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 #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>
  21 #include <linux/io.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>
  27
  28 #define KMMIO_PAGE_HASH_BITS 4
  29 #define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS)
  30
  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         pteval_t old_presence; /* page presence prior to arming */
  36         bool armed;
  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.
  41          * Used only by writers (RCU) and post_kmmio_handler().
  42          * Protected by kmmio_lock, when linked into kmmio_page_table.
  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
  52 struct kmmio_context {
  53         struct kmmio_fault_page *fpage;
  54         struct kmmio_probe *probe;
  55         unsigned long saved_flags;
  56         unsigned long addr;
  57         int active;
  58 };
  59
  60 static DEFINE_SPINLOCK(kmmio_lock);
  61
  62 /* Protected by kmmio_lock */
  63 unsigned int kmmio_count;
  64
  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);
  68
  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
  74 /* Accessed per-cpu */
  75 static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
  76
  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  */
  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)
  87 {
  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))
  91                         return p;
  92         }
  93         return NULL;
  94 }
  95
  96 /* You must be holding RCU read lock. */
  97 static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page)
  98 {
  99         struct list_head *head;
 100         struct kmmio_fault_page *f;
 101
 102         page &= PAGE_MASK;
 103         head = kmmio_page_list(page);
 104         list_for_each_entry_rcu(f, head, list) {
 105                 if (f->page == page)
 106                         return f;
 107         }
 108         return NULL;
 109 }
 110
 111 static void clear_pmd_presence(pmd_t *pmd, bool clear, pmdval_t *old)
 112 {
 113         pmdval_t v = pmd_val(*pmd);
 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;
 119         set_pmd(pmd, __pmd(v));
 120 }
 121
 122 static void clear_pte_presence(pte_t *pte, bool clear, pteval_t *old)
 123 {
 124         pteval_t v = pte_val(*pte);
 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;
 130         set_pte_atomic(pte, __pte(v));
 131 }
 132
 133 static int clear_page_presence(struct kmmio_fault_page *f, bool clear)
 134 {
 135         unsigned int level;
 136         pte_t *pte = lookup_address(f->page, &level);
 137
 138         if (!pte) {
 139                 pr_err("kmmio: no pte for page 0x%08lx\n", f->page);
 140                 return -1;
 141         }
 142
 143         switch (level) {
 144         case PG_LEVEL_2M:
 145                 clear_pmd_presence((pmd_t *)pte, clear, &f->old_presence);
 146                 break;
 147         case PG_LEVEL_4K:
 148                 clear_pte_presence(pte, clear, &f->old_presence);
 149                 break;
 150         default:
 151                 pr_err("kmmio: unexpected page level 0x%x.\n", level);
 152                 return -1;
 153         }
 154
 155         __flush_tlb_one(f->page);
 156         return 0;
 157 }
 158
 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)
 171 {
 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",
 176                                         f->page, f->count, !!f->old_presence);
 177         }
 178         ret = clear_page_presence(f, true);
 179         WARN_ONCE(ret < 0, KERN_ERR "kmmio arming 0x%08lx failed.\n", f->page);
 180         f->armed = true;
 181         return ret;
 182 }
 183
 184 /** Restore the given page to saved presence state. */
 185 static void disarm_kmmio_fault_page(struct kmmio_fault_page *f)
 186 {
 187         int ret = clear_page_presence(f, false);
 188         WARN_ONCE(ret < 0,
 189                         KERN_ERR "kmmio disarming 0x%08lx failed.\n", f->page);
 190         f->armed = false;
 191 }
 192
 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  */
 204 /*
 205  * Interrupts are disabled on entry as trap3 is an interrupt gate
 206  * and they remain disabled thorough out this function.
 207  */
 208 int kmmio_handler(struct pt_regs *regs, unsigned long addr)
 209 {
 210         struct kmmio_context *ctx;
 211         struct kmmio_fault_page *faultpage;
 212         int ret = 0; /* default to fault not handled */
 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
 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.
