blob: 1f32da13d39ee603cfd42e3d2af8f318f8bb0ea3 [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.
Christoph Hellwigbfe40372018-07-16 09:08:20 +02008 * Copyright 2018 Christoph Hellwig.
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 *
10 * See ../COPYING for licensing terms.
11 */
Kent Overstreetcaf41672013-05-07 16:18:35 -070012#define pr_fmt(fmt) "%s: " fmt, __func__
13
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/kernel.h>
15#include <linux/init.h>
16#include <linux/errno.h>
17#include <linux/time.h>
18#include <linux/aio_abi.h>
Paul Gortmaker630d9c42011-11-16 23:57:37 -050019#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#include <linux/syscalls.h>
Jens Axboeb9d128f2009-10-29 13:59:26 +010021#include <linux/backing-dev.h>
Christoph Hellwig9018ccc2018-07-24 11:36:37 +020022#include <linux/refcount.h>
Badari Pulavarty027445c2006-09-30 23:28:46 -070023#include <linux/uio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024
Ingo Molnar174cd4b2017-02-02 19:15:33 +010025#include <linux/sched/signal.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/fs.h>
27#include <linux/file.h>
28#include <linux/mm.h>
29#include <linux/mman.h>
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100030#include <linux/percpu.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031#include <linux/slab.h>
32#include <linux/timer.h>
33#include <linux/aio.h>
34#include <linux/highmem.h>
35#include <linux/workqueue.h>
36#include <linux/security.h>
Davide Libenzi9c3060b2007-05-10 22:23:21 -070037#include <linux/eventfd.h>
Jeff Moyercfb1e332009-10-02 18:57:36 -040038#include <linux/blkdev.h>
Jeff Moyer9d85cba72010-05-26 14:44:26 -070039#include <linux/compat.h>
Gu Zheng36bc08c2013-07-16 17:56:16 +080040#include <linux/migrate.h>
41#include <linux/ramfs.h>
Kent Overstreet723be6e2013-05-28 15:14:48 -070042#include <linux/percpu-refcount.h>
Benjamin LaHaise71ad7492013-09-17 10:18:25 -040043#include <linux/mount.h>
David Howells52db59d2019-03-25 16:38:23 +000044#include <linux/pseudo_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
Linus Torvalds7c0f6ba2016-12-24 11:46:01 -080046#include <linux/uaccess.h>
Jeff Moyera538e3f2018-12-11 12:37:49 -050047#include <linux/nospec.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070048
Al Viro68d70d02013-06-19 15:26:04 +040049#include "internal.h"
50
Christoph Hellwigf3a27522018-03-30 11:19:25 +020051#define KIOCB_KEY 0
52
Kent Overstreet4e179bc2013-05-07 16:18:33 -070053#define AIO_RING_MAGIC 0xa10a10a1
54#define AIO_RING_COMPAT_FEATURES 1
55#define AIO_RING_INCOMPAT_FEATURES 0
56struct aio_ring {
57 unsigned id; /* kernel internal index number */
58 unsigned nr; /* number of io_events */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -040059 unsigned head; /* Written to by userland or under ring_lock
60 * mutex by aio_read_events_ring(). */
Kent Overstreet4e179bc2013-05-07 16:18:33 -070061 unsigned tail;
62
63 unsigned magic;
64 unsigned compat_features;
65 unsigned incompat_features;
66 unsigned header_length; /* size of aio_ring */
67
68
Gustavo A. R. Silva241cb282020-05-28 09:35:11 -050069 struct io_event io_events[];
Kent Overstreet4e179bc2013-05-07 16:18:33 -070070}; /* 128 bytes + ring size */
71
Jens Axboea79d40e2018-12-04 09:45:32 -070072/*
73 * Plugging is meant to work with larger batches of IOs. If we don't
74 * have more than the below, then don't bother setting up a plug.
75 */
76#define AIO_PLUG_THRESHOLD 2
77
Kent Overstreet4e179bc2013-05-07 16:18:33 -070078#define AIO_RING_PAGES 8
Kent Overstreet4e179bc2013-05-07 16:18:33 -070079
Benjamin LaHaisedb446a02013-07-30 12:54:40 -040080struct kioctx_table {
Tejun Heod0264c02018-03-14 12:10:17 -070081 struct rcu_head rcu;
82 unsigned nr;
83 struct kioctx __rcu *table[];
Benjamin LaHaisedb446a02013-07-30 12:54:40 -040084};
85
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100086struct kioctx_cpu {
87 unsigned reqs_available;
88};
89
Jens Axboedc48e562015-04-15 11:17:23 -060090struct ctx_rq_wait {
91 struct completion comp;
92 atomic_t count;
93};
94
Kent Overstreet4e179bc2013-05-07 16:18:33 -070095struct kioctx {
Kent Overstreet723be6e2013-05-28 15:14:48 -070096 struct percpu_ref users;
Kent Overstreet36f55882013-05-07 16:18:41 -070097 atomic_t dead;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070098
Kent Overstreete34ecee2013-10-10 19:31:47 -070099 struct percpu_ref reqs;
100
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700101 unsigned long user_id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700102
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000103 struct __percpu kioctx_cpu *cpu;
104
105 /*
106 * For percpu reqs_available, number of slots we move to/from global
107 * counter at a time:
108 */
109 unsigned req_batch;
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700110 /*
111 * This is what userspace passed to io_setup(), it's not used for
112 * anything but counting against the global max_reqs quota.
113 *
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700114 * The real limit is nr_events - 1, which will be larger (see
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700115 * aio_setup_ring())
116 */
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700117 unsigned max_reqs;
118
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700119 /* Size of ringbuffer, in units of struct io_event */
120 unsigned nr_events;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700121
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700122 unsigned long mmap_base;
123 unsigned long mmap_size;
124
125 struct page **ring_pages;
126 long nr_pages;
127
Tejun Heof7298632018-03-14 12:45:15 -0700128 struct rcu_work free_rwork; /* see free_ioctx() */
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700129
Anatol Pomozove02ba722014-04-15 11:31:33 -0700130 /*
131 * signals when all in-flight requests are done
132 */
Jens Axboedc48e562015-04-15 11:17:23 -0600133 struct ctx_rq_wait *rq_wait;
Anatol Pomozove02ba722014-04-15 11:31:33 -0700134
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700135 struct {
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000136 /*
137 * This counts the number of available slots in the ringbuffer,
138 * so we avoid overflowing it: it's decremented (if positive)
139 * when allocating a kiocb and incremented when the resulting
140 * io_event is pulled off the ringbuffer.
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000141 *
142 * We batch accesses to it with a percpu version.
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000143 */
144 atomic_t reqs_available;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700145 } ____cacheline_aligned_in_smp;
146
147 struct {
148 spinlock_t ctx_lock;
149 struct list_head active_reqs; /* used for cancellation */
150 } ____cacheline_aligned_in_smp;
151
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700152 struct {
153 struct mutex ring_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700154 wait_queue_head_t wait;
155 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700156
157 struct {
158 unsigned tail;
Benjamin LaHaised856f322014-08-24 13:14:05 -0400159 unsigned completed_events;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700160 spinlock_t completion_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700161 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700162
163 struct page *internal_pages[AIO_RING_PAGES];
Gu Zheng36bc08c2013-07-16 17:56:16 +0800164 struct file *aio_ring_file;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400165
166 unsigned id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700167};
168
Linus Torvalds84c4e1f82019-03-03 14:23:33 -0800169/*
170 * First field must be the file pointer in all the
171 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
172 */
Christoph Hellwiga3c0d432018-03-27 19:18:57 +0200173struct fsync_iocb {
Christoph Hellwiga3c0d432018-03-27 19:18:57 +0200174 struct file *file;
Linus Torvalds84c4e1f82019-03-03 14:23:33 -0800175 struct work_struct work;
Christoph Hellwiga3c0d432018-03-27 19:18:57 +0200176 bool datasync;
Miklos Szeredi530f32f2020-05-14 16:44:24 +0200177 struct cred *creds;
Christoph Hellwiga3c0d432018-03-27 19:18:57 +0200178};
179
Christoph Hellwigbfe40372018-07-16 09:08:20 +0200180struct poll_iocb {
181 struct file *file;
182 struct wait_queue_head *head;
183 __poll_t events;
Al Viroaf5c72b2019-03-07 21:45:41 -0500184 bool done;
Christoph Hellwigbfe40372018-07-16 09:08:20 +0200185 bool cancelled;
186 struct wait_queue_entry wait;
187 struct work_struct work;
188};
189
Linus Torvalds84c4e1f82019-03-03 14:23:33 -0800190/*
191 * NOTE! Each of the iocb union members has the file pointer
192 * as the first entry in their struct definition. So you can
193 * access the file pointer through any of the sub-structs,
194 * or directly as just 'ki_filp' in this struct.
195 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100196struct aio_kiocb {
Christoph Hellwig54843f82018-05-02 19:57:21 +0200197 union {
Linus Torvalds84c4e1f82019-03-03 14:23:33 -0800198 struct file *ki_filp;
Christoph Hellwig54843f82018-05-02 19:57:21 +0200199 struct kiocb rw;
Christoph Hellwiga3c0d432018-03-27 19:18:57 +0200200 struct fsync_iocb fsync;
Christoph Hellwigbfe40372018-07-16 09:08:20 +0200201 struct poll_iocb poll;
Christoph Hellwig54843f82018-05-02 19:57:21 +0200202 };
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100203
204 struct kioctx *ki_ctx;
205 kiocb_cancel_fn *ki_cancel;
206
Al Viroa9339b72019-03-07 19:43:45 -0500207 struct io_event ki_res;
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100208
209 struct list_head ki_list; /* the aio core uses this
210 * for cancellation */
Christoph Hellwig9018ccc2018-07-24 11:36:37 +0200211 refcount_t ki_refcnt;
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100212
213 /*
214 * If the aio_resfd field of the userspace iocb is not zero,
215 * this is the underlying eventfd context to deliver events to.
