1 // SPDX-License-Identifier: GPL-2.0-only 1 // SPDX-License-Identifier: GPL-2.0-only
2 /* 2 /*
3 * fs/userfaultfd.c 3 * fs/userfaultfd.c
4 * 4 *
5 * Copyright (C) 2007 Davide Libenzi <davide 5 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
6 * Copyright (C) 2008-2009 Red Hat, Inc. 6 * Copyright (C) 2008-2009 Red Hat, Inc.
7 * Copyright (C) 2015 Red Hat, Inc. 7 * Copyright (C) 2015 Red Hat, Inc.
8 * 8 *
9 * Some part derived from fs/eventfd.c (anon 9 * Some part derived from fs/eventfd.c (anon inode setup) and
10 * mm/ksm.c (mm hashing). 10 * mm/ksm.c (mm hashing).
11 */ 11 */
12 12
13 #include <linux/list.h> 13 #include <linux/list.h>
14 #include <linux/hashtable.h> 14 #include <linux/hashtable.h>
15 #include <linux/sched/signal.h> 15 #include <linux/sched/signal.h>
16 #include <linux/sched/mm.h> 16 #include <linux/sched/mm.h>
17 #include <linux/mm.h> 17 #include <linux/mm.h>
18 #include <linux/mm_inline.h> 18 #include <linux/mm_inline.h>
19 #include <linux/mmu_notifier.h> 19 #include <linux/mmu_notifier.h>
20 #include <linux/poll.h> 20 #include <linux/poll.h>
21 #include <linux/slab.h> 21 #include <linux/slab.h>
22 #include <linux/seq_file.h> 22 #include <linux/seq_file.h>
23 #include <linux/file.h> 23 #include <linux/file.h>
24 #include <linux/bug.h> 24 #include <linux/bug.h>
25 #include <linux/anon_inodes.h> 25 #include <linux/anon_inodes.h>
26 #include <linux/syscalls.h> 26 #include <linux/syscalls.h>
27 #include <linux/userfaultfd_k.h> 27 #include <linux/userfaultfd_k.h>
28 #include <linux/mempolicy.h> 28 #include <linux/mempolicy.h>
29 #include <linux/ioctl.h> 29 #include <linux/ioctl.h>
30 #include <linux/security.h> 30 #include <linux/security.h>
31 #include <linux/hugetlb.h> 31 #include <linux/hugetlb.h>
32 #include <linux/swapops.h> 32 #include <linux/swapops.h>
33 #include <linux/miscdevice.h> 33 #include <linux/miscdevice.h>
34 #include <linux/uio.h> <<
35 34
36 static int sysctl_unprivileged_userfaultfd __r 35 static int sysctl_unprivileged_userfaultfd __read_mostly;
37 36
38 #ifdef CONFIG_SYSCTL 37 #ifdef CONFIG_SYSCTL
39 static struct ctl_table vm_userfaultfd_table[] 38 static struct ctl_table vm_userfaultfd_table[] = {
40 { 39 {
41 .procname = "unprivilege 40 .procname = "unprivileged_userfaultfd",
42 .data = &sysctl_unpr 41 .data = &sysctl_unprivileged_userfaultfd,
43 .maxlen = sizeof(sysct 42 .maxlen = sizeof(sysctl_unprivileged_userfaultfd),
44 .mode = 0644, 43 .mode = 0644,
45 .proc_handler = proc_dointve 44 .proc_handler = proc_dointvec_minmax,
46 .extra1 = SYSCTL_ZERO, 45 .extra1 = SYSCTL_ZERO,
47 .extra2 = SYSCTL_ONE, 46 .extra2 = SYSCTL_ONE,
48 }, 47 },
49 }; 48 };
50 #endif 49 #endif
51 50
52 static struct kmem_cache *userfaultfd_ctx_cach 51 static struct kmem_cache *userfaultfd_ctx_cachep __ro_after_init;
53 52
>> 53 /*
>> 54 * Start with fault_pending_wqh and fault_wqh so they're more likely
>> 55 * to be in the same cacheline.
>> 56 *
>> 57 * Locking order:
>> 58 * fd_wqh.lock
>> 59 * fault_pending_wqh.lock
>> 60 * fault_wqh.lock
>> 61 * event_wqh.lock
>> 62 *
>> 63 * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
>> 64 * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
>> 65 * also taken in IRQ context.
>> 66 */
>> 67 struct userfaultfd_ctx {
>> 68 /* waitqueue head for the pending (i.e. not read) userfaults */
>> 69 wait_queue_head_t fault_pending_wqh;
>> 70 /* waitqueue head for the userfaults */
>> 71 wait_queue_head_t fault_wqh;
>> 72 /* waitqueue head for the pseudo fd to wakeup poll/read */
>> 73 wait_queue_head_t fd_wqh;
>> 74 /* waitqueue head for events */
>> 75 wait_queue_head_t event_wqh;
>> 76 /* a refile sequence protected by fault_pending_wqh lock */
>> 77 seqcount_spinlock_t refile_seq;
>> 78 /* pseudo fd refcounting */
>> 79 refcount_t refcount;
>> 80 /* userfaultfd syscall flags */
>> 81 unsigned int flags;
>> 82 /* features requested from the userspace */
>> 83 unsigned int features;
>> 84 /* released */
>> 85 bool released;
>> 86 /* memory mappings are changing because of non-cooperative event */
>> 87 atomic_t mmap_changing;
>> 88 /* mm with one ore more vmas attached to this userfaultfd_ctx */
>> 89 struct mm_struct *mm;
>> 90 };
>> 91
54 struct userfaultfd_fork_ctx { 92 struct userfaultfd_fork_ctx {
55 struct userfaultfd_ctx *orig; 93 struct userfaultfd_ctx *orig;
56 struct userfaultfd_ctx *new; 94 struct userfaultfd_ctx *new;
57 struct list_head list; 95 struct list_head list;
58 }; 96 };
59 97
60 struct userfaultfd_unmap_ctx { 98 struct userfaultfd_unmap_ctx {
61 struct userfaultfd_ctx *ctx; 99 struct userfaultfd_ctx *ctx;
62 unsigned long start; 100 unsigned long start;
63 unsigned long end; 101 unsigned long end;
64 struct list_head list; 102 struct list_head list;
65 }; 103 };
66 104
67 struct userfaultfd_wait_queue { 105 struct userfaultfd_wait_queue {
68 struct uffd_msg msg; 106 struct uffd_msg msg;
69 wait_queue_entry_t wq; 107 wait_queue_entry_t wq;
70 struct userfaultfd_ctx *ctx; 108 struct userfaultfd_ctx *ctx;
71 bool waken; 109 bool waken;
72 }; 110 };
73 111
74 struct userfaultfd_wake_range { 112 struct userfaultfd_wake_range {
75 unsigned long start; 113 unsigned long start;
76 unsigned long len; 114 unsigned long len;
77 }; 115 };
78 116
79 /* internal indication that UFFD_API ioctl was 117 /* internal indication that UFFD_API ioctl was successfully executed */
80 #define UFFD_FEATURE_INITIALIZED 118 #define UFFD_FEATURE_INITIALIZED (1u << 31)
81 119
82 static bool userfaultfd_is_initialized(struct 120 static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx)
83 { 121 {
84 return ctx->features & UFFD_FEATURE_IN 122 return ctx->features & UFFD_FEATURE_INITIALIZED;
85 } 123 }
86 124
87 static bool userfaultfd_wp_async_ctx(struct us 125 static bool userfaultfd_wp_async_ctx(struct userfaultfd_ctx *ctx)
88 { 126 {
89 return ctx && (ctx->features & UFFD_FE 127 return ctx && (ctx->features & UFFD_FEATURE_WP_ASYNC);
90 } 128 }
91 129
92 /* 130 /*
93 * Whether WP_UNPOPULATED is enabled on the uf 131 * Whether WP_UNPOPULATED is enabled on the uffd context. It is only
94 * meaningful when userfaultfd_wp()==true on t 132 * meaningful when userfaultfd_wp()==true on the vma and when it's
95 * anonymous. 133 * anonymous.
96 */ 134 */
97 bool userfaultfd_wp_unpopulated(struct vm_area 135 bool userfaultfd_wp_unpopulated(struct vm_area_struct *vma)
98 { 136 {
99 struct userfaultfd_ctx *ctx = vma->vm_ 137 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
100 138
101 if (!ctx) 139 if (!ctx)
102 return false; 140 return false;
103 141
104 return ctx->features & UFFD_FEATURE_WP 142 return ctx->features & UFFD_FEATURE_WP_UNPOPULATED;
105 } 143 }
106 144
107 static void userfaultfd_set_vm_flags(struct vm 145 static void userfaultfd_set_vm_flags(struct vm_area_struct *vma,
108 vm_flags_ 146 vm_flags_t flags)
109 { 147 {
110 const bool uffd_wp_changed = (vma->vm_ 148 const bool uffd_wp_changed = (vma->vm_flags ^ flags) & VM_UFFD_WP;
111 149
112 vm_flags_reset(vma, flags); 150 vm_flags_reset(vma, flags);
113 /* 151 /*
114 * For shared mappings, we want to ena 152 * For shared mappings, we want to enable writenotify while
115 * userfaultfd-wp is enabled (see vma_ 153 * userfaultfd-wp is enabled (see vma_wants_writenotify()). We'll simply
116 * recalculate vma->vm_page_prot whene 154 * recalculate vma->vm_page_prot whenever userfaultfd-wp changes.
117 */ 155 */
118 if ((vma->vm_flags & VM_SHARED) && uff 156 if ((vma->vm_flags & VM_SHARED) && uffd_wp_changed)
119 vma_set_page_prot(vma); 157 vma_set_page_prot(vma);
120 } 158 }
121 159
122 static int userfaultfd_wake_function(wait_queu 160 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
123 int wake_ 161 int wake_flags, void *key)
124 { 162 {
125 struct userfaultfd_wake_range *range = 163 struct userfaultfd_wake_range *range = key;
126 int ret; 164 int ret;
127 struct userfaultfd_wait_queue *uwq; 165 struct userfaultfd_wait_queue *uwq;
128 unsigned long start, len; 166 unsigned long start, len;
129 167
130 uwq = container_of(wq, struct userfaul 168 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
131 ret = 0; 169 ret = 0;
132 /* len == 0 means wake all */ 170 /* len == 0 means wake all */
133 start = range->start; 171 start = range->start;
134 len = range->len; 172 len = range->len;
135 if (len && (start > uwq->msg.arg.pagef 173 if (len && (start > uwq->msg.arg.pagefault.address ||
136 start + len <= uwq->msg.ar 174 start + len <= uwq->msg.arg.pagefault.address))
137 goto out; 175 goto out;
138 WRITE_ONCE(uwq->waken, true); 176 WRITE_ONCE(uwq->waken, true);
139 /* 177 /*
140 * The Program-Order guarantees provid 178 * The Program-Order guarantees provided by the scheduler
141 * ensure uwq->waken is visible before 179 * ensure uwq->waken is visible before the task is woken.
142 */ 180 */
143 ret = wake_up_state(wq->private, mode) 181 ret = wake_up_state(wq->private, mode);
144 if (ret) { 182 if (ret) {
145 /* 183 /*
146 * Wake only once, autoremove 184 * Wake only once, autoremove behavior.
147 * 185 *
148 * After the effect of list_de 186 * After the effect of list_del_init is visible to the other
149 * CPUs, the waitqueue may dis 187 * CPUs, the waitqueue may disappear from under us, see the
150 * !list_empty_careful() in ha 188 * !list_empty_careful() in handle_userfault().
151 * 189 *
152 * try_to_wake_up() has an imp 190 * try_to_wake_up() has an implicit smp_mb(), and the
153 * wq->private is read before 191 * wq->private is read before calling the extern function
154 * "wake_up_state" (which in t 192 * "wake_up_state" (which in turns calls try_to_wake_up).
155 */ 193 */
156 list_del_init(&wq->entry); 194 list_del_init(&wq->entry);
157 } 195 }
158 out: 196 out:
159 return ret; 197 return ret;
160 } 198 }
161 199
162 /** 200 /**
163 * userfaultfd_ctx_get - Acquires a reference 201 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
164 * context. 202 * context.
165 * @ctx: [in] Pointer to the userfaultfd conte 203 * @ctx: [in] Pointer to the userfaultfd context.
166 */ 204 */
167 static void userfaultfd_ctx_get(struct userfau 205 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
168 { 206 {
169 refcount_inc(&ctx->refcount); 207 refcount_inc(&ctx->refcount);
170 } 208 }
171 209
172 /** 210 /**
173 * userfaultfd_ctx_put - Releases a reference 211 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
174 * context. 212 * context.
175 * @ctx: [in] Pointer to userfaultfd context. 213 * @ctx: [in] Pointer to userfaultfd context.
176 * 214 *
177 * The userfaultfd context reference must have 215 * The userfaultfd context reference must have been previously acquired either
178 * with userfaultfd_ctx_get() or userfaultfd_c 216 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
179 */ 217 */
180 static void userfaultfd_ctx_put(struct userfau 218 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
181 { 219 {
182 if (refcount_dec_and_test(&ctx->refcou 220 if (refcount_dec_and_test(&ctx->refcount)) {
183 VM_BUG_ON(spin_is_locked(&ctx- 221 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
184 VM_BUG_ON(waitqueue_active(&ct 222 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
185 VM_BUG_ON(spin_is_locked(&ctx- 223 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
186 VM_BUG_ON(waitqueue_active(&ct 224 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
187 VM_BUG_ON(spin_is_locked(&ctx- 225 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
188 VM_BUG_ON(waitqueue_active(&ct 226 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
189 VM_BUG_ON(spin_is_locked(&ctx- 227 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
190 VM_BUG_ON(waitqueue_active(&ct 228 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
191 mmdrop(ctx->mm); 229 mmdrop(ctx->mm);
192 kmem_cache_free(userfaultfd_ct 230 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
193 } 231 }
194 } 232 }
195 233
196 static inline void msg_init(struct uffd_msg *m 234 static inline void msg_init(struct uffd_msg *msg)
197 { 235 {
198 BUILD_BUG_ON(sizeof(struct uffd_msg) ! 236 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
199 /* 237 /*
200 * Must use memset to zero out the pad 238 * Must use memset to zero out the paddings or kernel data is
201 * leaked to userland. 239 * leaked to userland.
202 */ 240 */
203 memset(msg, 0, sizeof(struct uffd_msg) 241 memset(msg, 0, sizeof(struct uffd_msg));
204 } 242 }
205 243
206 static inline struct uffd_msg userfault_msg(un 244 static inline struct uffd_msg userfault_msg(unsigned long address,
207 un 245 unsigned long real_address,
208 un 246 unsigned int flags,
209 un 247 unsigned long reason,
210 un 248 unsigned int features)
211 { 249 {
212 struct uffd_msg msg; 250 struct uffd_msg msg;
213 251
214 msg_init(&msg); 252 msg_init(&msg);
215 msg.event = UFFD_EVENT_PAGEFAULT; 253 msg.event = UFFD_EVENT_PAGEFAULT;
216 254
217 msg.arg.pagefault.address = (features 255 msg.arg.pagefault.address = (features & UFFD_FEATURE_EXACT_ADDRESS) ?
218 real_addre 256 real_address : address;
219 257
220 /* 258 /*
221 * These flags indicate why the userfa 259 * These flags indicate why the userfault occurred:
222 * - UFFD_PAGEFAULT_FLAG_WP indicates 260 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
223 * - UFFD_PAGEFAULT_FLAG_MINOR indicat 261 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
224 * - Neither of these flags being set 262 * - Neither of these flags being set indicates a MISSING fault.
225 * 263 *
226 * Separately, UFFD_PAGEFAULT_FLAG_WRI 264 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
227 * fault. Otherwise, it was a read fau 265 * fault. Otherwise, it was a read fault.
228 */ 266 */
229 if (flags & FAULT_FLAG_WRITE) 267 if (flags & FAULT_FLAG_WRITE)
230 msg.arg.pagefault.flags |= UFF 268 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
231 if (reason & VM_UFFD_WP) 269 if (reason & VM_UFFD_WP)
232 msg.arg.pagefault.flags |= UFF 270 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
233 if (reason & VM_UFFD_MINOR) 271 if (reason & VM_UFFD_MINOR)
234 msg.arg.pagefault.flags |= UFF 272 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
235 if (features & UFFD_FEATURE_THREAD_ID) 273 if (features & UFFD_FEATURE_THREAD_ID)
236 msg.arg.pagefault.feat.ptid = 274 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
237 return msg; 275 return msg;
238 } 276 }
239 277
240 #ifdef CONFIG_HUGETLB_PAGE 278 #ifdef CONFIG_HUGETLB_PAGE
241 /* 279 /*
242 * Same functionality as userfaultfd_must_wait 280 * Same functionality as userfaultfd_must_wait below with modifications for
243 * hugepmd ranges. 281 * hugepmd ranges.
244 */ 282 */
245 static inline bool userfaultfd_huge_must_wait( 283 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
246 284 struct vm_fault *vmf,
247 285 unsigned long reason)
248 { 286 {
249 struct vm_area_struct *vma = vmf->vma; 287 struct vm_area_struct *vma = vmf->vma;
250 pte_t *ptep, pte; 288 pte_t *ptep, pte;
251 bool ret = true; 289 bool ret = true;
252 290
253 assert_fault_locked(vmf); 291 assert_fault_locked(vmf);
254 292
255 ptep = hugetlb_walk(vma, vmf->address, 293 ptep = hugetlb_walk(vma, vmf->address, vma_mmu_pagesize(vma));
256 if (!ptep) 294 if (!ptep)
257 goto out; 295 goto out;
258 296
259 ret = false; 297 ret = false;
260 pte = huge_ptep_get(vma->vm_mm, vmf->a !! 298 pte = huge_ptep_get(ptep);
261 299
262 /* 300 /*
263 * Lockless access: we're in a wait_ev 301 * Lockless access: we're in a wait_event so it's ok if it
264 * changes under us. PTE markers shou 302 * changes under us. PTE markers should be handled the same as none
265 * ptes here. 303 * ptes here.
