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 | |
2d337b71 Z |
35 | static int sysctl_unprivileged_userfaultfd __read_mostly; |
36 | ||
37 | #ifdef CONFIG_SYSCTL | |
38 | static struct ctl_table vm_userfaultfd_table[] = { | |
39 | { | |
40 | .procname = "unprivileged_userfaultfd", | |
41 | .data = &sysctl_unprivileged_userfaultfd, | |
42 | .maxlen = sizeof(sysctl_unprivileged_userfaultfd), | |
43 | .mode = 0644, | |
44 | .proc_handler = proc_dointvec_minmax, | |
45 | .extra1 = SYSCTL_ZERO, | |
46 | .extra2 = SYSCTL_ONE, | |
47 | }, | |
48 | { } | |
49 | }; | |
50 | #endif | |
cefdca0a | 51 | |
3004ec9c AA |
52 | static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly; |
53 | ||
3004ec9c AA |
54 | /* |
55 | * Start with fault_pending_wqh and fault_wqh so they're more likely | |
56 | * to be in the same cacheline. | |
cbcfa130 EB |
57 | * |
58 | * Locking order: | |
59 | * fd_wqh.lock | |
60 | * fault_pending_wqh.lock | |
61 | * fault_wqh.lock | |
62 | * event_wqh.lock | |
63 | * | |
64 | * To avoid deadlocks, IRQs must be disabled when taking any of the above locks, | |
65 | * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's | |
66 | * also taken in IRQ context. | |
3004ec9c | 67 | */ |
86039bd3 | 68 | struct userfaultfd_ctx { |
15b726ef AA |
69 | /* waitqueue head for the pending (i.e. not read) userfaults */ |
70 | wait_queue_head_t fault_pending_wqh; | |
71 | /* waitqueue head for the userfaults */ | |
86039bd3 AA |
72 | wait_queue_head_t fault_wqh; |
73 | /* waitqueue head for the pseudo fd to wakeup poll/read */ | |
74 | wait_queue_head_t fd_wqh; | |
9cd75c3c PE |
75 | /* waitqueue head for events */ |
76 | wait_queue_head_t event_wqh; | |
2c5b7e1b | 77 | /* a refile sequence protected by fault_pending_wqh lock */ |
2ca97ac8 | 78 | seqcount_spinlock_t refile_seq; |
3004ec9c | 79 | /* pseudo fd refcounting */ |
ca880420 | 80 | refcount_t refcount; |
86039bd3 AA |
81 | /* userfaultfd syscall flags */ |
82 | unsigned int flags; | |
9cd75c3c PE |
83 | /* features requested from the userspace */ |
84 | unsigned int features; | |
86039bd3 AA |
85 | /* released */ |
86 | bool released; | |
df2cc96e | 87 | /* memory mappings are changing because of non-cooperative event */ |
a759a909 | 88 | atomic_t mmap_changing; |
86039bd3 AA |
89 | /* mm with one ore more vmas attached to this userfaultfd_ctx */ |
90 | struct mm_struct *mm; | |
91 | }; | |
92 | ||
893e26e6 PE |
93 | struct userfaultfd_fork_ctx { |
94 | struct userfaultfd_ctx *orig; | |
95 | struct userfaultfd_ctx *new; | |
96 | struct list_head list; | |
97 | }; | |
98 | ||
897ab3e0 MR |
99 | struct userfaultfd_unmap_ctx { |
100 | struct userfaultfd_ctx *ctx; | |
101 | unsigned long start; | |
102 | unsigned long end; | |
103 | struct list_head list; | |
104 | }; | |
105 | ||
86039bd3 | 106 | struct userfaultfd_wait_queue { |
a9b85f94 | 107 | struct uffd_msg msg; |
ac6424b9 | 108 | wait_queue_entry_t wq; |
86039bd3 | 109 | struct userfaultfd_ctx *ctx; |
15a77c6f | 110 | bool waken; |
86039bd3 AA |
111 | }; |
112 | ||
113 | struct userfaultfd_wake_range { | |
114 | unsigned long start; | |
115 | unsigned long len; | |
116 | }; | |
117 | ||
22e5fe2a NA |
118 | /* internal indication that UFFD_API ioctl was successfully executed */ |
119 | #define UFFD_FEATURE_INITIALIZED (1u << 31) | |
120 | ||
121 | static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx) | |
122 | { | |
123 | return ctx->features & UFFD_FEATURE_INITIALIZED; | |
124 | } | |
125 | ||
2bad466c PX |
126 | /* |
127 | * Whether WP_UNPOPULATED is enabled on the uffd context. It is only | |
128 | * meaningful when userfaultfd_wp()==true on the vma and when it's | |
129 | * anonymous. | |
130 | */ | |
131 | bool userfaultfd_wp_unpopulated(struct vm_area_struct *vma) | |
132 | { | |
133 | struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx; | |
134 | ||
135 | if (!ctx) | |
136 | return false; | |
137 | ||
138 | return ctx->features & UFFD_FEATURE_WP_UNPOPULATED; | |
139 | } | |
140 | ||
51d3d5eb DH |
141 | static void userfaultfd_set_vm_flags(struct vm_area_struct *vma, |
142 | vm_flags_t flags) | |
143 | { | |
144 | const bool uffd_wp_changed = (vma->vm_flags ^ flags) & VM_UFFD_WP; | |
145 | ||
1c71222e | 146 | vm_flags_reset(vma, flags); |
51d3d5eb DH |
147 | /* |
148 | * For shared mappings, we want to enable writenotify while | |
149 | * userfaultfd-wp is enabled (see vma_wants_writenotify()). We'll simply | |
150 | * recalculate vma->vm_page_prot whenever userfaultfd-wp changes. | |
151 | */ | |
152 | if ((vma->vm_flags & VM_SHARED) && uffd_wp_changed) | |
153 | vma_set_page_prot(vma); | |
154 | } | |
155 | ||
ac6424b9 | 156 | static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode, |
86039bd3 AA |
157 | int wake_flags, void *key) |
158 | { | |
159 | struct userfaultfd_wake_range *range = key; | |
160 | int ret; | |
161 | struct userfaultfd_wait_queue *uwq; | |
162 | unsigned long start, len; | |
163 | ||
164 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
165 | ret = 0; | |
86039bd3 AA |
166 | /* len == 0 means wake all */ |
167 | start = range->start; | |
168 | len = range->len; | |
a9b85f94 AA |
169 | if (len && (start > uwq->msg.arg.pagefault.address || |
170 | start + len <= uwq->msg.arg.pagefault.address)) | |
86039bd3 | 171 | goto out; |
15a77c6f AA |
172 | WRITE_ONCE(uwq->waken, true); |
173 | /* | |
a9668cd6 PZ |
174 | * The Program-Order guarantees provided by the scheduler |
175 | * ensure uwq->waken is visible before the task is woken. | |
15a77c6f | 176 | */ |
86039bd3 | 177 | ret = wake_up_state(wq->private, mode); |
a9668cd6 | 178 | if (ret) { |
86039bd3 AA |
179 | /* |
180 | * Wake only once, autoremove behavior. | |
181 | * | |
a9668cd6 PZ |
182 | * After the effect of list_del_init is visible to the other |
183 | * CPUs, the waitqueue may disappear from under us, see the | |
184 | * !list_empty_careful() in handle_userfault(). | |
185 | * | |
186 | * try_to_wake_up() has an implicit smp_mb(), and the | |
187 | * wq->private is read before calling the extern function | |
188 | * "wake_up_state" (which in turns calls try_to_wake_up). | |
86039bd3 | 189 | */ |
2055da97 | 190 | list_del_init(&wq->entry); |
a9668cd6 | 191 | } |
86039bd3 AA |
192 | out: |
193 | return ret; | |
194 | } | |
195 | ||
196 | /** | |
197 | * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd | |
198 | * context. | |
199 | * @ctx: [in] Pointer to the userfaultfd context. | |
86039bd3 AA |
200 | */ |
201 | static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx) | |
202 | { | |
ca880420 | 203 | refcount_inc(&ctx->refcount); |
86039bd3 AA |
204 | } |
205 | ||
206 | /** | |
207 | * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd | |
208 | * context. | |
209 | * @ctx: [in] Pointer to userfaultfd context. | |
210 | * | |
211 | * The userfaultfd context reference must have been previously acquired either | |
212 | * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget(). | |
213 | */ | |
214 | static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx) | |
215 | { | |
ca880420 | 216 | if (refcount_dec_and_test(&ctx->refcount)) { |
86039bd3 AA |
217 | VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock)); |
218 | VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh)); | |
219 | VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock)); | |
220 | VM_BUG_ON(waitqueue_active(&ctx->fault_wqh)); | |
9cd75c3c PE |
221 | VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock)); |
222 | VM_BUG_ON(waitqueue_active(&ctx->event_wqh)); | |
86039bd3 AA |
223 | VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock)); |
224 | VM_BUG_ON(waitqueue_active(&ctx->fd_wqh)); | |
d2005e3f | 225 | mmdrop(ctx->mm); |
3004ec9c | 226 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
86039bd3 AA |
227 | } |
228 | } | |
229 | ||
a9b85f94 | 230 | static inline void msg_init(struct uffd_msg *msg) |
86039bd3 | 231 | { |
a9b85f94 AA |
232 | BUILD_BUG_ON(sizeof(struct uffd_msg) != 32); |
233 | /* | |
234 | * Must use memset to zero out the paddings or kernel data is | |
235 | * leaked to userland. | |
236 | */ | |
237 | memset(msg, 0, sizeof(struct uffd_msg)); | |
238 | } | |
239 | ||
240 | static inline struct uffd_msg userfault_msg(unsigned long address, | |
d172b1a3 | 241 | unsigned long real_address, |
a9b85f94 | 242 | unsigned int flags, |
9d4ac934 AP |
243 | unsigned long reason, |
244 | unsigned int features) | |
a9b85f94 AA |
245 | { |
246 | struct uffd_msg msg; | |
d172b1a3 | 247 | |
a9b85f94 AA |
248 | msg_init(&msg); |
249 | msg.event = UFFD_EVENT_PAGEFAULT; | |
824ddc60 | 250 | |
d172b1a3 NA |
251 | msg.arg.pagefault.address = (features & UFFD_FEATURE_EXACT_ADDRESS) ? |
252 | real_address : address; | |
253 | ||
7677f7fd AR |
254 | /* |
255 | * These flags indicate why the userfault occurred: | |
256 | * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault. | |
257 | * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault. | |
258 | * - Neither of these flags being set indicates a MISSING fault. | |
259 | * | |
260 | * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write | |
261 | * fault. Otherwise, it was a read fault. | |
262 | */ | |
86039bd3 | 263 | if (flags & FAULT_FLAG_WRITE) |
a9b85f94 | 264 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE; |
86039bd3 | 265 | if (reason & VM_UFFD_WP) |
a9b85f94 | 266 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP; |
7677f7fd AR |
267 | if (reason & VM_UFFD_MINOR) |
268 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR; | |
9d4ac934 | 269 | if (features & UFFD_FEATURE_THREAD_ID) |
a36985d3 | 270 | msg.arg.pagefault.feat.ptid = task_pid_vnr(current); |
a9b85f94 | 271 | return msg; |
86039bd3 AA |
272 | } |
273 | ||
369cd212 MK |
274 | #ifdef CONFIG_HUGETLB_PAGE |
275 | /* | |
276 | * Same functionality as userfaultfd_must_wait below with modifications for | |
277 | * hugepmd ranges. | |
278 | */ | |
279 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 280 | struct vm_area_struct *vma, |
369cd212 MK |
281 | unsigned long address, |
282 | unsigned long flags, | |
283 | unsigned long reason) | |
284 | { | |
1e2c0436 | 285 | pte_t *ptep, pte; |
369cd212 MK |
286 | bool ret = true; |
287 | ||
9c67a207 | 288 | mmap_assert_locked(ctx->mm); |
1e2c0436 | 289 | |
9c67a207 | 290 | ptep = hugetlb_walk(vma, address, vma_mmu_pagesize(vma)); |
1e2c0436 | 291 | if (!ptep) |
369cd212 MK |
292 | goto out; |
293 | ||
294 | ret = false; | |
1e2c0436 | 295 | pte = huge_ptep_get(ptep); |
369cd212 MK |
296 | |
297 | /* | |
298 | * Lockless access: we're in a wait_event so it's ok if it | |
5c041f5d PX |
