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. | |
8 | * | |
9 | * See ../COPYING for licensing terms. | |
10 | */ | |
caf4167a KO |
11 | #define pr_fmt(fmt) "%s: " fmt, __func__ |
12 | ||
1da177e4 LT |
13 | #include <linux/kernel.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/errno.h> | |
16 | #include <linux/time.h> | |
17 | #include <linux/aio_abi.h> | |
630d9c47 | 18 | #include <linux/export.h> |
1da177e4 | 19 | #include <linux/syscalls.h> |
b9d128f1 | 20 | #include <linux/backing-dev.h> |
027445c3 | 21 | #include <linux/uio.h> |
1da177e4 | 22 | |
1da177e4 LT |
23 | #include <linux/sched.h> |
24 | #include <linux/fs.h> | |
25 | #include <linux/file.h> | |
26 | #include <linux/mm.h> | |
27 | #include <linux/mman.h> | |
3d2d827f | 28 | #include <linux/mmu_context.h> |
1da177e4 LT |
29 | #include <linux/slab.h> |
30 | #include <linux/timer.h> | |
31 | #include <linux/aio.h> | |
32 | #include <linux/highmem.h> | |
33 | #include <linux/workqueue.h> | |
34 | #include <linux/security.h> | |
9c3060be | 35 | #include <linux/eventfd.h> |
cfb1e33e | 36 | #include <linux/blkdev.h> |
9d85cba7 | 37 | #include <linux/compat.h> |
36bc08cc GZ |
38 | #include <linux/anon_inodes.h> |
39 | #include <linux/migrate.h> | |
40 | #include <linux/ramfs.h> | |
1da177e4 LT |
41 | |
42 | #include <asm/kmap_types.h> | |
43 | #include <asm/uaccess.h> | |
1da177e4 | 44 | |
68d70d03 AV |
45 | #include "internal.h" |
46 | ||
4e179bca KO |
47 | #define AIO_RING_MAGIC 0xa10a10a1 |
48 | #define AIO_RING_COMPAT_FEATURES 1 | |
49 | #define AIO_RING_INCOMPAT_FEATURES 0 | |
50 | struct aio_ring { | |
51 | unsigned id; /* kernel internal index number */ | |
52 | unsigned nr; /* number of io_events */ | |
53 | unsigned head; | |
54 | unsigned tail; | |
55 | ||
56 | unsigned magic; | |
57 | unsigned compat_features; | |
58 | unsigned incompat_features; | |
59 | unsigned header_length; /* size of aio_ring */ | |
60 | ||
61 | ||
62 | struct io_event io_events[0]; | |
63 | }; /* 128 bytes + ring size */ | |
64 | ||
65 | #define AIO_RING_PAGES 8 | |
4e179bca | 66 | |
4e179bca KO |
67 | struct kioctx { |
68 | atomic_t users; | |
36f55889 | 69 | atomic_t dead; |
4e179bca KO |
70 | |
71 | /* This needs improving */ | |
72 | unsigned long user_id; | |
73 | struct hlist_node list; | |
74 | ||
3e845ce0 KO |
75 | /* |
76 | * This is what userspace passed to io_setup(), it's not used for | |
77 | * anything but counting against the global max_reqs quota. | |
78 | * | |
58c85dc2 | 79 | * The real limit is nr_events - 1, which will be larger (see |
3e845ce0 KO |
80 | * aio_setup_ring()) |
81 | */ | |
4e179bca KO |
82 | unsigned max_reqs; |
83 | ||
58c85dc2 KO |
84 | /* Size of ringbuffer, in units of struct io_event */ |
85 | unsigned nr_events; | |
4e179bca | 86 | |
58c85dc2 KO |
87 | unsigned long mmap_base; |
88 | unsigned long mmap_size; | |
89 | ||
90 | struct page **ring_pages; | |
91 | long nr_pages; | |
92 | ||
4e23bcae KO |
93 | struct rcu_head rcu_head; |
94 | struct work_struct rcu_work; | |
95 | ||
96 | struct { | |
97 | atomic_t reqs_active; | |
98 | } ____cacheline_aligned_in_smp; | |
99 | ||
100 | struct { | |
101 | spinlock_t ctx_lock; | |
102 | struct list_head active_reqs; /* used for cancellation */ | |
103 | } ____cacheline_aligned_in_smp; | |
104 | ||
58c85dc2 KO |
105 | struct { |
106 | struct mutex ring_lock; | |
4e23bcae KO |
107 | wait_queue_head_t wait; |
108 | } ____cacheline_aligned_in_smp; | |
58c85dc2 KO |
109 | |
110 | struct { | |
111 | unsigned tail; | |
112 | spinlock_t completion_lock; | |
4e23bcae | 113 | } ____cacheline_aligned_in_smp; |
58c85dc2 KO |
114 | |
115 | struct page *internal_pages[AIO_RING_PAGES]; | |
36bc08cc | 116 | struct file *aio_ring_file; |
4e179bca KO |
117 | }; |
118 | ||
1da177e4 | 119 | /*------ sysctl variables----*/ |
d55b5fda ZB |
120 | static DEFINE_SPINLOCK(aio_nr_lock); |
121 | unsigned long aio_nr; /* current system wide number of aio requests */ | |
122 | unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ | |
1da177e4 LT |
123 | /*----end sysctl variables---*/ |
124 | ||
e18b890b CL |
125 | static struct kmem_cache *kiocb_cachep; |
126 | static struct kmem_cache *kioctx_cachep; | |
1da177e4 | 127 | |
1da177e4 LT |
128 | /* aio_setup |
129 | * Creates the slab caches used by the aio routines, panic on | |
130 | * failure as this is done early during the boot sequence. | |
131 | */ | |
132 | static int __init aio_setup(void) | |
133 | { | |
0a31bd5f CL |
134 | kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
135 | kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); | |
1da177e4 | 136 | |
caf4167a | 137 | pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page)); |
1da177e4 LT |
138 | |
139 | return 0; | |
140 | } | |
385773e0 | 141 | __initcall(aio_setup); |
1da177e4 LT |
142 | |
143 | static void aio_free_ring(struct kioctx *ctx) | |
144 | { | |
36bc08cc GZ |
145 | int i; |
146 | struct file *aio_ring_file = ctx->aio_ring_file; | |
1da177e4 | 147 | |
36bc08cc GZ |
148 | for (i = 0; i < ctx->nr_pages; i++) { |
149 | pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i, | |
150 | page_count(ctx->ring_pages[i])); | |
58c85dc2 | 151 | put_page(ctx->ring_pages[i]); |
36bc08cc | 152 | } |
1da177e4 | 153 | |
58c85dc2 KO |
154 | if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) |
155 | kfree(ctx->ring_pages); | |
36bc08cc GZ |
156 | |
157 | if (aio_ring_file) { | |
158 | truncate_setsize(aio_ring_file->f_inode, 0); | |
159 | pr_debug("pid(%d) i_nlink=%u d_count=%d d_unhashed=%d i_count=%d\n", | |
160 | current->pid, aio_ring_file->f_inode->i_nlink, | |
161 | aio_ring_file->f_path.dentry->d_count, | |
162 | d_unhashed(aio_ring_file->f_path.dentry), | |
163 | atomic_read(&aio_ring_file->f_inode->i_count)); | |
164 | fput(aio_ring_file); | |
165 | ctx->aio_ring_file = NULL; | |
166 | } | |
167 | } | |
168 | ||
169 | static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma) | |
170 | { | |
171 | vma->vm_ops = &generic_file_vm_ops; | |
172 | return 0; | |
173 | } | |
174 | ||
175 | static const struct file_operations aio_ring_fops = { | |
176 | .mmap = aio_ring_mmap, | |
177 | }; | |
178 | ||
179 | static int aio_set_page_dirty(struct page *page) | |
180 | { | |
181 | return 0; | |
182 | } | |
183 | ||
184 | static int aio_migratepage(struct address_space *mapping, struct page *new, | |
185 | struct page *old, enum migrate_mode mode) | |
186 | { | |
187 | struct kioctx *ctx = mapping->private_data; | |
188 | unsigned long flags; | |
189 | unsigned idx = old->index; | |
190 | int rc; | |
191 | ||
192 | /* Writeback must be complete */ | |
193 | BUG_ON(PageWriteback(old)); | |
194 | put_page(old); | |
195 | ||
