blob: 2d97952e341a859ce977ddbd8968536a64c6290c [file] [log] [blame]
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001/*
2 * fs/userfaultfd.c
3 *
4 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
5 * Copyright (C) 2008-2009 Red Hat, Inc.
6 * Copyright (C) 2015 Red Hat, Inc.
7 *
8 * This work is licensed under the terms of the GNU GPL, version 2. See
9 * the COPYING file in the top-level directory.
10 *
11 * Some part derived from fs/eventfd.c (anon inode setup) and
12 * mm/ksm.c (mm hashing).
13 */
14
15#include <linux/hashtable.h>
16#include <linux/sched.h>
17#include <linux/mm.h>
18#include <linux/poll.h>
19#include <linux/slab.h>
20#include <linux/seq_file.h>
21#include <linux/file.h>
22#include <linux/bug.h>
23#include <linux/anon_inodes.h>
24#include <linux/syscalls.h>
25#include <linux/userfaultfd_k.h>
26#include <linux/mempolicy.h>
27#include <linux/ioctl.h>
28#include <linux/security.h>
29
Andrea Arcangeli3004ec92015-09-04 15:46:48 -070030static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
31
Andrea Arcangeli86039bd2015-09-04 15:46:31 -070032enum userfaultfd_state {
33 UFFD_STATE_WAIT_API,
34 UFFD_STATE_RUNNING,
35};
36
Andrea Arcangeli3004ec92015-09-04 15:46:48 -070037/*
38 * Start with fault_pending_wqh and fault_wqh so they're more likely
39 * to be in the same cacheline.
40 */
Andrea Arcangeli86039bd2015-09-04 15:46:31 -070041struct userfaultfd_ctx {
Andrea Arcangeli15b726e2015-09-04 15:46:44 -070042 /* waitqueue head for the pending (i.e. not read) userfaults */
43 wait_queue_head_t fault_pending_wqh;
44 /* waitqueue head for the userfaults */
Andrea Arcangeli86039bd2015-09-04 15:46:31 -070045 wait_queue_head_t fault_wqh;
46 /* waitqueue head for the pseudo fd to wakeup poll/read */
47 wait_queue_head_t fd_wqh;
Andrea Arcangeli2c5b7e12015-09-04 15:47:23 -070048 /* a refile sequence protected by fault_pending_wqh lock */
49 struct seqcount refile_seq;
Andrea Arcangeli3004ec92015-09-04 15:46:48 -070050 /* pseudo fd refcounting */
51 atomic_t refcount;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -070052 /* userfaultfd syscall flags */
53 unsigned int flags;
54 /* state machine */
55 enum userfaultfd_state state;
56 /* released */
57 bool released;
58 /* mm with one ore more vmas attached to this userfaultfd_ctx */
59 struct mm_struct *mm;
60};
61
62struct userfaultfd_wait_queue {
Andrea Arcangelia9b85f92015-09-04 15:46:37 -070063 struct uffd_msg msg;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -070064 wait_queue_t wq;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -070065 struct userfaultfd_ctx *ctx;
66};
67
68struct userfaultfd_wake_range {
69 unsigned long start;
70 unsigned long len;
71};
72
73static int userfaultfd_wake_function(wait_queue_t *wq, unsigned mode,
74 int wake_flags, void *key)
75{
76 struct userfaultfd_wake_range *range = key;
77 int ret;
78 struct userfaultfd_wait_queue *uwq;
79 unsigned long start, len;
80
81 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
82 ret = 0;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -070083 /* len == 0 means wake all */
84 start = range->start;
85 len = range->len;
Andrea Arcangelia9b85f92015-09-04 15:46:37 -070086 if (len && (start > uwq->msg.arg.pagefault.address ||
87 start + len <= uwq->msg.arg.pagefault.address))
Andrea Arcangeli86039bd2015-09-04 15:46:31 -070088 goto out;
89 ret = wake_up_state(wq->private, mode);
90 if (ret)
91 /*
92 * Wake only once, autoremove behavior.
93 *
94 * After the effect of list_del_init is visible to the
95 * other CPUs, the waitqueue may disappear from under
96 * us, see the !list_empty_careful() in
97 * handle_userfault(). try_to_wake_up() has an
98 * implicit smp_mb__before_spinlock, and the
99 * wq->private is read before calling the extern
100 * function "wake_up_state" (which in turns calls
101 * try_to_wake_up). While the spin_lock;spin_unlock;
102 * wouldn't be enough, the smp_mb__before_spinlock is
103 * enough to avoid an explicit smp_mb() here.
104 */
105 list_del_init(&wq->task_list);
106out:
107 return ret;
108}
109
110/**
111 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
112 * context.
113 * @ctx: [in] Pointer to the userfaultfd context.
114 *
115 * Returns: In case of success, returns not zero.
116 */
117static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
118{
119 if (!atomic_inc_not_zero(&ctx->refcount))
120 BUG();
121}
122
123/**
124 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
125 * context.
126 * @ctx: [in] Pointer to userfaultfd context.
127 *
128 * The userfaultfd context reference must have been previously acquired either
129 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
130 */
131static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
132{
133 if (atomic_dec_and_test(&ctx->refcount)) {
134 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
135 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
136 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
137 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
138 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
139 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
Oleg Nesterovd2005e32016-05-20 16:58:36 -0700140 mmdrop(ctx->mm);
Andrea Arcangeli3004ec92015-09-04 15:46:48 -0700141 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700142 }
143}
144
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700145static inline void msg_init(struct uffd_msg *msg)
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700146{
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700147 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
148 /*
149 * Must use memset to zero out the paddings or kernel data is
150 * leaked to userland.
