1 // SPDX-License-Identifier: GPL-2.0-only <<
2 /* 1 /*
3 * fs/userfaultfd.c 2 * fs/userfaultfd.c
4 * 3 *
5 * Copyright (C) 2007 Davide Libenzi <davide 4 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
6 * Copyright (C) 2008-2009 Red Hat, Inc. 5 * Copyright (C) 2008-2009 Red Hat, Inc.
7 * Copyright (C) 2015 Red Hat, Inc. 6 * Copyright (C) 2015 Red Hat, Inc.
8 * 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 *
9 * Some part derived from fs/eventfd.c (anon 11 * Some part derived from fs/eventfd.c (anon inode setup) and
10 * mm/ksm.c (mm hashing). 12 * mm/ksm.c (mm hashing).
11 */ 13 */
12 14
13 #include <linux/list.h> <<
14 #include <linux/hashtable.h> 15 #include <linux/hashtable.h>
15 #include <linux/sched/signal.h> !! 16 #include <linux/sched.h>
16 #include <linux/sched/mm.h> <<
17 #include <linux/mm.h> 17 #include <linux/mm.h>
18 #include <linux/mm_inline.h> <<
19 #include <linux/mmu_notifier.h> <<
20 #include <linux/poll.h> 18 #include <linux/poll.h>
21 #include <linux/slab.h> 19 #include <linux/slab.h>
22 #include <linux/seq_file.h> 20 #include <linux/seq_file.h>
23 #include <linux/file.h> 21 #include <linux/file.h>
24 #include <linux/bug.h> 22 #include <linux/bug.h>
25 #include <linux/anon_inodes.h> 23 #include <linux/anon_inodes.h>
26 #include <linux/syscalls.h> 24 #include <linux/syscalls.h>
27 #include <linux/userfaultfd_k.h> 25 #include <linux/userfaultfd_k.h>
28 #include <linux/mempolicy.h> 26 #include <linux/mempolicy.h>
29 #include <linux/ioctl.h> 27 #include <linux/ioctl.h>
30 #include <linux/security.h> 28 #include <linux/security.h>
31 #include <linux/hugetlb.h> <<
32 #include <linux/swapops.h> <<
33 #include <linux/miscdevice.h> <<
34 #include <linux/uio.h> <<
35 <<
36 static int sysctl_unprivileged_userfaultfd __r <<
37 <<
38 #ifdef CONFIG_SYSCTL <<
39 static struct ctl_table vm_userfaultfd_table[] <<
40 { <<
41 .procname = "unprivilege <<
42 .data = &sysctl_unpr <<
43 .maxlen = sizeof(sysct <<
44 .mode = 0644, <<
45 .proc_handler = proc_dointve <<
46 .extra1 = SYSCTL_ZERO, <<
47 .extra2 = SYSCTL_ONE, <<
48 }, <<
49 }; <<
50 #endif <<
51 29
52 static struct kmem_cache *userfaultfd_ctx_cach !! 30 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
53 31
54 struct userfaultfd_fork_ctx { !! 32 enum userfaultfd_state {
55 struct userfaultfd_ctx *orig; !! 33 UFFD_STATE_WAIT_API,
56 struct userfaultfd_ctx *new; !! 34 UFFD_STATE_RUNNING,
57 struct list_head list; <<
58 }; 35 };
59 36
60 struct userfaultfd_unmap_ctx { !! 37 /*
61 struct userfaultfd_ctx *ctx; !! 38 * Start with fault_pending_wqh and fault_wqh so they're more likely
62 unsigned long start; !! 39 * to be in the same cacheline.
63 unsigned long end; !! 40 */
64 struct list_head list; !! 41 struct userfaultfd_ctx {
>> 42 /* waitqueue head for the pending (i.e. not read) userfaults */
>> 43 wait_queue_head_t fault_pending_wqh;
>> 44 /* waitqueue head for the userfaults */
>> 45 wait_queue_head_t fault_wqh;
>> 46 /* waitqueue head for the pseudo fd to wakeup poll/read */
>> 47 wait_queue_head_t fd_wqh;
>> 48 /* a refile sequence protected by fault_pending_wqh lock */
>> 49 struct seqcount refile_seq;
>> 50 /* pseudo fd refcounting */
>> 51 atomic_t refcount;
>> 52 /* 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;
65 }; 60 };
66 61
67 struct userfaultfd_wait_queue { 62 struct userfaultfd_wait_queue {
68 struct uffd_msg msg; 63 struct uffd_msg msg;
69 wait_queue_entry_t wq; !! 64 wait_queue_t wq;
70 struct userfaultfd_ctx *ctx; 65 struct userfaultfd_ctx *ctx;
71 bool waken; 66 bool waken;
72 }; 67 };
73 68
74 struct userfaultfd_wake_range { 69 struct userfaultfd_wake_range {
75 unsigned long start; 70 unsigned long start;
76 unsigned long len; 71 unsigned long len;
77 }; 72 };
78 73
79 /* internal indication that UFFD_API ioctl was !! 74 static int userfaultfd_wake_function(wait_queue_t *wq, unsigned mode,
80 #define UFFD_FEATURE_INITIALIZED <<
81 <<
82 static bool userfaultfd_is_initialized(struct <<
83 { <<
84 return ctx->features & UFFD_FEATURE_IN <<
85 } <<
86 <<
87 static bool userfaultfd_wp_async_ctx(struct us <<
88 { <<
89 return ctx && (ctx->features & UFFD_FE <<
90 } <<
91 <<
92 /* <<
93 * Whether WP_UNPOPULATED is enabled on the uf <<
94 * meaningful when userfaultfd_wp()==true on t <<
95 * anonymous. <<
96 */ <<
97 bool userfaultfd_wp_unpopulated(struct vm_area <<
98 { <<
99 struct userfaultfd_ctx *ctx = vma->vm_ <<
100 <<
101 if (!ctx) <<
102 return false; <<
103 <<
104 return ctx->features & UFFD_FEATURE_WP <<
105 } <<
106 <<
107 static void userfaultfd_set_vm_flags(struct vm <<
108 vm_flags_ <<
109 { <<
110 const bool uffd_wp_changed = (vma->vm_ <<
111 <<
112 vm_flags_reset(vma, flags); <<
113 /* <<
114 * For shared mappings, we want to ena <<
115 * userfaultfd-wp is enabled (see vma_ <<
116 * recalculate vma->vm_page_prot whene <<
117 */ <<
118 if ((vma->vm_flags & VM_SHARED) && uff <<
119 vma_set_page_prot(vma); <<
120 } <<
121 <<
122 static int userfaultfd_wake_function(wait_queu <<
123 int wake_ 75 int wake_flags, void *key)
124 { 76 {
125 struct userfaultfd_wake_range *range = 77 struct userfaultfd_wake_range *range = key;
126 int ret; 78 int ret;
127 struct userfaultfd_wait_queue *uwq; 79 struct userfaultfd_wait_queue *uwq;
128 unsigned long start, len; 80 unsigned long start, len;
129 81
130 uwq = container_of(wq, struct userfaul 82 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
131 ret = 0; 83 ret = 0;
132 /* len == 0 means wake all */ 84 /* len == 0 means wake all */
133 start = range->start; 85 start = range->start;
134 len = range->len; 86 len = range->len;
135 if (len && (start > uwq->msg.arg.pagef 87 if (len && (start > uwq->msg.arg.pagefault.address ||
136 start + len <= uwq->msg.ar 88 start + len <= uwq->msg.arg.pagefault.address))
137 goto out; 89 goto out;
138 WRITE_ONCE(uwq->waken, true); 90 WRITE_ONCE(uwq->waken, true);
139 /* 91 /*
140 * The Program-Order guarantees provid !! 92 * The implicit smp_mb__before_spinlock in try_to_wake_up()
141 * ensure uwq->waken is visible before !! 93 * renders uwq->waken visible to other CPUs before the task is
>> 94 * waken.
142 */ 95 */
143 ret = wake_up_state(wq->private, mode) 96 ret = wake_up_state(wq->private, mode);
144 if (ret) { !! 97 if (ret)
145 /* 98 /*
146 * Wake only once, autoremove 99 * Wake only once, autoremove behavior.
147 * 100 *
148 * After the effect of list_de !! 101 * After the effect of list_del_init is visible to the
149 * CPUs, the waitqueue may dis !! 102 * other CPUs, the waitqueue may disappear from under
150 * !list_empty_careful() in ha !! 103 * us, see the !list_empty_careful() in
151 * !! 104 * handle_userfault(). try_to_wake_up() has an
152 * try_to_wake_up() has an imp !! 105 * implicit smp_mb__before_spinlock, and the
153 * wq->private is read before !! 106 * wq->private is read before calling the extern
154 * "wake_up_state" (which in t !! 107 * function "wake_up_state" (which in turns calls
>> 108 * try_to_wake_up). While the spin_lock;spin_unlock;
>> 109 * wouldn't be enough, the smp_mb__before_spinlock is
>> 110 * enough to avoid an explicit smp_mb() here.
155 */ 111 */
156 list_del_init(&wq->entry); !! 112 list_del_init(&wq->task_list);
157 } <<
158 out: 113 out:
159 return ret; 114 return ret;
160 } 115 }
161 116
162 /** 117 /**
163 * userfaultfd_ctx_get - Acquires a reference 118 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
164 * context. 119 * context.
165 * @ctx: [in] Pointer to the userfaultfd conte 120 * @ctx: [in] Pointer to the userfaultfd context.
>> 121 *
>> 122 * Returns: In case of success, returns not zero.
166 */ 123 */
167 static void userfaultfd_ctx_get(struct userfau 124 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
168 { 125 {
169 refcount_inc(&ctx->refcount); !! 126 if (!atomic_inc_not_zero(&ctx->refcount))
>> 127 BUG();
170 } 128 }
171 129
172 /** 130 /**
173 * userfaultfd_ctx_put - Releases a reference 131 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
174 * context. 132 * context.
175 * @ctx: [in] Pointer to userfaultfd context. 133 * @ctx: [in] Pointer to userfaultfd context.
176 * 134 *
177 * The userfaultfd context reference must have 135 * The userfaultfd context reference must have been previously acquired either
178 * with userfaultfd_ctx_get() or userfaultfd_c 136 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
179 */ 137 */
180 static void userfaultfd_ctx_put(struct userfau 138 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
181 { 139 {
182 if (refcount_dec_and_test(&ctx->refcou !! 140 if (atomic_dec_and_test(&ctx->refcount)) {
183 VM_BUG_ON(spin_is_locked(&ctx- 141 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
184 VM_BUG_ON(waitqueue_active(&ct 142 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
185 VM_BUG_ON(spin_is_locked(&ctx- 143 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
186 VM_BUG_ON(waitqueue_active(&ct 144 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
187 VM_BUG_ON(spin_is_locked(&ctx- <<
188 VM_BUG_ON(waitqueue_active(&ct <<
189 VM_BUG_ON(spin_is_locked(&ctx- 145 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
190 VM_BUG_ON(waitqueue_active(&ct 146 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
191 mmdrop(ctx->mm); 147 mmdrop(ctx->mm);
192 kmem_cache_free(userfaultfd_ct 148 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
193 } 149 }
194 } 150 }
195 151
196 static inline void msg_init(struct uffd_msg *m 152 static inline void msg_init(struct uffd_msg *msg)
197 { 153 {
198 BUILD_BUG_ON(sizeof(struct uffd_msg) ! 154 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
199 /* 155 /*
200 * Must use memset to zero out the pad 156 * Must use memset to zero out the paddings or kernel data is
201 * leaked to userland. 157 * leaked to userland.
202 */ 158 */
203 memset(msg, 0, sizeof(struct uffd_msg) 159 memset(msg, 0, sizeof(struct uffd_msg));
204 } 160 }
205 161
206 static inline struct uffd_msg userfault_msg(un 162 static inline struct uffd_msg userfault_msg(unsigned long address,
207 un <<
208 un 163 unsigned int flags,
209 un !! 164 unsigned long reason)
210 un <<
211 { 165 {
212 struct uffd_msg msg; 166 struct uffd_msg msg;
213 <<
214 msg_init(&msg); 167 msg_init(&msg);
215 msg.event = UFFD_EVENT_PAGEFAULT; 168 msg.event = UFFD_EVENT_PAGEFAULT;
216 !! 169 msg.arg.pagefault.address = address;
217 msg.arg.pagefault.address = (features <<
218 real_addre <<
219 <<
220 /* <<
221 * These flags indicate why the userfa <<
222 * - UFFD_PAGEFAULT_FLAG_WP indicates <<
223 * - UFFD_PAGEFAULT_FLAG_MINOR indicat <<
224 * - Neither of these flags being set <<
225 * <<
226 * Separately, UFFD_PAGEFAULT_FLAG_WRI <<
227 * fault. Otherwise, it was a read fau <<
228 */ <<
229 if (flags & FAULT_FLAG_WRITE) 170 if (flags & FAULT_FLAG_WRITE)
>> 171 /*
>> 172 * If UFFD_FEATURE_PAGEFAULT_FLAG_WRITE was set in the
>> 173 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
>> 174 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
>> 175 * was a read fault, otherwise if set it means it's
>> 176 * a write fault.
>> 177 */
230 msg.arg.pagefault.flags |= UFF 178 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
231 if (reason & VM_UFFD_WP) 179 if (reason & VM_UFFD_WP)
>> 180 /*
>> 181 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
>> 182 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
>> 183 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
>> 184 * a missing fault, otherwise if set it means it's a
>> 185 * write protect fault.
>> 186 */
232 msg.arg.pagefault.flags |= UFF 187 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
233 if (reason & VM_UFFD_MINOR) <<
234 msg.arg.pagefault.flags |= UFF <<
235 if (features & UFFD_FEATURE_THREAD_ID) <<
236 msg.arg.pagefault.feat.ptid = <<
237 return msg; 188 return msg;
238 } 189 }
239 190
240 #ifdef CONFIG_HUGETLB_PAGE <<
241 /* <<
242 * Same functionality as userfaultfd_must_wait <<
243 * hugepmd ranges. <<
244 */ <<
245 static inline bool userfaultfd_huge_must_wait( <<
246 <<
247 <<
248 { <<
249 struct vm_area_struct *vma = vmf->vma; <<
250 pte_t *ptep, pte; <<
251 bool ret = true; <<
252 <<
253 assert_fault_locked(vmf); <<
254 <<
255 ptep = hugetlb_walk(vma, vmf->address, <<
256 if (!ptep) <<
257 goto out; <<
258 <<
259 ret = false; <<
260 pte = huge_ptep_get(vma->vm_mm, vmf->a <<
261 <<
262 /* <<
263 * Lockless access: we're in a wait_ev <<
264 * changes under us. PTE markers shou <<
265 * ptes here. <<
266 */ <<
267 if (huge_pte_none_mostly(pte)) <<
268 ret = true; <<
269 if (!huge_pte_write(pte) && (reason & <<
270 ret = true; <<
271 out: <<
272 return ret; <<
273 } <<
274 #else <<
275 static inline bool userfaultfd_huge_must_wait( <<
276 <<
277 <<
278 { <<
279 return false; /* should never get he <<
280 } <<
281 #endif /* CONFIG_HUGETLB_PAGE */ <<
282 <<
283 /* 191 /*
284 * Verify the pagetables are still not ok afte 192 * Verify the pagetables are still not ok after having reigstered into
285 * the fault_pending_wqh to avoid userland hav 193 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
286 * userfault that has already been resolved, i !! 194 * userfault that has already been resolved, if userfaultfd_read and
287 * UFFDIO_COPY|ZEROPAGE are being run simultan 195 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
288 * threads. 196 * threads.
