Loading...
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * User-space Probes (UProbes)
4 *
5 * Copyright (C) IBM Corporation, 2008-2012
6 * Authors:
7 * Srikar Dronamraju
8 * Jim Keniston
9 * Copyright (C) 2011-2012 Red Hat, Inc., Peter Zijlstra
10 */
11
12#include <linux/kernel.h>
13#include <linux/highmem.h>
14#include <linux/pagemap.h> /* read_mapping_page */
15#include <linux/slab.h>
16#include <linux/sched.h>
17#include <linux/sched/mm.h>
18#include <linux/export.h>
19#include <linux/rmap.h> /* anon_vma_prepare */
20#include <linux/mmu_notifier.h>
21#include <linux/swap.h> /* folio_free_swap */
22#include <linux/ptrace.h> /* user_enable_single_step */
23#include <linux/kdebug.h> /* notifier mechanism */
24#include <linux/percpu-rwsem.h>
25#include <linux/task_work.h>
26#include <linux/shmem_fs.h>
27#include <linux/khugepaged.h>
28#include <linux/rcupdate_trace.h>
29#include <linux/workqueue.h>
30#include <linux/srcu.h>
31#include <linux/oom.h> /* check_stable_address_space */
32
33#include <linux/uprobes.h>
34
35#define UINSNS_PER_PAGE (PAGE_SIZE/UPROBE_XOL_SLOT_BYTES)
36#define MAX_UPROBE_XOL_SLOTS UINSNS_PER_PAGE
37
38static struct rb_root uprobes_tree = RB_ROOT;
39/*
40 * allows us to skip the uprobe_mmap if there are no uprobe events active
41 * at this time. Probably a fine grained per inode count is better?
42 */
43#define no_uprobe_events() RB_EMPTY_ROOT(&uprobes_tree)
44
45static DEFINE_RWLOCK(uprobes_treelock); /* serialize rbtree access */
46static seqcount_rwlock_t uprobes_seqcount = SEQCNT_RWLOCK_ZERO(uprobes_seqcount, &uprobes_treelock);
47
48#define UPROBES_HASH_SZ 13
49/* serialize uprobe->pending_list */
50static struct mutex uprobes_mmap_mutex[UPROBES_HASH_SZ];
51#define uprobes_mmap_hash(v) (&uprobes_mmap_mutex[((unsigned long)(v)) % UPROBES_HASH_SZ])
52
53DEFINE_STATIC_PERCPU_RWSEM(dup_mmap_sem);
54
55/* Covers return_instance's uprobe lifetime. */
56DEFINE_STATIC_SRCU(uretprobes_srcu);
57
58/* Have a copy of original instruction */
59#define UPROBE_COPY_INSN 0
60
61struct uprobe {
62 struct rb_node rb_node; /* node in the rb tree */
63 refcount_t ref;
64 struct rw_semaphore register_rwsem;
65 struct rw_semaphore consumer_rwsem;
66 struct list_head pending_list;
67 struct list_head consumers;
68 struct inode *inode; /* Also hold a ref to inode */
69 union {
70 struct rcu_head rcu;
71 struct work_struct work;
72 };
73 loff_t offset;
74 loff_t ref_ctr_offset;
75 unsigned long flags; /* "unsigned long" so bitops work */
76
77 /*
78 * The generic code assumes that it has two members of unknown type
79 * owned by the arch-specific code:
80 *
81 * insn - copy_insn() saves the original instruction here for
82 * arch_uprobe_analyze_insn().
83 *
84 * ixol - potentially modified instruction to execute out of
85 * line, copied to xol_area by xol_get_insn_slot().
86 */
87 struct arch_uprobe arch;
88};
89
90struct delayed_uprobe {
91 struct list_head list;
92 struct uprobe *uprobe;
93 struct mm_struct *mm;
94};
95
96static DEFINE_MUTEX(delayed_uprobe_lock);
97static LIST_HEAD(delayed_uprobe_list);
98
99/*
100 * Execute out of line area: anonymous executable mapping installed
101 * by the probed task to execute the copy of the original instruction
102 * mangled by set_swbp().
103 *
104 * On a breakpoint hit, thread contests for a slot. It frees the
105 * slot after singlestep. Currently a fixed number of slots are
106 * allocated.
107 */
108struct xol_area {
109 wait_queue_head_t wq; /* if all slots are busy */
110 unsigned long *bitmap; /* 0 = free slot */
111
112 struct page *page;
113 /*
114 * We keep the vma's vm_start rather than a pointer to the vma
115 * itself. The probed process or a naughty kernel module could make
116 * the vma go away, and we must handle that reasonably gracefully.
117 */
118 unsigned long vaddr; /* Page(s) of instruction slots */
119};
120
121static void uprobe_warn(struct task_struct *t, const char *msg)
122{
123 pr_warn("uprobe: %s:%d failed to %s\n", current->comm, current->pid, msg);
124}
125
126/*
127 * valid_vma: Verify if the specified vma is an executable vma
128 * Relax restrictions while unregistering: vm_flags might have
129 * changed after breakpoint was inserted.
130 * - is_register: indicates if we are in register context.
131 * - Return 1 if the specified virtual address is in an
132 * executable vma.
133 */
134static bool valid_vma(struct vm_area_struct *vma, bool is_register)
135{
136 vm_flags_t flags = VM_HUGETLB | VM_MAYEXEC | VM_MAYSHARE;
137
138 if (is_register)
139 flags |= VM_WRITE;
140
141 return vma->vm_file && (vma->vm_flags & flags) == VM_MAYEXEC;
142}
143
144static unsigned long offset_to_vaddr(struct vm_area_struct *vma, loff_t offset)
145{
146 return vma->vm_start + offset - ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
147}
148
149static loff_t vaddr_to_offset(struct vm_area_struct *vma, unsigned long vaddr)
150{
151 return ((loff_t)vma->vm_pgoff << PAGE_SHIFT) + (vaddr - vma->vm_start);
152}
153
154/**
155 * __replace_page - replace page in vma by new page.
156 * based on replace_page in mm/ksm.c
157 *
158 * @vma: vma that holds the pte pointing to page
159 * @addr: address the old @page is mapped at
160 * @old_page: the page we are replacing by new_page
161 * @new_page: the modified page we replace page by
162 *
163 * If @new_page is NULL, only unmap @old_page.
164 *
165 * Returns 0 on success, negative error code otherwise.
166 */
167static int __replace_page(struct vm_area_struct *vma, unsigned long addr,
168 struct page *old_page, struct page *new_page)
169{
170 struct folio *old_folio = page_folio(old_page);
171 struct folio *new_folio;
172 struct mm_struct *mm = vma->vm_mm;
173 DEFINE_FOLIO_VMA_WALK(pvmw, old_folio, vma, addr, 0);
174 int err;
175 struct mmu_notifier_range range;
176
177 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr,
178 addr + PAGE_SIZE);
179
180 if (new_page) {
181 new_folio = page_folio(new_page);
182 err = mem_cgroup_charge(new_folio, vma->vm_mm, GFP_KERNEL);
183 if (err)
184 return err;
185 }
186
187 /* For folio_free_swap() below */
188 folio_lock(old_folio);
189
190 mmu_notifier_invalidate_range_start(&range);
191 err = -EAGAIN;
192 if (!page_vma_mapped_walk(&pvmw))
193 goto unlock;
194 VM_BUG_ON_PAGE(addr != pvmw.address, old_page);
195
196 if (new_page) {
197 folio_get(new_folio);
198 folio_add_new_anon_rmap(new_folio, vma, addr, RMAP_EXCLUSIVE);
199 folio_add_lru_vma(new_folio, vma);
200 } else
201 /* no new page, just dec_mm_counter for old_page */
202 dec_mm_counter(mm, MM_ANONPAGES);
203
204 if (!folio_test_anon(old_folio)) {
205 dec_mm_counter(mm, mm_counter_file(old_folio));
206 inc_mm_counter(mm, MM_ANONPAGES);
207 }
208
209 flush_cache_page(vma, addr, pte_pfn(ptep_get(pvmw.pte)));
210 ptep_clear_flush(vma, addr, pvmw.pte);
211 if (new_page)
212 set_pte_at(mm, addr, pvmw.pte,
213 mk_pte(new_page, vma->vm_page_prot));
214
215 folio_remove_rmap_pte(old_folio, old_page, vma);
216 if (!folio_mapped(old_folio))
217 folio_free_swap(old_folio);
218 page_vma_mapped_walk_done(&pvmw);
219 folio_put(old_folio);
220
221 err = 0;
222 unlock:
223 mmu_notifier_invalidate_range_end(&range);
224 folio_unlock(old_folio);
225 return err;
226}
227
228/**
229 * is_swbp_insn - check if instruction is breakpoint instruction.
230 * @insn: instruction to be checked.
231 * Default implementation of is_swbp_insn
232 * Returns true if @insn is a breakpoint instruction.
233 */
234bool __weak is_swbp_insn(uprobe_opcode_t *insn)
235{
236 return *insn == UPROBE_SWBP_INSN;
237}
238
239/**
240 * is_trap_insn - check if instruction is breakpoint instruction.
241 * @insn: instruction to be checked.
242 * Default implementation of is_trap_insn
243 * Returns true if @insn is a breakpoint instruction.
244 *
245 * This function is needed for the case where an architecture has multiple
246 * trap instructions (like powerpc).
247 */
248bool __weak is_trap_insn(uprobe_opcode_t *insn)
249{
250 return is_swbp_insn(insn);
251}
252
253static void copy_from_page(struct page *page, unsigned long vaddr, void *dst, int len)
254{
255 void *kaddr = kmap_atomic(page);
256 memcpy(dst, kaddr + (vaddr & ~PAGE_MASK), len);
257 kunmap_atomic(kaddr);
258}
259
260static void copy_to_page(struct page *page, unsigned long vaddr, const void *src, int len)
261{
262 void *kaddr = kmap_atomic(page);
263 memcpy(kaddr + (vaddr & ~PAGE_MASK), src, len);
264 kunmap_atomic(kaddr);
265}
266
267static int verify_opcode(struct page *page, unsigned long vaddr, uprobe_opcode_t *new_opcode)
268{
269 uprobe_opcode_t old_opcode;
270 bool is_swbp;
271
272 /*
273 * Note: We only check if the old_opcode is UPROBE_SWBP_INSN here.
274 * We do not check if it is any other 'trap variant' which could
275 * be conditional trap instruction such as the one powerpc supports.
276 *
277 * The logic is that we do not care if the underlying instruction
278 * is a trap variant; uprobes always wins over any other (gdb)
279 * breakpoint.
280 */
281 copy_from_page(page, vaddr, &old_opcode, UPROBE_SWBP_INSN_SIZE);
282 is_swbp = is_swbp_insn(&old_opcode);
283
284 if (is_swbp_insn(new_opcode)) {
285 if (is_swbp) /* register: already installed? */
286 return 0;
287 } else {
288 if (!is_swbp) /* unregister: was it changed by us? */
289 return 0;
290 }
291
292 return 1;
293}
294
295static struct delayed_uprobe *
296delayed_uprobe_check(struct uprobe *uprobe, struct mm_struct *mm)
297{
298 struct delayed_uprobe *du;
299
300 list_for_each_entry(du, &delayed_uprobe_list, list)
301 if (du->uprobe == uprobe && du->mm == mm)
302 return du;
303 return NULL;
304}
305
306static int delayed_uprobe_add(struct uprobe *uprobe, struct mm_struct *mm)
307{
308 struct delayed_uprobe *du;
309
310 if (delayed_uprobe_check(uprobe, mm))
311 return 0;
312
313 du = kzalloc(sizeof(*du), GFP_KERNEL);
314 if (!du)
315 return -ENOMEM;
316
317 du->uprobe = uprobe;
318 du->mm = mm;
319 list_add(&du->list, &delayed_uprobe_list);
320 return 0;
321}
322
323static void delayed_uprobe_delete(struct delayed_uprobe *du)
324{
325 if (WARN_ON(!du))
326 return;
327 list_del(&du->list);
328 kfree(du);
329}
330
331static void delayed_uprobe_remove(struct uprobe *uprobe, struct mm_struct *mm)
332{
333 struct list_head *pos, *q;
334 struct delayed_uprobe *du;
335
336 if (!uprobe && !mm)
337 return;
338
339 list_for_each_safe(pos, q, &delayed_uprobe_list) {
340 du = list_entry(pos, struct delayed_uprobe, list);
341
342 if (uprobe && du->uprobe != uprobe)
343 continue;
344 if (mm && du->mm != mm)
345 continue;
346
347 delayed_uprobe_delete(du);
348 }
349}
350
351static bool valid_ref_ctr_vma(struct uprobe *uprobe,
352 struct vm_area_struct *vma)
353{
354 unsigned long vaddr = offset_to_vaddr(vma, uprobe->ref_ctr_offset);
355
356 return uprobe->ref_ctr_offset &&
357 vma->vm_file &&
358 file_inode(vma->vm_file) == uprobe->inode &&
359 (vma->vm_flags & (VM_WRITE|VM_SHARED)) == VM_WRITE &&
360 vma->vm_start <= vaddr &&
361 vma->vm_end > vaddr;
362}
363
364static struct vm_area_struct *
365find_ref_ctr_vma(struct uprobe *uprobe, struct mm_struct *mm)
366{
367 VMA_ITERATOR(vmi, mm, 0);
368 struct vm_area_struct *tmp;
369
370 for_each_vma(vmi, tmp)
371 if (valid_ref_ctr_vma(uprobe, tmp))
372 return tmp;
373
374 return NULL;
375}
376
377static int
378__update_ref_ctr(struct mm_struct *mm, unsigned long vaddr, short d)
379{
380 void *kaddr;
381 struct page *page;
382 int ret;
383 short *ptr;
384
385 if (!vaddr || !d)
386 return -EINVAL;
387
388 ret = get_user_pages_remote(mm, vaddr, 1,
389 FOLL_WRITE, &page, NULL);
390 if (unlikely(ret <= 0)) {
391 /*
392 * We are asking for 1 page. If get_user_pages_remote() fails,
393 * it may return 0, in that case we have to return error.
394 */
395 return ret == 0 ? -EBUSY : ret;
396 }
397
398 kaddr = kmap_atomic(page);
399 ptr = kaddr + (vaddr & ~PAGE_MASK);
400
401 if (unlikely(*ptr + d < 0)) {
402 pr_warn("ref_ctr going negative. vaddr: 0x%lx, "
403 "curr val: %d, delta: %d\n", vaddr, *ptr, d);
404 ret = -EINVAL;
405 goto out;
406 }
407
408 *ptr += d;
409 ret = 0;
410out:
411 kunmap_atomic(kaddr);
412 put_page(page);
413 return ret;
414}
415
416static void update_ref_ctr_warn(struct uprobe *uprobe,
417 struct mm_struct *mm, short d)
418{
419 pr_warn("ref_ctr %s failed for inode: 0x%lx offset: "
420 "0x%llx ref_ctr_offset: 0x%llx of mm: 0x%pK\n",
421 d > 0 ? "increment" : "decrement", uprobe->inode->i_ino,
422 (unsigned long long) uprobe->offset,
423 (unsigned long long) uprobe->ref_ctr_offset, mm);
424}
425
426static int update_ref_ctr(struct uprobe *uprobe, struct mm_struct *mm,
427 short d)
428{
429 struct vm_area_struct *rc_vma;
430 unsigned long rc_vaddr;
431 int ret = 0;
432
433 rc_vma = find_ref_ctr_vma(uprobe, mm);
434
435 if (rc_vma) {
436 rc_vaddr = offset_to_vaddr(rc_vma, uprobe->ref_ctr_offset);
437 ret = __update_ref_ctr(mm, rc_vaddr, d);
438 if (ret)
439 update_ref_ctr_warn(uprobe, mm, d);
440
441 if (d > 0)
442 return ret;
443 }
444
445 mutex_lock(&delayed_uprobe_lock);
446 if (d > 0)
447 ret = delayed_uprobe_add(uprobe, mm);
448 else
449 delayed_uprobe_remove(uprobe, mm);
450 mutex_unlock(&delayed_uprobe_lock);
451
452 return ret;
453}
454
455/*
456 * NOTE:
457 * Expect the breakpoint instruction to be the smallest size instruction for
458 * the architecture. If an arch has variable length instruction and the
459 * breakpoint instruction is not of the smallest length instruction
460 * supported by that architecture then we need to modify is_trap_at_addr and
461 * uprobe_write_opcode accordingly. This would never be a problem for archs
462 * that have fixed length instructions.
463 *
464 * uprobe_write_opcode - write the opcode at a given virtual address.
465 * @auprobe: arch specific probepoint information.
466 * @mm: the probed process address space.
467 * @vaddr: the virtual address to store the opcode.
468 * @opcode: opcode to be written at @vaddr.
469 *
470 * Called with mm->mmap_lock held for read or write.
471 * Return 0 (success) or a negative errno.
472 */
473int uprobe_write_opcode(struct arch_uprobe *auprobe, struct mm_struct *mm,
474 unsigned long vaddr, uprobe_opcode_t opcode)
475{
476 struct uprobe *uprobe;
477 struct page *old_page, *new_page;
478 struct vm_area_struct *vma;
479 int ret, is_register, ref_ctr_updated = 0;
480 bool orig_page_huge = false;
481 unsigned int gup_flags = FOLL_FORCE;
482
483 is_register = is_swbp_insn(&opcode);
484 uprobe = container_of(auprobe, struct uprobe, arch);
485
486retry:
487 if (is_register)
488 gup_flags |= FOLL_SPLIT_PMD;
489 /* Read the page with vaddr into memory */
490 old_page = get_user_page_vma_remote(mm, vaddr, gup_flags, &vma);
491 if (IS_ERR(old_page))
492 return PTR_ERR(old_page);
493
494 ret = verify_opcode(old_page, vaddr, &opcode);
495 if (ret <= 0)
496 goto put_old;
497
498 if (is_zero_page(old_page)) {
499 ret = -EINVAL;
500 goto put_old;
501 }
502
503 if (WARN(!is_register && PageCompound(old_page),
504 "uprobe unregister should never work on compound page\n")) {
505 ret = -EINVAL;
506 goto put_old;
507 }
508
509 /* We are going to replace instruction, update ref_ctr. */
510 if (!ref_ctr_updated && uprobe->ref_ctr_offset) {
511 ret = update_ref_ctr(uprobe, mm, is_register ? 1 : -1);
512 if (ret)
513 goto put_old;
514
515 ref_ctr_updated = 1;
516 }
517
518 ret = 0;
519 if (!is_register && !PageAnon(old_page))
520 goto put_old;
521
522 ret = anon_vma_prepare(vma);
523 if (ret)
524 goto put_old;
525
526 ret = -ENOMEM;
527 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vaddr);
528 if (!new_page)
529 goto put_old;
530
531 __SetPageUptodate(new_page);
532 copy_highpage(new_page, old_page);
533 copy_to_page(new_page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
534
535 if (!is_register) {
536 struct page *orig_page;
537 pgoff_t index;
538
539 VM_BUG_ON_PAGE(!PageAnon(old_page), old_page);
540
541 index = vaddr_to_offset(vma, vaddr & PAGE_MASK) >> PAGE_SHIFT;
542 orig_page = find_get_page(vma->vm_file->f_inode->i_mapping,
543 index);
544
545 if (orig_page) {
546 if (PageUptodate(orig_page) &&
547 pages_identical(new_page, orig_page)) {
548 /* let go new_page */
549 put_page(new_page);
550 new_page = NULL;
551
552 if (PageCompound(orig_page))
553 orig_page_huge = true;
554 }
555 put_page(orig_page);
556 }
557 }
558
559 ret = __replace_page(vma, vaddr & PAGE_MASK, old_page, new_page);
560 if (new_page)
561 put_page(new_page);
562put_old:
563 put_page(old_page);
564
565 if (unlikely(ret == -EAGAIN))
566 goto retry;
567
568 /* Revert back reference counter if instruction update failed. */
569 if (ret && is_register && ref_ctr_updated)
570 update_ref_ctr(uprobe, mm, -1);
571
572 /* try collapse pmd for compound page */
573 if (!ret && orig_page_huge)
574 collapse_pte_mapped_thp(mm, vaddr, false);
575
576 return ret;
577}
578
579/**
580 * set_swbp - store breakpoint at a given address.
581 * @auprobe: arch specific probepoint information.
582 * @mm: the probed process address space.
583 * @vaddr: the virtual address to insert the opcode.
584 *
585 * For mm @mm, store the breakpoint instruction at @vaddr.
586 * Return 0 (success) or a negative errno.
587 */
588int __weak set_swbp(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
589{
590 return uprobe_write_opcode(auprobe, mm, vaddr, UPROBE_SWBP_INSN);
591}
592
593/**
594 * set_orig_insn - Restore the original instruction.
595 * @mm: the probed process address space.
596 * @auprobe: arch specific probepoint information.
597 * @vaddr: the virtual address to insert the opcode.
598 *
599 * For mm @mm, restore the original opcode (opcode) at @vaddr.
600 * Return 0 (success) or a negative errno.
601 */
602int __weak
603set_orig_insn(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
604{
605 return uprobe_write_opcode(auprobe, mm, vaddr,
606 *(uprobe_opcode_t *)&auprobe->insn);
607}
608
609/* uprobe should have guaranteed positive refcount */
610static struct uprobe *get_uprobe(struct uprobe *uprobe)
611{
612 refcount_inc(&uprobe->ref);
613 return uprobe;
614}
615
616/*
617 * uprobe should have guaranteed lifetime, which can be either of:
618 * - caller already has refcount taken (and wants an extra one);
619 * - uprobe is RCU protected and won't be freed until after grace period;
620 * - we are holding uprobes_treelock (for read or write, doesn't matter).
