linux - kernel/events/uprobes.c

2912 lines · 74.2 KiB

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	#include <linux/pagewalk.h>
33
34	#include <linux/uprobes.h>
35
36	#define UINSNS_PER_PAGE (PAGE_SIZE/UPROBE_XOL_SLOT_BYTES)
37	#define MAX_UPROBE_XOL_SLOTS UINSNS_PER_PAGE
38
39	static struct rb_root uprobes_tree = RB_ROOT;
40	/*
41	* allows us to skip the uprobe_mmap if there are no uprobe events active
42	* at this time. Probably a fine grained per inode count is better?
43	*/
44	#define no_uprobe_events() RB_EMPTY_ROOT(&uprobes_tree)
45
46	static DEFINE_RWLOCK(uprobes_treelock); /* serialize rbtree access */
47	static seqcount_rwlock_t uprobes_seqcount = SEQCNT_RWLOCK_ZERO(uprobes_seqcount, &uprobes_treelock);
48
49	#define UPROBES_HASH_SZ 13
50	/* serialize uprobe->pending_list */
51	static struct mutex uprobes_mmap_mutex[UPROBES_HASH_SZ];
52	#define uprobes_mmap_hash(v) (&uprobes_mmap_mutex[((unsigned long)(v)) % UPROBES_HASH_SZ])
53
54	DEFINE_STATIC_PERCPU_RWSEM(dup_mmap_sem);
55
56	/* Covers return_instance's uprobe lifetime. */
57	DEFINE_STATIC_SRCU_FAST_UPDOWN(uretprobes_srcu);
58
59	/* Have a copy of original instruction */
60	#define UPROBE_COPY_INSN 0
61
62	struct uprobe {
63	struct rb_node rb_node; /* node in the rb tree */
64	refcount_t ref;
65	struct rw_semaphore register_rwsem;
66	struct rw_semaphore consumer_rwsem;
67	struct list_head pending_list;
68	struct list_head consumers;
69	struct inode inode; / Also hold a ref to inode */
70	union {
71	struct rcu_head rcu;
72	struct work_struct work;
73	};
74	loff_t offset;
75	loff_t ref_ctr_offset;
76	unsigned long flags; /* "unsigned long" so bitops work */
77
78	/*
79	* The generic code assumes that it has two members of unknown type
80	* owned by the arch-specific code:
81	*
82	* insn - copy_insn() saves the original instruction here for
83	* arch_uprobe_analyze_insn().
84	*
85	* ixol - potentially modified instruction to execute out of
86	* line, copied to xol_area by xol_get_insn_slot().
87	*/
88	struct arch_uprobe arch;
89	};
90
91	struct delayed_uprobe {
92	struct list_head list;
93	struct uprobe *uprobe;
94	struct mm_struct *mm;
95	};
96
97	static DEFINE_MUTEX(delayed_uprobe_lock);
98	static LIST_HEAD(delayed_uprobe_list);
99
100	/*
101	* Execute out of line area: anonymous executable mapping installed
102	* by the probed task to execute the copy of the original instruction
103	* mangled by set_swbp().
104	*
105	* On a breakpoint hit, thread contests for a slot. It frees the
106	* slot after singlestep. Currently a fixed number of slots are
107	* allocated.
108	*/
109	struct xol_area {
110	wait_queue_head_t wq; /* if all slots are busy */
111	unsigned long bitmap; / 0 = free slot */
112
113	struct page *page;
114	/*
115	* We keep the vma's vm_start rather than a pointer to the vma
116	* itself. The probed process or a naughty kernel module could make
117	* the vma go away, and we must handle that reasonably gracefully.
118	*/
119	unsigned long vaddr; /* Page(s) of instruction slots */
120	};
121
122	static void uprobe_warn(struct task_struct t, const char msg)
123	{
124	pr_warn("uprobe: %s:%d failed to %s\n", t->comm, t->pid, msg);
125	}
126
127	/*
128	* valid_vma: Verify if the specified vma is an executable vma
129	* Relax restrictions while unregistering: vm_flags might have
130	* changed after breakpoint was inserted.
131	* - is_register: indicates if we are in register context.
132	* - Return 1 if the specified virtual address is in an
133	* executable vma.
134	*/
135	static bool valid_vma(struct vm_area_struct *vma, bool is_register)
136	{
137	vm_flags_t flags = VM_HUGETLB \| VM_MAYEXEC \| VM_MAYSHARE;
138
139	if (is_register)
140	flags \|= VM_WRITE;
141
142	return vma->vm_file && (vma->vm_flags & flags) == VM_MAYEXEC;
143	}
144
145	static unsigned long offset_to_vaddr(struct vm_area_struct *vma, loff_t offset)
146	{
147	return vma->vm_start + offset -
148	((loff_t)vma_start_pgoff(vma) << PAGE_SHIFT);
149	}
150
151	static loff_t vaddr_to_offset(struct vm_area_struct *vma, unsigned long vaddr)
152	{
153	return ((loff_t)vma_start_pgoff(vma) << PAGE_SHIFT) +
154	(vaddr - vma->vm_start);
155	}
156
157	/**
158	* is_swbp_insn - check if instruction is breakpoint instruction.
159	* @insn: instruction to be checked.
160	* Default implementation of is_swbp_insn
161	* Returns true if @insn is a breakpoint instruction.
162	*/
163	bool __weak is_swbp_insn(uprobe_opcode_t *insn)
164	{
165	return *insn == UPROBE_SWBP_INSN;
166	}
167
168	/**
169	* is_trap_insn - check if instruction is breakpoint instruction.
170	* @insn: instruction to be checked.
171	* Default implementation of is_trap_insn
172	* Returns true if @insn is a breakpoint instruction.
173	*
174	* This function is needed for the case where an architecture has multiple
175	* trap instructions (like powerpc).
176	*/
177	bool __weak is_trap_insn(uprobe_opcode_t *insn)
178	{
179	return is_swbp_insn(insn);
180	}
181
182	void uprobe_copy_from_page(struct page page, unsigned long vaddr, void dst, int len)
183	{
184	void *kaddr = kmap_local_page(page);
185	memcpy(dst, kaddr + (vaddr & ~PAGE_MASK), len);
186	kunmap_local(kaddr);
187	}
188
189	static void copy_to_page(struct page page, unsigned long vaddr, const void src, int len)
190	{
191	void *kaddr = kmap_local_page(page);
192	memcpy(kaddr + (vaddr & ~PAGE_MASK), src, len);
193	kunmap_local(kaddr);
194	}
195
196	static int verify_opcode(struct page page, unsigned long vaddr, uprobe_opcode_t insn,
197	int nbytes, void *data)
198	{
199	uprobe_opcode_t old_opcode;
200	bool is_swbp;
201
202	/*
203	* Note: We only check if the old_opcode is UPROBE_SWBP_INSN here.
204	* We do not check if it is any other 'trap variant' which could
205	* be conditional trap instruction such as the one powerpc supports.
206	*
207	* The logic is that we do not care if the underlying instruction
208	* is a trap variant; uprobes always wins over any other (gdb)
209	* breakpoint.
210	*/
211	uprobe_copy_from_page(page, vaddr, &old_opcode, UPROBE_SWBP_INSN_SIZE);
212	is_swbp = is_swbp_insn(&old_opcode);
213
214	if (is_swbp_insn(insn)) {
215	if (is_swbp) /* register: already installed? */
216	return 0;
217	} else {
218	if (!is_swbp) /* unregister: was it changed by us? */
219	return 0;
220	}
221
222	return 1;
223	}
224
225	static struct delayed_uprobe *
226	delayed_uprobe_check(struct uprobe uprobe, struct mm_struct mm)
227	{
228	struct delayed_uprobe *du;
229
230	list_for_each_entry(du, &delayed_uprobe_list, list)
231	if (du->uprobe == uprobe && du->mm == mm)
232	return du;
233	return NULL;
234	}
235
236	static int delayed_uprobe_add(struct uprobe uprobe, struct mm_struct mm)
237	{
238	struct delayed_uprobe *du;
239
240	if (delayed_uprobe_check(uprobe, mm))
241	return 0;
242
243	du = kzalloc_obj(*du);
244	if (!du)
245	return -ENOMEM;
246
247	du->uprobe = uprobe;
248	du->mm = mm;
249	list_add(&du->list, &delayed_uprobe_list);
250	return 0;
251	}
252
253	static void delayed_uprobe_delete(struct delayed_uprobe *du)
254	{
255	if (WARN_ON(!du))
256	return;
257	list_del(&du->list);
258	kfree(du);
259	}
260
261	static void delayed_uprobe_remove(struct uprobe uprobe, struct mm_struct mm)
262	{
263	struct list_head pos, q;
264	struct delayed_uprobe *du;
265
266	if (!uprobe && !mm)
267	return;
268
269	list_for_each_safe(pos, q, &delayed_uprobe_list) {
270	du = list_entry(pos, struct delayed_uprobe, list);
271
272	if (uprobe && du->uprobe != uprobe)
273	continue;
274	if (mm && du->mm != mm)
275	continue;
276
277	delayed_uprobe_delete(du);
278	}
279	}
280
281	static bool valid_ref_ctr_vma(struct uprobe *uprobe,
282	struct vm_area_struct *vma)
283	{
284	unsigned long vaddr = offset_to_vaddr(vma, uprobe->ref_ctr_offset);
285
286	return uprobe->ref_ctr_offset &&
287	vma->vm_file &&
288	file_inode(vma->vm_file) == uprobe->inode &&
289	(vma->vm_flags & (VM_WRITE\|VM_SHARED)) == VM_WRITE &&
290	vma->vm_start <= vaddr &&
291	vma->vm_end > vaddr;
292	}
293
294	static struct vm_area_struct *
295	find_ref_ctr_vma(struct uprobe uprobe, struct mm_struct mm)
296	{
297	VMA_ITERATOR(vmi, mm, 0);
298	struct vm_area_struct *tmp;
299
300	for_each_vma(vmi, tmp)
301	if (valid_ref_ctr_vma(uprobe, tmp))
302	return tmp;
303
304	return NULL;
305	}
306
307	static int
308	__update_ref_ctr(struct mm_struct *mm, unsigned long vaddr, short d)
309	{
310	void *kaddr;
311	struct page *page;
312	int ret;
313	short *ptr;
314
315	if (!vaddr \|\| !d)
316	return -EINVAL;
317
318	ret = get_user_pages_remote(mm, vaddr, 1,
319	FOLL_WRITE, &page, NULL);
320	if (unlikely(ret <= 0)) {
321	/*
322	* We are asking for 1 page. If get_user_pages_remote() fails,
323	* it may return 0, in that case we have to return error.
324	*/
325	return ret == 0 ? -EBUSY : ret;
326	}
327
328	kaddr = kmap_local_page(page);
329	ptr = kaddr + (vaddr & ~PAGE_MASK);
330
331	if (unlikely(*ptr + d < 0)) {
332	pr_warn("ref_ctr going negative. vaddr: 0x%lx, "
333	"curr val: %d, delta: %d\n", vaddr, *ptr, d);
334	ret = -EINVAL;
335	goto out;
336	}
337
338	*ptr += d;
339	ret = 0;
340	out:
341	kunmap_local(kaddr);
342	put_page(page);
343	return ret;
344	}
345
346	static void update_ref_ctr_warn(struct uprobe *uprobe,
347	struct mm_struct *mm, short d)
348	{
349	pr_warn("ref_ctr %s failed for inode: 0x%llx offset: "
350	"0x%llx ref_ctr_offset: 0x%llx of mm: 0x%p\n",
351	d > 0 ? "increment" : "decrement", uprobe->inode->i_ino,
352	(unsigned long long) uprobe->offset,
353	(unsigned long long) uprobe->ref_ctr_offset, mm);
354	}
355
356	static int update_ref_ctr(struct uprobe uprobe, struct mm_struct mm,
357	short d)
358	{
359	struct vm_area_struct *rc_vma;
360	unsigned long rc_vaddr;
361	int ret = 0;
362
363	rc_vma = find_ref_ctr_vma(uprobe, mm);
364
365	if (rc_vma) {
366	rc_vaddr = offset_to_vaddr(rc_vma, uprobe->ref_ctr_offset);
367	ret = __update_ref_ctr(mm, rc_vaddr, d);
368	if (ret)
369	update_ref_ctr_warn(uprobe, mm, d);
370
371	if (d > 0)
372	return ret;
373	}
374
375	mutex_lock(&delayed_uprobe_lock);
376	if (d > 0)
377	ret = delayed_uprobe_add(uprobe, mm);
378	else
379	delayed_uprobe_remove(uprobe, mm);
380	mutex_unlock(&delayed_uprobe_lock);
381
382	return ret;
383	}
384
385	static bool orig_page_is_identical(struct vm_area_struct *vma,
386	unsigned long vaddr, struct page page, bool pmd_mappable)
387	{
388	const pgoff_t index = vaddr_to_offset(vma, vaddr) >> PAGE_SHIFT;
389	struct folio *orig_folio = filemap_get_folio(vma->vm_file->f_mapping,
390	index);
391	struct page *orig_page;
392	bool identical;
393
394	if (IS_ERR(orig_folio))
395	return false;
396	orig_page = folio_file_page(orig_folio, index);
397
398	*pmd_mappable = folio_test_pmd_mappable(orig_folio);
399	identical = folio_test_uptodate(orig_folio) &&
400	pages_identical(page, orig_page);
401	folio_put(orig_folio);
402	return identical;
403	}
404
405	static int __uprobe_write(struct vm_area_struct *vma,
406	struct folio_walk fw, struct folio folio,
407	unsigned long insn_vaddr, uprobe_opcode_t *insn, int nbytes,
408	bool is_register)
409	{
410	const unsigned long vaddr = insn_vaddr & PAGE_MASK;
411	bool pmd_mappable;
412
413	/* For now, we'll only handle PTE-mapped folios. */
414	if (fw->level != FW_LEVEL_PTE)
415	return -EFAULT;
416
417	/*
418	* See can_follow_write_pte(): we'd actually prefer a writable PTE here,
419	* but the VMA might not be writable.
