blob: c19755187ca592eeaeabf7c21d765897aba802aa [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
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 };
Jann Horn22f6b4d2016-09-16 00:31:22 +0200242 struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops,
243 AIO_RING_MAGIC);
244
245 if (!IS_ERR(root))
246 root->d_sb->s_iflags |= SB_I_NOEXEC;
247 return root;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400248}
249
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250/* aio_setup
251 * Creates the slab caches used by the aio routines, panic on
252 * failure as this is done early during the boot sequence.
253 */
254static int __init aio_setup(void)
255{
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400256 static struct file_system_type aio_fs = {
257 .name = "aio",
258 .mount = aio_mount,
259 .kill_sb = kill_anon_super,
260 };
261 aio_mnt = kern_mount(&aio_fs);
262 if (IS_ERR(aio_mnt))
263 panic("Failed to create aio fs mount.");
264
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100265 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Christoph Lameter0a31bd52007-05-06 14:49:57 -0700266 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267
Kent Overstreetcaf41672013-05-07 16:18:35 -0700268 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
270 return 0;
271}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700272__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400274static void put_aio_ring_file(struct kioctx *ctx)
275{
276 struct file *aio_ring_file = ctx->aio_ring_file;
Rasmus Villemoesde04e762016-09-15 00:25:03 +0200277 struct address_space *i_mapping;
278
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400279 if (aio_ring_file) {
280 truncate_setsize(aio_ring_file->f_inode, 0);
281
282 /* Prevent further access to the kioctx from migratepages */
Rasmus Villemoesde04e762016-09-15 00:25:03 +0200283 i_mapping = aio_ring_file->f_inode->i_mapping;
284 spin_lock(&i_mapping->private_lock);
285 i_mapping->private_data = NULL;
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400286 ctx->aio_ring_file = NULL;
Rasmus Villemoesde04e762016-09-15 00:25:03 +0200287 spin_unlock(&i_mapping->private_lock);
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400288
289 fput(aio_ring_file);
290 }
291}
292
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293static void aio_free_ring(struct kioctx *ctx)
294{
Gu Zheng36bc08c2013-07-16 17:56:16 +0800295 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400297 /* Disconnect the kiotx from the ring file. This prevents future
298 * accesses to the kioctx from page migration.
299 */
300 put_aio_ring_file(ctx);
301
Gu Zheng36bc08c2013-07-16 17:56:16 +0800302 for (i = 0; i < ctx->nr_pages; i++) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500303 struct page *page;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800304 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
305 page_count(ctx->ring_pages[i]));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500306 page = ctx->ring_pages[i];
307 if (!page)
308 continue;
309 ctx->ring_pages[i] = NULL;
310 put_page(page);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800311 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312
Sasha Levinddb8c452013-11-19 17:33:03 -0500313 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700314 kfree(ctx->ring_pages);
Sasha Levinddb8c452013-11-19 17:33:03 -0500315 ctx->ring_pages = NULL;
316 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317}
318
Oleg Nesterov5477e702015-09-04 15:48:04 -0700319static int aio_ring_mremap(struct vm_area_struct *vma)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800320{
Oleg Nesterov5477e702015-09-04 15:48:04 -0700321 struct file *file = vma->vm_file;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400322 struct mm_struct *mm = vma->vm_mm;
323 struct kioctx_table *table;
Al Virob2edffd2015-04-06 17:48:54 -0400324 int i, res = -EINVAL;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400325
326 spin_lock(&mm->ioctx_lock);
327 rcu_read_lock();
328 table = rcu_dereference(mm->ioctx_table);
329 for (i = 0; i < table->nr; i++) {
330 struct kioctx *ctx;
331
332 ctx = table->table[i];
333 if (ctx && ctx->aio_ring_file == file) {
Al Virob2edffd2015-04-06 17:48:54 -0400334 if (!atomic_read(&ctx->dead)) {
335 ctx->user_id = ctx->mmap_base = vma->vm_start;
336 res = 0;
337 }
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400338 break;
339 }
340 }
341
342 rcu_read_unlock();
343 spin_unlock(&mm->ioctx_lock);
Al Virob2edffd2015-04-06 17:48:54 -0400344 return res;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400345}
346
Oleg Nesterov5477e702015-09-04 15:48:04 -0700347static const struct vm_operations_struct aio_ring_vm_ops = {
348 .mremap = aio_ring_mremap,
349#if IS_ENABLED(CONFIG_MMU)
350 .fault = filemap_fault,
351 .map_pages = filemap_map_pages,
352 .page_mkwrite = filemap_page_mkwrite,
353#endif
354};
355
356static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
357{
358 vma->vm_flags |= VM_DONTEXPAND;
359 vma->vm_ops = &aio_ring_vm_ops;
360 return 0;
361}
362
Gu Zheng36bc08c2013-07-16 17:56:16 +0800363static const struct file_operations aio_ring_fops = {
364 .mmap = aio_ring_mmap,
365};
366
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400367#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800368static int aio_migratepage(struct address_space *mapping, struct page *new,
369 struct page *old, enum migrate_mode mode)
370{
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400371 struct kioctx *ctx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800372 unsigned long flags;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400373 pgoff_t idx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800374 int rc;
375
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500376 rc = 0;
377
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400378 /* mapping->private_lock here protects against the kioctx teardown. */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500379 spin_lock(&mapping->private_lock);
380 ctx = mapping->private_data;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400381 if (!ctx) {
382 rc = -EINVAL;
383 goto out;
384 }
385
386 /* The ring_lock mutex. The prevents aio_read_events() from writing
387 * to the ring's head, and prevents page migration from mucking in
388 * a partially initialized kiotx.
389 */
390 if (!mutex_trylock(&ctx->ring_lock)) {
391 rc = -EAGAIN;
392 goto out;
393 }
394
395 idx = old->index;
396 if (idx < (pgoff_t)ctx->nr_pages) {
397 /* Make sure the old page hasn't already been changed */
398 if (ctx->ring_pages[idx] != old)
399 rc = -EAGAIN;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500400 } else
401 rc = -EINVAL;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500402
403 if (rc != 0)
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400404 goto out_unlock;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500405
Gu Zheng36bc08c2013-07-16 17:56:16 +0800406 /* Writeback must be complete */
407 BUG_ON(PageWriteback(old));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500408 get_page(new);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800409
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500410 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800411 if (rc != MIGRATEPAGE_SUCCESS) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500412 put_page(new);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400413 goto out_unlock;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800414 }
415
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400416 /* Take completion_lock to prevent other writes to the ring buffer
417 * while the old page is copied to the new. This prevents new
418 * events from being lost.