621 */
622static struct uprobe *try_get_uprobe(struct uprobe *uprobe)
623{
624 if (refcount_inc_not_zero(&uprobe->ref))
625 return uprobe;
626 return NULL;
627}
628
629static inline bool uprobe_is_active(struct uprobe *uprobe)
630{
631 return !RB_EMPTY_NODE(&uprobe->rb_node);
632}
633
634static void uprobe_free_rcu_tasks_trace(struct rcu_head *rcu)
635{
636 struct uprobe *uprobe = container_of(rcu, struct uprobe, rcu);
637
638 kfree(uprobe);
639}
640
641static void uprobe_free_srcu(struct rcu_head *rcu)
642{
643 struct uprobe *uprobe = container_of(rcu, struct uprobe, rcu);
644
645 call_rcu_tasks_trace(&uprobe->rcu, uprobe_free_rcu_tasks_trace);
646}
647
648static void uprobe_free_deferred(struct work_struct *work)
649{
650 struct uprobe *uprobe = container_of(work, struct uprobe, work);
651
652 write_lock(&uprobes_treelock);
653
654 if (uprobe_is_active(uprobe)) {
655 write_seqcount_begin(&uprobes_seqcount);
656 rb_erase(&uprobe->rb_node, &uprobes_tree);
657 write_seqcount_end(&uprobes_seqcount);
658 }
659
660 write_unlock(&uprobes_treelock);
661
662 /*
663 * If application munmap(exec_vma) before uprobe_unregister()
664 * gets called, we don't get a chance to remove uprobe from
665 * delayed_uprobe_list from remove_breakpoint(). Do it here.
666 */
667 mutex_lock(&delayed_uprobe_lock);
668 delayed_uprobe_remove(uprobe, NULL);
669 mutex_unlock(&delayed_uprobe_lock);
670
671 /* start srcu -> rcu_tasks_trace -> kfree chain */
672 call_srcu(&uretprobes_srcu, &uprobe->rcu, uprobe_free_srcu);
673}
674
675static void put_uprobe(struct uprobe *uprobe)
676{
677 if (!refcount_dec_and_test(&uprobe->ref))
678 return;
679
680 INIT_WORK(&uprobe->work, uprobe_free_deferred);
681 schedule_work(&uprobe->work);
682}
683
684/* Initialize hprobe as SRCU-protected "leased" uprobe */
685static void hprobe_init_leased(struct hprobe *hprobe, struct uprobe *uprobe, int srcu_idx)
686{
687 WARN_ON(!uprobe);
688 hprobe->state = HPROBE_LEASED;
689 hprobe->uprobe = uprobe;
690 hprobe->srcu_idx = srcu_idx;
691}
692
693/* Initialize hprobe as refcounted ("stable") uprobe (uprobe can be NULL). */
694static void hprobe_init_stable(struct hprobe *hprobe, struct uprobe *uprobe)
695{
696 hprobe->state = uprobe ? HPROBE_STABLE : HPROBE_GONE;
697 hprobe->uprobe = uprobe;
698 hprobe->srcu_idx = -1;
699}
700
701/*
702 * hprobe_consume() fetches hprobe's underlying uprobe and detects whether
703 * uprobe is SRCU protected or is refcounted. hprobe_consume() can be
704 * used only once for a given hprobe.
705 *
706 * Caller has to call hprobe_finalize() and pass previous hprobe_state, so
707 * that hprobe_finalize() can perform SRCU unlock or put uprobe, whichever
708 * is appropriate.
709 */
710static inline struct uprobe *hprobe_consume(struct hprobe *hprobe, enum hprobe_state *hstate)
711{
712 *hstate = xchg(&hprobe->state, HPROBE_CONSUMED);
713 switch (*hstate) {
714 case HPROBE_LEASED:
715 case HPROBE_STABLE:
716 return hprobe->uprobe;
717 case HPROBE_GONE: /* uprobe is NULL, no SRCU */
718 case HPROBE_CONSUMED: /* uprobe was finalized already, do nothing */
719 return NULL;
720 default:
721 WARN(1, "hprobe invalid state %d", *hstate);
722 return NULL;
723 }
724}
725
726/*
727 * Reset hprobe state and, if hprobe was LEASED, release SRCU lock.
728 * hprobe_finalize() can only be used from current context after
729 * hprobe_consume() call (which determines uprobe and hstate value).
730 */
731static void hprobe_finalize(struct hprobe *hprobe, enum hprobe_state hstate)
732{
733 switch (hstate) {
734 case HPROBE_LEASED:
735 __srcu_read_unlock(&uretprobes_srcu, hprobe->srcu_idx);
736 break;
737 case HPROBE_STABLE:
738 put_uprobe(hprobe->uprobe);
739 break;
740 case HPROBE_GONE:
741 case HPROBE_CONSUMED:
742 break;
743 default:
744 WARN(1, "hprobe invalid state %d", hstate);
745 break;
746 }
747}
748
749/*
750 * Attempt to switch (atomically) uprobe from being SRCU protected (LEASED)
751 * to refcounted (STABLE) state. Competes with hprobe_consume(); only one of
752 * them can win the race to perform SRCU unlocking. Whoever wins must perform
753 * SRCU unlock.
754 *
755 * Returns underlying valid uprobe or NULL, if there was no underlying uprobe
756 * to begin with or we failed to bump its refcount and it's going away.
757 *
758 * Returned non-NULL uprobe can be still safely used within an ongoing SRCU
759 * locked region. If `get` is true, it's guaranteed that non-NULL uprobe has
760 * an extra refcount for caller to assume and use. Otherwise, it's not
761 * guaranteed that returned uprobe has a positive refcount, so caller has to
762 * attempt try_get_uprobe(), if it needs to preserve uprobe beyond current
763 * SRCU lock region. See dup_utask().
764 */
765static struct uprobe *hprobe_expire(struct hprobe *hprobe, bool get)
766{
767 enum hprobe_state hstate;
768
769 /*
770 * Caller should guarantee that return_instance is not going to be
771 * freed from under us. This can be achieved either through holding
772 * rcu_read_lock() or by owning return_instance in the first place.
773 *
774 * Underlying uprobe is itself protected from reuse by SRCU, so ensure
775 * SRCU lock is held properly.
776 */
777 lockdep_assert(srcu_read_lock_held(&uretprobes_srcu));
778
779 hstate = READ_ONCE(hprobe->state);
780 switch (hstate) {
781 case HPROBE_STABLE:
782 /* uprobe has positive refcount, bump refcount, if necessary */
783 return get ? get_uprobe(hprobe->uprobe) : hprobe->uprobe;
784 case HPROBE_GONE:
785 /*
786 * SRCU was unlocked earlier and we didn't manage to take
787 * uprobe refcnt, so it's effectively NULL
788 */
789 return NULL;
790 case HPROBE_CONSUMED:
791 /*
792 * uprobe was consumed, so it's effectively NULL as far as
793 * uretprobe processing logic is concerned
794 */
795 return NULL;
796 case HPROBE_LEASED: {
797 struct uprobe *uprobe = try_get_uprobe(hprobe->uprobe);
798 /*
799 * Try to switch hprobe state, guarding against
800 * hprobe_consume() or another hprobe_expire() racing with us.
801 * Note, if we failed to get uprobe refcount, we use special
802 * HPROBE_GONE state to signal that hprobe->uprobe shouldn't
803 * be used as it will be freed after SRCU is unlocked.
804 */
805 if (try_cmpxchg(&hprobe->state, &hstate, uprobe ? HPROBE_STABLE : HPROBE_GONE)) {
806 /* We won the race, we are the ones to unlock SRCU */
807 __srcu_read_unlock(&uretprobes_srcu, hprobe->srcu_idx);
808 return get ? get_uprobe(uprobe) : uprobe;
809 }
810
811 /*
812 * We lost the race, undo refcount bump (if it ever happened),
813 * unless caller would like an extra refcount anyways.
814 */
815 if (uprobe && !get)
816 put_uprobe(uprobe);
817 /*
818 * Even if hprobe_consume() or another hprobe_expire() wins
819 * the state update race and unlocks SRCU from under us, we
820 * still have a guarantee that underyling uprobe won't be
821 * freed due to ongoing caller's SRCU lock region, so we can
822 * return it regardless. Also, if `get` was true, we also have
823 * an extra ref for the caller to own. This is used in dup_utask().
824 */
825 return uprobe;
826 }
827 default:
828 WARN(1, "unknown hprobe state %d", hstate);
829 return NULL;
830 }
831}
832
833static __always_inline
834int uprobe_cmp(const struct inode *l_inode, const loff_t l_offset,
835 const struct uprobe *r)
836{
837 if (l_inode < r->inode)
838 return -1;
839
840 if (l_inode > r->inode)
841 return 1;
842
843 if (l_offset < r->offset)
844 return -1;
845
846 if (l_offset > r->offset)
847 return 1;
848
849 return 0;
850}
851
852#define __node_2_uprobe(node) \
853 rb_entry((node), struct uprobe, rb_node)
854
855struct __uprobe_key {
856 struct inode *inode;
857 loff_t offset;
858};
859
860static inline int __uprobe_cmp_key(const void *key, const struct rb_node *b)
861{
862 const struct __uprobe_key *a = key;
863 return uprobe_cmp(a->inode, a->offset, __node_2_uprobe(b));
864}
865
866static inline int __uprobe_cmp(struct rb_node *a, const struct rb_node *b)
867{
868 struct uprobe *u = __node_2_uprobe(a);
869 return uprobe_cmp(u->inode, u->offset, __node_2_uprobe(b));
870}
871
872/*
873 * Assumes being inside RCU protected region.
874 * No refcount is taken on returned uprobe.
875 */
876static struct uprobe *find_uprobe_rcu(struct inode *inode, loff_t offset)
877{
878 struct __uprobe_key key = {
879 .inode = inode,
880 .offset = offset,
881 };
882 struct rb_node *node;
883 unsigned int seq;
884
885 lockdep_assert(rcu_read_lock_trace_held());
886
887 do {
888 seq = read_seqcount_begin(&uprobes_seqcount);
889 node = rb_find_rcu(&key, &uprobes_tree, __uprobe_cmp_key);
890 /*
891 * Lockless RB-tree lookups can result only in false negatives.
892 * If the element is found, it is correct and can be returned
893 * under RCU protection. If we find nothing, we need to
894 * validate that seqcount didn't change. If it did, we have to
895 * try again as we might have missed the element (false
896 * negative). If seqcount is unchanged, search truly failed.
897 */
898 if (node)
899 return __node_2_uprobe(node);
900 } while (read_seqcount_retry(&uprobes_seqcount, seq));
901
902 return NULL;
903}
904
905/*
906 * Attempt to insert a new uprobe into uprobes_tree.
907 *
908 * If uprobe already exists (for given inode+offset), we just increment
909 * refcount of previously existing uprobe.
910 *
911 * If not, a provided new instance of uprobe is inserted into the tree (with
912 * assumed initial refcount == 1).
913 *
914 * In any case, we return a uprobe instance that ends up being in uprobes_tree.
915 * Caller has to clean up new uprobe instance, if it ended up not being
916 * inserted into the tree.
917 *
918 * We assume that uprobes_treelock is held for writing.
919 */
920static struct uprobe *__insert_uprobe(struct uprobe *uprobe)
921{
922 struct rb_node *node;
923again:
924 node = rb_find_add_rcu(&uprobe->rb_node, &uprobes_tree, __uprobe_cmp);
925 if (node) {
926 struct uprobe *u = __node_2_uprobe(node);
927
928 if (!try_get_uprobe(u)) {
929 rb_erase(node, &uprobes_tree);
930 RB_CLEAR_NODE(&u->rb_node);
931 goto again;
932 }
933
934 return u;
935 }
936
937 return uprobe;
938}
939
940/*
941 * Acquire uprobes_treelock and insert uprobe into uprobes_tree
942 * (or reuse existing one, see __insert_uprobe() comments above).
943 */
944static struct uprobe *insert_uprobe(struct uprobe *uprobe)
945{
946 struct uprobe *u;
947
948 write_lock(&uprobes_treelock);
949 write_seqcount_begin(&uprobes_seqcount);
950 u = __insert_uprobe(uprobe);
951 write_seqcount_end(&uprobes_seqcount);
952 write_unlock(&uprobes_treelock);
953
954 return u;
955}
956
957static void
958ref_ctr_mismatch_warn(struct uprobe *cur_uprobe, struct uprobe *uprobe)
959{
960 pr_warn("ref_ctr_offset mismatch. inode: 0x%lx offset: 0x%llx "
961 "ref_ctr_offset(old): 0x%llx ref_ctr_offset(new): 0x%llx\n",
962 uprobe->inode->i_ino, (unsigned long long) uprobe->offset,
963 (unsigned long long) cur_uprobe->ref_ctr_offset,
964 (unsigned long long) uprobe->ref_ctr_offset);
965}
966
967static struct uprobe *alloc_uprobe(struct inode *inode, loff_t offset,
968 loff_t ref_ctr_offset)
969{
970 struct uprobe *uprobe, *cur_uprobe;
971
972 uprobe = kzalloc(sizeof(struct uprobe), GFP_KERNEL);
973 if (!uprobe)
974 return ERR_PTR(-ENOMEM);
975
976 uprobe->inode = inode;
977 uprobe->offset = offset;
978 uprobe->ref_ctr_offset = ref_ctr_offset;
979 INIT_LIST_HEAD(&uprobe->consumers);
980 init_rwsem(&uprobe->register_rwsem);
981 init_rwsem(&uprobe->consumer_rwsem);
982 RB_CLEAR_NODE(&uprobe->rb_node);
983 refcount_set(&uprobe->ref, 1);
984
985 /* add to uprobes_tree, sorted on inode:offset */
986 cur_uprobe = insert_uprobe(uprobe);
987 /* a uprobe exists for this inode:offset combination */
988 if (cur_uprobe != uprobe) {
989 if (cur_uprobe->ref_ctr_offset != uprobe->ref_ctr_offset) {
990 ref_ctr_mismatch_warn(cur_uprobe, uprobe);
991 put_uprobe(cur_uprobe);
992 kfree(uprobe);
993 return ERR_PTR(-EINVAL);
994 }
995 kfree(uprobe);
996 uprobe = cur_uprobe;
997 }
998
999 return uprobe;
1000}
1001
1002static void consumer_add(struct uprobe *uprobe, struct uprobe_consumer *uc)
1003{
1004 static atomic64_t id;
1005
1006 down_write(&uprobe->consumer_rwsem);
1007 list_add_rcu(&uc->cons_node, &uprobe->consumers);
1008 uc->id = (__u64) atomic64_inc_return(&id);
1009 up_write(&uprobe->consumer_rwsem);
1010}
1011
1012/*
1013 * For uprobe @uprobe, delete the consumer @uc.
1014 * Should never be called with consumer that's not part of @uprobe->consumers.
1015 */
1016static void consumer_del(struct uprobe *uprobe, struct uprobe_consumer *uc)
1017{
1018 down_write(&uprobe->consumer_rwsem);
1019 list_del_rcu(&uc->cons_node);
1020 up_write(&uprobe->consumer_rwsem);
1021}
1022
1023static int __copy_insn(struct address_space *mapping, struct file *filp,
1024 void *insn, int nbytes, loff_t offset)
1025{
1026 struct page *page;
1027 /*
1028 * Ensure that the page that has the original instruction is populated
1029 * and in page-cache. If ->read_folio == NULL it must be shmem_mapping(),
1030 * see uprobe_register().
1031 */
1032 if (mapping->a_ops->read_folio)
1033 page = read_mapping_page(mapping, offset >> PAGE_SHIFT, filp);
1034 else
1035 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
1036 if (IS_ERR(page))
1037 return PTR_ERR(page);
1038
1039 copy_from_page(page, offset, insn, nbytes);
1040 put_page(page);
1041
1042 return 0;
1043}
1044
1045static int copy_insn(struct uprobe *uprobe, struct file *filp)
1046{
1047 struct address_space *mapping = uprobe->inode->i_mapping;
1048 loff_t offs = uprobe->offset;
1049 void *insn = &uprobe->arch.insn;
1050 int size = sizeof(uprobe->arch.insn);
1051 int len, err = -EIO;
1052
1053 /* Copy only available bytes, -EIO if nothing was read */
1054 do {
1055 if (offs >= i_size_read(uprobe->inode))
1056 break;
1057
1058 len = min_t(int, size, PAGE_SIZE - (offs & ~PAGE_MASK));
1059 err = __copy_insn(mapping, filp, insn, len, offs);
1060 if (err)
1061 break;
1062
1063 insn += len;
1064 offs += len;
1065 size -= len;
1066 } while (size);
1067
1068 return err;
1069}
1070
1071static int prepare_uprobe(struct uprobe *uprobe, struct file *file,
1072 struct mm_struct *mm, unsigned long vaddr)
1073{
1074 int ret = 0;
1075
1076 if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
1077 return ret;
1078
1079 /* TODO: move this into _register, until then we abuse this sem. */
1080 down_write(&uprobe->consumer_rwsem);
1081 if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
1082 goto out;
1083
1084 ret = copy_insn(uprobe, file);
1085 if (ret)
1086 goto out;
1087
1088 ret = -ENOTSUPP;
1089 if (is_trap_insn((uprobe_opcode_t *)&uprobe->arch.insn))
1090 goto out;
1091
1092 ret = arch_uprobe_analyze_insn(&uprobe->arch, mm, vaddr);
1093 if (ret)
1094 goto out;
1095
1096 smp_wmb(); /* pairs with the smp_rmb() in handle_swbp() */
1097 set_bit(UPROBE_COPY_INSN, &uprobe->flags);
1098
1099 out:
1100 up_write(&uprobe->consumer_rwsem);
1101
1102 return ret;
1103}
1104
1105static inline bool consumer_filter(struct uprobe_consumer *uc, struct mm_struct *mm)
1106{
1107 return !uc->filter || uc->filter(uc, mm);
1108}
1109
1110static bool filter_chain(struct uprobe *uprobe, struct mm_struct *mm)
1111{
1112 struct uprobe_consumer *uc;
1113 bool ret = false;
1114
1115 down_read(&uprobe->consumer_rwsem);
1116 list_for_each_entry_rcu(uc, &uprobe->consumers, cons_node, rcu_read_lock_trace_held()) {
1117 ret = consumer_filter(uc, mm);
1118 if (ret)
1119 break;
1120 }
1121 up_read(&uprobe->consumer_rwsem);
1122
1123 return ret;
1124}
1125
1126static int
1127install_breakpoint(struct uprobe *uprobe, struct mm_struct *mm,
1128 struct vm_area_struct *vma, unsigned long vaddr)
1129{
1130 bool first_uprobe;
1131 int ret;
1132
1133 ret = prepare_uprobe(uprobe, vma->vm_file, mm, vaddr);
1134 if (ret)
1135 return ret;
1136
1137 /*
1138 * set MMF_HAS_UPROBES in advance for uprobe_pre_sstep_notifier(),
1139 * the task can hit this breakpoint right after __replace_page().
1140 */
1141 first_uprobe = !test_bit(MMF_HAS_UPROBES, &mm->flags);
1142 if (first_uprobe)
1143 set_bit(MMF_HAS_UPROBES, &mm->flags);
1144
1145 ret = set_swbp(&uprobe->arch, mm, vaddr);
1146 if (!ret)
1147 clear_bit(MMF_RECALC_UPROBES, &mm->flags);
1148 else if (first_uprobe)
1149 clear_bit(MMF_HAS_UPROBES, &mm->flags);
1150
1151 return ret;
1152}
1153
1154static int
1155remove_breakpoint(struct uprobe *uprobe, struct mm_struct *mm, unsigned long vaddr)
1156{
1157 set_bit(MMF_RECALC_UPROBES, &mm->flags);
1158 return set_orig_insn(&uprobe->arch, mm, vaddr);
1159}
1160
1161struct map_info {
1162 struct map_info *next;
1163 struct mm_struct *mm;
1164 unsigned long vaddr;
1165};
1166
1167static inline struct map_info *free_map_info(struct map_info *info)
1168{
1169 struct map_info *next = info->next;
1170 kfree(info);
1171 return next;
1172}
1173
1174static struct map_info *
1175build_map_info(struct address_space *mapping, loff_t offset, bool is_register)
1176{
1177 unsigned long pgoff = offset >> PAGE_SHIFT;
1178 struct vm_area_struct *vma;
1179 struct map_info *curr = NULL;
1180 struct map_info *prev = NULL;
1181 struct map_info *info;
1182 int more = 0;
1183
1184 again:
1185 i_mmap_lock_read(mapping);
1186 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1187 if (!valid_vma(vma, is_register))
1188 continue;
1189
1190 if (!prev && !more) {
1191 /*
1192 * Needs GFP_NOWAIT to avoid i_mmap_rwsem recursion through
1193 * reclaim. This is optimistic, no harm done if it fails.
1194 */
1195 prev = kmalloc(sizeof(struct map_info),
1196 GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
1197 if (prev)
1198 prev->next = NULL;
1199 }
1200 if (!prev) {
1201 more++;
1202 continue;
1203 }
1204
1205 if (!mmget_not_zero(vma->vm_mm))
1206 continue;
1207
1208 info = prev;
1209 prev = prev->next;
1210 info->next = curr;
1211 curr = info;
1212
1213 info->mm = vma->vm_mm;
1214 info->vaddr = offset_to_vaddr(vma, offset);
1215 }
1216 i_mmap_unlock_read(mapping);
1217
1218 if (!more)
1219 goto out;
1220
1221 prev = curr;
1222 while (curr) {
1223 mmput(curr->mm);
1224 curr = curr->next;
1225 }
1226
1227 do {
1228 info = kmalloc(sizeof(struct map_info), GFP_KERNEL);
1229 if (!info) {
1230 curr = ERR_PTR(-ENOMEM);
1231 goto out;
1232 }
1233 info->next = prev;
1234 prev = info;
1235 } while (--more);
1236
1237 goto again;
1238 out:
1239 while (prev)
1240 prev = free_map_info(prev);
1241 return curr;
1242}
1243
1244static int
1245register_for_each_vma(struct uprobe *uprobe, struct uprobe_consumer *new)
1246{
1247 bool is_register = !!new;
1248 struct map_info *info;
1249 int err = 0;
1250
1251 percpu_down_write(&dup_mmap_sem);
1252 info = build_map_info(uprobe->inode->i_mapping,
1253 uprobe->offset, is_register);
1254 if (IS_ERR(info)) {
1255 err = PTR_ERR(info);
1256 goto out;
1257 }
1258
1259 while (info) {
1260 struct mm_struct *mm = info->mm;
1261 struct vm_area_struct *vma;
1262
1263 if (err && is_register)
1264 goto free;
1265 /*
1266 * We take mmap_lock for writing to avoid the race with
1267 * find_active_uprobe_rcu() which takes mmap_lock for reading.
