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>
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 #include <linux/list.h>
11 #include <linux/rculist.h>
12 #include <linux/spinlock.h>
13 #include <linux/hash.h>
14 #include <linux/export.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>
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>
30 #define KMMIO_PAGE_HASH_BITS 4
31 #define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS)
33 struct kmmio_fault_page {
34 struct list_head list;
35 struct kmmio_fault_page *release_next;
36 unsigned long addr; /* the requested address */
37 pteval_t old_presence; /* page presence prior to arming */
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.
48 bool scheduled_for_release;
51 struct kmmio_delayed_release {
53 struct kmmio_fault_page *release_list;
56 struct kmmio_context {
57 struct kmmio_fault_page *fpage;
58 struct kmmio_probe *probe;
59 unsigned long saved_flags;
64 static DEFINE_SPINLOCK(kmmio_lock);
66 /* Protected by kmmio_lock */
67 unsigned int kmmio_count;
69 /* Read-protected by RCU, write-protected by kmmio_lock. */
70 static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
71 static LIST_HEAD(kmmio_probes);
73 static struct list_head *kmmio_page_list(unsigned long addr)
76 pte_t *pte = lookup_address(addr, &l);
80 addr &= page_level_mask(l);
82 return &kmmio_page_table[hash_long(addr, KMMIO_PAGE_HASH_BITS)];
85 /* Accessed per-cpu */
86 static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
89 * this is basically a dynamic stabbing problem:
90 * Could use the existing prio tree code or
91 * Possible better implementations:
92 * The Interval Skip List: A Data Structure for Finding All Intervals That
93 * Overlap a Point (might be simple)
94 * Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup
96 /* Get the kmmio at this addr (if any). You must be holding RCU read lock. */
97 static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
99 struct kmmio_probe *p;
100 list_for_each_entry_rcu(p, &kmmio_probes, list) {
101 if (addr >= p->addr && addr < (p->addr + p->len))
107 /* You must be holding RCU read lock. */
108 static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long addr)
110 struct list_head *head;
111 struct kmmio_fault_page *f;
113 pte_t *pte = lookup_address(addr, &l);
117 addr &= page_level_mask(l);
118 head = kmmio_page_list(addr);
119 list_for_each_entry_rcu(f, head, list) {
126 static void clear_pmd_presence(pmd_t *pmd, bool clear, pmdval_t *old)
129 pmdval_t v = pmd_val(*pmd);
132 new_pmd = pmd_mknotpresent(*pmd);
134 /* Presume this has been called with clear==true previously */
135 new_pmd = __pmd(*old);
137 set_pmd(pmd, new_pmd);
140 static void clear_pte_presence(pte_t *pte, bool clear, pteval_t *old)
142 pteval_t v = pte_val(*pte);
145 /* Nothing should care about address */
146 pte_clear(&init_mm, 0, pte);
148 /* Presume this has been called with clear==true previously */
149 set_pte_atomic(pte, __pte(*old));
153 static int clear_page_presence(struct kmmio_fault_page *f, bool clear)
156 pte_t *pte = lookup_address(f->addr, &level);
159 pr_err("no pte for addr 0x%08lx\n", f->addr);
165 clear_pmd_presence((pmd_t *)pte, clear, &f->old_presence);
168 clear_pte_presence(pte, clear, &f->old_presence);
171 pr_err("unexpected page level 0x%x.\n", level);
175 __flush_tlb_one(f->addr);
180 * Mark the given page as not present. Access to it will trigger a fault.
182 * Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the
183 * protection is ignored here. RCU read lock is assumed held, so the struct
184 * will not disappear unexpectedly. Furthermore, the caller must guarantee,
185 * that double arming the same virtual address (page) cannot occur.
187 * Double disarming on the other hand is allowed, and may occur when a fault
188 * and mmiotrace shutdown happen simultaneously.
