]>
Commit | Line | Data |
---|---|---|
20c8ccb1 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
86039bd3 AA |
2 | /* |
3 | * fs/userfaultfd.c | |
4 | * | |
5 | * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org> | |
6 | * Copyright (C) 2008-2009 Red Hat, Inc. | |
7 | * Copyright (C) 2015 Red Hat, Inc. | |
8 | * | |
86039bd3 AA |
9 | * Some part derived from fs/eventfd.c (anon inode setup) and |
10 | * mm/ksm.c (mm hashing). | |
11 | */ | |
12 | ||
9cd75c3c | 13 | #include <linux/list.h> |
86039bd3 | 14 | #include <linux/hashtable.h> |
174cd4b1 | 15 | #include <linux/sched/signal.h> |
6e84f315 | 16 | #include <linux/sched/mm.h> |
86039bd3 AA |
17 | #include <linux/mm.h> |
18 | #include <linux/poll.h> | |
19 | #include <linux/slab.h> | |
20 | #include <linux/seq_file.h> | |
21 | #include <linux/file.h> | |
22 | #include <linux/bug.h> | |
23 | #include <linux/anon_inodes.h> | |
24 | #include <linux/syscalls.h> | |
25 | #include <linux/userfaultfd_k.h> | |
26 | #include <linux/mempolicy.h> | |
27 | #include <linux/ioctl.h> | |
28 | #include <linux/security.h> | |
cab350af | 29 | #include <linux/hugetlb.h> |
86039bd3 | 30 | |
cefdca0a PX |
31 | int sysctl_unprivileged_userfaultfd __read_mostly = 1; |
32 | ||
3004ec9c AA |
33 | static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly; |
34 | ||
86039bd3 AA |
35 | enum userfaultfd_state { |
36 | UFFD_STATE_WAIT_API, | |
37 | UFFD_STATE_RUNNING, | |
38 | }; | |
39 | ||
3004ec9c AA |
40 | /* |
41 | * Start with fault_pending_wqh and fault_wqh so they're more likely | |
42 | * to be in the same cacheline. | |
cbcfa130 EB |
43 | * |
44 | * Locking order: | |
45 | * fd_wqh.lock | |
46 | * fault_pending_wqh.lock | |
47 | * fault_wqh.lock | |
48 | * event_wqh.lock | |
49 | * | |
50 | * To avoid deadlocks, IRQs must be disabled when taking any of the above locks, | |
51 | * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's | |
52 | * also taken in IRQ context. | |
3004ec9c | 53 | */ |
86039bd3 | 54 | struct userfaultfd_ctx { |
15b726ef AA |
55 | /* waitqueue head for the pending (i.e. not read) userfaults */ |
56 | wait_queue_head_t fault_pending_wqh; | |
57 | /* waitqueue head for the userfaults */ | |
86039bd3 AA |
58 | wait_queue_head_t fault_wqh; |
59 | /* waitqueue head for the pseudo fd to wakeup poll/read */ | |
60 | wait_queue_head_t fd_wqh; | |
9cd75c3c PE |
61 | /* waitqueue head for events */ |
62 | wait_queue_head_t event_wqh; | |
2c5b7e1b AA |
63 | /* a refile sequence protected by fault_pending_wqh lock */ |
64 | struct seqcount refile_seq; | |
3004ec9c | 65 | /* pseudo fd refcounting */ |
ca880420 | 66 | refcount_t refcount; |
86039bd3 AA |
67 | /* userfaultfd syscall flags */ |
68 | unsigned int flags; | |
9cd75c3c PE |
69 | /* features requested from the userspace */ |
70 | unsigned int features; | |
86039bd3 AA |
71 | /* state machine */ |
72 | enum userfaultfd_state state; | |
73 | /* released */ | |
74 | bool released; | |
df2cc96e MR |
75 | /* memory mappings are changing because of non-cooperative event */ |
76 | bool mmap_changing; | |
86039bd3 AA |
77 | /* mm with one ore more vmas attached to this userfaultfd_ctx */ |
78 | struct mm_struct *mm; | |
79 | }; | |
80 | ||
893e26e6 PE |
81 | struct userfaultfd_fork_ctx { |
82 | struct userfaultfd_ctx *orig; | |
83 | struct userfaultfd_ctx *new; | |
84 | struct list_head list; | |
85 | }; | |
86 | ||
897ab3e0 MR |
87 | struct userfaultfd_unmap_ctx { |
88 | struct userfaultfd_ctx *ctx; | |
89 | unsigned long start; | |
90 | unsigned long end; | |
91 | struct list_head list; | |
92 | }; | |
93 | ||
86039bd3 | 94 | struct userfaultfd_wait_queue { |
a9b85f94 | 95 | struct uffd_msg msg; |
ac6424b9 | 96 | wait_queue_entry_t wq; |
86039bd3 | 97 | struct userfaultfd_ctx *ctx; |
15a77c6f | 98 | bool waken; |
86039bd3 AA |
99 | }; |
100 | ||
101 | struct userfaultfd_wake_range { | |
102 | unsigned long start; | |
103 | unsigned long len; | |
104 | }; | |
105 | ||
ac6424b9 | 106 | static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode, |
86039bd3 AA |
107 | int wake_flags, void *key) |
108 | { | |
109 | struct userfaultfd_wake_range *range = key; | |
110 | int ret; | |
111 | struct userfaultfd_wait_queue *uwq; | |
112 | unsigned long start, len; | |
113 | ||
114 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
115 | ret = 0; | |
86039bd3 AA |
116 | /* len == 0 means wake all */ |
117 | start = range->start; | |
118 | len = range->len; | |
a9b85f94 AA |
119 | if (len && (start > uwq->msg.arg.pagefault.address || |
120 | start + len <= uwq->msg.arg.pagefault.address)) | |
86039bd3 | 121 | goto out; |
15a77c6f AA |
122 | WRITE_ONCE(uwq->waken, true); |
123 | /* | |
a9668cd6 PZ |
124 | * The Program-Order guarantees provided by the scheduler |
125 | * ensure uwq->waken is visible before the task is woken. | |
15a77c6f | 126 | */ |
86039bd3 | 127 | ret = wake_up_state(wq->private, mode); |
a9668cd6 | 128 | if (ret) { |
86039bd3 AA |
129 | /* |
130 | * Wake only once, autoremove behavior. | |
131 | * | |
a9668cd6 PZ |
132 | * After the effect of list_del_init is visible to the other |
133 | * CPUs, the waitqueue may disappear from under us, see the | |
134 | * !list_empty_careful() in handle_userfault(). | |
135 | * | |
136 | * try_to_wake_up() has an implicit smp_mb(), and the | |
137 | * wq->private is read before calling the extern function | |
138 | * "wake_up_state" (which in turns calls try_to_wake_up). | |
86039bd3 | 139 | */ |
2055da97 | 140 | list_del_init(&wq->entry); |
a9668cd6 | 141 | } |
86039bd3 AA |
142 | out: |
143 | return ret; | |
144 | } | |
145 | ||
146 | /** | |
147 | * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd | |
148 | * context. | |
149 | * @ctx: [in] Pointer to the userfaultfd context. | |
86039bd3 AA |
150 | */ |
151 | static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx) | |
152 | { | |
ca880420 | 153 | refcount_inc(&ctx->refcount); |
86039bd3 AA |
154 | } |
155 | ||
156 | /** | |
157 | * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd | |
158 | * context. | |
159 | * @ctx: [in] Pointer to userfaultfd context. | |
160 | * | |
161 | * The userfaultfd context reference must have been previously acquired either | |
162 | * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget(). | |
163 | */ | |
164 | static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx) | |
165 | { | |
ca880420 | 166 | if (refcount_dec_and_test(&ctx->refcount)) { |
86039bd3 AA |
167 | VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock)); |
168 | VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh)); | |
169 | VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock)); | |
170 | VM_BUG_ON(waitqueue_active(&ctx->fault_wqh)); | |
9cd75c3c PE |
171 | VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock)); |
172 | VM_BUG_ON(waitqueue_active(&ctx->event_wqh)); | |
86039bd3 AA |
173 | VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock)); |
174 | VM_BUG_ON(waitqueue_active(&ctx->fd_wqh)); | |
d2005e3f | 175 | mmdrop(ctx->mm); |
3004ec9c | 176 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
86039bd3 AA |
177 | } |
178 | } | |
179 | ||
a9b85f94 | 180 | static inline void msg_init(struct uffd_msg *msg) |
86039bd3 | 181 | { |
a9b85f94 AA |
182 | BUILD_BUG_ON(sizeof(struct uffd_msg) != 32); |
183 | /* | |
184 | * Must use memset to zero out the paddings or kernel data is | |
185 | * leaked to userland. | |
186 | */ | |
187 | memset(msg, 0, sizeof(struct uffd_msg)); | |
188 | } | |
189 | ||
190 | static inline struct uffd_msg userfault_msg(unsigned long address, | |
191 | unsigned int flags, | |
9d4ac934 AP |
192 | unsigned long reason, |
193 | unsigned int features) | |
a9b85f94 AA |
194 | { |
195 | struct uffd_msg msg; | |
196 | msg_init(&msg); | |
197 | msg.event = UFFD_EVENT_PAGEFAULT; | |
198 | msg.arg.pagefault.address = address; | |
86039bd3 AA |
199 | if (flags & FAULT_FLAG_WRITE) |
200 | /* | |
a4605a61 | 201 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the |
a9b85f94 AA |
202 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE |
203 | * was not set in a UFFD_EVENT_PAGEFAULT, it means it | |
204 | * was a read fault, otherwise if set it means it's | |
205 | * a write fault. | |
86039bd3 | 206 | */ |
a9b85f94 | 207 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE; |
86039bd3 AA |
208 | if (reason & VM_UFFD_WP) |
209 | /* | |
a9b85f94 AA |
210 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the |
211 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was | |
212 | * not set in a UFFD_EVENT_PAGEFAULT, it means it was | |
213 | * a missing fault, otherwise if set it means it's a | |
214 | * write protect fault. | |
86039bd3 | 215 | */ |
a9b85f94 | 216 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP; |
9d4ac934 | 217 | if (features & UFFD_FEATURE_THREAD_ID) |
a36985d3 | 218 | msg.arg.pagefault.feat.ptid = task_pid_vnr(current); |
a9b85f94 | 219 | return msg; |
86039bd3 AA |
220 | } |
221 | ||
369cd212 MK |
222 | #ifdef CONFIG_HUGETLB_PAGE |
223 | /* | |
224 | * Same functionality as userfaultfd_must_wait below with modifications for | |
225 | * hugepmd ranges. | |
226 | */ | |
227 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 228 | struct vm_area_struct *vma, |
369cd212 MK |
229 | unsigned long address, |
230 | unsigned long flags, | |
231 | unsigned long reason) | |
232 | { | |
233 | struct mm_struct *mm = ctx->mm; | |
1e2c0436 | 234 | pte_t *ptep, pte; |
