arch/arm64/kernel/sdei.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 // Copyright (C) 2017 Arm Ltd.
   3 #define pr_fmt(fmt) "sdei: " fmt
   4
   5 #include <linux/arm_sdei.h>
   6 #include <linux/hardirq.h>
   7 #include <linux/irqflags.h>
   8 #include <linux/sched/task_stack.h>
   9 #include <linux/uaccess.h>
  10
  11 #include <asm/alternative.h>
  12 #include <asm/kprobes.h>
  13 #include <asm/mmu.h>
  14 #include <asm/ptrace.h>
  15 #include <asm/sections.h>
  16 #include <asm/sysreg.h>
  17 #include <asm/vmap_stack.h>
  18
  19 unsigned long sdei_exit_mode;
  20
  21 /*
  22  * VMAP'd stacks checking for stack overflow on exception using sp as a scratch
  23  * register, meaning SDEI has to switch to its own stack. We need two stacks as
  24  * a critical event may interrupt a normal event that has just taken a
  25  * synchronous exception, and is using sp as scratch register. For a critical
  26  * event interrupting a normal event, we can't reliably tell if we were on the
  27  * sdei stack.
  28  * For now, we allocate stacks when the driver is probed.
  29  */
  30 DECLARE_PER_CPU(unsigned long *, sdei_stack_normal_ptr);
  31 DECLARE_PER_CPU(unsigned long *, sdei_stack_critical_ptr);
  32
  33 #ifdef CONFIG_VMAP_STACK
  34 DEFINE_PER_CPU(unsigned long *, sdei_stack_normal_ptr);
  35 DEFINE_PER_CPU(unsigned long *, sdei_stack_critical_ptr);
  36 #endif
  37
  38 static void _free_sdei_stack(unsigned long * __percpu *ptr, int cpu)
  39 {
  40         unsigned long *p;
  41
  42         p = per_cpu(*ptr, cpu);
  43         if (p) {
  44                 per_cpu(*ptr, cpu) = NULL;
  45                 vfree(p);
  46         }
  47 }
  48
  49 static void free_sdei_stacks(void)
  50 {
  51         int cpu;
  52
  53         for_each_possible_cpu(cpu) {
  54                 _free_sdei_stack(&sdei_stack_normal_ptr, cpu);
  55                 _free_sdei_stack(&sdei_stack_critical_ptr, cpu);
  56         }
  57 }
  58
  59 static int _init_sdei_stack(unsigned long * __percpu *ptr, int cpu)
  60 {
  61         unsigned long *p;
  62
  63         p = arch_alloc_vmap_stack(SDEI_STACK_SIZE, cpu_to_node(cpu));
  64         if (!p)
  65                 return -ENOMEM;
  66         per_cpu(*ptr, cpu) = p;
  67
  68         return 0;
  69 }
  70
  71 static int init_sdei_stacks(void)
  72 {
  73         int cpu;
  74         int err = 0;
  75
  76         for_each_possible_cpu(cpu) {
  77                 err = _init_sdei_stack(&sdei_stack_normal_ptr, cpu);
  78                 if (err)
  79                         break;
  80                 err = _init_sdei_stack(&sdei_stack_critical_ptr, cpu);
  81                 if (err)
  82                         break;
  83         }
  84
  85         if (err)
  86                 free_sdei_stacks();
  87
  88         return err;
  89 }
  90
  91 bool _on_sdei_stack(unsigned long sp)
  92 {
  93         unsigned long low, high;
  94
  95         if (!IS_ENABLED(CONFIG_VMAP_STACK))
  96                 return false;
  97
  98         low = (unsigned long)raw_cpu_read(sdei_stack_critical_ptr);
  99         high = low + SDEI_STACK_SIZE;
 100
 101         if (low <= sp && sp < high)
 102                 return true;
 103
 104         low = (unsigned long)raw_cpu_read(sdei_stack_normal_ptr);
 105         high = low + SDEI_STACK_SIZE;
 106
 107         return (low <= sp && sp < high);
 108 }
 109
 110 unsigned long sdei_arch_get_entry_point(int conduit)
 111 {
 112         /*
 113          * SDEI works between adjacent exception levels. If we booted at EL1 we
 114          * assume a hypervisor is marshalling events. If we booted at EL2 and
 115          * dropped to EL1 because we don't support VHE, then we can't support
 116          * SDEI.
