blob: 84a751005f5b8ad0f0f5cd6e3c5ec67df5f7440c [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
168static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
169{
170 struct qstr this = QSTR_INIT("[aio]", 5);
171 struct file *file;
172 struct path path;
173 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
Dan Carpenter7f626562013-11-13 10:49:40 +0300174 if (IS_ERR(inode))
175 return ERR_CAST(inode);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400176
177 inode->i_mapping->a_ops = &aio_ctx_aops;
178 inode->i_mapping->private_data = ctx;
179 inode->i_size = PAGE_SIZE * nr_pages;
180
181 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
182 if (!path.dentry) {
183 iput(inode);
184 return ERR_PTR(-ENOMEM);
185 }
186 path.mnt = mntget(aio_mnt);
187
188 d_instantiate(path.dentry, inode);
189 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
190 if (IS_ERR(file)) {
191 path_put(&path);
192 return file;
193 }
194
195 file->f_flags = O_RDWR;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400196 return file;
197}
198
199static struct dentry *aio_mount(struct file_system_type *fs_type,
200 int flags, const char *dev_name, void *data)
201{
202 static const struct dentry_operations ops = {
203 .d_dname = simple_dname,
204 };
Gu Zheng8dc43792014-07-23 18:03:52 +0800205 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400206}
207
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208/* aio_setup
209 * Creates the slab caches used by the aio routines, panic on
210 * failure as this is done early during the boot sequence.
211 */
212static int __init aio_setup(void)
213{
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400214 static struct file_system_type aio_fs = {
215 .name = "aio",
216 .mount = aio_mount,
217 .kill_sb = kill_anon_super,
218 };
219 aio_mnt = kern_mount(&aio_fs);
220 if (IS_ERR(aio_mnt))
221 panic("Failed to create aio fs mount.");
222
Christoph Lameter0a31bd52007-05-06 14:49:57 -0700223 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
224 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225
Kent Overstreetcaf41672013-05-07 16:18:35 -0700226 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227
228 return 0;
229}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700230__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400232static void put_aio_ring_file(struct kioctx *ctx)
233{
234 struct file *aio_ring_file = ctx->aio_ring_file;
235 if (aio_ring_file) {
236 truncate_setsize(aio_ring_file->f_inode, 0);
237
238 /* Prevent further access to the kioctx from migratepages */
239 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
240 aio_ring_file->f_inode->i_mapping->private_data = NULL;
241 ctx->aio_ring_file = NULL;
242 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
243
244 fput(aio_ring_file);
245 }
246}
247
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248static void aio_free_ring(struct kioctx *ctx)
249{
Gu Zheng36bc08c2013-07-16 17:56:16 +0800250 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400252 /* Disconnect the kiotx from the ring file. This prevents future
253 * accesses to the kioctx from page migration.
254 */
255 put_aio_ring_file(ctx);
256
Gu Zheng36bc08c2013-07-16 17:56:16 +0800257 for (i = 0; i < ctx->nr_pages; i++) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500258 struct page *page;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800259 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
260 page_count(ctx->ring_pages[i]));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500261 page = ctx->ring_pages[i];
262 if (!page)
263 continue;
264 ctx->ring_pages[i] = NULL;
265 put_page(page);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800266 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267
Sasha Levinddb8c452013-11-19 17:33:03 -0500268 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700269 kfree(ctx->ring_pages);
Sasha Levinddb8c452013-11-19 17:33:03 -0500270 ctx->ring_pages = NULL;
271 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272}
273
Gu Zheng36bc08c2013-07-16 17:56:16 +0800274static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
275{
276 vma->vm_ops = &generic_file_vm_ops;
277 return 0;
278}
279
280static const struct file_operations aio_ring_fops = {
281 .mmap = aio_ring_mmap,
282};
283
284static int aio_set_page_dirty(struct page *page)
285{
286 return 0;
287}
288
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400289#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800290static int aio_migratepage(struct address_space *mapping, struct page *new,
291 struct page *old, enum migrate_mode mode)
292{
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400293 struct kioctx *ctx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800294 unsigned long flags;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400295 pgoff_t idx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800296 int rc;
297
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500298 rc = 0;
299
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400300 /* mapping->private_lock here protects against the kioctx teardown. */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500301 spin_lock(&mapping->private_lock);
302 ctx = mapping->private_data;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400303 if (!ctx) {
304 rc = -EINVAL;
305 goto out;
306 }
307
308 /* The ring_lock mutex. The prevents aio_read_events() from writing
309 * to the ring's head, and prevents page migration from mucking in
310 * a partially initialized kiotx.
311 */
312 if (!mutex_trylock(&ctx->ring_lock)) {
313 rc = -EAGAIN;
314 goto out;
315 }
316
317 idx = old->index;
318 if (idx < (pgoff_t)ctx->nr_pages) {
319 /* Make sure the old page hasn't already been changed */
320 if (ctx->ring_pages[idx] != old)
321 rc = -EAGAIN;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500322 } else
323 rc = -EINVAL;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500324
325 if (rc != 0)
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400326 goto out_unlock;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500327
Gu Zheng36bc08c2013-07-16 17:56:16 +0800328 /* Writeback must be complete */
329 BUG_ON(PageWriteback(old));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500330 get_page(new);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800331
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500332 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800333 if (rc != MIGRATEPAGE_SUCCESS) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500334 put_page(new);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400335 goto out_unlock;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800336 }
337
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400338 /* Take completion_lock to prevent other writes to the ring buffer
339 * while the old page is copied to the new. This prevents new
340 * events from being lost.
