2bf792d22087caa979b7de851097aae90b0fd040
[muen/linux.git] / kernel / events / uprobes.c
1 /*
2  * User-space Probes (UProbes)
3  *
4  * This program is free software; you can redistribute it and/or modify
5  * it under the terms of the GNU General Public License as published by
6  * the Free Software Foundation; either version 2 of the License, or
7  * (at your option) any later version.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write to the Free Software
16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
17  *
18  * Copyright (C) IBM Corporation, 2008-2012
19  * Authors:
20  *      Srikar Dronamraju
21  *      Jim Keniston
22  * Copyright (C) 2011-2012 Red Hat, Inc., Peter Zijlstra
23  */
24
25 #include <linux/kernel.h>
26 #include <linux/highmem.h>
27 #include <linux/pagemap.h>      /* read_mapping_page */
28 #include <linux/slab.h>
29 #include <linux/sched.h>
30 #include <linux/sched/mm.h>
31 #include <linux/sched/coredump.h>
32 #include <linux/export.h>
33 #include <linux/rmap.h>         /* anon_vma_prepare */
34 #include <linux/mmu_notifier.h> /* set_pte_at_notify */
35 #include <linux/swap.h>         /* try_to_free_swap */
36 #include <linux/ptrace.h>       /* user_enable_single_step */
37 #include <linux/kdebug.h>       /* notifier mechanism */
38 #include "../../mm/internal.h"  /* munlock_vma_page */
39 #include <linux/percpu-rwsem.h>
40 #include <linux/task_work.h>
41 #include <linux/shmem_fs.h>
42
43 #include <linux/uprobes.h>
44
45 #define UINSNS_PER_PAGE                 (PAGE_SIZE/UPROBE_XOL_SLOT_BYTES)
46 #define MAX_UPROBE_XOL_SLOTS            UINSNS_PER_PAGE
47
48 static struct rb_root uprobes_tree = RB_ROOT;
49 /*
50  * allows us to skip the uprobe_mmap if there are no uprobe events active
51  * at this time.  Probably a fine grained per inode count is better?
52  */
53 #define no_uprobe_events()      RB_EMPTY_ROOT(&uprobes_tree)
54
55 static DEFINE_SPINLOCK(uprobes_treelock);       /* serialize rbtree access */
56
57 #define UPROBES_HASH_SZ 13
58 /* serialize uprobe->pending_list */
59 static struct mutex uprobes_mmap_mutex[UPROBES_HASH_SZ];
60 #define uprobes_mmap_hash(v)    (&uprobes_mmap_mutex[((unsigned long)(v)) % UPROBES_HASH_SZ])
61
62 static struct percpu_rw_semaphore dup_mmap_sem;
63
64 /* Have a copy of original instruction */
65 #define UPROBE_COPY_INSN        0
66
67 struct uprobe {
68         struct rb_node          rb_node;        /* node in the rb tree */
69         atomic_t                ref;
70         struct rw_semaphore     register_rwsem;
71         struct rw_semaphore     consumer_rwsem;
72         struct list_head        pending_list;
73         struct uprobe_consumer  *consumers;
74         struct inode            *inode;         /* Also hold a ref to inode */
75         loff_t                  offset;
76         unsigned long           flags;
77
78         /*
79          * The generic code assumes that it has two members of unknown type
80          * owned by the arch-specific code:
81          *
82          *      insn -  copy_insn() saves the original instruction here for
83          *              arch_uprobe_analyze_insn().
84          *
85          *      ixol -  potentially modified instruction to execute out of
86          *              line, copied to xol_area by xol_get_insn_slot().
87          */
88         struct arch_uprobe      arch;
89 };
90
91 /*
92  * Execute out of line area: anonymous executable mapping installed
93  * by the probed task to execute the copy of the original instruction
94  * mangled by set_swbp().
95  *
96  * On a breakpoint hit, thread contests for a slot.  It frees the
97  * slot after singlestep. Currently a fixed number of slots are
98  * allocated.
99  */
100 struct xol_area {
101         wait_queue_head_t               wq;             /* if all slots are busy */
102         atomic_t                        slot_count;     /* number of in-use slots */
103         unsigned long                   *bitmap;        /* 0 = free slot */
104
105         struct vm_special_mapping       xol_mapping;
106         struct page                     *pages[2];
107         /*
108          * We keep the vma's vm_start rather than a pointer to the vma
109          * itself.  The probed process or a naughty kernel module could make
110          * the vma go away, and we must handle that reasonably gracefully.
111          */
112         unsigned long                   vaddr;          /* Page(s) of instruction slots */
113 };
114
115 /*
116  * valid_vma: Verify if the specified vma is an executable vma
117  * Relax restrictions while unregistering: vm_flags might have
118  * changed after breakpoint was inserted.
119  *      - is_register: indicates if we are in register context.
120  *      - Return 1 if the specified virtual address is in an
121  *        executable vma.
122  */
123 static bool valid_vma(struct vm_area_struct *vma, bool is_register)
124 {
125         vm_flags_t flags = VM_HUGETLB | VM_MAYEXEC | VM_MAYSHARE;
126
127         if (is_register)
128                 flags |= VM_WRITE;
129
130         return vma->vm_file && (vma->vm_flags & flags) == VM_MAYEXEC;
131 }
132
133 static unsigned long offset_to_vaddr(struct vm_area_struct *vma, loff_t offset)
134 {
135         return vma->vm_start + offset - ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
136 }
137
138 static loff_t vaddr_to_offset(struct vm_area_struct *vma, unsigned long vaddr)
139 {
140         return ((loff_t)vma->vm_pgoff << PAGE_SHIFT) + (vaddr - vma->vm_start);
141 }
142
143 /**
144  * __replace_page - replace page in vma by new page.
145  * based on replace_page in mm/ksm.c
146  *
147  * @vma:      vma that holds the pte pointing to page
148  * @addr:     address the old @page is mapped at
149  * @page:     the cowed page we are replacing by kpage
150  * @kpage:    the modified page we replace page by
151  *
152  * Returns 0 on success, -EFAULT on failure.
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 = old_page,
160                 .vma = vma,
161                 .address = addr,
162         };
163         int err;
164         /* For mmu_notifiers */
165         const unsigned long mmun_start = addr;
166         const unsigned long mmun_end   = addr + PAGE_SIZE;
167         struct mem_cgroup *memcg;
168
169         VM_BUG_ON_PAGE(PageTransHuge(old_page), old_page);
170
171         err = mem_cgroup_try_charge(new_page, vma->vm_mm, GFP_KERNEL, &memcg,
172                         false);
173         if (err)
174                 return err;
175
176         /* For try_to_free_swap() and munlock_vma_page() below */
177         lock_page(old_page);
178
179         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
180         err = -EAGAIN;
181         if (!page_vma_mapped_walk(&pvmw)) {
182                 mem_cgroup_cancel_charge(new_page, memcg, false);
183                 goto unlock;
184         }
185         VM_BUG_ON_PAGE(addr != pvmw.address, old_page);
186
187         get_page(new_page);
188         page_add_new_anon_rmap(new_page, vma, addr, false);
189         mem_cgroup_commit_charge(new_page, memcg, false, false);
190         lru_cache_add_active_or_unevictable(new_page, vma);
191
192         if (!PageAnon(old_page)) {
193                 dec_mm_counter(mm, mm_counter_file(old_page));
194                 inc_mm_counter(mm, MM_ANONPAGES);
195         }
196
197         flush_cache_page(vma, addr, pte_pfn(*pvmw.pte));
198         ptep_clear_flush_notify(vma, addr, pvmw.pte);
199         set_pte_at_notify(mm, addr, pvmw.pte,
200                         mk_pte(new_page, vma->vm_page_prot));
201
202         page_remove_rmap(old_page, false);
203         if (!page_mapped(old_page))
204                 try_to_free_swap(old_page);
205         page_vma_mapped_walk_done(&pvmw);
206
207         if (vma->vm_flags & VM_LOCKED)
208                 munlock_vma_page(old_page);
209         put_page(old_page);
210
211         err = 0;
212  unlock:
213         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
214         unlock_page(old_page);
215         return err;
216 }
217
218 /**
219  * is_swbp_insn - check if instruction is breakpoint instruction.
220  * @insn: instruction to be checked.
221  * Default implementation of is_swbp_insn
222  * Returns true if @insn is a breakpoint instruction.
223  */
224 bool __weak is_swbp_insn(uprobe_opcode_t *insn)
225 {
226         return *insn == UPROBE_SWBP_INSN;
227 }
228
229 /**
230  * is_trap_insn - check if instruction is breakpoint instruction.
231  * @insn: instruction to be checked.
232  * Default implementation of is_trap_insn
233  * Returns true if @insn is a breakpoint instruction.
234  *
235  * This function is needed for the case where an architecture has multiple
236  * trap instructions (like powerpc).
