]>
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> |
e1bdd5f2 | 29 | #include <linux/percpu.h> |
1da177e4 LT |
30 | #include <linux/slab.h> |
31 | #include <linux/timer.h> | |
32 | #include <linux/aio.h> | |
33 | #include <linux/highmem.h> | |
34 | #include <linux/workqueue.h> | |
35 | #include <linux/security.h> | |
9c3060be | 36 | #include <linux/eventfd.h> |
cfb1e33e | 37 | #include <linux/blkdev.h> |
9d85cba7 | 38 | #include <linux/compat.h> |
36bc08cc GZ |
39 | #include <linux/anon_inodes.h> |
40 | #include <linux/migrate.h> | |
41 | #include <linux/ramfs.h> | |
723be6e3 | 42 | #include <linux/percpu-refcount.h> |
1da177e4 LT |
43 | |
44 | #include <asm/kmap_types.h> | |
45 | #include <asm/uaccess.h> | |
1da177e4 | 46 | |
68d70d03 AV |
47 | #include "internal.h" |
48 | ||
4e179bca KO |
49 | #define AIO_RING_MAGIC 0xa10a10a1 |
50 | #define AIO_RING_COMPAT_FEATURES 1 | |
51 | #define AIO_RING_INCOMPAT_FEATURES 0 | |
52 | struct aio_ring { | |
53 | unsigned id; /* kernel internal index number */ | |
54 | unsigned nr; /* number of io_events */ | |
55 | unsigned head; | |
56 | unsigned tail; | |
57 | ||
58 | unsigned magic; | |
59 | unsigned compat_features; | |
60 | unsigned incompat_features; | |
61 | unsigned header_length; /* size of aio_ring */ | |
62 | ||
63 | ||
64 | struct io_event io_events[0]; | |
65 | }; /* 128 bytes + ring size */ | |
66 | ||
67 | #define AIO_RING_PAGES 8 | |
4e179bca | 68 | |
e1bdd5f2 KO |
69 | struct kioctx_cpu { |
70 | unsigned reqs_available; | |
71 | }; | |
72 | ||
4e179bca | 73 | struct kioctx { |
723be6e3 | 74 | struct percpu_ref users; |
36f55889 | 75 | atomic_t dead; |
4e179bca KO |
76 | |
77 | /* This needs improving */ | |
78 | unsigned long user_id; | |
79 | struct hlist_node list; | |
80 | ||
e1bdd5f2 KO |
81 | struct __percpu kioctx_cpu *cpu; |
82 | ||
83 | /* | |
84 | * For percpu reqs_available, number of slots we move to/from global | |
85 | * counter at a time: | |
86 | */ | |
87 | unsigned req_batch; | |
3e845ce0 KO |
88 | /* |
89 | * This is what userspace passed to io_setup(), it's not used for | |
90 | * anything but counting against the global max_reqs quota. | |
91 | * | |
58c85dc2 | 92 | * The real limit is nr_events - 1, which will be larger (see |
3e845ce0 KO |
93 | * aio_setup_ring()) |
94 | */ | |
4e179bca KO |
95 | unsigned max_reqs; |
96 | ||
58c85dc2 KO |
97 | /* Size of ringbuffer, in units of struct io_event */ |
98 | unsigned nr_events; | |
4e179bca | 99 | |
58c85dc2 KO |
100 | unsigned long mmap_base; |
101 | unsigned long mmap_size; | |
102 | ||
103 | struct page **ring_pages; | |
104 | long nr_pages; | |
105 | ||
4e23bcae | 106 | struct rcu_head rcu_head; |
723be6e3 | 107 | struct work_struct free_work; |
4e23bcae KO |
108 | |
109 | struct { | |
34e83fc6 KO |
110 | /* |
111 | * This counts the number of available slots in the ringbuffer, | |
112 | * so we avoid overflowing it: it's decremented (if positive) | |
113 | * when allocating a kiocb and incremented when the resulting | |
114 | * io_event is pulled off the ringbuffer. | |
e1bdd5f2 KO |
115 | * |
116 | * We batch accesses to it with a percpu version. | |
34e83fc6 KO |
117 | */ |
118 | atomic_t reqs_available; | |
4e23bcae KO |
119 | } ____cacheline_aligned_in_smp; |
120 | ||
121 | struct { | |
122 | spinlock_t ctx_lock; | |
123 | struct list_head active_reqs; /* used for cancellation */ | |
124 | } ____cacheline_aligned_in_smp; | |
125 | ||
58c85dc2 KO |
126 | struct { |
127 | struct mutex ring_lock; | |
4e23bcae KO |
128 | wait_queue_head_t wait; |
129 | } ____cacheline_aligned_in_smp; | |
58c85dc2 KO |
130 | |
131 | struct { | |
132 | unsigned tail; | |
133 | spinlock_t completion_lock; | |
4e23bcae | 134 | } ____cacheline_aligned_in_smp; |
58c85dc2 KO |
135 | |
136 | struct page *internal_pages[AIO_RING_PAGES]; | |
36bc08cc | 137 | struct file *aio_ring_file; |
4e179bca KO |
138 | }; |
139 | ||
1da177e4 | 140 | /*------ sysctl variables----*/ |
d55b5fda ZB |
141 | static DEFINE_SPINLOCK(aio_nr_lock); |
142 | unsigned long aio_nr; /* current system wide number of aio requests */ | |
143 | unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ | |
1da177e4 LT |
144 | /*----end sysctl variables---*/ |
145 | ||
e18b890b CL |
146 | static struct kmem_cache *kiocb_cachep; |
147 | static struct kmem_cache *kioctx_cachep; | |
1da177e4 | 148 | |
1da177e4 LT |
149 | /* aio_setup |
150 | * Creates the slab caches used by the aio routines, panic on | |
151 | * failure as this is done early during the boot sequence. | |
152 | */ | |
153 | static int __init aio_setup(void) | |
154 | { | |
0a31bd5f CL |
155 | kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
156 | kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); | |
1da177e4 | 157 | |
caf4167a | 158 | pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page)); |
1da177e4 LT |
159 | |
160 | return 0; | |
161 | } | |
385773e0 | 162 | __initcall(aio_setup); |
1da177e4 LT |
163 | |
164 | static void aio_free_ring(struct kioctx *ctx) | |
165 | { | |
36bc08cc GZ |
166 | int i; |
167 | struct file *aio_ring_file = ctx->aio_ring_file; | |
1da177e4 | 168 | |
36bc08cc GZ |
169 | for (i = 0; i < ctx->nr_pages; i++) { |
170 | pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i, | |
171 | page_count(ctx->ring_pages[i])); | |
58c85dc2 | 172 | put_page(ctx->ring_pages[i]); |
36bc08cc | 173 | } |
1da177e4 | 174 | |
58c85dc2 KO |
175 | if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) |
176 | kfree(ctx->ring_pages); | |
36bc08cc GZ |
177 | |
178 | if (aio_ring_file) { | |
179 | truncate_setsize(aio_ring_file->f_inode, 0); | |
180 | pr_debug("pid(%d) i_nlink=%u d_count=%d d_unhashed=%d i_count=%d\n", | |
181 | current->pid, aio_ring_file->f_inode->i_nlink, | |
182 | aio_ring_file->f_path.dentry->d_count, | |
183 | d_unhashed(aio_ring_file->f_path.dentry), | |
184 | atomic_read(&aio_ring_file->f_inode->i_count)); | |
185 | fput(aio_ring_file); | |
186 | ctx->aio_ring_file = NULL; | |
187 | } | |
188 | } | |
189 | ||
190 | static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma) | |
191 | { | |
192 | vma->vm_ops = &generic_file_vm_ops; | |
193 | return 0; | |
194 | } | |
195 | ||
196 | static const struct file_operations aio_ring_fops = { | |
197 | .mmap = aio_ring_mmap, | |
198 | }; | |
199 | ||
200 | static int aio_set_page_dirty(struct page *page) | |
201 | { | |
202 | return 0; | |
203 | } | |
204 | ||
0c45355f | 205 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc GZ |