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400341 */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400342 spin_lock_irqsave(&ctx->completion_lock, flags);
343 migrate_page_copy(new, old);
344 BUG_ON(ctx->ring_pages[idx] != old);
345 ctx->ring_pages[idx] = new;
346 spin_unlock_irqrestore(&ctx->completion_lock, flags);
347
348 /* The old page is no longer accessible. */
349 put_page(old);
350
351out_unlock:
352 mutex_unlock(&ctx->ring_lock);
353out:
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400354 spin_unlock(&mapping->private_lock);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800355 return rc;
356}
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400357#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800358
359static const struct address_space_operations aio_ctx_aops = {
360 .set_page_dirty = aio_set_page_dirty,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400361#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800362 .migratepage = aio_migratepage,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400363#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800364};
365
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366static int aio_setup_ring(struct kioctx *ctx)
367{
368 struct aio_ring *ring;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 unsigned nr_events = ctx->max_reqs;
Zach Brown41003a72013-05-07 16:18:25 -0700370 struct mm_struct *mm = current->mm;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900371 unsigned long size, unused;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 int nr_pages;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800373 int i;
374 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375
376 /* Compensate for the ring buffer's head/tail overlap entry */
377 nr_events += 2; /* 1 is required, 2 for good luck */
378
379 size = sizeof(struct aio_ring);
380 size += sizeof(struct io_event) * nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381
Gu Zheng36bc08c2013-07-16 17:56:16 +0800382 nr_pages = PFN_UP(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 if (nr_pages < 0)
384 return -EINVAL;
385
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400386 file = aio_private_file(ctx, nr_pages);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800387 if (IS_ERR(file)) {
388 ctx->aio_ring_file = NULL;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400389 return -ENOMEM;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800390 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391
Gu Zheng36bc08c2013-07-16 17:56:16 +0800392 ctx->aio_ring_file = file;
393 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
394 / sizeof(struct io_event);
395
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700396 ctx->ring_pages = ctx->internal_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 if (nr_pages > AIO_RING_PAGES) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700398 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
399 GFP_KERNEL);
Gu Zhengd1b94322013-12-04 18:19:06 +0800400 if (!ctx->ring_pages) {
401 put_aio_ring_file(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402 return -ENOMEM;
Gu Zhengd1b94322013-12-04 18:19:06 +0800403 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404 }
405
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900406 for (i = 0; i < nr_pages; i++) {
407 struct page *page;
408 page = find_or_create_page(file->f_inode->i_mapping,
409 i, GFP_HIGHUSER | __GFP_ZERO);
410 if (!page)
411 break;
412 pr_debug("pid(%d) page[%d]->count=%d\n",
413 current->pid, i, page_count(page));
414 SetPageUptodate(page);
415 SetPageDirty(page);
416 unlock_page(page);
417
418 ctx->ring_pages[i] = page;
419 }
420 ctx->nr_pages = i;
421
422 if (unlikely(i != nr_pages)) {
423 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400424 return -ENOMEM;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900425 }
426
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700427 ctx->mmap_size = nr_pages * PAGE_SIZE;
428 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800429
Zach Brown41003a72013-05-07 16:18:25 -0700430 down_write(&mm->mmap_sem);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800431 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
432 PROT_READ | PROT_WRITE,
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900433 MAP_SHARED, 0, &unused);
434 up_write(&mm->mmap_sem);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700435 if (IS_ERR((void *)ctx->mmap_base)) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700436 ctx->mmap_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400438 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 }
440
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700441 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400442
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700443 ctx->user_id = ctx->mmap_base;
444 ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700446 ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 ring->nr = nr_events; /* user copy */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400448 ring->id = ~0U;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 ring->head = ring->tail = 0;
450 ring->magic = AIO_RING_MAGIC;
451 ring->compat_features = AIO_RING_COMPAT_FEATURES;
452 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
453 ring->header_length = sizeof(struct aio_ring);
Cong Wange8e3c3d2011-11-25 23:14:27 +0800454 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700455 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456
457 return 0;
458}
459
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
461#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
462#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
463
Kent Overstreet0460fef2013-05-07 16:18:49 -0700464void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
465{
466 struct kioctx *ctx = req->ki_ctx;
467 unsigned long flags;
468
469 spin_lock_irqsave(&ctx->ctx_lock, flags);
470
471 if (!req->ki_list.next)
472 list_add(&req->ki_list, &ctx->active_reqs);
473
474 req->ki_cancel = cancel;
475
476 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
477}
478EXPORT_SYMBOL(kiocb_set_cancel_fn);
479
Fabian Frederickd52a8f92014-04-22 07:26:58 +0200480static int kiocb_cancel(struct kiocb *kiocb)
Kent Overstreet906b9732013-05-07 16:18:31 -0700481{
Kent Overstreet0460fef2013-05-07 16:18:49 -0700482 kiocb_cancel_fn *old, *cancel;
Kent Overstreet906b9732013-05-07 16:18:31 -0700483
Kent Overstreet0460fef2013-05-07 16:18:49 -0700484 /*
485 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
486 * actually has a cancel function, hence the cmpxchg()
487 */
Kent Overstreet906b9732013-05-07 16:18:31 -0700488
Kent Overstreet0460fef2013-05-07 16:18:49 -0700489 cancel = ACCESS_ONCE(kiocb->ki_cancel);
490 do {
491 if (!cancel || cancel == KIOCB_CANCELLED)
Kent Overstreet57282d82013-05-13 13:42:52 -0700492 return -EINVAL;
Kent Overstreet906b9732013-05-07 16:18:31 -0700493
Kent Overstreet0460fef2013-05-07 16:18:49 -0700494 old = cancel;
495 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
496 } while (cancel != old);
497
Kent Overstreet57282d82013-05-13 13:42:52 -0700498 return cancel(kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -0700499}
500
Kent Overstreete34ecee2013-10-10 19:31:47 -0700501static void free_ioctx(struct work_struct *work)
Kent Overstreet36f55882013-05-07 16:18:41 -0700502{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700503 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000504
Kent Overstreete34ecee2013-10-10 19:31:47 -0700505 pr_debug("freeing %p\n", ctx);
506
507 aio_free_ring(ctx);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000508 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400509 percpu_ref_exit(&ctx->reqs);
510 percpu_ref_exit(&ctx->users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700511 kmem_cache_free(kioctx_cachep, ctx);
512}
513
Kent Overstreete34ecee2013-10-10 19:31:47 -0700514static void free_ioctx_reqs(struct percpu_ref *ref)
515{
516 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
517
Anatol Pomozove02ba722014-04-15 11:31:33 -0700518 /* At this point we know that there are no any in-flight requests */
519 if (ctx->requests_done)
520 complete(ctx->requests_done);
521
Kent Overstreete34ecee2013-10-10 19:31:47 -0700522 INIT_WORK(&ctx->free_work, free_ioctx);
523 schedule_work(&ctx->free_work);
524}
525
Kent Overstreet36f55882013-05-07 16:18:41 -0700526/*
527 * When this function runs, the kioctx has been removed from the "hash table"
528 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
529 * now it's safe to cancel any that need to be.
530 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700531static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet36f55882013-05-07 16:18:41 -0700532{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700533 struct kioctx *ctx = container_of(ref, struct kioctx, users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700534 struct kiocb *req;
535
536 spin_lock_irq(&ctx->ctx_lock);
537
538 while (!list_empty(&ctx->active_reqs)) {
539 req = list_first_entry(&ctx->active_reqs,
540 struct kiocb, ki_list);
541
542 list_del_init(&req->ki_list);
Fabian Frederickd52a8f92014-04-22 07:26:58 +0200543 kiocb_cancel(req);
Kent Overstreet36f55882013-05-07 16:18:41 -0700544 }
545
546 spin_unlock_irq(&ctx->ctx_lock);
547
Kent Overstreete34ecee2013-10-10 19:31:47 -0700548 percpu_ref_kill(&ctx->reqs);
549 percpu_ref_put(&ctx->reqs);
Kent Overstreet36f55882013-05-07 16:18:41 -0700550}
551
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400552static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
553{
554 unsigned i, new_nr;
555 struct kioctx_table *table, *old;
556 struct aio_ring *ring;
557
558 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200559 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400560
561 while (1) {
562 if (table)
563 for (i = 0; i < table->nr; i++)
564 if (!table->table[i]) {
565 ctx->id = i;
566 table->table[i] = ctx;
567 spin_unlock(&mm->ioctx_lock);
568
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400569 /* While kioctx setup is in progress,
570 * we are protected from page migration
571 * changes ring_pages by ->ring_lock.
