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 | |
174cd4b1 | 23 | #include <linux/sched/signal.h> |
1da177e4 LT |
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/migrate.h> |
40 | #include <linux/ramfs.h> | |
723be6e3 | 41 | #include <linux/percpu-refcount.h> |
71ad7490 | 42 | #include <linux/mount.h> |
1da177e4 LT |
43 | |
44 | #include <asm/kmap_types.h> | |
7c0f6ba6 | 45 | #include <linux/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 */ | |
fa8a53c3 BL |
55 | unsigned head; /* Written to by userland or under ring_lock |
56 | * mutex by aio_read_events_ring(). */ | |
4e179bca KO |
57 | unsigned tail; |
58 | ||
59 | unsigned magic; | |
60 | unsigned compat_features; | |
61 | unsigned incompat_features; | |
62 | unsigned header_length; /* size of aio_ring */ | |
63 | ||
64 | ||
65 | struct io_event io_events[0]; | |
66 | }; /* 128 bytes + ring size */ | |
67 | ||
68 | #define AIO_RING_PAGES 8 | |
4e179bca | 69 | |
db446a08 | 70 | struct kioctx_table { |
d0264c01 TH |
71 | struct rcu_head rcu; |
72 | unsigned nr; | |
73 | struct kioctx __rcu *table[]; | |
db446a08 BL |
74 | }; |
75 | ||
e1bdd5f2 KO |
76 | struct kioctx_cpu { |
77 | unsigned reqs_available; | |
78 | }; | |
79 | ||
dc48e56d JA |
80 | struct ctx_rq_wait { |
81 | struct completion comp; | |
82 | atomic_t count; | |
83 | }; | |
84 | ||
4e179bca | 85 | struct kioctx { |
723be6e3 | 86 | struct percpu_ref users; |
36f55889 | 87 | atomic_t dead; |
4e179bca | 88 | |
e34ecee2 KO |
89 | struct percpu_ref reqs; |
90 | ||
4e179bca | 91 | unsigned long user_id; |
4e179bca | 92 | |
e1bdd5f2 KO |
93 | struct __percpu kioctx_cpu *cpu; |
94 | ||
95 | /* | |
96 | * For percpu reqs_available, number of slots we move to/from global | |
97 | * counter at a time: | |
98 | */ | |
99 | unsigned req_batch; | |
3e845ce0 KO |
100 | /* |
101 | * This is what userspace passed to io_setup(), it's not used for | |
102 | * anything but counting against the global max_reqs quota. | |
103 | * | |
58c85dc2 | 104 | * The real limit is nr_events - 1, which will be larger (see |
3e845ce0 KO |
105 | * aio_setup_ring()) |
106 | */ | |
4e179bca KO |
107 | unsigned max_reqs; |
108 | ||
58c85dc2 KO |
109 | /* Size of ringbuffer, in units of struct io_event */ |
110 | unsigned nr_events; | |
4e179bca | 111 | |
58c85dc2 KO |
112 | unsigned long mmap_base; |
113 | unsigned long mmap_size; | |
114 | ||
115 | struct page **ring_pages; | |
116 | long nr_pages; | |
117 | ||
a6d7cff4 TH |
118 | struct rcu_head free_rcu; |
119 | struct work_struct free_work; /* see free_ioctx() */ | |
4e23bcae | 120 | |
e02ba72a AP |
121 | /* |
122 | * signals when all in-flight requests are done | |
123 | */ | |
dc48e56d | 124 | struct ctx_rq_wait *rq_wait; |
e02ba72a | 125 | |
4e23bcae | 126 | struct { |
34e83fc6 KO |
127 | /* |
128 | * This counts the number of available slots in the ringbuffer, | |
129 | * so we avoid overflowing it: it's decremented (if positive) | |
130 | * when allocating a kiocb and incremented when the resulting | |
131 | * io_event is pulled off the ringbuffer. | |
e1bdd5f2 KO |
132 | * |
133 | * We batch accesses to it with a percpu version. | |
34e83fc6 KO |
134 | */ |
135 | atomic_t reqs_available; | |
4e23bcae KO |
136 | } ____cacheline_aligned_in_smp; |
137 | ||
138 | struct { | |
139 | spinlock_t ctx_lock; | |
140 | struct list_head active_reqs; /* used for cancellation */ | |
141 | } ____cacheline_aligned_in_smp; | |
142 | ||
58c85dc2 KO |
143 | struct { |
144 | struct mutex ring_lock; | |
4e23bcae KO |
145 | wait_queue_head_t wait; |
146 | } ____cacheline_aligned_in_smp; | |
58c85dc2 KO |
147 | |
148 | struct { | |
149 | unsigned tail; | |
d856f32a | 150 | unsigned completed_events; |
58c85dc2 | 151 | spinlock_t completion_lock; |
4e23bcae | 152 | } ____cacheline_aligned_in_smp; |
58c85dc2 KO |
153 | |
154 | struct page *internal_pages[AIO_RING_PAGES]; | |
36bc08cc | 155 | struct file *aio_ring_file; |
db446a08 BL |
156 | |
157 | unsigned id; | |
4e179bca KO |
158 | }; |
159 | ||
04b2fa9f CH |
160 | /* |
161 | * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either | |
162 | * cancelled or completed (this makes a certain amount of sense because | |
163 | * successful cancellation - io_cancel() - does deliver the completion to | |
164 | * userspace). | |
165 | * | |
166 | * And since most things don't implement kiocb cancellation and we'd really like | |
167 | * kiocb completion to be lockless when possible, we use ki_cancel to | |
168 | * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED | |
169 | * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel(). | |
170 | */ | |
171 | #define KIOCB_CANCELLED ((void *) (~0ULL)) | |
172 | ||
173 | struct aio_kiocb { | |
174 | struct kiocb common; | |
175 | ||
176 | struct kioctx *ki_ctx; | |
177 | kiocb_cancel_fn *ki_cancel; | |
178 | ||
179 | struct iocb __user *ki_user_iocb; /* user's aiocb */ | |
180 | __u64 ki_user_data; /* user's data for completion */ | |
181 | ||
182 | struct list_head ki_list; /* the aio core uses this | |
183 | * for cancellation */ | |
184 | ||
185 | /* | |
186 | * If the aio_resfd field of the userspace iocb is not zero, | |
187 | * this is the underlying eventfd context to deliver events to. | |
188 | */ | |
189 | struct eventfd_ctx *ki_eventfd; | |
190 | }; | |
191 | ||
1da177e4 | 192 | /*------ sysctl variables----*/ |
d55b5fda ZB |
193 | static DEFINE_SPINLOCK(aio_nr_lock); |
194 | unsigned long aio_nr; /* current system wide number of aio requests */ | |
195 | unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ | |
1da177e4 LT |
196 | /*----end sysctl variables---*/ |
197 | ||
e18b890b CL |
198 | static struct kmem_cache *kiocb_cachep; |
199 | static struct kmem_cache *kioctx_cachep; | |
1da177e4 | 200 | |
71ad7490 BL |
201 | static struct vfsmount *aio_mnt; |
202 | ||
203 | static const struct file_operations aio_ring_fops; | |
204 | static const struct address_space_operations aio_ctx_aops; | |
205 | ||
206 | static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages) | |
207 | { | |
208 | struct qstr this = QSTR_INIT("[aio]", 5); | |
209 | struct file *file; | |
210 | struct path path; | |
211 | struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb); | |
7f62656b DC |
212 | if (IS_ERR(inode)) |
213 | return ERR_CAST(inode); | |
71ad7490 BL |
214 | |
215 | inode->i_mapping->a_ops = &aio_ctx_aops; | |
216 | inode->i_mapping->private_data = ctx; | |
217 | inode->i_size = PAGE_SIZE * nr_pages; | |
218 | ||
219 | path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this); | |
220 | if (!path.dentry) { | |
221 | iput(inode); | |
222 | return ERR_PTR(-ENOMEM); | |
223 | } | |
224 | path.mnt = mntget(aio_mnt); | |
225 | ||
226 | d_instantiate(path.dentry, inode); | |
227 | file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops); | |
228 | if (IS_ERR(file)) { | |
229 | path_put(&path); | |
230 | return file; | |
231 | } | |
232 | ||
233 | file->f_flags = O_RDWR; | |
71ad7490 BL |
234 | return file; |
235 | } | |
236 | ||
237 | static struct dentry *aio_mount(struct file_system_type *fs_type, | |
238 | int flags, const char *dev_name, void *data) | |
239 | { | |
240 | static const struct dentry_operations ops = { | |
241 | .d_dname = simple_dname, | |
242 | }; | |
22f6b4d3 JH |
243 | struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops, |
244 | AIO_RING_MAGIC); | |
245 | ||
246 | if (!IS_ERR(root)) | |
247 | root->d_sb->s_iflags |= SB_I_NOEXEC; | |
248 | return root; | |
71ad7490 BL |
249 | } |
250 | ||
1da177e4 LT |
251 | /* aio_setup |
252 | * Creates the slab caches used by the aio routines, panic on | |
253 | * failure as this is done early during the boot sequence. | |
254 | */ | |
255 | static int __init aio_setup(void) | |
256 | { | |
71ad7490 BL |
257 | static struct file_system_type aio_fs = { |
258 | .name = "aio", | |
259 | .mount = aio_mount, | |
260 | .kill_sb = kill_anon_super, | |
261 | }; | |
262 | aio_mnt = kern_mount(&aio_fs); | |
263 | if (IS_ERR(aio_mnt)) | |
264 | panic("Failed to create aio fs mount."); | |
265 | ||
04b2fa9f | 266 | kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
0a31bd5f | 267 | kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
1da177e4 | 268 | |
caf4167a | 269 | pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page)); |
1da177e4 LT |
270 | |
271 | return 0; | |
272 | } | |
385773e0 | 273 | __initcall(aio_setup); |
1da177e4 | 274 | |
5e9ae2e5 BL |
275 | static void put_aio_ring_file(struct kioctx *ctx) |
276 | { | |
277 | struct file *aio_ring_file = ctx->aio_ring_file; | |
de04e769 RV |
278 | struct address_space *i_mapping; |
279 | ||
5e9ae2e5 | 280 | if (aio_ring_file) { |
45063097 | 281 | truncate_setsize(file_inode(aio_ring_file), 0); |
5e9ae2e5 BL |
282 | |
283 | /* Prevent further access to the kioctx from migratepages */ | |
45063097 | 284 | i_mapping = aio_ring_file->f_mapping; |