216 */
217 struct eventfd_ctx *ki_eventfd;
218};
219
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -0800221static DEFINE_SPINLOCK(aio_nr_lock);
222unsigned long aio_nr; /* current system wide number of aio requests */
223unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224/*----end sysctl variables---*/
225
Christoph Lametere18b8902006-12-06 20:33:20 -0800226static struct kmem_cache *kiocb_cachep;
227static struct kmem_cache *kioctx_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400229static struct vfsmount *aio_mnt;
230
231static const struct file_operations aio_ring_fops;
232static const struct address_space_operations aio_ctx_aops;
233
234static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
235{
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400236 struct file *file;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400237 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
Dan Carpenter7f626562013-11-13 10:49:40 +0300238 if (IS_ERR(inode))
239 return ERR_CAST(inode);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400240
241 inode->i_mapping->a_ops = &aio_ctx_aops;
242 inode->i_mapping->private_data = ctx;
243 inode->i_size = PAGE_SIZE * nr_pages;
244
Al Virod93aa9d2018-06-09 09:40:05 -0400245 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
246 O_RDWR, &aio_ring_fops);
Al Viroc9c554f2018-07-11 14:19:04 -0400247 if (IS_ERR(file))
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400248 iput(inode);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400249 return file;
250}
251
David Howells52db59d2019-03-25 16:38:23 +0000252static int aio_init_fs_context(struct fs_context *fc)
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400253{
David Howells52db59d2019-03-25 16:38:23 +0000254 if (!init_pseudo(fc, AIO_RING_MAGIC))
255 return -ENOMEM;
256 fc->s_iflags |= SB_I_NOEXEC;
257 return 0;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400258}
259
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260/* aio_setup
261 * Creates the slab caches used by the aio routines, panic on
262 * failure as this is done early during the boot sequence.
263 */
264static int __init aio_setup(void)
265{
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400266 static struct file_system_type aio_fs = {
267 .name = "aio",
David Howells52db59d2019-03-25 16:38:23 +0000268 .init_fs_context = aio_init_fs_context,
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400269 .kill_sb = kill_anon_super,
270 };
271 aio_mnt = kern_mount(&aio_fs);
272 if (IS_ERR(aio_mnt))
273 panic("Failed to create aio fs mount.");
274
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100275 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Christoph Lameter0a31bd52007-05-06 14:49:57 -0700276 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277 return 0;
278}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700279__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400281static void put_aio_ring_file(struct kioctx *ctx)
282{
283 struct file *aio_ring_file = ctx->aio_ring_file;
Rasmus Villemoesde04e762016-09-15 00:25:03 +0200284 struct address_space *i_mapping;
285
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400286 if (aio_ring_file) {
Al Viro45063092016-12-04 18:24:56 -0500287 truncate_setsize(file_inode(aio_ring_file), 0);
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400288
289 /* Prevent further access to the kioctx from migratepages */
Al Viro45063092016-12-04 18:24:56 -0500290 i_mapping = aio_ring_file->f_mapping;
Rasmus Villemoesde04e762016-09-15 00:25:03 +0200291 spin_lock(&i_mapping->private_lock);
292 i_mapping->private_data = NULL;
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400293 ctx->aio_ring_file = NULL;
Rasmus Villemoesde04e762016-09-15 00:25:03 +0200294 spin_unlock(&i_mapping->private_lock);
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400295
296 fput(aio_ring_file);
297 }
298}
299
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300static void aio_free_ring(struct kioctx *ctx)
301{
Gu Zheng36bc08c2013-07-16 17:56:16 +0800302 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400304 /* Disconnect the kiotx from the ring file. This prevents future
305 * accesses to the kioctx from page migration.
306 */
307 put_aio_ring_file(ctx);
308
Gu Zheng36bc08c2013-07-16 17:56:16 +0800309 for (i = 0; i < ctx->nr_pages; i++) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500310 struct page *page;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800311 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
312 page_count(ctx->ring_pages[i]));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500313 page = ctx->ring_pages[i];
314 if (!page)
315 continue;
316 ctx->ring_pages[i] = NULL;
317 put_page(page);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800318 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319
Sasha Levinddb8c452013-11-19 17:33:03 -0500320 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700321 kfree(ctx->ring_pages);
Sasha Levinddb8c452013-11-19 17:33:03 -0500322 ctx->ring_pages = NULL;
323 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324}
325
Dmitry Safonovcd544fd2020-12-14 19:08:13 -0800326static int aio_ring_mremap(struct vm_area_struct *vma, unsigned long flags)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800327{
Oleg Nesterov5477e702015-09-04 15:48:04 -0700328 struct file *file = vma->vm_file;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400329 struct mm_struct *mm = vma->vm_mm;
330 struct kioctx_table *table;
Al Virob2edffd2015-04-06 17:48:54 -0400331 int i, res = -EINVAL;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400332
Dmitry Safonovcd544fd2020-12-14 19:08:13 -0800333 if (flags & MREMAP_DONTUNMAP)
334 return -EINVAL;
335
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400336 spin_lock(&mm->ioctx_lock);
337 rcu_read_lock();
338 table = rcu_dereference(mm->ioctx_table);
339 for (i = 0; i < table->nr; i++) {
340 struct kioctx *ctx;
341
Tejun Heod0264c02018-03-14 12:10:17 -0700342 ctx = rcu_dereference(table->table[i]);
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400343 if (ctx && ctx->aio_ring_file == file) {
Al Virob2edffd2015-04-06 17:48:54 -0400344 if (!atomic_read(&ctx->dead)) {
345 ctx->user_id = ctx->mmap_base = vma->vm_start;
346 res = 0;
347 }
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400348 break;
349 }
350 }
351
352 rcu_read_unlock();
353 spin_unlock(&mm->ioctx_lock);
Al Virob2edffd2015-04-06 17:48:54 -0400354 return res;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400355}
356
Oleg Nesterov5477e702015-09-04 15:48:04 -0700357static const struct vm_operations_struct aio_ring_vm_ops = {
358 .mremap = aio_ring_mremap,
359#if IS_ENABLED(CONFIG_MMU)
360 .fault = filemap_fault,
361 .map_pages = filemap_map_pages,
362 .page_mkwrite = filemap_page_mkwrite,
363#endif
364};
365
366static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
367{
368 vma->vm_flags |= VM_DONTEXPAND;
369 vma->vm_ops = &aio_ring_vm_ops;
370 return 0;
371}
372
Gu Zheng36bc08c2013-07-16 17:56:16 +0800373static const struct file_operations aio_ring_fops = {
374 .mmap = aio_ring_mmap,
375};
376
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400377#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800378static int aio_migratepage(struct address_space *mapping, struct page *new,
379 struct page *old, enum migrate_mode mode)
380{
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400381 struct kioctx *ctx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800382 unsigned long flags;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400383 pgoff_t idx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800384 int rc;
385
Jérôme Glisse2916ecc2017-09-08 16:12:06 -0700386 /*
387 * We cannot support the _NO_COPY case here, because copy needs to
388 * happen under the ctx->completion_lock. That does not work with the
389 * migration workflow of MIGRATE_SYNC_NO_COPY.
390 */
391 if (mode == MIGRATE_SYNC_NO_COPY)
392 return -EINVAL;
393
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500394 rc = 0;
395
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400396 /* mapping->private_lock here protects against the kioctx teardown. */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500397 spin_lock(&mapping->private_lock);
398 ctx = mapping->private_data;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400399 if (!ctx) {
400 rc = -EINVAL;
401 goto out;
402 }
403
404 /* The ring_lock mutex. The prevents aio_read_events() from writing
405 * to the ring's head, and prevents page migration from mucking in
406 * a partially initialized kiotx.
407 */
408 if (!mutex_trylock(&ctx->ring_lock)) {
409 rc = -EAGAIN;
410 goto out;
411 }
412
413 idx = old->index;
414 if (idx < (pgoff_t)ctx->nr_pages) {
415 /* Make sure the old page hasn't already been changed */
416 if (ctx->ring_pages[idx] != old)
417 rc = -EAGAIN;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500418 } else
419 rc = -EINVAL;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500420
421 if (rc != 0)
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400422 goto out_unlock;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500423
Gu Zheng36bc08c2013-07-16 17:56:16 +0800424 /* Writeback must be complete */
425 BUG_ON(PageWriteback(old));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500426 get_page(new);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800427
Keith Busch37109692019-07-18 15:58:46 -0700428 rc = migrate_page_move_mapping(mapping, new, old, 1);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800429 if (rc != MIGRATEPAGE_SUCCESS) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500430 put_page(new);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400431 goto out_unlock;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800432 }
433
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400434 /* Take completion_lock to prevent other writes to the ring buffer
435 * while the old page is copied to the new. This prevents new
436 * events from being lost.
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400437 */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400438 spin_lock_irqsave(&ctx->completion_lock, flags);
439 migrate_page_copy(new, old);
440 BUG_ON(ctx->ring_pages[idx] != old);
441 ctx->ring_pages[idx] = new;
442 spin_unlock_irqrestore(&ctx->completion_lock, flags);
443
444 /* The old page is no longer accessible. */
445 put_page(old);
446
447out_unlock:
448 mutex_unlock(&ctx->ring_lock);
449out:
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400450 spin_unlock(&mapping->private_lock);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800451 return rc;
452}
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400453#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800454
455static const struct address_space_operations aio_ctx_aops = {
Gu Zheng835f2522014-11-06 17:46:21 +0800456 .set_page_dirty = __set_page_dirty_no_writeback,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400457#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800458 .migratepage = aio_migratepage,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400459#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800460};
461
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300462static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463{
464 struct aio_ring *ring;
Zach Brown41003a72013-05-07 16:18:25 -0700465 struct mm_struct *mm = current->mm;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900466 unsigned long size, unused;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 int nr_pages;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800468 int i;
469 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470
471 /* Compensate for the ring buffer's head/tail overlap entry */
472 nr_events += 2; /* 1 is required, 2 for good luck */
473
474 size = sizeof(struct aio_ring);
475 size += sizeof(struct io_event) * nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476
Gu Zheng36bc08c2013-07-16 17:56:16 +0800477 nr_pages = PFN_UP(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 if (nr_pages < 0)
479 return -EINVAL;
480
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400481 file = aio_private_file(ctx, nr_pages);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800482 if (IS_ERR(file)) {
483 ctx->aio_ring_file = NULL;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400484 return -ENOMEM;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800485 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486
Gu Zheng36bc08c2013-07-16 17:56:16 +0800487 ctx->aio_ring_file = file;
488 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
489 / sizeof(struct io_event);
490
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700491 ctx->ring_pages = ctx->internal_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 if (nr_pages > AIO_RING_PAGES) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700493 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
494 GFP_KERNEL);
Gu Zhengd1b94322013-12-04 18:19:06 +0800495 if (!ctx->ring_pages) {
496 put_aio_ring_file(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 return -ENOMEM;
Gu Zhengd1b94322013-12-04 18:19:06 +0800498 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 }
500
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900501 for (i = 0; i < nr_pages; i++) {
502 struct page *page;
Al Viro45063092016-12-04 18:24:56 -0500503 page = find_or_create_page(file->f_mapping,
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900504 i, GFP_HIGHUSER | __GFP_ZERO);
505 if (!page)
506 break;
507 pr_debug("pid(%d) page[%d]->count=%d\n",
508 current->pid, i, page_count(page));
509 SetPageUptodate(page);
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900510 unlock_page(page);
511
512 ctx->ring_pages[i] = page;
513 }
514 ctx->nr_pages = i;
515
516 if (unlikely(i != nr_pages)) {
517 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400518 return -ENOMEM;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900519 }
520
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700521 ctx->mmap_size = nr_pages * PAGE_SIZE;
522 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800523
Michel Lespinassed8ed45c2020-06-08 21:33:25 -0700524 if (mmap_write_lock_killable(mm)) {
Michal Hocko013373e2016-05-23 16:25:59 -0700525 ctx->mmap_size = 0;
526 aio_free_ring(ctx);
527 return -EINTR;
528 }
529
Peter Collingbourne45e55302020-08-06 23:23:37 -0700530 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
531 PROT_READ | PROT_WRITE,
532 MAP_SHARED, 0, &unused, NULL);
Michel Lespinassed8ed45c2020-06-08 21:33:25 -0700533 mmap_write_unlock(mm);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700534 if (IS_ERR((void *)ctx->mmap_base)) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700535 ctx->mmap_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400537 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 }
539
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700540 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400541
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700542 ctx->user_id = ctx->mmap_base;
543 ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700545 ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 ring->nr = nr_events; /* user copy */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400547 ring->id = ~0U;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 ring->head = ring->tail = 0;
549 ring->magic = AIO_RING_MAGIC;
550 ring->compat_features = AIO_RING_COMPAT_FEATURES;
551 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
552 ring->header_length = sizeof(struct aio_ring);
Cong Wange8e3c3d2011-11-25 23:14:27 +0800553 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700554 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555
556 return 0;
557}
558
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
560#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
561#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
562
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100563void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
Kent Overstreet0460fef2013-05-07 16:18:49 -0700564{
Christoph Hellwig54843f82018-05-02 19:57:21 +0200565 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
Kent Overstreet0460fef2013-05-07 16:18:49 -0700566 struct kioctx *ctx = req->ki_ctx;
567 unsigned long flags;
568
Christoph Hellwig75321b52018-04-09 14:57:56 +0200569 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
570 return;
571
Kent Overstreet0460fef2013-05-07 16:18:49 -0700572 spin_lock_irqsave(&ctx->ctx_lock, flags);
Christoph Hellwig75321b52018-04-09 14:57:56 +0200573 list_add_tail(&req->ki_list, &ctx->active_reqs);
Kent Overstreet0460fef2013-05-07 16:18:49 -0700574 req->ki_cancel = cancel;
Kent Overstreet0460fef2013-05-07 16:18:49 -0700575 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
576}
577EXPORT_SYMBOL(kiocb_set_cancel_fn);
578
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700579/*
580 * free_ioctx() should be RCU delayed to synchronize against the RCU
581 * protected lookup_ioctx() and also needs process context to call
Tejun Heof7298632018-03-14 12:45:15 -0700582 * aio_free_ring(). Use rcu_work.
