1 /* 1 /*
2 * An async IO implementation for Linux 2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvac 3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 * 4 *
5 * Implements an efficient asynchronous i 5 * Implements an efficient asynchronous io interface.
6 * 6 *
7 * Copyright 2000, 2001, 2002 Red Hat, In 7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 * Copyright 2018 Christoph Hellwig. 8 * Copyright 2018 Christoph Hellwig.
9 * 9 *
10 * See ../COPYING for licensing terms. 10 * See ../COPYING for licensing terms.
11 */ 11 */
12 #define pr_fmt(fmt) "%s: " fmt, __func__ 12 #define pr_fmt(fmt) "%s: " fmt, __func__
13 13
14 #include <linux/kernel.h> 14 #include <linux/kernel.h>
15 #include <linux/init.h> 15 #include <linux/init.h>
16 #include <linux/errno.h> 16 #include <linux/errno.h>
17 #include <linux/time.h> 17 #include <linux/time.h>
18 #include <linux/aio_abi.h> 18 #include <linux/aio_abi.h>
19 #include <linux/export.h> 19 #include <linux/export.h>
20 #include <linux/syscalls.h> 20 #include <linux/syscalls.h>
21 #include <linux/backing-dev.h> 21 #include <linux/backing-dev.h>
22 #include <linux/refcount.h> 22 #include <linux/refcount.h>
23 #include <linux/uio.h> 23 #include <linux/uio.h>
24 24
25 #include <linux/sched/signal.h> 25 #include <linux/sched/signal.h>
26 #include <linux/fs.h> 26 #include <linux/fs.h>
27 #include <linux/file.h> 27 #include <linux/file.h>
28 #include <linux/mm.h> 28 #include <linux/mm.h>
29 #include <linux/mman.h> 29 #include <linux/mman.h>
30 #include <linux/percpu.h> 30 #include <linux/percpu.h>
31 #include <linux/slab.h> 31 #include <linux/slab.h>
32 #include <linux/timer.h> 32 #include <linux/timer.h>
33 #include <linux/aio.h> 33 #include <linux/aio.h>
34 #include <linux/highmem.h> 34 #include <linux/highmem.h>
35 #include <linux/workqueue.h> 35 #include <linux/workqueue.h>
36 #include <linux/security.h> 36 #include <linux/security.h>
37 #include <linux/eventfd.h> 37 #include <linux/eventfd.h>
38 #include <linux/blkdev.h> 38 #include <linux/blkdev.h>
39 #include <linux/compat.h> 39 #include <linux/compat.h>
40 #include <linux/migrate.h> 40 #include <linux/migrate.h>
41 #include <linux/ramfs.h> 41 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.h> 42 #include <linux/percpu-refcount.h>
43 #include <linux/mount.h> 43 #include <linux/mount.h>
44 #include <linux/pseudo_fs.h> 44 #include <linux/pseudo_fs.h>
45 45
46 #include <linux/uaccess.h> 46 #include <linux/uaccess.h>
47 #include <linux/nospec.h> 47 #include <linux/nospec.h>
48 48
49 #include "internal.h" 49 #include "internal.h"
50 50
51 #define KIOCB_KEY 0 51 #define KIOCB_KEY 0
52 52
53 #define AIO_RING_MAGIC 0xa10a 53 #define AIO_RING_MAGIC 0xa10a10a1
54 #define AIO_RING_COMPAT_FEATURES 1 54 #define AIO_RING_COMPAT_FEATURES 1
55 #define AIO_RING_INCOMPAT_FEATURES 0 55 #define AIO_RING_INCOMPAT_FEATURES 0
56 struct aio_ring { 56 struct aio_ring {
57 unsigned id; /* kernel inte 57 unsigned id; /* kernel internal index number */
58 unsigned nr; /* number of i 58 unsigned nr; /* number of io_events */
59 unsigned head; /* Written to 59 unsigned head; /* Written to by userland or under ring_lock
60 * mutex by ai 60 * mutex by aio_read_events_ring(). */
61 unsigned tail; 61 unsigned tail;
62 62
63 unsigned magic; 63 unsigned magic;
64 unsigned compat_features; 64 unsigned compat_features;
65 unsigned incompat_features; 65 unsigned incompat_features;
66 unsigned header_length; /* siz 66 unsigned header_length; /* size of aio_ring */
67 67
68 68
69 struct io_event io_events[]; 69 struct io_event io_events[];
70 }; /* 128 bytes + ring size */ 70 }; /* 128 bytes + ring size */
71 71
72 /* 72 /*
73 * Plugging is meant to work with larger batch 73 * Plugging is meant to work with larger batches of IOs. If we don't
74 * have more than the below, then don't bother 74 * have more than the below, then don't bother setting up a plug.
75 */ 75 */
76 #define AIO_PLUG_THRESHOLD 2 76 #define AIO_PLUG_THRESHOLD 2
77 77
78 #define AIO_RING_PAGES 8 78 #define AIO_RING_PAGES 8
79 79
80 struct kioctx_table { 80 struct kioctx_table {
81 struct rcu_head rcu; 81 struct rcu_head rcu;
82 unsigned nr; 82 unsigned nr;
83 struct kioctx __rcu *table[] __cou 83 struct kioctx __rcu *table[] __counted_by(nr);
84 }; 84 };
85 85
86 struct kioctx_cpu { 86 struct kioctx_cpu {
87 unsigned reqs_available 87 unsigned reqs_available;
88 }; 88 };
89 89
90 struct ctx_rq_wait { 90 struct ctx_rq_wait {
91 struct completion comp; 91 struct completion comp;
92 atomic_t count; 92 atomic_t count;
93 }; 93 };
94 94
95 struct kioctx { 95 struct kioctx {
96 struct percpu_ref users; 96 struct percpu_ref users;
97 atomic_t dead; 97 atomic_t dead;
98 98
99 struct percpu_ref reqs; 99 struct percpu_ref reqs;
100 100
101 unsigned long user_id; 101 unsigned long user_id;
102 102
103 struct kioctx_cpu __percpu *cpu; 103 struct kioctx_cpu __percpu *cpu;
104 104
105 /* 105 /*
106 * For percpu reqs_available, number o 106 * For percpu reqs_available, number of slots we move to/from global
107 * counter at a time: 107 * counter at a time:
108 */ 108 */
109 unsigned req_batch; 109 unsigned req_batch;
110 /* 110 /*
111 * This is what userspace passed to io 111 * This is what userspace passed to io_setup(), it's not used for
112 * anything but counting against the g 112 * anything but counting against the global max_reqs quota.
113 * 113 *
114 * The real limit is nr_events - 1, wh 114 * The real limit is nr_events - 1, which will be larger (see
115 * aio_setup_ring()) 115 * aio_setup_ring())
116 */ 116 */
117 unsigned max_reqs; 117 unsigned max_reqs;
118 118
119 /* Size of ringbuffer, in units of str 119 /* Size of ringbuffer, in units of struct io_event */
120 unsigned nr_events; 120 unsigned nr_events;
121 121
122 unsigned long mmap_base; 122 unsigned long mmap_base;
123 unsigned long mmap_size; 123 unsigned long mmap_size;
124 124
125 struct folio **ring_folios; 125 struct folio **ring_folios;
126 long nr_pages; 126 long nr_pages;
127 127
128 struct rcu_work free_rwork; 128 struct rcu_work free_rwork; /* see free_ioctx() */
129 129
130 /* 130 /*
131 * signals when all in-flight requests 131 * signals when all in-flight requests are done
132 */ 132 */
133 struct ctx_rq_wait *rq_wait; 133 struct ctx_rq_wait *rq_wait;
134 134
135 struct { 135 struct {
136 /* 136 /*
137 * This counts the number of a 137 * This counts the number of available slots in the ringbuffer,
138 * so we avoid overflowing it: 138 * so we avoid overflowing it: it's decremented (if positive)
139 * when allocating a kiocb and 139 * when allocating a kiocb and incremented when the resulting
140 * io_event is pulled off the 140 * io_event is pulled off the ringbuffer.
141 * 141 *
142 * We batch accesses to it wit 142 * We batch accesses to it with a percpu version.
143 */ 143 */
144 atomic_t reqs_available 144 atomic_t reqs_available;
145 } ____cacheline_aligned_in_smp; 145 } ____cacheline_aligned_in_smp;
146 146
147 struct { 147 struct {
148 spinlock_t ctx_lock; 148 spinlock_t ctx_lock;
149 struct list_head active_reqs; 149 struct list_head active_reqs; /* used for cancellation */
150 } ____cacheline_aligned_in_smp; 150 } ____cacheline_aligned_in_smp;
151 151
152 struct { 152 struct {
153 struct mutex ring_lock; 153 struct mutex ring_lock;
154 wait_queue_head_t wait; 154 wait_queue_head_t wait;
155 } ____cacheline_aligned_in_smp; 155 } ____cacheline_aligned_in_smp;
156 156
157 struct { 157 struct {
158 unsigned tail; 158 unsigned tail;
159 unsigned completed_even 159 unsigned completed_events;
160 spinlock_t completion_loc 160 spinlock_t completion_lock;
161 } ____cacheline_aligned_in_smp; 161 } ____cacheline_aligned_in_smp;
162 162
163 struct folio *internal_foli 163 struct folio *internal_folios[AIO_RING_PAGES];
164 struct file *aio_ring_file 164 struct file *aio_ring_file;
165 165
166 unsigned id; 166 unsigned id;
167 }; 167 };
168 168
169 /* 169 /*
170 * First field must be the file pointer in all 170 * First field must be the file pointer in all the
171 * iocb unions! See also 'struct kiocb' in <li 171 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
172 */ 172 */
173 struct fsync_iocb { 173 struct fsync_iocb {
174 struct file *file; 174 struct file *file;
175 struct work_struct work; 175 struct work_struct work;
176 bool datasync; 176 bool datasync;
177 struct cred *creds; 177 struct cred *creds;
178 }; 178 };
179 179
180 struct poll_iocb { 180 struct poll_iocb {
181 struct file *file; 181 struct file *file;
182 struct wait_queue_head *head; 182 struct wait_queue_head *head;
183 __poll_t events; 183 __poll_t events;
184 bool cancelled; 184 bool cancelled;
185 bool work_scheduled 185 bool work_scheduled;
186 bool work_need_resc 186 bool work_need_resched;
187 struct wait_queue_entry wait; 187 struct wait_queue_entry wait;
188 struct work_struct work; 188 struct work_struct work;
189 }; 189 };
190 190
191 /* 191 /*
192 * NOTE! Each of the iocb union members has th 192 * NOTE! Each of the iocb union members has the file pointer
193 * as the first entry in their struct definiti 193 * as the first entry in their struct definition. So you can
194 * access the file pointer through any of the 194 * access the file pointer through any of the sub-structs,
195 * or directly as just 'ki_filp' in this struc 195 * or directly as just 'ki_filp' in this struct.
196 */ 196 */
197 struct aio_kiocb { 197 struct aio_kiocb {
198 union { 198 union {
199 struct file *ki_fi 199 struct file *ki_filp;
200 struct kiocb rw; 200 struct kiocb rw;
201 struct fsync_iocb fsync; 201 struct fsync_iocb fsync;
202 struct poll_iocb poll; 202 struct poll_iocb poll;
203 }; 203 };
204 204
205 struct kioctx *ki_ctx; 205 struct kioctx *ki_ctx;
206 kiocb_cancel_fn *ki_cancel; 206 kiocb_cancel_fn *ki_cancel;
207 207
208 struct io_event ki_res; 208 struct io_event ki_res;
209 209
210 struct list_head ki_list; 210 struct list_head ki_list; /* the aio core uses this
211 211 * for cancellation */
212 refcount_t ki_refcnt; 212 refcount_t ki_refcnt;
213 213
214 /* 214 /*
215 * If the aio_resfd field of the users 215 * If the aio_resfd field of the userspace iocb is not zero,
216 * this is the underlying eventfd cont 216 * this is the underlying eventfd context to deliver events to.
217 */ 217 */
218 struct eventfd_ctx *ki_eventfd; 218 struct eventfd_ctx *ki_eventfd;
219 }; 219 };
220 220
221 /*------ sysctl variables----*/ 221 /*------ sysctl variables----*/
222 static DEFINE_SPINLOCK(aio_nr_lock); 222 static DEFINE_SPINLOCK(aio_nr_lock);
223 static unsigned long aio_nr; /* cur 223 static unsigned long aio_nr; /* current system wide number of aio requests */
224 static unsigned long aio_max_nr = 0x10000; /* 224 static unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/ 225 /*----end sysctl variables---*/
226 #ifdef CONFIG_SYSCTL 226 #ifdef CONFIG_SYSCTL
227 static struct ctl_table aio_sysctls[] = { 227 static struct ctl_table aio_sysctls[] = {
228 { 228 {
229 .procname = "aio-nr", 229 .procname = "aio-nr",
230 .data = &aio_nr, 230 .data = &aio_nr,
231 .maxlen = sizeof(aio_n 231 .maxlen = sizeof(aio_nr),
232 .mode = 0444, 232 .mode = 0444,
233 .proc_handler = proc_doulong 233 .proc_handler = proc_doulongvec_minmax,
234 }, 234 },
235 { 235 {
236 .procname = "aio-max-nr" 236 .procname = "aio-max-nr",
237 .data = &aio_max_nr, 237 .data = &aio_max_nr,
238 .maxlen = sizeof(aio_m 238 .maxlen = sizeof(aio_max_nr),
239 .mode = 0644, 239 .mode = 0644,
240 .proc_handler = proc_doulong 240 .proc_handler = proc_doulongvec_minmax,
241 }, 241 },
242 }; 242 };
243 243
244 static void __init aio_sysctl_init(void) 244 static void __init aio_sysctl_init(void)
245 { 245 {
246 register_sysctl_init("fs", aio_sysctls 246 register_sysctl_init("fs", aio_sysctls);
247 } 247 }
248 #else 248 #else
249 #define aio_sysctl_init() do { } while (0) 249 #define aio_sysctl_init() do { } while (0)
250 #endif 250 #endif
251 251
252 static struct kmem_cache *kiocb_cachep; 252 static struct kmem_cache *kiocb_cachep;
253 static struct kmem_cache *kioctx_cachep 253 static struct kmem_cache *kioctx_cachep;
254 254
255 static struct vfsmount *aio_mnt; 255 static struct vfsmount *aio_mnt;
256 256
257 static const struct file_operations aio_ring_f 257 static const struct file_operations aio_ring_fops;
258 static const struct address_space_operations a 258 static const struct address_space_operations aio_ctx_aops;
259 259
260 static struct file *aio_private_file(struct ki 260 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
261 { 261 {
262 struct file *file; 262 struct file *file;
263 struct inode *inode = alloc_anon_inode 263 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
264 if (IS_ERR(inode)) 264 if (IS_ERR(inode))
265 return ERR_CAST(inode); 265 return ERR_CAST(inode);
266 266
267 inode->i_mapping->a_ops = &aio_ctx_aop 267 inode->i_mapping->a_ops = &aio_ctx_aops;
268 inode->i_mapping->i_private_data = ctx 268 inode->i_mapping->i_private_data = ctx;
269 inode->i_size = PAGE_SIZE * nr_pages; 269 inode->i_size = PAGE_SIZE * nr_pages;
270 270
271 file = alloc_file_pseudo(inode, aio_mn 271 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
272 O_RDWR, &aio_r 272 O_RDWR, &aio_ring_fops);
273 if (IS_ERR(file)) 273 if (IS_ERR(file))
274 iput(inode); 274 iput(inode);
275 return file; 275 return file;
276 } 276 }
277 277
278 static int aio_init_fs_context(struct fs_conte 278 static int aio_init_fs_context(struct fs_context *fc)
279 { 279 {
280 if (!init_pseudo(fc, AIO_RING_MAGIC)) 280 if (!init_pseudo(fc, AIO_RING_MAGIC))
281 return -ENOMEM; 281 return -ENOMEM;
282 fc->s_iflags |= SB_I_NOEXEC; 282 fc->s_iflags |= SB_I_NOEXEC;
283 return 0; 283 return 0;
284 } 284 }
285 285
286 /* aio_setup 286 /* aio_setup
287 * Creates the slab caches used by the ai 287 * Creates the slab caches used by the aio routines, panic on
288 * failure as this is done early during t 288 * failure as this is done early during the boot sequence.
289 */ 289 */
290 static int __init aio_setup(void) 290 static int __init aio_setup(void)
291 { 291 {
292 static struct file_system_type aio_fs 292 static struct file_system_type aio_fs = {
293 .name = "aio", 293 .name = "aio",
294 .init_fs_context = aio_init_fs 294 .init_fs_context = aio_init_fs_context,
295 .kill_sb = kill_anon_su 295 .kill_sb = kill_anon_super,
296 }; 296 };
297 aio_mnt = kern_mount(&aio_fs); 297 aio_mnt = kern_mount(&aio_fs);
298 if (IS_ERR(aio_mnt)) 298 if (IS_ERR(aio_mnt))
299 panic("Failed to create aio fs 299 panic("Failed to create aio fs mount.");
300 300
301 kiocb_cachep = KMEM_CACHE(aio_kiocb, S 301 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
302 kioctx_cachep = KMEM_CACHE(kioctx,SLAB 302 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
303 aio_sysctl_init(); 303 aio_sysctl_init();
304 return 0; 304 return 0;
305 } 305 }
306 __initcall(aio_setup); 306 __initcall(aio_setup);
307 307
308 static void put_aio_ring_file(struct kioctx *c 308 static void put_aio_ring_file(struct kioctx *ctx)
309 { 309 {
310 struct file *aio_ring_file = ctx->aio_ 310 struct file *aio_ring_file = ctx->aio_ring_file;
311 struct address_space *i_mapping; 311 struct address_space *i_mapping;
312 312
313 if (aio_ring_file) { 313 if (aio_ring_file) {
314 truncate_setsize(file_inode(ai 314 truncate_setsize(file_inode(aio_ring_file), 0);
315 315
316 /* Prevent further access to t 316 /* Prevent further access to the kioctx from migratepages */
317 i_mapping = aio_ring_file->f_m 317 i_mapping = aio_ring_file->f_mapping;
318 spin_lock(&i_mapping->i_privat 318 spin_lock(&i_mapping->i_private_lock);
319 i_mapping->i_private_data = NU 319 i_mapping->i_private_data = NULL;
320 ctx->aio_ring_file = NULL; 320 ctx->aio_ring_file = NULL;
321 spin_unlock(&i_mapping->i_priv 321 spin_unlock(&i_mapping->i_private_lock);
322 322
323 fput(aio_ring_file); 323 fput(aio_ring_file);
324 } 324 }
325 } 325 }
326 326
327 static void aio_free_ring(struct kioctx *ctx) 327 static void aio_free_ring(struct kioctx *ctx)
328 { 328 {
329 int i; 329 int i;
330 330
331 /* Disconnect the kiotx from the ring 331 /* Disconnect the kiotx from the ring file. This prevents future
332 * accesses to the kioctx from page mi 332 * accesses to the kioctx from page migration.
