blob: 14b93159ef83a140483bb31a4a6c70286d209ed8 [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/migrate.h>
40#include <linux/ramfs.h>
Kent Overstreet723be6e2013-05-28 15:14:48 -070041#include <linux/percpu-refcount.h>
Benjamin LaHaise71ad7492013-09-17 10:18:25 -040042#include <linux/mount.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 */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -040055 unsigned head; /* Written to by userland or under ring_lock
56 * mutex by aio_read_events_ring(). */
Kent Overstreet4e179bc2013-05-07 16:18:33 -070057 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
Kent Overstreet4e179bc2013-05-07 16:18:33 -070069
Benjamin LaHaisedb446a02013-07-30 12:54:40 -040070struct kioctx_table {
71 struct rcu_head rcu;
72 unsigned nr;
73 struct kioctx *table[];
74};
75
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100076struct kioctx_cpu {
77 unsigned reqs_available;
78};
79
Kent Overstreet4e179bc2013-05-07 16:18:33 -070080struct kioctx {
Kent Overstreet723be6e2013-05-28 15:14:48 -070081 struct percpu_ref users;
Kent Overstreet36f55882013-05-07 16:18:41 -070082 atomic_t dead;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070083
Kent Overstreete34ecee2013-10-10 19:31:47 -070084 struct percpu_ref reqs;
85
Kent Overstreet4e179bc2013-05-07 16:18:33 -070086 unsigned long user_id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070087
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100088 struct __percpu kioctx_cpu *cpu;
89
90 /*
91 * For percpu reqs_available, number of slots we move to/from global
92 * counter at a time:
93 */
94 unsigned req_batch;
Kent Overstreet3e845ce2013-05-07 16:18:51 -070095 /*
96 * This is what userspace passed to io_setup(), it's not used for
97 * anything but counting against the global max_reqs quota.
98 *
Kent Overstreet58c85dc2013-05-07 16:18:55 -070099 * The real limit is nr_events - 1, which will be larger (see
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700100 * aio_setup_ring())
101 */
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700102 unsigned max_reqs;
103
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700104 /* Size of ringbuffer, in units of struct io_event */
105 unsigned nr_events;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700106
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700107 unsigned long mmap_base;
108 unsigned long mmap_size;
109
110 struct page **ring_pages;
111 long nr_pages;
112
Kent Overstreet723be6e2013-05-28 15:14:48 -0700113 struct work_struct free_work;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700114
Anatol Pomozove02ba722014-04-15 11:31:33 -0700115 /*
116 * signals when all in-flight requests are done
117 */
118 struct completion *requests_done;
119
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700120 struct {
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000121 /*
122 * This counts the number of available slots in the ringbuffer,
123 * so we avoid overflowing it: it's decremented (if positive)
124 * when allocating a kiocb and incremented when the resulting
125 * io_event is pulled off the ringbuffer.
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000126 *
127 * We batch accesses to it with a percpu version.
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000128 */
129 atomic_t reqs_available;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700130 } ____cacheline_aligned_in_smp;
131
132 struct {
133 spinlock_t ctx_lock;
134 struct list_head active_reqs; /* used for cancellation */
135 } ____cacheline_aligned_in_smp;
136
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700137 struct {
138 struct mutex ring_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700139 wait_queue_head_t wait;
140 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700141
142 struct {
143 unsigned tail;
Benjamin LaHaised856f322014-08-24 13:14:05 -0400144 unsigned completed_events;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700145 spinlock_t completion_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700146 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700147
148 struct page *internal_pages[AIO_RING_PAGES];
Gu Zheng36bc08c2013-07-16 17:56:16 +0800149 struct file *aio_ring_file;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400150
151 unsigned id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700152};
153
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -0800155static DEFINE_SPINLOCK(aio_nr_lock);
156unsigned long aio_nr; /* current system wide number of aio requests */
157unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158/*----end sysctl variables---*/
159
Christoph Lametere18b8902006-12-06 20:33:20 -0800160static struct kmem_cache *kiocb_cachep;
161static struct kmem_cache *kioctx_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400163static struct vfsmount *aio_mnt;
164
165static const struct file_operations aio_ring_fops;
166static const struct address_space_operations aio_ctx_aops;
167
Gu Zheng835f2522014-11-06 17:46:21 +0800168/* Backing dev info for aio fs.
169 * -no dirty page accounting or writeback happens
170 */
171static struct backing_dev_info aio_fs_backing_dev_info = {
172 .name = "aiofs",
173 .state = 0,
174 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_MAP_COPY,
175};
176
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400177static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
178{
179 struct qstr this = QSTR_INIT("[aio]", 5);
180 struct file *file;
181 struct path path;
182 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
Dan Carpenter7f626562013-11-13 10:49:40 +0300183 if (IS_ERR(inode))
184 return ERR_CAST(inode);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400185
186 inode->i_mapping->a_ops = &aio_ctx_aops;
187 inode->i_mapping->private_data = ctx;
Gu Zheng835f2522014-11-06 17:46:21 +0800188 inode->i_mapping->backing_dev_info = &aio_fs_backing_dev_info;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400189 inode->i_size = PAGE_SIZE * nr_pages;
190
191 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
192 if (!path.dentry) {
193 iput(inode);
194 return ERR_PTR(-ENOMEM);
195 }
196 path.mnt = mntget(aio_mnt);
197
198 d_instantiate(path.dentry, inode);
199 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
200 if (IS_ERR(file)) {
201 path_put(&path);
202 return file;
203 }
204
205 file->f_flags = O_RDWR;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400206 return file;
207}
208
209static struct dentry *aio_mount(struct file_system_type *fs_type,
210 int flags, const char *dev_name, void *data)
211{
212 static const struct dentry_operations ops = {
213 .d_dname = simple_dname,
214 };
Gu Zheng8dc43792014-07-23 18:03:52 +0800215 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400216}
217
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218/* aio_setup
219 * Creates the slab caches used by the aio routines, panic on
220 * failure as this is done early during the boot sequence.
221 */
222static int __init aio_setup(void)
223{
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400224 static struct file_system_type aio_fs = {
225 .name = "aio",
226 .mount = aio_mount,
227 .kill_sb = kill_anon_super,
228 };
229 aio_mnt = kern_mount(&aio_fs);
230 if (IS_ERR(aio_mnt))
231 panic("Failed to create aio fs mount.");
232
Gu Zheng835f2522014-11-06 17:46:21 +0800233 if (bdi_init(&aio_fs_backing_dev_info))
234 panic("Failed to init aio fs backing dev info.");
235
Christoph Lameter0a31bd52007-05-06 14:49:57 -0700236 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
237 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238
Kent Overstreetcaf41672013-05-07 16:18:35 -0700239 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240
241 return 0;
242}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700243__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400245static void put_aio_ring_file(struct kioctx *ctx)
246{
247 struct file *aio_ring_file = ctx->aio_ring_file;
248 if (aio_ring_file) {
249 truncate_setsize(aio_ring_file->f_inode, 0);
250
251 /* Prevent further access to the kioctx from migratepages */
252 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
253 aio_ring_file->f_inode->i_mapping->private_data = NULL;
254 ctx->aio_ring_file = NULL;
255 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
256
257 fput(aio_ring_file);
258 }
259}
260
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261static void aio_free_ring(struct kioctx *ctx)
262{
Gu Zheng36bc08c2013-07-16 17:56:16 +0800263 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400265 /* Disconnect the kiotx from the ring file. This prevents future
266 * accesses to the kioctx from page migration.
