kernel/events/uprobes.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * User-space Probes (UProbes)
   4  *
   5  * Copyright (C) IBM Corporation, 2008-2012
   6  * Authors:
   7  *      Srikar Dronamraju
   8  *      Jim Keniston
   9  * Copyright (C) 2011-2012 Red Hat, Inc., Peter Zijlstra
  10  */
  11
  12 #include <linux/kernel.h>
  13 #include <linux/highmem.h>
  14 #include <linux/pagemap.h>      /* read_mapping_page */
  15 #include <linux/slab.h>
  16 #include <linux/sched.h>
  17 #include <linux/sched/mm.h>
  18 #include <linux/sched/coredump.h>
  19 #include <linux/export.h>
  20 #include <linux/rmap.h>         /* anon_vma_prepare */
  21 #include <linux/mmu_notifier.h> /* set_pte_at_notify */
  22 #include <linux/swap.h>         /* folio_free_swap */
  23 #include <linux/ptrace.h>       /* user_enable_single_step */
  24 #include <linux/kdebug.h>       /* notifier mechanism */
  25 #include "../../mm/internal.h"  /* munlock_vma_page */
  26 #include <linux/percpu-rwsem.h>
  27 #include <linux/task_work.h>
  28 #include <linux/shmem_fs.h>
  29 #include <linux/khugepaged.h>
  30
  31 #include <linux/uprobes.h>
  32
  33 #define UINSNS_PER_PAGE                 (PAGE_SIZE/UPROBE_XOL_SLOT_BYTES)
  34 #define MAX_UPROBE_XOL_SLOTS            UINSNS_PER_PAGE
  35
  36 static struct rb_root uprobes_tree = RB_ROOT;
  37 /*
  38  * allows us to skip the uprobe_mmap if there are no uprobe events active
  39  * at this time.  Probably a fine grained per inode count is better?
  40  */
  41 #define no_uprobe_events()      RB_EMPTY_ROOT(&uprobes_tree)
  42
  43 static DEFINE_SPINLOCK(uprobes_treelock);       /* serialize rbtree access */
  44
  45 #define UPROBES_HASH_SZ 13
  46 /* serialize uprobe->pending_list */
  47 static struct mutex uprobes_mmap_mutex[UPROBES_HASH_SZ];
  48 #define uprobes_mmap_hash(v)    (&uprobes_mmap_mutex[((unsigned long)(v)) % UPROBES_HASH_SZ])
  49
  50 DEFINE_STATIC_PERCPU_RWSEM(dup_mmap_sem);
  51
  52 /* Have a copy of original instruction */
  53 #define UPROBE_COPY_INSN        0
  54
  55 struct uprobe {
  56         struct rb_node          rb_node;        /* node in the rb tree */
  57         refcount_t              ref;
  58         struct rw_semaphore     register_rwsem;
  59         struct rw_semaphore     consumer_rwsem;
  60         struct list_head        pending_list;
  61         struct uprobe_consumer  *consumers;
  62         struct inode            *inode;         /* Also hold a ref to inode */
  63         loff_t                  offset;
  64         loff_t                  ref_ctr_offset;
  65         unsigned long           flags;
  66
  67         /*
  68          * The generic code assumes that it has two members of unknown type
  69          * owned by the arch-specific code:
  70          *
  71          *      insn -  copy_insn() saves the original instruction here for
  72          *              arch_uprobe_analyze_insn().
  73          *
  74          *      ixol -  potentially modified instruction to execute out of
  75          *              line, copied to xol_area by xol_get_insn_slot().
  76          */
  77         struct arch_uprobe      arch;
  78 };
  79
  80 struct delayed_uprobe {
  81         struct list_head list;
  82         struct uprobe *uprobe;
  83         struct mm_struct *mm;
  84 };
  85
  86 static DEFINE_MUTEX(delayed_uprobe_lock);
  87 static LIST_HEAD(delayed_uprobe_list);
  88
  89 /*
  90  * Execute out of line area: anonymous executable mapping installed
  91  * by the probed task to execute the copy of the original instruction
  92  * mangled by set_swbp().
  93  *
  94  * On a breakpoint hit, thread contests for a slot.  It frees the
  95  * slot after singlestep. Currently a fixed number of slots are
  96  * allocated.
  97  */
  98 struct xol_area {
  99         wait_queue_head_t               wq;             /* if all slots are busy */
 100         atomic_t                        slot_count;     /* number of in-use slots */
 101         unsigned long                   *bitmap;        /* 0 = free slot */
 102
 103         struct vm_special_mapping       xol_mapping;
 104         struct page                     *pages[2];
 105         /*
 106          * We keep the vma's vm_start rather than a pointer to the vma
 107          * itself.  The probed process or a naughty kernel module could make
 108          * the vma go away, and we must handle that reasonably gracefully.
 109          */
 110         unsigned long                   vaddr;          /* Page(s) of instruction slots */
 111 };
 112
 113 /*
 114  * valid_vma: Verify if the specified vma is an executable vma
 115  * Relax restrictions while unregistering: vm_flags might have
 116  * changed after breakpoint was inserted.
 117  *      - is_register: indicates if we are in register context.
 118  *      - Return 1 if the specified virtual address is in an
 119  *        executable vma.
 120  */
 121 static bool valid_vma(struct vm_area_struct *vma, bool is_register)
 122 {
 123         vm_flags_t flags = VM_HUGETLB | VM_MAYEXEC | VM_MAYSHARE;
 124
 125         if (is_register)
 126                 flags |= VM_WRITE;
 127
 128         return vma->vm_file && (vma->vm_flags & flags) == VM_MAYEXEC;
 129 }
 130
 131 static unsigned long offset_to_vaddr(struct vm_area_struct *vma, loff_t offset)
 132 {
 133         return vma->vm_start + offset - ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 134 }
 135
 136 static loff_t vaddr_to_offset(struct vm_area_struct *vma, unsigned long vaddr)
 137 {
 138         return ((loff_t)vma->vm_pgoff << PAGE_SHIFT) + (vaddr - vma->vm_start);
 139 }
 140
 141 /**
 142  * __replace_page - replace page in vma by new page.
 143  * based on replace_page in mm/ksm.c
 144  *
 145  * @vma:      vma that holds the pte pointing to page
 146  * @addr:     address the old @page is mapped at
 147  * @old_page: the page we are replacing by new_page
 148  * @new_page: the modified page we replace page by
 149  *
 150  * If @new_page is NULL, only unmap @old_page.
 151  *
 152  * Returns 0 on success, negative error code otherwise.
 153  */
 154 static int __replace_page(struct vm_area_struct *vma, unsigned long addr,
 155                                 struct page *old_page, struct page *new_page)
 156 {
 157         struct folio *old_folio = page_folio(old_page);
 158         struct folio *new_folio;
 159         struct mm_struct *mm = vma->vm_mm;
 160         DEFINE_FOLIO_VMA_WALK(pvmw, old_folio, vma, addr, 0);
 161         int err;
 162         struct mmu_notifier_range range;
 163
 164         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr,
 165                                 addr + PAGE_SIZE);
 166
 167         if (new_page) {
 168                 new_folio = page_folio(new_page);
 169                 err = mem_cgroup_charge(new_folio, vma->vm_mm, GFP_KERNEL);
 170                 if (err)
 171                         return err;
 172         }
 173
 174         /* For folio_free_swap() below */
 175         folio_lock(old_folio);
 176
 177         mmu_notifier_invalidate_range_start(&range);
 178         err = -EAGAIN;
 179         if (!page_vma_mapped_walk(&pvmw))
 180                 goto unlock;
 181         VM_BUG_ON_PAGE(addr != pvmw.address, old_page);
 182
 183         if (new_page) {
 184                 folio_get(new_folio);
 185                 page_add_new_anon_rmap(new_page, vma, addr);
 186                 folio_add_lru_vma(new_folio, vma);
 187         } else
 188                 /* no new page, just dec_mm_counter for old_page */
 189                 dec_mm_counter(mm, MM_ANONPAGES);
 190
 191         if (!folio_test_anon(old_folio)) {
 192                 dec_mm_counter(mm, mm_counter_file(old_page));
 193                 inc_mm_counter(mm, MM_ANONPAGES);
 194         }
 195
 196         flush_cache_page(vma, addr, pte_pfn(*pvmw.pte));
 197         ptep_clear_flush_notify(vma, addr, pvmw.pte);
 198         if (new_page)
 199                 set_pte_at_notify(mm, addr, pvmw.pte,
 200                                   mk_pte(new_page, vma->vm_page_prot));
 201
 202         page_remove_rmap(old_page, vma, false);
 203         if (!folio_mapped(old_folio))
 204                 folio_free_swap(old_folio);
 205         page_vma_mapped_walk_done(&pvmw);
 206         folio_put(old_folio);
 207
 208         err = 0;
 209  unlock:
 210         mmu_notifier_invalidate_range_end(&range);
 211         folio_unlock(old_folio);
 212         return err;
 213 }
 214
 215 /**
 216  * is_swbp_insn - check if instruction is breakpoint instruction.
 217  * @insn: instruction to be checked.
 218  * Default implementation of is_swbp_insn
 219  * Returns true if @insn is a breakpoint instruction.
 220  */
 221 bool __weak is_swbp_insn(uprobe_opcode_t *insn)
 222 {
 223         return *insn == UPROBE_SWBP_INSN;
 224 }
 225
 226 /**
 227  * is_trap_insn - check if instruction is breakpoint instruction.
 228  * @insn: instruction to be checked.
 229  * Default implementation of is_trap_insn
 230  * Returns true if @insn is a breakpoint instruction.
 231  *
 232  * This function is needed for the case where an architecture has multiple
 233  * trap instructions (like powerpc).
 234  */
 235 bool __weak is_trap_insn(uprobe_opcode_t *insn)
 236 {
 237         return is_swbp_insn(insn);
 238 }
 239
 240 static void copy_from_page(struct page *page, unsigned long vaddr, void *dst, int len)
 241 {
 242         void *kaddr = kmap_atomic(page);
 243         memcpy(dst, kaddr + (vaddr & ~PAGE_MASK), len);
 244         kunmap_atomic(kaddr);
 245 }
 246
 247 static void copy_to_page(struct page *page, unsigned long vaddr, const void *src, int len)
 248 {
 249         void *kaddr = kmap_atomic(page);
 250         memcpy(kaddr + (vaddr & ~PAGE_MASK), src, len);
 251         kunmap_atomic(kaddr);
 252 }
 253
 254 static int verify_opcode(struct page *page, unsigned long vaddr, uprobe_opcode_t *new_opcode)
 255 {
 256         uprobe_opcode_t old_opcode;
 257         bool is_swbp;
 258
 259         /*
 260          * Note: We only check if the old_opcode is UPROBE_SWBP_INSN here.
 261          * We do not check if it is any other 'trap variant' which could
 262          * be conditional trap instruction such as the one powerpc supports.
