• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <asm/asm-offsets.h>
8#include <asm/current.h>
9#include <asm/errno.h>
10#include <asm/ia32_unistd.h>
11#include <asm/thread_info.h>
12#include <asm/segment.h>
13#include <asm/irqflags.h>
14#include <asm/asm.h>
15#include <asm/smap.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 __kernel_vsyscall enter here
27 * on 64-bit kernels running on Intel CPUs.
28 *
29 * The SYSENTER instruction, in principle, should *only* occur in the
30 * vDSO.  In practice, a small number of Android devices were shipped
31 * with a copy of Bionic that inlined a SYSENTER instruction.  This
32 * never happened in any of Google's Bionic versions -- it only happened
33 * in a narrow range of Intel-provided versions.
34 *
35 * SYSENTER loads SS, RSP, CS, and RIP from previously programmed MSRs.
36 * IF and VM in RFLAGS are cleared (IOW: interrupts are off).
37 * SYSENTER does not save anything on the stack,
38 * and does not save old RIP (!!!), RSP, or RFLAGS.
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 */
50ENTRY(entry_SYSENTER_compat)
51	/* Interrupts are off on entry. */
52	SWAPGS
53
54	/* We are about to clobber %rsp anyway, clobbering here is OK */
55	SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp
56
57	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
58
59	/*
60	 * User tracing code (ptrace or signal handlers) might assume that
61	 * the saved RAX contains a 32-bit number when we're invoking a 32-bit
62	 * syscall.  Just in case the high bits are nonzero, zero-extend
63	 * the syscall number.  (This could almost certainly be deleted
64	 * with no ill effects.)
65	 */
66	movl	%eax, %eax
67
68	/* Construct struct pt_regs on stack */
69	pushq	$__USER32_DS		/* pt_regs->ss */
70	pushq	%rbp			/* pt_regs->sp (stashed in bp) */
71
72	/*
73	 * Push flags.  This is nasty.  First, interrupts are currently
74	 * off, but we need pt_regs->flags to have IF set.  Second, even
75	 * if TF was set when SYSENTER started, it's clear by now.  We fix
76	 * that later using TIF_SINGLESTEP.
77	 */
78	pushfq				/* pt_regs->flags (except IF = 0) */
79	orl	$X86_EFLAGS_IF, (%rsp)	/* Fix saved flags */
80	pushq	$__USER32_CS		/* pt_regs->cs */
81	pushq	$0			/* pt_regs->ip = 0 (placeholder) */
82	pushq	%rax			/* pt_regs->orig_ax */
83	pushq	%rdi			/* pt_regs->di */
84	pushq	%rsi			/* pt_regs->si */
85	pushq	%rdx			/* pt_regs->dx */
86	pushq	%rcx			/* pt_regs->cx */
87	pushq	$-ENOSYS		/* pt_regs->ax */
88	pushq   $0			/* pt_regs->r8  = 0 */
89	xorl	%r8d, %r8d		/* nospec   r8 */
90	pushq   $0			/* pt_regs->r9  = 0 */
91	xorl	%r9d, %r9d		/* nospec   r9 */
92	pushq   $0			/* pt_regs->r10 = 0 */
93	xorl	%r10d, %r10d		/* nospec   r10 */
94	pushq   $0			/* pt_regs->r11 = 0 */
95	xorl	%r11d, %r11d		/* nospec   r11 */
96	pushq   %rbx                    /* pt_regs->rbx */
97	xorl	%ebx, %ebx		/* nospec   rbx */
98	pushq   %rbp                    /* pt_regs->rbp (will be overwritten) */
99	xorl	%ebp, %ebp		/* nospec   rbp */
100	pushq   $0			/* pt_regs->r12 = 0 */
101	xorl	%r12d, %r12d		/* nospec   r12 */
102	pushq   $0			/* pt_regs->r13 = 0 */
103	xorl	%r13d, %r13d		/* nospec   r13 */
104	pushq   $0			/* pt_regs->r14 = 0 */
105	xorl	%r14d, %r14d		/* nospec   r14 */
106	pushq   $0			/* pt_regs->r15 = 0 */
107	xorl	%r15d, %r15d		/* nospec   r15 */
108	cld
109
110	IBRS_ENTER
111
112	/*
113	 * SYSENTER doesn't filter flags, so we need to clear NT and AC
114	 * ourselves.  To save a few cycles, we can check whether
115	 * either was set instead of doing an unconditional popfq.
116	 * This needs to happen before enabling interrupts so that
117	 * we don't get preempted with NT set.
118	 *
119	 * If TF is set, we will single-step all the way to here -- do_debug
120	 * will ignore all the traps.  (Yes, this is slow, but so is
121	 * single-stepping in general.  This allows us to avoid having
122	 * a more complicated code to handle the case where a user program
123	 * forces us to single-step through the SYSENTER entry code.)
