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. | |
2c14fa83 | 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> |
027445c3 | 22 | #include <linux/uio.h> |
1da177e4 | 23 | |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
1da177e4 LT |
25 | #include <linux/fs.h> |
26 | #include <linux/file.h> | |
27 | #include <linux/mm.h> | |
28 | #include <linux/mman.h> | |
3d2d827f | 29 | #include <linux/mmu_context.h> |
e1bdd5f2 | 30 | #include <linux/percpu.h> |
1da177e4 LT |
31 | #include <linux/slab.h> |
32 | #include <linux/timer.h> | |
33 | #include <linux/aio.h> | |
34 | #include <linux/highmem.h> | |
35 | #include <linux/workqueue.h> | |
36 | #include <linux/security.h> | |
9c3060be | 37 | #include <linux/eventfd.h> |
cfb1e33e | 38 | #include <linux/blkdev.h> |
9d85cba7 | 39 | #include <linux/compat.h> |
36bc08cc GZ |
40 | #include <linux/migrate.h> |
41 | #include <linux/ramfs.h> | |
723be6e3 | 42 | #include <linux/percpu-refcount.h> |
71ad7490 | 43 | #include <linux/mount.h> |
1da177e4 LT |
44 | |
45 | #include <asm/kmap_types.h> | |
7c0f6ba6 | 46 | #include <linux/uaccess.h> |
1da177e4 | 47 | |
68d70d03 AV |
48 | #include "internal.h" |
49 | ||
f3a2752a CH |
50 | #define KIOCB_KEY 0 |
51 | ||
4e179bca KO |
52 | #define AIO_RING_MAGIC 0xa10a10a1 |
53 | #define AIO_RING_COMPAT_FEATURES 1 | |
54 | #define AIO_RING_INCOMPAT_FEATURES 0 | |
55 | struct aio_ring { | |
56 | unsigned id; /* kernel internal index number */ | |
57 | unsigned nr; /* number of io_events */ | |
fa8a53c3 BL |
58 | unsigned head; /* Written to by userland or under ring_lock |
59 | * mutex by aio_read_events_ring(). */ | |
4e179bca KO |
60 | unsigned tail; |
61 | ||
62 | unsigned magic; | |
63 | unsigned compat_features; | |
64 | unsigned incompat_features; | |
65 | unsigned header_length; /* size of aio_ring */ | |
66 | ||
67 | ||
68 | struct io_event io_events[0]; | |
69 | }; /* 128 bytes + ring size */ | |
70 | ||
71 | #define AIO_RING_PAGES 8 | |
4e179bca | 72 | |
db446a08 | 73 | struct kioctx_table { |
d0264c01 TH |
74 | struct rcu_head rcu; |
75 | unsigned nr; | |
76 | struct kioctx __rcu *table[]; | |
db446a08 BL |
77 | }; |
78 | ||
e1bdd5f2 KO |
79 | struct kioctx_cpu { |
80 | unsigned reqs_available; | |
81 | }; | |
82 | ||
dc48e56d JA |
83 | struct ctx_rq_wait { |
84 | struct completion comp; | |
85 | atomic_t count; | |
86 | }; | |
87 | ||
4e179bca | 88 | struct kioctx { |
723be6e3 | 89 | struct percpu_ref users; |
36f55889 | 90 | atomic_t dead; |
4e179bca | 91 | |
e34ecee2 KO |
92 | struct percpu_ref reqs; |
93 | ||
4e179bca | 94 | unsigned long user_id; |
4e179bca | 95 | |
e1bdd5f2 KO |
96 | struct __percpu kioctx_cpu *cpu; |
97 | ||
98 | /* | |
99 | * For percpu reqs_available, number of slots we move to/from global | |
100 | * counter at a time: | |
101 | */ | |
102 | unsigned req_batch; | |
3e845ce0 KO |
103 | /* |
104 | * This is what userspace passed to io_setup(), it's not used for | |
105 | * anything but counting against the global max_reqs quota. | |
106 | * | |
58c85dc2 | 107 | * The real limit is nr_events - 1, which will be larger (see |
3e845ce0 KO |
108 | * aio_setup_ring()) |
109 | */ | |
4e179bca KO |
110 | unsigned max_reqs; |
111 | ||
58c85dc2 KO |
112 | /* Size of ringbuffer, in units of struct io_event */ |
113 | unsigned nr_events; | |
4e179bca | 114 | |
58c85dc2 KO |
115 | unsigned long mmap_base; |
116 | unsigned long mmap_size; | |
117 | ||
118 | struct page **ring_pages; | |
119 | long nr_pages; | |
120 | ||
f729863a | 121 | struct rcu_work free_rwork; /* see free_ioctx() */ |
4e23bcae | 122 | |
e02ba72a AP |
123 | /* |
124 | * signals when all in-flight requests are done | |
125 | */ | |
dc48e56d | 126 | struct ctx_rq_wait *rq_wait; |
e02ba72a | 127 | |
4e23bcae | 128 | struct { |
34e83fc6 KO |
129 | /* |
130 | * This counts the number of available slots in the ringbuffer, | |
131 | * so we avoid overflowing it: it's decremented (if positive) | |
132 | * when allocating a kiocb and incremented when the resulting | |
133 | * io_event is pulled off the ringbuffer. | |
e1bdd5f2 KO |
134 | * |
135 | * We batch accesses to it with a percpu version. | |
34e83fc6 KO |
136 | */ |
137 | atomic_t reqs_available; | |
4e23bcae KO |
138 | } ____cacheline_aligned_in_smp; |
139 | ||
140 | struct { | |
141 | spinlock_t ctx_lock; | |
142 | struct list_head active_reqs; /* used for cancellation */ | |
143 | } ____cacheline_aligned_in_smp; | |
144 | ||
58c85dc2 KO |
145 | struct { |
146 | struct mutex ring_lock; | |
4e23bcae KO |
147 | wait_queue_head_t wait; |
148 | } ____cacheline_aligned_in_smp; | |
58c85dc2 KO |
149 | |
150 | struct { | |
151 | unsigned tail; | |
d856f32a | 152 | unsigned completed_events; |
58c85dc2 | 153 | spinlock_t completion_lock; |
4e23bcae | 154 | } ____cacheline_aligned_in_smp; |
58c85dc2 KO |
155 | |
156 | struct page *internal_pages[AIO_RING_PAGES]; | |
36bc08cc | 157 | struct file *aio_ring_file; |
db446a08 BL |
158 | |
159 | unsigned id; | |
4e179bca KO |
160 | }; |
161 | ||
a3c0d439 CH |
162 | struct fsync_iocb { |
163 | struct work_struct work; | |
164 | struct file *file; | |
165 | bool datasync; | |
166 | }; | |
167 | ||
2c14fa83 CH |
168 | struct poll_iocb { |
169 | struct file *file; | |
170 | __poll_t events; | |
171 | struct wait_queue_head *head; | |
172 | ||
173 | union { | |
174 | struct wait_queue_entry wait; | |
175 | struct work_struct work; | |
176 | }; | |
177 | }; | |
178 | ||
04b2fa9f | 179 | struct aio_kiocb { |
54843f87 CH |
180 | union { |
181 | struct kiocb rw; | |
a3c0d439 | 182 | struct fsync_iocb fsync; |
2c14fa83 | 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 */ | |
194 | ||
195 | /* | |
196 | * If the aio_resfd field of the userspace iocb is not zero, | |
197 | * this is the underlying eventfd context to deliver events to. | |
198 | */ | |
199 | struct eventfd_ctx *ki_eventfd; | |
200 | }; | |
201 | ||
1da177e4 | 202 | /*------ sysctl variables----*/ |
d55b5fda ZB |
203 | static DEFINE_SPINLOCK(aio_nr_lock); |
204 | unsigned long aio_nr; /* current system wide number of aio requests */ | |
205 | unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ | |
1da177e4 LT |
206 | /*----end sysctl variables---*/ |
207 | ||
e18b890b CL |
208 | static struct kmem_cache *kiocb_cachep; |
209 | static struct kmem_cache *kioctx_cachep; | |
1da177e4 | 210 | |
71ad7490 BL |
211 | static struct vfsmount *aio_mnt; |
212 | ||
213 | static const struct file_operations aio_ring_fops; | |
214 | static const struct address_space_operations aio_ctx_aops; | |
215 | ||
216 | static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages) | |
217 | { | |
218 | struct qstr this = QSTR_INIT("[aio]", 5); | |
219 | struct file *file; | |
220 | struct path path; | |
221 | struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb); | |
7f62656b DC |
222 | if (IS_ERR(inode)) |
223 | return ERR_CAST(inode); | |
71ad7490 BL |
224 | |
225 | inode->i_mapping->a_ops = &aio_ctx_aops; | |
226 | inode->i_mapping->private_data = ctx; | |
227 | inode->i_size = PAGE_SIZE * nr_pages; | |
228 | ||
229 | path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this); | |
230 | if (!path.dentry) { | |
231 | iput(inode); | |
232 | return ERR_PTR(-ENOMEM); | |
233 | } | |
234 | path.mnt = mntget(aio_mnt); | |
235 | ||
236 | d_instantiate(path.dentry, inode); | |
c9c554f2 AV |
237 | file = alloc_file(&path, O_RDWR, &aio_ring_fops); |
238 | if (IS_ERR(file)) | |
71ad7490 | 239 | path_put(&path); |
71ad7490 BL |
240 | return file; |
241 | } | |
242 | ||
243 | static struct dentry *aio_mount(struct file_system_type *fs_type, | |
244 | int flags, const char *dev_name, void *data) | |
245 | { | |
246 | static const struct dentry_operations ops = { | |
247 | .d_dname = simple_dname, | |
248 | }; | |
22f6b4d3 JH |
249 | struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops, |
250 | AIO_RING_MAGIC); | |
251 | ||
252 | if (!IS_ERR(root)) | |
253 | root->d_sb->s_iflags |= SB_I_NOEXEC; | |
254 | return root; | |
71ad7490 BL |
255 | } |
256 | ||
1da177e4 LT |
257 | /* aio_setup |
258 | * Creates the slab caches used by the aio routines, panic on | |
259 | * failure as this is done early during the boot sequence. | |
260 | */ | |
261 | static int __init aio_setup(void) | |
262 | { | |
71ad7490 BL |
263 | static struct file_system_type aio_fs = { |
264 | .name = "aio", | |
265 | .mount = aio_mount, | |
266 | .kill_sb = kill_anon_super, | |
267 | }; | |
268 | aio_mnt = kern_mount(&aio_fs); | |
269 | if (IS_ERR(aio_mnt)) | |
270 | panic("Failed to create aio fs mount."); | |
271 | ||
04b2fa9f | 272 | kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
0a31bd5f | 273 | kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
1da177e4 LT |
274 | return 0; |
275 | } | |
385773e0 | 276 | __initcall(aio_setup); |
1da177e4 | 277 | |
5e9ae2e5 BL |
278 | static void put_aio_ring_file(struct kioctx *ctx) |
279 | { | |
280 | struct file *aio_ring_file = ctx->aio_ring_file; | |
de04e769 RV |
281 | struct address_space *i_mapping; |
282 | ||
5e9ae2e5 | 283 | if (aio_ring_file) { |
45063097 | 284 | truncate_setsize(file_inode(aio_ring_file), 0); |
5e9ae2e5 BL |
285 | |
286 | /* Prevent further access to the kioctx from migratepages */ | |
45063097 | 287 | i_mapping = aio_ring_file->f_mapping; |
de04e769 RV |
288 | spin_lock(&i_mapping->private_lock); |
289 | i_mapping->private_data = NULL; | |
5e9ae2e5 | 290 | ctx->aio_ring_file = NULL; |
de04e769 | 291 | spin_unlock(&i_mapping->private_lock); |
5e9ae2e5 BL |
292 | |
293 | fput(aio_ring_file); | |
294 | } | |
295 | } | |
296 | ||
1da177e4 LT |
297 | static void aio_free_ring(struct kioctx *ctx) |
298 | { | |
36bc08cc | 299 | int i; |
1da177e4 | 300 | |
fa8a53c3 BL |
301 | /* Disconnect the kiotx from the ring file. This prevents future |
302 | * accesses to the kioctx from page migration. | |
303 | */ | |
304 | put_aio_ring_file(ctx); | |
305 | ||
36bc08cc | 306 | for (i = 0; i < ctx->nr_pages; i++) { |
8e321fef | 307 | struct page *page; |
36bc08cc GZ |
308 | pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i, |
309 | page_count(ctx->ring_pages[i])); | |
8e321fef BL |
