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