xref: /openbmc/linux/arch/arm64/kernel/syscall.c (revision 0352f880)
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 	syscall_set_return_value(current, regs, 0, ret);
58 
59 	/*
60 	 * Ultimately, this value will get limited by KSTACK_OFFSET_MAX(),
61 	 * but not enough for arm64 stack utilization comfort. To keep
62 	 * reasonable stack head room, reduce the maximum offset to 9 bits.
63 	 *
64 	 * The actual entropy will be further reduced by the compiler when
65 	 * applying stack alignment constraints: the AAPCS mandates a
66 	 * 16-byte (i.e. 4-bit) aligned SP at function boundaries.
67 	 *
68 	 * The resulting 5 bits of entropy is seen in SP[8:4].
69 	 */
70 	choose_random_kstack_offset(get_random_u16() & 0x1FF);
71 }
72 
73 static inline bool has_syscall_work(unsigned long flags)
74 {
75 	return unlikely(flags & _TIF_SYSCALL_WORK);
76 }
77 
78 int syscall_trace_enter(struct pt_regs *regs);
79 void syscall_trace_exit(struct pt_regs *regs);
80 
81 static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr,
82 			   const syscall_fn_t syscall_table[])
83 {
84 	unsigned long flags = read_thread_flags();
85 
86 	regs->orig_x0 = regs->regs[0];
87 	regs->syscallno = scno;
88 
89 	/*
90 	 * BTI note:
91 	 * The architecture does not guarantee that SPSR.BTYPE is zero
92 	 * on taking an SVC, so we could return to userspace with a
93 	 * non-zero BTYPE after the syscall.
94 	 *
95 	 * This shouldn't matter except when userspace is explicitly
96 	 * doing something stupid, such as setting PROT_BTI on a page
97 	 * that lacks conforming BTI/PACIxSP instructions, falling
98 	 * through from one executable page to another with differing
99 	 * PROT_BTI, or messing with BTYPE via ptrace: in such cases,
100 	 * userspace should not be surprised if a SIGILL occurs on
101 	 * syscall return.
102 	 *
103 	 * So, don't touch regs->pstate & PSR_BTYPE_MASK here.
104 	 * (Similarly for HVC and SMC elsewhere.)
105 	 */
106 
107 	local_daif_restore(DAIF_PROCCTX);
108 
109 	if (flags & _TIF_MTE_ASYNC_FAULT) {
110 		/*
111 		 * Process the asynchronous tag check fault before the actual
112 		 * syscall. do_notify_resume() will send a signal to userspace
113 		 * before the syscall is restarted.
114 		 */
115 		syscall_set_return_value(current, regs, -ERESTARTNOINTR, 0);
116 		return;
117 	}
118 
119 	if (has_syscall_work(flags)) {
120 		/*
121 		 * The de-facto standard way to skip a system call using ptrace
122 		 * is to set the system call to -1 (NO_SYSCALL) and set x0 to a
123 		 * suitable error code for consumption by userspace. However,
124 		 * this cannot be distinguished from a user-issued syscall(-1)
125 		 * and so we must set x0 to -ENOSYS here in case the tracer doesn't
126 		 * issue the skip and we fall into trace_exit with x0 preserved.
127 		 *
128 		 * This is slightly odd because it also means that if a tracer
129 		 * sets the system call number to -1 but does not initialise x0,
130 		 * then x0 will be preserved for all system calls apart from a
131 		 * user-issued syscall(-1). However, requesting a skip and not
132 		 * setting the return value is unlikely to do anything sensible
133 		 * anyway.
134 		 */
135 		if (scno == NO_SYSCALL)
136 			syscall_set_return_value(current, regs, -ENOSYS, 0);
137 		scno = syscall_trace_enter(regs);
138 		if (scno == NO_SYSCALL)
139 			goto trace_exit;
140 	}
141 
142 	invoke_syscall(regs, scno, sc_nr, syscall_table);
143 
144 	/*
145 	 * The tracing status may have changed under our feet, so we have to
146 	 * check again. However, if we were tracing entry, then we always trace
147 	 * exit regardless, as the old entry assembly did.
148 	 */
149 	if (!has_syscall_work(flags) && !IS_ENABLED(CONFIG_DEBUG_RSEQ)) {
150 		local_daif_mask();
151 		flags = read_thread_flags();
152 		if (!has_syscall_work(flags) && !(flags & _TIF_SINGLESTEP))
153 			return;
154 		local_daif_restore(DAIF_PROCCTX);
155 	}
156 
157 trace_exit:
158 	syscall_trace_exit(regs);
159 }
160 
161 /*
162  * As per the ABI exit SME streaming mode and clear the SVE state not
163  * shared with FPSIMD on syscall entry.
164  */
165 static inline void fp_user_discard(void)
166 {
167 	/*
168 	 * If SME is active then exit streaming mode.  If ZA is active
169 	 * then flush the SVE registers but leave userspace access to
170 	 * both SVE and SME enabled, otherwise disable SME for the
171 	 * task and fall through to disabling SVE too.  This means
172 	 * that after a syscall we never have any streaming mode
173 	 * register state to track, if this changes the KVM code will
174 	 * need updating.
175 	 */
176 	if (system_supports_sme() && test_thread_flag(TIF_SME)) {
177 		u64 svcr = read_sysreg_s(SYS_SVCR);
178 
179 		if (svcr & SVCR_SM_MASK)
180 			sme_smstop_sm();
181 	}
182 
183 	if (!system_supports_sve())
184 		return;
185 
186 	/*
187 	 * If SME is not active then disable SVE, the registers will
188 	 * be cleared when userspace next attempts to access them and
189 	 * we do not need to track the SVE register state until then.
190 	 */
191 	clear_thread_flag(TIF_SVE);
192 
193 	/*
194 	 * task_fpsimd_load() won't be called to update CPACR_EL1 in
195 	 * ret_to_user unless TIF_FOREIGN_FPSTATE is still set, which only
196 	 * happens if a context switch or kernel_neon_begin() or context
197 	 * modification (sigreturn, ptrace) intervenes.
198 	 * So, ensure that CPACR_EL1 is already correct for the fast-path case.
199 	 */
200 	sve_user_disable();
201 }
202 
203 void do_el0_svc(struct pt_regs *regs)
204 {
205 	fp_user_discard();
206 	el0_svc_common(regs, regs->regs[8], __NR_syscalls, sys_call_table);
207 }
208 
209 #ifdef CONFIG_COMPAT
210 void do_el0_svc_compat(struct pt_regs *regs)
211 {
212 	el0_svc_common(regs, regs->regs[7], __NR_compat_syscalls,
213 		       compat_sys_call_table);
214 }
215 #endif
216