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 static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr, 79 const syscall_fn_t syscall_table[]) 80 { 81 unsigned long flags = read_thread_flags(); 82 83 regs->orig_x0 = regs->regs[0]; 84 regs->syscallno = scno; 85 86 /* 87 * BTI note: 88 * The architecture does not guarantee that SPSR.BTYPE is zero 89 * on taking an SVC, so we could return to userspace with a 90 * non-zero BTYPE after the syscall. 91 * 92 * This shouldn't matter except when userspace is explicitly 93 * doing something stupid, such as setting PROT_BTI on a page 94 * that lacks conforming BTI/PACIxSP instructions, falling 95 * through from one executable page to another with differing 96 * PROT_BTI, or messing with BTYPE via ptrace: in such cases, 97 * userspace should not be surprised if a SIGILL occurs on 98 * syscall return. 99 * 100 * So, don't touch regs->pstate & PSR_BTYPE_MASK here. 101 * (Similarly for HVC and SMC elsewhere.) 102 */ 103 104 local_daif_restore(DAIF_PROCCTX); 105 106 if (flags & _TIF_MTE_ASYNC_FAULT) { 107 /* 108 * Process the asynchronous tag check fault before the actual 109 * syscall. do_notify_resume() will send a signal to userspace 110 * before the syscall is restarted. 111 */ 112 syscall_set_return_value(current, regs, -ERESTARTNOINTR, 0); 113 return; 114 } 115 116 if (has_syscall_work(flags)) { 117 /* 118 * The de-facto standard way to skip a system call using ptrace 119 * is to set the system call to -1 (NO_SYSCALL) and set x0 to a 120 * suitable error code for consumption by userspace. However, 121 * this cannot be distinguished from a user-issued syscall(-1) 122 * and so we must set x0 to -ENOSYS here in case the tracer doesn't 123 * issue the skip and we fall into trace_exit with x0 preserved. 124 * 125 * This is slightly odd because it also means that if a tracer 126 * sets the system call number to -1 but does not initialise x0, 127 * then x0 will be preserved for all system calls apart from a 128 * user-issued syscall(-1). However, requesting a skip and not 129 * setting the return value is unlikely to do anything sensible 130 * anyway. 131 */ 132 if (scno == NO_SYSCALL) 133 syscall_set_return_value(current, regs, -ENOSYS, 0); 134 scno = syscall_trace_enter(regs); 135 if (scno == NO_SYSCALL) 136 goto trace_exit; 137 } 138 139 invoke_syscall(regs, scno, sc_nr, syscall_table); 140 141 /* 142 * The tracing status may have changed under our feet, so we have to 143 * check again. However, if we were tracing entry, then we always trace 144 * exit regardless, as the old entry assembly did. 145 */ 146 if (!has_syscall_work(flags) && !IS_ENABLED(CONFIG_DEBUG_RSEQ)) { 147 flags = read_thread_flags(); 148 if (!has_syscall_work(flags) && !(flags & _TIF_SINGLESTEP)) 149 return; 150 } 151 152 trace_exit: 153 syscall_trace_exit(regs); 154 } 155 156 /* 157 * As per the ABI exit SME streaming mode and clear the SVE state not 158 * shared with FPSIMD on syscall entry. 159 */ 160 static inline void fp_user_discard(void) 161 { 162 /* 163 * If SME is active then exit streaming mode. If ZA is active 164 * then flush the SVE registers but leave userspace access to 165 * both SVE and SME enabled, otherwise disable SME for the 166 * task and fall through to disabling SVE too. This means 167 * that after a syscall we never have any streaming mode 168 * register state to track, if this changes the KVM code will 169 * need updating. 170 */ 171 if (system_supports_sme()) 172 sme_smstop_sm(); 173 174 if (!system_supports_sve()) 175 return; 176 177 if (test_thread_flag(TIF_SVE)) { 178 unsigned int sve_vq_minus_one; 179 180 sve_vq_minus_one = sve_vq_from_vl(task_get_sve_vl(current)) - 1; 181 sve_flush_live(true, sve_vq_minus_one); 182 } 183 } 184 185 void do_el0_svc(struct pt_regs *regs) 186 { 187 fp_user_discard(); 188 el0_svc_common(regs, regs->regs[8], __NR_syscalls, sys_call_table); 189 } 190 191 #ifdef CONFIG_COMPAT 192 void do_el0_svc_compat(struct pt_regs *regs) 193 { 194 el0_svc_common(regs, regs->regs[7], __NR_compat_syscalls, 195 compat_sys_call_table); 196 } 197 #endif 198