]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * An async IO implementation for Linux | |
3 | * Written by Benjamin LaHaise <bcrl@kvack.org> | |
4 | * | |
5 | * Implements an efficient asynchronous io interface. | |
6 | * | |
7 | * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved. | |
bfe4037e | 8 | * Copyright 2018 Christoph Hellwig. |
1da177e4 LT |
9 | * |
10 | * See ../COPYING for licensing terms. | |
11 | */ | |
caf4167a KO |
12 | #define pr_fmt(fmt) "%s: " fmt, __func__ |
13 | ||
1da177e4 LT |
14 | #include <linux/kernel.h> |
15 | #include <linux/init.h> | |
16 | #include <linux/errno.h> | |
17 | #include <linux/time.h> | |
18 | #include <linux/aio_abi.h> | |
630d9c47 | 19 | #include <linux/export.h> |
1da177e4 | 20 | #include <linux/syscalls.h> |
b9d128f1 | 21 | #include <linux/backing-dev.h> |
9018ccc4 | 22 | #include <linux/refcount.h> |
027445c3 | 23 | #include <linux/uio.h> |
1da177e4 | 24 | |
174cd4b1 | 25 | #include <linux/sched/signal.h> |
1da177e4 LT |
26 | #include <linux/fs.h> |
27 | #include <linux/file.h> | |
28 | #include <linux/mm.h> | |
29 | #include <linux/mman.h> | |
3d2d827f | 30 | #include <linux/mmu_context.h> |
e1bdd5f2 | 31 | #include <linux/percpu.h> |
1da177e4 LT |
32 | #include <linux/slab.h> |
33 | #include <linux/timer.h> | |
34 | #include <linux/aio.h> | |
35 | #include <linux/highmem.h> | |
36 | #include <linux/workqueue.h> | |
37 | #include <linux/security.h> | |
9c3060be | 38 | #include <linux/eventfd.h> |
cfb1e33e | 39 | #include <linux/blkdev.h> |
9d85cba7 | 40 | #include <linux/compat.h> |
36bc08cc GZ |
41 | #include <linux/migrate.h> |
42 | #include <linux/ramfs.h> | |
723be6e3 | 43 | #include <linux/percpu-refcount.h> |
71ad7490 | 44 | #include <linux/mount.h> |
1da177e4 LT |
45 | |
46 | #include <asm/kmap_types.h> | |
7c0f6ba6 | 47 | #include <linux/uaccess.h> |
1da177e4 | 48 | |
68d70d03 AV |
49 | #include "internal.h" |
50 | ||
f3a2752a CH |
51 | #define KIOCB_KEY 0 |
52 | ||
4e179bca KO |
53 | #define AIO_RING_MAGIC 0xa10a10a1 |
54 | #define AIO_RING_COMPAT_FEATURES 1 | |
55 | #define AIO_RING_INCOMPAT_FEATURES 0 | |
56 | struct aio_ring { | |
57 | unsigned id; /* kernel internal index number */ | |
58 | unsigned nr; /* number of io_events */ | |
fa8a53c3 BL |
59 | unsigned head; /* Written to by userland or under ring_lock |
60 | * mutex by aio_read_events_ring(). */ | |
4e179bca KO |
61 | unsigned tail; |
62 | ||
63 | unsigned magic; | |
64 | unsigned compat_features; | |
65 | unsigned incompat_features; | |
66 | unsigned header_length; /* size of aio_ring */ | |
67 | ||
68 | ||
69 | struct io_event io_events[0]; | |
70 | }; /* 128 bytes + ring size */ | |
71 | ||
72 | #define AIO_RING_PAGES 8 | |
4e179bca | 73 | |
db446a08 | 74 | struct kioctx_table { |
d0264c01 TH |
75 | struct rcu_head rcu; |
76 | unsigned nr; | |
77 | struct kioctx __rcu *table[]; | |
db446a08 BL |
78 | }; |
79 | ||
e1bdd5f2 KO |
80 | struct kioctx_cpu { |
81 | unsigned reqs_available; | |
82 | }; | |
83 | ||
dc48e56d JA |
84 | struct ctx_rq_wait { |
85 | struct completion comp; | |
86 | atomic_t count; | |
87 | }; | |
88 | ||
4e179bca | 89 | struct kioctx { |
723be6e3 | 90 | struct percpu_ref users; |
36f55889 | 91 | atomic_t dead; |
4e179bca | 92 | |
e34ecee2 KO |
93 | struct percpu_ref reqs; |
94 | ||
4e179bca | 95 | unsigned long user_id; |
4e179bca | 96 | |
e1bdd5f2 KO |
97 | struct __percpu kioctx_cpu *cpu; |
98 | ||
99 | /* | |
100 | * For percpu reqs_available, number of slots we move to/from global | |
101 | * counter at a time: | |
102 | */ | |
103 | unsigned req_batch; | |
3e845ce0 KO |
104 | /* |
105 | * This is what userspace passed to io_setup(), it's not used for | |
106 | * anything but counting against the global max_reqs quota. | |
107 | * | |
58c85dc2 | 108 | * The real limit is nr_events - 1, which will be larger (see |
3e845ce0 KO |
109 | * aio_setup_ring()) |
110 | */ | |
4e179bca KO |
111 | unsigned max_reqs; |
112 | ||
58c85dc2 KO |
113 | /* Size of ringbuffer, in units of struct io_event */ |
114 | unsigned nr_events; | |
4e179bca | 115 | |
58c85dc2 KO |
116 | unsigned long mmap_base; |
117 | unsigned long mmap_size; | |
118 | ||
119 | struct page **ring_pages; | |
120 | long nr_pages; | |
121 | ||
f729863a | 122 | struct rcu_work free_rwork; /* see free_ioctx() */ |
4e23bcae | 123 | |
e02ba72a AP |
124 | /* |
125 | * signals when all in-flight requests are done | |
126 | */ | |
dc48e56d | 127 | struct ctx_rq_wait *rq_wait; |
e02ba72a | 128 | |
4e23bcae | 129 | struct { |
34e83fc6 KO |
130 | /* |
131 | * This counts the number of available slots in the ringbuffer, | |
132 | * so we avoid overflowing it: it's decremented (if positive) | |
133 | * when allocating a kiocb and incremented when the resulting | |
134 | * io_event is pulled off the ringbuffer. | |
e1bdd5f2 KO |
135 | * |
136 | * We batch accesses to it with a percpu version. | |
34e83fc6 KO |
137 | */ |
138 | atomic_t reqs_available; | |
4e23bcae KO |
139 | } ____cacheline_aligned_in_smp; |
140 | ||
141 | struct { | |
142 | spinlock_t ctx_lock; | |
143 | struct list_head active_reqs; /* used for cancellation */ | |
144 | } ____cacheline_aligned_in_smp; | |
145 | ||
58c85dc2 KO |
146 | struct { |
147 | struct mutex ring_lock; | |
4e23bcae KO |
148 | wait_queue_head_t wait; |
149 | } ____cacheline_aligned_in_smp; | |
58c85dc2 KO |
150 | |
151 | struct { | |
152 | unsigned tail; | |
d856f32a | 153 | unsigned completed_events; |
58c85dc2 | 154 | spinlock_t completion_lock; |
4e23bcae | 155 | } ____cacheline_aligned_in_smp; |
58c85dc2 KO |
156 | |
157 | struct page *internal_pages[AIO_RING_PAGES]; | |
36bc08cc | 158 | struct file *aio_ring_file; |
db446a08 BL |
159 | |
160 | unsigned id; | |
4e179bca KO |
161 | }; |
162 | ||
a3c0d439 CH |
163 | struct fsync_iocb { |
164 | struct work_struct work; | |
165 | struct file *file; | |
166 | bool datasync; | |
167 | }; | |
168 | ||
bfe4037e CH |
169 | struct poll_iocb { |
170 | struct file *file; | |
171 | struct wait_queue_head *head; | |
172 | __poll_t events; | |
173 | bool woken; | |
174 | bool cancelled; | |
175 | struct wait_queue_entry wait; | |
176 | struct work_struct work; | |
177 | }; | |
178 | ||
04b2fa9f | 179 | struct aio_kiocb { |
54843f87 CH |
180 | union { |
181 | struct kiocb rw; | |
a3c0d439 | 182 | struct fsync_iocb fsync; |
bfe4037e | 183 | struct poll_iocb poll; |
54843f87 | 184 | }; |
04b2fa9f CH |
185 | |
186 | struct kioctx *ki_ctx; | |
187 | kiocb_cancel_fn *ki_cancel; | |
188 | ||
189 | struct iocb __user *ki_user_iocb; /* user's aiocb */ | |
190 | __u64 ki_user_data; /* user's data for completion */ | |
191 | ||
192 | struct list_head ki_list; /* the aio core uses this | |
193 | * for cancellation */ | |
9018ccc4 | 194 | refcount_t ki_refcnt; |
04b2fa9f CH |
195 | |
196 | /* | |
197 | * If the aio_resfd field of the userspace iocb is not zero, | |
198 | * this is the underlying eventfd context to deliver events to. | |
199 | */ | |
200 | struct eventfd_ctx *ki_eventfd; | |
201 | }; | |
202 | ||
1da177e4 | 203 | /*------ sysctl variables----*/ |
d55b5fda ZB |
204 | static DEFINE_SPINLOCK(aio_nr_lock); |
205 | unsigned long aio_nr; /* current system wide number of aio requests */ | |
206 | unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ | |
1da177e4 LT |
207 | /*----end sysctl variables---*/ |
208 | ||
e18b890b CL |
209 | static struct kmem_cache *kiocb_cachep; |
210 | static struct kmem_cache *kioctx_cachep; | |
1da177e4 | 211 | |
71ad7490 BL |
212 | static struct vfsmount *aio_mnt; |
213 | ||
214 | static const struct file_operations aio_ring_fops; | |
215 | static const struct address_space_operations aio_ctx_aops; | |
216 | ||
217 | static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages) | |
218 | { | |
219 | struct qstr this = QSTR_INIT("[aio]", 5); | |
220 | struct file *file; | |
221 | struct path path; | |
222 | struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb); | |
7f62656b DC |
223 | if (IS_ERR(inode)) |
224 | return ERR_CAST(inode); | |
71ad7490 BL |
225 | |
226 | inode->i_mapping->a_ops = &aio_ctx_aops; | |
227 | inode->i_mapping->private_data = ctx; | |
228 | inode->i_size = PAGE_SIZE * nr_pages; | |
229 | ||
230 | path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this); | |
231 | if (!path.dentry) { | |
232 | iput(inode); | |
233 | return ERR_PTR(-ENOMEM); | |
234 | } | |
235 | path.mnt = mntget(aio_mnt); | |
236 | ||
237 | d_instantiate(path.dentry, inode); | |
238 | file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops); | |
239 | if (IS_ERR(file)) { | |
240 | path_put(&path); | |
241 | return file; | |
242 | } | |
243 | ||
244 | file->f_flags = O_RDWR; | |
71ad7490 BL |
245 | return file; |
246 | } | |
247 | ||
248 | static struct dentry *aio_mount(struct file_system_type *fs_type, | |
249 | int flags, const char *dev_name, void *data) | |
250 | { | |
251 | static const struct dentry_operations ops = { | |
252 | .d_dname = simple_dname, | |
253 | }; | |
22f6b4d3 JH |
254 | struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops, |
255 | AIO_RING_MAGIC); | |
256 | ||
257 | if (!IS_ERR(root)) | |
258 | root->d_sb->s_iflags |= SB_I_NOEXEC; | |
259 | return root; | |
71ad7490 BL |
260 | } |
261 | ||
1da177e4 LT |
262 | /* aio_setup |
263 | * Creates the slab caches used by the aio routines, panic on | |
264 | * failure as this is done early during the boot sequence. | |
265 | */ | |
266 | static int __init aio_setup(void) | |
267 | { | |
71ad7490 BL |
268 | static struct file_system_type aio_fs = { |
269 | .name = "aio", | |
270 | .mount = aio_mount, | |
271 | .kill_sb = kill_anon_super, | |
272 | }; | |
273 | aio_mnt = kern_mount(&aio_fs); | |
274 | if (IS_ERR(aio_mnt)) | |
275 | panic("Failed to create aio fs mount."); | |
276 | ||
04b2fa9f | 277 | kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
0a31bd5f | 278 | kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
1da177e4 LT |
279 | return 0; |
280 | } | |
385773e0 | 281 | __initcall(aio_setup); |
1da177e4 | 282 | |
5e9ae2e5 BL |
283 | static void put_aio_ring_file(struct kioctx *ctx) |
284 | { | |
285 | struct file *aio_ring_file = ctx->aio_ring_file; | |
de04e769 RV |
286 | struct address_space *i_mapping; |
287 | ||
5e9ae2e5 | 288 | if (aio_ring_file) { |
45063097 | 289 | truncate_setsize(file_inode(aio_ring_file), 0); |
5e9ae2e5 BL |
290 | |
291 | /* Prevent further access to the kioctx from migratepages */ | |
45063097 | 292 | i_mapping = aio_ring_file->f_mapping; |