299 | * changes under us. PTE markers should be handled the same as none |
300 | * ptes here. | |
369cd212 | 301 | */ |
5c041f5d | 302 | if (huge_pte_none_mostly(pte)) |
369cd212 | 303 | ret = true; |
1e2c0436 | 304 | if (!huge_pte_write(pte) && (reason & VM_UFFD_WP)) |
369cd212 MK |
305 | ret = true; |
306 | out: | |
307 | return ret; | |
308 | } | |
309 | #else | |
310 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 311 | struct vm_area_struct *vma, |
369cd212 MK |
312 | unsigned long address, |
313 | unsigned long flags, | |
314 | unsigned long reason) | |
315 | { | |
316 | return false; /* should never get here */ | |
317 | } | |
318 | #endif /* CONFIG_HUGETLB_PAGE */ | |
319 | ||
8d2afd96 AA |
320 | /* |
321 | * Verify the pagetables are still not ok after having reigstered into | |
322 | * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any | |
323 | * userfault that has already been resolved, if userfaultfd_read and | |
324 | * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different | |
325 | * threads. | |
326 | */ | |
327 | static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx, | |
328 | unsigned long address, | |
329 | unsigned long flags, | |
330 | unsigned long reason) | |
331 | { | |
332 | struct mm_struct *mm = ctx->mm; | |
333 | pgd_t *pgd; | |
c2febafc | 334 | p4d_t *p4d; |
8d2afd96 AA |
335 | pud_t *pud; |
336 | pmd_t *pmd, _pmd; | |
337 | pte_t *pte; | |
338 | bool ret = true; | |
339 | ||
42fc5414 | 340 | mmap_assert_locked(mm); |
8d2afd96 AA |
341 | |
342 | pgd = pgd_offset(mm, address); | |
343 | if (!pgd_present(*pgd)) | |
344 | goto out; | |
c2febafc KS |
345 | p4d = p4d_offset(pgd, address); |
346 | if (!p4d_present(*p4d)) | |
347 | goto out; | |
348 | pud = pud_offset(p4d, address); | |
8d2afd96 AA |
349 | if (!pud_present(*pud)) |
350 | goto out; | |
351 | pmd = pmd_offset(pud, address); | |
352 | /* | |
353 | * READ_ONCE must function as a barrier with narrower scope | |
354 | * and it must be equivalent to: | |
355 | * _pmd = *pmd; barrier(); | |
356 | * | |
357 | * This is to deal with the instability (as in | |
358 | * pmd_trans_unstable) of the pmd. | |
359 | */ | |
360 | _pmd = READ_ONCE(*pmd); | |
a365ac09 | 361 | if (pmd_none(_pmd)) |
8d2afd96 AA |
362 | goto out; |
363 | ||
364 | ret = false; | |
a365ac09 HY |
365 | if (!pmd_present(_pmd)) |
366 | goto out; | |
367 | ||
63b2d417 AA |
368 | if (pmd_trans_huge(_pmd)) { |
369 | if (!pmd_write(_pmd) && (reason & VM_UFFD_WP)) | |
370 | ret = true; | |
8d2afd96 | 371 | goto out; |
63b2d417 | 372 | } |
8d2afd96 AA |
373 | |
374 | /* | |
375 | * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it | |
376 | * and use the standard pte_offset_map() instead of parsing _pmd. | |
377 | */ | |
378 | pte = pte_offset_map(pmd, address); | |
379 | /* | |
380 | * Lockless access: we're in a wait_event so it's ok if it | |
5c041f5d PX |
381 | * changes under us. PTE markers should be handled the same as none |
382 | * ptes here. | |
8d2afd96 | 383 | */ |
5c041f5d | 384 | if (pte_none_mostly(*pte)) |
8d2afd96 | 385 | ret = true; |
63b2d417 AA |
386 | if (!pte_write(*pte) && (reason & VM_UFFD_WP)) |
387 | ret = true; | |
8d2afd96 AA |
388 | pte_unmap(pte); |
389 | ||
390 | out: | |
391 | return ret; | |
392 | } | |
393 | ||
2f064a59 | 394 | static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags) |
3e69ad08 PX |
395 | { |
396 | if (flags & FAULT_FLAG_INTERRUPTIBLE) | |
397 | return TASK_INTERRUPTIBLE; | |
398 | ||
399 | if (flags & FAULT_FLAG_KILLABLE) | |
400 | return TASK_KILLABLE; | |
401 | ||
402 | return TASK_UNINTERRUPTIBLE; | |
403 | } | |
404 | ||
86039bd3 AA |
405 | /* |
406 | * The locking rules involved in returning VM_FAULT_RETRY depending on | |
407 | * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and | |
408 | * FAULT_FLAG_KILLABLE are not straightforward. The "Caution" | |
409 | * recommendation in __lock_page_or_retry is not an understatement. | |
410 | * | |
c1e8d7c6 | 411 | * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released |
86039bd3 AA |
412 | * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is |
413 | * not set. | |
414 | * | |
415 | * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not | |
416 | * set, VM_FAULT_RETRY can still be returned if and only if there are | |
c1e8d7c6 | 417 | * fatal_signal_pending()s, and the mmap_lock must be released before |
86039bd3 AA |
418 | * returning it. |
419 | */ | |
2b740303 | 420 | vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason) |
86039bd3 | 421 | { |
b8da2e46 PX |
422 | struct vm_area_struct *vma = vmf->vma; |
423 | struct mm_struct *mm = vma->vm_mm; | |
86039bd3 AA |
424 | struct userfaultfd_ctx *ctx; |
425 | struct userfaultfd_wait_queue uwq; | |
2b740303 | 426 | vm_fault_t ret = VM_FAULT_SIGBUS; |
3e69ad08 | 427 | bool must_wait; |
2f064a59 | 428 | unsigned int blocking_state; |
86039bd3 | 429 | |
64c2b203 AA |
430 | /* |
431 | * We don't do userfault handling for the final child pid update. | |
432 | * | |
433 | * We also don't do userfault handling during | |
434 | * coredumping. hugetlbfs has the special | |
435 | * follow_hugetlb_page() to skip missing pages in the | |
436 | * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with | |
437 | * the no_page_table() helper in follow_page_mask(), but the | |
438 | * shmem_vm_ops->fault method is invoked even during | |
c1e8d7c6 | 439 | * coredumping without mmap_lock and it ends up here. |
64c2b203 AA |
440 | */ |
441 | if (current->flags & (PF_EXITING|PF_DUMPCORE)) | |
442 | goto out; | |
443 | ||
444 | /* | |
c1e8d7c6 ML |
445 | * Coredumping runs without mmap_lock so we can only check that |
446 | * the mmap_lock is held, if PF_DUMPCORE was not set. | |
64c2b203 | 447 | */ |
42fc5414 | 448 | mmap_assert_locked(mm); |
64c2b203 | 449 | |
b8da2e46 | 450 | ctx = vma->vm_userfaultfd_ctx.ctx; |
86039bd3 | 451 | if (!ctx) |
ba85c702 | 452 | goto out; |
86039bd3 AA |
453 | |
454 | BUG_ON(ctx->mm != mm); | |
455 | ||
7677f7fd AR |
456 | /* Any unrecognized flag is a bug. */ |
457 | VM_BUG_ON(reason & ~__VM_UFFD_FLAGS); | |
458 | /* 0 or > 1 flags set is a bug; we expect exactly 1. */ | |
459 | VM_BUG_ON(!reason || (reason & (reason - 1))); | |
86039bd3 | 460 | |
2d6d6f5a PS |
461 | if (ctx->features & UFFD_FEATURE_SIGBUS) |
462 | goto out; | |
2d5de004 | 463 | if (!(vmf->flags & FAULT_FLAG_USER) && (ctx->flags & UFFD_USER_MODE_ONLY)) |
37cd0575 | 464 | goto out; |
2d6d6f5a | 465 | |
86039bd3 AA |
466 | /* |
467 | * If it's already released don't get it. This avoids to loop | |
468 | * in __get_user_pages if userfaultfd_release waits on the | |
c1e8d7c6 | 469 | * caller of handle_userfault to release the mmap_lock. |
86039bd3 | 470 | */ |
6aa7de05 | 471 | if (unlikely(READ_ONCE(ctx->released))) { |
656710a6 AA |
472 | /* |
473 | * Don't return VM_FAULT_SIGBUS in this case, so a non | |
474 | * cooperative manager can close the uffd after the | |
475 | * last UFFDIO_COPY, without risking to trigger an | |
476 | * involuntary SIGBUS if the process was starting the | |
477 | * userfaultfd while the userfaultfd was still armed | |
478 | * (but after the last UFFDIO_COPY). If the uffd | |
479 | * wasn't already closed when the userfault reached | |
480 | * this point, that would normally be solved by | |
481 | * userfaultfd_must_wait returning 'false'. | |
482 | * | |
483 | * If we were to return VM_FAULT_SIGBUS here, the non | |
484 | * cooperative manager would be instead forced to | |
485 | * always call UFFDIO_UNREGISTER before it can safely | |
486 | * close the uffd. | |
487 | */ | |
488 | ret = VM_FAULT_NOPAGE; | |
ba85c702 | 489 | goto out; |
656710a6 | 490 | } |
86039bd3 AA |
491 | |
492 | /* | |
493 | * Check that we can return VM_FAULT_RETRY. | |
494 | * | |
495 | * NOTE: it should become possible to return VM_FAULT_RETRY | |
496 | * even if FAULT_FLAG_TRIED is set without leading to gup() | |
497 | * -EBUSY failures, if the userfaultfd is to be extended for | |
498 | * VM_UFFD_WP tracking and we intend to arm the userfault | |
499 | * without first stopping userland access to the memory. For | |
500 | * VM_UFFD_MISSING userfaults this is enough for now. | |
501 | */ | |
82b0f8c3 | 502 | if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) { |
86039bd3 AA |
503 | /* |
504 | * Validate the invariant that nowait must allow retry | |
505 | * to be sure not to return SIGBUS erroneously on | |
506 | * nowait invocations. | |
507 | */ | |
82b0f8c3 | 508 | BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT); |
86039bd3 AA |
509 | #ifdef CONFIG_DEBUG_VM |
510 | if (printk_ratelimit()) { | |
511 | printk(KERN_WARNING | |
82b0f8c3 JK |
512 | "FAULT_FLAG_ALLOW_RETRY missing %x\n", |
513 | vmf->flags); | |
86039bd3 AA |
514 | dump_stack(); |
515 | } | |
516 | #endif | |
ba85c702 | 517 | goto out; |
86039bd3 AA |
518 | } |
519 | ||
520 | /* | |
521 | * Handle nowait, not much to do other than tell it to retry | |
522 | * and wait. | |
523 | */ | |
ba85c702 | 524 | ret = VM_FAULT_RETRY; |
82b0f8c3 | 525 | if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) |
ba85c702 | 526 | goto out; |
86039bd3 | 527 | |
c1e8d7c6 | 528 | /* take the reference before dropping the mmap_lock */ |
86039bd3 AA |
529 | userfaultfd_ctx_get(ctx); |
530 | ||
86039bd3 AA |
531 | init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function); |
532 | uwq.wq.private = current; | |
d172b1a3 NA |
533 | uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags, |
534 | reason, ctx->features); | |
86039bd3 | 535 | uwq.ctx = ctx; |
15a77c6f | 536 | uwq.waken = false; |
86039bd3 | 537 | |
3e69ad08 | 538 | blocking_state = userfaultfd_get_blocking_state(vmf->flags); |
dfa37dc3 | 539 | |
b8da2e46 PX |
540 | /* |
541 | * Take the vma lock now, in order to safely call | |
542 | * userfaultfd_huge_must_wait() later. Since acquiring the | |
543 | * (sleepable) vma lock can modify the current task state, that | |
544 | * must be before explicitly calling set_current_state(). | |
545 | */ | |
546 | if (is_vm_hugetlb_page(vma)) | |
547 | hugetlb_vma_lock_read(vma); | |
548 | ||
cbcfa130 | 549 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
550 | /* |
551 | * After the __add_wait_queue the uwq is visible to userland | |
552 | * through poll/read(). | |
553 | */ | |
15b726ef AA |
554 | __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq); |
555 | /* | |
556 | * The smp_mb() after __set_current_state prevents the reads | |
557 | * following the spin_unlock to happen before the list_add in | |
558 | * __add_wait_queue. | |
559 | */ | |
15a77c6f | 560 | set_current_state(blocking_state); |