310 | page = ctx->ring_pages[i]; |
311 | if (!page) | |
312 | continue; | |
313 | ctx->ring_pages[i] = NULL; | |
314 | put_page(page); | |
36bc08cc | 315 | } |
1da177e4 | 316 | |
ddb8c45b | 317 | if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) { |
58c85dc2 | 318 | kfree(ctx->ring_pages); |
ddb8c45b SL |
319 | ctx->ring_pages = NULL; |
320 | } | |
36bc08cc GZ |
321 | } |
322 | ||
5477e70a | 323 | static int aio_ring_mremap(struct vm_area_struct *vma) |
e4a0d3e7 | 324 | { |
5477e70a | 325 | struct file *file = vma->vm_file; |
e4a0d3e7 PE |
326 | struct mm_struct *mm = vma->vm_mm; |
327 | struct kioctx_table *table; | |
b2edffdd | 328 | int i, res = -EINVAL; |
e4a0d3e7 PE |
329 | |
330 | spin_lock(&mm->ioctx_lock); | |
331 | rcu_read_lock(); | |
332 | table = rcu_dereference(mm->ioctx_table); | |
333 | for (i = 0; i < table->nr; i++) { | |
334 | struct kioctx *ctx; | |
335 | ||
d0264c01 | 336 | ctx = rcu_dereference(table->table[i]); |
e4a0d3e7 | 337 | if (ctx && ctx->aio_ring_file == file) { |
b2edffdd AV |
338 | if (!atomic_read(&ctx->dead)) { |
339 | ctx->user_id = ctx->mmap_base = vma->vm_start; | |
340 | res = 0; | |
341 | } | |
e4a0d3e7 PE |
342 | break; |
343 | } | |
344 | } | |
345 | ||
346 | rcu_read_unlock(); | |
347 | spin_unlock(&mm->ioctx_lock); | |
b2edffdd | 348 | return res; |
e4a0d3e7 PE |
349 | } |
350 | ||
5477e70a ON |
351 | static const struct vm_operations_struct aio_ring_vm_ops = { |
352 | .mremap = aio_ring_mremap, | |
353 | #if IS_ENABLED(CONFIG_MMU) | |
354 | .fault = filemap_fault, | |
355 | .map_pages = filemap_map_pages, | |
356 | .page_mkwrite = filemap_page_mkwrite, | |
357 | #endif | |
358 | }; | |
359 | ||
360 | static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma) | |
361 | { | |
362 | vma->vm_flags |= VM_DONTEXPAND; | |
363 | vma->vm_ops = &aio_ring_vm_ops; | |
364 | return 0; | |
365 | } | |
366 | ||
36bc08cc GZ |
367 | static const struct file_operations aio_ring_fops = { |
368 | .mmap = aio_ring_mmap, | |
369 | }; | |
370 | ||
0c45355f | 371 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc GZ |
372 | static int aio_migratepage(struct address_space *mapping, struct page *new, |
373 | struct page *old, enum migrate_mode mode) | |
374 | { | |
5e9ae2e5 | 375 | struct kioctx *ctx; |
36bc08cc | 376 | unsigned long flags; |
fa8a53c3 | 377 | pgoff_t idx; |
36bc08cc GZ |
378 | int rc; |
379 | ||
2916ecc0 JG |
380 | /* |
381 | * We cannot support the _NO_COPY case here, because copy needs to | |
382 | * happen under the ctx->completion_lock. That does not work with the | |
383 | * migration workflow of MIGRATE_SYNC_NO_COPY. | |
384 | */ | |
385 | if (mode == MIGRATE_SYNC_NO_COPY) | |
386 | return -EINVAL; | |
387 | ||
8e321fef BL |
388 | rc = 0; |
389 | ||
fa8a53c3 | 390 | /* mapping->private_lock here protects against the kioctx teardown. */ |
8e321fef BL |
391 | spin_lock(&mapping->private_lock); |
392 | ctx = mapping->private_data; | |
fa8a53c3 BL |
393 | if (!ctx) { |
394 | rc = -EINVAL; | |
395 | goto out; | |
396 | } | |
397 | ||
398 | /* The ring_lock mutex. The prevents aio_read_events() from writing | |
399 | * to the ring's head, and prevents page migration from mucking in | |
400 | * a partially initialized kiotx. | |
401 | */ | |
402 | if (!mutex_trylock(&ctx->ring_lock)) { | |
403 | rc = -EAGAIN; | |
404 | goto out; | |
405 | } | |
406 | ||
407 | idx = old->index; | |
408 | if (idx < (pgoff_t)ctx->nr_pages) { | |
409 | /* Make sure the old page hasn't already been changed */ | |
410 | if (ctx->ring_pages[idx] != old) | |
411 | rc = -EAGAIN; | |
8e321fef BL |
412 | } else |
413 | rc = -EINVAL; | |
8e321fef BL |
414 | |
415 | if (rc != 0) | |
fa8a53c3 | 416 | goto out_unlock; |
8e321fef | 417 | |
36bc08cc GZ |
418 | /* Writeback must be complete */ |
419 | BUG_ON(PageWriteback(old)); | |
8e321fef | 420 | get_page(new); |
36bc08cc | 421 | |
8e321fef | 422 | rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1); |
36bc08cc | 423 | if (rc != MIGRATEPAGE_SUCCESS) { |
8e321fef | 424 | put_page(new); |
fa8a53c3 | 425 | goto out_unlock; |
36bc08cc GZ |
426 | } |
427 | ||
fa8a53c3 BL |
428 | /* Take completion_lock to prevent other writes to the ring buffer |
429 | * while the old page is copied to the new. This prevents new | |
430 | * events from being lost. | |
5e9ae2e5 | 431 | */ |
fa8a53c3 BL |
432 | spin_lock_irqsave(&ctx->completion_lock, flags); |
433 | migrate_page_copy(new, old); | |
434 | BUG_ON(ctx->ring_pages[idx] != old); | |
435 | ctx->ring_pages[idx] = new; | |
436 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
36bc08cc | 437 | |
fa8a53c3 BL |
438 | /* The old page is no longer accessible. */ |
439 | put_page(old); | |
8e321fef | 440 | |
fa8a53c3 BL |
441 | out_unlock: |
442 | mutex_unlock(&ctx->ring_lock); | |
443 | out: | |
444 | spin_unlock(&mapping->private_lock); | |
36bc08cc | 445 | return rc; |
1da177e4 | 446 | } |
0c45355f | 447 | #endif |
1da177e4 | 448 | |
36bc08cc | 449 | static const struct address_space_operations aio_ctx_aops = { |
835f252c | 450 | .set_page_dirty = __set_page_dirty_no_writeback, |
0c45355f | 451 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc | 452 | .migratepage = aio_migratepage, |
0c45355f | 453 | #endif |
36bc08cc GZ |
454 | }; |
455 | ||
2a8a9867 | 456 | static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events) |
1da177e4 LT |
457 | { |
458 | struct aio_ring *ring; | |
41003a7b | 459 | struct mm_struct *mm = current->mm; |
3dc9acb6 | 460 | unsigned long size, unused; |
1da177e4 | 461 | int nr_pages; |
36bc08cc GZ |
462 | int i; |
463 | struct file *file; | |
1da177e4 LT |
464 | |
465 | /* Compensate for the ring buffer's head/tail overlap entry */ | |
466 | nr_events += 2; /* 1 is required, 2 for good luck */ | |
467 | ||
468 | size = sizeof(struct aio_ring); | |
469 | size += sizeof(struct io_event) * nr_events; | |
1da177e4 | 470 | |
36bc08cc | 471 | nr_pages = PFN_UP(size); |
1da177e4 LT |
472 | if (nr_pages < 0) |
473 | return -EINVAL; | |
474 | ||
71ad7490 | 475 | file = aio_private_file(ctx, nr_pages); |
36bc08cc GZ |
476 | if (IS_ERR(file)) { |
477 | ctx->aio_ring_file = NULL; | |
fa8a53c3 | 478 | return -ENOMEM; |
36bc08cc GZ |
479 | } |
480 | ||
3dc9acb6 LT |
481 | ctx->aio_ring_file = file; |
482 | nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) | |
483 | / sizeof(struct io_event); | |
484 | ||
485 | ctx->ring_pages = ctx->internal_pages; | |
486 | if (nr_pages > AIO_RING_PAGES) { | |
487 | ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *), | |
488 | GFP_KERNEL); | |
489 | if (!ctx->ring_pages) { | |
490 | put_aio_ring_file(ctx); | |
491 | return -ENOMEM; | |
492 | } | |
493 | } | |
494 | ||
36bc08cc GZ |
495 | for (i = 0; i < nr_pages; i++) { |
496 | struct page *page; | |
45063097 | 497 | page = find_or_create_page(file->f_mapping, |
36bc08cc GZ |
498 | i, GFP_HIGHUSER | __GFP_ZERO); |
499 | if (!page) | |
500 | break; | |
501 | pr_debug("pid(%d) page[%d]->count=%d\n", | |
502 | current->pid, i, page_count(page)); | |
503 | SetPageUptodate(page); | |
36bc08cc | 504 | unlock_page(page); |
3dc9acb6 LT |
505 | |
506 | ctx->ring_pages[i] = page; | |
36bc08cc | 507 | } |
3dc9acb6 | 508 | ctx->nr_pages = i; |
1da177e4 | 509 | |
3dc9acb6 LT |
510 | if (unlikely(i != nr_pages)) { |
511 | aio_free_ring(ctx); | |
fa8a53c3 | 512 | return -ENOMEM; |
1da177e4 LT |
513 | } |
514 | ||
58c85dc2 KO |
515 | ctx->mmap_size = nr_pages * PAGE_SIZE; |
516 | pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size); | |
36bc08cc | 517 | |
013373e8 MH |
518 | if (down_write_killable(&mm->mmap_sem)) { |
519 | ctx->mmap_size = 0; | |
520 | aio_free_ring(ctx); | |
521 | return -EINTR; | |
522 | } | |
523 | ||
36bc08cc GZ |
524 | ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size, |
525 | PROT_READ | PROT_WRITE, | |
897ab3e0 | 526 | MAP_SHARED, 0, &unused, NULL); |
3dc9acb6 | 527 | up_write(&mm->mmap_sem); |
58c85dc2 | 528 | if (IS_ERR((void *)ctx->mmap_base)) { |
58c85dc2 | 529 | ctx->mmap_size = 0; |
1da177e4 | 530 | aio_free_ring(ctx); |
fa8a53c3 | 531 | return -ENOMEM; |
1da177e4 LT |
532 | } |
533 | ||
58c85dc2 | 534 | pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base); |
d6c355c7 | 535 | |
58c85dc2 KO |
536 | ctx->user_id = ctx->mmap_base; |
537 | ctx->nr_events = nr_events; /* trusted copy */ | |
1da177e4 | 538 | |
58c85dc2 | 539 | ring = kmap_atomic(ctx->ring_pages[0]); |
1da177e4 | 540 | ring->nr = nr_events; /* user copy */ |
db446a08 | 541 | ring->id = ~0U; |
1da177e4 LT |
542 | ring->head = ring->tail = 0; |
543 | ring->magic = AIO_RING_MAGIC; | |
544 | ring->compat_features = AIO_RING_COMPAT_FEATURES; | |
545 | ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; | |
546 | ring->header_length = sizeof(struct aio_ring); | |
e8e3c3d6 | 547 | kunmap_atomic(ring); |
58c85dc2 | 548 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 LT |
549 | |
550 | return 0; | |
551 | } | |
552 | ||
1da177e4 LT |
553 | #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) |
554 | #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) | |
555 | #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) | |
556 | ||
04b2fa9f | 557 | void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel) |
0460fef2 | 558 | { |
54843f87 | 559 | struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw); |
0460fef2 KO |
560 | struct kioctx *ctx = req->ki_ctx; |
561 | unsigned long flags; | |
562 | ||
75321b50 CH |
563 | if (WARN_ON_ONCE(!list_empty(&req->ki_list))) |
564 | return; | |
0460fef2 | 565 | |
75321b50 CH |
566 | spin_lock_irqsave(&ctx->ctx_lock, flags); |
567 | list_add_tail(&req->ki_list, &ctx->active_reqs); | |
0460fef2 | 568 | req->ki_cancel = cancel; |
0460fef2 KO |
569 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); |
570 | } | |
571 | EXPORT_SYMBOL(kiocb_set_cancel_fn); | |
572 | ||
a6d7cff4 TH |
573 | /* |
574 | * free_ioctx() should be RCU delayed to synchronize against the RCU | |
575 | * protected lookup_ioctx() and also needs process context to call | |
f729863a | 576 | * aio_free_ring(). Use rcu_work. |
a6d7cff4 | 577 | */ |
e34ecee2 | 578 | static void free_ioctx(struct work_struct *work) |
