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-smccc.h>
   6 #include <linux/arm_sdei.h>
   7 #include <linux/hardirq.h>
   8 #include <linux/irqflags.h>
   9 #include <linux/sched/task_stack.h>
  10 #include <linux/scs.h>
  11 #include <linux/uaccess.h>
  12
  13 #include <asm/alternative.h>
  14 #include <asm/exception.h>
  15 #include <asm/kprobes.h>
  16 #include <asm/mmu.h>
  17 #include <asm/ptrace.h>
  18 #include <asm/sections.h>
  19 #include <asm/stacktrace.h>
  20 #include <asm/sysreg.h>
  21 #include <asm/vmap_stack.h>
  22
  23 unsigned long sdei_exit_mode;
  24
  25 /*
  26  * VMAP'd stacks checking for stack overflow on exception using sp as a scratch
  27  * register, meaning SDEI has to switch to its own stack. We need two stacks as
  28  * a critical event may interrupt a normal event that has just taken a
  29  * synchronous exception, and is using sp as scratch register. For a critical
  30  * event interrupting a normal event, we can't reliably tell if we were on the
  31  * sdei stack.
  32  * For now, we allocate stacks when the driver is probed.
  33  */
  34 DECLARE_PER_CPU(unsigned long *, sdei_stack_normal_ptr);
  35 DECLARE_PER_CPU(unsigned long *, sdei_stack_critical_ptr);
  36
  37 #ifdef CONFIG_VMAP_STACK
  38 DEFINE_PER_CPU(unsigned long *, sdei_stack_normal_ptr);
  39 DEFINE_PER_CPU(unsigned long *, sdei_stack_critical_ptr);
  40 #endif
  41
  42 DECLARE_PER_CPU(unsigned long *, sdei_shadow_call_stack_normal_ptr);
  43 DECLARE_PER_CPU(unsigned long *, sdei_shadow_call_stack_critical_ptr);
  44
  45 #ifdef CONFIG_SHADOW_CALL_STACK
  46 DEFINE_PER_CPU(unsigned long *, sdei_shadow_call_stack_normal_ptr);
  47 DEFINE_PER_CPU(unsigned long *, sdei_shadow_call_stack_critical_ptr);
  48 #endif
  49
  50 DEFINE_PER_CPU(struct sdei_registered_event *, sdei_active_normal_event);
  51 DEFINE_PER_CPU(struct sdei_registered_event *, sdei_active_critical_event);
  52
  53 static void _free_sdei_stack(unsigned long * __percpu *ptr, int cpu)
  54 {
  55         unsigned long *p;
  56
  57         p = per_cpu(*ptr, cpu);
  58         if (p) {
  59                 per_cpu(*ptr, cpu) = NULL;
  60                 vfree(p);
  61         }
  62 }
  63
  64 static void free_sdei_stacks(void)
  65 {
  66         int cpu;
  67
  68         if (!IS_ENABLED(CONFIG_VMAP_STACK))
  69                 return;
  70
  71         for_each_possible_cpu(cpu) {
  72                 _free_sdei_stack(&sdei_stack_normal_ptr, cpu);
  73                 _free_sdei_stack(&sdei_stack_critical_ptr, cpu);
  74         }
  75 }
  76
  77 static int _init_sdei_stack(unsigned long * __percpu *ptr, int cpu)
  78 {
  79         unsigned long *p;
  80
  81         p = arch_alloc_vmap_stack(SDEI_STACK_SIZE, cpu_to_node(cpu));
  82         if (!p)
  83                 return -ENOMEM;
  84         per_cpu(*ptr, cpu) = p;
  85
  86         return 0;
  87 }
  88
  89 static int init_sdei_stacks(void)
  90 {
  91         int cpu;
  92         int err = 0;
  93
  94         if (!IS_ENABLED(CONFIG_VMAP_STACK))
  95                 return 0;
  96
  97         for_each_possible_cpu(cpu) {
  98                 err = _init_sdei_stack(&sdei_stack_normal_ptr, cpu);
  99                 if (err)
 100                         break;
 101                 err = _init_sdei_stack(&sdei_stack_critical_ptr, cpu);
 102                 if (err)
 103                         break;
 104         }
 105
 106         if (err)
 107                 free_sdei_stacks();
 108
 109         return err;
 110 }
 111
 112 static void _free_sdei_scs(unsigned long * __percpu *ptr, int cpu)
 113 {
 114         void *s;
 115
 116         s = per_cpu(*ptr, cpu);
 117         if (s) {
 118                 per_cpu(*ptr, cpu) = NULL;
 119                 scs_free(s);
 120         }
 121 }
 122
 123 static void free_sdei_scs(void)
 124 {
 125         int cpu;
 126
 127         for_each_possible_cpu(cpu) {
 128                 _free_sdei_scs(&sdei_shadow_call_stack_normal_ptr, cpu);
 129                 _free_sdei_scs(&sdei_shadow_call_stack_critical_ptr, cpu);
 130         }
 131 }
 132
 133 static int _init_sdei_scs(unsigned long * __percpu *ptr, int cpu)
 134 {
 135         void *s;
 136
 137         s = scs_alloc(cpu_to_node(cpu));
 138         if (!s)
 139                 return -ENOMEM;
 140         per_cpu(*ptr, cpu) = s;
 141
 142         return 0;
 143 }
 144
 145 static int init_sdei_scs(void)
 146 {
 147         int cpu;
 148         int err = 0;
 149
 150         if (!scs_is_enabled())
 151                 return 0;
 152
 153         for_each_possible_cpu(cpu) {
 154                 err = _init_sdei_scs(&sdei_shadow_call_stack_normal_ptr, cpu);
 155                 if (err)
 156                         break;
 157                 err = _init_sdei_scs(&sdei_shadow_call_stack_critical_ptr, cpu);
 158                 if (err)
 159                         break;
 160         }
 161
 162         if (err)
 163                 free_sdei_scs();
 164
 165         return err;
 166 }
 167
 168 unsigned long sdei_arch_get_entry_point(int conduit)
 169 {
 170         /*
 171          * SDEI works between adjacent exception levels. If we booted at EL1 we
 172          * assume a hypervisor is marshalling events. If we booted at EL2 and
 173          * dropped to EL1 because we don't support VHE, then we can't support
 174          * SDEI.