196 | rc = migrate_page_move_mapping(mapping, new, old, NULL, mode); | |
197 | if (rc != MIGRATEPAGE_SUCCESS) { | |
198 | get_page(old); | |
199 | return rc; | |
200 | } | |
201 | ||
202 | get_page(new); | |
203 | ||
204 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
205 | migrate_page_copy(new, old); | |
206 | ctx->ring_pages[idx] = new; | |
207 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
208 | ||
209 | return rc; | |
1da177e4 LT |
210 | } |
211 | ||
36bc08cc GZ |
212 | static const struct address_space_operations aio_ctx_aops = { |
213 | .set_page_dirty = aio_set_page_dirty, | |
214 | .migratepage = aio_migratepage, | |
215 | }; | |
216 | ||
1da177e4 LT |
217 | static int aio_setup_ring(struct kioctx *ctx) |
218 | { | |
219 | struct aio_ring *ring; | |
1da177e4 | 220 | unsigned nr_events = ctx->max_reqs; |
41003a7b | 221 | struct mm_struct *mm = current->mm; |
41badc15 | 222 | unsigned long size, populate; |
1da177e4 | 223 | int nr_pages; |
36bc08cc GZ |
224 | int i; |
225 | struct file *file; | |
1da177e4 LT |
226 | |
227 | /* Compensate for the ring buffer's head/tail overlap entry */ | |
228 | nr_events += 2; /* 1 is required, 2 for good luck */ | |
229 | ||
230 | size = sizeof(struct aio_ring); | |
231 | size += sizeof(struct io_event) * nr_events; | |
1da177e4 | 232 | |
36bc08cc | 233 | nr_pages = PFN_UP(size); |
1da177e4 LT |
234 | if (nr_pages < 0) |
235 | return -EINVAL; | |
236 | ||
36bc08cc GZ |
237 | file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR); |
238 | if (IS_ERR(file)) { | |
239 | ctx->aio_ring_file = NULL; | |
240 | return -EAGAIN; | |
241 | } | |
242 | ||
243 | file->f_inode->i_mapping->a_ops = &aio_ctx_aops; | |
244 | file->f_inode->i_mapping->private_data = ctx; | |
245 | file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages; | |
246 | ||
247 | for (i = 0; i < nr_pages; i++) { | |
248 | struct page *page; | |
249 | page = find_or_create_page(file->f_inode->i_mapping, | |
250 | i, GFP_HIGHUSER | __GFP_ZERO); | |
251 | if (!page) | |
252 | break; | |
253 | pr_debug("pid(%d) page[%d]->count=%d\n", | |
254 | current->pid, i, page_count(page)); | |
255 | SetPageUptodate(page); | |
256 | SetPageDirty(page); | |
257 | unlock_page(page); | |
258 | } | |
259 | ctx->aio_ring_file = file; | |
260 | nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) | |
261 | / sizeof(struct io_event); | |
1da177e4 | 262 | |
58c85dc2 | 263 | ctx->ring_pages = ctx->internal_pages; |
1da177e4 | 264 | if (nr_pages > AIO_RING_PAGES) { |
58c85dc2 KO |
265 | ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *), |
266 | GFP_KERNEL); | |
267 | if (!ctx->ring_pages) | |
1da177e4 | 268 | return -ENOMEM; |
1da177e4 LT |
269 | } |
270 | ||
58c85dc2 KO |
271 | ctx->mmap_size = nr_pages * PAGE_SIZE; |
272 | pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size); | |
36bc08cc | 273 | |
41003a7b | 274 | down_write(&mm->mmap_sem); |
36bc08cc GZ |
275 | ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size, |
276 | PROT_READ | PROT_WRITE, | |
277 | MAP_SHARED | MAP_POPULATE, 0, &populate); | |
58c85dc2 | 278 | if (IS_ERR((void *)ctx->mmap_base)) { |
41003a7b | 279 | up_write(&mm->mmap_sem); |
58c85dc2 | 280 | ctx->mmap_size = 0; |
1da177e4 LT |
281 | aio_free_ring(ctx); |
282 | return -EAGAIN; | |
283 | } | |
36bc08cc GZ |
284 | up_write(&mm->mmap_sem); |
285 | ||
286 | mm_populate(ctx->mmap_base, populate); | |
1da177e4 | 287 | |
58c85dc2 KO |
288 | pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base); |
289 | ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages, | |
290 | 1, 0, ctx->ring_pages, NULL); | |
36bc08cc GZ |
291 | for (i = 0; i < ctx->nr_pages; i++) |
292 | put_page(ctx->ring_pages[i]); | |
1da177e4 | 293 | |
58c85dc2 | 294 | if (unlikely(ctx->nr_pages != nr_pages)) { |
1da177e4 LT |
295 | aio_free_ring(ctx); |
296 | return -EAGAIN; | |
297 | } | |
298 | ||
58c85dc2 KO |
299 | ctx->user_id = ctx->mmap_base; |
300 | ctx->nr_events = nr_events; /* trusted copy */ | |
1da177e4 | 301 | |
58c85dc2 | 302 | ring = kmap_atomic(ctx->ring_pages[0]); |
1da177e4 LT |
303 | ring->nr = nr_events; /* user copy */ |
304 | ring->id = ctx->user_id; | |
305 | ring->head = ring->tail = 0; | |
306 | ring->magic = AIO_RING_MAGIC; | |
307 | ring->compat_features = AIO_RING_COMPAT_FEATURES; | |
308 | ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; | |
309 | ring->header_length = sizeof(struct aio_ring); | |
e8e3c3d6 | 310 | kunmap_atomic(ring); |
58c85dc2 | 311 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 LT |
312 | |
313 | return 0; | |
314 | } | |
315 | ||
1da177e4 LT |
316 | #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) |
317 | #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) | |
318 | #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) | |
319 | ||
0460fef2 KO |
320 | void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel) |
321 | { | |
322 | struct kioctx *ctx = req->ki_ctx; | |
323 | unsigned long flags; | |
324 | ||
325 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
326 | ||
327 | if (!req->ki_list.next) | |
328 | list_add(&req->ki_list, &ctx->active_reqs); | |
329 | ||
330 | req->ki_cancel = cancel; | |
331 | ||
332 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
333 | } | |
334 | EXPORT_SYMBOL(kiocb_set_cancel_fn); | |
335 | ||
906b973c KO |
336 | static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb, |
337 | struct io_event *res) | |
338 | { | |
0460fef2 | 339 | kiocb_cancel_fn *old, *cancel; |
906b973c KO |
340 | int ret = -EINVAL; |
341 | ||
0460fef2 KO |
342 | /* |
343 | * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it | |
344 | * actually has a cancel function, hence the cmpxchg() | |
345 | */ | |
346 | ||
347 | cancel = ACCESS_ONCE(kiocb->ki_cancel); | |
348 | do { | |
349 | if (!cancel || cancel == KIOCB_CANCELLED) | |
350 | return ret; | |
906b973c | 351 | |
0460fef2 KO |
352 | old = cancel; |
353 | cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED); | |
354 | } while (cancel != old); | |
906b973c | 355 | |
0460fef2 KO |
356 | atomic_inc(&kiocb->ki_users); |
357 | spin_unlock_irq(&ctx->ctx_lock); | |
358 | ||
359 | memset(res, 0, sizeof(*res)); | |
360 | res->obj = (u64)(unsigned long)kiocb->ki_obj.user; | |
361 | res->data = kiocb->ki_user_data; | |
362 | ret = cancel(kiocb, res); | |
363 | ||
364 | spin_lock_irq(&ctx->ctx_lock); | |
906b973c KO |
365 | |
366 | return ret; | |
367 | } | |
368 | ||
36f55889 KO |
369 | static void free_ioctx_rcu(struct rcu_head *head) |
370 | { | |
371 | struct kioctx *ctx = container_of(head, struct kioctx, rcu_head); | |
372 | kmem_cache_free(kioctx_cachep, ctx); | |
373 | } | |
374 | ||
375 | /* | |