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400419 */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400420 spin_lock_irqsave(&ctx->completion_lock, flags);
421 migrate_page_copy(new, old);
422 BUG_ON(ctx->ring_pages[idx] != old);
423 ctx->ring_pages[idx] = new;
424 spin_unlock_irqrestore(&ctx->completion_lock, flags);
425
426 /* The old page is no longer accessible. */
427 put_page(old);
428
429out_unlock:
430 mutex_unlock(&ctx->ring_lock);
431out:
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400432 spin_unlock(&mapping->private_lock);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800433 return rc;
434}
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400435#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800436
437static const struct address_space_operations aio_ctx_aops = {
Gu Zheng835f2522014-11-06 17:46:21 +0800438 .set_page_dirty = __set_page_dirty_no_writeback,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400439#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800440 .migratepage = aio_migratepage,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400441#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800442};
443
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444static int aio_setup_ring(struct kioctx *ctx)
445{
446 struct aio_ring *ring;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 unsigned nr_events = ctx->max_reqs;
Zach Brown41003a72013-05-07 16:18:25 -0700448 struct mm_struct *mm = current->mm;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900449 unsigned long size, unused;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 int nr_pages;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800451 int i;
452 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453
454 /* Compensate for the ring buffer's head/tail overlap entry */
455 nr_events += 2; /* 1 is required, 2 for good luck */
456
457 size = sizeof(struct aio_ring);
458 size += sizeof(struct io_event) * nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459
Gu Zheng36bc08c2013-07-16 17:56:16 +0800460 nr_pages = PFN_UP(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 if (nr_pages < 0)
462 return -EINVAL;
463
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400464 file = aio_private_file(ctx, nr_pages);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800465 if (IS_ERR(file)) {
466 ctx->aio_ring_file = NULL;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400467 return -ENOMEM;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800468 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469
Gu Zheng36bc08c2013-07-16 17:56:16 +0800470 ctx->aio_ring_file = file;
471 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
472 / sizeof(struct io_event);
473
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700474 ctx->ring_pages = ctx->internal_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475 if (nr_pages > AIO_RING_PAGES) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700476 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
477 GFP_KERNEL);
Gu Zhengd1b94322013-12-04 18:19:06 +0800478 if (!ctx->ring_pages) {
479 put_aio_ring_file(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480 return -ENOMEM;
Gu Zhengd1b94322013-12-04 18:19:06 +0800481 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 }
483
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900484 for (i = 0; i < nr_pages; i++) {
485 struct page *page;
486 page = find_or_create_page(file->f_inode->i_mapping,
487 i, GFP_HIGHUSER | __GFP_ZERO);
488 if (!page)
489 break;
490 pr_debug("pid(%d) page[%d]->count=%d\n",
491 current->pid, i, page_count(page));
492 SetPageUptodate(page);
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900493 unlock_page(page);
494
495 ctx->ring_pages[i] = page;
496 }
497 ctx->nr_pages = i;
498
499 if (unlikely(i != nr_pages)) {
500 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400501 return -ENOMEM;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900502 }
503
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700504 ctx->mmap_size = nr_pages * PAGE_SIZE;
505 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800506
Michal Hocko013373e2016-05-23 16:25:59 -0700507 if (down_write_killable(&mm->mmap_sem)) {
508 ctx->mmap_size = 0;
509 aio_free_ring(ctx);
510 return -EINTR;
511 }
512
Gu Zheng36bc08c2013-07-16 17:56:16 +0800513 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
514 PROT_READ | PROT_WRITE,
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900515 MAP_SHARED, 0, &unused);
516 up_write(&mm->mmap_sem);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700517 if (IS_ERR((void *)ctx->mmap_base)) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700518 ctx->mmap_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400520 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 }
522
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700523 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400524
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700525 ctx->user_id = ctx->mmap_base;
526 ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700528 ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 ring->nr = nr_events; /* user copy */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400530 ring->id = ~0U;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531 ring->head = ring->tail = 0;
532 ring->magic = AIO_RING_MAGIC;
533 ring->compat_features = AIO_RING_COMPAT_FEATURES;
534 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
535 ring->header_length = sizeof(struct aio_ring);
Cong Wange8e3c3d2011-11-25 23:14:27 +0800536 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700537 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538
539 return 0;
540}
541
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
543#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
544#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
545
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100546void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
Kent Overstreet0460fef2013-05-07 16:18:49 -0700547{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100548 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
Kent Overstreet0460fef2013-05-07 16:18:49 -0700549 struct kioctx *ctx = req->ki_ctx;
550 unsigned long flags;
551
552 spin_lock_irqsave(&ctx->ctx_lock, flags);
553
554 if (!req->ki_list.next)
555 list_add(&req->ki_list, &ctx->active_reqs);
556
557 req->ki_cancel = cancel;
558
559 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
560}
561EXPORT_SYMBOL(kiocb_set_cancel_fn);
562
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100563static int kiocb_cancel(struct aio_kiocb *kiocb)
Kent Overstreet906b9732013-05-07 16:18:31 -0700564{
Kent Overstreet0460fef2013-05-07 16:18:49 -0700565 kiocb_cancel_fn *old, *cancel;
Kent Overstreet906b9732013-05-07 16:18:31 -0700566
Kent Overstreet0460fef2013-05-07 16:18:49 -0700567 /*
568 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
569 * actually has a cancel function, hence the cmpxchg()
570 */
Kent Overstreet906b9732013-05-07 16:18:31 -0700571
Kent Overstreet0460fef2013-05-07 16:18:49 -0700572 cancel = ACCESS_ONCE(kiocb->ki_cancel);
573 do {
574 if (!cancel || cancel == KIOCB_CANCELLED)
Kent Overstreet57282d82013-05-13 13:42:52 -0700575 return -EINVAL;
Kent Overstreet906b9732013-05-07 16:18:31 -0700576
Kent Overstreet0460fef2013-05-07 16:18:49 -0700577 old = cancel;
578 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
579 } while (cancel != old);
580
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100581 return cancel(&kiocb->common);
Kent Overstreet906b9732013-05-07 16:18:31 -0700582}
583
Kent Overstreete34ecee2013-10-10 19:31:47 -0700584static void free_ioctx(struct work_struct *work)
Kent Overstreet36f55882013-05-07 16:18:41 -0700585{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700586 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000587
Kent Overstreete34ecee2013-10-10 19:31:47 -0700588 pr_debug("freeing %p\n", ctx);
589
590 aio_free_ring(ctx);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000591 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400592 percpu_ref_exit(&ctx->reqs);
593 percpu_ref_exit(&ctx->users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700594 kmem_cache_free(kioctx_cachep, ctx);
595}
596
Kent Overstreete34ecee2013-10-10 19:31:47 -0700597static void free_ioctx_reqs(struct percpu_ref *ref)
598{
599 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
600
Anatol Pomozove02ba722014-04-15 11:31:33 -0700601 /* At this point we know that there are no any in-flight requests */
Jens Axboedc48e562015-04-15 11:17:23 -0600602 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
603 complete(&ctx->rq_wait->comp);
Anatol Pomozove02ba722014-04-15 11:31:33 -0700604
Kent Overstreete34ecee2013-10-10 19:31:47 -0700605 INIT_WORK(&ctx->free_work, free_ioctx);
606 schedule_work(&ctx->free_work);
607}
608
Kent Overstreet36f55882013-05-07 16:18:41 -0700609/*
610 * When this function runs, the kioctx has been removed from the "hash table"
611 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
612 * now it's safe to cancel any that need to be.