420	*/
421	if (!pte_write(fw->pte)) {
422	if (!PageAnonExclusive(fw->page))
423	return -EFAULT;
424	if (unlikely(userfaultfd_pte_wp(vma, fw->pte)))
425	return -EFAULT;
426	/* SOFTDIRTY is handled via pte_mkdirty() below. */
427	}
428
429	/*
430	* We'll temporarily unmap the page and flush the TLB, such that we can
431	* modify the page atomically.
432	*/
433	flush_cache_page(vma, vaddr, pte_pfn(fw->pte));
434	fw->pte = ptep_clear_flush(vma, vaddr, fw->ptep);
435	copy_to_page(fw->page, insn_vaddr, insn, nbytes);
436
437	/*
438	* When unregistering, we may only zap a PTE if uffd is disabled and
439	* there are no unexpected folio references ...
440	*/
441	if (is_register \|\| userfaultfd_missing(vma) \|\|
442	(folio_ref_count(folio) != folio_expected_ref_count(folio) + 1))
443	goto remap;
444
445	/*
446	* ... and the mapped page is identical to the original page that
447	* would get faulted in on next access.
448	*/
449	if (!orig_page_is_identical(vma, vaddr, fw->page, &pmd_mappable))
450	goto remap;
451
452	dec_mm_counter(vma->vm_mm, MM_ANONPAGES);
453	folio_remove_rmap_pte(folio, fw->page, vma);
454	if (!folio_mapped(folio) && folio_test_swapcache(folio) &&
455	folio_trylock(folio)) {
456	folio_free_swap(folio);
457	folio_unlock(folio);
458	}
459	folio_put(folio);
460
461	return pmd_mappable;
462	remap:
463	/*
464	* Make sure that our copy_to_page() changes become visible before the
465	* set_pte_at() write.
466	*/
467	smp_wmb();
468	/* We modified the page. Make sure to mark the PTE dirty. */
469	set_pte_at(vma->vm_mm, vaddr, fw->ptep, pte_mkdirty(fw->pte));
470	return 0;
471	}
472
473	/*
474	* NOTE:
475	* Expect the breakpoint instruction to be the smallest size instruction for
476	* the architecture. If an arch has variable length instruction and the
477	* breakpoint instruction is not of the smallest length instruction
478	* supported by that architecture then we need to modify is_trap_at_addr and
479	* uprobe_write_opcode accordingly. This would never be a problem for archs
480	* that have fixed length instructions.
481	*
482	* uprobe_write_opcode - write the opcode at a given virtual address.
483	* @auprobe: arch specific probepoint information.
484	* @vma: the probed virtual memory area.
485	* @opcode_vaddr: the virtual address to store the opcode.
486	* @opcode: opcode to be written at @opcode_vaddr.
487	*
488	* Called with mm->mmap_lock held for write.
489	* Return 0 (success) or a negative errno.
490	*/
491	int uprobe_write_opcode(struct arch_uprobe auprobe, struct vm_area_struct vma,
492	const unsigned long opcode_vaddr, uprobe_opcode_t opcode,
493	bool is_register)
494	{
495	return uprobe_write(auprobe, vma, opcode_vaddr, &opcode, UPROBE_SWBP_INSN_SIZE,
496	verify_opcode, is_register, true /* do_update_ref_ctr */, NULL);
497	}
498
499	int uprobe_write(struct arch_uprobe auprobe, struct vm_area_struct vma,
500	const unsigned long insn_vaddr, uprobe_opcode_t *insn, int nbytes,
501	uprobe_write_verify_t verify, bool is_register, bool do_update_ref_ctr,
502	void *data)
503	{
504	const unsigned long vaddr = insn_vaddr & PAGE_MASK;
505	struct mm_struct *mm = vma->vm_mm;
506	struct uprobe *uprobe;
507	int ret, ref_ctr_updated = 0;
508	unsigned int gup_flags = FOLL_FORCE;
509	struct mmu_notifier_range range;
510	struct folio_walk fw;
511	struct folio *folio;
512	struct page *page;
513
514	uprobe = container_of(auprobe, struct uprobe, arch);
515
516	if (WARN_ON_ONCE(!vma_is_cow_mapping(vma)))
517	return -EINVAL;
518
519	/*
520	* When registering, we have to break COW to get an exclusive anonymous
521	* page that we can safely modify. Use FOLL_WRITE to trigger a write
522	* fault if required. When unregistering, we might be lucky and the
523	* anon page is already gone. So defer write faults until really
524	* required. Use FOLL_SPLIT_PMD, because __uprobe_write()
525	* cannot deal with PMDs yet.
526	*/
527	if (is_register)
528	gup_flags \|= FOLL_WRITE \| FOLL_SPLIT_PMD;
529
530	retry:
531	ret = get_user_pages_remote(mm, vaddr, 1, gup_flags, &page, NULL);
532	if (ret <= 0)
533	goto out;
534	folio = page_folio(page);
535
536	ret = verify(page, insn_vaddr, insn, nbytes, data);
537	if (ret <= 0) {
538	folio_put(folio);
539	goto out;
540	}
541
542	/* We are going to replace instruction, update ref_ctr. */
543	if (do_update_ref_ctr && !ref_ctr_updated && uprobe->ref_ctr_offset) {
544	ret = update_ref_ctr(uprobe, mm, is_register ? 1 : -1);
545	if (ret) {
546	folio_put(folio);
547	goto out;
548	}
549
550	ref_ctr_updated = 1;
551	}
552
553	ret = 0;
554	if (unlikely(!folio_test_anon(folio) \|\| folio_is_zone_device(folio))) {
555	VM_WARN_ON_ONCE(is_register);
556	folio_put(folio);
557	goto out;
558	}
559
560	if (!is_register) {
561	/*
562	* In the common case, we'll be able to zap the page when
563	* unregistering. So trigger MMU notifiers now, as we won't
564	* be able to do it under PTL.
565	*/
566	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
567	vaddr, vaddr + PAGE_SIZE);
568	mmu_notifier_invalidate_range_start(&range);
569	}
570
571	ret = -EAGAIN;
572	/* Walk the page tables again, to perform the actual update. */
573	if (folio_walk_start(&fw, vma, vaddr, 0)) {
574	if (fw.page == page)
575	ret = __uprobe_write(vma, &fw, folio, insn_vaddr, insn, nbytes, is_register);
576	folio_walk_end(&fw, vma);
577	}
578
579	if (!is_register)
580	mmu_notifier_invalidate_range_end(&range);
581
582	folio_put(folio);
583	switch (ret) {
584	case -EFAULT:
585	gup_flags \|= FOLL_WRITE \| FOLL_SPLIT_PMD;
586	fallthrough;
587	case -EAGAIN:
588	goto retry;
589	default:
590	break;
591	}
592
593	out:
594	/* Revert back reference counter if instruction update failed. */
595	if (do_update_ref_ctr && ret < 0 && ref_ctr_updated)
596	update_ref_ctr(uprobe, mm, is_register ? -1 : 1);
597
598	/* try collapse pmd for compound page */
599	if (ret > 0)
600	collapse_pte_mapped_thp(mm, vaddr, false);
601
602	return ret < 0 ? ret : 0;
603	}
604
605	/**
606	* set_swbp - store breakpoint at a given address.
607	* @auprobe: arch specific probepoint information.
608	* @vma: the probed virtual memory area.
609	* @vaddr: the virtual address to insert the opcode.
610	*
611	* For mm @mm, store the breakpoint instruction at @vaddr.
612	* Return 0 (success) or a negative errno.
613	*/
614	int __weak set_swbp(struct arch_uprobe auprobe, struct vm_area_struct vma,
615	unsigned long vaddr)
616	{
617	return uprobe_write_opcode(auprobe, vma, vaddr, UPROBE_SWBP_INSN, true);
618	}
619
620	/**
621	* set_orig_insn - Restore the original instruction.
622	* @vma: the probed virtual memory area.
623	* @auprobe: arch specific probepoint information.
624	* @vaddr: the virtual address to insert the opcode.
625	*
626	* For mm @mm, restore the original opcode (opcode) at @vaddr.
627	* Return 0 (success) or a negative errno.
628	*/
629	int __weak set_orig_insn(struct arch_uprobe *auprobe,
630	struct vm_area_struct *vma, unsigned long vaddr)
631	{
632	return uprobe_write_opcode(auprobe, vma, vaddr,
633	(uprobe_opcode_t )&auprobe->insn, false);
634	}
635
636	/* uprobe should have guaranteed positive refcount */
637	static struct uprobe get_uprobe(struct uprobe uprobe)
638	{
639	refcount_inc(&uprobe->ref);
640	return uprobe;
641	}
642
643	/*
644	* uprobe should have guaranteed lifetime, which can be either of:
645	* - caller already has refcount taken (and wants an extra one);
646	* - uprobe is RCU protected and won't be freed until after grace period;
647	* - we are holding uprobes_treelock (for read or write, doesn't matter).
648	*/
649	static struct uprobe try_get_uprobe(struct uprobe uprobe)
650	{
651	if (refcount_inc_not_zero(&uprobe->ref))
652	return uprobe;
653	return NULL;
654	}
655
656	static inline bool uprobe_is_active(struct uprobe *uprobe)
657	{
658	return !RB_EMPTY_NODE(&uprobe->rb_node);
659	}
660
661	static void uprobe_free_rcu_tasks_trace(struct rcu_head *rcu)
662	{
663	struct uprobe *uprobe = container_of(rcu, struct uprobe, rcu);
664
665	kfree(uprobe);
666	}
667
668	static void uprobe_free_srcu(struct rcu_head *rcu)
669	{
670	struct uprobe *uprobe = container_of(rcu, struct uprobe, rcu);
671
672	call_rcu_tasks_trace(&uprobe->rcu, uprobe_free_rcu_tasks_trace);
673	}
674
675	static void uprobe_free_deferred(struct work_struct *work)
676	{
677	struct uprobe *uprobe = container_of(work, struct uprobe, work);
678
679	write_lock(&uprobes_treelock);
680
681	if (uprobe_is_active(uprobe)) {
682	write_seqcount_begin(&uprobes_seqcount);
683	rb_erase(&uprobe->rb_node, &uprobes_tree);
684	write_seqcount_end(&uprobes_seqcount);
685	}
686
687	write_unlock(&uprobes_treelock);
688
689	/*
690	* If application munmap(exec_vma) before uprobe_unregister()
691	* gets called, we don't get a chance to remove uprobe from
692	* delayed_uprobe_list from remove_breakpoint(). Do it here.
693	*/
694	mutex_lock(&delayed_uprobe_lock);
695	delayed_uprobe_remove(uprobe, NULL);
696	mutex_unlock(&delayed_uprobe_lock);
697
698	/* start srcu -> rcu_tasks_trace -> kfree chain */
699	call_srcu(&uretprobes_srcu, &uprobe->rcu, uprobe_free_srcu);
700	}
701
702	static void put_uprobe(struct uprobe *uprobe)
703	{
704	if (!refcount_dec_and_test(&uprobe->ref))
705	return;
706
707	INIT_WORK(&uprobe->work, uprobe_free_deferred);
708	schedule_work(&uprobe->work);
709	}
710
711	/* Initialize hprobe as SRCU-protected "leased" uprobe */
712	static void hprobe_init_leased(struct hprobe hprobe, struct uprobe uprobe,
713	struct srcu_ctr __percpu *srcu_scp)
714	{
715	WARN_ON(!uprobe);
716	hprobe->state = HPROBE_LEASED;
717	hprobe->uprobe = uprobe;
718	hprobe->srcu_scp = srcu_scp;
719	}
720
721	/* Initialize hprobe as refcounted ("stable") uprobe (uprobe can be NULL). */
722	static void hprobe_init_stable(struct hprobe hprobe, struct uprobe uprobe)
723	{
724	hprobe->state = uprobe ? HPROBE_STABLE : HPROBE_GONE;
725	hprobe->uprobe = uprobe;
726	hprobe->srcu_scp = NULL;
727	}
728
729	/*
730	* hprobe_consume() fetches hprobe's underlying uprobe and detects whether
731	* uprobe is SRCU protected or is refcounted. hprobe_consume() can be
732	* used only once for a given hprobe.
733	*
734	* Caller has to call hprobe_finalize() and pass previous hprobe_state, so
735	* that hprobe_finalize() can perform SRCU unlock or put uprobe, whichever
736	* is appropriate.
