blob: 118a2e0088d8fdd8391654a44edb06157dad5629 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 *
5 * Implements an efficient asynchronous io interface.
6 *
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 *
9 * See ../COPYING for licensing terms.
10 */
Kent Overstreetcaf41672013-05-07 16:18:35 -070011#define pr_fmt(fmt) "%s: " fmt, __func__
12
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/kernel.h>
14#include <linux/init.h>
15#include <linux/errno.h>
16#include <linux/time.h>
17#include <linux/aio_abi.h>
Paul Gortmaker630d9c42011-11-16 23:57:37 -050018#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/syscalls.h>
Jens Axboeb9d128f2009-10-29 13:59:26 +010020#include <linux/backing-dev.h>
Badari Pulavarty027445c2006-09-30 23:28:46 -070021#include <linux/uio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070022
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/sched.h>
24#include <linux/fs.h>
25#include <linux/file.h>
26#include <linux/mm.h>
27#include <linux/mman.h>
Michael S. Tsirkin3d2d8272009-09-21 17:03:51 -070028#include <linux/mmu_context.h>
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100029#include <linux/percpu.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#include <linux/slab.h>
31#include <linux/timer.h>
32#include <linux/aio.h>
33#include <linux/highmem.h>
34#include <linux/workqueue.h>
35#include <linux/security.h>
Davide Libenzi9c3060b2007-05-10 22:23:21 -070036#include <linux/eventfd.h>
Jeff Moyercfb1e332009-10-02 18:57:36 -040037#include <linux/blkdev.h>
Jeff Moyer9d85cba2010-05-26 14:44:26 -070038#include <linux/compat.h>
Gu Zheng36bc08c2013-07-16 17:56:16 +080039#include <linux/migrate.h>
40#include <linux/ramfs.h>
Kent Overstreet723be6e2013-05-28 15:14:48 -070041#include <linux/percpu-refcount.h>
Benjamin LaHaise71ad7492013-09-17 10:18:25 -040042#include <linux/mount.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44#include <asm/kmap_types.h>
45#include <asm/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
Al Viro68d70d02013-06-19 15:26:04 +040047#include "internal.h"
48
Kent Overstreet4e179bc2013-05-07 16:18:33 -070049#define AIO_RING_MAGIC 0xa10a10a1
50#define AIO_RING_COMPAT_FEATURES 1
51#define AIO_RING_INCOMPAT_FEATURES 0
52struct aio_ring {
53 unsigned id; /* kernel internal index number */
54 unsigned nr; /* number of io_events */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -040055 unsigned head; /* Written to by userland or under ring_lock
56 * mutex by aio_read_events_ring(). */
Kent Overstreet4e179bc2013-05-07 16:18:33 -070057 unsigned tail;
58
59 unsigned magic;
60 unsigned compat_features;
61 unsigned incompat_features;
62 unsigned header_length; /* size of aio_ring */
63
64
65 struct io_event io_events[0];
66}; /* 128 bytes + ring size */
67
68#define AIO_RING_PAGES 8
Kent Overstreet4e179bc2013-05-07 16:18:33 -070069
Benjamin LaHaisedb446a02013-07-30 12:54:40 -040070struct kioctx_table {
71 struct rcu_head rcu;
72 unsigned nr;
73 struct kioctx *table[];
74};
75
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100076struct kioctx_cpu {
77 unsigned reqs_available;
78};
79
Kent Overstreet4e179bc2013-05-07 16:18:33 -070080struct kioctx {
Kent Overstreet723be6e2013-05-28 15:14:48 -070081 struct percpu_ref users;
Kent Overstreet36f55882013-05-07 16:18:41 -070082 atomic_t dead;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070083
Kent Overstreete34ecee2013-10-10 19:31:47 -070084 struct percpu_ref reqs;
85
Kent Overstreet4e179bc2013-05-07 16:18:33 -070086 unsigned long user_id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070087
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100088 struct __percpu kioctx_cpu *cpu;
89
90 /*
91 * For percpu reqs_available, number of slots we move to/from global
92 * counter at a time:
93 */
94 unsigned req_batch;
Kent Overstreet3e845ce2013-05-07 16:18:51 -070095 /*
96 * This is what userspace passed to io_setup(), it's not used for
97 * anything but counting against the global max_reqs quota.
98 *
Kent Overstreet58c85dc2013-05-07 16:18:55 -070099 * The real limit is nr_events - 1, which will be larger (see
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700100 * aio_setup_ring())
101 */
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700102 unsigned max_reqs;
103
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700104 /* Size of ringbuffer, in units of struct io_event */
105 unsigned nr_events;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700106
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700107 unsigned long mmap_base;
108 unsigned long mmap_size;
109
110 struct page **ring_pages;
111 long nr_pages;
112
Kent Overstreet723be6e2013-05-28 15:14:48 -0700113 struct work_struct free_work;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700114
Anatol Pomozove02ba722014-04-15 11:31:33 -0700115 /*
116 * signals when all in-flight requests are done
117 */
118 struct completion *requests_done;
119
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700120 struct {
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000121 /*
122 * This counts the number of available slots in the ringbuffer,
123 * so we avoid overflowing it: it's decremented (if positive)
124 * when allocating a kiocb and incremented when the resulting
125 * io_event is pulled off the ringbuffer.
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000126 *
127 * We batch accesses to it with a percpu version.
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000128 */
129 atomic_t reqs_available;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700130 } ____cacheline_aligned_in_smp;
131
132 struct {
133 spinlock_t ctx_lock;
134 struct list_head active_reqs; /* used for cancellation */
135 } ____cacheline_aligned_in_smp;
136
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700137 struct {
138 struct mutex ring_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700139 wait_queue_head_t wait;
140 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700141
142 struct {
143 unsigned tail;
Benjamin LaHaised856f322014-08-24 13:14:05 -0400144 unsigned completed_events;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700145 spinlock_t completion_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700146 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700147
148 struct page *internal_pages[AIO_RING_PAGES];
Gu Zheng36bc08c2013-07-16 17:56:16 +0800149 struct file *aio_ring_file;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400150
151 unsigned id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700152};
153
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -0800155static DEFINE_SPINLOCK(aio_nr_lock);
156unsigned long aio_nr; /* current system wide number of aio requests */
157unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158/*----end sysctl variables---*/
159
Christoph Lametere18b8902006-12-06 20:33:20 -0800160static struct kmem_cache *kiocb_cachep;
161static struct kmem_cache *kioctx_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400163static struct vfsmount *aio_mnt;
164
165static const struct file_operations aio_ring_fops;
166static const struct address_space_operations aio_ctx_aops;
167
168static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
169{
170 struct qstr this = QSTR_INIT("[aio]", 5);
171 struct file *file;
172 struct path path;
173 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
Dan Carpenter7f626562013-11-13 10:49:40 +0300174 if (IS_ERR(inode))
175 return ERR_CAST(inode);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400176
177 inode->i_mapping->a_ops = &aio_ctx_aops;
178 inode->i_mapping->private_data = ctx;
179 inode->i_size = PAGE_SIZE * nr_pages;
180
181 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
182 if (!path.dentry) {
183 iput(inode);
184 return ERR_PTR(-ENOMEM);
185 }
186 path.mnt = mntget(aio_mnt);
187
188 d_instantiate(path.dentry, inode);
189 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
190 if (IS_ERR(file)) {
191 path_put(&path);
192 return file;
193 }
194
195 file->f_flags = O_RDWR;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400196 return file;
197}
198
199static struct dentry *aio_mount(struct file_system_type *fs_type,
200 int flags, const char *dev_name, void *data)
201{
202 static const struct dentry_operations ops = {
203 .d_dname = simple_dname,
204 };
Gu Zheng8dc43792014-07-23 18:03:52 +0800205 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400206}
207
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208/* aio_setup
209 * Creates the slab caches used by the aio routines, panic on
210 * failure as this is done early during the boot sequence.
