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; | |
c33c7948 | 338 | pte_t ptent; |
8d2afd96 AA |
339 | bool ret = true; |
340 | ||
42fc5414 | 341 | mmap_assert_locked(mm); |
8d2afd96 AA |
342 | |
343 | pgd = pgd_offset(mm, address); | |
344 | if (!pgd_present(*pgd)) | |
345 | goto out; | |
c2febafc KS |
346 | p4d = p4d_offset(pgd, address); |
347 | if (!p4d_present(*p4d)) | |
348 | goto out; | |
349 | pud = pud_offset(p4d, address); | |
8d2afd96 AA |
350 | if (!pud_present(*pud)) |
351 | goto out; | |
352 | pmd = pmd_offset(pud, address); | |
2b683a4f | 353 | again: |
26e1a0c3 | 354 | _pmd = pmdp_get_lockless(pmd); |
a365ac09 | 355 | if (pmd_none(_pmd)) |
8d2afd96 AA |
356 | goto out; |
357 | ||
358 | ret = false; | |
2b683a4f | 359 | if (!pmd_present(_pmd) || pmd_devmap(_pmd)) |
a365ac09 HY |
360 | goto out; |
361 | ||
63b2d417 AA |
362 | if (pmd_trans_huge(_pmd)) { |
363 | if (!pmd_write(_pmd) && (reason & VM_UFFD_WP)) | |
364 | ret = true; | |
8d2afd96 | 365 | goto out; |
63b2d417 | 366 | } |
8d2afd96 | 367 | |
8d2afd96 | 368 | pte = pte_offset_map(pmd, address); |
2b683a4f HD |
369 | if (!pte) { |
370 | ret = true; | |
371 | goto again; | |
372 | } | |
8d2afd96 AA |
373 | /* |
374 | * Lockless access: we're in a wait_event so it's ok if it | |
5c041f5d PX |
375 | * changes under us. PTE markers should be handled the same as none |
376 | * ptes here. | |
8d2afd96 | 377 | */ |
c33c7948 RR |
378 | ptent = ptep_get(pte); |
379 | if (pte_none_mostly(ptent)) | |
8d2afd96 | 380 | ret = true; |
c33c7948 | 381 | if (!pte_write(ptent) && (reason & VM_UFFD_WP)) |
63b2d417 | 382 | ret = true; |
8d2afd96 AA |
383 | pte_unmap(pte); |
384 | ||
385 | out: | |
386 | return ret; | |
387 | } | |
388 | ||
2f064a59 | 389 | static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags) |
3e69ad08 PX |
390 | { |
391 | if (flags & FAULT_FLAG_INTERRUPTIBLE) | |
392 | return TASK_INTERRUPTIBLE; | |
393 | ||
394 | if (flags & FAULT_FLAG_KILLABLE) | |
395 | return TASK_KILLABLE; | |
396 | ||
397 | return TASK_UNINTERRUPTIBLE; | |
398 | } | |
399 | ||
86039bd3 AA |
400 | /* |
401 | * The locking rules involved in returning VM_FAULT_RETRY depending on | |
402 | * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and | |
403 | * FAULT_FLAG_KILLABLE are not straightforward. The "Caution" | |
404 | * recommendation in __lock_page_or_retry is not an understatement. | |
405 | * | |
c1e8d7c6 | 406 | * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released |
86039bd3 AA |
407 | * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is |
408 | * not set. | |
409 | * | |
410 | * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not | |
411 | * set, VM_FAULT_RETRY can still be returned if and only if there are | |
c1e8d7c6 | 412 | * fatal_signal_pending()s, and the mmap_lock must be released before |
86039bd3 AA |
413 | * returning it. |
414 | */ | |
2b740303 | 415 | vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason) |
86039bd3 | 416 | { |
b8da2e46 PX |
417 | struct vm_area_struct *vma = vmf->vma; |
418 | struct mm_struct *mm = vma->vm_mm; | |
86039bd3 AA |
419 | struct userfaultfd_ctx *ctx; |
420 | struct userfaultfd_wait_queue uwq; | |
2b740303 | 421 | vm_fault_t ret = VM_FAULT_SIGBUS; |
3e69ad08 | 422 | bool must_wait; |
2f064a59 | 423 | unsigned int blocking_state; |
86039bd3 | 424 | |
64c2b203 AA |
425 | /* |
426 | * We don't do userfault handling for the final child pid update. | |
427 | * | |
428 | * We also don't do userfault handling during | |
429 | * coredumping. hugetlbfs has the special | |
430 | * follow_hugetlb_page() to skip missing pages in the | |
431 | * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with | |
432 | * the no_page_table() helper in follow_page_mask(), but the | |
433 | * shmem_vm_ops->fault method is invoked even during | |
c1e8d7c6 | 434 | * coredumping without mmap_lock and it ends up here. |
64c2b203 AA |
435 | */ |
436 | if (current->flags & (PF_EXITING|PF_DUMPCORE)) | |
437 | goto out; | |
438 | ||
439 | /* | |
c1e8d7c6 ML |
440 | * Coredumping runs without mmap_lock so we can only check that |
441 | * the mmap_lock is held, if PF_DUMPCORE was not set. | |
64c2b203 | 442 | */ |
42fc5414 | 443 | mmap_assert_locked(mm); |
64c2b203 | 444 | |
b8da2e46 | 445 | ctx = vma->vm_userfaultfd_ctx.ctx; |
86039bd3 | 446 | if (!ctx) |
ba85c702 | 447 | goto out; |
86039bd3 AA |
448 | |
449 | BUG_ON(ctx->mm != mm); | |
450 | ||
7677f7fd AR |
451 | /* Any unrecognized flag is a bug. */ |
452 | VM_BUG_ON(reason & ~__VM_UFFD_FLAGS); | |
453 | /* 0 or > 1 flags set is a bug; we expect exactly 1. */ | |
454 | VM_BUG_ON(!reason || (reason & (reason - 1))); | |
86039bd3 | 455 | |
2d6d6f5a PS |
456 | if (ctx->features & UFFD_FEATURE_SIGBUS) |
457 | goto out; | |
2d5de004 | 458 | if (!(vmf->flags & FAULT_FLAG_USER) && (ctx->flags & UFFD_USER_MODE_ONLY)) |
37cd0575 | 459 | goto out; |
2d6d6f5a | 460 | |
86039bd3 AA |
461 | /* |
462 | * If it's already released don't get it. This avoids to loop | |
463 | * in __get_user_pages if userfaultfd_release waits on the | |
c1e8d7c6 | 464 | * caller of handle_userfault to release the mmap_lock. |
86039bd3 | 465 | */ |
6aa7de05 | 466 | if (unlikely(READ_ONCE(ctx->released))) { |
656710a6 AA |
467 | /* |
468 | * Don't return VM_FAULT_SIGBUS in this case, so a non | |
469 | * cooperative manager can close the uffd after the | |
470 | * last UFFDIO_COPY, without risking to trigger an | |
471 | * involuntary SIGBUS if the process was starting the | |
472 | * userfaultfd while the userfaultfd was still armed | |
473 | * (but after the last UFFDIO_COPY). If the uffd | |
474 | * wasn't already closed when the userfault reached | |
475 | * this point, that would normally be solved by | |
476 | * userfaultfd_must_wait returning 'false'. | |
477 | * | |
478 | * If we were to return VM_FAULT_SIGBUS here, the non | |
479 | * cooperative manager would be instead forced to | |
480 | * always call UFFDIO_UNREGISTER before it can safely | |
481 | * close the uffd. | |
482 | */ | |
483 | ret = VM_FAULT_NOPAGE; | |
ba85c702 | 484 | goto out; |
656710a6 | 485 | } |
86039bd3 AA |
486 | |
487 | /* | |
488 | * Check that we can return VM_FAULT_RETRY. | |
489 | * | |
490 | * NOTE: it should become possible to return VM_FAULT_RETRY | |
491 | * even if FAULT_FLAG_TRIED is set without leading to gup() | |
492 | * -EBUSY failures, if the userfaultfd is to be extended for | |
493 | * VM_UFFD_WP tracking and we intend to arm the userfault | |
494 | * without first stopping userland access to the memory. For | |
495 | * VM_UFFD_MISSING userfaults this is enough for now. | |
496 | */ | |
82b0f8c3 | 497 | if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) { |
86039bd3 AA |
498 | /* |
499 | * Validate the invariant that nowait must allow retry | |
500 | * to be sure not to return SIGBUS erroneously on | |
501 | * nowait invocations. | |
502 | */ | |
82b0f8c3 | 503 | BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT); |
86039bd3 AA |
504 | #ifdef CONFIG_DEBUG_VM |
505 | if (printk_ratelimit()) { | |
506 | printk(KERN_WARNING | |
82b0f8c3 JK |
507 | "FAULT_FLAG_ALLOW_RETRY missing %x\n", |
508 | vmf->flags); | |
86039bd3 AA |
509 | dump_stack(); |
510 | } | |
511 | #endif | |
ba85c702 | 512 | goto out; |
86039bd3 AA |
513 | } |
514 | ||
515 | /* | |
516 | * Handle nowait, not much to do other than tell it to retry | |
517 | * and wait. | |
518 | */ | |
ba85c702 | 519 | ret = VM_FAULT_RETRY; |
82b0f8c3 | 520 | if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) |
ba85c702 | 521 | goto out; |
86039bd3 | 522 | |
c1e8d7c6 | 523 | /* take the reference before dropping the mmap_lock */ |
86039bd3 AA |
524 | userfaultfd_ctx_get(ctx); |
525 | ||
86039bd3 AA |
526 | init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function); |
527 | uwq.wq.private = current; | |
d172b1a3 NA |
528 | uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags, |
529 | reason, ctx->features); | |
86039bd3 | 530 | uwq.ctx = ctx; |
15a77c6f | 531 | uwq.waken = false; |
86039bd3 | 532 | |
3e69ad08 | 533 | blocking_state = userfaultfd_get_blocking_state(vmf->flags); |
dfa37dc3 | 534 | |
b8da2e46 PX |
535 | /* |
536 | * Take the vma lock now, in order to safely call | |
537 | * userfaultfd_huge_must_wait() later. Since acquiring the | |
538 | * (sleepable) vma lock can modify the current task state, that | |
539 | * must be before explicitly calling set_current_state(). | |
540 | */ | |
541 | if (is_vm_hugetlb_page(vma)) | |
542 | hugetlb_vma_lock_read(vma); | |
543 | ||
cbcfa130 | 544 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
545 | /* |
546 | * After the __add_wait_queue the uwq is visible to userland | |
547 | * through poll/read(). | |
548 | */ | |
15b726ef AA |
549 | __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq); |
550 | /* | |
551 | * The smp_mb() after __set_current_state prevents the reads | |
552 | * following the spin_unlock to happen before the list_add in | |
553 | * __add_wait_queue. | |
554 | */ | |
15a77c6f | 555 | set_current_state(blocking_state); |