289 */ 197 */
290 static inline bool userfaultfd_must_wait(struc 198 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
291 struc !! 199 unsigned long address,
>> 200 unsigned long flags,
292 unsig 201 unsigned long reason)
293 { 202 {
294 struct mm_struct *mm = ctx->mm; 203 struct mm_struct *mm = ctx->mm;
295 unsigned long address = vmf->address; <<
296 pgd_t *pgd; 204 pgd_t *pgd;
297 p4d_t *p4d; <<
298 pud_t *pud; 205 pud_t *pud;
299 pmd_t *pmd, _pmd; 206 pmd_t *pmd, _pmd;
300 pte_t *pte; 207 pte_t *pte;
301 pte_t ptent; <<
302 bool ret = true; 208 bool ret = true;
303 209
304 assert_fault_locked(vmf); !! 210 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
305 211
306 pgd = pgd_offset(mm, address); 212 pgd = pgd_offset(mm, address);
307 if (!pgd_present(*pgd)) 213 if (!pgd_present(*pgd))
308 goto out; 214 goto out;
309 p4d = p4d_offset(pgd, address); !! 215 pud = pud_offset(pgd, address);
310 if (!p4d_present(*p4d)) <<
311 goto out; <<
312 pud = pud_offset(p4d, address); <<
313 if (!pud_present(*pud)) 216 if (!pud_present(*pud))
314 goto out; 217 goto out;
315 pmd = pmd_offset(pud, address); 218 pmd = pmd_offset(pud, address);
316 again: !! 219 /*
317 _pmd = pmdp_get_lockless(pmd); !! 220 * READ_ONCE must function as a barrier with narrower scope
318 if (pmd_none(_pmd)) !! 221 * and it must be equivalent to:
>> 222 * _pmd = *pmd; barrier();
>> 223 *
>> 224 * This is to deal with the instability (as in
>> 225 * pmd_trans_unstable) of the pmd.
>> 226 */
>> 227 _pmd = READ_ONCE(*pmd);
>> 228 if (!pmd_present(_pmd))
319 goto out; 229 goto out;
320 230
321 ret = false; 231 ret = false;
322 if (!pmd_present(_pmd) || pmd_devmap(_ !! 232 if (pmd_trans_huge(_pmd))
323 goto out; 233 goto out;
324 234
325 if (pmd_trans_huge(_pmd)) { !! 235 /*
326 if (!pmd_write(_pmd) && (reaso !! 236 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
327 ret = true; !! 237 * and use the standard pte_offset_map() instead of parsing _pmd.
328 goto out; !! 238 */
329 } <<
330 <<
331 pte = pte_offset_map(pmd, address); 239 pte = pte_offset_map(pmd, address);
332 if (!pte) { <<
333 ret = true; <<
334 goto again; <<
335 } <<
336 /* 240 /*
337 * Lockless access: we're in a wait_ev 241 * Lockless access: we're in a wait_event so it's ok if it
338 * changes under us. PTE markers shou !! 242 * changes under us.
339 * ptes here. <<
340 */ 243 */
341 ptent = ptep_get(pte); !! 244 if (pte_none(*pte))
342 if (pte_none_mostly(ptent)) <<
343 ret = true; <<
344 if (!pte_write(ptent) && (reason & VM_ <<
345 ret = true; 245 ret = true;
346 pte_unmap(pte); 246 pte_unmap(pte);
347 247
348 out: 248 out:
349 return ret; 249 return ret;
350 } 250 }
351 251
352 static inline unsigned int userfaultfd_get_blo <<
353 { <<
354 if (flags & FAULT_FLAG_INTERRUPTIBLE) <<
355 return TASK_INTERRUPTIBLE; <<
356 <<
357 if (flags & FAULT_FLAG_KILLABLE) <<
358 return TASK_KILLABLE; <<
359 <<
360 return TASK_UNINTERRUPTIBLE; <<
361 } <<
362 <<
363 /* 252 /*
364 * The locking rules involved in returning VM_ 253 * The locking rules involved in returning VM_FAULT_RETRY depending on
365 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NO 254 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
366 * FAULT_FLAG_KILLABLE are not straightforward 255 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
367 * recommendation in __lock_page_or_retry is n 256 * recommendation in __lock_page_or_retry is not an understatement.
368 * 257 *
369 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_ !! 258 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
370 * before returning VM_FAULT_RETRY only if FAU 259 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
371 * not set. 260 * not set.
372 * 261 *
373 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_ 262 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
374 * set, VM_FAULT_RETRY can still be returned i 263 * set, VM_FAULT_RETRY can still be returned if and only if there are
375 * fatal_signal_pending()s, and the mmap_lock !! 264 * fatal_signal_pending()s, and the mmap_sem must be released before
376 * returning it. 265 * returning it.
377 */ 266 */
378 vm_fault_t handle_userfault(struct vm_fault *v !! 267 int handle_userfault(struct fault_env *fe, unsigned long reason)
379 { 268 {
380 struct vm_area_struct *vma = vmf->vma; !! 269 struct mm_struct *mm = fe->vma->vm_mm;
381 struct mm_struct *mm = vma->vm_mm; <<
382 struct userfaultfd_ctx *ctx; 270 struct userfaultfd_ctx *ctx;
383 struct userfaultfd_wait_queue uwq; 271 struct userfaultfd_wait_queue uwq;
384 vm_fault_t ret = VM_FAULT_SIGBUS; !! 272 int ret;
385 bool must_wait; !! 273 bool must_wait, return_to_userland;
386 unsigned int blocking_state; !! 274 long blocking_state;
387 <<
388 /* <<
389 * We don't do userfault handling for <<
390 * <<
391 * We also don't do userfault handling <<
392 * coredumping. hugetlbfs has the spec <<
393 * hugetlb_follow_page_mask() to skip <<
394 * FOLL_DUMP case, anon memory also ch <<
395 * the no_page_table() helper in follo <<
396 * shmem_vm_ops->fault method is invok <<
397 * coredumping and it ends up here. <<
398 */ <<
399 if (current->flags & (PF_EXITING|PF_DU <<
400 goto out; <<
401 275
402 assert_fault_locked(vmf); !! 276 BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
403 277
404 ctx = vma->vm_userfaultfd_ctx.ctx; !! 278 ret = VM_FAULT_SIGBUS;
>> 279 ctx = fe->vma->vm_userfaultfd_ctx.ctx;
405 if (!ctx) 280 if (!ctx)
406 goto out; 281 goto out;
407 282
408 BUG_ON(ctx->mm != mm); 283 BUG_ON(ctx->mm != mm);
409 284
410 /* Any unrecognized flag is a bug. */ !! 285 VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
411 VM_BUG_ON(reason & ~__VM_UFFD_FLAGS); !! 286 VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
412 /* 0 or > 1 flags set is a bug; we exp <<
413 VM_BUG_ON(!reason || (reason & (reason <<
414 <<
415 if (ctx->features & UFFD_FEATURE_SIGBU <<
416 goto out; <<
417 if (!(vmf->flags & FAULT_FLAG_USER) && <<
418 goto out; <<
419 287
420 /* 288 /*
421 * If it's already released don't get 289 * If it's already released don't get it. This avoids to loop
422 * in __get_user_pages if userfaultfd_ 290 * in __get_user_pages if userfaultfd_release waits on the
423 * caller of handle_userfault to relea !! 291 * caller of handle_userfault to release the mmap_sem.
424 */ 292 */
425 if (unlikely(READ_ONCE(ctx->released)) !! 293 if (unlikely(ACCESS_ONCE(ctx->released)))
426 /* !! 294 goto out;
427 * Don't return VM_FAULT_SIGBU !! 295
428 * cooperative manager can clo !! 296 /*
429 * last UFFDIO_COPY, without r !! 297 * We don't do userfault handling for the final child pid update.
430 * involuntary SIGBUS if the p !! 298 */
431 * userfaultfd while the userf !! 299 if (current->flags & PF_EXITING)
432 * (but after the last UFFDIO_ <<
433 * wasn't already closed when <<
434 * this point, that would norm <<
435 * userfaultfd_must_wait retur <<
436 * <<
437 * If we were to return VM_FAU <<
438 * cooperative manager would b <<
439 * always call UFFDIO_UNREGIST <<
440 * close the uffd. <<
441 */ <<
442 ret = VM_FAULT_NOPAGE; <<
443 goto out; 300 goto out;
444 } <<
445 301
446 /* 302 /*
447 * Check that we can return VM_FAULT_R 303 * Check that we can return VM_FAULT_RETRY.
448 * 304 *
449 * NOTE: it should become possible to 305 * NOTE: it should become possible to return VM_FAULT_RETRY
450 * even if FAULT_FLAG_TRIED is set wit 306 * even if FAULT_FLAG_TRIED is set without leading to gup()
451 * -EBUSY failures, if the userfaultfd 307 * -EBUSY failures, if the userfaultfd is to be extended for
452 * VM_UFFD_WP tracking and we intend t 308 * VM_UFFD_WP tracking and we intend to arm the userfault
453 * without first stopping userland acc 309 * without first stopping userland access to the memory. For
454 * VM_UFFD_MISSING userfaults this is 310 * VM_UFFD_MISSING userfaults this is enough for now.
455 */ 311 */
456 if (unlikely(!(vmf->flags & FAULT_FLAG !! 312 if (unlikely(!(fe->flags & FAULT_FLAG_ALLOW_RETRY))) {
457 /* 313 /*
458 * Validate the invariant that 314 * Validate the invariant that nowait must allow retry
459 * to be sure not to return SI 315 * to be sure not to return SIGBUS erroneously on
460 * nowait invocations. 316 * nowait invocations.
461 */ 317 */
462 BUG_ON(vmf->flags & FAULT_FLAG !! 318 BUG_ON(fe->flags & FAULT_FLAG_RETRY_NOWAIT);
463 #ifdef CONFIG_DEBUG_VM 319 #ifdef CONFIG_DEBUG_VM
464 if (printk_ratelimit()) { 320 if (printk_ratelimit()) {
465 printk(KERN_WARNING 321 printk(KERN_WARNING
466 "FAULT_FLAG_ALL !! 322 "FAULT_FLAG_ALLOW_RETRY missing %x\n", fe->flags);
467 vmf->flags); <<
468 dump_stack(); 323 dump_stack();
469 } 324 }
470 #endif 325 #endif
471 goto out; 326 goto out;
472 } 327 }
473 328
474 /* 329 /*
475 * Handle nowait, not much to do other 330 * Handle nowait, not much to do other than tell it to retry
476 * and wait. 331 * and wait.
477 */ 332 */
478 ret = VM_FAULT_RETRY; 333 ret = VM_FAULT_RETRY;
479 if (vmf->flags & FAULT_FLAG_RETRY_NOWA !! 334 if (fe->flags & FAULT_FLAG_RETRY_NOWAIT)
480 goto out; 335 goto out;
481 336
482 /* take the reference before dropping !! 337 /* take the reference before dropping the mmap_sem */
483 userfaultfd_ctx_get(ctx); 338 userfaultfd_ctx_get(ctx);
484 339
485 init_waitqueue_func_entry(&uwq.wq, use 340 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
486 uwq.wq.private = current; 341 uwq.wq.private = current;
487 uwq.msg = userfault_msg(vmf->address, !! 342 uwq.msg = userfault_msg(fe->address, fe->flags, reason);
488 reason, ctx->f <<
489 uwq.ctx = ctx; 343 uwq.ctx = ctx;
490 uwq.waken = false; 344 uwq.waken = false;
491 345
492 blocking_state = userfaultfd_get_block !! 346 return_to_userland =
493 !! 347 (fe->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
494 /* !! 348 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
495 * Take the vma lock now, in order to !! 349 blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
496 * userfaultfd_huge_must_wait() later. !! 350 TASK_KILLABLE;
497 * (sleepable) vma lock can modify the <<
498 * must be before explicitly calling s <<
499 */ <<
500 if (is_vm_hugetlb_page(vma)) <<
501 hugetlb_vma_lock_read(vma); <<
502 351
503 spin_lock_irq(&ctx->fault_pending_wqh. !! 352 spin_lock(&ctx->fault_pending_wqh.lock);
504 /* 353 /*
505 * After the __add_wait_queue the uwq 354 * After the __add_wait_queue the uwq is visible to userland
506 * through poll/read(). 355 * through poll/read().
507 */ 356 */
508 __add_wait_queue(&ctx->fault_pending_w 357 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
509 /* 358 /*
510 * The smp_mb() after __set_current_st 359 * The smp_mb() after __set_current_state prevents the reads
511 * following the spin_unlock to happen 360 * following the spin_unlock to happen before the list_add in
512 * __add_wait_queue. 361 * __add_wait_queue.
513 */ 362 */
514 set_current_state(blocking_state); 363 set_current_state(blocking_state);
515 spin_unlock_irq(&ctx->fault_pending_wq !! 364 spin_unlock(&ctx->fault_pending_wqh.lock);
516 365
517 if (!is_vm_hugetlb_page(vma)) !! 366 must_wait = userfaultfd_must_wait(ctx, fe->address, fe->flags, reason);
518 must_wait = userfaultfd_must_w !! 367 up_read(&mm->mmap_sem);
519 else <<
520 must_wait = userfaultfd_huge_m <<
521 if (is_vm_hugetlb_page(vma)) <<
522 hugetlb_vma_unlock_read(vma); <<
523 release_fault_lock(vmf); <<
524 368
525 if (likely(must_wait && !READ_ONCE(ctx !! 369 if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
526 wake_up_poll(&ctx->fd_wqh, EPO !! 370 (return_to_userland ? !signal_pending(current) :
>> 371 !fatal_signal_pending(current)))) {
>> 372 wake_up_poll(&ctx->fd_wqh, POLLIN);
527 schedule(); 373 schedule();
>> 374 ret |= VM_FAULT_MAJOR;
>> 375
>> 376 /*
>> 377 * False wakeups can orginate even from rwsem before
>> 378 * up_read() however userfaults will wait either for a
>> 379 * targeted wakeup on the specific uwq waitqueue from
>> 380 * wake_userfault() or for signals or for uffd
>> 381 * release.
>> 382 */
>> 383 while (!READ_ONCE(uwq.waken)) {
>> 384 /*
>> 385 * This needs the full smp_store_mb()
>> 386 * guarantee as the state write must be
>> 387 * visible to other CPUs before reading
>> 388 * uwq.waken from other CPUs.
