Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * common.c - C code for kernel entry and exit
4 * Copyright (c) 2015 Andrew Lutomirski
5 *
6 * Based on asm and ptrace code by many authors. The code here originated
7 * in ptrace.c and signal.c.
8 */
9
10#include <linux/kernel.h>
11#include <linux/sched.h>
12#include <linux/sched/task_stack.h>
13#include <linux/entry-common.h>
14#include <linux/mm.h>
15#include <linux/smp.h>
16#include <linux/errno.h>
17#include <linux/ptrace.h>
18#include <linux/export.h>
19#include <linux/nospec.h>
20#include <linux/syscalls.h>
21#include <linux/uaccess.h>
22
23#ifdef CONFIG_XEN_PV
24#include <xen/xen-ops.h>
25#include <xen/events.h>
26#endif
27
28#include <asm/desc.h>
29#include <asm/traps.h>
30#include <asm/vdso.h>
31#include <asm/cpufeature.h>
32#include <asm/fpu/api.h>
33#include <asm/nospec-branch.h>
34#include <asm/io_bitmap.h>
35#include <asm/syscall.h>
36#include <asm/irq_stack.h>
37
38#ifdef CONFIG_X86_64
39
40static __always_inline bool do_syscall_x64(struct pt_regs *regs, int nr)
41{
42 /*
43 * Convert negative numbers to very high and thus out of range
44 * numbers for comparisons.
45 */
46 unsigned int unr = nr;
47
48 if (likely(unr < NR_syscalls)) {
49 unr = array_index_nospec(unr, NR_syscalls);
50 regs->ax = sys_call_table[unr](regs);
51 return true;
52 }
53 return false;
54}
55
56static __always_inline bool do_syscall_x32(struct pt_regs *regs, int nr)
57{
58 /*
59 * Adjust the starting offset of the table, and convert numbers
60 * < __X32_SYSCALL_BIT to very high and thus out of range
61 * numbers for comparisons.
62 */
63 unsigned int xnr = nr - __X32_SYSCALL_BIT;
64
65 if (IS_ENABLED(CONFIG_X86_X32_ABI) && likely(xnr < X32_NR_syscalls)) {
66 xnr = array_index_nospec(xnr, X32_NR_syscalls);
67 regs->ax = x32_sys_call_table[xnr](regs);
68 return true;
69 }
70 return false;
71}
72
73__visible noinstr void do_syscall_64(struct pt_regs *regs, int nr)
74{
75 add_random_kstack_offset();
76 nr = syscall_enter_from_user_mode(regs, nr);
77
78 instrumentation_begin();
79
80 if (!do_syscall_x64(regs, nr) && !do_syscall_x32(regs, nr) && nr != -1) {
81 /* Invalid system call, but still a system call. */
82 regs->ax = __x64_sys_ni_syscall(regs);
83 }
84
85 instrumentation_end();
86 syscall_exit_to_user_mode(regs);
87}
88#endif
89
90#if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION)
91static __always_inline int syscall_32_enter(struct pt_regs *regs)
92{
93 if (IS_ENABLED(CONFIG_IA32_EMULATION))
94 current_thread_info()->status |= TS_COMPAT;
95
96 return (int)regs->orig_ax;
97}
98
99/*
100 * Invoke a 32-bit syscall. Called with IRQs on in CONTEXT_KERNEL.
101 */
102static __always_inline void do_syscall_32_irqs_on(struct pt_regs *regs, int nr)
103{
104 /*
105 * Convert negative numbers to very high and thus out of range
106 * numbers for comparisons.
107 */
108 unsigned int unr = nr;
109
110 if (likely(unr < IA32_NR_syscalls)) {
111 unr = array_index_nospec(unr, IA32_NR_syscalls);
112 regs->ax = ia32_sys_call_table[unr](regs);
113 } else if (nr != -1) {
114 regs->ax = __ia32_sys_ni_syscall(regs);
115 }
116}
117
118/* Handles int $0x80 */
119__visible noinstr void do_int80_syscall_32(struct pt_regs *regs)
120{
121 int nr = syscall_32_enter(regs);
122
123 add_random_kstack_offset();
124 /*
125 * Subtlety here: if ptrace pokes something larger than 2^31-1 into
126 * orig_ax, the int return value truncates it. This matches
127 * the semantics of syscall_get_nr().
