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 { | |
34e83fc6 KO |
97 | /* |
98 | * This counts the number of available slots in the ringbuffer, | |
99 | * so we avoid overflowing it: it's decremented (if positive) | |
100 | * when allocating a kiocb and incremented when the resulting | |
101 | * io_event is pulled off the ringbuffer. | |
102 | */ | |
103 | atomic_t reqs_available; | |
4e23bcae KO |
104 | } ____cacheline_aligned_in_smp; |
105 | ||
106 | struct { | |
107 | spinlock_t ctx_lock; | |
108 | struct list_head active_reqs; /* used for cancellation */ | |
109 | } ____cacheline_aligned_in_smp; | |
110 | ||
58c85dc2 KO |
111 | struct { |
112 | struct mutex ring_lock; | |
4e23bcae KO |
113 | wait_queue_head_t wait; |
114 | } ____cacheline_aligned_in_smp; | |
58c85dc2 KO |
115 | |
116 | struct { | |
117 | unsigned tail; | |
118 | spinlock_t completion_lock; | |
4e23bcae | 119 | } ____cacheline_aligned_in_smp; |
58c85dc2 KO |
120 | |
121 | struct page *internal_pages[AIO_RING_PAGES]; | |
36bc08cc | 122 | struct file *aio_ring_file; |
4e179bca KO |
123 | }; |
124 | ||
1da177e4 | 125 | /*------ sysctl variables----*/ |
d55b5fda ZB |
126 | static DEFINE_SPINLOCK(aio_nr_lock); |
127 | unsigned long aio_nr; /* current system wide number of aio requests */ | |
128 | unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ | |
1da177e4 LT |
129 | /*----end sysctl variables---*/ |
130 | ||
e18b890b CL |
131 | static struct kmem_cache *kiocb_cachep; |
132 | static struct kmem_cache *kioctx_cachep; | |
1da177e4 | 133 | |
1da177e4 LT |
134 | /* aio_setup |
135 | * Creates the slab caches used by the aio routines, panic on | |
136 | * failure as this is done early during the boot sequence. | |
137 | */ | |
138 | static int __init aio_setup(void) | |
139 | { | |
0a31bd5f CL |
140 | kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
141 | kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); | |
1da177e4 | 142 | |
caf4167a | 143 | pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page)); |
1da177e4 LT |
144 | |
145 | return 0; | |
146 | } | |
385773e0 | 147 | __initcall(aio_setup); |
1da177e4 LT |
148 | |
149 | static void aio_free_ring(struct kioctx *ctx) | |
150 | { | |
36bc08cc GZ |
151 | int i; |
152 | struct file *aio_ring_file = ctx->aio_ring_file; | |
1da177e4 | 153 | |
36bc08cc GZ |
154 | for (i = 0; i < ctx->nr_pages; i++) { |
155 | pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i, | |
156 | page_count(ctx->ring_pages[i])); | |
58c85dc2 | 157 | put_page(ctx->ring_pages[i]); |
36bc08cc | 158 | } |
1da177e4 | 159 | |
58c85dc2 KO |
160 | if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) |
161 | kfree(ctx->ring_pages); | |
36bc08cc GZ |
162 | |
163 | if (aio_ring_file) { | |
164 | truncate_setsize(aio_ring_file->f_inode, 0); | |
165 | pr_debug("pid(%d) i_nlink=%u d_count=%d d_unhashed=%d i_count=%d\n", | |
166 | current->pid, aio_ring_file->f_inode->i_nlink, | |
167 | aio_ring_file->f_path.dentry->d_count, | |
168 | d_unhashed(aio_ring_file->f_path.dentry), | |
169 | atomic_read(&aio_ring_file->f_inode->i_count)); | |
170 | fput(aio_ring_file); | |
171 | ctx->aio_ring_file = NULL; | |
172 | } | |
173 | } | |
174 | ||
175 | static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma) | |
176 | { | |
177 | vma->vm_ops = &generic_file_vm_ops; | |
178 | return 0; | |
179 | } | |
180 | ||
181 | static const struct file_operations aio_ring_fops = { | |
182 | .mmap = aio_ring_mmap, | |
183 | }; | |
184 | ||
185 | static int aio_set_page_dirty(struct page *page) | |
186 | { | |
187 | return 0; | |
188 | } | |
189 | ||
0c45355f | 190 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc GZ |
191 | static int aio_migratepage(struct address_space *mapping, struct page *new, |
192 | struct page *old, enum migrate_mode mode) | |
193 | { | |
194 | struct kioctx *ctx = mapping->private_data; | |
195 | unsigned long flags; | |
196 | unsigned idx = old->index; | |
197 | int rc; | |
198 | ||
199 | /* Writeback must be complete */ | |
200 | BUG_ON(PageWriteback(old)); | |
201 | put_page(old); | |
202 | ||
203 | rc = migrate_page_move_mapping(mapping, new, old, NULL, mode); | |
204 | if (rc != MIGRATEPAGE_SUCCESS) { | |
205 | get_page(old); | |
206 | return rc; | |
207 | } | |
208 | ||
209 | get_page(new); | |
210 | ||
211 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
212 | migrate_page_copy(new, old); | |
213 | ctx->ring_pages[idx] = new; | |
214 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
215 | ||
216 | return rc; | |
1da177e4 | 217 | } |
0c45355f | 218 | #endif |
1da177e4 | 219 | |
36bc08cc GZ |
220 | static const struct address_space_operations aio_ctx_aops = { |
221 | .set_page_dirty = aio_set_page_dirty, | |
0c45355f | 222 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc | 223 | .migratepage = aio_migratepage, |
0c45355f | 224 | #endif |
36bc08cc GZ |
225 | }; |
226 | ||
1da177e4 LT |
227 | static int aio_setup_ring(struct kioctx *ctx) |
228 | { | |
229 | struct aio_ring *ring; | |
1da177e4 | 230 | unsigned nr_events = ctx->max_reqs; |
41003a7b | 231 | struct mm_struct *mm = current->mm; |
41badc15 | 232 | unsigned long size, populate; |
1da177e4 | 233 | int nr_pages; |
36bc08cc GZ |
234 | int i; |
235 | struct file *file; | |
1da177e4 LT |
236 | |
237 | /* Compensate for the ring buffer's head/tail overlap entry */ | |
238 | nr_events += 2; /* 1 is required, 2 for good luck */ | |
239 | ||
240 | size = sizeof(struct aio_ring); | |
241 | size += sizeof(struct io_event) * nr_events; | |
1da177e4 | 242 | |
36bc08cc | 243 | nr_pages = PFN_UP(size); |
1da177e4 LT |
244 | if (nr_pages < 0) |
245 | return -EINVAL; | |
246 | ||
36bc08cc GZ |
247 | file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR); |
248 | if (IS_ERR(file)) { | |
249 | ctx->aio_ring_file = NULL; | |
250 | return -EAGAIN; | |
251 | } | |
252 | ||
253 | file->f_inode->i_mapping->a_ops = &aio_ctx_aops; | |
254 | file->f_inode->i_mapping->private_data = ctx; | |
255 | file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages; | |
256 | ||
257 | for (i = 0; i < nr_pages; i++) { | |
258 | struct page *page; | |
259 | page = find_or_create_page(file->f_inode->i_mapping, | |
260 | i, GFP_HIGHUSER | __GFP_ZERO); | |
261 | if (!page) | |
262 | break; | |
263 | pr_debug("pid(%d) page[%d]->count=%d\n", | |
264 | current->pid, i, page_count(page)); | |
265 | SetPageUptodate(page); | |