cbcfa130 | 556 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 557 | |
b8da2e46 | 558 | if (!is_vm_hugetlb_page(vma)) |
369cd212 MK |
559 | must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags, |
560 | reason); | |
561 | else | |
b8da2e46 | 562 | must_wait = userfaultfd_huge_must_wait(ctx, vma, |
7868a208 | 563 | vmf->address, |
369cd212 | 564 | vmf->flags, reason); |
b8da2e46 PX |
565 | if (is_vm_hugetlb_page(vma)) |
566 | hugetlb_vma_unlock_read(vma); | |
d8ed45c5 | 567 | mmap_read_unlock(mm); |
8d2afd96 | 568 | |
f9bf3522 | 569 | if (likely(must_wait && !READ_ONCE(ctx->released))) { |
a9a08845 | 570 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
86039bd3 | 571 | schedule(); |
ba85c702 | 572 | } |
86039bd3 | 573 | |
ba85c702 | 574 | __set_current_state(TASK_RUNNING); |
15b726ef AA |
575 | |
576 | /* | |
577 | * Here we race with the list_del; list_add in | |
578 | * userfaultfd_ctx_read(), however because we don't ever run | |
579 | * list_del_init() to refile across the two lists, the prev | |
580 | * and next pointers will never point to self. list_add also | |
581 | * would never let any of the two pointers to point to | |
582 | * self. So list_empty_careful won't risk to see both pointers | |
583 | * pointing to self at any time during the list refile. The | |
584 | * only case where list_del_init() is called is the full | |
585 | * removal in the wake function and there we don't re-list_add | |
586 | * and it's fine not to block on the spinlock. The uwq on this | |
587 | * kernel stack can be released after the list_del_init. | |
588 | */ | |
2055da97 | 589 | if (!list_empty_careful(&uwq.wq.entry)) { |
cbcfa130 | 590 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
15b726ef AA |
591 | /* |
592 | * No need of list_del_init(), the uwq on the stack | |
593 | * will be freed shortly anyway. | |
594 | */ | |
2055da97 | 595 | list_del(&uwq.wq.entry); |
cbcfa130 | 596 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 597 | } |
86039bd3 AA |
598 | |
599 | /* | |
600 | * ctx may go away after this if the userfault pseudo fd is | |
601 | * already released. | |
602 | */ | |
603 | userfaultfd_ctx_put(ctx); | |
604 | ||
ba85c702 AA |
605 | out: |
606 | return ret; | |
86039bd3 AA |
607 | } |
608 | ||
8c9e7bb7 AA |
609 | static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx, |
610 | struct userfaultfd_wait_queue *ewq) | |
9cd75c3c | 611 | { |
0cbb4b4f AA |
612 | struct userfaultfd_ctx *release_new_ctx; |
613 | ||
9a69a829 AA |
614 | if (WARN_ON_ONCE(current->flags & PF_EXITING)) |
615 | goto out; | |
9cd75c3c PE |
616 | |
617 | ewq->ctx = ctx; | |
618 | init_waitqueue_entry(&ewq->wq, current); | |
0cbb4b4f | 619 | release_new_ctx = NULL; |
9cd75c3c | 620 | |
cbcfa130 | 621 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
622 | /* |
623 | * After the __add_wait_queue the uwq is visible to userland | |
624 | * through poll/read(). | |
625 | */ | |
626 | __add_wait_queue(&ctx->event_wqh, &ewq->wq); | |
627 | for (;;) { | |
628 | set_current_state(TASK_KILLABLE); | |
629 | if (ewq->msg.event == 0) | |
630 | break; | |
6aa7de05 | 631 | if (READ_ONCE(ctx->released) || |
9cd75c3c | 632 | fatal_signal_pending(current)) { |
384632e6 AA |
633 | /* |
634 | * &ewq->wq may be queued in fork_event, but | |
635 | * __remove_wait_queue ignores the head | |
636 | * parameter. It would be a problem if it | |
637 | * didn't. | |
638 | */ | |
9cd75c3c | 639 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); |
7eb76d45 MR |
640 | if (ewq->msg.event == UFFD_EVENT_FORK) { |
641 | struct userfaultfd_ctx *new; | |
642 | ||
643 | new = (struct userfaultfd_ctx *) | |
644 | (unsigned long) | |
645 | ewq->msg.arg.reserved.reserved1; | |
0cbb4b4f | 646 | release_new_ctx = new; |
7eb76d45 | 647 | } |
9cd75c3c PE |
648 | break; |
649 | } | |
650 | ||
cbcfa130 | 651 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 652 | |
a9a08845 | 653 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
9cd75c3c PE |
654 | schedule(); |
655 | ||
cbcfa130 | 656 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
657 | } |
658 | __set_current_state(TASK_RUNNING); | |
cbcfa130 | 659 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 660 | |
0cbb4b4f AA |
661 | if (release_new_ctx) { |
662 | struct vm_area_struct *vma; | |
663 | struct mm_struct *mm = release_new_ctx->mm; | |
69dbe6da | 664 | VMA_ITERATOR(vmi, mm, 0); |
0cbb4b4f AA |
665 | |
666 | /* the various vma->vm_userfaultfd_ctx still points to it */ | |
d8ed45c5 | 667 | mmap_write_lock(mm); |
69dbe6da | 668 | for_each_vma(vmi, vma) { |
31e810aa | 669 | if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) { |
0cbb4b4f | 670 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
51d3d5eb DH |
671 | userfaultfd_set_vm_flags(vma, |
672 | vma->vm_flags & ~__VM_UFFD_FLAGS); | |
31e810aa | 673 | } |
69dbe6da | 674 | } |
d8ed45c5 | 675 | mmap_write_unlock(mm); |
0cbb4b4f AA |
676 | |
677 | userfaultfd_ctx_put(release_new_ctx); | |
678 | } | |
679 | ||
9cd75c3c PE |
680 | /* |
681 | * ctx may go away after this if the userfault pseudo fd is | |
682 | * already released. | |
683 | */ | |
9a69a829 | 684 | out: |
a759a909 NA |
685 | atomic_dec(&ctx->mmap_changing); |
686 | VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0); | |
9cd75c3c | 687 | userfaultfd_ctx_put(ctx); |
9cd75c3c PE |
688 | } |
689 | ||
690 | static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx, | |
691 | struct userfaultfd_wait_queue *ewq) | |
692 | { | |
693 | ewq->msg.event = 0; | |
694 | wake_up_locked(&ctx->event_wqh); | |
695 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); | |
696 | } | |
697 | ||
893e26e6 PE |
698 | int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs) |
699 | { | |
700 | struct userfaultfd_ctx *ctx = NULL, *octx; | |
701 | struct userfaultfd_fork_ctx *fctx; | |
702 | ||
703 | octx = vma->vm_userfaultfd_ctx.ctx; | |
704 | if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) { | |
705 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
51d3d5eb | 706 | userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS); |
893e26e6 PE |
707 | return 0; |
708 | } | |
709 | ||
710 | list_for_each_entry(fctx, fcs, list) | |
711 | if (fctx->orig == octx) { | |
712 | ctx = fctx->new; | |
713 | break; | |
714 | } | |
715 | ||
716 | if (!ctx) { | |
717 | fctx = kmalloc(sizeof(*fctx), GFP_KERNEL); | |
718 | if (!fctx) | |
719 | return -ENOMEM; | |
720 | ||
721 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); | |
722 | if (!ctx) { | |
723 | kfree(fctx); | |
724 | return -ENOMEM; | |
725 | } | |
726 | ||
ca880420 | 727 | refcount_set(&ctx->refcount, 1); |
893e26e6 | 728 | ctx->flags = octx->flags; |
893e26e6 PE |
729 | ctx->features = octx->features; |
730 | ctx->released = false; | |
a759a909 | 731 | atomic_set(&ctx->mmap_changing, 0); |
893e26e6 | 732 | ctx->mm = vma->vm_mm; |
00bb31fa | 733 | mmgrab(ctx->mm); |
893e26e6 PE |
734 | |
735 | userfaultfd_ctx_get(octx); | |
a759a909 | 736 | atomic_inc(&octx->mmap_changing); |
893e26e6 PE |
737 | fctx->orig = octx; |
738 | fctx->new = ctx; | |
739 | list_add_tail(&fctx->list, fcs); | |
740 | } | |
741 | ||
742 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
743 | return 0; | |
744 | } | |
745 | ||
8c9e7bb7 | 746 | static void dup_fctx(struct userfaultfd_fork_ctx *fctx) |
893e26e6 PE |
747 | { |
748 | struct userfaultfd_ctx *ctx = fctx->orig; | |
749 | struct userfaultfd_wait_queue ewq; | |
750 | ||
751 | msg_init(&ewq.msg); | |
752 | ||
753 | ewq.msg.event = UFFD_EVENT_FORK; | |
754 | ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new; | |
755 | ||
8c9e7bb7 | 756 | userfaultfd_event_wait_completion(ctx, &ewq); |
893e26e6 PE |
757 | } |
758 | ||
759 | void dup_userfaultfd_complete(struct list_head *fcs) | |
760 | { | |
893e26e6 PE |
761 | struct userfaultfd_fork_ctx *fctx, *n; |
762 | ||
763 | list_for_each_entry_safe(fctx, n, fcs, list) { | |
8c9e7bb7 | 764 | dup_fctx(fctx); |
893e26e6 PE |
765 | list_del(&fctx->list); |
766 | kfree(fctx); | |
767 | } | |
768 | } | |
769 | ||
72f87654 PE |
770 | void mremap_userfaultfd_prep(struct vm_area_struct *vma, |
771 | struct vm_userfaultfd_ctx *vm_ctx) | |
772 | { | |
773 | struct userfaultfd_ctx *ctx; | |
774 | ||
775 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
3cfd22be PX |
776 | |
777 | if (!ctx) | |
778 | return; | |
779 | ||
780 | if (ctx->features & UFFD_FEATURE_EVENT_REMAP) { | |
72f87654 PE |
781 | vm_ctx->ctx = ctx; |
782 | userfaultfd_ctx_get(ctx); | |
a759a909 | 783 | atomic_inc(&ctx->mmap_changing); |
3cfd22be PX |
784 | } else { |
785 | /* Drop uffd context if remap feature not enabled */ | |
786 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
51d3d5eb | 787 | userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS); |
72f87654 PE |
788 | } |
789 | } | |
790 | ||
90794bf1 | 791 | void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx, |
72f87654 PE |
792 | unsigned long from, unsigned long to, |
793 | unsigned long len) | |
794 | { | |
90794bf1 | 795 | struct userfaultfd_ctx *ctx = vm_ctx->ctx; |
72f87654 PE |
