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).
46 struct kmmio_delayed_release {
48 struct kmmio_fault_page *release_list;
51 struct kmmio_context {
52 struct kmmio_fault_page *fpage;
53 struct kmmio_probe *probe;
54 unsigned long saved_flags;
59 static DEFINE_SPINLOCK(kmmio_lock);
61 /* Protected by kmmio_lock */
62 unsigned int kmmio_count;
64 /* Read-protected by RCU, write-protected by kmmio_lock. */
65 static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
66 static LIST_HEAD(kmmio_probes);
68 static struct list_head *kmmio_page_list(unsigned long page)
70 return &kmmio_page_table[hash_long(page, KMMIO_PAGE_HASH_BITS)];
73 /* Accessed per-cpu */
74 static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
77 * this is basically a dynamic stabbing problem:
78 * Could use the existing prio tree code or
79 * Possible better implementations:
80 * The Interval Skip List: A Data Structure for Finding All Intervals That
81 * Overlap a Point (might be simple)
82 * Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup
84 /* Get the kmmio at this addr (if any). You must be holding RCU read lock. */
85 static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
87 struct kmmio_probe *p;
88 list_for_each_entry_rcu(p, &kmmio_probes, list) {
89 if (addr >= p->addr && addr <= (p->addr + p->len))
95 /* You must be holding RCU read lock. */
96 static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page)
98 struct list_head *head;
99 struct kmmio_fault_page *p;
102 head = kmmio_page_list(page);
103 list_for_each_entry_rcu(p, head, list) {
110 static int set_page_presence(unsigned long addr, bool present, bool *old)
116 pte_t *pte = lookup_address(addr, &level);
119 pr_err("kmmio: no pte for page 0x%08lx\n", addr);
126 pmdval = pmd_val(*pmd);
127 *old = !!(pmdval & _PAGE_PRESENT);
128 pmdval &= ~_PAGE_PRESENT;
130 pmdval |= _PAGE_PRESENT;
131 set_pmd(pmd, __pmd(pmdval));
135 pteval = pte_val(*pte);
136 *old = !!(pteval & _PAGE_PRESENT);
137 pteval &= ~_PAGE_PRESENT;
139 pteval |= _PAGE_PRESENT;
140 set_pte_atomic(pte, __pte(pteval));
144 pr_err("kmmio: unexpected page level 0x%x.\n", level);
148 __flush_tlb_one(addr);
154 * Mark the given page as not present. Access to it will trigger a fault.
156 * Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the
157 * protection is ignored here. RCU read lock is assumed held, so the struct
158 * will not disappear unexpectedly. Furthermore, the caller must guarantee,
159 * that double arming the same virtual address (page) cannot occur.
161 * Double disarming on the other hand is allowed, and may occur when a fault
162 * and mmiotrace shutdown happen simultaneously.
164 static int arm_kmmio_fault_page(struct kmmio_fault_page *f)
167 WARN_ONCE(f->armed, KERN_ERR "kmmio page already armed.\n");
169 pr_warning("kmmio double-arm: page 0x%08lx, ref %d, old %d\n",
170 f->page, f->count, f->old_presence);
172 ret = set_page_presence(f->page, false, &f->old_presence);
173 WARN_ONCE(ret < 0, KERN_ERR "kmmio arming 0x%08lx failed.\n", f->page);
178 /** Restore the given page to saved presence state. */
179 static void disarm_kmmio_fault_page(struct kmmio_fault_page *f)
182 int ret = set_page_presence(f->page, f->old_presence, &tmp);
184 KERN_ERR "kmmio disarming 0x%08lx failed.\n", f->page);
189 * This is being called from do_page_fault().
191 * We may be in an interrupt or a critical section. Also prefecthing may
192 * trigger a page fault. We may be in the middle of process switch.
193 * We cannot take any locks, because we could be executing especially
194 * within a kmmio critical section.
196 * Local interrupts are disabled, so preemption cannot happen.
197 * Do not enable interrupts, do not sleep, and watch out for other CPUs.
200 * Interrupts are disabled on entry as trap3 is an interrupt gate
201 * and they remain disabled thorough out this function.
