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 8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 9 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> 23#include <linux/io.h> 24#include <linux/slab.h> 25#include <asm/cacheflush.h> 26#include <asm/tlbflush.h> 27#include <linux/errno.h> 28#include <asm/debugreg.h> 29#include <linux/mmiotrace.h> 30 31#define KMMIO_PAGE_HASH_BITS 4 32#define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS) 33 34struct kmmio_fault_page { 35 struct list_head list; 36 struct kmmio_fault_page *release_next; 37 unsigned long page; /* location of the fault page */ 38 pteval_t old_presence; /* page presence prior to arming */ 39 bool armed; 40 41 /* 42 * Number of times this page has been registered as a part 43 * of a probe. If zero, page is disarmed and this may be freed. 44 * Used only by writers (RCU) and post_kmmio_handler(). 45 * Protected by kmmio_lock, when linked into kmmio_page_table. 46 */ 47 int count; 48 49 bool scheduled_for_release; 50}; 51 52struct kmmio_delayed_release { 53 struct rcu_head rcu; 54 struct kmmio_fault_page *release_list; 55}; 56 57struct kmmio_context { 58 struct kmmio_fault_page *fpage; 59 struct kmmio_probe *probe; 60 unsigned long saved_flags; 61 unsigned long addr; 62 int active; 63}; 64 65static DEFINE_SPINLOCK(kmmio_lock); 66 67/* Protected by kmmio_lock */ 68unsigned int kmmio_count; 69 70/* Read-protected by RCU, write-protected by kmmio_lock. */ 71static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE]; 72static LIST_HEAD(kmmio_probes); 73 74static struct list_head *kmmio_page_list(unsigned long page) 75{ 76 return &kmmio_page_table[hash_long(page, KMMIO_PAGE_HASH_BITS)]; 77} 78 79/* Accessed per-cpu */ 80static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx); 81 82/* 83 * this is basically a dynamic stabbing problem: 84 * Could use the existing prio tree code or 85 * Possible better implementations: 86 * The Interval Skip List: A Data Structure for Finding All Intervals That 87 * Overlap a Point (might be simple) 88 * Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup 89 */ 90/* Get the kmmio at this addr (if any). You must be holding RCU read lock. */ 91static struct kmmio_probe *get_kmmio_probe(unsigned long addr) 92{ 93 struct kmmio_probe *p; 94 list_for_each_entry_rcu(p, &kmmio_probes, list) { 95 if (addr >= p->addr && addr < (p->addr + p->len)) 96 return p; 97 } 98 return NULL; 99} 100 101/* You must be holding RCU read lock. */ 102static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page) 103{ 104 struct list_head *head; 105 struct kmmio_fault_page *f; 106 107 page &= PAGE_MASK; 108 head = kmmio_page_list(page); 109 list_for_each_entry_rcu(f, head, list) { 110 if (f->page == page) 111 return f; 112 } 113 return NULL; 114} 115 116static void clear_pmd_presence(pmd_t *pmd, bool clear, pmdval_t *old) 117{ 118 pmdval_t v = pmd_val(*pmd); 119 if (clear) { 120 *old = v & _PAGE_PRESENT; 121 v &= ~_PAGE_PRESENT; 122 } else /* presume this has been called with clear==true previously */ 123 v |= *old; 124 set_pmd(pmd, __pmd(v)); 125} 126 127static void clear_pte_presence(pte_t *pte, bool clear, pteval_t *old) 128{ 129 pteval_t v = pte_val(*pte); 130 if (clear) { 131 *old = v & _PAGE_PRESENT; 132 v &= ~_PAGE_PRESENT; 133 } else /* presume this has been called with clear==true previously */ 134 v |= *old; 135 set_pte_atomic(pte, __pte(v)); 136} 137 138static int clear_page_presence(struct kmmio_fault_page *f, bool clear) 139{ 140 unsigned int level; 141 pte_t *pte = lookup_address(f->page, &level); 142 143 if (!pte) { 144 pr_err("no pte for page 0x%08lx\n", f->page); 145 return -1; 146 } 147 148 switch (level) { 149 case PG_LEVEL_2M: 150 clear_pmd_presence((pmd_t *)pte, clear, &f->old_presence); 151 break; 152 case PG_LEVEL_4K: 153 clear_pte_presence(pte, clear, &f->old_presence); 154 break; 155 default: 156 pr_err("unexpected page level 0x%x.\n", level); 157 return -1; 158 } 159 160 __flush_tlb_one(f->page); 161 return 0; 162} 163 164/* 165 * Mark the given page as not present. Access to it will trigger a fault. 