arch/x86/mm/kmmio.c

   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/module.h>
  15 #include <linux/kernel.h>
  16 #include <linux/uaccess.h>
  17 #include <linux/ptrace.h>
  18 #include <linux/preempt.h>
  19 #include <linux/percpu.h>
  20 #include <linux/kdebug.h>
  21 #include <linux/mutex.h>
  22 #include <linux/io.h>
  23 #include <linux/slab.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>
  29
  30 #define KMMIO_PAGE_HASH_BITS 4
  31 #define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS)
  32
  33 struct kmmio_fault_page {
  34         struct list_head list;
  35         struct kmmio_fault_page *release_next;
  36         unsigned long addr; /* the requested address */
  37         pteval_t old_presence; /* page presence prior to arming */
  38         bool armed;
  39
  40         /*
  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.
  45          */
  46         int count;
  47
  48         bool scheduled_for_release;
  49 };
  50
  51 struct kmmio_delayed_release {
  52         struct rcu_head rcu;
  53         struct kmmio_fault_page *release_list;
  54 };
  55
  56 struct kmmio_context {
  57         struct kmmio_fault_page *fpage;
  58         struct kmmio_probe *probe;
  59         unsigned long saved_flags;
  60         unsigned long addr;
  61         int active;
  62 };
  63
  64 static DEFINE_SPINLOCK(kmmio_lock);
  65
  66 /* Protected by kmmio_lock */
  67 unsigned int kmmio_count;
  68
  69 /* Read-protected by RCU, write-protected by kmmio_lock. */
  70 static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
  71 static LIST_HEAD(kmmio_probes);
  72
  73 static struct list_head *kmmio_page_list(unsigned long addr)
  74 {
  75         unsigned int l;
  76         pte_t *pte = lookup_address(addr, &l);
  77
  78         if (!pte)
  79                 return NULL;
  80         addr &= page_level_mask(l);
  81
  82         return &kmmio_page_table[hash_long(addr, KMMIO_PAGE_HASH_BITS)];
  83 }
  84
  85 /* Accessed per-cpu */
  86 static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
  87
  88 /*
  89  * this is basically a dynamic stabbing problem:
  90  * Could use the existing prio tree code or
  91  * Possible better implementations:
  92  * The Interval Skip List: A Data Structure for Finding All Intervals That
  93  * Overlap a Point (might be simple)
  94  * Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup
  95  */
  96 /* Get the kmmio at this addr (if any). You must be holding RCU read lock. */
  97 static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
  98 {
  99         struct kmmio_probe *p;
 100         list_for_each_entry_rcu(p, &kmmio_probes, list) {
 101                 if (addr >= p->addr && addr < (p->addr + p->len))
 102                         return p;
 103         }
 104         return NULL;
 105 }
 106
 107 /* You must be holding RCU read lock. */
 108 static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long addr)
 109 {
 110         struct list_head *head;
 111         struct kmmio_fault_page *f;
 112         unsigned int l;
 113         pte_t *pte = lookup_address(addr, &l);
 114
 115         if (!pte)
 116                 return NULL;
 117         addr &= page_level_mask(l);
 118         head = kmmio_page_list(addr);
 119         list_for_each_entry_rcu(f, head, list) {
 120                 if (f->addr == addr)
 121                         return f;
 122         }
 123         return NULL;
 124 }
 125
 126 static void clear_pmd_presence(pmd_t *pmd, bool clear, pmdval_t *old)
 127 {
 128         pmdval_t v = pmd_val(*pmd);
 129         if (clear) {
 130                 *old = v & _PAGE_PRESENT;
 131                 v &= ~_PAGE_PRESENT;
 132         } else  /* presume this has been called with clear==true previously */
 133                 v |= *old;
 134         set_pmd(pmd, __pmd(v));
 135 }
 136
 137 static void clear_pte_presence(pte_t *pte, bool clear, pteval_t *old)
 138 {
 139         pteval_t v = pte_val(*pte);
 140         if (clear) {
 141                 *old = v & _PAGE_PRESENT;
 142                 v &= ~_PAGE_PRESENT;
 143         } else  /* presume this has been called with clear==true previously */
 144                 v |= *old;
 145         set_pte_atomic(pte, __pte(v));
 146 }
 147
 148 static int clear_page_presence(struct kmmio_fault_page *f, bool clear)
 149 {
 150         unsigned int level;
 151         pte_t *pte = lookup_address(f->addr, &level);
 152
 153         if (!pte) {
 154                 pr_err("no pte for addr 0x%08lx\n", f->addr);
 155                 return -1;
 156         }
 157
 158         switch (level) {
 159         case PG_LEVEL_2M:
 160                 clear_pmd_presence((pmd_t *)pte, clear, &f->old_presence);
 161                 break;
 162         case PG_LEVEL_4K:
 163                 clear_pte_presence(pte, clear, &f->old_presence);
 164                 break;
 165         default:
 166                 pr_err("unexpected page level 0x%x.\n", level);
 167                 return -1;
 168         }
 169
 170         __flush_tlb_one(f->addr);
 171         return 0;
 172 }
 173
 174 /*
 175  * Mark the given page as not present. Access to it will trigger a fault.
