]>
Commit | Line | Data |
---|---|---|
1a59d1b8 | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
8b7d89d0 | 2 | /* |
8b7d89d0 PP |
3 | * |
4 | * Copyright (C) IBM Corporation, 2005 | |
5 | * Jeff Muizelaar, 2006, 2007 | |
6 | * Pekka Paalanen, 2008 <pq@iki.fi> | |
7 | * | |
8 | * Derived from the read-mod example from relay-examples by Tom Zanussi. | |
9 | */ | |
3a0340be | 10 | |
eba11d6d | 11 | #define pr_fmt(fmt) "mmiotrace: " fmt |
3a0340be | 12 | |
d61fc448 PP |
13 | #define DEBUG 1 |
14 | ||
4b599fed | 15 | #include <linux/moduleparam.h> |
8b7d89d0 | 16 | #include <linux/debugfs.h> |
5a0e3ad6 | 17 | #include <linux/slab.h> |
f984b51e | 18 | #include <linux/uaccess.h> |
970e6fa0 | 19 | #include <linux/io.h> |
8b7d89d0 PP |
20 | #include <asm/pgtable.h> |
21 | #include <linux/mmiotrace.h> | |
66441bd3 | 22 | #include <asm/e820/api.h> /* for ISA_START_ADDRESS */ |
60063497 | 23 | #include <linux/atomic.h> |
f5136380 | 24 | #include <linux/percpu.h> |
7423d111 | 25 | #include <linux/cpu.h> |
8b7d89d0 | 26 | |
8b7d89d0 PP |
27 | #include "pf_in.h" |
28 | ||
8b7d89d0 PP |
29 | struct trap_reason { |
30 | unsigned long addr; | |
31 | unsigned long ip; | |
32 | enum reason_type type; | |
33 | int active_traces; | |
34 | }; | |
35 | ||
d61fc448 PP |
36 | struct remap_trace { |
37 | struct list_head list; | |
38 | struct kmmio_probe probe; | |
dee310d0 | 39 | resource_size_t phys; |
d61fc448 PP |
40 | unsigned long id; |
41 | }; | |
42 | ||
fe1ffafa | 43 | /* Accessed per-cpu. */ |
f5136380 | 44 | static DEFINE_PER_CPU(struct trap_reason, pf_reason); |
bd8ac686 | 45 | static DEFINE_PER_CPU(struct mmiotrace_rw, cpu_trace); |
8b7d89d0 | 46 | |
d61fc448 PP |
47 | static DEFINE_MUTEX(mmiotrace_mutex); |
48 | static DEFINE_SPINLOCK(trace_lock); | |
49 | static atomic_t mmiotrace_enabled; | |
50 | static LIST_HEAD(trace_list); /* struct remap_trace */ | |
d61fc448 PP |
51 | |
52 | /* | |
53 | * Locking in this file: | |
54 | * - mmiotrace_mutex enforces enable/disable_mmiotrace() critical sections. | |
55 | * - mmiotrace_enabled may be modified only when holding mmiotrace_mutex | |
56 | * and trace_lock. | |
57 | * - Routines depending on is_enabled() must take trace_lock. | |
58 | * - trace_list users must hold trace_lock. | |
9e57fb35 | 59 | * - is_enabled() guarantees that mmio_trace_{rw,mapping} are allowed. |
d61fc448 PP |
60 | * - pre/post callbacks assume the effect of is_enabled() being true. |
61 | */ | |
8b7d89d0 PP |
62 | |
63 | /* module parameters */ | |
d61fc448 | 64 | static unsigned long filter_offset; |
476bc001 RR |
65 | static bool nommiotrace; |
66 | static bool trace_pc; | |
8b7d89d0 | 67 | |
8b7d89d0 PP |
68 | module_param(filter_offset, ulong, 0); |
69 | module_param(nommiotrace, bool, 0); | |
8b7d89d0 PP |
70 | module_param(trace_pc, bool, 0); |
71 | ||
8b7d89d0 PP |
72 | MODULE_PARM_DESC(filter_offset, "Start address of traced mappings."); |