613 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700614static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet36f55882013-05-07 16:18:41 -0700615{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700616 struct kioctx *ctx = container_of(ref, struct kioctx, users);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100617 struct aio_kiocb *req;
Kent Overstreet36f55882013-05-07 16:18:41 -0700618
619 spin_lock_irq(&ctx->ctx_lock);
620
621 while (!list_empty(&ctx->active_reqs)) {
622 req = list_first_entry(&ctx->active_reqs,
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100623 struct aio_kiocb, ki_list);
Kent Overstreet36f55882013-05-07 16:18:41 -0700624
625 list_del_init(&req->ki_list);
Fabian Frederickd52a8f92014-04-22 07:26:58 +0200626 kiocb_cancel(req);
Kent Overstreet36f55882013-05-07 16:18:41 -0700627 }
628
629 spin_unlock_irq(&ctx->ctx_lock);
630
Kent Overstreete34ecee2013-10-10 19:31:47 -0700631 percpu_ref_kill(&ctx->reqs);
632 percpu_ref_put(&ctx->reqs);
Kent Overstreet36f55882013-05-07 16:18:41 -0700633}
634
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400635static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
636{
637 unsigned i, new_nr;
638 struct kioctx_table *table, *old;
639 struct aio_ring *ring;
640
641 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200642 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400643
644 while (1) {
645 if (table)
646 for (i = 0; i < table->nr; i++)
647 if (!table->table[i]) {
648 ctx->id = i;
649 table->table[i] = ctx;
650 spin_unlock(&mm->ioctx_lock);
651
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400652 /* While kioctx setup is in progress,
653 * we are protected from page migration
654 * changes ring_pages by ->ring_lock.
655 */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400656 ring = kmap_atomic(ctx->ring_pages[0]);
657 ring->id = ctx->id;
658 kunmap_atomic(ring);
659 return 0;
660 }
661
662 new_nr = (table ? table->nr : 1) * 4;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400663 spin_unlock(&mm->ioctx_lock);
664
665 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
666 new_nr, GFP_KERNEL);
667 if (!table)
668 return -ENOMEM;
669
670 table->nr = new_nr;
671
672 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200673 old = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400674
675 if (!old) {
676 rcu_assign_pointer(mm->ioctx_table, table);
677 } else if (table->nr > old->nr) {
678 memcpy(table->table, old->table,
679 old->nr * sizeof(struct kioctx *));
680
681 rcu_assign_pointer(mm->ioctx_table, table);
682 kfree_rcu(old, rcu);
683 } else {
684 kfree(table);
685 table = old;
686 }
687 }
688}
689
Kent Overstreete34ecee2013-10-10 19:31:47 -0700690static void aio_nr_sub(unsigned nr)
691{
692 spin_lock(&aio_nr_lock);
693 if (WARN_ON(aio_nr - nr > aio_nr))
694 aio_nr = 0;
695 else
696 aio_nr -= nr;
697 spin_unlock(&aio_nr_lock);
698}
699
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700/* ioctx_alloc
701 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
702 */
703static struct kioctx *ioctx_alloc(unsigned nr_events)
704{
Zach Brown41003a72013-05-07 16:18:25 -0700705 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 struct kioctx *ctx;
Al Viroe23754f2012-03-06 14:33:22 -0500707 int err = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000709 /*
710 * We keep track of the number of available ringbuffer slots, to prevent
711 * overflow (reqs_available), and we also use percpu counters for this.
712 *
713 * So since up to half the slots might be on other cpu's percpu counters
714 * and unavailable, double nr_events so userspace sees what they
715 * expected: additionally, we move req_batch slots to/from percpu
716 * counters at a time, so make sure that isn't 0:
717 */
718 nr_events = max(nr_events, num_possible_cpus() * 4);
719 nr_events *= 2;
720
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 /* Prevent overflows */
Al Viro08397ac2015-03-31 11:43:52 -0400722 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 pr_debug("ENOMEM: nr_events too high\n");
724 return ERR_PTR(-EINVAL);
725 }
726
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400727 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728 return ERR_PTR(-EAGAIN);
729
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800730 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731 if (!ctx)
732 return ERR_PTR(-ENOMEM);
733
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734 ctx->max_reqs = nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400736 spin_lock_init(&ctx->ctx_lock);
737 spin_lock_init(&ctx->completion_lock);
738 mutex_init(&ctx->ring_lock);
739 /* Protect against page migration throughout kiotx setup by keeping
740 * the ring_lock mutex held until setup is complete. */
741 mutex_lock(&ctx->ring_lock);
742 init_waitqueue_head(&ctx->wait);
743
744 INIT_LIST_HEAD(&ctx->active_reqs);
745
Tejun Heo2aad2a82014-09-24 13:31:50 -0400746 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700747 goto err;
748
Tejun Heo2aad2a82014-09-24 13:31:50 -0400749 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700750 goto err;
Kent Overstreet723be6e2013-05-28 15:14:48 -0700751
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000752 ctx->cpu = alloc_percpu(struct kioctx_cpu);
753 if (!ctx->cpu)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700754 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400756 err = aio_setup_ring(ctx);
757 if (err < 0)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700758 goto err;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000759
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000760 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000761 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
Benjamin LaHaise6878ea72013-07-31 10:34:18 -0400762 if (ctx->req_batch < 1)
763 ctx->req_batch = 1;
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000764
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 /* limit the number of system wide aios */
Al Viro9fa1cb32012-03-10 23:14:05 -0500766 spin_lock(&aio_nr_lock);
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400767 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
Al Viro2dd542b2012-03-10 23:10:35 -0500768 aio_nr + nr_events < aio_nr) {
Al Viro9fa1cb32012-03-10 23:14:05 -0500769 spin_unlock(&aio_nr_lock);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700770 err = -EAGAIN;
Gu Zhengd1b94322013-12-04 18:19:06 +0800771 goto err_ctx;
Al Viro2dd542b2012-03-10 23:10:35 -0500772 }
773 aio_nr += ctx->max_reqs;
Al Viro9fa1cb32012-03-10 23:14:05 -0500774 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775
Benjamin LaHaise18816862013-12-21 15:49:28 -0500776 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
777 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
Kent Overstreet723be6e2013-05-28 15:14:48 -0700778
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400779 err = ioctx_add_table(ctx, mm);
780 if (err)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700781 goto err_cleanup;
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400782
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400783 /* Release the ring_lock mutex now that all setup is complete. */
784 mutex_unlock(&ctx->ring_lock);
785
Kent Overstreetcaf41672013-05-07 16:18:35 -0700786 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700787 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 return ctx;
789
Kent Overstreete34ecee2013-10-10 19:31:47 -0700790err_cleanup:
791 aio_nr_sub(ctx->max_reqs);
Gu Zhengd1b94322013-12-04 18:19:06 +0800792err_ctx:
Al Virodeeb8522015-04-06 17:57:44 -0400793 atomic_set(&ctx->dead, 1);
794 if (ctx->mmap_size)
795 vm_munmap(ctx->mmap_base, ctx->mmap_size);
Gu Zhengd1b94322013-12-04 18:19:06 +0800796 aio_free_ring(ctx);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700797err:
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400798 mutex_unlock(&ctx->ring_lock);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000799 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400800 percpu_ref_exit(&ctx->reqs);
801 percpu_ref_exit(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreetcaf41672013-05-07 16:18:35 -0700803 pr_debug("error allocating ioctx %d\n", err);
Al Viroe23754f2012-03-06 14:33:22 -0500804 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805}
806
Kent Overstreet36f55882013-05-07 16:18:41 -0700807/* kill_ioctx
808 * Cancels all outstanding aio requests on an aio context. Used
809 * when the processes owning a context have all exited to encourage
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 * the rapid destruction of the kioctx.