572 */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400573 ring = kmap_atomic(ctx->ring_pages[0]);
574 ring->id = ctx->id;
575 kunmap_atomic(ring);
576 return 0;
577 }
578
579 new_nr = (table ? table->nr : 1) * 4;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400580 spin_unlock(&mm->ioctx_lock);
581
582 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
583 new_nr, GFP_KERNEL);
584 if (!table)
585 return -ENOMEM;
586
587 table->nr = new_nr;
588
589 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200590 old = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400591
592 if (!old) {
593 rcu_assign_pointer(mm->ioctx_table, table);
594 } else if (table->nr > old->nr) {
595 memcpy(table->table, old->table,
596 old->nr * sizeof(struct kioctx *));
597
598 rcu_assign_pointer(mm->ioctx_table, table);
599 kfree_rcu(old, rcu);
600 } else {
601 kfree(table);
602 table = old;
603 }
604 }
605}
606
Kent Overstreete34ecee2013-10-10 19:31:47 -0700607static void aio_nr_sub(unsigned nr)
608{
609 spin_lock(&aio_nr_lock);
610 if (WARN_ON(aio_nr - nr > aio_nr))
611 aio_nr = 0;
612 else
613 aio_nr -= nr;
614 spin_unlock(&aio_nr_lock);
615}
616
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617/* ioctx_alloc
618 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
619 */
620static struct kioctx *ioctx_alloc(unsigned nr_events)
621{
Zach Brown41003a72013-05-07 16:18:25 -0700622 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623 struct kioctx *ctx;
Al Viroe23754f2012-03-06 14:33:22 -0500624 int err = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000626 /*
627 * We keep track of the number of available ringbuffer slots, to prevent
628 * overflow (reqs_available), and we also use percpu counters for this.
629 *
630 * So since up to half the slots might be on other cpu's percpu counters
631 * and unavailable, double nr_events so userspace sees what they
632 * expected: additionally, we move req_batch slots to/from percpu
633 * counters at a time, so make sure that isn't 0:
634 */
635 nr_events = max(nr_events, num_possible_cpus() * 4);
636 nr_events *= 2;
637
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 /* Prevent overflows */
639 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
640 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
641 pr_debug("ENOMEM: nr_events too high\n");
642 return ERR_PTR(-EINVAL);
643 }
644
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400645 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646 return ERR_PTR(-EAGAIN);
647
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800648 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 if (!ctx)
650 return ERR_PTR(-ENOMEM);
651
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 ctx->max_reqs = nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400654 spin_lock_init(&ctx->ctx_lock);
655 spin_lock_init(&ctx->completion_lock);
656 mutex_init(&ctx->ring_lock);
657 /* Protect against page migration throughout kiotx setup by keeping
658 * the ring_lock mutex held until setup is complete. */
659 mutex_lock(&ctx->ring_lock);
660 init_waitqueue_head(&ctx->wait);
661
662 INIT_LIST_HEAD(&ctx->active_reqs);
663
Tejun Heo2aad2a82014-09-24 13:31:50 -0400664 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700665 goto err;
666
Tejun Heo2aad2a82014-09-24 13:31:50 -0400667 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700668 goto err;
Kent Overstreet723be6e2013-05-28 15:14:48 -0700669
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000670 ctx->cpu = alloc_percpu(struct kioctx_cpu);
671 if (!ctx->cpu)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700672 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400674 err = aio_setup_ring(ctx);
675 if (err < 0)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700676 goto err;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000677
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000678 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000679 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
Benjamin LaHaise6878ea72013-07-31 10:34:18 -0400680 if (ctx->req_batch < 1)
681 ctx->req_batch = 1;
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000682
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683 /* limit the number of system wide aios */
Al Viro9fa1cb32012-03-10 23:14:05 -0500684 spin_lock(&aio_nr_lock);
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400685 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
Al Viro2dd542b2012-03-10 23:10:35 -0500686 aio_nr + nr_events < aio_nr) {
Al Viro9fa1cb32012-03-10 23:14:05 -0500687 spin_unlock(&aio_nr_lock);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700688 err = -EAGAIN;
Gu Zhengd1b94322013-12-04 18:19:06 +0800689 goto err_ctx;
Al Viro2dd542b2012-03-10 23:10:35 -0500690 }
691 aio_nr += ctx->max_reqs;
Al Viro9fa1cb32012-03-10 23:14:05 -0500692 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693
Benjamin LaHaise18816862013-12-21 15:49:28 -0500694 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
695 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
Kent Overstreet723be6e2013-05-28 15:14:48 -0700696
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400697 err = ioctx_add_table(ctx, mm);
698 if (err)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700699 goto err_cleanup;
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400700
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400701 /* Release the ring_lock mutex now that all setup is complete. */
702 mutex_unlock(&ctx->ring_lock);
703
Kent Overstreetcaf41672013-05-07 16:18:35 -0700704 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700705 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 return ctx;
707
Kent Overstreete34ecee2013-10-10 19:31:47 -0700708err_cleanup:
709 aio_nr_sub(ctx->max_reqs);
Gu Zhengd1b94322013-12-04 18:19:06 +0800710err_ctx:
711 aio_free_ring(ctx);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700712err:
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400713 mutex_unlock(&ctx->ring_lock);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000714 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400715 percpu_ref_exit(&ctx->reqs);
716 percpu_ref_exit(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreetcaf41672013-05-07 16:18:35 -0700718 pr_debug("error allocating ioctx %d\n", err);
Al Viroe23754f2012-03-06 14:33:22 -0500719 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720}
721
Kent Overstreet36f55882013-05-07 16:18:41 -0700722/* kill_ioctx
723 * Cancels all outstanding aio requests on an aio context. Used
724 * when the processes owning a context have all exited to encourage
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 * the rapid destruction of the kioctx.
726 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400727static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
Anatol Pomozove02ba722014-04-15 11:31:33 -0700728 struct completion *requests_done)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729{
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400730 struct kioctx_table *table;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400731
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400732 if (atomic_xchg(&ctx->dead, 1))
733 return -EINVAL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400734
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400735
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400736 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200737 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400738 WARN_ON(ctx != table->table[ctx->id]);
739 table->table[ctx->id] = NULL;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400740 spin_unlock(&mm->ioctx_lock);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700741
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400742 /* percpu_ref_kill() will do the necessary call_rcu() */
743 wake_up_all(&ctx->wait);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700744
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400745 /*
746 * It'd be more correct to do this in free_ioctx(), after all
747 * the outstanding kiocbs have finished - but by then io_destroy
748 * has already returned, so io_setup() could potentially return
749 * -EAGAIN with no ioctxs actually in use (as far as userspace
750 * could tell).
751 */
752 aio_nr_sub(ctx->max_reqs);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400753
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400754 if (ctx->mmap_size)
755 vm_munmap(ctx->mmap_base, ctx->mmap_size);
756
757 ctx->requests_done = requests_done;
758 percpu_ref_kill(&ctx->users);
759 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760}
761
762/* wait_on_sync_kiocb:
763 * Waits on the given sync kiocb to complete.