1268 * Thus this install_breakpoint() can not make
1269 * is_trap_at_addr() true right after find_uprobe_rcu()
1270 * returns NULL in find_active_uprobe_rcu().
1271 */
1272 mmap_write_lock(mm);
1273 if (check_stable_address_space(mm))
1274 goto unlock;
1275
1276 vma = find_vma(mm, info->vaddr);
1277 if (!vma || !valid_vma(vma, is_register) ||
1278 file_inode(vma->vm_file) != uprobe->inode)
1279 goto unlock;
1280
1281 if (vma->vm_start > info->vaddr ||
1282 vaddr_to_offset(vma, info->vaddr) != uprobe->offset)
1283 goto unlock;
1284
1285 if (is_register) {
1286 /* consult only the "caller", new consumer. */
1287 if (consumer_filter(new, mm))
1288 err = install_breakpoint(uprobe, mm, vma, info->vaddr);
1289 } else if (test_bit(MMF_HAS_UPROBES, &mm->flags)) {
1290 if (!filter_chain(uprobe, mm))
1291 err |= remove_breakpoint(uprobe, mm, info->vaddr);
1292 }
1293
1294 unlock:
1295 mmap_write_unlock(mm);
1296 free:
1297 mmput(mm);
1298 info = free_map_info(info);
1299 }
1300 out:
1301 percpu_up_write(&dup_mmap_sem);
1302 return err;
1303}
1304
1305/**
1306 * uprobe_unregister_nosync - unregister an already registered probe.
1307 * @uprobe: uprobe to remove
1308 * @uc: identify which probe if multiple probes are colocated.
1309 */
1310void uprobe_unregister_nosync(struct uprobe *uprobe, struct uprobe_consumer *uc)
1311{
1312 int err;
1313
1314 down_write(&uprobe->register_rwsem);
1315 consumer_del(uprobe, uc);
1316 err = register_for_each_vma(uprobe, NULL);
1317 up_write(&uprobe->register_rwsem);
1318
1319 /* TODO : cant unregister? schedule a worker thread */
1320 if (unlikely(err)) {
1321 uprobe_warn(current, "unregister, leaking uprobe");
1322 return;
1323 }
1324
1325 put_uprobe(uprobe);
1326}
1327EXPORT_SYMBOL_GPL(uprobe_unregister_nosync);
1328
1329void uprobe_unregister_sync(void)
1330{
1331 /*
1332 * Now that handler_chain() and handle_uretprobe_chain() iterate over
1333 * uprobe->consumers list under RCU protection without holding
1334 * uprobe->register_rwsem, we need to wait for RCU grace period to
1335 * make sure that we can't call into just unregistered
1336 * uprobe_consumer's callbacks anymore. If we don't do that, fast and
1337 * unlucky enough caller can free consumer's memory and cause
1338 * handler_chain() or handle_uretprobe_chain() to do an use-after-free.
1339 */
1340 synchronize_rcu_tasks_trace();
1341 synchronize_srcu(&uretprobes_srcu);
1342}
1343EXPORT_SYMBOL_GPL(uprobe_unregister_sync);
1344
1345/**
1346 * uprobe_register - register a probe
1347 * @inode: the file in which the probe has to be placed.
1348 * @offset: offset from the start of the file.
1349 * @ref_ctr_offset: offset of SDT marker / reference counter
1350 * @uc: information on howto handle the probe..
1351 *
1352 * Apart from the access refcount, uprobe_register() takes a creation
1353 * refcount (thro alloc_uprobe) if and only if this @uprobe is getting
1354 * inserted into the rbtree (i.e first consumer for a @inode:@offset
1355 * tuple). Creation refcount stops uprobe_unregister from freeing the
1356 * @uprobe even before the register operation is complete. Creation
1357 * refcount is released when the last @uc for the @uprobe
1358 * unregisters. Caller of uprobe_register() is required to keep @inode
1359 * (and the containing mount) referenced.
1360 *
1361 * Return: pointer to the new uprobe on success or an ERR_PTR on failure.
1362 */
1363struct uprobe *uprobe_register(struct inode *inode,
1364 loff_t offset, loff_t ref_ctr_offset,
1365 struct uprobe_consumer *uc)
1366{
1367 struct uprobe *uprobe;
1368 int ret;
1369
1370 /* Uprobe must have at least one set consumer */
1371 if (!uc->handler && !uc->ret_handler)
1372 return ERR_PTR(-EINVAL);
1373
1374 /* copy_insn() uses read_mapping_page() or shmem_read_mapping_page() */
1375 if (!inode->i_mapping->a_ops->read_folio &&
1376 !shmem_mapping(inode->i_mapping))
1377 return ERR_PTR(-EIO);
1378 /* Racy, just to catch the obvious mistakes */
1379 if (offset > i_size_read(inode))
1380 return ERR_PTR(-EINVAL);
1381
1382 /*
1383 * This ensures that copy_from_page(), copy_to_page() and
1384 * __update_ref_ctr() can't cross page boundary.
1385 */
1386 if (!IS_ALIGNED(offset, UPROBE_SWBP_INSN_SIZE))
1387 return ERR_PTR(-EINVAL);
1388 if (!IS_ALIGNED(ref_ctr_offset, sizeof(short)))
1389 return ERR_PTR(-EINVAL);
1390
1391 uprobe = alloc_uprobe(inode, offset, ref_ctr_offset);
1392 if (IS_ERR(uprobe))
1393 return uprobe;
1394
1395 down_write(&uprobe->register_rwsem);
1396 consumer_add(uprobe, uc);
1397 ret = register_for_each_vma(uprobe, uc);
1398 up_write(&uprobe->register_rwsem);
1399
1400 if (ret) {
1401 uprobe_unregister_nosync(uprobe, uc);
1402 /*
1403 * Registration might have partially succeeded, so we can have
1404 * this consumer being called right at this time. We need to
1405 * sync here. It's ok, it's unlikely slow path.
1406 */
1407 uprobe_unregister_sync();
1408 return ERR_PTR(ret);
1409 }
1410
1411 return uprobe;
1412}
1413EXPORT_SYMBOL_GPL(uprobe_register);
1414
1415/**
1416 * uprobe_apply - add or remove the breakpoints according to @uc->filter
1417 * @uprobe: uprobe which "owns" the breakpoint
1418 * @uc: consumer which wants to add more or remove some breakpoints
1419 * @add: add or remove the breakpoints
1420 * Return: 0 on success or negative error code.
1421 */
1422int uprobe_apply(struct uprobe *uprobe, struct uprobe_consumer *uc, bool add)
1423{
1424 struct uprobe_consumer *con;
1425 int ret = -ENOENT;
1426
1427 down_write(&uprobe->register_rwsem);
1428
1429 rcu_read_lock_trace();
1430 list_for_each_entry_rcu(con, &uprobe->consumers, cons_node, rcu_read_lock_trace_held()) {
1431 if (con == uc) {
1432 ret = register_for_each_vma(uprobe, add ? uc : NULL);
1433 break;
1434 }
1435 }
1436 rcu_read_unlock_trace();
1437
1438 up_write(&uprobe->register_rwsem);
1439
1440 return ret;
1441}
1442
1443static int unapply_uprobe(struct uprobe *uprobe, struct mm_struct *mm)
1444{
1445 VMA_ITERATOR(vmi, mm, 0);
1446 struct vm_area_struct *vma;
1447 int err = 0;
1448
1449 mmap_read_lock(mm);
1450 for_each_vma(vmi, vma) {
1451 unsigned long vaddr;
1452 loff_t offset;
1453
1454 if (!valid_vma(vma, false) ||
1455 file_inode(vma->vm_file) != uprobe->inode)
1456 continue;
1457
1458 offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1459 if (uprobe->offset < offset ||
1460 uprobe->offset >= offset + vma->vm_end - vma->vm_start)
1461 continue;
1462
1463 vaddr = offset_to_vaddr(vma, uprobe->offset);
1464 err |= remove_breakpoint(uprobe, mm, vaddr);
1465 }
1466 mmap_read_unlock(mm);
1467
1468 return err;
1469}
1470
1471static struct rb_node *
1472find_node_in_range(struct inode *inode, loff_t min, loff_t max)
1473{
1474 struct rb_node *n = uprobes_tree.rb_node;
1475
1476 while (n) {
1477 struct uprobe *u = rb_entry(n, struct uprobe, rb_node);
1478
1479 if (inode < u->inode) {
1480 n = n->rb_left;
1481 } else if (inode > u->inode) {
1482 n = n->rb_right;
1483 } else {
1484 if (max < u->offset)
1485 n = n->rb_left;
1486 else if (min > u->offset)
1487 n = n->rb_right;
1488 else
1489 break;
1490 }
1491 }
1492
1493 return n;
1494}
1495
1496/*
1497 * For a given range in vma, build a list of probes that need to be inserted.
1498 */
1499static void build_probe_list(struct inode *inode,
1500 struct vm_area_struct *vma,
1501 unsigned long start, unsigned long end,
1502 struct list_head *head)
1503{
1504 loff_t min, max;
1505 struct rb_node *n, *t;
1506 struct uprobe *u;
1507
1508 INIT_LIST_HEAD(head);
1509 min = vaddr_to_offset(vma, start);
1510 max = min + (end - start) - 1;
1511
1512 read_lock(&uprobes_treelock);
1513 n = find_node_in_range(inode, min, max);
1514 if (n) {
1515 for (t = n; t; t = rb_prev(t)) {
1516 u = rb_entry(t, struct uprobe, rb_node);
1517 if (u->inode != inode || u->offset < min)
1518 break;
1519 /* if uprobe went away, it's safe to ignore it */
1520 if (try_get_uprobe(u))
1521 list_add(&u->pending_list, head);
1522 }
1523 for (t = n; (t = rb_next(t)); ) {
1524 u = rb_entry(t, struct uprobe, rb_node);
1525 if (u->inode != inode || u->offset > max)
1526 break;
1527 /* if uprobe went away, it's safe to ignore it */
1528 if (try_get_uprobe(u))
1529 list_add(&u->pending_list, head);
1530 }
1531 }
1532 read_unlock(&uprobes_treelock);
1533}
1534
1535/* @vma contains reference counter, not the probed instruction. */
1536static int delayed_ref_ctr_inc(struct vm_area_struct *vma)
1537{
1538 struct list_head *pos, *q;
1539 struct delayed_uprobe *du;
1540 unsigned long vaddr;
1541 int ret = 0, err = 0;
1542
1543 mutex_lock(&delayed_uprobe_lock);
1544 list_for_each_safe(pos, q, &delayed_uprobe_list) {
1545 du = list_entry(pos, struct delayed_uprobe, list);
1546
1547 if (du->mm != vma->vm_mm ||
1548 !valid_ref_ctr_vma(du->uprobe, vma))
1549 continue;
1550
1551 vaddr = offset_to_vaddr(vma, du->uprobe->ref_ctr_offset);
1552 ret = __update_ref_ctr(vma->vm_mm, vaddr, 1);
1553 if (ret) {
1554 update_ref_ctr_warn(du->uprobe, vma->vm_mm, 1);
1555 if (!err)
1556 err = ret;
1557 }
1558 delayed_uprobe_delete(du);
1559 }
1560 mutex_unlock(&delayed_uprobe_lock);
1561 return err;
1562}
1563
1564/*
1565 * Called from mmap_region/vma_merge with mm->mmap_lock acquired.
1566 *
1567 * Currently we ignore all errors and always return 0, the callers
1568 * can't handle the failure anyway.
1569 */
1570int uprobe_mmap(struct vm_area_struct *vma)
1571{
1572 struct list_head tmp_list;
1573 struct uprobe *uprobe, *u;
1574 struct inode *inode;
1575
1576 if (no_uprobe_events())
1577 return 0;
1578
1579 if (vma->vm_file &&
1580 (vma->vm_flags & (VM_WRITE|VM_SHARED)) == VM_WRITE &&
1581 test_bit(MMF_HAS_UPROBES, &vma->vm_mm->flags))
1582 delayed_ref_ctr_inc(vma);
1583
1584 if (!valid_vma(vma, true))
1585 return 0;
1586
1587 inode = file_inode(vma->vm_file);
1588 if (!inode)
1589 return 0;
1590
1591 mutex_lock(uprobes_mmap_hash(inode));
1592 build_probe_list(inode, vma, vma->vm_start, vma->vm_end, &tmp_list);
1593 /*
1594 * We can race with uprobe_unregister(), this uprobe can be already
1595 * removed. But in this case filter_chain() must return false, all
1596 * consumers have gone away.
1597 */
1598 list_for_each_entry_safe(uprobe, u, &tmp_list, pending_list) {
1599 if (!fatal_signal_pending(current) &&
1600 filter_chain(uprobe, vma->vm_mm)) {
1601 unsigned long vaddr = offset_to_vaddr(vma, uprobe->offset);
1602 install_breakpoint(uprobe, vma->vm_mm, vma, vaddr);
1603 }
1604 put_uprobe(uprobe);
1605 }
1606 mutex_unlock(uprobes_mmap_hash(inode));
1607
1608 return 0;
1609}
1610
1611static bool
1612vma_has_uprobes(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1613{
1614 loff_t min, max;
1615 struct inode *inode;
1616 struct rb_node *n;
1617
1618 inode = file_inode(vma->vm_file);
1619
1620 min = vaddr_to_offset(vma, start);
1621 max = min + (end - start) - 1;
1622
1623 read_lock(&uprobes_treelock);
1624 n = find_node_in_range(inode, min, max);
1625 read_unlock(&uprobes_treelock);
1626
1627 return !!n;
1628}
1629
1630/*
1631 * Called in context of a munmap of a vma.
1632 */
1633void uprobe_munmap(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1634{
1635 if (no_uprobe_events() || !valid_vma(vma, false))
1636 return;
1637
1638 if (!atomic_read(&vma->vm_mm->mm_users)) /* called by mmput() ? */
1639 return;
1640
1641 if (!test_bit(MMF_HAS_UPROBES, &vma->vm_mm->flags) ||
1642 test_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags))
1643 return;
1644
1645 if (vma_has_uprobes(vma, start, end))
1646 set_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags);
1647}
1648
1649static vm_fault_t xol_fault(const struct vm_special_mapping *sm,
1650 struct vm_area_struct *vma, struct vm_fault *vmf)
1651{
1652 struct xol_area *area = vma->vm_mm->uprobes_state.xol_area;
1653
1654 vmf->page = area->page;
1655 get_page(vmf->page);
1656 return 0;
1657}
1658
1659static int xol_mremap(const struct vm_special_mapping *sm, struct vm_area_struct *new_vma)
1660{
1661 return -EPERM;
1662}
1663
1664static const struct vm_special_mapping xol_mapping = {
1665 .name = "[uprobes]",
1666 .fault = xol_fault,
1667 .mremap = xol_mremap,
1668};
1669
1670/* Slot allocation for XOL */
1671static int xol_add_vma(struct mm_struct *mm, struct xol_area *area)
1672{
1673 struct vm_area_struct *vma;
1674 int ret;
1675
1676 if (mmap_write_lock_killable(mm))
1677 return -EINTR;
1678
1679 if (mm->uprobes_state.xol_area) {
1680 ret = -EALREADY;
1681 goto fail;
1682 }
1683
1684 if (!area->vaddr) {
1685 /* Try to map as high as possible, this is only a hint. */
1686 area->vaddr = get_unmapped_area(NULL, TASK_SIZE - PAGE_SIZE,
1687 PAGE_SIZE, 0, 0);
1688 if (IS_ERR_VALUE(area->vaddr)) {
1689 ret = area->vaddr;
1690 goto fail;
1691 }
1692 }
1693
1694 vma = _install_special_mapping(mm, area->vaddr, PAGE_SIZE,
1695 VM_EXEC|VM_MAYEXEC|VM_DONTCOPY|VM_IO,
1696 &xol_mapping);
1697 if (IS_ERR(vma)) {
1698 ret = PTR_ERR(vma);
1699 goto fail;
1700 }
1701
1702 ret = 0;
1703 /* pairs with get_xol_area() */
1704 smp_store_release(&mm->uprobes_state.xol_area, area); /* ^^^ */
1705 fail:
1706 mmap_write_unlock(mm);
1707
1708 return ret;
1709}
1710
1711void * __weak arch_uprobe_trampoline(unsigned long *psize)
1712{
1713 static uprobe_opcode_t insn = UPROBE_SWBP_INSN;
1714
1715 *psize = UPROBE_SWBP_INSN_SIZE;
1716 return &insn;
1717}
1718
1719static struct xol_area *__create_xol_area(unsigned long vaddr)
1720{
1721 struct mm_struct *mm = current->mm;
1722 unsigned long insns_size;
1723 struct xol_area *area;
1724 void *insns;
1725
1726 area = kzalloc(sizeof(*area), GFP_KERNEL);
1727 if (unlikely(!area))
1728 goto out;
1729
1730 area->bitmap = kcalloc(BITS_TO_LONGS(UINSNS_PER_PAGE), sizeof(long),
1731 GFP_KERNEL);
1732 if (!area->bitmap)
1733 goto free_area;
1734
1735 area->page = alloc_page(GFP_HIGHUSER | __GFP_ZERO);
1736 if (!area->page)
1737 goto free_bitmap;
1738
1739 area->vaddr = vaddr;
1740 init_waitqueue_head(&area->wq);
1741 /* Reserve the 1st slot for get_trampoline_vaddr() */
1742 set_bit(0, area->bitmap);
1743 insns = arch_uprobe_trampoline(&insns_size);
1744 arch_uprobe_copy_ixol(area->page, 0, insns, insns_size);
1745
1746 if (!xol_add_vma(mm, area))
1747 return area;
1748
1749 __free_page(area->page);
1750 free_bitmap:
1751 kfree(area->bitmap);
1752 free_area:
1753 kfree(area);
1754 out:
1755 return NULL;
1756}
1757
1758/*
1759 * get_xol_area - Allocate process's xol_area if necessary.
1760 * This area will be used for storing instructions for execution out of line.
1761 *
1762 * Returns the allocated area or NULL.
1763 */
1764static struct xol_area *get_xol_area(void)
1765{
1766 struct mm_struct *mm = current->mm;
1767 struct xol_area *area;
1768
1769 if (!mm->uprobes_state.xol_area)
1770 __create_xol_area(0);
1771
1772 /* Pairs with xol_add_vma() smp_store_release() */
1773 area = READ_ONCE(mm->uprobes_state.xol_area); /* ^^^ */
1774 return area;
1775}
1776
1777/*
1778 * uprobe_clear_state - Free the area allocated for slots.
1779 */
1780void uprobe_clear_state(struct mm_struct *mm)
1781{
1782 struct xol_area *area = mm->uprobes_state.xol_area;
1783
1784 mutex_lock(&delayed_uprobe_lock);
1785 delayed_uprobe_remove(NULL, mm);
1786 mutex_unlock(&delayed_uprobe_lock);
1787
1788 if (!area)
1789 return;
1790
1791 put_page(area->page);
1792 kfree(area->bitmap);
1793 kfree(area);
1794}
1795
1796void uprobe_start_dup_mmap(void)
1797{
1798 percpu_down_read(&dup_mmap_sem);
1799}
1800
1801void uprobe_end_dup_mmap(void)
1802{
1803 percpu_up_read(&dup_mmap_sem);
1804}
1805
1806void uprobe_dup_mmap(struct mm_struct *oldmm, struct mm_struct *newmm)
1807{
1808 if (test_bit(MMF_HAS_UPROBES, &oldmm->flags)) {
1809 set_bit(MMF_HAS_UPROBES, &newmm->flags);
1810 /* unconditionally, dup_mmap() skips VM_DONTCOPY vmas */
1811 set_bit(MMF_RECALC_UPROBES, &newmm->flags);
1812 }
1813}
1814
1815static unsigned long xol_get_slot_nr(struct xol_area *area)
1816{
1817 unsigned long slot_nr;
1818
1819 slot_nr = find_first_zero_bit(area->bitmap, UINSNS_PER_PAGE);
1820 if (slot_nr < UINSNS_PER_PAGE) {
1821 if (!test_and_set_bit(slot_nr, area->bitmap))
1822 return slot_nr;
1823 }
1824
1825 return UINSNS_PER_PAGE;
1826}
1827
1828/*
1829 * xol_get_insn_slot - allocate a slot for xol.
1830 */
1831static bool xol_get_insn_slot(struct uprobe *uprobe, struct uprobe_task *utask)
1832{
1833 struct xol_area *area = get_xol_area();
1834 unsigned long slot_nr;
1835
1836 if (!area)
1837 return false;
1838
1839 wait_event(area->wq, (slot_nr = xol_get_slot_nr(area)) < UINSNS_PER_PAGE);
1840
1841 utask->xol_vaddr = area->vaddr + slot_nr * UPROBE_XOL_SLOT_BYTES;
1842 arch_uprobe_copy_ixol(area->page, utask->xol_vaddr,
1843 &uprobe->arch.ixol, sizeof(uprobe->arch.ixol));
1844 return true;
1845}
1846
1847/*
1848 * xol_free_insn_slot - free the slot allocated by xol_get_insn_slot()
1849 */
1850static void xol_free_insn_slot(struct uprobe_task *utask)
1851{
1852 struct xol_area *area = current->mm->uprobes_state.xol_area;
1853 unsigned long offset = utask->xol_vaddr - area->vaddr;
1854 unsigned int slot_nr;
1855
1856 utask->xol_vaddr = 0;
1857 /* xol_vaddr must fit into [area->vaddr, area->vaddr + PAGE_SIZE) */
1858 if (WARN_ON_ONCE(offset >= PAGE_SIZE))
1859 return;
1860
1861 slot_nr = offset / UPROBE_XOL_SLOT_BYTES;
1862 clear_bit(slot_nr, area->bitmap);
1863 smp_mb__after_atomic(); /* pairs with prepare_to_wait() */
1864 if (waitqueue_active(&area->wq))
1865 wake_up(&area->wq);
1866}
1867
1868void __weak arch_uprobe_copy_ixol(struct page *page, unsigned long vaddr,
1869 void *src, unsigned long len)
1870{
1871 /* Initialize the slot */
1872 copy_to_page(page, vaddr, src, len);
1873
1874 /*
1875 * We probably need flush_icache_user_page() but it needs vma.