266 | SetPageDirty(page); | |
267 | unlock_page(page); | |
268 | } | |
269 | ctx->aio_ring_file = file; | |
270 | nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) | |
271 | / sizeof(struct io_event); | |
1da177e4 | 272 | |
58c85dc2 | 273 | ctx->ring_pages = ctx->internal_pages; |
1da177e4 | 274 | if (nr_pages > AIO_RING_PAGES) { |
58c85dc2 KO |
275 | ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *), |
276 | GFP_KERNEL); | |
277 | if (!ctx->ring_pages) | |
1da177e4 | 278 | return -ENOMEM; |
1da177e4 LT |
279 | } |
280 | ||
58c85dc2 KO |
281 | ctx->mmap_size = nr_pages * PAGE_SIZE; |
282 | pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size); | |
36bc08cc | 283 | |
41003a7b | 284 | down_write(&mm->mmap_sem); |
36bc08cc GZ |
285 | ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size, |
286 | PROT_READ | PROT_WRITE, | |
287 | MAP_SHARED | MAP_POPULATE, 0, &populate); | |
58c85dc2 | 288 | if (IS_ERR((void *)ctx->mmap_base)) { |
41003a7b | 289 | up_write(&mm->mmap_sem); |
58c85dc2 | 290 | ctx->mmap_size = 0; |
1da177e4 LT |
291 | aio_free_ring(ctx); |
292 | return -EAGAIN; | |
293 | } | |
36bc08cc GZ |
294 | up_write(&mm->mmap_sem); |
295 | ||
296 | mm_populate(ctx->mmap_base, populate); | |
1da177e4 | 297 | |
58c85dc2 KO |
298 | pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base); |
299 | ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages, | |
300 | 1, 0, ctx->ring_pages, NULL); | |
36bc08cc GZ |
301 | for (i = 0; i < ctx->nr_pages; i++) |
302 | put_page(ctx->ring_pages[i]); | |
1da177e4 | 303 | |
58c85dc2 | 304 | if (unlikely(ctx->nr_pages != nr_pages)) { |
1da177e4 LT |
305 | aio_free_ring(ctx); |
306 | return -EAGAIN; | |
307 | } | |
308 | ||
58c85dc2 KO |
309 | ctx->user_id = ctx->mmap_base; |
310 | ctx->nr_events = nr_events; /* trusted copy */ | |
1da177e4 | 311 | |
58c85dc2 | 312 | ring = kmap_atomic(ctx->ring_pages[0]); |
1da177e4 LT |
313 | ring->nr = nr_events; /* user copy */ |
314 | ring->id = ctx->user_id; | |
315 | ring->head = ring->tail = 0; | |
316 | ring->magic = AIO_RING_MAGIC; | |
317 | ring->compat_features = AIO_RING_COMPAT_FEATURES; | |
318 | ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; | |
319 | ring->header_length = sizeof(struct aio_ring); | |
e8e3c3d6 | 320 | kunmap_atomic(ring); |
58c85dc2 | 321 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 LT |
322 | |
323 | return 0; | |
324 | } | |
325 | ||
1da177e4 LT |
326 | #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) |
327 | #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) | |
328 | #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) | |
329 | ||
0460fef2 KO |
330 | void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel) |
331 | { | |
332 | struct kioctx *ctx = req->ki_ctx; | |
333 | unsigned long flags; | |
334 | ||
335 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
336 | ||
337 | if (!req->ki_list.next) | |
338 | list_add(&req->ki_list, &ctx->active_reqs); | |
339 | ||
340 | req->ki_cancel = cancel; | |
341 | ||
342 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
343 | } | |
344 | EXPORT_SYMBOL(kiocb_set_cancel_fn); | |
345 | ||
906b973c KO |
346 | static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb, |
347 | struct io_event *res) | |
348 | { | |
0460fef2 | 349 | kiocb_cancel_fn *old, *cancel; |
906b973c KO |
350 | int ret = -EINVAL; |
351 | ||
0460fef2 KO |
352 | /* |
353 | * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it | |
354 | * actually has a cancel function, hence the cmpxchg() | |
355 | */ | |
356 | ||
357 | cancel = ACCESS_ONCE(kiocb->ki_cancel); | |
358 | do { | |
359 | if (!cancel || cancel == KIOCB_CANCELLED) | |
360 | return ret; | |
906b973c | 361 | |
0460fef2 KO |
362 | old = cancel; |
363 | cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED); | |
364 | } while (cancel != old); | |
906b973c | 365 | |
0460fef2 KO |
366 | atomic_inc(&kiocb->ki_users); |
367 | spin_unlock_irq(&ctx->ctx_lock); | |
368 | ||
369 | memset(res, 0, sizeof(*res)); | |
370 | res->obj = (u64)(unsigned long)kiocb->ki_obj.user; | |
371 | res->data = kiocb->ki_user_data; | |
372 | ret = cancel(kiocb, res); | |
373 | ||
374 | spin_lock_irq(&ctx->ctx_lock); | |
906b973c KO |
375 | |
376 | return ret; | |
377 | } | |
378 | ||
36f55889 KO |
379 | static void free_ioctx_rcu(struct rcu_head *head) |
380 | { | |
381 | struct kioctx *ctx = container_of(head, struct kioctx, rcu_head); | |
382 | kmem_cache_free(kioctx_cachep, ctx); | |
383 | } | |
384 | ||
385 | /* | |
386 | * When this function runs, the kioctx has been removed from the "hash table" | |
387 | * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - | |
388 | * now it's safe to cancel any that need to be. | |
389 | */ | |
390 | static void free_ioctx(struct kioctx *ctx) | |
391 | { | |
3e845ce0 | 392 | struct aio_ring *ring; |
36f55889 KO |
393 | struct io_event res; |
394 | struct kiocb *req; | |
3e845ce0 | 395 | unsigned head, avail; |
36f55889 KO |
396 | |
397 | spin_lock_irq(&ctx->ctx_lock); | |
398 | ||
399 | while (!list_empty(&ctx->active_reqs)) { | |
400 | req = list_first_entry(&ctx->active_reqs, | |
401 | struct kiocb, ki_list); | |
402 | ||
403 | list_del_init(&req->ki_list); | |
404 | kiocb_cancel(ctx, req, &res); | |
405 | } | |
406 | ||
407 | spin_unlock_irq(&ctx->ctx_lock); | |
408 | ||
58c85dc2 | 409 | ring = kmap_atomic(ctx->ring_pages[0]); |
3e845ce0 KO |
410 | head = ring->head; |
411 | kunmap_atomic(ring); | |
412 | ||
34e83fc6 | 413 | while (atomic_read(&ctx->reqs_available) < ctx->nr_events - 1) { |
03e04f04 | 414 | wait_event(ctx->wait, |
34e83fc6 KO |
415 | (head != ctx->tail) || |
416 | (atomic_read(&ctx->reqs_available) >= | |
417 | ctx->nr_events - 1)); | |
3e845ce0 | 418 | |
58c85dc2 | 419 | avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head; |
3e845ce0 | 420 | |
34e83fc6 | 421 | atomic_add(avail, &ctx->reqs_available); |
3e845ce0 | 422 | head += avail; |
58c85dc2 | 423 | head %= ctx->nr_events; |
3e845ce0 KO |
424 | } |
425 | ||
34e83fc6 | 426 | WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr_events - 1); |
36f55889 KO |
427 | |
428 | aio_free_ring(ctx); | |
429 | ||
36f55889 KO |
430 | pr_debug("freeing %p\n", ctx); |
431 | ||
432 | /* | |
433 | * Here the call_rcu() is between the wait_event() for reqs_active to | |
434 | * hit 0, and freeing the ioctx. | |
435 | * | |
436 | * aio_complete() decrements reqs_active, but it has to touch the ioctx | |
