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. <<
9 * 8 *
10 * See ../COPYING for licensing terms. 9 * See ../COPYING for licensing terms.
11 */ 10 */
12 #define pr_fmt(fmt) "%s: " fmt, __func__ 11 #define pr_fmt(fmt) "%s: " fmt, __func__
13 12
14 #include <linux/kernel.h> 13 #include <linux/kernel.h>
15 #include <linux/init.h> 14 #include <linux/init.h>
16 #include <linux/errno.h> 15 #include <linux/errno.h>
17 #include <linux/time.h> 16 #include <linux/time.h>
18 #include <linux/aio_abi.h> 17 #include <linux/aio_abi.h>
19 #include <linux/export.h> 18 #include <linux/export.h>
20 #include <linux/syscalls.h> 19 #include <linux/syscalls.h>
21 #include <linux/backing-dev.h> 20 #include <linux/backing-dev.h>
22 #include <linux/refcount.h> <<
23 #include <linux/uio.h> 21 #include <linux/uio.h>
24 22
25 #include <linux/sched/signal.h> !! 23 #include <linux/sched.h>
26 #include <linux/fs.h> 24 #include <linux/fs.h>
27 #include <linux/file.h> 25 #include <linux/file.h>
28 #include <linux/mm.h> 26 #include <linux/mm.h>
29 #include <linux/mman.h> 27 #include <linux/mman.h>
30 #include <linux/percpu.h> !! 28 #include <linux/mmu_context.h>
31 #include <linux/slab.h> 29 #include <linux/slab.h>
32 #include <linux/timer.h> 30 #include <linux/timer.h>
33 #include <linux/aio.h> 31 #include <linux/aio.h>
34 #include <linux/highmem.h> 32 #include <linux/highmem.h>
35 #include <linux/workqueue.h> 33 #include <linux/workqueue.h>
36 #include <linux/security.h> 34 #include <linux/security.h>
37 #include <linux/eventfd.h> 35 #include <linux/eventfd.h>
38 #include <linux/blkdev.h> 36 #include <linux/blkdev.h>
39 #include <linux/compat.h> 37 #include <linux/compat.h>
40 #include <linux/migrate.h> <<
41 #include <linux/ramfs.h> <<
42 #include <linux/percpu-refcount.h> <<
43 #include <linux/mount.h> <<
44 #include <linux/pseudo_fs.h> <<
45 38
46 #include <linux/uaccess.h> !! 39 #include <asm/kmap_types.h>
47 #include <linux/nospec.h> !! 40 #include <asm/uaccess.h>
48 <<
49 #include "internal.h" <<
50 <<
51 #define KIOCB_KEY 0 <<
52 41
53 #define AIO_RING_MAGIC 0xa10a 42 #define AIO_RING_MAGIC 0xa10a10a1
54 #define AIO_RING_COMPAT_FEATURES 1 43 #define AIO_RING_COMPAT_FEATURES 1
55 #define AIO_RING_INCOMPAT_FEATURES 0 44 #define AIO_RING_INCOMPAT_FEATURES 0
56 struct aio_ring { 45 struct aio_ring {
57 unsigned id; /* kernel inte 46 unsigned id; /* kernel internal index number */
58 unsigned nr; /* number of i 47 unsigned nr; /* number of io_events */
59 unsigned head; /* Written to !! 48 unsigned head;
60 * mutex by ai <<
61 unsigned tail; 49 unsigned tail;
62 50
63 unsigned magic; 51 unsigned magic;
64 unsigned compat_features; 52 unsigned compat_features;
65 unsigned incompat_features; 53 unsigned incompat_features;
66 unsigned header_length; /* siz 54 unsigned header_length; /* size of aio_ring */
67 55
68 56
69 struct io_event io_events[]; !! 57 struct io_event io_events[0];
70 }; /* 128 bytes + ring size */ 58 }; /* 128 bytes + ring size */
71 59
72 /* <<
73 * Plugging is meant to work with larger batch <<
74 * have more than the below, then don't bother <<
75 */ <<
76 #define AIO_PLUG_THRESHOLD 2 <<
77 <<
78 #define AIO_RING_PAGES 8 60 #define AIO_RING_PAGES 8
79 61
80 struct kioctx_table { <<
81 struct rcu_head rcu; <<
82 unsigned nr; <<
83 struct kioctx __rcu *table[] __cou <<
84 }; <<
85 <<
86 struct kioctx_cpu { <<
87 unsigned reqs_available <<
88 }; <<
89 <<
90 struct ctx_rq_wait { <<
91 struct completion comp; <<
92 atomic_t count; <<
93 }; <<
94 <<
95 struct kioctx { 62 struct kioctx {
96 struct percpu_ref users; !! 63 atomic_t users;
97 atomic_t dead; 64 atomic_t dead;
98 65
99 struct percpu_ref reqs; !! 66 /* This needs improving */
100 <<
101 unsigned long user_id; 67 unsigned long user_id;
102 !! 68 struct hlist_node list;
103 struct kioctx_cpu __percpu *cpu; <<
104 69
105 /* 70 /*
106 * For percpu reqs_available, number o <<
107 * counter at a time: <<
108 */ <<
109 unsigned req_batch; <<
110 /* <<
111 * This is what userspace passed to io 71 * This is what userspace passed to io_setup(), it's not used for
112 * anything but counting against the g 72 * anything but counting against the global max_reqs quota.
113 * 73 *
114 * The real limit is nr_events - 1, wh 74 * The real limit is nr_events - 1, which will be larger (see
115 * aio_setup_ring()) 75 * aio_setup_ring())
116 */ 76 */
117 unsigned max_reqs; 77 unsigned max_reqs;
118 78
119 /* Size of ringbuffer, in units of str 79 /* Size of ringbuffer, in units of struct io_event */
120 unsigned nr_events; 80 unsigned nr_events;
121 81
122 unsigned long mmap_base; 82 unsigned long mmap_base;
123 unsigned long mmap_size; 83 unsigned long mmap_size;
124 84
125 struct folio **ring_folios; !! 85 struct page **ring_pages;
126 long nr_pages; 86 long nr_pages;
127 87
128 struct rcu_work free_rwork; !! 88 struct rcu_head rcu_head;
129 !! 89 struct work_struct rcu_work;
130 /* <<
131 * signals when all in-flight requests <<
132 */ <<
133 struct ctx_rq_wait *rq_wait; <<
134 90
135 struct { 91 struct {
136 /* !! 92 atomic_t reqs_active;
137 * This counts the number of a <<
138 * so we avoid overflowing it: <<
139 * when allocating a kiocb and <<
140 * io_event is pulled off the <<
141 * <<
142 * We batch accesses to it wit <<
143 */ <<
144 atomic_t reqs_available <<
145 } ____cacheline_aligned_in_smp; 93 } ____cacheline_aligned_in_smp;
146 94
147 struct { 95 struct {
148 spinlock_t ctx_lock; 96 spinlock_t ctx_lock;
149 struct list_head active_reqs; 97 struct list_head active_reqs; /* used for cancellation */
150 } ____cacheline_aligned_in_smp; 98 } ____cacheline_aligned_in_smp;
151 99
152 struct { 100 struct {
153 struct mutex ring_lock; 101 struct mutex ring_lock;
154 wait_queue_head_t wait; 102 wait_queue_head_t wait;
155 } ____cacheline_aligned_in_smp; 103 } ____cacheline_aligned_in_smp;
156 104
157 struct { 105 struct {
158 unsigned tail; 106 unsigned tail;
159 unsigned completed_even <<
160 spinlock_t completion_loc 107 spinlock_t completion_lock;
161 } ____cacheline_aligned_in_smp; 108 } ____cacheline_aligned_in_smp;
162 109
163 struct folio *internal_foli !! 110 struct page *internal_pages[AIO_RING_PAGES];
164 struct file *aio_ring_file <<
165 <<
166 unsigned id; <<
167 }; <<
168 <<
169 /* <<
170 * First field must be the file pointer in all <<
171 * iocb unions! See also 'struct kiocb' in <li <<
172 */ <<
173 struct fsync_iocb { <<
174 struct file *file; <<
175 struct work_struct work; <<
176 bool datasync; <<
177 struct cred *creds; <<
178 }; <<
179 <<
180 struct poll_iocb { <<
181 struct file *file; <<
182 struct wait_queue_head *head; <<
183 __poll_t events; <<
184 bool cancelled; <<
185 bool work_scheduled <<
186 bool work_need_resc <<
187 struct wait_queue_entry wait; <<
188 struct work_struct work; <<
189 }; <<
190 <<
191 /* <<
192 * NOTE! Each of the iocb union members has th <<
193 * as the first entry in their struct definiti <<
194 * access the file pointer through any of the <<
195 * or directly as just 'ki_filp' in this struc <<
196 */ <<
197 struct aio_kiocb { <<
198 union { <<
199 struct file *ki_fi <<
200 struct kiocb rw; <<
201 struct fsync_iocb fsync; <<
202 struct poll_iocb poll; <<
203 }; <<
204 <<
205 struct kioctx *ki_ctx; <<
206 kiocb_cancel_fn *ki_cancel; <<
207 <<
208 struct io_event ki_res; <<
209 <<
210 struct list_head ki_list; <<
211 <<
212 refcount_t ki_refcnt; <<
213 <<
214 /* <<
215 * If the aio_resfd field of the users <<
216 * this is the underlying eventfd cont <<
217 */ <<
218 struct eventfd_ctx *ki_eventfd; <<
219 }; 111 };
220 112
221 /*------ sysctl variables----*/ 113 /*------ sysctl variables----*/
222 static DEFINE_SPINLOCK(aio_nr_lock); 114 static DEFINE_SPINLOCK(aio_nr_lock);
223 static unsigned long aio_nr; /* cur !! 115 unsigned long aio_nr; /* current system wide number of aio requests */
224 static unsigned long aio_max_nr = 0x10000; /* !! 116 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/ 117 /*----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 <<
251 118
252 static struct kmem_cache *kiocb_cachep; 119 static struct kmem_cache *kiocb_cachep;
253 static struct kmem_cache *kioctx_cachep 120 static struct kmem_cache *kioctx_cachep;
254 121
255 static struct vfsmount *aio_mnt; <<
256 <<
257 static const struct file_operations aio_ring_f <<
258 static const struct address_space_operations a <<
259 <<
260 static struct file *aio_private_file(struct ki <<
261 { <<
262 struct file *file; <<
263 struct inode *inode = alloc_anon_inode <<
264 if (IS_ERR(inode)) <<
265 return ERR_CAST(inode); <<
266 <<
267 inode->i_mapping->a_ops = &aio_ctx_aop <<
268 inode->i_mapping->i_private_data = ctx <<
269 inode->i_size = PAGE_SIZE * nr_pages; <<
270 <<
271 file = alloc_file_pseudo(inode, aio_mn <<
272 O_RDWR, &aio_r <<
273 if (IS_ERR(file)) <<
274 iput(inode); <<
275 return file; <<
276 } <<
277 <<
278 static int aio_init_fs_context(struct fs_conte <<
279 { <<
280 if (!init_pseudo(fc, AIO_RING_MAGIC)) <<
281 return -ENOMEM; <<
282 fc->s_iflags |= SB_I_NOEXEC; <<
283 return 0; <<
284 } <<
285 <<
286 /* aio_setup 122 /* aio_setup
287 * Creates the slab caches used by the ai 123 * Creates the slab caches used by the aio routines, panic on
288 * failure as this is done early during t 124 * failure as this is done early during the boot sequence.
