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         bool 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 *p;
 101
 102         page &= PAGE_MASK;
 103         head = kmmio_page_list(page);
 104         list_for_each_entry_rcu(p, head, list) {
 105                 if (p->page == page)
 106                         return p;
 107         }
 108         return NULL;
 109 }
 110
 111 static void set_pmd_presence(pmd_t *pmd, bool present, bool *old)
 112 {
 113         pmdval_t v = pmd_val(*pmd);
 114         *old = !!(v & _PAGE_PRESENT);
 115         v &= ~_PAGE_PRESENT;
 116         if (present)
 117                 v |= _PAGE_PRESENT;
 118         set_pmd(pmd, __pmd(v));
 119 }
 120
 121 static void set_pte_presence(pte_t *pte, bool present, bool *old)
 122 {
 123         pteval_t v = pte_val(*pte);
 124         *old = !!(v & _PAGE_PRESENT);
 125         v &= ~_PAGE_PRESENT;
 126         if (present)
 127                 v |= _PAGE_PRESENT;
 128         set_pte_atomic(pte, __pte(v));
 129 }
 130
 131 static int set_page_presence(unsigned long addr, bool present, bool *old)
 132 {
 133         unsigned int level;
 134         pte_t *pte = lookup_address(addr, &level);
 135
 136         if (!pte) {
 137                 pr_err("kmmio: no pte for page 0x%08lx\n", addr);
 138                 return -1;
 139         }
 140
 141         switch (level) {
 142         case PG_LEVEL_2M:
 143                 set_pmd_presence((pmd_t *)pte, present, old);
 144                 break;
 145         case PG_LEVEL_4K:
 146                 set_pte_presence(pte, present, old);
 147                 break;
 148         default:
 149                 pr_err("kmmio: unexpected page level 0x%x.\n", level);
 150                 return -1;
 151         }
 152
 153         __flush_tlb_one(addr);
 154         return 0;
 155 }
 156
 157 /*
 158  * Mark the given page as not present. Access to it will trigger a fault.
 159  *
 160  * Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the
 161  * protection is ignored here. RCU read lock is assumed held, so the struct
 162  * will not disappear unexpectedly. Furthermore, the caller must guarantee,
 163  * that double arming the same virtual address (page) cannot occur.
 164  *
 165  * Double disarming on the other hand is allowed, and may occur when a fault
 166  * and mmiotrace shutdown happen simultaneously.
 167  */
 168 static int arm_kmmio_fault_page(struct kmmio_fault_page *f)
 169 {
 170         int ret;
 171         WARN_ONCE(f->armed, KERN_ERR "kmmio page already armed.\n");
 172         if (f->armed) {
 173                 pr_warning("kmmio double-arm: page 0x%08lx, ref %d, old %d\n",
 174                                         f->page, f->count, f->old_presence);
 175         }
 176         ret = set_page_presence(f->page, false, &f->old_presence);
 177         WARN_ONCE(ret < 0, KERN_ERR "kmmio arming 0x%08lx failed.\n", f->page);
 178         f->armed = true;
 179         return ret;
 180 }
 181
 182 /** Restore the given page to saved presence state. */
 183 static void disarm_kmmio_fault_page(struct kmmio_fault_page *f)
 184 {
 185         bool tmp;
 186         int ret = set_page_presence(f->page, f->old_presence, &tmp);
 187         WARN_ONCE(ret < 0,
 188                         KERN_ERR "kmmio disarming 0x%08lx failed.\n", f->page);
 189         f->armed = false;
 190 }
 191
 192 /*
 193  * This is being called from do_page_fault().
 194  *
 195  * We may be in an interrupt or a critical section. Also prefecthing may
 196  * trigger a page fault. We may be in the middle of process switch.
 197  * We cannot take any locks, because we could be executing especially
 198  * within a kmmio critical section.
 199  *
 200  * Local interrupts are disabled, so preemption cannot happen.
 201  * Do not enable interrupts, do not sleep, and watch out for other CPUs.
 202  */
 203 /*
 204  * Interrupts are disabled on entry as trap3 is an interrupt gate
 205  * and they remain disabled thorough out this function.