73 | MODULE_PARM_DESC(nommiotrace, "Disable actual MMIO tracing."); | |
8b7d89d0 | 74 | MODULE_PARM_DESC(trace_pc, "Record address of faulting instructions."); |
d61fc448 PP |
75 | |
76 | static bool is_enabled(void) | |
77 | { | |
78 | return atomic_read(&mmiotrace_enabled); | |
79 | } | |
8b7d89d0 | 80 | |
8b7d89d0 PP |
81 | static void print_pte(unsigned long address) |
82 | { | |
790e2a29 | 83 | unsigned int level; |
75bb8835 PP |
84 | pte_t *pte = lookup_address(address, &level); |
85 | ||
86 | if (!pte) { | |
3a0340be JP |
87 | pr_err("Error in %s: no pte for page 0x%08lx\n", |
88 | __func__, address); | |
75bb8835 PP |
89 | return; |
90 | } | |
91 | ||
92 | if (level == PG_LEVEL_2M) { | |
3a0340be JP |
93 | pr_emerg("4MB pages are not currently supported: 0x%08lx\n", |
94 | address); | |
8b7d89d0 PP |
95 | BUG(); |
96 | } | |
3a0340be JP |
97 | pr_info("pte for 0x%lx: 0x%llx 0x%llx\n", |
98 | address, | |
0663bb6c RD |
99 | (unsigned long long)pte_val(*pte), |
100 | (unsigned long long)pte_val(*pte) & _PAGE_PRESENT); | |
8b7d89d0 PP |
101 | } |
102 | ||
103 | /* | |
104 | * For some reason the pre/post pairs have been called in an | |
105 | * unmatched order. Report and die. | |
106 | */ | |
107 | static void die_kmmio_nesting_error(struct pt_regs *regs, unsigned long addr) | |
108 | { | |
f5136380 | 109 | const struct trap_reason *my_reason = &get_cpu_var(pf_reason); |
3a0340be JP |
110 | pr_emerg("unexpected fault for address: 0x%08lx, last fault for address: 0x%08lx\n", |
111 | addr, my_reason->addr); | |
8b7d89d0 | 112 | print_pte(addr); |
7b606162 SS |
113 | pr_emerg("faulting IP is at %pS\n", (void *)regs->ip); |
114 | pr_emerg("last faulting IP was at %pS\n", (void *)my_reason->ip); | |
8b7d89d0 | 115 | #ifdef __i386__ |
0fd0e3da | 116 | pr_emerg("eax: %08lx ebx: %08lx ecx: %08lx edx: %08lx\n", |
3a0340be | 117 | regs->ax, regs->bx, regs->cx, regs->dx); |
0fd0e3da | 118 | pr_emerg("esi: %08lx edi: %08lx ebp: %08lx esp: %08lx\n", |
3a0340be | 119 | regs->si, regs->di, regs->bp, regs->sp); |
8b7d89d0 | 120 | #else |
0fd0e3da | 121 | pr_emerg("rax: %016lx rcx: %016lx rdx: %016lx\n", |
3a0340be | 122 | regs->ax, regs->cx, regs->dx); |
0fd0e3da | 123 | pr_emerg("rsi: %016lx rdi: %016lx rbp: %016lx rsp: %016lx\n", |
3a0340be | 124 | regs->si, regs->di, regs->bp, regs->sp); |
8b7d89d0 | 125 | #endif |
f5136380 | 126 | put_cpu_var(pf_reason); |
8b7d89d0 PP |
127 | BUG(); |
128 | } | |
129 | ||
130 | static void pre(struct kmmio_probe *p, struct pt_regs *regs, | |
131 | unsigned long addr) | |
132 | { | |
f5136380 | 133 | struct trap_reason *my_reason = &get_cpu_var(pf_reason); |
bd8ac686 | 134 | struct mmiotrace_rw *my_trace = &get_cpu_var(cpu_trace); |
8b7d89d0 PP |
135 | const unsigned long instptr = instruction_pointer(regs); |
136 | const enum reason_type type = get_ins_type(instptr); | |
a50445d7 | 137 | struct remap_trace *trace = p->private; |
8b7d89d0 PP |
138 | |