211 */
212static int __init aio_setup(void)
213{
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400214 static struct file_system_type aio_fs = {
215 .name = "aio",
216 .mount = aio_mount,
217 .kill_sb = kill_anon_super,
218 };
219 aio_mnt = kern_mount(&aio_fs);
220 if (IS_ERR(aio_mnt))
221 panic("Failed to create aio fs mount.");
222
Christoph Lameter0a31bd52007-05-06 14:49:57 -0700223 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
224 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225
Kent Overstreetcaf41672013-05-07 16:18:35 -0700226 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227
228 return 0;
229}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700230__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400232static void put_aio_ring_file(struct kioctx *ctx)
233{
234 struct file *aio_ring_file = ctx->aio_ring_file;
235 if (aio_ring_file) {
236 truncate_setsize(aio_ring_file->f_inode, 0);
237
238 /* Prevent further access to the kioctx from migratepages */
239 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
240 aio_ring_file->f_inode->i_mapping->private_data = NULL;
241 ctx->aio_ring_file = NULL;
242 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
243
244 fput(aio_ring_file);
245 }
246}
247
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248static void aio_free_ring(struct kioctx *ctx)
249{
Gu Zheng36bc08c2013-07-16 17:56:16 +0800250 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400252 /* Disconnect the kiotx from the ring file. This prevents future
253 * accesses to the kioctx from page migration.
254 */
255 put_aio_ring_file(ctx);
256
Gu Zheng36bc08c2013-07-16 17:56:16 +0800257 for (i = 0; i < ctx->nr_pages; i++) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500258 struct page *page;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800259 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
260 page_count(ctx->ring_pages[i]));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500261 page = ctx->ring_pages[i];
262 if (!page)
263 continue;
264 ctx->ring_pages[i] = NULL;
265 put_page(page);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800266 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267
Sasha Levinddb8c452013-11-19 17:33:03 -0500268 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700269 kfree(ctx->ring_pages);
Sasha Levinddb8c452013-11-19 17:33:03 -0500270 ctx->ring_pages = NULL;
271 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272}
273
Gu Zheng36bc08c2013-07-16 17:56:16 +0800274static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
275{
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400276 vma->vm_flags |= VM_DONTEXPAND;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800277 vma->vm_ops = &generic_file_vm_ops;
278 return 0;
279}
280
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400281static void aio_ring_remap(struct file *file, struct vm_area_struct *vma)
282{
283 struct mm_struct *mm = vma->vm_mm;
284 struct kioctx_table *table;
285 int i;
286
287 spin_lock(&mm->ioctx_lock);
288 rcu_read_lock();
289 table = rcu_dereference(mm->ioctx_table);
290 for (i = 0; i < table->nr; i++) {
291 struct kioctx *ctx;
292
293 ctx = table->table[i];
294 if (ctx && ctx->aio_ring_file == file) {
295 ctx->user_id = ctx->mmap_base = vma->vm_start;
296 break;
297 }
298 }
299
300 rcu_read_unlock();
301 spin_unlock(&mm->ioctx_lock);
302}
303
Gu Zheng36bc08c2013-07-16 17:56:16 +0800304static const struct file_operations aio_ring_fops = {
305 .mmap = aio_ring_mmap,
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400306 .mremap = aio_ring_remap,
Gu Zheng36bc08c2013-07-16 17:56:16 +0800307};
308
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400309#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800310static int aio_migratepage(struct address_space *mapping, struct page *new,
311 struct page *old, enum migrate_mode mode)
312{
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400313 struct kioctx *ctx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800314 unsigned long flags;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400315 pgoff_t idx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800316 int rc;
317
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500318 rc = 0;
319
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400320 /* mapping->private_lock here protects against the kioctx teardown. */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500321 spin_lock(&mapping->private_lock);
322 ctx = mapping->private_data;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400323 if (!ctx) {
324 rc = -EINVAL;
325 goto out;
326 }
327
328 /* The ring_lock mutex. The prevents aio_read_events() from writing
329 * to the ring's head, and prevents page migration from mucking in
330 * a partially initialized kiotx.
331 */
332 if (!mutex_trylock(&ctx->ring_lock)) {
333 rc = -EAGAIN;
334 goto out;
335 }
336
337 idx = old->index;
338 if (idx < (pgoff_t)ctx->nr_pages) {
339 /* Make sure the old page hasn't already been changed */
340 if (ctx->ring_pages[idx] != old)
341 rc = -EAGAIN;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500342 } else
343 rc = -EINVAL;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500344
345 if (rc != 0)
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400346 goto out_unlock;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500347
Gu Zheng36bc08c2013-07-16 17:56:16 +0800348 /* Writeback must be complete */
349 BUG_ON(PageWriteback(old));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500350 get_page(new);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800351
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500352 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800353 if (rc != MIGRATEPAGE_SUCCESS) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500354 put_page(new);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400355 goto out_unlock;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800356 }
357
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400358 /* Take completion_lock to prevent other writes to the ring buffer
359 * while the old page is copied to the new. This prevents new
360 * events from being lost.
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400361 */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400362 spin_lock_irqsave(&ctx->completion_lock, flags);
363 migrate_page_copy(new, old);
364 BUG_ON(ctx->ring_pages[idx] != old);
365 ctx->ring_pages[idx] = new;
366 spin_unlock_irqrestore(&ctx->completion_lock, flags);
367
368 /* The old page is no longer accessible. */
369 put_page(old);
370
371out_unlock:
372 mutex_unlock(&ctx->ring_lock);
373out:
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400374 spin_unlock(&mapping->private_lock);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800375 return rc;
376}
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400377#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800378
379static const struct address_space_operations aio_ctx_aops = {
Gu Zheng835f2522014-11-06 17:46:21 +0800380 .set_page_dirty = __set_page_dirty_no_writeback,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400381#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800382 .migratepage = aio_migratepage,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400383#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800384};
385
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386static int aio_setup_ring(struct kioctx *ctx)
387{
388 struct aio_ring *ring;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389 unsigned nr_events = ctx->max_reqs;
Zach Brown41003a72013-05-07 16:18:25 -0700390 struct mm_struct *mm = current->mm;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900391 unsigned long size, unused;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 int nr_pages;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800393 int i;
394 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395
396 /* Compensate for the ring buffer's head/tail overlap entry */
397 nr_events += 2; /* 1 is required, 2 for good luck */
398
399 size = sizeof(struct aio_ring);
400 size += sizeof(struct io_event) * nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
Gu Zheng36bc08c2013-07-16 17:56:16 +0800402 nr_pages = PFN_UP(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 if (nr_pages < 0)
404 return -EINVAL;
405
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400406 file = aio_private_file(ctx, nr_pages);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800407 if (IS_ERR(file)) {
408 ctx->aio_ring_file = NULL;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400409 return -ENOMEM;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800410 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
Gu Zheng36bc08c2013-07-16 17:56:16 +0800412 ctx->aio_ring_file = file;
413 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
414 / sizeof(struct io_event);
415
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700416 ctx->ring_pages = ctx->internal_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 if (nr_pages > AIO_RING_PAGES) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700418 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
