blob: fb8e45b88cd4ecf0aa4b3c485051d7f437ba7bd8 [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 Moyer9d85cba72010-05-26 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
Jens Axboedc48e562015-04-15 11:17:23 -060080struct ctx_rq_wait {
81 struct completion comp;
82 atomic_t count;
83};
84
Kent Overstreet4e179bc2013-05-07 16:18:33 -070085struct kioctx {
Kent Overstreet723be6e2013-05-28 15:14:48 -070086 struct percpu_ref users;
Kent Overstreet36f55882013-05-07 16:18:41 -070087 atomic_t dead;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070088
Kent Overstreete34ecee2013-10-10 19:31:47 -070089 struct percpu_ref reqs;
90
Kent Overstreet4e179bc2013-05-07 16:18:33 -070091 unsigned long user_id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070092
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100093 struct __percpu kioctx_cpu *cpu;
94
95 /*
96 * For percpu reqs_available, number of slots we move to/from global
97 * counter at a time:
98 */
99 unsigned req_batch;
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700100 /*
101 * This is what userspace passed to io_setup(), it's not used for
102 * anything but counting against the global max_reqs quota.
103 *
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700104 * The real limit is nr_events - 1, which will be larger (see
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700105 * aio_setup_ring())
106 */
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700107 unsigned max_reqs;
108
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700109 /* Size of ringbuffer, in units of struct io_event */
110 unsigned nr_events;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700111
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700112 unsigned long mmap_base;
113 unsigned long mmap_size;
114
115 struct page **ring_pages;
116 long nr_pages;
117
Kent Overstreet723be6e2013-05-28 15:14:48 -0700118 struct work_struct free_work;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700119
Anatol Pomozove02ba722014-04-15 11:31:33 -0700120 /*
121 * signals when all in-flight requests are done
122 */
Jens Axboedc48e562015-04-15 11:17:23 -0600123 struct ctx_rq_wait *rq_wait;
Anatol Pomozove02ba722014-04-15 11:31:33 -0700124
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700125 struct {
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000126 /*
127 * This counts the number of available slots in the ringbuffer,
128 * so we avoid overflowing it: it's decremented (if positive)
129 * when allocating a kiocb and incremented when the resulting
130 * io_event is pulled off the ringbuffer.
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000131 *
132 * We batch accesses to it with a percpu version.
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000133 */
134 atomic_t reqs_available;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700135 } ____cacheline_aligned_in_smp;
136
137 struct {
138 spinlock_t ctx_lock;
139 struct list_head active_reqs; /* used for cancellation */
140 } ____cacheline_aligned_in_smp;
141
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700142 struct {
143 struct mutex ring_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700144 wait_queue_head_t wait;
145 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700146
147 struct {
148 unsigned tail;
Benjamin LaHaised856f322014-08-24 13:14:05 -0400149 unsigned completed_events;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700150 spinlock_t completion_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700151 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700152
153 struct page *internal_pages[AIO_RING_PAGES];
Gu Zheng36bc08c2013-07-16 17:56:16 +0800154 struct file *aio_ring_file;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400155
156 unsigned id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700157};
158
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100159/*
160 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
161 * cancelled or completed (this makes a certain amount of sense because
162 * successful cancellation - io_cancel() - does deliver the completion to
163 * userspace).
164 *
165 * And since most things don't implement kiocb cancellation and we'd really like
166 * kiocb completion to be lockless when possible, we use ki_cancel to
167 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
168 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
169 */
170#define KIOCB_CANCELLED ((void *) (~0ULL))
171
172struct aio_kiocb {
173 struct kiocb common;
174
175 struct kioctx *ki_ctx;
176 kiocb_cancel_fn *ki_cancel;
177
178 struct iocb __user *ki_user_iocb; /* user's aiocb */
179 __u64 ki_user_data; /* user's data for completion */
180
181 struct list_head ki_list; /* the aio core uses this
182 * for cancellation */
183
184 /*
185 * If the aio_resfd field of the userspace iocb is not zero,
186 * this is the underlying eventfd context to deliver events to.
187 */
188 struct eventfd_ctx *ki_eventfd;
189};
190
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -0800192static DEFINE_SPINLOCK(aio_nr_lock);
193unsigned long aio_nr; /* current system wide number of aio requests */
194unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195/*----end sysctl variables---*/
196
Christoph Lametere18b8902006-12-06 20:33:20 -0800197static struct kmem_cache *kiocb_cachep;
198static struct kmem_cache *kioctx_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400200static struct vfsmount *aio_mnt;
201
202static const struct file_operations aio_ring_fops;
203static const struct address_space_operations aio_ctx_aops;
204
205static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
206{
207 struct qstr this = QSTR_INIT("[aio]", 5);
208 struct file *file;
209 struct path path;
210 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
Dan Carpenter7f626562013-11-13 10:49:40 +0300211 if (IS_ERR(inode))
212 return ERR_CAST(inode);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400213
214 inode->i_mapping->a_ops = &aio_ctx_aops;
215 inode->i_mapping->private_data = ctx;
216 inode->i_size = PAGE_SIZE * nr_pages;
217
218 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
219 if (!path.dentry) {
220 iput(inode);
221 return ERR_PTR(-ENOMEM);
222 }
223 path.mnt = mntget(aio_mnt);
224
225 d_instantiate(path.dentry, inode);
226 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
227 if (IS_ERR(file)) {
228 path_put(&path);
229 return file;
230 }
231
232 file->f_flags = O_RDWR;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400233 return file;
234}
235
236static struct dentry *aio_mount(struct file_system_type *fs_type,
237 int flags, const char *dev_name, void *data)
238{
239 static const struct dentry_operations ops = {
240 .d_dname = simple_dname,
241 };
Gu Zheng8dc43792014-07-23 18:03:52 +0800242 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400243}
244
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245/* aio_setup
246 * Creates the slab caches used by the aio routines, panic on
247 * failure as this is done early during the boot sequence.
248 */
249static int __init aio_setup(void)
250{
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400251 static struct file_system_type aio_fs = {
252 .name = "aio",
253 .mount = aio_mount,
254 .kill_sb = kill_anon_super,
255 };
256 aio_mnt = kern_mount(&aio_fs);
257 if (IS_ERR(aio_mnt))
258 panic("Failed to create aio fs mount.");
259
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100260 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Christoph Lameter0a31bd52007-05-06 14:49:57 -0700261 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262
Kent Overstreetcaf41672013-05-07 16:18:35 -0700263 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264
265 return 0;
266}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700267__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400269static void put_aio_ring_file(struct kioctx *ctx)
270{
271 struct file *aio_ring_file = ctx->aio_ring_file;
272 if (aio_ring_file) {
273 truncate_setsize(aio_ring_file->f_inode, 0);
274
275 /* Prevent further access to the kioctx from migratepages */
276 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
277 aio_ring_file->f_inode->i_mapping->private_data = NULL;
278 ctx->aio_ring_file = NULL;
279 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
280
281 fput(aio_ring_file);
282 }
283}
284
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285static void aio_free_ring(struct kioctx *ctx)
286{
Gu Zheng36bc08c2013-07-16 17:56:16 +0800287 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400289 /* Disconnect the kiotx from the ring file. This prevents future
290 * accesses to the kioctx from page migration.