737	*/
738	static inline struct uprobe hprobe_consume(struct hprobe hprobe, enum hprobe_state *hstate)
739	{
740	*hstate = xchg(&hprobe->state, HPROBE_CONSUMED);
741	switch (*hstate) {
742	case HPROBE_LEASED:
743	case HPROBE_STABLE:
744	return hprobe->uprobe;
745	case HPROBE_GONE: /* uprobe is NULL, no SRCU */
746	case HPROBE_CONSUMED: /* uprobe was finalized already, do nothing */
747	return NULL;
748	default:
749	WARN(1, "hprobe invalid state %d", *hstate);
750	return NULL;
751	}
752	}
753
754	/*
755	* Reset hprobe state and, if hprobe was LEASED, release SRCU lock.
756	* hprobe_finalize() can only be used from current context after
757	* hprobe_consume() call (which determines uprobe and hstate value).
758	*/
759	static void hprobe_finalize(struct hprobe *hprobe, enum hprobe_state hstate)
760	{
761	switch (hstate) {
762	case HPROBE_LEASED:
763	srcu_up_read_fast(&uretprobes_srcu, hprobe->srcu_scp);
764	break;
765	case HPROBE_STABLE:
766	put_uprobe(hprobe->uprobe);
767	break;
768	case HPROBE_GONE:
769	case HPROBE_CONSUMED:
770	break;
771	default:
772	WARN(1, "hprobe invalid state %d", hstate);
773	break;
774	}
775	}
776
777	/*
778	* Attempt to switch (atomically) uprobe from being SRCU protected (LEASED)
779	* to refcounted (STABLE) state. Competes with hprobe_consume(); only one of
780	* them can win the race to perform SRCU unlocking. Whoever wins must perform
781	* SRCU unlock.
782	*
783	* Returns underlying valid uprobe or NULL, if there was no underlying uprobe
784	* to begin with or we failed to bump its refcount and it's going away.
785	*
786	* Returned non-NULL uprobe can be still safely used within an ongoing SRCU
787	* locked region. If `get` is true, it's guaranteed that non-NULL uprobe has
788	* an extra refcount for caller to assume and use. Otherwise, it's not
789	* guaranteed that returned uprobe has a positive refcount, so caller has to
790	* attempt try_get_uprobe(), if it needs to preserve uprobe beyond current
791	* SRCU lock region. See dup_utask().
792	*/
793	static struct uprobe hprobe_expire(struct hprobe hprobe, bool get)
794	{
795	enum hprobe_state hstate;
796
797	/*
798	* Caller should guarantee that return_instance is not going to be
799	* freed from under us. This can be achieved either through holding
800	* rcu_read_lock() or by owning return_instance in the first place.
801	*
802	* Underlying uprobe is itself protected from reuse by SRCU, so ensure
803	* SRCU lock is held properly.
804	*/
805	lockdep_assert(srcu_read_lock_held(&uretprobes_srcu));
806
807	hstate = READ_ONCE(hprobe->state);
808	switch (hstate) {
809	case HPROBE_STABLE:
810	/* uprobe has positive refcount, bump refcount, if necessary */
811	return get ? get_uprobe(hprobe->uprobe) : hprobe->uprobe;
812	case HPROBE_GONE:
813	/*
814	* SRCU was unlocked earlier and we didn't manage to take
815	* uprobe refcnt, so it's effectively NULL
816	*/
817	return NULL;
818	case HPROBE_CONSUMED:
819	/*
820	* uprobe was consumed, so it's effectively NULL as far as
821	* uretprobe processing logic is concerned
822	*/
823	return NULL;
824	case HPROBE_LEASED: {
825	struct uprobe *uprobe = try_get_uprobe(hprobe->uprobe);
826	/*
827	* Try to switch hprobe state, guarding against
828	* hprobe_consume() or another hprobe_expire() racing with us.
829	* Note, if we failed to get uprobe refcount, we use special
830	* HPROBE_GONE state to signal that hprobe->uprobe shouldn't
831	* be used as it will be freed after SRCU is unlocked.
832	*/
833	if (try_cmpxchg(&hprobe->state, &hstate, uprobe ? HPROBE_STABLE : HPROBE_GONE)) {
834	/* We won the race, we are the ones to unlock SRCU */
835	srcu_up_read_fast(&uretprobes_srcu, hprobe->srcu_scp);
836	return get && uprobe ? get_uprobe(uprobe) : uprobe;
837	}
838
839	/*
840	* We lost the race, undo refcount bump (if it ever happened),
841	* unless caller would like an extra refcount anyways.
842	*/
843	if (uprobe && !get)
844	put_uprobe(uprobe);
845	/*
846	* Even if hprobe_consume() or another hprobe_expire() wins
847	* the state update race and unlocks SRCU from under us, we
848	* still have a guarantee that underyling uprobe won't be
849	* freed due to ongoing caller's SRCU lock region, so we can
850	* return it regardless. Also, if `get` was true, we also have
851	* an extra ref for the caller to own. This is used in dup_utask().
852	*/
853	return uprobe;
854	}
855	default:
856	WARN(1, "unknown hprobe state %d", hstate);
857	return NULL;
858	}
859	}
860
861	static __always_inline
862	int uprobe_cmp(const struct inode *l_inode, const loff_t l_offset,
863	const struct uprobe *r)
864	{
865	if (l_inode < r->inode)
866	return -1;
867
868	if (l_inode > r->inode)
869	return 1;
870
871	if (l_offset < r->offset)
872	return -1;
873
874	if (l_offset > r->offset)
875	return 1;
876
877	return 0;
878	}
879
880	#define __node_2_uprobe(node) \
881	rb_entry((node), struct uprobe, rb_node)
882
883	struct __uprobe_key {
884	struct inode *inode;
885	loff_t offset;
886	};
887
888	static inline int __uprobe_cmp_key(const void key, const struct rb_node b)
889	{
890	const struct __uprobe_key *a = key;
891	return uprobe_cmp(a->inode, a->offset, __node_2_uprobe(b));
892	}
893
894	static inline int __uprobe_cmp(struct rb_node a, const struct rb_node b)
895	{
896	struct uprobe *u = __node_2_uprobe(a);
897	return uprobe_cmp(u->inode, u->offset, __node_2_uprobe(b));
898	}
899
900	/*
901	* Assumes being inside RCU protected region.
902	* No refcount is taken on returned uprobe.
903	*/
904	static struct uprobe find_uprobe_rcu(struct inode inode, loff_t offset)
905	{
906	struct __uprobe_key key = {
907	.inode = inode,
908	.offset = offset,
909	};
910	struct rb_node *node;
911	unsigned int seq;
912
913	lockdep_assert(rcu_read_lock_trace_held());
914
915	do {
916	seq = read_seqcount_begin(&uprobes_seqcount);
917	node = rb_find_rcu(&key, &uprobes_tree, __uprobe_cmp_key);
918	/*
919	* Lockless RB-tree lookups can result only in false negatives.
920	* If the element is found, it is correct and can be returned
921	* under RCU protection. If we find nothing, we need to
922	* validate that seqcount didn't change. If it did, we have to
923	* try again as we might have missed the element (false
924	* negative). If seqcount is unchanged, search truly failed.
925	*/
926	if (node)
927	return __node_2_uprobe(node);
928	} while (read_seqcount_retry(&uprobes_seqcount, seq));
929
930	return NULL;
931	}
932
933	/*
934	* Attempt to insert a new uprobe into uprobes_tree.
935	*
936	* If uprobe already exists (for given inode+offset), we just increment
937	* refcount of previously existing uprobe.
938	*
939	* If not, a provided new instance of uprobe is inserted into the tree (with
940	* assumed initial refcount == 1).
941	*
942	* In any case, we return a uprobe instance that ends up being in uprobes_tree.
943	* Caller has to clean up new uprobe instance, if it ended up not being
944	* inserted into the tree.
945	*
946	* We assume that uprobes_treelock is held for writing.
947	*/
948	static struct uprobe __insert_uprobe(struct uprobe uprobe)
949	{
950	struct rb_node *node;
951	again:
952	node = rb_find_add_rcu(&uprobe->rb_node, &uprobes_tree, __uprobe_cmp);
953	if (node) {
954	struct uprobe *u = __node_2_uprobe(node);
955
956	if (!try_get_uprobe(u)) {
957	rb_erase(node, &uprobes_tree);
958	RB_CLEAR_NODE(&u->rb_node);
959	goto again;
960	}
961
962	return u;
963	}
964
965	return uprobe;
966	}
967
968	/*
969	* Acquire uprobes_treelock and insert uprobe into uprobes_tree
970	* (or reuse existing one, see __insert_uprobe() comments above).
971	*/
972	static struct uprobe insert_uprobe(struct uprobe uprobe)
973	{
974	struct uprobe *u;
975
976	write_lock(&uprobes_treelock);
977	write_seqcount_begin(&uprobes_seqcount);
978	u = __insert_uprobe(uprobe);
979	write_seqcount_end(&uprobes_seqcount);
980	write_unlock(&uprobes_treelock);
981
982	return u;
983	}
984
985	static void
986	ref_ctr_mismatch_warn(struct uprobe cur_uprobe, struct uprobe uprobe)
987	{
988	pr_warn("ref_ctr_offset mismatch. inode: 0x%llx offset: 0x%llx "
989	"ref_ctr_offset(old): 0x%llx ref_ctr_offset(new): 0x%llx\n",
990	uprobe->inode->i_ino, (unsigned long long) uprobe->offset,
991	(unsigned long long) cur_uprobe->ref_ctr_offset,
992	(unsigned long long) uprobe->ref_ctr_offset);
993	}
994
995	static struct uprobe alloc_uprobe(struct inode inode, loff_t offset,
996	loff_t ref_ctr_offset)
997	{
998	struct uprobe uprobe, cur_uprobe;
999
1000	uprobe = kzalloc_obj(struct uprobe);
1001	if (!uprobe)
1002	return ERR_PTR(-ENOMEM);
1003
1004	uprobe->inode = inode;
1005	uprobe->offset = offset;
1006	uprobe->ref_ctr_offset = ref_ctr_offset;
1007	INIT_LIST_HEAD(&uprobe->consumers);
1008	init_rwsem(&uprobe->register_rwsem);
1009	init_rwsem(&uprobe->consumer_rwsem);
1010	RB_CLEAR_NODE(&uprobe->rb_node);
1011	refcount_set(&uprobe->ref, 1);
1012
1013	/* add to uprobes_tree, sorted on inode:offset */
1014	cur_uprobe = insert_uprobe(uprobe);
1015	/* a uprobe exists for this inode:offset combination */
1016	if (cur_uprobe != uprobe) {
1017	if (cur_uprobe->ref_ctr_offset != uprobe->ref_ctr_offset) {
1018	ref_ctr_mismatch_warn(cur_uprobe, uprobe);
1019	put_uprobe(cur_uprobe);
1020	kfree(uprobe);
1021	return ERR_PTR(-EINVAL);
1022	}
1023	kfree(uprobe);
1024	uprobe = cur_uprobe;
1025	}
1026
1027	return uprobe;
1028	}
1029
1030	static void consumer_add(struct uprobe uprobe, struct uprobe_consumer uc)
1031	{
1032	static atomic64_t id;
1033
1034	down_write(&uprobe->consumer_rwsem);
1035	list_add_rcu(&uc->cons_node, &uprobe->consumers);
1036	uc->id = (__u64) atomic64_inc_return(&id);
1037	up_write(&uprobe->consumer_rwsem);
1038	}
1039
1040	/*
1041	* For uprobe @uprobe, delete the consumer @uc.
1042	* Should never be called with consumer that's not part of @uprobe->consumers.
1043	*/
1044	static void consumer_del(struct uprobe uprobe, struct uprobe_consumer uc)
1045	{
1046	down_write(&uprobe->consumer_rwsem);
1047	list_del_rcu(&uc->cons_node);
1048	up_write(&uprobe->consumer_rwsem);
1049	}
1050
1051	static int __copy_insn(struct address_space mapping, struct file filp,
1052	void *insn, int nbytes, loff_t offset)
1053	{
1054	struct page *page;
1055	/*
1056	* Ensure that the page that has the original instruction is populated
1057	* and in page-cache. If ->read_folio == NULL it must be shmem_mapping(),
1058	* see uprobe_register().
