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 DEFINE_PER_CPU(struct sdei_registered_event *, sdei_active_normal_event);
43 DEFINE_PER_CPU(struct sdei_registered_event *, sdei_active_critical_event);
44
45 DECLARE_PER_CPU(unsigned long *, sdei_shadow_call_stack_normal_ptr);
46 DECLARE_PER_CPU(unsigned long *, sdei_shadow_call_stack_critical_ptr);
47
48 #ifdef CONFIG_SHADOW_CALL_STACK
49 DEFINE_PER_CPU(unsigned long *, sdei_shadow_call_stack_normal_ptr);
50 DEFINE_PER_CPU(unsigned long *, sdei_shadow_call_stack_critical_ptr);
51 #endif
52
_free_sdei_stack(unsigned long * __percpu * ptr,int cpu)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
free_sdei_stacks(void)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
_init_sdei_stack(unsigned long * __percpu * ptr,int cpu)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
init_sdei_stacks(void)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
_free_sdei_scs(unsigned long * __percpu * ptr,int cpu)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
free_sdei_scs(void)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
_init_sdei_scs(unsigned long * __percpu * ptr,int cpu)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
init_sdei_scs(void)145 static int init_sdei_scs(void)
146 {
147 int cpu;
148 int err = 0;
149
150 if (!IS_ENABLED(CONFIG_SHADOW_CALL_STACK))
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
on_sdei_normal_stack(unsigned long sp,struct stack_info * info)168 static bool on_sdei_normal_stack(unsigned long sp, struct stack_info *info)
169 {
170 unsigned long low = (unsigned long)raw_cpu_read(sdei_stack_normal_ptr);
171 unsigned long high = low + SDEI_STACK_SIZE;
172
173 return on_stack(sp, low, high, STACK_TYPE_SDEI_NORMAL, info);
174 }
175
on_sdei_critical_stack(unsigned long sp,struct stack_info * info)176 static bool on_sdei_critical_stack(unsigned long sp, struct stack_info *info)
177 {
178 unsigned long low = (unsigned long)raw_cpu_read(sdei_stack_critical_ptr);
179 unsigned long high = low + SDEI_STACK_SIZE;
180
181 return on_stack(sp, low, high, STACK_TYPE_SDEI_CRITICAL, info);
182 }
183
_on_sdei_stack(unsigned long sp,struct stack_info * info)184 bool _on_sdei_stack(unsigned long sp, struct stack_info *info)
185 {
186 if (!IS_ENABLED(CONFIG_VMAP_STACK))
187 return false;
188
189 if (on_sdei_critical_stack(sp, info))
190 return true;
191
192 if (on_sdei_normal_stack(sp, info))
193 return true;
194
195 return false;
196 }
197
sdei_arch_get_entry_point(int conduit)198 unsigned long sdei_arch_get_entry_point(int conduit)
199 {
200 /*
201 * SDEI works between adjacent exception levels. If we booted at EL1 we
202 * assume a hypervisor is marshalling events. If we booted at EL2 and
203 * dropped to EL1 because we don't support VHE, then we can't support
204 * SDEI.
205 */
206 if (is_hyp_mode_available() && !is_kernel_in_hyp_mode()) {
207 pr_err("Not supported on this hardware/boot configuration\n");
208 goto out_err;
209 }
210
211 if (init_sdei_stacks())
212 goto out_err;
213
214 if (init_sdei_scs())
215 goto out_err_free_stacks;
216
217 sdei_exit_mode = (conduit == SMCCC_CONDUIT_HVC) ? SDEI_EXIT_HVC : SDEI_EXIT_SMC;
218
219 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
220 if (arm64_kernel_unmapped_at_el0()) {
221 unsigned long offset;
222
223 offset = (unsigned long)__sdei_asm_entry_trampoline -
224 (unsigned long)__entry_tramp_text_start;
225 return TRAMP_VALIAS + offset;
226 } else
227 #endif /* CONFIG_UNMAP_KERNEL_AT_EL0 */
228 return (unsigned long)__sdei_asm_handler;
229
230 out_err_free_stacks:
231 free_sdei_stacks();
232 out_err:
233 return 0;
234 }
235
236 /*
237 * __sdei_handler() returns one of:
238 * SDEI_EV_HANDLED - success, return to the interrupted context.
239 * SDEI_EV_FAILED - failure, return this error code to firmare.
240 * virtual-address - success, return to this address.
241 */
_sdei_handler(struct pt_regs * regs,struct sdei_registered_event * arg)242 static __kprobes unsigned long _sdei_handler(struct pt_regs *regs,
243 struct sdei_registered_event *arg)
244 {
245 u32 mode;
246 int i, err = 0;
247 int clobbered_registers = 4;
248 u64 elr = read_sysreg(elr_el1);
249 u32 kernel_mode = read_sysreg(CurrentEL) | 1; /* +SPSel */
250 unsigned long vbar = read_sysreg(vbar_el1);
251
252 if (arm64_kernel_unmapped_at_el0())
253 clobbered_registers++;
254
255 /* Retrieve the missing registers values */
256 for (i = 0; i < clobbered_registers; i++) {
257 /* from within the handler, this call always succeeds */
258 sdei_api_event_context(i, ®s->regs[i]);
259 }
260
261 /*
262 * We didn't take an exception to get here, set PAN. UAO will be cleared
263 * by sdei_event_handler()s force_uaccess_begin() call.
264 */
265 __uaccess_enable_hw_pan();
266
267 err = sdei_event_handler(regs, arg);
268 if (err)
269 return SDEI_EV_FAILED;
270
271 if (elr != read_sysreg(elr_el1)) {
272 /*
273 * We took a synchronous exception from the SDEI handler.
274 * This could deadlock, and if you interrupt KVM it will
275 * hyp-panic instead.
276 */
277 pr_warn("unsafe: exception during handler\n");
278 }
279
280 mode = regs->pstate & (PSR_MODE32_BIT | PSR_MODE_MASK);
281
282 /*
283 * If we interrupted the kernel with interrupts masked, we always go
284 * back to wherever we came from.
285 */
286 if (mode == kernel_mode && !interrupts_enabled(regs))
287 return SDEI_EV_HANDLED;
288
289 /*
290 * Otherwise, we pretend this was an IRQ. This lets user space tasks
291 * receive signals before we return to them, and KVM to invoke it's
292 * world switch to do the same.
293 *
294 * See DDI0487B.a Table D1-7 'Vector offsets from vector table base
295 * address'.
296 */
297 if (mode == kernel_mode)
298 return vbar + 0x280;
299 else if (mode & PSR_MODE32_BIT)
300 return vbar + 0x680;
301
302 return vbar + 0x480;
303 }
304
305
306 asmlinkage noinstr unsigned long
__sdei_handler(struct pt_regs * regs,struct sdei_registered_event * arg)307 __sdei_handler(struct pt_regs *regs, struct sdei_registered_event *arg)
308 {
309 unsigned long ret;
310
311 arm64_enter_nmi(regs);
312
313 ret = _sdei_handler(regs, arg);
314
315 arm64_exit_nmi(regs);
316
317 return ret;
318 }
319