blob: 6b868f0e0c4c019c3b68712ca230deb979711a98 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 */
Kent Overstreetcaf41672013-05-07 16:18:35 -070011#define pr_fmt(fmt) "%s: " fmt, __func__
12
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#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>
Paul Gortmaker630d9c42011-11-16 23:57:37 -050018#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/syscalls.h>
Jens Axboeb9d128f2009-10-29 13:59:26 +010020#include <linux/backing-dev.h>
Badari Pulavarty027445c2006-09-30 23:28:46 -070021#include <linux/uio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070022
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/sched.h>
24#include <linux/fs.h>
25#include <linux/file.h>
26#include <linux/mm.h>
27#include <linux/mman.h>
Michael S. Tsirkin3d2d8272009-09-21 17:03:51 -070028#include <linux/mmu_context.h>
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100029#include <linux/percpu.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#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>
Davide Libenzi9c3060b2007-05-10 22:23:21 -070036#include <linux/eventfd.h>
Jeff Moyercfb1e332009-10-02 18:57:36 -040037#include <linux/blkdev.h>
Jeff Moyer9d85cba2010-05-26 14:44:26 -070038#include <linux/compat.h>
Gu Zheng36bc08c2013-07-16 17:56:16 +080039#include <linux/anon_inodes.h>
40#include <linux/migrate.h>
41#include <linux/ramfs.h>
Kent Overstreet723be6e2013-05-28 15:14:48 -070042#include <linux/percpu-refcount.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44#include <asm/kmap_types.h>
45#include <asm/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
Al Viro68d70d02013-06-19 15:26:04 +040047#include "internal.h"
48
Kent Overstreet4e179bc2013-05-07 16:18:33 -070049#define AIO_RING_MAGIC 0xa10a10a1
50#define AIO_RING_COMPAT_FEATURES 1
51#define AIO_RING_INCOMPAT_FEATURES 0
52struct 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
Kent Overstreet4e179bc2013-05-07 16:18:33 -070068
Benjamin LaHaisedb446a02013-07-30 12:54:40 -040069struct kioctx_table {
70 struct rcu_head rcu;
71 unsigned nr;
72 struct kioctx *table[];
73};
74
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100075struct kioctx_cpu {
76 unsigned reqs_available;
77};
78
Kent Overstreet4e179bc2013-05-07 16:18:33 -070079struct kioctx {
Kent Overstreet723be6e2013-05-28 15:14:48 -070080 struct percpu_ref users;
Kent Overstreet36f55882013-05-07 16:18:41 -070081 atomic_t dead;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070082
Kent Overstreet4e179bc2013-05-07 16:18:33 -070083 unsigned long user_id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070084
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100085 struct __percpu kioctx_cpu *cpu;
86
87 /*
88 * For percpu reqs_available, number of slots we move to/from global
89 * counter at a time:
90 */
91 unsigned req_batch;
Kent Overstreet3e845ce2013-05-07 16:18:51 -070092 /*
93 * This is what userspace passed to io_setup(), it's not used for
94 * anything but counting against the global max_reqs quota.
95 *
Kent Overstreet58c85dc2013-05-07 16:18:55 -070096 * The real limit is nr_events - 1, which will be larger (see
Kent Overstreet3e845ce2013-05-07 16:18:51 -070097 * aio_setup_ring())
98 */
Kent Overstreet4e179bc2013-05-07 16:18:33 -070099 unsigned max_reqs;
100
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700101 /* Size of ringbuffer, in units of struct io_event */
102 unsigned nr_events;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700103
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700104 unsigned long mmap_base;
105 unsigned long mmap_size;
106
107 struct page **ring_pages;
108 long nr_pages;
109
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700110 struct rcu_head rcu_head;
Kent Overstreet723be6e2013-05-28 15:14:48 -0700111 struct work_struct free_work;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700112
113 struct {
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000114 /*
115 * This counts the number of available slots in the ringbuffer,
116 * so we avoid overflowing it: it's decremented (if positive)
117 * when allocating a kiocb and incremented when the resulting
118 * io_event is pulled off the ringbuffer.
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000119 *
120 * We batch accesses to it with a percpu version.
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000121 */
122 atomic_t reqs_available;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700123 } ____cacheline_aligned_in_smp;
124
125 struct {
126 spinlock_t ctx_lock;
127 struct list_head active_reqs; /* used for cancellation */
128 } ____cacheline_aligned_in_smp;
129
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700130 struct {
131 struct mutex ring_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700132 wait_queue_head_t wait;
133 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700134
135 struct {
136 unsigned tail;
137 spinlock_t completion_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700138 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700139
140 struct page *internal_pages[AIO_RING_PAGES];
Gu Zheng36bc08c2013-07-16 17:56:16 +0800141 struct file *aio_ring_file;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400142
143 unsigned id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700144};
145
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -0800147static DEFINE_SPINLOCK(aio_nr_lock);
148unsigned long aio_nr; /* current system wide number of aio requests */
149unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150/*----end sysctl variables---*/
151
Christoph Lametere18b8902006-12-06 20:33:20 -0800152static struct kmem_cache *kiocb_cachep;
153static struct kmem_cache *kioctx_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155/* aio_setup
156 * Creates the slab caches used by the aio routines, panic on
157 * failure as this is done early during the boot sequence.
