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>         /* try_to_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 mm_struct *mm = vma->vm_mm;
 158         struct page_vma_mapped_walk pvmw = {
 159                 .page = compound_head(old_page),
 160                 .vma = vma,
 161                 .address = addr,
 162         };
 163         int err;
 164         struct mmu_notifier_range range;
 165         struct mem_cgroup *memcg;
 166
 167         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr,
 168                                 addr + PAGE_SIZE);
 169
 170         if (new_page) {
 171                 err = mem_cgroup_try_charge(new_page, vma->vm_mm, GFP_KERNEL,
 172                                             &memcg, false);
 173                 if (err)
 174                         return err;
 175         }
 176
 177         /* For try_to_free_swap() and munlock_vma_page() below */
 178         lock_page(old_page);
 179
 180         mmu_notifier_invalidate_range_start(&range);
 181         err = -EAGAIN;
 182         if (!page_vma_mapped_walk(&pvmw)) {
 183                 if (new_page)
 184                         mem_cgroup_cancel_charge(new_page, memcg, false);
 185                 goto unlock;
 186         }
 187         VM_BUG_ON_PAGE(addr != pvmw.address, old_page);
 188
 189         if (new_page) {
 190                 get_page(new_page);
 191                 page_add_new_anon_rmap(new_page, vma, addr, false);
 192                 mem_cgroup_commit_charge(new_page, memcg, false, false);
 193                 lru_cache_add_active_or_unevictable(new_page, vma);
 194         } else
 195                 /* no new page, just dec_mm_counter for old_page */
 196                 dec_mm_counter(mm, MM_ANONPAGES);
 197
 198         if (!PageAnon(old_page)) {
 199                 dec_mm_counter(mm, mm_counter_file(old_page));
 200                 inc_mm_counter(mm, MM_ANONPAGES);
 201         }
 202
 203         flush_cache_page(vma, addr, pte_pfn(*pvmw.pte));
 204         ptep_clear_flush_notify(vma, addr, pvmw.pte);
 205         if (new_page)
 206                 set_pte_at_notify(mm, addr, pvmw.pte,
 207                                   mk_pte(new_page, vma->vm_page_prot));
 208
 209         page_remove_rmap(old_page, false);
 210         if (!page_mapped(old_page))
 211                 try_to_free_swap(old_page);
 212         page_vma_mapped_walk_done(&pvmw);
 213
 214         if ((vma->vm_flags & VM_LOCKED) && !PageCompound(old_page))
 215                 munlock_vma_page(old_page);
 216         put_page(old_page);
 217
 218         err = 0;
 219  unlock:
 220         mmu_notifier_invalidate_range_end(&range);
 221         unlock_page(old_page);
 222         return err;
 223 }
 224
 225 /**
 226  * is_swbp_insn - check if instruction is breakpoint instruction.
 227  * @insn: instruction to be checked.
 228  * Default implementation of is_swbp_insn
 229  * Returns true if @insn is a breakpoint instruction.
 230  */
 231 bool __weak is_swbp_insn(uprobe_opcode_t *insn)
 232 {
 233         return *insn == UPROBE_SWBP_INSN;
 234 }
 235
 236 /**
 237  * is_trap_insn - check if instruction is breakpoint instruction.
 238  * @insn: instruction to be checked.
 239  * Default implementation of is_trap_insn
 240  * Returns true if @insn is a breakpoint instruction.
 241  *
 242  * This function is needed for the case where an architecture has multiple
 243  * trap instructions (like powerpc).
 244  */
 245 bool __weak is_trap_insn(uprobe_opcode_t *insn)
 246 {
 247         return is_swbp_insn(insn);
 248 }
 249
 250 static void copy_from_page(struct page *page, unsigned long vaddr, void *dst, int len)
 251 {
 252         void *kaddr = kmap_atomic(page);
 253         memcpy(dst, kaddr + (vaddr & ~PAGE_MASK), len);
 254         kunmap_atomic(kaddr);
 255 }
 256
 257 static void copy_to_page(struct page *page, unsigned long vaddr, const void *src, int len)
 258 {
 259         void *kaddr = kmap_atomic(page);
 260         memcpy(kaddr + (vaddr & ~PAGE_MASK), src, len);
 261         kunmap_atomic(kaddr);
 262 }
 263
 264 static int verify_opcode(struct page *page, unsigned long vaddr, uprobe_opcode_t *new_opcode)
 265 {
 266         uprobe_opcode_t old_opcode;
 267         bool is_swbp;
 268
 269         /*
 270          * Note: We only check if the old_opcode is UPROBE_SWBP_INSN here.
 271          * We do not check if it is any other 'trap variant' which could
 272          * be conditional trap instruction such as the one powerpc supports.
 273          *
 274          * The logic is that we do not care if the underlying instruction
 275          * is a trap variant; uprobes always wins over any other (gdb)
 276          * breakpoint.
 277          */
 278         copy_from_page(page, vaddr, &old_opcode, UPROBE_SWBP_INSN_SIZE);
 279         is_swbp = is_swbp_insn(&old_opcode);
 280
 281         if (is_swbp_insn(new_opcode)) {
 282                 if (is_swbp)            /* register: already installed? */
 283                         return 0;
 284         } else {
 285                 if (!is_swbp)           /* unregister: was it changed by us? */
 286                         return 0;
 287         }
 288
 289         return 1;
 290 }
 291
 292 static struct delayed_uprobe *
 293 delayed_uprobe_check(struct uprobe *uprobe, struct mm_struct *mm)
 294 {
 295         struct delayed_uprobe *du;
 296
 297         list_for_each_entry(du, &delayed_uprobe_list, list)
 298                 if (du->uprobe == uprobe && du->mm == mm)
 299                         return du;
 300         return NULL;
 301 }
 302
 303 static int delayed_uprobe_add(struct uprobe *uprobe, struct mm_struct *mm)
 304 {
 305         struct delayed_uprobe *du;
 306
 307         if (delayed_uprobe_check(uprobe, mm))
 308                 return 0;
 309
 310         du  = kzalloc(sizeof(*du), GFP_KERNEL);
 311         if (!du)
 312                 return -ENOMEM;
 313
 314         du->uprobe = uprobe;
 315         du->mm = mm;
 316         list_add(&du->list, &delayed_uprobe_list);
 317         return 0;
 318 }
 319
 320 static void delayed_uprobe_delete(struct delayed_uprobe *du)
 321 {
 322         if (WARN_ON(!du))
 323                 return;
 324         list_del(&du->list);
 325         kfree(du);
 326 }
 327
 328 static void delayed_uprobe_remove(struct uprobe *uprobe, struct mm_struct *mm)
 329 {
 330         struct list_head *pos, *q;
 331         struct delayed_uprobe *du;
 332
 333         if (!uprobe && !mm)
 334                 return;
 335
 336         list_for_each_safe(pos, q, &delayed_uprobe_list) {
 337                 du = list_entry(pos, struct delayed_uprobe, list);
 338
 339                 if (uprobe && du->uprobe != uprobe)
 340                         continue;
 341                 if (mm && du->mm != mm)
 342                         continue;
 343
 344                 delayed_uprobe_delete(du);
 345         }
 346 }
 347
 348 static bool valid_ref_ctr_vma(struct uprobe *uprobe,
 349                               struct vm_area_struct *vma)
 350 {
 351         unsigned long vaddr = offset_to_vaddr(vma, uprobe->ref_ctr_offset);
 352
 353         return uprobe->ref_ctr_offset &&
 354                 vma->vm_file &&
 355                 file_inode(vma->vm_file) == uprobe->inode &&
 356                 (vma->vm_flags & (VM_WRITE|VM_SHARED)) == VM_WRITE &&
 357                 vma->vm_start <= vaddr &&
 358                 vma->vm_end > vaddr;
 359 }
 360
 361 static struct vm_area_struct *
 362 find_ref_ctr_vma(struct uprobe *uprobe, struct mm_struct *mm)
 363 {
 364         struct vm_area_struct *tmp;
 365
 366         for (tmp = mm->mmap; tmp; tmp = tmp->vm_next)
 367                 if (valid_ref_ctr_vma(uprobe, tmp))
 368                         return tmp;
 369
 370         return NULL;
 371 }
 372
 373 static int
 374 __update_ref_ctr(struct mm_struct *mm, unsigned long vaddr, short d)
 375 {
 376         void *kaddr;
 377         struct page *page;
 378         struct vm_area_struct *vma;
 379         int ret;
 380         short *ptr;
 381
 382         if (!vaddr || !d)
 383                 return -EINVAL;
 384
 385         ret = get_user_pages_remote(NULL, mm, vaddr, 1,
 386                         FOLL_WRITE, &page, &vma, NULL);
 387         if (unlikely(ret <= 0)) {
 388                 /*
 389                  * We are asking for 1 page. If get_user_pages_remote() fails,
 390                  * it may return 0, in that case we have to return error.
