Commit | Line | Data |
---|---|---|
8b7d89d0 PP |
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> | |
6 | */ | |
7 | ||
0fd0e3da | 8 | #include <linux/list.h> |
668a6c36 | 9 | #include <linux/rculist.h> |
8b7d89d0 PP |
10 | #include <linux/spinlock.h> |
11 | #include <linux/hash.h> | |
12 | #include <linux/init.h> | |
13 | #include <linux/module.h> | |
8b7d89d0 | 14 | #include <linux/kernel.h> |
8b7d89d0 PP |
15 | #include <linux/uaccess.h> |
16 | #include <linux/ptrace.h> | |
17 | #include <linux/preempt.h> | |
f5136380 | 18 | #include <linux/percpu.h> |
0fd0e3da | 19 | #include <linux/kdebug.h> |
d61fc448 | 20 | #include <linux/mutex.h> |
970e6fa0 | 21 | #include <linux/io.h> |
8b7d89d0 | 22 | #include <asm/cacheflush.h> |
8b7d89d0 | 23 | #include <asm/tlbflush.h> |
970e6fa0 | 24 | #include <linux/errno.h> |
13829537 | 25 | #include <asm/debugreg.h> |
0fd0e3da | 26 | #include <linux/mmiotrace.h> |
8b7d89d0 | 27 | |
8b7d89d0 PP |
28 | #define KMMIO_PAGE_HASH_BITS 4 |
29 | #define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS) | |
30 | ||
0fd0e3da PP |
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 */ | |
46e91d00 | 35 | pteval_t old_presence; /* page presence prior to arming */ |
5359b585 | 36 | bool armed; |
0fd0e3da PP |
37 | |
38 | /* | |
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. | |
340430c5 PP |
41 | * Used only by writers (RCU) and post_kmmio_handler(). |
42 | * Protected by kmmio_lock, when linked into kmmio_page_table. | |
0fd0e3da PP |
43 | */ |
44 | int count; | |
45 | }; | |
46 | ||
47 | struct kmmio_delayed_release { | |
48 | struct rcu_head rcu; | |
49 | struct kmmio_fault_page *release_list; | |
50 | }; | |
51 | ||
8b7d89d0 PP |
52 | struct kmmio_context { |
53 | struct kmmio_fault_page *fpage; | |
54 | struct kmmio_probe *probe; | |
55 | unsigned long saved_flags; | |
0fd0e3da | 56 | unsigned long addr; |
8b7d89d0 PP |
57 | int active; |
58 | }; | |
59 | ||
8b7d89d0 PP |
60 | static DEFINE_SPINLOCK(kmmio_lock); |
61 | ||
13829537 | 62 | /* Protected by kmmio_lock */ |
8b7d89d0 | 63 | unsigned int kmmio_count; |
0fd0e3da PP |
64 | |
65 | /* Read-protected by RCU, write-protected by kmmio_lock. */ | |
8b7d89d0 PP |
66 | static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE]; |
67 | static LIST_HEAD(kmmio_probes); | |
68 | ||
0fd0e3da PP |
69 | static struct list_head *kmmio_page_list(unsigned long page) |
70 | { | |
71 | return &kmmio_page_table[hash_long(page, KMMIO_PAGE_HASH_BITS)]; | |
72 | } | |
73 | ||
f5136380 PP |
74 | /* Accessed per-cpu */ |
75 | static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx); | |
8b7d89d0 | 76 | |
8b7d89d0 PP |
77 | /* |
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 | |
84 | */ | |
0fd0e3da | 85 | /* Get the kmmio at this addr (if any). You must be holding RCU read lock. */ |
8b7d89d0 PP |
86 | static struct kmmio_probe *get_kmmio_probe(unsigned long addr) |
87 | { | |
88 | struct kmmio_probe *p; | |
0fd0e3da | 89 | list_for_each_entry_rcu(p, &kmmio_probes, list) { |
33015c85 | 90 | if (addr >= p->addr && addr < (p->addr + p->len)) |
8b7d89d0 PP |
91 | return p; |
92 | } | |
93 | return NULL; | |
94 | } | |
95 | ||
0fd0e3da | 96 | /* You must be holding RCU read lock. */ |
8b7d89d0 PP |
97 | static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page) |
98 | { | |
0fd0e3da | 99 | struct list_head *head; |
0492e1bb | 100 | struct kmmio_fault_page *f; |
8b7d89d0 PP |
101 | |
102 | page &= PAGE_MASK; | |