de04e769 RV |
285 | spin_lock(&i_mapping->private_lock); |
286 | i_mapping->private_data = NULL; | |
5e9ae2e5 | 287 | ctx->aio_ring_file = NULL; |
de04e769 | 288 | spin_unlock(&i_mapping->private_lock); |
5e9ae2e5 BL |
289 | |
290 | fput(aio_ring_file); | |
291 | } | |
292 | } | |
293 | ||
1da177e4 LT |
294 | static void aio_free_ring(struct kioctx *ctx) |
295 | { | |
36bc08cc | 296 | int i; |
1da177e4 | 297 | |
fa8a53c3 BL |
298 | /* Disconnect the kiotx from the ring file. This prevents future |
299 | * accesses to the kioctx from page migration. | |
300 | */ | |
301 | put_aio_ring_file(ctx); | |
302 | ||
36bc08cc | 303 | for (i = 0; i < ctx->nr_pages; i++) { |
8e321fef | 304 | struct page *page; |
36bc08cc GZ |
305 | pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i, |
306 | page_count(ctx->ring_pages[i])); | |
8e321fef BL |
307 | page = ctx->ring_pages[i]; |
308 | if (!page) | |
309 | continue; | |
310 | ctx->ring_pages[i] = NULL; | |
311 | put_page(page); | |
36bc08cc | 312 | } |
1da177e4 | 313 | |
ddb8c45b | 314 | if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) { |
58c85dc2 | 315 | kfree(ctx->ring_pages); |
ddb8c45b SL |
316 | ctx->ring_pages = NULL; |
317 | } | |
36bc08cc GZ |
318 | } |
319 | ||
5477e70a | 320 | static int aio_ring_mremap(struct vm_area_struct *vma) |
e4a0d3e7 | 321 | { |
5477e70a | 322 | struct file *file = vma->vm_file; |
e4a0d3e7 PE |
323 | struct mm_struct *mm = vma->vm_mm; |
324 | struct kioctx_table *table; | |
b2edffdd | 325 | int i, res = -EINVAL; |
e4a0d3e7 PE |
326 | |
327 | spin_lock(&mm->ioctx_lock); | |
328 | rcu_read_lock(); | |
329 | table = rcu_dereference(mm->ioctx_table); | |
330 | for (i = 0; i < table->nr; i++) { | |
331 | struct kioctx *ctx; | |
332 | ||
d0264c01 | 333 | ctx = rcu_dereference(table->table[i]); |
e4a0d3e7 | 334 | if (ctx && ctx->aio_ring_file == file) { |
b2edffdd AV |
335 | if (!atomic_read(&ctx->dead)) { |
336 | ctx->user_id = ctx->mmap_base = vma->vm_start; | |
337 | res = 0; | |
338 | } | |
e4a0d3e7 PE |
339 | break; |
340 | } | |
341 | } | |
342 | ||
343 | rcu_read_unlock(); | |
344 | spin_unlock(&mm->ioctx_lock); | |
b2edffdd | 345 | return res; |
e4a0d3e7 PE |
346 | } |
347 | ||
5477e70a ON |
348 | static const struct vm_operations_struct aio_ring_vm_ops = { |
349 | .mremap = aio_ring_mremap, | |
350 | #if IS_ENABLED(CONFIG_MMU) | |
351 | .fault = filemap_fault, | |
352 | .map_pages = filemap_map_pages, | |
353 | .page_mkwrite = filemap_page_mkwrite, | |
354 | #endif | |
355 | }; | |
356 | ||
357 | static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma) | |
358 | { | |
359 | vma->vm_flags |= VM_DONTEXPAND; | |
360 | vma->vm_ops = &aio_ring_vm_ops; | |
361 | return 0; | |
362 | } | |
363 | ||
36bc08cc GZ |
364 | static const struct file_operations aio_ring_fops = { |
365 | .mmap = aio_ring_mmap, | |
366 | }; | |
367 | ||
0c45355f | 368 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc GZ |
369 | static int aio_migratepage(struct address_space *mapping, struct page *new, |
370 | struct page *old, enum migrate_mode mode) | |
371 | { | |
5e9ae2e5 | 372 | struct kioctx *ctx; |
36bc08cc | 373 | unsigned long flags; |
fa8a53c3 | 374 | pgoff_t idx; |
36bc08cc GZ |
375 | int rc; |
376 | ||
2916ecc0 JG |
377 | /* |
378 | * We cannot support the _NO_COPY case here, because copy needs to | |
379 | * happen under the ctx->completion_lock. That does not work with the | |
380 | * migration workflow of MIGRATE_SYNC_NO_COPY. | |
381 | */ | |
382 | if (mode == MIGRATE_SYNC_NO_COPY) | |
383 | return -EINVAL; | |
384 | ||
8e321fef BL |
385 | rc = 0; |
386 | ||
fa8a53c3 | 387 | /* mapping->private_lock here protects against the kioctx teardown. */ |
8e321fef BL |
388 | spin_lock(&mapping->private_lock); |
389 | ctx = mapping->private_data; | |
fa8a53c3 BL |
390 | if (!ctx) { |
391 | rc = -EINVAL; | |
392 | goto out; | |
393 | } | |
394 | ||
395 | /* The ring_lock mutex. The prevents aio_read_events() from writing | |
396 | * to the ring's head, and prevents page migration from mucking in | |
397 | * a partially initialized kiotx. | |
398 | */ | |
399 | if (!mutex_trylock(&ctx->ring_lock)) { | |
400 | rc = -EAGAIN; | |
401 | goto out; | |
402 | } | |
403 | ||
404 | idx = old->index; | |
405 | if (idx < (pgoff_t)ctx->nr_pages) { | |
406 | /* Make sure the old page hasn't already been changed */ | |
407 | if (ctx->ring_pages[idx] != old) | |
408 | rc = -EAGAIN; | |
8e321fef BL |
409 | } else |
410 | rc = -EINVAL; | |
8e321fef BL |
411 | |
412 | if (rc != 0) | |
fa8a53c3 | 413 | goto out_unlock; |
8e321fef | 414 | |
36bc08cc GZ |
415 | /* Writeback must be complete */ |
416 | BUG_ON(PageWriteback(old)); | |
8e321fef | 417 | get_page(new); |
36bc08cc | 418 | |
8e321fef | 419 | rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1); |
36bc08cc | 420 | if (rc != MIGRATEPAGE_SUCCESS) { |
8e321fef | 421 | put_page(new); |
fa8a53c3 | 422 | goto out_unlock; |
36bc08cc GZ |
423 | } |
424 | ||
fa8a53c3 BL |
425 | /* Take completion_lock to prevent other writes to the ring buffer |
426 | * while the old page is copied to the new. This prevents new | |
427 | * events from being lost. | |
5e9ae2e5 | 428 | */ |
fa8a53c3 BL |
429 | spin_lock_irqsave(&ctx->completion_lock, flags); |
430 | migrate_page_copy(new, old); | |
431 | BUG_ON(ctx->ring_pages[idx] != old); | |
432 | ctx->ring_pages[idx] = new; | |
433 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
36bc08cc | 434 | |
fa8a53c3 BL |
435 | /* The old page is no longer accessible. */ |
436 | put_page(old); | |
8e321fef | 437 | |
fa8a53c3 BL |
438 | out_unlock: |
439 | mutex_unlock(&ctx->ring_lock); | |
440 | out: | |
441 | spin_unlock(&mapping->private_lock); | |
36bc08cc | 442 | return rc; |
1da177e4 | 443 | } |
0c45355f | 444 | #endif |
1da177e4 | 445 | |
36bc08cc | 446 | static const struct address_space_operations aio_ctx_aops = { |
835f252c | 447 | .set_page_dirty = __set_page_dirty_no_writeback, |
0c45355f | 448 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc | 449 | .migratepage = aio_migratepage, |
0c45355f | 450 | #endif |
36bc08cc GZ |
451 | }; |
452 | ||
2a8a9867 | 453 | static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events) |
1da177e4 LT |
454 | { |
455 | struct aio_ring *ring; | |
41003a7b | 456 | struct mm_struct *mm = current->mm; |
3dc9acb6 | 457 | unsigned long size, unused; |
1da177e4 | 458 | int nr_pages; |
36bc08cc GZ |
459 | int i; |
460 | struct file *file; | |
1da177e4 LT |
461 | |
462 | /* Compensate for the ring buffer's head/tail overlap entry */ | |
463 | nr_events += 2; /* 1 is required, 2 for good luck */ | |
464 | ||
465 | size = sizeof(struct aio_ring); | |
466 | size += sizeof(struct io_event) * nr_events; | |
1da177e4 | 467 | |
36bc08cc | 468 | nr_pages = PFN_UP(size); |
1da177e4 LT |
469 | if (nr_pages < 0) |
470 | return -EINVAL; | |
471 | ||
71ad7490 | 472 | file = aio_private_file(ctx, nr_pages); |
36bc08cc GZ |
473 | if (IS_ERR(file)) { |
474 | ctx->aio_ring_file = NULL; | |
fa8a53c3 | 475 | return -ENOMEM; |
36bc08cc GZ |
476 | } |
477 | ||
3dc9acb6 LT |
478 | ctx->aio_ring_file = file; |
479 | nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) | |
480 | / sizeof(struct io_event); | |
481 | ||
482 | ctx->ring_pages = ctx->internal_pages; | |
483 | if (nr_pages > AIO_RING_PAGES) { | |
484 | ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *), | |
485 | GFP_KERNEL); | |
486 | if (!ctx->ring_pages) { | |
487 | put_aio_ring_file(ctx); | |
488 | return -ENOMEM; | |
489 | } | |
490 | } | |
491 | ||
36bc08cc GZ |
492 | for (i = 0; i < nr_pages; i++) { |
493 | struct page *page; | |
45063097 | 494 | page = find_or_create_page(file->f_mapping, |
36bc08cc GZ |
495 | i, GFP_HIGHUSER | __GFP_ZERO); |
496 | if (!page) | |
497 | break; | |
498 | pr_debug("pid(%d) page[%d]->count=%d\n", | |
499 | current->pid, i, page_count(page)); | |
500 | SetPageUptodate(page); | |
36bc08cc | 501 | unlock_page(page); |
3dc9acb6 LT |
502 | |
503 | ctx->ring_pages[i] = page; | |
36bc08cc | 504 | } |
3dc9acb6 | 505 | ctx->nr_pages = i; |
1da177e4 | 506 | |
3dc9acb6 LT |
507 | if (unlikely(i != nr_pages)) { |
508 | aio_free_ring(ctx); | |
fa8a53c3 | 509 | return -ENOMEM; |
1da177e4 LT |
510 | } |
511 | ||
58c85dc2 KO |
512 | ctx->mmap_size = nr_pages * PAGE_SIZE; |
513 | pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size); | |
36bc08cc | 514 | |
013373e8 MH |
515 | if (down_write_killable(&mm->mmap_sem)) { |
516 | ctx->mmap_size = 0; | |
517 | aio_free_ring(ctx); | |
518 | return -EINTR; | |
519 | } | |
520 | ||
36bc08cc GZ |
521 | ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size, |
522 | PROT_READ | PROT_WRITE, | |
897ab3e0 | 523 | MAP_SHARED, 0, &unused, NULL); |
3dc9acb6 | 524 | up_write(&mm->mmap_sem); |
58c85dc2 | 525 | if (IS_ERR((void *)ctx->mmap_base)) { |
58c85dc2 | 526 | ctx->mmap_size = 0; |
1da177e4 | 527 | aio_free_ring(ctx); |