333 */ 333 */
334 put_aio_ring_file(ctx); 334 put_aio_ring_file(ctx);
335 335
336 for (i = 0; i < ctx->nr_pages; i++) { 336 for (i = 0; i < ctx->nr_pages; i++) {
337 struct folio *folio = ctx->rin 337 struct folio *folio = ctx->ring_folios[i];
338 338
339 if (!folio) 339 if (!folio)
340 continue; 340 continue;
341 341
342 pr_debug("pid(%d) [%d] folio-> 342 pr_debug("pid(%d) [%d] folio->count=%d\n", current->pid, i,
343 folio_ref_count(folio 343 folio_ref_count(folio));
344 ctx->ring_folios[i] = NULL; 344 ctx->ring_folios[i] = NULL;
345 folio_put(folio); 345 folio_put(folio);
346 } 346 }
347 347
348 if (ctx->ring_folios && ctx->ring_foli 348 if (ctx->ring_folios && ctx->ring_folios != ctx->internal_folios) {
349 kfree(ctx->ring_folios); 349 kfree(ctx->ring_folios);
350 ctx->ring_folios = NULL; 350 ctx->ring_folios = NULL;
351 } 351 }
352 } 352 }
353 353
354 static int aio_ring_mremap(struct vm_area_stru 354 static int aio_ring_mremap(struct vm_area_struct *vma)
355 { 355 {
356 struct file *file = vma->vm_file; 356 struct file *file = vma->vm_file;
357 struct mm_struct *mm = vma->vm_mm; 357 struct mm_struct *mm = vma->vm_mm;
358 struct kioctx_table *table; 358 struct kioctx_table *table;
359 int i, res = -EINVAL; 359 int i, res = -EINVAL;
360 360
361 spin_lock(&mm->ioctx_lock); 361 spin_lock(&mm->ioctx_lock);
362 rcu_read_lock(); 362 rcu_read_lock();
363 table = rcu_dereference(mm->ioctx_tabl 363 table = rcu_dereference(mm->ioctx_table);
364 if (!table) 364 if (!table)
365 goto out_unlock; 365 goto out_unlock;
366 366
367 for (i = 0; i < table->nr; i++) { 367 for (i = 0; i < table->nr; i++) {
368 struct kioctx *ctx; 368 struct kioctx *ctx;
369 369
370 ctx = rcu_dereference(table->t 370 ctx = rcu_dereference(table->table[i]);
371 if (ctx && ctx->aio_ring_file 371 if (ctx && ctx->aio_ring_file == file) {
372 if (!atomic_read(&ctx- 372 if (!atomic_read(&ctx->dead)) {
373 ctx->user_id = 373 ctx->user_id = ctx->mmap_base = vma->vm_start;
374 res = 0; 374 res = 0;
375 } 375 }
376 break; 376 break;
377 } 377 }
378 } 378 }
379 379
380 out_unlock: 380 out_unlock:
381 rcu_read_unlock(); 381 rcu_read_unlock();
382 spin_unlock(&mm->ioctx_lock); 382 spin_unlock(&mm->ioctx_lock);
383 return res; 383 return res;
384 } 384 }
385 385
386 static const struct vm_operations_struct aio_r 386 static const struct vm_operations_struct aio_ring_vm_ops = {
387 .mremap = aio_ring_mremap, 387 .mremap = aio_ring_mremap,
388 #if IS_ENABLED(CONFIG_MMU) 388 #if IS_ENABLED(CONFIG_MMU)
389 .fault = filemap_fault, 389 .fault = filemap_fault,
390 .map_pages = filemap_map_pages, 390 .map_pages = filemap_map_pages,
391 .page_mkwrite = filemap_page_mkwrite 391 .page_mkwrite = filemap_page_mkwrite,
392 #endif 392 #endif
393 }; 393 };
394 394
395 static int aio_ring_mmap(struct file *file, st 395 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
396 { 396 {
397 vm_flags_set(vma, VM_DONTEXPAND); 397 vm_flags_set(vma, VM_DONTEXPAND);
398 vma->vm_ops = &aio_ring_vm_ops; 398 vma->vm_ops = &aio_ring_vm_ops;
399 return 0; 399 return 0;
400 } 400 }
401 401
402 static const struct file_operations aio_ring_f 402 static const struct file_operations aio_ring_fops = {
403 .mmap = aio_ring_mmap, 403 .mmap = aio_ring_mmap,
404 }; 404 };
405 405
406 #if IS_ENABLED(CONFIG_MIGRATION) 406 #if IS_ENABLED(CONFIG_MIGRATION)
407 static int aio_migrate_folio(struct address_sp 407 static int aio_migrate_folio(struct address_space *mapping, struct folio *dst,
408 struct folio *src, enu 408 struct folio *src, enum migrate_mode mode)
409 { 409 {
410 struct kioctx *ctx; 410 struct kioctx *ctx;
411 unsigned long flags; 411 unsigned long flags;
412 pgoff_t idx; 412 pgoff_t idx;
413 int rc = 0; 413 int rc = 0;
414 414
415 /* mapping->i_private_lock here protec 415 /* mapping->i_private_lock here protects against the kioctx teardown. */
416 spin_lock(&mapping->i_private_lock); 416 spin_lock(&mapping->i_private_lock);
417 ctx = mapping->i_private_data; 417 ctx = mapping->i_private_data;
418 if (!ctx) { 418 if (!ctx) {
419 rc = -EINVAL; 419 rc = -EINVAL;
420 goto out; 420 goto out;
421 } 421 }
422 422
423 /* The ring_lock mutex. The prevents 423 /* The ring_lock mutex. The prevents aio_read_events() from writing
424 * to the ring's head, and prevents pa 424 * to the ring's head, and prevents page migration from mucking in
425 * a partially initialized kiotx. 425 * a partially initialized kiotx.
426 */ 426 */
427 if (!mutex_trylock(&ctx->ring_lock)) { 427 if (!mutex_trylock(&ctx->ring_lock)) {
428 rc = -EAGAIN; 428 rc = -EAGAIN;
429 goto out; 429 goto out;
430 } 430 }
431 431
432 idx = src->index; 432 idx = src->index;
433 if (idx < (pgoff_t)ctx->nr_pages) { 433 if (idx < (pgoff_t)ctx->nr_pages) {
434 /* Make sure the old folio has 434 /* Make sure the old folio hasn't already been changed */
435 if (ctx->ring_folios[idx] != s 435 if (ctx->ring_folios[idx] != src)
436 rc = -EAGAIN; 436 rc = -EAGAIN;
437 } else 437 } else
438 rc = -EINVAL; 438 rc = -EINVAL;
439 439
440 if (rc != 0) 440 if (rc != 0)
441 goto out_unlock; 441 goto out_unlock;
442 442
443 /* Writeback must be complete */ 443 /* Writeback must be complete */
444 BUG_ON(folio_test_writeback(src)); 444 BUG_ON(folio_test_writeback(src));
445 folio_get(dst); 445 folio_get(dst);
446 446
447 rc = folio_migrate_mapping(mapping, ds 447 rc = folio_migrate_mapping(mapping, dst, src, 1);
448 if (rc != MIGRATEPAGE_SUCCESS) { 448 if (rc != MIGRATEPAGE_SUCCESS) {
449 folio_put(dst); 449 folio_put(dst);
450 goto out_unlock; 450 goto out_unlock;
451 } 451 }
452 452
453 /* Take completion_lock to prevent oth 453 /* Take completion_lock to prevent other writes to the ring buffer
454 * while the old folio is copied to th 454 * while the old folio is copied to the new. This prevents new
455 * events from being lost. 455 * events from being lost.
456 */ 456 */
457 spin_lock_irqsave(&ctx->completion_loc 457 spin_lock_irqsave(&ctx->completion_lock, flags);
458 folio_copy(dst, src); 458 folio_copy(dst, src);
459 folio_migrate_flags(dst, src); 459 folio_migrate_flags(dst, src);
460 BUG_ON(ctx->ring_folios[idx] != src); 460 BUG_ON(ctx->ring_folios[idx] != src);
461 ctx->ring_folios[idx] = dst; 461 ctx->ring_folios[idx] = dst;
462 spin_unlock_irqrestore(&ctx->completio 462 spin_unlock_irqrestore(&ctx->completion_lock, flags);
463 463
464 /* The old folio is no longer accessib 464 /* The old folio is no longer accessible. */
465 folio_put(src); 465 folio_put(src);
466 466
467 out_unlock: 467 out_unlock:
468 mutex_unlock(&ctx->ring_lock); 468 mutex_unlock(&ctx->ring_lock);
469 out: 469 out:
470 spin_unlock(&mapping->i_private_lock); 470 spin_unlock(&mapping->i_private_lock);
471 return rc; 471 return rc;
472 } 472 }
473 #else 473 #else
474 #define aio_migrate_folio NULL 474 #define aio_migrate_folio NULL
475 #endif 475 #endif
476 476
477 static const struct address_space_operations a 477 static const struct address_space_operations aio_ctx_aops = {
478 .dirty_folio = noop_dirty_folio, 478 .dirty_folio = noop_dirty_folio,
479 .migrate_folio = aio_migrate_folio, 479 .migrate_folio = aio_migrate_folio,
480 }; 480 };
481 481
482 static int aio_setup_ring(struct kioctx *ctx, 482 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
483 { 483 {
484 struct aio_ring *ring; 484 struct aio_ring *ring;
485 struct mm_struct *mm = current->mm; 485 struct mm_struct *mm = current->mm;
486 unsigned long size, unused; 486 unsigned long size, unused;
487 int nr_pages; 487 int nr_pages;
488 int i; 488 int i;
489 struct file *file; 489 struct file *file;
490 490
491 /* Compensate for the ring buffer's he 491 /* Compensate for the ring buffer's head/tail overlap entry */
492 nr_events += 2; /* 1 is required, 2 fo 492 nr_events += 2; /* 1 is required, 2 for good luck */
493 493
494 size = sizeof(struct aio_ring); 494 size = sizeof(struct aio_ring);
495 size += sizeof(struct io_event) * nr_e 495 size += sizeof(struct io_event) * nr_events;
496 496
497 nr_pages = PFN_UP(size); 497 nr_pages = PFN_UP(size);
498 if (nr_pages < 0) 498 if (nr_pages < 0)
499 return -EINVAL; 499 return -EINVAL;
500 500
501 file = aio_private_file(ctx, nr_pages) 501 file = aio_private_file(ctx, nr_pages);
502 if (IS_ERR(file)) { 502 if (IS_ERR(file)) {
503 ctx->aio_ring_file = NULL; 503 ctx->aio_ring_file = NULL;
504 return -ENOMEM; 504 return -ENOMEM;
505 } 505 }
506 506
507 ctx->aio_ring_file = file; 507 ctx->aio_ring_file = file;
508 nr_events = (PAGE_SIZE * nr_pages - si 508 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
509 / sizeof(struct io_eve 509 / sizeof(struct io_event);
510 510
511 ctx->ring_folios = ctx->internal_folio 511 ctx->ring_folios = ctx->internal_folios;
512 if (nr_pages > AIO_RING_PAGES) { 512 if (nr_pages > AIO_RING_PAGES) {
513 ctx->ring_folios = kcalloc(nr_ 513 ctx->ring_folios = kcalloc(nr_pages, sizeof(struct folio *),
514 GFP 514 GFP_KERNEL);
515 if (!ctx->ring_folios) { 515 if (!ctx->ring_folios) {
516 put_aio_ring_file(ctx) 516 put_aio_ring_file(ctx);
517 return -ENOMEM; 517 return -ENOMEM;
518 } 518 }
519 } 519 }
520 520
521 for (i = 0; i < nr_pages; i++) { 521 for (i = 0; i < nr_pages; i++) {
522 struct folio *folio; 522 struct folio *folio;
523 523
524 folio = __filemap_get_folio(fi 524 folio = __filemap_get_folio(file->f_mapping, i,
525 FG 525 FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
526 GF 526 GFP_USER | __GFP_ZERO);
527 if (IS_ERR(folio)) 527 if (IS_ERR(folio))
528 break; 528 break;
529 529
530 pr_debug("pid(%d) [%d] folio-> 530 pr_debug("pid(%d) [%d] folio->count=%d\n", current->pid, i,
531 folio_ref_count(folio 531 folio_ref_count(folio));
532 folio_end_read(folio, true); 532 folio_end_read(folio, true);
533 533
534 ctx->ring_folios[i] = folio; 534 ctx->ring_folios[i] = folio;
535 } 535 }
536 ctx->nr_pages = i; 536 ctx->nr_pages = i;
537 537
538 if (unlikely(i != nr_pages)) { 538 if (unlikely(i != nr_pages)) {
539 aio_free_ring(ctx); 539 aio_free_ring(ctx);
540 return -ENOMEM; 540 return -ENOMEM;
541 } 541 }
542 542
543 ctx->mmap_size = nr_pages * PAGE_SIZE; 543 ctx->mmap_size = nr_pages * PAGE_SIZE;
544 pr_debug("attempting mmap of %lu bytes 544 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
545 545
546 if (mmap_write_lock_killable(mm)) { 546 if (mmap_write_lock_killable(mm)) {
547 ctx->mmap_size = 0; 547 ctx->mmap_size = 0;
548 aio_free_ring(ctx); 548 aio_free_ring(ctx);
549 return -EINTR; 549 return -EINTR;
550 } 550 }
551 551
552 ctx->mmap_base = do_mmap(ctx->aio_ring 552 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
553 PROT_READ | P 553 PROT_READ | PROT_WRITE,
554 MAP_SHARED, 0 554 MAP_SHARED, 0, 0, &unused, NULL);
555 mmap_write_unlock(mm); 555 mmap_write_unlock(mm);
556 if (IS_ERR((void *)ctx->mmap_base)) { 556 if (IS_ERR((void *)ctx->mmap_base)) {
557 ctx->mmap_size = 0; 557 ctx->mmap_size = 0;
558 aio_free_ring(ctx); 558 aio_free_ring(ctx);
559 return -ENOMEM; 559 return -ENOMEM;
560 } 560 }
561 561
562 pr_debug("mmap address: 0x%08lx\n", ct 562 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
563 563
564 ctx->user_id = ctx->mmap_base; 564 ctx->user_id = ctx->mmap_base;
565 ctx->nr_events = nr_events; /* trusted 565 ctx->nr_events = nr_events; /* trusted copy */
566 566
567 ring = folio_address(ctx->ring_folios[ 567 ring = folio_address(ctx->ring_folios[0]);
568 ring->nr = nr_events; /* user copy * 568 ring->nr = nr_events; /* user copy */
569 ring->id = ~0U; 569 ring->id = ~0U;
570 ring->head = ring->tail = 0; 570 ring->head = ring->tail = 0;
571 ring->magic = AIO_RING_MAGIC; 571 ring->magic = AIO_RING_MAGIC;
572 ring->compat_features = AIO_RING_COMPA 572 ring->compat_features = AIO_RING_COMPAT_FEATURES;
573 ring->incompat_features = AIO_RING_INC 573 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
574 ring->header_length = sizeof(struct ai 574 ring->header_length = sizeof(struct aio_ring);
575 flush_dcache_folio(ctx->ring_folios[0] 575 flush_dcache_folio(ctx->ring_folios[0]);
576 576
577 return 0; 577 return 0;
578 } 578 }
579 579
580 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / s 580 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
581 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - 581 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
582 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PE 582 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
583 583
584 void kiocb_set_cancel_fn(struct kiocb *iocb, k 584 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
585 { 585 {
586 struct aio_kiocb *req; 586 struct aio_kiocb *req;
587 struct kioctx *ctx; 587 struct kioctx *ctx;
588 unsigned long flags; 588 unsigned long flags;
589 589
590 /* 590 /*
591 * kiocb didn't come from aio or is ne 591 * kiocb didn't come from aio or is neither a read nor a write, hence
592 * ignore it. 592 * ignore it.
593 */ 593 */
594 if (!(iocb->ki_flags & IOCB_AIO_RW)) 594 if (!(iocb->ki_flags & IOCB_AIO_RW))
595 return; 595 return;
596 596
597 req = container_of(iocb, struct aio_ki 597 req = container_of(iocb, struct aio_kiocb, rw);
598 598
599 if (WARN_ON_ONCE(!list_empty(&req->ki_ 599 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
600 return; 600 return;
601 601
602 ctx = req->ki_ctx; 602 ctx = req->ki_ctx;
603 603
604 spin_lock_irqsave(&ctx->ctx_lock, flag 604 spin_lock_irqsave(&ctx->ctx_lock, flags);
605 list_add_tail(&req->ki_list, &ctx->act 605 list_add_tail(&req->ki_list, &ctx->active_reqs);
606 req->ki_cancel = cancel; 606 req->ki_cancel = cancel;
607 spin_unlock_irqrestore(&ctx->ctx_lock, 607 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
608 } 608 }
609 EXPORT_SYMBOL(kiocb_set_cancel_fn); 609 EXPORT_SYMBOL(kiocb_set_cancel_fn);
610 610
611 /* 611 /*
612 * free_ioctx() should be RCU delayed to synch 612 * free_ioctx() should be RCU delayed to synchronize against the RCU
613 * protected lookup_ioctx() and also needs pro 613 * protected lookup_ioctx() and also needs process context to call
614 * aio_free_ring(). Use rcu_work. 614 * aio_free_ring(). Use rcu_work.
615 */ 615 */
616 static void free_ioctx(struct work_struct *wor 616 static void free_ioctx(struct work_struct *work)
617 { 617 {
618 struct kioctx *ctx = container_of(to_r 618 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
619 free 619 free_rwork);
620 pr_debug("freeing %p\n", ctx); 620 pr_debug("freeing %p\n", ctx);
621 621
622 aio_free_ring(ctx); 622 aio_free_ring(ctx);
623 free_percpu(ctx->cpu); 623 free_percpu(ctx->cpu);
624 percpu_ref_exit(&ctx->reqs); 624 percpu_ref_exit(&ctx->reqs);
625 percpu_ref_exit(&ctx->users); 625 percpu_ref_exit(&ctx->users);
626 kmem_cache_free(kioctx_cachep, ctx); 626 kmem_cache_free(kioctx_cachep, ctx);
627 } 627 }
628 628
629 static void free_ioctx_reqs(struct percpu_ref 629 static void free_ioctx_reqs(struct percpu_ref *ref)
630 { 630 {
631 struct kioctx *ctx = container_of(ref, 631 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
632 632
633 /* At this point we know that there ar 633 /* At this point we know that there are no any in-flight requests */
634 if (ctx->rq_wait && atomic_dec_and_tes 634 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
635 complete(&ctx->rq_wait->comp); 635 complete(&ctx->rq_wait->comp);
636 636
637 /* Synchronize against RCU protected t 637 /* Synchronize against RCU protected table->table[] dereferences */
638 INIT_RCU_WORK(&ctx->free_rwork, free_i 638 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
639 queue_rcu_work(system_wq, &ctx->free_r 639 queue_rcu_work(system_wq, &ctx->free_rwork);
640 } 640 }
641 641
642 /* 642 /*
643 * When this function runs, the kioctx has bee 643 * When this function runs, the kioctx has been removed from the "hash table"
644 * and ctx->users has dropped to 0, so we know 644 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
645 * now it's safe to cancel any that need to be 645 * now it's safe to cancel any that need to be.