 221          */
 222         preempt_disable();
 223         rcu_read_lock();
 224
 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
 230                  * our feet. The latter case should not be possible.
 231                  */
 232                 goto no_kmmio;
 233         }
 234
 235         ctx = &get_cpu_var(kmmio_ctx);
 236         if (ctx->active) {
 237                 if (addr == ctx->addr) {
 238                         /*
 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.
 242                          */
 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, "
 257                                         "for address 0x%08lx. Ignoring.\n",
 258                                         smp_processor_id(), addr);
 259                         pr_emerg("kmmio: previous hit was at 0x%08lx.\n",
 260                                                 ctx->addr);
 261                         disarm_kmmio_fault_page(faultpage);
 262                 }
 263                 goto no_kmmio_ctx;
 264         }
 265         ctx->active++;
 266
 267         ctx->fpage = faultpage;
 268         ctx->probe = get_kmmio_probe(addr);
 269         ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
 270         ctx->addr = addr;
 271
 272         if (ctx->probe && ctx->probe->pre_handler)
 273                 ctx->probe->pre_handler(ctx->probe, regs, addr);
 274
 275         /*
 276          * Enable single-stepping and disable interrupts for the faulting
 277          * context. Local interrupts must not get enabled during stepping.
 278          */
 279         regs->flags |= X86_EFLAGS_TF;
 280         regs->flags &= ~X86_EFLAGS_IF;
 281
 282         /* Now we set present bit in PTE and single step. */
 283         disarm_kmmio_fault_page(ctx->fpage);
 284
 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
 292         put_cpu_var(kmmio_ctx);
 293         return 1; /* fault handled */
 294
 295 no_kmmio_ctx:
 296         put_cpu_var(kmmio_ctx);
 297 no_kmmio:
 298         rcu_read_unlock();
 299         preempt_enable_no_resched();
 300         return ret;
 301 }
 302
 303 /*
 304  * Interrupts are disabled on entry as trap1 is an interrupt gate
 305  * and they remain disabled thorough out this function.
 306  * This must always get called as the pair to kmmio_handler().
 307  */
 308 static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
 309 {
 310         int ret = 0;
 311         struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
 312
 313         if (!ctx->active) {
 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",
 320                                                         smp_processor_id());
 321                 goto out;
 322         }
 323
 324         if (ctx->probe && ctx->probe->post_handler)
 325                 ctx->probe->post_handler(ctx->probe, condition, regs);
 326
 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);
 332
 333         regs->flags &= ~X86_EFLAGS_TF;
 334         regs->flags |= ctx->saved_flags;
 335
 336         /* These were acquired in kmmio_handler(). */
 337         ctx->active--;
 338         BUG_ON(ctx->active);
 339         rcu_read_unlock();
 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          */
 347         if (!(regs->flags & X86_EFLAGS_TF))
 348                 ret = 1;
 349 out:
 350         put_cpu_var(kmmio_ctx);
 351         return ret;
 352 }
 353
 354 /* You must be holding kmmio_lock. */
 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) {
 362                 if (!f->count)
 363                         arm_kmmio_fault_page(f);
 364                 f->count++;
 365                 return 0;
 366         }
 367
 368         f = kzalloc(sizeof(*f), GFP_ATOMIC);
 369         if (!f)
 370                 return -1;
 371
 372         f->count = 1;
 373         f->page = page;
 374
 375         if (arm_kmmio_fault_page(f)) {
 376                 kfree(f);
 377                 return -1;
 378         }
 379
 380         list_add_rcu(&f->list, kmmio_page_list(f->page));
 381
 382         return 0;
 383 }
 384
 385 /* You must be holding kmmio_lock. */
 386 static void release_kmmio_fault_page(unsigned long page,
 387                                 struct kmmio_fault_page **release_list)
 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--;
 397         BUG_ON(f->count < 0);
 398         if (!f->count) {
 399                 disarm_kmmio_fault_page(f);
 400                 f->release_next = *release_list;
 401                 *release_list = f;
 402         }
 403 }
 404
 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  */
 412 int register_kmmio_probe(struct kmmio_probe *p)
 413 {
 414         unsigned long flags;
 415         int ret = 0;
 416         unsigned long size = 0;
 417         const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
 418
 419         spin_lock_irqsave(&kmmio_lock, flags);
 420         if (get_kmmio_probe(p->addr)) {
 421                 ret = -EEXIST;
 422                 goto out;
 423         }
 424         kmmio_count++;
 425         list_add_rcu(&p->list, &kmmio_probes);
 426         while (size < size_lim) {
 427                 if (add_kmmio_fault_page(p->addr + size))
 428                         pr_err("kmmio: Unable to set page fault.\n");
 429                 size += PAGE_SIZE;
 430         }
 431 out:
 432         spin_unlock_irqrestore(&kmmio_lock, flags);
 433         /*
 434          * XXX: What should I do here?