 175          */
 176         if (is_hyp_nvhe()) {
 177                 pr_err("Not supported on this hardware/boot configuration\n");
 178                 goto out_err;
 179         }
 180
 181         if (init_sdei_stacks())
 182                 goto out_err;
 183
 184         if (init_sdei_scs())
 185                 goto out_err_free_stacks;
 186
 187         sdei_exit_mode = (conduit == SMCCC_CONDUIT_HVC) ? SDEI_EXIT_HVC : SDEI_EXIT_SMC;
 188
 189 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
 190         if (arm64_kernel_unmapped_at_el0()) {
 191                 unsigned long offset;
 192
 193                 offset = (unsigned long)__sdei_asm_entry_trampoline -
 194                          (unsigned long)__entry_tramp_text_start;
 195                 return TRAMP_VALIAS + offset;
 196         } else
 197 #endif /* CONFIG_UNMAP_KERNEL_AT_EL0 */
 198                 return (unsigned long)__sdei_asm_handler;
 199
 200 out_err_free_stacks:
 201         free_sdei_stacks();
 202 out_err:
 203         return 0;
 204 }
 205
 206 /*
 207  * do_sdei_event() returns one of:
 208  *  SDEI_EV_HANDLED -  success, return to the interrupted context.
 209  *  SDEI_EV_FAILED  -  failure, return this error code to firmare.
 210  *  virtual-address -  success, return to this address.
 211  */
 212 unsigned long __kprobes do_sdei_event(struct pt_regs *regs,
 213                                       struct sdei_registered_event *arg)
 214 {
 215         u32 mode;
 216         int i, err = 0;
 217         int clobbered_registers = 4;
 218         u64 elr = read_sysreg(elr_el1);
 219         u32 kernel_mode = read_sysreg(CurrentEL) | 1;   /* +SPSel */
 220         unsigned long vbar = read_sysreg(vbar_el1);
 221
 222         if (arm64_kernel_unmapped_at_el0())
 223                 clobbered_registers++;
 224
 225         /* Retrieve the missing registers values */
 226         for (i = 0; i < clobbered_registers; i++) {
 227                 /* from within the handler, this call always succeeds */
 228                 sdei_api_event_context(i, &regs->regs[i]);
 229         }
 230
 231         err = sdei_event_handler(regs, arg);
 232         if (err)
 233                 return SDEI_EV_FAILED;
 234
 235         if (elr != read_sysreg(elr_el1)) {
 236                 /*
 237                  * We took a synchronous exception from the SDEI handler.
 238                  * This could deadlock, and if you interrupt KVM it will
 239                  * hyp-panic instead.
 240                  */
 241                 pr_warn("unsafe: exception during handler\n");
 242         }
 243
 244         mode = regs->pstate & (PSR_MODE32_BIT | PSR_MODE_MASK);
 245
 246         /*
 247          * If we interrupted the kernel with interrupts masked, we always go
 248          * back to wherever we came from.
 249          */
 250         if (mode == kernel_mode && !interrupts_enabled(regs))
 251                 return SDEI_EV_HANDLED;
 252
 253         /*
 254          * Otherwise, we pretend this was an IRQ. This lets user space tasks
 255          * receive signals before we return to them, and KVM to invoke it's
 256          * world switch to do the same.
 257          *
 258          * See DDI0487B.a Table D1-7 'Vector offsets from vector table base
 259          * address'.
 260          */
 261         if (mode == kernel_mode)
 262                 return vbar + 0x280;
 263         else if (mode & PSR_MODE32_BIT)
 264                 return vbar + 0x680;
 265
 266         return vbar + 0x480;
 267 }