>> 389 */
>> 390 set_current_state(blocking_state);
>> 391 if (READ_ONCE(uwq.waken) ||
>> 392 READ_ONCE(ctx->released) ||
>> 393 (return_to_userland ? signal_pending(current) :
>> 394 fatal_signal_pending(current)))
>> 395 break;
>> 396 schedule();
>> 397 }
528 } 398 }
529 399
530 __set_current_state(TASK_RUNNING); 400 __set_current_state(TASK_RUNNING);
531 401
>> 402 if (return_to_userland) {
>> 403 if (signal_pending(current) &&
>> 404 !fatal_signal_pending(current)) {
>> 405 /*
>> 406 * If we got a SIGSTOP or SIGCONT and this is
>> 407 * a normal userland page fault, just let
>> 408 * userland return so the signal will be
>> 409 * handled and gdb debugging works. The page
>> 410 * fault code immediately after we return from
>> 411 * this function is going to release the
>> 412 * mmap_sem and it's not depending on it
>> 413 * (unlike gup would if we were not to return
>> 414 * VM_FAULT_RETRY).
>> 415 *
>> 416 * If a fatal signal is pending we still take
>> 417 * the streamlined VM_FAULT_RETRY failure path
>> 418 * and there's no need to retake the mmap_sem
>> 419 * in such case.
>> 420 */
>> 421 down_read(&mm->mmap_sem);
>> 422 ret = VM_FAULT_NOPAGE;
>> 423 }
>> 424 }
>> 425
532 /* 426 /*
533 * Here we race with the list_del; lis 427 * Here we race with the list_del; list_add in
534 * userfaultfd_ctx_read(), however bec 428 * userfaultfd_ctx_read(), however because we don't ever run
535 * list_del_init() to refile across th 429 * list_del_init() to refile across the two lists, the prev
536 * and next pointers will never point 430 * and next pointers will never point to self. list_add also
537 * would never let any of the two poin 431 * would never let any of the two pointers to point to
538 * self. So list_empty_careful won't r 432 * self. So list_empty_careful won't risk to see both pointers
539 * pointing to self at any time during 433 * pointing to self at any time during the list refile. The
540 * only case where list_del_init() is 434 * only case where list_del_init() is called is the full
541 * removal in the wake function and th 435 * removal in the wake function and there we don't re-list_add
542 * and it's fine not to block on the s 436 * and it's fine not to block on the spinlock. The uwq on this
543 * kernel stack can be released after 437 * kernel stack can be released after the list_del_init.
544 */ 438 */
545 if (!list_empty_careful(&uwq.wq.entry) !! 439 if (!list_empty_careful(&uwq.wq.task_list)) {
546 spin_lock_irq(&ctx->fault_pend !! 440 spin_lock(&ctx->fault_pending_wqh.lock);
547 /* 441 /*
548 * No need of list_del_init(), 442 * No need of list_del_init(), the uwq on the stack
549 * will be freed shortly anywa 443 * will be freed shortly anyway.
550 */ 444 */
551 list_del(&uwq.wq.entry); !! 445 list_del(&uwq.wq.task_list);
552 spin_unlock_irq(&ctx->fault_pe !! 446 spin_unlock(&ctx->fault_pending_wqh.lock);
553 } 447 }
554 448
555 /* 449 /*
556 * ctx may go away after this if the u 450 * ctx may go away after this if the userfault pseudo fd is
557 * already released. 451 * already released.
558 */ 452 */
559 userfaultfd_ctx_put(ctx); 453 userfaultfd_ctx_put(ctx);
560 454
561 out: 455 out:
562 return ret; 456 return ret;
563 } 457 }
564 458
565 static void userfaultfd_event_wait_completion( <<
566 <<
567 { <<
568 struct userfaultfd_ctx *release_new_ct <<
569 <<
570 if (WARN_ON_ONCE(current->flags & PF_E <<
571 goto out; <<
572 <<
573 ewq->ctx = ctx; <<
574 init_waitqueue_entry(&ewq->wq, current <<
575 release_new_ctx = NULL; <<
576 <<
577 spin_lock_irq(&ctx->event_wqh.lock); <<
578 /* <<
579 * After the __add_wait_queue the uwq <<
580 * through poll/read(). <<
581 */ <<
582 __add_wait_queue(&ctx->event_wqh, &ewq <<
583 for (;;) { <<
584 set_current_state(TASK_KILLABL <<
585 if (ewq->msg.event == 0) <<
586 break; <<
587 if (READ_ONCE(ctx->released) | <<
588 fatal_signal_pending(curre <<
589 /* <<
590 * &ewq->wq may be que <<
591 * __remove_wait_queue <<
592 * parameter. It would <<
593 * didn't. <<
594 */ <<
595 __remove_wait_queue(&c <<
596 if (ewq->msg.event == <<
597 struct userfau <<
598 <<
599 new = (struct <<
600 (unsig <<
601 ewq->m <<
602 release_new_ct <<
603 } <<
604 break; <<
605 } <<
606 <<
607 spin_unlock_irq(&ctx->event_wq <<
608 <<
609 wake_up_poll(&ctx->fd_wqh, EPO <<
610 schedule(); <<
611 <<
612 spin_lock_irq(&ctx->event_wqh. <<
613 } <<
614 __set_current_state(TASK_RUNNING); <<
615 spin_unlock_irq(&ctx->event_wqh.lock); <<
616 <<
617 if (release_new_ctx) { <<
618 struct vm_area_struct *vma; <<
619 struct mm_struct *mm = release <<
620 VMA_ITERATOR(vmi, mm, 0); <<
621 <<
622 /* the various vma->vm_userfau <<
623 mmap_write_lock(mm); <<
624 for_each_vma(vmi, vma) { <<
625 if (vma->vm_userfaultf <<
626 vma_start_writ <<
627 vma->vm_userfa <<
628 userfaultfd_se <<
629 <<
630 } <<
631 } <<
632 mmap_write_unlock(mm); <<
633 <<
634 userfaultfd_ctx_put(release_ne <<
635 } <<
636 <<
637 /* <<
638 * ctx may go away after this if the u <<
639 * already released. <<
640 */ <<
641 out: <<
642 atomic_dec(&ctx->mmap_changing); <<
643 VM_BUG_ON(atomic_read(&ctx->mmap_chang <<
644 userfaultfd_ctx_put(ctx); <<
645 } <<
646 <<
647 static void userfaultfd_event_complete(struct <<
648 struct <<
649 { <<
650 ewq->msg.event = 0; <<
651 wake_up_locked(&ctx->event_wqh); <<
652 __remove_wait_queue(&ctx->event_wqh, & <<
653 } <<
654 <<
655 int dup_userfaultfd(struct vm_area_struct *vma <<
656 { <<
657 struct userfaultfd_ctx *ctx = NULL, *o <<
658 struct userfaultfd_fork_ctx *fctx; <<
659 <<
660 octx = vma->vm_userfaultfd_ctx.ctx; <<
661 if (!octx) <<
662 return 0; <<
663 <<
664 if (!(octx->features & UFFD_FEATURE_EV <<
665 vma_start_write(vma); <<
666 vma->vm_userfaultfd_ctx = NULL <<
667 userfaultfd_set_vm_flags(vma, <<
668 return 0; <<
669 } <<
670 <<
671 list_for_each_entry(fctx, fcs, list) <<
672 if (fctx->orig == octx) { <<
673 ctx = fctx->new; <<
674 break; <<
675 } <<
676 <<
677 if (!ctx) { <<
678 fctx = kmalloc(sizeof(*fctx), <<
679 if (!fctx) <<
680 return -ENOMEM; <<
681 <<
682 ctx = kmem_cache_alloc(userfau <<
683 if (!ctx) { <<
684 kfree(fctx); <<
685 return -ENOMEM; <<
686 } <<
687 <<
688 refcount_set(&ctx->refcount, 1 <<
689 ctx->flags = octx->flags; <<
690 ctx->features = octx->features <<
691 ctx->released = false; <<
692 init_rwsem(&ctx->map_changing_ <<
693 atomic_set(&ctx->mmap_changing <<
694 ctx->mm = vma->vm_mm; <<
695 mmgrab(ctx->mm); <<
696 <<
697 userfaultfd_ctx_get(octx); <<
698 down_write(&octx->map_changing <<
699 atomic_inc(&octx->mmap_changin <<
700 up_write(&octx->map_changing_l <<
701 fctx->orig = octx; <<
702 fctx->new = ctx; <<
703 list_add_tail(&fctx->list, fcs <<
704 } <<
705 <<
706 vma->vm_userfaultfd_ctx.ctx = ctx; <<
707 return 0; <<
708 } <<
709 <<
710 static void dup_fctx(struct userfaultfd_fork_c <<
711 { <<
712 struct userfaultfd_ctx *ctx = fctx->or <<
713 struct userfaultfd_wait_queue ewq; <<
714 <<
715 msg_init(&ewq.msg); <<
716 <<
717 ewq.msg.event = UFFD_EVENT_FORK; <<
718 ewq.msg.arg.reserved.reserved1 = (unsi <<
719 <<
720 userfaultfd_event_wait_completion(ctx, <<
721 } <<
722 <<
723 void dup_userfaultfd_complete(struct list_head <<
724 { <<
725 struct userfaultfd_fork_ctx *fctx, *n; <<
726 <<
727 list_for_each_entry_safe(fctx, n, fcs, <<
728 dup_fctx(fctx); <<
729 list_del(&fctx->list); <<
730 kfree(fctx); <<
731 } <<
732 } <<
733 <<
734 void mremap_userfaultfd_prep(struct vm_area_st <<
735 struct vm_userfau <<
736 { <<
737 struct userfaultfd_ctx *ctx; <<
738 <<
739 ctx = vma->vm_userfaultfd_ctx.ctx; <<
740 <<
741 if (!ctx) <<
742 return; <<
743 <<
744 if (ctx->features & UFFD_FEATURE_EVENT <<
745 vm_ctx->ctx = ctx; <<
746 userfaultfd_ctx_get(ctx); <<
747 down_write(&ctx->map_changing_ <<
748 atomic_inc(&ctx->mmap_changing <<
749 up_write(&ctx->map_changing_lo <<
750 } else { <<
751 /* Drop uffd context if remap <<
752 vma_start_write(vma); <<
753 vma->vm_userfaultfd_ctx = NULL <<
754 userfaultfd_set_vm_flags(vma, <<
755 } <<
756 } <<
757 <<
758 void mremap_userfaultfd_complete(struct vm_use <<
759 unsigned long <<
760 unsigned long <<
761 { <<
762 struct userfaultfd_ctx *ctx = vm_ctx-> <<
763 struct userfaultfd_wait_queue ewq; <<
764 <<
765 if (!ctx) <<
766 return; <<
767 <<
768 if (to & ~PAGE_MASK) { <<
769 userfaultfd_ctx_put(ctx); <<
770 return; <<
771 } <<
772 <<
773 msg_init(&ewq.msg); <<
774 <<
775 ewq.msg.event = UFFD_EVENT_REMAP; <<
776 ewq.msg.arg.remap.from = from; <<
777 ewq.msg.arg.remap.to = to; <<
778 ewq.msg.arg.remap.len = len; <<
779 <<
780 userfaultfd_event_wait_completion(ctx, <<
781 } <<
782 <<
783 bool userfaultfd_remove(struct vm_area_struct <<
784 unsigned long start, u <<
785 { <<
786 struct mm_struct *mm = vma->vm_mm; <<
787 struct userfaultfd_ctx *ctx; <<
788 struct userfaultfd_wait_queue ewq; <<
789 <<
790 ctx = vma->vm_userfaultfd_ctx.ctx; <<
791 if (!ctx || !(ctx->features & UFFD_FEA <<
792 return true; <<
793 <<
794 userfaultfd_ctx_get(ctx); <<
795 down_write(&ctx->map_changing_lock); <<
796 atomic_inc(&ctx->mmap_changing); <<
797 up_write(&ctx->map_changing_lock); <<
798 mmap_read_unlock(mm); <<
799 <<
800 msg_init(&ewq.msg); <<
801 <<
802 ewq.msg.event = UFFD_EVENT_REMOVE; <<
803 ewq.msg.arg.remove.start = start; <<
804 ewq.msg.arg.remove.end = end; <<
805 <<
806 userfaultfd_event_wait_completion(ctx, <<
807 <<
808 return false; <<
809 } <<
810 <<
811 static bool has_unmap_ctx(struct userfaultfd_c <<
812 unsigned long start, <<
813 { <<
814 struct userfaultfd_unmap_ctx *unmap_ct <<
815 <<
816 list_for_each_entry(unmap_ctx, unmaps, <<
817 if (unmap_ctx->ctx == ctx && u <<
818 unmap_ctx->end == end) <<
819 return true; <<
820 <<
821 return false; <<
822 } <<
823 <<
824 int userfaultfd_unmap_prep(struct vm_area_stru <<
825 unsigned long end, <<
826 { <<
827 struct userfaultfd_unmap_ctx *unmap_ct <<
828 struct userfaultfd_ctx *ctx = vma->vm_ <<
829 <<
830 if (!ctx || !(ctx->features & UFFD_FEA <<
831 has_unmap_ctx(ctx, unmaps, start, <<
832 return 0; <<
833 <<
834 unmap_ctx = kzalloc(sizeof(*unmap_ctx) <<
835 if (!unmap_ctx) <<
836 return -ENOMEM; <<
837 <<
838 userfaultfd_ctx_get(ctx); <<
839 down_write(&ctx->map_changing_lock); <<
840 atomic_inc(&ctx->mmap_changing); <<
841 up_write(&ctx->map_changing_lock); <<
842 unmap_ctx->ctx = ctx; <<
843 unmap_ctx->start = start; <<
844 unmap_ctx->end = end; <<
845 list_add_tail(&unmap_ctx->list, unmaps <<
846 <<
847 return 0; <<
848 } <<
849 <<
850 void userfaultfd_unmap_complete(struct mm_stru <<
851 { <<
852 struct userfaultfd_unmap_ctx *ctx, *n; <<
853 struct userfaultfd_wait_queue ewq; <<
854 <<
855 list_for_each_entry_safe(ctx, n, uf, l <<
856 msg_init(&ewq.msg); <<
857 <<
858 ewq.msg.event = UFFD_EVENT_UNM <<
859 ewq.msg.arg.remove.start = ctx <<
860 ewq.msg.arg.remove.end = ctx-> <<
861 <<
862 userfaultfd_event_wait_complet <<
863 <<
864 list_del(&ctx->list); <<
865 kfree(ctx); <<
866 } <<
867 } <<
868 <<
869 static int userfaultfd_release(struct inode *i 459 static int userfaultfd_release(struct inode *inode, struct file *file)
870 { 460 {
871 struct userfaultfd_ctx *ctx = file->pr 461 struct userfaultfd_ctx *ctx = file->private_data;
872 struct mm_struct *mm = ctx->mm; 462 struct mm_struct *mm = ctx->mm;
873 struct vm_area_struct *vma, *prev; 463 struct vm_area_struct *vma, *prev;
874 /* len == 0 means wake all */ 464 /* len == 0 means wake all */
875 struct userfaultfd_wake_range range = 465 struct userfaultfd_wake_range range = { .len = 0, };
876 unsigned long new_flags; 466 unsigned long new_flags;
877 VMA_ITERATOR(vmi, mm, 0); !! 467 bool still_valid;
878 468
879 WRITE_ONCE(ctx->released, true); !! 469 ACCESS_ONCE(ctx->released) = true;
880 470
881 if (!mmget_not_zero(mm)) 471 if (!mmget_not_zero(mm))
882 goto wakeup; 472 goto wakeup;
883 473
884 /* 474 /*
885 * Flush page faults out of all CPUs. 475 * Flush page faults out of all CPUs. NOTE: all page faults
886 * must be retried without returning V 476 * must be retried without returning VM_FAULT_SIGBUS if
887 * userfaultfd_ctx_get() succeeds but 477 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
888 * changes while handle_userfault rele !! 478 * changes while handle_userfault released the mmap_sem. So
889 * it's critical that released is set 479 * it's critical that released is set to true (above), before
890 * taking the mmap_lock for writing. !! 480 * taking the mmap_sem for writing.