158 */
159static int __init aio_setup(void)
160{
Christoph Lameter0a31bd52007-05-06 14:49:57 -0700161 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
162 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163
Kent Overstreetcaf41672013-05-07 16:18:35 -0700164 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165
166 return 0;
167}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700168__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169
170static void aio_free_ring(struct kioctx *ctx)
171{
Gu Zheng36bc08c2013-07-16 17:56:16 +0800172 int i;
173 struct file *aio_ring_file = ctx->aio_ring_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174
Gu Zheng36bc08c2013-07-16 17:56:16 +0800175 for (i = 0; i < ctx->nr_pages; i++) {
176 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
177 page_count(ctx->ring_pages[i]));
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700178 put_page(ctx->ring_pages[i]);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800179 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700181 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
182 kfree(ctx->ring_pages);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800183
184 if (aio_ring_file) {
185 truncate_setsize(aio_ring_file->f_inode, 0);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800186 fput(aio_ring_file);
187 ctx->aio_ring_file = NULL;
188 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189}
190
Gu Zheng36bc08c2013-07-16 17:56:16 +0800191static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
192{
193 vma->vm_ops = &generic_file_vm_ops;
194 return 0;
195}
196
197static const struct file_operations aio_ring_fops = {
198 .mmap = aio_ring_mmap,
199};
200
201static int aio_set_page_dirty(struct page *page)
202{
203 return 0;
204}
205
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400206#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800207static int aio_migratepage(struct address_space *mapping, struct page *new,
208 struct page *old, enum migrate_mode mode)
209{
210 struct kioctx *ctx = mapping->private_data;
211 unsigned long flags;
212 unsigned idx = old->index;
213 int rc;
214
215 /* Writeback must be complete */
216 BUG_ON(PageWriteback(old));
217 put_page(old);
218
219 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode);
220 if (rc != MIGRATEPAGE_SUCCESS) {
221 get_page(old);
222 return rc;
223 }
224
225 get_page(new);
226
227 spin_lock_irqsave(&ctx->completion_lock, flags);
228 migrate_page_copy(new, old);
229 ctx->ring_pages[idx] = new;
230 spin_unlock_irqrestore(&ctx->completion_lock, flags);
231
232 return rc;
233}
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400234#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800235
236static const struct address_space_operations aio_ctx_aops = {
237 .set_page_dirty = aio_set_page_dirty,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400238#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800239 .migratepage = aio_migratepage,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400240#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800241};
242
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243static int aio_setup_ring(struct kioctx *ctx)
244{
245 struct aio_ring *ring;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246 unsigned nr_events = ctx->max_reqs;
Zach Brown41003a72013-05-07 16:18:25 -0700247 struct mm_struct *mm = current->mm;
Michel Lespinasse41badc12013-02-22 16:32:47 -0800248 unsigned long size, populate;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 int nr_pages;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800250 int i;
251 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252
253 /* Compensate for the ring buffer's head/tail overlap entry */
254 nr_events += 2; /* 1 is required, 2 for good luck */
255
256 size = sizeof(struct aio_ring);
257 size += sizeof(struct io_event) * nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258
Gu Zheng36bc08c2013-07-16 17:56:16 +0800259 nr_pages = PFN_UP(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 if (nr_pages < 0)
261 return -EINVAL;
262
Gu Zheng36bc08c2013-07-16 17:56:16 +0800263 file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR);
264 if (IS_ERR(file)) {
265 ctx->aio_ring_file = NULL;
266 return -EAGAIN;
267 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
Gu Zheng36bc08c2013-07-16 17:56:16 +0800269 file->f_inode->i_mapping->a_ops = &aio_ctx_aops;
270 file->f_inode->i_mapping->private_data = ctx;
271 file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages;
272
273 for (i = 0; i < nr_pages; i++) {
274 struct page *page;
275 page = find_or_create_page(file->f_inode->i_mapping,
276 i, GFP_HIGHUSER | __GFP_ZERO);
277 if (!page)
278 break;
279 pr_debug("pid(%d) page[%d]->count=%d\n",
280 current->pid, i, page_count(page));
281 SetPageUptodate(page);
282 SetPageDirty(page);
283 unlock_page(page);
284 }
285 ctx->aio_ring_file = file;
286 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
287 / sizeof(struct io_event);
288
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700289 ctx->ring_pages = ctx->internal_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 if (nr_pages > AIO_RING_PAGES) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700291 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
292 GFP_KERNEL);
293 if (!ctx->ring_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295 }
296
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700297 ctx->mmap_size = nr_pages * PAGE_SIZE;
298 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800299
Zach Brown41003a72013-05-07 16:18:25 -0700300 down_write(&mm->mmap_sem);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800301 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
302 PROT_READ | PROT_WRITE,
303 MAP_SHARED | MAP_POPULATE, 0, &populate);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700304 if (IS_ERR((void *)ctx->mmap_base)) {
Zach Brown41003a72013-05-07 16:18:25 -0700305 up_write(&mm->mmap_sem);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700306 ctx->mmap_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 aio_free_ring(ctx);
308 return -EAGAIN;
309 }
310
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700311 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400312
313 /* We must do this while still holding mmap_sem for write, as we
314 * need to be protected against userspace attempting to mremap()
315 * or munmap() the ring buffer.
316 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700317 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
318 1, 0, ctx->ring_pages, NULL);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400319
320 /* Dropping the reference here is safe as the page cache will hold
321 * onto the pages for us. It is also required so that page migration
322 * can unmap the pages and get the right reference count.
323 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800324 for (i = 0; i < ctx->nr_pages; i++)
325 put_page(ctx->ring_pages[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400327 up_write(&mm->mmap_sem);
328
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700329 if (unlikely(ctx->nr_pages != nr_pages)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330 aio_free_ring(ctx);
331 return -EAGAIN;
332 }
333
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700334 ctx->user_id = ctx->mmap_base;
335 ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700337 ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 ring->nr = nr_events; /* user copy */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400339 ring->id = ~0U;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 ring->head = ring->tail = 0;
341 ring->magic = AIO_RING_MAGIC;
342 ring->compat_features = AIO_RING_COMPAT_FEATURES;
343 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
344 ring->header_length = sizeof(struct aio_ring);
Cong Wange8e3c3d2011-11-25 23:14:27 +0800345 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700346 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347
348 return 0;
349}
350
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
352#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
353#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
354
Kent Overstreet0460fef2013-05-07 16:18:49 -0700355void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
356{
357 struct kioctx *ctx = req->ki_ctx;
358 unsigned long flags;
359
360 spin_lock_irqsave(&ctx->ctx_lock, flags);
361
362 if (!req->ki_list.next)
363 list_add(&req->ki_list, &ctx->active_reqs);
364
365 req->ki_cancel = cancel;
366
367 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
368}
369EXPORT_SYMBOL(kiocb_set_cancel_fn);
370
Kent Overstreetbec68faa2013-05-13 14:45:08 -0700371static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
Kent Overstreet906b9732013-05-07 16:18:31 -0700372{
Kent Overstreet0460fef2013-05-07 16:18:49 -0700373 kiocb_cancel_fn *old, *cancel;
Kent Overstreet906b9732013-05-07 16:18:31 -0700374
Kent Overstreet0460fef2013-05-07 16:18:49 -0700375 /*
376 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
377 * actually has a cancel function, hence the cmpxchg()
378 */
Kent Overstreet906b9732013-05-07 16:18:31 -0700379
Kent Overstreet0460fef2013-05-07 16:18:49 -0700380 cancel = ACCESS_ONCE(kiocb->ki_cancel);
381 do {
382 if (!cancel || cancel == KIOCB_CANCELLED)
Kent Overstreet57282d82013-05-13 13:42:52 -0700383 return -EINVAL;
Kent Overstreet906b9732013-05-07 16:18:31 -0700384
Kent Overstreet0460fef2013-05-07 16:18:49 -0700385 old = cancel;
386 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
387 } while (cancel != old);
388
Kent Overstreet57282d82013-05-13 13:42:52 -0700389 return cancel(kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -0700390}
391
Kent Overstreet36f55882013-05-07 16:18:41 -0700392static void free_ioctx_rcu(struct rcu_head *head)
393{
394 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000395
396 free_percpu(ctx->cpu);
Kent Overstreet36f55882013-05-07 16:18:41 -0700397 kmem_cache_free(kioctx_cachep, ctx);
398}
399
400/*
401 * When this function runs, the kioctx has been removed from the "hash table"
402 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
403 * now it's safe to cancel any that need to be.