 263          *
 264          * The logic is that we do not care if the underlying instruction
 265          * is a trap variant; uprobes always wins over any other (gdb)
 266          * breakpoint.
 267          */
 268         copy_from_page(page, vaddr, &old_opcode, UPROBE_SWBP_INSN_SIZE);
 269         is_swbp = is_swbp_insn(&old_opcode);
 270
 271         if (is_swbp_insn(new_opcode)) {
 272                 if (is_swbp)            /* register: already installed? */
 273                         return 0;
 274         } else {
 275                 if (!is_swbp)           /* unregister: was it changed by us? */
 276                         return 0;
 277         }
 278
 279         return 1;
 280 }
 281
 282 static struct delayed_uprobe *
 283 delayed_uprobe_check(struct uprobe *uprobe, struct mm_struct *mm)
 284 {
 285         struct delayed_uprobe *du;
 286
 287         list_for_each_entry(du, &delayed_uprobe_list, list)
 288                 if (du->uprobe == uprobe && du->mm == mm)
 289                         return du;
 290         return NULL;
 291 }
 292
 293 static int delayed_uprobe_add(struct uprobe *uprobe, struct mm_struct *mm)
 294 {
 295         struct delayed_uprobe *du;
 296
 297         if (delayed_uprobe_check(uprobe, mm))
 298                 return 0;
 299
 300         du  = kzalloc(sizeof(*du), GFP_KERNEL);
 301         if (!du)
 302                 return -ENOMEM;
 303
 304         du->uprobe = uprobe;
 305         du->mm = mm;
 306         list_add(&du->list, &delayed_uprobe_list);
 307         return 0;
 308 }
 309
 310 static void delayed_uprobe_delete(struct delayed_uprobe *du)
 311 {
 312         if (WARN_ON(!du))
 313                 return;
 314         list_del(&du->list);
 315         kfree(du);
 316 }
 317
 318 static void delayed_uprobe_remove(struct uprobe *uprobe, struct mm_struct *mm)
 319 {
 320         struct list_head *pos, *q;
 321         struct delayed_uprobe *du;
 322
 323         if (!uprobe && !mm)
 324                 return;
 325
 326         list_for_each_safe(pos, q, &delayed_uprobe_list) {
 327                 du = list_entry(pos, struct delayed_uprobe, list);
 328
 329                 if (uprobe && du->uprobe != uprobe)
 330                         continue;
 331                 if (mm && du->mm != mm)
 332                         continue;
 333
 334                 delayed_uprobe_delete(du);
 335         }
 336 }
 337
 338 static bool valid_ref_ctr_vma(struct uprobe *uprobe,
 339                               struct vm_area_struct *vma)
 340 {
 341         unsigned long vaddr = offset_to_vaddr(vma, uprobe->ref_ctr_offset);
 342
 343         return uprobe->ref_ctr_offset &&
 344                 vma->vm_file &&
 345                 file_inode(vma->vm_file) == uprobe->inode &&
 346                 (vma->vm_flags & (VM_WRITE|VM_SHARED)) == VM_WRITE &&
 347                 vma->vm_start <= vaddr &&
 348                 vma->vm_end > vaddr;
 349 }
 350
 351 static struct vm_area_struct *
 352 find_ref_ctr_vma(struct uprobe *uprobe, struct mm_struct *mm)
 353 {
 354         VMA_ITERATOR(vmi, mm, 0);
 355         struct vm_area_struct *tmp;
 356
 357         for_each_vma(vmi, tmp)
 358                 if (valid_ref_ctr_vma(uprobe, tmp))
 359                         return tmp;
 360
 361         return NULL;
 362 }
 363
 364 static int
 365 __update_ref_ctr(struct mm_struct *mm, unsigned long vaddr, short d)
 366 {
 367         void *kaddr;
 368         struct page *page;
 369         struct vm_area_struct *vma;
 370         int ret;
 371         short *ptr;
 372
 373         if (!vaddr || !d)
 374                 return -EINVAL;
 375
 376         ret = get_user_pages_remote(mm, vaddr, 1,
 377                         FOLL_WRITE, &page, &vma, NULL);
 378         if (unlikely(ret <= 0)) {
 379                 /*
 380                  * We are asking for 1 page. If get_user_pages_remote() fails,
 381                  * it may return 0, in that case we have to return error.
 382                  */
 383                 return ret == 0 ? -EBUSY : ret;
 384         }
 385
 386         kaddr = kmap_atomic(page);
 387         ptr = kaddr + (vaddr & ~PAGE_MASK);
 388
 389         if (unlikely(*ptr + d < 0)) {
 390                 pr_warn("ref_ctr going negative. vaddr: 0x%lx, "
 391                         "curr val: %d, delta: %d\n", vaddr, *ptr, d);
 392                 ret = -EINVAL;
 393                 goto out;
 394         }
 395
 396         *ptr += d;
 397         ret = 0;
 398 out:
 399         kunmap_atomic(kaddr);
 400         put_page(page);
 401         return ret;
 402 }
 403
 404 static void update_ref_ctr_warn(struct uprobe *uprobe,
 405                                 struct mm_struct *mm, short d)
 406 {
 407         pr_warn("ref_ctr %s failed for inode: 0x%lx offset: "
 408                 "0x%llx ref_ctr_offset: 0x%llx of mm: 0x%pK\n",
 409                 d > 0 ? "increment" : "decrement", uprobe->inode->i_ino,
 410                 (unsigned long long) uprobe->offset,
 411                 (unsigned long long) uprobe->ref_ctr_offset, mm);
 412 }
 413
 414 static int update_ref_ctr(struct uprobe *uprobe, struct mm_struct *mm,
 415                           short d)
 416 {
 417         struct vm_area_struct *rc_vma;
 418         unsigned long rc_vaddr;
 419         int ret = 0;
 420
 421         rc_vma = find_ref_ctr_vma(uprobe, mm);
 422
 423         if (rc_vma) {
 424                 rc_vaddr = offset_to_vaddr(rc_vma, uprobe->ref_ctr_offset);
 425                 ret = __update_ref_ctr(mm, rc_vaddr, d);
 426                 if (ret)
 427                         update_ref_ctr_warn(uprobe, mm, d);
 428
 429                 if (d > 0)
 430                         return ret;
 431         }
 432
 433         mutex_lock(&delayed_uprobe_lock);
 434         if (d > 0)
 435                 ret = delayed_uprobe_add(uprobe, mm);
 436         else
 437                 delayed_uprobe_remove(uprobe, mm);
 438         mutex_unlock(&delayed_uprobe_lock);
 439
 440         return ret;
 441 }
 442
 443 /*
 444  * NOTE:
 445  * Expect the breakpoint instruction to be the smallest size instruction for
 446  * the architecture. If an arch has variable length instruction and the
 447  * breakpoint instruction is not of the smallest length instruction
 448  * supported by that architecture then we need to modify is_trap_at_addr and
 449  * uprobe_write_opcode accordingly. This would never be a problem for archs
 450  * that have fixed length instructions.
 451  *
 452  * uprobe_write_opcode - write the opcode at a given virtual address.
 453  * @auprobe: arch specific probepoint information.
 454  * @mm: the probed process address space.
 455  * @vaddr: the virtual address to store the opcode.
 456  * @opcode: opcode to be written at @vaddr.
 457  *
 458  * Called with mm->mmap_lock held for write.
 459  * Return 0 (success) or a negative errno.
 460  */
 461 int uprobe_write_opcode(struct arch_uprobe *auprobe, struct mm_struct *mm,
 462                         unsigned long vaddr, uprobe_opcode_t opcode)
 463 {
 464         struct uprobe *uprobe;
 465         struct page *old_page, *new_page;
 466         struct vm_area_struct *vma;
 467         int ret, is_register, ref_ctr_updated = 0;
 468         bool orig_page_huge = false;
 469         unsigned int gup_flags = FOLL_FORCE;
 470
 471         is_register = is_swbp_insn(&opcode);
 472         uprobe = container_of(auprobe, struct uprobe, arch);
 473
 474 retry:
 475         if (is_register)
 476                 gup_flags |= FOLL_SPLIT_PMD;
 477         /* Read the page with vaddr into memory */
 478         ret = get_user_pages_remote(mm, vaddr, 1, gup_flags,
 479                                     &old_page, &vma, NULL);
 480         if (ret <= 0)
 481                 return ret;
 482
 483         ret = verify_opcode(old_page, vaddr, &opcode);
 484         if (ret <= 0)
 485                 goto put_old;
 486
 487         if (WARN(!is_register && PageCompound(old_page),
 488                  "uprobe unregister should never work on compound page\n")) {
 489                 ret = -EINVAL;
 490                 goto put_old;
 491         }
 492
 493         /* We are going to replace instruction, update ref_ctr. */
 494         if (!ref_ctr_updated && uprobe->ref_ctr_offset) {
 495                 ret = update_ref_ctr(uprobe, mm, is_register ? 1 : -1);
 496                 if (ret)
 497                         goto put_old;
 498
 499                 ref_ctr_updated = 1;
 500         }
 501
 502         ret = 0;
 503         if (!is_register && !PageAnon(old_page))
 504                 goto put_old;
 505
 506         ret = anon_vma_prepare(vma);
 507         if (ret)
 508                 goto put_old;
 509
 510         ret = -ENOMEM;
 511         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vaddr);
 512         if (!new_page)
 513                 goto put_old;
 514
 515         __SetPageUptodate(new_page);
 516         copy_highpage(new_page, old_page);
 517         copy_to_page(new_page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
 518
 519         if (!is_register) {
 520                 struct page *orig_page;
 521                 pgoff_t index;
 522
 523                 VM_BUG_ON_PAGE(!PageAnon(old_page), old_page);
 524
 525                 index = vaddr_to_offset(vma, vaddr & PAGE_MASK) >> PAGE_SHIFT;
 526                 orig_page = find_get_page(vma->vm_file->f_inode->i_mapping,
 527                                           index);
 528
 529                 if (orig_page) {
 530                         if (PageUptodate(orig_page) &&
 531                             pages_identical(new_page, orig_page)) {
 532                                 /* let go new_page */
 533                                 put_page(new_page);
 534                                 new_page = NULL;
 535
 536                                 if (PageCompound(orig_page))
 537                                         orig_page_huge = true;
 538                         }
 539                         put_page(orig_page);
 540                 }
 541         }
 542
 543         ret = __replace_page(vma, vaddr, old_page, new_page);
 544         if (new_page)
 545                 put_page(new_page);
 546 put_old:
 547         put_page(old_page);
 548
 549         if (unlikely(ret == -EAGAIN))
 550                 goto retry;
 551
 552         /* Revert back reference counter if instruction update failed. */
 553         if (ret && is_register && ref_ctr_updated)
 554                 update_ref_ctr(uprobe, mm, -1);
 555
 556         /* try collapse pmd for compound page */
 557         if (!ret && orig_page_huge)
 558                 collapse_pte_mapped_thp(mm, vaddr, false);
 559
 560         return ret;
 561 }
 562
 563 /**
 564  * set_swbp - store breakpoint at a given address.
