1 // SPDX-License-Identifier: GPL-2.0
2 // Copyright (C) 2017 Arm Ltd.
3 #define pr_fmt(fmt) "sdei: " fmt
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>
13 #include <asm/alternative.h>
14 #include <asm/exception.h>
15 #include <asm/kprobes.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>
23 unsigned long sdei_exit_mode;
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
32 * For now, we allocate stacks when the driver is probed.
34 DECLARE_PER_CPU(unsigned long *, sdei_stack_normal_ptr);
35 DECLARE_PER_CPU(unsigned long *, sdei_stack_critical_ptr);
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);
42 DECLARE_PER_CPU(unsigned long *, sdei_shadow_call_stack_normal_ptr);
43 DECLARE_PER_CPU(unsigned long *, sdei_shadow_call_stack_critical_ptr);
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);
50 DEFINE_PER_CPU(struct sdei_registered_event *, sdei_active_normal_event);
51 DEFINE_PER_CPU(struct sdei_registered_event *, sdei_active_critical_event);
53 static void _free_sdei_stack(unsigned long * __percpu *ptr, int cpu)
57 p = per_cpu(*ptr, cpu);
59 per_cpu(*ptr, cpu) = NULL;
64 static void free_sdei_stacks(void)
68 if (!IS_ENABLED(CONFIG_VMAP_STACK))
71 for_each_possible_cpu(cpu) {
72 _free_sdei_stack(&sdei_stack_normal_ptr, cpu);
73 _free_sdei_stack(&sdei_stack_critical_ptr, cpu);
77 static int _init_sdei_stack(unsigned long * __percpu *ptr, int cpu)
81 p = arch_alloc_vmap_stack(SDEI_STACK_SIZE, cpu_to_node(cpu));
84 per_cpu(*ptr, cpu) = p;
89 static int init_sdei_stacks(void)
94 if (!IS_ENABLED(CONFIG_VMAP_STACK))
97 for_each_possible_cpu(cpu) {
98 err = _init_sdei_stack(&sdei_stack_normal_ptr, cpu);
101 err = _init_sdei_stack(&sdei_stack_critical_ptr, cpu);
112 static void _free_sdei_scs(unsigned long * __percpu *ptr, int cpu)
116 s = per_cpu(*ptr, cpu);
118 per_cpu(*ptr, cpu) = NULL;
123 static void free_sdei_scs(void)
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);
133 static int _init_sdei_scs(unsigned long * __percpu *ptr, int cpu)
137 s = scs_alloc(cpu_to_node(cpu));
140 per_cpu(*ptr, cpu) = s;
145 static int init_sdei_scs(void)
150 if (!IS_ENABLED(CONFIG_SHADOW_CALL_STACK))
153 for_each_possible_cpu(cpu) {
154 err = _init_sdei_scs(&sdei_shadow_call_stack_normal_ptr, cpu);
157 err = _init_sdei_scs(&sdei_shadow_call_stack_critical_ptr, cpu);
168 static bool on_sdei_normal_stack(unsigned long sp, unsigned long size,
169 struct stack_info *info)
171 unsigned long low = (unsigned long)raw_cpu_read(sdei_stack_normal_ptr);
172 unsigned long high = low + SDEI_STACK_SIZE;
174 return on_stack(sp, size, low, high, STACK_TYPE_SDEI_NORMAL, info);
177 static bool on_sdei_critical_stack(unsigned long sp, unsigned long size,
178 struct stack_info *info)
180 unsigned long low = (unsigned long)raw_cpu_read(sdei_stack_critical_ptr);
181 unsigned long high = low + SDEI_STACK_SIZE;
183 return on_stack(sp, size, low, high, STACK_TYPE_SDEI_CRITICAL, info);
186 bool _on_sdei_stack(unsigned long sp, unsigned long size, struct stack_info *info)
188 if (!IS_ENABLED(CONFIG_VMAP_STACK))
191 if (on_sdei_critical_stack(sp, size, info))
194 if (on_sdei_normal_stack(sp, size, info))
200 unsigned long sdei_arch_get_entry_point(int conduit)
203 * SDEI works between adjacent exception levels. If we booted at EL1 we
204 * assume a hypervisor is marshalling events. If we booted at EL2 and
205 * dropped to EL1 because we don't support VHE, then we can't support
208 if (is_hyp_mode_available() && !is_kernel_in_hyp_mode()) {
209 pr_err("Not supported on this hardware/boot configuration\n");
213 if (init_sdei_stacks())
217 goto out_err_free_stacks;
219 sdei_exit_mode = (conduit == SMCCC_CONDUIT_HVC) ? SDEI_EXIT_HVC : SDEI_EXIT_SMC;
221 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
222 if (arm64_kernel_unmapped_at_el0()) {
223 unsigned long offset;
225 offset = (unsigned long)__sdei_asm_entry_trampoline -
226 (unsigned long)__entry_tramp_text_start;
227 return TRAMP_VALIAS + offset;
229 #endif /* CONFIG_UNMAP_KERNEL_AT_EL0 */
230 return (unsigned long)__sdei_asm_handler;
239 * do_sdei_event() returns one of:
240 * SDEI_EV_HANDLED - success, return to the interrupted context.
241 * SDEI_EV_FAILED - failure, return this error code to firmare.
242 * virtual-address - success, return to this address.
244 unsigned long __kprobes do_sdei_event(struct pt_regs *regs,
245 struct sdei_registered_event *arg)
249 int clobbered_registers = 4;
250 u64 elr = read_sysreg(elr_el1);
251 u32 kernel_mode = read_sysreg(CurrentEL) | 1; /* +SPSel */
252 unsigned long vbar = read_sysreg(vbar_el1);
254 if (arm64_kernel_unmapped_at_el0())
255 clobbered_registers++;
257 /* Retrieve the missing registers values */
258 for (i = 0; i < clobbered_registers; i++) {
259 /* from within the handler, this call always succeeds */
260 sdei_api_event_context(i, ®s->regs[i]);
263 err = sdei_event_handler(regs, arg);
265 return SDEI_EV_FAILED;
267 if (elr != read_sysreg(elr_el1)) {
269 * We took a synchronous exception from the SDEI handler.
270 * This could deadlock, and if you interrupt KVM it will
273 pr_warn("unsafe: exception during handler\n");
276 mode = regs->pstate & (PSR_MODE32_BIT | PSR_MODE_MASK);
279 * If we interrupted the kernel with interrupts masked, we always go
280 * back to wherever we came from.
282 if (mode == kernel_mode && !interrupts_enabled(regs))
283 return SDEI_EV_HANDLED;
286 * Otherwise, we pretend this was an IRQ. This lets user space tasks
287 * receive signals before we return to them, and KVM to invoke it's
288 * world switch to do the same.
290 * See DDI0487B.a Table D1-7 'Vector offsets from vector table base
293 if (mode == kernel_mode)
295 else if (mode & PSR_MODE32_BIT)