237  */
238 bool __weak is_trap_insn(uprobe_opcode_t *insn)
239 {
240         return is_swbp_insn(insn);
241 }
242
243 static void copy_from_page(struct page *page, unsigned long vaddr, void *dst, int len)
244 {
245         void *kaddr = kmap_atomic(page);
246         memcpy(dst, kaddr + (vaddr & ~PAGE_MASK), len);
247         kunmap_atomic(kaddr);
248 }
249
250 static void copy_to_page(struct page *page, unsigned long vaddr, const void *src, int len)
251 {
252         void *kaddr = kmap_atomic(page);
253         memcpy(kaddr + (vaddr & ~PAGE_MASK), src, len);
254         kunmap_atomic(kaddr);
255 }
256
257 static int verify_opcode(struct page *page, unsigned long vaddr, uprobe_opcode_t *new_opcode)
258 {
259         uprobe_opcode_t old_opcode;
260         bool is_swbp;
261
262         /*
263          * Note: We only check if the old_opcode is UPROBE_SWBP_INSN here.
264          * We do not check if it is any other 'trap variant' which could
265          * be conditional trap instruction such as the one powerpc supports.
266          *
267          * The logic is that we do not care if the underlying instruction
268          * is a trap variant; uprobes always wins over any other (gdb)
269          * breakpoint.
270          */
271         copy_from_page(page, vaddr, &old_opcode, UPROBE_SWBP_INSN_SIZE);
272         is_swbp = is_swbp_insn(&old_opcode);
273
274         if (is_swbp_insn(new_opcode)) {
275                 if (is_swbp)            /* register: already installed? */
276                         return 0;
277         } else {
278                 if (!is_swbp)           /* unregister: was it changed by us? */
279                         return 0;
280         }
281
282         return 1;
283 }
284
285 /*
286  * NOTE:
287  * Expect the breakpoint instruction to be the smallest size instruction for
288  * the architecture. If an arch has variable length instruction and the
289  * breakpoint instruction is not of the smallest length instruction
290  * supported by that architecture then we need to modify is_trap_at_addr and
291  * uprobe_write_opcode accordingly. This would never be a problem for archs
292  * that have fixed length instructions.
293  *
294  * uprobe_write_opcode - write the opcode at a given virtual address.
295  * @mm: the probed process address space.
296  * @vaddr: the virtual address to store the opcode.
297  * @opcode: opcode to be written at @vaddr.
298  *
299  * Called with mm->mmap_sem held for write.
300  * Return 0 (success) or a negative errno.
301  */
302 int uprobe_write_opcode(struct arch_uprobe *auprobe, struct mm_struct *mm,
303                         unsigned long vaddr, uprobe_opcode_t opcode)
304 {
305         struct page *old_page, *new_page;
306         struct vm_area_struct *vma;
307         int ret;
308
309 retry:
310         /* Read the page with vaddr into memory */
311         ret = get_user_pages_remote(NULL, mm, vaddr, 1,
312                         FOLL_FORCE | FOLL_SPLIT, &old_page, &vma, NULL);
313         if (ret <= 0)
314                 return ret;
315
316         ret = verify_opcode(old_page, vaddr, &opcode);
317         if (ret <= 0)
318                 goto put_old;
319
320         ret = anon_vma_prepare(vma);
321         if (ret)
322                 goto put_old;
323
324         ret = -ENOMEM;
325         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vaddr);
326         if (!new_page)
327                 goto put_old;
328
329         __SetPageUptodate(new_page);
330         copy_highpage(new_page, old_page);
331         copy_to_page(new_page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
332
333         ret = __replace_page(vma, vaddr, old_page, new_page);
334         put_page(new_page);
335 put_old:
336         put_page(old_page);
337
338         if (unlikely(ret == -EAGAIN))
339                 goto retry;
340         return ret;
341 }
342
343 /**
344  * set_swbp - store breakpoint at a given address.
345  * @auprobe: arch specific probepoint information.
346  * @mm: the probed process address space.
347  * @vaddr: the virtual address to insert the opcode.
348  *
349  * For mm @mm, store the breakpoint instruction at @vaddr.
350  * Return 0 (success) or a negative errno.
351  */
352 int __weak set_swbp(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
353 {
354         return uprobe_write_opcode(auprobe, mm, vaddr, UPROBE_SWBP_INSN);
355 }
356
357 /**
358  * set_orig_insn - Restore the original instruction.
359  * @mm: the probed process address space.
360  * @auprobe: arch specific probepoint information.
361  * @vaddr: the virtual address to insert the opcode.
362  *
363  * For mm @mm, restore the original opcode (opcode) at @vaddr.
364  * Return 0 (success) or a negative errno.
365  */
366 int __weak
367 set_orig_insn(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
368 {
369         return uprobe_write_opcode(auprobe, mm, vaddr,
370                         *(uprobe_opcode_t *)&auprobe->insn);
371 }
372
373 static struct uprobe *get_uprobe(struct uprobe *uprobe)
374 {
375         atomic_inc(&uprobe->ref);
376         return uprobe;
377 }
378
379 static void put_uprobe(struct uprobe *uprobe)
380 {
381         if (atomic_dec_and_test(&uprobe->ref))
382                 kfree(uprobe);
383 }
384
385 static int match_uprobe(struct uprobe *l, struct uprobe *r)
386 {
387         if (l->inode < r->inode)
388                 return -1;
389
390         if (l->inode > r->inode)
391                 return 1;
392
393         if (l->offset < r->offset)
394                 return -1;
395
396         if (l->offset > r->offset)
397                 return 1;
398
399         return 0;
400 }
401
402 static struct uprobe *__find_uprobe(struct inode *inode, loff_t offset)
403 {
404         struct uprobe u = { .inode = inode, .offset = offset };
405         struct rb_node *n = uprobes_tree.rb_node;
406         struct uprobe *uprobe;
407         int match;
408
409         while (n) {
410                 uprobe = rb_entry(n, struct uprobe, rb_node);
411                 match = match_uprobe(&u, uprobe);
412                 if (!match)
413                         return get_uprobe(uprobe);
414
415                 if (match < 0)
416                         n = n->rb_left;
417                 else
418                         n = n->rb_right;
419         }
420         return NULL;
421 }
422
423 /*
424  * Find a uprobe corresponding to a given inode:offset
425  * Acquires uprobes_treelock
426  */
427 static struct uprobe *find_uprobe(struct inode *inode, loff_t offset)
428 {
429         struct uprobe *uprobe;
430
431         spin_lock(&uprobes_treelock);
432         uprobe = __find_uprobe(inode, offset);
433         spin_unlock(&uprobes_treelock);
434
435         return uprobe;
436 }
437
438 static struct uprobe *__insert_uprobe(struct uprobe *uprobe)
439 {
440         struct rb_node **p = &uprobes_tree.rb_node;
441         struct rb_node *parent = NULL;
442         struct uprobe *u;
443         int match;
444
445         while (*p) {
446                 parent = *p;
447                 u = rb_entry(parent, struct uprobe, rb_node);
448                 match = match_uprobe(uprobe, u);
449                 if (!match)
450                         return get_uprobe(u);
451
452                 if (match < 0)
453                         p = &parent->rb_left;
454                 else
455                         p = &parent->rb_right;
456
457         }
458
459         u = NULL;
460         rb_link_node(&uprobe->rb_node, parent, p);
461         rb_insert_color(&uprobe->rb_node, &uprobes_tree);
462         /* get access + creation ref */
463         atomic_set(&uprobe->ref, 2);
464
465         return u;
466 }
467
468 /*
469  * Acquire uprobes_treelock.
470  * Matching uprobe already exists in rbtree;
471  *      increment (access refcount) and return the matching uprobe.
472  *
473  * No matching uprobe; insert the uprobe in rb_tree;
474  *      get a double refcount (access + creation) and return NULL.
475  */
476 static struct uprobe *insert_uprobe(struct uprobe *uprobe)
477 {
478         struct uprobe *u;
479
480         spin_lock(&uprobes_treelock);
481         u = __insert_uprobe(uprobe);
482         spin_unlock(&uprobes_treelock);
483
484         return u;
485 }
486
487 static struct uprobe *alloc_uprobe(struct inode *inode, loff_t offset)
488 {
489         struct uprobe *uprobe, *cur_uprobe;
490
491         uprobe = kzalloc(sizeof(struct uprobe), GFP_KERNEL);
492         if (!uprobe)
493                 return NULL;
494
495         uprobe->inode = inode;
496         uprobe->offset = offset;
497         init_rwsem(&uprobe->register_rwsem);
498         init_rwsem(&uprobe->consumer_rwsem);
499
500         /* add to uprobes_tree, sorted on inode:offset */
501         cur_uprobe = insert_uprobe(uprobe);
502         /* a uprobe exists for this inode:offset combination */
503         if (cur_uprobe) {
504                 kfree(uprobe);
505                 uprobe = cur_uprobe;
506         }
507
508         return uprobe;
509 }
510
511 static void consumer_add(struct uprobe *uprobe, struct uprobe_consumer *uc)
512 {
513         down_write(&uprobe->consumer_rwsem);
514         uc->next = uprobe->consumers;
515         uprobe->consumers = uc;
516         up_write(&uprobe->consumer_rwsem);
517 }
518
519 /*
520  * For uprobe @uprobe, delete the consumer @uc.
