root/arch/x86/mm/kmmio.c

DEFINITIONS

1. kmmio_page_list
2. get_kmmio_probe
3. get_kmmio_fault_page
4. clear_pmd_presence
5. clear_pte_presence
6. clear_page_presence
7. arm_kmmio_fault_page
8. disarm_kmmio_fault_page
9. kmmio_handler
10. post_kmmio_handler
11. add_kmmio_fault_page
12. release_kmmio_fault_page
13. register_kmmio_probe
14. rcu_free_kmmio_fault_pages
15. remove_kmmio_fault_pages
16. unregister_kmmio_probe
17. kmmio_die_notifier
18. kmmio_init
19. kmmio_cleanup

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Support for MMIO probes.
   3  * Benfit many code from kprobes
   4  * (C) 2002 Louis Zhuang <louis.zhuang@intel.com>.
   5  *     2007 Alexander Eichner
   6  *     2008 Pekka Paalanen <pq@iki.fi>
   7  */
   8 
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10 
  11 #include <linux/list.h>
  12 #include <linux/rculist.h>
  13 #include <linux/spinlock.h>
  14 #include <linux/hash.h>
  15 #include <linux/export.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 
  34 struct kmmio_fault_page {
  35         struct list_head list;
  36         struct kmmio_fault_page *release_next;
  37         unsigned long addr; /* the requested address */
  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 
  52 struct kmmio_delayed_release {
  53         struct rcu_head rcu;
  54         struct kmmio_fault_page *release_list;
  55 };
  56 
  57 struct 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 
  65 static DEFINE_SPINLOCK(kmmio_lock);
  66 
  67 /* Protected by kmmio_lock */
  68 unsigned int kmmio_count;
  69 
  70 /* Read-protected by RCU, write-protected by kmmio_lock. */
  71 static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
  72 static LIST_HEAD(kmmio_probes);
  73 
  74 static struct list_head *kmmio_page_list(unsigned long addr)
  75 {
  76         unsigned int l;
  77         pte_t *pte = lookup_address(addr, &l);
  78 
  79         if (!pte)
  80                 return NULL;
  81         addr &= page_level_mask(l);
  82 
  83         return &kmmio_page_table[hash_long(addr, KMMIO_PAGE_HASH_BITS)];
  84 }
  85 
  86 /* Accessed per-cpu */
  87 static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
  88 
  89 /*
  90  * this is basically a dynamic stabbing problem:
  91  * Could use the existing prio tree code or
  92  * Possible better implementations:
  93  * The Interval Skip List: A Data Structure for Finding All Intervals That
  94  * Overlap a Point (might be simple)
  95  * Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup
  96  */
  97 /* Get the kmmio at this addr (if any). You must be holding RCU read lock. */
  98 static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
  99 {
 100         struct kmmio_probe *p;
 101         list_for_each_entry_rcu(p, &kmmio_probes, list) {
 102                 if (addr >= p->addr && addr < (p->addr + p->len))
 103                         return p;
 104         }
 105         return NULL;
 106 }
 107 
 108 /* You must be holding RCU read lock. */
 109 static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long addr)
 110 {
 111         struct list_head *head;
 112         struct kmmio_fault_page *f;
 113         unsigned int l;
 114         pte_t *pte = lookup_address(addr, &l);
 115 
 116         if (!pte)
 117                 return NULL;
 118         addr &= page_level_mask(l);
 119         head = kmmio_page_list(addr);
 120         list_for_each_entry_rcu(f, head, list) {
 121                 if (f->addr == addr)
 122                         return f;
 123         }
 124         return NULL;
 125 }
 126 
 127 static void clear_pmd_presence(pmd_t *pmd, bool clear, pmdval_t *old)
 128 {
 129         pmd_t new_pmd;
 130         pmdval_t v = pmd_val(*pmd);
 131         if (clear) {
 132                 *old = v;
 133                 new_pmd = pmd_mknotpresent(*pmd);
 134         } else {
 135                 /* Presume this has been called with clear==true previously */
 136                 new_pmd = __pmd(*old);
 137         }
 138         set_pmd(pmd, new_pmd);
 139 }
 140 
 141 static void clear_pte_presence(pte_t *pte, bool clear, pteval_t *old)
 142 {
 143         pteval_t v = pte_val(*pte);
 144         if (clear) {
 145                 *old = v;
 146                 /* Nothing should care about address */
 147                 pte_clear(&init_mm, 0, pte);
 148         } else {
 149                 /* Presume this has been called with clear==true previously */
 150                 set_pte_atomic(pte, __pte(*old));
 151         }
 152 }
 153 
 154 static int clear_page_presence(struct kmmio_fault_page *f, bool clear)
 155 {
 156         unsigned int level;
 157         pte_t *pte = lookup_address(f->addr, &level);
 158 
 159         if (!pte) {
 160                 pr_err("no pte for addr 0x%08lx\n", f->addr);
 161                 return -1;
 162         }
 163 
 164         switch (level) {
 165         case PG_LEVEL_2M:
 166                 clear_pmd_presence((pmd_t *)pte, clear, &f->old_presence);
 167                 break;
 168         case PG_LEVEL_4K:
 169                 clear_pte_presence(pte, clear, &f->old_presence);
 170                 break;
 171         default:
 172                 pr_err("unexpected page level 0x%x.\n", level);
 173                 return -1;
 174         }
 175 
 176         __flush_tlb_one_kernel(f->addr);
 177         return 0;
 178 }
 179 
 180 /*
 181  * Mark the given page as not present. Access to it will trigger a fault.
