uprobes: Fix utask->depth accounting in handle_trampoline()
[linux-3.10.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 <pzijlstr@redhat.com>
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/export.h>
31 #include <linux/rmap.h>         /* anon_vma_prepare */
32 #include <linux/mmu_notifier.h> /* set_pte_at_notify */
33 #include <linux/swap.h>         /* try_to_free_swap */
34 #include <linux/ptrace.h>       /* user_enable_single_step */
35 #include <linux/kdebug.h>       /* notifier mechanism */
36 #include "../../mm/internal.h"  /* munlock_vma_page */
37 #include <linux/percpu-rwsem.h>
38
39 #include <linux/uprobes.h>
40
41 #define UINSNS_PER_PAGE                 (PAGE_SIZE/UPROBE_XOL_SLOT_BYTES)
42 #define MAX_UPROBE_XOL_SLOTS            UINSNS_PER_PAGE
43
44 static struct rb_root uprobes_tree = RB_ROOT;
45 /*
46  * allows us to skip the uprobe_mmap if there are no uprobe events active
47  * at this time.  Probably a fine grained per inode count is better?
48  */
49 #define no_uprobe_events()      RB_EMPTY_ROOT(&uprobes_tree)
50
51 static DEFINE_SPINLOCK(uprobes_treelock);       /* serialize rbtree access */
52
53 #define UPROBES_HASH_SZ 13
54 /* serialize uprobe->pending_list */
55 static struct mutex uprobes_mmap_mutex[UPROBES_HASH_SZ];
56 #define uprobes_mmap_hash(v)    (&uprobes_mmap_mutex[((unsigned long)(v)) % UPROBES_HASH_SZ])
57
58 static struct percpu_rw_semaphore dup_mmap_sem;
59
60 /* Have a copy of original instruction */
61 #define UPROBE_COPY_INSN        0
62 /* Can skip singlestep */
63 #define UPROBE_SKIP_SSTEP       1
64
65 struct uprobe {
66         struct rb_node          rb_node;        /* node in the rb tree */
67         atomic_t                ref;
68         struct rw_semaphore     register_rwsem;
69         struct rw_semaphore     consumer_rwsem;
70         struct list_head        pending_list;
71         struct uprobe_consumer  *consumers;
72         struct inode            *inode;         /* Also hold a ref to inode */
73         loff_t                  offset;
74         unsigned long           flags;
75         struct arch_uprobe      arch;
76 };
77
78 struct return_instance {
79         struct uprobe           *uprobe;
80         unsigned long           func;
81         unsigned long           orig_ret_vaddr; /* original return address */
82         bool                    chained;        /* true, if instance is nested */
83
84         struct return_instance  *next;          /* keep as stack */
85 };
86
87 /*
88  * valid_vma: Verify if the specified vma is an executable vma
89  * Relax restrictions while unregistering: vm_flags might have
90  * changed after breakpoint was inserted.
91  *      - is_register: indicates if we are in register context.
92  *      - Return 1 if the specified virtual address is in an
93  *        executable vma.
94  */
95 static bool valid_vma(struct vm_area_struct *vma, bool is_register)
96 {
97         vm_flags_t flags = VM_HUGETLB | VM_MAYEXEC | VM_SHARED;
98
99         if (is_register)
100                 flags |= VM_WRITE;
101
102         return vma->vm_file && (vma->vm_flags & flags) == VM_MAYEXEC;
103 }
104
105 static unsigned long offset_to_vaddr(struct vm_area_struct *vma, loff_t offset)
106 {
107         return vma->vm_start + offset - ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
108 }
109
110 static loff_t vaddr_to_offset(struct vm_area_struct *vma, unsigned long vaddr)
111 {
112         return ((loff_t)vma->vm_pgoff << PAGE_SHIFT) + (vaddr - vma->vm_start);
113 }
114
115 /**
116  * __replace_page - replace page in vma by new page.
117  * based on replace_page in mm/ksm.c
118  *
119  * @vma:      vma that holds the pte pointing to page
120  * @addr:     address the old @page is mapped at
121  * @page:     the cowed page we are replacing by kpage
122  * @kpage:    the modified page we replace page by
123  *
124  * Returns 0 on success, -EFAULT on failure.
125  */
126 static int __replace_page(struct vm_area_struct *vma, unsigned long addr,
127                                 struct page *page, struct page *kpage)
128 {
129         struct mm_struct *mm = vma->vm_mm;
130         spinlock_t *ptl;
131         pte_t *ptep;
132         int err;
133         /* For mmu_notifiers */
134         const unsigned long mmun_start = addr;
135         const unsigned long mmun_end   = addr + PAGE_SIZE;
136
137         /* For try_to_free_swap() and munlock_vma_page() below */
138         lock_page(page);
139
140         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
141         err = -EAGAIN;
142         ptep = page_check_address(page, mm, addr, &ptl, 0);
143         if (!ptep)
144                 goto unlock;
145
146         get_page(kpage);
147         page_add_new_anon_rmap(kpage, vma, addr);
148
149         if (!PageAnon(page)) {
150                 dec_mm_counter(mm, MM_FILEPAGES);
151                 inc_mm_counter(mm, MM_ANONPAGES);
152         }
153
154         flush_cache_page(vma, addr, pte_pfn(*ptep));
155         ptep_clear_flush(vma, addr, ptep);
156         set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
157
158         page_remove_rmap(page);
159         if (!page_mapped(page))
160                 try_to_free_swap(page);
161         pte_unmap_unlock(ptep, ptl);
162
163         if (vma->vm_flags & VM_LOCKED)
164                 munlock_vma_page(page);
165         put_page(page);
166
167         err = 0;
168  unlock:
169         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
170         unlock_page(page);
171         return err;
172 }
173
174 /**
175  * is_swbp_insn - check if instruction is breakpoint instruction.
176  * @insn: instruction to be checked.
177  * Default implementation of is_swbp_insn
178  * Returns true if @insn is a breakpoint instruction.
179  */
180 bool __weak is_swbp_insn(uprobe_opcode_t *insn)
181 {
182         return *insn == UPROBE_SWBP_INSN;
183 }
184
185 /**
186  * is_trap_insn - check if instruction is breakpoint instruction.
187  * @insn: instruction to be checked.
188  * Default implementation of is_trap_insn
189  * Returns true if @insn is a breakpoint instruction.
190  *
191  * This function is needed for the case where an architecture has multiple
192  * trap instructions (like powerpc).
193  */
194 bool __weak is_trap_insn(uprobe_opcode_t *insn)
195 {
196         return is_swbp_insn(insn);
197 }
198
199 static void copy_from_page(struct page *page, unsigned long vaddr, void *dst, int len)
200 {
201         void *kaddr = kmap_atomic(page);
202         memcpy(dst, kaddr + (vaddr & ~PAGE_MASK), len);
203         kunmap_atomic(kaddr);
204 }
205
206 static void copy_to_page(struct page *page, unsigned long vaddr, const void *src, int len)
207 {
208         void *kaddr = kmap_atomic(page);
209         memcpy(kaddr + (vaddr & ~PAGE_MASK), src, len);
210         kunmap_atomic(kaddr);
211 }
212
213 static int verify_opcode(struct page *page, unsigned long vaddr, uprobe_opcode_t *new_opcode)
214 {
215         uprobe_opcode_t old_opcode;
216         bool is_swbp;
217
218         /*
219          * Note: We only check if the old_opcode is UPROBE_SWBP_INSN here.
220          * We do not check if it is any other 'trap variant' which could
221          * be conditional trap instruction such as the one powerpc supports.
222          *
223          * The logic is that we do not care if the underlying instruction
224          * is a trap variant; uprobes always wins over any other (gdb)
225          * breakpoint.