206 | static int aio_migratepage(struct address_space *mapping, struct page *new, |
207 | struct page *old, enum migrate_mode mode) | |
208 | { | |
209 | struct kioctx *ctx = mapping->private_data; | |
210 | unsigned long flags; | |
211 | unsigned idx = old->index; | |
212 | int rc; | |
213 | ||
214 | /* Writeback must be complete */ | |
215 | BUG_ON(PageWriteback(old)); | |
216 | put_page(old); | |
217 | ||
218 | rc = migrate_page_move_mapping(mapping, new, old, NULL, mode); | |
219 | if (rc != MIGRATEPAGE_SUCCESS) { | |
220 | get_page(old); | |
221 | return rc; | |
222 | } | |
223 | ||
224 | get_page(new); | |
225 | ||
226 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
227 | migrate_page_copy(new, old); | |
228 | ctx->ring_pages[idx] = new; | |
229 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
230 | ||
231 | return rc; | |
1da177e4 | 232 | } |
0c45355f | 233 | #endif |
1da177e4 | 234 | |
36bc08cc GZ |
235 | static const struct address_space_operations aio_ctx_aops = { |
236 | .set_page_dirty = aio_set_page_dirty, | |
0c45355f | 237 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc | 238 | .migratepage = aio_migratepage, |
0c45355f | 239 | #endif |
36bc08cc GZ |
240 | }; |
241 | ||
1da177e4 LT |
242 | static int aio_setup_ring(struct kioctx *ctx) |
243 | { | |
244 | struct aio_ring *ring; | |
1da177e4 | 245 | unsigned nr_events = ctx->max_reqs; |
41003a7b | 246 | struct mm_struct *mm = current->mm; |
41badc15 | 247 | unsigned long size, populate; |
1da177e4 | 248 | int nr_pages; |
36bc08cc GZ |
249 | int i; |
250 | struct file *file; | |
1da177e4 LT |
251 | |
252 | /* Compensate for the ring buffer's head/tail overlap entry */ | |
253 | nr_events += 2; /* 1 is required, 2 for good luck */ | |
254 | ||
255 | size = sizeof(struct aio_ring); | |
256 | size += sizeof(struct io_event) * nr_events; | |
1da177e4 | 257 | |
36bc08cc | 258 | nr_pages = PFN_UP(size); |
1da177e4 LT |
259 | if (nr_pages < 0) |
260 | return -EINVAL; | |
261 | ||
36bc08cc GZ |
262 | file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR); |
263 | if (IS_ERR(file)) { | |
264 | ctx->aio_ring_file = NULL; | |
265 | return -EAGAIN; | |
266 | } | |
267 | ||
268 | file->f_inode->i_mapping->a_ops = &aio_ctx_aops; | |
269 | file->f_inode->i_mapping->private_data = ctx; | |
270 | file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages; | |
271 | ||
272 | for (i = 0; i < nr_pages; i++) { | |
273 | struct page *page; | |
274 | page = find_or_create_page(file->f_inode->i_mapping, | |
275 | i, GFP_HIGHUSER | __GFP_ZERO); | |
276 | if (!page) | |
277 | break; | |
278 | pr_debug("pid(%d) page[%d]->count=%d\n", | |
279 | current->pid, i, page_count(page)); | |
280 | SetPageUptodate(page); | |
281 | SetPageDirty(page); | |
282 | unlock_page(page); | |
283 | } | |
284 | ctx->aio_ring_file = file; | |
285 | nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) | |
286 | / sizeof(struct io_event); | |
1da177e4 | 287 | |
58c85dc2 | 288 | ctx->ring_pages = ctx->internal_pages; |
1da177e4 | 289 | if (nr_pages > AIO_RING_PAGES) { |
58c85dc2 KO |
290 | ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *), |
291 | GFP_KERNEL); | |
292 | if (!ctx->ring_pages) | |
1da177e4 | 293 | return -ENOMEM; |
1da177e4 LT |
294 | } |
295 | ||
58c85dc2 KO |
296 | ctx->mmap_size = nr_pages * PAGE_SIZE; |
297 | pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size); | |
36bc08cc | 298 | |
41003a7b | 299 | down_write(&mm->mmap_sem); |
36bc08cc GZ |
300 | ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size, |
301 | PROT_READ | PROT_WRITE, | |
302 | MAP_SHARED | MAP_POPULATE, 0, &populate); | |
58c85dc2 | 303 | if (IS_ERR((void *)ctx->mmap_base)) { |
41003a7b | 304 | up_write(&mm->mmap_sem); |
58c85dc2 | 305 | ctx->mmap_size = 0; |
1da177e4 LT |
306 | aio_free_ring(ctx); |
307 | return -EAGAIN; | |
308 | } | |
36bc08cc GZ |
309 | up_write(&mm->mmap_sem); |
310 | ||
311 | mm_populate(ctx->mmap_base, populate); | |
1da177e4 | 312 | |
58c85dc2 KO |
313 | pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base); |
314 | ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages, | |
315 | 1, 0, ctx->ring_pages, NULL); | |
36bc08cc GZ |
316 | for (i = 0; i < ctx->nr_pages; i++) |
317 | put_page(ctx->ring_pages[i]); | |
1da177e4 | 318 | |
58c85dc2 | 319 | if (unlikely(ctx->nr_pages != nr_pages)) { |
1da177e4 LT |
320 | aio_free_ring(ctx); |
321 | return -EAGAIN; | |
322 | } | |
323 | ||
58c85dc2 KO |
324 | ctx->user_id = ctx->mmap_base; |
325 | ctx->nr_events = nr_events; /* trusted copy */ | |
1da177e4 | 326 | |
58c85dc2 | 327 | ring = kmap_atomic(ctx->ring_pages[0]); |
1da177e4 LT |
328 | ring->nr = nr_events; /* user copy */ |
329 | ring->id = ctx->user_id; | |
330 | ring->head = ring->tail = 0; | |
331 | ring->magic = AIO_RING_MAGIC; | |
332 | ring->compat_features = AIO_RING_COMPAT_FEATURES; | |
333 | ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; | |
334 | ring->header_length = sizeof(struct aio_ring); | |
e8e3c3d6 | 335 | kunmap_atomic(ring); |
58c85dc2 | 336 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 LT |
337 | |
338 | return 0; | |
339 | } | |
340 | ||
1da177e4 LT |
341 | #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) |
342 | #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) | |
343 | #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) | |
344 | ||
0460fef2 KO |
345 | void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel) |
346 | { | |
347 | struct kioctx *ctx = req->ki_ctx; | |
348 | unsigned long flags; | |
349 | ||
350 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
351 | ||
352 | if (!req->ki_list.next) | |
353 | list_add(&req->ki_list, &ctx->active_reqs); | |
354 | ||
355 | req->ki_cancel = cancel; | |
356 | ||
357 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
358 | } | |
359 | EXPORT_SYMBOL(kiocb_set_cancel_fn); | |
360 | ||
bec68faa | 361 | static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb) |
906b973c | 362 | { |
0460fef2 | 363 | kiocb_cancel_fn *old, *cancel; |
906b973c | 364 | |
0460fef2 KO |
365 | /* |
366 | * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it | |
367 | * actually has a cancel function, hence the cmpxchg() | |
368 | */ | |
369 | ||
370 | cancel = ACCESS_ONCE(kiocb->ki_cancel); | |
371 | do { | |
372 | if (!cancel || cancel == KIOCB_CANCELLED) | |
57282d8f | 373 | return -EINVAL; |
906b973c | 374 | |
0460fef2 KO |
375 | old = cancel; |
376 | cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED); | |
377 | } while (cancel != old); | |
906b973c | 378 | |