1059	*/
1060	if (mapping->a_ops->read_folio)
1061	page = read_mapping_page(mapping, offset >> PAGE_SHIFT, filp);
1062	else
1063	page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
1064	if (IS_ERR(page))
1065	return PTR_ERR(page);
1066
1067	uprobe_copy_from_page(page, offset, insn, nbytes);
1068	put_page(page);
1069
1070	return 0;
1071	}
1072
1073	static int copy_insn(struct uprobe uprobe, struct file filp)
1074	{
1075	struct address_space *mapping = uprobe->inode->i_mapping;
1076	loff_t offs = uprobe->offset;
1077	void *insn = &uprobe->arch.insn;
1078	int size = sizeof(uprobe->arch.insn);
1079	int len, err = -EIO;
1080
1081	/* Copy only available bytes, -EIO if nothing was read */
1082	do {
1083	if (offs >= i_size_read(uprobe->inode))
1084	break;
1085
1086	len = min_t(int, size, PAGE_SIZE - (offs & ~PAGE_MASK));
1087	err = __copy_insn(mapping, filp, insn, len, offs);
1088	if (err)
1089	break;
1090
1091	insn += len;
1092	offs += len;
1093	size -= len;
1094	} while (size);
1095
1096	return err;
1097	}
1098
1099	static int prepare_uprobe(struct uprobe uprobe, struct file file,
1100	struct mm_struct *mm, unsigned long vaddr)
1101	{
1102	int ret = 0;
1103
1104	if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
1105	return ret;
1106
1107	/* TODO: move this into _register, until then we abuse this sem. */
1108	down_write(&uprobe->consumer_rwsem);
1109	if (test_bit(UPROBE_COPY_INSN, &uprobe->flags))
1110	goto out;
1111
1112	ret = copy_insn(uprobe, file);
1113	if (ret)
1114	goto out;
1115
1116	ret = -ENOTSUPP;
1117	if (is_trap_insn((uprobe_opcode_t *)&uprobe->arch.insn))
1118	goto out;
1119
1120	ret = arch_uprobe_analyze_insn(&uprobe->arch, mm, vaddr);
1121	if (ret)
1122	goto out;
1123
1124	smp_wmb(); /* pairs with the smp_rmb() in handle_swbp() */
1125	set_bit(UPROBE_COPY_INSN, &uprobe->flags);
1126
1127	out:
1128	up_write(&uprobe->consumer_rwsem);
1129
1130	return ret;
1131	}
1132
1133	static inline bool consumer_filter(struct uprobe_consumer uc, struct mm_struct mm)
1134	{
1135	return !uc->filter \|\| uc->filter(uc, mm);
1136	}
1137
1138	static bool filter_chain(struct uprobe uprobe, struct mm_struct mm)
1139	{
1140	struct uprobe_consumer *uc;
1141	bool ret = false;
1142
1143	down_read(&uprobe->consumer_rwsem);
1144	list_for_each_entry(uc, &uprobe->consumers, cons_node) {
1145	ret = consumer_filter(uc, mm);
1146	if (ret)
1147	break;
1148	}
1149	up_read(&uprobe->consumer_rwsem);
1150
1151	return ret;
1152	}
1153
1154	static int install_breakpoint(struct uprobe uprobe, struct vm_area_struct vma,
1155	unsigned long vaddr)
1156	{
1157	struct mm_struct *mm = vma->vm_mm;
1158	bool first_uprobe;
1159	int ret;
1160
1161	ret = prepare_uprobe(uprobe, vma->vm_file, mm, vaddr);
1162	if (ret)
1163	return ret;
1164
1165	/*
1166	* set MMF_HAS_UPROBES in advance for uprobe_pre_sstep_notifier(),
1167	* the task can hit this breakpoint right after __replace_page().
1168	*/
1169	first_uprobe = !mm_flags_test(MMF_HAS_UPROBES, mm);
1170	if (first_uprobe)
1171	mm_flags_set(MMF_HAS_UPROBES, mm);
1172
1173	ret = set_swbp(&uprobe->arch, vma, vaddr);
1174	if (!ret)
1175	mm_flags_clear(MMF_RECALC_UPROBES, mm);
1176	else if (first_uprobe)
1177	mm_flags_clear(MMF_HAS_UPROBES, mm);
1178
1179	return ret;
1180	}
1181
1182	static int remove_breakpoint(struct uprobe uprobe, struct vm_area_struct vma,
1183	unsigned long vaddr)
1184	{
1185	struct mm_struct *mm = vma->vm_mm;
1186
1187	mm_flags_set(MMF_RECALC_UPROBES, mm);
1188	return set_orig_insn(&uprobe->arch, vma, vaddr);
1189	}
1190
1191	struct map_info {
1192	struct map_info *next;
1193	struct mm_struct *mm;
1194	unsigned long vaddr;
1195	};
1196
1197	static inline struct map_info free_map_info(struct map_info info)
1198	{
1199	struct map_info *next = info->next;
1200	kfree(info);
1201	return next;
1202	}
1203
1204	static struct map_info *
1205	build_map_info(struct address_space *mapping, loff_t offset, bool is_register)
1206	{
1207	unsigned long pgoff = offset >> PAGE_SHIFT;
1208	struct vm_area_struct *vma;
1209	struct map_info *curr = NULL;
1210	struct map_info *prev = NULL;
1211	struct map_info *info;
1212	int more = 0;
1213
1214	again:
1215	i_mmap_lock_read(mapping);
1216	mapping_rmap_tree_foreach(vma, mapping, pgoff, pgoff) {
1217	if (!valid_vma(vma, is_register))
1218	continue;
1219
1220	if (!prev && !more) {
1221	/*
1222	* Needs GFP_NOWAIT to avoid i_mmap_rwsem recursion through
1223	* reclaim. This is optimistic, no harm done if it fails.
1224	*/
1225	prev = kmalloc_obj(struct map_info,
1226	GFP_NOWAIT \| __GFP_NOMEMALLOC);
1227	if (prev)
1228	prev->next = NULL;
1229	}
1230	if (!prev) {
1231	more++;
1232	continue;
1233	}
1234
1235	if (!mmget_not_zero(vma->vm_mm))
1236	continue;
1237
1238	info = prev;
1239	prev = prev->next;
1240	info->next = curr;
1241	curr = info;
1242
1243	info->mm = vma->vm_mm;
1244	info->vaddr = offset_to_vaddr(vma, offset);
1245	}
1246	i_mmap_unlock_read(mapping);
1247
1248	if (!more)
1249	goto out;
1250
1251	prev = curr;
1252	while (curr) {
1253	mmput(curr->mm);
1254	curr = curr->next;
1255	}
1256
1257	do {
1258	info = kmalloc_obj(struct map_info);
1259	if (!info) {
1260	curr = ERR_PTR(-ENOMEM);
1261	goto out;
1262	}
1263	info->next = prev;
1264	prev = info;
1265	} while (--more);
1266
1267	goto again;
1268	out:
1269	while (prev)
1270	prev = free_map_info(prev);
1271	return curr;
1272	}
1273
1274	static int
1275	register_for_each_vma(struct uprobe uprobe, struct uprobe_consumer new)
1276	{
1277	bool is_register = !!new;
1278	struct map_info *info;
1279	int err = 0;
1280
1281	percpu_down_write(&dup_mmap_sem);
1282	info = build_map_info(uprobe->inode->i_mapping,
1283	uprobe->offset, is_register);
1284	if (IS_ERR(info)) {
1285	err = PTR_ERR(info);
1286	goto out;
1287	}
1288
1289	while (info) {
1290	struct mm_struct *mm = info->mm;
1291	struct vm_area_struct *vma;
1292
1293	if (err && is_register)
1294	goto free;
1295	/*
1296	* We take mmap_lock for writing to avoid the race with
1297	* find_active_uprobe_rcu() which takes mmap_lock for reading.
1298	* Thus this install_breakpoint() can not make
1299	* is_trap_at_addr() true right after find_uprobe_rcu()
1300	* returns NULL in find_active_uprobe_rcu().
1301	*/
1302	mmap_write_lock(mm);
1303	if (check_stable_address_space(mm))
1304	goto unlock;
1305
1306	vma = find_vma(mm, info->vaddr);
1307	if (!vma \|\| !valid_vma(vma, is_register) \|\|
1308	file_inode(vma->vm_file) != uprobe->inode)
1309	goto unlock;
1310
1311	if (vma->vm_start > info->vaddr \|\|
1312	vaddr_to_offset(vma, info->vaddr) != uprobe->offset)
1313	goto unlock;
1314
1315	if (is_register) {
1316	/* consult only the "caller", new consumer. */
1317	if (consumer_filter(new, mm))
1318	err = install_breakpoint(uprobe, vma, info->vaddr);
1319	} else if (mm_flags_test(MMF_HAS_UPROBES, mm)) {
1320	if (!filter_chain(uprobe, mm))
1321	err \|= remove_breakpoint(uprobe, vma, info->vaddr);
1322	}
1323
1324	unlock:
1325	mmap_write_unlock(mm);
1326	free:
1327	mmput(mm);
1328	info = free_map_info(info);
1329	}
1330	out:
1331	percpu_up_write(&dup_mmap_sem);
1332	return err;
1333	}
1334
1335	/**
1336	* uprobe_unregister_nosync - unregister an already registered probe.
1337	* @uprobe: uprobe to remove
1338	* @uc: identify which probe if multiple probes are colocated.
1339	*/
1340	void uprobe_unregister_nosync(struct uprobe uprobe, struct uprobe_consumer uc)
1341	{
1342	int err;
1343
1344	down_write(&uprobe->register_rwsem);
1345	consumer_del(uprobe, uc);
1346	err = register_for_each_vma(uprobe, NULL);
1347	up_write(&uprobe->register_rwsem);
1348
1349	/* TODO : cant unregister? schedule a worker thread */
1350	if (unlikely(err)) {
1351	uprobe_warn(current, "unregister, leaking uprobe");
1352	return;
1353	}
1354
1355	put_uprobe(uprobe);
1356	}
1357	EXPORT_SYMBOL_GPL(uprobe_unregister_nosync);
1358
1359	void uprobe_unregister_sync(void)
1360	{
1361	/*
1362	* Now that handler_chain() and handle_uretprobe_chain() iterate over
1363	* uprobe->consumers list under RCU protection without holding
1364	* uprobe->register_rwsem, we need to wait for RCU grace period to
1365	* make sure that we can't call into just unregistered
1366	* uprobe_consumer's callbacks anymore. If we don't do that, fast and
1367	* unlucky enough caller can free consumer's memory and cause
1368	* handler_chain() or handle_uretprobe_chain() to do an use-after-free.
1369	*/
1370	synchronize_rcu_tasks_trace();
1371	synchronize_srcu(&uretprobes_srcu);
1372	}
1373	EXPORT_SYMBOL_GPL(uprobe_unregister_sync);
1374
1375	/**
1376	* uprobe_register - register a probe
1377	* @inode: the file in which the probe has to be placed.
1378	* @offset: offset from the start of the file.
1379	* @ref_ctr_offset: offset of SDT marker / reference counter
1380	* @uc: information on howto handle the probe..
1381	*
1382	* Apart from the access refcount, uprobe_register() takes a creation
1383	* refcount (thro alloc_uprobe) if and only if this @uprobe is getting
1384	* inserted into the rbtree (i.e first consumer for a @inode:@offset
1385	* tuple). Creation refcount stops uprobe_unregister from freeing the
1386	* @uprobe even before the register operation is complete. Creation
1387	* refcount is released when the last @uc for the @uprobe
1388	* unregisters. Caller of uprobe_register() is required to keep @inode
1389	* (and the containing mount) referenced.
1390	*
1391	* Return: pointer to the new uprobe on success or an ERR_PTR on failure.
1392	*/
1393	struct uprobe uprobe_register(struct inode inode,
1394	loff_t offset, loff_t ref_ctr_offset,
1395	struct uprobe_consumer *uc)
1396	{
1397	struct uprobe *uprobe;
1398	int ret;
1399
1400	/* Uprobe must have at least one set consumer */
1401	if (!uc->handler && !uc->ret_handler)
1402	return ERR_PTR(-EINVAL);
1403
1404	/* copy_insn() uses read_mapping_page() or shmem_read_mapping_page() */
1405	if (!inode->i_mapping->a_ops->read_folio &&
1406	!shmem_mapping(inode->i_mapping))
1407	return ERR_PTR(-EIO);
1408	/* Racy, just to catch the obvious mistakes */
1409	if (offset > i_size_read(inode))
1410	return ERR_PTR(-EINVAL);
1411
1412	/*
1413	* This ensures that uprobe_copy_from_page(), copy_to_page() and
1414	* __update_ref_ctr() can't cross page boundary.
1415	*/
1416	if (!IS_ALIGNED(offset, UPROBE_SWBP_INSN_SIZE))
1417	return ERR_PTR(-EINVAL);
1418	if (!IS_ALIGNED(ref_ctr_offset, sizeof(short)))
1419	return ERR_PTR(-EINVAL);
1420
1421	uprobe = alloc_uprobe(inode, offset, ref_ctr_offset);
1422	if (IS_ERR(uprobe))
1423	return uprobe;
1424
1425	down_write(&uprobe->register_rwsem);
1426	consumer_add(uprobe, uc);
1427	ret = register_for_each_vma(uprobe, uc);
1428	up_write(&uprobe->register_rwsem);
1429
1430	if (ret) {
1431	uprobe_unregister_nosync(uprobe, uc);
1432	/*
1433	* Registration might have partially succeeded, so we can have
1434	* this consumer being called right at this time. We need to
1435	* sync here. It's ok, it's unlikely slow path.
1436	*/
1437	uprobe_unregister_sync();
1438	return ERR_PTR(ret);
1439	}
1440
1441	return uprobe;
1442	}
1443	EXPORT_SYMBOL_GPL(uprobe_register);
1444
1445	/**
1446	* uprobe_apply - add or remove the breakpoints according to @uc->filter
1447	* @uprobe: uprobe which "owns" the breakpoint
1448	* @uc: consumer which wants to add more or remove some breakpoints
1449	* @add: add or remove the breakpoints
1450	* Return: 0 on success or negative error code.