 565  * @auprobe: arch specific probepoint information.
 566  * @mm: the probed process address space.
 567  * @vaddr: the virtual address to insert the opcode.
 568  *
 569  * For mm @mm, store the breakpoint instruction at @vaddr.
 570  * Return 0 (success) or a negative errno.
 571  */
 572 int __weak set_swbp(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
 573 {
 574         return uprobe_write_opcode(auprobe, mm, vaddr, UPROBE_SWBP_INSN);
 575 }
 576
 577 /**
 578  * set_orig_insn - Restore the original instruction.
 579  * @mm: the probed process address space.
 580  * @auprobe: arch specific probepoint information.
 581  * @vaddr: the virtual address to insert the opcode.
 582  *
 583  * For mm @mm, restore the original opcode (opcode) at @vaddr.
 584  * Return 0 (success) or a negative errno.
 585  */
 586 int __weak
 587 set_orig_insn(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
 588 {
 589         return uprobe_write_opcode(auprobe, mm, vaddr,
 590                         *(uprobe_opcode_t *)&auprobe->insn);
 591 }
 592
 593 static struct uprobe *get_uprobe(struct uprobe *uprobe)
 594 {
 595         refcount_inc(&uprobe->ref);
 596         return uprobe;
 597 }
 598
 599 static void put_uprobe(struct uprobe *uprobe)
 600 {
 601         if (refcount_dec_and_test(&uprobe->ref)) {
 602                 /*
 603                  * If application munmap(exec_vma) before uprobe_unregister()
 604                  * gets called, we don't get a chance to remove uprobe from
 605                  * delayed_uprobe_list from remove_breakpoint(). Do it here.
 606                  */
 607                 mutex_lock(&delayed_uprobe_lock);
 608                 delayed_uprobe_remove(uprobe, NULL);
 609                 mutex_unlock(&delayed_uprobe_lock);
 610                 kfree(uprobe);
 611         }
 612 }
 613
 614 static __always_inline
 615 int uprobe_cmp(const struct inode *l_inode, const loff_t l_offset,
 616                const struct uprobe *r)
 617 {
 618         if (l_inode < r->inode)
 619                 return -1;
 620
 621         if (l_inode > r->inode)
 622                 return 1;
 623
 624         if (l_offset < r->offset)
 625                 return -1;
 626
 627         if (l_offset > r->offset)
 628                 return 1;
 629
 630         return 0;
 631 }
 632
 633 #define __node_2_uprobe(node) \
 634         rb_entry((node), struct uprobe, rb_node)
 635
 636 struct __uprobe_key {
 637         struct inode *inode;
 638         loff_t offset;
 639 };
 640
 641 static inline int __uprobe_cmp_key(const void *key, const struct rb_node *b)
 642 {
 643         const struct __uprobe_key *a = key;
 644         return uprobe_cmp(a->inode, a->offset, __node_2_uprobe(b));
 645 }
 646
 647 static inline int __uprobe_cmp(struct rb_node *a, const struct rb_node *b)
 648 {
 649         struct uprobe *u = __node_2_uprobe(a);
 650         return uprobe_cmp(u->inode, u->offset, __node_2_uprobe(b));
 651 }
 652
 653 static struct uprobe *__find_uprobe(struct inode *inode, loff_t offset)
 654 {
 655         struct __uprobe_key key = {
 656                 .inode = inode,
 657                 .offset = offset,
 658         };
 659         struct rb_node *node = rb_find(&key, &uprobes_tree, __uprobe_cmp_key);
 660
 661         if (node)
 662                 return get_uprobe(__node_2_uprobe(node));
 663
 664         return NULL;
 665 }
 666
 667 /*
 668  * Find a uprobe corresponding to a given inode:offset
 669  * Acquires uprobes_treelock
 670  */
 671 static struct uprobe *find_uprobe(struct inode *inode, loff_t offset)
 672 {
 673         struct uprobe *uprobe;
 674
 675         spin_lock(&uprobes_treelock);
 676         uprobe = __find_uprobe(inode, offset);
 677         spin_unlock(&uprobes_treelock);
 678
 679         return uprobe;
 680 }
 681
 682 static struct uprobe *__insert_uprobe(struct uprobe *uprobe)
 683 {
 684         struct rb_node *node;
 685
 686         node = rb_find_add(&uprobe->rb_node, &uprobes_tree, __uprobe_cmp);
 687         if (node)
 688                 return get_uprobe(__node_2_uprobe(node));
 689
 690         /* get access + creation ref */
 691         refcount_set(&uprobe->ref, 2);
 692         return NULL;
 693 }
 694
 695 /*
 696  * Acquire uprobes_treelock.
 697  * Matching uprobe already exists in rbtree;
 698  *      increment (access refcount) and return the matching uprobe.
 699  *
 700  * No matching uprobe; insert the uprobe in rb_tree;
 701  *      get a double refcount (access + creation) and return NULL.
 702  */
 703 static struct uprobe *insert_uprobe(struct uprobe *uprobe)
 704 {
 705         struct uprobe *u;
 706
 707         spin_lock(&uprobes_treelock);
 708         u = __insert_uprobe(uprobe);
 709         spin_unlock(&uprobes_treelock);
 710
 711         return u;
 712 }
 713
 714 static void
 715 ref_ctr_mismatch_warn(struct uprobe *cur_uprobe, struct uprobe *uprobe)
 716 {
 717         pr_warn("ref_ctr_offset mismatch. inode: 0x%lx offset: 0x%llx "
 718                 "ref_ctr_offset(old): 0x%llx ref_ctr_offset(new): 0x%llx\n",
 719                 uprobe->inode->i_ino, (unsigned long long) uprobe->offset,
 720                 (unsigned long long) cur_uprobe->ref_ctr_offset,
 721                 (unsigned long long) uprobe->ref_ctr_offset);
 722 }
 723
 724 static struct uprobe *alloc_uprobe(struct inode *inode, loff_t offset,
 725                                    loff_t ref_ctr_offset)
 726 {
 727         struct uprobe *uprobe, *cur_uprobe;
 728
 729         uprobe = kzalloc(sizeof(struct uprobe), GFP_KERNEL);
 730         if (!uprobe)
 731                 return NULL;
 732
 733         uprobe->inode = inode;
 734         uprobe->offset = offset;
 735         uprobe->ref_ctr_offset = ref_ctr_offset;
 736         init_rwsem(&uprobe->register_rwsem);
 737         init_rwsem(&uprobe->consumer_rwsem);
 738
 739         /* add to uprobes_tree, sorted on inode:offset */
 740         cur_uprobe = insert_uprobe(uprobe);
 741         /* a uprobe exists for this inode:offset combination */
 742         if (cur_uprobe) {
 743                 if (cur_uprobe->ref_ctr_offset != uprobe->ref_ctr_offset) {
 744                         ref_ctr_mismatch_warn(cur_uprobe, uprobe);
 745                         put_uprobe(cur_uprobe);
 746                         kfree(uprobe);
 747                         return ERR_PTR(-EINVAL);
 748                 }
 749                 kfree(uprobe);
 750                 uprobe = cur_uprobe;
 751         }
 752
 753         return uprobe;
 754 }
 755
 756 static void consumer_add(struct uprobe *uprobe, struct uprobe_consumer *uc)
 757 {
 758         down_write(&uprobe->consumer_rwsem);
 759         uc->next = uprobe->consumers;
 760         uprobe->consumers = uc;
 761         up_write(&uprobe->consumer_rwsem);
 762 }
 763
 764 /*
 765  * For uprobe @uprobe, delete the consumer @uc.
 766  * Return true if the @uc is deleted successfully
 767  * or return false.
 768  */
 769 static bool consumer_del(struct uprobe *uprobe, struct uprobe_consumer *uc)
 770 {
 771         struct uprobe_consumer **con;
 772         bool ret = false;
 773
 774         down_write(&uprobe->consumer_rwsem);
 775         for (con = &uprobe->consumers; *con; con = &(*con)->next) {
 776                 if (*con == uc) {
 777                         *con = uc->next;
 778                         ret = true;
 779                         break;
 780                 }
 781         }
 782         up_write(&uprobe->consumer_rwsem);
 783
 784         return ret;
 785 }
 786
 787 static int __copy_insn(struct address_space *mapping, struct file *filp,
 788                         void *insn, int nbytes, loff_t offset)
 789 {
 790         struct page *page;
 791         /*
 792          * Ensure that the page that has the original instruction is populated
 793          * and in page-cache. If ->read_folio == NULL it must be shmem_mapping(),
 794          * see uprobe_register().
 795          */
 796         if (mapping->a_ops->read_folio)
 797                 page = read_mapping_page(mapping, offset >> PAGE_SHIFT, filp);
 798         else
 799                 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
 800         if (IS_ERR(page))
 801                 return PTR_ERR(page);
 802
 803         copy_from_page(page, offset, insn, nbytes);
 804         put_page(page);
 805
 806         return 0;
 807 }
 808
 809 static int copy_insn(struct uprobe *uprobe, struct file *filp)
 810 {
 811         struct address_space *mapping = uprobe->inode->i_mapping;
 812         loff_t offs = uprobe->offset;
 813         void *insn = &uprobe->arch.insn;
 814         int size = sizeof(uprobe->arch.insn);
 815         int len, err = -EIO;
 816
 817         /* Copy only available bytes, -EIO if nothing was read */
 818         do {
 819                 if (offs >= i_size_read(uprobe->inode))
 820                         break;
 821
 822                 len = min_t(int, size, PAGE_SIZE - (offs & ~PAGE_MASK));
 823                 err = __copy_insn(mapping, filp, insn, len, offs);
 824                 if (err)
 825                         break;
 826
 827                 insn += len;
 828                 offs += len;
 829                 size -= len;
 830         } while (size);
 831
 832         return err;
 833 }
 834
 835 static int prepare_uprobe(struct uprobe *uprobe, struct file *file,
 836                                 struct mm_struct *mm, unsigned long vaddr)
 837 {
 838         int ret = 0;
 839
 840         if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
 841                 return ret;
 842
 843         /* TODO: move this into _register, until then we abuse this sem. */
 844         down_write(&uprobe->consumer_rwsem);
 845         if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
 846                 goto out;
 847
 848         ret = copy_insn(uprobe, file);
 849         if (ret)
 850                 goto out;
 851
 852         ret = -ENOTSUPP;
 853         if (is_trap_insn((uprobe_opcode_t *)&uprobe->arch.insn))
 854                 goto out;
 855
 856         ret = arch_uprobe_analyze_insn(&uprobe->arch, mm, vaddr);
 857         if (ret)
 858                 goto out;
 859
 860         smp_wmb(); /* pairs with the smp_rmb() in handle_swbp() */
 861         set_bit(UPROBE_COPY_INSN, &uprobe->flags);
 862
 863  out:
 864         up_write(&uprobe->consumer_rwsem);
 865
 866         return ret;
 867 }
 868
 869 static inline bool consumer_filter(struct uprobe_consumer *uc,
 870                                    enum uprobe_filter_ctx ctx, struct mm_struct *mm)
 871 {
 872         return !uc->filter || uc->filter(uc, ctx, mm);
 873 }
 874
 875 static bool filter_chain(struct uprobe *uprobe,
 876                          enum uprobe_filter_ctx ctx, struct mm_struct *mm)
 877 {
 878         struct uprobe_consumer *uc;
 879         bool ret = false;
 880
 881         down_read(&uprobe->consumer_rwsem);
 882         for (uc = uprobe->consumers; uc; uc = uc->next) {
 883                 ret = consumer_filter(uc, ctx, mm);
 884                 if (ret)
 885                         break;
 886         }
 887         up_read(&uprobe->consumer_rwsem);
 888
 889         return ret;
 890 }
 891
 892 static int
 893 install_breakpoint(struct uprobe *uprobe, struct mm_struct *mm,
 894                         struct vm_area_struct *vma, unsigned long vaddr)
 895 {
 896         bool first_uprobe;
 897         int ret;
 898
 899         ret = prepare_uprobe(uprobe, vma->vm_file, mm, vaddr);
 900         if (ret)
 901                 return ret;
 902
 903         /*
 904          * set MMF_HAS_UPROBES in advance for uprobe_pre_sstep_notifier(),
 905          * the task can hit this breakpoint right after __replace_page().
