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> |
2d5de004 | 33 | #include <linux/miscdevice.h> |
86039bd3 | 34 | |
d0d4730a | 35 | int sysctl_unprivileged_userfaultfd __read_mostly; |
cefdca0a | 36 | |
3004ec9c AA |
37 | static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly; |
38 | ||
3004ec9c AA |
39 | /* |
40 | * Start with fault_pending_wqh and fault_wqh so they're more likely | |
41 | * to be in the same cacheline. | |
cbcfa130 EB |
42 | * |
43 | * Locking order: | |
44 | * fd_wqh.lock | |
45 | * fault_pending_wqh.lock | |
46 | * fault_wqh.lock | |
47 | * event_wqh.lock | |
48 | * | |
49 | * To avoid deadlocks, IRQs must be disabled when taking any of the above locks, | |
50 | * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's | |
51 | * also taken in IRQ context. | |
3004ec9c | 52 | */ |
86039bd3 | 53 | struct userfaultfd_ctx { |
15b726ef AA |
54 | /* waitqueue head for the pending (i.e. not read) userfaults */ |
55 | wait_queue_head_t fault_pending_wqh; | |
56 | /* waitqueue head for the userfaults */ | |
86039bd3 AA |
57 | wait_queue_head_t fault_wqh; |
58 | /* waitqueue head for the pseudo fd to wakeup poll/read */ | |
59 | wait_queue_head_t fd_wqh; | |
9cd75c3c PE |
60 | /* waitqueue head for events */ |
61 | wait_queue_head_t event_wqh; | |
2c5b7e1b | 62 | /* a refile sequence protected by fault_pending_wqh lock */ |
2ca97ac8 | 63 | seqcount_spinlock_t refile_seq; |
3004ec9c | 64 | /* pseudo fd refcounting */ |
ca880420 | 65 | refcount_t refcount; |
86039bd3 AA |
66 | /* userfaultfd syscall flags */ |
67 | unsigned int flags; | |
9cd75c3c PE |
68 | /* features requested from the userspace */ |
69 | unsigned int features; | |
86039bd3 AA |
70 | /* released */ |
71 | bool released; | |
df2cc96e | 72 | /* memory mappings are changing because of non-cooperative event */ |
a759a909 | 73 | atomic_t mmap_changing; |
86039bd3 AA |
74 | /* mm with one ore more vmas attached to this userfaultfd_ctx */ |
75 | struct mm_struct *mm; | |
76 | }; | |
77 | ||
893e26e6 PE |
78 | struct userfaultfd_fork_ctx { |
79 | struct userfaultfd_ctx *orig; | |
80 | struct userfaultfd_ctx *new; | |
81 | struct list_head list; | |
82 | }; | |
83 | ||
897ab3e0 MR |
84 | struct userfaultfd_unmap_ctx { |
85 | struct userfaultfd_ctx *ctx; | |
86 | unsigned long start; | |
87 | unsigned long end; | |
88 | struct list_head list; | |
89 | }; | |
90 | ||
86039bd3 | 91 | struct userfaultfd_wait_queue { |
a9b85f94 | 92 | struct uffd_msg msg; |
ac6424b9 | 93 | wait_queue_entry_t wq; |
86039bd3 | 94 | struct userfaultfd_ctx *ctx; |
15a77c6f | 95 | bool waken; |
86039bd3 AA |
96 | }; |
97 | ||
98 | struct userfaultfd_wake_range { | |
99 | unsigned long start; | |
100 | unsigned long len; | |
101 | }; | |
102 | ||
22e5fe2a NA |
103 | /* internal indication that UFFD_API ioctl was successfully executed */ |
104 | #define UFFD_FEATURE_INITIALIZED (1u << 31) | |
105 | ||
106 | static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx) | |
107 | { | |
108 | return ctx->features & UFFD_FEATURE_INITIALIZED; | |
109 | } | |
110 | ||
2bad466c PX |
111 | /* |
112 | * Whether WP_UNPOPULATED is enabled on the uffd context. It is only | |
113 | * meaningful when userfaultfd_wp()==true on the vma and when it's | |
114 | * anonymous. | |
115 | */ | |
116 | bool userfaultfd_wp_unpopulated(struct vm_area_struct *vma) | |
117 | { | |
118 | struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx; | |
119 | ||
120 | if (!ctx) | |
121 | return false; | |
122 | ||
123 | return ctx->features & UFFD_FEATURE_WP_UNPOPULATED; | |
124 | } | |
125 | ||
51d3d5eb DH |
126 | static void userfaultfd_set_vm_flags(struct vm_area_struct *vma, |
127 | vm_flags_t flags) | |
128 | { | |
129 | const bool uffd_wp_changed = (vma->vm_flags ^ flags) & VM_UFFD_WP; | |
130 | ||
1c71222e | 131 | vm_flags_reset(vma, flags); |
51d3d5eb DH |
132 | /* |
133 | * For shared mappings, we want to enable writenotify while | |
134 | * userfaultfd-wp is enabled (see vma_wants_writenotify()). We'll simply | |
135 | * recalculate vma->vm_page_prot whenever userfaultfd-wp changes. | |
136 | */ | |
137 | if ((vma->vm_flags & VM_SHARED) && uffd_wp_changed) | |
138 | vma_set_page_prot(vma); | |
139 | } | |
140 | ||
ac6424b9 | 141 | static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode, |
86039bd3 AA |
142 | int wake_flags, void *key) |
143 | { | |
144 | struct userfaultfd_wake_range *range = key; | |
145 | int ret; | |
146 | struct userfaultfd_wait_queue *uwq; | |
147 | unsigned long start, len; | |
148 | ||
149 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
150 | ret = 0; | |
86039bd3 AA |
151 | /* len == 0 means wake all */ |
152 | start = range->start; | |
153 | len = range->len; | |
a9b85f94 AA |
154 | if (len && (start > uwq->msg.arg.pagefault.address || |
155 | start + len <= uwq->msg.arg.pagefault.address)) | |
86039bd3 | 156 | goto out; |
15a77c6f AA |
157 | WRITE_ONCE(uwq->waken, true); |
158 | /* | |
a9668cd6 PZ |
159 | * The Program-Order guarantees provided by the scheduler |
160 | * ensure uwq->waken is visible before the task is woken. | |
15a77c6f | 161 | */ |
86039bd3 | 162 | ret = wake_up_state(wq->private, mode); |
a9668cd6 | 163 | if (ret) { |
86039bd3 AA |
164 | /* |
165 | * Wake only once, autoremove behavior. | |
166 | * | |
a9668cd6 PZ |
167 | * After the effect of list_del_init is visible to the other |
168 | * CPUs, the waitqueue may disappear from under us, see the | |
169 | * !list_empty_careful() in handle_userfault(). | |
170 | * | |
171 | * try_to_wake_up() has an implicit smp_mb(), and the | |
172 | * wq->private is read before calling the extern function | |
173 | * "wake_up_state" (which in turns calls try_to_wake_up). | |
86039bd3 | 174 | */ |
2055da97 | 175 | list_del_init(&wq->entry); |
a9668cd6 | 176 | } |
86039bd3 AA |
177 | out: |
178 | return ret; | |
179 | } | |
180 | ||
181 | /** | |
182 | * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd | |
183 | * context. | |
184 | * @ctx: [in] Pointer to the userfaultfd context. | |
86039bd3 AA |
185 | */ |
186 | static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx) | |
187 | { | |
ca880420 | 188 | refcount_inc(&ctx->refcount); |
86039bd3 AA |
189 | } |
190 | ||
191 | /** | |
192 | * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd | |
193 | * context. | |
194 | * @ctx: [in] Pointer to userfaultfd context. | |
195 | * | |
196 | * The userfaultfd context reference must have been previously acquired either | |
197 | * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget(). | |
198 | */ | |
199 | static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx) | |
200 | { | |
ca880420 | 201 | if (refcount_dec_and_test(&ctx->refcount)) { |
86039bd3 AA |
202 | VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock)); |
203 | VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh)); | |
204 | VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock)); | |
205 | VM_BUG_ON(waitqueue_active(&ctx->fault_wqh)); | |
9cd75c3c PE |
206 | VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock)); |
207 | VM_BUG_ON(waitqueue_active(&ctx->event_wqh)); | |
86039bd3 AA |
208 | VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock)); |
209 | VM_BUG_ON(waitqueue_active(&ctx->fd_wqh)); | |
d2005e3f | 210 | mmdrop(ctx->mm); |
3004ec9c | 211 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
86039bd3 AA |
212 | } |
213 | } | |
214 | ||
a9b85f94 | 215 | static inline void msg_init(struct uffd_msg *msg) |
86039bd3 | 216 | { |
a9b85f94 AA |
217 | BUILD_BUG_ON(sizeof(struct uffd_msg) != 32); |
218 | /* | |
219 | * Must use memset to zero out the paddings or kernel data is | |
220 | * leaked to userland. | |
221 | */ | |
222 | memset(msg, 0, sizeof(struct uffd_msg)); | |
223 | } | |
224 | ||
225 | static inline struct uffd_msg userfault_msg(unsigned long address, | |
d172b1a3 | 226 | unsigned long real_address, |
a9b85f94 | 227 | unsigned int flags, |
9d4ac934 AP |
228 | unsigned long reason, |
229 | unsigned int features) | |
a9b85f94 AA |
230 | { |
231 | struct uffd_msg msg; | |
d172b1a3 | 232 | |
a9b85f94 AA |
233 | msg_init(&msg); |
234 | msg.event = UFFD_EVENT_PAGEFAULT; | |
824ddc60 | 235 | |
d172b1a3 NA |
236 | msg.arg.pagefault.address = (features & UFFD_FEATURE_EXACT_ADDRESS) ? |
237 | real_address : address; | |
238 | ||
7677f7fd AR |
239 | /* |
240 | * These flags indicate why the userfault occurred: | |
241 | * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault. | |
242 | * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault. | |
243 | * - Neither of these flags being set indicates a MISSING fault. | |
244 | * | |
245 | * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write | |
246 | * fault. Otherwise, it was a read fault. | |
247 | */ | |
86039bd3 | 248 | if (flags & FAULT_FLAG_WRITE) |
a9b85f94 | 249 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE; |
86039bd3 | 250 | if (reason & VM_UFFD_WP) |
a9b85f94 | 251 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP; |
7677f7fd AR |
252 | if (reason & VM_UFFD_MINOR) |
253 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR; | |
9d4ac934 | 254 | if (features & UFFD_FEATURE_THREAD_ID) |
a36985d3 | 255 | msg.arg.pagefault.feat.ptid = task_pid_vnr(current); |
a9b85f94 | 256 | return msg; |
86039bd3 AA |
257 | } |
258 | ||
369cd212 MK |
259 | #ifdef CONFIG_HUGETLB_PAGE |
260 | /* | |
261 | * Same functionality as userfaultfd_must_wait below with modifications for | |
262 | * hugepmd ranges. | |
263 | */ | |
264 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 265 | struct vm_area_struct *vma, |
369cd212 MK |
266 | unsigned long address, |
267 | unsigned long flags, | |
268 | unsigned long reason) | |
269 | { | |
1e2c0436 | 270 | pte_t *ptep, pte; |
369cd212 MK |
271 | bool ret = true; |
272 | ||
9c67a207 | 273 | mmap_assert_locked(ctx->mm); |
1e2c0436 | 274 | |
9c67a207 | 275 | ptep = hugetlb_walk(vma, address, vma_mmu_pagesize(vma)); |
1e2c0436 | 276 | if (!ptep) |
369cd212 MK |
277 | goto out; |
278 | ||
279 | ret = false; | |
1e2c0436 | 280 | pte = huge_ptep_get(ptep); |
369cd212 MK |
281 | |
282 | /* | |
283 | * Lockless access: we're in a wait_event so it's ok if it | |
5c041f5d PX |
284 | * changes under us. PTE markers should be handled the same as none |
285 | * ptes here. | |
369cd212 | 286 | */ |
5c041f5d | 287 | if (huge_pte_none_mostly(pte)) |
369cd212 | 288 | ret = true; |
1e2c0436 | 289 | if (!huge_pte_write(pte) && (reason & VM_UFFD_WP)) |