646 */ 646 */
647 static void free_ioctx_users(struct percpu_ref 647 static void free_ioctx_users(struct percpu_ref *ref)
648 { 648 {
649 struct kioctx *ctx = container_of(ref, 649 struct kioctx *ctx = container_of(ref, struct kioctx, users);
650 struct aio_kiocb *req; 650 struct aio_kiocb *req;
651 651
652 spin_lock_irq(&ctx->ctx_lock); 652 spin_lock_irq(&ctx->ctx_lock);
653 653
654 while (!list_empty(&ctx->active_reqs)) 654 while (!list_empty(&ctx->active_reqs)) {
655 req = list_first_entry(&ctx->a 655 req = list_first_entry(&ctx->active_reqs,
656 struct 656 struct aio_kiocb, ki_list);
657 req->ki_cancel(&req->rw); 657 req->ki_cancel(&req->rw);
658 list_del_init(&req->ki_list); 658 list_del_init(&req->ki_list);
659 } 659 }
660 660
661 spin_unlock_irq(&ctx->ctx_lock); 661 spin_unlock_irq(&ctx->ctx_lock);
662 662
663 percpu_ref_kill(&ctx->reqs); 663 percpu_ref_kill(&ctx->reqs);
664 percpu_ref_put(&ctx->reqs); 664 percpu_ref_put(&ctx->reqs);
665 } 665 }
666 666
667 static int ioctx_add_table(struct kioctx *ctx, 667 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
668 { 668 {
669 unsigned i, new_nr; 669 unsigned i, new_nr;
670 struct kioctx_table *table, *old; 670 struct kioctx_table *table, *old;
671 struct aio_ring *ring; 671 struct aio_ring *ring;
672 672
673 spin_lock(&mm->ioctx_lock); 673 spin_lock(&mm->ioctx_lock);
674 table = rcu_dereference_raw(mm->ioctx_ 674 table = rcu_dereference_raw(mm->ioctx_table);
675 675
676 while (1) { 676 while (1) {
677 if (table) 677 if (table)
678 for (i = 0; i < table- 678 for (i = 0; i < table->nr; i++)
679 if (!rcu_acces 679 if (!rcu_access_pointer(table->table[i])) {
680 ctx->i 680 ctx->id = i;
681 rcu_as 681 rcu_assign_pointer(table->table[i], ctx);
682 spin_u 682 spin_unlock(&mm->ioctx_lock);
683 683
684 /* Whi 684 /* While kioctx setup is in progress,
685 * we 685 * we are protected from page migration
686 * cha 686 * changes ring_folios by ->ring_lock.
687 */ 687 */
688 ring = 688 ring = folio_address(ctx->ring_folios[0]);
689 ring-> 689 ring->id = ctx->id;
690 return 690 return 0;
691 } 691 }
692 692
693 new_nr = (table ? table->nr : 693 new_nr = (table ? table->nr : 1) * 4;
694 spin_unlock(&mm->ioctx_lock); 694 spin_unlock(&mm->ioctx_lock);
695 695
696 table = kzalloc(struct_size(ta 696 table = kzalloc(struct_size(table, table, new_nr), GFP_KERNEL);
697 if (!table) 697 if (!table)
698 return -ENOMEM; 698 return -ENOMEM;
699 699
700 table->nr = new_nr; 700 table->nr = new_nr;
701 701
702 spin_lock(&mm->ioctx_lock); 702 spin_lock(&mm->ioctx_lock);
703 old = rcu_dereference_raw(mm-> 703 old = rcu_dereference_raw(mm->ioctx_table);
704 704
705 if (!old) { 705 if (!old) {
706 rcu_assign_pointer(mm- 706 rcu_assign_pointer(mm->ioctx_table, table);
707 } else if (table->nr > old->nr 707 } else if (table->nr > old->nr) {
708 memcpy(table->table, o 708 memcpy(table->table, old->table,
709 old->nr * sizeo 709 old->nr * sizeof(struct kioctx *));
710 710
711 rcu_assign_pointer(mm- 711 rcu_assign_pointer(mm->ioctx_table, table);
712 kfree_rcu(old, rcu); 712 kfree_rcu(old, rcu);
713 } else { 713 } else {
714 kfree(table); 714 kfree(table);
715 table = old; 715 table = old;
716 } 716 }
717 } 717 }
718 } 718 }
719 719
720 static void aio_nr_sub(unsigned nr) 720 static void aio_nr_sub(unsigned nr)
721 { 721 {
722 spin_lock(&aio_nr_lock); 722 spin_lock(&aio_nr_lock);
723 if (WARN_ON(aio_nr - nr > aio_nr)) 723 if (WARN_ON(aio_nr - nr > aio_nr))
724 aio_nr = 0; 724 aio_nr = 0;
725 else 725 else
726 aio_nr -= nr; 726 aio_nr -= nr;
727 spin_unlock(&aio_nr_lock); 727 spin_unlock(&aio_nr_lock);
728 } 728 }
729 729
730 /* ioctx_alloc 730 /* ioctx_alloc
731 * Allocates and initializes an ioctx. R 731 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
732 */ 732 */
733 static struct kioctx *ioctx_alloc(unsigned nr_ 733 static struct kioctx *ioctx_alloc(unsigned nr_events)
734 { 734 {
735 struct mm_struct *mm = current->mm; 735 struct mm_struct *mm = current->mm;
736 struct kioctx *ctx; 736 struct kioctx *ctx;
737 int err = -ENOMEM; 737 int err = -ENOMEM;
738 738
739 /* 739 /*
740 * Store the original nr_events -- wha 740 * Store the original nr_events -- what userspace passed to io_setup(),
741 * for counting against the global lim 741 * for counting against the global limit -- before it changes.
742 */ 742 */
743 unsigned int max_reqs = nr_events; 743 unsigned int max_reqs = nr_events;
744 744
745 /* 745 /*
746 * We keep track of the number of avai 746 * We keep track of the number of available ringbuffer slots, to prevent
747 * overflow (reqs_available), and we a 747 * overflow (reqs_available), and we also use percpu counters for this.
748 * 748 *
749 * So since up to half the slots might 749 * So since up to half the slots might be on other cpu's percpu counters
750 * and unavailable, double nr_events s 750 * and unavailable, double nr_events so userspace sees what they
751 * expected: additionally, we move req 751 * expected: additionally, we move req_batch slots to/from percpu
752 * counters at a time, so make sure th 752 * counters at a time, so make sure that isn't 0:
753 */ 753 */
754 nr_events = max(nr_events, num_possibl 754 nr_events = max(nr_events, num_possible_cpus() * 4);
755 nr_events *= 2; 755 nr_events *= 2;
756 756
757 /* Prevent overflows */ 757 /* Prevent overflows */
758 if (nr_events > (0x10000000U / sizeof( 758 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
759 pr_debug("ENOMEM: nr_events to 759 pr_debug("ENOMEM: nr_events too high\n");
760 return ERR_PTR(-EINVAL); 760 return ERR_PTR(-EINVAL);
761 } 761 }
762 762
763 if (!nr_events || (unsigned long)max_r 763 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
764 return ERR_PTR(-EAGAIN); 764 return ERR_PTR(-EAGAIN);
765 765
766 ctx = kmem_cache_zalloc(kioctx_cachep, 766 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
767 if (!ctx) 767 if (!ctx)
768 return ERR_PTR(-ENOMEM); 768 return ERR_PTR(-ENOMEM);
769 769
770 ctx->max_reqs = max_reqs; 770 ctx->max_reqs = max_reqs;
771 771
772 spin_lock_init(&ctx->ctx_lock); 772 spin_lock_init(&ctx->ctx_lock);
773 spin_lock_init(&ctx->completion_lock); 773 spin_lock_init(&ctx->completion_lock);
774 mutex_init(&ctx->ring_lock); 774 mutex_init(&ctx->ring_lock);
775 /* Protect against page migration thro 775 /* Protect against page migration throughout kiotx setup by keeping
776 * the ring_lock mutex held until setu 776 * the ring_lock mutex held until setup is complete. */
777 mutex_lock(&ctx->ring_lock); 777 mutex_lock(&ctx->ring_lock);
778 init_waitqueue_head(&ctx->wait); 778 init_waitqueue_head(&ctx->wait);
779 779
780 INIT_LIST_HEAD(&ctx->active_reqs); 780 INIT_LIST_HEAD(&ctx->active_reqs);
781 781
782 if (percpu_ref_init(&ctx->users, free_ 782 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
783 goto err; 783 goto err;
784 784
785 if (percpu_ref_init(&ctx->reqs, free_i 785 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
786 goto err; 786 goto err;
787 787
788 ctx->cpu = alloc_percpu(struct kioctx_ 788 ctx->cpu = alloc_percpu(struct kioctx_cpu);
789 if (!ctx->cpu) 789 if (!ctx->cpu)
790 goto err; 790 goto err;
791 791
792 err = aio_setup_ring(ctx, nr_events); 792 err = aio_setup_ring(ctx, nr_events);
793 if (err < 0) 793 if (err < 0)
794 goto err; 794 goto err;
795 795
796 atomic_set(&ctx->reqs_available, ctx-> 796 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
797 ctx->req_batch = (ctx->nr_events - 1) 797 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
798 if (ctx->req_batch < 1) 798 if (ctx->req_batch < 1)
799 ctx->req_batch = 1; 799 ctx->req_batch = 1;
800 800
801 /* limit the number of system wide aio 801 /* limit the number of system wide aios */
802 spin_lock(&aio_nr_lock); 802 spin_lock(&aio_nr_lock);
803 if (aio_nr + ctx->max_reqs > aio_max_n 803 if (aio_nr + ctx->max_reqs > aio_max_nr ||
804 aio_nr + ctx->max_reqs < aio_nr) { 804 aio_nr + ctx->max_reqs < aio_nr) {
805 spin_unlock(&aio_nr_lock); 805 spin_unlock(&aio_nr_lock);
806 err = -EAGAIN; 806 err = -EAGAIN;
807 goto err_ctx; 807 goto err_ctx;
808 } 808 }
809 aio_nr += ctx->max_reqs; 809 aio_nr += ctx->max_reqs;
810 spin_unlock(&aio_nr_lock); 810 spin_unlock(&aio_nr_lock);
811 811
812 percpu_ref_get(&ctx->users); /* io_ 812 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
813 percpu_ref_get(&ctx->reqs); /* fre 813 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
814 814
815 err = ioctx_add_table(ctx, mm); 815 err = ioctx_add_table(ctx, mm);
816 if (err) 816 if (err)
817 goto err_cleanup; 817 goto err_cleanup;
818 818
819 /* Release the ring_lock mutex now tha 819 /* Release the ring_lock mutex now that all setup is complete. */
820 mutex_unlock(&ctx->ring_lock); 820 mutex_unlock(&ctx->ring_lock);
821 821
822 pr_debug("allocated ioctx %p[%ld]: mm= 822 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
823 ctx, ctx->user_id, mm, ctx->n 823 ctx, ctx->user_id, mm, ctx->nr_events);
824 return ctx; 824 return ctx;
825 825
826 err_cleanup: 826 err_cleanup:
827 aio_nr_sub(ctx->max_reqs); 827 aio_nr_sub(ctx->max_reqs);
828 err_ctx: 828 err_ctx:
829 atomic_set(&ctx->dead, 1); 829 atomic_set(&ctx->dead, 1);
830 if (ctx->mmap_size) 830 if (ctx->mmap_size)
831 vm_munmap(ctx->mmap_base, ctx- 831 vm_munmap(ctx->mmap_base, ctx->mmap_size);
832 aio_free_ring(ctx); 832 aio_free_ring(ctx);
833 err: 833 err:
834 mutex_unlock(&ctx->ring_lock); 834 mutex_unlock(&ctx->ring_lock);
835 free_percpu(ctx->cpu); 835 free_percpu(ctx->cpu);
836 percpu_ref_exit(&ctx->reqs); 836 percpu_ref_exit(&ctx->reqs);
837 percpu_ref_exit(&ctx->users); 837 percpu_ref_exit(&ctx->users);
838 kmem_cache_free(kioctx_cachep, ctx); 838 kmem_cache_free(kioctx_cachep, ctx);
839 pr_debug("error allocating ioctx %d\n" 839 pr_debug("error allocating ioctx %d\n", err);
840 return ERR_PTR(err); 840 return ERR_PTR(err);
841 } 841 }
842 842
843 /* kill_ioctx 843 /* kill_ioctx
844 * Cancels all outstanding aio requests o 844 * Cancels all outstanding aio requests on an aio context. Used
845 * when the processes owning a context ha 845 * when the processes owning a context have all exited to encourage
846 * the rapid destruction of the kioctx. 846 * the rapid destruction of the kioctx.
847 */ 847 */
848 static int kill_ioctx(struct mm_struct *mm, st 848 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
849 struct ctx_rq_wait *wait 849 struct ctx_rq_wait *wait)
850 { 850 {
851 struct kioctx_table *table; 851 struct kioctx_table *table;
852 852
853 spin_lock(&mm->ioctx_lock); 853 spin_lock(&mm->ioctx_lock);
854 if (atomic_xchg(&ctx->dead, 1)) { 854 if (atomic_xchg(&ctx->dead, 1)) {
855 spin_unlock(&mm->ioctx_lock); 855 spin_unlock(&mm->ioctx_lock);
856 return -EINVAL; 856 return -EINVAL;
857 } 857 }
858 858
859 table = rcu_dereference_raw(mm->ioctx_ 859 table = rcu_dereference_raw(mm->ioctx_table);
860 WARN_ON(ctx != rcu_access_pointer(tabl 860 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
861 RCU_INIT_POINTER(table->table[ctx->id] 861 RCU_INIT_POINTER(table->table[ctx->id], NULL);
862 spin_unlock(&mm->ioctx_lock); 862 spin_unlock(&mm->ioctx_lock);
863 863
864 /* free_ioctx_reqs() will do the neces 864 /* free_ioctx_reqs() will do the necessary RCU synchronization */
865 wake_up_all(&ctx->wait); 865 wake_up_all(&ctx->wait);
866 866
867 /* 867 /*
868 * It'd be more correct to do this in 868 * It'd be more correct to do this in free_ioctx(), after all
869 * the outstanding kiocbs have finishe 869 * the outstanding kiocbs have finished - but by then io_destroy
870 * has already returned, so io_setup() 870 * has already returned, so io_setup() could potentially return
871 * -EAGAIN with no ioctxs actually in 871 * -EAGAIN with no ioctxs actually in use (as far as userspace
872 * could tell). 872 * could tell).
873 */ 873 */
874 aio_nr_sub(ctx->max_reqs); 874 aio_nr_sub(ctx->max_reqs);
875 875
876 if (ctx->mmap_size) 876 if (ctx->mmap_size)
877 vm_munmap(ctx->mmap_base, ctx- 877 vm_munmap(ctx->mmap_base, ctx->mmap_size);
878 878
879 ctx->rq_wait = wait; 879 ctx->rq_wait = wait;
880 percpu_ref_kill(&ctx->users); 880 percpu_ref_kill(&ctx->users);
881 return 0; 881 return 0;
882 } 882 }
883 883
884 /* 884 /*
885 * exit_aio: called when the last user of mm g 885 * exit_aio: called when the last user of mm goes away. At this point, there is
886 * no way for any new requests to be submited 886 * no way for any new requests to be submited or any of the io_* syscalls to be
887 * called on the context. 887 * called on the context.
888 * 888 *
889 * There may be outstanding kiocbs, but free_i 889 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
890 * them. 890 * them.
891 */ 891 */
892 void exit_aio(struct mm_struct *mm) 892 void exit_aio(struct mm_struct *mm)
893 { 893 {
894 struct kioctx_table *table = rcu_deref 894 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
895 struct ctx_rq_wait wait; 895 struct ctx_rq_wait wait;
896 int i, skipped; 896 int i, skipped;
897 897
898 if (!table) 898 if (!table)
899 return; 899 return;
900 900
901 atomic_set(&wait.count, table->nr); 901 atomic_set(&wait.count, table->nr);
902 init_completion(&wait.comp); 902 init_completion(&wait.comp);
903 903
904 skipped = 0; 904 skipped = 0;
905 for (i = 0; i < table->nr; ++i) { 905 for (i = 0; i < table->nr; ++i) {
906 struct kioctx *ctx = 906 struct kioctx *ctx =
907 rcu_dereference_protec 907 rcu_dereference_protected(table->table[i], true);
908 908
909 if (!ctx) { 909 if (!ctx) {
910 skipped++; 910 skipped++;
911 continue; 911 continue;
912 } 912 }
913 913
914 /* 914 /*
915 * We don't need to bother wit 915 * We don't need to bother with munmap() here - exit_mmap(mm)
916 * is coming and it'll unmap e 916 * is coming and it'll unmap everything. And we simply can't,
917 * this is not necessarily our 917 * this is not necessarily our ->mm.
918 * Since kill_ioctx() uses non 918 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
919 * that it needs to unmap the 919 * that it needs to unmap the area, just set it to 0.