128 */
129 nr = syscall_enter_from_user_mode(regs, nr);
130 instrumentation_begin();
131
132 do_syscall_32_irqs_on(regs, nr);
133
134 instrumentation_end();
135 syscall_exit_to_user_mode(regs);
136}
137
138static noinstr bool __do_fast_syscall_32(struct pt_regs *regs)
139{
140 int nr = syscall_32_enter(regs);
141 int res;
142
143 add_random_kstack_offset();
144 /*
145 * This cannot use syscall_enter_from_user_mode() as it has to
146 * fetch EBP before invoking any of the syscall entry work
147 * functions.
148 */
149 syscall_enter_from_user_mode_prepare(regs);
150
151 instrumentation_begin();
152 /* Fetch EBP from where the vDSO stashed it. */
153 if (IS_ENABLED(CONFIG_X86_64)) {
154 /*
155 * Micro-optimization: the pointer we're following is
156 * explicitly 32 bits, so it can't be out of range.
157 */
158 res = __get_user(*(u32 *)®s->bp,
159 (u32 __user __force *)(unsigned long)(u32)regs->sp);
160 } else {
161 res = get_user(*(u32 *)®s->bp,
162 (u32 __user __force *)(unsigned long)(u32)regs->sp);
163 }
164
165 if (res) {
166 /* User code screwed up. */
167 regs->ax = -EFAULT;
168
169 local_irq_disable();
170 instrumentation_end();
171 irqentry_exit_to_user_mode(regs);
172 return false;
173 }
174
175 nr = syscall_enter_from_user_mode_work(regs, nr);
176
177 /* Now this is just like a normal syscall. */
178 do_syscall_32_irqs_on(regs, nr);
179
180 instrumentation_end();
181 syscall_exit_to_user_mode(regs);
182 return true;
183}
184
185/* Returns 0 to return using IRET or 1 to return using SYSEXIT/SYSRETL. */
186__visible noinstr long do_fast_syscall_32(struct pt_regs *regs)
187{
188 /*
189 * Called using the internal vDSO SYSENTER/SYSCALL32 calling
190 * convention. Adjust regs so it looks like we entered using int80.
191 */
192 unsigned long landing_pad = (unsigned long)current->mm->context.vdso +
193 vdso_image_32.sym_int80_landing_pad;
194
195 /*
196 * SYSENTER loses EIP, and even SYSCALL32 needs us to skip forward
197 * so that 'regs->ip -= 2' lands back on an int $0x80 instruction.
198 * Fix it up.
199 */
200 regs->ip = landing_pad;
201
202 /* Invoke the syscall. If it failed, keep it simple: use IRET. */
203 if (!__do_fast_syscall_32(regs))
204 return 0;
205
206#ifdef CONFIG_X86_64
207 /*
208 * Opportunistic SYSRETL: if possible, try to return using SYSRETL.
209 * SYSRETL is available on all 64-bit CPUs, so we don't need to
210 * bother with SYSEXIT.
211 *
212 * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
213 * because the ECX fixup above will ensure that this is essentially
214 * never the case.
215 */
216 return regs->cs == __USER32_CS && regs->ss == __USER_DS &&
217 regs->ip == landing_pad &&
218 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF)) == 0;
219#else
220 /*
221 * Opportunistic SYSEXIT: if possible, try to return using SYSEXIT.
222 *
223 * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
224 * because the ECX fixup above will ensure that this is essentially
225 * never the case.
226 *
227 * We don't allow syscalls at all from VM86 mode, but we still
228 * need to check VM, because we might be returning from sys_vm86.
229 */
230 return static_cpu_has(X86_FEATURE_SEP) &&
231 regs->cs == __USER_CS && regs->ss == __USER_DS &&
232 regs->ip == landing_pad &&
233 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF | X86_EFLAGS_VM)) == 0;
234#endif
235}
236
237/* Returns 0 to return using IRET or 1 to return using SYSEXIT/SYSRETL. */
238__visible noinstr long do_SYSENTER_32(struct pt_regs *regs)
239{
240 /* SYSENTER loses RSP, but the vDSO saved it in RBP. */
241 regs->sp = regs->bp;
242
243 /* SYSENTER clobbers EFLAGS.IF. Assume it was set in usermode. */
244 regs->flags |= X86_EFLAGS_IF;
245
246 return do_fast_syscall_32(regs);
247}
248#endif
249
250SYSCALL_DEFINE0(ni_syscall)
251{
252 return -ENOSYS;
253}
254
255#ifdef CONFIG_XEN_PV
256#ifndef CONFIG_PREEMPTION
257/*
258 * Some hypercalls issued by the toolstack can take many 10s of
259 * seconds. Allow tasks running hypercalls via the privcmd driver to
260 * be voluntarily preempted even if full kernel preemption is
261 * disabled.