369cd212 MK |
290 | ret = true; |
291 | out: | |
292 | return ret; | |
293 | } | |
294 | #else | |
295 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 296 | struct vm_area_struct *vma, |
369cd212 MK |
297 | unsigned long address, |
298 | unsigned long flags, | |
299 | unsigned long reason) | |
300 | { | |
301 | return false; /* should never get here */ | |
302 | } | |
303 | #endif /* CONFIG_HUGETLB_PAGE */ | |
304 | ||
8d2afd96 AA |
305 | /* |
306 | * Verify the pagetables are still not ok after having reigstered into | |
307 | * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any | |
308 | * userfault that has already been resolved, if userfaultfd_read and | |
309 | * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different | |
310 | * threads. | |
311 | */ | |
312 | static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx, | |
313 | unsigned long address, | |
314 | unsigned long flags, | |
315 | unsigned long reason) | |
316 | { | |
317 | struct mm_struct *mm = ctx->mm; | |
318 | pgd_t *pgd; | |
c2febafc | 319 | p4d_t *p4d; |
8d2afd96 AA |
320 | pud_t *pud; |
321 | pmd_t *pmd, _pmd; | |
322 | pte_t *pte; | |
323 | bool ret = true; | |
324 | ||
42fc5414 | 325 | mmap_assert_locked(mm); |
8d2afd96 AA |
326 | |
327 | pgd = pgd_offset(mm, address); | |
328 | if (!pgd_present(*pgd)) | |
329 | goto out; | |
c2febafc KS |
330 | p4d = p4d_offset(pgd, address); |
331 | if (!p4d_present(*p4d)) | |
332 | goto out; | |
333 | pud = pud_offset(p4d, address); | |
8d2afd96 AA |
334 | if (!pud_present(*pud)) |
335 | goto out; | |
336 | pmd = pmd_offset(pud, address); | |
337 | /* | |
338 | * READ_ONCE must function as a barrier with narrower scope | |
339 | * and it must be equivalent to: | |
340 | * _pmd = *pmd; barrier(); | |
341 | * | |
342 | * This is to deal with the instability (as in | |
343 | * pmd_trans_unstable) of the pmd. | |
344 | */ | |
345 | _pmd = READ_ONCE(*pmd); | |
a365ac09 | 346 | if (pmd_none(_pmd)) |
8d2afd96 AA |
347 | goto out; |
348 | ||
349 | ret = false; | |
a365ac09 HY |
350 | if (!pmd_present(_pmd)) |
351 | goto out; | |
352 | ||
63b2d417 AA |
353 | if (pmd_trans_huge(_pmd)) { |
354 | if (!pmd_write(_pmd) && (reason & VM_UFFD_WP)) | |
355 | ret = true; | |
8d2afd96 | 356 | goto out; |
63b2d417 | 357 | } |
8d2afd96 AA |
358 | |
359 | /* | |
360 | * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it | |
361 | * and use the standard pte_offset_map() instead of parsing _pmd. | |
362 | */ | |
363 | pte = pte_offset_map(pmd, address); | |
364 | /* | |
365 | * Lockless access: we're in a wait_event so it's ok if it | |
5c041f5d PX |
366 | * changes under us. PTE markers should be handled the same as none |
367 | * ptes here. | |
8d2afd96 | 368 | */ |
5c041f5d | 369 | if (pte_none_mostly(*pte)) |
8d2afd96 | 370 | ret = true; |
63b2d417 AA |
371 | if (!pte_write(*pte) && (reason & VM_UFFD_WP)) |
372 | ret = true; | |
8d2afd96 AA |
373 | pte_unmap(pte); |
374 | ||
375 | out: | |
376 | return ret; | |
377 | } | |
378 | ||
2f064a59 | 379 | static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags) |
3e69ad08 PX |
380 | { |
381 | if (flags & FAULT_FLAG_INTERRUPTIBLE) | |
382 | return TASK_INTERRUPTIBLE; | |
383 | ||
384 | if (flags & FAULT_FLAG_KILLABLE) | |
385 | return TASK_KILLABLE; | |
386 | ||
387 | return TASK_UNINTERRUPTIBLE; | |
388 | } | |
389 | ||
86039bd3 AA |
390 | /* |
391 | * The locking rules involved in returning VM_FAULT_RETRY depending on | |
392 | * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and | |
393 | * FAULT_FLAG_KILLABLE are not straightforward. The "Caution" | |
394 | * recommendation in __lock_page_or_retry is not an understatement. | |
395 | * | |
c1e8d7c6 | 396 | * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released |
86039bd3 AA |
397 | * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is |
398 | * not set. | |
399 | * | |
400 | * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not | |
401 | * set, VM_FAULT_RETRY can still be returned if and only if there are | |
c1e8d7c6 | 402 | * fatal_signal_pending()s, and the mmap_lock must be released before |
86039bd3 AA |
403 | * returning it. |
404 | */ | |
2b740303 | 405 | vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason) |
86039bd3 | 406 | { |
b8da2e46 PX |
407 | struct vm_area_struct *vma = vmf->vma; |
408 | struct mm_struct *mm = vma->vm_mm; | |
86039bd3 AA |
409 | struct userfaultfd_ctx *ctx; |
410 | struct userfaultfd_wait_queue uwq; | |
2b740303 | 411 | vm_fault_t ret = VM_FAULT_SIGBUS; |
3e69ad08 | 412 | bool must_wait; |
2f064a59 | 413 | unsigned int blocking_state; |
86039bd3 | 414 | |
64c2b203 AA |
415 | /* |
416 | * We don't do userfault handling for the final child pid update. | |
417 | * | |
418 | * We also don't do userfault handling during | |
419 | * coredumping. hugetlbfs has the special | |
420 | * follow_hugetlb_page() to skip missing pages in the | |
421 | * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with | |
422 | * the no_page_table() helper in follow_page_mask(), but the | |
423 | * shmem_vm_ops->fault method is invoked even during | |
c1e8d7c6 | 424 | * coredumping without mmap_lock and it ends up here. |
64c2b203 AA |
425 | */ |
426 | if (current->flags & (PF_EXITING|PF_DUMPCORE)) | |
427 | goto out; | |
428 | ||
429 | /* | |
c1e8d7c6 ML |
430 | * Coredumping runs without mmap_lock so we can only check that |
431 | * the mmap_lock is held, if PF_DUMPCORE was not set. | |
64c2b203 | 432 | */ |
42fc5414 | 433 | mmap_assert_locked(mm); |
64c2b203 | 434 | |
b8da2e46 | 435 | ctx = vma->vm_userfaultfd_ctx.ctx; |
86039bd3 | 436 | if (!ctx) |
ba85c702 | 437 | goto out; |
86039bd3 AA |
438 | |
439 | BUG_ON(ctx->mm != mm); | |
440 | ||
7677f7fd AR |
441 | /* Any unrecognized flag is a bug. */ |
442 | VM_BUG_ON(reason & ~__VM_UFFD_FLAGS); | |
443 | /* 0 or > 1 flags set is a bug; we expect exactly 1. */ | |
444 | VM_BUG_ON(!reason || (reason & (reason - 1))); | |
86039bd3 | 445 | |
2d6d6f5a PS |
446 | if (ctx->features & UFFD_FEATURE_SIGBUS) |
447 | goto out; | |
2d5de004 | 448 | if (!(vmf->flags & FAULT_FLAG_USER) && (ctx->flags & UFFD_USER_MODE_ONLY)) |
37cd0575 | 449 | goto out; |
2d6d6f5a | 450 | |
86039bd3 AA |
451 | /* |
452 | * If it's already released don't get it. This avoids to loop | |
453 | * in __get_user_pages if userfaultfd_release waits on the | |
c1e8d7c6 | 454 | * caller of handle_userfault to release the mmap_lock. |
86039bd3 | 455 | */ |
6aa7de05 | 456 | if (unlikely(READ_ONCE(ctx->released))) { |
656710a6 AA |
457 | /* |
458 | * Don't return VM_FAULT_SIGBUS in this case, so a non | |
459 | * cooperative manager can close the uffd after the | |
460 | * last UFFDIO_COPY, without risking to trigger an | |
461 | * involuntary SIGBUS if the process was starting the | |
462 | * userfaultfd while the userfaultfd was still armed | |
463 | * (but after the last UFFDIO_COPY). If the uffd | |
464 | * wasn't already closed when the userfault reached | |
465 | * this point, that would normally be solved by | |
466 | * userfaultfd_must_wait returning 'false'. | |
467 | * | |
468 | * If we were to return VM_FAULT_SIGBUS here, the non | |
469 | * cooperative manager would be instead forced to | |
470 | * always call UFFDIO_UNREGISTER before it can safely | |
471 | * close the uffd. | |
472 | */ | |
473 | ret = VM_FAULT_NOPAGE; | |
ba85c702 | 474 | goto out; |
656710a6 | 475 | } |
86039bd3 AA |
476 | |
477 | /* | |
478 | * Check that we can return VM_FAULT_RETRY. | |
479 | * | |
480 | * NOTE: it should become possible to return VM_FAULT_RETRY | |
481 | * even if FAULT_FLAG_TRIED is set without leading to gup() | |
482 | * -EBUSY failures, if the userfaultfd is to be extended for | |
483 | * VM_UFFD_WP tracking and we intend to arm the userfault | |
484 | * without first stopping userland access to the memory. For | |
485 | * VM_UFFD_MISSING userfaults this is enough for now. | |
486 | */ | |
82b0f8c3 | 487 | if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) { |
86039bd3 AA |
488 | /* |
489 | * Validate the invariant that nowait must allow retry | |
490 | * to be sure not to return SIGBUS erroneously on | |
491 | * nowait invocations. | |
492 | */ | |
82b0f8c3 | 493 | BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT); |
86039bd3 AA |
494 | #ifdef CONFIG_DEBUG_VM |
495 | if (printk_ratelimit()) { | |
496 | printk(KERN_WARNING | |
82b0f8c3 JK |
497 | "FAULT_FLAG_ALLOW_RETRY missing %x\n", |
498 | vmf->flags); | |
86039bd3 AA |
499 | dump_stack(); |
500 | } | |
501 | #endif | |
ba85c702 | 502 | goto out; |
86039bd3 AA |
503 | } |
504 | ||
505 | /* | |
506 | * Handle nowait, not much to do other than tell it to retry | |
507 | * and wait. | |
508 | */ | |
ba85c702 | 509 | ret = VM_FAULT_RETRY; |
82b0f8c3 | 510 | if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) |
ba85c702 | 511 | goto out; |
86039bd3 | 512 | |
c1e8d7c6 | 513 | /* take the reference before dropping the mmap_lock */ |
86039bd3 AA |
514 | userfaultfd_ctx_get(ctx); |
515 | ||
86039bd3 AA |
516 | init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function); |
517 | uwq.wq.private = current; | |
d172b1a3 NA |
518 | uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags, |
519 | reason, ctx->features); | |
86039bd3 | 520 | uwq.ctx = ctx; |
15a77c6f | 521 | uwq.waken = false; |
86039bd3 | 522 | |
3e69ad08 | 523 | blocking_state = userfaultfd_get_blocking_state(vmf->flags); |
dfa37dc3 | 524 | |
b8da2e46 PX |
525 | /* |
526 | * Take the vma lock now, in order to safely call | |
527 | * userfaultfd_huge_must_wait() later. Since acquiring the | |
528 | * (sleepable) vma lock can modify the current task state, that | |
529 | * must be before explicitly calling set_current_state(). | |
530 | */ | |
531 | if (is_vm_hugetlb_page(vma)) | |
532 | hugetlb_vma_lock_read(vma); | |
533 | ||
cbcfa130 | 534 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
535 | /* |
536 | * After the __add_wait_queue the uwq is visible to userland | |
537 | * through poll/read(). | |
538 | */ | |
15b726ef AA |
539 | __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq); |
540 | /* | |
541 | * The smp_mb() after __set_current_state prevents the reads | |
542 | * following the spin_unlock to happen before the list_add in | |
543 | * __add_wait_queue. | |
544 | */ | |
15a77c6f | 545 | set_current_state(blocking_state); |
cbcfa130 | 546 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 547 | |