 182  *
 183  * Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the
 184  * protection is ignored here. RCU read lock is assumed held, so the struct
 185  * will not disappear unexpectedly. Furthermore, the caller must guarantee,
 186  * that double arming the same virtual address (page) cannot occur.
 187  *
 188  * Double disarming on the other hand is allowed, and may occur when a fault
 189  * and mmiotrace shutdown happen simultaneously.
 190  */
 191 static int arm_kmmio_fault_page(struct kmmio_fault_page *f)
 192 {
 193         int ret;
 194         WARN_ONCE(f->armed, KERN_ERR pr_fmt("kmmio page already armed.\n"));
 195         if (f->armed) {
 196                 pr_warning("double-arm: addr 0x%08lx, ref %d, old %d\n",
 197                            f->addr, f->count, !!f->old_presence);
 198         }
 199         ret = clear_page_presence(f, true);
 200         WARN_ONCE(ret < 0, KERN_ERR pr_fmt("arming at 0x%08lx failed.\n"),
 201                   f->addr);
 202         f->armed = true;
 203         return ret;
 204 }
 205 
 206 /** Restore the given page to saved presence state. */
 207 static void disarm_kmmio_fault_page(struct kmmio_fault_page *f)
 208 {
 209         int ret = clear_page_presence(f, false);
 210         WARN_ONCE(ret < 0,
 211                         KERN_ERR "kmmio disarming at 0x%08lx failed.\n", f->addr);
 212         f->armed = false;
 213 }
 214 
 215 /*
 216  * This is being called from do_page_fault().
 217  *
 218  * We may be in an interrupt or a critical section. Also prefecthing may
 219  * trigger a page fault. We may be in the middle of process switch.
 220  * We cannot take any locks, because we could be executing especially
 221  * within a kmmio critical section.
 222  *
 223  * Local interrupts are disabled, so preemption cannot happen.
 224  * Do not enable interrupts, do not sleep, and watch out for other CPUs.
 225  */
 226 /*
 227  * Interrupts are disabled on entry as trap3 is an interrupt gate
 228  * and they remain disabled throughout this function.
 229  */
 230 int kmmio_handler(struct pt_regs *regs, unsigned long addr)
 231 {
 232         struct kmmio_context *ctx;
 233         struct kmmio_fault_page *faultpage;
 234         int ret = 0; /* default to fault not handled */
 235         unsigned long page_base = addr;
 236         unsigned int l;
 237         pte_t *pte = lookup_address(addr, &l);
 238         if (!pte)
 239                 return -EINVAL;
 240         page_base &= page_level_mask(l);
 241 
 242         /*
 243          * Preemption is now disabled to prevent process switch during
 244          * single stepping. We can only handle one active kmmio trace
 245          * per cpu, so ensure that we finish it before something else
 246          * gets to run. We also hold the RCU read lock over single
 247          * stepping to avoid looking up the probe and kmmio_fault_page
 248          * again.
 249          */
 250         preempt_disable();
 251         rcu_read_lock();
 252 
 253         faultpage = get_kmmio_fault_page(page_base);
 254         if (!faultpage) {
 255                 /*
 256                  * Either this page fault is not caused by kmmio, or
 257                  * another CPU just pulled the kmmio probe from under
 258                  * our feet. The latter case should not be possible.