521  * Return true if the @uc is deleted successfully
522  * or return false.
523  */
524 static bool consumer_del(struct uprobe *uprobe, struct uprobe_consumer *uc)
525 {
526         struct uprobe_consumer **con;
527         bool ret = false;
528
529         down_write(&uprobe->consumer_rwsem);
530         for (con = &uprobe->consumers; *con; con = &(*con)->next) {
531                 if (*con == uc) {
532                         *con = uc->next;
533                         ret = true;
534                         break;
535                 }
536         }
537         up_write(&uprobe->consumer_rwsem);
538
539         return ret;
540 }
541
542 static int __copy_insn(struct address_space *mapping, struct file *filp,
543                         void *insn, int nbytes, loff_t offset)
544 {
545         struct page *page;
546         /*
547          * Ensure that the page that has the original instruction is populated
548          * and in page-cache. If ->readpage == NULL it must be shmem_mapping(),
549          * see uprobe_register().
550          */
551         if (mapping->a_ops->readpage)
552                 page = read_mapping_page(mapping, offset >> PAGE_SHIFT, filp);
553         else
554                 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
555         if (IS_ERR(page))
556                 return PTR_ERR(page);
557
558         copy_from_page(page, offset, insn, nbytes);
559         put_page(page);
560
561         return 0;
562 }
563
564 static int copy_insn(struct uprobe *uprobe, struct file *filp)
565 {
566         struct address_space *mapping = uprobe->inode->i_mapping;
567         loff_t offs = uprobe->offset;
568         void *insn = &uprobe->arch.insn;
569         int size = sizeof(uprobe->arch.insn);
570         int len, err = -EIO;
571
572         /* Copy only available bytes, -EIO if nothing was read */
573         do {
574                 if (offs >= i_size_read(uprobe->inode))
575                         break;
576
577                 len = min_t(int, size, PAGE_SIZE - (offs & ~PAGE_MASK));
578                 err = __copy_insn(mapping, filp, insn, len, offs);
579                 if (err)
580                         break;
581
582                 insn += len;
583                 offs += len;
584                 size -= len;
585         } while (size);
586
587         return err;
588 }
589
590 static int prepare_uprobe(struct uprobe *uprobe, struct file *file,
591                                 struct mm_struct *mm, unsigned long vaddr)
592 {
593         int ret = 0;
594
595         if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
596                 return ret;
597
598         /* TODO: move this into _register, until then we abuse this sem. */
599         down_write(&uprobe->consumer_rwsem);
600         if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
601                 goto out;
602
603         ret = copy_insn(uprobe, file);
604         if (ret)
605                 goto out;
606
607         ret = -ENOTSUPP;
608         if (is_trap_insn((uprobe_opcode_t *)&uprobe->arch.insn))
609                 goto out;
610
611         ret = arch_uprobe_analyze_insn(&uprobe->arch, mm, vaddr);
612         if (ret)
613                 goto out;
614
615         /* uprobe_write_opcode() assumes we don't cross page boundary */
616         BUG_ON((uprobe->offset & ~PAGE_MASK) +
617                         UPROBE_SWBP_INSN_SIZE > PAGE_SIZE);
618
619         smp_wmb(); /* pairs with rmb() in find_active_uprobe() */
620         set_bit(UPROBE_COPY_INSN, &uprobe->flags);
621
622  out:
623         up_write(&uprobe->consumer_rwsem);
624
625         return ret;
626 }
627
628 static inline bool consumer_filter(struct uprobe_consumer *uc,
629                                    enum uprobe_filter_ctx ctx, struct mm_struct *mm)
630 {
631         return !uc->filter || uc->filter(uc, ctx, mm);
632 }
633
634 static bool filter_chain(struct uprobe *uprobe,
635                          enum uprobe_filter_ctx ctx, struct mm_struct *mm)
636 {
637         struct uprobe_consumer *uc;
638         bool ret = false;
639
640         down_read(&uprobe->consumer_rwsem);
641         for (uc = uprobe->consumers; uc; uc = uc->next) {
642                 ret = consumer_filter(uc, ctx, mm);
643                 if (ret)
644                         break;
645         }
646         up_read(&uprobe->consumer_rwsem);
647
648         return ret;
649 }
650
651 static int
652 install_breakpoint(struct uprobe *uprobe, struct mm_struct *mm,
653                         struct vm_area_struct *vma, unsigned long vaddr)
654 {
655         bool first_uprobe;
656         int ret;
657
658         ret = prepare_uprobe(uprobe, vma->vm_file, mm, vaddr);
659         if (ret)
660                 return ret;
661
662         /*
663          * set MMF_HAS_UPROBES in advance for uprobe_pre_sstep_notifier(),
664          * the task can hit this breakpoint right after __replace_page().
665          */
666         first_uprobe = !test_bit(MMF_HAS_UPROBES, &mm->flags);
667         if (first_uprobe)
668                 set_bit(MMF_HAS_UPROBES, &mm->flags);
669
670         ret = set_swbp(&uprobe->arch, mm, vaddr);
671         if (!ret)
672                 clear_bit(MMF_RECALC_UPROBES, &mm->flags);
673         else if (first_uprobe)
674                 clear_bit(MMF_HAS_UPROBES, &mm->flags);
675
676         return ret;
677 }
678
679 static int
680 remove_breakpoint(struct uprobe *uprobe, struct mm_struct *mm, unsigned long vaddr)
681 {
682         set_bit(MMF_RECALC_UPROBES, &mm->flags);
683         return set_orig_insn(&uprobe->arch, mm, vaddr);
684 }
685
686 static inline bool uprobe_is_active(struct uprobe *uprobe)
687 {
688         return !RB_EMPTY_NODE(&uprobe->rb_node);
689 }
690 /*
691  * There could be threads that have already hit the breakpoint. They
692  * will recheck the current insn and restart if find_uprobe() fails.
693  * See find_active_uprobe().
694  */
695 static void delete_uprobe(struct uprobe *uprobe)
696 {
697         if (WARN_ON(!uprobe_is_active(uprobe)))
698                 return;
699
700         spin_lock(&uprobes_treelock);
701         rb_erase(&uprobe->rb_node, &uprobes_tree);
702         spin_unlock(&uprobes_treelock);
703         RB_CLEAR_NODE(&uprobe->rb_node); /* for uprobe_is_active() */
704         put_uprobe(uprobe);
705 }
706
707 struct map_info {
708         struct map_info *next;
709         struct mm_struct *mm;
710         unsigned long vaddr;
711 };
712
713 static inline struct map_info *free_map_info(struct map_info *info)
714 {
715         struct map_info *next = info->next;
716         kfree(info);
717         return next;
718 }
719
720 static struct map_info *
721 build_map_info(struct address_space *mapping, loff_t offset, bool is_register)
722 {
723         unsigned long pgoff = offset >> PAGE_SHIFT;
724         struct vm_area_struct *vma;
725         struct map_info *curr = NULL;
726         struct map_info *prev = NULL;
727         struct map_info *info;
728         int more = 0;
729
730  again:
731         i_mmap_lock_read(mapping);
732         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
733                 if (!valid_vma(vma, is_register))
734                         continue;
735
736                 if (!prev && !more) {
737                         /*
738                          * Needs GFP_NOWAIT to avoid i_mmap_rwsem recursion through
739                          * reclaim. This is optimistic, no harm done if it fails.
