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