151 */
152 memset(msg, 0, sizeof(struct uffd_msg));
153}
154
155static inline struct uffd_msg userfault_msg(unsigned long address,
156 unsigned int flags,
157 unsigned long reason)
158{
159 struct uffd_msg msg;
160 msg_init(&msg);
161 msg.event = UFFD_EVENT_PAGEFAULT;
162 msg.arg.pagefault.address = address;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700163 if (flags & FAULT_FLAG_WRITE)
164 /*
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700165 * If UFFD_FEATURE_PAGEFAULT_FLAG_WRITE was set in the
166 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
167 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
168 * was a read fault, otherwise if set it means it's
169 * a write fault.
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700170 */
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700171 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700172 if (reason & VM_UFFD_WP)
173 /*
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700174 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
175 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
176 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
177 * a missing fault, otherwise if set it means it's a
178 * write protect fault.
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700179 */
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700180 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
181 return msg;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700182}
183
184/*
Andrea Arcangeli8d2afd92015-09-04 15:46:51 -0700185 * Verify the pagetables are still not ok after having reigstered into
186 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
187 * userfault that has already been resolved, if userfaultfd_read and
188 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
189 * threads.
190 */
191static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
192 unsigned long address,
193 unsigned long flags,
194 unsigned long reason)
195{
196 struct mm_struct *mm = ctx->mm;
197 pgd_t *pgd;
198 pud_t *pud;
199 pmd_t *pmd, _pmd;
200 pte_t *pte;
201 bool ret = true;
202
203 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
204
205 pgd = pgd_offset(mm, address);
206 if (!pgd_present(*pgd))
207 goto out;
208 pud = pud_offset(pgd, address);
209 if (!pud_present(*pud))
210 goto out;
211 pmd = pmd_offset(pud, address);
212 /*
213 * READ_ONCE must function as a barrier with narrower scope
214 * and it must be equivalent to:
215 * _pmd = *pmd; barrier();
216 *
217 * This is to deal with the instability (as in
218 * pmd_trans_unstable) of the pmd.
219 */
220 _pmd = READ_ONCE(*pmd);
221 if (!pmd_present(_pmd))
222 goto out;
223
224 ret = false;
225 if (pmd_trans_huge(_pmd))
226 goto out;
227
228 /*
229 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
230 * and use the standard pte_offset_map() instead of parsing _pmd.
231 */
232 pte = pte_offset_map(pmd, address);
233 /*
234 * Lockless access: we're in a wait_event so it's ok if it
235 * changes under us.
236 */
237 if (pte_none(*pte))
238 ret = true;
239 pte_unmap(pte);
240
241out:
242 return ret;
243}
244
245/*
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700246 * The locking rules involved in returning VM_FAULT_RETRY depending on
247 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
248 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
249 * recommendation in __lock_page_or_retry is not an understatement.
250 *
251 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
252 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
253 * not set.
254 *
255 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
256 * set, VM_FAULT_RETRY can still be returned if and only if there are
257 * fatal_signal_pending()s, and the mmap_sem must be released before
258 * returning it.
259 */
260int handle_userfault(struct vm_area_struct *vma, unsigned long address,
261 unsigned int flags, unsigned long reason)
262{
263 struct mm_struct *mm = vma->vm_mm;
264 struct userfaultfd_ctx *ctx;
265 struct userfaultfd_wait_queue uwq;
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700266 int ret;
Andrea Arcangelidfa37dc2015-09-04 15:47:18 -0700267 bool must_wait, return_to_userland;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700268
269 BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
270
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700271 ret = VM_FAULT_SIGBUS;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700272 ctx = vma->vm_userfaultfd_ctx.ctx;
273 if (!ctx)
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700274 goto out;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700275
276 BUG_ON(ctx->mm != mm);
277
278 VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
279 VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
280
281 /*
282 * If it's already released don't get it. This avoids to loop
283 * in __get_user_pages if userfaultfd_release waits on the
284 * caller of handle_userfault to release the mmap_sem.
285 */
286 if (unlikely(ACCESS_ONCE(ctx->released)))
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700287 goto out;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700288
289 /*
Linus Torvalds39680f52016-03-01 11:56:22 -0800290 * We don't do userfault handling for the final child pid update.
291 */
292 if (current->flags & PF_EXITING)
293 goto out;
294
295 /*
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700296 * Check that we can return VM_FAULT_RETRY.
297 *
298 * NOTE: it should become possible to return VM_FAULT_RETRY
299 * even if FAULT_FLAG_TRIED is set without leading to gup()
300 * -EBUSY failures, if the userfaultfd is to be extended for
301 * VM_UFFD_WP tracking and we intend to arm the userfault
302 * without first stopping userland access to the memory. For
303 * VM_UFFD_MISSING userfaults this is enough for now.
304 */
305 if (unlikely(!(flags & FAULT_FLAG_ALLOW_RETRY))) {
306 /*
307 * Validate the invariant that nowait must allow retry
308 * to be sure not to return SIGBUS erroneously on
309 * nowait invocations.
310 */
311 BUG_ON(flags & FAULT_FLAG_RETRY_NOWAIT);
312#ifdef CONFIG_DEBUG_VM
313 if (printk_ratelimit()) {
314 printk(KERN_WARNING
315 "FAULT_FLAG_ALLOW_RETRY missing %x\n", flags);
316 dump_stack();
317 }
318#endif
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700319 goto out;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700320 }
321
322 /*
323 * Handle nowait, not much to do other than tell it to retry
324 * and wait.
325 */
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700326 ret = VM_FAULT_RETRY;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700327 if (flags & FAULT_FLAG_RETRY_NOWAIT)
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700328 goto out;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700329
330 /* take the reference before dropping the mmap_sem */
331 userfaultfd_ctx_get(ctx);
332
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700333 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
334 uwq.wq.private = current;
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700335 uwq.msg = userfault_msg(address, flags, reason);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700336 uwq.ctx = ctx;
337
Andrea Arcangelidfa37dc2015-09-04 15:47:18 -0700338 return_to_userland = (flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
339 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
340
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700341 spin_lock(&ctx->fault_pending_wqh.lock);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700342 /*
343 * After the __add_wait_queue the uwq is visible to userland
344 * through poll/read().