de04e769 RV |
293 | spin_lock(&i_mapping->private_lock); |
294 | i_mapping->private_data = NULL; | |
5e9ae2e5 | 295 | ctx->aio_ring_file = NULL; |
de04e769 | 296 | spin_unlock(&i_mapping->private_lock); |
5e9ae2e5 BL |
297 | |
298 | fput(aio_ring_file); | |
299 | } | |
300 | } | |
301 | ||
1da177e4 LT |
302 | static void aio_free_ring(struct kioctx *ctx) |
303 | { | |
36bc08cc | 304 | int i; |
1da177e4 | 305 | |
fa8a53c3 BL |
306 | /* Disconnect the kiotx from the ring file. This prevents future |
307 | * accesses to the kioctx from page migration. | |
308 | */ | |
309 | put_aio_ring_file(ctx); | |
310 | ||
36bc08cc | 311 | for (i = 0; i < ctx->nr_pages; i++) { |
8e321fef | 312 | struct page *page; |
36bc08cc GZ |
313 | pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i, |
314 | page_count(ctx->ring_pages[i])); | |
8e321fef BL |
315 | page = ctx->ring_pages[i]; |
316 | if (!page) | |
317 | continue; | |
318 | ctx->ring_pages[i] = NULL; | |
319 | put_page(page); | |
36bc08cc | 320 | } |
1da177e4 | 321 | |
ddb8c45b | 322 | if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) { |
58c85dc2 | 323 | kfree(ctx->ring_pages); |
ddb8c45b SL |
324 | ctx->ring_pages = NULL; |
325 | } | |
36bc08cc GZ |
326 | } |
327 | ||
5477e70a | 328 | static int aio_ring_mremap(struct vm_area_struct *vma) |
e4a0d3e7 | 329 | { |
5477e70a | 330 | struct file *file = vma->vm_file; |
e4a0d3e7 PE |
331 | struct mm_struct *mm = vma->vm_mm; |
332 | struct kioctx_table *table; | |
b2edffdd | 333 | int i, res = -EINVAL; |
e4a0d3e7 PE |
334 | |
335 | spin_lock(&mm->ioctx_lock); | |
336 | rcu_read_lock(); | |
337 | table = rcu_dereference(mm->ioctx_table); | |
338 | for (i = 0; i < table->nr; i++) { | |
339 | struct kioctx *ctx; | |
340 | ||
d0264c01 | 341 | ctx = rcu_dereference(table->table[i]); |
e4a0d3e7 | 342 | if (ctx && ctx->aio_ring_file == file) { |
b2edffdd AV |
343 | if (!atomic_read(&ctx->dead)) { |
344 | ctx->user_id = ctx->mmap_base = vma->vm_start; | |
345 | res = 0; | |
346 | } | |
e4a0d3e7 PE |
347 | break; |
348 | } | |
349 | } | |
350 | ||
351 | rcu_read_unlock(); | |
352 | spin_unlock(&mm->ioctx_lock); | |
b2edffdd | 353 | return res; |
e4a0d3e7 PE |
354 | } |
355 | ||
5477e70a ON |
356 | static const struct vm_operations_struct aio_ring_vm_ops = { |
357 | .mremap = aio_ring_mremap, | |
358 | #if IS_ENABLED(CONFIG_MMU) | |
359 | .fault = filemap_fault, | |
360 | .map_pages = filemap_map_pages, | |
361 | .page_mkwrite = filemap_page_mkwrite, | |
362 | #endif | |
363 | }; | |
364 | ||
365 | static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma) | |
366 | { | |
367 | vma->vm_flags |= VM_DONTEXPAND; | |
368 | vma->vm_ops = &aio_ring_vm_ops; | |
369 | return 0; | |
370 | } | |
371 | ||
36bc08cc GZ |
372 | static const struct file_operations aio_ring_fops = { |
373 | .mmap = aio_ring_mmap, | |
374 | }; | |
375 | ||
0c45355f | 376 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc GZ |
377 | static int aio_migratepage(struct address_space *mapping, struct page *new, |
378 | struct page *old, enum migrate_mode mode) | |
379 | { | |
5e9ae2e5 | 380 | struct kioctx *ctx; |
36bc08cc | 381 | unsigned long flags; |
fa8a53c3 | 382 | pgoff_t idx; |
36bc08cc GZ |
383 | int rc; |
384 | ||
2916ecc0 JG |
385 | /* |
386 | * We cannot support the _NO_COPY case here, because copy needs to | |
387 | * happen under the ctx->completion_lock. That does not work with the | |
388 | * migration workflow of MIGRATE_SYNC_NO_COPY. | |
389 | */ | |
390 | if (mode == MIGRATE_SYNC_NO_COPY) | |
391 | return -EINVAL; | |
392 | ||
8e321fef BL |
393 | rc = 0; |
394 | ||
fa8a53c3 | 395 | /* mapping->private_lock here protects against the kioctx teardown. */ |
8e321fef BL |
396 | spin_lock(&mapping->private_lock); |
397 | ctx = mapping->private_data; | |
fa8a53c3 BL |
398 | if (!ctx) { |
399 | rc = -EINVAL; | |
400 | goto out; | |
401 | } | |
402 | ||
403 | /* The ring_lock mutex. The prevents aio_read_events() from writing | |
404 | * to the ring's head, and prevents page migration from mucking in | |
405 | * a partially initialized kiotx. | |
406 | */ | |
407 | if (!mutex_trylock(&ctx->ring_lock)) { | |
408 | rc = -EAGAIN; | |
409 | goto out; | |
410 | } | |
411 | ||
412 | idx = old->index; | |
413 | if (idx < (pgoff_t)ctx->nr_pages) { | |
414 | /* Make sure the old page hasn't already been changed */ | |
415 | if (ctx->ring_pages[idx] != old) | |
416 | rc = -EAGAIN; | |
8e321fef BL |
417 | } else |
418 | rc = -EINVAL; | |
8e321fef BL |
419 | |
420 | if (rc != 0) | |
fa8a53c3 | 421 | goto out_unlock; |
8e321fef | 422 | |
36bc08cc GZ |
423 | /* Writeback must be complete */ |
424 | BUG_ON(PageWriteback(old)); | |
8e321fef | 425 | get_page(new); |
36bc08cc | 426 | |
8e321fef | 427 | rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1); |
36bc08cc | 428 | if (rc != MIGRATEPAGE_SUCCESS) { |
8e321fef | 429 | put_page(new); |
fa8a53c3 | 430 | goto out_unlock; |
36bc08cc GZ |
431 | } |
432 | ||
fa8a53c3 BL |
433 | /* Take completion_lock to prevent other writes to the ring buffer |
434 | * while the old page is copied to the new. This prevents new | |
435 | * events from being lost. | |
5e9ae2e5 | 436 | */ |
fa8a53c3 BL |
437 | spin_lock_irqsave(&ctx->completion_lock, flags); |
438 | migrate_page_copy(new, old); | |
439 | BUG_ON(ctx->ring_pages[idx] != old); | |
440 | ctx->ring_pages[idx] = new; | |
441 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
36bc08cc | 442 | |
fa8a53c3 BL |
443 | /* The old page is no longer accessible. */ |
444 | put_page(old); | |
8e321fef | 445 | |
fa8a53c3 BL |
446 | out_unlock: |
447 | mutex_unlock(&ctx->ring_lock); | |
448 | out: | |
449 | spin_unlock(&mapping->private_lock); | |
36bc08cc | 450 | return rc; |
1da177e4 | 451 | } |
0c45355f | 452 | #endif |
1da177e4 | 453 | |
36bc08cc | 454 | static const struct address_space_operations aio_ctx_aops = { |
835f252c | 455 | .set_page_dirty = __set_page_dirty_no_writeback, |
0c45355f | 456 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc | 457 | .migratepage = aio_migratepage, |
0c45355f | 458 | #endif |
36bc08cc GZ |
459 | }; |
460 | ||
2a8a9867 | 461 | static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events) |
1da177e4 LT |
462 | { |
463 | struct aio_ring *ring; | |
41003a7b | 464 | struct mm_struct *mm = current->mm; |
3dc9acb6 | 465 | unsigned long size, unused; |
1da177e4 | 466 | int nr_pages; |
36bc08cc GZ |
467 | int i; |
468 | struct file *file; | |
1da177e4 LT |
469 | |
470 | /* Compensate for the ring buffer's head/tail overlap entry */ | |
471 | nr_events += 2; /* 1 is required, 2 for good luck */ | |
472 | ||
473 | size = sizeof(struct aio_ring); | |
474 | size += sizeof(struct io_event) * nr_events; | |
1da177e4 | 475 | |
36bc08cc | 476 | nr_pages = PFN_UP(size); |
1da177e4 LT |
477 | if (nr_pages < 0) |
478 | return -EINVAL; | |
479 | ||
71ad7490 | 480 | file = aio_private_file(ctx, nr_pages); |
36bc08cc GZ |
481 | if (IS_ERR(file)) { |
482 | ctx->aio_ring_file = NULL; | |
fa8a53c3 | 483 | return -ENOMEM; |
36bc08cc GZ |
484 | } |
485 | ||
3dc9acb6 LT |
486 | ctx->aio_ring_file = file; |
487 | nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) | |
488 | / sizeof(struct io_event); | |
489 | ||
490 | ctx->ring_pages = ctx->internal_pages; | |
491 | if (nr_pages > AIO_RING_PAGES) { | |
492 | ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *), | |
493 | GFP_KERNEL); | |
494 | if (!ctx->ring_pages) { | |
495 | put_aio_ring_file(ctx); | |
496 | return -ENOMEM; | |
497 | } | |
498 | } | |
499 | ||
36bc08cc GZ |
500 | for (i = 0; i < nr_pages; i++) { |
501 | struct page *page; | |
45063097 | 502 | page = find_or_create_page(file->f_mapping, |
36bc08cc GZ |
503 | i, GFP_HIGHUSER | __GFP_ZERO); |
504 | if (!page) | |
505 | break; | |
506 | pr_debug("pid(%d) page[%d]->count=%d\n", | |
507 | current->pid, i, page_count(page)); | |
508 | SetPageUptodate(page); | |
36bc08cc | 509 | unlock_page(page); |
3dc9acb6 LT |
510 | |
511 | ctx->ring_pages[i] = page; | |
36bc08cc | 512 | } |
3dc9acb6 | 513 | ctx->nr_pages = i; |
1da177e4 | 514 | |
3dc9acb6 LT |
515 | if (unlikely(i != nr_pages)) { |
516 | aio_free_ring(ctx); | |
fa8a53c3 | 517 | return -ENOMEM; |
1da177e4 LT |
518 | } |
519 | ||
58c85dc2 KO |
520 | ctx->mmap_size = nr_pages * PAGE_SIZE; |
521 | pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size); | |
36bc08cc | 522 | |
013373e8 MH |
523 | if (down_write_killable(&mm->mmap_sem)) { |
524 | ctx->mmap_size = 0; | |
525 | aio_free_ring(ctx); | |
526 | return -EINTR; | |
527 | } | |
528 | ||
36bc08cc GZ |
529 | ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size, |
530 | PROT_READ | PROT_WRITE, | |
897ab3e0 | 531 | MAP_SHARED, 0, &unused, NULL); |
3dc9acb6 | 532 | up_write(&mm->mmap_sem); |
58c85dc2 | 533 | if (IS_ERR((void *)ctx->mmap_base)) { |
58c85dc2 | 534 | ctx->mmap_size = 0; |
1da177e4 | 535 | aio_free_ring(ctx); |
fa8a53c3 | 536 | return -ENOMEM; |
1da177e4 LT |
537 | } |
538 | ||
58c85dc2 | 539 | pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base); |
d6c355c7 | 540 | |
58c85dc2 KO |
541 | ctx->user_id = ctx->mmap_base; |
542 | ctx->nr_events = nr_events; /* trusted copy */ | |
1da177e4 | 543 | |
58c85dc2 | 544 | ring = kmap_atomic(ctx->ring_pages[0]); |
1da177e4 | 545 | ring->nr = nr_events; /* user copy */ |
db446a08 | 546 | ring->id = ~0U; |
1da177e4 LT |
547 | ring->head = ring->tail = 0; |
548 | ring->magic = AIO_RING_MAGIC; | |
549 | ring->compat_features = AIO_RING_COMPAT_FEATURES; | |
550 | ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; | |
551 | ring->header_length = sizeof(struct aio_ring); | |
e8e3c3d6 | 552 | kunmap_atomic(ring); |
58c85dc2 | 553 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 LT |
554 | |
555 | return 0; | |
556 | } | |
557 | ||
1da177e4 LT |
558 | #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) |
559 | #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) | |
560 | #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) | |
561 | ||
04b2fa9f | 562 | void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel) |
0460fef2 | 563 | { |
54843f87 | 564 | struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw); |
0460fef2 KO |
565 | struct kioctx *ctx = req->ki_ctx; |
566 | unsigned long flags; | |
567 | ||
75321b50 CH |
568 | if (WARN_ON_ONCE(!list_empty(&req->ki_list))) |
569 | return; | |
0460fef2 | 570 | |
75321b50 CH |
571 | spin_lock_irqsave(&ctx->ctx_lock, flags); |