266 */ 304 */
267 if (huge_pte_none_mostly(pte)) 305 if (huge_pte_none_mostly(pte))
268 ret = true; 306 ret = true;
269 if (!huge_pte_write(pte) && (reason & 307 if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
270 ret = true; 308 ret = true;
271 out: 309 out:
272 return ret; 310 return ret;
273 } 311 }
274 #else 312 #else
275 static inline bool userfaultfd_huge_must_wait( 313 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
276 314 struct vm_fault *vmf,
277 315 unsigned long reason)
278 { 316 {
279 return false; /* should never get he 317 return false; /* should never get here */
280 } 318 }
281 #endif /* CONFIG_HUGETLB_PAGE */ 319 #endif /* CONFIG_HUGETLB_PAGE */
282 320
283 /* 321 /*
284 * Verify the pagetables are still not ok afte 322 * Verify the pagetables are still not ok after having reigstered into
285 * the fault_pending_wqh to avoid userland hav 323 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
286 * userfault that has already been resolved, i !! 324 * userfault that has already been resolved, if userfaultfd_read and
287 * UFFDIO_COPY|ZEROPAGE are being run simultan 325 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
288 * threads. 326 * threads.
289 */ 327 */
290 static inline bool userfaultfd_must_wait(struc 328 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
291 struc 329 struct vm_fault *vmf,
292 unsig 330 unsigned long reason)
293 { 331 {
294 struct mm_struct *mm = ctx->mm; 332 struct mm_struct *mm = ctx->mm;
295 unsigned long address = vmf->address; 333 unsigned long address = vmf->address;
296 pgd_t *pgd; 334 pgd_t *pgd;
297 p4d_t *p4d; 335 p4d_t *p4d;
298 pud_t *pud; 336 pud_t *pud;
299 pmd_t *pmd, _pmd; 337 pmd_t *pmd, _pmd;
300 pte_t *pte; 338 pte_t *pte;
301 pte_t ptent; 339 pte_t ptent;
302 bool ret = true; 340 bool ret = true;
303 341
304 assert_fault_locked(vmf); 342 assert_fault_locked(vmf);
305 343
306 pgd = pgd_offset(mm, address); 344 pgd = pgd_offset(mm, address);
307 if (!pgd_present(*pgd)) 345 if (!pgd_present(*pgd))
308 goto out; 346 goto out;
309 p4d = p4d_offset(pgd, address); 347 p4d = p4d_offset(pgd, address);
310 if (!p4d_present(*p4d)) 348 if (!p4d_present(*p4d))
311 goto out; 349 goto out;
312 pud = pud_offset(p4d, address); 350 pud = pud_offset(p4d, address);
313 if (!pud_present(*pud)) 351 if (!pud_present(*pud))
314 goto out; 352 goto out;
315 pmd = pmd_offset(pud, address); 353 pmd = pmd_offset(pud, address);
316 again: 354 again:
317 _pmd = pmdp_get_lockless(pmd); 355 _pmd = pmdp_get_lockless(pmd);
318 if (pmd_none(_pmd)) 356 if (pmd_none(_pmd))
319 goto out; 357 goto out;
320 358
321 ret = false; 359 ret = false;
322 if (!pmd_present(_pmd) || pmd_devmap(_ 360 if (!pmd_present(_pmd) || pmd_devmap(_pmd))
323 goto out; 361 goto out;
324 362
325 if (pmd_trans_huge(_pmd)) { 363 if (pmd_trans_huge(_pmd)) {
326 if (!pmd_write(_pmd) && (reaso 364 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
327 ret = true; 365 ret = true;
328 goto out; 366 goto out;
329 } 367 }
330 368
331 pte = pte_offset_map(pmd, address); 369 pte = pte_offset_map(pmd, address);
332 if (!pte) { 370 if (!pte) {
333 ret = true; 371 ret = true;
334 goto again; 372 goto again;
335 } 373 }
336 /* 374 /*
337 * Lockless access: we're in a wait_ev 375 * Lockless access: we're in a wait_event so it's ok if it
338 * changes under us. PTE markers shou 376 * changes under us. PTE markers should be handled the same as none
339 * ptes here. 377 * ptes here.
340 */ 378 */
341 ptent = ptep_get(pte); 379 ptent = ptep_get(pte);
342 if (pte_none_mostly(ptent)) 380 if (pte_none_mostly(ptent))
343 ret = true; 381 ret = true;
344 if (!pte_write(ptent) && (reason & VM_ 382 if (!pte_write(ptent) && (reason & VM_UFFD_WP))
345 ret = true; 383 ret = true;
346 pte_unmap(pte); 384 pte_unmap(pte);
347 385
348 out: 386 out:
349 return ret; 387 return ret;
350 } 388 }
351 389
352 static inline unsigned int userfaultfd_get_blo 390 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags)
353 { 391 {
354 if (flags & FAULT_FLAG_INTERRUPTIBLE) 392 if (flags & FAULT_FLAG_INTERRUPTIBLE)
355 return TASK_INTERRUPTIBLE; 393 return TASK_INTERRUPTIBLE;
356 394
357 if (flags & FAULT_FLAG_KILLABLE) 395 if (flags & FAULT_FLAG_KILLABLE)
358 return TASK_KILLABLE; 396 return TASK_KILLABLE;
359 397
360 return TASK_UNINTERRUPTIBLE; 398 return TASK_UNINTERRUPTIBLE;
361 } 399 }
362 400
363 /* 401 /*
364 * The locking rules involved in returning VM_ 402 * The locking rules involved in returning VM_FAULT_RETRY depending on
365 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NO 403 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
366 * FAULT_FLAG_KILLABLE are not straightforward 404 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
367 * recommendation in __lock_page_or_retry is n 405 * recommendation in __lock_page_or_retry is not an understatement.
368 * 406 *
369 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_ 407 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
370 * before returning VM_FAULT_RETRY only if FAU 408 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
371 * not set. 409 * not set.
372 * 410 *
373 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_ 411 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
374 * set, VM_FAULT_RETRY can still be returned i 412 * set, VM_FAULT_RETRY can still be returned if and only if there are
375 * fatal_signal_pending()s, and the mmap_lock 413 * fatal_signal_pending()s, and the mmap_lock must be released before
376 * returning it. 414 * returning it.
377 */ 415 */
378 vm_fault_t handle_userfault(struct vm_fault *v 416 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
379 { 417 {
380 struct vm_area_struct *vma = vmf->vma; 418 struct vm_area_struct *vma = vmf->vma;
381 struct mm_struct *mm = vma->vm_mm; 419 struct mm_struct *mm = vma->vm_mm;
382 struct userfaultfd_ctx *ctx; 420 struct userfaultfd_ctx *ctx;
383 struct userfaultfd_wait_queue uwq; 421 struct userfaultfd_wait_queue uwq;
384 vm_fault_t ret = VM_FAULT_SIGBUS; 422 vm_fault_t ret = VM_FAULT_SIGBUS;
385 bool must_wait; 423 bool must_wait;
386 unsigned int blocking_state; 424 unsigned int blocking_state;
387 425
388 /* 426 /*
389 * We don't do userfault handling for 427 * We don't do userfault handling for the final child pid update.
390 * 428 *
391 * We also don't do userfault handling 429 * We also don't do userfault handling during
392 * coredumping. hugetlbfs has the spec 430 * coredumping. hugetlbfs has the special
393 * hugetlb_follow_page_mask() to skip 431 * hugetlb_follow_page_mask() to skip missing pages in the
394 * FOLL_DUMP case, anon memory also ch 432 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
395 * the no_page_table() helper in follo 433 * the no_page_table() helper in follow_page_mask(), but the
396 * shmem_vm_ops->fault method is invok 434 * shmem_vm_ops->fault method is invoked even during
397 * coredumping and it ends up here. 435 * coredumping and it ends up here.
398 */ 436 */
399 if (current->flags & (PF_EXITING|PF_DU 437 if (current->flags & (PF_EXITING|PF_DUMPCORE))
400 goto out; 438 goto out;
401 439
402 assert_fault_locked(vmf); 440 assert_fault_locked(vmf);
403 441
404 ctx = vma->vm_userfaultfd_ctx.ctx; 442 ctx = vma->vm_userfaultfd_ctx.ctx;
405 if (!ctx) 443 if (!ctx)
406 goto out; 444 goto out;
407 445
408 BUG_ON(ctx->mm != mm); 446 BUG_ON(ctx->mm != mm);
409 447
410 /* Any unrecognized flag is a bug. */ 448 /* Any unrecognized flag is a bug. */
411 VM_BUG_ON(reason & ~__VM_UFFD_FLAGS); 449 VM_BUG_ON(reason & ~__VM_UFFD_FLAGS);
412 /* 0 or > 1 flags set is a bug; we exp 450 /* 0 or > 1 flags set is a bug; we expect exactly 1. */
413 VM_BUG_ON(!reason || (reason & (reason 451 VM_BUG_ON(!reason || (reason & (reason - 1)));
414 452
415 if (ctx->features & UFFD_FEATURE_SIGBU 453 if (ctx->features & UFFD_FEATURE_SIGBUS)
416 goto out; 454 goto out;
417 if (!(vmf->flags & FAULT_FLAG_USER) && 455 if (!(vmf->flags & FAULT_FLAG_USER) && (ctx->flags & UFFD_USER_MODE_ONLY))
418 goto out; 456 goto out;
419 457
420 /* 458 /*
421 * If it's already released don't get 459 * If it's already released don't get it. This avoids to loop
422 * in __get_user_pages if userfaultfd_ 460 * in __get_user_pages if userfaultfd_release waits on the
423 * caller of handle_userfault to relea 461 * caller of handle_userfault to release the mmap_lock.
424 */ 462 */
425 if (unlikely(READ_ONCE(ctx->released)) 463 if (unlikely(READ_ONCE(ctx->released))) {
426 /* 464 /*
427 * Don't return VM_FAULT_SIGBU 465 * Don't return VM_FAULT_SIGBUS in this case, so a non
428 * cooperative manager can clo 466 * cooperative manager can close the uffd after the
429 * last UFFDIO_COPY, without r 467 * last UFFDIO_COPY, without risking to trigger an
430 * involuntary SIGBUS if the p 468 * involuntary SIGBUS if the process was starting the
431 * userfaultfd while the userf 469 * userfaultfd while the userfaultfd was still armed
432 * (but after the last UFFDIO_ 470 * (but after the last UFFDIO_COPY). If the uffd
433 * wasn't already closed when 471 * wasn't already closed when the userfault reached
434 * this point, that would norm 472 * this point, that would normally be solved by
435 * userfaultfd_must_wait retur 473 * userfaultfd_must_wait returning 'false'.
436 * 474 *
437 * If we were to return VM_FAU 475 * If we were to return VM_FAULT_SIGBUS here, the non
438 * cooperative manager would b 476 * cooperative manager would be instead forced to
439 * always call UFFDIO_UNREGIST 477 * always call UFFDIO_UNREGISTER before it can safely
440 * close the uffd. 478 * close the uffd.
441 */ 479 */
442 ret = VM_FAULT_NOPAGE; 480 ret = VM_FAULT_NOPAGE;
443 goto out; 481 goto out;
444 } 482 }
445 483
446 /* 484 /*
447 * Check that we can return VM_FAULT_R 485 * Check that we can return VM_FAULT_RETRY.
448 * 486 *
449 * NOTE: it should become possible to 487 * NOTE: it should become possible to return VM_FAULT_RETRY
450 * even if FAULT_FLAG_TRIED is set wit 488 * even if FAULT_FLAG_TRIED is set without leading to gup()
451 * -EBUSY failures, if the userfaultfd 489 * -EBUSY failures, if the userfaultfd is to be extended for
452 * VM_UFFD_WP tracking and we intend t 490 * VM_UFFD_WP tracking and we intend to arm the userfault
453 * without first stopping userland acc 491 * without first stopping userland access to the memory. For
454 * VM_UFFD_MISSING userfaults this is 492 * VM_UFFD_MISSING userfaults this is enough for now.
455 */ 493 */
456 if (unlikely(!(vmf->flags & FAULT_FLAG 494 if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
457 /* 495 /*
458 * Validate the invariant that 496 * Validate the invariant that nowait must allow retry
459 * to be sure not to return SI 497 * to be sure not to return SIGBUS erroneously on
460 * nowait invocations. 498 * nowait invocations.
461 */ 499 */
462 BUG_ON(vmf->flags & FAULT_FLAG 500 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
463 #ifdef CONFIG_DEBUG_VM 501 #ifdef CONFIG_DEBUG_VM
464 if (printk_ratelimit()) { 502 if (printk_ratelimit()) {
465 printk(KERN_WARNING 503 printk(KERN_WARNING
466 "FAULT_FLAG_ALL 504 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
467 vmf->flags); 505 vmf->flags);
468 dump_stack(); 506 dump_stack();
469 } 507 }
470 #endif 508 #endif
471 goto out; 509 goto out;
472 } 510 }
473 511
474 /* 512 /*
475 * Handle nowait, not much to do other 513 * Handle nowait, not much to do other than tell it to retry
476 * and wait. 514 * and wait.
477 */ 515 */
478 ret = VM_FAULT_RETRY; 516 ret = VM_FAULT_RETRY;
479 if (vmf->flags & FAULT_FLAG_RETRY_NOWA 517 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
480 goto out; 518 goto out;
481 519
482 /* take the reference before dropping 520 /* take the reference before dropping the mmap_lock */
483 userfaultfd_ctx_get(ctx); 521 userfaultfd_ctx_get(ctx);
484 522
485 init_waitqueue_func_entry(&uwq.wq, use 523 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
486 uwq.wq.private = current; 524 uwq.wq.private = current;
487 uwq.msg = userfault_msg(vmf->address, 525 uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags,
488 reason, ctx->f 526 reason, ctx->features);
489 uwq.ctx = ctx; 527 uwq.ctx = ctx;
490 uwq.waken = false; 528 uwq.waken = false;
491 529
492 blocking_state = userfaultfd_get_block 530 blocking_state = userfaultfd_get_blocking_state(vmf->flags);
493 531
494 /* 532 /*
495 * Take the vma lock now, in order to 533 * Take the vma lock now, in order to safely call
496 * userfaultfd_huge_must_wait() later. 534 * userfaultfd_huge_must_wait() later. Since acquiring the
497 * (sleepable) vma lock can modify the 535 * (sleepable) vma lock can modify the current task state, that
498 * must be before explicitly calling s 536 * must be before explicitly calling set_current_state().
499 */ 537 */
500 if (is_vm_hugetlb_page(vma)) 538 if (is_vm_hugetlb_page(vma))
501 hugetlb_vma_lock_read(vma); 539 hugetlb_vma_lock_read(vma);
502 540
503 spin_lock_irq(&ctx->fault_pending_wqh. 541 spin_lock_irq(&ctx->fault_pending_wqh.lock);
504 /* 542 /*
505 * After the __add_wait_queue the uwq 543 * After the __add_wait_queue the uwq is visible to userland
506 * through poll/read(). 544 * through poll/read().
507 */ 545 */
508 __add_wait_queue(&ctx->fault_pending_w 546 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
509 /* 547 /*
510 * The smp_mb() after __set_current_st 548 * The smp_mb() after __set_current_state prevents the reads
511 * following the spin_unlock to happen 549 * following the spin_unlock to happen before the list_add in
512 * __add_wait_queue. 550 * __add_wait_queue.
513 */ 551 */
514 set_current_state(blocking_state); 552 set_current_state(blocking_state);
515 spin_unlock_irq(&ctx->fault_pending_wq 553 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
516 554
517 if (!is_vm_hugetlb_page(vma)) 555 if (!is_vm_hugetlb_page(vma))
518 must_wait = userfaultfd_must_w 556 must_wait = userfaultfd_must_wait(ctx, vmf, reason);
519 else 557 else
520 must_wait = userfaultfd_huge_m 558 must_wait = userfaultfd_huge_must_wait(ctx, vmf, reason);
521 if (is_vm_hugetlb_page(vma)) 559 if (is_vm_hugetlb_page(vma))
522 hugetlb_vma_unlock_read(vma); 560 hugetlb_vma_unlock_read(vma);
523 release_fault_lock(vmf); 561 release_fault_lock(vmf);
524 562
525 if (likely(must_wait && !READ_ONCE(ctx 563 if (likely(must_wait && !READ_ONCE(ctx->released))) {
526 wake_up_poll(&ctx->fd_wqh, EPO 564 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
527 schedule(); 565 schedule();
528 } 566 }
529 567
530 __set_current_state(TASK_RUNNING); 568 __set_current_state(TASK_RUNNING);
531 569
532 /* 570 /*
533 * Here we race with the list_del; lis 571 * Here we race with the list_del; list_add in
534 * userfaultfd_ctx_read(), however bec 572 * userfaultfd_ctx_read(), however because we don't ever run
535 * list_del_init() to refile across th 573 * list_del_init() to refile across the two lists, the prev
536 * and next pointers will never point 574 * and next pointers will never point to self. list_add also
537 * would never let any of the two poin 575 * would never let any of the two pointers to point to
538 * self. So list_empty_careful won't r 576 * self. So list_empty_careful won't risk to see both pointers
539 * pointing to self at any time during 577 * pointing to self at any time during the list refile. The
540 * only case where list_del_init() is 578 * only case where list_del_init() is called is the full
541 * removal in the wake function and th 579 * removal in the wake function and there we don't re-list_add
542 * and it's fine not to block on the s 580 * and it's fine not to block on the spinlock. The uwq on this
543 * kernel stack can be released after 581 * kernel stack can be released after the list_del_init.
544 */ 582 */
545 if (!list_empty_careful(&uwq.wq.entry) 583 if (!list_empty_careful(&uwq.wq.entry)) {
546 spin_lock_irq(&ctx->fault_pend 584 spin_lock_irq(&ctx->fault_pending_wqh.lock);
547 /* 585 /*
548 * No need of list_del_init(), 586 * No need of list_del_init(), the uwq on the stack
549 * will be freed shortly anywa 587 * will be freed shortly anyway.
550 */ 588 */
551 list_del(&uwq.wq.entry); 589 list_del(&uwq.wq.entry);
552 spin_unlock_irq(&ctx->fault_pe 590 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
553 } 591 }
554 592
555 /* 593 /*
556 * ctx may go away after this if the u 594 * ctx may go away after this if the userfault pseudo fd is
557 * already released. 595 * already released.