291 */
292 put_aio_ring_file(ctx);
293
Gu Zheng36bc08c2013-07-16 17:56:16 +0800294 for (i = 0; i < ctx->nr_pages; i++) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500295 struct page *page;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800296 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
297 page_count(ctx->ring_pages[i]));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500298 page = ctx->ring_pages[i];
299 if (!page)
300 continue;
301 ctx->ring_pages[i] = NULL;
302 put_page(page);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800303 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304
Sasha Levinddb8c452013-11-19 17:33:03 -0500305 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700306 kfree(ctx->ring_pages);
Sasha Levinddb8c452013-11-19 17:33:03 -0500307 ctx->ring_pages = NULL;
308 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309}
310
Oleg Nesterov5477e702015-09-04 15:48:04 -0700311static int aio_ring_mremap(struct vm_area_struct *vma)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800312{
Oleg Nesterov5477e702015-09-04 15:48:04 -0700313 struct file *file = vma->vm_file;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400314 struct mm_struct *mm = vma->vm_mm;
315 struct kioctx_table *table;
Al Virob2edffd2015-04-06 17:48:54 -0400316 int i, res = -EINVAL;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400317
318 spin_lock(&mm->ioctx_lock);
319 rcu_read_lock();
320 table = rcu_dereference(mm->ioctx_table);
321 for (i = 0; i < table->nr; i++) {
322 struct kioctx *ctx;
323
324 ctx = table->table[i];
325 if (ctx && ctx->aio_ring_file == file) {
Al Virob2edffd2015-04-06 17:48:54 -0400326 if (!atomic_read(&ctx->dead)) {
327 ctx->user_id = ctx->mmap_base = vma->vm_start;
328 res = 0;
329 }
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400330 break;
331 }
332 }
333
334 rcu_read_unlock();
335 spin_unlock(&mm->ioctx_lock);
Al Virob2edffd2015-04-06 17:48:54 -0400336 return res;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400337}
338
Oleg Nesterov5477e702015-09-04 15:48:04 -0700339static const struct vm_operations_struct aio_ring_vm_ops = {
340 .mremap = aio_ring_mremap,
341#if IS_ENABLED(CONFIG_MMU)
342 .fault = filemap_fault,
343 .map_pages = filemap_map_pages,
344 .page_mkwrite = filemap_page_mkwrite,
345#endif
346};
347
348static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
349{
350 vma->vm_flags |= VM_DONTEXPAND;
351 vma->vm_ops = &aio_ring_vm_ops;
352 return 0;
353}
354
Gu Zheng36bc08c2013-07-16 17:56:16 +0800355static const struct file_operations aio_ring_fops = {
356 .mmap = aio_ring_mmap,
357};
358
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400359#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800360static int aio_migratepage(struct address_space *mapping, struct page *new,
361 struct page *old, enum migrate_mode mode)
362{
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400363 struct kioctx *ctx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800364 unsigned long flags;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400365 pgoff_t idx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800366 int rc;
367
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500368 rc = 0;
369
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400370 /* mapping->private_lock here protects against the kioctx teardown. */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500371 spin_lock(&mapping->private_lock);
372 ctx = mapping->private_data;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400373 if (!ctx) {
374 rc = -EINVAL;
375 goto out;
376 }
377
378 /* The ring_lock mutex. The prevents aio_read_events() from writing
379 * to the ring's head, and prevents page migration from mucking in
380 * a partially initialized kiotx.
381 */
382 if (!mutex_trylock(&ctx->ring_lock)) {
383 rc = -EAGAIN;
384 goto out;
385 }
386
387 idx = old->index;
388 if (idx < (pgoff_t)ctx->nr_pages) {
389 /* Make sure the old page hasn't already been changed */
390 if (ctx->ring_pages[idx] != old)
391 rc = -EAGAIN;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500392 } else
393 rc = -EINVAL;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500394
395 if (rc != 0)
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400396 goto out_unlock;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500397
Gu Zheng36bc08c2013-07-16 17:56:16 +0800398 /* Writeback must be complete */
399 BUG_ON(PageWriteback(old));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500400 get_page(new);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800401
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500402 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800403 if (rc != MIGRATEPAGE_SUCCESS) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500404 put_page(new);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400405 goto out_unlock;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800406 }
407
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400408 /* Take completion_lock to prevent other writes to the ring buffer
409 * while the old page is copied to the new. This prevents new
410 * events from being lost.
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400411 */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400412 spin_lock_irqsave(&ctx->completion_lock, flags);
413 migrate_page_copy(new, old);
414 BUG_ON(ctx->ring_pages[idx] != old);
415 ctx->ring_pages[idx] = new;
416 spin_unlock_irqrestore(&ctx->completion_lock, flags);
417
418 /* The old page is no longer accessible. */
419 put_page(old);
420
421out_unlock:
422 mutex_unlock(&ctx->ring_lock);
423out:
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400424 spin_unlock(&mapping->private_lock);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800425 return rc;
426}
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400427#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800428
429static const struct address_space_operations aio_ctx_aops = {
Gu Zheng835f2522014-11-06 17:46:21 +0800430 .set_page_dirty = __set_page_dirty_no_writeback,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400431#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800432 .migratepage = aio_migratepage,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400433#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800434};
435
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436static int aio_setup_ring(struct kioctx *ctx)
437{
438 struct aio_ring *ring;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 unsigned nr_events = ctx->max_reqs;
Zach Brown41003a72013-05-07 16:18:25 -0700440 struct mm_struct *mm = current->mm;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900441 unsigned long size, unused;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 int nr_pages;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800443 int i;
444 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445
446 /* Compensate for the ring buffer's head/tail overlap entry */
447 nr_events += 2; /* 1 is required, 2 for good luck */
448
449 size = sizeof(struct aio_ring);
450 size += sizeof(struct io_event) * nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451
Gu Zheng36bc08c2013-07-16 17:56:16 +0800452 nr_pages = PFN_UP(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 if (nr_pages < 0)
454 return -EINVAL;
455
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400456 file = aio_private_file(ctx, nr_pages);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800457 if (IS_ERR(file)) {
458 ctx->aio_ring_file = NULL;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400459 return -ENOMEM;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800460 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461
Gu Zheng36bc08c2013-07-16 17:56:16 +0800462 ctx->aio_ring_file = file;
463 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
464 / sizeof(struct io_event);
465
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700466 ctx->ring_pages = ctx->internal_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 if (nr_pages > AIO_RING_PAGES) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700468 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
469 GFP_KERNEL);
Gu Zhengd1b94322013-12-04 18:19:06 +0800470 if (!ctx->ring_pages) {
471 put_aio_ring_file(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 return -ENOMEM;
Gu Zhengd1b94322013-12-04 18:19:06 +0800473 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474 }
475
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900476 for (i = 0; i < nr_pages; i++) {
477 struct page *page;
478 page = find_or_create_page(file->f_inode->i_mapping,
479 i, GFP_HIGHUSER | __GFP_ZERO);
480 if (!page)
481 break;
482 pr_debug("pid(%d) page[%d]->count=%d\n",
483 current->pid, i, page_count(page));
484 SetPageUptodate(page);
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900485 unlock_page(page);