1876 * This should work on most of architectures by default. If
1877 * architecture needs to do something different it can define
1878 * its own version of the function.
1879 */
1880 flush_dcache_page(page);
1881}
1882
1883/**
1884 * uprobe_get_swbp_addr - compute address of swbp given post-swbp regs
1885 * @regs: Reflects the saved state of the task after it has hit a breakpoint
1886 * instruction.
1887 * Return the address of the breakpoint instruction.
1888 */
1889unsigned long __weak uprobe_get_swbp_addr(struct pt_regs *regs)
1890{
1891 return instruction_pointer(regs) - UPROBE_SWBP_INSN_SIZE;
1892}
1893
1894unsigned long uprobe_get_trap_addr(struct pt_regs *regs)
1895{
1896 struct uprobe_task *utask = current->utask;
1897
1898 if (unlikely(utask && utask->active_uprobe))
1899 return utask->vaddr;
1900
1901 return instruction_pointer(regs);
1902}
1903
1904static struct return_instance *free_ret_instance(struct return_instance *ri, bool cleanup_hprobe)
1905{
1906 struct return_instance *next = ri->next;
1907
1908 if (cleanup_hprobe) {
1909 enum hprobe_state hstate;
1910
1911 (void)hprobe_consume(&ri->hprobe, &hstate);
1912 hprobe_finalize(&ri->hprobe, hstate);
1913 }
1914
1915 kfree_rcu(ri, rcu);
1916 return next;
1917}
1918
1919/*
1920 * Called with no locks held.
1921 * Called in context of an exiting or an exec-ing thread.
1922 */
1923void uprobe_free_utask(struct task_struct *t)
1924{
1925 struct uprobe_task *utask = t->utask;
1926 struct return_instance *ri;
1927
1928 if (!utask)
1929 return;
1930
1931 t->utask = NULL;
1932 WARN_ON_ONCE(utask->active_uprobe || utask->xol_vaddr);
1933
1934 timer_delete_sync(&utask->ri_timer);
1935
1936 ri = utask->return_instances;
1937 while (ri)
1938 ri = free_ret_instance(ri, true /* cleanup_hprobe */);
1939
1940 kfree(utask);
1941}
1942
1943#define RI_TIMER_PERIOD (HZ / 10) /* 100 ms */
1944
1945#define for_each_ret_instance_rcu(pos, head) \
1946 for (pos = rcu_dereference_raw(head); pos; pos = rcu_dereference_raw(pos->next))
1947
1948static void ri_timer(struct timer_list *timer)
1949{
1950 struct uprobe_task *utask = container_of(timer, struct uprobe_task, ri_timer);
1951 struct return_instance *ri;
1952
1953 /* SRCU protects uprobe from reuse for the cmpxchg() inside hprobe_expire(). */
1954 guard(srcu)(&uretprobes_srcu);
1955 /* RCU protects return_instance from freeing. */
1956 guard(rcu)();
1957
1958 for_each_ret_instance_rcu(ri, utask->return_instances)
1959 hprobe_expire(&ri->hprobe, false);
1960}
1961
1962static struct uprobe_task *alloc_utask(void)
1963{
1964 struct uprobe_task *utask;
1965
1966 utask = kzalloc(sizeof(*utask), GFP_KERNEL);
1967 if (!utask)
1968 return NULL;
1969
1970 timer_setup(&utask->ri_timer, ri_timer, 0);
1971
1972 return utask;
1973}
1974
1975/*
1976 * Allocate a uprobe_task object for the task if necessary.
1977 * Called when the thread hits a breakpoint.
1978 *
1979 * Returns:
1980 * - pointer to new uprobe_task on success
1981 * - NULL otherwise
1982 */
1983static struct uprobe_task *get_utask(void)
1984{
1985 if (!current->utask)
1986 current->utask = alloc_utask();
1987 return current->utask;
1988}
1989
1990static size_t ri_size(int consumers_cnt)
1991{
1992 struct return_instance *ri;
1993
1994 return sizeof(*ri) + sizeof(ri->consumers[0]) * consumers_cnt;
1995}
1996
1997#define DEF_CNT 4
1998
1999static struct return_instance *alloc_return_instance(void)
2000{
2001 struct return_instance *ri;
2002
2003 ri = kzalloc(ri_size(DEF_CNT), GFP_KERNEL);
2004 if (!ri)
2005 return ZERO_SIZE_PTR;
2006
2007 ri->consumers_cnt = DEF_CNT;
2008 return ri;
2009}
2010
2011static struct return_instance *dup_return_instance(struct return_instance *old)
2012{
2013 size_t size = ri_size(old->consumers_cnt);
2014
2015 return kmemdup(old, size, GFP_KERNEL);
2016}
2017
2018static int dup_utask(struct task_struct *t, struct uprobe_task *o_utask)
2019{
2020 struct uprobe_task *n_utask;
2021 struct return_instance **p, *o, *n;
2022 struct uprobe *uprobe;
2023
2024 n_utask = alloc_utask();
2025 if (!n_utask)
2026 return -ENOMEM;
2027 t->utask = n_utask;
2028
2029 /* protect uprobes from freeing, we'll need try_get_uprobe() them */
2030 guard(srcu)(&uretprobes_srcu);
2031
2032 p = &n_utask->return_instances;
2033 for (o = o_utask->return_instances; o; o = o->next) {
2034 n = dup_return_instance(o);
2035 if (!n)
2036 return -ENOMEM;
2037
2038 /* if uprobe is non-NULL, we'll have an extra refcount for uprobe */
2039 uprobe = hprobe_expire(&o->hprobe, true);
2040
2041 /*
2042 * New utask will have stable properly refcounted uprobe or
2043 * NULL. Even if we failed to get refcounted uprobe, we still
2044 * need to preserve full set of return_instances for proper
2045 * uretprobe handling and nesting in forked task.
2046 */
2047 hprobe_init_stable(&n->hprobe, uprobe);
2048
2049 n->next = NULL;
2050 rcu_assign_pointer(*p, n);
2051 p = &n->next;
2052
2053 n_utask->depth++;
2054 }
2055
2056 return 0;
2057}
2058
2059static void dup_xol_work(struct callback_head *work)
2060{
2061 if (current->flags & PF_EXITING)
2062 return;
2063
2064 if (!__create_xol_area(current->utask->dup_xol_addr) &&
2065 !fatal_signal_pending(current))
2066 uprobe_warn(current, "dup xol area");
2067}
2068
2069/*
2070 * Called in context of a new clone/fork from copy_process.
2071 */
2072void uprobe_copy_process(struct task_struct *t, unsigned long flags)
2073{
2074 struct uprobe_task *utask = current->utask;
2075 struct mm_struct *mm = current->mm;
2076 struct xol_area *area;
2077
2078 t->utask = NULL;
2079
2080 if (!utask || !utask->return_instances)
2081 return;
2082
2083 if (mm == t->mm && !(flags & CLONE_VFORK))
2084 return;
2085
2086 if (dup_utask(t, utask))
2087 return uprobe_warn(t, "dup ret instances");
2088
2089 /* The task can fork() after dup_xol_work() fails */
2090 area = mm->uprobes_state.xol_area;
2091 if (!area)
2092 return uprobe_warn(t, "dup xol area");
2093
2094 if (mm == t->mm)
2095 return;
2096
2097 t->utask->dup_xol_addr = area->vaddr;
2098 init_task_work(&t->utask->dup_xol_work, dup_xol_work);
2099 task_work_add(t, &t->utask->dup_xol_work, TWA_RESUME);
2100}
2101
2102/*
2103 * Current area->vaddr notion assume the trampoline address is always
2104 * equal area->vaddr.
2105 *
2106 * Returns -1 in case the xol_area is not allocated.
2107 */
2108unsigned long uprobe_get_trampoline_vaddr(void)
2109{
2110 struct xol_area *area;
2111 unsigned long trampoline_vaddr = -1;
2112
2113 /* Pairs with xol_add_vma() smp_store_release() */
2114 area = READ_ONCE(current->mm->uprobes_state.xol_area); /* ^^^ */
2115 if (area)
2116 trampoline_vaddr = area->vaddr;
2117
2118 return trampoline_vaddr;
2119}
2120
2121static void cleanup_return_instances(struct uprobe_task *utask, bool chained,
2122 struct pt_regs *regs)
2123{
2124 struct return_instance *ri = utask->return_instances;
2125 enum rp_check ctx = chained ? RP_CHECK_CHAIN_CALL : RP_CHECK_CALL;
2126
2127 while (ri && !arch_uretprobe_is_alive(ri, ctx, regs)) {
2128 ri = free_ret_instance(ri, true /* cleanup_hprobe */);
2129 utask->depth--;
2130 }
2131 rcu_assign_pointer(utask->return_instances, ri);
2132}
2133
2134static void prepare_uretprobe(struct uprobe *uprobe, struct pt_regs *regs,
2135 struct return_instance *ri)
2136{
2137 struct uprobe_task *utask = current->utask;
2138 unsigned long orig_ret_vaddr, trampoline_vaddr;
2139 bool chained;
2140 int srcu_idx;
2141
2142 if (!get_xol_area())
2143 goto free;
2144
2145 if (utask->depth >= MAX_URETPROBE_DEPTH) {
2146 printk_ratelimited(KERN_INFO "uprobe: omit uretprobe due to"
2147 " nestedness limit pid/tgid=%d/%d\n",
2148 current->pid, current->tgid);
2149 goto free;
2150 }
2151
2152 trampoline_vaddr = uprobe_get_trampoline_vaddr();
2153 orig_ret_vaddr = arch_uretprobe_hijack_return_addr(trampoline_vaddr, regs);
2154 if (orig_ret_vaddr == -1)
2155 goto free;
2156
2157 /* drop the entries invalidated by longjmp() */
2158 chained = (orig_ret_vaddr == trampoline_vaddr);
2159 cleanup_return_instances(utask, chained, regs);
2160
2161 /*
2162 * We don't want to keep trampoline address in stack, rather keep the
2163 * original return address of first caller thru all the consequent
2164 * instances. This also makes breakpoint unwrapping easier.
2165 */
2166 if (chained) {
2167 if (!utask->return_instances) {
2168 /*
2169 * This situation is not possible. Likely we have an
2170 * attack from user-space.
2171 */
2172 uprobe_warn(current, "handle tail call");
2173 goto free;
2174 }
2175 orig_ret_vaddr = utask->return_instances->orig_ret_vaddr;
2176 }
2177
2178 /* __srcu_read_lock() because SRCU lock survives switch to user space */
2179 srcu_idx = __srcu_read_lock(&uretprobes_srcu);
2180
2181 ri->func = instruction_pointer(regs);
2182 ri->stack = user_stack_pointer(regs);
2183 ri->orig_ret_vaddr = orig_ret_vaddr;
2184 ri->chained = chained;
2185
2186 utask->depth++;
2187
2188 hprobe_init_leased(&ri->hprobe, uprobe, srcu_idx);
2189 ri->next = utask->return_instances;
2190 rcu_assign_pointer(utask->return_instances, ri);
2191
2192 mod_timer(&utask->ri_timer, jiffies + RI_TIMER_PERIOD);
2193
2194 return;
2195free:
2196 kfree(ri);
2197}
2198
2199/* Prepare to single-step probed instruction out of line. */
2200static int
2201pre_ssout(struct uprobe *uprobe, struct pt_regs *regs, unsigned long bp_vaddr)
2202{
2203 struct uprobe_task *utask = current->utask;
2204 int err;
2205
2206 if (!try_get_uprobe(uprobe))
2207 return -EINVAL;
2208
2209 if (!xol_get_insn_slot(uprobe, utask)) {
2210 err = -ENOMEM;
2211 goto err_out;
2212 }
2213
2214 utask->vaddr = bp_vaddr;
2215 err = arch_uprobe_pre_xol(&uprobe->arch, regs);
2216 if (unlikely(err)) {
2217 xol_free_insn_slot(utask);
2218 goto err_out;
2219 }
2220
2221 utask->active_uprobe = uprobe;
2222 utask->state = UTASK_SSTEP;
2223 return 0;
2224err_out:
2225 put_uprobe(uprobe);
2226 return err;
2227}
2228
2229/*
2230 * If we are singlestepping, then ensure this thread is not connected to
2231 * non-fatal signals until completion of singlestep. When xol insn itself
2232 * triggers the signal, restart the original insn even if the task is
2233 * already SIGKILL'ed (since coredump should report the correct ip). This
2234 * is even more important if the task has a handler for SIGSEGV/etc, The
2235 * _same_ instruction should be repeated again after return from the signal
2236 * handler, and SSTEP can never finish in this case.
2237 */
2238bool uprobe_deny_signal(void)
2239{
2240 struct task_struct *t = current;
2241 struct uprobe_task *utask = t->utask;
2242
2243 if (likely(!utask || !utask->active_uprobe))
2244 return false;
2245
2246 WARN_ON_ONCE(utask->state != UTASK_SSTEP);
2247
2248 if (task_sigpending(t)) {
2249 spin_lock_irq(&t->sighand->siglock);
2250 clear_tsk_thread_flag(t, TIF_SIGPENDING);
2251 spin_unlock_irq(&t->sighand->siglock);
2252
2253 if (__fatal_signal_pending(t) || arch_uprobe_xol_was_trapped(t)) {
2254 utask->state = UTASK_SSTEP_TRAPPED;
2255 set_tsk_thread_flag(t, TIF_UPROBE);
2256 }
2257 }
2258
2259 return true;
2260}
2261
2262static void mmf_recalc_uprobes(struct mm_struct *mm)
2263{
2264 VMA_ITERATOR(vmi, mm, 0);
2265 struct vm_area_struct *vma;
2266
2267 for_each_vma(vmi, vma) {
2268 if (!valid_vma(vma, false))
2269 continue;
2270 /*
2271 * This is not strictly accurate, we can race with
2272 * uprobe_unregister() and see the already removed
2273 * uprobe if delete_uprobe() was not yet called.
2274 * Or this uprobe can be filtered out.
2275 */
2276 if (vma_has_uprobes(vma, vma->vm_start, vma->vm_end))
2277 return;
2278 }
2279
2280 clear_bit(MMF_HAS_UPROBES, &mm->flags);
2281}
2282
2283static int is_trap_at_addr(struct mm_struct *mm, unsigned long vaddr)
2284{
2285 struct page *page;
2286 uprobe_opcode_t opcode;
2287 int result;
2288
2289 if (WARN_ON_ONCE(!IS_ALIGNED(vaddr, UPROBE_SWBP_INSN_SIZE)))
2290 return -EINVAL;
2291
2292 pagefault_disable();
2293 result = __get_user(opcode, (uprobe_opcode_t __user *)vaddr);
2294 pagefault_enable();
2295
2296 if (likely(result == 0))
2297 goto out;
2298
2299 result = get_user_pages(vaddr, 1, FOLL_FORCE, &page);
2300 if (result < 0)
2301 return result;
2302
2303 copy_from_page(page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
2304 put_page(page);
2305 out:
2306 /* This needs to return true for any variant of the trap insn */
2307 return is_trap_insn(&opcode);
2308}
2309
2310/* assumes being inside RCU protected region */
2311static struct uprobe *find_active_uprobe_rcu(unsigned long bp_vaddr, int *is_swbp)
2312{
2313 struct mm_struct *mm = current->mm;
2314 struct uprobe *uprobe = NULL;
2315 struct vm_area_struct *vma;
2316
2317 mmap_read_lock(mm);
2318 vma = vma_lookup(mm, bp_vaddr);
2319 if (vma) {
2320 if (valid_vma(vma, false)) {
2321 struct inode *inode = file_inode(vma->vm_file);
2322 loff_t offset = vaddr_to_offset(vma, bp_vaddr);
2323
2324 uprobe = find_uprobe_rcu(inode, offset);
2325 }
2326
2327 if (!uprobe)
2328 *is_swbp = is_trap_at_addr(mm, bp_vaddr);
2329 } else {
2330 *is_swbp = -EFAULT;
2331 }
2332
2333 if (!uprobe && test_and_clear_bit(MMF_RECALC_UPROBES, &mm->flags))
2334 mmf_recalc_uprobes(mm);
2335 mmap_read_unlock(mm);
2336
2337 return uprobe;
2338}
2339
2340static struct return_instance*
2341push_consumer(struct return_instance *ri, int idx, __u64 id, __u64 cookie)
2342{
2343 if (unlikely(ri == ZERO_SIZE_PTR))
2344 return ri;
2345
2346 if (unlikely(idx >= ri->consumers_cnt)) {
2347 struct return_instance *old_ri = ri;
2348
2349 ri->consumers_cnt += DEF_CNT;
2350 ri = krealloc(old_ri, ri_size(old_ri->consumers_cnt), GFP_KERNEL);
2351 if (!ri) {
2352 kfree(old_ri);
2353 return ZERO_SIZE_PTR;
2354 }
2355 }
2356
2357 ri->consumers[idx].id = id;
2358 ri->consumers[idx].cookie = cookie;
2359 return ri;
2360}
2361
2362static struct return_consumer *
2363return_consumer_find(struct return_instance *ri, int *iter, int id)
2364{
2365 struct return_consumer *ric;
2366 int idx = *iter;
2367
2368 for (ric = &ri->consumers[idx]; idx < ri->consumers_cnt; idx++, ric++) {
2369 if (ric->id == id) {
2370 *iter = idx + 1;
2371 return ric;
2372 }
2373 }
2374 return NULL;
2375}
2376
2377static bool ignore_ret_handler(int rc)
2378{
2379 return rc == UPROBE_HANDLER_REMOVE || rc == UPROBE_HANDLER_IGNORE;
2380}
2381
2382static void handler_chain(struct uprobe *uprobe, struct pt_regs *regs)
2383{
2384 struct uprobe_consumer *uc;
2385 bool has_consumers = false, remove = true;
2386 struct return_instance *ri = NULL;
2387 int push_idx = 0;
2388
2389 current->utask->auprobe = &uprobe->arch;
2390
2391 list_for_each_entry_rcu(uc, &uprobe->consumers, cons_node, rcu_read_lock_trace_held()) {
2392 bool session = uc->handler && uc->ret_handler;
2393 __u64 cookie = 0;
2394 int rc = 0;
2395
2396 if (uc->handler) {
2397 rc = uc->handler(uc, regs, &cookie);
2398 WARN(rc < 0 || rc > 2,
2399 "bad rc=0x%x from %ps()\n", rc, uc->handler);
2400 }
2401
2402 remove &= rc == UPROBE_HANDLER_REMOVE;
2403 has_consumers = true;
2404
2405 if (!uc->ret_handler || ignore_ret_handler(rc))
2406 continue;
2407
2408 if (!ri)
2409 ri = alloc_return_instance();
2410
2411 if (session)
2412 ri = push_consumer(ri, push_idx++, uc->id, cookie);
2413 }
2414 current->utask->auprobe = NULL;
2415
2416 if (!ZERO_OR_NULL_PTR(ri)) {
2417 /*
2418 * The push_idx value has the final number of return consumers,
2419 * and ri->consumers_cnt has number of allocated consumers.
2420 */
2421 ri->consumers_cnt = push_idx;
2422 prepare_uretprobe(uprobe, regs, ri);
2423 }
2424
2425 if (remove && has_consumers) {
2426 down_read(&uprobe->register_rwsem);
2427
2428 /* re-check that removal is still required, this time under lock */
2429 if (!filter_chain(uprobe, current->mm)) {
2430 WARN_ON(!uprobe_is_active(uprobe));
2431 unapply_uprobe(uprobe, current->mm);
2432 }
2433
2434 up_read(&uprobe->register_rwsem);
2435 }
2436}
2437
2438static void
2439handle_uretprobe_chain(struct return_instance *ri, struct uprobe *uprobe, struct pt_regs *regs)
2440{
2441 struct return_consumer *ric;
2442 struct uprobe_consumer *uc;
2443 int ric_idx = 0;
2444
2445 /* all consumers unsubscribed meanwhile */
2446 if (unlikely(!uprobe))
2447 return;
2448
2449 rcu_read_lock_trace();
2450 list_for_each_entry_rcu(uc, &uprobe->consumers, cons_node, rcu_read_lock_trace_held()) {
2451 bool session = uc->handler && uc->ret_handler;
2452
2453 if (uc->ret_handler) {
2454 ric = return_consumer_find(ri, &ric_idx, uc->id);
2455 if (!session || ric)
2456 uc->ret_handler(uc, ri->func, regs, ric ? &ric->cookie : NULL);
2457 }
2458 }
2459 rcu_read_unlock_trace();
2460}
2461
2462static struct return_instance *find_next_ret_chain(struct return_instance *ri)
2463{
2464 bool chained;
2465
2466 do {
2467 chained = ri->chained;
2468 ri = ri->next; /* can't be NULL if chained */
2469 } while (chained);
2470
2471 return ri;
2472}
2473
2474void uprobe_handle_trampoline(struct pt_regs *regs)
2475{
2476 struct uprobe_task *utask;
2477 struct return_instance *ri, *next;
2478 struct uprobe *uprobe;
2479 enum hprobe_state hstate;
2480 bool valid;
2481
2482 utask = current->utask;
2483 if (!utask)
2484 goto sigill;
2485
2486 ri = utask->return_instances;
2487 if (!ri)
2488 goto sigill;
2489
2490 do {
2491 /*
2492 * We should throw out the frames invalidated by longjmp().
2493 * If this chain is valid, then the next one should be alive
2494 * or NULL; the latter case means that nobody but ri->func
2495 * could hit this trampoline on return. TODO: sigaltstack().