920 */ 920 */
921 ctx->mmap_size = 0; 921 ctx->mmap_size = 0;
922 kill_ioctx(mm, ctx, &wait); 922 kill_ioctx(mm, ctx, &wait);
923 } 923 }
924 924
925 if (!atomic_sub_and_test(skipped, &wai 925 if (!atomic_sub_and_test(skipped, &wait.count)) {
926 /* Wait until all IO for the c 926 /* Wait until all IO for the context are done. */
927 wait_for_completion(&wait.comp 927 wait_for_completion(&wait.comp);
928 } 928 }
929 929
930 RCU_INIT_POINTER(mm->ioctx_table, NULL 930 RCU_INIT_POINTER(mm->ioctx_table, NULL);
931 kfree(table); 931 kfree(table);
932 } 932 }
933 933
934 static void put_reqs_available(struct kioctx * 934 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
935 { 935 {
936 struct kioctx_cpu *kcpu; 936 struct kioctx_cpu *kcpu;
937 unsigned long flags; 937 unsigned long flags;
938 938
939 local_irq_save(flags); 939 local_irq_save(flags);
940 kcpu = this_cpu_ptr(ctx->cpu); 940 kcpu = this_cpu_ptr(ctx->cpu);
941 kcpu->reqs_available += nr; 941 kcpu->reqs_available += nr;
942 942
943 while (kcpu->reqs_available >= ctx->re 943 while (kcpu->reqs_available >= ctx->req_batch * 2) {
944 kcpu->reqs_available -= ctx->r 944 kcpu->reqs_available -= ctx->req_batch;
945 atomic_add(ctx->req_batch, &ct 945 atomic_add(ctx->req_batch, &ctx->reqs_available);
946 } 946 }
947 947
948 local_irq_restore(flags); 948 local_irq_restore(flags);
949 } 949 }
950 950
951 static bool __get_reqs_available(struct kioctx 951 static bool __get_reqs_available(struct kioctx *ctx)
952 { 952 {
953 struct kioctx_cpu *kcpu; 953 struct kioctx_cpu *kcpu;
954 bool ret = false; 954 bool ret = false;
955 unsigned long flags; 955 unsigned long flags;
956 956
957 local_irq_save(flags); 957 local_irq_save(flags);
958 kcpu = this_cpu_ptr(ctx->cpu); 958 kcpu = this_cpu_ptr(ctx->cpu);
959 if (!kcpu->reqs_available) { 959 if (!kcpu->reqs_available) {
960 int avail = atomic_read(&ctx-> 960 int avail = atomic_read(&ctx->reqs_available);
961 961
962 do { 962 do {
963 if (avail < ctx->req_b 963 if (avail < ctx->req_batch)
964 goto out; 964 goto out;
965 } while (!atomic_try_cmpxchg(& 965 } while (!atomic_try_cmpxchg(&ctx->reqs_available,
966 & 966 &avail, avail - ctx->req_batch));
967 967
968 kcpu->reqs_available += ctx->r 968 kcpu->reqs_available += ctx->req_batch;
969 } 969 }
970 970
971 ret = true; 971 ret = true;
972 kcpu->reqs_available--; 972 kcpu->reqs_available--;
973 out: 973 out:
974 local_irq_restore(flags); 974 local_irq_restore(flags);
975 return ret; 975 return ret;
976 } 976 }
977 977
978 /* refill_reqs_available 978 /* refill_reqs_available
979 * Updates the reqs_available reference c 979 * Updates the reqs_available reference counts used for tracking the
980 * number of free slots in the completion 980 * number of free slots in the completion ring. This can be called
981 * from aio_complete() (to optimistically 981 * from aio_complete() (to optimistically update reqs_available) or
982 * from aio_get_req() (the we're out of e 982 * from aio_get_req() (the we're out of events case). It must be
983 * called holding ctx->completion_lock. 983 * called holding ctx->completion_lock.
984 */ 984 */
985 static void refill_reqs_available(struct kioct 985 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
986 unsigned tai 986 unsigned tail)
987 { 987 {
988 unsigned events_in_ring, completed; 988 unsigned events_in_ring, completed;
989 989
990 /* Clamp head since userland can write 990 /* Clamp head since userland can write to it. */
991 head %= ctx->nr_events; 991 head %= ctx->nr_events;
992 if (head <= tail) 992 if (head <= tail)
993 events_in_ring = tail - head; 993 events_in_ring = tail - head;
994 else 994 else
995 events_in_ring = ctx->nr_event 995 events_in_ring = ctx->nr_events - (head - tail);
996 996
997 completed = ctx->completed_events; 997 completed = ctx->completed_events;
998 if (events_in_ring < completed) 998 if (events_in_ring < completed)
999 completed -= events_in_ring; 999 completed -= events_in_ring;
1000 else 1000 else
1001 completed = 0; 1001 completed = 0;
1002 1002
1003 if (!completed) 1003 if (!completed)
1004 return; 1004 return;
1005 1005
1006 ctx->completed_events -= completed; 1006 ctx->completed_events -= completed;
1007 put_reqs_available(ctx, completed); 1007 put_reqs_available(ctx, completed);
1008 } 1008 }
1009 1009
1010 /* user_refill_reqs_available 1010 /* user_refill_reqs_available
1011 * Called to refill reqs_available when 1011 * Called to refill reqs_available when aio_get_req() encounters an
1012 * out of space in the completion ring. 1012 * out of space in the completion ring.
1013 */ 1013 */
1014 static void user_refill_reqs_available(struct 1014 static void user_refill_reqs_available(struct kioctx *ctx)
1015 { 1015 {
1016 spin_lock_irq(&ctx->completion_lock); 1016 spin_lock_irq(&ctx->completion_lock);
1017 if (ctx->completed_events) { 1017 if (ctx->completed_events) {
1018 struct aio_ring *ring; 1018 struct aio_ring *ring;
1019 unsigned head; 1019 unsigned head;
1020 1020
1021 /* Access of ring->head may r 1021 /* Access of ring->head may race with aio_read_events_ring()
1022 * here, but that's okay sinc 1022 * here, but that's okay since whether we read the old version
1023 * or the new version, and ei 1023 * or the new version, and either will be valid. The important
1024 * part is that head cannot p 1024 * part is that head cannot pass tail since we prevent
1025 * aio_complete() from updati 1025 * aio_complete() from updating tail by holding
1026 * ctx->completion_lock. Eve 1026 * ctx->completion_lock. Even if head is invalid, the check
1027 * against ctx->completed_eve 1027 * against ctx->completed_events below will make sure we do the
1028 * safe/right thing. 1028 * safe/right thing.
1029 */ 1029 */
1030 ring = folio_address(ctx->rin 1030 ring = folio_address(ctx->ring_folios[0]);
1031 head = ring->head; 1031 head = ring->head;
1032 1032
1033 refill_reqs_available(ctx, he 1033 refill_reqs_available(ctx, head, ctx->tail);
1034 } 1034 }
1035 1035
1036 spin_unlock_irq(&ctx->completion_lock 1036 spin_unlock_irq(&ctx->completion_lock);
1037 } 1037 }
1038 1038
1039 static bool get_reqs_available(struct kioctx 1039 static bool get_reqs_available(struct kioctx *ctx)
1040 { 1040 {
1041 if (__get_reqs_available(ctx)) 1041 if (__get_reqs_available(ctx))
1042 return true; 1042 return true;
1043 user_refill_reqs_available(ctx); 1043 user_refill_reqs_available(ctx);
1044 return __get_reqs_available(ctx); 1044 return __get_reqs_available(ctx);
1045 } 1045 }
1046 1046
1047 /* aio_get_req 1047 /* aio_get_req
1048 * Allocate a slot for an aio request. 1048 * Allocate a slot for an aio request.
1049 * Returns NULL if no requests are free. 1049 * Returns NULL if no requests are free.
1050 * 1050 *
1051 * The refcount is initialized to 2 - one for 1051 * The refcount is initialized to 2 - one for the async op completion,
1052 * one for the synchronous code that does thi 1052 * one for the synchronous code that does this.
1053 */ 1053 */
1054 static inline struct aio_kiocb *aio_get_req(s 1054 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1055 { 1055 {
1056 struct aio_kiocb *req; 1056 struct aio_kiocb *req;
1057 1057
1058 req = kmem_cache_alloc(kiocb_cachep, 1058 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1059 if (unlikely(!req)) 1059 if (unlikely(!req))
1060 return NULL; 1060 return NULL;
1061 1061
1062 if (unlikely(!get_reqs_available(ctx) 1062 if (unlikely(!get_reqs_available(ctx))) {
1063 kmem_cache_free(kiocb_cachep, 1063 kmem_cache_free(kiocb_cachep, req);
1064 return NULL; 1064 return NULL;
1065 } 1065 }
1066 1066
1067 percpu_ref_get(&ctx->reqs); 1067 percpu_ref_get(&ctx->reqs);
1068 req->ki_ctx = ctx; 1068 req->ki_ctx = ctx;
1069 INIT_LIST_HEAD(&req->ki_list); 1069 INIT_LIST_HEAD(&req->ki_list);
1070 refcount_set(&req->ki_refcnt, 2); 1070 refcount_set(&req->ki_refcnt, 2);
1071 req->ki_eventfd = NULL; 1071 req->ki_eventfd = NULL;
1072 return req; 1072 return req;
1073 } 1073 }
1074 1074
1075 static struct kioctx *lookup_ioctx(unsigned l 1075 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1076 { 1076 {
1077 struct aio_ring __user *ring = (void 1077 struct aio_ring __user *ring = (void __user *)ctx_id;
1078 struct mm_struct *mm = current->mm; 1078 struct mm_struct *mm = current->mm;
1079 struct kioctx *ctx, *ret = NULL; 1079 struct kioctx *ctx, *ret = NULL;
1080 struct kioctx_table *table; 1080 struct kioctx_table *table;
1081 unsigned id; 1081 unsigned id;
1082 1082
1083 if (get_user(id, &ring->id)) 1083 if (get_user(id, &ring->id))
1084 return NULL; 1084 return NULL;
1085 1085
1086 rcu_read_lock(); 1086 rcu_read_lock();
1087 table = rcu_dereference(mm->ioctx_tab 1087 table = rcu_dereference(mm->ioctx_table);
1088 1088
1089 if (!table || id >= table->nr) 1089 if (!table || id >= table->nr)
1090 goto out; 1090 goto out;
1091 1091
1092 id = array_index_nospec(id, table->nr 1092 id = array_index_nospec(id, table->nr);
1093 ctx = rcu_dereference(table->table[id 1093 ctx = rcu_dereference(table->table[id]);
1094 if (ctx && ctx->user_id == ctx_id) { 1094 if (ctx && ctx->user_id == ctx_id) {
1095 if (percpu_ref_tryget_live(&c 1095 if (percpu_ref_tryget_live(&ctx->users))
1096 ret = ctx; 1096 ret = ctx;
1097 } 1097 }
1098 out: 1098 out:
1099 rcu_read_unlock(); 1099 rcu_read_unlock();
1100 return ret; 1100 return ret;
1101 } 1101 }
1102 1102
1103 static inline void iocb_destroy(struct aio_ki 1103 static inline void iocb_destroy(struct aio_kiocb *iocb)
1104 { 1104 {
1105 if (iocb->ki_eventfd) 1105 if (iocb->ki_eventfd)
1106 eventfd_ctx_put(iocb->ki_even 1106 eventfd_ctx_put(iocb->ki_eventfd);
1107 if (iocb->ki_filp) 1107 if (iocb->ki_filp)
1108 fput(iocb->ki_filp); 1108 fput(iocb->ki_filp);
1109 percpu_ref_put(&iocb->ki_ctx->reqs); 1109 percpu_ref_put(&iocb->ki_ctx->reqs);
1110 kmem_cache_free(kiocb_cachep, iocb); 1110 kmem_cache_free(kiocb_cachep, iocb);
1111 } 1111 }
1112 1112
1113 struct aio_waiter { 1113 struct aio_waiter {
1114 struct wait_queue_entry w; 1114 struct wait_queue_entry w;
1115 size_t min_nr; 1115 size_t min_nr;
1116 }; 1116 };
1117 1117
1118 /* aio_complete 1118 /* aio_complete
1119 * Called when the io request on the giv 1119 * Called when the io request on the given iocb is complete.
1120 */ 1120 */
1121 static void aio_complete(struct aio_kiocb *io 1121 static void aio_complete(struct aio_kiocb *iocb)
1122 { 1122 {
1123 struct kioctx *ctx = iocb->ki_ctx; 1123 struct kioctx *ctx = iocb->ki_ctx;
1124 struct aio_ring *ring; 1124 struct aio_ring *ring;
1125 struct io_event *ev_page, *event; 1125 struct io_event *ev_page, *event;
1126 unsigned tail, pos, head, avail; 1126 unsigned tail, pos, head, avail;
1127 unsigned long flags; 1127 unsigned long flags;
1128 1128
1129 /* 1129 /*
1130 * Add a completion event to the ring 1130 * Add a completion event to the ring buffer. Must be done holding
1131 * ctx->completion_lock to prevent ot 1131 * ctx->completion_lock to prevent other code from messing with the tail
1132 * pointer since we might be called f 1132 * pointer since we might be called from irq context.
1133 */ 1133 */
1134 spin_lock_irqsave(&ctx->completion_lo 1134 spin_lock_irqsave(&ctx->completion_lock, flags);
1135 1135
1136 tail = ctx->tail; 1136 tail = ctx->tail;
1137 pos = tail + AIO_EVENTS_OFFSET; 1137 pos = tail + AIO_EVENTS_OFFSET;
1138 1138
1139 if (++tail >= ctx->nr_events) 1139 if (++tail >= ctx->nr_events)
1140 tail = 0; 1140 tail = 0;
1141 1141
1142 ev_page = folio_address(ctx->ring_fol 1142 ev_page = folio_address(ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE]);
1143 event = ev_page + pos % AIO_EVENTS_PE 1143 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1144 1144
1145 *event = iocb->ki_res; 1145 *event = iocb->ki_res;
1146 1146
1147 flush_dcache_folio(ctx->ring_folios[p 1147 flush_dcache_folio(ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE]);
1148 1148
1149 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\ 1149 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1150 (void __user *)(unsigned lon 1150 (void __user *)(unsigned long)iocb->ki_res.obj,
1151 iocb->ki_res.data, iocb->ki_ 1151 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1152 1152
1153 /* after flagging the request as done 1153 /* after flagging the request as done, we
1154 * must never even look at it again 1154 * must never even look at it again
1155 */ 1155 */
1156 smp_wmb(); /* make event visible 1156 smp_wmb(); /* make event visible before updating tail */
1157 1157
1158 ctx->tail = tail; 1158 ctx->tail = tail;
1159 1159
1160 ring = folio_address(ctx->ring_folios 1160 ring = folio_address(ctx->ring_folios[0]);
1161 head = ring->head; 1161 head = ring->head;
1162 ring->tail = tail; 1162 ring->tail = tail;
1163 flush_dcache_folio(ctx->ring_folios[0 1163 flush_dcache_folio(ctx->ring_folios[0]);
1164 1164
1165 ctx->completed_events++; 1165 ctx->completed_events++;
1166 if (ctx->completed_events > 1) 1166 if (ctx->completed_events > 1)
1167 refill_reqs_available(ctx, he 1167 refill_reqs_available(ctx, head, tail);
1168 1168
1169 avail = tail > head 1169 avail = tail > head
1170 ? tail - head 1170 ? tail - head
1171 : tail + ctx->nr_events - hea 1171 : tail + ctx->nr_events - head;
1172 spin_unlock_irqrestore(&ctx->completi 1172 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1173 1173
1174 pr_debug("added to ring %p at [%u]\n" 1174 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1175 1175
1176 /* 1176 /*
1177 * Check if the user asked us to deli 1177 * Check if the user asked us to deliver the result through an
1178 * eventfd. The eventfd_signal() func 1178 * eventfd. The eventfd_signal() function is safe to be called
1179 * from IRQ context. 1179 * from IRQ context.
1180 */ 1180 */
1181 if (iocb->ki_eventfd) 1181 if (iocb->ki_eventfd)
1182 eventfd_signal(iocb->ki_event 1182 eventfd_signal(iocb->ki_eventfd);
1183 1183
1184 /* 1184 /*
1185 * We have to order our ring_info tai 1185 * We have to order our ring_info tail store above and test
1186 * of the wait list below outside the 1186 * of the wait list below outside the wait lock. This is
1187 * like in wake_up_bit() where cleari 1187 * like in wake_up_bit() where clearing a bit has to be
1188 * ordered with the unlocked test. 1188 * ordered with the unlocked test.
1189 */ 1189 */
1190 smp_mb(); 1190 smp_mb();
1191 1191
1192 if (waitqueue_active(&ctx->wait)) { 1192 if (waitqueue_active(&ctx->wait)) {
1193 struct aio_waiter *curr, *nex 1193 struct aio_waiter *curr, *next;
1194 unsigned long flags; 1194 unsigned long flags;
1195 1195
1196 spin_lock_irqsave(&ctx->wait. 1196 spin_lock_irqsave(&ctx->wait.lock, flags);
1197 list_for_each_entry_safe(curr 1197 list_for_each_entry_safe(curr, next, &ctx->wait.head, w.entry)
1198 if (avail >= curr->mi 1198 if (avail >= curr->min_nr) {
1199 wake_up_proce 1199 wake_up_process(curr->w.private);
1200 list_del_init 1200 list_del_init_careful(&curr->w.entry);
1201 } 1201 }
1202 spin_unlock_irqrestore(&ctx-> 1202 spin_unlock_irqrestore(&ctx->wait.lock, flags);
1203 } 1203 }
1204 } 1204 }
1205 1205
1206 static inline void iocb_put(struct aio_kiocb 1206 static inline void iocb_put(struct aio_kiocb *iocb)
1207 { 1207 {
1208 if (refcount_dec_and_test(&iocb->ki_r 1208 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1209 aio_complete(iocb); 1209 aio_complete(iocb);
1210 iocb_destroy(iocb); 1210 iocb_destroy(iocb);
1211 } 1211 }
1212 } 1212 }
1213 1213
1214 /* aio_read_events_ring 1214 /* aio_read_events_ring
1215 * Pull an event off of the ioctx's even 1215 * Pull an event off of the ioctx's event ring. Returns the number of
1216 * events fetched 1216 * events fetched
1217 */ 1217 */
1218 static long aio_read_events_ring(struct kioct 1218 static long aio_read_events_ring(struct kioctx *ctx,
1219 struct io_ev 1219 struct io_event __user *event, long nr)
1220 { 1220 {
1221 struct aio_ring *ring; 1221 struct aio_ring *ring;
1222 unsigned head, tail, pos; 1222 unsigned head, tail, pos;
1223 long ret = 0; 1223 long ret = 0;
1224 int copy_ret; 1224 int copy_ret;
1225 1225
1226 /* 1226 /*
1227 * The mutex can block and wake us up 1227 * The mutex can block and wake us up and that will cause
1228 * wait_event_interruptible_hrtimeout 1228 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1229 * and repeat. This should be rare en 1229 * and repeat. This should be rare enough that it doesn't cause
1230 * peformance issues. See the comment 1230 * peformance issues. See the comment in read_events() for more detail.
1231 */ 1231 */
1232 sched_annotate_sleep(); 1232 sched_annotate_sleep();
1233 mutex_lock(&ctx->ring_lock); 1233 mutex_lock(&ctx->ring_lock);
1234 1234
1235 /* Access to ->ring_folios here is pr 1235 /* Access to ->ring_folios here is protected by ctx->ring_lock. */
1236 ring = folio_address(ctx->ring_folios 1236 ring = folio_address(ctx->ring_folios[0]);
1237 head = ring->head; 1237 head = ring->head;
1238 tail = ring->tail; 1238 tail = ring->tail;
1239 1239
1240 /* 1240 /*
1241 * Ensure that once we've read the cu 1241 * Ensure that once we've read the current tail pointer, that
1242 * we also see the events that were s 1242 * we also see the events that were stored up to the tail.