891 */ 481 */
892 mmap_write_lock(mm); !! 482 down_write(&mm->mmap_sem);
>> 483 still_valid = mmget_still_valid(mm);
893 prev = NULL; 484 prev = NULL;
894 for_each_vma(vmi, vma) { !! 485 for (vma = mm->mmap; vma; vma = vma->vm_next) {
895 cond_resched(); 486 cond_resched();
896 BUG_ON(!!vma->vm_userfaultfd_c 487 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
897 !!(vma->vm_flags & __VM !! 488 !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
898 if (vma->vm_userfaultfd_ctx.ct 489 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
899 prev = vma; 490 prev = vma;
900 continue; 491 continue;
901 } 492 }
902 /* Reset ptes for the whole vm !! 493 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
903 if (userfaultfd_wp(vma)) !! 494 if (still_valid) {
904 uffd_wp_range(vma, vma !! 495 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
905 vma->vm_ !! 496 new_flags, vma->anon_vma,
906 new_flags = vma->vm_flags & ~_ !! 497 vma->vm_file, vma->vm_pgoff,
907 vma = vma_modify_flags_uffd(&v !! 498 vma_policy(vma),
908 vm !! 499 NULL_VM_UFFD_CTX);
909 NU !! 500 if (prev)
910 !! 501 vma = prev;
911 vma_start_write(vma); !! 502 else
912 userfaultfd_set_vm_flags(vma, !! 503 prev = vma;
>> 504 }
>> 505 vma->vm_flags = new_flags;
913 vma->vm_userfaultfd_ctx = NULL 506 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
914 <<
915 prev = vma; <<
916 } 507 }
917 mmap_write_unlock(mm); !! 508 up_write(&mm->mmap_sem);
918 mmput(mm); 509 mmput(mm);
919 wakeup: 510 wakeup:
920 /* 511 /*
921 * After no new page faults can wait o 512 * After no new page faults can wait on this fault_*wqh, flush
922 * the last page faults that may have 513 * the last page faults that may have been already waiting on
923 * the fault_*wqh. 514 * the fault_*wqh.
924 */ 515 */
925 spin_lock_irq(&ctx->fault_pending_wqh. !! 516 spin_lock(&ctx->fault_pending_wqh.lock);
926 __wake_up_locked_key(&ctx->fault_pendi 517 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
927 __wake_up(&ctx->fault_wqh, TASK_NORMAL !! 518 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
928 spin_unlock_irq(&ctx->fault_pending_wq !! 519 spin_unlock(&ctx->fault_pending_wqh.lock);
929 520
930 /* Flush pending events that may still !! 521 wake_up_poll(&ctx->fd_wqh, POLLHUP);
931 wake_up_all(&ctx->event_wqh); <<
932 <<
933 wake_up_poll(&ctx->fd_wqh, EPOLLHUP); <<
934 userfaultfd_ctx_put(ctx); 522 userfaultfd_ctx_put(ctx);
935 return 0; 523 return 0;
936 } 524 }
937 525
938 /* fault_pending_wqh.lock must be hold by the 526 /* fault_pending_wqh.lock must be hold by the caller */
939 static inline struct userfaultfd_wait_queue *f !! 527 static inline struct userfaultfd_wait_queue *find_userfault(
940 wait_queue_head_t *wqh) !! 528 struct userfaultfd_ctx *ctx)
941 { 529 {
942 wait_queue_entry_t *wq; !! 530 wait_queue_t *wq;
943 struct userfaultfd_wait_queue *uwq; 531 struct userfaultfd_wait_queue *uwq;
944 532
945 lockdep_assert_held(&wqh->lock); !! 533 VM_BUG_ON(!spin_is_locked(&ctx->fault_pending_wqh.lock));
946 534
947 uwq = NULL; 535 uwq = NULL;
948 if (!waitqueue_active(wqh)) !! 536 if (!waitqueue_active(&ctx->fault_pending_wqh))
949 goto out; 537 goto out;
950 /* walk in reverse to provide FIFO beh 538 /* walk in reverse to provide FIFO behavior to read userfaults */
951 wq = list_last_entry(&wqh->head, typeo !! 539 wq = list_last_entry(&ctx->fault_pending_wqh.task_list,
>> 540 typeof(*wq), task_list);
952 uwq = container_of(wq, struct userfaul 541 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
953 out: 542 out:
954 return uwq; 543 return uwq;
955 } 544 }
956 545
957 static inline struct userfaultfd_wait_queue *f !! 546 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
958 struct userfaultfd_ctx *ctx) <<
959 { <<
960 return find_userfault_in(&ctx->fault_p <<
961 } <<
962 <<
963 static inline struct userfaultfd_wait_queue *f <<
964 struct userfaultfd_ctx *ctx) <<
965 { <<
966 return find_userfault_in(&ctx->event_w <<
967 } <<
968 <<
969 static __poll_t userfaultfd_poll(struct file * <<
970 { 547 {
971 struct userfaultfd_ctx *ctx = file->pr 548 struct userfaultfd_ctx *ctx = file->private_data;
972 __poll_t ret; !! 549 unsigned int ret;
973 550
974 poll_wait(file, &ctx->fd_wqh, wait); 551 poll_wait(file, &ctx->fd_wqh, wait);
975 552
976 if (!userfaultfd_is_initialized(ctx)) !! 553 switch (ctx->state) {
977 return EPOLLERR; !! 554 case UFFD_STATE_WAIT_API:
978 !! 555 return POLLERR;
979 /* !! 556 case UFFD_STATE_RUNNING:
980 * poll() never guarantees that read w !! 557 /*
981 * userfaults can be waken before they !! 558 * poll() never guarantees that read won't block.
982 */ !! 559 * userfaults can be waken before they're read().
983 if (unlikely(!(file->f_flags & O_NONBL !! 560 */
984 return EPOLLERR; !! 561 if (unlikely(!(file->f_flags & O_NONBLOCK)))
985 /* !! 562 return POLLERR;
986 * lockless access to see if there are !! 563 /*
987 * __pollwait last action is the add_w !! 564 * lockless access to see if there are pending faults
988 * the spin_unlock would allow the wai !! 565 * __pollwait last action is the add_wait_queue but
989 * pass above the actual list_add insi !! 566 * the spin_unlock would allow the waitqueue_active to
990 * add_wait_queue critical section. So !! 567 * pass above the actual list_add inside
991 * memory barrier to serialize the lis !! 568 * add_wait_queue critical section. So use a full
992 * add_wait_queue() with the waitqueue !! 569 * memory barrier to serialize the list_add write of
993 * below. !! 570 * add_wait_queue() with the waitqueue_active read
994 */ !! 571 * below.
995 ret = 0; !! 572 */
996 smp_mb(); !! 573 ret = 0;
997 if (waitqueue_active(&ctx->fault_pendi !! 574 smp_mb();
998 ret = EPOLLIN; !! 575 if (waitqueue_active(&ctx->fault_pending_wqh))
999 else if (waitqueue_active(&ctx->event_ !! 576 ret = POLLIN;
1000 ret = EPOLLIN; !! 577 return ret;
1001 !! 578 default:
1002 return ret; !! 579 BUG();
1003 } !! 580 }
1004 <<
1005 static const struct file_operations userfault <<
1006 <<
1007 static int resolve_userfault_fork(struct user <<
1008 struct inod <<
1009 struct uffd <<
1010 { <<
1011 int fd; <<
1012 <<
1013 fd = anon_inode_create_getfd("[userfa <<
1014 O_RDONLY | (new->flag <<
1015 if (fd < 0) <<
1016 return fd; <<
1017 <<
1018 msg->arg.reserved.reserved1 = 0; <<
1019 msg->arg.fork.ufd = fd; <<
1020 return 0; <<
1021 } 581 }
1022 582
1023 static ssize_t userfaultfd_ctx_read(struct us 583 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1024 struct uf !! 584 struct uffd_msg *msg)
1025 { 585 {
1026 ssize_t ret; 586 ssize_t ret;
1027 DECLARE_WAITQUEUE(wait, current); 587 DECLARE_WAITQUEUE(wait, current);
1028 struct userfaultfd_wait_queue *uwq; 588 struct userfaultfd_wait_queue *uwq;
1029 /* <<
1030 * Handling fork event requires sleep <<
1031 * we drop the event_wqh lock, then d <<
1032 * lock it back and wake up the waite <<
1033 * dropped the ewq may go away so we <<
1034 * carefully. <<
1035 */ <<
1036 LIST_HEAD(fork_event); <<
1037 struct userfaultfd_ctx *fork_nctx = N <<
1038 589
1039 /* always take the fd_wqh lock before 590 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1040 spin_lock_irq(&ctx->fd_wqh.lock); !! 591 spin_lock(&ctx->fd_wqh.lock);
1041 __add_wait_queue(&ctx->fd_wqh, &wait) 592 __add_wait_queue(&ctx->fd_wqh, &wait);
1042 for (;;) { 593 for (;;) {
1043 set_current_state(TASK_INTERR 594 set_current_state(TASK_INTERRUPTIBLE);
1044 spin_lock(&ctx->fault_pending 595 spin_lock(&ctx->fault_pending_wqh.lock);
1045 uwq = find_userfault(ctx); 596 uwq = find_userfault(ctx);
1046 if (uwq) { 597 if (uwq) {
1047 /* 598 /*
1048 * Use a seqcount to 599 * Use a seqcount to repeat the lockless check
1049 * in wake_userfault( 600 * in wake_userfault() to avoid missing
1050 * wakeups because du 601 * wakeups because during the refile both
1051 * waitqueue could be 602 * waitqueue could become empty if this is the
1052 * only userfault. 603 * only userfault.
1053 */ 604 */
1054 write_seqcount_begin( 605 write_seqcount_begin(&ctx->refile_seq);
1055 606
1056 /* 607 /*
1057 * The fault_pending_ 608 * The fault_pending_wqh.lock prevents the uwq
1058 * to disappear from 609 * to disappear from under us.
1059 * 610 *
1060 * Refile this userfa 611 * Refile this userfault from
1061 * fault_pending_wqh 612 * fault_pending_wqh to fault_wqh, it's not
1062 * pending anymore af 613 * pending anymore after we read it.
1063 * 614 *
1064 * Use list_del() by 615 * Use list_del() by hand (as
1065 * userfaultfd_wake_f 616 * userfaultfd_wake_function also uses
1066 * list_del_init() by 617 * list_del_init() by hand) to be sure nobody
1067 * changes __remove_w 618 * changes __remove_wait_queue() to use
1068 * list_del_init() in 619 * list_del_init() in turn breaking the
1069 * !list_empty_carefu 620 * !list_empty_careful() check in
1070 * handle_userfault() !! 621 * handle_userfault(). The uwq->wq.task_list
1071 * must never be empt 622 * must never be empty at any time during the
1072 * refile, or the wai 623 * refile, or the waitqueue could disappear
1073 * from under us. The 624 * from under us. The "wait_queue_head_t"
1074 * parameter of __rem 625 * parameter of __remove_wait_queue() is unused
1075 * anyway. 626 * anyway.
1076 */ 627 */
1077 list_del(&uwq->wq.ent !! 628 list_del(&uwq->wq.task_list);
1078 add_wait_queue(&ctx-> !! 629 __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1079 630
1080 write_seqcount_end(&c 631 write_seqcount_end(&ctx->refile_seq);
1081 632
1082 /* careful to always 633 /* careful to always initialize msg if ret == 0 */
1083 *msg = uwq->msg; 634 *msg = uwq->msg;
1084 spin_unlock(&ctx->fau 635 spin_unlock(&ctx->fault_pending_wqh.lock);
1085 ret = 0; 636 ret = 0;
1086 break; 637 break;
1087 } 638 }
1088 spin_unlock(&ctx->fault_pendi 639 spin_unlock(&ctx->fault_pending_wqh.lock);
1089 <<
1090 spin_lock(&ctx->event_wqh.loc <<
1091 uwq = find_userfault_evt(ctx) <<
1092 if (uwq) { <<
1093 *msg = uwq->msg; <<
1094 <<
1095 if (uwq->msg.event == <<
1096 fork_nctx = ( <<
1097 (unsi <<
1098 uwq-> <<
1099 list_move(&uw <<
1100 /* <<
1101 * fork_nctx <<
1102 * we drop th <<
1103 * reference <<
1104 */ <<
1105 userfaultfd_c <<
1106 spin_unlock(& <<
1107 ret = 0; <<
1108 break; <<
1109 } <<
1110 <<
1111 userfaultfd_event_com <<
1112 spin_unlock(&ctx->eve <<
1113 ret = 0; <<
1114 break; <<
1115 } <<
1116 spin_unlock(&ctx->event_wqh.l <<
1117 <<
1118 if (signal_pending(current)) 640 if (signal_pending(current)) {
1119 ret = -ERESTARTSYS; 641 ret = -ERESTARTSYS;
1120 break; 642 break;
1121 } 643 }
1122 if (no_wait) { 644 if (no_wait) {
1123 ret = -EAGAIN; 645 ret = -EAGAIN;
1124 break; 646 break;
1125 } 647 }
1126 spin_unlock_irq(&ctx->fd_wqh. !! 648 spin_unlock(&ctx->fd_wqh.lock);
1127 schedule(); 649 schedule();
1128 spin_lock_irq(&ctx->fd_wqh.lo !! 650 spin_lock(&ctx->fd_wqh.lock);
1129 } 651 }
1130 __remove_wait_queue(&ctx->fd_wqh, &wa 652 __remove_wait_queue(&ctx->fd_wqh, &wait);
1131 __set_current_state(TASK_RUNNING); 653 __set_current_state(TASK_RUNNING);
1132 spin_unlock_irq(&ctx->fd_wqh.lock); !! 654 spin_unlock(&ctx->fd_wqh.lock);
1133 <<
1134 if (!ret && msg->event == UFFD_EVENT_ <<
1135 ret = resolve_userfault_fork( <<
1136 spin_lock_irq(&ctx->event_wqh <<
1137 if (!list_empty(&fork_event)) <<
1138 /* <<
1139 * The fork thread di <<
1140 * drop the temporary <<
1141 */ <<
1142 userfaultfd_ctx_put(f <<
1143 <<
1144 uwq = list_first_entr <<
1145 <<
1146 <<
1147 /* <<
1148 * If fork_event list <<
1149 * the event wasn't a <<
1150 * (the event is allo <<
1151 * stack), put the ev <<
1152 * the event_wq. fork <<
1153 * as soon as we retu <<
1154 * stay queued there <<
1155 * "ret" value. <<
1156 */ <<
1157 list_del(&uwq->wq.ent <<
1158 __add_wait_queue(&ctx <<
1159 <<
1160 /* <<
1161 * Leave the event in <<
1162 * error to userland <<
1163 * the userfault fork <<
1164 */ <<
1165 if (likely(!ret)) <<
1166 userfaultfd_e <<
1167 } else { <<
1168 /* <<
1169 * Here the fork thre <<
1170 * refcount from the <<
1171 * has already been r <<
1172 * the reference we t <<
1173 * lock above. If res <<
1174 * failed we've to dr <<
1175 * fork_nctx has to b <<
1176 * it succeeded we'll <<
1177 * uffd references it <<
1178 */ <<
1179 if (ret) <<
1180 userfaultfd_c <<
1181 } <<
1182 spin_unlock_irq(&ctx->event_w <<
1183 } <<
1184 655
1185 return ret; 656 return ret;
1186 } 657 }
1187 658
1188 static ssize_t userfaultfd_read_iter(struct k !! 659 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
>> 660 size_t count, loff_t *ppos)
1189 { 661 {
1190 struct file *file = iocb->ki_filp; <<
1191 struct userfaultfd_ctx *ctx = file->p 662 struct userfaultfd_ctx *ctx = file->private_data;
1192 ssize_t _ret, ret = 0; 663 ssize_t _ret, ret = 0;
1193 struct uffd_msg msg; 664 struct uffd_msg msg;
1194 struct inode *inode = file_inode(file !! 665 int no_wait = file->f_flags & O_NONBLOCK;
1195 bool no_wait; <<
1196 666
1197 if (!userfaultfd_is_initialized(ctx)) !! 667 if (ctx->state == UFFD_STATE_WAIT_API)
1198 return -EINVAL; 668 return -EINVAL;
1199 669
1200 no_wait = file->f_flags & O_NONBLOCK <<
1201 for (;;) { 670 for (;;) {
1202 if (iov_iter_count(to) < size !! 671 if (count < sizeof(msg))
1203 return ret ? ret : -E 672 return ret ? ret : -EINVAL;
1204 _ret = userfaultfd_ctx_read(c !! 673 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1205 if (_ret < 0) 674 if (_ret < 0)
1206 return ret ? ret : _r 675 return ret ? ret : _ret;
1207 _ret = !copy_to_iter_full(&ms !! 676 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1208 if (_ret) <<
1209 return ret ? ret : -E 677 return ret ? ret : -EFAULT;
1210 ret += sizeof(msg); 678 ret += sizeof(msg);
>> 679 buf += sizeof(msg);
>> 680 count -= sizeof(msg);
1211 /* 681 /*
1212 * Allow to read more than on 682 * Allow to read more than one fault at time but only
1213 * block if waiting for the v 683 * block if waiting for the very first one.