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700583 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700584static void free_ioctx(struct work_struct *work)
Kent Overstreet36f55882013-05-07 16:18:41 -0700585{
Tejun Heof7298632018-03-14 12:45:15 -0700586 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
587 free_rwork);
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
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700605 /* Synchronize against RCU protected table->table[] dereferences */
Tejun Heof7298632018-03-14 12:45:15 -0700606 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
607 queue_rcu_work(system_wq, &ctx->free_rwork);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700608}
609
Kent Overstreet36f55882013-05-07 16:18:41 -0700610/*
611 * When this function runs, the kioctx has been removed from the "hash table"
612 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
613 * now it's safe to cancel any that need to be.
614 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700615static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet36f55882013-05-07 16:18:41 -0700616{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700617 struct kioctx *ctx = container_of(ref, struct kioctx, users);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100618 struct aio_kiocb *req;
Kent Overstreet36f55882013-05-07 16:18:41 -0700619
620 spin_lock_irq(&ctx->ctx_lock);
621
622 while (!list_empty(&ctx->active_reqs)) {
623 req = list_first_entry(&ctx->active_reqs,
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100624 struct aio_kiocb, ki_list);
Christoph Hellwig888933f2018-05-23 14:11:02 +0200625 req->ki_cancel(&req->rw);
Al Viro4faa99962018-05-23 22:53:22 -0400626 list_del_init(&req->ki_list);
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++)
Tejun Heod0264c02018-03-14 12:10:17 -0700647 if (!rcu_access_pointer(table->table[i])) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400648 ctx->id = i;
Tejun Heod0264c02018-03-14 12:10:17 -0700649 rcu_assign_pointer(table->table[i], ctx);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400650 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 /*
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300710 * Store the original nr_events -- what userspace passed to io_setup(),
711 * for counting against the global limit -- before it changes.
712 */
713 unsigned int max_reqs = nr_events;
714
715 /*
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000716 * We keep track of the number of available ringbuffer slots, to prevent
717 * overflow (reqs_available), and we also use percpu counters for this.
718 *
719 * So since up to half the slots might be on other cpu's percpu counters
720 * and unavailable, double nr_events so userspace sees what they
721 * expected: additionally, we move req_batch slots to/from percpu
722 * counters at a time, so make sure that isn't 0:
723 */
724 nr_events = max(nr_events, num_possible_cpus() * 4);
725 nr_events *= 2;
726
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 /* Prevent overflows */
Al Viro08397ac2015-03-31 11:43:52 -0400728 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 pr_debug("ENOMEM: nr_events too high\n");
730 return ERR_PTR(-EINVAL);
731 }
732
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300733 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734 return ERR_PTR(-EAGAIN);
735
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800736 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 if (!ctx)
738 return ERR_PTR(-ENOMEM);
739
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300740 ctx->max_reqs = max_reqs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400742 spin_lock_init(&ctx->ctx_lock);
743 spin_lock_init(&ctx->completion_lock);
744 mutex_init(&ctx->ring_lock);
745 /* Protect against page migration throughout kiotx setup by keeping
746 * the ring_lock mutex held until setup is complete. */
747 mutex_lock(&ctx->ring_lock);
748 init_waitqueue_head(&ctx->wait);
749
750 INIT_LIST_HEAD(&ctx->active_reqs);
751
Tejun Heo2aad2a82014-09-24 13:31:50 -0400752 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700753 goto err;
754
Tejun Heo2aad2a82014-09-24 13:31:50 -0400755 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700756 goto err;
Kent Overstreet723be6e2013-05-28 15:14:48 -0700757
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000758 ctx->cpu = alloc_percpu(struct kioctx_cpu);
759 if (!ctx->cpu)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700760 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300762 err = aio_setup_ring(ctx, nr_events);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400763 if (err < 0)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700764 goto err;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000765
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000766 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000767 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
Benjamin LaHaise6878ea72013-07-31 10:34:18 -0400768 if (ctx->req_batch < 1)
769 ctx->req_batch = 1;
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000770
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771 /* limit the number of system wide aios */
Al Viro9fa1cb32012-03-10 23:14:05 -0500772 spin_lock(&aio_nr_lock);
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300773 if (aio_nr + ctx->max_reqs > aio_max_nr ||
774 aio_nr + ctx->max_reqs < aio_nr) {
Al Viro9fa1cb32012-03-10 23:14:05 -0500775 spin_unlock(&aio_nr_lock);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700776 err = -EAGAIN;
Gu Zhengd1b94322013-12-04 18:19:06 +0800777 goto err_ctx;
Al Viro2dd542b72012-03-10 23:10:35 -0500778 }
779 aio_nr += ctx->max_reqs;
Al Viro9fa1cb32012-03-10 23:14:05 -0500780 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781
Benjamin LaHaise18816862013-12-21 15:49:28 -0500782 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
783 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
Kent Overstreet723be6e2013-05-28 15:14:48 -0700784
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400785 err = ioctx_add_table(ctx, mm);
786 if (err)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700787 goto err_cleanup;
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400788
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400789 /* Release the ring_lock mutex now that all setup is complete. */
790 mutex_unlock(&ctx->ring_lock);
791
Kent Overstreetcaf41672013-05-07 16:18:35 -0700792 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700793 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 return ctx;
795
Kent Overstreete34ecee2013-10-10 19:31:47 -0700796err_cleanup:
797 aio_nr_sub(ctx->max_reqs);
Gu Zhengd1b94322013-12-04 18:19:06 +0800798err_ctx:
Al Virodeeb8522015-04-06 17:57:44 -0400799 atomic_set(&ctx->dead, 1);
800 if (ctx->mmap_size)
801 vm_munmap(ctx->mmap_base, ctx->mmap_size);
Gu Zhengd1b94322013-12-04 18:19:06 +0800802 aio_free_ring(ctx);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700803err:
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400804 mutex_unlock(&ctx->ring_lock);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000805 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400806 percpu_ref_exit(&ctx->reqs);
807 percpu_ref_exit(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreetcaf41672013-05-07 16:18:35 -0700809 pr_debug("error allocating ioctx %d\n", err);
Al Viroe23754f2012-03-06 14:33:22 -0500810 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811}
812
Kent Overstreet36f55882013-05-07 16:18:41 -0700813/* kill_ioctx
814 * Cancels all outstanding aio requests on an aio context. Used
815 * when the processes owning a context have all exited to encourage
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 * the rapid destruction of the kioctx.
817 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400818static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
Jens Axboedc48e562015-04-15 11:17:23 -0600819 struct ctx_rq_wait *wait)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820{
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400821 struct kioctx_table *table;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400822
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400823 spin_lock(&mm->ioctx_lock);
Al Virob2edffd2015-04-06 17:48:54 -0400824 if (atomic_xchg(&ctx->dead, 1)) {
825 spin_unlock(&mm->ioctx_lock);
826 return -EINVAL;
827 }
828
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200829 table = rcu_dereference_raw(mm->ioctx_table);
Tejun Heod0264c02018-03-14 12:10:17 -0700830 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
831 RCU_INIT_POINTER(table->table[ctx->id], NULL);
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400832 spin_unlock(&mm->ioctx_lock);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700833
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700834 /* free_ioctx_reqs() will do the necessary RCU synchronization */
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400835 wake_up_all(&ctx->wait);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700836
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400837 /*
838 * It'd be more correct to do this in free_ioctx(), after all
839 * the outstanding kiocbs have finished - but by then io_destroy
840 * has already returned, so io_setup() could potentially return
841 * -EAGAIN with no ioctxs actually in use (as far as userspace
842 * could tell).
843 */
844 aio_nr_sub(ctx->max_reqs);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400845
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400846 if (ctx->mmap_size)
847 vm_munmap(ctx->mmap_base, ctx->mmap_size);
848
Jens Axboedc48e562015-04-15 11:17:23 -0600849 ctx->rq_wait = wait;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400850 percpu_ref_kill(&ctx->users);
851 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852}
853
Kent Overstreet36f55882013-05-07 16:18:41 -0700854/*
855 * exit_aio: called when the last user of mm goes away. At this point, there is
856 * no way for any new requests to be submited or any of the io_* syscalls to be
857 * called on the context.