36f55889 | 579 | { |
f729863a TH |
580 | struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx, |
581 | free_rwork); | |
e34ecee2 | 582 | pr_debug("freeing %p\n", ctx); |
e1bdd5f2 | 583 | |
e34ecee2 | 584 | aio_free_ring(ctx); |
e1bdd5f2 | 585 | free_percpu(ctx->cpu); |
9a1049da TH |
586 | percpu_ref_exit(&ctx->reqs); |
587 | percpu_ref_exit(&ctx->users); | |
36f55889 KO |
588 | kmem_cache_free(kioctx_cachep, ctx); |
589 | } | |
590 | ||
e34ecee2 KO |
591 | static void free_ioctx_reqs(struct percpu_ref *ref) |
592 | { | |
593 | struct kioctx *ctx = container_of(ref, struct kioctx, reqs); | |
594 | ||
e02ba72a | 595 | /* At this point we know that there are no any in-flight requests */ |
dc48e56d JA |
596 | if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count)) |
597 | complete(&ctx->rq_wait->comp); | |
e02ba72a | 598 | |
a6d7cff4 | 599 | /* Synchronize against RCU protected table->table[] dereferences */ |
f729863a TH |
600 | INIT_RCU_WORK(&ctx->free_rwork, free_ioctx); |
601 | queue_rcu_work(system_wq, &ctx->free_rwork); | |
e34ecee2 KO |
602 | } |
603 | ||
36f55889 KO |
604 | /* |
605 | * When this function runs, the kioctx has been removed from the "hash table" | |
606 | * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - | |
607 | * now it's safe to cancel any that need to be. | |
608 | */ | |
e34ecee2 | 609 | static void free_ioctx_users(struct percpu_ref *ref) |
36f55889 | 610 | { |
e34ecee2 | 611 | struct kioctx *ctx = container_of(ref, struct kioctx, users); |
04b2fa9f | 612 | struct aio_kiocb *req; |
36f55889 KO |
613 | |
614 | spin_lock_irq(&ctx->ctx_lock); | |
615 | ||
616 | while (!list_empty(&ctx->active_reqs)) { | |
617 | req = list_first_entry(&ctx->active_reqs, | |
04b2fa9f | 618 | struct aio_kiocb, ki_list); |
888933f8 | 619 | req->ki_cancel(&req->rw); |
4faa9996 | 620 | list_del_init(&req->ki_list); |
36f55889 KO |
621 | } |
622 | ||
623 | spin_unlock_irq(&ctx->ctx_lock); | |
624 | ||
e34ecee2 KO |
625 | percpu_ref_kill(&ctx->reqs); |
626 | percpu_ref_put(&ctx->reqs); | |
36f55889 KO |
627 | } |
628 | ||
db446a08 BL |
629 | static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm) |
630 | { | |
631 | unsigned i, new_nr; | |
632 | struct kioctx_table *table, *old; | |
633 | struct aio_ring *ring; | |
634 | ||
635 | spin_lock(&mm->ioctx_lock); | |
855ef0de | 636 | table = rcu_dereference_raw(mm->ioctx_table); |
db446a08 BL |
637 | |
638 | while (1) { | |
639 | if (table) | |
640 | for (i = 0; i < table->nr; i++) | |
d0264c01 | 641 | if (!rcu_access_pointer(table->table[i])) { |
db446a08 | 642 | ctx->id = i; |
d0264c01 | 643 | rcu_assign_pointer(table->table[i], ctx); |
db446a08 BL |
644 | spin_unlock(&mm->ioctx_lock); |
645 | ||
fa8a53c3 BL |
646 | /* While kioctx setup is in progress, |
647 | * we are protected from page migration | |
648 | * changes ring_pages by ->ring_lock. | |
649 | */ | |
db446a08 BL |
650 | ring = kmap_atomic(ctx->ring_pages[0]); |
651 | ring->id = ctx->id; | |
652 | kunmap_atomic(ring); | |
653 | return 0; | |
654 | } | |
655 | ||
656 | new_nr = (table ? table->nr : 1) * 4; | |
db446a08 BL |
657 | spin_unlock(&mm->ioctx_lock); |
658 | ||
659 | table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) * | |
660 | new_nr, GFP_KERNEL); | |
661 | if (!table) | |
662 | return -ENOMEM; | |
663 | ||
664 | table->nr = new_nr; | |
665 | ||
666 | spin_lock(&mm->ioctx_lock); | |
855ef0de | 667 | old = rcu_dereference_raw(mm->ioctx_table); |
db446a08 BL |
668 | |
669 | if (!old) { | |
670 | rcu_assign_pointer(mm->ioctx_table, table); | |
671 | } else if (table->nr > old->nr) { | |
672 | memcpy(table->table, old->table, | |
673 | old->nr * sizeof(struct kioctx *)); | |
674 | ||
675 | rcu_assign_pointer(mm->ioctx_table, table); | |
676 | kfree_rcu(old, rcu); | |
677 | } else { | |
678 | kfree(table); | |
679 | table = old; | |
680 | } | |
681 | } | |
682 | } | |
683 | ||
e34ecee2 KO |
684 | static void aio_nr_sub(unsigned nr) |
685 | { | |
686 | spin_lock(&aio_nr_lock); | |
687 | if (WARN_ON(aio_nr - nr > aio_nr)) | |
688 | aio_nr = 0; | |
689 | else | |
690 | aio_nr -= nr; | |
691 | spin_unlock(&aio_nr_lock); | |
692 | } | |
693 | ||
1da177e4 LT |
694 | /* ioctx_alloc |
695 | * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. | |
696 | */ | |
697 | static struct kioctx *ioctx_alloc(unsigned nr_events) | |
698 | { | |
41003a7b | 699 | struct mm_struct *mm = current->mm; |
1da177e4 | 700 | struct kioctx *ctx; |
e23754f8 | 701 | int err = -ENOMEM; |
1da177e4 | 702 | |
2a8a9867 MFO |
703 | /* |
704 | * Store the original nr_events -- what userspace passed to io_setup(), | |
705 | * for counting against the global limit -- before it changes. | |
706 | */ | |
707 | unsigned int max_reqs = nr_events; | |
708 | ||
e1bdd5f2 KO |
709 | /* |
710 | * We keep track of the number of available ringbuffer slots, to prevent | |
711 | * overflow (reqs_available), and we also use percpu counters for this. | |
712 | * | |
713 | * So since up to half the slots might be on other cpu's percpu counters | |
714 | * and unavailable, double nr_events so userspace sees what they | |
715 | * expected: additionally, we move req_batch slots to/from percpu | |
716 | * counters at a time, so make sure that isn't 0: | |
717 | */ | |
718 | nr_events = max(nr_events, num_possible_cpus() * 4); | |
719 | nr_events *= 2; | |
720 | ||
1da177e4 | 721 | /* Prevent overflows */ |
08397acd | 722 | if (nr_events > (0x10000000U / sizeof(struct io_event))) { |
1da177e4 LT |
723 | pr_debug("ENOMEM: nr_events too high\n"); |
724 | return ERR_PTR(-EINVAL); | |
725 | } | |
726 | ||
2a8a9867 | 727 | if (!nr_events || (unsigned long)max_reqs > aio_max_nr) |
1da177e4 LT |
728 | return ERR_PTR(-EAGAIN); |
729 | ||
c3762229 | 730 | ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); |
1da177e4 LT |
731 | if (!ctx) |
732 | return ERR_PTR(-ENOMEM); | |
733 | ||
2a8a9867 | 734 | ctx->max_reqs = max_reqs; |
1da177e4 | 735 | |
1da177e4 | 736 | spin_lock_init(&ctx->ctx_lock); |
0460fef2 | 737 | spin_lock_init(&ctx->completion_lock); |
58c85dc2 | 738 | mutex_init(&ctx->ring_lock); |
fa8a53c3 BL |
739 | /* Protect against page migration throughout kiotx setup by keeping |
740 | * the ring_lock mutex held until setup is complete. */ | |
741 | mutex_lock(&ctx->ring_lock); | |
1da177e4 LT |
742 | init_waitqueue_head(&ctx->wait); |
743 | ||
744 | INIT_LIST_HEAD(&ctx->active_reqs); | |
1da177e4 | 745 | |
2aad2a86 | 746 | if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL)) |
fa8a53c3 BL |
747 | goto err; |
748 | ||
2aad2a86 | 749 | if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL)) |
fa8a53c3 BL |
750 | goto err; |
751 | ||
e1bdd5f2 KO |
752 | ctx->cpu = alloc_percpu(struct kioctx_cpu); |
753 | if (!ctx->cpu) | |
e34ecee2 | 754 | goto err; |
1da177e4 | 755 | |
2a8a9867 | 756 | err = aio_setup_ring(ctx, nr_events); |
fa8a53c3 | 757 | if (err < 0) |
e34ecee2 | 758 | goto err; |
e1bdd5f2 | 759 | |
34e83fc6 | 760 | atomic_set(&ctx->reqs_available, ctx->nr_events - 1); |
e1bdd5f2 | 761 | ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4); |
6878ea72 BL |
762 | if (ctx->req_batch < 1) |
763 | ctx->req_batch = 1; | |
34e83fc6 | 764 | |
1da177e4 | 765 | /* limit the number of system wide aios */ |
9fa1cb39 | 766 | spin_lock(&aio_nr_lock); |
2a8a9867 MFO |
767 | if (aio_nr + ctx->max_reqs > aio_max_nr || |
768 | aio_nr + ctx->max_reqs < aio_nr) { | |
9fa1cb39 | 769 | spin_unlock(&aio_nr_lock); |
e34ecee2 | 770 | err = -EAGAIN; |
d1b94327 | 771 | goto err_ctx; |
2dd542b7 AV |
772 | } |
773 | aio_nr += ctx->max_reqs; | |
9fa1cb39 | 774 | spin_unlock(&aio_nr_lock); |
1da177e4 | 775 | |
1881686f BL |
776 | percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */ |
777 | percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */ | |
723be6e3 | 778 | |
da90382c BL |
779 | err = ioctx_add_table(ctx, mm); |
780 | if (err) | |
e34ecee2 | 781 | goto err_cleanup; |
da90382c | 782 | |
fa8a53c3 BL |
783 | /* Release the ring_lock mutex now that all setup is complete. */ |
784 | mutex_unlock(&ctx->ring_lock); | |
785 | ||
caf4167a | 786 | pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", |
58c85dc2 | 787 | ctx, ctx->user_id, mm, ctx->nr_events); |
1da177e4 LT |
788 | return ctx; |
789 | ||
e34ecee2 KO |
790 | err_cleanup: |
791 | aio_nr_sub(ctx->max_reqs); | |
d1b94327 | 792 | err_ctx: |
deeb8525 AV |
793 | atomic_set(&ctx->dead, 1); |
794 | if (ctx->mmap_size) | |
795 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
d1b94327 | 796 | aio_free_ring(ctx); |
e34ecee2 | 797 | err: |
fa8a53c3 | 798 | mutex_unlock(&ctx->ring_lock); |
e1bdd5f2 | 799 | free_percpu(ctx->cpu); |
9a1049da TH |
800 | percpu_ref_exit(&ctx->reqs); |
801 | percpu_ref_exit(&ctx->users); | |
1da177e4 | 802 | kmem_cache_free(kioctx_cachep, ctx); |
caf4167a | 803 | pr_debug("error allocating ioctx %d\n", err); |
e23754f8 | 804 | return ERR_PTR(err); |
1da177e4 LT |
805 | } |
806 | ||
36f55889 KO |
807 | /* kill_ioctx |
808 | * Cancels all outstanding aio requests on an aio context. Used | |
809 | * when the processes owning a context have all exited to encourage | |
810 | * the rapid destruction of the kioctx. | |
811 | */ | |
fb2d4483 | 812 | static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx, |
dc48e56d | 813 | struct ctx_rq_wait *wait) |
36f55889 | 814 | { |
fa88b6f8 | 815 | struct kioctx_table *table; |
db446a08 | 816 | |
b2edffdd AV |
817 | spin_lock(&mm->ioctx_lock); |
818 | if (atomic_xchg(&ctx->dead, 1)) { | |
819 | spin_unlock(&mm->ioctx_lock); | |
fa88b6f8 | 820 | return -EINVAL; |
b2edffdd | 821 | } |
db446a08 | 822 | |
855ef0de | 823 | table = rcu_dereference_raw(mm->ioctx_table); |
d0264c01 TH |
824 | WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id])); |
825 | RCU_INIT_POINTER(table->table[ctx->id], NULL); | |
fa88b6f8 | 826 | spin_unlock(&mm->ioctx_lock); |
4fcc712f | 827 | |
a6d7cff4 | 828 | /* free_ioctx_reqs() will do the necessary RCU synchronization */ |
fa88b6f8 | 829 | wake_up_all(&ctx->wait); |
4fcc712f | 830 | |
fa88b6f8 BL |
831 | /* |
832 | * It'd be more correct to do this in free_ioctx(), after all | |