226          */
227         copy_from_page(page, vaddr, &old_opcode, UPROBE_SWBP_INSN_SIZE);
228         is_swbp = is_swbp_insn(&old_opcode);
229
230         if (is_swbp_insn(new_opcode)) {
231                 if (is_swbp)            /* register: already installed? */
232                         return 0;
233         } else {
234                 if (!is_swbp)           /* unregister: was it changed by us? */
235                         return 0;
236         }
237
238         return 1;
239 }
240
241 /*
242  * NOTE:
243  * Expect the breakpoint instruction to be the smallest size instruction for
244  * the architecture. If an arch has variable length instruction and the
245  * breakpoint instruction is not of the smallest length instruction
246  * supported by that architecture then we need to modify is_trap_at_addr and
247  * write_opcode accordingly. This would never be a problem for archs that
248  * have fixed length instructions.
249  */
250
251 /*
252  * write_opcode - write the opcode at a given virtual address.
253  * @mm: the probed process address space.
254  * @vaddr: the virtual address to store the opcode.
255  * @opcode: opcode to be written at @vaddr.
256  *
257  * Called with mm->mmap_sem held (for read and with a reference to
258  * mm).
259  *
260  * For mm @mm, write the opcode at @vaddr.
261  * Return 0 (success) or a negative errno.
262  */
263 static int write_opcode(struct mm_struct *mm, unsigned long vaddr,
264                         uprobe_opcode_t opcode)
265 {
266         struct page *old_page, *new_page;
267         struct vm_area_struct *vma;
268         int ret;
269
270 retry:
271         /* Read the page with vaddr into memory */
272         ret = get_user_pages(NULL, mm, vaddr, 1, 0, 1, &old_page, &vma);
273         if (ret <= 0)
274                 return ret;
275
276         ret = verify_opcode(old_page, vaddr, &opcode);
277         if (ret <= 0)
278                 goto put_old;
279
280         ret = -ENOMEM;
281         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vaddr);
282         if (!new_page)
283                 goto put_old;
284
285         __SetPageUptodate(new_page);
286
287         copy_highpage(new_page, old_page);
288         copy_to_page(new_page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
289
290         ret = anon_vma_prepare(vma);
291         if (ret)
292                 goto put_new;
293
294         ret = __replace_page(vma, vaddr, old_page, new_page);
295
296 put_new:
297         page_cache_release(new_page);
298 put_old:
299         put_page(old_page);
300
301         if (unlikely(ret == -EAGAIN))
302                 goto retry;
303         return ret;
304 }
305
306 /**
307  * set_swbp - store breakpoint at a given address.
308  * @auprobe: arch specific probepoint information.
309  * @mm: the probed process address space.
310  * @vaddr: the virtual address to insert the opcode.
311  *
312  * For mm @mm, store the breakpoint instruction at @vaddr.
313  * Return 0 (success) or a negative errno.
314  */
315 int __weak set_swbp(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
316 {
317         return write_opcode(mm, vaddr, UPROBE_SWBP_INSN);
318 }
319
320 /**
321  * set_orig_insn - Restore the original instruction.
322  * @mm: the probed process address space.
323  * @auprobe: arch specific probepoint information.
324  * @vaddr: the virtual address to insert the opcode.
325  *
326  * For mm @mm, restore the original opcode (opcode) at @vaddr.
327  * Return 0 (success) or a negative errno.
328  */
329 int __weak
330 set_orig_insn(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
331 {
332         return write_opcode(mm, vaddr, *(uprobe_opcode_t *)auprobe->insn);
333 }
334
335 static int match_uprobe(struct uprobe *l, struct uprobe *r)
336 {
337         if (l->inode < r->inode)
338                 return -1;
339
340         if (l->inode > r->inode)
341                 return 1;
342
343         if (l->offset < r->offset)
344                 return -1;
345
346         if (l->offset > r->offset)
347                 return 1;
348
349         return 0;
350 }
351
352 static struct uprobe *__find_uprobe(struct inode *inode, loff_t offset)
353 {
354         struct uprobe u = { .inode = inode, .offset = offset };
355         struct rb_node *n = uprobes_tree.rb_node;
356         struct uprobe *uprobe;
357         int match;
358
359         while (n) {
360                 uprobe = rb_entry(n, struct uprobe, rb_node);
361                 match = match_uprobe(&u, uprobe);
362                 if (!match) {
363                         atomic_inc(&uprobe->ref);
364                         return uprobe;
365                 }
366
367                 if (match < 0)
368                         n = n->rb_left;
369                 else
370                         n = n->rb_right;
371         }
372         return NULL;
373 }
374
375 /*
376  * Find a uprobe corresponding to a given inode:offset
377  * Acquires uprobes_treelock
378  */
379 static struct uprobe *find_uprobe(struct inode *inode, loff_t offset)
380 {
381         struct uprobe *uprobe;
382
383         spin_lock(&uprobes_treelock);
384         uprobe = __find_uprobe(inode, offset);
385         spin_unlock(&uprobes_treelock);
386
387         return uprobe;
388 }
389
390 static struct uprobe *__insert_uprobe(struct uprobe *uprobe)
391 {
392         struct rb_node **p = &uprobes_tree.rb_node;
393         struct rb_node *parent = NULL;
394         struct uprobe *u;
395         int match;
396
397         while (*p) {
398                 parent = *p;
399                 u = rb_entry(parent, struct uprobe, rb_node);
400                 match = match_uprobe(uprobe, u);
401                 if (!match) {
402                         atomic_inc(&u->ref);
403                         return u;
404                 }
405
406                 if (match < 0)
407                         p = &parent->rb_left;
408                 else
409                         p = &parent->rb_right;
410
411         }
412
413         u = NULL;
414         rb_link_node(&uprobe->rb_node, parent, p);
415         rb_insert_color(&uprobe->rb_node, &uprobes_tree);
416         /* get access + creation ref */
417         atomic_set(&uprobe->ref, 2);
418
419         return u;
420 }
421
422 /*
423  * Acquire uprobes_treelock.
424  * Matching uprobe already exists in rbtree;
425  *      increment (access refcount) and return the matching uprobe.
426  *
427  * No matching uprobe; insert the uprobe in rb_tree;
428  *      get a double refcount (access + creation) and return NULL.
429  */
430 static struct uprobe *insert_uprobe(struct uprobe *uprobe)
431 {
432         struct uprobe *u;
433
434         spin_lock(&uprobes_treelock);
435         u = __insert_uprobe(uprobe);
436         spin_unlock(&uprobes_treelock);
437
438         return u;
439 }
440
441 static void put_uprobe(struct uprobe *uprobe)
442 {
443         if (atomic_dec_and_test(&uprobe->ref))
444                 kfree(uprobe);
445 }
446
447 static struct uprobe *alloc_uprobe(struct inode *inode, loff_t offset)
448 {
449         struct uprobe *uprobe, *cur_uprobe;
450
451         uprobe = kzalloc(sizeof(struct uprobe), GFP_KERNEL);
452         if (!uprobe)
453                 return NULL;
454
455         uprobe->inode = igrab(inode);
456         uprobe->offset = offset;
457         init_rwsem(&uprobe->register_rwsem);
458         init_rwsem(&uprobe->consumer_rwsem);
459         /* For now assume that the instruction need not be single-stepped */
460         __set_bit(UPROBE_SKIP_SSTEP, &uprobe->flags);
461
462         /* add to uprobes_tree, sorted on inode:offset */
463         cur_uprobe = insert_uprobe(uprobe);
464
465         /* a uprobe exists for this inode:offset combination */
466         if (cur_uprobe) {
467                 kfree(uprobe);
468                 uprobe = cur_uprobe;
469                 iput(inode);
470         }
471
472         return uprobe;
473 }
474
475 static void consumer_add(struct uprobe *uprobe, struct uprobe_consumer *uc)
476 {
477         down_write(&uprobe->consumer_rwsem);
478         uc->next = uprobe->consumers;
479         uprobe->consumers = uc;
480         up_write(&uprobe->consumer_rwsem);
481 }
482
483 /*
484  * For uprobe @uprobe, delete the consumer @uc.
485  * Return true if the @uc is deleted successfully
486  * or return false.