b8da2e46 | 548 | if (!is_vm_hugetlb_page(vma)) |
369cd212 MK |
549 | must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags, |
550 | reason); | |
551 | else | |
b8da2e46 | 552 | must_wait = userfaultfd_huge_must_wait(ctx, vma, |
7868a208 | 553 | vmf->address, |
369cd212 | 554 | vmf->flags, reason); |
b8da2e46 PX |
555 | if (is_vm_hugetlb_page(vma)) |
556 | hugetlb_vma_unlock_read(vma); | |
d8ed45c5 | 557 | mmap_read_unlock(mm); |
8d2afd96 | 558 | |
f9bf3522 | 559 | if (likely(must_wait && !READ_ONCE(ctx->released))) { |
a9a08845 | 560 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
86039bd3 | 561 | schedule(); |
ba85c702 | 562 | } |
86039bd3 | 563 | |
ba85c702 | 564 | __set_current_state(TASK_RUNNING); |
15b726ef AA |
565 | |
566 | /* | |
567 | * Here we race with the list_del; list_add in | |
568 | * userfaultfd_ctx_read(), however because we don't ever run | |
569 | * list_del_init() to refile across the two lists, the prev | |
570 | * and next pointers will never point to self. list_add also | |
571 | * would never let any of the two pointers to point to | |
572 | * self. So list_empty_careful won't risk to see both pointers | |
573 | * pointing to self at any time during the list refile. The | |
574 | * only case where list_del_init() is called is the full | |
575 | * removal in the wake function and there we don't re-list_add | |
576 | * and it's fine not to block on the spinlock. The uwq on this | |
577 | * kernel stack can be released after the list_del_init. | |
578 | */ | |
2055da97 | 579 | if (!list_empty_careful(&uwq.wq.entry)) { |
cbcfa130 | 580 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
15b726ef AA |
581 | /* |
582 | * No need of list_del_init(), the uwq on the stack | |
583 | * will be freed shortly anyway. | |
584 | */ | |
2055da97 | 585 | list_del(&uwq.wq.entry); |
cbcfa130 | 586 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 587 | } |
86039bd3 AA |
588 | |
589 | /* | |
590 | * ctx may go away after this if the userfault pseudo fd is | |
591 | * already released. | |
592 | */ | |
593 | userfaultfd_ctx_put(ctx); | |
594 | ||
ba85c702 AA |
595 | out: |
596 | return ret; | |
86039bd3 AA |
597 | } |
598 | ||
8c9e7bb7 AA |
599 | static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx, |
600 | struct userfaultfd_wait_queue *ewq) | |
9cd75c3c | 601 | { |
0cbb4b4f AA |
602 | struct userfaultfd_ctx *release_new_ctx; |
603 | ||
9a69a829 AA |
604 | if (WARN_ON_ONCE(current->flags & PF_EXITING)) |
605 | goto out; | |
9cd75c3c PE |
606 | |
607 | ewq->ctx = ctx; | |
608 | init_waitqueue_entry(&ewq->wq, current); | |
0cbb4b4f | 609 | release_new_ctx = NULL; |
9cd75c3c | 610 | |
cbcfa130 | 611 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
612 | /* |
613 | * After the __add_wait_queue the uwq is visible to userland | |
614 | * through poll/read(). | |
615 | */ | |
616 | __add_wait_queue(&ctx->event_wqh, &ewq->wq); | |
617 | for (;;) { | |
618 | set_current_state(TASK_KILLABLE); | |
619 | if (ewq->msg.event == 0) | |
620 | break; | |
6aa7de05 | 621 | if (READ_ONCE(ctx->released) || |
9cd75c3c | 622 | fatal_signal_pending(current)) { |
384632e6 AA |
623 | /* |
624 | * &ewq->wq may be queued in fork_event, but | |
625 | * __remove_wait_queue ignores the head | |
626 | * parameter. It would be a problem if it | |
627 | * didn't. | |
628 | */ | |
9cd75c3c | 629 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); |
7eb76d45 MR |
630 | if (ewq->msg.event == UFFD_EVENT_FORK) { |
631 | struct userfaultfd_ctx *new; | |
632 | ||
633 | new = (struct userfaultfd_ctx *) | |
634 | (unsigned long) | |
635 | ewq->msg.arg.reserved.reserved1; | |
0cbb4b4f | 636 | release_new_ctx = new; |
7eb76d45 | 637 | } |
9cd75c3c PE |
638 | break; |
639 | } | |
640 | ||
cbcfa130 | 641 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 642 | |
a9a08845 | 643 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
9cd75c3c PE |
644 | schedule(); |
645 | ||
cbcfa130 | 646 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
647 | } |
648 | __set_current_state(TASK_RUNNING); | |
cbcfa130 | 649 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 650 | |
0cbb4b4f AA |
651 | if (release_new_ctx) { |
652 | struct vm_area_struct *vma; | |
653 | struct mm_struct *mm = release_new_ctx->mm; | |
69dbe6da | 654 | VMA_ITERATOR(vmi, mm, 0); |
0cbb4b4f AA |
655 | |
656 | /* the various vma->vm_userfaultfd_ctx still points to it */ | |
d8ed45c5 | 657 | mmap_write_lock(mm); |
69dbe6da | 658 | for_each_vma(vmi, vma) { |
31e810aa | 659 | if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) { |
0cbb4b4f | 660 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
51d3d5eb DH |
661 | userfaultfd_set_vm_flags(vma, |
662 | vma->vm_flags & ~__VM_UFFD_FLAGS); | |
31e810aa | 663 | } |
69dbe6da | 664 | } |
d8ed45c5 | 665 | mmap_write_unlock(mm); |
0cbb4b4f AA |
666 | |
667 | userfaultfd_ctx_put(release_new_ctx); | |
668 | } | |
669 | ||
9cd75c3c PE |
670 | /* |
671 | * ctx may go away after this if the userfault pseudo fd is | |
672 | * already released. | |
673 | */ | |
9a69a829 | 674 | out: |
a759a909 NA |
675 | atomic_dec(&ctx->mmap_changing); |
676 | VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0); | |
9cd75c3c | 677 | userfaultfd_ctx_put(ctx); |
9cd75c3c PE |
678 | } |
679 | ||
680 | static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx, | |
681 | struct userfaultfd_wait_queue *ewq) | |
682 | { | |
683 | ewq->msg.event = 0; | |
684 | wake_up_locked(&ctx->event_wqh); | |
685 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); | |
686 | } | |
687 | ||
893e26e6 PE |
688 | int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs) |
689 | { | |
690 | struct userfaultfd_ctx *ctx = NULL, *octx; | |
691 | struct userfaultfd_fork_ctx *fctx; | |
692 | ||
693 | octx = vma->vm_userfaultfd_ctx.ctx; | |
694 | if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) { | |
695 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
51d3d5eb | 696 | userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS); |
893e26e6 PE |
697 | return 0; |
698 | } | |
699 | ||
700 | list_for_each_entry(fctx, fcs, list) | |
701 | if (fctx->orig == octx) { | |
702 | ctx = fctx->new; | |
703 | break; | |
704 | } | |
705 | ||
706 | if (!ctx) { | |
707 | fctx = kmalloc(sizeof(*fctx), GFP_KERNEL); | |
708 | if (!fctx) | |
709 | return -ENOMEM; | |
710 | ||
711 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); | |
712 | if (!ctx) { | |
713 | kfree(fctx); | |
714 | return -ENOMEM; | |
715 | } | |
716 | ||
ca880420 | 717 | refcount_set(&ctx->refcount, 1); |
893e26e6 | 718 | ctx->flags = octx->flags; |
893e26e6 PE |
719 | ctx->features = octx->features; |
720 | ctx->released = false; | |
a759a909 | 721 | atomic_set(&ctx->mmap_changing, 0); |
893e26e6 | 722 | ctx->mm = vma->vm_mm; |
00bb31fa | 723 | mmgrab(ctx->mm); |
893e26e6 PE |
724 | |
725 | userfaultfd_ctx_get(octx); | |
a759a909 | 726 | atomic_inc(&octx->mmap_changing); |
893e26e6 PE |
727 | fctx->orig = octx; |
728 | fctx->new = ctx; | |
729 | list_add_tail(&fctx->list, fcs); | |
730 | } | |
731 | ||
732 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
733 | return 0; | |
734 | } | |
735 | ||
8c9e7bb7 | 736 | static void dup_fctx(struct userfaultfd_fork_ctx *fctx) |
893e26e6 PE |
737 | { |
738 | struct userfaultfd_ctx *ctx = fctx->orig; | |
739 | struct userfaultfd_wait_queue ewq; | |
740 | ||
741 | msg_init(&ewq.msg); | |
742 | ||
743 | ewq.msg.event = UFFD_EVENT_FORK; | |
744 | ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new; | |
745 | ||
8c9e7bb7 | 746 | userfaultfd_event_wait_completion(ctx, &ewq); |
893e26e6 PE |
747 | } |
748 | ||
749 | void dup_userfaultfd_complete(struct list_head *fcs) | |
750 | { | |
893e26e6 PE |
751 | struct userfaultfd_fork_ctx *fctx, *n; |
752 | ||
753 | list_for_each_entry_safe(fctx, n, fcs, list) { | |
8c9e7bb7 | 754 | dup_fctx(fctx); |
893e26e6 PE |
755 | list_del(&fctx->list); |
756 | kfree(fctx); | |
757 | } | |
758 | } | |
759 | ||
72f87654 PE |
760 | void mremap_userfaultfd_prep(struct vm_area_struct *vma, |
761 | struct vm_userfaultfd_ctx *vm_ctx) | |
762 | { | |
763 | struct userfaultfd_ctx *ctx; | |
764 | ||
765 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
3cfd22be PX |
766 | |
767 | if (!ctx) | |
768 | return; | |
769 | ||
770 | if (ctx->features & UFFD_FEATURE_EVENT_REMAP) { | |
72f87654 PE |
771 | vm_ctx->ctx = ctx; |
772 | userfaultfd_ctx_get(ctx); | |
a759a909 | 773 | atomic_inc(&ctx->mmap_changing); |
3cfd22be PX |
774 | } else { |
775 | /* Drop uffd context if remap feature not enabled */ | |
776 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
51d3d5eb | 777 | userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS); |
72f87654 PE |
778 | } |
779 | } | |
780 | ||
90794bf1 | 781 | void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx, |
72f87654 PE |
782 | unsigned long from, unsigned long to, |
783 | unsigned long len) | |
784 | { | |
90794bf1 | 785 | struct userfaultfd_ctx *ctx = vm_ctx->ctx; |
72f87654 PE |
786 | struct userfaultfd_wait_queue ewq; |
787 | ||
788 | if (!ctx) | |
789 | return; | |
790 | ||
791 | if (to & ~PAGE_MASK) { | |
792 | userfaultfd_ctx_put(ctx); | |
793 | return; | |
794 | } | |
795 | ||
796 | msg_init(&ewq.msg); | |
797 | ||
798 | ewq.msg.event = UFFD_EVENT_REMAP; | |
799 | ewq.msg.arg.remap.from = from; | |
800 | ewq.msg.arg.remap.to = to; | |
801 | ewq.msg.arg.remap.len = len; | |
802 | ||
803 | userfaultfd_event_wait_completion(ctx, &ewq); | |
804 | } | |
805 | ||
70ccb92f | 806 | bool userfaultfd_remove(struct vm_area_struct *vma, |
d811914d | 807 | unsigned long start, unsigned long end) |
05ce7724 PE |
808 | { |
809 | struct mm_struct *mm = vma->vm_mm; | |
810 | struct userfaultfd_ctx *ctx; | |
811 | struct userfaultfd_wait_queue ewq; | |
812 | ||
813 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
d811914d | 814 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE)) |
70ccb92f | 815 | return true; |
05ce7724 PE |
816 | |
817 | userfaultfd_ctx_get(ctx); | |
a759a909 | 818 | atomic_inc(&ctx->mmap_changing); |
d8ed45c5 | 819 | mmap_read_unlock(mm); |
05ce7724 | 820 | |
05ce7724 PE |
821 | msg_init(&ewq.msg); |
822 | ||
d811914d MR |
823 | ewq.msg.event = UFFD_EVENT_REMOVE; |
824 | ewq.msg.arg.remove.start = start; | |
825 | ewq.msg.arg.remove.end = end; | |
05ce7724 PE |
826 | |
827 | userfaultfd_event_wait_completion(ctx, &ewq); | |
828 | ||
70ccb92f | 829 | return false; |