369cd212 MK |
235 | bool ret = true; |
236 | ||
237 | VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem)); | |
238 | ||
1e2c0436 JF |
239 | ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma)); |
240 | ||
241 | if (!ptep) | |
369cd212 MK |
242 | goto out; |
243 | ||
244 | ret = false; | |
1e2c0436 | 245 | pte = huge_ptep_get(ptep); |
369cd212 MK |
246 | |
247 | /* | |
248 | * Lockless access: we're in a wait_event so it's ok if it | |
249 | * changes under us. | |
250 | */ | |
1e2c0436 | 251 | if (huge_pte_none(pte)) |
369cd212 | 252 | ret = true; |
1e2c0436 | 253 | if (!huge_pte_write(pte) && (reason & VM_UFFD_WP)) |
369cd212 MK |
254 | ret = true; |
255 | out: | |
256 | return ret; | |
257 | } | |
258 | #else | |
259 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 260 | struct vm_area_struct *vma, |
369cd212 MK |
261 | unsigned long address, |
262 | unsigned long flags, | |
263 | unsigned long reason) | |
264 | { | |
265 | return false; /* should never get here */ | |
266 | } | |
267 | #endif /* CONFIG_HUGETLB_PAGE */ | |
268 | ||
8d2afd96 AA |
269 | /* |
270 | * Verify the pagetables are still not ok after having reigstered into | |
271 | * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any | |
272 | * userfault that has already been resolved, if userfaultfd_read and | |
273 | * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different | |
274 | * threads. | |
275 | */ | |
276 | static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx, | |
277 | unsigned long address, | |
278 | unsigned long flags, | |
279 | unsigned long reason) | |
280 | { | |
281 | struct mm_struct *mm = ctx->mm; | |
282 | pgd_t *pgd; | |
c2febafc | 283 | p4d_t *p4d; |
8d2afd96 AA |
284 | pud_t *pud; |
285 | pmd_t *pmd, _pmd; | |
286 | pte_t *pte; | |
287 | bool ret = true; | |
288 | ||
289 | VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem)); | |
290 | ||
291 | pgd = pgd_offset(mm, address); | |
292 | if (!pgd_present(*pgd)) | |
293 | goto out; | |
c2febafc KS |
294 | p4d = p4d_offset(pgd, address); |
295 | if (!p4d_present(*p4d)) | |
296 | goto out; | |
297 | pud = pud_offset(p4d, address); | |
8d2afd96 AA |
298 | if (!pud_present(*pud)) |
299 | goto out; | |
300 | pmd = pmd_offset(pud, address); | |
301 | /* | |
302 | * READ_ONCE must function as a barrier with narrower scope | |
303 | * and it must be equivalent to: | |
304 | * _pmd = *pmd; barrier(); | |
305 | * | |
306 | * This is to deal with the instability (as in | |
307 | * pmd_trans_unstable) of the pmd. | |
308 | */ | |
309 | _pmd = READ_ONCE(*pmd); | |
a365ac09 | 310 | if (pmd_none(_pmd)) |
8d2afd96 AA |
311 | goto out; |
312 | ||
313 | ret = false; | |
a365ac09 HY |
314 | if (!pmd_present(_pmd)) |
315 | goto out; | |
316 | ||
8d2afd96 AA |
317 | if (pmd_trans_huge(_pmd)) |
318 | goto out; | |
319 | ||
320 | /* | |
321 | * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it | |
322 | * and use the standard pte_offset_map() instead of parsing _pmd. | |
323 | */ | |
324 | pte = pte_offset_map(pmd, address); | |
325 | /* | |
326 | * Lockless access: we're in a wait_event so it's ok if it | |
327 | * changes under us. | |
328 | */ | |
329 | if (pte_none(*pte)) | |
330 | ret = true; | |
331 | pte_unmap(pte); | |
332 | ||
333 | out: | |
334 | return ret; | |
335 | } | |
336 | ||
86039bd3 AA |
337 | /* |
338 | * The locking rules involved in returning VM_FAULT_RETRY depending on | |
339 | * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and | |
340 | * FAULT_FLAG_KILLABLE are not straightforward. The "Caution" | |
341 | * recommendation in __lock_page_or_retry is not an understatement. | |
342 | * | |
343 | * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released | |
344 | * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is | |
345 | * not set. | |
346 | * | |
347 | * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not | |
348 | * set, VM_FAULT_RETRY can still be returned if and only if there are | |
349 | * fatal_signal_pending()s, and the mmap_sem must be released before | |
350 | * returning it. | |
351 | */ | |
2b740303 | 352 | vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason) |
86039bd3 | 353 | { |
82b0f8c3 | 354 | struct mm_struct *mm = vmf->vma->vm_mm; |
86039bd3 AA |
355 | struct userfaultfd_ctx *ctx; |
356 | struct userfaultfd_wait_queue uwq; | |
2b740303 | 357 | vm_fault_t ret = VM_FAULT_SIGBUS; |
dfa37dc3 | 358 | bool must_wait, return_to_userland; |
15a77c6f | 359 | long blocking_state; |
86039bd3 | 360 | |
64c2b203 AA |
361 | /* |
362 | * We don't do userfault handling for the final child pid update. | |
363 | * | |
364 | * We also don't do userfault handling during | |
365 | * coredumping. hugetlbfs has the special | |
366 | * follow_hugetlb_page() to skip missing pages in the | |
367 | * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with | |
368 | * the no_page_table() helper in follow_page_mask(), but the | |
369 | * shmem_vm_ops->fault method is invoked even during | |
370 | * coredumping without mmap_sem and it ends up here. | |
371 | */ | |
372 | if (current->flags & (PF_EXITING|PF_DUMPCORE)) | |
373 | goto out; | |
374 | ||
375 | /* | |
376 | * Coredumping runs without mmap_sem so we can only check that | |
377 | * the mmap_sem is held, if PF_DUMPCORE was not set. | |
378 | */ | |
379 | WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem)); | |
380 | ||
82b0f8c3 | 381 | ctx = vmf->vma->vm_userfaultfd_ctx.ctx; |
86039bd3 | 382 | if (!ctx) |
ba85c702 | 383 | goto out; |
86039bd3 AA |
384 | |
385 | BUG_ON(ctx->mm != mm); | |
386 | ||
387 | VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP)); | |
388 | VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP)); | |
389 | ||
2d6d6f5a PS |
390 | if (ctx->features & UFFD_FEATURE_SIGBUS) |
391 | goto out; | |
392 | ||
86039bd3 AA |
393 | /* |
394 | * If it's already released don't get it. This avoids to loop | |
395 | * in __get_user_pages if userfaultfd_release waits on the | |
396 | * caller of handle_userfault to release the mmap_sem. | |
397 | */ | |
6aa7de05 | 398 | if (unlikely(READ_ONCE(ctx->released))) { |
656710a6 AA |
399 | /* |
400 | * Don't return VM_FAULT_SIGBUS in this case, so a non | |
401 | * cooperative manager can close the uffd after the | |
402 | * last UFFDIO_COPY, without risking to trigger an | |
403 | * involuntary SIGBUS if the process was starting the | |
404 | * userfaultfd while the userfaultfd was still armed | |
405 | * (but after the last UFFDIO_COPY). If the uffd | |
406 | * wasn't already closed when the userfault reached | |
407 | * this point, that would normally be solved by | |
408 | * userfaultfd_must_wait returning 'false'. | |
409 | * | |
410 | * If we were to return VM_FAULT_SIGBUS here, the non | |
411 | * cooperative manager would be instead forced to | |
412 | * always call UFFDIO_UNREGISTER before it can safely | |
413 | * close the uffd. | |
414 | */ | |
415 | ret = VM_FAULT_NOPAGE; | |
ba85c702 | 416 | goto out; |
656710a6 | 417 | } |
86039bd3 AA |
418 | |
419 | /* | |
420 | * Check that we can return VM_FAULT_RETRY. | |
421 | * | |
422 | * NOTE: it should become possible to return VM_FAULT_RETRY | |
423 | * even if FAULT_FLAG_TRIED is set without leading to gup() | |
424 | * -EBUSY failures, if the userfaultfd is to be extended for | |
425 | * VM_UFFD_WP tracking and we intend to arm the userfault | |
426 | * without first stopping userland access to the memory. For | |
427 | * VM_UFFD_MISSING userfaults this is enough for now. | |
428 | */ | |
82b0f8c3 | 429 | if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) { |
86039bd3 AA |
430 | /* |
431 | * Validate the invariant that nowait must allow retry | |
432 | * to be sure not to return SIGBUS erroneously on | |
433 | * nowait invocations. | |
434 | */ | |
82b0f8c3 | 435 | BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT); |
86039bd3 AA |
436 | #ifdef CONFIG_DEBUG_VM |
437 | if (printk_ratelimit()) { | |
438 | printk(KERN_WARNING | |
82b0f8c3 JK |
439 | "FAULT_FLAG_ALLOW_RETRY missing %x\n", |
440 | vmf->flags); | |
86039bd3 AA |
441 | dump_stack(); |
442 | } | |
443 | #endif | |
ba85c702 | 444 | goto out; |
86039bd3 AA |
445 | } |
446 | ||
447 | /* | |
448 | * Handle nowait, not much to do other than tell it to retry | |
449 | * and wait. | |
450 | */ | |
ba85c702 | 451 | ret = VM_FAULT_RETRY; |
82b0f8c3 | 452 | if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) |
ba85c702 | 453 | goto out; |
86039bd3 AA |
454 | |
455 | /* take the reference before dropping the mmap_sem */ | |
456 | userfaultfd_ctx_get(ctx); | |
457 | ||
86039bd3 AA |
458 | init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function); |
459 | uwq.wq.private = current; | |
9d4ac934 AP |
460 | uwq.msg = userfault_msg(vmf->address, vmf->flags, reason, |
461 | ctx->features); | |
86039bd3 | 462 | uwq.ctx = ctx; |
15a77c6f | 463 | uwq.waken = false; |
86039bd3 | 464 | |
bae473a4 | 465 | return_to_userland = |
82b0f8c3 | 466 | (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) == |
dfa37dc3 | 467 | (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE); |
15a77c6f AA |
468 | blocking_state = return_to_userland ? TASK_INTERRUPTIBLE : |
469 | TASK_KILLABLE; | |
dfa37dc3 | 470 | |
cbcfa130 | 471 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
472 | /* |
473 | * After the __add_wait_queue the uwq is visible to userland | |
474 | * through poll/read(). | |
475 | */ | |