cbcfa130 | 561 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 562 | |
b8da2e46 | 563 | if (!is_vm_hugetlb_page(vma)) |
369cd212 MK |
564 | must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags, |
565 | reason); | |
566 | else | |
b8da2e46 | 567 | must_wait = userfaultfd_huge_must_wait(ctx, vma, |
7868a208 | 568 | vmf->address, |
369cd212 | 569 | vmf->flags, reason); |
b8da2e46 PX |
570 | if (is_vm_hugetlb_page(vma)) |
571 | hugetlb_vma_unlock_read(vma); | |
d8ed45c5 | 572 | mmap_read_unlock(mm); |
8d2afd96 | 573 | |
f9bf3522 | 574 | if (likely(must_wait && !READ_ONCE(ctx->released))) { |
a9a08845 | 575 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
86039bd3 | 576 | schedule(); |
ba85c702 | 577 | } |
86039bd3 | 578 | |
ba85c702 | 579 | __set_current_state(TASK_RUNNING); |
15b726ef AA |
580 | |
581 | /* | |
582 | * Here we race with the list_del; list_add in | |
583 | * userfaultfd_ctx_read(), however because we don't ever run | |
584 | * list_del_init() to refile across the two lists, the prev | |
585 | * and next pointers will never point to self. list_add also | |
586 | * would never let any of the two pointers to point to | |
587 | * self. So list_empty_careful won't risk to see both pointers | |
588 | * pointing to self at any time during the list refile. The | |
589 | * only case where list_del_init() is called is the full | |
590 | * removal in the wake function and there we don't re-list_add | |
591 | * and it's fine not to block on the spinlock. The uwq on this | |
592 | * kernel stack can be released after the list_del_init. | |
593 | */ | |
2055da97 | 594 | if (!list_empty_careful(&uwq.wq.entry)) { |
cbcfa130 | 595 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
15b726ef AA |
596 | /* |
597 | * No need of list_del_init(), the uwq on the stack | |
598 | * will be freed shortly anyway. | |
599 | */ | |
2055da97 | 600 | list_del(&uwq.wq.entry); |
cbcfa130 | 601 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 602 | } |
86039bd3 AA |
603 | |
604 | /* | |
605 | * ctx may go away after this if the userfault pseudo fd is | |
606 | * already released. | |
607 | */ | |
608 | userfaultfd_ctx_put(ctx); | |
609 | ||
ba85c702 AA |
610 | out: |
611 | return ret; | |
86039bd3 AA |
612 | } |
613 | ||
8c9e7bb7 AA |
614 | static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx, |
615 | struct userfaultfd_wait_queue *ewq) | |
9cd75c3c | 616 | { |
0cbb4b4f AA |
617 | struct userfaultfd_ctx *release_new_ctx; |
618 | ||
9a69a829 AA |
619 | if (WARN_ON_ONCE(current->flags & PF_EXITING)) |
620 | goto out; | |
9cd75c3c PE |
621 | |
622 | ewq->ctx = ctx; | |
623 | init_waitqueue_entry(&ewq->wq, current); | |
0cbb4b4f | 624 | release_new_ctx = NULL; |
9cd75c3c | 625 | |
cbcfa130 | 626 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
627 | /* |
628 | * After the __add_wait_queue the uwq is visible to userland | |
629 | * through poll/read(). | |
630 | */ | |
631 | __add_wait_queue(&ctx->event_wqh, &ewq->wq); | |
632 | for (;;) { | |
633 | set_current_state(TASK_KILLABLE); | |
634 | if (ewq->msg.event == 0) | |
635 | break; | |
6aa7de05 | 636 | if (READ_ONCE(ctx->released) || |
9cd75c3c | 637 | fatal_signal_pending(current)) { |
384632e6 AA |
638 | /* |
639 | * &ewq->wq may be queued in fork_event, but | |
640 | * __remove_wait_queue ignores the head | |
641 | * parameter. It would be a problem if it | |
642 | * didn't. | |
643 | */ | |
9cd75c3c | 644 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); |
7eb76d45 MR |
645 | if (ewq->msg.event == UFFD_EVENT_FORK) { |
646 | struct userfaultfd_ctx *new; | |
647 | ||
648 | new = (struct userfaultfd_ctx *) | |
649 | (unsigned long) | |
650 | ewq->msg.arg.reserved.reserved1; | |
0cbb4b4f | 651 | release_new_ctx = new; |
7eb76d45 | 652 | } |
9cd75c3c PE |
653 | break; |
654 | } | |
655 | ||
cbcfa130 | 656 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 657 | |
a9a08845 | 658 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
9cd75c3c PE |
659 | schedule(); |
660 | ||
cbcfa130 | 661 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
662 | } |
663 | __set_current_state(TASK_RUNNING); | |
cbcfa130 | 664 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 665 | |
0cbb4b4f AA |
666 | if (release_new_ctx) { |
667 | struct vm_area_struct *vma; | |
668 | struct mm_struct *mm = release_new_ctx->mm; | |
69dbe6da | 669 | VMA_ITERATOR(vmi, mm, 0); |
0cbb4b4f AA |
670 | |
671 | /* the various vma->vm_userfaultfd_ctx still points to it */ | |
d8ed45c5 | 672 | mmap_write_lock(mm); |
69dbe6da | 673 | for_each_vma(vmi, vma) { |
31e810aa | 674 | if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) { |
0cbb4b4f | 675 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
51d3d5eb DH |
676 | userfaultfd_set_vm_flags(vma, |
677 | vma->vm_flags & ~__VM_UFFD_FLAGS); | |
31e810aa | 678 | } |
69dbe6da | 679 | } |
d8ed45c5 | 680 | mmap_write_unlock(mm); |
0cbb4b4f AA |
681 | |
682 | userfaultfd_ctx_put(release_new_ctx); | |
683 | } | |
684 | ||
9cd75c3c PE |
685 | /* |
686 | * ctx may go away after this if the userfault pseudo fd is | |
687 | * already released. | |
688 | */ | |
9a69a829 | 689 | out: |
a759a909 NA |
690 | atomic_dec(&ctx->mmap_changing); |
691 | VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0); | |
9cd75c3c | 692 | userfaultfd_ctx_put(ctx); |
9cd75c3c PE |
693 | } |
694 | ||
695 | static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx, | |
696 | struct userfaultfd_wait_queue *ewq) | |
697 | { | |
698 | ewq->msg.event = 0; | |
699 | wake_up_locked(&ctx->event_wqh); | |
700 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); | |
701 | } | |
702 | ||
893e26e6 PE |
703 | int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs) |
704 | { | |
705 | struct userfaultfd_ctx *ctx = NULL, *octx; | |
706 | struct userfaultfd_fork_ctx *fctx; | |
707 | ||
708 | octx = vma->vm_userfaultfd_ctx.ctx; | |
709 | if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) { | |
710 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
51d3d5eb | 711 | userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS); |
893e26e6 PE |
712 | return 0; |
713 | } | |
714 | ||
715 | list_for_each_entry(fctx, fcs, list) | |
716 | if (fctx->orig == octx) { | |
717 | ctx = fctx->new; | |
718 | break; | |
719 | } | |
720 | ||
721 | if (!ctx) { | |
722 | fctx = kmalloc(sizeof(*fctx), GFP_KERNEL); | |
723 | if (!fctx) | |
724 | return -ENOMEM; | |
725 | ||
726 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); | |
727 | if (!ctx) { | |
728 | kfree(fctx); | |
729 | return -ENOMEM; | |
730 | } | |
731 | ||
ca880420 | 732 | refcount_set(&ctx->refcount, 1); |
893e26e6 | 733 | ctx->flags = octx->flags; |
893e26e6 PE |
734 | ctx->features = octx->features; |
735 | ctx->released = false; | |
a759a909 | 736 | atomic_set(&ctx->mmap_changing, 0); |
893e26e6 | 737 | ctx->mm = vma->vm_mm; |
00bb31fa | 738 | mmgrab(ctx->mm); |
893e26e6 PE |
739 | |
740 | userfaultfd_ctx_get(octx); | |
a759a909 | 741 | atomic_inc(&octx->mmap_changing); |
893e26e6 PE |
742 | fctx->orig = octx; |
743 | fctx->new = ctx; | |
744 | list_add_tail(&fctx->list, fcs); | |
745 | } | |
746 | ||
747 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
748 | return 0; | |
749 | } | |
750 | ||
8c9e7bb7 | 751 | static void dup_fctx(struct userfaultfd_fork_ctx *fctx) |
893e26e6 PE |
752 | { |
753 | struct userfaultfd_ctx *ctx = fctx->orig; | |
754 | struct userfaultfd_wait_queue ewq; | |
755 | ||
756 | msg_init(&ewq.msg); | |
757 | ||
758 | ewq.msg.event = UFFD_EVENT_FORK; | |
759 | ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new; | |
760 | ||
8c9e7bb7 | 761 | userfaultfd_event_wait_completion(ctx, &ewq); |
893e26e6 PE |
762 | } |
763 | ||
764 | void dup_userfaultfd_complete(struct list_head *fcs) | |
765 | { | |
893e26e6 PE |
766 | struct userfaultfd_fork_ctx *fctx, *n; |
767 | ||
768 | list_for_each_entry_safe(fctx, n, fcs, list) { | |
8c9e7bb7 | 769 | dup_fctx(fctx); |
893e26e6 PE |
770 | list_del(&fctx->list); |
771 | kfree(fctx); | |
772 | } | |
773 | } | |
774 | ||
72f87654 PE |
775 | void mremap_userfaultfd_prep(struct vm_area_struct *vma, |
776 | struct vm_userfaultfd_ctx *vm_ctx) | |
777 | { | |
778 | struct userfaultfd_ctx *ctx; | |
779 | ||
780 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
3cfd22be PX |
781 | |
782 | if (!ctx) | |
783 | return; | |
784 | ||
785 | if (ctx->features & UFFD_FEATURE_EVENT_REMAP) { | |
72f87654 PE |
786 | vm_ctx->ctx = ctx; |
787 | userfaultfd_ctx_get(ctx); | |
a759a909 | 788 | atomic_inc(&ctx->mmap_changing); |
3cfd22be PX |
789 | } else { |
790 | /* Drop uffd context if remap feature not enabled */ | |
791 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
51d3d5eb | 792 | userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS); |
72f87654 PE |
793 | } |
794 | } | |
795 | ||
90794bf1 | 796 | void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx, |
72f87654 PE |
797 | unsigned long from, unsigned long to, |
798 | unsigned long len) | |
799 | { | |
90794bf1 | 800 | struct userfaultfd_ctx *ctx = vm_ctx->ctx; |
72f87654 PE |
801 | struct userfaultfd_wait_queue ewq; |
802 | ||
803 | if (!ctx) | |
804 | return; | |
805 | ||
806 | if (to & ~PAGE_MASK) { | |
807 | userfaultfd_ctx_put(ctx); | |
808 | return; | |
809 | } | |
810 | ||
811 | msg_init(&ewq.msg); | |
812 | ||
813 | ewq.msg.event = UFFD_EVENT_REMAP; | |
814 | ewq.msg.arg.remap.from = from; | |
815 | ewq.msg.arg.remap.to = to; | |
816 | ewq.msg.arg.remap.len = len; | |
817 | ||
818 | userfaultfd_event_wait_completion(ctx, &ewq); | |
819 | } | |
820 | ||
70ccb92f | 821 | bool userfaultfd_remove(struct vm_area_struct *vma, |
d811914d | 822 | unsigned long start, unsigned long end) |
05ce7724 PE |
823 | { |
824 | struct mm_struct *mm = vma->vm_mm; | |
825 | struct userfaultfd_ctx *ctx; | |
826 | struct userfaultfd_wait_queue ewq; | |
827 | ||
828 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
d811914d | 829 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE)) |
70ccb92f | 830 | return true; |
05ce7724 PE |
831 | |
832 | userfaultfd_ctx_get(ctx); | |
a759a909 | 833 | atomic_inc(&ctx->mmap_changing); |
d8ed45c5 | 834 | mmap_read_unlock(mm); |
05ce7724 | 835 | |
05ce7724 PE |
836 | msg_init(&ewq.msg); |
837 | ||
d811914d MR |
838 | ewq.msg.event = UFFD_EVENT_REMOVE; |
839 | ewq.msg.arg.remove.start = start; | |
840 | ewq.msg.arg.remove.end = end; | |
05ce7724 PE |
841 | |
842 | userfaultfd_event_wait_completion(ctx, &ewq); | |
843 | ||
70ccb92f | 844 | return false; |
05ce7724 PE |
845 | } |
846 | ||
897ab3e0 MR |
847 | static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps, |