1243 */ 1243 */
1244 smp_rmb(); 1244 smp_rmb();
1245 1245
1246 pr_debug("h%u t%u m%u\n", head, tail, 1246 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1247 1247
1248 if (head == tail) 1248 if (head == tail)
1249 goto out; 1249 goto out;
1250 1250
1251 head %= ctx->nr_events; 1251 head %= ctx->nr_events;
1252 tail %= ctx->nr_events; 1252 tail %= ctx->nr_events;
1253 1253
1254 while (ret < nr) { 1254 while (ret < nr) {
1255 long avail; 1255 long avail;
1256 struct io_event *ev; 1256 struct io_event *ev;
1257 struct folio *folio; 1257 struct folio *folio;
1258 1258
1259 avail = (head <= tail ? tail 1259 avail = (head <= tail ? tail : ctx->nr_events) - head;
1260 if (head == tail) 1260 if (head == tail)
1261 break; 1261 break;
1262 1262
1263 pos = head + AIO_EVENTS_OFFSE 1263 pos = head + AIO_EVENTS_OFFSET;
1264 folio = ctx->ring_folios[pos 1264 folio = ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE];
1265 pos %= AIO_EVENTS_PER_PAGE; 1265 pos %= AIO_EVENTS_PER_PAGE;
1266 1266
1267 avail = min(avail, nr - ret); 1267 avail = min(avail, nr - ret);
1268 avail = min_t(long, avail, AI 1268 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1269 1269
1270 ev = folio_address(folio); 1270 ev = folio_address(folio);
1271 copy_ret = copy_to_user(event 1271 copy_ret = copy_to_user(event + ret, ev + pos,
1272 sizeo 1272 sizeof(*ev) * avail);
1273 1273
1274 if (unlikely(copy_ret)) { 1274 if (unlikely(copy_ret)) {
1275 ret = -EFAULT; 1275 ret = -EFAULT;
1276 goto out; 1276 goto out;
1277 } 1277 }
1278 1278
1279 ret += avail; 1279 ret += avail;
1280 head += avail; 1280 head += avail;
1281 head %= ctx->nr_events; 1281 head %= ctx->nr_events;
1282 } 1282 }
1283 1283
1284 ring = folio_address(ctx->ring_folios 1284 ring = folio_address(ctx->ring_folios[0]);
1285 ring->head = head; 1285 ring->head = head;
1286 flush_dcache_folio(ctx->ring_folios[0 1286 flush_dcache_folio(ctx->ring_folios[0]);
1287 1287
1288 pr_debug("%li h%u t%u\n", ret, head, 1288 pr_debug("%li h%u t%u\n", ret, head, tail);
1289 out: 1289 out:
1290 mutex_unlock(&ctx->ring_lock); 1290 mutex_unlock(&ctx->ring_lock);
1291 1291
1292 return ret; 1292 return ret;
1293 } 1293 }
1294 1294
1295 static bool aio_read_events(struct kioctx *ct 1295 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1296 struct io_event _ 1296 struct io_event __user *event, long *i)
1297 { 1297 {
1298 long ret = aio_read_events_ring(ctx, 1298 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1299 1299
1300 if (ret > 0) 1300 if (ret > 0)
1301 *i += ret; 1301 *i += ret;
1302 1302
1303 if (unlikely(atomic_read(&ctx->dead)) 1303 if (unlikely(atomic_read(&ctx->dead)))
1304 ret = -EINVAL; 1304 ret = -EINVAL;
1305 1305
1306 if (!*i) 1306 if (!*i)
1307 *i = ret; 1307 *i = ret;
1308 1308
1309 return ret < 0 || *i >= min_nr; 1309 return ret < 0 || *i >= min_nr;
1310 } 1310 }
1311 1311
1312 static long read_events(struct kioctx *ctx, l 1312 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1313 struct io_event __use 1313 struct io_event __user *event,
1314 ktime_t until) 1314 ktime_t until)
1315 { 1315 {
1316 struct hrtimer_sleeper t; 1316 struct hrtimer_sleeper t;
1317 struct aio_waiter w; 1317 struct aio_waiter w;
1318 long ret = 0, ret2 = 0; 1318 long ret = 0, ret2 = 0;
1319 1319
1320 /* 1320 /*
1321 * Note that aio_read_events() is bei 1321 * Note that aio_read_events() is being called as the conditional - i.e.
1322 * we're calling it after prepare_to_ 1322 * we're calling it after prepare_to_wait() has set task state to
1323 * TASK_INTERRUPTIBLE. 1323 * TASK_INTERRUPTIBLE.
1324 * 1324 *
1325 * But aio_read_events() can block, a 1325 * But aio_read_events() can block, and if it blocks it's going to flip
1326 * the task state back to TASK_RUNNIN 1326 * the task state back to TASK_RUNNING.
1327 * 1327 *
1328 * This should be ok, provided it doe 1328 * This should be ok, provided it doesn't flip the state back to
1329 * TASK_RUNNING and return 0 too much 1329 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1330 * will only happen if the mutex_lock 1330 * will only happen if the mutex_lock() call blocks, and we then find
1331 * the ringbuffer empty. So in practi 1331 * the ringbuffer empty. So in practice we should be ok, but it's
1332 * something to be aware of when touc 1332 * something to be aware of when touching this code.
1333 */ 1333 */
1334 aio_read_events(ctx, min_nr, nr, even 1334 aio_read_events(ctx, min_nr, nr, event, &ret);
1335 if (until == 0 || ret < 0 || ret >= m 1335 if (until == 0 || ret < 0 || ret >= min_nr)
1336 return ret; 1336 return ret;
1337 1337
1338 hrtimer_init_sleeper_on_stack(&t, CLO 1338 hrtimer_init_sleeper_on_stack(&t, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1339 if (until != KTIME_MAX) { 1339 if (until != KTIME_MAX) {
1340 hrtimer_set_expires_range_ns( 1340 hrtimer_set_expires_range_ns(&t.timer, until, current->timer_slack_ns);
1341 hrtimer_sleeper_start_expires 1341 hrtimer_sleeper_start_expires(&t, HRTIMER_MODE_REL);
1342 } 1342 }
1343 1343
1344 init_wait(&w.w); 1344 init_wait(&w.w);
1345 1345
1346 while (1) { 1346 while (1) {
1347 unsigned long nr_got = ret; 1347 unsigned long nr_got = ret;
1348 1348
1349 w.min_nr = min_nr - ret; 1349 w.min_nr = min_nr - ret;
1350 1350
1351 ret2 = prepare_to_wait_event( 1351 ret2 = prepare_to_wait_event(&ctx->wait, &w.w, TASK_INTERRUPTIBLE);
1352 if (!ret2 && !t.task) 1352 if (!ret2 && !t.task)
1353 ret2 = -ETIME; 1353 ret2 = -ETIME;
1354 1354
1355 if (aio_read_events(ctx, min_ 1355 if (aio_read_events(ctx, min_nr, nr, event, &ret) || ret2)
1356 break; 1356 break;
1357 1357
1358 if (nr_got == ret) 1358 if (nr_got == ret)
1359 schedule(); 1359 schedule();
1360 } 1360 }
1361 1361
1362 finish_wait(&ctx->wait, &w.w); 1362 finish_wait(&ctx->wait, &w.w);
1363 hrtimer_cancel(&t.timer); 1363 hrtimer_cancel(&t.timer);
1364 destroy_hrtimer_on_stack(&t.timer); 1364 destroy_hrtimer_on_stack(&t.timer);
1365 1365
1366 return ret; 1366 return ret;
1367 } 1367 }
1368 1368
1369 /* sys_io_setup: 1369 /* sys_io_setup:
1370 * Create an aio_context capable of rece 1370 * Create an aio_context capable of receiving at least nr_events.
1371 * ctxp must not point to an aio_context 1371 * ctxp must not point to an aio_context that already exists, and
1372 * must be initialized to 0 prior to the 1372 * must be initialized to 0 prior to the call. On successful
1373 * creation of the aio_context, *ctxp is 1373 * creation of the aio_context, *ctxp is filled in with the resulting
1374 * handle. May fail with -EINVAL if *ct 1374 * handle. May fail with -EINVAL if *ctxp is not initialized,
1375 * if the specified nr_events exceeds in 1375 * if the specified nr_events exceeds internal limits. May fail
1376 * with -EAGAIN if the specified nr_even 1376 * with -EAGAIN if the specified nr_events exceeds the user's limit
1377 * of available events. May fail with - 1377 * of available events. May fail with -ENOMEM if insufficient kernel
1378 * resources are available. May fail wi 1378 * resources are available. May fail with -EFAULT if an invalid
1379 * pointer is passed for ctxp. Will fai 1379 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1380 * implemented. 1380 * implemented.
1381 */ 1381 */
1382 SYSCALL_DEFINE2(io_setup, unsigned, nr_events 1382 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1383 { 1383 {
1384 struct kioctx *ioctx = NULL; 1384 struct kioctx *ioctx = NULL;
1385 unsigned long ctx; 1385 unsigned long ctx;
1386 long ret; 1386 long ret;
1387 1387
1388 ret = get_user(ctx, ctxp); 1388 ret = get_user(ctx, ctxp);
1389 if (unlikely(ret)) 1389 if (unlikely(ret))
1390 goto out; 1390 goto out;
1391 1391
1392 ret = -EINVAL; 1392 ret = -EINVAL;
1393 if (unlikely(ctx || nr_events == 0)) 1393 if (unlikely(ctx || nr_events == 0)) {
1394 pr_debug("EINVAL: ctx %lu nr_ 1394 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1395 ctx, nr_events); 1395 ctx, nr_events);
1396 goto out; 1396 goto out;
1397 } 1397 }
1398 1398
1399 ioctx = ioctx_alloc(nr_events); 1399 ioctx = ioctx_alloc(nr_events);
1400 ret = PTR_ERR(ioctx); 1400 ret = PTR_ERR(ioctx);
1401 if (!IS_ERR(ioctx)) { 1401 if (!IS_ERR(ioctx)) {
1402 ret = put_user(ioctx->user_id 1402 ret = put_user(ioctx->user_id, ctxp);
1403 if (ret) 1403 if (ret)
1404 kill_ioctx(current->m 1404 kill_ioctx(current->mm, ioctx, NULL);
1405 percpu_ref_put(&ioctx->users) 1405 percpu_ref_put(&ioctx->users);
1406 } 1406 }
1407 1407
1408 out: 1408 out:
1409 return ret; 1409 return ret;
1410 } 1410 }
1411 1411
1412 #ifdef CONFIG_COMPAT 1412 #ifdef CONFIG_COMPAT
1413 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr 1413 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1414 { 1414 {
1415 struct kioctx *ioctx = NULL; 1415 struct kioctx *ioctx = NULL;
1416 unsigned long ctx; 1416 unsigned long ctx;
1417 long ret; 1417 long ret;
1418 1418
1419 ret = get_user(ctx, ctx32p); 1419 ret = get_user(ctx, ctx32p);
1420 if (unlikely(ret)) 1420 if (unlikely(ret))
1421 goto out; 1421 goto out;
1422 1422
1423 ret = -EINVAL; 1423 ret = -EINVAL;
1424 if (unlikely(ctx || nr_events == 0)) 1424 if (unlikely(ctx || nr_events == 0)) {
1425 pr_debug("EINVAL: ctx %lu nr_ 1425 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1426 ctx, nr_events); 1426 ctx, nr_events);
1427 goto out; 1427 goto out;
1428 } 1428 }
1429 1429
1430 ioctx = ioctx_alloc(nr_events); 1430 ioctx = ioctx_alloc(nr_events);
1431 ret = PTR_ERR(ioctx); 1431 ret = PTR_ERR(ioctx);
1432 if (!IS_ERR(ioctx)) { 1432 if (!IS_ERR(ioctx)) {
1433 /* truncating is ok because i 1433 /* truncating is ok because it's a user address */
1434 ret = put_user((u32)ioctx->us 1434 ret = put_user((u32)ioctx->user_id, ctx32p);
1435 if (ret) 1435 if (ret)
1436 kill_ioctx(current->m 1436 kill_ioctx(current->mm, ioctx, NULL);
1437 percpu_ref_put(&ioctx->users) 1437 percpu_ref_put(&ioctx->users);
1438 } 1438 }
1439 1439
1440 out: 1440 out:
1441 return ret; 1441 return ret;
1442 } 1442 }
1443 #endif 1443 #endif
1444 1444
1445 /* sys_io_destroy: 1445 /* sys_io_destroy:
1446 * Destroy the aio_context specified. M 1446 * Destroy the aio_context specified. May cancel any outstanding
1447 * AIOs and block on completion. Will f 1447 * AIOs and block on completion. Will fail with -ENOSYS if not
1448 * implemented. May fail with -EINVAL i 1448 * implemented. May fail with -EINVAL if the context pointed to
1449 * is invalid. 1449 * is invalid.
1450 */ 1450 */
1451 SYSCALL_DEFINE1(io_destroy, aio_context_t, ct 1451 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1452 { 1452 {
1453 struct kioctx *ioctx = lookup_ioctx(c 1453 struct kioctx *ioctx = lookup_ioctx(ctx);
1454 if (likely(NULL != ioctx)) { 1454 if (likely(NULL != ioctx)) {
1455 struct ctx_rq_wait wait; 1455 struct ctx_rq_wait wait;
1456 int ret; 1456 int ret;
1457 1457
1458 init_completion(&wait.comp); 1458 init_completion(&wait.comp);
1459 atomic_set(&wait.count, 1); 1459 atomic_set(&wait.count, 1);
1460 1460
1461 /* Pass requests_done to kill 1461 /* Pass requests_done to kill_ioctx() where it can be set
1462 * in a thread-safe way. If w 1462 * in a thread-safe way. If we try to set it here then we have
1463 * a race condition if two io 1463 * a race condition if two io_destroy() called simultaneously.
1464 */ 1464 */
1465 ret = kill_ioctx(current->mm, 1465 ret = kill_ioctx(current->mm, ioctx, &wait);
1466 percpu_ref_put(&ioctx->users) 1466 percpu_ref_put(&ioctx->users);
1467 1467
1468 /* Wait until all IO for the 1468 /* Wait until all IO for the context are done. Otherwise kernel
1469 * keep using user-space buff 1469 * keep using user-space buffers even if user thinks the context
1470 * is destroyed. 1470 * is destroyed.
1471 */ 1471 */
1472 if (!ret) 1472 if (!ret)
1473 wait_for_completion(& 1473 wait_for_completion(&wait.comp);
1474 1474
1475 return ret; 1475 return ret;
1476 } 1476 }
1477 pr_debug("EINVAL: invalid context id\ 1477 pr_debug("EINVAL: invalid context id\n");
1478 return -EINVAL; 1478 return -EINVAL;
1479 } 1479 }
1480 1480
1481 static void aio_remove_iocb(struct aio_kiocb 1481 static void aio_remove_iocb(struct aio_kiocb *iocb)
1482 { 1482 {
1483 struct kioctx *ctx = iocb->ki_ctx; 1483 struct kioctx *ctx = iocb->ki_ctx;
1484 unsigned long flags; 1484 unsigned long flags;
1485 1485
1486 spin_lock_irqsave(&ctx->ctx_lock, fla 1486 spin_lock_irqsave(&ctx->ctx_lock, flags);
1487 list_del(&iocb->ki_list); 1487 list_del(&iocb->ki_list);
1488 spin_unlock_irqrestore(&ctx->ctx_lock 1488 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1489 } 1489 }
1490 1490
1491 static void aio_complete_rw(struct kiocb *kio 1491 static void aio_complete_rw(struct kiocb *kiocb, long res)
1492 { 1492 {
1493 struct aio_kiocb *iocb = container_of 1493 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1494 1494
1495 if (!list_empty_careful(&iocb->ki_lis 1495 if (!list_empty_careful(&iocb->ki_list))
1496 aio_remove_iocb(iocb); 1496 aio_remove_iocb(iocb);
1497 1497
1498 if (kiocb->ki_flags & IOCB_WRITE) { 1498 if (kiocb->ki_flags & IOCB_WRITE) {
1499 struct inode *inode = file_in 1499 struct inode *inode = file_inode(kiocb->ki_filp);
1500 1500
1501 if (S_ISREG(inode->i_mode)) 1501 if (S_ISREG(inode->i_mode))
1502 kiocb_end_write(kiocb 1502 kiocb_end_write(kiocb);
1503 } 1503 }
1504 1504
1505 iocb->ki_res.res = res; 1505 iocb->ki_res.res = res;
1506 iocb->ki_res.res2 = 0; 1506 iocb->ki_res.res2 = 0;
1507 iocb_put(iocb); 1507 iocb_put(iocb);
1508 } 1508 }
1509 1509
1510 static int aio_prep_rw(struct kiocb *req, con 1510 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb, int rw_type)
1511 { 1511 {
1512 int ret; 1512 int ret;
1513 1513
1514 req->ki_complete = aio_complete_rw; 1514 req->ki_complete = aio_complete_rw;
1515 req->private = NULL; 1515 req->private = NULL;
1516 req->ki_pos = iocb->aio_offset; 1516 req->ki_pos = iocb->aio_offset;
1517 req->ki_flags = req->ki_filp->f_iocb_ 1517 req->ki_flags = req->ki_filp->f_iocb_flags | IOCB_AIO_RW;
1518 if (iocb->aio_flags & IOCB_FLAG_RESFD 1518 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1519 req->ki_flags |= IOCB_EVENTFD 1519 req->ki_flags |= IOCB_EVENTFD;
1520 if (iocb->aio_flags & IOCB_FLAG_IOPRI 1520 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1521 /* 1521 /*
1522 * If the IOCB_FLAG_IOPRIO fl 1522 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1523 * aio_reqprio is interpreted 1523 * aio_reqprio is interpreted as an I/O scheduling
1524 * class and priority. 1524 * class and priority.