811 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400812static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
Jens Axboedc48e562015-04-15 11:17:23 -0600813 struct ctx_rq_wait *wait)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814{
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400815 struct kioctx_table *table;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400816
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400817 spin_lock(&mm->ioctx_lock);
Al Virob2edffd2015-04-06 17:48:54 -0400818 if (atomic_xchg(&ctx->dead, 1)) {
819 spin_unlock(&mm->ioctx_lock);
820 return -EINVAL;
821 }
822
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200823 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400824 WARN_ON(ctx != table->table[ctx->id]);
825 table->table[ctx->id] = NULL;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400826 spin_unlock(&mm->ioctx_lock);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700827
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400828 /* percpu_ref_kill() will do the necessary call_rcu() */
829 wake_up_all(&ctx->wait);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700830
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400831 /*
832 * It'd be more correct to do this in free_ioctx(), after all
833 * the outstanding kiocbs have finished - but by then io_destroy
834 * has already returned, so io_setup() could potentially return
835 * -EAGAIN with no ioctxs actually in use (as far as userspace
836 * could tell).
837 */
838 aio_nr_sub(ctx->max_reqs);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400839
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400840 if (ctx->mmap_size)
841 vm_munmap(ctx->mmap_base, ctx->mmap_size);
842
Jens Axboedc48e562015-04-15 11:17:23 -0600843 ctx->rq_wait = wait;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400844 percpu_ref_kill(&ctx->users);
845 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846}
847
Kent Overstreet36f55882013-05-07 16:18:41 -0700848/*
849 * exit_aio: called when the last user of mm goes away. At this point, there is
850 * no way for any new requests to be submited or any of the io_* syscalls to be
851 * called on the context.
852 *
853 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
854 * them.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800856void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857{
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200858 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
Jens Axboedc48e562015-04-15 11:17:23 -0600859 struct ctx_rq_wait wait;
860 int i, skipped;
Jens Axboeabf137d2008-12-09 08:11:22 +0100861
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200862 if (!table)
863 return;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400864
Jens Axboedc48e562015-04-15 11:17:23 -0600865 atomic_set(&wait.count, table->nr);
866 init_completion(&wait.comp);
867
868 skipped = 0;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200869 for (i = 0; i < table->nr; ++i) {
870 struct kioctx *ctx = table->table[i];
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400871
Jens Axboedc48e562015-04-15 11:17:23 -0600872 if (!ctx) {
873 skipped++;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200874 continue;
Jens Axboedc48e562015-04-15 11:17:23 -0600875 }
876
Al Viro936af152012-04-20 21:49:41 -0400877 /*
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200878 * We don't need to bother with munmap() here - exit_mmap(mm)
879 * is coming and it'll unmap everything. And we simply can't,
880 * this is not necessarily our ->mm.
881 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
882 * that it needs to unmap the area, just set it to 0.
Al Viro936af152012-04-20 21:49:41 -0400883 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700884 ctx->mmap_size = 0;
Jens Axboedc48e562015-04-15 11:17:23 -0600885 kill_ioctx(mm, ctx, &wait);
886 }
Kent Overstreet36f55882013-05-07 16:18:41 -0700887
Jens Axboedc48e562015-04-15 11:17:23 -0600888 if (!atomic_sub_and_test(skipped, &wait.count)) {
Gu Zheng6098b452014-09-03 17:45:44 +0800889 /* Wait until all IO for the context are done. */
Jens Axboedc48e562015-04-15 11:17:23 -0600890 wait_for_completion(&wait.comp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 }
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200892
893 RCU_INIT_POINTER(mm->ioctx_table, NULL);
894 kfree(table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895}
896
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000897static void put_reqs_available(struct kioctx *ctx, unsigned nr)
898{
899 struct kioctx_cpu *kcpu;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400900 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000901
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400902 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400903 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000904 kcpu->reqs_available += nr;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400905
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000906 while (kcpu->reqs_available >= ctx->req_batch * 2) {
907 kcpu->reqs_available -= ctx->req_batch;
908 atomic_add(ctx->req_batch, &ctx->reqs_available);
909 }
910
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400911 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000912}
913
914static bool get_reqs_available(struct kioctx *ctx)
915{
916 struct kioctx_cpu *kcpu;
917 bool ret = false;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400918 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000919
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400920 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400921 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000922 if (!kcpu->reqs_available) {
923 int old, avail = atomic_read(&ctx->reqs_available);
924
925 do {
926 if (avail < ctx->req_batch)
927 goto out;
928
929 old = avail;
930 avail = atomic_cmpxchg(&ctx->reqs_available,
931 avail, avail - ctx->req_batch);
932 } while (avail != old);
933
934 kcpu->reqs_available += ctx->req_batch;
935 }
936
937 ret = true;
938 kcpu->reqs_available--;
939out:
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400940 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000941 return ret;
942}
943
Benjamin LaHaised856f322014-08-24 13:14:05 -0400944/* refill_reqs_available
945 * Updates the reqs_available reference counts used for tracking the
946 * number of free slots in the completion ring. This can be called
947 * from aio_complete() (to optimistically update reqs_available) or
948 * from aio_get_req() (the we're out of events case). It must be
949 * called holding ctx->completion_lock.
950 */
951static void refill_reqs_available(struct kioctx *ctx, unsigned head,
952 unsigned tail)
953{
954 unsigned events_in_ring, completed;
955
956 /* Clamp head since userland can write to it. */
957 head %= ctx->nr_events;
958 if (head <= tail)
959 events_in_ring = tail - head;
960 else
961 events_in_ring = ctx->nr_events - (head - tail);
962
963 completed = ctx->completed_events;
964 if (events_in_ring < completed)
965 completed -= events_in_ring;
966 else
967 completed = 0;
968
969 if (!completed)
970 return;
971
972 ctx->completed_events -= completed;
973 put_reqs_available(ctx, completed);
974}
975
976/* user_refill_reqs_available
977 * Called to refill reqs_available when aio_get_req() encounters an
978 * out of space in the completion ring.
979 */
980static void user_refill_reqs_available(struct kioctx *ctx)
981{
982 spin_lock_irq(&ctx->completion_lock);
983 if (ctx->completed_events) {
984 struct aio_ring *ring;
985 unsigned head;
986
987 /* Access of ring->head may race with aio_read_events_ring()
988 * here, but that's okay since whether we read the old version
989 * or the new version, and either will be valid. The important
990 * part is that head cannot pass tail since we prevent
991 * aio_complete() from updating tail by holding
992 * ctx->completion_lock. Even if head is invalid, the check
993 * against ctx->completed_events below will make sure we do the
994 * safe/right thing.