262 *
263 * Such preemptible hypercalls are bracketed by
264 * xen_preemptible_hcall_begin() and xen_preemptible_hcall_end()
265 * calls.
266 */
267DEFINE_PER_CPU(bool, xen_in_preemptible_hcall);
268EXPORT_SYMBOL_GPL(xen_in_preemptible_hcall);
269
270/*
271 * In case of scheduling the flag must be cleared and restored after
272 * returning from schedule as the task might move to a different CPU.
273 */
274static __always_inline bool get_and_clear_inhcall(void)
275{
276 bool inhcall = __this_cpu_read(xen_in_preemptible_hcall);
277
278 __this_cpu_write(xen_in_preemptible_hcall, false);
279 return inhcall;
280}
281
282static __always_inline void restore_inhcall(bool inhcall)
283{
284 __this_cpu_write(xen_in_preemptible_hcall, inhcall);
285}
286#else
287static __always_inline bool get_and_clear_inhcall(void) { return false; }
288static __always_inline void restore_inhcall(bool inhcall) { }
289#endif
290
291static void __xen_pv_evtchn_do_upcall(struct pt_regs *regs)
292{
293 struct pt_regs *old_regs = set_irq_regs(regs);
294
295 inc_irq_stat(irq_hv_callback_count);
296
297 xen_hvm_evtchn_do_upcall();
298
299 set_irq_regs(old_regs);
300}
301
302__visible noinstr void xen_pv_evtchn_do_upcall(struct pt_regs *regs)
303{
304 irqentry_state_t state = irqentry_enter(regs);
305 bool inhcall;
306
307 instrumentation_begin();
308 run_sysvec_on_irqstack_cond(__xen_pv_evtchn_do_upcall, regs);
309
310 inhcall = get_and_clear_inhcall();
311 if (inhcall && !WARN_ON_ONCE(state.exit_rcu)) {
312 irqentry_exit_cond_resched();
313 instrumentation_end();
314 restore_inhcall(inhcall);
315 } else {
316 instrumentation_end();
317 irqentry_exit(regs, state);
318 }
319}
320#endif /* CONFIG_XEN_PV */
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * common.c - C code for kernel entry and exit
4 * Copyright (c) 2015 Andrew Lutomirski
5 *
6 * Based on asm and ptrace code by many authors. The code here originated
7 * in ptrace.c and signal.c.
8 */
9
10#include <linux/kernel.h>
11#include <linux/sched.h>
12#include <linux/sched/task_stack.h>
13#include <linux/entry-common.h>
14#include <linux/mm.h>
15#include <linux/smp.h>
16#include <linux/errno.h>
17#include <linux/ptrace.h>
18#include <linux/export.h>
19#include <linux/nospec.h>
20#include <linux/syscalls.h>
21#include <linux/uaccess.h>
22
23#ifdef CONFIG_XEN_PV
24#include <xen/xen-ops.h>
25#include <xen/events.h>
26#endif
27
28#include <asm/desc.h>
29#include <asm/traps.h>
30#include <asm/vdso.h>
31#include <asm/cpufeature.h>
32#include <asm/fpu/api.h>
33#include <asm/nospec-branch.h>
34#include <asm/io_bitmap.h>
35#include <asm/syscall.h>
36#include <asm/irq_stack.h>
37
38#ifdef CONFIG_X86_64
39__visible noinstr void do_syscall_64(unsigned long nr, struct pt_regs *regs)
40{
41 nr = syscall_enter_from_user_mode(regs, nr);
42
43 instrumentation_begin();
44 if (likely(nr < NR_syscalls)) {
45 nr = array_index_nospec(nr, NR_syscalls);
46 regs->ax = sys_call_table[nr](regs);
47#ifdef CONFIG_X86_X32_ABI
48 } else if (likely((nr & __X32_SYSCALL_BIT) &&
49 (nr & ~__X32_SYSCALL_BIT) < X32_NR_syscalls)) {
50 nr = array_index_nospec(nr & ~__X32_SYSCALL_BIT,
51 X32_NR_syscalls);
52 regs->ax = x32_sys_call_table[nr](regs);
53#endif
54 }
55 instrumentation_end();
56 syscall_exit_to_user_mode(regs);
57}
58#endif
59
60#if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION)
61static __always_inline unsigned int syscall_32_enter(struct pt_regs *regs)
62{
63 if (IS_ENABLED(CONFIG_IA32_EMULATION))
64 current_thread_info()->status |= TS_COMPAT;
65
66 return (unsigned int)regs->orig_ax;
67}
68
69/*
70 * Invoke a 32-bit syscall. Called with IRQs on in CONTEXT_KERNEL.