0fd0e3da | 103 | head = kmmio_page_list(page); |
0492e1bb SB |
104 | list_for_each_entry_rcu(f, head, list) { |
105 | if (f->page == page) | |
106 | return f; | |
8b7d89d0 | 107 | } |
8b7d89d0 PP |
108 | return NULL; |
109 | } | |
110 | ||
46e91d00 | 111 | static void clear_pmd_presence(pmd_t *pmd, bool clear, pmdval_t *old) |
0b700a6a PP |
112 | { |
113 | pmdval_t v = pmd_val(*pmd); | |
46e91d00 SB |
114 | if (clear) { |
115 | *old = v & _PAGE_PRESENT; | |
116 | v &= ~_PAGE_PRESENT; | |
117 | } else /* presume this has been called with clear==true previously */ | |
118 | v |= *old; | |
0b700a6a PP |
119 | set_pmd(pmd, __pmd(v)); |
120 | } | |
121 | ||
46e91d00 | 122 | static void clear_pte_presence(pte_t *pte, bool clear, pteval_t *old) |
0b700a6a PP |
123 | { |
124 | pteval_t v = pte_val(*pte); | |
46e91d00 SB |
125 | if (clear) { |
126 | *old = v & _PAGE_PRESENT; | |
127 | v &= ~_PAGE_PRESENT; | |
128 | } else /* presume this has been called with clear==true previously */ | |
129 | v |= *old; | |
0b700a6a PP |
130 | set_pte_atomic(pte, __pte(v)); |
131 | } | |
132 | ||
46e91d00 | 133 | static int clear_page_presence(struct kmmio_fault_page *f, bool clear) |
8b7d89d0 | 134 | { |
790e2a29 | 135 | unsigned int level; |
46e91d00 | 136 | pte_t *pte = lookup_address(f->page, &level); |
8b7d89d0 | 137 | |
75bb8835 | 138 | if (!pte) { |
46e91d00 | 139 | pr_err("kmmio: no pte for page 0x%08lx\n", f->page); |
e9d54cae | 140 | return -1; |
75bb8835 PP |
141 | } |
142 | ||
13829537 PP |
143 | switch (level) { |
144 | case PG_LEVEL_2M: | |
46e91d00 | 145 | clear_pmd_presence((pmd_t *)pte, clear, &f->old_presence); |
13829537 | 146 | break; |
13829537 | 147 | case PG_LEVEL_4K: |
46e91d00 | 148 | clear_pte_presence(pte, clear, &f->old_presence); |
13829537 | 149 | break; |
13829537 PP |
150 | default: |
151 | pr_err("kmmio: unexpected page level 0x%x.\n", level); | |
e9d54cae | 152 | return -1; |
8b7d89d0 PP |
153 | } |
154 | ||
46e91d00 | 155 | __flush_tlb_one(f->page); |
e9d54cae | 156 | return 0; |
13829537 | 157 | } |
75bb8835 | 158 | |
5359b585 PP |
159 | /* |
160 | * Mark the given page as not present. Access to it will trigger a fault. | |
161 | * | |
162 | * Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the | |
163 | * protection is ignored here. RCU read lock is assumed held, so the struct | |
164 | * will not disappear unexpectedly. Furthermore, the caller must guarantee, | |
165 | * that double arming the same virtual address (page) cannot occur. | |
166 | * | |
167 | * Double disarming on the other hand is allowed, and may occur when a fault | |
168 | * and mmiotrace shutdown happen simultaneously. | |
169 | */ | |
170 | static int arm_kmmio_fault_page(struct kmmio_fault_page *f) | |
13829537 | 171 | { |
5359b585 PP |
172 | int ret; |
173 | WARN_ONCE(f->armed, KERN_ERR "kmmio page already armed.\n"); | |
174 | if (f->armed) { | |
175 | pr_warning("kmmio double-arm: page 0x%08lx, ref %d, old %d\n", | |
46e91d00 | 176 | f->page, f->count, !!f->old_presence); |
5359b585 | 177 | } |
46e91d00 | 178 | ret = clear_page_presence(f, true); |
5359b585 PP |
179 | WARN_ONCE(ret < 0, KERN_ERR "kmmio arming 0x%08lx failed.\n", f->page); |
180 | f->armed = true; | |
e9d54cae | 181 | return ret; |
8b7d89d0 PP |
182 | } |
183 | ||
5359b585 PP |
184 | /** Restore the given page to saved presence state. */ |
185 | static void disarm_kmmio_fault_page(struct kmmio_fault_page *f) | |
8b7d89d0 | 186 | { |