995 */
996 ring = kmap_atomic(ctx->ring_pages[0]);
997 head = ring->head;
998 kunmap_atomic(ring);
999
1000 refill_reqs_available(ctx, head, ctx->tail);
1001 }
1002
1003 spin_unlock_irq(&ctx->completion_lock);
1004}
1005
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006/* aio_get_req
Kent Overstreet57282d82013-05-13 13:42:52 -07001007 * Allocate a slot for an aio request.
1008 * Returns NULL if no requests are free.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001010static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001012 struct aio_kiocb *req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001013
Benjamin LaHaised856f322014-08-24 13:14:05 -04001014 if (!get_reqs_available(ctx)) {
1015 user_refill_reqs_available(ctx);
1016 if (!get_reqs_available(ctx))
1017 return NULL;
1018 }
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001019
Kent Overstreet0460fef2013-05-07 16:18:49 -07001020 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 if (unlikely(!req))
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001022 goto out_put;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023
Kent Overstreete34ecee2013-10-10 19:31:47 -07001024 percpu_ref_get(&ctx->reqs);
1025
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026 req->ki_ctx = ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 return req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001028out_put:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001029 put_reqs_available(ctx, 1);
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001030 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031}
1032
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001033static void kiocb_free(struct aio_kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001035 if (req->common.ki_filp)
1036 fput(req->common.ki_filp);
Davide Libenzi13389012009-06-30 11:41:11 -07001037 if (req->ki_eventfd != NULL)
1038 eventfd_ctx_put(req->ki_eventfd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039 kmem_cache_free(kiocb_cachep, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040}
1041
Adrian Bunkd5470b52008-04-29 00:58:57 -07001042static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001044 struct aio_ring __user *ring = (void __user *)ctx_id;
Jens Axboeabf137d2008-12-09 08:11:22 +01001045 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -07001046 struct kioctx *ctx, *ret = NULL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001047 struct kioctx_table *table;
1048 unsigned id;
1049
1050 if (get_user(id, &ring->id))
1051 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052
Jens Axboeabf137d2008-12-09 08:11:22 +01001053 rcu_read_lock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001054 table = rcu_dereference(mm->ioctx_table);
Jens Axboeabf137d2008-12-09 08:11:22 +01001055
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001056 if (!table || id >= table->nr)
1057 goto out;
1058
1059 ctx = table->table[id];
Benjamin LaHaisef30d7042013-08-07 18:23:48 -04001060 if (ctx && ctx->user_id == ctx_id) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001061 percpu_ref_get(&ctx->users);
1062 ret = ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +01001063 }
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001064out:
Jens Axboeabf137d2008-12-09 08:11:22 +01001065 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -07001066 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067}
1068
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069/* aio_complete
1070 * Called when the io request on the given iocb is complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001072static void aio_complete(struct kiocb *kiocb, long res, long res2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001074 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 struct kioctx *ctx = iocb->ki_ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 struct aio_ring *ring;
Kent Overstreet21b40202013-05-07 16:18:47 -07001077 struct io_event *ev_page, *event;
Benjamin LaHaised856f322014-08-24 13:14:05 -04001078 unsigned tail, pos, head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080
Zach Brown20dcae32005-11-13 16:07:33 -08001081 /*
1082 * Special case handling for sync iocbs:
1083 * - events go directly into the iocb for fast handling
1084 * - the sync task with the iocb in its stack holds the single iocb
1085 * ref, no other paths have a way to get another ref
1086 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001088 BUG_ON(is_sync_kiocb(kiocb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089
Kent Overstreet0460fef2013-05-07 16:18:49 -07001090 if (iocb->ki_list.next) {
1091 unsigned long flags;
1092
1093 spin_lock_irqsave(&ctx->ctx_lock, flags);
1094 list_del(&iocb->ki_list);
1095 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1096 }
Kent Overstreet11599eb2013-05-07 16:18:39 -07001097
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 /*
Kent Overstreet0460fef2013-05-07 16:18:49 -07001099 * Add a completion event to the ring buffer. Must be done holding
Tang Chen4b30f072013-07-03 15:09:16 -07001100 * ctx->completion_lock to prevent other code from messing with the tail
Kent Overstreet0460fef2013-05-07 16:18:49 -07001101 * pointer since we might be called from irq context.
1102 */
1103 spin_lock_irqsave(&ctx->completion_lock, flags);
1104
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001105 tail = ctx->tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001106 pos = tail + AIO_EVENTS_OFFSET;
1107
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001108 if (++tail >= ctx->nr_events)
Ken Chen4bf69b22005-05-01 08:59:15 -07001109 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001111 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001112 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1113
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001114 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 event->data = iocb->ki_user_data;
1116 event->res = res;
1117 event->res2 = res2;
1118
Kent Overstreet21b40202013-05-07 16:18:47 -07001119 kunmap_atomic(ev_page);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001120 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001121
1122 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001123 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
Kent Overstreetcaf41672013-05-07 16:18:35 -07001124 res, res2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125
1126 /* after flagging the request as done, we
1127 * must never even look at it again
1128 */
1129 smp_wmb(); /* make event visible before updating tail */
1130
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001131 ctx->tail = tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001132
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001133 ring = kmap_atomic(ctx->ring_pages[0]);
Benjamin LaHaised856f322014-08-24 13:14:05 -04001134 head = ring->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 ring->tail = tail;
Cong Wange8e3c3d2011-11-25 23:14:27 +08001136 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001137 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138
Benjamin LaHaised856f322014-08-24 13:14:05 -04001139 ctx->completed_events++;
1140 if (ctx->completed_events > 1)
1141 refill_reqs_available(ctx, head, tail);
Kent Overstreet0460fef2013-05-07 16:18:49 -07001142 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1143
Kent Overstreet21b40202013-05-07 16:18:47 -07001144 pr_debug("added to ring %p at [%u]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001145
1146 /*
1147 * Check if the user asked us to deliver the result through an
1148 * eventfd. The eventfd_signal() function is safe to be called
1149 * from IRQ context.
1150 */
Davide Libenzi87c3a862009-03-18 17:04:19 -07001151 if (iocb->ki_eventfd != NULL)
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001152 eventfd_signal(iocb->ki_eventfd, 1);
1153
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 /* everything turned out well, dispose of the aiocb. */
Kent Overstreet57282d82013-05-13 13:42:52 -07001155 kiocb_free(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156
Quentin Barnes6cb2a212008-03-19 17:00:39 -07001157 /*
1158 * We have to order our ring_info tail store above and test
1159 * of the wait list below outside the wait lock. This is
1160 * like in wake_up_bit() where clearing a bit has to be
1161 * ordered with the unlocked test.