 259                  */
 260                 goto no_kmmio;
 261         }
 262 
 263         ctx = &get_cpu_var(kmmio_ctx);
 264         if (ctx->active) {
 265                 if (page_base == ctx->addr) {
 266                         /*
 267                          * A second fault on the same page means some other
 268                          * condition needs handling by do_page_fault(), the
 269                          * page really not being present is the most common.
 270                          */
 271                         pr_debug("secondary hit for 0x%08lx CPU %d.\n",
 272                                  addr, smp_processor_id());
 273 
 274                         if (!faultpage->old_presence)
 275                                 pr_info("unexpected secondary hit for address 0x%08lx on CPU %d.\n",
 276                                         addr, smp_processor_id());
 277                 } else {
 278                         /*
 279                          * Prevent overwriting already in-flight context.
 280                          * This should not happen, let's hope disarming at
 281                          * least prevents a panic.
 282                          */
 283                         pr_emerg("recursive probe hit on CPU %d, for address 0x%08lx. Ignoring.\n",
 284                                  smp_processor_id(), addr);
 285                         pr_emerg("previous hit was at 0x%08lx.\n", ctx->addr);
 286                         disarm_kmmio_fault_page(faultpage);
 287                 }
 288                 goto no_kmmio_ctx;
 289         }
 290         ctx->active++;
 291 
 292         ctx->fpage = faultpage;
 293         ctx->probe = get_kmmio_probe(page_base);
 294         ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
 295         ctx->addr = page_base;
 296 
 297         if (ctx->probe && ctx->probe->pre_handler)
 298                 ctx->probe->pre_handler(ctx->probe, regs, addr);
 299 
 300         /*
 301          * Enable single-stepping and disable interrupts for the faulting
 302          * context. Local interrupts must not get enabled during stepping.
 303          */
 304         regs->flags |= X86_EFLAGS_TF;
 305         regs->flags &= ~X86_EFLAGS_IF;
 306 
 307         /* Now we set present bit in PTE and single step. */
 308         disarm_kmmio_fault_page(ctx->fpage);
 309 
 310         /*
 311          * If another cpu accesses the same page while we are stepping,
 312          * the access will not be caught. It will simply succeed and the
 313          * only downside is we lose the event. If this becomes a problem,
 314          * the user should drop to single cpu before tracing.
 315          */
 316 
 317         put_cpu_var(kmmio_ctx);
 318         return 1; /* fault handled */
 319 
 320 no_kmmio_ctx:
 321         put_cpu_var(kmmio_ctx);
 322 no_kmmio:
 323         rcu_read_unlock();
 324         preempt_enable_no_resched();
 325         return ret;
 326 }
 327 
 328 /*
 329  * Interrupts are disabled on entry as trap1 is an interrupt gate
 330  * and they remain disabled throughout this function.
 331  * This must always get called as the pair to kmmio_handler().
 332  */
 333 static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
 334 {
 335         int ret = 0;
 336         struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
 337 
 338         if (!ctx->active) {
 339                 /*
 340                  * debug traps without an active context are due to either
 341                  * something external causing them (f.e. using a debugger while
 342                  * mmio tracing enabled), or erroneous behaviour
 343                  */
 344                 pr_warning("unexpected debug trap on CPU %d.\n",
 345                            smp_processor_id());
 346                 goto out;
 347         }
 348 
 349         if (ctx->probe && ctx->probe->post_handler)
 350                 ctx->probe->post_handler(ctx->probe, condition, regs);
 351 
 352         /* Prevent racing against release_kmmio_fault_page(). */
 353         spin_lock(&kmmio_lock);
 354         if (ctx->fpage->count)
 355                 arm_kmmio_fault_page(ctx->fpage);
 356         spin_unlock(&kmmio_lock);
 357 
 358         regs->flags &= ~X86_EFLAGS_TF;
 359         regs->flags |= ctx->saved_flags;
 360 
 361         /* These were acquired in kmmio_handler(). */
 362         ctx->active--;
 363         BUG_ON(ctx->active);
 364         rcu_read_unlock();
 365         preempt_enable_no_resched();
 366 
 367         /*
 368          * if somebody else is singlestepping across a probe point, flags
 369          * will have TF set, in which case, continue the remaining processing
 370          * of do_debug, as if this is not a probe hit.