376 | * When this function runs, the kioctx has been removed from the "hash table" | |
377 | * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - | |
378 | * now it's safe to cancel any that need to be. | |
379 | */ | |
380 | static void free_ioctx(struct kioctx *ctx) | |
381 | { | |
3e845ce0 | 382 | struct aio_ring *ring; |
36f55889 KO |
383 | struct io_event res; |
384 | struct kiocb *req; | |
3e845ce0 | 385 | unsigned head, avail; |
36f55889 KO |
386 | |
387 | spin_lock_irq(&ctx->ctx_lock); | |
388 | ||
389 | while (!list_empty(&ctx->active_reqs)) { | |
390 | req = list_first_entry(&ctx->active_reqs, | |
391 | struct kiocb, ki_list); | |
392 | ||
393 | list_del_init(&req->ki_list); | |
394 | kiocb_cancel(ctx, req, &res); | |
395 | } | |
396 | ||
397 | spin_unlock_irq(&ctx->ctx_lock); | |
398 | ||
58c85dc2 | 399 | ring = kmap_atomic(ctx->ring_pages[0]); |
3e845ce0 KO |
400 | head = ring->head; |
401 | kunmap_atomic(ring); | |
402 | ||
403 | while (atomic_read(&ctx->reqs_active) > 0) { | |
03e04f04 BL |
404 | wait_event(ctx->wait, |
405 | head != ctx->tail || | |
406 | atomic_read(&ctx->reqs_active) <= 0); | |
3e845ce0 | 407 | |
58c85dc2 | 408 | avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head; |
3e845ce0 KO |
409 | |
410 | atomic_sub(avail, &ctx->reqs_active); | |
411 | head += avail; | |
58c85dc2 | 412 | head %= ctx->nr_events; |
3e845ce0 KO |
413 | } |
414 | ||
415 | WARN_ON(atomic_read(&ctx->reqs_active) < 0); | |
36f55889 KO |
416 | |
417 | aio_free_ring(ctx); | |
418 | ||
36f55889 KO |
419 | pr_debug("freeing %p\n", ctx); |
420 | ||
421 | /* | |
422 | * Here the call_rcu() is between the wait_event() for reqs_active to | |
423 | * hit 0, and freeing the ioctx. | |
424 | * | |
425 | * aio_complete() decrements reqs_active, but it has to touch the ioctx | |
426 | * after to issue a wakeup so we use rcu. | |
427 | */ | |
428 | call_rcu(&ctx->rcu_head, free_ioctx_rcu); | |
429 | } | |
430 | ||
431 | static void put_ioctx(struct kioctx *ctx) | |
432 | { | |
433 | if (unlikely(atomic_dec_and_test(&ctx->users))) | |
434 | free_ioctx(ctx); | |
435 | } | |
436 | ||
1da177e4 LT |
437 | /* ioctx_alloc |
438 | * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. | |
439 | */ | |
440 | static struct kioctx *ioctx_alloc(unsigned nr_events) | |
441 | { | |
41003a7b | 442 | struct mm_struct *mm = current->mm; |
1da177e4 | 443 | struct kioctx *ctx; |
e23754f8 | 444 | int err = -ENOMEM; |
1da177e4 LT |
445 | |
446 | /* Prevent overflows */ | |
447 | if ((nr_events > (0x10000000U / sizeof(struct io_event))) || | |
448 | (nr_events > (0x10000000U / sizeof(struct kiocb)))) { | |
449 | pr_debug("ENOMEM: nr_events too high\n"); | |
450 | return ERR_PTR(-EINVAL); | |
451 | } | |
452 | ||
2dd542b7 | 453 | if (!nr_events || (unsigned long)nr_events > aio_max_nr) |
1da177e4 LT |
454 | return ERR_PTR(-EAGAIN); |
455 | ||
c3762229 | 456 | ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); |
1da177e4 LT |
457 | if (!ctx) |
458 | return ERR_PTR(-ENOMEM); | |
459 | ||
1da177e4 | 460 | ctx->max_reqs = nr_events; |
1da177e4 | 461 | |
86b62a2c | 462 | atomic_set(&ctx->users, 2); |
36f55889 | 463 | atomic_set(&ctx->dead, 0); |
1da177e4 | 464 | spin_lock_init(&ctx->ctx_lock); |
0460fef2 | 465 | spin_lock_init(&ctx->completion_lock); |
58c85dc2 | 466 | mutex_init(&ctx->ring_lock); |
1da177e4 LT |
467 | init_waitqueue_head(&ctx->wait); |
468 | ||
469 | INIT_LIST_HEAD(&ctx->active_reqs); | |
1da177e4 LT |
470 | |
471 | if (aio_setup_ring(ctx) < 0) | |
472 | goto out_freectx; | |
473 | ||
474 | /* limit the number of system wide aios */ | |
9fa1cb39 | 475 | spin_lock(&aio_nr_lock); |
2dd542b7 AV |
476 | if (aio_nr + nr_events > aio_max_nr || |
477 | aio_nr + nr_events < aio_nr) { | |
9fa1cb39 | 478 | spin_unlock(&aio_nr_lock); |
1da177e4 | 479 | goto out_cleanup; |
2dd542b7 AV |
480 | } |
481 | aio_nr += ctx->max_reqs; | |
9fa1cb39 | 482 | spin_unlock(&aio_nr_lock); |
1da177e4 | 483 | |
39fa0031 | 484 | /* now link into global list. */ |
abf137dd JA |
485 | spin_lock(&mm->ioctx_lock); |
486 | hlist_add_head_rcu(&ctx->list, &mm->ioctx_list); | |
487 | spin_unlock(&mm->ioctx_lock); | |
1da177e4 | 488 | |
caf4167a | 489 | pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", |
58c85dc2 | 490 | ctx, ctx->user_id, mm, ctx->nr_events); |
1da177e4 LT |
491 | return ctx; |
492 | ||
493 | out_cleanup: | |
e23754f8 AV |
494 | err = -EAGAIN; |
495 | aio_free_ring(ctx); | |
1da177e4 | 496 | out_freectx: |
36bc08cc GZ |
497 | if (ctx->aio_ring_file) |
498 | fput(ctx->aio_ring_file); | |
1da177e4 | 499 | kmem_cache_free(kioctx_cachep, ctx); |
caf4167a | 500 | pr_debug("error allocating ioctx %d\n", err); |
e23754f8 | 501 | return ERR_PTR(err); |
1da177e4 LT |
502 | } |
503 | ||
36f55889 | 504 | static void kill_ioctx_work(struct work_struct *work) |
1da177e4 | 505 | { |
36f55889 | 506 | struct kioctx *ctx = container_of(work, struct kioctx, rcu_work); |
06af121e | 507 | |
36f55889 KO |
508 | wake_up_all(&ctx->wait); |
509 | put_ioctx(ctx); | |
510 | } | |
906b973c | 511 | |
36f55889 KO |
512 | static void kill_ioctx_rcu(struct rcu_head *head) |
513 | { | |
514 | struct kioctx *ctx = container_of(head, struct kioctx, rcu_head); | |
1da177e4 | 515 | |
36f55889 KO |
516 | INIT_WORK(&ctx->rcu_work, kill_ioctx_work); |
517 | schedule_work(&ctx->rcu_work); | |
518 | } | |
1da177e4 | 519 | |
36f55889 KO |
520 | /* kill_ioctx |
521 | * Cancels all outstanding aio requests on an aio context. Used | |
522 | * when the processes owning a context have all exited to encourage | |
523 | * the rapid destruction of the kioctx. | |
524 | */ | |
525 | static void kill_ioctx(struct kioctx *ctx) | |
526 | { | |
527 | if (!atomic_xchg(&ctx->dead, 1)) { | |
528 | hlist_del_rcu(&ctx->list); | |
dee11c23 | 529 | |
36f55889 | 530 | /* |
4fcc712f KO |
531 | * It'd be more correct to do this in free_ioctx(), after all |
532 | * the outstanding kiocbs have finished - but by then io_destroy | |
533 | * has already returned, so io_setup() could potentially return | |
534 | * -EAGAIN with no ioctxs actually in use (as far as userspace | |
535 | * could tell). | |
36f55889 | 536 | */ |
4fcc712f KO |
537 | spin_lock(&aio_nr_lock); |
538 | BUG_ON(aio_nr - ctx->max_reqs > aio_nr); | |
539 | aio_nr -= ctx->max_reqs; | |
540 | spin_unlock(&aio_nr_lock); | |
541 | ||
542 | if (ctx->mmap_size) | |
543 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
544 | ||
545 | /* Between hlist_del_rcu() and dropping the initial ref */ | |
546 | call_rcu(&ctx->rcu_head, kill_ioctx_rcu); | |
36f55889 | 547 | } |
1da177e4 LT |
548 | } |
549 | ||
550 | /* wait_on_sync_kiocb: | |
551 | * Waits on the given sync kiocb to complete. | |
552 | */ | |