848 | unsigned long start, unsigned long end) | |
849 | { | |
850 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
851 | ||
852 | list_for_each_entry(unmap_ctx, unmaps, list) | |
853 | if (unmap_ctx->ctx == ctx && unmap_ctx->start == start && | |
854 | unmap_ctx->end == end) | |
855 | return true; | |
856 | ||
857 | return false; | |
858 | } | |
859 | ||
69dbe6da LH |
860 | int userfaultfd_unmap_prep(struct mm_struct *mm, unsigned long start, |
861 | unsigned long end, struct list_head *unmaps) | |
897ab3e0 | 862 | { |
69dbe6da LH |
863 | VMA_ITERATOR(vmi, mm, start); |
864 | struct vm_area_struct *vma; | |
865 | ||
866 | for_each_vma_range(vmi, vma, end) { | |
897ab3e0 MR |
867 | struct userfaultfd_unmap_ctx *unmap_ctx; |
868 | struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx; | |
869 | ||
870 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) || | |
871 | has_unmap_ctx(ctx, unmaps, start, end)) | |
872 | continue; | |
873 | ||
874 | unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL); | |
875 | if (!unmap_ctx) | |
876 | return -ENOMEM; | |
877 | ||
878 | userfaultfd_ctx_get(ctx); | |
a759a909 | 879 | atomic_inc(&ctx->mmap_changing); |
897ab3e0 MR |
880 | unmap_ctx->ctx = ctx; |
881 | unmap_ctx->start = start; | |
882 | unmap_ctx->end = end; | |
883 | list_add_tail(&unmap_ctx->list, unmaps); | |
884 | } | |
885 | ||
886 | return 0; | |
887 | } | |
888 | ||
889 | void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf) | |
890 | { | |
891 | struct userfaultfd_unmap_ctx *ctx, *n; | |
892 | struct userfaultfd_wait_queue ewq; | |
893 | ||
894 | list_for_each_entry_safe(ctx, n, uf, list) { | |
895 | msg_init(&ewq.msg); | |
896 | ||
897 | ewq.msg.event = UFFD_EVENT_UNMAP; | |
898 | ewq.msg.arg.remove.start = ctx->start; | |
899 | ewq.msg.arg.remove.end = ctx->end; | |
900 | ||
901 | userfaultfd_event_wait_completion(ctx->ctx, &ewq); | |
902 | ||
903 | list_del(&ctx->list); | |
904 | kfree(ctx); | |
905 | } | |
906 | } | |
907 | ||
86039bd3 AA |
908 | static int userfaultfd_release(struct inode *inode, struct file *file) |
909 | { | |
910 | struct userfaultfd_ctx *ctx = file->private_data; | |
911 | struct mm_struct *mm = ctx->mm; | |
912 | struct vm_area_struct *vma, *prev; | |
913 | /* len == 0 means wake all */ | |
914 | struct userfaultfd_wake_range range = { .len = 0, }; | |
915 | unsigned long new_flags; | |
11a9b902 | 916 | VMA_ITERATOR(vmi, mm, 0); |
86039bd3 | 917 | |
6aa7de05 | 918 | WRITE_ONCE(ctx->released, true); |
86039bd3 | 919 | |
d2005e3f ON |
920 | if (!mmget_not_zero(mm)) |
921 | goto wakeup; | |
922 | ||
86039bd3 AA |
923 | /* |
924 | * Flush page faults out of all CPUs. NOTE: all page faults | |
925 | * must be retried without returning VM_FAULT_SIGBUS if | |
926 | * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx | |
c1e8d7c6 | 927 | * changes while handle_userfault released the mmap_lock. So |
86039bd3 | 928 | * it's critical that released is set to true (above), before |
c1e8d7c6 | 929 | * taking the mmap_lock for writing. |
86039bd3 | 930 | */ |
d8ed45c5 | 931 | mmap_write_lock(mm); |
86039bd3 | 932 | prev = NULL; |
11a9b902 | 933 | for_each_vma(vmi, vma) { |
86039bd3 AA |
934 | cond_resched(); |
935 | BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 936 | !!(vma->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
937 | if (vma->vm_userfaultfd_ctx.ctx != ctx) { |
938 | prev = vma; | |
939 | continue; | |
940 | } | |
7677f7fd | 941 | new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS; |
9760ebff | 942 | prev = vma_merge(&vmi, mm, prev, vma->vm_start, vma->vm_end, |
4d45e75a JH |
943 | new_flags, vma->anon_vma, |
944 | vma->vm_file, vma->vm_pgoff, | |
945 | vma_policy(vma), | |
5c26f6ac | 946 | NULL_VM_UFFD_CTX, anon_vma_name(vma)); |
69dbe6da | 947 | if (prev) { |
4d45e75a | 948 | vma = prev; |
69dbe6da | 949 | } else { |
4d45e75a | 950 | prev = vma; |
69dbe6da LH |
951 | } |
952 | ||
51d3d5eb | 953 | userfaultfd_set_vm_flags(vma, new_flags); |
86039bd3 AA |
954 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
955 | } | |
d8ed45c5 | 956 | mmap_write_unlock(mm); |
d2005e3f ON |
957 | mmput(mm); |
958 | wakeup: | |
86039bd3 | 959 | /* |
15b726ef | 960 | * After no new page faults can wait on this fault_*wqh, flush |
86039bd3 | 961 | * the last page faults that may have been already waiting on |
15b726ef | 962 | * the fault_*wqh. |
86039bd3 | 963 | */ |
cbcfa130 | 964 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
ac5be6b4 | 965 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range); |
c430d1e8 | 966 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range); |
cbcfa130 | 967 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 968 | |
5a18b64e MR |
969 | /* Flush pending events that may still wait on event_wqh */ |
970 | wake_up_all(&ctx->event_wqh); | |
971 | ||
a9a08845 | 972 | wake_up_poll(&ctx->fd_wqh, EPOLLHUP); |
86039bd3 AA |
973 | userfaultfd_ctx_put(ctx); |
974 | return 0; | |
975 | } | |
976 | ||
15b726ef | 977 | /* fault_pending_wqh.lock must be hold by the caller */ |
6dcc27fd PE |
978 | static inline struct userfaultfd_wait_queue *find_userfault_in( |
979 | wait_queue_head_t *wqh) | |
86039bd3 | 980 | { |
ac6424b9 | 981 | wait_queue_entry_t *wq; |
15b726ef | 982 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 983 | |
456a7378 | 984 | lockdep_assert_held(&wqh->lock); |
86039bd3 | 985 | |
15b726ef | 986 | uwq = NULL; |
6dcc27fd | 987 | if (!waitqueue_active(wqh)) |
15b726ef AA |
988 | goto out; |
989 | /* walk in reverse to provide FIFO behavior to read userfaults */ | |
2055da97 | 990 | wq = list_last_entry(&wqh->head, typeof(*wq), entry); |
15b726ef AA |
991 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
992 | out: | |
993 | return uwq; | |
86039bd3 | 994 | } |
6dcc27fd PE |
995 | |
996 | static inline struct userfaultfd_wait_queue *find_userfault( | |
997 | struct userfaultfd_ctx *ctx) | |
998 | { | |
999 | return find_userfault_in(&ctx->fault_pending_wqh); | |
1000 | } | |
86039bd3 | 1001 | |
9cd75c3c PE |
1002 | static inline struct userfaultfd_wait_queue *find_userfault_evt( |
1003 | struct userfaultfd_ctx *ctx) | |
1004 | { | |
1005 | return find_userfault_in(&ctx->event_wqh); | |
1006 | } | |
1007 | ||
076ccb76 | 1008 | static __poll_t userfaultfd_poll(struct file *file, poll_table *wait) |
86039bd3 AA |
1009 | { |
1010 | struct userfaultfd_ctx *ctx = file->private_data; | |
076ccb76 | 1011 | __poll_t ret; |
86039bd3 AA |
1012 | |
1013 | poll_wait(file, &ctx->fd_wqh, wait); | |
1014 | ||
22e5fe2a | 1015 | if (!userfaultfd_is_initialized(ctx)) |
a9a08845 | 1016 | return EPOLLERR; |
9cd75c3c | 1017 | |
22e5fe2a NA |
1018 | /* |
1019 | * poll() never guarantees that read won't block. | |
1020 | * userfaults can be waken before they're read(). | |
1021 | */ | |
1022 | if (unlikely(!(file->f_flags & O_NONBLOCK))) | |
a9a08845 | 1023 | return EPOLLERR; |
22e5fe2a NA |
1024 | /* |
1025 | * lockless access to see if there are pending faults | |
1026 | * __pollwait last action is the add_wait_queue but | |
1027 | * the spin_unlock would allow the waitqueue_active to | |
1028 | * pass above the actual list_add inside | |
1029 | * add_wait_queue critical section. So use a full | |
1030 | * memory barrier to serialize the list_add write of | |
1031 | * add_wait_queue() with the waitqueue_active read | |
1032 | * below. | |
1033 | */ | |
1034 | ret = 0; | |
1035 | smp_mb(); | |
1036 | if (waitqueue_active(&ctx->fault_pending_wqh)) | |
1037 | ret = EPOLLIN; | |
1038 | else if (waitqueue_active(&ctx->event_wqh)) | |
1039 | ret = EPOLLIN; | |
1040 | ||
1041 | return ret; | |
86039bd3 AA |
1042 | } |
1043 | ||
893e26e6 PE |
1044 | static const struct file_operations userfaultfd_fops; |
1045 | ||
b537900f DC |
1046 | static int resolve_userfault_fork(struct userfaultfd_ctx *new, |
1047 | struct inode *inode, | |
893e26e6 PE |
1048 | struct uffd_msg *msg) |
1049 | { | |
1050 | int fd; | |
893e26e6 | 1051 | |
b537900f | 1052 | fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new, |
abec3d01 | 1053 | O_RDONLY | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode); |
893e26e6 PE |
1054 | if (fd < 0) |
1055 | return fd; | |
1056 | ||
893e26e6 PE |
1057 | msg->arg.reserved.reserved1 = 0; |
1058 | msg->arg.fork.ufd = fd; | |
893e26e6 PE |
1059 | return 0; |
1060 | } | |
1061 | ||
86039bd3 | 1062 | static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait, |
b537900f | 1063 | struct uffd_msg *msg, struct inode *inode) |
86039bd3 AA |
1064 | { |
1065 | ssize_t ret; | |
1066 | DECLARE_WAITQUEUE(wait, current); | |
15b726ef | 1067 | struct userfaultfd_wait_queue *uwq; |
893e26e6 PE |
1068 | /* |
1069 | * Handling fork event requires sleeping operations, so | |
1070 | * we drop the event_wqh lock, then do these ops, then | |
1071 | * lock it back and wake up the waiter. While the lock is | |
1072 | * dropped the ewq may go away so we keep track of it | |
1073 | * carefully. | |
1074 | */ | |
1075 | LIST_HEAD(fork_event); | |
1076 | struct userfaultfd_ctx *fork_nctx = NULL; | |
86039bd3 | 1077 | |
15b726ef | 1078 | /* always take the fd_wqh lock before the fault_pending_wqh lock */ |
ae62c16e | 1079 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1080 | __add_wait_queue(&ctx->fd_wqh, &wait); |
1081 | for (;;) { | |
1082 | set_current_state(TASK_INTERRUPTIBLE); | |
15b726ef AA |
1083 | spin_lock(&ctx->fault_pending_wqh.lock); |
1084 | uwq = find_userfault(ctx); | |
1085 | if (uwq) { | |
2c5b7e1b AA |
1086 | /* |
1087 | * Use a seqcount to repeat the lockless check | |
1088 | * in wake_userfault() to avoid missing | |
1089 | * wakeups because during the refile both | |
1090 | * waitqueue could become empty if this is the | |
1091 | * only userfault. | |
1092 | */ | |
1093 | write_seqcount_begin(&ctx->refile_seq); | |
1094 | ||
86039bd3 | 1095 | /* |
15b726ef AA |
1096 | * The fault_pending_wqh.lock prevents the uwq |
1097 | * to disappear from under us. | |
1098 | * | |
1099 | * Refile this userfault from | |
1100 | * fault_pending_wqh to fault_wqh, it's not | |
1101 | * pending anymore after we read it. | |
1102 | * | |
1103 | * Use list_del() by hand (as | |
1104 | * userfaultfd_wake_function also uses | |
1105 | * list_del_init() by hand) to be sure nobody | |
1106 | * changes __remove_wait_queue() to use | |
1107 | * list_del_init() in turn breaking the | |
1108 | * !list_empty_careful() check in | |
2055da97 | 1109 | * handle_userfault(). The uwq->wq.head list |
15b726ef AA |
1110 | * must never be empty at any time during the |
1111 | * refile, or the waitqueue could disappear | |