2496 */
2497 next = find_next_ret_chain(ri);
2498 valid = !next || arch_uretprobe_is_alive(next, RP_CHECK_RET, regs);
2499
2500 instruction_pointer_set(regs, ri->orig_ret_vaddr);
2501 do {
2502 /* pop current instance from the stack of pending return instances,
2503 * as it's not pending anymore: we just fixed up original
2504 * instruction pointer in regs and are about to call handlers;
2505 * this allows fixup_uretprobe_trampoline_entries() to properly fix up
2506 * captured stack traces from uretprobe handlers, in which pending
2507 * trampoline addresses on the stack are replaced with correct
2508 * original return addresses
2509 */
2510 rcu_assign_pointer(utask->return_instances, ri->next);
2511
2512 uprobe = hprobe_consume(&ri->hprobe, &hstate);
2513 if (valid)
2514 handle_uretprobe_chain(ri, uprobe, regs);
2515 hprobe_finalize(&ri->hprobe, hstate);
2516
2517 /* We already took care of hprobe, no need to waste more time on that. */
2518 ri = free_ret_instance(ri, false /* !cleanup_hprobe */);
2519 utask->depth--;
2520 } while (ri != next);
2521 } while (!valid);
2522
2523 return;
2524
2525sigill:
2526 uprobe_warn(current, "handle uretprobe, sending SIGILL.");
2527 force_sig(SIGILL);
2528}
2529
2530bool __weak arch_uprobe_ignore(struct arch_uprobe *aup, struct pt_regs *regs)
2531{
2532 return false;
2533}
2534
2535bool __weak arch_uretprobe_is_alive(struct return_instance *ret, enum rp_check ctx,
2536 struct pt_regs *regs)
2537{
2538 return true;
2539}
2540
2541/*
2542 * Run handler and ask thread to singlestep.
2543 * Ensure all non-fatal signals cannot interrupt thread while it singlesteps.
2544 */
2545static void handle_swbp(struct pt_regs *regs)
2546{
2547 struct uprobe *uprobe;
2548 unsigned long bp_vaddr;
2549 int is_swbp;
2550
2551 bp_vaddr = uprobe_get_swbp_addr(regs);
2552 if (bp_vaddr == uprobe_get_trampoline_vaddr())
2553 return uprobe_handle_trampoline(regs);
2554
2555 rcu_read_lock_trace();
2556
2557 uprobe = find_active_uprobe_rcu(bp_vaddr, &is_swbp);
2558 if (!uprobe) {
2559 if (is_swbp > 0) {
2560 /* No matching uprobe; signal SIGTRAP. */
2561 force_sig(SIGTRAP);
2562 } else {
2563 /*
2564 * Either we raced with uprobe_unregister() or we can't
2565 * access this memory. The latter is only possible if
2566 * another thread plays with our ->mm. In both cases
2567 * we can simply restart. If this vma was unmapped we
2568 * can pretend this insn was not executed yet and get
2569 * the (correct) SIGSEGV after restart.
2570 */
2571 instruction_pointer_set(regs, bp_vaddr);
2572 }
2573 goto out;
2574 }
2575
2576 /* change it in advance for ->handler() and restart */
2577 instruction_pointer_set(regs, bp_vaddr);
2578
2579 /*
2580 * TODO: move copy_insn/etc into _register and remove this hack.
2581 * After we hit the bp, _unregister + _register can install the
2582 * new and not-yet-analyzed uprobe at the same address, restart.
2583 */
2584 if (unlikely(!test_bit(UPROBE_COPY_INSN, &uprobe->flags)))
2585 goto out;
2586
2587 /*
2588 * Pairs with the smp_wmb() in prepare_uprobe().
2589 *
2590 * Guarantees that if we see the UPROBE_COPY_INSN bit set, then
2591 * we must also see the stores to &uprobe->arch performed by the
2592 * prepare_uprobe() call.
2593 */
2594 smp_rmb();
2595
2596 /* Tracing handlers use ->utask to communicate with fetch methods */
2597 if (!get_utask())
2598 goto out;
2599
2600 if (arch_uprobe_ignore(&uprobe->arch, regs))
2601 goto out;
2602
2603 handler_chain(uprobe, regs);
2604
2605 if (arch_uprobe_skip_sstep(&uprobe->arch, regs))
2606 goto out;
2607
2608 if (pre_ssout(uprobe, regs, bp_vaddr))
2609 goto out;
2610
2611out:
2612 /* arch_uprobe_skip_sstep() succeeded, or restart if can't singlestep */
2613 rcu_read_unlock_trace();
2614}
2615
2616/*
2617 * Perform required fix-ups and disable singlestep.
2618 * Allow pending signals to take effect.
2619 */
2620static void handle_singlestep(struct uprobe_task *utask, struct pt_regs *regs)
2621{
2622 struct uprobe *uprobe;
2623 int err = 0;
2624
2625 uprobe = utask->active_uprobe;
2626 if (utask->state == UTASK_SSTEP_ACK)
2627 err = arch_uprobe_post_xol(&uprobe->arch, regs);
2628 else if (utask->state == UTASK_SSTEP_TRAPPED)
2629 arch_uprobe_abort_xol(&uprobe->arch, regs);
2630 else
2631 WARN_ON_ONCE(1);
2632
2633 put_uprobe(uprobe);
2634 utask->active_uprobe = NULL;
2635 utask->state = UTASK_RUNNING;
2636 xol_free_insn_slot(utask);
2637
2638 spin_lock_irq(¤t->sighand->siglock);
2639 recalc_sigpending(); /* see uprobe_deny_signal() */
2640 spin_unlock_irq(¤t->sighand->siglock);
2641
2642 if (unlikely(err)) {
2643 uprobe_warn(current, "execute the probed insn, sending SIGILL.");
2644 force_sig(SIGILL);
2645 }
2646}
2647
2648/*
2649 * On breakpoint hit, breakpoint notifier sets the TIF_UPROBE flag and
2650 * allows the thread to return from interrupt. After that handle_swbp()
2651 * sets utask->active_uprobe.
2652 *
2653 * On singlestep exception, singlestep notifier sets the TIF_UPROBE flag
2654 * and allows the thread to return from interrupt.
2655 *
2656 * While returning to userspace, thread notices the TIF_UPROBE flag and calls
2657 * uprobe_notify_resume().
2658 */
2659void uprobe_notify_resume(struct pt_regs *regs)
2660{
2661 struct uprobe_task *utask;
2662
2663 clear_thread_flag(TIF_UPROBE);
2664
2665 utask = current->utask;
2666 if (utask && utask->active_uprobe)
2667 handle_singlestep(utask, regs);
2668 else
2669 handle_swbp(regs);
2670}
2671
2672/*
2673 * uprobe_pre_sstep_notifier gets called from interrupt context as part of
2674 * notifier mechanism. Set TIF_UPROBE flag and indicate breakpoint hit.
2675 */
2676int uprobe_pre_sstep_notifier(struct pt_regs *regs)
2677{
2678 if (!current->mm)
2679 return 0;
2680
2681 if (!test_bit(MMF_HAS_UPROBES, ¤t->mm->flags) &&
2682 (!current->utask || !current->utask->return_instances))
2683 return 0;
2684
2685 set_thread_flag(TIF_UPROBE);
2686 return 1;
2687}
2688
2689/*
2690 * uprobe_post_sstep_notifier gets called in interrupt context as part of notifier
2691 * mechanism. Set TIF_UPROBE flag and indicate completion of singlestep.
2692 */
2693int uprobe_post_sstep_notifier(struct pt_regs *regs)
2694{
2695 struct uprobe_task *utask = current->utask;
2696
2697 if (!current->mm || !utask || !utask->active_uprobe)
2698 /* task is currently not uprobed */
2699 return 0;
2700
2701 utask->state = UTASK_SSTEP_ACK;
2702 set_thread_flag(TIF_UPROBE);
2703 return 1;
2704}
2705
2706static struct notifier_block uprobe_exception_nb = {
2707 .notifier_call = arch_uprobe_exception_notify,
2708 .priority = INT_MAX-1, /* notified after kprobes, kgdb */
2709};
2710
2711void __init uprobes_init(void)
2712{
2713 int i;
2714
2715 for (i = 0; i < UPROBES_HASH_SZ; i++)
2716 mutex_init(&uprobes_mmap_mutex[i]);
2717
2718 BUG_ON(register_die_notifier(&uprobe_exception_nb));
2719}
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * User-space Probes (UProbes)
4 *
5 * Copyright (C) IBM Corporation, 2008-2012
6 * Authors:
7 * Srikar Dronamraju
8 * Jim Keniston
9 * Copyright (C) 2011-2012 Red Hat, Inc., Peter Zijlstra
10 */
11
12#include <linux/kernel.h>
13#include <linux/highmem.h>
14#include <linux/pagemap.h> /* read_mapping_page */
15#include <linux/slab.h>
16#include <linux/sched.h>
17#include <linux/sched/mm.h>
18#include <linux/sched/coredump.h>
19#include <linux/export.h>
20#include <linux/rmap.h> /* anon_vma_prepare */
21#include <linux/mmu_notifier.h> /* set_pte_at_notify */
22#include <linux/swap.h> /* folio_free_swap */
23#include <linux/ptrace.h> /* user_enable_single_step */
24#include <linux/kdebug.h> /* notifier mechanism */
25#include <linux/percpu-rwsem.h>
26#include <linux/task_work.h>
27#include <linux/shmem_fs.h>
28#include <linux/khugepaged.h>
29
30#include <linux/uprobes.h>
31
32#define UINSNS_PER_PAGE (PAGE_SIZE/UPROBE_XOL_SLOT_BYTES)
33#define MAX_UPROBE_XOL_SLOTS UINSNS_PER_PAGE
34
35static struct rb_root uprobes_tree = RB_ROOT;
36/*
37 * allows us to skip the uprobe_mmap if there are no uprobe events active
38 * at this time. Probably a fine grained per inode count is better?
39 */
40#define no_uprobe_events() RB_EMPTY_ROOT(&uprobes_tree)
41
42static DEFINE_SPINLOCK(uprobes_treelock); /* serialize rbtree access */
43
44#define UPROBES_HASH_SZ 13
45/* serialize uprobe->pending_list */
46static struct mutex uprobes_mmap_mutex[UPROBES_HASH_SZ];
47#define uprobes_mmap_hash(v) (&uprobes_mmap_mutex[((unsigned long)(v)) % UPROBES_HASH_SZ])
48
49DEFINE_STATIC_PERCPU_RWSEM(dup_mmap_sem);
50
51/* Have a copy of original instruction */
52#define UPROBE_COPY_INSN 0
53
54struct uprobe {
55 struct rb_node rb_node; /* node in the rb tree */
56 refcount_t ref;
57 struct rw_semaphore register_rwsem;
58 struct rw_semaphore consumer_rwsem;
59 struct list_head pending_list;
60 struct uprobe_consumer *consumers;
61 struct inode *inode; /* Also hold a ref to inode */
62 loff_t offset;
63 loff_t ref_ctr_offset;
64 unsigned long flags;
65
66 /*
67 * The generic code assumes that it has two members of unknown type
68 * owned by the arch-specific code:
69 *
70 * insn - copy_insn() saves the original instruction here for
71 * arch_uprobe_analyze_insn().
72 *
73 * ixol - potentially modified instruction to execute out of
74 * line, copied to xol_area by xol_get_insn_slot().
75 */
76 struct arch_uprobe arch;
77};
78
79struct delayed_uprobe {
80 struct list_head list;
81 struct uprobe *uprobe;
82 struct mm_struct *mm;
83};
84
85static DEFINE_MUTEX(delayed_uprobe_lock);
86static LIST_HEAD(delayed_uprobe_list);
87
88/*
89 * Execute out of line area: anonymous executable mapping installed
90 * by the probed task to execute the copy of the original instruction
91 * mangled by set_swbp().
92 *
93 * On a breakpoint hit, thread contests for a slot. It frees the
94 * slot after singlestep. Currently a fixed number of slots are
95 * allocated.
96 */
97struct xol_area {
98 wait_queue_head_t wq; /* if all slots are busy */
99 atomic_t slot_count; /* number of in-use slots */
100 unsigned long *bitmap; /* 0 = free slot */
101
102 struct vm_special_mapping xol_mapping;
103 struct page *pages[2];
104 /*
105 * We keep the vma's vm_start rather than a pointer to the vma
106 * itself. The probed process or a naughty kernel module could make
107 * the vma go away, and we must handle that reasonably gracefully.
108 */
109 unsigned long vaddr; /* Page(s) of instruction slots */
110};
111
112/*
113 * valid_vma: Verify if the specified vma is an executable vma
114 * Relax restrictions while unregistering: vm_flags might have
115 * changed after breakpoint was inserted.
116 * - is_register: indicates if we are in register context.
117 * - Return 1 if the specified virtual address is in an
118 * executable vma.
119 */
120static bool valid_vma(struct vm_area_struct *vma, bool is_register)
121{
122 vm_flags_t flags = VM_HUGETLB | VM_MAYEXEC | VM_MAYSHARE;
123
124 if (is_register)
125 flags |= VM_WRITE;
126
127 return vma->vm_file && (vma->vm_flags & flags) == VM_MAYEXEC;
128}
129
130static unsigned long offset_to_vaddr(struct vm_area_struct *vma, loff_t offset)
131{
132 return vma->vm_start + offset - ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
133}
134
135static loff_t vaddr_to_offset(struct vm_area_struct *vma, unsigned long vaddr)
136{
137 return ((loff_t)vma->vm_pgoff << PAGE_SHIFT) + (vaddr - vma->vm_start);
138}
139
140/**
141 * __replace_page - replace page in vma by new page.
142 * based on replace_page in mm/ksm.c
143 *
144 * @vma: vma that holds the pte pointing to page
145 * @addr: address the old @page is mapped at
146 * @old_page: the page we are replacing by new_page
147 * @new_page: the modified page we replace page by
148 *
149 * If @new_page is NULL, only unmap @old_page.
150 *
151 * Returns 0 on success, negative error code otherwise.
152 */
153static int __replace_page(struct vm_area_struct *vma, unsigned long addr,
154 struct page *old_page, struct page *new_page)
155{
156 struct folio *old_folio = page_folio(old_page);
157 struct folio *new_folio;
158 struct mm_struct *mm = vma->vm_mm;
159 DEFINE_FOLIO_VMA_WALK(pvmw, old_folio, vma, addr, 0);
160 int err;
161 struct mmu_notifier_range range;
162
163 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr,
164 addr + PAGE_SIZE);
165
166 if (new_page) {
167 new_folio = page_folio(new_page);
168 err = mem_cgroup_charge(new_folio, vma->vm_mm, GFP_KERNEL);
169 if (err)
170 return err;
171 }
172
173 /* For folio_free_swap() below */
174 folio_lock(old_folio);
175
176 mmu_notifier_invalidate_range_start(&range);
177 err = -EAGAIN;
178 if (!page_vma_mapped_walk(&pvmw))
179 goto unlock;
180 VM_BUG_ON_PAGE(addr != pvmw.address, old_page);
181
182 if (new_page) {
183 folio_get(new_folio);
184 folio_add_new_anon_rmap(new_folio, vma, addr);
185 folio_add_lru_vma(new_folio, vma);
186 } else
187 /* no new page, just dec_mm_counter for old_page */
188 dec_mm_counter(mm, MM_ANONPAGES);
189
190 if (!folio_test_anon(old_folio)) {
191 dec_mm_counter(mm, mm_counter_file(old_folio));
192 inc_mm_counter(mm, MM_ANONPAGES);
193 }
194
195 flush_cache_page(vma, addr, pte_pfn(ptep_get(pvmw.pte)));
196 ptep_clear_flush(vma, addr, pvmw.pte);
197 if (new_page)
198 set_pte_at_notify(mm, addr, pvmw.pte,
199 mk_pte(new_page, vma->vm_page_prot));
200
201 folio_remove_rmap_pte(old_folio, old_page, vma);
202 if (!folio_mapped(old_folio))
203 folio_free_swap(old_folio);
204 page_vma_mapped_walk_done(&pvmw);
205 folio_put(old_folio);
206
207 err = 0;
208 unlock:
209 mmu_notifier_invalidate_range_end(&range);
210 folio_unlock(old_folio);
211 return err;
212}
213
214/**
215 * is_swbp_insn - check if instruction is breakpoint instruction.
216 * @insn: instruction to be checked.
217 * Default implementation of is_swbp_insn
218 * Returns true if @insn is a breakpoint instruction.
219 */
220bool __weak is_swbp_insn(uprobe_opcode_t *insn)
221{
222 return *insn == UPROBE_SWBP_INSN;
223}
224
225/**
226 * is_trap_insn - check if instruction is breakpoint instruction.
227 * @insn: instruction to be checked.
228 * Default implementation of is_trap_insn
229 * Returns true if @insn is a breakpoint instruction.
230 *
231 * This function is needed for the case where an architecture has multiple
232 * trap instructions (like powerpc).
233 */
234bool __weak is_trap_insn(uprobe_opcode_t *insn)
235{
236 return is_swbp_insn(insn);
237}
238
239static void copy_from_page(struct page *page, unsigned long vaddr, void *dst, int len)
240{
241 void *kaddr = kmap_atomic(page);
242 memcpy(dst, kaddr + (vaddr & ~PAGE_MASK), len);
243 kunmap_atomic(kaddr);
244}
245
246static void copy_to_page(struct page *page, unsigned long vaddr, const void *src, int len)
247{
248 void *kaddr = kmap_atomic(page);
249 memcpy(kaddr + (vaddr & ~PAGE_MASK), src, len);
250 kunmap_atomic(kaddr);
251}
252
253static int verify_opcode(struct page *page, unsigned long vaddr, uprobe_opcode_t *new_opcode)
254{
255 uprobe_opcode_t old_opcode;
256 bool is_swbp;
257
258 /*
259 * Note: We only check if the old_opcode is UPROBE_SWBP_INSN here.
260 * We do not check if it is any other 'trap variant' which could
261 * be conditional trap instruction such as the one powerpc supports.
262 *
263 * The logic is that we do not care if the underlying instruction
264 * is a trap variant; uprobes always wins over any other (gdb)
265 * breakpoint.
266 */
267 copy_from_page(page, vaddr, &old_opcode, UPROBE_SWBP_INSN_SIZE);
268 is_swbp = is_swbp_insn(&old_opcode);
269
270 if (is_swbp_insn(new_opcode)) {
271 if (is_swbp) /* register: already installed? */
272 return 0;
273 } else {
274 if (!is_swbp) /* unregister: was it changed by us? */
275 return 0;
276 }
277
278 return 1;
279}
280
281static struct delayed_uprobe *
282delayed_uprobe_check(struct uprobe *uprobe, struct mm_struct *mm)
283{
284 struct delayed_uprobe *du;
285
286 list_for_each_entry(du, &delayed_uprobe_list, list)
287 if (du->uprobe == uprobe && du->mm == mm)
288 return du;
289 return NULL;
290}
291
292static int delayed_uprobe_add(struct uprobe *uprobe, struct mm_struct *mm)
293{
294 struct delayed_uprobe *du;
295
296 if (delayed_uprobe_check(uprobe, mm))
297 return 0;
298
299 du = kzalloc(sizeof(*du), GFP_KERNEL);
300 if (!du)
301 return -ENOMEM;
302
303 du->uprobe = uprobe;
304 du->mm = mm;
305 list_add(&du->list, &delayed_uprobe_list);
306 return 0;
307}
308
309static void delayed_uprobe_delete(struct delayed_uprobe *du)
310{
311 if (WARN_ON(!du))
312 return;
313 list_del(&du->list);
314 kfree(du);
315}
316
317static void delayed_uprobe_remove(struct uprobe *uprobe, struct mm_struct *mm)
318{
319 struct list_head *pos, *q;
320 struct delayed_uprobe *du;
321
322 if (!uprobe && !mm)
323 return;
324
325 list_for_each_safe(pos, q, &delayed_uprobe_list) {
326 du = list_entry(pos, struct delayed_uprobe, list);
327
328 if (uprobe && du->uprobe != uprobe)
329 continue;
330 if (mm && du->mm != mm)
331 continue;
332
333 delayed_uprobe_delete(du);
334 }
335}
336
337static bool valid_ref_ctr_vma(struct uprobe *uprobe,
338 struct vm_area_struct *vma)
339{
340 unsigned long vaddr = offset_to_vaddr(vma, uprobe->ref_ctr_offset);
341
342 return uprobe->ref_ctr_offset &&
343 vma->vm_file &&
344 file_inode(vma->vm_file) == uprobe->inode &&
345 (vma->vm_flags & (VM_WRITE|VM_SHARED)) == VM_WRITE &&
346 vma->vm_start <= vaddr &&
347 vma->vm_end > vaddr;
348}
349
350static struct vm_area_struct *
351find_ref_ctr_vma(struct uprobe *uprobe, struct mm_struct *mm)
352{
353 VMA_ITERATOR(vmi, mm, 0);
354 struct vm_area_struct *tmp;
355
356 for_each_vma(vmi, tmp)
357 if (valid_ref_ctr_vma(uprobe, tmp))
358 return tmp;
359
360 return NULL;
361}
362
363static int
364__update_ref_ctr(struct mm_struct *mm, unsigned long vaddr, short d)
365{
366 void *kaddr;
367 struct page *page;
368 int ret;
369 short *ptr;
370
371 if (!vaddr || !d)
372 return -EINVAL;
373
374 ret = get_user_pages_remote(mm, vaddr, 1,
375 FOLL_WRITE, &page, NULL);
376 if (unlikely(ret <= 0)) {
377 /*
378 * We are asking for 1 page. If get_user_pages_remote() fails,
379 * it may return 0, in that case we have to return error.
