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/syscalls.h>
   9
  10 #include <asm/daifflags.h>
  11 #include <asm/debug-monitors.h>
  12 #include <asm/fpsimd.h>
  13 #include <asm/syscall.h>
  14 #include <asm/thread_info.h>
  15 #include <asm/unistd.h>
  16
  17 long compat_arm_syscall(struct pt_regs *regs, int scno);
  18 long sys_ni_syscall(void);
  19
  20 static long do_ni_syscall(struct pt_regs *regs, int scno)
  21 {
  22 #ifdef CONFIG_COMPAT
  23         long ret;
  24         if (is_compat_task()) {
  25                 ret = compat_arm_syscall(regs, scno);
  26                 if (ret != -ENOSYS)
  27                         return ret;
  28         }
  29 #endif
  30
  31         return sys_ni_syscall();
  32 }
  33
  34 static long __invoke_syscall(struct pt_regs *regs, syscall_fn_t syscall_fn)
  35 {
  36         return syscall_fn(regs);
  37 }
  38
  39 static void invoke_syscall(struct pt_regs *regs, unsigned int scno,
  40                            unsigned int sc_nr,
  41                            const syscall_fn_t syscall_table[])
  42 {
  43         long ret;
  44
  45         if (scno < sc_nr) {
  46                 syscall_fn_t syscall_fn;
  47                 syscall_fn = syscall_table[array_index_nospec(scno, sc_nr)];
  48                 ret = __invoke_syscall(regs, syscall_fn);
  49         } else {
  50                 ret = do_ni_syscall(regs, scno);
  51         }
  52
  53         if (is_compat_task())
  54                 ret = lower_32_bits(ret);
  55
  56         regs->regs[0] = ret;
  57 }
  58
  59 static inline bool has_syscall_work(unsigned long flags)
  60 {
  61         return unlikely(flags & _TIF_SYSCALL_WORK);
  62 }
  63
  64 int syscall_trace_enter(struct pt_regs *regs);
  65 void syscall_trace_exit(struct pt_regs *regs);
  66
  67 #ifdef CONFIG_ARM64_ERRATUM_1463225
  68 DECLARE_PER_CPU(int, __in_cortex_a76_erratum_1463225_wa);
  69
  70 static void cortex_a76_erratum_1463225_svc_handler(void)
  71 {
  72         u32 reg, val;
  73
  74         if (!unlikely(test_thread_flag(TIF_SINGLESTEP)))
  75                 return;
  76
  77         if (!unlikely(this_cpu_has_cap(ARM64_WORKAROUND_1463225)))
  78                 return;
  79
  80         __this_cpu_write(__in_cortex_a76_erratum_1463225_wa, 1);
  81         reg = read_sysreg(mdscr_el1);
  82         val = reg | DBG_MDSCR_SS | DBG_MDSCR_KDE;
  83         write_sysreg(val, mdscr_el1);
  84         asm volatile("msr daifclr, #8");
  85         isb();
  86
  87         /* We will have taken a single-step exception by this point */
  88
  89         write_sysreg(reg, mdscr_el1);
  90         __this_cpu_write(__in_cortex_a76_erratum_1463225_wa, 0);
  91 }
  92 #else
  93 static void cortex_a76_erratum_1463225_svc_handler(void) { }
  94 #endif /* CONFIG_ARM64_ERRATUM_1463225 */
  95
  96 static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr,
  97                            const syscall_fn_t syscall_table[])
  98 {
  99         unsigned long flags = current_thread_info()->flags;
 100
 101         regs->orig_x0 = regs->regs[0];
 102         regs->syscallno = scno;
 103
 104         /*
 105          * BTI note:
 106          * The architecture does not guarantee that SPSR.BTYPE is zero
 107          * on taking an SVC, so we could return to userspace with a
 108          * non-zero BTYPE after the syscall.
 109          *
 110          * This shouldn't matter except when userspace is explicitly
 111          * doing something stupid, such as setting PROT_BTI on a page
 112          * that lacks conforming BTI/PACIxSP instructions, falling
 113          * through from one executable page to another with differing
 114          * PROT_BTI, or messing with BTYPE via ptrace: in such cases,
 115          * userspace should not be surprised if a SIGILL occurs on
 116          * syscall return.