858 *
859 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
860 * them.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800862void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863{
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200864 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
Jens Axboedc48e562015-04-15 11:17:23 -0600865 struct ctx_rq_wait wait;
866 int i, skipped;
Jens Axboeabf137d2008-12-09 08:11:22 +0100867
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200868 if (!table)
869 return;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400870
Jens Axboedc48e562015-04-15 11:17:23 -0600871 atomic_set(&wait.count, table->nr);
872 init_completion(&wait.comp);
873
874 skipped = 0;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200875 for (i = 0; i < table->nr; ++i) {
Tejun Heod0264c02018-03-14 12:10:17 -0700876 struct kioctx *ctx =
877 rcu_dereference_protected(table->table[i], true);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400878
Jens Axboedc48e562015-04-15 11:17:23 -0600879 if (!ctx) {
880 skipped++;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200881 continue;
Jens Axboedc48e562015-04-15 11:17:23 -0600882 }
883
Al Viro936af152012-04-20 21:49:41 -0400884 /*
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200885 * We don't need to bother with munmap() here - exit_mmap(mm)
886 * is coming and it'll unmap everything. And we simply can't,
887 * this is not necessarily our ->mm.
888 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
889 * that it needs to unmap the area, just set it to 0.
Al Viro936af152012-04-20 21:49:41 -0400890 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700891 ctx->mmap_size = 0;
Jens Axboedc48e562015-04-15 11:17:23 -0600892 kill_ioctx(mm, ctx, &wait);
893 }
Kent Overstreet36f55882013-05-07 16:18:41 -0700894
Jens Axboedc48e562015-04-15 11:17:23 -0600895 if (!atomic_sub_and_test(skipped, &wait.count)) {
Gu Zheng6098b452014-09-03 17:45:44 +0800896 /* Wait until all IO for the context are done. */
Jens Axboedc48e562015-04-15 11:17:23 -0600897 wait_for_completion(&wait.comp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 }
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200899
900 RCU_INIT_POINTER(mm->ioctx_table, NULL);
901 kfree(table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902}
903
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000904static void put_reqs_available(struct kioctx *ctx, unsigned nr)
905{
906 struct kioctx_cpu *kcpu;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400907 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000908
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400909 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400910 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000911 kcpu->reqs_available += nr;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400912
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000913 while (kcpu->reqs_available >= ctx->req_batch * 2) {
914 kcpu->reqs_available -= ctx->req_batch;
915 atomic_add(ctx->req_batch, &ctx->reqs_available);
916 }
917
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400918 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000919}
920
Christoph Hellwig432c7992018-11-19 15:57:42 -0700921static bool __get_reqs_available(struct kioctx *ctx)
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000922{
923 struct kioctx_cpu *kcpu;
924 bool ret = false;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400925 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000926
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400927 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400928 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000929 if (!kcpu->reqs_available) {
930 int old, avail = atomic_read(&ctx->reqs_available);
931
932 do {
933 if (avail < ctx->req_batch)
934 goto out;
935
936 old = avail;
937 avail = atomic_cmpxchg(&ctx->reqs_available,
938 avail, avail - ctx->req_batch);
939 } while (avail != old);
940
941 kcpu->reqs_available += ctx->req_batch;
942 }
943
944 ret = true;
945 kcpu->reqs_available--;
946out:
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400947 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000948 return ret;
949}
950
Benjamin LaHaised856f322014-08-24 13:14:05 -0400951/* refill_reqs_available
952 * Updates the reqs_available reference counts used for tracking the
953 * number of free slots in the completion ring. This can be called
954 * from aio_complete() (to optimistically update reqs_available) or
955 * from aio_get_req() (the we're out of events case). It must be
956 * called holding ctx->completion_lock.
957 */
958static void refill_reqs_available(struct kioctx *ctx, unsigned head,
959 unsigned tail)
960{
961 unsigned events_in_ring, completed;
962
963 /* Clamp head since userland can write to it. */
964 head %= ctx->nr_events;
965 if (head <= tail)
966 events_in_ring = tail - head;
967 else
968 events_in_ring = ctx->nr_events - (head - tail);
969
970 completed = ctx->completed_events;
971 if (events_in_ring < completed)
972 completed -= events_in_ring;
973 else
974 completed = 0;
975
976 if (!completed)
977 return;
978
979 ctx->completed_events -= completed;
980 put_reqs_available(ctx, completed);
981}
982
983/* user_refill_reqs_available
984 * Called to refill reqs_available when aio_get_req() encounters an
985 * out of space in the completion ring.
986 */
987static void user_refill_reqs_available(struct kioctx *ctx)
988{
989 spin_lock_irq(&ctx->completion_lock);
990 if (ctx->completed_events) {
991 struct aio_ring *ring;
992 unsigned head;
993
994 /* Access of ring->head may race with aio_read_events_ring()
995 * here, but that's okay since whether we read the old version
996 * or the new version, and either will be valid. The important
997 * part is that head cannot pass tail since we prevent
998 * aio_complete() from updating tail by holding
999 * ctx->completion_lock. Even if head is invalid, the check
1000 * against ctx->completed_events below will make sure we do the
1001 * safe/right thing.
1002 */
1003 ring = kmap_atomic(ctx->ring_pages[0]);
1004 head = ring->head;
1005 kunmap_atomic(ring);
1006
1007 refill_reqs_available(ctx, head, ctx->tail);
1008 }
1009
1010 spin_unlock_irq(&ctx->completion_lock);
1011}
1012
Christoph Hellwig432c7992018-11-19 15:57:42 -07001013static bool get_reqs_available(struct kioctx *ctx)
1014{
1015 if (__get_reqs_available(ctx))
1016 return true;
1017 user_refill_reqs_available(ctx);
1018 return __get_reqs_available(ctx);
1019}
1020
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021/* aio_get_req
Kent Overstreet57282d82013-05-13 13:42:52 -07001022 * Allocate a slot for an aio request.
1023 * Returns NULL if no requests are free.
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001024 *
1025 * The refcount is initialized to 2 - one for the async op completion,
1026 * one for the synchronous code that does this.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001028static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001030 struct aio_kiocb *req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001031
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001032 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033 if (unlikely(!req))
Christoph Hellwig432c7992018-11-19 15:57:42 -07001034 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035
Al Virofa0ca2a2019-03-06 18:21:08 -05001036 if (unlikely(!get_reqs_available(ctx))) {
Wei Yongjun6af1c842019-04-04 08:44:05 +00001037 kmem_cache_free(kiocb_cachep, req);
Al Virofa0ca2a2019-03-06 18:21:08 -05001038 return NULL;
1039 }
1040
Kent Overstreete34ecee2013-10-10 19:31:47 -07001041 percpu_ref_get(&ctx->reqs);
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001042 req->ki_ctx = ctx;
Christoph Hellwig75321b52018-04-09 14:57:56 +02001043 INIT_LIST_HEAD(&req->ki_list);
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001044 refcount_set(&req->ki_refcnt, 2);
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001045 req->ki_eventfd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046 return req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047}
1048
Adrian Bunkd5470b52008-04-29 00:58:57 -07001049static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001051 struct aio_ring __user *ring = (void __user *)ctx_id;
Jens Axboeabf137d2008-12-09 08:11:22 +01001052 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -07001053 struct kioctx *ctx, *ret = NULL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001054 struct kioctx_table *table;
1055 unsigned id;
1056
1057 if (get_user(id, &ring->id))
1058 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059
Jens Axboeabf137d2008-12-09 08:11:22 +01001060 rcu_read_lock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001061 table = rcu_dereference(mm->ioctx_table);
Jens Axboeabf137d2008-12-09 08:11:22 +01001062
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001063 if (!table || id >= table->nr)
1064 goto out;
1065
Jeff Moyera538e3f2018-12-11 12:37:49 -05001066 id = array_index_nospec(id, table->nr);
Tejun Heod0264c02018-03-14 12:10:17 -07001067 ctx = rcu_dereference(table->table[id]);
Benjamin LaHaisef30d7042013-08-07 18:23:48 -04001068 if (ctx && ctx->user_id == ctx_id) {
Al Virobaf10562018-05-20 16:46:23 -04001069 if (percpu_ref_tryget_live(&ctx->users))
1070 ret = ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +01001071 }
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001072out:
Jens Axboeabf137d2008-12-09 08:11:22 +01001073 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -07001074 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075}
1076
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001077static inline void iocb_destroy(struct aio_kiocb *iocb)
1078{
Al Viro74259702019-03-06 18:18:31 -05001079 if (iocb->ki_eventfd)
1080 eventfd_ctx_put(iocb->ki_eventfd);
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001081 if (iocb->ki_filp)
1082 fput(iocb->ki_filp);
1083 percpu_ref_put(&iocb->ki_ctx->reqs);
1084 kmem_cache_free(kiocb_cachep, iocb);
1085}
1086
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087/* aio_complete
1088 * Called when the io request on the given iocb is complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 */
Al Viro2bb874c2019-03-07 19:49:55 -05001090static void aio_complete(struct aio_kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091{
1092 struct kioctx *ctx = iocb->ki_ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 struct aio_ring *ring;
Kent Overstreet21b40202013-05-07 16:18:47 -07001094 struct io_event *ev_page, *event;
Benjamin LaHaised856f322014-08-24 13:14:05 -04001095 unsigned tail, pos, head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -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
Al Viroa9339b72019-03-07 19:43:45 -05001114 *event = iocb->ki_res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115
Kent Overstreet21b40202013-05-07 16:18:47 -07001116 kunmap_atomic(ev_page);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001117 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001118
Al Viroa9339b72019-03-07 19:43:45 -05001119 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1120 (void __user *)(unsigned long)iocb->ki_res.obj,
1121 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122
1123 /* after flagging the request as done, we
1124 * must never even look at it again
1125 */
1126 smp_wmb(); /* make event visible before updating tail */
1127
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001128 ctx->tail = tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001129
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001130 ring = kmap_atomic(ctx->ring_pages[0]);
Benjamin LaHaised856f322014-08-24 13:14:05 -04001131 head = ring->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 ring->tail = tail;
Cong Wange8e3c3d2011-11-25 23:14:27 +08001133 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001134 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135
Benjamin LaHaised856f322014-08-24 13:14:05 -04001136 ctx->completed_events++;
1137 if (ctx->completed_events > 1)
1138 refill_reqs_available(ctx, head, tail);
Kent Overstreet0460fef2013-05-07 16:18:49 -07001139 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1140
Kent Overstreet21b40202013-05-07 16:18:47 -07001141 pr_debug("added to ring %p at [%u]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001142
1143 /*
1144 * Check if the user asked us to deliver the result through an
1145 * eventfd. The eventfd_signal() function is safe to be called
1146 * from IRQ context.
