kmmio.c 15.4 KB
Newer Older
1 2 3 4 5 6 7
/* Support for MMIO probes.
 * Benfit many code from kprobes
 * (C) 2002 Louis Zhuang <louis.zhuang@intel.com>.
 *     2007 Alexander Eichner
 *     2008 Pekka Paalanen <pq@iki.fi>
 */

8 9
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt

10
#include <linux/list.h>
11
#include <linux/rculist.h>
12 13 14 15 16 17 18 19
#include <linux/spinlock.h>
#include <linux/hash.h>
#include <linux/init.h>
#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/uaccess.h>
#include <linux/ptrace.h>
#include <linux/preempt.h>
20
#include <linux/percpu.h>
21
#include <linux/kdebug.h>
P
Pekka Paalanen 已提交
22
#include <linux/mutex.h>
P
Pekka Paalanen 已提交
23
#include <linux/io.h>
24 25
#include <asm/cacheflush.h>
#include <asm/tlbflush.h>
P
Pekka Paalanen 已提交
26
#include <linux/errno.h>
P
Pekka Paalanen 已提交
27
#include <asm/debugreg.h>
28
#include <linux/mmiotrace.h>
29 30 31 32

#define KMMIO_PAGE_HASH_BITS 4
#define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS)

33 34 35 36
struct kmmio_fault_page {
	struct list_head list;
	struct kmmio_fault_page *release_next;
	unsigned long page; /* location of the fault page */
37
	pteval_t old_presence; /* page presence prior to arming */
38
	bool armed;
39 40 41 42

	/*
	 * Number of times this page has been registered as a part
	 * of a probe. If zero, page is disarmed and this may be freed.
43 44
	 * Used only by writers (RCU) and post_kmmio_handler().
	 * Protected by kmmio_lock, when linked into kmmio_page_table.
45 46 47 48 49 50 51 52 53
	 */
	int count;
};

struct kmmio_delayed_release {
	struct rcu_head rcu;
	struct kmmio_fault_page *release_list;
};

54 55 56 57
struct kmmio_context {
	struct kmmio_fault_page *fpage;
	struct kmmio_probe *probe;
	unsigned long saved_flags;
58
	unsigned long addr;
59 60 61 62 63
	int active;
};

static DEFINE_SPINLOCK(kmmio_lock);

P
Pekka Paalanen 已提交
64
/* Protected by kmmio_lock */
65
unsigned int kmmio_count;
66 67

/* Read-protected by RCU, write-protected by kmmio_lock. */
68 69 70
static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
static LIST_HEAD(kmmio_probes);

71 72 73 74 75
static struct list_head *kmmio_page_list(unsigned long page)
{
	return &kmmio_page_table[hash_long(page, KMMIO_PAGE_HASH_BITS)];
}

76 77
/* Accessed per-cpu */
static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
78 79 80 81 82 83 84 85 86

/*
 * this is basically a dynamic stabbing problem:
 * Could use the existing prio tree code or
 * Possible better implementations:
 * The Interval Skip List: A Data Structure for Finding All Intervals That
 * Overlap a Point (might be simple)
 * Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup
 */
87
/* Get the kmmio at this addr (if any). You must be holding RCU read lock. */
88 89 90
static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
{
	struct kmmio_probe *p;
91
	list_for_each_entry_rcu(p, &kmmio_probes, list) {
92
		if (addr >= p->addr && addr < (p->addr + p->len))
93 94 95 96 97
			return p;
	}
	return NULL;
}

98
/* You must be holding RCU read lock. */
99 100
static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page)
{
101
	struct list_head *head;
102
	struct kmmio_fault_page *f;
103 104

	page &= PAGE_MASK;
105
	head = kmmio_page_list(page);
106 107 108
	list_for_each_entry_rcu(f, head, list) {
		if (f->page == page)
			return f;
109 110 111 112
	}
	return NULL;
}

113
static void clear_pmd_presence(pmd_t *pmd, bool clear, pmdval_t *old)
114 115
{
	pmdval_t v = pmd_val(*pmd);
116 117 118 119 120
	if (clear) {
		*old = v & _PAGE_PRESENT;
		v &= ~_PAGE_PRESENT;
	} else	/* presume this has been called with clear==true previously */
		v |= *old;
121 122 123
	set_pmd(pmd, __pmd(v));
}

