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