xref: /openbmc/linux/arch/x86/mm/kmmio.c (revision d23015c1)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Support for MMIO probes.
3  * Benfit many code from kprobes
4  * (C) 2002 Louis Zhuang <louis.zhuang@intel.com>.
5  *     2007 Alexander Eichner
6  *     2008 Pekka Paalanen <pq@iki.fi>
7  */
8 
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 
11 #include <linux/list.h>
12 #include <linux/rculist.h>
13 #include <linux/spinlock.h>
14 #include <linux/hash.h>
15 #include <linux/export.h>
16 #include <linux/kernel.h>
17 #include <linux/uaccess.h>
18 #include <linux/ptrace.h>
19 #include <linux/preempt.h>
20 #include <linux/percpu.h>
21 #include <linux/kdebug.h>
22 #include <linux/mutex.h>
23 #include <linux/io.h>
24 #include <linux/slab.h>
25 #include <asm/cacheflush.h>
26 #include <asm/tlbflush.h>
27 #include <linux/errno.h>
28 #include <asm/debugreg.h>
29 #include <linux/mmiotrace.h>
30 
31 #define KMMIO_PAGE_HASH_BITS 4
32 #define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS)
33 
34 struct kmmio_fault_page {
35 	struct list_head list;
36 	struct kmmio_fault_page *release_next;
37 	unsigned long addr; /* the requested address */
38 	pteval_t old_presence; /* page presence prior to arming */
39 	bool armed;
40 
41 	/*
42 	 * Number of times this page has been registered as a part
43 	 * of a probe. If zero, page is disarmed and this may be freed.
44 	 * Used only by writers (RCU) and post_kmmio_handler().
45 	 * Protected by kmmio_lock, when linked into kmmio_page_table.
46 	 */
47 	int count;
48 
49 	bool scheduled_for_release;
50 };
51 
52 struct kmmio_delayed_release {
53 	struct rcu_head rcu;
54 	struct kmmio_fault_page *release_list;
55 };
56 
57 struct kmmio_context {
58 	struct kmmio_fault_page *fpage;
59 	struct kmmio_probe *probe;
60 	unsigned long saved_flags;
61 	unsigned long addr;
62 	int active;
63 };
64 
65 static DEFINE_SPINLOCK(kmmio_lock);
66 
67 /* Protected by kmmio_lock */
68 unsigned int kmmio_count;
69 
70 /* Read-protected by RCU, write-protected by kmmio_lock. */
71 static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
72 static LIST_HEAD(kmmio_probes);
73 
74 static struct list_head *kmmio_page_list(unsigned long addr)
75 {
76 	unsigned int l;
77 	pte_t *pte = lookup_address(addr, &l);
78 
79 	if (!pte)
80 		return NULL;
81 	addr &= page_level_mask(l);
82 
83 	return &kmmio_page_table[hash_long(addr, KMMIO_PAGE_HASH_BITS)];
84 }
85 
86 /* Accessed per-cpu */
87 static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
88 
89 /*
90  * this is basically a dynamic stabbing problem:
91  * Could use the existing prio tree code or
92  * Possible better implementations:
93  * The Interval Skip List: A Data Structure for Finding All Intervals That
94  * Overlap a Point (might be simple)
95  * Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup
96  */
97 /* Get the kmmio at this addr (if any). You must be holding RCU read lock. */
98 static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
99 {
100 	struct kmmio_probe *p;
101 	list_for_each_entry_rcu(p, &kmmio_probes, list) {
102 		if (addr >= p->addr && addr < (p->addr + p->len))
103 			return p;
104 	}
105 	return NULL;
106 }
107 
108 /* You must be holding RCU read lock. */
109 static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long addr)
110 {
111 	struct list_head *head;
112 	struct kmmio_fault_page *f;
113 	unsigned int l;
114 	pte_t *pte = lookup_address(addr, &l);
115 
116 	if (!pte)
117 		return NULL;
118 	addr &= page_level_mask(l);
119 	head = kmmio_page_list(addr);
120 	list_for_each_entry_rcu(f, head, list) {
121 		if (f->addr == addr)
122 			return f;
123 	}
124 	return NULL;
125 }
126 
127 static void clear_pmd_presence(pmd_t *pmd, bool clear, pmdval_t *old)
128 {
129 	pmd_t new_pmd;
130 	pmdval_t v = pmd_val(*pmd);
131 	if (clear) {
132 		*old = v;