124	 *
125	 * NB.: .Lsysenter_fix_flags is a label with the code under it moved
126	 * out-of-line as an optimization: NT is unlikely to be set in the
127	 * majority of the cases and instead of polluting the I$ unnecessarily,
128	 * we're keeping that code behind a branch which will predict as
129	 * not-taken and therefore its instructions won't be fetched.
130	 */
131	testl	$X86_EFLAGS_NT|X86_EFLAGS_AC|X86_EFLAGS_TF, EFLAGS(%rsp)
132	jnz	.Lsysenter_fix_flags
133.Lsysenter_flags_fixed:
134
135	/*
136	 * User mode is traced as though IRQs are on, and SYSENTER
137	 * turned them off.
138	 */
139	TRACE_IRQS_OFF
140
141	movq	%rsp, %rdi
142	call	do_fast_syscall_32
143	/* XEN PV guests always use IRET path */
144	ALTERNATIVE "testl %eax, %eax; jz .Lsyscall_32_done", \
145		    "jmp .Lsyscall_32_done", X86_FEATURE_XENPV
146	jmp	sysret32_from_system_call
147
148.Lsysenter_fix_flags:
149	pushq	$X86_EFLAGS_FIXED
150	popfq
151	jmp	.Lsysenter_flags_fixed
152GLOBAL(__end_entry_SYSENTER_compat)
153ENDPROC(entry_SYSENTER_compat)
154
155/*
156 * 32-bit SYSCALL entry.
157 *
158 * 32-bit system calls through the vDSO's __kernel_vsyscall enter here
159 * on 64-bit kernels running on AMD CPUs.
160 *
161 * The SYSCALL instruction, in principle, should *only* occur in the
162 * vDSO.  In practice, it appears that this really is the case.
163 * As evidence:
164 *
165 *  - The calling convention for SYSCALL has changed several times without
166 *    anyone noticing.
167 *
168 *  - Prior to the in-kernel X86_BUG_SYSRET_SS_ATTRS fixup, anything
169 *    user task that did SYSCALL without immediately reloading SS
170 *    would randomly crash.
171 *
172 *  - Most programmers do not directly target AMD CPUs, and the 32-bit
173 *    SYSCALL instruction does not exist on Intel CPUs.  Even on AMD
174 *    CPUs, Linux disables the SYSCALL instruction on 32-bit kernels
175 *    because the SYSCALL instruction in legacy/native 32-bit mode (as
176 *    opposed to compat mode) is sufficiently poorly designed as to be
177 *    essentially unusable.
178 *
179 * 32-bit SYSCALL saves RIP to RCX, clears RFLAGS.RF, then saves
180 * RFLAGS to R11, then loads new SS, CS, and RIP from previously
181 * programmed MSRs.  RFLAGS gets masked by a value from another MSR
182 * (so CLD and CLAC are not needed).  SYSCALL does not save anything on
183 * the stack and does not change RSP.
184 *
185 * Note: RFLAGS saving+masking-with-MSR happens only in Long mode
186 * (in legacy 32-bit mode, IF, RF and VM bits are cleared and that's it).
187 * Don't get confused: RFLAGS saving+masking depends on Long Mode Active bit
188 * (EFER.LMA=1), NOT on bitness of userspace where SYSCALL executes
189 * or target CS descriptor's L bit (SYSCALL does not read segment descriptors).