46e91d00 | 187 | int ret = clear_page_presence(f, false); |
5359b585 PP |
188 | WARN_ONCE(ret < 0, |
189 | KERN_ERR "kmmio disarming 0x%08lx failed.\n", f->page); | |
190 | f->armed = false; | |
8b7d89d0 PP |
191 | } |
192 | ||
0fd0e3da PP |
193 | /* |
194 | * This is being called from do_page_fault(). | |
195 | * | |
196 | * We may be in an interrupt or a critical section. Also prefecthing may | |
197 | * trigger a page fault. We may be in the middle of process switch. | |
198 | * We cannot take any locks, because we could be executing especially | |
199 | * within a kmmio critical section. | |
200 | * | |
201 | * Local interrupts are disabled, so preemption cannot happen. | |
202 | * Do not enable interrupts, do not sleep, and watch out for other CPUs. | |
203 | */ | |
8b7d89d0 PP |
204 | /* |
205 | * Interrupts are disabled on entry as trap3 is an interrupt gate | |
206 | * and they remain disabled thorough out this function. | |
207 | */ | |
0fd0e3da | 208 | int kmmio_handler(struct pt_regs *regs, unsigned long addr) |
8b7d89d0 | 209 | { |
0fd0e3da PP |
210 | struct kmmio_context *ctx; |
211 | struct kmmio_fault_page *faultpage; | |
13829537 | 212 | int ret = 0; /* default to fault not handled */ |
8b7d89d0 PP |
213 | |
214 | /* | |
215 | * Preemption is now disabled to prevent process switch during | |
216 | * single stepping. We can only handle one active kmmio trace | |
217 | * per cpu, so ensure that we finish it before something else | |
d61fc448 PP |
218 | * gets to run. We also hold the RCU read lock over single |
219 | * stepping to avoid looking up the probe and kmmio_fault_page | |
220 | * again. | |
8b7d89d0 PP |
221 | */ |
222 | preempt_disable(); | |
0fd0e3da | 223 | rcu_read_lock(); |
d61fc448 | 224 | |
0fd0e3da PP |
225 | faultpage = get_kmmio_fault_page(addr); |
226 | if (!faultpage) { | |
227 | /* | |
228 | * Either this page fault is not caused by kmmio, or | |
229 | * another CPU just pulled the kmmio probe from under | |
13829537 | 230 | * our feet. The latter case should not be possible. |
0fd0e3da PP |
231 | */ |
232 | goto no_kmmio; | |
233 | } | |
234 | ||
235 | ctx = &get_cpu_var(kmmio_ctx); | |
8b7d89d0 | 236 | if (ctx->active) { |
13829537 PP |
237 | if (addr == ctx->addr) { |
238 | /* | |
3e39aa15 SB |
239 | * A second fault on the same page means some other |
240 | * condition needs handling by do_page_fault(), the | |
241 | * page really not being present is the most common. | |
13829537 | 242 | */ |
3e39aa15 SB |
243 | pr_debug("kmmio: secondary hit for 0x%08lx CPU %d.\n", |
244 | addr, smp_processor_id()); | |
245 | ||
246 | if (!faultpage->old_presence) | |
247 | pr_info("kmmio: unexpected secondary hit for " | |
248 | "address 0x%08lx on CPU %d.\n", addr, | |
249 | smp_processor_id()); | |
250 | } else { | |
251 | /* | |
252 | * Prevent overwriting already in-flight context. | |
253 | * This should not happen, let's hope disarming at | |
254 | * least prevents a panic. | |
255 | */ | |
256 | pr_emerg("kmmio: recursive probe hit on CPU %d, " | |
0fd0e3da | 257 | "for address 0x%08lx. Ignoring.\n", |
f5136380 | 258 | smp_processor_id(), addr); |
3e39aa15 SB |
259 | pr_emerg("kmmio: previous hit was at 0x%08lx.\n", |
260 | ctx->addr); | |
261 | disarm_kmmio_fault_page(faultpage); | |
262 | } | |
0fd0e3da | 263 | goto no_kmmio_ctx; |
8b7d89d0 PP |
264 | } |
265 | ctx->active++; | |
266 | ||
0fd0e3da | 267 | ctx->fpage = faultpage; |
8b7d89d0 | 268 | ctx->probe = get_kmmio_probe(addr); |
