]>
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 | 17 | #include <linux/mm.h> |
17fca131 | 18 | #include <linux/mm_inline.h> |
6dfeaff9 | 19 | #include <linux/mmu_notifier.h> |
86039bd3 AA |
20 | #include <linux/poll.h> |
21 | #include <linux/slab.h> | |
22 | #include <linux/seq_file.h> | |
23 | #include <linux/file.h> | |
24 | #include <linux/bug.h> | |
25 | #include <linux/anon_inodes.h> | |
26 | #include <linux/syscalls.h> | |
27 | #include <linux/userfaultfd_k.h> | |
28 | #include <linux/mempolicy.h> | |
29 | #include <linux/ioctl.h> | |
30 | #include <linux/security.h> | |
cab350af | 31 | #include <linux/hugetlb.h> |
5c041f5d | 32 | #include <linux/swapops.h> |
86039bd3 | 33 | |
d0d4730a | 34 | int sysctl_unprivileged_userfaultfd __read_mostly; |
cefdca0a | 35 | |
3004ec9c AA |
36 | static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly; |
37 | ||
3004ec9c AA |
38 | /* |
39 | * Start with fault_pending_wqh and fault_wqh so they're more likely | |
40 | * to be in the same cacheline. | |
cbcfa130 EB |
41 | * |
42 | * Locking order: | |
43 | * fd_wqh.lock | |
44 | * fault_pending_wqh.lock | |
45 | * fault_wqh.lock | |
46 | * event_wqh.lock | |
47 | * | |
48 | * To avoid deadlocks, IRQs must be disabled when taking any of the above locks, | |
49 | * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's | |
50 | * also taken in IRQ context. | |
3004ec9c | 51 | */ |
86039bd3 | 52 | struct userfaultfd_ctx { |
15b726ef AA |
53 | /* waitqueue head for the pending (i.e. not read) userfaults */ |
54 | wait_queue_head_t fault_pending_wqh; | |
55 | /* waitqueue head for the userfaults */ | |
86039bd3 AA |
56 | wait_queue_head_t fault_wqh; |
57 | /* waitqueue head for the pseudo fd to wakeup poll/read */ | |
58 | wait_queue_head_t fd_wqh; | |
9cd75c3c PE |
59 | /* waitqueue head for events */ |
60 | wait_queue_head_t event_wqh; | |
2c5b7e1b | 61 | /* a refile sequence protected by fault_pending_wqh lock */ |
2ca97ac8 | 62 | seqcount_spinlock_t refile_seq; |
3004ec9c | 63 | /* pseudo fd refcounting */ |
ca880420 | 64 | refcount_t refcount; |
86039bd3 AA |
65 | /* userfaultfd syscall flags */ |
66 | unsigned int flags; | |
9cd75c3c PE |
67 | /* features requested from the userspace */ |
68 | unsigned int features; | |
86039bd3 AA |
69 | /* released */ |
70 | bool released; | |
df2cc96e | 71 | /* memory mappings are changing because of non-cooperative event */ |
a759a909 | 72 | atomic_t mmap_changing; |
86039bd3 AA |
73 | /* mm with one ore more vmas attached to this userfaultfd_ctx */ |
74 | struct mm_struct *mm; | |
75 | }; | |
76 | ||
893e26e6 PE |
77 | struct userfaultfd_fork_ctx { |
78 | struct userfaultfd_ctx *orig; | |
79 | struct userfaultfd_ctx *new; | |
80 | struct list_head list; | |
81 | }; | |
82 | ||
897ab3e0 MR |
83 | struct userfaultfd_unmap_ctx { |
84 | struct userfaultfd_ctx *ctx; | |
85 | unsigned long start; | |
86 | unsigned long end; | |
87 | struct list_head list; | |
88 | }; | |
89 | ||
86039bd3 | 90 | struct userfaultfd_wait_queue { |
a9b85f94 | 91 | struct uffd_msg msg; |
ac6424b9 | 92 | wait_queue_entry_t wq; |
86039bd3 | 93 | struct userfaultfd_ctx *ctx; |
15a77c6f | 94 | bool waken; |
86039bd3 AA |
95 | }; |
96 | ||
97 | struct userfaultfd_wake_range { | |
98 | unsigned long start; | |
99 | unsigned long len; | |
100 | }; | |
101 | ||
22e5fe2a NA |
102 | /* internal indication that UFFD_API ioctl was successfully executed */ |
103 | #define UFFD_FEATURE_INITIALIZED (1u << 31) | |
104 | ||
105 | static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx) | |
106 | { | |
107 | return ctx->features & UFFD_FEATURE_INITIALIZED; | |
108 | } | |
109 | ||
ac6424b9 | 110 | static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode, |
86039bd3 AA |
111 | int wake_flags, void *key) |
112 | { | |
113 | struct userfaultfd_wake_range *range = key; | |
114 | int ret; | |
115 | struct userfaultfd_wait_queue *uwq; | |
116 | unsigned long start, len; | |
117 | ||
118 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
119 | ret = 0; | |
86039bd3 AA |
120 | /* len == 0 means wake all */ |
121 | start = range->start; | |
122 | len = range->len; | |
a9b85f94 AA |
123 | if (len && (start > uwq->msg.arg.pagefault.address || |
124 | start + len <= uwq->msg.arg.pagefault.address)) | |
86039bd3 | 125 | goto out; |
15a77c6f AA |
126 | WRITE_ONCE(uwq->waken, true); |
127 | /* | |
a9668cd6 PZ |
128 | * The Program-Order guarantees provided by the scheduler |
129 | * ensure uwq->waken is visible before the task is woken. | |
15a77c6f | 130 | */ |
86039bd3 | 131 | ret = wake_up_state(wq->private, mode); |
a9668cd6 | 132 | if (ret) { |
86039bd3 AA |
133 | /* |
134 | * Wake only once, autoremove behavior. | |
135 | * | |
a9668cd6 PZ |
136 | * After the effect of list_del_init is visible to the other |
137 | * CPUs, the waitqueue may disappear from under us, see the | |
138 | * !list_empty_careful() in handle_userfault(). | |
139 | * | |
140 | * try_to_wake_up() has an implicit smp_mb(), and the | |
141 | * wq->private is read before calling the extern function | |
142 | * "wake_up_state" (which in turns calls try_to_wake_up). | |
86039bd3 | 143 | */ |
2055da97 | 144 | list_del_init(&wq->entry); |
a9668cd6 | 145 | } |
86039bd3 AA |
146 | out: |
147 | return ret; | |
148 | } | |
149 | ||
150 | /** | |
151 | * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd | |
152 | * context. | |
153 | * @ctx: [in] Pointer to the userfaultfd context. | |
86039bd3 AA |
154 | */ |
155 | static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx) | |
156 | { | |
ca880420 | 157 | refcount_inc(&ctx->refcount); |
86039bd3 AA |
158 | } |
159 | ||
160 | /** | |
161 | * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd | |
162 | * context. | |
163 | * @ctx: [in] Pointer to userfaultfd context. | |
164 | * | |
165 | * The userfaultfd context reference must have been previously acquired either | |
166 | * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget(). | |
167 | */ | |
168 | static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx) | |
169 | { | |
ca880420 | 170 | if (refcount_dec_and_test(&ctx->refcount)) { |
86039bd3 AA |
171 | VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock)); |
172 | VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh)); | |
173 | VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock)); | |
174 | VM_BUG_ON(waitqueue_active(&ctx->fault_wqh)); | |
9cd75c3c PE |
175 | VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock)); |
176 | VM_BUG_ON(waitqueue_active(&ctx->event_wqh)); | |
86039bd3 AA |
177 | VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock)); |
178 | VM_BUG_ON(waitqueue_active(&ctx->fd_wqh)); | |
d2005e3f | 179 | mmdrop(ctx->mm); |
3004ec9c | 180 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
86039bd3 AA |
181 | } |
182 | } | |
183 | ||
a9b85f94 | 184 | static inline void msg_init(struct uffd_msg *msg) |
86039bd3 | 185 | { |
a9b85f94 AA |
186 | BUILD_BUG_ON(sizeof(struct uffd_msg) != 32); |
187 | /* | |
188 | * Must use memset to zero out the paddings or kernel data is | |
189 | * leaked to userland. | |
190 | */ | |
191 | memset(msg, 0, sizeof(struct uffd_msg)); | |
192 | } | |
193 | ||
194 | static inline struct uffd_msg userfault_msg(unsigned long address, | |
d172b1a3 | 195 | unsigned long real_address, |
a9b85f94 | 196 | unsigned int flags, |
9d4ac934 AP |
197 | unsigned long reason, |
198 | unsigned int features) | |
a9b85f94 AA |
199 | { |
200 | struct uffd_msg msg; | |
d172b1a3 | 201 | |
a9b85f94 AA |
202 | msg_init(&msg); |
203 | msg.event = UFFD_EVENT_PAGEFAULT; | |
824ddc60 | 204 | |
d172b1a3 NA |
205 | msg.arg.pagefault.address = (features & UFFD_FEATURE_EXACT_ADDRESS) ? |
206 | real_address : address; | |
207 | ||
7677f7fd AR |
208 | /* |
209 | * These flags indicate why the userfault occurred: | |
210 | * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault. | |
211 | * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault. | |
212 | * - Neither of these flags being set indicates a MISSING fault. | |
213 | * | |
214 | * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write | |
215 | * fault. Otherwise, it was a read fault. | |
216 | */ | |
86039bd3 | 217 | if (flags & FAULT_FLAG_WRITE) |
a9b85f94 | 218 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE; |
86039bd3 | 219 | if (reason & VM_UFFD_WP) |
a9b85f94 | 220 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP; |
7677f7fd AR |
221 | if (reason & VM_UFFD_MINOR) |
222 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR; | |
9d4ac934 | 223 | if (features & UFFD_FEATURE_THREAD_ID) |
a36985d3 | 224 | msg.arg.pagefault.feat.ptid = task_pid_vnr(current); |
a9b85f94 | 225 | return msg; |
86039bd3 AA |
226 | } |
227 | ||
369cd212 MK |
228 | #ifdef CONFIG_HUGETLB_PAGE |
229 | /* | |
230 | * Same functionality as userfaultfd_must_wait below with modifications for | |
231 | * hugepmd ranges. | |
232 | */ | |
233 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 234 | struct vm_area_struct *vma, |
369cd212 MK |
235 | unsigned long address, |
236 | unsigned long flags, | |
237 | unsigned long reason) | |
238 | { | |
239 | struct mm_struct *mm = ctx->mm; | |
1e2c0436 | 240 | pte_t *ptep, pte; |
369cd212 MK |
241 | bool ret = true; |
242 | ||
42fc5414 | 243 | mmap_assert_locked(mm); |
369cd212 | 244 | |
1e2c0436 JF |
245 | ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma)); |
246 | ||
247 | if (!ptep) | |
369cd212 MK |
248 | goto out; |
249 | ||
250 | ret = false; | |
1e2c0436 | 251 | pte = huge_ptep_get(ptep); |
369cd212 MK |
252 | |
253 | /* | |
254 | * Lockless access: we're in a wait_event so it's ok if it | |
5c041f5d PX |
255 | * changes under us. PTE markers should be handled the same as none |
256 | * ptes here. | |
369cd212 | 257 | */ |
5c041f5d | 258 | if (huge_pte_none_mostly(pte)) |
369cd212 | 259 | ret = true; |
1e2c0436 | 260 | if (!huge_pte_write(pte) && (reason & VM_UFFD_WP)) |
369cd212 MK |
261 | ret = true; |
262 | out: | |
263 | return ret; | |
264 | } | |
265 | #else | |