404 */
Kent Overstreet723be6e2013-05-28 15:14:48 -0700405static void free_ioctx(struct work_struct *work)
Kent Overstreet36f55882013-05-07 16:18:41 -0700406{
Kent Overstreet723be6e2013-05-28 15:14:48 -0700407 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700408 struct aio_ring *ring;
Kent Overstreet36f55882013-05-07 16:18:41 -0700409 struct kiocb *req;
Kent Overstreet5ffac122013-05-09 15:36:07 -0700410 unsigned cpu, avail;
411 DEFINE_WAIT(wait);
Kent Overstreet36f55882013-05-07 16:18:41 -0700412
413 spin_lock_irq(&ctx->ctx_lock);
414
415 while (!list_empty(&ctx->active_reqs)) {
416 req = list_first_entry(&ctx->active_reqs,
417 struct kiocb, ki_list);
418
419 list_del_init(&req->ki_list);
Kent Overstreetbec68faa2013-05-13 14:45:08 -0700420 kiocb_cancel(ctx, req);
Kent Overstreet36f55882013-05-07 16:18:41 -0700421 }
422
423 spin_unlock_irq(&ctx->ctx_lock);
424
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000425 for_each_possible_cpu(cpu) {
426 struct kioctx_cpu *kcpu = per_cpu_ptr(ctx->cpu, cpu);
427
428 atomic_add(kcpu->reqs_available, &ctx->reqs_available);
429 kcpu->reqs_available = 0;
430 }
431
Kent Overstreet5ffac122013-05-09 15:36:07 -0700432 while (1) {
433 prepare_to_wait(&ctx->wait, &wait, TASK_UNINTERRUPTIBLE);
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700434
Kent Overstreet5ffac122013-05-09 15:36:07 -0700435 ring = kmap_atomic(ctx->ring_pages[0]);
436 avail = (ring->head <= ring->tail)
437 ? ring->tail - ring->head
438 : ctx->nr_events - ring->head + ring->tail;
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700439
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000440 atomic_add(avail, &ctx->reqs_available);
Kent Overstreet5ffac122013-05-09 15:36:07 -0700441 ring->head = ring->tail;
442 kunmap_atomic(ring);
443
444 if (atomic_read(&ctx->reqs_available) >= ctx->nr_events - 1)
445 break;
446
447 schedule();
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700448 }
Kent Overstreet5ffac122013-05-09 15:36:07 -0700449 finish_wait(&ctx->wait, &wait);
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700450
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000451 WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr_events - 1);
Kent Overstreet36f55882013-05-07 16:18:41 -0700452
453 aio_free_ring(ctx);
454
Kent Overstreet36f55882013-05-07 16:18:41 -0700455 pr_debug("freeing %p\n", ctx);
456
457 /*
458 * Here the call_rcu() is between the wait_event() for reqs_active to
459 * hit 0, and freeing the ioctx.
460 *
461 * aio_complete() decrements reqs_active, but it has to touch the ioctx
462 * after to issue a wakeup so we use rcu.
463 */
464 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
465}
466
Kent Overstreet723be6e2013-05-28 15:14:48 -0700467static void free_ioctx_ref(struct percpu_ref *ref)
Kent Overstreet36f55882013-05-07 16:18:41 -0700468{
Kent Overstreet723be6e2013-05-28 15:14:48 -0700469 struct kioctx *ctx = container_of(ref, struct kioctx, users);
470
471 INIT_WORK(&ctx->free_work, free_ioctx);
472 schedule_work(&ctx->free_work);
Kent Overstreet36f55882013-05-07 16:18:41 -0700473}
474
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400475static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
476{
477 unsigned i, new_nr;
478 struct kioctx_table *table, *old;
479 struct aio_ring *ring;
480
481 spin_lock(&mm->ioctx_lock);
Artem Savkovd9b2c872013-09-09 12:29:35 -0400482 rcu_read_lock();
Benjamin LaHaise77d30b12013-08-30 11:12:50 -0400483 table = rcu_dereference(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400484
485 while (1) {
486 if (table)
487 for (i = 0; i < table->nr; i++)
488 if (!table->table[i]) {
489 ctx->id = i;
490 table->table[i] = ctx;
Artem Savkovd9b2c872013-09-09 12:29:35 -0400491 rcu_read_unlock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400492 spin_unlock(&mm->ioctx_lock);
493
494 ring = kmap_atomic(ctx->ring_pages[0]);
495 ring->id = ctx->id;
496 kunmap_atomic(ring);
497 return 0;
498 }
499
500 new_nr = (table ? table->nr : 1) * 4;
501
Artem Savkovd9b2c872013-09-09 12:29:35 -0400502 rcu_read_unlock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400503 spin_unlock(&mm->ioctx_lock);
504
505 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
506 new_nr, GFP_KERNEL);
507 if (!table)
508 return -ENOMEM;
509
510 table->nr = new_nr;
511
512 spin_lock(&mm->ioctx_lock);
Artem Savkovd9b2c872013-09-09 12:29:35 -0400513 rcu_read_lock();
Benjamin LaHaise77d30b12013-08-30 11:12:50 -0400514 old = rcu_dereference(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400515
516 if (!old) {
517 rcu_assign_pointer(mm->ioctx_table, table);
518 } else if (table->nr > old->nr) {
519 memcpy(table->table, old->table,
520 old->nr * sizeof(struct kioctx *));
521
522 rcu_assign_pointer(mm->ioctx_table, table);
523 kfree_rcu(old, rcu);
524 } else {
525 kfree(table);
526 table = old;
527 }
528 }
529}
530
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531/* ioctx_alloc
532 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
533 */
534static struct kioctx *ioctx_alloc(unsigned nr_events)
535{
Zach Brown41003a72013-05-07 16:18:25 -0700536 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 struct kioctx *ctx;
Al Viroe23754f2012-03-06 14:33:22 -0500538 int err = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000540 /*
541 * We keep track of the number of available ringbuffer slots, to prevent
542 * overflow (reqs_available), and we also use percpu counters for this.