267 */
268 put_aio_ring_file(ctx);
269
Gu Zheng36bc08c2013-07-16 17:56:16 +0800270 for (i = 0; i < ctx->nr_pages; i++) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500271 struct page *page;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800272 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
273 page_count(ctx->ring_pages[i]));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500274 page = ctx->ring_pages[i];
275 if (!page)
276 continue;
277 ctx->ring_pages[i] = NULL;
278 put_page(page);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800279 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280
Sasha Levinddb8c452013-11-19 17:33:03 -0500281 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700282 kfree(ctx->ring_pages);
Sasha Levinddb8c452013-11-19 17:33:03 -0500283 ctx->ring_pages = NULL;
284 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285}
286
Gu Zheng36bc08c2013-07-16 17:56:16 +0800287static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
288{
289 vma->vm_ops = &generic_file_vm_ops;
290 return 0;
291}
292
293static const struct file_operations aio_ring_fops = {
294 .mmap = aio_ring_mmap,
295};
296
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400297#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800298static int aio_migratepage(struct address_space *mapping, struct page *new,
299 struct page *old, enum migrate_mode mode)
300{
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400301 struct kioctx *ctx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800302 unsigned long flags;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400303 pgoff_t idx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800304 int rc;
305
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500306 rc = 0;
307
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400308 /* mapping->private_lock here protects against the kioctx teardown. */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500309 spin_lock(&mapping->private_lock);
310 ctx = mapping->private_data;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400311 if (!ctx) {
312 rc = -EINVAL;
313 goto out;
314 }
315
316 /* The ring_lock mutex. The prevents aio_read_events() from writing
317 * to the ring's head, and prevents page migration from mucking in
318 * a partially initialized kiotx.
319 */
320 if (!mutex_trylock(&ctx->ring_lock)) {
321 rc = -EAGAIN;
322 goto out;
323 }
324
325 idx = old->index;
326 if (idx < (pgoff_t)ctx->nr_pages) {
327 /* Make sure the old page hasn't already been changed */
328 if (ctx->ring_pages[idx] != old)
329 rc = -EAGAIN;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500330 } else
331 rc = -EINVAL;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500332
333 if (rc != 0)
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400334 goto out_unlock;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500335
Gu Zheng36bc08c2013-07-16 17:56:16 +0800336 /* Writeback must be complete */
337 BUG_ON(PageWriteback(old));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500338 get_page(new);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800339
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500340 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800341 if (rc != MIGRATEPAGE_SUCCESS) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500342 put_page(new);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400343 goto out_unlock;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800344 }
345
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400346 /* Take completion_lock to prevent other writes to the ring buffer
347 * while the old page is copied to the new. This prevents new
348 * events from being lost.
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400349 */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400350 spin_lock_irqsave(&ctx->completion_lock, flags);
351 migrate_page_copy(new, old);
352 BUG_ON(ctx->ring_pages[idx] != old);
353 ctx->ring_pages[idx] = new;
354 spin_unlock_irqrestore(&ctx->completion_lock, flags);
355
356 /* The old page is no longer accessible. */
357 put_page(old);
358
359out_unlock:
360 mutex_unlock(&ctx->ring_lock);
361out:
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400362 spin_unlock(&mapping->private_lock);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800363 return rc;
364}
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400365#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800366
367static const struct address_space_operations aio_ctx_aops = {
Gu Zheng835f2522014-11-06 17:46:21 +0800368 .set_page_dirty = __set_page_dirty_no_writeback,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400369#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800370 .migratepage = aio_migratepage,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400371#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800372};
373
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374static int aio_setup_ring(struct kioctx *ctx)
375{
376 struct aio_ring *ring;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 unsigned nr_events = ctx->max_reqs;
Zach Brown41003a72013-05-07 16:18:25 -0700378 struct mm_struct *mm = current->mm;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900379 unsigned long size, unused;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 int nr_pages;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800381 int i;
382 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383
384 /* Compensate for the ring buffer's head/tail overlap entry */
385 nr_events += 2; /* 1 is required, 2 for good luck */
386
387 size = sizeof(struct aio_ring);
388 size += sizeof(struct io_event) * nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
Gu Zheng36bc08c2013-07-16 17:56:16 +0800390 nr_pages = PFN_UP(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 if (nr_pages < 0)
392 return -EINVAL;
393
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400394 file = aio_private_file(ctx, nr_pages);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800395 if (IS_ERR(file)) {
396 ctx->aio_ring_file = NULL;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400397 return -ENOMEM;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800398 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399
Gu Zheng36bc08c2013-07-16 17:56:16 +0800400 ctx->aio_ring_file = file;
401 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
402 / sizeof(struct io_event);
403
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700404 ctx->ring_pages = ctx->internal_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 if (nr_pages > AIO_RING_PAGES) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700406 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
407 GFP_KERNEL);
Gu Zhengd1b94322013-12-04 18:19:06 +0800408 if (!ctx->ring_pages) {
409 put_aio_ring_file(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 return -ENOMEM;
Gu Zhengd1b94322013-12-04 18:19:06 +0800411 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 }
413
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900414 for (i = 0; i < nr_pages; i++) {
415 struct page *page;
416 page = find_or_create_page(file->f_inode->i_mapping,
417 i, GFP_HIGHUSER | __GFP_ZERO);
418 if (!page)
419 break;
420 pr_debug("pid(%d) page[%d]->count=%d\n",
421 current->pid, i, page_count(page));
422 SetPageUptodate(page);
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900423 unlock_page(page);
424
425 ctx->ring_pages[i] = page;
426 }
427 ctx->nr_pages = i;
428
429 if (unlikely(i != nr_pages)) {
430 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400431 return -ENOMEM;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900432 }
433
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700434 ctx->mmap_size = nr_pages * PAGE_SIZE;
435 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800436
Zach Brown41003a72013-05-07 16:18:25 -0700437 down_write(&mm->mmap_sem);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800438 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
439 PROT_READ | PROT_WRITE,
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900440 MAP_SHARED, 0, &unused);
441 up_write(&mm->mmap_sem);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700442 if (IS_ERR((void *)ctx->mmap_base)) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700443 ctx->mmap_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400445 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 }
447
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700448 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400449
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700450 ctx->user_id = ctx->mmap_base;
451 ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700453 ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 ring->nr = nr_events; /* user copy */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400455 ring->id = ~0U;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 ring->head = ring->tail = 0;
457 ring->magic = AIO_RING_MAGIC;
458 ring->compat_features = AIO_RING_COMPAT_FEATURES;
459 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
460 ring->header_length = sizeof(struct aio_ring);
Cong Wange8e3c3d2011-11-25 23:14:27 +0800461 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700462 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463
464 return 0;
465}
466
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
468#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
469#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
470
Kent Overstreet0460fef2013-05-07 16:18:49 -0700471void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
472{
473 struct kioctx *ctx = req->ki_ctx;
474 unsigned long flags;
475
476 spin_lock_irqsave(&ctx->ctx_lock, flags);
477
478 if (!req->ki_list.next)
479 list_add(&req->ki_list, &ctx->active_reqs);
480
481 req->ki_cancel = cancel;
482
483 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
484}
485EXPORT_SYMBOL(kiocb_set_cancel_fn);
486
Fabian Frederickd52a8f92014-04-22 07:26:58 +0200487static int kiocb_cancel(struct kiocb *kiocb)
Kent Overstreet906b9732013-05-07 16:18:31 -0700488{
Kent Overstreet0460fef2013-05-07 16:18:49 -0700489 kiocb_cancel_fn *old, *cancel;
Kent Overstreet906b9732013-05-07 16:18:31 -0700490
Kent Overstreet0460fef2013-05-07 16:18:49 -0700491 /*
492 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
493 * actually has a cancel function, hence the cmpxchg()
494 */
Kent Overstreet906b9732013-05-07 16:18:31 -0700495
Kent Overstreet0460fef2013-05-07 16:18:49 -0700496 cancel = ACCESS_ONCE(kiocb->ki_cancel);
497 do {
498 if (!cancel || cancel == KIOCB_CANCELLED)
Kent Overstreet57282d82013-05-13 13:42:52 -0700499 return -EINVAL;
Kent Overstreet906b9732013-05-07 16:18:31 -0700500
Kent Overstreet0460fef2013-05-07 16:18:49 -0700501 old = cancel;
502 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
503 } while (cancel != old);
504
Kent Overstreet57282d82013-05-13 13:42:52 -0700505 return cancel(kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -0700506}
507
Kent Overstreete34ecee2013-10-10 19:31:47 -0700508static void free_ioctx(struct work_struct *work)
Kent Overstreet36f55882013-05-07 16:18:41 -0700509{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700510 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000511
Kent Overstreete34ecee2013-10-10 19:31:47 -0700512 pr_debug("freeing %p\n", ctx);
513
514 aio_free_ring(ctx);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000515 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400516 percpu_ref_exit(&ctx->reqs);
517 percpu_ref_exit(&ctx->users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700518 kmem_cache_free(kioctx_cachep, ctx);
519}
520
Kent Overstreete34ecee2013-10-10 19:31:47 -0700521static void free_ioctx_reqs(struct percpu_ref *ref)
522{
523 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
524
Anatol Pomozove02ba722014-04-15 11:31:33 -0700525 /* At this point we know that there are no any in-flight requests */
526 if (ctx->requests_done)
527 complete(ctx->requests_done);
528
Kent Overstreete34ecee2013-10-10 19:31:47 -0700529 INIT_WORK(&ctx->free_work, free_ioctx);
530 schedule_work(&ctx->free_work);
531}
532
Kent Overstreet36f55882013-05-07 16:18:41 -0700533/*
534 * When this function runs, the kioctx has been removed from the "hash table"
535 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
536 * now it's safe to cancel any that need to be.