1112 | * from under us. The "wait_queue_head_t" | |
1113 | * parameter of __remove_wait_queue() is unused | |
1114 | * anyway. | |
86039bd3 | 1115 | */ |
2055da97 | 1116 | list_del(&uwq->wq.entry); |
c430d1e8 | 1117 | add_wait_queue(&ctx->fault_wqh, &uwq->wq); |
15b726ef | 1118 | |
2c5b7e1b AA |
1119 | write_seqcount_end(&ctx->refile_seq); |
1120 | ||
a9b85f94 AA |
1121 | /* careful to always initialize msg if ret == 0 */ |
1122 | *msg = uwq->msg; | |
15b726ef | 1123 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1124 | ret = 0; |
1125 | break; | |
1126 | } | |
15b726ef | 1127 | spin_unlock(&ctx->fault_pending_wqh.lock); |
9cd75c3c PE |
1128 | |
1129 | spin_lock(&ctx->event_wqh.lock); | |
1130 | uwq = find_userfault_evt(ctx); | |
1131 | if (uwq) { | |
1132 | *msg = uwq->msg; | |
1133 | ||
893e26e6 PE |
1134 | if (uwq->msg.event == UFFD_EVENT_FORK) { |
1135 | fork_nctx = (struct userfaultfd_ctx *) | |
1136 | (unsigned long) | |
1137 | uwq->msg.arg.reserved.reserved1; | |
2055da97 | 1138 | list_move(&uwq->wq.entry, &fork_event); |
384632e6 AA |
1139 | /* |
1140 | * fork_nctx can be freed as soon as | |
1141 | * we drop the lock, unless we take a | |
1142 | * reference on it. | |
1143 | */ | |
1144 | userfaultfd_ctx_get(fork_nctx); | |
893e26e6 PE |
1145 | spin_unlock(&ctx->event_wqh.lock); |
1146 | ret = 0; | |
1147 | break; | |
1148 | } | |
1149 | ||
9cd75c3c PE |
1150 | userfaultfd_event_complete(ctx, uwq); |
1151 | spin_unlock(&ctx->event_wqh.lock); | |
1152 | ret = 0; | |
1153 | break; | |
1154 | } | |
1155 | spin_unlock(&ctx->event_wqh.lock); | |
1156 | ||
86039bd3 AA |
1157 | if (signal_pending(current)) { |
1158 | ret = -ERESTARTSYS; | |
1159 | break; | |
1160 | } | |
1161 | if (no_wait) { | |
1162 | ret = -EAGAIN; | |
1163 | break; | |
1164 | } | |
ae62c16e | 1165 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1166 | schedule(); |
ae62c16e | 1167 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1168 | } |
1169 | __remove_wait_queue(&ctx->fd_wqh, &wait); | |
1170 | __set_current_state(TASK_RUNNING); | |
ae62c16e | 1171 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1172 | |
893e26e6 | 1173 | if (!ret && msg->event == UFFD_EVENT_FORK) { |
b537900f | 1174 | ret = resolve_userfault_fork(fork_nctx, inode, msg); |
cbcfa130 | 1175 | spin_lock_irq(&ctx->event_wqh.lock); |
384632e6 AA |
1176 | if (!list_empty(&fork_event)) { |
1177 | /* | |
1178 | * The fork thread didn't abort, so we can | |
1179 | * drop the temporary refcount. | |
1180 | */ | |
1181 | userfaultfd_ctx_put(fork_nctx); | |
1182 | ||
1183 | uwq = list_first_entry(&fork_event, | |
1184 | typeof(*uwq), | |
1185 | wq.entry); | |
1186 | /* | |
1187 | * If fork_event list wasn't empty and in turn | |
1188 | * the event wasn't already released by fork | |
1189 | * (the event is allocated on fork kernel | |
1190 | * stack), put the event back to its place in | |
1191 | * the event_wq. fork_event head will be freed | |
1192 | * as soon as we return so the event cannot | |
1193 | * stay queued there no matter the current | |
1194 | * "ret" value. | |
1195 | */ | |
1196 | list_del(&uwq->wq.entry); | |
1197 | __add_wait_queue(&ctx->event_wqh, &uwq->wq); | |
893e26e6 | 1198 | |
384632e6 AA |
1199 | /* |
1200 | * Leave the event in the waitqueue and report | |
1201 | * error to userland if we failed to resolve | |
1202 | * the userfault fork. | |
1203 | */ | |
1204 | if (likely(!ret)) | |
893e26e6 | 1205 | userfaultfd_event_complete(ctx, uwq); |
384632e6 AA |
1206 | } else { |
1207 | /* | |
1208 | * Here the fork thread aborted and the | |
1209 | * refcount from the fork thread on fork_nctx | |
1210 | * has already been released. We still hold | |
1211 | * the reference we took before releasing the | |
1212 | * lock above. If resolve_userfault_fork | |
1213 | * failed we've to drop it because the | |
1214 | * fork_nctx has to be freed in such case. If | |
1215 | * it succeeded we'll hold it because the new | |
1216 | * uffd references it. | |
1217 | */ | |
1218 | if (ret) | |
1219 | userfaultfd_ctx_put(fork_nctx); | |
893e26e6 | 1220 | } |
cbcfa130 | 1221 | spin_unlock_irq(&ctx->event_wqh.lock); |
893e26e6 PE |
1222 | } |
1223 | ||
86039bd3 AA |
1224 | return ret; |
1225 | } | |
1226 | ||
1227 | static ssize_t userfaultfd_read(struct file *file, char __user *buf, | |
1228 | size_t count, loff_t *ppos) | |
1229 | { | |
1230 | struct userfaultfd_ctx *ctx = file->private_data; | |
1231 | ssize_t _ret, ret = 0; | |
a9b85f94 | 1232 | struct uffd_msg msg; |
86039bd3 | 1233 | int no_wait = file->f_flags & O_NONBLOCK; |
b537900f | 1234 | struct inode *inode = file_inode(file); |
86039bd3 | 1235 | |
22e5fe2a | 1236 | if (!userfaultfd_is_initialized(ctx)) |
86039bd3 | 1237 | return -EINVAL; |
86039bd3 AA |
1238 | |
1239 | for (;;) { | |
a9b85f94 | 1240 | if (count < sizeof(msg)) |
86039bd3 | 1241 | return ret ? ret : -EINVAL; |
b537900f | 1242 | _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode); |
86039bd3 AA |
1243 | if (_ret < 0) |
1244 | return ret ? ret : _ret; | |
a9b85f94 | 1245 | if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg))) |
86039bd3 | 1246 | return ret ? ret : -EFAULT; |
a9b85f94 AA |
1247 | ret += sizeof(msg); |
1248 | buf += sizeof(msg); | |
1249 | count -= sizeof(msg); | |
86039bd3 AA |
1250 | /* |
1251 | * Allow to read more than one fault at time but only | |
1252 | * block if waiting for the very first one. | |
1253 | */ | |
1254 | no_wait = O_NONBLOCK; | |
1255 | } | |
1256 | } | |
1257 | ||
1258 | static void __wake_userfault(struct userfaultfd_ctx *ctx, | |
1259 | struct userfaultfd_wake_range *range) | |
1260 | { | |
cbcfa130 | 1261 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 1262 | /* wake all in the range and autoremove */ |
15b726ef | 1263 | if (waitqueue_active(&ctx->fault_pending_wqh)) |
ac5be6b4 | 1264 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, |
15b726ef AA |
1265 | range); |
1266 | if (waitqueue_active(&ctx->fault_wqh)) | |
c430d1e8 | 1267 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range); |
cbcfa130 | 1268 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1269 | } |
1270 | ||
1271 | static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx, | |
1272 | struct userfaultfd_wake_range *range) | |
1273 | { | |
2c5b7e1b AA |
1274 | unsigned seq; |
1275 | bool need_wakeup; | |
1276 | ||
86039bd3 AA |
1277 | /* |
1278 | * To be sure waitqueue_active() is not reordered by the CPU | |
1279 | * before the pagetable update, use an explicit SMP memory | |
3e4e28c5 | 1280 | * barrier here. PT lock release or mmap_read_unlock(mm) still |
86039bd3 AA |
1281 | * have release semantics that can allow the |
1282 | * waitqueue_active() to be reordered before the pte update. | |
1283 | */ | |
1284 | smp_mb(); | |
1285 | ||
1286 | /* | |
1287 | * Use waitqueue_active because it's very frequent to | |
1288 | * change the address space atomically even if there are no | |
1289 | * userfaults yet. So we take the spinlock only when we're | |
1290 | * sure we've userfaults to wake. | |
1291 | */ | |
2c5b7e1b AA |
1292 | do { |
1293 | seq = read_seqcount_begin(&ctx->refile_seq); | |
1294 | need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) || | |
1295 | waitqueue_active(&ctx->fault_wqh); | |
1296 | cond_resched(); | |
1297 | } while (read_seqcount_retry(&ctx->refile_seq, seq)); | |
1298 | if (need_wakeup) | |
86039bd3 AA |
1299 | __wake_userfault(ctx, range); |
1300 | } | |
1301 | ||
1302 | static __always_inline int validate_range(struct mm_struct *mm, | |
e71e2ace | 1303 | __u64 start, __u64 len) |
86039bd3 AA |
1304 | { |
1305 | __u64 task_size = mm->task_size; | |
1306 | ||
e71e2ace | 1307 | if (start & ~PAGE_MASK) |
86039bd3 AA |
1308 | return -EINVAL; |
1309 | if (len & ~PAGE_MASK) | |
1310 | return -EINVAL; | |
1311 | if (!len) | |
1312 | return -EINVAL; | |
e71e2ace | 1313 | if (start < mmap_min_addr) |
86039bd3 | 1314 | return -EINVAL; |
e71e2ace | 1315 | if (start >= task_size) |
86039bd3 | 1316 | return -EINVAL; |
e71e2ace | 1317 | if (len > task_size - start) |
86039bd3 AA |
1318 | return -EINVAL; |
1319 | return 0; | |
1320 | } | |
1321 | ||
1322 | static int userfaultfd_register(struct userfaultfd_ctx *ctx, | |
1323 | unsigned long arg) | |
1324 | { | |
1325 | struct mm_struct *mm = ctx->mm; | |
1326 | struct vm_area_struct *vma, *prev, *cur; | |
1327 | int ret; | |
1328 | struct uffdio_register uffdio_register; | |
1329 | struct uffdio_register __user *user_uffdio_register; | |
1330 | unsigned long vm_flags, new_flags; | |
1331 | bool found; | |
ce53e8e6 | 1332 | bool basic_ioctls; |
86039bd3 | 1333 | unsigned long start, end, vma_end; |
11a9b902 | 1334 | struct vma_iterator vmi; |
5543d3c4 | 1335 | pgoff_t pgoff; |
86039bd3 AA |
1336 | |
1337 | user_uffdio_register = (struct uffdio_register __user *) arg; | |
1338 | ||
1339 | ret = -EFAULT; | |
1340 | if (copy_from_user(&uffdio_register, user_uffdio_register, | |
1341 | sizeof(uffdio_register)-sizeof(__u64))) | |
1342 | goto out; | |
1343 | ||
1344 | ret = -EINVAL; | |
1345 | if (!uffdio_register.mode) | |
1346 | goto out; | |
7677f7fd | 1347 | if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES) |
86039bd3 AA |
1348 | goto out; |
1349 | vm_flags = 0; | |
1350 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING) | |
1351 | vm_flags |= VM_UFFD_MISSING; | |
00b151f2 PX |
1352 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) { |
1353 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP | |
1354 | goto out; | |
1355 | #endif | |
86039bd3 | 1356 | vm_flags |= VM_UFFD_WP; |
00b151f2 | 1357 | } |
7677f7fd AR |
1358 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) { |
1359 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR | |
1360 | goto out; | |
1361 | #endif | |
1362 | vm_flags |= VM_UFFD_MINOR; | |
1363 | } | |
86039bd3 | 1364 | |
e71e2ace | 1365 | ret = validate_range(mm, uffdio_register.range.start, |
86039bd3 AA |
1366 | uffdio_register.range.len); |
1367 | if (ret) | |
1368 | goto out; | |
1369 | ||
1370 | start = uffdio_register.range.start; | |
1371 | end = start + uffdio_register.range.len; | |
1372 | ||
d2005e3f ON |
1373 | ret = -ENOMEM; |
1374 | if (!mmget_not_zero(mm)) | |
1375 | goto out; | |
1376 | ||
11a9b902 | 1377 | ret = -EINVAL; |
d8ed45c5 | 1378 | mmap_write_lock(mm); |
11a9b902 LH |
1379 | vma_iter_init(&vmi, mm, start); |
1380 | vma = vma_find(&vmi, end); | |
86039bd3 AA |
1381 | if (!vma) |
1382 | goto out_unlock; | |
1383 | ||
cab350af MK |
1384 | /* |
1385 | * If the first vma contains huge pages, make sure start address | |
1386 | * is aligned to huge page size. | |
