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/init.h>
15 #include <linux/module.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>
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 page; /* location of the fault page */
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.
49 struct kmmio_delayed_release {
51 struct kmmio_fault_page *release_list;
54 struct kmmio_context {
55 struct kmmio_fault_page *fpage;
56 struct kmmio_probe *probe;
57 unsigned long saved_flags;
62 static DEFINE_SPINLOCK(kmmio_lock);
64 /* Protected by kmmio_lock */
65 unsigned int kmmio_count;
67 /* Read-protected by RCU, write-protected by kmmio_lock. */
68 static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
69 static LIST_HEAD(kmmio_probes);
71 static struct list_head *kmmio_page_list(unsigned long page)
73 return &kmmio_page_table[hash_long(page, KMMIO_PAGE_HASH_BITS)];
76 /* Accessed per-cpu */
77 static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
80 * this is basically a dynamic stabbing problem:
81 * Could use the existing prio tree code or
82 * Possible better implementations:
83 * The Interval Skip List: A Data Structure for Finding All Intervals That
84 * Overlap a Point (might be simple)
85 * 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 static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
90 struct kmmio_probe *p;
91 list_for_each_entry_rcu(p, &kmmio_probes, list) {
92 if (addr >= p->addr && addr < (p->addr + p->len))
98 /* You must be holding RCU read lock. */
99 static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page)
101 struct list_head *head;
102 struct kmmio_fault_page *f;
105 head = kmmio_page_list(page);
106 list_for_each_entry_rcu(f, head, list) {
113 static void clear_pmd_presence(pmd_t *pmd, bool clear, pmdval_t *old)
115 pmdval_t v = pmd_val(*pmd);
117 *old = v & _PAGE_PRESENT;
119 } else /* presume this has been called with clear==true previously */
121 set_pmd(pmd, __pmd(v));
124 static void clear_pte_presence(pte_t *pte, bool clear, pteval_t *old)
126 pteval_t v = pte_val(*pte);
128 *old = v & _PAGE_PRESENT;
130 } else /* presume this has been called with clear==true previously */
132 set_pte_atomic(pte, __pte(v));
135 static int clear_page_presence(struct kmmio_fault_page *f, bool clear)
138 pte_t *pte = lookup_address(f->page, &level);
141 pr_err("no pte for page 0x%08lx\n", f->page);
147 clear_pmd_presence((pmd_t *)pte, clear, &f->old_presence);
150 clear_pte_presence(pte, clear, &f->old_presence);
153 pr_err("unexpected page level 0x%x.\n", level);
157 __flush_tlb_one(f->page);
162 * Mark the given page as not present. Access to it will trigger a fault.
164 * Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the
165 * protection is ignored here. RCU read lock is assumed held, so the struct
166 * will not disappear unexpectedly. Furthermore, the caller must guarantee,
167 * that double arming the same virtual address (page) cannot occur.
169 * Double disarming on the other hand is allowed, and may occur when a fault
170 * and mmiotrace shutdown happen simultaneously.
172 static int arm_kmmio_fault_page(struct kmmio_fault_page *f)
175 WARN_ONCE(f->armed, KERN_ERR pr_fmt("kmmio page already armed.\n"));
177 pr_warning("double-arm: page 0x%08lx, ref %d, old %d\n",
178 f->page, f->count, !!f->old_presence);
180 ret = clear_page_presence(f, true);
181 WARN_ONCE(ret < 0, KERN_ERR pr_fmt("arming 0x%08lx failed.\n"),
187 /** Restore the given page to saved presence state. */
188 static void disarm_kmmio_fault_page(struct kmmio_fault_page *f)
190 int ret = clear_page_presence(f, false);
192 KERN_ERR "kmmio disarming 0x%08lx failed.\n", f->page);
197 * This is being called from do_page_fault().
199 * We may be in an interrupt or a critical section. Also prefecthing may
200 * trigger a page fault. We may be in the middle of process switch.
201 * We cannot take any locks, because we could be executing especially
202 * within a kmmio critical section.
204 * Local interrupts are disabled, so preemption cannot happen.
205 * Do not enable interrupts, do not sleep, and watch out for other CPUs.
208 * Interrupts are disabled on entry as trap3 is an interrupt gate
209 * and they remain disabled throughout this function.
