to the in-kernel X86_BUG_SYSRET_SS
150 * user task that did SYSCALL without immed
151 * would randomly crash.
152 *
153 * - Most programmers do not directly target
154 * SYSCALL instruction does not exist on In
155 * CPUs, Linux disables the SYSCALL instruc
156 * because the SYSCALL instruction in legac
157 * opposed to compat mode) is sufficiently
158 * essentially unusable.
159 *
160 * 32-bit SYSCALL saves RIP to RCX, clears RFL
161 * RFLAGS to R11, then loads new SS, CS, and R
162 * programmed MSRs. RFLAGS gets masked by a v
163 * (so CLD and CLAC are not needed). SYSCALL
164 * the stack and does not change RSP.
165 *
166 * Note: RFLAGS saving+masking-with-MSR happen
167 * (in legacy 32-bit mode, IF, RF and VM bits
168 * Don't get confused: RFLAGS saving+masking d
169 * (EFER.LMA=1), NOT on bitness of userspace w
170 * or target CS descriptor's L bit (SYSCALL do
171 *
172 * Arguments:
173 * eax system call number
174 * ecx return address
175 * ebx arg1
176 * ebp arg2 (note: not saved in the stack
177 * edx arg3
178 * esi arg4
179 * edi arg5
180 * esp user stack
181 * 0(%esp) arg6
182 */
183 SYM_CODE_START(entry_SYSCALL_compat)
184 UNWIND_HINT_ENTRY
185 ENDBR
186 /* Interrupts are off on entry. */
187 swapgs
188
189 /* Stash user ESP */
190 movl %esp, %r8d
191
192 /* Use %rsp as scratch reg. User ESP i
193 SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp
194
195 /* Switch to the kernel stack */
196 movq PER_CPU_VAR(pcpu_hot + X86_top
197
198 SYM_INNER_LABEL(entry_SYSCALL_compat_safe_stac
199 ANNOTATE_NOENDBR
200
201 /* Construct struct pt_regs on stack *
202 pushq $__USER_DS /* pt_
203 pushq %r8 /* pt_
204 pushq %r11 /* pt_
205 pushq $__USER32_CS /* pt_
206 pushq %rcx /* pt_
207 SYM_INNER_LABEL(entry_SYSCALL_compat_after_hwf
208 movl %eax, %eax /* dis
209 pushq %rax /* pt_
210 PUSH_AND_CLEAR_REGS rcx=%rbp rax=$-ENO
211 UNWIND_HINT_REGS
212
213 IBRS_ENTER
214 UNTRAIN_RET
215 CLEAR_BRANCH_HISTORY
216
217 movq %rsp, %rdi
218 call do_fast_syscall_32
219
220 sysret32_from_system_call:
221 /* XEN PV guests always use IRET path
222 ALTERNATIVE "testb %al, %al; jz swapgs
223 "jmp swapgs_restore_regs_a
224
225 /*
226 * Opportunistic SYSRET
227 *
228 * We are not going to return to users
229 * stack. So let's erase the thread st
230 */
231 STACKLEAK_ERASE
232
233 IBRS_EXIT
234
235 movq RBX(%rsp), %rbx /* pt_
236 movq RBP(%rsp), %rbp /* pt_
237 movq EFLAGS(%rsp), %r11 /* pt_
238 movq RIP(%rsp), %rcx /* pt_
239 addq $RAX, %rsp /* Ski
240 popq %rax /* pt_
241 popq %rdx /* Ski
242 popq %rdx /* pt_
243 popq %rsi /* pt_
244 popq %rdi /* pt_
245
246 /*
247 * USERGS_SYSRET32 does:
248 * GSBASE = user's GS base
249 * EIP = ECX
250 * RFLAGS = R11
251 * CS = __USER32_CS
252 * SS = __USER_DS
253 *
254 * ECX will not match pt_regs->cx, but
255 * trampoline that will fix up RCX, so
256 *
257 * R12-R15 are callee-saved, so they c
258 * when the system call started, which
259 * code. We zero R8-R10 to avoid info
260 */
261 movq RSP-ORIG_RAX(%rsp), %rsp
262 SYM_INNER_LABEL(entry_SYSRETL_compat_unsafe_st
263 ANNOTATE_NOENDBR
264
265 /*
266 * The original userspace %rsp (RSP-OR
267 * on the process stack which is not m
268 * not readable after we SWITCH_TO_USE
269 * switch until after after the last r
270 * stack.
271 *
272 * %r8/%r9 are zeroed before the sysre
273 */
274 SWITCH_TO_USER_CR3_NOSTACK scratch_reg
275
276 xorl %r8d, %r8d
277 xorl %r9d, %r9d
278 xorl %r10d, %r10d
279 swapgs
280 CLEAR_CPU_BUFFERS
281 sysretl
282 SYM_INNER_LABEL(entry_SYSRETL_compat_end, SYM_
283 ANNOTATE_NOENDBR
284 int3
285 SYM_CODE_END(entry_SYSCALL_compat)
286
287 /*
288 * int 0x80 is used by 32 bit mode as a system
289 * point to C routines, however since this is
290 * history needs to be scrubbed to protect aga
291 * scrubbing needs to take place in assembly c
292 * routines.
293 */
294 SYM_CODE_START(int80_emulation)
295 ANNOTATE_NOENDBR
296 UNWIND_HINT_FUNC
297 CLEAR_BRANCH_HISTORY
298 jmp do_int80_emulation
299 SYM_CODE_END(int80_emulation)
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.