 391                  */
 392                 return ret == 0 ? -EBUSY : ret;
 393         }
 394
 395         kaddr = kmap_atomic(page);
 396         ptr = kaddr + (vaddr & ~PAGE_MASK);
 397
 398         if (unlikely(*ptr + d < 0)) {
 399                 pr_warn("ref_ctr going negative. vaddr: 0x%lx, "
 400                         "curr val: %d, delta: %d\n", vaddr, *ptr, d);
 401                 ret = -EINVAL;
 402                 goto out;
 403         }
 404
 405         *ptr += d;
 406         ret = 0;
 407 out:
 408         kunmap_atomic(kaddr);
 409         put_page(page);
 410         return ret;
 411 }
 412
 413 static void update_ref_ctr_warn(struct uprobe *uprobe,
 414                                 struct mm_struct *mm, short d)
 415 {
 416         pr_warn("ref_ctr %s failed for inode: 0x%lx offset: "
 417                 "0x%llx ref_ctr_offset: 0x%llx of mm: 0x%pK\n",
 418                 d > 0 ? "increment" : "decrement", uprobe->inode->i_ino,
 419                 (unsigned long long) uprobe->offset,
 420                 (unsigned long long) uprobe->ref_ctr_offset, mm);
 421 }
 422
 423 static int update_ref_ctr(struct uprobe *uprobe, struct mm_struct *mm,
 424                           short d)
 425 {
 426         struct vm_area_struct *rc_vma;
 427         unsigned long rc_vaddr;
 428         int ret = 0;
 429
 430         rc_vma = find_ref_ctr_vma(uprobe, mm);
 431
 432         if (rc_vma) {
 433                 rc_vaddr = offset_to_vaddr(rc_vma, uprobe->ref_ctr_offset);
 434                 ret = __update_ref_ctr(mm, rc_vaddr, d);
 435                 if (ret)
 436                         update_ref_ctr_warn(uprobe, mm, d);
 437
 438                 if (d > 0)
 439                         return ret;
 440         }
 441
 442         mutex_lock(&delayed_uprobe_lock);
 443         if (d > 0)
 444                 ret = delayed_uprobe_add(uprobe, mm);
 445         else
 446                 delayed_uprobe_remove(uprobe, mm);
 447         mutex_unlock(&delayed_uprobe_lock);
 448
 449         return ret;
 450 }
 451
 452 /*
 453  * NOTE:
 454  * Expect the breakpoint instruction to be the smallest size instruction for
 455  * the architecture. If an arch has variable length instruction and the
 456  * breakpoint instruction is not of the smallest length instruction
 457  * supported by that architecture then we need to modify is_trap_at_addr and
 458  * uprobe_write_opcode accordingly. This would never be a problem for archs
 459  * that have fixed length instructions.
 460  *
 461  * uprobe_write_opcode - write the opcode at a given virtual address.
 462  * @mm: the probed process address space.
 463  * @vaddr: the virtual address to store the opcode.
 464  * @opcode: opcode to be written at @vaddr.
 465  *
 466  * Called with mm->mmap_sem held for write.
 467  * Return 0 (success) or a negative errno.
 468  */
 469 int uprobe_write_opcode(struct arch_uprobe *auprobe, struct mm_struct *mm,
 470                         unsigned long vaddr, uprobe_opcode_t opcode)
 471 {
 472         struct uprobe *uprobe;
 473         struct page *old_page, *new_page;
 474         struct vm_area_struct *vma;
 475         int ret, is_register, ref_ctr_updated = 0;
 476         bool orig_page_huge = false;
 477         unsigned int gup_flags = FOLL_FORCE;
 478
 479         is_register = is_swbp_insn(&opcode);
 480         uprobe = container_of(auprobe, struct uprobe, arch);
 481
 482 retry:
 483         if (is_register)
 484                 gup_flags |= FOLL_SPLIT_PMD;
 485         /* Read the page with vaddr into memory */
 486         ret = get_user_pages_remote(NULL, mm, vaddr, 1, gup_flags,
 487                                     &old_page, &vma, NULL);
 488         if (ret <= 0)
 489                 return ret;
 490
 491         ret = verify_opcode(old_page, vaddr, &opcode);
 492         if (ret <= 0)
 493                 goto put_old;
 494
 495         if (WARN(!is_register && PageCompound(old_page),
 496                  "uprobe unregister should never work on compound page\n")) {
 497                 ret = -EINVAL;
 498                 goto put_old;
 499         }
 500
 501         /* We are going to replace instruction, update ref_ctr. */
 502         if (!ref_ctr_updated && uprobe->ref_ctr_offset) {
 503                 ret = update_ref_ctr(uprobe, mm, is_register ? 1 : -1);
 504                 if (ret)
 505                         goto put_old;
 506
 507                 ref_ctr_updated = 1;
 508         }
 509
 510         ret = 0;
 511         if (!is_register && !PageAnon(old_page))
 512                 goto put_old;
 513
 514         ret = anon_vma_prepare(vma);
 515         if (ret)
 516                 goto put_old;
 517
 518         ret = -ENOMEM;
 519         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vaddr);
 520         if (!new_page)
 521                 goto put_old;
 522
 523         __SetPageUptodate(new_page);
 524         copy_highpage(new_page, old_page);
 525         copy_to_page(new_page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
 526
 527         if (!is_register) {
 528                 struct page *orig_page;
 529                 pgoff_t index;
 530
 531                 VM_BUG_ON_PAGE(!PageAnon(old_page), old_page);
 532
 533                 index = vaddr_to_offset(vma, vaddr & PAGE_MASK) >> PAGE_SHIFT;
 534                 orig_page = find_get_page(vma->vm_file->f_inode->i_mapping,
 535                                           index);
 536
 537                 if (orig_page) {
 538                         if (PageUptodate(orig_page) &&
 539                             pages_identical(new_page, orig_page)) {
 540                                 /* let go new_page */
 541                                 put_page(new_page);
 542                                 new_page = NULL;
 543
 544                                 if (PageCompound(orig_page))
 545                                         orig_page_huge = true;
 546                         }
 547                         put_page(orig_page);
 548                 }
 549         }
 550
 551         ret = __replace_page(vma, vaddr, old_page, new_page);
 552         if (new_page)
 553                 put_page(new_page);
 554 put_old:
 555         put_page(old_page);
 556
 557         if (unlikely(ret == -EAGAIN))
 558                 goto retry;
 559
 560         /* Revert back reference counter if instruction update failed. */
 561         if (ret && is_register && ref_ctr_updated)
 562                 update_ref_ctr(uprobe, mm, -1);
 563
 564         /* try collapse pmd for compound page */
 565         if (!ret && orig_page_huge)
 566                 collapse_pte_mapped_thp(mm, vaddr);
 567
 568         return ret;
 569 }
 570
 571 /**
 572  * set_swbp - store breakpoint at a given address.
 573  * @auprobe: arch specific probepoint information.
 574  * @mm: the probed process address space.
 575  * @vaddr: the virtual address to insert the opcode.
 576  *
 577  * For mm @mm, store the breakpoint instruction at @vaddr.
 578  * Return 0 (success) or a negative errno.
 579  */
 580 int __weak set_swbp(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
 581 {
 582         return uprobe_write_opcode(auprobe, mm, vaddr, UPROBE_SWBP_INSN);
 583 }
 584
 585 /**
 586  * set_orig_insn - Restore the original instruction.
 587  * @mm: the probed process address space.
 588  * @auprobe: arch specific probepoint information.
 589  * @vaddr: the virtual address to insert the opcode.
 590  *
 591  * For mm @mm, restore the original opcode (opcode) at @vaddr.
 592  * Return 0 (success) or a negative errno.
 593  */
 594 int __weak
 595 set_orig_insn(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
 596 {
 597         return uprobe_write_opcode(auprobe, mm, vaddr,
 598                         *(uprobe_opcode_t *)&auprobe->insn);
 599 }
 600
 601 static struct uprobe *get_uprobe(struct uprobe *uprobe)
 602 {
 603         refcount_inc(&uprobe->ref);
 604         return uprobe;
 605 }
 606
 607 static void put_uprobe(struct uprobe *uprobe)
 608 {
 609         if (refcount_dec_and_test(&uprobe->ref)) {
 610                 /*
 611                  * If application munmap(exec_vma) before uprobe_unregister()
 612                  * gets called, we don't get a chance to remove uprobe from
 613                  * delayed_uprobe_list from remove_breakpoint(). Do it here.
 614                  */
 615                 mutex_lock(&delayed_uprobe_lock);
 616                 delayed_uprobe_remove(uprobe, NULL);
 617                 mutex_unlock(&delayed_uprobe_lock);
 618                 kfree(uprobe);
 619         }
 620 }
 621
 622 static int match_uprobe(struct uprobe *l, struct uprobe *r)
 623 {
 624         if (l->inode < r->inode)
 625                 return -1;
 626
 627         if (l->inode > r->inode)
 628                 return 1;
 629
 630         if (l->offset < r->offset)
 631                 return -1;
 632
 633         if (l->offset > r->offset)
 634                 return 1;
 635
 636         return 0;
 637 }
 638
 639 static struct uprobe *__find_uprobe(struct inode *inode, loff_t offset)
 640 {
 641         struct uprobe u = { .inode = inode, .offset = offset };
 642         struct rb_node *n = uprobes_tree.rb_node;
 643         struct uprobe *uprobe;
 644         int match;
 645
 646         while (n) {
 647                 uprobe = rb_entry(n, struct uprobe, rb_node);
 648                 match = match_uprobe(&u, uprobe);
 649                 if (!match)
 650                         return get_uprobe(uprobe);
 651
 652                 if (match < 0)
 653                         n = n->rb_left;
 654                 else
 655                         n = n->rb_right;
 656         }
 657         return NULL;
 658 }
 659
 660 /*
 661  * Find a uprobe corresponding to a given inode:offset
 662  * Acquires uprobes_treelock
 663  */
 664 static struct uprobe *find_uprobe(struct inode *inode, loff_t offset)
 665 {
 666         struct uprobe *uprobe;
 667
 668         spin_lock(&uprobes_treelock);
 669         uprobe = __find_uprobe(inode, offset);
 670         spin_unlock(&uprobes_treelock);
 671
 672         return uprobe;
 673 }
 674
 675 static struct uprobe *__insert_uprobe(struct uprobe *uprobe)
 676 {
 677         struct rb_node **p = &uprobes_tree.rb_node;
 678         struct rb_node *parent = NULL;
 679         struct uprobe *u;
 680         int match;
 681
 682         while (*p) {
 683                 parent = *p;
 684                 u = rb_entry(parent, struct uprobe, rb_node);
 685                 match = match_uprobe(uprobe, u);
 686                 if (!match)
 687                         return get_uprobe(u);
 688
 689                 if (match < 0)
 690                         p = &parent->rb_left;
 691                 else
 692                         p = &parent->rb_right;
 693
 694         }
 695
 696         u = NULL;
 697         rb_link_node(&uprobe->rb_node, parent, p);
 698         rb_insert_color(&uprobe->rb_node, &uprobes_tree);
 699         /* get access + creation ref */
 700         refcount_set(&uprobe->ref, 2);
 701
 702         return u;
 703 }
 704
 705 /*
 706  * Acquire uprobes_treelock.