133 		new_pmd = pmd_mknotpresent(*pmd);
134 	} else {
135 		/* Presume this has been called with clear==true previously */
136 		new_pmd = __pmd(*old);
137 	}
138 	set_pmd(pmd, new_pmd);
139 }
140 
141 static void clear_pte_presence(pte_t *pte, bool clear, pteval_t *old)
142 {
143 	pteval_t v = pte_val(*pte);
144 	if (clear) {
145 		*old = v;
146 		/* Nothing should care about address */
147 		pte_clear(&init_mm, 0, pte);
148 	} else {
149 		/* Presume this has been called with clear==true previously */
150 		set_pte_atomic(pte, __pte(*old));
151 	}
152 }
153 
154 static int clear_page_presence(struct kmmio_fault_page *f, bool clear)
155 {
156 	unsigned int level;
157 	pte_t *pte = lookup_address(f->addr, &level);
158 
159 	if (!pte) {
160 		pr_err("no pte for addr 0x%08lx\n", f->addr);
161 		return -1;
162 	}
163 
164 	switch (level) {
165 	case PG_LEVEL_2M:
166 		clear_pmd_presence((pmd_t *)pte, clear, &f->old_presence);
167 		break;
168 	case PG_LEVEL_4K:
169 		clear_pte_presence(pte, clear, &f->old_presence);
170 		break;
171 	default:
172 		pr_err("unexpected page level 0x%x.\n", level);
173 		return -1;
174 	}
175 
176 	__flush_tlb_one_kernel(f->addr);
177 	return 0;
178 }
179 
180 /*
181  * Mark the given page as not present. Access to it will trigger a fault.
182  *
183  * Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the
184  * protection is ignored here. RCU read lock is assumed held, so the struct
185  * will not disappear unexpectedly. Furthermore, the caller must guarantee,
186  * that double arming the same virtual address (page) cannot occur.
187  *
188  * Double disarming on the other hand is allowed, and may occur when a fault
189  * and mmiotrace shutdown happen simultaneously.
190  */
191 static int arm_kmmio_fault_page(struct kmmio_fault_page *f)
192 {
193 	int ret;
194 	WARN_ONCE(f->armed, KERN_ERR pr_fmt("kmmio page already armed.\n"));
195 	if (f->armed) {
196 		pr_warn("double-arm: addr 0x%08lx, ref %d, old %d\n",
197 			f->addr, f->count, !!f->old_presence);
198 	}
199 	ret = clear_page_presence(f, true);
200 	WARN_ONCE(ret < 0, KERN_ERR pr_fmt("arming at 0x%08lx failed.\n"),
201 		  f->addr);
202 	f->armed = true;
203 	return ret;
204 }
205 
206 /** Restore the given page to saved presence state. */
207 static void disarm_kmmio_fault_page(struct kmmio_fault_page *f)
208 {
209 	int ret = clear_page_presence(f, false);
210 	WARN_ONCE(ret < 0,
211 			KERN_ERR "kmmio disarming at 0x%08lx failed.\n", f->addr);
212 	f->armed = false;
213 }
214 
215 /*
216  * This is being called from do_page_fault().
217  *
218  * We may be in an interrupt or a critical section. Also prefecthing may
219  * trigger a page fault. We may be in the middle of process switch.
220  * We cannot take any locks, because we could be executing especially
221  * within a kmmio critical section.
222  *
223  * Local interrupts are disabled, so preemption cannot happen.
224  * Do not enable interrupts, do not sleep, and watch out for other CPUs.
225  */
226 /*
227  * Interrupts are disabled on entry as trap3 is an interrupt gate
228  * and they remain disabled throughout this function.
229  */
230 int kmmio_handler(struct pt_regs *regs, unsigned long addr)
231 {
232 	struct kmmio_context *ctx;
233 	struct kmmio_fault_page *faultpage;
234 	int ret = 0; /* default to fault not handled */
235 	unsigned long page_base = addr;
236 	unsigned int l;
237 	pte_t *pte = lookup_address(addr, &l);
238 	if (!pte)
239 		return -EINVAL;
240 	page_base &= page_level_mask(l);
241 
242 	/*
243 	 * Preemption is now disabled to prevent process switch during
244 	 * single stepping. We can only handle one active kmmio trace
245 	 * per cpu, so ensure that we finish it before something else
246 	 * gets to run. We also hold the RCU read lock over single
247 	 * stepping to avoid looking up the probe and kmmio_fault_page
248 	 * again.