572 | list_add_tail(&req->ki_list, &ctx->active_reqs); | |
0460fef2 | 573 | req->ki_cancel = cancel; |
0460fef2 KO |
574 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); |
575 | } | |
576 | EXPORT_SYMBOL(kiocb_set_cancel_fn); | |
577 | ||
a6d7cff4 TH |
578 | /* |
579 | * free_ioctx() should be RCU delayed to synchronize against the RCU | |
580 | * protected lookup_ioctx() and also needs process context to call | |
f729863a | 581 | * aio_free_ring(). Use rcu_work. |
a6d7cff4 | 582 | */ |
e34ecee2 | 583 | static void free_ioctx(struct work_struct *work) |
36f55889 | 584 | { |
f729863a TH |
585 | struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx, |
586 | free_rwork); | |
e34ecee2 | 587 | pr_debug("freeing %p\n", ctx); |
e1bdd5f2 | 588 | |
e34ecee2 | 589 | aio_free_ring(ctx); |
e1bdd5f2 | 590 | free_percpu(ctx->cpu); |
9a1049da TH |
591 | percpu_ref_exit(&ctx->reqs); |
592 | percpu_ref_exit(&ctx->users); | |
36f55889 KO |
593 | kmem_cache_free(kioctx_cachep, ctx); |
594 | } | |
595 | ||
e34ecee2 KO |
596 | static void free_ioctx_reqs(struct percpu_ref *ref) |
597 | { | |
598 | struct kioctx *ctx = container_of(ref, struct kioctx, reqs); | |
599 | ||
e02ba72a | 600 | /* At this point we know that there are no any in-flight requests */ |
dc48e56d JA |
601 | if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count)) |
602 | complete(&ctx->rq_wait->comp); | |
e02ba72a | 603 | |
a6d7cff4 | 604 | /* Synchronize against RCU protected table->table[] dereferences */ |
f729863a TH |
605 | INIT_RCU_WORK(&ctx->free_rwork, free_ioctx); |
606 | queue_rcu_work(system_wq, &ctx->free_rwork); | |
e34ecee2 KO |
607 | } |
608 | ||
36f55889 KO |
609 | /* |
610 | * When this function runs, the kioctx has been removed from the "hash table" | |
611 | * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - | |
612 | * now it's safe to cancel any that need to be. | |
613 | */ | |
e34ecee2 | 614 | static void free_ioctx_users(struct percpu_ref *ref) |
36f55889 | 615 | { |
e34ecee2 | 616 | struct kioctx *ctx = container_of(ref, struct kioctx, users); |
04b2fa9f | 617 | struct aio_kiocb *req; |
36f55889 KO |
618 | |
619 | spin_lock_irq(&ctx->ctx_lock); | |
620 | ||
621 | while (!list_empty(&ctx->active_reqs)) { | |
622 | req = list_first_entry(&ctx->active_reqs, | |
04b2fa9f | 623 | struct aio_kiocb, ki_list); |
888933f8 | 624 | req->ki_cancel(&req->rw); |
4faa9996 | 625 | list_del_init(&req->ki_list); |
36f55889 KO |
626 | } |
627 | ||
628 | spin_unlock_irq(&ctx->ctx_lock); | |
629 | ||
e34ecee2 KO |
630 | percpu_ref_kill(&ctx->reqs); |
631 | percpu_ref_put(&ctx->reqs); | |
36f55889 KO |
632 | } |
633 | ||
db446a08 BL |
634 | static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm) |
635 | { | |
636 | unsigned i, new_nr; | |
637 | struct kioctx_table *table, *old; | |
638 | struct aio_ring *ring; | |
639 | ||
640 | spin_lock(&mm->ioctx_lock); | |
855ef0de | 641 | table = rcu_dereference_raw(mm->ioctx_table); |
db446a08 BL |
642 | |
643 | while (1) { | |
644 | if (table) | |
645 | for (i = 0; i < table->nr; i++) | |
d0264c01 | 646 | if (!rcu_access_pointer(table->table[i])) { |
db446a08 | 647 | ctx->id = i; |
d0264c01 | 648 | rcu_assign_pointer(table->table[i], ctx); |
db446a08 BL |
649 | spin_unlock(&mm->ioctx_lock); |
650 | ||
fa8a53c3 BL |
651 | /* While kioctx setup is in progress, |
652 | * we are protected from page migration | |
653 | * changes ring_pages by ->ring_lock. | |
654 | */ | |
db446a08 BL |
655 | ring = kmap_atomic(ctx->ring_pages[0]); |
656 | ring->id = ctx->id; | |
657 | kunmap_atomic(ring); | |
658 | return 0; | |
659 | } | |
660 | ||
661 | new_nr = (table ? table->nr : 1) * 4; | |
db446a08 BL |
662 | spin_unlock(&mm->ioctx_lock); |
663 | ||
664 | table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) * | |
665 | new_nr, GFP_KERNEL); | |
666 | if (!table) | |
667 | return -ENOMEM; | |
668 | ||
669 | table->nr = new_nr; | |
670 | ||
671 | spin_lock(&mm->ioctx_lock); | |
855ef0de | 672 | old = rcu_dereference_raw(mm->ioctx_table); |
db446a08 BL |
673 | |
674 | if (!old) { | |
675 | rcu_assign_pointer(mm->ioctx_table, table); | |
676 | } else if (table->nr > old->nr) { | |
677 | memcpy(table->table, old->table, | |
678 | old->nr * sizeof(struct kioctx *)); | |
679 | ||
680 | rcu_assign_pointer(mm->ioctx_table, table); | |
681 | kfree_rcu(old, rcu); | |
682 | } else { | |
683 | kfree(table); | |
684 | table = old; | |
685 | } | |
686 | } | |
687 | } | |
688 | ||
e34ecee2 KO |
689 | static void aio_nr_sub(unsigned nr) |
690 | { | |
691 | spin_lock(&aio_nr_lock); | |
692 | if (WARN_ON(aio_nr - nr > aio_nr)) | |
693 | aio_nr = 0; | |
694 | else | |
695 | aio_nr -= nr; | |
696 | spin_unlock(&aio_nr_lock); | |
697 | } | |
698 | ||
1da177e4 LT |
699 | /* ioctx_alloc |
700 | * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. | |
701 | */ | |
702 | static struct kioctx *ioctx_alloc(unsigned nr_events) | |
703 | { | |
41003a7b | 704 | struct mm_struct *mm = current->mm; |
1da177e4 | 705 | struct kioctx *ctx; |
e23754f8 | 706 | int err = -ENOMEM; |
1da177e4 | 707 | |
2a8a9867 MFO |
708 | /* |
709 | * Store the original nr_events -- what userspace passed to io_setup(), | |
710 | * for counting against the global limit -- before it changes. | |
711 | */ | |
712 | unsigned int max_reqs = nr_events; | |
713 | ||
e1bdd5f2 KO |
714 | /* |
715 | * We keep track of the number of available ringbuffer slots, to prevent | |
716 | * overflow (reqs_available), and we also use percpu counters for this. | |
717 | * | |
718 | * So since up to half the slots might be on other cpu's percpu counters | |
719 | * and unavailable, double nr_events so userspace sees what they | |
720 | * expected: additionally, we move req_batch slots to/from percpu | |
721 | * counters at a time, so make sure that isn't 0: | |
722 | */ | |
723 | nr_events = max(nr_events, num_possible_cpus() * 4); | |
724 | nr_events *= 2; | |
725 | ||
1da177e4 | 726 | /* Prevent overflows */ |
08397acd | 727 | if (nr_events > (0x10000000U / sizeof(struct io_event))) { |
1da177e4 LT |
728 | pr_debug("ENOMEM: nr_events too high\n"); |
729 | return ERR_PTR(-EINVAL); | |
730 | } | |
731 | ||
2a8a9867 | 732 | if (!nr_events || (unsigned long)max_reqs > aio_max_nr) |
1da177e4 LT |
733 | return ERR_PTR(-EAGAIN); |
734 | ||
c3762229 | 735 | ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); |
1da177e4 LT |
736 | if (!ctx) |
737 | return ERR_PTR(-ENOMEM); | |
738 | ||
2a8a9867 | 739 | ctx->max_reqs = max_reqs; |
1da177e4 | 740 | |
1da177e4 | 741 | spin_lock_init(&ctx->ctx_lock); |
0460fef2 | 742 | spin_lock_init(&ctx->completion_lock); |
58c85dc2 | 743 | mutex_init(&ctx->ring_lock); |
fa8a53c3 BL |
744 | /* Protect against page migration throughout kiotx setup by keeping |
745 | * the ring_lock mutex held until setup is complete. */ | |
746 | mutex_lock(&ctx->ring_lock); | |
1da177e4 LT |
747 | init_waitqueue_head(&ctx->wait); |
748 | ||
749 | INIT_LIST_HEAD(&ctx->active_reqs); | |
1da177e4 | 750 | |
2aad2a86 | 751 | if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL)) |
fa8a53c3 BL |
752 | goto err; |
753 | ||
2aad2a86 | 754 | if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL)) |
fa8a53c3 BL |
755 | goto err; |
756 | ||
e1bdd5f2 KO |
757 | ctx->cpu = alloc_percpu(struct kioctx_cpu); |
758 | if (!ctx->cpu) | |
e34ecee2 | 759 | goto err; |
1da177e4 | 760 | |
2a8a9867 | 761 | err = aio_setup_ring(ctx, nr_events); |
fa8a53c3 | 762 | if (err < 0) |
e34ecee2 | 763 | goto err; |
e1bdd5f2 | 764 | |
34e83fc6 | 765 | atomic_set(&ctx->reqs_available, ctx->nr_events - 1); |
e1bdd5f2 | 766 | ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4); |
6878ea72 BL |
767 | if (ctx->req_batch < 1) |
768 | ctx->req_batch = 1; | |
34e83fc6 | 769 | |
1da177e4 | 770 | /* limit the number of system wide aios */ |
9fa1cb39 | 771 | spin_lock(&aio_nr_lock); |
2a8a9867 MFO |
772 | if (aio_nr + ctx->max_reqs > aio_max_nr || |
773 | aio_nr + ctx->max_reqs < aio_nr) { | |
9fa1cb39 | 774 | spin_unlock(&aio_nr_lock); |
e34ecee2 | 775 | err = -EAGAIN; |
d1b94327 | 776 | goto err_ctx; |
2dd542b7 AV |
777 | } |
778 | aio_nr += ctx->max_reqs; | |
9fa1cb39 | 779 | spin_unlock(&aio_nr_lock); |
1da177e4 | 780 | |
1881686f BL |
781 | percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */ |
782 | percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */ | |
723be6e3 | 783 | |
da90382c BL |
784 | err = ioctx_add_table(ctx, mm); |
785 | if (err) | |
e34ecee2 | 786 | goto err_cleanup; |
da90382c | 787 | |
fa8a53c3 BL |
788 | /* Release the ring_lock mutex now that all setup is complete. */ |
789 | mutex_unlock(&ctx->ring_lock); | |
790 | ||
caf4167a | 791 | pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", |
58c85dc2 | 792 | ctx, ctx->user_id, mm, ctx->nr_events); |
1da177e4 LT |
793 | return ctx; |
794 | ||
e34ecee2 KO |
795 | err_cleanup: |
796 | aio_nr_sub(ctx->max_reqs); | |
d1b94327 | 797 | err_ctx: |
deeb8525 AV |
798 | atomic_set(&ctx->dead, 1); |
799 | if (ctx->mmap_size) | |
800 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
d1b94327 | 801 | aio_free_ring(ctx); |
e34ecee2 | 802 | err: |
fa8a53c3 | 803 | mutex_unlock(&ctx->ring_lock); |
e1bdd5f2 | 804 | free_percpu(ctx->cpu); |
9a1049da TH |
805 | percpu_ref_exit(&ctx->reqs); |
806 | percpu_ref_exit(&ctx->users); | |
1da177e4 | 807 | kmem_cache_free(kioctx_cachep, ctx); |
caf4167a | 808 | pr_debug("error allocating ioctx %d\n", err); |
e23754f8 | 809 | return ERR_PTR(err); |
1da177e4 LT |
810 | } |
811 | ||
36f55889 KO |
812 | /* kill_ioctx |
813 | * Cancels all outstanding aio requests on an aio context. Used | |
814 | * when the processes owning a context have all exited to encourage | |
815 | * the rapid destruction of the kioctx. | |
816 | */ | |
fb2d4483 | 817 | static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx, |
dc48e56d | 818 | struct ctx_rq_wait *wait) |
36f55889 | 819 | { |
fa88b6f8 | 820 | struct kioctx_table *table; |
db446a08 | 821 | |
b2edffdd AV |
822 | spin_lock(&mm->ioctx_lock); |
823 | if (atomic_xchg(&ctx->dead, 1)) { | |
824 | spin_unlock(&mm->ioctx_lock); | |
fa88b6f8 | 825 | return -EINVAL; |
b2edffdd | 826 | } |
db446a08 | 827 | |
855ef0de | 828 | table = rcu_dereference_raw(mm->ioctx_table); |
d0264c01 TH |
829 | WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id])); |
830 | RCU_INIT_POINTER(table->table[ctx->id], NULL); | |
fa88b6f8 | 831 | spin_unlock(&mm->ioctx_lock); |
4fcc712f | 832 | |