 707  * Matching uprobe already exists in rbtree;
 708  *      increment (access refcount) and return the matching uprobe.
 709  *
 710  * No matching uprobe; insert the uprobe in rb_tree;
 711  *      get a double refcount (access + creation) and return NULL.
 712  */
 713 static struct uprobe *insert_uprobe(struct uprobe *uprobe)
 714 {
 715         struct uprobe *u;
 716
 717         spin_lock(&uprobes_treelock);
 718         u = __insert_uprobe(uprobe);
 719         spin_unlock(&uprobes_treelock);
 720
 721         return u;
 722 }
 723
 724 static void
 725 ref_ctr_mismatch_warn(struct uprobe *cur_uprobe, struct uprobe *uprobe)
 726 {
 727         pr_warn("ref_ctr_offset mismatch. inode: 0x%lx offset: 0x%llx "
 728                 "ref_ctr_offset(old): 0x%llx ref_ctr_offset(new): 0x%llx\n",
 729                 uprobe->inode->i_ino, (unsigned long long) uprobe->offset,
 730                 (unsigned long long) cur_uprobe->ref_ctr_offset,
 731                 (unsigned long long) uprobe->ref_ctr_offset);
 732 }
 733
 734 static struct uprobe *alloc_uprobe(struct inode *inode, loff_t offset,
 735                                    loff_t ref_ctr_offset)
 736 {
 737         struct uprobe *uprobe, *cur_uprobe;
 738
 739         uprobe = kzalloc(sizeof(struct uprobe), GFP_KERNEL);
 740         if (!uprobe)
 741                 return NULL;
 742
 743         uprobe->inode = inode;
 744         uprobe->offset = offset;
 745         uprobe->ref_ctr_offset = ref_ctr_offset;
 746         init_rwsem(&uprobe->register_rwsem);
 747         init_rwsem(&uprobe->consumer_rwsem);
 748
 749         /* add to uprobes_tree, sorted on inode:offset */
 750         cur_uprobe = insert_uprobe(uprobe);
 751         /* a uprobe exists for this inode:offset combination */
 752         if (cur_uprobe) {
 753                 if (cur_uprobe->ref_ctr_offset != uprobe->ref_ctr_offset) {
 754                         ref_ctr_mismatch_warn(cur_uprobe, uprobe);
 755                         put_uprobe(cur_uprobe);
 756                         kfree(uprobe);
 757                         return ERR_PTR(-EINVAL);
 758                 }
 759                 kfree(uprobe);
 760                 uprobe = cur_uprobe;
 761         }
 762
 763         return uprobe;
 764 }
 765
 766 static void consumer_add(struct uprobe *uprobe, struct uprobe_consumer *uc)
 767 {
 768         down_write(&uprobe->consumer_rwsem);
 769         uc->next = uprobe->consumers;
 770         uprobe->consumers = uc;
 771         up_write(&uprobe->consumer_rwsem);
 772 }
 773
 774 /*
 775  * For uprobe @uprobe, delete the consumer @uc.
 776  * Return true if the @uc is deleted successfully
 777  * or return false.
 778  */
 779 static bool consumer_del(struct uprobe *uprobe, struct uprobe_consumer *uc)
 780 {
 781         struct uprobe_consumer **con;
 782         bool ret = false;
 783
 784         down_write(&uprobe->consumer_rwsem);
 785         for (con = &uprobe->consumers; *con; con = &(*con)->next) {
 786                 if (*con == uc) {
 787                         *con = uc->next;
 788                         ret = true;
 789                         break;
 790                 }
 791         }
 792         up_write(&uprobe->consumer_rwsem);
 793
 794         return ret;
 795 }
 796
 797 static int __copy_insn(struct address_space *mapping, struct file *filp,
 798                         void *insn, int nbytes, loff_t offset)
 799 {
 800         struct page *page;
 801         /*
 802          * Ensure that the page that has the original instruction is populated
 803          * and in page-cache. If ->readpage == NULL it must be shmem_mapping(),
 804          * see uprobe_register().
 805          */
 806         if (mapping->a_ops->readpage)
 807                 page = read_mapping_page(mapping, offset >> PAGE_SHIFT, filp);
 808         else
 809                 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
 810         if (IS_ERR(page))
 811                 return PTR_ERR(page);
 812
 813         copy_from_page(page, offset, insn, nbytes);
 814         put_page(page);
 815
 816         return 0;
 817 }
 818
 819 static int copy_insn(struct uprobe *uprobe, struct file *filp)
 820 {
 821         struct address_space *mapping = uprobe->inode->i_mapping;
 822         loff_t offs = uprobe->offset;
 823         void *insn = &uprobe->arch.insn;
 824         int size = sizeof(uprobe->arch.insn);
 825         int len, err = -EIO;
 826
 827         /* Copy only available bytes, -EIO if nothing was read */
 828         do {
 829                 if (offs >= i_size_read(uprobe->inode))
 830                         break;
 831
 832                 len = min_t(int, size, PAGE_SIZE - (offs & ~PAGE_MASK));
 833                 err = __copy_insn(mapping, filp, insn, len, offs);
 834                 if (err)
 835                         break;
 836
 837                 insn += len;
 838                 offs += len;
 839                 size -= len;
 840         } while (size);
 841
 842         return err;
 843 }
 844
 845 static int prepare_uprobe(struct uprobe *uprobe, struct file *file,
 846                                 struct mm_struct *mm, unsigned long vaddr)
 847 {
 848         int ret = 0;
 849
 850         if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
 851                 return ret;
 852
 853         /* TODO: move this into _register, until then we abuse this sem. */
 854         down_write(&uprobe->consumer_rwsem);
 855         if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
 856                 goto out;
 857
 858         ret = copy_insn(uprobe, file);
 859         if (ret)
 860                 goto out;
 861
 862         ret = -ENOTSUPP;
 863         if (is_trap_insn((uprobe_opcode_t *)&uprobe->arch.insn))
 864                 goto out;
 865
 866         ret = arch_uprobe_analyze_insn(&uprobe->arch, mm, vaddr);
 867         if (ret)
 868                 goto out;
 869
 870         smp_wmb(); /* pairs with the smp_rmb() in handle_swbp() */
 871         set_bit(UPROBE_COPY_INSN, &uprobe->flags);
 872
 873  out:
 874         up_write(&uprobe->consumer_rwsem);
 875
 876         return ret;
 877 }
 878
 879 static inline bool consumer_filter(struct uprobe_consumer *uc,
 880                                    enum uprobe_filter_ctx ctx, struct mm_struct *mm)
 881 {
 882         return !uc->filter || uc->filter(uc, ctx, mm);
 883 }
 884
 885 static bool filter_chain(struct uprobe *uprobe,
 886                          enum uprobe_filter_ctx ctx, struct mm_struct *mm)
 887 {
 888         struct uprobe_consumer *uc;
 889         bool ret = false;
 890
 891         down_read(&uprobe->consumer_rwsem);
 892         for (uc = uprobe->consumers; uc; uc = uc->next) {
 893                 ret = consumer_filter(uc, ctx, mm);
 894                 if (ret)
 895                         break;
 896         }
 897         up_read(&uprobe->consumer_rwsem);
 898
 899         return ret;
 900 }
 901
 902 static int
 903 install_breakpoint(struct uprobe *uprobe, struct mm_struct *mm,
 904                         struct vm_area_struct *vma, unsigned long vaddr)
 905 {
 906         bool first_uprobe;
 907         int ret;
 908
 909         ret = prepare_uprobe(uprobe, vma->vm_file, mm, vaddr);
 910         if (ret)
 911                 return ret;
 912
 913         /*
 914          * set MMF_HAS_UPROBES in advance for uprobe_pre_sstep_notifier(),
 915          * the task can hit this breakpoint right after __replace_page().
 916          */
 917         first_uprobe = !test_bit(MMF_HAS_UPROBES, &mm->flags);
 918         if (first_uprobe)
 919                 set_bit(MMF_HAS_UPROBES, &mm->flags);
 920
 921         ret = set_swbp(&uprobe->arch, mm, vaddr);
 922         if (!ret)
 923                 clear_bit(MMF_RECALC_UPROBES, &mm->flags);
 924         else if (first_uprobe)
 925                 clear_bit(MMF_HAS_UPROBES, &mm->flags);
 926
 927         return ret;
 928 }
 929
 930 static int
 931 remove_breakpoint(struct uprobe *uprobe, struct mm_struct *mm, unsigned long vaddr)
 932 {
 933         set_bit(MMF_RECALC_UPROBES, &mm->flags);
 934         return set_orig_insn(&uprobe->arch, mm, vaddr);
 935 }
 936
 937 static inline bool uprobe_is_active(struct uprobe *uprobe)
 938 {
 939         return !RB_EMPTY_NODE(&uprobe->rb_node);
 940 }
 941 /*
 942  * There could be threads that have already hit the breakpoint. They
 943  * will recheck the current insn and restart if find_uprobe() fails.
 944  * See find_active_uprobe().
 945  */
 946 static void delete_uprobe(struct uprobe *uprobe)
 947 {
 948         if (WARN_ON(!uprobe_is_active(uprobe)))
 949                 return;
 950
 951         spin_lock(&uprobes_treelock);
 952         rb_erase(&uprobe->rb_node, &uprobes_tree);
 953         spin_unlock(&uprobes_treelock);
 954         RB_CLEAR_NODE(&uprobe->rb_node); /* for uprobe_is_active() */
 955         put_uprobe(uprobe);
 956 }
 957
 958 struct map_info {
 959         struct map_info *next;
 960         struct mm_struct *mm;
 961         unsigned long vaddr;
 962 };
 963
 964 static inline struct map_info *free_map_info(struct map_info *info)
 965 {
 966         struct map_info *next = info->next;
 967         kfree(info);
 968         return next;
 969 }
 970
 971 static struct map_info *
 972 build_map_info(struct address_space *mapping, loff_t offset, bool is_register)
 973 {
 974         unsigned long pgoff = offset >> PAGE_SHIFT;
 975         struct vm_area_struct *vma;
 976         struct map_info *curr = NULL;
 977         struct map_info *prev = NULL;
 978         struct map_info *info;
 979         int more = 0;
 980
 981  again:
 982         i_mmap_lock_read(mapping);
 983         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
 984                 if (!valid_vma(vma, is_register))
 985                         continue;
 986
 987                 if (!prev && !more) {
 988                         /*
 989                          * Needs GFP_NOWAIT to avoid i_mmap_rwsem recursion through
 990                          * reclaim. This is optimistic, no harm done if it fails.