558 */ 596 */
559 userfaultfd_ctx_put(ctx); 597 userfaultfd_ctx_put(ctx);
560 598
561 out: 599 out:
562 return ret; 600 return ret;
563 } 601 }
564 602
565 static void userfaultfd_event_wait_completion( 603 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
566 604 struct userfaultfd_wait_queue *ewq)
567 { 605 {
568 struct userfaultfd_ctx *release_new_ct 606 struct userfaultfd_ctx *release_new_ctx;
569 607
570 if (WARN_ON_ONCE(current->flags & PF_E 608 if (WARN_ON_ONCE(current->flags & PF_EXITING))
571 goto out; 609 goto out;
572 610
573 ewq->ctx = ctx; 611 ewq->ctx = ctx;
574 init_waitqueue_entry(&ewq->wq, current 612 init_waitqueue_entry(&ewq->wq, current);
575 release_new_ctx = NULL; 613 release_new_ctx = NULL;
576 614
577 spin_lock_irq(&ctx->event_wqh.lock); 615 spin_lock_irq(&ctx->event_wqh.lock);
578 /* 616 /*
579 * After the __add_wait_queue the uwq 617 * After the __add_wait_queue the uwq is visible to userland
580 * through poll/read(). 618 * through poll/read().
581 */ 619 */
582 __add_wait_queue(&ctx->event_wqh, &ewq 620 __add_wait_queue(&ctx->event_wqh, &ewq->wq);
583 for (;;) { 621 for (;;) {
584 set_current_state(TASK_KILLABL 622 set_current_state(TASK_KILLABLE);
585 if (ewq->msg.event == 0) 623 if (ewq->msg.event == 0)
586 break; 624 break;
587 if (READ_ONCE(ctx->released) | 625 if (READ_ONCE(ctx->released) ||
588 fatal_signal_pending(curre 626 fatal_signal_pending(current)) {
589 /* 627 /*
590 * &ewq->wq may be que 628 * &ewq->wq may be queued in fork_event, but
591 * __remove_wait_queue 629 * __remove_wait_queue ignores the head
592 * parameter. It would 630 * parameter. It would be a problem if it
593 * didn't. 631 * didn't.
594 */ 632 */
595 __remove_wait_queue(&c 633 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
596 if (ewq->msg.event == 634 if (ewq->msg.event == UFFD_EVENT_FORK) {
597 struct userfau 635 struct userfaultfd_ctx *new;
598 636
599 new = (struct 637 new = (struct userfaultfd_ctx *)
600 (unsig 638 (unsigned long)
601 ewq->m 639 ewq->msg.arg.reserved.reserved1;
602 release_new_ct 640 release_new_ctx = new;
603 } 641 }
604 break; 642 break;
605 } 643 }
606 644
607 spin_unlock_irq(&ctx->event_wq 645 spin_unlock_irq(&ctx->event_wqh.lock);
608 646
609 wake_up_poll(&ctx->fd_wqh, EPO 647 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
610 schedule(); 648 schedule();
611 649
612 spin_lock_irq(&ctx->event_wqh. 650 spin_lock_irq(&ctx->event_wqh.lock);
613 } 651 }
614 __set_current_state(TASK_RUNNING); 652 __set_current_state(TASK_RUNNING);
615 spin_unlock_irq(&ctx->event_wqh.lock); 653 spin_unlock_irq(&ctx->event_wqh.lock);
616 654
617 if (release_new_ctx) { 655 if (release_new_ctx) {
618 struct vm_area_struct *vma; 656 struct vm_area_struct *vma;
619 struct mm_struct *mm = release 657 struct mm_struct *mm = release_new_ctx->mm;
620 VMA_ITERATOR(vmi, mm, 0); 658 VMA_ITERATOR(vmi, mm, 0);
621 659
622 /* the various vma->vm_userfau 660 /* the various vma->vm_userfaultfd_ctx still points to it */
623 mmap_write_lock(mm); 661 mmap_write_lock(mm);
624 for_each_vma(vmi, vma) { 662 for_each_vma(vmi, vma) {
625 if (vma->vm_userfaultf 663 if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
626 vma_start_writ 664 vma_start_write(vma);
627 vma->vm_userfa 665 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
628 userfaultfd_se 666 userfaultfd_set_vm_flags(vma,
629 667 vma->vm_flags & ~__VM_UFFD_FLAGS);
630 } 668 }
631 } 669 }
632 mmap_write_unlock(mm); 670 mmap_write_unlock(mm);
633 671
634 userfaultfd_ctx_put(release_ne 672 userfaultfd_ctx_put(release_new_ctx);
635 } 673 }
636 674
637 /* 675 /*
638 * ctx may go away after this if the u 676 * ctx may go away after this if the userfault pseudo fd is
639 * already released. 677 * already released.
640 */ 678 */
641 out: 679 out:
642 atomic_dec(&ctx->mmap_changing); 680 atomic_dec(&ctx->mmap_changing);
643 VM_BUG_ON(atomic_read(&ctx->mmap_chang 681 VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0);
644 userfaultfd_ctx_put(ctx); 682 userfaultfd_ctx_put(ctx);
645 } 683 }
646 684
647 static void userfaultfd_event_complete(struct 685 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
648 struct 686 struct userfaultfd_wait_queue *ewq)
649 { 687 {
650 ewq->msg.event = 0; 688 ewq->msg.event = 0;
651 wake_up_locked(&ctx->event_wqh); 689 wake_up_locked(&ctx->event_wqh);
652 __remove_wait_queue(&ctx->event_wqh, & 690 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
653 } 691 }
654 692
655 int dup_userfaultfd(struct vm_area_struct *vma 693 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
656 { 694 {
657 struct userfaultfd_ctx *ctx = NULL, *o 695 struct userfaultfd_ctx *ctx = NULL, *octx;
658 struct userfaultfd_fork_ctx *fctx; 696 struct userfaultfd_fork_ctx *fctx;
659 697
660 octx = vma->vm_userfaultfd_ctx.ctx; 698 octx = vma->vm_userfaultfd_ctx.ctx;
661 if (!octx) !! 699 if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
662 return 0; <<
663 <<
664 if (!(octx->features & UFFD_FEATURE_EV <<
665 vma_start_write(vma); 700 vma_start_write(vma);
666 vma->vm_userfaultfd_ctx = NULL 701 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
667 userfaultfd_set_vm_flags(vma, 702 userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
668 return 0; 703 return 0;
669 } 704 }
670 705
671 list_for_each_entry(fctx, fcs, list) 706 list_for_each_entry(fctx, fcs, list)
672 if (fctx->orig == octx) { 707 if (fctx->orig == octx) {
673 ctx = fctx->new; 708 ctx = fctx->new;
674 break; 709 break;
675 } 710 }
676 711
677 if (!ctx) { 712 if (!ctx) {
678 fctx = kmalloc(sizeof(*fctx), 713 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
679 if (!fctx) 714 if (!fctx)
680 return -ENOMEM; 715 return -ENOMEM;
681 716
682 ctx = kmem_cache_alloc(userfau 717 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
683 if (!ctx) { 718 if (!ctx) {
684 kfree(fctx); 719 kfree(fctx);
685 return -ENOMEM; 720 return -ENOMEM;
686 } 721 }
687 722
688 refcount_set(&ctx->refcount, 1 723 refcount_set(&ctx->refcount, 1);
689 ctx->flags = octx->flags; 724 ctx->flags = octx->flags;
690 ctx->features = octx->features 725 ctx->features = octx->features;
691 ctx->released = false; 726 ctx->released = false;
692 init_rwsem(&ctx->map_changing_ <<
693 atomic_set(&ctx->mmap_changing 727 atomic_set(&ctx->mmap_changing, 0);
694 ctx->mm = vma->vm_mm; 728 ctx->mm = vma->vm_mm;
695 mmgrab(ctx->mm); 729 mmgrab(ctx->mm);
696 730
697 userfaultfd_ctx_get(octx); 731 userfaultfd_ctx_get(octx);
698 down_write(&octx->map_changing <<
699 atomic_inc(&octx->mmap_changin 732 atomic_inc(&octx->mmap_changing);
700 up_write(&octx->map_changing_l <<
701 fctx->orig = octx; 733 fctx->orig = octx;
702 fctx->new = ctx; 734 fctx->new = ctx;
703 list_add_tail(&fctx->list, fcs 735 list_add_tail(&fctx->list, fcs);
704 } 736 }
705 737
706 vma->vm_userfaultfd_ctx.ctx = ctx; 738 vma->vm_userfaultfd_ctx.ctx = ctx;
707 return 0; 739 return 0;
708 } 740 }
709 741
710 static void dup_fctx(struct userfaultfd_fork_c 742 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
711 { 743 {
712 struct userfaultfd_ctx *ctx = fctx->or 744 struct userfaultfd_ctx *ctx = fctx->orig;
713 struct userfaultfd_wait_queue ewq; 745 struct userfaultfd_wait_queue ewq;
714 746
715 msg_init(&ewq.msg); 747 msg_init(&ewq.msg);
716 748
717 ewq.msg.event = UFFD_EVENT_FORK; 749 ewq.msg.event = UFFD_EVENT_FORK;
718 ewq.msg.arg.reserved.reserved1 = (unsi 750 ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
719 751
720 userfaultfd_event_wait_completion(ctx, 752 userfaultfd_event_wait_completion(ctx, &ewq);
721 } 753 }
722 754
723 void dup_userfaultfd_complete(struct list_head 755 void dup_userfaultfd_complete(struct list_head *fcs)
724 { 756 {
725 struct userfaultfd_fork_ctx *fctx, *n; 757 struct userfaultfd_fork_ctx *fctx, *n;
726 758
727 list_for_each_entry_safe(fctx, n, fcs, 759 list_for_each_entry_safe(fctx, n, fcs, list) {
728 dup_fctx(fctx); 760 dup_fctx(fctx);
729 list_del(&fctx->list); 761 list_del(&fctx->list);
730 kfree(fctx); 762 kfree(fctx);
731 } 763 }
732 } 764 }
733 765
734 void mremap_userfaultfd_prep(struct vm_area_st 766 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
735 struct vm_userfau 767 struct vm_userfaultfd_ctx *vm_ctx)
736 { 768 {
737 struct userfaultfd_ctx *ctx; 769 struct userfaultfd_ctx *ctx;
738 770
739 ctx = vma->vm_userfaultfd_ctx.ctx; 771 ctx = vma->vm_userfaultfd_ctx.ctx;
740 772
741 if (!ctx) 773 if (!ctx)
742 return; 774 return;
743 775
744 if (ctx->features & UFFD_FEATURE_EVENT 776 if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
745 vm_ctx->ctx = ctx; 777 vm_ctx->ctx = ctx;
746 userfaultfd_ctx_get(ctx); 778 userfaultfd_ctx_get(ctx);
747 down_write(&ctx->map_changing_ <<
748 atomic_inc(&ctx->mmap_changing 779 atomic_inc(&ctx->mmap_changing);
749 up_write(&ctx->map_changing_lo <<
750 } else { 780 } else {
751 /* Drop uffd context if remap 781 /* Drop uffd context if remap feature not enabled */
752 vma_start_write(vma); 782 vma_start_write(vma);
753 vma->vm_userfaultfd_ctx = NULL 783 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
754 userfaultfd_set_vm_flags(vma, 784 userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
755 } 785 }
756 } 786 }
757 787
758 void mremap_userfaultfd_complete(struct vm_use 788 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
759 unsigned long 789 unsigned long from, unsigned long to,
760 unsigned long 790 unsigned long len)
761 { 791 {
762 struct userfaultfd_ctx *ctx = vm_ctx-> 792 struct userfaultfd_ctx *ctx = vm_ctx->ctx;
763 struct userfaultfd_wait_queue ewq; 793 struct userfaultfd_wait_queue ewq;
764 794
765 if (!ctx) 795 if (!ctx)
766 return; 796 return;
767 797
768 if (to & ~PAGE_MASK) { 798 if (to & ~PAGE_MASK) {
769 userfaultfd_ctx_put(ctx); 799 userfaultfd_ctx_put(ctx);
770 return; 800 return;
771 } 801 }
772 802
773 msg_init(&ewq.msg); 803 msg_init(&ewq.msg);
774 804
775 ewq.msg.event = UFFD_EVENT_REMAP; 805 ewq.msg.event = UFFD_EVENT_REMAP;
776 ewq.msg.arg.remap.from = from; 806 ewq.msg.arg.remap.from = from;
777 ewq.msg.arg.remap.to = to; 807 ewq.msg.arg.remap.to = to;
778 ewq.msg.arg.remap.len = len; 808 ewq.msg.arg.remap.len = len;
779 809
780 userfaultfd_event_wait_completion(ctx, 810 userfaultfd_event_wait_completion(ctx, &ewq);
781 } 811 }
782 812
783 bool userfaultfd_remove(struct vm_area_struct 813 bool userfaultfd_remove(struct vm_area_struct *vma,
784 unsigned long start, u 814 unsigned long start, unsigned long end)
785 { 815 {
786 struct mm_struct *mm = vma->vm_mm; 816 struct mm_struct *mm = vma->vm_mm;
787 struct userfaultfd_ctx *ctx; 817 struct userfaultfd_ctx *ctx;
788 struct userfaultfd_wait_queue ewq; 818 struct userfaultfd_wait_queue ewq;
789 819
790 ctx = vma->vm_userfaultfd_ctx.ctx; 820 ctx = vma->vm_userfaultfd_ctx.ctx;
791 if (!ctx || !(ctx->features & UFFD_FEA 821 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
792 return true; 822 return true;
793 823
794 userfaultfd_ctx_get(ctx); 824 userfaultfd_ctx_get(ctx);
795 down_write(&ctx->map_changing_lock); <<
796 atomic_inc(&ctx->mmap_changing); 825 atomic_inc(&ctx->mmap_changing);
797 up_write(&ctx->map_changing_lock); <<
798 mmap_read_unlock(mm); 826 mmap_read_unlock(mm);
799 827
800 msg_init(&ewq.msg); 828 msg_init(&ewq.msg);
801 829
802 ewq.msg.event = UFFD_EVENT_REMOVE; 830 ewq.msg.event = UFFD_EVENT_REMOVE;
803 ewq.msg.arg.remove.start = start; 831 ewq.msg.arg.remove.start = start;
804 ewq.msg.arg.remove.end = end; 832 ewq.msg.arg.remove.end = end;
805 833
806 userfaultfd_event_wait_completion(ctx, 834 userfaultfd_event_wait_completion(ctx, &ewq);
807 835
808 return false; 836 return false;
809 } 837 }
810 838
811 static bool has_unmap_ctx(struct userfaultfd_c 839 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
812 unsigned long start, 840 unsigned long start, unsigned long end)
813 { 841 {
814 struct userfaultfd_unmap_ctx *unmap_ct 842 struct userfaultfd_unmap_ctx *unmap_ctx;
815 843
816 list_for_each_entry(unmap_ctx, unmaps, 844 list_for_each_entry(unmap_ctx, unmaps, list)
817 if (unmap_ctx->ctx == ctx && u 845 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
818 unmap_ctx->end == end) 846 unmap_ctx->end == end)
819 return true; 847 return true;
820 848
821 return false; 849 return false;
822 } 850 }
823 851
824 int userfaultfd_unmap_prep(struct vm_area_stru 852 int userfaultfd_unmap_prep(struct vm_area_struct *vma, unsigned long start,
825 unsigned long end, 853 unsigned long end, struct list_head *unmaps)
826 { 854 {
827 struct userfaultfd_unmap_ctx *unmap_ct 855 struct userfaultfd_unmap_ctx *unmap_ctx;
828 struct userfaultfd_ctx *ctx = vma->vm_ 856 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
829 857
830 if (!ctx || !(ctx->features & UFFD_FEA 858 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
831 has_unmap_ctx(ctx, unmaps, start, 859 has_unmap_ctx(ctx, unmaps, start, end))
832 return 0; 860 return 0;
833 861
834 unmap_ctx = kzalloc(sizeof(*unmap_ctx) 862 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
835 if (!unmap_ctx) 863 if (!unmap_ctx)
836 return -ENOMEM; 864 return -ENOMEM;
837 865
838 userfaultfd_ctx_get(ctx); 866 userfaultfd_ctx_get(ctx);
839 down_write(&ctx->map_changing_lock); <<
840 atomic_inc(&ctx->mmap_changing); 867 atomic_inc(&ctx->mmap_changing);
841 up_write(&ctx->map_changing_lock); <<
842 unmap_ctx->ctx = ctx; 868 unmap_ctx->ctx = ctx;
843 unmap_ctx->start = start; 869 unmap_ctx->start = start;
844 unmap_ctx->end = end; 870 unmap_ctx->end = end;
845 list_add_tail(&unmap_ctx->list, unmaps 871 list_add_tail(&unmap_ctx->list, unmaps);
846 872
847 return 0; 873 return 0;
848 } 874 }
849 875
850 void userfaultfd_unmap_complete(struct mm_stru 876 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
851 { 877 {
852 struct userfaultfd_unmap_ctx *ctx, *n; 878 struct userfaultfd_unmap_ctx *ctx, *n;
853 struct userfaultfd_wait_queue ewq; 879 struct userfaultfd_wait_queue ewq;
854 880
855 list_for_each_entry_safe(ctx, n, uf, l 881 list_for_each_entry_safe(ctx, n, uf, list) {
856 msg_init(&ewq.msg); 882 msg_init(&ewq.msg);
857 883
858 ewq.msg.event = UFFD_EVENT_UNM 884 ewq.msg.event = UFFD_EVENT_UNMAP;
859 ewq.msg.arg.remove.start = ctx 885 ewq.msg.arg.remove.start = ctx->start;
860 ewq.msg.arg.remove.end = ctx-> 886 ewq.msg.arg.remove.end = ctx->end;
861 887
862 userfaultfd_event_wait_complet 888 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
863 889
864 list_del(&ctx->list); 890 list_del(&ctx->list);
865 kfree(ctx); 891 kfree(ctx);
866 } 892 }
867 } 893 }
868 894
869 static int userfaultfd_release(struct inode *i 895 static int userfaultfd_release(struct inode *inode, struct file *file)
870 { 896 {
871 struct userfaultfd_ctx *ctx = file->pr 897 struct userfaultfd_ctx *ctx = file->private_data;
872 struct mm_struct *mm = ctx->mm; 898 struct mm_struct *mm = ctx->mm;
873 struct vm_area_struct *vma, *prev; 899 struct vm_area_struct *vma, *prev;
874 /* len == 0 means wake all */ 900 /* len == 0 means wake all */
875 struct userfaultfd_wake_range range = 901 struct userfaultfd_wake_range range = { .len = 0, };
876 unsigned long new_flags; 902 unsigned long new_flags;
877 VMA_ITERATOR(vmi, mm, 0); 903 VMA_ITERATOR(vmi, mm, 0);
878 904
879 WRITE_ONCE(ctx->released, true); 905 WRITE_ONCE(ctx->released, true);
880 906
881 if (!mmget_not_zero(mm)) 907 if (!mmget_not_zero(mm))
882 goto wakeup; 908 goto wakeup;
883 909
884 /* 910 /*
885 * Flush page faults out of all CPUs. 911 * Flush page faults out of all CPUs. NOTE: all page faults
886 * must be retried without returning V 912 * must be retried without returning VM_FAULT_SIGBUS if
887 * userfaultfd_ctx_get() succeeds but 913 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
888 * changes while handle_userfault rele 914 * changes while handle_userfault released the mmap_lock. So
889 * it's critical that released is set 915 * it's critical that released is set to true (above), before
890 * taking the mmap_lock for writing. 916 * taking the mmap_lock for writing.