57282d8f | 379 | return cancel(kiocb); |
906b973c KO |
380 | } |
381 | ||
36f55889 KO |
382 | static void free_ioctx_rcu(struct rcu_head *head) |
383 | { | |
384 | struct kioctx *ctx = container_of(head, struct kioctx, rcu_head); | |
e1bdd5f2 KO |
385 | |
386 | free_percpu(ctx->cpu); | |
36f55889 KO |
387 | kmem_cache_free(kioctx_cachep, ctx); |
388 | } | |
389 | ||
390 | /* | |
391 | * When this function runs, the kioctx has been removed from the "hash table" | |
392 | * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - | |
393 | * now it's safe to cancel any that need to be. | |
394 | */ | |
723be6e3 | 395 | static void free_ioctx(struct work_struct *work) |
36f55889 | 396 | { |
723be6e3 | 397 | struct kioctx *ctx = container_of(work, struct kioctx, free_work); |
3e845ce0 | 398 | struct aio_ring *ring; |
36f55889 | 399 | struct kiocb *req; |
5ffac122 KO |
400 | unsigned cpu, avail; |
401 | DEFINE_WAIT(wait); | |
36f55889 KO |
402 | |
403 | spin_lock_irq(&ctx->ctx_lock); | |
404 | ||
405 | while (!list_empty(&ctx->active_reqs)) { | |
406 | req = list_first_entry(&ctx->active_reqs, | |
407 | struct kiocb, ki_list); | |
408 | ||
409 | list_del_init(&req->ki_list); | |
bec68faa | 410 | kiocb_cancel(ctx, req); |
36f55889 KO |
411 | } |
412 | ||
413 | spin_unlock_irq(&ctx->ctx_lock); | |
414 | ||
e1bdd5f2 KO |
415 | for_each_possible_cpu(cpu) { |
416 | struct kioctx_cpu *kcpu = per_cpu_ptr(ctx->cpu, cpu); | |
417 | ||
418 | atomic_add(kcpu->reqs_available, &ctx->reqs_available); | |
419 | kcpu->reqs_available = 0; | |
420 | } | |
421 | ||
5ffac122 KO |
422 | while (1) { |
423 | prepare_to_wait(&ctx->wait, &wait, TASK_UNINTERRUPTIBLE); | |
3e845ce0 | 424 | |
5ffac122 KO |
425 | ring = kmap_atomic(ctx->ring_pages[0]); |
426 | avail = (ring->head <= ring->tail) | |
427 | ? ring->tail - ring->head | |
428 | : ctx->nr_events - ring->head + ring->tail; | |
3e845ce0 | 429 | |
34e83fc6 | 430 | atomic_add(avail, &ctx->reqs_available); |
5ffac122 KO |
431 | ring->head = ring->tail; |
432 | kunmap_atomic(ring); | |
433 | ||
434 | if (atomic_read(&ctx->reqs_available) >= ctx->nr_events - 1) | |
435 | break; | |
436 | ||
437 | schedule(); | |
3e845ce0 | 438 | } |
5ffac122 | 439 | finish_wait(&ctx->wait, &wait); |
3e845ce0 | 440 | |
34e83fc6 | 441 | WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr_events - 1); |
36f55889 KO |
442 | |
443 | aio_free_ring(ctx); | |
444 | ||
36f55889 KO |
445 | pr_debug("freeing %p\n", ctx); |
446 | ||
447 | /* | |
448 | * Here the call_rcu() is between the wait_event() for reqs_active to | |
449 | * hit 0, and freeing the ioctx. | |
450 | * | |
451 | * aio_complete() decrements reqs_active, but it has to touch the ioctx | |
452 | * after to issue a wakeup so we use rcu. | |
453 | */ | |
454 | call_rcu(&ctx->rcu_head, free_ioctx_rcu); | |
455 | } | |
456 | ||
723be6e3 | 457 | static void free_ioctx_ref(struct percpu_ref *ref) |
36f55889 | 458 | { |
723be6e3 KO |
459 | struct kioctx *ctx = container_of(ref, struct kioctx, users); |
460 | ||
461 | INIT_WORK(&ctx->free_work, free_ioctx); | |
462 | schedule_work(&ctx->free_work); | |
36f55889 KO |
463 | } |
464 | ||
1da177e4 LT |
465 | /* ioctx_alloc |
466 | * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. | |
467 | */ | |
468 | static struct kioctx *ioctx_alloc(unsigned nr_events) | |
469 | { | |
41003a7b | 470 | struct mm_struct *mm = current->mm; |
1da177e4 | 471 | struct kioctx *ctx; |
e23754f8 | 472 | int err = -ENOMEM; |
1da177e4 | 473 | |
e1bdd5f2 KO |
474 | /* |
475 | * We keep track of the number of available ringbuffer slots, to prevent | |
476 | * overflow (reqs_available), and we also use percpu counters for this. | |
477 | * | |
478 | * So since up to half the slots might be on other cpu's percpu counters | |
479 | * and unavailable, double nr_events so userspace sees what they | |
480 | * expected: additionally, we move req_batch slots to/from percpu | |
481 | * counters at a time, so make sure that isn't 0: | |
482 | */ | |
483 | nr_events = max(nr_events, num_possible_cpus() * 4); | |
484 | nr_events *= 2; | |
485 | ||
1da177e4 LT |
486 | /* Prevent overflows */ |
487 | if ((nr_events > (0x10000000U / sizeof(struct io_event))) || | |
488 | (nr_events > (0x10000000U / sizeof(struct kiocb)))) { | |
489 | pr_debug("ENOMEM: nr_events too high\n"); | |
490 | return ERR_PTR(-EINVAL); | |
491 | } | |
492 | ||
4cd81c3d | 493 | if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL)) |
1da177e4 LT |
494 | return ERR_PTR(-EAGAIN); |
495 | ||
c3762229 | 496 | ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); |
1da177e4 LT |
497 | if (!ctx) |
498 | return ERR_PTR(-ENOMEM); | |
499 | ||
1da177e4 | 500 | ctx->max_reqs = nr_events; |
1da177e4 | 501 | |
723be6e3 KO |
502 | if (percpu_ref_init(&ctx->users, free_ioctx_ref)) |
503 | goto out_freectx; | |
504 | ||
1da177e4 | 505 | spin_lock_init(&ctx->ctx_lock); |
0460fef2 | 506 | spin_lock_init(&ctx->completion_lock); |
58c85dc2 | 507 | mutex_init(&ctx->ring_lock); |
1da177e4 LT |
508 | init_waitqueue_head(&ctx->wait); |
509 | ||
510 | INIT_LIST_HEAD(&ctx->active_reqs); | |
1da177e4 | 511 | |
e1bdd5f2 KO |
512 | ctx->cpu = alloc_percpu(struct kioctx_cpu); |
513 | if (!ctx->cpu) | |
723be6e3 | 514 | goto out_freeref; |
1da177e4 | 515 | |
e1bdd5f2 KO |
516 | if (aio_setup_ring(ctx) < 0) |
517 | goto out_freepcpu; | |
518 | ||
34e83fc6 | 519 | atomic_set(&ctx->reqs_available, ctx->nr_events - 1); |
e1bdd5f2 KO |
520 | ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4); |
521 | BUG_ON(!ctx->req_batch); | |
34e83fc6 | 522 | |
1da177e4 | 523 | /* limit the number of system wide aios */ |
9fa1cb39 | 524 | spin_lock(&aio_nr_lock); |
4cd81c3d | 525 | if (aio_nr + nr_events > (aio_max_nr * 2UL) || |
2dd542b7 | 526 | aio_nr + nr_events < aio_nr) { |
9fa1cb39 | 527 | spin_unlock(&aio_nr_lock); |
1da177e4 | 528 | goto out_cleanup; |
2dd542b7 AV |
529 | } |
530 | aio_nr += ctx->max_reqs; | |
9fa1cb39 | 531 | spin_unlock(&aio_nr_lock); |
1da177e4 | 532 | |
723be6e3 KO |
533 | percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */ |
534 | ||
39fa0031 | 535 | /* now link into global list. */ |
abf137dd JA |
536 | spin_lock(&mm->ioctx_lock); |
537 | hlist_add_head_rcu(&ctx->list, &mm->ioctx_list); | |
538 | spin_unlock(&mm->ioctx_lock); | |
1da177e4 | 539 | |
caf4167a | 540 | pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", |
58c85dc2 | 541 | ctx, ctx->user_id, mm, ctx->nr_events); |
1da177e4 LT |
542 | return ctx; |
543 | ||
544 | out_cleanup: | |
e23754f8 AV |