249 	 */
250 	preempt_disable();
251 	rcu_read_lock();
252 
253 	faultpage = get_kmmio_fault_page(page_base);
254 	if (!faultpage) {
255 		/*
256 		 * Either this page fault is not caused by kmmio, or
257 		 * another CPU just pulled the kmmio probe from under
258 		 * our feet. The latter case should not be possible.
259 		 */
260 		goto no_kmmio;
261 	}
262 
263 	ctx = &get_cpu_var(kmmio_ctx);
264 	if (ctx->active) {
265 		if (page_base == ctx->addr) {
266 			/*
267 			 * A second fault on the same page means some other
268 			 * condition needs handling by do_page_fault(), the
269 			 * page really not being present is the most common.
270 			 */
271 			pr_debug("secondary hit for 0x%08lx CPU %d.\n",
272 				 addr, smp_processor_id());
273 
274 			if (!faultpage->old_presence)
275 				pr_info("unexpected secondary hit for address 0x%08lx on CPU %d.\n",
276 					addr, smp_processor_id());
277 		} else {
278 			/*
279 			 * Prevent overwriting already in-flight context.
280 			 * This should not happen, let's hope disarming at
281 			 * least prevents a panic.
282 			 */
283 			pr_emerg("recursive probe hit on CPU %d, for address 0x%08lx. Ignoring.\n",
284 				 smp_processor_id(), addr);
285 			pr_emerg("previous hit was at 0x%08lx.\n", ctx->addr);
286 			disarm_kmmio_fault_page(faultpage);
287 		}
288 		goto no_kmmio_ctx;
289 	}
290 	ctx->active++;
291 
292 	ctx->fpage = faultpage;
293 	ctx->probe = get_kmmio_probe(page_base);
294 	ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
295 	ctx->addr = page_base;
296 
297 	if (ctx->probe && ctx->probe->pre_handler)
298 		ctx->probe->pre_handler(ctx->probe, regs, addr);
299 
300 	/*
301 	 * Enable single-stepping and disable interrupts for the faulting
302 	 * context. Local interrupts must not get enabled during stepping.
303 	 */
304 	regs->flags |= X86_EFLAGS_TF;
305 	regs->flags &= ~X86_EFLAGS_IF;
306 
307 	/* Now we set present bit in PTE and single step. */
308 	disarm_kmmio_fault_page(ctx->fpage);
309 
310 	/*
311 	 * If another cpu accesses the same page while we are stepping,
312 	 * the access will not be caught. It will simply succeed and the
313 	 * only downside is we lose the event. If this becomes a problem,
314 	 * the user should drop to single cpu before tracing.
315 	 */
316 
317 	put_cpu_var(kmmio_ctx);
318 	return 1; /* fault handled */
319 
320 no_kmmio_ctx:
321 	put_cpu_var(kmmio_ctx);
322 no_kmmio:
323 	rcu_read_unlock();
324 	preempt_enable_no_resched();
325 	return ret;
326 }
327 
328 /*
329  * Interrupts are disabled on entry as trap1 is an interrupt gate
330  * and they remain disabled throughout this function.
331  * This must always get called as the pair to kmmio_handler().
332  */
333 static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
334 {
335 	int ret = 0;
336 	struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
337 
338 	if (!ctx->active) {
339 		/*
340 		 * debug traps without an active context are due to either
341 		 * something external causing them (f.e. using a debugger while
342 		 * mmio tracing enabled), or erroneous behaviour
343 		 */
344 		pr_warn("unexpected debug trap on CPU %d.\n", smp_processor_id());
345 		goto out;
346 	}
347 
348 	if (ctx->probe && ctx->probe->post_handler)
349 		ctx->probe->post_handler(ctx->probe, condition, regs);
350 
351 	/* Prevent racing against release_kmmio_fault_page(). */
352 	spin_lock(&kmmio_lock);
353 	if (ctx->fpage->count)
354 		arm_kmmio_fault_page(ctx->fpage);
355 	spin_unlock(&kmmio_lock);
356 
357 	regs->flags &= ~X86_EFLAGS_TF;
358 	regs->flags |= ctx->saved_flags;
359 
360 	/* These were acquired in kmmio_handler(). */
361 	ctx->active--;
362 	BUG_ON(ctx->active);
363 	rcu_read_unlock();
364 	preempt_enable_no_resched();
365 
366 	/*
367 	 * if somebody else is singlestepping across a probe point, flags
368 	 * will have TF set, in which case, continue the remaining processing
369 	 * of do_debug, as if this is not a probe hit.