211 int kmmio_handler(struct pt_regs *regs, unsigned long addr)
213 struct kmmio_context *ctx;
214 struct kmmio_fault_page *faultpage;
215 int ret = 0; /* default to fault not handled */
218 * Preemption is now disabled to prevent process switch during
219 * single stepping. We can only handle one active kmmio trace
220 * per cpu, so ensure that we finish it before something else
221 * gets to run. We also hold the RCU read lock over single
222 * stepping to avoid looking up the probe and kmmio_fault_page
228 faultpage = get_kmmio_fault_page(addr);
231 * Either this page fault is not caused by kmmio, or
232 * another CPU just pulled the kmmio probe from under
233 * our feet. The latter case should not be possible.
238 ctx = &get_cpu_var(kmmio_ctx);
240 if (addr == ctx->addr) {
242 * A second fault on the same page means some other
243 * condition needs handling by do_page_fault(), the
244 * page really not being present is the most common.
246 pr_debug("secondary hit for 0x%08lx CPU %d.\n",
247 addr, smp_processor_id());
249 if (!faultpage->old_presence)
250 pr_info("unexpected secondary hit for address 0x%08lx on CPU %d.\n",
251 addr, smp_processor_id());
254 * Prevent overwriting already in-flight context.
255 * This should not happen, let's hope disarming at
256 * least prevents a panic.
258 pr_emerg("recursive probe hit on CPU %d, for address 0x%08lx. Ignoring.\n",
259 smp_processor_id(), addr);
260 pr_emerg("previous hit was at 0x%08lx.\n", ctx->addr);
261 disarm_kmmio_fault_page(faultpage);
267 ctx->fpage = faultpage;
268 ctx->probe = get_kmmio_probe(addr);
269 ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
272 if (ctx->probe && ctx->probe->pre_handler)
273 ctx->probe->pre_handler(ctx->probe, regs, addr);
276 * Enable single-stepping and disable interrupts for the faulting
277 * context. Local interrupts must not get enabled during stepping.
279 regs->flags |= X86_EFLAGS_TF;
280 regs->flags &= ~X86_EFLAGS_IF;
282 /* Now we set present bit in PTE and single step. */
283 disarm_kmmio_fault_page(ctx->fpage);
286 * If another cpu accesses the same page while we are stepping,
287 * the access will not be caught. It will simply succeed and the
288 * only downside is we lose the event. If this becomes a problem,
289 * the user should drop to single cpu before tracing.
292 put_cpu_var(kmmio_ctx);
293 return 1; /* fault handled */
296 put_cpu_var(kmmio_ctx);
299 preempt_enable_no_resched();
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().
308 static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
311 struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
315 * debug traps without an active context are due to either
316 * something external causing them (f.e. using a debugger while
317 * mmio tracing enabled), or erroneous behaviour
319 pr_warning("unexpected debug trap on CPU %d.\n",
324 if (ctx->probe && ctx->probe->post_handler)
325 ctx->probe->post_handler(ctx->probe, condition, regs);
327 /* Prevent racing against release_kmmio_fault_page(). */
328 spin_lock(&kmmio_lock);
329 if (ctx->fpage->count)
330 arm_kmmio_fault_page(ctx->fpage);
331 spin_unlock(&kmmio_lock);
333 regs->flags &= ~X86_EFLAGS_TF;
334 regs->flags |= ctx->saved_flags;
336 /* These were acquired in kmmio_handler(). */
340 preempt_enable_no_resched();
343 * if somebody else is singlestepping across a probe point, flags
344 * will have TF set, in which case, continue the remaining processing
345 * of do_debug, as if this is not a probe hit.
347 if (!(regs->flags & X86_EFLAGS_TF))
350 put_cpu_var(kmmio_ctx);
354 /* You must be holding kmmio_lock. */
355 static int add_kmmio_fault_page(unsigned long page)
357 struct kmmio_fault_page *f;
360 f = get_kmmio_fault_page(page);
363 arm_kmmio_fault_page(f);
368 f = kzalloc(sizeof(*f), GFP_ATOMIC);
375 if (arm_kmmio_fault_page(f)) {
380 list_add_rcu(&f->list, kmmio_page_list(f->page));
385 /* You must be holding kmmio_lock. */
386 static void release_kmmio_fault_page(unsigned long page,
387 struct kmmio_fault_page **release_list)
389 struct kmmio_fault_page *f;
392 f = get_kmmio_fault_page(page);
397 BUG_ON(f->count < 0);
399 disarm_kmmio_fault_page(f);
400 f->release_next = *release_list;
406 * With page-unaligned ioremaps, one or two armed pages may contain
407 * addresses from outside the intended mapping. Events for these addresses
408 * are currently silently dropped. The events may result only from programming
409 * mistakes by accessing addresses before the beginning or past the end of a
412 int register_kmmio_probe(struct kmmio_probe *p)
416 unsigned long size = 0;
417 const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
419 spin_lock_irqsave(&kmmio_lock, flags);
420 if (get_kmmio_probe(p->addr)) {
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");
432 spin_unlock_irqrestore(&kmmio_lock, flags);
434 * XXX: What should I do here?
