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/*
2 * common.c - C code for kernel entry and exit
3 * Copyright (c) 2015 Andrew Lutomirski
4 * GPL v2
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/mm.h>
13#include <linux/smp.h>
14#include <linux/errno.h>
15#include <linux/ptrace.h>
16#include <linux/tracehook.h>
17#include <linux/audit.h>
18#include <linux/seccomp.h>
19#include <linux/signal.h>
20#include <linux/export.h>
21#include <linux/context_tracking.h>
22#include <linux/user-return-notifier.h>
23#include <linux/uprobes.h>
24
25#include <asm/desc.h>
26#include <asm/traps.h>
27#include <asm/vdso.h>
28#include <linux/uaccess.h>
29#include <asm/cpufeature.h>
30
31#define CREATE_TRACE_POINTS
32#include <trace/events/syscalls.h>
33
34#ifdef CONFIG_CONTEXT_TRACKING
35/* Called on entry from user mode with IRQs off. */
36__visible inline void enter_from_user_mode(void)
37{
38 CT_WARN_ON(ct_state() != CONTEXT_USER);
39 user_exit_irqoff();
40}
41#else
42static inline void enter_from_user_mode(void) {}
43#endif
44
45static void do_audit_syscall_entry(struct pt_regs *regs, u32 arch)
46{
47#ifdef CONFIG_X86_64
48 if (arch == AUDIT_ARCH_X86_64) {
49 audit_syscall_entry(regs->orig_ax, regs->di,
50 regs->si, regs->dx, regs->r10);
51 } else
52#endif
53 {
54 audit_syscall_entry(regs->orig_ax, regs->bx,
55 regs->cx, regs->dx, regs->si);
56 }
57}
58
59/*
60 * Returns the syscall nr to run (which should match regs->orig_ax) or -1
61 * to skip the syscall.
62 */
63static long syscall_trace_enter(struct pt_regs *regs)
64{
65 u32 arch = in_ia32_syscall() ? AUDIT_ARCH_I386 : AUDIT_ARCH_X86_64;
66
67 struct thread_info *ti = current_thread_info();
68 unsigned long ret = 0;
69 bool emulated = false;
70 u32 work;
71
72 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
73 BUG_ON(regs != task_pt_regs(current));
74
75 work = ACCESS_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY;
76
77 if (unlikely(work & _TIF_SYSCALL_EMU))
78 emulated = true;
79
80 if ((emulated || (work & _TIF_SYSCALL_TRACE)) &&
81 tracehook_report_syscall_entry(regs))
82 return -1L;
83
84 if (emulated)
85 return -1L;
86
87#ifdef CONFIG_SECCOMP
88 /*
89 * Do seccomp after ptrace, to catch any tracer changes.
90 */
91 if (work & _TIF_SECCOMP) {
92 struct seccomp_data sd;
93
94 sd.arch = arch;
95 sd.nr = regs->orig_ax;
96 sd.instruction_pointer = regs->ip;
97#ifdef CONFIG_X86_64
98 if (arch == AUDIT_ARCH_X86_64) {
99 sd.args[0] = regs->di;
100 sd.args[1] = regs->si;
101 sd.args[2] = regs->dx;
102 sd.args[3] = regs->r10;
103 sd.args[4] = regs->r8;
104 sd.args[5] = regs->r9;
105 } else
106#endif
107 {
108 sd.args[0] = regs->bx;
109 sd.args[1] = regs->cx;
110 sd.args[2] = regs->dx;
111 sd.args[3] = regs->si;
112 sd.args[4] = regs->di;
113 sd.args[5] = regs->bp;
114 }
115
116 ret = __secure_computing(&sd);
117 if (ret == -1)
118 return ret;
119 }
120#endif
121
122 if (unlikely(test_thread_flag(TIF_SYSCALL_TRACEPOINT)))
123 trace_sys_enter(regs, regs->orig_ax);
124
125 do_audit_syscall_entry(regs, arch);
126
127 return ret ?: regs->orig_ax;
128}
129
130#define EXIT_TO_USERMODE_LOOP_FLAGS \
131 (_TIF_SIGPENDING | _TIF_NOTIFY_RESUME | _TIF_UPROBE | \
132 _TIF_NEED_RESCHED | _TIF_USER_RETURN_NOTIFY)
133
134static void exit_to_usermode_loop(struct pt_regs *regs, u32 cached_flags)
135{
136 /*
137 * In order to return to user mode, we need to have IRQs off with
138 * none of _TIF_SIGPENDING, _TIF_NOTIFY_RESUME, _TIF_USER_RETURN_NOTIFY,
139 * _TIF_UPROBE, or _TIF_NEED_RESCHED set. Several of these flags
140 * can be set at any time on preemptable kernels if we have IRQs on,
141 * so we need to loop. Disabling preemption wouldn't help: doing the
142 * work to clear some of the flags can sleep.