1387 | */ | |
1388 | if (is_vm_hugetlb_page(vma)) { | |
1389 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1390 | ||
1391 | if (start & (vma_hpagesize - 1)) | |
1392 | goto out_unlock; | |
1393 | } | |
1394 | ||
86039bd3 AA |
1395 | /* |
1396 | * Search for not compatible vmas. | |
86039bd3 AA |
1397 | */ |
1398 | found = false; | |
ce53e8e6 | 1399 | basic_ioctls = false; |
11a9b902 LH |
1400 | cur = vma; |
1401 | do { | |
86039bd3 AA |
1402 | cond_resched(); |
1403 | ||
1404 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 1405 | !!(cur->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
1406 | |
1407 | /* check not compatible vmas */ | |
1408 | ret = -EINVAL; | |
63b2d417 | 1409 | if (!vma_can_userfault(cur, vm_flags)) |
86039bd3 | 1410 | goto out_unlock; |
29ec9066 AA |
1411 | |
1412 | /* | |
1413 | * UFFDIO_COPY will fill file holes even without | |
1414 | * PROT_WRITE. This check enforces that if this is a | |
1415 | * MAP_SHARED, the process has write permission to the backing | |
1416 | * file. If VM_MAYWRITE is set it also enforces that on a | |
1417 | * MAP_SHARED vma: there is no F_WRITE_SEAL and no further | |
1418 | * F_WRITE_SEAL can be taken until the vma is destroyed. | |
1419 | */ | |
1420 | ret = -EPERM; | |
1421 | if (unlikely(!(cur->vm_flags & VM_MAYWRITE))) | |
1422 | goto out_unlock; | |
1423 | ||
cab350af MK |
1424 | /* |
1425 | * If this vma contains ending address, and huge pages | |
1426 | * check alignment. | |
1427 | */ | |
1428 | if (is_vm_hugetlb_page(cur) && end <= cur->vm_end && | |
1429 | end > cur->vm_start) { | |
1430 | unsigned long vma_hpagesize = vma_kernel_pagesize(cur); | |
1431 | ||
1432 | ret = -EINVAL; | |
1433 | ||
1434 | if (end & (vma_hpagesize - 1)) | |
1435 | goto out_unlock; | |
1436 | } | |
63b2d417 AA |
1437 | if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE)) |
1438 | goto out_unlock; | |
86039bd3 AA |
1439 | |
1440 | /* | |
1441 | * Check that this vma isn't already owned by a | |
1442 | * different userfaultfd. We can't allow more than one | |
1443 | * userfaultfd to own a single vma simultaneously or we | |
1444 | * wouldn't know which one to deliver the userfaults to. | |
1445 | */ | |
1446 | ret = -EBUSY; | |
1447 | if (cur->vm_userfaultfd_ctx.ctx && | |
1448 | cur->vm_userfaultfd_ctx.ctx != ctx) | |
1449 | goto out_unlock; | |
1450 | ||
cab350af MK |
1451 | /* |
1452 | * Note vmas containing huge pages | |
1453 | */ | |
ce53e8e6 MR |
1454 | if (is_vm_hugetlb_page(cur)) |
1455 | basic_ioctls = true; | |
cab350af | 1456 | |
86039bd3 | 1457 | found = true; |
11a9b902 | 1458 | } for_each_vma_range(vmi, cur, end); |
86039bd3 AA |
1459 | BUG_ON(!found); |
1460 | ||
11a9b902 LH |
1461 | vma_iter_set(&vmi, start); |
1462 | prev = vma_prev(&vmi); | |
270aa010 PX |
1463 | if (vma->vm_start < start) |
1464 | prev = vma; | |
86039bd3 AA |
1465 | |
1466 | ret = 0; | |
11a9b902 | 1467 | for_each_vma_range(vmi, vma, end) { |
86039bd3 AA |
1468 | cond_resched(); |
1469 | ||
63b2d417 | 1470 | BUG_ON(!vma_can_userfault(vma, vm_flags)); |
86039bd3 AA |
1471 | BUG_ON(vma->vm_userfaultfd_ctx.ctx && |
1472 | vma->vm_userfaultfd_ctx.ctx != ctx); | |
29ec9066 | 1473 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
86039bd3 AA |
1474 | |
1475 | /* | |
1476 | * Nothing to do: this vma is already registered into this | |
1477 | * userfaultfd and with the right tracking mode too. | |
1478 | */ | |
1479 | if (vma->vm_userfaultfd_ctx.ctx == ctx && | |
1480 | (vma->vm_flags & vm_flags) == vm_flags) | |
1481 | goto skip; | |
1482 | ||
1483 | if (vma->vm_start > start) | |
1484 | start = vma->vm_start; | |
1485 | vma_end = min(end, vma->vm_end); | |
1486 | ||
7677f7fd | 1487 | new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags; |
5543d3c4 | 1488 | pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); |
9760ebff | 1489 | prev = vma_merge(&vmi, mm, prev, start, vma_end, new_flags, |
5543d3c4 | 1490 | vma->anon_vma, vma->vm_file, pgoff, |
86039bd3 | 1491 | vma_policy(vma), |
9a10064f | 1492 | ((struct vm_userfaultfd_ctx){ ctx }), |
5c26f6ac | 1493 | anon_vma_name(vma)); |
86039bd3 | 1494 | if (prev) { |
69dbe6da | 1495 | /* vma_merge() invalidated the mas */ |
86039bd3 AA |
1496 | vma = prev; |
1497 | goto next; | |
1498 | } | |
1499 | if (vma->vm_start < start) { | |
9760ebff | 1500 | ret = split_vma(&vmi, vma, start, 1); |
86039bd3 AA |
1501 | if (ret) |
1502 | break; | |
1503 | } | |
1504 | if (vma->vm_end > end) { | |
9760ebff | 1505 | ret = split_vma(&vmi, vma, end, 0); |
86039bd3 AA |
1506 | if (ret) |
1507 | break; | |
1508 | } | |
1509 | next: | |
1510 | /* | |
1511 | * In the vma_merge() successful mprotect-like case 8: | |
1512 | * the next vma was merged into the current one and | |
1513 | * the current one has not been updated yet. | |
1514 | */ | |
51d3d5eb | 1515 | userfaultfd_set_vm_flags(vma, new_flags); |
86039bd3 AA |
1516 | vma->vm_userfaultfd_ctx.ctx = ctx; |
1517 | ||
6dfeaff9 PX |
1518 | if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma)) |
1519 | hugetlb_unshare_all_pmds(vma); | |
1520 | ||
86039bd3 AA |
1521 | skip: |
1522 | prev = vma; | |
1523 | start = vma->vm_end; | |
11a9b902 LH |
1524 | } |
1525 | ||
86039bd3 | 1526 | out_unlock: |
d8ed45c5 | 1527 | mmap_write_unlock(mm); |
d2005e3f | 1528 | mmput(mm); |
86039bd3 | 1529 | if (!ret) { |
14819305 PX |
1530 | __u64 ioctls_out; |
1531 | ||
1532 | ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC : | |
1533 | UFFD_API_RANGE_IOCTLS; | |
1534 | ||
1535 | /* | |
1536 | * Declare the WP ioctl only if the WP mode is | |
1537 | * specified and all checks passed with the range | |
1538 | */ | |
1539 | if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP)) | |
1540 | ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT); | |
1541 | ||
f6191471 AR |
1542 | /* CONTINUE ioctl is only supported for MINOR ranges. */ |
1543 | if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR)) | |
1544 | ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE); | |
1545 | ||
86039bd3 AA |
1546 | /* |
1547 | * Now that we scanned all vmas we can already tell | |
1548 | * userland which ioctls methods are guaranteed to | |
1549 | * succeed on this range. | |
1550 | */ | |
14819305 | 1551 | if (put_user(ioctls_out, &user_uffdio_register->ioctls)) |
86039bd3 AA |
1552 | ret = -EFAULT; |
1553 | } | |
1554 | out: | |
1555 | return ret; | |
1556 | } | |
1557 | ||
1558 | static int userfaultfd_unregister(struct userfaultfd_ctx *ctx, | |
1559 | unsigned long arg) | |
1560 | { | |
1561 | struct mm_struct *mm = ctx->mm; | |
1562 | struct vm_area_struct *vma, *prev, *cur; | |
1563 | int ret; | |
1564 | struct uffdio_range uffdio_unregister; | |
1565 | unsigned long new_flags; | |
1566 | bool found; | |
1567 | unsigned long start, end, vma_end; | |
1568 | const void __user *buf = (void __user *)arg; | |
11a9b902 | 1569 | struct vma_iterator vmi; |
5543d3c4 | 1570 | pgoff_t pgoff; |
86039bd3 AA |
1571 | |
1572 | ret = -EFAULT; | |
1573 | if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister))) | |
1574 | goto out; | |
1575 | ||
e71e2ace | 1576 | ret = validate_range(mm, uffdio_unregister.start, |
86039bd3 AA |
1577 | uffdio_unregister.len); |
1578 | if (ret) | |
1579 | goto out; | |
1580 | ||
1581 | start = uffdio_unregister.start; | |
1582 | end = start + uffdio_unregister.len; | |
1583 | ||
d2005e3f ON |
1584 | ret = -ENOMEM; |
1585 | if (!mmget_not_zero(mm)) | |
1586 | goto out; | |
1587 | ||
d8ed45c5 | 1588 | mmap_write_lock(mm); |
86039bd3 | 1589 | ret = -EINVAL; |
11a9b902 LH |
1590 | vma_iter_init(&vmi, mm, start); |
1591 | vma = vma_find(&vmi, end); | |
1592 | if (!vma) | |
86039bd3 AA |
1593 | goto out_unlock; |
1594 | ||
cab350af MK |
1595 | /* |
1596 | * If the first vma contains huge pages, make sure start address | |
1597 | * is aligned to huge page size. | |
1598 | */ | |
1599 | if (is_vm_hugetlb_page(vma)) { | |
1600 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1601 | ||
1602 | if (start & (vma_hpagesize - 1)) | |
1603 | goto out_unlock; | |
1604 | } | |
1605 | ||
86039bd3 AA |
1606 | /* |
1607 | * Search for not compatible vmas. | |
86039bd3 AA |
1608 | */ |
1609 | found = false; | |
11a9b902 LH |
1610 | cur = vma; |
1611 | do { | |
86039bd3 AA |
1612 | cond_resched(); |
1613 | ||
1614 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 1615 | !!(cur->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
1616 | |
1617 | /* | |
1618 | * Check not compatible vmas, not strictly required | |
1619 | * here as not compatible vmas cannot have an | |
1620 | * userfaultfd_ctx registered on them, but this | |
1621 | * provides for more strict behavior to notice | |
1622 | * unregistration errors. | |
1623 | */ | |
63b2d417 | 1624 | if (!vma_can_userfault(cur, cur->vm_flags)) |
86039bd3 AA |
1625 | goto out_unlock; |
1626 | ||
1627 | found = true; | |
11a9b902 | 1628 | } for_each_vma_range(vmi, cur, end); |
86039bd3 AA |
1629 | BUG_ON(!found); |
1630 | ||
11a9b902 LH |
1631 | vma_iter_set(&vmi, start); |
1632 | prev = vma_prev(&vmi); | |
270aa010 PX |
1633 | if (vma->vm_start < start) |
1634 | prev = vma; | |
1635 | ||
86039bd3 | 1636 | ret = 0; |
11a9b902 | 1637 | for_each_vma_range(vmi, vma, end) { |
86039bd3 AA |
1638 | cond_resched(); |
1639 | ||
63b2d417 | 1640 | BUG_ON(!vma_can_userfault(vma, vma->vm_flags)); |
86039bd3 AA |
1641 | |
1642 | /* | |
1643 | * Nothing to do: this vma is already registered into this | |
1644 | * userfaultfd and with the right tracking mode too. | |
1645 | */ | |
1646 | if (!vma->vm_userfaultfd_ctx.ctx) | |
1647 | goto skip; | |
1648 | ||
01e881f5 AA |
1649 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
1650 | ||
86039bd3 AA |
1651 | if (vma->vm_start > start) |
1652 | start = vma->vm_start; | |
1653 | vma_end = min(end, vma->vm_end); | |
1654 | ||
09fa5296 AA |
1655 | if (userfaultfd_missing(vma)) { |
1656 | /* | |
1657 | * Wake any concurrent pending userfault while | |
1658 | * we unregister, so they will not hang | |
1659 | * permanently and it avoids userland to call | |
1660 | * UFFDIO_WAKE explicitly. | |
1661 | */ | |
1662 | struct userfaultfd_wake_range range; | |
1663 | range.start = start; | |
1664 | range.len = vma_end - start; | |
1665 | wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range); | |
1666 | } | |
1667 | ||
f369b07c PX |
1668 | /* Reset ptes for the whole vma range if wr-protected */ |
1669 | if (userfaultfd_wp(vma)) | |
61c50040 | 1670 | uffd_wp_range(vma, start, vma_end - start, false); |
f369b07c | 1671 | |
7677f7fd | 1672 | new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS; |
