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