166 * 167 * Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the 168 * protection is ignored here. RCU read lock is assumed held, so the struct 169 * will not disappear unexpectedly. Furthermore, the caller must guarantee, 170 * that double arming the same virtual address (page) cannot occur. 171 * 172 * Double disarming on the other hand is allowed, and may occur when a fault 173 * and mmiotrace shutdown happen simultaneously. 174 */ 175static int arm_kmmio_fault_page(struct kmmio_fault_page *f) 176{ 177 int ret; 178 WARN_ONCE(f->armed, KERN_ERR pr_fmt("kmmio page already armed.\n")); 179 if (f->armed) { 180 pr_warning("double-arm: page 0x%08lx, ref %d, old %d\n", 181 f->page, f->count, !!f->old_presence); 182 } 183 ret = clear_page_presence(f, true); 184 WARN_ONCE(ret < 0, KERN_ERR pr_fmt("arming 0x%08lx failed.\n"), 185 f->page); 186 f->armed = true; 187 return ret; 188} 189 190/** Restore the given page to saved presence state. */ 191static void disarm_kmmio_fault_page(struct kmmio_fault_page *f) 192{ 193 int ret = clear_page_presence(f, false); 194 WARN_ONCE(ret < 0, 195 KERN_ERR "kmmio disarming 0x%08lx failed.\n", f->page); 196 f->armed = false; 197} 198 199/* 200 * This is being called from do_page_fault(). 201 * 202 * We may be in an interrupt or a critical section. Also prefecthing may 203 * trigger a page fault. We may be in the middle of process switch. 204 * We cannot take any locks, because we could be executing especially 205 * within a kmmio critical section. 206 * 207 * Local interrupts are disabled, so preemption cannot happen. 208 * Do not enable interrupts, do not sleep, and watch out for other CPUs. 209 */ 210/* 211 * Interrupts are disabled on entry as trap3 is an interrupt gate 212 * and they remain disabled throughout this function. 213 */ 214int kmmio_handler(struct pt_regs *regs, unsigned long addr) 215{ 216 struct kmmio_context *ctx; 217 struct kmmio_fault_page *faultpage; 218 int ret = 0; /* default to fault not handled */ 219 220 /* 221 * Preemption is now disabled to prevent process switch during 222 * single stepping. We can only handle one active kmmio trace 223 * per cpu, so ensure that we finish it before something else 224 * gets to run. We also hold the RCU read lock over single 225 * stepping to avoid looking up the probe and kmmio_fault_page 226 * again. 227 */ 228 preempt_disable(); 229 rcu_read_lock(); 230 231 faultpage = get_kmmio_fault_page(addr); 232 if (!faultpage) { 233 /* 234 * Either this page fault is not caused by kmmio, or 235 * another CPU just pulled the kmmio probe from under 236 * our feet. The latter case should not be possible. 237 */ 238 goto no_kmmio; 239 } 240 241 ctx = &get_cpu_var(kmmio_ctx); 242 if (ctx->active) { 243 if (addr == ctx->addr) { 244 /* 245 * A second fault on the same page means some other 246 * condition needs handling by do_page_fault(), the 247 * page really not being present is the most common. 248 */ 249 pr_debug("secondary hit for 0x%08lx CPU %d.\n", 250 addr, smp_processor_id()); 251 252 if (!faultpage->old_presence) 253 pr_info("unexpected secondary hit for address 0x%08lx on CPU %d.\n", 254 addr, smp_processor_id()); 255 } else { 256 /* 257 * Prevent overwriting already in-flight context. 258 * This should not happen, let's hope disarming at 259 * least prevents a panic. 260 */ 261 pr_emerg("recursive probe hit on CPU %d, for address 0x%08lx. Ignoring.\n", 262 smp_processor_id(), addr); 263 pr_emerg("previous hit was at 0x%08lx.\n", ctx->addr); 264 disarm_kmmio_fault_page(faultpage); 265 } 266 goto no_kmmio_ctx; 267 } 268 ctx->active++; 269 270 ctx->fpage = faultpage; 271 ctx->probe = get_kmmio_probe(addr); 272 ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF)); 273 ctx->addr = addr; 274 275 if (ctx->probe && ctx->probe->pre_handler) 276 ctx->probe->pre_handler(ctx->probe, regs, addr); 277 278 /* 279 * Enable single-stepping and disable interrupts for the faulting 280 * context. Local interrupts must not get enabled during stepping. 281 */ 282 regs->flags |= X86_EFLAGS_TF; 283 regs->flags &= ~X86_EFLAGS_IF; 284 285 /* Now we set present bit in PTE and single step. */ 286 disarm_kmmio_fault_page(ctx->fpage); 287 288 /* 289 * If another cpu accesses the same page while we are stepping, 290 * the access will not be caught. It will simply succeed and the 291 * only downside is we lose the event. If this becomes a problem, 292 * the user should drop to single cpu before tracing. 