190 *
191 * Arguments:
192 * eax  system call number
193 * ecx  return address
194 * ebx  arg1
195 * ebp  arg2	(note: not saved in the stack frame, should not be touched)
196 * edx  arg3
197 * esi  arg4
198 * edi  arg5
199 * esp  user stack
200 * 0(%esp) arg6
201 */
202ENTRY(entry_SYSCALL_compat)
203	/* Interrupts are off on entry. */
204	swapgs
205
206	/* Stash user ESP */
207	movl	%esp, %r8d
208
209	/* Use %rsp as scratch reg. User ESP is stashed in r8 */
210	SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp
211
212	/* Switch to the kernel stack */
213	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
214
215	/* Construct struct pt_regs on stack */
216	pushq	$__USER32_DS		/* pt_regs->ss */
217	pushq	%r8			/* pt_regs->sp */
218	pushq	%r11			/* pt_regs->flags */
219	pushq	$__USER32_CS		/* pt_regs->cs */
220	pushq	%rcx			/* pt_regs->ip */
221GLOBAL(entry_SYSCALL_compat_after_hwframe)
222	movl	%eax, %eax		/* discard orig_ax high bits */
223	pushq	%rax			/* pt_regs->orig_ax */
224	pushq	%rdi			/* pt_regs->di */
225	pushq	%rsi			/* pt_regs->si */
226	xorl	%esi, %esi		/* nospec   si */
227	pushq	%rdx			/* pt_regs->dx */
228	xorl	%edx, %edx		/* nospec   dx */
229	pushq	%rbp			/* pt_regs->cx (stashed in bp) */
230	xorl	%ecx, %ecx		/* nospec   cx */
231	pushq	$-ENOSYS		/* pt_regs->ax */
232	pushq   $0			/* pt_regs->r8  = 0 */
233	xorl	%r8d, %r8d		/* nospec   r8 */
234	pushq   $0			/* pt_regs->r9  = 0 */
235	xorl	%r9d, %r9d		/* nospec   r9 */
236	pushq   $0			/* pt_regs->r10 = 0 */
237	xorl	%r10d, %r10d		/* nospec   r10 */
238	pushq   $0			/* pt_regs->r11 = 0 */
239	xorl	%r11d, %r11d		/* nospec   r11 */
240	pushq   %rbx                    /* pt_regs->rbx */
241	xorl	%ebx, %ebx		/* nospec   rbx */
242	pushq   %rbp                    /* pt_regs->rbp (will be overwritten) */
243	xorl	%ebp, %ebp		/* nospec   rbp */
244	pushq   $0			/* pt_regs->r12 = 0 */
245	xorl	%r12d, %r12d		/* nospec   r12 */
246	pushq   $0			/* pt_regs->r13 = 0 */
247	xorl	%r13d, %r13d		/* nospec   r13 */
248	pushq   $0			/* pt_regs->r14 = 0 */
249	xorl	%r14d, %r14d		/* nospec   r14 */
250	pushq   $0			/* pt_regs->r15 = 0 */
251	xorl	%r15d, %r15d		/* nospec   r15 */
252
253	/*
254	 * User mode is traced as though IRQs are on, and SYSENTER
255	 * turned them off.
256	 */
257	TRACE_IRQS_OFF
258
259	IBRS_ENTER
260
261	movq	%rsp, %rdi
262	call	do_fast_syscall_32
263	/* XEN PV guests always use IRET path */
264	ALTERNATIVE "testl %eax, %eax; jz .Lsyscall_32_done", \
265		    "jmp .Lsyscall_32_done", X86_FEATURE_XENPV
266
267	/* Opportunistic SYSRET */
268sysret32_from_system_call:
269	/*
270	 * We are not going to return to userspace from the trampoline
271	 * stack. So let's erase the thread stack right now.
272	 */
273	STACKLEAK_ERASE
274	TRACE_IRQS_ON			/* User mode traces as IRQs on. */
275
276	IBRS_EXIT
277
278	movq	RBX(%rsp), %rbx		/* pt_regs->rbx */
279	movq	RBP(%rsp), %rbp		/* pt_regs->rbp */
280	movq	EFLAGS(%rsp), %r11	/* pt_regs->flags (in r11) */
281	movq	RIP(%rsp), %rcx		/* pt_regs->ip (in rcx) */
282	addq	$RAX, %rsp		/* Skip r8-r15 */
283	popq	%rax			/* pt_regs->rax */
284	popq	%rdx			/* Skip pt_regs->cx */
285	popq	%rdx			/* pt_regs->dx */
286	popq	%rsi			/* pt_regs->si */
287	popq	%rdi			/* pt_regs->di */
288
289        /*
290         * USERGS_SYSRET32 does:
291         *  GSBASE = user's GS base
292         *  EIP = ECX
293         *  RFLAGS = R11
294         *  CS = __USER32_CS
295         *  SS = __USER_DS
296         *
297	 * ECX will not match pt_regs->cx, but we're returning to a vDSO
298	 * trampoline that will fix up RCX, so this is okay.
299	 *
300	 * R12-R15 are callee-saved, so they contain whatever was in them
301	 * when the system call started, which is already known to user
302	 * code.  We zero R8-R10 to avoid info leaks.
303         */
304	movq	RSP-ORIG_RAX(%rsp), %rsp
305
306	/*
307	 * The original userspace %rsp (RSP-ORIG_RAX(%rsp)) is stored
308	 * on the process stack which is not mapped to userspace and
309	 * not readable after we SWITCH_TO_USER_CR3.  Delay the CR3
310	 * switch until after after the last reference to the process
311	 * stack.
312	 *
313	 * %r8/%r9 are zeroed before the sysret, thus safe to clobber.
314	 */
315	SWITCH_TO_USER_CR3_NOSTACK scratch_reg=%r8 scratch_reg2=%r9
316
317	xorl	%r8d, %r8d
318	xorl	%r9d, %r9d
319	xorl	%r10d, %r10d
320	swapgs
321	sysretl
322END(entry_SYSCALL_compat)
323
324/*
325 * 32-bit legacy system call entry.