1214 */ 684 */
1215 no_wait = true; !! 685 no_wait = O_NONBLOCK;
1216 } 686 }
1217 } 687 }
1218 688
1219 static void __wake_userfault(struct userfault 689 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1220 struct userfault 690 struct userfaultfd_wake_range *range)
1221 { 691 {
1222 spin_lock_irq(&ctx->fault_pending_wqh !! 692 unsigned long start, end;
>> 693
>> 694 start = range->start;
>> 695 end = range->start + range->len;
>> 696
>> 697 spin_lock(&ctx->fault_pending_wqh.lock);
1223 /* wake all in the range and autoremo 698 /* wake all in the range and autoremove */
1224 if (waitqueue_active(&ctx->fault_pend 699 if (waitqueue_active(&ctx->fault_pending_wqh))
1225 __wake_up_locked_key(&ctx->fa 700 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1226 range); 701 range);
1227 if (waitqueue_active(&ctx->fault_wqh) 702 if (waitqueue_active(&ctx->fault_wqh))
1228 __wake_up(&ctx->fault_wqh, TA !! 703 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1229 spin_unlock_irq(&ctx->fault_pending_w !! 704 spin_unlock(&ctx->fault_pending_wqh.lock);
1230 } 705 }
1231 706
1232 static __always_inline void wake_userfault(st 707 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1233 st 708 struct userfaultfd_wake_range *range)
1234 { 709 {
1235 unsigned seq; 710 unsigned seq;
1236 bool need_wakeup; 711 bool need_wakeup;
1237 712
1238 /* 713 /*
1239 * To be sure waitqueue_active() is n 714 * To be sure waitqueue_active() is not reordered by the CPU
1240 * before the pagetable update, use a 715 * before the pagetable update, use an explicit SMP memory
1241 * barrier here. PT lock release or m !! 716 * barrier here. PT lock release or up_read(mmap_sem) still
1242 * have release semantics that can al 717 * have release semantics that can allow the
1243 * waitqueue_active() to be reordered 718 * waitqueue_active() to be reordered before the pte update.
1244 */ 719 */
1245 smp_mb(); 720 smp_mb();
1246 721
1247 /* 722 /*
1248 * Use waitqueue_active because it's 723 * Use waitqueue_active because it's very frequent to
1249 * change the address space atomicall 724 * change the address space atomically even if there are no
1250 * userfaults yet. So we take the spi 725 * userfaults yet. So we take the spinlock only when we're
1251 * sure we've userfaults to wake. 726 * sure we've userfaults to wake.
1252 */ 727 */
1253 do { 728 do {
1254 seq = read_seqcount_begin(&ct 729 seq = read_seqcount_begin(&ctx->refile_seq);
1255 need_wakeup = waitqueue_activ 730 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1256 waitqueue_active(&ctx 731 waitqueue_active(&ctx->fault_wqh);
1257 cond_resched(); 732 cond_resched();
1258 } while (read_seqcount_retry(&ctx->re 733 } while (read_seqcount_retry(&ctx->refile_seq, seq));
1259 if (need_wakeup) 734 if (need_wakeup)
1260 __wake_userfault(ctx, range); 735 __wake_userfault(ctx, range);
1261 } 736 }
1262 737
1263 static __always_inline int validate_unaligned !! 738 static __always_inline int validate_range(struct mm_struct *mm,
1264 struct mm_struct *mm, __u64 start, __ !! 739 __u64 start, __u64 len)
1265 { 740 {
1266 __u64 task_size = mm->task_size; 741 __u64 task_size = mm->task_size;
1267 742
>> 743 if (start & ~PAGE_MASK)
>> 744 return -EINVAL;
1268 if (len & ~PAGE_MASK) 745 if (len & ~PAGE_MASK)
1269 return -EINVAL; 746 return -EINVAL;
1270 if (!len) 747 if (!len)
1271 return -EINVAL; 748 return -EINVAL;
1272 if (start < mmap_min_addr) 749 if (start < mmap_min_addr)
1273 return -EINVAL; 750 return -EINVAL;
1274 if (start >= task_size) 751 if (start >= task_size)
1275 return -EINVAL; 752 return -EINVAL;
1276 if (len > task_size - start) 753 if (len > task_size - start)
1277 return -EINVAL; 754 return -EINVAL;
1278 if (start + len <= start) <<
1279 return -EINVAL; <<
1280 return 0; 755 return 0;
1281 } 756 }
1282 757
1283 static __always_inline int validate_range(str <<
1284 __u <<
1285 { <<
1286 if (start & ~PAGE_MASK) <<
1287 return -EINVAL; <<
1288 <<
1289 return validate_unaligned_range(mm, s <<
1290 } <<
1291 <<
1292 static int userfaultfd_register(struct userfa 758 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1293 unsigned long 759 unsigned long arg)
1294 { 760 {
1295 struct mm_struct *mm = ctx->mm; 761 struct mm_struct *mm = ctx->mm;
1296 struct vm_area_struct *vma, *prev, *c 762 struct vm_area_struct *vma, *prev, *cur;
1297 int ret; 763 int ret;
1298 struct uffdio_register uffdio_registe 764 struct uffdio_register uffdio_register;
1299 struct uffdio_register __user *user_u 765 struct uffdio_register __user *user_uffdio_register;
1300 unsigned long vm_flags, new_flags; 766 unsigned long vm_flags, new_flags;
1301 bool found; 767 bool found;
1302 bool basic_ioctls; <<
1303 unsigned long start, end, vma_end; 768 unsigned long start, end, vma_end;
1304 struct vma_iterator vmi; <<
1305 bool wp_async = userfaultfd_wp_async_ <<
1306 769
1307 user_uffdio_register = (struct uffdio 770 user_uffdio_register = (struct uffdio_register __user *) arg;
1308 771
1309 ret = -EFAULT; 772 ret = -EFAULT;
1310 if (copy_from_user(&uffdio_register, 773 if (copy_from_user(&uffdio_register, user_uffdio_register,
1311 sizeof(uffdio_regi 774 sizeof(uffdio_register)-sizeof(__u64)))
1312 goto out; 775 goto out;
1313 776
1314 ret = -EINVAL; 777 ret = -EINVAL;
1315 if (!uffdio_register.mode) 778 if (!uffdio_register.mode)
1316 goto out; 779 goto out;
1317 if (uffdio_register.mode & ~UFFD_API_ !! 780 if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
>> 781 UFFDIO_REGISTER_MODE_WP))
1318 goto out; 782 goto out;
1319 vm_flags = 0; 783 vm_flags = 0;
1320 if (uffdio_register.mode & UFFDIO_REG 784 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1321 vm_flags |= VM_UFFD_MISSING; 785 vm_flags |= VM_UFFD_MISSING;
1322 if (uffdio_register.mode & UFFDIO_REG 786 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1323 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP <<
1324 goto out; <<
1325 #endif <<
1326 vm_flags |= VM_UFFD_WP; 787 vm_flags |= VM_UFFD_WP;
1327 } !! 788 /*
1328 if (uffdio_register.mode & UFFDIO_REG !! 789 * FIXME: remove the below error constraint by
1329 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR !! 790 * implementing the wprotect tracking mode.
>> 791 */
>> 792 ret = -EINVAL;
1330 goto out; 793 goto out;
1331 #endif <<
1332 vm_flags |= VM_UFFD_MINOR; <<
1333 } 794 }
1334 795
1335 ret = validate_range(mm, uffdio_regis 796 ret = validate_range(mm, uffdio_register.range.start,
1336 uffdio_register. 797 uffdio_register.range.len);
1337 if (ret) 798 if (ret)
1338 goto out; 799 goto out;
1339 800
1340 start = uffdio_register.range.start; 801 start = uffdio_register.range.start;
1341 end = start + uffdio_register.range.l 802 end = start + uffdio_register.range.len;
1342 803
1343 ret = -ENOMEM; 804 ret = -ENOMEM;
1344 if (!mmget_not_zero(mm)) 805 if (!mmget_not_zero(mm))
1345 goto out; 806 goto out;
1346 807
1347 ret = -EINVAL; !! 808 down_write(&mm->mmap_sem);
1348 mmap_write_lock(mm); !! 809 if (!mmget_still_valid(mm))
1349 vma_iter_init(&vmi, mm, start); <<
1350 vma = vma_find(&vmi, end); <<
1351 if (!vma) <<
1352 goto out_unlock; 810 goto out_unlock;
1353 811
1354 /* !! 812 vma = find_vma_prev(mm, start, &prev);
1355 * If the first vma contains huge pag !! 813 if (!vma)
1356 * is aligned to huge page size. !! 814 goto out_unlock;
1357 */ <<
1358 if (is_vm_hugetlb_page(vma)) { <<
1359 unsigned long vma_hpagesize = <<
1360 815
1361 if (start & (vma_hpagesize - !! 816 /* check that there's at least one vma in the range */
1362 goto out_unlock; !! 817 ret = -EINVAL;
1363 } !! 818 if (vma->vm_start >= end)
>> 819 goto out_unlock;
1364 820
1365 /* 821 /*
1366 * Search for not compatible vmas. 822 * Search for not compatible vmas.
>> 823 *
>> 824 * FIXME: this shall be relaxed later so that it doesn't fail
>> 825 * on tmpfs backed vmas (in addition to the current allowance
>> 826 * on anonymous vmas).
1367 */ 827 */
1368 found = false; 828 found = false;
1369 basic_ioctls = false; !! 829 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1370 cur = vma; <<
1371 do { <<
1372 cond_resched(); 830 cond_resched();
1373 831
1374 BUG_ON(!!cur->vm_userfaultfd_ 832 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1375 !!(cur->vm_flags & __V !! 833 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1376 834
1377 /* check not compatible vmas 835 /* check not compatible vmas */
1378 ret = -EINVAL; 836 ret = -EINVAL;
1379 if (!vma_can_userfault(cur, v !! 837 if (cur->vm_ops)
1380 goto out_unlock; <<
1381 <<
1382 /* <<
1383 * UFFDIO_COPY will fill file <<
1384 * PROT_WRITE. This check enf <<
1385 * MAP_SHARED, the process ha <<
1386 * file. If VM_MAYWRITE is se <<
1387 * MAP_SHARED vma: there is n <<
1388 * F_WRITE_SEAL can be taken <<
1389 */ <<
1390 ret = -EPERM; <<
1391 if (unlikely(!(cur->vm_flags <<
1392 goto out_unlock; <<
1393 <<
1394 /* <<
1395 * If this vma contains endin <<
1396 * check alignment. <<
1397 */ <<
1398 if (is_vm_hugetlb_page(cur) & <<
1399 end > cur->vm_start) { <<
1400 unsigned long vma_hpa <<
1401 <<
1402 ret = -EINVAL; <<
1403 <<
1404 if (end & (vma_hpages <<
1405 goto out_unlo <<
1406 } <<
1407 if ((vm_flags & VM_UFFD_WP) & <<
1408 goto out_unlock; 838 goto out_unlock;
1409 839
1410 /* 840 /*
1411 * Check that this vma isn't 841 * Check that this vma isn't already owned by a
1412 * different userfaultfd. We 842 * different userfaultfd. We can't allow more than one
1413 * userfaultfd to own a singl 843 * userfaultfd to own a single vma simultaneously or we
1414 * wouldn't know which one to 844 * wouldn't know which one to deliver the userfaults to.
1415 */ 845 */
1416 ret = -EBUSY; 846 ret = -EBUSY;
1417 if (cur->vm_userfaultfd_ctx.c 847 if (cur->vm_userfaultfd_ctx.ctx &&
1418 cur->vm_userfaultfd_ctx.c 848 cur->vm_userfaultfd_ctx.ctx != ctx)
1419 goto out_unlock; 849 goto out_unlock;
1420 850
1421 /* <<
1422 * Note vmas containing huge <<
1423 */ <<
1424 if (is_vm_hugetlb_page(cur)) <<
1425 basic_ioctls = true; <<
1426 <<
1427 found = true; 851 found = true;
1428 } for_each_vma_range(vmi, cur, end); !! 852 }
1429 BUG_ON(!found); 853 BUG_ON(!found);
1430 854
1431 vma_iter_set(&vmi, start); <<
1432 prev = vma_prev(&vmi); <<
1433 if (vma->vm_start < start) 855 if (vma->vm_start < start)
1434 prev = vma; 856 prev = vma;
1435 857
1436 ret = 0; 858 ret = 0;
1437 for_each_vma_range(vmi, vma, end) { !! 859 do {
1438 cond_resched(); 860 cond_resched();
1439 861
1440 BUG_ON(!vma_can_userfault(vma !! 862 BUG_ON(vma->vm_ops);
1441 BUG_ON(vma->vm_userfaultfd_ct 863 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1442 vma->vm_userfaultfd_ct 864 vma->vm_userfaultfd_ctx.ctx != ctx);
1443 WARN_ON(!(vma->vm_flags & VM_ <<
1444 865
1445 /* 866 /*
1446 * Nothing to do: this vma is 867 * Nothing to do: this vma is already registered into this
1447 * userfaultfd and with the r 868 * userfaultfd and with the right tracking mode too.