437 | * after to issue a wakeup so we use rcu. | |
438 | */ | |
439 | call_rcu(&ctx->rcu_head, free_ioctx_rcu); | |
440 | } | |
441 | ||
442 | static void put_ioctx(struct kioctx *ctx) | |
443 | { | |
444 | if (unlikely(atomic_dec_and_test(&ctx->users))) | |
445 | free_ioctx(ctx); | |
446 | } | |
447 | ||
1da177e4 LT |
448 | /* ioctx_alloc |
449 | * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. | |
450 | */ | |
451 | static struct kioctx *ioctx_alloc(unsigned nr_events) | |
452 | { | |
41003a7b | 453 | struct mm_struct *mm = current->mm; |
1da177e4 | 454 | struct kioctx *ctx; |
e23754f8 | 455 | int err = -ENOMEM; |
1da177e4 LT |
456 | |
457 | /* Prevent overflows */ | |
458 | if ((nr_events > (0x10000000U / sizeof(struct io_event))) || | |
459 | (nr_events > (0x10000000U / sizeof(struct kiocb)))) { | |
460 | pr_debug("ENOMEM: nr_events too high\n"); | |
461 | return ERR_PTR(-EINVAL); | |
462 | } | |
463 | ||
2dd542b7 | 464 | if (!nr_events || (unsigned long)nr_events > aio_max_nr) |
1da177e4 LT |
465 | return ERR_PTR(-EAGAIN); |
466 | ||
c3762229 | 467 | ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); |
1da177e4 LT |
468 | if (!ctx) |
469 | return ERR_PTR(-ENOMEM); | |
470 | ||
1da177e4 | 471 | ctx->max_reqs = nr_events; |
1da177e4 | 472 | |
86b62a2c | 473 | atomic_set(&ctx->users, 2); |
36f55889 | 474 | atomic_set(&ctx->dead, 0); |
1da177e4 | 475 | spin_lock_init(&ctx->ctx_lock); |
0460fef2 | 476 | spin_lock_init(&ctx->completion_lock); |
58c85dc2 | 477 | mutex_init(&ctx->ring_lock); |
1da177e4 LT |
478 | init_waitqueue_head(&ctx->wait); |
479 | ||
480 | INIT_LIST_HEAD(&ctx->active_reqs); | |
1da177e4 LT |
481 | |
482 | if (aio_setup_ring(ctx) < 0) | |
483 | goto out_freectx; | |
484 | ||
34e83fc6 KO |
485 | atomic_set(&ctx->reqs_available, ctx->nr_events - 1); |
486 | ||
1da177e4 | 487 | /* limit the number of system wide aios */ |
9fa1cb39 | 488 | spin_lock(&aio_nr_lock); |
2dd542b7 AV |
489 | if (aio_nr + nr_events > aio_max_nr || |
490 | aio_nr + nr_events < aio_nr) { | |
9fa1cb39 | 491 | spin_unlock(&aio_nr_lock); |
1da177e4 | 492 | goto out_cleanup; |
2dd542b7 AV |
493 | } |
494 | aio_nr += ctx->max_reqs; | |
9fa1cb39 | 495 | spin_unlock(&aio_nr_lock); |
1da177e4 | 496 | |
39fa0031 | 497 | /* now link into global list. */ |
abf137dd JA |
498 | spin_lock(&mm->ioctx_lock); |
499 | hlist_add_head_rcu(&ctx->list, &mm->ioctx_list); | |
500 | spin_unlock(&mm->ioctx_lock); | |
1da177e4 | 501 | |
caf4167a | 502 | pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", |
58c85dc2 | 503 | ctx, ctx->user_id, mm, ctx->nr_events); |
1da177e4 LT |
504 | return ctx; |
505 | ||
506 | out_cleanup: | |
e23754f8 AV |
507 | err = -EAGAIN; |
508 | aio_free_ring(ctx); | |
1da177e4 | 509 | out_freectx: |
36bc08cc GZ |
510 | if (ctx->aio_ring_file) |
511 | fput(ctx->aio_ring_file); | |
1da177e4 | 512 | kmem_cache_free(kioctx_cachep, ctx); |
caf4167a | 513 | pr_debug("error allocating ioctx %d\n", err); |
e23754f8 | 514 | return ERR_PTR(err); |
1da177e4 LT |
515 | } |
516 | ||
36f55889 | 517 | static void kill_ioctx_work(struct work_struct *work) |
1da177e4 | 518 | { |
36f55889 | 519 | struct kioctx *ctx = container_of(work, struct kioctx, rcu_work); |
06af121e | 520 | |
36f55889 KO |
521 | wake_up_all(&ctx->wait); |
522 | put_ioctx(ctx); | |
523 | } | |
906b973c | 524 | |
36f55889 KO |
525 | static void kill_ioctx_rcu(struct rcu_head *head) |
526 | { | |
527 | struct kioctx *ctx = container_of(head, struct kioctx, rcu_head); | |
1da177e4 | 528 | |
36f55889 KO |
529 | INIT_WORK(&ctx->rcu_work, kill_ioctx_work); |
530 | schedule_work(&ctx->rcu_work); | |
531 | } | |
1da177e4 | 532 | |
36f55889 KO |
533 | /* kill_ioctx |
534 | * Cancels all outstanding aio requests on an aio context. Used | |
535 | * when the processes owning a context have all exited to encourage | |
536 | * the rapid destruction of the kioctx. | |
537 | */ | |
538 | static void kill_ioctx(struct kioctx *ctx) | |
539 | { | |
540 | if (!atomic_xchg(&ctx->dead, 1)) { | |
541 | hlist_del_rcu(&ctx->list); | |
dee11c23 | 542 | |
36f55889 | 543 | /* |
4fcc712f KO |
544 | * It'd be more correct to do this in free_ioctx(), after all |
545 | * the outstanding kiocbs have finished - but by then io_destroy | |
546 | * has already returned, so io_setup() could potentially return | |
547 | * -EAGAIN with no ioctxs actually in use (as far as userspace | |
548 | * could tell). | |
36f55889 | 549 | */ |
4fcc712f KO |
550 | spin_lock(&aio_nr_lock); |
551 | BUG_ON(aio_nr - ctx->max_reqs > aio_nr); | |
552 | aio_nr -= ctx->max_reqs; | |
553 | spin_unlock(&aio_nr_lock); | |
554 | ||
555 | if (ctx->mmap_size) | |
556 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
557 | ||
558 | /* Between hlist_del_rcu() and dropping the initial ref */ | |
559 | call_rcu(&ctx->rcu_head, kill_ioctx_rcu); | |
36f55889 | 560 | } |
1da177e4 LT |
561 | } |
562 | ||
563 | /* wait_on_sync_kiocb: | |
564 | * Waits on the given sync kiocb to complete. | |
565 | */ | |
fc9b52cd | 566 | ssize_t wait_on_sync_kiocb(struct kiocb *iocb) |
1da177e4 | 567 | { |
11599eba | 568 | while (atomic_read(&iocb->ki_users)) { |
1da177e4 | 569 | set_current_state(TASK_UNINTERRUPTIBLE); |
11599eba | 570 | if (!atomic_read(&iocb->ki_users)) |
1da177e4 | 571 | break; |
41d10da3 | 572 | io_schedule(); |
1da177e4 LT |
573 | } |
574 | __set_current_state(TASK_RUNNING); | |
575 | return iocb->ki_user_data; | |
576 | } | |
385773e0 | 577 | EXPORT_SYMBOL(wait_on_sync_kiocb); |
1da177e4 | 578 | |
36f55889 KO |
579 | /* |
580 | * exit_aio: called when the last user of mm goes away. At this point, there is | |
581 | * no way for any new requests to be submited or any of the io_* syscalls to be | |
582 | * called on the context. | |
583 | * | |
584 | * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on | |
585 | * them. | |
1da177e4 | 586 | */ |
fc9b52cd | 587 | void exit_aio(struct mm_struct *mm) |
1da177e4 | 588 | { |
abf137dd | 589 | struct kioctx *ctx; |
36f55889 | 590 | struct hlist_node *n; |
abf137dd | 591 | |
36f55889 | 592 | hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) { |
1da177e4 LT |
593 | if (1 != atomic_read(&ctx->users)) |
594 | printk(KERN_DEBUG | |
595 | "exit_aio:ioctx still alive: %d %d %d\n", | |
36f55889 KO |
596 | atomic_read(&ctx->users), |
597 | atomic_read(&ctx->dead), | |
34e83fc6 | 598 | atomic_read(&ctx->reqs_available)); |
936af157 AV |