543 *
544 * So since up to half the slots might be on other cpu's percpu counters
545 * and unavailable, double nr_events so userspace sees what they
546 * expected: additionally, we move req_batch slots to/from percpu
547 * counters at a time, so make sure that isn't 0:
548 */
549 nr_events = max(nr_events, num_possible_cpus() * 4);
550 nr_events *= 2;
551
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 /* Prevent overflows */
553 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
554 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
555 pr_debug("ENOMEM: nr_events too high\n");
556 return ERR_PTR(-EINVAL);
557 }
558
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400559 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560 return ERR_PTR(-EAGAIN);
561
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800562 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 if (!ctx)
564 return ERR_PTR(-ENOMEM);
565
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 ctx->max_reqs = nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567
Kent Overstreet723be6e2013-05-28 15:14:48 -0700568 if (percpu_ref_init(&ctx->users, free_ioctx_ref))
569 goto out_freectx;
570
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 spin_lock_init(&ctx->ctx_lock);
Kent Overstreet0460fef2013-05-07 16:18:49 -0700572 spin_lock_init(&ctx->completion_lock);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700573 mutex_init(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 init_waitqueue_head(&ctx->wait);
575
576 INIT_LIST_HEAD(&ctx->active_reqs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000578 ctx->cpu = alloc_percpu(struct kioctx_cpu);
579 if (!ctx->cpu)
Kent Overstreet723be6e2013-05-28 15:14:48 -0700580 goto out_freeref;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000582 if (aio_setup_ring(ctx) < 0)
583 goto out_freepcpu;
584
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000585 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000586 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
Benjamin LaHaise6878ea72013-07-31 10:34:18 -0400587 if (ctx->req_batch < 1)
588 ctx->req_batch = 1;
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000589
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 /* limit the number of system wide aios */
Al Viro9fa1cb32012-03-10 23:14:05 -0500591 spin_lock(&aio_nr_lock);
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400592 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
Al Viro2dd542b2012-03-10 23:10:35 -0500593 aio_nr + nr_events < aio_nr) {
Al Viro9fa1cb32012-03-10 23:14:05 -0500594 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595 goto out_cleanup;
Al Viro2dd542b2012-03-10 23:10:35 -0500596 }
597 aio_nr += ctx->max_reqs;
Al Viro9fa1cb32012-03-10 23:14:05 -0500598 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599
Kent Overstreet723be6e2013-05-28 15:14:48 -0700600 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
601
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400602 err = ioctx_add_table(ctx, mm);
603 if (err)
604 goto out_cleanup_put;
605
Kent Overstreetcaf41672013-05-07 16:18:35 -0700606 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700607 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608 return ctx;
609
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400610out_cleanup_put:
611 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612out_cleanup:
Al Viroe23754f2012-03-06 14:33:22 -0500613 err = -EAGAIN;
614 aio_free_ring(ctx);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000615out_freepcpu:
616 free_percpu(ctx->cpu);
Kent Overstreet723be6e2013-05-28 15:14:48 -0700617out_freeref:
618 free_percpu(ctx->users.pcpu_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619out_freectx:
Gu Zheng36bc08c2013-07-16 17:56:16 +0800620 if (ctx->aio_ring_file)
621 fput(ctx->aio_ring_file);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreetcaf41672013-05-07 16:18:35 -0700623 pr_debug("error allocating ioctx %d\n", err);
Al Viroe23754f2012-03-06 14:33:22 -0500624 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625}
626
Kent Overstreet36f55882013-05-07 16:18:41 -0700627/* kill_ioctx
628 * Cancels all outstanding aio requests on an aio context. Used
629 * when the processes owning a context have all exited to encourage
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 * the rapid destruction of the kioctx.
631 */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400632static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633{
Kent Overstreet36f55882013-05-07 16:18:41 -0700634 if (!atomic_xchg(&ctx->dead, 1)) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400635 struct kioctx_table *table;
636
637 spin_lock(&mm->ioctx_lock);
Artem Savkovd9b2c872013-09-09 12:29:35 -0400638 rcu_read_lock();
Benjamin LaHaise77d30b12013-08-30 11:12:50 -0400639 table = rcu_dereference(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400640
641 WARN_ON(ctx != table->table[ctx->id]);
642 table->table[ctx->id] = NULL;
Artem Savkovd9b2c872013-09-09 12:29:35 -0400643 rcu_read_unlock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400644 spin_unlock(&mm->ioctx_lock);
645
Kent Overstreet723be6e2013-05-28 15:14:48 -0700646 /* percpu_ref_kill() will do the necessary call_rcu() */
647 wake_up_all(&ctx->wait);
Al Viro06af1212012-03-20 16:26:24 -0400648
Kent Overstreet36f55882013-05-07 16:18:41 -0700649 /*
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700650 * It'd be more correct to do this in free_ioctx(), after all
651 * the outstanding kiocbs have finished - but by then io_destroy
652 * has already returned, so io_setup() could potentially return
653 * -EAGAIN with no ioctxs actually in use (as far as userspace
654 * could tell).
Kent Overstreet36f55882013-05-07 16:18:41 -0700655 */
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700656 spin_lock(&aio_nr_lock);
657 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
658 aio_nr -= ctx->max_reqs;
659 spin_unlock(&aio_nr_lock);
660
661 if (ctx->mmap_size)
662 vm_munmap(ctx->mmap_base, ctx->mmap_size);
663
Kent Overstreet723be6e2013-05-28 15:14:48 -0700664 percpu_ref_kill(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666}
667
668/* wait_on_sync_kiocb:
669 * Waits on the given sync kiocb to complete.
670 */
Kent Overstreet57282d82013-05-13 13:42:52 -0700671ssize_t wait_on_sync_kiocb(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672{
Kent Overstreet57282d82013-05-13 13:42:52 -0700673 while (!req->ki_ctx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 set_current_state(TASK_UNINTERRUPTIBLE);
Kent Overstreet57282d82013-05-13 13:42:52 -0700675 if (req->ki_ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 break;
Jeff Moyer41d10da2007-10-16 23:27:20 -0700677 io_schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 }
679 __set_current_state(TASK_RUNNING);
Kent Overstreet57282d82013-05-13 13:42:52 -0700680 return req->ki_user_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700682EXPORT_SYMBOL(wait_on_sync_kiocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
Kent Overstreet36f55882013-05-07 16:18:41 -0700684/*
685 * exit_aio: called when the last user of mm goes away. At this point, there is
686 * no way for any new requests to be submited or any of the io_* syscalls to be
687 * called on the context.
688 *
689 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
690 * them.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800692void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400694 struct kioctx_table *table;
Jens Axboeabf137d2008-12-09 08:11:22 +0100695 struct kioctx *ctx;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400696 unsigned i = 0;
Jens Axboeabf137d2008-12-09 08:11:22 +0100697
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400698 while (1) {
699 rcu_read_lock();
700 table = rcu_dereference(mm->ioctx_table);
701
702 do {
703 if (!table || i >= table->nr) {
704 rcu_read_unlock();
705 rcu_assign_pointer(mm->ioctx_table, NULL);
706 if (table)
707 kfree(table);
708 return;
709 }
710
711 ctx = table->table[i++];
712 } while (!ctx);
713
714 rcu_read_unlock();
715
Al Viro936af152012-04-20 21:49:41 -0400716 /*
717 * We don't need to bother with munmap() here -
718 * exit_mmap(mm) is coming and it'll unmap everything.
719 * Since aio_free_ring() uses non-zero ->mmap_size
720 * as indicator that it needs to unmap the area,
721 * just set it to 0; aio_free_ring() is the only
722 * place that uses ->mmap_size, so it's safe.
