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