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