345 */
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700346 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
347 /*
348 * The smp_mb() after __set_current_state prevents the reads
349 * following the spin_unlock to happen before the list_add in
350 * __add_wait_queue.
351 */
Andrea Arcangelidfa37dc2015-09-04 15:47:18 -0700352 set_current_state(return_to_userland ? TASK_INTERRUPTIBLE :
353 TASK_KILLABLE);
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700354 spin_unlock(&ctx->fault_pending_wqh.lock);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700355
Andrea Arcangeli8d2afd92015-09-04 15:46:51 -0700356 must_wait = userfaultfd_must_wait(ctx, address, flags, reason);
357 up_read(&mm->mmap_sem);
358
359 if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
Andrea Arcangelidfa37dc2015-09-04 15:47:18 -0700360 (return_to_userland ? !signal_pending(current) :
361 !fatal_signal_pending(current)))) {
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700362 wake_up_poll(&ctx->fd_wqh, POLLIN);
363 schedule();
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700364 ret |= VM_FAULT_MAJOR;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700365 }
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700366
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700367 __set_current_state(TASK_RUNNING);
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700368
Andrea Arcangelidfa37dc2015-09-04 15:47:18 -0700369 if (return_to_userland) {
370 if (signal_pending(current) &&
371 !fatal_signal_pending(current)) {
372 /*
373 * If we got a SIGSTOP or SIGCONT and this is
374 * a normal userland page fault, just let
375 * userland return so the signal will be
376 * handled and gdb debugging works. The page
377 * fault code immediately after we return from
378 * this function is going to release the
379 * mmap_sem and it's not depending on it
380 * (unlike gup would if we were not to return
381 * VM_FAULT_RETRY).
382 *
383 * If a fatal signal is pending we still take
384 * the streamlined VM_FAULT_RETRY failure path
385 * and there's no need to retake the mmap_sem
386 * in such case.
387 */
388 down_read(&mm->mmap_sem);
389 ret = 0;
390 }
391 }
392
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700393 /*
394 * Here we race with the list_del; list_add in
395 * userfaultfd_ctx_read(), however because we don't ever run
396 * list_del_init() to refile across the two lists, the prev
397 * and next pointers will never point to self. list_add also
398 * would never let any of the two pointers to point to
399 * self. So list_empty_careful won't risk to see both pointers
400 * pointing to self at any time during the list refile. The
401 * only case where list_del_init() is called is the full
402 * removal in the wake function and there we don't re-list_add
403 * and it's fine not to block on the spinlock. The uwq on this
404 * kernel stack can be released after the list_del_init.
405 */
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700406 if (!list_empty_careful(&uwq.wq.task_list)) {
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700407 spin_lock(&ctx->fault_pending_wqh.lock);
408 /*
409 * No need of list_del_init(), the uwq on the stack
410 * will be freed shortly anyway.
411 */
412 list_del(&uwq.wq.task_list);
413 spin_unlock(&ctx->fault_pending_wqh.lock);
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700414 }
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700415
416 /*
417 * ctx may go away after this if the userfault pseudo fd is
418 * already released.
419 */
420 userfaultfd_ctx_put(ctx);
421
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700422out:
423 return ret;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700424}
425
426static int userfaultfd_release(struct inode *inode, struct file *file)
427{
428 struct userfaultfd_ctx *ctx = file->private_data;
429 struct mm_struct *mm = ctx->mm;
430 struct vm_area_struct *vma, *prev;
431 /* len == 0 means wake all */
432 struct userfaultfd_wake_range range = { .len = 0, };
433 unsigned long new_flags;
434
435 ACCESS_ONCE(ctx->released) = true;
436
Oleg Nesterovd2005e32016-05-20 16:58:36 -0700437 if (!mmget_not_zero(mm))
438 goto wakeup;
439
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700440 /*
441 * Flush page faults out of all CPUs. NOTE: all page faults
442 * must be retried without returning VM_FAULT_SIGBUS if
443 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
444 * changes while handle_userfault released the mmap_sem. So
445 * it's critical that released is set to true (above), before
446 * taking the mmap_sem for writing.
447 */
448 down_write(&mm->mmap_sem);
449 prev = NULL;
450 for (vma = mm->mmap; vma; vma = vma->vm_next) {
451 cond_resched();
452 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
453 !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
454 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
455 prev = vma;
456 continue;
457 }
458 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
459 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
460 new_flags, vma->anon_vma,
461 vma->vm_file, vma->vm_pgoff,
462 vma_policy(vma),
463 NULL_VM_UFFD_CTX);
464 if (prev)
465 vma = prev;
466 else
467 prev = vma;
468 vma->vm_flags = new_flags;
469 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
470 }
471 up_write(&mm->mmap_sem);
Oleg Nesterovd2005e32016-05-20 16:58:36 -0700472 mmput(mm);
473wakeup:
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700474 /*
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700475 * After no new page faults can wait on this fault_*wqh, flush
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700476 * the last page faults that may have been already waiting on
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700477 * the fault_*wqh.