139 | /* it doesn't make sense to have more than one active trace per cpu */ | |
f5136380 | 140 | if (my_reason->active_traces) |
8b7d89d0 PP |
141 | die_kmmio_nesting_error(regs, addr); |
142 | else | |
f5136380 | 143 | my_reason->active_traces++; |
8b7d89d0 | 144 | |
f5136380 PP |
145 | my_reason->type = type; |
146 | my_reason->addr = addr; | |
147 | my_reason->ip = instptr; | |
8b7d89d0 | 148 | |
bd8ac686 PP |
149 | my_trace->phys = addr - trace->probe.addr + trace->phys; |
150 | my_trace->map_id = trace->id; | |
8b7d89d0 PP |
151 | |
152 | /* | |
153 | * Only record the program counter when requested. | |
154 | * It may taint clean-room reverse engineering. | |
155 | */ | |
156 | if (trace_pc) | |
bd8ac686 | 157 | my_trace->pc = instptr; |
8b7d89d0 | 158 | else |
bd8ac686 | 159 | my_trace->pc = 0; |
8b7d89d0 | 160 | |
f984b51e PP |
161 | /* |
162 | * XXX: the timestamp recorded will be *after* the tracing has been | |
163 | * done, not at the time we hit the instruction. SMP implications | |
164 | * on event ordering? | |
165 | */ | |
8b7d89d0 PP |
166 | |
167 | switch (type) { | |
168 | case REG_READ: | |
bd8ac686 PP |
169 | my_trace->opcode = MMIO_READ; |
170 | my_trace->width = get_ins_mem_width(instptr); | |
8b7d89d0 PP |
171 | break; |
172 | case REG_WRITE: | |
bd8ac686 PP |
173 | my_trace->opcode = MMIO_WRITE; |
174 | my_trace->width = get_ins_mem_width(instptr); | |
175 | my_trace->value = get_ins_reg_val(instptr, regs); | |
8b7d89d0 PP |
176 | break; |
177 | case IMM_WRITE: | |
bd8ac686 PP |
178 | my_trace->opcode = MMIO_WRITE; |
179 | my_trace->width = get_ins_mem_width(instptr); | |
180 | my_trace->value = get_ins_imm_val(instptr); | |
8b7d89d0 PP |
181 | break; |
182 | default: | |
183 | { | |
184 | unsigned char *ip = (unsigned char *)instptr; | |
bd8ac686 PP |
185 | my_trace->opcode = MMIO_UNKNOWN_OP; |
186 | my_trace->width = 0; | |
187 | my_trace->value = (*ip) << 16 | *(ip + 1) << 8 | | |
f5136380 | 188 | *(ip + 2); |
8b7d89d0 PP |
189 | } |
190 | } | |
f5136380 PP |
191 | put_cpu_var(cpu_trace); |
192 | put_cpu_var(pf_reason); | |
8b7d89d0 PP |
193 | } |
194 | ||
195 | static void post(struct kmmio_probe *p, unsigned long condition, | |
196 | struct pt_regs *regs) | |
197 | { | |
f5136380 | 198 | struct trap_reason *my_reason = &get_cpu_var(pf_reason); |
bd8ac686 | 199 | struct mmiotrace_rw *my_trace = &get_cpu_var(cpu_trace); |
8b7d89d0 PP |
200 | |
201 | /* this should always return the active_trace count to 0 */ | |
f5136380 PP |
202 | my_reason->active_traces--; |
203 | if (my_reason->active_traces) { | |
3a0340be | 204 | pr_emerg("unexpected post handler"); |
8b7d89d0 PP |
205 | BUG(); |
206 | } | |
207 | ||
f5136380 | 208 | switch (my_reason->type) { |
8b7d89d0 | 209 | case REG_READ: |
bd8ac686 | 210 | my_trace->value = get_ins_reg_val(my_reason->ip, regs); |
8b7d89d0 PP |
211 | break; |
212 | default: | |
213 | break; | |
214 | } | |
f984b51e | 215 | |
bd8ac686 | 216 | mmio_trace_rw(my_trace); |