486
487 ctx->ring_pages[i] = page;
488 }
489 ctx->nr_pages = i;
490
491 if (unlikely(i != nr_pages)) {
492 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400493 return -ENOMEM;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900494 }
495
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700496 ctx->mmap_size = nr_pages * PAGE_SIZE;
497 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800498
Michal Hocko013373e2016-05-23 16:25:59 -0700499 if (down_write_killable(&mm->mmap_sem)) {
500 ctx->mmap_size = 0;
501 aio_free_ring(ctx);
502 return -EINTR;
503 }
504
Gu Zheng36bc08c2013-07-16 17:56:16 +0800505 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
506 PROT_READ | PROT_WRITE,
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900507 MAP_SHARED, 0, &unused);
508 up_write(&mm->mmap_sem);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700509 if (IS_ERR((void *)ctx->mmap_base)) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700510 ctx->mmap_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400512 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 }
514
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700515 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400516
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700517 ctx->user_id = ctx->mmap_base;
518 ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700520 ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 ring->nr = nr_events; /* user copy */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400522 ring->id = ~0U;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 ring->head = ring->tail = 0;
524 ring->magic = AIO_RING_MAGIC;
525 ring->compat_features = AIO_RING_COMPAT_FEATURES;
526 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
527 ring->header_length = sizeof(struct aio_ring);
Cong Wange8e3c3d2011-11-25 23:14:27 +0800528 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700529 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530
531 return 0;
532}
533
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
535#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
536#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
537
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100538void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
Kent Overstreet0460fef2013-05-07 16:18:49 -0700539{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100540 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
Kent Overstreet0460fef2013-05-07 16:18:49 -0700541 struct kioctx *ctx = req->ki_ctx;
542 unsigned long flags;
543
544 spin_lock_irqsave(&ctx->ctx_lock, flags);
545
546 if (!req->ki_list.next)
547 list_add(&req->ki_list, &ctx->active_reqs);
548
549 req->ki_cancel = cancel;
550
551 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
552}
553EXPORT_SYMBOL(kiocb_set_cancel_fn);
554
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100555static int kiocb_cancel(struct aio_kiocb *kiocb)
Kent Overstreet906b9732013-05-07 16:18:31 -0700556{
Kent Overstreet0460fef2013-05-07 16:18:49 -0700557 kiocb_cancel_fn *old, *cancel;
Kent Overstreet906b9732013-05-07 16:18:31 -0700558
Kent Overstreet0460fef2013-05-07 16:18:49 -0700559 /*
560 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
561 * actually has a cancel function, hence the cmpxchg()
562 */
Kent Overstreet906b9732013-05-07 16:18:31 -0700563
Kent Overstreet0460fef2013-05-07 16:18:49 -0700564 cancel = ACCESS_ONCE(kiocb->ki_cancel);
565 do {
566 if (!cancel || cancel == KIOCB_CANCELLED)
Kent Overstreet57282d82013-05-13 13:42:52 -0700567 return -EINVAL;
Kent Overstreet906b9732013-05-07 16:18:31 -0700568
Kent Overstreet0460fef2013-05-07 16:18:49 -0700569 old = cancel;
570 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
571 } while (cancel != old);
572
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100573 return cancel(&kiocb->common);
Kent Overstreet906b9732013-05-07 16:18:31 -0700574}
575
Kent Overstreete34ecee2013-10-10 19:31:47 -0700576static void free_ioctx(struct work_struct *work)
Kent Overstreet36f55882013-05-07 16:18:41 -0700577{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700578 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000579
Kent Overstreete34ecee2013-10-10 19:31:47 -0700580 pr_debug("freeing %p\n", ctx);
581
582 aio_free_ring(ctx);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000583 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400584 percpu_ref_exit(&ctx->reqs);
585 percpu_ref_exit(&ctx->users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700586 kmem_cache_free(kioctx_cachep, ctx);
587}
588
Kent Overstreete34ecee2013-10-10 19:31:47 -0700589static void free_ioctx_reqs(struct percpu_ref *ref)
590{
591 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
592
Anatol Pomozove02ba722014-04-15 11:31:33 -0700593 /* At this point we know that there are no any in-flight requests */
Jens Axboedc48e562015-04-15 11:17:23 -0600594 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
595 complete(&ctx->rq_wait->comp);
Anatol Pomozove02ba722014-04-15 11:31:33 -0700596
Kent Overstreete34ecee2013-10-10 19:31:47 -0700597 INIT_WORK(&ctx->free_work, free_ioctx);
598 schedule_work(&ctx->free_work);
599}
600
Kent Overstreet36f55882013-05-07 16:18:41 -0700601/*
602 * When this function runs, the kioctx has been removed from the "hash table"
603 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
604 * now it's safe to cancel any that need to be.
605 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700606static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet36f55882013-05-07 16:18:41 -0700607{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700608 struct kioctx *ctx = container_of(ref, struct kioctx, users);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100609 struct aio_kiocb *req;
Kent Overstreet36f55882013-05-07 16:18:41 -0700610
611 spin_lock_irq(&ctx->ctx_lock);
612
613 while (!list_empty(&ctx->active_reqs)) {
614 req = list_first_entry(&ctx->active_reqs,
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100615 struct aio_kiocb, ki_list);
Kent Overstreet36f55882013-05-07 16:18:41 -0700616
617 list_del_init(&req->ki_list);
Fabian Frederickd52a8f92014-04-22 07:26:58 +0200618 kiocb_cancel(req);
Kent Overstreet36f55882013-05-07 16:18:41 -0700619 }
620
621 spin_unlock_irq(&ctx->ctx_lock);
622
Kent Overstreete34ecee2013-10-10 19:31:47 -0700623 percpu_ref_kill(&ctx->reqs);
624 percpu_ref_put(&ctx->reqs);
Kent Overstreet36f55882013-05-07 16:18:41 -0700625}
626
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400627static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
628{
629 unsigned i, new_nr;
630 struct kioctx_table *table, *old;
631 struct aio_ring *ring;
632
633 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200634 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400635
636 while (1) {
637 if (table)
638 for (i = 0; i < table->nr; i++)
639 if (!table->table[i]) {
640 ctx->id = i;
641 table->table[i] = ctx;
642 spin_unlock(&mm->ioctx_lock);
643
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400644 /* While kioctx setup is in progress,
645 * we are protected from page migration
646 * changes ring_pages by ->ring_lock.
647 */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400648 ring = kmap_atomic(ctx->ring_pages[0]);
649 ring->id = ctx->id;
650 kunmap_atomic(ring);
651 return 0;
652 }
653
654 new_nr = (table ? table->nr : 1) * 4;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400655 spin_unlock(&mm->ioctx_lock);
656
657 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
658 new_nr, GFP_KERNEL);
659 if (!table)
660 return -ENOMEM;
661
662 table->nr = new_nr;
663
664 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200665 old = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400666
667 if (!old) {
668 rcu_assign_pointer(mm->ioctx_table, table);
669 } else if (table->nr > old->nr) {
670 memcpy(table->table, old->table,
671 old->nr * sizeof(struct kioctx *));
672
673 rcu_assign_pointer(mm->ioctx_table, table);
674 kfree_rcu(old, rcu);
675 } else {
676 kfree(table);
677 table = old;
678 }
679 }
680}
681
Kent Overstreete34ecee2013-10-10 19:31:47 -0700682static void aio_nr_sub(unsigned nr)
683{
684 spin_lock(&aio_nr_lock);
685 if (WARN_ON(aio_nr - nr > aio_nr))
686 aio_nr = 0;
687 else
688 aio_nr -= nr;
689 spin_unlock(&aio_nr_lock);
690}
691
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692/* ioctx_alloc
693 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
694 */
695static struct kioctx *ioctx_alloc(unsigned nr_events)
696{
Zach Brown41003a72013-05-07 16:18:25 -0700697 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 struct kioctx *ctx;
Al Viroe23754f2012-03-06 14:33:22 -0500699 int err = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000701 /*
702 * We keep track of the number of available ringbuffer slots, to prevent
703 * overflow (reqs_available), and we also use percpu counters for this.
