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