f5136380 PP |
217 | put_cpu_var(cpu_trace); |
218 | put_cpu_var(pf_reason); | |
8b7d89d0 PP |
219 | } |
220 | ||
dee310d0 | 221 | static void ioremap_trace_core(resource_size_t offset, unsigned long size, |
8b7d89d0 PP |
222 | void __iomem *addr) |
223 | { | |
d61fc448 | 224 | static atomic_t next_id; |
8b7d89d0 | 225 | struct remap_trace *trace = kmalloc(sizeof(*trace), GFP_KERNEL); |
87e547fe | 226 | /* These are page-unaligned. */ |
bd8ac686 PP |
227 | struct mmiotrace_map map = { |
228 | .phys = offset, | |
229 | .virt = (unsigned long)addr, | |
230 | .len = size, | |
231 | .opcode = MMIO_PROBE | |
8b7d89d0 | 232 | }; |
8b7d89d0 | 233 | |
d61fc448 | 234 | if (!trace) { |
3a0340be | 235 | pr_err("kmalloc failed in ioremap\n"); |
d61fc448 PP |
236 | return; |
237 | } | |
238 | ||
8b7d89d0 PP |
239 | *trace = (struct remap_trace) { |
240 | .probe = { | |
241 | .addr = (unsigned long)addr, | |
242 | .len = size, | |
243 | .pre_handler = pre, | |
244 | .post_handler = post, | |
a50445d7 | 245 | .private = trace |
d61fc448 PP |
246 | }, |
247 | .phys = offset, | |
248 | .id = atomic_inc_return(&next_id) | |
8b7d89d0 | 249 | }; |
bd8ac686 | 250 | map.map_id = trace->id; |
8b7d89d0 | 251 | |
d61fc448 | 252 | spin_lock_irq(&trace_lock); |
bbe5c783 PP |
253 | if (!is_enabled()) { |
254 | kfree(trace); | |
d61fc448 | 255 | goto not_enabled; |
bbe5c783 | 256 | } |
d61fc448 | 257 | |
bd8ac686 | 258 | mmio_trace_mapping(&map); |
8b7d89d0 | 259 | list_add_tail(&trace->list, &trace_list); |
8b7d89d0 PP |
260 | if (!nommiotrace) |
261 | register_kmmio_probe(&trace->probe); | |
d61fc448 PP |
262 | |
263 | not_enabled: | |
264 | spin_unlock_irq(&trace_lock); | |
8b7d89d0 PP |
265 | } |
266 | ||
dee310d0 PP |
267 | void mmiotrace_ioremap(resource_size_t offset, unsigned long size, |
268 | void __iomem *addr) | |
8b7d89d0 | 269 | { |
d61fc448 | 270 | if (!is_enabled()) /* recheck and proper locking in *_core() */ |
8b7d89d0 PP |
271 | return; |
272 | ||
3a0340be JP |
273 | pr_debug("ioremap_*(0x%llx, 0x%lx) = %p\n", |
274 | (unsigned long long)offset, size, addr); | |
d61fc448 | 275 | if ((filter_offset) && (offset != filter_offset)) |
8b7d89d0 | 276 | return; |
d61fc448 | 277 | ioremap_trace_core(offset, size, addr); |
8b7d89d0 | 278 | } |
8b7d89d0 | 279 | |
d61fc448 | 280 | static void iounmap_trace_core(volatile void __iomem *addr) |
8b7d89d0 | 281 | { |
bd8ac686 PP |
282 | struct mmiotrace_map map = { |
283 | .phys = 0, | |
284 | .virt = (unsigned long)addr, | |
285 | .len = 0, | |
286 | .opcode = MMIO_UNPROBE | |
8b7d89d0 PP |
287 | }; |
288 | struct remap_trace *trace; | |
289 | struct remap_trace *tmp; | |
d61fc448 PP |
290 | struct remap_trace *found_trace = NULL; |
291 | ||
3a0340be | 292 | pr_debug("Unmapping %p.\n", addr); |
8b7d89d0 | 293 | |
d61fc448 PP |
294 | spin_lock_irq(&trace_lock); |
295 | if (!is_enabled()) | |
296 | goto not_enabled; | |
297 | ||
8b7d89d0 PP |