487  */
488 static bool consumer_del(struct uprobe *uprobe, struct uprobe_consumer *uc)
489 {
490         struct uprobe_consumer **con;
491         bool ret = false;
492
493         down_write(&uprobe->consumer_rwsem);
494         for (con = &uprobe->consumers; *con; con = &(*con)->next) {
495                 if (*con == uc) {
496                         *con = uc->next;
497                         ret = true;
498                         break;
499                 }
500         }
501         up_write(&uprobe->consumer_rwsem);
502
503         return ret;
504 }
505
506 static int
507 __copy_insn(struct address_space *mapping, struct file *filp, char *insn,
508                         unsigned long nbytes, loff_t offset)
509 {
510         struct page *page;
511
512         if (!mapping->a_ops->readpage)
513                 return -EIO;
514         /*
515          * Ensure that the page that has the original instruction is
516          * populated and in page-cache.
517          */
518         page = read_mapping_page(mapping, offset >> PAGE_CACHE_SHIFT, filp);
519         if (IS_ERR(page))
520                 return PTR_ERR(page);
521
522         copy_from_page(page, offset, insn, nbytes);
523         page_cache_release(page);
524
525         return 0;
526 }
527
528 static int copy_insn(struct uprobe *uprobe, struct file *filp)
529 {
530         struct address_space *mapping;
531         unsigned long nbytes;
532         int bytes;
533
534         nbytes = PAGE_SIZE - (uprobe->offset & ~PAGE_MASK);
535         mapping = uprobe->inode->i_mapping;
536
537         /* Instruction at end of binary; copy only available bytes */
538         if (uprobe->offset + MAX_UINSN_BYTES > uprobe->inode->i_size)
539                 bytes = uprobe->inode->i_size - uprobe->offset;
540         else
541                 bytes = MAX_UINSN_BYTES;
542
543         /* Instruction at the page-boundary; copy bytes in second page */
544         if (nbytes < bytes) {
545                 int err = __copy_insn(mapping, filp, uprobe->arch.insn + nbytes,
546                                 bytes - nbytes, uprobe->offset + nbytes);
547                 if (err)
548                         return err;
549                 bytes = nbytes;
550         }
551         return __copy_insn(mapping, filp, uprobe->arch.insn, bytes, uprobe->offset);
552 }
553
554 static int prepare_uprobe(struct uprobe *uprobe, struct file *file,
555                                 struct mm_struct *mm, unsigned long vaddr)
556 {
557         int ret = 0;
558
559         if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
560                 return ret;
561
562         /* TODO: move this into _register, until then we abuse this sem. */
563         down_write(&uprobe->consumer_rwsem);
564         if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
565                 goto out;
566
567         ret = copy_insn(uprobe, file);
568         if (ret)
569                 goto out;
570
571         ret = -ENOTSUPP;
572         if (is_trap_insn((uprobe_opcode_t *)uprobe->arch.insn))
573                 goto out;
574
575         ret = arch_uprobe_analyze_insn(&uprobe->arch, mm, vaddr);
576         if (ret)
577                 goto out;
578
579         /* write_opcode() assumes we don't cross page boundary */
580         BUG_ON((uprobe->offset & ~PAGE_MASK) +
581                         UPROBE_SWBP_INSN_SIZE > PAGE_SIZE);
582
583         smp_wmb(); /* pairs with rmb() in find_active_uprobe() */
584         set_bit(UPROBE_COPY_INSN, &uprobe->flags);
585
586  out:
587         up_write(&uprobe->consumer_rwsem);
588
589         return ret;
590 }
591
592 static inline bool consumer_filter(struct uprobe_consumer *uc,
593                                    enum uprobe_filter_ctx ctx, struct mm_struct *mm)
594 {
595         return !uc->filter || uc->filter(uc, ctx, mm);
596 }
597
598 static bool filter_chain(struct uprobe *uprobe,
599                          enum uprobe_filter_ctx ctx, struct mm_struct *mm)
600 {
601         struct uprobe_consumer *uc;
602         bool ret = false;
603
604         down_read(&uprobe->consumer_rwsem);
605         for (uc = uprobe->consumers; uc; uc = uc->next) {
606                 ret = consumer_filter(uc, ctx, mm);
607                 if (ret)
608                         break;
609         }
610         up_read(&uprobe->consumer_rwsem);
611
612         return ret;
613 }
614
615 static int
616 install_breakpoint(struct uprobe *uprobe, struct mm_struct *mm,
617                         struct vm_area_struct *vma, unsigned long vaddr)
618 {
619         bool first_uprobe;
620         int ret;
621
622         ret = prepare_uprobe(uprobe, vma->vm_file, mm, vaddr);
623         if (ret)
624                 return ret;
625
626         /*
627          * set MMF_HAS_UPROBES in advance for uprobe_pre_sstep_notifier(),
628          * the task can hit this breakpoint right after __replace_page().
629          */
630         first_uprobe = !test_bit(MMF_HAS_UPROBES, &mm->flags);
631         if (first_uprobe)
632                 set_bit(MMF_HAS_UPROBES, &mm->flags);
633
634         ret = set_swbp(&uprobe->arch, mm, vaddr);
635         if (!ret)
636                 clear_bit(MMF_RECALC_UPROBES, &mm->flags);
637         else if (first_uprobe)
638                 clear_bit(MMF_HAS_UPROBES, &mm->flags);
639
640         return ret;
641 }
642
643 static int
644 remove_breakpoint(struct uprobe *uprobe, struct mm_struct *mm, unsigned long vaddr)
645 {
646         set_bit(MMF_RECALC_UPROBES, &mm->flags);
647         return set_orig_insn(&uprobe->arch, mm, vaddr);
648 }
649
650 static inline bool uprobe_is_active(struct uprobe *uprobe)
651 {
652         return !RB_EMPTY_NODE(&uprobe->rb_node);
653 }
654 /*
655  * There could be threads that have already hit the breakpoint. They
656  * will recheck the current insn and restart if find_uprobe() fails.
657  * See find_active_uprobe().
658  */
659 static void delete_uprobe(struct uprobe *uprobe)
660 {
661         if (WARN_ON(!uprobe_is_active(uprobe)))
662                 return;
663
664         spin_lock(&uprobes_treelock);
665         rb_erase(&uprobe->rb_node, &uprobes_tree);
666         spin_unlock(&uprobes_treelock);
667         RB_CLEAR_NODE(&uprobe->rb_node); /* for uprobe_is_active() */
668         iput(uprobe->inode);
669         put_uprobe(uprobe);
670 }
671
672 struct map_info {
673         struct map_info *next;
674         struct mm_struct *mm;
675         unsigned long vaddr;
676 };
677
678 static inline struct map_info *free_map_info(struct map_info *info)
679 {
680         struct map_info *next = info->next;
681         kfree(info);
682         return next;
683 }
684
685 static struct map_info *
686 build_map_info(struct address_space *mapping, loff_t offset, bool is_register)
687 {
688         unsigned long pgoff = offset >> PAGE_SHIFT;
689         struct vm_area_struct *vma;
690         struct map_info *curr = NULL;
691         struct map_info *prev = NULL;
692         struct map_info *info;
693         int more = 0;
694
695  again:
696         mutex_lock(&mapping->i_mmap_mutex);
697         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
698                 if (!valid_vma(vma, is_register))
699                         continue;
700
701                 if (!prev && !more) {
702                         /*
703                          * Needs GFP_NOWAIT to avoid i_mmap_mutex recursion through
704                          * reclaim. This is optimistic, no harm done if it fails.