 117          */
 118         if (is_hyp_mode_available() && !is_kernel_in_hyp_mode()) {
 119                 pr_err("Not supported on this hardware/boot configuration\n");
 120                 return 0;
 121         }
 122
 123         if (IS_ENABLED(CONFIG_VMAP_STACK)) {
 124                 if (init_sdei_stacks())
 125                         return 0;
 126         }
 127
 128         sdei_exit_mode = (conduit == CONDUIT_HVC) ? SDEI_EXIT_HVC : SDEI_EXIT_SMC;
 129
 130 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
 131         if (arm64_kernel_unmapped_at_el0()) {
 132                 unsigned long offset;
 133
 134                 offset = (unsigned long)__sdei_asm_entry_trampoline -
 135                          (unsigned long)__entry_tramp_text_start;
 136                 return TRAMP_VALIAS + offset;
 137         } else
 138 #endif /* CONFIG_UNMAP_KERNEL_AT_EL0 */
 139                 return (unsigned long)__sdei_asm_handler;
 140
 141 }
 142
 143 /*
 144  * __sdei_handler() returns one of:
 145  *  SDEI_EV_HANDLED -  success, return to the interrupted context.
 146  *  SDEI_EV_FAILED  -  failure, return this error code to firmare.
 147  *  virtual-address -  success, return to this address.
 148  */
 149 static __kprobes unsigned long _sdei_handler(struct pt_regs *regs,
 150                                              struct sdei_registered_event *arg)
 151 {
 152         u32 mode;
 153         int i, err = 0;
 154         int clobbered_registers = 4;
 155         u64 elr = read_sysreg(elr_el1);
 156         u32 kernel_mode = read_sysreg(CurrentEL) | 1;   /* +SPSel */
 157         unsigned long vbar = read_sysreg(vbar_el1);
 158
 159         if (arm64_kernel_unmapped_at_el0())
 160                 clobbered_registers++;
 161
 162         /* Retrieve the missing registers values */
 163         for (i = 0; i < clobbered_registers; i++) {
 164                 /* from within the handler, this call always succeeds */
 165                 sdei_api_event_context(i, &regs->regs[i]);
 166         }
 167
 168         /*
 169          * We didn't take an exception to get here, set PAN. UAO will be cleared
 170          * by sdei_event_handler()s set_fs(USER_DS) call.
 171          */
 172         __uaccess_enable_hw_pan();
 173
 174         err = sdei_event_handler(regs, arg);
 175         if (err)
 176                 return SDEI_EV_FAILED;
 177
 178         if (elr != read_sysreg(elr_el1)) {
 179                 /*
 180                  * We took a synchronous exception from the SDEI handler.
 181                  * This could deadlock, and if you interrupt KVM it will
 182                  * hyp-panic instead.
 183                  */
 184                 pr_warn("unsafe: exception during handler\n");
 185         }
 186
 187         mode = regs->pstate & (PSR_MODE32_BIT | PSR_MODE_MASK);
 188
 189         /*
 190          * If we interrupted the kernel with interrupts masked, we always go
 191          * back to wherever we came from.
 192          */
 193         if (mode == kernel_mode && !interrupts_enabled(regs))
 194                 return SDEI_EV_HANDLED;
 195
 196         /*
 197          * Otherwise, we pretend this was an IRQ. This lets user space tasks
 198          * receive signals before we return to them, and KVM to invoke it's
 199          * world switch to do the same.
 200          *
 201          * See DDI0487B.a Table D1-7 'Vector offsets from vector table base
 202          * address'.
 203          */
 204         if (mode == kernel_mode)
 205                 return vbar + 0x280;
 206         else if (mode & PSR_MODE32_BIT)
 207                 return vbar + 0x680;
 208
 209         return vbar + 0x480;
 210 }
 211
 212
 213 asmlinkage __kprobes notrace unsigned long
 214 __sdei_handler(struct pt_regs *regs, struct sdei_registered_event *arg)
 215 {
 216         unsigned long ret;
 217         bool do_nmi_exit = false;
 218
 219         /*
 220          * nmi_enter() deals with printk() re-entrance and use of RCU when
 221          * RCU believed this CPU was idle. Because critical events can
 222          * interrupt normal events, we may already be in_nmi().
 223          */
 224         if (!in_nmi()) {
 225                 nmi_enter();
 226                 do_nmi_exit = true;
 227         }
 228
 229         ret = _sdei_handler(regs, arg);
 230
 231         if (do_nmi_exit)
 232                 nmi_exit();
 233
 234         return ret;
 235 }