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