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700478 */
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700479 spin_lock(&ctx->fault_pending_wqh.lock);
Andrea Arcangeliac5be6b2015-09-22 14:58:49 -0700480 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
481 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700482 spin_unlock(&ctx->fault_pending_wqh.lock);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700483
484 wake_up_poll(&ctx->fd_wqh, POLLHUP);
485 userfaultfd_ctx_put(ctx);
486 return 0;
487}
488
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700489/* fault_pending_wqh.lock must be hold by the caller */
490static inline struct userfaultfd_wait_queue *find_userfault(
491 struct userfaultfd_ctx *ctx)
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700492{
493 wait_queue_t *wq;
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700494 struct userfaultfd_wait_queue *uwq;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700495
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700496 VM_BUG_ON(!spin_is_locked(&ctx->fault_pending_wqh.lock));
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700497
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700498 uwq = NULL;
499 if (!waitqueue_active(&ctx->fault_pending_wqh))
500 goto out;
501 /* walk in reverse to provide FIFO behavior to read userfaults */
502 wq = list_last_entry(&ctx->fault_pending_wqh.task_list,
503 typeof(*wq), task_list);
504 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
505out:
506 return uwq;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700507}
508
509static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
510{
511 struct userfaultfd_ctx *ctx = file->private_data;
512 unsigned int ret;
513
514 poll_wait(file, &ctx->fd_wqh, wait);
515
516 switch (ctx->state) {
517 case UFFD_STATE_WAIT_API:
518 return POLLERR;
519 case UFFD_STATE_RUNNING:
Andrea Arcangeliba85c702015-09-04 15:46:41 -0700520 /*
521 * poll() never guarantees that read won't block.
522 * userfaults can be waken before they're read().
523 */
524 if (unlikely(!(file->f_flags & O_NONBLOCK)))
525 return POLLERR;
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700526 /*
527 * lockless access to see if there are pending faults
528 * __pollwait last action is the add_wait_queue but
529 * the spin_unlock would allow the waitqueue_active to
530 * pass above the actual list_add inside
531 * add_wait_queue critical section. So use a full
532 * memory barrier to serialize the list_add write of
533 * add_wait_queue() with the waitqueue_active read
534 * below.
535 */
536 ret = 0;
537 smp_mb();
538 if (waitqueue_active(&ctx->fault_pending_wqh))
539 ret = POLLIN;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700540 return ret;
541 default:
542 BUG();
543 }
544}
545
546static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700547 struct uffd_msg *msg)
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700548{
549 ssize_t ret;
550 DECLARE_WAITQUEUE(wait, current);
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700551 struct userfaultfd_wait_queue *uwq;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700552
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700553 /* always take the fd_wqh lock before the fault_pending_wqh lock */
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700554 spin_lock(&ctx->fd_wqh.lock);
555 __add_wait_queue(&ctx->fd_wqh, &wait);
556 for (;;) {
557 set_current_state(TASK_INTERRUPTIBLE);
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700558 spin_lock(&ctx->fault_pending_wqh.lock);
559 uwq = find_userfault(ctx);
560 if (uwq) {
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700561 /*
Andrea Arcangeli2c5b7e12015-09-04 15:47:23 -0700562 * Use a seqcount to repeat the lockless check
563 * in wake_userfault() to avoid missing
564 * wakeups because during the refile both
565 * waitqueue could become empty if this is the
566 * only userfault.
567 */
568 write_seqcount_begin(&ctx->refile_seq);
569
570 /*
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700571 * The fault_pending_wqh.lock prevents the uwq
572 * to disappear from under us.
573 *
574 * Refile this userfault from
575 * fault_pending_wqh to fault_wqh, it's not
576 * pending anymore after we read it.
577 *
578 * Use list_del() by hand (as
579 * userfaultfd_wake_function also uses
580 * list_del_init() by hand) to be sure nobody
581 * changes __remove_wait_queue() to use
582 * list_del_init() in turn breaking the
583 * !list_empty_careful() check in
584 * handle_userfault(). The uwq->wq.task_list
585 * must never be empty at any time during the
586 * refile, or the waitqueue could disappear
587 * from under us. The "wait_queue_head_t"
588 * parameter of __remove_wait_queue() is unused
589 * anyway.
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700590 */
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700591 list_del(&uwq->wq.task_list);
592 __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
593
Andrea Arcangeli2c5b7e12015-09-04 15:47:23 -0700594 write_seqcount_end(&ctx->refile_seq);
595
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700596 /* careful to always initialize msg if ret == 0 */
597 *msg = uwq->msg;
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700598 spin_unlock(&ctx->fault_pending_wqh.lock);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700599 ret = 0;
600 break;
601 }
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700602 spin_unlock(&ctx->fault_pending_wqh.lock);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700603 if (signal_pending(current)) {
604 ret = -ERESTARTSYS;
605 break;
606 }
607 if (no_wait) {
608 ret = -EAGAIN;
609 break;
610 }
611 spin_unlock(&ctx->fd_wqh.lock);
612 schedule();
613 spin_lock(&ctx->fd_wqh.lock);
614 }
615 __remove_wait_queue(&ctx->fd_wqh, &wait);
616 __set_current_state(TASK_RUNNING);
617 spin_unlock(&ctx->fd_wqh.lock);
618
619 return ret;
620}
621
622static ssize_t userfaultfd_read(struct file *file, char __user *buf,
623 size_t count, loff_t *ppos)
624{
625 struct userfaultfd_ctx *ctx = file->private_data;
626 ssize_t _ret, ret = 0;
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700627 struct uffd_msg msg;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700628 int no_wait = file->f_flags & O_NONBLOCK;
629
630 if (ctx->state == UFFD_STATE_WAIT_API)
631 return -EINVAL;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700632
633 for (;;) {
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700634 if (count < sizeof(msg))
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700635 return ret ? ret : -EINVAL;
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700636 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700637 if (_ret < 0)
638 return ret ? ret : _ret;
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700639 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700640 return ret ? ret : -EFAULT;
Andrea Arcangelia9b85f92015-09-04 15:46:37 -0700641 ret += sizeof(msg);
642 buf += sizeof(msg);
643 count -= sizeof(msg);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700644 /*
645 * Allow to read more than one fault at time but only
646 * block if waiting for the very first one.