143 */
144 while (true) {
145 /* We have work to do. */
146 local_irq_enable();
147
148 if (cached_flags & _TIF_NEED_RESCHED)
149 schedule();
150
151 if (cached_flags & _TIF_UPROBE)
152 uprobe_notify_resume(regs);
153
154 /* deal with pending signal delivery */
155 if (cached_flags & _TIF_SIGPENDING)
156 do_signal(regs);
157
158 if (cached_flags & _TIF_NOTIFY_RESUME) {
159 clear_thread_flag(TIF_NOTIFY_RESUME);
160 tracehook_notify_resume(regs);
161 }
162
163 if (cached_flags & _TIF_USER_RETURN_NOTIFY)
164 fire_user_return_notifiers();
165
166 /* Disable IRQs and retry */
167 local_irq_disable();
168
169 cached_flags = READ_ONCE(current_thread_info()->flags);
170
171 if (!(cached_flags & EXIT_TO_USERMODE_LOOP_FLAGS))
172 break;
173 }
174}
175
176/* Called with IRQs disabled. */
177__visible inline void prepare_exit_to_usermode(struct pt_regs *regs)
178{
179 struct thread_info *ti = current_thread_info();
180 u32 cached_flags;
181
182 if (IS_ENABLED(CONFIG_PROVE_LOCKING) && WARN_ON(!irqs_disabled()))
183 local_irq_disable();
184
185 lockdep_sys_exit();
186
187 cached_flags = READ_ONCE(ti->flags);
188
189 if (unlikely(cached_flags & EXIT_TO_USERMODE_LOOP_FLAGS))
190 exit_to_usermode_loop(regs, cached_flags);
191
192#ifdef CONFIG_COMPAT
193 /*
194 * Compat syscalls set TS_COMPAT. Make sure we clear it before
195 * returning to user mode. We need to clear it *after* signal
196 * handling, because syscall restart has a fixup for compat
197 * syscalls. The fixup is exercised by the ptrace_syscall_32
198 * selftest.
199 *
200 * We also need to clear TS_REGS_POKED_I386: the 32-bit tracer
201 * special case only applies after poking regs and before the
202 * very next return to user mode.
203 */
204 current->thread.status &= ~(TS_COMPAT|TS_I386_REGS_POKED);
205#endif
206
207 user_enter_irqoff();
208}
209
210#define SYSCALL_EXIT_WORK_FLAGS \
211 (_TIF_SYSCALL_TRACE | _TIF_SYSCALL_AUDIT | \
212 _TIF_SINGLESTEP | _TIF_SYSCALL_TRACEPOINT)
213
214static void syscall_slow_exit_work(struct pt_regs *regs, u32 cached_flags)
215{
216 bool step;
217
218 audit_syscall_exit(regs);
219
220 if (cached_flags & _TIF_SYSCALL_TRACEPOINT)
221 trace_sys_exit(regs, regs->ax);
222
223 /*
224 * If TIF_SYSCALL_EMU is set, we only get here because of
225 * TIF_SINGLESTEP (i.e. this is PTRACE_SYSEMU_SINGLESTEP).
226 * We already reported this syscall instruction in
227 * syscall_trace_enter().
228 */
229 step = unlikely(
230 (cached_flags & (_TIF_SINGLESTEP | _TIF_SYSCALL_EMU))
231 == _TIF_SINGLESTEP);
232 if (step || cached_flags & _TIF_SYSCALL_TRACE)
233 tracehook_report_syscall_exit(regs, step);
234}
235
236/*
237 * Called with IRQs on and fully valid regs. Returns with IRQs off in a
238 * state such that we can immediately switch to user mode.
239 */
240__visible inline void syscall_return_slowpath(struct pt_regs *regs)
241{
242 struct thread_info *ti = current_thread_info();
243 u32 cached_flags = READ_ONCE(ti->flags);
244
245 CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
246
247 if (IS_ENABLED(CONFIG_PROVE_LOCKING) &&
248 WARN(irqs_disabled(), "syscall %ld left IRQs disabled", regs->orig_ax))
249 local_irq_enable();
250
251 /*
252 * First do one-time work. If these work items are enabled, we
253 * want to run them exactly once per syscall exit with IRQs on.
254 */
255 if (unlikely(cached_flags & SYSCALL_EXIT_WORK_FLAGS))
256 syscall_slow_exit_work(regs, cached_flags);
257
258 local_irq_disable();
259 prepare_exit_to_usermode(regs);
260}
261
262#ifdef CONFIG_X86_64
263__visible void do_syscall_64(struct pt_regs *regs)
264{
265 struct thread_info *ti = current_thread_info();
266 unsigned long nr = regs->orig_ax;
267
268 enter_from_user_mode();
269 local_irq_enable();
270
271 if (READ_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY)
272 nr = syscall_trace_enter(regs);
273
274 /*
275 * NB: Native and x32 syscalls are dispatched from the same
276 * table. The only functional difference is the x32 bit in
277 * regs->orig_ax, which changes the behavior of some syscalls.