891 */ 917 */
892 mmap_write_lock(mm); 918 mmap_write_lock(mm);
893 prev = NULL; 919 prev = NULL;
894 for_each_vma(vmi, vma) { 920 for_each_vma(vmi, vma) {
895 cond_resched(); 921 cond_resched();
896 BUG_ON(!!vma->vm_userfaultfd_c 922 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
897 !!(vma->vm_flags & __VM 923 !!(vma->vm_flags & __VM_UFFD_FLAGS));
898 if (vma->vm_userfaultfd_ctx.ct 924 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
899 prev = vma; 925 prev = vma;
900 continue; 926 continue;
901 } 927 }
902 /* Reset ptes for the whole vm 928 /* Reset ptes for the whole vma range if wr-protected */
903 if (userfaultfd_wp(vma)) 929 if (userfaultfd_wp(vma))
904 uffd_wp_range(vma, vma 930 uffd_wp_range(vma, vma->vm_start,
905 vma->vm_ 931 vma->vm_end - vma->vm_start, false);
906 new_flags = vma->vm_flags & ~_ 932 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
907 vma = vma_modify_flags_uffd(&v 933 vma = vma_modify_flags_uffd(&vmi, prev, vma, vma->vm_start,
908 vm 934 vma->vm_end, new_flags,
909 NU 935 NULL_VM_UFFD_CTX);
910 936
911 vma_start_write(vma); 937 vma_start_write(vma);
912 userfaultfd_set_vm_flags(vma, 938 userfaultfd_set_vm_flags(vma, new_flags);
913 vma->vm_userfaultfd_ctx = NULL 939 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
914 940
915 prev = vma; 941 prev = vma;
916 } 942 }
917 mmap_write_unlock(mm); 943 mmap_write_unlock(mm);
918 mmput(mm); 944 mmput(mm);
919 wakeup: 945 wakeup:
920 /* 946 /*
921 * After no new page faults can wait o 947 * After no new page faults can wait on this fault_*wqh, flush
922 * the last page faults that may have 948 * the last page faults that may have been already waiting on
923 * the fault_*wqh. 949 * the fault_*wqh.
924 */ 950 */
925 spin_lock_irq(&ctx->fault_pending_wqh. 951 spin_lock_irq(&ctx->fault_pending_wqh.lock);
926 __wake_up_locked_key(&ctx->fault_pendi 952 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
927 __wake_up(&ctx->fault_wqh, TASK_NORMAL 953 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
928 spin_unlock_irq(&ctx->fault_pending_wq 954 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
929 955
930 /* Flush pending events that may still 956 /* Flush pending events that may still wait on event_wqh */
931 wake_up_all(&ctx->event_wqh); 957 wake_up_all(&ctx->event_wqh);
932 958
933 wake_up_poll(&ctx->fd_wqh, EPOLLHUP); 959 wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
934 userfaultfd_ctx_put(ctx); 960 userfaultfd_ctx_put(ctx);
935 return 0; 961 return 0;
936 } 962 }
937 963
938 /* fault_pending_wqh.lock must be hold by the 964 /* fault_pending_wqh.lock must be hold by the caller */
939 static inline struct userfaultfd_wait_queue *f 965 static inline struct userfaultfd_wait_queue *find_userfault_in(
940 wait_queue_head_t *wqh) 966 wait_queue_head_t *wqh)
941 { 967 {
942 wait_queue_entry_t *wq; 968 wait_queue_entry_t *wq;
943 struct userfaultfd_wait_queue *uwq; 969 struct userfaultfd_wait_queue *uwq;
944 970
945 lockdep_assert_held(&wqh->lock); 971 lockdep_assert_held(&wqh->lock);
946 972
947 uwq = NULL; 973 uwq = NULL;
948 if (!waitqueue_active(wqh)) 974 if (!waitqueue_active(wqh))
949 goto out; 975 goto out;
950 /* walk in reverse to provide FIFO beh 976 /* walk in reverse to provide FIFO behavior to read userfaults */
951 wq = list_last_entry(&wqh->head, typeo 977 wq = list_last_entry(&wqh->head, typeof(*wq), entry);
952 uwq = container_of(wq, struct userfaul 978 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
953 out: 979 out:
954 return uwq; 980 return uwq;
955 } 981 }
956 982
957 static inline struct userfaultfd_wait_queue *f 983 static inline struct userfaultfd_wait_queue *find_userfault(
958 struct userfaultfd_ctx *ctx) 984 struct userfaultfd_ctx *ctx)
959 { 985 {
960 return find_userfault_in(&ctx->fault_p 986 return find_userfault_in(&ctx->fault_pending_wqh);
961 } 987 }
962 988
963 static inline struct userfaultfd_wait_queue *f 989 static inline struct userfaultfd_wait_queue *find_userfault_evt(
964 struct userfaultfd_ctx *ctx) 990 struct userfaultfd_ctx *ctx)
965 { 991 {
966 return find_userfault_in(&ctx->event_w 992 return find_userfault_in(&ctx->event_wqh);
967 } 993 }
968 994
969 static __poll_t userfaultfd_poll(struct file * 995 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
970 { 996 {
971 struct userfaultfd_ctx *ctx = file->pr 997 struct userfaultfd_ctx *ctx = file->private_data;
972 __poll_t ret; 998 __poll_t ret;
973 999
974 poll_wait(file, &ctx->fd_wqh, wait); 1000 poll_wait(file, &ctx->fd_wqh, wait);
975 1001
976 if (!userfaultfd_is_initialized(ctx)) 1002 if (!userfaultfd_is_initialized(ctx))
977 return EPOLLERR; 1003 return EPOLLERR;
978 1004
979 /* 1005 /*
980 * poll() never guarantees that read w 1006 * poll() never guarantees that read won't block.
981 * userfaults can be waken before they 1007 * userfaults can be waken before they're read().
982 */ 1008 */
983 if (unlikely(!(file->f_flags & O_NONBL 1009 if (unlikely(!(file->f_flags & O_NONBLOCK)))
984 return EPOLLERR; 1010 return EPOLLERR;
985 /* 1011 /*
986 * lockless access to see if there are 1012 * lockless access to see if there are pending faults
987 * __pollwait last action is the add_w 1013 * __pollwait last action is the add_wait_queue but
988 * the spin_unlock would allow the wai 1014 * the spin_unlock would allow the waitqueue_active to
989 * pass above the actual list_add insi 1015 * pass above the actual list_add inside
990 * add_wait_queue critical section. So 1016 * add_wait_queue critical section. So use a full
991 * memory barrier to serialize the lis 1017 * memory barrier to serialize the list_add write of
992 * add_wait_queue() with the waitqueue 1018 * add_wait_queue() with the waitqueue_active read
993 * below. 1019 * below.
994 */ 1020 */
995 ret = 0; 1021 ret = 0;
996 smp_mb(); 1022 smp_mb();
997 if (waitqueue_active(&ctx->fault_pendi 1023 if (waitqueue_active(&ctx->fault_pending_wqh))
998 ret = EPOLLIN; 1024 ret = EPOLLIN;
999 else if (waitqueue_active(&ctx->event_ 1025 else if (waitqueue_active(&ctx->event_wqh))
1000 ret = EPOLLIN; 1026 ret = EPOLLIN;
1001 1027
1002 return ret; 1028 return ret;
1003 } 1029 }
1004 1030
1005 static const struct file_operations userfault 1031 static const struct file_operations userfaultfd_fops;
1006 1032
1007 static int resolve_userfault_fork(struct user 1033 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
1008 struct inod 1034 struct inode *inode,
1009 struct uffd 1035 struct uffd_msg *msg)
1010 { 1036 {
1011 int fd; 1037 int fd;
1012 1038
1013 fd = anon_inode_create_getfd("[userfa 1039 fd = anon_inode_create_getfd("[userfaultfd]", &userfaultfd_fops, new,
1014 O_RDONLY | (new->flag 1040 O_RDONLY | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
1015 if (fd < 0) 1041 if (fd < 0)
1016 return fd; 1042 return fd;
1017 1043
1018 msg->arg.reserved.reserved1 = 0; 1044 msg->arg.reserved.reserved1 = 0;
1019 msg->arg.fork.ufd = fd; 1045 msg->arg.fork.ufd = fd;
1020 return 0; 1046 return 0;
1021 } 1047 }
1022 1048
1023 static ssize_t userfaultfd_ctx_read(struct us 1049 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1024 struct uf 1050 struct uffd_msg *msg, struct inode *inode)
1025 { 1051 {
1026 ssize_t ret; 1052 ssize_t ret;
1027 DECLARE_WAITQUEUE(wait, current); 1053 DECLARE_WAITQUEUE(wait, current);
1028 struct userfaultfd_wait_queue *uwq; 1054 struct userfaultfd_wait_queue *uwq;
1029 /* 1055 /*
1030 * Handling fork event requires sleep 1056 * Handling fork event requires sleeping operations, so
1031 * we drop the event_wqh lock, then d 1057 * we drop the event_wqh lock, then do these ops, then
1032 * lock it back and wake up the waite 1058 * lock it back and wake up the waiter. While the lock is
1033 * dropped the ewq may go away so we 1059 * dropped the ewq may go away so we keep track of it
1034 * carefully. 1060 * carefully.
1035 */ 1061 */
1036 LIST_HEAD(fork_event); 1062 LIST_HEAD(fork_event);
1037 struct userfaultfd_ctx *fork_nctx = N 1063 struct userfaultfd_ctx *fork_nctx = NULL;
1038 1064
1039 /* always take the fd_wqh lock before 1065 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1040 spin_lock_irq(&ctx->fd_wqh.lock); 1066 spin_lock_irq(&ctx->fd_wqh.lock);
1041 __add_wait_queue(&ctx->fd_wqh, &wait) 1067 __add_wait_queue(&ctx->fd_wqh, &wait);
1042 for (;;) { 1068 for (;;) {
1043 set_current_state(TASK_INTERR 1069 set_current_state(TASK_INTERRUPTIBLE);
1044 spin_lock(&ctx->fault_pending 1070 spin_lock(&ctx->fault_pending_wqh.lock);
1045 uwq = find_userfault(ctx); 1071 uwq = find_userfault(ctx);
1046 if (uwq) { 1072 if (uwq) {
1047 /* 1073 /*
1048 * Use a seqcount to 1074 * Use a seqcount to repeat the lockless check
1049 * in wake_userfault( 1075 * in wake_userfault() to avoid missing
1050 * wakeups because du 1076 * wakeups because during the refile both
1051 * waitqueue could be 1077 * waitqueue could become empty if this is the
1052 * only userfault. 1078 * only userfault.
1053 */ 1079 */
1054 write_seqcount_begin( 1080 write_seqcount_begin(&ctx->refile_seq);
1055 1081
1056 /* 1082 /*
1057 * The fault_pending_ 1083 * The fault_pending_wqh.lock prevents the uwq
1058 * to disappear from 1084 * to disappear from under us.
1059 * 1085 *
1060 * Refile this userfa 1086 * Refile this userfault from
1061 * fault_pending_wqh 1087 * fault_pending_wqh to fault_wqh, it's not
1062 * pending anymore af 1088 * pending anymore after we read it.
1063 * 1089 *
1064 * Use list_del() by 1090 * Use list_del() by hand (as
1065 * userfaultfd_wake_f 1091 * userfaultfd_wake_function also uses
1066 * list_del_init() by 1092 * list_del_init() by hand) to be sure nobody
1067 * changes __remove_w 1093 * changes __remove_wait_queue() to use
1068 * list_del_init() in 1094 * list_del_init() in turn breaking the
1069 * !list_empty_carefu 1095 * !list_empty_careful() check in
1070 * handle_userfault() 1096 * handle_userfault(). The uwq->wq.head list
1071 * must never be empt 1097 * must never be empty at any time during the
1072 * refile, or the wai 1098 * refile, or the waitqueue could disappear
1073 * from under us. The 1099 * from under us. The "wait_queue_head_t"
1074 * parameter of __rem 1100 * parameter of __remove_wait_queue() is unused
1075 * anyway. 1101 * anyway.
1076 */ 1102 */
1077 list_del(&uwq->wq.ent 1103 list_del(&uwq->wq.entry);
1078 add_wait_queue(&ctx-> 1104 add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1079 1105
1080 write_seqcount_end(&c 1106 write_seqcount_end(&ctx->refile_seq);
1081 1107
1082 /* careful to always 1108 /* careful to always initialize msg if ret == 0 */
1083 *msg = uwq->msg; 1109 *msg = uwq->msg;
1084 spin_unlock(&ctx->fau 1110 spin_unlock(&ctx->fault_pending_wqh.lock);
1085 ret = 0; 1111 ret = 0;
1086 break; 1112 break;
1087 } 1113 }
1088 spin_unlock(&ctx->fault_pendi 1114 spin_unlock(&ctx->fault_pending_wqh.lock);
1089 1115
1090 spin_lock(&ctx->event_wqh.loc 1116 spin_lock(&ctx->event_wqh.lock);
1091 uwq = find_userfault_evt(ctx) 1117 uwq = find_userfault_evt(ctx);
1092 if (uwq) { 1118 if (uwq) {
1093 *msg = uwq->msg; 1119 *msg = uwq->msg;
1094 1120
1095 if (uwq->msg.event == 1121 if (uwq->msg.event == UFFD_EVENT_FORK) {
1096 fork_nctx = ( 1122 fork_nctx = (struct userfaultfd_ctx *)
1097 (unsi 1123 (unsigned long)
1098 uwq-> 1124 uwq->msg.arg.reserved.reserved1;
1099 list_move(&uw 1125 list_move(&uwq->wq.entry, &fork_event);
1100 /* 1126 /*
1101 * fork_nctx 1127 * fork_nctx can be freed as soon as
1102 * we drop th 1128 * we drop the lock, unless we take a
1103 * reference 1129 * reference on it.
1104 */ 1130 */
1105 userfaultfd_c 1131 userfaultfd_ctx_get(fork_nctx);
1106 spin_unlock(& 1132 spin_unlock(&ctx->event_wqh.lock);
1107 ret = 0; 1133 ret = 0;
1108 break; 1134 break;
1109 } 1135 }
1110 1136
1111 userfaultfd_event_com 1137 userfaultfd_event_complete(ctx, uwq);
1112 spin_unlock(&ctx->eve 1138 spin_unlock(&ctx->event_wqh.lock);
1113 ret = 0; 1139 ret = 0;
1114 break; 1140 break;
1115 } 1141 }
1116 spin_unlock(&ctx->event_wqh.l 1142 spin_unlock(&ctx->event_wqh.lock);
1117 1143
1118 if (signal_pending(current)) 1144 if (signal_pending(current)) {
1119 ret = -ERESTARTSYS; 1145 ret = -ERESTARTSYS;
1120 break; 1146 break;
1121 } 1147 }
1122 if (no_wait) { 1148 if (no_wait) {
1123 ret = -EAGAIN; 1149 ret = -EAGAIN;
1124 break; 1150 break;
1125 } 1151 }
1126 spin_unlock_irq(&ctx->fd_wqh. 1152 spin_unlock_irq(&ctx->fd_wqh.lock);
1127 schedule(); 1153 schedule();
1128 spin_lock_irq(&ctx->fd_wqh.lo 1154 spin_lock_irq(&ctx->fd_wqh.lock);
1129 } 1155 }
1130 __remove_wait_queue(&ctx->fd_wqh, &wa 1156 __remove_wait_queue(&ctx->fd_wqh, &wait);
1131 __set_current_state(TASK_RUNNING); 1157 __set_current_state(TASK_RUNNING);
1132 spin_unlock_irq(&ctx->fd_wqh.lock); 1158 spin_unlock_irq(&ctx->fd_wqh.lock);
1133 1159
1134 if (!ret && msg->event == UFFD_EVENT_ 1160 if (!ret && msg->event == UFFD_EVENT_FORK) {
1135 ret = resolve_userfault_fork( 1161 ret = resolve_userfault_fork(fork_nctx, inode, msg);
1136 spin_lock_irq(&ctx->event_wqh 1162 spin_lock_irq(&ctx->event_wqh.lock);
1137 if (!list_empty(&fork_event)) 1163 if (!list_empty(&fork_event)) {
1138 /* 1164 /*
1139 * The fork thread di 1165 * The fork thread didn't abort, so we can
1140 * drop the temporary 1166 * drop the temporary refcount.
1141 */ 1167 */
1142 userfaultfd_ctx_put(f 1168 userfaultfd_ctx_put(fork_nctx);
1143 1169
1144 uwq = list_first_entr 1170 uwq = list_first_entry(&fork_event,
1145 1171 typeof(*uwq),
1146 1172 wq.entry);
1147 /* 1173 /*
1148 * If fork_event list 1174 * If fork_event list wasn't empty and in turn
1149 * the event wasn't a 1175 * the event wasn't already released by fork
1150 * (the event is allo 1176 * (the event is allocated on fork kernel
1151 * stack), put the ev 1177 * stack), put the event back to its place in
1152 * the event_wq. fork 1178 * the event_wq. fork_event head will be freed
1153 * as soon as we retu 1179 * as soon as we return so the event cannot
1154 * stay queued there 1180 * stay queued there no matter the current
1155 * "ret" value. 1181 * "ret" value.