419 GFP_KERNEL);
Gu Zhengd1b94322013-12-04 18:19:06 +0800420 if (!ctx->ring_pages) {
421 put_aio_ring_file(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 return -ENOMEM;
Gu Zhengd1b94322013-12-04 18:19:06 +0800423 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 }
425
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900426 for (i = 0; i < nr_pages; i++) {
427 struct page *page;
428 page = find_or_create_page(file->f_inode->i_mapping,
429 i, GFP_HIGHUSER | __GFP_ZERO);
430 if (!page)
431 break;
432 pr_debug("pid(%d) page[%d]->count=%d\n",
433 current->pid, i, page_count(page));
434 SetPageUptodate(page);
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900435 unlock_page(page);
436
437 ctx->ring_pages[i] = page;
438 }
439 ctx->nr_pages = i;
440
441 if (unlikely(i != nr_pages)) {
442 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400443 return -ENOMEM;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900444 }
445
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700446 ctx->mmap_size = nr_pages * PAGE_SIZE;
447 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800448
Zach Brown41003a72013-05-07 16:18:25 -0700449 down_write(&mm->mmap_sem);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800450 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
451 PROT_READ | PROT_WRITE,
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900452 MAP_SHARED, 0, &unused);
453 up_write(&mm->mmap_sem);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700454 if (IS_ERR((void *)ctx->mmap_base)) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700455 ctx->mmap_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400457 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 }
459
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700460 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400461
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700462 ctx->user_id = ctx->mmap_base;
463 ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700465 ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 ring->nr = nr_events; /* user copy */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400467 ring->id = ~0U;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 ring->head = ring->tail = 0;
469 ring->magic = AIO_RING_MAGIC;
470 ring->compat_features = AIO_RING_COMPAT_FEATURES;
471 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
472 ring->header_length = sizeof(struct aio_ring);
Cong Wange8e3c3d2011-11-25 23:14:27 +0800473 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700474 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475
476 return 0;
477}
478
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
480#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
481#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
482
Kent Overstreet0460fef2013-05-07 16:18:49 -0700483void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
484{
485 struct kioctx *ctx = req->ki_ctx;
486 unsigned long flags;
487
488 spin_lock_irqsave(&ctx->ctx_lock, flags);
489
490 if (!req->ki_list.next)
491 list_add(&req->ki_list, &ctx->active_reqs);
492
493 req->ki_cancel = cancel;
494
495 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
496}
497EXPORT_SYMBOL(kiocb_set_cancel_fn);
498
Fabian Frederickd52a8f92014-04-22 07:26:58 +0200499static int kiocb_cancel(struct kiocb *kiocb)
Kent Overstreet906b9732013-05-07 16:18:31 -0700500{
Kent Overstreet0460fef2013-05-07 16:18:49 -0700501 kiocb_cancel_fn *old, *cancel;
Kent Overstreet906b9732013-05-07 16:18:31 -0700502
Kent Overstreet0460fef2013-05-07 16:18:49 -0700503 /*
504 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
505 * actually has a cancel function, hence the cmpxchg()
506 */
Kent Overstreet906b9732013-05-07 16:18:31 -0700507
Kent Overstreet0460fef2013-05-07 16:18:49 -0700508 cancel = ACCESS_ONCE(kiocb->ki_cancel);
509 do {
510 if (!cancel || cancel == KIOCB_CANCELLED)
Kent Overstreet57282d82013-05-13 13:42:52 -0700511 return -EINVAL;
Kent Overstreet906b9732013-05-07 16:18:31 -0700512
Kent Overstreet0460fef2013-05-07 16:18:49 -0700513 old = cancel;
514 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
515 } while (cancel != old);
516
Kent Overstreet57282d82013-05-13 13:42:52 -0700517 return cancel(kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -0700518}
519
Kent Overstreete34ecee2013-10-10 19:31:47 -0700520static void free_ioctx(struct work_struct *work)
Kent Overstreet36f55882013-05-07 16:18:41 -0700521{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700522 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000523
Kent Overstreete34ecee2013-10-10 19:31:47 -0700524 pr_debug("freeing %p\n", ctx);
525
526 aio_free_ring(ctx);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000527 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400528 percpu_ref_exit(&ctx->reqs);
529 percpu_ref_exit(&ctx->users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700530 kmem_cache_free(kioctx_cachep, ctx);
531}
532
Kent Overstreete34ecee2013-10-10 19:31:47 -0700533static void free_ioctx_reqs(struct percpu_ref *ref)
534{
535 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
536
Anatol Pomozove02ba722014-04-15 11:31:33 -0700537 /* At this point we know that there are no any in-flight requests */
538 if (ctx->requests_done)
539 complete(ctx->requests_done);
540
Kent Overstreete34ecee2013-10-10 19:31:47 -0700541 INIT_WORK(&ctx->free_work, free_ioctx);
542 schedule_work(&ctx->free_work);
543}
544
Kent Overstreet36f55882013-05-07 16:18:41 -0700545/*
546 * When this function runs, the kioctx has been removed from the "hash table"
547 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
548 * now it's safe to cancel any that need to be.
549 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700550static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet36f55882013-05-07 16:18:41 -0700551{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700552 struct kioctx *ctx = container_of(ref, struct kioctx, users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700553 struct kiocb *req;
554
555 spin_lock_irq(&ctx->ctx_lock);
556
557 while (!list_empty(&ctx->active_reqs)) {
558 req = list_first_entry(&ctx->active_reqs,
559 struct kiocb, ki_list);
560
561 list_del_init(&req->ki_list);
Fabian Frederickd52a8f92014-04-22 07:26:58 +0200562 kiocb_cancel(req);
Kent Overstreet36f55882013-05-07 16:18:41 -0700563 }
564
565 spin_unlock_irq(&ctx->ctx_lock);
566
Kent Overstreete34ecee2013-10-10 19:31:47 -0700567 percpu_ref_kill(&ctx->reqs);
568 percpu_ref_put(&ctx->reqs);
Kent Overstreet36f55882013-05-07 16:18:41 -0700569}
570
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400571static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
572{
573 unsigned i, new_nr;
574 struct kioctx_table *table, *old;
575 struct aio_ring *ring;
576
577 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200578 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400579
580 while (1) {
581 if (table)
582 for (i = 0; i < table->nr; i++)
583 if (!table->table[i]) {
584 ctx->id = i;
585 table->table[i] = ctx;
586 spin_unlock(&mm->ioctx_lock);
587
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400588 /* While kioctx setup is in progress,
589 * we are protected from page migration
590 * changes ring_pages by ->ring_lock.
591 */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400592 ring = kmap_atomic(ctx->ring_pages[0]);
593 ring->id = ctx->id;
594 kunmap_atomic(ring);
595 return 0;
596 }
597
598 new_nr = (table ? table->nr : 1) * 4;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400599 spin_unlock(&mm->ioctx_lock);
600
601 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
602 new_nr, GFP_KERNEL);
603 if (!table)
604 return -ENOMEM;
605
606 table->nr = new_nr;
607
608 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200609 old = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400610
611 if (!old) {
612 rcu_assign_pointer(mm->ioctx_table, table);
613 } else if (table->nr > old->nr) {
614 memcpy(table->table, old->table,
615 old->nr * sizeof(struct kioctx *));
616
617 rcu_assign_pointer(mm->ioctx_table, table);
618 kfree_rcu(old, rcu);
619 } else {
620 kfree(table);
621 table = old;
622 }
623 }
624}
625
Kent Overstreete34ecee2013-10-10 19:31:47 -0700626static void aio_nr_sub(unsigned nr)
627{
628 spin_lock(&aio_nr_lock);
629 if (WARN_ON(aio_nr - nr > aio_nr))
630 aio_nr = 0;
631 else
632 aio_nr -= nr;
633 spin_unlock(&aio_nr_lock);
634}
635
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636/* ioctx_alloc
637 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
638 */
639static struct kioctx *ioctx_alloc(unsigned nr_events)
640{
Zach Brown41003a72013-05-07 16:18:25 -0700641 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642 struct kioctx *ctx;
Al Viroe23754f2012-03-06 14:33:22 -0500643 int err = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000645 /*
646 * We keep track of the number of available ringbuffer slots, to prevent
647 * overflow (reqs_available), and we also use percpu counters for this.