289 */ 125 */
290 static int __init aio_setup(void) 126 static int __init aio_setup(void)
291 { 127 {
292 static struct file_system_type aio_fs !! 128 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
293 .name = "aio", <<
294 .init_fs_context = aio_init_fs <<
295 .kill_sb = kill_anon_su <<
296 }; <<
297 aio_mnt = kern_mount(&aio_fs); <<
298 if (IS_ERR(aio_mnt)) <<
299 panic("Failed to create aio fs <<
300 <<
301 kiocb_cachep = KMEM_CACHE(aio_kiocb, S <<
302 kioctx_cachep = KMEM_CACHE(kioctx,SLAB 129 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
303 aio_sysctl_init(); <<
304 return 0; <<
305 } <<
306 __initcall(aio_setup); <<
307 <<
308 static void put_aio_ring_file(struct kioctx *c <<
309 { <<
310 struct file *aio_ring_file = ctx->aio_ <<
311 struct address_space *i_mapping; <<
312 130
313 if (aio_ring_file) { !! 131 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
314 truncate_setsize(file_inode(ai <<
315 132
316 /* Prevent further access to t !! 133 return 0;
317 i_mapping = aio_ring_file->f_m <<
318 spin_lock(&i_mapping->i_privat <<
319 i_mapping->i_private_data = NU <<
320 ctx->aio_ring_file = NULL; <<
321 spin_unlock(&i_mapping->i_priv <<
322 <<
323 fput(aio_ring_file); <<
324 } <<
325 } 134 }
>> 135 __initcall(aio_setup);
326 136
327 static void aio_free_ring(struct kioctx *ctx) 137 static void aio_free_ring(struct kioctx *ctx)
328 { 138 {
329 int i; !! 139 long i;
330 <<
331 /* Disconnect the kiotx from the ring <<
332 * accesses to the kioctx from page mi <<
333 */ <<
334 put_aio_ring_file(ctx); <<
335 140
336 for (i = 0; i < ctx->nr_pages; i++) { !! 141 for (i = 0; i < ctx->nr_pages; i++)
337 struct folio *folio = ctx->rin !! 142 put_page(ctx->ring_pages[i]);
338 143
339 if (!folio) !! 144 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
340 continue; !! 145 kfree(ctx->ring_pages);
341 <<
342 pr_debug("pid(%d) [%d] folio-> <<
343 folio_ref_count(folio <<
344 ctx->ring_folios[i] = NULL; <<
345 folio_put(folio); <<
346 } <<
347 <<
348 if (ctx->ring_folios && ctx->ring_foli <<
349 kfree(ctx->ring_folios); <<
350 ctx->ring_folios = NULL; <<
351 } <<
352 } 146 }
353 147
354 static int aio_ring_mremap(struct vm_area_stru !! 148 static int aio_setup_ring(struct kioctx *ctx)
355 { <<
356 struct file *file = vma->vm_file; <<
357 struct mm_struct *mm = vma->vm_mm; <<
358 struct kioctx_table *table; <<
359 int i, res = -EINVAL; <<
360 <<
361 spin_lock(&mm->ioctx_lock); <<
362 rcu_read_lock(); <<
363 table = rcu_dereference(mm->ioctx_tabl <<
364 if (!table) <<
365 goto out_unlock; <<
366 <<
367 for (i = 0; i < table->nr; i++) { <<
368 struct kioctx *ctx; <<
369 <<
370 ctx = rcu_dereference(table->t <<
371 if (ctx && ctx->aio_ring_file <<
372 if (!atomic_read(&ctx- <<
373 ctx->user_id = <<
374 res = 0; <<
375 } <<
376 break; <<
377 } <<
378 } <<
379 <<
380 out_unlock: <<
381 rcu_read_unlock(); <<
382 spin_unlock(&mm->ioctx_lock); <<
383 return res; <<
384 } <<
385 <<
386 static const struct vm_operations_struct aio_r <<
387 .mremap = aio_ring_mremap, <<
388 #if IS_ENABLED(CONFIG_MMU) <<
389 .fault = filemap_fault, <<
390 .map_pages = filemap_map_pages, <<
391 .page_mkwrite = filemap_page_mkwrite <<
392 #endif <<
393 }; <<
394 <<
395 static int aio_ring_mmap(struct file *file, st <<
396 { <<
397 vm_flags_set(vma, VM_DONTEXPAND); <<
398 vma->vm_ops = &aio_ring_vm_ops; <<
399 return 0; <<
400 } <<
401 <<
402 static const struct file_operations aio_ring_f <<
403 .mmap = aio_ring_mmap, <<
404 }; <<
405 <<
406 #if IS_ENABLED(CONFIG_MIGRATION) <<
407 static int aio_migrate_folio(struct address_sp <<
408 struct folio *src, enu <<
409 { <<
410 struct kioctx *ctx; <<
411 unsigned long flags; <<
412 pgoff_t idx; <<
413 int rc = 0; <<
414 <<
415 /* mapping->i_private_lock here protec <<
416 spin_lock(&mapping->i_private_lock); <<
417 ctx = mapping->i_private_data; <<
418 if (!ctx) { <<
419 rc = -EINVAL; <<
420 goto out; <<
421 } <<
422 <<
423 /* The ring_lock mutex. The prevents <<
424 * to the ring's head, and prevents pa <<
425 * a partially initialized kiotx. <<
426 */ <<
427 if (!mutex_trylock(&ctx->ring_lock)) { <<
428 rc = -EAGAIN; <<
429 goto out; <<
430 } <<
431 <<
432 idx = src->index; <<
433 if (idx < (pgoff_t)ctx->nr_pages) { <<
434 /* Make sure the old folio has <<
435 if (ctx->ring_folios[idx] != s <<
436 rc = -EAGAIN; <<
437 } else <<
438 rc = -EINVAL; <<
439 <<
440 if (rc != 0) <<
441 goto out_unlock; <<
442 <<
443 /* Writeback must be complete */ <<
444 BUG_ON(folio_test_writeback(src)); <<
445 folio_get(dst); <<
446 <<
447 rc = folio_migrate_mapping(mapping, ds <<
448 if (rc != MIGRATEPAGE_SUCCESS) { <<
449 folio_put(dst); <<
450 goto out_unlock; <<
451 } <<
452 <<
453 /* Take completion_lock to prevent oth <<
454 * while the old folio is copied to th <<
455 * events from being lost. <<
456 */ <<
457 spin_lock_irqsave(&ctx->completion_loc <<
458 folio_copy(dst, src); <<
459 folio_migrate_flags(dst, src); <<
460 BUG_ON(ctx->ring_folios[idx] != src); <<
461 ctx->ring_folios[idx] = dst; <<
462 spin_unlock_irqrestore(&ctx->completio <<
463 <<
464 /* The old folio is no longer accessib <<
465 folio_put(src); <<
466 <<
467 out_unlock: <<
468 mutex_unlock(&ctx->ring_lock); <<
469 out: <<
470 spin_unlock(&mapping->i_private_lock); <<
471 return rc; <<
472 } <<
473 #else <<
474 #define aio_migrate_folio NULL <<
475 #endif <<
476 <<
477 static const struct address_space_operations a <<
478 .dirty_folio = noop_dirty_folio, <<
479 .migrate_folio = aio_migrate_folio, <<
480 }; <<
481 <<
482 static int aio_setup_ring(struct kioctx *ctx, <<
483 { 149 {
484 struct aio_ring *ring; 150 struct aio_ring *ring;
>> 151 unsigned nr_events = ctx->max_reqs;
485 struct mm_struct *mm = current->mm; 152 struct mm_struct *mm = current->mm;
486 unsigned long size, unused; !! 153 unsigned long size, populate;
487 int nr_pages; 154 int nr_pages;
488 int i; <<
489 struct file *file; <<
490 155
491 /* Compensate for the ring buffer's he 156 /* Compensate for the ring buffer's head/tail overlap entry */
492 nr_events += 2; /* 1 is required, 2 fo 157 nr_events += 2; /* 1 is required, 2 for good luck */
493 158
494 size = sizeof(struct aio_ring); 159 size = sizeof(struct aio_ring);
495 size += sizeof(struct io_event) * nr_e 160 size += sizeof(struct io_event) * nr_events;
>> 161 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
496 162
497 nr_pages = PFN_UP(size); <<
498 if (nr_pages < 0) 163 if (nr_pages < 0)
499 return -EINVAL; 164 return -EINVAL;
500 165
501 file = aio_private_file(ctx, nr_pages) !! 166 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
502 if (IS_ERR(file)) { <<
503 ctx->aio_ring_file = NULL; <<
504 return -ENOMEM; <<
505 } <<
506 167
507 ctx->aio_ring_file = file; !! 168 ctx->nr_events = 0;
508 nr_events = (PAGE_SIZE * nr_pages - si !! 169 ctx->ring_pages = ctx->internal_pages;
509 / sizeof(struct io_eve <<
510 <<
511 ctx->ring_folios = ctx->internal_folio <<
512 if (nr_pages > AIO_RING_PAGES) { 170 if (nr_pages > AIO_RING_PAGES) {
513 ctx->ring_folios = kcalloc(nr_ !! 171 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
514 GFP !! 172 GFP_KERNEL);
515 if (!ctx->ring_folios) { !! 173 if (!ctx->ring_pages)
516 put_aio_ring_file(ctx) <<
517 return -ENOMEM; 174 return -ENOMEM;
518 } <<
519 } <<
520 <<
521 for (i = 0; i < nr_pages; i++) { <<
522 struct folio *folio; <<
523 <<
524 folio = __filemap_get_folio(fi <<
525 FG <<
526 GF <<
527 if (IS_ERR(folio)) <<
528 break; <<
529 <<
530 pr_debug("pid(%d) [%d] folio-> <<
531 folio_ref_count(folio <<
532 folio_end_read(folio, true); <<
533 <<
534 ctx->ring_folios[i] = folio; <<
535 } <<
536 ctx->nr_pages = i; <<
537 <<
538 if (unlikely(i != nr_pages)) { <<
539 aio_free_ring(ctx); <<
540 return -ENOMEM; <<
541 } 175 }
542 176
543 ctx->mmap_size = nr_pages * PAGE_SIZE; 177 ctx->mmap_size = nr_pages * PAGE_SIZE;
544 pr_debug("attempting mmap of %lu bytes 178 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
545 !! 179 down_write(&mm->mmap_sem);
546 if (mmap_write_lock_killable(mm)) { !! 180 ctx->mmap_base = do_mmap_pgoff(NULL, 0, ctx->mmap_size,
547 ctx->mmap_size = 0; !! 181 PROT_READ|PROT_WRITE,
548 aio_free_ring(ctx); !! 182 MAP_ANONYMOUS|MAP_PRIVATE, 0, &populate);
549 return -EINTR; <<
550 } <<
551 <<
552 ctx->mmap_base = do_mmap(ctx->aio_ring <<
553 PROT_READ | P <<
554 MAP_SHARED, 0 <<
555 mmap_write_unlock(mm); <<
556 if (IS_ERR((void *)ctx->mmap_base)) { 183 if (IS_ERR((void *)ctx->mmap_base)) {
>> 184 up_write(&mm->mmap_sem);
557 ctx->mmap_size = 0; 185 ctx->mmap_size = 0;
558 aio_free_ring(ctx); 186 aio_free_ring(ctx);
559 return -ENOMEM; !! 187 return -EAGAIN;
560 } 188 }
561 189
562 pr_debug("mmap address: 0x%08lx\n", ct 190 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
>> 191 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
>> 192 1, 0, ctx->ring_pages, NULL);
>> 193 up_write(&mm->mmap_sem);
>> 194
>> 195 if (unlikely(ctx->nr_pages != nr_pages)) {
>> 196 aio_free_ring(ctx);
>> 197 return -EAGAIN;
>> 198 }
>> 199 if (populate)
>> 200 mm_populate(ctx->mmap_base, populate);
563 201
564 ctx->user_id = ctx->mmap_base; 202 ctx->user_id = ctx->mmap_base;
565 ctx->nr_events = nr_events; /* trusted 203 ctx->nr_events = nr_events; /* trusted copy */
566 204
567 ring = folio_address(ctx->ring_folios[ !! 205 ring = kmap_atomic(ctx->ring_pages[0]);
568 ring->nr = nr_events; /* user copy * 206 ring->nr = nr_events; /* user copy */
569 ring->id = ~0U; !! 207 ring->id = ctx->user_id;
570 ring->head = ring->tail = 0; 208 ring->head = ring->tail = 0;
571 ring->magic = AIO_RING_MAGIC; 209 ring->magic = AIO_RING_MAGIC;
572 ring->compat_features = AIO_RING_COMPA 210 ring->compat_features = AIO_RING_COMPAT_FEATURES;
573 ring->incompat_features = AIO_RING_INC 211 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
574 ring->header_length = sizeof(struct ai 212 ring->header_length = sizeof(struct aio_ring);
575 flush_dcache_folio(ctx->ring_folios[0] !! 213 kunmap_atomic(ring);
>> 214 flush_dcache_page(ctx->ring_pages[0]);
576 215
577 return 0; 216 return 0;
578 } 217 }
579 218
580 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / s 219 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
581 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - 220 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
582 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PE 221 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
583 222
584 void kiocb_set_cancel_fn(struct kiocb *iocb, k !! 223 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
585 { 224 {
586 struct aio_kiocb *req; !! 225 struct kioctx *ctx = req->ki_ctx;
587 struct kioctx *ctx; <<
588 unsigned long flags; 226 unsigned long flags;
589 227
590 /* !! 228 spin_lock_irqsave(&ctx->ctx_lock, flags);
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_ <<
600 return; <<
601 229
602 ctx = req->ki_ctx; !! 230 if (!req->ki_list.next)
>> 231 list_add(&req->ki_list, &ctx->active_reqs);
603 232
604 spin_lock_irqsave(&ctx->ctx_lock, flag <<
605 list_add_tail(&req->ki_list, &ctx->act <<
606 req->ki_cancel = cancel; 233 req->ki_cancel = cancel;
>> 234
607 spin_unlock_irqrestore(&ctx->ctx_lock, 235 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
608 } 236 }
609 EXPORT_SYMBOL(kiocb_set_cancel_fn); 237 EXPORT_SYMBOL(kiocb_set_cancel_fn);
610 238
611 /* !! 239 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
612 * free_ioctx() should be RCU delayed to synch !! 240 struct io_event *res)
613 * protected lookup_ioctx() and also needs pro <<
614 * aio_free_ring(). Use rcu_work. <<
615 */ <<
616 static void free_ioctx(struct work_struct *wor <<
617 { 241 {
618 struct kioctx *ctx = container_of(to_r !! 242 kiocb_cancel_fn *old, *cancel;
619 free !! 243 int ret = -EINVAL;
620 pr_debug("freeing %p\n", ctx); <<
621 244
622 aio_free_ring(ctx); !! 245 /*
623 free_percpu(ctx->cpu); !! 246 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
624 percpu_ref_exit(&ctx->reqs); !! 247 * actually has a cancel function, hence the cmpxchg()
625 percpu_ref_exit(&ctx->users); !! 248 */
626 kmem_cache_free(kioctx_cachep, ctx); <<
627 } <<
628 249
629 static void free_ioctx_reqs(struct percpu_ref !! 250 cancel = ACCESS_ONCE(kiocb->ki_cancel);
630 { !! 251 do {
631 struct kioctx *ctx = container_of(ref, !! 252 if (!cancel || cancel == KIOCB_CANCELLED)
>> 253 return ret;
>> 254
>> 255 old = cancel;
>> 256 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
>> 257 } while (cancel != old);
>> 258
>> 259 atomic_inc(&kiocb->ki_users);
>> 260 spin_unlock_irq(&ctx->ctx_lock);
>> 261
>> 262 memset(res, 0, sizeof(*res));
>> 263 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
>> 264 res->data = kiocb->ki_user_data;
>> 265 ret = cancel(kiocb, res);
>> 266
>> 267 spin_lock_irq(&ctx->ctx_lock);
632 268
633 /* At this point we know that there ar !! 269 return ret;
634 if (ctx->rq_wait && atomic_dec_and_tes !! 270 }
635 complete(&ctx->rq_wait->comp); <<
636 271
637 /* Synchronize against RCU protected t !! 272 static void free_ioctx_rcu(struct rcu_head *head)
638 INIT_RCU_WORK(&ctx->free_rwork, free_i !! 273 {
639 queue_rcu_work(system_wq, &ctx->free_r !! 274 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
>> 275 kmem_cache_free(kioctx_cachep, ctx);
640 } 276 }
641 277
642 /* 278 /*
643 * When this function runs, the kioctx has bee 279 * When this function runs, the kioctx has been removed from the "hash table"
644 * and ctx->users has dropped to 0, so we know 280 * 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 281 * now it's safe to cancel any that need to be.
646 */ 282 */
647 static void free_ioctx_users(struct percpu_ref !! 283 static void free_ioctx(struct kioctx *ctx)
648 { 284 {
649 struct kioctx *ctx = container_of(ref, !! 285 struct aio_ring *ring;
650 struct aio_kiocb *req; !! 286 struct io_event res;
>> 287 struct kiocb *req;
>> 288 unsigned head, avail;
651 289
652 spin_lock_irq(&ctx->ctx_lock); 290 spin_lock_irq(&ctx->ctx_lock);
653 291
654 while (!list_empty(&ctx->active_reqs)) 292 while (!list_empty(&ctx->active_reqs)) {
655 req = list_first_entry(&ctx->a 293 req = list_first_entry(&ctx->active_reqs,
656 struct !! 294 struct kiocb, ki_list);
657 req->ki_cancel(&req->rw); !! 295
658 list_del_init(&req->ki_list); 296 list_del_init(&req->ki_list);
>> 297 kiocb_cancel(ctx, req, &res);
659 } 298 }
660 299
661 spin_unlock_irq(&ctx->ctx_lock); 300 spin_unlock_irq(&ctx->ctx_lock);
662 301
663 percpu_ref_kill(&ctx->reqs); !! 302 ring = kmap_atomic(ctx->ring_pages[0]);
664 percpu_ref_put(&ctx->reqs); !! 303 head = ring->head;
665 } !! 304 kunmap_atomic(ring);
666 <<
667 static int ioctx_add_table(struct kioctx *ctx, <<
668 { <<
669 unsigned i, new_nr; <<
670 struct kioctx_table *table, *old; <<
671 struct aio_ring *ring; <<
672 305
673 spin_lock(&mm->ioctx_lock); !! 306 while (atomic_read(&ctx->reqs_active) > 0) {
674 table = rcu_dereference_raw(mm->ioctx_ !! 307 wait_event(ctx->wait,
>> 308 head != ctx->tail ||
>> 309 atomic_read(&ctx->reqs_active) <= 0);
675 310
676 while (1) { !! 311 avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head;
677 if (table) <<
678 for (i = 0; i < table- <<
679 if (!rcu_acces <<
680 ctx->i <<
681 rcu_as <<
682 spin_u <<
683 <<
684 /* Whi <<
685 * we <<
686 * cha <<
687 */ <<
688 ring = <<
689 ring-> <<
690 return <<
691 } <<
692 312
693 new_nr = (table ? table->nr : !! 313 head += avail;
694 spin_unlock(&mm->ioctx_lock); !! 314 head %= ctx->nr_events;
>> 315 }
695 316
696 table = kzalloc(struct_size(ta !! 317 WARN_ON(atomic_read(&ctx->reqs_active) < 0);
697 if (!table) <<
698 return -ENOMEM; <<
699 318
700 table->nr = new_nr; !! 319 aio_free_ring(ctx);
701 320
702 spin_lock(&mm->ioctx_lock); !! 321 pr_debug("freeing %p\n", ctx);
703 old = rcu_dereference_raw(mm-> <<
704 322
705 if (!old) { !! 323 /*
706 rcu_assign_pointer(mm- !! 324 * Here the call_rcu() is between the wait_event() for reqs_active to
707 } else if (table->nr > old->nr !! 325 * hit 0, and freeing the ioctx.