 991                          */
 992                         prev = kmalloc(sizeof(struct map_info),
 993                                         GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
 994                         if (prev)
 995                                 prev->next = NULL;
 996                 }
 997                 if (!prev) {
 998                         more++;
 999                         continue;
1000                 }
1001
1002                 if (!mmget_not_zero(vma->vm_mm))
1003                         continue;
1004
1005                 info = prev;
1006                 prev = prev->next;
1007                 info->next = curr;
1008                 curr = info;
1009
1010                 info->mm = vma->vm_mm;
1011                 info->vaddr = offset_to_vaddr(vma, offset);
1012         }
1013         i_mmap_unlock_read(mapping);
1014
1015         if (!more)
1016                 goto out;
1017
1018         prev = curr;
1019         while (curr) {
1020                 mmput(curr->mm);
1021                 curr = curr->next;
1022         }
1023
1024         do {
1025                 info = kmalloc(sizeof(struct map_info), GFP_KERNEL);
1026                 if (!info) {
1027                         curr = ERR_PTR(-ENOMEM);
1028                         goto out;
1029                 }
1030                 info->next = prev;
1031                 prev = info;
1032         } while (--more);
1033
1034         goto again;
1035  out:
1036         while (prev)
1037                 prev = free_map_info(prev);
1038         return curr;
1039 }
1040
1041 static int
1042 register_for_each_vma(struct uprobe *uprobe, struct uprobe_consumer *new)
1043 {
1044         bool is_register = !!new;
1045         struct map_info *info;
1046         int err = 0;
1047
1048         percpu_down_write(&dup_mmap_sem);
1049         info = build_map_info(uprobe->inode->i_mapping,
1050                                         uprobe->offset, is_register);
1051         if (IS_ERR(info)) {
1052                 err = PTR_ERR(info);
1053                 goto out;
1054         }
1055
1056         while (info) {
1057                 struct mm_struct *mm = info->mm;
1058                 struct vm_area_struct *vma;
1059
1060                 if (err && is_register)
1061                         goto free;
1062
1063                 down_write(&mm->mmap_sem);
1064                 vma = find_vma(mm, info->vaddr);
1065                 if (!vma || !valid_vma(vma, is_register) ||
1066                     file_inode(vma->vm_file) != uprobe->inode)
1067                         goto unlock;
1068
1069                 if (vma->vm_start > info->vaddr ||
1070                     vaddr_to_offset(vma, info->vaddr) != uprobe->offset)
1071                         goto unlock;
1072
1073                 if (is_register) {
1074                         /* consult only the "caller", new consumer. */
1075                         if (consumer_filter(new,
1076                                         UPROBE_FILTER_REGISTER, mm))
1077                                 err = install_breakpoint(uprobe, mm, vma, info->vaddr);
1078                 } else if (test_bit(MMF_HAS_UPROBES, &mm->flags)) {
1079                         if (!filter_chain(uprobe,
1080                                         UPROBE_FILTER_UNREGISTER, mm))
1081                                 err |= remove_breakpoint(uprobe, mm, info->vaddr);
1082                 }
1083
1084  unlock:
1085                 up_write(&mm->mmap_sem);
1086  free:
1087                 mmput(mm);
1088                 info = free_map_info(info);
1089         }
1090  out:
1091         percpu_up_write(&dup_mmap_sem);
1092         return err;
1093 }
1094
1095 static void
1096 __uprobe_unregister(struct uprobe *uprobe, struct uprobe_consumer *uc)
1097 {
1098         int err;
1099
1100         if (WARN_ON(!consumer_del(uprobe, uc)))
1101                 return;
1102
1103         err = register_for_each_vma(uprobe, NULL);
1104         /* TODO : cant unregister? schedule a worker thread */
1105         if (!uprobe->consumers && !err)
1106                 delete_uprobe(uprobe);
1107 }
1108
1109 /*
1110  * uprobe_unregister - unregister an already registered probe.
1111  * @inode: the file in which the probe has to be removed.
1112  * @offset: offset from the start of the file.
1113  * @uc: identify which probe if multiple probes are colocated.
1114  */
1115 void uprobe_unregister(struct inode *inode, loff_t offset, struct uprobe_consumer *uc)
1116 {
1117         struct uprobe *uprobe;
1118
1119         uprobe = find_uprobe(inode, offset);
1120         if (WARN_ON(!uprobe))
1121                 return;
1122
1123         down_write(&uprobe->register_rwsem);
1124         __uprobe_unregister(uprobe, uc);
1125         up_write(&uprobe->register_rwsem);
1126         put_uprobe(uprobe);
1127 }
1128 EXPORT_SYMBOL_GPL(uprobe_unregister);
1129
1130 /*
1131  * __uprobe_register - register a probe
1132  * @inode: the file in which the probe has to be placed.
1133  * @offset: offset from the start of the file.
1134  * @uc: information on howto handle the probe..
1135  *
1136  * Apart from the access refcount, __uprobe_register() takes a creation
1137  * refcount (thro alloc_uprobe) if and only if this @uprobe is getting
1138  * inserted into the rbtree (i.e first consumer for a @inode:@offset
1139  * tuple).  Creation refcount stops uprobe_unregister from freeing the
1140  * @uprobe even before the register operation is complete. Creation
1141  * refcount is released when the last @uc for the @uprobe
1142  * unregisters. Caller of __uprobe_register() is required to keep @inode
1143  * (and the containing mount) referenced.
1144  *
1145  * Return errno if it cannot successully install probes
1146  * else return 0 (success)
1147  */
1148 static int __uprobe_register(struct inode *inode, loff_t offset,
1149                              loff_t ref_ctr_offset, struct uprobe_consumer *uc)
1150 {
1151         struct uprobe *uprobe;
1152         int ret;
1153
1154         /* Uprobe must have at least one set consumer */
1155         if (!uc->handler && !uc->ret_handler)
1156                 return -EINVAL;
1157
1158         /* copy_insn() uses read_mapping_page() or shmem_read_mapping_page() */
1159         if (!inode->i_mapping->a_ops->readpage && !shmem_mapping(inode->i_mapping))
1160                 return -EIO;
1161         /* Racy, just to catch the obvious mistakes */
1162         if (offset > i_size_read(inode))
1163                 return -EINVAL;
1164
1165         /*
1166          * This ensures that copy_from_page(), copy_to_page() and
1167          * __update_ref_ctr() can't cross page boundary.
1168          */
1169         if (!IS_ALIGNED(offset, UPROBE_SWBP_INSN_SIZE))
1170                 return -EINVAL;
1171         if (!IS_ALIGNED(ref_ctr_offset, sizeof(short)))
1172                 return -EINVAL;
1173
1174  retry:
1175         uprobe = alloc_uprobe(inode, offset, ref_ctr_offset);
1176         if (!uprobe)
1177                 return -ENOMEM;
1178         if (IS_ERR(uprobe))
1179                 return PTR_ERR(uprobe);
1180
1181         /*
1182          * We can race with uprobe_unregister()->delete_uprobe().
1183          * Check uprobe_is_active() and retry if it is false.
1184          */
1185         down_write(&uprobe->register_rwsem);
1186         ret = -EAGAIN;
1187         if (likely(uprobe_is_active(uprobe))) {
1188                 consumer_add(uprobe, uc);
1189                 ret = register_for_each_vma(uprobe, uc);
1190                 if (ret)
1191                         __uprobe_unregister(uprobe, uc);
1192         }
1193         up_write(&uprobe->register_rwsem);
1194         put_uprobe(uprobe);
1195
1196         if (unlikely(ret == -EAGAIN))
1197                 goto retry;
1198         return ret;
1199 }
1200
1201 int uprobe_register(struct inode *inode, loff_t offset,
1202                     struct uprobe_consumer *uc)
1203 {
1204         return __uprobe_register(inode, offset, 0, uc);
1205 }
1206 EXPORT_SYMBOL_GPL(uprobe_register);
1207
1208 int uprobe_register_refctr(struct inode *inode, loff_t offset,
1209                            loff_t ref_ctr_offset, struct uprobe_consumer *uc)
1210 {
1211         return __uprobe_register(inode, offset, ref_ctr_offset, uc);
1212 }
1213 EXPORT_SYMBOL_GPL(uprobe_register_refctr);
1214
1215 /*
1216  * uprobe_apply - unregister an already registered probe.
1217  * @inode: the file in which the probe has to be removed.
1218  * @offset: offset from the start of the file.