5543d3c4 | 1673 | pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); |
9760ebff | 1674 | prev = vma_merge(&vmi, mm, prev, start, vma_end, new_flags, |
5543d3c4 | 1675 | vma->anon_vma, vma->vm_file, pgoff, |
86039bd3 | 1676 | vma_policy(vma), |
5c26f6ac | 1677 | NULL_VM_UFFD_CTX, anon_vma_name(vma)); |
86039bd3 AA |
1678 | if (prev) { |
1679 | vma = prev; | |
1680 | goto next; | |
1681 | } | |
1682 | if (vma->vm_start < start) { | |
9760ebff | 1683 | ret = split_vma(&vmi, vma, start, 1); |
86039bd3 AA |
1684 | if (ret) |
1685 | break; | |
1686 | } | |
1687 | if (vma->vm_end > end) { | |
9760ebff | 1688 | ret = split_vma(&vmi, vma, end, 0); |
86039bd3 AA |
1689 | if (ret) |
1690 | break; | |
1691 | } | |
1692 | next: | |
1693 | /* | |
1694 | * In the vma_merge() successful mprotect-like case 8: | |
1695 | * the next vma was merged into the current one and | |
1696 | * the current one has not been updated yet. | |
1697 | */ | |
51d3d5eb | 1698 | userfaultfd_set_vm_flags(vma, new_flags); |
86039bd3 AA |
1699 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
1700 | ||
1701 | skip: | |
1702 | prev = vma; | |
1703 | start = vma->vm_end; | |
11a9b902 LH |
1704 | } |
1705 | ||
86039bd3 | 1706 | out_unlock: |
d8ed45c5 | 1707 | mmap_write_unlock(mm); |
d2005e3f | 1708 | mmput(mm); |
86039bd3 AA |
1709 | out: |
1710 | return ret; | |
1711 | } | |
1712 | ||
1713 | /* | |
ba85c702 AA |
1714 | * userfaultfd_wake may be used in combination with the |
1715 | * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches. | |
86039bd3 AA |
1716 | */ |
1717 | static int userfaultfd_wake(struct userfaultfd_ctx *ctx, | |
1718 | unsigned long arg) | |
1719 | { | |
1720 | int ret; | |
1721 | struct uffdio_range uffdio_wake; | |
1722 | struct userfaultfd_wake_range range; | |
1723 | const void __user *buf = (void __user *)arg; | |
1724 | ||
1725 | ret = -EFAULT; | |
1726 | if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake))) | |
1727 | goto out; | |
1728 | ||
e71e2ace | 1729 | ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len); |
86039bd3 AA |
1730 | if (ret) |
1731 | goto out; | |
1732 | ||
1733 | range.start = uffdio_wake.start; | |
1734 | range.len = uffdio_wake.len; | |
1735 | ||
1736 | /* | |
1737 | * len == 0 means wake all and we don't want to wake all here, | |
1738 | * so check it again to be sure. | |
1739 | */ | |
1740 | VM_BUG_ON(!range.len); | |
1741 | ||
1742 | wake_userfault(ctx, &range); | |
1743 | ret = 0; | |
1744 | ||
1745 | out: | |
1746 | return ret; | |
1747 | } | |
1748 | ||
ad465cae AA |
1749 | static int userfaultfd_copy(struct userfaultfd_ctx *ctx, |
1750 | unsigned long arg) | |
1751 | { | |
1752 | __s64 ret; | |
1753 | struct uffdio_copy uffdio_copy; | |
1754 | struct uffdio_copy __user *user_uffdio_copy; | |
1755 | struct userfaultfd_wake_range range; | |
d9712937 | 1756 | uffd_flags_t flags = 0; |
ad465cae AA |
1757 | |
1758 | user_uffdio_copy = (struct uffdio_copy __user *) arg; | |
1759 | ||
df2cc96e | 1760 | ret = -EAGAIN; |
a759a909 | 1761 | if (atomic_read(&ctx->mmap_changing)) |
df2cc96e MR |
1762 | goto out; |
1763 | ||
ad465cae AA |
1764 | ret = -EFAULT; |
1765 | if (copy_from_user(&uffdio_copy, user_uffdio_copy, | |
1766 | /* don't copy "copy" last field */ | |
1767 | sizeof(uffdio_copy)-sizeof(__s64))) | |
1768 | goto out; | |
1769 | ||
e71e2ace | 1770 | ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len); |
ad465cae AA |
1771 | if (ret) |
1772 | goto out; | |
1773 | /* | |
1774 | * double check for wraparound just in case. copy_from_user() | |
1775 | * will later check uffdio_copy.src + uffdio_copy.len to fit | |
1776 | * in the userland range. | |
1777 | */ | |
1778 | ret = -EINVAL; | |
1779 | if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src) | |
1780 | goto out; | |
72981e0e | 1781 | if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP)) |
ad465cae | 1782 | goto out; |
d9712937 AR |
1783 | if (uffdio_copy.mode & UFFDIO_COPY_MODE_WP) |
1784 | flags |= MFILL_ATOMIC_WP; | |
d2005e3f | 1785 | if (mmget_not_zero(ctx->mm)) { |
a734991c AR |
1786 | ret = mfill_atomic_copy(ctx->mm, uffdio_copy.dst, uffdio_copy.src, |
1787 | uffdio_copy.len, &ctx->mmap_changing, | |
d9712937 | 1788 | flags); |
d2005e3f | 1789 | mmput(ctx->mm); |
96333187 | 1790 | } else { |
e86b298b | 1791 | return -ESRCH; |
d2005e3f | 1792 | } |
ad465cae AA |
1793 | if (unlikely(put_user(ret, &user_uffdio_copy->copy))) |
1794 | return -EFAULT; | |
1795 | if (ret < 0) | |
1796 | goto out; | |
1797 | BUG_ON(!ret); | |
1798 | /* len == 0 would wake all */ | |
1799 | range.len = ret; | |
1800 | if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) { | |
1801 | range.start = uffdio_copy.dst; | |
1802 | wake_userfault(ctx, &range); | |
1803 | } | |
1804 | ret = range.len == uffdio_copy.len ? 0 : -EAGAIN; | |
1805 | out: | |
1806 | return ret; | |
1807 | } | |
1808 | ||
1809 | static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx, | |
1810 | unsigned long arg) | |
1811 | { | |
1812 | __s64 ret; | |
1813 | struct uffdio_zeropage uffdio_zeropage; | |
1814 | struct uffdio_zeropage __user *user_uffdio_zeropage; | |
1815 | struct userfaultfd_wake_range range; | |
1816 | ||
1817 | user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg; | |
1818 | ||
df2cc96e | 1819 | ret = -EAGAIN; |
a759a909 | 1820 | if (atomic_read(&ctx->mmap_changing)) |
df2cc96e MR |
1821 | goto out; |
1822 | ||
ad465cae AA |
1823 | ret = -EFAULT; |
1824 | if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage, | |
1825 | /* don't copy "zeropage" last field */ | |
1826 | sizeof(uffdio_zeropage)-sizeof(__s64))) | |
1827 | goto out; | |
1828 | ||
e71e2ace | 1829 | ret = validate_range(ctx->mm, uffdio_zeropage.range.start, |
ad465cae AA |
1830 | uffdio_zeropage.range.len); |
1831 | if (ret) | |
1832 | goto out; | |
1833 | ret = -EINVAL; | |
1834 | if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE) | |
1835 | goto out; | |
1836 | ||
d2005e3f | 1837 | if (mmget_not_zero(ctx->mm)) { |
a734991c AR |
1838 | ret = mfill_atomic_zeropage(ctx->mm, uffdio_zeropage.range.start, |
1839 | uffdio_zeropage.range.len, | |
1840 | &ctx->mmap_changing); | |
d2005e3f | 1841 | mmput(ctx->mm); |
9d95aa4b | 1842 | } else { |
e86b298b | 1843 | return -ESRCH; |
d2005e3f | 1844 | } |
ad465cae AA |
1845 | if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage))) |
1846 | return -EFAULT; | |
1847 | if (ret < 0) | |
1848 | goto out; | |
1849 | /* len == 0 would wake all */ | |
1850 | BUG_ON(!ret); | |
1851 | range.len = ret; | |
1852 | if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) { | |
1853 | range.start = uffdio_zeropage.range.start; | |
1854 | wake_userfault(ctx, &range); | |
1855 | } | |
1856 | ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN; | |
1857 | out: | |
1858 | return ret; | |
1859 | } | |
1860 | ||
63b2d417 AA |
1861 | static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx, |
1862 | unsigned long arg) | |
1863 | { | |
1864 | int ret; | |
1865 | struct uffdio_writeprotect uffdio_wp; | |
1866 | struct uffdio_writeprotect __user *user_uffdio_wp; | |
1867 | struct userfaultfd_wake_range range; | |
23080e27 | 1868 | bool mode_wp, mode_dontwake; |
63b2d417 | 1869 | |
a759a909 | 1870 | if (atomic_read(&ctx->mmap_changing)) |
63b2d417 AA |
1871 | return -EAGAIN; |
1872 | ||
1873 | user_uffdio_wp = (struct uffdio_writeprotect __user *) arg; | |
1874 | ||
1875 | if (copy_from_user(&uffdio_wp, user_uffdio_wp, | |
1876 | sizeof(struct uffdio_writeprotect))) | |
1877 | return -EFAULT; | |
1878 | ||
e71e2ace | 1879 | ret = validate_range(ctx->mm, uffdio_wp.range.start, |
63b2d417 AA |
1880 | uffdio_wp.range.len); |
1881 | if (ret) | |
1882 | return ret; | |
1883 | ||
1884 | if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE | | |
1885 | UFFDIO_WRITEPROTECT_MODE_WP)) | |
1886 | return -EINVAL; | |
23080e27 PX |
1887 | |
1888 | mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP; | |
1889 | mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE; | |
1890 | ||
1891 | if (mode_wp && mode_dontwake) | |
63b2d417 AA |
1892 | return -EINVAL; |
1893 | ||
cb185d5f NA |
1894 | if (mmget_not_zero(ctx->mm)) { |
1895 | ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start, | |
1896 | uffdio_wp.range.len, mode_wp, | |
1897 | &ctx->mmap_changing); | |
1898 | mmput(ctx->mm); | |
1899 | } else { | |
1900 | return -ESRCH; | |
1901 | } | |
1902 | ||
63b2d417 AA |
1903 | if (ret) |
1904 | return ret; | |
1905 | ||
23080e27 | 1906 | if (!mode_wp && !mode_dontwake) { |
63b2d417 AA |
1907 | range.start = uffdio_wp.range.start; |
1908 | range.len = uffdio_wp.range.len; | |
1909 | wake_userfault(ctx, &range); | |
1910 | } | |
1911 | return ret; | |
1912 | } | |
1913 | ||
f6191471 AR |
1914 | static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg) |
1915 | { | |
1916 | __s64 ret; | |
1917 | struct uffdio_continue uffdio_continue; | |
1918 | struct uffdio_continue __user *user_uffdio_continue; | |
1919 | struct userfaultfd_wake_range range; | |
02891844 | 1920 | uffd_flags_t flags = 0; |
f6191471 AR |
1921 | |
1922 | user_uffdio_continue = (struct uffdio_continue __user *)arg; | |
1923 | ||
1924 | ret = -EAGAIN; | |
a759a909 | 1925 | if (atomic_read(&ctx->mmap_changing)) |
f6191471 AR |
1926 | goto out; |
1927 | ||
1928 | ret = -EFAULT; | |
1929 | if (copy_from_user(&uffdio_continue, user_uffdio_continue, | |
1930 | /* don't copy the output fields */ | |
1931 | sizeof(uffdio_continue) - (sizeof(__s64)))) | |
1932 | goto out; | |
1933 | ||
e71e2ace | 1934 | ret = validate_range(ctx->mm, uffdio_continue.range.start, |
f6191471 AR |
1935 | uffdio_continue.range.len); |
1936 | if (ret) | |
1937 | goto out; | |
1938 | ||
1939 | ret = -EINVAL; | |
1940 | /* double check for wraparound just in case. */ | |
1941 | if (uffdio_continue.range.start + uffdio_continue.range.len <= | |
1942 | uffdio_continue.range.start) { | |
1943 | goto out; | |
1944 | } | |
02891844 AR |
1945 | if (uffdio_continue.mode & ~(UFFDIO_CONTINUE_MODE_DONTWAKE | |
1946 | UFFDIO_CONTINUE_MODE_WP)) | |
f6191471 | 1947 | goto out; |
02891844 AR |
1948 | if (uffdio_continue.mode & UFFDIO_CONTINUE_MODE_WP) |
1949 | flags |= MFILL_ATOMIC_WP; | |
f6191471 AR |
1950 | |
1951 | if (mmget_not_zero(ctx->mm)) { | |
a734991c AR |
1952 | ret = mfill_atomic_continue(ctx->mm, uffdio_continue.range.start, |
1953 | uffdio_continue.range.len, | |
02891844 | 1954 | &ctx->mmap_changing, flags); |
f6191471 AR |
1955 | mmput(ctx->mm); |
1956 | } else { | |
1957 | return -ESRCH; | |
1958 | } | |
1959 | ||
1960 | if (unlikely(put_user(ret, &user_uffdio_continue->mapped))) | |