545 | err = -EAGAIN; |
546 | aio_free_ring(ctx); | |
e1bdd5f2 KO |
547 | out_freepcpu: |
548 | free_percpu(ctx->cpu); | |
723be6e3 KO |
549 | out_freeref: |
550 | free_percpu(ctx->users.pcpu_count); | |
1da177e4 | 551 | out_freectx: |
36bc08cc GZ |
552 | if (ctx->aio_ring_file) |
553 | fput(ctx->aio_ring_file); | |
1da177e4 | 554 | kmem_cache_free(kioctx_cachep, ctx); |
caf4167a | 555 | pr_debug("error allocating ioctx %d\n", err); |
e23754f8 | 556 | return ERR_PTR(err); |
1da177e4 LT |
557 | } |
558 | ||
36f55889 KO |
559 | /* kill_ioctx |
560 | * Cancels all outstanding aio requests on an aio context. Used | |
561 | * when the processes owning a context have all exited to encourage | |
562 | * the rapid destruction of the kioctx. | |
563 | */ | |
564 | static void kill_ioctx(struct kioctx *ctx) | |
565 | { | |
566 | if (!atomic_xchg(&ctx->dead, 1)) { | |
567 | hlist_del_rcu(&ctx->list); | |
723be6e3 KO |
568 | /* percpu_ref_kill() will do the necessary call_rcu() */ |
569 | wake_up_all(&ctx->wait); | |
dee11c23 | 570 | |
36f55889 | 571 | /* |
4fcc712f KO |
572 | * It'd be more correct to do this in free_ioctx(), after all |
573 | * the outstanding kiocbs have finished - but by then io_destroy | |
574 | * has already returned, so io_setup() could potentially return | |
575 | * -EAGAIN with no ioctxs actually in use (as far as userspace | |
576 | * could tell). | |
36f55889 | 577 | */ |
4fcc712f KO |
578 | spin_lock(&aio_nr_lock); |
579 | BUG_ON(aio_nr - ctx->max_reqs > aio_nr); | |
580 | aio_nr -= ctx->max_reqs; | |
581 | spin_unlock(&aio_nr_lock); | |
582 | ||
583 | if (ctx->mmap_size) | |
584 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
585 | ||
723be6e3 | 586 | percpu_ref_kill(&ctx->users); |
36f55889 | 587 | } |
1da177e4 LT |
588 | } |
589 | ||
590 | /* wait_on_sync_kiocb: | |
591 | * Waits on the given sync kiocb to complete. | |
592 | */ | |
57282d8f | 593 | ssize_t wait_on_sync_kiocb(struct kiocb *req) |
1da177e4 | 594 | { |
57282d8f | 595 | while (!req->ki_ctx) { |
1da177e4 | 596 | set_current_state(TASK_UNINTERRUPTIBLE); |
57282d8f | 597 | if (req->ki_ctx) |
1da177e4 | 598 | break; |
41d10da3 | 599 | io_schedule(); |
1da177e4 LT |
600 | } |
601 | __set_current_state(TASK_RUNNING); | |
57282d8f | 602 | return req->ki_user_data; |
1da177e4 | 603 | } |
385773e0 | 604 | EXPORT_SYMBOL(wait_on_sync_kiocb); |
1da177e4 | 605 | |
36f55889 KO |
606 | /* |
607 | * exit_aio: called when the last user of mm goes away. At this point, there is | |
608 | * no way for any new requests to be submited or any of the io_* syscalls to be | |
609 | * called on the context. | |
610 | * | |
611 | * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on | |
612 | * them. | |
1da177e4 | 613 | */ |
fc9b52cd | 614 | void exit_aio(struct mm_struct *mm) |
1da177e4 | 615 | { |
abf137dd | 616 | struct kioctx *ctx; |
36f55889 | 617 | struct hlist_node *n; |
abf137dd | 618 | |
36f55889 | 619 | hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) { |
936af157 AV |
620 | /* |
621 | * We don't need to bother with munmap() here - | |
622 | * exit_mmap(mm) is coming and it'll unmap everything. | |
623 | * Since aio_free_ring() uses non-zero ->mmap_size | |
624 | * as indicator that it needs to unmap the area, | |
625 | * just set it to 0; aio_free_ring() is the only | |
626 | * place that uses ->mmap_size, so it's safe. | |
936af157 | 627 | */ |
58c85dc2 | 628 | ctx->mmap_size = 0; |
36f55889 | 629 | |
4fcc712f | 630 | kill_ioctx(ctx); |
1da177e4 LT |
631 | } |
632 | } | |
633 | ||
e1bdd5f2 KO |
634 | static void put_reqs_available(struct kioctx *ctx, unsigned nr) |
635 | { | |
636 | struct kioctx_cpu *kcpu; | |
637 | ||
638 | preempt_disable(); | |
639 | kcpu = this_cpu_ptr(ctx->cpu); | |
640 | ||
641 | kcpu->reqs_available += nr; | |
642 | while (kcpu->reqs_available >= ctx->req_batch * 2) { | |
643 | kcpu->reqs_available -= ctx->req_batch; | |
644 | atomic_add(ctx->req_batch, &ctx->reqs_available); | |
645 | } | |
646 | ||
647 | preempt_enable(); | |
648 | } | |
649 | ||
650 | static bool get_reqs_available(struct kioctx *ctx) | |
651 | { | |
652 | struct kioctx_cpu *kcpu; | |
653 | bool ret = false; | |
654 | ||
655 | preempt_disable(); | |
656 | kcpu = this_cpu_ptr(ctx->cpu); | |
657 | ||
658 | if (!kcpu->reqs_available) { | |
659 | int old, avail = atomic_read(&ctx->reqs_available); | |
660 | ||
661 | do { | |
662 | if (avail < ctx->req_batch) | |
663 | goto out; | |
664 | ||
665 | old = avail; | |
666 | avail = atomic_cmpxchg(&ctx->reqs_available, | |
667 | avail, avail - ctx->req_batch); | |
668 | } while (avail != old); | |
669 | ||
670 | kcpu->reqs_available += ctx->req_batch; | |
671 | } | |
672 | ||
673 | ret = true; | |
674 | kcpu->reqs_available--; | |
675 | out: | |
676 | preempt_enable(); | |
677 | return ret; | |
678 | } | |
679 | ||
1da177e4 | 680 | /* aio_get_req |
57282d8f KO |
681 | * Allocate a slot for an aio request. |
682 | * Returns NULL if no requests are free. | |
1da177e4 | 683 | */ |
a1c8eae7 | 684 | static inline struct kiocb *aio_get_req(struct kioctx *ctx) |
1da177e4 | 685 | { |
a1c8eae7 KO |
686 | struct kiocb *req; |
687 | ||
e1bdd5f2 | 688 | if (!get_reqs_available(ctx)) |
a1c8eae7 KO |
689 | return NULL; |
690 | ||
0460fef2 | 691 | req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO); |
1da177e4 | 692 | if (unlikely(!req)) |
a1c8eae7 | 693 | goto out_put; |
1da177e4 | 694 | |
1da177e4 | 695 | req->ki_ctx = ctx; |
080d676d | 696 | return req; |
a1c8eae7 | 697 | out_put: |
e1bdd5f2 | 698 | put_reqs_available(ctx, 1); |
a1c8eae7 | 699 | return NULL; |
1da177e4 LT |
700 | } |
701 | ||
11599eba | 702 | static void kiocb_free(struct kiocb *req) |
1da177e4 | 703 | { |
1d98ebfc KO |
704 | if (req->ki_filp) |
705 | fput(req->ki_filp); | |
13389010 DL |
706 | if (req->ki_eventfd != NULL) |
707 | eventfd_ctx_put(req->ki_eventfd); | |
1da177e4 | 708 | kmem_cache_free(kiocb_cachep, req); |
1da177e4 LT |
709 | } |
710 | ||
d5470b59 | 711 | static struct kioctx *lookup_ioctx(unsigned long ctx_id) |
1da177e4 | 712 | { |
abf137dd | 713 | struct mm_struct *mm = current->mm; |
65c24491 | 714 | struct kioctx *ctx, *ret = NULL; |
1da177e4 | 715 | |
abf137dd JA |
716 | rcu_read_lock(); |
717 | ||
b67bfe0d | 718 | hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) { |
36f55889 | 719 | if (ctx->user_id == ctx_id) { |
723be6e3 | 720 | percpu_ref_get(&ctx->users); |
65c24491 | 721 | ret = ctx; |
1da177e4 LT |
722 | break; |
723 | } | |
abf137dd | 724 | } |
1da177e4 | 725 | |
abf137dd | 726 | rcu_read_unlock(); |