1219  * @uc: consumer which wants to add more or remove some breakpoints
1220  * @add: add or remove the breakpoints
1221  */
1222 int uprobe_apply(struct inode *inode, loff_t offset,
1223                         struct uprobe_consumer *uc, bool add)
1224 {
1225         struct uprobe *uprobe;
1226         struct uprobe_consumer *con;
1227         int ret = -ENOENT;
1228
1229         uprobe = find_uprobe(inode, offset);
1230         if (WARN_ON(!uprobe))
1231                 return ret;
1232
1233         down_write(&uprobe->register_rwsem);
1234         for (con = uprobe->consumers; con && con != uc ; con = con->next)
1235                 ;
1236         if (con)
1237                 ret = register_for_each_vma(uprobe, add ? uc : NULL);
1238         up_write(&uprobe->register_rwsem);
1239         put_uprobe(uprobe);
1240
1241         return ret;
1242 }
1243
1244 static int unapply_uprobe(struct uprobe *uprobe, struct mm_struct *mm)
1245 {
1246         struct vm_area_struct *vma;
1247         int err = 0;
1248
1249         down_read(&mm->mmap_sem);
1250         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1251                 unsigned long vaddr;
1252                 loff_t offset;
1253
1254                 if (!valid_vma(vma, false) ||
1255                     file_inode(vma->vm_file) != uprobe->inode)
1256                         continue;
1257
1258                 offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1259                 if (uprobe->offset <  offset ||
1260                     uprobe->offset >= offset + vma->vm_end - vma->vm_start)
1261                         continue;
1262
1263                 vaddr = offset_to_vaddr(vma, uprobe->offset);
1264                 err |= remove_breakpoint(uprobe, mm, vaddr);
1265         }
1266         up_read(&mm->mmap_sem);
1267
1268         return err;
1269 }
1270
1271 static struct rb_node *
1272 find_node_in_range(struct inode *inode, loff_t min, loff_t max)
1273 {
1274         struct rb_node *n = uprobes_tree.rb_node;
1275
1276         while (n) {
1277                 struct uprobe *u = rb_entry(n, struct uprobe, rb_node);
1278
1279                 if (inode < u->inode) {
1280                         n = n->rb_left;
1281                 } else if (inode > u->inode) {
1282                         n = n->rb_right;
1283                 } else {
1284                         if (max < u->offset)
1285                                 n = n->rb_left;
1286                         else if (min > u->offset)
1287                                 n = n->rb_right;
1288                         else
1289                                 break;
1290                 }
1291         }
1292
1293         return n;
1294 }
1295
1296 /*
1297  * For a given range in vma, build a list of probes that need to be inserted.
1298  */
1299 static void build_probe_list(struct inode *inode,
1300                                 struct vm_area_struct *vma,
1301                                 unsigned long start, unsigned long end,
1302                                 struct list_head *head)
1303 {
1304         loff_t min, max;
1305         struct rb_node *n, *t;
1306         struct uprobe *u;
1307
1308         INIT_LIST_HEAD(head);
1309         min = vaddr_to_offset(vma, start);
1310         max = min + (end - start) - 1;
1311
1312         spin_lock(&uprobes_treelock);
1313         n = find_node_in_range(inode, min, max);
1314         if (n) {
1315                 for (t = n; t; t = rb_prev(t)) {
1316                         u = rb_entry(t, struct uprobe, rb_node);
1317                         if (u->inode != inode || u->offset < min)
1318                                 break;
1319                         list_add(&u->pending_list, head);
1320                         get_uprobe(u);
1321                 }
1322                 for (t = n; (t = rb_next(t)); ) {
1323                         u = rb_entry(t, struct uprobe, rb_node);
1324                         if (u->inode != inode || u->offset > max)
1325                                 break;
1326                         list_add(&u->pending_list, head);
1327                         get_uprobe(u);
1328                 }
1329         }
1330         spin_unlock(&uprobes_treelock);
1331 }
1332
1333 /* @vma contains reference counter, not the probed instruction. */
1334 static int delayed_ref_ctr_inc(struct vm_area_struct *vma)
1335 {
1336         struct list_head *pos, *q;
1337         struct delayed_uprobe *du;
1338         unsigned long vaddr;
1339         int ret = 0, err = 0;
1340
1341         mutex_lock(&delayed_uprobe_lock);
1342         list_for_each_safe(pos, q, &delayed_uprobe_list) {
1343                 du = list_entry(pos, struct delayed_uprobe, list);
1344
1345                 if (du->mm != vma->vm_mm ||
1346                     !valid_ref_ctr_vma(du->uprobe, vma))
1347                         continue;
1348
1349                 vaddr = offset_to_vaddr(vma, du->uprobe->ref_ctr_offset);
1350                 ret = __update_ref_ctr(vma->vm_mm, vaddr, 1);
1351                 if (ret) {
1352                         update_ref_ctr_warn(du->uprobe, vma->vm_mm, 1);
1353                         if (!err)
1354                                 err = ret;
1355                 }
1356                 delayed_uprobe_delete(du);
1357         }
1358         mutex_unlock(&delayed_uprobe_lock);
1359         return err;
1360 }
1361
1362 /*
1363  * Called from mmap_region/vma_adjust with mm->mmap_sem acquired.
1364  *
1365  * Currently we ignore all errors and always return 0, the callers
1366  * can't handle the failure anyway.
1367  */
1368 int uprobe_mmap(struct vm_area_struct *vma)
1369 {
1370         struct list_head tmp_list;
1371         struct uprobe *uprobe, *u;
1372         struct inode *inode;
1373
1374         if (no_uprobe_events())
1375                 return 0;
1376
1377         if (vma->vm_file &&
1378             (vma->vm_flags & (VM_WRITE|VM_SHARED)) == VM_WRITE &&
1379             test_bit(MMF_HAS_UPROBES, &vma->vm_mm->flags))
1380                 delayed_ref_ctr_inc(vma);
1381
1382         if (!valid_vma(vma, true))
1383                 return 0;
1384
1385         inode = file_inode(vma->vm_file);
1386         if (!inode)
1387                 return 0;
1388
1389         mutex_lock(uprobes_mmap_hash(inode));
1390         build_probe_list(inode, vma, vma->vm_start, vma->vm_end, &tmp_list);
1391         /*
1392          * We can race with uprobe_unregister(), this uprobe can be already
1393          * removed. But in this case filter_chain() must return false, all
1394          * consumers have gone away.
1395          */
1396         list_for_each_entry_safe(uprobe, u, &tmp_list, pending_list) {
1397                 if (!fatal_signal_pending(current) &&
1398                     filter_chain(uprobe, UPROBE_FILTER_MMAP, vma->vm_mm)) {
1399                         unsigned long vaddr = offset_to_vaddr(vma, uprobe->offset);
1400                         install_breakpoint(uprobe, vma->vm_mm, vma, vaddr);
1401                 }
1402                 put_uprobe(uprobe);
1403         }
1404         mutex_unlock(uprobes_mmap_hash(inode));
1405
1406         return 0;
1407 }
1408
1409 static bool
1410 vma_has_uprobes(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1411 {
1412         loff_t min, max;
1413         struct inode *inode;
1414         struct rb_node *n;
1415
1416         inode = file_inode(vma->vm_file);
1417
1418         min = vaddr_to_offset(vma, start);
1419         max = min + (end - start) - 1;
1420
1421         spin_lock(&uprobes_treelock);
1422         n = find_node_in_range(inode, min, max);
1423         spin_unlock(&uprobes_treelock);
1424
1425         return !!n;
1426 }
1427
1428 /*
1429  * Called in context of a munmap of a vma.
1430  */
1431 void uprobe_munmap(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1432 {
1433         if (no_uprobe_events() || !valid_vma(vma, false))
1434                 return;
1435
1436         if (!atomic_read(&vma->vm_mm->mm_users)) /* called by mmput() ? */
1437                 return;
1438
1439         if (!test_bit(MMF_HAS_UPROBES, &vma->vm_mm->flags) ||
1440              test_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags))
1441                 return;
1442
1443         if (vma_has_uprobes(vma, start, end))
1444                 set_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags);
1445 }
1446
1447 /* Slot allocation for XOL */
1448 static int xol_add_vma(struct mm_struct *mm, struct xol_area *area)
1449 {
1450         struct vm_area_struct *vma;
1451         int ret;
1452
1453         if (down_write_killable(&mm->mmap_sem))
1454                 return -EINTR;
1455
1456         if (mm->uprobes_state.xol_area) {
1457                 ret = -EALREADY;
1458                 goto fail;
1459         }
1460
1461         if (!area->vaddr) {
1462                 /* Try to map as high as possible, this is only a hint. */
1463                 area->vaddr = get_unmapped_area(NULL, TASK_SIZE - PAGE_SIZE,
1464                                                 PAGE_SIZE, 0, 0);
1465                 if (area->vaddr & ~PAGE_MASK) {
1466                         ret = area->vaddr;
1467                         goto fail;
1468                 }
1469         }
1470
1471         vma = _install_special_mapping(mm, area->vaddr, PAGE_SIZE,
1472                                 VM_EXEC|VM_MAYEXEC|VM_DONTCOPY|VM_IO,
1473                                 &area->xol_mapping);
1474         if (IS_ERR(vma)) {
1475                 ret = PTR_ERR(vma);
1476                 goto fail;
1477         }
1478
1479         ret = 0;
1480         /* pairs with get_xol_area() */
1481         smp_store_release(&mm->uprobes_state.xol_area, area); /* ^^^ */
1482  fail:
1483         up_write(&mm->mmap_sem);
1484
1485         return ret;
1486 }
1487
1488 static struct xol_area *__create_xol_area(unsigned long vaddr)
1489 {
1490         struct mm_struct *mm = current->mm;
1491         uprobe_opcode_t insn = UPROBE_SWBP_INSN;
1492         struct xol_area *area;
1493
1494         area = kmalloc(sizeof(*area), GFP_KERNEL);
1495         if (unlikely(!area))
1496                 goto out;
1497
1498         area->bitmap = kcalloc(BITS_TO_LONGS(UINSNS_PER_PAGE), sizeof(long),
1499                                GFP_KERNEL);
1500         if (!area->bitmap)
1501                 goto free_area;
1502
1503         area->xol_mapping.name = "[uprobes]";
1504         area->xol_mapping.fault = NULL;
1505         area->xol_mapping.pages = area->pages;
1506         area->pages[0] = alloc_page(GFP_HIGHUSER);
1507         if (!area->pages[0])
1508                 goto free_bitmap;
1509         area->pages[1] = NULL;
1510
1511         area->vaddr = vaddr;
1512         init_waitqueue_head(&area->wq);
1513         /* Reserve the 1st slot for get_trampoline_vaddr() */
1514         set_bit(0, area->bitmap);
1515         atomic_set(&area->slot_count, 1);
1516         arch_uprobe_copy_ixol(area->pages[0], 0, &insn, UPROBE_SWBP_INSN_SIZE);
1517
1518         if (!xol_add_vma(mm, area))
1519                 return area;
1520
1521         __free_page(area->pages[0]);
1522  free_bitmap:
1523         kfree(area->bitmap);
1524  free_area:
1525         kfree(area);
1526  out:
1527         return NULL;
1528 }
1529
1530 /*
1531  * get_xol_area - Allocate process's xol_area if necessary.