05ce7724 PE |
830 | } |
831 | ||
897ab3e0 MR |
832 | static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps, |
833 | unsigned long start, unsigned long end) | |
834 | { | |
835 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
836 | ||
837 | list_for_each_entry(unmap_ctx, unmaps, list) | |
838 | if (unmap_ctx->ctx == ctx && unmap_ctx->start == start && | |
839 | unmap_ctx->end == end) | |
840 | return true; | |
841 | ||
842 | return false; | |
843 | } | |
844 | ||
69dbe6da LH |
845 | int userfaultfd_unmap_prep(struct mm_struct *mm, unsigned long start, |
846 | unsigned long end, struct list_head *unmaps) | |
897ab3e0 | 847 | { |
69dbe6da LH |
848 | VMA_ITERATOR(vmi, mm, start); |
849 | struct vm_area_struct *vma; | |
850 | ||
851 | for_each_vma_range(vmi, vma, end) { | |
897ab3e0 MR |
852 | struct userfaultfd_unmap_ctx *unmap_ctx; |
853 | struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx; | |
854 | ||
855 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) || | |
856 | has_unmap_ctx(ctx, unmaps, start, end)) | |
857 | continue; | |
858 | ||
859 | unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL); | |
860 | if (!unmap_ctx) | |
861 | return -ENOMEM; | |
862 | ||
863 | userfaultfd_ctx_get(ctx); | |
a759a909 | 864 | atomic_inc(&ctx->mmap_changing); |
897ab3e0 MR |
865 | unmap_ctx->ctx = ctx; |
866 | unmap_ctx->start = start; | |
867 | unmap_ctx->end = end; | |
868 | list_add_tail(&unmap_ctx->list, unmaps); | |
869 | } | |
870 | ||
871 | return 0; | |
872 | } | |
873 | ||
874 | void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf) | |
875 | { | |
876 | struct userfaultfd_unmap_ctx *ctx, *n; | |
877 | struct userfaultfd_wait_queue ewq; | |
878 | ||
879 | list_for_each_entry_safe(ctx, n, uf, list) { | |
880 | msg_init(&ewq.msg); | |
881 | ||
882 | ewq.msg.event = UFFD_EVENT_UNMAP; | |
883 | ewq.msg.arg.remove.start = ctx->start; | |
884 | ewq.msg.arg.remove.end = ctx->end; | |
885 | ||
886 | userfaultfd_event_wait_completion(ctx->ctx, &ewq); | |
887 | ||
888 | list_del(&ctx->list); | |
889 | kfree(ctx); | |
890 | } | |
891 | } | |
892 | ||
86039bd3 AA |
893 | static int userfaultfd_release(struct inode *inode, struct file *file) |
894 | { | |
895 | struct userfaultfd_ctx *ctx = file->private_data; | |
896 | struct mm_struct *mm = ctx->mm; | |
897 | struct vm_area_struct *vma, *prev; | |
898 | /* len == 0 means wake all */ | |
899 | struct userfaultfd_wake_range range = { .len = 0, }; | |
900 | unsigned long new_flags; | |
11a9b902 | 901 | VMA_ITERATOR(vmi, mm, 0); |
86039bd3 | 902 | |
6aa7de05 | 903 | WRITE_ONCE(ctx->released, true); |
86039bd3 | 904 | |
d2005e3f ON |
905 | if (!mmget_not_zero(mm)) |
906 | goto wakeup; | |
907 | ||
86039bd3 AA |
908 | /* |
909 | * Flush page faults out of all CPUs. NOTE: all page faults | |
910 | * must be retried without returning VM_FAULT_SIGBUS if | |
911 | * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx | |
c1e8d7c6 | 912 | * changes while handle_userfault released the mmap_lock. So |
86039bd3 | 913 | * it's critical that released is set to true (above), before |
c1e8d7c6 | 914 | * taking the mmap_lock for writing. |
86039bd3 | 915 | */ |
d8ed45c5 | 916 | mmap_write_lock(mm); |
86039bd3 | 917 | prev = NULL; |
11a9b902 | 918 | for_each_vma(vmi, vma) { |
86039bd3 AA |
919 | cond_resched(); |
920 | BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 921 | !!(vma->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
922 | if (vma->vm_userfaultfd_ctx.ctx != ctx) { |
923 | prev = vma; | |
924 | continue; | |
925 | } | |
7677f7fd | 926 | new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS; |
9760ebff | 927 | prev = vma_merge(&vmi, mm, prev, vma->vm_start, vma->vm_end, |
4d45e75a JH |
928 | new_flags, vma->anon_vma, |
929 | vma->vm_file, vma->vm_pgoff, | |
930 | vma_policy(vma), | |
5c26f6ac | 931 | NULL_VM_UFFD_CTX, anon_vma_name(vma)); |
69dbe6da | 932 | if (prev) { |
4d45e75a | 933 | vma = prev; |
69dbe6da | 934 | } else { |
4d45e75a | 935 | prev = vma; |
69dbe6da LH |
936 | } |
937 | ||
51d3d5eb | 938 | userfaultfd_set_vm_flags(vma, new_flags); |
86039bd3 AA |
939 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
940 | } | |
d8ed45c5 | 941 | mmap_write_unlock(mm); |
d2005e3f ON |
942 | mmput(mm); |
943 | wakeup: | |
86039bd3 | 944 | /* |
15b726ef | 945 | * After no new page faults can wait on this fault_*wqh, flush |
86039bd3 | 946 | * the last page faults that may have been already waiting on |
15b726ef | 947 | * the fault_*wqh. |
86039bd3 | 948 | */ |
cbcfa130 | 949 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
ac5be6b4 | 950 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range); |
c430d1e8 | 951 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range); |
cbcfa130 | 952 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 953 | |
5a18b64e MR |
954 | /* Flush pending events that may still wait on event_wqh */ |
955 | wake_up_all(&ctx->event_wqh); | |
956 | ||
a9a08845 | 957 | wake_up_poll(&ctx->fd_wqh, EPOLLHUP); |
86039bd3 AA |
958 | userfaultfd_ctx_put(ctx); |
959 | return 0; | |
960 | } | |
961 | ||
15b726ef | 962 | /* fault_pending_wqh.lock must be hold by the caller */ |
6dcc27fd PE |
963 | static inline struct userfaultfd_wait_queue *find_userfault_in( |
964 | wait_queue_head_t *wqh) | |
86039bd3 | 965 | { |
ac6424b9 | 966 | wait_queue_entry_t *wq; |
15b726ef | 967 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 968 | |
456a7378 | 969 | lockdep_assert_held(&wqh->lock); |
86039bd3 | 970 | |
15b726ef | 971 | uwq = NULL; |
6dcc27fd | 972 | if (!waitqueue_active(wqh)) |
15b726ef AA |
973 | goto out; |
974 | /* walk in reverse to provide FIFO behavior to read userfaults */ | |
2055da97 | 975 | wq = list_last_entry(&wqh->head, typeof(*wq), entry); |
15b726ef AA |
976 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
977 | out: | |
978 | return uwq; | |
86039bd3 | 979 | } |
6dcc27fd PE |
980 | |
981 | static inline struct userfaultfd_wait_queue *find_userfault( | |
982 | struct userfaultfd_ctx *ctx) | |
983 | { | |
984 | return find_userfault_in(&ctx->fault_pending_wqh); | |
985 | } | |
86039bd3 | 986 | |
9cd75c3c PE |
987 | static inline struct userfaultfd_wait_queue *find_userfault_evt( |
988 | struct userfaultfd_ctx *ctx) | |
989 | { | |
990 | return find_userfault_in(&ctx->event_wqh); | |
991 | } | |
992 | ||
076ccb76 | 993 | static __poll_t userfaultfd_poll(struct file *file, poll_table *wait) |
86039bd3 AA |
994 | { |
995 | struct userfaultfd_ctx *ctx = file->private_data; | |
076ccb76 | 996 | __poll_t ret; |
86039bd3 AA |
997 | |
998 | poll_wait(file, &ctx->fd_wqh, wait); | |
999 | ||
22e5fe2a | 1000 | if (!userfaultfd_is_initialized(ctx)) |
a9a08845 | 1001 | return EPOLLERR; |
9cd75c3c | 1002 | |
22e5fe2a NA |
1003 | /* |
1004 | * poll() never guarantees that read won't block. | |
1005 | * userfaults can be waken before they're read(). | |
1006 | */ | |
1007 | if (unlikely(!(file->f_flags & O_NONBLOCK))) | |
a9a08845 | 1008 | return EPOLLERR; |
22e5fe2a NA |
1009 | /* |
1010 | * lockless access to see if there are pending faults | |
1011 | * __pollwait last action is the add_wait_queue but | |
1012 | * the spin_unlock would allow the waitqueue_active to | |
1013 | * pass above the actual list_add inside | |
1014 | * add_wait_queue critical section. So use a full | |
1015 | * memory barrier to serialize the list_add write of | |
1016 | * add_wait_queue() with the waitqueue_active read | |
1017 | * below. | |
1018 | */ | |
1019 | ret = 0; | |
1020 | smp_mb(); | |
1021 | if (waitqueue_active(&ctx->fault_pending_wqh)) | |
1022 | ret = EPOLLIN; | |
1023 | else if (waitqueue_active(&ctx->event_wqh)) | |
1024 | ret = EPOLLIN; | |
1025 | ||
1026 | return ret; | |
86039bd3 AA |
1027 | } |
1028 | ||
893e26e6 PE |
1029 | static const struct file_operations userfaultfd_fops; |
1030 | ||
b537900f DC |
1031 | static int resolve_userfault_fork(struct userfaultfd_ctx *new, |
1032 | struct inode *inode, | |
893e26e6 PE |
1033 | struct uffd_msg *msg) |
1034 | { | |
1035 | int fd; | |
893e26e6 | 1036 | |
b537900f | 1037 | fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new, |
abec3d01 | 1038 | O_RDONLY | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode); |
893e26e6 PE |
1039 | if (fd < 0) |
1040 | return fd; | |
1041 | ||
893e26e6 PE |
1042 | msg->arg.reserved.reserved1 = 0; |
1043 | msg->arg.fork.ufd = fd; | |
893e26e6 PE |
1044 | return 0; |
1045 | } | |
1046 | ||
86039bd3 | 1047 | static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait, |
b537900f | 1048 | struct uffd_msg *msg, struct inode *inode) |
86039bd3 AA |
1049 | { |
1050 | ssize_t ret; | |
1051 | DECLARE_WAITQUEUE(wait, current); | |
15b726ef | 1052 | struct userfaultfd_wait_queue *uwq; |
893e26e6 PE |
1053 | /* |
1054 | * Handling fork event requires sleeping operations, so | |
1055 | * we drop the event_wqh lock, then do these ops, then | |
1056 | * lock it back and wake up the waiter. While the lock is | |
1057 | * dropped the ewq may go away so we keep track of it | |
1058 | * carefully. | |
1059 | */ | |
1060 | LIST_HEAD(fork_event); | |
1061 | struct userfaultfd_ctx *fork_nctx = NULL; | |
86039bd3 | 1062 | |
15b726ef | 1063 | /* always take the fd_wqh lock before the fault_pending_wqh lock */ |
ae62c16e | 1064 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1065 | __add_wait_queue(&ctx->fd_wqh, &wait); |
1066 | for (;;) { | |
1067 | set_current_state(TASK_INTERRUPTIBLE); | |
15b726ef AA |
1068 | spin_lock(&ctx->fault_pending_wqh.lock); |
1069 | uwq = find_userfault(ctx); | |
1070 | if (uwq) { | |
2c5b7e1b AA |
1071 | /* |
1072 | * Use a seqcount to repeat the lockless check | |
1073 | * in wake_userfault() to avoid missing | |
1074 | * wakeups because during the refile both | |
1075 | * waitqueue could become empty if this is the | |
1076 | * only userfault. | |
1077 | */ | |
1078 | write_seqcount_begin(&ctx->refile_seq); | |
1079 | ||
86039bd3 | 1080 | /* |
15b726ef AA |
1081 | * The fault_pending_wqh.lock prevents the uwq |
1082 | * to disappear from under us. | |
1083 | * | |
1084 | * Refile this userfault from | |
1085 | * fault_pending_wqh to fault_wqh, it's not | |
1086 | * pending anymore after we read it. | |
1087 | * | |
1088 | * Use list_del() by hand (as | |
1089 | * userfaultfd_wake_function also uses | |
1090 | * list_del_init() by hand) to be sure nobody | |
1091 | * changes __remove_wait_queue() to use | |
1092 | * list_del_init() in turn breaking the | |