704 *
705 * So since up to half the slots might be on other cpu's percpu counters
706 * and unavailable, double nr_events so userspace sees what they
707 * expected: additionally, we move req_batch slots to/from percpu
708 * counters at a time, so make sure that isn't 0:
709 */
710 nr_events = max(nr_events, num_possible_cpus() * 4);
711 nr_events *= 2;
712
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 /* Prevent overflows */
Al Viro08397ac2015-03-31 11:43:52 -0400714 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 pr_debug("ENOMEM: nr_events too high\n");
716 return ERR_PTR(-EINVAL);
717 }
718
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400719 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 return ERR_PTR(-EAGAIN);
721
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800722 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 if (!ctx)
724 return ERR_PTR(-ENOMEM);
725
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 ctx->max_reqs = nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400728 spin_lock_init(&ctx->ctx_lock);
729 spin_lock_init(&ctx->completion_lock);
730 mutex_init(&ctx->ring_lock);
731 /* Protect against page migration throughout kiotx setup by keeping
732 * the ring_lock mutex held until setup is complete. */
733 mutex_lock(&ctx->ring_lock);
734 init_waitqueue_head(&ctx->wait);
735
736 INIT_LIST_HEAD(&ctx->active_reqs);
737
Tejun Heo2aad2a82014-09-24 13:31:50 -0400738 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700739 goto err;
740
Tejun Heo2aad2a82014-09-24 13:31:50 -0400741 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700742 goto err;
Kent Overstreet723be6e2013-05-28 15:14:48 -0700743
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000744 ctx->cpu = alloc_percpu(struct kioctx_cpu);
745 if (!ctx->cpu)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700746 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400748 err = aio_setup_ring(ctx);
749 if (err < 0)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700750 goto err;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000751
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000752 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000753 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
Benjamin LaHaise6878ea72013-07-31 10:34:18 -0400754 if (ctx->req_batch < 1)
755 ctx->req_batch = 1;
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000756
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 /* limit the number of system wide aios */
Al Viro9fa1cb32012-03-10 23:14:05 -0500758 spin_lock(&aio_nr_lock);
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400759 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
Al Viro2dd542b72012-03-10 23:10:35 -0500760 aio_nr + nr_events < aio_nr) {
Al Viro9fa1cb32012-03-10 23:14:05 -0500761 spin_unlock(&aio_nr_lock);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700762 err = -EAGAIN;
Gu Zhengd1b94322013-12-04 18:19:06 +0800763 goto err_ctx;
Al Viro2dd542b72012-03-10 23:10:35 -0500764 }
765 aio_nr += ctx->max_reqs;
Al Viro9fa1cb32012-03-10 23:14:05 -0500766 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767
Benjamin LaHaise18816862013-12-21 15:49:28 -0500768 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
769 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
Kent Overstreet723be6e2013-05-28 15:14:48 -0700770
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400771 err = ioctx_add_table(ctx, mm);
772 if (err)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700773 goto err_cleanup;
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400774
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400775 /* Release the ring_lock mutex now that all setup is complete. */
776 mutex_unlock(&ctx->ring_lock);
777
Kent Overstreetcaf41672013-05-07 16:18:35 -0700778 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700779 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 return ctx;
781
Kent Overstreete34ecee2013-10-10 19:31:47 -0700782err_cleanup:
783 aio_nr_sub(ctx->max_reqs);
Gu Zhengd1b94322013-12-04 18:19:06 +0800784err_ctx:
Al Virodeeb8522015-04-06 17:57:44 -0400785 atomic_set(&ctx->dead, 1);
786 if (ctx->mmap_size)
787 vm_munmap(ctx->mmap_base, ctx->mmap_size);
Gu Zhengd1b94322013-12-04 18:19:06 +0800788 aio_free_ring(ctx);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700789err:
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400790 mutex_unlock(&ctx->ring_lock);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000791 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400792 percpu_ref_exit(&ctx->reqs);
793 percpu_ref_exit(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreetcaf41672013-05-07 16:18:35 -0700795 pr_debug("error allocating ioctx %d\n", err);
Al Viroe23754f2012-03-06 14:33:22 -0500796 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797}
798
Kent Overstreet36f55882013-05-07 16:18:41 -0700799/* kill_ioctx
800 * Cancels all outstanding aio requests on an aio context. Used
801 * when the processes owning a context have all exited to encourage
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 * the rapid destruction of the kioctx.
803 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400804static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
Jens Axboedc48e562015-04-15 11:17:23 -0600805 struct ctx_rq_wait *wait)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806{
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400807 struct kioctx_table *table;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400808
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400809 spin_lock(&mm->ioctx_lock);
Al Virob2edffd2015-04-06 17:48:54 -0400810 if (atomic_xchg(&ctx->dead, 1)) {
811 spin_unlock(&mm->ioctx_lock);
812 return -EINVAL;
813 }
814
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200815 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400816 WARN_ON(ctx != table->table[ctx->id]);
817 table->table[ctx->id] = NULL;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400818 spin_unlock(&mm->ioctx_lock);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700819
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400820 /* percpu_ref_kill() will do the necessary call_rcu() */
821 wake_up_all(&ctx->wait);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700822
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400823 /*
824 * It'd be more correct to do this in free_ioctx(), after all
825 * the outstanding kiocbs have finished - but by then io_destroy
826 * has already returned, so io_setup() could potentially return
827 * -EAGAIN with no ioctxs actually in use (as far as userspace
828 * could tell).
829 */
830 aio_nr_sub(ctx->max_reqs);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400831
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400832 if (ctx->mmap_size)
833 vm_munmap(ctx->mmap_base, ctx->mmap_size);
834
Jens Axboedc48e562015-04-15 11:17:23 -0600835 ctx->rq_wait = wait;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400836 percpu_ref_kill(&ctx->users);
837 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838}
839
Kent Overstreet36f55882013-05-07 16:18:41 -0700840/*
841 * exit_aio: called when the last user of mm goes away. At this point, there is
842 * no way for any new requests to be submited or any of the io_* syscalls to be
843 * called on the context.
844 *
845 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
846 * them.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800848void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849{
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200850 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
Jens Axboedc48e562015-04-15 11:17:23 -0600851 struct ctx_rq_wait wait;
852 int i, skipped;
Jens Axboeabf137d2008-12-09 08:11:22 +0100853
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200854 if (!table)
855 return;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400856
Jens Axboedc48e562015-04-15 11:17:23 -0600857 atomic_set(&wait.count, table->nr);
858 init_completion(&wait.comp);
859
860 skipped = 0;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200861 for (i = 0; i < table->nr; ++i) {
862 struct kioctx *ctx = table->table[i];
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400863
Jens Axboedc48e562015-04-15 11:17:23 -0600864 if (!ctx) {
865 skipped++;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200866 continue;
Jens Axboedc48e562015-04-15 11:17:23 -0600867 }
868
Al Viro936af152012-04-20 21:49:41 -0400869 /*
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200870 * We don't need to bother with munmap() here - exit_mmap(mm)
871 * is coming and it'll unmap everything. And we simply can't,
872 * this is not necessarily our ->mm.
873 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
874 * that it needs to unmap the area, just set it to 0.
Al Viro936af152012-04-20 21:49:41 -0400875 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700876 ctx->mmap_size = 0;
Jens Axboedc48e562015-04-15 11:17:23 -0600877 kill_ioctx(mm, ctx, &wait);
878 }
Kent Overstreet36f55882013-05-07 16:18:41 -0700879
Jens Axboedc48e562015-04-15 11:17:23 -0600880 if (!atomic_sub_and_test(skipped, &wait.count)) {
Gu Zheng6098b452014-09-03 17:45:44 +0800881 /* Wait until all IO for the context are done. */
Jens Axboedc48e562015-04-15 11:17:23 -0600882 wait_for_completion(&wait.comp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 }
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200884
885 RCU_INIT_POINTER(mm->ioctx_table, NULL);
886 kfree(table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887}
888
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000889static void put_reqs_available(struct kioctx *ctx, unsigned nr)
890{
891 struct kioctx_cpu *kcpu;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400892 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000893
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400894 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400895 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000896 kcpu->reqs_available += nr;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400897
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000898 while (kcpu->reqs_available >= ctx->req_batch * 2) {
899 kcpu->reqs_available -= ctx->req_batch;
900 atomic_add(ctx->req_batch, &ctx->reqs_available);
901 }
902
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400903 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000904}
905
906static bool get_reqs_available(struct kioctx *ctx)
907{
908 struct kioctx_cpu *kcpu;
909 bool ret = false;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400910 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000911
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400912 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400913 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000914 if (!kcpu->reqs_available) {
915 int old, avail = atomic_read(&ctx->reqs_available);
916
917 do {
918 if (avail < ctx->req_batch)
919 goto out;
920
921 old = avail;
922 avail = atomic_cmpxchg(&ctx->reqs_available,
923 avail, avail - ctx->req_batch);
924 } while (avail != old);
925
926 kcpu->reqs_available += ctx->req_batch;
927 }
928
929 ret = true;
930 kcpu->reqs_available--;
931out:
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400932 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000933 return ret;
934}
935
Benjamin LaHaised856f322014-08-24 13:14:05 -0400936/* refill_reqs_available
937 * Updates the reqs_available reference counts used for tracking the
938 * number of free slots in the completion ring. This can be called
939 * from aio_complete() (to optimistically update reqs_available) or
940 * from aio_get_req() (the we're out of events case). It must be
941 * called holding ctx->completion_lock.