fc9b52cd | 553 | ssize_t wait_on_sync_kiocb(struct kiocb *iocb) |
1da177e4 | 554 | { |
11599eba | 555 | while (atomic_read(&iocb->ki_users)) { |
1da177e4 | 556 | set_current_state(TASK_UNINTERRUPTIBLE); |
11599eba | 557 | if (!atomic_read(&iocb->ki_users)) |
1da177e4 | 558 | break; |
41d10da3 | 559 | io_schedule(); |
1da177e4 LT |
560 | } |
561 | __set_current_state(TASK_RUNNING); | |
562 | return iocb->ki_user_data; | |
563 | } | |
385773e0 | 564 | EXPORT_SYMBOL(wait_on_sync_kiocb); |
1da177e4 | 565 | |
36f55889 KO |
566 | /* |
567 | * exit_aio: called when the last user of mm goes away. At this point, there is | |
568 | * no way for any new requests to be submited or any of the io_* syscalls to be | |
569 | * called on the context. | |
570 | * | |
571 | * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on | |
572 | * them. | |
1da177e4 | 573 | */ |
fc9b52cd | 574 | void exit_aio(struct mm_struct *mm) |
1da177e4 | 575 | { |
abf137dd | 576 | struct kioctx *ctx; |
36f55889 | 577 | struct hlist_node *n; |
abf137dd | 578 | |
36f55889 | 579 | hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) { |
1da177e4 LT |
580 | if (1 != atomic_read(&ctx->users)) |
581 | printk(KERN_DEBUG | |
582 | "exit_aio:ioctx still alive: %d %d %d\n", | |
36f55889 KO |
583 | atomic_read(&ctx->users), |
584 | atomic_read(&ctx->dead), | |
11599eba | 585 | atomic_read(&ctx->reqs_active)); |
936af157 AV |
586 | /* |
587 | * We don't need to bother with munmap() here - | |
588 | * exit_mmap(mm) is coming and it'll unmap everything. | |
589 | * Since aio_free_ring() uses non-zero ->mmap_size | |
590 | * as indicator that it needs to unmap the area, | |
591 | * just set it to 0; aio_free_ring() is the only | |
592 | * place that uses ->mmap_size, so it's safe. | |
936af157 | 593 | */ |
58c85dc2 | 594 | ctx->mmap_size = 0; |
36f55889 | 595 | |
4fcc712f | 596 | kill_ioctx(ctx); |
1da177e4 LT |
597 | } |
598 | } | |
599 | ||
1da177e4 | 600 | /* aio_get_req |
11599eba | 601 | * Allocate a slot for an aio request. Increments the ki_users count |
1da177e4 LT |
602 | * of the kioctx so that the kioctx stays around until all requests are |
603 | * complete. Returns NULL if no requests are free. | |
604 | * | |
11599eba | 605 | * Returns with kiocb->ki_users set to 2. The io submit code path holds |
1da177e4 LT |
606 | * an extra reference while submitting the i/o. |
607 | * This prevents races between the aio code path referencing the | |
608 | * req (after submitting it) and aio_complete() freeing the req. | |
609 | */ | |
a1c8eae7 | 610 | static inline struct kiocb *aio_get_req(struct kioctx *ctx) |
1da177e4 | 611 | { |
a1c8eae7 KO |
612 | struct kiocb *req; |
613 | ||
58c85dc2 | 614 | if (atomic_read(&ctx->reqs_active) >= ctx->nr_events) |
a1c8eae7 KO |
615 | return NULL; |
616 | ||
58c85dc2 | 617 | if (atomic_inc_return(&ctx->reqs_active) > ctx->nr_events - 1) |
a1c8eae7 | 618 | goto out_put; |
1da177e4 | 619 | |
0460fef2 | 620 | req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO); |
1da177e4 | 621 | if (unlikely(!req)) |
a1c8eae7 | 622 | goto out_put; |
1da177e4 | 623 | |
11599eba | 624 | atomic_set(&req->ki_users, 2); |
1da177e4 | 625 | req->ki_ctx = ctx; |
1da177e4 | 626 | |
080d676d | 627 | return req; |
a1c8eae7 KO |
628 | out_put: |
629 | atomic_dec(&ctx->reqs_active); | |
630 | return NULL; | |
1da177e4 LT |
631 | } |
632 | ||
11599eba | 633 | static void kiocb_free(struct kiocb *req) |
1da177e4 | 634 | { |
1d98ebfc KO |
635 | if (req->ki_filp) |
636 | fput(req->ki_filp); | |
13389010 DL |
637 | if (req->ki_eventfd != NULL) |
638 | eventfd_ctx_put(req->ki_eventfd); | |
1da177e4 LT |
639 | if (req->ki_dtor) |
640 | req->ki_dtor(req); | |
eed4e51f BP |
641 | if (req->ki_iovec != &req->ki_inline_vec) |
642 | kfree(req->ki_iovec); | |
1da177e4 | 643 | kmem_cache_free(kiocb_cachep, req); |
1da177e4 LT |
644 | } |
645 | ||
2d68449e | 646 | void aio_put_req(struct kiocb *req) |
1da177e4 | 647 | { |
11599eba KO |
648 | if (atomic_dec_and_test(&req->ki_users)) |
649 | kiocb_free(req); | |
1da177e4 | 650 | } |
385773e0 | 651 | EXPORT_SYMBOL(aio_put_req); |
1da177e4 | 652 | |
d5470b59 | 653 | static struct kioctx *lookup_ioctx(unsigned long ctx_id) |
1da177e4 | 654 | { |
abf137dd | 655 | struct mm_struct *mm = current->mm; |
65c24491 | 656 | struct kioctx *ctx, *ret = NULL; |
1da177e4 | 657 | |
abf137dd JA |
658 | rcu_read_lock(); |
659 | ||
b67bfe0d | 660 | hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) { |
36f55889 KO |
661 | if (ctx->user_id == ctx_id) { |
662 | atomic_inc(&ctx->users); | |
65c24491 | 663 | ret = ctx; |
1da177e4 LT |
664 | break; |
665 | } | |
abf137dd | 666 | } |
1da177e4 | 667 | |
abf137dd | 668 | rcu_read_unlock(); |
65c24491 | 669 | return ret; |
1da177e4 LT |
670 | } |
671 | ||
1da177e4 LT |
672 | /* aio_complete |
673 | * Called when the io request on the given iocb is complete. | |
1da177e4 | 674 | */ |
2d68449e | 675 | void aio_complete(struct kiocb *iocb, long res, long res2) |
1da177e4 LT |
676 | { |
677 | struct kioctx *ctx = iocb->ki_ctx; | |
1da177e4 | 678 | struct aio_ring *ring; |
21b40200 | 679 | struct io_event *ev_page, *event; |
1da177e4 | 680 | unsigned long flags; |
21b40200 | 681 | unsigned tail, pos; |
1da177e4 | 682 | |
20dcae32 ZB |
683 | /* |
684 | * Special case handling for sync iocbs: | |
685 | * - events go directly into the iocb for fast handling | |
686 | * - the sync task with the iocb in its stack holds the single iocb | |
687 | * ref, no other paths have a way to get another ref | |
688 | * - the sync task helpfully left a reference to itself in the iocb | |
1da177e4 LT |
689 | */ |
690 | if (is_sync_kiocb(iocb)) { | |
11599eba | 691 | BUG_ON(atomic_read(&iocb->ki_users) != 1); |
1da177e4 | 692 | iocb->ki_user_data = res; |
11599eba | 693 | atomic_set(&iocb->ki_users, 0); |
1da177e4 | 694 | wake_up_process(iocb->ki_obj.tsk); |
2d68449e | 695 | return; |
1da177e4 LT |
696 | } |
697 | ||
36f55889 | 698 | /* |
36f55889 KO |
699 | * Take rcu_read_lock() in case the kioctx is being destroyed, as we |
700 | * need to issue a wakeup after decrementing reqs_active. | |
1da177e4 | 701 | */ |
36f55889 | 702 | rcu_read_lock(); |
1da177e4 | 703 | |
0460fef2 KO |
704 | if (iocb->ki_list.next) { |
705 | unsigned long flags; | |
706 | ||
707 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
708 | list_del(&iocb->ki_list); | |
709 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
710 | } | |
11599eba | 711 | |
1da177e4 LT |
712 | /* |
713 | * cancelled requests don't get events, userland was given one | |
714 | * when the event got cancelled. | |
715 | */ | |
0460fef2 | 716 | if (unlikely(xchg(&iocb->ki_cancel, |
3e845ce0 KO |
717 | KIOCB_CANCELLED) == KIOCB_CANCELLED)) { |