 176  *
 177  * Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the
 178  * protection is ignored here. RCU read lock is assumed held, so the struct
 179  * will not disappear unexpectedly. Furthermore, the caller must guarantee,
 180  * that double arming the same virtual address (page) cannot occur.
 181  *
 182  * Double disarming on the other hand is allowed, and may occur when a fault
 183  * and mmiotrace shutdown happen simultaneously.
 184  */
 185 static int arm_kmmio_fault_page(struct kmmio_fault_page *f)
 186 {
 187         int ret;
 188         WARN_ONCE(f->armed, KERN_ERR pr_fmt("kmmio page already armed.\n"));
 189         if (f->armed) {
 190                 pr_warning("double-arm: addr 0x%08lx, ref %d, old %d\n",
 191                            f->addr, f->count, !!f->old_presence);
 192         }
 193         ret = clear_page_presence(f, true);
 194         WARN_ONCE(ret < 0, KERN_ERR pr_fmt("arming at 0x%08lx failed.\n"),
 195                   f->addr);
 196         f->armed = true;
 197         return ret;
 198 }
 199
 200 /** Restore the given page to saved presence state. */
 201 static void disarm_kmmio_fault_page(struct kmmio_fault_page *f)
 202 {
 203         int ret = clear_page_presence(f, false);
 204         WARN_ONCE(ret < 0,
 205                         KERN_ERR "kmmio disarming at 0x%08lx failed.\n", f->addr);
 206         f->armed = false;
 207 }
 208
 209 /*
 210  * This is being called from do_page_fault().
 211  *
 212  * We may be in an interrupt or a critical section. Also prefecthing may
 213  * trigger a page fault. We may be in the middle of process switch.
 214  * We cannot take any locks, because we could be executing especially
 215  * within a kmmio critical section.
 216  *
 217  * Local interrupts are disabled, so preemption cannot happen.
 218  * Do not enable interrupts, do not sleep, and watch out for other CPUs.
 219  */
 220 /*
 221  * Interrupts are disabled on entry as trap3 is an interrupt gate
 222  * and they remain disabled throughout this function.
 223  */
 224 int kmmio_handler(struct pt_regs *regs, unsigned long addr)
 225 {
 226         struct kmmio_context *ctx;
 227         struct kmmio_fault_page *faultpage;
 228         int ret = 0; /* default to fault not handled */
 229         unsigned long page_base = addr;
 230         unsigned int l;
 231         pte_t *pte = lookup_address(addr, &l);
 232         if (!pte)
 233                 return -EINVAL;
 234         page_base &= page_level_mask(l);
 235
 236         /*
 237          * Preemption is now disabled to prevent process switch during
 238          * single stepping. We can only handle one active kmmio trace
 239          * per cpu, so ensure that we finish it before something else
 240          * gets to run. We also hold the RCU read lock over single
 241          * stepping to avoid looking up the probe and kmmio_fault_page
 242          * again.
 243          */
 244         preempt_disable();
 245         rcu_read_lock();
 246
 247         faultpage = get_kmmio_fault_page(page_base);
 248         if (!faultpage) {
 249                 /*
 250                  * Either this page fault is not caused by kmmio, or
 251                  * another CPU just pulled the kmmio probe from under
 252                  * our feet. The latter case should not be possible.