708 memcpy(table->table, o !! 326 *
709 old->nr * sizeo !! 327 * aio_complete() decrements reqs_active, but it has to touch the ioctx
710 !! 328 * after to issue a wakeup so we use rcu.
711 rcu_assign_pointer(mm- !! 329 */
712 kfree_rcu(old, rcu); !! 330 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
713 } else { <<
714 kfree(table); <<
715 table = old; <<
716 } <<
717 } <<
718 } 331 }
719 332
720 static void aio_nr_sub(unsigned nr) !! 333 static void put_ioctx(struct kioctx *ctx)
721 { 334 {
722 spin_lock(&aio_nr_lock); !! 335 if (unlikely(atomic_dec_and_test(&ctx->users)))
723 if (WARN_ON(aio_nr - nr > aio_nr)) !! 336 free_ioctx(ctx);
724 aio_nr = 0; <<
725 else <<
726 aio_nr -= nr; <<
727 spin_unlock(&aio_nr_lock); <<
728 } 337 }
729 338
730 /* ioctx_alloc 339 /* ioctx_alloc
731 * Allocates and initializes an ioctx. R 340 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
732 */ 341 */
733 static struct kioctx *ioctx_alloc(unsigned nr_ 342 static struct kioctx *ioctx_alloc(unsigned nr_events)
734 { 343 {
735 struct mm_struct *mm = current->mm; 344 struct mm_struct *mm = current->mm;
736 struct kioctx *ctx; 345 struct kioctx *ctx;
737 int err = -ENOMEM; 346 int err = -ENOMEM;
738 347
739 /* <<
740 * Store the original nr_events -- wha <<
741 * for counting against the global lim <<
742 */ <<
743 unsigned int max_reqs = nr_events; <<
744 <<
745 /* <<
746 * We keep track of the number of avai <<
747 * overflow (reqs_available), and we a <<
748 * <<
749 * So since up to half the slots might <<
750 * and unavailable, double nr_events s <<
751 * expected: additionally, we move req <<
752 * counters at a time, so make sure th <<
753 */ <<
754 nr_events = max(nr_events, num_possibl <<
755 nr_events *= 2; <<
756 <<
757 /* Prevent overflows */ 348 /* Prevent overflows */
758 if (nr_events > (0x10000000U / sizeof( !! 349 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
>> 350 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
759 pr_debug("ENOMEM: nr_events to 351 pr_debug("ENOMEM: nr_events too high\n");
760 return ERR_PTR(-EINVAL); 352 return ERR_PTR(-EINVAL);
761 } 353 }
762 354
763 if (!nr_events || (unsigned long)max_r !! 355 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
764 return ERR_PTR(-EAGAIN); 356 return ERR_PTR(-EAGAIN);
765 357
766 ctx = kmem_cache_zalloc(kioctx_cachep, 358 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
767 if (!ctx) 359 if (!ctx)
768 return ERR_PTR(-ENOMEM); 360 return ERR_PTR(-ENOMEM);
769 361
770 ctx->max_reqs = max_reqs; !! 362 ctx->max_reqs = nr_events;
771 363
>> 364 atomic_set(&ctx->users, 2);
>> 365 atomic_set(&ctx->dead, 0);
772 spin_lock_init(&ctx->ctx_lock); 366 spin_lock_init(&ctx->ctx_lock);
773 spin_lock_init(&ctx->completion_lock); 367 spin_lock_init(&ctx->completion_lock);
774 mutex_init(&ctx->ring_lock); 368 mutex_init(&ctx->ring_lock);
775 /* Protect against page migration thro <<
776 * the ring_lock mutex held until setu <<
777 mutex_lock(&ctx->ring_lock); <<
778 init_waitqueue_head(&ctx->wait); 369 init_waitqueue_head(&ctx->wait);
779 370
780 INIT_LIST_HEAD(&ctx->active_reqs); 371 INIT_LIST_HEAD(&ctx->active_reqs);
781 372
782 if (percpu_ref_init(&ctx->users, free_ !! 373 if (aio_setup_ring(ctx) < 0)
783 goto err; !! 374 goto out_freectx;
784 <<
785 if (percpu_ref_init(&ctx->reqs, free_i <<
786 goto err; <<
787 <<
788 ctx->cpu = alloc_percpu(struct kioctx_ <<
789 if (!ctx->cpu) <<
790 goto err; <<
791 <<
792 err = aio_setup_ring(ctx, nr_events); <<
793 if (err < 0) <<
794 goto err; <<
795 <<
796 atomic_set(&ctx->reqs_available, ctx-> <<
797 ctx->req_batch = (ctx->nr_events - 1) <<
798 if (ctx->req_batch < 1) <<
799 ctx->req_batch = 1; <<
800 375
801 /* limit the number of system wide aio 376 /* limit the number of system wide aios */
802 spin_lock(&aio_nr_lock); 377 spin_lock(&aio_nr_lock);
803 if (aio_nr + ctx->max_reqs > aio_max_n !! 378 if (aio_nr + nr_events > aio_max_nr ||
804 aio_nr + ctx->max_reqs < aio_nr) { !! 379 aio_nr + nr_events < aio_nr) {
805 spin_unlock(&aio_nr_lock); 380 spin_unlock(&aio_nr_lock);
806 err = -EAGAIN; !! 381 goto out_cleanup;
807 goto err_ctx; <<
808 } 382 }
809 aio_nr += ctx->max_reqs; 383 aio_nr += ctx->max_reqs;
810 spin_unlock(&aio_nr_lock); 384 spin_unlock(&aio_nr_lock);
811 385
812 percpu_ref_get(&ctx->users); /* io_ !! 386 /* now link into global list. */
813 percpu_ref_get(&ctx->reqs); /* fre !! 387 spin_lock(&mm->ioctx_lock);
814 !! 388 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
815 err = ioctx_add_table(ctx, mm); !! 389 spin_unlock(&mm->ioctx_lock);
816 if (err) <<
817 goto err_cleanup; <<
818 <<
819 /* Release the ring_lock mutex now tha <<
820 mutex_unlock(&ctx->ring_lock); <<
821 390
822 pr_debug("allocated ioctx %p[%ld]: mm= 391 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
823 ctx, ctx->user_id, mm, ctx->n 392 ctx, ctx->user_id, mm, ctx->nr_events);
824 return ctx; 393 return ctx;
825 394
826 err_cleanup: !! 395 out_cleanup:
827 aio_nr_sub(ctx->max_reqs); !! 396 err = -EAGAIN;
828 err_ctx: <<
829 atomic_set(&ctx->dead, 1); <<
830 if (ctx->mmap_size) <<
831 vm_munmap(ctx->mmap_base, ctx- <<
832 aio_free_ring(ctx); 397 aio_free_ring(ctx);
833 err: !! 398 out_freectx:
834 mutex_unlock(&ctx->ring_lock); <<
835 free_percpu(ctx->cpu); <<
836 percpu_ref_exit(&ctx->reqs); <<
837 percpu_ref_exit(&ctx->users); <<
838 kmem_cache_free(kioctx_cachep, ctx); 399 kmem_cache_free(kioctx_cachep, ctx);
839 pr_debug("error allocating ioctx %d\n" 400 pr_debug("error allocating ioctx %d\n", err);
840 return ERR_PTR(err); 401 return ERR_PTR(err);
841 } 402 }
842 403
>> 404 static void kill_ioctx_work(struct work_struct *work)
>> 405 {
>> 406 struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
>> 407
>> 408 wake_up_all(&ctx->wait);
>> 409 put_ioctx(ctx);
>> 410 }
>> 411
>> 412 static void kill_ioctx_rcu(struct rcu_head *head)
>> 413 {
>> 414 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
>> 415
>> 416 INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
>> 417 schedule_work(&ctx->rcu_work);
>> 418 }
>> 419
843 /* kill_ioctx 420 /* kill_ioctx
844 * Cancels all outstanding aio requests o 421 * Cancels all outstanding aio requests on an aio context. Used
845 * when the processes owning a context ha 422 * when the processes owning a context have all exited to encourage
846 * the rapid destruction of the kioctx. 423 * the rapid destruction of the kioctx.
847 */ 424 */
848 static int kill_ioctx(struct mm_struct *mm, st !! 425 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
849 struct ctx_rq_wait *wait <<
850 { 426 {
851 struct kioctx_table *table; !! 427 if (!atomic_xchg(&ctx->dead, 1)) {
852 !! 428 spin_lock(&mm->ioctx_lock);
853 spin_lock(&mm->ioctx_lock); !! 429 hlist_del_rcu(&ctx->list);
854 if (atomic_xchg(&ctx->dead, 1)) { <<
855 spin_unlock(&mm->ioctx_lock); 430 spin_unlock(&mm->ioctx_lock);
856 return -EINVAL; <<
857 } <<
858 431
859 table = rcu_dereference_raw(mm->ioctx_ !! 432 /*
860 WARN_ON(ctx != rcu_access_pointer(tabl !! 433 * It'd be more correct to do this in free_ioctx(), after all
861 RCU_INIT_POINTER(table->table[ctx->id] !! 434 * the outstanding kiocbs have finished - but by then io_destroy
862 spin_unlock(&mm->ioctx_lock); !! 435 * has already returned, so io_setup() could potentially return
863 !! 436 * -EAGAIN with no ioctxs actually in use (as far as userspace
864 /* free_ioctx_reqs() will do the neces !! 437 * could tell).
865 wake_up_all(&ctx->wait); !! 438 */
>> 439 spin_lock(&aio_nr_lock);
>> 440 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
>> 441 aio_nr -= ctx->max_reqs;
>> 442 spin_unlock(&aio_nr_lock);
866 443
867 /* !! 444 if (ctx->mmap_size)
868 * It'd be more correct to do this in !! 445 vm_munmap(ctx->mmap_base, ctx->mmap_size);
869 * the outstanding kiocbs have finishe <<
870 * has already returned, so io_setup() <<
871 * -EAGAIN with no ioctxs actually in <<
872 * could tell). <<
873 */ <<
874 aio_nr_sub(ctx->max_reqs); <<
875 446
876 if (ctx->mmap_size) !! 447 /* Between hlist_del_rcu() and dropping the initial ref */
877 vm_munmap(ctx->mmap_base, ctx- !! 448 call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
>> 449 }
>> 450 }
878 451
879 ctx->rq_wait = wait; !! 452 /* wait_on_sync_kiocb:
880 percpu_ref_kill(&ctx->users); !! 453 * Waits on the given sync kiocb to complete.
881 return 0; !! 454 */
>> 455 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
>> 456 {
>> 457 while (atomic_read(&iocb->ki_users)) {
>> 458 set_current_state(TASK_UNINTERRUPTIBLE);
>> 459 if (!atomic_read(&iocb->ki_users))
>> 460 break;
>> 461 io_schedule();
>> 462 }
>> 463 __set_current_state(TASK_RUNNING);
>> 464 return iocb->ki_user_data;
882 } 465 }
>> 466 EXPORT_SYMBOL(wait_on_sync_kiocb);
883 467
884 /* 468 /*
885 * exit_aio: called when the last user of mm g 469 * 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 470 * no way for any new requests to be submited or any of the io_* syscalls to be
887 * called on the context. 471 * called on the context.
888 * 472 *
889 * There may be outstanding kiocbs, but free_i 473 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
890 * them. 474 * them.
891 */ 475 */
892 void exit_aio(struct mm_struct *mm) 476 void exit_aio(struct mm_struct *mm)
893 { 477 {
894 struct kioctx_table *table = rcu_deref !! 478 struct kioctx *ctx;
895 struct ctx_rq_wait wait; !! 479 struct hlist_node *n;
896 int i, skipped; <<
897 <<
898 if (!table) <<
899 return; <<
900 <<
901 atomic_set(&wait.count, table->nr); <<
902 init_completion(&wait.comp); <<
903 <<
904 skipped = 0; <<
905 for (i = 0; i < table->nr; ++i) { <<
906 struct kioctx *ctx = <<
907 rcu_dereference_protec <<
908 <<
909 if (!ctx) { <<
910 skipped++; <<
911 continue; <<
912 } <<
913 480
>> 481 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
>> 482 if (1 != atomic_read(&ctx->users))
>> 483 printk(KERN_DEBUG
>> 484 "exit_aio:ioctx still alive: %d %d %d\n",
>> 485 atomic_read(&ctx->users),
>> 486 atomic_read(&ctx->dead),
>> 487 atomic_read(&ctx->reqs_active));
914 /* 488 /*
915 * We don't need to bother wit !! 489 * We don't need to bother with munmap() here -
916 * is coming and it'll unmap e !! 490 * exit_mmap(mm) is coming and it'll unmap everything.
917 * this is not necessarily our !! 491 * Since aio_free_ring() uses non-zero ->mmap_size
918 * Since kill_ioctx() uses non !! 492 * as indicator that it needs to unmap the area,
919 * that it needs to unmap the !! 493 * just set it to 0; aio_free_ring() is the only
>> 494 * place that uses ->mmap_size, so it's safe.
