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 #include <linux/list.h>
9 #include <linux/rculist.h>
10 #include <linux/spinlock.h>
11 #include <linux/hash.h>
12 #include <linux/init.h>
13 #include <linux/module.h>
14 #include <linux/kernel.h>
15 #include <linux/uaccess.h>
16 #include <linux/ptrace.h>
17 #include <linux/preempt.h>
18 #include <linux/percpu.h>
19 #include <linux/kdebug.h>
20 #include <linux/mutex.h>
22 #include <asm/cacheflush.h>
23 #include <asm/tlbflush.h>
24 #include <linux/errno.h>
25 #include <asm/debugreg.h>
26 #include <linux/mmiotrace.h>
28 #define KMMIO_PAGE_HASH_BITS 4
29 #define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS)
31 struct kmmio_fault_page {
32 struct list_head list;
33 struct kmmio_fault_page *release_next;
34 unsigned long page; /* location of the fault page */
35 bool old_presence; /* page presence prior to arming */
39 * Number of times this page has been registered as a part
40 * of a probe. If zero, page is disarmed and this may be freed.
41 * Used only by writers (RCU) and post_kmmio_handler().
42 * Protected by kmmio_lock, when linked into kmmio_page_table.
47 struct kmmio_delayed_release {
49 struct kmmio_fault_page *release_list;
52 struct kmmio_context {
53 struct kmmio_fault_page *fpage;
54 struct kmmio_probe *probe;
55 unsigned long saved_flags;
60 static DEFINE_SPINLOCK(kmmio_lock);
62 /* Protected by kmmio_lock */
63 unsigned int kmmio_count;
65 /* Read-protected by RCU, write-protected by kmmio_lock. */
66 static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
67 static LIST_HEAD(kmmio_probes);
69 static struct list_head *kmmio_page_list(unsigned long page)
71 return &kmmio_page_table[hash_long(page, KMMIO_PAGE_HASH_BITS)];
74 /* Accessed per-cpu */
75 static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
78 * this is basically a dynamic stabbing problem:
79 * Could use the existing prio tree code or
80 * Possible better implementations:
81 * The Interval Skip List: A Data Structure for Finding All Intervals That
82 * Overlap a Point (might be simple)
83 * Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup
85 /* Get the kmmio at this addr (if any). You must be holding RCU read lock. */
86 static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
88 struct kmmio_probe *p;
89 list_for_each_entry_rcu(p, &kmmio_probes, list) {
90 if (addr >= p->addr && addr <= (p->addr + p->len))
96 /* You must be holding RCU read lock. */
97 static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page)
99 struct list_head *head;
100 struct kmmio_fault_page *p;
103 head = kmmio_page_list(page);
104 list_for_each_entry_rcu(p, head, list) {
111 static void set_pmd_presence(pmd_t *pmd, bool present, bool *old)
113 pmdval_t v = pmd_val(*pmd);
114 *old = !!(v & _PAGE_PRESENT);
118 set_pmd(pmd, __pmd(v));
121 static void set_pte_presence(pte_t *pte, bool present, bool *old)
123 pteval_t v = pte_val(*pte);
124 *old = !!(v & _PAGE_PRESENT);
128 set_pte_atomic(pte, __pte(v));
131 static int set_page_presence(unsigned long addr, bool present, bool *old)
134 pte_t *pte = lookup_address(addr, &level);
137 pr_err("kmmio: no pte for page 0x%08lx\n", addr);
143 set_pmd_presence((pmd_t *)pte, present, old);
146 set_pte_presence(pte, present, old);
149 pr_err("kmmio: unexpected page level 0x%x.\n", level);
153 __flush_tlb_one(addr);
158 * Mark the given page as not present. Access to it will trigger a fault.
160 * Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the
161 * protection is ignored here. RCU read lock is assumed held, so the struct
162 * will not disappear unexpectedly. Furthermore, the caller must guarantee,
163 * that double arming the same virtual address (page) cannot occur.
165 * Double disarming on the other hand is allowed, and may occur when a fault
166 * and mmiotrace shutdown happen simultaneously.
168 static int arm_kmmio_fault_page(struct kmmio_fault_page *f)
171 WARN_ONCE(f->armed, KERN_ERR "kmmio page already armed.\n");
173 pr_warning("kmmio double-arm: page 0x%08lx, ref %d, old %d\n",
174 f->page, f->count, f->old_presence);
176 ret = set_page_presence(f->page, false, &f->old_presence);
177 WARN_ONCE(ret < 0, KERN_ERR "kmmio arming 0x%08lx failed.\n", f->page);
182 /** Restore the given page to saved presence state. */
183 static void disarm_kmmio_fault_page(struct kmmio_fault_page *f)
186 int ret = set_page_presence(f->page, f->old_presence, &tmp);
188 KERN_ERR "kmmio disarming 0x%08lx failed.\n", f->page);
193 * This is being called from do_page_fault().