190 static int arm_kmmio_fault_page(struct kmmio_fault_page *f)
193 WARN_ONCE(f->armed, KERN_ERR pr_fmt("kmmio page already armed.\n"));
195 pr_warning("double-arm: addr 0x%08lx, ref %d, old %d\n",
196 f->addr, f->count, !!f->old_presence);
198 ret = clear_page_presence(f, true);
199 WARN_ONCE(ret < 0, KERN_ERR pr_fmt("arming at 0x%08lx failed.\n"),
205 /** Restore the given page to saved presence state. */
206 static void disarm_kmmio_fault_page(struct kmmio_fault_page *f)
208 int ret = clear_page_presence(f, false);
210 KERN_ERR "kmmio disarming at 0x%08lx failed.\n", f->addr);
215 * This is being called from do_page_fault().
217 * We may be in an interrupt or a critical section. Also prefecthing may
218 * trigger a page fault. We may be in the middle of process switch.
219 * We cannot take any locks, because we could be executing especially
220 * within a kmmio critical section.
222 * Local interrupts are disabled, so preemption cannot happen.
223 * Do not enable interrupts, do not sleep, and watch out for other CPUs.
226 * Interrupts are disabled on entry as trap3 is an interrupt gate
227 * and they remain disabled throughout this function.
229 int kmmio_handler(struct pt_regs *regs, unsigned long addr)
231 struct kmmio_context *ctx;
232 struct kmmio_fault_page *faultpage;
233 int ret = 0; /* default to fault not handled */
234 unsigned long page_base = addr;
236 pte_t *pte = lookup_address(addr, &l);
239 page_base &= page_level_mask(l);
242 * Preemption is now disabled to prevent process switch during
243 * single stepping. We can only handle one active kmmio trace
244 * per cpu, so ensure that we finish it before something else
245 * gets to run. We also hold the RCU read lock over single
246 * stepping to avoid looking up the probe and kmmio_fault_page
252 faultpage = get_kmmio_fault_page(page_base);
255 * Either this page fault is not caused by kmmio, or
256 * another CPU just pulled the kmmio probe from under
257 * our feet. The latter case should not be possible.
262 ctx = &get_cpu_var(kmmio_ctx);
264 if (page_base == ctx->addr) {
266 * A second fault on the same page means some other
267 * condition needs handling by do_page_fault(), the
268 * page really not being present is the most common.
270 pr_debug("secondary hit for 0x%08lx CPU %d.\n",
271 addr, smp_processor_id());
273 if (!faultpage->old_presence)
274 pr_info("unexpected secondary hit for address 0x%08lx on CPU %d.\n",
275 addr, smp_processor_id());
278 * Prevent overwriting already in-flight context.
279 * This should not happen, let's hope disarming at
280 * least prevents a panic.
282 pr_emerg("recursive probe hit on CPU %d, for address 0x%08lx. Ignoring.\n",
283 smp_processor_id(), addr);
284 pr_emerg("previous hit was at 0x%08lx.\n", ctx->addr);
285 disarm_kmmio_fault_page(faultpage);
291 ctx->fpage = faultpage;
292 ctx->probe = get_kmmio_probe(page_base);
293 ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
294 ctx->addr = page_base;
296 if (ctx->probe && ctx->probe->pre_handler)
297 ctx->probe->pre_handler(ctx->probe, regs, addr);
300 * Enable single-stepping and disable interrupts for the faulting
301 * context. Local interrupts must not get enabled during stepping.
303 regs->flags |= X86_EFLAGS_TF;
304 regs->flags &= ~X86_EFLAGS_IF;
306 /* Now we set present bit in PTE and single step. */
307 disarm_kmmio_fault_page(ctx->fpage);
310 * If another cpu accesses the same page while we are stepping,
311 * the access will not be caught. It will simply succeed and the
312 * only downside is we lose the event. If this becomes a problem,
313 * the user should drop to single cpu before tracing.
316 put_cpu_var(kmmio_ctx);
317 return 1; /* fault handled */
320 put_cpu_var(kmmio_ctx);
323 preempt_enable_no_resched();
328 * Interrupts are disabled on entry as trap1 is an interrupt gate
329 * and they remain disabled throughout this function.