fa8a53c3 | 528 | return -ENOMEM; |
1da177e4 LT |
529 | } |
530 | ||
58c85dc2 | 531 | pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base); |
d6c355c7 | 532 | |
58c85dc2 KO |
533 | ctx->user_id = ctx->mmap_base; |
534 | ctx->nr_events = nr_events; /* trusted copy */ | |
1da177e4 | 535 | |
58c85dc2 | 536 | ring = kmap_atomic(ctx->ring_pages[0]); |
1da177e4 | 537 | ring->nr = nr_events; /* user copy */ |
db446a08 | 538 | ring->id = ~0U; |
1da177e4 LT |
539 | ring->head = ring->tail = 0; |
540 | ring->magic = AIO_RING_MAGIC; | |
541 | ring->compat_features = AIO_RING_COMPAT_FEATURES; | |
542 | ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; | |
543 | ring->header_length = sizeof(struct aio_ring); | |
e8e3c3d6 | 544 | kunmap_atomic(ring); |
58c85dc2 | 545 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 LT |
546 | |
547 | return 0; | |
548 | } | |
549 | ||
1da177e4 LT |
550 | #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) |
551 | #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) | |
552 | #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) | |
553 | ||
04b2fa9f | 554 | void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel) |
0460fef2 | 555 | { |
04b2fa9f | 556 | struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common); |
0460fef2 KO |
557 | struct kioctx *ctx = req->ki_ctx; |
558 | unsigned long flags; | |
559 | ||
560 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
561 | ||
562 | if (!req->ki_list.next) | |
563 | list_add(&req->ki_list, &ctx->active_reqs); | |
564 | ||
565 | req->ki_cancel = cancel; | |
566 | ||
567 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
568 | } | |
569 | EXPORT_SYMBOL(kiocb_set_cancel_fn); | |
570 | ||
04b2fa9f | 571 | static int kiocb_cancel(struct aio_kiocb *kiocb) |
906b973c | 572 | { |
0460fef2 | 573 | kiocb_cancel_fn *old, *cancel; |
906b973c | 574 | |
0460fef2 KO |
575 | /* |
576 | * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it | |
577 | * actually has a cancel function, hence the cmpxchg() | |
578 | */ | |
579 | ||
6aa7de05 | 580 | cancel = READ_ONCE(kiocb->ki_cancel); |
0460fef2 KO |
581 | do { |
582 | if (!cancel || cancel == KIOCB_CANCELLED) | |
57282d8f | 583 | return -EINVAL; |
906b973c | 584 | |
0460fef2 KO |
585 | old = cancel; |
586 | cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED); | |
587 | } while (cancel != old); | |
906b973c | 588 | |
04b2fa9f | 589 | return cancel(&kiocb->common); |
906b973c KO |
590 | } |
591 | ||
a6d7cff4 TH |
592 | /* |
593 | * free_ioctx() should be RCU delayed to synchronize against the RCU | |
594 | * protected lookup_ioctx() and also needs process context to call | |
595 | * aio_free_ring(), so the double bouncing through kioctx->free_rcu and | |
596 | * ->free_work. | |
597 | */ | |
e34ecee2 | 598 | static void free_ioctx(struct work_struct *work) |
36f55889 | 599 | { |
e34ecee2 | 600 | struct kioctx *ctx = container_of(work, struct kioctx, free_work); |
e1bdd5f2 | 601 | |
e34ecee2 | 602 | pr_debug("freeing %p\n", ctx); |
e1bdd5f2 | 603 | |
e34ecee2 | 604 | aio_free_ring(ctx); |
e1bdd5f2 | 605 | free_percpu(ctx->cpu); |
9a1049da TH |
606 | percpu_ref_exit(&ctx->reqs); |
607 | percpu_ref_exit(&ctx->users); | |
36f55889 KO |
608 | kmem_cache_free(kioctx_cachep, ctx); |
609 | } | |
610 | ||
a6d7cff4 TH |
611 | static void free_ioctx_rcufn(struct rcu_head *head) |
612 | { | |
613 | struct kioctx *ctx = container_of(head, struct kioctx, free_rcu); | |
614 | ||
615 | INIT_WORK(&ctx->free_work, free_ioctx); | |
616 | schedule_work(&ctx->free_work); | |
617 | } | |
618 | ||
e34ecee2 KO |
619 | static void free_ioctx_reqs(struct percpu_ref *ref) |
620 | { | |
621 | struct kioctx *ctx = container_of(ref, struct kioctx, reqs); | |
622 | ||
e02ba72a | 623 | /* At this point we know that there are no any in-flight requests */ |
dc48e56d JA |
624 | if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count)) |
625 | complete(&ctx->rq_wait->comp); | |
e02ba72a | 626 | |
a6d7cff4 TH |
627 | /* Synchronize against RCU protected table->table[] dereferences */ |
628 | call_rcu(&ctx->free_rcu, free_ioctx_rcufn); | |
e34ecee2 KO |
629 | } |
630 | ||
36f55889 KO |
631 | /* |
632 | * When this function runs, the kioctx has been removed from the "hash table" | |
633 | * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - | |
634 | * now it's safe to cancel any that need to be. | |
635 | */ | |
e34ecee2 | 636 | static void free_ioctx_users(struct percpu_ref *ref) |
36f55889 | 637 | { |
e34ecee2 | 638 | struct kioctx *ctx = container_of(ref, struct kioctx, users); |
04b2fa9f | 639 | struct aio_kiocb *req; |
36f55889 KO |
640 | |
641 | spin_lock_irq(&ctx->ctx_lock); | |
642 | ||
643 | while (!list_empty(&ctx->active_reqs)) { | |
644 | req = list_first_entry(&ctx->active_reqs, | |
04b2fa9f | 645 | struct aio_kiocb, ki_list); |
36f55889 KO |
646 | |
647 | list_del_init(&req->ki_list); | |
d52a8f9e | 648 | kiocb_cancel(req); |
36f55889 KO |
649 | } |
650 | ||
651 | spin_unlock_irq(&ctx->ctx_lock); | |
652 | ||
e34ecee2 KO |
653 | percpu_ref_kill(&ctx->reqs); |
654 | percpu_ref_put(&ctx->reqs); | |
36f55889 KO |
655 | } |
656 | ||
db446a08 BL |
657 | static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm) |
658 | { | |
659 | unsigned i, new_nr; | |
660 | struct kioctx_table *table, *old; | |
661 | struct aio_ring *ring; | |
662 | ||
663 | spin_lock(&mm->ioctx_lock); | |
855ef0de | 664 | table = rcu_dereference_raw(mm->ioctx_table); |
db446a08 BL |
665 | |
666 | while (1) { | |
667 | if (table) | |
668 | for (i = 0; i < table->nr; i++) | |
d0264c01 | 669 | if (!rcu_access_pointer(table->table[i])) { |
db446a08 | 670 | ctx->id = i; |
d0264c01 | 671 | rcu_assign_pointer(table->table[i], ctx); |
db446a08 BL |
672 | spin_unlock(&mm->ioctx_lock); |
673 | ||
fa8a53c3 BL |
674 | /* While kioctx setup is in progress, |
675 | * we are protected from page migration | |
676 | * changes ring_pages by ->ring_lock. | |
677 | */ | |
db446a08 BL |
678 | ring = kmap_atomic(ctx->ring_pages[0]); |
679 | ring->id = ctx->id; | |
680 | kunmap_atomic(ring); | |
681 | return 0; | |
682 | } | |
683 | ||
684 | new_nr = (table ? table->nr : 1) * 4; | |
db446a08 BL |
685 | spin_unlock(&mm->ioctx_lock); |
686 | ||
687 | table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) * | |
688 | new_nr, GFP_KERNEL); | |
689 | if (!table) | |
690 | return -ENOMEM; | |
691 | ||
692 | table->nr = new_nr; | |
693 | ||
694 | spin_lock(&mm->ioctx_lock); | |
855ef0de | 695 | old = rcu_dereference_raw(mm->ioctx_table); |
db446a08 BL |
696 | |
697 | if (!old) { | |
698 | rcu_assign_pointer(mm->ioctx_table, table); | |
699 | } else if (table->nr > old->nr) { | |
700 | memcpy(table->table, old->table, | |
701 | old->nr * sizeof(struct kioctx *)); | |
702 | ||
703 | rcu_assign_pointer(mm->ioctx_table, table); | |
704 | kfree_rcu(old, rcu); | |
705 | } else { | |
706 | kfree(table); | |
707 | table = old; | |
708 | } | |
709 | } | |
710 | } | |
711 | ||
e34ecee2 KO |
712 | static void aio_nr_sub(unsigned nr) |
713 | { | |
714 | spin_lock(&aio_nr_lock); | |
715 | if (WARN_ON(aio_nr - nr > aio_nr)) | |
716 | aio_nr = 0; | |
717 | else | |
718 | aio_nr -= nr; | |
719 | spin_unlock(&aio_nr_lock); | |
720 | } | |
721 | ||
1da177e4 LT |
722 | /* ioctx_alloc |
723 | * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. | |
724 | */ | |
725 | static struct kioctx *ioctx_alloc(unsigned nr_events) | |
726 | { | |
41003a7b | 727 | struct mm_struct *mm = current->mm; |
1da177e4 | 728 | struct kioctx *ctx; |
e23754f8 | 729 | int err = -ENOMEM; |
1da177e4 | 730 | |
2a8a9867 MFO |
731 | /* |
732 | * Store the original nr_events -- what userspace passed to io_setup(), | |
733 | * for counting against the global limit -- before it changes. | |
734 | */ | |
735 | unsigned int max_reqs = nr_events; | |
736 | ||
e1bdd5f2 KO |
737 | /* |
738 | * We keep track of the number of available ringbuffer slots, to prevent | |
739 | * overflow (reqs_available), and we also use percpu counters for this. | |
740 | * | |
741 | * So since up to half the slots might be on other cpu's percpu counters | |
742 | * and unavailable, double nr_events so userspace sees what they | |
743 | * expected: additionally, we move req_batch slots to/from percpu | |
744 | * counters at a time, so make sure that isn't 0: | |
745 | */ | |
746 | nr_events = max(nr_events, num_possible_cpus() * 4); | |
747 | nr_events *= 2; | |
748 | ||
1da177e4 | 749 | /* Prevent overflows */ |
08397acd | 750 | if (nr_events > (0x10000000U / sizeof(struct io_event))) { |
1da177e4 LT |
751 | pr_debug("ENOMEM: nr_events too high\n"); |
752 | return ERR_PTR(-EINVAL); | |
753 | } | |
754 | ||
2a8a9867 | 755 | if (!nr_events || (unsigned long)max_reqs > aio_max_nr) |
1da177e4 LT |
756 | return ERR_PTR(-EAGAIN); |
757 | ||
c3762229 | 758 | ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); |