1448 */ 869 */
1449 if (vma->vm_userfaultfd_ctx.c 870 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1450 (vma->vm_flags & vm_flags 871 (vma->vm_flags & vm_flags) == vm_flags)
1451 goto skip; 872 goto skip;
1452 873
1453 if (vma->vm_start > start) 874 if (vma->vm_start > start)
1454 start = vma->vm_start 875 start = vma->vm_start;
1455 vma_end = min(end, vma->vm_en 876 vma_end = min(end, vma->vm_end);
1456 877
1457 new_flags = (vma->vm_flags & !! 878 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1458 vma = vma_modify_flags_uffd(& !! 879 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1459 n !! 880 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1460 ( !! 881 vma_policy(vma),
1461 if (IS_ERR(vma)) { !! 882 ((struct vm_userfaultfd_ctx){ ctx }));
1462 ret = PTR_ERR(vma); !! 883 if (prev) {
1463 break; !! 884 vma = prev;
>> 885 goto next;
1464 } 886 }
1465 !! 887 if (vma->vm_start < start) {
>> 888 ret = split_vma(mm, vma, start, 1);
>> 889 if (ret)
>> 890 break;
>> 891 }
>> 892 if (vma->vm_end > end) {
>> 893 ret = split_vma(mm, vma, end, 0);
>> 894 if (ret)
>> 895 break;
>> 896 }
>> 897 next:
1466 /* 898 /*
1467 * In the vma_merge() success 899 * In the vma_merge() successful mprotect-like case 8:
1468 * the next vma was merged in 900 * the next vma was merged into the current one and
1469 * the current one has not be 901 * the current one has not been updated yet.
1470 */ 902 */
1471 vma_start_write(vma); !! 903 vma->vm_flags = new_flags;
1472 userfaultfd_set_vm_flags(vma, <<
1473 vma->vm_userfaultfd_ctx.ctx = 904 vma->vm_userfaultfd_ctx.ctx = ctx;
1474 905
1475 if (is_vm_hugetlb_page(vma) & <<
1476 hugetlb_unshare_all_p <<
1477 <<
1478 skip: 906 skip:
1479 prev = vma; 907 prev = vma;
1480 start = vma->vm_end; 908 start = vma->vm_end;
1481 } !! 909 vma = vma->vm_next;
1482 !! 910 } while (vma && vma->vm_start < end);
1483 out_unlock: 911 out_unlock:
1484 mmap_write_unlock(mm); !! 912 up_write(&mm->mmap_sem);
1485 mmput(mm); 913 mmput(mm);
1486 if (!ret) { 914 if (!ret) {
1487 __u64 ioctls_out; <<
1488 <<
1489 ioctls_out = basic_ioctls ? U <<
1490 UFFD_API_RANGE_IOCTLS; <<
1491 <<
1492 /* <<
1493 * Declare the WP ioctl only <<
1494 * specified and all checks p <<
1495 */ <<
1496 if (!(uffdio_register.mode & <<
1497 ioctls_out &= ~((__u6 <<
1498 <<
1499 /* CONTINUE ioctl is only sup <<
1500 if (!(uffdio_register.mode & <<
1501 ioctls_out &= ~((__u6 <<
1502 <<
1503 /* 915 /*
1504 * Now that we scanned all vm 916 * Now that we scanned all vmas we can already tell
1505 * userland which ioctls meth 917 * userland which ioctls methods are guaranteed to
1506 * succeed on this range. 918 * succeed on this range.
1507 */ 919 */
1508 if (put_user(ioctls_out, &use !! 920 if (put_user(UFFD_API_RANGE_IOCTLS,
>> 921 &user_uffdio_register->ioctls))
1509 ret = -EFAULT; 922 ret = -EFAULT;
1510 } 923 }
1511 out: 924 out:
1512 return ret; 925 return ret;
1513 } 926 }
1514 927
1515 static int userfaultfd_unregister(struct user 928 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1516 unsigned lo 929 unsigned long arg)
1517 { 930 {
1518 struct mm_struct *mm = ctx->mm; 931 struct mm_struct *mm = ctx->mm;
1519 struct vm_area_struct *vma, *prev, *c 932 struct vm_area_struct *vma, *prev, *cur;
1520 int ret; 933 int ret;
1521 struct uffdio_range uffdio_unregister 934 struct uffdio_range uffdio_unregister;
1522 unsigned long new_flags; 935 unsigned long new_flags;
1523 bool found; 936 bool found;
1524 unsigned long start, end, vma_end; 937 unsigned long start, end, vma_end;
1525 const void __user *buf = (void __user 938 const void __user *buf = (void __user *)arg;
1526 struct vma_iterator vmi; <<
1527 bool wp_async = userfaultfd_wp_async_ <<
1528 939
1529 ret = -EFAULT; 940 ret = -EFAULT;
1530 if (copy_from_user(&uffdio_unregister 941 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1531 goto out; 942 goto out;
1532 943
1533 ret = validate_range(mm, uffdio_unreg 944 ret = validate_range(mm, uffdio_unregister.start,
1534 uffdio_unregiste 945 uffdio_unregister.len);
1535 if (ret) 946 if (ret)
1536 goto out; 947 goto out;
1537 948
1538 start = uffdio_unregister.start; 949 start = uffdio_unregister.start;
1539 end = start + uffdio_unregister.len; 950 end = start + uffdio_unregister.len;
1540 951
1541 ret = -ENOMEM; 952 ret = -ENOMEM;
1542 if (!mmget_not_zero(mm)) 953 if (!mmget_not_zero(mm))
1543 goto out; 954 goto out;
1544 955
1545 mmap_write_lock(mm); !! 956 down_write(&mm->mmap_sem);
1546 ret = -EINVAL; !! 957 if (!mmget_still_valid(mm))
1547 vma_iter_init(&vmi, mm, start); <<
1548 vma = vma_find(&vmi, end); <<
1549 if (!vma) <<
1550 goto out_unlock; 958 goto out_unlock;
1551 959
1552 /* !! 960 vma = find_vma_prev(mm, start, &prev);
1553 * If the first vma contains huge pag !! 961 if (!vma)
1554 * is aligned to huge page size. !! 962 goto out_unlock;
1555 */ <<
1556 if (is_vm_hugetlb_page(vma)) { <<
1557 unsigned long vma_hpagesize = <<
1558 963
1559 if (start & (vma_hpagesize - !! 964 /* check that there's at least one vma in the range */
1560 goto out_unlock; !! 965 ret = -EINVAL;
1561 } !! 966 if (vma->vm_start >= end)
>> 967 goto out_unlock;
1562 968
1563 /* 969 /*
1564 * Search for not compatible vmas. 970 * Search for not compatible vmas.
>> 971 *
>> 972 * FIXME: this shall be relaxed later so that it doesn't fail
>> 973 * on tmpfs backed vmas (in addition to the current allowance
>> 974 * on anonymous vmas).
1565 */ 975 */
1566 found = false; 976 found = false;
1567 cur = vma; !! 977 ret = -EINVAL;
1568 do { !! 978 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1569 cond_resched(); 979 cond_resched();
1570 980
1571 BUG_ON(!!cur->vm_userfaultfd_ 981 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1572 !!(cur->vm_flags & __V !! 982 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1573 983
1574 /* 984 /*
1575 * Check not compatible vmas, 985 * Check not compatible vmas, not strictly required
1576 * here as not compatible vma 986 * here as not compatible vmas cannot have an
1577 * userfaultfd_ctx registered 987 * userfaultfd_ctx registered on them, but this
1578 * provides for more strict b 988 * provides for more strict behavior to notice
1579 * unregistration errors. 989 * unregistration errors.
1580 */ 990 */
1581 if (!vma_can_userfault(cur, c !! 991 if (cur->vm_ops)
1582 goto out_unlock; 992 goto out_unlock;
1583 993
1584 found = true; 994 found = true;
1585 } for_each_vma_range(vmi, cur, end); !! 995 }
1586 BUG_ON(!found); 996 BUG_ON(!found);
1587 997
1588 vma_iter_set(&vmi, start); <<
1589 prev = vma_prev(&vmi); <<
1590 if (vma->vm_start < start) 998 if (vma->vm_start < start)
1591 prev = vma; 999 prev = vma;
1592 1000
1593 ret = 0; 1001 ret = 0;
1594 for_each_vma_range(vmi, vma, end) { !! 1002 do {
1595 cond_resched(); 1003 cond_resched();
1596 1004
1597 BUG_ON(!vma_can_userfault(vma !! 1005 BUG_ON(vma->vm_ops);
1598 1006
1599 /* 1007 /*
1600 * Nothing to do: this vma is 1008 * Nothing to do: this vma is already registered into this
1601 * userfaultfd and with the r 1009 * userfaultfd and with the right tracking mode too.
1602 */ 1010 */
1603 if (!vma->vm_userfaultfd_ctx. 1011 if (!vma->vm_userfaultfd_ctx.ctx)
1604 goto skip; 1012 goto skip;
1605 1013
1606 WARN_ON(!(vma->vm_flags & VM_ <<
1607 <<
1608 if (vma->vm_start > start) 1014 if (vma->vm_start > start)
1609 start = vma->vm_start 1015 start = vma->vm_start;
1610 vma_end = min(end, vma->vm_en 1016 vma_end = min(end, vma->vm_end);
1611 1017
1612 if (userfaultfd_missing(vma)) !! 1018 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1613 /* !! 1019 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1614 * Wake any concurren !! 1020 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1615 * we unregister, so !! 1021 vma_policy(vma),
1616 * permanently and it !! 1022 NULL_VM_UFFD_CTX);
1617 * UFFDIO_WAKE explic !! 1023 if (prev) {
1618 */ !! 1024 vma = prev;
1619 struct userfaultfd_wa !! 1025 goto next;
1620 range.start = start; <<
1621 range.len = vma_end - <<
1622 wake_userfault(vma->v <<
1623 } 1026 }
1624 !! 1027 if (vma->vm_start < start) {
1625 /* Reset ptes for the whole v !! 1028 ret = split_vma(mm, vma, start, 1);
1626 if (userfaultfd_wp(vma)) !! 1029 if (ret)
1627 uffd_wp_range(vma, st !! 1030 break;
1628 <<
1629 new_flags = vma->vm_flags & ~ <<
1630 vma = vma_modify_flags_uffd(& <<
1631 n <<
1632 if (IS_ERR(vma)) { <<
1633 ret = PTR_ERR(vma); <<
1634 break; <<
1635 } 1031 }
1636 !! 1032 if (vma->vm_end > end) {
>> 1033 ret = split_vma(mm, vma, end, 0);
>> 1034 if (ret)
>> 1035 break;
>> 1036 }
>> 1037 next:
1637 /* 1038 /*
1638 * In the vma_merge() success 1039 * In the vma_merge() successful mprotect-like case 8:
1639 * the next vma was merged in 1040 * the next vma was merged into the current one and
1640 * the current one has not be 1041 * the current one has not been updated yet.
1641 */ 1042 */
1642 vma_start_write(vma); !! 1043 vma->vm_flags = new_flags;
1643 userfaultfd_set_vm_flags(vma, <<
1644 vma->vm_userfaultfd_ctx = NUL 1044 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1645 1045
1646 skip: 1046 skip:
1647 prev = vma; 1047 prev = vma;
1648 start = vma->vm_end; 1048 start = vma->vm_end;
1649 } !! 1049 vma = vma->vm_next;
1650 !! 1050 } while (vma && vma->vm_start < end);
1651 out_unlock: 1051 out_unlock:
1652 mmap_write_unlock(mm); !! 1052 up_write(&mm->mmap_sem);
1653 mmput(mm); 1053 mmput(mm);
1654 out: 1054 out:
1655 return ret; 1055 return ret;
1656 } 1056 }
1657 1057
1658 /* 1058 /*
1659 * userfaultfd_wake may be used in combinatio 1059 * userfaultfd_wake may be used in combination with the
1660 * UFFDIO_*_MODE_DONTWAKE to wakeup userfault 1060 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1661 */ 1061 */
1662 static int userfaultfd_wake(struct userfaultf 1062 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1663 unsigned long arg 1063 unsigned long arg)
1664 { 1064 {
1665 int ret; 1065 int ret;
1666 struct uffdio_range uffdio_wake; 1066 struct uffdio_range uffdio_wake;
1667 struct userfaultfd_wake_range range; 1067 struct userfaultfd_wake_range range;
1668 const void __user *buf = (void __user 1068 const void __user *buf = (void __user *)arg;
1669 1069
1670 ret = -EFAULT; 1070 ret = -EFAULT;
1671 if (copy_from_user(&uffdio_wake, buf, 1071 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1672 goto out; 1072 goto out;
1673 1073
1674 ret = validate_range(ctx->mm, uffdio_ 1074 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1675 if (ret) 1075 if (ret)
1676 goto out; 1076 goto out;
1677 1077
1678 range.start = uffdio_wake.start; 1078 range.start = uffdio_wake.start;
1679 range.len = uffdio_wake.len; 1079 range.len = uffdio_wake.len;
1680 1080
1681 /* 1081 /*
1682 * len == 0 means wake all and we don 1082 * len == 0 means wake all and we don't want to wake all here,
1683 * so check it again to be sure. 1083 * so check it again to be sure.
1684 */ 1084 */
1685 VM_BUG_ON(!range.len); 1085 VM_BUG_ON(!range.len);
1686 1086
1687 wake_userfault(ctx, &range); 1087 wake_userfault(ctx, &range);
1688 ret = 0; 1088 ret = 0;
1689 1089
1690 out: 1090 out:
1691 return ret; 1091 return ret;
1692 } 1092 }
1693 1093
1694 static int userfaultfd_copy(struct userfaultf 1094 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1695 unsigned long arg 1095 unsigned long arg)
1696 { 1096 {
1697 __s64 ret; 1097 __s64 ret;
1698 struct uffdio_copy uffdio_copy; 1098 struct uffdio_copy uffdio_copy;
1699 struct uffdio_copy __user *user_uffdi 1099 struct uffdio_copy __user *user_uffdio_copy;
1700 struct userfaultfd_wake_range range; 1100 struct userfaultfd_wake_range range;
1701 uffd_flags_t flags = 0; <<
1702 1101
1703 user_uffdio_copy = (struct uffdio_cop 1102 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1704 1103
1705 ret = -EAGAIN; <<
1706 if (atomic_read(&ctx->mmap_changing)) <<
1707 goto out; <<
1708 <<
1709 ret = -EFAULT; 1104 ret = -EFAULT;
1710 if (copy_from_user(&uffdio_copy, user 1105 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1711 /* don't copy "cop 1106 /* don't copy "copy" last field */
1712 sizeof(uffdio_copy 1107 sizeof(uffdio_copy)-sizeof(__s64)))
1713 goto out; 1108 goto out;
1714 1109
1715 ret = validate_unaligned_range(ctx->m <<
1716 uffdio <<
1717 if (ret) <<
1718 goto out; <<
1719 ret = validate_range(ctx->mm, uffdio_ 1110 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1720 if (ret) 1111 if (ret)
1721 goto out; 1112 goto out;
1722 !! 1113 /*
>> 1114 * double check for wraparound just in case. copy_from_user()
>> 1115 * will later check uffdio_copy.src + uffdio_copy.len to fit
>> 1116 * in the userland range.