195 * We may be in an interrupt or a critical section. Also prefecthing may
196 * trigger a page fault. We may be in the middle of process switch.
197 * We cannot take any locks, because we could be executing especially
198 * within a kmmio critical section.
200 * Local interrupts are disabled, so preemption cannot happen.
201 * Do not enable interrupts, do not sleep, and watch out for other CPUs.
204 * Interrupts are disabled on entry as trap3 is an interrupt gate
205 * and they remain disabled thorough out this function.
207 int kmmio_handler(struct pt_regs *regs, unsigned long addr)
209 struct kmmio_context *ctx;
210 struct kmmio_fault_page *faultpage;
211 int ret = 0; /* default to fault not handled */
214 * Preemption is now disabled to prevent process switch during
215 * single stepping. We can only handle one active kmmio trace
216 * per cpu, so ensure that we finish it before something else
217 * gets to run. We also hold the RCU read lock over single
218 * stepping to avoid looking up the probe and kmmio_fault_page
224 faultpage = get_kmmio_fault_page(addr);
227 * Either this page fault is not caused by kmmio, or
228 * another CPU just pulled the kmmio probe from under
229 * our feet. The latter case should not be possible.
234 ctx = &get_cpu_var(kmmio_ctx);
236 if (addr == ctx->addr) {
238 * A second fault on the same page means some other
239 * condition needs handling by do_page_fault(), the
240 * page really not being present is the most common.
242 pr_debug("kmmio: secondary hit for 0x%08lx CPU %d.\n",
243 addr, smp_processor_id());
245 if (!faultpage->old_presence)
246 pr_info("kmmio: unexpected secondary hit for "
247 "address 0x%08lx on CPU %d.\n", addr,
251 * Prevent overwriting already in-flight context.
252 * This should not happen, let's hope disarming at
253 * least prevents a panic.
255 pr_emerg("kmmio: recursive probe hit on CPU %d, "
256 "for address 0x%08lx. Ignoring.\n",
257 smp_processor_id(), addr);
258 pr_emerg("kmmio: previous hit was at 0x%08lx.\n",
260 disarm_kmmio_fault_page(faultpage);
266 ctx->fpage = faultpage;
267 ctx->probe = get_kmmio_probe(addr);
268 ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
271 if (ctx->probe && ctx->probe->pre_handler)
272 ctx->probe->pre_handler(ctx->probe, regs, addr);
275 * Enable single-stepping and disable interrupts for the faulting
276 * context. Local interrupts must not get enabled during stepping.
278 regs->flags |= X86_EFLAGS_TF;
279 regs->flags &= ~X86_EFLAGS_IF;
281 /* Now we set present bit in PTE and single step. */
282 disarm_kmmio_fault_page(ctx->fpage);
285 * If another cpu accesses the same page while we are stepping,
286 * the access will not be caught. It will simply succeed and the
287 * only downside is we lose the event. If this becomes a problem,
288 * the user should drop to single cpu before tracing.
291 put_cpu_var(kmmio_ctx);
292 return 1; /* fault handled */
295 put_cpu_var(kmmio_ctx);
298 preempt_enable_no_resched();
303 * Interrupts are disabled on entry as trap1 is an interrupt gate
304 * and they remain disabled thorough out this function.
305 * This must always get called as the pair to kmmio_handler().
307 static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
310 struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
313 pr_warning("kmmio: spurious debug trap on CPU %d.\n",
318 if (ctx->probe && ctx->probe->post_handler)
319 ctx->probe->post_handler(ctx->probe, condition, regs);
321 /* Prevent racing against release_kmmio_fault_page(). */
322 spin_lock(&kmmio_lock);
323 if (ctx->fpage->count)
324 arm_kmmio_fault_page(ctx->fpage);
325 spin_unlock(&kmmio_lock);
327 regs->flags &= ~X86_EFLAGS_TF;
328 regs->flags |= ctx->saved_flags;
330 /* These were acquired in kmmio_handler(). */
334 preempt_enable_no_resched();
337 * if somebody else is singlestepping across a probe point, flags
338 * will have TF set, in which case, continue the remaining processing
339 * of do_debug, as if this is not a probe hit.