740                          */
741                         prev = kmalloc(sizeof(struct map_info),
742                                         GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
743                         if (prev)
744                                 prev->next = NULL;
745                 }
746                 if (!prev) {
747                         more++;
748                         continue;
749                 }
750
751                 if (!mmget_not_zero(vma->vm_mm))
752                         continue;
753
754                 info = prev;
755                 prev = prev->next;
756                 info->next = curr;
757                 curr = info;
758
759                 info->mm = vma->vm_mm;
760                 info->vaddr = offset_to_vaddr(vma, offset);
761         }
762         i_mmap_unlock_read(mapping);
763
764         if (!more)
765                 goto out;
766
767         prev = curr;
768         while (curr) {
769                 mmput(curr->mm);
770                 curr = curr->next;
771         }
772
773         do {
774                 info = kmalloc(sizeof(struct map_info), GFP_KERNEL);
775                 if (!info) {
776                         curr = ERR_PTR(-ENOMEM);
777                         goto out;
778                 }
779                 info->next = prev;
780                 prev = info;
781         } while (--more);
782
783         goto again;
784  out:
785         while (prev)
786                 prev = free_map_info(prev);
787         return curr;
788 }
789
790 static int
791 register_for_each_vma(struct uprobe *uprobe, struct uprobe_consumer *new)
792 {
793         bool is_register = !!new;
794         struct map_info *info;
795         int err = 0;
796
797         percpu_down_write(&dup_mmap_sem);
798         info = build_map_info(uprobe->inode->i_mapping,
799                                         uprobe->offset, is_register);
800         if (IS_ERR(info)) {
801                 err = PTR_ERR(info);
802                 goto out;
803         }
804
805         while (info) {
806                 struct mm_struct *mm = info->mm;
807                 struct vm_area_struct *vma;
808
809                 if (err && is_register)
810                         goto free;
811
812                 down_write(&mm->mmap_sem);
813                 vma = find_vma(mm, info->vaddr);
814                 if (!vma || !valid_vma(vma, is_register) ||
815                     file_inode(vma->vm_file) != uprobe->inode)
816                         goto unlock;
817
818                 if (vma->vm_start > info->vaddr ||
819                     vaddr_to_offset(vma, info->vaddr) != uprobe->offset)
820                         goto unlock;
821
822                 if (is_register) {
823                         /* consult only the "caller", new consumer. */
824                         if (consumer_filter(new,
825                                         UPROBE_FILTER_REGISTER, mm))
826                                 err = install_breakpoint(uprobe, mm, vma, info->vaddr);
827                 } else if (test_bit(MMF_HAS_UPROBES, &mm->flags)) {
828                         if (!filter_chain(uprobe,
829                                         UPROBE_FILTER_UNREGISTER, mm))
830                                 err |= remove_breakpoint(uprobe, mm, info->vaddr);
831                 }
832
833  unlock:
834                 up_write(&mm->mmap_sem);
835  free:
836                 mmput(mm);
837                 info = free_map_info(info);
838         }
839  out:
840         percpu_up_write(&dup_mmap_sem);
841         return err;
842 }
843
844 static void
845 __uprobe_unregister(struct uprobe *uprobe, struct uprobe_consumer *uc)
846 {
847         int err;
848
849         if (WARN_ON(!consumer_del(uprobe, uc)))
850                 return;
851
852         err = register_for_each_vma(uprobe, NULL);
853         /* TODO : cant unregister? schedule a worker thread */
854         if (!uprobe->consumers && !err)
855                 delete_uprobe(uprobe);
856 }
857
858 /*
859  * uprobe_unregister - unregister an already registered probe.
860  * @inode: the file in which the probe has to be removed.
861  * @offset: offset from the start of the file.
862  * @uc: identify which probe if multiple probes are colocated.
863  */
864 void uprobe_unregister(struct inode *inode, loff_t offset, struct uprobe_consumer *uc)
865 {
866         struct uprobe *uprobe;
867
868         uprobe = find_uprobe(inode, offset);
869         if (WARN_ON(!uprobe))
870                 return;
871
872         down_write(&uprobe->register_rwsem);
873         __uprobe_unregister(uprobe, uc);
874         up_write(&uprobe->register_rwsem);
875         put_uprobe(uprobe);
876 }
877 EXPORT_SYMBOL_GPL(uprobe_unregister);
878
879 /*
880  * __uprobe_register - register a probe
881  * @inode: the file in which the probe has to be placed.
882  * @offset: offset from the start of the file.
883  * @uc: information on howto handle the probe..
884  *
885  * Apart from the access refcount, __uprobe_register() takes a creation
886  * refcount (thro alloc_uprobe) if and only if this @uprobe is getting
887  * inserted into the rbtree (i.e first consumer for a @inode:@offset
888  * tuple).  Creation refcount stops uprobe_unregister from freeing the
889  * @uprobe even before the register operation is complete. Creation
890  * refcount is released when the last @uc for the @uprobe
891  * unregisters. Caller of __uprobe_register() is required to keep @inode
892  * (and the containing mount) referenced.
893  *
894  * Return errno if it cannot successully install probes
895  * else return 0 (success)
896  */
897 static int __uprobe_register(struct inode *inode, loff_t offset,
898                              struct uprobe_consumer *uc)
899 {
900         struct uprobe *uprobe;
901         int ret;
902
903         /* Uprobe must have at least one set consumer */
904         if (!uc->handler && !uc->ret_handler)
905                 return -EINVAL;
906
907         /* copy_insn() uses read_mapping_page() or shmem_read_mapping_page() */
908         if (!inode->i_mapping->a_ops->readpage && !shmem_mapping(inode->i_mapping))
909                 return -EIO;
910         /* Racy, just to catch the obvious mistakes */
911         if (offset > i_size_read(inode))
912                 return -EINVAL;
913
914  retry:
915         uprobe = alloc_uprobe(inode, offset);
916         if (!uprobe)
917                 return -ENOMEM;
918         /*
919          * We can race with uprobe_unregister()->delete_uprobe().
920          * Check uprobe_is_active() and retry if it is false.
921          */
922         down_write(&uprobe->register_rwsem);
923         ret = -EAGAIN;
924         if (likely(uprobe_is_active(uprobe))) {
925                 consumer_add(uprobe, uc);
926                 ret = register_for_each_vma(uprobe, uc);
927                 if (ret)
928                         __uprobe_unregister(uprobe, uc);
929         }
930         up_write(&uprobe->register_rwsem);
931         put_uprobe(uprobe);
932
933         if (unlikely(ret == -EAGAIN))
934                 goto retry;
935         return ret;
936 }
937
938 int uprobe_register(struct inode *inode, loff_t offset,
939                     struct uprobe_consumer *uc)
940 {
941         return __uprobe_register(inode, offset, uc);
942 }
943 EXPORT_SYMBOL_GPL(uprobe_register);
944
945 /*
946  * uprobe_apply - unregister an already registered probe.
947  * @inode: the file in which the probe has to be removed.
948  * @offset: offset from the start of the file.
949  * @uc: consumer which wants to add more or remove some breakpoints
950  * @add: add or remove the breakpoints
951  */
952 int uprobe_apply(struct inode *inode, loff_t offset,
953                         struct uprobe_consumer *uc, bool add)
954 {
955         struct uprobe *uprobe;
956         struct uprobe_consumer *con;
957         int ret = -ENOENT;
958
959         uprobe = find_uprobe(inode, offset);
960         if (WARN_ON(!uprobe))
961                 return ret;
962
963         down_write(&uprobe->register_rwsem);
964         for (con = uprobe->consumers; con && con != uc ; con = con->next)
965                 ;
966         if (con)
967                 ret = register_for_each_vma(uprobe, add ? uc : NULL);
968         up_write(&uprobe->register_rwsem);
969         put_uprobe(uprobe);
970
971         return ret;
972 }
973
974 static int unapply_uprobe(struct uprobe *uprobe, struct mm_struct *mm)
975 {
976         struct vm_area_struct *vma;
977         int err = 0;
978
979         down_read(&mm->mmap_sem);
980         for (vma = mm->mmap; vma; vma = vma->vm_next) {
981                 unsigned long vaddr;
982                 loff_t offset;
983
984                 if (!valid_vma(vma, false) ||
985                     file_inode(vma->vm_file) != uprobe->inode)
986                         continue;
987
988                 offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
989                 if (uprobe->offset <  offset ||
990                     uprobe->offset >= offset + vma->vm_end - vma->vm_start)
991                         continue;
992
993                 vaddr = offset_to_vaddr(vma, uprobe->offset);
994                 err |= remove_breakpoint(uprobe, mm, vaddr);
995         }
996         up_read(&mm->mmap_sem);
997
998         return err;
999 }
1000
1001 static struct rb_node *
1002 find_node_in_range(struct inode *inode, loff_t min, loff_t max)
1003 {
1004         struct rb_node *n = uprobes_tree.rb_node;
1005
1006         while (n) {
1007                 struct uprobe *u = rb_entry(n, struct uprobe, rb_node);
1008
1009                 if (inode < u->inode) {
1010                         n = n->rb_left;
1011                 } else if (inode > u->inode) {
1012                         n = n->rb_right;
1013                 } else {
1014                         if (max < u->offset)
1015                                 n = n->rb_left;
1016                         else if (min > u->offset)
1017                                 n = n->rb_right;
1018                         else
1019                                 break;
1020                 }
1021         }
1022
1023         return n;
1024 }
1025
1026 /*
1027  * For a given range in vma, build a list of probes that need to be inserted.
1028  */
1029 static void build_probe_list(struct inode *inode,
1030                                 struct vm_area_struct *vma,
1031                                 unsigned long start, unsigned long end,
1032                                 struct list_head *head)
1033 {
1034         loff_t min, max;
1035         struct rb_node *n, *t;
1036         struct uprobe *u;
1037
1038         INIT_LIST_HEAD(head);
1039         min = vaddr_to_offset(vma, start);
1040         max = min + (end - start) - 1;
1041
1042         spin_lock(&uprobes_treelock);
1043         n = find_node_in_range(inode, min, max);
1044         if (n) {
1045                 for (t = n; t; t = rb_prev(t)) {
1046                         u = rb_entry(t, struct uprobe, rb_node);
1047                         if (u->inode != inode || u->offset < min)
1048                                 break;
1049                         list_add(&u->pending_list, head);
1050                         get_uprobe(u);
1051                 }
1052                 for (t = n; (t = rb_next(t)); ) {
1053                         u = rb_entry(t, struct uprobe, rb_node);
1054                         if (u->inode != inode || u->offset > max)
1055                                 break;
1056                         list_add(&u->pending_list, head);
1057                         get_uprobe(u);
1058                 }
1059         }
1060         spin_unlock(&uprobes_treelock);
1061 }
1062
1063 /*
1064  * Called from mmap_region/vma_adjust with mm->mmap_sem acquired.