326 *
327 * 32-bit x86 Linux system calls traditionally used the INT $0x80
328 * instruction.  INT $0x80 lands here.
329 *
330 * This entry point can be used by 32-bit and 64-bit programs to perform
331 * 32-bit system calls.  Instances of INT $0x80 can be found inline in
332 * various programs and libraries.  It is also used by the vDSO's
333 * __kernel_vsyscall fallback for hardware that doesn't support a faster
334 * entry method.  Restarted 32-bit system calls also fall back to INT
335 * $0x80 regardless of what instruction was originally used to do the
336 * system call.
337 *
338 * This is considered a slow path.  It is not used by most libc
339 * implementations on modern hardware except during process startup.
340 *
341 * Arguments:
342 * eax  system call number
343 * ebx  arg1
344 * ecx  arg2
345 * edx  arg3
346 * esi  arg4
347 * edi  arg5
348 * ebp  arg6
349 */
350ENTRY(entry_INT80_compat)
351	/*
352	 * Interrupts are off on entry.
353	 */
354	ASM_CLAC			/* Do this early to minimize exposure */
355	SWAPGS
356
357	/*
358	 * User tracing code (ptrace or signal handlers) might assume that
359	 * the saved RAX contains a 32-bit number when we're invoking a 32-bit
360	 * syscall.  Just in case the high bits are nonzero, zero-extend
361	 * the syscall number.  (This could almost certainly be deleted
362	 * with no ill effects.)
363	 */
364	movl	%eax, %eax
365
366	/* switch to thread stack expects orig_ax and rdi to be pushed */
367	pushq	%rax			/* pt_regs->orig_ax */
368	pushq	%rdi			/* pt_regs->di */
369
370	/* Need to switch before accessing the thread stack. */
371	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi
372	/* In the Xen PV case we already run on the thread stack. */
373	ALTERNATIVE "movq %rsp, %rdi", "jmp .Lint80_keep_stack", X86_FEATURE_XENPV
374	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
375
376	pushq	6*8(%rdi)		/* regs->ss */
377	pushq	5*8(%rdi)		/* regs->rsp */
378	pushq	4*8(%rdi)		/* regs->eflags */
379	pushq	3*8(%rdi)		/* regs->cs */
380	pushq	2*8(%rdi)		/* regs->ip */
381	pushq	1*8(%rdi)		/* regs->orig_ax */
382	pushq	(%rdi)			/* pt_regs->di */
383.Lint80_keep_stack:
384
385	pushq	%rsi			/* pt_regs->si */
386	xorl	%esi, %esi		/* nospec   si */
387	pushq	%rdx			/* pt_regs->dx */
388	xorl	%edx, %edx		/* nospec   dx */
389	pushq	%rcx			/* pt_regs->cx */
390	xorl	%ecx, %ecx		/* nospec   cx */
391	pushq	$-ENOSYS		/* pt_regs->ax */
392	pushq   %r8			/* pt_regs->r8 */
393	xorl	%r8d, %r8d		/* nospec   r8 */
394	pushq   %r9			/* pt_regs->r9 */
395	xorl	%r9d, %r9d		/* nospec   r9 */
396	pushq   %r10			/* pt_regs->r10*/
397	xorl	%r10d, %r10d		/* nospec   r10 */
398	pushq   %r11			/* pt_regs->r11 */
399	xorl	%r11d, %r11d		/* nospec   r11 */
400	pushq   %rbx                    /* pt_regs->rbx */
401	xorl	%ebx, %ebx		/* nospec   rbx */
402	pushq   %rbp                    /* pt_regs->rbp */
403	xorl	%ebp, %ebp		/* nospec   rbp */
404	pushq   %r12                    /* pt_regs->r12 */
405	xorl	%r12d, %r12d		/* nospec   r12 */
406	pushq   %r13                    /* pt_regs->r13 */
407	xorl	%r13d, %r13d		/* nospec   r13 */
408	pushq   %r14                    /* pt_regs->r14 */
409	xorl	%r14d, %r14d		/* nospec   r14 */
410	pushq   %r15                    /* pt_regs->r15 */
411	xorl	%r15d, %r15d		/* nospec   r15 */
412	cld
413
414	/*
415	 * User mode is traced as though IRQs are on, and the interrupt
416	 * gate turned them off.
417	 */
418	TRACE_IRQS_OFF
419	IBRS_ENTER
420
421	movq	%rsp, %rdi
422	call	do_int80_syscall_32
423.Lsyscall_32_done:
424
425	/* Go back to user mode. */
426	TRACE_IRQS_ON
427	jmp	swapgs_restore_regs_and_return_to_usermode
428END(entry_INT80_compat)
429