942 */
943static void refill_reqs_available(struct kioctx *ctx, unsigned head,
944 unsigned tail)
945{
946 unsigned events_in_ring, completed;
947
948 /* Clamp head since userland can write to it. */
949 head %= ctx->nr_events;
950 if (head <= tail)
951 events_in_ring = tail - head;
952 else
953 events_in_ring = ctx->nr_events - (head - tail);
954
955 completed = ctx->completed_events;
956 if (events_in_ring < completed)
957 completed -= events_in_ring;
958 else
959 completed = 0;
960
961 if (!completed)
962 return;
963
964 ctx->completed_events -= completed;
965 put_reqs_available(ctx, completed);
966}
967
968/* user_refill_reqs_available
969 * Called to refill reqs_available when aio_get_req() encounters an
970 * out of space in the completion ring.
971 */
972static void user_refill_reqs_available(struct kioctx *ctx)
973{
974 spin_lock_irq(&ctx->completion_lock);
975 if (ctx->completed_events) {
976 struct aio_ring *ring;
977 unsigned head;
978
979 /* Access of ring->head may race with aio_read_events_ring()
980 * here, but that's okay since whether we read the old version
981 * or the new version, and either will be valid. The important
982 * part is that head cannot pass tail since we prevent
983 * aio_complete() from updating tail by holding
984 * ctx->completion_lock. Even if head is invalid, the check
985 * against ctx->completed_events below will make sure we do the
986 * safe/right thing.
987 */
988 ring = kmap_atomic(ctx->ring_pages[0]);
989 head = ring->head;
990 kunmap_atomic(ring);
991
992 refill_reqs_available(ctx, head, ctx->tail);
993 }
994
995 spin_unlock_irq(&ctx->completion_lock);
996}
997
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998/* aio_get_req
Kent Overstreet57282d82013-05-13 13:42:52 -0700999 * Allocate a slot for an aio request.
1000 * Returns NULL if no requests are free.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001002static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001004 struct aio_kiocb *req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001005
Benjamin LaHaised856f322014-08-24 13:14:05 -04001006 if (!get_reqs_available(ctx)) {
1007 user_refill_reqs_available(ctx);
1008 if (!get_reqs_available(ctx))
1009 return NULL;
1010 }
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001011
Kent Overstreet0460fef2013-05-07 16:18:49 -07001012 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 if (unlikely(!req))
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001014 goto out_put;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015
Kent Overstreete34ecee2013-10-10 19:31:47 -07001016 percpu_ref_get(&ctx->reqs);
1017
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018 req->ki_ctx = ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019 return req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001020out_put:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001021 put_reqs_available(ctx, 1);
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001022 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023}
1024
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001025static void kiocb_free(struct aio_kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001027 if (req->common.ki_filp)
1028 fput(req->common.ki_filp);
Davide Libenzi13389012009-06-30 11:41:11 -07001029 if (req->ki_eventfd != NULL)
1030 eventfd_ctx_put(req->ki_eventfd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 kmem_cache_free(kiocb_cachep, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032}
1033
Adrian Bunkd5470b52008-04-29 00:58:57 -07001034static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001036 struct aio_ring __user *ring = (void __user *)ctx_id;
Jens Axboeabf137d2008-12-09 08:11:22 +01001037 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -07001038 struct kioctx *ctx, *ret = NULL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001039 struct kioctx_table *table;
1040 unsigned id;
1041
1042 if (get_user(id, &ring->id))
1043 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044
Jens Axboeabf137d2008-12-09 08:11:22 +01001045 rcu_read_lock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001046 table = rcu_dereference(mm->ioctx_table);
Jens Axboeabf137d2008-12-09 08:11:22 +01001047
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001048 if (!table || id >= table->nr)
1049 goto out;
1050
1051 ctx = table->table[id];
Benjamin LaHaisef30d7042013-08-07 18:23:48 -04001052 if (ctx && ctx->user_id == ctx_id) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001053 percpu_ref_get(&ctx->users);
1054 ret = ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +01001055 }
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001056out:
Jens Axboeabf137d2008-12-09 08:11:22 +01001057 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -07001058 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059}
1060
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061/* aio_complete
1062 * Called when the io request on the given iocb is complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001064static void aio_complete(struct kiocb *kiocb, long res, long res2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001066 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067 struct kioctx *ctx = iocb->ki_ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068 struct aio_ring *ring;
Kent Overstreet21b40202013-05-07 16:18:47 -07001069 struct io_event *ev_page, *event;
Benjamin LaHaised856f322014-08-24 13:14:05 -04001070 unsigned tail, pos, head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072
Zach Brown20dcae32005-11-13 16:07:33 -08001073 /*
1074 * Special case handling for sync iocbs:
1075 * - events go directly into the iocb for fast handling
1076 * - the sync task with the iocb in its stack holds the single iocb
1077 * ref, no other paths have a way to get another ref
1078 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001080 BUG_ON(is_sync_kiocb(kiocb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081
Kent Overstreet0460fef2013-05-07 16:18:49 -07001082 if (iocb->ki_list.next) {
1083 unsigned long flags;
1084
1085 spin_lock_irqsave(&ctx->ctx_lock, flags);
1086 list_del(&iocb->ki_list);
1087 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1088 }
Kent Overstreet11599eb2013-05-07 16:18:39 -07001089
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 /*
Kent Overstreet0460fef2013-05-07 16:18:49 -07001091 * Add a completion event to the ring buffer. Must be done holding
Tang Chen4b30f072013-07-03 15:09:16 -07001092 * ctx->completion_lock to prevent other code from messing with the tail
Kent Overstreet0460fef2013-05-07 16:18:49 -07001093 * pointer since we might be called from irq context.
1094 */
1095 spin_lock_irqsave(&ctx->completion_lock, flags);
1096
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001097 tail = ctx->tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001098 pos = tail + AIO_EVENTS_OFFSET;
1099
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001100 if (++tail >= ctx->nr_events)
Ken Chen4bf69b22005-05-01 08:59:15 -07001101 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001103 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001104 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1105
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001106 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 event->data = iocb->ki_user_data;
1108 event->res = res;
1109 event->res2 = res2;
1110
Kent Overstreet21b40202013-05-07 16:18:47 -07001111 kunmap_atomic(ev_page);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001112 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001113
1114 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001115 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
Kent Overstreetcaf41672013-05-07 16:18:35 -07001116 res, res2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117
1118 /* after flagging the request as done, we
1119 * must never even look at it again
1120 */
1121 smp_wmb(); /* make event visible before updating tail */
1122
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001123 ctx->tail = tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001124
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001125 ring = kmap_atomic(ctx->ring_pages[0]);
Benjamin LaHaised856f322014-08-24 13:14:05 -04001126 head = ring->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 ring->tail = tail;
Cong Wange8e3c3d2011-11-25 23:14:27 +08001128 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001129 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
Benjamin LaHaised856f322014-08-24 13:14:05 -04001131 ctx->completed_events++;
1132 if (ctx->completed_events > 1)
1133 refill_reqs_available(ctx, head, tail);
Kent Overstreet0460fef2013-05-07 16:18:49 -07001134 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1135
Kent Overstreet21b40202013-05-07 16:18:47 -07001136 pr_debug("added to ring %p at [%u]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001137
1138 /*
1139 * Check if the user asked us to deliver the result through an
1140 * eventfd. The eventfd_signal() function is safe to be called
1141 * from IRQ context.