330 * This must always get called as the pair to kmmio_handler().
332 static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
335 struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
339 * debug traps without an active context are due to either
340 * something external causing them (f.e. using a debugger while
341 * mmio tracing enabled), or erroneous behaviour
343 pr_warning("unexpected debug trap on CPU %d.\n",
348 if (ctx->probe && ctx->probe->post_handler)
349 ctx->probe->post_handler(ctx->probe, condition, regs);
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);
357 regs->flags &= ~X86_EFLAGS_TF;
358 regs->flags |= ctx->saved_flags;
360 /* These were acquired in kmmio_handler(). */
364 preempt_enable_no_resched();
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.
371 if (!(regs->flags & X86_EFLAGS_TF))
374 put_cpu_var(kmmio_ctx);
378 /* You must be holding kmmio_lock. */
379 static int add_kmmio_fault_page(unsigned long addr)
381 struct kmmio_fault_page *f;
383 f = get_kmmio_fault_page(addr);
386 arm_kmmio_fault_page(f);
391 f = kzalloc(sizeof(*f), GFP_ATOMIC);
398 if (arm_kmmio_fault_page(f)) {
403 list_add_rcu(&f->list, kmmio_page_list(f->addr));
408 /* You must be holding kmmio_lock. */
409 static void release_kmmio_fault_page(unsigned long addr,
410 struct kmmio_fault_page **release_list)
412 struct kmmio_fault_page *f;
414 f = get_kmmio_fault_page(addr);
419 BUG_ON(f->count < 0);
421 disarm_kmmio_fault_page(f);
422 if (!f->scheduled_for_release) {
423 f->release_next = *release_list;
425 f->scheduled_for_release = true;
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
437 int register_kmmio_probe(struct kmmio_probe *p)
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);
447 spin_lock_irqsave(&kmmio_lock, flags);
448 if (get_kmmio_probe(addr)) {
453 pte = lookup_address(addr, &l);
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);
467 spin_unlock_irqrestore(&kmmio_lock, flags);
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.
475 EXPORT_SYMBOL(register_kmmio_probe);
477 static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
479 struct kmmio_delayed_release *dr = container_of(
481 struct kmmio_delayed_release,
483 struct kmmio_fault_page *f = dr->release_list;
485 struct kmmio_fault_page *next = f->release_next;
493 static void remove_kmmio_fault_pages(struct rcu_head *head)
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;
501 spin_lock_irqsave(&kmmio_lock, flags);
504 list_del_rcu(&f->list);
505 prevp = &f->release_next;
507 *prevp = f->release_next;
508 f->release_next = NULL;
509 f->scheduled_for_release = false;
513 spin_unlock_irqrestore(&kmmio_lock, flags);
515 /* This is the real RCU destroy call. */
516 call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
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.
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.
532 void unregister_kmmio_probe(struct kmmio_probe *p)
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;
543 pte = lookup_address(addr, &l);
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);
552 list_del_rcu(&p->list);
554 spin_unlock_irqrestore(&kmmio_lock, flags);
559 drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
561 pr_crit("leaking kmmio_fault_page objects.\n");
564 drelease->release_list = release_list;
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
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.
580 call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
582 EXPORT_SYMBOL(unregister_kmmio_probe);
585 kmmio_die_notifier(struct notifier_block *nb, unsigned long val, void *args)
587 struct die_args *arg = args;
588 unsigned long* dr6_p = (unsigned long *)ERR_PTR(arg->err);
590 if (val == DIE_DEBUG && (*dr6_p & DR_STEP))
591 if (post_kmmio_handler(*dr6_p, arg->regs) == 1) {
593 * Reset the BS bit in dr6 (pointed by args->err) to
594 * denote completion of processing
603 static struct notifier_block nb_die = {
604 .notifier_call = kmmio_die_notifier
611 for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
612 INIT_LIST_HEAD(&kmmio_page_table[i]);
614 return register_die_notifier(&nb_die);
617 void kmmio_cleanup(void)
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");