537 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700538static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet36f55882013-05-07 16:18:41 -0700539{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700540 struct kioctx *ctx = container_of(ref, struct kioctx, users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700541 struct kiocb *req;
542
543 spin_lock_irq(&ctx->ctx_lock);
544
545 while (!list_empty(&ctx->active_reqs)) {
546 req = list_first_entry(&ctx->active_reqs,
547 struct kiocb, ki_list);
548
549 list_del_init(&req->ki_list);
Fabian Frederickd52a8f92014-04-22 07:26:58 +0200550 kiocb_cancel(req);
Kent Overstreet36f55882013-05-07 16:18:41 -0700551 }
552
553 spin_unlock_irq(&ctx->ctx_lock);
554
Kent Overstreete34ecee2013-10-10 19:31:47 -0700555 percpu_ref_kill(&ctx->reqs);
556 percpu_ref_put(&ctx->reqs);
Kent Overstreet36f55882013-05-07 16:18:41 -0700557}
558
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400559static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
560{
561 unsigned i, new_nr;
562 struct kioctx_table *table, *old;
563 struct aio_ring *ring;
564
565 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200566 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400567
568 while (1) {
569 if (table)
570 for (i = 0; i < table->nr; i++)
571 if (!table->table[i]) {
572 ctx->id = i;
573 table->table[i] = ctx;
574 spin_unlock(&mm->ioctx_lock);
575
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400576 /* While kioctx setup is in progress,
577 * we are protected from page migration
578 * changes ring_pages by ->ring_lock.
579 */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400580 ring = kmap_atomic(ctx->ring_pages[0]);
581 ring->id = ctx->id;
582 kunmap_atomic(ring);
583 return 0;
584 }
585
586 new_nr = (table ? table->nr : 1) * 4;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400587 spin_unlock(&mm->ioctx_lock);
588
589 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
590 new_nr, GFP_KERNEL);
591 if (!table)
592 return -ENOMEM;
593
594 table->nr = new_nr;
595
596 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200597 old = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400598
599 if (!old) {
600 rcu_assign_pointer(mm->ioctx_table, table);
601 } else if (table->nr > old->nr) {
602 memcpy(table->table, old->table,
603 old->nr * sizeof(struct kioctx *));
604
605 rcu_assign_pointer(mm->ioctx_table, table);
606 kfree_rcu(old, rcu);
607 } else {
608 kfree(table);
609 table = old;
610 }
611 }
612}
613
Kent Overstreete34ecee2013-10-10 19:31:47 -0700614static void aio_nr_sub(unsigned nr)
615{
616 spin_lock(&aio_nr_lock);
617 if (WARN_ON(aio_nr - nr > aio_nr))
618 aio_nr = 0;
619 else
620 aio_nr -= nr;
621 spin_unlock(&aio_nr_lock);
622}
623
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624/* ioctx_alloc
625 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
626 */
627static struct kioctx *ioctx_alloc(unsigned nr_events)
628{
Zach Brown41003a72013-05-07 16:18:25 -0700629 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 struct kioctx *ctx;
Al Viroe23754f2012-03-06 14:33:22 -0500631 int err = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000633 /*
634 * We keep track of the number of available ringbuffer slots, to prevent
635 * overflow (reqs_available), and we also use percpu counters for this.
636 *
637 * So since up to half the slots might be on other cpu's percpu counters
638 * and unavailable, double nr_events so userspace sees what they
639 * expected: additionally, we move req_batch slots to/from percpu
640 * counters at a time, so make sure that isn't 0:
641 */
642 nr_events = max(nr_events, num_possible_cpus() * 4);
643 nr_events *= 2;
644
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 /* Prevent overflows */
646 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
647 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
648 pr_debug("ENOMEM: nr_events too high\n");
649 return ERR_PTR(-EINVAL);
650 }
651
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400652 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 return ERR_PTR(-EAGAIN);
654
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800655 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 if (!ctx)
657 return ERR_PTR(-ENOMEM);
658
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 ctx->max_reqs = nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400661 spin_lock_init(&ctx->ctx_lock);
662 spin_lock_init(&ctx->completion_lock);
663 mutex_init(&ctx->ring_lock);
664 /* Protect against page migration throughout kiotx setup by keeping
665 * the ring_lock mutex held until setup is complete. */
666 mutex_lock(&ctx->ring_lock);
667 init_waitqueue_head(&ctx->wait);
668
669 INIT_LIST_HEAD(&ctx->active_reqs);
670
Tejun Heo2aad2a82014-09-24 13:31:50 -0400671 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700672 goto err;
673
Tejun Heo2aad2a82014-09-24 13:31:50 -0400674 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700675 goto err;
Kent Overstreet723be6e2013-05-28 15:14:48 -0700676
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000677 ctx->cpu = alloc_percpu(struct kioctx_cpu);
678 if (!ctx->cpu)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700679 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400681 err = aio_setup_ring(ctx);
682 if (err < 0)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700683 goto err;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000684
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000685 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000686 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
Benjamin LaHaise6878ea72013-07-31 10:34:18 -0400687 if (ctx->req_batch < 1)
688 ctx->req_batch = 1;
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000689
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 /* limit the number of system wide aios */
Al Viro9fa1cb32012-03-10 23:14:05 -0500691 spin_lock(&aio_nr_lock);
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400692 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
Al Viro2dd542b2012-03-10 23:10:35 -0500693 aio_nr + nr_events < aio_nr) {
Al Viro9fa1cb32012-03-10 23:14:05 -0500694 spin_unlock(&aio_nr_lock);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700695 err = -EAGAIN;
Gu Zhengd1b94322013-12-04 18:19:06 +0800696 goto err_ctx;
Al Viro2dd542b2012-03-10 23:10:35 -0500697 }
698 aio_nr += ctx->max_reqs;
Al Viro9fa1cb32012-03-10 23:14:05 -0500699 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700
Benjamin LaHaise18816862013-12-21 15:49:28 -0500701 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
702 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
Kent Overstreet723be6e2013-05-28 15:14:48 -0700703
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400704 err = ioctx_add_table(ctx, mm);
705 if (err)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700706 goto err_cleanup;
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400707
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400708 /* Release the ring_lock mutex now that all setup is complete. */
709 mutex_unlock(&ctx->ring_lock);
710
Kent Overstreetcaf41672013-05-07 16:18:35 -0700711 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700712 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 return ctx;
714
Kent Overstreete34ecee2013-10-10 19:31:47 -0700715err_cleanup:
716 aio_nr_sub(ctx->max_reqs);
Gu Zhengd1b94322013-12-04 18:19:06 +0800717err_ctx:
718 aio_free_ring(ctx);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700719err:
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400720 mutex_unlock(&ctx->ring_lock);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000721 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400722 percpu_ref_exit(&ctx->reqs);
723 percpu_ref_exit(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724 kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreetcaf41672013-05-07 16:18:35 -0700725 pr_debug("error allocating ioctx %d\n", err);
Al Viroe23754f2012-03-06 14:33:22 -0500726 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727}
728
Kent Overstreet36f55882013-05-07 16:18:41 -0700729/* kill_ioctx
730 * Cancels all outstanding aio requests on an aio context. Used
731 * when the processes owning a context have all exited to encourage
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 * the rapid destruction of the kioctx.