370 	 */
371 	if (!(regs->flags & X86_EFLAGS_TF))
372 		ret = 1;
373 out:
374 	put_cpu_var(kmmio_ctx);
375 	return ret;
376 }
377 
378 /* You must be holding kmmio_lock. */
379 static int add_kmmio_fault_page(unsigned long addr)
380 {
381 	struct kmmio_fault_page *f;
382 
383 	f = get_kmmio_fault_page(addr);
384 	if (f) {
385 		if (!f->count)
386 			arm_kmmio_fault_page(f);
387 		f->count++;
388 		return 0;
389 	}
390 
391 	f = kzalloc(sizeof(*f), GFP_ATOMIC);
392 	if (!f)
393 		return -1;
394 
395 	f->count = 1;
396 	f->addr = addr;
397 
398 	if (arm_kmmio_fault_page(f)) {
399 		kfree(f);
400 		return -1;
401 	}
402 
403 	list_add_rcu(&f->list, kmmio_page_list(f->addr));
404 
405 	return 0;
406 }
407 
408 /* You must be holding kmmio_lock. */
409 static void release_kmmio_fault_page(unsigned long addr,
410 				struct kmmio_fault_page **release_list)
411 {
412 	struct kmmio_fault_page *f;
413 
414 	f = get_kmmio_fault_page(addr);
415 	if (!f)
416 		return;
417 
418 	f->count--;
419 	BUG_ON(f->count < 0);
420 	if (!f->count) {
421 		disarm_kmmio_fault_page(f);
422 		if (!f->scheduled_for_release) {
423 			f->release_next = *release_list;
424 			*release_list = f;
425 			f->scheduled_for_release = true;
426 		}
427 	}
428 }
429 
430 /*
431  * With page-unaligned ioremaps, one or two armed pages may contain
432  * addresses from outside the intended mapping. Events for these addresses
433  * are currently silently dropped. The events may result only from programming
434  * mistakes by accessing addresses before the beginning or past the end of a
435  * mapping.
436  */
437 int register_kmmio_probe(struct kmmio_probe *p)
438 {
439 	unsigned long flags;
440 	int ret = 0;
441 	unsigned long size = 0;
442 	unsigned long addr = p->addr & PAGE_MASK;
443 	const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
444 	unsigned int l;
445 	pte_t *pte;
446 
447 	spin_lock_irqsave(&kmmio_lock, flags);
448 	if (get_kmmio_probe(addr)) {
449 		ret = -EEXIST;
450 		goto out;
451 	}
452 
453 	pte = lookup_address(addr, &l);
454 	if (!pte) {
455 		ret = -EINVAL;
456 		goto out;
457 	}
458 
459 	kmmio_count++;
460 	list_add_rcu(&p->list, &kmmio_probes);
461 	while (size < size_lim) {
462 		if (add_kmmio_fault_page(addr + size))
463 			pr_err("Unable to set page fault.\n");
464 		size += page_level_size(l);
465 	}
466 out:
467 	spin_unlock_irqrestore(&kmmio_lock, flags);
468 	/*
469 	 * XXX: What should I do here?
470 	 * Here was a call to global_flush_tlb(), but it does not exist
471 	 * anymore. It seems it's not needed after all.
472 	 */
473 	return ret;
474 }
475 EXPORT_SYMBOL(register_kmmio_probe);
476 
477 static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
478 {
479 	struct kmmio_delayed_release *dr = container_of(
480 						head,
481 						struct kmmio_delayed_release,
482 						rcu);
483 	struct kmmio_fault_page *f = dr->release_list;
484 	while (f) {
485 		struct kmmio_fault_page *next = f->release_next;
486 		BUG_ON(f->count);
487 		kfree(f);
488 		f = next;
489 	}
490 	kfree(dr);
491 }
492 
493 static void remove_kmmio_fault_pages(struct rcu_head *head)
494 {
495 	struct kmmio_delayed_release *dr =
496 		container_of(head, struct kmmio_delayed_release, rcu);
497 	struct kmmio_fault_page *f = dr->release_list;
498 	struct kmmio_fault_page **prevp = &dr->release_list;
499 	unsigned long flags;
500 
501 	spin_lock_irqsave(&kmmio_lock, flags);
502 	while (f) {
503 		if (!f->count) {
504 			list_del_rcu(&f->list);
505 			prevp = &f->release_next;
506 		} else {
507 			*prevp = f->release_next;
508 			f->release_next = NULL;
509 			f->scheduled_for_release = false;
510 		}
511 		f = *prevp;
512 	}
513 	spin_unlock_irqrestore(&kmmio_lock, flags);
514 
515 	/* This is the real RCU destroy call. */
516 	call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
517 }
518 
519 /*
520  * Remove a kmmio probe. You have to synchronize_rcu() before you can be
521  * sure that the callbacks will not be called anymore. Only after that
522  * you may actually release your struct kmmio_probe.