1162 */
1163 smp_mb();
1164
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165 if (waitqueue_active(&ctx->wait))
1166 wake_up(&ctx->wait);
1167
Kent Overstreete34ecee2013-10-10 19:31:47 -07001168 percpu_ref_put(&ctx->reqs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169}
1170
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001171/* aio_read_events_ring
Kent Overstreeta31ad382013-05-07 16:18:45 -07001172 * Pull an event off of the ioctx's event ring. Returns the number of
1173 * events fetched
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 */
Kent Overstreeta31ad382013-05-07 16:18:45 -07001175static long aio_read_events_ring(struct kioctx *ctx,
1176 struct io_event __user *event, long nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178 struct aio_ring *ring;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001179 unsigned head, tail, pos;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001180 long ret = 0;
1181 int copy_ret;
1182
Dave Chinner9c9ce762015-02-03 19:29:05 -05001183 /*
1184 * The mutex can block and wake us up and that will cause
1185 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1186 * and repeat. This should be rare enough that it doesn't cause
1187 * peformance issues. See the comment in read_events() for more detail.
1188 */
1189 sched_annotate_sleep();
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001190 mutex_lock(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -04001192 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001193 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001194 head = ring->head;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001195 tail = ring->tail;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001196 kunmap_atomic(ring);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197
Jeff Moyer2ff396be2014-09-02 13:17:00 -04001198 /*
1199 * Ensure that once we've read the current tail pointer, that
1200 * we also see the events that were stored up to the tail.
1201 */
1202 smp_rmb();
1203
Kent Overstreet5ffac122013-05-09 15:36:07 -07001204 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001205
Kent Overstreet5ffac122013-05-09 15:36:07 -07001206 if (head == tail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207 goto out;
1208
Benjamin LaHaiseedfbbf32014-06-24 13:32:51 -04001209 head %= ctx->nr_events;
1210 tail %= ctx->nr_events;
1211
Kent Overstreeta31ad382013-05-07 16:18:45 -07001212 while (ret < nr) {
1213 long avail;
1214 struct io_event *ev;
1215 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216
Kent Overstreet5ffac122013-05-09 15:36:07 -07001217 avail = (head <= tail ? tail : ctx->nr_events) - head;
1218 if (head == tail)
Kent Overstreeta31ad382013-05-07 16:18:45 -07001219 break;
1220
1221 avail = min(avail, nr - ret);
1222 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1223 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1224
1225 pos = head + AIO_EVENTS_OFFSET;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001226 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
Kent Overstreeta31ad382013-05-07 16:18:45 -07001227 pos %= AIO_EVENTS_PER_PAGE;
1228
1229 ev = kmap(page);
1230 copy_ret = copy_to_user(event + ret, ev + pos,
1231 sizeof(*ev) * avail);
1232 kunmap(page);
1233
1234 if (unlikely(copy_ret)) {
1235 ret = -EFAULT;
1236 goto out;
1237 }
1238
1239 ret += avail;
1240 head += avail;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001241 head %= ctx->nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001244 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001245 ring->head = head;
Zhao Hongjiang91d80a82013-04-26 11:03:53 +08001246 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001247 flush_dcache_page(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001248
Kent Overstreet5ffac122013-05-09 15:36:07 -07001249 pr_debug("%li h%u t%u\n", ret, head, tail);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001250out:
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001251 mutex_unlock(&ctx->ring_lock);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001252
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 return ret;
1254}
1255
Kent Overstreeta31ad382013-05-07 16:18:45 -07001256static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1257 struct io_event __user *event, long *i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001259 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260
Kent Overstreeta31ad382013-05-07 16:18:45 -07001261 if (ret > 0)
1262 *i += ret;
1263
1264 if (unlikely(atomic_read(&ctx->dead)))
1265 ret = -EINVAL;
1266
1267 if (!*i)
1268 *i = ret;
1269
1270 return ret < 0 || *i >= min_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271}
1272
Kent Overstreeta31ad382013-05-07 16:18:45 -07001273static long read_events(struct kioctx *ctx, long min_nr, long nr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274 struct io_event __user *event,
1275 struct timespec __user *timeout)
1276{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001277 ktime_t until = { .tv64 = KTIME_MAX };
1278 long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 if (timeout) {
1281 struct timespec ts;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001282
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
Kent Overstreeta31ad382013-05-07 16:18:45 -07001284 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285
Kent Overstreeta31ad382013-05-07 16:18:45 -07001286 until = timespec_to_ktime(ts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 }
1288
Kent Overstreeta31ad382013-05-07 16:18:45 -07001289 /*
1290 * Note that aio_read_events() is being called as the conditional - i.e.
1291 * we're calling it after prepare_to_wait() has set task state to
1292 * TASK_INTERRUPTIBLE.
1293 *
1294 * But aio_read_events() can block, and if it blocks it's going to flip
1295 * the task state back to TASK_RUNNING.
1296 *
1297 * This should be ok, provided it doesn't flip the state back to
1298 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1299 * will only happen if the mutex_lock() call blocks, and we then find
1300 * the ringbuffer empty. So in practice we should be ok, but it's
1301 * something to be aware of when touching this code.
1302 */
Fam Zheng5f785de2014-11-06 20:44:36 +08001303 if (until.tv64 == 0)
1304 aio_read_events(ctx, min_nr, nr, event, &ret);
1305 else
1306 wait_event_interruptible_hrtimeout(ctx->wait,
1307 aio_read_events(ctx, min_nr, nr, event, &ret),
1308 until);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309
Kent Overstreeta31ad382013-05-07 16:18:45 -07001310 if (!ret && signal_pending(current))
1311 ret = -EINTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312
Kent Overstreeta31ad382013-05-07 16:18:45 -07001313 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314}
1315
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316/* sys_io_setup:
1317 * Create an aio_context capable of receiving at least nr_events.
1318 * ctxp must not point to an aio_context that already exists, and
1319 * must be initialized to 0 prior to the call. On successful
1320 * creation of the aio_context, *ctxp is filled in with the resulting
1321 * handle. May fail with -EINVAL if *ctxp is not initialized,
1322 * if the specified nr_events exceeds internal limits. May fail
1323 * with -EAGAIN if the specified nr_events exceeds the user's limit
1324 * of available events. May fail with -ENOMEM if insufficient kernel
1325 * resources are available. May fail with -EFAULT if an invalid
1326 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1327 * implemented.
1328 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001329SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330{
1331 struct kioctx *ioctx = NULL;
1332 unsigned long ctx;
1333 long ret;
1334
1335 ret = get_user(ctx, ctxp);
1336 if (unlikely(ret))
1337 goto out;
1338
1339 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001340 if (unlikely(ctx || nr_events == 0)) {
Kinglong Meeacd88d42015-02-04 21:15:59 +08001341 pr_debug("EINVAL: ctx %lu nr_events %u\n",
Zach Brownd55b5fd2005-11-07 00:59:31 -08001342 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 goto out;
1344 }
1345
1346 ioctx = ioctx_alloc(nr_events);
1347 ret = PTR_ERR(ioctx);
1348 if (!IS_ERR(ioctx)) {
1349 ret = put_user(ioctx->user_id, ctxp);
Al Viroa2e18592012-03-20 16:27:57 -04001350 if (ret)
Anatol Pomozove02ba722014-04-15 11:31:33 -07001351 kill_ioctx(current->mm, ioctx, NULL);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001352 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 }
1354
1355out:
1356 return ret;
1357}
1358
1359/* sys_io_destroy:
1360 * Destroy the aio_context specified. May cancel any outstanding
1361 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001362 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363 * is invalid.