 206  */
 207 int kmmio_handler(struct pt_regs *regs, unsigned long addr)
 208 {
 209         struct kmmio_context *ctx;
 210         struct kmmio_fault_page *faultpage;
 211         int ret = 0; /* default to fault not handled */
 212
 213         /*
 214          * Preemption is now disabled to prevent process switch during
 215          * single stepping. We can only handle one active kmmio trace
 216          * per cpu, so ensure that we finish it before something else
 217          * gets to run. We also hold the RCU read lock over single
 218          * stepping to avoid looking up the probe and kmmio_fault_page
 219          * again.
 220          */
 221         preempt_disable();
 222         rcu_read_lock();
 223
 224         faultpage = get_kmmio_fault_page(addr);
 225         if (!faultpage) {
 226                 /*
 227                  * Either this page fault is not caused by kmmio, or
 228                  * another CPU just pulled the kmmio probe from under
 229                  * our feet. The latter case should not be possible.
 230                  */
 231                 goto no_kmmio;
 232         }
 233
 234         ctx = &get_cpu_var(kmmio_ctx);
 235         if (ctx->active) {
 236                 if (addr == ctx->addr) {
 237                         /*
 238                          * A second fault on the same page means some other
 239                          * condition needs handling by do_page_fault(), the
 240                          * page really not being present is the most common.
 241                          */
 242                         pr_debug("kmmio: secondary hit for 0x%08lx CPU %d.\n",
 243                                         addr, smp_processor_id());
 244
 245                         if (!faultpage->old_presence)
 246                                 pr_info("kmmio: unexpected secondary hit for "
 247                                         "address 0x%08lx on CPU %d.\n", addr,
 248                                         smp_processor_id());
 249                 } else {
 250                         /*
 251                          * Prevent overwriting already in-flight context.
 252                          * This should not happen, let's hope disarming at
 253                          * least prevents a panic.
 254                          */
 255                         pr_emerg("kmmio: recursive probe hit on CPU %d, "
 256                                         "for address 0x%08lx. Ignoring.\n",
 257                                         smp_processor_id(), addr);
 258                         pr_emerg("kmmio: previous hit was at 0x%08lx.\n",
 259                                                 ctx->addr);
 260                         disarm_kmmio_fault_page(faultpage);
 261                 }
 262                 goto no_kmmio_ctx;
 263         }
 264         ctx->active++;
 265
 266         ctx->fpage = faultpage;
 267         ctx->probe = get_kmmio_probe(addr);
 268         ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
 269         ctx->addr = addr;
 270
 271         if (ctx->probe && ctx->probe->pre_handler)
 272                 ctx->probe->pre_handler(ctx->probe, regs, addr);
 273
 274         /*
 275          * Enable single-stepping and disable interrupts for the faulting
 276          * context. Local interrupts must not get enabled during stepping.
 277          */
 278         regs->flags |= X86_EFLAGS_TF;
 279         regs->flags &= ~X86_EFLAGS_IF;
 280
 281         /* Now we set present bit in PTE and single step. */
 282         disarm_kmmio_fault_page(ctx->fpage);
 283
 284         /*
 285          * If another cpu accesses the same page while we are stepping,
 286          * the access will not be caught. It will simply succeed and the
 287          * only downside is we lose the event. If this becomes a problem,
 288          * the user should drop to single cpu before tracing.
 289          */
 290
 291         put_cpu_var(kmmio_ctx);
 292         return 1; /* fault handled */
 293
 294 no_kmmio_ctx:
 295         put_cpu_var(kmmio_ctx);
 296 no_kmmio:
 297         rcu_read_unlock();
 298         preempt_enable_no_resched();
 299         return ret;
 300 }
 301
 302 /*
 303  * Interrupts are disabled on entry as trap1 is an interrupt gate
 304  * and they remain disabled thorough out this function.
 305  * This must always get called as the pair to kmmio_handler().