733 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400734static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
Anatol Pomozove02ba722014-04-15 11:31:33 -0700735 struct completion *requests_done)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736{
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400737 struct kioctx_table *table;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400738
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400739 if (atomic_xchg(&ctx->dead, 1))
740 return -EINVAL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400741
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400742
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400743 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200744 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400745 WARN_ON(ctx != table->table[ctx->id]);
746 table->table[ctx->id] = NULL;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400747 spin_unlock(&mm->ioctx_lock);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700748
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400749 /* percpu_ref_kill() will do the necessary call_rcu() */
750 wake_up_all(&ctx->wait);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700751
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400752 /*
753 * It'd be more correct to do this in free_ioctx(), after all
754 * the outstanding kiocbs have finished - but by then io_destroy
755 * has already returned, so io_setup() could potentially return
756 * -EAGAIN with no ioctxs actually in use (as far as userspace
757 * could tell).
758 */
759 aio_nr_sub(ctx->max_reqs);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400760
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400761 if (ctx->mmap_size)
762 vm_munmap(ctx->mmap_base, ctx->mmap_size);
763
764 ctx->requests_done = requests_done;
765 percpu_ref_kill(&ctx->users);
766 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767}
768
769/* wait_on_sync_kiocb:
770 * Waits on the given sync kiocb to complete.
771 */
Kent Overstreet57282d82013-05-13 13:42:52 -0700772ssize_t wait_on_sync_kiocb(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773{
Kent Overstreet57282d82013-05-13 13:42:52 -0700774 while (!req->ki_ctx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 set_current_state(TASK_UNINTERRUPTIBLE);
Kent Overstreet57282d82013-05-13 13:42:52 -0700776 if (req->ki_ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 break;
Jeff Moyer41d10da2007-10-16 23:27:20 -0700778 io_schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 }
780 __set_current_state(TASK_RUNNING);
Kent Overstreet57282d82013-05-13 13:42:52 -0700781 return req->ki_user_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700783EXPORT_SYMBOL(wait_on_sync_kiocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784
Kent Overstreet36f55882013-05-07 16:18:41 -0700785/*
786 * exit_aio: called when the last user of mm goes away. At this point, there is
787 * no way for any new requests to be submited or any of the io_* syscalls to be
788 * called on the context.
789 *
790 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
791 * them.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800793void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794{
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200795 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
796 int i;
Jens Axboeabf137d2008-12-09 08:11:22 +0100797
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200798 if (!table)
799 return;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400800
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200801 for (i = 0; i < table->nr; ++i) {
802 struct kioctx *ctx = table->table[i];
Gu Zheng6098b452014-09-03 17:45:44 +0800803 struct completion requests_done =
804 COMPLETION_INITIALIZER_ONSTACK(requests_done);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400805
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200806 if (!ctx)
807 continue;
Al Viro936af152012-04-20 21:49:41 -0400808 /*
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200809 * We don't need to bother with munmap() here - exit_mmap(mm)
810 * is coming and it'll unmap everything. And we simply can't,
811 * this is not necessarily our ->mm.
812 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
813 * that it needs to unmap the area, just set it to 0.
Al Viro936af152012-04-20 21:49:41 -0400814 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700815 ctx->mmap_size = 0;
Gu Zheng6098b452014-09-03 17:45:44 +0800816 kill_ioctx(mm, ctx, &requests_done);
Kent Overstreet36f55882013-05-07 16:18:41 -0700817
Gu Zheng6098b452014-09-03 17:45:44 +0800818 /* Wait until all IO for the context are done. */
819 wait_for_completion(&requests_done);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 }
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200821
822 RCU_INIT_POINTER(mm->ioctx_table, NULL);
823 kfree(table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824}
825
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000826static void put_reqs_available(struct kioctx *ctx, unsigned nr)
827{
828 struct kioctx_cpu *kcpu;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400829 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000830
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400831 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400832 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000833 kcpu->reqs_available += nr;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400834
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000835 while (kcpu->reqs_available >= ctx->req_batch * 2) {
836 kcpu->reqs_available -= ctx->req_batch;
837 atomic_add(ctx->req_batch, &ctx->reqs_available);
838 }
839
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400840 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000841}
842
843static bool get_reqs_available(struct kioctx *ctx)
844{
845 struct kioctx_cpu *kcpu;
846 bool ret = false;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400847 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000848
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400849 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400850 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000851 if (!kcpu->reqs_available) {
852 int old, avail = atomic_read(&ctx->reqs_available);
853
854 do {
855 if (avail < ctx->req_batch)
856 goto out;
857
858 old = avail;
859 avail = atomic_cmpxchg(&ctx->reqs_available,
860 avail, avail - ctx->req_batch);
861 } while (avail != old);
862
863 kcpu->reqs_available += ctx->req_batch;
864 }
865
866 ret = true;
867 kcpu->reqs_available--;
868out:
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400869 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000870 return ret;
871}
872
Benjamin LaHaised856f322014-08-24 13:14:05 -0400873/* refill_reqs_available
874 * Updates the reqs_available reference counts used for tracking the
875 * number of free slots in the completion ring. This can be called
876 * from aio_complete() (to optimistically update reqs_available) or
877 * from aio_get_req() (the we're out of events case). It must be
878 * called holding ctx->completion_lock.
879 */
880static void refill_reqs_available(struct kioctx *ctx, unsigned head,
881 unsigned tail)
882{
883 unsigned events_in_ring, completed;
884
885 /* Clamp head since userland can write to it. */
886 head %= ctx->nr_events;
887 if (head <= tail)
888 events_in_ring = tail - head;
889 else
890 events_in_ring = ctx->nr_events - (head - tail);
891
892 completed = ctx->completed_events;
893 if (events_in_ring < completed)
894 completed -= events_in_ring;
895 else
896 completed = 0;
897
898 if (!completed)
899 return;
900
901 ctx->completed_events -= completed;
902 put_reqs_available(ctx, completed);
903}
904
905/* user_refill_reqs_available
906 * Called to refill reqs_available when aio_get_req() encounters an
907 * out of space in the completion ring.