1525 */ 1525 */
1526 ret = ioprio_check_cap(iocb-> 1526 ret = ioprio_check_cap(iocb->aio_reqprio);
1527 if (ret) { 1527 if (ret) {
1528 pr_debug("aio ioprio 1528 pr_debug("aio ioprio check cap error: %d\n", ret);
1529 return ret; 1529 return ret;
1530 } 1530 }
1531 1531
1532 req->ki_ioprio = iocb->aio_re 1532 req->ki_ioprio = iocb->aio_reqprio;
1533 } else 1533 } else
1534 req->ki_ioprio = get_current_ 1534 req->ki_ioprio = get_current_ioprio();
1535 1535
1536 ret = kiocb_set_rw_flags(req, iocb->a 1536 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags, rw_type);
1537 if (unlikely(ret)) 1537 if (unlikely(ret))
1538 return ret; 1538 return ret;
1539 1539
1540 req->ki_flags &= ~IOCB_HIPRI; /* no o 1540 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1541 return 0; 1541 return 0;
1542 } 1542 }
1543 1543
1544 static ssize_t aio_setup_rw(int rw, const str 1544 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1545 struct iovec **iovec, bool ve 1545 struct iovec **iovec, bool vectored, bool compat,
1546 struct iov_iter *iter) 1546 struct iov_iter *iter)
1547 { 1547 {
1548 void __user *buf = (void __user *)(ui 1548 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1549 size_t len = iocb->aio_nbytes; 1549 size_t len = iocb->aio_nbytes;
1550 1550
1551 if (!vectored) { 1551 if (!vectored) {
1552 ssize_t ret = import_ubuf(rw, 1552 ssize_t ret = import_ubuf(rw, buf, len, iter);
1553 *iovec = NULL; 1553 *iovec = NULL;
1554 return ret; 1554 return ret;
1555 } 1555 }
1556 1556
1557 return __import_iovec(rw, buf, len, U 1557 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1558 } 1558 }
1559 1559
1560 static inline void aio_rw_done(struct kiocb * 1560 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1561 { 1561 {
1562 switch (ret) { 1562 switch (ret) {
1563 case -EIOCBQUEUED: 1563 case -EIOCBQUEUED:
1564 break; 1564 break;
1565 case -ERESTARTSYS: 1565 case -ERESTARTSYS:
1566 case -ERESTARTNOINTR: 1566 case -ERESTARTNOINTR:
1567 case -ERESTARTNOHAND: 1567 case -ERESTARTNOHAND:
1568 case -ERESTART_RESTARTBLOCK: 1568 case -ERESTART_RESTARTBLOCK:
1569 /* 1569 /*
1570 * There's no easy way to res 1570 * There's no easy way to restart the syscall since other AIO's
1571 * may be already running. Ju 1571 * may be already running. Just fail this IO with EINTR.
1572 */ 1572 */
1573 ret = -EINTR; 1573 ret = -EINTR;
1574 fallthrough; 1574 fallthrough;
1575 default: 1575 default:
1576 req->ki_complete(req, ret); 1576 req->ki_complete(req, ret);
1577 } 1577 }
1578 } 1578 }
1579 1579
1580 static int aio_read(struct kiocb *req, const 1580 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1581 bool vectored, bool c 1581 bool vectored, bool compat)
1582 { 1582 {
1583 struct iovec inline_vecs[UIO_FASTIOV] 1583 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1584 struct iov_iter iter; 1584 struct iov_iter iter;
1585 struct file *file; 1585 struct file *file;
1586 int ret; 1586 int ret;
1587 1587
1588 ret = aio_prep_rw(req, iocb, READ); 1588 ret = aio_prep_rw(req, iocb, READ);
1589 if (ret) 1589 if (ret)
1590 return ret; 1590 return ret;
1591 file = req->ki_filp; 1591 file = req->ki_filp;
1592 if (unlikely(!(file->f_mode & FMODE_R 1592 if (unlikely(!(file->f_mode & FMODE_READ)))
1593 return -EBADF; 1593 return -EBADF;
1594 if (unlikely(!file->f_op->read_iter)) 1594 if (unlikely(!file->f_op->read_iter))
1595 return -EINVAL; 1595 return -EINVAL;
1596 1596
1597 ret = aio_setup_rw(ITER_DEST, iocb, & 1597 ret = aio_setup_rw(ITER_DEST, iocb, &iovec, vectored, compat, &iter);
1598 if (ret < 0) 1598 if (ret < 0)
1599 return ret; 1599 return ret;
1600 ret = rw_verify_area(READ, file, &req 1600 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1601 if (!ret) 1601 if (!ret)
1602 aio_rw_done(req, file->f_op-> 1602 aio_rw_done(req, file->f_op->read_iter(req, &iter));
1603 kfree(iovec); 1603 kfree(iovec);
1604 return ret; 1604 return ret;
1605 } 1605 }
1606 1606
1607 static int aio_write(struct kiocb *req, const 1607 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1608 bool vectored, bool 1608 bool vectored, bool compat)
1609 { 1609 {
1610 struct iovec inline_vecs[UIO_FASTIOV] 1610 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1611 struct iov_iter iter; 1611 struct iov_iter iter;
1612 struct file *file; 1612 struct file *file;
1613 int ret; 1613 int ret;
1614 1614
1615 ret = aio_prep_rw(req, iocb, WRITE); 1615 ret = aio_prep_rw(req, iocb, WRITE);
1616 if (ret) 1616 if (ret)
1617 return ret; 1617 return ret;
1618 file = req->ki_filp; 1618 file = req->ki_filp;
1619 1619
1620 if (unlikely(!(file->f_mode & FMODE_W 1620 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1621 return -EBADF; 1621 return -EBADF;
1622 if (unlikely(!file->f_op->write_iter) 1622 if (unlikely(!file->f_op->write_iter))
1623 return -EINVAL; 1623 return -EINVAL;
1624 1624
1625 ret = aio_setup_rw(ITER_SOURCE, iocb, 1625 ret = aio_setup_rw(ITER_SOURCE, iocb, &iovec, vectored, compat, &iter);
1626 if (ret < 0) 1626 if (ret < 0)
1627 return ret; 1627 return ret;
1628 ret = rw_verify_area(WRITE, file, &re 1628 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1629 if (!ret) { 1629 if (!ret) {
1630 if (S_ISREG(file_inode(file)- 1630 if (S_ISREG(file_inode(file)->i_mode))
1631 kiocb_start_write(req 1631 kiocb_start_write(req);
1632 req->ki_flags |= IOCB_WRITE; 1632 req->ki_flags |= IOCB_WRITE;
1633 aio_rw_done(req, file->f_op-> 1633 aio_rw_done(req, file->f_op->write_iter(req, &iter));
1634 } 1634 }
1635 kfree(iovec); 1635 kfree(iovec);
1636 return ret; 1636 return ret;
1637 } 1637 }
1638 1638
1639 static void aio_fsync_work(struct work_struct 1639 static void aio_fsync_work(struct work_struct *work)
1640 { 1640 {
1641 struct aio_kiocb *iocb = container_of 1641 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1642 const struct cred *old_cred = overrid 1642 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1643 1643
1644 iocb->ki_res.res = vfs_fsync(iocb->fs 1644 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1645 revert_creds(old_cred); 1645 revert_creds(old_cred);
1646 put_cred(iocb->fsync.creds); 1646 put_cred(iocb->fsync.creds);
1647 iocb_put(iocb); 1647 iocb_put(iocb);
1648 } 1648 }
1649 1649
1650 static int aio_fsync(struct fsync_iocb *req, 1650 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1651 bool datasync) 1651 bool datasync)
1652 { 1652 {
1653 if (unlikely(iocb->aio_buf || iocb->a 1653 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1654 iocb->aio_rw_flags)) 1654 iocb->aio_rw_flags))
1655 return -EINVAL; 1655 return -EINVAL;
1656 1656
1657 if (unlikely(!req->file->f_op->fsync) 1657 if (unlikely(!req->file->f_op->fsync))
1658 return -EINVAL; 1658 return -EINVAL;
1659 1659
1660 req->creds = prepare_creds(); 1660 req->creds = prepare_creds();
1661 if (!req->creds) 1661 if (!req->creds)
1662 return -ENOMEM; 1662 return -ENOMEM;
1663 1663
1664 req->datasync = datasync; 1664 req->datasync = datasync;
1665 INIT_WORK(&req->work, aio_fsync_work) 1665 INIT_WORK(&req->work, aio_fsync_work);
1666 schedule_work(&req->work); 1666 schedule_work(&req->work);
1667 return 0; 1667 return 0;
1668 } 1668 }
1669 1669
1670 static void aio_poll_put_work(struct work_str 1670 static void aio_poll_put_work(struct work_struct *work)
1671 { 1671 {
1672 struct poll_iocb *req = container_of( 1672 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1673 struct aio_kiocb *iocb = container_of 1673 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1674 1674
1675 iocb_put(iocb); 1675 iocb_put(iocb);
1676 } 1676 }
1677 1677
1678 /* 1678 /*
1679 * Safely lock the waitqueue which the reques 1679 * Safely lock the waitqueue which the request is on, synchronizing with the
1680 * case where the ->poll() provider decides t 1680 * case where the ->poll() provider decides to free its waitqueue early.
1681 * 1681 *
1682 * Returns true on success, meaning that req- 1682 * Returns true on success, meaning that req->head->lock was locked, req->wait
1683 * is on req->head, and an RCU read lock was 1683 * is on req->head, and an RCU read lock was taken. Returns false if the
1684 * request was already removed from its waitq 1684 * request was already removed from its waitqueue (which might no longer exist).
1685 */ 1685 */
1686 static bool poll_iocb_lock_wq(struct poll_ioc 1686 static bool poll_iocb_lock_wq(struct poll_iocb *req)
1687 { 1687 {
1688 wait_queue_head_t *head; 1688 wait_queue_head_t *head;
1689 1689
1690 /* 1690 /*
1691 * While we hold the waitqueue lock a 1691 * While we hold the waitqueue lock and the waitqueue is nonempty,
1692 * wake_up_pollfree() will wait for u 1692 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1693 * lock in the first place can race w 1693 * lock in the first place can race with the waitqueue being freed.
1694 * 1694 *
1695 * We solve this as eventpoll does: b 1695 * We solve this as eventpoll does: by taking advantage of the fact that
1696 * all users of wake_up_pollfree() wi 1696 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1697 * we enter rcu_read_lock() and see t 1697 * we enter rcu_read_lock() and see that the pointer to the queue is
1698 * non-NULL, we can then lock it with 1698 * non-NULL, we can then lock it without the memory being freed out from
1699 * under us, then check whether the r 1699 * under us, then check whether the request is still on the queue.
1700 * 1700 *
1701 * Keep holding rcu_read_lock() as lo 1701 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1702 * case the caller deletes the entry 1702 * case the caller deletes the entry from the queue, leaving it empty.
1703 * In that case, only RCU prevents th 1703 * In that case, only RCU prevents the queue memory from being freed.
1704 */ 1704 */
1705 rcu_read_lock(); 1705 rcu_read_lock();
1706 head = smp_load_acquire(&req->head); 1706 head = smp_load_acquire(&req->head);
1707 if (head) { 1707 if (head) {
1708 spin_lock(&head->lock); 1708 spin_lock(&head->lock);
1709 if (!list_empty(&req->wait.en 1709 if (!list_empty(&req->wait.entry))
1710 return true; 1710 return true;
1711 spin_unlock(&head->lock); 1711 spin_unlock(&head->lock);
1712 } 1712 }
1713 rcu_read_unlock(); 1713 rcu_read_unlock();
1714 return false; 1714 return false;
1715 } 1715 }
1716 1716
1717 static void poll_iocb_unlock_wq(struct poll_i 1717 static void poll_iocb_unlock_wq(struct poll_iocb *req)
1718 { 1718 {
1719 spin_unlock(&req->head->lock); 1719 spin_unlock(&req->head->lock);
1720 rcu_read_unlock(); 1720 rcu_read_unlock();
1721 } 1721 }
1722 1722
1723 static void aio_poll_complete_work(struct wor 1723 static void aio_poll_complete_work(struct work_struct *work)
1724 { 1724 {
1725 struct poll_iocb *req = container_of( 1725 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1726 struct aio_kiocb *iocb = container_of 1726 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1727 struct poll_table_struct pt = { ._key 1727 struct poll_table_struct pt = { ._key = req->events };
1728 struct kioctx *ctx = iocb->ki_ctx; 1728 struct kioctx *ctx = iocb->ki_ctx;
1729 __poll_t mask = 0; 1729 __poll_t mask = 0;
1730 1730
1731 if (!READ_ONCE(req->cancelled)) 1731 if (!READ_ONCE(req->cancelled))
1732 mask = vfs_poll(req->file, &p 1732 mask = vfs_poll(req->file, &pt) & req->events;
1733 1733
1734 /* 1734 /*
1735 * Note that ->ki_cancel callers also 1735 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1736 * calling ->ki_cancel. We need the 1736 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1737 * synchronize with them. In the can 1737 * synchronize with them. In the cancellation case the list_del_init
1738 * itself is not actually needed, but 1738 * itself is not actually needed, but harmless so we keep it in to
1739 * avoid further branches in the fast 1739 * avoid further branches in the fast path.
1740 */ 1740 */
1741 spin_lock_irq(&ctx->ctx_lock); 1741 spin_lock_irq(&ctx->ctx_lock);
1742 if (poll_iocb_lock_wq(req)) { 1742 if (poll_iocb_lock_wq(req)) {
1743 if (!mask && !READ_ONCE(req-> 1743 if (!mask && !READ_ONCE(req->cancelled)) {
1744 /* 1744 /*
1745 * The request isn't 1745 * The request isn't actually ready to be completed yet.
1746 * Reschedule complet 1746 * Reschedule completion if another wakeup came in.
1747 */ 1747 */
1748 if (req->work_need_re 1748 if (req->work_need_resched) {
1749 schedule_work 1749 schedule_work(&req->work);
1750 req->work_nee 1750 req->work_need_resched = false;
1751 } else { 1751 } else {
1752 req->work_sch 1752 req->work_scheduled = false;
1753 } 1753 }
1754 poll_iocb_unlock_wq(r 1754 poll_iocb_unlock_wq(req);
1755 spin_unlock_irq(&ctx- 1755 spin_unlock_irq(&ctx->ctx_lock);
1756 return; 1756 return;
1757 } 1757 }
1758 list_del_init(&req->wait.entr 1758 list_del_init(&req->wait.entry);
1759 poll_iocb_unlock_wq(req); 1759 poll_iocb_unlock_wq(req);
1760 } /* else, POLLFREE has freed the wai 1760 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1761 list_del_init(&iocb->ki_list); 1761 list_del_init(&iocb->ki_list);
1762 iocb->ki_res.res = mangle_poll(mask); 1762 iocb->ki_res.res = mangle_poll(mask);
1763 spin_unlock_irq(&ctx->ctx_lock); 1763 spin_unlock_irq(&ctx->ctx_lock);
1764 1764
1765 iocb_put(iocb); 1765 iocb_put(iocb);
1766 } 1766 }
1767 1767
1768 /* assumes we are called with irqs disabled * 1768 /* assumes we are called with irqs disabled */
1769 static int aio_poll_cancel(struct kiocb *iocb 1769 static int aio_poll_cancel(struct kiocb *iocb)
1770 { 1770 {
1771 struct aio_kiocb *aiocb = container_o 1771 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1772 struct poll_iocb *req = &aiocb->poll; 1772 struct poll_iocb *req = &aiocb->poll;
1773 1773
1774 if (poll_iocb_lock_wq(req)) { 1774 if (poll_iocb_lock_wq(req)) {
1775 WRITE_ONCE(req->cancelled, tr 1775 WRITE_ONCE(req->cancelled, true);
1776 if (!req->work_scheduled) { 1776 if (!req->work_scheduled) {
1777 schedule_work(&aiocb- 1777 schedule_work(&aiocb->poll.work);
1778 req->work_scheduled = 1778 req->work_scheduled = true;
1779 } 1779 }
1780 poll_iocb_unlock_wq(req); 1780 poll_iocb_unlock_wq(req);
1781 } /* else, the request was force-canc 1781 } /* else, the request was force-cancelled by POLLFREE already */
1782 1782
1783 return 0; 1783 return 0;
1784 } 1784 }
1785 1785
1786 static int aio_poll_wake(struct wait_queue_en 1786 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1787 void *key) 1787 void *key)
1788 { 1788 {
1789 struct poll_iocb *req = container_of( 1789 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1790 struct aio_kiocb *iocb = container_of 1790 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1791 __poll_t mask = key_to_poll(key); 1791 __poll_t mask = key_to_poll(key);
1792 unsigned long flags; 1792 unsigned long flags;
1793 1793
1794 /* for instances that support it chec 1794 /* for instances that support it check for an event match first: */
1795 if (mask && !(mask & req->events)) 1795 if (mask && !(mask & req->events))
1796 return 0; 1796 return 0;
1797 1797
1798 /* 1798 /*
1799 * Complete the request inline if pos 1799 * Complete the request inline if possible. This requires that three
1800 * conditions be met: 1800 * conditions be met:
1801 * 1. An event mask must have been 1801 * 1. An event mask must have been passed. If a plain wakeup was done
1802 * instead, then mask == 0 and w 1802 * instead, then mask == 0 and we have to call vfs_poll() to get
1803 * the events, so inline complet 1803 * the events, so inline completion isn't possible.
1804 * 2. The completion work must not 1804 * 2. The completion work must not have already been scheduled.
1805 * 3. ctx_lock must not be busy. W 1805 * 3. ctx_lock must not be busy. We have to use trylock because we
1806 * already hold the waitqueue lo 1806 * already hold the waitqueue lock, so this inverts the normal
1807 * locking order. Use irqsave/i 1807 * locking order. Use irqsave/irqrestore because not all
1808 * filesystems (e.g. fuse) call 1808 * filesystems (e.g. fuse) call this function with IRQs disabled,
1809 * yet IRQs have to be disabled 1809 * yet IRQs have to be disabled before ctx_lock is obtained.
1810 */ 1810 */
1811 if (mask && !req->work_scheduled && 1811 if (mask && !req->work_scheduled &&
1812 spin_trylock_irqsave(&iocb->ki_ct 1812 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1813 struct kioctx *ctx = iocb->ki 1813 struct kioctx *ctx = iocb->ki_ctx;
1814 1814
1815 list_del_init(&req->wait.entr 1815 list_del_init(&req->wait.entry);
1816 list_del(&iocb->ki_list); 1816 list_del(&iocb->ki_list);
1817 iocb->ki_res.res = mangle_pol 1817 iocb->ki_res.res = mangle_poll(mask);
1818 if (iocb->ki_eventfd && !even 1818 if (iocb->ki_eventfd && !eventfd_signal_allowed()) {
1819 iocb = NULL; 1819 iocb = NULL;
1820 INIT_WORK(&req->work, 1820 INIT_WORK(&req->work, aio_poll_put_work);
1821 schedule_work(&req->w 1821 schedule_work(&req->work);
1822 } 1822 }
1823 spin_unlock_irqrestore(&ctx-> 1823 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1824 if (iocb) 1824 if (iocb)
1825 iocb_put(iocb); 1825 iocb_put(iocb);
1826 } else { 1826 } else {
1827 /* 1827 /*
1828 * Schedule the completion wo 1828 * Schedule the completion work if needed. If it was already
1829 * scheduled, record that ano 1829 * scheduled, record that another wakeup came in.
1830 * 1830 *
1831 * Don't remove the request f 1831 * Don't remove the request from the waitqueue here, as it might
1832 * not actually be complete y 1832 * not actually be complete yet (we won't know until vfs_poll()
1833 * is called), and we must no 1833 * is called), and we must not miss any wakeups. POLLFREE is an
1834 * exception to this; see bel 1834 * exception to this; see below.