648 *
649 * So since up to half the slots might be on other cpu's percpu counters
650 * and unavailable, double nr_events so userspace sees what they
651 * expected: additionally, we move req_batch slots to/from percpu
652 * counters at a time, so make sure that isn't 0:
653 */
654 nr_events = max(nr_events, num_possible_cpus() * 4);
655 nr_events *= 2;
656
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 /* Prevent overflows */
658 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
659 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
660 pr_debug("ENOMEM: nr_events too high\n");
661 return ERR_PTR(-EINVAL);
662 }
663
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400664 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665 return ERR_PTR(-EAGAIN);
666
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800667 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 if (!ctx)
669 return ERR_PTR(-ENOMEM);
670
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 ctx->max_reqs = nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400673 spin_lock_init(&ctx->ctx_lock);
674 spin_lock_init(&ctx->completion_lock);
675 mutex_init(&ctx->ring_lock);
676 /* Protect against page migration throughout kiotx setup by keeping
677 * the ring_lock mutex held until setup is complete. */
678 mutex_lock(&ctx->ring_lock);
679 init_waitqueue_head(&ctx->wait);
680
681 INIT_LIST_HEAD(&ctx->active_reqs);
682
Tejun Heo2aad2a82014-09-24 13:31:50 -0400683 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700684 goto err;
685
Tejun Heo2aad2a82014-09-24 13:31:50 -0400686 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700687 goto err;
Kent Overstreet723be6e2013-05-28 15:14:48 -0700688
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000689 ctx->cpu = alloc_percpu(struct kioctx_cpu);
690 if (!ctx->cpu)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700691 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400693 err = aio_setup_ring(ctx);
694 if (err < 0)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700695 goto err;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000696
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000697 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000698 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
Benjamin LaHaise6878ea72013-07-31 10:34:18 -0400699 if (ctx->req_batch < 1)
700 ctx->req_batch = 1;
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000701
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702 /* limit the number of system wide aios */
Al Viro9fa1cb32012-03-10 23:14:05 -0500703 spin_lock(&aio_nr_lock);
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400704 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
Al Viro2dd542b2012-03-10 23:10:35 -0500705 aio_nr + nr_events < aio_nr) {
Al Viro9fa1cb32012-03-10 23:14:05 -0500706 spin_unlock(&aio_nr_lock);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700707 err = -EAGAIN;
Gu Zhengd1b94322013-12-04 18:19:06 +0800708 goto err_ctx;
Al Viro2dd542b2012-03-10 23:10:35 -0500709 }
710 aio_nr += ctx->max_reqs;
Al Viro9fa1cb32012-03-10 23:14:05 -0500711 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712
Benjamin LaHaise18816862013-12-21 15:49:28 -0500713 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
714 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
Kent Overstreet723be6e2013-05-28 15:14:48 -0700715
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400716 err = ioctx_add_table(ctx, mm);
717 if (err)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700718 goto err_cleanup;
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400719
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400720 /* Release the ring_lock mutex now that all setup is complete. */
721 mutex_unlock(&ctx->ring_lock);
722
Kent Overstreetcaf41672013-05-07 16:18:35 -0700723 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700724 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 return ctx;
726
Kent Overstreete34ecee2013-10-10 19:31:47 -0700727err_cleanup:
728 aio_nr_sub(ctx->max_reqs);
Gu Zhengd1b94322013-12-04 18:19:06 +0800729err_ctx:
730 aio_free_ring(ctx);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700731err:
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400732 mutex_unlock(&ctx->ring_lock);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000733 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400734 percpu_ref_exit(&ctx->reqs);
735 percpu_ref_exit(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreetcaf41672013-05-07 16:18:35 -0700737 pr_debug("error allocating ioctx %d\n", err);
Al Viroe23754f2012-03-06 14:33:22 -0500738 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739}
740
Kent Overstreet36f55882013-05-07 16:18:41 -0700741/* kill_ioctx
742 * Cancels all outstanding aio requests on an aio context. Used
743 * when the processes owning a context have all exited to encourage
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 * the rapid destruction of the kioctx.
745 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400746static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
Anatol Pomozove02ba722014-04-15 11:31:33 -0700747 struct completion *requests_done)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748{
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400749 struct kioctx_table *table;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400750
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400751 if (atomic_xchg(&ctx->dead, 1))
752 return -EINVAL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400753
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400754
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400755 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200756 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400757 WARN_ON(ctx != table->table[ctx->id]);
758 table->table[ctx->id] = NULL;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400759 spin_unlock(&mm->ioctx_lock);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700760
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400761 /* percpu_ref_kill() will do the necessary call_rcu() */
762 wake_up_all(&ctx->wait);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700763
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400764 /*
765 * It'd be more correct to do this in free_ioctx(), after all
766 * the outstanding kiocbs have finished - but by then io_destroy
767 * has already returned, so io_setup() could potentially return
768 * -EAGAIN with no ioctxs actually in use (as far as userspace
769 * could tell).
770 */
771 aio_nr_sub(ctx->max_reqs);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400772
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400773 if (ctx->mmap_size)
774 vm_munmap(ctx->mmap_base, ctx->mmap_size);
775
776 ctx->requests_done = requests_done;
777 percpu_ref_kill(&ctx->users);
778 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779}
780
781/* wait_on_sync_kiocb:
782 * Waits on the given sync kiocb to complete.
783 */
Kent Overstreet57282d82013-05-13 13:42:52 -0700784ssize_t wait_on_sync_kiocb(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785{
Kent Overstreet57282d82013-05-13 13:42:52 -0700786 while (!req->ki_ctx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 set_current_state(TASK_UNINTERRUPTIBLE);
Kent Overstreet57282d82013-05-13 13:42:52 -0700788 if (req->ki_ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 break;
Jeff Moyer41d10da2007-10-16 23:27:20 -0700790 io_schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 }
792 __set_current_state(TASK_RUNNING);
Kent Overstreet57282d82013-05-13 13:42:52 -0700793 return req->ki_user_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700795EXPORT_SYMBOL(wait_on_sync_kiocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796
Kent Overstreet36f55882013-05-07 16:18:41 -0700797/*
798 * exit_aio: called when the last user of mm goes away. At this point, there is
799 * no way for any new requests to be submited or any of the io_* syscalls to be
800 * called on the context.
801 *
802 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
803 * them.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800805void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806{
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200807 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
808 int i;
Jens Axboeabf137d2008-12-09 08:11:22 +0100809
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200810 if (!table)
811 return;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400812
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200813 for (i = 0; i < table->nr; ++i) {
814 struct kioctx *ctx = table->table[i];
Gu Zheng6098b452014-09-03 17:45:44 +0800815 struct completion requests_done =
816 COMPLETION_INITIALIZER_ONSTACK(requests_done);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400817
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200818 if (!ctx)
819 continue;
Al Viro936af152012-04-20 21:49:41 -0400820 /*
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200821 * We don't need to bother with munmap() here - exit_mmap(mm)
822 * is coming and it'll unmap everything. And we simply can't,
823 * this is not necessarily our ->mm.
824 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
825 * that it needs to unmap the area, just set it to 0.
Al Viro936af152012-04-20 21:49:41 -0400826 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700827 ctx->mmap_size = 0;
Gu Zheng6098b452014-09-03 17:45:44 +0800828 kill_ioctx(mm, ctx, &requests_done);
Kent Overstreet36f55882013-05-07 16:18:41 -0700829
Gu Zheng6098b452014-09-03 17:45:44 +0800830 /* Wait until all IO for the context are done. */
831 wait_for_completion(&requests_done);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 }
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200833
834 RCU_INIT_POINTER(mm->ioctx_table, NULL);
835 kfree(table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836}
837
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000838static void put_reqs_available(struct kioctx *ctx, unsigned nr)
839{
840 struct kioctx_cpu *kcpu;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400841 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000842
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400843 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400844 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000845 kcpu->reqs_available += nr;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400846
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000847 while (kcpu->reqs_available >= ctx->req_batch * 2) {
848 kcpu->reqs_available -= ctx->req_batch;
849 atomic_add(ctx->req_batch, &ctx->reqs_available);
850 }
851
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400852 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000853}
854
855static bool get_reqs_available(struct kioctx *ctx)
856{
857 struct kioctx_cpu *kcpu;
858 bool ret = false;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400859 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000860
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400861 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400862 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000863 if (!kcpu->reqs_available) {
864 int old, avail = atomic_read(&ctx->reqs_available);
865
866 do {
867 if (avail < ctx->req_batch)
868 goto out;
869
870 old = avail;
871 avail = atomic_cmpxchg(&ctx->reqs_available,
872 avail, avail - ctx->req_batch);
873 } while (avail != old);
874
875 kcpu->reqs_available += ctx->req_batch;
876 }
877
878 ret = true;
879 kcpu->reqs_available--;
880out:
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400881 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000882 return ret;
883}
884
Benjamin LaHaised856f322014-08-24 13:14:05 -0400885/* refill_reqs_available
886 * Updates the reqs_available reference counts used for tracking the
887 * number of free slots in the completion ring. This can be called
888 * from aio_complete() (to optimistically update reqs_available) or
889 * from aio_get_req() (the we're out of events case). It must be
890 * called holding ctx->completion_lock.