>> 1117 */
1723 ret = -EINVAL; 1118 ret = -EINVAL;
1724 if (uffdio_copy.mode & ~(UFFDIO_COPY_ !! 1119 if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
>> 1120 goto out;
>> 1121 if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1725 goto out; 1122 goto out;
1726 if (uffdio_copy.mode & UFFDIO_COPY_MO <<
1727 flags |= MFILL_ATOMIC_WP; <<
1728 if (mmget_not_zero(ctx->mm)) { 1123 if (mmget_not_zero(ctx->mm)) {
1729 ret = mfill_atomic_copy(ctx, !! 1124 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1730 uffdi !! 1125 uffdio_copy.len);
1731 mmput(ctx->mm); 1126 mmput(ctx->mm);
1732 } else { <<
1733 return -ESRCH; <<
1734 } 1127 }
1735 if (unlikely(put_user(ret, &user_uffd 1128 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1736 return -EFAULT; 1129 return -EFAULT;
1737 if (ret < 0) 1130 if (ret < 0)
1738 goto out; 1131 goto out;
1739 BUG_ON(!ret); 1132 BUG_ON(!ret);
1740 /* len == 0 would wake all */ 1133 /* len == 0 would wake all */
1741 range.len = ret; 1134 range.len = ret;
1742 if (!(uffdio_copy.mode & UFFDIO_COPY_ 1135 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1743 range.start = uffdio_copy.dst 1136 range.start = uffdio_copy.dst;
1744 wake_userfault(ctx, &range); 1137 wake_userfault(ctx, &range);
1745 } 1138 }
1746 ret = range.len == uffdio_copy.len ? 1139 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1747 out: 1140 out:
1748 return ret; 1141 return ret;
1749 } 1142 }
1750 1143
1751 static int userfaultfd_zeropage(struct userfa 1144 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1752 unsigned long 1145 unsigned long arg)
1753 { 1146 {
1754 __s64 ret; 1147 __s64 ret;
1755 struct uffdio_zeropage uffdio_zeropag 1148 struct uffdio_zeropage uffdio_zeropage;
1756 struct uffdio_zeropage __user *user_u 1149 struct uffdio_zeropage __user *user_uffdio_zeropage;
1757 struct userfaultfd_wake_range range; 1150 struct userfaultfd_wake_range range;
1758 1151
1759 user_uffdio_zeropage = (struct uffdio 1152 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1760 1153
1761 ret = -EAGAIN; <<
1762 if (atomic_read(&ctx->mmap_changing)) <<
1763 goto out; <<
1764 <<
1765 ret = -EFAULT; 1154 ret = -EFAULT;
1766 if (copy_from_user(&uffdio_zeropage, 1155 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1767 /* don't copy "zer 1156 /* don't copy "zeropage" last field */
1768 sizeof(uffdio_zero 1157 sizeof(uffdio_zeropage)-sizeof(__s64)))
1769 goto out; 1158 goto out;
1770 1159
1771 ret = validate_range(ctx->mm, uffdio_ 1160 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1772 uffdio_zeropage. 1161 uffdio_zeropage.range.len);
1773 if (ret) 1162 if (ret)
1774 goto out; 1163 goto out;
1775 ret = -EINVAL; 1164 ret = -EINVAL;
1776 if (uffdio_zeropage.mode & ~UFFDIO_ZE 1165 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1777 goto out; 1166 goto out;
1778 1167
1779 if (mmget_not_zero(ctx->mm)) { 1168 if (mmget_not_zero(ctx->mm)) {
1780 ret = mfill_atomic_zeropage(c !! 1169 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1781 uf !! 1170 uffdio_zeropage.range.len);
1782 mmput(ctx->mm); 1171 mmput(ctx->mm);
1783 } else { <<
1784 return -ESRCH; <<
1785 } 1172 }
1786 if (unlikely(put_user(ret, &user_uffd 1173 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1787 return -EFAULT; 1174 return -EFAULT;
1788 if (ret < 0) 1175 if (ret < 0)
1789 goto out; 1176 goto out;
1790 /* len == 0 would wake all */ 1177 /* len == 0 would wake all */
1791 BUG_ON(!ret); 1178 BUG_ON(!ret);
1792 range.len = ret; 1179 range.len = ret;
1793 if (!(uffdio_zeropage.mode & UFFDIO_Z 1180 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1794 range.start = uffdio_zeropage 1181 range.start = uffdio_zeropage.range.start;
1795 wake_userfault(ctx, &range); 1182 wake_userfault(ctx, &range);
1796 } 1183 }
1797 ret = range.len == uffdio_zeropage.ra 1184 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1798 out: 1185 out:
1799 return ret; 1186 return ret;
1800 } 1187 }
1801 1188
1802 static int userfaultfd_writeprotect(struct us <<
1803 unsigned <<
1804 { <<
1805 int ret; <<
1806 struct uffdio_writeprotect uffdio_wp; <<
1807 struct uffdio_writeprotect __user *us <<
1808 struct userfaultfd_wake_range range; <<
1809 bool mode_wp, mode_dontwake; <<
1810 <<
1811 if (atomic_read(&ctx->mmap_changing)) <<
1812 return -EAGAIN; <<
1813 <<
1814 user_uffdio_wp = (struct uffdio_write <<
1815 <<
1816 if (copy_from_user(&uffdio_wp, user_u <<
1817 sizeof(struct uffd <<
1818 return -EFAULT; <<
1819 <<
1820 ret = validate_range(ctx->mm, uffdio_ <<
1821 uffdio_wp.range. <<
1822 if (ret) <<
1823 return ret; <<
1824 <<
1825 if (uffdio_wp.mode & ~(UFFDIO_WRITEPR <<
1826 UFFDIO_WRITEPR <<
1827 return -EINVAL; <<
1828 <<
1829 mode_wp = uffdio_wp.mode & UFFDIO_WRI <<
1830 mode_dontwake = uffdio_wp.mode & UFFD <<
1831 <<
1832 if (mode_wp && mode_dontwake) <<
1833 return -EINVAL; <<
1834 <<
1835 if (mmget_not_zero(ctx->mm)) { <<
1836 ret = mwriteprotect_range(ctx <<
1837 uff <<
1838 mmput(ctx->mm); <<
1839 } else { <<
1840 return -ESRCH; <<
1841 } <<
1842 <<
1843 if (ret) <<
1844 return ret; <<
1845 <<
1846 if (!mode_wp && !mode_dontwake) { <<
1847 range.start = uffdio_wp.range <<
1848 range.len = uffdio_wp.range.l <<
1849 wake_userfault(ctx, &range); <<
1850 } <<
1851 return ret; <<
1852 } <<
1853 <<
1854 static int userfaultfd_continue(struct userfa <<
1855 { <<
1856 __s64 ret; <<
1857 struct uffdio_continue uffdio_continu <<
1858 struct uffdio_continue __user *user_u <<
1859 struct userfaultfd_wake_range range; <<
1860 uffd_flags_t flags = 0; <<
1861 <<
1862 user_uffdio_continue = (struct uffdio <<
1863 <<
1864 ret = -EAGAIN; <<
1865 if (atomic_read(&ctx->mmap_changing)) <<
1866 goto out; <<
1867 <<
1868 ret = -EFAULT; <<
1869 if (copy_from_user(&uffdio_continue, <<
1870 /* don't copy the <<
1871 sizeof(uffdio_cont <<
1872 goto out; <<
1873 <<
1874 ret = validate_range(ctx->mm, uffdio_ <<
1875 uffdio_continue. <<
1876 if (ret) <<
1877 goto out; <<
1878 <<
1879 ret = -EINVAL; <<
1880 if (uffdio_continue.mode & ~(UFFDIO_C <<
1881 UFFDIO_C <<
1882 goto out; <<
1883 if (uffdio_continue.mode & UFFDIO_CON <<
1884 flags |= MFILL_ATOMIC_WP; <<
1885 <<
1886 if (mmget_not_zero(ctx->mm)) { <<
1887 ret = mfill_atomic_continue(c <<
1888 u <<
1889 mmput(ctx->mm); <<
1890 } else { <<
1891 return -ESRCH; <<
1892 } <<
1893 <<
1894 if (unlikely(put_user(ret, &user_uffd <<
1895 return -EFAULT; <<
1896 if (ret < 0) <<
1897 goto out; <<
1898 <<
1899 /* len == 0 would wake all */ <<
1900 BUG_ON(!ret); <<
1901 range.len = ret; <<
1902 if (!(uffdio_continue.mode & UFFDIO_C <<
1903 range.start = uffdio_continue <<
1904 wake_userfault(ctx, &range); <<
1905 } <<
1906 ret = range.len == uffdio_continue.ra <<
1907 <<
1908 out: <<
1909 return ret; <<
1910 } <<
1911 <<
1912 static inline int userfaultfd_poison(struct u <<
1913 { <<
1914 __s64 ret; <<
1915 struct uffdio_poison uffdio_poison; <<
1916 struct uffdio_poison __user *user_uff <<
1917 struct userfaultfd_wake_range range; <<
1918 <<
1919 user_uffdio_poison = (struct uffdio_p <<
1920 <<
1921 ret = -EAGAIN; <<
1922 if (atomic_read(&ctx->mmap_changing)) <<
1923 goto out; <<
1924 <<
1925 ret = -EFAULT; <<
1926 if (copy_from_user(&uffdio_poison, us <<
1927 /* don't copy the <<
1928 sizeof(uffdio_pois <<
1929 goto out; <<
1930 <<
1931 ret = validate_range(ctx->mm, uffdio_ <<
1932 uffdio_poison.ra <<
1933 if (ret) <<
1934 goto out; <<
1935 <<
1936 ret = -EINVAL; <<
1937 if (uffdio_poison.mode & ~UFFDIO_POIS <<
1938 goto out; <<
1939 <<
1940 if (mmget_not_zero(ctx->mm)) { <<
1941 ret = mfill_atomic_poison(ctx <<
1942 uff <<
1943 mmput(ctx->mm); <<
1944 } else { <<
1945 return -ESRCH; <<
1946 } <<
1947 <<
1948 if (unlikely(put_user(ret, &user_uffd <<
1949 return -EFAULT; <<
1950 if (ret < 0) <<
1951 goto out; <<
1952 <<
1953 /* len == 0 would wake all */ <<
1954 BUG_ON(!ret); <<
1955 range.len = ret; <<
1956 if (!(uffdio_poison.mode & UFFDIO_POI <<
1957 range.start = uffdio_poison.r <<
1958 wake_userfault(ctx, &range); <<
1959 } <<
1960 ret = range.len == uffdio_poison.rang <<
1961 <<
1962 out: <<
1963 return ret; <<
1964 } <<
1965 <<
1966 bool userfaultfd_wp_async(struct vm_area_stru <<
1967 { <<
1968 return userfaultfd_wp_async_ctx(vma-> <<
1969 } <<
1970 <<
1971 static inline unsigned int uffd_ctx_features( <<
1972 { <<
1973 /* <<
1974 * For the current set of features th <<
1975 * UFFD_FEATURE_INITIALIZED to mark t <<
1976 */ <<
1977 return (unsigned int)user_features | <<
1978 } <<
1979 <<
1980 static int userfaultfd_move(struct userfaultf <<
1981 unsigned long arg <<
1982 { <<
1983 __s64 ret; <<
1984 struct uffdio_move uffdio_move; <<
1985 struct uffdio_move __user *user_uffdi <<
1986 struct userfaultfd_wake_range range; <<
1987 struct mm_struct *mm = ctx->mm; <<
1988 <<
1989 user_uffdio_move = (struct uffdio_mov <<
1990 <<
1991 if (atomic_read(&ctx->mmap_changing)) <<
1992 return -EAGAIN; <<
1993 <<
1994 if (copy_from_user(&uffdio_move, user <<
1995 /* don't copy "mov <<
1996 sizeof(uffdio_move <<
1997 return -EFAULT; <<
1998 <<
1999 /* Do not allow cross-mm moves. */ <<
2000 if (mm != current->mm) <<
2001 return -EINVAL; <<
2002 <<
2003 ret = validate_range(mm, uffdio_move. <<
2004 if (ret) <<
2005 return ret; <<
2006 <<
2007 ret = validate_range(mm, uffdio_move. <<
2008 if (ret) <<
2009 return ret; <<
2010 <<
2011 if (uffdio_move.mode & ~(UFFDIO_MOVE_ <<
2012 UFFDIO_MOVE <<
2013 return -EINVAL; <<
2014 <<
2015 if (mmget_not_zero(mm)) { <<
2016 ret = move_pages(ctx, uffdio_ <<
2017 uffdio_move. <<
2018 mmput(mm); <<
2019 } else { <<
2020 return -ESRCH; <<
2021 } <<
2022 <<
2023 if (unlikely(put_user(ret, &user_uffd <<
2024 return -EFAULT; <<
2025 if (ret < 0) <<
2026 goto out; <<
2027 <<
2028 /* len == 0 would wake all */ <<
2029 VM_WARN_ON(!ret); <<
2030 range.len = ret; <<
2031 if (!(uffdio_move.mode & UFFDIO_MOVE_ <<
2032 range.start = uffdio_move.dst <<
2033 wake_userfault(ctx, &range); <<
2034 } <<
2035 ret = range.len == uffdio_move.len ? <<
2036 <<
2037 out: <<
2038 return ret; <<
2039 } <<
2040 <<
2041 /* 1189 /*
2042 * userland asks for a certain API version an 1190 * userland asks for a certain API version and we return which bits
2043 * and ioctl commands are implemented in this 1191 * and ioctl commands are implemented in this kernel for such API
2044 * version or -EINVAL if unknown. 1192 * version or -EINVAL if unknown.
