개념 설명 전체 · v6.18.37 / arch/arm64/kernel/syscall.c

    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 #include <asm/unistd_compat_32.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 	if (is_compat_task()) {
   25 		long ret = compat_arm_syscall(regs, scno);
   26 		if (ret != -ENOSYS)
   27 			return ret;
   28 	}
   29 
   30 	return sys_ni_syscall();
   31 }
   32 
   33 static long __invoke_syscall(struct pt_regs *regs, syscall_fn_t syscall_fn)
   34 {
   35 	return syscall_fn(regs);
   36 }
   37 
   38 static void invoke_syscall(struct pt_regs *regs, unsigned int scno,
   39 			   unsigned int sc_nr,
   40 			   const syscall_fn_t syscall_table[])
   41 {
   42 	long ret;
   43 
   44 	add_random_kstack_offset();
   45 
   46 	if (likely(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 	syscall_set_return_value(current, regs, 0, ret);
   55 
   56 	/*
   57 	 * This value will get limited by KSTACK_OFFSET_MAX(), which is 10
   58 	 * bits. The actual entropy will be further reduced by the compiler
   59 	 * when applying stack alignment constraints: the AAPCS mandates a
   60 	 * 16-byte aligned SP at function boundaries, which will remove the
   61 	 * 4 low bits from any entropy chosen here.
   62 	 *
   63 	 * The resulting 6 bits of entropy is seen in SP[9:4].
   64 	 */
   65 	choose_random_kstack_offset(get_random_u16());
   66 }
   67 
   68 static inline bool has_syscall_work(unsigned long flags)
   69 {
   70 	return unlikely(flags & _TIF_SYSCALL_WORK);
   71 }
   72 
   73 static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr,
   74 			   const syscall_fn_t syscall_table[])
   75 {
   76 	unsigned long flags = read_thread_flags();
   77 
   78 	regs->orig_x0 = regs->regs[0];
   79 	regs->syscallno = scno;
   80 
   81 	/*
   82 	 * BTI note:
   83 	 * The architecture does not guarantee that SPSR.BTYPE is zero
   84 	 * on taking an SVC, so we could return to userspace with a
   85 	 * non-zero BTYPE after the syscall.
   86 	 *
   87 	 * This shouldn't matter except when userspace is explicitly
   88 	 * doing something stupid, such as setting PROT_BTI on a page
   89 	 * that lacks conforming BTI/PACIxSP instructions, falling
   90 	 * through from one executable page to another with differing
   91 	 * PROT_BTI, or messing with BTYPE via ptrace: in such cases,
   92 	 * userspace should not be surprised if a SIGILL occurs on
   93 	 * syscall return.
   94 	 *
   95 	 * So, don't touch regs->pstate & PSR_BTYPE_MASK here.
   96 	 * (Similarly for HVC and SMC elsewhere.)
   97 	 */
   98 
   99 	if (flags & _TIF_MTE_ASYNC_FAULT) {
  100 		/*
  101 		 * Process the asynchronous tag check fault before the actual
  102 		 * syscall. do_notify_resume() will send a signal to userspace
  103 		 * before the syscall is restarted.
  104 		 */
  105 		syscall_set_return_value(current, regs, -ERESTARTNOINTR, 0);
  106 		return;
  107 	}
  108 
  109 	if (has_syscall_work(flags)) {
  110 		/*
  111 		 * The de-facto standard way to skip a system call using ptrace
  112 		 * is to set the system call to -1 (NO_SYSCALL) and set x0 to a
  113 		 * suitable error code for consumption by userspace. However,
  114 		 * this cannot be distinguished from a user-issued syscall(-1)
  115 		 * and so we must set x0 to -ENOSYS here in case the tracer doesn't
  116 		 * issue the skip and we fall into trace_exit with x0 preserved.
  117 		 *
  118 		 * This is slightly odd because it also means that if a tracer
  119 		 * sets the system call number to -1 but does not initialise x0,
  120 		 * then x0 will be preserved for all system calls apart from a
  121 		 * user-issued syscall(-1). However, requesting a skip and not
  122 		 * setting the return value is unlikely to do anything sensible
  123 		 * anyway.
  124 		 */
  125 		if (scno == NO_SYSCALL)
  126 			syscall_set_return_value(current, regs, -ENOSYS, 0);
  127 		scno = syscall_trace_enter(regs);
  128 		if (scno == NO_SYSCALL)
  129 			goto trace_exit;
  130 	}
  131 
  132 	invoke_syscall(regs, scno, sc_nr, syscall_table);
  133 
  134 	/*
  135 	 * The tracing status may have changed under our feet, so we have to
  136 	 * check again. However, if we were tracing entry, then we always trace
  137 	 * exit regardless, as the old entry assembly did.
  138 	 */
  139 	if (!has_syscall_work(flags) && !IS_ENABLED(CONFIG_DEBUG_RSEQ)) {
  140 		flags = read_thread_flags();
  141 		if (!has_syscall_work(flags) && !(flags & _TIF_SINGLESTEP))
  142 			return;
  143 	}
  144 
  145 trace_exit:
  146 	syscall_trace_exit(regs);
  147 }
  148 
  149 void do_el0_svc(struct pt_regs *regs)
  150 {
  151 	el0_svc_common(regs, regs->regs[8], __NR_syscalls, sys_call_table);
  152 }
  153 
  154 #ifdef CONFIG_COMPAT
  155 void do_el0_svc_compat(struct pt_regs *regs)
  156 {
  157 	el0_svc_common(regs, regs->regs[7], __NR_compat32_syscalls,
  158 		       compat_sys_call_table);
  159 }
  160 #endif