266 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 267 | struct vm_area_struct *vma, |
369cd212 MK |
268 | unsigned long address, |
269 | unsigned long flags, | |
270 | unsigned long reason) | |
271 | { | |
272 | return false; /* should never get here */ | |
273 | } | |
274 | #endif /* CONFIG_HUGETLB_PAGE */ | |
275 | ||
8d2afd96 AA |
276 | /* |
277 | * Verify the pagetables are still not ok after having reigstered into | |
278 | * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any | |
279 | * userfault that has already been resolved, if userfaultfd_read and | |
280 | * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different | |
281 | * threads. | |
282 | */ | |
283 | static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx, | |
284 | unsigned long address, | |
285 | unsigned long flags, | |
286 | unsigned long reason) | |
287 | { | |
288 | struct mm_struct *mm = ctx->mm; | |
289 | pgd_t *pgd; | |
c2febafc | 290 | p4d_t *p4d; |
8d2afd96 AA |
291 | pud_t *pud; |
292 | pmd_t *pmd, _pmd; | |
293 | pte_t *pte; | |
294 | bool ret = true; | |
295 | ||
42fc5414 | 296 | mmap_assert_locked(mm); |
8d2afd96 AA |
297 | |
298 | pgd = pgd_offset(mm, address); | |
299 | if (!pgd_present(*pgd)) | |
300 | goto out; | |
c2febafc KS |
301 | p4d = p4d_offset(pgd, address); |
302 | if (!p4d_present(*p4d)) | |
303 | goto out; | |
304 | pud = pud_offset(p4d, address); | |
8d2afd96 AA |
305 | if (!pud_present(*pud)) |
306 | goto out; | |
307 | pmd = pmd_offset(pud, address); | |
308 | /* | |
309 | * READ_ONCE must function as a barrier with narrower scope | |
310 | * and it must be equivalent to: | |
311 | * _pmd = *pmd; barrier(); | |
312 | * | |
313 | * This is to deal with the instability (as in | |
314 | * pmd_trans_unstable) of the pmd. | |
315 | */ | |
316 | _pmd = READ_ONCE(*pmd); | |
a365ac09 | 317 | if (pmd_none(_pmd)) |
8d2afd96 AA |
318 | goto out; |
319 | ||
320 | ret = false; | |
a365ac09 HY |
321 | if (!pmd_present(_pmd)) |
322 | goto out; | |
323 | ||
63b2d417 AA |
324 | if (pmd_trans_huge(_pmd)) { |
325 | if (!pmd_write(_pmd) && (reason & VM_UFFD_WP)) | |
326 | ret = true; | |
8d2afd96 | 327 | goto out; |
63b2d417 | 328 | } |
8d2afd96 AA |
329 | |
330 | /* | |
331 | * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it | |
332 | * and use the standard pte_offset_map() instead of parsing _pmd. | |
333 | */ | |
334 | pte = pte_offset_map(pmd, address); | |
335 | /* | |
336 | * Lockless access: we're in a wait_event so it's ok if it | |
5c041f5d PX |
337 | * changes under us. PTE markers should be handled the same as none |
338 | * ptes here. | |
8d2afd96 | 339 | */ |
5c041f5d | 340 | if (pte_none_mostly(*pte)) |
8d2afd96 | 341 | ret = true; |
63b2d417 AA |
342 | if (!pte_write(*pte) && (reason & VM_UFFD_WP)) |
343 | ret = true; | |
8d2afd96 AA |
344 | pte_unmap(pte); |
345 | ||
346 | out: | |
347 | return ret; | |
348 | } | |
349 | ||
2f064a59 | 350 | static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags) |
3e69ad08 PX |
351 | { |
352 | if (flags & FAULT_FLAG_INTERRUPTIBLE) | |
353 | return TASK_INTERRUPTIBLE; | |
354 | ||
355 | if (flags & FAULT_FLAG_KILLABLE) | |
356 | return TASK_KILLABLE; | |
357 | ||
358 | return TASK_UNINTERRUPTIBLE; | |
359 | } | |
360 | ||
86039bd3 AA |
361 | /* |
362 | * The locking rules involved in returning VM_FAULT_RETRY depending on | |
363 | * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and | |
364 | * FAULT_FLAG_KILLABLE are not straightforward. The "Caution" | |
365 | * recommendation in __lock_page_or_retry is not an understatement. | |
366 | * | |
c1e8d7c6 | 367 | * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released |
86039bd3 AA |
368 | * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is |
369 | * not set. | |
370 | * | |
371 | * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not | |
372 | * set, VM_FAULT_RETRY can still be returned if and only if there are | |
c1e8d7c6 | 373 | * fatal_signal_pending()s, and the mmap_lock must be released before |
86039bd3 AA |
374 | * returning it. |
375 | */ | |
2b740303 | 376 | vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason) |
86039bd3 | 377 | { |
82b0f8c3 | 378 | struct mm_struct *mm = vmf->vma->vm_mm; |
86039bd3 AA |
379 | struct userfaultfd_ctx *ctx; |
380 | struct userfaultfd_wait_queue uwq; | |
2b740303 | 381 | vm_fault_t ret = VM_FAULT_SIGBUS; |
3e69ad08 | 382 | bool must_wait; |
2f064a59 | 383 | unsigned int blocking_state; |
86039bd3 | 384 | |
64c2b203 AA |
385 | /* |
386 | * We don't do userfault handling for the final child pid update. | |
387 | * | |
388 | * We also don't do userfault handling during | |
389 | * coredumping. hugetlbfs has the special | |
390 | * follow_hugetlb_page() to skip missing pages in the | |
391 | * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with | |
392 | * the no_page_table() helper in follow_page_mask(), but the | |
393 | * shmem_vm_ops->fault method is invoked even during | |
c1e8d7c6 | 394 | * coredumping without mmap_lock and it ends up here. |
64c2b203 AA |
395 | */ |
396 | if (current->flags & (PF_EXITING|PF_DUMPCORE)) | |
397 | goto out; | |
398 | ||
399 | /* | |
c1e8d7c6 ML |
400 | * Coredumping runs without mmap_lock so we can only check that |
401 | * the mmap_lock is held, if PF_DUMPCORE was not set. | |
64c2b203 | 402 | */ |
42fc5414 | 403 | mmap_assert_locked(mm); |
64c2b203 | 404 | |
82b0f8c3 | 405 | ctx = vmf->vma->vm_userfaultfd_ctx.ctx; |
86039bd3 | 406 | if (!ctx) |
ba85c702 | 407 | goto out; |
86039bd3 AA |
408 | |
409 | BUG_ON(ctx->mm != mm); | |
410 | ||
7677f7fd AR |
411 | /* Any unrecognized flag is a bug. */ |
412 | VM_BUG_ON(reason & ~__VM_UFFD_FLAGS); | |
413 | /* 0 or > 1 flags set is a bug; we expect exactly 1. */ | |
414 | VM_BUG_ON(!reason || (reason & (reason - 1))); | |
86039bd3 | 415 | |
2d6d6f5a PS |
416 | if (ctx->features & UFFD_FEATURE_SIGBUS) |
417 | goto out; | |
37cd0575 LG |
418 | if ((vmf->flags & FAULT_FLAG_USER) == 0 && |
419 | ctx->flags & UFFD_USER_MODE_ONLY) { | |
420 | printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd " | |
421 | "sysctl knob to 1 if kernel faults must be handled " | |
422 | "without obtaining CAP_SYS_PTRACE capability\n"); | |
423 | goto out; | |
424 | } | |
2d6d6f5a | 425 | |
86039bd3 AA |
426 | /* |
427 | * If it's already released don't get it. This avoids to loop | |
428 | * in __get_user_pages if userfaultfd_release waits on the | |
c1e8d7c6 | 429 | * caller of handle_userfault to release the mmap_lock. |
86039bd3 | 430 | */ |
6aa7de05 | 431 | if (unlikely(READ_ONCE(ctx->released))) { |
656710a6 AA |
432 | /* |
433 | * Don't return VM_FAULT_SIGBUS in this case, so a non | |
434 | * cooperative manager can close the uffd after the | |
435 | * last UFFDIO_COPY, without risking to trigger an | |
436 | * involuntary SIGBUS if the process was starting the | |
437 | * userfaultfd while the userfaultfd was still armed | |
438 | * (but after the last UFFDIO_COPY). If the uffd | |
439 | * wasn't already closed when the userfault reached | |
440 | * this point, that would normally be solved by | |
441 | * userfaultfd_must_wait returning 'false'. | |
442 | * | |
443 | * If we were to return VM_FAULT_SIGBUS here, the non | |
444 | * cooperative manager would be instead forced to | |
445 | * always call UFFDIO_UNREGISTER before it can safely | |
446 | * close the uffd. | |
447 | */ | |
448 | ret = VM_FAULT_NOPAGE; | |
ba85c702 | 449 | goto out; |
656710a6 | 450 | } |
86039bd3 AA |
451 | |
452 | /* | |
453 | * Check that we can return VM_FAULT_RETRY. | |
454 | * | |
455 | * NOTE: it should become possible to return VM_FAULT_RETRY | |
456 | * even if FAULT_FLAG_TRIED is set without leading to gup() | |
457 | * -EBUSY failures, if the userfaultfd is to be extended for | |
458 | * VM_UFFD_WP tracking and we intend to arm the userfault | |
459 | * without first stopping userland access to the memory. For | |
460 | * VM_UFFD_MISSING userfaults this is enough for now. | |
461 | */ | |
82b0f8c3 | 462 | if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) { |
86039bd3 AA |
463 | /* |
464 | * Validate the invariant that nowait must allow retry | |
465 | * to be sure not to return SIGBUS erroneously on | |
466 | * nowait invocations. | |
467 | */ | |
82b0f8c3 | 468 | BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT); |
86039bd3 AA |
469 | #ifdef CONFIG_DEBUG_VM |
470 | if (printk_ratelimit()) { | |
471 | printk(KERN_WARNING | |
82b0f8c3 JK |
472 | "FAULT_FLAG_ALLOW_RETRY missing %x\n", |
473 | vmf->flags); | |
86039bd3 AA |
474 | dump_stack(); |
475 | } | |
476 | #endif | |
ba85c702 | 477 | goto out; |
86039bd3 AA |
478 | } |
479 | ||
480 | /* | |
481 | * Handle nowait, not much to do other than tell it to retry | |
482 | * and wait. | |
483 | */ | |
ba85c702 | 484 | ret = VM_FAULT_RETRY; |
82b0f8c3 | 485 | if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) |
ba85c702 | 486 | goto out; |
86039bd3 | 487 | |
c1e8d7c6 | 488 | /* take the reference before dropping the mmap_lock */ |
86039bd3 AA |
489 | userfaultfd_ctx_get(ctx); |
490 | ||
86039bd3 AA |
491 | init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function); |
492 | uwq.wq.private = current; | |
d172b1a3 NA |
493 | uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags, |
494 | reason, ctx->features); | |
86039bd3 | 495 | uwq.ctx = ctx; |
15a77c6f | 496 | uwq.waken = false; |
86039bd3 | 497 | |
3e69ad08 | 498 | blocking_state = userfaultfd_get_blocking_state(vmf->flags); |
dfa37dc3 | 499 | |
cbcfa130 | 500 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
501 | /* |
502 | * After the __add_wait_queue the uwq is visible to userland | |
503 | * through poll/read(). | |
504 | */ | |
15b726ef AA |
505 | __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq); |
506 | /* | |
507 | * The smp_mb() after __set_current_state prevents the reads | |
508 | * following the spin_unlock to happen before the list_add in | |
509 | * __add_wait_queue. | |
510 | */ | |
15a77c6f | 511 | set_current_state(blocking_state); |
cbcfa130 | 512 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 513 | |