71 */
72static __always_inline void do_syscall_32_irqs_on(struct pt_regs *regs,
73 unsigned int nr)
74{
75 if (likely(nr < IA32_NR_syscalls)) {
76 instrumentation_begin();
77 nr = array_index_nospec(nr, IA32_NR_syscalls);
78 regs->ax = ia32_sys_call_table[nr](regs);
79 instrumentation_end();
80 }
81}
82
83/* Handles int $0x80 */
84__visible noinstr void do_int80_syscall_32(struct pt_regs *regs)
85{
86 unsigned int nr = syscall_32_enter(regs);
87
88 /*
89 * Subtlety here: if ptrace pokes something larger than 2^32-1 into
90 * orig_ax, the unsigned int return value truncates it. This may
91 * or may not be necessary, but it matches the old asm behavior.
92 */
93 nr = (unsigned int)syscall_enter_from_user_mode(regs, nr);
94
95 do_syscall_32_irqs_on(regs, nr);
96 syscall_exit_to_user_mode(regs);
97}
98
99static noinstr bool __do_fast_syscall_32(struct pt_regs *regs)
100{
101 unsigned int nr = syscall_32_enter(regs);
102 int res;
103
104 /*
105 * This cannot use syscall_enter_from_user_mode() as it has to
106 * fetch EBP before invoking any of the syscall entry work
107 * functions.
108 */
109 syscall_enter_from_user_mode_prepare(regs);
110
111 instrumentation_begin();
112 /* Fetch EBP from where the vDSO stashed it. */
113 if (IS_ENABLED(CONFIG_X86_64)) {
114 /*
115 * Micro-optimization: the pointer we're following is
116 * explicitly 32 bits, so it can't be out of range.
117 */
118 res = __get_user(*(u32 *)®s->bp,
119 (u32 __user __force *)(unsigned long)(u32)regs->sp);
120 } else {
121 res = get_user(*(u32 *)®s->bp,
122 (u32 __user __force *)(unsigned long)(u32)regs->sp);
123 }
124 instrumentation_end();
125
126 if (res) {
127 /* User code screwed up. */
128 regs->ax = -EFAULT;
129 syscall_exit_to_user_mode(regs);
130 return false;
131 }
132
133 /* The case truncates any ptrace induced syscall nr > 2^32 -1 */
134 nr = (unsigned int)syscall_enter_from_user_mode_work(regs, nr);
135
136 /* Now this is just like a normal syscall. */
137 do_syscall_32_irqs_on(regs, nr);
138 syscall_exit_to_user_mode(regs);
139 return true;
140}
141
142/* Returns 0 to return using IRET or 1 to return using SYSEXIT/SYSRETL. */
143__visible noinstr long do_fast_syscall_32(struct pt_regs *regs)
144{
145 /*
146 * Called using the internal vDSO SYSENTER/SYSCALL32 calling
147 * convention. Adjust regs so it looks like we entered using int80.
148 */
149 unsigned long landing_pad = (unsigned long)current->mm->context.vdso +
150 vdso_image_32.sym_int80_landing_pad;
151
152 /*
153 * SYSENTER loses EIP, and even SYSCALL32 needs us to skip forward
154 * so that 'regs->ip -= 2' lands back on an int $0x80 instruction.
155 * Fix it up.
156 */
157 regs->ip = landing_pad;
158
159 /* Invoke the syscall. If it failed, keep it simple: use IRET. */
160 if (!__do_fast_syscall_32(regs))
161 return 0;
162
163#ifdef CONFIG_X86_64
164 /*
165 * Opportunistic SYSRETL: if possible, try to return using SYSRETL.
166 * SYSRETL is available on all 64-bit CPUs, so we don't need to
167 * bother with SYSEXIT.
168 *
169 * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
170 * because the ECX fixup above will ensure that this is essentially
171 * never the case.
172 */
173 return regs->cs == __USER32_CS && regs->ss == __USER_DS &&
174 regs->ip == landing_pad &&
175 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF)) == 0;
176#else
177 /*
178 * Opportunistic SYSEXIT: if possible, try to return using SYSEXIT.
179 *
180 * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
181 * because the ECX fixup above will ensure that this is essentially
182 * never the case.
183 *
184 * We don't allow syscalls at all from VM86 mode, but we still
185 * need to check VM, because we might be returning from sys_vm86.