124
static void clear_pte_presence(pte_t *pte, bool clear, pteval_t *old)
125 126
{
	pteval_t v = pte_val(*pte);
127 128 129 130 131
	if (clear) {
		*old = v & _PAGE_PRESENT;
		v &= ~_PAGE_PRESENT;
	} else	/* presume this has been called with clear==true previously */
		v |= *old;
132 133 134
	set_pte_atomic(pte, __pte(v));
}

135
static int clear_page_presence(struct kmmio_fault_page *f, bool clear)
136
{
137
	unsigned int level;
138
	pte_t *pte = lookup_address(f->page, &level);
139

140
	if (!pte) {
141
		pr_err("no pte for page 0x%08lx\n", f->page);
142
		return -1;
143 144
	}

P
Pekka Paalanen 已提交
145 146
	switch (level) {
	case PG_LEVEL_2M:
147
		clear_pmd_presence((pmd_t *)pte, clear, &f->old_presence);
P
Pekka Paalanen 已提交
148 149
		break;
	case PG_LEVEL_4K:
150
		clear_pte_presence(pte, clear, &f->old_presence);
P
Pekka Paalanen 已提交
151 152
		break;
	default:
153
		pr_err("unexpected page level 0x%x.\n", level);
154
		return -1;
155 156
	}

157
	__flush_tlb_one(f->page);
158
	return 0;
P
Pekka Paalanen 已提交
159
}
160

161 162 163 164 165 166 167 168 169 170 171 172
/*
 * Mark the given page as not present. Access to it will trigger a fault.
 *
 * Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the
 * protection is ignored here. RCU read lock is assumed held, so the struct
 * will not disappear unexpectedly. Furthermore, the caller must guarantee,
 * that double arming the same virtual address (page) cannot occur.
 *
 * Double disarming on the other hand is allowed, and may occur when a fault
 * and mmiotrace shutdown happen simultaneously.
 */
static int arm_kmmio_fault_page(struct kmmio_fault_page *f)
P
Pekka Paalanen 已提交
173
{
174
	int ret;
175
	WARN_ONCE(f->armed, KERN_ERR pr_fmt("kmmio page already armed.\n"));
176
	if (f->armed) {
177 178
		pr_warning("double-arm: page 0x%08lx, ref %d, old %d\n",
			   f->page, f->count, !!f->old_presence);
179
	}
180
	ret = clear_page_presence(f, true);
181 182
	WARN_ONCE(ret < 0, KERN_ERR pr_fmt("arming 0x%08lx failed.\n"),
		  f->page);
183
	f->armed = true;
184
	return ret;
185 186
}

187 188
/** Restore the given page to saved presence state. */
static void disarm_kmmio_fault_page(struct kmmio_fault_page *f)
189
{
190
	int ret = clear_page_presence(f, false);
191 192 193
	WARN_ONCE(ret < 0,
			KERN_ERR "kmmio disarming 0x%08lx failed.\n", f->page);
	f->armed = false;
194 195
}

196 197 198 199 200 201 202 203 204 205 206
/*
 * This is being called from do_page_fault().
 *
 * We may be in an interrupt or a critical section. Also prefecthing may
 * trigger a page fault. We may be in the middle of process switch.
 * We cannot take any locks, because we could be executing especially
 * within a kmmio critical section.
 *
 * Local interrupts are disabled, so preemption cannot happen.
 * Do not enable interrupts, do not sleep, and watch out for other CPUs.
 */
207 208
/*
 * Interrupts are disabled on entry as trap3 is an interrupt gate
209
 * and they remain disabled throughout this function.
210
 */
211
int kmmio_handler(struct pt_regs *regs, unsigned long addr)
212
{
213 214
	struct kmmio_context *ctx;
	struct kmmio_fault_page *faultpage;
P
Pekka Paalanen 已提交
215
	int ret = 0; /* default to fault not handled */
216 217 218 219 220

	/*
	 * Preemption is now disabled to prevent process switch during
	 * single stepping. We can only handle one active kmmio trace
	 * per cpu, so ensure that we finish it before something else
P
Pekka Paalanen 已提交
221 222 223
	 * gets to run. We also hold the RCU read lock over single
	 * stepping to avoid looking up the probe and kmmio_fault_page
	 * again.
224 225
	 */
	preempt_disable();
226
	rcu_read_lock();
P
Pekka Paalanen 已提交
227