49023168 | 269 | ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF)); |
0fd0e3da | 270 | ctx->addr = addr; |
8b7d89d0 PP |
271 | |
272 | if (ctx->probe && ctx->probe->pre_handler) | |
273 | ctx->probe->pre_handler(ctx->probe, regs, addr); | |
274 | ||
d61fc448 PP |
275 | /* |
276 | * Enable single-stepping and disable interrupts for the faulting | |
277 | * context. Local interrupts must not get enabled during stepping. | |
278 | */ | |
49023168 IM |
279 | regs->flags |= X86_EFLAGS_TF; |
280 | regs->flags &= ~X86_EFLAGS_IF; | |
8b7d89d0 | 281 | |
0fd0e3da | 282 | /* Now we set present bit in PTE and single step. */ |
5359b585 | 283 | disarm_kmmio_fault_page(ctx->fpage); |
8b7d89d0 | 284 | |
d61fc448 PP |
285 | /* |
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. | |
290 | */ | |
291 | ||
f5136380 | 292 | put_cpu_var(kmmio_ctx); |
13829537 | 293 | return 1; /* fault handled */ |
8b7d89d0 | 294 | |
0fd0e3da PP |
295 | no_kmmio_ctx: |
296 | put_cpu_var(kmmio_ctx); | |
8b7d89d0 | 297 | no_kmmio: |
0fd0e3da | 298 | rcu_read_unlock(); |
8b7d89d0 | 299 | preempt_enable_no_resched(); |
13829537 | 300 | return ret; |
8b7d89d0 PP |
301 | } |
302 | ||
303 | /* | |
304 | * Interrupts are disabled on entry as trap1 is an interrupt gate | |
305 | * and they remain disabled thorough out this function. | |
0fd0e3da | 306 | * This must always get called as the pair to kmmio_handler(). |
8b7d89d0 PP |
307 | */ |
308 | static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs) | |
309 | { | |
f5136380 PP |
310 | int ret = 0; |
311 | struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx); | |
8b7d89d0 | 312 | |
13829537 | 313 | if (!ctx->active) { |
0f9a623d SB |
314 | /* |
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 | |
318 | */ | |
319 | pr_warning("kmmio: unexpected debug trap on CPU %d.\n", | |
13829537 | 320 | smp_processor_id()); |
f5136380 | 321 | goto out; |
13829537 | 322 | } |
8b7d89d0 PP |
323 | |
324 | if (ctx->probe && ctx->probe->post_handler) | |
325 | ctx->probe->post_handler(ctx->probe, condition, regs); | |
326 | ||
340430c5 PP |
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); | |
8b7d89d0 | 332 | |
49023168 | 333 | regs->flags &= ~X86_EFLAGS_TF; |
8b7d89d0 PP |
334 | regs->flags |= ctx->saved_flags; |
335 | ||
336 | /* These were acquired in kmmio_handler(). */ | |
337 | ctx->active--; | |
0fd0e3da | 338 | BUG_ON(ctx->active); |
d61fc448 | 339 | rcu_read_unlock(); |
8b7d89d0 PP |
340 | preempt_enable_no_resched(); |
341 | ||
342 | /* | |
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. | |
346 | */ | |
49023168 | 347 | if (!(regs->flags & X86_EFLAGS_TF)) |
f5136380 | 348 | ret = 1; |
f5136380 PP |
349 | out: |
350 | put_cpu_var(kmmio_ctx); | |
351 | return ret; | |
8b7d89d0 PP |
352 | } |
353 | ||
0fd0e3da | 354 | /* You must be holding kmmio_lock. */ |
8b7d89d0 PP |
355 | static int add_kmmio_fault_page(unsigned long page) |
356 | { | |
357 | struct kmmio_fault_page *f; | |
358 | ||
359 | page &= PAGE_MASK; | |
360 | f = get_kmmio_fault_page(page); | |
361 | if (f) { | |
0fd0e3da | 362 | if (!f->count) |
5359b585 | 363 | arm_kmmio_fault_page(f); |
8b7d89d0 PP |
364 | f->count++; |
365 | return 0; | |
366 | } | |
367 | ||
5359b585 | 368 | f = kzalloc(sizeof(*f), GFP_ATOMIC); |
8b7d89d0 PP |
369 | if (!f) |
370 | return -1; | |
371 | ||
372 | f->count = 1; | |
373 | f->page = page; | |
8b7d89d0 | 374 | |
5359b585 | 375 | if (arm_kmmio_fault_page(f)) { |