1835 */ 1835 */
1836 if (req->work_scheduled) { 1836 if (req->work_scheduled) {
1837 req->work_need_resche 1837 req->work_need_resched = true;
1838 } else { 1838 } else {
1839 schedule_work(&req->w 1839 schedule_work(&req->work);
1840 req->work_scheduled = 1840 req->work_scheduled = true;
1841 } 1841 }
1842 1842
1843 /* 1843 /*
1844 * If the waitqueue is being 1844 * If the waitqueue is being freed early but we can't complete
1845 * the request inline, we hav 1845 * the request inline, we have to tear down the request as best
1846 * we can. That means immedi 1846 * we can. That means immediately removing the request from its
1847 * waitqueue and preventing a 1847 * waitqueue and preventing all further accesses to the
1848 * waitqueue via the request. 1848 * waitqueue via the request. We also need to schedule the
1849 * completion work (done abov 1849 * completion work (done above). Also mark the request as
1850 * cancelled, to potentially 1850 * cancelled, to potentially skip an unneeded call to ->poll().
1851 */ 1851 */
1852 if (mask & POLLFREE) { 1852 if (mask & POLLFREE) {
1853 WRITE_ONCE(req->cance 1853 WRITE_ONCE(req->cancelled, true);
1854 list_del_init(&req->w 1854 list_del_init(&req->wait.entry);
1855 1855
1856 /* 1856 /*
1857 * Careful: this *mus 1857 * Careful: this *must* be the last step, since as soon
1858 * as req->head is NU 1858 * as req->head is NULL'ed out, the request can be
1859 * completed and free 1859 * completed and freed, since aio_poll_complete_work()
1860 * will no longer nee 1860 * will no longer need to take the waitqueue lock.
1861 */ 1861 */
1862 smp_store_release(&re 1862 smp_store_release(&req->head, NULL);
1863 } 1863 }
1864 } 1864 }
1865 return 1; 1865 return 1;
1866 } 1866 }
1867 1867
1868 struct aio_poll_table { 1868 struct aio_poll_table {
1869 struct poll_table_struct pt; 1869 struct poll_table_struct pt;
1870 struct aio_kiocb *iocb 1870 struct aio_kiocb *iocb;
1871 bool queue 1871 bool queued;
1872 int error 1872 int error;
1873 }; 1873 };
1874 1874
1875 static void 1875 static void
1876 aio_poll_queue_proc(struct file *file, struct 1876 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1877 struct poll_table_struct *p) 1877 struct poll_table_struct *p)
1878 { 1878 {
1879 struct aio_poll_table *pt = container 1879 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1880 1880
1881 /* multiple wait queues per file are 1881 /* multiple wait queues per file are not supported */
1882 if (unlikely(pt->queued)) { 1882 if (unlikely(pt->queued)) {
1883 pt->error = -EINVAL; 1883 pt->error = -EINVAL;
1884 return; 1884 return;
1885 } 1885 }
1886 1886
1887 pt->queued = true; 1887 pt->queued = true;
1888 pt->error = 0; 1888 pt->error = 0;
1889 pt->iocb->poll.head = head; 1889 pt->iocb->poll.head = head;
1890 add_wait_queue(head, &pt->iocb->poll. 1890 add_wait_queue(head, &pt->iocb->poll.wait);
1891 } 1891 }
1892 1892
1893 static int aio_poll(struct aio_kiocb *aiocb, 1893 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1894 { 1894 {
1895 struct kioctx *ctx = aiocb->ki_ctx; 1895 struct kioctx *ctx = aiocb->ki_ctx;
1896 struct poll_iocb *req = &aiocb->poll; 1896 struct poll_iocb *req = &aiocb->poll;
1897 struct aio_poll_table apt; 1897 struct aio_poll_table apt;
1898 bool cancel = false; 1898 bool cancel = false;
1899 __poll_t mask; 1899 __poll_t mask;
1900 1900
1901 /* reject any unknown events outside 1901 /* reject any unknown events outside the normal event mask. */
1902 if ((u16)iocb->aio_buf != iocb->aio_b 1902 if ((u16)iocb->aio_buf != iocb->aio_buf)
1903 return -EINVAL; 1903 return -EINVAL;
1904 /* reject fields that are not defined 1904 /* reject fields that are not defined for poll */
1905 if (iocb->aio_offset || iocb->aio_nby 1905 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1906 return -EINVAL; 1906 return -EINVAL;
1907 1907
1908 INIT_WORK(&req->work, aio_poll_comple 1908 INIT_WORK(&req->work, aio_poll_complete_work);
1909 req->events = demangle_poll(iocb->aio 1909 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1910 1910
1911 req->head = NULL; 1911 req->head = NULL;
1912 req->cancelled = false; 1912 req->cancelled = false;
1913 req->work_scheduled = false; 1913 req->work_scheduled = false;
1914 req->work_need_resched = false; 1914 req->work_need_resched = false;
1915 1915
1916 apt.pt._qproc = aio_poll_queue_proc; 1916 apt.pt._qproc = aio_poll_queue_proc;
1917 apt.pt._key = req->events; 1917 apt.pt._key = req->events;
1918 apt.iocb = aiocb; 1918 apt.iocb = aiocb;
1919 apt.queued = false; 1919 apt.queued = false;
1920 apt.error = -EINVAL; /* same as no su 1920 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1921 1921
1922 /* initialized the list so that we ca 1922 /* initialized the list so that we can do list_empty checks */
1923 INIT_LIST_HEAD(&req->wait.entry); 1923 INIT_LIST_HEAD(&req->wait.entry);
1924 init_waitqueue_func_entry(&req->wait, 1924 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1925 1925
1926 mask = vfs_poll(req->file, &apt.pt) & 1926 mask = vfs_poll(req->file, &apt.pt) & req->events;
1927 spin_lock_irq(&ctx->ctx_lock); 1927 spin_lock_irq(&ctx->ctx_lock);
1928 if (likely(apt.queued)) { 1928 if (likely(apt.queued)) {
1929 bool on_queue = poll_iocb_loc 1929 bool on_queue = poll_iocb_lock_wq(req);
1930 1930
1931 if (!on_queue || req->work_sc 1931 if (!on_queue || req->work_scheduled) {
1932 /* 1932 /*
1933 * aio_poll_wake() al 1933 * aio_poll_wake() already either scheduled the async
1934 * completion work, o 1934 * completion work, or completed the request inline.
1935 */ 1935 */
1936 if (apt.error) /* uns 1936 if (apt.error) /* unsupported case: multiple queues */
1937 cancel = true 1937 cancel = true;
1938 apt.error = 0; 1938 apt.error = 0;
1939 mask = 0; 1939 mask = 0;
1940 } 1940 }
1941 if (mask || apt.error) { 1941 if (mask || apt.error) {
1942 /* Steal to complete 1942 /* Steal to complete synchronously. */
1943 list_del_init(&req->w 1943 list_del_init(&req->wait.entry);
1944 } else if (cancel) { 1944 } else if (cancel) {
1945 /* Cancel if possible 1945 /* Cancel if possible (may be too late though). */
1946 WRITE_ONCE(req->cance 1946 WRITE_ONCE(req->cancelled, true);
1947 } else if (on_queue) { 1947 } else if (on_queue) {
1948 /* 1948 /*
1949 * Actually waiting f 1949 * Actually waiting for an event, so add the request to
1950 * active_reqs so tha 1950 * active_reqs so that it can be cancelled if needed.
1951 */ 1951 */
1952 list_add_tail(&aiocb- 1952 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1953 aiocb->ki_cancel = ai 1953 aiocb->ki_cancel = aio_poll_cancel;
1954 } 1954 }
1955 if (on_queue) 1955 if (on_queue)
1956 poll_iocb_unlock_wq(r 1956 poll_iocb_unlock_wq(req);
1957 } 1957 }
1958 if (mask) { /* no async, we'd stolen 1958 if (mask) { /* no async, we'd stolen it */
1959 aiocb->ki_res.res = mangle_po 1959 aiocb->ki_res.res = mangle_poll(mask);
1960 apt.error = 0; 1960 apt.error = 0;
1961 } 1961 }
1962 spin_unlock_irq(&ctx->ctx_lock); 1962 spin_unlock_irq(&ctx->ctx_lock);
1963 if (mask) 1963 if (mask)
1964 iocb_put(aiocb); 1964 iocb_put(aiocb);
1965 return apt.error; 1965 return apt.error;
1966 } 1966 }
1967 1967
1968 static int __io_submit_one(struct kioctx *ctx 1968 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1969 struct iocb __user 1969 struct iocb __user *user_iocb, struct aio_kiocb *req,
1970 bool compat) 1970 bool compat)
1971 { 1971 {
1972 req->ki_filp = fget(iocb->aio_fildes) 1972 req->ki_filp = fget(iocb->aio_fildes);
1973 if (unlikely(!req->ki_filp)) 1973 if (unlikely(!req->ki_filp))
1974 return -EBADF; 1974 return -EBADF;
1975 1975
1976 if (iocb->aio_flags & IOCB_FLAG_RESFD 1976 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1977 struct eventfd_ctx *eventfd; 1977 struct eventfd_ctx *eventfd;
1978 /* 1978 /*
1979 * If the IOCB_FLAG_RESFD fla 1979 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1980 * instance of the file* now. 1980 * instance of the file* now. The file descriptor must be
1981 * an eventfd() fd, and will 1981 * an eventfd() fd, and will be signaled for each completed
1982 * event using the eventfd_si 1982 * event using the eventfd_signal() function.
1983 */ 1983 */
1984 eventfd = eventfd_ctx_fdget(i 1984 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1985 if (IS_ERR(eventfd)) 1985 if (IS_ERR(eventfd))
1986 return PTR_ERR(eventf 1986 return PTR_ERR(eventfd);
1987 1987
1988 req->ki_eventfd = eventfd; 1988 req->ki_eventfd = eventfd;
1989 } 1989 }
1990 1990
1991 if (unlikely(put_user(KIOCB_KEY, &use 1991 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1992 pr_debug("EFAULT: aio_key\n") 1992 pr_debug("EFAULT: aio_key\n");
1993 return -EFAULT; 1993 return -EFAULT;
1994 } 1994 }
1995 1995
1996 req->ki_res.obj = (u64)(unsigned long 1996 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1997 req->ki_res.data = iocb->aio_data; 1997 req->ki_res.data = iocb->aio_data;
1998 req->ki_res.res = 0; 1998 req->ki_res.res = 0;
1999 req->ki_res.res2 = 0; 1999 req->ki_res.res2 = 0;
2000 2000
2001 switch (iocb->aio_lio_opcode) { 2001 switch (iocb->aio_lio_opcode) {
2002 case IOCB_CMD_PREAD: 2002 case IOCB_CMD_PREAD:
2003 return aio_read(&req->rw, ioc 2003 return aio_read(&req->rw, iocb, false, compat);
2004 case IOCB_CMD_PWRITE: 2004 case IOCB_CMD_PWRITE:
2005 return aio_write(&req->rw, io 2005 return aio_write(&req->rw, iocb, false, compat);
2006 case IOCB_CMD_PREADV: 2006 case IOCB_CMD_PREADV:
2007 return aio_read(&req->rw, ioc 2007 return aio_read(&req->rw, iocb, true, compat);
2008 case IOCB_CMD_PWRITEV: 2008 case IOCB_CMD_PWRITEV:
2009 return aio_write(&req->rw, io 2009 return aio_write(&req->rw, iocb, true, compat);
2010 case IOCB_CMD_FSYNC: 2010 case IOCB_CMD_FSYNC:
2011 return aio_fsync(&req->fsync, 2011 return aio_fsync(&req->fsync, iocb, false);
2012 case IOCB_CMD_FDSYNC: 2012 case IOCB_CMD_FDSYNC:
2013 return aio_fsync(&req->fsync, 2013 return aio_fsync(&req->fsync, iocb, true);
2014 case IOCB_CMD_POLL: 2014 case IOCB_CMD_POLL:
2015 return aio_poll(req, iocb); 2015 return aio_poll(req, iocb);
2016 default: 2016 default:
2017 pr_debug("invalid aio operati 2017 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
2018 return -EINVAL; 2018 return -EINVAL;
2019 } 2019 }
2020 } 2020 }
2021 2021
2022 static int io_submit_one(struct kioctx *ctx, 2022 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
2023 bool compat) 2023 bool compat)
2024 { 2024 {
2025 struct aio_kiocb *req; 2025 struct aio_kiocb *req;
2026 struct iocb iocb; 2026 struct iocb iocb;
2027 int err; 2027 int err;
2028 2028
2029 if (unlikely(copy_from_user(&iocb, us 2029 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
2030 return -EFAULT; 2030 return -EFAULT;
2031 2031
2032 /* enforce forwards compatibility on 2032 /* enforce forwards compatibility on users */
2033 if (unlikely(iocb.aio_reserved2)) { 2033 if (unlikely(iocb.aio_reserved2)) {
2034 pr_debug("EINVAL: reserve fie 2034 pr_debug("EINVAL: reserve field set\n");
2035 return -EINVAL; 2035 return -EINVAL;
2036 } 2036 }
2037 2037
2038 /* prevent overflows */ 2038 /* prevent overflows */
2039 if (unlikely( 2039 if (unlikely(
2040 (iocb.aio_buf != (unsigned long)i 2040 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2041 (iocb.aio_nbytes != (size_t)iocb. 2041 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2042 ((ssize_t)iocb.aio_nbytes < 0) 2042 ((ssize_t)iocb.aio_nbytes < 0)
2043 )) { 2043 )) {
2044 pr_debug("EINVAL: overflow ch 2044 pr_debug("EINVAL: overflow check\n");
2045 return -EINVAL; 2045 return -EINVAL;
2046 } 2046 }
2047 2047
2048 req = aio_get_req(ctx); 2048 req = aio_get_req(ctx);
2049 if (unlikely(!req)) 2049 if (unlikely(!req))
2050 return -EAGAIN; 2050 return -EAGAIN;
2051 2051
2052 err = __io_submit_one(ctx, &iocb, use 2052 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2053 2053
2054 /* Done with the synchronous referenc 2054 /* Done with the synchronous reference */
2055 iocb_put(req); 2055 iocb_put(req);
2056 2056
2057 /* 2057 /*
2058 * If err is 0, we'd either done aio_ 2058 * If err is 0, we'd either done aio_complete() ourselves or have
2059 * arranged for that to be done async 2059 * arranged for that to be done asynchronously. Anything non-zero
2060 * means that we need to destroy req 2060 * means that we need to destroy req ourselves.
2061 */ 2061 */
2062 if (unlikely(err)) { 2062 if (unlikely(err)) {
2063 iocb_destroy(req); 2063 iocb_destroy(req);
2064 put_reqs_available(ctx, 1); 2064 put_reqs_available(ctx, 1);
2065 } 2065 }
2066 return err; 2066 return err;
2067 } 2067 }
2068 2068
2069 /* sys_io_submit: 2069 /* sys_io_submit:
2070 * Queue the nr iocbs pointed to by iocb 2070 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2071 * the number of iocbs queued. May retu 2071 * the number of iocbs queued. May return -EINVAL if the aio_context
2072 * specified by ctx_id is invalid, if nr 2072 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2073 * *iocbpp[0] is not properly initialize 2073 * *iocbpp[0] is not properly initialized, if the operation specified
2074 * is invalid for the file descriptor in 2074 * is invalid for the file descriptor in the iocb. May fail with
2075 * -EFAULT if any of the data structures 2075 * -EFAULT if any of the data structures point to invalid data. May
2076 * fail with -EBADF if the file descript 2076 * fail with -EBADF if the file descriptor specified in the first
2077 * iocb is invalid. May fail with -EAGA 2077 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2078 * are available to queue any iocbs. Wi 2078 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2079 * fail with -ENOSYS if not implemented. 2079 * fail with -ENOSYS if not implemented.
2080 */ 2080 */
2081 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx 2081 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2082 struct iocb __user * __user * 2082 struct iocb __user * __user *, iocbpp)
2083 { 2083 {
2084 struct kioctx *ctx; 2084 struct kioctx *ctx;
2085 long ret = 0; 2085 long ret = 0;
2086 int i = 0; 2086 int i = 0;
2087 struct blk_plug plug; 2087 struct blk_plug plug;
2088 2088
2089 if (unlikely(nr < 0)) 2089 if (unlikely(nr < 0))
2090 return -EINVAL; 2090 return -EINVAL;
2091 2091
2092 ctx = lookup_ioctx(ctx_id); 2092 ctx = lookup_ioctx(ctx_id);
2093 if (unlikely(!ctx)) { 2093 if (unlikely(!ctx)) {
2094 pr_debug("EINVAL: invalid con 2094 pr_debug("EINVAL: invalid context id\n");
2095 return -EINVAL; 2095 return -EINVAL;
2096 } 2096 }
2097 2097
2098 if (nr > ctx->nr_events) 2098 if (nr > ctx->nr_events)
2099 nr = ctx->nr_events; 2099 nr = ctx->nr_events;
2100 2100
2101 if (nr > AIO_PLUG_THRESHOLD) 2101 if (nr > AIO_PLUG_THRESHOLD)
2102 blk_start_plug(&plug); 2102 blk_start_plug(&plug);
2103 for (i = 0; i < nr; i++) { 2103 for (i = 0; i < nr; i++) {
2104 struct iocb __user *user_iocb 2104 struct iocb __user *user_iocb;
2105 2105
2106 if (unlikely(get_user(user_io 2106 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2107 ret = -EFAULT; 2107 ret = -EFAULT;
2108 break; 2108 break;
2109 } 2109 }
2110 2110
2111 ret = io_submit_one(ctx, user 2111 ret = io_submit_one(ctx, user_iocb, false);
2112 if (ret) 2112 if (ret)
2113 break; 2113 break;
2114 } 2114 }
2115 if (nr > AIO_PLUG_THRESHOLD) 2115 if (nr > AIO_PLUG_THRESHOLD)
2116 blk_finish_plug(&plug); 2116 blk_finish_plug(&plug);
2117 2117
2118 percpu_ref_put(&ctx->users); 2118 percpu_ref_put(&ctx->users);
2119 return i ? i : ret; 2119 return i ? i : ret;
2120 } 2120 }
2121 2121
2122 #ifdef CONFIG_COMPAT 2122 #ifdef CONFIG_COMPAT
2123 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_ 2123 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2124 int, nr, compat_uptr_t 2124 int, nr, compat_uptr_t __user *, iocbpp)
2125 { 2125 {
2126 struct kioctx *ctx; 2126 struct kioctx *ctx;
2127 long ret = 0; 2127 long ret = 0;
2128 int i = 0; 2128 int i = 0;
2129 struct blk_plug plug; 2129 struct blk_plug plug;
2130 2130
2131 if (unlikely(nr < 0)) 2131 if (unlikely(nr < 0))
2132 return -EINVAL; 2132 return -EINVAL;
2133 2133
2134 ctx = lookup_ioctx(ctx_id); 2134 ctx = lookup_ioctx(ctx_id);
2135 if (unlikely(!ctx)) { 2135 if (unlikely(!ctx)) {
2136 pr_debug("EINVAL: invalid con 2136 pr_debug("EINVAL: invalid context id\n");
2137 return -EINVAL; 2137 return -EINVAL;
2138 } 2138 }
2139 2139
2140 if (nr > ctx->nr_events) 2140 if (nr > ctx->nr_events)
2141 nr = ctx->nr_events; 2141 nr = ctx->nr_events;
2142 2142
2143 if (nr > AIO_PLUG_THRESHOLD) 2143 if (nr > AIO_PLUG_THRESHOLD)
2144 blk_start_plug(&plug); 2144 blk_start_plug(&plug);
2145 for (i = 0; i < nr; i++) { 2145 for (i = 0; i < nr; i++) {
2146 compat_uptr_t user_iocb; 2146 compat_uptr_t user_iocb;
2147 2147
2148 if (unlikely(get_user(user_io 2148 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2149 ret = -EFAULT; 2149 ret = -EFAULT;
2150 break; 2150 break;
2151 } 2151 }
2152 2152
2153 ret = io_submit_one(ctx, comp 2153 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2154 if (ret) 2154 if (ret)
2155 break; 2155 break;
2156 } 2156 }
2157 if (nr > AIO_PLUG_THRESHOLD) 2157 if (nr > AIO_PLUG_THRESHOLD)
2158 blk_finish_plug(&plug); 2158 blk_finish_plug(&plug);
2159 2159
2160 percpu_ref_put(&ctx->users); 2160 percpu_ref_put(&ctx->users);
2161 return i ? i : ret; 2161 return i ? i : ret;
2162 } 2162 }
2163 #endif 2163 #endif
2164 2164
2165 /* sys_io_cancel: 2165 /* sys_io_cancel:
2166 * Attempts to cancel an iocb previously 2166 * Attempts to cancel an iocb previously passed to io_submit. If
2167 * the operation is successfully cancell 2167 * the operation is successfully cancelled, the resulting event is
2168 * copied into the memory pointed to by 2168 * copied into the memory pointed to by result without being placed
2169 * into the completion queue and 0 is re 2169 * into the completion queue and 0 is returned. May fail with
2170 * -EFAULT if any of the data structures 2170 * -EFAULT if any of the data structures pointed to are invalid.