1147 */
Al Viro74259702019-03-06 18:18:31 -05001148 if (iocb->ki_eventfd)
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001149 eventfd_signal(iocb->ki_eventfd, 1);
1150
Quentin Barnes6cb2a212008-03-19 17:00:39 -07001151 /*
1152 * We have to order our ring_info tail store above and test
1153 * of the wait list below outside the wait lock. This is
1154 * like in wake_up_bit() where clearing a bit has to be
1155 * ordered with the unlocked test.
1156 */
1157 smp_mb();
1158
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 if (waitqueue_active(&ctx->wait))
1160 wake_up(&ctx->wait);
Al Viro2bb874c2019-03-07 19:49:55 -05001161}
1162
1163static inline void iocb_put(struct aio_kiocb *iocb)
1164{
1165 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1166 aio_complete(iocb);
1167 iocb_destroy(iocb);
1168 }
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
Kent Overstreeta31ad382013-05-07 16:18:45 -07001221 pos = head + AIO_EVENTS_OFFSET;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001222 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
Kent Overstreeta31ad382013-05-07 16:18:45 -07001223 pos %= AIO_EVENTS_PER_PAGE;
1224
Al Virod2988bd42018-05-26 19:13:10 -04001225 avail = min(avail, nr - ret);
1226 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1227
Kent Overstreeta31ad382013-05-07 16:18:45 -07001228 ev = kmap(page);
1229 copy_ret = copy_to_user(event + ret, ev + pos,
1230 sizeof(*ev) * avail);
1231 kunmap(page);
1232
1233 if (unlikely(copy_ret)) {
1234 ret = -EFAULT;
1235 goto out;
1236 }
1237
1238 ret += avail;
1239 head += avail;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001240 head %= ctx->nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001243 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001244 ring->head = head;
Zhao Hongjiang91d80a82013-04-26 11:03:53 +08001245 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001246 flush_dcache_page(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001247
Kent Overstreet5ffac122013-05-09 15:36:07 -07001248 pr_debug("%li h%u t%u\n", ret, head, tail);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001249out:
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001250 mutex_unlock(&ctx->ring_lock);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001251
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 return ret;
1253}
1254
Kent Overstreeta31ad382013-05-07 16:18:45 -07001255static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1256 struct io_event __user *event, long *i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001258 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259
Kent Overstreeta31ad382013-05-07 16:18:45 -07001260 if (ret > 0)
1261 *i += ret;
1262
1263 if (unlikely(atomic_read(&ctx->dead)))
1264 ret = -EINVAL;
1265
1266 if (!*i)
1267 *i = ret;
1268
1269 return ret < 0 || *i >= min_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270}
1271
Kent Overstreeta31ad382013-05-07 16:18:45 -07001272static long read_events(struct kioctx *ctx, long min_nr, long nr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273 struct io_event __user *event,
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07001274 ktime_t until)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001276 long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277
Kent Overstreeta31ad382013-05-07 16:18:45 -07001278 /*
1279 * Note that aio_read_events() is being called as the conditional - i.e.
1280 * we're calling it after prepare_to_wait() has set task state to
1281 * TASK_INTERRUPTIBLE.
1282 *
1283 * But aio_read_events() can block, and if it blocks it's going to flip
1284 * the task state back to TASK_RUNNING.
1285 *
1286 * This should be ok, provided it doesn't flip the state back to
1287 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1288 * will only happen if the mutex_lock() call blocks, and we then find
1289 * the ringbuffer empty. So in practice we should be ok, but it's
1290 * something to be aware of when touching this code.
1291 */
Thomas Gleixner2456e852016-12-25 11:38:40 +01001292 if (until == 0)
Fam Zheng5f785de2014-11-06 20:44:36 +08001293 aio_read_events(ctx, min_nr, nr, event, &ret);
1294 else
1295 wait_event_interruptible_hrtimeout(ctx->wait,
1296 aio_read_events(ctx, min_nr, nr, event, &ret),
1297 until);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001298 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299}
1300
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301/* sys_io_setup:
1302 * Create an aio_context capable of receiving at least nr_events.
1303 * ctxp must not point to an aio_context that already exists, and
1304 * must be initialized to 0 prior to the call. On successful
1305 * creation of the aio_context, *ctxp is filled in with the resulting
1306 * handle. May fail with -EINVAL if *ctxp is not initialized,
1307 * if the specified nr_events exceeds internal limits. May fail
1308 * with -EAGAIN if the specified nr_events exceeds the user's limit
1309 * of available events. May fail with -ENOMEM if insufficient kernel
1310 * resources are available. May fail with -EFAULT if an invalid
1311 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1312 * implemented.
1313 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001314SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315{
1316 struct kioctx *ioctx = NULL;
1317 unsigned long ctx;
1318 long ret;
1319
1320 ret = get_user(ctx, ctxp);
1321 if (unlikely(ret))
1322 goto out;
1323
1324 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001325 if (unlikely(ctx || nr_events == 0)) {
Kinglong Meeacd88d42015-02-04 21:15:59 +08001326 pr_debug("EINVAL: ctx %lu nr_events %u\n",
Zach Brownd55b5fd2005-11-07 00:59:31 -08001327 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 goto out;
1329 }
1330
1331 ioctx = ioctx_alloc(nr_events);
1332 ret = PTR_ERR(ioctx);
1333 if (!IS_ERR(ioctx)) {
1334 ret = put_user(ioctx->user_id, ctxp);
Al Viroa2e18592012-03-20 16:27:57 -04001335 if (ret)
Anatol Pomozove02ba722014-04-15 11:31:33 -07001336 kill_ioctx(current->mm, ioctx, NULL);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001337 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 }
1339
1340out:
1341 return ret;
1342}
1343
Al Viroc00d2c72016-12-20 07:04:57 -05001344#ifdef CONFIG_COMPAT
1345COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1346{
1347 struct kioctx *ioctx = NULL;
1348 unsigned long ctx;
1349 long ret;
1350
1351 ret = get_user(ctx, ctx32p);
1352 if (unlikely(ret))
1353 goto out;
1354
1355 ret = -EINVAL;
1356 if (unlikely(ctx || nr_events == 0)) {
1357 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1358 ctx, nr_events);
1359 goto out;
1360 }
1361
1362 ioctx = ioctx_alloc(nr_events);
1363 ret = PTR_ERR(ioctx);
1364 if (!IS_ERR(ioctx)) {
1365 /* truncating is ok because it's a user address */
1366 ret = put_user((u32)ioctx->user_id, ctx32p);
1367 if (ret)
1368 kill_ioctx(current->mm, ioctx, NULL);
1369 percpu_ref_put(&ioctx->users);
1370 }
1371
1372out:
1373 return ret;
1374}
1375#endif
1376
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377/* sys_io_destroy:
1378 * Destroy the aio_context specified. May cancel any outstanding
1379 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001380 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 * is invalid.
1382 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001383SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384{
1385 struct kioctx *ioctx = lookup_ioctx(ctx);
1386 if (likely(NULL != ioctx)) {
Jens Axboedc48e562015-04-15 11:17:23 -06001387 struct ctx_rq_wait wait;
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001388 int ret;
Anatol Pomozove02ba722014-04-15 11:31:33 -07001389
Jens Axboedc48e562015-04-15 11:17:23 -06001390 init_completion(&wait.comp);
1391 atomic_set(&wait.count, 1);
1392
Anatol Pomozove02ba722014-04-15 11:31:33 -07001393 /* Pass requests_done to kill_ioctx() where it can be set
1394 * in a thread-safe way. If we try to set it here then we have
1395 * a race condition if two io_destroy() called simultaneously.
1396 */
Jens Axboedc48e562015-04-15 11:17:23 -06001397 ret = kill_ioctx(current->mm, ioctx, &wait);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001398 percpu_ref_put(&ioctx->users);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001399
1400 /* Wait until all IO for the context are done. Otherwise kernel
1401 * keep using user-space buffers even if user thinks the context
1402 * is destroyed.
1403 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001404 if (!ret)
Jens Axboedc48e562015-04-15 11:17:23 -06001405 wait_for_completion(&wait.comp);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001406
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001407 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408 }
Kinglong Meeacd88d42015-02-04 21:15:59 +08001409 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 return -EINVAL;
1411}
1412
Al Viro3c96c7f2018-05-28 13:37:43 -04001413static void aio_remove_iocb(struct aio_kiocb *iocb)
1414{
1415 struct kioctx *ctx = iocb->ki_ctx;
1416 unsigned long flags;
1417
1418 spin_lock_irqsave(&ctx->ctx_lock, flags);
1419 list_del(&iocb->ki_list);
1420 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1421}
1422
Christoph Hellwig54843f82018-05-02 19:57:21 +02001423static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1424{
1425 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1426
Al Viro3c96c7f2018-05-28 13:37:43 -04001427 if (!list_empty_careful(&iocb->ki_list))
1428 aio_remove_iocb(iocb);
1429
Christoph Hellwig54843f82018-05-02 19:57:21 +02001430 if (kiocb->ki_flags & IOCB_WRITE) {
1431 struct inode *inode = file_inode(kiocb->ki_filp);
1432
1433 /*
1434 * Tell lockdep we inherited freeze protection from submission
1435 * thread.
1436 */
1437 if (S_ISREG(inode->i_mode))
1438 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1439 file_end_write(kiocb->ki_filp);
1440 }
1441
Al Viro2bb874c2019-03-07 19:49:55 -05001442 iocb->ki_res.res = res;
1443 iocb->ki_res.res2 = res2;
1444 iocb_put(iocb);
Christoph Hellwig54843f82018-05-02 19:57:21 +02001445}
1446
Jens Axboe88a6f182018-11-24 14:46:14 -07001447static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
Christoph Hellwig54843f82018-05-02 19:57:21 +02001448{
1449 int ret;
1450
Christoph Hellwig54843f82018-05-02 19:57:21 +02001451 req->ki_complete = aio_complete_rw;
Mike Marshallec51f8e2019-02-05 14:13:35 -05001452 req->private = NULL;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001453 req->ki_pos = iocb->aio_offset;
1454 req->ki_flags = iocb_flags(req->ki_filp);
1455 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1456 req->ki_flags |= IOCB_EVENTFD;
Adam Manzanaresfc287242018-05-22 10:52:18 -07001457 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
Adam Manzanaresd9a08a92018-05-22 10:52:19 -07001458 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1459 /*
1460 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1461 * aio_reqprio is interpreted as an I/O scheduling
1462 * class and priority.