1da177e4 LT |
759 | if (!ctx) |
760 | return ERR_PTR(-ENOMEM); | |
761 | ||
2a8a9867 | 762 | ctx->max_reqs = max_reqs; |
1da177e4 | 763 | |
1da177e4 | 764 | spin_lock_init(&ctx->ctx_lock); |
0460fef2 | 765 | spin_lock_init(&ctx->completion_lock); |
58c85dc2 | 766 | mutex_init(&ctx->ring_lock); |
fa8a53c3 BL |
767 | /* Protect against page migration throughout kiotx setup by keeping |
768 | * the ring_lock mutex held until setup is complete. */ | |
769 | mutex_lock(&ctx->ring_lock); | |
1da177e4 LT |
770 | init_waitqueue_head(&ctx->wait); |
771 | ||
772 | INIT_LIST_HEAD(&ctx->active_reqs); | |
1da177e4 | 773 | |
2aad2a86 | 774 | if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL)) |
fa8a53c3 BL |
775 | goto err; |
776 | ||
2aad2a86 | 777 | if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL)) |
fa8a53c3 BL |
778 | goto err; |
779 | ||
e1bdd5f2 KO |
780 | ctx->cpu = alloc_percpu(struct kioctx_cpu); |
781 | if (!ctx->cpu) | |
e34ecee2 | 782 | goto err; |
1da177e4 | 783 | |
2a8a9867 | 784 | err = aio_setup_ring(ctx, nr_events); |
fa8a53c3 | 785 | if (err < 0) |
e34ecee2 | 786 | goto err; |
e1bdd5f2 | 787 | |
34e83fc6 | 788 | atomic_set(&ctx->reqs_available, ctx->nr_events - 1); |
e1bdd5f2 | 789 | ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4); |
6878ea72 BL |
790 | if (ctx->req_batch < 1) |
791 | ctx->req_batch = 1; | |
34e83fc6 | 792 | |
1da177e4 | 793 | /* limit the number of system wide aios */ |
9fa1cb39 | 794 | spin_lock(&aio_nr_lock); |
2a8a9867 MFO |
795 | if (aio_nr + ctx->max_reqs > aio_max_nr || |
796 | aio_nr + ctx->max_reqs < aio_nr) { | |
9fa1cb39 | 797 | spin_unlock(&aio_nr_lock); |
e34ecee2 | 798 | err = -EAGAIN; |
d1b94327 | 799 | goto err_ctx; |
2dd542b7 AV |
800 | } |
801 | aio_nr += ctx->max_reqs; | |
9fa1cb39 | 802 | spin_unlock(&aio_nr_lock); |
1da177e4 | 803 | |
1881686f BL |
804 | percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */ |
805 | percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */ | |
723be6e3 | 806 | |
da90382c BL |
807 | err = ioctx_add_table(ctx, mm); |
808 | if (err) | |
e34ecee2 | 809 | goto err_cleanup; |
da90382c | 810 | |
fa8a53c3 BL |
811 | /* Release the ring_lock mutex now that all setup is complete. */ |
812 | mutex_unlock(&ctx->ring_lock); | |
813 | ||
caf4167a | 814 | pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", |
58c85dc2 | 815 | ctx, ctx->user_id, mm, ctx->nr_events); |
1da177e4 LT |
816 | return ctx; |
817 | ||
e34ecee2 KO |
818 | err_cleanup: |
819 | aio_nr_sub(ctx->max_reqs); | |
d1b94327 | 820 | err_ctx: |
deeb8525 AV |
821 | atomic_set(&ctx->dead, 1); |
822 | if (ctx->mmap_size) | |
823 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
d1b94327 | 824 | aio_free_ring(ctx); |
e34ecee2 | 825 | err: |
fa8a53c3 | 826 | mutex_unlock(&ctx->ring_lock); |
e1bdd5f2 | 827 | free_percpu(ctx->cpu); |
9a1049da TH |
828 | percpu_ref_exit(&ctx->reqs); |
829 | percpu_ref_exit(&ctx->users); | |
1da177e4 | 830 | kmem_cache_free(kioctx_cachep, ctx); |
caf4167a | 831 | pr_debug("error allocating ioctx %d\n", err); |
e23754f8 | 832 | return ERR_PTR(err); |
1da177e4 LT |
833 | } |
834 | ||
36f55889 KO |
835 | /* kill_ioctx |
836 | * Cancels all outstanding aio requests on an aio context. Used | |
837 | * when the processes owning a context have all exited to encourage | |
838 | * the rapid destruction of the kioctx. | |
839 | */ | |
fb2d4483 | 840 | static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx, |
dc48e56d | 841 | struct ctx_rq_wait *wait) |
36f55889 | 842 | { |
fa88b6f8 | 843 | struct kioctx_table *table; |
db446a08 | 844 | |
b2edffdd AV |
845 | spin_lock(&mm->ioctx_lock); |
846 | if (atomic_xchg(&ctx->dead, 1)) { | |
847 | spin_unlock(&mm->ioctx_lock); | |
fa88b6f8 | 848 | return -EINVAL; |
b2edffdd | 849 | } |
db446a08 | 850 | |
855ef0de | 851 | table = rcu_dereference_raw(mm->ioctx_table); |
d0264c01 TH |
852 | WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id])); |
853 | RCU_INIT_POINTER(table->table[ctx->id], NULL); | |
fa88b6f8 | 854 | spin_unlock(&mm->ioctx_lock); |
4fcc712f | 855 | |
a6d7cff4 | 856 | /* free_ioctx_reqs() will do the necessary RCU synchronization */ |
fa88b6f8 | 857 | wake_up_all(&ctx->wait); |
4fcc712f | 858 | |
fa88b6f8 BL |
859 | /* |
860 | * It'd be more correct to do this in free_ioctx(), after all | |
861 | * the outstanding kiocbs have finished - but by then io_destroy | |
862 | * has already returned, so io_setup() could potentially return | |
863 | * -EAGAIN with no ioctxs actually in use (as far as userspace | |
864 | * could tell). | |
865 | */ | |
866 | aio_nr_sub(ctx->max_reqs); | |
4fcc712f | 867 | |
fa88b6f8 BL |
868 | if (ctx->mmap_size) |
869 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
fb2d4483 | 870 | |
dc48e56d | 871 | ctx->rq_wait = wait; |
fa88b6f8 BL |
872 | percpu_ref_kill(&ctx->users); |
873 | return 0; | |
1da177e4 LT |
874 | } |
875 | ||
36f55889 KO |
876 | /* |
877 | * exit_aio: called when the last user of mm goes away. At this point, there is | |
878 | * no way for any new requests to be submited or any of the io_* syscalls to be | |
879 | * called on the context. | |
880 | * | |
881 | * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on | |
882 | * them. | |
1da177e4 | 883 | */ |
fc9b52cd | 884 | void exit_aio(struct mm_struct *mm) |
1da177e4 | 885 | { |
4b70ac5f | 886 | struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table); |
dc48e56d JA |
887 | struct ctx_rq_wait wait; |
888 | int i, skipped; | |
db446a08 | 889 | |
4b70ac5f ON |
890 | if (!table) |
891 | return; | |
db446a08 | 892 | |
dc48e56d JA |
893 | atomic_set(&wait.count, table->nr); |
894 | init_completion(&wait.comp); | |
895 | ||
896 | skipped = 0; | |
4b70ac5f | 897 | for (i = 0; i < table->nr; ++i) { |
d0264c01 TH |
898 | struct kioctx *ctx = |
899 | rcu_dereference_protected(table->table[i], true); | |
abf137dd | 900 | |
dc48e56d JA |
901 | if (!ctx) { |
902 | skipped++; | |
4b70ac5f | 903 | continue; |
dc48e56d JA |
904 | } |
905 | ||
936af157 | 906 | /* |
4b70ac5f ON |
907 | * We don't need to bother with munmap() here - exit_mmap(mm) |
908 | * is coming and it'll unmap everything. And we simply can't, | |
909 | * this is not necessarily our ->mm. | |
910 | * Since kill_ioctx() uses non-zero ->mmap_size as indicator | |
911 | * that it needs to unmap the area, just set it to 0. | |
936af157 | 912 | */ |
58c85dc2 | 913 | ctx->mmap_size = 0; |
dc48e56d JA |
914 | kill_ioctx(mm, ctx, &wait); |
915 | } | |
36f55889 | 916 | |
dc48e56d | 917 | if (!atomic_sub_and_test(skipped, &wait.count)) { |
6098b45b | 918 | /* Wait until all IO for the context are done. */ |
dc48e56d | 919 | wait_for_completion(&wait.comp); |
1da177e4 | 920 | } |
4b70ac5f ON |
921 | |
922 | RCU_INIT_POINTER(mm->ioctx_table, NULL); | |
923 | kfree(table); | |
1da177e4 LT |
924 | } |
925 | ||
e1bdd5f2 KO |
926 | static void put_reqs_available(struct kioctx *ctx, unsigned nr) |
927 | { | |
928 | struct kioctx_cpu *kcpu; | |
263782c1 | 929 | unsigned long flags; |
e1bdd5f2 | 930 | |
263782c1 | 931 | local_irq_save(flags); |
be6fb451 | 932 | kcpu = this_cpu_ptr(ctx->cpu); |
e1bdd5f2 | 933 | kcpu->reqs_available += nr; |
263782c1 | 934 | |
e1bdd5f2 KO |
935 | while (kcpu->reqs_available >= ctx->req_batch * 2) { |
936 | kcpu->reqs_available -= ctx->req_batch; | |
937 | atomic_add(ctx->req_batch, &ctx->reqs_available); | |
938 | } | |
939 | ||
263782c1 | 940 | local_irq_restore(flags); |
e1bdd5f2 KO |
941 | } |
942 | ||
943 | static bool get_reqs_available(struct kioctx *ctx) | |
944 | { | |
945 | struct kioctx_cpu *kcpu; | |
946 | bool ret = false; | |
263782c1 | 947 | unsigned long flags; |
e1bdd5f2 | 948 | |
263782c1 | 949 | local_irq_save(flags); |
be6fb451 | 950 | kcpu = this_cpu_ptr(ctx->cpu); |
e1bdd5f2 KO |
951 | if (!kcpu->reqs_available) { |
952 | int old, avail = atomic_read(&ctx->reqs_available); | |
953 | ||
954 | do { | |
955 | if (avail < ctx->req_batch) | |
956 | goto out; | |
957 | ||
958 | old = avail; | |
959 | avail = atomic_cmpxchg(&ctx->reqs_available, | |
960 | avail, avail - ctx->req_batch); | |
961 | } while (avail != old); | |
962 | ||
963 | kcpu->reqs_available += ctx->req_batch; | |
964 | } | |
965 | ||
966 | ret = true; | |
967 | kcpu->reqs_available--; | |
968 | out: | |
263782c1 | 969 | local_irq_restore(flags); |
e1bdd5f2 KO |
970 | return ret; |
971 | } | |
972 | ||
d856f32a BL |
973 | /* refill_reqs_available |
974 | * Updates the reqs_available reference counts used for tracking the | |
975 | * number of free slots in the completion ring. This can be called | |
976 | * from aio_complete() (to optimistically update reqs_available) or | |
977 | * from aio_get_req() (the we're out of events case). It must be | |
978 | * called holding ctx->completion_lock. | |
979 | */ | |
980 | static void refill_reqs_available(struct kioctx *ctx, unsigned head, | |