599 | /* |
600 | * We don't need to bother with munmap() here - | |
601 | * exit_mmap(mm) is coming and it'll unmap everything. | |
602 | * Since aio_free_ring() uses non-zero ->mmap_size | |
603 | * as indicator that it needs to unmap the area, | |
604 | * just set it to 0; aio_free_ring() is the only | |
605 | * place that uses ->mmap_size, so it's safe. | |
936af157 | 606 | */ |
58c85dc2 | 607 | ctx->mmap_size = 0; |
36f55889 | 608 | |
4fcc712f | 609 | kill_ioctx(ctx); |
1da177e4 LT |
610 | } |
611 | } | |
612 | ||
1da177e4 | 613 | /* aio_get_req |
11599eba | 614 | * Allocate a slot for an aio request. Increments the ki_users count |
1da177e4 LT |
615 | * of the kioctx so that the kioctx stays around until all requests are |
616 | * complete. Returns NULL if no requests are free. | |
617 | * | |
11599eba | 618 | * Returns with kiocb->ki_users set to 2. The io submit code path holds |
1da177e4 LT |
619 | * an extra reference while submitting the i/o. |
620 | * This prevents races between the aio code path referencing the | |
621 | * req (after submitting it) and aio_complete() freeing the req. | |
622 | */ | |
a1c8eae7 | 623 | static inline struct kiocb *aio_get_req(struct kioctx *ctx) |
1da177e4 | 624 | { |
a1c8eae7 KO |
625 | struct kiocb *req; |
626 | ||
34e83fc6 | 627 | if (atomic_dec_if_positive(&ctx->reqs_available) <= 0) |
a1c8eae7 KO |
628 | return NULL; |
629 | ||
0460fef2 | 630 | req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO); |
1da177e4 | 631 | if (unlikely(!req)) |
a1c8eae7 | 632 | goto out_put; |
1da177e4 | 633 | |
11599eba | 634 | atomic_set(&req->ki_users, 2); |
1da177e4 | 635 | req->ki_ctx = ctx; |
1da177e4 | 636 | |
080d676d | 637 | return req; |
a1c8eae7 | 638 | out_put: |
34e83fc6 | 639 | atomic_inc(&ctx->reqs_available); |
a1c8eae7 | 640 | return NULL; |
1da177e4 LT |
641 | } |
642 | ||
11599eba | 643 | static void kiocb_free(struct kiocb *req) |
1da177e4 | 644 | { |
1d98ebfc KO |
645 | if (req->ki_filp) |
646 | fput(req->ki_filp); | |
13389010 DL |
647 | if (req->ki_eventfd != NULL) |
648 | eventfd_ctx_put(req->ki_eventfd); | |
1da177e4 LT |
649 | if (req->ki_dtor) |
650 | req->ki_dtor(req); | |
eed4e51f BP |
651 | if (req->ki_iovec != &req->ki_inline_vec) |
652 | kfree(req->ki_iovec); | |
1da177e4 | 653 | kmem_cache_free(kiocb_cachep, req); |
1da177e4 LT |
654 | } |
655 | ||
2d68449e | 656 | void aio_put_req(struct kiocb *req) |
1da177e4 | 657 | { |
11599eba KO |
658 | if (atomic_dec_and_test(&req->ki_users)) |
659 | kiocb_free(req); | |
1da177e4 | 660 | } |
385773e0 | 661 | EXPORT_SYMBOL(aio_put_req); |
1da177e4 | 662 | |
d5470b59 | 663 | static struct kioctx *lookup_ioctx(unsigned long ctx_id) |
1da177e4 | 664 | { |
abf137dd | 665 | struct mm_struct *mm = current->mm; |
65c24491 | 666 | struct kioctx *ctx, *ret = NULL; |
1da177e4 | 667 | |
abf137dd JA |
668 | rcu_read_lock(); |
669 | ||
b67bfe0d | 670 | hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) { |
36f55889 KO |
671 | if (ctx->user_id == ctx_id) { |
672 | atomic_inc(&ctx->users); | |
65c24491 | 673 | ret = ctx; |
1da177e4 LT |
674 | break; |
675 | } | |
abf137dd | 676 | } |
1da177e4 | 677 | |
abf137dd | 678 | rcu_read_unlock(); |
65c24491 | 679 | return ret; |
1da177e4 LT |
680 | } |
681 | ||
1da177e4 LT |
682 | /* aio_complete |
683 | * Called when the io request on the given iocb is complete. | |
1da177e4 | 684 | */ |
2d68449e | 685 | void aio_complete(struct kiocb *iocb, long res, long res2) |
1da177e4 LT |
686 | { |
687 | struct kioctx *ctx = iocb->ki_ctx; | |
1da177e4 | 688 | struct aio_ring *ring; |
21b40200 | 689 | struct io_event *ev_page, *event; |
1da177e4 | 690 | unsigned long flags; |
21b40200 | 691 | unsigned tail, pos; |
1da177e4 | 692 | |
20dcae32 ZB |
693 | /* |
694 | * Special case handling for sync iocbs: | |
695 | * - events go directly into the iocb for fast handling | |
696 | * - the sync task with the iocb in its stack holds the single iocb | |
697 | * ref, no other paths have a way to get another ref | |
698 | * - the sync task helpfully left a reference to itself in the iocb | |
1da177e4 LT |
699 | */ |
700 | if (is_sync_kiocb(iocb)) { | |
11599eba | 701 | BUG_ON(atomic_read(&iocb->ki_users) != 1); |
1da177e4 | 702 | iocb->ki_user_data = res; |
11599eba | 703 | atomic_set(&iocb->ki_users, 0); |
1da177e4 | 704 | wake_up_process(iocb->ki_obj.tsk); |
2d68449e | 705 | return; |
1da177e4 LT |
706 | } |
707 | ||
36f55889 | 708 | /* |
36f55889 | 709 | * Take rcu_read_lock() in case the kioctx is being destroyed, as we |
34e83fc6 | 710 | * need to issue a wakeup after incrementing reqs_available. |
1da177e4 | 711 | */ |
36f55889 | 712 | rcu_read_lock(); |
1da177e4 | 713 | |
0460fef2 KO |
714 | if (iocb->ki_list.next) { |
715 | unsigned long flags; | |
716 | ||
717 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
718 | list_del(&iocb->ki_list); | |
719 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
720 | } | |
11599eba | 721 | |
1da177e4 LT |
722 | /* |
723 | * cancelled requests don't get events, userland was given one | |
724 | * when the event got cancelled. | |
725 | */ | |
0460fef2 | 726 | if (unlikely(xchg(&iocb->ki_cancel, |
3e845ce0 | 727 | KIOCB_CANCELLED) == KIOCB_CANCELLED)) { |
34e83fc6 | 728 | atomic_inc(&ctx->reqs_available); |
3e845ce0 | 729 | /* Still need the wake_up in case free_ioctx is waiting */ |
1da177e4 | 730 | goto put_rq; |
3e845ce0 | 731 | } |
1da177e4 | 732 | |
0460fef2 KO |
733 | /* |
734 | * Add a completion event to the ring buffer. Must be done holding | |
4b30f07e | 735 | * ctx->completion_lock to prevent other code from messing with the tail |
0460fef2 KO |
736 | * pointer since we might be called from irq context. |
737 | */ | |
738 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
739 | ||
58c85dc2 | 740 | tail = ctx->tail; |
21b40200 KO |
741 | pos = tail + AIO_EVENTS_OFFSET; |
742 | ||
58c85dc2 | 743 | if (++tail >= ctx->nr_events) |
4bf69b2a | 744 | tail = 0; |
1da177e4 | 745 | |
58c85dc2 | 746 | ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
747 | event = ev_page + pos % AIO_EVENTS_PER_PAGE; |
748 | ||
1da177e4 LT |
749 | event->obj = (u64)(unsigned long)iocb->ki_obj.user; |
750 | event->data = iocb->ki_user_data; | |
751 | event->res = res; | |
752 | event->res2 = res2; | |
753 | ||
21b40200 | 754 | kunmap_atomic(ev_page); |
58c85dc2 | 755 | flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
756 | |