796 | struct userfaultfd_wait_queue ewq; |
797 | ||
798 | if (!ctx) | |
799 | return; | |
800 | ||
801 | if (to & ~PAGE_MASK) { | |
802 | userfaultfd_ctx_put(ctx); | |
803 | return; | |
804 | } | |
805 | ||
806 | msg_init(&ewq.msg); | |
807 | ||
808 | ewq.msg.event = UFFD_EVENT_REMAP; | |
809 | ewq.msg.arg.remap.from = from; | |
810 | ewq.msg.arg.remap.to = to; | |
811 | ewq.msg.arg.remap.len = len; | |
812 | ||
813 | userfaultfd_event_wait_completion(ctx, &ewq); | |
814 | } | |
815 | ||
70ccb92f | 816 | bool userfaultfd_remove(struct vm_area_struct *vma, |
d811914d | 817 | unsigned long start, unsigned long end) |
05ce7724 PE |
818 | { |
819 | struct mm_struct *mm = vma->vm_mm; | |
820 | struct userfaultfd_ctx *ctx; | |
821 | struct userfaultfd_wait_queue ewq; | |
822 | ||
823 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
d811914d | 824 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE)) |
70ccb92f | 825 | return true; |
05ce7724 PE |
826 | |
827 | userfaultfd_ctx_get(ctx); | |
a759a909 | 828 | atomic_inc(&ctx->mmap_changing); |
d8ed45c5 | 829 | mmap_read_unlock(mm); |
05ce7724 | 830 | |
05ce7724 PE |
831 | msg_init(&ewq.msg); |
832 | ||
d811914d MR |
833 | ewq.msg.event = UFFD_EVENT_REMOVE; |
834 | ewq.msg.arg.remove.start = start; | |
835 | ewq.msg.arg.remove.end = end; | |
05ce7724 PE |
836 | |
837 | userfaultfd_event_wait_completion(ctx, &ewq); | |
838 | ||
70ccb92f | 839 | return false; |
05ce7724 PE |
840 | } |
841 | ||
897ab3e0 MR |
842 | static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps, |
843 | unsigned long start, unsigned long end) | |
844 | { | |
845 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
846 | ||
847 | list_for_each_entry(unmap_ctx, unmaps, list) | |
848 | if (unmap_ctx->ctx == ctx && unmap_ctx->start == start && | |
849 | unmap_ctx->end == end) | |
850 | return true; | |
851 | ||
852 | return false; | |
853 | } | |
854 | ||
65ac1320 | 855 | int userfaultfd_unmap_prep(struct vm_area_struct *vma, unsigned long start, |
69dbe6da | 856 | unsigned long end, struct list_head *unmaps) |
897ab3e0 | 857 | { |
65ac1320 LH |
858 | struct userfaultfd_unmap_ctx *unmap_ctx; |
859 | struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx; | |
897ab3e0 | 860 | |
65ac1320 LH |
861 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) || |
862 | has_unmap_ctx(ctx, unmaps, start, end)) | |
863 | return 0; | |
897ab3e0 | 864 | |
65ac1320 LH |
865 | unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL); |
866 | if (!unmap_ctx) | |
867 | return -ENOMEM; | |
897ab3e0 | 868 | |
65ac1320 LH |
869 | userfaultfd_ctx_get(ctx); |
870 | atomic_inc(&ctx->mmap_changing); | |
871 | unmap_ctx->ctx = ctx; | |
872 | unmap_ctx->start = start; | |
873 | unmap_ctx->end = end; | |
874 | list_add_tail(&unmap_ctx->list, unmaps); | |
897ab3e0 MR |
875 | |
876 | return 0; | |
877 | } | |
878 | ||
879 | void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf) | |
880 | { | |
881 | struct userfaultfd_unmap_ctx *ctx, *n; | |
882 | struct userfaultfd_wait_queue ewq; | |
883 | ||
884 | list_for_each_entry_safe(ctx, n, uf, list) { | |
885 | msg_init(&ewq.msg); | |
886 | ||
887 | ewq.msg.event = UFFD_EVENT_UNMAP; | |
888 | ewq.msg.arg.remove.start = ctx->start; | |
889 | ewq.msg.arg.remove.end = ctx->end; | |
890 | ||
891 | userfaultfd_event_wait_completion(ctx->ctx, &ewq); | |
892 | ||
893 | list_del(&ctx->list); | |
894 | kfree(ctx); | |
895 | } | |
896 | } | |
897 | ||
86039bd3 AA |
898 | static int userfaultfd_release(struct inode *inode, struct file *file) |
899 | { | |
900 | struct userfaultfd_ctx *ctx = file->private_data; | |
901 | struct mm_struct *mm = ctx->mm; | |
902 | struct vm_area_struct *vma, *prev; | |
903 | /* len == 0 means wake all */ | |
904 | struct userfaultfd_wake_range range = { .len = 0, }; | |
905 | unsigned long new_flags; | |
11a9b902 | 906 | VMA_ITERATOR(vmi, mm, 0); |
86039bd3 | 907 | |
6aa7de05 | 908 | WRITE_ONCE(ctx->released, true); |
86039bd3 | 909 | |
d2005e3f ON |
910 | if (!mmget_not_zero(mm)) |
911 | goto wakeup; | |
912 | ||
86039bd3 AA |
913 | /* |
914 | * Flush page faults out of all CPUs. NOTE: all page faults | |
915 | * must be retried without returning VM_FAULT_SIGBUS if | |
916 | * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx | |
c1e8d7c6 | 917 | * changes while handle_userfault released the mmap_lock. So |
86039bd3 | 918 | * it's critical that released is set to true (above), before |
c1e8d7c6 | 919 | * taking the mmap_lock for writing. |
86039bd3 | 920 | */ |
d8ed45c5 | 921 | mmap_write_lock(mm); |
86039bd3 | 922 | prev = NULL; |
11a9b902 | 923 | for_each_vma(vmi, vma) { |
86039bd3 AA |
924 | cond_resched(); |
925 | BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 926 | !!(vma->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
927 | if (vma->vm_userfaultfd_ctx.ctx != ctx) { |
928 | prev = vma; | |
929 | continue; | |
930 | } | |
7677f7fd | 931 | new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS; |
9760ebff | 932 | prev = vma_merge(&vmi, mm, prev, vma->vm_start, vma->vm_end, |
4d45e75a JH |
933 | new_flags, vma->anon_vma, |
934 | vma->vm_file, vma->vm_pgoff, | |
935 | vma_policy(vma), | |
5c26f6ac | 936 | NULL_VM_UFFD_CTX, anon_vma_name(vma)); |
69dbe6da | 937 | if (prev) { |
4d45e75a | 938 | vma = prev; |
69dbe6da | 939 | } else { |
4d45e75a | 940 | prev = vma; |
69dbe6da LH |
941 | } |
942 | ||
51d3d5eb | 943 | userfaultfd_set_vm_flags(vma, new_flags); |
86039bd3 AA |
944 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
945 | } | |
d8ed45c5 | 946 | mmap_write_unlock(mm); |
d2005e3f ON |
947 | mmput(mm); |
948 | wakeup: | |
86039bd3 | 949 | /* |
15b726ef | 950 | * After no new page faults can wait on this fault_*wqh, flush |
86039bd3 | 951 | * the last page faults that may have been already waiting on |
15b726ef | 952 | * the fault_*wqh. |
86039bd3 | 953 | */ |
cbcfa130 | 954 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
ac5be6b4 | 955 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range); |
c430d1e8 | 956 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range); |
cbcfa130 | 957 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 958 | |
5a18b64e MR |
959 | /* Flush pending events that may still wait on event_wqh */ |
960 | wake_up_all(&ctx->event_wqh); | |
961 | ||
a9a08845 | 962 | wake_up_poll(&ctx->fd_wqh, EPOLLHUP); |
86039bd3 AA |
963 | userfaultfd_ctx_put(ctx); |
964 | return 0; | |
965 | } | |
966 | ||
15b726ef | 967 | /* fault_pending_wqh.lock must be hold by the caller */ |
6dcc27fd PE |
968 | static inline struct userfaultfd_wait_queue *find_userfault_in( |
969 | wait_queue_head_t *wqh) | |
86039bd3 | 970 | { |
ac6424b9 | 971 | wait_queue_entry_t *wq; |
15b726ef | 972 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 973 | |
456a7378 | 974 | lockdep_assert_held(&wqh->lock); |
86039bd3 | 975 | |
15b726ef | 976 | uwq = NULL; |
6dcc27fd | 977 | if (!waitqueue_active(wqh)) |
15b726ef AA |
978 | goto out; |
979 | /* walk in reverse to provide FIFO behavior to read userfaults */ | |
2055da97 | 980 | wq = list_last_entry(&wqh->head, typeof(*wq), entry); |
15b726ef AA |
981 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
982 | out: | |
983 | return uwq; | |
86039bd3 | 984 | } |
6dcc27fd PE |
985 | |
986 | static inline struct userfaultfd_wait_queue *find_userfault( | |
987 | struct userfaultfd_ctx *ctx) | |
988 | { | |
989 | return find_userfault_in(&ctx->fault_pending_wqh); | |
990 | } | |
86039bd3 | 991 | |
9cd75c3c PE |
992 | static inline struct userfaultfd_wait_queue *find_userfault_evt( |
993 | struct userfaultfd_ctx *ctx) | |
994 | { | |
995 | return find_userfault_in(&ctx->event_wqh); | |
996 | } | |
997 | ||
076ccb76 | 998 | static __poll_t userfaultfd_poll(struct file *file, poll_table *wait) |
86039bd3 AA |
999 | { |
1000 | struct userfaultfd_ctx *ctx = file->private_data; | |
076ccb76 | 1001 | __poll_t ret; |
86039bd3 AA |
1002 | |
1003 | poll_wait(file, &ctx->fd_wqh, wait); | |
1004 | ||
22e5fe2a | 1005 | if (!userfaultfd_is_initialized(ctx)) |
a9a08845 | 1006 | return EPOLLERR; |
9cd75c3c | 1007 | |
22e5fe2a NA |
1008 | /* |
1009 | * poll() never guarantees that read won't block. | |
1010 | * userfaults can be waken before they're read(). | |
1011 | */ | |
1012 | if (unlikely(!(file->f_flags & O_NONBLOCK))) | |
a9a08845 | 1013 | return EPOLLERR; |
22e5fe2a NA |
1014 | /* |
1015 | * lockless access to see if there are pending faults | |
1016 | * __pollwait last action is the add_wait_queue but | |
1017 | * the spin_unlock would allow the waitqueue_active to | |
1018 | * pass above the actual list_add inside | |
1019 | * add_wait_queue critical section. So use a full | |
1020 | * memory barrier to serialize the list_add write of | |
1021 | * add_wait_queue() with the waitqueue_active read | |
1022 | * below. | |
1023 | */ | |
1024 | ret = 0; | |
1025 | smp_mb(); | |
1026 | if (waitqueue_active(&ctx->fault_pending_wqh)) | |
1027 | ret = EPOLLIN; | |
1028 | else if (waitqueue_active(&ctx->event_wqh)) | |
1029 | ret = EPOLLIN; | |
1030 | ||
1031 | return ret; | |
86039bd3 AA |
1032 | } |