1364 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001365SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366{
1367 struct kioctx *ioctx = lookup_ioctx(ctx);
1368 if (likely(NULL != ioctx)) {
Jens Axboedc48e562015-04-15 11:17:23 -06001369 struct ctx_rq_wait wait;
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001370 int ret;
Anatol Pomozove02ba722014-04-15 11:31:33 -07001371
Jens Axboedc48e562015-04-15 11:17:23 -06001372 init_completion(&wait.comp);
1373 atomic_set(&wait.count, 1);
1374
Anatol Pomozove02ba722014-04-15 11:31:33 -07001375 /* Pass requests_done to kill_ioctx() where it can be set
1376 * in a thread-safe way. If we try to set it here then we have
1377 * a race condition if two io_destroy() called simultaneously.
1378 */
Jens Axboedc48e562015-04-15 11:17:23 -06001379 ret = kill_ioctx(current->mm, ioctx, &wait);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001380 percpu_ref_put(&ioctx->users);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001381
1382 /* Wait until all IO for the context are done. Otherwise kernel
1383 * keep using user-space buffers even if user thinks the context
1384 * is destroyed.
1385 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001386 if (!ret)
Jens Axboedc48e562015-04-15 11:17:23 -06001387 wait_for_completion(&wait.comp);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001388
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001389 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 }
Kinglong Meeacd88d42015-02-04 21:15:59 +08001391 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 return -EINVAL;
1393}
1394
Christoph Hellwig89319d312016-10-30 11:42:03 -05001395static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec,
1396 bool vectored, bool compat, struct iov_iter *iter)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001397{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001398 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1399 size_t len = iocb->aio_nbytes;
1400
1401 if (!vectored) {
1402 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1403 *iovec = NULL;
1404 return ret;
1405 }
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001406#ifdef CONFIG_COMPAT
1407 if (compat)
Christoph Hellwig89319d312016-10-30 11:42:03 -05001408 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1409 iter);
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001410#endif
Christoph Hellwig89319d312016-10-30 11:42:03 -05001411 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001412}
1413
Christoph Hellwig89319d312016-10-30 11:42:03 -05001414static inline ssize_t aio_ret(struct kiocb *req, ssize_t ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001416 switch (ret) {
1417 case -EIOCBQUEUED:
1418 return ret;
1419 case -ERESTARTSYS:
1420 case -ERESTARTNOINTR:
1421 case -ERESTARTNOHAND:
1422 case -ERESTART_RESTARTBLOCK:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001423 /*
1424 * There's no easy way to restart the syscall since other AIO's
1425 * may be already running. Just fail this IO with EINTR.
1426 */
Christoph Hellwig89319d312016-10-30 11:42:03 -05001427 ret = -EINTR;
1428 /*FALLTHRU*/
1429 default:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001430 aio_complete(req, ret, 0);
Christoph Hellwig89319d312016-10-30 11:42:03 -05001431 return 0;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001432 }
Christoph Hellwig89319d312016-10-30 11:42:03 -05001433}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
Christoph Hellwig89319d312016-10-30 11:42:03 -05001435static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
1436 bool compat)
1437{
1438 struct file *file = req->ki_filp;
1439 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1440 struct iov_iter iter;
1441 ssize_t ret;
1442
1443 if (unlikely(!(file->f_mode & FMODE_READ)))
1444 return -EBADF;
1445 if (unlikely(!file->f_op->read_iter))
1446 return -EINVAL;
1447
1448 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1449 if (ret)
1450 return ret;
1451 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1452 if (!ret)
1453 ret = aio_ret(req, file->f_op->read_iter(req, &iter));
1454 kfree(iovec);
1455 return ret;
1456}
1457
1458static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored,
1459 bool compat)
1460{
1461 struct file *file = req->ki_filp;
1462 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1463 struct iov_iter iter;
1464 ssize_t ret;
1465
1466 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1467 return -EBADF;
1468 if (unlikely(!file->f_op->write_iter))
1469 return -EINVAL;
1470
1471 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1472 if (ret)
1473 return ret;
1474 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1475 if (!ret) {
1476 file_start_write(file);
1477 ret = aio_ret(req, file->f_op->write_iter(req, &iter));
1478 file_end_write(file);
1479 }
1480 kfree(iovec);
1481 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482}
1483
Adrian Bunkd5470b52008-04-29 00:58:57 -07001484static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001485 struct iocb *iocb, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001487 struct aio_kiocb *req;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001488 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 ssize_t ret;
1490
1491 /* enforce forwards compatibility on users */
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001492 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001493 pr_debug("EINVAL: reserve field set\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 return -EINVAL;
1495 }
1496
1497 /* prevent overflows */
1498 if (unlikely(
1499 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1500 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1501 ((ssize_t)iocb->aio_nbytes < 0)
1502 )) {
Kinglong Meeacd88d42015-02-04 21:15:59 +08001503 pr_debug("EINVAL: overflow check\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504 return -EINVAL;
1505 }
1506
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001507 req = aio_get_req(ctx);
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001508 if (unlikely(!req))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 return -EAGAIN;
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001510
Christoph Hellwig89319d312016-10-30 11:42:03 -05001511 req->common.ki_filp = file = fget(iocb->aio_fildes);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001512 if (unlikely(!req->common.ki_filp)) {
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001513 ret = -EBADF;
1514 goto out_put_req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515 }
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001516 req->common.ki_pos = iocb->aio_offset;
1517 req->common.ki_complete = aio_complete;
Al Viro2ba48ce2015-04-09 13:52:01 -04001518 req->common.ki_flags = iocb_flags(req->common.ki_filp);
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001519
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001520 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1521 /*
1522 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1523 * instance of the file* now. The file descriptor must be
1524 * an eventfd() fd, and will be signaled for each completed
1525 * event using the eventfd_signal() function.