 253                  */
 254                 goto no_kmmio;
 255         }
 256
 257         ctx = &get_cpu_var(kmmio_ctx);
 258         if (ctx->active) {
 259                 if (page_base == ctx->addr) {
 260                         /*
 261                          * A second fault on the same page means some other
 262                          * condition needs handling by do_page_fault(), the
 263                          * page really not being present is the most common.
 264                          */
 265                         pr_debug("secondary hit for 0x%08lx CPU %d.\n",
 266                                  addr, smp_processor_id());
 267
 268                         if (!faultpage->old_presence)
 269                                 pr_info("unexpected secondary hit for address 0x%08lx on CPU %d.\n",
 270                                         addr, smp_processor_id());
 271                 } else {
 272                         /*
 273                          * Prevent overwriting already in-flight context.
 274                          * This should not happen, let's hope disarming at
 275                          * least prevents a panic.
 276                          */
 277                         pr_emerg("recursive probe hit on CPU %d, for address 0x%08lx. Ignoring.\n",
 278                                  smp_processor_id(), addr);
 279                         pr_emerg("previous hit was at 0x%08lx.\n", ctx->addr);
 280                         disarm_kmmio_fault_page(faultpage);
 281                 }
 282                 goto no_kmmio_ctx;
 283         }
 284         ctx->active++;
 285
 286         ctx->fpage = faultpage;
 287         ctx->probe = get_kmmio_probe(page_base);
 288         ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
 289         ctx->addr = page_base;
 290
 291         if (ctx->probe && ctx->probe->pre_handler)
 292                 ctx->probe->pre_handler(ctx->probe, regs, addr);
 293
 294         /*
 295          * Enable single-stepping and disable interrupts for the faulting
 296          * context. Local interrupts must not get enabled during stepping.
 297          */
 298         regs->flags |= X86_EFLAGS_TF;
 299         regs->flags &= ~X86_EFLAGS_IF;
 300
 301         /* Now we set present bit in PTE and single step. */
 302         disarm_kmmio_fault_page(ctx->fpage);
 303
 304         /*
 305          * If another cpu accesses the same page while we are stepping,
 306          * the access will not be caught. It will simply succeed and the
 307          * only downside is we lose the event. If this becomes a problem,
 308          * the user should drop to single cpu before tracing.
 309          */
 310
 311         put_cpu_var(kmmio_ctx);
 312         return 1; /* fault handled */
 313
 314 no_kmmio_ctx:
 315         put_cpu_var(kmmio_ctx);
 316 no_kmmio:
 317         rcu_read_unlock();
 318         preempt_enable_no_resched();
 319         return ret;
 320 }
 321
 322 /*
 323  * Interrupts are disabled on entry as trap1 is an interrupt gate
 324  * and they remain disabled throughout this function.
 325  * This must always get called as the pair to kmmio_handler().
 326  */
 327 static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
 328 {
 329         int ret = 0;
 330         struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
 331
 332         if (!ctx->active) {
 333                 /*
 334                  * debug traps without an active context are due to either
 335                  * something external causing them (f.e. using a debugger while
 336                  * mmio tracing enabled), or erroneous behaviour
 337                  */
 338                 pr_warning("unexpected debug trap on CPU %d.\n",
 339                            smp_processor_id());
 340                 goto out;
 341         }
 342
 343         if (ctx->probe && ctx->probe->post_handler)
 344                 ctx->probe->post_handler(ctx->probe, condition, regs);
 345
 346         /* Prevent racing against release_kmmio_fault_page(). */
 347         spin_lock(&kmmio_lock);
 348         if (ctx->fpage->count)
 349                 arm_kmmio_fault_page(ctx->fpage);
 350         spin_unlock(&kmmio_lock);
 351
 352         regs->flags &= ~X86_EFLAGS_TF;
 353         regs->flags |= ctx->saved_flags;
 354
 355         /* These were acquired in kmmio_handler(). */
 356         ctx->active--;
 357         BUG_ON(ctx->active);
 358         rcu_read_unlock();
 359         preempt_enable_no_resched();
 360
 361         /*
 362          * if somebody else is singlestepping across a probe point, flags
 363          * will have TF set, in which case, continue the remaining processing
 364          * of do_debug, as if this is not a probe hit.