764 */
Kent Overstreet57282d82013-05-13 13:42:52 -0700765ssize_t wait_on_sync_kiocb(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766{
Kent Overstreet57282d82013-05-13 13:42:52 -0700767 while (!req->ki_ctx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 set_current_state(TASK_UNINTERRUPTIBLE);
Kent Overstreet57282d82013-05-13 13:42:52 -0700769 if (req->ki_ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 break;
Jeff Moyer41d10da2007-10-16 23:27:20 -0700771 io_schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 }
773 __set_current_state(TASK_RUNNING);
Kent Overstreet57282d82013-05-13 13:42:52 -0700774 return req->ki_user_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700776EXPORT_SYMBOL(wait_on_sync_kiocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777
Kent Overstreet36f55882013-05-07 16:18:41 -0700778/*
779 * exit_aio: called when the last user of mm goes away. At this point, there is
780 * no way for any new requests to be submited or any of the io_* syscalls to be
781 * called on the context.
782 *
783 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
784 * them.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800786void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787{
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200788 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
789 int i;
Jens Axboeabf137d2008-12-09 08:11:22 +0100790
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200791 if (!table)
792 return;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400793
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200794 for (i = 0; i < table->nr; ++i) {
795 struct kioctx *ctx = table->table[i];
Gu Zheng6098b452014-09-03 17:45:44 +0800796 struct completion requests_done =
797 COMPLETION_INITIALIZER_ONSTACK(requests_done);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400798
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200799 if (!ctx)
800 continue;
Al Viro936af152012-04-20 21:49:41 -0400801 /*
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200802 * We don't need to bother with munmap() here - exit_mmap(mm)
803 * is coming and it'll unmap everything. And we simply can't,
804 * this is not necessarily our ->mm.
805 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
806 * that it needs to unmap the area, just set it to 0.
Al Viro936af152012-04-20 21:49:41 -0400807 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700808 ctx->mmap_size = 0;
Gu Zheng6098b452014-09-03 17:45:44 +0800809 kill_ioctx(mm, ctx, &requests_done);
Kent Overstreet36f55882013-05-07 16:18:41 -0700810
Gu Zheng6098b452014-09-03 17:45:44 +0800811 /* Wait until all IO for the context are done. */
812 wait_for_completion(&requests_done);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 }
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200814
815 RCU_INIT_POINTER(mm->ioctx_table, NULL);
816 kfree(table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817}
818
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000819static void put_reqs_available(struct kioctx *ctx, unsigned nr)
820{
821 struct kioctx_cpu *kcpu;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400822 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000823
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400824 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400825 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000826 kcpu->reqs_available += nr;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400827
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000828 while (kcpu->reqs_available >= ctx->req_batch * 2) {
829 kcpu->reqs_available -= ctx->req_batch;
830 atomic_add(ctx->req_batch, &ctx->reqs_available);
831 }
832
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400833 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000834}
835
836static bool get_reqs_available(struct kioctx *ctx)
837{
838 struct kioctx_cpu *kcpu;
839 bool ret = false;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400840 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000841
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400842 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400843 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000844 if (!kcpu->reqs_available) {
845 int old, avail = atomic_read(&ctx->reqs_available);
846
847 do {
848 if (avail < ctx->req_batch)
849 goto out;
850
851 old = avail;
852 avail = atomic_cmpxchg(&ctx->reqs_available,
853 avail, avail - ctx->req_batch);
854 } while (avail != old);
855
856 kcpu->reqs_available += ctx->req_batch;
857 }
858
859 ret = true;
860 kcpu->reqs_available--;
861out:
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400862 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000863 return ret;
864}
865
Benjamin LaHaised856f322014-08-24 13:14:05 -0400866/* refill_reqs_available
867 * Updates the reqs_available reference counts used for tracking the
868 * number of free slots in the completion ring. This can be called
869 * from aio_complete() (to optimistically update reqs_available) or
870 * from aio_get_req() (the we're out of events case). It must be
871 * called holding ctx->completion_lock.
872 */
873static void refill_reqs_available(struct kioctx *ctx, unsigned head,
874 unsigned tail)
875{
876 unsigned events_in_ring, completed;
877
878 /* Clamp head since userland can write to it. */
879 head %= ctx->nr_events;
880 if (head <= tail)
881 events_in_ring = tail - head;
882 else
883 events_in_ring = ctx->nr_events - (head - tail);
884
885 completed = ctx->completed_events;
886 if (events_in_ring < completed)
887 completed -= events_in_ring;
888 else
889 completed = 0;
890
891 if (!completed)
892 return;
893
894 ctx->completed_events -= completed;
895 put_reqs_available(ctx, completed);
896}
897
898/* user_refill_reqs_available
899 * Called to refill reqs_available when aio_get_req() encounters an
900 * out of space in the completion ring.
901 */
902static void user_refill_reqs_available(struct kioctx *ctx)
903{
904 spin_lock_irq(&ctx->completion_lock);
905 if (ctx->completed_events) {
906 struct aio_ring *ring;
907 unsigned head;
908
909 /* Access of ring->head may race with aio_read_events_ring()
910 * here, but that's okay since whether we read the old version
911 * or the new version, and either will be valid. The important
912 * part is that head cannot pass tail since we prevent
913 * aio_complete() from updating tail by holding
914 * ctx->completion_lock. Even if head is invalid, the check
915 * against ctx->completed_events below will make sure we do the
916 * safe/right thing.
917 */
918 ring = kmap_atomic(ctx->ring_pages[0]);
919 head = ring->head;
920 kunmap_atomic(ring);
921
922 refill_reqs_available(ctx, head, ctx->tail);
923 }
924
925 spin_unlock_irq(&ctx->completion_lock);
926}
927
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928/* aio_get_req
Kent Overstreet57282d82013-05-13 13:42:52 -0700929 * Allocate a slot for an aio request.