757 | pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n", | |
caf4167a KO |
758 | ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, |
759 | res, res2); | |
1da177e4 LT |
760 | |
761 | /* after flagging the request as done, we | |
762 | * must never even look at it again | |
763 | */ | |
764 | smp_wmb(); /* make event visible before updating tail */ | |
765 | ||
58c85dc2 | 766 | ctx->tail = tail; |
1da177e4 | 767 | |
58c85dc2 | 768 | ring = kmap_atomic(ctx->ring_pages[0]); |
21b40200 | 769 | ring->tail = tail; |
e8e3c3d6 | 770 | kunmap_atomic(ring); |
58c85dc2 | 771 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 | 772 | |
0460fef2 KO |
773 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
774 | ||
21b40200 | 775 | pr_debug("added to ring %p at [%u]\n", iocb, tail); |
8d1c98b0 DL |
776 | |
777 | /* | |
778 | * Check if the user asked us to deliver the result through an | |
779 | * eventfd. The eventfd_signal() function is safe to be called | |
780 | * from IRQ context. | |
781 | */ | |
87c3a86e | 782 | if (iocb->ki_eventfd != NULL) |
8d1c98b0 DL |
783 | eventfd_signal(iocb->ki_eventfd, 1); |
784 | ||
1da177e4 LT |
785 | put_rq: |
786 | /* everything turned out well, dispose of the aiocb. */ | |
11599eba | 787 | aio_put_req(iocb); |
1da177e4 | 788 | |
6cb2a210 QB |
789 | /* |
790 | * We have to order our ring_info tail store above and test | |
791 | * of the wait list below outside the wait lock. This is | |
792 | * like in wake_up_bit() where clearing a bit has to be | |
793 | * ordered with the unlocked test. | |
794 | */ | |
795 | smp_mb(); | |
796 | ||
1da177e4 LT |
797 | if (waitqueue_active(&ctx->wait)) |
798 | wake_up(&ctx->wait); | |
799 | ||
36f55889 | 800 | rcu_read_unlock(); |
1da177e4 | 801 | } |
385773e0 | 802 | EXPORT_SYMBOL(aio_complete); |
1da177e4 | 803 | |
a31ad380 KO |
804 | /* aio_read_events |
805 | * Pull an event off of the ioctx's event ring. Returns the number of | |
806 | * events fetched | |
1da177e4 | 807 | */ |
a31ad380 KO |
808 | static long aio_read_events_ring(struct kioctx *ctx, |
809 | struct io_event __user *event, long nr) | |
1da177e4 | 810 | { |
1da177e4 | 811 | struct aio_ring *ring; |
a31ad380 KO |
812 | unsigned head, pos; |
813 | long ret = 0; | |
814 | int copy_ret; | |
815 | ||
58c85dc2 | 816 | mutex_lock(&ctx->ring_lock); |
1da177e4 | 817 | |
58c85dc2 | 818 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 KO |
819 | head = ring->head; |
820 | kunmap_atomic(ring); | |
821 | ||
58c85dc2 | 822 | pr_debug("h%u t%u m%u\n", head, ctx->tail, ctx->nr_events); |
1da177e4 | 823 | |
58c85dc2 | 824 | if (head == ctx->tail) |
1da177e4 LT |
825 | goto out; |
826 | ||
a31ad380 KO |
827 | while (ret < nr) { |
828 | long avail; | |
829 | struct io_event *ev; | |
830 | struct page *page; | |
831 | ||
58c85dc2 KO |
832 | avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head; |
833 | if (head == ctx->tail) | |
a31ad380 KO |
834 | break; |
835 | ||
836 | avail = min(avail, nr - ret); | |
837 | avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - | |
838 | ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE)); | |
839 | ||
840 | pos = head + AIO_EVENTS_OFFSET; | |
58c85dc2 | 841 | page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]; |
a31ad380 KO |
842 | pos %= AIO_EVENTS_PER_PAGE; |
843 | ||
844 | ev = kmap(page); | |
845 | copy_ret = copy_to_user(event + ret, ev + pos, | |
846 | sizeof(*ev) * avail); | |
847 | kunmap(page); | |
848 | ||
849 | if (unlikely(copy_ret)) { | |
850 | ret = -EFAULT; | |
851 | goto out; | |
852 | } | |
853 | ||
854 | ret += avail; | |
855 | head += avail; | |
58c85dc2 | 856 | head %= ctx->nr_events; |
1da177e4 | 857 | } |
1da177e4 | 858 | |
58c85dc2 | 859 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 860 | ring->head = head; |
91d80a84 | 861 | kunmap_atomic(ring); |
58c85dc2 | 862 | flush_dcache_page(ctx->ring_pages[0]); |
a31ad380 | 863 | |
58c85dc2 | 864 | pr_debug("%li h%u t%u\n", ret, head, ctx->tail); |
3e845ce0 | 865 | |
34e83fc6 | 866 | atomic_add(ret, &ctx->reqs_available); |
a31ad380 | 867 | out: |
58c85dc2 | 868 | mutex_unlock(&ctx->ring_lock); |
a31ad380 | 869 | |
1da177e4 LT |
870 | return ret; |
871 | } | |
872 | ||
a31ad380 KO |
873 | static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr, |
874 | struct io_event __user *event, long *i) | |
1da177e4 | 875 | { |
a31ad380 | 876 | long ret = aio_read_events_ring(ctx, event + *i, nr - *i); |
1da177e4 | 877 | |
a31ad380 KO |
878 | if (ret > 0) |
879 | *i += ret; | |
1da177e4 | 880 | |
a31ad380 KO |
881 | if (unlikely(atomic_read(&ctx->dead))) |
882 | ret = -EINVAL; | |
1da177e4 | 883 | |
a31ad380 KO |
884 | if (!*i) |
885 | *i = ret; | |
1da177e4 | 886 | |
a31ad380 | 887 | return ret < 0 || *i >= min_nr; |
1da177e4 LT |
888 | } |
889 | ||
a31ad380 | 890 | static long read_events(struct kioctx *ctx, long min_nr, long nr, |
1da177e4 LT |
891 | struct io_event __user *event, |
892 | struct timespec __user *timeout) | |
893 | { | |
a31ad380 KO |
894 | ktime_t until = { .tv64 = KTIME_MAX }; |
895 | long ret = 0; | |
1da177e4 | 896 | |
1da177e4 LT |
897 | if (timeout) { |
898 | struct timespec ts; | |
a31ad380 | 899 | |
1da177e4 | 900 | if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) |
a31ad380 | 901 | return -EFAULT; |
1da177e4 | 902 | |
a31ad380 | 903 | until = timespec_to_ktime(ts); |
1da177e4 LT |
904 | } |
905 | ||
a31ad380 KO |
906 | /* |
907 | * Note that aio_read_events() is being called as the conditional - i.e. | |
908 | * we're calling it after prepare_to_wait() has set task state to | |
909 | * TASK_INTERRUPTIBLE. | |
910 | * | |
911 | * But aio_read_events() can block, and if it blocks it's going to flip | |
912 | * the task state back to TASK_RUNNING. | |
913 | * | |
914 | * This should be ok, provided it doesn't flip the state back to | |
915 | * TASK_RUNNING and return 0 too much - that causes us to spin. That | |
916 | * will only happen if the mutex_lock() call blocks, and we then find | |
917 | * the ringbuffer empty. So in practice we should be ok, but it's | |
918 | * something to be aware of when touching this code. | |
919 | */ | |
920 | wait_event_interruptible_hrtimeout(ctx->wait, | |
921 | aio_read_events(ctx, min_nr, nr, event, &ret), until); | |
1da177e4 | 922 | |
a31ad380 KO |
923 | if (!ret && signal_pending(current)) |
924 | ret = -EINTR; | |
1da177e4 | 925 | |
a31ad380 | 926 | return ret; |