435 * Here was a call to global_flush_tlb(), but it does not exist
436 * anymore. It seems it's not needed after all.
440 EXPORT_SYMBOL(register_kmmio_probe);
442 static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
444 struct kmmio_delayed_release *dr = container_of(
446 struct kmmio_delayed_release,
448 struct kmmio_fault_page *f = dr->release_list;
450 struct kmmio_fault_page *next = f->release_next;
458 static void remove_kmmio_fault_pages(struct rcu_head *head)
460 struct kmmio_delayed_release *dr =
461 container_of(head, struct kmmio_delayed_release, rcu);
462 struct kmmio_fault_page *f = dr->release_list;
463 struct kmmio_fault_page **prevp = &dr->release_list;
466 spin_lock_irqsave(&kmmio_lock, flags);
469 list_del_rcu(&f->list);
470 prevp = &f->release_next;
472 *prevp = f->release_next;
476 spin_unlock_irqrestore(&kmmio_lock, flags);
478 /* This is the real RCU destroy call. */
479 call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
483 * Remove a kmmio probe. You have to synchronize_rcu() before you can be
484 * sure that the callbacks will not be called anymore. Only after that
485 * you may actually release your struct kmmio_probe.
487 * Unregistering a kmmio fault page has three steps:
488 * 1. release_kmmio_fault_page()
489 * Disarm the page, wait a grace period to let all faults finish.
490 * 2. remove_kmmio_fault_pages()
491 * Remove the pages from kmmio_page_table.
492 * 3. rcu_free_kmmio_fault_pages()
493 * Actually free the kmmio_fault_page structs as with RCU.
495 void unregister_kmmio_probe(struct kmmio_probe *p)
498 unsigned long size = 0;
499 const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
500 struct kmmio_fault_page *release_list = NULL;
501 struct kmmio_delayed_release *drelease;
503 spin_lock_irqsave(&kmmio_lock, flags);
504 while (size < size_lim) {
505 release_kmmio_fault_page(p->addr + size, &release_list);
508 list_del_rcu(&p->list);
510 spin_unlock_irqrestore(&kmmio_lock, flags);
512 drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
514 pr_crit("leaking kmmio_fault_page objects.\n");
517 drelease->release_list = release_list;
520 * This is not really RCU here. We have just disarmed a set of
521 * pages so that they cannot trigger page faults anymore. However,
522 * we cannot remove the pages from kmmio_page_table,
523 * because a probe hit might be in flight on another CPU. The
524 * pages are collected into a list, and they will be removed from
525 * kmmio_page_table when it is certain that no probe hit related to
526 * these pages can be in flight. RCU grace period sounds like a
529 * If we removed the pages too early, kmmio page fault handler might
530 * not find the respective kmmio_fault_page and determine it's not
531 * a kmmio fault, when it actually is. This would lead to madness.
533 call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
535 EXPORT_SYMBOL(unregister_kmmio_probe);
538 kmmio_die_notifier(struct notifier_block *nb, unsigned long val, void *args)
540 struct die_args *arg = args;
541 unsigned long* dr6_p = (unsigned long *)ERR_PTR(arg->err);
543 if (val == DIE_DEBUG && (*dr6_p & DR_STEP))
544 if (post_kmmio_handler(*dr6_p, arg->regs) == 1) {
546 * Reset the BS bit in dr6 (pointed by args->err) to
547 * denote completion of processing
556 static struct notifier_block nb_die = {
557 .notifier_call = kmmio_die_notifier
564 for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
565 INIT_LIST_HEAD(&kmmio_page_table[i]);
567 return register_die_notifier(&nb_die);
570 void kmmio_cleanup(void)
574 unregister_die_notifier(&nb_die);
575 for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) {
576 WARN_ONCE(!list_empty(&kmmio_page_table[i]),
577 KERN_ERR "kmmio_page_table not empty at cleanup, any further tracing will leak memory.\n");