Al Viro936af152012-04-20 21:49:41 -0400723 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700724 ctx->mmap_size = 0;
Kent Overstreet36f55882013-05-07 16:18:41 -0700725
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400726 kill_ioctx(mm, ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 }
728}
729
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000730static void put_reqs_available(struct kioctx *ctx, unsigned nr)
731{
732 struct kioctx_cpu *kcpu;
733
734 preempt_disable();
735 kcpu = this_cpu_ptr(ctx->cpu);
736
737 kcpu->reqs_available += nr;
738 while (kcpu->reqs_available >= ctx->req_batch * 2) {
739 kcpu->reqs_available -= ctx->req_batch;
740 atomic_add(ctx->req_batch, &ctx->reqs_available);
741 }
742
743 preempt_enable();
744}
745
746static bool get_reqs_available(struct kioctx *ctx)
747{
748 struct kioctx_cpu *kcpu;
749 bool ret = false;
750
751 preempt_disable();
752 kcpu = this_cpu_ptr(ctx->cpu);
753
754 if (!kcpu->reqs_available) {
755 int old, avail = atomic_read(&ctx->reqs_available);
756
757 do {
758 if (avail < ctx->req_batch)
759 goto out;
760
761 old = avail;
762 avail = atomic_cmpxchg(&ctx->reqs_available,
763 avail, avail - ctx->req_batch);
764 } while (avail != old);
765
766 kcpu->reqs_available += ctx->req_batch;
767 }
768
769 ret = true;
770 kcpu->reqs_available--;
771out:
772 preempt_enable();
773 return ret;
774}
775
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776/* aio_get_req
Kent Overstreet57282d82013-05-13 13:42:52 -0700777 * Allocate a slot for an aio request.
778 * Returns NULL if no requests are free.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 */
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700780static inline struct kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781{
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700782 struct kiocb *req;
783
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000784 if (!get_reqs_available(ctx))
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700785 return NULL;
786
Kent Overstreet0460fef2013-05-07 16:18:49 -0700787 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 if (unlikely(!req))
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700789 goto out_put;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 req->ki_ctx = ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 return req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700793out_put:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000794 put_reqs_available(ctx, 1);
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700795 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796}
797
Kent Overstreet11599eb2013-05-07 16:18:39 -0700798static void kiocb_free(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799{
Kent Overstreet1d98ebf2013-05-07 16:18:37 -0700800 if (req->ki_filp)
801 fput(req->ki_filp);
Davide Libenzi13389012009-06-30 11:41:11 -0700802 if (req->ki_eventfd != NULL)
803 eventfd_ctx_put(req->ki_eventfd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 kmem_cache_free(kiocb_cachep, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805}
806
Adrian Bunkd5470b52008-04-29 00:58:57 -0700807static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400809 struct aio_ring __user *ring = (void __user *)ctx_id;
Jens Axboeabf137d2008-12-09 08:11:22 +0100810 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -0700811 struct kioctx *ctx, *ret = NULL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400812 struct kioctx_table *table;
813 unsigned id;
814
815 if (get_user(id, &ring->id))
816 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817
Jens Axboeabf137d2008-12-09 08:11:22 +0100818 rcu_read_lock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400819 table = rcu_dereference(mm->ioctx_table);
Jens Axboeabf137d2008-12-09 08:11:22 +0100820
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400821 if (!table || id >= table->nr)
822 goto out;
823
824 ctx = table->table[id];
Benjamin LaHaisef30d7042013-08-07 18:23:48 -0400825 if (ctx && ctx->user_id == ctx_id) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400826 percpu_ref_get(&ctx->users);
827 ret = ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +0100828 }
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400829out:
Jens Axboeabf137d2008-12-09 08:11:22 +0100830 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -0700831 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832}
833
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834/* aio_complete
835 * Called when the io request on the given iocb is complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836 */
Kent Overstreet2d684492013-05-07 16:18:29 -0700837void aio_complete(struct kiocb *iocb, long res, long res2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838{
839 struct kioctx *ctx = iocb->ki_ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840 struct aio_ring *ring;
Kent Overstreet21b40202013-05-07 16:18:47 -0700841 struct io_event *ev_page, *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842 unsigned long flags;
Kent Overstreet21b40202013-05-07 16:18:47 -0700843 unsigned tail, pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844
Zach Brown20dcae32005-11-13 16:07:33 -0800845 /*
846 * Special case handling for sync iocbs:
847 * - events go directly into the iocb for fast handling
848 * - the sync task with the iocb in its stack holds the single iocb
849 * ref, no other paths have a way to get another ref
850 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 */
852 if (is_sync_kiocb(iocb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 iocb->ki_user_data = res;
Kent Overstreet57282d82013-05-13 13:42:52 -0700854 smp_wmb();
855 iocb->ki_ctx = ERR_PTR(-EXDEV);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 wake_up_process(iocb->ki_obj.tsk);
Kent Overstreet2d684492013-05-07 16:18:29 -0700857 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858 }
859
Kent Overstreet36f55882013-05-07 16:18:41 -0700860 /*
Kent Overstreet36f55882013-05-07 16:18:41 -0700861 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000862 * need to issue a wakeup after incrementing reqs_available.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 */
Kent Overstreet36f55882013-05-07 16:18:41 -0700864 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865
Kent Overstreet0460fef2013-05-07 16:18:49 -0700866 if (iocb->ki_list.next) {
867 unsigned long flags;
868
869 spin_lock_irqsave(&ctx->ctx_lock, flags);
870 list_del(&iocb->ki_list);
871 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
872 }
Kent Overstreet11599eb2013-05-07 16:18:39 -0700873
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 /*
Kent Overstreet0460fef2013-05-07 16:18:49 -0700875 * Add a completion event to the ring buffer. Must be done holding
Tang Chen4b30f072013-07-03 15:09:16 -0700876 * ctx->completion_lock to prevent other code from messing with the tail
Kent Overstreet0460fef2013-05-07 16:18:49 -0700877 * pointer since we might be called from irq context.
878 */
879 spin_lock_irqsave(&ctx->completion_lock, flags);
880
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700881 tail = ctx->tail;
Kent Overstreet21b40202013-05-07 16:18:47 -0700882 pos = tail + AIO_EVENTS_OFFSET;
883
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700884 if (++tail >= ctx->nr_events)
Ken Chen4bf69b22005-05-01 08:59:15 -0700885 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700887 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -0700888 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
889
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
891 event->data = iocb->ki_user_data;
892 event->res = res;
893 event->res2 = res2;
894
Kent Overstreet21b40202013-05-07 16:18:47 -0700895 kunmap_atomic(ev_page);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700896 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -0700897
898 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
Kent Overstreetcaf41672013-05-07 16:18:35 -0700899 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
900 res, res2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901
902 /* after flagging the request as done, we
903 * must never even look at it again
904 */
905 smp_wmb(); /* make event visible before updating tail */
906
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700907 ctx->tail = tail;
Kent Overstreet21b40202013-05-07 16:18:47 -0700908
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700909 ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 ring->tail = tail;
Cong Wange8e3c3d2011-11-25 23:14:27 +0800911 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700912 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913
Kent Overstreet0460fef2013-05-07 16:18:49 -0700914 spin_unlock_irqrestore(&ctx->completion_lock, flags);
915
Kent Overstreet21b40202013-05-07 16:18:47 -0700916 pr_debug("added to ring %p at [%u]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -0700917
918 /*
919 * Check if the user asked us to deliver the result through an
920 * eventfd. The eventfd_signal() function is safe to be called
921 * from IRQ context.