380 */
381 return ret == 0 ? -EBUSY : ret;
382 }
383
384 kaddr = kmap_atomic(page);
385 ptr = kaddr + (vaddr & ~PAGE_MASK);
386
387 if (unlikely(*ptr + d < 0)) {
388 pr_warn("ref_ctr going negative. vaddr: 0x%lx, "
389 "curr val: %d, delta: %d\n", vaddr, *ptr, d);
390 ret = -EINVAL;
391 goto out;
392 }
393
394 *ptr += d;
395 ret = 0;
396out:
397 kunmap_atomic(kaddr);
398 put_page(page);
399 return ret;
400}
401
402static void update_ref_ctr_warn(struct uprobe *uprobe,
403 struct mm_struct *mm, short d)
404{
405 pr_warn("ref_ctr %s failed for inode: 0x%lx offset: "
406 "0x%llx ref_ctr_offset: 0x%llx of mm: 0x%pK\n",
407 d > 0 ? "increment" : "decrement", uprobe->inode->i_ino,
408 (unsigned long long) uprobe->offset,
409 (unsigned long long) uprobe->ref_ctr_offset, mm);
410}
411
412static int update_ref_ctr(struct uprobe *uprobe, struct mm_struct *mm,
413 short d)
414{
415 struct vm_area_struct *rc_vma;
416 unsigned long rc_vaddr;
417 int ret = 0;
418
419 rc_vma = find_ref_ctr_vma(uprobe, mm);
420
421 if (rc_vma) {
422 rc_vaddr = offset_to_vaddr(rc_vma, uprobe->ref_ctr_offset);
423 ret = __update_ref_ctr(mm, rc_vaddr, d);
424 if (ret)
425 update_ref_ctr_warn(uprobe, mm, d);
426
427 if (d > 0)
428 return ret;
429 }
430
431 mutex_lock(&delayed_uprobe_lock);
432 if (d > 0)
433 ret = delayed_uprobe_add(uprobe, mm);
434 else
435 delayed_uprobe_remove(uprobe, mm);
436 mutex_unlock(&delayed_uprobe_lock);
437
438 return ret;
439}
440
441/*
442 * NOTE:
443 * Expect the breakpoint instruction to be the smallest size instruction for
444 * the architecture. If an arch has variable length instruction and the
445 * breakpoint instruction is not of the smallest length instruction
446 * supported by that architecture then we need to modify is_trap_at_addr and
447 * uprobe_write_opcode accordingly. This would never be a problem for archs
448 * that have fixed length instructions.
449 *
450 * uprobe_write_opcode - write the opcode at a given virtual address.
451 * @auprobe: arch specific probepoint information.
452 * @mm: the probed process address space.
453 * @vaddr: the virtual address to store the opcode.
454 * @opcode: opcode to be written at @vaddr.
455 *
456 * Called with mm->mmap_lock held for write.
457 * Return 0 (success) or a negative errno.
458 */
459int uprobe_write_opcode(struct arch_uprobe *auprobe, struct mm_struct *mm,
460 unsigned long vaddr, uprobe_opcode_t opcode)
461{
462 struct uprobe *uprobe;
463 struct page *old_page, *new_page;
464 struct vm_area_struct *vma;
465 int ret, is_register, ref_ctr_updated = 0;
466 bool orig_page_huge = false;
467 unsigned int gup_flags = FOLL_FORCE;
468
469 is_register = is_swbp_insn(&opcode);
470 uprobe = container_of(auprobe, struct uprobe, arch);
471
472retry:
473 if (is_register)
474 gup_flags |= FOLL_SPLIT_PMD;
475 /* Read the page with vaddr into memory */
476 old_page = get_user_page_vma_remote(mm, vaddr, gup_flags, &vma);
477 if (IS_ERR(old_page))
478 return PTR_ERR(old_page);
479
480 ret = verify_opcode(old_page, vaddr, &opcode);
481 if (ret <= 0)
482 goto put_old;
483
484 if (WARN(!is_register && PageCompound(old_page),
485 "uprobe unregister should never work on compound page\n")) {
486 ret = -EINVAL;
487 goto put_old;
488 }
489
490 /* We are going to replace instruction, update ref_ctr. */
491 if (!ref_ctr_updated && uprobe->ref_ctr_offset) {
492 ret = update_ref_ctr(uprobe, mm, is_register ? 1 : -1);
493 if (ret)
494 goto put_old;
495
496 ref_ctr_updated = 1;
497 }
498
499 ret = 0;
500 if (!is_register && !PageAnon(old_page))
501 goto put_old;
502
503 ret = anon_vma_prepare(vma);
504 if (ret)
505 goto put_old;
506
507 ret = -ENOMEM;
508 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vaddr);
509 if (!new_page)
510 goto put_old;
511
512 __SetPageUptodate(new_page);
513 copy_highpage(new_page, old_page);
514 copy_to_page(new_page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
515
516 if (!is_register) {
517 struct page *orig_page;
518 pgoff_t index;
519
520 VM_BUG_ON_PAGE(!PageAnon(old_page), old_page);
521
522 index = vaddr_to_offset(vma, vaddr & PAGE_MASK) >> PAGE_SHIFT;
523 orig_page = find_get_page(vma->vm_file->f_inode->i_mapping,
524 index);
525
526 if (orig_page) {
527 if (PageUptodate(orig_page) &&
528 pages_identical(new_page, orig_page)) {
529 /* let go new_page */
530 put_page(new_page);
531 new_page = NULL;
532
533 if (PageCompound(orig_page))
534 orig_page_huge = true;
535 }
536 put_page(orig_page);
537 }
538 }
539
540 ret = __replace_page(vma, vaddr & PAGE_MASK, old_page, new_page);
541 if (new_page)
542 put_page(new_page);
543put_old:
544 put_page(old_page);
545
546 if (unlikely(ret == -EAGAIN))
547 goto retry;
548
549 /* Revert back reference counter if instruction update failed. */
550 if (ret && is_register && ref_ctr_updated)
551 update_ref_ctr(uprobe, mm, -1);
552
553 /* try collapse pmd for compound page */
554 if (!ret && orig_page_huge)
555 collapse_pte_mapped_thp(mm, vaddr, false);
556
557 return ret;
558}
559
560/**
561 * set_swbp - store breakpoint at a given address.
562 * @auprobe: arch specific probepoint information.
563 * @mm: the probed process address space.
564 * @vaddr: the virtual address to insert the opcode.
565 *
566 * For mm @mm, store the breakpoint instruction at @vaddr.
567 * Return 0 (success) or a negative errno.
568 */
569int __weak set_swbp(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
570{
571 return uprobe_write_opcode(auprobe, mm, vaddr, UPROBE_SWBP_INSN);
572}
573
574/**
575 * set_orig_insn - Restore the original instruction.
576 * @mm: the probed process address space.
577 * @auprobe: arch specific probepoint information.
578 * @vaddr: the virtual address to insert the opcode.
579 *
580 * For mm @mm, restore the original opcode (opcode) at @vaddr.
581 * Return 0 (success) or a negative errno.
582 */
583int __weak
584set_orig_insn(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
585{
586 return uprobe_write_opcode(auprobe, mm, vaddr,
587 *(uprobe_opcode_t *)&auprobe->insn);
588}
589
590static struct uprobe *get_uprobe(struct uprobe *uprobe)
591{
592 refcount_inc(&uprobe->ref);
593 return uprobe;
594}
595
596static void put_uprobe(struct uprobe *uprobe)
597{
598 if (refcount_dec_and_test(&uprobe->ref)) {
599 /*
600 * If application munmap(exec_vma) before uprobe_unregister()
601 * gets called, we don't get a chance to remove uprobe from
602 * delayed_uprobe_list from remove_breakpoint(). Do it here.
603 */
604 mutex_lock(&delayed_uprobe_lock);
605 delayed_uprobe_remove(uprobe, NULL);
606 mutex_unlock(&delayed_uprobe_lock);
607 kfree(uprobe);
608 }
609}
610
611static __always_inline
612int uprobe_cmp(const struct inode *l_inode, const loff_t l_offset,
613 const struct uprobe *r)
614{
615 if (l_inode < r->inode)
616 return -1;
617
618 if (l_inode > r->inode)
619 return 1;
620
621 if (l_offset < r->offset)
622 return -1;
623
624 if (l_offset > r->offset)
625 return 1;
626
627 return 0;
628}
629
630#define __node_2_uprobe(node) \
631 rb_entry((node), struct uprobe, rb_node)
632
633struct __uprobe_key {
634 struct inode *inode;
635 loff_t offset;
636};
637
638static inline int __uprobe_cmp_key(const void *key, const struct rb_node *b)
639{
640 const struct __uprobe_key *a = key;
641 return uprobe_cmp(a->inode, a->offset, __node_2_uprobe(b));
642}
643
644static inline int __uprobe_cmp(struct rb_node *a, const struct rb_node *b)
645{
646 struct uprobe *u = __node_2_uprobe(a);
647 return uprobe_cmp(u->inode, u->offset, __node_2_uprobe(b));
648}
649
650static struct uprobe *__find_uprobe(struct inode *inode, loff_t offset)
651{
652 struct __uprobe_key key = {
653 .inode = inode,
654 .offset = offset,
655 };
656 struct rb_node *node = rb_find(&key, &uprobes_tree, __uprobe_cmp_key);
657
658 if (node)
659 return get_uprobe(__node_2_uprobe(node));
660
661 return NULL;
662}
663
664/*
665 * Find a uprobe corresponding to a given inode:offset
666 * Acquires uprobes_treelock
667 */
668static struct uprobe *find_uprobe(struct inode *inode, loff_t offset)
669{
670 struct uprobe *uprobe;
671
672 spin_lock(&uprobes_treelock);
673 uprobe = __find_uprobe(inode, offset);
674 spin_unlock(&uprobes_treelock);
675
676 return uprobe;
677}
678
679static struct uprobe *__insert_uprobe(struct uprobe *uprobe)
680{
681 struct rb_node *node;
682
683 node = rb_find_add(&uprobe->rb_node, &uprobes_tree, __uprobe_cmp);
684 if (node)
685 return get_uprobe(__node_2_uprobe(node));
686
687 /* get access + creation ref */
688 refcount_set(&uprobe->ref, 2);
689 return NULL;
690}
691
692/*
693 * Acquire uprobes_treelock.
694 * Matching uprobe already exists in rbtree;
695 * increment (access refcount) and return the matching uprobe.
696 *
697 * No matching uprobe; insert the uprobe in rb_tree;
698 * get a double refcount (access + creation) and return NULL.
699 */
700static struct uprobe *insert_uprobe(struct uprobe *uprobe)
701{
702 struct uprobe *u;
703
704 spin_lock(&uprobes_treelock);
705 u = __insert_uprobe(uprobe);
706 spin_unlock(&uprobes_treelock);
707
708 return u;
709}
710
711static void
712ref_ctr_mismatch_warn(struct uprobe *cur_uprobe, struct uprobe *uprobe)
713{
714 pr_warn("ref_ctr_offset mismatch. inode: 0x%lx offset: 0x%llx "
715 "ref_ctr_offset(old): 0x%llx ref_ctr_offset(new): 0x%llx\n",
716 uprobe->inode->i_ino, (unsigned long long) uprobe->offset,
717 (unsigned long long) cur_uprobe->ref_ctr_offset,
718 (unsigned long long) uprobe->ref_ctr_offset);
719}
720
721static struct uprobe *alloc_uprobe(struct inode *inode, loff_t offset,
722 loff_t ref_ctr_offset)
723{
724 struct uprobe *uprobe, *cur_uprobe;
725
726 uprobe = kzalloc(sizeof(struct uprobe), GFP_KERNEL);
727 if (!uprobe)
728 return NULL;
729
730 uprobe->inode = inode;
731 uprobe->offset = offset;
732 uprobe->ref_ctr_offset = ref_ctr_offset;
733 init_rwsem(&uprobe->register_rwsem);
734 init_rwsem(&uprobe->consumer_rwsem);
735
736 /* add to uprobes_tree, sorted on inode:offset */
737 cur_uprobe = insert_uprobe(uprobe);
738 /* a uprobe exists for this inode:offset combination */
739 if (cur_uprobe) {
740 if (cur_uprobe->ref_ctr_offset != uprobe->ref_ctr_offset) {
741 ref_ctr_mismatch_warn(cur_uprobe, uprobe);
742 put_uprobe(cur_uprobe);
743 kfree(uprobe);
744 return ERR_PTR(-EINVAL);
745 }
746 kfree(uprobe);
747 uprobe = cur_uprobe;
748 }
749
750 return uprobe;
751}
752
753static void consumer_add(struct uprobe *uprobe, struct uprobe_consumer *uc)
754{
755 down_write(&uprobe->consumer_rwsem);
756 uc->next = uprobe->consumers;
757 uprobe->consumers = uc;
758 up_write(&uprobe->consumer_rwsem);
759}
760
761/*
762 * For uprobe @uprobe, delete the consumer @uc.
763 * Return true if the @uc is deleted successfully
764 * or return false.
765 */
766static bool consumer_del(struct uprobe *uprobe, struct uprobe_consumer *uc)
767{
768 struct uprobe_consumer **con;
769 bool ret = false;
770
771 down_write(&uprobe->consumer_rwsem);
772 for (con = &uprobe->consumers; *con; con = &(*con)->next) {
773 if (*con == uc) {
774 *con = uc->next;
775 ret = true;
776 break;
777 }
778 }
779 up_write(&uprobe->consumer_rwsem);
780
781 return ret;
782}
783
784static int __copy_insn(struct address_space *mapping, struct file *filp,
785 void *insn, int nbytes, loff_t offset)
786{
787 struct page *page;
788 /*
789 * Ensure that the page that has the original instruction is populated
790 * and in page-cache. If ->read_folio == NULL it must be shmem_mapping(),
791 * see uprobe_register().
792 */
793 if (mapping->a_ops->read_folio)
794 page = read_mapping_page(mapping, offset >> PAGE_SHIFT, filp);
795 else
796 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
797 if (IS_ERR(page))
798 return PTR_ERR(page);
799
800 copy_from_page(page, offset, insn, nbytes);
801 put_page(page);
802
803 return 0;
804}
805
806static int copy_insn(struct uprobe *uprobe, struct file *filp)
807{
808 struct address_space *mapping = uprobe->inode->i_mapping;
809 loff_t offs = uprobe->offset;
810 void *insn = &uprobe->arch.insn;
811 int size = sizeof(uprobe->arch.insn);
812 int len, err = -EIO;
813
814 /* Copy only available bytes, -EIO if nothing was read */
815 do {
816 if (offs >= i_size_read(uprobe->inode))
817 break;
818
819 len = min_t(int, size, PAGE_SIZE - (offs & ~PAGE_MASK));
820 err = __copy_insn(mapping, filp, insn, len, offs);
821 if (err)
822 break;
823
824 insn += len;
825 offs += len;
826 size -= len;
827 } while (size);
828
829 return err;
830}
831
832static int prepare_uprobe(struct uprobe *uprobe, struct file *file,
833 struct mm_struct *mm, unsigned long vaddr)
834{
835 int ret = 0;
836
837 if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
838 return ret;
839
840 /* TODO: move this into _register, until then we abuse this sem. */
841 down_write(&uprobe->consumer_rwsem);
842 if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
843 goto out;
844
845 ret = copy_insn(uprobe, file);
846 if (ret)
847 goto out;
848
849 ret = -ENOTSUPP;
850 if (is_trap_insn((uprobe_opcode_t *)&uprobe->arch.insn))
851 goto out;
852
853 ret = arch_uprobe_analyze_insn(&uprobe->arch, mm, vaddr);
854 if (ret)
855 goto out;
856
857 smp_wmb(); /* pairs with the smp_rmb() in handle_swbp() */
858 set_bit(UPROBE_COPY_INSN, &uprobe->flags);
859
860 out:
861 up_write(&uprobe->consumer_rwsem);
862
863 return ret;
864}
865
866static inline bool consumer_filter(struct uprobe_consumer *uc,
867 enum uprobe_filter_ctx ctx, struct mm_struct *mm)
868{
869 return !uc->filter || uc->filter(uc, ctx, mm);
870}
871
872static bool filter_chain(struct uprobe *uprobe,
873 enum uprobe_filter_ctx ctx, struct mm_struct *mm)
874{
875 struct uprobe_consumer *uc;
876 bool ret = false;
877
878 down_read(&uprobe->consumer_rwsem);
879 for (uc = uprobe->consumers; uc; uc = uc->next) {
880 ret = consumer_filter(uc, ctx, mm);
881 if (ret)
882 break;
883 }
884 up_read(&uprobe->consumer_rwsem);
885
886 return ret;
887}
888
889static int
890install_breakpoint(struct uprobe *uprobe, struct mm_struct *mm,
891 struct vm_area_struct *vma, unsigned long vaddr)
892{
893 bool first_uprobe;
894 int ret;
895
896 ret = prepare_uprobe(uprobe, vma->vm_file, mm, vaddr);
897 if (ret)
898 return ret;
899
900 /*
901 * set MMF_HAS_UPROBES in advance for uprobe_pre_sstep_notifier(),
902 * the task can hit this breakpoint right after __replace_page().
903 */
904 first_uprobe = !test_bit(MMF_HAS_UPROBES, &mm->flags);
905 if (first_uprobe)
906 set_bit(MMF_HAS_UPROBES, &mm->flags);
907
908 ret = set_swbp(&uprobe->arch, mm, vaddr);
909 if (!ret)
910 clear_bit(MMF_RECALC_UPROBES, &mm->flags);
911 else if (first_uprobe)
912 clear_bit(MMF_HAS_UPROBES, &mm->flags);
913
914 return ret;
915}
916
917static int
918remove_breakpoint(struct uprobe *uprobe, struct mm_struct *mm, unsigned long vaddr)
919{
920 set_bit(MMF_RECALC_UPROBES, &mm->flags);
921 return set_orig_insn(&uprobe->arch, mm, vaddr);
922}
923
924static inline bool uprobe_is_active(struct uprobe *uprobe)
925{
926 return !RB_EMPTY_NODE(&uprobe->rb_node);
927}
928/*
929 * There could be threads that have already hit the breakpoint. They
930 * will recheck the current insn and restart if find_uprobe() fails.
931 * See find_active_uprobe().
932 */
933static void delete_uprobe(struct uprobe *uprobe)
934{
935 if (WARN_ON(!uprobe_is_active(uprobe)))
936 return;
937
938 spin_lock(&uprobes_treelock);
939 rb_erase(&uprobe->rb_node, &uprobes_tree);
940 spin_unlock(&uprobes_treelock);
941 RB_CLEAR_NODE(&uprobe->rb_node); /* for uprobe_is_active() */
942 put_uprobe(uprobe);
943}
944
945struct map_info {
946 struct map_info *next;
947 struct mm_struct *mm;
948 unsigned long vaddr;
949};
950
951static inline struct map_info *free_map_info(struct map_info *info)
952{
953 struct map_info *next = info->next;
954 kfree(info);
955 return next;
956}
957
958static struct map_info *
959build_map_info(struct address_space *mapping, loff_t offset, bool is_register)
960{
961 unsigned long pgoff = offset >> PAGE_SHIFT;
962 struct vm_area_struct *vma;
963 struct map_info *curr = NULL;
964 struct map_info *prev = NULL;
965 struct map_info *info;
966 int more = 0;
967
968 again:
969 i_mmap_lock_read(mapping);
970 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
971 if (!valid_vma(vma, is_register))
972 continue;
973
974 if (!prev && !more) {
975 /*
976 * Needs GFP_NOWAIT to avoid i_mmap_rwsem recursion through
977 * reclaim. This is optimistic, no harm done if it fails.
978 */
979 prev = kmalloc(sizeof(struct map_info),
980 GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
981 if (prev)
982 prev->next = NULL;
983 }
984 if (!prev) {
985 more++;
986 continue;
987 }
988
989 if (!mmget_not_zero(vma->vm_mm))
990 continue;
991
992 info = prev;
993 prev = prev->next;
994 info->next = curr;
995 curr = info;
996
997 info->mm = vma->vm_mm;
998 info->vaddr = offset_to_vaddr(vma, offset);
999 }
1000 i_mmap_unlock_read(mapping);
1001
1002 if (!more)
1003 goto out;
1004
1005 prev = curr;
1006 while (curr) {
1007 mmput(curr->mm);
1008 curr = curr->next;
1009 }
1010
1011 do {
1012 info = kmalloc(sizeof(struct map_info), GFP_KERNEL);
1013 if (!info) {
1014 curr = ERR_PTR(-ENOMEM);
1015 goto out;
1016 }
1017 info->next = prev;
1018 prev = info;
1019 } while (--more);
1020
1021 goto again;
1022 out:
1023 while (prev)
1024 prev = free_map_info(prev);
1025 return curr;
1026}
1027
1028static int
1029register_for_each_vma(struct uprobe *uprobe, struct uprobe_consumer *new)
1030{
1031 bool is_register = !!new;
1032 struct map_info *info;
1033 int err = 0;
1034
1035 percpu_down_write(&dup_mmap_sem);
1036 info = build_map_info(uprobe->inode->i_mapping,
1037 uprobe->offset, is_register);
1038 if (IS_ERR(info)) {
1039 err = PTR_ERR(info);
1040 goto out;
1041 }
1042
1043 while (info) {
1044 struct mm_struct *mm = info->mm;
1045 struct vm_area_struct *vma;
1046
1047 if (err && is_register)
1048 goto free;
1049
1050 mmap_write_lock(mm);
1051 vma = find_vma(mm, info->vaddr);
1052 if (!vma || !valid_vma(vma, is_register) ||
1053 file_inode(vma->vm_file) != uprobe->inode)
1054 goto unlock;
1055
1056 if (vma->vm_start > info->vaddr ||
1057 vaddr_to_offset(vma, info->vaddr) != uprobe->offset)
1058 goto unlock;
1059
1060 if (is_register) {
1061 /* consult only the "caller", new consumer. */
1062 if (consumer_filter(new,
1063 UPROBE_FILTER_REGISTER, mm))
1064 err = install_breakpoint(uprobe, mm, vma, info->vaddr);
1065 } else if (test_bit(MMF_HAS_UPROBES, &mm->flags)) {
1066 if (!filter_chain(uprobe,
1067 UPROBE_FILTER_UNREGISTER, mm))
1068 err |= remove_breakpoint(uprobe, mm, info->vaddr);
1069 }
1070
1071 unlock:
1072 mmap_write_unlock(mm);
1073 free:
1074 mmput(mm);
1075 info = free_map_info(info);
1076 }
1077 out:
1078 percpu_up_write(&dup_mmap_sem);
1079 return err;
1080}
1081
1082static void
1083__uprobe_unregister(struct uprobe *uprobe, struct uprobe_consumer *uc)
1084{
1085 int err;
1086
1087 if (WARN_ON(!consumer_del(uprobe, uc)))
1088 return;
1089
1090 err = register_for_each_vma(uprobe, NULL);
1091 /* TODO : cant unregister? schedule a worker thread */
1092 if (!uprobe->consumers && !err)
1093 delete_uprobe(uprobe);
1094}
1095
1096/*
1097 * uprobe_unregister - unregister an already registered probe.