920 */ 495 */
921 ctx->mmap_size = 0; 496 ctx->mmap_size = 0;
922 kill_ioctx(mm, ctx, &wait); <<
923 } <<
924 <<
925 if (!atomic_sub_and_test(skipped, &wai <<
926 /* Wait until all IO for the c <<
927 wait_for_completion(&wait.comp <<
928 } <<
929 <<
930 RCU_INIT_POINTER(mm->ioctx_table, NULL <<
931 kfree(table); <<
932 } <<
933 <<
934 static void put_reqs_available(struct kioctx * <<
935 { <<
936 struct kioctx_cpu *kcpu; <<
937 unsigned long flags; <<
938 497
939 local_irq_save(flags); !! 498 kill_ioctx(mm, ctx);
940 kcpu = this_cpu_ptr(ctx->cpu); <<
941 kcpu->reqs_available += nr; <<
942 <<
943 while (kcpu->reqs_available >= ctx->re <<
944 kcpu->reqs_available -= ctx->r <<
945 atomic_add(ctx->req_batch, &ct <<
946 } 499 }
947 <<
948 local_irq_restore(flags); <<
949 } 500 }
950 501
951 static bool __get_reqs_available(struct kioctx !! 502 /* aio_get_req
952 { !! 503 * Allocate a slot for an aio request. Increments the ki_users count
953 struct kioctx_cpu *kcpu; !! 504 * of the kioctx so that the kioctx stays around until all requests are
954 bool ret = false; !! 505 * complete. Returns NULL if no requests are free.
955 unsigned long flags; !! 506 *
956 !! 507 * Returns with kiocb->ki_users set to 2. The io submit code path holds
957 local_irq_save(flags); !! 508 * an extra reference while submitting the i/o.
958 kcpu = this_cpu_ptr(ctx->cpu); !! 509 * This prevents races between the aio code path referencing the
959 if (!kcpu->reqs_available) { !! 510 * req (after submitting it) and aio_complete() freeing the req.
960 int avail = atomic_read(&ctx-> <<
961 <<
962 do { <<
963 if (avail < ctx->req_b <<
964 goto out; <<
965 } while (!atomic_try_cmpxchg(& <<
966 & <<
967 <<
968 kcpu->reqs_available += ctx->r <<
969 } <<
970 <<
971 ret = true; <<
972 kcpu->reqs_available--; <<
973 out: <<
974 local_irq_restore(flags); <<
975 return ret; <<
976 } <<
977 <<
978 /* refill_reqs_available <<
979 * Updates the reqs_available reference c <<
980 * number of free slots in the completion <<
981 * from aio_complete() (to optimistically <<
982 * from aio_get_req() (the we're out of e <<
983 * called holding ctx->completion_lock. <<
984 */ 511 */
985 static void refill_reqs_available(struct kioct !! 512 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
986 unsigned tai <<
987 { 513 {
988 unsigned events_in_ring, completed; !! 514 struct kiocb *req;
989 <<
990 /* Clamp head since userland can write <<
991 head %= ctx->nr_events; <<
992 if (head <= tail) <<
993 events_in_ring = tail - head; <<
994 else <<
995 events_in_ring = ctx->nr_event <<
996 515
997 completed = ctx->completed_events; !! 516 if (atomic_read(&ctx->reqs_active) >= ctx->nr_events)
998 if (events_in_ring < completed) !! 517 return NULL;
999 completed -= events_in_ring; <<
1000 else <<
1001 completed = 0; <<
1002 <<
1003 if (!completed) <<
1004 return; <<
1005 518
1006 ctx->completed_events -= completed; !! 519 if (atomic_inc_return(&ctx->reqs_active) > ctx->nr_events - 1)
1007 put_reqs_available(ctx, completed); !! 520 goto out_put;
1008 } <<
1009 521
1010 /* user_refill_reqs_available !! 522 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1011 * Called to refill reqs_available when !! 523 if (unlikely(!req))
1012 * out of space in the completion ring. !! 524 goto out_put;
1013 */ <<
1014 static void user_refill_reqs_available(struct <<
1015 { <<
1016 spin_lock_irq(&ctx->completion_lock); <<
1017 if (ctx->completed_events) { <<
1018 struct aio_ring *ring; <<
1019 unsigned head; <<
1020 <<
1021 /* Access of ring->head may r <<
1022 * here, but that's okay sinc <<
1023 * or the new version, and ei <<
1024 * part is that head cannot p <<
1025 * aio_complete() from updati <<
1026 * ctx->completion_lock. Eve <<
1027 * against ctx->completed_eve <<
1028 * safe/right thing. <<
1029 */ <<
1030 ring = folio_address(ctx->rin <<
1031 head = ring->head; <<
1032 525
1033 refill_reqs_available(ctx, he !! 526 atomic_set(&req->ki_users, 2);
1034 } !! 527 req->ki_ctx = ctx;
1035 528
1036 spin_unlock_irq(&ctx->completion_lock !! 529 return req;
>> 530 out_put:
>> 531 atomic_dec(&ctx->reqs_active);
>> 532 return NULL;
1037 } 533 }
1038 534
1039 static bool get_reqs_available(struct kioctx !! 535 static void kiocb_free(struct kiocb *req)
1040 { 536 {
1041 if (__get_reqs_available(ctx)) !! 537 if (req->ki_filp)
1042 return true; !! 538 fput(req->ki_filp);
1043 user_refill_reqs_available(ctx); !! 539 if (req->ki_eventfd != NULL)
1044 return __get_reqs_available(ctx); !! 540 eventfd_ctx_put(req->ki_eventfd);
>> 541 if (req->ki_dtor)
>> 542 req->ki_dtor(req);
>> 543 if (req->ki_iovec != &req->ki_inline_vec)
>> 544 kfree(req->ki_iovec);
>> 545 kmem_cache_free(kiocb_cachep, req);
1045 } 546 }
1046 547
1047 /* aio_get_req !! 548 void aio_put_req(struct kiocb *req)
1048 * Allocate a slot for an aio request. <<
1049 * Returns NULL if no requests are free. <<
1050 * <<
1051 * The refcount is initialized to 2 - one for <<
1052 * one for the synchronous code that does thi <<
1053 */ <<
1054 static inline struct aio_kiocb *aio_get_req(s <<
1055 { 549 {
1056 struct aio_kiocb *req; !! 550 if (atomic_dec_and_test(&req->ki_users))
1057 !! 551 kiocb_free(req);
1058 req = kmem_cache_alloc(kiocb_cachep, <<
1059 if (unlikely(!req)) <<
1060 return NULL; <<
1061 <<
1062 if (unlikely(!get_reqs_available(ctx) <<
1063 kmem_cache_free(kiocb_cachep, <<
1064 return NULL; <<
1065 } <<
1066 <<
1067 percpu_ref_get(&ctx->reqs); <<
1068 req->ki_ctx = ctx; <<
1069 INIT_LIST_HEAD(&req->ki_list); <<
1070 refcount_set(&req->ki_refcnt, 2); <<
1071 req->ki_eventfd = NULL; <<
1072 return req; <<
1073 } 552 }
>> 553 EXPORT_SYMBOL(aio_put_req);
1074 554
1075 static struct kioctx *lookup_ioctx(unsigned l 555 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1076 { 556 {
1077 struct aio_ring __user *ring = (void <<
1078 struct mm_struct *mm = current->mm; 557 struct mm_struct *mm = current->mm;
1079 struct kioctx *ctx, *ret = NULL; 558 struct kioctx *ctx, *ret = NULL;
1080 struct kioctx_table *table; <<
1081 unsigned id; <<
1082 <<
1083 if (get_user(id, &ring->id)) <<
1084 return NULL; <<
1085 559
1086 rcu_read_lock(); 560 rcu_read_lock();
1087 table = rcu_dereference(mm->ioctx_tab <<
1088 <<
1089 if (!table || id >= table->nr) <<
1090 goto out; <<
1091 561
1092 id = array_index_nospec(id, table->nr !! 562 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
1093 ctx = rcu_dereference(table->table[id !! 563 if (ctx->user_id == ctx_id) {
1094 if (ctx && ctx->user_id == ctx_id) { !! 564 atomic_inc(&ctx->users);
1095 if (percpu_ref_tryget_live(&c <<
1096 ret = ctx; 565 ret = ctx;
>> 566 break;
>> 567 }
1097 } 568 }
1098 out: !! 569
1099 rcu_read_unlock(); 570 rcu_read_unlock();
1100 return ret; 571 return ret;
1101 } 572 }
1102 573
1103 static inline void iocb_destroy(struct aio_ki <<
1104 { <<
1105 if (iocb->ki_eventfd) <<
1106 eventfd_ctx_put(iocb->ki_even <<
1107 if (iocb->ki_filp) <<
1108 fput(iocb->ki_filp); <<
1109 percpu_ref_put(&iocb->ki_ctx->reqs); <<
1110 kmem_cache_free(kiocb_cachep, iocb); <<
1111 } <<
1112 <<
1113 struct aio_waiter { <<
1114 struct wait_queue_entry w; <<
1115 size_t min_nr; <<
1116 }; <<
1117 <<
1118 /* aio_complete 574 /* aio_complete
1119 * Called when the io request on the giv 575 * Called when the io request on the given iocb is complete.
1120 */ 576 */
1121 static void aio_complete(struct aio_kiocb *io !! 577 void aio_complete(struct kiocb *iocb, long res, long res2)
1122 { 578 {
1123 struct kioctx *ctx = iocb->ki_ctx; 579 struct kioctx *ctx = iocb->ki_ctx;
1124 struct aio_ring *ring; 580 struct aio_ring *ring;
1125 struct io_event *ev_page, *event; 581 struct io_event *ev_page, *event;
1126 unsigned tail, pos, head, avail; <<
1127 unsigned long flags; 582 unsigned long flags;
>> 583 unsigned tail, pos;
>> 584
>> 585 /*
>> 586 * Special case handling for sync iocbs:
>> 587 * - events go directly into the iocb for fast handling
>> 588 * - the sync task with the iocb in its stack holds the single iocb
>> 589 * ref, no other paths have a way to get another ref
>> 590 * - the sync task helpfully left a reference to itself in the iocb
>> 591 */
>> 592 if (is_sync_kiocb(iocb)) {
>> 593 BUG_ON(atomic_read(&iocb->ki_users) != 1);
>> 594 iocb->ki_user_data = res;
>> 595 atomic_set(&iocb->ki_users, 0);
>> 596 wake_up_process(iocb->ki_obj.tsk);
>> 597 return;
>> 598 }
>> 599
>> 600 /*
>> 601 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
>> 602 * need to issue a wakeup after decrementing reqs_active.
>> 603 */
>> 604 rcu_read_lock();
>> 605
>> 606 if (iocb->ki_list.next) {
>> 607 unsigned long flags;
>> 608
>> 609 spin_lock_irqsave(&ctx->ctx_lock, flags);
>> 610 list_del(&iocb->ki_list);
>> 611 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
>> 612 }
>> 613
>> 614 /*
>> 615 * cancelled requests don't get events, userland was given one
>> 616 * when the event got cancelled.
>> 617 */
>> 618 if (unlikely(xchg(&iocb->ki_cancel,
>> 619 KIOCB_CANCELLED) == KIOCB_CANCELLED)) {
>> 620 atomic_dec(&ctx->reqs_active);
>> 621 /* Still need the wake_up in case free_ioctx is waiting */
>> 622 goto put_rq;
>> 623 }
1128 624
1129 /* 625 /*
1130 * Add a completion event to the ring 626 * Add a completion event to the ring buffer. Must be done holding
1131 * ctx->completion_lock to prevent ot !! 627 * ctx->ctx_lock to prevent other code from messing with the tail
1132 * pointer since we might be called f 628 * pointer since we might be called from irq context.
1133 */ 629 */
1134 spin_lock_irqsave(&ctx->completion_lo 630 spin_lock_irqsave(&ctx->completion_lock, flags);
1135 631
1136 tail = ctx->tail; 632 tail = ctx->tail;
1137 pos = tail + AIO_EVENTS_OFFSET; 633 pos = tail + AIO_EVENTS_OFFSET;
1138 634
1139 if (++tail >= ctx->nr_events) 635 if (++tail >= ctx->nr_events)
1140 tail = 0; 636 tail = 0;
1141 637
1142 ev_page = folio_address(ctx->ring_fol !! 638 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1143 event = ev_page + pos % AIO_EVENTS_PE 639 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1144 640
1145 *event = iocb->ki_res; !! 641 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1146 !! 642 event->data = iocb->ki_user_data;
1147 flush_dcache_folio(ctx->ring_folios[p !! 643 event->res = res;
1148 !! 644 event->res2 = res2;
1149 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\ !! 645
1150 (void __user *)(unsigned lon !! 646 kunmap_atomic(ev_page);
1151 iocb->ki_res.data, iocb->ki_ !! 647 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
>> 648
>> 649 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
>> 650 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
>> 651 res, res2);
1152 652
1153 /* after flagging the request as done 653 /* after flagging the request as done, we
1154 * must never even look at it again 654 * must never even look at it again
1155 */ 655 */
1156 smp_wmb(); /* make event visible 656 smp_wmb(); /* make event visible before updating tail */
1157 657
1158 ctx->tail = tail; 658 ctx->tail = tail;
1159 659
1160 ring = folio_address(ctx->ring_folios !! 660 ring = kmap_atomic(ctx->ring_pages[0]);
1161 head = ring->head; <<
1162 ring->tail = tail; 661 ring->tail = tail;
1163 flush_dcache_folio(ctx->ring_folios[0 !! 662 kunmap_atomic(ring);
>> 663 flush_dcache_page(ctx->ring_pages[0]);
1164 664
1165 ctx->completed_events++; <<
1166 if (ctx->completed_events > 1) <<
1167 refill_reqs_available(ctx, he <<
1168 <<
1169 avail = tail > head <<
1170 ? tail - head <<
1171 : tail + ctx->nr_events - hea <<
1172 spin_unlock_irqrestore(&ctx->completi 665 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1173 666
1174 pr_debug("added to ring %p at [%u]\n" 667 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1175 668
1176 /* 669 /*
1177 * Check if the user asked us to deli 670 * Check if the user asked us to deliver the result through an
1178 * eventfd. The eventfd_signal() func 671 * eventfd. The eventfd_signal() function is safe to be called
1179 * from IRQ context. 672 * from IRQ context.
1180 */ 673 */
1181 if (iocb->ki_eventfd) !! 674 if (iocb->ki_eventfd != NULL)
1182 eventfd_signal(iocb->ki_event !! 675 eventfd_signal(iocb->ki_eventfd, 1);
>> 676
>> 677 put_rq:
>> 678 /* everything turned out well, dispose of the aiocb. */
>> 679 aio_put_req(iocb);
>> 680 atomic_dec(&ctx->reqs_active);
1183 681
1184 /* 682 /*
1185 * We have to order our ring_info tai 683 * We have to order our ring_info tail store above and test
1186 * of the wait list below outside the 684 * of the wait list below outside the wait lock. This is
1187 * like in wake_up_bit() where cleari 685 * like in wake_up_bit() where clearing a bit has to be
1188 * ordered with the unlocked test. 686 * ordered with the unlocked test.