891 */
892static void refill_reqs_available(struct kioctx *ctx, unsigned head,
893 unsigned tail)
894{
895 unsigned events_in_ring, completed;
896
897 /* Clamp head since userland can write to it. */
898 head %= ctx->nr_events;
899 if (head <= tail)
900 events_in_ring = tail - head;
901 else
902 events_in_ring = ctx->nr_events - (head - tail);
903
904 completed = ctx->completed_events;
905 if (events_in_ring < completed)
906 completed -= events_in_ring;
907 else
908 completed = 0;
909
910 if (!completed)
911 return;
912
913 ctx->completed_events -= completed;
914 put_reqs_available(ctx, completed);
915}
916
917/* user_refill_reqs_available
918 * Called to refill reqs_available when aio_get_req() encounters an
919 * out of space in the completion ring.
920 */
921static void user_refill_reqs_available(struct kioctx *ctx)
922{
923 spin_lock_irq(&ctx->completion_lock);
924 if (ctx->completed_events) {
925 struct aio_ring *ring;
926 unsigned head;
927
928 /* Access of ring->head may race with aio_read_events_ring()
929 * here, but that's okay since whether we read the old version
930 * or the new version, and either will be valid. The important
931 * part is that head cannot pass tail since we prevent
932 * aio_complete() from updating tail by holding
933 * ctx->completion_lock. Even if head is invalid, the check
934 * against ctx->completed_events below will make sure we do the
935 * safe/right thing.
936 */
937 ring = kmap_atomic(ctx->ring_pages[0]);
938 head = ring->head;
939 kunmap_atomic(ring);
940
941 refill_reqs_available(ctx, head, ctx->tail);
942 }
943
944 spin_unlock_irq(&ctx->completion_lock);
945}
946
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947/* aio_get_req
Kent Overstreet57282d82013-05-13 13:42:52 -0700948 * Allocate a slot for an aio request.
949 * Returns NULL if no requests are free.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 */
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700951static inline struct kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952{
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700953 struct kiocb *req;
954
Benjamin LaHaised856f322014-08-24 13:14:05 -0400955 if (!get_reqs_available(ctx)) {
956 user_refill_reqs_available(ctx);
957 if (!get_reqs_available(ctx))
958 return NULL;
959 }
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700960
Kent Overstreet0460fef2013-05-07 16:18:49 -0700961 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 if (unlikely(!req))
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700963 goto out_put;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964
Kent Overstreete34ecee2013-10-10 19:31:47 -0700965 percpu_ref_get(&ctx->reqs);
966
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 req->ki_ctx = ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 return req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700969out_put:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000970 put_reqs_available(ctx, 1);
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700971 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972}
973
Kent Overstreet11599eb2013-05-07 16:18:39 -0700974static void kiocb_free(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975{
Kent Overstreet1d98ebf2013-05-07 16:18:37 -0700976 if (req->ki_filp)
977 fput(req->ki_filp);
Davide Libenzi13389012009-06-30 11:41:11 -0700978 if (req->ki_eventfd != NULL)
979 eventfd_ctx_put(req->ki_eventfd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980 kmem_cache_free(kiocb_cachep, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981}
982
Adrian Bunkd5470b52008-04-29 00:58:57 -0700983static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400985 struct aio_ring __user *ring = (void __user *)ctx_id;
Jens Axboeabf137d2008-12-09 08:11:22 +0100986 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -0700987 struct kioctx *ctx, *ret = NULL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400988 struct kioctx_table *table;
989 unsigned id;
990
991 if (get_user(id, &ring->id))
992 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993
Jens Axboeabf137d2008-12-09 08:11:22 +0100994 rcu_read_lock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400995 table = rcu_dereference(mm->ioctx_table);
Jens Axboeabf137d2008-12-09 08:11:22 +0100996
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400997 if (!table || id >= table->nr)
998 goto out;
999
1000 ctx = table->table[id];
Benjamin LaHaisef30d7042013-08-07 18:23:48 -04001001 if (ctx && ctx->user_id == ctx_id) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001002 percpu_ref_get(&ctx->users);
1003 ret = ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +01001004 }
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001005out:
Jens Axboeabf137d2008-12-09 08:11:22 +01001006 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -07001007 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008}
1009
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010/* aio_complete
1011 * Called when the io request on the given iocb is complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 */
Kent Overstreet2d684492013-05-07 16:18:29 -07001013void aio_complete(struct kiocb *iocb, long res, long res2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014{
1015 struct kioctx *ctx = iocb->ki_ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016 struct aio_ring *ring;
Kent Overstreet21b40202013-05-07 16:18:47 -07001017 struct io_event *ev_page, *event;
Benjamin LaHaised856f322014-08-24 13:14:05 -04001018 unsigned tail, pos, head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020
Zach Brown20dcae32005-11-13 16:07:33 -08001021 /*
1022 * Special case handling for sync iocbs:
1023 * - events go directly into the iocb for fast handling
1024 * - the sync task with the iocb in its stack holds the single iocb
1025 * ref, no other paths have a way to get another ref
1026 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 */
1028 if (is_sync_kiocb(iocb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029 iocb->ki_user_data = res;
Kent Overstreet57282d82013-05-13 13:42:52 -07001030 smp_wmb();
1031 iocb->ki_ctx = ERR_PTR(-EXDEV);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032 wake_up_process(iocb->ki_obj.tsk);
Kent Overstreet2d684492013-05-07 16:18:29 -07001033 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 }
1035
Kent Overstreet0460fef2013-05-07 16:18:49 -07001036 if (iocb->ki_list.next) {
1037 unsigned long flags;
1038
1039 spin_lock_irqsave(&ctx->ctx_lock, flags);
1040 list_del(&iocb->ki_list);
1041 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1042 }
Kent Overstreet11599eb2013-05-07 16:18:39 -07001043
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 /*
Kent Overstreet0460fef2013-05-07 16:18:49 -07001045 * Add a completion event to the ring buffer. Must be done holding
Tang Chen4b30f072013-07-03 15:09:16 -07001046 * ctx->completion_lock to prevent other code from messing with the tail
Kent Overstreet0460fef2013-05-07 16:18:49 -07001047 * pointer since we might be called from irq context.
1048 */
1049 spin_lock_irqsave(&ctx->completion_lock, flags);
1050
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001051 tail = ctx->tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001052 pos = tail + AIO_EVENTS_OFFSET;
1053
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001054 if (++tail >= ctx->nr_events)
Ken Chen4bf69b22005-05-01 08:59:15 -07001055 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001057 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001058 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1059
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1061 event->data = iocb->ki_user_data;
1062 event->res = res;
1063 event->res2 = res2;
1064
Kent Overstreet21b40202013-05-07 16:18:47 -07001065 kunmap_atomic(ev_page);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001066 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001067
1068 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
Kent Overstreetcaf41672013-05-07 16:18:35 -07001069 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1070 res, res2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071
1072 /* after flagging the request as done, we
1073 * must never even look at it again
1074 */
1075 smp_wmb(); /* make event visible before updating tail */
1076
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001077 ctx->tail = tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001078
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001079 ring = kmap_atomic(ctx->ring_pages[0]);
Benjamin LaHaised856f322014-08-24 13:14:05 -04001080 head = ring->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081 ring->tail = tail;
Cong Wange8e3c3d2011-11-25 23:14:27 +08001082 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001083 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084
Benjamin LaHaised856f322014-08-24 13:14:05 -04001085 ctx->completed_events++;
1086 if (ctx->completed_events > 1)
1087 refill_reqs_available(ctx, head, tail);
Kent Overstreet0460fef2013-05-07 16:18:49 -07001088 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1089
Kent Overstreet21b40202013-05-07 16:18:47 -07001090 pr_debug("added to ring %p at [%u]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001091
1092 /*
1093 * Check if the user asked us to deliver the result through an
1094 * eventfd. The eventfd_signal() function is safe to be called
1095 * from IRQ context.