1463 */
1464 ret = ioprio_check_cap(iocb->aio_reqprio);
1465 if (ret) {
Adam Manzanares9a6d9a62018-06-04 10:59:57 -07001466 pr_debug("aio ioprio check cap error: %d\n", ret);
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001467 return ret;
Adam Manzanaresd9a08a92018-05-22 10:52:19 -07001468 }
1469
1470 req->ki_ioprio = iocb->aio_reqprio;
1471 } else
Damien Le Moal76dc8912018-11-20 10:52:36 +09001472 req->ki_ioprio = get_current_ioprio();
Adam Manzanaresd9a08a92018-05-22 10:52:19 -07001473
Christoph Hellwig54843f82018-05-02 19:57:21 +02001474 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1475 if (unlikely(ret))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001476 return ret;
Christoph Hellwig154989e2018-11-22 16:44:07 +01001477
1478 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1479 return 0;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001480}
1481
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001482static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1483 struct iovec **iovec, bool vectored, bool compat,
1484 struct iov_iter *iter)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001485{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001486 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1487 size_t len = iocb->aio_nbytes;
1488
1489 if (!vectored) {
1490 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1491 *iovec = NULL;
1492 return ret;
1493 }
Christoph Hellwig89cd35c2020-09-25 06:51:41 +02001494
1495 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001496}
1497
Al Viro9061d142018-05-26 19:11:40 -04001498static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001500 switch (ret) {
1501 case -EIOCBQUEUED:
Al Viro9061d142018-05-26 19:11:40 -04001502 break;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001503 case -ERESTARTSYS:
1504 case -ERESTARTNOINTR:
1505 case -ERESTARTNOHAND:
1506 case -ERESTART_RESTARTBLOCK:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001507 /*
1508 * There's no easy way to restart the syscall since other AIO's
1509 * may be already running. Just fail this IO with EINTR.
1510 */
Christoph Hellwig89319d312016-10-30 11:42:03 -05001511 ret = -EINTR;
Gustavo A. R. Silvadf561f662020-08-23 17:36:59 -05001512 fallthrough;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001513 default:
Jens Axboebc9bff62018-11-06 14:27:13 -07001514 req->ki_complete(req, ret, 0);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001515 }
Christoph Hellwig89319d312016-10-30 11:42:03 -05001516}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517
Al Viro958c13c2019-03-06 18:13:00 -05001518static int aio_read(struct kiocb *req, const struct iocb *iocb,
Jens Axboe88a6f182018-11-24 14:46:14 -07001519 bool vectored, bool compat)
Christoph Hellwig89319d312016-10-30 11:42:03 -05001520{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001521 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1522 struct iov_iter iter;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001523 struct file *file;
Al Viro958c13c2019-03-06 18:13:00 -05001524 int ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001525
Christoph Hellwig54843f82018-05-02 19:57:21 +02001526 ret = aio_prep_rw(req, iocb);
1527 if (ret)
1528 return ret;
1529 file = req->ki_filp;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001530 if (unlikely(!(file->f_mode & FMODE_READ)))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001531 return -EBADF;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001532 ret = -EINVAL;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001533 if (unlikely(!file->f_op->read_iter))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001534 return -EINVAL;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001535
1536 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001537 if (ret < 0)
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001538 return ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001539 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1540 if (!ret)
Al Viro9061d142018-05-26 19:11:40 -04001541 aio_rw_done(req, call_read_iter(file, req, &iter));
Christoph Hellwig89319d312016-10-30 11:42:03 -05001542 kfree(iovec);
1543 return ret;
1544}
1545
Al Viro958c13c2019-03-06 18:13:00 -05001546static int aio_write(struct kiocb *req, const struct iocb *iocb,
Jens Axboe88a6f182018-11-24 14:46:14 -07001547 bool vectored, bool compat)
Christoph Hellwig89319d312016-10-30 11:42:03 -05001548{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001549 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1550 struct iov_iter iter;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001551 struct file *file;
Al Viro958c13c2019-03-06 18:13:00 -05001552 int ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001553
Christoph Hellwig54843f82018-05-02 19:57:21 +02001554 ret = aio_prep_rw(req, iocb);
1555 if (ret)
1556 return ret;
1557 file = req->ki_filp;
1558
Christoph Hellwig89319d312016-10-30 11:42:03 -05001559 if (unlikely(!(file->f_mode & FMODE_WRITE)))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001560 return -EBADF;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001561 if (unlikely(!file->f_op->write_iter))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001562 return -EINVAL;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001563
1564 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001565 if (ret < 0)
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001566 return ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001567 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1568 if (!ret) {
Jan Kara70fe2f42016-10-30 11:42:04 -05001569 /*
Christoph Hellwig92ce4722018-04-06 09:28:17 +02001570 * Open-code file_start_write here to grab freeze protection,
Christoph Hellwig54843f82018-05-02 19:57:21 +02001571 * which will be released by another thread in
1572 * aio_complete_rw(). Fool lockdep by telling it the lock got
1573 * released so that it doesn't complain about the held lock when
1574 * we return to userspace.
Jan Kara70fe2f42016-10-30 11:42:04 -05001575 */
Christoph Hellwig92ce4722018-04-06 09:28:17 +02001576 if (S_ISREG(file_inode(file)->i_mode)) {
Darrick J. Wong8a3c84b2020-11-10 16:50:21 -08001577 sb_start_write(file_inode(file)->i_sb);
Shaohua Lia12f1ae2016-12-13 12:09:56 -08001578 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
Christoph Hellwig92ce4722018-04-06 09:28:17 +02001579 }
1580 req->ki_flags |= IOCB_WRITE;
Al Viro9061d142018-05-26 19:11:40 -04001581 aio_rw_done(req, call_write_iter(file, req, &iter));
Christoph Hellwig89319d312016-10-30 11:42:03 -05001582 }
1583 kfree(iovec);
1584 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585}
1586
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001587static void aio_fsync_work(struct work_struct *work)
1588{
Al Viro2bb874c2019-03-07 19:49:55 -05001589 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
Miklos Szeredi530f32f2020-05-14 16:44:24 +02001590 const struct cred *old_cred = override_creds(iocb->fsync.creds);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001591
Al Viro2bb874c2019-03-07 19:49:55 -05001592 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
Miklos Szeredi530f32f2020-05-14 16:44:24 +02001593 revert_creds(old_cred);
1594 put_cred(iocb->fsync.creds);
Al Viro2bb874c2019-03-07 19:49:55 -05001595 iocb_put(iocb);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001596}
1597
Jens Axboe88a6f182018-11-24 14:46:14 -07001598static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1599 bool datasync)
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001600{
1601 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1602 iocb->aio_rw_flags))
1603 return -EINVAL;
Linus Torvaldsa11e1d42018-06-28 09:43:44 -07001604
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001605 if (unlikely(!req->file->f_op->fsync))
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001606 return -EINVAL;
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001607
Miklos Szeredi530f32f2020-05-14 16:44:24 +02001608 req->creds = prepare_creds();
1609 if (!req->creds)
1610 return -ENOMEM;
1611
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001612 req->datasync = datasync;
1613 INIT_WORK(&req->work, aio_fsync_work);
1614 schedule_work(&req->work);
Al Viro9061d142018-05-26 19:11:40 -04001615 return 0;
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001616}
1617
Jens Axboe01d7a352020-02-03 10:33:42 -07001618static void aio_poll_put_work(struct work_struct *work)
1619{
1620 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1621 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1622
1623 iocb_put(iocb);
1624}
1625
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001626static void aio_poll_complete_work(struct work_struct *work)
1627{
1628 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1629 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1630 struct poll_table_struct pt = { ._key = req->events };
1631 struct kioctx *ctx = iocb->ki_ctx;
1632 __poll_t mask = 0;
1633
1634 if (!READ_ONCE(req->cancelled))
1635 mask = vfs_poll(req->file, &pt) & req->events;
1636
1637 /*
1638 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1639 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1640 * synchronize with them. In the cancellation case the list_del_init
1641 * itself is not actually needed, but harmless so we keep it in to
1642 * avoid further branches in the fast path.
1643 */
1644 spin_lock_irq(&ctx->ctx_lock);
1645 if (!mask && !READ_ONCE(req->cancelled)) {
1646 add_wait_queue(req->head, &req->wait);
1647 spin_unlock_irq(&ctx->ctx_lock);
1648 return;
1649 }
1650 list_del_init(&iocb->ki_list);
Al Viroaf5c72b2019-03-07 21:45:41 -05001651 iocb->ki_res.res = mangle_poll(mask);
1652 req->done = true;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001653 spin_unlock_irq(&ctx->ctx_lock);
1654
Al Viroaf5c72b2019-03-07 21:45:41 -05001655 iocb_put(iocb);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001656}
1657
1658/* assumes we are called with irqs disabled */
1659static int aio_poll_cancel(struct kiocb *iocb)
1660{
1661 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1662 struct poll_iocb *req = &aiocb->poll;
1663
1664 spin_lock(&req->head->lock);
1665 WRITE_ONCE(req->cancelled, true);
1666 if (!list_empty(&req->wait.entry)) {
1667 list_del_init(&req->wait.entry);
1668 schedule_work(&aiocb->poll.work);
1669 }
1670 spin_unlock(&req->head->lock);
1671
1672 return 0;
1673}
1674
1675static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1676 void *key)
1677{
1678 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
Christoph Hellwige8693bc2018-07-16 12:25:17 +02001679 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001680 __poll_t mask = key_to_poll(key);
Bart Van Assched3d6a182019-02-08 16:59:49 -08001681 unsigned long flags;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001682
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001683 /* for instances that support it check for an event match first: */
Al Viroaf5c72b2019-03-07 21:45:41 -05001684 if (mask && !(mask & req->events))
1685 return 0;
Christoph Hellwige8693bc2018-07-16 12:25:17 +02001686
Al Viroaf5c72b2019-03-07 21:45:41 -05001687 list_del_init(&req->wait.entry);
1688
1689 if (mask && spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
Jens Axboe01d7a352020-02-03 10:33:42 -07001690 struct kioctx *ctx = iocb->ki_ctx;
1691
Bart Van Assched3d6a182019-02-08 16:59:49 -08001692 /*
1693 * Try to complete the iocb inline if we can. Use
1694 * irqsave/irqrestore because not all filesystems (e.g. fuse)
1695 * call this function with IRQs disabled and because IRQs
1696 * have to be disabled before ctx_lock is obtained.