 306  */
 307 static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
 308 {
 309         int ret = 0;
 310         struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
 311
 312         if (!ctx->active) {
 313                 pr_debug("kmmio: spurious debug trap on CPU %d.\n",
 314                                                         smp_processor_id());
 315                 goto out;
 316         }
 317
 318         if (ctx->probe && ctx->probe->post_handler)
 319                 ctx->probe->post_handler(ctx->probe, condition, regs);
 320
 321         /* Prevent racing against release_kmmio_fault_page(). */
 322         spin_lock(&kmmio_lock);
 323         if (ctx->fpage->count)
 324                 arm_kmmio_fault_page(ctx->fpage);
 325         spin_unlock(&kmmio_lock);
 326
 327         regs->flags &= ~X86_EFLAGS_TF;
 328         regs->flags |= ctx->saved_flags;
 329
 330         /* These were acquired in kmmio_handler(). */
 331         ctx->active--;
 332         BUG_ON(ctx->active);
 333         rcu_read_unlock();
 334         preempt_enable_no_resched();
 335
 336         /*
 337          * if somebody else is singlestepping across a probe point, flags
 338          * will have TF set, in which case, continue the remaining processing
 339          * of do_debug, as if this is not a probe hit.
 340          */
 341         if (!(regs->flags & X86_EFLAGS_TF))
 342                 ret = 1;
 343 out:
 344         put_cpu_var(kmmio_ctx);
 345         return ret;
 346 }
 347
 348 /* You must be holding kmmio_lock. */
 349 static int add_kmmio_fault_page(unsigned long page)
 350 {
 351         struct kmmio_fault_page *f;
 352
 353         page &= PAGE_MASK;
 354         f = get_kmmio_fault_page(page);
 355         if (f) {
 356                 if (!f->count)
 357                         arm_kmmio_fault_page(f);
 358                 f->count++;
 359                 return 0;
 360         }
 361
 362         f = kzalloc(sizeof(*f), GFP_ATOMIC);
 363         if (!f)
 364                 return -1;
 365
 366         f->count = 1;
 367         f->page = page;
 368
 369         if (arm_kmmio_fault_page(f)) {
 370                 kfree(f);
 371                 return -1;
 372         }
 373
 374         list_add_rcu(&f->list, kmmio_page_list(f->page));
 375
 376         return 0;
 377 }
 378
 379 /* You must be holding kmmio_lock. */
 380 static void release_kmmio_fault_page(unsigned long page,
 381                                 struct kmmio_fault_page **release_list)
 382 {
 383         struct kmmio_fault_page *f;
 384
 385         page &= PAGE_MASK;
 386         f = get_kmmio_fault_page(page);
 387         if (!f)
 388                 return;
 389
 390         f->count--;
 391         BUG_ON(f->count < 0);
 392         if (!f->count) {
 393                 disarm_kmmio_fault_page(f);
 394                 f->release_next = *release_list;
 395                 *release_list = f;
 396         }
 397 }
 398
 399 /*
 400  * With page-unaligned ioremaps, one or two armed pages may contain
 401  * addresses from outside the intended mapping. Events for these addresses
 402  * are currently silently dropped. The events may result only from programming
 403  * mistakes by accessing addresses before the beginning or past the end of a
 404  * mapping.
 405  */
 406 int register_kmmio_probe(struct kmmio_probe *p)
 407 {
 408         unsigned long flags;
 409         int ret = 0;
 410         unsigned long size = 0;
 411         const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
 412
 413         spin_lock_irqsave(&kmmio_lock, flags);
 414         if (get_kmmio_probe(p->addr)) {
 415                 ret = -EEXIST;
 416                 goto out;
 417         }
 418         kmmio_count++;
 419         list_add_rcu(&p->list, &kmmio_probes);
 420         while (size < size_lim) {
 421                 if (add_kmmio_fault_page(p->addr + size))
 422                         pr_err("kmmio: Unable to set page fault.\n");
 423                 size += PAGE_SIZE;
 424         }
 425 out:
 426         spin_unlock_irqrestore(&kmmio_lock, flags);
 427         /*
 428          * XXX: What should I do here?
 429          * Here was a call to global_flush_tlb(), but it does not exist
 430          * anymore. It seems it's not needed after all.