1189 */ 687 */
1190 smp_mb(); 688 smp_mb();
1191 689
1192 if (waitqueue_active(&ctx->wait)) { !! 690 if (waitqueue_active(&ctx->wait))
1193 struct aio_waiter *curr, *nex !! 691 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 } <<
1205 692
1206 static inline void iocb_put(struct aio_kiocb !! 693 rcu_read_unlock();
1207 { <<
1208 if (refcount_dec_and_test(&iocb->ki_r <<
1209 aio_complete(iocb); <<
1210 iocb_destroy(iocb); <<
1211 } <<
1212 } 694 }
>> 695 EXPORT_SYMBOL(aio_complete);
1213 696
1214 /* aio_read_events_ring !! 697 /* aio_read_events
1215 * Pull an event off of the ioctx's even 698 * Pull an event off of the ioctx's event ring. Returns the number of
1216 * events fetched 699 * events fetched
1217 */ 700 */
1218 static long aio_read_events_ring(struct kioct 701 static long aio_read_events_ring(struct kioctx *ctx,
1219 struct io_ev 702 struct io_event __user *event, long nr)
1220 { 703 {
1221 struct aio_ring *ring; 704 struct aio_ring *ring;
1222 unsigned head, tail, pos; !! 705 unsigned head, pos;
1223 long ret = 0; 706 long ret = 0;
1224 int copy_ret; 707 int copy_ret;
1225 708
1226 /* <<
1227 * The mutex can block and wake us up <<
1228 * wait_event_interruptible_hrtimeout <<
1229 * and repeat. This should be rare en <<
1230 * peformance issues. See the comment <<
1231 */ <<
1232 sched_annotate_sleep(); <<
1233 mutex_lock(&ctx->ring_lock); 709 mutex_lock(&ctx->ring_lock);
1234 710
1235 /* Access to ->ring_folios here is pr !! 711 ring = kmap_atomic(ctx->ring_pages[0]);
1236 ring = folio_address(ctx->ring_folios <<
1237 head = ring->head; 712 head = ring->head;
1238 tail = ring->tail; !! 713 kunmap_atomic(ring);
1239 <<
1240 /* <<
1241 * Ensure that once we've read the cu <<
1242 * we also see the events that were s <<
1243 */ <<
1244 smp_rmb(); <<
1245 714
1246 pr_debug("h%u t%u m%u\n", head, tail, !! 715 pr_debug("h%u t%u m%u\n", head, ctx->tail, ctx->nr_events);
1247 716
1248 if (head == tail) !! 717 if (head == ctx->tail)
1249 goto out; 718 goto out;
1250 719
1251 head %= ctx->nr_events; 720 head %= ctx->nr_events;
1252 tail %= ctx->nr_events; <<
1253 721
1254 while (ret < nr) { 722 while (ret < nr) {
1255 long avail; 723 long avail;
1256 struct io_event *ev; 724 struct io_event *ev;
1257 struct folio *folio; !! 725 struct page *page;
1258 726
1259 avail = (head <= tail ? tail !! 727 avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head;
1260 if (head == tail) !! 728 if (head == ctx->tail)
1261 break; 729 break;
1262 730
>> 731 avail = min(avail, nr - ret);
>> 732 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
>> 733 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
>> 734
1263 pos = head + AIO_EVENTS_OFFSE 735 pos = head + AIO_EVENTS_OFFSET;
1264 folio = ctx->ring_folios[pos !! 736 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1265 pos %= AIO_EVENTS_PER_PAGE; 737 pos %= AIO_EVENTS_PER_PAGE;
1266 738
1267 avail = min(avail, nr - ret); !! 739 ev = kmap(page);
1268 avail = min_t(long, avail, AI <<
1269 <<
1270 ev = folio_address(folio); <<
1271 copy_ret = copy_to_user(event 740 copy_ret = copy_to_user(event + ret, ev + pos,
1272 sizeo 741 sizeof(*ev) * avail);
>> 742 kunmap(page);
1273 743
1274 if (unlikely(copy_ret)) { 744 if (unlikely(copy_ret)) {
1275 ret = -EFAULT; 745 ret = -EFAULT;
1276 goto out; 746 goto out;
1277 } 747 }
1278 748
1279 ret += avail; 749 ret += avail;
1280 head += avail; 750 head += avail;
1281 head %= ctx->nr_events; 751 head %= ctx->nr_events;
1282 } 752 }
1283 753
1284 ring = folio_address(ctx->ring_folios !! 754 ring = kmap_atomic(ctx->ring_pages[0]);
1285 ring->head = head; 755 ring->head = head;
1286 flush_dcache_folio(ctx->ring_folios[0 !! 756 kunmap_atomic(ring);
>> 757 flush_dcache_page(ctx->ring_pages[0]);
1287 758
1288 pr_debug("%li h%u t%u\n", ret, head, !! 759 pr_debug("%li h%u t%u\n", ret, head, ctx->tail);
1289 out: 760 out:
1290 mutex_unlock(&ctx->ring_lock); 761 mutex_unlock(&ctx->ring_lock);
1291 762
1292 return ret; 763 return ret;
1293 } 764 }
1294 765
1295 static bool aio_read_events(struct kioctx *ct 766 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1296 struct io_event _ 767 struct io_event __user *event, long *i)
1297 { 768 {
1298 long ret = aio_read_events_ring(ctx, 769 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1299 770
1300 if (ret > 0) 771 if (ret > 0)
1301 *i += ret; 772 *i += ret;
1302 773
1303 if (unlikely(atomic_read(&ctx->dead)) 774 if (unlikely(atomic_read(&ctx->dead)))
1304 ret = -EINVAL; 775 ret = -EINVAL;
1305 776
1306 if (!*i) 777 if (!*i)
1307 *i = ret; 778 *i = ret;
1308 779
1309 return ret < 0 || *i >= min_nr; 780 return ret < 0 || *i >= min_nr;
1310 } 781 }
1311 782
1312 static long read_events(struct kioctx *ctx, l 783 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1313 struct io_event __use 784 struct io_event __user *event,
1314 ktime_t until) !! 785 struct timespec __user *timeout)
1315 { 786 {
1316 struct hrtimer_sleeper t; !! 787 ktime_t until = { .tv64 = KTIME_MAX };
1317 struct aio_waiter w; !! 788 long ret = 0;
1318 long ret = 0, ret2 = 0; !! 789
>> 790 if (timeout) {
>> 791 struct timespec ts;
>> 792
>> 793 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
>> 794 return -EFAULT;
>> 795
>> 796 until = timespec_to_ktime(ts);
>> 797 }
1319 798
1320 /* 799 /*
1321 * Note that aio_read_events() is bei 800 * Note that aio_read_events() is being called as the conditional - i.e.
1322 * we're calling it after prepare_to_ 801 * we're calling it after prepare_to_wait() has set task state to
1323 * TASK_INTERRUPTIBLE. 802 * TASK_INTERRUPTIBLE.
1324 * 803 *
1325 * But aio_read_events() can block, a 804 * But aio_read_events() can block, and if it blocks it's going to flip
1326 * the task state back to TASK_RUNNIN 805 * the task state back to TASK_RUNNING.
1327 * 806 *
1328 * This should be ok, provided it doe 807 * This should be ok, provided it doesn't flip the state back to
1329 * TASK_RUNNING and return 0 too much 808 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1330 * will only happen if the mutex_lock 809 * will only happen if the mutex_lock() call blocks, and we then find
1331 * the ringbuffer empty. So in practi 810 * the ringbuffer empty. So in practice we should be ok, but it's
1332 * something to be aware of when touc 811 * something to be aware of when touching this code.
1333 */ 812 */
1334 aio_read_events(ctx, min_nr, nr, even !! 813 wait_event_interruptible_hrtimeout(ctx->wait,
1335 if (until == 0 || ret < 0 || ret >= m !! 814 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1336 return ret; <<
1337 <<
1338 hrtimer_init_sleeper_on_stack(&t, CLO <<
1339 if (until != KTIME_MAX) { <<
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 815
1362 finish_wait(&ctx->wait, &w.w); !! 816 if (!ret && signal_pending(current))
1363 hrtimer_cancel(&t.timer); !! 817 ret = -EINTR;
1364 destroy_hrtimer_on_stack(&t.timer); <<
1365 818
1366 return ret; 819 return ret;
1367 } 820 }
1368 821
1369 /* sys_io_setup: 822 /* sys_io_setup:
1370 * Create an aio_context capable of rece 823 * Create an aio_context capable of receiving at least nr_events.
1371 * ctxp must not point to an aio_context 824 * ctxp must not point to an aio_context that already exists, and
1372 * must be initialized to 0 prior to the 825 * must be initialized to 0 prior to the call. On successful
1373 * creation of the aio_context, *ctxp is 826 * creation of the aio_context, *ctxp is filled in with the resulting
1374 * handle. May fail with -EINVAL if *ct 827 * handle. May fail with -EINVAL if *ctxp is not initialized,
1375 * if the specified nr_events exceeds in 828 * if the specified nr_events exceeds internal limits. May fail
1376 * with -EAGAIN if the specified nr_even 829 * with -EAGAIN if the specified nr_events exceeds the user's limit
1377 * of available events. May fail with - 830 * of available events. May fail with -ENOMEM if insufficient kernel
1378 * resources are available. May fail wi 831 * resources are available. May fail with -EFAULT if an invalid
1379 * pointer is passed for ctxp. Will fai 832 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1380 * implemented. 833 * implemented.
1381 */ 834 */
1382 SYSCALL_DEFINE2(io_setup, unsigned, nr_events 835 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1383 { 836 {
1384 struct kioctx *ioctx = NULL; 837 struct kioctx *ioctx = NULL;
1385 unsigned long ctx; 838 unsigned long ctx;
1386 long ret; 839 long ret;
1387 840
1388 ret = get_user(ctx, ctxp); 841 ret = get_user(ctx, ctxp);
1389 if (unlikely(ret)) 842 if (unlikely(ret))
1390 goto out; 843 goto out;
1391 844
1392 ret = -EINVAL; 845 ret = -EINVAL;
1393 if (unlikely(ctx || nr_events == 0)) 846 if (unlikely(ctx || nr_events == 0)) {
1394 pr_debug("EINVAL: ctx %lu nr_ !! 847 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1395 ctx, nr_events); 848 ctx, nr_events);
1396 goto out; 849 goto out;
1397 } 850 }
1398 851
1399 ioctx = ioctx_alloc(nr_events); 852 ioctx = ioctx_alloc(nr_events);
1400 ret = PTR_ERR(ioctx); 853 ret = PTR_ERR(ioctx);
1401 if (!IS_ERR(ioctx)) { 854 if (!IS_ERR(ioctx)) {
1402 ret = put_user(ioctx->user_id 855 ret = put_user(ioctx->user_id, ctxp);
1403 if (ret) 856 if (ret)
1404 kill_ioctx(current->m !! 857 kill_ioctx(current->mm, ioctx);
1405 percpu_ref_put(&ioctx->users) !! 858 put_ioctx(ioctx);
1406 } <<
1407 <<
1408 out: <<
1409 return ret; <<
1410 } <<
1411 <<
1412 #ifdef CONFIG_COMPAT <<
1413 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr <<
1414 { <<
1415 struct kioctx *ioctx = NULL; <<
1416 unsigned long ctx; <<
1417 long ret; <<
1418 <<
1419 ret = get_user(ctx, ctx32p); <<
1420 if (unlikely(ret)) <<
1421 goto out; <<
1422 <<
1423 ret = -EINVAL; <<
1424 if (unlikely(ctx || nr_events == 0)) <<
1425 pr_debug("EINVAL: ctx %lu nr_ <<
1426 ctx, nr_events); <<
1427 goto out; <<
1428 } <<
1429 <<
1430 ioctx = ioctx_alloc(nr_events); <<
1431 ret = PTR_ERR(ioctx); <<
1432 if (!IS_ERR(ioctx)) { <<
1433 /* truncating is ok because i <<
1434 ret = put_user((u32)ioctx->us <<
1435 if (ret) <<
1436 kill_ioctx(current->m <<
1437 percpu_ref_put(&ioctx->users) <<
1438 } 859 }
1439 860
1440 out: 861 out:
1441 return ret; 862 return ret;
1442 } 863 }
1443 #endif <<
1444 864
1445 /* sys_io_destroy: 865 /* sys_io_destroy:
1446 * Destroy the aio_context specified. M 866 * Destroy the aio_context specified. May cancel any outstanding
1447 * AIOs and block on completion. Will f 867 * AIOs and block on completion. Will fail with -ENOSYS if not
1448 * implemented. May fail with -EINVAL i 868 * implemented. May fail with -EINVAL if the context pointed to
1449 * is invalid. 869 * is invalid.