833 | * the outstanding kiocbs have finished - but by then io_destroy | |
834 | * has already returned, so io_setup() could potentially return | |
835 | * -EAGAIN with no ioctxs actually in use (as far as userspace | |
836 | * could tell). | |
837 | */ | |
838 | aio_nr_sub(ctx->max_reqs); | |
4fcc712f | 839 | |
fa88b6f8 BL |
840 | if (ctx->mmap_size) |
841 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
fb2d4483 | 842 | |
dc48e56d | 843 | ctx->rq_wait = wait; |
fa88b6f8 BL |
844 | percpu_ref_kill(&ctx->users); |
845 | return 0; | |
1da177e4 LT |
846 | } |
847 | ||
36f55889 KO |
848 | /* |
849 | * exit_aio: called when the last user of mm goes away. At this point, there is | |
850 | * no way for any new requests to be submited or any of the io_* syscalls to be | |
851 | * called on the context. | |
852 | * | |
853 | * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on | |
854 | * them. | |
1da177e4 | 855 | */ |
fc9b52cd | 856 | void exit_aio(struct mm_struct *mm) |
1da177e4 | 857 | { |
4b70ac5f | 858 | struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table); |
dc48e56d JA |
859 | struct ctx_rq_wait wait; |
860 | int i, skipped; | |
db446a08 | 861 | |
4b70ac5f ON |
862 | if (!table) |
863 | return; | |
db446a08 | 864 | |
dc48e56d JA |
865 | atomic_set(&wait.count, table->nr); |
866 | init_completion(&wait.comp); | |
867 | ||
868 | skipped = 0; | |
4b70ac5f | 869 | for (i = 0; i < table->nr; ++i) { |
d0264c01 TH |
870 | struct kioctx *ctx = |
871 | rcu_dereference_protected(table->table[i], true); | |
abf137dd | 872 | |
dc48e56d JA |
873 | if (!ctx) { |
874 | skipped++; | |
4b70ac5f | 875 | continue; |
dc48e56d JA |
876 | } |
877 | ||
936af157 | 878 | /* |
4b70ac5f ON |
879 | * We don't need to bother with munmap() here - exit_mmap(mm) |
880 | * is coming and it'll unmap everything. And we simply can't, | |
881 | * this is not necessarily our ->mm. | |
882 | * Since kill_ioctx() uses non-zero ->mmap_size as indicator | |
883 | * that it needs to unmap the area, just set it to 0. | |
936af157 | 884 | */ |
58c85dc2 | 885 | ctx->mmap_size = 0; |
dc48e56d JA |
886 | kill_ioctx(mm, ctx, &wait); |
887 | } | |
36f55889 | 888 | |
dc48e56d | 889 | if (!atomic_sub_and_test(skipped, &wait.count)) { |
6098b45b | 890 | /* Wait until all IO for the context are done. */ |
dc48e56d | 891 | wait_for_completion(&wait.comp); |
1da177e4 | 892 | } |
4b70ac5f ON |
893 | |
894 | RCU_INIT_POINTER(mm->ioctx_table, NULL); | |
895 | kfree(table); | |
1da177e4 LT |
896 | } |
897 | ||
e1bdd5f2 KO |
898 | static void put_reqs_available(struct kioctx *ctx, unsigned nr) |
899 | { | |
900 | struct kioctx_cpu *kcpu; | |
263782c1 | 901 | unsigned long flags; |
e1bdd5f2 | 902 | |
263782c1 | 903 | local_irq_save(flags); |
be6fb451 | 904 | kcpu = this_cpu_ptr(ctx->cpu); |
e1bdd5f2 | 905 | kcpu->reqs_available += nr; |
263782c1 | 906 | |
e1bdd5f2 KO |
907 | while (kcpu->reqs_available >= ctx->req_batch * 2) { |
908 | kcpu->reqs_available -= ctx->req_batch; | |
909 | atomic_add(ctx->req_batch, &ctx->reqs_available); | |
910 | } | |
911 | ||
263782c1 | 912 | local_irq_restore(flags); |
e1bdd5f2 KO |
913 | } |
914 | ||
915 | static bool get_reqs_available(struct kioctx *ctx) | |
916 | { | |
917 | struct kioctx_cpu *kcpu; | |
918 | bool ret = false; | |
263782c1 | 919 | unsigned long flags; |
e1bdd5f2 | 920 | |
263782c1 | 921 | local_irq_save(flags); |
be6fb451 | 922 | kcpu = this_cpu_ptr(ctx->cpu); |
e1bdd5f2 KO |
923 | if (!kcpu->reqs_available) { |
924 | int old, avail = atomic_read(&ctx->reqs_available); | |
925 | ||
926 | do { | |
927 | if (avail < ctx->req_batch) | |
928 | goto out; | |
929 | ||
930 | old = avail; | |
931 | avail = atomic_cmpxchg(&ctx->reqs_available, | |
932 | avail, avail - ctx->req_batch); | |
933 | } while (avail != old); | |
934 | ||
935 | kcpu->reqs_available += ctx->req_batch; | |
936 | } | |
937 | ||
938 | ret = true; | |
939 | kcpu->reqs_available--; | |
940 | out: | |
263782c1 | 941 | local_irq_restore(flags); |
e1bdd5f2 KO |
942 | return ret; |
943 | } | |
944 | ||
d856f32a BL |
945 | /* refill_reqs_available |
946 | * Updates the reqs_available reference counts used for tracking the | |
947 | * number of free slots in the completion ring. This can be called | |
948 | * from aio_complete() (to optimistically update reqs_available) or | |
949 | * from aio_get_req() (the we're out of events case). It must be | |
950 | * called holding ctx->completion_lock. | |
951 | */ | |
952 | static void refill_reqs_available(struct kioctx *ctx, unsigned head, | |
953 | unsigned tail) | |
954 | { | |
955 | unsigned events_in_ring, completed; | |
956 | ||
957 | /* Clamp head since userland can write to it. */ | |
958 | head %= ctx->nr_events; | |
959 | if (head <= tail) | |
960 | events_in_ring = tail - head; | |
961 | else | |
962 | events_in_ring = ctx->nr_events - (head - tail); | |
963 | ||
964 | completed = ctx->completed_events; | |
965 | if (events_in_ring < completed) | |
966 | completed -= events_in_ring; | |
967 | else | |
968 | completed = 0; | |
969 | ||
970 | if (!completed) | |
971 | return; | |
972 | ||
973 | ctx->completed_events -= completed; | |
974 | put_reqs_available(ctx, completed); | |
975 | } | |
976 | ||
977 | /* user_refill_reqs_available | |
978 | * Called to refill reqs_available when aio_get_req() encounters an | |
979 | * out of space in the completion ring. | |
980 | */ | |
981 | static void user_refill_reqs_available(struct kioctx *ctx) | |
982 | { | |
983 | spin_lock_irq(&ctx->completion_lock); | |
984 | if (ctx->completed_events) { | |
985 | struct aio_ring *ring; | |
986 | unsigned head; | |
987 | ||
988 | /* Access of ring->head may race with aio_read_events_ring() | |
989 | * here, but that's okay since whether we read the old version | |
990 | * or the new version, and either will be valid. The important | |
991 | * part is that head cannot pass tail since we prevent | |
992 | * aio_complete() from updating tail by holding | |
993 | * ctx->completion_lock. Even if head is invalid, the check | |
994 | * against ctx->completed_events below will make sure we do the | |
995 | * safe/right thing. | |
996 | */ | |
997 | ring = kmap_atomic(ctx->ring_pages[0]); | |
998 | head = ring->head; | |
999 | kunmap_atomic(ring); | |
1000 | ||
1001 | refill_reqs_available(ctx, head, ctx->tail); | |
1002 | } | |
1003 | ||
1004 | spin_unlock_irq(&ctx->completion_lock); | |
1005 | } | |
1006 | ||
1da177e4 | 1007 | /* aio_get_req |
57282d8f KO |
1008 | * Allocate a slot for an aio request. |
1009 | * Returns NULL if no requests are free. | |
1da177e4 | 1010 | */ |
04b2fa9f | 1011 | static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx) |
1da177e4 | 1012 | { |
04b2fa9f | 1013 | struct aio_kiocb *req; |
a1c8eae7 | 1014 | |
d856f32a BL |
1015 | if (!get_reqs_available(ctx)) { |
1016 | user_refill_reqs_available(ctx); | |
1017 | if (!get_reqs_available(ctx)) | |
1018 | return NULL; | |
1019 | } | |
a1c8eae7 | 1020 | |
0460fef2 | 1021 | req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO); |
1da177e4 | 1022 | if (unlikely(!req)) |
a1c8eae7 | 1023 | goto out_put; |
1da177e4 | 1024 | |
e34ecee2 | 1025 | percpu_ref_get(&ctx->reqs); |
75321b50 | 1026 | INIT_LIST_HEAD(&req->ki_list); |
1da177e4 | 1027 | req->ki_ctx = ctx; |
080d676d | 1028 | return req; |
a1c8eae7 | 1029 | out_put: |
e1bdd5f2 | 1030 | put_reqs_available(ctx, 1); |
a1c8eae7 | 1031 | return NULL; |
1da177e4 LT |
1032 | } |
1033 | ||
d5470b59 | 1034 | static struct kioctx *lookup_ioctx(unsigned long ctx_id) |
1da177e4 | 1035 | { |
db446a08 | 1036 | struct aio_ring __user *ring = (void __user *)ctx_id; |
abf137dd | 1037 | struct mm_struct *mm = current->mm; |
65c24491 | 1038 | struct kioctx *ctx, *ret = NULL; |
db446a08 BL |
1039 | struct kioctx_table *table; |
1040 | unsigned id; | |
1041 | ||
1042 | if (get_user(id, &ring->id)) | |
1043 | return NULL; | |
1da177e4 | 1044 | |
abf137dd | 1045 | rcu_read_lock(); |
db446a08 | 1046 | table = rcu_dereference(mm->ioctx_table); |
abf137dd | 1047 | |
db446a08 BL |
1048 | if (!table || id >= table->nr) |
1049 | goto out; | |
1da177e4 | 1050 | |
d0264c01 | 1051 | ctx = rcu_dereference(table->table[id]); |
f30d704f | 1052 | if (ctx && ctx->user_id == ctx_id) { |
baf10564 AV |
1053 | if (percpu_ref_tryget_live(&ctx->users)) |
1054 | ret = ctx; | |
db446a08 BL |
1055 | } |
1056 | out: | |
abf137dd | 1057 | rcu_read_unlock(); |
65c24491 | 1058 | return ret; |
1da177e4 LT |
1059 | } |
1060 | ||
1da177e4 LT |
1061 | /* aio_complete |
1062 | * Called when the io request on the given iocb is complete. | |
1da177e4 | 1063 | */ |
54843f87 | 1064 | static void aio_complete(struct aio_kiocb *iocb, long res, long res2) |
1da177e4 LT |
1065 | { |
1066 | struct kioctx *ctx = iocb->ki_ctx; | |
1da177e4 | 1067 | struct aio_ring *ring; |
21b40200 | 1068 | struct io_event *ev_page, *event; |
d856f32a | 1069 | unsigned tail, pos, head; |
1da177e4 | 1070 | unsigned long flags; |
1da177e4 | 1071 | |
0460fef2 KO |
1072 | /* |
1073 | * Add a completion event to the ring buffer. Must be done holding | |
4b30f07e | 1074 | * ctx->completion_lock to prevent other code from messing with the tail |
0460fef2 KO |
1075 | * pointer since we might be called from irq context. |
1076 | */ | |
1077 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1078 | ||
58c85dc2 | 1079 | tail = ctx->tail; |
21b40200 KO |
1080 | pos = tail + AIO_EVENTS_OFFSET; |
1081 | ||
58c85dc2 | 1082 | if (++tail >= ctx->nr_events) |
4bf69b2a | 1083 | tail = 0; |
1da177e4 | 1084 | |
58c85dc2 | 1085 | ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
1086 | event = ev_page + pos % AIO_EVENTS_PER_PAGE; |
1087 | ||
04b2fa9f | 1088 | event->obj = (u64)(unsigned long)iocb->ki_user_iocb; |
1da177e4 LT |
1089 | event->data = iocb->ki_user_data; |
1090 | event->res = res; | |
1091 | event->res2 = res2; | |
1092 | ||
21b40200 | 1093 | kunmap_atomic(ev_page); |
58c85dc2 | 1094 | flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
1095 | |
1096 | pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n", | |
04b2fa9f | 1097 | ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data, |