15b726ef AA |
476 | __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq); |
477 | /* | |
478 | * The smp_mb() after __set_current_state prevents the reads | |
479 | * following the spin_unlock to happen before the list_add in | |
480 | * __add_wait_queue. | |
481 | */ | |
15a77c6f | 482 | set_current_state(blocking_state); |
cbcfa130 | 483 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 484 | |
369cd212 MK |
485 | if (!is_vm_hugetlb_page(vmf->vma)) |
486 | must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags, | |
487 | reason); | |
488 | else | |
7868a208 PA |
489 | must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma, |
490 | vmf->address, | |
369cd212 | 491 | vmf->flags, reason); |
8d2afd96 AA |
492 | up_read(&mm->mmap_sem); |
493 | ||
6aa7de05 | 494 | if (likely(must_wait && !READ_ONCE(ctx->released) && |
dfa37dc3 AA |
495 | (return_to_userland ? !signal_pending(current) : |
496 | !fatal_signal_pending(current)))) { | |
a9a08845 | 497 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
86039bd3 | 498 | schedule(); |
ba85c702 | 499 | ret |= VM_FAULT_MAJOR; |
15a77c6f AA |
500 | |
501 | /* | |
502 | * False wakeups can orginate even from rwsem before | |
503 | * up_read() however userfaults will wait either for a | |
504 | * targeted wakeup on the specific uwq waitqueue from | |
505 | * wake_userfault() or for signals or for uffd | |
506 | * release. | |
507 | */ | |
508 | while (!READ_ONCE(uwq.waken)) { | |
509 | /* | |
510 | * This needs the full smp_store_mb() | |
511 | * guarantee as the state write must be | |
512 | * visible to other CPUs before reading | |
513 | * uwq.waken from other CPUs. | |
514 | */ | |
515 | set_current_state(blocking_state); | |
516 | if (READ_ONCE(uwq.waken) || | |
517 | READ_ONCE(ctx->released) || | |
518 | (return_to_userland ? signal_pending(current) : | |
519 | fatal_signal_pending(current))) | |
520 | break; | |
521 | schedule(); | |
522 | } | |
ba85c702 | 523 | } |
86039bd3 | 524 | |
ba85c702 | 525 | __set_current_state(TASK_RUNNING); |
15b726ef | 526 | |
dfa37dc3 AA |
527 | if (return_to_userland) { |
528 | if (signal_pending(current) && | |
529 | !fatal_signal_pending(current)) { | |
530 | /* | |
531 | * If we got a SIGSTOP or SIGCONT and this is | |
532 | * a normal userland page fault, just let | |
533 | * userland return so the signal will be | |
534 | * handled and gdb debugging works. The page | |
535 | * fault code immediately after we return from | |
536 | * this function is going to release the | |
537 | * mmap_sem and it's not depending on it | |
538 | * (unlike gup would if we were not to return | |
539 | * VM_FAULT_RETRY). | |
540 | * | |
541 | * If a fatal signal is pending we still take | |
542 | * the streamlined VM_FAULT_RETRY failure path | |
543 | * and there's no need to retake the mmap_sem | |
544 | * in such case. | |
545 | */ | |
546 | down_read(&mm->mmap_sem); | |
6bbc4a41 | 547 | ret = VM_FAULT_NOPAGE; |
dfa37dc3 AA |
548 | } |
549 | } | |
550 | ||
15b726ef AA |
551 | /* |
552 | * Here we race with the list_del; list_add in | |
553 | * userfaultfd_ctx_read(), however because we don't ever run | |
554 | * list_del_init() to refile across the two lists, the prev | |
555 | * and next pointers will never point to self. list_add also | |
556 | * would never let any of the two pointers to point to | |
557 | * self. So list_empty_careful won't risk to see both pointers | |
558 | * pointing to self at any time during the list refile. The | |
559 | * only case where list_del_init() is called is the full | |
560 | * removal in the wake function and there we don't re-list_add | |
561 | * and it's fine not to block on the spinlock. The uwq on this | |
562 | * kernel stack can be released after the list_del_init. | |
563 | */ | |
2055da97 | 564 | if (!list_empty_careful(&uwq.wq.entry)) { |
cbcfa130 | 565 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
15b726ef AA |
566 | /* |
567 | * No need of list_del_init(), the uwq on the stack | |
568 | * will be freed shortly anyway. | |
569 | */ | |
2055da97 | 570 | list_del(&uwq.wq.entry); |
cbcfa130 | 571 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 572 | } |
86039bd3 AA |
573 | |
574 | /* | |
575 | * ctx may go away after this if the userfault pseudo fd is | |
576 | * already released. | |
577 | */ | |
578 | userfaultfd_ctx_put(ctx); | |
579 | ||
ba85c702 AA |
580 | out: |
581 | return ret; | |
86039bd3 AA |
582 | } |
583 | ||
8c9e7bb7 AA |
584 | static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx, |
585 | struct userfaultfd_wait_queue *ewq) | |
9cd75c3c | 586 | { |
0cbb4b4f AA |
587 | struct userfaultfd_ctx *release_new_ctx; |
588 | ||
9a69a829 AA |
589 | if (WARN_ON_ONCE(current->flags & PF_EXITING)) |
590 | goto out; | |
9cd75c3c PE |
591 | |
592 | ewq->ctx = ctx; | |
593 | init_waitqueue_entry(&ewq->wq, current); | |
0cbb4b4f | 594 | release_new_ctx = NULL; |
9cd75c3c | 595 | |
cbcfa130 | 596 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
597 | /* |
598 | * After the __add_wait_queue the uwq is visible to userland | |
599 | * through poll/read(). | |
600 | */ | |
601 | __add_wait_queue(&ctx->event_wqh, &ewq->wq); | |
602 | for (;;) { | |
603 | set_current_state(TASK_KILLABLE); | |
604 | if (ewq->msg.event == 0) | |
605 | break; | |
6aa7de05 | 606 | if (READ_ONCE(ctx->released) || |
9cd75c3c | 607 | fatal_signal_pending(current)) { |
384632e6 AA |
608 | /* |
609 | * &ewq->wq may be queued in fork_event, but | |
610 | * __remove_wait_queue ignores the head | |
611 | * parameter. It would be a problem if it | |
612 | * didn't. | |
613 | */ | |
9cd75c3c | 614 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); |
7eb76d45 MR |
615 | if (ewq->msg.event == UFFD_EVENT_FORK) { |
616 | struct userfaultfd_ctx *new; | |
617 | ||
618 | new = (struct userfaultfd_ctx *) | |
619 | (unsigned long) | |
620 | ewq->msg.arg.reserved.reserved1; | |
0cbb4b4f | 621 | release_new_ctx = new; |
7eb76d45 | 622 | } |
9cd75c3c PE |
623 | break; |
624 | } | |
625 | ||
cbcfa130 | 626 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 627 | |
a9a08845 | 628 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
9cd75c3c PE |
629 | schedule(); |
630 | ||
cbcfa130 | 631 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
632 | } |
633 | __set_current_state(TASK_RUNNING); | |
cbcfa130 | 634 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 635 | |
0cbb4b4f AA |
636 | if (release_new_ctx) { |
637 | struct vm_area_struct *vma; | |
638 | struct mm_struct *mm = release_new_ctx->mm; | |
639 | ||
640 | /* the various vma->vm_userfaultfd_ctx still points to it */ | |
641 | down_write(&mm->mmap_sem); | |
04f5866e AA |
642 | /* no task can run (and in turn coredump) yet */ |
643 | VM_WARN_ON(!mmget_still_valid(mm)); | |
0cbb4b4f | 644 | for (vma = mm->mmap; vma; vma = vma->vm_next) |
31e810aa | 645 | if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) { |
0cbb4b4f | 646 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
31e810aa MR |
647 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); |
648 | } | |
0cbb4b4f AA |
649 | up_write(&mm->mmap_sem); |
650 | ||
651 | userfaultfd_ctx_put(release_new_ctx); | |
652 | } | |
653 | ||
9cd75c3c PE |
654 | /* |
655 | * ctx may go away after this if the userfault pseudo fd is | |
656 | * already released. | |
657 | */ | |
9a69a829 | 658 | out: |
df2cc96e | 659 | WRITE_ONCE(ctx->mmap_changing, false); |
9cd75c3c | 660 | userfaultfd_ctx_put(ctx); |
9cd75c3c PE |
661 | } |
662 | ||
663 | static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx, | |
664 | struct userfaultfd_wait_queue *ewq) | |
665 | { | |
666 | ewq->msg.event = 0; | |
667 | wake_up_locked(&ctx->event_wqh); | |
668 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); | |
669 | } | |
670 | ||
893e26e6 PE |
671 | int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs) |
672 | { | |
673 | struct userfaultfd_ctx *ctx = NULL, *octx; | |
674 | struct userfaultfd_fork_ctx *fctx; | |
675 | ||
676 | octx = vma->vm_userfaultfd_ctx.ctx; | |
677 | if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) { | |
678 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
679 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); | |
680 | return 0; | |
681 | } | |
682 | ||
683 | list_for_each_entry(fctx, fcs, list) | |
684 | if (fctx->orig == octx) { | |
685 | ctx = fctx->new; | |
686 | break; | |
687 | } | |
688 | ||
689 | if (!ctx) { | |
690 | fctx = kmalloc(sizeof(*fctx), GFP_KERNEL); | |
691 | if (!fctx) | |
692 | return -ENOMEM; | |
693 | ||
694 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); | |
695 | if (!ctx) { | |
696 | kfree(fctx); | |
697 | return -ENOMEM; | |
698 | } | |
699 | ||
ca880420 | 700 | refcount_set(&ctx->refcount, 1); |
893e26e6 PE |
701 | ctx->flags = octx->flags; |
702 | ctx->state = UFFD_STATE_RUNNING; | |
703 | ctx->features = octx->features; | |
704 | ctx->released = false; | |
df2cc96e | 705 | ctx->mmap_changing = false; |
893e26e6 | 706 | ctx->mm = vma->vm_mm; |
00bb31fa | 707 | mmgrab(ctx->mm); |
893e26e6 PE |
708 | |
709 | userfaultfd_ctx_get(octx); | |
df2cc96e | 710 | WRITE_ONCE(octx->mmap_changing, true); |
893e26e6 PE |
711 | fctx->orig = octx; |
712 | fctx->new = ctx; | |
713 | list_add_tail(&fctx->list, fcs); | |
714 | } | |
715 | ||
716 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
717 | return 0; | |
718 | } | |
719 | ||
8c9e7bb7 | 720 | static void dup_fctx(struct userfaultfd_fork_ctx *fctx) |
