xref: /openbmc/linux/arch/arm64/kernel/syscall.c (revision c4f7ac64)
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/daifflags.h>
12 #include <asm/debug-monitors.h>
13 #include <asm/exception.h>
14 #include <asm/fpsimd.h>
15 #include <asm/syscall.h>
16 #include <asm/thread_info.h>
17 #include <asm/unistd.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 #ifdef CONFIG_COMPAT
25 	long ret;
26 	if (is_compat_task()) {
27 		ret = compat_arm_syscall(regs, scno);
28 		if (ret != -ENOSYS)
29 			return ret;
30 	}
31 #endif
32 
33 	return sys_ni_syscall();
34 }
35 
36 static long __invoke_syscall(struct pt_regs *regs, syscall_fn_t syscall_fn)
37 {
38 	return syscall_fn(regs);
39 }
40 
41 static void invoke_syscall(struct pt_regs *regs, unsigned int scno,
42 			   unsigned int sc_nr,
43 			   const syscall_fn_t syscall_table[])
44 {
45 	long ret;
46 
47 	add_random_kstack_offset();
48 
49 	if (scno < sc_nr) {
50 		syscall_fn_t syscall_fn;
51 		syscall_fn = syscall_table[array_index_nospec(scno, sc_nr)];
52 		ret = __invoke_syscall(regs, syscall_fn);
53 	} else {
54 		ret = do_ni_syscall(regs, scno);
55 	}
56 
57 	if (is_compat_task())
58 		ret = lower_32_bits(ret);
59 
60 	regs->regs[0] = ret;
61 
62 	/*
63 	 * Ultimately, this value will get limited by KSTACK_OFFSET_MAX(),
64 	 * but not enough for arm64 stack utilization comfort. To keep
65 	 * reasonable stack head room, reduce the maximum offset to 9 bits.
66 	 *
67 	 * The actual entropy will be further reduced by the compiler when
68 	 * applying stack alignment constraints: the AAPCS mandates a
69 	 * 16-byte (i.e. 4-bit) aligned SP at function boundaries.
70 	 *
71 	 * The resulting 5 bits of entropy is seen in SP[8:4].
72 	 */
73 	choose_random_kstack_offset(get_random_int() & 0x1FF);
74 }
75 
76 static inline bool has_syscall_work(unsigned long flags)
77 {
78 	return unlikely(flags & _TIF_SYSCALL_WORK);
79 }
80 
81 int syscall_trace_enter(struct pt_regs *regs);
82 void syscall_trace_exit(struct pt_regs *regs);
83 
84 static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr,
85 			   const syscall_fn_t syscall_table[])
86 {
87 	unsigned long flags = current_thread_info()->flags;
88 
89 	regs->orig_x0 = regs->regs[0];
90 	regs->syscallno = scno;
91 
92 	/*
93 	 * BTI note:
94 	 * The architecture does not guarantee that SPSR.BTYPE is zero
95 	 * on taking an SVC, so we could return to userspace with a
96 	 * non-zero BTYPE after the syscall.
97 	 *
98 	 * This shouldn't matter except when userspace is explicitly
99 	 * doing something stupid, such as setting PROT_BTI on a page
100 	 * that lacks conforming BTI/PACIxSP instructions, falling
101 	 * through from one executable page to another with differing
102 	 * PROT_BTI, or messing with BTYPE via ptrace: in such cases,
103 	 * userspace should not be surprised if a SIGILL occurs on
104 	 * syscall return.
105 	 *
106 	 * So, don't touch regs->pstate & PSR_BTYPE_MASK here.
107 	 * (Similarly for HVC and SMC elsewhere.)
108 	 */
109 
110 	local_daif_restore(DAIF_PROCCTX);
111 
112 	if (flags & _TIF_MTE_ASYNC_FAULT) {
113 		/*
114 		 * Process the asynchronous tag check fault before the actual
115 		 * syscall. do_notify_resume() will send a signal to userspace
116 		 * before the syscall is restarted.
117 		 */
118 		regs->regs[0] = -ERESTARTNOINTR;
119 		return;
120 	}
121 
122 	if (has_syscall_work(flags)) {
123 		/*
124 		 * The de-facto standard way to skip a system call using ptrace
125 		 * is to set the system call to -1 (NO_SYSCALL) and set x0 to a
126 		 * suitable error code for consumption by userspace. However,
127 		 * this cannot be distinguished from a user-issued syscall(-1)
128 		 * and so we must set x0 to -ENOSYS here in case the tracer doesn't
129 		 * issue the skip and we fall into trace_exit with x0 preserved.
130 		 *
131 		 * This is slightly odd because it also means that if a tracer
132 		 * sets the system call number to -1 but does not initialise x0,
133 		 * then x0 will be preserved for all system calls apart from a
134 		 * user-issued syscall(-1). However, requesting a skip and not
135 		 * setting the return value is unlikely to do anything sensible
136 		 * anyway.
137 		 */
138 		if (scno == NO_SYSCALL)
139 			regs->regs[0] = -ENOSYS;
140 		scno = syscall_trace_enter(regs);
141 		if (scno == NO_SYSCALL)
142 			goto trace_exit;
143 	}
144 
145 	invoke_syscall(regs, scno, sc_nr, syscall_table);
146 
147 	/*
148 	 * The tracing status may have changed under our feet, so we have to
149 	 * check again. However, if we were tracing entry, then we always trace
150 	 * exit regardless, as the old entry assembly did.
151 	 */
152 	if (!has_syscall_work(flags) && !IS_ENABLED(CONFIG_DEBUG_RSEQ)) {
153 		local_daif_mask();
154 		flags = current_thread_info()->flags;
155 		if (!has_syscall_work(flags) && !(flags & _TIF_SINGLESTEP))
156 			return;
157 		local_daif_restore(DAIF_PROCCTX);
158 	}
159 
160 trace_exit:
161 	syscall_trace_exit(regs);
162 }
163 
164 static inline void sve_user_discard(void)
165 {
166 	if (!system_supports_sve())
167 		return;
168 
169 	clear_thread_flag(TIF_SVE);
170 
171 	/*
172 	 * task_fpsimd_load() won't be called to update CPACR_EL1 in
173 	 * ret_to_user unless TIF_FOREIGN_FPSTATE is still set, which only
174 	 * happens if a context switch or kernel_neon_begin() or context
175 	 * modification (sigreturn, ptrace) intervenes.
176 	 * So, ensure that CPACR_EL1 is already correct for the fast-path case.
177 	 */
178 	sve_user_disable();
179 }
180 
181 void do_el0_svc(struct pt_regs *regs)
182 {
183 	sve_user_discard();
184 	el0_svc_common(regs, regs->regs[8], __NR_syscalls, sys_call_table);
185 }
186 
187 #ifdef CONFIG_COMPAT
188 void do_el0_svc_compat(struct pt_regs *regs)
189 {
190 	el0_svc_common(regs, regs->regs[7], __NR_compat_syscalls,
191 		       compat_sys_call_table);
192 }
193 #endif
194