caf4167a | 1098 | res, res2); |
1da177e4 LT |
1099 | |
1100 | /* after flagging the request as done, we | |
1101 | * must never even look at it again | |
1102 | */ | |
1103 | smp_wmb(); /* make event visible before updating tail */ | |
1104 | ||
58c85dc2 | 1105 | ctx->tail = tail; |
1da177e4 | 1106 | |
58c85dc2 | 1107 | ring = kmap_atomic(ctx->ring_pages[0]); |
d856f32a | 1108 | head = ring->head; |
21b40200 | 1109 | ring->tail = tail; |
e8e3c3d6 | 1110 | kunmap_atomic(ring); |
58c85dc2 | 1111 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 | 1112 | |
d856f32a BL |
1113 | ctx->completed_events++; |
1114 | if (ctx->completed_events > 1) | |
1115 | refill_reqs_available(ctx, head, tail); | |
0460fef2 KO |
1116 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1117 | ||
21b40200 | 1118 | pr_debug("added to ring %p at [%u]\n", iocb, tail); |
8d1c98b0 DL |
1119 | |
1120 | /* | |
1121 | * Check if the user asked us to deliver the result through an | |
1122 | * eventfd. The eventfd_signal() function is safe to be called | |
1123 | * from IRQ context. | |
1124 | */ | |
54843f87 | 1125 | if (iocb->ki_eventfd) { |
8d1c98b0 | 1126 | eventfd_signal(iocb->ki_eventfd, 1); |
54843f87 CH |
1127 | eventfd_ctx_put(iocb->ki_eventfd); |
1128 | } | |
8d1c98b0 | 1129 | |
54843f87 | 1130 | kmem_cache_free(kiocb_cachep, iocb); |
1da177e4 | 1131 | |
6cb2a210 QB |
1132 | /* |
1133 | * We have to order our ring_info tail store above and test | |
1134 | * of the wait list below outside the wait lock. This is | |
1135 | * like in wake_up_bit() where clearing a bit has to be | |
1136 | * ordered with the unlocked test. | |
1137 | */ | |
1138 | smp_mb(); | |
1139 | ||
1da177e4 LT |
1140 | if (waitqueue_active(&ctx->wait)) |
1141 | wake_up(&ctx->wait); | |
1142 | ||
e34ecee2 | 1143 | percpu_ref_put(&ctx->reqs); |
1da177e4 LT |
1144 | } |
1145 | ||
2be4e7de | 1146 | /* aio_read_events_ring |
a31ad380 KO |
1147 | * Pull an event off of the ioctx's event ring. Returns the number of |
1148 | * events fetched | |
1da177e4 | 1149 | */ |
a31ad380 KO |
1150 | static long aio_read_events_ring(struct kioctx *ctx, |
1151 | struct io_event __user *event, long nr) | |
1da177e4 | 1152 | { |
1da177e4 | 1153 | struct aio_ring *ring; |
5ffac122 | 1154 | unsigned head, tail, pos; |
a31ad380 KO |
1155 | long ret = 0; |
1156 | int copy_ret; | |
1157 | ||
9c9ce763 DC |
1158 | /* |
1159 | * The mutex can block and wake us up and that will cause | |
1160 | * wait_event_interruptible_hrtimeout() to schedule without sleeping | |
1161 | * and repeat. This should be rare enough that it doesn't cause | |
1162 | * peformance issues. See the comment in read_events() for more detail. | |
1163 | */ | |
1164 | sched_annotate_sleep(); | |
58c85dc2 | 1165 | mutex_lock(&ctx->ring_lock); |
1da177e4 | 1166 | |
fa8a53c3 | 1167 | /* Access to ->ring_pages here is protected by ctx->ring_lock. */ |
58c85dc2 | 1168 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 1169 | head = ring->head; |
5ffac122 | 1170 | tail = ring->tail; |
a31ad380 KO |
1171 | kunmap_atomic(ring); |
1172 | ||
2ff396be JM |
1173 | /* |
1174 | * Ensure that once we've read the current tail pointer, that | |
1175 | * we also see the events that were stored up to the tail. | |
1176 | */ | |
1177 | smp_rmb(); | |
1178 | ||
5ffac122 | 1179 | pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events); |
1da177e4 | 1180 | |
5ffac122 | 1181 | if (head == tail) |
1da177e4 LT |
1182 | goto out; |
1183 | ||
edfbbf38 BL |
1184 | head %= ctx->nr_events; |
1185 | tail %= ctx->nr_events; | |
1186 | ||
a31ad380 KO |
1187 | while (ret < nr) { |
1188 | long avail; | |
1189 | struct io_event *ev; | |
1190 | struct page *page; | |
1191 | ||
5ffac122 KO |
1192 | avail = (head <= tail ? tail : ctx->nr_events) - head; |
1193 | if (head == tail) | |
a31ad380 KO |
1194 | break; |
1195 | ||
a31ad380 | 1196 | pos = head + AIO_EVENTS_OFFSET; |
58c85dc2 | 1197 | page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]; |
a31ad380 KO |
1198 | pos %= AIO_EVENTS_PER_PAGE; |
1199 | ||
d2988bd4 AV |
1200 | avail = min(avail, nr - ret); |
1201 | avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos); | |
1202 | ||
a31ad380 KO |
1203 | ev = kmap(page); |
1204 | copy_ret = copy_to_user(event + ret, ev + pos, | |
1205 | sizeof(*ev) * avail); | |
1206 | kunmap(page); | |
1207 | ||
1208 | if (unlikely(copy_ret)) { | |
1209 | ret = -EFAULT; | |
1210 | goto out; | |
1211 | } | |
1212 | ||
1213 | ret += avail; | |
1214 | head += avail; | |
58c85dc2 | 1215 | head %= ctx->nr_events; |
1da177e4 | 1216 | } |
1da177e4 | 1217 | |
58c85dc2 | 1218 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 1219 | ring->head = head; |
91d80a84 | 1220 | kunmap_atomic(ring); |
58c85dc2 | 1221 | flush_dcache_page(ctx->ring_pages[0]); |
a31ad380 | 1222 | |
5ffac122 | 1223 | pr_debug("%li h%u t%u\n", ret, head, tail); |
a31ad380 | 1224 | out: |
58c85dc2 | 1225 | mutex_unlock(&ctx->ring_lock); |
a31ad380 | 1226 | |
1da177e4 LT |
1227 | return ret; |
1228 | } | |
1229 | ||
a31ad380 KO |
1230 | static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr, |
1231 | struct io_event __user *event, long *i) | |
1da177e4 | 1232 | { |
a31ad380 | 1233 | long ret = aio_read_events_ring(ctx, event + *i, nr - *i); |
1da177e4 | 1234 | |
a31ad380 KO |
1235 | if (ret > 0) |
1236 | *i += ret; | |
1da177e4 | 1237 | |
a31ad380 KO |
1238 | if (unlikely(atomic_read(&ctx->dead))) |
1239 | ret = -EINVAL; | |
1da177e4 | 1240 | |
a31ad380 KO |
1241 | if (!*i) |
1242 | *i = ret; | |
1da177e4 | 1243 | |
a31ad380 | 1244 | return ret < 0 || *i >= min_nr; |
1da177e4 LT |
1245 | } |
1246 | ||
a31ad380 | 1247 | static long read_events(struct kioctx *ctx, long min_nr, long nr, |
1da177e4 | 1248 | struct io_event __user *event, |
fa2e62a5 | 1249 | ktime_t until) |
1da177e4 | 1250 | { |
a31ad380 | 1251 | long ret = 0; |
1da177e4 | 1252 | |
a31ad380 KO |
1253 | /* |
1254 | * Note that aio_read_events() is being called as the conditional - i.e. | |
1255 | * we're calling it after prepare_to_wait() has set task state to | |
1256 | * TASK_INTERRUPTIBLE. | |
1257 | * | |
1258 | * But aio_read_events() can block, and if it blocks it's going to flip | |
1259 | * the task state back to TASK_RUNNING. | |
1260 | * | |
1261 | * This should be ok, provided it doesn't flip the state back to | |
1262 | * TASK_RUNNING and return 0 too much - that causes us to spin. That | |
1263 | * will only happen if the mutex_lock() call blocks, and we then find | |
1264 | * the ringbuffer empty. So in practice we should be ok, but it's | |
1265 | * something to be aware of when touching this code. | |
1266 | */ | |
2456e855 | 1267 | if (until == 0) |
5f785de5 FZ |
1268 | aio_read_events(ctx, min_nr, nr, event, &ret); |
1269 | else | |
1270 | wait_event_interruptible_hrtimeout(ctx->wait, | |
1271 | aio_read_events(ctx, min_nr, nr, event, &ret), | |
1272 | until); | |
a31ad380 | 1273 | return ret; |
1da177e4 LT |
1274 | } |
1275 | ||
1da177e4 LT |
1276 | /* sys_io_setup: |
1277 | * Create an aio_context capable of receiving at least nr_events. | |
1278 | * ctxp must not point to an aio_context that already exists, and | |
1279 | * must be initialized to 0 prior to the call. On successful | |
1280 | * creation of the aio_context, *ctxp is filled in with the resulting | |
1281 | * handle. May fail with -EINVAL if *ctxp is not initialized, | |
1282 | * if the specified nr_events exceeds internal limits. May fail | |
1283 | * with -EAGAIN if the specified nr_events exceeds the user's limit | |
1284 | * of available events. May fail with -ENOMEM if insufficient kernel | |
1285 | * resources are available. May fail with -EFAULT if an invalid | |
1286 | * pointer is passed for ctxp. Will fail with -ENOSYS if not | |
1287 | * implemented. | |
1288 | */ | |
002c8976 | 1289 | SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) |
1da177e4 LT |
1290 | { |
1291 | struct kioctx *ioctx = NULL; | |
1292 | unsigned long ctx; | |
1293 | long ret; | |
1294 | ||
1295 | ret = get_user(ctx, ctxp); | |
1296 | if (unlikely(ret)) | |
1297 | goto out; | |
1298 | ||
1299 | ret = -EINVAL; | |
d55b5fda | 1300 | if (unlikely(ctx || nr_events == 0)) { |
acd88d4e | 1301 | pr_debug("EINVAL: ctx %lu nr_events %u\n", |
d55b5fda | 1302 | ctx, nr_events); |
1da177e4 LT |
1303 | goto out; |
1304 | } | |
1305 | ||
1306 | ioctx = ioctx_alloc(nr_events); | |
1307 | ret = PTR_ERR(ioctx); | |
1308 | if (!IS_ERR(ioctx)) { | |
1309 | ret = put_user(ioctx->user_id, ctxp); | |
a2e1859a | 1310 | if (ret) |
e02ba72a | 1311 | kill_ioctx(current->mm, ioctx, NULL); |
723be6e3 | 1312 | percpu_ref_put(&ioctx->users); |
1da177e4 LT |
1313 | } |
1314 | ||
1315 | out: | |
1316 | return ret; | |
1317 | } | |
1318 | ||
c00d2c7e AV |
1319 | #ifdef CONFIG_COMPAT |
1320 | COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p) | |
1321 | { | |
1322 | struct kioctx *ioctx = NULL; | |
1323 | unsigned long ctx; | |
1324 | long ret; | |
1325 | ||
1326 | ret = get_user(ctx, ctx32p); | |
1327 | if (unlikely(ret)) | |
1328 | goto out; | |
1329 | ||
1330 | ret = -EINVAL; | |
1331 | if (unlikely(ctx || nr_events == 0)) { | |
1332 | pr_debug("EINVAL: ctx %lu nr_events %u\n", | |
1333 | ctx, nr_events); | |
1334 | goto out; | |
1335 | } | |
1336 | ||
1337 | ioctx = ioctx_alloc(nr_events); | |
1338 | ret = PTR_ERR(ioctx); | |
1339 | if (!IS_ERR(ioctx)) { | |
1340 | /* truncating is ok because it's a user address */ | |
1341 | ret = put_user((u32)ioctx->user_id, ctx32p); | |
1342 | if (ret) | |
1343 | kill_ioctx(current->mm, ioctx, NULL); | |
1344 | percpu_ref_put(&ioctx->users); | |
1345 | } | |
1346 | ||
1347 | out: | |
1348 | return ret; | |
1349 | } | |
1350 | #endif | |
1351 | ||
1da177e4 LT |
1352 | /* sys_io_destroy: |
1353 | * Destroy the aio_context specified. May cancel any outstanding | |
1354 | * AIOs and block on completion. Will fail with -ENOSYS if not | |
642b5123 | 1355 | * implemented. May fail with -EINVAL if the context pointed to |
1da177e4 LT |
1356 | * is invalid. |
1357 | */ | |
002c8976 | 1358 | SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) |
1da177e4 LT |
1359 | { |