228 229 230 231 232
	faultpage = get_kmmio_fault_page(addr);
	if (!faultpage) {
		/*
		 * Either this page fault is not caused by kmmio, or
		 * another CPU just pulled the kmmio probe from under
P
Pekka Paalanen 已提交
233
		 * our feet. The latter case should not be possible.
234 235 236 237 238
		 */
		goto no_kmmio;
	}

	ctx = &get_cpu_var(kmmio_ctx);
239
	if (ctx->active) {
P
Pekka Paalanen 已提交
240 241
		if (addr == ctx->addr) {
			/*
242 243 244
			 * A second fault on the same page means some other
			 * condition needs handling by do_page_fault(), the
			 * page really not being present is the most common.
P
Pekka Paalanen 已提交
245
			 */
246 247
			pr_debug("secondary hit for 0x%08lx CPU %d.\n",
				 addr, smp_processor_id());
248 249

			if (!faultpage->old_presence)
250 251
				pr_info("unexpected secondary hit for address 0x%08lx on CPU %d.\n",
					addr, smp_processor_id());
252 253 254 255 256 257
		} else {
			/*
			 * Prevent overwriting already in-flight context.
			 * This should not happen, let's hope disarming at
			 * least prevents a panic.
			 */
258 259 260
			pr_emerg("recursive probe hit on CPU %d, for address 0x%08lx. Ignoring.\n",
				 smp_processor_id(), addr);
			pr_emerg("previous hit was at 0x%08lx.\n", ctx->addr);
261 262
			disarm_kmmio_fault_page(faultpage);
		}
263
		goto no_kmmio_ctx;
264 265 266
	}
	ctx->active++;

267
	ctx->fpage = faultpage;
268
	ctx->probe = get_kmmio_probe(addr);
I
Ingo Molnar 已提交
269
	ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
270
	ctx->addr = addr;
271 272 273 274

	if (ctx->probe && ctx->probe->pre_handler)
		ctx->probe->pre_handler(ctx->probe, regs, addr);

P
Pekka Paalanen 已提交
275 276 277 278
	/*
	 * Enable single-stepping and disable interrupts for the faulting
	 * context. Local interrupts must not get enabled during stepping.
	 */
I
Ingo Molnar 已提交
279 280
	regs->flags |= X86_EFLAGS_TF;
	regs->flags &= ~X86_EFLAGS_IF;
281

282
	/* Now we set present bit in PTE and single step. */
283
	disarm_kmmio_fault_page(ctx->fpage);
284

P
Pekka Paalanen 已提交
285 286 287 288 289 290 291
	/*
	 * If another cpu accesses the same page while we are stepping,
	 * the access will not be caught. It will simply succeed and the
	 * only downside is we lose the event. If this becomes a problem,
	 * the user should drop to single cpu before tracing.
	 */

292
	put_cpu_var(kmmio_ctx);
P
Pekka Paalanen 已提交
293
	return 1; /* fault handled */
294

295 296
no_kmmio_ctx:
	put_cpu_var(kmmio_ctx);
297
no_kmmio:
298
	rcu_read_unlock();
299
	preempt_enable_no_resched();
P
Pekka Paalanen 已提交
300
	return ret;
301 302 303 304
}

/*
 * Interrupts are disabled on entry as trap1 is an interrupt gate
305
 * and they remain disabled throughout this function.
306
 * This must always get called as the pair to kmmio_handler().
307 308 309
 */
static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
{
310 311
	int ret = 0;
	struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
312

P
Pekka Paalanen 已提交
313
	if (!ctx->active) {
314 315 316 317 318
		/*
		 * debug traps without an active context are due to either
		 * something external causing them (f.e. using a debugger while
		 * mmio tracing enabled), or erroneous behaviour
		 */
319 320
		pr_warning("unexpected debug trap on CPU %d.\n",
			   smp_processor_id());
321
		goto out;
P
Pekka Paalanen 已提交
322
	}
323 324 325 326

	if (ctx->probe && ctx->probe->post_handler)
		ctx->probe->post_handler(ctx->probe, condition, regs);

327 328 329 330 331
	/* Prevent racing against release_kmmio_fault_page(). */
	spin_lock(&kmmio_lock);
	if (ctx->fpage->count)
		arm_kmmio_fault_page(ctx->fpage);
	spin_unlock(&kmmio_lock);
332

I
Ingo Molnar 已提交
333
	regs->flags &= ~X86_EFLAGS_TF;
334 335 336 337
	regs->flags |= ctx->saved_flags;