1da177e4 LT |
927 | } |
928 | ||
1da177e4 LT |
929 | /* sys_io_setup: |
930 | * Create an aio_context capable of receiving at least nr_events. | |
931 | * ctxp must not point to an aio_context that already exists, and | |
932 | * must be initialized to 0 prior to the call. On successful | |
933 | * creation of the aio_context, *ctxp is filled in with the resulting | |
934 | * handle. May fail with -EINVAL if *ctxp is not initialized, | |
935 | * if the specified nr_events exceeds internal limits. May fail | |
936 | * with -EAGAIN if the specified nr_events exceeds the user's limit | |
937 | * of available events. May fail with -ENOMEM if insufficient kernel | |
938 | * resources are available. May fail with -EFAULT if an invalid | |
939 | * pointer is passed for ctxp. Will fail with -ENOSYS if not | |
940 | * implemented. | |
941 | */ | |
002c8976 | 942 | SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) |
1da177e4 LT |
943 | { |
944 | struct kioctx *ioctx = NULL; | |
945 | unsigned long ctx; | |
946 | long ret; | |
947 | ||
948 | ret = get_user(ctx, ctxp); | |
949 | if (unlikely(ret)) | |
950 | goto out; | |
951 | ||
952 | ret = -EINVAL; | |
d55b5fda ZB |
953 | if (unlikely(ctx || nr_events == 0)) { |
954 | pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", | |
955 | ctx, nr_events); | |
1da177e4 LT |
956 | goto out; |
957 | } | |
958 | ||
959 | ioctx = ioctx_alloc(nr_events); | |
960 | ret = PTR_ERR(ioctx); | |
961 | if (!IS_ERR(ioctx)) { | |
962 | ret = put_user(ioctx->user_id, ctxp); | |
a2e1859a | 963 | if (ret) |
36f55889 | 964 | kill_ioctx(ioctx); |
a2e1859a | 965 | put_ioctx(ioctx); |
1da177e4 LT |
966 | } |
967 | ||
968 | out: | |
969 | return ret; | |
970 | } | |
971 | ||
972 | /* sys_io_destroy: | |
973 | * Destroy the aio_context specified. May cancel any outstanding | |
974 | * AIOs and block on completion. Will fail with -ENOSYS if not | |
642b5123 | 975 | * implemented. May fail with -EINVAL if the context pointed to |
1da177e4 LT |
976 | * is invalid. |
977 | */ | |
002c8976 | 978 | SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) |
1da177e4 LT |
979 | { |
980 | struct kioctx *ioctx = lookup_ioctx(ctx); | |
981 | if (likely(NULL != ioctx)) { | |
36f55889 | 982 | kill_ioctx(ioctx); |
a2e1859a | 983 | put_ioctx(ioctx); |
1da177e4 LT |
984 | return 0; |
985 | } | |
986 | pr_debug("EINVAL: io_destroy: invalid context id\n"); | |
987 | return -EINVAL; | |
988 | } | |
989 | ||
eed4e51f | 990 | static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret) |
1da177e4 | 991 | { |
eed4e51f BP |
992 | struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg]; |
993 | ||
994 | BUG_ON(ret <= 0); | |
995 | ||
996 | while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) { | |
997 | ssize_t this = min((ssize_t)iov->iov_len, ret); | |
998 | iov->iov_base += this; | |
999 | iov->iov_len -= this; | |
1000 | iocb->ki_left -= this; | |
1001 | ret -= this; | |
1002 | if (iov->iov_len == 0) { | |
1003 | iocb->ki_cur_seg++; | |
1004 | iov++; | |
897f15fb | 1005 | } |
eed4e51f | 1006 | } |
1da177e4 | 1007 | |
eed4e51f BP |
1008 | /* the caller should not have done more io than what fit in |
1009 | * the remaining iovecs */ | |
1010 | BUG_ON(ret > 0 && iocb->ki_left == 0); | |
1da177e4 LT |
1011 | } |
1012 | ||
41ef4eb8 KO |
1013 | typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *, |
1014 | unsigned long, loff_t); | |
1015 | ||
1016 | static ssize_t aio_rw_vect_retry(struct kiocb *iocb, int rw, aio_rw_op *rw_op) | |
1da177e4 LT |
1017 | { |
1018 | struct file *file = iocb->ki_filp; | |
eed4e51f BP |
1019 | struct address_space *mapping = file->f_mapping; |
1020 | struct inode *inode = mapping->host; | |
1da177e4 LT |
1021 | ssize_t ret = 0; |
1022 | ||
c2ec6682 RR |
1023 | /* This matches the pread()/pwrite() logic */ |
1024 | if (iocb->ki_pos < 0) | |
1025 | return -EINVAL; | |
1026 | ||
41ef4eb8 | 1027 | if (rw == WRITE) |
8d71db4f | 1028 | file_start_write(file); |
897f15fb | 1029 | do { |
eed4e51f BP |
1030 | ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg], |
1031 | iocb->ki_nr_segs - iocb->ki_cur_seg, | |
1032 | iocb->ki_pos); | |
1033 | if (ret > 0) | |
1034 | aio_advance_iovec(iocb, ret); | |
1035 | ||
1036 | /* retry all partial writes. retry partial reads as long as its a | |
1037 | * regular file. */ | |
1038 | } while (ret > 0 && iocb->ki_left > 0 && | |
41ef4eb8 | 1039 | (rw == WRITE || |
eed4e51f | 1040 | (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)))); |
41ef4eb8 | 1041 | if (rw == WRITE) |
8d71db4f | 1042 | file_end_write(file); |
1da177e4 | 1043 | |
eed4e51f BP |
1044 | /* This means we must have transferred all that we could */ |
1045 | /* No need to retry anymore */ | |
1da177e4 LT |
1046 | if ((ret == 0) || (iocb->ki_left == 0)) |
1047 | ret = iocb->ki_nbytes - iocb->ki_left; | |
1048 | ||
7adfa2ff RR |
1049 | /* If we managed to write some out we return that, rather than |
1050 | * the eventual error. */ | |
41ef4eb8 | 1051 | if (rw == WRITE |
41003a7b | 1052 | && ret < 0 && ret != -EIOCBQUEUED |
7adfa2ff RR |
1053 | && iocb->ki_nbytes - iocb->ki_left) |
1054 | ret = iocb->ki_nbytes - iocb->ki_left; | |
1055 | ||
1da177e4 LT |
1056 | return ret; |
1057 | } | |
1058 | ||
41ef4eb8 | 1059 | static ssize_t aio_setup_vectored_rw(int rw, struct kiocb *kiocb, bool compat) |
eed4e51f BP |
1060 | { |
1061 | ssize_t ret; | |
1062 | ||
41ef4eb8 KO |
1063 | kiocb->ki_nr_segs = kiocb->ki_nbytes; |
1064 | ||
9d85cba7 JM |
1065 | #ifdef CONFIG_COMPAT |
1066 | if (compat) | |
41ef4eb8 | 1067 | ret = compat_rw_copy_check_uvector(rw, |
9d85cba7 | 1068 | (struct compat_iovec __user *)kiocb->ki_buf, |
41ef4eb8 | 1069 | kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec, |
ac34ebb3 | 1070 | &kiocb->ki_iovec); |
9d85cba7 JM |
1071 | else |
1072 | #endif | |
41ef4eb8 | 1073 | ret = rw_copy_check_uvector(rw, |
9d85cba7 | 1074 | (struct iovec __user *)kiocb->ki_buf, |
41ef4eb8 | 1075 | kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec, |
ac34ebb3 | 1076 | &kiocb->ki_iovec); |
eed4e51f | 1077 | if (ret < 0) |
41ef4eb8 | 1078 | return ret; |
a70b52ec | 1079 | |
41ef4eb8 | 1080 | /* ki_nbytes now reflect bytes instead of segs */ |
eed4e51f | 1081 | kiocb->ki_nbytes = ret; |
41ef4eb8 | 1082 | return 0; |
eed4e51f BP |
1083 | } |
1084 | ||
41ef4eb8 | 1085 | static ssize_t aio_setup_single_vector(int rw, struct kiocb *kiocb) |
eed4e51f | 1086 | { |