1360 | struct kioctx *ioctx = lookup_ioctx(ctx); | |
1361 | if (likely(NULL != ioctx)) { | |
dc48e56d | 1362 | struct ctx_rq_wait wait; |
fb2d4483 | 1363 | int ret; |
e02ba72a | 1364 | |
dc48e56d JA |
1365 | init_completion(&wait.comp); |
1366 | atomic_set(&wait.count, 1); | |
1367 | ||
e02ba72a AP |
1368 | /* Pass requests_done to kill_ioctx() where it can be set |
1369 | * in a thread-safe way. If we try to set it here then we have | |
1370 | * a race condition if two io_destroy() called simultaneously. | |
1371 | */ | |
dc48e56d | 1372 | ret = kill_ioctx(current->mm, ioctx, &wait); |
723be6e3 | 1373 | percpu_ref_put(&ioctx->users); |
e02ba72a AP |
1374 | |
1375 | /* Wait until all IO for the context are done. Otherwise kernel | |
1376 | * keep using user-space buffers even if user thinks the context | |
1377 | * is destroyed. | |
1378 | */ | |
fb2d4483 | 1379 | if (!ret) |
dc48e56d | 1380 | wait_for_completion(&wait.comp); |
e02ba72a | 1381 | |
fb2d4483 | 1382 | return ret; |
1da177e4 | 1383 | } |
acd88d4e | 1384 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1385 | return -EINVAL; |
1386 | } | |
1387 | ||
3c96c7f4 AV |
1388 | static void aio_remove_iocb(struct aio_kiocb *iocb) |
1389 | { | |
1390 | struct kioctx *ctx = iocb->ki_ctx; | |
1391 | unsigned long flags; | |
1392 | ||
1393 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
1394 | list_del(&iocb->ki_list); | |
1395 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
1396 | } | |
1397 | ||
54843f87 CH |
1398 | static void aio_complete_rw(struct kiocb *kiocb, long res, long res2) |
1399 | { | |
1400 | struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw); | |
1401 | ||
3c96c7f4 AV |
1402 | if (!list_empty_careful(&iocb->ki_list)) |
1403 | aio_remove_iocb(iocb); | |
1404 | ||
54843f87 CH |
1405 | if (kiocb->ki_flags & IOCB_WRITE) { |
1406 | struct inode *inode = file_inode(kiocb->ki_filp); | |
1407 | ||
1408 | /* | |
1409 | * Tell lockdep we inherited freeze protection from submission | |
1410 | * thread. | |
1411 | */ | |
1412 | if (S_ISREG(inode->i_mode)) | |
1413 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); | |
1414 | file_end_write(kiocb->ki_filp); | |
1415 | } | |
1416 | ||
1417 | fput(kiocb->ki_filp); | |
1418 | aio_complete(iocb, res, res2); | |
1419 | } | |
1420 | ||
1421 | static int aio_prep_rw(struct kiocb *req, struct iocb *iocb) | |
1422 | { | |
1423 | int ret; | |
1424 | ||
1425 | req->ki_filp = fget(iocb->aio_fildes); | |
1426 | if (unlikely(!req->ki_filp)) | |
1427 | return -EBADF; | |
1428 | req->ki_complete = aio_complete_rw; | |
1429 | req->ki_pos = iocb->aio_offset; | |
1430 | req->ki_flags = iocb_flags(req->ki_filp); | |
1431 | if (iocb->aio_flags & IOCB_FLAG_RESFD) | |
1432 | req->ki_flags |= IOCB_EVENTFD; | |
fc28724d | 1433 | req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp)); |
d9a08a9e AM |
1434 | if (iocb->aio_flags & IOCB_FLAG_IOPRIO) { |
1435 | /* | |
1436 | * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then | |
1437 | * aio_reqprio is interpreted as an I/O scheduling | |
1438 | * class and priority. | |
1439 | */ | |
1440 | ret = ioprio_check_cap(iocb->aio_reqprio); | |
1441 | if (ret) { | |
9a6d9a62 AM |
1442 | pr_debug("aio ioprio check cap error: %d\n", ret); |
1443 | return ret; | |
d9a08a9e AM |
1444 | } |
1445 | ||
1446 | req->ki_ioprio = iocb->aio_reqprio; | |
1447 | } else | |
1448 | req->ki_ioprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_NONE, 0); | |
1449 | ||
54843f87 CH |
1450 | ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags); |
1451 | if (unlikely(ret)) | |
1452 | fput(req->ki_filp); | |
1453 | return ret; | |
1454 | } | |
1455 | ||
89319d31 CH |
1456 | static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec, |
1457 | bool vectored, bool compat, struct iov_iter *iter) | |
eed4e51f | 1458 | { |
89319d31 CH |
1459 | void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf; |
1460 | size_t len = iocb->aio_nbytes; | |
1461 | ||
1462 | if (!vectored) { | |
1463 | ssize_t ret = import_single_range(rw, buf, len, *iovec, iter); | |
1464 | *iovec = NULL; | |
1465 | return ret; | |
1466 | } | |
9d85cba7 JM |
1467 | #ifdef CONFIG_COMPAT |
1468 | if (compat) | |
89319d31 CH |
1469 | return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec, |
1470 | iter); | |
9d85cba7 | 1471 | #endif |
89319d31 | 1472 | return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter); |
eed4e51f BP |
1473 | } |
1474 | ||
9061d14a | 1475 | static inline void aio_rw_done(struct kiocb *req, ssize_t ret) |
89319d31 CH |
1476 | { |
1477 | switch (ret) { | |
1478 | case -EIOCBQUEUED: | |
9061d14a | 1479 | break; |
89319d31 CH |
1480 | case -ERESTARTSYS: |
1481 | case -ERESTARTNOINTR: | |
1482 | case -ERESTARTNOHAND: | |
1483 | case -ERESTART_RESTARTBLOCK: | |
1484 | /* | |
1485 | * There's no easy way to restart the syscall since other AIO's | |
1486 | * may be already running. Just fail this IO with EINTR. | |
1487 | */ | |
1488 | ret = -EINTR; | |
1489 | /*FALLTHRU*/ | |
1490 | default: | |
54843f87 | 1491 | aio_complete_rw(req, ret, 0); |
89319d31 CH |
1492 | } |
1493 | } | |
1494 | ||
1495 | static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored, | |
1496 | bool compat) | |
1da177e4 | 1497 | { |
00fefb9c | 1498 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; |
293bc982 | 1499 | struct iov_iter iter; |
54843f87 | 1500 | struct file *file; |
89319d31 | 1501 | ssize_t ret; |
1da177e4 | 1502 | |
54843f87 CH |
1503 | ret = aio_prep_rw(req, iocb); |
1504 | if (ret) | |
1505 | return ret; | |
1506 | file = req->ki_filp; | |
1507 | ||
1508 | ret = -EBADF; | |
89319d31 | 1509 | if (unlikely(!(file->f_mode & FMODE_READ))) |
54843f87 CH |
1510 | goto out_fput; |
1511 | ret = -EINVAL; | |
89319d31 | 1512 | if (unlikely(!file->f_op->read_iter)) |
54843f87 | 1513 | goto out_fput; |
73a7075e | 1514 | |
89319d31 CH |
1515 | ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter); |
1516 | if (ret) | |
54843f87 | 1517 | goto out_fput; |
89319d31 CH |
1518 | ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter)); |
1519 | if (!ret) | |
9061d14a | 1520 | aio_rw_done(req, call_read_iter(file, req, &iter)); |
89319d31 | 1521 | kfree(iovec); |
54843f87 | 1522 | out_fput: |
9061d14a | 1523 | if (unlikely(ret)) |
54843f87 | 1524 | fput(file); |
89319d31 CH |
1525 | return ret; |
1526 | } | |
73a7075e | 1527 | |
89319d31 CH |
1528 | static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored, |
1529 | bool compat) | |
1530 | { | |
89319d31 CH |
1531 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; |
1532 | struct iov_iter iter; | |
54843f87 | 1533 | struct file *file; |
89319d31 | 1534 | ssize_t ret; |
41ef4eb8 | 1535 | |
54843f87 CH |
1536 | ret = aio_prep_rw(req, iocb); |
1537 | if (ret) | |
1538 | return ret; | |
1539 | file = req->ki_filp; | |
1540 | ||
1541 | ret = -EBADF; | |
89319d31 | 1542 | if (unlikely(!(file->f_mode & FMODE_WRITE))) |
54843f87 CH |
1543 | goto out_fput; |
1544 | ret = -EINVAL; | |
89319d31 | 1545 | if (unlikely(!file->f_op->write_iter)) |
54843f87 | 1546 | goto out_fput; |
1da177e4 | 1547 | |
89319d31 CH |
1548 | ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter); |
1549 | if (ret) | |
54843f87 | 1550 | goto out_fput; |
89319d31 CH |
1551 | ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter)); |
1552 | if (!ret) { | |
70fe2f48 | 1553 | /* |
92ce4728 | 1554 | * Open-code file_start_write here to grab freeze protection, |
54843f87 CH |
1555 | * which will be released by another thread in |
1556 | * aio_complete_rw(). Fool lockdep by telling it the lock got | |
1557 | * released so that it doesn't complain about the held lock when | |
1558 | * we return to userspace. | |
70fe2f48 | 1559 | */ |
92ce4728 CH |
1560 | if (S_ISREG(file_inode(file)->i_mode)) { |
1561 | __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true); | |
a12f1ae6 | 1562 | __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE); |
92ce4728 CH |
1563 | } |
1564 | req->ki_flags |= IOCB_WRITE; | |
9061d14a | 1565 | aio_rw_done(req, call_write_iter(file, req, &iter)); |
41ef4eb8 | 1566 | } |
89319d31 | 1567 | kfree(iovec); |
54843f87 | 1568 | out_fput: |
9061d14a | 1569 | if (unlikely(ret)) |
54843f87 | 1570 | fput(file); |
89319d31 | 1571 | return ret; |
1da177e4 LT |
1572 | } |
1573 | ||
a3c0d439 CH |
1574 | static void aio_fsync_work(struct work_struct *work) |
1575 | { | |
1576 | struct fsync_iocb *req = container_of(work, struct fsync_iocb, work); | |
1577 | int ret; | |
1578 | ||
1579 | ret = vfs_fsync(req->file, req->datasync); | |
1580 | fput(req->file); | |
1581 | aio_complete(container_of(req, struct aio_kiocb, fsync), ret, 0); | |
1582 | } | |
1583 | ||
1584 | static int aio_fsync(struct fsync_iocb *req, struct iocb *iocb, bool datasync) | |
1585 | { | |
1586 | if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes || | |
1587 | iocb->aio_rw_flags)) | |
1588 | return -EINVAL; | |
a3c0d439 CH |
1589 | req->file = fget(iocb->aio_fildes); |
1590 | if (unlikely(!req->file)) | |
1591 | return -EBADF; | |
1592 | if (unlikely(!req->file->f_op->fsync)) { | |
1593 | fput(req->file); | |
1594 | return -EINVAL; | |
1595 | } | |
1596 | ||
1597 | req->datasync = datasync; | |
1598 | INIT_WORK(&req->work, aio_fsync_work); | |
1599 | schedule_work(&req->work); | |
9061d14a | 1600 | return 0; |
a3c0d439 CH |
1601 | } |
1602 | ||
2c14fa83 CH |
1603 | /* need to use list_del_init so we can check if item was present */ |
1604 | static inline bool __aio_poll_remove(struct poll_iocb *req) | |
1605 | { | |
1606 | if (list_empty(&req->wait.entry)) | |
1607 | return false; | |
1608 | list_del_init(&req->wait.entry); | |
1609 | return true; | |
1610 | } | |
1611 | ||
3c96c7f4 | 1612 | static inline void __aio_poll_complete(struct aio_kiocb *iocb, __poll_t mask) |
2c14fa83 | 1613 | { |
3c96c7f4 | 1614 | fput(iocb->poll.file); |
2c14fa83 | 1615 | aio_complete(iocb, mangle_poll(mask), 0); |
2c14fa83 CH |
1616 | } |
1617 | ||
1618 | static void aio_poll_work(struct work_struct *work) | |
1619 | { | |
3c96c7f4 | 1620 | struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, poll.work); |
2c14fa83 | 1621 | |