1697 */
Al Viroaf5c72b2019-03-07 21:45:41 -05001698 list_del(&iocb->ki_list);
1699 iocb->ki_res.res = mangle_poll(mask);
1700 req->done = true;
Jens Axboe01d7a352020-02-03 10:33:42 -07001701 if (iocb->ki_eventfd && eventfd_signal_count()) {
1702 iocb = NULL;
1703 INIT_WORK(&req->work, aio_poll_put_work);
1704 schedule_work(&req->work);
1705 }
1706 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1707 if (iocb)
1708 iocb_put(iocb);
Al Viroaf5c72b2019-03-07 21:45:41 -05001709 } else {
1710 schedule_work(&req->work);
Christoph Hellwige8693bc2018-07-16 12:25:17 +02001711 }
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001712 return 1;
1713}
1714
1715struct aio_poll_table {
1716 struct poll_table_struct pt;
1717 struct aio_kiocb *iocb;
1718 int error;
1719};
1720
1721static void
1722aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1723 struct poll_table_struct *p)
1724{
1725 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1726
1727 /* multiple wait queues per file are not supported */
1728 if (unlikely(pt->iocb->poll.head)) {
1729 pt->error = -EINVAL;
1730 return;
1731 }
1732
1733 pt->error = 0;
1734 pt->iocb->poll.head = head;
1735 add_wait_queue(head, &pt->iocb->poll.wait);
1736}
1737
Al Viro958c13c2019-03-06 18:13:00 -05001738static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001739{
1740 struct kioctx *ctx = aiocb->ki_ctx;
1741 struct poll_iocb *req = &aiocb->poll;
1742 struct aio_poll_table apt;
Al Viroaf5c72b2019-03-07 21:45:41 -05001743 bool cancel = false;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001744 __poll_t mask;
1745
1746 /* reject any unknown events outside the normal event mask. */
1747 if ((u16)iocb->aio_buf != iocb->aio_buf)
1748 return -EINVAL;
1749 /* reject fields that are not defined for poll */
1750 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1751 return -EINVAL;
1752
1753 INIT_WORK(&req->work, aio_poll_complete_work);
1754 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001755
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001756 req->head = NULL;
Al Viroaf5c72b2019-03-07 21:45:41 -05001757 req->done = false;
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001758 req->cancelled = false;
1759
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001760 apt.pt._qproc = aio_poll_queue_proc;
1761 apt.pt._key = req->events;
1762 apt.iocb = aiocb;
1763 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1764
1765 /* initialized the list so that we can do list_empty checks */
1766 INIT_LIST_HEAD(&req->wait.entry);
1767 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1768
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001769 mask = vfs_poll(req->file, &apt.pt) & req->events;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001770 spin_lock_irq(&ctx->ctx_lock);
Al Viroaf5c72b2019-03-07 21:45:41 -05001771 if (likely(req->head)) {
1772 spin_lock(&req->head->lock);
1773 if (unlikely(list_empty(&req->wait.entry))) {
1774 if (apt.error)
1775 cancel = true;
1776 apt.error = 0;
1777 mask = 0;
1778 }
1779 if (mask || apt.error) {
1780 list_del_init(&req->wait.entry);
1781 } else if (cancel) {
1782 WRITE_ONCE(req->cancelled, true);
1783 } else if (!req->done) { /* actually waiting for an event */
1784 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1785 aiocb->ki_cancel = aio_poll_cancel;
1786 }
1787 spin_unlock(&req->head->lock);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001788 }
Al Viroaf5c72b2019-03-07 21:45:41 -05001789 if (mask) { /* no async, we'd stolen it */
1790 aiocb->ki_res.res = mangle_poll(mask);
1791 apt.error = 0;
1792 }
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001793 spin_unlock_irq(&ctx->ctx_lock);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001794 if (mask)
Al Viroaf5c72b2019-03-07 21:45:41 -05001795 iocb_put(aiocb);
1796 return apt.error;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001797}
1798
Jens Axboe88a6f182018-11-24 14:46:14 -07001799static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
Al Viro7316b492019-03-06 18:24:51 -05001800 struct iocb __user *user_iocb, struct aio_kiocb *req,
1801 bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802{
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001803 req->ki_filp = fget(iocb->aio_fildes);
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001804 if (unlikely(!req->ki_filp))
Al Viro7316b492019-03-06 18:24:51 -05001805 return -EBADF;
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001806
Jens Axboe88a6f182018-11-24 14:46:14 -07001807 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
Al Viro74259702019-03-06 18:18:31 -05001808 struct eventfd_ctx *eventfd;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001809 /*
1810 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1811 * instance of the file* now. The file descriptor must be
1812 * an eventfd() fd, and will be signaled for each completed
1813 * event using the eventfd_signal() function.
1814 */
Al Viro74259702019-03-06 18:18:31 -05001815 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
Al Viro7316b492019-03-06 18:24:51 -05001816 if (IS_ERR(eventfd))
Dan Carpenter18bfb9c2019-04-03 09:22:35 +03001817 return PTR_ERR(eventfd);
Al Viro7316b492019-03-06 18:24:51 -05001818
Al Viro74259702019-03-06 18:18:31 -05001819 req->ki_eventfd = eventfd;
Goldwyn Rodrigues9830f4b2017-06-20 07:05:42 -05001820 }
1821
Al Viro7316b492019-03-06 18:24:51 -05001822 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001823 pr_debug("EFAULT: aio_key\n");
Al Viro7316b492019-03-06 18:24:51 -05001824 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 }
1826
Al Viroa9339b72019-03-07 19:43:45 -05001827 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1828 req->ki_res.data = iocb->aio_data;
1829 req->ki_res.res = 0;
1830 req->ki_res.res2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831
Jens Axboe88a6f182018-11-24 14:46:14 -07001832 switch (iocb->aio_lio_opcode) {
Christoph Hellwig89319d312016-10-30 11:42:03 -05001833 case IOCB_CMD_PREAD:
Al Viro7316b492019-03-06 18:24:51 -05001834 return aio_read(&req->rw, iocb, false, compat);
Christoph Hellwig89319d312016-10-30 11:42:03 -05001835 case IOCB_CMD_PWRITE:
Al Viro7316b492019-03-06 18:24:51 -05001836 return aio_write(&req->rw, iocb, false, compat);
Christoph Hellwig89319d312016-10-30 11:42:03 -05001837 case IOCB_CMD_PREADV:
Al Viro7316b492019-03-06 18:24:51 -05001838 return aio_read(&req->rw, iocb, true, compat);
Christoph Hellwig89319d312016-10-30 11:42:03 -05001839 case IOCB_CMD_PWRITEV:
Al Viro7316b492019-03-06 18:24:51 -05001840 return aio_write(&req->rw, iocb, true, compat);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001841 case IOCB_CMD_FSYNC:
Al Viro7316b492019-03-06 18:24:51 -05001842 return aio_fsync(&req->fsync, iocb, false);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001843 case IOCB_CMD_FDSYNC:
Al Viro7316b492019-03-06 18:24:51 -05001844 return aio_fsync(&req->fsync, iocb, true);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001845 case IOCB_CMD_POLL:
Al Viro7316b492019-03-06 18:24:51 -05001846 return aio_poll(req, iocb);
Christoph Hellwig89319d312016-10-30 11:42:03 -05001847 default:
Jens Axboe88a6f182018-11-24 14:46:14 -07001848 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
Al Viro7316b492019-03-06 18:24:51 -05001849 return -EINVAL;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001850 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851}
1852
Jens Axboe88a6f182018-11-24 14:46:14 -07001853static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1854 bool compat)
1855{
Al Viro7316b492019-03-06 18:24:51 -05001856 struct aio_kiocb *req;
Jens Axboe88a6f182018-11-24 14:46:14 -07001857 struct iocb iocb;
Al Viro7316b492019-03-06 18:24:51 -05001858 int err;
Jens Axboe88a6f182018-11-24 14:46:14 -07001859
1860 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1861 return -EFAULT;
1862
Al Viro7316b492019-03-06 18:24:51 -05001863 /* enforce forwards compatibility on users */
1864 if (unlikely(iocb.aio_reserved2)) {
1865 pr_debug("EINVAL: reserve field set\n");
1866 return -EINVAL;
1867 }
1868
1869 /* prevent overflows */
1870 if (unlikely(
1871 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1872 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1873 ((ssize_t)iocb.aio_nbytes < 0)
1874 )) {
1875 pr_debug("EINVAL: overflow check\n");
1876 return -EINVAL;
1877 }
1878
1879 req = aio_get_req(ctx);
1880 if (unlikely(!req))
1881 return -EAGAIN;
1882
1883 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
1884
1885 /* Done with the synchronous reference */
1886 iocb_put(req);
1887
1888 /*
1889 * If err is 0, we'd either done aio_complete() ourselves or have
1890 * arranged for that to be done asynchronously. Anything non-zero
1891 * means that we need to destroy req ourselves.
1892 */
1893 if (unlikely(err)) {
1894 iocb_destroy(req);
1895 put_reqs_available(ctx, 1);
1896 }
1897 return err;
Jens Axboe88a6f182018-11-24 14:46:14 -07001898}
1899
Jeff Moyer9d85cba72010-05-26 14:44:26 -07001900/* sys_io_submit:
1901 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1902 * the number of iocbs queued. May return -EINVAL if the aio_context
1903 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1904 * *iocbpp[0] is not properly initialized, if the operation specified
1905 * is invalid for the file descriptor in the iocb. May fail with
1906 * -EFAULT if any of the data structures point to invalid data. May
1907 * fail with -EBADF if the file descriptor specified in the first
1908 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1909 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1910 * fail with -ENOSYS if not implemented.