 365          */
 366         if (!(regs->flags & X86_EFLAGS_TF))
 367                 ret = 1;
 368 out:
 369         put_cpu_var(kmmio_ctx);
 370         return ret;
 371 }
 372
 373 /* You must be holding kmmio_lock. */
 374 static int add_kmmio_fault_page(unsigned long addr)
 375 {
 376         struct kmmio_fault_page *f;
 377
 378         f = get_kmmio_fault_page(addr);
 379         if (f) {
 380                 if (!f->count)
 381                         arm_kmmio_fault_page(f);
 382                 f->count++;
 383                 return 0;
 384         }
 385
 386         f = kzalloc(sizeof(*f), GFP_ATOMIC);
 387         if (!f)
 388                 return -1;
 389
 390         f->count = 1;
 391         f->addr = addr;
 392
 393         if (arm_kmmio_fault_page(f)) {
 394                 kfree(f);
 395                 return -1;
 396         }
 397
 398         list_add_rcu(&f->list, kmmio_page_list(f->addr));
 399
 400         return 0;
 401 }
 402
 403 /* You must be holding kmmio_lock. */
 404 static void release_kmmio_fault_page(unsigned long addr,
 405                                 struct kmmio_fault_page **release_list)
 406 {
 407         struct kmmio_fault_page *f;
 408
 409         f = get_kmmio_fault_page(addr);
 410         if (!f)
 411                 return;
 412
 413         f->count--;
 414         BUG_ON(f->count < 0);
 415         if (!f->count) {
 416                 disarm_kmmio_fault_page(f);
 417                 if (!f->scheduled_for_release) {
 418                         f->release_next = *release_list;
 419                         *release_list = f;
 420                         f->scheduled_for_release = true;
 421                 }
 422         }
 423 }
 424
 425 /*
 426  * With page-unaligned ioremaps, one or two armed pages may contain
 427  * addresses from outside the intended mapping. Events for these addresses
 428  * are currently silently dropped. The events may result only from programming
 429  * mistakes by accessing addresses before the beginning or past the end of a
 430  * mapping.
 431  */
 432 int register_kmmio_probe(struct kmmio_probe *p)
 433 {
 434         unsigned long flags;
 435         int ret = 0;
 436         unsigned long size = 0;
 437         const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
 438         unsigned int l;
 439         pte_t *pte;
 440
 441         spin_lock_irqsave(&kmmio_lock, flags);
 442         if (get_kmmio_probe(p->addr)) {
 443                 ret = -EEXIST;
 444                 goto out;
 445         }
 446
 447         pte = lookup_address(p->addr, &l);
 448         if (!pte) {
 449                 ret = -EINVAL;
 450                 goto out;
 451         }
 452
 453         kmmio_count++;
 454         list_add_rcu(&p->list, &kmmio_probes);
 455         while (size < size_lim) {
 456                 if (add_kmmio_fault_page(p->addr + size))
 457                         pr_err("Unable to set page fault.\n");
 458                 size += page_level_size(l);
 459         }
 460 out:
 461         spin_unlock_irqrestore(&kmmio_lock, flags);
 462         /*
 463          * XXX: What should I do here?
 464          * Here was a call to global_flush_tlb(), but it does not exist
 465          * anymore. It seems it's not needed after all.
 466          */
 467         return ret;
 468 }
 469 EXPORT_SYMBOL(register_kmmio_probe);
 470
 471 static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
 472 {
 473         struct kmmio_delayed_release *dr = container_of(
 474                                                 head,
 475                                                 struct kmmio_delayed_release,
 476                                                 rcu);
 477         struct kmmio_fault_page *f = dr->release_list;
 478         while (f) {
 479                 struct kmmio_fault_page *next = f->release_next;
 480                 BUG_ON(f->count);
 481                 kfree(f);
 482                 f = next;
 483         }
 484         kfree(dr);
 485 }
 486
 487 static void remove_kmmio_fault_pages(struct rcu_head *head)
 488 {
 489         struct kmmio_delayed_release *dr =
 490                 container_of(head, struct kmmio_delayed_release, rcu);
 491         struct kmmio_fault_page *f = dr->release_list;
 492         struct kmmio_fault_page **prevp = &dr->release_list;
 493         unsigned long flags;
 494
 495         spin_lock_irqsave(&kmmio_lock, flags);
 496         while (f) {
 497                 if (!f->count) {
 498                         list_del_rcu(&f->list);
 499                         prevp = &f->release_next;
 500                 } else {
 501                         *prevp = f->release_next;
 502                         f->release_next = NULL;
 503                         f->scheduled_for_release = false;
 504                 }
 505                 f = *prevp;
 506         }
 507         spin_unlock_irqrestore(&kmmio_lock, flags);
 508
 509         /* This is the real RCU destroy call. */
 510         call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
 511 }
 512
 513 /*
 514  * Remove a kmmio probe. You have to synchronize_rcu() before you can be
 515  * sure that the callbacks will not be called anymore. Only after that
 516  * you may actually release your struct kmmio_probe.