186 */
187 return static_cpu_has(X86_FEATURE_SEP) &&
188 regs->cs == __USER_CS && regs->ss == __USER_DS &&
189 regs->ip == landing_pad &&
190 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF | X86_EFLAGS_VM)) == 0;
191#endif
192}
193
194/* Returns 0 to return using IRET or 1 to return using SYSEXIT/SYSRETL. */
195__visible noinstr long do_SYSENTER_32(struct pt_regs *regs)
196{
197 /* SYSENTER loses RSP, but the vDSO saved it in RBP. */
198 regs->sp = regs->bp;
199
200 /* SYSENTER clobbers EFLAGS.IF. Assume it was set in usermode. */
201 regs->flags |= X86_EFLAGS_IF;
202
203 return do_fast_syscall_32(regs);
204}
205#endif
206
207SYSCALL_DEFINE0(ni_syscall)
208{
209 return -ENOSYS;
210}
211
212noinstr bool idtentry_enter_nmi(struct pt_regs *regs)
213{
214 bool irq_state = lockdep_hardirqs_enabled();
215
216 __nmi_enter();
217 lockdep_hardirqs_off(CALLER_ADDR0);
218 lockdep_hardirq_enter();
219 rcu_nmi_enter();
220
221 instrumentation_begin();
222 trace_hardirqs_off_finish();
223 ftrace_nmi_enter();
224 instrumentation_end();
225
226 return irq_state;
227}
228
229noinstr void idtentry_exit_nmi(struct pt_regs *regs, bool restore)
230{
231 instrumentation_begin();
232 ftrace_nmi_exit();
233 if (restore) {
234 trace_hardirqs_on_prepare();
235 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
236 }
237 instrumentation_end();
238
239 rcu_nmi_exit();
240 lockdep_hardirq_exit();
241 if (restore)
242 lockdep_hardirqs_on(CALLER_ADDR0);
243 __nmi_exit();
244}
245
246#ifdef CONFIG_XEN_PV
247#ifndef CONFIG_PREEMPTION
248/*
249 * Some hypercalls issued by the toolstack can take many 10s of
250 * seconds. Allow tasks running hypercalls via the privcmd driver to
251 * be voluntarily preempted even if full kernel preemption is
252 * disabled.
253 *
254 * Such preemptible hypercalls are bracketed by
255 * xen_preemptible_hcall_begin() and xen_preemptible_hcall_end()
256 * calls.
257 */
258DEFINE_PER_CPU(bool, xen_in_preemptible_hcall);
259EXPORT_SYMBOL_GPL(xen_in_preemptible_hcall);
260
261/*
262 * In case of scheduling the flag must be cleared and restored after
263 * returning from schedule as the task might move to a different CPU.
264 */
265static __always_inline bool get_and_clear_inhcall(void)
266{
267 bool inhcall = __this_cpu_read(xen_in_preemptible_hcall);
268
269 __this_cpu_write(xen_in_preemptible_hcall, false);
270 return inhcall;
271}
272
273static __always_inline void restore_inhcall(bool inhcall)
274{
275 __this_cpu_write(xen_in_preemptible_hcall, inhcall);
276}
277#else
278static __always_inline bool get_and_clear_inhcall(void) { return false; }
279static __always_inline void restore_inhcall(bool inhcall) { }
280#endif
281
282static void __xen_pv_evtchn_do_upcall(void)
283{
284 irq_enter_rcu();
285 inc_irq_stat(irq_hv_callback_count);
286
287 xen_hvm_evtchn_do_upcall();
288
289 irq_exit_rcu();
290}
291
292__visible noinstr void xen_pv_evtchn_do_upcall(struct pt_regs *regs)
293{
294 struct pt_regs *old_regs;
295 bool inhcall;
296 irqentry_state_t state;
297
298 state = irqentry_enter(regs);
299 old_regs = set_irq_regs(regs);
300
301 instrumentation_begin();
302 run_on_irqstack_cond(__xen_pv_evtchn_do_upcall, regs);
303 instrumentation_begin();
304
305 set_irq_regs(old_regs);
306
307 inhcall = get_and_clear_inhcall();
308 if (inhcall && !WARN_ON_ONCE(state.exit_rcu)) {
309 instrumentation_begin();
310 irqentry_exit_cond_resched();
311 instrumentation_end();
312 restore_inhcall(inhcall);
313 } else {
314 irqentry_exit(regs, state);
315 }
316}
317#endif /* CONFIG_XEN_PV */