1065  *
1066  * Currently we ignore all errors and always return 0, the callers
1067  * can't handle the failure anyway.
1068  */
1069 int uprobe_mmap(struct vm_area_struct *vma)
1070 {
1071         struct list_head tmp_list;
1072         struct uprobe *uprobe, *u;
1073         struct inode *inode;
1074
1075         if (no_uprobe_events() || !valid_vma(vma, true))
1076                 return 0;
1077
1078         inode = file_inode(vma->vm_file);
1079         if (!inode)
1080                 return 0;
1081
1082         mutex_lock(uprobes_mmap_hash(inode));
1083         build_probe_list(inode, vma, vma->vm_start, vma->vm_end, &tmp_list);
1084         /*
1085          * We can race with uprobe_unregister(), this uprobe can be already
1086          * removed. But in this case filter_chain() must return false, all
1087          * consumers have gone away.
1088          */
1089         list_for_each_entry_safe(uprobe, u, &tmp_list, pending_list) {
1090                 if (!fatal_signal_pending(current) &&
1091                     filter_chain(uprobe, UPROBE_FILTER_MMAP, vma->vm_mm)) {
1092                         unsigned long vaddr = offset_to_vaddr(vma, uprobe->offset);
1093                         install_breakpoint(uprobe, vma->vm_mm, vma, vaddr);
1094                 }
1095                 put_uprobe(uprobe);
1096         }
1097         mutex_unlock(uprobes_mmap_hash(inode));
1098
1099         return 0;
1100 }
1101
1102 static bool
1103 vma_has_uprobes(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1104 {
1105         loff_t min, max;
1106         struct inode *inode;
1107         struct rb_node *n;
1108
1109         inode = file_inode(vma->vm_file);
1110
1111         min = vaddr_to_offset(vma, start);
1112         max = min + (end - start) - 1;
1113
1114         spin_lock(&uprobes_treelock);
1115         n = find_node_in_range(inode, min, max);
1116         spin_unlock(&uprobes_treelock);
1117
1118         return !!n;
1119 }
1120
1121 /*
1122  * Called in context of a munmap of a vma.
1123  */
1124 void uprobe_munmap(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1125 {
1126         if (no_uprobe_events() || !valid_vma(vma, false))
1127                 return;
1128
1129         if (!atomic_read(&vma->vm_mm->mm_users)) /* called by mmput() ? */
1130                 return;
1131
1132         if (!test_bit(MMF_HAS_UPROBES, &vma->vm_mm->flags) ||
1133              test_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags))
1134                 return;
1135
1136         if (vma_has_uprobes(vma, start, end))
1137                 set_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags);
1138 }
1139
1140 /* Slot allocation for XOL */
1141 static int xol_add_vma(struct mm_struct *mm, struct xol_area *area)
1142 {
1143         struct vm_area_struct *vma;
1144         int ret;
1145
1146         if (down_write_killable(&mm->mmap_sem))
1147                 return -EINTR;
1148
1149         if (mm->uprobes_state.xol_area) {
1150                 ret = -EALREADY;
1151                 goto fail;
1152         }
1153
1154         if (!area->vaddr) {
1155                 /* Try to map as high as possible, this is only a hint. */
1156                 area->vaddr = get_unmapped_area(NULL, TASK_SIZE - PAGE_SIZE,
1157                                                 PAGE_SIZE, 0, 0);
1158                 if (area->vaddr & ~PAGE_MASK) {
1159                         ret = area->vaddr;
1160                         goto fail;
1161                 }
1162         }
1163
1164         vma = _install_special_mapping(mm, area->vaddr, PAGE_SIZE,
1165                                 VM_EXEC|VM_MAYEXEC|VM_DONTCOPY|VM_IO,
1166                                 &area->xol_mapping);
1167         if (IS_ERR(vma)) {
1168                 ret = PTR_ERR(vma);
1169                 goto fail;
1170         }
1171
1172         ret = 0;
1173         /* pairs with get_xol_area() */
1174         smp_store_release(&mm->uprobes_state.xol_area, area); /* ^^^ */
1175  fail:
1176         up_write(&mm->mmap_sem);
1177
1178         return ret;
1179 }
1180
1181 static struct xol_area *__create_xol_area(unsigned long vaddr)
1182 {
1183         struct mm_struct *mm = current->mm;
1184         uprobe_opcode_t insn = UPROBE_SWBP_INSN;
1185         struct xol_area *area;
1186
1187         area = kmalloc(sizeof(*area), GFP_KERNEL);
1188         if (unlikely(!area))
1189                 goto out;
1190
1191         area->bitmap = kcalloc(BITS_TO_LONGS(UINSNS_PER_PAGE), sizeof(long),
1192                                GFP_KERNEL);
1193         if (!area->bitmap)
1194                 goto free_area;
1195
1196         area->xol_mapping.name = "[uprobes]";
1197         area->xol_mapping.fault = NULL;
1198         area->xol_mapping.pages = area->pages;
1199         area->pages[0] = alloc_page(GFP_HIGHUSER);
1200         if (!area->pages[0])
1201                 goto free_bitmap;
1202         area->pages[1] = NULL;
1203
1204         area->vaddr = vaddr;
1205         init_waitqueue_head(&area->wq);
1206         /* Reserve the 1st slot for get_trampoline_vaddr() */
1207         set_bit(0, area->bitmap);
1208         atomic_set(&area->slot_count, 1);
1209         arch_uprobe_copy_ixol(area->pages[0], 0, &insn, UPROBE_SWBP_INSN_SIZE);
1210
1211         if (!xol_add_vma(mm, area))
1212                 return area;
1213
1214         __free_page(area->pages[0]);
1215  free_bitmap:
1216         kfree(area->bitmap);
1217  free_area:
1218         kfree(area);
1219  out:
1220         return NULL;
1221 }
1222
1223 /*
1224  * get_xol_area - Allocate process's xol_area if necessary.
1225  * This area will be used for storing instructions for execution out of line.
1226  *
1227  * Returns the allocated area or NULL.
1228  */
1229 static struct xol_area *get_xol_area(void)
1230 {
1231         struct mm_struct *mm = current->mm;
1232         struct xol_area *area;
1233
1234         if (!mm->uprobes_state.xol_area)
1235                 __create_xol_area(0);
1236
1237         /* Pairs with xol_add_vma() smp_store_release() */
1238         area = READ_ONCE(mm->uprobes_state.xol_area); /* ^^^ */
1239         return area;
1240 }
1241
1242 /*
1243  * uprobe_clear_state - Free the area allocated for slots.
1244  */
1245 void uprobe_clear_state(struct mm_struct *mm)
1246 {
1247         struct xol_area *area = mm->uprobes_state.xol_area;
1248
1249         if (!area)
1250                 return;
1251
1252         put_page(area->pages[0]);
1253         kfree(area->bitmap);
1254         kfree(area);
1255 }
1256
1257 void uprobe_start_dup_mmap(void)
1258 {
1259         percpu_down_read(&dup_mmap_sem);
1260 }
1261
1262 void uprobe_end_dup_mmap(void)
1263 {
1264         percpu_up_read(&dup_mmap_sem);
1265 }
1266
1267 void uprobe_dup_mmap(struct mm_struct *oldmm, struct mm_struct *newmm)
1268 {
1269         if (test_bit(MMF_HAS_UPROBES, &oldmm->flags)) {
1270                 set_bit(MMF_HAS_UPROBES, &newmm->flags);
1271                 /* unconditionally, dup_mmap() skips VM_DONTCOPY vmas */
1272                 set_bit(MMF_RECALC_UPROBES, &newmm->flags);
1273         }
1274 }
1275
1276 /*
1277  *  - search for a free slot.
1278  */
1279 static unsigned long xol_take_insn_slot(struct xol_area *area)
1280 {
1281         unsigned long slot_addr;
1282         int slot_nr;
1283
1284         do {
1285                 slot_nr = find_first_zero_bit(area->bitmap, UINSNS_PER_PAGE);
1286                 if (slot_nr < UINSNS_PER_PAGE) {
1287                         if (!test_and_set_bit(slot_nr, area->bitmap))
1288                                 break;
1289
1290                         slot_nr = UINSNS_PER_PAGE;
1291                         continue;
1292                 }
1293                 wait_event(area->wq, (atomic_read(&area->slot_count) < UINSNS_PER_PAGE));
1294         } while (slot_nr >= UINSNS_PER_PAGE);
1295
1296         slot_addr = area->vaddr + (slot_nr * UPROBE_XOL_SLOT_BYTES);
1297         atomic_inc(&area->slot_count);
1298
1299         return slot_addr;
1300 }
1301
1302 /*
1303  * xol_get_insn_slot - allocate a slot for xol.