1451	*/
1452	int uprobe_apply(struct uprobe uprobe, struct uprobe_consumer uc, bool add)
1453	{
1454	struct uprobe_consumer *con;
1455	int ret = -ENOENT;
1456
1457	down_write(&uprobe->register_rwsem);
1458
1459	rcu_read_lock_trace();
1460	list_for_each_entry_rcu(con, &uprobe->consumers, cons_node, rcu_read_lock_trace_held()) {
1461	if (con == uc) {
1462	ret = register_for_each_vma(uprobe, add ? uc : NULL);
1463	break;
1464	}
1465	}
1466	rcu_read_unlock_trace();
1467
1468	up_write(&uprobe->register_rwsem);
1469
1470	return ret;
1471	}
1472
1473	static int unapply_uprobe(struct uprobe uprobe, struct mm_struct mm)
1474	{
1475	VMA_ITERATOR(vmi, mm, 0);
1476	struct vm_area_struct *vma;
1477	int err = 0;
1478
1479	mmap_write_lock(mm);
1480	for_each_vma(vmi, vma) {
1481	unsigned long vaddr;
1482	loff_t offset;
1483
1484	if (!valid_vma(vma, false) \|\|
1485	file_inode(vma->vm_file) != uprobe->inode)
1486	continue;
1487
1488	offset = (loff_t)vma_start_pgoff(vma) << PAGE_SHIFT;
1489	if (uprobe->offset < offset \|\|
1490	uprobe->offset >= offset + vma->vm_end - vma->vm_start)
1491	continue;
1492
1493	vaddr = offset_to_vaddr(vma, uprobe->offset);
1494	err \|= remove_breakpoint(uprobe, vma, vaddr);
1495	}
1496	mmap_write_unlock(mm);
1497
1498	return err;
1499	}
1500
1501	static struct rb_node *
1502	find_node_in_range(struct inode *inode, loff_t min, loff_t max)
1503	{
1504	struct rb_node *n = uprobes_tree.rb_node;
1505
1506	while (n) {
1507	struct uprobe *u = rb_entry(n, struct uprobe, rb_node);
1508
1509	if (inode < u->inode) {
1510	n = n->rb_left;
1511	} else if (inode > u->inode) {
1512	n = n->rb_right;
1513	} else {
1514	if (max < u->offset)
1515	n = n->rb_left;
1516	else if (min > u->offset)
1517	n = n->rb_right;
1518	else
1519	break;
1520	}
1521	}
1522
1523	return n;
1524	}
1525
1526	/*
1527	* For a given range in vma, build a list of probes that need to be inserted.
1528	*/
1529	static void build_probe_list(struct inode *inode,
1530	struct vm_area_struct *vma,
1531	unsigned long start, unsigned long end,
1532	struct list_head *head)
1533	{
1534	loff_t min, max;
1535	struct rb_node n, t;
1536	struct uprobe *u;
1537
1538	INIT_LIST_HEAD(head);
1539	min = vaddr_to_offset(vma, start);
1540	max = min + (end - start) - 1;
1541
1542	read_lock(&uprobes_treelock);
1543	n = find_node_in_range(inode, min, max);
1544	if (n) {
1545	for (t = n; t; t = rb_prev(t)) {
1546	u = rb_entry(t, struct uprobe, rb_node);
1547	if (u->inode != inode \|\| u->offset < min)
1548	break;
1549	/* if uprobe went away, it's safe to ignore it */
1550	if (try_get_uprobe(u))
1551	list_add(&u->pending_list, head);
1552	}
1553	for (t = n; (t = rb_next(t)); ) {
1554	u = rb_entry(t, struct uprobe, rb_node);
1555	if (u->inode != inode \|\| u->offset > max)
1556	break;
1557	/* if uprobe went away, it's safe to ignore it */
1558	if (try_get_uprobe(u))
1559	list_add(&u->pending_list, head);
1560	}
1561	}
1562	read_unlock(&uprobes_treelock);
1563	}
1564
1565	/* @vma contains reference counter, not the probed instruction. */
1566	static int delayed_ref_ctr_inc(struct vm_area_struct *vma)
1567	{
1568	struct list_head pos, q;
1569	struct delayed_uprobe *du;
1570	unsigned long vaddr;
1571	int ret = 0, err = 0;
1572
1573	mutex_lock(&delayed_uprobe_lock);
1574	list_for_each_safe(pos, q, &delayed_uprobe_list) {
1575	du = list_entry(pos, struct delayed_uprobe, list);
1576
1577	if (du->mm != vma->vm_mm \|\|
1578	!valid_ref_ctr_vma(du->uprobe, vma))
1579	continue;
1580
1581	vaddr = offset_to_vaddr(vma, du->uprobe->ref_ctr_offset);
1582	ret = __update_ref_ctr(vma->vm_mm, vaddr, 1);
1583	if (ret) {
1584	update_ref_ctr_warn(du->uprobe, vma->vm_mm, 1);
1585	if (!err)
1586	err = ret;
1587	}
1588	delayed_uprobe_delete(du);
1589	}
1590	mutex_unlock(&delayed_uprobe_lock);
1591	return err;
1592	}
1593
1594	/*
1595	* Called from mmap_region/vma_merge with mm->mmap_lock acquired.
1596	*
1597	* Currently we ignore all errors and always return 0, the callers
1598	* can't handle the failure anyway.
1599	*/
1600	int uprobe_mmap(struct vm_area_struct *vma)
1601	{
1602	struct list_head tmp_list;
1603	struct uprobe uprobe, u;
1604	struct inode *inode;
1605
1606	if (no_uprobe_events())
1607	return 0;
1608
1609	if (vma->vm_file &&
1610	(vma->vm_flags & (VM_WRITE\|VM_SHARED)) == VM_WRITE &&
1611	mm_flags_test(MMF_HAS_UPROBES, vma->vm_mm))
1612	delayed_ref_ctr_inc(vma);
1613
1614	if (!valid_vma(vma, true))
1615	return 0;
1616
1617	inode = file_inode(vma->vm_file);
1618	if (!inode)
1619	return 0;
1620
1621	mutex_lock(uprobes_mmap_hash(inode));
1622	build_probe_list(inode, vma, vma->vm_start, vma->vm_end, &tmp_list);
1623	/*
1624	* We can race with uprobe_unregister(), this uprobe can be already
1625	* removed. But in this case filter_chain() must return false, all
1626	* consumers have gone away.
1627	*/
1628	list_for_each_entry_safe(uprobe, u, &tmp_list, pending_list) {
1629	if (!fatal_signal_pending(current) &&
1630	filter_chain(uprobe, vma->vm_mm)) {
1631	unsigned long vaddr = offset_to_vaddr(vma, uprobe->offset);
1632	install_breakpoint(uprobe, vma, vaddr);
1633	}
1634	put_uprobe(uprobe);
1635	}
1636	mutex_unlock(uprobes_mmap_hash(inode));
1637
1638	return 0;
1639	}
1640
1641	static bool
1642	vma_has_uprobes(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1643	{
1644	loff_t min, max;
1645	struct inode *inode;
1646	struct rb_node *n;
1647
1648	inode = file_inode(vma->vm_file);
1649
1650	min = vaddr_to_offset(vma, start);
1651	max = min + (end - start) - 1;
1652
1653	read_lock(&uprobes_treelock);
1654	n = find_node_in_range(inode, min, max);
1655	read_unlock(&uprobes_treelock);
1656
1657	return !!n;
1658	}
1659
1660	/*
1661	* Called in context of a munmap of a vma.
1662	*/
1663	void uprobe_munmap(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1664	{
1665	if (no_uprobe_events() \|\| !valid_vma(vma, false))
1666	return;
1667
1668	if (!atomic_read(&vma->vm_mm->mm_users)) /* called by mmput() ? */
1669	return;
1670
1671	if (!mm_flags_test(MMF_HAS_UPROBES, vma->vm_mm) \|\|
1672	mm_flags_test(MMF_RECALC_UPROBES, vma->vm_mm))
1673	return;
1674
1675	if (vma_has_uprobes(vma, start, end))
1676	mm_flags_set(MMF_RECALC_UPROBES, vma->vm_mm);
1677	}
1678
1679	static vm_fault_t xol_fault(const struct vm_special_mapping *sm,
1680	struct vm_area_struct vma, struct vm_fault vmf)
1681	{
1682	struct xol_area *area = vma->vm_mm->uprobes_state.xol_area;
1683
1684	vmf->page = area->page;
1685	get_page(vmf->page);
1686	return 0;
1687	}
1688
1689	static int xol_mremap(const struct vm_special_mapping sm, struct vm_area_struct new_vma)
1690	{
1691	return -EPERM;
1692	}
1693
1694	static const struct vm_special_mapping xol_mapping = {
1695	.name = "[uprobes]",
1696	.fault = xol_fault,
1697	.mremap = xol_mremap,
1698	};
1699
1700	unsigned long __weak arch_uprobe_get_xol_area(void)
1701	{
1702	/* Try to map as high as possible, this is only a hint. */
1703	return get_unmapped_area(NULL, TASK_SIZE - PAGE_SIZE, PAGE_SIZE, 0, 0);
1704	}
1705
1706	/* Slot allocation for XOL */
1707	static int xol_add_vma(struct mm_struct mm, struct xol_area area)
1708	{
1709	struct vm_area_struct *vma;
1710	int ret;
1711
1712	if (mmap_write_lock_killable(mm))
1713	return -EINTR;
1714
1715	if (mm->uprobes_state.xol_area) {
1716	ret = -EALREADY;
1717	goto fail;
1718	}
1719
1720	if (!area->vaddr) {
1721	area->vaddr = arch_uprobe_get_xol_area();
1722	if (IS_ERR_VALUE(area->vaddr)) {
1723	ret = area->vaddr;
1724	goto fail;
1725	}
1726	}
1727
1728	vma = _install_special_mapping(mm, area->vaddr, PAGE_SIZE,
1729	VM_EXEC\|VM_MAYEXEC\|VM_DONTCOPY\|VM_IO\|
1730	VM_SEALED_SYSMAP,
1731	&xol_mapping);
1732	if (IS_ERR(vma)) {
1733	ret = PTR_ERR(vma);
1734	goto fail;
1735	}
1736
1737	ret = 0;
1738	/* pairs with get_xol_area() */
1739	smp_store_release(&mm->uprobes_state.xol_area, area); /* ^^^ */
1740	fail:
1741	mmap_write_unlock(mm);
1742
1743	return ret;
1744	}
1745
1746	void * __weak arch_uretprobe_trampoline(unsigned long *psize)
1747	{
1748	static uprobe_opcode_t insn = UPROBE_SWBP_INSN;
1749
1750	*psize = UPROBE_SWBP_INSN_SIZE;
1751	return &insn;
1752	}
1753
1754	static struct xol_area *__create_xol_area(unsigned long vaddr)
1755	{
1756	struct mm_struct *mm = current->mm;
1757	unsigned long insns_size;
1758	struct xol_area *area;
1759	void *insns;
1760
1761	area = kzalloc_obj(*area);
1762	if (unlikely(!area))
1763	goto out;
1764
1765	area->bitmap = kcalloc(BITS_TO_LONGS(UINSNS_PER_PAGE), sizeof(long),
1766	GFP_KERNEL);
1767	if (!area->bitmap)
1768	goto free_area;
1769
1770	area->page = alloc_page(GFP_HIGHUSER \| __GFP_ZERO);
1771	if (!area->page)
1772	goto free_bitmap;
1773
1774	area->vaddr = vaddr;
1775	init_waitqueue_head(&area->wq);
1776	/* Reserve the 1st slot for get_trampoline_vaddr() */
1777	set_bit(0, area->bitmap);
1778	insns = arch_uretprobe_trampoline(&insns_size);
1779	arch_uprobe_copy_ixol(area->page, 0, insns, insns_size);
1780
1781	if (!xol_add_vma(mm, area))
1782	return area;
1783
1784	__free_page(area->page);
1785	free_bitmap:
1786	kfree(area->bitmap);
1787	free_area:
1788	kfree(area);
1789	out:
1790	return NULL;
1791	}
1792
1793	/*
1794	* get_xol_area - Allocate process's xol_area if necessary.
1795	* This area will be used for storing instructions for execution out of line.
1796	*
1797	* Returns the allocated area or NULL.
1798	*/
1799	static struct xol_area *get_xol_area(void)
1800	{
1801	struct mm_struct *mm = current->mm;
1802	struct xol_area *area;
1803
1804	if (!mm->uprobes_state.xol_area)
1805	__create_xol_area(0);
1806
1807	/* Pairs with xol_add_vma() smp_store_release() */
1808	area = READ_ONCE(mm->uprobes_state.xol_area); /* ^^^ */
1809	return area;
1810	}
1811
1812	/*
1813	* uprobe_clear_state - Free the area allocated for slots.
