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