1033 | ||
893e26e6 PE |
1034 | static const struct file_operations userfaultfd_fops; |
1035 | ||
b537900f DC |
1036 | static int resolve_userfault_fork(struct userfaultfd_ctx *new, |
1037 | struct inode *inode, | |
893e26e6 PE |
1038 | struct uffd_msg *msg) |
1039 | { | |
1040 | int fd; | |
893e26e6 | 1041 | |
b537900f | 1042 | fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new, |
abec3d01 | 1043 | O_RDONLY | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode); |
893e26e6 PE |
1044 | if (fd < 0) |
1045 | return fd; | |
1046 | ||
893e26e6 PE |
1047 | msg->arg.reserved.reserved1 = 0; |
1048 | msg->arg.fork.ufd = fd; | |
893e26e6 PE |
1049 | return 0; |
1050 | } | |
1051 | ||
86039bd3 | 1052 | static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait, |
b537900f | 1053 | struct uffd_msg *msg, struct inode *inode) |
86039bd3 AA |
1054 | { |
1055 | ssize_t ret; | |
1056 | DECLARE_WAITQUEUE(wait, current); | |
15b726ef | 1057 | struct userfaultfd_wait_queue *uwq; |
893e26e6 PE |
1058 | /* |
1059 | * Handling fork event requires sleeping operations, so | |
1060 | * we drop the event_wqh lock, then do these ops, then | |
1061 | * lock it back and wake up the waiter. While the lock is | |
1062 | * dropped the ewq may go away so we keep track of it | |
1063 | * carefully. | |
1064 | */ | |
1065 | LIST_HEAD(fork_event); | |
1066 | struct userfaultfd_ctx *fork_nctx = NULL; | |
86039bd3 | 1067 | |
15b726ef | 1068 | /* always take the fd_wqh lock before the fault_pending_wqh lock */ |
ae62c16e | 1069 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1070 | __add_wait_queue(&ctx->fd_wqh, &wait); |
1071 | for (;;) { | |
1072 | set_current_state(TASK_INTERRUPTIBLE); | |
15b726ef AA |
1073 | spin_lock(&ctx->fault_pending_wqh.lock); |
1074 | uwq = find_userfault(ctx); | |
1075 | if (uwq) { | |
2c5b7e1b AA |
1076 | /* |
1077 | * Use a seqcount to repeat the lockless check | |
1078 | * in wake_userfault() to avoid missing | |
1079 | * wakeups because during the refile both | |
1080 | * waitqueue could become empty if this is the | |
1081 | * only userfault. | |
1082 | */ | |
1083 | write_seqcount_begin(&ctx->refile_seq); | |
1084 | ||
86039bd3 | 1085 | /* |
15b726ef AA |
1086 | * The fault_pending_wqh.lock prevents the uwq |
1087 | * to disappear from under us. | |
1088 | * | |
1089 | * Refile this userfault from | |
1090 | * fault_pending_wqh to fault_wqh, it's not | |
1091 | * pending anymore after we read it. | |
1092 | * | |
1093 | * Use list_del() by hand (as | |
1094 | * userfaultfd_wake_function also uses | |
1095 | * list_del_init() by hand) to be sure nobody | |
1096 | * changes __remove_wait_queue() to use | |
1097 | * list_del_init() in turn breaking the | |
1098 | * !list_empty_careful() check in | |
2055da97 | 1099 | * handle_userfault(). The uwq->wq.head list |
15b726ef AA |
1100 | * must never be empty at any time during the |
1101 | * refile, or the waitqueue could disappear | |
1102 | * from under us. The "wait_queue_head_t" | |
1103 | * parameter of __remove_wait_queue() is unused | |
1104 | * anyway. | |
86039bd3 | 1105 | */ |
2055da97 | 1106 | list_del(&uwq->wq.entry); |
c430d1e8 | 1107 | add_wait_queue(&ctx->fault_wqh, &uwq->wq); |
15b726ef | 1108 | |
2c5b7e1b AA |
1109 | write_seqcount_end(&ctx->refile_seq); |
1110 | ||
a9b85f94 AA |
1111 | /* careful to always initialize msg if ret == 0 */ |
1112 | *msg = uwq->msg; | |
15b726ef | 1113 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1114 | ret = 0; |
1115 | break; | |
1116 | } | |
15b726ef | 1117 | spin_unlock(&ctx->fault_pending_wqh.lock); |
9cd75c3c PE |
1118 | |
1119 | spin_lock(&ctx->event_wqh.lock); | |
1120 | uwq = find_userfault_evt(ctx); | |
1121 | if (uwq) { | |
1122 | *msg = uwq->msg; | |
1123 | ||
893e26e6 PE |
1124 | if (uwq->msg.event == UFFD_EVENT_FORK) { |
1125 | fork_nctx = (struct userfaultfd_ctx *) | |
1126 | (unsigned long) | |
1127 | uwq->msg.arg.reserved.reserved1; | |
2055da97 | 1128 | list_move(&uwq->wq.entry, &fork_event); |
384632e6 AA |
1129 | /* |
1130 | * fork_nctx can be freed as soon as | |
1131 | * we drop the lock, unless we take a | |
1132 | * reference on it. | |
1133 | */ | |
1134 | userfaultfd_ctx_get(fork_nctx); | |
893e26e6 PE |
1135 | spin_unlock(&ctx->event_wqh.lock); |
1136 | ret = 0; | |
1137 | break; | |
1138 | } | |
1139 | ||
9cd75c3c PE |
1140 | userfaultfd_event_complete(ctx, uwq); |
1141 | spin_unlock(&ctx->event_wqh.lock); | |
1142 | ret = 0; | |
1143 | break; | |
1144 | } | |
1145 | spin_unlock(&ctx->event_wqh.lock); | |
1146 | ||
86039bd3 AA |
1147 | if (signal_pending(current)) { |
1148 | ret = -ERESTARTSYS; | |
1149 | break; | |
1150 | } | |
1151 | if (no_wait) { | |
1152 | ret = -EAGAIN; | |
1153 | break; | |
1154 | } | |
ae62c16e | 1155 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1156 | schedule(); |
ae62c16e | 1157 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1158 | } |
1159 | __remove_wait_queue(&ctx->fd_wqh, &wait); | |
1160 | __set_current_state(TASK_RUNNING); | |
ae62c16e | 1161 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1162 | |
893e26e6 | 1163 | if (!ret && msg->event == UFFD_EVENT_FORK) { |
b537900f | 1164 | ret = resolve_userfault_fork(fork_nctx, inode, msg); |
cbcfa130 | 1165 | spin_lock_irq(&ctx->event_wqh.lock); |
384632e6 AA |
1166 | if (!list_empty(&fork_event)) { |
1167 | /* | |
1168 | * The fork thread didn't abort, so we can | |
1169 | * drop the temporary refcount. | |
1170 | */ | |
1171 | userfaultfd_ctx_put(fork_nctx); | |
1172 | ||
1173 | uwq = list_first_entry(&fork_event, | |
1174 | typeof(*uwq), | |
1175 | wq.entry); | |
1176 | /* | |
1177 | * If fork_event list wasn't empty and in turn | |
1178 | * the event wasn't already released by fork | |
1179 | * (the event is allocated on fork kernel | |
1180 | * stack), put the event back to its place in | |
1181 | * the event_wq. fork_event head will be freed | |
1182 | * as soon as we return so the event cannot | |
1183 | * stay queued there no matter the current | |
1184 | * "ret" value. | |
1185 | */ | |
1186 | list_del(&uwq->wq.entry); | |
1187 | __add_wait_queue(&ctx->event_wqh, &uwq->wq); | |
893e26e6 | 1188 | |
384632e6 AA |
1189 | /* |
1190 | * Leave the event in the waitqueue and report | |
1191 | * error to userland if we failed to resolve | |
1192 | * the userfault fork. | |
1193 | */ | |
1194 | if (likely(!ret)) | |
893e26e6 | 1195 | userfaultfd_event_complete(ctx, uwq); |
384632e6 AA |
1196 | } else { |
1197 | /* | |
1198 | * Here the fork thread aborted and the | |
1199 | * refcount from the fork thread on fork_nctx | |
1200 | * has already been released. We still hold | |
1201 | * the reference we took before releasing the | |
1202 | * lock above. If resolve_userfault_fork | |
1203 | * failed we've to drop it because the | |
1204 | * fork_nctx has to be freed in such case. If | |
1205 | * it succeeded we'll hold it because the new | |
1206 | * uffd references it. | |
1207 | */ | |
1208 | if (ret) | |
1209 | userfaultfd_ctx_put(fork_nctx); | |
893e26e6 | 1210 | } |
cbcfa130 | 1211 | spin_unlock_irq(&ctx->event_wqh.lock); |
893e26e6 PE |
1212 | } |
1213 | ||
86039bd3 AA |
1214 | return ret; |
1215 | } | |
1216 | ||
1217 | static ssize_t userfaultfd_read(struct file *file, char __user *buf, | |
1218 | size_t count, loff_t *ppos) | |
1219 | { | |
1220 | struct userfaultfd_ctx *ctx = file->private_data; | |
1221 | ssize_t _ret, ret = 0; | |
a9b85f94 | 1222 | struct uffd_msg msg; |
86039bd3 | 1223 | int no_wait = file->f_flags & O_NONBLOCK; |
b537900f | 1224 | struct inode *inode = file_inode(file); |
86039bd3 | 1225 | |
22e5fe2a | 1226 | if (!userfaultfd_is_initialized(ctx)) |
86039bd3 | 1227 | return -EINVAL; |
86039bd3 AA |
1228 | |
1229 | for (;;) { | |
a9b85f94 | 1230 | if (count < sizeof(msg)) |
86039bd3 | 1231 | return ret ? ret : -EINVAL; |
b537900f | 1232 | _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode); |
86039bd3 AA |
1233 | if (_ret < 0) |
1234 | return ret ? ret : _ret; | |
a9b85f94 | 1235 | if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg))) |
86039bd3 | 1236 | return ret ? ret : -EFAULT; |
a9b85f94 AA |
1237 | ret += sizeof(msg); |
1238 | buf += sizeof(msg); | |
1239 | count -= sizeof(msg); | |
86039bd3 AA |
1240 | /* |
1241 | * Allow to read more than one fault at time but only | |
1242 | * block if waiting for the very first one. | |
1243 | */ | |
1244 | no_wait = O_NONBLOCK; | |
1245 | } | |
1246 | } | |
1247 | ||
1248 | static void __wake_userfault(struct userfaultfd_ctx *ctx, | |
1249 | struct userfaultfd_wake_range *range) | |
1250 | { | |
cbcfa130 | 1251 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 1252 | /* wake all in the range and autoremove */ |
15b726ef | 1253 | if (waitqueue_active(&ctx->fault_pending_wqh)) |
ac5be6b4 | 1254 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, |
15b726ef AA |
1255 | range); |
1256 | if (waitqueue_active(&ctx->fault_wqh)) | |
c430d1e8 | 1257 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range); |
cbcfa130 | 1258 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1259 | } |