705                          */
706                         prev = kmalloc(sizeof(struct map_info),
707                                         GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
708                         if (prev)
709                                 prev->next = NULL;
710                 }
711                 if (!prev) {
712                         more++;
713                         continue;
714                 }
715
716                 if (!atomic_inc_not_zero(&vma->vm_mm->mm_users))
717                         continue;
718
719                 info = prev;
720                 prev = prev->next;
721                 info->next = curr;
722                 curr = info;
723
724                 info->mm = vma->vm_mm;
725                 info->vaddr = offset_to_vaddr(vma, offset);
726         }
727         mutex_unlock(&mapping->i_mmap_mutex);
728
729         if (!more)
730                 goto out;
731
732         prev = curr;
733         while (curr) {
734                 mmput(curr->mm);
735                 curr = curr->next;
736         }
737
738         do {
739                 info = kmalloc(sizeof(struct map_info), GFP_KERNEL);
740                 if (!info) {
741                         curr = ERR_PTR(-ENOMEM);
742                         goto out;
743                 }
744                 info->next = prev;
745                 prev = info;
746         } while (--more);
747
748         goto again;
749  out:
750         while (prev)
751                 prev = free_map_info(prev);
752         return curr;
753 }
754
755 static int
756 register_for_each_vma(struct uprobe *uprobe, struct uprobe_consumer *new)
757 {
758         bool is_register = !!new;
759         struct map_info *info;
760         int err = 0;
761
762         percpu_down_write(&dup_mmap_sem);
763         info = build_map_info(uprobe->inode->i_mapping,
764                                         uprobe->offset, is_register);
765         if (IS_ERR(info)) {
766                 err = PTR_ERR(info);
767                 goto out;
768         }
769
770         while (info) {
771                 struct mm_struct *mm = info->mm;
772                 struct vm_area_struct *vma;
773
774                 if (err && is_register)
775                         goto free;
776
777                 down_write(&mm->mmap_sem);
778                 vma = find_vma(mm, info->vaddr);
779                 if (!vma || !valid_vma(vma, is_register) ||
780                     file_inode(vma->vm_file) != uprobe->inode)
781                         goto unlock;
782
783                 if (vma->vm_start > info->vaddr ||
784                     vaddr_to_offset(vma, info->vaddr) != uprobe->offset)
785                         goto unlock;
786
787                 if (is_register) {
788                         /* consult only the "caller", new consumer. */
789                         if (consumer_filter(new,
790                                         UPROBE_FILTER_REGISTER, mm))
791                                 err = install_breakpoint(uprobe, mm, vma, info->vaddr);
792                 } else if (test_bit(MMF_HAS_UPROBES, &mm->flags)) {
793                         if (!filter_chain(uprobe,
794                                         UPROBE_FILTER_UNREGISTER, mm))
795                                 err |= remove_breakpoint(uprobe, mm, info->vaddr);
796                 }
797
798  unlock:
799                 up_write(&mm->mmap_sem);
800  free:
801                 mmput(mm);
802                 info = free_map_info(info);
803         }
804  out:
805         percpu_up_write(&dup_mmap_sem);
806         return err;
807 }
808
809 static int __uprobe_register(struct uprobe *uprobe, struct uprobe_consumer *uc)
810 {
811         consumer_add(uprobe, uc);
812         return register_for_each_vma(uprobe, uc);
813 }
814
815 static void __uprobe_unregister(struct uprobe *uprobe, struct uprobe_consumer *uc)
816 {
817         int err;
818
819         if (!consumer_del(uprobe, uc))  /* WARN? */
820                 return;
821
822         err = register_for_each_vma(uprobe, NULL);
823         /* TODO : cant unregister? schedule a worker thread */
824         if (!uprobe->consumers && !err)
825                 delete_uprobe(uprobe);
826 }
827
828 /*
829  * uprobe_register - register a probe
830  * @inode: the file in which the probe has to be placed.
831  * @offset: offset from the start of the file.
832  * @uc: information on howto handle the probe..
833  *
834  * Apart from the access refcount, uprobe_register() takes a creation
835  * refcount (thro alloc_uprobe) if and only if this @uprobe is getting
836  * inserted into the rbtree (i.e first consumer for a @inode:@offset
837  * tuple).  Creation refcount stops uprobe_unregister from freeing the
838  * @uprobe even before the register operation is complete. Creation
839  * refcount is released when the last @uc for the @uprobe
840  * unregisters.
841  *
842  * Return errno if it cannot successully install probes
843  * else return 0 (success)
844  */
845 int uprobe_register(struct inode *inode, loff_t offset, struct uprobe_consumer *uc)
846 {
847         struct uprobe *uprobe;
848         int ret;
849
850         /* Uprobe must have at least one set consumer */
851         if (!uc->handler && !uc->ret_handler)
852                 return -EINVAL;
853
854         /* Racy, just to catch the obvious mistakes */
855         if (offset > i_size_read(inode))
856                 return -EINVAL;
857
858  retry:
859         uprobe = alloc_uprobe(inode, offset);
860         if (!uprobe)
861                 return -ENOMEM;
862         /*
863          * We can race with uprobe_unregister()->delete_uprobe().
864          * Check uprobe_is_active() and retry if it is false.
865          */
866         down_write(&uprobe->register_rwsem);
867         ret = -EAGAIN;
868         if (likely(uprobe_is_active(uprobe))) {
869                 ret = __uprobe_register(uprobe, uc);
870                 if (ret)
871                         __uprobe_unregister(uprobe, uc);
872         }
873         up_write(&uprobe->register_rwsem);
874         put_uprobe(uprobe);
875
876         if (unlikely(ret == -EAGAIN))
877                 goto retry;
878         return ret;
879 }
880 EXPORT_SYMBOL_GPL(uprobe_register);
881
882 /*
883  * uprobe_apply - unregister a already registered probe.
884  * @inode: the file in which the probe has to be removed.
885  * @offset: offset from the start of the file.
886  * @uc: consumer which wants to add more or remove some breakpoints
887  * @add: add or remove the breakpoints
888  */
889 int uprobe_apply(struct inode *inode, loff_t offset,
890                         struct uprobe_consumer *uc, bool add)
891 {
892         struct uprobe *uprobe;
893         struct uprobe_consumer *con;
894         int ret = -ENOENT;
895
896         uprobe = find_uprobe(inode, offset);
897         if (!uprobe)
898                 return ret;
899
900         down_write(&uprobe->register_rwsem);
901         for (con = uprobe->consumers; con && con != uc ; con = con->next)
902                 ;
903         if (con)
904                 ret = register_for_each_vma(uprobe, add ? uc : NULL);
905         up_write(&uprobe->register_rwsem);
906         put_uprobe(uprobe);
907
908         return ret;
909 }
910
911 /*
912  * uprobe_unregister - unregister a already registered probe.
913  * @inode: the file in which the probe has to be removed.
914  * @offset: offset from the start of the file.
915  * @uc: identify which probe if multiple probes are colocated.
916  */
917 void uprobe_unregister(struct inode *inode, loff_t offset, struct uprobe_consumer *uc)
918 {
919         struct uprobe *uprobe;
920
921         uprobe = find_uprobe(inode, offset);
922         if (!uprobe)
923                 return;
924
925         down_write(&uprobe->register_rwsem);
926         __uprobe_unregister(uprobe, uc);
927         up_write(&uprobe->register_rwsem);
928         put_uprobe(uprobe);
929 }
930 EXPORT_SYMBOL_GPL(uprobe_unregister);
931
932 static int unapply_uprobe(struct uprobe *uprobe, struct mm_struct *mm)
933 {
934         struct vm_area_struct *vma;
935         int err = 0;
936
937         down_read(&mm->mmap_sem);
938         for (vma = mm->mmap; vma; vma = vma->vm_next) {
939                 unsigned long vaddr;
940                 loff_t offset;
941
942                 if (!valid_vma(vma, false) ||
943                     file_inode(vma->vm_file) != uprobe->inode)
944                         continue;
945
946                 offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
947                 if (uprobe->offset <  offset ||
948                     uprobe->offset >= offset + vma->vm_end - vma->vm_start)
949                         continue;
950
951                 vaddr = offset_to_vaddr(vma, uprobe->offset);
952                 err |= remove_breakpoint(uprobe, mm, vaddr);
953         }
954         up_read(&mm->mmap_sem);
955
956         return err;
957 }
958
959 static struct rb_node *
960 find_node_in_range(struct inode *inode, loff_t min, loff_t max)
961 {
962         struct rb_node *n = uprobes_tree.rb_node;
963
964         while (n) {
965                 struct uprobe *u = rb_entry(n, struct uprobe, rb_node);
966
967                 if (inode < u->inode) {
968                         n = n->rb_left;
969                 } else if (inode > u->inode) {
970                         n = n->rb_right;
971                 } else {
972                         if (max < u->offset)
973                                 n = n->rb_left;
974                         else if (min > u->offset)
975                                 n = n->rb_right;
976                         else
977                                 break;
978                 }
979         }
980
981         return n;
982 }
983
984 /*
985  * For a given range in vma, build a list of probes that need to be inserted.