1450 */ 870 */
1451 SYSCALL_DEFINE1(io_destroy, aio_context_t, ct 871 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1452 { 872 {
1453 struct kioctx *ioctx = lookup_ioctx(c 873 struct kioctx *ioctx = lookup_ioctx(ctx);
1454 if (likely(NULL != ioctx)) { 874 if (likely(NULL != ioctx)) {
1455 struct ctx_rq_wait wait; !! 875 kill_ioctx(current->mm, ioctx);
1456 int ret; !! 876 put_ioctx(ioctx);
1457 !! 877 return 0;
1458 init_completion(&wait.comp); <<
1459 atomic_set(&wait.count, 1); <<
1460 <<
1461 /* Pass requests_done to kill <<
1462 * in a thread-safe way. If w <<
1463 * a race condition if two io <<
1464 */ <<
1465 ret = kill_ioctx(current->mm, <<
1466 percpu_ref_put(&ioctx->users) <<
1467 <<
1468 /* Wait until all IO for the <<
1469 * keep using user-space buff <<
1470 * is destroyed. <<
1471 */ <<
1472 if (!ret) <<
1473 wait_for_completion(& <<
1474 <<
1475 return ret; <<
1476 } 878 }
1477 pr_debug("EINVAL: invalid context id\ !! 879 pr_debug("EINVAL: io_destroy: invalid context id\n");
1478 return -EINVAL; 880 return -EINVAL;
1479 } 881 }
1480 882
1481 static void aio_remove_iocb(struct aio_kiocb !! 883 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1482 { 884 {
1483 struct kioctx *ctx = iocb->ki_ctx; !! 885 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1484 unsigned long flags; <<
1485 <<
1486 spin_lock_irqsave(&ctx->ctx_lock, fla <<
1487 list_del(&iocb->ki_list); <<
1488 spin_unlock_irqrestore(&ctx->ctx_lock <<
1489 } <<
1490 <<
1491 static void aio_complete_rw(struct kiocb *kio <<
1492 { <<
1493 struct aio_kiocb *iocb = container_of <<
1494 886
1495 if (!list_empty_careful(&iocb->ki_lis !! 887 BUG_ON(ret <= 0);
1496 aio_remove_iocb(iocb); <<
1497 888
1498 if (kiocb->ki_flags & IOCB_WRITE) { !! 889 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1499 struct inode *inode = file_in !! 890 ssize_t this = min((ssize_t)iov->iov_len, ret);
1500 !! 891 iov->iov_base += this;
1501 if (S_ISREG(inode->i_mode)) !! 892 iov->iov_len -= this;
1502 kiocb_end_write(kiocb !! 893 iocb->ki_left -= this;
1503 } !! 894 ret -= this;
1504 !! 895 if (iov->iov_len == 0) {
1505 iocb->ki_res.res = res; !! 896 iocb->ki_cur_seg++;
1506 iocb->ki_res.res2 = 0; !! 897 iov++;
1507 iocb_put(iocb); <<
1508 } <<
1509 <<
1510 static int aio_prep_rw(struct kiocb *req, con <<
1511 { <<
1512 int ret; <<
1513 <<
1514 req->ki_complete = aio_complete_rw; <<
1515 req->private = NULL; <<
1516 req->ki_pos = iocb->aio_offset; <<
1517 req->ki_flags = req->ki_filp->f_iocb_ <<
1518 if (iocb->aio_flags & IOCB_FLAG_RESFD <<
1519 req->ki_flags |= IOCB_EVENTFD <<
1520 if (iocb->aio_flags & IOCB_FLAG_IOPRI <<
1521 /* <<
1522 * If the IOCB_FLAG_IOPRIO fl <<
1523 * aio_reqprio is interpreted <<
1524 * class and priority. <<
1525 */ <<
1526 ret = ioprio_check_cap(iocb-> <<
1527 if (ret) { <<
1528 pr_debug("aio ioprio <<
1529 return ret; <<
1530 } 898 }
1531 <<
1532 req->ki_ioprio = iocb->aio_re <<
1533 } else <<
1534 req->ki_ioprio = get_current_ <<
1535 <<
1536 ret = kiocb_set_rw_flags(req, iocb->a <<
1537 if (unlikely(ret)) <<
1538 return ret; <<
1539 <<
1540 req->ki_flags &= ~IOCB_HIPRI; /* no o <<
1541 return 0; <<
1542 } <<
1543 <<
1544 static ssize_t aio_setup_rw(int rw, const str <<
1545 struct iovec **iovec, bool ve <<
1546 struct iov_iter *iter) <<
1547 { <<
1548 void __user *buf = (void __user *)(ui <<
1549 size_t len = iocb->aio_nbytes; <<
1550 <<
1551 if (!vectored) { <<
1552 ssize_t ret = import_ubuf(rw, <<
1553 *iovec = NULL; <<
1554 return ret; <<
1555 } 899 }
1556 900
1557 return __import_iovec(rw, buf, len, U !! 901 /* the caller should not have done more io than what fit in
>> 902 * the remaining iovecs */
>> 903 BUG_ON(ret > 0 && iocb->ki_left == 0);
1558 } 904 }
1559 905
1560 static inline void aio_rw_done(struct kiocb * !! 906 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1561 { !! 907 unsigned long, loff_t);
1562 switch (ret) { <<
1563 case -EIOCBQUEUED: <<
1564 break; <<
1565 case -ERESTARTSYS: <<
1566 case -ERESTARTNOINTR: <<
1567 case -ERESTARTNOHAND: <<
1568 case -ERESTART_RESTARTBLOCK: <<
1569 /* <<
1570 * There's no easy way to res <<
1571 * may be already running. Ju <<
1572 */ <<
1573 ret = -EINTR; <<
1574 fallthrough; <<
1575 default: <<
1576 req->ki_complete(req, ret); <<
1577 } <<
1578 } <<
1579 908
1580 static int aio_read(struct kiocb *req, const !! 909 static ssize_t aio_rw_vect_retry(struct kiocb *iocb, int rw, aio_rw_op *rw_op)
1581 bool vectored, bool c <<
1582 { 910 {
1583 struct iovec inline_vecs[UIO_FASTIOV] !! 911 struct file *file = iocb->ki_filp;
1584 struct iov_iter iter; !! 912 struct address_space *mapping = file->f_mapping;
1585 struct file *file; !! 913 struct inode *inode = mapping->host;
1586 int ret; !! 914 ssize_t ret = 0;
1587 915
1588 ret = aio_prep_rw(req, iocb, READ); !! 916 /* This matches the pread()/pwrite() logic */
1589 if (ret) !! 917 if (iocb->ki_pos < 0)
1590 return ret; <<
1591 file = req->ki_filp; <<
1592 if (unlikely(!(file->f_mode & FMODE_R <<
1593 return -EBADF; <<
1594 if (unlikely(!file->f_op->read_iter)) <<
1595 return -EINVAL; 918 return -EINVAL;
1596 919
1597 ret = aio_setup_rw(ITER_DEST, iocb, & !! 920 if (rw == WRITE)
1598 if (ret < 0) !! 921 file_start_write(file);
1599 return ret; !! 922 do {
1600 ret = rw_verify_area(READ, file, &req !! 923 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1601 if (!ret) !! 924 iocb->ki_nr_segs - iocb->ki_cur_seg,
1602 aio_rw_done(req, file->f_op-> !! 925 iocb->ki_pos);
1603 kfree(iovec); !! 926 if (ret > 0)
>> 927 aio_advance_iovec(iocb, ret);
>> 928
>> 929 /* retry all partial writes. retry partial reads as long as its a
>> 930 * regular file. */
>> 931 } while (ret > 0 && iocb->ki_left > 0 &&
>> 932 (rw == WRITE ||
>> 933 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
>> 934 if (rw == WRITE)
>> 935 file_end_write(file);
>> 936
>> 937 /* This means we must have transferred all that we could */
>> 938 /* No need to retry anymore */
>> 939 if ((ret == 0) || (iocb->ki_left == 0))
>> 940 ret = iocb->ki_nbytes - iocb->ki_left;
>> 941
>> 942 /* If we managed to write some out we return that, rather than
>> 943 * the eventual error. */
>> 944 if (rw == WRITE
>> 945 && ret < 0 && ret != -EIOCBQUEUED
>> 946 && iocb->ki_nbytes - iocb->ki_left)
>> 947 ret = iocb->ki_nbytes - iocb->ki_left;
>> 948
1604 return ret; 949 return ret;
1605 } 950 }
1606 951
1607 static int aio_write(struct kiocb *req, const !! 952 static ssize_t aio_setup_vectored_rw(int rw, struct kiocb *kiocb, bool compat)
1608 bool vectored, bool <<
1609 { 953 {
1610 struct iovec inline_vecs[UIO_FASTIOV] !! 954 ssize_t ret;
1611 struct iov_iter iter; <<
1612 struct file *file; <<
1613 int ret; <<
1614 <<
1615 ret = aio_prep_rw(req, iocb, WRITE); <<
1616 if (ret) <<
1617 return ret; <<
1618 file = req->ki_filp; <<
1619 955
1620 if (unlikely(!(file->f_mode & FMODE_W !! 956 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1621 return -EBADF; <<
1622 if (unlikely(!file->f_op->write_iter) <<
1623 return -EINVAL; <<
1624 957
1625 ret = aio_setup_rw(ITER_SOURCE, iocb, !! 958 #ifdef CONFIG_COMPAT
>> 959 if (compat)
>> 960 ret = compat_rw_copy_check_uvector(rw,
>> 961 (struct compat_iovec __user *)kiocb->ki_buf,
>> 962 kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec,
>> 963 &kiocb->ki_iovec);
>> 964 else
>> 965 #endif
>> 966 ret = rw_copy_check_uvector(rw,
>> 967 (struct iovec __user *)kiocb->ki_buf,
>> 968 kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec,
>> 969 &kiocb->ki_iovec);
1626 if (ret < 0) 970 if (ret < 0)
1627 return ret; 971 return ret;
1628 ret = rw_verify_area(WRITE, file, &re <<
1629 if (!ret) { <<
1630 if (S_ISREG(file_inode(file)- <<
1631 kiocb_start_write(req <<
1632 req->ki_flags |= IOCB_WRITE; <<
1633 aio_rw_done(req, file->f_op-> <<
1634 } <<
1635 kfree(iovec); <<
1636 return ret; <<
1637 } <<
1638 972
1639 static void aio_fsync_work(struct work_struct !! 973 /* ki_nbytes now reflect bytes instead of segs */
1640 { !! 974 kiocb->ki_nbytes = ret;
1641 struct aio_kiocb *iocb = container_of <<
1642 const struct cred *old_cred = overrid <<
1643 <<
1644 iocb->ki_res.res = vfs_fsync(iocb->fs <<
1645 revert_creds(old_cred); <<
1646 put_cred(iocb->fsync.creds); <<
1647 iocb_put(iocb); <<
1648 } <<
1649 <<
1650 static int aio_fsync(struct fsync_iocb *req, <<
1651 bool datasync) <<
1652 { <<
1653 if (unlikely(iocb->aio_buf || iocb->a <<
1654 iocb->aio_rw_flags)) <<
1655 return -EINVAL; <<
1656 <<
1657 if (unlikely(!req->file->f_op->fsync) <<
1658 return -EINVAL; <<
1659 <<
1660 req->creds = prepare_creds(); <<
1661 if (!req->creds) <<
1662 return -ENOMEM; <<
1663 <<
1664 req->datasync = datasync; <<
1665 INIT_WORK(&req->work, aio_fsync_work) <<
1666 schedule_work(&req->work); <<
1667 return 0; 975 return 0;
1668 } 976 }
1669 977
1670 static void aio_poll_put_work(struct work_str !! 978 static ssize_t aio_setup_single_vector(int rw, struct kiocb *kiocb)
1671 { 979 {
1672 struct poll_iocb *req = container_of( !! 980 size_t len = kiocb->ki_nbytes;
1673 struct aio_kiocb *iocb = container_of <<
1674 981
1675 iocb_put(iocb); !! 982 if (len > MAX_RW_COUNT)
1676 } !! 983 len = MAX_RW_COUNT;
1677 <<
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 984
1690 /* !! 985 if (unlikely(!access_ok(!rw, kiocb->ki_buf, len)))
1691 * While we hold the waitqueue lock a !! 986 return -EFAULT;
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 <<
1724 { <<
1725 struct poll_iocb *req = container_of( <<
1726 struct aio_kiocb *iocb = container_of <<
1727 struct poll_table_struct pt = { ._key <<
1728 struct kioctx *ctx = iocb->ki_ctx; <<
1729 __poll_t mask = 0; <<
1730 <<
1731 if (!READ_ONCE(req->cancelled)) <<
1732 mask = vfs_poll(req->file, &p <<
1733 <<
1734 /* <<
1735 * Note that ->ki_cancel callers also <<
1736 * calling ->ki_cancel. We need the <<
1737 * synchronize with them. In the can <<
1738 * itself is not actually needed, but <<
1739 * avoid further branches in the fast <<
1740 */ <<
1741 spin_lock_irq(&ctx->ctx_lock); <<
1742 if (poll_iocb_lock_wq(req)) { <<
1743 if (!mask && !READ_ONCE(req-> <<
1744 /* <<
1745 * The request isn't <<
1746 * Reschedule complet <<
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); <<
1762 iocb->ki_res.res = mangle_poll(mask); <<
1763 spin_unlock_irq(&ctx->ctx_lock); <<
1764 <<
1765 iocb_put(iocb); <<
1766 } <<
1767 <<
1768 /* assumes we are called with irqs disabled * <<
1769 static int aio_poll_cancel(struct kiocb *iocb <<
1770 { <<
1771 struct aio_kiocb *aiocb = container_o <<
1772 struct poll_iocb *req = &aiocb->poll; <<
1773 <<
1774 if (poll_iocb_lock_wq(req)) { <<
1775 WRITE_ONCE(req->cancelled, tr <<
1776 if (!req->work_scheduled) { <<
1777 schedule_work(&aiocb- <<
1778 req->work_scheduled = <<
1779 } <<
1780 poll_iocb_unlock_wq(req); <<
1781 } /* else, the request was force-canc <<
1782 987
>> 988 kiocb->ki_iovec = &kiocb->ki_inline_vec;
>> 989 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
>> 990 kiocb->ki_iovec->iov_len = len;
>> 991 kiocb->ki_nr_segs = 1;
1783 return 0; 992 return 0;
1784 } 993 }
1785 994
1786 static int aio_poll_wake(struct wait_queue_en !! 995 /*
1787 void *key) !! 996 * aio_setup_iocb:
1788 { !! 997 * Performs the initial checks and aio retry method
1789 struct poll_iocb *req = container_of( !! 998 * setup for the kiocb at the time of io submission.