41ef4eb8 KO |
1087 | if (unlikely(!access_ok(!rw, kiocb->ki_buf, kiocb->ki_nbytes))) |
1088 | return -EFAULT; | |
a70b52ec | 1089 | |
eed4e51f BP |
1090 | kiocb->ki_iovec = &kiocb->ki_inline_vec; |
1091 | kiocb->ki_iovec->iov_base = kiocb->ki_buf; | |
41ef4eb8 | 1092 | kiocb->ki_iovec->iov_len = kiocb->ki_nbytes; |
eed4e51f | 1093 | kiocb->ki_nr_segs = 1; |
eed4e51f BP |
1094 | return 0; |
1095 | } | |
1096 | ||
1da177e4 LT |
1097 | /* |
1098 | * aio_setup_iocb: | |
1099 | * Performs the initial checks and aio retry method | |
1100 | * setup for the kiocb at the time of io submission. | |
1101 | */ | |
41ef4eb8 | 1102 | static ssize_t aio_run_iocb(struct kiocb *req, bool compat) |
1da177e4 | 1103 | { |
41ef4eb8 KO |
1104 | struct file *file = req->ki_filp; |
1105 | ssize_t ret; | |
1106 | int rw; | |
1107 | fmode_t mode; | |
1108 | aio_rw_op *rw_op; | |
1da177e4 | 1109 | |
41ef4eb8 | 1110 | switch (req->ki_opcode) { |
1da177e4 | 1111 | case IOCB_CMD_PREAD: |
eed4e51f | 1112 | case IOCB_CMD_PREADV: |
41ef4eb8 KO |
1113 | mode = FMODE_READ; |
1114 | rw = READ; | |
1115 | rw_op = file->f_op->aio_read; | |
1116 | goto rw_common; | |
1117 | ||
1118 | case IOCB_CMD_PWRITE: | |
eed4e51f | 1119 | case IOCB_CMD_PWRITEV: |
41ef4eb8 KO |
1120 | mode = FMODE_WRITE; |
1121 | rw = WRITE; | |
1122 | rw_op = file->f_op->aio_write; | |
1123 | goto rw_common; | |
1124 | rw_common: | |
1125 | if (unlikely(!(file->f_mode & mode))) | |
1126 | return -EBADF; | |
1127 | ||
1128 | if (!rw_op) | |
1129 | return -EINVAL; | |
1130 | ||
1131 | ret = (req->ki_opcode == IOCB_CMD_PREADV || | |
1132 | req->ki_opcode == IOCB_CMD_PWRITEV) | |
1133 | ? aio_setup_vectored_rw(rw, req, compat) | |
1134 | : aio_setup_single_vector(rw, req); | |
eed4e51f | 1135 | if (ret) |
41ef4eb8 KO |
1136 | return ret; |
1137 | ||
1138 | ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes); | |
1139 | if (ret < 0) | |
1140 | return ret; | |
1141 | ||
1142 | req->ki_nbytes = ret; | |
1143 | req->ki_left = ret; | |
1144 | ||
1145 | ret = aio_rw_vect_retry(req, rw, rw_op); | |
1da177e4 | 1146 | break; |
41ef4eb8 | 1147 | |
1da177e4 | 1148 | case IOCB_CMD_FDSYNC: |
41ef4eb8 KO |
1149 | if (!file->f_op->aio_fsync) |
1150 | return -EINVAL; | |
1151 | ||
1152 | ret = file->f_op->aio_fsync(req, 1); | |
1da177e4 | 1153 | break; |
41ef4eb8 | 1154 | |
1da177e4 | 1155 | case IOCB_CMD_FSYNC: |
41ef4eb8 KO |
1156 | if (!file->f_op->aio_fsync) |
1157 | return -EINVAL; | |
1158 | ||
1159 | ret = file->f_op->aio_fsync(req, 0); | |
1da177e4 | 1160 | break; |
41ef4eb8 | 1161 | |
1da177e4 | 1162 | default: |
caf4167a | 1163 | pr_debug("EINVAL: no operation provided\n"); |
41ef4eb8 | 1164 | return -EINVAL; |
1da177e4 LT |
1165 | } |
1166 | ||
41ef4eb8 KO |
1167 | if (ret != -EIOCBQUEUED) { |
1168 | /* | |
1169 | * There's no easy way to restart the syscall since other AIO's | |
1170 | * may be already running. Just fail this IO with EINTR. | |
1171 | */ | |
1172 | if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR || | |
1173 | ret == -ERESTARTNOHAND || | |
1174 | ret == -ERESTART_RESTARTBLOCK)) | |
1175 | ret = -EINTR; | |
1176 | aio_complete(req, ret, 0); | |
1177 | } | |
1da177e4 LT |
1178 | |
1179 | return 0; | |
1180 | } | |
1181 | ||
d5470b59 | 1182 | static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, |
a1c8eae7 | 1183 | struct iocb *iocb, bool compat) |
1da177e4 LT |
1184 | { |
1185 | struct kiocb *req; | |
1da177e4 LT |
1186 | ssize_t ret; |
1187 | ||
1188 | /* enforce forwards compatibility on users */ | |
9c3060be | 1189 | if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { |
caf4167a | 1190 | pr_debug("EINVAL: reserve field set\n"); |
1da177e4 LT |
1191 | return -EINVAL; |
1192 | } | |
1193 | ||
1194 | /* prevent overflows */ | |
1195 | if (unlikely( | |
1196 | (iocb->aio_buf != (unsigned long)iocb->aio_buf) || | |
1197 | (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || | |
1198 | ((ssize_t)iocb->aio_nbytes < 0) | |
1199 | )) { | |
1200 | pr_debug("EINVAL: io_submit: overflow check\n"); | |
1201 | return -EINVAL; | |
1202 | } | |
1203 | ||
41ef4eb8 | 1204 | req = aio_get_req(ctx); |
1d98ebfc | 1205 | if (unlikely(!req)) |
1da177e4 | 1206 | return -EAGAIN; |
1d98ebfc KO |
1207 | |
1208 | req->ki_filp = fget(iocb->aio_fildes); | |
1209 | if (unlikely(!req->ki_filp)) { | |
1210 | ret = -EBADF; | |
1211 | goto out_put_req; | |
1da177e4 | 1212 | } |
1d98ebfc | 1213 | |
9c3060be DL |
1214 | if (iocb->aio_flags & IOCB_FLAG_RESFD) { |
1215 | /* | |
1216 | * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an | |
1217 | * instance of the file* now. The file descriptor must be | |
1218 | * an eventfd() fd, and will be signaled for each completed | |
1219 | * event using the eventfd_signal() function. | |
1220 | */ | |
13389010 | 1221 | req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); |
801678c5 | 1222 | if (IS_ERR(req->ki_eventfd)) { |
9c3060be | 1223 | ret = PTR_ERR(req->ki_eventfd); |
87c3a86e | 1224 | req->ki_eventfd = NULL; |
9c3060be DL |
1225 | goto out_put_req; |
1226 | } | |
1227 | } | |
1da177e4 | 1228 | |
8a660890 | 1229 | ret = put_user(KIOCB_KEY, &user_iocb->aio_key); |
1da177e4 | 1230 | if (unlikely(ret)) { |
caf4167a | 1231 | pr_debug("EFAULT: aio_key\n"); |
1da177e4 LT |
1232 | goto out_put_req; |
1233 | } | |
1234 | ||
1235 | req->ki_obj.user = user_iocb; | |
1236 | req->ki_user_data = iocb->aio_data; | |
1237 | req->ki_pos = iocb->aio_offset; | |
1238 | ||
1239 | req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; | |
1240 | req->ki_left = req->ki_nbytes = iocb->aio_nbytes; | |
1241 | req->ki_opcode = iocb->aio_lio_opcode; | |
1da177e4 | 1242 | |
41ef4eb8 | 1243 | ret = aio_run_iocb(req, compat); |
41003a7b | 1244 | if (ret) |
7137c6bd | 1245 | goto out_put_req; |
41003a7b | 1246 | |
1da177e4 LT |
1247 | aio_put_req(req); /* drop extra ref to req */ |
1248 | return 0; | |
1da177e4 | 1249 | out_put_req: |
34e83fc6 | 1250 | atomic_inc(&ctx->reqs_available); |
1da177e4 LT |
1251 | aio_put_req(req); /* drop extra ref to req */ |
1252 | aio_put_req(req); /* drop i/o ref to req */ | |
1253 | return ret; | |
1254 | } | |
1255 | ||
9d85cba7 JM |
1256 | long do_io_submit(aio_context_t ctx_id, long nr, |
1257 | struct iocb __user *__user *iocbpp, bool compat) | |
1da177e4 LT |
1258 | { |
1259 | struct kioctx *ctx; | |
1260 | long ret = 0; | |
080d676d | 1261 | int i = 0; |
9f5b9425 | 1262 | struct blk_plug plug; |