	/* These were acquired in kmmio_handler(). */
	ctx->active--;
338
	BUG_ON(ctx->active);
P
Pekka Paalanen 已提交
339
	rcu_read_unlock();
340 341 342 343 344 345 346
	preempt_enable_no_resched();

	/*
	 * if somebody else is singlestepping across a probe point, flags
	 * will have TF set, in which case, continue the remaining processing
	 * of do_debug, as if this is not a probe hit.
	 */
I
Ingo Molnar 已提交
347
	if (!(regs->flags & X86_EFLAGS_TF))
348 349 350 351
		ret = 1;
out:
	put_cpu_var(kmmio_ctx);
	return ret;
352 353
}

354
/* You must be holding kmmio_lock. */
355 356 357 358 359 360 361
static int add_kmmio_fault_page(unsigned long page)
{
	struct kmmio_fault_page *f;

	page &= PAGE_MASK;
	f = get_kmmio_fault_page(page);
	if (f) {
362
		if (!f->count)
363
			arm_kmmio_fault_page(f);
364 365 366 367
		f->count++;
		return 0;
	}

368
	f = kzalloc(sizeof(*f), GFP_ATOMIC);
369 370 371 372 373 374
	if (!f)
		return -1;

	f->count = 1;
	f->page = page;

375
	if (arm_kmmio_fault_page(f)) {
376 377 378 379 380
		kfree(f);
		return -1;
	}

	list_add_rcu(&f->list, kmmio_page_list(f->page));
381 382 383 384

	return 0;
}

385 386 387
/* You must be holding kmmio_lock. */
static void release_kmmio_fault_page(unsigned long page,
				struct kmmio_fault_page **release_list)
388 389 390 391 392 393 394 395 396
{
	struct kmmio_fault_page *f;

	page &= PAGE_MASK;
	f = get_kmmio_fault_page(page);
	if (!f)
		return;

	f->count--;
397
	BUG_ON(f->count < 0);
398
	if (!f->count) {
399
		disarm_kmmio_fault_page(f);
400 401
		f->release_next = *release_list;
		*release_list = f;
402 403 404
	}
}

405 406 407 408 409 410 411
/*
 * With page-unaligned ioremaps, one or two armed pages may contain
 * addresses from outside the intended mapping. Events for these addresses
 * are currently silently dropped. The events may result only from programming
 * mistakes by accessing addresses before the beginning or past the end of a
 * mapping.
 */
412 413
int register_kmmio_probe(struct kmmio_probe *p)
{
P
Pekka Paalanen 已提交
414
	unsigned long flags;
415 416
	int ret = 0;
	unsigned long size = 0;
417
	const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
418

P
Pekka Paalanen 已提交
419
	spin_lock_irqsave(&kmmio_lock, flags);
420 421 422 423
	if (get_kmmio_probe(p->addr)) {
		ret = -EEXIST;
		goto out;
	}
P
Pekka Paalanen 已提交
424
	kmmio_count++;
425
	list_add_rcu(&p->list, &kmmio_probes);
426
	while (size < size_lim) {
427
		if (add_kmmio_fault_page(p->addr + size))
428
			pr_err("Unable to set page fault.\n");
429 430 431
		size += PAGE_SIZE;
	}
out:
P
Pekka Paalanen 已提交
432
	spin_unlock_irqrestore(&kmmio_lock, flags);
433 434 435
	/*
	 * XXX: What should I do here?
	 * Here was a call to global_flush_tlb(), but it does not exist
436
	 * anymore. It seems it's not needed after all.
437 438 439
	 */
	return ret;
}
440
EXPORT_SYMBOL(register_kmmio_probe);
441

442 443 444 445 446 447
static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
{
	struct kmmio_delayed_release *dr = container_of(
						head,
						struct kmmio_delayed_release,
						rcu);
448 449 450 451 452 453
	struct kmmio_fault_page *f = dr->release_list;
	while (f) {
		struct kmmio_fault_page *next = f->release_next;
		BUG_ON(f->count);
		kfree(f);
		f = next;
454 455 456 457 458 459
	}
	kfree(dr);
}

static void remove_kmmio_fault_pages(struct rcu_head *head)
{
460 461
	struct kmmio_delayed_release *dr =
		container_of(head, struct kmmio_delayed_release, rcu);
462
	struct kmmio_fault_page *f = dr->release_list;
463 464
	struct kmmio_fault_page **prevp = &dr->release_list;
	unsigned long flags;
465