930 * Returns NULL if no requests are free.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 */
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700932static inline struct kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933{
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700934 struct kiocb *req;
935
Benjamin LaHaised856f322014-08-24 13:14:05 -0400936 if (!get_reqs_available(ctx)) {
937 user_refill_reqs_available(ctx);
938 if (!get_reqs_available(ctx))
939 return NULL;
940 }
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700941
Kent Overstreet0460fef2013-05-07 16:18:49 -0700942 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 if (unlikely(!req))
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700944 goto out_put;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945
Kent Overstreete34ecee2013-10-10 19:31:47 -0700946 percpu_ref_get(&ctx->reqs);
947
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 req->ki_ctx = ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 return req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700950out_put:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000951 put_reqs_available(ctx, 1);
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700952 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953}
954
Kent Overstreet11599eb2013-05-07 16:18:39 -0700955static void kiocb_free(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956{
Kent Overstreet1d98ebf2013-05-07 16:18:37 -0700957 if (req->ki_filp)
958 fput(req->ki_filp);
Davide Libenzi13389012009-06-30 11:41:11 -0700959 if (req->ki_eventfd != NULL)
960 eventfd_ctx_put(req->ki_eventfd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 kmem_cache_free(kiocb_cachep, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962}
963
Adrian Bunkd5470b52008-04-29 00:58:57 -0700964static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400966 struct aio_ring __user *ring = (void __user *)ctx_id;
Jens Axboeabf137d2008-12-09 08:11:22 +0100967 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -0700968 struct kioctx *ctx, *ret = NULL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400969 struct kioctx_table *table;
970 unsigned id;
971
972 if (get_user(id, &ring->id))
973 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974
Jens Axboeabf137d2008-12-09 08:11:22 +0100975 rcu_read_lock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400976 table = rcu_dereference(mm->ioctx_table);
Jens Axboeabf137d2008-12-09 08:11:22 +0100977
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400978 if (!table || id >= table->nr)
979 goto out;
980
981 ctx = table->table[id];
Benjamin LaHaisef30d7042013-08-07 18:23:48 -0400982 if (ctx && ctx->user_id == ctx_id) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400983 percpu_ref_get(&ctx->users);
984 ret = ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +0100985 }
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400986out:
Jens Axboeabf137d2008-12-09 08:11:22 +0100987 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -0700988 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989}
990
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991/* aio_complete
992 * Called when the io request on the given iocb is complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 */
Kent Overstreet2d684492013-05-07 16:18:29 -0700994void aio_complete(struct kiocb *iocb, long res, long res2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995{
996 struct kioctx *ctx = iocb->ki_ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 struct aio_ring *ring;
Kent Overstreet21b40202013-05-07 16:18:47 -0700998 struct io_event *ev_page, *event;
Benjamin LaHaised856f322014-08-24 13:14:05 -0400999 unsigned tail, pos, head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001
Zach Brown20dcae32005-11-13 16:07:33 -08001002 /*
1003 * Special case handling for sync iocbs:
1004 * - events go directly into the iocb for fast handling
1005 * - the sync task with the iocb in its stack holds the single iocb
1006 * ref, no other paths have a way to get another ref
1007 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 */
1009 if (is_sync_kiocb(iocb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010 iocb->ki_user_data = res;
Kent Overstreet57282d82013-05-13 13:42:52 -07001011 smp_wmb();
1012 iocb->ki_ctx = ERR_PTR(-EXDEV);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 wake_up_process(iocb->ki_obj.tsk);
Kent Overstreet2d684492013-05-07 16:18:29 -07001014 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015 }
1016
Kent Overstreet0460fef2013-05-07 16:18:49 -07001017 if (iocb->ki_list.next) {
1018 unsigned long flags;
1019
1020 spin_lock_irqsave(&ctx->ctx_lock, flags);
1021 list_del(&iocb->ki_list);
1022 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1023 }
Kent Overstreet11599eb2013-05-07 16:18:39 -07001024
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025 /*
Kent Overstreet0460fef2013-05-07 16:18:49 -07001026 * Add a completion event to the ring buffer. Must be done holding
Tang Chen4b30f072013-07-03 15:09:16 -07001027 * ctx->completion_lock to prevent other code from messing with the tail
Kent Overstreet0460fef2013-05-07 16:18:49 -07001028 * pointer since we might be called from irq context.
1029 */
1030 spin_lock_irqsave(&ctx->completion_lock, flags);
1031
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001032 tail = ctx->tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001033 pos = tail + AIO_EVENTS_OFFSET;
1034
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001035 if (++tail >= ctx->nr_events)
Ken Chen4bf69b22005-05-01 08:59:15 -07001036 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001038 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001039 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1040
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1042 event->data = iocb->ki_user_data;
1043 event->res = res;
1044 event->res2 = res2;
1045
Kent Overstreet21b40202013-05-07 16:18:47 -07001046 kunmap_atomic(ev_page);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001047 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001048
1049 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
Kent Overstreetcaf41672013-05-07 16:18:35 -07001050 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1051 res, res2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052
1053 /* after flagging the request as done, we
1054 * must never even look at it again
1055 */
1056 smp_wmb(); /* make event visible before updating tail */
1057
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001058 ctx->tail = tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001059
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001060 ring = kmap_atomic(ctx->ring_pages[0]);
Benjamin LaHaised856f322014-08-24 13:14:05 -04001061 head = ring->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 ring->tail = tail;
Cong Wange8e3c3d2011-11-25 23:14:27 +08001063 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001064 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065
Benjamin LaHaised856f322014-08-24 13:14:05 -04001066 ctx->completed_events++;
1067 if (ctx->completed_events > 1)
1068 refill_reqs_available(ctx, head, tail);
Kent Overstreet0460fef2013-05-07 16:18:49 -07001069 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1070
Kent Overstreet21b40202013-05-07 16:18:47 -07001071 pr_debug("added to ring %p at [%u]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001072
1073 /*
1074 * Check if the user asked us to deliver the result through an
1075 * eventfd. The eventfd_signal() function is safe to be called
1076 * from IRQ context.
1077 */
Davide Libenzi87c3a862009-03-18 17:04:19 -07001078 if (iocb->ki_eventfd != NULL)
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001079 eventfd_signal(iocb->ki_eventfd, 1);
1080
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081 /* everything turned out well, dispose of the aiocb. */
Kent Overstreet57282d82013-05-13 13:42:52 -07001082 kiocb_free(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083
Quentin Barnes6cb2a212008-03-19 17:00:39 -07001084 /*
1085 * We have to order our ring_info tail store above and test
1086 * of the wait list below outside the wait lock. This is
1087 * like in wake_up_bit() where clearing a bit has to be
1088 * ordered with the unlocked test.
1089 */
1090 smp_mb();
1091
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 if (waitqueue_active(&ctx->wait))
1093 wake_up(&ctx->wait);
1094
Kent Overstreete34ecee2013-10-10 19:31:47 -07001095 percpu_ref_put(&ctx->reqs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096}
H Hartley Sweeten385773e2009-09-22 16:43:53 -07001097EXPORT_SYMBOL(aio_complete);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001099/* aio_read_events_ring
Kent Overstreeta31ad382013-05-07 16:18:45 -07001100 * Pull an event off of the ioctx's event ring. Returns the number of
1101 * events fetched
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 */
Kent Overstreeta31ad382013-05-07 16:18:45 -07001103static long aio_read_events_ring(struct kioctx *ctx,
1104 struct io_event __user *event, long nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106 struct aio_ring *ring;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001107 unsigned head, tail, pos;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001108 long ret = 0;
1109 int copy_ret;
1110
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001111 mutex_lock(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -04001113 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001114 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001115 head = ring->head;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001116 tail = ring->tail;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001117 kunmap_atomic(ring);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118
Jeff Moyer2ff396be2014-09-02 13:17:00 -04001119 /*
1120 * Ensure that once we've read the current tail pointer, that
1121 * we also see the events that were stored up to the tail.