922 */
Davide Libenzi87c3a862009-03-18 17:04:19 -0700923 if (iocb->ki_eventfd != NULL)
Davide Libenzi8d1c98b2008-04-10 21:29:19 -0700924 eventfd_signal(iocb->ki_eventfd, 1);
925
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 /* everything turned out well, dispose of the aiocb. */
Kent Overstreet57282d82013-05-13 13:42:52 -0700927 kiocb_free(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928
Quentin Barnes6cb2a212008-03-19 17:00:39 -0700929 /*
930 * We have to order our ring_info tail store above and test
931 * of the wait list below outside the wait lock. This is
932 * like in wake_up_bit() where clearing a bit has to be
933 * ordered with the unlocked test.
934 */
935 smp_mb();
936
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 if (waitqueue_active(&ctx->wait))
938 wake_up(&ctx->wait);
939
Kent Overstreet36f55882013-05-07 16:18:41 -0700940 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700942EXPORT_SYMBOL(aio_complete);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943
Kent Overstreeta31ad382013-05-07 16:18:45 -0700944/* aio_read_events
945 * Pull an event off of the ioctx's event ring. Returns the number of
946 * events fetched
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 */
Kent Overstreeta31ad382013-05-07 16:18:45 -0700948static long aio_read_events_ring(struct kioctx *ctx,
949 struct io_event __user *event, long nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 struct aio_ring *ring;
Kent Overstreet5ffac122013-05-09 15:36:07 -0700952 unsigned head, tail, pos;
Kent Overstreeta31ad382013-05-07 16:18:45 -0700953 long ret = 0;
954 int copy_ret;
955
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700956 mutex_lock(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700958 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -0700959 head = ring->head;
Kent Overstreet5ffac122013-05-09 15:36:07 -0700960 tail = ring->tail;
Kent Overstreeta31ad382013-05-07 16:18:45 -0700961 kunmap_atomic(ring);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962
Kent Overstreet5ffac122013-05-09 15:36:07 -0700963 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
Kent Overstreeta31ad382013-05-07 16:18:45 -0700964
Kent Overstreet5ffac122013-05-09 15:36:07 -0700965 if (head == tail)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 goto out;
967
Kent Overstreeta31ad382013-05-07 16:18:45 -0700968 while (ret < nr) {
969 long avail;
970 struct io_event *ev;
971 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972
Kent Overstreet5ffac122013-05-09 15:36:07 -0700973 avail = (head <= tail ? tail : ctx->nr_events) - head;
974 if (head == tail)
Kent Overstreeta31ad382013-05-07 16:18:45 -0700975 break;
976
977 avail = min(avail, nr - ret);
978 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
979 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
980
981 pos = head + AIO_EVENTS_OFFSET;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700982 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
Kent Overstreeta31ad382013-05-07 16:18:45 -0700983 pos %= AIO_EVENTS_PER_PAGE;
984
985 ev = kmap(page);
986 copy_ret = copy_to_user(event + ret, ev + pos,
987 sizeof(*ev) * avail);
988 kunmap(page);
989
990 if (unlikely(copy_ret)) {
991 ret = -EFAULT;
992 goto out;
993 }
994
995 ret += avail;
996 head += avail;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700997 head %= ctx->nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001000 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001001 ring->head = head;
Zhao Hongjiang91d80a82013-04-26 11:03:53 +08001002 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001003 flush_dcache_page(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001004
Kent Overstreet5ffac122013-05-09 15:36:07 -07001005 pr_debug("%li h%u t%u\n", ret, head, tail);
Kent Overstreet3e845ce2013-05-07 16:18:51 -07001006
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001007 put_reqs_available(ctx, ret);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001008out:
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001009 mutex_unlock(&ctx->ring_lock);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001010
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011 return ret;
1012}
1013
Kent Overstreeta31ad382013-05-07 16:18:45 -07001014static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1015 struct io_event __user *event, long *i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001017 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018
Kent Overstreeta31ad382013-05-07 16:18:45 -07001019 if (ret > 0)
1020 *i += ret;
1021
1022 if (unlikely(atomic_read(&ctx->dead)))
1023 ret = -EINVAL;
1024
1025 if (!*i)
1026 *i = ret;
1027
1028 return ret < 0 || *i >= min_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029}
1030
Kent Overstreeta31ad382013-05-07 16:18:45 -07001031static long read_events(struct kioctx *ctx, long min_nr, long nr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032 struct io_event __user *event,
1033 struct timespec __user *timeout)
1034{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001035 ktime_t until = { .tv64 = KTIME_MAX };
1036 long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 if (timeout) {
1039 struct timespec ts;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001040
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
Kent Overstreeta31ad382013-05-07 16:18:45 -07001042 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043
Kent Overstreeta31ad382013-05-07 16:18:45 -07001044 until = timespec_to_ktime(ts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 }
1046
Kent Overstreeta31ad382013-05-07 16:18:45 -07001047 /*
1048 * Note that aio_read_events() is being called as the conditional - i.e.
1049 * we're calling it after prepare_to_wait() has set task state to
1050 * TASK_INTERRUPTIBLE.
1051 *
1052 * But aio_read_events() can block, and if it blocks it's going to flip
1053 * the task state back to TASK_RUNNING.
1054 *
1055 * This should be ok, provided it doesn't flip the state back to
1056 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1057 * will only happen if the mutex_lock() call blocks, and we then find
1058 * the ringbuffer empty. So in practice we should be ok, but it's
1059 * something to be aware of when touching this code.
1060 */
1061 wait_event_interruptible_hrtimeout(ctx->wait,
1062 aio_read_events(ctx, min_nr, nr, event, &ret), until);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063
Kent Overstreeta31ad382013-05-07 16:18:45 -07001064 if (!ret && signal_pending(current))
1065 ret = -EINTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066
Kent Overstreeta31ad382013-05-07 16:18:45 -07001067 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068}
1069
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070/* sys_io_setup:
1071 * Create an aio_context capable of receiving at least nr_events.
1072 * ctxp must not point to an aio_context that already exists, and
1073 * must be initialized to 0 prior to the call. On successful
1074 * creation of the aio_context, *ctxp is filled in with the resulting
1075 * handle. May fail with -EINVAL if *ctxp is not initialized,
1076 * if the specified nr_events exceeds internal limits. May fail
1077 * with -EAGAIN if the specified nr_events exceeds the user's limit
1078 * of available events. May fail with -ENOMEM if insufficient kernel
1079 * resources are available. May fail with -EFAULT if an invalid
1080 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1081 * implemented.