523  *
524  * Unregistering a kmmio fault page has three steps:
525  * 1. release_kmmio_fault_page()
526  *    Disarm the page, wait a grace period to let all faults finish.
527  * 2. remove_kmmio_fault_pages()
528  *    Remove the pages from kmmio_page_table.
529  * 3. rcu_free_kmmio_fault_pages()
530  *    Actually free the kmmio_fault_page structs as with RCU.
531  */
532 void unregister_kmmio_probe(struct kmmio_probe *p)
533 {
534 	unsigned long flags;
535 	unsigned long size = 0;
536 	unsigned long addr = p->addr & PAGE_MASK;
537 	const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
538 	struct kmmio_fault_page *release_list = NULL;
539 	struct kmmio_delayed_release *drelease;
540 	unsigned int l;
541 	pte_t *pte;
542 
543 	pte = lookup_address(addr, &l);
544 	if (!pte)
545 		return;
546 
547 	spin_lock_irqsave(&kmmio_lock, flags);
548 	while (size < size_lim) {
549 		release_kmmio_fault_page(addr + size, &release_list);
550 		size += page_level_size(l);
551 	}
552 	list_del_rcu(&p->list);
553 	kmmio_count--;
554 	spin_unlock_irqrestore(&kmmio_lock, flags);
555 
556 	if (!release_list)
557 		return;
558 
559 	drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
560 	if (!drelease) {
561 		pr_crit("leaking kmmio_fault_page objects.\n");
562 		return;
563 	}
564 	drelease->release_list = release_list;
565 
566 	/*
567 	 * This is not really RCU here. We have just disarmed a set of
568 	 * pages so that they cannot trigger page faults anymore. However,
569 	 * we cannot remove the pages from kmmio_page_table,
570 	 * because a probe hit might be in flight on another CPU. The
571 	 * pages are collected into a list, and they will be removed from
572 	 * kmmio_page_table when it is certain that no probe hit related to
573 	 * these pages can be in flight. RCU grace period sounds like a
574 	 * good choice.
575 	 *
576 	 * If we removed the pages too early, kmmio page fault handler might
577 	 * not find the respective kmmio_fault_page and determine it's not
578 	 * a kmmio fault, when it actually is. This would lead to madness.
579 	 */
580 	call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
581 }
582 EXPORT_SYMBOL(unregister_kmmio_probe);
583 
584 static int
585 kmmio_die_notifier(struct notifier_block *nb, unsigned long val, void *args)
586 {
587 	struct die_args *arg = args;
588 	unsigned long* dr6_p = (unsigned long *)ERR_PTR(arg->err);
589 
590 	if (val == DIE_DEBUG && (*dr6_p & DR_STEP))
591 		if (post_kmmio_handler(*dr6_p, arg->regs) == 1) {
592 			/*
593 			 * Reset the BS bit in dr6 (pointed by args->err) to
594 			 * denote completion of processing
595 			 */
596 			*dr6_p &= ~DR_STEP;
597 			return NOTIFY_STOP;
598 		}
599 
600 	return NOTIFY_DONE;
601 }
602 
603 static struct notifier_block nb_die = {
604 	.notifier_call = kmmio_die_notifier
605 };
606 
607 int kmmio_init(void)
608 {
609 	int i;
610 
611 	for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
612 		INIT_LIST_HEAD(&kmmio_page_table[i]);
613 
614 	return register_die_notifier(&nb_die);
615 }
616 
617 void kmmio_cleanup(void)
618 {
619 	int i;
620 
621 	unregister_die_notifier(&nb_die);
622 	for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) {
623 		WARN_ONCE(!list_empty(&kmmio_page_table[i]),
624 			KERN_ERR "kmmio_page_table not empty at cleanup, any further tracing will leak memory.\n");
625 	}
626 }
627