466
	spin_lock_irqsave(&kmmio_lock, flags);
467 468 469 470
	while (f) {
		if (!f->count) {
			list_del_rcu(&f->list);
			prevp = &f->release_next;
471
		} else {
472
			*prevp = f->release_next;
473
		}
474
		f = f->release_next;
475 476
	}
	spin_unlock_irqrestore(&kmmio_lock, flags);
477

478 479 480 481 482 483
	/* This is the real RCU destroy call. */
	call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
}

/*
 * Remove a kmmio probe. You have to synchronize_rcu() before you can be
P
Pekka Paalanen 已提交
484 485
 * sure that the callbacks will not be called anymore. Only after that
 * you may actually release your struct kmmio_probe.
486 487 488 489 490 491 492
 *
 * Unregistering a kmmio fault page has three steps:
 * 1. release_kmmio_fault_page()
 *    Disarm the page, wait a grace period to let all faults finish.
 * 2. remove_kmmio_fault_pages()
 *    Remove the pages from kmmio_page_table.
 * 3. rcu_free_kmmio_fault_pages()
493
 *    Actually free the kmmio_fault_page structs as with RCU.
494
 */
495 496
void unregister_kmmio_probe(struct kmmio_probe *p)
{
P
Pekka Paalanen 已提交
497
	unsigned long flags;
498
	unsigned long size = 0;
499
	const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
500 501
	struct kmmio_fault_page *release_list = NULL;
	struct kmmio_delayed_release *drelease;
502

P
Pekka Paalanen 已提交
503
	spin_lock_irqsave(&kmmio_lock, flags);
504
	while (size < size_lim) {
505
		release_kmmio_fault_page(p->addr + size, &release_list);
506 507
		size += PAGE_SIZE;
	}
508
	list_del_rcu(&p->list);
509
	kmmio_count--;
P
Pekka Paalanen 已提交
510
	spin_unlock_irqrestore(&kmmio_lock, flags);
511

512 513
	drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
	if (!drelease) {
514
		pr_crit("leaking kmmio_fault_page objects.\n");
515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533
		return;
	}
	drelease->release_list = release_list;

	/*
	 * This is not really RCU here. We have just disarmed a set of
	 * pages so that they cannot trigger page faults anymore. However,
	 * we cannot remove the pages from kmmio_page_table,
	 * because a probe hit might be in flight on another CPU. The
	 * pages are collected into a list, and they will be removed from
	 * kmmio_page_table when it is certain that no probe hit related to
	 * these pages can be in flight. RCU grace period sounds like a
	 * good choice.
	 *
	 * If we removed the pages too early, kmmio page fault handler might
	 * not find the respective kmmio_fault_page and determine it's not
	 * a kmmio fault, when it actually is. This would lead to madness.
	 */
	call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
534
}
535
EXPORT_SYMBOL(unregister_kmmio_probe);
536

537 538
static int
kmmio_die_notifier(struct notifier_block *nb, unsigned long val, void *args)
539 540
{
	struct die_args *arg = args;
541
	unsigned long* dr6_p = (unsigned long *)ERR_PTR(arg->err);
542

543 544
	if (val == DIE_DEBUG && (*dr6_p & DR_STEP))
		if (post_kmmio_handler(*dr6_p, arg->regs) == 1) {
545 546 547 548
			/*
			 * Reset the BS bit in dr6 (pointed by args->err) to
			 * denote completion of processing
			 */
549
			*dr6_p &= ~DR_STEP;
550
			return NOTIFY_STOP;
551
		}
552 553 554

	return NOTIFY_DONE;
}
P
Pekka Paalanen 已提交
555 556 557 558 559

static struct notifier_block nb_die = {
	.notifier_call = kmmio_die_notifier
};

560
int kmmio_init(void)
P
Pekka Paalanen 已提交
561 562
{
	int i;
563

P
Pekka Paalanen 已提交
564 565
	for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
		INIT_LIST_HEAD(&kmmio_page_table[i]);
566

P
Pekka Paalanen 已提交
567 568
	return register_die_notifier(&nb_die);
}
569 570 571 572 573 574 575 576 577 578 579

void kmmio_cleanup(void)
{
	int i;

	unregister_die_notifier(&nb_die);
	for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) {
		WARN_ONCE(!list_empty(&kmmio_page_table[i]),
			KERN_ERR "kmmio_page_table not empty at cleanup, any further tracing will leak memory.\n");
	}
}