 371          */
 372         if (!(regs->flags & X86_EFLAGS_TF))
 373                 ret = 1;
 374 out:
 375         put_cpu_var(kmmio_ctx);
 376         return ret;
 377 }
 378 
 379 /* You must be holding kmmio_lock. */
 380 static int add_kmmio_fault_page(unsigned long addr)
 381 {
 382         struct kmmio_fault_page *f;
 383 
 384         f = get_kmmio_fault_page(addr);
 385         if (f) {
 386                 if (!f->count)
 387                         arm_kmmio_fault_page(f);
 388                 f->count++;
 389                 return 0;
 390         }
 391 
 392         f = kzalloc(sizeof(*f), GFP_ATOMIC);
 393         if (!f)
 394                 return -1;
 395 
 396         f->count = 1;
 397         f->addr = addr;
 398 
 399         if (arm_kmmio_fault_page(f)) {
 400                 kfree(f);
 401                 return -1;
 402         }
 403 
 404         list_add_rcu(&f->list, kmmio_page_list(f->addr));
 405 
 406         return 0;
 407 }
 408 
 409 /* You must be holding kmmio_lock. */
 410 static void release_kmmio_fault_page(unsigned long addr,
 411                                 struct kmmio_fault_page **release_list)
 412 {
 413         struct kmmio_fault_page *f;
 414 
 415         f = get_kmmio_fault_page(addr);
 416         if (!f)
 417                 return;
 418 
 419         f->count--;
 420         BUG_ON(f->count < 0);
 421         if (!f->count) {
 422                 disarm_kmmio_fault_page(f);
 423                 if (!f->scheduled_for_release) {
 424                         f->release_next = *release_list;
 425                         *release_list = f;
 426                         f->scheduled_for_release = true;
 427                 }
 428         }
 429 }
 430 
 431 /*
 432  * With page-unaligned ioremaps, one or two armed pages may contain
 433  * addresses from outside the intended mapping. Events for these addresses
 434  * are currently silently dropped. The events may result only from programming
 435  * mistakes by accessing addresses before the beginning or past the end of a
 436  * mapping.
 437  */
 438 int register_kmmio_probe(struct kmmio_probe *p)
 439 {
 440         unsigned long flags;
 441         int ret = 0;
 442         unsigned long size = 0;
 443         unsigned long addr = p->addr & PAGE_MASK;
 444         const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
 445         unsigned int l;
 446         pte_t *pte;
 447 
 448         spin_lock_irqsave(&kmmio_lock, flags);
 449         if (get_kmmio_probe(addr)) {
 450                 ret = -EEXIST;
 451                 goto out;
 452         }
 453 
 454         pte = lookup_address(addr, &l);
 455         if (!pte) {
 456                 ret = -EINVAL;
 457                 goto out;
 458         }
 459 
 460         kmmio_count++;
 461         list_add_rcu(&p->list, &kmmio_probes);
 462         while (size < size_lim) {
 463                 if (add_kmmio_fault_page(addr + size))
 464                         pr_err("Unable to set page fault.\n");
 465                 size += page_level_size(l);
 466         }
 467 out:
 468         spin_unlock_irqrestore(&kmmio_lock, flags);
 469         /*
 470          * XXX: What should I do here?
 471          * Here was a call to global_flush_tlb(), but it does not exist
 472          * anymore. It seems it's not needed after all.
 473          */
 474         return ret;
 475 }
 476 EXPORT_SYMBOL(register_kmmio_probe);
 477 
 478 static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
 479 {
 480         struct kmmio_delayed_release *dr = container_of(
 481                                                 head,
 482                                                 struct kmmio_delayed_release,
 483                                                 rcu);
 484         struct kmmio_fault_page *f = dr->release_list;
 485         while (f) {
 486                 struct kmmio_fault_page *next = f->release_next;
 487                 BUG_ON(f->count);
 488                 kfree(f);
 489                 f = next;
 490         }
 491         kfree(dr);
 492 }
 493 
 494 static void remove_kmmio_fault_pages(struct rcu_head *head)
 495 {
 496         struct kmmio_delayed_release *dr =
 497                 container_of(head, struct kmmio_delayed_release, rcu);
 498         struct kmmio_fault_page *f = dr->release_list;
 499         struct kmmio_fault_page **prevp = &dr->release_list;
 500         unsigned long flags;
 501 
 502         spin_lock_irqsave(&kmmio_lock, flags);
 503         while (f) {
 504                 if (!f->count) {
 505                         list_del_rcu(&f->list);
 506                         prevp = &f->release_next;
 507                 } else {
 508                         *prevp = f->release_next;
 509                         f->release_next = NULL;
 510                         f->scheduled_for_release = false;
 511                 }
 512                 f = *prevp;
 513         }
 514         spin_unlock_irqrestore(&kmmio_lock, flags);
 515 
 516         /* This is the real RCU destroy call. */
 517         call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
 518 }
 519 
 520 /*
 521  * Remove a kmmio probe. You have to synchronize_rcu() before you can be
 522  * sure that the callbacks will not be called anymore. Only after that
 523  * you may actually release your struct kmmio_probe.