647 */
648 no_wait = O_NONBLOCK;
649 }
650}
651
652static void __wake_userfault(struct userfaultfd_ctx *ctx,
653 struct userfaultfd_wake_range *range)
654{
655 unsigned long start, end;
656
657 start = range->start;
658 end = range->start + range->len;
659
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700660 spin_lock(&ctx->fault_pending_wqh.lock);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700661 /* wake all in the range and autoremove */
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700662 if (waitqueue_active(&ctx->fault_pending_wqh))
Andrea Arcangeliac5be6b2015-09-22 14:58:49 -0700663 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700664 range);
665 if (waitqueue_active(&ctx->fault_wqh))
Andrea Arcangeliac5be6b2015-09-22 14:58:49 -0700666 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
Andrea Arcangeli15b726e2015-09-04 15:46:44 -0700667 spin_unlock(&ctx->fault_pending_wqh.lock);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700668}
669
670static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
671 struct userfaultfd_wake_range *range)
672{
Andrea Arcangeli2c5b7e12015-09-04 15:47:23 -0700673 unsigned seq;
674 bool need_wakeup;
675
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700676 /*
677 * To be sure waitqueue_active() is not reordered by the CPU
678 * before the pagetable update, use an explicit SMP memory
679 * barrier here. PT lock release or up_read(mmap_sem) still
680 * have release semantics that can allow the
681 * waitqueue_active() to be reordered before the pte update.
682 */
683 smp_mb();
684
685 /*
686 * Use waitqueue_active because it's very frequent to
687 * change the address space atomically even if there are no
688 * userfaults yet. So we take the spinlock only when we're
689 * sure we've userfaults to wake.
690 */
Andrea Arcangeli2c5b7e12015-09-04 15:47:23 -0700691 do {
692 seq = read_seqcount_begin(&ctx->refile_seq);
693 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
694 waitqueue_active(&ctx->fault_wqh);
695 cond_resched();
696 } while (read_seqcount_retry(&ctx->refile_seq, seq));
697 if (need_wakeup)
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700698 __wake_userfault(ctx, range);
699}
700
701static __always_inline int validate_range(struct mm_struct *mm,
702 __u64 start, __u64 len)
703{
704 __u64 task_size = mm->task_size;
705
706 if (start & ~PAGE_MASK)
707 return -EINVAL;
708 if (len & ~PAGE_MASK)
709 return -EINVAL;
710 if (!len)
711 return -EINVAL;
712 if (start < mmap_min_addr)
713 return -EINVAL;
714 if (start >= task_size)
715 return -EINVAL;
716 if (len > task_size - start)
717 return -EINVAL;
718 return 0;
719}
720
721static int userfaultfd_register(struct userfaultfd_ctx *ctx,
722 unsigned long arg)
723{
724 struct mm_struct *mm = ctx->mm;
725 struct vm_area_struct *vma, *prev, *cur;
726 int ret;
727 struct uffdio_register uffdio_register;
728 struct uffdio_register __user *user_uffdio_register;
729 unsigned long vm_flags, new_flags;
730 bool found;
731 unsigned long start, end, vma_end;
732
733 user_uffdio_register = (struct uffdio_register __user *) arg;
734
735 ret = -EFAULT;
736 if (copy_from_user(&uffdio_register, user_uffdio_register,
737 sizeof(uffdio_register)-sizeof(__u64)))
738 goto out;
739
740 ret = -EINVAL;
741 if (!uffdio_register.mode)
742 goto out;
743 if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
744 UFFDIO_REGISTER_MODE_WP))
745 goto out;
746 vm_flags = 0;
747 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
748 vm_flags |= VM_UFFD_MISSING;
749 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
750 vm_flags |= VM_UFFD_WP;
751 /*
752 * FIXME: remove the below error constraint by
753 * implementing the wprotect tracking mode.
754 */
755 ret = -EINVAL;
756 goto out;
757 }
758
759 ret = validate_range(mm, uffdio_register.range.start,
760 uffdio_register.range.len);
761 if (ret)
762 goto out;
763
764 start = uffdio_register.range.start;
765 end = start + uffdio_register.range.len;
766
Oleg Nesterovd2005e32016-05-20 16:58:36 -0700767 ret = -ENOMEM;
768 if (!mmget_not_zero(mm))
769 goto out;
770
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700771 down_write(&mm->mmap_sem);
772 vma = find_vma_prev(mm, start, &prev);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700773 if (!vma)
774 goto out_unlock;
775
776 /* check that there's at least one vma in the range */
777 ret = -EINVAL;
778 if (vma->vm_start >= end)
779 goto out_unlock;
780
781 /*
782 * Search for not compatible vmas.
783 *
784 * FIXME: this shall be relaxed later so that it doesn't fail
785 * on tmpfs backed vmas (in addition to the current allowance
786 * on anonymous vmas).
787 */
788 found = false;
789 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
790 cond_resched();
791
792 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
793 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
794
795 /* check not compatible vmas */
796 ret = -EINVAL;
797 if (cur->vm_ops)
798 goto out_unlock;
799
800 /*
801 * Check that this vma isn't already owned by a
802 * different userfaultfd. We can't allow more than one
803 * userfaultfd to own a single vma simultaneously or we
804 * wouldn't know which one to deliver the userfaults to.
805 */
806 ret = -EBUSY;
807 if (cur->vm_userfaultfd_ctx.ctx &&
808 cur->vm_userfaultfd_ctx.ctx != ctx)
809 goto out_unlock;
810
811 found = true;
812 }
813 BUG_ON(!found);
814
815 if (vma->vm_start < start)
816 prev = vma;
817
818 ret = 0;
819 do {
820 cond_resched();
821
822 BUG_ON(vma->vm_ops);
823 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
824 vma->vm_userfaultfd_ctx.ctx != ctx);
825
826 /*
827 * Nothing to do: this vma is already registered into this
828 * userfaultfd and with the right tracking mode too.