65c24491 | 727 | return ret; |
1da177e4 LT |
728 | } |
729 | ||
1da177e4 LT |
730 | /* aio_complete |
731 | * Called when the io request on the given iocb is complete. | |
1da177e4 | 732 | */ |
2d68449e | 733 | void aio_complete(struct kiocb *iocb, long res, long res2) |
1da177e4 LT |
734 | { |
735 | struct kioctx *ctx = iocb->ki_ctx; | |
1da177e4 | 736 | struct aio_ring *ring; |
21b40200 | 737 | struct io_event *ev_page, *event; |
1da177e4 | 738 | unsigned long flags; |
21b40200 | 739 | unsigned tail, pos; |
1da177e4 | 740 | |
20dcae32 ZB |
741 | /* |
742 | * Special case handling for sync iocbs: | |
743 | * - events go directly into the iocb for fast handling | |
744 | * - the sync task with the iocb in its stack holds the single iocb | |
745 | * ref, no other paths have a way to get another ref | |
746 | * - the sync task helpfully left a reference to itself in the iocb | |
1da177e4 LT |
747 | */ |
748 | if (is_sync_kiocb(iocb)) { | |
1da177e4 | 749 | iocb->ki_user_data = res; |
57282d8f KO |
750 | smp_wmb(); |
751 | iocb->ki_ctx = ERR_PTR(-EXDEV); | |
1da177e4 | 752 | wake_up_process(iocb->ki_obj.tsk); |
2d68449e | 753 | return; |
1da177e4 LT |
754 | } |
755 | ||
36f55889 | 756 | /* |
36f55889 | 757 | * Take rcu_read_lock() in case the kioctx is being destroyed, as we |
34e83fc6 | 758 | * need to issue a wakeup after incrementing reqs_available. |
1da177e4 | 759 | */ |
36f55889 | 760 | rcu_read_lock(); |
1da177e4 | 761 | |
0460fef2 KO |
762 | if (iocb->ki_list.next) { |
763 | unsigned long flags; | |
764 | ||
765 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
766 | list_del(&iocb->ki_list); | |
767 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
768 | } | |
11599eba | 769 | |
0460fef2 KO |
770 | /* |
771 | * Add a completion event to the ring buffer. Must be done holding | |
4b30f07e | 772 | * ctx->completion_lock to prevent other code from messing with the tail |
0460fef2 KO |
773 | * pointer since we might be called from irq context. |
774 | */ | |
775 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
776 | ||
58c85dc2 | 777 | tail = ctx->tail; |
21b40200 KO |
778 | pos = tail + AIO_EVENTS_OFFSET; |
779 | ||
58c85dc2 | 780 | if (++tail >= ctx->nr_events) |
4bf69b2a | 781 | tail = 0; |
1da177e4 | 782 | |
58c85dc2 | 783 | ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
784 | event = ev_page + pos % AIO_EVENTS_PER_PAGE; |
785 | ||
1da177e4 LT |
786 | event->obj = (u64)(unsigned long)iocb->ki_obj.user; |
787 | event->data = iocb->ki_user_data; | |
788 | event->res = res; | |
789 | event->res2 = res2; | |
790 | ||
21b40200 | 791 | kunmap_atomic(ev_page); |
58c85dc2 | 792 | flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
793 | |
794 | pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n", | |
caf4167a KO |
795 | ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, |
796 | res, res2); | |
1da177e4 LT |
797 | |
798 | /* after flagging the request as done, we | |
799 | * must never even look at it again | |
800 | */ | |
801 | smp_wmb(); /* make event visible before updating tail */ | |
802 | ||
58c85dc2 | 803 | ctx->tail = tail; |
1da177e4 | 804 | |
58c85dc2 | 805 | ring = kmap_atomic(ctx->ring_pages[0]); |
21b40200 | 806 | ring->tail = tail; |
e8e3c3d6 | 807 | kunmap_atomic(ring); |
58c85dc2 | 808 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 | 809 | |
0460fef2 KO |
810 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
811 | ||
21b40200 | 812 | pr_debug("added to ring %p at [%u]\n", iocb, tail); |
8d1c98b0 DL |
813 | |
814 | /* | |
815 | * Check if the user asked us to deliver the result through an | |
816 | * eventfd. The eventfd_signal() function is safe to be called | |
817 | * from IRQ context. | |
818 | */ | |
87c3a86e | 819 | if (iocb->ki_eventfd != NULL) |
8d1c98b0 DL |
820 | eventfd_signal(iocb->ki_eventfd, 1); |
821 | ||
1da177e4 | 822 | /* everything turned out well, dispose of the aiocb. */ |
57282d8f | 823 | kiocb_free(iocb); |
1da177e4 | 824 | |
6cb2a210 QB |
825 | /* |
826 | * We have to order our ring_info tail store above and test | |
827 | * of the wait list below outside the wait lock. This is | |
828 | * like in wake_up_bit() where clearing a bit has to be | |
829 | * ordered with the unlocked test. | |
830 | */ | |
831 | smp_mb(); | |
832 | ||
1da177e4 LT |
833 | if (waitqueue_active(&ctx->wait)) |
834 | wake_up(&ctx->wait); | |
835 | ||
36f55889 | 836 | rcu_read_unlock(); |
1da177e4 | 837 | } |
385773e0 | 838 | EXPORT_SYMBOL(aio_complete); |
1da177e4 | 839 | |
a31ad380 KO |
840 | /* aio_read_events |
841 | * Pull an event off of the ioctx's event ring. Returns the number of | |
842 | * events fetched | |
1da177e4 | 843 | */ |
a31ad380 KO |
844 | static long aio_read_events_ring(struct kioctx *ctx, |
845 | struct io_event __user *event, long nr) | |
1da177e4 | 846 | { |
1da177e4 | 847 | struct aio_ring *ring; |
5ffac122 | 848 | unsigned head, tail, pos; |
a31ad380 KO |
849 | long ret = 0; |
850 | int copy_ret; | |
851 | ||
58c85dc2 | 852 | mutex_lock(&ctx->ring_lock); |
1da177e4 | 853 | |
58c85dc2 | 854 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 855 | head = ring->head; |
5ffac122 | 856 | tail = ring->tail; |
a31ad380 KO |
857 | kunmap_atomic(ring); |
858 | ||
5ffac122 | 859 | pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events); |
1da177e4 | 860 | |
5ffac122 | 861 | if (head == tail) |
1da177e4 LT |
862 | goto out; |
863 | ||
a31ad380 KO |
864 | while (ret < nr) { |
865 | long avail; | |
866 | struct io_event *ev; | |
867 | struct page *page; | |
868 | ||
5ffac122 KO |
869 | avail = (head <= tail ? tail : ctx->nr_events) - head; |
870 | if (head == tail) | |
a31ad380 KO |
871 | break; |
872 | ||
873 | avail = min(avail, nr - ret); | |
874 | avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - | |
875 | ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE)); | |
876 | ||
877 | pos = head + AIO_EVENTS_OFFSET; | |
58c85dc2 | 878 | page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]; |
a31ad380 KO |
879 | pos %= AIO_EVENTS_PER_PAGE; |
880 | ||
881 | ev = kmap(page); | |
882 | copy_ret = copy_to_user(event + ret, ev + pos, | |
883 | sizeof(*ev) * avail); | |
884 | kunmap(page); | |
885 | ||
886 | if (unlikely(copy_ret)) { | |
887 | ret = -EFAULT; | |
888 | goto out; | |
889 | } | |
890 | ||
891 | ret += avail; | |
892 | head += avail; | |
58c85dc2 | 893 | head %= ctx->nr_events; |
1da177e4 | 894 | } |
1da177e4 | 895 | |
58c85dc2 | 896 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 897 | ring->head = head; |