718 | atomic_dec(&ctx->reqs_active); | |
719 | /* Still need the wake_up in case free_ioctx is waiting */ | |
1da177e4 | 720 | goto put_rq; |
3e845ce0 | 721 | } |
1da177e4 | 722 | |
0460fef2 KO |
723 | /* |
724 | * Add a completion event to the ring buffer. Must be done holding | |
4b30f07e | 725 | * ctx->completion_lock to prevent other code from messing with the tail |
0460fef2 KO |
726 | * pointer since we might be called from irq context. |
727 | */ | |
728 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
729 | ||
58c85dc2 | 730 | tail = ctx->tail; |
21b40200 KO |
731 | pos = tail + AIO_EVENTS_OFFSET; |
732 | ||
58c85dc2 | 733 | if (++tail >= ctx->nr_events) |
4bf69b2a | 734 | tail = 0; |
1da177e4 | 735 | |
58c85dc2 | 736 | ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
737 | event = ev_page + pos % AIO_EVENTS_PER_PAGE; |
738 | ||
1da177e4 LT |
739 | event->obj = (u64)(unsigned long)iocb->ki_obj.user; |
740 | event->data = iocb->ki_user_data; | |
741 | event->res = res; | |
742 | event->res2 = res2; | |
743 | ||
21b40200 | 744 | kunmap_atomic(ev_page); |
58c85dc2 | 745 | flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
746 | |
747 | pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n", | |
caf4167a KO |
748 | ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, |
749 | res, res2); | |
1da177e4 LT |
750 | |
751 | /* after flagging the request as done, we | |
752 | * must never even look at it again | |
753 | */ | |
754 | smp_wmb(); /* make event visible before updating tail */ | |
755 | ||
58c85dc2 | 756 | ctx->tail = tail; |
1da177e4 | 757 | |
58c85dc2 | 758 | ring = kmap_atomic(ctx->ring_pages[0]); |
21b40200 | 759 | ring->tail = tail; |
e8e3c3d6 | 760 | kunmap_atomic(ring); |
58c85dc2 | 761 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 | 762 | |
0460fef2 KO |
763 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
764 | ||
21b40200 | 765 | pr_debug("added to ring %p at [%u]\n", iocb, tail); |
8d1c98b0 DL |
766 | |
767 | /* | |
768 | * Check if the user asked us to deliver the result through an | |
769 | * eventfd. The eventfd_signal() function is safe to be called | |
770 | * from IRQ context. | |
771 | */ | |
87c3a86e | 772 | if (iocb->ki_eventfd != NULL) |
8d1c98b0 DL |
773 | eventfd_signal(iocb->ki_eventfd, 1); |
774 | ||
1da177e4 LT |
775 | put_rq: |
776 | /* everything turned out well, dispose of the aiocb. */ | |
11599eba | 777 | aio_put_req(iocb); |
1da177e4 | 778 | |
6cb2a210 QB |
779 | /* |
780 | * We have to order our ring_info tail store above and test | |
781 | * of the wait list below outside the wait lock. This is | |
782 | * like in wake_up_bit() where clearing a bit has to be | |
783 | * ordered with the unlocked test. | |
784 | */ | |
785 | smp_mb(); | |
786 | ||
1da177e4 LT |
787 | if (waitqueue_active(&ctx->wait)) |
788 | wake_up(&ctx->wait); | |
789 | ||
36f55889 | 790 | rcu_read_unlock(); |
1da177e4 | 791 | } |
385773e0 | 792 | EXPORT_SYMBOL(aio_complete); |
1da177e4 | 793 | |
a31ad380 KO |
794 | /* aio_read_events |
795 | * Pull an event off of the ioctx's event ring. Returns the number of | |
796 | * events fetched | |
1da177e4 | 797 | */ |
a31ad380 KO |
798 | static long aio_read_events_ring(struct kioctx *ctx, |
799 | struct io_event __user *event, long nr) | |
1da177e4 | 800 | { |
1da177e4 | 801 | struct aio_ring *ring; |
a31ad380 KO |
802 | unsigned head, pos; |
803 | long ret = 0; | |
804 | int copy_ret; | |
805 | ||
58c85dc2 | 806 | mutex_lock(&ctx->ring_lock); |
1da177e4 | 807 | |
58c85dc2 | 808 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 KO |
809 | head = ring->head; |
810 | kunmap_atomic(ring); | |
811 | ||
58c85dc2 | 812 | pr_debug("h%u t%u m%u\n", head, ctx->tail, ctx->nr_events); |
1da177e4 | 813 | |
58c85dc2 | 814 | if (head == ctx->tail) |
1da177e4 LT |
815 | goto out; |
816 | ||
a31ad380 KO |
817 | while (ret < nr) { |
818 | long avail; | |
819 | struct io_event *ev; | |
820 | struct page *page; | |
821 | ||
58c85dc2 KO |
822 | avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head; |
823 | if (head == ctx->tail) | |
a31ad380 KO |
824 | break; |
825 | ||
826 | avail = min(avail, nr - ret); | |
827 | avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - | |
828 | ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE)); | |
829 | ||
830 | pos = head + AIO_EVENTS_OFFSET; | |
58c85dc2 | 831 | page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]; |
a31ad380 KO |
832 | pos %= AIO_EVENTS_PER_PAGE; |
833 | ||
834 | ev = kmap(page); | |
835 | copy_ret = copy_to_user(event + ret, ev + pos, | |
836 | sizeof(*ev) * avail); | |
837 | kunmap(page); | |
838 | ||
839 | if (unlikely(copy_ret)) { | |
840 | ret = -EFAULT; | |
841 | goto out; | |
842 | } | |
843 | ||
844 | ret += avail; | |
845 | head += avail; | |
58c85dc2 | 846 | head %= ctx->nr_events; |
1da177e4 | 847 | } |
1da177e4 | 848 | |
58c85dc2 | 849 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 850 | ring->head = head; |
91d80a84 | 851 | kunmap_atomic(ring); |
58c85dc2 | 852 | flush_dcache_page(ctx->ring_pages[0]); |
a31ad380 | 853 | |
58c85dc2 | 854 | pr_debug("%li h%u t%u\n", ret, head, ctx->tail); |
3e845ce0 KO |
855 | |
856 | atomic_sub(ret, &ctx->reqs_active); | |
a31ad380 | 857 | out: |
58c85dc2 | 858 | mutex_unlock(&ctx->ring_lock); |
a31ad380 | 859 | |
1da177e4 LT |
860 | return ret; |
861 | } | |
862 | ||
a31ad380 KO |
863 | static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr, |
864 | struct io_event __user *event, long *i) | |
1da177e4 | 865 | { |
a31ad380 | 866 | long ret = aio_read_events_ring(ctx, event + *i, nr - *i); |
1da177e4 | 867 | |
a31ad380 KO |
868 | if (ret > 0) |
869 | *i += ret; | |
1da177e4 | 870 | |
a31ad380 KO |
871 | if (unlikely(atomic_read(&ctx->dead))) |
872 | ret = -EINVAL; | |
1da177e4 | 873 | |
a31ad380 KO |
874 | if (!*i) |
875 | *i = ret; | |
1da177e4 | 876 | |
a31ad380 | 877 | return ret < 0 || *i >= min_nr; |
1da177e4 LT |
878 | } |
879 | ||
a31ad380 | 880 | static long read_events(struct kioctx *ctx, long min_nr, long nr, |
1da177e4 LT |
881 | struct io_event __user *event, |
882 | struct timespec __user *timeout) | |
883 | { | |
a31ad380 KO |
884 | ktime_t until = { .tv64 = KTIME_MAX }; |
885 | long ret = 0; | |
1da177e4 | 886 | |
1da177e4 LT |
887 | if (timeout) { |
888 | struct timespec ts; | |
a31ad380 | 889 | |
1da177e4 | 890 | if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) |
a31ad380 | 891 | return -EFAULT; |
1da177e4 | 892 | |
a31ad380 | 893 | until = timespec_to_ktime(ts); |
1da177e4 LT |
894 | } |
895 | ||
a31ad380 KO |
896 | /* |
897 | * Note that aio_read_events() is being called as the conditional - i.e. | |
898 | * we're calling it after prepare_to_wait() has set task state to | |