203 int kmmio_handler(struct pt_regs *regs, unsigned long addr)
205 struct kmmio_context *ctx;
206 struct kmmio_fault_page *faultpage;
207 int ret = 0; /* default to fault not handled */
210 * Preemption is now disabled to prevent process switch during
211 * single stepping. We can only handle one active kmmio trace
212 * per cpu, so ensure that we finish it before something else
213 * gets to run. We also hold the RCU read lock over single
214 * stepping to avoid looking up the probe and kmmio_fault_page
220 faultpage = get_kmmio_fault_page(addr);
223 * Either this page fault is not caused by kmmio, or
224 * another CPU just pulled the kmmio probe from under
225 * our feet. The latter case should not be possible.
230 ctx = &get_cpu_var(kmmio_ctx);
232 disarm_kmmio_fault_page(faultpage);
233 if (addr == ctx->addr) {
235 * On SMP we sometimes get recursive probe hits on the
236 * same address. Context is already saved, fall out.
238 pr_debug("kmmio: duplicate probe hit on CPU %d, for "
239 "address 0x%08lx.\n",
240 smp_processor_id(), addr);
245 * Prevent overwriting already in-flight context.
246 * This should not happen, let's hope disarming at least
249 pr_emerg("kmmio: recursive probe hit on CPU %d, "
250 "for address 0x%08lx. Ignoring.\n",
251 smp_processor_id(), addr);
252 pr_emerg("kmmio: previous hit was at 0x%08lx.\n",
258 ctx->fpage = faultpage;
259 ctx->probe = get_kmmio_probe(addr);
260 ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
263 if (ctx->probe && ctx->probe->pre_handler)
264 ctx->probe->pre_handler(ctx->probe, regs, addr);
267 * Enable single-stepping and disable interrupts for the faulting
268 * context. Local interrupts must not get enabled during stepping.
270 regs->flags |= X86_EFLAGS_TF;
271 regs->flags &= ~X86_EFLAGS_IF;
273 /* Now we set present bit in PTE and single step. */
274 disarm_kmmio_fault_page(ctx->fpage);
277 * If another cpu accesses the same page while we are stepping,
278 * the access will not be caught. It will simply succeed and the
279 * only downside is we lose the event. If this becomes a problem,
280 * the user should drop to single cpu before tracing.
283 put_cpu_var(kmmio_ctx);
284 return 1; /* fault handled */
287 put_cpu_var(kmmio_ctx);
290 preempt_enable_no_resched();
295 * Interrupts are disabled on entry as trap1 is an interrupt gate
296 * and they remain disabled thorough out this function.
297 * This must always get called as the pair to kmmio_handler().
299 static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
302 struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
305 pr_debug("kmmio: spurious debug trap on CPU %d.\n",
310 if (ctx->probe && ctx->probe->post_handler)
311 ctx->probe->post_handler(ctx->probe, condition, regs);
313 arm_kmmio_fault_page(ctx->fpage);
315 regs->flags &= ~X86_EFLAGS_TF;
316 regs->flags |= ctx->saved_flags;
318 /* These were acquired in kmmio_handler(). */
322 preempt_enable_no_resched();
325 * if somebody else is singlestepping across a probe point, flags
326 * will have TF set, in which case, continue the remaining processing
327 * of do_debug, as if this is not a probe hit.
329 if (!(regs->flags & X86_EFLAGS_TF))
332 put_cpu_var(kmmio_ctx);
336 /* You must be holding kmmio_lock. */
337 static int add_kmmio_fault_page(unsigned long page)
339 struct kmmio_fault_page *f;
342 f = get_kmmio_fault_page(page);
345 arm_kmmio_fault_page(f);
350 f = kzalloc(sizeof(*f), GFP_ATOMIC);
357 if (arm_kmmio_fault_page(f)) {
362 list_add_rcu(&f->list, kmmio_page_list(f->page));
367 /* You must be holding kmmio_lock. */
368 static void release_kmmio_fault_page(unsigned long page,
369 struct kmmio_fault_page **release_list)
371 struct kmmio_fault_page *f;
374 f = get_kmmio_fault_page(page);
379 BUG_ON(f->count < 0);
381 disarm_kmmio_fault_page(f);
382 f->release_next = *release_list;
388 * With page-unaligned ioremaps, one or two armed pages may contain
389 * addresses from outside the intended mapping. Events for these addresses
390 * are currently silently dropped. The events may result only from programming
391 * mistakes by accessing addresses before the beginning or past the end of a
394 int register_kmmio_probe(struct kmmio_probe *p)
398 unsigned long size = 0;
399 const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
401 spin_lock_irqsave(&kmmio_lock, flags);
402 if (get_kmmio_probe(p->addr)) {
407 list_add_rcu(&p->list, &kmmio_probes);
408 while (size < size_lim) {
409 if (add_kmmio_fault_page(p->addr + size))
410 pr_err("kmmio: Unable to set page fault.\n");
414 spin_unlock_irqrestore(&kmmio_lock, flags);
416 * XXX: What should I do here?