369cd212 MK |
514 | if (!is_vm_hugetlb_page(vmf->vma)) |
515 | must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags, | |
516 | reason); | |
517 | else | |
7868a208 PA |
518 | must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma, |
519 | vmf->address, | |
369cd212 | 520 | vmf->flags, reason); |
d8ed45c5 | 521 | mmap_read_unlock(mm); |
8d2afd96 | 522 | |
f9bf3522 | 523 | if (likely(must_wait && !READ_ONCE(ctx->released))) { |
a9a08845 | 524 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
86039bd3 | 525 | schedule(); |
ba85c702 | 526 | } |
86039bd3 | 527 | |
ba85c702 | 528 | __set_current_state(TASK_RUNNING); |
15b726ef AA |
529 | |
530 | /* | |
531 | * Here we race with the list_del; list_add in | |
532 | * userfaultfd_ctx_read(), however because we don't ever run | |
533 | * list_del_init() to refile across the two lists, the prev | |
534 | * and next pointers will never point to self. list_add also | |
535 | * would never let any of the two pointers to point to | |
536 | * self. So list_empty_careful won't risk to see both pointers | |
537 | * pointing to self at any time during the list refile. The | |
538 | * only case where list_del_init() is called is the full | |
539 | * removal in the wake function and there we don't re-list_add | |
540 | * and it's fine not to block on the spinlock. The uwq on this | |
541 | * kernel stack can be released after the list_del_init. | |
542 | */ | |
2055da97 | 543 | if (!list_empty_careful(&uwq.wq.entry)) { |
cbcfa130 | 544 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
15b726ef AA |
545 | /* |
546 | * No need of list_del_init(), the uwq on the stack | |
547 | * will be freed shortly anyway. | |
548 | */ | |
2055da97 | 549 | list_del(&uwq.wq.entry); |
cbcfa130 | 550 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 551 | } |
86039bd3 AA |
552 | |
553 | /* | |
554 | * ctx may go away after this if the userfault pseudo fd is | |
555 | * already released. | |
556 | */ | |
557 | userfaultfd_ctx_put(ctx); | |
558 | ||
ba85c702 AA |
559 | out: |
560 | return ret; | |
86039bd3 AA |
561 | } |
562 | ||
8c9e7bb7 AA |
563 | static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx, |
564 | struct userfaultfd_wait_queue *ewq) | |
9cd75c3c | 565 | { |
0cbb4b4f AA |
566 | struct userfaultfd_ctx *release_new_ctx; |
567 | ||
9a69a829 AA |
568 | if (WARN_ON_ONCE(current->flags & PF_EXITING)) |
569 | goto out; | |
9cd75c3c PE |
570 | |
571 | ewq->ctx = ctx; | |
572 | init_waitqueue_entry(&ewq->wq, current); | |
0cbb4b4f | 573 | release_new_ctx = NULL; |
9cd75c3c | 574 | |
cbcfa130 | 575 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
576 | /* |
577 | * After the __add_wait_queue the uwq is visible to userland | |
578 | * through poll/read(). | |
579 | */ | |
580 | __add_wait_queue(&ctx->event_wqh, &ewq->wq); | |
581 | for (;;) { | |
582 | set_current_state(TASK_KILLABLE); | |
583 | if (ewq->msg.event == 0) | |
584 | break; | |
6aa7de05 | 585 | if (READ_ONCE(ctx->released) || |
9cd75c3c | 586 | fatal_signal_pending(current)) { |
384632e6 AA |
587 | /* |
588 | * &ewq->wq may be queued in fork_event, but | |
589 | * __remove_wait_queue ignores the head | |
590 | * parameter. It would be a problem if it | |
591 | * didn't. | |
592 | */ | |
9cd75c3c | 593 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); |
7eb76d45 MR |
594 | if (ewq->msg.event == UFFD_EVENT_FORK) { |
595 | struct userfaultfd_ctx *new; | |
596 | ||
597 | new = (struct userfaultfd_ctx *) | |
598 | (unsigned long) | |
599 | ewq->msg.arg.reserved.reserved1; | |
0cbb4b4f | 600 | release_new_ctx = new; |
7eb76d45 | 601 | } |
9cd75c3c PE |
602 | break; |
603 | } | |
604 | ||
cbcfa130 | 605 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 606 | |
a9a08845 | 607 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
9cd75c3c PE |
608 | schedule(); |
609 | ||
cbcfa130 | 610 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
611 | } |
612 | __set_current_state(TASK_RUNNING); | |
cbcfa130 | 613 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 614 | |
0cbb4b4f AA |
615 | if (release_new_ctx) { |
616 | struct vm_area_struct *vma; | |
617 | struct mm_struct *mm = release_new_ctx->mm; | |
618 | ||
619 | /* the various vma->vm_userfaultfd_ctx still points to it */ | |
d8ed45c5 | 620 | mmap_write_lock(mm); |
0cbb4b4f | 621 | for (vma = mm->mmap; vma; vma = vma->vm_next) |
31e810aa | 622 | if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) { |
0cbb4b4f | 623 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
7677f7fd | 624 | vma->vm_flags &= ~__VM_UFFD_FLAGS; |
31e810aa | 625 | } |
d8ed45c5 | 626 | mmap_write_unlock(mm); |
0cbb4b4f AA |
627 | |
628 | userfaultfd_ctx_put(release_new_ctx); | |
629 | } | |
630 | ||
9cd75c3c PE |
631 | /* |
632 | * ctx may go away after this if the userfault pseudo fd is | |
633 | * already released. | |
634 | */ | |
9a69a829 | 635 | out: |
a759a909 NA |
636 | atomic_dec(&ctx->mmap_changing); |
637 | VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0); | |
9cd75c3c | 638 | userfaultfd_ctx_put(ctx); |
9cd75c3c PE |
639 | } |
640 | ||
641 | static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx, | |
642 | struct userfaultfd_wait_queue *ewq) | |
643 | { | |
644 | ewq->msg.event = 0; | |
645 | wake_up_locked(&ctx->event_wqh); | |
646 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); | |
647 | } | |
648 | ||
893e26e6 PE |
649 | int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs) |
650 | { | |
651 | struct userfaultfd_ctx *ctx = NULL, *octx; | |
652 | struct userfaultfd_fork_ctx *fctx; | |
653 | ||
654 | octx = vma->vm_userfaultfd_ctx.ctx; | |
655 | if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) { | |
656 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
7677f7fd | 657 | vma->vm_flags &= ~__VM_UFFD_FLAGS; |
893e26e6 PE |
658 | return 0; |
659 | } | |
660 | ||
661 | list_for_each_entry(fctx, fcs, list) | |
662 | if (fctx->orig == octx) { | |
663 | ctx = fctx->new; | |
664 | break; | |
665 | } | |
666 | ||
667 | if (!ctx) { | |
668 | fctx = kmalloc(sizeof(*fctx), GFP_KERNEL); | |
669 | if (!fctx) | |
670 | return -ENOMEM; | |
671 | ||
672 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); | |
673 | if (!ctx) { | |
674 | kfree(fctx); | |
675 | return -ENOMEM; | |
676 | } | |
677 | ||
ca880420 | 678 | refcount_set(&ctx->refcount, 1); |
893e26e6 | 679 | ctx->flags = octx->flags; |
893e26e6 PE |
680 | ctx->features = octx->features; |
681 | ctx->released = false; | |
a759a909 | 682 | atomic_set(&ctx->mmap_changing, 0); |
893e26e6 | 683 | ctx->mm = vma->vm_mm; |
00bb31fa | 684 | mmgrab(ctx->mm); |
893e26e6 PE |
685 | |
686 | userfaultfd_ctx_get(octx); | |
a759a909 | 687 | atomic_inc(&octx->mmap_changing); |
893e26e6 PE |
688 | fctx->orig = octx; |
689 | fctx->new = ctx; | |
690 | list_add_tail(&fctx->list, fcs); | |
691 | } | |
692 | ||
693 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
694 | return 0; | |
695 | } | |
696 | ||
8c9e7bb7 | 697 | static void dup_fctx(struct userfaultfd_fork_ctx *fctx) |
893e26e6 PE |
698 | { |
699 | struct userfaultfd_ctx *ctx = fctx->orig; | |
700 | struct userfaultfd_wait_queue ewq; | |
701 | ||
702 | msg_init(&ewq.msg); | |
703 | ||
704 | ewq.msg.event = UFFD_EVENT_FORK; | |
705 | ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new; | |
706 | ||
8c9e7bb7 | 707 | userfaultfd_event_wait_completion(ctx, &ewq); |
893e26e6 PE |
708 | } |
709 | ||
710 | void dup_userfaultfd_complete(struct list_head *fcs) | |
711 | { | |
893e26e6 PE |
712 | struct userfaultfd_fork_ctx *fctx, *n; |
713 | ||
714 | list_for_each_entry_safe(fctx, n, fcs, list) { | |
8c9e7bb7 | 715 | dup_fctx(fctx); |
893e26e6 PE |
716 | list_del(&fctx->list); |
717 | kfree(fctx); | |
718 | } | |
719 | } | |
720 | ||
72f87654 PE |
721 | void mremap_userfaultfd_prep(struct vm_area_struct *vma, |
722 | struct vm_userfaultfd_ctx *vm_ctx) | |
723 | { | |
724 | struct userfaultfd_ctx *ctx; | |
725 | ||
726 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
3cfd22be PX |
727 | |
728 | if (!ctx) | |
729 | return; | |
730 | ||
731 | if (ctx->features & UFFD_FEATURE_EVENT_REMAP) { | |
72f87654 PE |
732 | vm_ctx->ctx = ctx; |
733 | userfaultfd_ctx_get(ctx); | |
a759a909 | 734 | atomic_inc(&ctx->mmap_changing); |
3cfd22be PX |
735 | } else { |
736 | /* Drop uffd context if remap feature not enabled */ | |
737 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
7677f7fd | 738 | vma->vm_flags &= ~__VM_UFFD_FLAGS; |
72f87654 PE |
739 | } |
740 | } | |
741 | ||
90794bf1 | 742 | void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx, |
72f87654 PE |
743 | unsigned long from, unsigned long to, |
744 | unsigned long len) | |
745 | { | |
90794bf1 | 746 | struct userfaultfd_ctx *ctx = vm_ctx->ctx; |
72f87654 PE |
747 | struct userfaultfd_wait_queue ewq; |
748 | ||
749 | if (!ctx) | |
750 | return; | |
751 | ||
752 | if (to & ~PAGE_MASK) { | |
753 | userfaultfd_ctx_put(ctx); | |
754 | return; | |
755 | } | |
756 | ||
757 | msg_init(&ewq.msg); | |
758 | ||
759 | ewq.msg.event = UFFD_EVENT_REMAP; | |
760 | ewq.msg.arg.remap.from = from; | |
761 | ewq.msg.arg.remap.to = to; | |
762 | ewq.msg.arg.remap.len = len; | |
763 | ||
764 | userfaultfd_event_wait_completion(ctx, &ewq); | |
765 | } | |
766 | ||
70ccb92f | 767 | bool userfaultfd_remove(struct vm_area_struct *vma, |
d811914d | 768 | unsigned long start, unsigned long end) |
05ce7724 PE |
769 | { |
770 | struct mm_struct *mm = vma->vm_mm; | |
771 | struct userfaultfd_ctx *ctx; | |
772 | struct userfaultfd_wait_queue ewq; | |
773 | ||
774 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
d811914d | 775 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE)) |
70ccb92f | 776 | return true; |
05ce7724 PE |
777 | |
778 | userfaultfd_ctx_get(ctx); | |
a759a909 | 779 | atomic_inc(&ctx->mmap_changing); |
d8ed45c5 | 780 | mmap_read_unlock(mm); |
05ce7724 | 781 | |
05ce7724 PE |
782 | msg_init(&ewq.msg); |
783 | ||
d811914d MR |
784 | ewq.msg.event = UFFD_EVENT_REMOVE; |
785 | ewq.msg.arg.remove.start = start; | |