899 | * TASK_INTERRUPTIBLE. | |
900 | * | |
901 | * But aio_read_events() can block, and if it blocks it's going to flip | |
902 | * the task state back to TASK_RUNNING. | |
903 | * | |
904 | * This should be ok, provided it doesn't flip the state back to | |
905 | * TASK_RUNNING and return 0 too much - that causes us to spin. That | |
906 | * will only happen if the mutex_lock() call blocks, and we then find | |
907 | * the ringbuffer empty. So in practice we should be ok, but it's | |
908 | * something to be aware of when touching this code. | |
909 | */ | |
910 | wait_event_interruptible_hrtimeout(ctx->wait, | |
911 | aio_read_events(ctx, min_nr, nr, event, &ret), until); | |
1da177e4 | 912 | |
a31ad380 KO |
913 | if (!ret && signal_pending(current)) |
914 | ret = -EINTR; | |
1da177e4 | 915 | |
a31ad380 | 916 | return ret; |
1da177e4 LT |
917 | } |
918 | ||
1da177e4 LT |
919 | /* sys_io_setup: |
920 | * Create an aio_context capable of receiving at least nr_events. | |
921 | * ctxp must not point to an aio_context that already exists, and | |
922 | * must be initialized to 0 prior to the call. On successful | |
923 | * creation of the aio_context, *ctxp is filled in with the resulting | |
924 | * handle. May fail with -EINVAL if *ctxp is not initialized, | |
925 | * if the specified nr_events exceeds internal limits. May fail | |
926 | * with -EAGAIN if the specified nr_events exceeds the user's limit | |
927 | * of available events. May fail with -ENOMEM if insufficient kernel | |
928 | * resources are available. May fail with -EFAULT if an invalid | |
929 | * pointer is passed for ctxp. Will fail with -ENOSYS if not | |
930 | * implemented. | |
931 | */ | |
002c8976 | 932 | SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) |
1da177e4 LT |
933 | { |
934 | struct kioctx *ioctx = NULL; | |
935 | unsigned long ctx; | |
936 | long ret; | |
937 | ||
938 | ret = get_user(ctx, ctxp); | |
939 | if (unlikely(ret)) | |
940 | goto out; | |
941 | ||
942 | ret = -EINVAL; | |
d55b5fda ZB |
943 | if (unlikely(ctx || nr_events == 0)) { |
944 | pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", | |
945 | ctx, nr_events); | |
1da177e4 LT |
946 | goto out; |
947 | } | |
948 | ||
949 | ioctx = ioctx_alloc(nr_events); | |
950 | ret = PTR_ERR(ioctx); | |
951 | if (!IS_ERR(ioctx)) { | |
952 | ret = put_user(ioctx->user_id, ctxp); | |
a2e1859a | 953 | if (ret) |
36f55889 | 954 | kill_ioctx(ioctx); |
a2e1859a | 955 | put_ioctx(ioctx); |
1da177e4 LT |
956 | } |
957 | ||
958 | out: | |
959 | return ret; | |
960 | } | |
961 | ||
962 | /* sys_io_destroy: | |
963 | * Destroy the aio_context specified. May cancel any outstanding | |
964 | * AIOs and block on completion. Will fail with -ENOSYS if not | |
642b5123 | 965 | * implemented. May fail with -EINVAL if the context pointed to |
1da177e4 LT |
966 | * is invalid. |
967 | */ | |
002c8976 | 968 | SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) |
1da177e4 LT |
969 | { |
970 | struct kioctx *ioctx = lookup_ioctx(ctx); | |
971 | if (likely(NULL != ioctx)) { | |
36f55889 | 972 | kill_ioctx(ioctx); |
a2e1859a | 973 | put_ioctx(ioctx); |
1da177e4 LT |
974 | return 0; |
975 | } | |
976 | pr_debug("EINVAL: io_destroy: invalid context id\n"); | |
977 | return -EINVAL; | |
978 | } | |
979 | ||
eed4e51f | 980 | static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret) |
1da177e4 | 981 | { |
eed4e51f BP |
982 | struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg]; |
983 | ||
984 | BUG_ON(ret <= 0); | |
985 | ||
986 | while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) { | |
987 | ssize_t this = min((ssize_t)iov->iov_len, ret); | |
988 | iov->iov_base += this; | |
989 | iov->iov_len -= this; | |
990 | iocb->ki_left -= this; | |
991 | ret -= this; | |
992 | if (iov->iov_len == 0) { | |
993 | iocb->ki_cur_seg++; | |
994 | iov++; | |
897f15fb | 995 | } |
eed4e51f | 996 | } |
1da177e4 | 997 | |
eed4e51f BP |
998 | /* the caller should not have done more io than what fit in |
999 | * the remaining iovecs */ | |
1000 | BUG_ON(ret > 0 && iocb->ki_left == 0); | |
1da177e4 LT |
1001 | } |
1002 | ||
41ef4eb8 KO |
1003 | typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *, |
1004 | unsigned long, loff_t); | |
1005 | ||
1006 | static ssize_t aio_rw_vect_retry(struct kiocb *iocb, int rw, aio_rw_op *rw_op) | |
1da177e4 LT |
1007 | { |
1008 | struct file *file = iocb->ki_filp; | |
eed4e51f BP |
1009 | struct address_space *mapping = file->f_mapping; |
1010 | struct inode *inode = mapping->host; | |
1da177e4 LT |
1011 | ssize_t ret = 0; |
1012 | ||
c2ec6682 RR |
1013 | /* This matches the pread()/pwrite() logic */ |
1014 | if (iocb->ki_pos < 0) | |
1015 | return -EINVAL; | |
1016 | ||
41ef4eb8 | 1017 | if (rw == WRITE) |
8d71db4f | 1018 | file_start_write(file); |
897f15fb | 1019 | do { |
eed4e51f BP |
1020 | ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg], |
1021 | iocb->ki_nr_segs - iocb->ki_cur_seg, | |
1022 | iocb->ki_pos); | |
1023 | if (ret > 0) | |
1024 | aio_advance_iovec(iocb, ret); | |
1025 | ||
1026 | /* retry all partial writes. retry partial reads as long as its a | |
1027 | * regular file. */ | |
1028 | } while (ret > 0 && iocb->ki_left > 0 && | |
41ef4eb8 | 1029 | (rw == WRITE || |
eed4e51f | 1030 | (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)))); |
41ef4eb8 | 1031 | if (rw == WRITE) |
8d71db4f | 1032 | file_end_write(file); |
1da177e4 | 1033 | |
eed4e51f BP |
1034 | /* This means we must have transferred all that we could */ |
1035 | /* No need to retry anymore */ | |
1da177e4 LT |
1036 | if ((ret == 0) || (iocb->ki_left == 0)) |
1037 | ret = iocb->ki_nbytes - iocb->ki_left; | |
1038 | ||
7adfa2ff RR |
1039 | /* If we managed to write some out we return that, rather than |
1040 | * the eventual error. */ | |
41ef4eb8 | 1041 | if (rw == WRITE |
41003a7b | 1042 | && ret < 0 && ret != -EIOCBQUEUED |
7adfa2ff RR |
1043 | && iocb->ki_nbytes - iocb->ki_left) |
1044 | ret = iocb->ki_nbytes - iocb->ki_left; | |
1045 | ||
1da177e4 LT |
1046 | return ret; |
1047 | } | |
1048 | ||
41ef4eb8 | 1049 | static ssize_t aio_setup_vectored_rw(int rw, struct kiocb *kiocb, bool compat) |
eed4e51f BP |
1050 | { |
1051 | ssize_t ret; | |
1052 | ||
41ef4eb8 KO |
1053 | kiocb->ki_nr_segs = kiocb->ki_nbytes; |
1054 | ||
9d85cba7 JM |
1055 | #ifdef CONFIG_COMPAT |
1056 | if (compat) | |
41ef4eb8 | 1057 | ret = compat_rw_copy_check_uvector(rw, |
9d85cba7 | 1058 | (struct compat_iovec __user *)kiocb->ki_buf, |
41ef4eb8 | 1059 | kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec, |
ac34ebb3 | 1060 | &kiocb->ki_iovec); |
9d85cba7 JM |
1061 | else |
1062 | #endif | |
41ef4eb8 | 1063 | ret = rw_copy_check_uvector(rw, |