986  */
987 static void build_probe_list(struct inode *inode,
988                                 struct vm_area_struct *vma,
989                                 unsigned long start, unsigned long end,
990                                 struct list_head *head)
991 {
992         loff_t min, max;
993         struct rb_node *n, *t;
994         struct uprobe *u;
995
996         INIT_LIST_HEAD(head);
997         min = vaddr_to_offset(vma, start);
998         max = min + (end - start) - 1;
999
1000         spin_lock(&uprobes_treelock);
1001         n = find_node_in_range(inode, min, max);
1002         if (n) {
1003                 for (t = n; t; t = rb_prev(t)) {
1004                         u = rb_entry(t, struct uprobe, rb_node);
1005                         if (u->inode != inode || u->offset < min)
1006                                 break;
1007                         list_add(&u->pending_list, head);
1008                         atomic_inc(&u->ref);
1009                 }
1010                 for (t = n; (t = rb_next(t)); ) {
1011                         u = rb_entry(t, struct uprobe, rb_node);
1012                         if (u->inode != inode || u->offset > max)
1013                                 break;
1014                         list_add(&u->pending_list, head);
1015                         atomic_inc(&u->ref);
1016                 }
1017         }
1018         spin_unlock(&uprobes_treelock);
1019 }
1020
1021 /*
1022  * Called from mmap_region/vma_adjust with mm->mmap_sem acquired.
1023  *
1024  * Currently we ignore all errors and always return 0, the callers
1025  * can't handle the failure anyway.
1026  */
1027 int uprobe_mmap(struct vm_area_struct *vma)
1028 {
1029         struct list_head tmp_list;
1030         struct uprobe *uprobe, *u;
1031         struct inode *inode;
1032
1033         if (no_uprobe_events() || !valid_vma(vma, true))
1034                 return 0;
1035
1036         inode = file_inode(vma->vm_file);
1037         if (!inode)
1038                 return 0;
1039
1040         mutex_lock(uprobes_mmap_hash(inode));
1041         build_probe_list(inode, vma, vma->vm_start, vma->vm_end, &tmp_list);
1042         /*
1043          * We can race with uprobe_unregister(), this uprobe can be already
1044          * removed. But in this case filter_chain() must return false, all
1045          * consumers have gone away.
1046          */
1047         list_for_each_entry_safe(uprobe, u, &tmp_list, pending_list) {
1048                 if (!fatal_signal_pending(current) &&
1049                     filter_chain(uprobe, UPROBE_FILTER_MMAP, vma->vm_mm)) {
1050                         unsigned long vaddr = offset_to_vaddr(vma, uprobe->offset);
1051                         install_breakpoint(uprobe, vma->vm_mm, vma, vaddr);
1052                 }
1053                 put_uprobe(uprobe);
1054         }
1055         mutex_unlock(uprobes_mmap_hash(inode));
1056
1057         return 0;
1058 }
1059
1060 static bool
1061 vma_has_uprobes(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1062 {
1063         loff_t min, max;
1064         struct inode *inode;
1065         struct rb_node *n;
1066
1067         inode = file_inode(vma->vm_file);
1068
1069         min = vaddr_to_offset(vma, start);
1070         max = min + (end - start) - 1;
1071
1072         spin_lock(&uprobes_treelock);
1073         n = find_node_in_range(inode, min, max);
1074         spin_unlock(&uprobes_treelock);
1075
1076         return !!n;
1077 }
1078
1079 /*
1080  * Called in context of a munmap of a vma.
1081  */
1082 void uprobe_munmap(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1083 {
1084         if (no_uprobe_events() || !valid_vma(vma, false))
1085                 return;
1086
1087         if (!atomic_read(&vma->vm_mm->mm_users)) /* called by mmput() ? */
1088                 return;
1089
1090         if (!test_bit(MMF_HAS_UPROBES, &vma->vm_mm->flags) ||
1091              test_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags))
1092                 return;
1093
1094         if (vma_has_uprobes(vma, start, end))
1095                 set_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags);
1096 }
1097
1098 /* Slot allocation for XOL */
1099 static int xol_add_vma(struct xol_area *area)
1100 {
1101         struct mm_struct *mm = current->mm;
1102         int ret = -EALREADY;
1103
1104         down_write(&mm->mmap_sem);
1105         if (mm->uprobes_state.xol_area)
1106                 goto fail;
1107
1108         ret = -ENOMEM;
1109         /* Try to map as high as possible, this is only a hint. */
1110         area->vaddr = get_unmapped_area(NULL, TASK_SIZE - PAGE_SIZE, PAGE_SIZE, 0, 0);
1111         if (area->vaddr & ~PAGE_MASK) {
1112                 ret = area->vaddr;
1113                 goto fail;
1114         }
1115
1116         ret = install_special_mapping(mm, area->vaddr, PAGE_SIZE,
1117                                 VM_EXEC|VM_MAYEXEC|VM_DONTCOPY|VM_IO, &area->page);
1118         if (ret)
1119                 goto fail;
1120
1121         smp_wmb();      /* pairs with get_xol_area() */
1122         mm->uprobes_state.xol_area = area;
1123         ret = 0;
1124  fail:
1125         up_write(&mm->mmap_sem);
1126
1127         return ret;
1128 }
1129
1130 /*
1131  * get_xol_area - Allocate process's xol_area if necessary.
1132  * This area will be used for storing instructions for execution out of line.
1133  *
1134  * Returns the allocated area or NULL.
1135  */
1136 static struct xol_area *get_xol_area(void)
1137 {
1138         struct mm_struct *mm = current->mm;
1139         struct xol_area *area;
1140         uprobe_opcode_t insn = UPROBE_SWBP_INSN;
1141
1142         area = mm->uprobes_state.xol_area;
1143         if (area)
1144                 goto ret;
1145
1146         area = kzalloc(sizeof(*area), GFP_KERNEL);
1147         if (unlikely(!area))
1148                 goto out;
1149
1150         area->bitmap = kzalloc(BITS_TO_LONGS(UINSNS_PER_PAGE) * sizeof(long), GFP_KERNEL);
1151         if (!area->bitmap)
1152                 goto free_area;
1153
1154         area->page = alloc_page(GFP_HIGHUSER);
1155         if (!area->page)
1156                 goto free_bitmap;
1157
1158         /* allocate first slot of task's xol_area for the return probes */
1159         set_bit(0, area->bitmap);
1160         copy_to_page(area->page, 0, &insn, UPROBE_SWBP_INSN_SIZE);
1161         atomic_set(&area->slot_count, 1);
1162         init_waitqueue_head(&area->wq);
1163
1164         if (!xol_add_vma(area))
1165                 return area;
1166
1167         __free_page(area->page);
1168  free_bitmap:
1169         kfree(area->bitmap);
1170  free_area:
1171         kfree(area);
1172  out:
1173         area = mm->uprobes_state.xol_area;
1174  ret:
1175         smp_read_barrier_depends();     /* pairs with wmb in xol_add_vma() */
1176         return area;
1177 }
1178
1179 /*
1180  * uprobe_clear_state - Free the area allocated for slots.