1098 * @inode: the file in which the probe has to be removed.
1099 * @offset: offset from the start of the file.
1100 * @uc: identify which probe if multiple probes are colocated.
1101 */
1102void uprobe_unregister(struct inode *inode, loff_t offset, struct uprobe_consumer *uc)
1103{
1104 struct uprobe *uprobe;
1105
1106 uprobe = find_uprobe(inode, offset);
1107 if (WARN_ON(!uprobe))
1108 return;
1109
1110 down_write(&uprobe->register_rwsem);
1111 __uprobe_unregister(uprobe, uc);
1112 up_write(&uprobe->register_rwsem);
1113 put_uprobe(uprobe);
1114}
1115EXPORT_SYMBOL_GPL(uprobe_unregister);
1116
1117/*
1118 * __uprobe_register - register a probe
1119 * @inode: the file in which the probe has to be placed.
1120 * @offset: offset from the start of the file.
1121 * @uc: information on howto handle the probe..
1122 *
1123 * Apart from the access refcount, __uprobe_register() takes a creation
1124 * refcount (thro alloc_uprobe) if and only if this @uprobe is getting
1125 * inserted into the rbtree (i.e first consumer for a @inode:@offset
1126 * tuple). Creation refcount stops uprobe_unregister from freeing the
1127 * @uprobe even before the register operation is complete. Creation
1128 * refcount is released when the last @uc for the @uprobe
1129 * unregisters. Caller of __uprobe_register() is required to keep @inode
1130 * (and the containing mount) referenced.
1131 *
1132 * Return errno if it cannot successully install probes
1133 * else return 0 (success)
1134 */
1135static int __uprobe_register(struct inode *inode, loff_t offset,
1136 loff_t ref_ctr_offset, struct uprobe_consumer *uc)
1137{
1138 struct uprobe *uprobe;
1139 int ret;
1140
1141 /* Uprobe must have at least one set consumer */
1142 if (!uc->handler && !uc->ret_handler)
1143 return -EINVAL;
1144
1145 /* copy_insn() uses read_mapping_page() or shmem_read_mapping_page() */
1146 if (!inode->i_mapping->a_ops->read_folio &&
1147 !shmem_mapping(inode->i_mapping))
1148 return -EIO;
1149 /* Racy, just to catch the obvious mistakes */
1150 if (offset > i_size_read(inode))
1151 return -EINVAL;
1152
1153 /*
1154 * This ensures that copy_from_page(), copy_to_page() and
1155 * __update_ref_ctr() can't cross page boundary.
1156 */
1157 if (!IS_ALIGNED(offset, UPROBE_SWBP_INSN_SIZE))
1158 return -EINVAL;
1159 if (!IS_ALIGNED(ref_ctr_offset, sizeof(short)))
1160 return -EINVAL;
1161
1162 retry:
1163 uprobe = alloc_uprobe(inode, offset, ref_ctr_offset);
1164 if (!uprobe)
1165 return -ENOMEM;
1166 if (IS_ERR(uprobe))
1167 return PTR_ERR(uprobe);
1168
1169 /*
1170 * We can race with uprobe_unregister()->delete_uprobe().
1171 * Check uprobe_is_active() and retry if it is false.
1172 */
1173 down_write(&uprobe->register_rwsem);
1174 ret = -EAGAIN;
1175 if (likely(uprobe_is_active(uprobe))) {
1176 consumer_add(uprobe, uc);
1177 ret = register_for_each_vma(uprobe, uc);
1178 if (ret)
1179 __uprobe_unregister(uprobe, uc);
1180 }
1181 up_write(&uprobe->register_rwsem);
1182 put_uprobe(uprobe);
1183
1184 if (unlikely(ret == -EAGAIN))
1185 goto retry;
1186 return ret;
1187}
1188
1189int uprobe_register(struct inode *inode, loff_t offset,
1190 struct uprobe_consumer *uc)
1191{
1192 return __uprobe_register(inode, offset, 0, uc);
1193}
1194EXPORT_SYMBOL_GPL(uprobe_register);
1195
1196int uprobe_register_refctr(struct inode *inode, loff_t offset,
1197 loff_t ref_ctr_offset, struct uprobe_consumer *uc)
1198{
1199 return __uprobe_register(inode, offset, ref_ctr_offset, uc);
1200}
1201EXPORT_SYMBOL_GPL(uprobe_register_refctr);
1202
1203/*
1204 * uprobe_apply - unregister an already registered probe.
1205 * @inode: the file in which the probe has to be removed.
1206 * @offset: offset from the start of the file.
1207 * @uc: consumer which wants to add more or remove some breakpoints
1208 * @add: add or remove the breakpoints
1209 */
1210int uprobe_apply(struct inode *inode, loff_t offset,
1211 struct uprobe_consumer *uc, bool add)
1212{
1213 struct uprobe *uprobe;
1214 struct uprobe_consumer *con;
1215 int ret = -ENOENT;
1216
1217 uprobe = find_uprobe(inode, offset);
1218 if (WARN_ON(!uprobe))
1219 return ret;
1220
1221 down_write(&uprobe->register_rwsem);
1222 for (con = uprobe->consumers; con && con != uc ; con = con->next)
1223 ;
1224 if (con)
1225 ret = register_for_each_vma(uprobe, add ? uc : NULL);
1226 up_write(&uprobe->register_rwsem);
1227 put_uprobe(uprobe);
1228
1229 return ret;
1230}
1231
1232static int unapply_uprobe(struct uprobe *uprobe, struct mm_struct *mm)
1233{
1234 VMA_ITERATOR(vmi, mm, 0);
1235 struct vm_area_struct *vma;
1236 int err = 0;
1237
1238 mmap_read_lock(mm);
1239 for_each_vma(vmi, vma) {
1240 unsigned long vaddr;
1241 loff_t offset;
1242
1243 if (!valid_vma(vma, false) ||
1244 file_inode(vma->vm_file) != uprobe->inode)
1245 continue;
1246
1247 offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1248 if (uprobe->offset < offset ||
1249 uprobe->offset >= offset + vma->vm_end - vma->vm_start)
1250 continue;
1251
1252 vaddr = offset_to_vaddr(vma, uprobe->offset);
1253 err |= remove_breakpoint(uprobe, mm, vaddr);
1254 }
1255 mmap_read_unlock(mm);
1256
1257 return err;
1258}
1259
1260static struct rb_node *
1261find_node_in_range(struct inode *inode, loff_t min, loff_t max)
1262{
1263 struct rb_node *n = uprobes_tree.rb_node;
1264
1265 while (n) {
1266 struct uprobe *u = rb_entry(n, struct uprobe, rb_node);
1267
1268 if (inode < u->inode) {
1269 n = n->rb_left;
1270 } else if (inode > u->inode) {
1271 n = n->rb_right;
1272 } else {
1273 if (max < u->offset)
1274 n = n->rb_left;
1275 else if (min > u->offset)
1276 n = n->rb_right;
1277 else
1278 break;
1279 }
1280 }
1281
1282 return n;
1283}
1284
1285/*
1286 * For a given range in vma, build a list of probes that need to be inserted.
1287 */
1288static void build_probe_list(struct inode *inode,
1289 struct vm_area_struct *vma,
1290 unsigned long start, unsigned long end,
1291 struct list_head *head)
1292{
1293 loff_t min, max;
1294 struct rb_node *n, *t;
1295 struct uprobe *u;
1296
1297 INIT_LIST_HEAD(head);
1298 min = vaddr_to_offset(vma, start);
1299 max = min + (end - start) - 1;
1300
1301 spin_lock(&uprobes_treelock);
1302 n = find_node_in_range(inode, min, max);
1303 if (n) {
1304 for (t = n; t; t = rb_prev(t)) {
1305 u = rb_entry(t, struct uprobe, rb_node);
1306 if (u->inode != inode || u->offset < min)
1307 break;
1308 list_add(&u->pending_list, head);
1309 get_uprobe(u);
1310 }
1311 for (t = n; (t = rb_next(t)); ) {
1312 u = rb_entry(t, struct uprobe, rb_node);
1313 if (u->inode != inode || u->offset > max)
1314 break;
1315 list_add(&u->pending_list, head);
1316 get_uprobe(u);
1317 }
1318 }
1319 spin_unlock(&uprobes_treelock);
1320}
1321
1322/* @vma contains reference counter, not the probed instruction. */
1323static int delayed_ref_ctr_inc(struct vm_area_struct *vma)
1324{
1325 struct list_head *pos, *q;
1326 struct delayed_uprobe *du;
1327 unsigned long vaddr;
1328 int ret = 0, err = 0;
1329
1330 mutex_lock(&delayed_uprobe_lock);
1331 list_for_each_safe(pos, q, &delayed_uprobe_list) {
1332 du = list_entry(pos, struct delayed_uprobe, list);
1333
1334 if (du->mm != vma->vm_mm ||
1335 !valid_ref_ctr_vma(du->uprobe, vma))
1336 continue;
1337
1338 vaddr = offset_to_vaddr(vma, du->uprobe->ref_ctr_offset);
1339 ret = __update_ref_ctr(vma->vm_mm, vaddr, 1);
1340 if (ret) {
1341 update_ref_ctr_warn(du->uprobe, vma->vm_mm, 1);
1342 if (!err)
1343 err = ret;
1344 }
1345 delayed_uprobe_delete(du);
1346 }
1347 mutex_unlock(&delayed_uprobe_lock);
1348 return err;
1349}
1350
1351/*
1352 * Called from mmap_region/vma_merge with mm->mmap_lock acquired.
1353 *
1354 * Currently we ignore all errors and always return 0, the callers
1355 * can't handle the failure anyway.
1356 */
1357int uprobe_mmap(struct vm_area_struct *vma)
1358{
1359 struct list_head tmp_list;
1360 struct uprobe *uprobe, *u;
1361 struct inode *inode;
1362
1363 if (no_uprobe_events())
1364 return 0;
1365
1366 if (vma->vm_file &&
1367 (vma->vm_flags & (VM_WRITE|VM_SHARED)) == VM_WRITE &&
1368 test_bit(MMF_HAS_UPROBES, &vma->vm_mm->flags))
1369 delayed_ref_ctr_inc(vma);
1370
1371 if (!valid_vma(vma, true))
1372 return 0;
1373
1374 inode = file_inode(vma->vm_file);
1375 if (!inode)
1376 return 0;
1377
1378 mutex_lock(uprobes_mmap_hash(inode));
1379 build_probe_list(inode, vma, vma->vm_start, vma->vm_end, &tmp_list);
1380 /*
1381 * We can race with uprobe_unregister(), this uprobe can be already
1382 * removed. But in this case filter_chain() must return false, all
1383 * consumers have gone away.
1384 */
1385 list_for_each_entry_safe(uprobe, u, &tmp_list, pending_list) {
1386 if (!fatal_signal_pending(current) &&
1387 filter_chain(uprobe, UPROBE_FILTER_MMAP, vma->vm_mm)) {
1388 unsigned long vaddr = offset_to_vaddr(vma, uprobe->offset);
1389 install_breakpoint(uprobe, vma->vm_mm, vma, vaddr);
1390 }
1391 put_uprobe(uprobe);
1392 }
1393 mutex_unlock(uprobes_mmap_hash(inode));
1394
1395 return 0;
1396}
1397
1398static bool
1399vma_has_uprobes(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1400{
1401 loff_t min, max;
1402 struct inode *inode;
1403 struct rb_node *n;
1404
1405 inode = file_inode(vma->vm_file);
1406
1407 min = vaddr_to_offset(vma, start);
1408 max = min + (end - start) - 1;
1409
1410 spin_lock(&uprobes_treelock);
1411 n = find_node_in_range(inode, min, max);
1412 spin_unlock(&uprobes_treelock);
1413
1414 return !!n;
1415}
1416
1417/*
1418 * Called in context of a munmap of a vma.
1419 */
1420void uprobe_munmap(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1421{
1422 if (no_uprobe_events() || !valid_vma(vma, false))
1423 return;
1424
1425 if (!atomic_read(&vma->vm_mm->mm_users)) /* called by mmput() ? */
1426 return;
1427
1428 if (!test_bit(MMF_HAS_UPROBES, &vma->vm_mm->flags) ||
1429 test_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags))
1430 return;
1431
1432 if (vma_has_uprobes(vma, start, end))
1433 set_bit(MMF_RECALC_UPROBES, &vma->vm_mm->flags);
1434}
1435
1436/* Slot allocation for XOL */
1437static int xol_add_vma(struct mm_struct *mm, struct xol_area *area)
1438{
1439 struct vm_area_struct *vma;
1440 int ret;
1441
1442 if (mmap_write_lock_killable(mm))
1443 return -EINTR;
1444
1445 if (mm->uprobes_state.xol_area) {
1446 ret = -EALREADY;
1447 goto fail;
1448 }
1449
1450 if (!area->vaddr) {
1451 /* Try to map as high as possible, this is only a hint. */
1452 area->vaddr = get_unmapped_area(NULL, TASK_SIZE - PAGE_SIZE,
1453 PAGE_SIZE, 0, 0);
1454 if (IS_ERR_VALUE(area->vaddr)) {
1455 ret = area->vaddr;
1456 goto fail;
1457 }
1458 }
1459
1460 vma = _install_special_mapping(mm, area->vaddr, PAGE_SIZE,
1461 VM_EXEC|VM_MAYEXEC|VM_DONTCOPY|VM_IO,
1462 &area->xol_mapping);
1463 if (IS_ERR(vma)) {
1464 ret = PTR_ERR(vma);
1465 goto fail;
1466 }
1467
1468 ret = 0;
1469 /* pairs with get_xol_area() */
1470 smp_store_release(&mm->uprobes_state.xol_area, area); /* ^^^ */
1471 fail:
1472 mmap_write_unlock(mm);
1473
1474 return ret;
1475}
1476
1477static struct xol_area *__create_xol_area(unsigned long vaddr)
1478{
1479 struct mm_struct *mm = current->mm;
1480 uprobe_opcode_t insn = UPROBE_SWBP_INSN;
1481 struct xol_area *area;
1482
1483 area = kmalloc(sizeof(*area), GFP_KERNEL);
1484 if (unlikely(!area))
1485 goto out;
1486
1487 area->bitmap = kcalloc(BITS_TO_LONGS(UINSNS_PER_PAGE), sizeof(long),
1488 GFP_KERNEL);
1489 if (!area->bitmap)
1490 goto free_area;
1491
1492 area->xol_mapping.name = "[uprobes]";
1493 area->xol_mapping.fault = NULL;
1494 area->xol_mapping.pages = area->pages;
1495 area->pages[0] = alloc_page(GFP_HIGHUSER);
1496 if (!area->pages[0])
1497 goto free_bitmap;
1498 area->pages[1] = NULL;
1499
1500 area->vaddr = vaddr;
1501 init_waitqueue_head(&area->wq);
1502 /* Reserve the 1st slot for get_trampoline_vaddr() */
1503 set_bit(0, area->bitmap);
1504 atomic_set(&area->slot_count, 1);
1505 arch_uprobe_copy_ixol(area->pages[0], 0, &insn, UPROBE_SWBP_INSN_SIZE);
1506
1507 if (!xol_add_vma(mm, area))
1508 return area;
1509
1510 __free_page(area->pages[0]);
1511 free_bitmap:
1512 kfree(area->bitmap);
1513 free_area:
1514 kfree(area);
1515 out:
1516 return NULL;
1517}
1518
1519/*
1520 * get_xol_area - Allocate process's xol_area if necessary.
1521 * This area will be used for storing instructions for execution out of line.
1522 *
1523 * Returns the allocated area or NULL.
1524 */
1525static struct xol_area *get_xol_area(void)
1526{
1527 struct mm_struct *mm = current->mm;
1528 struct xol_area *area;
1529
1530 if (!mm->uprobes_state.xol_area)
1531 __create_xol_area(0);
1532
1533 /* Pairs with xol_add_vma() smp_store_release() */
1534 area = READ_ONCE(mm->uprobes_state.xol_area); /* ^^^ */
1535 return area;
1536}
1537
1538/*
1539 * uprobe_clear_state - Free the area allocated for slots.
1540 */
1541void uprobe_clear_state(struct mm_struct *mm)
1542{
1543 struct xol_area *area = mm->uprobes_state.xol_area;
1544
1545 mutex_lock(&delayed_uprobe_lock);
1546 delayed_uprobe_remove(NULL, mm);
1547 mutex_unlock(&delayed_uprobe_lock);
1548
1549 if (!area)
1550 return;
1551
1552 put_page(area->pages[0]);
1553 kfree(area->bitmap);
1554 kfree(area);
1555}
1556
1557void uprobe_start_dup_mmap(void)
1558{
1559 percpu_down_read(&dup_mmap_sem);
1560}
1561
1562void uprobe_end_dup_mmap(void)
1563{
1564 percpu_up_read(&dup_mmap_sem);
1565}
1566
1567void uprobe_dup_mmap(struct mm_struct *oldmm, struct mm_struct *newmm)
1568{
1569 if (test_bit(MMF_HAS_UPROBES, &oldmm->flags)) {
1570 set_bit(MMF_HAS_UPROBES, &newmm->flags);
1571 /* unconditionally, dup_mmap() skips VM_DONTCOPY vmas */
1572 set_bit(MMF_RECALC_UPROBES, &newmm->flags);
1573 }
1574}
1575
1576/*
1577 * - search for a free slot.
1578 */
1579static unsigned long xol_take_insn_slot(struct xol_area *area)
1580{
1581 unsigned long slot_addr;
1582 int slot_nr;
1583
1584 do {
1585 slot_nr = find_first_zero_bit(area->bitmap, UINSNS_PER_PAGE);
1586 if (slot_nr < UINSNS_PER_PAGE) {
1587 if (!test_and_set_bit(slot_nr, area->bitmap))
1588 break;
1589
1590 slot_nr = UINSNS_PER_PAGE;
1591 continue;
1592 }
1593 wait_event(area->wq, (atomic_read(&area->slot_count) < UINSNS_PER_PAGE));
1594 } while (slot_nr >= UINSNS_PER_PAGE);
1595
1596 slot_addr = area->vaddr + (slot_nr * UPROBE_XOL_SLOT_BYTES);
1597 atomic_inc(&area->slot_count);
1598
1599 return slot_addr;
1600}
1601
1602/*
1603 * xol_get_insn_slot - allocate a slot for xol.
1604 * Returns the allocated slot address or 0.
1605 */
1606static unsigned long xol_get_insn_slot(struct uprobe *uprobe)
1607{
1608 struct xol_area *area;
1609 unsigned long xol_vaddr;
1610
1611 area = get_xol_area();
1612 if (!area)
1613 return 0;
1614
1615 xol_vaddr = xol_take_insn_slot(area);
1616 if (unlikely(!xol_vaddr))
1617 return 0;
1618
1619 arch_uprobe_copy_ixol(area->pages[0], xol_vaddr,
1620 &uprobe->arch.ixol, sizeof(uprobe->arch.ixol));
1621
1622 return xol_vaddr;
1623}
1624
1625/*
1626 * xol_free_insn_slot - If slot was earlier allocated by
1627 * @xol_get_insn_slot(), make the slot available for
1628 * subsequent requests.
1629 */
1630static void xol_free_insn_slot(struct task_struct *tsk)
1631{
1632 struct xol_area *area;
1633 unsigned long vma_end;
1634 unsigned long slot_addr;
1635
1636 if (!tsk->mm || !tsk->mm->uprobes_state.xol_area || !tsk->utask)
1637 return;
1638
1639 slot_addr = tsk->utask->xol_vaddr;
1640 if (unlikely(!slot_addr))
1641 return;
1642
1643 area = tsk->mm->uprobes_state.xol_area;
1644 vma_end = area->vaddr + PAGE_SIZE;
1645 if (area->vaddr <= slot_addr && slot_addr < vma_end) {
1646 unsigned long offset;
1647 int slot_nr;
1648
1649 offset = slot_addr - area->vaddr;
1650 slot_nr = offset / UPROBE_XOL_SLOT_BYTES;
1651 if (slot_nr >= UINSNS_PER_PAGE)
1652 return;
1653
1654 clear_bit(slot_nr, area->bitmap);
1655 atomic_dec(&area->slot_count);
1656 smp_mb__after_atomic(); /* pairs with prepare_to_wait() */
1657 if (waitqueue_active(&area->wq))
1658 wake_up(&area->wq);
1659
1660 tsk->utask->xol_vaddr = 0;
1661 }
1662}
1663
1664void __weak arch_uprobe_copy_ixol(struct page *page, unsigned long vaddr,
1665 void *src, unsigned long len)
1666{
1667 /* Initialize the slot */
1668 copy_to_page(page, vaddr, src, len);
1669
1670 /*
1671 * We probably need flush_icache_user_page() but it needs vma.
1672 * This should work on most of architectures by default. If
1673 * architecture needs to do something different it can define
1674 * its own version of the function.
1675 */
1676 flush_dcache_page(page);
1677}
1678
1679/**
1680 * uprobe_get_swbp_addr - compute address of swbp given post-swbp regs
1681 * @regs: Reflects the saved state of the task after it has hit a breakpoint
1682 * instruction.
1683 * Return the address of the breakpoint instruction.
1684 */
1685unsigned long __weak uprobe_get_swbp_addr(struct pt_regs *regs)
1686{
1687 return instruction_pointer(regs) - UPROBE_SWBP_INSN_SIZE;
1688}
1689
1690unsigned long uprobe_get_trap_addr(struct pt_regs *regs)
1691{
1692 struct uprobe_task *utask = current->utask;
1693
1694 if (unlikely(utask && utask->active_uprobe))
1695 return utask->vaddr;
1696
1697 return instruction_pointer(regs);
1698}
1699
1700static struct return_instance *free_ret_instance(struct return_instance *ri)
1701{
1702 struct return_instance *next = ri->next;
1703 put_uprobe(ri->uprobe);
1704 kfree(ri);
1705 return next;
1706}
1707
1708/*
1709 * Called with no locks held.
1710 * Called in context of an exiting or an exec-ing thread.