1260 | ||
1261 | static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx, | |
1262 | struct userfaultfd_wake_range *range) | |
1263 | { | |
2c5b7e1b AA |
1264 | unsigned seq; |
1265 | bool need_wakeup; | |
1266 | ||
86039bd3 AA |
1267 | /* |
1268 | * To be sure waitqueue_active() is not reordered by the CPU | |
1269 | * before the pagetable update, use an explicit SMP memory | |
3e4e28c5 | 1270 | * barrier here. PT lock release or mmap_read_unlock(mm) still |
86039bd3 AA |
1271 | * have release semantics that can allow the |
1272 | * waitqueue_active() to be reordered before the pte update. | |
1273 | */ | |
1274 | smp_mb(); | |
1275 | ||
1276 | /* | |
1277 | * Use waitqueue_active because it's very frequent to | |
1278 | * change the address space atomically even if there are no | |
1279 | * userfaults yet. So we take the spinlock only when we're | |
1280 | * sure we've userfaults to wake. | |
1281 | */ | |
2c5b7e1b AA |
1282 | do { |
1283 | seq = read_seqcount_begin(&ctx->refile_seq); | |
1284 | need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) || | |
1285 | waitqueue_active(&ctx->fault_wqh); | |
1286 | cond_resched(); | |
1287 | } while (read_seqcount_retry(&ctx->refile_seq, seq)); | |
1288 | if (need_wakeup) | |
86039bd3 AA |
1289 | __wake_userfault(ctx, range); |
1290 | } | |
1291 | ||
1292 | static __always_inline int validate_range(struct mm_struct *mm, | |
e71e2ace | 1293 | __u64 start, __u64 len) |
86039bd3 AA |
1294 | { |
1295 | __u64 task_size = mm->task_size; | |
1296 | ||
e71e2ace | 1297 | if (start & ~PAGE_MASK) |
86039bd3 AA |
1298 | return -EINVAL; |
1299 | if (len & ~PAGE_MASK) | |
1300 | return -EINVAL; | |
1301 | if (!len) | |
1302 | return -EINVAL; | |
e71e2ace | 1303 | if (start < mmap_min_addr) |
86039bd3 | 1304 | return -EINVAL; |
e71e2ace | 1305 | if (start >= task_size) |
86039bd3 | 1306 | return -EINVAL; |
e71e2ace | 1307 | if (len > task_size - start) |
86039bd3 AA |
1308 | return -EINVAL; |
1309 | return 0; | |
1310 | } | |
1311 | ||
1312 | static int userfaultfd_register(struct userfaultfd_ctx *ctx, | |
1313 | unsigned long arg) | |
1314 | { | |
1315 | struct mm_struct *mm = ctx->mm; | |
1316 | struct vm_area_struct *vma, *prev, *cur; | |
1317 | int ret; | |
1318 | struct uffdio_register uffdio_register; | |
1319 | struct uffdio_register __user *user_uffdio_register; | |
1320 | unsigned long vm_flags, new_flags; | |
1321 | bool found; | |
ce53e8e6 | 1322 | bool basic_ioctls; |
86039bd3 | 1323 | unsigned long start, end, vma_end; |
11a9b902 | 1324 | struct vma_iterator vmi; |
5543d3c4 | 1325 | pgoff_t pgoff; |
86039bd3 AA |
1326 | |
1327 | user_uffdio_register = (struct uffdio_register __user *) arg; | |
1328 | ||
1329 | ret = -EFAULT; | |
1330 | if (copy_from_user(&uffdio_register, user_uffdio_register, | |
1331 | sizeof(uffdio_register)-sizeof(__u64))) | |
1332 | goto out; | |
1333 | ||
1334 | ret = -EINVAL; | |
1335 | if (!uffdio_register.mode) | |
1336 | goto out; | |
7677f7fd | 1337 | if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES) |
86039bd3 AA |
1338 | goto out; |
1339 | vm_flags = 0; | |
1340 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING) | |
1341 | vm_flags |= VM_UFFD_MISSING; | |
00b151f2 PX |
1342 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) { |
1343 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP | |
1344 | goto out; | |
1345 | #endif | |
86039bd3 | 1346 | vm_flags |= VM_UFFD_WP; |
00b151f2 | 1347 | } |
7677f7fd AR |
1348 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) { |
1349 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR | |
1350 | goto out; | |
1351 | #endif | |
1352 | vm_flags |= VM_UFFD_MINOR; | |
1353 | } | |
86039bd3 | 1354 | |
e71e2ace | 1355 | ret = validate_range(mm, uffdio_register.range.start, |
86039bd3 AA |
1356 | uffdio_register.range.len); |
1357 | if (ret) | |
1358 | goto out; | |
1359 | ||
1360 | start = uffdio_register.range.start; | |
1361 | end = start + uffdio_register.range.len; | |
1362 | ||
d2005e3f ON |
1363 | ret = -ENOMEM; |
1364 | if (!mmget_not_zero(mm)) | |
1365 | goto out; | |
1366 | ||
11a9b902 | 1367 | ret = -EINVAL; |
d8ed45c5 | 1368 | mmap_write_lock(mm); |
11a9b902 LH |
1369 | vma_iter_init(&vmi, mm, start); |
1370 | vma = vma_find(&vmi, end); | |
86039bd3 AA |
1371 | if (!vma) |
1372 | goto out_unlock; | |
1373 | ||
cab350af MK |
1374 | /* |
1375 | * If the first vma contains huge pages, make sure start address | |
1376 | * is aligned to huge page size. | |
1377 | */ | |
1378 | if (is_vm_hugetlb_page(vma)) { | |
1379 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1380 | ||
1381 | if (start & (vma_hpagesize - 1)) | |
1382 | goto out_unlock; | |
1383 | } | |
1384 | ||
86039bd3 AA |
1385 | /* |
1386 | * Search for not compatible vmas. | |
86039bd3 AA |
1387 | */ |
1388 | found = false; | |
ce53e8e6 | 1389 | basic_ioctls = false; |
11a9b902 LH |
1390 | cur = vma; |
1391 | do { | |
86039bd3 AA |
1392 | cond_resched(); |
1393 | ||
1394 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 1395 | !!(cur->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
1396 | |
1397 | /* check not compatible vmas */ | |
1398 | ret = -EINVAL; | |
63b2d417 | 1399 | if (!vma_can_userfault(cur, vm_flags)) |
86039bd3 | 1400 | goto out_unlock; |
29ec9066 AA |
1401 | |
1402 | /* | |
1403 | * UFFDIO_COPY will fill file holes even without | |
1404 | * PROT_WRITE. This check enforces that if this is a | |
1405 | * MAP_SHARED, the process has write permission to the backing | |
1406 | * file. If VM_MAYWRITE is set it also enforces that on a | |
1407 | * MAP_SHARED vma: there is no F_WRITE_SEAL and no further | |
1408 | * F_WRITE_SEAL can be taken until the vma is destroyed. | |
1409 | */ | |
1410 | ret = -EPERM; | |
1411 | if (unlikely(!(cur->vm_flags & VM_MAYWRITE))) | |
1412 | goto out_unlock; | |
1413 | ||
cab350af MK |
1414 | /* |
1415 | * If this vma contains ending address, and huge pages | |
1416 | * check alignment. | |
1417 | */ | |
1418 | if (is_vm_hugetlb_page(cur) && end <= cur->vm_end && | |
1419 | end > cur->vm_start) { | |
1420 | unsigned long vma_hpagesize = vma_kernel_pagesize(cur); | |
1421 | ||
1422 | ret = -EINVAL; | |
1423 | ||
1424 | if (end & (vma_hpagesize - 1)) | |
1425 | goto out_unlock; | |
1426 | } | |
63b2d417 AA |
1427 | if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE)) |
1428 | goto out_unlock; | |
86039bd3 AA |
1429 | |
1430 | /* | |
1431 | * Check that this vma isn't already owned by a | |
1432 | * different userfaultfd. We can't allow more than one | |
1433 | * userfaultfd to own a single vma simultaneously or we | |
1434 | * wouldn't know which one to deliver the userfaults to. | |
1435 | */ | |
1436 | ret = -EBUSY; | |
1437 | if (cur->vm_userfaultfd_ctx.ctx && | |
1438 | cur->vm_userfaultfd_ctx.ctx != ctx) | |
1439 | goto out_unlock; | |
1440 | ||
cab350af MK |
1441 | /* |
1442 | * Note vmas containing huge pages | |
1443 | */ | |
ce53e8e6 MR |
1444 | if (is_vm_hugetlb_page(cur)) |
1445 | basic_ioctls = true; | |
cab350af | 1446 | |
86039bd3 | 1447 | found = true; |
11a9b902 | 1448 | } for_each_vma_range(vmi, cur, end); |
86039bd3 AA |
1449 | BUG_ON(!found); |
1450 | ||
11a9b902 LH |
1451 | vma_iter_set(&vmi, start); |
1452 | prev = vma_prev(&vmi); | |
270aa010 PX |
1453 | if (vma->vm_start < start) |
1454 | prev = vma; | |
86039bd3 AA |
1455 | |
1456 | ret = 0; | |
11a9b902 | 1457 | for_each_vma_range(vmi, vma, end) { |
86039bd3 AA |
1458 | cond_resched(); |
1459 | ||
63b2d417 | 1460 | BUG_ON(!vma_can_userfault(vma, vm_flags)); |
86039bd3 AA |
1461 | BUG_ON(vma->vm_userfaultfd_ctx.ctx && |
1462 | vma->vm_userfaultfd_ctx.ctx != ctx); | |
29ec9066 | 1463 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
86039bd3 AA |
1464 | |
1465 | /* | |
1466 | * Nothing to do: this vma is already registered into this | |
1467 | * userfaultfd and with the right tracking mode too. | |
1468 | */ | |
1469 | if (vma->vm_userfaultfd_ctx.ctx == ctx && | |
1470 | (vma->vm_flags & vm_flags) == vm_flags) | |
1471 | goto skip; | |
1472 | ||
1473 | if (vma->vm_start > start) | |
1474 | start = vma->vm_start; | |
1475 | vma_end = min(end, vma->vm_end); | |
1476 | ||
7677f7fd | 1477 | new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags; |
5543d3c4 | 1478 | pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); |
9760ebff | 1479 | prev = vma_merge(&vmi, mm, prev, start, vma_end, new_flags, |
5543d3c4 | 1480 | vma->anon_vma, vma->vm_file, pgoff, |
86039bd3 | 1481 | vma_policy(vma), |
9a10064f | 1482 | ((struct vm_userfaultfd_ctx){ ctx }), |
5c26f6ac | 1483 | anon_vma_name(vma)); |
86039bd3 | 1484 | if (prev) { |
69dbe6da | 1485 | /* vma_merge() invalidated the mas */ |
86039bd3 AA |
1486 | vma = prev; |
1487 | goto next; | |
1488 | } | |
1489 | if (vma->vm_start < start) { | |
9760ebff | 1490 | ret = split_vma(&vmi, vma, start, 1); |
86039bd3 AA |
1491 | if (ret) |
1492 | break; | |
1493 | } | |
1494 | if (vma->vm_end > end) { | |
9760ebff | 1495 | ret = split_vma(&vmi, vma, end, 0); |
86039bd3 AA |
1496 | if (ret) |
1497 | break; | |
1498 | } | |
1499 | next: | |