298 | list_for_each_entry_safe(trace, tmp, &trace_list, list) { |
299 | if ((unsigned long)addr == trace->probe.addr) { | |
300 | if (!nommiotrace) | |
301 | unregister_kmmio_probe(&trace->probe); | |
302 | list_del(&trace->list); | |
d61fc448 | 303 | found_trace = trace; |
8b7d89d0 PP |
304 | break; |
305 | } | |
306 | } | |
bd8ac686 PP |
307 | map.map_id = (found_trace) ? found_trace->id : -1; |
308 | mmio_trace_mapping(&map); | |
d61fc448 PP |
309 | |
310 | not_enabled: | |
311 | spin_unlock_irq(&trace_lock); | |
312 | if (found_trace) { | |
313 | synchronize_rcu(); /* unregister_kmmio_probe() requirement */ | |
314 | kfree(found_trace); | |
315 | } | |
316 | } | |
317 | ||
318 | void mmiotrace_iounmap(volatile void __iomem *addr) | |
319 | { | |
320 | might_sleep(); | |
321 | if (is_enabled()) /* recheck and proper locking in *_core() */ | |
322 | iounmap_trace_core(addr); | |
8b7d89d0 | 323 | } |
8b7d89d0 | 324 | |
9e57fb35 PP |
325 | int mmiotrace_printk(const char *fmt, ...) |
326 | { | |
327 | int ret = 0; | |
328 | va_list args; | |
329 | unsigned long flags; | |
330 | va_start(args, fmt); | |
331 | ||
332 | spin_lock_irqsave(&trace_lock, flags); | |
333 | if (is_enabled()) | |
334 | ret = mmio_trace_printk(fmt, args); | |
335 | spin_unlock_irqrestore(&trace_lock, flags); | |
336 | ||
337 | va_end(args); | |
338 | return ret; | |
339 | } | |
340 | EXPORT_SYMBOL(mmiotrace_printk); | |
341 | ||
8b7d89d0 PP |
342 | static void clear_trace_list(void) |
343 | { | |
344 | struct remap_trace *trace; | |
345 | struct remap_trace *tmp; | |
346 | ||
d61fc448 PP |
347 | /* |
348 | * No locking required, because the caller ensures we are in a | |
349 | * critical section via mutex, and is_enabled() is false, | |
350 | * i.e. nothing can traverse or modify this list. | |
351 | * Caller also ensures is_enabled() cannot change. | |
352 | */ | |
353 | list_for_each_entry(trace, &trace_list, list) { | |
3a0340be JP |
354 | pr_notice("purging non-iounmapped trace @0x%08lx, size 0x%lx.\n", |
355 | trace->probe.addr, trace->probe.len); | |
8b7d89d0 PP |
356 | if (!nommiotrace) |
357 | unregister_kmmio_probe(&trace->probe); | |
d61fc448 PP |
358 | } |
359 | synchronize_rcu(); /* unregister_kmmio_probe() requirement */ | |
360 | ||
361 | list_for_each_entry_safe(trace, tmp, &trace_list, list) { | |
8b7d89d0 PP |
362 | list_del(&trace->list); |
363 | kfree(trace); | |
d61fc448 PP |
364 | } |
365 | } | |
366 | ||
7423d111 | 367 | #ifdef CONFIG_HOTPLUG_CPU |
c38da569 | 368 | static cpumask_var_t downed_cpus; |
7423d111 PP |
369 | |
370 | static void enter_uniprocessor(void) | |
371 | { | |
372 | int cpu; | |
373 | int err; | |
374 | ||
c38da569 RR |
375 | if (downed_cpus == NULL && |
376 | !alloc_cpumask_var(&downed_cpus, GFP_KERNEL)) { | |
3a0340be | 377 | pr_notice("Failed to allocate mask\n"); |
c38da569 RR |
378 | goto out; |
379 | } | |
380 | ||
7423d111 | 381 | get_online_cpus(); |
c38da569 RR |
382 | cpumask_copy(downed_cpus, cpu_online_mask); |