e9d54cae SB |
376 | kfree(f); |
377 | return -1; | |
378 | } | |
379 | ||
380 | list_add_rcu(&f->list, kmmio_page_list(f->page)); | |
8b7d89d0 PP |
381 | |
382 | return 0; | |
383 | } | |
384 | ||
0fd0e3da PP |
385 | /* You must be holding kmmio_lock. */ |
386 | static void release_kmmio_fault_page(unsigned long page, | |
387 | struct kmmio_fault_page **release_list) | |
8b7d89d0 PP |
388 | { |
389 | struct kmmio_fault_page *f; | |
390 | ||
391 | page &= PAGE_MASK; | |
392 | f = get_kmmio_fault_page(page); | |
393 | if (!f) | |
394 | return; | |
395 | ||
396 | f->count--; | |
0fd0e3da | 397 | BUG_ON(f->count < 0); |
8b7d89d0 | 398 | if (!f->count) { |
5359b585 | 399 | disarm_kmmio_fault_page(f); |
0fd0e3da PP |
400 | f->release_next = *release_list; |
401 | *release_list = f; | |
8b7d89d0 PP |
402 | } |
403 | } | |
404 | ||
87e547fe PP |
405 | /* |
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 | |
410 | * mapping. | |
411 | */ | |
8b7d89d0 PP |
412 | int register_kmmio_probe(struct kmmio_probe *p) |
413 | { | |
d61fc448 | 414 | unsigned long flags; |
8b7d89d0 PP |
415 | int ret = 0; |
416 | unsigned long size = 0; | |
87e547fe | 417 | const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK); |
8b7d89d0 | 418 | |
d61fc448 | 419 | spin_lock_irqsave(&kmmio_lock, flags); |
8b7d89d0 PP |
420 | if (get_kmmio_probe(p->addr)) { |
421 | ret = -EEXIST; | |
422 | goto out; | |
423 | } | |
d61fc448 | 424 | kmmio_count++; |
0fd0e3da | 425 | list_add_rcu(&p->list, &kmmio_probes); |
87e547fe | 426 | while (size < size_lim) { |
8b7d89d0 | 427 | if (add_kmmio_fault_page(p->addr + size)) |
0fd0e3da | 428 | pr_err("kmmio: Unable to set page fault.\n"); |
8b7d89d0 PP |
429 | size += PAGE_SIZE; |
430 | } | |
8b7d89d0 | 431 | out: |
d61fc448 | 432 | spin_unlock_irqrestore(&kmmio_lock, flags); |
8b7d89d0 PP |
433 | /* |
434 | * XXX: What should I do here? | |
435 | * Here was a call to global_flush_tlb(), but it does not exist | |
0fd0e3da | 436 | * anymore. It seems it's not needed after all. |
8b7d89d0 PP |
437 | */ |
438 | return ret; | |
439 | } | |
0fd0e3da | 440 | EXPORT_SYMBOL(register_kmmio_probe); |
8b7d89d0 | 441 | |
0fd0e3da PP |
442 | static void rcu_free_kmmio_fault_pages(struct rcu_head *head) |
443 | { | |
444 | struct kmmio_delayed_release *dr = container_of( | |
445 | head, | |
446 | struct kmmio_delayed_release, | |
447 | rcu); | |
0492e1bb SB |
448 | struct kmmio_fault_page *f = dr->release_list; |
449 | while (f) { | |
450 | struct kmmio_fault_page *next = f->release_next; | |
451 | BUG_ON(f->count); | |
452 | kfree(f); | |
453 | f = next; | |
0fd0e3da PP |
454 | } |
455 | kfree(dr); | |
456 | } | |
457 | ||
458 | static void remove_kmmio_fault_pages(struct rcu_head *head) | |
459 | { | |
d0fc63f7 SB |
460 | struct kmmio_delayed_release *dr = |
461 | container_of(head, struct kmmio_delayed_release, rcu); | |
0492e1bb | 462 | struct kmmio_fault_page *f = dr->release_list; |
0fd0e3da PP |
463 | struct kmmio_fault_page **prevp = &dr->release_list; |
464 | unsigned long flags; | |
d0fc63f7 | 465 | |
0fd0e3da | 466 | spin_lock_irqsave(&kmmio_lock, flags); |
0492e1bb SB |
467 | while (f) { |
468 | if (!f->count) { | |
469 | list_del_rcu(&f->list); | |
470 | prevp = &f->release_next; | |
d0fc63f7 | 471 | } else { |
0492e1bb | 472 | *prevp = f->release_next; |
d0fc63f7 | 473 | } |
0492e1bb | 474 | f = f->release_next; |
0fd0e3da PP |
475 | } |
476 | spin_unlock_irqrestore(&kmmio_lock, flags); | |