a6d7cff4 | 833 | /* free_ioctx_reqs() will do the necessary RCU synchronization */ |
fa88b6f8 | 834 | wake_up_all(&ctx->wait); |
4fcc712f | 835 | |
fa88b6f8 BL |
836 | /* |
837 | * It'd be more correct to do this in free_ioctx(), after all | |
838 | * the outstanding kiocbs have finished - but by then io_destroy | |
839 | * has already returned, so io_setup() could potentially return | |
840 | * -EAGAIN with no ioctxs actually in use (as far as userspace | |
841 | * could tell). | |
842 | */ | |
843 | aio_nr_sub(ctx->max_reqs); | |
4fcc712f | 844 | |
fa88b6f8 BL |
845 | if (ctx->mmap_size) |
846 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
fb2d4483 | 847 | |
dc48e56d | 848 | ctx->rq_wait = wait; |
fa88b6f8 BL |
849 | percpu_ref_kill(&ctx->users); |
850 | return 0; | |
1da177e4 LT |
851 | } |
852 | ||
36f55889 KO |
853 | /* |
854 | * exit_aio: called when the last user of mm goes away. At this point, there is | |
855 | * no way for any new requests to be submited or any of the io_* syscalls to be | |
856 | * called on the context. | |
857 | * | |
858 | * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on | |
859 | * them. | |
1da177e4 | 860 | */ |
fc9b52cd | 861 | void exit_aio(struct mm_struct *mm) |
1da177e4 | 862 | { |
4b70ac5f | 863 | struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table); |
dc48e56d JA |
864 | struct ctx_rq_wait wait; |
865 | int i, skipped; | |
db446a08 | 866 | |
4b70ac5f ON |
867 | if (!table) |
868 | return; | |
db446a08 | 869 | |
dc48e56d JA |
870 | atomic_set(&wait.count, table->nr); |
871 | init_completion(&wait.comp); | |
872 | ||
873 | skipped = 0; | |
4b70ac5f | 874 | for (i = 0; i < table->nr; ++i) { |
d0264c01 TH |
875 | struct kioctx *ctx = |
876 | rcu_dereference_protected(table->table[i], true); | |
abf137dd | 877 | |
dc48e56d JA |
878 | if (!ctx) { |
879 | skipped++; | |
4b70ac5f | 880 | continue; |
dc48e56d JA |
881 | } |
882 | ||
936af157 | 883 | /* |
4b70ac5f ON |
884 | * We don't need to bother with munmap() here - exit_mmap(mm) |
885 | * is coming and it'll unmap everything. And we simply can't, | |
886 | * this is not necessarily our ->mm. | |
887 | * Since kill_ioctx() uses non-zero ->mmap_size as indicator | |
888 | * that it needs to unmap the area, just set it to 0. | |
936af157 | 889 | */ |
58c85dc2 | 890 | ctx->mmap_size = 0; |
dc48e56d JA |
891 | kill_ioctx(mm, ctx, &wait); |
892 | } | |
36f55889 | 893 | |
dc48e56d | 894 | if (!atomic_sub_and_test(skipped, &wait.count)) { |
6098b45b | 895 | /* Wait until all IO for the context are done. */ |
dc48e56d | 896 | wait_for_completion(&wait.comp); |
1da177e4 | 897 | } |
4b70ac5f ON |
898 | |
899 | RCU_INIT_POINTER(mm->ioctx_table, NULL); | |
900 | kfree(table); | |
1da177e4 LT |
901 | } |
902 | ||
e1bdd5f2 KO |
903 | static void put_reqs_available(struct kioctx *ctx, unsigned nr) |
904 | { | |
905 | struct kioctx_cpu *kcpu; | |
263782c1 | 906 | unsigned long flags; |
e1bdd5f2 | 907 | |
263782c1 | 908 | local_irq_save(flags); |
be6fb451 | 909 | kcpu = this_cpu_ptr(ctx->cpu); |
e1bdd5f2 | 910 | kcpu->reqs_available += nr; |
263782c1 | 911 | |
e1bdd5f2 KO |
912 | while (kcpu->reqs_available >= ctx->req_batch * 2) { |
913 | kcpu->reqs_available -= ctx->req_batch; | |
914 | atomic_add(ctx->req_batch, &ctx->reqs_available); | |
915 | } | |
916 | ||
263782c1 | 917 | local_irq_restore(flags); |
e1bdd5f2 KO |
918 | } |
919 | ||
920 | static bool get_reqs_available(struct kioctx *ctx) | |
921 | { | |
922 | struct kioctx_cpu *kcpu; | |
923 | bool ret = false; | |
263782c1 | 924 | unsigned long flags; |
e1bdd5f2 | 925 | |
263782c1 | 926 | local_irq_save(flags); |
be6fb451 | 927 | kcpu = this_cpu_ptr(ctx->cpu); |
e1bdd5f2 KO |
928 | if (!kcpu->reqs_available) { |
929 | int old, avail = atomic_read(&ctx->reqs_available); | |
930 | ||
931 | do { | |
932 | if (avail < ctx->req_batch) | |
933 | goto out; | |
934 | ||
935 | old = avail; | |
936 | avail = atomic_cmpxchg(&ctx->reqs_available, | |
937 | avail, avail - ctx->req_batch); | |
938 | } while (avail != old); | |
939 | ||
940 | kcpu->reqs_available += ctx->req_batch; | |
941 | } | |
942 | ||
943 | ret = true; | |
944 | kcpu->reqs_available--; | |
945 | out: | |
263782c1 | 946 | local_irq_restore(flags); |
e1bdd5f2 KO |
947 | return ret; |
948 | } | |
949 | ||
d856f32a BL |
950 | /* refill_reqs_available |
951 | * Updates the reqs_available reference counts used for tracking the | |
952 | * number of free slots in the completion ring. This can be called | |
953 | * from aio_complete() (to optimistically update reqs_available) or | |
954 | * from aio_get_req() (the we're out of events case). It must be | |
955 | * called holding ctx->completion_lock. | |
956 | */ | |
957 | static void refill_reqs_available(struct kioctx *ctx, unsigned head, | |
958 | unsigned tail) | |
959 | { | |
960 | unsigned events_in_ring, completed; | |
961 | ||
962 | /* Clamp head since userland can write to it. */ | |
963 | head %= ctx->nr_events; | |
964 | if (head <= tail) | |
965 | events_in_ring = tail - head; | |
966 | else | |
967 | events_in_ring = ctx->nr_events - (head - tail); | |
968 | ||
969 | completed = ctx->completed_events; | |
970 | if (events_in_ring < completed) | |
971 | completed -= events_in_ring; | |
972 | else | |
973 | completed = 0; | |
974 | ||
975 | if (!completed) | |
976 | return; | |
977 | ||
978 | ctx->completed_events -= completed; | |
979 | put_reqs_available(ctx, completed); | |
980 | } | |
981 | ||
982 | /* user_refill_reqs_available | |
983 | * Called to refill reqs_available when aio_get_req() encounters an | |
984 | * out of space in the completion ring. | |
985 | */ | |
986 | static void user_refill_reqs_available(struct kioctx *ctx) | |
987 | { | |
988 | spin_lock_irq(&ctx->completion_lock); | |
989 | if (ctx->completed_events) { | |
990 | struct aio_ring *ring; | |
991 | unsigned head; | |
992 | ||
993 | /* Access of ring->head may race with aio_read_events_ring() | |
994 | * here, but that's okay since whether we read the old version | |
995 | * or the new version, and either will be valid. The important | |
996 | * part is that head cannot pass tail since we prevent | |
997 | * aio_complete() from updating tail by holding | |
998 | * ctx->completion_lock. Even if head is invalid, the check | |
999 | * against ctx->completed_events below will make sure we do the | |
1000 | * safe/right thing. | |
1001 | */ | |
1002 | ring = kmap_atomic(ctx->ring_pages[0]); | |
1003 | head = ring->head; | |
1004 | kunmap_atomic(ring); | |
1005 | ||
1006 | refill_reqs_available(ctx, head, ctx->tail); | |
1007 | } | |
1008 | ||
1009 | spin_unlock_irq(&ctx->completion_lock); | |
1010 | } | |
1011 | ||
1da177e4 | 1012 | /* aio_get_req |
57282d8f KO |
1013 | * Allocate a slot for an aio request. |
1014 | * Returns NULL if no requests are free. | |
1da177e4 | 1015 | */ |
04b2fa9f | 1016 | static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx) |
1da177e4 | 1017 | { |
04b2fa9f | 1018 | struct aio_kiocb *req; |
a1c8eae7 | 1019 | |
d856f32a BL |
1020 | if (!get_reqs_available(ctx)) { |
1021 | user_refill_reqs_available(ctx); | |
1022 | if (!get_reqs_available(ctx)) | |
1023 | return NULL; | |
1024 | } | |
a1c8eae7 | 1025 | |
0460fef2 | 1026 | req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO); |
1da177e4 | 1027 | if (unlikely(!req)) |
a1c8eae7 | 1028 | goto out_put; |
1da177e4 | 1029 | |
e34ecee2 | 1030 | percpu_ref_get(&ctx->reqs); |
75321b50 | 1031 | INIT_LIST_HEAD(&req->ki_list); |
9018ccc4 | 1032 | refcount_set(&req->ki_refcnt, 0); |
1da177e4 | 1033 | req->ki_ctx = ctx; |
080d676d | 1034 | return req; |
a1c8eae7 | 1035 | out_put: |
e1bdd5f2 | 1036 | put_reqs_available(ctx, 1); |
a1c8eae7 | 1037 | return NULL; |
1da177e4 LT |
1038 | } |
1039 | ||
d5470b59 | 1040 | static struct kioctx *lookup_ioctx(unsigned long ctx_id) |
1da177e4 | 1041 | { |
db446a08 | 1042 | struct aio_ring __user *ring = (void __user *)ctx_id; |
abf137dd | 1043 | struct mm_struct *mm = current->mm; |
65c24491 | 1044 | struct kioctx *ctx, *ret = NULL; |
db446a08 BL |
1045 | struct kioctx_table *table; |
1046 | unsigned id; | |
1047 | ||
1048 | if (get_user(id, &ring->id)) | |
1049 | return NULL; | |
1da177e4 | 1050 | |
abf137dd | 1051 | rcu_read_lock(); |
db446a08 | 1052 | table = rcu_dereference(mm->ioctx_table); |
abf137dd | 1053 | |
db446a08 BL |
1054 | if (!table || id >= table->nr) |
1055 | goto out; | |
1da177e4 | 1056 | |
d0264c01 | 1057 | ctx = rcu_dereference(table->table[id]); |
f30d704f | 1058 | if (ctx && ctx->user_id == ctx_id) { |
baf10564 AV |
1059 | if (percpu_ref_tryget_live(&ctx->users)) |
1060 | ret = ctx; | |
db446a08 BL |
1061 | } |
1062 | out: | |
abf137dd | 1063 | rcu_read_unlock(); |
65c24491 | 1064 | return ret; |
1da177e4 LT |
1065 | } |
1066 | ||
9018ccc4 CH |
1067 | static inline void iocb_put(struct aio_kiocb *iocb) |
1068 | { | |
1069 | if (refcount_read(&iocb->ki_refcnt) == 0 || | |
1070 | refcount_dec_and_test(&iocb->ki_refcnt)) { | |
1071 | percpu_ref_put(&iocb->ki_ctx->reqs); | |
1072 | kmem_cache_free(kiocb_cachep, iocb); | |
1073 | } | |
1074 | } | |
1075 | ||
1da177e4 LT |
1076 | /* aio_complete |
1077 | * Called when the io request on the given iocb is complete. | |
1da177e4 | 1078 | */ |
54843f87 | 1079 | static void aio_complete(struct aio_kiocb *iocb, long res, long res2) |
1da177e4 LT |
1080 | { |
1081 | struct kioctx *ctx = iocb->ki_ctx; | |
1da177e4 | 1082 | struct aio_ring *ring; |
21b40200 | 1083 | struct io_event *ev_page, *event; |
d856f32a | 1084 | unsigned tail, pos, head; |
1da177e4 | 1085 | unsigned long flags; |
1da177e4 | 1086 | |
0460fef2 KO |
1087 | /* |
1088 | * Add a completion event to the ring buffer. Must be done holding | |
4b30f07e | 1089 | * ctx->completion_lock to prevent other code from messing with the tail |
0460fef2 KO |
1090 | * pointer since we might be called from irq context. |
1091 | */ | |
1092 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1093 | ||
58c85dc2 | 1094 | tail = ctx->tail; |
21b40200 KO |
1095 | pos = tail + AIO_EVENTS_OFFSET; |
1096 | ||
58c85dc2 | 1097 | if (++tail >= ctx->nr_events) |
4bf69b2a | 1098 | tail = 0; |
1da177e4 | 1099 | |
58c85dc2 | 1100 | ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
1101 | event = ev_page + pos % AIO_EVENTS_PER_PAGE; |
1102 | ||
04b2fa9f | 1103 | event->obj = (u64)(unsigned long)iocb->ki_user_iocb; |
1da177e4 LT |
1104 | event->data = iocb->ki_user_data; |
1105 | event->res = res; | |
1106 | event->res2 = res2; | |
1107 | ||