341 if (!(regs->flags & X86_EFLAGS_TF))
344 put_cpu_var(kmmio_ctx);
348 /* You must be holding kmmio_lock. */
349 static int add_kmmio_fault_page(unsigned long page)
351 struct kmmio_fault_page *f;
354 f = get_kmmio_fault_page(page);
357 arm_kmmio_fault_page(f);
362 f = kzalloc(sizeof(*f), GFP_ATOMIC);
369 if (arm_kmmio_fault_page(f)) {
374 list_add_rcu(&f->list, kmmio_page_list(f->page));
379 /* You must be holding kmmio_lock. */
380 static void release_kmmio_fault_page(unsigned long page,
381 struct kmmio_fault_page **release_list)
383 struct kmmio_fault_page *f;
386 f = get_kmmio_fault_page(page);
391 BUG_ON(f->count < 0);
393 disarm_kmmio_fault_page(f);
394 f->release_next = *release_list;
400 * With page-unaligned ioremaps, one or two armed pages may contain
401 * addresses from outside the intended mapping. Events for these addresses
402 * are currently silently dropped. The events may result only from programming
403 * mistakes by accessing addresses before the beginning or past the end of a
406 int register_kmmio_probe(struct kmmio_probe *p)
410 unsigned long size = 0;
411 const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
413 spin_lock_irqsave(&kmmio_lock, flags);
414 if (get_kmmio_probe(p->addr)) {
419 list_add_rcu(&p->list, &kmmio_probes);
420 while (size < size_lim) {
421 if (add_kmmio_fault_page(p->addr + size))
422 pr_err("kmmio: Unable to set page fault.\n");
426 spin_unlock_irqrestore(&kmmio_lock, flags);
428 * XXX: What should I do here?
429 * Here was a call to global_flush_tlb(), but it does not exist
430 * anymore. It seems it's not needed after all.
434 EXPORT_SYMBOL(register_kmmio_probe);
436 static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
438 struct kmmio_delayed_release *dr = container_of(
440 struct kmmio_delayed_release,
442 struct kmmio_fault_page *p = dr->release_list;
444 struct kmmio_fault_page *next = p->release_next;
452 static void remove_kmmio_fault_pages(struct rcu_head *head)
454 struct kmmio_delayed_release *dr = container_of(
456 struct kmmio_delayed_release,
458 struct kmmio_fault_page *p = dr->release_list;
459 struct kmmio_fault_page **prevp = &dr->release_list;
461 spin_lock_irqsave(&kmmio_lock, flags);
464 list_del_rcu(&p->list);
466 *prevp = p->release_next;
467 prevp = &p->release_next;
470 spin_unlock_irqrestore(&kmmio_lock, flags);
471 /* This is the real RCU destroy call. */
472 call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
476 * Remove a kmmio probe. You have to synchronize_rcu() before you can be
477 * sure that the callbacks will not be called anymore. Only after that
478 * you may actually release your struct kmmio_probe.
480 * Unregistering a kmmio fault page has three steps:
481 * 1. release_kmmio_fault_page()
482 * Disarm the page, wait a grace period to let all faults finish.
483 * 2. remove_kmmio_fault_pages()
484 * Remove the pages from kmmio_page_table.
485 * 3. rcu_free_kmmio_fault_pages()
486 * Actally free the kmmio_fault_page structs as with RCU.
488 void unregister_kmmio_probe(struct kmmio_probe *p)
491 unsigned long size = 0;
492 const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
493 struct kmmio_fault_page *release_list = NULL;
494 struct kmmio_delayed_release *drelease;
496 spin_lock_irqsave(&kmmio_lock, flags);
497 while (size < size_lim) {
498 release_kmmio_fault_page(p->addr + size, &release_list);
501 list_del_rcu(&p->list);
503 spin_unlock_irqrestore(&kmmio_lock, flags);
505 drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
507 pr_crit("kmmio: leaking kmmio_fault_page objects.\n");
510 drelease->release_list = release_list;
513 * This is not really RCU here. We have just disarmed a set of
514 * pages so that they cannot trigger page faults anymore. However,
515 * we cannot remove the pages from kmmio_page_table,
516 * because a probe hit might be in flight on another CPU. The
517 * pages are collected into a list, and they will be removed from
518 * kmmio_page_table when it is certain that no probe hit related to
519 * these pages can be in flight. RCU grace period sounds like a
522 * If we removed the pages too early, kmmio page fault handler might
523 * not find the respective kmmio_fault_page and determine it's not
524 * a kmmio fault, when it actually is. This would lead to madness.
526 call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
528 EXPORT_SYMBOL(unregister_kmmio_probe);
530 static int kmmio_die_notifier(struct notifier_block *nb, unsigned long val,
533 struct die_args *arg = args;
535 if (val == DIE_DEBUG && (arg->err & DR_STEP))
536 if (post_kmmio_handler(arg->err, arg->regs) == 1)
542 static struct notifier_block nb_die = {
543 .notifier_call = kmmio_die_notifier
546 static int __init init_kmmio(void)
549 for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
550 INIT_LIST_HEAD(&kmmio_page_table[i]);
551 return register_die_notifier(&nb_die);
553 fs_initcall(init_kmmio); /* should be before device_initcall() */