786 | ewq.msg.arg.remove.end = end; | |
05ce7724 PE |
787 | |
788 | userfaultfd_event_wait_completion(ctx, &ewq); | |
789 | ||
70ccb92f | 790 | return false; |
05ce7724 PE |
791 | } |
792 | ||
897ab3e0 MR |
793 | static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps, |
794 | unsigned long start, unsigned long end) | |
795 | { | |
796 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
797 | ||
798 | list_for_each_entry(unmap_ctx, unmaps, list) | |
799 | if (unmap_ctx->ctx == ctx && unmap_ctx->start == start && | |
800 | unmap_ctx->end == end) | |
801 | return true; | |
802 | ||
803 | return false; | |
804 | } | |
805 | ||
806 | int userfaultfd_unmap_prep(struct vm_area_struct *vma, | |
807 | unsigned long start, unsigned long end, | |
808 | struct list_head *unmaps) | |
809 | { | |
810 | for ( ; vma && vma->vm_start < end; vma = vma->vm_next) { | |
811 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
812 | struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx; | |
813 | ||
814 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) || | |
815 | has_unmap_ctx(ctx, unmaps, start, end)) | |
816 | continue; | |
817 | ||
818 | unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL); | |
819 | if (!unmap_ctx) | |
820 | return -ENOMEM; | |
821 | ||
822 | userfaultfd_ctx_get(ctx); | |
a759a909 | 823 | atomic_inc(&ctx->mmap_changing); |
897ab3e0 MR |
824 | unmap_ctx->ctx = ctx; |
825 | unmap_ctx->start = start; | |
826 | unmap_ctx->end = end; | |
827 | list_add_tail(&unmap_ctx->list, unmaps); | |
828 | } | |
829 | ||
830 | return 0; | |
831 | } | |
832 | ||
833 | void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf) | |
834 | { | |
835 | struct userfaultfd_unmap_ctx *ctx, *n; | |
836 | struct userfaultfd_wait_queue ewq; | |
837 | ||
838 | list_for_each_entry_safe(ctx, n, uf, list) { | |
839 | msg_init(&ewq.msg); | |
840 | ||
841 | ewq.msg.event = UFFD_EVENT_UNMAP; | |
842 | ewq.msg.arg.remove.start = ctx->start; | |
843 | ewq.msg.arg.remove.end = ctx->end; | |
844 | ||
845 | userfaultfd_event_wait_completion(ctx->ctx, &ewq); | |
846 | ||
847 | list_del(&ctx->list); | |
848 | kfree(ctx); | |
849 | } | |
850 | } | |
851 | ||
86039bd3 AA |
852 | static int userfaultfd_release(struct inode *inode, struct file *file) |
853 | { | |
854 | struct userfaultfd_ctx *ctx = file->private_data; | |
855 | struct mm_struct *mm = ctx->mm; | |
856 | struct vm_area_struct *vma, *prev; | |
857 | /* len == 0 means wake all */ | |
858 | struct userfaultfd_wake_range range = { .len = 0, }; | |
859 | unsigned long new_flags; | |
860 | ||
6aa7de05 | 861 | WRITE_ONCE(ctx->released, true); |
86039bd3 | 862 | |
d2005e3f ON |
863 | if (!mmget_not_zero(mm)) |
864 | goto wakeup; | |
865 | ||
86039bd3 AA |
866 | /* |
867 | * Flush page faults out of all CPUs. NOTE: all page faults | |
868 | * must be retried without returning VM_FAULT_SIGBUS if | |
869 | * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx | |
c1e8d7c6 | 870 | * changes while handle_userfault released the mmap_lock. So |
86039bd3 | 871 | * it's critical that released is set to true (above), before |
c1e8d7c6 | 872 | * taking the mmap_lock for writing. |
86039bd3 | 873 | */ |
d8ed45c5 | 874 | mmap_write_lock(mm); |
86039bd3 AA |
875 | prev = NULL; |
876 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
877 | cond_resched(); | |
878 | BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 879 | !!(vma->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
880 | if (vma->vm_userfaultfd_ctx.ctx != ctx) { |
881 | prev = vma; | |
882 | continue; | |
883 | } | |
7677f7fd | 884 | new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS; |
4d45e75a JH |
885 | prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end, |
886 | new_flags, vma->anon_vma, | |
887 | vma->vm_file, vma->vm_pgoff, | |
888 | vma_policy(vma), | |
5c26f6ac | 889 | NULL_VM_UFFD_CTX, anon_vma_name(vma)); |
4d45e75a JH |
890 | if (prev) |
891 | vma = prev; | |
892 | else | |
893 | prev = vma; | |
86039bd3 AA |
894 | vma->vm_flags = new_flags; |
895 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
896 | } | |
d8ed45c5 | 897 | mmap_write_unlock(mm); |
d2005e3f ON |
898 | mmput(mm); |
899 | wakeup: | |
86039bd3 | 900 | /* |
15b726ef | 901 | * After no new page faults can wait on this fault_*wqh, flush |
86039bd3 | 902 | * the last page faults that may have been already waiting on |
15b726ef | 903 | * the fault_*wqh. |
86039bd3 | 904 | */ |
cbcfa130 | 905 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
ac5be6b4 | 906 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range); |
c430d1e8 | 907 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range); |
cbcfa130 | 908 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 909 | |
5a18b64e MR |
910 | /* Flush pending events that may still wait on event_wqh */ |
911 | wake_up_all(&ctx->event_wqh); | |
912 | ||
a9a08845 | 913 | wake_up_poll(&ctx->fd_wqh, EPOLLHUP); |
86039bd3 AA |
914 | userfaultfd_ctx_put(ctx); |
915 | return 0; | |
916 | } | |
917 | ||
15b726ef | 918 | /* fault_pending_wqh.lock must be hold by the caller */ |
6dcc27fd PE |
919 | static inline struct userfaultfd_wait_queue *find_userfault_in( |
920 | wait_queue_head_t *wqh) | |
86039bd3 | 921 | { |
ac6424b9 | 922 | wait_queue_entry_t *wq; |
15b726ef | 923 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 924 | |
456a7378 | 925 | lockdep_assert_held(&wqh->lock); |
86039bd3 | 926 | |
15b726ef | 927 | uwq = NULL; |
6dcc27fd | 928 | if (!waitqueue_active(wqh)) |
15b726ef AA |
929 | goto out; |
930 | /* walk in reverse to provide FIFO behavior to read userfaults */ | |
2055da97 | 931 | wq = list_last_entry(&wqh->head, typeof(*wq), entry); |
15b726ef AA |
932 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
933 | out: | |
934 | return uwq; | |
86039bd3 | 935 | } |
6dcc27fd PE |
936 | |
937 | static inline struct userfaultfd_wait_queue *find_userfault( | |
938 | struct userfaultfd_ctx *ctx) | |
939 | { | |
940 | return find_userfault_in(&ctx->fault_pending_wqh); | |
941 | } | |
86039bd3 | 942 | |
9cd75c3c PE |
943 | static inline struct userfaultfd_wait_queue *find_userfault_evt( |
944 | struct userfaultfd_ctx *ctx) | |
945 | { | |
946 | return find_userfault_in(&ctx->event_wqh); | |
947 | } | |
948 | ||
076ccb76 | 949 | static __poll_t userfaultfd_poll(struct file *file, poll_table *wait) |
86039bd3 AA |
950 | { |
951 | struct userfaultfd_ctx *ctx = file->private_data; | |
076ccb76 | 952 | __poll_t ret; |
86039bd3 AA |
953 | |
954 | poll_wait(file, &ctx->fd_wqh, wait); | |
955 | ||
22e5fe2a | 956 | if (!userfaultfd_is_initialized(ctx)) |
a9a08845 | 957 | return EPOLLERR; |
9cd75c3c | 958 | |
22e5fe2a NA |
959 | /* |
960 | * poll() never guarantees that read won't block. | |
961 | * userfaults can be waken before they're read(). | |
962 | */ | |
963 | if (unlikely(!(file->f_flags & O_NONBLOCK))) | |
a9a08845 | 964 | return EPOLLERR; |
22e5fe2a NA |
965 | /* |
966 | * lockless access to see if there are pending faults | |
967 | * __pollwait last action is the add_wait_queue but | |
968 | * the spin_unlock would allow the waitqueue_active to | |
969 | * pass above the actual list_add inside | |
970 | * add_wait_queue critical section. So use a full | |
971 | * memory barrier to serialize the list_add write of | |
972 | * add_wait_queue() with the waitqueue_active read | |
973 | * below. | |
974 | */ | |
975 | ret = 0; | |
976 | smp_mb(); | |
977 | if (waitqueue_active(&ctx->fault_pending_wqh)) | |
978 | ret = EPOLLIN; | |
979 | else if (waitqueue_active(&ctx->event_wqh)) | |
980 | ret = EPOLLIN; | |
981 | ||
982 | return ret; | |
86039bd3 AA |
983 | } |
984 | ||
893e26e6 PE |
985 | static const struct file_operations userfaultfd_fops; |
986 | ||
b537900f DC |
987 | static int resolve_userfault_fork(struct userfaultfd_ctx *new, |
988 | struct inode *inode, | |
893e26e6 PE |
989 | struct uffd_msg *msg) |
990 | { | |
991 | int fd; | |
893e26e6 | 992 | |
b537900f DC |
993 | fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new, |
994 | O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode); | |
893e26e6 PE |
995 | if (fd < 0) |
996 | return fd; | |
997 | ||
893e26e6 PE |
998 | msg->arg.reserved.reserved1 = 0; |
999 | msg->arg.fork.ufd = fd; | |
893e26e6 PE |
1000 | return 0; |
1001 | } | |
1002 | ||
86039bd3 | 1003 | static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait, |
b537900f | 1004 | struct uffd_msg *msg, struct inode *inode) |
86039bd3 AA |
1005 | { |
1006 | ssize_t ret; | |
1007 | DECLARE_WAITQUEUE(wait, current); | |
15b726ef | 1008 | struct userfaultfd_wait_queue *uwq; |
893e26e6 PE |
1009 | /* |
1010 | * Handling fork event requires sleeping operations, so | |
1011 | * we drop the event_wqh lock, then do these ops, then | |
1012 | * lock it back and wake up the waiter. While the lock is | |
1013 | * dropped the ewq may go away so we keep track of it | |
1014 | * carefully. | |
1015 | */ | |
1016 | LIST_HEAD(fork_event); | |
1017 | struct userfaultfd_ctx *fork_nctx = NULL; | |
86039bd3 | 1018 | |
15b726ef | 1019 | /* always take the fd_wqh lock before the fault_pending_wqh lock */ |
ae62c16e | 1020 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1021 | __add_wait_queue(&ctx->fd_wqh, &wait); |
1022 | for (;;) { | |
1023 | set_current_state(TASK_INTERRUPTIBLE); | |
15b726ef AA |
1024 | spin_lock(&ctx->fault_pending_wqh.lock); |
1025 | uwq = find_userfault(ctx); | |
1026 | if (uwq) { | |
2c5b7e1b AA |
1027 | /* |
1028 | * Use a seqcount to repeat the lockless check | |
1029 | * in wake_userfault() to avoid missing | |
1030 | * wakeups because during the refile both | |
1031 | * waitqueue could become empty if this is the | |
1032 | * only userfault. | |
1033 | */ | |
1034 | write_seqcount_begin(&ctx->refile_seq); | |
1035 | ||
86039bd3 | 1036 | /* |
15b726ef AA |
1037 | * The fault_pending_wqh.lock prevents the uwq |
1038 | * to disappear from under us. | |
1039 | * | |
1040 | * Refile this userfault from | |
1041 | * fault_pending_wqh to fault_wqh, it's not | |
1042 | * pending anymore after we read it. | |