981 | unsigned tail) | |
982 | { | |
983 | unsigned events_in_ring, completed; | |
984 | ||
985 | /* Clamp head since userland can write to it. */ | |
986 | head %= ctx->nr_events; | |
987 | if (head <= tail) | |
988 | events_in_ring = tail - head; | |
989 | else | |
990 | events_in_ring = ctx->nr_events - (head - tail); | |
991 | ||
992 | completed = ctx->completed_events; | |
993 | if (events_in_ring < completed) | |
994 | completed -= events_in_ring; | |
995 | else | |
996 | completed = 0; | |
997 | ||
998 | if (!completed) | |
999 | return; | |
1000 | ||
1001 | ctx->completed_events -= completed; | |
1002 | put_reqs_available(ctx, completed); | |
1003 | } | |
1004 | ||
1005 | /* user_refill_reqs_available | |
1006 | * Called to refill reqs_available when aio_get_req() encounters an | |
1007 | * out of space in the completion ring. | |
1008 | */ | |
1009 | static void user_refill_reqs_available(struct kioctx *ctx) | |
1010 | { | |
1011 | spin_lock_irq(&ctx->completion_lock); | |
1012 | if (ctx->completed_events) { | |
1013 | struct aio_ring *ring; | |
1014 | unsigned head; | |
1015 | ||
1016 | /* Access of ring->head may race with aio_read_events_ring() | |
1017 | * here, but that's okay since whether we read the old version | |
1018 | * or the new version, and either will be valid. The important | |
1019 | * part is that head cannot pass tail since we prevent | |
1020 | * aio_complete() from updating tail by holding | |
1021 | * ctx->completion_lock. Even if head is invalid, the check | |
1022 | * against ctx->completed_events below will make sure we do the | |
1023 | * safe/right thing. | |
1024 | */ | |
1025 | ring = kmap_atomic(ctx->ring_pages[0]); | |
1026 | head = ring->head; | |
1027 | kunmap_atomic(ring); | |
1028 | ||
1029 | refill_reqs_available(ctx, head, ctx->tail); | |
1030 | } | |
1031 | ||
1032 | spin_unlock_irq(&ctx->completion_lock); | |
1033 | } | |
1034 | ||
1da177e4 | 1035 | /* aio_get_req |
57282d8f KO |
1036 | * Allocate a slot for an aio request. |
1037 | * Returns NULL if no requests are free. | |
1da177e4 | 1038 | */ |
04b2fa9f | 1039 | static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx) |
1da177e4 | 1040 | { |
04b2fa9f | 1041 | struct aio_kiocb *req; |
a1c8eae7 | 1042 | |
d856f32a BL |
1043 | if (!get_reqs_available(ctx)) { |
1044 | user_refill_reqs_available(ctx); | |
1045 | if (!get_reqs_available(ctx)) | |
1046 | return NULL; | |
1047 | } | |
a1c8eae7 | 1048 | |
0460fef2 | 1049 | req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO); |
1da177e4 | 1050 | if (unlikely(!req)) |
a1c8eae7 | 1051 | goto out_put; |
1da177e4 | 1052 | |
e34ecee2 KO |
1053 | percpu_ref_get(&ctx->reqs); |
1054 | ||
1da177e4 | 1055 | req->ki_ctx = ctx; |
080d676d | 1056 | return req; |
a1c8eae7 | 1057 | out_put: |
e1bdd5f2 | 1058 | put_reqs_available(ctx, 1); |
a1c8eae7 | 1059 | return NULL; |
1da177e4 LT |
1060 | } |
1061 | ||
04b2fa9f | 1062 | static void kiocb_free(struct aio_kiocb *req) |
1da177e4 | 1063 | { |
04b2fa9f CH |
1064 | if (req->common.ki_filp) |
1065 | fput(req->common.ki_filp); | |
13389010 DL |
1066 | if (req->ki_eventfd != NULL) |
1067 | eventfd_ctx_put(req->ki_eventfd); | |
1da177e4 | 1068 | kmem_cache_free(kiocb_cachep, req); |
1da177e4 LT |
1069 | } |
1070 | ||
d5470b59 | 1071 | static struct kioctx *lookup_ioctx(unsigned long ctx_id) |
1da177e4 | 1072 | { |
db446a08 | 1073 | struct aio_ring __user *ring = (void __user *)ctx_id; |
abf137dd | 1074 | struct mm_struct *mm = current->mm; |
65c24491 | 1075 | struct kioctx *ctx, *ret = NULL; |
db446a08 BL |
1076 | struct kioctx_table *table; |
1077 | unsigned id; | |
1078 | ||
1079 | if (get_user(id, &ring->id)) | |
1080 | return NULL; | |
1da177e4 | 1081 | |
abf137dd | 1082 | rcu_read_lock(); |
db446a08 | 1083 | table = rcu_dereference(mm->ioctx_table); |
abf137dd | 1084 | |
db446a08 BL |
1085 | if (!table || id >= table->nr) |
1086 | goto out; | |
1da177e4 | 1087 | |
d0264c01 | 1088 | ctx = rcu_dereference(table->table[id]); |
f30d704f | 1089 | if (ctx && ctx->user_id == ctx_id) { |
db446a08 BL |
1090 | percpu_ref_get(&ctx->users); |
1091 | ret = ctx; | |
1092 | } | |
1093 | out: | |
abf137dd | 1094 | rcu_read_unlock(); |
65c24491 | 1095 | return ret; |
1da177e4 LT |
1096 | } |
1097 | ||
1da177e4 LT |
1098 | /* aio_complete |
1099 | * Called when the io request on the given iocb is complete. | |
1da177e4 | 1100 | */ |
04b2fa9f | 1101 | static void aio_complete(struct kiocb *kiocb, long res, long res2) |
1da177e4 | 1102 | { |
04b2fa9f | 1103 | struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common); |
1da177e4 | 1104 | struct kioctx *ctx = iocb->ki_ctx; |
1da177e4 | 1105 | struct aio_ring *ring; |
21b40200 | 1106 | struct io_event *ev_page, *event; |
d856f32a | 1107 | unsigned tail, pos, head; |
1da177e4 | 1108 | unsigned long flags; |
1da177e4 | 1109 | |
70fe2f48 JK |
1110 | if (kiocb->ki_flags & IOCB_WRITE) { |
1111 | struct file *file = kiocb->ki_filp; | |
1112 | ||
1113 | /* | |
1114 | * Tell lockdep we inherited freeze protection from submission | |
1115 | * thread. | |
1116 | */ | |
a12f1ae6 SL |
1117 | if (S_ISREG(file_inode(file)->i_mode)) |
1118 | __sb_writers_acquired(file_inode(file)->i_sb, SB_FREEZE_WRITE); | |
70fe2f48 JK |
1119 | file_end_write(file); |
1120 | } | |
1121 | ||
20dcae32 ZB |
1122 | /* |
1123 | * Special case handling for sync iocbs: | |
1124 | * - events go directly into the iocb for fast handling | |
1125 | * - the sync task with the iocb in its stack holds the single iocb | |
1126 | * ref, no other paths have a way to get another ref | |
1127 | * - the sync task helpfully left a reference to itself in the iocb | |
1da177e4 | 1128 | */ |
04b2fa9f | 1129 | BUG_ON(is_sync_kiocb(kiocb)); |
1da177e4 | 1130 | |
0460fef2 KO |
1131 | if (iocb->ki_list.next) { |
1132 | unsigned long flags; | |
1133 | ||
1134 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
1135 | list_del(&iocb->ki_list); | |
1136 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
1137 | } | |
11599eba | 1138 | |
0460fef2 KO |
1139 | /* |
1140 | * Add a completion event to the ring buffer. Must be done holding | |
4b30f07e | 1141 | * ctx->completion_lock to prevent other code from messing with the tail |
0460fef2 KO |
1142 | * pointer since we might be called from irq context. |
1143 | */ | |
1144 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1145 | ||
58c85dc2 | 1146 | tail = ctx->tail; |
21b40200 KO |
1147 | pos = tail + AIO_EVENTS_OFFSET; |
1148 | ||
58c85dc2 | 1149 | if (++tail >= ctx->nr_events) |
4bf69b2a | 1150 | tail = 0; |
1da177e4 | 1151 | |
58c85dc2 | 1152 | ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
1153 | event = ev_page + pos % AIO_EVENTS_PER_PAGE; |
1154 | ||
04b2fa9f | 1155 | event->obj = (u64)(unsigned long)iocb->ki_user_iocb; |
1da177e4 LT |
1156 | event->data = iocb->ki_user_data; |
1157 | event->res = res; | |
1158 | event->res2 = res2; | |
1159 | ||
21b40200 | 1160 | kunmap_atomic(ev_page); |
58c85dc2 | 1161 | flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
1162 | |
1163 | pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n", | |
04b2fa9f | 1164 | ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data, |
caf4167a | 1165 | res, res2); |
1da177e4 LT |
1166 | |
1167 | /* after flagging the request as done, we | |
1168 | * must never even look at it again | |
1169 | */ | |
1170 | smp_wmb(); /* make event visible before updating tail */ | |
1171 | ||
58c85dc2 | 1172 | ctx->tail = tail; |
1da177e4 | 1173 | |
58c85dc2 | 1174 | ring = kmap_atomic(ctx->ring_pages[0]); |
d856f32a | 1175 | head = ring->head; |
21b40200 | 1176 | ring->tail = tail; |
e8e3c3d6 | 1177 | kunmap_atomic(ring); |
58c85dc2 | 1178 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 | 1179 | |
d856f32a BL |
1180 | ctx->completed_events++; |
1181 | if (ctx->completed_events > 1) | |
1182 | refill_reqs_available(ctx, head, tail); | |
0460fef2 KO |
1183 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1184 | ||
21b40200 | 1185 | pr_debug("added to ring %p at [%u]\n", iocb, tail); |
8d1c98b0 DL |
1186 | |
1187 | /* | |
1188 | * Check if the user asked us to deliver the result through an | |
1189 | * eventfd. The eventfd_signal() function is safe to be called | |
1190 | * from IRQ context. | |
1191 | */ | |
87c3a86e | 1192 | if (iocb->ki_eventfd != NULL) |
8d1c98b0 DL |
1193 | eventfd_signal(iocb->ki_eventfd, 1); |
1194 | ||
1da177e4 | 1195 | /* everything turned out well, dispose of the aiocb. */ |
57282d8f | 1196 | kiocb_free(iocb); |
1da177e4 | 1197 | |
6cb2a210 QB |
1198 | /* |
1199 | * We have to order our ring_info tail store above and test | |
1200 | * of the wait list below outside the wait lock. This is | |
1201 | * like in wake_up_bit() where clearing a bit has to be | |
1202 | * ordered with the unlocked test. | |
1203 | */ | |
1204 | smp_mb(); | |
1205 | ||
1da177e4 LT |
1206 | if (waitqueue_active(&ctx->wait)) |