893e26e6 PE |
721 | { |
722 | struct userfaultfd_ctx *ctx = fctx->orig; | |
723 | struct userfaultfd_wait_queue ewq; | |
724 | ||
725 | msg_init(&ewq.msg); | |
726 | ||
727 | ewq.msg.event = UFFD_EVENT_FORK; | |
728 | ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new; | |
729 | ||
8c9e7bb7 | 730 | userfaultfd_event_wait_completion(ctx, &ewq); |
893e26e6 PE |
731 | } |
732 | ||
733 | void dup_userfaultfd_complete(struct list_head *fcs) | |
734 | { | |
893e26e6 PE |
735 | struct userfaultfd_fork_ctx *fctx, *n; |
736 | ||
737 | list_for_each_entry_safe(fctx, n, fcs, list) { | |
8c9e7bb7 | 738 | dup_fctx(fctx); |
893e26e6 PE |
739 | list_del(&fctx->list); |
740 | kfree(fctx); | |
741 | } | |
742 | } | |
743 | ||
72f87654 PE |
744 | void mremap_userfaultfd_prep(struct vm_area_struct *vma, |
745 | struct vm_userfaultfd_ctx *vm_ctx) | |
746 | { | |
747 | struct userfaultfd_ctx *ctx; | |
748 | ||
749 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
3cfd22be PX |
750 | |
751 | if (!ctx) | |
752 | return; | |
753 | ||
754 | if (ctx->features & UFFD_FEATURE_EVENT_REMAP) { | |
72f87654 PE |
755 | vm_ctx->ctx = ctx; |
756 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 757 | WRITE_ONCE(ctx->mmap_changing, true); |
3cfd22be PX |
758 | } else { |
759 | /* Drop uffd context if remap feature not enabled */ | |
760 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
761 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); | |
72f87654 PE |
762 | } |
763 | } | |
764 | ||
90794bf1 | 765 | void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx, |
72f87654 PE |
766 | unsigned long from, unsigned long to, |
767 | unsigned long len) | |
768 | { | |
90794bf1 | 769 | struct userfaultfd_ctx *ctx = vm_ctx->ctx; |
72f87654 PE |
770 | struct userfaultfd_wait_queue ewq; |
771 | ||
772 | if (!ctx) | |
773 | return; | |
774 | ||
775 | if (to & ~PAGE_MASK) { | |
776 | userfaultfd_ctx_put(ctx); | |
777 | return; | |
778 | } | |
779 | ||
780 | msg_init(&ewq.msg); | |
781 | ||
782 | ewq.msg.event = UFFD_EVENT_REMAP; | |
783 | ewq.msg.arg.remap.from = from; | |
784 | ewq.msg.arg.remap.to = to; | |
785 | ewq.msg.arg.remap.len = len; | |
786 | ||
787 | userfaultfd_event_wait_completion(ctx, &ewq); | |
788 | } | |
789 | ||
70ccb92f | 790 | bool userfaultfd_remove(struct vm_area_struct *vma, |
d811914d | 791 | unsigned long start, unsigned long end) |
05ce7724 PE |
792 | { |
793 | struct mm_struct *mm = vma->vm_mm; | |
794 | struct userfaultfd_ctx *ctx; | |
795 | struct userfaultfd_wait_queue ewq; | |
796 | ||
797 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
d811914d | 798 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE)) |
70ccb92f | 799 | return true; |
05ce7724 PE |
800 | |
801 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 802 | WRITE_ONCE(ctx->mmap_changing, true); |
05ce7724 PE |
803 | up_read(&mm->mmap_sem); |
804 | ||
05ce7724 PE |
805 | msg_init(&ewq.msg); |
806 | ||
d811914d MR |
807 | ewq.msg.event = UFFD_EVENT_REMOVE; |
808 | ewq.msg.arg.remove.start = start; | |
809 | ewq.msg.arg.remove.end = end; | |
05ce7724 PE |
810 | |
811 | userfaultfd_event_wait_completion(ctx, &ewq); | |
812 | ||
70ccb92f | 813 | return false; |
05ce7724 PE |
814 | } |
815 | ||
897ab3e0 MR |
816 | static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps, |
817 | unsigned long start, unsigned long end) | |
818 | { | |
819 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
820 | ||
821 | list_for_each_entry(unmap_ctx, unmaps, list) | |
822 | if (unmap_ctx->ctx == ctx && unmap_ctx->start == start && | |
823 | unmap_ctx->end == end) | |
824 | return true; | |
825 | ||
826 | return false; | |
827 | } | |
828 | ||
829 | int userfaultfd_unmap_prep(struct vm_area_struct *vma, | |
830 | unsigned long start, unsigned long end, | |
831 | struct list_head *unmaps) | |
832 | { | |
833 | for ( ; vma && vma->vm_start < end; vma = vma->vm_next) { | |
834 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
835 | struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx; | |
836 | ||
837 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) || | |
838 | has_unmap_ctx(ctx, unmaps, start, end)) | |
839 | continue; | |
840 | ||
841 | unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL); | |
842 | if (!unmap_ctx) | |
843 | return -ENOMEM; | |
844 | ||
845 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 846 | WRITE_ONCE(ctx->mmap_changing, true); |
897ab3e0 MR |
847 | unmap_ctx->ctx = ctx; |
848 | unmap_ctx->start = start; | |
849 | unmap_ctx->end = end; | |
850 | list_add_tail(&unmap_ctx->list, unmaps); | |
851 | } | |
852 | ||
853 | return 0; | |
854 | } | |
855 | ||
856 | void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf) | |
857 | { | |
858 | struct userfaultfd_unmap_ctx *ctx, *n; | |
859 | struct userfaultfd_wait_queue ewq; | |
860 | ||
861 | list_for_each_entry_safe(ctx, n, uf, list) { | |
862 | msg_init(&ewq.msg); | |
863 | ||
864 | ewq.msg.event = UFFD_EVENT_UNMAP; | |
865 | ewq.msg.arg.remove.start = ctx->start; | |
866 | ewq.msg.arg.remove.end = ctx->end; | |
867 | ||
868 | userfaultfd_event_wait_completion(ctx->ctx, &ewq); | |
869 | ||
870 | list_del(&ctx->list); | |
871 | kfree(ctx); | |
872 | } | |
873 | } | |
874 | ||
86039bd3 AA |
875 | static int userfaultfd_release(struct inode *inode, struct file *file) |
876 | { | |
877 | struct userfaultfd_ctx *ctx = file->private_data; | |
878 | struct mm_struct *mm = ctx->mm; | |
879 | struct vm_area_struct *vma, *prev; | |
880 | /* len == 0 means wake all */ | |
881 | struct userfaultfd_wake_range range = { .len = 0, }; | |
882 | unsigned long new_flags; | |
46d0b24c | 883 | bool still_valid; |
86039bd3 | 884 | |
6aa7de05 | 885 | WRITE_ONCE(ctx->released, true); |
86039bd3 | 886 | |
d2005e3f ON |
887 | if (!mmget_not_zero(mm)) |
888 | goto wakeup; | |
889 | ||
86039bd3 AA |
890 | /* |
891 | * Flush page faults out of all CPUs. NOTE: all page faults | |
892 | * must be retried without returning VM_FAULT_SIGBUS if | |
893 | * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx | |
894 | * changes while handle_userfault released the mmap_sem. So | |
895 | * it's critical that released is set to true (above), before | |
896 | * taking the mmap_sem for writing. | |
897 | */ | |
898 | down_write(&mm->mmap_sem); | |
46d0b24c | 899 | still_valid = mmget_still_valid(mm); |
86039bd3 AA |
900 | prev = NULL; |
901 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
902 | cond_resched(); | |
903 | BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^ | |
904 | !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
905 | if (vma->vm_userfaultfd_ctx.ctx != ctx) { | |
906 | prev = vma; | |
907 | continue; | |
908 | } | |
909 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); | |
46d0b24c ON |
910 | if (still_valid) { |
911 | prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end, | |
912 | new_flags, vma->anon_vma, | |
913 | vma->vm_file, vma->vm_pgoff, | |
914 | vma_policy(vma), | |
915 | NULL_VM_UFFD_CTX); | |
916 | if (prev) | |
917 | vma = prev; | |
918 | else | |
919 | prev = vma; | |
920 | } | |
86039bd3 AA |
921 | vma->vm_flags = new_flags; |
922 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
923 | } | |
924 | up_write(&mm->mmap_sem); | |
d2005e3f ON |
925 | mmput(mm); |
926 | wakeup: | |
86039bd3 | 927 | /* |
15b726ef | 928 | * After no new page faults can wait on this fault_*wqh, flush |
86039bd3 | 929 | * the last page faults that may have been already waiting on |
15b726ef | 930 | * the fault_*wqh. |
86039bd3 | 931 | */ |
cbcfa130 | 932 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
ac5be6b4 | 933 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range); |
c430d1e8 | 934 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range); |
cbcfa130 | 935 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 936 | |
5a18b64e MR |
937 | /* Flush pending events that may still wait on event_wqh */ |
938 | wake_up_all(&ctx->event_wqh); | |
939 | ||
a9a08845 | 940 | wake_up_poll(&ctx->fd_wqh, EPOLLHUP); |
86039bd3 AA |
941 | userfaultfd_ctx_put(ctx); |
942 | return 0; | |
943 | } | |
944 | ||
15b726ef | 945 | /* fault_pending_wqh.lock must be hold by the caller */ |
6dcc27fd PE |
946 | static inline struct userfaultfd_wait_queue *find_userfault_in( |
947 | wait_queue_head_t *wqh) | |
86039bd3 | 948 | { |
ac6424b9 | 949 | wait_queue_entry_t *wq; |
15b726ef | 950 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 951 | |
456a7378 | 952 | lockdep_assert_held(&wqh->lock); |
86039bd3 | 953 | |
15b726ef | 954 | uwq = NULL; |
6dcc27fd | 955 | if (!waitqueue_active(wqh)) |
15b726ef AA |
956 | goto out; |
957 | /* walk in reverse to provide FIFO behavior to read userfaults */ | |
2055da97 | 958 | wq = list_last_entry(&wqh->head, typeof(*wq), entry); |
15b726ef AA |
959 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
960 | out: | |
961 | return uwq; | |
86039bd3 | 962 | } |
6dcc27fd PE |
963 | |
964 | static inline struct userfaultfd_wait_queue *find_userfault( | |
965 | struct userfaultfd_ctx *ctx) | |
966 | { | |
967 | return find_userfault_in(&ctx->fault_pending_wqh); | |
968 | } | |
86039bd3 | 969 | |
9cd75c3c PE |
970 | static inline struct userfaultfd_wait_queue *find_userfault_evt( |
971 | struct userfaultfd_ctx *ctx) | |
972 | { | |
973 | return find_userfault_in(&ctx->event_wqh); | |
974 | } | |
975 | ||