908 */
909static void user_refill_reqs_available(struct kioctx *ctx)
910{
911 spin_lock_irq(&ctx->completion_lock);
912 if (ctx->completed_events) {
913 struct aio_ring *ring;
914 unsigned head;
915
916 /* Access of ring->head may race with aio_read_events_ring()
917 * here, but that's okay since whether we read the old version
918 * or the new version, and either will be valid. The important
919 * part is that head cannot pass tail since we prevent
920 * aio_complete() from updating tail by holding
921 * ctx->completion_lock. Even if head is invalid, the check
922 * against ctx->completed_events below will make sure we do the
923 * safe/right thing.
924 */
925 ring = kmap_atomic(ctx->ring_pages[0]);
926 head = ring->head;
927 kunmap_atomic(ring);
928
929 refill_reqs_available(ctx, head, ctx->tail);
930 }
931
932 spin_unlock_irq(&ctx->completion_lock);
933}
934
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935/* aio_get_req
Kent Overstreet57282d82013-05-13 13:42:52 -0700936 * Allocate a slot for an aio request.
937 * Returns NULL if no requests are free.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 */
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700939static inline struct kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940{
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700941 struct kiocb *req;
942
Benjamin LaHaised856f322014-08-24 13:14:05 -0400943 if (!get_reqs_available(ctx)) {
944 user_refill_reqs_available(ctx);
945 if (!get_reqs_available(ctx))
946 return NULL;
947 }
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700948
Kent Overstreet0460fef2013-05-07 16:18:49 -0700949 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 if (unlikely(!req))
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700951 goto out_put;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952
Kent Overstreete34ecee2013-10-10 19:31:47 -0700953 percpu_ref_get(&ctx->reqs);
954
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 req->ki_ctx = ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 return req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700957out_put:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000958 put_reqs_available(ctx, 1);
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700959 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960}
961
Kent Overstreet11599eb2013-05-07 16:18:39 -0700962static void kiocb_free(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963{
Kent Overstreet1d98ebf2013-05-07 16:18:37 -0700964 if (req->ki_filp)
965 fput(req->ki_filp);
Davide Libenzi13389012009-06-30 11:41:11 -0700966 if (req->ki_eventfd != NULL)
967 eventfd_ctx_put(req->ki_eventfd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 kmem_cache_free(kiocb_cachep, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969}
970
Adrian Bunkd5470b52008-04-29 00:58:57 -0700971static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400973 struct aio_ring __user *ring = (void __user *)ctx_id;
Jens Axboeabf137d2008-12-09 08:11:22 +0100974 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -0700975 struct kioctx *ctx, *ret = NULL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400976 struct kioctx_table *table;
977 unsigned id;
978
979 if (get_user(id, &ring->id))
980 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981
Jens Axboeabf137d2008-12-09 08:11:22 +0100982 rcu_read_lock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400983 table = rcu_dereference(mm->ioctx_table);
Jens Axboeabf137d2008-12-09 08:11:22 +0100984
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400985 if (!table || id >= table->nr)
986 goto out;
987
988 ctx = table->table[id];
Benjamin LaHaisef30d7042013-08-07 18:23:48 -0400989 if (ctx && ctx->user_id == ctx_id) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400990 percpu_ref_get(&ctx->users);
991 ret = ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +0100992 }
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400993out:
Jens Axboeabf137d2008-12-09 08:11:22 +0100994 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -0700995 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996}
997
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998/* aio_complete
999 * Called when the io request on the given iocb is complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 */
Kent Overstreet2d684492013-05-07 16:18:29 -07001001void aio_complete(struct kiocb *iocb, long res, long res2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002{
1003 struct kioctx *ctx = iocb->ki_ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 struct aio_ring *ring;
Kent Overstreet21b40202013-05-07 16:18:47 -07001005 struct io_event *ev_page, *event;
Benjamin LaHaised856f322014-08-24 13:14:05 -04001006 unsigned tail, pos, head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008
Zach Brown20dcae32005-11-13 16:07:33 -08001009 /*
1010 * Special case handling for sync iocbs:
1011 * - events go directly into the iocb for fast handling
1012 * - the sync task with the iocb in its stack holds the single iocb
1013 * ref, no other paths have a way to get another ref
1014 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015 */
1016 if (is_sync_kiocb(iocb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017 iocb->ki_user_data = res;
Kent Overstreet57282d82013-05-13 13:42:52 -07001018 smp_wmb();
1019 iocb->ki_ctx = ERR_PTR(-EXDEV);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020 wake_up_process(iocb->ki_obj.tsk);
Kent Overstreet2d684492013-05-07 16:18:29 -07001021 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 }
1023
Kent Overstreet0460fef2013-05-07 16:18:49 -07001024 if (iocb->ki_list.next) {
1025 unsigned long flags;
1026
1027 spin_lock_irqsave(&ctx->ctx_lock, flags);
1028 list_del(&iocb->ki_list);
1029 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1030 }
Kent Overstreet11599eb2013-05-07 16:18:39 -07001031
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032 /*
Kent Overstreet0460fef2013-05-07 16:18:49 -07001033 * Add a completion event to the ring buffer. Must be done holding
Tang Chen4b30f072013-07-03 15:09:16 -07001034 * ctx->completion_lock to prevent other code from messing with the tail
Kent Overstreet0460fef2013-05-07 16:18:49 -07001035 * pointer since we might be called from irq context.
1036 */
1037 spin_lock_irqsave(&ctx->completion_lock, flags);
1038
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001039 tail = ctx->tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001040 pos = tail + AIO_EVENTS_OFFSET;
1041
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001042 if (++tail >= ctx->nr_events)
Ken Chen4bf69b22005-05-01 08:59:15 -07001043 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001045 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001046 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1047
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1049 event->data = iocb->ki_user_data;
1050 event->res = res;
1051 event->res2 = res2;
1052
Kent Overstreet21b40202013-05-07 16:18:47 -07001053 kunmap_atomic(ev_page);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001054 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001055
1056 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
Kent Overstreetcaf41672013-05-07 16:18:35 -07001057 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1058 res, res2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059
1060 /* after flagging the request as done, we
1061 * must never even look at it again
1062 */
1063 smp_wmb(); /* make event visible before updating tail */
1064
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001065 ctx->tail = tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001066
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001067 ring = kmap_atomic(ctx->ring_pages[0]);
Benjamin LaHaised856f322014-08-24 13:14:05 -04001068 head = ring->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 ring->tail = tail;
Cong Wange8e3c3d2011-11-25 23:14:27 +08001070 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001071 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072
Benjamin LaHaised856f322014-08-24 13:14:05 -04001073 ctx->completed_events++;
1074 if (ctx->completed_events > 1)
1075 refill_reqs_available(ctx, head, tail);
Kent Overstreet0460fef2013-05-07 16:18:49 -07001076 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1077
Kent Overstreet21b40202013-05-07 16:18:47 -07001078 pr_debug("added to ring %p at [%u]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001079
1080 /*
1081 * Check if the user asked us to deliver the result through an
1082 * eventfd. The eventfd_signal() function is safe to be called
1083 * from IRQ context.