 431          */
 432         return ret;
 433 }
 434 EXPORT_SYMBOL(register_kmmio_probe);
 435
 436 static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
 437 {
 438         struct kmmio_delayed_release *dr = container_of(
 439                                                 head,
 440                                                 struct kmmio_delayed_release,
 441                                                 rcu);
 442         struct kmmio_fault_page *p = dr->release_list;
 443         while (p) {
 444                 struct kmmio_fault_page *next = p->release_next;
 445                 BUG_ON(p->count);
 446                 kfree(p);
 447                 p = next;
 448         }
 449         kfree(dr);
 450 }
 451
 452 static void remove_kmmio_fault_pages(struct rcu_head *head)
 453 {
 454         struct kmmio_delayed_release *dr =
 455                 container_of(head, struct kmmio_delayed_release, rcu);
 456         struct kmmio_fault_page *p = dr->release_list;
 457         struct kmmio_fault_page **prevp = &dr->release_list;
 458         unsigned long flags;
 459
 460         spin_lock_irqsave(&kmmio_lock, flags);
 461         while (p) {
 462                 if (!p->count) {
 463                         list_del_rcu(&p->list);
 464                         prevp = &p->release_next;
 465                 } else {
 466                         *prevp = p->release_next;
 467                 }
 468                 p = p->release_next;
 469         }
 470         spin_unlock_irqrestore(&kmmio_lock, flags);
 471
 472         /* This is the real RCU destroy call. */
 473         call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
 474 }
 475
 476 /*
 477  * Remove a kmmio probe. You have to synchronize_rcu() before you can be
 478  * sure that the callbacks will not be called anymore. Only after that
 479  * you may actually release your struct kmmio_probe.
 480  *
 481  * Unregistering a kmmio fault page has three steps:
 482  * 1. release_kmmio_fault_page()
 483  *    Disarm the page, wait a grace period to let all faults finish.
 484  * 2. remove_kmmio_fault_pages()
 485  *    Remove the pages from kmmio_page_table.
 486  * 3. rcu_free_kmmio_fault_pages()
 487  *    Actally free the kmmio_fault_page structs as with RCU.
 488  */
 489 void unregister_kmmio_probe(struct kmmio_probe *p)
 490 {
 491         unsigned long flags;
 492         unsigned long size = 0;
 493         const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
 494         struct kmmio_fault_page *release_list = NULL;
 495         struct kmmio_delayed_release *drelease;
 496
 497         spin_lock_irqsave(&kmmio_lock, flags);
 498         while (size < size_lim) {
 499                 release_kmmio_fault_page(p->addr + size, &release_list);
 500                 size += PAGE_SIZE;
 501         }
 502         list_del_rcu(&p->list);
 503         kmmio_count--;
 504         spin_unlock_irqrestore(&kmmio_lock, flags);
 505
 506         drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
 507         if (!drelease) {
 508                 pr_crit("kmmio: leaking kmmio_fault_page objects.\n");
 509                 return;
 510         }
 511         drelease->release_list = release_list;
 512
 513         /*
 514          * This is not really RCU here. We have just disarmed a set of
 515          * pages so that they cannot trigger page faults anymore. However,
 516          * we cannot remove the pages from kmmio_page_table,
 517          * because a probe hit might be in flight on another CPU. The
 518          * pages are collected into a list, and they will be removed from
 519          * kmmio_page_table when it is certain that no probe hit related to
 520          * these pages can be in flight. RCU grace period sounds like a
 521          * good choice.
 522          *
 523          * If we removed the pages too early, kmmio page fault handler might
 524          * not find the respective kmmio_fault_page and determine it's not
 525          * a kmmio fault, when it actually is. This would lead to madness.
 526          */
 527         call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
 528 }
 529 EXPORT_SYMBOL(unregister_kmmio_probe);
 530
 531 static int kmmio_die_notifier(struct notifier_block *nb, unsigned long val,
 532                                                                 void *args)
 533 {
 534         struct die_args *arg = args;
 535
 536         if (val == DIE_DEBUG && (arg->err & DR_STEP))
 537                 if (post_kmmio_handler(arg->err, arg->regs) == 1)
 538                         return NOTIFY_STOP;
 539
 540         return NOTIFY_DONE;
 541 }
 542
 543 static struct notifier_block nb_die = {
 544         .notifier_call = kmmio_die_notifier
 545 };
 546
 547 static int __init init_kmmio(void)
 548 {
 549         int i;
 550         for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
 551                 INIT_LIST_HEAD(&kmmio_page_table[i]);
 552         return register_die_notifier(&nb_die);
 553 }
 554 fs_initcall(init_kmmio); /* should be before device_initcall() */