d0fc63f7 | 477 | |
0fd0e3da PP |
478 | /* This is the real RCU destroy call. */ |
479 | call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages); | |
480 | } | |
481 | ||
482 | /* | |
483 | * Remove a kmmio probe. You have to synchronize_rcu() before you can be | |
d61fc448 PP |
484 | * sure that the callbacks will not be called anymore. Only after that |
485 | * you may actually release your struct kmmio_probe. | |
0fd0e3da PP |
486 | * |
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 | * Actally free the kmmio_fault_page structs as with RCU. | |
494 | */ | |
8b7d89d0 PP |
495 | void unregister_kmmio_probe(struct kmmio_probe *p) |
496 | { | |
d61fc448 | 497 | unsigned long flags; |
8b7d89d0 | 498 | unsigned long size = 0; |
87e547fe | 499 | const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK); |
0fd0e3da PP |
500 | struct kmmio_fault_page *release_list = NULL; |
501 | struct kmmio_delayed_release *drelease; | |
8b7d89d0 | 502 | |
d61fc448 | 503 | spin_lock_irqsave(&kmmio_lock, flags); |
87e547fe | 504 | while (size < size_lim) { |
0fd0e3da | 505 | release_kmmio_fault_page(p->addr + size, &release_list); |
8b7d89d0 PP |
506 | size += PAGE_SIZE; |
507 | } | |
0fd0e3da | 508 | list_del_rcu(&p->list); |
8b7d89d0 | 509 | kmmio_count--; |
d61fc448 | 510 | spin_unlock_irqrestore(&kmmio_lock, flags); |
8b7d89d0 | 511 | |
0fd0e3da PP |
512 | drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC); |
513 | if (!drelease) { | |
514 | pr_crit("kmmio: leaking kmmio_fault_page objects.\n"); | |
515 | return; | |
516 | } | |
517 | drelease->release_list = release_list; | |
518 | ||
519 | /* | |
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 | |
527 | * good choice. | |
528 | * | |
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. | |
532 | */ | |
533 | call_rcu(&drelease->rcu, remove_kmmio_fault_pages); | |
8b7d89d0 | 534 | } |
0fd0e3da | 535 | EXPORT_SYMBOL(unregister_kmmio_probe); |
8b7d89d0 | 536 | |
0f9a623d SB |
537 | static int |
538 | kmmio_die_notifier(struct notifier_block *nb, unsigned long val, void *args) | |
8b7d89d0 PP |
539 | { |
540 | struct die_args *arg = args; | |
541 | ||
13829537 | 542 | if (val == DIE_DEBUG && (arg->err & DR_STEP)) |
62edab90 P |
543 | if (post_kmmio_handler(arg->err, arg->regs) == 1) { |
544 | /* | |
545 | * Reset the BS bit in dr6 (pointed by args->err) to | |
546 | * denote completion of processing | |
547 | */ | |
548 | (*(unsigned long *)ERR_PTR(arg->err)) &= ~DR_STEP; | |
8b7d89d0 | 549 | return NOTIFY_STOP; |
62edab90 | 550 | } |
8b7d89d0 PP |
551 | |
552 | return NOTIFY_DONE; | |
553 | } | |
13829537 PP |
554 | |
555 | static struct notifier_block nb_die = { | |
556 | .notifier_call = kmmio_die_notifier | |
557 | }; | |
558 | ||
0f9a623d | 559 | int kmmio_init(void) |
13829537 PP |
560 | { |
561 | int i; | |
0f9a623d | 562 | |
13829537 PP |
563 | for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) |
564 | INIT_LIST_HEAD(&kmmio_page_table[i]); | |
0f9a623d | 565 | |
13829537 PP |
566 | return register_die_notifier(&nb_die); |
567 | } | |
0f9a623d SB |
568 | |
569 | void kmmio_cleanup(void) | |
570 | { | |
571 | int i; | |
572 | ||
573 | unregister_die_notifier(&nb_die); | |
574 | for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) { | |
575 | WARN_ONCE(!list_empty(&kmmio_page_table[i]), | |
576 | KERN_ERR "kmmio_page_table not empty at cleanup, any further tracing will leak memory.\n"); | |
577 | } | |
578 | } |