293 */ 294 295 put_cpu_var(kmmio_ctx); 296 return 1; /* fault handled */ 297 298no_kmmio_ctx: 299 put_cpu_var(kmmio_ctx); 300no_kmmio: 301 rcu_read_unlock(); 302 preempt_enable_no_resched(); 303 return ret; 304} 305 306/* 307 * Interrupts are disabled on entry as trap1 is an interrupt gate 308 * and they remain disabled throughout this function. 309 * This must always get called as the pair to kmmio_handler(). 310 */ 311static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs) 312{ 313 int ret = 0; 314 struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx); 315 316 if (!ctx->active) { 317 /* 318 * debug traps without an active context are due to either 319 * something external causing them (f.e. using a debugger while 320 * mmio tracing enabled), or erroneous behaviour 321 */ 322 pr_warning("unexpected debug trap on CPU %d.\n", 323 smp_processor_id()); 324 goto out; 325 } 326 327 if (ctx->probe && ctx->probe->post_handler) 328 ctx->probe->post_handler(ctx->probe, condition, regs); 329 330 /* Prevent racing against release_kmmio_fault_page(). */ 331 spin_lock(&kmmio_lock); 332 if (ctx->fpage->count) 333 arm_kmmio_fault_page(ctx->fpage); 334 spin_unlock(&kmmio_lock); 335 336 regs->flags &= ~X86_EFLAGS_TF; 337 regs->flags |= ctx->saved_flags; 338 339 /* These were acquired in kmmio_handler(). */ 340 ctx->active--; 341 BUG_ON(ctx->active); 342 rcu_read_unlock(); 343 preempt_enable_no_resched(); 344 345 /* 346 * if somebody else is singlestepping across a probe point, flags 347 * will have TF set, in which case, continue the remaining processing 348 * of do_debug, as if this is not a probe hit. 349 */ 350 if (!(regs->flags & X86_EFLAGS_TF)) 351 ret = 1; 352out: 353 put_cpu_var(kmmio_ctx); 354 return ret; 355} 356 357/* You must be holding kmmio_lock. */ 358static int add_kmmio_fault_page(unsigned long page) 359{ 360 struct kmmio_fault_page *f; 361 362 page &= PAGE_MASK; 363 f = get_kmmio_fault_page(page); 364 if (f) { 365 if (!f->count) 366 arm_kmmio_fault_page(f); 367 f->count++; 368 return 0; 369 } 370 371 f = kzalloc(sizeof(*f), GFP_ATOMIC); 372 if (!f) 373 return -1; 374 375 f->count = 1; 376 f->page = page; 377 378 if (arm_kmmio_fault_page(f)) { 379 kfree(f); 380 return -1; 381 } 382 383 list_add_rcu(&f->list, kmmio_page_list(f->page)); 384 385 return 0; 386} 387 388/* You must be holding kmmio_lock. */ 389static void release_kmmio_fault_page(unsigned long page, 390 struct kmmio_fault_page **release_list) 391{ 392 struct kmmio_fault_page *f; 393 394 page &= PAGE_MASK; 395 f = get_kmmio_fault_page(page); 396 if (!f) 397 return; 398 399 f->count--; 400 BUG_ON(f->count < 0); 401 if (!f->count) { 402 disarm_kmmio_fault_page(f); 403 if (!f->scheduled_for_release) { 404 f->release_next = *release_list; 405 *release_list = f; 406 f->scheduled_for_release = true; 407 } 408 } 409} 410 411/* 412 * With page-unaligned ioremaps, one or two armed pages may contain 413 * addresses from outside the intended mapping. Events for these addresses 414 * are currently silently dropped. The events may result only from programming 415 * mistakes by accessing addresses before the beginning or past the end of a 416 * mapping. 417 */ 418int register_kmmio_probe(struct kmmio_probe *p) 419{ 420 unsigned long flags; 421 int ret = 0; 422 unsigned long size = 0; 423 const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK); 424 425 spin_lock_irqsave(&kmmio_lock, flags); 426 if (get_kmmio_probe(p->addr)) { 427 ret = -EEXIST; 428 goto out; 429 } 430 kmmio_count++; 431 list_add_rcu(&p->list, &kmmio_probes); 432 while (size < size_lim) { 433 if (add_kmmio_fault_page(p->addr + size)) 434 pr_err("Unable to set page fault.