1122 */
1123 smp_rmb();
1124
Kent Overstreet5ffac122013-05-09 15:36:07 -07001125 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001126
Kent Overstreet5ffac122013-05-09 15:36:07 -07001127 if (head == tail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 goto out;
1129
Benjamin LaHaiseedfbbf32014-06-24 13:32:51 -04001130 head %= ctx->nr_events;
1131 tail %= ctx->nr_events;
1132
Kent Overstreeta31ad382013-05-07 16:18:45 -07001133 while (ret < nr) {
1134 long avail;
1135 struct io_event *ev;
1136 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137
Kent Overstreet5ffac122013-05-09 15:36:07 -07001138 avail = (head <= tail ? tail : ctx->nr_events) - head;
1139 if (head == tail)
Kent Overstreeta31ad382013-05-07 16:18:45 -07001140 break;
1141
1142 avail = min(avail, nr - ret);
1143 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1144 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1145
1146 pos = head + AIO_EVENTS_OFFSET;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001147 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
Kent Overstreeta31ad382013-05-07 16:18:45 -07001148 pos %= AIO_EVENTS_PER_PAGE;
1149
1150 ev = kmap(page);
1151 copy_ret = copy_to_user(event + ret, ev + pos,
1152 sizeof(*ev) * avail);
1153 kunmap(page);
1154
1155 if (unlikely(copy_ret)) {
1156 ret = -EFAULT;
1157 goto out;
1158 }
1159
1160 ret += avail;
1161 head += avail;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001162 head %= ctx->nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001165 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001166 ring->head = head;
Zhao Hongjiang91d80a82013-04-26 11:03:53 +08001167 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001168 flush_dcache_page(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001169
Kent Overstreet5ffac122013-05-09 15:36:07 -07001170 pr_debug("%li h%u t%u\n", ret, head, tail);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001171out:
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001172 mutex_unlock(&ctx->ring_lock);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001173
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 return ret;
1175}
1176
Kent Overstreeta31ad382013-05-07 16:18:45 -07001177static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1178 struct io_event __user *event, long *i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001180 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181
Kent Overstreeta31ad382013-05-07 16:18:45 -07001182 if (ret > 0)
1183 *i += ret;
1184
1185 if (unlikely(atomic_read(&ctx->dead)))
1186 ret = -EINVAL;
1187
1188 if (!*i)
1189 *i = ret;
1190
1191 return ret < 0 || *i >= min_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192}
1193
Kent Overstreeta31ad382013-05-07 16:18:45 -07001194static long read_events(struct kioctx *ctx, long min_nr, long nr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 struct io_event __user *event,
1196 struct timespec __user *timeout)
1197{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001198 ktime_t until = { .tv64 = KTIME_MAX };
1199 long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201 if (timeout) {
1202 struct timespec ts;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001203
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
Kent Overstreeta31ad382013-05-07 16:18:45 -07001205 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206
Kent Overstreeta31ad382013-05-07 16:18:45 -07001207 until = timespec_to_ktime(ts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 }
1209
Kent Overstreeta31ad382013-05-07 16:18:45 -07001210 /*
1211 * Note that aio_read_events() is being called as the conditional - i.e.
1212 * we're calling it after prepare_to_wait() has set task state to
1213 * TASK_INTERRUPTIBLE.
1214 *
1215 * But aio_read_events() can block, and if it blocks it's going to flip
1216 * the task state back to TASK_RUNNING.
1217 *
1218 * This should be ok, provided it doesn't flip the state back to
1219 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1220 * will only happen if the mutex_lock() call blocks, and we then find
1221 * the ringbuffer empty. So in practice we should be ok, but it's
1222 * something to be aware of when touching this code.
1223 */
1224 wait_event_interruptible_hrtimeout(ctx->wait,
1225 aio_read_events(ctx, min_nr, nr, event, &ret), until);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226
Kent Overstreeta31ad382013-05-07 16:18:45 -07001227 if (!ret && signal_pending(current))
1228 ret = -EINTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229
Kent Overstreeta31ad382013-05-07 16:18:45 -07001230 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231}
1232
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233/* sys_io_setup:
1234 * Create an aio_context capable of receiving at least nr_events.
1235 * ctxp must not point to an aio_context that already exists, and
1236 * must be initialized to 0 prior to the call. On successful
1237 * creation of the aio_context, *ctxp is filled in with the resulting
1238 * handle. May fail with -EINVAL if *ctxp is not initialized,
1239 * if the specified nr_events exceeds internal limits. May fail
1240 * with -EAGAIN if the specified nr_events exceeds the user's limit
1241 * of available events. May fail with -ENOMEM if insufficient kernel
1242 * resources are available. May fail with -EFAULT if an invalid
1243 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1244 * implemented.
1245 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001246SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247{
1248 struct kioctx *ioctx = NULL;
1249 unsigned long ctx;
1250 long ret;
1251
1252 ret = get_user(ctx, ctxp);
1253 if (unlikely(ret))
1254 goto out;
1255
1256 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001257 if (unlikely(ctx || nr_events == 0)) {
1258 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1259 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260 goto out;
1261 }
1262
1263 ioctx = ioctx_alloc(nr_events);
1264 ret = PTR_ERR(ioctx);
1265 if (!IS_ERR(ioctx)) {
1266 ret = put_user(ioctx->user_id, ctxp);
Al Viroa2e18592012-03-20 16:27:57 -04001267 if (ret)
Anatol Pomozove02ba722014-04-15 11:31:33 -07001268 kill_ioctx(current->mm, ioctx, NULL);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001269 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 }
1271
1272out:
1273 return ret;
1274}
1275
1276/* sys_io_destroy:
1277 * Destroy the aio_context specified. May cancel any outstanding
1278 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001279 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 * is invalid.
1281 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001282SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283{
1284 struct kioctx *ioctx = lookup_ioctx(ctx);
1285 if (likely(NULL != ioctx)) {
Anatol Pomozove02ba722014-04-15 11:31:33 -07001286 struct completion requests_done =
1287 COMPLETION_INITIALIZER_ONSTACK(requests_done);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001288 int ret;
Anatol Pomozove02ba722014-04-15 11:31:33 -07001289
1290 /* Pass requests_done to kill_ioctx() where it can be set
1291 * in a thread-safe way. If we try to set it here then we have
1292 * a race condition if two io_destroy() called simultaneously.
1293 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001294 ret = kill_ioctx(current->mm, ioctx, &requests_done);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001295 percpu_ref_put(&ioctx->users);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001296
1297 /* Wait until all IO for the context are done. Otherwise kernel
1298 * keep using user-space buffers even if user thinks the context
1299 * is destroyed.