076ccb76 | 976 | static __poll_t userfaultfd_poll(struct file *file, poll_table *wait) |
86039bd3 AA |
977 | { |
978 | struct userfaultfd_ctx *ctx = file->private_data; | |
076ccb76 | 979 | __poll_t ret; |
86039bd3 AA |
980 | |
981 | poll_wait(file, &ctx->fd_wqh, wait); | |
982 | ||
983 | switch (ctx->state) { | |
984 | case UFFD_STATE_WAIT_API: | |
a9a08845 | 985 | return EPOLLERR; |
86039bd3 | 986 | case UFFD_STATE_RUNNING: |
ba85c702 AA |
987 | /* |
988 | * poll() never guarantees that read won't block. | |
989 | * userfaults can be waken before they're read(). | |
990 | */ | |
991 | if (unlikely(!(file->f_flags & O_NONBLOCK))) | |
a9a08845 | 992 | return EPOLLERR; |
15b726ef AA |
993 | /* |
994 | * lockless access to see if there are pending faults | |
995 | * __pollwait last action is the add_wait_queue but | |
996 | * the spin_unlock would allow the waitqueue_active to | |
997 | * pass above the actual list_add inside | |
998 | * add_wait_queue critical section. So use a full | |
999 | * memory barrier to serialize the list_add write of | |
1000 | * add_wait_queue() with the waitqueue_active read | |
1001 | * below. | |
1002 | */ | |
1003 | ret = 0; | |
1004 | smp_mb(); | |
1005 | if (waitqueue_active(&ctx->fault_pending_wqh)) | |
a9a08845 | 1006 | ret = EPOLLIN; |
9cd75c3c | 1007 | else if (waitqueue_active(&ctx->event_wqh)) |
a9a08845 | 1008 | ret = EPOLLIN; |
9cd75c3c | 1009 | |
86039bd3 AA |
1010 | return ret; |
1011 | default: | |
8474901a | 1012 | WARN_ON_ONCE(1); |
a9a08845 | 1013 | return EPOLLERR; |
86039bd3 AA |
1014 | } |
1015 | } | |
1016 | ||
893e26e6 PE |
1017 | static const struct file_operations userfaultfd_fops; |
1018 | ||
1019 | static int resolve_userfault_fork(struct userfaultfd_ctx *ctx, | |
1020 | struct userfaultfd_ctx *new, | |
1021 | struct uffd_msg *msg) | |
1022 | { | |
1023 | int fd; | |
893e26e6 | 1024 | |
284cd241 EB |
1025 | fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, new, |
1026 | O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS)); | |
893e26e6 PE |
1027 | if (fd < 0) |
1028 | return fd; | |
1029 | ||
893e26e6 PE |
1030 | msg->arg.reserved.reserved1 = 0; |
1031 | msg->arg.fork.ufd = fd; | |
893e26e6 PE |
1032 | return 0; |
1033 | } | |
1034 | ||
86039bd3 | 1035 | static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait, |
a9b85f94 | 1036 | struct uffd_msg *msg) |
86039bd3 AA |
1037 | { |
1038 | ssize_t ret; | |
1039 | DECLARE_WAITQUEUE(wait, current); | |
15b726ef | 1040 | struct userfaultfd_wait_queue *uwq; |
893e26e6 PE |
1041 | /* |
1042 | * Handling fork event requires sleeping operations, so | |
1043 | * we drop the event_wqh lock, then do these ops, then | |
1044 | * lock it back and wake up the waiter. While the lock is | |
1045 | * dropped the ewq may go away so we keep track of it | |
1046 | * carefully. | |
1047 | */ | |
1048 | LIST_HEAD(fork_event); | |
1049 | struct userfaultfd_ctx *fork_nctx = NULL; | |
86039bd3 | 1050 | |
15b726ef | 1051 | /* always take the fd_wqh lock before the fault_pending_wqh lock */ |
ae62c16e | 1052 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1053 | __add_wait_queue(&ctx->fd_wqh, &wait); |
1054 | for (;;) { | |
1055 | set_current_state(TASK_INTERRUPTIBLE); | |
15b726ef AA |
1056 | spin_lock(&ctx->fault_pending_wqh.lock); |
1057 | uwq = find_userfault(ctx); | |
1058 | if (uwq) { | |
2c5b7e1b AA |
1059 | /* |
1060 | * Use a seqcount to repeat the lockless check | |
1061 | * in wake_userfault() to avoid missing | |
1062 | * wakeups because during the refile both | |
1063 | * waitqueue could become empty if this is the | |
1064 | * only userfault. | |
1065 | */ | |
1066 | write_seqcount_begin(&ctx->refile_seq); | |
1067 | ||
86039bd3 | 1068 | /* |
15b726ef AA |
1069 | * The fault_pending_wqh.lock prevents the uwq |
1070 | * to disappear from under us. | |
1071 | * | |
1072 | * Refile this userfault from | |
1073 | * fault_pending_wqh to fault_wqh, it's not | |
1074 | * pending anymore after we read it. | |
1075 | * | |
1076 | * Use list_del() by hand (as | |
1077 | * userfaultfd_wake_function also uses | |
1078 | * list_del_init() by hand) to be sure nobody | |
1079 | * changes __remove_wait_queue() to use | |
1080 | * list_del_init() in turn breaking the | |
1081 | * !list_empty_careful() check in | |
2055da97 | 1082 | * handle_userfault(). The uwq->wq.head list |
15b726ef AA |
1083 | * must never be empty at any time during the |
1084 | * refile, or the waitqueue could disappear | |
1085 | * from under us. The "wait_queue_head_t" | |
1086 | * parameter of __remove_wait_queue() is unused | |
1087 | * anyway. | |
86039bd3 | 1088 | */ |
2055da97 | 1089 | list_del(&uwq->wq.entry); |
c430d1e8 | 1090 | add_wait_queue(&ctx->fault_wqh, &uwq->wq); |
15b726ef | 1091 | |
2c5b7e1b AA |
1092 | write_seqcount_end(&ctx->refile_seq); |
1093 | ||
a9b85f94 AA |
1094 | /* careful to always initialize msg if ret == 0 */ |
1095 | *msg = uwq->msg; | |
15b726ef | 1096 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1097 | ret = 0; |
1098 | break; | |
1099 | } | |
15b726ef | 1100 | spin_unlock(&ctx->fault_pending_wqh.lock); |
9cd75c3c PE |
1101 | |
1102 | spin_lock(&ctx->event_wqh.lock); | |
1103 | uwq = find_userfault_evt(ctx); | |
1104 | if (uwq) { | |
1105 | *msg = uwq->msg; | |
1106 | ||
893e26e6 PE |
1107 | if (uwq->msg.event == UFFD_EVENT_FORK) { |
1108 | fork_nctx = (struct userfaultfd_ctx *) | |
1109 | (unsigned long) | |
1110 | uwq->msg.arg.reserved.reserved1; | |
2055da97 | 1111 | list_move(&uwq->wq.entry, &fork_event); |
384632e6 AA |
1112 | /* |
1113 | * fork_nctx can be freed as soon as | |
1114 | * we drop the lock, unless we take a | |
1115 | * reference on it. | |
1116 | */ | |
1117 | userfaultfd_ctx_get(fork_nctx); | |
893e26e6 PE |
1118 | spin_unlock(&ctx->event_wqh.lock); |
1119 | ret = 0; | |
1120 | break; | |
1121 | } | |
1122 | ||
9cd75c3c PE |
1123 | userfaultfd_event_complete(ctx, uwq); |
1124 | spin_unlock(&ctx->event_wqh.lock); | |
1125 | ret = 0; | |
1126 | break; | |
1127 | } | |
1128 | spin_unlock(&ctx->event_wqh.lock); | |
1129 | ||
86039bd3 AA |
1130 | if (signal_pending(current)) { |
1131 | ret = -ERESTARTSYS; | |
1132 | break; | |
1133 | } | |
1134 | if (no_wait) { | |
1135 | ret = -EAGAIN; | |
1136 | break; | |
1137 | } | |
ae62c16e | 1138 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1139 | schedule(); |
ae62c16e | 1140 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1141 | } |
1142 | __remove_wait_queue(&ctx->fd_wqh, &wait); | |
1143 | __set_current_state(TASK_RUNNING); | |
ae62c16e | 1144 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1145 | |
893e26e6 PE |
1146 | if (!ret && msg->event == UFFD_EVENT_FORK) { |
1147 | ret = resolve_userfault_fork(ctx, fork_nctx, msg); | |
cbcfa130 | 1148 | spin_lock_irq(&ctx->event_wqh.lock); |
384632e6 AA |
1149 | if (!list_empty(&fork_event)) { |
1150 | /* | |
1151 | * The fork thread didn't abort, so we can | |
1152 | * drop the temporary refcount. | |
1153 | */ | |
1154 | userfaultfd_ctx_put(fork_nctx); | |
1155 | ||
1156 | uwq = list_first_entry(&fork_event, | |
1157 | typeof(*uwq), | |
1158 | wq.entry); | |
1159 | /* | |
1160 | * If fork_event list wasn't empty and in turn | |
1161 | * the event wasn't already released by fork | |
1162 | * (the event is allocated on fork kernel | |
1163 | * stack), put the event back to its place in | |
1164 | * the event_wq. fork_event head will be freed | |
1165 | * as soon as we return so the event cannot | |
1166 | * stay queued there no matter the current | |
1167 | * "ret" value. | |
1168 | */ | |
1169 | list_del(&uwq->wq.entry); | |
1170 | __add_wait_queue(&ctx->event_wqh, &uwq->wq); | |
893e26e6 | 1171 | |
384632e6 AA |
1172 | /* |
1173 | * Leave the event in the waitqueue and report | |
1174 | * error to userland if we failed to resolve | |
1175 | * the userfault fork. | |
1176 | */ | |
1177 | if (likely(!ret)) | |
893e26e6 | 1178 | userfaultfd_event_complete(ctx, uwq); |
384632e6 AA |
1179 | } else { |
1180 | /* | |
1181 | * Here the fork thread aborted and the | |
1182 | * refcount from the fork thread on fork_nctx | |
1183 | * has already been released. We still hold | |
1184 | * the reference we took before releasing the | |
1185 | * lock above. If resolve_userfault_fork | |
1186 | * failed we've to drop it because the | |
1187 | * fork_nctx has to be freed in such case. If | |
1188 | * it succeeded we'll hold it because the new | |
1189 | * uffd references it. | |
1190 | */ | |
1191 | if (ret) | |
1192 | userfaultfd_ctx_put(fork_nctx); | |
893e26e6 | 1193 | } |
cbcfa130 | 1194 | spin_unlock_irq(&ctx->event_wqh.lock); |
893e26e6 PE |
1195 | } |
1196 | ||
86039bd3 AA |
1197 | return ret; |
1198 | } | |
1199 | ||
1200 | static ssize_t userfaultfd_read(struct file *file, char __user *buf, | |
1201 | size_t count, loff_t *ppos) | |
1202 | { | |
1203 | struct userfaultfd_ctx *ctx = file->private_data; | |
1204 | ssize_t _ret, ret = 0; | |
a9b85f94 | 1205 | struct uffd_msg msg; |
86039bd3 AA |
1206 | int no_wait = file->f_flags & O_NONBLOCK; |
1207 | ||
1208 | if (ctx->state == UFFD_STATE_WAIT_API) | |
1209 | return -EINVAL; | |
86039bd3 AA |
1210 | |
1211 | for (;;) { | |
a9b85f94 | 1212 | if (count < sizeof(msg)) |
86039bd3 | 1213 | return ret ? ret : -EINVAL; |
a9b85f94 | 1214 | _ret = userfaultfd_ctx_read(ctx, no_wait, &msg); |
86039bd3 AA |
1215 | if (_ret < 0) |
1216 | return ret ? ret : _ret; | |
a9b85f94 | 1217 | if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg))) |