1156 */ 1182 */
1157 list_del(&uwq->wq.ent 1183 list_del(&uwq->wq.entry);
1158 __add_wait_queue(&ctx 1184 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1159 1185
1160 /* 1186 /*
1161 * Leave the event in 1187 * Leave the event in the waitqueue and report
1162 * error to userland 1188 * error to userland if we failed to resolve
1163 * the userfault fork 1189 * the userfault fork.
1164 */ 1190 */
1165 if (likely(!ret)) 1191 if (likely(!ret))
1166 userfaultfd_e 1192 userfaultfd_event_complete(ctx, uwq);
1167 } else { 1193 } else {
1168 /* 1194 /*
1169 * Here the fork thre 1195 * Here the fork thread aborted and the
1170 * refcount from the 1196 * refcount from the fork thread on fork_nctx
1171 * has already been r 1197 * has already been released. We still hold
1172 * the reference we t 1198 * the reference we took before releasing the
1173 * lock above. If res 1199 * lock above. If resolve_userfault_fork
1174 * failed we've to dr 1200 * failed we've to drop it because the
1175 * fork_nctx has to b 1201 * fork_nctx has to be freed in such case. If
1176 * it succeeded we'll 1202 * it succeeded we'll hold it because the new
1177 * uffd references it 1203 * uffd references it.
1178 */ 1204 */
1179 if (ret) 1205 if (ret)
1180 userfaultfd_c 1206 userfaultfd_ctx_put(fork_nctx);
1181 } 1207 }
1182 spin_unlock_irq(&ctx->event_w 1208 spin_unlock_irq(&ctx->event_wqh.lock);
1183 } 1209 }
1184 1210
1185 return ret; 1211 return ret;
1186 } 1212 }
1187 1213
1188 static ssize_t userfaultfd_read_iter(struct k !! 1214 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
>> 1215 size_t count, loff_t *ppos)
1189 { 1216 {
1190 struct file *file = iocb->ki_filp; <<
1191 struct userfaultfd_ctx *ctx = file->p 1217 struct userfaultfd_ctx *ctx = file->private_data;
1192 ssize_t _ret, ret = 0; 1218 ssize_t _ret, ret = 0;
1193 struct uffd_msg msg; 1219 struct uffd_msg msg;
>> 1220 int no_wait = file->f_flags & O_NONBLOCK;
1194 struct inode *inode = file_inode(file 1221 struct inode *inode = file_inode(file);
1195 bool no_wait; <<
1196 1222
1197 if (!userfaultfd_is_initialized(ctx)) 1223 if (!userfaultfd_is_initialized(ctx))
1198 return -EINVAL; 1224 return -EINVAL;
1199 1225
1200 no_wait = file->f_flags & O_NONBLOCK <<
1201 for (;;) { 1226 for (;;) {
1202 if (iov_iter_count(to) < size !! 1227 if (count < sizeof(msg))
1203 return ret ? ret : -E 1228 return ret ? ret : -EINVAL;
1204 _ret = userfaultfd_ctx_read(c 1229 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1205 if (_ret < 0) 1230 if (_ret < 0)
1206 return ret ? ret : _r 1231 return ret ? ret : _ret;
1207 _ret = !copy_to_iter_full(&ms !! 1232 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1208 if (_ret) <<
1209 return ret ? ret : -E 1233 return ret ? ret : -EFAULT;
1210 ret += sizeof(msg); 1234 ret += sizeof(msg);
>> 1235 buf += sizeof(msg);
>> 1236 count -= sizeof(msg);
1211 /* 1237 /*
1212 * Allow to read more than on 1238 * Allow to read more than one fault at time but only
1213 * block if waiting for the v 1239 * block if waiting for the very first one.
1214 */ 1240 */
1215 no_wait = true; !! 1241 no_wait = O_NONBLOCK;
1216 } 1242 }
1217 } 1243 }
1218 1244
1219 static void __wake_userfault(struct userfault 1245 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1220 struct userfault 1246 struct userfaultfd_wake_range *range)
1221 { 1247 {
1222 spin_lock_irq(&ctx->fault_pending_wqh 1248 spin_lock_irq(&ctx->fault_pending_wqh.lock);
1223 /* wake all in the range and autoremo 1249 /* wake all in the range and autoremove */
1224 if (waitqueue_active(&ctx->fault_pend 1250 if (waitqueue_active(&ctx->fault_pending_wqh))
1225 __wake_up_locked_key(&ctx->fa 1251 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1226 range); 1252 range);
1227 if (waitqueue_active(&ctx->fault_wqh) 1253 if (waitqueue_active(&ctx->fault_wqh))
1228 __wake_up(&ctx->fault_wqh, TA 1254 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1229 spin_unlock_irq(&ctx->fault_pending_w 1255 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1230 } 1256 }
1231 1257
1232 static __always_inline void wake_userfault(st 1258 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1233 st 1259 struct userfaultfd_wake_range *range)
1234 { 1260 {
1235 unsigned seq; 1261 unsigned seq;
1236 bool need_wakeup; 1262 bool need_wakeup;
1237 1263
1238 /* 1264 /*
1239 * To be sure waitqueue_active() is n 1265 * To be sure waitqueue_active() is not reordered by the CPU
1240 * before the pagetable update, use a 1266 * before the pagetable update, use an explicit SMP memory
1241 * barrier here. PT lock release or m 1267 * barrier here. PT lock release or mmap_read_unlock(mm) still
1242 * have release semantics that can al 1268 * have release semantics that can allow the
1243 * waitqueue_active() to be reordered 1269 * waitqueue_active() to be reordered before the pte update.
1244 */ 1270 */
1245 smp_mb(); 1271 smp_mb();
1246 1272
1247 /* 1273 /*
1248 * Use waitqueue_active because it's 1274 * Use waitqueue_active because it's very frequent to
1249 * change the address space atomicall 1275 * change the address space atomically even if there are no
1250 * userfaults yet. So we take the spi 1276 * userfaults yet. So we take the spinlock only when we're
1251 * sure we've userfaults to wake. 1277 * sure we've userfaults to wake.
1252 */ 1278 */
1253 do { 1279 do {
1254 seq = read_seqcount_begin(&ct 1280 seq = read_seqcount_begin(&ctx->refile_seq);
1255 need_wakeup = waitqueue_activ 1281 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1256 waitqueue_active(&ctx 1282 waitqueue_active(&ctx->fault_wqh);
1257 cond_resched(); 1283 cond_resched();
1258 } while (read_seqcount_retry(&ctx->re 1284 } while (read_seqcount_retry(&ctx->refile_seq, seq));
1259 if (need_wakeup) 1285 if (need_wakeup)
1260 __wake_userfault(ctx, range); 1286 __wake_userfault(ctx, range);
1261 } 1287 }
1262 1288
1263 static __always_inline int validate_unaligned 1289 static __always_inline int validate_unaligned_range(
1264 struct mm_struct *mm, __u64 start, __ 1290 struct mm_struct *mm, __u64 start, __u64 len)
1265 { 1291 {
1266 __u64 task_size = mm->task_size; 1292 __u64 task_size = mm->task_size;
1267 1293
1268 if (len & ~PAGE_MASK) 1294 if (len & ~PAGE_MASK)
1269 return -EINVAL; 1295 return -EINVAL;
1270 if (!len) 1296 if (!len)
1271 return -EINVAL; 1297 return -EINVAL;
1272 if (start < mmap_min_addr) 1298 if (start < mmap_min_addr)
1273 return -EINVAL; 1299 return -EINVAL;
1274 if (start >= task_size) 1300 if (start >= task_size)
1275 return -EINVAL; 1301 return -EINVAL;
1276 if (len > task_size - start) 1302 if (len > task_size - start)
1277 return -EINVAL; 1303 return -EINVAL;
1278 if (start + len <= start) 1304 if (start + len <= start)
1279 return -EINVAL; 1305 return -EINVAL;
1280 return 0; 1306 return 0;
1281 } 1307 }
1282 1308
1283 static __always_inline int validate_range(str 1309 static __always_inline int validate_range(struct mm_struct *mm,
1284 __u 1310 __u64 start, __u64 len)
1285 { 1311 {
1286 if (start & ~PAGE_MASK) 1312 if (start & ~PAGE_MASK)
1287 return -EINVAL; 1313 return -EINVAL;
1288 1314
1289 return validate_unaligned_range(mm, s 1315 return validate_unaligned_range(mm, start, len);
1290 } 1316 }
1291 1317
1292 static int userfaultfd_register(struct userfa 1318 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1293 unsigned long 1319 unsigned long arg)
1294 { 1320 {
1295 struct mm_struct *mm = ctx->mm; 1321 struct mm_struct *mm = ctx->mm;
1296 struct vm_area_struct *vma, *prev, *c 1322 struct vm_area_struct *vma, *prev, *cur;
1297 int ret; 1323 int ret;
1298 struct uffdio_register uffdio_registe 1324 struct uffdio_register uffdio_register;
1299 struct uffdio_register __user *user_u 1325 struct uffdio_register __user *user_uffdio_register;
1300 unsigned long vm_flags, new_flags; 1326 unsigned long vm_flags, new_flags;
1301 bool found; 1327 bool found;
1302 bool basic_ioctls; 1328 bool basic_ioctls;
1303 unsigned long start, end, vma_end; 1329 unsigned long start, end, vma_end;
1304 struct vma_iterator vmi; 1330 struct vma_iterator vmi;
1305 bool wp_async = userfaultfd_wp_async_ 1331 bool wp_async = userfaultfd_wp_async_ctx(ctx);
1306 1332
1307 user_uffdio_register = (struct uffdio 1333 user_uffdio_register = (struct uffdio_register __user *) arg;
1308 1334
1309 ret = -EFAULT; 1335 ret = -EFAULT;
1310 if (copy_from_user(&uffdio_register, 1336 if (copy_from_user(&uffdio_register, user_uffdio_register,
1311 sizeof(uffdio_regi 1337 sizeof(uffdio_register)-sizeof(__u64)))
1312 goto out; 1338 goto out;
1313 1339
1314 ret = -EINVAL; 1340 ret = -EINVAL;
1315 if (!uffdio_register.mode) 1341 if (!uffdio_register.mode)
1316 goto out; 1342 goto out;
1317 if (uffdio_register.mode & ~UFFD_API_ 1343 if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1318 goto out; 1344 goto out;
1319 vm_flags = 0; 1345 vm_flags = 0;
1320 if (uffdio_register.mode & UFFDIO_REG 1346 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1321 vm_flags |= VM_UFFD_MISSING; 1347 vm_flags |= VM_UFFD_MISSING;
1322 if (uffdio_register.mode & UFFDIO_REG 1348 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1323 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP 1349 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1324 goto out; 1350 goto out;
1325 #endif 1351 #endif
1326 vm_flags |= VM_UFFD_WP; 1352 vm_flags |= VM_UFFD_WP;
1327 } 1353 }
1328 if (uffdio_register.mode & UFFDIO_REG 1354 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1329 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR 1355 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1330 goto out; 1356 goto out;
1331 #endif 1357 #endif
1332 vm_flags |= VM_UFFD_MINOR; 1358 vm_flags |= VM_UFFD_MINOR;
1333 } 1359 }
1334 1360
1335 ret = validate_range(mm, uffdio_regis 1361 ret = validate_range(mm, uffdio_register.range.start,
1336 uffdio_register. 1362 uffdio_register.range.len);
1337 if (ret) 1363 if (ret)
1338 goto out; 1364 goto out;
1339 1365
1340 start = uffdio_register.range.start; 1366 start = uffdio_register.range.start;
1341 end = start + uffdio_register.range.l 1367 end = start + uffdio_register.range.len;
1342 1368
1343 ret = -ENOMEM; 1369 ret = -ENOMEM;
1344 if (!mmget_not_zero(mm)) 1370 if (!mmget_not_zero(mm))
1345 goto out; 1371 goto out;
1346 1372
1347 ret = -EINVAL; 1373 ret = -EINVAL;
1348 mmap_write_lock(mm); 1374 mmap_write_lock(mm);
1349 vma_iter_init(&vmi, mm, start); 1375 vma_iter_init(&vmi, mm, start);
1350 vma = vma_find(&vmi, end); 1376 vma = vma_find(&vmi, end);
1351 if (!vma) 1377 if (!vma)
1352 goto out_unlock; 1378 goto out_unlock;
1353 1379
1354 /* 1380 /*
1355 * If the first vma contains huge pag 1381 * If the first vma contains huge pages, make sure start address
1356 * is aligned to huge page size. 1382 * is aligned to huge page size.
1357 */ 1383 */
1358 if (is_vm_hugetlb_page(vma)) { 1384 if (is_vm_hugetlb_page(vma)) {
1359 unsigned long vma_hpagesize = 1385 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1360 1386
1361 if (start & (vma_hpagesize - 1387 if (start & (vma_hpagesize - 1))
1362 goto out_unlock; 1388 goto out_unlock;
1363 } 1389 }
1364 1390
1365 /* 1391 /*
1366 * Search for not compatible vmas. 1392 * Search for not compatible vmas.
1367 */ 1393 */
1368 found = false; 1394 found = false;
1369 basic_ioctls = false; 1395 basic_ioctls = false;
1370 cur = vma; 1396 cur = vma;
1371 do { 1397 do {
1372 cond_resched(); 1398 cond_resched();
1373 1399
1374 BUG_ON(!!cur->vm_userfaultfd_ 1400 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1375 !!(cur->vm_flags & __V 1401 !!(cur->vm_flags & __VM_UFFD_FLAGS));
1376 1402
1377 /* check not compatible vmas 1403 /* check not compatible vmas */
1378 ret = -EINVAL; 1404 ret = -EINVAL;
1379 if (!vma_can_userfault(cur, v 1405 if (!vma_can_userfault(cur, vm_flags, wp_async))
1380 goto out_unlock; 1406 goto out_unlock;
1381 1407
1382 /* 1408 /*
1383 * UFFDIO_COPY will fill file 1409 * UFFDIO_COPY will fill file holes even without
1384 * PROT_WRITE. This check enf 1410 * PROT_WRITE. This check enforces that if this is a
1385 * MAP_SHARED, the process ha 1411 * MAP_SHARED, the process has write permission to the backing
1386 * file. If VM_MAYWRITE is se 1412 * file. If VM_MAYWRITE is set it also enforces that on a
1387 * MAP_SHARED vma: there is n 1413 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1388 * F_WRITE_SEAL can be taken 1414 * F_WRITE_SEAL can be taken until the vma is destroyed.
1389 */ 1415 */
1390 ret = -EPERM; 1416 ret = -EPERM;
1391 if (unlikely(!(cur->vm_flags 1417 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1392 goto out_unlock; 1418 goto out_unlock;
1393 1419
1394 /* 1420 /*
1395 * If this vma contains endin 1421 * If this vma contains ending address, and huge pages
1396 * check alignment. 1422 * check alignment.
1397 */ 1423 */
1398 if (is_vm_hugetlb_page(cur) & 1424 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1399 end > cur->vm_start) { 1425 end > cur->vm_start) {
1400 unsigned long vma_hpa 1426 unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1401 1427
1402 ret = -EINVAL; 1428 ret = -EINVAL;
1403 1429
1404 if (end & (vma_hpages 1430 if (end & (vma_hpagesize - 1))
1405 goto out_unlo 1431 goto out_unlock;
1406 } 1432 }
1407 if ((vm_flags & VM_UFFD_WP) & 1433 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1408 goto out_unlock; 1434 goto out_unlock;
1409 1435
1410 /* 1436 /*
1411 * Check that this vma isn't 1437 * Check that this vma isn't already owned by a
1412 * different userfaultfd. We 1438 * different userfaultfd. We can't allow more than one
1413 * userfaultfd to own a singl 1439 * userfaultfd to own a single vma simultaneously or we
1414 * wouldn't know which one to 1440 * wouldn't know which one to deliver the userfaults to.
1415 */ 1441 */
1416 ret = -EBUSY; 1442 ret = -EBUSY;
1417 if (cur->vm_userfaultfd_ctx.c 1443 if (cur->vm_userfaultfd_ctx.ctx &&
1418 cur->vm_userfaultfd_ctx.c 1444 cur->vm_userfaultfd_ctx.ctx != ctx)
1419 goto out_unlock; 1445 goto out_unlock;
1420 1446
1421 /* 1447 /*
1422 * Note vmas containing huge 1448 * Note vmas containing huge pages
1423 */ 1449 */
1424 if (is_vm_hugetlb_page(cur)) 1450 if (is_vm_hugetlb_page(cur))
1425 basic_ioctls = true; 1451 basic_ioctls = true;
1426 1452
1427 found = true; 1453 found = true;
1428 } for_each_vma_range(vmi, cur, end); 1454 } for_each_vma_range(vmi, cur, end);
1429 BUG_ON(!found); 1455 BUG_ON(!found);
1430 1456
1431 vma_iter_set(&vmi, start); 1457 vma_iter_set(&vmi, start);
1432 prev = vma_prev(&vmi); 1458 prev = vma_prev(&vmi);
1433 if (vma->vm_start < start) 1459 if (vma->vm_start < start)
1434 prev = vma; 1460 prev = vma;
1435 1461
1436 ret = 0; 1462 ret = 0;
1437 for_each_vma_range(vmi, vma, end) { 1463 for_each_vma_range(vmi, vma, end) {
1438 cond_resched(); 1464 cond_resched();
1439 1465
1440 BUG_ON(!vma_can_userfault(vma 1466 BUG_ON(!vma_can_userfault(vma, vm_flags, wp_async));
1441 BUG_ON(vma->vm_userfaultfd_ct 1467 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1442 vma->vm_userfaultfd_ct 1468 vma->vm_userfaultfd_ctx.ctx != ctx);
1443 WARN_ON(!(vma->vm_flags & VM_ 1469 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1444 1470
1445 /* 1471 /*
1446 * Nothing to do: this vma is 1472 * Nothing to do: this vma is already registered into this
1447 * userfaultfd and with the r 1473 * userfaultfd and with the right tracking mode too.