1532  * This area will be used for storing instructions for execution out of line.
1533  *
1534  * Returns the allocated area or NULL.
1535  */
1536 static struct xol_area *get_xol_area(void)
1537 {
1538         struct mm_struct *mm = current->mm;
1539         struct xol_area *area;
1540
1541         if (!mm->uprobes_state.xol_area)
1542                 __create_xol_area(0);
1543
1544         /* Pairs with xol_add_vma() smp_store_release() */
1545         area = READ_ONCE(mm->uprobes_state.xol_area); /* ^^^ */
1546         return area;
1547 }
1548
1549 /*
1550  * uprobe_clear_state - Free the area allocated for slots.
1551  */
1552 void uprobe_clear_state(struct mm_struct *mm)
1553 {
1554         struct xol_area *area = mm->uprobes_state.xol_area;
1555
1556         mutex_lock(&delayed_uprobe_lock);
1557         delayed_uprobe_remove(NULL, mm);
1558         mutex_unlock(&delayed_uprobe_lock);
1559
1560         if (!area)
1561                 return;
1562
1563         put_page(area->pages[0]);
1564         kfree(area->bitmap);
1565         kfree(area);
1566 }
1567
1568 void uprobe_start_dup_mmap(void)
1569 {
1570         percpu_down_read(&dup_mmap_sem);
1571 }
1572
1573 void uprobe_end_dup_mmap(void)
1574 {
1575         percpu_up_read(&dup_mmap_sem);
1576 }
1577
1578 void uprobe_dup_mmap(struct mm_struct *oldmm, struct mm_struct *newmm)
1579 {
1580         if (test_bit(MMF_HAS_UPROBES, &oldmm->flags)) {
1581                 set_bit(MMF_HAS_UPROBES, &newmm->flags);
1582                 /* unconditionally, dup_mmap() skips VM_DONTCOPY vmas */
1583                 set_bit(MMF_RECALC_UPROBES, &newmm->flags);
1584         }
1585 }
1586
1587 /*
1588  *  - search for a free slot.
1589  */
1590 static unsigned long xol_take_insn_slot(struct xol_area *area)
1591 {
1592         unsigned long slot_addr;
1593         int slot_nr;
1594
1595         do {
1596                 slot_nr = find_first_zero_bit(area->bitmap, UINSNS_PER_PAGE);
1597                 if (slot_nr < UINSNS_PER_PAGE) {
1598                         if (!test_and_set_bit(slot_nr, area->bitmap))
1599                                 break;
1600
1601                         slot_nr = UINSNS_PER_PAGE;
1602                         continue;
1603                 }
1604                 wait_event(area->wq, (atomic_read(&area->slot_count) < UINSNS_PER_PAGE));
1605         } while (slot_nr >= UINSNS_PER_PAGE);
1606
1607         slot_addr = area->vaddr + (slot_nr * UPROBE_XOL_SLOT_BYTES);
1608         atomic_inc(&area->slot_count);
1609
1610         return slot_addr;
1611 }
1612
1613 /*
1614  * xol_get_insn_slot - allocate a slot for xol.
1615  * Returns the allocated slot address or 0.
1616  */
1617 static unsigned long xol_get_insn_slot(struct uprobe *uprobe)
1618 {
1619         struct xol_area *area;
1620         unsigned long xol_vaddr;
1621
1622         area = get_xol_area();
1623         if (!area)
1624                 return 0;
1625
1626         xol_vaddr = xol_take_insn_slot(area);
1627         if (unlikely(!xol_vaddr))
1628                 return 0;
1629
1630         arch_uprobe_copy_ixol(area->pages[0], xol_vaddr,
1631                               &uprobe->arch.ixol, sizeof(uprobe->arch.ixol));
1632
1633         return xol_vaddr;
1634 }
1635
1636 /*
1637  * xol_free_insn_slot - If slot was earlier allocated by
1638  * @xol_get_insn_slot(), make the slot available for
1639  * subsequent requests.
1640  */
1641 static void xol_free_insn_slot(struct task_struct *tsk)
1642 {
1643         struct xol_area *area;
1644         unsigned long vma_end;
1645         unsigned long slot_addr;
1646
1647         if (!tsk->mm || !tsk->mm->uprobes_state.xol_area || !tsk->utask)
1648                 return;
1649
1650         slot_addr = tsk->utask->xol_vaddr;
1651         if (unlikely(!slot_addr))
1652                 return;
1653
1654         area = tsk->mm->uprobes_state.xol_area;
1655         vma_end = area->vaddr + PAGE_SIZE;
1656         if (area->vaddr <= slot_addr && slot_addr < vma_end) {
1657                 unsigned long offset;
1658                 int slot_nr;
1659
1660                 offset = slot_addr - area->vaddr;
1661                 slot_nr = offset / UPROBE_XOL_SLOT_BYTES;
1662                 if (slot_nr >= UINSNS_PER_PAGE)
1663                         return;
1664
1665                 clear_bit(slot_nr, area->bitmap);
1666                 atomic_dec(&area->slot_count);
1667                 smp_mb__after_atomic(); /* pairs with prepare_to_wait() */
1668                 if (waitqueue_active(&area->wq))
1669                         wake_up(&area->wq);
1670
1671                 tsk->utask->xol_vaddr = 0;
1672         }
1673 }
1674
1675 void __weak arch_uprobe_copy_ixol(struct page *page, unsigned long vaddr,
1676                                   void *src, unsigned long len)
1677 {
1678         /* Initialize the slot */
1679         copy_to_page(page, vaddr, src, len);
1680
1681         /*
1682          * We probably need flush_icache_user_range() but it needs vma.
1683          * This should work on most of architectures by default. If
1684          * architecture needs to do something different it can define
1685          * its own version of the function.
1686          */
1687         flush_dcache_page(page);
1688 }
1689
1690 /**
1691  * uprobe_get_swbp_addr - compute address of swbp given post-swbp regs
1692  * @regs: Reflects the saved state of the task after it has hit a breakpoint
1693  * instruction.
1694  * Return the address of the breakpoint instruction.
1695  */
1696 unsigned long __weak uprobe_get_swbp_addr(struct pt_regs *regs)
1697 {
1698         return instruction_pointer(regs) - UPROBE_SWBP_INSN_SIZE;
1699 }
1700
1701 unsigned long uprobe_get_trap_addr(struct pt_regs *regs)
1702 {
1703         struct uprobe_task *utask = current->utask;
1704
1705         if (unlikely(utask && utask->active_uprobe))
1706                 return utask->vaddr;
1707
1708         return instruction_pointer(regs);
1709 }
1710
1711 static struct return_instance *free_ret_instance(struct return_instance *ri)
1712 {
1713         struct return_instance *next = ri->next;
1714         put_uprobe(ri->uprobe);
1715         kfree(ri);
1716         return next;
1717 }
1718
1719 /*
1720  * Called with no locks held.
1721  * Called in context of an exiting or an exec-ing thread.
1722  */
1723 void uprobe_free_utask(struct task_struct *t)
1724 {
1725         struct uprobe_task *utask = t->utask;
1726         struct return_instance *ri;
1727
1728         if (!utask)
1729                 return;
1730
1731         if (utask->active_uprobe)
1732                 put_uprobe(utask->active_uprobe);
1733
1734         ri = utask->return_instances;
1735         while (ri)
1736                 ri = free_ret_instance(ri);
1737
1738         xol_free_insn_slot(t);
1739         kfree(utask);
1740         t->utask = NULL;
1741 }
1742
1743 /*
1744  * Allocate a uprobe_task object for the task if if necessary.
1745  * Called when the thread hits a breakpoint.
1746  *
1747  * Returns:
1748  * - pointer to new uprobe_task on success
1749  * - NULL otherwise
1750  */
1751 static struct uprobe_task *get_utask(void)
1752 {
1753         if (!current->utask)
1754                 current->utask = kzalloc(sizeof(struct uprobe_task), GFP_KERNEL);
1755         return current->utask;
1756 }
1757
1758 static int dup_utask(struct task_struct *t, struct uprobe_task *o_utask)
1759 {
1760         struct uprobe_task *n_utask;
1761         struct return_instance **p, *o, *n;
1762
1763         n_utask = kzalloc(sizeof(struct uprobe_task), GFP_KERNEL);
1764         if (!n_utask)
1765                 return -ENOMEM;
1766         t->utask = n_utask;
1767
1768         p = &n_utask->return_instances;
1769         for (o = o_utask->return_instances; o; o = o->next) {
1770                 n = kmalloc(sizeof(struct return_instance), GFP_KERNEL);
1771                 if (!n)
1772                         return -ENOMEM;
1773
1774                 *n = *o;
1775                 get_uprobe(n->uprobe);
1776                 n->next = NULL;
1777
1778                 *p = n;
1779                 p = &n->next;
1780                 n_utask->depth++;
1781         }
1782
1783         return 0;
1784 }
1785
1786 static void uprobe_warn(struct task_struct *t, const char *msg)
1787 {
1788         pr_warn("uprobe: %s:%d failed to %s\n",
1789                         current->comm, current->pid, msg);
1790 }
1791
1792 static void dup_xol_work(struct callback_head *work)
1793 {
1794         if (current->flags & PF_EXITING)
1795                 return;
1796
1797         if (!__create_xol_area(current->utask->dup_xol_addr) &&
1798                         !fatal_signal_pending(current))
1799                 uprobe_warn(current, "dup xol area");
1800 }
1801
1802 /*
1803  * Called in context of a new clone/fork from copy_process.