1096 */
Davide Libenzi87c3a862009-03-18 17:04:19 -07001097 if (iocb->ki_eventfd != NULL)
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001098 eventfd_signal(iocb->ki_eventfd, 1);
1099
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 /* everything turned out well, dispose of the aiocb. */
Kent Overstreet57282d82013-05-13 13:42:52 -07001101 kiocb_free(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102
Quentin Barnes6cb2a212008-03-19 17:00:39 -07001103 /*
1104 * We have to order our ring_info tail store above and test
1105 * of the wait list below outside the wait lock. This is
1106 * like in wake_up_bit() where clearing a bit has to be
1107 * ordered with the unlocked test.
1108 */
1109 smp_mb();
1110
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 if (waitqueue_active(&ctx->wait))
1112 wake_up(&ctx->wait);
1113
Kent Overstreete34ecee2013-10-10 19:31:47 -07001114 percpu_ref_put(&ctx->reqs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115}
H Hartley Sweeten385773e2009-09-22 16:43:53 -07001116EXPORT_SYMBOL(aio_complete);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001118/* aio_read_events_ring
Kent Overstreeta31ad382013-05-07 16:18:45 -07001119 * Pull an event off of the ioctx's event ring. Returns the number of
1120 * events fetched
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 */
Kent Overstreeta31ad382013-05-07 16:18:45 -07001122static long aio_read_events_ring(struct kioctx *ctx,
1123 struct io_event __user *event, long nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 struct aio_ring *ring;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001126 unsigned head, tail, pos;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001127 long ret = 0;
1128 int copy_ret;
1129
Dave Chinner9c9ce762015-02-03 19:29:05 -05001130 /*
1131 * The mutex can block and wake us up and that will cause
1132 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1133 * and repeat. This should be rare enough that it doesn't cause
1134 * peformance issues. See the comment in read_events() for more detail.
1135 */
1136 sched_annotate_sleep();
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001137 mutex_lock(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -04001139 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001140 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001141 head = ring->head;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001142 tail = ring->tail;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001143 kunmap_atomic(ring);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144
Jeff Moyer2ff396be2014-09-02 13:17:00 -04001145 /*
1146 * Ensure that once we've read the current tail pointer, that
1147 * we also see the events that were stored up to the tail.
1148 */
1149 smp_rmb();
1150
Kent Overstreet5ffac122013-05-09 15:36:07 -07001151 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001152
Kent Overstreet5ffac122013-05-09 15:36:07 -07001153 if (head == tail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 goto out;
1155
Benjamin LaHaiseedfbbf32014-06-24 13:32:51 -04001156 head %= ctx->nr_events;
1157 tail %= ctx->nr_events;
1158
Kent Overstreeta31ad382013-05-07 16:18:45 -07001159 while (ret < nr) {
1160 long avail;
1161 struct io_event *ev;
1162 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163
Kent Overstreet5ffac122013-05-09 15:36:07 -07001164 avail = (head <= tail ? tail : ctx->nr_events) - head;
1165 if (head == tail)
Kent Overstreeta31ad382013-05-07 16:18:45 -07001166 break;
1167
1168 avail = min(avail, nr - ret);
1169 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1170 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1171
1172 pos = head + AIO_EVENTS_OFFSET;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001173 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
Kent Overstreeta31ad382013-05-07 16:18:45 -07001174 pos %= AIO_EVENTS_PER_PAGE;
1175
1176 ev = kmap(page);
1177 copy_ret = copy_to_user(event + ret, ev + pos,
1178 sizeof(*ev) * avail);
1179 kunmap(page);
1180
1181 if (unlikely(copy_ret)) {
1182 ret = -EFAULT;
1183 goto out;
1184 }
1185
1186 ret += avail;
1187 head += avail;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001188 head %= ctx->nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001191 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001192 ring->head = head;
Zhao Hongjiang91d80a82013-04-26 11:03:53 +08001193 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001194 flush_dcache_page(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001195
Kent Overstreet5ffac122013-05-09 15:36:07 -07001196 pr_debug("%li h%u t%u\n", ret, head, tail);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001197out:
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001198 mutex_unlock(&ctx->ring_lock);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001199
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200 return ret;
1201}
1202
Kent Overstreeta31ad382013-05-07 16:18:45 -07001203static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1204 struct io_event __user *event, long *i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001206 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207
Kent Overstreeta31ad382013-05-07 16:18:45 -07001208 if (ret > 0)
1209 *i += ret;
1210
1211 if (unlikely(atomic_read(&ctx->dead)))
1212 ret = -EINVAL;
1213
1214 if (!*i)
1215 *i = ret;
1216
1217 return ret < 0 || *i >= min_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218}
1219
Kent Overstreeta31ad382013-05-07 16:18:45 -07001220static long read_events(struct kioctx *ctx, long min_nr, long nr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 struct io_event __user *event,
1222 struct timespec __user *timeout)
1223{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001224 ktime_t until = { .tv64 = KTIME_MAX };
1225 long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 if (timeout) {
1228 struct timespec ts;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001229
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
Kent Overstreeta31ad382013-05-07 16:18:45 -07001231 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232
Kent Overstreeta31ad382013-05-07 16:18:45 -07001233 until = timespec_to_ktime(ts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 }
1235
Kent Overstreeta31ad382013-05-07 16:18:45 -07001236 /*
1237 * Note that aio_read_events() is being called as the conditional - i.e.
1238 * we're calling it after prepare_to_wait() has set task state to
1239 * TASK_INTERRUPTIBLE.
1240 *
1241 * But aio_read_events() can block, and if it blocks it's going to flip
1242 * the task state back to TASK_RUNNING.
1243 *
1244 * This should be ok, provided it doesn't flip the state back to
1245 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1246 * will only happen if the mutex_lock() call blocks, and we then find
1247 * the ringbuffer empty. So in practice we should be ok, but it's
1248 * something to be aware of when touching this code.
1249 */
Fam Zheng5f785de2014-11-06 20:44:36 +08001250 if (until.tv64 == 0)
1251 aio_read_events(ctx, min_nr, nr, event, &ret);
1252 else
1253 wait_event_interruptible_hrtimeout(ctx->wait,
1254 aio_read_events(ctx, min_nr, nr, event, &ret),
1255 until);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256
Kent Overstreeta31ad382013-05-07 16:18:45 -07001257 if (!ret && signal_pending(current))
1258 ret = -EINTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259
Kent Overstreeta31ad382013-05-07 16:18:45 -07001260 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261}
1262
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263/* sys_io_setup:
1264 * Create an aio_context capable of receiving at least nr_events.
1265 * ctxp must not point to an aio_context that already exists, and
1266 * must be initialized to 0 prior to the call. On successful
1267 * creation of the aio_context, *ctxp is filled in with the resulting
1268 * handle. May fail with -EINVAL if *ctxp is not initialized,
1269 * if the specified nr_events exceeds internal limits. May fail
1270 * with -EAGAIN if the specified nr_events exceeds the user's limit
1271 * of available events. May fail with -ENOMEM if insufficient kernel
1272 * resources are available. May fail with -EFAULT if an invalid
1273 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1274 * implemented.