1448 */ 1474 */
1449 if (vma->vm_userfaultfd_ctx.c 1475 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1450 (vma->vm_flags & vm_flags 1476 (vma->vm_flags & vm_flags) == vm_flags)
1451 goto skip; 1477 goto skip;
1452 1478
1453 if (vma->vm_start > start) 1479 if (vma->vm_start > start)
1454 start = vma->vm_start 1480 start = vma->vm_start;
1455 vma_end = min(end, vma->vm_en 1481 vma_end = min(end, vma->vm_end);
1456 1482
1457 new_flags = (vma->vm_flags & 1483 new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
1458 vma = vma_modify_flags_uffd(& 1484 vma = vma_modify_flags_uffd(&vmi, prev, vma, start, vma_end,
1459 n 1485 new_flags,
1460 ( 1486 (struct vm_userfaultfd_ctx){ctx});
1461 if (IS_ERR(vma)) { 1487 if (IS_ERR(vma)) {
1462 ret = PTR_ERR(vma); 1488 ret = PTR_ERR(vma);
1463 break; 1489 break;
1464 } 1490 }
1465 1491
1466 /* 1492 /*
1467 * In the vma_merge() success 1493 * In the vma_merge() successful mprotect-like case 8:
1468 * the next vma was merged in 1494 * the next vma was merged into the current one and
1469 * the current one has not be 1495 * the current one has not been updated yet.
1470 */ 1496 */
1471 vma_start_write(vma); 1497 vma_start_write(vma);
1472 userfaultfd_set_vm_flags(vma, 1498 userfaultfd_set_vm_flags(vma, new_flags);
1473 vma->vm_userfaultfd_ctx.ctx = 1499 vma->vm_userfaultfd_ctx.ctx = ctx;
1474 1500
1475 if (is_vm_hugetlb_page(vma) & 1501 if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1476 hugetlb_unshare_all_p 1502 hugetlb_unshare_all_pmds(vma);
1477 1503
1478 skip: 1504 skip:
1479 prev = vma; 1505 prev = vma;
1480 start = vma->vm_end; 1506 start = vma->vm_end;
1481 } 1507 }
1482 1508
1483 out_unlock: 1509 out_unlock:
1484 mmap_write_unlock(mm); 1510 mmap_write_unlock(mm);
1485 mmput(mm); 1511 mmput(mm);
1486 if (!ret) { 1512 if (!ret) {
1487 __u64 ioctls_out; 1513 __u64 ioctls_out;
1488 1514
1489 ioctls_out = basic_ioctls ? U 1515 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1490 UFFD_API_RANGE_IOCTLS; 1516 UFFD_API_RANGE_IOCTLS;
1491 1517
1492 /* 1518 /*
1493 * Declare the WP ioctl only 1519 * Declare the WP ioctl only if the WP mode is
1494 * specified and all checks p 1520 * specified and all checks passed with the range
1495 */ 1521 */
1496 if (!(uffdio_register.mode & 1522 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1497 ioctls_out &= ~((__u6 1523 ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1498 1524
1499 /* CONTINUE ioctl is only sup 1525 /* CONTINUE ioctl is only supported for MINOR ranges. */
1500 if (!(uffdio_register.mode & 1526 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1501 ioctls_out &= ~((__u6 1527 ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1502 1528
1503 /* 1529 /*
1504 * Now that we scanned all vm 1530 * Now that we scanned all vmas we can already tell
1505 * userland which ioctls meth 1531 * userland which ioctls methods are guaranteed to
1506 * succeed on this range. 1532 * succeed on this range.
1507 */ 1533 */
1508 if (put_user(ioctls_out, &use 1534 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1509 ret = -EFAULT; 1535 ret = -EFAULT;
1510 } 1536 }
1511 out: 1537 out:
1512 return ret; 1538 return ret;
1513 } 1539 }
1514 1540
1515 static int userfaultfd_unregister(struct user 1541 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1516 unsigned lo 1542 unsigned long arg)
1517 { 1543 {
1518 struct mm_struct *mm = ctx->mm; 1544 struct mm_struct *mm = ctx->mm;
1519 struct vm_area_struct *vma, *prev, *c 1545 struct vm_area_struct *vma, *prev, *cur;
1520 int ret; 1546 int ret;
1521 struct uffdio_range uffdio_unregister 1547 struct uffdio_range uffdio_unregister;
1522 unsigned long new_flags; 1548 unsigned long new_flags;
1523 bool found; 1549 bool found;
1524 unsigned long start, end, vma_end; 1550 unsigned long start, end, vma_end;
1525 const void __user *buf = (void __user 1551 const void __user *buf = (void __user *)arg;
1526 struct vma_iterator vmi; 1552 struct vma_iterator vmi;
1527 bool wp_async = userfaultfd_wp_async_ 1553 bool wp_async = userfaultfd_wp_async_ctx(ctx);
1528 1554
1529 ret = -EFAULT; 1555 ret = -EFAULT;
1530 if (copy_from_user(&uffdio_unregister 1556 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1531 goto out; 1557 goto out;
1532 1558
1533 ret = validate_range(mm, uffdio_unreg 1559 ret = validate_range(mm, uffdio_unregister.start,
1534 uffdio_unregiste 1560 uffdio_unregister.len);
1535 if (ret) 1561 if (ret)
1536 goto out; 1562 goto out;
1537 1563
1538 start = uffdio_unregister.start; 1564 start = uffdio_unregister.start;
1539 end = start + uffdio_unregister.len; 1565 end = start + uffdio_unregister.len;
1540 1566
1541 ret = -ENOMEM; 1567 ret = -ENOMEM;
1542 if (!mmget_not_zero(mm)) 1568 if (!mmget_not_zero(mm))
1543 goto out; 1569 goto out;
1544 1570
1545 mmap_write_lock(mm); 1571 mmap_write_lock(mm);
1546 ret = -EINVAL; 1572 ret = -EINVAL;
1547 vma_iter_init(&vmi, mm, start); 1573 vma_iter_init(&vmi, mm, start);
1548 vma = vma_find(&vmi, end); 1574 vma = vma_find(&vmi, end);
1549 if (!vma) 1575 if (!vma)
1550 goto out_unlock; 1576 goto out_unlock;
1551 1577
1552 /* 1578 /*
1553 * If the first vma contains huge pag 1579 * If the first vma contains huge pages, make sure start address
1554 * is aligned to huge page size. 1580 * is aligned to huge page size.
1555 */ 1581 */
1556 if (is_vm_hugetlb_page(vma)) { 1582 if (is_vm_hugetlb_page(vma)) {
1557 unsigned long vma_hpagesize = 1583 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1558 1584
1559 if (start & (vma_hpagesize - 1585 if (start & (vma_hpagesize - 1))
1560 goto out_unlock; 1586 goto out_unlock;
1561 } 1587 }
1562 1588
1563 /* 1589 /*
1564 * Search for not compatible vmas. 1590 * Search for not compatible vmas.
1565 */ 1591 */
1566 found = false; 1592 found = false;
1567 cur = vma; 1593 cur = vma;
1568 do { 1594 do {
1569 cond_resched(); 1595 cond_resched();
1570 1596
1571 BUG_ON(!!cur->vm_userfaultfd_ 1597 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1572 !!(cur->vm_flags & __V 1598 !!(cur->vm_flags & __VM_UFFD_FLAGS));
1573 1599
1574 /* 1600 /*
1575 * Check not compatible vmas, 1601 * Check not compatible vmas, not strictly required
1576 * here as not compatible vma 1602 * here as not compatible vmas cannot have an
1577 * userfaultfd_ctx registered 1603 * userfaultfd_ctx registered on them, but this
1578 * provides for more strict b 1604 * provides for more strict behavior to notice
1579 * unregistration errors. 1605 * unregistration errors.
1580 */ 1606 */
1581 if (!vma_can_userfault(cur, c 1607 if (!vma_can_userfault(cur, cur->vm_flags, wp_async))
1582 goto out_unlock; 1608 goto out_unlock;
1583 1609
1584 found = true; 1610 found = true;
1585 } for_each_vma_range(vmi, cur, end); 1611 } for_each_vma_range(vmi, cur, end);
1586 BUG_ON(!found); 1612 BUG_ON(!found);
1587 1613
1588 vma_iter_set(&vmi, start); 1614 vma_iter_set(&vmi, start);
1589 prev = vma_prev(&vmi); 1615 prev = vma_prev(&vmi);
1590 if (vma->vm_start < start) 1616 if (vma->vm_start < start)
1591 prev = vma; 1617 prev = vma;
1592 1618
1593 ret = 0; 1619 ret = 0;
1594 for_each_vma_range(vmi, vma, end) { 1620 for_each_vma_range(vmi, vma, end) {
1595 cond_resched(); 1621 cond_resched();
1596 1622
1597 BUG_ON(!vma_can_userfault(vma 1623 BUG_ON(!vma_can_userfault(vma, vma->vm_flags, wp_async));
1598 1624
1599 /* 1625 /*
1600 * Nothing to do: this vma is 1626 * Nothing to do: this vma is already registered into this
1601 * userfaultfd and with the r 1627 * userfaultfd and with the right tracking mode too.
1602 */ 1628 */
1603 if (!vma->vm_userfaultfd_ctx. 1629 if (!vma->vm_userfaultfd_ctx.ctx)
1604 goto skip; 1630 goto skip;
1605 1631
1606 WARN_ON(!(vma->vm_flags & VM_ 1632 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1607 1633
1608 if (vma->vm_start > start) 1634 if (vma->vm_start > start)
1609 start = vma->vm_start 1635 start = vma->vm_start;
1610 vma_end = min(end, vma->vm_en 1636 vma_end = min(end, vma->vm_end);
1611 1637
1612 if (userfaultfd_missing(vma)) 1638 if (userfaultfd_missing(vma)) {
1613 /* 1639 /*
1614 * Wake any concurren 1640 * Wake any concurrent pending userfault while
1615 * we unregister, so 1641 * we unregister, so they will not hang
1616 * permanently and it 1642 * permanently and it avoids userland to call
1617 * UFFDIO_WAKE explic 1643 * UFFDIO_WAKE explicitly.
1618 */ 1644 */
1619 struct userfaultfd_wa 1645 struct userfaultfd_wake_range range;
1620 range.start = start; 1646 range.start = start;
1621 range.len = vma_end - 1647 range.len = vma_end - start;
1622 wake_userfault(vma->v 1648 wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1623 } 1649 }
1624 1650
1625 /* Reset ptes for the whole v 1651 /* Reset ptes for the whole vma range if wr-protected */
1626 if (userfaultfd_wp(vma)) 1652 if (userfaultfd_wp(vma))
1627 uffd_wp_range(vma, st 1653 uffd_wp_range(vma, start, vma_end - start, false);
1628 1654
1629 new_flags = vma->vm_flags & ~ 1655 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
1630 vma = vma_modify_flags_uffd(& 1656 vma = vma_modify_flags_uffd(&vmi, prev, vma, start, vma_end,
1631 n 1657 new_flags, NULL_VM_UFFD_CTX);
1632 if (IS_ERR(vma)) { 1658 if (IS_ERR(vma)) {
1633 ret = PTR_ERR(vma); 1659 ret = PTR_ERR(vma);
1634 break; 1660 break;
1635 } 1661 }
1636 1662
1637 /* 1663 /*
1638 * In the vma_merge() success 1664 * In the vma_merge() successful mprotect-like case 8:
1639 * the next vma was merged in 1665 * the next vma was merged into the current one and
1640 * the current one has not be 1666 * the current one has not been updated yet.
1641 */ 1667 */
1642 vma_start_write(vma); 1668 vma_start_write(vma);
1643 userfaultfd_set_vm_flags(vma, 1669 userfaultfd_set_vm_flags(vma, new_flags);
1644 vma->vm_userfaultfd_ctx = NUL 1670 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1645 1671
1646 skip: 1672 skip:
1647 prev = vma; 1673 prev = vma;
1648 start = vma->vm_end; 1674 start = vma->vm_end;
1649 } 1675 }
1650 1676
1651 out_unlock: 1677 out_unlock:
1652 mmap_write_unlock(mm); 1678 mmap_write_unlock(mm);
1653 mmput(mm); 1679 mmput(mm);
1654 out: 1680 out:
1655 return ret; 1681 return ret;
1656 } 1682 }
1657 1683
1658 /* 1684 /*
1659 * userfaultfd_wake may be used in combinatio 1685 * userfaultfd_wake may be used in combination with the
1660 * UFFDIO_*_MODE_DONTWAKE to wakeup userfault 1686 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1661 */ 1687 */
1662 static int userfaultfd_wake(struct userfaultf 1688 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1663 unsigned long arg 1689 unsigned long arg)
1664 { 1690 {
1665 int ret; 1691 int ret;
1666 struct uffdio_range uffdio_wake; 1692 struct uffdio_range uffdio_wake;
1667 struct userfaultfd_wake_range range; 1693 struct userfaultfd_wake_range range;
1668 const void __user *buf = (void __user 1694 const void __user *buf = (void __user *)arg;
1669 1695
1670 ret = -EFAULT; 1696 ret = -EFAULT;
1671 if (copy_from_user(&uffdio_wake, buf, 1697 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1672 goto out; 1698 goto out;
1673 1699
1674 ret = validate_range(ctx->mm, uffdio_ 1700 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1675 if (ret) 1701 if (ret)
1676 goto out; 1702 goto out;
1677 1703
1678 range.start = uffdio_wake.start; 1704 range.start = uffdio_wake.start;
1679 range.len = uffdio_wake.len; 1705 range.len = uffdio_wake.len;
1680 1706
1681 /* 1707 /*
1682 * len == 0 means wake all and we don 1708 * len == 0 means wake all and we don't want to wake all here,
1683 * so check it again to be sure. 1709 * so check it again to be sure.