\n"); 435 size += PAGE_SIZE; 436 } 437out: 438 spin_unlock_irqrestore(&kmmio_lock, flags); 439 return ret; 440} 441EXPORT_SYMBOL(register_kmmio_probe); 442 443static void rcu_free_kmmio_fault_pages(struct rcu_head *head) 444{ 445 struct kmmio_delayed_release *dr = container_of( 446 head, 447 struct kmmio_delayed_release, 448 rcu); 449 struct kmmio_fault_page *f = dr->release_list; 450 while (f) { 451 struct kmmio_fault_page *next = f->release_next; 452 BUG_ON(f->count); 453 kfree(f); 454 f = next; 455 } 456 kfree(dr); 457} 458 459static void remove_kmmio_fault_pages(struct rcu_head *head) 460{ 461 struct kmmio_delayed_release *dr = 462 container_of(head, struct kmmio_delayed_release, rcu); 463 struct kmmio_fault_page *f = dr->release_list; 464 struct kmmio_fault_page **prevp = &dr->release_list; 465 unsigned long flags; 466 467 spin_lock_irqsave(&kmmio_lock, flags); 468 while (f) { 469 if (!f->count) { 470 list_del_rcu(&f->list); 471 prevp = &f->release_next; 472 } else { 473 *prevp = f->release_next; 474 f->release_next = NULL; 475 f->scheduled_for_release = false; 476 } 477 f = *prevp; 478 } 479 spin_unlock_irqrestore(&kmmio_lock, flags); 480 481 /* This is the real RCU destroy call. */ 482 call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages); 483} 484 485/* 486 * Remove a kmmio probe. You have to synchronize_rcu() before you can be 487 * sure that the callbacks will not be called anymore. Only after that 488 * you may actually release your struct kmmio_probe. 489 * 490 * Unregistering a kmmio fault page has three steps: 491 * 1. release_kmmio_fault_page() 492 * Disarm the page, wait a grace period to let all faults finish. 493 * 2. remove_kmmio_fault_pages() 494 * Remove the pages from kmmio_page_table. 495 * 3. rcu_free_kmmio_fault_pages() 496 * Actually free the kmmio_fault_page structs as with RCU. 497 */ 498void unregister_kmmio_probe(struct kmmio_probe *p) 499{ 500 unsigned long flags; 501 unsigned long size = 0; 502 const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK); 503 struct kmmio_fault_page *release_list = NULL; 504 struct kmmio_delayed_release *drelease; 505 506 spin_lock_irqsave(&kmmio_lock, flags); 507 while (size < size_lim) { 508 release_kmmio_fault_page(p->addr + size, &release_list); 509 size += PAGE_SIZE; 510 } 511 list_del_rcu(&p->list); 512 kmmio_count--; 513 spin_unlock_irqrestore(&kmmio_lock, flags); 514 515 if (!release_list) 516 return; 517 518 drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC); 519 if (!drelease) { 520 pr_crit("leaking kmmio_fault_page objects.\n"); 521 return; 522 } 523 drelease->release_list = release_list; 524 525 /* 526 * This is not really RCU here. We have just disarmed a set of 527 * pages so that they cannot trigger page faults anymore. However, 528 * we cannot remove the pages from kmmio_page_table, 529 * because a probe hit might be in flight on another CPU. The 530 * pages are collected into a list, and they will be removed from 531 * kmmio_page_table when it is certain that no probe hit related to 532 * these pages can be in flight. RCU grace period sounds like a 533 * good choice. 534 * 535 * If we removed the pages too early, kmmio page fault handler might 536 * not find the respective kmmio_fault_page and determine it's not 537 * a kmmio fault, when it actually is. This would lead to madness. 538 */ 539 call_rcu(&drelease->rcu, remove_kmmio_fault_pages); 540} 541EXPORT_SYMBOL(unregister_kmmio_probe); 542 543static int 544kmmio_die_notifier(struct notifier_block *nb, unsigned long val, void *args) 545{ 546 struct die_args *arg = args; 547 unsigned long* dr6_p = (unsigned long *)ERR_PTR(arg->err); 548 549 if (val == DIE_DEBUG && (*dr6_p & DR_STEP)) 550 if (post_kmmio_handler(*dr6_p, arg->regs) == 1) { 551 /* 552 * Reset the BS bit in dr6 (pointed by args->err) to 553 * denote completion of processing 554 */ 555 *dr6_p &= ~DR_STEP; 556 return NOTIFY_STOP; 557 } 558 559 return NOTIFY_DONE; 560} 561 562static struct notifier_block nb_die = { 563 .notifier_call = kmmio_die_notifier 564}; 565 566int kmmio_init(void) 567{ 568 int i; 569 570 for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) 571 INIT_LIST_HEAD(&kmmio_page_table[i]); 572 573 return register_die_notifier(&nb_die); 574} 575 576void kmmio_cleanup(void) 577{ 578 int i; 579 580 unregister_die_notifier(&nb_die); 581 for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) { 582 WARN_ONCE(!list_empty(&kmmio_page_table[i]), 583 KERN_ERR "kmmio_page_table not empty at cleanup, any further tracing will leak memory.\n"); 584 } 585} 586