1790 struct aio_kiocb *iocb = container_of !! 999 */
1791 __poll_t mask = key_to_poll(key); !! 1000 static ssize_t aio_run_iocb(struct kiocb *req, bool compat)
1792 unsigned long flags; !! 1001 {
1793 !! 1002 struct file *file = req->ki_filp;
1794 /* for instances that support it chec !! 1003 ssize_t ret;
1795 if (mask && !(mask & req->events)) !! 1004 int rw;
1796 return 0; !! 1005 fmode_t mode;
1797 !! 1006 aio_rw_op *rw_op;
1798 /* <<
1799 * Complete the request inline if pos <<
1800 * conditions be met: <<
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 <<
1814 1007
1815 list_del_init(&req->wait.entr !! 1008 switch (req->ki_opcode) {
1816 list_del(&iocb->ki_list); !! 1009 case IOCB_CMD_PREAD:
1817 iocb->ki_res.res = mangle_pol !! 1010 case IOCB_CMD_PREADV:
1818 if (iocb->ki_eventfd && !even !! 1011 mode = FMODE_READ;
1819 iocb = NULL; !! 1012 rw = READ;
1820 INIT_WORK(&req->work, !! 1013 rw_op = file->f_op->aio_read;
1821 schedule_work(&req->w !! 1014 goto rw_common;
1822 } <<
1823 spin_unlock_irqrestore(&ctx-> <<
1824 if (iocb) <<
1825 iocb_put(iocb); <<
1826 } else { <<
1827 /* <<
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 1015
1843 /* !! 1016 case IOCB_CMD_PWRITE:
1844 * If the waitqueue is being !! 1017 case IOCB_CMD_PWRITEV:
1845 * the request inline, we hav !! 1018 mode = FMODE_WRITE;
1846 * we can. That means immedi !! 1019 rw = WRITE;
1847 * waitqueue and preventing a !! 1020 rw_op = file->f_op->aio_write;
1848 * waitqueue via the request. !! 1021 goto rw_common;
1849 * completion work (done abov !! 1022 rw_common:
1850 * cancelled, to potentially !! 1023 if (unlikely(!(file->f_mode & mode)))
1851 */ !! 1024 return -EBADF;
1852 if (mask & POLLFREE) { !! 1025
1853 WRITE_ONCE(req->cance !! 1026 if (!rw_op)
1854 list_del_init(&req->w !! 1027 return -EINVAL;
1855 !! 1028
1856 /* !! 1029 ret = (req->ki_opcode == IOCB_CMD_PREADV ||
1857 * Careful: this *mus !! 1030 req->ki_opcode == IOCB_CMD_PWRITEV)
1858 * as req->head is NU !! 1031 ? aio_setup_vectored_rw(rw, req, compat)
1859 * completed and free !! 1032 : aio_setup_single_vector(rw, req);
1860 * will no longer nee !! 1033 if (ret)
1861 */ !! 1034 return ret;
1862 smp_store_release(&re <<
1863 } <<
1864 } <<
1865 return 1; <<
1866 } <<
1867 1035
1868 struct aio_poll_table { !! 1036 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1869 struct poll_table_struct pt; !! 1037 if (ret < 0)
1870 struct aio_kiocb *iocb !! 1038 return ret;
1871 bool queue <<
1872 int error <<
1873 }; <<
1874 1039
1875 static void !! 1040 req->ki_nbytes = ret;
1876 aio_poll_queue_proc(struct file *file, struct !! 1041 req->ki_left = ret;
1877 struct poll_table_struct *p) <<
1878 { <<
1879 struct aio_poll_table *pt = container <<
1880 <<
1881 /* multiple wait queues per file are <<
1882 if (unlikely(pt->queued)) { <<
1883 pt->error = -EINVAL; <<
1884 return; <<
1885 } <<
1886 1042
1887 pt->queued = true; !! 1043 ret = aio_rw_vect_retry(req, rw, rw_op);
1888 pt->error = 0; !! 1044 break;
1889 pt->iocb->poll.head = head; <<
1890 add_wait_queue(head, &pt->iocb->poll. <<
1891 } <<
1892 <<
1893 static int aio_poll(struct aio_kiocb *aiocb, <<
1894 { <<
1895 struct kioctx *ctx = aiocb->ki_ctx; <<
1896 struct poll_iocb *req = &aiocb->poll; <<
1897 struct aio_poll_table apt; <<
1898 bool cancel = false; <<
1899 __poll_t mask; <<
1900 1045
1901 /* reject any unknown events outside !! 1046 case IOCB_CMD_FDSYNC:
1902 if ((u16)iocb->aio_buf != iocb->aio_b !! 1047 if (!file->f_op->aio_fsync)
1903 return -EINVAL; !! 1048 return -EINVAL;
1904 /* reject fields that are not defined <<
1905 if (iocb->aio_offset || iocb->aio_nby <<
1906 return -EINVAL; <<
1907 1049
1908 INIT_WORK(&req->work, aio_poll_comple !! 1050 ret = file->f_op->aio_fsync(req, 1);
1909 req->events = demangle_poll(iocb->aio !! 1051 break;
1910 1052
1911 req->head = NULL; !! 1053 case IOCB_CMD_FSYNC:
1912 req->cancelled = false; !! 1054 if (!file->f_op->aio_fsync)
1913 req->work_scheduled = false; !! 1055 return -EINVAL;
1914 req->work_need_resched = false; <<
1915 <<
1916 apt.pt._qproc = aio_poll_queue_proc; <<
1917 apt.pt._key = req->events; <<
1918 apt.iocb = aiocb; <<
1919 apt.queued = false; <<
1920 apt.error = -EINVAL; /* same as no su <<
1921 <<
1922 /* initialized the list so that we ca <<
1923 INIT_LIST_HEAD(&req->wait.entry); <<
1924 init_waitqueue_func_entry(&req->wait, <<
1925 1056
1926 mask = vfs_poll(req->file, &apt.pt) & !! 1057 ret = file->f_op->aio_fsync(req, 0);
1927 spin_lock_irq(&ctx->ctx_lock); !! 1058 break;
1928 if (likely(apt.queued)) { <<
1929 bool on_queue = poll_iocb_loc <<
1930 1059
1931 if (!on_queue || req->work_sc !! 1060 default:
1932 /* !! 1061 pr_debug("EINVAL: no operation provided\n");
1933 * aio_poll_wake() al !! 1062 return -EINVAL;
1934 * completion work, o <<
1935 */ <<
1936 if (apt.error) /* uns <<
1937 cancel = true <<
1938 apt.error = 0; <<
1939 mask = 0; <<
1940 } <<
1941 if (mask || apt.error) { <<
1942 /* Steal to complete <<
1943 list_del_init(&req->w <<
1944 } else if (cancel) { <<
1945 /* Cancel if possible <<
1946 WRITE_ONCE(req->cance <<
1947 } else if (on_queue) { <<
1948 /* <<
1949 * Actually waiting f <<
1950 * active_reqs so tha <<
1951 */ <<
1952 list_add_tail(&aiocb- <<
1953 aiocb->ki_cancel = ai <<
1954 } <<
1955 if (on_queue) <<
1956 poll_iocb_unlock_wq(r <<
1957 } <<
1958 if (mask) { /* no async, we'd stolen <<
1959 aiocb->ki_res.res = mangle_po <<
1960 apt.error = 0; <<
1961 } 1063 }
1962 spin_unlock_irq(&ctx->ctx_lock); <<
1963 if (mask) <<
1964 iocb_put(aiocb); <<
1965 return apt.error; <<
1966 } <<
1967 <<
1968 static int __io_submit_one(struct kioctx *ctx <<
1969 struct iocb __user <<
1970 bool compat) <<
1971 { <<
1972 req->ki_filp = fget(iocb->aio_fildes) <<
1973 if (unlikely(!req->ki_filp)) <<
1974 return -EBADF; <<
1975 1064
1976 if (iocb->aio_flags & IOCB_FLAG_RESFD !! 1065 if (ret != -EIOCBQUEUED) {
1977 struct eventfd_ctx *eventfd; <<
1978 /* 1066 /*
1979 * If the IOCB_FLAG_RESFD fla !! 1067 * There's no easy way to restart the syscall since other AIO's
1980 * instance of the file* now. !! 1068 * may be already running. Just fail this IO with EINTR.
1981 * an eventfd() fd, and will <<
1982 * event using the eventfd_si <<
1983 */ 1069 */
1984 eventfd = eventfd_ctx_fdget(i !! 1070 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1985 if (IS_ERR(eventfd)) !! 1071 ret == -ERESTARTNOHAND ||
1986 return PTR_ERR(eventf !! 1072 ret == -ERESTART_RESTARTBLOCK))
1987 !! 1073 ret = -EINTR;
1988 req->ki_eventfd = eventfd; !! 1074 aio_complete(req, ret, 0);
1989 } 1075 }
1990 1076
1991 if (unlikely(put_user(KIOCB_KEY, &use !! 1077 return 0;
1992 pr_debug("EFAULT: aio_key\n") <<
1993 return -EFAULT; <<
1994 } <<
1995 <<
1996 req->ki_res.obj = (u64)(unsigned long <<
1997 req->ki_res.data = iocb->aio_data; <<
1998 req->ki_res.res = 0; <<
1999 req->ki_res.res2 = 0; <<
2000 <<
2001 switch (iocb->aio_lio_opcode) { <<
2002 case IOCB_CMD_PREAD: <<
2003 return aio_read(&req->rw, ioc <<
2004 case IOCB_CMD_PWRITE: <<
2005 return aio_write(&req->rw, io <<
2006 case IOCB_CMD_PREADV: <<
2007 return aio_read(&req->rw, ioc <<
2008 case IOCB_CMD_PWRITEV: <<
2009 return aio_write(&req->rw, io <<
2010 case IOCB_CMD_FSYNC: <<
2011 return aio_fsync(&req->fsync, <<
2012 case IOCB_CMD_FDSYNC: <<
2013 return aio_fsync(&req->fsync, <<
2014 case IOCB_CMD_POLL: <<
2015 return aio_poll(req, iocb); <<
2016 default: <<
2017 pr_debug("invalid aio operati <<
2018 return -EINVAL; <<
2019 } <<
2020 } 1078 }
2021 1079
2022 static int io_submit_one(struct kioctx *ctx, 1080 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
2023 bool compat) !! 1081 struct iocb *iocb, bool compat)
2024 { 1082 {
2025 struct aio_kiocb *req; !! 1083 struct kiocb *req;
2026 struct iocb iocb; !! 1084 ssize_t ret;
2027 int err; <<
2028 <<
2029 if (unlikely(copy_from_user(&iocb, us <<
2030 return -EFAULT; <<
2031 1085
2032 /* enforce forwards compatibility on 1086 /* enforce forwards compatibility on users */
2033 if (unlikely(iocb.aio_reserved2)) { !! 1087 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
2034 pr_debug("EINVAL: reserve fie 1088 pr_debug("EINVAL: reserve field set\n");
2035 return -EINVAL; 1089 return -EINVAL;
2036 } 1090 }
2037 1091
2038 /* prevent overflows */ 1092 /* prevent overflows */
2039 if (unlikely( 1093 if (unlikely(
2040 (iocb.aio_buf != (unsigned long)i !! 1094 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
2041 (iocb.aio_nbytes != (size_t)iocb. !! 1095 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
2042 ((ssize_t)iocb.aio_nbytes < 0) !! 1096 ((ssize_t)iocb->aio_nbytes < 0)
2043 )) { 1097 )) {
2044 pr_debug("EINVAL: overflow ch !! 1098 pr_debug("EINVAL: io_submit: overflow check\n");
2045 return -EINVAL; 1099 return -EINVAL;
2046 } 1100 }
2047 1101
2048 req = aio_get_req(ctx); 1102 req = aio_get_req(ctx);
2049 if (unlikely(!req)) 1103 if (unlikely(!req))
2050 return -EAGAIN; 1104 return -EAGAIN;
2051 1105
2052 err = __io_submit_one(ctx, &iocb, use !! 1106 req->ki_filp = fget(iocb->aio_fildes);
>> 1107 if (unlikely(!req->ki_filp)) {
>> 1108 ret = -EBADF;
>> 1109 goto out_put_req;
>> 1110 }
2053 1111
2054 /* Done with the synchronous referenc !! 1112 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
2055 iocb_put(req); !! 1113 /*
>> 1114 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
>> 1115 * instance of the file* now. The file descriptor must be
>> 1116 * an eventfd() fd, and will be signaled for each completed
>> 1117 * event using the eventfd_signal() function.
>> 1118 */
>> 1119 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
>> 1120 if (IS_ERR(req->ki_eventfd)) {
>> 1121 ret = PTR_ERR(req->ki_eventfd);
>> 1122 req->ki_eventfd = NULL;
>> 1123 goto out_put_req;
>> 1124 }
>> 1125 }
2056 1126
2057 /* !! 1127 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
2058 * If err is 0, we'd either done aio_ !! 1128 if (unlikely(ret)) {
2059 * arranged for that to be done async !! 1129 pr_debug("EFAULT: aio_key\n");
2060 * means that we need to destroy req !! 1130 goto out_put_req;
2061 */ <<
2062 if (unlikely(err)) { <<
2063 iocb_destroy(req); <<
2064 put_reqs_available(ctx, 1); <<
2065 } 1131 }
2066 return err; !! 1132
>> 1133 req->ki_obj.user = user_iocb;
>> 1134 req->ki_user_data = iocb->aio_data;
>> 1135 req->ki_pos = iocb->aio_offset;
>> 1136
>> 1137 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
>> 1138 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
>> 1139 req->ki_opcode = iocb->aio_lio_opcode;
>> 1140
>> 1141 ret = aio_run_iocb(req, compat);
>> 1142 if (ret)
>> 1143 goto out_put_req;
>> 1144
>> 1145 aio_put_req(req); /* drop extra ref to req */
>> 1146 return 0;
>> 1147 out_put_req:
>> 1148 atomic_dec(&ctx->reqs_active);
>> 1149 aio_put_req(req); /* drop extra ref to req */
>> 1150 aio_put_req(req); /* drop i/o ref to req */
>> 1151 return ret;
2067 } 1152 }
2068 1153
2069 /* sys_io_submit: !! 1154 long do_io_submit(aio_context_t ctx_id, long nr,
2070 * Queue the nr iocbs pointed to by iocb !! 1155 struct iocb __user *__user *iocbpp, bool compat)
2071 * the number of iocbs queued. May retu <<
2072 * specified by ctx_id is invalid, if nr <<
2073 * *iocbpp[0] is not properly initialize <<
2074 * is invalid for the file descriptor in <<
2075 * -EFAULT if any of the data structures <<
2076 * fail with -EBADF if the file descript <<
2077 * iocb is invalid. May fail with -EAGA <<
2078 * are available to queue any iocbs. Wi <<
2079 * fail with -ENOSYS if not implemented. <<
2080 */ <<
2081 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx <<
2082 struct iocb __user * __user * <<
2083 { 1156 {
2084 struct kioctx *ctx; 1157 struct kioctx *ctx;
2085 long ret = 0; 1158 long ret = 0;
2086 int i = 0; 1159 int i = 0;
2087 struct blk_plug plug; 1160 struct blk_plug plug;
2088 1161
2089 if (unlikely(nr < 0)) 1162 if (unlikely(nr < 0))
2090 return -EINVAL; 1163 return -EINVAL;
2091 1164
>> 1165 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
>> 1166 nr = LONG_MAX/sizeof(*iocbpp);
>> 1167
>> 1168 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
>> 1169 return -EFAULT;
>> 1170
2092 ctx = lookup_ioctx(ctx_id); 1171 ctx = lookup_ioctx(ctx_id);
2093 if (unlikely(!ctx)) { 1172 if (unlikely(!ctx)) {
2094 pr_debug("EINVAL: invalid con 1173 pr_debug("EINVAL: invalid context id\n");
2095 return -EINVAL; 1174 return -EINVAL;
2096 } 1175 }
2097 1176
2098 if (nr > ctx->nr_events) !! 1177 blk_start_plug(&plug);
2099 nr = ctx->nr_events; <<
2100 1178
2101 if (nr > AIO_PLUG_THRESHOLD) !! 1179 /*
2102 blk_start_plug(&plug); !! 1180 * AKPM: should this return a partial result if some of the IOs were
2103 for (i = 0; i < nr; i++) { !! 1181 * successfully submitted?
>> 1182 */
>> 1183 for (i=0; i<nr; i++) {
2104 struct iocb __user *user_iocb 1184 struct iocb __user *user_iocb;
>> 1185 struct iocb tmp;
>> 1186
>> 1187 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
>> 1188 ret = -EFAULT;
>> 1189 break;
>> 1190 }
2105 1191
2106 if (unlikely(get_user(user_io !! 1192 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
2107 ret = -EFAULT; 1193 ret = -EFAULT;
2108 break; 1194 break;
2109 } 1195 }
2110 1196
2111 ret = io_submit_one(ctx, user !! 1197 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
2112 if (ret) 1198 if (ret)
2113 break; 1199 break;
2114 } 1200 }
2115 if (nr > AIO_PLUG_THRESHOLD) !! 1201 blk_finish_plug(&plug);
2116 blk_finish_plug(&plug); <<
2117 1202
2118 percpu_ref_put(&ctx->users); !! 1203 put_ioctx(ctx);
2119 return i ? i : ret; 1204 return i ? i : ret;
2120 } 1205 }
2121 1206
2122 #ifdef CONFIG_COMPAT !! 1207 /* sys_io_submit:
2123 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_ !! 1208 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2124 int, nr, compat_uptr_t !! 1209 * the number of iocbs queued. May return -EINVAL if the aio_context
>> 1210 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
>> 1211 * *iocbpp[0] is not properly initialized, if the operation specified
>> 1212 * is invalid for the file descriptor in the iocb. May fail with
>> 1213 * -EFAULT if any of the data structures point to invalid data. May
>> 1214 * fail with -EBADF if the file descriptor specified in the first
>> 1215 * iocb is invalid. May fail with -EAGAIN if insufficient resources
>> 1216 * are available to queue any iocbs. Will return 0 if nr is 0. Will
>> 1217 * fail with -ENOSYS if not implemented.
