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