383 | cpumask_clear_cpu(cpumask_first(cpu_online_mask), downed_cpus); | |
7423d111 | 384 | if (num_online_cpus() > 1) |
3a0340be | 385 | pr_notice("Disabling non-boot CPUs...\n"); |
7423d111 PP |
386 | put_online_cpus(); |
387 | ||
c38da569 | 388 | for_each_cpu(cpu, downed_cpus) { |
7423d111 | 389 | err = cpu_down(cpu); |
970e6fa0 | 390 | if (!err) |
3a0340be | 391 | pr_info("CPU%d is down.\n", cpu); |
970e6fa0 | 392 | else |
3a0340be | 393 | pr_err("Error taking CPU%d down: %d\n", cpu, err); |
7423d111 | 394 | } |
c38da569 | 395 | out: |
7423d111 | 396 | if (num_online_cpus() > 1) |
3a0340be | 397 | pr_warning("multiple CPUs still online, may miss events.\n"); |
7423d111 PP |
398 | } |
399 | ||
148f9bb8 | 400 | static void leave_uniprocessor(void) |
7423d111 PP |
401 | { |
402 | int cpu; | |
403 | int err; | |
404 | ||
c38da569 | 405 | if (downed_cpus == NULL || cpumask_weight(downed_cpus) == 0) |
7423d111 | 406 | return; |
3a0340be | 407 | pr_notice("Re-enabling CPUs...\n"); |
c38da569 | 408 | for_each_cpu(cpu, downed_cpus) { |
7423d111 PP |
409 | err = cpu_up(cpu); |
410 | if (!err) | |
3a0340be | 411 | pr_info("enabled CPU%d.\n", cpu); |
7423d111 | 412 | else |
3a0340be | 413 | pr_err("cannot re-enable CPU%d: %d\n", cpu, err); |
7423d111 PP |
414 | } |
415 | } | |
416 | ||
417 | #else /* !CONFIG_HOTPLUG_CPU */ | |
418 | static void enter_uniprocessor(void) | |
419 | { | |
420 | if (num_online_cpus() > 1) | |
3a0340be JP |
421 | pr_warning("multiple CPUs are online, may miss events. " |
422 | "Suggest booting with maxcpus=1 kernel argument.\n"); | |
7423d111 PP |
423 | } |
424 | ||
425 | static void leave_uniprocessor(void) | |
426 | { | |
427 | } | |
428 | #endif | |
429 | ||
f984b51e | 430 | void enable_mmiotrace(void) |
d61fc448 PP |
431 | { |
432 | mutex_lock(&mmiotrace_mutex); | |
433 | if (is_enabled()) | |
434 | goto out; | |
435 | ||
d61fc448 | 436 | if (nommiotrace) |
3a0340be | 437 | pr_info("MMIO tracing disabled.\n"); |
0f9a623d | 438 | kmmio_init(); |
7423d111 | 439 | enter_uniprocessor(); |
d61fc448 PP |
440 | spin_lock_irq(&trace_lock); |
441 | atomic_inc(&mmiotrace_enabled); | |
442 | spin_unlock_irq(&trace_lock); | |
3a0340be | 443 | pr_info("enabled.\n"); |
d61fc448 PP |
444 | out: |
445 | mutex_unlock(&mmiotrace_mutex); | |
446 | } | |
447 | ||
f984b51e | 448 | void disable_mmiotrace(void) |
d61fc448 PP |
449 | { |
450 | mutex_lock(&mmiotrace_mutex); | |
451 | if (!is_enabled()) | |
452 | goto out; | |
453 | ||
454 | spin_lock_irq(&trace_lock); | |
455 | atomic_dec(&mmiotrace_enabled); | |
456 | BUG_ON(is_enabled()); | |
457 | spin_unlock_irq(&trace_lock); | |
458 | ||
459 | clear_trace_list(); /* guarantees: no more kmmio callbacks */ | |
7423d111 | 460 | leave_uniprocessor(); |
0f9a623d | 461 | kmmio_cleanup(); |
3a0340be | 462 | pr_info("disabled.\n"); |
d61fc448 PP |
463 | out: |
464 | mutex_unlock(&mmiotrace_mutex); | |
8b7d89d0 | 465 | } |