1275 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001276SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277{
1278 struct kioctx *ioctx = NULL;
1279 unsigned long ctx;
1280 long ret;
1281
1282 ret = get_user(ctx, ctxp);
1283 if (unlikely(ret))
1284 goto out;
1285
1286 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001287 if (unlikely(ctx || nr_events == 0)) {
1288 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1289 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290 goto out;
1291 }
1292
1293 ioctx = ioctx_alloc(nr_events);
1294 ret = PTR_ERR(ioctx);
1295 if (!IS_ERR(ioctx)) {
1296 ret = put_user(ioctx->user_id, ctxp);
Al Viroa2e18592012-03-20 16:27:57 -04001297 if (ret)
Anatol Pomozove02ba722014-04-15 11:31:33 -07001298 kill_ioctx(current->mm, ioctx, NULL);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001299 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 }
1301
1302out:
1303 return ret;
1304}
1305
1306/* sys_io_destroy:
1307 * Destroy the aio_context specified. May cancel any outstanding
1308 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001309 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 * is invalid.
1311 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001312SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313{
1314 struct kioctx *ioctx = lookup_ioctx(ctx);
1315 if (likely(NULL != ioctx)) {
Anatol Pomozove02ba722014-04-15 11:31:33 -07001316 struct completion requests_done =
1317 COMPLETION_INITIALIZER_ONSTACK(requests_done);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001318 int ret;
Anatol Pomozove02ba722014-04-15 11:31:33 -07001319
1320 /* Pass requests_done to kill_ioctx() where it can be set
1321 * in a thread-safe way. If we try to set it here then we have
1322 * a race condition if two io_destroy() called simultaneously.
1323 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001324 ret = kill_ioctx(current->mm, ioctx, &requests_done);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001325 percpu_ref_put(&ioctx->users);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001326
1327 /* Wait until all IO for the context are done. Otherwise kernel
1328 * keep using user-space buffers even if user thinks the context
1329 * is destroyed.
1330 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001331 if (!ret)
1332 wait_for_completion(&requests_done);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001333
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001334 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 }
1336 pr_debug("EINVAL: io_destroy: invalid context id\n");
1337 return -EINVAL;
1338}
1339
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001340typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1341 unsigned long, loff_t);
Al Viro293bc982014-02-11 18:37:41 -05001342typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001343
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001344static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1345 int rw, char __user *buf,
1346 unsigned long *nr_segs,
1347 struct iovec **iovec,
1348 bool compat)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001349{
1350 ssize_t ret;
1351
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001352 *nr_segs = kiocb->ki_nbytes;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001353
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001354#ifdef CONFIG_COMPAT
1355 if (compat)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001356 ret = compat_rw_copy_check_uvector(rw,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001357 (struct compat_iovec __user *)buf,
Gu Zheng00fefb92014-07-23 18:03:54 +08001358 *nr_segs, UIO_FASTIOV, *iovec, iovec);
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001359 else
1360#endif
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001361 ret = rw_copy_check_uvector(rw,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001362 (struct iovec __user *)buf,
Gu Zheng00fefb92014-07-23 18:03:54 +08001363 *nr_segs, UIO_FASTIOV, *iovec, iovec);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001364 if (ret < 0)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001365 return ret;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001366
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001367 /* ki_nbytes now reflect bytes instead of segs */
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001368 kiocb->ki_nbytes = ret;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001369 return 0;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001370}
1371
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001372static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1373 int rw, char __user *buf,
1374 unsigned long *nr_segs,
1375 struct iovec *iovec)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001376{
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001377 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001378 return -EFAULT;
Linus Torvaldsa70b52e2012-05-21 16:06:20 -07001379
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001380 iovec->iov_base = buf;
1381 iovec->iov_len = kiocb->ki_nbytes;
1382 *nr_segs = 1;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001383 return 0;
1384}
1385
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386/*
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001387 * aio_run_iocb:
1388 * Performs the initial checks and io submission.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 */
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001390static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1391 char __user *buf, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392{
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001393 struct file *file = req->ki_filp;
1394 ssize_t ret;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001395 unsigned long nr_segs;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001396 int rw;
1397 fmode_t mode;
1398 aio_rw_op *rw_op;
Al Viro293bc982014-02-11 18:37:41 -05001399 rw_iter_op *iter_op;
Gu Zheng00fefb92014-07-23 18:03:54 +08001400 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
Al Viro293bc982014-02-11 18:37:41 -05001401 struct iov_iter iter;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001403 switch (opcode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 case IOCB_CMD_PREAD:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001405 case IOCB_CMD_PREADV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001406 mode = FMODE_READ;
1407 rw = READ;
1408 rw_op = file->f_op->aio_read;
Al Viro293bc982014-02-11 18:37:41 -05001409 iter_op = file->f_op->read_iter;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001410 goto rw_common;
1411
1412 case IOCB_CMD_PWRITE:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001413 case IOCB_CMD_PWRITEV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001414 mode = FMODE_WRITE;
1415 rw = WRITE;
1416 rw_op = file->f_op->aio_write;
Al Viro293bc982014-02-11 18:37:41 -05001417 iter_op = file->f_op->write_iter;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001418 goto rw_common;
1419rw_common:
1420 if (unlikely(!(file->f_mode & mode)))
1421 return -EBADF;
1422
Al Viro293bc982014-02-11 18:37:41 -05001423 if (!rw_op && !iter_op)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001424 return -EINVAL;
1425
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001426 ret = (opcode == IOCB_CMD_PREADV ||
1427 opcode == IOCB_CMD_PWRITEV)
1428 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1429 &iovec, compat)
1430 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1431 iovec);
Leon Yu754320d2014-05-01 03:31:28 +00001432 if (!ret)
1433 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001434 if (ret < 0) {
Gu Zheng00fefb92014-07-23 18:03:54 +08001435 if (iovec != inline_vecs)
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001436 kfree(iovec);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001437 return ret;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001438 }
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001439
1440 req->ki_nbytes = ret;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001441
Kent Overstreet73a70752013-05-09 15:03:42 -07001442 /* XXX: move/kill - rw_verify_area()? */
1443 /* This matches the pread()/pwrite() logic */
1444 if (req->ki_pos < 0) {
1445 ret = -EINVAL;
1446 break;
1447 }
1448
1449 if (rw == WRITE)
1450 file_start_write(file);
1451
Al Viro293bc982014-02-11 18:37:41 -05001452 if (iter_op) {
1453 iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes);
1454 ret = iter_op(req, &iter);
1455 } else {
1456 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1457 }
Kent Overstreet73a70752013-05-09 15:03:42 -07001458
1459 if (rw == WRITE)
1460 file_end_write(file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001462
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463 case IOCB_CMD_FDSYNC:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001464 if (!file->f_op->aio_fsync)
1465 return -EINVAL;
1466
1467 ret = file->f_op->aio_fsync(req, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001469
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 case IOCB_CMD_FSYNC:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001471 if (!file->f_op->aio_fsync)
1472 return -EINVAL;
1473
1474 ret = file->f_op->aio_fsync(req, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001476
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 default:
Kent Overstreetcaf41672013-05-07 16:18:35 -07001478 pr_debug("EINVAL: no operation provided\n");
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001479 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 }
1481
Gu Zheng00fefb92014-07-23 18:03:54 +08001482 if (iovec != inline_vecs)
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001483 kfree(iovec);
1484
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001485 if (ret != -EIOCBQUEUED) {
1486 /*
1487 * There's no easy way to restart the syscall since other AIO's
1488 * may be already running. Just fail this IO with EINTR.
1489 */
1490 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1491 ret == -ERESTARTNOHAND ||
1492 ret == -ERESTART_RESTARTBLOCK))
1493 ret = -EINTR;
1494 aio_complete(req, ret, 0);
1495 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496
1497 return 0;
1498}
1499
Adrian Bunkd5470b52008-04-29 00:58:57 -07001500static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001501 struct iocb *iocb, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502{
1503 struct kiocb *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504 ssize_t ret;
1505
1506 /* enforce forwards compatibility on users */
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001507 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001508 pr_debug("EINVAL: reserve field set\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 return -EINVAL;
1510 }
1511
1512 /* prevent overflows */
1513 if (unlikely(
1514 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1515 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1516 ((ssize_t)iocb->aio_nbytes < 0)
1517 )) {
1518 pr_debug("EINVAL: io_submit: overflow check\n");
1519 return -EINVAL;
1520 }
1521
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001522 req = aio_get_req(ctx);
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001523 if (unlikely(!req))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 return -EAGAIN;
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001525
1526 req->ki_filp = fget(iocb->aio_fildes);
1527 if (unlikely(!req->ki_filp)) {
1528 ret = -EBADF;
1529 goto out_put_req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 }
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001531
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001532 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1533 /*
1534 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1535 * instance of the file* now. The file descriptor must be
1536 * an eventfd() fd, and will be signaled for each completed
1537 * event using the eventfd_signal() function.