1093 | * !list_empty_careful() check in | |
2055da97 | 1094 | * handle_userfault(). The uwq->wq.head list |
15b726ef AA |
1095 | * must never be empty at any time during the |
1096 | * refile, or the waitqueue could disappear | |
1097 | * from under us. The "wait_queue_head_t" | |
1098 | * parameter of __remove_wait_queue() is unused | |
1099 | * anyway. | |
86039bd3 | 1100 | */ |
2055da97 | 1101 | list_del(&uwq->wq.entry); |
c430d1e8 | 1102 | add_wait_queue(&ctx->fault_wqh, &uwq->wq); |
15b726ef | 1103 | |
2c5b7e1b AA |
1104 | write_seqcount_end(&ctx->refile_seq); |
1105 | ||
a9b85f94 AA |
1106 | /* careful to always initialize msg if ret == 0 */ |
1107 | *msg = uwq->msg; | |
15b726ef | 1108 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1109 | ret = 0; |
1110 | break; | |
1111 | } | |
15b726ef | 1112 | spin_unlock(&ctx->fault_pending_wqh.lock); |
9cd75c3c PE |
1113 | |
1114 | spin_lock(&ctx->event_wqh.lock); | |
1115 | uwq = find_userfault_evt(ctx); | |
1116 | if (uwq) { | |
1117 | *msg = uwq->msg; | |
1118 | ||
893e26e6 PE |
1119 | if (uwq->msg.event == UFFD_EVENT_FORK) { |
1120 | fork_nctx = (struct userfaultfd_ctx *) | |
1121 | (unsigned long) | |
1122 | uwq->msg.arg.reserved.reserved1; | |
2055da97 | 1123 | list_move(&uwq->wq.entry, &fork_event); |
384632e6 AA |
1124 | /* |
1125 | * fork_nctx can be freed as soon as | |
1126 | * we drop the lock, unless we take a | |
1127 | * reference on it. | |
1128 | */ | |
1129 | userfaultfd_ctx_get(fork_nctx); | |
893e26e6 PE |
1130 | spin_unlock(&ctx->event_wqh.lock); |
1131 | ret = 0; | |
1132 | break; | |
1133 | } | |
1134 | ||
9cd75c3c PE |
1135 | userfaultfd_event_complete(ctx, uwq); |
1136 | spin_unlock(&ctx->event_wqh.lock); | |
1137 | ret = 0; | |
1138 | break; | |
1139 | } | |
1140 | spin_unlock(&ctx->event_wqh.lock); | |
1141 | ||
86039bd3 AA |
1142 | if (signal_pending(current)) { |
1143 | ret = -ERESTARTSYS; | |
1144 | break; | |
1145 | } | |
1146 | if (no_wait) { | |
1147 | ret = -EAGAIN; | |
1148 | break; | |
1149 | } | |
ae62c16e | 1150 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1151 | schedule(); |
ae62c16e | 1152 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1153 | } |
1154 | __remove_wait_queue(&ctx->fd_wqh, &wait); | |
1155 | __set_current_state(TASK_RUNNING); | |
ae62c16e | 1156 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1157 | |
893e26e6 | 1158 | if (!ret && msg->event == UFFD_EVENT_FORK) { |
b537900f | 1159 | ret = resolve_userfault_fork(fork_nctx, inode, msg); |
cbcfa130 | 1160 | spin_lock_irq(&ctx->event_wqh.lock); |
384632e6 AA |
1161 | if (!list_empty(&fork_event)) { |
1162 | /* | |
1163 | * The fork thread didn't abort, so we can | |
1164 | * drop the temporary refcount. | |
1165 | */ | |
1166 | userfaultfd_ctx_put(fork_nctx); | |
1167 | ||
1168 | uwq = list_first_entry(&fork_event, | |
1169 | typeof(*uwq), | |
1170 | wq.entry); | |
1171 | /* | |
1172 | * If fork_event list wasn't empty and in turn | |
1173 | * the event wasn't already released by fork | |
1174 | * (the event is allocated on fork kernel | |
1175 | * stack), put the event back to its place in | |
1176 | * the event_wq. fork_event head will be freed | |
1177 | * as soon as we return so the event cannot | |
1178 | * stay queued there no matter the current | |
1179 | * "ret" value. | |
1180 | */ | |
1181 | list_del(&uwq->wq.entry); | |
1182 | __add_wait_queue(&ctx->event_wqh, &uwq->wq); | |
893e26e6 | 1183 | |
384632e6 AA |
1184 | /* |
1185 | * Leave the event in the waitqueue and report | |
1186 | * error to userland if we failed to resolve | |
1187 | * the userfault fork. | |
1188 | */ | |
1189 | if (likely(!ret)) | |
893e26e6 | 1190 | userfaultfd_event_complete(ctx, uwq); |
384632e6 AA |
1191 | } else { |
1192 | /* | |
1193 | * Here the fork thread aborted and the | |
1194 | * refcount from the fork thread on fork_nctx | |
1195 | * has already been released. We still hold | |
1196 | * the reference we took before releasing the | |
1197 | * lock above. If resolve_userfault_fork | |
1198 | * failed we've to drop it because the | |
1199 | * fork_nctx has to be freed in such case. If | |
1200 | * it succeeded we'll hold it because the new | |
1201 | * uffd references it. | |
1202 | */ | |
1203 | if (ret) | |
1204 | userfaultfd_ctx_put(fork_nctx); | |
893e26e6 | 1205 | } |
cbcfa130 | 1206 | spin_unlock_irq(&ctx->event_wqh.lock); |
893e26e6 PE |
1207 | } |
1208 | ||
86039bd3 AA |
1209 | return ret; |
1210 | } | |
1211 | ||
1212 | static ssize_t userfaultfd_read(struct file *file, char __user *buf, | |
1213 | size_t count, loff_t *ppos) | |
1214 | { | |
1215 | struct userfaultfd_ctx *ctx = file->private_data; | |
1216 | ssize_t _ret, ret = 0; | |
a9b85f94 | 1217 | struct uffd_msg msg; |
86039bd3 | 1218 | int no_wait = file->f_flags & O_NONBLOCK; |
b537900f | 1219 | struct inode *inode = file_inode(file); |
86039bd3 | 1220 | |
22e5fe2a | 1221 | if (!userfaultfd_is_initialized(ctx)) |
86039bd3 | 1222 | return -EINVAL; |
86039bd3 AA |
1223 | |
1224 | for (;;) { | |
a9b85f94 | 1225 | if (count < sizeof(msg)) |
86039bd3 | 1226 | return ret ? ret : -EINVAL; |
b537900f | 1227 | _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode); |
86039bd3 AA |
1228 | if (_ret < 0) |
1229 | return ret ? ret : _ret; | |
a9b85f94 | 1230 | if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg))) |
86039bd3 | 1231 | return ret ? ret : -EFAULT; |
a9b85f94 AA |
1232 | ret += sizeof(msg); |
1233 | buf += sizeof(msg); | |
1234 | count -= sizeof(msg); | |
86039bd3 AA |
1235 | /* |
1236 | * Allow to read more than one fault at time but only | |
1237 | * block if waiting for the very first one. | |
1238 | */ | |
1239 | no_wait = O_NONBLOCK; | |
1240 | } | |
1241 | } | |
1242 | ||
1243 | static void __wake_userfault(struct userfaultfd_ctx *ctx, | |
1244 | struct userfaultfd_wake_range *range) | |
1245 | { | |
cbcfa130 | 1246 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 1247 | /* wake all in the range and autoremove */ |
15b726ef | 1248 | if (waitqueue_active(&ctx->fault_pending_wqh)) |
ac5be6b4 | 1249 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, |
15b726ef AA |
1250 | range); |
1251 | if (waitqueue_active(&ctx->fault_wqh)) | |
c430d1e8 | 1252 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range); |
cbcfa130 | 1253 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1254 | } |
1255 | ||
1256 | static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx, | |
1257 | struct userfaultfd_wake_range *range) | |
1258 | { | |
2c5b7e1b AA |
1259 | unsigned seq; |
1260 | bool need_wakeup; | |
1261 | ||
86039bd3 AA |
1262 | /* |
1263 | * To be sure waitqueue_active() is not reordered by the CPU | |
1264 | * before the pagetable update, use an explicit SMP memory | |
3e4e28c5 | 1265 | * barrier here. PT lock release or mmap_read_unlock(mm) still |
86039bd3 AA |
1266 | * have release semantics that can allow the |
1267 | * waitqueue_active() to be reordered before the pte update. | |
1268 | */ | |
1269 | smp_mb(); | |
1270 | ||
1271 | /* | |
1272 | * Use waitqueue_active because it's very frequent to | |
1273 | * change the address space atomically even if there are no | |
1274 | * userfaults yet. So we take the spinlock only when we're | |
1275 | * sure we've userfaults to wake. | |
1276 | */ | |
2c5b7e1b AA |
1277 | do { |
1278 | seq = read_seqcount_begin(&ctx->refile_seq); | |
1279 | need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) || | |
1280 | waitqueue_active(&ctx->fault_wqh); | |
1281 | cond_resched(); | |
1282 | } while (read_seqcount_retry(&ctx->refile_seq, seq)); | |
1283 | if (need_wakeup) | |
86039bd3 AA |
1284 | __wake_userfault(ctx, range); |
1285 | } | |
1286 | ||
1287 | static __always_inline int validate_range(struct mm_struct *mm, | |
e71e2ace | 1288 | __u64 start, __u64 len) |
86039bd3 AA |
1289 | { |
1290 | __u64 task_size = mm->task_size; | |
1291 | ||
e71e2ace | 1292 | if (start & ~PAGE_MASK) |
86039bd3 AA |
1293 | return -EINVAL; |
1294 | if (len & ~PAGE_MASK) | |
1295 | return -EINVAL; | |
1296 | if (!len) | |
1297 | return -EINVAL; | |
e71e2ace | 1298 | if (start < mmap_min_addr) |
86039bd3 | 1299 | return -EINVAL; |
e71e2ace | 1300 | if (start >= task_size) |
86039bd3 | 1301 | return -EINVAL; |
e71e2ace | 1302 | if (len > task_size - start) |
86039bd3 AA |
1303 | return -EINVAL; |
1304 | return 0; | |
1305 | } | |
1306 | ||
1307 | static int userfaultfd_register(struct userfaultfd_ctx *ctx, | |
1308 | unsigned long arg) | |
1309 | { | |
1310 | struct mm_struct *mm = ctx->mm; | |
1311 | struct vm_area_struct *vma, *prev, *cur; | |
1312 | int ret; | |
1313 | struct uffdio_register uffdio_register; | |
1314 | struct uffdio_register __user *user_uffdio_register; | |
1315 | unsigned long vm_flags, new_flags; | |
1316 | bool found; | |
ce53e8e6 | 1317 | bool basic_ioctls; |
86039bd3 | 1318 | unsigned long start, end, vma_end; |
11a9b902 | 1319 | struct vma_iterator vmi; |
86039bd3 AA |
1320 | |
1321 | user_uffdio_register = (struct uffdio_register __user *) arg; | |
1322 | ||
1323 | ret = -EFAULT; | |
1324 | if (copy_from_user(&uffdio_register, user_uffdio_register, | |
1325 | sizeof(uffdio_register)-sizeof(__u64))) | |
1326 | goto out; | |
1327 | ||
1328 | ret = -EINVAL; | |
1329 | if (!uffdio_register.mode) | |
1330 | goto out; | |
7677f7fd | 1331 | if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES) |
86039bd3 AA |
1332 | goto out; |
1333 | vm_flags = 0; | |
1334 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING) | |
1335 | vm_flags |= VM_UFFD_MISSING; | |
00b151f2 PX |
1336 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) { |
1337 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP | |
1338 | goto out; | |
1339 | #endif | |
86039bd3 | 1340 | vm_flags |= VM_UFFD_WP; |
00b151f2 | 1341 | } |
7677f7fd AR |
1342 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) { |
1343 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR | |
1344 | goto out; | |
1345 | #endif | |
1346 | vm_flags |= VM_UFFD_MINOR; | |
1347 | } | |
86039bd3 | 1348 | |
e71e2ace | 1349 | ret = validate_range(mm, uffdio_register.range.start, |
86039bd3 AA |
1350 | uffdio_register.range.len); |
1351 | if (ret) | |
1352 | goto out; | |
1353 | ||
1354 | start = uffdio_register.range.start; | |
1355 | end = start + uffdio_register.range.len; | |
1356 | ||
d2005e3f ON |
1357 | ret = -ENOMEM; |
1358 | if (!mmget_not_zero(mm)) | |
1359 | goto out; | |
1360 | ||