3c96c7f4 AV |
1622 | if (!list_empty_careful(&iocb->ki_list)) |
1623 | aio_remove_iocb(iocb); | |
1624 | __aio_poll_complete(iocb, iocb->poll.events); | |
2c14fa83 CH |
1625 | } |
1626 | ||
1627 | static int aio_poll_cancel(struct kiocb *iocb) | |
1628 | { | |
1629 | struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw); | |
1630 | struct poll_iocb *req = &aiocb->poll; | |
1631 | struct wait_queue_head *head = req->head; | |
1632 | bool found = false; | |
1633 | ||
1634 | spin_lock(&head->lock); | |
1635 | found = __aio_poll_remove(req); | |
1636 | spin_unlock(&head->lock); | |
1637 | ||
1638 | if (found) { | |
1639 | req->events = 0; | |
1640 | INIT_WORK(&req->work, aio_poll_work); | |
1641 | schedule_work(&req->work); | |
1642 | } | |
1643 | return 0; | |
1644 | } | |
1645 | ||
1646 | static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
1647 | void *key) | |
1648 | { | |
1649 | struct poll_iocb *req = container_of(wait, struct poll_iocb, wait); | |
1962da0d | 1650 | struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll); |
2c14fa83 CH |
1651 | struct file *file = req->file; |
1652 | __poll_t mask = key_to_poll(key); | |
1653 | ||
1654 | assert_spin_locked(&req->head->lock); | |
1655 | ||
1656 | /* for instances that support it check for an event match first: */ | |
1657 | if (mask && !(mask & req->events)) | |
1658 | return 0; | |
1659 | ||
2739b807 | 1660 | mask = file->f_op->poll_mask(file, req->events) & req->events; |
2c14fa83 CH |
1661 | if (!mask) |
1662 | return 0; | |
1663 | ||
1664 | __aio_poll_remove(req); | |
1665 | ||
1962da0d CH |
1666 | /* |
1667 | * Try completing without a context switch if we can acquire ctx_lock | |
1668 | * without spinning. Otherwise we need to defer to a workqueue to | |
1669 | * avoid a deadlock due to the lock order. | |
1670 | */ | |
1671 | if (spin_trylock(&iocb->ki_ctx->ctx_lock)) { | |
1672 | list_del_init(&iocb->ki_list); | |
1673 | spin_unlock(&iocb->ki_ctx->ctx_lock); | |
1674 | ||
3c96c7f4 | 1675 | __aio_poll_complete(iocb, mask); |
1962da0d CH |
1676 | } else { |
1677 | req->events = mask; | |
1678 | INIT_WORK(&req->work, aio_poll_work); | |
1679 | schedule_work(&req->work); | |
1680 | } | |
1681 | ||
2c14fa83 CH |
1682 | return 1; |
1683 | } | |
1684 | ||
1685 | static ssize_t aio_poll(struct aio_kiocb *aiocb, struct iocb *iocb) | |
1686 | { | |
1687 | struct kioctx *ctx = aiocb->ki_ctx; | |
1688 | struct poll_iocb *req = &aiocb->poll; | |
1689 | __poll_t mask; | |
1690 | ||
1691 | /* reject any unknown events outside the normal event mask. */ | |
1692 | if ((u16)iocb->aio_buf != iocb->aio_buf) | |
1693 | return -EINVAL; | |
1694 | /* reject fields that are not defined for poll */ | |
1695 | if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags) | |
1696 | return -EINVAL; | |
1697 | ||
1698 | req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP; | |
1699 | req->file = fget(iocb->aio_fildes); | |
1700 | if (unlikely(!req->file)) | |
1701 | return -EBADF; | |
1702 | if (!file_has_poll_mask(req->file)) | |
1703 | goto out_fail; | |
1704 | ||
1705 | req->head = req->file->f_op->get_poll_head(req->file, req->events); | |
1706 | if (!req->head) | |
1707 | goto out_fail; | |
1708 | if (IS_ERR(req->head)) { | |
1709 | mask = EPOLLERR; | |
1710 | goto done; | |
1711 | } | |
1712 | ||
1713 | init_waitqueue_func_entry(&req->wait, aio_poll_wake); | |
1714 | aiocb->ki_cancel = aio_poll_cancel; | |
1715 | ||
1716 | spin_lock_irq(&ctx->ctx_lock); | |
1717 | spin_lock(&req->head->lock); | |
2739b807 | 1718 | mask = req->file->f_op->poll_mask(req->file, req->events) & req->events; |
2c14fa83 CH |
1719 | if (!mask) { |
1720 | __add_wait_queue(req->head, &req->wait); | |
1721 | list_add_tail(&aiocb->ki_list, &ctx->active_reqs); | |
1722 | } | |
1723 | spin_unlock(&req->head->lock); | |
1724 | spin_unlock_irq(&ctx->ctx_lock); | |
1725 | done: | |
1726 | if (mask) | |
3c96c7f4 | 1727 | __aio_poll_complete(aiocb, mask); |
9061d14a | 1728 | return 0; |
2c14fa83 CH |
1729 | out_fail: |
1730 | fput(req->file); | |
1731 | return -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
1732 | } | |
1733 | ||
d5470b59 | 1734 | static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, |
95af8496 | 1735 | bool compat) |
1da177e4 | 1736 | { |
04b2fa9f | 1737 | struct aio_kiocb *req; |
95af8496 | 1738 | struct iocb iocb; |
1da177e4 LT |
1739 | ssize_t ret; |
1740 | ||
95af8496 AV |
1741 | if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb)))) |
1742 | return -EFAULT; | |
1743 | ||
1da177e4 | 1744 | /* enforce forwards compatibility on users */ |
95af8496 | 1745 | if (unlikely(iocb.aio_reserved2)) { |
caf4167a | 1746 | pr_debug("EINVAL: reserve field set\n"); |
1da177e4 LT |
1747 | return -EINVAL; |
1748 | } | |
1749 | ||
1750 | /* prevent overflows */ | |
1751 | if (unlikely( | |
95af8496 AV |
1752 | (iocb.aio_buf != (unsigned long)iocb.aio_buf) || |
1753 | (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) || | |
1754 | ((ssize_t)iocb.aio_nbytes < 0) | |
1da177e4 | 1755 | )) { |
acd88d4e | 1756 | pr_debug("EINVAL: overflow check\n"); |
1da177e4 LT |
1757 | return -EINVAL; |
1758 | } | |
1759 | ||
41ef4eb8 | 1760 | req = aio_get_req(ctx); |
1d98ebfc | 1761 | if (unlikely(!req)) |
1da177e4 | 1762 | return -EAGAIN; |
1d98ebfc | 1763 | |
95af8496 | 1764 | if (iocb.aio_flags & IOCB_FLAG_RESFD) { |
9c3060be DL |
1765 | /* |
1766 | * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an | |
1767 | * instance of the file* now. The file descriptor must be | |
1768 | * an eventfd() fd, and will be signaled for each completed | |
1769 | * event using the eventfd_signal() function. | |
1770 | */ | |
95af8496 | 1771 | req->ki_eventfd = eventfd_ctx_fdget((int) iocb.aio_resfd); |
801678c5 | 1772 | if (IS_ERR(req->ki_eventfd)) { |
9c3060be | 1773 | ret = PTR_ERR(req->ki_eventfd); |
87c3a86e | 1774 | req->ki_eventfd = NULL; |
9c3060be DL |
1775 | goto out_put_req; |
1776 | } | |
9830f4be GR |
1777 | } |
1778 | ||
8a660890 | 1779 | ret = put_user(KIOCB_KEY, &user_iocb->aio_key); |
1da177e4 | 1780 | if (unlikely(ret)) { |
caf4167a | 1781 | pr_debug("EFAULT: aio_key\n"); |
1da177e4 LT |
1782 | goto out_put_req; |
1783 | } | |
1784 | ||
04b2fa9f | 1785 | req->ki_user_iocb = user_iocb; |
95af8496 | 1786 | req->ki_user_data = iocb.aio_data; |
1da177e4 | 1787 | |
95af8496 | 1788 | switch (iocb.aio_lio_opcode) { |
89319d31 | 1789 | case IOCB_CMD_PREAD: |
95af8496 | 1790 | ret = aio_read(&req->rw, &iocb, false, compat); |
89319d31 CH |
1791 | break; |
1792 | case IOCB_CMD_PWRITE: | |
95af8496 | 1793 | ret = aio_write(&req->rw, &iocb, false, compat); |
89319d31 CH |
1794 | break; |
1795 | case IOCB_CMD_PREADV: | |
95af8496 | 1796 | ret = aio_read(&req->rw, &iocb, true, compat); |
89319d31 CH |
1797 | break; |
1798 | case IOCB_CMD_PWRITEV: | |
95af8496 | 1799 | ret = aio_write(&req->rw, &iocb, true, compat); |
89319d31 | 1800 | break; |
a3c0d439 | 1801 | case IOCB_CMD_FSYNC: |
95af8496 | 1802 | ret = aio_fsync(&req->fsync, &iocb, false); |
a3c0d439 CH |
1803 | break; |
1804 | case IOCB_CMD_FDSYNC: | |
95af8496 | 1805 | ret = aio_fsync(&req->fsync, &iocb, true); |
ac060cba | 1806 | break; |
2c14fa83 | 1807 | case IOCB_CMD_POLL: |
95af8496 | 1808 | ret = aio_poll(req, &iocb); |
a3c0d439 | 1809 | break; |
89319d31 | 1810 | default: |
95af8496 | 1811 | pr_debug("invalid aio operation %d\n", iocb.aio_lio_opcode); |
89319d31 CH |
1812 | ret = -EINVAL; |
1813 | break; | |
1814 | } | |
41003a7b | 1815 | |
92ce4728 | 1816 | /* |
9061d14a AV |
1817 | * If ret is 0, we'd either done aio_complete() ourselves or have |
1818 | * arranged for that to be done asynchronously. Anything non-zero | |
1819 | * means that we need to destroy req ourselves. | |
92ce4728 | 1820 | */ |
9061d14a | 1821 | if (ret) |
89319d31 | 1822 | goto out_put_req; |
1da177e4 | 1823 | return 0; |
1da177e4 | 1824 | out_put_req: |
e1bdd5f2 | 1825 | put_reqs_available(ctx, 1); |
e34ecee2 | 1826 | percpu_ref_put(&ctx->reqs); |
54843f87 CH |
1827 | if (req->ki_eventfd) |
1828 | eventfd_ctx_put(req->ki_eventfd); | |
1829 | kmem_cache_free(kiocb_cachep, req); | |
1da177e4 LT |
1830 | return ret; |
1831 | } | |
1832 | ||
67ba049f AV |
1833 | /* sys_io_submit: |
1834 | * Queue the nr iocbs pointed to by iocbpp for processing. Returns | |
1835 | * the number of iocbs queued. May return -EINVAL if the aio_context | |
1836 | * specified by ctx_id is invalid, if nr is < 0, if the iocb at | |
1837 | * *iocbpp[0] is not properly initialized, if the operation specified | |
1838 | * is invalid for the file descriptor in the iocb. May fail with | |
1839 | * -EFAULT if any of the data structures point to invalid data. May | |
1840 | * fail with -EBADF if the file descriptor specified in the first | |
1841 | * iocb is invalid. May fail with -EAGAIN if insufficient resources | |
1842 | * are available to queue any iocbs. Will return 0 if nr is 0. Will | |
1843 | * fail with -ENOSYS if not implemented. | |
1844 | */ | |
1845 | SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, | |
1846 | struct iocb __user * __user *, iocbpp) | |
1da177e4 LT |
1847 | { |
1848 | struct kioctx *ctx; | |
1849 | long ret = 0; | |
080d676d | 1850 | int i = 0; |
9f5b9425 | 1851 | struct blk_plug plug; |
1da177e4 LT |
1852 | |
1853 | if (unlikely(nr < 0)) | |
1854 | return -EINVAL; | |
1855 | ||
1da177e4 LT |
1856 | ctx = lookup_ioctx(ctx_id); |
1857 | if (unlikely(!ctx)) { | |
caf4167a | 1858 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1859 | return -EINVAL; |
1860 | } | |
1861 | ||
1da92779 AV |
1862 | if (nr > ctx->nr_events) |
1863 | nr = ctx->nr_events; | |
1864 | ||
9f5b9425 | 1865 | blk_start_plug(&plug); |
67ba049f | 1866 | for (i = 0; i < nr; i++) { |
1da177e4 | 1867 | struct iocb __user *user_iocb; |
1da177e4 | 1868 | |
67ba049f | 1869 | if (unlikely(get_user(user_iocb, iocbpp + i))) { |
1da177e4 LT |
1870 | ret = -EFAULT; |
1871 | break; | |
1872 | } | |
1873 | ||
67ba049f | 1874 | ret = io_submit_one(ctx, user_iocb, false); |
1da177e4 LT |
1875 | if (ret) |
1876 | break; | |
1877 | } | |
9f5b9425 | 1878 | blk_finish_plug(&plug); |
1da177e4 | 1879 | |