1300 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001301 if (!ret)
1302 wait_for_completion(&requests_done);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001303
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001304 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 }
1306 pr_debug("EINVAL: io_destroy: invalid context id\n");
1307 return -EINVAL;
1308}
1309
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001310typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1311 unsigned long, loff_t);
Al Viro293bc982014-02-11 18:37:41 -05001312typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001313
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001314static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1315 int rw, char __user *buf,
1316 unsigned long *nr_segs,
1317 struct iovec **iovec,
1318 bool compat)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001319{
1320 ssize_t ret;
1321
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001322 *nr_segs = kiocb->ki_nbytes;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001323
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001324#ifdef CONFIG_COMPAT
1325 if (compat)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001326 ret = compat_rw_copy_check_uvector(rw,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001327 (struct compat_iovec __user *)buf,
Gu Zheng00fefb92014-07-23 18:03:54 +08001328 *nr_segs, UIO_FASTIOV, *iovec, iovec);
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001329 else
1330#endif
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001331 ret = rw_copy_check_uvector(rw,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001332 (struct iovec __user *)buf,
Gu Zheng00fefb92014-07-23 18:03:54 +08001333 *nr_segs, UIO_FASTIOV, *iovec, iovec);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001334 if (ret < 0)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001335 return ret;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001336
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001337 /* ki_nbytes now reflect bytes instead of segs */
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001338 kiocb->ki_nbytes = ret;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001339 return 0;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001340}
1341
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001342static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1343 int rw, char __user *buf,
1344 unsigned long *nr_segs,
1345 struct iovec *iovec)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001346{
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001347 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001348 return -EFAULT;
Linus Torvaldsa70b52e2012-05-21 16:06:20 -07001349
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001350 iovec->iov_base = buf;
1351 iovec->iov_len = kiocb->ki_nbytes;
1352 *nr_segs = 1;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001353 return 0;
1354}
1355
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356/*
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001357 * aio_run_iocb:
1358 * Performs the initial checks and io submission.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 */
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001360static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1361 char __user *buf, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362{
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001363 struct file *file = req->ki_filp;
1364 ssize_t ret;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001365 unsigned long nr_segs;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001366 int rw;
1367 fmode_t mode;
1368 aio_rw_op *rw_op;
Al Viro293bc982014-02-11 18:37:41 -05001369 rw_iter_op *iter_op;
Gu Zheng00fefb92014-07-23 18:03:54 +08001370 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
Al Viro293bc982014-02-11 18:37:41 -05001371 struct iov_iter iter;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001373 switch (opcode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 case IOCB_CMD_PREAD:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001375 case IOCB_CMD_PREADV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001376 mode = FMODE_READ;
1377 rw = READ;
1378 rw_op = file->f_op->aio_read;
Al Viro293bc982014-02-11 18:37:41 -05001379 iter_op = file->f_op->read_iter;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001380 goto rw_common;
1381
1382 case IOCB_CMD_PWRITE:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001383 case IOCB_CMD_PWRITEV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001384 mode = FMODE_WRITE;
1385 rw = WRITE;
1386 rw_op = file->f_op->aio_write;
Al Viro293bc982014-02-11 18:37:41 -05001387 iter_op = file->f_op->write_iter;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001388 goto rw_common;
1389rw_common:
1390 if (unlikely(!(file->f_mode & mode)))
1391 return -EBADF;
1392
Al Viro293bc982014-02-11 18:37:41 -05001393 if (!rw_op && !iter_op)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001394 return -EINVAL;
1395
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001396 ret = (opcode == IOCB_CMD_PREADV ||
1397 opcode == IOCB_CMD_PWRITEV)
1398 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1399 &iovec, compat)
1400 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1401 iovec);
Leon Yu754320d2014-05-01 03:31:28 +00001402 if (!ret)
1403 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001404 if (ret < 0) {
Gu Zheng00fefb92014-07-23 18:03:54 +08001405 if (iovec != inline_vecs)
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001406 kfree(iovec);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001407 return ret;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001408 }
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001409
1410 req->ki_nbytes = ret;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001411
Kent Overstreet73a70752013-05-09 15:03:42 -07001412 /* XXX: move/kill - rw_verify_area()? */
1413 /* This matches the pread()/pwrite() logic */
1414 if (req->ki_pos < 0) {
1415 ret = -EINVAL;
1416 break;
1417 }
1418
1419 if (rw == WRITE)
1420 file_start_write(file);
1421
Al Viro293bc982014-02-11 18:37:41 -05001422 if (iter_op) {
1423 iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes);
1424 ret = iter_op(req, &iter);
1425 } else {
1426 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1427 }
Kent Overstreet73a70752013-05-09 15:03:42 -07001428
1429 if (rw == WRITE)
1430 file_end_write(file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001432
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 case IOCB_CMD_FDSYNC:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001434 if (!file->f_op->aio_fsync)
1435 return -EINVAL;
1436
1437 ret = file->f_op->aio_fsync(req, 1);
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_FSYNC:
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, 0);
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 default:
Kent Overstreetcaf41672013-05-07 16:18:35 -07001448 pr_debug("EINVAL: no operation provided\n");
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001449 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 }
1451
Gu Zheng00fefb92014-07-23 18:03:54 +08001452 if (iovec != inline_vecs)
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001453 kfree(iovec);
1454
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001455 if (ret != -EIOCBQUEUED) {
1456 /*
1457 * There's no easy way to restart the syscall since other AIO's
1458 * may be already running. Just fail this IO with EINTR.
1459 */
1460 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1461 ret == -ERESTARTNOHAND ||
1462 ret == -ERESTART_RESTARTBLOCK))
1463 ret = -EINTR;
1464 aio_complete(req, ret, 0);
1465 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466
1467 return 0;
1468}
1469
Adrian Bunkd5470b52008-04-29 00:58:57 -07001470static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001471 struct iocb *iocb, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472{
1473 struct kiocb *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474 ssize_t ret;
1475
1476 /* enforce forwards compatibility on users */
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001477 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001478 pr_debug("EINVAL: reserve field set\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 return -EINVAL;
1480 }
1481
1482 /* prevent overflows */
1483 if (unlikely(
1484 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1485 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1486 ((ssize_t)iocb->aio_nbytes < 0)
1487 )) {
1488 pr_debug("EINVAL: io_submit: overflow check\n");
1489 return -EINVAL;
1490 }
1491
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001492 req = aio_get_req(ctx);
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001493 if (unlikely(!req))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 return -EAGAIN;
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001495
1496 req->ki_filp = fget(iocb->aio_fildes);
1497 if (unlikely(!req->ki_filp)) {
1498 ret = -EBADF;
1499 goto out_put_req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 }
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001501
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001502 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1503 /*
1504 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1505 * instance of the file* now. The file descriptor must be
1506 * an eventfd() fd, and will be signaled for each completed
1507 * event using the eventfd_signal() function.