86039bd3 | 1218 | return ret ? ret : -EFAULT; |
a9b85f94 AA |
1219 | ret += sizeof(msg); |
1220 | buf += sizeof(msg); | |
1221 | count -= sizeof(msg); | |
86039bd3 AA |
1222 | /* |
1223 | * Allow to read more than one fault at time but only | |
1224 | * block if waiting for the very first one. | |
1225 | */ | |
1226 | no_wait = O_NONBLOCK; | |
1227 | } | |
1228 | } | |
1229 | ||
1230 | static void __wake_userfault(struct userfaultfd_ctx *ctx, | |
1231 | struct userfaultfd_wake_range *range) | |
1232 | { | |
cbcfa130 | 1233 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 1234 | /* wake all in the range and autoremove */ |
15b726ef | 1235 | if (waitqueue_active(&ctx->fault_pending_wqh)) |
ac5be6b4 | 1236 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, |
15b726ef AA |
1237 | range); |
1238 | if (waitqueue_active(&ctx->fault_wqh)) | |
c430d1e8 | 1239 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range); |
cbcfa130 | 1240 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1241 | } |
1242 | ||
1243 | static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx, | |
1244 | struct userfaultfd_wake_range *range) | |
1245 | { | |
2c5b7e1b AA |
1246 | unsigned seq; |
1247 | bool need_wakeup; | |
1248 | ||
86039bd3 AA |
1249 | /* |
1250 | * To be sure waitqueue_active() is not reordered by the CPU | |
1251 | * before the pagetable update, use an explicit SMP memory | |
1252 | * barrier here. PT lock release or up_read(mmap_sem) still | |
1253 | * have release semantics that can allow the | |
1254 | * waitqueue_active() to be reordered before the pte update. | |
1255 | */ | |
1256 | smp_mb(); | |
1257 | ||
1258 | /* | |
1259 | * Use waitqueue_active because it's very frequent to | |
1260 | * change the address space atomically even if there are no | |
1261 | * userfaults yet. So we take the spinlock only when we're | |
1262 | * sure we've userfaults to wake. | |
1263 | */ | |
2c5b7e1b AA |
1264 | do { |
1265 | seq = read_seqcount_begin(&ctx->refile_seq); | |
1266 | need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) || | |
1267 | waitqueue_active(&ctx->fault_wqh); | |
1268 | cond_resched(); | |
1269 | } while (read_seqcount_retry(&ctx->refile_seq, seq)); | |
1270 | if (need_wakeup) | |
86039bd3 AA |
1271 | __wake_userfault(ctx, range); |
1272 | } | |
1273 | ||
1274 | static __always_inline int validate_range(struct mm_struct *mm, | |
7d032574 | 1275 | __u64 *start, __u64 len) |
86039bd3 AA |
1276 | { |
1277 | __u64 task_size = mm->task_size; | |
1278 | ||
7d032574 AK |
1279 | *start = untagged_addr(*start); |
1280 | ||
1281 | if (*start & ~PAGE_MASK) | |
86039bd3 AA |
1282 | return -EINVAL; |
1283 | if (len & ~PAGE_MASK) | |
1284 | return -EINVAL; | |
1285 | if (!len) | |
1286 | return -EINVAL; | |
7d032574 | 1287 | if (*start < mmap_min_addr) |
86039bd3 | 1288 | return -EINVAL; |
7d032574 | 1289 | if (*start >= task_size) |
86039bd3 | 1290 | return -EINVAL; |
7d032574 | 1291 | if (len > task_size - *start) |
86039bd3 AA |
1292 | return -EINVAL; |
1293 | return 0; | |
1294 | } | |
1295 | ||
ba6907db MR |
1296 | static inline bool vma_can_userfault(struct vm_area_struct *vma) |
1297 | { | |
cac67329 MR |
1298 | return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) || |
1299 | vma_is_shmem(vma); | |
ba6907db MR |
1300 | } |
1301 | ||
86039bd3 AA |
1302 | static int userfaultfd_register(struct userfaultfd_ctx *ctx, |
1303 | unsigned long arg) | |
1304 | { | |
1305 | struct mm_struct *mm = ctx->mm; | |
1306 | struct vm_area_struct *vma, *prev, *cur; | |
1307 | int ret; | |
1308 | struct uffdio_register uffdio_register; | |
1309 | struct uffdio_register __user *user_uffdio_register; | |
1310 | unsigned long vm_flags, new_flags; | |
1311 | bool found; | |
ce53e8e6 | 1312 | bool basic_ioctls; |
86039bd3 AA |
1313 | unsigned long start, end, vma_end; |
1314 | ||
1315 | user_uffdio_register = (struct uffdio_register __user *) arg; | |
1316 | ||
1317 | ret = -EFAULT; | |
1318 | if (copy_from_user(&uffdio_register, user_uffdio_register, | |
1319 | sizeof(uffdio_register)-sizeof(__u64))) | |
1320 | goto out; | |
1321 | ||
1322 | ret = -EINVAL; | |
1323 | if (!uffdio_register.mode) | |
1324 | goto out; | |
1325 | if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING| | |
1326 | UFFDIO_REGISTER_MODE_WP)) | |
1327 | goto out; | |
1328 | vm_flags = 0; | |
1329 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING) | |
1330 | vm_flags |= VM_UFFD_MISSING; | |
1331 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) { | |
1332 | vm_flags |= VM_UFFD_WP; | |
1333 | /* | |
1334 | * FIXME: remove the below error constraint by | |
1335 | * implementing the wprotect tracking mode. | |
1336 | */ | |
1337 | ret = -EINVAL; | |
1338 | goto out; | |
1339 | } | |
1340 | ||
7d032574 | 1341 | ret = validate_range(mm, &uffdio_register.range.start, |
86039bd3 AA |
1342 | uffdio_register.range.len); |
1343 | if (ret) | |
1344 | goto out; | |
1345 | ||
1346 | start = uffdio_register.range.start; | |
1347 | end = start + uffdio_register.range.len; | |
1348 | ||
d2005e3f ON |
1349 | ret = -ENOMEM; |
1350 | if (!mmget_not_zero(mm)) | |
1351 | goto out; | |
1352 | ||
86039bd3 | 1353 | down_write(&mm->mmap_sem); |
04f5866e AA |
1354 | if (!mmget_still_valid(mm)) |
1355 | goto out_unlock; | |
86039bd3 | 1356 | vma = find_vma_prev(mm, start, &prev); |
86039bd3 AA |
1357 | if (!vma) |
1358 | goto out_unlock; | |
1359 | ||
1360 | /* check that there's at least one vma in the range */ | |
1361 | ret = -EINVAL; | |
1362 | if (vma->vm_start >= end) | |
1363 | goto out_unlock; | |
1364 | ||
cab350af MK |
1365 | /* |
1366 | * If the first vma contains huge pages, make sure start address | |
1367 | * is aligned to huge page size. | |
1368 | */ | |
1369 | if (is_vm_hugetlb_page(vma)) { | |
1370 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1371 | ||
1372 | if (start & (vma_hpagesize - 1)) | |
1373 | goto out_unlock; | |
1374 | } | |
1375 | ||
86039bd3 AA |
1376 | /* |
1377 | * Search for not compatible vmas. | |
86039bd3 AA |
1378 | */ |
1379 | found = false; | |
ce53e8e6 | 1380 | basic_ioctls = false; |
86039bd3 AA |
1381 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { |
1382 | cond_resched(); | |
1383 | ||
1384 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
1385 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
1386 | ||
1387 | /* check not compatible vmas */ | |
1388 | ret = -EINVAL; | |
ba6907db | 1389 | if (!vma_can_userfault(cur)) |
86039bd3 | 1390 | goto out_unlock; |
29ec9066 AA |
1391 | |
1392 | /* | |
1393 | * UFFDIO_COPY will fill file holes even without | |
1394 | * PROT_WRITE. This check enforces that if this is a | |
1395 | * MAP_SHARED, the process has write permission to the backing | |
1396 | * file. If VM_MAYWRITE is set it also enforces that on a | |
1397 | * MAP_SHARED vma: there is no F_WRITE_SEAL and no further | |
1398 | * F_WRITE_SEAL can be taken until the vma is destroyed. | |
1399 | */ | |
1400 | ret = -EPERM; | |
1401 | if (unlikely(!(cur->vm_flags & VM_MAYWRITE))) | |
1402 | goto out_unlock; | |
1403 | ||
cab350af MK |
1404 | /* |
1405 | * If this vma contains ending address, and huge pages | |
1406 | * check alignment. | |
1407 | */ | |
1408 | if (is_vm_hugetlb_page(cur) && end <= cur->vm_end && | |
1409 | end > cur->vm_start) { | |
1410 | unsigned long vma_hpagesize = vma_kernel_pagesize(cur); | |
1411 | ||
1412 | ret = -EINVAL; | |
1413 | ||
1414 | if (end & (vma_hpagesize - 1)) | |
1415 | goto out_unlock; | |
1416 | } | |
86039bd3 AA |
1417 | |
1418 | /* | |
1419 | * Check that this vma isn't already owned by a | |
1420 | * different userfaultfd. We can't allow more than one | |
1421 | * userfaultfd to own a single vma simultaneously or we | |
1422 | * wouldn't know which one to deliver the userfaults to. | |
1423 | */ | |
1424 | ret = -EBUSY; | |
1425 | if (cur->vm_userfaultfd_ctx.ctx && | |
1426 | cur->vm_userfaultfd_ctx.ctx != ctx) | |
1427 | goto out_unlock; | |
1428 | ||
cab350af MK |
1429 | /* |
1430 | * Note vmas containing huge pages | |
1431 | */ | |
ce53e8e6 MR |
1432 | if (is_vm_hugetlb_page(cur)) |
1433 | basic_ioctls = true; | |
cab350af | 1434 | |
86039bd3 AA |
1435 | found = true; |
1436 | } | |
1437 | BUG_ON(!found); | |
1438 | ||
1439 | if (vma->vm_start < start) | |
1440 | prev = vma; | |
1441 | ||
1442 | ret = 0; | |
1443 | do { | |
1444 | cond_resched(); | |
1445 | ||
ba6907db | 1446 | BUG_ON(!vma_can_userfault(vma)); |
86039bd3 AA |
1447 | BUG_ON(vma->vm_userfaultfd_ctx.ctx && |
1448 | vma->vm_userfaultfd_ctx.ctx != ctx); | |
29ec9066 | 1449 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
86039bd3 AA |
1450 | |
1451 | /* | |
1452 | * Nothing to do: this vma is already registered into this | |
1453 | * userfaultfd and with the right tracking mode too. | |
1454 | */ | |
1455 | if (vma->vm_userfaultfd_ctx.ctx == ctx && | |
1456 | (vma->vm_flags & vm_flags) == vm_flags) | |
1457 | goto skip; | |
1458 | ||
1459 | if (vma->vm_start > start) | |
1460 | start = vma->vm_start; | |
1461 | vma_end = min(end, vma->vm_end); | |
1462 | ||
9d4678eb AA |
1463 | new_flags = (vma->vm_flags & |
1464 | ~(VM_UFFD_MISSING|VM_UFFD_WP)) | vm_flags; | |
86039bd3 AA |
1465 | prev = vma_merge(mm, prev, start, vma_end, new_flags, |
1466 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1467 | vma_policy(vma), | |
1468 | ((struct vm_userfaultfd_ctx){ ctx })); | |
1469 | if (prev) { | |