417 * Here was a call to global_flush_tlb(), but it does not exist
418 * anymore. It seems it's not needed after all.
422 EXPORT_SYMBOL(register_kmmio_probe);
424 static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
426 struct kmmio_delayed_release *dr = container_of(
428 struct kmmio_delayed_release,
430 struct kmmio_fault_page *p = dr->release_list;
432 struct kmmio_fault_page *next = p->release_next;
440 static void remove_kmmio_fault_pages(struct rcu_head *head)
442 struct kmmio_delayed_release *dr = container_of(
444 struct kmmio_delayed_release,
446 struct kmmio_fault_page *p = dr->release_list;
447 struct kmmio_fault_page **prevp = &dr->release_list;
449 spin_lock_irqsave(&kmmio_lock, flags);
452 list_del_rcu(&p->list);
454 *prevp = p->release_next;
455 prevp = &p->release_next;
458 spin_unlock_irqrestore(&kmmio_lock, flags);
459 /* This is the real RCU destroy call. */
460 call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
464 * Remove a kmmio probe. You have to synchronize_rcu() before you can be
465 * sure that the callbacks will not be called anymore. Only after that
466 * you may actually release your struct kmmio_probe.
468 * Unregistering a kmmio fault page has three steps:
469 * 1. release_kmmio_fault_page()
470 * Disarm the page, wait a grace period to let all faults finish.
471 * 2. remove_kmmio_fault_pages()
472 * Remove the pages from kmmio_page_table.
473 * 3. rcu_free_kmmio_fault_pages()
474 * Actally free the kmmio_fault_page structs as with RCU.
476 void unregister_kmmio_probe(struct kmmio_probe *p)
479 unsigned long size = 0;
480 const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
481 struct kmmio_fault_page *release_list = NULL;
482 struct kmmio_delayed_release *drelease;
484 spin_lock_irqsave(&kmmio_lock, flags);
485 while (size < size_lim) {
486 release_kmmio_fault_page(p->addr + size, &release_list);
489 list_del_rcu(&p->list);
491 spin_unlock_irqrestore(&kmmio_lock, flags);
493 drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
495 pr_crit("kmmio: leaking kmmio_fault_page objects.\n");
498 drelease->release_list = release_list;
501 * This is not really RCU here. We have just disarmed a set of
502 * pages so that they cannot trigger page faults anymore. However,
503 * we cannot remove the pages from kmmio_page_table,
504 * because a probe hit might be in flight on another CPU. The
505 * pages are collected into a list, and they will be removed from
506 * kmmio_page_table when it is certain that no probe hit related to
507 * these pages can be in flight. RCU grace period sounds like a
510 * If we removed the pages too early, kmmio page fault handler might
511 * not find the respective kmmio_fault_page and determine it's not
512 * a kmmio fault, when it actually is. This would lead to madness.
514 call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
516 EXPORT_SYMBOL(unregister_kmmio_probe);
518 static int kmmio_die_notifier(struct notifier_block *nb, unsigned long val,
521 struct die_args *arg = args;
523 if (val == DIE_DEBUG && (arg->err & DR_STEP))
524 if (post_kmmio_handler(arg->err, arg->regs) == 1)
530 static struct notifier_block nb_die = {
531 .notifier_call = kmmio_die_notifier
534 static int __init init_kmmio(void)
537 for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
538 INIT_LIST_HEAD(&kmmio_page_table[i]);
539 return register_die_notifier(&nb_die);
541 fs_initcall(init_kmmio); /* should be before device_initcall() */