1500 | /* | |
1501 | * In the vma_merge() successful mprotect-like case 8: | |
1502 | * the next vma was merged into the current one and | |
1503 | * the current one has not been updated yet. | |
1504 | */ | |
51d3d5eb | 1505 | userfaultfd_set_vm_flags(vma, new_flags); |
86039bd3 AA |
1506 | vma->vm_userfaultfd_ctx.ctx = ctx; |
1507 | ||
6dfeaff9 PX |
1508 | if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma)) |
1509 | hugetlb_unshare_all_pmds(vma); | |
1510 | ||
86039bd3 AA |
1511 | skip: |
1512 | prev = vma; | |
1513 | start = vma->vm_end; | |
11a9b902 LH |
1514 | } |
1515 | ||
86039bd3 | 1516 | out_unlock: |
d8ed45c5 | 1517 | mmap_write_unlock(mm); |
d2005e3f | 1518 | mmput(mm); |
86039bd3 | 1519 | if (!ret) { |
14819305 PX |
1520 | __u64 ioctls_out; |
1521 | ||
1522 | ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC : | |
1523 | UFFD_API_RANGE_IOCTLS; | |
1524 | ||
1525 | /* | |
1526 | * Declare the WP ioctl only if the WP mode is | |
1527 | * specified and all checks passed with the range | |
1528 | */ | |
1529 | if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP)) | |
1530 | ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT); | |
1531 | ||
f6191471 AR |
1532 | /* CONTINUE ioctl is only supported for MINOR ranges. */ |
1533 | if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR)) | |
1534 | ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE); | |
1535 | ||
86039bd3 AA |
1536 | /* |
1537 | * Now that we scanned all vmas we can already tell | |
1538 | * userland which ioctls methods are guaranteed to | |
1539 | * succeed on this range. | |
1540 | */ | |
14819305 | 1541 | if (put_user(ioctls_out, &user_uffdio_register->ioctls)) |
86039bd3 AA |
1542 | ret = -EFAULT; |
1543 | } | |
1544 | out: | |
1545 | return ret; | |
1546 | } | |
1547 | ||
1548 | static int userfaultfd_unregister(struct userfaultfd_ctx *ctx, | |
1549 | unsigned long arg) | |
1550 | { | |
1551 | struct mm_struct *mm = ctx->mm; | |
1552 | struct vm_area_struct *vma, *prev, *cur; | |
1553 | int ret; | |
1554 | struct uffdio_range uffdio_unregister; | |
1555 | unsigned long new_flags; | |
1556 | bool found; | |
1557 | unsigned long start, end, vma_end; | |
1558 | const void __user *buf = (void __user *)arg; | |
11a9b902 | 1559 | struct vma_iterator vmi; |
5543d3c4 | 1560 | pgoff_t pgoff; |
86039bd3 AA |
1561 | |
1562 | ret = -EFAULT; | |
1563 | if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister))) | |
1564 | goto out; | |
1565 | ||
e71e2ace | 1566 | ret = validate_range(mm, uffdio_unregister.start, |
86039bd3 AA |
1567 | uffdio_unregister.len); |
1568 | if (ret) | |
1569 | goto out; | |
1570 | ||
1571 | start = uffdio_unregister.start; | |
1572 | end = start + uffdio_unregister.len; | |
1573 | ||
d2005e3f ON |
1574 | ret = -ENOMEM; |
1575 | if (!mmget_not_zero(mm)) | |
1576 | goto out; | |
1577 | ||
d8ed45c5 | 1578 | mmap_write_lock(mm); |
86039bd3 | 1579 | ret = -EINVAL; |
11a9b902 LH |
1580 | vma_iter_init(&vmi, mm, start); |
1581 | vma = vma_find(&vmi, end); | |
1582 | if (!vma) | |
86039bd3 AA |
1583 | goto out_unlock; |
1584 | ||
cab350af MK |
1585 | /* |
1586 | * If the first vma contains huge pages, make sure start address | |
1587 | * is aligned to huge page size. | |
1588 | */ | |
1589 | if (is_vm_hugetlb_page(vma)) { | |
1590 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1591 | ||
1592 | if (start & (vma_hpagesize - 1)) | |
1593 | goto out_unlock; | |
1594 | } | |
1595 | ||
86039bd3 AA |
1596 | /* |
1597 | * Search for not compatible vmas. | |
86039bd3 AA |
1598 | */ |
1599 | found = false; | |
11a9b902 LH |
1600 | cur = vma; |
1601 | do { | |
86039bd3 AA |
1602 | cond_resched(); |
1603 | ||
1604 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
7677f7fd | 1605 | !!(cur->vm_flags & __VM_UFFD_FLAGS)); |
86039bd3 AA |
1606 | |
1607 | /* | |
1608 | * Check not compatible vmas, not strictly required | |
1609 | * here as not compatible vmas cannot have an | |
1610 | * userfaultfd_ctx registered on them, but this | |
1611 | * provides for more strict behavior to notice | |
1612 | * unregistration errors. | |
1613 | */ | |
63b2d417 | 1614 | if (!vma_can_userfault(cur, cur->vm_flags)) |
86039bd3 AA |
1615 | goto out_unlock; |
1616 | ||
1617 | found = true; | |
11a9b902 | 1618 | } for_each_vma_range(vmi, cur, end); |
86039bd3 AA |
1619 | BUG_ON(!found); |
1620 | ||
11a9b902 LH |
1621 | vma_iter_set(&vmi, start); |
1622 | prev = vma_prev(&vmi); | |
270aa010 PX |
1623 | if (vma->vm_start < start) |
1624 | prev = vma; | |
1625 | ||
86039bd3 | 1626 | ret = 0; |
11a9b902 | 1627 | for_each_vma_range(vmi, vma, end) { |
86039bd3 AA |
1628 | cond_resched(); |
1629 | ||
63b2d417 | 1630 | BUG_ON(!vma_can_userfault(vma, vma->vm_flags)); |
86039bd3 AA |
1631 | |
1632 | /* | |
1633 | * Nothing to do: this vma is already registered into this | |
1634 | * userfaultfd and with the right tracking mode too. | |
1635 | */ | |
1636 | if (!vma->vm_userfaultfd_ctx.ctx) | |
1637 | goto skip; | |
1638 | ||
01e881f5 AA |
1639 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
1640 | ||
86039bd3 AA |
1641 | if (vma->vm_start > start) |
1642 | start = vma->vm_start; | |
1643 | vma_end = min(end, vma->vm_end); | |
1644 | ||
09fa5296 AA |
1645 | if (userfaultfd_missing(vma)) { |
1646 | /* | |
1647 | * Wake any concurrent pending userfault while | |
1648 | * we unregister, so they will not hang | |
1649 | * permanently and it avoids userland to call | |
1650 | * UFFDIO_WAKE explicitly. | |
1651 | */ | |
1652 | struct userfaultfd_wake_range range; | |
1653 | range.start = start; | |
1654 | range.len = vma_end - start; | |
1655 | wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range); | |
1656 | } | |
1657 | ||
f369b07c PX |
1658 | /* Reset ptes for the whole vma range if wr-protected */ |
1659 | if (userfaultfd_wp(vma)) | |
61c50040 | 1660 | uffd_wp_range(vma, start, vma_end - start, false); |
f369b07c | 1661 | |
7677f7fd | 1662 | new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS; |
5543d3c4 | 1663 | pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); |
9760ebff | 1664 | prev = vma_merge(&vmi, mm, prev, start, vma_end, new_flags, |
5543d3c4 | 1665 | vma->anon_vma, vma->vm_file, pgoff, |
86039bd3 | 1666 | vma_policy(vma), |
5c26f6ac | 1667 | NULL_VM_UFFD_CTX, anon_vma_name(vma)); |
86039bd3 AA |
1668 | if (prev) { |
1669 | vma = prev; | |
1670 | goto next; | |
1671 | } | |
1672 | if (vma->vm_start < start) { | |
9760ebff | 1673 | ret = split_vma(&vmi, vma, start, 1); |
86039bd3 AA |
1674 | if (ret) |
1675 | break; | |
1676 | } | |
1677 | if (vma->vm_end > end) { | |
9760ebff | 1678 | ret = split_vma(&vmi, vma, end, 0); |
86039bd3 AA |
1679 | if (ret) |
1680 | break; | |
1681 | } | |
1682 | next: | |
1683 | /* | |
1684 | * In the vma_merge() successful mprotect-like case 8: | |
1685 | * the next vma was merged into the current one and | |
1686 | * the current one has not been updated yet. | |
1687 | */ | |
51d3d5eb | 1688 | userfaultfd_set_vm_flags(vma, new_flags); |
86039bd3 AA |
1689 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
1690 | ||
1691 | skip: | |
1692 | prev = vma; | |
1693 | start = vma->vm_end; | |
11a9b902 LH |
1694 | } |
1695 | ||
86039bd3 | 1696 | out_unlock: |
d8ed45c5 | 1697 | mmap_write_unlock(mm); |
d2005e3f | 1698 | mmput(mm); |
86039bd3 AA |
1699 | out: |
1700 | return ret; | |
1701 | } | |
1702 | ||
1703 | /* | |
ba85c702 AA |
1704 | * userfaultfd_wake may be used in combination with the |
1705 | * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches. | |
86039bd3 AA |
1706 | */ |
1707 | static int userfaultfd_wake(struct userfaultfd_ctx *ctx, | |
1708 | unsigned long arg) | |
1709 | { | |
1710 | int ret; | |
1711 | struct uffdio_range uffdio_wake; | |
1712 | struct userfaultfd_wake_range range; | |
1713 | const void __user *buf = (void __user *)arg; | |
1714 | ||
1715 | ret = -EFAULT; | |
1716 | if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake))) | |
1717 | goto out; | |
1718 | ||
e71e2ace | 1719 | ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len); |
86039bd3 AA |
1720 | if (ret) |
1721 | goto out; | |
1722 | ||
1723 | range.start = uffdio_wake.start; | |
1724 | range.len = uffdio_wake.len; | |
1725 | ||
1726 | /* | |
1727 | * len == 0 means wake all and we don't want to wake all here, | |
1728 | * so check it again to be sure. | |
1729 | */ | |
1730 | VM_BUG_ON(!range.len); | |
1731 | ||
1732 | wake_userfault(ctx, &range); | |
1733 | ret = 0; | |
1734 | ||
1735 | out: | |
1736 | return ret; | |
1737 | } | |
1738 | ||
ad465cae AA |
1739 | static int userfaultfd_copy(struct userfaultfd_ctx *ctx, |
1740 | unsigned long arg) | |
1741 | { | |
1742 | __s64 ret; | |
1743 | struct uffdio_copy uffdio_copy; | |
1744 | struct uffdio_copy __user *user_uffdio_copy; | |
1745 | struct userfaultfd_wake_range range; | |
d9712937 | 1746 | uffd_flags_t flags = 0; |
ad465cae AA |
1747 | |
1748 | user_uffdio_copy = (struct uffdio_copy __user *) arg; | |
1749 | ||
df2cc96e | 1750 | ret = -EAGAIN; |
a759a909 | 1751 | if (atomic_read(&ctx->mmap_changing)) |
