xref: /openbmc/linux/arch/arm64/kernel/syscall.c (revision 0bf49ffb)
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/syscalls.h>
9 
10 #include <asm/daifflags.h>
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 #ifdef CONFIG_COMPAT
24 	long ret;
25 	if (is_compat_task()) {
26 		ret = compat_arm_syscall(regs, scno);
27 		if (ret != -ENOSYS)
28 			return ret;
29 	}
30 #endif
31 
32 	return sys_ni_syscall();
33 }
34 
35 static long __invoke_syscall(struct pt_regs *regs, syscall_fn_t syscall_fn)
36 {
37 	return syscall_fn(regs);
38 }
39 
40 static void invoke_syscall(struct pt_regs *regs, unsigned int scno,
41 			   unsigned int sc_nr,
42 			   const syscall_fn_t syscall_table[])
43 {
44 	long ret;
45 
46 	if (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 	if (is_compat_task())
55 		ret = lower_32_bits(ret);
56 
57 	regs->regs[0] = ret;
58 }
59 
60 static inline bool has_syscall_work(unsigned long flags)
61 {
62 	return unlikely(flags & _TIF_SYSCALL_WORK);
63 }
64 
65 int syscall_trace_enter(struct pt_regs *regs);
66 void syscall_trace_exit(struct pt_regs *regs);
67 
68 static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr,
69 			   const syscall_fn_t syscall_table[])
70 {
71 	unsigned long flags = current_thread_info()->flags;
72 
73 	regs->orig_x0 = regs->regs[0];
74 	regs->syscallno = scno;
75 
76 	/*
77 	 * BTI note:
78 	 * The architecture does not guarantee that SPSR.BTYPE is zero
79 	 * on taking an SVC, so we could return to userspace with a
80 	 * non-zero BTYPE after the syscall.
81 	 *
82 	 * This shouldn't matter except when userspace is explicitly
83 	 * doing something stupid, such as setting PROT_BTI on a page
84 	 * that lacks conforming BTI/PACIxSP instructions, falling
85 	 * through from one executable page to another with differing
86 	 * PROT_BTI, or messing with BTYPE via ptrace: in such cases,
87 	 * userspace should not be surprised if a SIGILL occurs on
88 	 * syscall return.
89 	 *
90 	 * So, don't touch regs->pstate & PSR_BTYPE_MASK here.
91 	 * (Similarly for HVC and SMC elsewhere.)
92 	 */
93 
94 	local_daif_restore(DAIF_PROCCTX);
95 
96 	if (flags & _TIF_MTE_ASYNC_FAULT) {
97 		/*
98 		 * Process the asynchronous tag check fault before the actual
99 		 * syscall. do_notify_resume() will send a signal to userspace
100 		 * before the syscall is restarted.
101 		 */
102 		regs->regs[0] = -ERESTARTNOINTR;
103 		return;
104 	}
105 
106 	if (has_syscall_work(flags)) {
107 		/*
108 		 * The de-facto standard way to skip a system call using ptrace
109 		 * is to set the system call to -1 (NO_SYSCALL) and set x0 to a
110 		 * suitable error code for consumption by userspace. However,
111 		 * this cannot be distinguished from a user-issued syscall(-1)
112 		 * and so we must set x0 to -ENOSYS here in case the tracer doesn't
113 		 * issue the skip and we fall into trace_exit with x0 preserved.
114 		 *
115 		 * This is slightly odd because it also means that if a tracer
116 		 * sets the system call number to -1 but does not initialise x0,
117 		 * then x0 will be preserved for all system calls apart from a
118 		 * user-issued syscall(-1). However, requesting a skip and not
119 		 * setting the return value is unlikely to do anything sensible
120 		 * anyway.
121 		 */
122 		if (scno == NO_SYSCALL)
123 			regs->regs[0] = -ENOSYS;
124 		scno = syscall_trace_enter(regs);
125 		if (scno == NO_SYSCALL)
126 			goto trace_exit;
127 	}
128 
129 	invoke_syscall(regs, scno, sc_nr, syscall_table);
130 
131 	/*
132 	 * The tracing status may have changed under our feet, so we have to
133 	 * check again. However, if we were tracing entry, then we always trace
134 	 * exit regardless, as the old entry assembly did.
135 	 */
136 	if (!has_syscall_work(flags) && !IS_ENABLED(CONFIG_DEBUG_RSEQ)) {
137 		local_daif_mask();
138 		flags = current_thread_info()->flags;
139 		if (!has_syscall_work(flags) && !(flags & _TIF_SINGLESTEP))
140 			return;
141 		local_daif_restore(DAIF_PROCCTX);
142 	}
143 
144 trace_exit:
145 	syscall_trace_exit(regs);
146 }
147 
148 static inline void sve_user_discard(void)
149 {
150 	if (!system_supports_sve())
151 		return;
152 
153 	clear_thread_flag(TIF_SVE);
154 
155 	/*
156 	 * task_fpsimd_load() won't be called to update CPACR_EL1 in
157 	 * ret_to_user unless TIF_FOREIGN_FPSTATE is still set, which only
158 	 * happens if a context switch or kernel_neon_begin() or context
159 	 * modification (sigreturn, ptrace) intervenes.
160 	 * So, ensure that CPACR_EL1 is already correct for the fast-path case.
161 	 */
162 	sve_user_disable();
163 }
164 
165 void do_el0_svc(struct pt_regs *regs)
166 {
167 	sve_user_discard();
168 	el0_svc_common(regs, regs->regs[8], __NR_syscalls, sys_call_table);
169 }
170 
171 #ifdef CONFIG_COMPAT
172 void do_el0_svc_compat(struct pt_regs *regs)
173 {
174 	el0_svc_common(regs, regs->regs[7], __NR_compat_syscalls,
175 		       compat_sys_call_table);
176 }
177 #endif
178