xref: /linux/arch/arm64/kernel/syscall.c (revision 249ebf3f65f8530beb2cbfb91bff1d83ba88d23c)
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 
18 long compat_arm_syscall(struct pt_regs *regs, int scno);
19 long sys_ni_syscall(void);
20 
21 static long do_ni_syscall(struct pt_regs *regs, int scno)
22 {
23 	if (is_compat_task()) {
24 		long ret = compat_arm_syscall(regs, scno);
25 		if (ret != -ENOSYS)
26 			return ret;
27 	}
28 
29 	return sys_ni_syscall();
30 }
31 
32 static long __invoke_syscall(struct pt_regs *regs, syscall_fn_t syscall_fn)
33 {
34 	return syscall_fn(regs);
35 }
36 
37 static void invoke_syscall(struct pt_regs *regs, unsigned int scno,
38 			   unsigned int sc_nr,
39 			   const syscall_fn_t syscall_table[])
40 {
41 	long ret;
42 
43 	add_random_kstack_offset();
44 
45 	if (scno < sc_nr) {
46 		syscall_fn_t syscall_fn;
47 		syscall_fn = syscall_table[array_index_nospec(scno, sc_nr)];
48 		ret = __invoke_syscall(regs, syscall_fn);
49 	} else {
50 		ret = do_ni_syscall(regs, scno);
51 	}
52 
53 	syscall_set_return_value(current, regs, 0, ret);
54 
55 	/*
56 	 * Ultimately, this value will get limited by KSTACK_OFFSET_MAX(),
57 	 * but not enough for arm64 stack utilization comfort. To keep
58 	 * reasonable stack head room, reduce the maximum offset to 9 bits.
59 	 *
60 	 * The actual entropy will be further reduced by the compiler when
61 	 * applying stack alignment constraints: the AAPCS mandates a
62 	 * 16-byte (i.e. 4-bit) aligned SP at function boundaries.
63 	 *
64 	 * The resulting 5 bits of entropy is seen in SP[8:4].
65 	 */
66 	choose_random_kstack_offset(get_random_u16() & 0x1FF);
67 }
68 
69 static inline bool has_syscall_work(unsigned long flags)
70 {
71 	return unlikely(flags & _TIF_SYSCALL_WORK);
72 }
73 
74 static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr,
75 			   const syscall_fn_t syscall_table[])
76 {
77 	unsigned long flags = read_thread_flags();
78 
79 	regs->orig_x0 = regs->regs[0];
80 	regs->syscallno = scno;
81 
82 	/*
83 	 * BTI note:
84 	 * The architecture does not guarantee that SPSR.BTYPE is zero
85 	 * on taking an SVC, so we could return to userspace with a
86 	 * non-zero BTYPE after the syscall.
87 	 *
88 	 * This shouldn't matter except when userspace is explicitly
89 	 * doing something stupid, such as setting PROT_BTI on a page
90 	 * that lacks conforming BTI/PACIxSP instructions, falling
91 	 * through from one executable page to another with differing
92 	 * PROT_BTI, or messing with BTYPE via ptrace: in such cases,
93 	 * userspace should not be surprised if a SIGILL occurs on
94 	 * syscall return.
95 	 *
96 	 * So, don't touch regs->pstate & PSR_BTYPE_MASK here.
97 	 * (Similarly for HVC and SMC elsewhere.)
98 	 */
99 
100 	if (flags & _TIF_MTE_ASYNC_FAULT) {
101 		/*
102 		 * Process the asynchronous tag check fault before the actual
103 		 * syscall. do_notify_resume() will send a signal to userspace
104 		 * before the syscall is restarted.
105 		 */
106 		syscall_set_return_value(current, regs, -ERESTARTNOINTR, 0);
107 		return;
108 	}
109 
110 	if (has_syscall_work(flags)) {
111 		/*
112 		 * The de-facto standard way to skip a system call using ptrace
113 		 * is to set the system call to -1 (NO_SYSCALL) and set x0 to a
114 		 * suitable error code for consumption by userspace. However,
115 		 * this cannot be distinguished from a user-issued syscall(-1)
116 		 * and so we must set x0 to -ENOSYS here in case the tracer doesn't
117 		 * issue the skip and we fall into trace_exit with x0 preserved.
118 		 *
119 		 * This is slightly odd because it also means that if a tracer
120 		 * sets the system call number to -1 but does not initialise x0,
121 		 * then x0 will be preserved for all system calls apart from a
122 		 * user-issued syscall(-1). However, requesting a skip and not
123 		 * setting the return value is unlikely to do anything sensible
124 		 * anyway.
125 		 */
126 		if (scno == NO_SYSCALL)
127 			syscall_set_return_value(current, regs, -ENOSYS, 0);
128 		scno = syscall_trace_enter(regs);
129 		if (scno == NO_SYSCALL)
130 			goto trace_exit;
131 	}
132 
133 	invoke_syscall(regs, scno, sc_nr, syscall_table);
134 
135 	/*
136 	 * The tracing status may have changed under our feet, so we have to
137 	 * check again. However, if we were tracing entry, then we always trace
138 	 * exit regardless, as the old entry assembly did.
139 	 */
140 	if (!has_syscall_work(flags) && !IS_ENABLED(CONFIG_DEBUG_RSEQ)) {
141 		flags = read_thread_flags();
142 		if (!has_syscall_work(flags) && !(flags & _TIF_SINGLESTEP))
143 			return;
144 	}
145 
146 trace_exit:
147 	syscall_trace_exit(regs);
148 }
149 
150 void do_el0_svc(struct pt_regs *regs)
151 {
152 	el0_svc_common(regs, regs->regs[8], __NR_syscalls, sys_call_table);
153 }
154 
155 #ifdef CONFIG_COMPAT
156 void do_el0_svc_compat(struct pt_regs *regs)
157 {
158 	el0_svc_common(regs, regs->regs[7], __NR_compat_syscalls,
159 		       compat_sys_call_table);
160 }
161 #endif
162