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