91d80a84 | 898 | kunmap_atomic(ring); |
58c85dc2 | 899 | flush_dcache_page(ctx->ring_pages[0]); |
a31ad380 | 900 | |
5ffac122 | 901 | pr_debug("%li h%u t%u\n", ret, head, tail); |
3e845ce0 | 902 | |
e1bdd5f2 | 903 | put_reqs_available(ctx, ret); |
a31ad380 | 904 | out: |
58c85dc2 | 905 | mutex_unlock(&ctx->ring_lock); |
a31ad380 | 906 | |
1da177e4 LT |
907 | return ret; |
908 | } | |
909 | ||
a31ad380 KO |
910 | static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr, |
911 | struct io_event __user *event, long *i) | |
1da177e4 | 912 | { |
a31ad380 | 913 | long ret = aio_read_events_ring(ctx, event + *i, nr - *i); |
1da177e4 | 914 | |
a31ad380 KO |
915 | if (ret > 0) |
916 | *i += ret; | |
1da177e4 | 917 | |
a31ad380 KO |
918 | if (unlikely(atomic_read(&ctx->dead))) |
919 | ret = -EINVAL; | |
1da177e4 | 920 | |
a31ad380 KO |
921 | if (!*i) |
922 | *i = ret; | |
1da177e4 | 923 | |
a31ad380 | 924 | return ret < 0 || *i >= min_nr; |
1da177e4 LT |
925 | } |
926 | ||
a31ad380 | 927 | static long read_events(struct kioctx *ctx, long min_nr, long nr, |
1da177e4 LT |
928 | struct io_event __user *event, |
929 | struct timespec __user *timeout) | |
930 | { | |
a31ad380 KO |
931 | ktime_t until = { .tv64 = KTIME_MAX }; |
932 | long ret = 0; | |
1da177e4 | 933 | |
1da177e4 LT |
934 | if (timeout) { |
935 | struct timespec ts; | |
a31ad380 | 936 | |
1da177e4 | 937 | if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) |
a31ad380 | 938 | return -EFAULT; |
1da177e4 | 939 | |
a31ad380 | 940 | until = timespec_to_ktime(ts); |
1da177e4 LT |
941 | } |
942 | ||
a31ad380 KO |
943 | /* |
944 | * Note that aio_read_events() is being called as the conditional - i.e. | |
945 | * we're calling it after prepare_to_wait() has set task state to | |
946 | * TASK_INTERRUPTIBLE. | |
947 | * | |
948 | * But aio_read_events() can block, and if it blocks it's going to flip | |
949 | * the task state back to TASK_RUNNING. | |
950 | * | |
951 | * This should be ok, provided it doesn't flip the state back to | |
952 | * TASK_RUNNING and return 0 too much - that causes us to spin. That | |
953 | * will only happen if the mutex_lock() call blocks, and we then find | |
954 | * the ringbuffer empty. So in practice we should be ok, but it's | |
955 | * something to be aware of when touching this code. | |
956 | */ | |
957 | wait_event_interruptible_hrtimeout(ctx->wait, | |
958 | aio_read_events(ctx, min_nr, nr, event, &ret), until); | |
1da177e4 | 959 | |
a31ad380 KO |
960 | if (!ret && signal_pending(current)) |
961 | ret = -EINTR; | |
1da177e4 | 962 | |
a31ad380 | 963 | return ret; |
1da177e4 LT |
964 | } |
965 | ||
1da177e4 LT |
966 | /* sys_io_setup: |
967 | * Create an aio_context capable of receiving at least nr_events. | |
968 | * ctxp must not point to an aio_context that already exists, and | |
969 | * must be initialized to 0 prior to the call. On successful | |
970 | * creation of the aio_context, *ctxp is filled in with the resulting | |
971 | * handle. May fail with -EINVAL if *ctxp is not initialized, | |
972 | * if the specified nr_events exceeds internal limits. May fail | |
973 | * with -EAGAIN if the specified nr_events exceeds the user's limit | |
974 | * of available events. May fail with -ENOMEM if insufficient kernel | |
975 | * resources are available. May fail with -EFAULT if an invalid | |
976 | * pointer is passed for ctxp. Will fail with -ENOSYS if not | |
977 | * implemented. | |
978 | */ | |
002c8976 | 979 | SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) |
1da177e4 LT |
980 | { |
981 | struct kioctx *ioctx = NULL; | |
982 | unsigned long ctx; | |
983 | long ret; | |
984 | ||
985 | ret = get_user(ctx, ctxp); | |
986 | if (unlikely(ret)) | |
987 | goto out; | |
988 | ||
989 | ret = -EINVAL; | |
d55b5fda ZB |
990 | if (unlikely(ctx || nr_events == 0)) { |
991 | pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", | |
992 | ctx, nr_events); | |
1da177e4 LT |
993 | goto out; |
994 | } | |
995 | ||
996 | ioctx = ioctx_alloc(nr_events); | |
997 | ret = PTR_ERR(ioctx); | |
998 | if (!IS_ERR(ioctx)) { | |
999 | ret = put_user(ioctx->user_id, ctxp); | |
a2e1859a | 1000 | if (ret) |
36f55889 | 1001 | kill_ioctx(ioctx); |
723be6e3 | 1002 | percpu_ref_put(&ioctx->users); |
1da177e4 LT |
1003 | } |
1004 | ||
1005 | out: | |
1006 | return ret; | |
1007 | } | |
1008 | ||
1009 | /* sys_io_destroy: | |
1010 | * Destroy the aio_context specified. May cancel any outstanding | |
1011 | * AIOs and block on completion. Will fail with -ENOSYS if not | |
642b5123 | 1012 | * implemented. May fail with -EINVAL if the context pointed to |
1da177e4 LT |
1013 | * is invalid. |
1014 | */ | |
002c8976 | 1015 | SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) |
1da177e4 LT |
1016 | { |
1017 | struct kioctx *ioctx = lookup_ioctx(ctx); | |
1018 | if (likely(NULL != ioctx)) { | |
36f55889 | 1019 | kill_ioctx(ioctx); |
723be6e3 | 1020 | percpu_ref_put(&ioctx->users); |
1da177e4 LT |
1021 | return 0; |
1022 | } | |
1023 | pr_debug("EINVAL: io_destroy: invalid context id\n"); | |
1024 | return -EINVAL; | |
1025 | } | |
1026 | ||
41ef4eb8 KO |
1027 | typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *, |
1028 | unsigned long, loff_t); | |
1029 | ||
8bc92afc KO |
1030 | static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb, |
1031 | int rw, char __user *buf, | |
1032 | unsigned long *nr_segs, | |
1033 | struct iovec **iovec, | |
1034 | bool compat) | |
eed4e51f BP |
1035 | { |
1036 | ssize_t ret; | |
1037 | ||
8bc92afc | 1038 | *nr_segs = kiocb->ki_nbytes; |
41ef4eb8 | 1039 | |
9d85cba7 JM |
1040 | #ifdef CONFIG_COMPAT |
1041 | if (compat) | |
41ef4eb8 | 1042 | ret = compat_rw_copy_check_uvector(rw, |
8bc92afc KO |
1043 | (struct compat_iovec __user *)buf, |
1044 | *nr_segs, 1, *iovec, iovec); | |
9d85cba7 JM |
1045 | else |
1046 | #endif | |
41ef4eb8 | 1047 | ret = rw_copy_check_uvector(rw, |
8bc92afc KO |
1048 | (struct iovec __user *)buf, |
1049 | *nr_segs, 1, *iovec, iovec); | |
eed4e51f | 1050 | if (ret < 0) |
41ef4eb8 | 1051 | return ret; |
a70b52ec | 1052 | |
41ef4eb8 | 1053 | /* ki_nbytes now reflect bytes instead of segs */ |
eed4e51f | 1054 | kiocb->ki_nbytes = ret; |
41ef4eb8 | 1055 | return 0; |
eed4e51f BP |
1056 | } |
1057 | ||
8bc92afc KO |
1058 | static ssize_t aio_setup_single_vector(struct kiocb *kiocb, |
1059 | int rw, char __user *buf, | |
1060 | unsigned long *nr_segs, | |
1061 | struct iovec *iovec) | |
eed4e51f | 1062 | { |