1043 | * | |
1044 | * Use list_del() by hand (as | |
1045 | * userfaultfd_wake_function also uses | |
1046 | * list_del_init() by hand) to be sure nobody | |
1047 | * changes __remove_wait_queue() to use | |
1048 | * list_del_init() in turn breaking the | |
1049 | * !list_empty_careful() check in | |
2055da97 | 1050 | * handle_userfault(). The uwq->wq.head list |
15b726ef AA |
1051 | * must never be empty at any time during the |
1052 | * refile, or the waitqueue could disappear | |
1053 | * from under us. The "wait_queue_head_t" | |
1054 | * parameter of __remove_wait_queue() is unused | |
1055 | * anyway. | |
86039bd3 | 1056 | */ |
2055da97 | 1057 | list_del(&uwq->wq.entry); |
c430d1e8 | 1058 | add_wait_queue(&ctx->fault_wqh, &uwq->wq); |
15b726ef | 1059 | |
2c5b7e1b AA |
1060 | write_seqcount_end(&ctx->refile_seq); |
1061 | ||
a9b85f94 AA |
1062 | /* careful to always initialize msg if ret == 0 */ |
1063 | *msg = uwq->msg; | |
15b726ef | 1064 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1065 | ret = 0; |
1066 | break; | |
1067 | } | |
15b726ef | 1068 | spin_unlock(&ctx->fault_pending_wqh.lock); |
9cd75c3c PE |
1069 | |
1070 | spin_lock(&ctx->event_wqh.lock); | |
1071 | uwq = find_userfault_evt(ctx); | |
1072 | if (uwq) { | |
1073 | *msg = uwq->msg; | |
1074 | ||
893e26e6 PE |
1075 | if (uwq->msg.event == UFFD_EVENT_FORK) { |
1076 | fork_nctx = (struct userfaultfd_ctx *) | |
1077 | (unsigned long) | |
1078 | uwq->msg.arg.reserved.reserved1; | |
2055da97 | 1079 | list_move(&uwq->wq.entry, &fork_event); |
384632e6 AA |
1080 | /* |
1081 | * fork_nctx can be freed as soon as | |
1082 | * we drop the lock, unless we take a | |
1083 | * reference on it. | |
1084 | */ | |
1085 | userfaultfd_ctx_get(fork_nctx); | |
893e26e6 PE |
1086 | spin_unlock(&ctx->event_wqh.lock); |
1087 | ret = 0; | |
1088 | break; | |
1089 | } | |
1090 | ||
9cd75c3c PE |
1091 | userfaultfd_event_complete(ctx, uwq); |
1092 | spin_unlock(&ctx->event_wqh.lock); | |
1093 | ret = 0; | |
1094 | break; | |
1095 | } | |
1096 | spin_unlock(&ctx->event_wqh.lock); | |
1097 | ||
86039bd3 AA |
1098 | if (signal_pending(current)) { |
1099 | ret = -ERESTARTSYS; | |
1100 | break; | |
1101 | } | |
1102 | if (no_wait) { | |
1103 | ret = -EAGAIN; | |
1104 | break; | |
1105 | } | |
ae62c16e | 1106 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1107 | schedule(); |
ae62c16e | 1108 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1109 | } |
1110 | __remove_wait_queue(&ctx->fd_wqh, &wait); | |
1111 | __set_current_state(TASK_RUNNING); | |
ae62c16e | 1112 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1113 | |
893e26e6 | 1114 | if (!ret && msg->event == UFFD_EVENT_FORK) { |
b537900f | 1115 | ret = resolve_userfault_fork(fork_nctx, inode, msg); |
cbcfa130 | 1116 | spin_lock_irq(&ctx->event_wqh.lock); |
384632e6 AA |
1117 | if (!list_empty(&fork_event)) { |
1118 | /* | |
1119 | * The fork thread didn't abort, so we can | |
1120 | * drop the temporary refcount. | |
1121 | */ | |
1122 | userfaultfd_ctx_put(fork_nctx); | |
1123 | ||
1124 | uwq = list_first_entry(&fork_event, | |
1125 | typeof(*uwq), | |
1126 | wq.entry); | |
1127 | /* | |
1128 | * If fork_event list wasn't empty and in turn | |
1129 | * the event wasn't already released by fork | |
1130 | * (the event is allocated on fork kernel | |
1131 | * stack), put the event back to its place in | |
1132 | * the event_wq. fork_event head will be freed | |
1133 | * as soon as we return so the event cannot | |
1134 | * stay queued there no matter the current | |
1135 | * "ret" value. | |
1136 | */ | |
1137 | list_del(&uwq->wq.entry); | |
1138 | __add_wait_queue(&ctx->event_wqh, &uwq->wq); | |
893e26e6 | 1139 | |
384632e6 AA |
1140 | /* |
1141 | * Leave the event in the waitqueue and report | |
1142 | * error to userland if we failed to resolve | |
1143 | * the userfault fork. | |
1144 | */ | |
1145 | if (likely(!ret)) | |
893e26e6 | 1146 | userfaultfd_event_complete(ctx, uwq); |
384632e6 AA |
1147 | } else { |
1148 | /* | |
1149 | * Here the fork thread aborted and the | |
1150 | * refcount from the fork thread on fork_nctx | |
1151 | * has already been released. We still hold | |
1152 | * the reference we took before releasing the | |
1153 | * lock above. If resolve_userfault_fork | |
1154 | * failed we've to drop it because the | |
1155 | * fork_nctx has to be freed in such case. If | |
1156 | * it succeeded we'll hold it because the new | |
1157 | * uffd references it. | |
1158 | */ | |
1159 | if (ret) | |
1160 | userfaultfd_ctx_put(fork_nctx); | |
893e26e6 | 1161 | } |
cbcfa130 | 1162 | spin_unlock_irq(&ctx->event_wqh.lock); |
893e26e6 PE |
1163 | } |
1164 | ||
86039bd3 AA |
1165 | return ret; |
1166 | } | |
1167 | ||
1168 | static ssize_t userfaultfd_read(struct file *file, char __user *buf, | |
1169 | size_t count, loff_t *ppos) | |
1170 | { | |
1171 | struct userfaultfd_ctx *ctx = file->private_data; | |
1172 | ssize_t _ret, ret = 0; | |
a9b85f94 | 1173 | struct uffd_msg msg; |
86039bd3 | 1174 | int no_wait = file->f_flags & O_NONBLOCK; |
b537900f | 1175 | struct inode *inode = file_inode(file); |
86039bd3 | 1176 | |
22e5fe2a | 1177 | if (!userfaultfd_is_initialized(ctx)) |
86039bd3 | 1178 | return -EINVAL; |
86039bd3 AA |
1179 | |
1180 | for (;;) { | |
a9b85f94 | 1181 | if (count < sizeof(msg)) |
86039bd3 | 1182 | return ret ? ret : -EINVAL; |
b537900f | 1183 | _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode); |
86039bd3 AA |
1184 | if (_ret < 0) |
1185 | return ret ? ret : _ret; | |
a9b85f94 | 1186 | if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg))) |
86039bd3 | 1187 | return ret ? ret : -EFAULT; |
a9b85f94 AA |
1188 | ret += sizeof(msg); |
1189 | buf += sizeof(msg); | |
1190 | count -= sizeof(msg); | |
86039bd3 AA |
1191 | /* |
1192 | * Allow to read more than one fault at time but only | |
1193 | * block if waiting for the very first one. | |
1194 | */ | |
1195 | no_wait = O_NONBLOCK; | |
1196 | } | |
1197 | } | |
1198 | ||
1199 | static void __wake_userfault(struct userfaultfd_ctx *ctx, | |
1200 | struct userfaultfd_wake_range *range) | |
1201 | { | |
cbcfa130 | 1202 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 1203 | /* wake all in the range and autoremove */ |
15b726ef | 1204 | if (waitqueue_active(&ctx->fault_pending_wqh)) |
ac5be6b4 | 1205 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, |
15b726ef AA |
1206 | range); |
1207 | if (waitqueue_active(&ctx->fault_wqh)) | |
c430d1e8 | 1208 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range); |
cbcfa130 | 1209 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1210 | } |
1211 | ||
1212 | static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx, | |
1213 | struct userfaultfd_wake_range *range) | |
1214 | { | |
2c5b7e1b AA |
1215 | unsigned seq; |
1216 | bool need_wakeup; | |
1217 | ||
86039bd3 AA |
1218 | /* |
1219 | * To be sure waitqueue_active() is not reordered by the CPU | |
1220 | * before the pagetable update, use an explicit SMP memory | |
3e4e28c5 | 1221 | * barrier here. PT lock release or mmap_read_unlock(mm) still |
86039bd3 AA |
1222 | * have release semantics that can allow the |
1223 | * waitqueue_active() to be reordered before the pte update. | |
1224 | */ | |
1225 | smp_mb(); | |
1226 | ||
1227 | /* | |
1228 | * Use waitqueue_active because it's very frequent to | |
1229 | * change the address space atomically even if there are no | |
1230 | * userfaults yet. So we take the spinlock only when we're | |
1231 | * sure we've userfaults to wake. | |
1232 | */ | |
2c5b7e1b AA |
1233 | do { |
1234 | seq = read_seqcount_begin(&ctx->refile_seq); | |
1235 | need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) || | |
1236 | waitqueue_active(&ctx->fault_wqh); | |
1237 | cond_resched(); | |
1238 | } while (read_seqcount_retry(&ctx->refile_seq, seq)); | |
1239 | if (need_wakeup) | |
86039bd3 AA |
1240 | __wake_userfault(ctx, range); |
1241 | } | |
1242 | ||
1243 | static __always_inline int validate_range(struct mm_struct *mm, | |
e71e2ace | 1244 | __u64 start, __u64 len) |
86039bd3 AA |
1245 | { |
1246 | __u64 task_size = mm->task_size; | |
1247 | ||
e71e2ace | 1248 | if (start & ~PAGE_MASK) |
86039bd3 AA |
1249 | return -EINVAL; |
1250 | if (len & ~PAGE_MASK) | |
1251 | return -EINVAL; | |
1252 | if (!len) | |
1253 | return -EINVAL; | |
e71e2ace | 1254 | if (start < mmap_min_addr) |
86039bd3 | 1255 | return -EINVAL; |
e71e2ace | 1256 | if (start >= task_size) |
86039bd3 | 1257 | return -EINVAL; |
e71e2ace | 1258 | if (len > task_size - start) |
86039bd3 AA |
1259 | return -EINVAL; |
1260 | return 0; | |
1261 | } | |
1262 | ||
1263 | static int userfaultfd_register(struct userfaultfd_ctx *ctx, | |
1264 | unsigned long arg) | |
1265 | { | |
1266 | struct mm_struct *mm = ctx->mm; | |
1267 | struct vm_area_struct *vma, *prev, *cur; | |
1268 | int ret; | |
1269 | struct uffdio_register uffdio_register; | |
1270 | struct uffdio_register __user *user_uffdio_register; | |
1271 | unsigned long vm_flags, new_flags; | |
1272 | bool found; | |
ce53e8e6 | 1273 | bool basic_ioctls; |
86039bd3 AA |
1274 | unsigned long start, end, vma_end; |
1275 | ||
1276 | user_uffdio_register = (struct uffdio_register __user *) arg; | |
1277 | ||
1278 | ret = -EFAULT; | |
1279 | if (copy_from_user(&uffdio_register, user_uffdio_register, | |
1280 | sizeof(uffdio_register)-sizeof(__u64))) | |
1281 | goto out; | |
1282 | ||
1283 | ret = -EINVAL; | |
1284 | if (!uffdio_register.mode) | |
1285 | goto out; | |
7677f7fd | 1286 | if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES) |
86039bd3 AA |
1287 | goto out; |
1288 | vm_flags = 0; | |
1289 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING) | |
1290 | vm_flags |= VM_UFFD_MISSING; | |
00b151f2 PX |
1291 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) { |
1292 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP | |
1293 | goto out; | |
1294 | #endif | |
86039bd3 | 1295 | vm_flags |= VM_UFFD_WP; |
00b151f2 | 1296 | } |
7677f7fd AR |
1297 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) { |