9d85cba7 | 1064 | (struct iovec __user *)kiocb->ki_buf, |
41ef4eb8 | 1065 | kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec, |
ac34ebb3 | 1066 | &kiocb->ki_iovec); |
eed4e51f | 1067 | if (ret < 0) |
41ef4eb8 | 1068 | return ret; |
a70b52ec | 1069 | |
41ef4eb8 | 1070 | /* ki_nbytes now reflect bytes instead of segs */ |
eed4e51f | 1071 | kiocb->ki_nbytes = ret; |
41ef4eb8 | 1072 | return 0; |
eed4e51f BP |
1073 | } |
1074 | ||
41ef4eb8 | 1075 | static ssize_t aio_setup_single_vector(int rw, struct kiocb *kiocb) |
eed4e51f | 1076 | { |
41ef4eb8 KO |
1077 | if (unlikely(!access_ok(!rw, kiocb->ki_buf, kiocb->ki_nbytes))) |
1078 | return -EFAULT; | |
a70b52ec | 1079 | |
eed4e51f BP |
1080 | kiocb->ki_iovec = &kiocb->ki_inline_vec; |
1081 | kiocb->ki_iovec->iov_base = kiocb->ki_buf; | |
41ef4eb8 | 1082 | kiocb->ki_iovec->iov_len = kiocb->ki_nbytes; |
eed4e51f | 1083 | kiocb->ki_nr_segs = 1; |
eed4e51f BP |
1084 | return 0; |
1085 | } | |
1086 | ||
1da177e4 LT |
1087 | /* |
1088 | * aio_setup_iocb: | |
1089 | * Performs the initial checks and aio retry method | |
1090 | * setup for the kiocb at the time of io submission. | |
1091 | */ | |
41ef4eb8 | 1092 | static ssize_t aio_run_iocb(struct kiocb *req, bool compat) |
1da177e4 | 1093 | { |
41ef4eb8 KO |
1094 | struct file *file = req->ki_filp; |
1095 | ssize_t ret; | |
1096 | int rw; | |
1097 | fmode_t mode; | |
1098 | aio_rw_op *rw_op; | |
1da177e4 | 1099 | |
41ef4eb8 | 1100 | switch (req->ki_opcode) { |
1da177e4 | 1101 | case IOCB_CMD_PREAD: |
eed4e51f | 1102 | case IOCB_CMD_PREADV: |
41ef4eb8 KO |
1103 | mode = FMODE_READ; |
1104 | rw = READ; | |
1105 | rw_op = file->f_op->aio_read; | |
1106 | goto rw_common; | |
1107 | ||
1108 | case IOCB_CMD_PWRITE: | |
eed4e51f | 1109 | case IOCB_CMD_PWRITEV: |
41ef4eb8 KO |
1110 | mode = FMODE_WRITE; |
1111 | rw = WRITE; | |
1112 | rw_op = file->f_op->aio_write; | |
1113 | goto rw_common; | |
1114 | rw_common: | |
1115 | if (unlikely(!(file->f_mode & mode))) | |
1116 | return -EBADF; | |
1117 | ||
1118 | if (!rw_op) | |
1119 | return -EINVAL; | |
1120 | ||
1121 | ret = (req->ki_opcode == IOCB_CMD_PREADV || | |
1122 | req->ki_opcode == IOCB_CMD_PWRITEV) | |
1123 | ? aio_setup_vectored_rw(rw, req, compat) | |
1124 | : aio_setup_single_vector(rw, req); | |
eed4e51f | 1125 | if (ret) |
41ef4eb8 KO |
1126 | return ret; |
1127 | ||
1128 | ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes); | |
1129 | if (ret < 0) | |
1130 | return ret; | |
1131 | ||
1132 | req->ki_nbytes = ret; | |
1133 | req->ki_left = ret; | |
1134 | ||
1135 | ret = aio_rw_vect_retry(req, rw, rw_op); | |
1da177e4 | 1136 | break; |
41ef4eb8 | 1137 | |
1da177e4 | 1138 | case IOCB_CMD_FDSYNC: |
41ef4eb8 KO |
1139 | if (!file->f_op->aio_fsync) |
1140 | return -EINVAL; | |
1141 | ||
1142 | ret = file->f_op->aio_fsync(req, 1); | |
1da177e4 | 1143 | break; |
41ef4eb8 | 1144 | |
1da177e4 | 1145 | case IOCB_CMD_FSYNC: |
41ef4eb8 KO |
1146 | if (!file->f_op->aio_fsync) |
1147 | return -EINVAL; | |
1148 | ||
1149 | ret = file->f_op->aio_fsync(req, 0); | |
1da177e4 | 1150 | break; |
41ef4eb8 | 1151 | |
1da177e4 | 1152 | default: |
caf4167a | 1153 | pr_debug("EINVAL: no operation provided\n"); |
41ef4eb8 | 1154 | return -EINVAL; |
1da177e4 LT |
1155 | } |
1156 | ||
41ef4eb8 KO |
1157 | if (ret != -EIOCBQUEUED) { |
1158 | /* | |
1159 | * There's no easy way to restart the syscall since other AIO's | |
1160 | * may be already running. Just fail this IO with EINTR. | |
1161 | */ | |
1162 | if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR || | |
1163 | ret == -ERESTARTNOHAND || | |
1164 | ret == -ERESTART_RESTARTBLOCK)) | |
1165 | ret = -EINTR; | |
1166 | aio_complete(req, ret, 0); | |
1167 | } | |
1da177e4 LT |
1168 | |
1169 | return 0; | |
1170 | } | |
1171 | ||
d5470b59 | 1172 | static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, |
a1c8eae7 | 1173 | struct iocb *iocb, bool compat) |
1da177e4 LT |
1174 | { |
1175 | struct kiocb *req; | |
1da177e4 LT |
1176 | ssize_t ret; |
1177 | ||
1178 | /* enforce forwards compatibility on users */ | |
9c3060be | 1179 | if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { |
caf4167a | 1180 | pr_debug("EINVAL: reserve field set\n"); |
1da177e4 LT |
1181 | return -EINVAL; |
1182 | } | |
1183 | ||
1184 | /* prevent overflows */ | |
1185 | if (unlikely( | |
1186 | (iocb->aio_buf != (unsigned long)iocb->aio_buf) || | |
1187 | (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || | |
1188 | ((ssize_t)iocb->aio_nbytes < 0) | |
1189 | )) { | |
1190 | pr_debug("EINVAL: io_submit: overflow check\n"); | |
1191 | return -EINVAL; | |
1192 | } | |
1193 | ||
41ef4eb8 | 1194 | req = aio_get_req(ctx); |
1d98ebfc | 1195 | if (unlikely(!req)) |
1da177e4 | 1196 | return -EAGAIN; |
1d98ebfc KO |
1197 | |
1198 | req->ki_filp = fget(iocb->aio_fildes); | |
1199 | if (unlikely(!req->ki_filp)) { | |
1200 | ret = -EBADF; | |
1201 | goto out_put_req; | |
1da177e4 | 1202 | } |
1d98ebfc | 1203 | |
9c3060be DL |
1204 | if (iocb->aio_flags & IOCB_FLAG_RESFD) { |
1205 | /* | |
1206 | * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an | |
1207 | * instance of the file* now. The file descriptor must be | |
1208 | * an eventfd() fd, and will be signaled for each completed | |
1209 | * event using the eventfd_signal() function. | |
1210 | */ | |
13389010 | 1211 | req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); |
801678c5 | 1212 | if (IS_ERR(req->ki_eventfd)) { |
9c3060be | 1213 | ret = PTR_ERR(req->ki_eventfd); |
87c3a86e | 1214 | req->ki_eventfd = NULL; |
9c3060be DL |
1215 | goto out_put_req; |
1216 | } | |
1217 | } | |
1da177e4 | 1218 | |
8a660890 | 1219 | ret = put_user(KIOCB_KEY, &user_iocb->aio_key); |
1da177e4 | 1220 | if (unlikely(ret)) { |
caf4167a | 1221 | pr_debug("EFAULT: aio_key\n"); |
1da177e4 LT |
1222 | goto out_put_req; |
1223 | } | |
1224 | ||
1225 | req->ki_obj.user = user_iocb; | |
1226 | req->ki_user_data = iocb->aio_data; | |
1227 | req->ki_pos = iocb->aio_offset; | |
1228 | ||
1229 | req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; | |
1230 | req->ki_left = req->ki_nbytes = iocb->aio_nbytes; | |
1231 | req->ki_opcode = iocb->aio_lio_opcode; | |
1da177e4 | 1232 | |
41ef4eb8 | 1233 | ret = aio_run_iocb(req, compat); |
41003a7b | 1234 | if (ret) |
7137c6bd | 1235 | goto out_put_req; |
41003a7b | 1236 | |
1da177e4 LT |
1237 | aio_put_req(req); /* drop extra ref to req */ |
1238 | return 0; | |
1da177e4 | 1239 | out_put_req: |
11599eba | 1240 | atomic_dec(&ctx->reqs_active); |
1da177e4 LT |
1241 | aio_put_req(req); /* drop extra ref to req */ |
1242 | aio_put_req(req); /* drop i/o ref to req */ | |
1243 | return ret; | |
1244 | } | |