21b40200 | 1108 | kunmap_atomic(ev_page); |
58c85dc2 | 1109 | flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
1110 | |
1111 | pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n", | |
04b2fa9f | 1112 | ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data, |
caf4167a | 1113 | res, res2); |
1da177e4 LT |
1114 | |
1115 | /* after flagging the request as done, we | |
1116 | * must never even look at it again | |
1117 | */ | |
1118 | smp_wmb(); /* make event visible before updating tail */ | |
1119 | ||
58c85dc2 | 1120 | ctx->tail = tail; |
1da177e4 | 1121 | |
58c85dc2 | 1122 | ring = kmap_atomic(ctx->ring_pages[0]); |
d856f32a | 1123 | head = ring->head; |
21b40200 | 1124 | ring->tail = tail; |
e8e3c3d6 | 1125 | kunmap_atomic(ring); |
58c85dc2 | 1126 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 | 1127 | |
d856f32a BL |
1128 | ctx->completed_events++; |
1129 | if (ctx->completed_events > 1) | |
1130 | refill_reqs_available(ctx, head, tail); | |
0460fef2 KO |
1131 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1132 | ||
21b40200 | 1133 | pr_debug("added to ring %p at [%u]\n", iocb, tail); |
8d1c98b0 DL |
1134 | |
1135 | /* | |
1136 | * Check if the user asked us to deliver the result through an | |
1137 | * eventfd. The eventfd_signal() function is safe to be called | |
1138 | * from IRQ context. | |
1139 | */ | |
54843f87 | 1140 | if (iocb->ki_eventfd) { |
8d1c98b0 | 1141 | eventfd_signal(iocb->ki_eventfd, 1); |
54843f87 CH |
1142 | eventfd_ctx_put(iocb->ki_eventfd); |
1143 | } | |
8d1c98b0 | 1144 | |
6cb2a210 QB |
1145 | /* |
1146 | * We have to order our ring_info tail store above and test | |
1147 | * of the wait list below outside the wait lock. This is | |
1148 | * like in wake_up_bit() where clearing a bit has to be | |
1149 | * ordered with the unlocked test. | |
1150 | */ | |
1151 | smp_mb(); | |
1152 | ||
1da177e4 LT |
1153 | if (waitqueue_active(&ctx->wait)) |
1154 | wake_up(&ctx->wait); | |
9018ccc4 | 1155 | iocb_put(iocb); |
1da177e4 LT |
1156 | } |
1157 | ||
2be4e7de | 1158 | /* aio_read_events_ring |
a31ad380 KO |
1159 | * Pull an event off of the ioctx's event ring. Returns the number of |
1160 | * events fetched | |
1da177e4 | 1161 | */ |
a31ad380 KO |
1162 | static long aio_read_events_ring(struct kioctx *ctx, |
1163 | struct io_event __user *event, long nr) | |
1da177e4 | 1164 | { |
1da177e4 | 1165 | struct aio_ring *ring; |
5ffac122 | 1166 | unsigned head, tail, pos; |
a31ad380 KO |
1167 | long ret = 0; |
1168 | int copy_ret; | |
1169 | ||
9c9ce763 DC |
1170 | /* |
1171 | * The mutex can block and wake us up and that will cause | |
1172 | * wait_event_interruptible_hrtimeout() to schedule without sleeping | |
1173 | * and repeat. This should be rare enough that it doesn't cause | |
1174 | * peformance issues. See the comment in read_events() for more detail. | |
1175 | */ | |
1176 | sched_annotate_sleep(); | |
58c85dc2 | 1177 | mutex_lock(&ctx->ring_lock); |
1da177e4 | 1178 | |
fa8a53c3 | 1179 | /* Access to ->ring_pages here is protected by ctx->ring_lock. */ |
58c85dc2 | 1180 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 1181 | head = ring->head; |
5ffac122 | 1182 | tail = ring->tail; |
a31ad380 KO |
1183 | kunmap_atomic(ring); |
1184 | ||
2ff396be JM |
1185 | /* |
1186 | * Ensure that once we've read the current tail pointer, that | |
1187 | * we also see the events that were stored up to the tail. | |
1188 | */ | |
1189 | smp_rmb(); | |
1190 | ||
5ffac122 | 1191 | pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events); |
1da177e4 | 1192 | |
5ffac122 | 1193 | if (head == tail) |
1da177e4 LT |
1194 | goto out; |
1195 | ||
edfbbf38 BL |
1196 | head %= ctx->nr_events; |
1197 | tail %= ctx->nr_events; | |
1198 | ||
a31ad380 KO |
1199 | while (ret < nr) { |
1200 | long avail; | |
1201 | struct io_event *ev; | |
1202 | struct page *page; | |
1203 | ||
5ffac122 KO |
1204 | avail = (head <= tail ? tail : ctx->nr_events) - head; |
1205 | if (head == tail) | |
a31ad380 KO |
1206 | break; |
1207 | ||
a31ad380 | 1208 | pos = head + AIO_EVENTS_OFFSET; |
58c85dc2 | 1209 | page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]; |
a31ad380 KO |
1210 | pos %= AIO_EVENTS_PER_PAGE; |
1211 | ||
d2988bd4 AV |
1212 | avail = min(avail, nr - ret); |
1213 | avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos); | |
1214 | ||
a31ad380 KO |
1215 | ev = kmap(page); |
1216 | copy_ret = copy_to_user(event + ret, ev + pos, | |
1217 | sizeof(*ev) * avail); | |
1218 | kunmap(page); | |
1219 | ||
1220 | if (unlikely(copy_ret)) { | |
1221 | ret = -EFAULT; | |
1222 | goto out; | |
1223 | } | |
1224 | ||
1225 | ret += avail; | |
1226 | head += avail; | |
58c85dc2 | 1227 | head %= ctx->nr_events; |
1da177e4 | 1228 | } |
1da177e4 | 1229 | |
58c85dc2 | 1230 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 1231 | ring->head = head; |
91d80a84 | 1232 | kunmap_atomic(ring); |
58c85dc2 | 1233 | flush_dcache_page(ctx->ring_pages[0]); |
a31ad380 | 1234 | |
5ffac122 | 1235 | pr_debug("%li h%u t%u\n", ret, head, tail); |
a31ad380 | 1236 | out: |
58c85dc2 | 1237 | mutex_unlock(&ctx->ring_lock); |
a31ad380 | 1238 | |
1da177e4 LT |
1239 | return ret; |
1240 | } | |
1241 | ||
a31ad380 KO |
1242 | static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr, |
1243 | struct io_event __user *event, long *i) | |
1da177e4 | 1244 | { |
a31ad380 | 1245 | long ret = aio_read_events_ring(ctx, event + *i, nr - *i); |
1da177e4 | 1246 | |
a31ad380 KO |
1247 | if (ret > 0) |
1248 | *i += ret; | |
1da177e4 | 1249 | |
a31ad380 KO |
1250 | if (unlikely(atomic_read(&ctx->dead))) |
1251 | ret = -EINVAL; | |
1da177e4 | 1252 | |
a31ad380 KO |
1253 | if (!*i) |
1254 | *i = ret; | |
1da177e4 | 1255 | |
a31ad380 | 1256 | return ret < 0 || *i >= min_nr; |
1da177e4 LT |
1257 | } |
1258 | ||
a31ad380 | 1259 | static long read_events(struct kioctx *ctx, long min_nr, long nr, |
1da177e4 | 1260 | struct io_event __user *event, |
fa2e62a5 | 1261 | ktime_t until) |
1da177e4 | 1262 | { |
a31ad380 | 1263 | long ret = 0; |
1da177e4 | 1264 | |
a31ad380 KO |
1265 | /* |
1266 | * Note that aio_read_events() is being called as the conditional - i.e. | |
1267 | * we're calling it after prepare_to_wait() has set task state to | |
1268 | * TASK_INTERRUPTIBLE. | |
1269 | * | |
1270 | * But aio_read_events() can block, and if it blocks it's going to flip | |
1271 | * the task state back to TASK_RUNNING. | |
1272 | * | |
1273 | * This should be ok, provided it doesn't flip the state back to | |
1274 | * TASK_RUNNING and return 0 too much - that causes us to spin. That | |
1275 | * will only happen if the mutex_lock() call blocks, and we then find | |
1276 | * the ringbuffer empty. So in practice we should be ok, but it's | |
1277 | * something to be aware of when touching this code. | |
1278 | */ | |
2456e855 | 1279 | if (until == 0) |
5f785de5 FZ |
1280 | aio_read_events(ctx, min_nr, nr, event, &ret); |
1281 | else | |
1282 | wait_event_interruptible_hrtimeout(ctx->wait, | |
1283 | aio_read_events(ctx, min_nr, nr, event, &ret), | |
1284 | until); | |
a31ad380 | 1285 | return ret; |
1da177e4 LT |
1286 | } |
1287 | ||
1da177e4 LT |
1288 | /* sys_io_setup: |
1289 | * Create an aio_context capable of receiving at least nr_events. | |
1290 | * ctxp must not point to an aio_context that already exists, and | |
1291 | * must be initialized to 0 prior to the call. On successful | |
1292 | * creation of the aio_context, *ctxp is filled in with the resulting | |
1293 | * handle. May fail with -EINVAL if *ctxp is not initialized, | |
1294 | * if the specified nr_events exceeds internal limits. May fail | |
1295 | * with -EAGAIN if the specified nr_events exceeds the user's limit | |
1296 | * of available events. May fail with -ENOMEM if insufficient kernel | |
1297 | * resources are available. May fail with -EFAULT if an invalid | |
1298 | * pointer is passed for ctxp. Will fail with -ENOSYS if not | |
1299 | * implemented. | |
1300 | */ | |
002c8976 | 1301 | SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) |
1da177e4 LT |
1302 | { |
1303 | struct kioctx *ioctx = NULL; | |
1304 | unsigned long ctx; | |
1305 | long ret; | |
1306 | ||
1307 | ret = get_user(ctx, ctxp); | |
1308 | if (unlikely(ret)) | |
1309 | goto out; | |
1310 | ||
1311 | ret = -EINVAL; | |
d55b5fda | 1312 | if (unlikely(ctx || nr_events == 0)) { |
acd88d4e | 1313 | pr_debug("EINVAL: ctx %lu nr_events %u\n", |
d55b5fda | 1314 | ctx, nr_events); |
1da177e4 LT |
1315 | goto out; |
1316 | } | |
1317 | ||
1318 | ioctx = ioctx_alloc(nr_events); | |
1319 | ret = PTR_ERR(ioctx); | |
1320 | if (!IS_ERR(ioctx)) { | |
1321 | ret = put_user(ioctx->user_id, ctxp); | |
a2e1859a | 1322 | if (ret) |
e02ba72a | 1323 | kill_ioctx(current->mm, ioctx, NULL); |
723be6e3 | 1324 | percpu_ref_put(&ioctx->users); |
1da177e4 LT |
1325 | } |
1326 | ||
1327 | out: | |
1328 | return ret; | |
1329 | } | |
1330 | ||
c00d2c7e AV |
1331 | #ifdef CONFIG_COMPAT |
1332 | COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p) | |
1333 | { | |
1334 | struct kioctx *ioctx = NULL; | |
1335 | unsigned long ctx; | |
1336 | long ret; | |
1337 | ||
1338 | ret = get_user(ctx, ctx32p); | |
1339 | if (unlikely(ret)) | |
1340 | goto out; | |
1341 | ||
1342 | ret = -EINVAL; | |
1343 | if (unlikely(ctx || nr_events == 0)) { | |
1344 | pr_debug("EINVAL: ctx %lu nr_events %u\n", | |
1345 | ctx, nr_events); | |
1346 | goto out; | |
1347 | } | |
1348 | ||
1349 | ioctx = ioctx_alloc(nr_events); | |
1350 | ret = PTR_ERR(ioctx); | |
1351 | if (!IS_ERR(ioctx)) { | |
1352 | /* truncating is ok because it's a user address */ | |
1353 | ret = put_user((u32)ioctx->user_id, ctx32p); | |
1354 | if (ret) | |
1355 | kill_ioctx(current->mm, ioctx, NULL); | |
1356 | percpu_ref_put(&ioctx->users); | |
1357 | } | |
1358 | ||
1359 | out: | |
1360 | return ret; | |
1361 | } | |
1362 | #endif | |
1363 | ||
1da177e4 LT |
1364 | /* sys_io_destroy: |
1365 | * Destroy the aio_context specified. May cancel any outstanding | |
1366 | * AIOs and block on completion. Will fail with -ENOSYS if not | |
642b5123 | 1367 | * implemented. May fail with -EINVAL if the context pointed to |
1da177e4 LT |
1368 | * is invalid. |
1369 | */ | |
002c8976 | 1370 | SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) |
1da177e4 LT |
1371 | { |
1372 | struct kioctx *ioctx = lookup_ioctx(ctx); | |
1373 | if (likely(NULL != ioctx)) { | |
dc48e56d | 1374 | struct ctx_rq_wait wait; |