278 */
279 if (likely((nr & __SYSCALL_MASK) < NR_syscalls)) {
280 regs->ax = sys_call_table[nr & __SYSCALL_MASK](
281 regs->di, regs->si, regs->dx,
282 regs->r10, regs->r8, regs->r9);
283 }
284
285 syscall_return_slowpath(regs);
286}
287#endif
288
289#if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION)
290/*
291 * Does a 32-bit syscall. Called with IRQs on in CONTEXT_KERNEL. Does
292 * all entry and exit work and returns with IRQs off. This function is
293 * extremely hot in workloads that use it, and it's usually called from
294 * do_fast_syscall_32, so forcibly inline it to improve performance.
295 */
296static __always_inline void do_syscall_32_irqs_on(struct pt_regs *regs)
297{
298 struct thread_info *ti = current_thread_info();
299 unsigned int nr = (unsigned int)regs->orig_ax;
300
301#ifdef CONFIG_IA32_EMULATION
302 current->thread.status |= TS_COMPAT;
303#endif
304
305 if (READ_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY) {
306 /*
307 * Subtlety here: if ptrace pokes something larger than
308 * 2^32-1 into orig_ax, this truncates it. This may or
309 * may not be necessary, but it matches the old asm
310 * behavior.
311 */
312 nr = syscall_trace_enter(regs);
313 }
314
315 if (likely(nr < IA32_NR_syscalls)) {
316 /*
317 * It's possible that a 32-bit syscall implementation
318 * takes a 64-bit parameter but nonetheless assumes that
319 * the high bits are zero. Make sure we zero-extend all
320 * of the args.
321 */
322 regs->ax = ia32_sys_call_table[nr](
323 (unsigned int)regs->bx, (unsigned int)regs->cx,
324 (unsigned int)regs->dx, (unsigned int)regs->si,
325 (unsigned int)regs->di, (unsigned int)regs->bp);
326 }
327
328 syscall_return_slowpath(regs);
329}
330
331/* Handles int $0x80 */
332__visible void do_int80_syscall_32(struct pt_regs *regs)
333{
334 enter_from_user_mode();
335 local_irq_enable();
336 do_syscall_32_irqs_on(regs);
337}
338
339/* Returns 0 to return using IRET or 1 to return using SYSEXIT/SYSRETL. */
340__visible long do_fast_syscall_32(struct pt_regs *regs)
341{
342 /*
343 * Called using the internal vDSO SYSENTER/SYSCALL32 calling
344 * convention. Adjust regs so it looks like we entered using int80.
345 */
346
347 unsigned long landing_pad = (unsigned long)current->mm->context.vdso +
348 vdso_image_32.sym_int80_landing_pad;
349
350 /*
351 * SYSENTER loses EIP, and even SYSCALL32 needs us to skip forward
352 * so that 'regs->ip -= 2' lands back on an int $0x80 instruction.
353 * Fix it up.
354 */
355 regs->ip = landing_pad;
356
357 enter_from_user_mode();
358
359 local_irq_enable();
360
361 /* Fetch EBP from where the vDSO stashed it. */
362 if (
363#ifdef CONFIG_X86_64
364 /*
365 * Micro-optimization: the pointer we're following is explicitly
366 * 32 bits, so it can't be out of range.
367 */
368 __get_user(*(u32 *)®s->bp,
369 (u32 __user __force *)(unsigned long)(u32)regs->sp)
370#else
371 get_user(*(u32 *)®s->bp,
372 (u32 __user __force *)(unsigned long)(u32)regs->sp)
373#endif
374 ) {
375
376 /* User code screwed up. */
377 local_irq_disable();
378 regs->ax = -EFAULT;
379 prepare_exit_to_usermode(regs);
380 return 0; /* Keep it simple: use IRET. */
381 }
382
383 /* Now this is just like a normal syscall. */
384 do_syscall_32_irqs_on(regs);
385
386#ifdef CONFIG_X86_64
387 /*
388 * Opportunistic SYSRETL: if possible, try to return using SYSRETL.
389 * SYSRETL is available on all 64-bit CPUs, so we don't need to
390 * bother with SYSEXIT.
391 *
392 * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
393 * because the ECX fixup above will ensure that this is essentially
394 * never the case.
395 */
396 return regs->cs == __USER32_CS && regs->ss == __USER_DS &&
397 regs->ip == landing_pad &&
398 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF)) == 0;
399#else
400 /*
401 * Opportunistic SYSEXIT: if possible, try to return using SYSEXIT.
402 *
403 * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
404 * because the ECX fixup above will ensure that this is essentially
405 * never the case.
406 *
407 * We don't allow syscalls at all from VM86 mode, but we still
408 * need to check VM, because we might be returning from sys_vm86.
409 */
410 return static_cpu_has(X86_FEATURE_SEP) &&
411 regs->cs == __USER_CS && regs->ss == __USER_DS &&
412 regs->ip == landing_pad &&
413 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF | X86_EFLAGS_VM)) == 0;
414#endif
415}
416#endif