1684 */ 1710 */
1685 VM_BUG_ON(!range.len); 1711 VM_BUG_ON(!range.len);
1686 1712
1687 wake_userfault(ctx, &range); 1713 wake_userfault(ctx, &range);
1688 ret = 0; 1714 ret = 0;
1689 1715
1690 out: 1716 out:
1691 return ret; 1717 return ret;
1692 } 1718 }
1693 1719
1694 static int userfaultfd_copy(struct userfaultf 1720 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1695 unsigned long arg 1721 unsigned long arg)
1696 { 1722 {
1697 __s64 ret; 1723 __s64 ret;
1698 struct uffdio_copy uffdio_copy; 1724 struct uffdio_copy uffdio_copy;
1699 struct uffdio_copy __user *user_uffdi 1725 struct uffdio_copy __user *user_uffdio_copy;
1700 struct userfaultfd_wake_range range; 1726 struct userfaultfd_wake_range range;
1701 uffd_flags_t flags = 0; 1727 uffd_flags_t flags = 0;
1702 1728
1703 user_uffdio_copy = (struct uffdio_cop 1729 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1704 1730
1705 ret = -EAGAIN; 1731 ret = -EAGAIN;
1706 if (atomic_read(&ctx->mmap_changing)) 1732 if (atomic_read(&ctx->mmap_changing))
1707 goto out; 1733 goto out;
1708 1734
1709 ret = -EFAULT; 1735 ret = -EFAULT;
1710 if (copy_from_user(&uffdio_copy, user 1736 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1711 /* don't copy "cop 1737 /* don't copy "copy" last field */
1712 sizeof(uffdio_copy 1738 sizeof(uffdio_copy)-sizeof(__s64)))
1713 goto out; 1739 goto out;
1714 1740
1715 ret = validate_unaligned_range(ctx->m 1741 ret = validate_unaligned_range(ctx->mm, uffdio_copy.src,
1716 uffdio 1742 uffdio_copy.len);
1717 if (ret) 1743 if (ret)
1718 goto out; 1744 goto out;
1719 ret = validate_range(ctx->mm, uffdio_ 1745 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1720 if (ret) 1746 if (ret)
1721 goto out; 1747 goto out;
1722 1748
1723 ret = -EINVAL; 1749 ret = -EINVAL;
1724 if (uffdio_copy.mode & ~(UFFDIO_COPY_ 1750 if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1725 goto out; 1751 goto out;
1726 if (uffdio_copy.mode & UFFDIO_COPY_MO 1752 if (uffdio_copy.mode & UFFDIO_COPY_MODE_WP)
1727 flags |= MFILL_ATOMIC_WP; 1753 flags |= MFILL_ATOMIC_WP;
1728 if (mmget_not_zero(ctx->mm)) { 1754 if (mmget_not_zero(ctx->mm)) {
1729 ret = mfill_atomic_copy(ctx, !! 1755 ret = mfill_atomic_copy(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1730 uffdi !! 1756 uffdio_copy.len, &ctx->mmap_changing,
>> 1757 flags);
1731 mmput(ctx->mm); 1758 mmput(ctx->mm);
1732 } else { 1759 } else {
1733 return -ESRCH; 1760 return -ESRCH;
1734 } 1761 }
1735 if (unlikely(put_user(ret, &user_uffd 1762 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1736 return -EFAULT; 1763 return -EFAULT;
1737 if (ret < 0) 1764 if (ret < 0)
1738 goto out; 1765 goto out;
1739 BUG_ON(!ret); 1766 BUG_ON(!ret);
1740 /* len == 0 would wake all */ 1767 /* len == 0 would wake all */
1741 range.len = ret; 1768 range.len = ret;
1742 if (!(uffdio_copy.mode & UFFDIO_COPY_ 1769 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1743 range.start = uffdio_copy.dst 1770 range.start = uffdio_copy.dst;
1744 wake_userfault(ctx, &range); 1771 wake_userfault(ctx, &range);
1745 } 1772 }
1746 ret = range.len == uffdio_copy.len ? 1773 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1747 out: 1774 out:
1748 return ret; 1775 return ret;
1749 } 1776 }
1750 1777
1751 static int userfaultfd_zeropage(struct userfa 1778 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1752 unsigned long 1779 unsigned long arg)
1753 { 1780 {
1754 __s64 ret; 1781 __s64 ret;
1755 struct uffdio_zeropage uffdio_zeropag 1782 struct uffdio_zeropage uffdio_zeropage;
1756 struct uffdio_zeropage __user *user_u 1783 struct uffdio_zeropage __user *user_uffdio_zeropage;
1757 struct userfaultfd_wake_range range; 1784 struct userfaultfd_wake_range range;
1758 1785
1759 user_uffdio_zeropage = (struct uffdio 1786 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1760 1787
1761 ret = -EAGAIN; 1788 ret = -EAGAIN;
1762 if (atomic_read(&ctx->mmap_changing)) 1789 if (atomic_read(&ctx->mmap_changing))
1763 goto out; 1790 goto out;
1764 1791
1765 ret = -EFAULT; 1792 ret = -EFAULT;
1766 if (copy_from_user(&uffdio_zeropage, 1793 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1767 /* don't copy "zer 1794 /* don't copy "zeropage" last field */
1768 sizeof(uffdio_zero 1795 sizeof(uffdio_zeropage)-sizeof(__s64)))
1769 goto out; 1796 goto out;
1770 1797
1771 ret = validate_range(ctx->mm, uffdio_ 1798 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1772 uffdio_zeropage. 1799 uffdio_zeropage.range.len);
1773 if (ret) 1800 if (ret)
1774 goto out; 1801 goto out;
1775 ret = -EINVAL; 1802 ret = -EINVAL;
1776 if (uffdio_zeropage.mode & ~UFFDIO_ZE 1803 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1777 goto out; 1804 goto out;
1778 1805
1779 if (mmget_not_zero(ctx->mm)) { 1806 if (mmget_not_zero(ctx->mm)) {
1780 ret = mfill_atomic_zeropage(c !! 1807 ret = mfill_atomic_zeropage(ctx->mm, uffdio_zeropage.range.start,
1781 uf !! 1808 uffdio_zeropage.range.len,
>> 1809 &ctx->mmap_changing);
1782 mmput(ctx->mm); 1810 mmput(ctx->mm);
1783 } else { 1811 } else {
1784 return -ESRCH; 1812 return -ESRCH;
1785 } 1813 }
1786 if (unlikely(put_user(ret, &user_uffd 1814 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1787 return -EFAULT; 1815 return -EFAULT;
1788 if (ret < 0) 1816 if (ret < 0)
1789 goto out; 1817 goto out;
1790 /* len == 0 would wake all */ 1818 /* len == 0 would wake all */
1791 BUG_ON(!ret); 1819 BUG_ON(!ret);
1792 range.len = ret; 1820 range.len = ret;
1793 if (!(uffdio_zeropage.mode & UFFDIO_Z 1821 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1794 range.start = uffdio_zeropage 1822 range.start = uffdio_zeropage.range.start;
1795 wake_userfault(ctx, &range); 1823 wake_userfault(ctx, &range);
1796 } 1824 }
1797 ret = range.len == uffdio_zeropage.ra 1825 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1798 out: 1826 out:
1799 return ret; 1827 return ret;
1800 } 1828 }
1801 1829
1802 static int userfaultfd_writeprotect(struct us 1830 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1803 unsigned 1831 unsigned long arg)
1804 { 1832 {
1805 int ret; 1833 int ret;
1806 struct uffdio_writeprotect uffdio_wp; 1834 struct uffdio_writeprotect uffdio_wp;
1807 struct uffdio_writeprotect __user *us 1835 struct uffdio_writeprotect __user *user_uffdio_wp;
1808 struct userfaultfd_wake_range range; 1836 struct userfaultfd_wake_range range;
1809 bool mode_wp, mode_dontwake; 1837 bool mode_wp, mode_dontwake;
1810 1838
1811 if (atomic_read(&ctx->mmap_changing)) 1839 if (atomic_read(&ctx->mmap_changing))
1812 return -EAGAIN; 1840 return -EAGAIN;
1813 1841
1814 user_uffdio_wp = (struct uffdio_write 1842 user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1815 1843
1816 if (copy_from_user(&uffdio_wp, user_u 1844 if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1817 sizeof(struct uffd 1845 sizeof(struct uffdio_writeprotect)))
1818 return -EFAULT; 1846 return -EFAULT;
1819 1847
1820 ret = validate_range(ctx->mm, uffdio_ 1848 ret = validate_range(ctx->mm, uffdio_wp.range.start,
1821 uffdio_wp.range. 1849 uffdio_wp.range.len);
1822 if (ret) 1850 if (ret)
1823 return ret; 1851 return ret;
1824 1852
1825 if (uffdio_wp.mode & ~(UFFDIO_WRITEPR 1853 if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1826 UFFDIO_WRITEPR 1854 UFFDIO_WRITEPROTECT_MODE_WP))
1827 return -EINVAL; 1855 return -EINVAL;
1828 1856
1829 mode_wp = uffdio_wp.mode & UFFDIO_WRI 1857 mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1830 mode_dontwake = uffdio_wp.mode & UFFD 1858 mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1831 1859
1832 if (mode_wp && mode_dontwake) 1860 if (mode_wp && mode_dontwake)
1833 return -EINVAL; 1861 return -EINVAL;
1834 1862
1835 if (mmget_not_zero(ctx->mm)) { 1863 if (mmget_not_zero(ctx->mm)) {
1836 ret = mwriteprotect_range(ctx !! 1864 ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1837 uff !! 1865 uffdio_wp.range.len, mode_wp,
>> 1866 &ctx->mmap_changing);
1838 mmput(ctx->mm); 1867 mmput(ctx->mm);
1839 } else { 1868 } else {
1840 return -ESRCH; 1869 return -ESRCH;
1841 } 1870 }
1842 1871
1843 if (ret) 1872 if (ret)
1844 return ret; 1873 return ret;
1845 1874
1846 if (!mode_wp && !mode_dontwake) { 1875 if (!mode_wp && !mode_dontwake) {
1847 range.start = uffdio_wp.range 1876 range.start = uffdio_wp.range.start;
1848 range.len = uffdio_wp.range.l 1877 range.len = uffdio_wp.range.len;
1849 wake_userfault(ctx, &range); 1878 wake_userfault(ctx, &range);
1850 } 1879 }
1851 return ret; 1880 return ret;
1852 } 1881 }
1853 1882
1854 static int userfaultfd_continue(struct userfa 1883 static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1855 { 1884 {
1856 __s64 ret; 1885 __s64 ret;
1857 struct uffdio_continue uffdio_continu 1886 struct uffdio_continue uffdio_continue;
1858 struct uffdio_continue __user *user_u 1887 struct uffdio_continue __user *user_uffdio_continue;
1859 struct userfaultfd_wake_range range; 1888 struct userfaultfd_wake_range range;
1860 uffd_flags_t flags = 0; 1889 uffd_flags_t flags = 0;
1861 1890
1862 user_uffdio_continue = (struct uffdio 1891 user_uffdio_continue = (struct uffdio_continue __user *)arg;
1863 1892
1864 ret = -EAGAIN; 1893 ret = -EAGAIN;
1865 if (atomic_read(&ctx->mmap_changing)) 1894 if (atomic_read(&ctx->mmap_changing))
1866 goto out; 1895 goto out;
1867 1896
1868 ret = -EFAULT; 1897 ret = -EFAULT;
1869 if (copy_from_user(&uffdio_continue, 1898 if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1870 /* don't copy the 1899 /* don't copy the output fields */
1871 sizeof(uffdio_cont 1900 sizeof(uffdio_continue) - (sizeof(__s64))))
1872 goto out; 1901 goto out;
1873 1902
1874 ret = validate_range(ctx->mm, uffdio_ 1903 ret = validate_range(ctx->mm, uffdio_continue.range.start,
1875 uffdio_continue. 1904 uffdio_continue.range.len);
1876 if (ret) 1905 if (ret)
1877 goto out; 1906 goto out;
1878 1907
1879 ret = -EINVAL; 1908 ret = -EINVAL;
1880 if (uffdio_continue.mode & ~(UFFDIO_C 1909 if (uffdio_continue.mode & ~(UFFDIO_CONTINUE_MODE_DONTWAKE |
1881 UFFDIO_C 1910 UFFDIO_CONTINUE_MODE_WP))
1882 goto out; 1911 goto out;
1883 if (uffdio_continue.mode & UFFDIO_CON 1912 if (uffdio_continue.mode & UFFDIO_CONTINUE_MODE_WP)
1884 flags |= MFILL_ATOMIC_WP; 1913 flags |= MFILL_ATOMIC_WP;
1885 1914
1886 if (mmget_not_zero(ctx->mm)) { 1915 if (mmget_not_zero(ctx->mm)) {
1887 ret = mfill_atomic_continue(c !! 1916 ret = mfill_atomic_continue(ctx->mm, uffdio_continue.range.start,
1888 u !! 1917 uffdio_continue.range.len,
>> 1918 &ctx->mmap_changing, flags);
1889 mmput(ctx->mm); 1919 mmput(ctx->mm);
1890 } else { 1920 } else {
1891 return -ESRCH; 1921 return -ESRCH;
1892 } 1922 }
1893 1923
1894 if (unlikely(put_user(ret, &user_uffd 1924 if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1895 return -EFAULT; 1925 return -EFAULT;
1896 if (ret < 0) 1926 if (ret < 0)
1897 goto out; 1927 goto out;
1898 1928
1899 /* len == 0 would wake all */ 1929 /* len == 0 would wake all */
1900 BUG_ON(!ret); 1930 BUG_ON(!ret);
1901 range.len = ret; 1931 range.len = ret;
1902 if (!(uffdio_continue.mode & UFFDIO_C 1932 if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1903 range.start = uffdio_continue 1933 range.start = uffdio_continue.range.start;
1904 wake_userfault(ctx, &range); 1934 wake_userfault(ctx, &range);
1905 } 1935 }
1906 ret = range.len == uffdio_continue.ra 1936 ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1907 1937
1908 out: 1938 out:
1909 return ret; 1939 return ret;
1910 } 1940 }
1911 1941
1912 static inline int userfaultfd_poison(struct u 1942 static inline int userfaultfd_poison(struct userfaultfd_ctx *ctx, unsigned long arg)
1913 { 1943 {
1914 __s64 ret; 1944 __s64 ret;
1915 struct uffdio_poison uffdio_poison; 1945 struct uffdio_poison uffdio_poison;
1916 struct uffdio_poison __user *user_uff 1946 struct uffdio_poison __user *user_uffdio_poison;
1917 struct userfaultfd_wake_range range; 1947 struct userfaultfd_wake_range range;
1918 1948
1919 user_uffdio_poison = (struct uffdio_p 1949 user_uffdio_poison = (struct uffdio_poison __user *)arg;
1920 1950
1921 ret = -EAGAIN; 1951 ret = -EAGAIN;
1922 if (atomic_read(&ctx->mmap_changing)) 1952 if (atomic_read(&ctx->mmap_changing))
1923 goto out; 1953 goto out;
1924 1954
1925 ret = -EFAULT; 1955 ret = -EFAULT;
1926 if (copy_from_user(&uffdio_poison, us 1956 if (copy_from_user(&uffdio_poison, user_uffdio_poison,
1927 /* don't copy the 1957 /* don't copy the output fields */
1928 sizeof(uffdio_pois 1958 sizeof(uffdio_poison) - (sizeof(__s64))))
1929 goto out; 1959 goto out;
1930 1960
1931 ret = validate_range(ctx->mm, uffdio_ 1961 ret = validate_range(ctx->mm, uffdio_poison.range.start,
1932 uffdio_poison.ra 1962 uffdio_poison.range.len);
1933 if (ret) 1963 if (ret)
1934 goto out; 1964 goto out;
1935 1965
1936 ret = -EINVAL; 1966 ret = -EINVAL;
1937 if (uffdio_poison.mode & ~UFFDIO_POIS 1967 if (uffdio_poison.mode & ~UFFDIO_POISON_MODE_DONTWAKE)
1938 goto out; 1968 goto out;
1939 1969
1940 if (mmget_not_zero(ctx->mm)) { 1970 if (mmget_not_zero(ctx->mm)) {
1941 ret = mfill_atomic_poison(ctx !! 1971 ret = mfill_atomic_poison(ctx->mm, uffdio_poison.range.start,
1942 uff !! 1972 uffdio_poison.range.len,
>> 1973 &ctx->mmap_changing, 0);
1943 mmput(ctx->mm); 1974 mmput(ctx->mm);
1944 } else { 1975 } else {
1945 return -ESRCH; 1976 return -ESRCH;
1946 } 1977 }
1947 1978
1948 if (unlikely(put_user(ret, &user_uffd 1979 if (unlikely(put_user(ret, &user_uffdio_poison->updated)))
1949 return -EFAULT; 1980 return -EFAULT;
1950 if (ret < 0) 1981 if (ret < 0)
1951 goto out; 1982 goto out;
1952 1983
1953 /* len == 0 would wake all */ 1984 /* len == 0 would wake all */
1954 BUG_ON(!ret); 1985 BUG_ON(!ret);
1955 range.len = ret; 1986 range.len = ret;
1956 if (!(uffdio_poison.mode & UFFDIO_POI 1987 if (!(uffdio_poison.mode & UFFDIO_POISON_MODE_DONTWAKE)) {
1957 range.start = uffdio_poison.r 1988 range.start = uffdio_poison.range.start;
1958 wake_userfault(ctx, &range); 1989 wake_userfault(ctx, &range);
1959 } 1990 }
1960 ret = range.len == uffdio_poison.rang 1991 ret = range.len == uffdio_poison.range.len ? 0 : -EAGAIN;
1961 1992
1962 out: 1993 out:
1963 return ret; 1994 return ret;
1964 } 1995 }
1965 1996
1966 bool userfaultfd_wp_async(struct vm_area_stru 1997 bool userfaultfd_wp_async(struct vm_area_struct *vma)
1967 { 1998 {
1968 return userfaultfd_wp_async_ctx(vma-> 1999 return userfaultfd_wp_async_ctx(vma->vm_userfaultfd_ctx.ctx);
1969 } 2000 }
1970 2001
1971 static inline unsigned int uffd_ctx_features( 2002 static inline unsigned int uffd_ctx_features(__u64 user_features)
1972 { 2003 {
1973 /* 2004 /*
1974 * For the current set of features th 2005 * For the current set of features the bits just coincide. Set
1975 * UFFD_FEATURE_INITIALIZED to mark t 2006 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1976 */ 2007 */
1977 return (unsigned int)user_features | 2008 return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED;
1978 } 2009 }
1979 2010
1980 static int userfaultfd_move(struct userfaultf 2011 static int userfaultfd_move(struct userfaultfd_ctx *ctx,
1981 unsigned long arg 2012 unsigned long arg)
1982 { 2013 {
1983 __s64 ret; 2014 __s64 ret;
1984 struct uffdio_move uffdio_move; 2015 struct uffdio_move uffdio_move;
1985 struct uffdio_move __user *user_uffdi 2016 struct uffdio_move __user *user_uffdio_move;
1986 struct userfaultfd_wake_range range; 2017 struct userfaultfd_wake_range range;
1987 struct mm_struct *mm = ctx->mm; 2018 struct mm_struct *mm = ctx->mm;
1988 2019
1989 user_uffdio_move = (struct uffdio_mov 2020 user_uffdio_move = (struct uffdio_move __user *) arg;
1990 2021
1991 if (atomic_read(&ctx->mmap_changing)) 2022 if (atomic_read(&ctx->mmap_changing))
1992 return -EAGAIN; 2023 return -EAGAIN;
1993 2024
1994 if (copy_from_user(&uffdio_move, user 2025 if (copy_from_user(&uffdio_move, user_uffdio_move,
1995 /* don't copy "mov 2026 /* don't copy "move" last field */
1996 sizeof(uffdio_move 2027 sizeof(uffdio_move)-sizeof(__s64)))
1997 return -EFAULT; 2028 return -EFAULT;
1998 2029
1999 /* Do not allow cross-mm moves. */ 2030 /* Do not allow cross-mm moves. */
2000 if (mm != current->mm) 2031 if (mm != current->mm)
2001 return -EINVAL; 2032 return -EINVAL;
2002 2033
2003 ret = validate_range(mm, uffdio_move. 2034 ret = validate_range(mm, uffdio_move.dst, uffdio_move.len);
2004 if (ret) 2035 if (ret)
2005 return ret; 2036 return ret;
2006 2037
2007 ret = validate_range(mm, uffdio_move. 2038 ret = validate_range(mm, uffdio_move.src, uffdio_move.len);
2008 if (ret) 2039 if (ret)
2009 return ret; 2040 return ret;
2010 2041
2011 if (uffdio_move.mode & ~(UFFDIO_MOVE_ 2042 if (uffdio_move.mode & ~(UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES|
2012 UFFDIO_MOVE 2043 UFFDIO_MOVE_MODE_DONTWAKE))
2013 return -EINVAL; 2044 return -EINVAL;
2014 2045
2015 if (mmget_not_zero(mm)) { 2046 if (mmget_not_zero(mm)) {
2016 ret = move_pages(ctx, uffdio_ !! 2047 mmap_read_lock(mm);
2017 uffdio_move. !! 2048
>> 2049 /* Re-check after taking mmap_lock */
>> 2050 if (likely(!atomic_read(&ctx->mmap_changing)))
>> 2051 ret = move_pages(ctx, mm, uffdio_move.dst, uffdio_move.src,
>> 2052 uffdio_move.len, uffdio_move.mode);
>> 2053 else
>> 2054 ret = -EINVAL;
>> 2055
>> 2056 mmap_read_unlock(mm);
2018 mmput(mm); 2057 mmput(mm);
2019 } else { 2058 } else {
2020 return -ESRCH; 2059 return -ESRCH;
2021 } 2060 }
2022 2061
2023 if (unlikely(put_user(ret, &user_uffd 2062 if (unlikely(put_user(ret, &user_uffdio_move->move)))
2024 return -EFAULT; 2063 return -EFAULT;
2025 if (ret < 0) 2064 if (ret < 0)
2026 goto out; 2065 goto out;
2027 2066
2028 /* len == 0 would wake all */ 2067 /* len == 0 would wake all */
2029 VM_WARN_ON(!ret); 2068 VM_WARN_ON(!ret);
2030 range.len = ret; 2069 range.len = ret;
2031 if (!(uffdio_move.mode & UFFDIO_MOVE_ 2070 if (!(uffdio_move.mode & UFFDIO_MOVE_MODE_DONTWAKE)) {
2032 range.start = uffdio_move.dst 2071 range.start = uffdio_move.dst;
2033 wake_userfault(ctx, &range); 2072 wake_userfault(ctx, &range);
2034 } 2073 }
2035 ret = range.len == uffdio_move.len ? 2074 ret = range.len == uffdio_move.len ? 0 : -EAGAIN;
2036 2075
2037 out: 2076 out:
2038 return ret; 2077 return ret;
2039 } 2078 }
2040 2079
2041 /* 2080 /*
2042 * userland asks for a certain API version an 2081 * userland asks for a certain API version and we return which bits
2043 * and ioctl commands are implemented in this 2082 * and ioctl commands are implemented in this kernel for such API
2044 * version or -EINVAL if unknown. 2083 * version or -EINVAL if unknown.