fb2d4483 | 1375 | int ret; |
e02ba72a | 1376 | |
dc48e56d JA |
1377 | init_completion(&wait.comp); |
1378 | atomic_set(&wait.count, 1); | |
1379 | ||
e02ba72a AP |
1380 | /* Pass requests_done to kill_ioctx() where it can be set |
1381 | * in a thread-safe way. If we try to set it here then we have | |
1382 | * a race condition if two io_destroy() called simultaneously. | |
1383 | */ | |
dc48e56d | 1384 | ret = kill_ioctx(current->mm, ioctx, &wait); |
723be6e3 | 1385 | percpu_ref_put(&ioctx->users); |
e02ba72a AP |
1386 | |
1387 | /* Wait until all IO for the context are done. Otherwise kernel | |
1388 | * keep using user-space buffers even if user thinks the context | |
1389 | * is destroyed. | |
1390 | */ | |
fb2d4483 | 1391 | if (!ret) |
dc48e56d | 1392 | wait_for_completion(&wait.comp); |
e02ba72a | 1393 | |
fb2d4483 | 1394 | return ret; |
1da177e4 | 1395 | } |
acd88d4e | 1396 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1397 | return -EINVAL; |
1398 | } | |
1399 | ||
3c96c7f4 AV |
1400 | static void aio_remove_iocb(struct aio_kiocb *iocb) |
1401 | { | |
1402 | struct kioctx *ctx = iocb->ki_ctx; | |
1403 | unsigned long flags; | |
1404 | ||
1405 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
1406 | list_del(&iocb->ki_list); | |
1407 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
1408 | } | |
1409 | ||
54843f87 CH |
1410 | static void aio_complete_rw(struct kiocb *kiocb, long res, long res2) |
1411 | { | |
1412 | struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw); | |
1413 | ||
3c96c7f4 AV |
1414 | if (!list_empty_careful(&iocb->ki_list)) |
1415 | aio_remove_iocb(iocb); | |
1416 | ||
54843f87 CH |
1417 | if (kiocb->ki_flags & IOCB_WRITE) { |
1418 | struct inode *inode = file_inode(kiocb->ki_filp); | |
1419 | ||
1420 | /* | |
1421 | * Tell lockdep we inherited freeze protection from submission | |
1422 | * thread. | |
1423 | */ | |
1424 | if (S_ISREG(inode->i_mode)) | |
1425 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); | |
1426 | file_end_write(kiocb->ki_filp); | |
1427 | } | |
1428 | ||
1429 | fput(kiocb->ki_filp); | |
1430 | aio_complete(iocb, res, res2); | |
1431 | } | |
1432 | ||
1433 | static int aio_prep_rw(struct kiocb *req, struct iocb *iocb) | |
1434 | { | |
1435 | int ret; | |
1436 | ||
1437 | req->ki_filp = fget(iocb->aio_fildes); | |
1438 | if (unlikely(!req->ki_filp)) | |
1439 | return -EBADF; | |
1440 | req->ki_complete = aio_complete_rw; | |
1441 | req->ki_pos = iocb->aio_offset; | |
1442 | req->ki_flags = iocb_flags(req->ki_filp); | |
1443 | if (iocb->aio_flags & IOCB_FLAG_RESFD) | |
1444 | req->ki_flags |= IOCB_EVENTFD; | |
fc28724d | 1445 | req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp)); |
d9a08a9e AM |
1446 | if (iocb->aio_flags & IOCB_FLAG_IOPRIO) { |
1447 | /* | |
1448 | * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then | |
1449 | * aio_reqprio is interpreted as an I/O scheduling | |
1450 | * class and priority. | |
1451 | */ | |
1452 | ret = ioprio_check_cap(iocb->aio_reqprio); | |
1453 | if (ret) { | |
9a6d9a62 AM |
1454 | pr_debug("aio ioprio check cap error: %d\n", ret); |
1455 | return ret; | |
d9a08a9e AM |
1456 | } |
1457 | ||
1458 | req->ki_ioprio = iocb->aio_reqprio; | |
1459 | } else | |
1460 | req->ki_ioprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_NONE, 0); | |
1461 | ||
54843f87 CH |
1462 | ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags); |
1463 | if (unlikely(ret)) | |
1464 | fput(req->ki_filp); | |
1465 | return ret; | |
1466 | } | |
1467 | ||
89319d31 CH |
1468 | static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec, |
1469 | bool vectored, bool compat, struct iov_iter *iter) | |
eed4e51f | 1470 | { |
89319d31 CH |
1471 | void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf; |
1472 | size_t len = iocb->aio_nbytes; | |
1473 | ||
1474 | if (!vectored) { | |
1475 | ssize_t ret = import_single_range(rw, buf, len, *iovec, iter); | |
1476 | *iovec = NULL; | |
1477 | return ret; | |
1478 | } | |
9d85cba7 JM |
1479 | #ifdef CONFIG_COMPAT |
1480 | if (compat) | |
89319d31 CH |
1481 | return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec, |
1482 | iter); | |
9d85cba7 | 1483 | #endif |
89319d31 | 1484 | return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter); |
eed4e51f BP |
1485 | } |
1486 | ||
9061d14a | 1487 | static inline void aio_rw_done(struct kiocb *req, ssize_t ret) |
89319d31 CH |
1488 | { |
1489 | switch (ret) { | |
1490 | case -EIOCBQUEUED: | |
9061d14a | 1491 | break; |
89319d31 CH |
1492 | case -ERESTARTSYS: |
1493 | case -ERESTARTNOINTR: | |
1494 | case -ERESTARTNOHAND: | |
1495 | case -ERESTART_RESTARTBLOCK: | |
1496 | /* | |
1497 | * There's no easy way to restart the syscall since other AIO's | |
1498 | * may be already running. Just fail this IO with EINTR. | |
1499 | */ | |
1500 | ret = -EINTR; | |
1501 | /*FALLTHRU*/ | |
1502 | default: | |
54843f87 | 1503 | aio_complete_rw(req, ret, 0); |
89319d31 CH |
1504 | } |
1505 | } | |
1506 | ||
1507 | static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored, | |
1508 | bool compat) | |
1da177e4 | 1509 | { |
00fefb9c | 1510 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; |
293bc982 | 1511 | struct iov_iter iter; |
54843f87 | 1512 | struct file *file; |
89319d31 | 1513 | ssize_t ret; |
1da177e4 | 1514 | |
54843f87 CH |
1515 | ret = aio_prep_rw(req, iocb); |
1516 | if (ret) | |
1517 | return ret; | |
1518 | file = req->ki_filp; | |
1519 | ||
1520 | ret = -EBADF; | |
89319d31 | 1521 | if (unlikely(!(file->f_mode & FMODE_READ))) |
54843f87 CH |
1522 | goto out_fput; |
1523 | ret = -EINVAL; | |
89319d31 | 1524 | if (unlikely(!file->f_op->read_iter)) |
54843f87 | 1525 | goto out_fput; |
73a7075e | 1526 | |
89319d31 CH |
1527 | ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter); |
1528 | if (ret) | |
54843f87 | 1529 | goto out_fput; |
89319d31 CH |
1530 | ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter)); |
1531 | if (!ret) | |
9061d14a | 1532 | aio_rw_done(req, call_read_iter(file, req, &iter)); |
89319d31 | 1533 | kfree(iovec); |
54843f87 | 1534 | out_fput: |
9061d14a | 1535 | if (unlikely(ret)) |
54843f87 | 1536 | fput(file); |
89319d31 CH |
1537 | return ret; |
1538 | } | |
73a7075e | 1539 | |
89319d31 CH |
1540 | static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored, |
1541 | bool compat) | |
1542 | { | |
89319d31 CH |
1543 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; |
1544 | struct iov_iter iter; | |
54843f87 | 1545 | struct file *file; |
89319d31 | 1546 | ssize_t ret; |
41ef4eb8 | 1547 | |
54843f87 CH |
1548 | ret = aio_prep_rw(req, iocb); |
1549 | if (ret) | |
1550 | return ret; | |
1551 | file = req->ki_filp; | |
1552 | ||
1553 | ret = -EBADF; | |
89319d31 | 1554 | if (unlikely(!(file->f_mode & FMODE_WRITE))) |
54843f87 CH |
1555 | goto out_fput; |
1556 | ret = -EINVAL; | |
89319d31 | 1557 | if (unlikely(!file->f_op->write_iter)) |
54843f87 | 1558 | goto out_fput; |
1da177e4 | 1559 | |
89319d31 CH |
1560 | ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter); |
1561 | if (ret) | |
54843f87 | 1562 | goto out_fput; |
89319d31 CH |
1563 | ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter)); |
1564 | if (!ret) { | |
70fe2f48 | 1565 | /* |
92ce4728 | 1566 | * Open-code file_start_write here to grab freeze protection, |
54843f87 CH |
1567 | * which will be released by another thread in |
1568 | * aio_complete_rw(). Fool lockdep by telling it the lock got | |
1569 | * released so that it doesn't complain about the held lock when | |
1570 | * we return to userspace. | |
70fe2f48 | 1571 | */ |
92ce4728 CH |
1572 | if (S_ISREG(file_inode(file)->i_mode)) { |
1573 | __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true); | |
a12f1ae6 | 1574 | __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE); |
92ce4728 CH |
1575 | } |
1576 | req->ki_flags |= IOCB_WRITE; | |
9061d14a | 1577 | aio_rw_done(req, call_write_iter(file, req, &iter)); |
41ef4eb8 | 1578 | } |
89319d31 | 1579 | kfree(iovec); |
54843f87 | 1580 | out_fput: |
9061d14a | 1581 | if (unlikely(ret)) |
54843f87 | 1582 | fput(file); |
89319d31 | 1583 | return ret; |
1da177e4 LT |
1584 | } |
1585 | ||
a3c0d439 CH |
1586 | static void aio_fsync_work(struct work_struct *work) |
1587 | { | |
1588 | struct fsync_iocb *req = container_of(work, struct fsync_iocb, work); | |
1589 | int ret; | |
1590 | ||
1591 | ret = vfs_fsync(req->file, req->datasync); | |
1592 | fput(req->file); | |
1593 | aio_complete(container_of(req, struct aio_kiocb, fsync), ret, 0); | |
1594 | } | |
1595 | ||
1596 | static int aio_fsync(struct fsync_iocb *req, struct iocb *iocb, bool datasync) | |
1597 | { | |
1598 | if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes || | |
1599 | iocb->aio_rw_flags)) | |
1600 | return -EINVAL; | |
a11e1d43 | 1601 | |
a3c0d439 CH |
1602 | req->file = fget(iocb->aio_fildes); |
1603 | if (unlikely(!req->file)) | |
1604 | return -EBADF; | |
1605 | if (unlikely(!req->file->f_op->fsync)) { | |
1606 | fput(req->file); | |
1607 | return -EINVAL; | |
1608 | } | |
1609 | ||
1610 | req->datasync = datasync; | |
1611 | INIT_WORK(&req->work, aio_fsync_work); | |
1612 | schedule_work(&req->work); | |
9061d14a | 1613 | return 0; |
a3c0d439 CH |
1614 | } |
1615 | ||
bfe4037e CH |
1616 | static inline void aio_poll_complete(struct aio_kiocb *iocb, __poll_t mask) |
1617 | { | |
1618 | struct file *file = iocb->poll.file; | |
1619 | ||
1620 | aio_complete(iocb, mangle_poll(mask), 0); | |
1621 | fput(file); | |
1622 | } | |
1623 | ||
1624 | static void aio_poll_complete_work(struct work_struct *work) | |
1625 | { | |
1626 | struct poll_iocb *req = container_of(work, struct poll_iocb, work); | |
1627 | struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll); | |
1628 | struct poll_table_struct pt = { ._key = req->events }; | |
1629 | struct kioctx *ctx = iocb->ki_ctx; | |
1630 | __poll_t mask = 0; | |
1631 | ||
1632 | if (!READ_ONCE(req->cancelled)) | |
1633 | mask = vfs_poll(req->file, &pt) & req->events; | |
1634 | ||
1635 | /* | |
1636 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
1637 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
1638 | * synchronize with them. In the cancellation case the list_del_init | |
1639 | * itself is not actually needed, but harmless so we keep it in to | |
1640 | * avoid further branches in the fast path. | |
1641 | */ | |
1642 | spin_lock_irq(&ctx->ctx_lock); | |
1643 | if (!mask && !READ_ONCE(req->cancelled)) { | |