1207 | wake_up(&ctx->wait); | |
1208 | ||
e34ecee2 | 1209 | percpu_ref_put(&ctx->reqs); |
1da177e4 LT |
1210 | } |
1211 | ||
2be4e7de | 1212 | /* aio_read_events_ring |
a31ad380 KO |
1213 | * Pull an event off of the ioctx's event ring. Returns the number of |
1214 | * events fetched | |
1da177e4 | 1215 | */ |
a31ad380 KO |
1216 | static long aio_read_events_ring(struct kioctx *ctx, |
1217 | struct io_event __user *event, long nr) | |
1da177e4 | 1218 | { |
1da177e4 | 1219 | struct aio_ring *ring; |
5ffac122 | 1220 | unsigned head, tail, pos; |
a31ad380 KO |
1221 | long ret = 0; |
1222 | int copy_ret; | |
1223 | ||
9c9ce763 DC |
1224 | /* |
1225 | * The mutex can block and wake us up and that will cause | |
1226 | * wait_event_interruptible_hrtimeout() to schedule without sleeping | |
1227 | * and repeat. This should be rare enough that it doesn't cause | |
1228 | * peformance issues. See the comment in read_events() for more detail. | |
1229 | */ | |
1230 | sched_annotate_sleep(); | |
58c85dc2 | 1231 | mutex_lock(&ctx->ring_lock); |
1da177e4 | 1232 | |
fa8a53c3 | 1233 | /* Access to ->ring_pages here is protected by ctx->ring_lock. */ |
58c85dc2 | 1234 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 1235 | head = ring->head; |
5ffac122 | 1236 | tail = ring->tail; |
a31ad380 KO |
1237 | kunmap_atomic(ring); |
1238 | ||
2ff396be JM |
1239 | /* |
1240 | * Ensure that once we've read the current tail pointer, that | |
1241 | * we also see the events that were stored up to the tail. | |
1242 | */ | |
1243 | smp_rmb(); | |
1244 | ||
5ffac122 | 1245 | pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events); |
1da177e4 | 1246 | |
5ffac122 | 1247 | if (head == tail) |
1da177e4 LT |
1248 | goto out; |
1249 | ||
edfbbf38 BL |
1250 | head %= ctx->nr_events; |
1251 | tail %= ctx->nr_events; | |
1252 | ||
a31ad380 KO |
1253 | while (ret < nr) { |
1254 | long avail; | |
1255 | struct io_event *ev; | |
1256 | struct page *page; | |
1257 | ||
5ffac122 KO |
1258 | avail = (head <= tail ? tail : ctx->nr_events) - head; |
1259 | if (head == tail) | |
a31ad380 KO |
1260 | break; |
1261 | ||
1262 | avail = min(avail, nr - ret); | |
1263 | avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - | |
1264 | ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE)); | |
1265 | ||
1266 | pos = head + AIO_EVENTS_OFFSET; | |
58c85dc2 | 1267 | page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]; |
a31ad380 KO |
1268 | pos %= AIO_EVENTS_PER_PAGE; |
1269 | ||
1270 | ev = kmap(page); | |
1271 | copy_ret = copy_to_user(event + ret, ev + pos, | |
1272 | sizeof(*ev) * avail); | |
1273 | kunmap(page); | |
1274 | ||
1275 | if (unlikely(copy_ret)) { | |
1276 | ret = -EFAULT; | |
1277 | goto out; | |
1278 | } | |
1279 | ||
1280 | ret += avail; | |
1281 | head += avail; | |
58c85dc2 | 1282 | head %= ctx->nr_events; |
1da177e4 | 1283 | } |
1da177e4 | 1284 | |
58c85dc2 | 1285 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 1286 | ring->head = head; |
91d80a84 | 1287 | kunmap_atomic(ring); |
58c85dc2 | 1288 | flush_dcache_page(ctx->ring_pages[0]); |
a31ad380 | 1289 | |
5ffac122 | 1290 | pr_debug("%li h%u t%u\n", ret, head, tail); |
a31ad380 | 1291 | out: |
58c85dc2 | 1292 | mutex_unlock(&ctx->ring_lock); |
a31ad380 | 1293 | |
1da177e4 LT |
1294 | return ret; |
1295 | } | |
1296 | ||
a31ad380 KO |
1297 | static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr, |
1298 | struct io_event __user *event, long *i) | |
1da177e4 | 1299 | { |
a31ad380 | 1300 | long ret = aio_read_events_ring(ctx, event + *i, nr - *i); |
1da177e4 | 1301 | |
a31ad380 KO |
1302 | if (ret > 0) |
1303 | *i += ret; | |
1da177e4 | 1304 | |
a31ad380 KO |
1305 | if (unlikely(atomic_read(&ctx->dead))) |
1306 | ret = -EINVAL; | |
1da177e4 | 1307 | |
a31ad380 KO |
1308 | if (!*i) |
1309 | *i = ret; | |
1da177e4 | 1310 | |
a31ad380 | 1311 | return ret < 0 || *i >= min_nr; |
1da177e4 LT |
1312 | } |
1313 | ||
a31ad380 | 1314 | static long read_events(struct kioctx *ctx, long min_nr, long nr, |
1da177e4 | 1315 | struct io_event __user *event, |
fa2e62a5 | 1316 | ktime_t until) |
1da177e4 | 1317 | { |
a31ad380 | 1318 | long ret = 0; |
1da177e4 | 1319 | |
a31ad380 KO |
1320 | /* |
1321 | * Note that aio_read_events() is being called as the conditional - i.e. | |
1322 | * we're calling it after prepare_to_wait() has set task state to | |
1323 | * TASK_INTERRUPTIBLE. | |
1324 | * | |
1325 | * But aio_read_events() can block, and if it blocks it's going to flip | |
1326 | * the task state back to TASK_RUNNING. | |
1327 | * | |
1328 | * This should be ok, provided it doesn't flip the state back to | |
1329 | * TASK_RUNNING and return 0 too much - that causes us to spin. That | |
1330 | * will only happen if the mutex_lock() call blocks, and we then find | |
1331 | * the ringbuffer empty. So in practice we should be ok, but it's | |
1332 | * something to be aware of when touching this code. | |
1333 | */ | |
2456e855 | 1334 | if (until == 0) |
5f785de5 FZ |
1335 | aio_read_events(ctx, min_nr, nr, event, &ret); |
1336 | else | |
1337 | wait_event_interruptible_hrtimeout(ctx->wait, | |
1338 | aio_read_events(ctx, min_nr, nr, event, &ret), | |
1339 | until); | |
1da177e4 | 1340 | |
a31ad380 KO |
1341 | if (!ret && signal_pending(current)) |
1342 | ret = -EINTR; | |
1da177e4 | 1343 | |
a31ad380 | 1344 | return ret; |
1da177e4 LT |
1345 | } |
1346 | ||
1da177e4 LT |
1347 | /* sys_io_setup: |
1348 | * Create an aio_context capable of receiving at least nr_events. | |
1349 | * ctxp must not point to an aio_context that already exists, and | |
1350 | * must be initialized to 0 prior to the call. On successful | |
1351 | * creation of the aio_context, *ctxp is filled in with the resulting | |
1352 | * handle. May fail with -EINVAL if *ctxp is not initialized, | |
1353 | * if the specified nr_events exceeds internal limits. May fail | |
1354 | * with -EAGAIN if the specified nr_events exceeds the user's limit | |
1355 | * of available events. May fail with -ENOMEM if insufficient kernel | |
1356 | * resources are available. May fail with -EFAULT if an invalid | |
1357 | * pointer is passed for ctxp. Will fail with -ENOSYS if not | |
1358 | * implemented. | |
1359 | */ | |
002c8976 | 1360 | SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) |
1da177e4 LT |
1361 | { |
1362 | struct kioctx *ioctx = NULL; | |
1363 | unsigned long ctx; | |
1364 | long ret; | |
1365 | ||
1366 | ret = get_user(ctx, ctxp); | |
1367 | if (unlikely(ret)) | |
1368 | goto out; | |
1369 | ||
1370 | ret = -EINVAL; | |
d55b5fda | 1371 | if (unlikely(ctx || nr_events == 0)) { |
acd88d4e | 1372 | pr_debug("EINVAL: ctx %lu nr_events %u\n", |
d55b5fda | 1373 | ctx, nr_events); |
1da177e4 LT |
1374 | goto out; |
1375 | } | |
1376 | ||
1377 | ioctx = ioctx_alloc(nr_events); | |
1378 | ret = PTR_ERR(ioctx); | |
1379 | if (!IS_ERR(ioctx)) { | |
1380 | ret = put_user(ioctx->user_id, ctxp); | |
a2e1859a | 1381 | if (ret) |
e02ba72a | 1382 | kill_ioctx(current->mm, ioctx, NULL); |
723be6e3 | 1383 | percpu_ref_put(&ioctx->users); |
1da177e4 LT |
1384 | } |
1385 | ||
1386 | out: | |
1387 | return ret; | |
1388 | } | |
1389 | ||
c00d2c7e AV |
1390 | #ifdef CONFIG_COMPAT |
1391 | COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p) | |
1392 | { | |
1393 | struct kioctx *ioctx = NULL; | |
1394 | unsigned long ctx; | |
1395 | long ret; | |
1396 | ||
1397 | ret = get_user(ctx, ctx32p); | |
1398 | if (unlikely(ret)) | |
1399 | goto out; | |
1400 | ||
1401 | ret = -EINVAL; | |
1402 | if (unlikely(ctx || nr_events == 0)) { | |
1403 | pr_debug("EINVAL: ctx %lu nr_events %u\n", | |
1404 | ctx, nr_events); | |
1405 | goto out; | |
1406 | } | |
1407 | ||
1408 | ioctx = ioctx_alloc(nr_events); | |
1409 | ret = PTR_ERR(ioctx); | |
1410 | if (!IS_ERR(ioctx)) { | |
1411 | /* truncating is ok because it's a user address */ | |
1412 | ret = put_user((u32)ioctx->user_id, ctx32p); | |
1413 | if (ret) | |
1414 | kill_ioctx(current->mm, ioctx, NULL); | |
1415 | percpu_ref_put(&ioctx->users); | |
1416 | } | |
1417 | ||
1418 | out: | |
1419 | return ret; | |
1420 | } | |
1421 | #endif | |
1422 | ||
1da177e4 LT |
1423 | /* sys_io_destroy: |
1424 | * Destroy the aio_context specified. May cancel any outstanding | |
1425 | * AIOs and block on completion. Will fail with -ENOSYS if not | |
642b5123 | 1426 | * implemented. May fail with -EINVAL if the context pointed to |
1da177e4 LT |
1427 | * is invalid. |
1428 | */ | |
002c8976 | 1429 | SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) |
1da177e4 LT |
1430 | { |
1431 | struct kioctx *ioctx = lookup_ioctx(ctx); | |
1432 | if (likely(NULL != ioctx)) { | |
dc48e56d | 1433 | struct ctx_rq_wait wait; |
fb2d4483 | 1434 | int ret; |
e02ba72a | 1435 | |
dc48e56d JA |
1436 | init_completion(&wait.comp); |
1437 | atomic_set(&wait.count, 1); | |
1438 | ||
e02ba72a AP |