11a9b902 | 1361 | ret = -EINVAL; |
d8ed45c5 | 1362 | mmap_write_lock(mm); |
11a9b902 LH |
1363 | vma_iter_init(&vmi, mm, start); |
1364 | vma = vma_find(&vmi, end); | |
86039bd3 AA |
1365 | if (!vma) |
1366 | goto out_unlock; | |
1367 | ||
cab350af MK |
1368 | /* |
1369 | * If the first vma contains huge pages, make sure start address | |
1370 | * is aligned to huge page size. | |
1371 | */ | |
1372 | if (is_vm_hugetlb_page(vma)) { | |
1373 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1374 | ||
1375 | if (start & (vma_hpagesize - 1)) | |
1376 | goto out_unlock; | |
1377 | } | |
1378 | ||
86039bd3 AA |
1379 | /* |
1380 | * Search for not compatible vmas. | |
86039bd3 AA |
1381 | */ |
1382 | found = false; | |
ce53e8e6 | 1383 | basic_ioctls = false; |
11a9b902 LH |
1384 | cur = vma; |
1385 | do { | |
86039bd3 AA |
1386 | cond_resched(); |
1387 | ||
1388 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 1389 | !!(cur->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
1390 | |
1391 | /* check not compatible vmas */ | |
1392 | ret = -EINVAL; | |
63b2d417 | 1393 | if (!vma_can_userfault(cur, vm_flags)) |
86039bd3 | 1394 | goto out_unlock; |
29ec9066 AA |
1395 | |
1396 | /* | |
1397 | * UFFDIO_COPY will fill file holes even without | |
1398 | * PROT_WRITE. This check enforces that if this is a | |
1399 | * MAP_SHARED, the process has write permission to the backing | |
1400 | * file. If VM_MAYWRITE is set it also enforces that on a | |
1401 | * MAP_SHARED vma: there is no F_WRITE_SEAL and no further | |
1402 | * F_WRITE_SEAL can be taken until the vma is destroyed. | |
1403 | */ | |
1404 | ret = -EPERM; | |
1405 | if (unlikely(!(cur->vm_flags & VM_MAYWRITE))) | |
1406 | goto out_unlock; | |
1407 | ||
cab350af MK |
1408 | /* |
1409 | * If this vma contains ending address, and huge pages | |
1410 | * check alignment. | |
1411 | */ | |
1412 | if (is_vm_hugetlb_page(cur) && end <= cur->vm_end && | |
1413 | end > cur->vm_start) { | |
1414 | unsigned long vma_hpagesize = vma_kernel_pagesize(cur); | |
1415 | ||
1416 | ret = -EINVAL; | |
1417 | ||
1418 | if (end & (vma_hpagesize - 1)) | |
1419 | goto out_unlock; | |
1420 | } | |
63b2d417 AA |
1421 | if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE)) |
1422 | goto out_unlock; | |
86039bd3 AA |
1423 | |
1424 | /* | |
1425 | * Check that this vma isn't already owned by a | |
1426 | * different userfaultfd. We can't allow more than one | |
1427 | * userfaultfd to own a single vma simultaneously or we | |
1428 | * wouldn't know which one to deliver the userfaults to. | |
1429 | */ | |
1430 | ret = -EBUSY; | |
1431 | if (cur->vm_userfaultfd_ctx.ctx && | |
1432 | cur->vm_userfaultfd_ctx.ctx != ctx) | |
1433 | goto out_unlock; | |
1434 | ||
cab350af MK |
1435 | /* |
1436 | * Note vmas containing huge pages | |
1437 | */ | |
ce53e8e6 MR |
1438 | if (is_vm_hugetlb_page(cur)) |
1439 | basic_ioctls = true; | |
cab350af | 1440 | |
86039bd3 | 1441 | found = true; |
11a9b902 | 1442 | } for_each_vma_range(vmi, cur, end); |
86039bd3 AA |
1443 | BUG_ON(!found); |
1444 | ||
11a9b902 LH |
1445 | vma_iter_set(&vmi, start); |
1446 | prev = vma_prev(&vmi); | |
86039bd3 AA |
1447 | |
1448 | ret = 0; | |
11a9b902 | 1449 | for_each_vma_range(vmi, vma, end) { |
86039bd3 AA |
1450 | cond_resched(); |
1451 | ||
63b2d417 | 1452 | BUG_ON(!vma_can_userfault(vma, vm_flags)); |
86039bd3 AA |
1453 | BUG_ON(vma->vm_userfaultfd_ctx.ctx && |
1454 | vma->vm_userfaultfd_ctx.ctx != ctx); | |
29ec9066 | 1455 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
86039bd3 AA |
1456 | |
1457 | /* | |
1458 | * Nothing to do: this vma is already registered into this | |
1459 | * userfaultfd and with the right tracking mode too. | |
1460 | */ | |
1461 | if (vma->vm_userfaultfd_ctx.ctx == ctx && | |
1462 | (vma->vm_flags & vm_flags) == vm_flags) | |
1463 | goto skip; | |
1464 | ||
1465 | if (vma->vm_start > start) | |
1466 | start = vma->vm_start; | |
1467 | vma_end = min(end, vma->vm_end); | |
1468 | ||
7677f7fd | 1469 | new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags; |
9760ebff | 1470 | prev = vma_merge(&vmi, mm, prev, start, vma_end, new_flags, |
86039bd3 AA |
1471 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, |
1472 | vma_policy(vma), | |
9a10064f | 1473 | ((struct vm_userfaultfd_ctx){ ctx }), |
5c26f6ac | 1474 | anon_vma_name(vma)); |
86039bd3 | 1475 | if (prev) { |
69dbe6da | 1476 | /* vma_merge() invalidated the mas */ |
86039bd3 AA |
1477 | vma = prev; |
1478 | goto next; | |
1479 | } | |
1480 | if (vma->vm_start < start) { | |
9760ebff | 1481 | ret = split_vma(&vmi, vma, start, 1); |
86039bd3 AA |
1482 | if (ret) |
1483 | break; | |
1484 | } | |
1485 | if (vma->vm_end > end) { | |
9760ebff | 1486 | ret = split_vma(&vmi, vma, end, 0); |
86039bd3 AA |
1487 | if (ret) |
1488 | break; | |
1489 | } | |
1490 | next: | |
1491 | /* | |
1492 | * In the vma_merge() successful mprotect-like case 8: | |
1493 | * the next vma was merged into the current one and | |
1494 | * the current one has not been updated yet. | |
1495 | */ | |
51d3d5eb | 1496 | userfaultfd_set_vm_flags(vma, new_flags); |
86039bd3 AA |
1497 | vma->vm_userfaultfd_ctx.ctx = ctx; |
1498 | ||
6dfeaff9 PX |
1499 | if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma)) |
1500 | hugetlb_unshare_all_pmds(vma); | |
1501 | ||
86039bd3 AA |
1502 | skip: |
1503 | prev = vma; | |
1504 | start = vma->vm_end; | |
11a9b902 LH |
1505 | } |
1506 | ||
86039bd3 | 1507 | out_unlock: |
d8ed45c5 | 1508 | mmap_write_unlock(mm); |
d2005e3f | 1509 | mmput(mm); |
86039bd3 | 1510 | if (!ret) { |
14819305 PX |
1511 | __u64 ioctls_out; |
1512 | ||
1513 | ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC : | |
1514 | UFFD_API_RANGE_IOCTLS; | |
1515 | ||
1516 | /* | |
1517 | * Declare the WP ioctl only if the WP mode is | |
1518 | * specified and all checks passed with the range | |
1519 | */ | |
1520 | if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP)) | |
1521 | ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT); | |
1522 | ||
f6191471 AR |
1523 | /* CONTINUE ioctl is only supported for MINOR ranges. */ |
1524 | if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR)) | |
1525 | ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE); | |
1526 | ||
86039bd3 AA |
1527 | /* |
1528 | * Now that we scanned all vmas we can already tell | |
1529 | * userland which ioctls methods are guaranteed to | |
1530 | * succeed on this range. | |
1531 | */ | |
14819305 | 1532 | if (put_user(ioctls_out, &user_uffdio_register->ioctls)) |
86039bd3 AA |
1533 | ret = -EFAULT; |
1534 | } | |
1535 | out: | |
1536 | return ret; | |
1537 | } | |
1538 | ||
1539 | static int userfaultfd_unregister(struct userfaultfd_ctx *ctx, | |
1540 | unsigned long arg) | |
1541 | { | |
1542 | struct mm_struct *mm = ctx->mm; | |
1543 | struct vm_area_struct *vma, *prev, *cur; | |
1544 | int ret; | |
1545 | struct uffdio_range uffdio_unregister; | |
1546 | unsigned long new_flags; | |
1547 | bool found; | |
1548 | unsigned long start, end, vma_end; | |
1549 | const void __user *buf = (void __user *)arg; | |
11a9b902 | 1550 | struct vma_iterator vmi; |
86039bd3 AA |
1551 | |
1552 | ret = -EFAULT; | |
1553 | if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister))) | |
1554 | goto out; | |
1555 | ||
e71e2ace | 1556 | ret = validate_range(mm, uffdio_unregister.start, |
86039bd3 AA |
1557 | uffdio_unregister.len); |
1558 | if (ret) | |
1559 | goto out; | |
1560 | ||
1561 | start = uffdio_unregister.start; | |
1562 | end = start + uffdio_unregister.len; | |
1563 | ||
d2005e3f ON |
1564 | ret = -ENOMEM; |
1565 | if (!mmget_not_zero(mm)) | |
1566 | goto out; | |
1567 | ||
d8ed45c5 | 1568 | mmap_write_lock(mm); |
86039bd3 | 1569 | ret = -EINVAL; |
11a9b902 LH |
1570 | vma_iter_init(&vmi, mm, start); |
1571 | vma = vma_find(&vmi, end); | |
1572 | if (!vma) | |
86039bd3 AA |
1573 | goto out_unlock; |
1574 | ||
cab350af MK |
1575 | /* |
1576 | * If the first vma contains huge pages, make sure start address | |
1577 | * is aligned to huge page size. | |
1578 | */ | |
1579 | if (is_vm_hugetlb_page(vma)) { | |
1580 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1581 | ||
1582 | if (start & (vma_hpagesize - 1)) | |
1583 | goto out_unlock; | |
1584 | } | |
1585 | ||
86039bd3 AA |
1586 | /* |
1587 | * Search for not compatible vmas. | |
86039bd3 AA |
1588 | */ |
1589 | found = false; | |
11a9b902 LH |
1590 | cur = vma; |
1591 | do { | |
86039bd3 AA |
1592 | cond_resched(); |
1593 | ||
1594 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 1595 | !!(cur->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
1596 | |
1597 | /* | |
1598 | * Check not compatible vmas, not strictly required | |
1599 | * here as not compatible vmas cannot have an | |
1600 | * userfaultfd_ctx registered on them, but this | |
1601 | * provides for more strict behavior to notice | |
1602 | * unregistration errors. | |
1603 | */ | |
63b2d417 | 1604 | if (!vma_can_userfault(cur, cur->vm_flags)) |
86039bd3 AA |
1605 | goto out_unlock; |
1606 | ||
1607 | found = true; | |
11a9b902 | 1608 | } for_each_vma_range(vmi, cur, end); |
86039bd3 AA |
1609 | BUG_ON(!found); |
1610 | ||
11a9b902 LH |
1611 | vma_iter_set(&vmi, start); |
1612 | prev = vma_prev(&vmi); | |
86039bd3 | 1613 | ret = 0; |
11a9b902 | 1614 | for_each_vma_range(vmi, vma, end) { |
86039bd3 AA |
1615 | cond_resched(); |
1616 | ||
63b2d417 | 1617 | BUG_ON(!vma_can_userfault(vma, vma->vm_flags)); |
86039bd3 AA |
1618 | |
1619 | /* | |
1620 | * Nothing to do: this vma is already registered into this | |
1621 | * userfaultfd and with the right tracking mode too. | |
1622 | */ | |
1623 | if (!vma->vm_userfaultfd_ctx.ctx) | |
1624 | goto skip; | |
1625 | ||
01e881f5 AA |
1626 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
1627 | ||
86039bd3 AA |
1628 | if (vma->vm_start > start) |
1629 | start = vma->vm_start; | |
1630 | vma_end = min(end, vma->vm_end); | |
1631 | ||
09fa5296 AA |
1632 | if (userfaultfd_missing(vma)) { |
1633 | /* | |
1634 | * Wake any concurrent pending userfault while | |
1635 | * we unregister, so they will not hang | |
1636 | * permanently and it avoids userland to call | |
1637 | * UFFDIO_WAKE explicitly. | |
1638 | */ | |
1639 | struct userfaultfd_wake_range range; | |
1640 | range.start = start; | |
1641 | range.len = vma_end - start; | |
1642 | wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range); | |
1643 | } | |
1644 | ||