1181  */
1182 void uprobe_clear_state(struct mm_struct *mm)
1183 {
1184         struct xol_area *area = mm->uprobes_state.xol_area;
1185
1186         if (!area)
1187                 return;
1188
1189         put_page(area->page);
1190         kfree(area->bitmap);
1191         kfree(area);
1192 }
1193
1194 void uprobe_start_dup_mmap(void)
1195 {
1196         percpu_down_read(&dup_mmap_sem);
1197 }
1198
1199 void uprobe_end_dup_mmap(void)
1200 {
1201         percpu_up_read(&dup_mmap_sem);
1202 }
1203
1204 void uprobe_dup_mmap(struct mm_struct *oldmm, struct mm_struct *newmm)
1205 {
1206         newmm->uprobes_state.xol_area = NULL;
1207
1208         if (test_bit(MMF_HAS_UPROBES, &oldmm->flags)) {
1209                 set_bit(MMF_HAS_UPROBES, &newmm->flags);
1210                 /* unconditionally, dup_mmap() skips VM_DONTCOPY vmas */
1211                 set_bit(MMF_RECALC_UPROBES, &newmm->flags);
1212         }
1213 }
1214
1215 /*
1216  *  - search for a free slot.
1217  */
1218 static unsigned long xol_take_insn_slot(struct xol_area *area)
1219 {
1220         unsigned long slot_addr;
1221         int slot_nr;
1222
1223         do {
1224                 slot_nr = find_first_zero_bit(area->bitmap, UINSNS_PER_PAGE);
1225                 if (slot_nr < UINSNS_PER_PAGE) {
1226                         if (!test_and_set_bit(slot_nr, area->bitmap))
1227                                 break;
1228
1229                         slot_nr = UINSNS_PER_PAGE;
1230                         continue;
1231                 }
1232                 wait_event(area->wq, (atomic_read(&area->slot_count) < UINSNS_PER_PAGE));
1233         } while (slot_nr >= UINSNS_PER_PAGE);
1234
1235         slot_addr = area->vaddr + (slot_nr * UPROBE_XOL_SLOT_BYTES);
1236         atomic_inc(&area->slot_count);
1237
1238         return slot_addr;
1239 }
1240
1241 /*
1242  * xol_get_insn_slot - allocate a slot for xol.
1243  * Returns the allocated slot address or 0.
1244  */
1245 static unsigned long xol_get_insn_slot(struct uprobe *uprobe)
1246 {
1247         struct xol_area *area;
1248         unsigned long xol_vaddr;
1249
1250         area = get_xol_area();
1251         if (!area)
1252                 return 0;
1253
1254         xol_vaddr = xol_take_insn_slot(area);
1255         if (unlikely(!xol_vaddr))
1256                 return 0;
1257
1258         /* Initialize the slot */
1259         copy_to_page(area->page, xol_vaddr, uprobe->arch.insn, MAX_UINSN_BYTES);
1260         /*
1261          * We probably need flush_icache_user_range() but it needs vma.
1262          * This should work on supported architectures too.
1263          */
1264         flush_dcache_page(area->page);
1265
1266         return xol_vaddr;
1267 }
1268
1269 /*
1270  * xol_free_insn_slot - If slot was earlier allocated by
1271  * @xol_get_insn_slot(), make the slot available for
1272  * subsequent requests.
1273  */
1274 static void xol_free_insn_slot(struct task_struct *tsk)
1275 {
1276         struct xol_area *area;
1277         unsigned long vma_end;
1278         unsigned long slot_addr;
1279
1280         if (!tsk->mm || !tsk->mm->uprobes_state.xol_area || !tsk->utask)
1281                 return;
1282
1283         slot_addr = tsk->utask->xol_vaddr;
1284         if (unlikely(!slot_addr))
1285                 return;
1286
1287         area = tsk->mm->uprobes_state.xol_area;
1288         vma_end = area->vaddr + PAGE_SIZE;
1289         if (area->vaddr <= slot_addr && slot_addr < vma_end) {
1290                 unsigned long offset;
1291                 int slot_nr;
1292
1293                 offset = slot_addr - area->vaddr;
1294                 slot_nr = offset / UPROBE_XOL_SLOT_BYTES;
1295                 if (slot_nr >= UINSNS_PER_PAGE)
1296                         return;
1297
1298                 clear_bit(slot_nr, area->bitmap);
1299                 atomic_dec(&area->slot_count);
1300                 if (waitqueue_active(&area->wq))
1301                         wake_up(&area->wq);
1302
1303                 tsk->utask->xol_vaddr = 0;
1304         }
1305 }
1306
1307 /**
1308  * uprobe_get_swbp_addr - compute address of swbp given post-swbp regs
1309  * @regs: Reflects the saved state of the task after it has hit a breakpoint
1310  * instruction.
1311  * Return the address of the breakpoint instruction.
1312  */
1313 unsigned long __weak uprobe_get_swbp_addr(struct pt_regs *regs)
1314 {
1315         return instruction_pointer(regs) - UPROBE_SWBP_INSN_SIZE;
1316 }
1317
1318 /*
1319  * Called with no locks held.
1320  * Called in context of a exiting or a exec-ing thread.
1321  */
1322 void uprobe_free_utask(struct task_struct *t)
1323 {
1324         struct uprobe_task *utask = t->utask;
1325         struct return_instance *ri, *tmp;
1326
1327         if (!utask)
1328                 return;
1329
1330         if (utask->active_uprobe)
1331                 put_uprobe(utask->active_uprobe);
1332
1333         ri = utask->return_instances;
1334         while (ri) {
1335                 tmp = ri;
1336                 ri = ri->next;
1337
1338                 put_uprobe(tmp->uprobe);
1339                 kfree(tmp);
1340         }
1341
1342         xol_free_insn_slot(t);
1343         kfree(utask);
1344         t->utask = NULL;
1345 }
1346
1347 /*
1348  * Called in context of a new clone/fork from copy_process.
1349  */
1350 void uprobe_copy_process(struct task_struct *t)
1351 {
1352         t->utask = NULL;
1353 }
1354
1355 /*
1356  * Allocate a uprobe_task object for the task if if necessary.
1357  * Called when the thread hits a breakpoint.
1358  *
1359  * Returns:
1360  * - pointer to new uprobe_task on success
1361  * - NULL otherwise
1362  */
1363 static struct uprobe_task *get_utask(void)
1364 {
1365         if (!current->utask)
1366                 current->utask = kzalloc(sizeof(struct uprobe_task), GFP_KERNEL);
1367         return current->utask;
1368 }
1369
1370 /*
1371  * Current area->vaddr notion assume the trampoline address is always
1372  * equal area->vaddr.
1373  *
1374  * Returns -1 in case the xol_area is not allocated.
1375  */
1376 static unsigned long get_trampoline_vaddr(void)
1377 {
1378         struct xol_area *area;
1379         unsigned long trampoline_vaddr = -1;
1380
1381         area = current->mm->uprobes_state.xol_area;
1382         smp_read_barrier_depends();
1383         if (area)
1384                 trampoline_vaddr = area->vaddr;
1385
1386         return trampoline_vaddr;
1387 }
1388
1389 static void prepare_uretprobe(struct uprobe *uprobe, struct pt_regs *regs)
1390 {
1391         struct return_instance *ri;
1392         struct uprobe_task *utask;
1393         unsigned long orig_ret_vaddr, trampoline_vaddr;
1394         bool chained = false;
1395
1396         if (!get_xol_area())
1397                 return;
1398
1399         utask = get_utask();
1400         if (!utask)
1401                 return;
1402
1403         if (utask->depth >= MAX_URETPROBE_DEPTH) {
1404                 printk_ratelimited(KERN_INFO "uprobe: omit uretprobe due to"
1405                                 " nestedness limit pid/tgid=%d/%d\n",
1406                                 current->pid, current->tgid);
1407                 return;
1408         }
1409
1410         ri = kzalloc(sizeof(struct return_instance), GFP_KERNEL);
1411         if (!ri)
1412                 goto fail;
1413
1414         trampoline_vaddr = get_trampoline_vaddr();
1415         orig_ret_vaddr = arch_uretprobe_hijack_return_addr(trampoline_vaddr, regs);
1416         if (orig_ret_vaddr == -1)
1417                 goto fail;
1418
1419         /*
1420          * We don't want to keep trampoline address in stack, rather keep the
1421          * original return address of first caller thru all the consequent
1422          * instances. This also makes breakpoint unwrapping easier.