1711 */
1712void uprobe_free_utask(struct task_struct *t)
1713{
1714 struct uprobe_task *utask = t->utask;
1715 struct return_instance *ri;
1716
1717 if (!utask)
1718 return;
1719
1720 if (utask->active_uprobe)
1721 put_uprobe(utask->active_uprobe);
1722
1723 ri = utask->return_instances;
1724 while (ri)
1725 ri = free_ret_instance(ri);
1726
1727 xol_free_insn_slot(t);
1728 kfree(utask);
1729 t->utask = NULL;
1730}
1731
1732/*
1733 * Allocate a uprobe_task object for the task if necessary.
1734 * Called when the thread hits a breakpoint.
1735 *
1736 * Returns:
1737 * - pointer to new uprobe_task on success
1738 * - NULL otherwise
1739 */
1740static struct uprobe_task *get_utask(void)
1741{
1742 if (!current->utask)
1743 current->utask = kzalloc(sizeof(struct uprobe_task), GFP_KERNEL);
1744 return current->utask;
1745}
1746
1747static int dup_utask(struct task_struct *t, struct uprobe_task *o_utask)
1748{
1749 struct uprobe_task *n_utask;
1750 struct return_instance **p, *o, *n;
1751
1752 n_utask = kzalloc(sizeof(struct uprobe_task), GFP_KERNEL);
1753 if (!n_utask)
1754 return -ENOMEM;
1755 t->utask = n_utask;
1756
1757 p = &n_utask->return_instances;
1758 for (o = o_utask->return_instances; o; o = o->next) {
1759 n = kmalloc(sizeof(struct return_instance), GFP_KERNEL);
1760 if (!n)
1761 return -ENOMEM;
1762
1763 *n = *o;
1764 get_uprobe(n->uprobe);
1765 n->next = NULL;
1766
1767 *p = n;
1768 p = &n->next;
1769 n_utask->depth++;
1770 }
1771
1772 return 0;
1773}
1774
1775static void uprobe_warn(struct task_struct *t, const char *msg)
1776{
1777 pr_warn("uprobe: %s:%d failed to %s\n",
1778 current->comm, current->pid, msg);
1779}
1780
1781static void dup_xol_work(struct callback_head *work)
1782{
1783 if (current->flags & PF_EXITING)
1784 return;
1785
1786 if (!__create_xol_area(current->utask->dup_xol_addr) &&
1787 !fatal_signal_pending(current))
1788 uprobe_warn(current, "dup xol area");
1789}
1790
1791/*
1792 * Called in context of a new clone/fork from copy_process.
1793 */
1794void uprobe_copy_process(struct task_struct *t, unsigned long flags)
1795{
1796 struct uprobe_task *utask = current->utask;
1797 struct mm_struct *mm = current->mm;
1798 struct xol_area *area;
1799
1800 t->utask = NULL;
1801
1802 if (!utask || !utask->return_instances)
1803 return;
1804
1805 if (mm == t->mm && !(flags & CLONE_VFORK))
1806 return;
1807
1808 if (dup_utask(t, utask))
1809 return uprobe_warn(t, "dup ret instances");
1810
1811 /* The task can fork() after dup_xol_work() fails */
1812 area = mm->uprobes_state.xol_area;
1813 if (!area)
1814 return uprobe_warn(t, "dup xol area");
1815
1816 if (mm == t->mm)
1817 return;
1818
1819 t->utask->dup_xol_addr = area->vaddr;
1820 init_task_work(&t->utask->dup_xol_work, dup_xol_work);
1821 task_work_add(t, &t->utask->dup_xol_work, TWA_RESUME);
1822}
1823
1824/*
1825 * Current area->vaddr notion assume the trampoline address is always
1826 * equal area->vaddr.
1827 *
1828 * Returns -1 in case the xol_area is not allocated.
1829 */
1830static unsigned long get_trampoline_vaddr(void)
1831{
1832 struct xol_area *area;
1833 unsigned long trampoline_vaddr = -1;
1834
1835 /* Pairs with xol_add_vma() smp_store_release() */
1836 area = READ_ONCE(current->mm->uprobes_state.xol_area); /* ^^^ */
1837 if (area)
1838 trampoline_vaddr = area->vaddr;
1839
1840 return trampoline_vaddr;
1841}
1842
1843static void cleanup_return_instances(struct uprobe_task *utask, bool chained,
1844 struct pt_regs *regs)
1845{
1846 struct return_instance *ri = utask->return_instances;
1847 enum rp_check ctx = chained ? RP_CHECK_CHAIN_CALL : RP_CHECK_CALL;
1848
1849 while (ri && !arch_uretprobe_is_alive(ri, ctx, regs)) {
1850 ri = free_ret_instance(ri);
1851 utask->depth--;
1852 }
1853 utask->return_instances = ri;
1854}
1855
1856static void prepare_uretprobe(struct uprobe *uprobe, struct pt_regs *regs)
1857{
1858 struct return_instance *ri;
1859 struct uprobe_task *utask;
1860 unsigned long orig_ret_vaddr, trampoline_vaddr;
1861 bool chained;
1862
1863 if (!get_xol_area())
1864 return;
1865
1866 utask = get_utask();
1867 if (!utask)
1868 return;
1869
1870 if (utask->depth >= MAX_URETPROBE_DEPTH) {
1871 printk_ratelimited(KERN_INFO "uprobe: omit uretprobe due to"
1872 " nestedness limit pid/tgid=%d/%d\n",
1873 current->pid, current->tgid);
1874 return;
1875 }
1876
1877 ri = kmalloc(sizeof(struct return_instance), GFP_KERNEL);
1878 if (!ri)
1879 return;
1880
1881 trampoline_vaddr = get_trampoline_vaddr();
1882 orig_ret_vaddr = arch_uretprobe_hijack_return_addr(trampoline_vaddr, regs);
1883 if (orig_ret_vaddr == -1)
1884 goto fail;
1885
1886 /* drop the entries invalidated by longjmp() */
1887 chained = (orig_ret_vaddr == trampoline_vaddr);
1888 cleanup_return_instances(utask, chained, regs);
1889
1890 /*
1891 * We don't want to keep trampoline address in stack, rather keep the
1892 * original return address of first caller thru all the consequent
1893 * instances. This also makes breakpoint unwrapping easier.
1894 */
1895 if (chained) {
1896 if (!utask->return_instances) {
1897 /*
1898 * This situation is not possible. Likely we have an
1899 * attack from user-space.
1900 */
1901 uprobe_warn(current, "handle tail call");
1902 goto fail;
1903 }
1904 orig_ret_vaddr = utask->return_instances->orig_ret_vaddr;
1905 }
1906
1907 ri->uprobe = get_uprobe(uprobe);
1908 ri->func = instruction_pointer(regs);
1909 ri->stack = user_stack_pointer(regs);
1910 ri->orig_ret_vaddr = orig_ret_vaddr;
1911 ri->chained = chained;
1912
1913 utask->depth++;
1914 ri->next = utask->return_instances;
1915 utask->return_instances = ri;
1916
1917 return;
1918 fail:
1919 kfree(ri);
1920}
1921
1922/* Prepare to single-step probed instruction out of line. */
1923static int
1924pre_ssout(struct uprobe *uprobe, struct pt_regs *regs, unsigned long bp_vaddr)
1925{
1926 struct uprobe_task *utask;
1927 unsigned long xol_vaddr;
1928 int err;
1929
1930 utask = get_utask();
1931 if (!utask)
1932 return -ENOMEM;
1933
1934 xol_vaddr = xol_get_insn_slot(uprobe);
1935 if (!xol_vaddr)
1936 return -ENOMEM;
1937
1938 utask->xol_vaddr = xol_vaddr;
1939 utask->vaddr = bp_vaddr;
1940
1941 err = arch_uprobe_pre_xol(&uprobe->arch, regs);
1942 if (unlikely(err)) {
1943 xol_free_insn_slot(current);
1944 return err;
1945 }
1946
1947 utask->active_uprobe = uprobe;
1948 utask->state = UTASK_SSTEP;
1949 return 0;
1950}
1951
1952/*
1953 * If we are singlestepping, then ensure this thread is not connected to
1954 * non-fatal signals until completion of singlestep. When xol insn itself
1955 * triggers the signal, restart the original insn even if the task is
1956 * already SIGKILL'ed (since coredump should report the correct ip). This
1957 * is even more important if the task has a handler for SIGSEGV/etc, The
1958 * _same_ instruction should be repeated again after return from the signal
1959 * handler, and SSTEP can never finish in this case.
1960 */
1961bool uprobe_deny_signal(void)
1962{
1963 struct task_struct *t = current;
1964 struct uprobe_task *utask = t->utask;
1965
1966 if (likely(!utask || !utask->active_uprobe))
1967 return false;
1968
1969 WARN_ON_ONCE(utask->state != UTASK_SSTEP);
1970
1971 if (task_sigpending(t)) {
1972 spin_lock_irq(&t->sighand->siglock);
1973 clear_tsk_thread_flag(t, TIF_SIGPENDING);
1974 spin_unlock_irq(&t->sighand->siglock);
1975
1976 if (__fatal_signal_pending(t) || arch_uprobe_xol_was_trapped(t)) {
1977 utask->state = UTASK_SSTEP_TRAPPED;
1978 set_tsk_thread_flag(t, TIF_UPROBE);
1979 }
1980 }
1981
1982 return true;
1983}
1984
1985static void mmf_recalc_uprobes(struct mm_struct *mm)
1986{
1987 VMA_ITERATOR(vmi, mm, 0);
1988 struct vm_area_struct *vma;
1989
1990 for_each_vma(vmi, vma) {
1991 if (!valid_vma(vma, false))
1992 continue;
1993 /*
1994 * This is not strictly accurate, we can race with
1995 * uprobe_unregister() and see the already removed
1996 * uprobe if delete_uprobe() was not yet called.
1997 * Or this uprobe can be filtered out.
1998 */
1999 if (vma_has_uprobes(vma, vma->vm_start, vma->vm_end))
2000 return;
2001 }
2002
2003 clear_bit(MMF_HAS_UPROBES, &mm->flags);
2004}
2005
2006static int is_trap_at_addr(struct mm_struct *mm, unsigned long vaddr)
2007{
2008 struct page *page;
2009 uprobe_opcode_t opcode;
2010 int result;
2011
2012 if (WARN_ON_ONCE(!IS_ALIGNED(vaddr, UPROBE_SWBP_INSN_SIZE)))
2013 return -EINVAL;
2014
2015 pagefault_disable();
2016 result = __get_user(opcode, (uprobe_opcode_t __user *)vaddr);
2017 pagefault_enable();
2018
2019 if (likely(result == 0))
2020 goto out;
2021
2022 /*
2023 * The NULL 'tsk' here ensures that any faults that occur here
2024 * will not be accounted to the task. 'mm' *is* current->mm,
2025 * but we treat this as a 'remote' access since it is
2026 * essentially a kernel access to the memory.
2027 */
2028 result = get_user_pages_remote(mm, vaddr, 1, FOLL_FORCE, &page, NULL);
2029 if (result < 0)
2030 return result;
2031
2032 copy_from_page(page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
2033 put_page(page);
2034 out:
2035 /* This needs to return true for any variant of the trap insn */
2036 return is_trap_insn(&opcode);
2037}
2038
2039static struct uprobe *find_active_uprobe(unsigned long bp_vaddr, int *is_swbp)
2040{
2041 struct mm_struct *mm = current->mm;
2042 struct uprobe *uprobe = NULL;
2043 struct vm_area_struct *vma;
2044
2045 mmap_read_lock(mm);
2046 vma = vma_lookup(mm, bp_vaddr);
2047 if (vma) {
2048 if (valid_vma(vma, false)) {
2049 struct inode *inode = file_inode(vma->vm_file);
2050 loff_t offset = vaddr_to_offset(vma, bp_vaddr);
2051
2052 uprobe = find_uprobe(inode, offset);
2053 }
2054
2055 if (!uprobe)
2056 *is_swbp = is_trap_at_addr(mm, bp_vaddr);
2057 } else {
2058 *is_swbp = -EFAULT;
2059 }
2060
2061 if (!uprobe && test_and_clear_bit(MMF_RECALC_UPROBES, &mm->flags))
2062 mmf_recalc_uprobes(mm);
2063 mmap_read_unlock(mm);
2064
2065 return uprobe;
2066}
2067
2068static void handler_chain(struct uprobe *uprobe, struct pt_regs *regs)
2069{
2070 struct uprobe_consumer *uc;
2071 int remove = UPROBE_HANDLER_REMOVE;
2072 bool need_prep = false; /* prepare return uprobe, when needed */
2073
2074 down_read(&uprobe->register_rwsem);
2075 for (uc = uprobe->consumers; uc; uc = uc->next) {
2076 int rc = 0;
2077
2078 if (uc->handler) {
2079 rc = uc->handler(uc, regs);
2080 WARN(rc & ~UPROBE_HANDLER_MASK,
2081 "bad rc=0x%x from %ps()\n", rc, uc->handler);
2082 }
2083
2084 if (uc->ret_handler)
2085 need_prep = true;
2086
2087 remove &= rc;
2088 }
2089
2090 if (need_prep && !remove)
2091 prepare_uretprobe(uprobe, regs); /* put bp at return */
2092
2093 if (remove && uprobe->consumers) {
2094 WARN_ON(!uprobe_is_active(uprobe));
2095 unapply_uprobe(uprobe, current->mm);
2096 }
2097 up_read(&uprobe->register_rwsem);
2098}
2099
2100static void
2101handle_uretprobe_chain(struct return_instance *ri, struct pt_regs *regs)
2102{
2103 struct uprobe *uprobe = ri->uprobe;
2104 struct uprobe_consumer *uc;
2105
2106 down_read(&uprobe->register_rwsem);
2107 for (uc = uprobe->consumers; uc; uc = uc->next) {
2108 if (uc->ret_handler)
2109 uc->ret_handler(uc, ri->func, regs);
2110 }
2111 up_read(&uprobe->register_rwsem);
2112}
2113
2114static struct return_instance *find_next_ret_chain(struct return_instance *ri)
2115{
2116 bool chained;
2117
2118 do {
2119 chained = ri->chained;
2120 ri = ri->next; /* can't be NULL if chained */
2121 } while (chained);
2122
2123 return ri;
2124}
2125
2126static void handle_trampoline(struct pt_regs *regs)
2127{
2128 struct uprobe_task *utask;
2129 struct return_instance *ri, *next;
2130 bool valid;
2131
2132 utask = current->utask;
2133 if (!utask)
2134 goto sigill;
2135
2136 ri = utask->return_instances;
2137 if (!ri)
2138 goto sigill;
2139
2140 do {
2141 /*
2142 * We should throw out the frames invalidated by longjmp().
2143 * If this chain is valid, then the next one should be alive
2144 * or NULL; the latter case means that nobody but ri->func
2145 * could hit this trampoline on return. TODO: sigaltstack().
2146 */
2147 next = find_next_ret_chain(ri);
2148 valid = !next || arch_uretprobe_is_alive(next, RP_CHECK_RET, regs);
2149
2150 instruction_pointer_set(regs, ri->orig_ret_vaddr);
2151 do {
2152 if (valid)
2153 handle_uretprobe_chain(ri, regs);
2154 ri = free_ret_instance(ri);
2155 utask->depth--;
2156 } while (ri != next);
2157 } while (!valid);
2158
2159 utask->return_instances = ri;
2160 return;
2161
2162 sigill:
2163 uprobe_warn(current, "handle uretprobe, sending SIGILL.");
2164 force_sig(SIGILL);
2165
2166}
2167
2168bool __weak arch_uprobe_ignore(struct arch_uprobe *aup, struct pt_regs *regs)
2169{
2170 return false;
2171}
2172
2173bool __weak arch_uretprobe_is_alive(struct return_instance *ret, enum rp_check ctx,
2174 struct pt_regs *regs)
2175{
2176 return true;
2177}
2178
2179/*
2180 * Run handler and ask thread to singlestep.
2181 * Ensure all non-fatal signals cannot interrupt thread while it singlesteps.
2182 */
2183static void handle_swbp(struct pt_regs *regs)
2184{
2185 struct uprobe *uprobe;
2186 unsigned long bp_vaddr;
2187 int is_swbp;
2188
2189 bp_vaddr = uprobe_get_swbp_addr(regs);
2190 if (bp_vaddr == get_trampoline_vaddr())
2191 return handle_trampoline(regs);
2192
2193 uprobe = find_active_uprobe(bp_vaddr, &is_swbp);
2194 if (!uprobe) {
2195 if (is_swbp > 0) {
2196 /* No matching uprobe; signal SIGTRAP. */
2197 force_sig(SIGTRAP);
2198 } else {
2199 /*
2200 * Either we raced with uprobe_unregister() or we can't
2201 * access this memory. The latter is only possible if
2202 * another thread plays with our ->mm. In both cases
2203 * we can simply restart. If this vma was unmapped we
2204 * can pretend this insn was not executed yet and get
2205 * the (correct) SIGSEGV after restart.
2206 */
2207 instruction_pointer_set(regs, bp_vaddr);
2208 }
2209 return;
2210 }
2211
2212 /* change it in advance for ->handler() and restart */
2213 instruction_pointer_set(regs, bp_vaddr);
2214
2215 /*
2216 * TODO: move copy_insn/etc into _register and remove this hack.
2217 * After we hit the bp, _unregister + _register can install the
2218 * new and not-yet-analyzed uprobe at the same address, restart.
2219 */
2220 if (unlikely(!test_bit(UPROBE_COPY_INSN, &uprobe->flags)))
2221 goto out;
2222
2223 /*
2224 * Pairs with the smp_wmb() in prepare_uprobe().
2225 *
2226 * Guarantees that if we see the UPROBE_COPY_INSN bit set, then
2227 * we must also see the stores to &uprobe->arch performed by the
2228 * prepare_uprobe() call.
2229 */
2230 smp_rmb();
2231
2232 /* Tracing handlers use ->utask to communicate with fetch methods */
2233 if (!get_utask())
2234 goto out;
2235
2236 if (arch_uprobe_ignore(&uprobe->arch, regs))
2237 goto out;
2238
2239 handler_chain(uprobe, regs);
2240
2241 if (arch_uprobe_skip_sstep(&uprobe->arch, regs))
2242 goto out;
2243
2244 if (!pre_ssout(uprobe, regs, bp_vaddr))
2245 return;
2246
2247 /* arch_uprobe_skip_sstep() succeeded, or restart if can't singlestep */
2248out:
2249 put_uprobe(uprobe);
2250}
2251
2252/*
2253 * Perform required fix-ups and disable singlestep.
2254 * Allow pending signals to take effect.
2255 */
2256static void handle_singlestep(struct uprobe_task *utask, struct pt_regs *regs)
2257{
2258 struct uprobe *uprobe;
2259 int err = 0;
2260
2261 uprobe = utask->active_uprobe;
2262 if (utask->state == UTASK_SSTEP_ACK)
2263 err = arch_uprobe_post_xol(&uprobe->arch, regs);
2264 else if (utask->state == UTASK_SSTEP_TRAPPED)
2265 arch_uprobe_abort_xol(&uprobe->arch, regs);
2266 else
2267 WARN_ON_ONCE(1);
2268
2269 put_uprobe(uprobe);
2270 utask->active_uprobe = NULL;
2271 utask->state = UTASK_RUNNING;
2272 xol_free_insn_slot(current);
2273
2274 spin_lock_irq(¤t->sighand->siglock);
2275 recalc_sigpending(); /* see uprobe_deny_signal() */
2276 spin_unlock_irq(¤t->sighand->siglock);
2277
2278 if (unlikely(err)) {
2279 uprobe_warn(current, "execute the probed insn, sending SIGILL.");
2280 force_sig(SIGILL);
2281 }
2282}
2283
2284/*
2285 * On breakpoint hit, breakpoint notifier sets the TIF_UPROBE flag and
2286 * allows the thread to return from interrupt. After that handle_swbp()
2287 * sets utask->active_uprobe.
2288 *
2289 * On singlestep exception, singlestep notifier sets the TIF_UPROBE flag
2290 * and allows the thread to return from interrupt.
2291 *
2292 * While returning to userspace, thread notices the TIF_UPROBE flag and calls
2293 * uprobe_notify_resume().
2294 */
2295void uprobe_notify_resume(struct pt_regs *regs)
2296{
2297 struct uprobe_task *utask;
2298
2299 clear_thread_flag(TIF_UPROBE);
2300
2301 utask = current->utask;
2302 if (utask && utask->active_uprobe)
2303 handle_singlestep(utask, regs);
2304 else
2305 handle_swbp(regs);
2306}
2307
2308/*
2309 * uprobe_pre_sstep_notifier gets called from interrupt context as part of
2310 * notifier mechanism. Set TIF_UPROBE flag and indicate breakpoint hit.
2311 */
2312int uprobe_pre_sstep_notifier(struct pt_regs *regs)
2313{
2314 if (!current->mm)
2315 return 0;
2316
2317 if (!test_bit(MMF_HAS_UPROBES, ¤t->mm->flags) &&
2318 (!current->utask || !current->utask->return_instances))
2319 return 0;
2320
2321 set_thread_flag(TIF_UPROBE);
2322 return 1;
2323}
2324
2325/*
2326 * uprobe_post_sstep_notifier gets called in interrupt context as part of notifier
2327 * mechanism. Set TIF_UPROBE flag and indicate completion of singlestep.
2328 */
2329int uprobe_post_sstep_notifier(struct pt_regs *regs)
2330{
2331 struct uprobe_task *utask = current->utask;
2332
2333 if (!current->mm || !utask || !utask->active_uprobe)
2334 /* task is currently not uprobed */
2335 return 0;
2336
2337 utask->state = UTASK_SSTEP_ACK;
2338 set_thread_flag(TIF_UPROBE);
2339 return 1;
2340}
2341
2342static struct notifier_block uprobe_exception_nb = {
2343 .notifier_call = arch_uprobe_exception_notify,
2344 .priority = INT_MAX-1, /* notified after kprobes, kgdb */
2345};
2346
2347void __init uprobes_init(void)
2348{
2349 int i;
2350
2351 for (i = 0; i < UPROBES_HASH_SZ; i++)
2352 mutex_init(&uprobes_mmap_mutex[i]);
2353
2354 BUG_ON(register_die_notifier(&uprobe_exception_nb));
2355}