723be6e3 | 1880 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1881 | return i ? i : ret; |
1882 | } | |
1883 | ||
c00d2c7e | 1884 | #ifdef CONFIG_COMPAT |
c00d2c7e | 1885 | COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id, |
67ba049f | 1886 | int, nr, compat_uptr_t __user *, iocbpp) |
c00d2c7e | 1887 | { |
67ba049f AV |
1888 | struct kioctx *ctx; |
1889 | long ret = 0; | |
1890 | int i = 0; | |
1891 | struct blk_plug plug; | |
c00d2c7e AV |
1892 | |
1893 | if (unlikely(nr < 0)) | |
1894 | return -EINVAL; | |
1895 | ||
67ba049f AV |
1896 | ctx = lookup_ioctx(ctx_id); |
1897 | if (unlikely(!ctx)) { | |
1898 | pr_debug("EINVAL: invalid context id\n"); | |
1899 | return -EINVAL; | |
1900 | } | |
1901 | ||
1da92779 AV |
1902 | if (nr > ctx->nr_events) |
1903 | nr = ctx->nr_events; | |
1904 | ||
67ba049f AV |
1905 | blk_start_plug(&plug); |
1906 | for (i = 0; i < nr; i++) { | |
1907 | compat_uptr_t user_iocb; | |
1908 | ||
1909 | if (unlikely(get_user(user_iocb, iocbpp + i))) { | |
1910 | ret = -EFAULT; | |
1911 | break; | |
1912 | } | |
1913 | ||
1914 | ret = io_submit_one(ctx, compat_ptr(user_iocb), true); | |
1915 | if (ret) | |
1916 | break; | |
1917 | } | |
1918 | blk_finish_plug(&plug); | |
1919 | ||
1920 | percpu_ref_put(&ctx->users); | |
1921 | return i ? i : ret; | |
c00d2c7e AV |
1922 | } |
1923 | #endif | |
1924 | ||
1da177e4 LT |
1925 | /* lookup_kiocb |
1926 | * Finds a given iocb for cancellation. | |
1da177e4 | 1927 | */ |
04b2fa9f | 1928 | static struct aio_kiocb * |
f3a2752a | 1929 | lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb) |
1da177e4 | 1930 | { |
04b2fa9f | 1931 | struct aio_kiocb *kiocb; |
d00689af ZB |
1932 | |
1933 | assert_spin_locked(&ctx->ctx_lock); | |
1934 | ||
1da177e4 | 1935 | /* TODO: use a hash or array, this sucks. */ |
04b2fa9f CH |
1936 | list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) { |
1937 | if (kiocb->ki_user_iocb == iocb) | |
1da177e4 LT |
1938 | return kiocb; |
1939 | } | |
1940 | return NULL; | |
1941 | } | |
1942 | ||
1943 | /* sys_io_cancel: | |
1944 | * Attempts to cancel an iocb previously passed to io_submit. If | |
1945 | * the operation is successfully cancelled, the resulting event is | |
1946 | * copied into the memory pointed to by result without being placed | |
1947 | * into the completion queue and 0 is returned. May fail with | |
1948 | * -EFAULT if any of the data structures pointed to are invalid. | |
1949 | * May fail with -EINVAL if aio_context specified by ctx_id is | |
1950 | * invalid. May fail with -EAGAIN if the iocb specified was not | |
1951 | * cancelled. Will fail with -ENOSYS if not implemented. | |
1952 | */ | |
002c8976 HC |
1953 | SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, |
1954 | struct io_event __user *, result) | |
1da177e4 | 1955 | { |
1da177e4 | 1956 | struct kioctx *ctx; |
04b2fa9f | 1957 | struct aio_kiocb *kiocb; |
888933f8 | 1958 | int ret = -EINVAL; |
1da177e4 | 1959 | u32 key; |
1da177e4 | 1960 | |
f3a2752a | 1961 | if (unlikely(get_user(key, &iocb->aio_key))) |
1da177e4 | 1962 | return -EFAULT; |
f3a2752a CH |
1963 | if (unlikely(key != KIOCB_KEY)) |
1964 | return -EINVAL; | |
1da177e4 LT |
1965 | |
1966 | ctx = lookup_ioctx(ctx_id); | |
1967 | if (unlikely(!ctx)) | |
1968 | return -EINVAL; | |
1969 | ||
1970 | spin_lock_irq(&ctx->ctx_lock); | |
f3a2752a | 1971 | kiocb = lookup_kiocb(ctx, iocb); |
888933f8 CH |
1972 | if (kiocb) { |
1973 | ret = kiocb->ki_cancel(&kiocb->rw); | |
1974 | list_del_init(&kiocb->ki_list); | |
1975 | } | |
1da177e4 LT |
1976 | spin_unlock_irq(&ctx->ctx_lock); |
1977 | ||
906b973c | 1978 | if (!ret) { |
bec68faa KO |
1979 | /* |
1980 | * The result argument is no longer used - the io_event is | |
1981 | * always delivered via the ring buffer. -EINPROGRESS indicates | |
1982 | * cancellation is progress: | |
906b973c | 1983 | */ |
bec68faa | 1984 | ret = -EINPROGRESS; |
906b973c | 1985 | } |
1da177e4 | 1986 | |
723be6e3 | 1987 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1988 | |
1989 | return ret; | |
1990 | } | |
1991 | ||
fa2e62a5 DD |
1992 | static long do_io_getevents(aio_context_t ctx_id, |
1993 | long min_nr, | |
1994 | long nr, | |
1995 | struct io_event __user *events, | |
1996 | struct timespec64 *ts) | |
1997 | { | |
1998 | ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX; | |
1999 | struct kioctx *ioctx = lookup_ioctx(ctx_id); | |
2000 | long ret = -EINVAL; | |
2001 | ||
2002 | if (likely(ioctx)) { | |
2003 | if (likely(min_nr <= nr && min_nr >= 0)) | |
2004 | ret = read_events(ioctx, min_nr, nr, events, until); | |
2005 | percpu_ref_put(&ioctx->users); | |
2006 | } | |
2007 | ||
2008 | return ret; | |
2009 | } | |
2010 | ||
1da177e4 LT |
2011 | /* io_getevents: |
2012 | * Attempts to read at least min_nr events and up to nr events from | |
642b5123 ST |
2013 | * the completion queue for the aio_context specified by ctx_id. If |
2014 | * it succeeds, the number of read events is returned. May fail with | |
2015 | * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is | |
2016 | * out of range, if timeout is out of range. May fail with -EFAULT | |
2017 | * if any of the memory specified is invalid. May return 0 or | |
2018 | * < min_nr if the timeout specified by timeout has elapsed | |
2019 | * before sufficient events are available, where timeout == NULL | |
2020 | * specifies an infinite timeout. Note that the timeout pointed to by | |
6900807c | 2021 | * timeout is relative. Will fail with -ENOSYS if not implemented. |
1da177e4 | 2022 | */ |
002c8976 HC |
2023 | SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, |
2024 | long, min_nr, | |
2025 | long, nr, | |
2026 | struct io_event __user *, events, | |
2027 | struct timespec __user *, timeout) | |
1da177e4 | 2028 | { |
fa2e62a5 | 2029 | struct timespec64 ts; |
7a074e96 CH |
2030 | int ret; |
2031 | ||
2032 | if (timeout && unlikely(get_timespec64(&ts, timeout))) | |
2033 | return -EFAULT; | |
2034 | ||
2035 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL); | |
2036 | if (!ret && signal_pending(current)) | |
2037 | ret = -EINTR; | |
2038 | return ret; | |
2039 | } | |
1da177e4 | 2040 | |
7a074e96 CH |
2041 | SYSCALL_DEFINE6(io_pgetevents, |
2042 | aio_context_t, ctx_id, | |
2043 | long, min_nr, | |
2044 | long, nr, | |
2045 | struct io_event __user *, events, | |
2046 | struct timespec __user *, timeout, | |
2047 | const struct __aio_sigset __user *, usig) | |
2048 | { | |
2049 | struct __aio_sigset ksig = { NULL, }; | |
2050 | sigset_t ksigmask, sigsaved; | |
2051 | struct timespec64 ts; | |
2052 | int ret; | |
2053 | ||
2054 | if (timeout && unlikely(get_timespec64(&ts, timeout))) | |
2055 | return -EFAULT; | |
2056 | ||
2057 | if (usig && copy_from_user(&ksig, usig, sizeof(ksig))) | |
2058 | return -EFAULT; | |
2059 | ||
2060 | if (ksig.sigmask) { | |
2061 | if (ksig.sigsetsize != sizeof(sigset_t)) | |
2062 | return -EINVAL; | |
2063 | if (copy_from_user(&ksigmask, ksig.sigmask, sizeof(ksigmask))) | |
fa2e62a5 | 2064 | return -EFAULT; |
7a074e96 CH |
2065 | sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP)); |
2066 | sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); | |
2067 | } | |
2068 | ||
2069 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL); | |
2070 | if (signal_pending(current)) { | |
2071 | if (ksig.sigmask) { | |
2072 | current->saved_sigmask = sigsaved; | |
2073 | set_restore_sigmask(); | |
2074 | } | |
2075 | ||
2076 | if (!ret) | |
2077 | ret = -ERESTARTNOHAND; | |
2078 | } else { | |
2079 | if (ksig.sigmask) | |
2080 | sigprocmask(SIG_SETMASK, &sigsaved, NULL); | |
1da177e4 | 2081 | } |
fa2e62a5 | 2082 | |
7a074e96 | 2083 | return ret; |
1da177e4 | 2084 | } |
c00d2c7e AV |
2085 | |
2086 | #ifdef CONFIG_COMPAT | |
2087 | COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id, | |
2088 | compat_long_t, min_nr, | |
2089 | compat_long_t, nr, | |
2090 | struct io_event __user *, events, | |
2091 | struct compat_timespec __user *, timeout) | |
2092 | { | |
fa2e62a5 | 2093 | struct timespec64 t; |
7a074e96 CH |
2094 | int ret; |
2095 | ||
2096 | if (timeout && compat_get_timespec64(&t, timeout)) | |
2097 | return -EFAULT; | |
2098 | ||
2099 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL); | |
2100 | if (!ret && signal_pending(current)) | |
2101 | ret = -EINTR; | |
2102 | return ret; | |
2103 | } | |
2104 | ||
c00d2c7e | 2105 | |
7a074e96 CH |
2106 | struct __compat_aio_sigset { |
2107 | compat_sigset_t __user *sigmask; | |
2108 | compat_size_t sigsetsize; | |
2109 | }; | |
2110 | ||
2111 | COMPAT_SYSCALL_DEFINE6(io_pgetevents, | |
2112 | compat_aio_context_t, ctx_id, | |
2113 | compat_long_t, min_nr, | |
2114 | compat_long_t, nr, | |
2115 | struct io_event __user *, events, | |
2116 | struct compat_timespec __user *, timeout, | |
2117 | const struct __compat_aio_sigset __user *, usig) | |
2118 | { | |
2119 | struct __compat_aio_sigset ksig = { NULL, }; | |
2120 | sigset_t ksigmask, sigsaved; | |
2121 | struct timespec64 t; | |
2122 | int ret; | |
2123 | ||
2124 | if (timeout && compat_get_timespec64(&t, timeout)) | |
2125 | return -EFAULT; | |
2126 | ||
2127 | if (usig && copy_from_user(&ksig, usig, sizeof(ksig))) | |
2128 | return -EFAULT; | |
2129 | ||
2130 | if (ksig.sigmask) { | |
2131 | if (ksig.sigsetsize != sizeof(compat_sigset_t)) | |
2132 | return -EINVAL; | |
2133 | if (get_compat_sigset(&ksigmask, ksig.sigmask)) | |
c00d2c7e | 2134 | return -EFAULT; |
7a074e96 CH |
2135 | sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP)); |
2136 | sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); | |
2137 | } | |
c00d2c7e | 2138 | |
7a074e96 CH |
2139 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL); |
2140 | if (signal_pending(current)) { | |
2141 | if (ksig.sigmask) { | |
2142 | current->saved_sigmask = sigsaved; | |
2143 | set_restore_sigmask(); | |
2144 | } | |
2145 | if (!ret) | |
2146 | ret = -ERESTARTNOHAND; | |
2147 | } else { | |
2148 | if (ksig.sigmask) | |
2149 | sigprocmask(SIG_SETMASK, &sigsaved, NULL); | |
c00d2c7e | 2150 | } |
fa2e62a5 | 2151 | |
7a074e96 | 2152 | return ret; |
c00d2c7e AV |
2153 | } |
2154 | #endif |