1082 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001083SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084{
1085 struct kioctx *ioctx = NULL;
1086 unsigned long ctx;
1087 long ret;
1088
1089 ret = get_user(ctx, ctxp);
1090 if (unlikely(ret))
1091 goto out;
1092
1093 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001094 if (unlikely(ctx || nr_events == 0)) {
1095 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1096 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 goto out;
1098 }
1099
1100 ioctx = ioctx_alloc(nr_events);
1101 ret = PTR_ERR(ioctx);
1102 if (!IS_ERR(ioctx)) {
1103 ret = put_user(ioctx->user_id, ctxp);
Al Viroa2e18592012-03-20 16:27:57 -04001104 if (ret)
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001105 kill_ioctx(current->mm, ioctx);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001106 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 }
1108
1109out:
1110 return ret;
1111}
1112
1113/* sys_io_destroy:
1114 * Destroy the aio_context specified. May cancel any outstanding
1115 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001116 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 * is invalid.
1118 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001119SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120{
1121 struct kioctx *ioctx = lookup_ioctx(ctx);
1122 if (likely(NULL != ioctx)) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001123 kill_ioctx(current->mm, ioctx);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001124 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 return 0;
1126 }
1127 pr_debug("EINVAL: io_destroy: invalid context id\n");
1128 return -EINVAL;
1129}
1130
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001131typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1132 unsigned long, loff_t);
1133
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001134static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1135 int rw, char __user *buf,
1136 unsigned long *nr_segs,
1137 struct iovec **iovec,
1138 bool compat)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001139{
1140 ssize_t ret;
1141
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001142 *nr_segs = kiocb->ki_nbytes;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001143
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001144#ifdef CONFIG_COMPAT
1145 if (compat)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001146 ret = compat_rw_copy_check_uvector(rw,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001147 (struct compat_iovec __user *)buf,
1148 *nr_segs, 1, *iovec, iovec);
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001149 else
1150#endif
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001151 ret = rw_copy_check_uvector(rw,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001152 (struct iovec __user *)buf,
1153 *nr_segs, 1, *iovec, iovec);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001154 if (ret < 0)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001155 return ret;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001156
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001157 /* ki_nbytes now reflect bytes instead of segs */
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001158 kiocb->ki_nbytes = ret;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001159 return 0;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001160}
1161
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001162static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1163 int rw, char __user *buf,
1164 unsigned long *nr_segs,
1165 struct iovec *iovec)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001166{
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001167 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001168 return -EFAULT;
Linus Torvaldsa70b52e2012-05-21 16:06:20 -07001169
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001170 iovec->iov_base = buf;
1171 iovec->iov_len = kiocb->ki_nbytes;
1172 *nr_segs = 1;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001173 return 0;
1174}
1175
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176/*
1177 * aio_setup_iocb:
1178 * Performs the initial checks and aio retry method
1179 * setup for the kiocb at the time of io submission.
1180 */
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001181static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1182 char __user *buf, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183{
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001184 struct file *file = req->ki_filp;
1185 ssize_t ret;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001186 unsigned long nr_segs;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001187 int rw;
1188 fmode_t mode;
1189 aio_rw_op *rw_op;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001190 struct iovec inline_vec, *iovec = &inline_vec;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001192 switch (opcode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 case IOCB_CMD_PREAD:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001194 case IOCB_CMD_PREADV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001195 mode = FMODE_READ;
1196 rw = READ;
1197 rw_op = file->f_op->aio_read;
1198 goto rw_common;
1199
1200 case IOCB_CMD_PWRITE:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001201 case IOCB_CMD_PWRITEV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001202 mode = FMODE_WRITE;
1203 rw = WRITE;
1204 rw_op = file->f_op->aio_write;
1205 goto rw_common;
1206rw_common:
1207 if (unlikely(!(file->f_mode & mode)))
1208 return -EBADF;
1209
1210 if (!rw_op)
1211 return -EINVAL;
1212
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001213 ret = (opcode == IOCB_CMD_PREADV ||
1214 opcode == IOCB_CMD_PWRITEV)
1215 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1216 &iovec, compat)
1217 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1218 iovec);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001219 if (ret)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001220 return ret;
1221
1222 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001223 if (ret < 0) {
1224 if (iovec != &inline_vec)
1225 kfree(iovec);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001226 return ret;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001227 }
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001228
1229 req->ki_nbytes = ret;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001230
Kent Overstreet73a70752013-05-09 15:03:42 -07001231 /* XXX: move/kill - rw_verify_area()? */
1232 /* This matches the pread()/pwrite() logic */
1233 if (req->ki_pos < 0) {
1234 ret = -EINVAL;
1235 break;
1236 }
1237
1238 if (rw == WRITE)
1239 file_start_write(file);
1240
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001241 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
Kent Overstreet73a70752013-05-09 15:03:42 -07001242
1243 if (rw == WRITE)
1244 file_end_write(file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001246
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247 case IOCB_CMD_FDSYNC:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001248 if (!file->f_op->aio_fsync)
1249 return -EINVAL;
1250
1251 ret = file->f_op->aio_fsync(req, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001253
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 case IOCB_CMD_FSYNC:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001255 if (!file->f_op->aio_fsync)
1256 return -EINVAL;
1257
1258 ret = file->f_op->aio_fsync(req, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001260
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 default:
Kent Overstreetcaf41672013-05-07 16:18:35 -07001262 pr_debug("EINVAL: no operation provided\n");
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001263 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 }
1265
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001266 if (iovec != &inline_vec)
1267 kfree(iovec);
1268
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001269 if (ret != -EIOCBQUEUED) {
1270 /*
1271 * There's no easy way to restart the syscall since other AIO's
1272 * may be already running. Just fail this IO with EINTR.
1273 */
1274 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1275 ret == -ERESTARTNOHAND ||
1276 ret == -ERESTART_RESTARTBLOCK))
1277 ret = -EINTR;
1278 aio_complete(req, ret, 0);
1279 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280
1281 return 0;
1282}
1283
Adrian Bunkd5470b52008-04-29 00:58:57 -07001284static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001285 struct iocb *iocb, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286{
1287 struct kiocb *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 ssize_t ret;
1289
1290 /* enforce forwards compatibility on users */
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001291 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001292 pr_debug("EINVAL: reserve field set\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 return -EINVAL;
1294 }
1295
1296 /* prevent overflows */
1297 if (unlikely(
1298 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1299 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1300 ((ssize_t)iocb->aio_nbytes < 0)
1301 )) {
1302 pr_debug("EINVAL: io_submit: overflow check\n");
1303 return -EINVAL;
1304 }
1305
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001306 req = aio_get_req(ctx);
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001307 if (unlikely(!req))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 return -EAGAIN;
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001309
1310 req->ki_filp = fget(iocb->aio_fildes);
1311 if (unlikely(!req->ki_filp)) {
1312 ret = -EBADF;
1313 goto out_put_req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 }
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001315
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001316 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1317 /*
1318 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1319 * instance of the file* now. The file descriptor must be
1320 * an eventfd() fd, and will be signaled for each completed
1321 * event using the eventfd_signal() function.