1304  * Returns the allocated slot address or 0.
1305  */
1306 static unsigned long xol_get_insn_slot(struct uprobe *uprobe)
1307 {
1308         struct xol_area *area;
1309         unsigned long xol_vaddr;
1310
1311         area = get_xol_area();
1312         if (!area)
1313                 return 0;
1314
1315         xol_vaddr = xol_take_insn_slot(area);
1316         if (unlikely(!xol_vaddr))
1317                 return 0;
1318
1319         arch_uprobe_copy_ixol(area->pages[0], xol_vaddr,
1320                               &uprobe->arch.ixol, sizeof(uprobe->arch.ixol));
1321
1322         return xol_vaddr;
1323 }
1324
1325 /*
1326  * xol_free_insn_slot - If slot was earlier allocated by
1327  * @xol_get_insn_slot(), make the slot available for
1328  * subsequent requests.
1329  */
1330 static void xol_free_insn_slot(struct task_struct *tsk)
1331 {
1332         struct xol_area *area;
1333         unsigned long vma_end;
1334         unsigned long slot_addr;
1335
1336         if (!tsk->mm || !tsk->mm->uprobes_state.xol_area || !tsk->utask)
1337                 return;
1338
1339         slot_addr = tsk->utask->xol_vaddr;
1340         if (unlikely(!slot_addr))
1341                 return;
1342
1343         area = tsk->mm->uprobes_state.xol_area;
1344         vma_end = area->vaddr + PAGE_SIZE;
1345         if (area->vaddr <= slot_addr && slot_addr < vma_end) {
1346                 unsigned long offset;
1347                 int slot_nr;
1348
1349                 offset = slot_addr - area->vaddr;
1350                 slot_nr = offset / UPROBE_XOL_SLOT_BYTES;
1351                 if (slot_nr >= UINSNS_PER_PAGE)
1352                         return;
1353
1354                 clear_bit(slot_nr, area->bitmap);
1355                 atomic_dec(&area->slot_count);
1356                 smp_mb__after_atomic(); /* pairs with prepare_to_wait() */
1357                 if (waitqueue_active(&area->wq))
1358                         wake_up(&area->wq);
1359
1360                 tsk->utask->xol_vaddr = 0;
1361         }
1362 }
1363
1364 void __weak arch_uprobe_copy_ixol(struct page *page, unsigned long vaddr,
1365                                   void *src, unsigned long len)
1366 {
1367         /* Initialize the slot */
1368         copy_to_page(page, vaddr, src, len);
1369
1370         /*
1371          * We probably need flush_icache_user_range() but it needs vma.
1372          * This should work on most of architectures by default. If
1373          * architecture needs to do something different it can define
1374          * its own version of the function.
1375          */
1376         flush_dcache_page(page);
1377 }
1378
1379 /**
1380  * uprobe_get_swbp_addr - compute address of swbp given post-swbp regs
1381  * @regs: Reflects the saved state of the task after it has hit a breakpoint
1382  * instruction.
1383  * Return the address of the breakpoint instruction.
1384  */
1385 unsigned long __weak uprobe_get_swbp_addr(struct pt_regs *regs)
1386 {
1387         return instruction_pointer(regs) - UPROBE_SWBP_INSN_SIZE;
1388 }
1389
1390 unsigned long uprobe_get_trap_addr(struct pt_regs *regs)
1391 {
1392         struct uprobe_task *utask = current->utask;
1393
1394         if (unlikely(utask && utask->active_uprobe))
1395                 return utask->vaddr;
1396
1397         return instruction_pointer(regs);
1398 }
1399
1400 static struct return_instance *free_ret_instance(struct return_instance *ri)
1401 {
1402         struct return_instance *next = ri->next;
1403         put_uprobe(ri->uprobe);
1404         kfree(ri);
1405         return next;
1406 }
1407
1408 /*
1409  * Called with no locks held.
1410  * Called in context of an exiting or an exec-ing thread.
1411  */
1412 void uprobe_free_utask(struct task_struct *t)
1413 {
1414         struct uprobe_task *utask = t->utask;
1415         struct return_instance *ri;
1416
1417         if (!utask)
1418                 return;
1419
1420         if (utask->active_uprobe)
1421                 put_uprobe(utask->active_uprobe);
1422
1423         ri = utask->return_instances;
1424         while (ri)
1425                 ri = free_ret_instance(ri);
1426
1427         xol_free_insn_slot(t);
1428         kfree(utask);
1429         t->utask = NULL;
1430 }
1431
1432 /*
1433  * Allocate a uprobe_task object for the task if if necessary.
1434  * Called when the thread hits a breakpoint.
1435  *
1436  * Returns:
1437  * - pointer to new uprobe_task on success
1438  * - NULL otherwise
1439  */
1440 static struct uprobe_task *get_utask(void)
1441 {
1442         if (!current->utask)
1443                 current->utask = kzalloc(sizeof(struct uprobe_task), GFP_KERNEL);
1444         return current->utask;
1445 }
1446
1447 static int dup_utask(struct task_struct *t, struct uprobe_task *o_utask)
1448 {
1449         struct uprobe_task *n_utask;
1450         struct return_instance **p, *o, *n;
1451
1452         n_utask = kzalloc(sizeof(struct uprobe_task), GFP_KERNEL);
1453         if (!n_utask)
1454                 return -ENOMEM;
1455         t->utask = n_utask;
1456
1457         p = &n_utask->return_instances;
1458         for (o = o_utask->return_instances; o; o = o->next) {
1459                 n = kmalloc(sizeof(struct return_instance), GFP_KERNEL);
1460                 if (!n)
1461                         return -ENOMEM;
1462
1463                 *n = *o;
1464                 get_uprobe(n->uprobe);
1465                 n->next = NULL;
1466
1467                 *p = n;
1468                 p = &n->next;
1469                 n_utask->depth++;
1470         }
1471
1472         return 0;
1473 }
1474
1475 static void uprobe_warn(struct task_struct *t, const char *msg)
1476 {
1477         pr_warn("uprobe: %s:%d failed to %s\n",
1478                         current->comm, current->pid, msg);
1479 }
1480
1481 static void dup_xol_work(struct callback_head *work)
1482 {
1483         if (current->flags & PF_EXITING)
1484                 return;
1485
1486         if (!__create_xol_area(current->utask->dup_xol_addr) &&
1487                         !fatal_signal_pending(current))
1488                 uprobe_warn(current, "dup xol area");
1489 }
1490
1491 /*
1492  * Called in context of a new clone/fork from copy_process.
1493  */
1494 void uprobe_copy_process(struct task_struct *t, unsigned long flags)
1495 {
1496         struct uprobe_task *utask = current->utask;
1497         struct mm_struct *mm = current->mm;
1498         struct xol_area *area;
1499
1500         t->utask = NULL;
1501
1502         if (!utask || !utask->return_instances)
1503                 return;
1504
1505         if (mm == t->mm && !(flags & CLONE_VFORK))
1506                 return;
1507
1508         if (dup_utask(t, utask))
1509                 return uprobe_warn(t, "dup ret instances");
1510
1511         /* The task can fork() after dup_xol_work() fails */
1512         area = mm->uprobes_state.xol_area;
1513         if (!area)
1514                 return uprobe_warn(t, "dup xol area");
1515
1516         if (mm == t->mm)
1517                 return;
1518
1519         t->utask->dup_xol_addr = area->vaddr;
1520         init_task_work(&t->utask->dup_xol_work, dup_xol_work);
1521         task_work_add(t, &t->utask->dup_xol_work, true);
1522 }
1523
1524 /*
1525  * Current area->vaddr notion assume the trampoline address is always
1526  * equal area->vaddr.
1527  *
1528  * Returns -1 in case the xol_area is not allocated.