1470 | vma = prev; | |
1471 | goto next; | |
1472 | } | |
1473 | if (vma->vm_start < start) { | |
1474 | ret = split_vma(mm, vma, start, 1); | |
1475 | if (ret) | |
1476 | break; | |
1477 | } | |
1478 | if (vma->vm_end > end) { | |
1479 | ret = split_vma(mm, vma, end, 0); | |
1480 | if (ret) | |
1481 | break; | |
1482 | } | |
1483 | next: | |
1484 | /* | |
1485 | * In the vma_merge() successful mprotect-like case 8: | |
1486 | * the next vma was merged into the current one and | |
1487 | * the current one has not been updated yet. | |
1488 | */ | |
1489 | vma->vm_flags = new_flags; | |
1490 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
1491 | ||
1492 | skip: | |
1493 | prev = vma; | |
1494 | start = vma->vm_end; | |
1495 | vma = vma->vm_next; | |
1496 | } while (vma && vma->vm_start < end); | |
1497 | out_unlock: | |
1498 | up_write(&mm->mmap_sem); | |
d2005e3f | 1499 | mmput(mm); |
86039bd3 AA |
1500 | if (!ret) { |
1501 | /* | |
1502 | * Now that we scanned all vmas we can already tell | |
1503 | * userland which ioctls methods are guaranteed to | |
1504 | * succeed on this range. | |
1505 | */ | |
ce53e8e6 | 1506 | if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC : |
cab350af | 1507 | UFFD_API_RANGE_IOCTLS, |
86039bd3 AA |
1508 | &user_uffdio_register->ioctls)) |
1509 | ret = -EFAULT; | |
1510 | } | |
1511 | out: | |
1512 | return ret; | |
1513 | } | |
1514 | ||
1515 | static int userfaultfd_unregister(struct userfaultfd_ctx *ctx, | |
1516 | unsigned long arg) | |
1517 | { | |
1518 | struct mm_struct *mm = ctx->mm; | |
1519 | struct vm_area_struct *vma, *prev, *cur; | |
1520 | int ret; | |
1521 | struct uffdio_range uffdio_unregister; | |
1522 | unsigned long new_flags; | |
1523 | bool found; | |
1524 | unsigned long start, end, vma_end; | |
1525 | const void __user *buf = (void __user *)arg; | |
1526 | ||
1527 | ret = -EFAULT; | |
1528 | if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister))) | |
1529 | goto out; | |
1530 | ||
7d032574 | 1531 | ret = validate_range(mm, &uffdio_unregister.start, |
86039bd3 AA |
1532 | uffdio_unregister.len); |
1533 | if (ret) | |
1534 | goto out; | |
1535 | ||
1536 | start = uffdio_unregister.start; | |
1537 | end = start + uffdio_unregister.len; | |
1538 | ||
d2005e3f ON |
1539 | ret = -ENOMEM; |
1540 | if (!mmget_not_zero(mm)) | |
1541 | goto out; | |
1542 | ||
86039bd3 | 1543 | down_write(&mm->mmap_sem); |
04f5866e AA |
1544 | if (!mmget_still_valid(mm)) |
1545 | goto out_unlock; | |
86039bd3 | 1546 | vma = find_vma_prev(mm, start, &prev); |
86039bd3 AA |
1547 | if (!vma) |
1548 | goto out_unlock; | |
1549 | ||
1550 | /* check that there's at least one vma in the range */ | |
1551 | ret = -EINVAL; | |
1552 | if (vma->vm_start >= end) | |
1553 | goto out_unlock; | |
1554 | ||
cab350af MK |
1555 | /* |
1556 | * If the first vma contains huge pages, make sure start address | |
1557 | * is aligned to huge page size. | |
1558 | */ | |
1559 | if (is_vm_hugetlb_page(vma)) { | |
1560 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1561 | ||
1562 | if (start & (vma_hpagesize - 1)) | |
1563 | goto out_unlock; | |
1564 | } | |
1565 | ||
86039bd3 AA |
1566 | /* |
1567 | * Search for not compatible vmas. | |
86039bd3 AA |
1568 | */ |
1569 | found = false; | |
1570 | ret = -EINVAL; | |
1571 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { | |
1572 | cond_resched(); | |
1573 | ||
1574 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
1575 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
1576 | ||
1577 | /* | |
1578 | * Check not compatible vmas, not strictly required | |
1579 | * here as not compatible vmas cannot have an | |
1580 | * userfaultfd_ctx registered on them, but this | |
1581 | * provides for more strict behavior to notice | |
1582 | * unregistration errors. | |
1583 | */ | |
ba6907db | 1584 | if (!vma_can_userfault(cur)) |
86039bd3 AA |
1585 | goto out_unlock; |
1586 | ||
1587 | found = true; | |
1588 | } | |
1589 | BUG_ON(!found); | |
1590 | ||
1591 | if (vma->vm_start < start) | |
1592 | prev = vma; | |
1593 | ||
1594 | ret = 0; | |
1595 | do { | |
1596 | cond_resched(); | |
1597 | ||
ba6907db | 1598 | BUG_ON(!vma_can_userfault(vma)); |
86039bd3 AA |
1599 | |
1600 | /* | |
1601 | * Nothing to do: this vma is already registered into this | |
1602 | * userfaultfd and with the right tracking mode too. | |
1603 | */ | |
1604 | if (!vma->vm_userfaultfd_ctx.ctx) | |
1605 | goto skip; | |
1606 | ||
01e881f5 AA |
1607 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
1608 | ||
86039bd3 AA |
1609 | if (vma->vm_start > start) |
1610 | start = vma->vm_start; | |
1611 | vma_end = min(end, vma->vm_end); | |
1612 | ||
09fa5296 AA |
1613 | if (userfaultfd_missing(vma)) { |
1614 | /* | |
1615 | * Wake any concurrent pending userfault while | |
1616 | * we unregister, so they will not hang | |
1617 | * permanently and it avoids userland to call | |
1618 | * UFFDIO_WAKE explicitly. | |
1619 | */ | |
1620 | struct userfaultfd_wake_range range; | |
1621 | range.start = start; | |
1622 | range.len = vma_end - start; | |
1623 | wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range); | |
1624 | } | |
1625 | ||
86039bd3 AA |
1626 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); |
1627 | prev = vma_merge(mm, prev, start, vma_end, new_flags, | |
1628 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1629 | vma_policy(vma), | |
1630 | NULL_VM_UFFD_CTX); | |
1631 | if (prev) { | |
1632 | vma = prev; | |
1633 | goto next; | |
1634 | } | |
1635 | if (vma->vm_start < start) { | |
1636 | ret = split_vma(mm, vma, start, 1); | |
1637 | if (ret) | |
1638 | break; | |
1639 | } | |
1640 | if (vma->vm_end > end) { | |
1641 | ret = split_vma(mm, vma, end, 0); | |
1642 | if (ret) | |
1643 | break; | |
1644 | } | |
1645 | next: | |
1646 | /* | |
1647 | * In the vma_merge() successful mprotect-like case 8: | |
1648 | * the next vma was merged into the current one and | |
1649 | * the current one has not been updated yet. | |
1650 | */ | |
1651 | vma->vm_flags = new_flags; | |
1652 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
1653 | ||
1654 | skip: | |
1655 | prev = vma; | |
1656 | start = vma->vm_end; | |
1657 | vma = vma->vm_next; | |
1658 | } while (vma && vma->vm_start < end); | |
1659 | out_unlock: | |
1660 | up_write(&mm->mmap_sem); | |
d2005e3f | 1661 | mmput(mm); |
86039bd3 AA |
1662 | out: |
1663 | return ret; | |
1664 | } | |
1665 | ||
1666 | /* | |
ba85c702 AA |
1667 | * userfaultfd_wake may be used in combination with the |
1668 | * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches. | |
86039bd3 AA |
1669 | */ |
1670 | static int userfaultfd_wake(struct userfaultfd_ctx *ctx, | |
1671 | unsigned long arg) | |
1672 | { | |
1673 | int ret; | |
1674 | struct uffdio_range uffdio_wake; | |
1675 | struct userfaultfd_wake_range range; | |
1676 | const void __user *buf = (void __user *)arg; | |
1677 | ||
1678 | ret = -EFAULT; | |
1679 | if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake))) | |
1680 | goto out; | |
1681 | ||
7d032574 | 1682 | ret = validate_range(ctx->mm, &uffdio_wake.start, uffdio_wake.len); |
86039bd3 AA |
1683 | if (ret) |
1684 | goto out; | |
1685 | ||
1686 | range.start = uffdio_wake.start; | |
1687 | range.len = uffdio_wake.len; | |
1688 | ||
1689 | /* | |
1690 | * len == 0 means wake all and we don't want to wake all here, | |
1691 | * so check it again to be sure. | |
1692 | */ | |
1693 | VM_BUG_ON(!range.len); | |
1694 | ||
1695 | wake_userfault(ctx, &range); | |
1696 | ret = 0; | |
1697 | ||
1698 | out: | |
1699 | return ret; | |
1700 | } | |
1701 | ||
ad465cae AA |
1702 | static int userfaultfd_copy(struct userfaultfd_ctx *ctx, |
1703 | unsigned long arg) | |
1704 | { | |
1705 | __s64 ret; | |
1706 | struct uffdio_copy uffdio_copy; | |
1707 | struct uffdio_copy __user *user_uffdio_copy; | |
1708 | struct userfaultfd_wake_range range; | |
1709 | ||
1710 | user_uffdio_copy = (struct uffdio_copy __user *) arg; | |
1711 | ||
df2cc96e MR |
1712 | ret = -EAGAIN; |
1713 | if (READ_ONCE(ctx->mmap_changing)) | |
1714 | goto out; | |
1715 | ||
ad465cae AA |
1716 | ret = -EFAULT; |
1717 | if (copy_from_user(&uffdio_copy, user_uffdio_copy, | |
1718 | /* don't copy "copy" last field */ | |
1719 | sizeof(uffdio_copy)-sizeof(__s64))) | |
1720 | goto out; | |
1721 | ||
7d032574 | 1722 | ret = validate_range(ctx->mm, &uffdio_copy.dst, uffdio_copy.len); |
ad465cae AA |
1723 | if (ret) |
1724 | goto out; | |
1725 | /* | |
1726 | * double check for wraparound just in case. copy_from_user() | |
1727 | * will later check uffdio_copy.src + uffdio_copy.len to fit | |
1728 | * in the userland range. | |
1729 | */ | |
1730 | ret = -EINVAL; | |
1731 | if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src) | |
1732 | goto out; | |
1733 | if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE) | |
1734 | goto out; | |
d2005e3f ON |
1735 | if (mmget_not_zero(ctx->mm)) { |
1736 | ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src, | |
df2cc96e | 1737 | uffdio_copy.len, &ctx->mmap_changing); |
d2005e3f | 1738 | mmput(ctx->mm); |
96333187 | 1739 | } else { |
e86b298b | 1740 | return -ESRCH; |
d2005e3f | 1741 | } |
ad465cae AA |
1742 | if (unlikely(put_user(ret, &user_uffdio_copy->copy))) |
1743 | return -EFAULT; | |
1744 | if (ret < 0) | |
1745 | goto out; | |
1746 | BUG_ON(!ret); | |
1747 | /* len == 0 would wake all */ | |
1748 | range.len = ret; | |