1814	*/
1815	void uprobe_clear_state(struct mm_struct *mm)
1816	{
1817	struct xol_area *area = mm->uprobes_state.xol_area;
1818
1819	mutex_lock(&delayed_uprobe_lock);
1820	delayed_uprobe_remove(NULL, mm);
1821	mutex_unlock(&delayed_uprobe_lock);
1822
1823	if (!area)
1824	return;
1825
1826	put_page(area->page);
1827	kfree(area->bitmap);
1828	kfree(area);
1829	}
1830
1831	void uprobe_start_dup_mmap(void)
1832	{
1833	percpu_down_read(&dup_mmap_sem);
1834	}
1835
1836	void uprobe_end_dup_mmap(void)
1837	{
1838	percpu_up_read(&dup_mmap_sem);
1839	}
1840
1841	void uprobe_dup_mmap(struct mm_struct oldmm, struct mm_struct newmm)
1842	{
1843	if (mm_flags_test(MMF_HAS_UPROBES, oldmm)) {
1844	mm_flags_set(MMF_HAS_UPROBES, newmm);
1845	/* unconditionally, dup_mmap() skips VM_DONTCOPY vmas */
1846	mm_flags_set(MMF_RECALC_UPROBES, newmm);
1847	}
1848	}
1849
1850	static unsigned long xol_get_slot_nr(struct xol_area *area)
1851	{
1852	unsigned long slot_nr;
1853
1854	slot_nr = find_first_zero_bit(area->bitmap, UINSNS_PER_PAGE);
1855	if (slot_nr < UINSNS_PER_PAGE) {
1856	if (!test_and_set_bit(slot_nr, area->bitmap))
1857	return slot_nr;
1858	}
1859
1860	return UINSNS_PER_PAGE;
1861	}
1862
1863	/*
1864	* xol_get_insn_slot - allocate a slot for xol.
1865	*/
1866	static bool xol_get_insn_slot(struct uprobe uprobe, struct uprobe_task utask)
1867	{
1868	struct xol_area *area = get_xol_area();
1869	unsigned long slot_nr;
1870
1871	if (!area)
1872	return false;
1873
1874	wait_event(area->wq, (slot_nr = xol_get_slot_nr(area)) < UINSNS_PER_PAGE);
1875
1876	utask->xol_vaddr = area->vaddr + slot_nr * UPROBE_XOL_SLOT_BYTES;
1877	arch_uprobe_copy_ixol(area->page, utask->xol_vaddr,
1878	&uprobe->arch.ixol, sizeof(uprobe->arch.ixol));
1879	return true;
1880	}
1881
1882	/*
1883	* xol_free_insn_slot - free the slot allocated by xol_get_insn_slot()
1884	*/
1885	static void xol_free_insn_slot(struct uprobe_task *utask)
1886	{
1887	struct xol_area *area = current->mm->uprobes_state.xol_area;
1888	unsigned long offset = utask->xol_vaddr - area->vaddr;
1889	unsigned int slot_nr;
1890
1891	utask->xol_vaddr = 0;
1892	/* xol_vaddr must fit into [area->vaddr, area->vaddr + PAGE_SIZE) */
1893	if (WARN_ON_ONCE(offset >= PAGE_SIZE))
1894	return;
1895
1896	slot_nr = offset / UPROBE_XOL_SLOT_BYTES;
1897	clear_bit(slot_nr, area->bitmap);
1898	smp_mb__after_atomic(); /* pairs with prepare_to_wait() */
1899	if (waitqueue_active(&area->wq))
1900	wake_up(&area->wq);
1901	}
1902
1903	void __weak arch_uprobe_copy_ixol(struct page *page, unsigned long vaddr,
1904	void *src, unsigned long len)
1905	{
1906	/* Initialize the slot */
1907	copy_to_page(page, vaddr, src, len);
1908
1909	/*
1910	* We probably need flush_icache_user_page() but it needs vma.
1911	* This should work on most of architectures by default. If
1912	* architecture needs to do something different it can define
1913	* its own version of the function.
1914	*/
1915	flush_dcache_page(page);
1916	}
1917
1918	/**
1919	* uprobe_get_swbp_addr - compute address of swbp given post-swbp regs
1920	* @regs: Reflects the saved state of the task after it has hit a breakpoint
1921	* instruction.
1922	* Return the address of the breakpoint instruction.
1923	*/
1924	unsigned long __weak uprobe_get_swbp_addr(struct pt_regs *regs)
1925	{
1926	return instruction_pointer(regs) - UPROBE_SWBP_INSN_SIZE;
1927	}
1928
1929	unsigned long uprobe_get_trap_addr(struct pt_regs *regs)
1930	{
1931	struct uprobe_task *utask = current->utask;
1932
1933	if (unlikely(utask && utask->active_uprobe))
1934	return utask->vaddr;
1935
1936	return instruction_pointer(regs);
1937	}
1938
1939	static void ri_pool_push(struct uprobe_task utask, struct return_instance ri)
1940	{
1941	ri->cons_cnt = 0;
1942	ri->next = utask->ri_pool;
1943	utask->ri_pool = ri;
1944	}
1945
1946	static struct return_instance ri_pool_pop(struct uprobe_task utask)
1947	{
1948	struct return_instance *ri = utask->ri_pool;
1949
1950	if (likely(ri))
1951	utask->ri_pool = ri->next;
1952
1953	return ri;
1954	}
1955
1956	static void ri_free(struct return_instance *ri)
1957	{
1958	kfree(ri->extra_consumers);
1959	kfree_rcu(ri, rcu);
1960	}
1961
1962	static void free_ret_instance(struct uprobe_task *utask,
1963	struct return_instance *ri, bool cleanup_hprobe)
1964	{
1965	unsigned seq;
1966
1967	if (cleanup_hprobe) {
1968	enum hprobe_state hstate;
1969
1970	(void)hprobe_consume(&ri->hprobe, &hstate);
1971	hprobe_finalize(&ri->hprobe, hstate);
1972	}
1973
1974	/*
1975	* At this point return_instance is unlinked from utask's
1976	* return_instances list and this has become visible to ri_timer().
1977	* If seqcount now indicates that ri_timer's return instance
1978	* processing loop isn't active, we can return ri into the pool of
1979	* to-be-reused return instances for future uretprobes. If ri_timer()
1980	* happens to be running right now, though, we fallback to safety and
1981	* just perform RCU-delated freeing of ri.
1982	* Admittedly, this is a rather simple use of seqcount, but it nicely
1983	* abstracts away all the necessary memory barriers, so we use
1984	* a well-supported kernel primitive here.
1985	*/
1986	if (raw_seqcount_try_begin(&utask->ri_seqcount, seq)) {
1987	/* immediate reuse of ri without RCU GP is OK */
1988	ri_pool_push(utask, ri);
1989	} else {
1990	/* we might be racing with ri_timer(), so play it safe */
1991	ri_free(ri);
1992	}
1993	}
1994
1995	/*
1996	* Called with no locks held.
1997	* Called in context of an exiting or an exec-ing thread.
1998	*/
1999	void uprobe_free_utask(struct task_struct *t)
2000	{
2001	struct uprobe_task *utask = t->utask;
2002	struct return_instance ri, ri_next;
2003
2004	if (!utask)
2005	return;
2006
2007	t->utask = NULL;
2008	WARN_ON_ONCE(utask->active_uprobe \|\| utask->xol_vaddr);
2009
2010	timer_delete_sync(&utask->ri_timer);
2011
2012	ri = utask->return_instances;
2013	while (ri) {
2014	ri_next = ri->next;
2015	free_ret_instance(utask, ri, true /* cleanup_hprobe */);
2016	ri = ri_next;
2017	}
2018
2019	/* free_ret_instance() above might add to ri_pool, so this loop should come last */
2020	ri = utask->ri_pool;
2021	while (ri) {
2022	ri_next = ri->next;
2023	ri_free(ri);
2024	ri = ri_next;
2025	}
2026
2027	kfree(utask);
2028	}
2029
2030	#define RI_TIMER_PERIOD (HZ / 10) /* 100 ms */
2031
2032	#define for_each_ret_instance_rcu(pos, head) \
2033	for (pos = rcu_dereference_raw(head); pos; pos = rcu_dereference_raw(pos->next))
2034
2035	static void ri_timer(struct timer_list *timer)
2036	{
2037	struct uprobe_task *utask = container_of(timer, struct uprobe_task, ri_timer);
2038	struct return_instance *ri;
2039
2040	/* SRCU protects uprobe from reuse for the cmpxchg() inside hprobe_expire(). */
2041	guard(srcu_fast_updown)(&uretprobes_srcu);
2042	/* RCU protects return_instance from freeing. */
2043	guard(rcu)();
2044
2045	/*
2046	* See free_ret_instance() for notes on seqcount use.
2047	* We also employ raw API variants to avoid lockdep false-positive
2048	* warning complaining about enabled preemption. The timer can only be
2049	* invoked once for a uprobe_task. Therefore there can only be one
2050	* writer. The reader does not require an even sequence count to make
2051	* progress, so it is OK to remain preemptible on PREEMPT_RT.
2052	*/
2053	raw_write_seqcount_begin(&utask->ri_seqcount);
2054
2055	for_each_ret_instance_rcu(ri, utask->return_instances)
2056	hprobe_expire(&ri->hprobe, false);
2057
2058	raw_write_seqcount_end(&utask->ri_seqcount);
2059	}
2060
2061	static struct uprobe_task *alloc_utask(void)
2062	{
2063	struct uprobe_task *utask;
2064
2065	utask = kzalloc_obj(*utask);
2066	if (!utask)
2067	return NULL;
2068
2069	timer_setup(&utask->ri_timer, ri_timer, 0);
2070	seqcount_init(&utask->ri_seqcount);
2071
2072	return utask;
2073	}
2074
2075	/*
2076	* Allocate a uprobe_task object for the task if necessary.
2077	* Called when the thread hits a breakpoint.
2078	*
2079	* Returns:
2080	* - pointer to new uprobe_task on success
2081	* - NULL otherwise
2082	*/
2083	static struct uprobe_task *get_utask(void)
2084	{
2085	if (!current->utask)
2086	current->utask = alloc_utask();
2087	return current->utask;
2088	}
2089
2090	static struct return_instance alloc_return_instance(struct uprobe_task utask)
2091	{
2092	struct return_instance *ri;
2093
2094	ri = ri_pool_pop(utask);
2095	if (ri)
2096	return ri;
2097
2098	ri = kzalloc_obj(*ri);
2099	if (!ri)
2100	return ZERO_SIZE_PTR;
2101
2102	return ri;
2103	}
2104
2105	static struct return_instance dup_return_instance(struct return_instance old)
2106	{
2107	struct return_instance *ri;
2108
2109	ri = kmemdup(old, sizeof(*ri), GFP_KERNEL);
2110	if (!ri)
2111	return NULL;
2112
2113	if (unlikely(old->cons_cnt > 1)) {
2114	ri->extra_consumers = kmemdup(old->extra_consumers,
2115	sizeof(ri->extra_consumers[0]) * (old->cons_cnt - 1),
2116	GFP_KERNEL);
2117	if (!ri->extra_consumers) {
2118	kfree(ri);
2119	return NULL;
2120	}
2121	}
2122
2123	return ri;
2124	}
2125
2126	static int dup_utask(struct task_struct t, struct uprobe_task o_utask)
2127	{
2128	struct uprobe_task *n_utask;
2129	struct return_instance *p, o, *n;
2130	struct uprobe *uprobe;
2131
2132	n_utask = alloc_utask();
2133	if (!n_utask)
2134	return -ENOMEM;
2135	t->utask = n_utask;
2136
2137	/* protect uprobes from freeing, we'll need try_get_uprobe() them */
2138	guard(srcu_fast_updown)(&uretprobes_srcu);
2139
2140	p = &n_utask->return_instances;
2141	for (o = o_utask->return_instances; o; o = o->next) {
2142	n = dup_return_instance(o);
2143	if (!n)
2144	return -ENOMEM;
2145
2146	/* if uprobe is non-NULL, we'll have an extra refcount for uprobe */
2147	uprobe = hprobe_expire(&o->hprobe, true);
2148
2149	/*
2150	* New utask will have stable properly refcounted uprobe or
2151	* NULL. Even if we failed to get refcounted uprobe, we still
2152	* need to preserve full set of return_instances for proper
2153	* uretprobe handling and nesting in forked task.
2154	*/
2155	hprobe_init_stable(&n->hprobe, uprobe);
2156
2157	n->next = NULL;
2158	rcu_assign_pointer(*p, n);
2159	p = &n->next;
2160
2161	n_utask->depth++;
2162	}
2163
2164	return 0;
2165	}
2166
2167	static void dup_xol_work(struct callback_head *work)
2168	{
2169	if (current->flags & PF_EXITING)
2170	return;
2171
2172	if (!__create_xol_area(current->utask->dup_xol_addr) &&
2173	!fatal_signal_pending(current))
2174	uprobe_warn(current, "dup xol area");
2175	}
2176
2177	/*
2178	* Called in context of a new clone/fork from copy_process.