829 */
830 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
831 (vma->vm_flags & vm_flags) == vm_flags)
832 goto skip;
833
834 if (vma->vm_start > start)
835 start = vma->vm_start;
836 vma_end = min(end, vma->vm_end);
837
838 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
839 prev = vma_merge(mm, prev, start, vma_end, new_flags,
840 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
841 vma_policy(vma),
842 ((struct vm_userfaultfd_ctx){ ctx }));
843 if (prev) {
844 vma = prev;
845 goto next;
846 }
847 if (vma->vm_start < start) {
848 ret = split_vma(mm, vma, start, 1);
849 if (ret)
850 break;
851 }
852 if (vma->vm_end > end) {
853 ret = split_vma(mm, vma, end, 0);
854 if (ret)
855 break;
856 }
857 next:
858 /*
859 * In the vma_merge() successful mprotect-like case 8:
860 * the next vma was merged into the current one and
861 * the current one has not been updated yet.
862 */
863 vma->vm_flags = new_flags;
864 vma->vm_userfaultfd_ctx.ctx = ctx;
865
866 skip:
867 prev = vma;
868 start = vma->vm_end;
869 vma = vma->vm_next;
870 } while (vma && vma->vm_start < end);
871out_unlock:
872 up_write(&mm->mmap_sem);
Oleg Nesterovd2005e32016-05-20 16:58:36 -0700873 mmput(mm);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700874 if (!ret) {
875 /*
876 * Now that we scanned all vmas we can already tell
877 * userland which ioctls methods are guaranteed to
878 * succeed on this range.
879 */
880 if (put_user(UFFD_API_RANGE_IOCTLS,
881 &user_uffdio_register->ioctls))
882 ret = -EFAULT;
883 }
884out:
885 return ret;
886}
887
888static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
889 unsigned long arg)
890{
891 struct mm_struct *mm = ctx->mm;
892 struct vm_area_struct *vma, *prev, *cur;
893 int ret;
894 struct uffdio_range uffdio_unregister;
895 unsigned long new_flags;
896 bool found;
897 unsigned long start, end, vma_end;
898 const void __user *buf = (void __user *)arg;
899
900 ret = -EFAULT;
901 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
902 goto out;
903
904 ret = validate_range(mm, uffdio_unregister.start,
905 uffdio_unregister.len);
906 if (ret)
907 goto out;
908
909 start = uffdio_unregister.start;
910 end = start + uffdio_unregister.len;
911
Oleg Nesterovd2005e32016-05-20 16:58:36 -0700912 ret = -ENOMEM;
913 if (!mmget_not_zero(mm))
914 goto out;
915
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700916 down_write(&mm->mmap_sem);
917 vma = find_vma_prev(mm, start, &prev);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -0700918 if (!vma)
919 goto out_unlock;
920
921 /* check that there's at least one vma in the range */
922 ret = -EINVAL;
923 if (vma->vm_start >= end)
924 goto out_unlock;
925
926 /*
927 * Search for not compatible vmas.
928 *
929 * FIXME: this shall be relaxed later so that it doesn't fail
930 * on tmpfs backed vmas (in addition to the current allowance
931 * on anonymous vmas).
932 */
933 found = false;
934 ret = -EINVAL;
935 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
936 cond_resched();
937
938 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
939 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
940
941 /*
942 * Check not compatible vmas, not strictly required
943 * here as not compatible vmas cannot have an
944 * userfaultfd_ctx registered on them, but this
945 * provides for more strict behavior to notice
946 * unregistration errors.
947 */
948 if (cur->vm_ops)
949 goto out_unlock;
950
951 found = true;
952 }
953 BUG_ON(!found);
954
955 if (vma->vm_start < start)
956 prev = vma;
957
958 ret = 0;
959 do {
960 cond_resched();
961
962 BUG_ON(vma->vm_ops);
963
964 /*
965 * Nothing to do: this vma is already registered into this
966 * userfaultfd and with the right tracking mode too.
967 */
968 if (!vma->vm_userfaultfd_ctx.ctx)
969 goto skip;
970
971 if (vma->vm_start > start)
972 start = vma->vm_start;
973 vma_end = min(end, vma->vm_end);
974
975 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
976 prev = vma_merge(mm, prev, start, vma_end, new_flags,
977 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
978 vma_policy(vma),
979 NULL_VM_UFFD_CTX);
980 if (prev) {
981 vma = prev;
982 goto next;
983 }
984 if (vma->vm_start < start) {
985 ret = split_vma(mm, vma, start, 1);
986 if (ret)
987 break;
988 }
989 if (vma->vm_end > end) {
990 ret = split_vma(mm, vma, end, 0);
991 if (ret)
992 break;
993 }
994 next:
995 /*
996 * In the vma_merge() successful mprotect-like case 8:
997 * the next vma was merged into the current one and
998 * the current one has not been updated yet.
999 */
1000 vma->vm_flags = new_flags;
1001 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1002
1003 skip:
1004 prev = vma;
1005 start = vma->vm_end;
1006 vma = vma->vm_next;
1007 } while (vma && vma->vm_start < end);
1008out_unlock:
1009 up_write(&mm->mmap_sem);
Oleg Nesterovd2005e32016-05-20 16:58:36 -07001010 mmput(mm);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001011out:
1012 return ret;
1013}
1014
1015/*
Andrea Arcangeliba85c702015-09-04 15:46:41 -07001016 * userfaultfd_wake may be used in combination with the
1017 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001018 */
1019static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1020 unsigned long arg)
1021{
1022 int ret;
1023 struct uffdio_range uffdio_wake;
1024 struct userfaultfd_wake_range range;
1025 const void __user *buf = (void __user *)arg;
1026
1027 ret = -EFAULT;
1028 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1029 goto out;
1030
1031 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1032 if (ret)
1033 goto out;
1034
1035 range.start = uffdio_wake.start;
1036 range.len = uffdio_wake.len;
1037
1038 /*
1039 * len == 0 means wake all and we don't want to wake all here,
1040 * so check it again to be sure.