1084 */
Davide Libenzi87c3a862009-03-18 17:04:19 -07001085 if (iocb->ki_eventfd != NULL)
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001086 eventfd_signal(iocb->ki_eventfd, 1);
1087
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088 /* everything turned out well, dispose of the aiocb. */
Kent Overstreet57282d82013-05-13 13:42:52 -07001089 kiocb_free(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090
Quentin Barnes6cb2a212008-03-19 17:00:39 -07001091 /*
1092 * We have to order our ring_info tail store above and test
1093 * of the wait list below outside the wait lock. This is
1094 * like in wake_up_bit() where clearing a bit has to be
1095 * ordered with the unlocked test.
1096 */
1097 smp_mb();
1098
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 if (waitqueue_active(&ctx->wait))
1100 wake_up(&ctx->wait);
1101
Kent Overstreete34ecee2013-10-10 19:31:47 -07001102 percpu_ref_put(&ctx->reqs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103}
H Hartley Sweeten385773e2009-09-22 16:43:53 -07001104EXPORT_SYMBOL(aio_complete);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001106/* aio_read_events_ring
Kent Overstreeta31ad382013-05-07 16:18:45 -07001107 * Pull an event off of the ioctx's event ring. Returns the number of
1108 * events fetched
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 */
Kent Overstreeta31ad382013-05-07 16:18:45 -07001110static long aio_read_events_ring(struct kioctx *ctx,
1111 struct io_event __user *event, long nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 struct aio_ring *ring;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001114 unsigned head, tail, pos;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001115 long ret = 0;
1116 int copy_ret;
1117
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001118 mutex_lock(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -04001120 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001121 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001122 head = ring->head;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001123 tail = ring->tail;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001124 kunmap_atomic(ring);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125
Jeff Moyer2ff396be2014-09-02 13:17:00 -04001126 /*
1127 * Ensure that once we've read the current tail pointer, that
1128 * we also see the events that were stored up to the tail.
1129 */
1130 smp_rmb();
1131
Kent Overstreet5ffac122013-05-09 15:36:07 -07001132 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001133
Kent Overstreet5ffac122013-05-09 15:36:07 -07001134 if (head == tail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 goto out;
1136
Benjamin LaHaiseedfbbf32014-06-24 13:32:51 -04001137 head %= ctx->nr_events;
1138 tail %= ctx->nr_events;
1139
Kent Overstreeta31ad382013-05-07 16:18:45 -07001140 while (ret < nr) {
1141 long avail;
1142 struct io_event *ev;
1143 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144
Kent Overstreet5ffac122013-05-09 15:36:07 -07001145 avail = (head <= tail ? tail : ctx->nr_events) - head;
1146 if (head == tail)
Kent Overstreeta31ad382013-05-07 16:18:45 -07001147 break;
1148
1149 avail = min(avail, nr - ret);
1150 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1151 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1152
1153 pos = head + AIO_EVENTS_OFFSET;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001154 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
Kent Overstreeta31ad382013-05-07 16:18:45 -07001155 pos %= AIO_EVENTS_PER_PAGE;
1156
1157 ev = kmap(page);
1158 copy_ret = copy_to_user(event + ret, ev + pos,
1159 sizeof(*ev) * avail);
1160 kunmap(page);
1161
1162 if (unlikely(copy_ret)) {
1163 ret = -EFAULT;
1164 goto out;
1165 }
1166
1167 ret += avail;
1168 head += avail;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001169 head %= ctx->nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001172 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001173 ring->head = head;
Zhao Hongjiang91d80a82013-04-26 11:03:53 +08001174 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001175 flush_dcache_page(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001176
Kent Overstreet5ffac122013-05-09 15:36:07 -07001177 pr_debug("%li h%u t%u\n", ret, head, tail);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001178out:
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001179 mutex_unlock(&ctx->ring_lock);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001180
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 return ret;
1182}
1183
Kent Overstreeta31ad382013-05-07 16:18:45 -07001184static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1185 struct io_event __user *event, long *i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001187 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188
Kent Overstreeta31ad382013-05-07 16:18:45 -07001189 if (ret > 0)
1190 *i += ret;
1191
1192 if (unlikely(atomic_read(&ctx->dead)))
1193 ret = -EINVAL;
1194
1195 if (!*i)
1196 *i = ret;
1197
1198 return ret < 0 || *i >= min_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199}
1200
Kent Overstreeta31ad382013-05-07 16:18:45 -07001201static long read_events(struct kioctx *ctx, long min_nr, long nr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 struct io_event __user *event,
1203 struct timespec __user *timeout)
1204{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001205 ktime_t until = { .tv64 = KTIME_MAX };
1206 long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 if (timeout) {
1209 struct timespec ts;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001210
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
Kent Overstreeta31ad382013-05-07 16:18:45 -07001212 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213
Kent Overstreeta31ad382013-05-07 16:18:45 -07001214 until = timespec_to_ktime(ts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 }
1216
Kent Overstreeta31ad382013-05-07 16:18:45 -07001217 /*
1218 * Note that aio_read_events() is being called as the conditional - i.e.
1219 * we're calling it after prepare_to_wait() has set task state to
1220 * TASK_INTERRUPTIBLE.
1221 *
1222 * But aio_read_events() can block, and if it blocks it's going to flip
1223 * the task state back to TASK_RUNNING.
1224 *
1225 * This should be ok, provided it doesn't flip the state back to
1226 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1227 * will only happen if the mutex_lock() call blocks, and we then find
1228 * the ringbuffer empty. So in practice we should be ok, but it's
1229 * something to be aware of when touching this code.
1230 */
1231 wait_event_interruptible_hrtimeout(ctx->wait,
1232 aio_read_events(ctx, min_nr, nr, event, &ret), until);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233
Kent Overstreeta31ad382013-05-07 16:18:45 -07001234 if (!ret && signal_pending(current))
1235 ret = -EINTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236
Kent Overstreeta31ad382013-05-07 16:18:45 -07001237 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238}
1239
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240/* sys_io_setup:
1241 * Create an aio_context capable of receiving at least nr_events.
1242 * ctxp must not point to an aio_context that already exists, and
1243 * must be initialized to 0 prior to the call. On successful
1244 * creation of the aio_context, *ctxp is filled in with the resulting
1245 * handle. May fail with -EINVAL if *ctxp is not initialized,
1246 * if the specified nr_events exceeds internal limits. May fail
1247 * with -EAGAIN if the specified nr_events exceeds the user's limit
1248 * of available events. May fail with -ENOMEM if insufficient kernel
1249 * resources are available. May fail with -EFAULT if an invalid
1250 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1251 * implemented.
1252 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001253SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254{
1255 struct kioctx *ioctx = NULL;
1256 unsigned long ctx;
1257 long ret;
1258
1259 ret = get_user(ctx, ctxp);
1260 if (unlikely(ret))
1261 goto out;
1262
1263 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001264 if (unlikely(ctx || nr_events == 0)) {
1265 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1266 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 goto out;
1268 }
1269
1270 ioctx = ioctx_alloc(nr_events);
1271 ret = PTR_ERR(ioctx);
1272 if (!IS_ERR(ioctx)) {
1273 ret = put_user(ioctx->user_id, ctxp);
Al Viroa2e18592012-03-20 16:27:57 -04001274 if (ret)
Anatol Pomozove02ba722014-04-15 11:31:33 -07001275 kill_ioctx(current->mm, ioctx, NULL);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001276 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 }
1278
1279out:
1280 return ret;
1281}
1282
1283/* sys_io_destroy:
1284 * Destroy the aio_context specified. May cancel any outstanding
1285 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001286 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 * is invalid.