2045 */ 2084 */
2046 static int userfaultfd_api(struct userfaultfd 2085 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
2047 unsigned long arg) 2086 unsigned long arg)
2048 { 2087 {
2049 struct uffdio_api uffdio_api; 2088 struct uffdio_api uffdio_api;
2050 void __user *buf = (void __user *)arg 2089 void __user *buf = (void __user *)arg;
2051 unsigned int ctx_features; 2090 unsigned int ctx_features;
2052 int ret; 2091 int ret;
2053 __u64 features; 2092 __u64 features;
2054 2093
2055 ret = -EFAULT; 2094 ret = -EFAULT;
2056 if (copy_from_user(&uffdio_api, buf, 2095 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
2057 goto out; 2096 goto out;
2058 features = uffdio_api.features; 2097 features = uffdio_api.features;
2059 ret = -EINVAL; 2098 ret = -EINVAL;
2060 if (uffdio_api.api != UFFD_API) !! 2099 if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
2061 goto err_out; 2100 goto err_out;
2062 ret = -EPERM; 2101 ret = -EPERM;
2063 if ((features & UFFD_FEATURE_EVENT_FO 2102 if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
2064 goto err_out; 2103 goto err_out;
2065 2104
2066 /* WP_ASYNC relies on WP_UNPOPULATED, 2105 /* WP_ASYNC relies on WP_UNPOPULATED, choose it unconditionally */
2067 if (features & UFFD_FEATURE_WP_ASYNC) 2106 if (features & UFFD_FEATURE_WP_ASYNC)
2068 features |= UFFD_FEATURE_WP_U 2107 features |= UFFD_FEATURE_WP_UNPOPULATED;
2069 2108
2070 /* report all available features and 2109 /* report all available features and ioctls to userland */
2071 uffdio_api.features = UFFD_API_FEATUR 2110 uffdio_api.features = UFFD_API_FEATURES;
2072 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR 2111 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
2073 uffdio_api.features &= 2112 uffdio_api.features &=
2074 ~(UFFD_FEATURE_MINOR_HUGETLBF 2113 ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
2075 #endif 2114 #endif
2076 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP 2115 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
2077 uffdio_api.features &= ~UFFD_FEATURE_ 2116 uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
2078 #endif 2117 #endif
2079 #ifndef CONFIG_PTE_MARKER_UFFD_WP 2118 #ifndef CONFIG_PTE_MARKER_UFFD_WP
2080 uffdio_api.features &= ~UFFD_FEATURE_ 2119 uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM;
2081 uffdio_api.features &= ~UFFD_FEATURE_ 2120 uffdio_api.features &= ~UFFD_FEATURE_WP_UNPOPULATED;
2082 uffdio_api.features &= ~UFFD_FEATURE_ 2121 uffdio_api.features &= ~UFFD_FEATURE_WP_ASYNC;
2083 #endif 2122 #endif
2084 <<
2085 ret = -EINVAL; <<
2086 if (features & ~uffdio_api.features) <<
2087 goto err_out; <<
2088 <<
2089 uffdio_api.ioctls = UFFD_API_IOCTLS; 2123 uffdio_api.ioctls = UFFD_API_IOCTLS;
2090 ret = -EFAULT; 2124 ret = -EFAULT;
2091 if (copy_to_user(buf, &uffdio_api, si 2125 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2092 goto out; 2126 goto out;
2093 2127
2094 /* only enable the requested features 2128 /* only enable the requested features for this uffd context */
2095 ctx_features = uffd_ctx_features(feat 2129 ctx_features = uffd_ctx_features(features);
2096 ret = -EINVAL; 2130 ret = -EINVAL;
2097 if (cmpxchg(&ctx->features, 0, ctx_fe 2131 if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
2098 goto err_out; 2132 goto err_out;
2099 2133
2100 ret = 0; 2134 ret = 0;
2101 out: 2135 out:
2102 return ret; 2136 return ret;
2103 err_out: 2137 err_out:
2104 memset(&uffdio_api, 0, sizeof(uffdio_ 2138 memset(&uffdio_api, 0, sizeof(uffdio_api));
2105 if (copy_to_user(buf, &uffdio_api, si 2139 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2106 ret = -EFAULT; 2140 ret = -EFAULT;
2107 goto out; 2141 goto out;
2108 } 2142 }
2109 2143
2110 static long userfaultfd_ioctl(struct file *fi 2144 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
2111 unsigned long a 2145 unsigned long arg)
2112 { 2146 {
2113 int ret = -EINVAL; 2147 int ret = -EINVAL;
2114 struct userfaultfd_ctx *ctx = file->p 2148 struct userfaultfd_ctx *ctx = file->private_data;
2115 2149
2116 if (cmd != UFFDIO_API && !userfaultfd 2150 if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx))
2117 return -EINVAL; 2151 return -EINVAL;
2118 2152
2119 switch(cmd) { 2153 switch(cmd) {
2120 case UFFDIO_API: 2154 case UFFDIO_API:
2121 ret = userfaultfd_api(ctx, ar 2155 ret = userfaultfd_api(ctx, arg);
2122 break; 2156 break;
2123 case UFFDIO_REGISTER: 2157 case UFFDIO_REGISTER:
2124 ret = userfaultfd_register(ct 2158 ret = userfaultfd_register(ctx, arg);
2125 break; 2159 break;
2126 case UFFDIO_UNREGISTER: 2160 case UFFDIO_UNREGISTER:
2127 ret = userfaultfd_unregister( 2161 ret = userfaultfd_unregister(ctx, arg);
2128 break; 2162 break;
2129 case UFFDIO_WAKE: 2163 case UFFDIO_WAKE:
2130 ret = userfaultfd_wake(ctx, a 2164 ret = userfaultfd_wake(ctx, arg);
2131 break; 2165 break;
2132 case UFFDIO_COPY: 2166 case UFFDIO_COPY:
2133 ret = userfaultfd_copy(ctx, a 2167 ret = userfaultfd_copy(ctx, arg);
2134 break; 2168 break;
2135 case UFFDIO_ZEROPAGE: 2169 case UFFDIO_ZEROPAGE:
2136 ret = userfaultfd_zeropage(ct 2170 ret = userfaultfd_zeropage(ctx, arg);
2137 break; 2171 break;
2138 case UFFDIO_MOVE: 2172 case UFFDIO_MOVE:
2139 ret = userfaultfd_move(ctx, a 2173 ret = userfaultfd_move(ctx, arg);
2140 break; 2174 break;
2141 case UFFDIO_WRITEPROTECT: 2175 case UFFDIO_WRITEPROTECT:
2142 ret = userfaultfd_writeprotec 2176 ret = userfaultfd_writeprotect(ctx, arg);
2143 break; 2177 break;
2144 case UFFDIO_CONTINUE: 2178 case UFFDIO_CONTINUE:
2145 ret = userfaultfd_continue(ct 2179 ret = userfaultfd_continue(ctx, arg);
2146 break; 2180 break;
2147 case UFFDIO_POISON: 2181 case UFFDIO_POISON:
2148 ret = userfaultfd_poison(ctx, 2182 ret = userfaultfd_poison(ctx, arg);
2149 break; 2183 break;
2150 } 2184 }
2151 return ret; 2185 return ret;
2152 } 2186 }
2153 2187
2154 #ifdef CONFIG_PROC_FS 2188 #ifdef CONFIG_PROC_FS
2155 static void userfaultfd_show_fdinfo(struct se 2189 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2156 { 2190 {
2157 struct userfaultfd_ctx *ctx = f->priv 2191 struct userfaultfd_ctx *ctx = f->private_data;
2158 wait_queue_entry_t *wq; 2192 wait_queue_entry_t *wq;
2159 unsigned long pending = 0, total = 0; 2193 unsigned long pending = 0, total = 0;
2160 2194
2161 spin_lock_irq(&ctx->fault_pending_wqh 2195 spin_lock_irq(&ctx->fault_pending_wqh.lock);
2162 list_for_each_entry(wq, &ctx->fault_p 2196 list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2163 pending++; 2197 pending++;
2164 total++; 2198 total++;
2165 } 2199 }
2166 list_for_each_entry(wq, &ctx->fault_w 2200 list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2167 total++; 2201 total++;
2168 } 2202 }
2169 spin_unlock_irq(&ctx->fault_pending_w 2203 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2170 2204
2171 /* 2205 /*
2172 * If more protocols will be added, t 2206 * If more protocols will be added, there will be all shown
2173 * separated by a space. Like this: 2207 * separated by a space. Like this:
2174 * protocols: aa:... bb:... 2208 * protocols: aa:... bb:...
2175 */ 2209 */
2176 seq_printf(m, "pending:\t%lu\ntotal:\ 2210 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2177 pending, total, UFFD_API, 2211 pending, total, UFFD_API, ctx->features,
2178 UFFD_API_IOCTLS|UFFD_API_R 2212 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2179 } 2213 }
2180 #endif 2214 #endif
2181 2215
2182 static const struct file_operations userfault 2216 static const struct file_operations userfaultfd_fops = {
2183 #ifdef CONFIG_PROC_FS 2217 #ifdef CONFIG_PROC_FS
2184 .show_fdinfo = userfaultfd_show_fd 2218 .show_fdinfo = userfaultfd_show_fdinfo,
2185 #endif 2219 #endif
2186 .release = userfaultfd_release 2220 .release = userfaultfd_release,
2187 .poll = userfaultfd_poll, 2221 .poll = userfaultfd_poll,
2188 .read_iter = userfaultfd_read_it !! 2222 .read = userfaultfd_read,
2189 .unlocked_ioctl = userfaultfd_ioctl, 2223 .unlocked_ioctl = userfaultfd_ioctl,
2190 .compat_ioctl = compat_ptr_ioctl, 2224 .compat_ioctl = compat_ptr_ioctl,
2191 .llseek = noop_llseek, 2225 .llseek = noop_llseek,
2192 }; 2226 };
2193 2227
2194 static void init_once_userfaultfd_ctx(void *m 2228 static void init_once_userfaultfd_ctx(void *mem)
2195 { 2229 {
2196 struct userfaultfd_ctx *ctx = (struct 2230 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2197 2231
2198 init_waitqueue_head(&ctx->fault_pendi 2232 init_waitqueue_head(&ctx->fault_pending_wqh);
2199 init_waitqueue_head(&ctx->fault_wqh); 2233 init_waitqueue_head(&ctx->fault_wqh);
2200 init_waitqueue_head(&ctx->event_wqh); 2234 init_waitqueue_head(&ctx->event_wqh);
2201 init_waitqueue_head(&ctx->fd_wqh); 2235 init_waitqueue_head(&ctx->fd_wqh);
2202 seqcount_spinlock_init(&ctx->refile_s 2236 seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2203 } 2237 }
2204 2238
2205 static int new_userfaultfd(int flags) 2239 static int new_userfaultfd(int flags)
2206 { 2240 {
2207 struct userfaultfd_ctx *ctx; 2241 struct userfaultfd_ctx *ctx;
2208 struct file *file; <<
2209 int fd; 2242 int fd;
2210 2243
2211 BUG_ON(!current->mm); 2244 BUG_ON(!current->mm);
2212 2245
2213 /* Check the UFFD_* constants for con 2246 /* Check the UFFD_* constants for consistency. */
2214 BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UF 2247 BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2215 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXE 2248 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2216 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBL 2249 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2217 2250
2218 if (flags & ~(UFFD_SHARED_FCNTL_FLAGS 2251 if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2219 return -EINVAL; 2252 return -EINVAL;
2220 2253
2221 ctx = kmem_cache_alloc(userfaultfd_ct 2254 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2222 if (!ctx) 2255 if (!ctx)
2223 return -ENOMEM; 2256 return -ENOMEM;
2224 2257
2225 refcount_set(&ctx->refcount, 1); 2258 refcount_set(&ctx->refcount, 1);
2226 ctx->flags = flags; 2259 ctx->flags = flags;
2227 ctx->features = 0; 2260 ctx->features = 0;
2228 ctx->released = false; 2261 ctx->released = false;
2229 init_rwsem(&ctx->map_changing_lock); <<
2230 atomic_set(&ctx->mmap_changing, 0); 2262 atomic_set(&ctx->mmap_changing, 0);
2231 ctx->mm = current->mm; 2263 ctx->mm = current->mm;
2232 !! 2264 /* prevent the mm struct to be freed */
2233 fd = get_unused_fd_flags(flags & UFFD !! 2265 mmgrab(ctx->mm);
2234 if (fd < 0) <<
2235 goto err_out; <<
2236 2266
2237 /* Create a new inode so that the LSM 2267 /* Create a new inode so that the LSM can block the creation. */
2238 file = anon_inode_create_getfile("[us !! 2268 fd = anon_inode_create_getfd("[userfaultfd]", &userfaultfd_fops, ctx,
2239 O_RDONLY | (flags & U 2269 O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2240 if (IS_ERR(file)) { !! 2270 if (fd < 0) {
2241 put_unused_fd(fd); !! 2271 mmdrop(ctx->mm);
2242 fd = PTR_ERR(file); !! 2272 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2243 goto err_out; <<
2244 } 2273 }
2245 /* prevent the mm struct to be freed <<
2246 mmgrab(ctx->mm); <<
2247 file->f_mode |= FMODE_NOWAIT; <<
2248 fd_install(fd, file); <<
2249 return fd; <<
2250 err_out: <<
2251 kmem_cache_free(userfaultfd_ctx_cache <<
2252 return fd; 2274 return fd;
2253 } 2275 }
2254 2276
2255 static inline bool userfaultfd_syscall_allowe 2277 static inline bool userfaultfd_syscall_allowed(int flags)
2256 { 2278 {
2257 /* Userspace-only page faults are alw 2279 /* Userspace-only page faults are always allowed */
2258 if (flags & UFFD_USER_MODE_ONLY) 2280 if (flags & UFFD_USER_MODE_ONLY)
2259 return true; 2281 return true;
2260 2282
2261 /* 2283 /*
2262 * The user is requesting a userfault 2284 * The user is requesting a userfaultfd which can handle kernel faults.
2263 * Privileged users are always allowe 2285 * Privileged users are always allowed to do this.
2264 */ 2286 */
2265 if (capable(CAP_SYS_PTRACE)) 2287 if (capable(CAP_SYS_PTRACE))
2266 return true; 2288 return true;
2267 2289
2268 /* Otherwise, access to kernel fault 2290 /* Otherwise, access to kernel fault handling is sysctl controlled. */
2269 return sysctl_unprivileged_userfaultf 2291 return sysctl_unprivileged_userfaultfd;
2270 } 2292 }
2271 2293
2272 SYSCALL_DEFINE1(userfaultfd, int, flags) 2294 SYSCALL_DEFINE1(userfaultfd, int, flags)
2273 { 2295 {
2274 if (!userfaultfd_syscall_allowed(flag 2296 if (!userfaultfd_syscall_allowed(flags))
2275 return -EPERM; 2297 return -EPERM;
2276 2298
2277 return new_userfaultfd(flags); 2299 return new_userfaultfd(flags);
2278 } 2300 }
2279 2301
2280 static long userfaultfd_dev_ioctl(struct file 2302 static long userfaultfd_dev_ioctl(struct file *file, unsigned int cmd, unsigned long flags)
2281 { 2303 {
2282 if (cmd != USERFAULTFD_IOC_NEW) 2304 if (cmd != USERFAULTFD_IOC_NEW)
2283 return -EINVAL; 2305 return -EINVAL;
2284 2306
2285 return new_userfaultfd(flags); 2307 return new_userfaultfd(flags);
2286 } 2308 }
2287 2309
2288 static const struct file_operations userfault 2310 static const struct file_operations userfaultfd_dev_fops = {
2289 .unlocked_ioctl = userfaultfd_dev_ioc 2311 .unlocked_ioctl = userfaultfd_dev_ioctl,
2290 .compat_ioctl = userfaultfd_dev_ioctl 2312 .compat_ioctl = userfaultfd_dev_ioctl,
2291 .owner = THIS_MODULE, 2313 .owner = THIS_MODULE,
2292 .llseek = noop_llseek, 2314 .llseek = noop_llseek,
2293 }; 2315 };
2294 2316
2295 static struct miscdevice userfaultfd_misc = { 2317 static struct miscdevice userfaultfd_misc = {
2296 .minor = MISC_DYNAMIC_MINOR, 2318 .minor = MISC_DYNAMIC_MINOR,
2297 .name = "userfaultfd", 2319 .name = "userfaultfd",
2298 .fops = &userfaultfd_dev_fops 2320 .fops = &userfaultfd_dev_fops
2299 }; 2321 };
2300 2322
2301 static int __init userfaultfd_init(void) 2323 static int __init userfaultfd_init(void)
2302 { 2324 {
2303 int ret; 2325 int ret;
2304 2326
2305 ret = misc_register(&userfaultfd_misc 2327 ret = misc_register(&userfaultfd_misc);
2306 if (ret) 2328 if (ret)
2307 return ret; 2329 return ret;
2308 2330
2309 userfaultfd_ctx_cachep = kmem_cache_c 2331 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2310 2332 sizeof(struct userfaultfd_ctx),
2311 2333 0,
2312 2334 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2313 2335 init_once_userfaultfd_ctx);
2314 #ifdef CONFIG_SYSCTL 2336 #ifdef CONFIG_SYSCTL
2315 register_sysctl_init("vm", vm_userfau 2337 register_sysctl_init("vm", vm_userfaultfd_table);
2316 #endif 2338 #endif
2317 return 0; 2339 return 0;
2318 } 2340 }
2319 __initcall(userfaultfd_init); 2341 __initcall(userfaultfd_init);
2320 2342
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.