2171 * May fail with -EINVAL if aio_context 2171 * May fail with -EINVAL if aio_context specified by ctx_id is
2172 * invalid. May fail with -EAGAIN if th 2172 * invalid. May fail with -EAGAIN if the iocb specified was not
2173 * cancelled. Will fail with -ENOSYS if 2173 * cancelled. Will fail with -ENOSYS if not implemented.
2174 */ 2174 */
2175 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx 2175 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2176 struct io_event __user *, res 2176 struct io_event __user *, result)
2177 { 2177 {
2178 struct kioctx *ctx; 2178 struct kioctx *ctx;
2179 struct aio_kiocb *kiocb; 2179 struct aio_kiocb *kiocb;
2180 int ret = -EINVAL; 2180 int ret = -EINVAL;
2181 u32 key; 2181 u32 key;
2182 u64 obj = (u64)(unsigned long)iocb; 2182 u64 obj = (u64)(unsigned long)iocb;
2183 2183
2184 if (unlikely(get_user(key, &iocb->aio 2184 if (unlikely(get_user(key, &iocb->aio_key)))
2185 return -EFAULT; 2185 return -EFAULT;
2186 if (unlikely(key != KIOCB_KEY)) 2186 if (unlikely(key != KIOCB_KEY))
2187 return -EINVAL; 2187 return -EINVAL;
2188 2188
2189 ctx = lookup_ioctx(ctx_id); 2189 ctx = lookup_ioctx(ctx_id);
2190 if (unlikely(!ctx)) 2190 if (unlikely(!ctx))
2191 return -EINVAL; 2191 return -EINVAL;
2192 2192
2193 spin_lock_irq(&ctx->ctx_lock); 2193 spin_lock_irq(&ctx->ctx_lock);
2194 /* TODO: use a hash or array, this su 2194 /* TODO: use a hash or array, this sucks. */
2195 list_for_each_entry(kiocb, &ctx->acti 2195 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2196 if (kiocb->ki_res.obj == obj) 2196 if (kiocb->ki_res.obj == obj) {
2197 ret = kiocb->ki_cance 2197 ret = kiocb->ki_cancel(&kiocb->rw);
2198 list_del_init(&kiocb- 2198 list_del_init(&kiocb->ki_list);
2199 break; 2199 break;
2200 } 2200 }
2201 } 2201 }
2202 spin_unlock_irq(&ctx->ctx_lock); 2202 spin_unlock_irq(&ctx->ctx_lock);
2203 2203
2204 if (!ret) { 2204 if (!ret) {
2205 /* 2205 /*
2206 * The result argument is no 2206 * The result argument is no longer used - the io_event is
2207 * always delivered via the r 2207 * always delivered via the ring buffer. -EINPROGRESS indicates
2208 * cancellation is progress: 2208 * cancellation is progress:
2209 */ 2209 */
2210 ret = -EINPROGRESS; 2210 ret = -EINPROGRESS;
2211 } 2211 }
2212 2212
2213 percpu_ref_put(&ctx->users); 2213 percpu_ref_put(&ctx->users);
2214 2214
2215 return ret; 2215 return ret;
2216 } 2216 }
2217 2217
2218 static long do_io_getevents(aio_context_t ctx 2218 static long do_io_getevents(aio_context_t ctx_id,
2219 long min_nr, 2219 long min_nr,
2220 long nr, 2220 long nr,
2221 struct io_event __user *event 2221 struct io_event __user *events,
2222 struct timespec64 *ts) 2222 struct timespec64 *ts)
2223 { 2223 {
2224 ktime_t until = ts ? timespec64_to_kt 2224 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2225 struct kioctx *ioctx = lookup_ioctx(c 2225 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2226 long ret = -EINVAL; 2226 long ret = -EINVAL;
2227 2227
2228 if (likely(ioctx)) { 2228 if (likely(ioctx)) {
2229 if (likely(min_nr <= nr && mi 2229 if (likely(min_nr <= nr && min_nr >= 0))
2230 ret = read_events(ioc 2230 ret = read_events(ioctx, min_nr, nr, events, until);
2231 percpu_ref_put(&ioctx->users) 2231 percpu_ref_put(&ioctx->users);
2232 } 2232 }
2233 2233
2234 return ret; 2234 return ret;
2235 } 2235 }
2236 2236
2237 /* io_getevents: 2237 /* io_getevents:
2238 * Attempts to read at least min_nr even 2238 * Attempts to read at least min_nr events and up to nr events from
2239 * the completion queue for the aio_cont 2239 * the completion queue for the aio_context specified by ctx_id. If
2240 * it succeeds, the number of read event 2240 * it succeeds, the number of read events is returned. May fail with
2241 * -EINVAL if ctx_id is invalid, if min_ 2241 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2242 * out of range, if timeout is out of ra 2242 * out of range, if timeout is out of range. May fail with -EFAULT
2243 * if any of the memory specified is inv 2243 * if any of the memory specified is invalid. May return 0 or
2244 * < min_nr if the timeout specified by 2244 * < min_nr if the timeout specified by timeout has elapsed
2245 * before sufficient events are availabl 2245 * before sufficient events are available, where timeout == NULL
2246 * specifies an infinite timeout. Note t 2246 * specifies an infinite timeout. Note that the timeout pointed to by
2247 * timeout is relative. Will fail with 2247 * timeout is relative. Will fail with -ENOSYS if not implemented.
2248 */ 2248 */
2249 #ifdef CONFIG_64BIT 2249 #ifdef CONFIG_64BIT
2250 2250
2251 SYSCALL_DEFINE5(io_getevents, aio_context_t, 2251 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2252 long, min_nr, 2252 long, min_nr,
2253 long, nr, 2253 long, nr,
2254 struct io_event __user *, eve 2254 struct io_event __user *, events,
2255 struct __kernel_timespec __us 2255 struct __kernel_timespec __user *, timeout)
2256 { 2256 {
2257 struct timespec64 ts; 2257 struct timespec64 ts;
2258 int ret; 2258 int ret;
2259 2259
2260 if (timeout && unlikely(get_timespec6 2260 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2261 return -EFAULT; 2261 return -EFAULT;
2262 2262
2263 ret = do_io_getevents(ctx_id, min_nr, 2263 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2264 if (!ret && signal_pending(current)) 2264 if (!ret && signal_pending(current))
2265 ret = -EINTR; 2265 ret = -EINTR;
2266 return ret; 2266 return ret;
2267 } 2267 }
2268 2268
2269 #endif 2269 #endif
2270 2270
2271 struct __aio_sigset { 2271 struct __aio_sigset {
2272 const sigset_t __user *sigmask; 2272 const sigset_t __user *sigmask;
2273 size_t sigsetsize; 2273 size_t sigsetsize;
2274 }; 2274 };
2275 2275
2276 SYSCALL_DEFINE6(io_pgetevents, 2276 SYSCALL_DEFINE6(io_pgetevents,
2277 aio_context_t, ctx_id, 2277 aio_context_t, ctx_id,
2278 long, min_nr, 2278 long, min_nr,
2279 long, nr, 2279 long, nr,
2280 struct io_event __user *, eve 2280 struct io_event __user *, events,
2281 struct __kernel_timespec __us 2281 struct __kernel_timespec __user *, timeout,
2282 const struct __aio_sigset __u 2282 const struct __aio_sigset __user *, usig)
2283 { 2283 {
2284 struct __aio_sigset ksig = { NULL 2284 struct __aio_sigset ksig = { NULL, };
2285 struct timespec64 ts; 2285 struct timespec64 ts;
2286 bool interrupted; 2286 bool interrupted;
2287 int ret; 2287 int ret;
2288 2288
2289 if (timeout && unlikely(get_timespec6 2289 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2290 return -EFAULT; 2290 return -EFAULT;
2291 2291
2292 if (usig && copy_from_user(&ksig, usi 2292 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2293 return -EFAULT; 2293 return -EFAULT;
2294 2294
2295 ret = set_user_sigmask(ksig.sigmask, 2295 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2296 if (ret) 2296 if (ret)
2297 return ret; 2297 return ret;
2298 2298
2299 ret = do_io_getevents(ctx_id, min_nr, 2299 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2300 2300
2301 interrupted = signal_pending(current) 2301 interrupted = signal_pending(current);
2302 restore_saved_sigmask_unless(interrup 2302 restore_saved_sigmask_unless(interrupted);
2303 if (interrupted && !ret) 2303 if (interrupted && !ret)
2304 ret = -ERESTARTNOHAND; 2304 ret = -ERESTARTNOHAND;
2305 2305
2306 return ret; 2306 return ret;
2307 } 2307 }
2308 2308
2309 #if defined(CONFIG_COMPAT_32BIT_TIME) && !def 2309 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2310 2310
2311 SYSCALL_DEFINE6(io_pgetevents_time32, 2311 SYSCALL_DEFINE6(io_pgetevents_time32,
2312 aio_context_t, ctx_id, 2312 aio_context_t, ctx_id,
2313 long, min_nr, 2313 long, min_nr,
2314 long, nr, 2314 long, nr,
2315 struct io_event __user *, eve 2315 struct io_event __user *, events,
2316 struct old_timespec32 __user 2316 struct old_timespec32 __user *, timeout,
2317 const struct __aio_sigset __u 2317 const struct __aio_sigset __user *, usig)
2318 { 2318 {
2319 struct __aio_sigset ksig = { NULL 2319 struct __aio_sigset ksig = { NULL, };
2320 struct timespec64 ts; 2320 struct timespec64 ts;
2321 bool interrupted; 2321 bool interrupted;
2322 int ret; 2322 int ret;
2323 2323
2324 if (timeout && unlikely(get_old_times 2324 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2325 return -EFAULT; 2325 return -EFAULT;
2326 2326
2327 if (usig && copy_from_user(&ksig, usi 2327 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2328 return -EFAULT; 2328 return -EFAULT;
2329 2329
2330 2330
2331 ret = set_user_sigmask(ksig.sigmask, 2331 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2332 if (ret) 2332 if (ret)
2333 return ret; 2333 return ret;
2334 2334
2335 ret = do_io_getevents(ctx_id, min_nr, 2335 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2336 2336
2337 interrupted = signal_pending(current) 2337 interrupted = signal_pending(current);
2338 restore_saved_sigmask_unless(interrup 2338 restore_saved_sigmask_unless(interrupted);
2339 if (interrupted && !ret) 2339 if (interrupted && !ret)
2340 ret = -ERESTARTNOHAND; 2340 ret = -ERESTARTNOHAND;
2341 2341
2342 return ret; 2342 return ret;
2343 } 2343 }
2344 2344
2345 #endif 2345 #endif
2346 2346
2347 #if defined(CONFIG_COMPAT_32BIT_TIME) 2347 #if defined(CONFIG_COMPAT_32BIT_TIME)
2348 2348
2349 SYSCALL_DEFINE5(io_getevents_time32, __u32, c 2349 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2350 __s32, min_nr, 2350 __s32, min_nr,
2351 __s32, nr, 2351 __s32, nr,
2352 struct io_event __user *, eve 2352 struct io_event __user *, events,
2353 struct old_timespec32 __user 2353 struct old_timespec32 __user *, timeout)
2354 { 2354 {
2355 struct timespec64 t; 2355 struct timespec64 t;
2356 int ret; 2356 int ret;
2357 2357
2358 if (timeout && get_old_timespec32(&t, 2358 if (timeout && get_old_timespec32(&t, timeout))
2359 return -EFAULT; 2359 return -EFAULT;
2360 2360
2361 ret = do_io_getevents(ctx_id, min_nr, 2361 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2362 if (!ret && signal_pending(current)) 2362 if (!ret && signal_pending(current))
2363 ret = -EINTR; 2363 ret = -EINTR;
2364 return ret; 2364 return ret;
2365 } 2365 }
2366 2366
2367 #endif 2367 #endif
2368 2368
2369 #ifdef CONFIG_COMPAT 2369 #ifdef CONFIG_COMPAT
2370 2370
2371 struct __compat_aio_sigset { 2371 struct __compat_aio_sigset {
2372 compat_uptr_t sigmask; 2372 compat_uptr_t sigmask;
2373 compat_size_t sigsetsize; 2373 compat_size_t sigsetsize;
2374 }; 2374 };
2375 2375
2376 #if defined(CONFIG_COMPAT_32BIT_TIME) 2376 #if defined(CONFIG_COMPAT_32BIT_TIME)
2377 2377
2378 COMPAT_SYSCALL_DEFINE6(io_pgetevents, 2378 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2379 compat_aio_context_t, ctx_id, 2379 compat_aio_context_t, ctx_id,
2380 compat_long_t, min_nr, 2380 compat_long_t, min_nr,
2381 compat_long_t, nr, 2381 compat_long_t, nr,
2382 struct io_event __user *, eve 2382 struct io_event __user *, events,
2383 struct old_timespec32 __user 2383 struct old_timespec32 __user *, timeout,
2384 const struct __compat_aio_sig 2384 const struct __compat_aio_sigset __user *, usig)
2385 { 2385 {
2386 struct __compat_aio_sigset ksig = { 0 2386 struct __compat_aio_sigset ksig = { 0, };
2387 struct timespec64 t; 2387 struct timespec64 t;
2388 bool interrupted; 2388 bool interrupted;
2389 int ret; 2389 int ret;
2390 2390
2391 if (timeout && get_old_timespec32(&t, 2391 if (timeout && get_old_timespec32(&t, timeout))
2392 return -EFAULT; 2392 return -EFAULT;
2393 2393
2394 if (usig && copy_from_user(&ksig, usi 2394 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2395 return -EFAULT; 2395 return -EFAULT;
2396 2396
2397 ret = set_compat_user_sigmask(compat_ 2397 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2398 if (ret) 2398 if (ret)
2399 return ret; 2399 return ret;
2400 2400
2401 ret = do_io_getevents(ctx_id, min_nr, 2401 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2402 2402
2403 interrupted = signal_pending(current) 2403 interrupted = signal_pending(current);
2404 restore_saved_sigmask_unless(interrup 2404 restore_saved_sigmask_unless(interrupted);
2405 if (interrupted && !ret) 2405 if (interrupted && !ret)
2406 ret = -ERESTARTNOHAND; 2406 ret = -ERESTARTNOHAND;
2407 2407
2408 return ret; 2408 return ret;
2409 } 2409 }
2410 2410
2411 #endif 2411 #endif
2412 2412
2413 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64, 2413 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2414 compat_aio_context_t, ctx_id, 2414 compat_aio_context_t, ctx_id,
2415 compat_long_t, min_nr, 2415 compat_long_t, min_nr,
2416 compat_long_t, nr, 2416 compat_long_t, nr,
2417 struct io_event __user *, eve 2417 struct io_event __user *, events,
2418 struct __kernel_timespec __us 2418 struct __kernel_timespec __user *, timeout,
2419 const struct __compat_aio_sig 2419 const struct __compat_aio_sigset __user *, usig)
2420 { 2420 {
2421 struct __compat_aio_sigset ksig = { 0 2421 struct __compat_aio_sigset ksig = { 0, };
2422 struct timespec64 t; 2422 struct timespec64 t;
2423 bool interrupted; 2423 bool interrupted;
2424 int ret; 2424 int ret;
2425 2425
2426 if (timeout && get_timespec64(&t, tim 2426 if (timeout && get_timespec64(&t, timeout))
2427 return -EFAULT; 2427 return -EFAULT;
2428 2428
2429 if (usig && copy_from_user(&ksig, usi 2429 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2430 return -EFAULT; 2430 return -EFAULT;
2431 2431
2432 ret = set_compat_user_sigmask(compat_ 2432 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2433 if (ret) 2433 if (ret)
2434 return ret; 2434 return ret;
2435 2435
2436 ret = do_io_getevents(ctx_id, min_nr, 2436 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2437 2437
2438 interrupted = signal_pending(current) 2438 interrupted = signal_pending(current);
2439 restore_saved_sigmask_unless(interrup 2439 restore_saved_sigmask_unless(interrupted);
2440 if (interrupted && !ret) 2440 if (interrupted && !ret)
2441 ret = -ERESTARTNOHAND; 2441 ret = -ERESTARTNOHAND;
2442 2442
2443 return ret; 2443 return ret;
2444 } 2444 }
2445 #endif 2445 #endif
2446 2446
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.