1288 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001289SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290{
1291 struct kioctx *ioctx = lookup_ioctx(ctx);
1292 if (likely(NULL != ioctx)) {
Anatol Pomozove02ba722014-04-15 11:31:33 -07001293 struct completion requests_done =
1294 COMPLETION_INITIALIZER_ONSTACK(requests_done);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001295 int ret;
Anatol Pomozove02ba722014-04-15 11:31:33 -07001296
1297 /* Pass requests_done to kill_ioctx() where it can be set
1298 * in a thread-safe way. If we try to set it here then we have
1299 * a race condition if two io_destroy() called simultaneously.
1300 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001301 ret = kill_ioctx(current->mm, ioctx, &requests_done);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001302 percpu_ref_put(&ioctx->users);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001303
1304 /* Wait until all IO for the context are done. Otherwise kernel
1305 * keep using user-space buffers even if user thinks the context
1306 * is destroyed.
1307 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001308 if (!ret)
1309 wait_for_completion(&requests_done);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001310
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001311 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 }
1313 pr_debug("EINVAL: io_destroy: invalid context id\n");
1314 return -EINVAL;
1315}
1316
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001317typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1318 unsigned long, loff_t);
Al Viro293bc982014-02-11 18:37:41 -05001319typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001320
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001321static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1322 int rw, char __user *buf,
1323 unsigned long *nr_segs,
1324 struct iovec **iovec,
1325 bool compat)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001326{
1327 ssize_t ret;
1328
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001329 *nr_segs = kiocb->ki_nbytes;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001330
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001331#ifdef CONFIG_COMPAT
1332 if (compat)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001333 ret = compat_rw_copy_check_uvector(rw,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001334 (struct compat_iovec __user *)buf,
Gu Zheng00fefb92014-07-23 18:03:54 +08001335 *nr_segs, UIO_FASTIOV, *iovec, iovec);
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001336 else
1337#endif
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001338 ret = rw_copy_check_uvector(rw,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001339 (struct iovec __user *)buf,
Gu Zheng00fefb92014-07-23 18:03:54 +08001340 *nr_segs, UIO_FASTIOV, *iovec, iovec);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001341 if (ret < 0)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001342 return ret;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001343
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001344 /* ki_nbytes now reflect bytes instead of segs */
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001345 kiocb->ki_nbytes = ret;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001346 return 0;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001347}
1348
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001349static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1350 int rw, char __user *buf,
1351 unsigned long *nr_segs,
1352 struct iovec *iovec)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001353{
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001354 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001355 return -EFAULT;
Linus Torvaldsa70b52e2012-05-21 16:06:20 -07001356
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001357 iovec->iov_base = buf;
1358 iovec->iov_len = kiocb->ki_nbytes;
1359 *nr_segs = 1;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001360 return 0;
1361}
1362
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363/*
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001364 * aio_run_iocb:
1365 * Performs the initial checks and io submission.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366 */
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001367static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1368 char __user *buf, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369{
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001370 struct file *file = req->ki_filp;
1371 ssize_t ret;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001372 unsigned long nr_segs;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001373 int rw;
1374 fmode_t mode;
1375 aio_rw_op *rw_op;
Al Viro293bc982014-02-11 18:37:41 -05001376 rw_iter_op *iter_op;
Gu Zheng00fefb92014-07-23 18:03:54 +08001377 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
Al Viro293bc982014-02-11 18:37:41 -05001378 struct iov_iter iter;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001380 switch (opcode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 case IOCB_CMD_PREAD:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001382 case IOCB_CMD_PREADV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001383 mode = FMODE_READ;
1384 rw = READ;
1385 rw_op = file->f_op->aio_read;
Al Viro293bc982014-02-11 18:37:41 -05001386 iter_op = file->f_op->read_iter;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001387 goto rw_common;
1388
1389 case IOCB_CMD_PWRITE:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001390 case IOCB_CMD_PWRITEV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001391 mode = FMODE_WRITE;
1392 rw = WRITE;
1393 rw_op = file->f_op->aio_write;
Al Viro293bc982014-02-11 18:37:41 -05001394 iter_op = file->f_op->write_iter;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001395 goto rw_common;
1396rw_common:
1397 if (unlikely(!(file->f_mode & mode)))
1398 return -EBADF;
1399
Al Viro293bc982014-02-11 18:37:41 -05001400 if (!rw_op && !iter_op)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001401 return -EINVAL;
1402
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001403 ret = (opcode == IOCB_CMD_PREADV ||
1404 opcode == IOCB_CMD_PWRITEV)
1405 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1406 &iovec, compat)
1407 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1408 iovec);
Leon Yu754320d2014-05-01 03:31:28 +00001409 if (!ret)
1410 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001411 if (ret < 0) {
Gu Zheng00fefb92014-07-23 18:03:54 +08001412 if (iovec != inline_vecs)
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001413 kfree(iovec);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001414 return ret;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001415 }
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001416
1417 req->ki_nbytes = ret;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001418
Kent Overstreet73a70752013-05-09 15:03:42 -07001419 /* XXX: move/kill - rw_verify_area()? */
1420 /* This matches the pread()/pwrite() logic */
1421 if (req->ki_pos < 0) {
1422 ret = -EINVAL;
1423 break;
1424 }
1425
1426 if (rw == WRITE)
1427 file_start_write(file);
1428
Al Viro293bc982014-02-11 18:37:41 -05001429 if (iter_op) {
1430 iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes);
1431 ret = iter_op(req, &iter);
1432 } else {
1433 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1434 }
Kent Overstreet73a70752013-05-09 15:03:42 -07001435
1436 if (rw == WRITE)
1437 file_end_write(file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001439
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 case IOCB_CMD_FDSYNC:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001441 if (!file->f_op->aio_fsync)
1442 return -EINVAL;
1443
1444 ret = file->f_op->aio_fsync(req, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001446
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 case IOCB_CMD_FSYNC:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001448 if (!file->f_op->aio_fsync)
1449 return -EINVAL;
1450
1451 ret = file->f_op->aio_fsync(req, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001453
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 default:
Kent Overstreetcaf41672013-05-07 16:18:35 -07001455 pr_debug("EINVAL: no operation provided\n");
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001456 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 }
1458
Gu Zheng00fefb92014-07-23 18:03:54 +08001459 if (iovec != inline_vecs)
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001460 kfree(iovec);
1461
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001462 if (ret != -EIOCBQUEUED) {
1463 /*
1464 * There's no easy way to restart the syscall since other AIO's
1465 * may be already running. Just fail this IO with EINTR.
1466 */
1467 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1468 ret == -ERESTARTNOHAND ||
1469 ret == -ERESTART_RESTARTBLOCK))
1470 ret = -EINTR;
1471 aio_complete(req, ret, 0);
1472 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473
1474 return 0;
1475}
1476
Adrian Bunkd5470b52008-04-29 00:58:57 -07001477static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001478 struct iocb *iocb, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479{
1480 struct kiocb *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481 ssize_t ret;
1482
1483 /* enforce forwards compatibility on users */
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001484 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001485 pr_debug("EINVAL: reserve field set\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486 return -EINVAL;
1487 }
1488
1489 /* prevent overflows */
1490 if (unlikely(
1491 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1492 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1493 ((ssize_t)iocb->aio_nbytes < 0)
1494 )) {
1495 pr_debug("EINVAL: io_submit: overflow check\n");
1496 return -EINVAL;
1497 }
1498
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001499 req = aio_get_req(ctx);
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001500 if (unlikely(!req))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501 return -EAGAIN;
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001502
1503 req->ki_filp = fget(iocb->aio_fildes);
1504 if (unlikely(!req->ki_filp)) {
1505 ret = -EBADF;
1506 goto out_put_req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507 }
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001508
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001509 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1510 /*
1511 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1512 * instance of the file* now. The file descriptor must be
1513 * an eventfd() fd, and will be signaled for each completed
1514 * event using the eventfd_signal() function.