1298 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR | |
1299 | goto out; | |
1300 | #endif | |
1301 | vm_flags |= VM_UFFD_MINOR; | |
1302 | } | |
86039bd3 | 1303 | |
e71e2ace | 1304 | ret = validate_range(mm, uffdio_register.range.start, |
86039bd3 AA |
1305 | uffdio_register.range.len); |
1306 | if (ret) | |
1307 | goto out; | |
1308 | ||
1309 | start = uffdio_register.range.start; | |
1310 | end = start + uffdio_register.range.len; | |
1311 | ||
d2005e3f ON |
1312 | ret = -ENOMEM; |
1313 | if (!mmget_not_zero(mm)) | |
1314 | goto out; | |
1315 | ||
d8ed45c5 | 1316 | mmap_write_lock(mm); |
86039bd3 | 1317 | vma = find_vma_prev(mm, start, &prev); |
86039bd3 AA |
1318 | if (!vma) |
1319 | goto out_unlock; | |
1320 | ||
1321 | /* check that there's at least one vma in the range */ | |
1322 | ret = -EINVAL; | |
1323 | if (vma->vm_start >= end) | |
1324 | goto out_unlock; | |
1325 | ||
cab350af MK |
1326 | /* |
1327 | * If the first vma contains huge pages, make sure start address | |
1328 | * is aligned to huge page size. | |
1329 | */ | |
1330 | if (is_vm_hugetlb_page(vma)) { | |
1331 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1332 | ||
1333 | if (start & (vma_hpagesize - 1)) | |
1334 | goto out_unlock; | |
1335 | } | |
1336 | ||
86039bd3 AA |
1337 | /* |
1338 | * Search for not compatible vmas. | |
86039bd3 AA |
1339 | */ |
1340 | found = false; | |
ce53e8e6 | 1341 | basic_ioctls = false; |
86039bd3 AA |
1342 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { |
1343 | cond_resched(); | |
1344 | ||
1345 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 1346 | !!(cur->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
1347 | |
1348 | /* check not compatible vmas */ | |
1349 | ret = -EINVAL; | |
63b2d417 | 1350 | if (!vma_can_userfault(cur, vm_flags)) |
86039bd3 | 1351 | goto out_unlock; |
29ec9066 AA |
1352 | |
1353 | /* | |
1354 | * UFFDIO_COPY will fill file holes even without | |
1355 | * PROT_WRITE. This check enforces that if this is a | |
1356 | * MAP_SHARED, the process has write permission to the backing | |
1357 | * file. If VM_MAYWRITE is set it also enforces that on a | |
1358 | * MAP_SHARED vma: there is no F_WRITE_SEAL and no further | |
1359 | * F_WRITE_SEAL can be taken until the vma is destroyed. | |
1360 | */ | |
1361 | ret = -EPERM; | |
1362 | if (unlikely(!(cur->vm_flags & VM_MAYWRITE))) | |
1363 | goto out_unlock; | |
1364 | ||
cab350af MK |
1365 | /* |
1366 | * If this vma contains ending address, and huge pages | |
1367 | * check alignment. | |
1368 | */ | |
1369 | if (is_vm_hugetlb_page(cur) && end <= cur->vm_end && | |
1370 | end > cur->vm_start) { | |
1371 | unsigned long vma_hpagesize = vma_kernel_pagesize(cur); | |
1372 | ||
1373 | ret = -EINVAL; | |
1374 | ||
1375 | if (end & (vma_hpagesize - 1)) | |
1376 | goto out_unlock; | |
1377 | } | |
63b2d417 AA |
1378 | if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE)) |
1379 | goto out_unlock; | |
86039bd3 AA |
1380 | |
1381 | /* | |
1382 | * Check that this vma isn't already owned by a | |
1383 | * different userfaultfd. We can't allow more than one | |
1384 | * userfaultfd to own a single vma simultaneously or we | |
1385 | * wouldn't know which one to deliver the userfaults to. | |
1386 | */ | |
1387 | ret = -EBUSY; | |
1388 | if (cur->vm_userfaultfd_ctx.ctx && | |
1389 | cur->vm_userfaultfd_ctx.ctx != ctx) | |
1390 | goto out_unlock; | |
1391 | ||
cab350af MK |
1392 | /* |
1393 | * Note vmas containing huge pages | |
1394 | */ | |
ce53e8e6 MR |
1395 | if (is_vm_hugetlb_page(cur)) |
1396 | basic_ioctls = true; | |
cab350af | 1397 | |
86039bd3 AA |
1398 | found = true; |
1399 | } | |
1400 | BUG_ON(!found); | |
1401 | ||
1402 | if (vma->vm_start < start) | |
1403 | prev = vma; | |
1404 | ||
1405 | ret = 0; | |
1406 | do { | |
1407 | cond_resched(); | |
1408 | ||
63b2d417 | 1409 | BUG_ON(!vma_can_userfault(vma, vm_flags)); |
86039bd3 AA |
1410 | BUG_ON(vma->vm_userfaultfd_ctx.ctx && |
1411 | vma->vm_userfaultfd_ctx.ctx != ctx); | |
29ec9066 | 1412 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
86039bd3 AA |
1413 | |
1414 | /* | |
1415 | * Nothing to do: this vma is already registered into this | |
1416 | * userfaultfd and with the right tracking mode too. | |
1417 | */ | |
1418 | if (vma->vm_userfaultfd_ctx.ctx == ctx && | |
1419 | (vma->vm_flags & vm_flags) == vm_flags) | |
1420 | goto skip; | |
1421 | ||
1422 | if (vma->vm_start > start) | |
1423 | start = vma->vm_start; | |
1424 | vma_end = min(end, vma->vm_end); | |
1425 | ||
7677f7fd | 1426 | new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags; |
86039bd3 AA |
1427 | prev = vma_merge(mm, prev, start, vma_end, new_flags, |
1428 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1429 | vma_policy(vma), | |
9a10064f | 1430 | ((struct vm_userfaultfd_ctx){ ctx }), |
5c26f6ac | 1431 | anon_vma_name(vma)); |
86039bd3 AA |
1432 | if (prev) { |
1433 | vma = prev; | |
1434 | goto next; | |
1435 | } | |
1436 | if (vma->vm_start < start) { | |
1437 | ret = split_vma(mm, vma, start, 1); | |
1438 | if (ret) | |
1439 | break; | |
1440 | } | |
1441 | if (vma->vm_end > end) { | |
1442 | ret = split_vma(mm, vma, end, 0); | |
1443 | if (ret) | |
1444 | break; | |
1445 | } | |
1446 | next: | |
1447 | /* | |
1448 | * In the vma_merge() successful mprotect-like case 8: | |
1449 | * the next vma was merged into the current one and | |
1450 | * the current one has not been updated yet. | |
1451 | */ | |
1452 | vma->vm_flags = new_flags; | |
1453 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
1454 | ||
6dfeaff9 PX |
1455 | if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma)) |
1456 | hugetlb_unshare_all_pmds(vma); | |
1457 | ||
86039bd3 AA |
1458 | skip: |
1459 | prev = vma; | |
1460 | start = vma->vm_end; | |
1461 | vma = vma->vm_next; | |
1462 | } while (vma && vma->vm_start < end); | |
1463 | out_unlock: | |
d8ed45c5 | 1464 | mmap_write_unlock(mm); |
d2005e3f | 1465 | mmput(mm); |
86039bd3 | 1466 | if (!ret) { |
14819305 PX |
1467 | __u64 ioctls_out; |
1468 | ||
1469 | ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC : | |
1470 | UFFD_API_RANGE_IOCTLS; | |
1471 | ||
1472 | /* | |
1473 | * Declare the WP ioctl only if the WP mode is | |
1474 | * specified and all checks passed with the range | |
1475 | */ | |
1476 | if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP)) | |
1477 | ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT); | |
1478 | ||
f6191471 AR |
1479 | /* CONTINUE ioctl is only supported for MINOR ranges. */ |
1480 | if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR)) | |
1481 | ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE); | |
1482 | ||
86039bd3 AA |
1483 | /* |
1484 | * Now that we scanned all vmas we can already tell | |
1485 | * userland which ioctls methods are guaranteed to | |
1486 | * succeed on this range. | |
1487 | */ | |
14819305 | 1488 | if (put_user(ioctls_out, &user_uffdio_register->ioctls)) |
86039bd3 AA |
1489 | ret = -EFAULT; |
1490 | } | |
1491 | out: | |
1492 | return ret; | |
1493 | } | |
1494 | ||
1495 | static int userfaultfd_unregister(struct userfaultfd_ctx *ctx, | |
1496 | unsigned long arg) | |
1497 | { | |
1498 | struct mm_struct *mm = ctx->mm; | |
1499 | struct vm_area_struct *vma, *prev, *cur; | |
1500 | int ret; | |
1501 | struct uffdio_range uffdio_unregister; | |
1502 | unsigned long new_flags; | |
1503 | bool found; | |
1504 | unsigned long start, end, vma_end; | |
1505 | const void __user *buf = (void __user *)arg; | |
1506 | ||
1507 | ret = -EFAULT; | |
1508 | if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister))) | |
1509 | goto out; | |
1510 | ||
e71e2ace | 1511 | ret = validate_range(mm, uffdio_unregister.start, |
86039bd3 AA |
1512 | uffdio_unregister.len); |
1513 | if (ret) | |
1514 | goto out; | |
1515 | ||
1516 | start = uffdio_unregister.start; | |
1517 | end = start + uffdio_unregister.len; | |
1518 | ||
d2005e3f ON |
1519 | ret = -ENOMEM; |
1520 | if (!mmget_not_zero(mm)) | |
1521 | goto out; | |
1522 | ||
d8ed45c5 | 1523 | mmap_write_lock(mm); |
86039bd3 | 1524 | vma = find_vma_prev(mm, start, &prev); |
86039bd3 AA |
1525 | if (!vma) |
1526 | goto out_unlock; | |
1527 | ||
1528 | /* check that there's at least one vma in the range */ | |
1529 | ret = -EINVAL; | |
1530 | if (vma->vm_start >= end) | |
1531 | goto out_unlock; | |
1532 | ||
cab350af MK |
1533 | /* |
1534 | * If the first vma contains huge pages, make sure start address | |
1535 | * is aligned to huge page size. | |
1536 | */ | |
1537 | if (is_vm_hugetlb_page(vma)) { | |
1538 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1539 | ||
1540 | if (start & (vma_hpagesize - 1)) | |
1541 | goto out_unlock; | |
1542 | } | |
1543 | ||
86039bd3 AA |
1544 | /* |
1545 | * Search for not compatible vmas. | |
86039bd3 AA |
1546 | */ |
1547 | found = false; | |
1548 | ret = -EINVAL; | |
1549 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { | |
1550 | cond_resched(); | |
1551 | ||
1552 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 1553 | !!(cur->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
1554 | |
1555 | /* | |
1556 | * Check not compatible vmas, not strictly required | |
1557 | * here as not compatible vmas cannot have an | |
1558 | * userfaultfd_ctx registered on them, but this | |
1559 | * provides for more strict behavior to notice | |
1560 | * unregistration errors. | |
1561 | */ | |
63b2d417 | 1562 | if (!vma_can_userfault(cur, cur->vm_flags)) |
86039bd3 AA |
1563 | goto out_unlock; |
1564 | ||
1565 | found = true; | |
1566 | } | |
1567 | BUG_ON(!found); | |
1568 | ||
1569 | if (vma->vm_start < start) | |
1570 | prev = vma; | |
1571 | ||
1572 | ret = 0; | |
1573 | do { | |
1574 | cond_resched(); | |
1575 | ||
63b2d417 | 1576 | BUG_ON(!vma_can_userfault(vma, vma->vm_flags)); |
86039bd3 AA |
1577 | |
1578 | /* | |
1579 | * Nothing to do: this vma is already registered into this | |
1580 | * userfaultfd and with the right tracking mode too. | |
1581 | */ | |
1582 | if (!vma->vm_userfaultfd_ctx.ctx) | |