f369b07c PX |
1645 | /* Reset ptes for the whole vma range if wr-protected */ |
1646 | if (userfaultfd_wp(vma)) | |
61c50040 | 1647 | uffd_wp_range(vma, start, vma_end - start, false); |
f369b07c | 1648 | |
7677f7fd | 1649 | new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS; |
9760ebff | 1650 | prev = vma_merge(&vmi, mm, prev, start, vma_end, new_flags, |
86039bd3 AA |
1651 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, |
1652 | vma_policy(vma), | |
5c26f6ac | 1653 | NULL_VM_UFFD_CTX, anon_vma_name(vma)); |
86039bd3 AA |
1654 | if (prev) { |
1655 | vma = prev; | |
1656 | goto next; | |
1657 | } | |
1658 | if (vma->vm_start < start) { | |
9760ebff | 1659 | ret = split_vma(&vmi, vma, start, 1); |
86039bd3 AA |
1660 | if (ret) |
1661 | break; | |
1662 | } | |
1663 | if (vma->vm_end > end) { | |
9760ebff | 1664 | ret = split_vma(&vmi, vma, end, 0); |
86039bd3 AA |
1665 | if (ret) |
1666 | break; | |
1667 | } | |
1668 | next: | |
1669 | /* | |
1670 | * In the vma_merge() successful mprotect-like case 8: | |
1671 | * the next vma was merged into the current one and | |
1672 | * the current one has not been updated yet. | |
1673 | */ | |
51d3d5eb | 1674 | userfaultfd_set_vm_flags(vma, new_flags); |
86039bd3 AA |
1675 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
1676 | ||
1677 | skip: | |
1678 | prev = vma; | |
1679 | start = vma->vm_end; | |
11a9b902 LH |
1680 | } |
1681 | ||
86039bd3 | 1682 | out_unlock: |
d8ed45c5 | 1683 | mmap_write_unlock(mm); |
d2005e3f | 1684 | mmput(mm); |
86039bd3 AA |
1685 | out: |
1686 | return ret; | |
1687 | } | |
1688 | ||
1689 | /* | |
ba85c702 AA |
1690 | * userfaultfd_wake may be used in combination with the |
1691 | * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches. | |
86039bd3 AA |
1692 | */ |
1693 | static int userfaultfd_wake(struct userfaultfd_ctx *ctx, | |
1694 | unsigned long arg) | |
1695 | { | |
1696 | int ret; | |
1697 | struct uffdio_range uffdio_wake; | |
1698 | struct userfaultfd_wake_range range; | |
1699 | const void __user *buf = (void __user *)arg; | |
1700 | ||
1701 | ret = -EFAULT; | |
1702 | if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake))) | |
1703 | goto out; | |
1704 | ||
e71e2ace | 1705 | ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len); |
86039bd3 AA |
1706 | if (ret) |
1707 | goto out; | |
1708 | ||
1709 | range.start = uffdio_wake.start; | |
1710 | range.len = uffdio_wake.len; | |
1711 | ||
1712 | /* | |
1713 | * len == 0 means wake all and we don't want to wake all here, | |
1714 | * so check it again to be sure. | |
1715 | */ | |
1716 | VM_BUG_ON(!range.len); | |
1717 | ||
1718 | wake_userfault(ctx, &range); | |
1719 | ret = 0; | |
1720 | ||
1721 | out: | |
1722 | return ret; | |
1723 | } | |
1724 | ||
ad465cae AA |
1725 | static int userfaultfd_copy(struct userfaultfd_ctx *ctx, |
1726 | unsigned long arg) | |
1727 | { | |
1728 | __s64 ret; | |
1729 | struct uffdio_copy uffdio_copy; | |
1730 | struct uffdio_copy __user *user_uffdio_copy; | |
1731 | struct userfaultfd_wake_range range; | |
1732 | ||
1733 | user_uffdio_copy = (struct uffdio_copy __user *) arg; | |
1734 | ||
df2cc96e | 1735 | ret = -EAGAIN; |
a759a909 | 1736 | if (atomic_read(&ctx->mmap_changing)) |
df2cc96e MR |
1737 | goto out; |
1738 | ||
ad465cae AA |
1739 | ret = -EFAULT; |
1740 | if (copy_from_user(&uffdio_copy, user_uffdio_copy, | |
1741 | /* don't copy "copy" last field */ | |
1742 | sizeof(uffdio_copy)-sizeof(__s64))) | |
1743 | goto out; | |
1744 | ||
e71e2ace | 1745 | ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len); |
ad465cae AA |
1746 | if (ret) |
1747 | goto out; | |
1748 | /* | |
1749 | * double check for wraparound just in case. copy_from_user() | |
1750 | * will later check uffdio_copy.src + uffdio_copy.len to fit | |
1751 | * in the userland range. | |
1752 | */ | |
1753 | ret = -EINVAL; | |
1754 | if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src) | |
1755 | goto out; | |
72981e0e | 1756 | if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP)) |
ad465cae | 1757 | goto out; |
d2005e3f | 1758 | if (mmget_not_zero(ctx->mm)) { |
a734991c AR |
1759 | ret = mfill_atomic_copy(ctx->mm, uffdio_copy.dst, uffdio_copy.src, |
1760 | uffdio_copy.len, &ctx->mmap_changing, | |
1761 | uffdio_copy.mode); | |
d2005e3f | 1762 | mmput(ctx->mm); |
96333187 | 1763 | } else { |
e86b298b | 1764 | return -ESRCH; |
d2005e3f | 1765 | } |
ad465cae AA |
1766 | if (unlikely(put_user(ret, &user_uffdio_copy->copy))) |
1767 | return -EFAULT; | |
1768 | if (ret < 0) | |
1769 | goto out; | |
1770 | BUG_ON(!ret); | |
1771 | /* len == 0 would wake all */ | |
1772 | range.len = ret; | |
1773 | if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) { | |
1774 | range.start = uffdio_copy.dst; | |
1775 | wake_userfault(ctx, &range); | |
1776 | } | |
1777 | ret = range.len == uffdio_copy.len ? 0 : -EAGAIN; | |
1778 | out: | |
1779 | return ret; | |
1780 | } | |
1781 | ||
1782 | static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx, | |
1783 | unsigned long arg) | |
1784 | { | |
1785 | __s64 ret; | |
1786 | struct uffdio_zeropage uffdio_zeropage; | |
1787 | struct uffdio_zeropage __user *user_uffdio_zeropage; | |
1788 | struct userfaultfd_wake_range range; | |
1789 | ||
1790 | user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg; | |
1791 | ||
df2cc96e | 1792 | ret = -EAGAIN; |
a759a909 | 1793 | if (atomic_read(&ctx->mmap_changing)) |
df2cc96e MR |
1794 | goto out; |
1795 | ||
ad465cae AA |
1796 | ret = -EFAULT; |
1797 | if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage, | |
1798 | /* don't copy "zeropage" last field */ | |
1799 | sizeof(uffdio_zeropage)-sizeof(__s64))) | |
1800 | goto out; | |
1801 | ||
e71e2ace | 1802 | ret = validate_range(ctx->mm, uffdio_zeropage.range.start, |
ad465cae AA |
1803 | uffdio_zeropage.range.len); |
1804 | if (ret) | |
1805 | goto out; | |
1806 | ret = -EINVAL; | |
1807 | if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE) | |
1808 | goto out; | |
1809 | ||
d2005e3f | 1810 | if (mmget_not_zero(ctx->mm)) { |
a734991c AR |
1811 | ret = mfill_atomic_zeropage(ctx->mm, uffdio_zeropage.range.start, |
1812 | uffdio_zeropage.range.len, | |
1813 | &ctx->mmap_changing); | |
d2005e3f | 1814 | mmput(ctx->mm); |
9d95aa4b | 1815 | } else { |
e86b298b | 1816 | return -ESRCH; |
d2005e3f | 1817 | } |
ad465cae AA |
1818 | if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage))) |
1819 | return -EFAULT; | |
1820 | if (ret < 0) | |
1821 | goto out; | |
1822 | /* len == 0 would wake all */ | |
1823 | BUG_ON(!ret); | |
1824 | range.len = ret; | |
1825 | if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) { | |
1826 | range.start = uffdio_zeropage.range.start; | |
1827 | wake_userfault(ctx, &range); | |
1828 | } | |
1829 | ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN; | |
1830 | out: | |
1831 | return ret; | |
1832 | } | |
1833 | ||
63b2d417 AA |
1834 | static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx, |
1835 | unsigned long arg) | |
1836 | { | |
1837 | int ret; | |
1838 | struct uffdio_writeprotect uffdio_wp; | |
1839 | struct uffdio_writeprotect __user *user_uffdio_wp; | |
1840 | struct userfaultfd_wake_range range; | |
23080e27 | 1841 | bool mode_wp, mode_dontwake; |
63b2d417 | 1842 | |
a759a909 | 1843 | if (atomic_read(&ctx->mmap_changing)) |
63b2d417 AA |
1844 | return -EAGAIN; |
1845 | ||
1846 | user_uffdio_wp = (struct uffdio_writeprotect __user *) arg; | |
1847 | ||
1848 | if (copy_from_user(&uffdio_wp, user_uffdio_wp, | |
1849 | sizeof(struct uffdio_writeprotect))) | |
1850 | return -EFAULT; | |
1851 | ||
e71e2ace | 1852 | ret = validate_range(ctx->mm, uffdio_wp.range.start, |
63b2d417 AA |
1853 | uffdio_wp.range.len); |
1854 | if (ret) | |
1855 | return ret; | |
1856 | ||
1857 | if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE | | |
1858 | UFFDIO_WRITEPROTECT_MODE_WP)) | |
1859 | return -EINVAL; | |
23080e27 PX |
1860 | |
1861 | mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP; | |
1862 | mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE; | |
1863 | ||
1864 | if (mode_wp && mode_dontwake) | |
63b2d417 AA |
1865 | return -EINVAL; |
1866 | ||
cb185d5f NA |
1867 | if (mmget_not_zero(ctx->mm)) { |
1868 | ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start, | |
1869 | uffdio_wp.range.len, mode_wp, | |
1870 | &ctx->mmap_changing); | |
1871 | mmput(ctx->mm); | |
1872 | } else { | |
1873 | return -ESRCH; | |
1874 | } | |
1875 | ||
63b2d417 AA |
1876 | if (ret) |
1877 | return ret; | |
1878 | ||
23080e27 | 1879 | if (!mode_wp && !mode_dontwake) { |
63b2d417 AA |
1880 | range.start = uffdio_wp.range.start; |
1881 | range.len = uffdio_wp.range.len; | |
1882 | wake_userfault(ctx, &range); | |
1883 | } | |
1884 | return ret; | |
1885 | } | |
1886 | ||
f6191471 AR |
1887 | static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg) |
1888 | { | |
1889 | __s64 ret; | |
1890 | struct uffdio_continue uffdio_continue; | |
1891 | struct uffdio_continue __user *user_uffdio_continue; | |
1892 | struct userfaultfd_wake_range range; | |
1893 | ||
1894 | user_uffdio_continue = (struct uffdio_continue __user *)arg; | |
1895 | ||
1896 | ret = -EAGAIN; | |
a759a909 | 1897 | if (atomic_read(&ctx->mmap_changing)) |
f6191471 AR |
1898 | goto out; |
1899 | ||
1900 | ret = -EFAULT; | |
1901 | if (copy_from_user(&uffdio_continue, user_uffdio_continue, | |
1902 | /* don't copy the output fields */ | |
1903 | sizeof(uffdio_continue) - (sizeof(__s64)))) | |
1904 | goto out; | |
1905 | ||
e71e2ace | 1906 | ret = validate_range(ctx->mm, uffdio_continue.range.start, |
f6191471 AR |
1907 | uffdio_continue.range.len); |
1908 | if (ret) | |
1909 | goto out; | |
1910 | ||
1911 | ret = -EINVAL; | |
1912 | /* double check for wraparound just in case. */ | |
1913 | if (uffdio_continue.range.start + uffdio_continue.range.len <= | |
1914 | uffdio_continue.range.start) { | |
1915 | goto out; | |
1916 | } | |
1917 | if (uffdio_continue.mode & ~UFFDIO_CONTINUE_MODE_DONTWAKE) | |
1918 | goto out; | |
1919 | ||
1920 | if (mmget_not_zero(ctx->mm)) { | |
a734991c AR |
1921 | ret = mfill_atomic_continue(ctx->mm, uffdio_continue.range.start, |
1922 | uffdio_continue.range.len, | |
1923 | &ctx->mmap_changing); | |
f6191471 AR |
1924 | mmput(ctx->mm); |
1925 | } else { | |
1926 | return -ESRCH; | |
1927 | } | |
1928 | ||
1929 | if (unlikely(put_user(ret, &user_uffdio_continue->mapped))) | |
1930 | return -EFAULT; | |