1439 | /* Pass requests_done to kill_ioctx() where it can be set |
1440 | * in a thread-safe way. If we try to set it here then we have | |
1441 | * a race condition if two io_destroy() called simultaneously. | |
1442 | */ | |
dc48e56d | 1443 | ret = kill_ioctx(current->mm, ioctx, &wait); |
723be6e3 | 1444 | percpu_ref_put(&ioctx->users); |
e02ba72a AP |
1445 | |
1446 | /* Wait until all IO for the context are done. Otherwise kernel | |
1447 | * keep using user-space buffers even if user thinks the context | |
1448 | * is destroyed. | |
1449 | */ | |
fb2d4483 | 1450 | if (!ret) |
dc48e56d | 1451 | wait_for_completion(&wait.comp); |
e02ba72a | 1452 | |
fb2d4483 | 1453 | return ret; |
1da177e4 | 1454 | } |
acd88d4e | 1455 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1456 | return -EINVAL; |
1457 | } | |
1458 | ||
89319d31 CH |
1459 | static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec, |
1460 | bool vectored, bool compat, struct iov_iter *iter) | |
eed4e51f | 1461 | { |
89319d31 CH |
1462 | void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf; |
1463 | size_t len = iocb->aio_nbytes; | |
1464 | ||
1465 | if (!vectored) { | |
1466 | ssize_t ret = import_single_range(rw, buf, len, *iovec, iter); | |
1467 | *iovec = NULL; | |
1468 | return ret; | |
1469 | } | |
9d85cba7 JM |
1470 | #ifdef CONFIG_COMPAT |
1471 | if (compat) | |
89319d31 CH |
1472 | return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec, |
1473 | iter); | |
9d85cba7 | 1474 | #endif |
89319d31 | 1475 | return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter); |
eed4e51f BP |
1476 | } |
1477 | ||
89319d31 CH |
1478 | static inline ssize_t aio_ret(struct kiocb *req, ssize_t ret) |
1479 | { | |
1480 | switch (ret) { | |
1481 | case -EIOCBQUEUED: | |
1482 | return ret; | |
1483 | case -ERESTARTSYS: | |
1484 | case -ERESTARTNOINTR: | |
1485 | case -ERESTARTNOHAND: | |
1486 | case -ERESTART_RESTARTBLOCK: | |
1487 | /* | |
1488 | * There's no easy way to restart the syscall since other AIO's | |
1489 | * may be already running. Just fail this IO with EINTR. | |
1490 | */ | |
1491 | ret = -EINTR; | |
1492 | /*FALLTHRU*/ | |
1493 | default: | |
1494 | aio_complete(req, ret, 0); | |
1495 | return 0; | |
1496 | } | |
1497 | } | |
1498 | ||
1499 | static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored, | |
1500 | bool compat) | |
1da177e4 | 1501 | { |
41ef4eb8 | 1502 | struct file *file = req->ki_filp; |
00fefb9c | 1503 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; |
293bc982 | 1504 | struct iov_iter iter; |
89319d31 | 1505 | ssize_t ret; |
1da177e4 | 1506 | |
89319d31 CH |
1507 | if (unlikely(!(file->f_mode & FMODE_READ))) |
1508 | return -EBADF; | |
1509 | if (unlikely(!file->f_op->read_iter)) | |
1510 | return -EINVAL; | |
73a7075e | 1511 | |
89319d31 CH |
1512 | ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter); |
1513 | if (ret) | |
1514 | return ret; | |
1515 | ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter)); | |
1516 | if (!ret) | |
bb7462b6 | 1517 | ret = aio_ret(req, call_read_iter(file, req, &iter)); |
89319d31 CH |
1518 | kfree(iovec); |
1519 | return ret; | |
1520 | } | |
73a7075e | 1521 | |
89319d31 CH |
1522 | static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored, |
1523 | bool compat) | |
1524 | { | |
1525 | struct file *file = req->ki_filp; | |
1526 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1527 | struct iov_iter iter; | |
1528 | ssize_t ret; | |
41ef4eb8 | 1529 | |
89319d31 CH |
1530 | if (unlikely(!(file->f_mode & FMODE_WRITE))) |
1531 | return -EBADF; | |
1532 | if (unlikely(!file->f_op->write_iter)) | |
41ef4eb8 | 1533 | return -EINVAL; |
1da177e4 | 1534 | |
89319d31 CH |
1535 | ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter); |
1536 | if (ret) | |
1537 | return ret; | |
1538 | ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter)); | |
1539 | if (!ret) { | |
70fe2f48 | 1540 | req->ki_flags |= IOCB_WRITE; |
89319d31 | 1541 | file_start_write(file); |
bb7462b6 | 1542 | ret = aio_ret(req, call_write_iter(file, req, &iter)); |
70fe2f48 JK |
1543 | /* |
1544 | * We release freeze protection in aio_complete(). Fool lockdep | |
1545 | * by telling it the lock got released so that it doesn't | |
1546 | * complain about held lock when we return to userspace. | |
1547 | */ | |
a12f1ae6 SL |
1548 | if (S_ISREG(file_inode(file)->i_mode)) |
1549 | __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE); | |
41ef4eb8 | 1550 | } |
89319d31 CH |
1551 | kfree(iovec); |
1552 | return ret; | |
1da177e4 LT |
1553 | } |
1554 | ||
d5470b59 | 1555 | static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, |
a1c8eae7 | 1556 | struct iocb *iocb, bool compat) |
1da177e4 | 1557 | { |
04b2fa9f | 1558 | struct aio_kiocb *req; |
89319d31 | 1559 | struct file *file; |
1da177e4 LT |
1560 | ssize_t ret; |
1561 | ||
1562 | /* enforce forwards compatibility on users */ | |
9830f4be | 1563 | if (unlikely(iocb->aio_reserved2)) { |
caf4167a | 1564 | pr_debug("EINVAL: reserve field set\n"); |
1da177e4 LT |
1565 | return -EINVAL; |
1566 | } | |
1567 | ||
1568 | /* prevent overflows */ | |
1569 | if (unlikely( | |
1570 | (iocb->aio_buf != (unsigned long)iocb->aio_buf) || | |
1571 | (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || | |
1572 | ((ssize_t)iocb->aio_nbytes < 0) | |
1573 | )) { | |
acd88d4e | 1574 | pr_debug("EINVAL: overflow check\n"); |
1da177e4 LT |
1575 | return -EINVAL; |
1576 | } | |
1577 | ||
41ef4eb8 | 1578 | req = aio_get_req(ctx); |
1d98ebfc | 1579 | if (unlikely(!req)) |
1da177e4 | 1580 | return -EAGAIN; |
1d98ebfc | 1581 | |
89319d31 | 1582 | req->common.ki_filp = file = fget(iocb->aio_fildes); |
04b2fa9f | 1583 | if (unlikely(!req->common.ki_filp)) { |
1d98ebfc KO |
1584 | ret = -EBADF; |
1585 | goto out_put_req; | |
1da177e4 | 1586 | } |
04b2fa9f CH |
1587 | req->common.ki_pos = iocb->aio_offset; |
1588 | req->common.ki_complete = aio_complete; | |
2ba48ce5 | 1589 | req->common.ki_flags = iocb_flags(req->common.ki_filp); |
45d06cf7 | 1590 | req->common.ki_hint = file_write_hint(file); |
1d98ebfc | 1591 | |
9c3060be DL |
1592 | if (iocb->aio_flags & IOCB_FLAG_RESFD) { |
1593 | /* | |
1594 | * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an | |
1595 | * instance of the file* now. The file descriptor must be | |
1596 | * an eventfd() fd, and will be signaled for each completed | |
1597 | * event using the eventfd_signal() function. | |
1598 | */ | |
13389010 | 1599 | req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); |
801678c5 | 1600 | if (IS_ERR(req->ki_eventfd)) { |
9c3060be | 1601 | ret = PTR_ERR(req->ki_eventfd); |
87c3a86e | 1602 | req->ki_eventfd = NULL; |
9c3060be DL |
1603 | goto out_put_req; |
1604 | } | |
04b2fa9f CH |
1605 | |
1606 | req->common.ki_flags |= IOCB_EVENTFD; | |
9c3060be | 1607 | } |
1da177e4 | 1608 | |
9830f4be GR |
1609 | ret = kiocb_set_rw_flags(&req->common, iocb->aio_rw_flags); |
1610 | if (unlikely(ret)) { | |
1611 | pr_debug("EINVAL: aio_rw_flags\n"); | |
1612 | goto out_put_req; | |
1613 | } | |
1614 | ||
8a660890 | 1615 | ret = put_user(KIOCB_KEY, &user_iocb->aio_key); |
1da177e4 | 1616 | if (unlikely(ret)) { |
caf4167a | 1617 | pr_debug("EFAULT: aio_key\n"); |
1da177e4 LT |
1618 | goto out_put_req; |
1619 | } | |
1620 | ||
04b2fa9f | 1621 | req->ki_user_iocb = user_iocb; |
1da177e4 | 1622 | req->ki_user_data = iocb->aio_data; |
1da177e4 | 1623 | |
89319d31 CH |
1624 | get_file(file); |
1625 | switch (iocb->aio_lio_opcode) { | |
1626 | case IOCB_CMD_PREAD: | |
1627 | ret = aio_read(&req->common, iocb, false, compat); | |
1628 | break; | |
1629 | case IOCB_CMD_PWRITE: | |
1630 | ret = aio_write(&req->common, iocb, false, compat); | |
1631 | break; | |
1632 | case IOCB_CMD_PREADV: | |
1633 | ret = aio_read(&req->common, iocb, true, compat); | |
1634 | break; | |
1635 | case IOCB_CMD_PWRITEV: | |
1636 | ret = aio_write(&req->common, iocb, true, compat); | |
1637 | break; | |
1638 | default: | |
1639 | pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode); | |
1640 | ret = -EINVAL; | |
1641 | break; | |
1642 | } | |
1643 | fput(file); | |
41003a7b | 1644 | |
89319d31 CH |
1645 | if (ret && ret != -EIOCBQUEUED) |
1646 | goto out_put_req; | |
1da177e4 | 1647 | return 0; |
1da177e4 | 1648 | out_put_req: |
e1bdd5f2 | 1649 | put_reqs_available(ctx, 1); |
e34ecee2 | 1650 | percpu_ref_put(&ctx->reqs); |
57282d8f | 1651 | kiocb_free(req); |
1da177e4 LT |
1652 | return ret; |
1653 | } | |
1654 | ||
c00d2c7e AV |
1655 | static long do_io_submit(aio_context_t ctx_id, long nr, |
1656 | struct iocb __user *__user *iocbpp, bool compat) | |
1da177e4 LT |
1657 | { |
1658 | struct kioctx *ctx; | |
1659 | long ret = 0; | |
080d676d | 1660 | int i = 0; |
9f5b9425 | 1661 | struct blk_plug plug; |
1da177e4 LT |
1662 | |
1663 | if (unlikely(nr < 0)) | |
1664 | return -EINVAL; | |
1665 | ||
75e1c70f JM |