 906          */
 907         first_uprobe = !test_bit(MMF_HAS_UPROBES, &mm->flags);
 908         if (first_uprobe)
 909                 set_bit(MMF_HAS_UPROBES, &mm->flags);
 910
 911         ret = set_swbp(&uprobe->arch, mm, vaddr);
 912         if (!ret)
 913                 clear_bit(MMF_RECALC_UPROBES, &mm->flags);
 914         else if (first_uprobe)
 915                 clear_bit(MMF_HAS_UPROBES, &mm->flags);
 916
 917         return ret;
 918 }
 919
 920 static int
 921 remove_breakpoint(struct uprobe *uprobe, struct mm_struct *mm, unsigned long vaddr)
 922 {
 923         set_bit(MMF_RECALC_UPROBES, &mm->flags);
 924         return set_orig_insn(&uprobe->arch, mm, vaddr);
 925 }
 926
 927 static inline bool uprobe_is_active(struct uprobe *uprobe)
 928 {
 929         return !RB_EMPTY_NODE(&uprobe->rb_node);
 930 }
 931 /*
 932  * There could be threads that have already hit the breakpoint. They
 933  * will recheck the current insn and restart if find_uprobe() fails.
 934  * See find_active_uprobe().
 935  */
 936 static void delete_uprobe(struct uprobe *uprobe)
 937 {
 938         if (WARN_ON(!uprobe_is_active(uprobe)))
 939                 return;
 940
 941         spin_lock(&uprobes_treelock);
 942         rb_erase(&uprobe->rb_node, &uprobes_tree);
 943         spin_unlock(&uprobes_treelock);
 944         RB_CLEAR_NODE(&uprobe->rb_node); /* for uprobe_is_active() */
 945         put_uprobe(uprobe);
 946 }
 947
 948 struct map_info {
 949         struct map_info *next;
 950         struct mm_struct *mm;
 951         unsigned long vaddr;
 952 };
 953
 954 static inline struct map_info *free_map_info(struct map_info *info)
 955 {
 956         struct map_info *next = info->next;
 957         kfree(info);
 958         return next;
 959 }
 960
 961 static struct map_info *
 962 build_map_info(struct address_space *mapping, loff_t offset, bool is_register)
 963 {
 964         unsigned long pgoff = offset >> PAGE_SHIFT;
 965         struct vm_area_struct *vma;
 966         struct map_info *curr = NULL;
 967         struct map_info *prev = NULL;
 968         struct map_info *info;
 969         int more = 0;
 970
 971  again:
 972         i_mmap_lock_read(mapping);
 973         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
 974                 if (!valid_vma(vma, is_register))
 975                         continue;
 976
 977                 if (!prev && !more) {
 978                         /*
 979                          * Needs GFP_NOWAIT to avoid i_mmap_rwsem recursion through
 980                          * reclaim. This is optimistic, no harm done if it fails.
 981                          */
 982                         prev = kmalloc(sizeof(struct map_info),
 983                                         GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
 984                         if (prev)
 985                                 prev->next = NULL;
 986                 }
 987                 if (!prev) {
 988                         more++;
 989                         continue;
 990                 }
 991
 992                 if (!mmget_not_zero(vma->vm_mm))
 993                         continue;
 994
 995                 info = prev;
 996                 prev = prev->next;
 997                 info->next = curr;
 998                 curr = info;
 999
1000                 info->mm = vma->vm_mm;
1001                 info->vaddr = offset_to_vaddr(vma, offset);
1002         }
1003         i_mmap_unlock_read(mapping);
1004
1005         if (!more)
1006                 goto out;
1007
1008         prev = curr;
1009         while (curr) {
1010                 mmput(curr->mm);
1011                 curr = curr->next;
1012         }
1013
1014         do {
1015                 info = kmalloc(sizeof(struct map_info), GFP_KERNEL);
1016                 if (!info) {
1017                         curr = ERR_PTR(-ENOMEM);
1018                         goto out;
1019                 }
1020                 info->next = prev;
1021                 prev = info;
1022         } while (--more);
1023
1024         goto again;
1025  out:
1026         while (prev)
1027                 prev = free_map_info(prev);
1028         return curr;
1029 }
1030
1031 static int
1032 register_for_each_vma(struct uprobe *uprobe, struct uprobe_consumer *new)
1033 {
1034         bool is_register = !!new;
1035         struct map_info *info;
1036         int err = 0;
1037
1038         percpu_down_write(&dup_mmap_sem);
1039         info = build_map_info(uprobe->inode->i_mapping,
1040                                         uprobe->offset, is_register);
1041         if (IS_ERR(info)) {
1042                 err = PTR_ERR(info);
1043                 goto out;
1044         }
1045
1046         while (info) {
1047                 struct mm_struct *mm = info->mm;
1048                 struct vm_area_struct *vma;
1049
1050                 if (err && is_register)
1051                         goto free;
1052
1053                 mmap_write_lock(mm);
1054                 vma = find_vma(mm, info->vaddr);
1055                 if (!vma || !valid_vma(vma, is_register) ||
1056                     file_inode(vma->vm_file) != uprobe->inode)
1057                         goto unlock;
1058
1059                 if (vma->vm_start > info->vaddr ||
1060                     vaddr_to_offset(vma, info->vaddr) != uprobe->offset)
1061                         goto unlock;
1062
1063                 if (is_register) {
1064                         /* consult only the "caller", new consumer. */
1065                         if (consumer_filter(new,
1066                                         UPROBE_FILTER_REGISTER, mm))
1067                                 err = install_breakpoint(uprobe, mm, vma, info->vaddr);
1068                 } else if (test_bit(MMF_HAS_UPROBES, &mm->flags)) {
1069                         if (!filter_chain(uprobe,
1070                                         UPROBE_FILTER_UNREGISTER, mm))
1071                                 err |= remove_breakpoint(uprobe, mm, info->vaddr);
1072                 }
1073
1074  unlock:
1075                 mmap_write_unlock(mm);
1076  free:
1077                 mmput(mm);
1078                 info = free_map_info(info);
1079         }
1080  out:
1081         percpu_up_write(&dup_mmap_sem);
1082         return err;
1083 }
1084
1085 static void
1086 __uprobe_unregister(struct uprobe *uprobe, struct uprobe_consumer *uc)
1087 {
1088         int err;
1089
1090         if (WARN_ON(!consumer_del(uprobe, uc)))
1091                 return;
1092
1093         err = register_for_each_vma(uprobe, NULL);
1094         /* TODO : cant unregister? schedule a worker thread */
1095         if (!uprobe->consumers && !err)
1096                 delete_uprobe(uprobe);
1097 }
1098
1099 /*
1100  * uprobe_unregister - unregister an already registered probe.
1101  * @inode: the file in which the probe has to be removed.
1102  * @offset: offset from the start of the file.
1103  * @uc: identify which probe if multiple probes are colocated.
1104  */
1105 void uprobe_unregister(struct inode *inode, loff_t offset, struct uprobe_consumer *uc)
1106 {
1107         struct uprobe *uprobe;
1108
1109         uprobe = find_uprobe(inode, offset);
1110         if (WARN_ON(!uprobe))
1111                 return;
1112
1113         down_write(&uprobe->register_rwsem);
1114         __uprobe_unregister(uprobe, uc);
1115         up_write(&uprobe->register_rwsem);
1116         put_uprobe(uprobe);
1117 }
1118 EXPORT_SYMBOL_GPL(uprobe_unregister);
1119
1120 /*
1121  * __uprobe_register - register a probe
1122  * @inode: the file in which the probe has to be placed.
1123  * @offset: offset from the start of the file.
1124  * @uc: information on howto handle the probe..
1125  *
1126  * Apart from the access refcount, __uprobe_register() takes a creation
1127  * refcount (thro alloc_uprobe) if and only if this @uprobe is getting
1128  * inserted into the rbtree (i.e first consumer for a @inode:@offset
1129  * tuple).  Creation refcount stops uprobe_unregister from freeing the
1130  * @uprobe even before the register operation is complete. Creation
1131  * refcount is released when the last @uc for the @uprobe
1132  * unregisters. Caller of __uprobe_register() is required to keep @inode
1133  * (and the containing mount) referenced.
1134  *
1135  * Return errno if it cannot successully install probes
1136  * else return 0 (success)
1137  */
1138 static int __uprobe_register(struct inode *inode, loff_t offset,
1139                              loff_t ref_ctr_offset, struct uprobe_consumer *uc)
1140 {
1141         struct uprobe *uprobe;
1142         int ret;
1143
1144         /* Uprobe must have at least one set consumer */
1145         if (!uc->handler && !uc->ret_handler)
1146                 return -EINVAL;
1147
1148         /* copy_insn() uses read_mapping_page() or shmem_read_mapping_page() */
1149         if (!inode->i_mapping->a_ops->read_folio &&
1150             !shmem_mapping(inode->i_mapping))
1151                 return -EIO;
1152         /* Racy, just to catch the obvious mistakes */
1153         if (offset > i_size_read(inode))
1154                 return -EINVAL;
1155
1156         /*
1157          * This ensures that copy_from_page(), copy_to_page() and
1158          * __update_ref_ctr() can't cross page boundary.