1804  */
1805 void uprobe_copy_process(struct task_struct *t, unsigned long flags)
1806 {
1807         struct uprobe_task *utask = current->utask;
1808         struct mm_struct *mm = current->mm;
1809         struct xol_area *area;
1810
1811         t->utask = NULL;
1812
1813         if (!utask || !utask->return_instances)
1814                 return;
1815
1816         if (mm == t->mm && !(flags & CLONE_VFORK))
1817                 return;
1818
1819         if (dup_utask(t, utask))
1820                 return uprobe_warn(t, "dup ret instances");
1821
1822         /* The task can fork() after dup_xol_work() fails */
1823         area = mm->uprobes_state.xol_area;
1824         if (!area)
1825                 return uprobe_warn(t, "dup xol area");
1826
1827         if (mm == t->mm)
1828                 return;
1829
1830         t->utask->dup_xol_addr = area->vaddr;
1831         init_task_work(&t->utask->dup_xol_work, dup_xol_work);
1832         task_work_add(t, &t->utask->dup_xol_work, true);
1833 }
1834
1835 /*
1836  * Current area->vaddr notion assume the trampoline address is always
1837  * equal area->vaddr.
1838  *
1839  * Returns -1 in case the xol_area is not allocated.
1840  */
1841 static unsigned long get_trampoline_vaddr(void)
1842 {
1843         struct xol_area *area;
1844         unsigned long trampoline_vaddr = -1;
1845
1846         /* Pairs with xol_add_vma() smp_store_release() */
1847         area = READ_ONCE(current->mm->uprobes_state.xol_area); /* ^^^ */
1848         if (area)
1849                 trampoline_vaddr = area->vaddr;
1850
1851         return trampoline_vaddr;
1852 }
1853
1854 static void cleanup_return_instances(struct uprobe_task *utask, bool chained,
1855                                         struct pt_regs *regs)
1856 {
1857         struct return_instance *ri = utask->return_instances;
1858         enum rp_check ctx = chained ? RP_CHECK_CHAIN_CALL : RP_CHECK_CALL;
1859
1860         while (ri && !arch_uretprobe_is_alive(ri, ctx, regs)) {
1861                 ri = free_ret_instance(ri);
1862                 utask->depth--;
1863         }
1864         utask->return_instances = ri;
1865 }
1866
1867 static void prepare_uretprobe(struct uprobe *uprobe, struct pt_regs *regs)
1868 {
1869         struct return_instance *ri;
1870         struct uprobe_task *utask;
1871         unsigned long orig_ret_vaddr, trampoline_vaddr;
1872         bool chained;
1873
1874         if (!get_xol_area())
1875                 return;
1876
1877         utask = get_utask();
1878         if (!utask)
1879                 return;
1880
1881         if (utask->depth >= MAX_URETPROBE_DEPTH) {
1882                 printk_ratelimited(KERN_INFO "uprobe: omit uretprobe due to"
1883                                 " nestedness limit pid/tgid=%d/%d\n",
1884                                 current->pid, current->tgid);
1885                 return;
1886         }
1887
1888         ri = kmalloc(sizeof(struct return_instance), GFP_KERNEL);
1889         if (!ri)
1890                 return;
1891
1892         trampoline_vaddr = get_trampoline_vaddr();
1893         orig_ret_vaddr = arch_uretprobe_hijack_return_addr(trampoline_vaddr, regs);
1894         if (orig_ret_vaddr == -1)
1895                 goto fail;
1896
1897         /* drop the entries invalidated by longjmp() */
1898         chained = (orig_ret_vaddr == trampoline_vaddr);
1899         cleanup_return_instances(utask, chained, regs);
1900
1901         /*
1902          * We don't want to keep trampoline address in stack, rather keep the
1903          * original return address of first caller thru all the consequent
1904          * instances. This also makes breakpoint unwrapping easier.
1905          */
1906         if (chained) {
1907                 if (!utask->return_instances) {
1908                         /*
1909                          * This situation is not possible. Likely we have an
1910                          * attack from user-space.
1911                          */
1912                         uprobe_warn(current, "handle tail call");
1913                         goto fail;
1914                 }
1915                 orig_ret_vaddr = utask->return_instances->orig_ret_vaddr;
1916         }
1917
1918         ri->uprobe = get_uprobe(uprobe);
1919         ri->func = instruction_pointer(regs);
1920         ri->stack = user_stack_pointer(regs);
1921         ri->orig_ret_vaddr = orig_ret_vaddr;
1922         ri->chained = chained;
1923
1924         utask->depth++;
1925         ri->next = utask->return_instances;
1926         utask->return_instances = ri;
1927
1928         return;
1929  fail:
1930         kfree(ri);
1931 }
1932
1933 /* Prepare to single-step probed instruction out of line. */
1934 static int
1935 pre_ssout(struct uprobe *uprobe, struct pt_regs *regs, unsigned long bp_vaddr)
1936 {
1937         struct uprobe_task *utask;
1938         unsigned long xol_vaddr;
1939         int err;
1940
1941         utask = get_utask();
1942         if (!utask)
1943                 return -ENOMEM;
1944
1945         xol_vaddr = xol_get_insn_slot(uprobe);
1946         if (!xol_vaddr)
1947                 return -ENOMEM;
1948
1949         utask->xol_vaddr = xol_vaddr;
1950         utask->vaddr = bp_vaddr;
1951
1952         err = arch_uprobe_pre_xol(&uprobe->arch, regs);
1953         if (unlikely(err)) {
1954                 xol_free_insn_slot(current);
1955                 return err;
1956         }
1957
1958         utask->active_uprobe = uprobe;
1959         utask->state = UTASK_SSTEP;
1960         return 0;
1961 }
1962
1963 /*
1964  * If we are singlestepping, then ensure this thread is not connected to
1965  * non-fatal signals until completion of singlestep.  When xol insn itself
1966  * triggers the signal,  restart the original insn even if the task is
1967  * already SIGKILL'ed (since coredump should report the correct ip).  This
1968  * is even more important if the task has a handler for SIGSEGV/etc, The
1969  * _same_ instruction should be repeated again after return from the signal
1970  * handler, and SSTEP can never finish in this case.
1971  */
1972 bool uprobe_deny_signal(void)
1973 {
1974         struct task_struct *t = current;
1975         struct uprobe_task *utask = t->utask;
1976
1977         if (likely(!utask || !utask->active_uprobe))
1978                 return false;
1979
1980         WARN_ON_ONCE(utask->state != UTASK_SSTEP);
1981
1982         if (signal_pending(t)) {
1983                 spin_lock_irq(&t->sighand->siglock);
1984                 clear_tsk_thread_flag(t, TIF_SIGPENDING);
1985                 spin_unlock_irq(&t->sighand->siglock);
1986
1987                 if (__fatal_signal_pending(t) || arch_uprobe_xol_was_trapped(t)) {
1988                         utask->state = UTASK_SSTEP_TRAPPED;
1989                         set_tsk_thread_flag(t, TIF_UPROBE);
1990                 }
1991         }
1992
1993         return true;
1994 }
1995
1996 static void mmf_recalc_uprobes(struct mm_struct *mm)
1997 {
1998         struct vm_area_struct *vma;
1999
2000         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2001                 if (!valid_vma(vma, false))
2002                         continue;
2003                 /*
2004                  * This is not strictly accurate, we can race with
2005                  * uprobe_unregister() and see the already removed
2006                  * uprobe if delete_uprobe() was not yet called.
2007                  * Or this uprobe can be filtered out.
2008                  */
2009                 if (vma_has_uprobes(vma, vma->vm_start, vma->vm_end))
2010                         return;
2011         }
2012
2013         clear_bit(MMF_HAS_UPROBES, &mm->flags);
2014 }
2015
2016 static int is_trap_at_addr(struct mm_struct *mm, unsigned long vaddr)
2017 {
2018         struct page *page;
2019         uprobe_opcode_t opcode;
2020         int result;
2021
2022         if (WARN_ON_ONCE(!IS_ALIGNED(vaddr, UPROBE_SWBP_INSN_SIZE)))
2023                 return -EINVAL;
2024
2025         pagefault_disable();
2026         result = __get_user(opcode, (uprobe_opcode_t __user *)vaddr);
2027         pagefault_enable();
2028
2029         if (likely(result == 0))
2030                 goto out;
2031
2032         /*
2033          * The NULL 'tsk' here ensures that any faults that occur here
2034          * will not be accounted to the task.  'mm' *is* current->mm,
2035          * but we treat this as a 'remote' access since it is
2036          * essentially a kernel access to the memory.