 517  *
 518  * Unregistering a kmmio fault page has three steps:
 519  * 1. release_kmmio_fault_page()
 520  *    Disarm the page, wait a grace period to let all faults finish.
 521  * 2. remove_kmmio_fault_pages()
 522  *    Remove the pages from kmmio_page_table.
 523  * 3. rcu_free_kmmio_fault_pages()
 524  *    Actually free the kmmio_fault_page structs as with RCU.
 525  */
 526 void unregister_kmmio_probe(struct kmmio_probe *p)
 527 {
 528         unsigned long flags;
 529         unsigned long size = 0;
 530         const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
 531         struct kmmio_fault_page *release_list = NULL;
 532         struct kmmio_delayed_release *drelease;
 533         unsigned int l;
 534         pte_t *pte;
 535
 536         pte = lookup_address(p->addr, &l);
 537         if (!pte)
 538                 return;
 539
 540         spin_lock_irqsave(&kmmio_lock, flags);
 541         while (size < size_lim) {
 542                 release_kmmio_fault_page(p->addr + size, &release_list);
 543                 size += page_level_size(l);
 544         }
 545         list_del_rcu(&p->list);
 546         kmmio_count--;
 547         spin_unlock_irqrestore(&kmmio_lock, flags);
 548
 549         if (!release_list)
 550                 return;
 551
 552         drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
 553         if (!drelease) {
 554                 pr_crit("leaking kmmio_fault_page objects.\n");
 555                 return;
 556         }
 557         drelease->release_list = release_list;
 558
 559         /*
 560          * This is not really RCU here. We have just disarmed a set of
 561          * pages so that they cannot trigger page faults anymore. However,
 562          * we cannot remove the pages from kmmio_page_table,
 563          * because a probe hit might be in flight on another CPU. The
 564          * pages are collected into a list, and they will be removed from
 565          * kmmio_page_table when it is certain that no probe hit related to
 566          * these pages can be in flight. RCU grace period sounds like a
 567          * good choice.
 568          *
 569          * If we removed the pages too early, kmmio page fault handler might
 570          * not find the respective kmmio_fault_page and determine it's not
 571          * a kmmio fault, when it actually is. This would lead to madness.
 572          */
 573         call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
 574 }
 575 EXPORT_SYMBOL(unregister_kmmio_probe);
 576
 577 static int
 578 kmmio_die_notifier(struct notifier_block *nb, unsigned long val, void *args)
 579 {
 580         struct die_args *arg = args;
 581         unsigned long* dr6_p = (unsigned long *)ERR_PTR(arg->err);
 582
 583         if (val == DIE_DEBUG && (*dr6_p & DR_STEP))
 584                 if (post_kmmio_handler(*dr6_p, arg->regs) == 1) {
 585                         /*
 586                          * Reset the BS bit in dr6 (pointed by args->err) to
 587                          * denote completion of processing
 588                          */
 589                         *dr6_p &= ~DR_STEP;
 590                         return NOTIFY_STOP;
 591                 }
 592
 593         return NOTIFY_DONE;
 594 }
 595
 596 static struct notifier_block nb_die = {
 597         .notifier_call = kmmio_die_notifier
 598 };
 599
 600 int kmmio_init(void)
 601 {
 602         int i;
 603
 604         for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
 605                 INIT_LIST_HEAD(&kmmio_page_table[i]);
 606
 607         return register_die_notifier(&nb_die);
 608 }
 609
 610 void kmmio_cleanup(void)
 611 {
 612         int i;
 613
 614         unregister_die_notifier(&nb_die);
 615         for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) {
 616                 WARN_ONCE(!list_empty(&kmmio_page_table[i]),
 617                         KERN_ERR "kmmio_page_table not empty at cleanup, any further tracing will leak memory.\n");
 618         }
 619 }