df2cc96e MR |
1752 | goto out; |
1753 | ||
ad465cae AA |
1754 | ret = -EFAULT; |
1755 | if (copy_from_user(&uffdio_copy, user_uffdio_copy, | |
1756 | /* don't copy "copy" last field */ | |
1757 | sizeof(uffdio_copy)-sizeof(__s64))) | |
1758 | goto out; | |
1759 | ||
e71e2ace | 1760 | ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len); |
ad465cae AA |
1761 | if (ret) |
1762 | goto out; | |
1763 | /* | |
1764 | * double check for wraparound just in case. copy_from_user() | |
1765 | * will later check uffdio_copy.src + uffdio_copy.len to fit | |
1766 | * in the userland range. | |
1767 | */ | |
1768 | ret = -EINVAL; | |
1769 | if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src) | |
1770 | goto out; | |
72981e0e | 1771 | if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP)) |
ad465cae | 1772 | goto out; |
d9712937 AR |
1773 | if (uffdio_copy.mode & UFFDIO_COPY_MODE_WP) |
1774 | flags |= MFILL_ATOMIC_WP; | |
d2005e3f | 1775 | if (mmget_not_zero(ctx->mm)) { |
a734991c AR |
1776 | ret = mfill_atomic_copy(ctx->mm, uffdio_copy.dst, uffdio_copy.src, |
1777 | uffdio_copy.len, &ctx->mmap_changing, | |
d9712937 | 1778 | flags); |
d2005e3f | 1779 | mmput(ctx->mm); |
96333187 | 1780 | } else { |
e86b298b | 1781 | return -ESRCH; |
d2005e3f | 1782 | } |
ad465cae AA |
1783 | if (unlikely(put_user(ret, &user_uffdio_copy->copy))) |
1784 | return -EFAULT; | |
1785 | if (ret < 0) | |
1786 | goto out; | |
1787 | BUG_ON(!ret); | |
1788 | /* len == 0 would wake all */ | |
1789 | range.len = ret; | |
1790 | if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) { | |
1791 | range.start = uffdio_copy.dst; | |
1792 | wake_userfault(ctx, &range); | |
1793 | } | |
1794 | ret = range.len == uffdio_copy.len ? 0 : -EAGAIN; | |
1795 | out: | |
1796 | return ret; | |
1797 | } | |
1798 | ||
1799 | static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx, | |
1800 | unsigned long arg) | |
1801 | { | |
1802 | __s64 ret; | |
1803 | struct uffdio_zeropage uffdio_zeropage; | |
1804 | struct uffdio_zeropage __user *user_uffdio_zeropage; | |
1805 | struct userfaultfd_wake_range range; | |
1806 | ||
1807 | user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg; | |
1808 | ||
df2cc96e | 1809 | ret = -EAGAIN; |
a759a909 | 1810 | if (atomic_read(&ctx->mmap_changing)) |
df2cc96e MR |
1811 | goto out; |
1812 | ||
ad465cae AA |
1813 | ret = -EFAULT; |
1814 | if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage, | |
1815 | /* don't copy "zeropage" last field */ | |
1816 | sizeof(uffdio_zeropage)-sizeof(__s64))) | |
1817 | goto out; | |
1818 | ||
e71e2ace | 1819 | ret = validate_range(ctx->mm, uffdio_zeropage.range.start, |
ad465cae AA |
1820 | uffdio_zeropage.range.len); |
1821 | if (ret) | |
1822 | goto out; | |
1823 | ret = -EINVAL; | |
1824 | if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE) | |
1825 | goto out; | |
1826 | ||
d2005e3f | 1827 | if (mmget_not_zero(ctx->mm)) { |
a734991c AR |
1828 | ret = mfill_atomic_zeropage(ctx->mm, uffdio_zeropage.range.start, |
1829 | uffdio_zeropage.range.len, | |
1830 | &ctx->mmap_changing); | |
d2005e3f | 1831 | mmput(ctx->mm); |
9d95aa4b | 1832 | } else { |
e86b298b | 1833 | return -ESRCH; |
d2005e3f | 1834 | } |
ad465cae AA |
1835 | if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage))) |
1836 | return -EFAULT; | |
1837 | if (ret < 0) | |
1838 | goto out; | |
1839 | /* len == 0 would wake all */ | |
1840 | BUG_ON(!ret); | |
1841 | range.len = ret; | |
1842 | if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) { | |
1843 | range.start = uffdio_zeropage.range.start; | |
1844 | wake_userfault(ctx, &range); | |
1845 | } | |
1846 | ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN; | |
1847 | out: | |
1848 | return ret; | |
1849 | } | |
1850 | ||
63b2d417 AA |
1851 | static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx, |
1852 | unsigned long arg) | |
1853 | { | |
1854 | int ret; | |
1855 | struct uffdio_writeprotect uffdio_wp; | |
1856 | struct uffdio_writeprotect __user *user_uffdio_wp; | |
1857 | struct userfaultfd_wake_range range; | |
23080e27 | 1858 | bool mode_wp, mode_dontwake; |
63b2d417 | 1859 | |
a759a909 | 1860 | if (atomic_read(&ctx->mmap_changing)) |
63b2d417 AA |
1861 | return -EAGAIN; |
1862 | ||
1863 | user_uffdio_wp = (struct uffdio_writeprotect __user *) arg; | |
1864 | ||
1865 | if (copy_from_user(&uffdio_wp, user_uffdio_wp, | |
1866 | sizeof(struct uffdio_writeprotect))) | |
1867 | return -EFAULT; | |
1868 | ||
e71e2ace | 1869 | ret = validate_range(ctx->mm, uffdio_wp.range.start, |
63b2d417 AA |
1870 | uffdio_wp.range.len); |
1871 | if (ret) | |
1872 | return ret; | |
1873 | ||
1874 | if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE | | |
1875 | UFFDIO_WRITEPROTECT_MODE_WP)) | |
1876 | return -EINVAL; | |
23080e27 PX |
1877 | |
1878 | mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP; | |
1879 | mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE; | |
1880 | ||
1881 | if (mode_wp && mode_dontwake) | |
63b2d417 AA |
1882 | return -EINVAL; |
1883 | ||
cb185d5f NA |
1884 | if (mmget_not_zero(ctx->mm)) { |
1885 | ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start, | |
1886 | uffdio_wp.range.len, mode_wp, | |
1887 | &ctx->mmap_changing); | |
1888 | mmput(ctx->mm); | |
1889 | } else { | |
1890 | return -ESRCH; | |
1891 | } | |
1892 | ||
63b2d417 AA |
1893 | if (ret) |
1894 | return ret; | |
1895 | ||
23080e27 | 1896 | if (!mode_wp && !mode_dontwake) { |
63b2d417 AA |
1897 | range.start = uffdio_wp.range.start; |
1898 | range.len = uffdio_wp.range.len; | |
1899 | wake_userfault(ctx, &range); | |
1900 | } | |
1901 | return ret; | |
1902 | } | |
1903 | ||
f6191471 AR |
1904 | static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg) |
1905 | { | |
1906 | __s64 ret; | |
1907 | struct uffdio_continue uffdio_continue; | |
1908 | struct uffdio_continue __user *user_uffdio_continue; | |
1909 | struct userfaultfd_wake_range range; | |
02891844 | 1910 | uffd_flags_t flags = 0; |
f6191471 AR |
1911 | |
1912 | user_uffdio_continue = (struct uffdio_continue __user *)arg; | |
1913 | ||
1914 | ret = -EAGAIN; | |
a759a909 | 1915 | if (atomic_read(&ctx->mmap_changing)) |
f6191471 AR |
1916 | goto out; |
1917 | ||
1918 | ret = -EFAULT; | |
1919 | if (copy_from_user(&uffdio_continue, user_uffdio_continue, | |
1920 | /* don't copy the output fields */ | |
1921 | sizeof(uffdio_continue) - (sizeof(__s64)))) | |
1922 | goto out; | |
1923 | ||
e71e2ace | 1924 | ret = validate_range(ctx->mm, uffdio_continue.range.start, |
f6191471 AR |
1925 | uffdio_continue.range.len); |
1926 | if (ret) | |
1927 | goto out; | |
1928 | ||
1929 | ret = -EINVAL; | |
1930 | /* double check for wraparound just in case. */ | |
1931 | if (uffdio_continue.range.start + uffdio_continue.range.len <= | |
1932 | uffdio_continue.range.start) { | |
1933 | goto out; | |
1934 | } | |
02891844 AR |
1935 | if (uffdio_continue.mode & ~(UFFDIO_CONTINUE_MODE_DONTWAKE | |
1936 | UFFDIO_CONTINUE_MODE_WP)) | |
f6191471 | 1937 | goto out; |
02891844 AR |
1938 | if (uffdio_continue.mode & UFFDIO_CONTINUE_MODE_WP) |
1939 | flags |= MFILL_ATOMIC_WP; | |
f6191471 AR |
1940 | |
1941 | if (mmget_not_zero(ctx->mm)) { | |
a734991c AR |
1942 | ret = mfill_atomic_continue(ctx->mm, uffdio_continue.range.start, |
1943 | uffdio_continue.range.len, | |
02891844 | 1944 | &ctx->mmap_changing, flags); |
f6191471 AR |
1945 | mmput(ctx->mm); |
1946 | } else { | |
1947 | return -ESRCH; | |
1948 | } | |
1949 | ||
1950 | if (unlikely(put_user(ret, &user_uffdio_continue->mapped))) | |
1951 | return -EFAULT; | |
1952 | if (ret < 0) | |
1953 | goto out; | |
1954 | ||
1955 | /* len == 0 would wake all */ | |
1956 | BUG_ON(!ret); | |
1957 | range.len = ret; | |
1958 | if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) { | |
1959 | range.start = uffdio_continue.range.start; | |
1960 | wake_userfault(ctx, &range); | |
1961 | } | |
1962 | ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN; | |
1963 | ||
1964 | out: | |
1965 | return ret; | |
1966 | } | |
1967 | ||
9cd75c3c PE |
1968 | static inline unsigned int uffd_ctx_features(__u64 user_features) |
1969 | { | |
1970 | /* | |
22e5fe2a NA |
1971 | * For the current set of features the bits just coincide. Set |
1972 | * UFFD_FEATURE_INITIALIZED to mark the features as enabled. | |
9cd75c3c | 1973 | */ |
22e5fe2a | 1974 | return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED; |
9cd75c3c PE |
1975 | } |
1976 | ||
86039bd3 AA |
1977 | /* |
1978 | * userland asks for a certain API version and we return which bits | |
1979 | * and ioctl commands are implemented in this kernel for such API | |
1980 | * version or -EINVAL if unknown. | |
1981 | */ | |
1982 | static int userfaultfd_api(struct userfaultfd_ctx *ctx, | |
1983 | unsigned long arg) | |
1984 | { | |
1985 | struct uffdio_api uffdio_api; | |
1986 | void __user *buf = (void __user *)arg; | |
22e5fe2a | 1987 | unsigned int ctx_features; |
86039bd3 | 1988 | int ret; |
65603144 | 1989 | __u64 features; |
86039bd3 | 1990 | |
86039bd3 | 1991 | ret = -EFAULT; |