2179	*/
2180	void uprobe_copy_process(struct task_struct *t, u64 flags)
2181	{
2182	struct uprobe_task *utask = current->utask;
2183	struct mm_struct *mm = current->mm;
2184	struct xol_area *area;
2185
2186	t->utask = NULL;
2187
2188	if (!utask \|\| !utask->return_instances)
2189	return;
2190
2191	if (mm == t->mm && !(flags & CLONE_VFORK))
2192	return;
2193
2194	if (dup_utask(t, utask))
2195	return uprobe_warn(t, "dup ret instances");
2196
2197	/* The task can fork() after dup_xol_work() fails */
2198	area = mm->uprobes_state.xol_area;
2199	if (!area)
2200	return uprobe_warn(t, "dup xol area");
2201
2202	if (mm == t->mm)
2203	return;
2204
2205	t->utask->dup_xol_addr = area->vaddr;
2206	init_task_work(&t->utask->dup_xol_work, dup_xol_work);
2207	task_work_add(t, &t->utask->dup_xol_work, TWA_RESUME);
2208	}
2209
2210	/*
2211	* Current area->vaddr notion assume the trampoline address is always
2212	* equal area->vaddr.
2213	*
2214	* Returns -1 in case the xol_area is not allocated.
2215	*/
2216	unsigned long uprobe_get_trampoline_vaddr(void)
2217	{
2218	unsigned long trampoline_vaddr = UPROBE_NO_TRAMPOLINE_VADDR;
2219	struct xol_area *area;
2220
2221	/* Pairs with xol_add_vma() smp_store_release() */
2222	area = READ_ONCE(current->mm->uprobes_state.xol_area); /* ^^^ */
2223	if (area)
2224	trampoline_vaddr = area->vaddr;
2225
2226	return trampoline_vaddr;
2227	}
2228
2229	static void cleanup_return_instances(struct uprobe_task *utask, bool chained,
2230	struct pt_regs *regs)
2231	{
2232	struct return_instance ri = utask->return_instances, ri_next;
2233	enum rp_check ctx = chained ? RP_CHECK_CHAIN_CALL : RP_CHECK_CALL;
2234
2235	while (ri && !arch_uretprobe_is_alive(ri, ctx, regs)) {
2236	ri_next = ri->next;
2237	rcu_assign_pointer(utask->return_instances, ri_next);
2238	utask->depth--;
2239
2240	free_ret_instance(utask, ri, true /* cleanup_hprobe */);
2241	ri = ri_next;
2242	}
2243	}
2244
2245	static void prepare_uretprobe(struct uprobe uprobe, struct pt_regs regs,
2246	struct return_instance *ri)
2247	{
2248	struct uprobe_task *utask = current->utask;
2249	unsigned long orig_ret_vaddr, trampoline_vaddr;
2250	struct srcu_ctr __percpu *srcu_scp;
2251	bool chained;
2252
2253	if (!get_xol_area())
2254	goto free;
2255
2256	if (utask->depth >= MAX_URETPROBE_DEPTH) {
2257	printk_ratelimited(KERN_INFO "uprobe: omit uretprobe due to"
2258	" nestedness limit pid/tgid=%d/%d\n",
2259	current->pid, current->tgid);
2260	goto free;
2261	}
2262
2263	trampoline_vaddr = uprobe_get_trampoline_vaddr();
2264	orig_ret_vaddr = arch_uretprobe_hijack_return_addr(trampoline_vaddr, regs);
2265	if (orig_ret_vaddr == -1)
2266	goto free;
2267
2268	/* drop the entries invalidated by longjmp() */
2269	chained = (orig_ret_vaddr == trampoline_vaddr);
2270	cleanup_return_instances(utask, chained, regs);
2271
2272	/*
2273	* We don't want to keep trampoline address in stack, rather keep the
2274	* original return address of first caller thru all the consequent
2275	* instances. This also makes breakpoint unwrapping easier.
2276	*/
2277	if (chained) {
2278	if (!utask->return_instances) {
2279	/*
2280	* This situation is not possible. Likely we have an
2281	* attack from user-space.
2282	*/
2283	uprobe_warn(current, "handle tail call");
2284	goto free;
2285	}
2286	orig_ret_vaddr = utask->return_instances->orig_ret_vaddr;
2287	}
2288
2289	/*
2290	* Use srcu_down_read_fast() because the SRCU lock survives a switch to
2291	* user space and can be unlocked from a different context by ri_timer()
2292	* or dup_utask().
2293	*/
2294	srcu_scp = srcu_down_read_fast(&uretprobes_srcu);
2295
2296	ri->func = instruction_pointer(regs);
2297	ri->stack = user_stack_pointer(regs);
2298	ri->orig_ret_vaddr = orig_ret_vaddr;
2299	ri->chained = chained;
2300
2301	utask->depth++;
2302
2303	hprobe_init_leased(&ri->hprobe, uprobe, srcu_scp);
2304	ri->next = utask->return_instances;
2305	rcu_assign_pointer(utask->return_instances, ri);
2306
2307	mod_timer(&utask->ri_timer, jiffies + RI_TIMER_PERIOD);
2308
2309	return;
2310	free:
2311	ri_free(ri);
2312	}
2313
2314	/* Prepare to single-step probed instruction out of line. */
2315	static int
2316	pre_ssout(struct uprobe uprobe, struct pt_regs regs, unsigned long bp_vaddr)
2317	{
2318	struct uprobe_task *utask = current->utask;
2319	int err;
2320
2321	if (!try_get_uprobe(uprobe))
2322	return -EINVAL;
2323
2324	if (!xol_get_insn_slot(uprobe, utask)) {
2325	err = -ENOMEM;
2326	goto err_out;
2327	}
2328
2329	utask->vaddr = bp_vaddr;
2330	err = arch_uprobe_pre_xol(&uprobe->arch, regs);
2331	if (unlikely(err)) {
2332	xol_free_insn_slot(utask);
2333	goto err_out;
2334	}
2335
2336	utask->active_uprobe = uprobe;
2337	utask->state = UTASK_SSTEP;
2338	return 0;
2339	err_out:
2340	put_uprobe(uprobe);
2341	return err;
2342	}
2343
2344	/*
2345	* If we are singlestepping, then ensure this thread is not connected to
2346	* non-fatal signals until completion of singlestep. When xol insn itself
2347	* triggers the signal, restart the original insn even if the task is
2348	* already SIGKILL'ed (since coredump should report the correct ip). This
2349	* is even more important if the task has a handler for SIGSEGV/etc, The
2350	* _same_ instruction should be repeated again after return from the signal
2351	* handler, and SSTEP can never finish in this case.
2352	*/
2353	bool uprobe_deny_signal(void)
2354	{
2355	struct task_struct *t = current;
2356	struct uprobe_task *utask = t->utask;
2357
2358	if (likely(!utask \|\| !utask->active_uprobe))
2359	return false;
2360
2361	WARN_ON_ONCE(utask->state != UTASK_SSTEP);
2362
2363	if (task_sigpending(t)) {
2364	utask->signal_denied = true;
2365	clear_tsk_thread_flag(t, TIF_SIGPENDING);
2366
2367	if (__fatal_signal_pending(t) \|\| arch_uprobe_xol_was_trapped(t)) {
2368	utask->state = UTASK_SSTEP_TRAPPED;
2369	set_tsk_thread_flag(t, TIF_UPROBE);
2370	}
2371	}
2372
2373	return true;
2374	}
2375
2376	static void mmf_recalc_uprobes(struct mm_struct *mm)
2377	{
2378	VMA_ITERATOR(vmi, mm, 0);
2379	struct vm_area_struct *vma;
2380
2381	for_each_vma(vmi, vma) {
2382	if (!valid_vma(vma, false))
2383	continue;
2384	/*
2385	* This is not strictly accurate, we can race with
2386	* uprobe_unregister() and see the already removed
2387	* uprobe if delete_uprobe() was not yet called.
2388	* Or this uprobe can be filtered out.
2389	*/
2390	if (vma_has_uprobes(vma, vma->vm_start, vma->vm_end))
2391	return;
2392	}
2393
2394	mm_flags_clear(MMF_HAS_UPROBES, mm);
2395	}
2396
2397	static int is_trap_at_addr(struct mm_struct *mm, unsigned long vaddr)
2398	{
2399	struct page *page;
2400	uprobe_opcode_t opcode;
2401	int result;
2402
2403	if (WARN_ON_ONCE(!IS_ALIGNED(vaddr, UPROBE_SWBP_INSN_SIZE)))
2404	return -EINVAL;
2405
2406	pagefault_disable();
2407	result = __get_user(opcode, (uprobe_opcode_t __user *)vaddr);
2408	pagefault_enable();
2409
2410	if (likely(result == 0))
2411	goto out;
2412
2413	result = get_user_pages(vaddr, 1, FOLL_FORCE, &page);
2414	if (result < 0)
2415	return result;
2416
2417	uprobe_copy_from_page(page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
2418	put_page(page);
2419	out:
2420	/* This needs to return true for any variant of the trap insn */
2421	return is_trap_insn(&opcode);
2422	}
2423
2424	static struct uprobe *find_active_uprobe_speculative(unsigned long bp_vaddr)
2425	{
2426	struct mm_struct *mm = current->mm;
2427	struct uprobe *uprobe = NULL;
2428	struct vm_area_struct *vma;
2429	struct file *vm_file;
2430	loff_t offset;
2431	unsigned int seq;
2432
2433	guard(rcu)();
2434
2435	if (!mmap_lock_speculate_try_begin(mm, &seq))
2436	return NULL;
2437
2438	vma = vma_lookup(mm, bp_vaddr);
2439	if (!vma)
2440	return NULL;
2441
2442	/*
2443	* vm_file memory can be reused for another instance of struct file,
2444	* but can't be freed from under us, so it's safe to read fields from
2445	* it, even if the values are some garbage values; ultimately
2446	* find_uprobe_rcu() + mmap_lock_speculation_end() check will ensure
2447	* that whatever we speculatively found is correct
2448	*/
2449	vm_file = READ_ONCE(vma->vm_file);
2450	if (!vm_file)
2451	return NULL;
2452
2453	offset = (loff_t)(vma_start_pgoff(vma) << PAGE_SHIFT) +
2454	(bp_vaddr - vma->vm_start);
2455	uprobe = find_uprobe_rcu(vm_file->f_inode, offset);
2456	if (!uprobe)
2457	return NULL;
2458
2459	/* now double check that nothing about MM changed */
2460	if (mmap_lock_speculate_retry(mm, seq))
2461	return NULL;
2462
2463	return uprobe;
2464	}
2465
2466	/* assumes being inside RCU protected region */
2467	static struct uprobe find_active_uprobe_rcu(unsigned long bp_vaddr, int is_swbp)
2468	{
2469	struct mm_struct *mm = current->mm;
2470	struct uprobe *uprobe = NULL;
2471	struct vm_area_struct *vma;
2472
2473	uprobe = find_active_uprobe_speculative(bp_vaddr);
2474	if (uprobe)
2475	return uprobe;
2476
2477	mmap_read_lock(mm);
2478	vma = vma_lookup(mm, bp_vaddr);
2479	if (vma) {
2480	if (vma->vm_file) {
2481	struct inode *inode = file_inode(vma->vm_file);
2482	loff_t offset = vaddr_to_offset(vma, bp_vaddr);
2483
2484	uprobe = find_uprobe_rcu(inode, offset);
2485	}
2486
2487	if (!uprobe)
2488	*is_swbp = is_trap_at_addr(mm, bp_vaddr);
2489	} else {
2490	*is_swbp = -EFAULT;
2491	}
2492
2493	if (!uprobe && mm_flags_test_and_clear(MMF_RECALC_UPROBES, mm))
2494	mmf_recalc_uprobes(mm);
2495	mmap_read_unlock(mm);
2496
2497	return uprobe;
2498	}
2499
2500	static struct return_instance push_consumer(struct return_instance ri, __u64 id, __u64 cookie)
2501	{
2502	struct return_consumer *ric;
2503
2504	if (unlikely(ri == ZERO_SIZE_PTR))
2505	return ri;
2506
2507	if (unlikely(ri->cons_cnt > 0)) {
2508	ric = krealloc(ri->extra_consumers, sizeof(ric) ri->cons_cnt, GFP_KERNEL);
2509	if (!ric) {
2510	ri_free(ri);
2511	return ZERO_SIZE_PTR;
2512	}
2513	ri->extra_consumers = ric;
2514	}
2515
2516	ric = likely(ri->cons_cnt == 0) ? &ri->consumer : &ri->extra_consumers[ri->cons_cnt - 1];
2517	ric->id = id;
2518	ric->cookie = cookie;
2519
2520	ri->cons_cnt++;
2521	return ri;
2522	}
2523
2524	static struct return_consumer *
2525	return_consumer_find(struct return_instance ri, int iter, int id)
2526	{
2527	struct return_consumer *ric;
2528	int idx;
2529
2530	for (idx = *iter; idx < ri->cons_cnt; idx++)
2531	{
2532	ric = likely(idx == 0) ? &ri->consumer : &ri->extra_consumers[idx - 1];
2533	if (ric->id == id) {
2534	*iter = idx + 1;
2535	return ric;
2536	}
2537	}
2538
2539	return NULL;
2540	}
2541
2542	static bool ignore_ret_handler(int rc)
2543	{
2544	return rc == UPROBE_HANDLER_REMOVE \|\| rc == UPROBE_HANDLER_IGNORE;
2545	}
2546
2547	static void handler_chain(struct uprobe uprobe, struct pt_regs regs)
2548	{
2549	struct uprobe_consumer *uc;
2550	bool has_consumers = false, remove = true;
2551	struct return_instance *ri = NULL;
2552	struct uprobe_task *utask = current->utask;
2553
2554	utask->auprobe = &uprobe->arch;
2555
2556	list_for_each_entry_rcu(uc, &uprobe->consumers, cons_node, rcu_read_lock_trace_held()) {
2557	bool session = uc->handler && uc->ret_handler;
2558	__u64 cookie = 0;
2559	int rc = 0;
2560
2561	if (uc->handler) {
2562	rc = uc->handler(uc, regs, &cookie);
2563	WARN(rc < 0 \|\| rc > 2,
2564	"bad rc=0x%x from %ps()\n", rc, uc->handler);
2565	}
2566
2567	remove &= rc == UPROBE_HANDLER_REMOVE;
2568	has_consumers = true;
2569
2570	if (!uc->ret_handler \|\| ignore_ret_handler(rc))
2571	continue;
2572
2573	if (!ri)
2574	ri = alloc_return_instance(utask);
2575
2576	if (session)
2577	ri = push_consumer(ri, uc->id, cookie);
2578	}
2579	utask->auprobe = NULL;
2580
2581	if (!ZERO_OR_NULL_PTR(ri))
2582	prepare_uretprobe(uprobe, regs, ri);
2583
2584	if (remove && has_consumers) {
2585	down_read(&uprobe->register_rwsem);
2586
2587	/* re-check that removal is still required, this time under lock */
2588	if (!filter_chain(uprobe, current->mm)) {
2589	WARN_ON(!uprobe_is_active(uprobe));
2590	unapply_uprobe(uprobe, current->mm);
2591	}
2592
2593	up_read(&uprobe->register_rwsem);
2594	}
2595	}
2596
2597	static void
2598	handle_uretprobe_chain(struct return_instance ri, struct uprobe uprobe, struct pt_regs *regs)
2599	{
2600	struct return_consumer *ric;
2601	struct uprobe_consumer *uc;
2602	int ric_idx = 0;
2603
2604	/* all consumers unsubscribed meanwhile */
2605	if (unlikely(!uprobe))
2606	return;
2607
2608	rcu_read_lock_trace();
2609	list_for_each_entry_rcu(uc, &uprobe->consumers, cons_node, rcu_read_lock_trace_held()) {
2610	bool session = uc->handler && uc->ret_handler;
2611
2612	if (uc->ret_handler) {
2613	ric = return_consumer_find(ri, &ric_idx, uc->id);
2614	if (!session \|\| ric)
2615	uc->ret_handler(uc, ri->func, regs, ric ? &ric->cookie : NULL);
2616	}
2617	}
2618	rcu_read_unlock_trace();
2619	}
2620
2621	static struct return_instance find_next_ret_chain(struct return_instance ri)
2622	{
2623	bool chained;
2624
2625	do {
2626	chained = ri->chained;
2627	ri = ri->next; /* can't be NULL if chained */
2628	} while (chained);
2629
2630	return ri;
2631	}
2632
2633	void uprobe_handle_trampoline(struct pt_regs *regs)
2634	{
2635	struct uprobe_task *utask;
2636	struct return_instance ri, ri_next, *next_chain;
2637	struct uprobe *uprobe;
2638	enum hprobe_state hstate;
2639	bool valid;
2640
2641	utask = current->utask;
2642	if (!utask)
2643	goto sigill;
2644
2645	ri = utask->return_instances;
2646	if (!ri)
2647	goto sigill;
2648
2649	do {
2650	/*
2651	* We should throw out the frames invalidated by longjmp().