1508 */
Davide Libenzi13389012009-06-30 11:41:11 -07001509 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
Hirofumi Nakagawa801678c2008-04-29 01:03:09 -07001510 if (IS_ERR(req->ki_eventfd)) {
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001511 ret = PTR_ERR(req->ki_eventfd);
Davide Libenzi87c3a862009-03-18 17:04:19 -07001512 req->ki_eventfd = NULL;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001513 goto out_put_req;
1514 }
1515 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516
Kent Overstreet8a660892013-05-07 16:19:10 -07001517 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 if (unlikely(ret)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001519 pr_debug("EFAULT: aio_key\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520 goto out_put_req;
1521 }
1522
1523 req->ki_obj.user = user_iocb;
1524 req->ki_user_data = iocb->aio_data;
1525 req->ki_pos = iocb->aio_offset;
Kent Overstreet73a70752013-05-09 15:03:42 -07001526 req->ki_nbytes = iocb->aio_nbytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001528 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1529 (char __user *)(unsigned long)iocb->aio_buf,
1530 compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 if (ret)
1532 goto out_put_req;
1533
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535out_put_req:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001536 put_reqs_available(ctx, 1);
Kent Overstreete34ecee2013-10-10 19:31:47 -07001537 percpu_ref_put(&ctx->reqs);
Kent Overstreet57282d82013-05-13 13:42:52 -07001538 kiocb_free(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 return ret;
1540}
1541
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001542long do_io_submit(aio_context_t ctx_id, long nr,
1543 struct iocb __user *__user *iocbpp, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544{
1545 struct kioctx *ctx;
1546 long ret = 0;
Jeff Moyer080d6762011-11-02 13:40:10 -07001547 int i = 0;
Shaohua Li9f5b9422010-07-01 07:55:01 +02001548 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549
1550 if (unlikely(nr < 0))
1551 return -EINVAL;
1552
Jeff Moyer75e1c702010-09-10 14:16:00 -07001553 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1554 nr = LONG_MAX/sizeof(*iocbpp);
1555
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1557 return -EFAULT;
1558
1559 ctx = lookup_ioctx(ctx_id);
1560 if (unlikely(!ctx)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001561 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 return -EINVAL;
1563 }
1564
Shaohua Li9f5b9422010-07-01 07:55:01 +02001565 blk_start_plug(&plug);
1566
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567 /*
1568 * AKPM: should this return a partial result if some of the IOs were
1569 * successfully submitted?
1570 */
1571 for (i=0; i<nr; i++) {
1572 struct iocb __user *user_iocb;
1573 struct iocb tmp;
1574
1575 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1576 ret = -EFAULT;
1577 break;
1578 }
1579
1580 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1581 ret = -EFAULT;
1582 break;
1583 }
1584
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001585 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586 if (ret)
1587 break;
1588 }
Shaohua Li9f5b9422010-07-01 07:55:01 +02001589 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590
Kent Overstreet723be6e2013-05-28 15:14:48 -07001591 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592 return i ? i : ret;
1593}
1594
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001595/* sys_io_submit:
1596 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1597 * the number of iocbs queued. May return -EINVAL if the aio_context
1598 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1599 * *iocbpp[0] is not properly initialized, if the operation specified
1600 * is invalid for the file descriptor in the iocb. May fail with
1601 * -EFAULT if any of the data structures point to invalid data. May
1602 * fail with -EBADF if the file descriptor specified in the first
1603 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1604 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1605 * fail with -ENOSYS if not implemented.
1606 */
1607SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1608 struct iocb __user * __user *, iocbpp)
1609{
1610 return do_io_submit(ctx_id, nr, iocbpp, 0);
1611}
1612
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613/* lookup_kiocb
1614 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001616static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1617 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618{
1619 struct list_head *pos;
Zach Brownd00689a2005-11-13 16:07:34 -08001620
1621 assert_spin_locked(&ctx->ctx_lock);
1622
Kent Overstreet8a660892013-05-07 16:19:10 -07001623 if (key != KIOCB_KEY)
1624 return NULL;
1625
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 /* TODO: use a hash or array, this sucks. */
1627 list_for_each(pos, &ctx->active_reqs) {
1628 struct kiocb *kiocb = list_kiocb(pos);
Kent Overstreet8a660892013-05-07 16:19:10 -07001629 if (kiocb->ki_obj.user == iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 return kiocb;
1631 }
1632 return NULL;
1633}
1634
1635/* sys_io_cancel:
1636 * Attempts to cancel an iocb previously passed to io_submit. If
1637 * the operation is successfully cancelled, the resulting event is
1638 * copied into the memory pointed to by result without being placed
1639 * into the completion queue and 0 is returned. May fail with
1640 * -EFAULT if any of the data structures pointed to are invalid.
1641 * May fail with -EINVAL if aio_context specified by ctx_id is
1642 * invalid. May fail with -EAGAIN if the iocb specified was not
1643 * cancelled. Will fail with -ENOSYS if not implemented.
1644 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001645SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1646 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 struct kioctx *ctx;
1649 struct kiocb *kiocb;
1650 u32 key;
1651 int ret;
1652
1653 ret = get_user(key, &iocb->aio_key);
1654 if (unlikely(ret))
1655 return -EFAULT;
1656
1657 ctx = lookup_ioctx(ctx_id);
1658 if (unlikely(!ctx))
1659 return -EINVAL;
1660
1661 spin_lock_irq(&ctx->ctx_lock);
Kent Overstreet906b9732013-05-07 16:18:31 -07001662
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 kiocb = lookup_kiocb(ctx, iocb, key);
Kent Overstreet906b9732013-05-07 16:18:31 -07001664 if (kiocb)
Fabian Frederickd52a8f92014-04-22 07:26:58 +02001665 ret = kiocb_cancel(kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -07001666 else
1667 ret = -EINVAL;
1668
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669 spin_unlock_irq(&ctx->ctx_lock);
1670
Kent Overstreet906b9732013-05-07 16:18:31 -07001671 if (!ret) {
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001672 /*
1673 * The result argument is no longer used - the io_event is
1674 * always delivered via the ring buffer. -EINPROGRESS indicates
1675 * cancellation is progress:
Kent Overstreet906b9732013-05-07 16:18:31 -07001676 */
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001677 ret = -EINPROGRESS;
Kent Overstreet906b9732013-05-07 16:18:31 -07001678 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679
Kent Overstreet723be6e2013-05-28 15:14:48 -07001680 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681
1682 return ret;
1683}
1684
1685/* io_getevents:
1686 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001687 * the completion queue for the aio_context specified by ctx_id. If
1688 * it succeeds, the number of read events is returned. May fail with
1689 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1690 * out of range, if timeout is out of range. May fail with -EFAULT
1691 * if any of the memory specified is invalid. May return 0 or
1692 * < min_nr if the timeout specified by timeout has elapsed
1693 * before sufficient events are available, where timeout == NULL
1694 * specifies an infinite timeout. Note that the timeout pointed to by
Jeff Moyer69008072013-05-24 15:55:24 -07001695 * timeout is relative. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001697SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1698 long, min_nr,
1699 long, nr,
1700 struct io_event __user *, events,
1701 struct timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702{
1703 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1704 long ret = -EINVAL;
1705
1706 if (likely(ioctx)) {
Namhyung Kim2e410252011-01-12 17:01:08 -08001707 if (likely(min_nr <= nr && min_nr >= 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 ret = read_events(ioctx, min_nr, nr, events, timeout);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001709 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 return ret;
1712}