1583 | goto skip; | |
1584 | ||
01e881f5 AA |
1585 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
1586 | ||
86039bd3 AA |
1587 | if (vma->vm_start > start) |
1588 | start = vma->vm_start; | |
1589 | vma_end = min(end, vma->vm_end); | |
1590 | ||
09fa5296 AA |
1591 | if (userfaultfd_missing(vma)) { |
1592 | /* | |
1593 | * Wake any concurrent pending userfault while | |
1594 | * we unregister, so they will not hang | |
1595 | * permanently and it avoids userland to call | |
1596 | * UFFDIO_WAKE explicitly. | |
1597 | */ | |
1598 | struct userfaultfd_wake_range range; | |
1599 | range.start = start; | |
1600 | range.len = vma_end - start; | |
1601 | wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range); | |
1602 | } | |
1603 | ||
f369b07c PX |
1604 | /* Reset ptes for the whole vma range if wr-protected */ |
1605 | if (userfaultfd_wp(vma)) | |
1606 | uffd_wp_range(mm, vma, start, vma_end - start, false); | |
1607 | ||
7677f7fd | 1608 | new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS; |
86039bd3 AA |
1609 | prev = vma_merge(mm, prev, start, vma_end, new_flags, |
1610 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1611 | vma_policy(vma), | |
5c26f6ac | 1612 | NULL_VM_UFFD_CTX, anon_vma_name(vma)); |
86039bd3 AA |
1613 | if (prev) { |
1614 | vma = prev; | |
1615 | goto next; | |
1616 | } | |
1617 | if (vma->vm_start < start) { | |
1618 | ret = split_vma(mm, vma, start, 1); | |
1619 | if (ret) | |
1620 | break; | |
1621 | } | |
1622 | if (vma->vm_end > end) { | |
1623 | ret = split_vma(mm, vma, end, 0); | |
1624 | if (ret) | |
1625 | break; | |
1626 | } | |
1627 | next: | |
1628 | /* | |
1629 | * In the vma_merge() successful mprotect-like case 8: | |
1630 | * the next vma was merged into the current one and | |
1631 | * the current one has not been updated yet. | |
1632 | */ | |
1633 | vma->vm_flags = new_flags; | |
1634 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
1635 | ||
1636 | skip: | |
1637 | prev = vma; | |
1638 | start = vma->vm_end; | |
1639 | vma = vma->vm_next; | |
1640 | } while (vma && vma->vm_start < end); | |
1641 | out_unlock: | |
d8ed45c5 | 1642 | mmap_write_unlock(mm); |
d2005e3f | 1643 | mmput(mm); |
86039bd3 AA |
1644 | out: |
1645 | return ret; | |
1646 | } | |
1647 | ||
1648 | /* | |
ba85c702 AA |
1649 | * userfaultfd_wake may be used in combination with the |
1650 | * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches. | |
86039bd3 AA |
1651 | */ |
1652 | static int userfaultfd_wake(struct userfaultfd_ctx *ctx, | |
1653 | unsigned long arg) | |
1654 | { | |
1655 | int ret; | |
1656 | struct uffdio_range uffdio_wake; | |
1657 | struct userfaultfd_wake_range range; | |
1658 | const void __user *buf = (void __user *)arg; | |
1659 | ||
1660 | ret = -EFAULT; | |
1661 | if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake))) | |
1662 | goto out; | |
1663 | ||
e71e2ace | 1664 | ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len); |
86039bd3 AA |
1665 | if (ret) |
1666 | goto out; | |
1667 | ||
1668 | range.start = uffdio_wake.start; | |
1669 | range.len = uffdio_wake.len; | |
1670 | ||
1671 | /* | |
1672 | * len == 0 means wake all and we don't want to wake all here, | |
1673 | * so check it again to be sure. | |
1674 | */ | |
1675 | VM_BUG_ON(!range.len); | |
1676 | ||
1677 | wake_userfault(ctx, &range); | |
1678 | ret = 0; | |
1679 | ||
1680 | out: | |
1681 | return ret; | |
1682 | } | |
1683 | ||
ad465cae AA |
1684 | static int userfaultfd_copy(struct userfaultfd_ctx *ctx, |
1685 | unsigned long arg) | |
1686 | { | |
1687 | __s64 ret; | |
1688 | struct uffdio_copy uffdio_copy; | |
1689 | struct uffdio_copy __user *user_uffdio_copy; | |
1690 | struct userfaultfd_wake_range range; | |
1691 | ||
1692 | user_uffdio_copy = (struct uffdio_copy __user *) arg; | |
1693 | ||
df2cc96e | 1694 | ret = -EAGAIN; |
a759a909 | 1695 | if (atomic_read(&ctx->mmap_changing)) |
df2cc96e MR |
1696 | goto out; |
1697 | ||
ad465cae AA |
1698 | ret = -EFAULT; |
1699 | if (copy_from_user(&uffdio_copy, user_uffdio_copy, | |
1700 | /* don't copy "copy" last field */ | |
1701 | sizeof(uffdio_copy)-sizeof(__s64))) | |
1702 | goto out; | |
1703 | ||
e71e2ace | 1704 | ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len); |
ad465cae AA |
1705 | if (ret) |
1706 | goto out; | |
1707 | /* | |
1708 | * double check for wraparound just in case. copy_from_user() | |
1709 | * will later check uffdio_copy.src + uffdio_copy.len to fit | |
1710 | * in the userland range. | |
1711 | */ | |
1712 | ret = -EINVAL; | |
1713 | if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src) | |
1714 | goto out; | |
72981e0e | 1715 | if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP)) |
ad465cae | 1716 | goto out; |
d2005e3f ON |
1717 | if (mmget_not_zero(ctx->mm)) { |
1718 | ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src, | |
72981e0e AA |
1719 | uffdio_copy.len, &ctx->mmap_changing, |
1720 | uffdio_copy.mode); | |
d2005e3f | 1721 | mmput(ctx->mm); |
96333187 | 1722 | } else { |
e86b298b | 1723 | return -ESRCH; |
d2005e3f | 1724 | } |
ad465cae AA |
1725 | if (unlikely(put_user(ret, &user_uffdio_copy->copy))) |
1726 | return -EFAULT; | |
1727 | if (ret < 0) | |
1728 | goto out; | |
1729 | BUG_ON(!ret); | |
1730 | /* len == 0 would wake all */ | |
1731 | range.len = ret; | |
1732 | if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) { | |
1733 | range.start = uffdio_copy.dst; | |
1734 | wake_userfault(ctx, &range); | |
1735 | } | |
1736 | ret = range.len == uffdio_copy.len ? 0 : -EAGAIN; | |
1737 | out: | |
1738 | return ret; | |
1739 | } | |
1740 | ||
1741 | static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx, | |
1742 | unsigned long arg) | |
1743 | { | |
1744 | __s64 ret; | |
1745 | struct uffdio_zeropage uffdio_zeropage; | |
1746 | struct uffdio_zeropage __user *user_uffdio_zeropage; | |
1747 | struct userfaultfd_wake_range range; | |
1748 | ||
1749 | user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg; | |
1750 | ||
df2cc96e | 1751 | ret = -EAGAIN; |
a759a909 | 1752 | if (atomic_read(&ctx->mmap_changing)) |
df2cc96e MR |
1753 | goto out; |
1754 | ||
ad465cae AA |
1755 | ret = -EFAULT; |
1756 | if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage, | |
1757 | /* don't copy "zeropage" last field */ | |
1758 | sizeof(uffdio_zeropage)-sizeof(__s64))) | |
1759 | goto out; | |
1760 | ||
e71e2ace | 1761 | ret = validate_range(ctx->mm, uffdio_zeropage.range.start, |
ad465cae AA |
1762 | uffdio_zeropage.range.len); |
1763 | if (ret) | |
1764 | goto out; | |
1765 | ret = -EINVAL; | |
1766 | if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE) | |
1767 | goto out; | |
1768 | ||
d2005e3f ON |
1769 | if (mmget_not_zero(ctx->mm)) { |
1770 | ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start, | |
df2cc96e MR |
1771 | uffdio_zeropage.range.len, |
1772 | &ctx->mmap_changing); | |
d2005e3f | 1773 | mmput(ctx->mm); |
9d95aa4b | 1774 | } else { |
e86b298b | 1775 | return -ESRCH; |
d2005e3f | 1776 | } |
ad465cae AA |
1777 | if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage))) |
1778 | return -EFAULT; | |
1779 | if (ret < 0) | |
1780 | goto out; | |
1781 | /* len == 0 would wake all */ | |
1782 | BUG_ON(!ret); | |
1783 | range.len = ret; | |
1784 | if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) { | |
1785 | range.start = uffdio_zeropage.range.start; | |
1786 | wake_userfault(ctx, &range); | |
1787 | } | |
1788 | ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN; | |
1789 | out: | |
1790 | return ret; | |
1791 | } | |
1792 | ||
63b2d417 AA |
1793 | static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx, |
1794 | unsigned long arg) | |
1795 | { | |
1796 | int ret; | |
1797 | struct uffdio_writeprotect uffdio_wp; | |
1798 | struct uffdio_writeprotect __user *user_uffdio_wp; | |
1799 | struct userfaultfd_wake_range range; | |
23080e27 | 1800 | bool mode_wp, mode_dontwake; |
63b2d417 | 1801 | |
a759a909 | 1802 | if (atomic_read(&ctx->mmap_changing)) |
63b2d417 AA |
1803 | return -EAGAIN; |
1804 | ||
1805 | user_uffdio_wp = (struct uffdio_writeprotect __user *) arg; | |
1806 | ||
1807 | if (copy_from_user(&uffdio_wp, user_uffdio_wp, | |
1808 | sizeof(struct uffdio_writeprotect))) | |
1809 | return -EFAULT; | |
1810 | ||
e71e2ace | 1811 | ret = validate_range(ctx->mm, uffdio_wp.range.start, |
63b2d417 AA |
1812 | uffdio_wp.range.len); |
1813 | if (ret) | |
1814 | return ret; | |
1815 | ||
1816 | if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE | | |
1817 | UFFDIO_WRITEPROTECT_MODE_WP)) | |
1818 | return -EINVAL; | |
23080e27 PX |
1819 | |
1820 | mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP; | |
1821 | mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE; | |
1822 | ||
1823 | if (mode_wp && mode_dontwake) | |
63b2d417 AA |
1824 | return -EINVAL; |
1825 | ||
cb185d5f NA |
1826 | if (mmget_not_zero(ctx->mm)) { |
1827 | ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start, | |
1828 | uffdio_wp.range.len, mode_wp, | |
1829 | &ctx->mmap_changing); | |
1830 | mmput(ctx->mm); | |
1831 | } else { | |
1832 | return -ESRCH; | |
1833 | } | |
1834 | ||
63b2d417 AA |
1835 | if (ret) |
1836 | return ret; | |
1837 | ||
23080e27 | 1838 | if (!mode_wp && !mode_dontwake) { |
63b2d417 AA |
1839 | range.start = uffdio_wp.range.start; |
1840 | range.len = uffdio_wp.range.len; | |
1841 | wake_userfault(ctx, &range); | |
1842 | } | |
1843 | return ret; | |
1844 | } | |
1845 | ||
f6191471 AR |
1846 | static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg) |
1847 | { | |
1848 | __s64 ret; | |
1849 | struct uffdio_continue uffdio_continue; | |
1850 | struct uffdio_continue __user *user_uffdio_continue; | |
1851 | struct userfaultfd_wake_range range; | |
1852 | ||
1853 | user_uffdio_continue = (struct uffdio_continue __user *)arg; | |
1854 | ||
1855 | ret = -EAGAIN; | |
a759a909 | 1856 | if (atomic_read(&ctx->mmap_changing)) |
f6191471 AR |
1857 | goto out; |
1858 | ||
1859 | ret = -EFAULT; | |
1860 | if (copy_from_user(&uffdio_continue, user_uffdio_continue, | |
1861 | /* don't copy the output fields */ | |