2045 */ 1193 */
2046 static int userfaultfd_api(struct userfaultfd 1194 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
2047 unsigned long arg) 1195 unsigned long arg)
2048 { 1196 {
2049 struct uffdio_api uffdio_api; 1197 struct uffdio_api uffdio_api;
2050 void __user *buf = (void __user *)arg 1198 void __user *buf = (void __user *)arg;
2051 unsigned int ctx_features; <<
2052 int ret; 1199 int ret;
2053 __u64 features; <<
2054 1200
>> 1201 ret = -EINVAL;
>> 1202 if (ctx->state != UFFD_STATE_WAIT_API)
>> 1203 goto out;
2055 ret = -EFAULT; 1204 ret = -EFAULT;
2056 if (copy_from_user(&uffdio_api, buf, 1205 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
2057 goto out; 1206 goto out;
2058 features = uffdio_api.features; !! 1207 if (uffdio_api.api != UFFD_API || uffdio_api.features) {
2059 ret = -EINVAL; !! 1208 memset(&uffdio_api, 0, sizeof(uffdio_api));
2060 if (uffdio_api.api != UFFD_API) !! 1209 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2061 goto err_out; !! 1210 goto out;
2062 ret = -EPERM; !! 1211 ret = -EINVAL;
2063 if ((features & UFFD_FEATURE_EVENT_FO !! 1212 goto out;
2064 goto err_out; !! 1213 }
2065 <<
2066 /* WP_ASYNC relies on WP_UNPOPULATED, <<
2067 if (features & UFFD_FEATURE_WP_ASYNC) <<
2068 features |= UFFD_FEATURE_WP_U <<
2069 <<
2070 /* report all available features and <<
2071 uffdio_api.features = UFFD_API_FEATUR 1214 uffdio_api.features = UFFD_API_FEATURES;
2072 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR <<
2073 uffdio_api.features &= <<
2074 ~(UFFD_FEATURE_MINOR_HUGETLBF <<
2075 #endif <<
2076 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP <<
2077 uffdio_api.features &= ~UFFD_FEATURE_ <<
2078 #endif <<
2079 #ifndef CONFIG_PTE_MARKER_UFFD_WP <<
2080 uffdio_api.features &= ~UFFD_FEATURE_ <<
2081 uffdio_api.features &= ~UFFD_FEATURE_ <<
2082 uffdio_api.features &= ~UFFD_FEATURE_ <<
2083 #endif <<
2084 <<
2085 ret = -EINVAL; <<
2086 if (features & ~uffdio_api.features) <<
2087 goto err_out; <<
2088 <<
2089 uffdio_api.ioctls = UFFD_API_IOCTLS; 1215 uffdio_api.ioctls = UFFD_API_IOCTLS;
2090 ret = -EFAULT; 1216 ret = -EFAULT;
2091 if (copy_to_user(buf, &uffdio_api, si 1217 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2092 goto out; 1218 goto out;
2093 !! 1219 ctx->state = UFFD_STATE_RUNNING;
2094 /* only enable the requested features <<
2095 ctx_features = uffd_ctx_features(feat <<
2096 ret = -EINVAL; <<
2097 if (cmpxchg(&ctx->features, 0, ctx_fe <<
2098 goto err_out; <<
2099 <<
2100 ret = 0; 1220 ret = 0;
2101 out: 1221 out:
2102 return ret; 1222 return ret;
2103 err_out: <<
2104 memset(&uffdio_api, 0, sizeof(uffdio_ <<
2105 if (copy_to_user(buf, &uffdio_api, si <<
2106 ret = -EFAULT; <<
2107 goto out; <<
2108 } 1223 }
2109 1224
2110 static long userfaultfd_ioctl(struct file *fi 1225 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
2111 unsigned long a 1226 unsigned long arg)
2112 { 1227 {
2113 int ret = -EINVAL; 1228 int ret = -EINVAL;
2114 struct userfaultfd_ctx *ctx = file->p 1229 struct userfaultfd_ctx *ctx = file->private_data;
2115 1230
2116 if (cmd != UFFDIO_API && !userfaultfd !! 1231 if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
2117 return -EINVAL; 1232 return -EINVAL;
2118 1233
2119 switch(cmd) { 1234 switch(cmd) {
2120 case UFFDIO_API: 1235 case UFFDIO_API:
2121 ret = userfaultfd_api(ctx, ar 1236 ret = userfaultfd_api(ctx, arg);
2122 break; 1237 break;
2123 case UFFDIO_REGISTER: 1238 case UFFDIO_REGISTER:
2124 ret = userfaultfd_register(ct 1239 ret = userfaultfd_register(ctx, arg);
2125 break; 1240 break;
2126 case UFFDIO_UNREGISTER: 1241 case UFFDIO_UNREGISTER:
2127 ret = userfaultfd_unregister( 1242 ret = userfaultfd_unregister(ctx, arg);
2128 break; 1243 break;
2129 case UFFDIO_WAKE: 1244 case UFFDIO_WAKE:
2130 ret = userfaultfd_wake(ctx, a 1245 ret = userfaultfd_wake(ctx, arg);
2131 break; 1246 break;
2132 case UFFDIO_COPY: 1247 case UFFDIO_COPY:
2133 ret = userfaultfd_copy(ctx, a 1248 ret = userfaultfd_copy(ctx, arg);
2134 break; 1249 break;
2135 case UFFDIO_ZEROPAGE: 1250 case UFFDIO_ZEROPAGE:
2136 ret = userfaultfd_zeropage(ct 1251 ret = userfaultfd_zeropage(ctx, arg);
2137 break; 1252 break;
2138 case UFFDIO_MOVE: <<
2139 ret = userfaultfd_move(ctx, a <<
2140 break; <<
2141 case UFFDIO_WRITEPROTECT: <<
2142 ret = userfaultfd_writeprotec <<
2143 break; <<
2144 case UFFDIO_CONTINUE: <<
2145 ret = userfaultfd_continue(ct <<
2146 break; <<
2147 case UFFDIO_POISON: <<
2148 ret = userfaultfd_poison(ctx, <<
2149 break; <<
2150 } 1253 }
2151 return ret; 1254 return ret;
2152 } 1255 }
2153 1256
2154 #ifdef CONFIG_PROC_FS 1257 #ifdef CONFIG_PROC_FS
2155 static void userfaultfd_show_fdinfo(struct se 1258 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2156 { 1259 {
2157 struct userfaultfd_ctx *ctx = f->priv 1260 struct userfaultfd_ctx *ctx = f->private_data;
2158 wait_queue_entry_t *wq; !! 1261 wait_queue_t *wq;
>> 1262 struct userfaultfd_wait_queue *uwq;
2159 unsigned long pending = 0, total = 0; 1263 unsigned long pending = 0, total = 0;
2160 1264
2161 spin_lock_irq(&ctx->fault_pending_wqh !! 1265 spin_lock(&ctx->fault_pending_wqh.lock);
2162 list_for_each_entry(wq, &ctx->fault_p !! 1266 list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
>> 1267 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
2163 pending++; 1268 pending++;
2164 total++; 1269 total++;
2165 } 1270 }
2166 list_for_each_entry(wq, &ctx->fault_w !! 1271 list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
>> 1272 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
2167 total++; 1273 total++;
2168 } 1274 }
2169 spin_unlock_irq(&ctx->fault_pending_w !! 1275 spin_unlock(&ctx->fault_pending_wqh.lock);
2170 1276
2171 /* 1277 /*
2172 * If more protocols will be added, t 1278 * If more protocols will be added, there will be all shown
2173 * separated by a space. Like this: 1279 * separated by a space. Like this:
2174 * protocols: aa:... bb:... 1280 * protocols: aa:... bb:...
2175 */ 1281 */
2176 seq_printf(m, "pending:\t%lu\ntotal:\ 1282 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2177 pending, total, UFFD_API, !! 1283 pending, total, UFFD_API, UFFD_API_FEATURES,
2178 UFFD_API_IOCTLS|UFFD_API_R 1284 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2179 } 1285 }
2180 #endif 1286 #endif
2181 1287
2182 static const struct file_operations userfault 1288 static const struct file_operations userfaultfd_fops = {
2183 #ifdef CONFIG_PROC_FS 1289 #ifdef CONFIG_PROC_FS
2184 .show_fdinfo = userfaultfd_show_fd 1290 .show_fdinfo = userfaultfd_show_fdinfo,
2185 #endif 1291 #endif
2186 .release = userfaultfd_release 1292 .release = userfaultfd_release,
2187 .poll = userfaultfd_poll, 1293 .poll = userfaultfd_poll,
2188 .read_iter = userfaultfd_read_it !! 1294 .read = userfaultfd_read,
2189 .unlocked_ioctl = userfaultfd_ioctl, 1295 .unlocked_ioctl = userfaultfd_ioctl,
2190 .compat_ioctl = compat_ptr_ioctl, !! 1296 .compat_ioctl = userfaultfd_ioctl,
2191 .llseek = noop_llseek, 1297 .llseek = noop_llseek,
2192 }; 1298 };
2193 1299
2194 static void init_once_userfaultfd_ctx(void *m 1300 static void init_once_userfaultfd_ctx(void *mem)
2195 { 1301 {
2196 struct userfaultfd_ctx *ctx = (struct 1302 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2197 1303
2198 init_waitqueue_head(&ctx->fault_pendi 1304 init_waitqueue_head(&ctx->fault_pending_wqh);
2199 init_waitqueue_head(&ctx->fault_wqh); 1305 init_waitqueue_head(&ctx->fault_wqh);
2200 init_waitqueue_head(&ctx->event_wqh); <<
2201 init_waitqueue_head(&ctx->fd_wqh); 1306 init_waitqueue_head(&ctx->fd_wqh);
2202 seqcount_spinlock_init(&ctx->refile_s !! 1307 seqcount_init(&ctx->refile_seq);
2203 } 1308 }
2204 1309
2205 static int new_userfaultfd(int flags) !! 1310 /**
>> 1311 * userfaultfd_file_create - Creates an userfaultfd file pointer.
>> 1312 * @flags: Flags for the userfaultfd file.
>> 1313 *
>> 1314 * This function creates an userfaultfd file pointer, w/out installing
>> 1315 * it into the fd table. This is useful when the userfaultfd file is
>> 1316 * used during the initialization of data structures that require
>> 1317 * extra setup after the userfaultfd creation. So the userfaultfd
>> 1318 * creation is split into the file pointer creation phase, and the
>> 1319 * file descriptor installation phase. In this way races with
>> 1320 * userspace closing the newly installed file descriptor can be
>> 1321 * avoided. Returns an userfaultfd file pointer, or a proper error
>> 1322 * pointer.
>> 1323 */
>> 1324 static struct file *userfaultfd_file_create(int flags)
2206 { 1325 {
2207 struct userfaultfd_ctx *ctx; <<
2208 struct file *file; 1326 struct file *file;
2209 int fd; !! 1327 struct userfaultfd_ctx *ctx;
2210 1328
2211 BUG_ON(!current->mm); 1329 BUG_ON(!current->mm);
2212 1330
2213 /* Check the UFFD_* constants for con 1331 /* Check the UFFD_* constants for consistency. */
2214 BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UF <<
2215 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXE 1332 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2216 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBL 1333 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2217 1334
2218 if (flags & ~(UFFD_SHARED_FCNTL_FLAGS !! 1335 file = ERR_PTR(-EINVAL);
2219 return -EINVAL; !! 1336 if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
>> 1337 goto out;
2220 1338
>> 1339 file = ERR_PTR(-ENOMEM);
2221 ctx = kmem_cache_alloc(userfaultfd_ct 1340 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2222 if (!ctx) 1341 if (!ctx)
2223 return -ENOMEM; !! 1342 goto out;
2224 1343
2225 refcount_set(&ctx->refcount, 1); !! 1344 atomic_set(&ctx->refcount, 1);
2226 ctx->flags = flags; 1345 ctx->flags = flags;
2227 ctx->features = 0; !! 1346 ctx->state = UFFD_STATE_WAIT_API;
2228 ctx->released = false; 1347 ctx->released = false;
2229 init_rwsem(&ctx->map_changing_lock); <<
2230 atomic_set(&ctx->mmap_changing, 0); <<
2231 ctx->mm = current->mm; 1348 ctx->mm = current->mm;
>> 1349 /* prevent the mm struct to be freed */
>> 1350 atomic_inc(&ctx->mm->mm_count);
2232 1351
2233 fd = get_unused_fd_flags(flags & UFFD !! 1352 file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
2234 if (fd < 0) !! 1353 O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
2235 goto err_out; <<
2236 <<
2237 /* Create a new inode so that the LSM <<
2238 file = anon_inode_create_getfile("[us <<
2239 O_RDONLY | (flags & U <<
2240 if (IS_ERR(file)) { 1354 if (IS_ERR(file)) {
2241 put_unused_fd(fd); !! 1355 mmdrop(ctx->mm);
2242 fd = PTR_ERR(file); !! 1356 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2243 goto err_out; <<
2244 } 1357 }
2245 /* prevent the mm struct to be freed !! 1358 out:
2246 mmgrab(ctx->mm); !! 1359 return file;
2247 file->f_mode |= FMODE_NOWAIT; <<
2248 fd_install(fd, file); <<
2249 return fd; <<
2250 err_out: <<
2251 kmem_cache_free(userfaultfd_ctx_cache <<
2252 return fd; <<
2253 } <<
2254 <<
2255 static inline bool userfaultfd_syscall_allowe <<
2256 { <<
2257 /* Userspace-only page faults are alw <<
2258 if (flags & UFFD_USER_MODE_ONLY) <<
2259 return true; <<
2260 <<
2261 /* <<
2262 * The user is requesting a userfault <<
2263 * Privileged users are always allowe <<
2264 */ <<
2265 if (capable(CAP_SYS_PTRACE)) <<
2266 return true; <<
2267 <<
2268 /* Otherwise, access to kernel fault <<
2269 return sysctl_unprivileged_userfaultf <<
2270 } 1360 }
2271 1361
2272 SYSCALL_DEFINE1(userfaultfd, int, flags) 1362 SYSCALL_DEFINE1(userfaultfd, int, flags)
2273 { 1363 {
2274 if (!userfaultfd_syscall_allowed(flag !! 1364 int fd, error;
2275 return -EPERM; !! 1365 struct file *file;
2276 1366
2277 return new_userfaultfd(flags); !! 1367 error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
2278 } !! 1368 if (error < 0)
>> 1369 return error;
>> 1370 fd = error;
2279 1371
2280 static long userfaultfd_dev_ioctl(struct file !! 1372 file = userfaultfd_file_create(flags);
2281 { !! 1373 if (IS_ERR(file)) {
2282 if (cmd != USERFAULTFD_IOC_NEW) !! 1374 error = PTR_ERR(file);
2283 return -EINVAL; !! 1375 goto err_put_unused_fd;
>> 1376 }
>> 1377 fd_install(fd, file);
2284 1378
2285 return new_userfaultfd(flags); !! 1379 return fd;
2286 } <<
2287 1380
2288 static const struct file_operations userfault !! 1381 err_put_unused_fd:
2289 .unlocked_ioctl = userfaultfd_dev_ioc !! 1382 put_unused_fd(fd);
2290 .compat_ioctl = userfaultfd_dev_ioctl <<
2291 .owner = THIS_MODULE, <<
2292 .llseek = noop_llseek, <<
2293 }; <<
2294 1383
2295 static struct miscdevice userfaultfd_misc = { !! 1384 return error;
2296 .minor = MISC_DYNAMIC_MINOR, !! 1385 }
2297 .name = "userfaultfd", <<
2298 .fops = &userfaultfd_dev_fops <<
2299 }; <<
2300 1386
2301 static int __init userfaultfd_init(void) 1387 static int __init userfaultfd_init(void)
2302 { 1388 {
2303 int ret; <<
2304 <<
2305 ret = misc_register(&userfaultfd_misc <<
2306 if (ret) <<
2307 return ret; <<
2308 <<
2309 userfaultfd_ctx_cachep = kmem_cache_c 1389 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2310 1390 sizeof(struct userfaultfd_ctx),
2311 1391 0,
2312 1392 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2313 1393 init_once_userfaultfd_ctx);
2314 #ifdef CONFIG_SYSCTL <<
2315 register_sysctl_init("vm", vm_userfau <<
2316 #endif <<
2317 return 0; 1394 return 0;
2318 } 1395 }
2319 __initcall(userfaultfd_init); 1396 __initcall(userfaultfd_init);
2320 1397
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.