>> 1218 */
>> 1219 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
>> 1220 struct iocb __user * __user *, iocbpp)
2125 { 1221 {
2126 struct kioctx *ctx; !! 1222 return do_io_submit(ctx_id, nr, iocbpp, 0);
2127 long ret = 0; !! 1223 }
2128 int i = 0; <<
2129 struct blk_plug plug; <<
2130 <<
2131 if (unlikely(nr < 0)) <<
2132 return -EINVAL; <<
2133 <<
2134 ctx = lookup_ioctx(ctx_id); <<
2135 if (unlikely(!ctx)) { <<
2136 pr_debug("EINVAL: invalid con <<
2137 return -EINVAL; <<
2138 } <<
2139 1224
2140 if (nr > ctx->nr_events) !! 1225 /* lookup_kiocb
2141 nr = ctx->nr_events; !! 1226 * Finds a given iocb for cancellation.
>> 1227 */
>> 1228 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
>> 1229 u32 key)
>> 1230 {
>> 1231 struct list_head *pos;
2142 1232
2143 if (nr > AIO_PLUG_THRESHOLD) !! 1233 assert_spin_locked(&ctx->ctx_lock);
2144 blk_start_plug(&plug); <<
2145 for (i = 0; i < nr; i++) { <<
2146 compat_uptr_t user_iocb; <<
2147 1234
2148 if (unlikely(get_user(user_io !! 1235 if (key != KIOCB_KEY)
2149 ret = -EFAULT; !! 1236 return NULL;
2150 break; <<
2151 } <<
2152 1237
2153 ret = io_submit_one(ctx, comp !! 1238 /* TODO: use a hash or array, this sucks. */
2154 if (ret) !! 1239 list_for_each(pos, &ctx->active_reqs) {
2155 break; !! 1240 struct kiocb *kiocb = list_kiocb(pos);
>> 1241 if (kiocb->ki_obj.user == iocb)
>> 1242 return kiocb;
2156 } 1243 }
2157 if (nr > AIO_PLUG_THRESHOLD) !! 1244 return NULL;
2158 blk_finish_plug(&plug); <<
2159 <<
2160 percpu_ref_put(&ctx->users); <<
2161 return i ? i : ret; <<
2162 } 1245 }
2163 #endif <<
2164 1246
2165 /* sys_io_cancel: 1247 /* sys_io_cancel:
2166 * Attempts to cancel an iocb previously 1248 * Attempts to cancel an iocb previously passed to io_submit. If
2167 * the operation is successfully cancell 1249 * the operation is successfully cancelled, the resulting event is
2168 * copied into the memory pointed to by 1250 * copied into the memory pointed to by result without being placed
2169 * into the completion queue and 0 is re 1251 * into the completion queue and 0 is returned. May fail with
2170 * -EFAULT if any of the data structures 1252 * -EFAULT if any of the data structures pointed to are invalid.
2171 * May fail with -EINVAL if aio_context 1253 * May fail with -EINVAL if aio_context specified by ctx_id is
2172 * invalid. May fail with -EAGAIN if th 1254 * invalid. May fail with -EAGAIN if the iocb specified was not
2173 * cancelled. Will fail with -ENOSYS if 1255 * cancelled. Will fail with -ENOSYS if not implemented.
2174 */ 1256 */
2175 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx 1257 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2176 struct io_event __user *, res 1258 struct io_event __user *, result)
2177 { 1259 {
>> 1260 struct io_event res;
2178 struct kioctx *ctx; 1261 struct kioctx *ctx;
2179 struct aio_kiocb *kiocb; !! 1262 struct kiocb *kiocb;
2180 int ret = -EINVAL; <<
2181 u32 key; 1263 u32 key;
2182 u64 obj = (u64)(unsigned long)iocb; !! 1264 int ret;
2183 1265
2184 if (unlikely(get_user(key, &iocb->aio !! 1266 ret = get_user(key, &iocb->aio_key);
>> 1267 if (unlikely(ret))
2185 return -EFAULT; 1268 return -EFAULT;
2186 if (unlikely(key != KIOCB_KEY)) <<
2187 return -EINVAL; <<
2188 1269
2189 ctx = lookup_ioctx(ctx_id); 1270 ctx = lookup_ioctx(ctx_id);
2190 if (unlikely(!ctx)) 1271 if (unlikely(!ctx))
2191 return -EINVAL; 1272 return -EINVAL;
2192 1273
2193 spin_lock_irq(&ctx->ctx_lock); 1274 spin_lock_irq(&ctx->ctx_lock);
2194 /* TODO: use a hash or array, this su !! 1275
2195 list_for_each_entry(kiocb, &ctx->acti !! 1276 kiocb = lookup_kiocb(ctx, iocb, key);
2196 if (kiocb->ki_res.obj == obj) !! 1277 if (kiocb)
2197 ret = kiocb->ki_cance !! 1278 ret = kiocb_cancel(ctx, kiocb, &res);
2198 list_del_init(&kiocb- !! 1279 else
2199 break; !! 1280 ret = -EINVAL;
2200 } !! 1281
2201 } <<
2202 spin_unlock_irq(&ctx->ctx_lock); 1282 spin_unlock_irq(&ctx->ctx_lock);
2203 1283
2204 if (!ret) { 1284 if (!ret) {
2205 /* !! 1285 /* Cancellation succeeded -- copy the result
2206 * The result argument is no !! 1286 * into the user's buffer.
2207 * always delivered via the r <<
2208 * cancellation is progress: <<
2209 */ 1287 */
2210 ret = -EINPROGRESS; !! 1288 if (copy_to_user(result, &res, sizeof(res)))
>> 1289 ret = -EFAULT;
2211 } 1290 }
2212 1291
2213 percpu_ref_put(&ctx->users); !! 1292 put_ioctx(ctx);
2214 <<
2215 return ret; <<
2216 } <<
2217 <<
2218 static long do_io_getevents(aio_context_t ctx <<
2219 long min_nr, <<
2220 long nr, <<
2221 struct io_event __user *event <<
2222 struct timespec64 *ts) <<
2223 { <<
2224 ktime_t until = ts ? timespec64_to_kt <<
2225 struct kioctx *ioctx = lookup_ioctx(c <<
2226 long ret = -EINVAL; <<
2227 <<
2228 if (likely(ioctx)) { <<
2229 if (likely(min_nr <= nr && mi <<
2230 ret = read_events(ioc <<
2231 percpu_ref_put(&ioctx->users) <<
2232 } <<
2233 1293
2234 return ret; 1294 return ret;
2235 } 1295 }
2236 1296
2237 /* io_getevents: 1297 /* io_getevents:
2238 * Attempts to read at least min_nr even 1298 * Attempts to read at least min_nr events and up to nr events from
2239 * the completion queue for the aio_cont 1299 * the completion queue for the aio_context specified by ctx_id. If
2240 * it succeeds, the number of read event 1300 * it succeeds, the number of read events is returned. May fail with
2241 * -EINVAL if ctx_id is invalid, if min_ 1301 * -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 1302 * out of range, if timeout is out of range. May fail with -EFAULT
2243 * if any of the memory specified is inv 1303 * if any of the memory specified is invalid. May return 0 or
2244 * < min_nr if the timeout specified by 1304 * < min_nr if the timeout specified by timeout has elapsed
2245 * before sufficient events are availabl 1305 * before sufficient events are available, where timeout == NULL
2246 * specifies an infinite timeout. Note t 1306 * specifies an infinite timeout. Note that the timeout pointed to by
2247 * timeout is relative. Will fail with 1307 * timeout is relative. Will fail with -ENOSYS if not implemented.
2248 */ 1308 */
2249 #ifdef CONFIG_64BIT <<
2250 <<
2251 SYSCALL_DEFINE5(io_getevents, aio_context_t, 1309 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2252 long, min_nr, 1310 long, min_nr,
2253 long, nr, 1311 long, nr,
2254 struct io_event __user *, eve 1312 struct io_event __user *, events,
2255 struct __kernel_timespec __us !! 1313 struct timespec __user *, timeout)
2256 { <<
2257 struct timespec64 ts; <<
2258 int ret; <<
2259 <<
2260 if (timeout && unlikely(get_timespec6 <<
2261 return -EFAULT; <<
2262 <<
2263 ret = do_io_getevents(ctx_id, min_nr, <<
2264 if (!ret && signal_pending(current)) <<
2265 ret = -EINTR; <<
2266 return ret; <<
2267 } <<
2268 <<
2269 #endif <<
2270 <<
2271 struct __aio_sigset { <<
2272 const sigset_t __user *sigmask; <<
2273 size_t sigsetsize; <<
2274 }; <<
2275 <<
2276 SYSCALL_DEFINE6(io_pgetevents, <<
2277 aio_context_t, ctx_id, <<
2278 long, min_nr, <<
2279 long, nr, <<
2280 struct io_event __user *, eve <<
2281 struct __kernel_timespec __us <<
2282 const struct __aio_sigset __u <<
2283 { <<
2284 struct __aio_sigset ksig = { NULL <<
2285 struct timespec64 ts; <<
2286 bool interrupted; <<
2287 int ret; <<
2288 <<
2289 if (timeout && unlikely(get_timespec6 <<
2290 return -EFAULT; <<
2291 <<
2292 if (usig && copy_from_user(&ksig, usi <<
2293 return -EFAULT; <<
2294 <<
2295 ret = set_user_sigmask(ksig.sigmask, <<
2296 if (ret) <<
2297 return ret; <<
2298 <<
2299 ret = do_io_getevents(ctx_id, min_nr, <<
2300 <<
2301 interrupted = signal_pending(current) <<
2302 restore_saved_sigmask_unless(interrup <<
2303 if (interrupted && !ret) <<
2304 ret = -ERESTARTNOHAND; <<
2305 <<
2306 return ret; <<
2307 } <<
2308 <<
2309 #if defined(CONFIG_COMPAT_32BIT_TIME) && !def <<
2310 <<
2311 SYSCALL_DEFINE6(io_pgetevents_time32, <<
2312 aio_context_t, ctx_id, <<
2313 long, min_nr, <<
2314 long, nr, <<
2315 struct io_event __user *, eve <<
2316 struct old_timespec32 __user <<
2317 const struct __aio_sigset __u <<
2318 { 1314 {
2319 struct __aio_sigset ksig = { NULL !! 1315 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2320 struct timespec64 ts; !! 1316 long ret = -EINVAL;
2321 bool interrupted; <<
2322 int ret; <<
2323 <<
2324 if (timeout && unlikely(get_old_times <<
2325 return -EFAULT; <<
2326 <<
2327 if (usig && copy_from_user(&ksig, usi <<
2328 return -EFAULT; <<
2329 <<
2330 <<
2331 ret = set_user_sigmask(ksig.sigmask, <<
2332 if (ret) <<
2333 return ret; <<
2334 <<
2335 ret = do_io_getevents(ctx_id, min_nr, <<
2336 <<
2337 interrupted = signal_pending(current) <<
2338 restore_saved_sigmask_unless(interrup <<
2339 if (interrupted && !ret) <<
2340 ret = -ERESTARTNOHAND; <<
2341 <<
2342 return ret; <<
2343 } <<
2344 <<
2345 #endif <<
2346 <<
2347 #if defined(CONFIG_COMPAT_32BIT_TIME) <<
2348 <<
2349 SYSCALL_DEFINE5(io_getevents_time32, __u32, c <<
2350 __s32, min_nr, <<
2351 __s32, nr, <<
2352 struct io_event __user *, eve <<
2353 struct old_timespec32 __user <<
2354 { <<
2355 struct timespec64 t; <<
2356 int ret; <<
2357 <<
2358 if (timeout && get_old_timespec32(&t, <<
2359 return -EFAULT; <<
2360 <<
2361 ret = do_io_getevents(ctx_id, min_nr, <<
2362 if (!ret && signal_pending(current)) <<
2363 ret = -EINTR; <<
2364 return ret; <<
2365 } <<
2366 <<
2367 #endif <<
2368 <<
2369 #ifdef CONFIG_COMPAT <<
2370 <<
2371 struct __compat_aio_sigset { <<
2372 compat_uptr_t sigmask; <<
2373 compat_size_t sigsetsize; <<
2374 }; <<
2375 <<
2376 #if defined(CONFIG_COMPAT_32BIT_TIME) <<
2377 <<
2378 COMPAT_SYSCALL_DEFINE6(io_pgetevents, <<
2379 compat_aio_context_t, ctx_id, <<
2380 compat_long_t, min_nr, <<
2381 compat_long_t, nr, <<
2382 struct io_event __user *, eve <<
2383 struct old_timespec32 __user <<
2384 const struct __compat_aio_sig <<
2385 { <<
2386 struct __compat_aio_sigset ksig = { 0 <<
2387 struct timespec64 t; <<
2388 bool interrupted; <<
2389 int ret; <<
2390 <<
2391 if (timeout && get_old_timespec32(&t, <<
2392 return -EFAULT; <<
2393 <<
2394 if (usig && copy_from_user(&ksig, usi <<
2395 return -EFAULT; <<
2396 <<
2397 ret = set_compat_user_sigmask(compat_ <<
2398 if (ret) <<
2399 return ret; <<
2400 <<
2401 ret = do_io_getevents(ctx_id, min_nr, <<
2402 <<
2403 interrupted = signal_pending(current) <<
2404 restore_saved_sigmask_unless(interrup <<
2405 if (interrupted && !ret) <<
2406 ret = -ERESTARTNOHAND; <<
2407 <<
2408 return ret; <<
2409 } <<
2410 <<
2411 #endif <<
2412 <<
2413 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64, <<
2414 compat_aio_context_t, ctx_id, <<
2415 compat_long_t, min_nr, <<
2416 compat_long_t, nr, <<
2417 struct io_event __user *, eve <<
2418 struct __kernel_timespec __us <<
2419 const struct __compat_aio_sig <<
2420 { <<
2421 struct __compat_aio_sigset ksig = { 0 <<
2422 struct timespec64 t; <<
2423 bool interrupted; <<
2424 int ret; <<
2425 <<
2426 if (timeout && get_timespec64(&t, tim <<
2427 return -EFAULT; <<
2428 <<
2429 if (usig && copy_from_user(&ksig, usi <<
2430 return -EFAULT; <<
2431 <<
2432 ret = set_compat_user_sigmask(compat_ <<
2433 if (ret) <<
2434 return ret; <<
2435 <<
2436 ret = do_io_getevents(ctx_id, min_nr, <<
2437 <<
2438 interrupted = signal_pending(current) <<
2439 restore_saved_sigmask_unless(interrup <<
2440 if (interrupted && !ret) <<
2441 ret = -ERESTARTNOHAND; <<
2442 1317
>> 1318 if (likely(ioctx)) {
>> 1319 if (likely(min_nr <= nr && min_nr >= 0))
>> 1320 ret = read_events(ioctx, min_nr, nr, events, timeout);
>> 1321 put_ioctx(ioctx);
>> 1322 }
2443 return ret; 1323 return ret;
2444 } 1324 }
2445 #endif <<
2446 1325
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.