1041 */
1042 VM_BUG_ON(!range.len);
1043
1044 wake_userfault(ctx, &range);
1045 ret = 0;
1046
1047out:
1048 return ret;
1049}
1050
Andrea Arcangeliad465cae2015-09-04 15:47:11 -07001051static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1052 unsigned long arg)
1053{
1054 __s64 ret;
1055 struct uffdio_copy uffdio_copy;
1056 struct uffdio_copy __user *user_uffdio_copy;
1057 struct userfaultfd_wake_range range;
1058
1059 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1060
1061 ret = -EFAULT;
1062 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1063 /* don't copy "copy" last field */
1064 sizeof(uffdio_copy)-sizeof(__s64)))
1065 goto out;
1066
1067 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1068 if (ret)
1069 goto out;
1070 /*
1071 * double check for wraparound just in case. copy_from_user()
1072 * will later check uffdio_copy.src + uffdio_copy.len to fit
1073 * in the userland range.
1074 */
1075 ret = -EINVAL;
1076 if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1077 goto out;
1078 if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1079 goto out;
Oleg Nesterovd2005e32016-05-20 16:58:36 -07001080 if (mmget_not_zero(ctx->mm)) {
1081 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1082 uffdio_copy.len);
1083 mmput(ctx->mm);
1084 }
Andrea Arcangeliad465cae2015-09-04 15:47:11 -07001085 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1086 return -EFAULT;
1087 if (ret < 0)
1088 goto out;
1089 BUG_ON(!ret);
1090 /* len == 0 would wake all */
1091 range.len = ret;
1092 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1093 range.start = uffdio_copy.dst;
1094 wake_userfault(ctx, &range);
1095 }
1096 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1097out:
1098 return ret;
1099}
1100
1101static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1102 unsigned long arg)
1103{
1104 __s64 ret;
1105 struct uffdio_zeropage uffdio_zeropage;
1106 struct uffdio_zeropage __user *user_uffdio_zeropage;
1107 struct userfaultfd_wake_range range;
1108
1109 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1110
1111 ret = -EFAULT;
1112 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1113 /* don't copy "zeropage" last field */
1114 sizeof(uffdio_zeropage)-sizeof(__s64)))
1115 goto out;
1116
1117 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1118 uffdio_zeropage.range.len);
1119 if (ret)
1120 goto out;
1121 ret = -EINVAL;
1122 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1123 goto out;
1124
Oleg Nesterovd2005e32016-05-20 16:58:36 -07001125 if (mmget_not_zero(ctx->mm)) {
1126 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1127 uffdio_zeropage.range.len);
1128 mmput(ctx->mm);
1129 }
Andrea Arcangeliad465cae2015-09-04 15:47:11 -07001130 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1131 return -EFAULT;
1132 if (ret < 0)
1133 goto out;
1134 /* len == 0 would wake all */
1135 BUG_ON(!ret);
1136 range.len = ret;
1137 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1138 range.start = uffdio_zeropage.range.start;
1139 wake_userfault(ctx, &range);
1140 }
1141 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1142out:
1143 return ret;
1144}
1145
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001146/*
1147 * userland asks for a certain API version and we return which bits
1148 * and ioctl commands are implemented in this kernel for such API
1149 * version or -EINVAL if unknown.
1150 */
1151static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1152 unsigned long arg)
1153{
1154 struct uffdio_api uffdio_api;
1155 void __user *buf = (void __user *)arg;
1156 int ret;
1157
1158 ret = -EINVAL;
1159 if (ctx->state != UFFD_STATE_WAIT_API)
1160 goto out;
1161 ret = -EFAULT;
Andrea Arcangelia9b85f92015-09-04 15:46:37 -07001162 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001163 goto out;
Andrea Arcangelia9b85f92015-09-04 15:46:37 -07001164 if (uffdio_api.api != UFFD_API || uffdio_api.features) {
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001165 memset(&uffdio_api, 0, sizeof(uffdio_api));
1166 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1167 goto out;
1168 ret = -EINVAL;
1169 goto out;
1170 }
Pavel Emelyanov3f602d22015-09-04 15:46:34 -07001171 uffdio_api.features = UFFD_API_FEATURES;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001172 uffdio_api.ioctls = UFFD_API_IOCTLS;
1173 ret = -EFAULT;
1174 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1175 goto out;
1176 ctx->state = UFFD_STATE_RUNNING;
1177 ret = 0;
1178out:
1179 return ret;
1180}
1181
1182static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1183 unsigned long arg)
1184{
1185 int ret = -EINVAL;
1186 struct userfaultfd_ctx *ctx = file->private_data;
1187
Andrea Arcangelie6485a42015-09-04 15:47:15 -07001188 if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1189 return -EINVAL;
1190
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001191 switch(cmd) {
1192 case UFFDIO_API:
1193 ret = userfaultfd_api(ctx, arg);
1194 break;
1195 case UFFDIO_REGISTER:
1196 ret = userfaultfd_register(ctx, arg);
1197 break;
1198 case UFFDIO_UNREGISTER:
1199 ret = userfaultfd_unregister(ctx, arg);
1200 break;
1201 case UFFDIO_WAKE:
1202 ret = userfaultfd_wake(ctx, arg);
1203 break;
Andrea Arcangeliad465cae2015-09-04 15:47:11 -07001204 case UFFDIO_COPY:
1205 ret = userfaultfd_copy(ctx, arg);
1206 break;
1207 case UFFDIO_ZEROPAGE:
1208 ret = userfaultfd_zeropage(ctx, arg);
1209 break;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001210 }
1211 return ret;
1212}
1213
1214#ifdef CONFIG_PROC_FS
1215static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1216{
1217 struct userfaultfd_ctx *ctx = f->private_data;
1218 wait_queue_t *wq;
1219 struct userfaultfd_wait_queue *uwq;
1220 unsigned long pending = 0, total = 0;
1221
Andrea Arcangeli15b726e2015-09-04 15:46:44 -07001222 spin_lock(&ctx->fault_pending_wqh.lock);
1223 list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001224 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
Andrea Arcangeli15b726e2015-09-04 15:46:44 -07001225 pending++;
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001226 total++;
1227 }
Andrea Arcangeli15b726e2015-09-04 15:46:44 -07001228 list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1229 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1230 total++;
1231 }
1232 spin_unlock(&ctx->fault_pending_wqh.lock);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001233
1234 /*
1235 * If more protocols will be added, there will be all shown
1236 * separated by a space. Like this:
1237 * protocols: aa:... bb:...