1862 | sizeof(uffdio_continue) - (sizeof(__s64)))) | |
1863 | goto out; | |
1864 | ||
e71e2ace | 1865 | ret = validate_range(ctx->mm, uffdio_continue.range.start, |
f6191471 AR |
1866 | uffdio_continue.range.len); |
1867 | if (ret) | |
1868 | goto out; | |
1869 | ||
1870 | ret = -EINVAL; | |
1871 | /* double check for wraparound just in case. */ | |
1872 | if (uffdio_continue.range.start + uffdio_continue.range.len <= | |
1873 | uffdio_continue.range.start) { | |
1874 | goto out; | |
1875 | } | |
1876 | if (uffdio_continue.mode & ~UFFDIO_CONTINUE_MODE_DONTWAKE) | |
1877 | goto out; | |
1878 | ||
1879 | if (mmget_not_zero(ctx->mm)) { | |
1880 | ret = mcopy_continue(ctx->mm, uffdio_continue.range.start, | |
1881 | uffdio_continue.range.len, | |
1882 | &ctx->mmap_changing); | |
1883 | mmput(ctx->mm); | |
1884 | } else { | |
1885 | return -ESRCH; | |
1886 | } | |
1887 | ||
1888 | if (unlikely(put_user(ret, &user_uffdio_continue->mapped))) | |
1889 | return -EFAULT; | |
1890 | if (ret < 0) | |
1891 | goto out; | |
1892 | ||
1893 | /* len == 0 would wake all */ | |
1894 | BUG_ON(!ret); | |
1895 | range.len = ret; | |
1896 | if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) { | |
1897 | range.start = uffdio_continue.range.start; | |
1898 | wake_userfault(ctx, &range); | |
1899 | } | |
1900 | ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN; | |
1901 | ||
1902 | out: | |
1903 | return ret; | |
1904 | } | |
1905 | ||
9cd75c3c PE |
1906 | static inline unsigned int uffd_ctx_features(__u64 user_features) |
1907 | { | |
1908 | /* | |
22e5fe2a NA |
1909 | * For the current set of features the bits just coincide. Set |
1910 | * UFFD_FEATURE_INITIALIZED to mark the features as enabled. | |
9cd75c3c | 1911 | */ |
22e5fe2a | 1912 | return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED; |
9cd75c3c PE |
1913 | } |
1914 | ||
86039bd3 AA |
1915 | /* |
1916 | * userland asks for a certain API version and we return which bits | |
1917 | * and ioctl commands are implemented in this kernel for such API | |
1918 | * version or -EINVAL if unknown. | |
1919 | */ | |
1920 | static int userfaultfd_api(struct userfaultfd_ctx *ctx, | |
1921 | unsigned long arg) | |
1922 | { | |
1923 | struct uffdio_api uffdio_api; | |
1924 | void __user *buf = (void __user *)arg; | |
22e5fe2a | 1925 | unsigned int ctx_features; |
86039bd3 | 1926 | int ret; |
65603144 | 1927 | __u64 features; |
86039bd3 | 1928 | |
86039bd3 | 1929 | ret = -EFAULT; |
a9b85f94 | 1930 | if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api))) |
86039bd3 | 1931 | goto out; |
914eedcb AR |
1932 | /* Ignore unsupported features (userspace built against newer kernel) */ |
1933 | features = uffdio_api.features & UFFD_API_FEATURES; | |
3c1c24d9 MR |
1934 | ret = -EPERM; |
1935 | if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE)) | |
1936 | goto err_out; | |
65603144 AA |
1937 | /* report all available features and ioctls to userland */ |
1938 | uffdio_api.features = UFFD_API_FEATURES; | |
7677f7fd | 1939 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR |
964ab004 AR |
1940 | uffdio_api.features &= |
1941 | ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM); | |
00b151f2 PX |
1942 | #endif |
1943 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP | |
1944 | uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP; | |
b1f9e876 PX |
1945 | #endif |
1946 | #ifndef CONFIG_PTE_MARKER_UFFD_WP | |
1947 | uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM; | |
7677f7fd | 1948 | #endif |
86039bd3 AA |
1949 | uffdio_api.ioctls = UFFD_API_IOCTLS; |
1950 | ret = -EFAULT; | |
1951 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1952 | goto out; | |
22e5fe2a | 1953 | |
65603144 | 1954 | /* only enable the requested features for this uffd context */ |
22e5fe2a NA |
1955 | ctx_features = uffd_ctx_features(features); |
1956 | ret = -EINVAL; | |
1957 | if (cmpxchg(&ctx->features, 0, ctx_features) != 0) | |
1958 | goto err_out; | |
1959 | ||
86039bd3 AA |
1960 | ret = 0; |
1961 | out: | |
1962 | return ret; | |
3c1c24d9 MR |
1963 | err_out: |
1964 | memset(&uffdio_api, 0, sizeof(uffdio_api)); | |
1965 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1966 | ret = -EFAULT; | |
1967 | goto out; | |
86039bd3 AA |
1968 | } |
1969 | ||
1970 | static long userfaultfd_ioctl(struct file *file, unsigned cmd, | |
1971 | unsigned long arg) | |
1972 | { | |
1973 | int ret = -EINVAL; | |
1974 | struct userfaultfd_ctx *ctx = file->private_data; | |
1975 | ||
22e5fe2a | 1976 | if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx)) |
e6485a47 AA |
1977 | return -EINVAL; |
1978 | ||
86039bd3 AA |
1979 | switch(cmd) { |
1980 | case UFFDIO_API: | |
1981 | ret = userfaultfd_api(ctx, arg); | |
1982 | break; | |
1983 | case UFFDIO_REGISTER: | |
1984 | ret = userfaultfd_register(ctx, arg); | |
1985 | break; | |
1986 | case UFFDIO_UNREGISTER: | |
1987 | ret = userfaultfd_unregister(ctx, arg); | |
1988 | break; | |
1989 | case UFFDIO_WAKE: | |
1990 | ret = userfaultfd_wake(ctx, arg); | |
1991 | break; | |
ad465cae AA |
1992 | case UFFDIO_COPY: |
1993 | ret = userfaultfd_copy(ctx, arg); | |
1994 | break; | |
1995 | case UFFDIO_ZEROPAGE: | |
1996 | ret = userfaultfd_zeropage(ctx, arg); | |
1997 | break; | |
63b2d417 AA |
1998 | case UFFDIO_WRITEPROTECT: |
1999 | ret = userfaultfd_writeprotect(ctx, arg); | |
2000 | break; | |
f6191471 AR |
2001 | case UFFDIO_CONTINUE: |
2002 | ret = userfaultfd_continue(ctx, arg); | |
2003 | break; | |
86039bd3 AA |
2004 | } |
2005 | return ret; | |
2006 | } | |
2007 | ||
2008 | #ifdef CONFIG_PROC_FS | |
2009 | static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f) | |
2010 | { | |
2011 | struct userfaultfd_ctx *ctx = f->private_data; | |
ac6424b9 | 2012 | wait_queue_entry_t *wq; |
86039bd3 AA |
2013 | unsigned long pending = 0, total = 0; |
2014 | ||
cbcfa130 | 2015 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
2055da97 | 2016 | list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) { |
15b726ef AA |
2017 | pending++; |
2018 | total++; | |
2019 | } | |
2055da97 | 2020 | list_for_each_entry(wq, &ctx->fault_wqh.head, entry) { |
86039bd3 AA |
2021 | total++; |
2022 | } | |
cbcfa130 | 2023 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
2024 | |
2025 | /* | |
2026 | * If more protocols will be added, there will be all shown | |
2027 | * separated by a space. Like this: | |
2028 | * protocols: aa:... bb:... | |
2029 | */ | |
2030 | seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n", | |
045098e9 | 2031 | pending, total, UFFD_API, ctx->features, |
86039bd3 AA |
2032 | UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS); |
2033 | } | |
2034 | #endif | |
2035 | ||
2036 | static const struct file_operations userfaultfd_fops = { | |
2037 | #ifdef CONFIG_PROC_FS | |
2038 | .show_fdinfo = userfaultfd_show_fdinfo, | |
2039 | #endif | |
2040 | .release = userfaultfd_release, | |
2041 | .poll = userfaultfd_poll, | |
2042 | .read = userfaultfd_read, | |
2043 | .unlocked_ioctl = userfaultfd_ioctl, | |
1832f2d8 | 2044 | .compat_ioctl = compat_ptr_ioctl, |
86039bd3 AA |
2045 | .llseek = noop_llseek, |
2046 | }; | |
2047 | ||
3004ec9c AA |
2048 | static void init_once_userfaultfd_ctx(void *mem) |
2049 | { | |
2050 | struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem; | |
2051 | ||
2052 | init_waitqueue_head(&ctx->fault_pending_wqh); | |
2053 | init_waitqueue_head(&ctx->fault_wqh); | |
9cd75c3c | 2054 | init_waitqueue_head(&ctx->event_wqh); |
3004ec9c | 2055 | init_waitqueue_head(&ctx->fd_wqh); |
2ca97ac8 | 2056 | seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock); |
3004ec9c AA |
2057 | } |
2058 | ||
284cd241 | 2059 | SYSCALL_DEFINE1(userfaultfd, int, flags) |
86039bd3 | 2060 | { |
86039bd3 | 2061 | struct userfaultfd_ctx *ctx; |
284cd241 | 2062 | int fd; |
86039bd3 | 2063 | |
d0d4730a LG |
2064 | if (!sysctl_unprivileged_userfaultfd && |
2065 | (flags & UFFD_USER_MODE_ONLY) == 0 && | |
2066 | !capable(CAP_SYS_PTRACE)) { | |
2067 | printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd " | |
2068 | "sysctl knob to 1 if kernel faults must be handled " | |
2069 | "without obtaining CAP_SYS_PTRACE capability\n"); | |
cefdca0a | 2070 | return -EPERM; |
d0d4730a | 2071 | } |
cefdca0a | 2072 | |
86039bd3 AA |
2073 | BUG_ON(!current->mm); |
2074 | ||
2075 | /* Check the UFFD_* constants for consistency. */ | |
37cd0575 | 2076 | BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS); |
86039bd3 AA |
2077 | BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC); |
2078 | BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK); | |
2079 | ||
37cd0575 | 2080 | if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY)) |
284cd241 | 2081 | return -EINVAL; |
86039bd3 | 2082 | |
3004ec9c | 2083 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); |
86039bd3 | 2084 | if (!ctx) |
284cd241 | 2085 | return -ENOMEM; |
86039bd3 | 2086 | |
ca880420 | 2087 | refcount_set(&ctx->refcount, 1); |
86039bd3 | 2088 | ctx->flags = flags; |
9cd75c3c | 2089 | ctx->features = 0; |
86039bd3 | 2090 | ctx->released = false; |
a759a909 | 2091 | atomic_set(&ctx->mmap_changing, 0); |
86039bd3 AA |
2092 | ctx->mm = current->mm; |
2093 | /* prevent the mm struct to be freed */ | |
f1f10076 | 2094 | mmgrab(ctx->mm); |
86039bd3 | 2095 | |
b537900f DC |
2096 | fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx, |
2097 | O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL); | |
284cd241 | 2098 | if (fd < 0) { |
d2005e3f | 2099 | mmdrop(ctx->mm); |
3004ec9c | 2100 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
c03e946f | 2101 | } |
86039bd3 | 2102 | return fd; |
86039bd3 | 2103 | } |
3004ec9c AA |
2104 | |
2105 | static int __init userfaultfd_init(void) | |
2106 | { | |
2107 | userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache", | |
2108 | sizeof(struct userfaultfd_ctx), | |
2109 | 0, | |
2110 | SLAB_HWCACHE_ALIGN|SLAB_PANIC, | |
2111 | init_once_userfaultfd_ctx); | |
2112 | return 0; | |
2113 | } | |
2114 | __initcall(userfaultfd_init); |