1423          */
1424         if (orig_ret_vaddr == trampoline_vaddr) {
1425                 if (!utask->return_instances) {
1426                         /*
1427                          * This situation is not possible. Likely we have an
1428                          * attack from user-space.
1429                          */
1430                         pr_warn("uprobe: unable to set uretprobe pid/tgid=%d/%d\n",
1431                                                 current->pid, current->tgid);
1432                         goto fail;
1433                 }
1434
1435                 chained = true;
1436                 orig_ret_vaddr = utask->return_instances->orig_ret_vaddr;
1437         }
1438
1439         atomic_inc(&uprobe->ref);
1440         ri->uprobe = uprobe;
1441         ri->func = instruction_pointer(regs);
1442         ri->orig_ret_vaddr = orig_ret_vaddr;
1443         ri->chained = chained;
1444
1445         utask->depth++;
1446
1447         /* add instance to the stack */
1448         ri->next = utask->return_instances;
1449         utask->return_instances = ri;
1450
1451         return;
1452
1453  fail:
1454         kfree(ri);
1455 }
1456
1457 /* Prepare to single-step probed instruction out of line. */
1458 static int
1459 pre_ssout(struct uprobe *uprobe, struct pt_regs *regs, unsigned long bp_vaddr)
1460 {
1461         struct uprobe_task *utask;
1462         unsigned long xol_vaddr;
1463         int err;
1464
1465         utask = get_utask();
1466         if (!utask)
1467                 return -ENOMEM;
1468
1469         xol_vaddr = xol_get_insn_slot(uprobe);
1470         if (!xol_vaddr)
1471                 return -ENOMEM;
1472
1473         utask->xol_vaddr = xol_vaddr;
1474         utask->vaddr = bp_vaddr;
1475
1476         err = arch_uprobe_pre_xol(&uprobe->arch, regs);
1477         if (unlikely(err)) {
1478                 xol_free_insn_slot(current);
1479                 return err;
1480         }
1481
1482         utask->active_uprobe = uprobe;
1483         utask->state = UTASK_SSTEP;
1484         return 0;
1485 }
1486
1487 /*
1488  * If we are singlestepping, then ensure this thread is not connected to
1489  * non-fatal signals until completion of singlestep.  When xol insn itself
1490  * triggers the signal,  restart the original insn even if the task is
1491  * already SIGKILL'ed (since coredump should report the correct ip).  This
1492  * is even more important if the task has a handler for SIGSEGV/etc, The
1493  * _same_ instruction should be repeated again after return from the signal
1494  * handler, and SSTEP can never finish in this case.
1495  */
1496 bool uprobe_deny_signal(void)
1497 {
1498         struct task_struct *t = current;
1499         struct uprobe_task *utask = t->utask;
1500
1501         if (likely(!utask || !utask->active_uprobe))
1502                 return false;
1503
1504         WARN_ON_ONCE(utask->state != UTASK_SSTEP);
1505
1506         if (signal_pending(t)) {
1507                 spin_lock_irq(&t->sighand->siglock);
1508                 clear_tsk_thread_flag(t, TIF_SIGPENDING);
1509                 spin_unlock_irq(&t->sighand->siglock);
1510
1511                 if (__fatal_signal_pending(t) || arch_uprobe_xol_was_trapped(t)) {
1512                         utask->state = UTASK_SSTEP_TRAPPED;
1513                         set_tsk_thread_flag(t, TIF_UPROBE);
1514                         set_tsk_thread_flag(t, TIF_NOTIFY_RESUME);
1515                 }
1516         }
1517
1518         return true;
1519 }
1520
1521 /*
1522  * Avoid singlestepping the original instruction if the original instruction
1523  * is a NOP or can be emulated.
1524  */
1525 static bool can_skip_sstep(struct uprobe *uprobe, struct pt_regs *regs)
1526 {
1527         if (test_bit(UPROBE_SKIP_SSTEP, &uprobe->flags)) {
1528                 if (arch_uprobe_skip_sstep(&uprobe->arch, regs))
1529                         return true;
1530                 clear_bit(UPROBE_SKIP_SSTEP, &uprobe->flags);
1531         }
1532         return false;
1533 }
1534
1535 static void mmf_recalc_uprobes(struct mm_struct *mm)
1536 {
1537         struct vm_area_struct *vma;
1538
1539         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1540                 if (!valid_vma(vma, false))
1541                         continue;
1542                 /*
1543                  * This is not strictly accurate, we can race with
1544                  * uprobe_unregister() and see the already removed
1545                  * uprobe if delete_uprobe() was not yet called.
1546                  * Or this uprobe can be filtered out.
1547                  */
1548                 if (vma_has_uprobes(vma, vma->vm_start, vma->vm_end))
1549                         return;
1550         }
1551
1552         clear_bit(MMF_HAS_UPROBES, &mm->flags);
1553 }
1554
1555 static int is_trap_at_addr(struct mm_struct *mm, unsigned long vaddr)
1556 {
1557         struct page *page;
1558         uprobe_opcode_t opcode;
1559         int result;
1560
1561         pagefault_disable();
1562         result = __copy_from_user_inatomic(&opcode, (void __user*)vaddr,
1563                                                         sizeof(opcode));
1564         pagefault_enable();
1565
1566         if (likely(result == 0))
1567                 goto out;
1568
1569         result = get_user_pages(NULL, mm, vaddr, 1, 0, 1, &page, NULL);
1570         if (result < 0)
1571                 return result;
1572
1573         copy_from_page(page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
1574         put_page(page);
1575  out:
1576         /* This needs to return true for any variant of the trap insn */
1577         return is_trap_insn(&opcode);
1578 }
1579
1580 static struct uprobe *find_active_uprobe(unsigned long bp_vaddr, int *is_swbp)
1581 {
1582         struct mm_struct *mm = current->mm;
1583         struct uprobe *uprobe = NULL;
1584         struct vm_area_struct *vma;
1585
1586         down_read(&mm->mmap_sem);
1587         vma = find_vma(mm, bp_vaddr);
1588         if (vma && vma->vm_start <= bp_vaddr) {
1589                 if (valid_vma(vma, false)) {
1590                         struct inode *inode = file_inode(vma->vm_file);
1591                         loff_t offset = vaddr_to_offset(vma, bp_vaddr);
1592
1593                         uprobe = find_uprobe(inode, offset);
1594                 }
1595
1596                 if (!uprobe)
1597                         *is_swbp = is_trap_at_addr(mm, bp_vaddr);
1598         } else {
1599                 *is_swbp = -EFAULT;
1600         }
1601
1602         if (!uprobe && test_and_clear_bit(MMF_RECALC_UPROBES, &mm->flags))
1603                 mmf_recalc_uprobes(mm);
1604         up_read(&mm->mmap_sem);
1605
1606         return uprobe;
1607 }
1608
1609 static void handler_chain(struct uprobe *uprobe, struct pt_regs *regs)
1610 {
1611         struct uprobe_consumer *uc;
1612         int remove = UPROBE_HANDLER_REMOVE;
1613         bool need_prep = false; /* prepare return uprobe, when needed */
1614
1615         down_read(&uprobe->register_rwsem);
1616         for (uc = uprobe->consumers; uc; uc = uc->next) {
1617                 int rc = 0;
1618
1619                 if (uc->handler) {
1620                         rc = uc->handler(uc, regs);
1621                         WARN(rc & ~UPROBE_HANDLER_MASK,
1622                                 "bad rc=0x%x from %pf()\n", rc, uc->handler);
1623                 }
1624
1625                 if (uc->ret_handler)
1626                         need_prep = true;
1627
1628                 remove &= rc;
1629         }
1630
1631         if (need_prep && !remove)
1632                 prepare_uretprobe(uprobe, regs); /* put bp at return */
1633
1634         if (remove && uprobe->consumers) {
1635                 WARN_ON(!uprobe_is_active(uprobe));
1636                 unapply_uprobe(uprobe, current->mm);
1637         }
1638         up_read(&uprobe->register_rwsem);
1639 }
1640
1641 static void
1642 handle_uretprobe_chain(struct return_instance *ri, struct pt_regs *regs)
1643 {
1644         struct uprobe *uprobe = ri->uprobe;
1645         struct uprobe_consumer *uc;
1646
1647         down_read(&uprobe->register_rwsem);
1648         for (uc = uprobe->consumers; uc; uc = uc->next) {
1649                 if (uc->ret_handler)
1650                         uc->ret_handler(uc, ri->func, regs);
1651         }
1652         up_read(&uprobe->register_rwsem);
1653 }
1654
1655 static bool handle_trampoline(struct pt_regs *regs)
1656 {
1657         struct uprobe_task *utask;
1658         struct return_instance *ri, *tmp;
1659         bool chained;
1660
1661         utask = current->utask;
1662         if (!utask)
1663                 return false;
1664
1665         ri = utask->return_instances;
1666         if (!ri)
1667                 return false;
1668
1669         /*
1670          * TODO: we should throw out return_instance's invalidated by
1671          * longjmp(), currently we assume that the probed function always
1672          * returns.