1529  */
1530 static unsigned long get_trampoline_vaddr(void)
1531 {
1532         struct xol_area *area;
1533         unsigned long trampoline_vaddr = -1;
1534
1535         /* Pairs with xol_add_vma() smp_store_release() */
1536         area = READ_ONCE(current->mm->uprobes_state.xol_area); /* ^^^ */
1537         if (area)
1538                 trampoline_vaddr = area->vaddr;
1539
1540         return trampoline_vaddr;
1541 }
1542
1543 static void cleanup_return_instances(struct uprobe_task *utask, bool chained,
1544                                         struct pt_regs *regs)
1545 {
1546         struct return_instance *ri = utask->return_instances;
1547         enum rp_check ctx = chained ? RP_CHECK_CHAIN_CALL : RP_CHECK_CALL;
1548
1549         while (ri && !arch_uretprobe_is_alive(ri, ctx, regs)) {
1550                 ri = free_ret_instance(ri);
1551                 utask->depth--;
1552         }
1553         utask->return_instances = ri;
1554 }
1555
1556 static void prepare_uretprobe(struct uprobe *uprobe, struct pt_regs *regs)
1557 {
1558         struct return_instance *ri;
1559         struct uprobe_task *utask;
1560         unsigned long orig_ret_vaddr, trampoline_vaddr;
1561         bool chained;
1562
1563         if (!get_xol_area())
1564                 return;
1565
1566         utask = get_utask();
1567         if (!utask)
1568                 return;
1569
1570         if (utask->depth >= MAX_URETPROBE_DEPTH) {
1571                 printk_ratelimited(KERN_INFO "uprobe: omit uretprobe due to"
1572                                 " nestedness limit pid/tgid=%d/%d\n",
1573                                 current->pid, current->tgid);
1574                 return;
1575         }
1576
1577         ri = kmalloc(sizeof(struct return_instance), GFP_KERNEL);
1578         if (!ri)
1579                 return;
1580
1581         trampoline_vaddr = get_trampoline_vaddr();
1582         orig_ret_vaddr = arch_uretprobe_hijack_return_addr(trampoline_vaddr, regs);
1583         if (orig_ret_vaddr == -1)
1584                 goto fail;
1585
1586         /* drop the entries invalidated by longjmp() */
1587         chained = (orig_ret_vaddr == trampoline_vaddr);
1588         cleanup_return_instances(utask, chained, regs);
1589
1590         /*
1591          * We don't want to keep trampoline address in stack, rather keep the
1592          * original return address of first caller thru all the consequent
1593          * instances. This also makes breakpoint unwrapping easier.
1594          */
1595         if (chained) {
1596                 if (!utask->return_instances) {
1597                         /*
1598                          * This situation is not possible. Likely we have an
1599                          * attack from user-space.
1600                          */
1601                         uprobe_warn(current, "handle tail call");
1602                         goto fail;
1603                 }
1604                 orig_ret_vaddr = utask->return_instances->orig_ret_vaddr;
1605         }
1606
1607         ri->uprobe = get_uprobe(uprobe);
1608         ri->func = instruction_pointer(regs);
1609         ri->stack = user_stack_pointer(regs);
1610         ri->orig_ret_vaddr = orig_ret_vaddr;
1611         ri->chained = chained;
1612
1613         utask->depth++;
1614         ri->next = utask->return_instances;
1615         utask->return_instances = ri;
1616
1617         return;
1618  fail:
1619         kfree(ri);
1620 }
1621
1622 /* Prepare to single-step probed instruction out of line. */
1623 static int
1624 pre_ssout(struct uprobe *uprobe, struct pt_regs *regs, unsigned long bp_vaddr)
1625 {
1626         struct uprobe_task *utask;
1627         unsigned long xol_vaddr;
1628         int err;
1629
1630         utask = get_utask();
1631         if (!utask)
1632                 return -ENOMEM;
1633
1634         xol_vaddr = xol_get_insn_slot(uprobe);
1635         if (!xol_vaddr)
1636                 return -ENOMEM;
1637
1638         utask->xol_vaddr = xol_vaddr;
1639         utask->vaddr = bp_vaddr;
1640
1641         err = arch_uprobe_pre_xol(&uprobe->arch, regs);
1642         if (unlikely(err)) {
1643                 xol_free_insn_slot(current);
1644                 return err;
1645         }
1646
1647         utask->active_uprobe = uprobe;
1648         utask->state = UTASK_SSTEP;
1649         return 0;
1650 }
1651
1652 /*
1653  * If we are singlestepping, then ensure this thread is not connected to
1654  * non-fatal signals until completion of singlestep.  When xol insn itself
1655  * triggers the signal,  restart the original insn even if the task is
1656  * already SIGKILL'ed (since coredump should report the correct ip).  This
1657  * is even more important if the task has a handler for SIGSEGV/etc, The
1658  * _same_ instruction should be repeated again after return from the signal
1659  * handler, and SSTEP can never finish in this case.
1660  */
1661 bool uprobe_deny_signal(void)
1662 {
1663         struct task_struct *t = current;
1664         struct uprobe_task *utask = t->utask;
1665
1666         if (likely(!utask || !utask->active_uprobe))
1667                 return false;
1668
1669         WARN_ON_ONCE(utask->state != UTASK_SSTEP);
1670
1671         if (signal_pending(t)) {
1672                 spin_lock_irq(&t->sighand->siglock);
1673                 clear_tsk_thread_flag(t, TIF_SIGPENDING);
1674                 spin_unlock_irq(&t->sighand->siglock);
1675
1676                 if (__fatal_signal_pending(t) || arch_uprobe_xol_was_trapped(t)) {
1677                         utask->state = UTASK_SSTEP_TRAPPED;
1678                         set_tsk_thread_flag(t, TIF_UPROBE);
1679                 }
1680         }
1681
1682         return true;
1683 }
1684
1685 static void mmf_recalc_uprobes(struct mm_struct *mm)
1686 {
1687         struct vm_area_struct *vma;
1688
1689         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1690                 if (!valid_vma(vma, false))
1691                         continue;
1692                 /*
1693                  * This is not strictly accurate, we can race with
1694                  * uprobe_unregister() and see the already removed
1695                  * uprobe if delete_uprobe() was not yet called.
1696                  * Or this uprobe can be filtered out.
1697                  */
1698                 if (vma_has_uprobes(vma, vma->vm_start, vma->vm_end))
1699                         return;
1700         }
1701
1702         clear_bit(MMF_HAS_UPROBES, &mm->flags);
1703 }
1704
1705 static int is_trap_at_addr(struct mm_struct *mm, unsigned long vaddr)
1706 {
1707         struct page *page;
1708         uprobe_opcode_t opcode;
1709         int result;
1710
1711         pagefault_disable();
1712         result = __get_user(opcode, (uprobe_opcode_t __user *)vaddr);
1713         pagefault_enable();
1714
1715         if (likely(result == 0))
1716                 goto out;
1717
1718         /*
1719          * The NULL 'tsk' here ensures that any faults that occur here
1720          * will not be accounted to the task.  'mm' *is* current->mm,
1721          * but we treat this as a 'remote' access since it is
1722          * essentially a kernel access to the memory.
1723          */
1724         result = get_user_pages_remote(NULL, mm, vaddr, 1, FOLL_FORCE, &page,
1725                         NULL, NULL);
1726         if (result < 0)
1727                 return result;
1728
1729         copy_from_page(page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
1730         put_page(page);
1731  out:
1732         /* This needs to return true for any variant of the trap insn */
1733         return is_trap_insn(&opcode);
1734 }
1735
1736 static struct uprobe *find_active_uprobe(unsigned long bp_vaddr, int *is_swbp)
1737 {
1738         struct mm_struct *mm = current->mm;
1739         struct uprobe *uprobe = NULL;
1740         struct vm_area_struct *vma;
1741
1742         down_read(&mm->mmap_sem);
1743         vma = find_vma(mm, bp_vaddr);
1744         if (vma && vma->vm_start <= bp_vaddr) {
1745                 if (valid_vma(vma, false)) {
1746                         struct inode *inode = file_inode(vma->vm_file);
1747                         loff_t offset = vaddr_to_offset(vma, bp_vaddr);
1748
1749                         uprobe = find_uprobe(inode, offset);
1750                 }
1751
1752                 if (!uprobe)
1753                         *is_swbp = is_trap_at_addr(mm, bp_vaddr);
1754         } else {
1755                 *is_swbp = -EFAULT;
1756         }
1757
1758         if (!uprobe && test_and_clear_bit(MMF_RECALC_UPROBES, &mm->flags))
1759                 mmf_recalc_uprobes(mm);
1760         up_read(&mm->mmap_sem);
1761
1762         return uprobe;
1763 }
1764
1765 static void handler_chain(struct uprobe *uprobe, struct pt_regs *regs)
1766 {
1767         struct uprobe_consumer *uc;
1768         int remove = UPROBE_HANDLER_REMOVE;
1769         bool need_prep = false; /* prepare return uprobe, when needed */
1770
1771         down_read(&uprobe->register_rwsem);
1772         for (uc = uprobe->consumers; uc; uc = uc->next) {
1773                 int rc = 0;
1774
1775                 if (uc->handler) {
1776                         rc = uc->handler(uc, regs);
1777                         WARN(rc & ~UPROBE_HANDLER_MASK,
1778                                 "bad rc=0x%x from %pf()\n", rc, uc->handler);
1779                 }
1780
1781                 if (uc->ret_handler)
1782                         need_prep = true;
1783
1784                 remove &= rc;
1785         }
1786
1787         if (need_prep && !remove)
1788                 prepare_uretprobe(uprobe, regs); /* put bp at return */
1789
1790         if (remove && uprobe->consumers) {
1791                 WARN_ON(!uprobe_is_active(uprobe));
1792                 unapply_uprobe(uprobe, current->mm);
1793         }
1794         up_read(&uprobe->register_rwsem);
1795 }
1796
1797 static void
1798 handle_uretprobe_chain(struct return_instance *ri, struct pt_regs *regs)
1799 {
1800         struct uprobe *uprobe = ri->uprobe;
1801         struct uprobe_consumer *uc;
1802
1803         down_read(&uprobe->register_rwsem);
1804         for (uc = uprobe->consumers; uc; uc = uc->next) {
1805                 if (uc->ret_handler)
1806                         uc->ret_handler(uc, ri->func, regs);
1807         }
1808         up_read(&uprobe->register_rwsem);
1809 }
1810
1811 static struct return_instance *find_next_ret_chain(struct return_instance *ri)
1812 {
1813         bool chained;
1814
1815         do {
1816                 chained = ri->chained;
1817                 ri = ri->next;  /* can't be NULL if chained */
1818         } while (chained);
1819
1820         return ri;
1821 }
1822
1823 static void handle_trampoline(struct pt_regs *regs)
1824 {
1825         struct uprobe_task *utask;
1826         struct return_instance *ri, *next;
1827         bool valid;
1828
1829         utask = current->utask;
1830         if (!utask)
1831                 goto sigill;
1832
1833         ri = utask->return_instances;
1834         if (!ri)
1835                 goto sigill;
1836
1837         do {
1838                 /*
1839                  * We should throw out the frames invalidated by longjmp().