1931 | if (ret < 0) | |
1932 | goto out; | |
1933 | ||
1934 | /* len == 0 would wake all */ | |
1935 | BUG_ON(!ret); | |
1936 | range.len = ret; | |
1937 | if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) { | |
1938 | range.start = uffdio_continue.range.start; | |
1939 | wake_userfault(ctx, &range); | |
1940 | } | |
1941 | ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN; | |
1942 | ||
1943 | out: | |
1944 | return ret; | |
1945 | } | |
1946 | ||
9cd75c3c PE |
1947 | static inline unsigned int uffd_ctx_features(__u64 user_features) |
1948 | { | |
1949 | /* | |
22e5fe2a NA |
1950 | * For the current set of features the bits just coincide. Set |
1951 | * UFFD_FEATURE_INITIALIZED to mark the features as enabled. | |
9cd75c3c | 1952 | */ |
22e5fe2a | 1953 | return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED; |
9cd75c3c PE |
1954 | } |
1955 | ||
86039bd3 AA |
1956 | /* |
1957 | * userland asks for a certain API version and we return which bits | |
1958 | * and ioctl commands are implemented in this kernel for such API | |
1959 | * version or -EINVAL if unknown. | |
1960 | */ | |
1961 | static int userfaultfd_api(struct userfaultfd_ctx *ctx, | |
1962 | unsigned long arg) | |
1963 | { | |
1964 | struct uffdio_api uffdio_api; | |
1965 | void __user *buf = (void __user *)arg; | |
22e5fe2a | 1966 | unsigned int ctx_features; |
86039bd3 | 1967 | int ret; |
65603144 | 1968 | __u64 features; |
86039bd3 | 1969 | |
86039bd3 | 1970 | ret = -EFAULT; |
a9b85f94 | 1971 | if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api))) |
86039bd3 | 1972 | goto out; |
914eedcb AR |
1973 | /* Ignore unsupported features (userspace built against newer kernel) */ |
1974 | features = uffdio_api.features & UFFD_API_FEATURES; | |
3c1c24d9 MR |
1975 | ret = -EPERM; |
1976 | if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE)) | |
1977 | goto err_out; | |
65603144 AA |
1978 | /* report all available features and ioctls to userland */ |
1979 | uffdio_api.features = UFFD_API_FEATURES; | |
7677f7fd | 1980 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR |
964ab004 AR |
1981 | uffdio_api.features &= |
1982 | ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM); | |
00b151f2 PX |
1983 | #endif |
1984 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP | |
1985 | uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP; | |
b1f9e876 PX |
1986 | #endif |
1987 | #ifndef CONFIG_PTE_MARKER_UFFD_WP | |
1988 | uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM; | |
2bad466c | 1989 | uffdio_api.features &= ~UFFD_FEATURE_WP_UNPOPULATED; |
7677f7fd | 1990 | #endif |
86039bd3 AA |
1991 | uffdio_api.ioctls = UFFD_API_IOCTLS; |
1992 | ret = -EFAULT; | |
1993 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1994 | goto out; | |
22e5fe2a | 1995 | |
65603144 | 1996 | /* only enable the requested features for this uffd context */ |
22e5fe2a NA |
1997 | ctx_features = uffd_ctx_features(features); |
1998 | ret = -EINVAL; | |
1999 | if (cmpxchg(&ctx->features, 0, ctx_features) != 0) | |
2000 | goto err_out; | |
2001 | ||
86039bd3 AA |
2002 | ret = 0; |
2003 | out: | |
2004 | return ret; | |
3c1c24d9 MR |
2005 | err_out: |
2006 | memset(&uffdio_api, 0, sizeof(uffdio_api)); | |
2007 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
2008 | ret = -EFAULT; | |
2009 | goto out; | |
86039bd3 AA |
2010 | } |
2011 | ||
2012 | static long userfaultfd_ioctl(struct file *file, unsigned cmd, | |
2013 | unsigned long arg) | |
2014 | { | |
2015 | int ret = -EINVAL; | |
2016 | struct userfaultfd_ctx *ctx = file->private_data; | |
2017 | ||
22e5fe2a | 2018 | if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx)) |
e6485a47 AA |
2019 | return -EINVAL; |
2020 | ||
86039bd3 AA |
2021 | switch(cmd) { |
2022 | case UFFDIO_API: | |
2023 | ret = userfaultfd_api(ctx, arg); | |
2024 | break; | |
2025 | case UFFDIO_REGISTER: | |
2026 | ret = userfaultfd_register(ctx, arg); | |
2027 | break; | |
2028 | case UFFDIO_UNREGISTER: | |
2029 | ret = userfaultfd_unregister(ctx, arg); | |
2030 | break; | |
2031 | case UFFDIO_WAKE: | |
2032 | ret = userfaultfd_wake(ctx, arg); | |
2033 | break; | |
ad465cae AA |
2034 | case UFFDIO_COPY: |
2035 | ret = userfaultfd_copy(ctx, arg); | |
2036 | break; | |
2037 | case UFFDIO_ZEROPAGE: | |
2038 | ret = userfaultfd_zeropage(ctx, arg); | |
2039 | break; | |
63b2d417 AA |
2040 | case UFFDIO_WRITEPROTECT: |
2041 | ret = userfaultfd_writeprotect(ctx, arg); | |
2042 | break; | |
f6191471 AR |
2043 | case UFFDIO_CONTINUE: |
2044 | ret = userfaultfd_continue(ctx, arg); | |
2045 | break; | |
86039bd3 AA |
2046 | } |
2047 | return ret; | |
2048 | } | |
2049 | ||
2050 | #ifdef CONFIG_PROC_FS | |
2051 | static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f) | |
2052 | { | |
2053 | struct userfaultfd_ctx *ctx = f->private_data; | |
ac6424b9 | 2054 | wait_queue_entry_t *wq; |
86039bd3 AA |
2055 | unsigned long pending = 0, total = 0; |
2056 | ||
cbcfa130 | 2057 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
2055da97 | 2058 | list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) { |
15b726ef AA |
2059 | pending++; |
2060 | total++; | |
2061 | } | |
2055da97 | 2062 | list_for_each_entry(wq, &ctx->fault_wqh.head, entry) { |
86039bd3 AA |
2063 | total++; |
2064 | } | |
cbcfa130 | 2065 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
2066 | |
2067 | /* | |
2068 | * If more protocols will be added, there will be all shown | |
2069 | * separated by a space. Like this: | |
2070 | * protocols: aa:... bb:... | |
2071 | */ | |
2072 | seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n", | |
045098e9 | 2073 | pending, total, UFFD_API, ctx->features, |
86039bd3 AA |
2074 | UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS); |
2075 | } | |
2076 | #endif | |
2077 | ||
2078 | static const struct file_operations userfaultfd_fops = { | |
2079 | #ifdef CONFIG_PROC_FS | |
2080 | .show_fdinfo = userfaultfd_show_fdinfo, | |
2081 | #endif | |
2082 | .release = userfaultfd_release, | |
2083 | .poll = userfaultfd_poll, | |
2084 | .read = userfaultfd_read, | |
2085 | .unlocked_ioctl = userfaultfd_ioctl, | |
1832f2d8 | 2086 | .compat_ioctl = compat_ptr_ioctl, |
86039bd3 AA |
2087 | .llseek = noop_llseek, |
2088 | }; | |
2089 | ||
3004ec9c AA |
2090 | static void init_once_userfaultfd_ctx(void *mem) |
2091 | { | |
2092 | struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem; | |
2093 | ||
2094 | init_waitqueue_head(&ctx->fault_pending_wqh); | |
2095 | init_waitqueue_head(&ctx->fault_wqh); | |
9cd75c3c | 2096 | init_waitqueue_head(&ctx->event_wqh); |
3004ec9c | 2097 | init_waitqueue_head(&ctx->fd_wqh); |
2ca97ac8 | 2098 | seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock); |
3004ec9c AA |
2099 | } |
2100 | ||
2d5de004 | 2101 | static int new_userfaultfd(int flags) |
86039bd3 | 2102 | { |
86039bd3 | 2103 | struct userfaultfd_ctx *ctx; |
284cd241 | 2104 | int fd; |
86039bd3 AA |
2105 | |
2106 | BUG_ON(!current->mm); | |
2107 | ||
2108 | /* Check the UFFD_* constants for consistency. */ | |
37cd0575 | 2109 | BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS); |
86039bd3 AA |
2110 | BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC); |
2111 | BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK); | |
2112 | ||
37cd0575 | 2113 | if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY)) |
284cd241 | 2114 | return -EINVAL; |
86039bd3 | 2115 | |
3004ec9c | 2116 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); |
86039bd3 | 2117 | if (!ctx) |
284cd241 | 2118 | return -ENOMEM; |
86039bd3 | 2119 | |
ca880420 | 2120 | refcount_set(&ctx->refcount, 1); |
86039bd3 | 2121 | ctx->flags = flags; |
9cd75c3c | 2122 | ctx->features = 0; |
86039bd3 | 2123 | ctx->released = false; |
a759a909 | 2124 | atomic_set(&ctx->mmap_changing, 0); |
86039bd3 AA |
2125 | ctx->mm = current->mm; |
2126 | /* prevent the mm struct to be freed */ | |
f1f10076 | 2127 | mmgrab(ctx->mm); |
86039bd3 | 2128 | |
b537900f | 2129 | fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx, |
abec3d01 | 2130 | O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL); |
284cd241 | 2131 | if (fd < 0) { |
d2005e3f | 2132 | mmdrop(ctx->mm); |
3004ec9c | 2133 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
c03e946f | 2134 | } |
86039bd3 | 2135 | return fd; |
86039bd3 | 2136 | } |
3004ec9c | 2137 | |
2d5de004 AR |
2138 | static inline bool userfaultfd_syscall_allowed(int flags) |
2139 | { | |
2140 | /* Userspace-only page faults are always allowed */ | |
2141 | if (flags & UFFD_USER_MODE_ONLY) | |
2142 | return true; | |
2143 | ||
2144 | /* | |
2145 | * The user is requesting a userfaultfd which can handle kernel faults. | |
2146 | * Privileged users are always allowed to do this. | |
2147 | */ | |
2148 | if (capable(CAP_SYS_PTRACE)) | |
2149 | return true; | |
2150 | ||
2151 | /* Otherwise, access to kernel fault handling is sysctl controlled. */ | |
2152 | return sysctl_unprivileged_userfaultfd; | |
2153 | } | |
2154 | ||
2155 | SYSCALL_DEFINE1(userfaultfd, int, flags) | |
2156 | { | |
2157 | if (!userfaultfd_syscall_allowed(flags)) | |
2158 | return -EPERM; | |
2159 | ||
2160 | return new_userfaultfd(flags); | |
2161 | } | |
2162 | ||
2163 | static long userfaultfd_dev_ioctl(struct file *file, unsigned int cmd, unsigned long flags) | |
2164 | { | |
2165 | if (cmd != USERFAULTFD_IOC_NEW) | |
2166 | return -EINVAL; | |
2167 | ||
2168 | return new_userfaultfd(flags); | |
2169 | } | |
2170 | ||
2171 | static const struct file_operations userfaultfd_dev_fops = { | |
2172 | .unlocked_ioctl = userfaultfd_dev_ioctl, | |
2173 | .compat_ioctl = userfaultfd_dev_ioctl, | |
2174 | .owner = THIS_MODULE, | |
2175 | .llseek = noop_llseek, | |
2176 | }; | |
2177 | ||
2178 | static struct miscdevice userfaultfd_misc = { | |
2179 | .minor = MISC_DYNAMIC_MINOR, | |
2180 | .name = "userfaultfd", | |
2181 | .fops = &userfaultfd_dev_fops | |
2182 | }; | |
2183 | ||
3004ec9c AA |
2184 | static int __init userfaultfd_init(void) |
2185 | { | |
2d5de004 AR |
2186 | int ret; |
2187 | ||
2188 | ret = misc_register(&userfaultfd_misc); | |
2189 | if (ret) | |
2190 | return ret; | |
2191 | ||
3004ec9c AA |
2192 | userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache", |
2193 | sizeof(struct userfaultfd_ctx), | |
2194 | 0, | |
2195 | SLAB_HWCACHE_ALIGN|SLAB_PANIC, | |
2196 | init_once_userfaultfd_ctx); | |
2197 | return 0; | |
2198 | } | |
2199 | __initcall(userfaultfd_init); |