1142 */
Davide Libenzi87c3a862009-03-18 17:04:19 -07001143 if (iocb->ki_eventfd != NULL)
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001144 eventfd_signal(iocb->ki_eventfd, 1);
1145
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146 /* everything turned out well, dispose of the aiocb. */
Kent Overstreet57282d82013-05-13 13:42:52 -07001147 kiocb_free(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148
Quentin Barnes6cb2a212008-03-19 17:00:39 -07001149 /*
1150 * We have to order our ring_info tail store above and test
1151 * of the wait list below outside the wait lock. This is
1152 * like in wake_up_bit() where clearing a bit has to be
1153 * ordered with the unlocked test.
1154 */
1155 smp_mb();
1156
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 if (waitqueue_active(&ctx->wait))
1158 wake_up(&ctx->wait);
1159
Kent Overstreete34ecee2013-10-10 19:31:47 -07001160 percpu_ref_put(&ctx->reqs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161}
1162
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001163/* aio_read_events_ring
Kent Overstreeta31ad382013-05-07 16:18:45 -07001164 * Pull an event off of the ioctx's event ring. Returns the number of
1165 * events fetched
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 */
Kent Overstreeta31ad382013-05-07 16:18:45 -07001167static long aio_read_events_ring(struct kioctx *ctx,
1168 struct io_event __user *event, long nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 struct aio_ring *ring;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001171 unsigned head, tail, pos;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001172 long ret = 0;
1173 int copy_ret;
1174
Dave Chinner9c9ce762015-02-03 19:29:05 -05001175 /*
1176 * The mutex can block and wake us up and that will cause
1177 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1178 * and repeat. This should be rare enough that it doesn't cause
1179 * peformance issues. See the comment in read_events() for more detail.
1180 */
1181 sched_annotate_sleep();
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001182 mutex_lock(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -04001184 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001185 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001186 head = ring->head;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001187 tail = ring->tail;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001188 kunmap_atomic(ring);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189
Jeff Moyer2ff396be2014-09-02 13:17:00 -04001190 /*
1191 * Ensure that once we've read the current tail pointer, that
1192 * we also see the events that were stored up to the tail.
1193 */
1194 smp_rmb();
1195
Kent Overstreet5ffac122013-05-09 15:36:07 -07001196 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001197
Kent Overstreet5ffac122013-05-09 15:36:07 -07001198 if (head == tail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 goto out;
1200
Benjamin LaHaiseedfbbf32014-06-24 13:32:51 -04001201 head %= ctx->nr_events;
1202 tail %= ctx->nr_events;
1203
Kent Overstreeta31ad382013-05-07 16:18:45 -07001204 while (ret < nr) {
1205 long avail;
1206 struct io_event *ev;
1207 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208
Kent Overstreet5ffac122013-05-09 15:36:07 -07001209 avail = (head <= tail ? tail : ctx->nr_events) - head;
1210 if (head == tail)
Kent Overstreeta31ad382013-05-07 16:18:45 -07001211 break;
1212
1213 avail = min(avail, nr - ret);
1214 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1215 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1216
1217 pos = head + AIO_EVENTS_OFFSET;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001218 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
Kent Overstreeta31ad382013-05-07 16:18:45 -07001219 pos %= AIO_EVENTS_PER_PAGE;
1220
1221 ev = kmap(page);
1222 copy_ret = copy_to_user(event + ret, ev + pos,
1223 sizeof(*ev) * avail);
1224 kunmap(page);
1225
1226 if (unlikely(copy_ret)) {
1227 ret = -EFAULT;
1228 goto out;
1229 }
1230
1231 ret += avail;
1232 head += avail;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001233 head %= ctx->nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001236 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001237 ring->head = head;
Zhao Hongjiang91d80a82013-04-26 11:03:53 +08001238 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001239 flush_dcache_page(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001240
Kent Overstreet5ffac122013-05-09 15:36:07 -07001241 pr_debug("%li h%u t%u\n", ret, head, tail);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001242out:
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001243 mutex_unlock(&ctx->ring_lock);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001244
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 return ret;
1246}
1247
Kent Overstreeta31ad382013-05-07 16:18:45 -07001248static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1249 struct io_event __user *event, long *i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001251 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252
Kent Overstreeta31ad382013-05-07 16:18:45 -07001253 if (ret > 0)
1254 *i += ret;
1255
1256 if (unlikely(atomic_read(&ctx->dead)))
1257 ret = -EINVAL;
1258
1259 if (!*i)
1260 *i = ret;
1261
1262 return ret < 0 || *i >= min_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263}
1264
Kent Overstreeta31ad382013-05-07 16:18:45 -07001265static long read_events(struct kioctx *ctx, long min_nr, long nr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 struct io_event __user *event,
1267 struct timespec __user *timeout)
1268{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001269 ktime_t until = { .tv64 = KTIME_MAX };
1270 long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272 if (timeout) {
1273 struct timespec ts;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001274
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
Kent Overstreeta31ad382013-05-07 16:18:45 -07001276 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277
Kent Overstreeta31ad382013-05-07 16:18:45 -07001278 until = timespec_to_ktime(ts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279 }
1280
Kent Overstreeta31ad382013-05-07 16:18:45 -07001281 /*
1282 * Note that aio_read_events() is being called as the conditional - i.e.
1283 * we're calling it after prepare_to_wait() has set task state to
1284 * TASK_INTERRUPTIBLE.
1285 *
1286 * But aio_read_events() can block, and if it blocks it's going to flip
1287 * the task state back to TASK_RUNNING.
1288 *
1289 * This should be ok, provided it doesn't flip the state back to
1290 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1291 * will only happen if the mutex_lock() call blocks, and we then find
1292 * the ringbuffer empty. So in practice we should be ok, but it's
1293 * something to be aware of when touching this code.
1294 */
Fam Zheng5f785de2014-11-06 20:44:36 +08001295 if (until.tv64 == 0)
1296 aio_read_events(ctx, min_nr, nr, event, &ret);
1297 else
1298 wait_event_interruptible_hrtimeout(ctx->wait,
1299 aio_read_events(ctx, min_nr, nr, event, &ret),
1300 until);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301
Kent Overstreeta31ad382013-05-07 16:18:45 -07001302 if (!ret && signal_pending(current))
1303 ret = -EINTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304
Kent Overstreeta31ad382013-05-07 16:18:45 -07001305 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306}
1307
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308/* sys_io_setup:
1309 * Create an aio_context capable of receiving at least nr_events.
1310 * ctxp must not point to an aio_context that already exists, and
1311 * must be initialized to 0 prior to the call. On successful
1312 * creation of the aio_context, *ctxp is filled in with the resulting
1313 * handle. May fail with -EINVAL if *ctxp is not initialized,
1314 * if the specified nr_events exceeds internal limits. May fail
1315 * with -EAGAIN if the specified nr_events exceeds the user's limit
1316 * of available events. May fail with -ENOMEM if insufficient kernel
1317 * resources are available. May fail with -EFAULT if an invalid
1318 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1319 * implemented.
1320 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001321SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322{
1323 struct kioctx *ioctx = NULL;
1324 unsigned long ctx;
1325 long ret;
1326
1327 ret = get_user(ctx, ctxp);
1328 if (unlikely(ret))
1329 goto out;
1330
1331 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001332 if (unlikely(ctx || nr_events == 0)) {
Kinglong Meeacd88d42015-02-04 21:15:59 +08001333 pr_debug("EINVAL: ctx %lu nr_events %u\n",
Zach Brownd55b5fd2005-11-07 00:59:31 -08001334 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 goto out;
1336 }
1337
1338 ioctx = ioctx_alloc(nr_events);
1339 ret = PTR_ERR(ioctx);
1340 if (!IS_ERR(ioctx)) {
1341 ret = put_user(ioctx->user_id, ctxp);
Al Viroa2e18592012-03-20 16:27:57 -04001342 if (ret)
Anatol Pomozove02ba722014-04-15 11:31:33 -07001343 kill_ioctx(current->mm, ioctx, NULL);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001344 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 }
1346
1347out:
1348 return ret;
1349}
1350
1351/* sys_io_destroy:
1352 * Destroy the aio_context specified. May cancel any outstanding
1353 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001354 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 * is invalid.