1159          */
1160         if (!IS_ALIGNED(offset, UPROBE_SWBP_INSN_SIZE))
1161                 return -EINVAL;
1162         if (!IS_ALIGNED(ref_ctr_offset, sizeof(short)))
1163                 return -EINVAL;
1164
1165  retry:
1166         uprobe = alloc_uprobe(inode, offset, ref_ctr_offset);
1167         if (!uprobe)
1168                 return -ENOMEM;
1169         if (IS_ERR(uprobe))
1170                 return PTR_ERR(uprobe);
1171
1172         /*
1173          * We can race with uprobe_unregister()->delete_uprobe().
1174          * Check uprobe_is_active() and retry if it is false.
1175          */
1176         down_write(&uprobe->register_rwsem);
1177         ret = -EAGAIN;
1178         if (likely(uprobe_is_active(uprobe))) {
1179                 consumer_add(uprobe, uc);
1180                 ret = register_for_each_vma(uprobe, uc);
1181                 if (ret)
1182                         __uprobe_unregister(uprobe, uc);
1183         }
1184         up_write(&uprobe->register_rwsem);
1185         put_uprobe(uprobe);
1186
1187         if (unlikely(ret == -EAGAIN))
1188                 goto retry;
1189         return ret;
1190 }
1191
1192 int uprobe_register(struct inode *inode, loff_t offset,
1193                     struct uprobe_consumer *uc)
1194 {
1195         return __uprobe_register(inode, offset, 0, uc);
1196 }
1197 EXPORT_SYMBOL_GPL(uprobe_register);
1198
1199 int uprobe_register_refctr(struct inode *inode, loff_t offset,
1200                            loff_t ref_ctr_offset, struct uprobe_consumer *uc)
1201 {
1202         return __uprobe_register(inode, offset, ref_ctr_offset, uc);
1203 }
1204 EXPORT_SYMBOL_GPL(uprobe_register_refctr);
1205
1206 /*
1207  * uprobe_apply - unregister an already registered probe.
1208  * @inode: the file in which the probe has to be removed.
1209  * @offset: offset from the start of the file.
1210  * @uc: consumer which wants to add more or remove some breakpoints
1211  * @add: add or remove the breakpoints
1212  */
1213 int uprobe_apply(struct inode *inode, loff_t offset,
1214                         struct uprobe_consumer *uc, bool add)
1215 {
1216         struct uprobe *uprobe;
1217         struct uprobe_consumer *con;
1218         int ret = -ENOENT;
1219
1220         uprobe = find_uprobe(inode, offset);
1221         if (WARN_ON(!uprobe))
1222                 return ret;
1223
1224         down_write(&uprobe->register_rwsem);
1225         for (con = uprobe->consumers; con && con != uc ; con = con->next)
1226                 ;
1227         if (con)
1228                 ret = register_for_each_vma(uprobe, add ? uc : NULL);
1229         up_write(&uprobe->register_rwsem);
1230         put_uprobe(uprobe);
1231
1232         return ret;
1233 }
1234
1235 static int unapply_uprobe(struct uprobe *uprobe, struct mm_struct *mm)
1236 {
1237         VMA_ITERATOR(vmi, mm, 0);
1238         struct vm_area_struct *vma;
1239         int err = 0;
1240
1241         mmap_read_lock(mm);
1242         for_each_vma(vmi, vma) {
1243                 unsigned long vaddr;
1244                 loff_t offset;
1245
1246                 if (!valid_vma(vma, false) ||
1247                     file_inode(vma->vm_file) != uprobe->inode)
1248                         continue;
1249
1250                 offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1251                 if (uprobe->offset <  offset ||
1252                     uprobe->offset >= offset + vma->vm_end - vma->vm_start)
1253                         continue;
1254
1255                 vaddr = offset_to_vaddr(vma, uprobe->offset);
1256                 err |= remove_breakpoint(uprobe, mm, vaddr);
1257         }
1258         mmap_read_unlock(mm);
1259
1260         return err;
1261 }
1262
1263 static struct rb_node *
1264 find_node_in_range(struct inode *inode, loff_t min, loff_t max)
1265 {
1266         struct rb_node *n = uprobes_tree.rb_node;
1267
1268         while (n) {
1269                 struct uprobe *u = rb_entry(n, struct uprobe, rb_node);
1270
1271                 if (inode < u->inode) {
1272                         n = n->rb_left;
1273                 } else if (inode > u->inode) {
1274                         n = n->rb_right;
1275                 } else {
1276                         if (max < u->offset)
1277                                 n = n->rb_left;
1278                         else if (min > u->offset)
1279                                 n = n->rb_right;
1280                         else
1281                                 break;
1282                 }
1283         }
1284
1285         return n;
1286 }
1287
1288 /*
1289  * For a given range in vma, build a list of probes that need to be inserted.
1290  */
1291 static void build_probe_list(struct inode *inode,
1292                                 struct vm_area_struct *vma,
1293                                 unsigned long start, unsigned long end,
1294                                 struct list_head *head)
1295 {
1296         loff_t min, max;
1297         struct rb_node *n, *t;
1298         struct uprobe *u;
1299
1300         INIT_LIST_HEAD(head);
1301         min = vaddr_to_offset(vma, start);
1302         max = min + (end - start) - 1;
1303
1304         spin_lock(&uprobes_treelock);
1305         n = find_node_in_range(inode, min, max);
1306         if (n) {
1307                 for (t = n; t; t = rb_prev(t)) {
1308                         u = rb_entry(t, struct uprobe, rb_node);
1309                         if (u->inode != inode || u->offset < min)
1310                                 break;
1311                         list_add(&u->pending_list, head);
1312                         get_uprobe(u);
1313                 }
1314                 for (t = n; (t = rb_next(t)); ) {
1315                         u = rb_entry(t, struct uprobe, rb_node);
1316                         if (u->inode != inode || u->offset > max)
1317                                 break;
1318                         list_add(&u->pending_list, head);
1319                         get_uprobe(u);
1320                 }
1321         }
1322         spin_unlock(&uprobes_treelock);
1323 }
1324
1325 /* @vma contains reference counter, not the probed instruction. */
1326 static int delayed_ref_ctr_inc(struct vm_area_struct *vma)
1327 {
1328         struct list_head *pos, *q;
1329         struct delayed_uprobe *du;
1330         unsigned long vaddr;
1331         int ret = 0, err = 0;
1332
1333         mutex_lock(&delayed_uprobe_lock);
1334         list_for_each_safe(pos, q, &delayed_uprobe_list) {
1335                 du = list_entry(pos, struct delayed_uprobe, list);
1336
1337                 if (du->mm != vma->vm_mm ||
1338                     !valid_ref_ctr_vma(du->uprobe, vma))
1339                         continue;
1340
1341                 vaddr = offset_to_vaddr(vma, du->uprobe->ref_ctr_offset);
1342                 ret = __update_ref_ctr(vma->vm_mm, vaddr, 1);
1343                 if (ret) {
1344                         update_ref_ctr_warn(du->uprobe, vma->vm_mm, 1);
1345                         if (!err)
1346                                 err = ret;
1347                 }
1348                 delayed_uprobe_delete(du);
1349         }
1350         mutex_unlock(&delayed_uprobe_lock);
1351         return err;
1352 }
1353
1354 /*
1355  * Called from mmap_region/vma_adjust with mm->mmap_lock acquired.
1356  *
1357  * Currently we ignore all errors and always return 0, the callers
1358  * can't handle the failure anyway.
1359  */
1360 int uprobe_mmap(struct vm_area_struct *vma)
1361 {
1362         struct list_head tmp_list;
1363         struct uprobe *uprobe, *u;
1364         struct inode *inode;
1365
1366         if (no_uprobe_events())
1367                 return 0;
1368
1369         if (vma->vm_file &&
1370             (vma->vm_flags & (VM_WRITE|VM_SHARED)) == VM_WRITE &&
1371             test_bit(MMF_HAS_UPROBES, &vma->vm_mm->flags))
1372                 delayed_ref_ctr_inc(vma);
1373
1374         if (!valid_vma(vma, true))
1375                 return 0;
1376
1377         inode = file_inode(vma->vm_file);
1378         if (!inode)
1379                 return 0;
1380
1381         mutex_lock(uprobes_mmap_hash(inode));
1382         build_probe_list(inode, vma, vma->vm_start, vma->vm_end, &tmp_list);
1383         /*
1384          * We can race with uprobe_unregister(), this uprobe can be already
1385          * removed. But in this case filter_chain() must return false, all
1386          * consumers have gone away.
1387          */
1388         list_for_each_entry_safe(uprobe, u, &tmp_list, pending_list) {
1389                 if (!fatal_signal_pending(current) &&
1390                     filter_chain(uprobe, UPROBE_FILTER_MMAP, vma->vm_mm)) {
1391                         unsigned long vaddr = offset_to_vaddr(vma, uprobe->offset);
1392                         install_breakpoint(uprobe, vma->vm_mm, vma, vaddr);
1393                 }
1394                 put_uprobe(uprobe);
1395         }
1396         mutex_unlock(uprobes_mmap_hash(inode));
1397
1398         return 0;
1399 }
1400
1401 static bool
1402 vma_has_uprobes(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1403 {
1404         loff_t min, max;
1405         struct inode *inode;
1406         struct rb_node *n;
1407
1408         inode = file_inode(vma->vm_file);
1409
1410         min = vaddr_to_offset(vma, start);
1411         max = min + (end - start) - 1;
1412
1413         spin_lock(&uprobes_treelock);
1414         n = find_node_in_range(inode, min, max);
1415         spin_unlock(&uprobes_treelock);
1416
1417         return !!n;
1418 }
1419
1420 /*
1421  * Called in context of a munmap of a vma.