1322 */
Davide Libenzi13389012009-06-30 11:41:11 -07001323 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
Hirofumi Nakagawa801678c2008-04-29 01:03:09 -07001324 if (IS_ERR(req->ki_eventfd)) {
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001325 ret = PTR_ERR(req->ki_eventfd);
Davide Libenzi87c3a862009-03-18 17:04:19 -07001326 req->ki_eventfd = NULL;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001327 goto out_put_req;
1328 }
1329 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330
Kent Overstreet8a660892013-05-07 16:19:10 -07001331 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 if (unlikely(ret)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001333 pr_debug("EFAULT: aio_key\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 goto out_put_req;
1335 }
1336
1337 req->ki_obj.user = user_iocb;
1338 req->ki_user_data = iocb->aio_data;
1339 req->ki_pos = iocb->aio_offset;
Kent Overstreet73a70752013-05-09 15:03:42 -07001340 req->ki_nbytes = iocb->aio_nbytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001342 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1343 (char __user *)(unsigned long)iocb->aio_buf,
1344 compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 if (ret)
1346 goto out_put_req;
1347
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349out_put_req:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001350 put_reqs_available(ctx, 1);
Kent Overstreet57282d82013-05-13 13:42:52 -07001351 kiocb_free(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 return ret;
1353}
1354
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001355long do_io_submit(aio_context_t ctx_id, long nr,
1356 struct iocb __user *__user *iocbpp, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357{
1358 struct kioctx *ctx;
1359 long ret = 0;
Jeff Moyer080d6762011-11-02 13:40:10 -07001360 int i = 0;
Shaohua Li9f5b9422010-07-01 07:55:01 +02001361 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362
1363 if (unlikely(nr < 0))
1364 return -EINVAL;
1365
Jeff Moyer75e1c702010-09-10 14:16:00 -07001366 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1367 nr = LONG_MAX/sizeof(*iocbpp);
1368
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1370 return -EFAULT;
1371
1372 ctx = lookup_ioctx(ctx_id);
1373 if (unlikely(!ctx)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001374 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 return -EINVAL;
1376 }
1377
Shaohua Li9f5b9422010-07-01 07:55:01 +02001378 blk_start_plug(&plug);
1379
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 /*
1381 * AKPM: should this return a partial result if some of the IOs were
1382 * successfully submitted?
1383 */
1384 for (i=0; i<nr; i++) {
1385 struct iocb __user *user_iocb;
1386 struct iocb tmp;
1387
1388 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1389 ret = -EFAULT;
1390 break;
1391 }
1392
1393 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1394 ret = -EFAULT;
1395 break;
1396 }
1397
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001398 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 if (ret)
1400 break;
1401 }
Shaohua Li9f5b9422010-07-01 07:55:01 +02001402 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403
Kent Overstreet723be6e2013-05-28 15:14:48 -07001404 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405 return i ? i : ret;
1406}
1407
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001408/* sys_io_submit:
1409 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1410 * the number of iocbs queued. May return -EINVAL if the aio_context
1411 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1412 * *iocbpp[0] is not properly initialized, if the operation specified
1413 * is invalid for the file descriptor in the iocb. May fail with
1414 * -EFAULT if any of the data structures point to invalid data. May
1415 * fail with -EBADF if the file descriptor specified in the first
1416 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1417 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1418 * fail with -ENOSYS if not implemented.
1419 */
1420SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1421 struct iocb __user * __user *, iocbpp)
1422{
1423 return do_io_submit(ctx_id, nr, iocbpp, 0);
1424}
1425
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426/* lookup_kiocb
1427 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001429static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1430 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431{
1432 struct list_head *pos;
Zach Brownd00689a2005-11-13 16:07:34 -08001433
1434 assert_spin_locked(&ctx->ctx_lock);
1435
Kent Overstreet8a660892013-05-07 16:19:10 -07001436 if (key != KIOCB_KEY)
1437 return NULL;
1438
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 /* TODO: use a hash or array, this sucks. */
1440 list_for_each(pos, &ctx->active_reqs) {
1441 struct kiocb *kiocb = list_kiocb(pos);
Kent Overstreet8a660892013-05-07 16:19:10 -07001442 if (kiocb->ki_obj.user == iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443 return kiocb;
1444 }
1445 return NULL;
1446}
1447
1448/* sys_io_cancel:
1449 * Attempts to cancel an iocb previously passed to io_submit. If
1450 * the operation is successfully cancelled, the resulting event is
1451 * copied into the memory pointed to by result without being placed
1452 * into the completion queue and 0 is returned. May fail with
1453 * -EFAULT if any of the data structures pointed to are invalid.
1454 * May fail with -EINVAL if aio_context specified by ctx_id is
1455 * invalid. May fail with -EAGAIN if the iocb specified was not
1456 * cancelled. Will fail with -ENOSYS if not implemented.
1457 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001458SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1459 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 struct kioctx *ctx;
1462 struct kiocb *kiocb;
1463 u32 key;
1464 int ret;
1465
1466 ret = get_user(key, &iocb->aio_key);
1467 if (unlikely(ret))
1468 return -EFAULT;
1469
1470 ctx = lookup_ioctx(ctx_id);
1471 if (unlikely(!ctx))
1472 return -EINVAL;
1473
1474 spin_lock_irq(&ctx->ctx_lock);
Kent Overstreet906b9732013-05-07 16:18:31 -07001475
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476 kiocb = lookup_kiocb(ctx, iocb, key);
Kent Overstreet906b9732013-05-07 16:18:31 -07001477 if (kiocb)
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001478 ret = kiocb_cancel(ctx, kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -07001479 else
1480 ret = -EINVAL;
1481
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482 spin_unlock_irq(&ctx->ctx_lock);
1483
Kent Overstreet906b9732013-05-07 16:18:31 -07001484 if (!ret) {
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001485 /*
1486 * The result argument is no longer used - the io_event is
1487 * always delivered via the ring buffer. -EINPROGRESS indicates
1488 * cancellation is progress:
Kent Overstreet906b9732013-05-07 16:18:31 -07001489 */
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001490 ret = -EINPROGRESS;
Kent Overstreet906b9732013-05-07 16:18:31 -07001491 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492
Kent Overstreet723be6e2013-05-28 15:14:48 -07001493 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494
1495 return ret;
1496}
1497
1498/* io_getevents:
1499 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001500 * the completion queue for the aio_context specified by ctx_id. If
1501 * it succeeds, the number of read events is returned. May fail with
1502 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1503 * out of range, if timeout is out of range. May fail with -EFAULT
1504 * if any of the memory specified is invalid. May return 0 or
1505 * < min_nr if the timeout specified by timeout has elapsed
1506 * before sufficient events are available, where timeout == NULL
1507 * specifies an infinite timeout. Note that the timeout pointed to by
Jeff Moyer69008072013-05-24 15:55:24 -07001508 * timeout is relative. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001510SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1511 long, min_nr,
1512 long, nr,
1513 struct io_event __user *, events,
1514 struct timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515{
1516 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1517 long ret = -EINVAL;
1518
1519 if (likely(ioctx)) {
Namhyung Kim2e410252011-01-12 17:01:08 -08001520 if (likely(min_nr <= nr && min_nr >= 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 ret = read_events(ioctx, min_nr, nr, events, timeout);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001522 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 return ret;
1525}