8bc92afc | 1063 | if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes))) |
41ef4eb8 | 1064 | return -EFAULT; |
a70b52ec | 1065 | |
8bc92afc KO |
1066 | iovec->iov_base = buf; |
1067 | iovec->iov_len = kiocb->ki_nbytes; | |
1068 | *nr_segs = 1; | |
eed4e51f BP |
1069 | return 0; |
1070 | } | |
1071 | ||
1da177e4 LT |
1072 | /* |
1073 | * aio_setup_iocb: | |
1074 | * Performs the initial checks and aio retry method | |
1075 | * setup for the kiocb at the time of io submission. | |
1076 | */ | |
8bc92afc KO |
1077 | static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode, |
1078 | char __user *buf, bool compat) | |
1da177e4 | 1079 | { |
41ef4eb8 KO |
1080 | struct file *file = req->ki_filp; |
1081 | ssize_t ret; | |
8bc92afc | 1082 | unsigned long nr_segs; |
41ef4eb8 KO |
1083 | int rw; |
1084 | fmode_t mode; | |
1085 | aio_rw_op *rw_op; | |
8bc92afc | 1086 | struct iovec inline_vec, *iovec = &inline_vec; |
1da177e4 | 1087 | |
8bc92afc | 1088 | switch (opcode) { |
1da177e4 | 1089 | case IOCB_CMD_PREAD: |
eed4e51f | 1090 | case IOCB_CMD_PREADV: |
41ef4eb8 KO |
1091 | mode = FMODE_READ; |
1092 | rw = READ; | |
1093 | rw_op = file->f_op->aio_read; | |
1094 | goto rw_common; | |
1095 | ||
1096 | case IOCB_CMD_PWRITE: | |
eed4e51f | 1097 | case IOCB_CMD_PWRITEV: |
41ef4eb8 KO |
1098 | mode = FMODE_WRITE; |
1099 | rw = WRITE; | |
1100 | rw_op = file->f_op->aio_write; | |
1101 | goto rw_common; | |
1102 | rw_common: | |
1103 | if (unlikely(!(file->f_mode & mode))) | |
1104 | return -EBADF; | |
1105 | ||
1106 | if (!rw_op) | |
1107 | return -EINVAL; | |
1108 | ||
8bc92afc KO |
1109 | ret = (opcode == IOCB_CMD_PREADV || |
1110 | opcode == IOCB_CMD_PWRITEV) | |
1111 | ? aio_setup_vectored_rw(req, rw, buf, &nr_segs, | |
1112 | &iovec, compat) | |
1113 | : aio_setup_single_vector(req, rw, buf, &nr_segs, | |
1114 | iovec); | |
eed4e51f | 1115 | if (ret) |
41ef4eb8 KO |
1116 | return ret; |
1117 | ||
1118 | ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes); | |
8bc92afc KO |
1119 | if (ret < 0) { |
1120 | if (iovec != &inline_vec) | |
1121 | kfree(iovec); | |
41ef4eb8 | 1122 | return ret; |
8bc92afc | 1123 | } |
41ef4eb8 KO |
1124 | |
1125 | req->ki_nbytes = ret; | |
41ef4eb8 | 1126 | |
73a7075e KO |
1127 | /* XXX: move/kill - rw_verify_area()? */ |
1128 | /* This matches the pread()/pwrite() logic */ | |
1129 | if (req->ki_pos < 0) { | |
1130 | ret = -EINVAL; | |
1131 | break; | |
1132 | } | |
1133 | ||
1134 | if (rw == WRITE) | |
1135 | file_start_write(file); | |
1136 | ||
8bc92afc | 1137 | ret = rw_op(req, iovec, nr_segs, req->ki_pos); |
73a7075e KO |
1138 | |
1139 | if (rw == WRITE) | |
1140 | file_end_write(file); | |
1da177e4 | 1141 | break; |
41ef4eb8 | 1142 | |
1da177e4 | 1143 | case IOCB_CMD_FDSYNC: |
41ef4eb8 KO |
1144 | if (!file->f_op->aio_fsync) |
1145 | return -EINVAL; | |
1146 | ||
1147 | ret = file->f_op->aio_fsync(req, 1); | |
1da177e4 | 1148 | break; |
41ef4eb8 | 1149 | |
1da177e4 | 1150 | case IOCB_CMD_FSYNC: |
41ef4eb8 KO |
1151 | if (!file->f_op->aio_fsync) |
1152 | return -EINVAL; | |
1153 | ||
1154 | ret = file->f_op->aio_fsync(req, 0); | |
1da177e4 | 1155 | break; |
41ef4eb8 | 1156 | |
1da177e4 | 1157 | default: |
caf4167a | 1158 | pr_debug("EINVAL: no operation provided\n"); |
41ef4eb8 | 1159 | return -EINVAL; |
1da177e4 LT |
1160 | } |
1161 | ||
8bc92afc KO |
1162 | if (iovec != &inline_vec) |
1163 | kfree(iovec); | |
1164 | ||
41ef4eb8 KO |
1165 | if (ret != -EIOCBQUEUED) { |
1166 | /* | |
1167 | * There's no easy way to restart the syscall since other AIO's | |
1168 | * may be already running. Just fail this IO with EINTR. | |
1169 | */ | |
1170 | if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR || | |
1171 | ret == -ERESTARTNOHAND || | |
1172 | ret == -ERESTART_RESTARTBLOCK)) | |
1173 | ret = -EINTR; | |
1174 | aio_complete(req, ret, 0); | |
1175 | } | |
1da177e4 LT |
1176 | |
1177 | return 0; | |
1178 | } | |
1179 | ||
d5470b59 | 1180 | static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, |
a1c8eae7 | 1181 | struct iocb *iocb, bool compat) |
1da177e4 LT |
1182 | { |
1183 | struct kiocb *req; | |
1da177e4 LT |
1184 | ssize_t ret; |
1185 | ||
1186 | /* enforce forwards compatibility on users */ | |
9c3060be | 1187 | if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { |
caf4167a | 1188 | pr_debug("EINVAL: reserve field set\n"); |
1da177e4 LT |
1189 | return -EINVAL; |
1190 | } | |
1191 | ||
1192 | /* prevent overflows */ | |
1193 | if (unlikely( | |
1194 | (iocb->aio_buf != (unsigned long)iocb->aio_buf) || | |
1195 | (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || | |
1196 | ((ssize_t)iocb->aio_nbytes < 0) | |
1197 | )) { | |
1198 | pr_debug("EINVAL: io_submit: overflow check\n"); | |
1199 | return -EINVAL; | |
1200 | } | |
1201 | ||
41ef4eb8 | 1202 | req = aio_get_req(ctx); |
1d98ebfc | 1203 | if (unlikely(!req)) |
1da177e4 | 1204 | return -EAGAIN; |
1d98ebfc KO |
1205 | |
1206 | req->ki_filp = fget(iocb->aio_fildes); | |
1207 | if (unlikely(!req->ki_filp)) { | |
1208 | ret = -EBADF; | |
1209 | goto out_put_req; | |
1da177e4 | 1210 | } |
1d98ebfc | 1211 | |
9c3060be DL |
1212 | if (iocb->aio_flags & IOCB_FLAG_RESFD) { |
1213 | /* | |
1214 | * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an | |
1215 | * instance of the file* now. The file descriptor must be | |
1216 | * an eventfd() fd, and will be signaled for each completed | |
1217 | * event using the eventfd_signal() function. | |
1218 | */ | |
13389010 | 1219 | req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); |
801678c5 | 1220 | if (IS_ERR(req->ki_eventfd)) { |
9c3060be | 1221 | ret = PTR_ERR(req->ki_eventfd); |
87c3a86e | 1222 | req->ki_eventfd = NULL; |
9c3060be DL |
1223 | goto out_put_req; |
1224 | } | |
1225 | } | |
1da177e4 | 1226 | |
8a660890 | 1227 | ret = put_user(KIOCB_KEY, &user_iocb->aio_key); |
1da177e4 | 1228 | if (unlikely(ret)) { |
caf4167a | 1229 | pr_debug("EFAULT: aio_key\n"); |
1da177e4 LT |
1230 | goto out_put_req; |
1231 | } | |
1232 | ||
1233 | req->ki_obj.user = user_iocb; | |
1234 | req->ki_user_data = iocb->aio_data; | |
1235 | req->ki_pos = iocb->aio_offset; | |
73a7075e | 1236 | req->ki_nbytes = iocb->aio_nbytes; |
1da177e4 | 1237 | |
8bc92afc KO |
1238 | ret = aio_run_iocb(req, iocb->aio_lio_opcode, |
1239 | (char __user *)(unsigned long)iocb->aio_buf, | |
1240 | compat); | |
41003a7b | 1241 | if (ret) |
7137c6bd | 1242 | goto out_put_req; |
41003a7b | 1243 | |
1da177e4 | 1244 | return 0; |
1da177e4 | 1245 | out_put_req: |
e1bdd5f2 | 1246 | put_reqs_available(ctx, 1); |