2037          */
2038         result = get_user_pages_remote(NULL, mm, vaddr, 1, FOLL_FORCE, &page,
2039                         NULL, NULL);
2040         if (result < 0)
2041                 return result;
2042
2043         copy_from_page(page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
2044         put_page(page);
2045  out:
2046         /* This needs to return true for any variant of the trap insn */
2047         return is_trap_insn(&opcode);
2048 }
2049
2050 static struct uprobe *find_active_uprobe(unsigned long bp_vaddr, int *is_swbp)
2051 {
2052         struct mm_struct *mm = current->mm;
2053         struct uprobe *uprobe = NULL;
2054         struct vm_area_struct *vma;
2055
2056         down_read(&mm->mmap_sem);
2057         vma = find_vma(mm, bp_vaddr);
2058         if (vma && vma->vm_start <= bp_vaddr) {
2059                 if (valid_vma(vma, false)) {
2060                         struct inode *inode = file_inode(vma->vm_file);
2061                         loff_t offset = vaddr_to_offset(vma, bp_vaddr);
2062
2063                         uprobe = find_uprobe(inode, offset);
2064                 }
2065
2066                 if (!uprobe)
2067                         *is_swbp = is_trap_at_addr(mm, bp_vaddr);
2068         } else {
2069                 *is_swbp = -EFAULT;
2070         }
2071
2072         if (!uprobe && test_and_clear_bit(MMF_RECALC_UPROBES, &mm->flags))
2073                 mmf_recalc_uprobes(mm);
2074         up_read(&mm->mmap_sem);
2075
2076         return uprobe;
2077 }
2078
2079 static void handler_chain(struct uprobe *uprobe, struct pt_regs *regs)
2080 {
2081         struct uprobe_consumer *uc;
2082         int remove = UPROBE_HANDLER_REMOVE;
2083         bool need_prep = false; /* prepare return uprobe, when needed */
2084
2085         down_read(&uprobe->register_rwsem);
2086         for (uc = uprobe->consumers; uc; uc = uc->next) {
2087                 int rc = 0;
2088
2089                 if (uc->handler) {
2090                         rc = uc->handler(uc, regs);
2091                         WARN(rc & ~UPROBE_HANDLER_MASK,
2092                                 "bad rc=0x%x from %ps()\n", rc, uc->handler);
2093                 }
2094
2095                 if (uc->ret_handler)
2096                         need_prep = true;
2097
2098                 remove &= rc;
2099         }
2100
2101         if (need_prep && !remove)
2102                 prepare_uretprobe(uprobe, regs); /* put bp at return */
2103
2104         if (remove && uprobe->consumers) {
2105                 WARN_ON(!uprobe_is_active(uprobe));
2106                 unapply_uprobe(uprobe, current->mm);
2107         }
2108         up_read(&uprobe->register_rwsem);
2109 }
2110
2111 static void
2112 handle_uretprobe_chain(struct return_instance *ri, struct pt_regs *regs)
2113 {
2114         struct uprobe *uprobe = ri->uprobe;
2115         struct uprobe_consumer *uc;
2116
2117         down_read(&uprobe->register_rwsem);
2118         for (uc = uprobe->consumers; uc; uc = uc->next) {
2119                 if (uc->ret_handler)
2120                         uc->ret_handler(uc, ri->func, regs);
2121         }
2122         up_read(&uprobe->register_rwsem);
2123 }
2124
2125 static struct return_instance *find_next_ret_chain(struct return_instance *ri)
2126 {
2127         bool chained;
2128
2129         do {
2130                 chained = ri->chained;
2131                 ri = ri->next;  /* can't be NULL if chained */
2132         } while (chained);
2133
2134         return ri;
2135 }
2136
2137 static void handle_trampoline(struct pt_regs *regs)
2138 {
2139         struct uprobe_task *utask;
2140         struct return_instance *ri, *next;
2141         bool valid;
2142
2143         utask = current->utask;
2144         if (!utask)
2145                 goto sigill;
2146
2147         ri = utask->return_instances;
2148         if (!ri)
2149                 goto sigill;
2150
2151         do {
2152                 /*
2153                  * We should throw out the frames invalidated by longjmp().
2154                  * If this chain is valid, then the next one should be alive
2155                  * or NULL; the latter case means that nobody but ri->func
2156                  * could hit this trampoline on return. TODO: sigaltstack().
2157                  */
2158                 next = find_next_ret_chain(ri);
2159                 valid = !next || arch_uretprobe_is_alive(next, RP_CHECK_RET, regs);
2160
2161                 instruction_pointer_set(regs, ri->orig_ret_vaddr);
2162                 do {
2163                         if (valid)
2164                                 handle_uretprobe_chain(ri, regs);
2165                         ri = free_ret_instance(ri);
2166                         utask->depth--;
2167                 } while (ri != next);
2168         } while (!valid);
2169
2170         utask->return_instances = ri;
2171         return;
2172
2173  sigill:
2174         uprobe_warn(current, "handle uretprobe, sending SIGILL.");
2175         force_sig(SIGILL);
2176
2177 }
2178
2179 bool __weak arch_uprobe_ignore(struct arch_uprobe *aup, struct pt_regs *regs)
2180 {
2181         return false;
2182 }
2183
2184 bool __weak arch_uretprobe_is_alive(struct return_instance *ret, enum rp_check ctx,
2185                                         struct pt_regs *regs)
2186 {
2187         return true;
2188 }
2189
2190 /*
2191  * Run handler and ask thread to singlestep.
2192  * Ensure all non-fatal signals cannot interrupt thread while it singlesteps.
2193  */
2194 static void handle_swbp(struct pt_regs *regs)
2195 {
2196         struct uprobe *uprobe;
2197         unsigned long bp_vaddr;
2198         int is_swbp;
2199
2200         bp_vaddr = uprobe_get_swbp_addr(regs);
2201         if (bp_vaddr == get_trampoline_vaddr())
2202                 return handle_trampoline(regs);
2203
2204         uprobe = find_active_uprobe(bp_vaddr, &is_swbp);
2205         if (!uprobe) {
2206                 if (is_swbp > 0) {
2207                         /* No matching uprobe; signal SIGTRAP. */
2208                         force_sig(SIGTRAP);
2209                 } else {
2210                         /*
2211                          * Either we raced with uprobe_unregister() or we can't
2212                          * access this memory. The latter is only possible if
2213                          * another thread plays with our ->mm. In both cases
2214                          * we can simply restart. If this vma was unmapped we
2215                          * can pretend this insn was not executed yet and get
2216                          * the (correct) SIGSEGV after restart.
2217                          */
2218                         instruction_pointer_set(regs, bp_vaddr);
2219                 }
2220                 return;
2221         }
2222
2223         /* change it in advance for ->handler() and restart */
2224         instruction_pointer_set(regs, bp_vaddr);
2225
2226         /*
2227          * TODO: move copy_insn/etc into _register and remove this hack.
2228          * After we hit the bp, _unregister + _register can install the
2229          * new and not-yet-analyzed uprobe at the same address, restart.
2230          */
2231         if (unlikely(!test_bit(UPROBE_COPY_INSN, &uprobe->flags)))
2232                 goto out;
2233
2234         /*
2235          * Pairs with the smp_wmb() in prepare_uprobe().
2236          *
2237          * Guarantees that if we see the UPROBE_COPY_INSN bit set, then
2238          * we must also see the stores to &uprobe->arch performed by the
2239          * prepare_uprobe() call.
2240          */
2241         smp_rmb();
2242
2243         /* Tracing handlers use ->utask to communicate with fetch methods */
2244         if (!get_utask())
2245                 goto out;
2246
2247         if (arch_uprobe_ignore(&uprobe->arch, regs))
2248                 goto out;
2249
2250         handler_chain(uprobe, regs);
2251
2252         if (arch_uprobe_skip_sstep(&uprobe->arch, regs))
2253                 goto out;
2254
2255         if (!pre_ssout(uprobe, regs, bp_vaddr))
2256                 return;
2257
2258         /* arch_uprobe_skip_sstep() succeeded, or restart if can't singlestep */
2259 out:
2260         put_uprobe(uprobe);
2261 }
2262
2263 /*
2264  * Perform required fix-ups and disable singlestep.
2265  * Allow pending signals to take effect.
2266  */
2267 static void handle_singlestep(struct uprobe_task *utask, struct pt_regs *regs)
2268 {
2269         struct uprobe *uprobe;
2270         int err = 0;
2271
2272         uprobe = utask->active_uprobe;
2273         if (utask->state == UTASK_SSTEP_ACK)
2274                 err = arch_uprobe_post_xol(&uprobe->arch, regs);
2275         else if (utask->state == UTASK_SSTEP_TRAPPED)
2276                 arch_uprobe_abort_xol(&uprobe->arch, regs);
2277         else
2278                 WARN_ON_ONCE(1);
2279
2280         put_uprobe(uprobe);
2281         utask->active_uprobe = NULL;
2282         utask->state = UTASK_RUNNING;
2283         xol_free_insn_slot(current);
2284
2285         spin_lock_irq(&current->sighand->siglock);
2286         recalc_sigpending(); /* see uprobe_deny_signal() */
2287         spin_unlock_irq(&current->sighand->siglock);
2288
2289         if (unlikely(err)) {
2290                 uprobe_warn(current, "execute the probed insn, sending SIGILL.");
2291                 force_sig(SIGILL);
2292         }
2293 }
2294
2295 /*
2296  * On breakpoint hit, breakpoint notifier sets the TIF_UPROBE flag and
2297  * allows the thread to return from interrupt. After that handle_swbp()
2298  * sets utask->active_uprobe.
2299  *
2300  * On singlestep exception, singlestep notifier sets the TIF_UPROBE flag
2301  * and allows the thread to return from interrupt.
2302  *
2303  * While returning to userspace, thread notices the TIF_UPROBE flag and calls
2304  * uprobe_notify_resume().
2305  */
2306 void uprobe_notify_resume(struct pt_regs *regs)
2307 {
2308         struct uprobe_task *utask;
2309
2310         clear_thread_flag(TIF_UPROBE);
2311
2312         utask = current->utask;
2313         if (utask && utask->active_uprobe)
2314                 handle_singlestep(utask, regs);
2315         else
2316                 handle_swbp(regs);
2317 }
2318
2319 /*
2320  * uprobe_pre_sstep_notifier gets called from interrupt context as part of
2321  * notifier mechanism. Set TIF_UPROBE flag and indicate breakpoint hit.
2322  */
2323 int uprobe_pre_sstep_notifier(struct pt_regs *regs)
2324 {
2325         if (!current->mm)
2326                 return 0;
2327
2328         if (!test_bit(MMF_HAS_UPROBES, &current->mm->flags) &&
2329             (!current->utask || !current->utask->return_instances))
2330                 return 0;
2331
2332         set_thread_flag(TIF_UPROBE);
2333         return 1;
2334 }
2335
2336 /*
2337  * uprobe_post_sstep_notifier gets called in interrupt context as part of notifier
2338  * mechanism. Set TIF_UPROBE flag and indicate completion of singlestep.
2339  */
2340 int uprobe_post_sstep_notifier(struct pt_regs *regs)
2341 {
2342         struct uprobe_task *utask = current->utask;
2343
2344         if (!current->mm || !utask || !utask->active_uprobe)
2345                 /* task is currently not uprobed */
2346                 return 0;
2347
2348         utask->state = UTASK_SSTEP_ACK;
2349         set_thread_flag(TIF_UPROBE);
2350         return 1;
2351 }
2352
2353 static struct notifier_block uprobe_exception_nb = {
2354         .notifier_call          = arch_uprobe_exception_notify,
2355         .priority               = INT_MAX-1,    /* notified after kprobes, kgdb */
2356 };
2357
2358 void __init uprobes_init(void)
2359 {
2360         int i;
2361
2362         for (i = 0; i < UPROBES_HASH_SZ; i++)
2363                 mutex_init(&uprobes_mmap_mutex[i]);
2364
2365         BUG_ON(register_die_notifier(&uprobe_exception_nb));
2366 }