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