1840                  * If this chain is valid, then the next one should be alive
1841                  * or NULL; the latter case means that nobody but ri->func
1842                  * could hit this trampoline on return. TODO: sigaltstack().
1843                  */
1844                 next = find_next_ret_chain(ri);
1845                 valid = !next || arch_uretprobe_is_alive(next, RP_CHECK_RET, regs);
1846
1847                 instruction_pointer_set(regs, ri->orig_ret_vaddr);
1848                 do {
1849                         if (valid)
1850                                 handle_uretprobe_chain(ri, regs);
1851                         ri = free_ret_instance(ri);
1852                         utask->depth--;
1853                 } while (ri != next);
1854         } while (!valid);
1855
1856         utask->return_instances = ri;
1857         return;
1858
1859  sigill:
1860         uprobe_warn(current, "handle uretprobe, sending SIGILL.");
1861         force_sig(SIGILL, current);
1862
1863 }
1864
1865 bool __weak arch_uprobe_ignore(struct arch_uprobe *aup, struct pt_regs *regs)
1866 {
1867         return false;
1868 }
1869
1870 bool __weak arch_uretprobe_is_alive(struct return_instance *ret, enum rp_check ctx,
1871                                         struct pt_regs *regs)
1872 {
1873         return true;
1874 }
1875
1876 /*
1877  * Run handler and ask thread to singlestep.
1878  * Ensure all non-fatal signals cannot interrupt thread while it singlesteps.
1879  */
1880 static void handle_swbp(struct pt_regs *regs)
1881 {
1882         struct uprobe *uprobe;
1883         unsigned long bp_vaddr;
1884         int uninitialized_var(is_swbp);
1885
1886         bp_vaddr = uprobe_get_swbp_addr(regs);
1887         if (bp_vaddr == get_trampoline_vaddr())
1888                 return handle_trampoline(regs);
1889
1890         uprobe = find_active_uprobe(bp_vaddr, &is_swbp);
1891         if (!uprobe) {
1892                 if (is_swbp > 0) {
1893                         /* No matching uprobe; signal SIGTRAP. */
1894                         send_sig(SIGTRAP, current, 0);
1895                 } else {
1896                         /*
1897                          * Either we raced with uprobe_unregister() or we can't
1898                          * access this memory. The latter is only possible if
1899                          * another thread plays with our ->mm. In both cases
1900                          * we can simply restart. If this vma was unmapped we
1901                          * can pretend this insn was not executed yet and get
1902                          * the (correct) SIGSEGV after restart.
1903                          */
1904                         instruction_pointer_set(regs, bp_vaddr);
1905                 }
1906                 return;
1907         }
1908
1909         /* change it in advance for ->handler() and restart */
1910         instruction_pointer_set(regs, bp_vaddr);
1911
1912         /*
1913          * TODO: move copy_insn/etc into _register and remove this hack.
1914          * After we hit the bp, _unregister + _register can install the
1915          * new and not-yet-analyzed uprobe at the same address, restart.
1916          */
1917         smp_rmb(); /* pairs with wmb() in install_breakpoint() */
1918         if (unlikely(!test_bit(UPROBE_COPY_INSN, &uprobe->flags)))
1919                 goto out;
1920
1921         /* Tracing handlers use ->utask to communicate with fetch methods */
1922         if (!get_utask())
1923                 goto out;
1924
1925         if (arch_uprobe_ignore(&uprobe->arch, regs))
1926                 goto out;
1927
1928         handler_chain(uprobe, regs);
1929
1930         if (arch_uprobe_skip_sstep(&uprobe->arch, regs))
1931                 goto out;
1932
1933         if (!pre_ssout(uprobe, regs, bp_vaddr))
1934                 return;
1935
1936         /* arch_uprobe_skip_sstep() succeeded, or restart if can't singlestep */
1937 out:
1938         put_uprobe(uprobe);
1939 }
1940
1941 /*
1942  * Perform required fix-ups and disable singlestep.
1943  * Allow pending signals to take effect.
1944  */
1945 static void handle_singlestep(struct uprobe_task *utask, struct pt_regs *regs)
1946 {
1947         struct uprobe *uprobe;
1948         int err = 0;
1949
1950         uprobe = utask->active_uprobe;
1951         if (utask->state == UTASK_SSTEP_ACK)
1952                 err = arch_uprobe_post_xol(&uprobe->arch, regs);
1953         else if (utask->state == UTASK_SSTEP_TRAPPED)
1954                 arch_uprobe_abort_xol(&uprobe->arch, regs);
1955         else
1956                 WARN_ON_ONCE(1);
1957
1958         put_uprobe(uprobe);
1959         utask->active_uprobe = NULL;
1960         utask->state = UTASK_RUNNING;
1961         xol_free_insn_slot(current);
1962
1963         spin_lock_irq(&current->sighand->siglock);
1964         recalc_sigpending(); /* see uprobe_deny_signal() */
1965         spin_unlock_irq(&current->sighand->siglock);
1966
1967         if (unlikely(err)) {
1968                 uprobe_warn(current, "execute the probed insn, sending SIGILL.");
1969                 force_sig(SIGILL, current);
1970         }
1971 }
1972
1973 /*
1974  * On breakpoint hit, breakpoint notifier sets the TIF_UPROBE flag and
1975  * allows the thread to return from interrupt. After that handle_swbp()
1976  * sets utask->active_uprobe.
1977  *
1978  * On singlestep exception, singlestep notifier sets the TIF_UPROBE flag
1979  * and allows the thread to return from interrupt.
1980  *
1981  * While returning to userspace, thread notices the TIF_UPROBE flag and calls
1982  * uprobe_notify_resume().
1983  */
1984 void uprobe_notify_resume(struct pt_regs *regs)
1985 {
1986         struct uprobe_task *utask;
1987
1988         clear_thread_flag(TIF_UPROBE);
1989
1990         utask = current->utask;
1991         if (utask && utask->active_uprobe)
1992                 handle_singlestep(utask, regs);
1993         else
1994                 handle_swbp(regs);
1995 }
1996
1997 /*
1998  * uprobe_pre_sstep_notifier gets called from interrupt context as part of
1999  * notifier mechanism. Set TIF_UPROBE flag and indicate breakpoint hit.
2000  */
2001 int uprobe_pre_sstep_notifier(struct pt_regs *regs)
2002 {
2003         if (!current->mm)
2004                 return 0;
2005
2006         if (!test_bit(MMF_HAS_UPROBES, &current->mm->flags) &&
2007             (!current->utask || !current->utask->return_instances))
2008                 return 0;
2009
2010         set_thread_flag(TIF_UPROBE);
2011         return 1;
2012 }
2013
2014 /*
2015  * uprobe_post_sstep_notifier gets called in interrupt context as part of notifier
2016  * mechanism. Set TIF_UPROBE flag and indicate completion of singlestep.
2017  */
2018 int uprobe_post_sstep_notifier(struct pt_regs *regs)
2019 {
2020         struct uprobe_task *utask = current->utask;
2021
2022         if (!current->mm || !utask || !utask->active_uprobe)
2023                 /* task is currently not uprobed */
2024                 return 0;
2025
2026         utask->state = UTASK_SSTEP_ACK;
2027         set_thread_flag(TIF_UPROBE);
2028         return 1;
2029 }
2030
2031 static struct notifier_block uprobe_exception_nb = {
2032         .notifier_call          = arch_uprobe_exception_notify,
2033         .priority               = INT_MAX-1,    /* notified after kprobes, kgdb */
2034 };
2035
2036 static int __init init_uprobes(void)
2037 {
2038         int i;
2039
2040         for (i = 0; i < UPROBES_HASH_SZ; i++)
2041                 mutex_init(&uprobes_mmap_mutex[i]);
2042
2043         if (percpu_init_rwsem(&dup_mmap_sem))
2044                 return -ENOMEM;
2045
2046         return register_die_notifier(&uprobe_exception_nb);
2047 }
2048 __initcall(init_uprobes);