1245 | ||
9d85cba7 JM |
1246 | long do_io_submit(aio_context_t ctx_id, long nr, |
1247 | struct iocb __user *__user *iocbpp, bool compat) | |
1da177e4 LT |
1248 | { |
1249 | struct kioctx *ctx; | |
1250 | long ret = 0; | |
080d676d | 1251 | int i = 0; |
9f5b9425 | 1252 | struct blk_plug plug; |
1da177e4 LT |
1253 | |
1254 | if (unlikely(nr < 0)) | |
1255 | return -EINVAL; | |
1256 | ||
75e1c70f JM |
1257 | if (unlikely(nr > LONG_MAX/sizeof(*iocbpp))) |
1258 | nr = LONG_MAX/sizeof(*iocbpp); | |
1259 | ||
1da177e4 LT |
1260 | if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) |
1261 | return -EFAULT; | |
1262 | ||
1263 | ctx = lookup_ioctx(ctx_id); | |
1264 | if (unlikely(!ctx)) { | |
caf4167a | 1265 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1266 | return -EINVAL; |
1267 | } | |
1268 | ||
9f5b9425 SL |
1269 | blk_start_plug(&plug); |
1270 | ||
1da177e4 LT |
1271 | /* |
1272 | * AKPM: should this return a partial result if some of the IOs were | |
1273 | * successfully submitted? | |
1274 | */ | |
1275 | for (i=0; i<nr; i++) { | |
1276 | struct iocb __user *user_iocb; | |
1277 | struct iocb tmp; | |
1278 | ||
1279 | if (unlikely(__get_user(user_iocb, iocbpp + i))) { | |
1280 | ret = -EFAULT; | |
1281 | break; | |
1282 | } | |
1283 | ||
1284 | if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { | |
1285 | ret = -EFAULT; | |
1286 | break; | |
1287 | } | |
1288 | ||
a1c8eae7 | 1289 | ret = io_submit_one(ctx, user_iocb, &tmp, compat); |
1da177e4 LT |
1290 | if (ret) |
1291 | break; | |
1292 | } | |
9f5b9425 | 1293 | blk_finish_plug(&plug); |
1da177e4 LT |
1294 | |
1295 | put_ioctx(ctx); | |
1296 | return i ? i : ret; | |
1297 | } | |
1298 | ||
9d85cba7 JM |
1299 | /* sys_io_submit: |
1300 | * Queue the nr iocbs pointed to by iocbpp for processing. Returns | |
1301 | * the number of iocbs queued. May return -EINVAL if the aio_context | |
1302 | * specified by ctx_id is invalid, if nr is < 0, if the iocb at | |
1303 | * *iocbpp[0] is not properly initialized, if the operation specified | |
1304 | * is invalid for the file descriptor in the iocb. May fail with | |
1305 | * -EFAULT if any of the data structures point to invalid data. May | |
1306 | * fail with -EBADF if the file descriptor specified in the first | |
1307 | * iocb is invalid. May fail with -EAGAIN if insufficient resources | |
1308 | * are available to queue any iocbs. Will return 0 if nr is 0. Will | |
1309 | * fail with -ENOSYS if not implemented. | |
1310 | */ | |
1311 | SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, | |
1312 | struct iocb __user * __user *, iocbpp) | |
1313 | { | |
1314 | return do_io_submit(ctx_id, nr, iocbpp, 0); | |
1315 | } | |
1316 | ||
1da177e4 LT |
1317 | /* lookup_kiocb |
1318 | * Finds a given iocb for cancellation. | |
1da177e4 | 1319 | */ |
25ee7e38 AB |
1320 | static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, |
1321 | u32 key) | |
1da177e4 LT |
1322 | { |
1323 | struct list_head *pos; | |
d00689af ZB |
1324 | |
1325 | assert_spin_locked(&ctx->ctx_lock); | |
1326 | ||
8a660890 KO |
1327 | if (key != KIOCB_KEY) |
1328 | return NULL; | |
1329 | ||
1da177e4 LT |
1330 | /* TODO: use a hash or array, this sucks. */ |
1331 | list_for_each(pos, &ctx->active_reqs) { | |
1332 | struct kiocb *kiocb = list_kiocb(pos); | |
8a660890 | 1333 | if (kiocb->ki_obj.user == iocb) |
1da177e4 LT |
1334 | return kiocb; |
1335 | } | |
1336 | return NULL; | |
1337 | } | |
1338 | ||
1339 | /* sys_io_cancel: | |
1340 | * Attempts to cancel an iocb previously passed to io_submit. If | |
1341 | * the operation is successfully cancelled, the resulting event is | |
1342 | * copied into the memory pointed to by result without being placed | |
1343 | * into the completion queue and 0 is returned. May fail with | |
1344 | * -EFAULT if any of the data structures pointed to are invalid. | |
1345 | * May fail with -EINVAL if aio_context specified by ctx_id is | |
1346 | * invalid. May fail with -EAGAIN if the iocb specified was not | |
1347 | * cancelled. Will fail with -ENOSYS if not implemented. | |
1348 | */ | |
002c8976 HC |
1349 | SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, |
1350 | struct io_event __user *, result) | |
1da177e4 | 1351 | { |
906b973c | 1352 | struct io_event res; |
1da177e4 LT |
1353 | struct kioctx *ctx; |
1354 | struct kiocb *kiocb; | |
1355 | u32 key; | |
1356 | int ret; | |
1357 | ||
1358 | ret = get_user(key, &iocb->aio_key); | |
1359 | if (unlikely(ret)) | |
1360 | return -EFAULT; | |
1361 | ||
1362 | ctx = lookup_ioctx(ctx_id); | |
1363 | if (unlikely(!ctx)) | |
1364 | return -EINVAL; | |
1365 | ||
1366 | spin_lock_irq(&ctx->ctx_lock); | |
906b973c | 1367 | |
1da177e4 | 1368 | kiocb = lookup_kiocb(ctx, iocb, key); |
906b973c KO |
1369 | if (kiocb) |
1370 | ret = kiocb_cancel(ctx, kiocb, &res); | |
1371 | else | |
1372 | ret = -EINVAL; | |
1373 | ||
1da177e4 LT |
1374 | spin_unlock_irq(&ctx->ctx_lock); |
1375 | ||
906b973c KO |
1376 | if (!ret) { |
1377 | /* Cancellation succeeded -- copy the result | |
1378 | * into the user's buffer. | |
1379 | */ | |
1380 | if (copy_to_user(result, &res, sizeof(res))) | |
1381 | ret = -EFAULT; | |
1382 | } | |
1da177e4 LT |
1383 | |
1384 | put_ioctx(ctx); | |
1385 | ||
1386 | return ret; | |
1387 | } | |
1388 | ||
1389 | /* io_getevents: | |
1390 | * Attempts to read at least min_nr events and up to nr events from | |
642b5123 ST |
1391 | * the completion queue for the aio_context specified by ctx_id. If |
1392 | * it succeeds, the number of read events is returned. May fail with | |
1393 | * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is | |
1394 | * out of range, if timeout is out of range. May fail with -EFAULT | |
1395 | * if any of the memory specified is invalid. May return 0 or | |
1396 | * < min_nr if the timeout specified by timeout has elapsed | |
1397 | * before sufficient events are available, where timeout == NULL | |
1398 | * specifies an infinite timeout. Note that the timeout pointed to by | |
6900807c | 1399 | * timeout is relative. Will fail with -ENOSYS if not implemented. |
1da177e4 | 1400 | */ |
002c8976 HC |
1401 | SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, |
1402 | long, min_nr, | |
1403 | long, nr, | |
1404 | struct io_event __user *, events, | |
1405 | struct timespec __user *, timeout) | |
1da177e4 LT |
1406 | { |
1407 | struct kioctx *ioctx = lookup_ioctx(ctx_id); | |
1408 | long ret = -EINVAL; | |
1409 | ||
1410 | if (likely(ioctx)) { | |
2e410255 | 1411 | if (likely(min_nr <= nr && min_nr >= 0)) |
1da177e4 LT |
1412 | ret = read_events(ioctx, min_nr, nr, events, timeout); |
1413 | put_ioctx(ioctx); | |
1414 | } | |
1da177e4 LT |
1415 | return ret; |
1416 | } |