57282d8f | 1247 | kiocb_free(req); |
1da177e4 LT |
1248 | return ret; |
1249 | } | |
1250 | ||
9d85cba7 JM |
1251 | long do_io_submit(aio_context_t ctx_id, long nr, |
1252 | struct iocb __user *__user *iocbpp, bool compat) | |
1da177e4 LT |
1253 | { |
1254 | struct kioctx *ctx; | |
1255 | long ret = 0; | |
080d676d | 1256 | int i = 0; |
9f5b9425 | 1257 | struct blk_plug plug; |
1da177e4 LT |
1258 | |
1259 | if (unlikely(nr < 0)) | |
1260 | return -EINVAL; | |
1261 | ||
75e1c70f JM |
1262 | if (unlikely(nr > LONG_MAX/sizeof(*iocbpp))) |
1263 | nr = LONG_MAX/sizeof(*iocbpp); | |
1264 | ||
1da177e4 LT |
1265 | if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) |
1266 | return -EFAULT; | |
1267 | ||
1268 | ctx = lookup_ioctx(ctx_id); | |
1269 | if (unlikely(!ctx)) { | |
caf4167a | 1270 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1271 | return -EINVAL; |
1272 | } | |
1273 | ||
9f5b9425 SL |
1274 | blk_start_plug(&plug); |
1275 | ||
1da177e4 LT |
1276 | /* |
1277 | * AKPM: should this return a partial result if some of the IOs were | |
1278 | * successfully submitted? | |
1279 | */ | |
1280 | for (i=0; i<nr; i++) { | |
1281 | struct iocb __user *user_iocb; | |
1282 | struct iocb tmp; | |
1283 | ||
1284 | if (unlikely(__get_user(user_iocb, iocbpp + i))) { | |
1285 | ret = -EFAULT; | |
1286 | break; | |
1287 | } | |
1288 | ||
1289 | if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { | |
1290 | ret = -EFAULT; | |
1291 | break; | |
1292 | } | |
1293 | ||
a1c8eae7 | 1294 | ret = io_submit_one(ctx, user_iocb, &tmp, compat); |
1da177e4 LT |
1295 | if (ret) |
1296 | break; | |
1297 | } | |
9f5b9425 | 1298 | blk_finish_plug(&plug); |
1da177e4 | 1299 | |
723be6e3 | 1300 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1301 | return i ? i : ret; |
1302 | } | |
1303 | ||
9d85cba7 JM |
1304 | /* sys_io_submit: |
1305 | * Queue the nr iocbs pointed to by iocbpp for processing. Returns | |
1306 | * the number of iocbs queued. May return -EINVAL if the aio_context | |
1307 | * specified by ctx_id is invalid, if nr is < 0, if the iocb at | |
1308 | * *iocbpp[0] is not properly initialized, if the operation specified | |
1309 | * is invalid for the file descriptor in the iocb. May fail with | |
1310 | * -EFAULT if any of the data structures point to invalid data. May | |
1311 | * fail with -EBADF if the file descriptor specified in the first | |
1312 | * iocb is invalid. May fail with -EAGAIN if insufficient resources | |
1313 | * are available to queue any iocbs. Will return 0 if nr is 0. Will | |
1314 | * fail with -ENOSYS if not implemented. | |
1315 | */ | |
1316 | SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, | |
1317 | struct iocb __user * __user *, iocbpp) | |
1318 | { | |
1319 | return do_io_submit(ctx_id, nr, iocbpp, 0); | |
1320 | } | |
1321 | ||
1da177e4 LT |
1322 | /* lookup_kiocb |
1323 | * Finds a given iocb for cancellation. | |
1da177e4 | 1324 | */ |
25ee7e38 AB |
1325 | static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, |
1326 | u32 key) | |
1da177e4 LT |
1327 | { |
1328 | struct list_head *pos; | |
d00689af ZB |
1329 | |
1330 | assert_spin_locked(&ctx->ctx_lock); | |
1331 | ||
8a660890 KO |
1332 | if (key != KIOCB_KEY) |
1333 | return NULL; | |
1334 | ||
1da177e4 LT |
1335 | /* TODO: use a hash or array, this sucks. */ |
1336 | list_for_each(pos, &ctx->active_reqs) { | |
1337 | struct kiocb *kiocb = list_kiocb(pos); | |
8a660890 | 1338 | if (kiocb->ki_obj.user == iocb) |
1da177e4 LT |
1339 | return kiocb; |
1340 | } | |
1341 | return NULL; | |
1342 | } | |
1343 | ||
1344 | /* sys_io_cancel: | |
1345 | * Attempts to cancel an iocb previously passed to io_submit. If | |
1346 | * the operation is successfully cancelled, the resulting event is | |
1347 | * copied into the memory pointed to by result without being placed | |
1348 | * into the completion queue and 0 is returned. May fail with | |
1349 | * -EFAULT if any of the data structures pointed to are invalid. | |
1350 | * May fail with -EINVAL if aio_context specified by ctx_id is | |
1351 | * invalid. May fail with -EAGAIN if the iocb specified was not | |
1352 | * cancelled. Will fail with -ENOSYS if not implemented. | |
1353 | */ | |
002c8976 HC |
1354 | SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, |
1355 | struct io_event __user *, result) | |
1da177e4 | 1356 | { |
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 | 1373 | if (kiocb) |
bec68faa | 1374 | ret = kiocb_cancel(ctx, kiocb); |
906b973c KO |
1375 | else |
1376 | ret = -EINVAL; | |
1377 | ||
1da177e4 LT |
1378 | spin_unlock_irq(&ctx->ctx_lock); |
1379 | ||
906b973c | 1380 | if (!ret) { |
bec68faa KO |
1381 | /* |
1382 | * The result argument is no longer used - the io_event is | |
1383 | * always delivered via the ring buffer. -EINPROGRESS indicates | |
1384 | * cancellation is progress: | |
906b973c | 1385 | */ |
bec68faa | 1386 | ret = -EINPROGRESS; |
906b973c | 1387 | } |
1da177e4 | 1388 | |
723be6e3 | 1389 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1390 | |
1391 | return ret; | |
1392 | } | |
1393 | ||
1394 | /* io_getevents: | |
1395 | * Attempts to read at least min_nr events and up to nr events from | |
642b5123 ST |
1396 | * the completion queue for the aio_context specified by ctx_id. If |
1397 | * it succeeds, the number of read events is returned. May fail with | |
1398 | * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is | |
1399 | * out of range, if timeout is out of range. May fail with -EFAULT | |
1400 | * if any of the memory specified is invalid. May return 0 or | |
1401 | * < min_nr if the timeout specified by timeout has elapsed | |
1402 | * before sufficient events are available, where timeout == NULL | |
1403 | * specifies an infinite timeout. Note that the timeout pointed to by | |
6900807c | 1404 | * timeout is relative. Will fail with -ENOSYS if not implemented. |
1da177e4 | 1405 | */ |
002c8976 HC |
1406 | SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, |
1407 | long, min_nr, | |
1408 | long, nr, | |
1409 | struct io_event __user *, events, | |
1410 | struct timespec __user *, timeout) | |
1da177e4 LT |
1411 | { |
1412 | struct kioctx *ioctx = lookup_ioctx(ctx_id); | |
1413 | long ret = -EINVAL; | |
1414 | ||
1415 | if (likely(ioctx)) { | |
2e410255 | 1416 | if (likely(min_nr <= nr && min_nr >= 0)) |
1da177e4 | 1417 | ret = read_events(ioctx, min_nr, nr, events, timeout); |
723be6e3 | 1418 | percpu_ref_put(&ioctx->users); |
1da177e4 | 1419 | } |
1da177e4 LT |
1420 | return ret; |
1421 | } |