1da177e4 LT |
1263 | |
1264 | if (unlikely(nr < 0)) | |
1265 | return -EINVAL; | |
1266 | ||
75e1c70f JM |
1267 | if (unlikely(nr > LONG_MAX/sizeof(*iocbpp))) |
1268 | nr = LONG_MAX/sizeof(*iocbpp); | |
1269 | ||
1da177e4 LT |
1270 | if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) |
1271 | return -EFAULT; | |
1272 | ||
1273 | ctx = lookup_ioctx(ctx_id); | |
1274 | if (unlikely(!ctx)) { | |
caf4167a | 1275 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1276 | return -EINVAL; |
1277 | } | |
1278 | ||
9f5b9425 SL |
1279 | blk_start_plug(&plug); |
1280 | ||
1da177e4 LT |
1281 | /* |
1282 | * AKPM: should this return a partial result if some of the IOs were | |
1283 | * successfully submitted? | |
1284 | */ | |
1285 | for (i=0; i<nr; i++) { | |
1286 | struct iocb __user *user_iocb; | |
1287 | struct iocb tmp; | |
1288 | ||
1289 | if (unlikely(__get_user(user_iocb, iocbpp + i))) { | |
1290 | ret = -EFAULT; | |
1291 | break; | |
1292 | } | |
1293 | ||
1294 | if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { | |
1295 | ret = -EFAULT; | |
1296 | break; | |
1297 | } | |
1298 | ||
a1c8eae7 | 1299 | ret = io_submit_one(ctx, user_iocb, &tmp, compat); |
1da177e4 LT |
1300 | if (ret) |
1301 | break; | |
1302 | } | |
9f5b9425 | 1303 | blk_finish_plug(&plug); |
1da177e4 LT |
1304 | |
1305 | put_ioctx(ctx); | |
1306 | return i ? i : ret; | |
1307 | } | |
1308 | ||
9d85cba7 JM |
1309 | /* sys_io_submit: |
1310 | * Queue the nr iocbs pointed to by iocbpp for processing. Returns | |
1311 | * the number of iocbs queued. May return -EINVAL if the aio_context | |
1312 | * specified by ctx_id is invalid, if nr is < 0, if the iocb at | |
1313 | * *iocbpp[0] is not properly initialized, if the operation specified | |
1314 | * is invalid for the file descriptor in the iocb. May fail with | |
1315 | * -EFAULT if any of the data structures point to invalid data. May | |
1316 | * fail with -EBADF if the file descriptor specified in the first | |
1317 | * iocb is invalid. May fail with -EAGAIN if insufficient resources | |
1318 | * are available to queue any iocbs. Will return 0 if nr is 0. Will | |
1319 | * fail with -ENOSYS if not implemented. | |
1320 | */ | |
1321 | SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, | |
1322 | struct iocb __user * __user *, iocbpp) | |
1323 | { | |
1324 | return do_io_submit(ctx_id, nr, iocbpp, 0); | |
1325 | } | |
1326 | ||
1da177e4 LT |
1327 | /* lookup_kiocb |
1328 | * Finds a given iocb for cancellation. | |
1da177e4 | 1329 | */ |
25ee7e38 AB |
1330 | static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, |
1331 | u32 key) | |
1da177e4 LT |
1332 | { |
1333 | struct list_head *pos; | |
d00689af ZB |
1334 | |
1335 | assert_spin_locked(&ctx->ctx_lock); | |
1336 | ||
8a660890 KO |
1337 | if (key != KIOCB_KEY) |
1338 | return NULL; | |
1339 | ||
1da177e4 LT |
1340 | /* TODO: use a hash or array, this sucks. */ |
1341 | list_for_each(pos, &ctx->active_reqs) { | |
1342 | struct kiocb *kiocb = list_kiocb(pos); | |
8a660890 | 1343 | if (kiocb->ki_obj.user == iocb) |
1da177e4 LT |
1344 | return kiocb; |
1345 | } | |
1346 | return NULL; | |
1347 | } | |
1348 | ||
1349 | /* sys_io_cancel: | |
1350 | * Attempts to cancel an iocb previously passed to io_submit. If | |
1351 | * the operation is successfully cancelled, the resulting event is | |
1352 | * copied into the memory pointed to by result without being placed | |
1353 | * into the completion queue and 0 is returned. May fail with | |
1354 | * -EFAULT if any of the data structures pointed to are invalid. | |
1355 | * May fail with -EINVAL if aio_context specified by ctx_id is | |
1356 | * invalid. May fail with -EAGAIN if the iocb specified was not | |
1357 | * cancelled. Will fail with -ENOSYS if not implemented. | |
1358 | */ | |
002c8976 HC |
1359 | SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, |
1360 | struct io_event __user *, result) | |
1da177e4 | 1361 | { |
906b973c | 1362 | struct io_event res; |
1da177e4 LT |
1363 | struct kioctx *ctx; |
1364 | struct kiocb *kiocb; | |
1365 | u32 key; | |
1366 | int ret; | |
1367 | ||
1368 | ret = get_user(key, &iocb->aio_key); | |
1369 | if (unlikely(ret)) | |
1370 | return -EFAULT; | |
1371 | ||
1372 | ctx = lookup_ioctx(ctx_id); | |
1373 | if (unlikely(!ctx)) | |
1374 | return -EINVAL; | |
1375 | ||
1376 | spin_lock_irq(&ctx->ctx_lock); | |
906b973c | 1377 | |
1da177e4 | 1378 | kiocb = lookup_kiocb(ctx, iocb, key); |
906b973c KO |
1379 | if (kiocb) |
1380 | ret = kiocb_cancel(ctx, kiocb, &res); | |
1381 | else | |
1382 | ret = -EINVAL; | |
1383 | ||
1da177e4 LT |
1384 | spin_unlock_irq(&ctx->ctx_lock); |
1385 | ||
906b973c KO |
1386 | if (!ret) { |
1387 | /* Cancellation succeeded -- copy the result | |
1388 | * into the user's buffer. | |
1389 | */ | |
1390 | if (copy_to_user(result, &res, sizeof(res))) | |
1391 | ret = -EFAULT; | |
1392 | } | |
1da177e4 LT |
1393 | |
1394 | put_ioctx(ctx); | |
1395 | ||
1396 | return ret; | |
1397 | } | |
1398 | ||
1399 | /* io_getevents: | |
1400 | * Attempts to read at least min_nr events and up to nr events from | |
642b5123 ST |
1401 | * the completion queue for the aio_context specified by ctx_id. If |
1402 | * it succeeds, the number of read events is returned. May fail with | |
1403 | * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is | |
1404 | * out of range, if timeout is out of range. May fail with -EFAULT | |
1405 | * if any of the memory specified is invalid. May return 0 or | |
1406 | * < min_nr if the timeout specified by timeout has elapsed | |
1407 | * before sufficient events are available, where timeout == NULL | |
1408 | * specifies an infinite timeout. Note that the timeout pointed to by | |
6900807c | 1409 | * timeout is relative. Will fail with -ENOSYS if not implemented. |
1da177e4 | 1410 | */ |
002c8976 HC |
1411 | SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, |
1412 | long, min_nr, | |
1413 | long, nr, | |
1414 | struct io_event __user *, events, | |
1415 | struct timespec __user *, timeout) | |
1da177e4 LT |
1416 | { |
1417 | struct kioctx *ioctx = lookup_ioctx(ctx_id); | |
1418 | long ret = -EINVAL; | |
1419 | ||
1420 | if (likely(ioctx)) { | |
2e410255 | 1421 | if (likely(min_nr <= nr && min_nr >= 0)) |
1da177e4 LT |
1422 | ret = read_events(ioctx, min_nr, nr, events, timeout); |
1423 | put_ioctx(ioctx); | |
1424 | } | |
1da177e4 LT |
1425 | return ret; |
1426 | } |