1356 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001357SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358{
1359 struct kioctx *ioctx = lookup_ioctx(ctx);
1360 if (likely(NULL != ioctx)) {
Jens Axboedc48e562015-04-15 11:17:23 -06001361 struct ctx_rq_wait wait;
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001362 int ret;
Anatol Pomozove02ba722014-04-15 11:31:33 -07001363
Jens Axboedc48e562015-04-15 11:17:23 -06001364 init_completion(&wait.comp);
1365 atomic_set(&wait.count, 1);
1366
Anatol Pomozove02ba722014-04-15 11:31:33 -07001367 /* Pass requests_done to kill_ioctx() where it can be set
1368 * in a thread-safe way. If we try to set it here then we have
1369 * a race condition if two io_destroy() called simultaneously.
1370 */
Jens Axboedc48e562015-04-15 11:17:23 -06001371 ret = kill_ioctx(current->mm, ioctx, &wait);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001372 percpu_ref_put(&ioctx->users);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001373
1374 /* Wait until all IO for the context are done. Otherwise kernel
1375 * keep using user-space buffers even if user thinks the context
1376 * is destroyed.
1377 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001378 if (!ret)
Jens Axboedc48e562015-04-15 11:17:23 -06001379 wait_for_completion(&wait.comp);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001380
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001381 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 }
Kinglong Meeacd88d42015-02-04 21:15:59 +08001383 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384 return -EINVAL;
1385}
1386
Al Viro293bc982014-02-11 18:37:41 -05001387typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001388
Al Viroa96114f2015-03-20 20:40:18 -04001389static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1390 struct iovec **iovec,
1391 bool compat,
1392 struct iov_iter *iter)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001393{
Jeff Moyer9d85cba72010-05-26 14:44:26 -07001394#ifdef CONFIG_COMPAT
1395 if (compat)
Al Viro32a56af2015-03-21 19:34:53 -04001396 return compat_import_iovec(rw,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001397 (struct compat_iovec __user *)buf,
Al Viro32a56af2015-03-21 19:34:53 -04001398 len, UIO_FASTIOV, iovec, iter);
Jeff Moyer9d85cba72010-05-26 14:44:26 -07001399#endif
Al Viro32a56af2015-03-21 19:34:53 -04001400 return import_iovec(rw, (struct iovec __user *)buf,
1401 len, UIO_FASTIOV, iovec, iter);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001402}
1403
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404/*
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001405 * aio_run_iocb:
1406 * Performs the initial checks and io submission.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407 */
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001408static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
Christoph Hellwig66ee59a2015-02-11 19:56:46 +01001409 char __user *buf, size_t len, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410{
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001411 struct file *file = req->ki_filp;
1412 ssize_t ret;
1413 int rw;
1414 fmode_t mode;
Al Viro293bc982014-02-11 18:37:41 -05001415 rw_iter_op *iter_op;
Gu Zheng00fefb92014-07-23 18:03:54 +08001416 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
Al Viro293bc982014-02-11 18:37:41 -05001417 struct iov_iter iter;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001419 switch (opcode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420 case IOCB_CMD_PREAD:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001421 case IOCB_CMD_PREADV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001422 mode = FMODE_READ;
1423 rw = READ;
Al Viro293bc982014-02-11 18:37:41 -05001424 iter_op = file->f_op->read_iter;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001425 goto rw_common;
1426
1427 case IOCB_CMD_PWRITE:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001428 case IOCB_CMD_PWRITEV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001429 mode = FMODE_WRITE;
1430 rw = WRITE;
Al Viro293bc982014-02-11 18:37:41 -05001431 iter_op = file->f_op->write_iter;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001432 goto rw_common;
1433rw_common:
1434 if (unlikely(!(file->f_mode & mode)))
1435 return -EBADF;
1436
Al Viro84363182015-04-04 01:14:53 -04001437 if (!iter_op)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001438 return -EINVAL;
1439
Christoph Hellwig66ee59a2015-02-11 19:56:46 +01001440 if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
Al Viroa96114f2015-03-20 20:40:18 -04001441 ret = aio_setup_vectored_rw(rw, buf, len,
1442 &iovec, compat, &iter);
Al Viro32a56af2015-03-21 19:34:53 -04001443 else {
Al Virod4fb3922015-03-21 19:11:55 -04001444 ret = import_single_range(rw, buf, len, iovec, &iter);
Al Viro32a56af2015-03-21 19:34:53 -04001445 iovec = NULL;
1446 }
Leon Yu754320d2014-05-01 03:31:28 +00001447 if (!ret)
Al Viroa96114f2015-03-20 20:40:18 -04001448 ret = rw_verify_area(rw, file, &req->ki_pos,
1449 iov_iter_count(&iter));
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001450 if (ret < 0) {
Al Viro32a56af2015-03-21 19:34:53 -04001451 kfree(iovec);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001452 return ret;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001453 }
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001454
Kent Overstreet73a70752013-05-09 15:03:42 -07001455 if (rw == WRITE)
1456 file_start_write(file);
1457
Al Viro84363182015-04-04 01:14:53 -04001458 ret = iter_op(req, &iter);
Kent Overstreet73a70752013-05-09 15:03:42 -07001459
1460 if (rw == WRITE)
1461 file_end_write(file);
Al Viro32a56af2015-03-21 19:34:53 -04001462 kfree(iovec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001464
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 case IOCB_CMD_FDSYNC:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001466 if (!file->f_op->aio_fsync)
1467 return -EINVAL;
1468
1469 ret = file->f_op->aio_fsync(req, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001471
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 case IOCB_CMD_FSYNC:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001473 if (!file->f_op->aio_fsync)
1474 return -EINVAL;
1475
1476 ret = file->f_op->aio_fsync(req, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001478
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 default:
Kent Overstreetcaf41672013-05-07 16:18:35 -07001480 pr_debug("EINVAL: no operation provided\n");
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001481 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482 }
1483
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001484 if (ret != -EIOCBQUEUED) {
1485 /*
1486 * There's no easy way to restart the syscall since other AIO's
1487 * may be already running. Just fail this IO with EINTR.
1488 */
1489 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1490 ret == -ERESTARTNOHAND ||
1491 ret == -ERESTART_RESTARTBLOCK))
1492 ret = -EINTR;
1493 aio_complete(req, ret, 0);
1494 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495
1496 return 0;
1497}
1498
Adrian Bunkd5470b52008-04-29 00:58:57 -07001499static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001500 struct iocb *iocb, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001502 struct aio_kiocb *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 ssize_t ret;
1504
1505 /* enforce forwards compatibility on users */
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001506 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001507 pr_debug("EINVAL: reserve field set\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508 return -EINVAL;
1509 }
1510
1511 /* prevent overflows */
1512 if (unlikely(
1513 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1514 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1515 ((ssize_t)iocb->aio_nbytes < 0)
1516 )) {
Kinglong Meeacd88d42015-02-04 21:15:59 +08001517 pr_debug("EINVAL: overflow check\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 return -EINVAL;
1519 }
1520
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001521 req = aio_get_req(ctx);
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001522 if (unlikely(!req))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523 return -EAGAIN;
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001524
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001525 req->common.ki_filp = fget(iocb->aio_fildes);
1526 if (unlikely(!req->common.ki_filp)) {
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001527 ret = -EBADF;
1528 goto out_put_req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 }
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001530 req->common.ki_pos = iocb->aio_offset;
1531 req->common.ki_complete = aio_complete;
Al Viro2ba48ce2015-04-09 13:52:01 -04001532 req->common.ki_flags = iocb_flags(req->common.ki_filp);
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001533
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001534 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1535 /*
1536 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1537 * instance of the file* now. The file descriptor must be
1538 * an eventfd() fd, and will be signaled for each completed
1539 * event using the eventfd_signal() function.