1644 | add_wait_queue(req->head, &req->wait); | |
1645 | spin_unlock_irq(&ctx->ctx_lock); | |
1646 | return; | |
1647 | } | |
1648 | list_del_init(&iocb->ki_list); | |
1649 | spin_unlock_irq(&ctx->ctx_lock); | |
1650 | ||
1651 | aio_poll_complete(iocb, mask); | |
1652 | } | |
1653 | ||
1654 | /* assumes we are called with irqs disabled */ | |
1655 | static int aio_poll_cancel(struct kiocb *iocb) | |
1656 | { | |
1657 | struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw); | |
1658 | struct poll_iocb *req = &aiocb->poll; | |
1659 | ||
1660 | spin_lock(&req->head->lock); | |
1661 | WRITE_ONCE(req->cancelled, true); | |
1662 | if (!list_empty(&req->wait.entry)) { | |
1663 | list_del_init(&req->wait.entry); | |
1664 | schedule_work(&aiocb->poll.work); | |
1665 | } | |
1666 | spin_unlock(&req->head->lock); | |
1667 | ||
1668 | return 0; | |
1669 | } | |
1670 | ||
1671 | static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
1672 | void *key) | |
1673 | { | |
1674 | struct poll_iocb *req = container_of(wait, struct poll_iocb, wait); | |
1675 | __poll_t mask = key_to_poll(key); | |
1676 | ||
1677 | req->woken = true; | |
1678 | ||
1679 | /* for instances that support it check for an event match first: */ | |
1680 | if (mask && !(mask & req->events)) | |
1681 | return 0; | |
1682 | ||
1683 | list_del_init(&req->wait.entry); | |
1684 | schedule_work(&req->work); | |
1685 | return 1; | |
1686 | } | |
1687 | ||
1688 | struct aio_poll_table { | |
1689 | struct poll_table_struct pt; | |
1690 | struct aio_kiocb *iocb; | |
1691 | int error; | |
1692 | }; | |
1693 | ||
1694 | static void | |
1695 | aio_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
1696 | struct poll_table_struct *p) | |
1697 | { | |
1698 | struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt); | |
1699 | ||
1700 | /* multiple wait queues per file are not supported */ | |
1701 | if (unlikely(pt->iocb->poll.head)) { | |
1702 | pt->error = -EINVAL; | |
1703 | return; | |
1704 | } | |
1705 | ||
1706 | pt->error = 0; | |
1707 | pt->iocb->poll.head = head; | |
1708 | add_wait_queue(head, &pt->iocb->poll.wait); | |
1709 | } | |
1710 | ||
1711 | static ssize_t aio_poll(struct aio_kiocb *aiocb, struct iocb *iocb) | |
1712 | { | |
1713 | struct kioctx *ctx = aiocb->ki_ctx; | |
1714 | struct poll_iocb *req = &aiocb->poll; | |
1715 | struct aio_poll_table apt; | |
1716 | __poll_t mask; | |
1717 | ||
1718 | /* reject any unknown events outside the normal event mask. */ | |
1719 | if ((u16)iocb->aio_buf != iocb->aio_buf) | |
1720 | return -EINVAL; | |
1721 | /* reject fields that are not defined for poll */ | |
1722 | if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags) | |
1723 | return -EINVAL; | |
1724 | ||
1725 | INIT_WORK(&req->work, aio_poll_complete_work); | |
1726 | req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP; | |
1727 | req->file = fget(iocb->aio_fildes); | |
1728 | if (unlikely(!req->file)) | |
1729 | return -EBADF; | |
1730 | ||
1731 | apt.pt._qproc = aio_poll_queue_proc; | |
1732 | apt.pt._key = req->events; | |
1733 | apt.iocb = aiocb; | |
1734 | apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
1735 | ||
1736 | /* initialized the list so that we can do list_empty checks */ | |
1737 | INIT_LIST_HEAD(&req->wait.entry); | |
1738 | init_waitqueue_func_entry(&req->wait, aio_poll_wake); | |
1739 | ||
1740 | /* one for removal from waitqueue, one for this function */ | |
1741 | refcount_set(&aiocb->ki_refcnt, 2); | |
1742 | ||
1743 | mask = vfs_poll(req->file, &apt.pt) & req->events; | |
1744 | if (unlikely(!req->head)) { | |
1745 | /* we did not manage to set up a waitqueue, done */ | |
1746 | goto out; | |
1747 | } | |
1748 | ||
1749 | spin_lock_irq(&ctx->ctx_lock); | |
1750 | spin_lock(&req->head->lock); | |
1751 | if (req->woken) { | |
1752 | /* wake_up context handles the rest */ | |
1753 | mask = 0; | |
1754 | apt.error = 0; | |
1755 | } else if (mask || apt.error) { | |
1756 | /* if we get an error or a mask we are done */ | |
1757 | WARN_ON_ONCE(list_empty(&req->wait.entry)); | |
1758 | list_del_init(&req->wait.entry); | |
1759 | } else { | |
1760 | /* actually waiting for an event */ | |
1761 | list_add_tail(&aiocb->ki_list, &ctx->active_reqs); | |
1762 | aiocb->ki_cancel = aio_poll_cancel; | |
1763 | } | |
1764 | spin_unlock(&req->head->lock); | |
1765 | spin_unlock_irq(&ctx->ctx_lock); | |
1766 | ||
1767 | out: | |
1768 | if (unlikely(apt.error)) { | |
1769 | fput(req->file); | |
1770 | return apt.error; | |
1771 | } | |
1772 | ||
1773 | if (mask) | |
1774 | aio_poll_complete(aiocb, mask); | |
1775 | iocb_put(aiocb); | |
1776 | return 0; | |
1777 | } | |
1778 | ||
d5470b59 | 1779 | static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, |
95af8496 | 1780 | bool compat) |
1da177e4 | 1781 | { |
04b2fa9f | 1782 | struct aio_kiocb *req; |
95af8496 | 1783 | struct iocb iocb; |
1da177e4 LT |
1784 | ssize_t ret; |
1785 | ||
95af8496 AV |
1786 | if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb)))) |
1787 | return -EFAULT; | |
1788 | ||
1da177e4 | 1789 | /* enforce forwards compatibility on users */ |
95af8496 | 1790 | if (unlikely(iocb.aio_reserved2)) { |
caf4167a | 1791 | pr_debug("EINVAL: reserve field set\n"); |
1da177e4 LT |
1792 | return -EINVAL; |
1793 | } | |
1794 | ||
1795 | /* prevent overflows */ | |
1796 | if (unlikely( | |
95af8496 AV |
1797 | (iocb.aio_buf != (unsigned long)iocb.aio_buf) || |
1798 | (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) || | |
1799 | ((ssize_t)iocb.aio_nbytes < 0) | |
1da177e4 | 1800 | )) { |
acd88d4e | 1801 | pr_debug("EINVAL: overflow check\n"); |
1da177e4 LT |
1802 | return -EINVAL; |
1803 | } | |
1804 | ||
41ef4eb8 | 1805 | req = aio_get_req(ctx); |
1d98ebfc | 1806 | if (unlikely(!req)) |
1da177e4 | 1807 | return -EAGAIN; |
1d98ebfc | 1808 | |
95af8496 | 1809 | if (iocb.aio_flags & IOCB_FLAG_RESFD) { |
9c3060be DL |
1810 | /* |
1811 | * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an | |
1812 | * instance of the file* now. The file descriptor must be | |
1813 | * an eventfd() fd, and will be signaled for each completed | |
1814 | * event using the eventfd_signal() function. | |
1815 | */ | |
95af8496 | 1816 | req->ki_eventfd = eventfd_ctx_fdget((int) iocb.aio_resfd); |
801678c5 | 1817 | if (IS_ERR(req->ki_eventfd)) { |
9c3060be | 1818 | ret = PTR_ERR(req->ki_eventfd); |
87c3a86e | 1819 | req->ki_eventfd = NULL; |
9c3060be DL |
1820 | goto out_put_req; |
1821 | } | |
9830f4be GR |
1822 | } |
1823 | ||
8a660890 | 1824 | ret = put_user(KIOCB_KEY, &user_iocb->aio_key); |
1da177e4 | 1825 | if (unlikely(ret)) { |
caf4167a | 1826 | pr_debug("EFAULT: aio_key\n"); |
1da177e4 LT |
1827 | goto out_put_req; |
1828 | } | |
1829 | ||
04b2fa9f | 1830 | req->ki_user_iocb = user_iocb; |
95af8496 | 1831 | req->ki_user_data = iocb.aio_data; |
1da177e4 | 1832 | |
95af8496 | 1833 | switch (iocb.aio_lio_opcode) { |
89319d31 | 1834 | case IOCB_CMD_PREAD: |
95af8496 | 1835 | ret = aio_read(&req->rw, &iocb, false, compat); |
89319d31 CH |
1836 | break; |
1837 | case IOCB_CMD_PWRITE: | |
95af8496 | 1838 | ret = aio_write(&req->rw, &iocb, false, compat); |
89319d31 CH |
1839 | break; |
1840 | case IOCB_CMD_PREADV: | |
95af8496 | 1841 | ret = aio_read(&req->rw, &iocb, true, compat); |
89319d31 CH |
1842 | break; |
1843 | case IOCB_CMD_PWRITEV: | |
95af8496 | 1844 | ret = aio_write(&req->rw, &iocb, true, compat); |
89319d31 | 1845 | break; |
a3c0d439 | 1846 | case IOCB_CMD_FSYNC: |
95af8496 | 1847 | ret = aio_fsync(&req->fsync, &iocb, false); |
a3c0d439 CH |
1848 | break; |
1849 | case IOCB_CMD_FDSYNC: | |
95af8496 | 1850 | ret = aio_fsync(&req->fsync, &iocb, true); |
ac060cba | 1851 | break; |
bfe4037e CH |
1852 | case IOCB_CMD_POLL: |
1853 | ret = aio_poll(req, &iocb); | |
1854 | break; | |
89319d31 | 1855 | default: |
95af8496 | 1856 | pr_debug("invalid aio operation %d\n", iocb.aio_lio_opcode); |
89319d31 CH |
1857 | ret = -EINVAL; |
1858 | break; | |
1859 | } | |
41003a7b | 1860 | |
92ce4728 | 1861 | /* |
9061d14a AV |
1862 | * If ret is 0, we'd either done aio_complete() ourselves or have |
1863 | * arranged for that to be done asynchronously. Anything non-zero | |
1864 | * means that we need to destroy req ourselves. | |
92ce4728 | 1865 | */ |
9061d14a | 1866 | if (ret) |
89319d31 | 1867 | goto out_put_req; |
1da177e4 | 1868 | return 0; |
1da177e4 | 1869 | out_put_req: |
e1bdd5f2 | 1870 | put_reqs_available(ctx, 1); |
e34ecee2 | 1871 | percpu_ref_put(&ctx->reqs); |
54843f87 CH |
1872 | if (req->ki_eventfd) |
1873 | eventfd_ctx_put(req->ki_eventfd); | |
1874 | kmem_cache_free(kiocb_cachep, req); | |
1da177e4 LT |
1875 | return ret; |
1876 | } | |
1877 | ||
67ba049f AV |
1878 | /* sys_io_submit: |
1879 | * Queue the nr iocbs pointed to by iocbpp for processing. Returns | |
1880 | * the number of iocbs queued. May return -EINVAL if the aio_context | |
1881 | * specified by ctx_id is invalid, if nr is < 0, if the iocb at | |
1882 | * *iocbpp[0] is not properly initialized, if the operation specified | |
1883 | * is invalid for the file descriptor in the iocb. May fail with | |
1884 | * -EFAULT if any of the data structures point to invalid data. May | |
1885 | * fail with -EBADF if the file descriptor specified in the first | |
1886 | * iocb is invalid. May fail with -EAGAIN if insufficient resources | |
1887 | * are available to queue any iocbs. Will return 0 if nr is 0. Will | |
1888 | * fail with -ENOSYS if not implemented. | |
1889 | */ | |
1890 | SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, | |
1891 | struct iocb __user * __user *, iocbpp) | |
1da177e4 LT |
1892 | { |
1893 | struct kioctx *ctx; | |
1894 | long ret = 0; | |
080d676d | 1895 | int i = 0; |
9f5b9425 | 1896 | struct blk_plug plug; |
1da177e4 LT |
1897 | |
1898 | if (unlikely(nr < 0)) | |
1899 | return -EINVAL; | |
1900 | ||
1da177e4 LT |
1901 | ctx = lookup_ioctx(ctx_id); |
1902 | if (unlikely(!ctx)) { | |
caf4167a | 1903 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1904 | return -EINVAL; |
1905 | } | |
1906 | ||
1da92779 AV |
1907 | if (nr > ctx->nr_events) |
1908 | nr = ctx->nr_events; | |
1909 | ||
9f5b9425 | 1910 | blk_start_plug(&plug); |
67ba049f | 1911 | for (i = 0; i < nr; i++) { |
1da177e4 | 1912 | struct iocb __user *user_iocb; |
1da177e4 | 1913 | |
67ba049f | 1914 | if (unlikely(get_user(user_iocb, iocbpp + i))) { |
1da177e4 LT |
1915 | ret = -EFAULT; |
1916 | break; | |
1917 | } | |
1918 | ||