1749 | if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) { | |
1750 | range.start = uffdio_copy.dst; | |
1751 | wake_userfault(ctx, &range); | |
1752 | } | |
1753 | ret = range.len == uffdio_copy.len ? 0 : -EAGAIN; | |
1754 | out: | |
1755 | return ret; | |
1756 | } | |
1757 | ||
1758 | static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx, | |
1759 | unsigned long arg) | |
1760 | { | |
1761 | __s64 ret; | |
1762 | struct uffdio_zeropage uffdio_zeropage; | |
1763 | struct uffdio_zeropage __user *user_uffdio_zeropage; | |
1764 | struct userfaultfd_wake_range range; | |
1765 | ||
1766 | user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg; | |
1767 | ||
df2cc96e MR |
1768 | ret = -EAGAIN; |
1769 | if (READ_ONCE(ctx->mmap_changing)) | |
1770 | goto out; | |
1771 | ||
ad465cae AA |
1772 | ret = -EFAULT; |
1773 | if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage, | |
1774 | /* don't copy "zeropage" last field */ | |
1775 | sizeof(uffdio_zeropage)-sizeof(__s64))) | |
1776 | goto out; | |
1777 | ||
7d032574 | 1778 | ret = validate_range(ctx->mm, &uffdio_zeropage.range.start, |
ad465cae AA |
1779 | uffdio_zeropage.range.len); |
1780 | if (ret) | |
1781 | goto out; | |
1782 | ret = -EINVAL; | |
1783 | if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE) | |
1784 | goto out; | |
1785 | ||
d2005e3f ON |
1786 | if (mmget_not_zero(ctx->mm)) { |
1787 | ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start, | |
df2cc96e MR |
1788 | uffdio_zeropage.range.len, |
1789 | &ctx->mmap_changing); | |
d2005e3f | 1790 | mmput(ctx->mm); |
9d95aa4b | 1791 | } else { |
e86b298b | 1792 | return -ESRCH; |
d2005e3f | 1793 | } |
ad465cae AA |
1794 | if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage))) |
1795 | return -EFAULT; | |
1796 | if (ret < 0) | |
1797 | goto out; | |
1798 | /* len == 0 would wake all */ | |
1799 | BUG_ON(!ret); | |
1800 | range.len = ret; | |
1801 | if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) { | |
1802 | range.start = uffdio_zeropage.range.start; | |
1803 | wake_userfault(ctx, &range); | |
1804 | } | |
1805 | ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN; | |
1806 | out: | |
1807 | return ret; | |
1808 | } | |
1809 | ||
9cd75c3c PE |
1810 | static inline unsigned int uffd_ctx_features(__u64 user_features) |
1811 | { | |
1812 | /* | |
1813 | * For the current set of features the bits just coincide | |
1814 | */ | |
1815 | return (unsigned int)user_features; | |
1816 | } | |
1817 | ||
86039bd3 AA |
1818 | /* |
1819 | * userland asks for a certain API version and we return which bits | |
1820 | * and ioctl commands are implemented in this kernel for such API | |
1821 | * version or -EINVAL if unknown. | |
1822 | */ | |
1823 | static int userfaultfd_api(struct userfaultfd_ctx *ctx, | |
1824 | unsigned long arg) | |
1825 | { | |
1826 | struct uffdio_api uffdio_api; | |
1827 | void __user *buf = (void __user *)arg; | |
1828 | int ret; | |
65603144 | 1829 | __u64 features; |
86039bd3 AA |
1830 | |
1831 | ret = -EINVAL; | |
1832 | if (ctx->state != UFFD_STATE_WAIT_API) | |
1833 | goto out; | |
1834 | ret = -EFAULT; | |
a9b85f94 | 1835 | if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api))) |
86039bd3 | 1836 | goto out; |
65603144 | 1837 | features = uffdio_api.features; |
3c1c24d9 MR |
1838 | ret = -EINVAL; |
1839 | if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) | |
1840 | goto err_out; | |
1841 | ret = -EPERM; | |
1842 | if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE)) | |
1843 | goto err_out; | |
65603144 AA |
1844 | /* report all available features and ioctls to userland */ |
1845 | uffdio_api.features = UFFD_API_FEATURES; | |
86039bd3 AA |
1846 | uffdio_api.ioctls = UFFD_API_IOCTLS; |
1847 | ret = -EFAULT; | |
1848 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1849 | goto out; | |
1850 | ctx->state = UFFD_STATE_RUNNING; | |
65603144 AA |
1851 | /* only enable the requested features for this uffd context */ |
1852 | ctx->features = uffd_ctx_features(features); | |
86039bd3 AA |
1853 | ret = 0; |
1854 | out: | |
1855 | return ret; | |
3c1c24d9 MR |
1856 | err_out: |
1857 | memset(&uffdio_api, 0, sizeof(uffdio_api)); | |
1858 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1859 | ret = -EFAULT; | |
1860 | goto out; | |
86039bd3 AA |
1861 | } |
1862 | ||
1863 | static long userfaultfd_ioctl(struct file *file, unsigned cmd, | |
1864 | unsigned long arg) | |
1865 | { | |
1866 | int ret = -EINVAL; | |
1867 | struct userfaultfd_ctx *ctx = file->private_data; | |
1868 | ||
e6485a47 AA |
1869 | if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API) |
1870 | return -EINVAL; | |
1871 | ||
86039bd3 AA |
1872 | switch(cmd) { |
1873 | case UFFDIO_API: | |
1874 | ret = userfaultfd_api(ctx, arg); | |
1875 | break; | |
1876 | case UFFDIO_REGISTER: | |
1877 | ret = userfaultfd_register(ctx, arg); | |
1878 | break; | |
1879 | case UFFDIO_UNREGISTER: | |
1880 | ret = userfaultfd_unregister(ctx, arg); | |
1881 | break; | |
1882 | case UFFDIO_WAKE: | |
1883 | ret = userfaultfd_wake(ctx, arg); | |
1884 | break; | |
ad465cae AA |
1885 | case UFFDIO_COPY: |
1886 | ret = userfaultfd_copy(ctx, arg); | |
1887 | break; | |
1888 | case UFFDIO_ZEROPAGE: | |
1889 | ret = userfaultfd_zeropage(ctx, arg); | |
1890 | break; | |
86039bd3 AA |
1891 | } |
1892 | return ret; | |
1893 | } | |
1894 | ||
1895 | #ifdef CONFIG_PROC_FS | |
1896 | static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f) | |
1897 | { | |
1898 | struct userfaultfd_ctx *ctx = f->private_data; | |
ac6424b9 | 1899 | wait_queue_entry_t *wq; |
86039bd3 AA |
1900 | unsigned long pending = 0, total = 0; |
1901 | ||
cbcfa130 | 1902 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
2055da97 | 1903 | list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) { |
15b726ef AA |
1904 | pending++; |
1905 | total++; | |
1906 | } | |
2055da97 | 1907 | list_for_each_entry(wq, &ctx->fault_wqh.head, entry) { |
86039bd3 AA |
1908 | total++; |
1909 | } | |
cbcfa130 | 1910 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1911 | |
1912 | /* | |
1913 | * If more protocols will be added, there will be all shown | |
1914 | * separated by a space. Like this: | |
1915 | * protocols: aa:... bb:... | |
1916 | */ | |
1917 | seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n", | |
045098e9 | 1918 | pending, total, UFFD_API, ctx->features, |
86039bd3 AA |
1919 | UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS); |
1920 | } | |
1921 | #endif | |
1922 | ||
1923 | static const struct file_operations userfaultfd_fops = { | |
1924 | #ifdef CONFIG_PROC_FS | |
1925 | .show_fdinfo = userfaultfd_show_fdinfo, | |
1926 | #endif | |
1927 | .release = userfaultfd_release, | |
1928 | .poll = userfaultfd_poll, | |
1929 | .read = userfaultfd_read, | |
1930 | .unlocked_ioctl = userfaultfd_ioctl, | |
1832f2d8 | 1931 | .compat_ioctl = compat_ptr_ioctl, |
86039bd3 AA |
1932 | .llseek = noop_llseek, |
1933 | }; | |
1934 | ||
3004ec9c AA |
1935 | static void init_once_userfaultfd_ctx(void *mem) |
1936 | { | |
1937 | struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem; | |
1938 | ||
1939 | init_waitqueue_head(&ctx->fault_pending_wqh); | |
1940 | init_waitqueue_head(&ctx->fault_wqh); | |
9cd75c3c | 1941 | init_waitqueue_head(&ctx->event_wqh); |
3004ec9c | 1942 | init_waitqueue_head(&ctx->fd_wqh); |
2c5b7e1b | 1943 | seqcount_init(&ctx->refile_seq); |
3004ec9c AA |
1944 | } |
1945 | ||
284cd241 | 1946 | SYSCALL_DEFINE1(userfaultfd, int, flags) |
86039bd3 | 1947 | { |
86039bd3 | 1948 | struct userfaultfd_ctx *ctx; |
284cd241 | 1949 | int fd; |
86039bd3 | 1950 | |
cefdca0a PX |
1951 | if (!sysctl_unprivileged_userfaultfd && !capable(CAP_SYS_PTRACE)) |
1952 | return -EPERM; | |
1953 | ||
86039bd3 AA |
1954 | BUG_ON(!current->mm); |
1955 | ||
1956 | /* Check the UFFD_* constants for consistency. */ | |
1957 | BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC); | |
1958 | BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK); | |
1959 | ||
86039bd3 | 1960 | if (flags & ~UFFD_SHARED_FCNTL_FLAGS) |
284cd241 | 1961 | return -EINVAL; |
86039bd3 | 1962 | |
3004ec9c | 1963 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); |
86039bd3 | 1964 | if (!ctx) |
284cd241 | 1965 | return -ENOMEM; |
86039bd3 | 1966 | |
ca880420 | 1967 | refcount_set(&ctx->refcount, 1); |
86039bd3 | 1968 | ctx->flags = flags; |
9cd75c3c | 1969 | ctx->features = 0; |
86039bd3 AA |
1970 | ctx->state = UFFD_STATE_WAIT_API; |
1971 | ctx->released = false; | |
df2cc96e | 1972 | ctx->mmap_changing = false; |
86039bd3 AA |
1973 | ctx->mm = current->mm; |
1974 | /* prevent the mm struct to be freed */ | |
f1f10076 | 1975 | mmgrab(ctx->mm); |
86039bd3 | 1976 | |
284cd241 EB |
1977 | fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, ctx, |
1978 | O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS)); | |
1979 | if (fd < 0) { | |
d2005e3f | 1980 | mmdrop(ctx->mm); |
3004ec9c | 1981 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
c03e946f | 1982 | } |
86039bd3 | 1983 | return fd; |
86039bd3 | 1984 | } |
3004ec9c AA |
1985 | |
1986 | static int __init userfaultfd_init(void) | |
1987 | { | |
1988 | userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache", | |
1989 | sizeof(struct userfaultfd_ctx), | |
1990 | 0, | |
1991 | SLAB_HWCACHE_ALIGN|SLAB_PANIC, | |
1992 | init_once_userfaultfd_ctx); | |
1993 | return 0; | |
1994 | } | |
1995 | __initcall(userfaultfd_init); |