1422  */
1423 void uprobe_munmap(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1424 {
1425         if (no_uprobe_events() || !valid_vma(vma, false))
1426                 return;
1427
1428         if (!atomic_read(&vma->vm_mm->mm_users)) /* called by mmput() ? */
1429                 return;
1430
1431         if (!test_bit(MMF_HAS_UPROBES, &vma->vm_mm->flags) ||
1432              test_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags))
1433                 return;
1434
1435         if (vma_has_uprobes(vma, start, end))
1436                 set_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags);
1437 }
1438
1439 /* Slot allocation for XOL */
1440 static int xol_add_vma(struct mm_struct *mm, struct xol_area *area)
1441 {
1442         struct vm_area_struct *vma;
1443         int ret;
1444
1445         if (mmap_write_lock_killable(mm))
1446                 return -EINTR;
1447
1448         if (mm->uprobes_state.xol_area) {
1449                 ret = -EALREADY;
1450                 goto fail;
1451         }
1452
1453         if (!area->vaddr) {
1454                 /* Try to map as high as possible, this is only a hint. */
1455                 area->vaddr = get_unmapped_area(NULL, TASK_SIZE - PAGE_SIZE,
1456                                                 PAGE_SIZE, 0, 0);
1457                 if (IS_ERR_VALUE(area->vaddr)) {
1458                         ret = area->vaddr;
1459                         goto fail;
1460                 }
1461         }
1462
1463         vma = _install_special_mapping(mm, area->vaddr, PAGE_SIZE,
1464                                 VM_EXEC|VM_MAYEXEC|VM_DONTCOPY|VM_IO,
1465                                 &area->xol_mapping);
1466         if (IS_ERR(vma)) {
1467                 ret = PTR_ERR(vma);
1468                 goto fail;
1469         }
1470
1471         ret = 0;
1472         /* pairs with get_xol_area() */
1473         smp_store_release(&mm->uprobes_state.xol_area, area); /* ^^^ */
1474  fail:
1475         mmap_write_unlock(mm);
1476
1477         return ret;
1478 }
1479
1480 static struct xol_area *__create_xol_area(unsigned long vaddr)
1481 {
1482         struct mm_struct *mm = current->mm;
1483         uprobe_opcode_t insn = UPROBE_SWBP_INSN;
1484         struct xol_area *area;
1485
1486         area = kmalloc(sizeof(*area), GFP_KERNEL);
1487         if (unlikely(!area))
1488                 goto out;
1489
1490         area->bitmap = kcalloc(BITS_TO_LONGS(UINSNS_PER_PAGE), sizeof(long),
1491                                GFP_KERNEL);
1492         if (!area->bitmap)
1493                 goto free_area;
1494
1495         area->xol_mapping.name = "[uprobes]";
1496         area->xol_mapping.fault = NULL;
1497         area->xol_mapping.pages = area->pages;
1498         area->pages[0] = alloc_page(GFP_HIGHUSER);
1499         if (!area->pages[0])
1500                 goto free_bitmap;
1501         area->pages[1] = NULL;
1502
1503         area->vaddr = vaddr;
1504         init_waitqueue_head(&area->wq);
1505         /* Reserve the 1st slot for get_trampoline_vaddr() */
1506         set_bit(0, area->bitmap);
1507         atomic_set(&area->slot_count, 1);
1508         arch_uprobe_copy_ixol(area->pages[0], 0, &insn, UPROBE_SWBP_INSN_SIZE);
1509
1510         if (!xol_add_vma(mm, area))
1511                 return area;
1512
1513         __free_page(area->pages[0]);
1514  free_bitmap:
1515         kfree(area->bitmap);
1516  free_area:
1517         kfree(area);
1518  out:
1519         return NULL;
1520 }
1521
1522 /*
1523  * get_xol_area - Allocate process's xol_area if necessary.
1524  * This area will be used for storing instructions for execution out of line.
1525  *
1526  * Returns the allocated area or NULL.
1527  */
1528 static struct xol_area *get_xol_area(void)
1529 {
1530         struct mm_struct *mm = current->mm;
1531         struct xol_area *area;
1532
1533         if (!mm->uprobes_state.xol_area)
1534                 __create_xol_area(0);
1535
1536         /* Pairs with xol_add_vma() smp_store_release() */
1537         area = READ_ONCE(mm->uprobes_state.xol_area); /* ^^^ */
1538         return area;
1539 }
1540
1541 /*
1542  * uprobe_clear_state - Free the area allocated for slots.
1543  */
1544 void uprobe_clear_state(struct mm_struct *mm)
1545 {
1546         struct xol_area *area = mm->uprobes_state.xol_area;
1547
1548         mutex_lock(&delayed_uprobe_lock);
1549         delayed_uprobe_remove(NULL, mm);
1550         mutex_unlock(&delayed_uprobe_lock);
1551
1552         if (!area)
1553                 return;
1554
1555         put_page(area->pages[0]);
1556         kfree(area->bitmap);
1557         kfree(area);
1558 }
1559
1560 void uprobe_start_dup_mmap(void)
1561 {
1562         percpu_down_read(&dup_mmap_sem);
1563 }
1564
1565 void uprobe_end_dup_mmap(void)
1566 {
1567         percpu_up_read(&dup_mmap_sem);
1568 }
1569
1570 void uprobe_dup_mmap(struct mm_struct *oldmm, struct mm_struct *newmm)
1571 {
1572         if (test_bit(MMF_HAS_UPROBES, &oldmm->flags)) {
1573                 set_bit(MMF_HAS_UPROBES, &newmm->flags);
1574                 /* unconditionally, dup_mmap() skips VM_DONTCOPY vmas */
1575                 set_bit(MMF_RECALC_UPROBES, &newmm->flags);
1576         }
1577 }
1578
1579 /*
1580  *  - search for a free slot.
1581  */
1582 static unsigned long xol_take_insn_slot(struct xol_area *area)
1583 {
1584         unsigned long slot_addr;
1585         int slot_nr;
1586
1587         do {
1588                 slot_nr = find_first_zero_bit(area->bitmap, UINSNS_PER_PAGE);
1589                 if (slot_nr < UINSNS_PER_PAGE) {
1590                         if (!test_and_set_bit(slot_nr, area->bitmap))
1591                                 break;
1592
1593                         slot_nr = UINSNS_PER_PAGE;
1594                         continue;
1595                 }
1596                 wait_event(area->wq, (atomic_read(&area->slot_count) < UINSNS_PER_PAGE));
1597         } while (slot_nr >= UINSNS_PER_PAGE);
1598
1599         slot_addr = area->vaddr + (slot_nr * UPROBE_XOL_SLOT_BYTES);
1600         atomic_inc(&area->slot_count);
1601
1602         return slot_addr;
1603 }
1604
1605 /*
1606  * xol_get_insn_slot - allocate a slot for xol.
1607  * Returns the allocated slot address or 0.
1608  */
1609 static unsigned long xol_get_insn_slot(struct uprobe *uprobe)
1610 {
1611         struct xol_area *area;
1612         unsigned long xol_vaddr;
1613
1614         area = get_xol_area();
1615         if (!area)
1616                 return 0;
1617
1618         xol_vaddr = xol_take_insn_slot(area);
1619         if (unlikely(!xol_vaddr))
1620                 return 0;
1621
1622         arch_uprobe_copy_ixol(area->pages[0], xol_vaddr,
1623                               &uprobe->arch.ixol, sizeof(uprobe->arch.ixol));
1624
1625         return xol_vaddr;
1626 }
1627
1628 /*
1629  * xol_free_insn_slot - If slot was earlier allocated by
1630  * @xol_get_insn_slot(), make the slot available for
1631  * subsequent requests.
1632  */
1633 static void xol_free_insn_slot(struct task_struct *tsk)
1634 {
1635         struct xol_area *area;
1636         unsigned long vma_end;
1637         unsigned long slot_addr;
1638
1639         if (!tsk->mm || !tsk->mm->uprobes_state.xol_area || !tsk->utask)
1640                 return;
1641
1642         slot_addr = tsk->utask->xol_vaddr;
1643         if (unlikely(!slot_addr))
1644                 return;
1645
1646         area = tsk->mm->uprobes_state.xol_area;
1647         vma_end = area->vaddr + PAGE_SIZE;
1648         if (area->vaddr <= slot_addr && slot_addr < vma_end) {
1649                 unsigned long offset;
1650                 int slot_nr;
1651
1652                 offset = slot_addr - area->vaddr;
1653                 slot_nr = offset / UPROBE_XOL_SLOT_BYTES;
1654                 if (slot_nr >= UINSNS_PER_PAGE)
1655                         return;
1656
1657                 clear_bit(slot_nr, area->bitmap);
1658                 atomic_dec(&area->slot_count);
1659                 smp_mb__after_atomic(); /* pairs with prepare_to_wait() */
1660                 if (waitqueue_active(&area->wq))
1661                         wake_up(&area->wq);
1662
1663                 tsk->utask->xol_vaddr = 0;
1664         }
1665 }
1666
1667 void __weak arch_uprobe_copy_ixol(struct page *page, unsigned long vaddr,
1668                                   void *src, unsigned long len)
1669 {
1670         /* Initialize the slot */
1671         copy_to_page(page, vaddr, src, len);
1672
1673         /*
1674          * We probably need flush_icache_user_page() but it needs vma.
1675          * This should work on most of architectures by default. If
1676          * architecture needs to do something different it can define
1677          * its own version of the function.
1678          */
1679         flush_dcache_page(page);
1680 }
1681
1682 /**
1683  * uprobe_get_swbp_addr - compute address of swbp given post-swbp regs
1684  * @regs: Reflects the saved state of the task after it has hit a breakpoint
1685  * instruction.
1686  * Return the address of the breakpoint instruction.
1687  */
1688 unsigned long __weak uprobe_get_swbp_addr(struct pt_regs *regs)
1689 {
1690         return instruction_pointer(regs) - UPROBE_SWBP_INSN_SIZE;
1691 }
1692
1693 unsigned long uprobe_get_trap_addr(struct pt_regs *regs)
1694 {
1695         struct uprobe_task *utask = current->utask;
1696
1697         if (unlikely(utask && utask->active_uprobe))
1698                 return utask->vaddr;
1699
1700         return instruction_pointer(regs);
1701 }
1702
1703 static struct return_instance *free_ret_instance(struct return_instance *ri)
1704 {
1705         struct return_instance *next = ri->next;
1706         put_uprobe(ri->uprobe);
1707         kfree(ri);
1708         return next;
1709 }
1710
1711 /*
1712  * Called with no locks held.
1713  * Called in context of an exiting or an exec-ing thread.
1714  */
1715 void uprobe_free_utask(struct task_struct *t)
1716 {
1717         struct uprobe_task *utask = t->utask;
1718         struct return_instance *ri;
1719
1720         if (!utask)
1721                 return;
1722
1723         if (utask->active_uprobe)
1724                 put_uprobe(utask->active_uprobe);
1725
1726         ri = utask->return_instances;
1727         while (ri)
1728                 ri = free_ret_instance(ri);
1729
1730         xol_free_insn_slot(t);
1731         kfree(utask);
1732         t->utask = NULL;
1733 }
1734
1735 /*
1736  * Allocate a uprobe_task object for the task if necessary.
1737  * Called when the thread hits a breakpoint.