 117          *
 118          * So, don't touch regs->pstate & PSR_BTYPE_MASK here.
 119          * (Similarly for HVC and SMC elsewhere.)
 120          */
 121
 122         cortex_a76_erratum_1463225_svc_handler();
 123         local_daif_restore(DAIF_PROCCTX);
 124
 125         if (flags & _TIF_MTE_ASYNC_FAULT) {
 126                 /*
 127                  * Process the asynchronous tag check fault before the actual
 128                  * syscall. do_notify_resume() will send a signal to userspace
 129                  * before the syscall is restarted.
 130                  */
 131                 regs->regs[0] = -ERESTARTNOINTR;
 132                 return;
 133         }
 134
 135         if (has_syscall_work(flags)) {
 136                 /*
 137                  * The de-facto standard way to skip a system call using ptrace
 138                  * is to set the system call to -1 (NO_SYSCALL) and set x0 to a
 139                  * suitable error code for consumption by userspace. However,
 140                  * this cannot be distinguished from a user-issued syscall(-1)
 141                  * and so we must set x0 to -ENOSYS here in case the tracer doesn't
 142                  * issue the skip and we fall into trace_exit with x0 preserved.
 143                  *
 144                  * This is slightly odd because it also means that if a tracer
 145                  * sets the system call number to -1 but does not initialise x0,
 146                  * then x0 will be preserved for all system calls apart from a
 147                  * user-issued syscall(-1). However, requesting a skip and not
 148                  * setting the return value is unlikely to do anything sensible
 149                  * anyway.
 150                  */
 151                 if (scno == NO_SYSCALL)
 152                         regs->regs[0] = -ENOSYS;
 153                 scno = syscall_trace_enter(regs);
 154                 if (scno == NO_SYSCALL)
 155                         goto trace_exit;
 156         }
 157
 158         invoke_syscall(regs, scno, sc_nr, syscall_table);
 159
 160         /*
 161          * The tracing status may have changed under our feet, so we have to
 162          * check again. However, if we were tracing entry, then we always trace
 163          * exit regardless, as the old entry assembly did.
 164          */
 165         if (!has_syscall_work(flags) && !IS_ENABLED(CONFIG_DEBUG_RSEQ)) {
 166                 local_daif_mask();
 167                 flags = current_thread_info()->flags;
 168                 if (!has_syscall_work(flags) && !(flags & _TIF_SINGLESTEP)) {
 169                         /*
 170                          * We're off to userspace, where interrupts are
 171                          * always enabled after we restore the flags from
 172                          * the SPSR.
 173                          */
 174                         trace_hardirqs_on();
 175                         return;
 176                 }
 177                 local_daif_restore(DAIF_PROCCTX);
 178         }
 179
 180 trace_exit:
 181         syscall_trace_exit(regs);
 182 }
 183
 184 static inline void sve_user_discard(void)
 185 {
 186         if (!system_supports_sve())
 187                 return;
 188
 189         clear_thread_flag(TIF_SVE);
 190
 191         /*
 192          * task_fpsimd_load() won't be called to update CPACR_EL1 in
 193          * ret_to_user unless TIF_FOREIGN_FPSTATE is still set, which only
 194          * happens if a context switch or kernel_neon_begin() or context
 195          * modification (sigreturn, ptrace) intervenes.
 196          * So, ensure that CPACR_EL1 is already correct for the fast-path case.
 197          */
 198         sve_user_disable();
 199 }
 200
 201 void do_el0_svc(struct pt_regs *regs)
 202 {
 203         sve_user_discard();
 204         el0_svc_common(regs, regs->regs[8], __NR_syscalls, sys_call_table);
 205 }
 206
 207 #ifdef CONFIG_COMPAT
 208 void do_el0_svc_compat(struct pt_regs *regs)
 209 {
 210         el0_svc_common(regs, regs->regs[7], __NR_compat_syscalls,
 211                        compat_sys_call_table);
 212 }
 213 #endif