1515 */
Davide Libenzi13389012009-06-30 11:41:11 -07001516 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
Hirofumi Nakagawa801678c2008-04-29 01:03:09 -07001517 if (IS_ERR(req->ki_eventfd)) {
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001518 ret = PTR_ERR(req->ki_eventfd);
Davide Libenzi87c3a862009-03-18 17:04:19 -07001519 req->ki_eventfd = NULL;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001520 goto out_put_req;
1521 }
1522 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523
Kent Overstreet8a660892013-05-07 16:19:10 -07001524 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 if (unlikely(ret)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001526 pr_debug("EFAULT: aio_key\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527 goto out_put_req;
1528 }
1529
1530 req->ki_obj.user = user_iocb;
1531 req->ki_user_data = iocb->aio_data;
1532 req->ki_pos = iocb->aio_offset;
Kent Overstreet73a70752013-05-09 15:03:42 -07001533 req->ki_nbytes = iocb->aio_nbytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001535 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1536 (char __user *)(unsigned long)iocb->aio_buf,
1537 compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538 if (ret)
1539 goto out_put_req;
1540
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542out_put_req:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001543 put_reqs_available(ctx, 1);
Kent Overstreete34ecee2013-10-10 19:31:47 -07001544 percpu_ref_put(&ctx->reqs);
Kent Overstreet57282d82013-05-13 13:42:52 -07001545 kiocb_free(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546 return ret;
1547}
1548
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001549long do_io_submit(aio_context_t ctx_id, long nr,
1550 struct iocb __user *__user *iocbpp, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551{
1552 struct kioctx *ctx;
1553 long ret = 0;
Jeff Moyer080d6762011-11-02 13:40:10 -07001554 int i = 0;
Shaohua Li9f5b9422010-07-01 07:55:01 +02001555 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556
1557 if (unlikely(nr < 0))
1558 return -EINVAL;
1559
Jeff Moyer75e1c702010-09-10 14:16:00 -07001560 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1561 nr = LONG_MAX/sizeof(*iocbpp);
1562
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1564 return -EFAULT;
1565
1566 ctx = lookup_ioctx(ctx_id);
1567 if (unlikely(!ctx)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001568 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 return -EINVAL;
1570 }
1571
Shaohua Li9f5b9422010-07-01 07:55:01 +02001572 blk_start_plug(&plug);
1573
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574 /*
1575 * AKPM: should this return a partial result if some of the IOs were
1576 * successfully submitted?
1577 */
1578 for (i=0; i<nr; i++) {
1579 struct iocb __user *user_iocb;
1580 struct iocb tmp;
1581
1582 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1583 ret = -EFAULT;
1584 break;
1585 }
1586
1587 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1588 ret = -EFAULT;
1589 break;
1590 }
1591
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001592 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 if (ret)
1594 break;
1595 }
Shaohua Li9f5b9422010-07-01 07:55:01 +02001596 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597
Kent Overstreet723be6e2013-05-28 15:14:48 -07001598 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599 return i ? i : ret;
1600}
1601
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001602/* sys_io_submit:
1603 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1604 * the number of iocbs queued. May return -EINVAL if the aio_context
1605 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1606 * *iocbpp[0] is not properly initialized, if the operation specified
1607 * is invalid for the file descriptor in the iocb. May fail with
1608 * -EFAULT if any of the data structures point to invalid data. May
1609 * fail with -EBADF if the file descriptor specified in the first
1610 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1611 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1612 * fail with -ENOSYS if not implemented.
1613 */
1614SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1615 struct iocb __user * __user *, iocbpp)
1616{
1617 return do_io_submit(ctx_id, nr, iocbpp, 0);
1618}
1619
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620/* lookup_kiocb
1621 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001623static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1624 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625{
1626 struct list_head *pos;
Zach Brownd00689a2005-11-13 16:07:34 -08001627
1628 assert_spin_locked(&ctx->ctx_lock);
1629
Kent Overstreet8a660892013-05-07 16:19:10 -07001630 if (key != KIOCB_KEY)
1631 return NULL;
1632
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633 /* TODO: use a hash or array, this sucks. */
1634 list_for_each(pos, &ctx->active_reqs) {
1635 struct kiocb *kiocb = list_kiocb(pos);
Kent Overstreet8a660892013-05-07 16:19:10 -07001636 if (kiocb->ki_obj.user == iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637 return kiocb;
1638 }
1639 return NULL;
1640}
1641
1642/* sys_io_cancel:
1643 * Attempts to cancel an iocb previously passed to io_submit. If
1644 * the operation is successfully cancelled, the resulting event is
1645 * copied into the memory pointed to by result without being placed
1646 * into the completion queue and 0 is returned. May fail with
1647 * -EFAULT if any of the data structures pointed to are invalid.
1648 * May fail with -EINVAL if aio_context specified by ctx_id is
1649 * invalid. May fail with -EAGAIN if the iocb specified was not
1650 * cancelled. Will fail with -ENOSYS if not implemented.
1651 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001652SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1653 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 struct kioctx *ctx;
1656 struct kiocb *kiocb;
1657 u32 key;
1658 int ret;
1659
1660 ret = get_user(key, &iocb->aio_key);
1661 if (unlikely(ret))
1662 return -EFAULT;
1663
1664 ctx = lookup_ioctx(ctx_id);
1665 if (unlikely(!ctx))
1666 return -EINVAL;
1667
1668 spin_lock_irq(&ctx->ctx_lock);
Kent Overstreet906b9732013-05-07 16:18:31 -07001669
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670 kiocb = lookup_kiocb(ctx, iocb, key);
Kent Overstreet906b9732013-05-07 16:18:31 -07001671 if (kiocb)
Fabian Frederickd52a8f92014-04-22 07:26:58 +02001672 ret = kiocb_cancel(kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -07001673 else
1674 ret = -EINVAL;
1675
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676 spin_unlock_irq(&ctx->ctx_lock);
1677
Kent Overstreet906b9732013-05-07 16:18:31 -07001678 if (!ret) {
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001679 /*
1680 * The result argument is no longer used - the io_event is
1681 * always delivered via the ring buffer. -EINPROGRESS indicates
1682 * cancellation is progress:
Kent Overstreet906b9732013-05-07 16:18:31 -07001683 */
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001684 ret = -EINPROGRESS;
Kent Overstreet906b9732013-05-07 16:18:31 -07001685 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
Kent Overstreet723be6e2013-05-28 15:14:48 -07001687 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688
1689 return ret;
1690}
1691
1692/* io_getevents:
1693 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001694 * the completion queue for the aio_context specified by ctx_id. If
1695 * it succeeds, the number of read events is returned. May fail with
1696 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1697 * out of range, if timeout is out of range. May fail with -EFAULT
1698 * if any of the memory specified is invalid. May return 0 or
1699 * < min_nr if the timeout specified by timeout has elapsed
1700 * before sufficient events are available, where timeout == NULL
1701 * specifies an infinite timeout. Note that the timeout pointed to by
Jeff Moyer69008072013-05-24 15:55:24 -07001702 * timeout is relative. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001704SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1705 long, min_nr,
1706 long, nr,
1707 struct io_event __user *, events,
1708 struct timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709{
1710 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1711 long ret = -EINVAL;
1712
1713 if (likely(ioctx)) {
Namhyung Kim2e410252011-01-12 17:01:08 -08001714 if (likely(min_nr <= nr && min_nr >= 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 ret = read_events(ioctx, min_nr, nr, events, timeout);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001716 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 return ret;
1719}