1738  *
1739  * Returns:
1740  * - pointer to new uprobe_task on success
1741  * - NULL otherwise
1742  */
1743 static struct uprobe_task *get_utask(void)
1744 {
1745         if (!current->utask)
1746                 current->utask = kzalloc(sizeof(struct uprobe_task), GFP_KERNEL);
1747         return current->utask;
1748 }
1749
1750 static int dup_utask(struct task_struct *t, struct uprobe_task *o_utask)
1751 {
1752         struct uprobe_task *n_utask;
1753         struct return_instance **p, *o, *n;
1754
1755         n_utask = kzalloc(sizeof(struct uprobe_task), GFP_KERNEL);
1756         if (!n_utask)
1757                 return -ENOMEM;
1758         t->utask = n_utask;
1759
1760         p = &n_utask->return_instances;
1761         for (o = o_utask->return_instances; o; o = o->next) {
1762                 n = kmalloc(sizeof(struct return_instance), GFP_KERNEL);
1763                 if (!n)
1764                         return -ENOMEM;
1765
1766                 *n = *o;
1767                 get_uprobe(n->uprobe);
1768                 n->next = NULL;
1769
1770                 *p = n;
1771                 p = &n->next;
1772                 n_utask->depth++;
1773         }
1774
1775         return 0;
1776 }
1777
1778 static void uprobe_warn(struct task_struct *t, const char *msg)
1779 {
1780         pr_warn("uprobe: %s:%d failed to %s\n",
1781                         current->comm, current->pid, msg);
1782 }
1783
1784 static void dup_xol_work(struct callback_head *work)
1785 {
1786         if (current->flags & PF_EXITING)
1787                 return;
1788
1789         if (!__create_xol_area(current->utask->dup_xol_addr) &&
1790                         !fatal_signal_pending(current))
1791                 uprobe_warn(current, "dup xol area");
1792 }
1793
1794 /*
1795  * Called in context of a new clone/fork from copy_process.
1796  */
1797 void uprobe_copy_process(struct task_struct *t, unsigned long flags)
1798 {
1799         struct uprobe_task *utask = current->utask;
1800         struct mm_struct *mm = current->mm;
1801         struct xol_area *area;
1802
1803         t->utask = NULL;
1804
1805         if (!utask || !utask->return_instances)
1806                 return;
1807
1808         if (mm == t->mm && !(flags & CLONE_VFORK))
1809                 return;
1810
1811         if (dup_utask(t, utask))
1812                 return uprobe_warn(t, "dup ret instances");
1813
1814         /* The task can fork() after dup_xol_work() fails */
1815         area = mm->uprobes_state.xol_area;
1816         if (!area)
1817                 return uprobe_warn(t, "dup xol area");
1818
1819         if (mm == t->mm)
1820                 return;
1821
1822         t->utask->dup_xol_addr = area->vaddr;
1823         init_task_work(&t->utask->dup_xol_work, dup_xol_work);
1824         task_work_add(t, &t->utask->dup_xol_work, TWA_RESUME);
1825 }
1826
1827 /*
1828  * Current area->vaddr notion assume the trampoline address is always
1829  * equal area->vaddr.
1830  *
1831  * Returns -1 in case the xol_area is not allocated.
1832  */
1833 static unsigned long get_trampoline_vaddr(void)
1834 {
1835         struct xol_area *area;
1836         unsigned long trampoline_vaddr = -1;
1837
1838         /* Pairs with xol_add_vma() smp_store_release() */
1839         area = READ_ONCE(current->mm->uprobes_state.xol_area); /* ^^^ */
1840         if (area)
1841                 trampoline_vaddr = area->vaddr;
1842
1843         return trampoline_vaddr;
1844 }
1845
1846 static void cleanup_return_instances(struct uprobe_task *utask, bool chained,
1847                                         struct pt_regs *regs)
1848 {
1849         struct return_instance *ri = utask->return_instances;
1850         enum rp_check ctx = chained ? RP_CHECK_CHAIN_CALL : RP_CHECK_CALL;
1851
1852         while (ri && !arch_uretprobe_is_alive(ri, ctx, regs)) {
1853                 ri = free_ret_instance(ri);
1854                 utask->depth--;
1855         }
1856         utask->return_instances = ri;
1857 }
1858
1859 static void prepare_uretprobe(struct uprobe *uprobe, struct pt_regs *regs)
1860 {
1861         struct return_instance *ri;
1862         struct uprobe_task *utask;
1863         unsigned long orig_ret_vaddr, trampoline_vaddr;
1864         bool chained;
1865
1866         if (!get_xol_area())
1867                 return;
1868
1869         utask = get_utask();
1870         if (!utask)
1871                 return;
1872
1873         if (utask->depth >= MAX_URETPROBE_DEPTH) {
1874                 printk_ratelimited(KERN_INFO "uprobe: omit uretprobe due to"
1875                                 " nestedness limit pid/tgid=%d/%d\n",
1876                                 current->pid, current->tgid);
1877                 return;
1878         }
1879
1880         ri = kmalloc(sizeof(struct return_instance), GFP_KERNEL);
1881         if (!ri)
1882                 return;
1883
1884         trampoline_vaddr = get_trampoline_vaddr();
1885         orig_ret_vaddr = arch_uretprobe_hijack_return_addr(trampoline_vaddr, regs);
1886         if (orig_ret_vaddr == -1)
1887                 goto fail;
1888
1889         /* drop the entries invalidated by longjmp() */
1890         chained = (orig_ret_vaddr == trampoline_vaddr);
1891         cleanup_return_instances(utask, chained, regs);
1892
1893         /*
1894          * We don't want to keep trampoline address in stack, rather keep the
1895          * original return address of first caller thru all the consequent
1896          * instances. This also makes breakpoint unwrapping easier.
1897          */
1898         if (chained) {
1899                 if (!utask->return_instances) {
1900                         /*
1901                          * This situation is not possible. Likely we have an
1902                          * attack from user-space.
1903                          */
1904                         uprobe_warn(current, "handle tail call");
1905                         goto fail;
1906                 }
1907                 orig_ret_vaddr = utask->return_instances->orig_ret_vaddr;
1908         }
1909
1910         ri->uprobe = get_uprobe(uprobe);
1911         ri->func = instruction_pointer(regs);
1912         ri->stack = user_stack_pointer(regs);
1913         ri->orig_ret_vaddr = orig_ret_vaddr;
1914         ri->chained = chained;
1915
1916         utask->depth++;
1917         ri->next = utask->return_instances;
1918         utask->return_instances = ri;
1919
1920         return;
1921  fail:
1922         kfree(ri);
1923 }
1924
1925 /* Prepare to single-step probed instruction out of line. */
1926 static int
1927 pre_ssout(struct uprobe *uprobe, struct pt_regs *regs, unsigned long bp_vaddr)
1928 {
1929         struct uprobe_task *utask;
1930         unsigned long xol_vaddr;
1931         int err;
1932
1933         utask = get_utask();
1934         if (!utask)
1935                 return -ENOMEM;
1936
1937         xol_vaddr = xol_get_insn_slot(uprobe);
1938         if (!xol_vaddr)
1939                 return -ENOMEM;
1940
1941         utask->xol_vaddr = xol_vaddr;
1942         utask->vaddr = bp_vaddr;
1943
1944         err = arch_uprobe_pre_xol(&uprobe->arch, regs);
1945         if (unlikely(err)) {
1946                 xol_free_insn_slot(current);
1947                 return err;
1948         }
1949
1950         utask->active_uprobe = uprobe;
1951         utask->state = UTASK_SSTEP;
1952         return 0;
1953 }
1954
1955 /*
1956  * If we are singlestepping, then ensure this thread is not connected to
1957  * non-fatal signals until completion of singlestep.  When xol insn itself
1958  * triggers the signal,  restart the original insn even if the task is
1959  * already SIGKILL'ed (since coredump should report the correct ip).  This
1960  * is even more important if the task has a handler for SIGSEGV/etc, The
1961  * _same_ instruction should be repeated again after return from the signal
1962  * handler, and SSTEP can never finish in this case.
1963  */
1964 bool uprobe_deny_signal(void)
1965 {
1966         struct task_struct *t = current;
1967         struct uprobe_task *utask = t->utask;
1968
1969         if (likely(!utask || !utask->active_uprobe))
1970                 return false;
1971
1972         WARN_ON_ONCE(utask->state != UTASK_SSTEP);
1973
1974         if (task_sigpending(t)) {
1975                 spin_lock_irq(&t->sighand->siglock);
1976                 clear_tsk_thread_flag(t, TIF_SIGPENDING);
1977                 spin_unlock_irq(&t->sighand->siglock);
1978
1979                 if (__fatal_signal_pending(t) || arch_uprobe_xol_was_trapped(t)) {
1980                         utask->state = UTASK_SSTEP_TRAPPED;
1981                         set_tsk_thread_flag(t, TIF_UPROBE);
1982                 }
1983         }
1984
1985         return true;
1986 }
1987
1988 static void mmf_recalc_uprobes(struct mm_struct *mm)
1989 {
1990         VMA_ITERATOR(vmi, mm, 0);
1991         struct vm_area_struct *vma;
1992
1993         for_each_vma(vmi, vma) {
1994                 if (!valid_vma(vma, false))
1995                         continue;
1996                 /*
1997                  * This is not strictly accurate, we can race with
1998                  * uprobe_unregister() and see the already removed
1999                  * uprobe if delete_uprobe() was not yet called.
2000                  * Or this uprobe can be filtered out.
2001                  */
2002                 if (vma_has_uprobes(vma, vma->vm_start, vma->vm_end))
2003                         return;
2004         }
2005
2006         clear_bit(MMF_HAS_UPROBES, &mm->flags);
2007 }
2008
2009 static int is_trap_at_addr(struct mm_struct *mm, unsigned long vaddr)
2010 {
2011         struct page *page;
2012         uprobe_opcode_t opcode;
2013         int result;
2014
2015         if (WARN_ON_ONCE(!IS_ALIGNED(vaddr, UPROBE_SWBP_INSN_SIZE)))
2016                 return -EINVAL;
2017
2018         pagefault_disable();
2019         result = __get_user(opcode, (uprobe_opcode_t __user *)vaddr);
2020         pagefault_enable();
2021
2022         if (likely(result == 0))
2023                 goto out;
2024
2025         /*
2026          * The NULL 'tsk' here ensures that any faults that occur here
2027          * will not be accounted to the task.  'mm' *is* current->mm,
2028          * but we treat this as a 'remote' access since it is
2029          * essentially a kernel access to the memory.