1538 */
Davide Libenzi13389012009-06-30 11:41:11 -07001539 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
Hirofumi Nakagawa801678c2008-04-29 01:03:09 -07001540 if (IS_ERR(req->ki_eventfd)) {
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001541 ret = PTR_ERR(req->ki_eventfd);
Davide Libenzi87c3a862009-03-18 17:04:19 -07001542 req->ki_eventfd = NULL;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001543 goto out_put_req;
1544 }
1545 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546
Kent Overstreet8a660892013-05-07 16:19:10 -07001547 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548 if (unlikely(ret)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001549 pr_debug("EFAULT: aio_key\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550 goto out_put_req;
1551 }
1552
1553 req->ki_obj.user = user_iocb;
1554 req->ki_user_data = iocb->aio_data;
1555 req->ki_pos = iocb->aio_offset;
Kent Overstreet73a70752013-05-09 15:03:42 -07001556 req->ki_nbytes = iocb->aio_nbytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001558 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1559 (char __user *)(unsigned long)iocb->aio_buf,
1560 compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 if (ret)
1562 goto out_put_req;
1563
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565out_put_req:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001566 put_reqs_available(ctx, 1);
Kent Overstreete34ecee2013-10-10 19:31:47 -07001567 percpu_ref_put(&ctx->reqs);
Kent Overstreet57282d82013-05-13 13:42:52 -07001568 kiocb_free(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 return ret;
1570}
1571
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001572long do_io_submit(aio_context_t ctx_id, long nr,
1573 struct iocb __user *__user *iocbpp, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574{
1575 struct kioctx *ctx;
1576 long ret = 0;
Jeff Moyer080d6762011-11-02 13:40:10 -07001577 int i = 0;
Shaohua Li9f5b9422010-07-01 07:55:01 +02001578 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579
1580 if (unlikely(nr < 0))
1581 return -EINVAL;
1582
Jeff Moyer75e1c702010-09-10 14:16:00 -07001583 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1584 nr = LONG_MAX/sizeof(*iocbpp);
1585
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1587 return -EFAULT;
1588
1589 ctx = lookup_ioctx(ctx_id);
1590 if (unlikely(!ctx)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001591 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592 return -EINVAL;
1593 }
1594
Shaohua Li9f5b9422010-07-01 07:55:01 +02001595 blk_start_plug(&plug);
1596
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597 /*
1598 * AKPM: should this return a partial result if some of the IOs were
1599 * successfully submitted?
1600 */
1601 for (i=0; i<nr; i++) {
1602 struct iocb __user *user_iocb;
1603 struct iocb tmp;
1604
1605 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1606 ret = -EFAULT;
1607 break;
1608 }
1609
1610 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1611 ret = -EFAULT;
1612 break;
1613 }
1614
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001615 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616 if (ret)
1617 break;
1618 }
Shaohua Li9f5b9422010-07-01 07:55:01 +02001619 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620
Kent Overstreet723be6e2013-05-28 15:14:48 -07001621 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622 return i ? i : ret;
1623}
1624
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001625/* sys_io_submit:
1626 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1627 * the number of iocbs queued. May return -EINVAL if the aio_context
1628 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1629 * *iocbpp[0] is not properly initialized, if the operation specified
1630 * is invalid for the file descriptor in the iocb. May fail with
1631 * -EFAULT if any of the data structures point to invalid data. May
1632 * fail with -EBADF if the file descriptor specified in the first
1633 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1634 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1635 * fail with -ENOSYS if not implemented.
1636 */
1637SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1638 struct iocb __user * __user *, iocbpp)
1639{
1640 return do_io_submit(ctx_id, nr, iocbpp, 0);
1641}
1642
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643/* lookup_kiocb
1644 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001646static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1647 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648{
1649 struct list_head *pos;
Zach Brownd00689a2005-11-13 16:07:34 -08001650
1651 assert_spin_locked(&ctx->ctx_lock);
1652
Kent Overstreet8a660892013-05-07 16:19:10 -07001653 if (key != KIOCB_KEY)
1654 return NULL;
1655
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 /* TODO: use a hash or array, this sucks. */
1657 list_for_each(pos, &ctx->active_reqs) {
1658 struct kiocb *kiocb = list_kiocb(pos);
Kent Overstreet8a660892013-05-07 16:19:10 -07001659 if (kiocb->ki_obj.user == iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660 return kiocb;
1661 }
1662 return NULL;
1663}
1664
1665/* sys_io_cancel:
1666 * Attempts to cancel an iocb previously passed to io_submit. If
1667 * the operation is successfully cancelled, the resulting event is
1668 * copied into the memory pointed to by result without being placed
1669 * into the completion queue and 0 is returned. May fail with
1670 * -EFAULT if any of the data structures pointed to are invalid.
1671 * May fail with -EINVAL if aio_context specified by ctx_id is
1672 * invalid. May fail with -EAGAIN if the iocb specified was not
1673 * cancelled. Will fail with -ENOSYS if not implemented.
1674 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001675SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1676 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 struct kioctx *ctx;
1679 struct kiocb *kiocb;
1680 u32 key;
1681 int ret;
1682
1683 ret = get_user(key, &iocb->aio_key);
1684 if (unlikely(ret))
1685 return -EFAULT;
1686
1687 ctx = lookup_ioctx(ctx_id);
1688 if (unlikely(!ctx))
1689 return -EINVAL;
1690
1691 spin_lock_irq(&ctx->ctx_lock);
Kent Overstreet906b9732013-05-07 16:18:31 -07001692
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693 kiocb = lookup_kiocb(ctx, iocb, key);
Kent Overstreet906b9732013-05-07 16:18:31 -07001694 if (kiocb)
Fabian Frederickd52a8f92014-04-22 07:26:58 +02001695 ret = kiocb_cancel(kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -07001696 else
1697 ret = -EINVAL;
1698
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 spin_unlock_irq(&ctx->ctx_lock);
1700
Kent Overstreet906b9732013-05-07 16:18:31 -07001701 if (!ret) {
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001702 /*
1703 * The result argument is no longer used - the io_event is
1704 * always delivered via the ring buffer. -EINPROGRESS indicates
1705 * cancellation is progress:
Kent Overstreet906b9732013-05-07 16:18:31 -07001706 */
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001707 ret = -EINPROGRESS;
Kent Overstreet906b9732013-05-07 16:18:31 -07001708 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709
Kent Overstreet723be6e2013-05-28 15:14:48 -07001710 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711
1712 return ret;
1713}
1714
1715/* io_getevents:
1716 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001717 * the completion queue for the aio_context specified by ctx_id. If
1718 * it succeeds, the number of read events is returned. May fail with
1719 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1720 * out of range, if timeout is out of range. May fail with -EFAULT
1721 * if any of the memory specified is invalid. May return 0 or
1722 * < min_nr if the timeout specified by timeout has elapsed
1723 * before sufficient events are available, where timeout == NULL
1724 * specifies an infinite timeout. Note that the timeout pointed to by
Jeff Moyer69008072013-05-24 15:55:24 -07001725 * timeout is relative. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001727SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1728 long, min_nr,
1729 long, nr,
1730 struct io_event __user *, events,
1731 struct timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732{
1733 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1734 long ret = -EINVAL;
1735
1736 if (likely(ioctx)) {
Namhyung Kim2e410252011-01-12 17:01:08 -08001737 if (likely(min_nr <= nr && min_nr >= 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738 ret = read_events(ioctx, min_nr, nr, events, timeout);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001739 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 return ret;
1742}