1673          */
1674         instruction_pointer_set(regs, ri->orig_ret_vaddr);
1675
1676         for (;;) {
1677                 handle_uretprobe_chain(ri, regs);
1678
1679                 chained = ri->chained;
1680                 put_uprobe(ri->uprobe);
1681
1682                 tmp = ri;
1683                 ri = ri->next;
1684                 kfree(tmp);
1685                 utask->depth--;
1686
1687                 if (!chained)
1688                         break;
1689                 BUG_ON(!ri);
1690         }
1691
1692         utask->return_instances = ri;
1693
1694         return true;
1695 }
1696
1697 /*
1698  * Run handler and ask thread to singlestep.
1699  * Ensure all non-fatal signals cannot interrupt thread while it singlesteps.
1700  */
1701 static void handle_swbp(struct pt_regs *regs)
1702 {
1703         struct uprobe *uprobe;
1704         unsigned long bp_vaddr;
1705         int uninitialized_var(is_swbp);
1706
1707         bp_vaddr = uprobe_get_swbp_addr(regs);
1708         if (bp_vaddr == get_trampoline_vaddr()) {
1709                 if (handle_trampoline(regs))
1710                         return;
1711
1712                 pr_warn("uprobe: unable to handle uretprobe pid/tgid=%d/%d\n",
1713                                                 current->pid, current->tgid);
1714         }
1715
1716         uprobe = find_active_uprobe(bp_vaddr, &is_swbp);
1717         if (!uprobe) {
1718                 if (is_swbp > 0) {
1719                         /* No matching uprobe; signal SIGTRAP. */
1720                         send_sig(SIGTRAP, current, 0);
1721                 } else {
1722                         /*
1723                          * Either we raced with uprobe_unregister() or we can't
1724                          * access this memory. The latter is only possible if
1725                          * another thread plays with our ->mm. In both cases
1726                          * we can simply restart. If this vma was unmapped we
1727                          * can pretend this insn was not executed yet and get
1728                          * the (correct) SIGSEGV after restart.
1729                          */
1730                         instruction_pointer_set(regs, bp_vaddr);
1731                 }
1732                 return;
1733         }
1734
1735         /* change it in advance for ->handler() and restart */
1736         instruction_pointer_set(regs, bp_vaddr);
1737
1738         /*
1739          * TODO: move copy_insn/etc into _register and remove this hack.
1740          * After we hit the bp, _unregister + _register can install the
1741          * new and not-yet-analyzed uprobe at the same address, restart.
1742          */
1743         smp_rmb(); /* pairs with wmb() in install_breakpoint() */
1744         if (unlikely(!test_bit(UPROBE_COPY_INSN, &uprobe->flags)))
1745                 goto out;
1746
1747         handler_chain(uprobe, regs);
1748         if (can_skip_sstep(uprobe, regs))
1749                 goto out;
1750
1751         if (!pre_ssout(uprobe, regs, bp_vaddr))
1752                 return;
1753
1754         /* can_skip_sstep() succeeded, or restart if can't singlestep */
1755 out:
1756         put_uprobe(uprobe);
1757 }
1758
1759 /*
1760  * Perform required fix-ups and disable singlestep.
1761  * Allow pending signals to take effect.
1762  */
1763 static void handle_singlestep(struct uprobe_task *utask, struct pt_regs *regs)
1764 {
1765         struct uprobe *uprobe;
1766
1767         uprobe = utask->active_uprobe;
1768         if (utask->state == UTASK_SSTEP_ACK)
1769                 arch_uprobe_post_xol(&uprobe->arch, regs);
1770         else if (utask->state == UTASK_SSTEP_TRAPPED)
1771                 arch_uprobe_abort_xol(&uprobe->arch, regs);
1772         else
1773                 WARN_ON_ONCE(1);
1774
1775         put_uprobe(uprobe);
1776         utask->active_uprobe = NULL;
1777         utask->state = UTASK_RUNNING;
1778         xol_free_insn_slot(current);
1779
1780         spin_lock_irq(&current->sighand->siglock);
1781         recalc_sigpending(); /* see uprobe_deny_signal() */
1782         spin_unlock_irq(&current->sighand->siglock);
1783 }
1784
1785 /*
1786  * On breakpoint hit, breakpoint notifier sets the TIF_UPROBE flag and
1787  * allows the thread to return from interrupt. After that handle_swbp()
1788  * sets utask->active_uprobe.
1789  *
1790  * On singlestep exception, singlestep notifier sets the TIF_UPROBE flag
1791  * and allows the thread to return from interrupt.
1792  *
1793  * While returning to userspace, thread notices the TIF_UPROBE flag and calls
1794  * uprobe_notify_resume().
1795  */
1796 void uprobe_notify_resume(struct pt_regs *regs)
1797 {
1798         struct uprobe_task *utask;
1799
1800         clear_thread_flag(TIF_UPROBE);
1801
1802         utask = current->utask;
1803         if (utask && utask->active_uprobe)
1804                 handle_singlestep(utask, regs);
1805         else
1806                 handle_swbp(regs);
1807 }
1808
1809 /*
1810  * uprobe_pre_sstep_notifier gets called from interrupt context as part of
1811  * notifier mechanism. Set TIF_UPROBE flag and indicate breakpoint hit.
1812  */
1813 int uprobe_pre_sstep_notifier(struct pt_regs *regs)
1814 {
1815         if (!current->mm)
1816                 return 0;
1817
1818         if (!test_bit(MMF_HAS_UPROBES, &current->mm->flags) &&
1819             (!current->utask || !current->utask->return_instances))
1820                 return 0;
1821
1822         set_thread_flag(TIF_UPROBE);
1823         return 1;
1824 }
1825
1826 /*
1827  * uprobe_post_sstep_notifier gets called in interrupt context as part of notifier
1828  * mechanism. Set TIF_UPROBE flag and indicate completion of singlestep.
1829  */
1830 int uprobe_post_sstep_notifier(struct pt_regs *regs)
1831 {
1832         struct uprobe_task *utask = current->utask;
1833
1834         if (!current->mm || !utask || !utask->active_uprobe)
1835                 /* task is currently not uprobed */
1836                 return 0;
1837
1838         utask->state = UTASK_SSTEP_ACK;
1839         set_thread_flag(TIF_UPROBE);
1840         return 1;
1841 }
1842
1843 static struct notifier_block uprobe_exception_nb = {
1844         .notifier_call          = arch_uprobe_exception_notify,
1845         .priority               = INT_MAX-1,    /* notified after kprobes, kgdb */
1846 };
1847
1848 static int __init init_uprobes(void)
1849 {
1850         int i;
1851
1852         for (i = 0; i < UPROBES_HASH_SZ; i++)
1853                 mutex_init(&uprobes_mmap_mutex[i]);
1854
1855         if (percpu_init_rwsem(&dup_mmap_sem))
1856                 return -ENOMEM;
1857
1858         return register_die_notifier(&uprobe_exception_nb);
1859 }
1860 module_init(init_uprobes);
1861
1862 static void __exit exit_uprobes(void)
1863 {
1864 }
1865 module_exit(exit_uprobes);