2030          */
2031         result = get_user_pages_remote(mm, vaddr, 1, FOLL_FORCE, &page,
2032                         NULL, NULL);
2033         if (result < 0)
2034                 return result;
2035
2036         copy_from_page(page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
2037         put_page(page);
2038  out:
2039         /* This needs to return true for any variant of the trap insn */
2040         return is_trap_insn(&opcode);
2041 }
2042
2043 static struct uprobe *find_active_uprobe(unsigned long bp_vaddr, int *is_swbp)
2044 {
2045         struct mm_struct *mm = current->mm;
2046         struct uprobe *uprobe = NULL;
2047         struct vm_area_struct *vma;
2048
2049         mmap_read_lock(mm);
2050         vma = vma_lookup(mm, bp_vaddr);
2051         if (vma) {
2052                 if (valid_vma(vma, false)) {
2053                         struct inode *inode = file_inode(vma->vm_file);
2054                         loff_t offset = vaddr_to_offset(vma, bp_vaddr);
2055
2056                         uprobe = find_uprobe(inode, offset);
2057                 }
2058
2059                 if (!uprobe)
2060                         *is_swbp = is_trap_at_addr(mm, bp_vaddr);
2061         } else {
2062                 *is_swbp = -EFAULT;
2063         }
2064
2065         if (!uprobe && test_and_clear_bit(MMF_RECALC_UPROBES, &mm->flags))
2066                 mmf_recalc_uprobes(mm);
2067         mmap_read_unlock(mm);
2068
2069         return uprobe;
2070 }
2071
2072 static void handler_chain(struct uprobe *uprobe, struct pt_regs *regs)
2073 {
2074         struct uprobe_consumer *uc;
2075         int remove = UPROBE_HANDLER_REMOVE;
2076         bool need_prep = false; /* prepare return uprobe, when needed */
2077
2078         down_read(&uprobe->register_rwsem);
2079         for (uc = uprobe->consumers; uc; uc = uc->next) {
2080                 int rc = 0;
2081
2082                 if (uc->handler) {
2083                         rc = uc->handler(uc, regs);
2084                         WARN(rc & ~UPROBE_HANDLER_MASK,
2085                                 "bad rc=0x%x from %ps()\n", rc, uc->handler);
2086                 }
2087
2088                 if (uc->ret_handler)
2089                         need_prep = true;
2090
2091                 remove &= rc;
2092         }
2093
2094         if (need_prep && !remove)
2095                 prepare_uretprobe(uprobe, regs); /* put bp at return */
2096
2097         if (remove && uprobe->consumers) {
2098                 WARN_ON(!uprobe_is_active(uprobe));
2099                 unapply_uprobe(uprobe, current->mm);
2100         }
2101         up_read(&uprobe->register_rwsem);
2102 }
2103
2104 static void
2105 handle_uretprobe_chain(struct return_instance *ri, struct pt_regs *regs)
2106 {
2107         struct uprobe *uprobe = ri->uprobe;
2108         struct uprobe_consumer *uc;
2109
2110         down_read(&uprobe->register_rwsem);
2111         for (uc = uprobe->consumers; uc; uc = uc->next) {
2112                 if (uc->ret_handler)
2113                         uc->ret_handler(uc, ri->func, regs);
2114         }
2115         up_read(&uprobe->register_rwsem);
2116 }
2117
2118 static struct return_instance *find_next_ret_chain(struct return_instance *ri)
2119 {
2120         bool chained;
2121
2122         do {
2123                 chained = ri->chained;
2124                 ri = ri->next;  /* can't be NULL if chained */
2125         } while (chained);
2126
2127         return ri;
2128 }
2129
2130 static void handle_trampoline(struct pt_regs *regs)
2131 {
2132         struct uprobe_task *utask;
2133         struct return_instance *ri, *next;
2134         bool valid;
2135
2136         utask = current->utask;
2137         if (!utask)
2138                 goto sigill;
2139
2140         ri = utask->return_instances;
2141         if (!ri)
2142                 goto sigill;
2143
2144         do {
2145                 /*
2146                  * We should throw out the frames invalidated by longjmp().
2147                  * If this chain is valid, then the next one should be alive
2148                  * or NULL; the latter case means that nobody but ri->func
2149                  * could hit this trampoline on return. TODO: sigaltstack().
2150                  */
2151                 next = find_next_ret_chain(ri);
2152                 valid = !next || arch_uretprobe_is_alive(next, RP_CHECK_RET, regs);
2153
2154                 instruction_pointer_set(regs, ri->orig_ret_vaddr);
2155                 do {
2156                         if (valid)
2157                                 handle_uretprobe_chain(ri, regs);
2158                         ri = free_ret_instance(ri);
2159                         utask->depth--;
2160                 } while (ri != next);
2161         } while (!valid);
2162
2163         utask->return_instances = ri;
2164         return;
2165
2166  sigill:
2167         uprobe_warn(current, "handle uretprobe, sending SIGILL.");
2168         force_sig(SIGILL);
2169
2170 }
2171
2172 bool __weak arch_uprobe_ignore(struct arch_uprobe *aup, struct pt_regs *regs)
2173 {
2174         return false;
2175 }
2176
2177 bool __weak arch_uretprobe_is_alive(struct return_instance *ret, enum rp_check ctx,
2178                                         struct pt_regs *regs)
2179 {
2180         return true;
2181 }
2182
2183 /*
2184  * Run handler and ask thread to singlestep.
2185  * Ensure all non-fatal signals cannot interrupt thread while it singlesteps.
2186  */
2187 static void handle_swbp(struct pt_regs *regs)
2188 {
2189         struct uprobe *uprobe;
2190         unsigned long bp_vaddr;
2191         int is_swbp;
2192
2193         bp_vaddr = uprobe_get_swbp_addr(regs);
2194         if (bp_vaddr == get_trampoline_vaddr())
2195                 return handle_trampoline(regs);
2196
2197         uprobe = find_active_uprobe(bp_vaddr, &is_swbp);
2198         if (!uprobe) {
2199                 if (is_swbp > 0) {
2200                         /* No matching uprobe; signal SIGTRAP. */
2201                         force_sig(SIGTRAP);
2202                 } else {
2203                         /*
2204                          * Either we raced with uprobe_unregister() or we can't
2205                          * access this memory. The latter is only possible if
2206                          * another thread plays with our ->mm. In both cases
2207                          * we can simply restart. If this vma was unmapped we
2208                          * can pretend this insn was not executed yet and get
2209                          * the (correct) SIGSEGV after restart.
2210                          */
2211                         instruction_pointer_set(regs, bp_vaddr);
2212                 }
2213                 return;
2214         }
2215
2216         /* change it in advance for ->handler() and restart */
2217         instruction_pointer_set(regs, bp_vaddr);
2218
2219         /*
2220          * TODO: move copy_insn/etc into _register and remove this hack.
2221          * After we hit the bp, _unregister + _register can install the
2222          * new and not-yet-analyzed uprobe at the same address, restart.
2223          */
2224         if (unlikely(!test_bit(UPROBE_COPY_INSN, &uprobe->flags)))
2225                 goto out;
2226
2227         /*
2228          * Pairs with the smp_wmb() in prepare_uprobe().
2229          *
2230          * Guarantees that if we see the UPROBE_COPY_INSN bit set, then
2231          * we must also see the stores to &uprobe->arch performed by the
2232          * prepare_uprobe() call.
2233          */
2234         smp_rmb();
2235
2236         /* Tracing handlers use ->utask to communicate with fetch methods */
2237         if (!get_utask())
2238                 goto out;
2239
2240         if (arch_uprobe_ignore(&uprobe->arch, regs))
2241                 goto out;
2242
2243         handler_chain(uprobe, regs);
2244
2245         if (arch_uprobe_skip_sstep(&uprobe->arch, regs))
2246                 goto out;
2247
2248         if (!pre_ssout(uprobe, regs, bp_vaddr))
2249                 return;
2250
2251         /* arch_uprobe_skip_sstep() succeeded, or restart if can't singlestep */
2252 out:
2253         put_uprobe(uprobe);
2254 }
2255
2256 /*
2257  * Perform required fix-ups and disable singlestep.
2258  * Allow pending signals to take effect.
2259  */
2260 static void handle_singlestep(struct uprobe_task *utask, struct pt_regs *regs)
2261 {
2262         struct uprobe *uprobe;
2263         int err = 0;
2264
2265         uprobe = utask->active_uprobe;
2266         if (utask->state == UTASK_SSTEP_ACK)
2267                 err = arch_uprobe_post_xol(&uprobe->arch, regs);
2268         else if (utask->state == UTASK_SSTEP_TRAPPED)
2269                 arch_uprobe_abort_xol(&uprobe->arch, regs);
2270         else
2271                 WARN_ON_ONCE(1);
2272
2273         put_uprobe(uprobe);
2274         utask->active_uprobe = NULL;
2275         utask->state = UTASK_RUNNING;
2276         xol_free_insn_slot(current);
2277
2278         spin_lock_irq(&current->sighand->siglock);
2279         recalc_sigpending(); /* see uprobe_deny_signal() */
2280         spin_unlock_irq(&current->sighand->siglock);
2281
2282         if (unlikely(err)) {
2283                 uprobe_warn(current, "execute the probed insn, sending SIGILL.");
2284                 force_sig(SIGILL);
2285         }
2286 }
2287
2288 /*
2289  * On breakpoint hit, breakpoint notifier sets the TIF_UPROBE flag and
2290  * allows the thread to return from interrupt. After that handle_swbp()
2291  * sets utask->active_uprobe.
2292  *
2293  * On singlestep exception, singlestep notifier sets the TIF_UPROBE flag
2294  * and allows the thread to return from interrupt.
2295  *
2296  * While returning to userspace, thread notices the TIF_UPROBE flag and calls
2297  * uprobe_notify_resume().
2298  */
2299 void uprobe_notify_resume(struct pt_regs *regs)
2300 {
2301         struct uprobe_task *utask;
2302
2303         clear_thread_flag(TIF_UPROBE);
2304
2305         utask = current->utask;
2306         if (utask && utask->active_uprobe)
2307                 handle_singlestep(utask, regs);
2308         else
2309                 handle_swbp(regs);
2310 }
2311
2312 /*
2313  * uprobe_pre_sstep_notifier gets called from interrupt context as part of
2314  * notifier mechanism. Set TIF_UPROBE flag and indicate breakpoint hit.
2315  */
2316 int uprobe_pre_sstep_notifier(struct pt_regs *regs)
2317 {
2318         if (!current->mm)
2319                 return 0;
2320
2321         if (!test_bit(MMF_HAS_UPROBES, &current->mm->flags) &&
2322             (!current->utask || !current->utask->return_instances))
2323                 return 0;
2324
2325         set_thread_flag(TIF_UPROBE);
2326         return 1;
2327 }
2328
2329 /*
2330  * uprobe_post_sstep_notifier gets called in interrupt context as part of notifier
2331  * mechanism. Set TIF_UPROBE flag and indicate completion of singlestep.
2332  */
2333 int uprobe_post_sstep_notifier(struct pt_regs *regs)
2334 {
2335         struct uprobe_task *utask = current->utask;
2336
2337         if (!current->mm || !utask || !utask->active_uprobe)
2338                 /* task is currently not uprobed */
2339                 return 0;
2340
2341         utask->state = UTASK_SSTEP_ACK;
2342         set_thread_flag(TIF_UPROBE);
2343         return 1;
2344 }
2345
2346 static struct notifier_block uprobe_exception_nb = {
2347         .notifier_call          = arch_uprobe_exception_notify,
2348         .priority               = INT_MAX-1,    /* notified after kprobes, kgdb */
2349 };
2350
2351 void __init uprobes_init(void)
2352 {
2353         int i;
2354
2355         for (i = 0; i < UPROBES_HASH_SZ; i++)
2356                 mutex_init(&uprobes_mmap_mutex[i]);
2357
2358         BUG_ON(register_die_notifier(&uprobe_exception_nb));
2359 }