1526 */
Davide Libenzi13389012009-06-30 11:41:11 -07001527 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
Hirofumi Nakagawa801678c2008-04-29 01:03:09 -07001528 if (IS_ERR(req->ki_eventfd)) {
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001529 ret = PTR_ERR(req->ki_eventfd);
Davide Libenzi87c3a862009-03-18 17:04:19 -07001530 req->ki_eventfd = NULL;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001531 goto out_put_req;
1532 }
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001533
1534 req->common.ki_flags |= IOCB_EVENTFD;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001535 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536
Kent Overstreet8a660892013-05-07 16:19:10 -07001537 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538 if (unlikely(ret)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001539 pr_debug("EFAULT: aio_key\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 goto out_put_req;
1541 }
1542
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001543 req->ki_user_iocb = user_iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 req->ki_user_data = iocb->aio_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545
Christoph Hellwig89319d312016-10-30 11:42:03 -05001546 get_file(file);
1547 switch (iocb->aio_lio_opcode) {
1548 case IOCB_CMD_PREAD:
1549 ret = aio_read(&req->common, iocb, false, compat);
1550 break;
1551 case IOCB_CMD_PWRITE:
1552 ret = aio_write(&req->common, iocb, false, compat);
1553 break;
1554 case IOCB_CMD_PREADV:
1555 ret = aio_read(&req->common, iocb, true, compat);
1556 break;
1557 case IOCB_CMD_PWRITEV:
1558 ret = aio_write(&req->common, iocb, true, compat);
1559 break;
1560 default:
1561 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1562 ret = -EINVAL;
1563 break;
1564 }
1565 fput(file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566
Christoph Hellwig89319d312016-10-30 11:42:03 -05001567 if (ret && ret != -EIOCBQUEUED)
1568 goto out_put_req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570out_put_req:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001571 put_reqs_available(ctx, 1);
Kent Overstreete34ecee2013-10-10 19:31:47 -07001572 percpu_ref_put(&ctx->reqs);
Kent Overstreet57282d82013-05-13 13:42:52 -07001573 kiocb_free(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574 return ret;
1575}
1576
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001577long do_io_submit(aio_context_t ctx_id, long nr,
1578 struct iocb __user *__user *iocbpp, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579{
1580 struct kioctx *ctx;
1581 long ret = 0;
Jeff Moyer080d6762011-11-02 13:40:10 -07001582 int i = 0;
Shaohua Li9f5b9422010-07-01 07:55:01 +02001583 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584
1585 if (unlikely(nr < 0))
1586 return -EINVAL;
1587
Jeff Moyer75e1c702010-09-10 14:16:00 -07001588 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1589 nr = LONG_MAX/sizeof(*iocbpp);
1590
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1592 return -EFAULT;
1593
1594 ctx = lookup_ioctx(ctx_id);
1595 if (unlikely(!ctx)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001596 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597 return -EINVAL;
1598 }
1599
Shaohua Li9f5b9422010-07-01 07:55:01 +02001600 blk_start_plug(&plug);
1601
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 /*
1603 * AKPM: should this return a partial result if some of the IOs were
1604 * successfully submitted?
1605 */
1606 for (i=0; i<nr; i++) {
1607 struct iocb __user *user_iocb;
1608 struct iocb tmp;
1609
1610 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1611 ret = -EFAULT;
1612 break;
1613 }
1614
1615 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1616 ret = -EFAULT;
1617 break;
1618 }
1619
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001620 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 if (ret)
1622 break;
1623 }
Shaohua Li9f5b9422010-07-01 07:55:01 +02001624 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625
Kent Overstreet723be6e2013-05-28 15:14:48 -07001626 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627 return i ? i : ret;
1628}
1629
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001630/* sys_io_submit:
1631 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1632 * the number of iocbs queued. May return -EINVAL if the aio_context
1633 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1634 * *iocbpp[0] is not properly initialized, if the operation specified
1635 * is invalid for the file descriptor in the iocb. May fail with
1636 * -EFAULT if any of the data structures point to invalid data. May
1637 * fail with -EBADF if the file descriptor specified in the first
1638 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1639 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1640 * fail with -ENOSYS if not implemented.
1641 */
1642SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1643 struct iocb __user * __user *, iocbpp)
1644{
1645 return do_io_submit(ctx_id, nr, iocbpp, 0);
1646}
1647
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648/* lookup_kiocb
1649 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001651static struct aio_kiocb *
1652lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001654 struct aio_kiocb *kiocb;
Zach Brownd00689a2005-11-13 16:07:34 -08001655
1656 assert_spin_locked(&ctx->ctx_lock);
1657
Kent Overstreet8a660892013-05-07 16:19:10 -07001658 if (key != KIOCB_KEY)
1659 return NULL;
1660
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 /* TODO: use a hash or array, this sucks. */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001662 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1663 if (kiocb->ki_user_iocb == iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664 return kiocb;
1665 }
1666 return NULL;
1667}
1668
1669/* sys_io_cancel:
1670 * Attempts to cancel an iocb previously passed to io_submit. If
1671 * the operation is successfully cancelled, the resulting event is
1672 * copied into the memory pointed to by result without being placed
1673 * into the completion queue and 0 is returned. May fail with
1674 * -EFAULT if any of the data structures pointed to are invalid.
1675 * May fail with -EINVAL if aio_context specified by ctx_id is
1676 * invalid. May fail with -EAGAIN if the iocb specified was not
1677 * cancelled. Will fail with -ENOSYS if not implemented.
1678 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001679SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1680 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 struct kioctx *ctx;
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001683 struct aio_kiocb *kiocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 u32 key;
1685 int ret;
1686
1687 ret = get_user(key, &iocb->aio_key);
1688 if (unlikely(ret))
1689 return -EFAULT;
1690
1691 ctx = lookup_ioctx(ctx_id);
1692 if (unlikely(!ctx))
1693 return -EINVAL;
1694
1695 spin_lock_irq(&ctx->ctx_lock);
Kent Overstreet906b9732013-05-07 16:18:31 -07001696
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 kiocb = lookup_kiocb(ctx, iocb, key);
Kent Overstreet906b9732013-05-07 16:18:31 -07001698 if (kiocb)
Fabian Frederickd52a8f92014-04-22 07:26:58 +02001699 ret = kiocb_cancel(kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -07001700 else
1701 ret = -EINVAL;
1702
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 spin_unlock_irq(&ctx->ctx_lock);
1704
Kent Overstreet906b9732013-05-07 16:18:31 -07001705 if (!ret) {
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001706 /*
1707 * The result argument is no longer used - the io_event is
1708 * always delivered via the ring buffer. -EINPROGRESS indicates
1709 * cancellation is progress:
Kent Overstreet906b9732013-05-07 16:18:31 -07001710 */
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001711 ret = -EINPROGRESS;
Kent Overstreet906b9732013-05-07 16:18:31 -07001712 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713
Kent Overstreet723be6e2013-05-28 15:14:48 -07001714 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715
1716 return ret;
1717}
1718
1719/* io_getevents:
1720 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001721 * the completion queue for the aio_context specified by ctx_id. If
1722 * it succeeds, the number of read events is returned. May fail with
1723 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1724 * out of range, if timeout is out of range. May fail with -EFAULT
1725 * if any of the memory specified is invalid. May return 0 or
1726 * < min_nr if the timeout specified by timeout has elapsed
1727 * before sufficient events are available, where timeout == NULL
1728 * specifies an infinite timeout. Note that the timeout pointed to by
Jeff Moyer69008072013-05-24 15:55:24 -07001729 * timeout is relative. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001731SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1732 long, min_nr,
1733 long, nr,
1734 struct io_event __user *, events,
1735 struct timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736{
1737 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1738 long ret = -EINVAL;
1739
1740 if (likely(ioctx)) {
Namhyung Kim2e410252011-01-12 17:01:08 -08001741 if (likely(min_nr <= nr && min_nr >= 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 ret = read_events(ioctx, min_nr, nr, events, timeout);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001743 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 return ret;
1746}