xref: /linux/arch/arm64/kernel/syscall.c (revision 40286d6379aacfcc053253ef78dc78b09addffda)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include <linux/compiler.h>
4 #include <linux/context_tracking.h>
5 #include <linux/errno.h>
6 #include <linux/nospec.h>
7 #include <linux/ptrace.h>
8 #include <linux/randomize_kstack.h>
9 #include <linux/syscalls.h>
10 
11 #include <asm/debug-monitors.h>
12 #include <asm/exception.h>
13 #include <asm/fpsimd.h>
14 #include <asm/syscall.h>
15 #include <asm/thread_info.h>
16 #include <asm/unistd.h>
17 #include <asm/unistd_compat_32.h>
18 
19 long compat_arm_syscall(struct pt_regs *regs, int scno);
20 long sys_ni_syscall(void);
21 
22 static long do_ni_syscall(struct pt_regs *regs, int scno)
23 {
24 	if (is_compat_task()) {
25 		long ret = compat_arm_syscall(regs, scno);
26 		if (ret != -ENOSYS)
27 			return ret;
28 	}
29 
30 	return sys_ni_syscall();
31 }
32 
33 static long __invoke_syscall(struct pt_regs *regs, syscall_fn_t syscall_fn)
34 {
35 	return syscall_fn(regs);
36 }
37 
38 static void invoke_syscall(struct pt_regs *regs, unsigned int scno,
39 			   unsigned int sc_nr,
40 			   const syscall_fn_t syscall_table[])
41 {
42 	long ret;
43 
44 	add_random_kstack_offset();
45 
46 	if (likely(scno < sc_nr)) {
47 		syscall_fn_t syscall_fn;
48 		syscall_fn = syscall_table[array_index_nospec(scno, sc_nr)];
49 		ret = __invoke_syscall(regs, syscall_fn);
50 	} else {
51 		ret = do_ni_syscall(regs, scno);
52 	}
53 
54 	syscall_set_return_value(current, regs, 0, ret);
55 }
56 
57 static inline bool has_syscall_work(unsigned long flags)
58 {
59 	return unlikely(flags & _TIF_SYSCALL_WORK);
60 }
61 
62 static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr,
63 			   const syscall_fn_t syscall_table[])
64 {
65 	unsigned long flags = read_thread_flags();
66 
67 	regs->orig_x0 = regs->regs[0];
68 	regs->syscallno = scno;
69 
70 	/*
71 	 * BTI note:
72 	 * The architecture does not guarantee that SPSR.BTYPE is zero
73 	 * on taking an SVC, so we could return to userspace with a
74 	 * non-zero BTYPE after the syscall.
75 	 *
76 	 * This shouldn't matter except when userspace is explicitly
77 	 * doing something stupid, such as setting PROT_BTI on a page
78 	 * that lacks conforming BTI/PACIxSP instructions, falling
79 	 * through from one executable page to another with differing
80 	 * PROT_BTI, or messing with BTYPE via ptrace: in such cases,
81 	 * userspace should not be surprised if a SIGILL occurs on
82 	 * syscall return.
83 	 *
84 	 * So, don't touch regs->pstate & PSR_BTYPE_MASK here.
85 	 * (Similarly for HVC and SMC elsewhere.)
86 	 */
87 
88 	if (unlikely(flags & _TIF_MTE_ASYNC_FAULT)) {
89 		/*
90 		 * Process the asynchronous tag check fault before the actual
91 		 * syscall. do_notify_resume() will send a signal to userspace
92 		 * before the syscall is restarted.
93 		 */
94 		syscall_set_return_value(current, regs, -ERESTARTNOINTR, 0);
95 		return;
96 	}
97 
98 	if (has_syscall_work(flags)) {
99 		/*
100 		 * The de-facto standard way to skip a system call using ptrace
101 		 * is to set the system call to -1 (NO_SYSCALL) and set x0 to a
102 		 * suitable error code for consumption by userspace. However,
103 		 * this cannot be distinguished from a user-issued syscall(-1)
104 		 * and so we must set x0 to -ENOSYS here in case the tracer doesn't
105 		 * issue the skip and we fall into trace_exit with x0 preserved.
106 		 *
107 		 * This is slightly odd because it also means that if a tracer
108 		 * sets the system call number to -1 but does not initialise x0,
109 		 * then x0 will be preserved for all system calls apart from a
110 		 * user-issued syscall(-1). However, requesting a skip and not
111 		 * setting the return value is unlikely to do anything sensible
112 		 * anyway.
113 		 */
114 		if (scno == NO_SYSCALL)
115 			syscall_set_return_value(current, regs, -ENOSYS, 0);
116 		scno = syscall_trace_enter(regs);
117 		if (scno == NO_SYSCALL)
118 			goto trace_exit;
119 	}
120 
121 	invoke_syscall(regs, scno, sc_nr, syscall_table);
122 
123 	/*
124 	 * The tracing status may have changed under our feet, so we have to
125 	 * check again. However, if we were tracing entry, then we always trace
126 	 * exit regardless, as the old entry assembly did.
127 	 */
128 	if (!has_syscall_work(flags) && !IS_ENABLED(CONFIG_DEBUG_RSEQ)) {
129 		flags = read_thread_flags();
130 		if (!has_syscall_work(flags) && !(flags & _TIF_SINGLESTEP))
131 			return;
132 	}
133 
134 trace_exit:
135 	syscall_trace_exit(regs);
136 }
137 
138 void do_el0_svc(struct pt_regs *regs)
139 {
140 	el0_svc_common(regs, regs->regs[8], __NR_syscalls, sys_call_table);
141 }
142 
143 #ifdef CONFIG_COMPAT
144 void do_el0_svc_compat(struct pt_regs *regs)
145 {
146 	el0_svc_common(regs, regs->regs[7], __NR_compat32_syscalls,
147 		       compat_sys_call_table);
148 }
149 #endif
150