1540 */
Davide Libenzi13389012009-06-30 11:41:11 -07001541 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
Hirofumi Nakagawa801678c2008-04-29 01:03:09 -07001542 if (IS_ERR(req->ki_eventfd)) {
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001543 ret = PTR_ERR(req->ki_eventfd);
Davide Libenzi87c3a862009-03-18 17:04:19 -07001544 req->ki_eventfd = NULL;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001545 goto out_put_req;
1546 }
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001547
1548 req->common.ki_flags |= IOCB_EVENTFD;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001549 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550
Kent Overstreet8a660892013-05-07 16:19:10 -07001551 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552 if (unlikely(ret)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001553 pr_debug("EFAULT: aio_key\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554 goto out_put_req;
1555 }
1556
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001557 req->ki_user_iocb = user_iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558 req->ki_user_data = iocb->aio_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001560 ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001561 (char __user *)(unsigned long)iocb->aio_buf,
Christoph Hellwig66ee59a2015-02-11 19:56:46 +01001562 iocb->aio_nbytes,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001563 compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564 if (ret)
1565 goto out_put_req;
1566
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001568out_put_req:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001569 put_reqs_available(ctx, 1);
Kent Overstreete34ecee2013-10-10 19:31:47 -07001570 percpu_ref_put(&ctx->reqs);
Kent Overstreet57282d82013-05-13 13:42:52 -07001571 kiocb_free(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 return ret;
1573}
1574
Jeff Moyer9d85cba72010-05-26 14:44:26 -07001575long do_io_submit(aio_context_t ctx_id, long nr,
1576 struct iocb __user *__user *iocbpp, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577{
1578 struct kioctx *ctx;
1579 long ret = 0;
Jeff Moyer080d6762011-11-02 13:40:10 -07001580 int i = 0;
Shaohua Li9f5b9422010-07-01 07:55:01 +02001581 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582
1583 if (unlikely(nr < 0))
1584 return -EINVAL;
1585
Jeff Moyer75e1c702010-09-10 14:16:00 -07001586 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1587 nr = LONG_MAX/sizeof(*iocbpp);
1588
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1590 return -EFAULT;
1591
1592 ctx = lookup_ioctx(ctx_id);
1593 if (unlikely(!ctx)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001594 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595 return -EINVAL;
1596 }
1597
Shaohua Li9f5b9422010-07-01 07:55:01 +02001598 blk_start_plug(&plug);
1599
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600 /*
1601 * AKPM: should this return a partial result if some of the IOs were
1602 * successfully submitted?
1603 */
1604 for (i=0; i<nr; i++) {
1605 struct iocb __user *user_iocb;
1606 struct iocb tmp;
1607
1608 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1609 ret = -EFAULT;
1610 break;
1611 }
1612
1613 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1614 ret = -EFAULT;
1615 break;
1616 }
1617
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001618 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 if (ret)
1620 break;
1621 }
Shaohua Li9f5b9422010-07-01 07:55:01 +02001622 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623
Kent Overstreet723be6e2013-05-28 15:14:48 -07001624 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625 return i ? i : ret;
1626}
1627
Jeff Moyer9d85cba72010-05-26 14:44:26 -07001628/* sys_io_submit:
1629 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1630 * the number of iocbs queued. May return -EINVAL if the aio_context
1631 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1632 * *iocbpp[0] is not properly initialized, if the operation specified
1633 * is invalid for the file descriptor in the iocb. May fail with
1634 * -EFAULT if any of the data structures point to invalid data. May
1635 * fail with -EBADF if the file descriptor specified in the first
1636 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1637 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1638 * fail with -ENOSYS if not implemented.
1639 */
1640SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1641 struct iocb __user * __user *, iocbpp)
1642{
1643 return do_io_submit(ctx_id, nr, iocbpp, 0);
1644}
1645
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646/* lookup_kiocb
1647 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001649static struct aio_kiocb *
1650lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001652 struct aio_kiocb *kiocb;
Zach Brownd00689a2005-11-13 16:07:34 -08001653
1654 assert_spin_locked(&ctx->ctx_lock);
1655
Kent Overstreet8a660892013-05-07 16:19:10 -07001656 if (key != KIOCB_KEY)
1657 return NULL;
1658
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 /* TODO: use a hash or array, this sucks. */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001660 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1661 if (kiocb->ki_user_iocb == iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662 return kiocb;
1663 }
1664 return NULL;
1665}
1666
1667/* sys_io_cancel:
1668 * Attempts to cancel an iocb previously passed to io_submit. If
1669 * the operation is successfully cancelled, the resulting event is
1670 * copied into the memory pointed to by result without being placed
1671 * into the completion queue and 0 is returned. May fail with
1672 * -EFAULT if any of the data structures pointed to are invalid.
1673 * May fail with -EINVAL if aio_context specified by ctx_id is
1674 * invalid. May fail with -EAGAIN if the iocb specified was not
1675 * cancelled. Will fail with -ENOSYS if not implemented.
1676 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001677SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1678 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680 struct kioctx *ctx;
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001681 struct aio_kiocb *kiocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 u32 key;
1683 int ret;
1684
1685 ret = get_user(key, &iocb->aio_key);
1686 if (unlikely(ret))
1687 return -EFAULT;
1688
1689 ctx = lookup_ioctx(ctx_id);
1690 if (unlikely(!ctx))
1691 return -EINVAL;
1692
1693 spin_lock_irq(&ctx->ctx_lock);
Kent Overstreet906b9732013-05-07 16:18:31 -07001694
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695 kiocb = lookup_kiocb(ctx, iocb, key);
Kent Overstreet906b9732013-05-07 16:18:31 -07001696 if (kiocb)
Fabian Frederickd52a8f92014-04-22 07:26:58 +02001697 ret = kiocb_cancel(kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -07001698 else
1699 ret = -EINVAL;
1700
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 spin_unlock_irq(&ctx->ctx_lock);
1702
Kent Overstreet906b9732013-05-07 16:18:31 -07001703 if (!ret) {
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001704 /*
1705 * The result argument is no longer used - the io_event is
1706 * always delivered via the ring buffer. -EINPROGRESS indicates
1707 * cancellation is progress:
Kent Overstreet906b9732013-05-07 16:18:31 -07001708 */
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001709 ret = -EINPROGRESS;
Kent Overstreet906b9732013-05-07 16:18:31 -07001710 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711
Kent Overstreet723be6e2013-05-28 15:14:48 -07001712 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713
1714 return ret;
1715}
1716
1717/* io_getevents:
1718 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001719 * the completion queue for the aio_context specified by ctx_id. If
1720 * it succeeds, the number of read events is returned. May fail with
1721 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1722 * out of range, if timeout is out of range. May fail with -EFAULT
1723 * if any of the memory specified is invalid. May return 0 or
1724 * < min_nr if the timeout specified by timeout has elapsed
1725 * before sufficient events are available, where timeout == NULL
1726 * specifies an infinite timeout. Note that the timeout pointed to by
Jeff Moyer69008072013-05-24 15:55:24 -07001727 * timeout is relative. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001729SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1730 long, min_nr,
1731 long, nr,
1732 struct io_event __user *, events,
1733 struct timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734{
1735 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1736 long ret = -EINVAL;
1737
1738 if (likely(ioctx)) {
Namhyung Kim2e410252011-01-12 17:01:08 -08001739 if (likely(min_nr <= nr && min_nr >= 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740 ret = read_events(ioctx, min_nr, nr, events, timeout);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001741 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 return ret;
1744}