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