1961 | return -EFAULT; | |
1962 | if (ret < 0) | |
1963 | goto out; | |
1964 | ||
1965 | /* len == 0 would wake all */ | |
1966 | BUG_ON(!ret); | |
1967 | range.len = ret; | |
1968 | if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) { | |
1969 | range.start = uffdio_continue.range.start; | |
1970 | wake_userfault(ctx, &range); | |
1971 | } | |
1972 | ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN; | |
1973 | ||
1974 | out: | |
1975 | return ret; | |
1976 | } | |
1977 | ||
9cd75c3c PE |
1978 | static inline unsigned int uffd_ctx_features(__u64 user_features) |
1979 | { | |
1980 | /* | |
22e5fe2a NA |
1981 | * For the current set of features the bits just coincide. Set |
1982 | * UFFD_FEATURE_INITIALIZED to mark the features as enabled. | |
9cd75c3c | 1983 | */ |
22e5fe2a | 1984 | return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED; |
9cd75c3c PE |
1985 | } |
1986 | ||
86039bd3 AA |
1987 | /* |
1988 | * userland asks for a certain API version and we return which bits | |
1989 | * and ioctl commands are implemented in this kernel for such API | |
1990 | * version or -EINVAL if unknown. | |
1991 | */ | |
1992 | static int userfaultfd_api(struct userfaultfd_ctx *ctx, | |
1993 | unsigned long arg) | |
1994 | { | |
1995 | struct uffdio_api uffdio_api; | |
1996 | void __user *buf = (void __user *)arg; | |
22e5fe2a | 1997 | unsigned int ctx_features; |
86039bd3 | 1998 | int ret; |
65603144 | 1999 | __u64 features; |
86039bd3 | 2000 | |
86039bd3 | 2001 | ret = -EFAULT; |
a9b85f94 | 2002 | if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api))) |
86039bd3 | 2003 | goto out; |
2ff559f3 PX |
2004 | features = uffdio_api.features; |
2005 | ret = -EINVAL; | |
2006 | if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) | |
2007 | goto err_out; | |
3c1c24d9 MR |
2008 | ret = -EPERM; |
2009 | if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE)) | |
2010 | goto err_out; | |
65603144 AA |
2011 | /* report all available features and ioctls to userland */ |
2012 | uffdio_api.features = UFFD_API_FEATURES; | |
7677f7fd | 2013 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR |
964ab004 AR |
2014 | uffdio_api.features &= |
2015 | ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM); | |
00b151f2 PX |
2016 | #endif |
2017 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP | |
2018 | uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP; | |
b1f9e876 PX |
2019 | #endif |
2020 | #ifndef CONFIG_PTE_MARKER_UFFD_WP | |
2021 | uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM; | |
2bad466c | 2022 | uffdio_api.features &= ~UFFD_FEATURE_WP_UNPOPULATED; |
7677f7fd | 2023 | #endif |
86039bd3 AA |
2024 | uffdio_api.ioctls = UFFD_API_IOCTLS; |
2025 | ret = -EFAULT; | |
2026 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
2027 | goto out; | |
22e5fe2a | 2028 | |
65603144 | 2029 | /* only enable the requested features for this uffd context */ |
22e5fe2a NA |
2030 | ctx_features = uffd_ctx_features(features); |
2031 | ret = -EINVAL; | |
2032 | if (cmpxchg(&ctx->features, 0, ctx_features) != 0) | |
2033 | goto err_out; | |
2034 | ||
86039bd3 AA |
2035 | ret = 0; |
2036 | out: | |
2037 | return ret; | |
3c1c24d9 MR |
2038 | err_out: |
2039 | memset(&uffdio_api, 0, sizeof(uffdio_api)); | |
2040 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
2041 | ret = -EFAULT; | |
2042 | goto out; | |
86039bd3 AA |
2043 | } |
2044 | ||
2045 | static long userfaultfd_ioctl(struct file *file, unsigned cmd, | |
2046 | unsigned long arg) | |
2047 | { | |
2048 | int ret = -EINVAL; | |
2049 | struct userfaultfd_ctx *ctx = file->private_data; | |
2050 | ||
22e5fe2a | 2051 | if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx)) |
e6485a47 AA |
2052 | return -EINVAL; |
2053 | ||
86039bd3 AA |
2054 | switch(cmd) { |
2055 | case UFFDIO_API: | |
2056 | ret = userfaultfd_api(ctx, arg); | |
2057 | break; | |
2058 | case UFFDIO_REGISTER: | |
2059 | ret = userfaultfd_register(ctx, arg); | |
2060 | break; | |
2061 | case UFFDIO_UNREGISTER: | |
2062 | ret = userfaultfd_unregister(ctx, arg); | |
2063 | break; | |
2064 | case UFFDIO_WAKE: | |
2065 | ret = userfaultfd_wake(ctx, arg); | |
2066 | break; | |
ad465cae AA |
2067 | case UFFDIO_COPY: |
2068 | ret = userfaultfd_copy(ctx, arg); | |
2069 | break; | |
2070 | case UFFDIO_ZEROPAGE: | |
2071 | ret = userfaultfd_zeropage(ctx, arg); | |
2072 | break; | |
63b2d417 AA |
2073 | case UFFDIO_WRITEPROTECT: |
2074 | ret = userfaultfd_writeprotect(ctx, arg); | |
2075 | break; | |
f6191471 AR |
2076 | case UFFDIO_CONTINUE: |
2077 | ret = userfaultfd_continue(ctx, arg); | |
2078 | break; | |
86039bd3 AA |
2079 | } |
2080 | return ret; | |
2081 | } | |
2082 | ||
2083 | #ifdef CONFIG_PROC_FS | |
2084 | static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f) | |
2085 | { | |
2086 | struct userfaultfd_ctx *ctx = f->private_data; | |
ac6424b9 | 2087 | wait_queue_entry_t *wq; |
86039bd3 AA |
2088 | unsigned long pending = 0, total = 0; |
2089 | ||
cbcfa130 | 2090 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
2055da97 | 2091 | list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) { |
15b726ef AA |
2092 | pending++; |
2093 | total++; | |
2094 | } | |
2055da97 | 2095 | list_for_each_entry(wq, &ctx->fault_wqh.head, entry) { |
86039bd3 AA |
2096 | total++; |
2097 | } | |
cbcfa130 | 2098 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
2099 | |
2100 | /* | |
2101 | * If more protocols will be added, there will be all shown | |
2102 | * separated by a space. Like this: | |
2103 | * protocols: aa:... bb:... | |
2104 | */ | |
2105 | seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n", | |
045098e9 | 2106 | pending, total, UFFD_API, ctx->features, |
86039bd3 AA |
2107 | UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS); |
2108 | } | |
2109 | #endif | |
2110 | ||
2111 | static const struct file_operations userfaultfd_fops = { | |
2112 | #ifdef CONFIG_PROC_FS | |
2113 | .show_fdinfo = userfaultfd_show_fdinfo, | |
2114 | #endif | |
2115 | .release = userfaultfd_release, | |
2116 | .poll = userfaultfd_poll, | |
2117 | .read = userfaultfd_read, | |
2118 | .unlocked_ioctl = userfaultfd_ioctl, | |
1832f2d8 | 2119 | .compat_ioctl = compat_ptr_ioctl, |
86039bd3 AA |
2120 | .llseek = noop_llseek, |
2121 | }; | |
2122 | ||
3004ec9c AA |
2123 | static void init_once_userfaultfd_ctx(void *mem) |
2124 | { | |
2125 | struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem; | |
2126 | ||
2127 | init_waitqueue_head(&ctx->fault_pending_wqh); | |
2128 | init_waitqueue_head(&ctx->fault_wqh); | |
9cd75c3c | 2129 | init_waitqueue_head(&ctx->event_wqh); |
3004ec9c | 2130 | init_waitqueue_head(&ctx->fd_wqh); |
2ca97ac8 | 2131 | seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock); |
3004ec9c AA |
2132 | } |
2133 | ||
2d5de004 | 2134 | static int new_userfaultfd(int flags) |
86039bd3 | 2135 | { |
86039bd3 | 2136 | struct userfaultfd_ctx *ctx; |
284cd241 | 2137 | int fd; |
86039bd3 AA |
2138 | |
2139 | BUG_ON(!current->mm); | |
2140 | ||
2141 | /* Check the UFFD_* constants for consistency. */ | |
37cd0575 | 2142 | BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS); |
86039bd3 AA |
2143 | BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC); |
2144 | BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK); | |
2145 | ||
37cd0575 | 2146 | if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY)) |
284cd241 | 2147 | return -EINVAL; |
86039bd3 | 2148 | |
3004ec9c | 2149 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); |
86039bd3 | 2150 | if (!ctx) |
284cd241 | 2151 | return -ENOMEM; |
86039bd3 | 2152 | |
ca880420 | 2153 | refcount_set(&ctx->refcount, 1); |
86039bd3 | 2154 | ctx->flags = flags; |
9cd75c3c | 2155 | ctx->features = 0; |
86039bd3 | 2156 | ctx->released = false; |
a759a909 | 2157 | atomic_set(&ctx->mmap_changing, 0); |
86039bd3 AA |
2158 | ctx->mm = current->mm; |
2159 | /* prevent the mm struct to be freed */ | |
f1f10076 | 2160 | mmgrab(ctx->mm); |
86039bd3 | 2161 | |
b537900f | 2162 | fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx, |
abec3d01 | 2163 | O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL); |
284cd241 | 2164 | if (fd < 0) { |
d2005e3f | 2165 | mmdrop(ctx->mm); |
3004ec9c | 2166 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
c03e946f | 2167 | } |
86039bd3 | 2168 | return fd; |
86039bd3 | 2169 | } |
3004ec9c | 2170 | |
2d5de004 AR |
2171 | static inline bool userfaultfd_syscall_allowed(int flags) |
2172 | { | |
2173 | /* Userspace-only page faults are always allowed */ | |
2174 | if (flags & UFFD_USER_MODE_ONLY) | |
2175 | return true; | |
2176 | ||
2177 | /* | |
2178 | * The user is requesting a userfaultfd which can handle kernel faults. | |
2179 | * Privileged users are always allowed to do this. | |
2180 | */ | |
2181 | if (capable(CAP_SYS_PTRACE)) | |
2182 | return true; | |
2183 | ||
2184 | /* Otherwise, access to kernel fault handling is sysctl controlled. */ | |
2185 | return sysctl_unprivileged_userfaultfd; | |
2186 | } | |
2187 | ||
2188 | SYSCALL_DEFINE1(userfaultfd, int, flags) | |
2189 | { | |
2190 | if (!userfaultfd_syscall_allowed(flags)) | |
2191 | return -EPERM; | |
2192 | ||
2193 | return new_userfaultfd(flags); | |
2194 | } | |
2195 | ||
2196 | static long userfaultfd_dev_ioctl(struct file *file, unsigned int cmd, unsigned long flags) | |
2197 | { | |
2198 | if (cmd != USERFAULTFD_IOC_NEW) | |
2199 | return -EINVAL; | |
2200 | ||
2201 | return new_userfaultfd(flags); | |
2202 | } | |
2203 | ||
2204 | static const struct file_operations userfaultfd_dev_fops = { | |
2205 | .unlocked_ioctl = userfaultfd_dev_ioctl, | |
2206 | .compat_ioctl = userfaultfd_dev_ioctl, | |
2207 | .owner = THIS_MODULE, | |
2208 | .llseek = noop_llseek, | |
2209 | }; | |
2210 | ||
2211 | static struct miscdevice userfaultfd_misc = { | |
2212 | .minor = MISC_DYNAMIC_MINOR, | |
2213 | .name = "userfaultfd", | |
2214 | .fops = &userfaultfd_dev_fops | |
2215 | }; | |
2216 | ||
3004ec9c AA |
2217 | static int __init userfaultfd_init(void) |
2218 | { | |
2d5de004 AR |
2219 | int ret; |
2220 | ||
2221 | ret = misc_register(&userfaultfd_misc); | |
2222 | if (ret) | |
2223 | return ret; | |
2224 | ||
3004ec9c AA |
2225 | userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache", |
2226 | sizeof(struct userfaultfd_ctx), | |
2227 | 0, | |
2228 | SLAB_HWCACHE_ALIGN|SLAB_PANIC, | |
2229 | init_once_userfaultfd_ctx); | |
2d337b71 Z |
2230 | #ifdef CONFIG_SYSCTL |
2231 | register_sysctl_init("vm", vm_userfaultfd_table); | |
2232 | #endif | |
3004ec9c AA |
2233 | return 0; |
2234 | } | |
2235 | __initcall(userfaultfd_init); |