67ba049f | 1919 | ret = io_submit_one(ctx, user_iocb, false); |
1da177e4 LT |
1920 | if (ret) |
1921 | break; | |
1922 | } | |
9f5b9425 | 1923 | blk_finish_plug(&plug); |
1da177e4 | 1924 | |
723be6e3 | 1925 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1926 | return i ? i : ret; |
1927 | } | |
1928 | ||
c00d2c7e | 1929 | #ifdef CONFIG_COMPAT |
c00d2c7e | 1930 | COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id, |
67ba049f | 1931 | int, nr, compat_uptr_t __user *, iocbpp) |
c00d2c7e | 1932 | { |
67ba049f AV |
1933 | struct kioctx *ctx; |
1934 | long ret = 0; | |
1935 | int i = 0; | |
1936 | struct blk_plug plug; | |
c00d2c7e AV |
1937 | |
1938 | if (unlikely(nr < 0)) | |
1939 | return -EINVAL; | |
1940 | ||
67ba049f AV |
1941 | ctx = lookup_ioctx(ctx_id); |
1942 | if (unlikely(!ctx)) { | |
1943 | pr_debug("EINVAL: invalid context id\n"); | |
1944 | return -EINVAL; | |
1945 | } | |
1946 | ||
1da92779 AV |
1947 | if (nr > ctx->nr_events) |
1948 | nr = ctx->nr_events; | |
1949 | ||
67ba049f AV |
1950 | blk_start_plug(&plug); |
1951 | for (i = 0; i < nr; i++) { | |
1952 | compat_uptr_t user_iocb; | |
1953 | ||
1954 | if (unlikely(get_user(user_iocb, iocbpp + i))) { | |
1955 | ret = -EFAULT; | |
1956 | break; | |
1957 | } | |
1958 | ||
1959 | ret = io_submit_one(ctx, compat_ptr(user_iocb), true); | |
1960 | if (ret) | |
1961 | break; | |
1962 | } | |
1963 | blk_finish_plug(&plug); | |
1964 | ||
1965 | percpu_ref_put(&ctx->users); | |
1966 | return i ? i : ret; | |
c00d2c7e AV |
1967 | } |
1968 | #endif | |
1969 | ||
1da177e4 LT |
1970 | /* lookup_kiocb |
1971 | * Finds a given iocb for cancellation. | |
1da177e4 | 1972 | */ |
04b2fa9f | 1973 | static struct aio_kiocb * |
f3a2752a | 1974 | lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb) |
1da177e4 | 1975 | { |
04b2fa9f | 1976 | struct aio_kiocb *kiocb; |
d00689af ZB |
1977 | |
1978 | assert_spin_locked(&ctx->ctx_lock); | |
1979 | ||
1da177e4 | 1980 | /* TODO: use a hash or array, this sucks. */ |
04b2fa9f CH |
1981 | list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) { |
1982 | if (kiocb->ki_user_iocb == iocb) | |
1da177e4 LT |
1983 | return kiocb; |
1984 | } | |
1985 | return NULL; | |
1986 | } | |
1987 | ||
1988 | /* sys_io_cancel: | |
1989 | * Attempts to cancel an iocb previously passed to io_submit. If | |
1990 | * the operation is successfully cancelled, the resulting event is | |
1991 | * copied into the memory pointed to by result without being placed | |
1992 | * into the completion queue and 0 is returned. May fail with | |
1993 | * -EFAULT if any of the data structures pointed to are invalid. | |
1994 | * May fail with -EINVAL if aio_context specified by ctx_id is | |
1995 | * invalid. May fail with -EAGAIN if the iocb specified was not | |
1996 | * cancelled. Will fail with -ENOSYS if not implemented. | |
1997 | */ | |
002c8976 HC |
1998 | SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, |
1999 | struct io_event __user *, result) | |
1da177e4 | 2000 | { |
1da177e4 | 2001 | struct kioctx *ctx; |
04b2fa9f | 2002 | struct aio_kiocb *kiocb; |
888933f8 | 2003 | int ret = -EINVAL; |
1da177e4 | 2004 | u32 key; |
1da177e4 | 2005 | |
f3a2752a | 2006 | if (unlikely(get_user(key, &iocb->aio_key))) |
1da177e4 | 2007 | return -EFAULT; |
f3a2752a CH |
2008 | if (unlikely(key != KIOCB_KEY)) |
2009 | return -EINVAL; | |
1da177e4 LT |
2010 | |
2011 | ctx = lookup_ioctx(ctx_id); | |
2012 | if (unlikely(!ctx)) | |
2013 | return -EINVAL; | |
2014 | ||
2015 | spin_lock_irq(&ctx->ctx_lock); | |
f3a2752a | 2016 | kiocb = lookup_kiocb(ctx, iocb); |
888933f8 CH |
2017 | if (kiocb) { |
2018 | ret = kiocb->ki_cancel(&kiocb->rw); | |
2019 | list_del_init(&kiocb->ki_list); | |
2020 | } | |
1da177e4 LT |
2021 | spin_unlock_irq(&ctx->ctx_lock); |
2022 | ||
906b973c | 2023 | if (!ret) { |
bec68faa KO |
2024 | /* |
2025 | * The result argument is no longer used - the io_event is | |
2026 | * always delivered via the ring buffer. -EINPROGRESS indicates | |
2027 | * cancellation is progress: | |
906b973c | 2028 | */ |
bec68faa | 2029 | ret = -EINPROGRESS; |
906b973c | 2030 | } |
1da177e4 | 2031 | |
723be6e3 | 2032 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
2033 | |
2034 | return ret; | |
2035 | } | |
2036 | ||
fa2e62a5 DD |
2037 | static long do_io_getevents(aio_context_t ctx_id, |
2038 | long min_nr, | |
2039 | long nr, | |
2040 | struct io_event __user *events, | |
2041 | struct timespec64 *ts) | |
2042 | { | |
2043 | ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX; | |
2044 | struct kioctx *ioctx = lookup_ioctx(ctx_id); | |
2045 | long ret = -EINVAL; | |
2046 | ||
2047 | if (likely(ioctx)) { | |
2048 | if (likely(min_nr <= nr && min_nr >= 0)) | |
2049 | ret = read_events(ioctx, min_nr, nr, events, until); | |
2050 | percpu_ref_put(&ioctx->users); | |
2051 | } | |
2052 | ||
2053 | return ret; | |
2054 | } | |
2055 | ||
1da177e4 LT |
2056 | /* io_getevents: |
2057 | * Attempts to read at least min_nr events and up to nr events from | |
642b5123 ST |
2058 | * the completion queue for the aio_context specified by ctx_id. If |
2059 | * it succeeds, the number of read events is returned. May fail with | |
2060 | * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is | |
2061 | * out of range, if timeout is out of range. May fail with -EFAULT | |
2062 | * if any of the memory specified is invalid. May return 0 or | |
2063 | * < min_nr if the timeout specified by timeout has elapsed | |
2064 | * before sufficient events are available, where timeout == NULL | |
2065 | * specifies an infinite timeout. Note that the timeout pointed to by | |
6900807c | 2066 | * timeout is relative. Will fail with -ENOSYS if not implemented. |
1da177e4 | 2067 | */ |
002c8976 HC |
2068 | SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, |
2069 | long, min_nr, | |
2070 | long, nr, | |
2071 | struct io_event __user *, events, | |
2072 | struct timespec __user *, timeout) | |
1da177e4 | 2073 | { |
fa2e62a5 | 2074 | struct timespec64 ts; |
7a074e96 CH |
2075 | int ret; |
2076 | ||
2077 | if (timeout && unlikely(get_timespec64(&ts, timeout))) | |
2078 | return -EFAULT; | |
2079 | ||
2080 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL); | |
2081 | if (!ret && signal_pending(current)) | |
2082 | ret = -EINTR; | |
2083 | return ret; | |
2084 | } | |
1da177e4 | 2085 | |
9ba546c0 CH |
2086 | struct __aio_sigset { |
2087 | const sigset_t __user *sigmask; | |
2088 | size_t sigsetsize; | |
2089 | }; | |
2090 | ||
7a074e96 CH |
2091 | SYSCALL_DEFINE6(io_pgetevents, |
2092 | aio_context_t, ctx_id, | |
2093 | long, min_nr, | |
2094 | long, nr, | |
2095 | struct io_event __user *, events, | |
2096 | struct timespec __user *, timeout, | |
2097 | const struct __aio_sigset __user *, usig) | |
2098 | { | |
2099 | struct __aio_sigset ksig = { NULL, }; | |
2100 | sigset_t ksigmask, sigsaved; | |
2101 | struct timespec64 ts; | |
2102 | int ret; | |
2103 | ||
2104 | if (timeout && unlikely(get_timespec64(&ts, timeout))) | |
2105 | return -EFAULT; | |
2106 | ||
2107 | if (usig && copy_from_user(&ksig, usig, sizeof(ksig))) | |
2108 | return -EFAULT; | |
2109 | ||
2110 | if (ksig.sigmask) { | |
2111 | if (ksig.sigsetsize != sizeof(sigset_t)) | |
2112 | return -EINVAL; | |
2113 | if (copy_from_user(&ksigmask, ksig.sigmask, sizeof(ksigmask))) | |
fa2e62a5 | 2114 | return -EFAULT; |
7a074e96 CH |
2115 | sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP)); |
2116 | sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); | |
2117 | } | |
2118 | ||
2119 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL); | |
2120 | if (signal_pending(current)) { | |
2121 | if (ksig.sigmask) { | |
2122 | current->saved_sigmask = sigsaved; | |
2123 | set_restore_sigmask(); | |
2124 | } | |
2125 | ||
2126 | if (!ret) | |
2127 | ret = -ERESTARTNOHAND; | |
2128 | } else { | |
2129 | if (ksig.sigmask) | |
2130 | sigprocmask(SIG_SETMASK, &sigsaved, NULL); | |
1da177e4 | 2131 | } |
fa2e62a5 | 2132 | |
7a074e96 | 2133 | return ret; |
1da177e4 | 2134 | } |
c00d2c7e AV |
2135 | |
2136 | #ifdef CONFIG_COMPAT | |
2137 | COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id, | |
2138 | compat_long_t, min_nr, | |
2139 | compat_long_t, nr, | |
2140 | struct io_event __user *, events, | |
2141 | struct compat_timespec __user *, timeout) | |
2142 | { | |
fa2e62a5 | 2143 | struct timespec64 t; |
7a074e96 CH |
2144 | int ret; |
2145 | ||
2146 | if (timeout && compat_get_timespec64(&t, timeout)) | |
2147 | return -EFAULT; | |
2148 | ||
2149 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL); | |
2150 | if (!ret && signal_pending(current)) | |
2151 | ret = -EINTR; | |
2152 | return ret; | |
2153 | } | |
2154 | ||
c00d2c7e | 2155 | |
7a074e96 CH |
2156 | struct __compat_aio_sigset { |
2157 | compat_sigset_t __user *sigmask; | |
2158 | compat_size_t sigsetsize; | |
2159 | }; | |
2160 | ||
2161 | COMPAT_SYSCALL_DEFINE6(io_pgetevents, | |
2162 | compat_aio_context_t, ctx_id, | |
2163 | compat_long_t, min_nr, | |
2164 | compat_long_t, nr, | |
2165 | struct io_event __user *, events, | |
2166 | struct compat_timespec __user *, timeout, | |
2167 | const struct __compat_aio_sigset __user *, usig) | |
2168 | { | |
2169 | struct __compat_aio_sigset ksig = { NULL, }; | |
2170 | sigset_t ksigmask, sigsaved; | |
2171 | struct timespec64 t; | |
2172 | int ret; | |
2173 | ||
2174 | if (timeout && compat_get_timespec64(&t, timeout)) | |
2175 | return -EFAULT; | |
2176 | ||
2177 | if (usig && copy_from_user(&ksig, usig, sizeof(ksig))) | |
2178 | return -EFAULT; | |
2179 | ||
2180 | if (ksig.sigmask) { | |
2181 | if (ksig.sigsetsize != sizeof(compat_sigset_t)) | |
2182 | return -EINVAL; | |
2183 | if (get_compat_sigset(&ksigmask, ksig.sigmask)) | |
c00d2c7e | 2184 | return -EFAULT; |
7a074e96 CH |
2185 | sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP)); |
2186 | sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); | |
2187 | } | |
c00d2c7e | 2188 | |
7a074e96 CH |
2189 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL); |
2190 | if (signal_pending(current)) { | |
2191 | if (ksig.sigmask) { | |
2192 | current->saved_sigmask = sigsaved; | |
2193 | set_restore_sigmask(); | |
2194 | } | |
2195 | if (!ret) | |
2196 | ret = -ERESTARTNOHAND; | |
2197 | } else { | |
2198 | if (ksig.sigmask) | |
2199 | sigprocmask(SIG_SETMASK, &sigsaved, NULL); | |
c00d2c7e | 2200 | } |
fa2e62a5 | 2201 | |
7a074e96 | 2202 | return ret; |
c00d2c7e AV |
2203 | } |
2204 | #endif |