a9b85f94 | 1992 | if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api))) |
86039bd3 | 1993 | goto out; |
2ff559f3 PX |
1994 | features = uffdio_api.features; |
1995 | ret = -EINVAL; | |
1996 | if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) | |
1997 | goto err_out; | |
3c1c24d9 MR |
1998 | ret = -EPERM; |
1999 | if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE)) | |
2000 | goto err_out; | |
65603144 AA |
2001 | /* report all available features and ioctls to userland */ |
2002 | uffdio_api.features = UFFD_API_FEATURES; | |
7677f7fd | 2003 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR |
964ab004 AR |
2004 | uffdio_api.features &= |
2005 | ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM); | |
00b151f2 PX |
2006 | #endif |
2007 | #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP | |
2008 | uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP; | |
b1f9e876 PX |
2009 | #endif |
2010 | #ifndef CONFIG_PTE_MARKER_UFFD_WP | |
2011 | uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM; | |
2bad466c | 2012 | uffdio_api.features &= ~UFFD_FEATURE_WP_UNPOPULATED; |
7677f7fd | 2013 | #endif |
86039bd3 AA |
2014 | uffdio_api.ioctls = UFFD_API_IOCTLS; |
2015 | ret = -EFAULT; | |
2016 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
2017 | goto out; | |
22e5fe2a | 2018 | |
65603144 | 2019 | /* only enable the requested features for this uffd context */ |
22e5fe2a NA |
2020 | ctx_features = uffd_ctx_features(features); |
2021 | ret = -EINVAL; | |
2022 | if (cmpxchg(&ctx->features, 0, ctx_features) != 0) | |
2023 | goto err_out; | |
2024 | ||
86039bd3 AA |
2025 | ret = 0; |
2026 | out: | |
2027 | return ret; | |
3c1c24d9 MR |
2028 | err_out: |
2029 | memset(&uffdio_api, 0, sizeof(uffdio_api)); | |
2030 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
2031 | ret = -EFAULT; | |
2032 | goto out; | |
86039bd3 AA |
2033 | } |
2034 | ||
2035 | static long userfaultfd_ioctl(struct file *file, unsigned cmd, | |
2036 | unsigned long arg) | |
2037 | { | |
2038 | int ret = -EINVAL; | |
2039 | struct userfaultfd_ctx *ctx = file->private_data; | |
2040 | ||
22e5fe2a | 2041 | if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx)) |
e6485a47 AA |
2042 | return -EINVAL; |
2043 | ||
86039bd3 AA |
2044 | switch(cmd) { |
2045 | case UFFDIO_API: | |
2046 | ret = userfaultfd_api(ctx, arg); | |
2047 | break; | |
2048 | case UFFDIO_REGISTER: | |
2049 | ret = userfaultfd_register(ctx, arg); | |
2050 | break; | |
2051 | case UFFDIO_UNREGISTER: | |
2052 | ret = userfaultfd_unregister(ctx, arg); | |
2053 | break; | |
2054 | case UFFDIO_WAKE: | |
2055 | ret = userfaultfd_wake(ctx, arg); | |
2056 | break; | |
ad465cae AA |
2057 | case UFFDIO_COPY: |
2058 | ret = userfaultfd_copy(ctx, arg); | |
2059 | break; | |
2060 | case UFFDIO_ZEROPAGE: | |
2061 | ret = userfaultfd_zeropage(ctx, arg); | |
2062 | break; | |
63b2d417 AA |
2063 | case UFFDIO_WRITEPROTECT: |
2064 | ret = userfaultfd_writeprotect(ctx, arg); | |
2065 | break; | |
f6191471 AR |
2066 | case UFFDIO_CONTINUE: |
2067 | ret = userfaultfd_continue(ctx, arg); | |
2068 | break; | |
86039bd3 AA |
2069 | } |
2070 | return ret; | |
2071 | } | |
2072 | ||
2073 | #ifdef CONFIG_PROC_FS | |
2074 | static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f) | |
2075 | { | |
2076 | struct userfaultfd_ctx *ctx = f->private_data; | |
ac6424b9 | 2077 | wait_queue_entry_t *wq; |
86039bd3 AA |
2078 | unsigned long pending = 0, total = 0; |
2079 | ||
cbcfa130 | 2080 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
2055da97 | 2081 | list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) { |
15b726ef AA |
2082 | pending++; |
2083 | total++; | |
2084 | } | |
2055da97 | 2085 | list_for_each_entry(wq, &ctx->fault_wqh.head, entry) { |
86039bd3 AA |
2086 | total++; |
2087 | } | |
cbcfa130 | 2088 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
2089 | |
2090 | /* | |
2091 | * If more protocols will be added, there will be all shown | |
2092 | * separated by a space. Like this: | |
2093 | * protocols: aa:... bb:... | |
2094 | */ | |
2095 | seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n", | |
045098e9 | 2096 | pending, total, UFFD_API, ctx->features, |
86039bd3 AA |
2097 | UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS); |
2098 | } | |
2099 | #endif | |
2100 | ||
2101 | static const struct file_operations userfaultfd_fops = { | |
2102 | #ifdef CONFIG_PROC_FS | |
2103 | .show_fdinfo = userfaultfd_show_fdinfo, | |
2104 | #endif | |
2105 | .release = userfaultfd_release, | |
2106 | .poll = userfaultfd_poll, | |
2107 | .read = userfaultfd_read, | |
2108 | .unlocked_ioctl = userfaultfd_ioctl, | |
1832f2d8 | 2109 | .compat_ioctl = compat_ptr_ioctl, |
86039bd3 AA |
2110 | .llseek = noop_llseek, |
2111 | }; | |
2112 | ||
3004ec9c AA |
2113 | static void init_once_userfaultfd_ctx(void *mem) |
2114 | { | |
2115 | struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem; | |
2116 | ||
2117 | init_waitqueue_head(&ctx->fault_pending_wqh); | |
2118 | init_waitqueue_head(&ctx->fault_wqh); | |
9cd75c3c | 2119 | init_waitqueue_head(&ctx->event_wqh); |
3004ec9c | 2120 | init_waitqueue_head(&ctx->fd_wqh); |
2ca97ac8 | 2121 | seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock); |
3004ec9c AA |
2122 | } |
2123 | ||
2d5de004 | 2124 | static int new_userfaultfd(int flags) |
86039bd3 | 2125 | { |
86039bd3 | 2126 | struct userfaultfd_ctx *ctx; |
284cd241 | 2127 | int fd; |
86039bd3 AA |
2128 | |
2129 | BUG_ON(!current->mm); | |
2130 | ||
2131 | /* Check the UFFD_* constants for consistency. */ | |
37cd0575 | 2132 | BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS); |
86039bd3 AA |
2133 | BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC); |
2134 | BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK); | |
2135 | ||
37cd0575 | 2136 | if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY)) |
284cd241 | 2137 | return -EINVAL; |
86039bd3 | 2138 | |
3004ec9c | 2139 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); |
86039bd3 | 2140 | if (!ctx) |
284cd241 | 2141 | return -ENOMEM; |
86039bd3 | 2142 | |
ca880420 | 2143 | refcount_set(&ctx->refcount, 1); |
86039bd3 | 2144 | ctx->flags = flags; |
9cd75c3c | 2145 | ctx->features = 0; |
86039bd3 | 2146 | ctx->released = false; |
a759a909 | 2147 | atomic_set(&ctx->mmap_changing, 0); |
86039bd3 AA |
2148 | ctx->mm = current->mm; |
2149 | /* prevent the mm struct to be freed */ | |
f1f10076 | 2150 | mmgrab(ctx->mm); |
86039bd3 | 2151 | |
b537900f | 2152 | fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx, |
abec3d01 | 2153 | O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL); |
284cd241 | 2154 | if (fd < 0) { |
d2005e3f | 2155 | mmdrop(ctx->mm); |
3004ec9c | 2156 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
c03e946f | 2157 | } |
86039bd3 | 2158 | return fd; |
86039bd3 | 2159 | } |
3004ec9c | 2160 | |
2d5de004 AR |
2161 | static inline bool userfaultfd_syscall_allowed(int flags) |
2162 | { | |
2163 | /* Userspace-only page faults are always allowed */ | |
2164 | if (flags & UFFD_USER_MODE_ONLY) | |
2165 | return true; | |
2166 | ||
2167 | /* | |
2168 | * The user is requesting a userfaultfd which can handle kernel faults. | |
2169 | * Privileged users are always allowed to do this. | |
2170 | */ | |
2171 | if (capable(CAP_SYS_PTRACE)) | |
2172 | return true; | |
2173 | ||
2174 | /* Otherwise, access to kernel fault handling is sysctl controlled. */ | |
2175 | return sysctl_unprivileged_userfaultfd; | |
2176 | } | |
2177 | ||
2178 | SYSCALL_DEFINE1(userfaultfd, int, flags) | |
2179 | { | |
2180 | if (!userfaultfd_syscall_allowed(flags)) | |
2181 | return -EPERM; | |
2182 | ||
2183 | return new_userfaultfd(flags); | |
2184 | } | |
2185 | ||
2186 | static long userfaultfd_dev_ioctl(struct file *file, unsigned int cmd, unsigned long flags) | |
2187 | { | |
2188 | if (cmd != USERFAULTFD_IOC_NEW) | |
2189 | return -EINVAL; | |
2190 | ||
2191 | return new_userfaultfd(flags); | |
2192 | } | |
2193 | ||
2194 | static const struct file_operations userfaultfd_dev_fops = { | |
2195 | .unlocked_ioctl = userfaultfd_dev_ioctl, | |
2196 | .compat_ioctl = userfaultfd_dev_ioctl, | |
2197 | .owner = THIS_MODULE, | |
2198 | .llseek = noop_llseek, | |
2199 | }; | |
2200 | ||
2201 | static struct miscdevice userfaultfd_misc = { | |
2202 | .minor = MISC_DYNAMIC_MINOR, | |
2203 | .name = "userfaultfd", | |
2204 | .fops = &userfaultfd_dev_fops | |
2205 | }; | |
2206 | ||
3004ec9c AA |
2207 | static int __init userfaultfd_init(void) |
2208 | { | |
2d5de004 AR |
2209 | int ret; |
2210 | ||
2211 | ret = misc_register(&userfaultfd_misc); | |
2212 | if (ret) | |
2213 | return ret; | |
2214 | ||
3004ec9c AA |
2215 | userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache", |
2216 | sizeof(struct userfaultfd_ctx), | |
2217 | 0, | |
2218 | SLAB_HWCACHE_ALIGN|SLAB_PANIC, | |
2219 | init_once_userfaultfd_ctx); | |
2d337b71 Z |
2220 | #ifdef CONFIG_SYSCTL |
2221 | register_sysctl_init("vm", vm_userfaultfd_table); | |
2222 | #endif | |
3004ec9c AA |
2223 | return 0; |
2224 | } | |
2225 | __initcall(userfaultfd_init); |