 524  *
 525  * Unregistering a kmmio fault page has three steps:
 526  * 1. release_kmmio_fault_page()
 527  *    Disarm the page, wait a grace period to let all faults finish.
 528  * 2. remove_kmmio_fault_pages()
 529  *    Remove the pages from kmmio_page_table.
 530  * 3. rcu_free_kmmio_fault_pages()
 531  *    Actually free the kmmio_fault_page structs as with RCU.
 532  */
 533 void unregister_kmmio_probe(struct kmmio_probe *p)
 534 {
 535         unsigned long flags;
 536         unsigned long size = 0;
 537         unsigned long addr = p->addr & PAGE_MASK;
 538         const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
 539         struct kmmio_fault_page *release_list = NULL;
 540         struct kmmio_delayed_release *drelease;
 541         unsigned int l;
 542         pte_t *pte;
 543 
 544         pte = lookup_address(addr, &l);
 545         if (!pte)
 546                 return;
 547 
 548         spin_lock_irqsave(&kmmio_lock, flags);
 549         while (size < size_lim) {
 550                 release_kmmio_fault_page(addr + size, &release_list);
 551                 size += page_level_size(l);
 552         }
 553         list_del_rcu(&p->list);
 554         kmmio_count--;
 555         spin_unlock_irqrestore(&kmmio_lock, flags);
 556 
 557         if (!release_list)
 558                 return;
 559 
 560         drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
 561         if (!drelease) {
 562                 pr_crit("leaking kmmio_fault_page objects.\n");
 563                 return;
 564         }
 565         drelease->release_list = release_list;
 566 
 567         /*
 568          * This is not really RCU here. We have just disarmed a set of
 569          * pages so that they cannot trigger page faults anymore. However,
 570          * we cannot remove the pages from kmmio_page_table,
 571          * because a probe hit might be in flight on another CPU. The
 572          * pages are collected into a list, and they will be removed from
 573          * kmmio_page_table when it is certain that no probe hit related to
 574          * these pages can be in flight. RCU grace period sounds like a
 575          * good choice.
 576          *
 577          * If we removed the pages too early, kmmio page fault handler might
 578          * not find the respective kmmio_fault_page and determine it's not
 579          * a kmmio fault, when it actually is. This would lead to madness.
 580          */
 581         call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
 582 }
 583 EXPORT_SYMBOL(unregister_kmmio_probe);
 584 
 585 static int
 586 kmmio_die_notifier(struct notifier_block *nb, unsigned long val, void *args)
 587 {
 588         struct die_args *arg = args;
 589         unsigned long* dr6_p = (unsigned long *)ERR_PTR(arg->err);
 590 
 591         if (val == DIE_DEBUG && (*dr6_p & DR_STEP))
 592                 if (post_kmmio_handler(*dr6_p, arg->regs) == 1) {
 593                         /*
 594                          * Reset the BS bit in dr6 (pointed by args->err) to
 595                          * denote completion of processing
 596                          */
 597                         *dr6_p &= ~DR_STEP;
 598                         return NOTIFY_STOP;
 599                 }
 600 
 601         return NOTIFY_DONE;
 602 }
 603 
 604 static struct notifier_block nb_die = {
 605         .notifier_call = kmmio_die_notifier
 606 };
 607 
 608 int kmmio_init(void)
 609 {
 610         int i;
 611 
 612         for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
 613                 INIT_LIST_HEAD(&kmmio_page_table[i]);
 614 
 615         return register_die_notifier(&nb_die);
 616 }
 617 
 618 void kmmio_cleanup(void)
 619 {
 620         int i;
 621 
 622         unregister_die_notifier(&nb_die);
 623         for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) {
 624                 WARN_ONCE(!list_empty(&kmmio_page_table[i]),
 625                         KERN_ERR "kmmio_page_table not empty at cleanup, any further tracing will leak memory.\n");
 626         }
 627 }