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