1666 | if (unlikely(nr > LONG_MAX/sizeof(*iocbpp))) |
1667 | nr = LONG_MAX/sizeof(*iocbpp); | |
1668 | ||
1da177e4 LT |
1669 | if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) |
1670 | return -EFAULT; | |
1671 | ||
1672 | ctx = lookup_ioctx(ctx_id); | |
1673 | if (unlikely(!ctx)) { | |
caf4167a | 1674 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1675 | return -EINVAL; |
1676 | } | |
1677 | ||
9f5b9425 SL |
1678 | blk_start_plug(&plug); |
1679 | ||
1da177e4 LT |
1680 | /* |
1681 | * AKPM: should this return a partial result if some of the IOs were | |
1682 | * successfully submitted? | |
1683 | */ | |
1684 | for (i=0; i<nr; i++) { | |
1685 | struct iocb __user *user_iocb; | |
1686 | struct iocb tmp; | |
1687 | ||
1688 | if (unlikely(__get_user(user_iocb, iocbpp + i))) { | |
1689 | ret = -EFAULT; | |
1690 | break; | |
1691 | } | |
1692 | ||
1693 | if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { | |
1694 | ret = -EFAULT; | |
1695 | break; | |
1696 | } | |
1697 | ||
a1c8eae7 | 1698 | ret = io_submit_one(ctx, user_iocb, &tmp, compat); |
1da177e4 LT |
1699 | if (ret) |
1700 | break; | |
1701 | } | |
9f5b9425 | 1702 | blk_finish_plug(&plug); |
1da177e4 | 1703 | |
723be6e3 | 1704 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1705 | return i ? i : ret; |
1706 | } | |
1707 | ||
9d85cba7 JM |
1708 | /* sys_io_submit: |
1709 | * Queue the nr iocbs pointed to by iocbpp for processing. Returns | |
1710 | * the number of iocbs queued. May return -EINVAL if the aio_context | |
1711 | * specified by ctx_id is invalid, if nr is < 0, if the iocb at | |
1712 | * *iocbpp[0] is not properly initialized, if the operation specified | |
1713 | * is invalid for the file descriptor in the iocb. May fail with | |
1714 | * -EFAULT if any of the data structures point to invalid data. May | |
1715 | * fail with -EBADF if the file descriptor specified in the first | |
1716 | * iocb is invalid. May fail with -EAGAIN if insufficient resources | |
1717 | * are available to queue any iocbs. Will return 0 if nr is 0. Will | |
1718 | * fail with -ENOSYS if not implemented. | |
1719 | */ | |
1720 | SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, | |
1721 | struct iocb __user * __user *, iocbpp) | |
1722 | { | |
1723 | return do_io_submit(ctx_id, nr, iocbpp, 0); | |
1724 | } | |
1725 | ||
c00d2c7e AV |
1726 | #ifdef CONFIG_COMPAT |
1727 | static inline long | |
1728 | copy_iocb(long nr, u32 __user *ptr32, struct iocb __user * __user *ptr64) | |
1729 | { | |
1730 | compat_uptr_t uptr; | |
1731 | int i; | |
1732 | ||
1733 | for (i = 0; i < nr; ++i) { | |
1734 | if (get_user(uptr, ptr32 + i)) | |
1735 | return -EFAULT; | |
1736 | if (put_user(compat_ptr(uptr), ptr64 + i)) | |
1737 | return -EFAULT; | |
1738 | } | |
1739 | return 0; | |
1740 | } | |
1741 | ||
1742 | #define MAX_AIO_SUBMITS (PAGE_SIZE/sizeof(struct iocb *)) | |
1743 | ||
1744 | COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id, | |
1745 | int, nr, u32 __user *, iocb) | |
1746 | { | |
1747 | struct iocb __user * __user *iocb64; | |
1748 | long ret; | |
1749 | ||
1750 | if (unlikely(nr < 0)) | |
1751 | return -EINVAL; | |
1752 | ||
1753 | if (nr > MAX_AIO_SUBMITS) | |
1754 | nr = MAX_AIO_SUBMITS; | |
1755 | ||
1756 | iocb64 = compat_alloc_user_space(nr * sizeof(*iocb64)); | |
1757 | ret = copy_iocb(nr, iocb, iocb64); | |
1758 | if (!ret) | |
1759 | ret = do_io_submit(ctx_id, nr, iocb64, 1); | |
1760 | return ret; | |
1761 | } | |
1762 | #endif | |
1763 | ||
1da177e4 LT |
1764 | /* lookup_kiocb |
1765 | * Finds a given iocb for cancellation. | |
1da177e4 | 1766 | */ |
04b2fa9f CH |
1767 | static struct aio_kiocb * |
1768 | lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key) | |
1da177e4 | 1769 | { |
04b2fa9f | 1770 | struct aio_kiocb *kiocb; |
d00689af ZB |
1771 | |
1772 | assert_spin_locked(&ctx->ctx_lock); | |
1773 | ||
8a660890 KO |
1774 | if (key != KIOCB_KEY) |
1775 | return NULL; | |
1776 | ||
1da177e4 | 1777 | /* TODO: use a hash or array, this sucks. */ |
04b2fa9f CH |
1778 | list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) { |
1779 | if (kiocb->ki_user_iocb == iocb) | |
1da177e4 LT |
1780 | return kiocb; |
1781 | } | |
1782 | return NULL; | |
1783 | } | |
1784 | ||
1785 | /* sys_io_cancel: | |
1786 | * Attempts to cancel an iocb previously passed to io_submit. If | |
1787 | * the operation is successfully cancelled, the resulting event is | |
1788 | * copied into the memory pointed to by result without being placed | |
1789 | * into the completion queue and 0 is returned. May fail with | |
1790 | * -EFAULT if any of the data structures pointed to are invalid. | |
1791 | * May fail with -EINVAL if aio_context specified by ctx_id is | |
1792 | * invalid. May fail with -EAGAIN if the iocb specified was not | |
1793 | * cancelled. Will fail with -ENOSYS if not implemented. | |
1794 | */ | |
002c8976 HC |
1795 | SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, |
1796 | struct io_event __user *, result) | |
1da177e4 | 1797 | { |
1da177e4 | 1798 | struct kioctx *ctx; |
04b2fa9f | 1799 | struct aio_kiocb *kiocb; |
1da177e4 LT |
1800 | u32 key; |
1801 | int ret; | |
1802 | ||
1803 | ret = get_user(key, &iocb->aio_key); | |
1804 | if (unlikely(ret)) | |
1805 | return -EFAULT; | |
1806 | ||
1807 | ctx = lookup_ioctx(ctx_id); | |
1808 | if (unlikely(!ctx)) | |
1809 | return -EINVAL; | |
1810 | ||
1811 | spin_lock_irq(&ctx->ctx_lock); | |
906b973c | 1812 | |
1da177e4 | 1813 | kiocb = lookup_kiocb(ctx, iocb, key); |
906b973c | 1814 | if (kiocb) |
d52a8f9e | 1815 | ret = kiocb_cancel(kiocb); |
906b973c KO |
1816 | else |
1817 | ret = -EINVAL; | |
1818 | ||
1da177e4 LT |
1819 | spin_unlock_irq(&ctx->ctx_lock); |
1820 | ||
906b973c | 1821 | if (!ret) { |
bec68faa KO |
1822 | /* |
1823 | * The result argument is no longer used - the io_event is | |
1824 | * always delivered via the ring buffer. -EINPROGRESS indicates | |
1825 | * cancellation is progress: | |
906b973c | 1826 | */ |
bec68faa | 1827 | ret = -EINPROGRESS; |
906b973c | 1828 | } |
1da177e4 | 1829 | |
723be6e3 | 1830 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1831 | |
1832 | return ret; | |
1833 | } | |
1834 | ||
fa2e62a5 DD |
1835 | static long do_io_getevents(aio_context_t ctx_id, |
1836 | long min_nr, | |
1837 | long nr, | |
1838 | struct io_event __user *events, | |
1839 | struct timespec64 *ts) | |
1840 | { | |
1841 | ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX; | |
1842 | struct kioctx *ioctx = lookup_ioctx(ctx_id); | |
1843 | long ret = -EINVAL; | |
1844 | ||
1845 | if (likely(ioctx)) { | |
1846 | if (likely(min_nr <= nr && min_nr >= 0)) | |
1847 | ret = read_events(ioctx, min_nr, nr, events, until); | |
1848 | percpu_ref_put(&ioctx->users); | |
1849 | } | |
1850 | ||
1851 | return ret; | |
1852 | } | |
1853 | ||
1da177e4 LT |
1854 | /* io_getevents: |
1855 | * Attempts to read at least min_nr events and up to nr events from | |
642b5123 ST |
1856 | * the completion queue for the aio_context specified by ctx_id. If |
1857 | * it succeeds, the number of read events is returned. May fail with | |
1858 | * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is | |
1859 | * out of range, if timeout is out of range. May fail with -EFAULT | |
1860 | * if any of the memory specified is invalid. May return 0 or | |
1861 | * < min_nr if the timeout specified by timeout has elapsed | |
1862 | * before sufficient events are available, where timeout == NULL | |
1863 | * specifies an infinite timeout. Note that the timeout pointed to by | |
6900807c | 1864 | * timeout is relative. Will fail with -ENOSYS if not implemented. |
1da177e4 | 1865 | */ |
002c8976 HC |
1866 | SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, |
1867 | long, min_nr, | |
1868 | long, nr, | |
1869 | struct io_event __user *, events, | |
1870 | struct timespec __user *, timeout) | |
1da177e4 | 1871 | { |
fa2e62a5 | 1872 | struct timespec64 ts; |
1da177e4 | 1873 | |
fa2e62a5 DD |
1874 | if (timeout) { |
1875 | if (unlikely(get_timespec64(&ts, timeout))) | |
1876 | return -EFAULT; | |
1da177e4 | 1877 | } |
fa2e62a5 DD |
1878 | |
1879 | return do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL); | |
1da177e4 | 1880 | } |
c00d2c7e AV |
1881 | |
1882 | #ifdef CONFIG_COMPAT | |
1883 | COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id, | |
1884 | compat_long_t, min_nr, | |
1885 | compat_long_t, nr, | |
1886 | struct io_event __user *, events, | |
1887 | struct compat_timespec __user *, timeout) | |
1888 | { | |
fa2e62a5 | 1889 | struct timespec64 t; |
c00d2c7e AV |
1890 | |
1891 | if (timeout) { | |
fa2e62a5 | 1892 | if (compat_get_timespec64(&t, timeout)) |
c00d2c7e AV |
1893 | return -EFAULT; |
1894 | ||
c00d2c7e | 1895 | } |
fa2e62a5 DD |
1896 | |
1897 | return do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL); | |
c00d2c7e AV |
1898 | } |
1899 | #endif |