1238 */
1239 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
Pavel Emelyanov3f602d22015-09-04 15:46:34 -07001240 pending, total, UFFD_API, UFFD_API_FEATURES,
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001241 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1242}
1243#endif
1244
1245static const struct file_operations userfaultfd_fops = {
1246#ifdef CONFIG_PROC_FS
1247 .show_fdinfo = userfaultfd_show_fdinfo,
1248#endif
1249 .release = userfaultfd_release,
1250 .poll = userfaultfd_poll,
1251 .read = userfaultfd_read,
1252 .unlocked_ioctl = userfaultfd_ioctl,
1253 .compat_ioctl = userfaultfd_ioctl,
1254 .llseek = noop_llseek,
1255};
1256
Andrea Arcangeli3004ec92015-09-04 15:46:48 -07001257static void init_once_userfaultfd_ctx(void *mem)
1258{
1259 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1260
1261 init_waitqueue_head(&ctx->fault_pending_wqh);
1262 init_waitqueue_head(&ctx->fault_wqh);
1263 init_waitqueue_head(&ctx->fd_wqh);
Andrea Arcangeli2c5b7e12015-09-04 15:47:23 -07001264 seqcount_init(&ctx->refile_seq);
Andrea Arcangeli3004ec92015-09-04 15:46:48 -07001265}
1266
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001267/**
1268 * userfaultfd_file_create - Creates an userfaultfd file pointer.
1269 * @flags: Flags for the userfaultfd file.
1270 *
1271 * This function creates an userfaultfd file pointer, w/out installing
1272 * it into the fd table. This is useful when the userfaultfd file is
1273 * used during the initialization of data structures that require
1274 * extra setup after the userfaultfd creation. So the userfaultfd
1275 * creation is split into the file pointer creation phase, and the
1276 * file descriptor installation phase. In this way races with
1277 * userspace closing the newly installed file descriptor can be
1278 * avoided. Returns an userfaultfd file pointer, or a proper error
1279 * pointer.
1280 */
1281static struct file *userfaultfd_file_create(int flags)
1282{
1283 struct file *file;
1284 struct userfaultfd_ctx *ctx;
1285
1286 BUG_ON(!current->mm);
1287
1288 /* Check the UFFD_* constants for consistency. */
1289 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1290 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1291
1292 file = ERR_PTR(-EINVAL);
1293 if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1294 goto out;
1295
1296 file = ERR_PTR(-ENOMEM);
Andrea Arcangeli3004ec92015-09-04 15:46:48 -07001297 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001298 if (!ctx)
1299 goto out;
1300
1301 atomic_set(&ctx->refcount, 1);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001302 ctx->flags = flags;
1303 ctx->state = UFFD_STATE_WAIT_API;
1304 ctx->released = false;
1305 ctx->mm = current->mm;
1306 /* prevent the mm struct to be freed */
Oleg Nesterovd2005e32016-05-20 16:58:36 -07001307 atomic_inc(&ctx->mm->mm_count);
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001308
1309 file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1310 O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
Eric Biggersc03e946f2015-09-17 16:01:54 -07001311 if (IS_ERR(file)) {
Oleg Nesterovd2005e32016-05-20 16:58:36 -07001312 mmdrop(ctx->mm);
Andrea Arcangeli3004ec92015-09-04 15:46:48 -07001313 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
Eric Biggersc03e946f2015-09-17 16:01:54 -07001314 }
Andrea Arcangeli86039bd2015-09-04 15:46:31 -07001315out:
1316 return file;
1317}
1318
1319SYSCALL_DEFINE1(userfaultfd, int, flags)
1320{
1321 int fd, error;
1322 struct file *file;
1323
1324 error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1325 if (error < 0)
1326 return error;
1327 fd = error;
1328
1329 file = userfaultfd_file_create(flags);
1330 if (IS_ERR(file)) {
1331 error = PTR_ERR(file);
1332 goto err_put_unused_fd;
1333 }
1334 fd_install(fd, file);
1335
1336 return fd;
1337
1338err_put_unused_fd:
1339 put_unused_fd(fd);
1340
1341 return error;
1342}
Andrea Arcangeli3004ec92015-09-04 15:46:48 -07001343
1344static int __init userfaultfd_init(void)
1345{
1346 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1347 sizeof(struct userfaultfd_ctx),
1348 0,
1349 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1350 init_once_userfaultfd_ctx);
1351 return 0;
1352}
1353__initcall(userfaultfd_init);