 435          * Here was a call to global_flush_tlb(), but it does not exist
 436          * anymore. It seems it's not needed after all.
 437          */
 438         return ret;
 439 }
 440 EXPORT_SYMBOL(register_kmmio_probe);
 441
 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);
 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;
 454         }
 455         kfree(dr);
 456 }
 457
 458 static void remove_kmmio_fault_pages(struct rcu_head *head)
 459 {
 460         struct kmmio_delayed_release *dr =
 461                 container_of(head, struct kmmio_delayed_release, rcu);
 462         struct kmmio_fault_page *f = dr->release_list;
 463         struct kmmio_fault_page **prevp = &dr->release_list;
 464         unsigned long flags;
 465
 466         spin_lock_irqsave(&kmmio_lock, flags);
 467         while (f) {
 468                 if (!f->count) {
 469                         list_del_rcu(&f->list);
 470                         prevp = &f->release_next;
 471                 } else {
 472                         *prevp = f->release_next;
 473                 }
 474                 f = f->release_next;
 475         }
 476         spin_unlock_irqrestore(&kmmio_lock, flags);
 477
 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
 484  * sure that the callbacks will not be called anymore. Only after that
 485  * you may actually release your struct kmmio_probe.
 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  */
 495 void unregister_kmmio_probe(struct kmmio_probe *p)
 496 {
 497         unsigned long flags;
 498         unsigned long size = 0;
 499         const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
 500         struct kmmio_fault_page *release_list = NULL;
 501         struct kmmio_delayed_release *drelease;
 502
 503         spin_lock_irqsave(&kmmio_lock, flags);
 504         while (size < size_lim) {
 505                 release_kmmio_fault_page(p->addr + size, &release_list);
 506                 size += PAGE_SIZE;
 507         }
 508         list_del_rcu(&p->list);
 509         kmmio_count--;
 510         spin_unlock_irqrestore(&kmmio_lock, flags);
 511
 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);
 534 }
 535 EXPORT_SYMBOL(unregister_kmmio_probe);
 536
 537 static int
 538 kmmio_die_notifier(struct notifier_block *nb, unsigned long val, void *args)
 539 {
 540         struct die_args *arg = args;
 541
 542         if (val == DIE_DEBUG && (arg->err & DR_STEP))
 543                 if (post_kmmio_handler(arg->err, arg->regs) == 1)
 544                         return NOTIFY_STOP;
 545
 546         return NOTIFY_DONE;
 547 }
 548
 549 static struct notifier_block nb_die = {
 550         .notifier_call = kmmio_die_notifier
 551 };
 552
 553 int kmmio_init(void)
 554 {
 555         int i;
 556
 557         for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
 558                 INIT_LIST_HEAD(&kmmio_page_table[i]);
 559
 560         return register_die_notifier(&nb_die);
 561 }
 562
 563 void kmmio_cleanup(void)
 564 {
 565         int i;
 566
 567         unregister_die_notifier(&nb_die);
 568         for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) {
 569                 WARN_ONCE(!list_empty(&kmmio_page_table[i]),
 570                         KERN_ERR "kmmio_page_table not empty at cleanup, any further tracing will leak memory.\n");
 571         }
 572 }