2652	* If this chain is valid, then the next one should be alive
2653	* or NULL; the latter case means that nobody but ri->func
2654	* could hit this trampoline on return. TODO: sigaltstack().
2655	*/
2656	next_chain = find_next_ret_chain(ri);
2657	valid = !next_chain \|\| arch_uretprobe_is_alive(next_chain, RP_CHECK_RET, regs);
2658
2659	instruction_pointer_set(regs, ri->orig_ret_vaddr);
2660	do {
2661	/* pop current instance from the stack of pending return instances,
2662	* as it's not pending anymore: we just fixed up original
2663	* instruction pointer in regs and are about to call handlers;
2664	* this allows fixup_uretprobe_trampoline_entries() to properly fix up
2665	* captured stack traces from uretprobe handlers, in which pending
2666	* trampoline addresses on the stack are replaced with correct
2667	* original return addresses
2668	*/
2669	ri_next = ri->next;
2670	rcu_assign_pointer(utask->return_instances, ri_next);
2671	utask->depth--;
2672
2673	uprobe = hprobe_consume(&ri->hprobe, &hstate);
2674	if (valid)
2675	handle_uretprobe_chain(ri, uprobe, regs);
2676	hprobe_finalize(&ri->hprobe, hstate);
2677
2678	/* We already took care of hprobe, no need to waste more time on that. */
2679	free_ret_instance(utask, ri, false /* !cleanup_hprobe */);
2680	ri = ri_next;
2681	} while (ri != next_chain);
2682	} while (!valid);
2683
2684	return;
2685
2686	sigill:
2687	uprobe_warn(current, "handle uretprobe, sending SIGILL.");
2688	force_sig(SIGILL);
2689	}
2690
2691	bool __weak arch_uprobe_ignore(struct arch_uprobe aup, struct pt_regs regs)
2692	{
2693	return false;
2694	}
2695
2696	bool __weak arch_uretprobe_is_alive(struct return_instance *ret, enum rp_check ctx,
2697	struct pt_regs *regs)
2698	{
2699	return true;
2700	}
2701
2702	void __weak arch_uprobe_optimize(struct arch_uprobe *auprobe, unsigned long vaddr)
2703	{
2704	}
2705
2706	/*
2707	* Run handler and ask thread to singlestep.
2708	* Ensure all non-fatal signals cannot interrupt thread while it singlesteps.
2709	*/
2710	static void handle_swbp(struct pt_regs *regs)
2711	{
2712	struct uprobe *uprobe;
2713	unsigned long bp_vaddr;
2714	int is_swbp;
2715
2716	bp_vaddr = uprobe_get_swbp_addr(regs);
2717	if (bp_vaddr == uprobe_get_trampoline_vaddr())
2718	return uprobe_handle_trampoline(regs);
2719
2720	rcu_read_lock_trace();
2721
2722	uprobe = find_active_uprobe_rcu(bp_vaddr, &is_swbp);
2723	if (!uprobe) {
2724	if (is_swbp > 0) {
2725	/* No matching uprobe; signal SIGTRAP. */
2726	force_sig(SIGTRAP);
2727	} else {
2728	/*
2729	* Either we raced with uprobe_unregister() or we can't
2730	* access this memory. The latter is only possible if
2731	* another thread plays with our ->mm. In both cases
2732	* we can simply restart. If this vma was unmapped we
2733	* can pretend this insn was not executed yet and get
2734	* the (correct) SIGSEGV after restart.
2735	*/
2736	instruction_pointer_set(regs, bp_vaddr);
2737	}
2738	goto out;
2739	}
2740
2741	/* change it in advance for ->handler() and restart */
2742	instruction_pointer_set(regs, bp_vaddr);
2743
2744	/*
2745	* TODO: move copy_insn/etc into _register and remove this hack.
2746	* After we hit the bp, _unregister + _register can install the
2747	* new and not-yet-analyzed uprobe at the same address, restart.
2748	*/
2749	if (unlikely(!test_bit(UPROBE_COPY_INSN, &uprobe->flags)))
2750	goto out;
2751
2752	/*
2753	* Pairs with the smp_wmb() in prepare_uprobe().
2754	*
2755	* Guarantees that if we see the UPROBE_COPY_INSN bit set, then
2756	* we must also see the stores to &uprobe->arch performed by the
2757	* prepare_uprobe() call.
2758	*/
2759	smp_rmb();
2760
2761	/* Tracing handlers use ->utask to communicate with fetch methods */
2762	if (!get_utask())
2763	goto out;
2764
2765	if (arch_uprobe_ignore(&uprobe->arch, regs))
2766	goto out;
2767
2768	handler_chain(uprobe, regs);
2769
2770	/* Try to optimize after first hit. */
2771	arch_uprobe_optimize(&uprobe->arch, bp_vaddr);
2772
2773	/*
2774	* If user decided to take execution elsewhere, it makes little sense
2775	* to execute the original instruction, so let's skip it.
2776	*/
2777	if (instruction_pointer(regs) != bp_vaddr)
2778	goto out;
2779
2780	if (arch_uprobe_skip_sstep(&uprobe->arch, regs))
2781	goto out;
2782
2783	if (pre_ssout(uprobe, regs, bp_vaddr))
2784	goto out;
2785
2786	out:
2787	/* arch_uprobe_skip_sstep() succeeded, or restart if can't singlestep */
2788	rcu_read_unlock_trace();
2789	}
2790
2791	void handle_syscall_uprobe(struct pt_regs *regs, unsigned long bp_vaddr)
2792	{
2793	struct uprobe *uprobe;
2794	int is_swbp;
2795
2796	guard(rcu_tasks_trace)();
2797
2798	uprobe = find_active_uprobe_rcu(bp_vaddr, &is_swbp);
2799	if (!uprobe)
2800	return;
2801	if (!get_utask())
2802	return;
2803	if (arch_uprobe_ignore(&uprobe->arch, regs))
2804	return;
2805	handler_chain(uprobe, regs);
2806	}
2807
2808	/*
2809	* Perform required fix-ups and disable singlestep.
2810	* Allow pending signals to take effect.
2811	*/
2812	static void handle_singlestep(struct uprobe_task utask, struct pt_regs regs)
2813	{
2814	struct uprobe *uprobe;
2815	int err = 0;
2816
2817	uprobe = utask->active_uprobe;
2818	if (utask->state == UTASK_SSTEP_ACK)
2819	err = arch_uprobe_post_xol(&uprobe->arch, regs);
2820	else if (utask->state == UTASK_SSTEP_TRAPPED)
2821	arch_uprobe_abort_xol(&uprobe->arch, regs);
2822	else
2823	WARN_ON_ONCE(1);
2824
2825	put_uprobe(uprobe);
2826	utask->active_uprobe = NULL;
2827	utask->state = UTASK_RUNNING;
2828	xol_free_insn_slot(utask);
2829
2830	if (utask->signal_denied) {
2831	set_thread_flag(TIF_SIGPENDING);
2832	utask->signal_denied = false;
2833	}
2834
2835	if (unlikely(err)) {
2836	uprobe_warn(current, "execute the probed insn, sending SIGILL.");
2837	force_sig(SIGILL);
2838	}
2839	}
2840
2841	/*
2842	* On breakpoint hit, breakpoint notifier sets the TIF_UPROBE flag and
2843	* allows the thread to return from interrupt. After that handle_swbp()
2844	* sets utask->active_uprobe.
2845	*
2846	* On singlestep exception, singlestep notifier sets the TIF_UPROBE flag
2847	* and allows the thread to return from interrupt.
2848	*
2849	* While returning to userspace, thread notices the TIF_UPROBE flag and calls
2850	* uprobe_notify_resume().
2851	*/
2852	void uprobe_notify_resume(struct pt_regs *regs)
2853	{
2854	struct uprobe_task *utask;
2855
2856	clear_thread_flag(TIF_UPROBE);
2857
2858	utask = current->utask;
2859	if (utask && utask->active_uprobe)
2860	handle_singlestep(utask, regs);
2861	else
2862	handle_swbp(regs);
2863	}
2864
2865	/*
2866	* uprobe_pre_sstep_notifier gets called from interrupt context as part of
2867	* notifier mechanism. Set TIF_UPROBE flag and indicate breakpoint hit.
2868	*/
2869	int uprobe_pre_sstep_notifier(struct pt_regs *regs)
2870	{
2871	if (!current->mm)
2872	return 0;
2873
2874	if (!mm_flags_test(MMF_HAS_UPROBES, current->mm) &&
2875	(!current->utask \|\| !current->utask->return_instances))
2876	return 0;
2877
2878	set_thread_flag(TIF_UPROBE);
2879	return 1;
2880	}
2881
2882	/*
2883	* uprobe_post_sstep_notifier gets called in interrupt context as part of notifier
2884	* mechanism. Set TIF_UPROBE flag and indicate completion of singlestep.
2885	*/
2886	int uprobe_post_sstep_notifier(struct pt_regs *regs)
2887	{
2888	struct uprobe_task *utask = current->utask;
2889
2890	if (!current->mm \|\| !utask \|\| !utask->active_uprobe)
2891	/* task is currently not uprobed */
2892	return 0;
2893
2894	utask->state = UTASK_SSTEP_ACK;
2895	set_thread_flag(TIF_UPROBE);
2896	return 1;
2897	}
2898
2899	static struct notifier_block uprobe_exception_nb = {
2900	.notifier_call = arch_uprobe_exception_notify,
2901	.priority = INT_MAX-1, /* notified after kprobes, kgdb */
2902	};
2903
2904	void __init uprobes_init(void)
2905	{
2906	int i;
2907
2908	for (i = 0; i < UPROBES_HASH_SZ; i++)
2909	mutex_init(&uprobes_mmap_mutex[i]);
2910
2911	BUG_ON(register_die_notifier(&uprobe_exception_nb));
2912	}