1911 */
1912SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1913 struct iocb __user * __user *, iocbpp)
1914{
Al Viro67ba0492018-05-26 20:10:07 -04001915 struct kioctx *ctx;
1916 long ret = 0;
1917 int i = 0;
1918 struct blk_plug plug;
1919
1920 if (unlikely(nr < 0))
1921 return -EINVAL;
1922
Al Viro67ba0492018-05-26 20:10:07 -04001923 ctx = lookup_ioctx(ctx_id);
1924 if (unlikely(!ctx)) {
1925 pr_debug("EINVAL: invalid context id\n");
1926 return -EINVAL;
1927 }
1928
Al Viro1da92772018-05-26 20:10:07 -04001929 if (nr > ctx->nr_events)
1930 nr = ctx->nr_events;
1931
Jens Axboea79d40e2018-12-04 09:45:32 -07001932 if (nr > AIO_PLUG_THRESHOLD)
1933 blk_start_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04001934 for (i = 0; i < nr; i++) {
1935 struct iocb __user *user_iocb;
1936
1937 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1938 ret = -EFAULT;
1939 break;
1940 }
1941
1942 ret = io_submit_one(ctx, user_iocb, false);
1943 if (ret)
1944 break;
1945 }
Jens Axboea79d40e2018-12-04 09:45:32 -07001946 if (nr > AIO_PLUG_THRESHOLD)
1947 blk_finish_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04001948
1949 percpu_ref_put(&ctx->users);
1950 return i ? i : ret;
Jeff Moyer9d85cba72010-05-26 14:44:26 -07001951}
1952
Al Viroc00d2c72016-12-20 07:04:57 -05001953#ifdef CONFIG_COMPAT
Al Viroc00d2c72016-12-20 07:04:57 -05001954COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
Al Viro67ba0492018-05-26 20:10:07 -04001955 int, nr, compat_uptr_t __user *, iocbpp)
Al Viroc00d2c72016-12-20 07:04:57 -05001956{
Al Viro67ba0492018-05-26 20:10:07 -04001957 struct kioctx *ctx;
1958 long ret = 0;
1959 int i = 0;
1960 struct blk_plug plug;
Al Viroc00d2c72016-12-20 07:04:57 -05001961
1962 if (unlikely(nr < 0))
1963 return -EINVAL;
1964
Al Viro67ba0492018-05-26 20:10:07 -04001965 ctx = lookup_ioctx(ctx_id);
1966 if (unlikely(!ctx)) {
1967 pr_debug("EINVAL: invalid context id\n");
1968 return -EINVAL;
1969 }
1970
Al Viro1da92772018-05-26 20:10:07 -04001971 if (nr > ctx->nr_events)
1972 nr = ctx->nr_events;
1973
Jens Axboea79d40e2018-12-04 09:45:32 -07001974 if (nr > AIO_PLUG_THRESHOLD)
1975 blk_start_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04001976 for (i = 0; i < nr; i++) {
1977 compat_uptr_t user_iocb;
1978
1979 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1980 ret = -EFAULT;
1981 break;
1982 }
1983
1984 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
1985 if (ret)
1986 break;
1987 }
Jens Axboea79d40e2018-12-04 09:45:32 -07001988 if (nr > AIO_PLUG_THRESHOLD)
1989 blk_finish_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04001990
1991 percpu_ref_put(&ctx->users);
1992 return i ? i : ret;
Al Viroc00d2c72016-12-20 07:04:57 -05001993}
1994#endif
1995
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996/* sys_io_cancel:
1997 * Attempts to cancel an iocb previously passed to io_submit. If
1998 * the operation is successfully cancelled, the resulting event is
1999 * copied into the memory pointed to by result without being placed
2000 * into the completion queue and 0 is returned. May fail with
2001 * -EFAULT if any of the data structures pointed to are invalid.
2002 * May fail with -EINVAL if aio_context specified by ctx_id is
2003 * invalid. May fail with -EAGAIN if the iocb specified was not
2004 * cancelled. Will fail with -ENOSYS if not implemented.
2005 */
Heiko Carstens002c8972009-01-14 14:14:18 +01002006SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2007 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009 struct kioctx *ctx;
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01002010 struct aio_kiocb *kiocb;
Christoph Hellwig888933f2018-05-23 14:11:02 +02002011 int ret = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012 u32 key;
Al Viroa9339b72019-03-07 19:43:45 -05002013 u64 obj = (u64)(unsigned long)iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014
Christoph Hellwigf3a27522018-03-30 11:19:25 +02002015 if (unlikely(get_user(key, &iocb->aio_key)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016 return -EFAULT;
Christoph Hellwigf3a27522018-03-30 11:19:25 +02002017 if (unlikely(key != KIOCB_KEY))
2018 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019
2020 ctx = lookup_ioctx(ctx_id);
2021 if (unlikely(!ctx))
2022 return -EINVAL;
2023
2024 spin_lock_irq(&ctx->ctx_lock);
Al Viro833f4152019-03-11 19:00:36 -04002025 /* TODO: use a hash or array, this sucks. */
2026 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
Al Viroa9339b72019-03-07 19:43:45 -05002027 if (kiocb->ki_res.obj == obj) {
Al Viro833f4152019-03-11 19:00:36 -04002028 ret = kiocb->ki_cancel(&kiocb->rw);
2029 list_del_init(&kiocb->ki_list);
2030 break;
2031 }
Christoph Hellwig888933f2018-05-23 14:11:02 +02002032 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 spin_unlock_irq(&ctx->ctx_lock);
2034
Kent Overstreet906b9732013-05-07 16:18:31 -07002035 if (!ret) {
Kent Overstreetbec68faa2013-05-13 14:45:08 -07002036 /*
2037 * The result argument is no longer used - the io_event is
2038 * always delivered via the ring buffer. -EINPROGRESS indicates
2039 * cancellation is progress:
Kent Overstreet906b9732013-05-07 16:18:31 -07002040 */
Kent Overstreetbec68faa2013-05-13 14:45:08 -07002041 ret = -EINPROGRESS;
Kent Overstreet906b9732013-05-07 16:18:31 -07002042 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043
Kent Overstreet723be6e2013-05-28 15:14:48 -07002044 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045
2046 return ret;
2047}
2048
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002049static long do_io_getevents(aio_context_t ctx_id,
2050 long min_nr,
2051 long nr,
2052 struct io_event __user *events,
2053 struct timespec64 *ts)
2054{
2055 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2056 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2057 long ret = -EINVAL;
2058
2059 if (likely(ioctx)) {
2060 if (likely(min_nr <= nr && min_nr >= 0))
2061 ret = read_events(ioctx, min_nr, nr, events, until);
2062 percpu_ref_put(&ioctx->users);
2063 }
2064
2065 return ret;
2066}
2067
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068/* io_getevents:
2069 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07002070 * the completion queue for the aio_context specified by ctx_id. If
2071 * it succeeds, the number of read events is returned. May fail with
2072 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2073 * out of range, if timeout is out of range. May fail with -EFAULT
2074 * if any of the memory specified is invalid. May return 0 or
2075 * < min_nr if the timeout specified by timeout has elapsed
2076 * before sufficient events are available, where timeout == NULL
2077 * specifies an infinite timeout. Note that the timeout pointed to by
Jeff Moyer69008072013-05-24 15:55:24 -07002078 * timeout is relative. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 */
Arnd Bergmann3ca47e92019-04-23 17:43:50 +02002080#ifdef CONFIG_64BIT
Deepa Dinamani7a353972018-09-19 21:41:08 -07002081
Heiko Carstens002c8972009-01-14 14:14:18 +01002082SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2083 long, min_nr,
2084 long, nr,
2085 struct io_event __user *, events,
Deepa Dinamani7a353972018-09-19 21:41:08 -07002086 struct __kernel_timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087{
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002088 struct timespec64 ts;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002089 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002091 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2092 return -EFAULT;
2093
2094 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2095 if (!ret && signal_pending(current))
2096 ret = -EINTR;
2097 return ret;
2098}
2099
Deepa Dinamani7a353972018-09-19 21:41:08 -07002100#endif
2101
Christoph Hellwig9ba546c2018-07-11 15:48:46 +02002102struct __aio_sigset {
2103 const sigset_t __user *sigmask;
2104 size_t sigsetsize;
2105};
2106
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002107SYSCALL_DEFINE6(io_pgetevents,
2108 aio_context_t, ctx_id,
2109 long, min_nr,
2110 long, nr,
2111 struct io_event __user *, events,
Deepa Dinamani7a353972018-09-19 21:41:08 -07002112 struct __kernel_timespec __user *, timeout,
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002113 const struct __aio_sigset __user *, usig)
2114{
2115 struct __aio_sigset ksig = { NULL, };
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002116 struct timespec64 ts;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002117 bool interrupted;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002118 int ret;
2119
2120 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2121 return -EFAULT;
2122
2123 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2124 return -EFAULT;
2125
Oleg Nesterovb7724342019-07-16 16:29:53 -07002126 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
Deepa Dinamani7a353972018-09-19 21:41:08 -07002127 if (ret)
2128 return ret;
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002129
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002130 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002131
2132 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002133 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002134 if (interrupted && !ret)
Deepa Dinamani7a353972018-09-19 21:41:08 -07002135 ret = -ERESTARTNOHAND;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002136
2137 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138}
Al Viroc00d2c72016-12-20 07:04:57 -05002139
Deepa Dinamani7a353972018-09-19 21:41:08 -07002140#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2141
2142SYSCALL_DEFINE6(io_pgetevents_time32,
2143 aio_context_t, ctx_id,
2144 long, min_nr,
2145 long, nr,
2146 struct io_event __user *, events,
2147 struct old_timespec32 __user *, timeout,
2148 const struct __aio_sigset __user *, usig)
2149{
2150 struct __aio_sigset ksig = { NULL, };
Deepa Dinamani7a353972018-09-19 21:41:08 -07002151 struct timespec64 ts;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002152 bool interrupted;
Deepa Dinamani7a353972018-09-19 21:41:08 -07002153 int ret;
2154
2155 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2156 return -EFAULT;
2157
2158 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2159 return -EFAULT;
2160
Deepa Dinamanided653c2018-09-19 21:41:04 -07002161
Oleg Nesterovb7724342019-07-16 16:29:53 -07002162 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
Deepa Dinamanided653c2018-09-19 21:41:04 -07002163 if (ret)
2164 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165
2166 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002167
2168 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002169 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002170 if (interrupted && !ret)
Deepa Dinamani854a6ed2018-09-19 21:41:05 -07002171 ret = -ERESTARTNOHAND;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172
2173 return ret;
2174}
Al Viroc00d2c72016-12-20 07:04:57 -05002175
Deepa Dinamani7a353972018-09-19 21:41:08 -07002176#endif
2177
2178#if defined(CONFIG_COMPAT_32BIT_TIME)
2179
Arnd Bergmann8dabe722019-01-07 00:33:08 +01002180SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2181 __s32, min_nr,
2182 __s32, nr,
2183 struct io_event __user *, events,
2184 struct old_timespec32 __user *, timeout)
Al Viroc00d2c72016-12-20 07:04:57 -05002185{
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002186 struct timespec64 t;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002187 int ret;
Al Viroc00d2c72016-12-20 07:04:57 -05002188
Arnd Bergmann9afc5ee2018-07-13 12:52:28 +02002189 if (timeout && get_old_timespec32(&t, timeout))
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002190 return -EFAULT;
2191
2192 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2193 if (!ret && signal_pending(current))
2194 ret = -EINTR;
2195 return ret;
2196}
2197
Deepa Dinamani7a353972018-09-19 21:41:08 -07002198#endif
2199
2200#ifdef CONFIG_COMPAT
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002201
2202struct __compat_aio_sigset {
Guillem Jover97eba802019-08-21 05:38:20 +02002203 compat_uptr_t sigmask;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002204 compat_size_t sigsetsize;
2205};
2206
Deepa Dinamani7a353972018-09-19 21:41:08 -07002207#if defined(CONFIG_COMPAT_32BIT_TIME)
2208
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002209COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2210 compat_aio_context_t, ctx_id,
2211 compat_long_t, min_nr,
2212 compat_long_t, nr,
2213 struct io_event __user *, events,
Arnd Bergmann9afc5ee2018-07-13 12:52:28 +02002214 struct old_timespec32 __user *, timeout,
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002215 const struct __compat_aio_sigset __user *, usig)
2216{
Guillem Jover97eba802019-08-21 05:38:20 +02002217 struct __compat_aio_sigset ksig = { 0, };
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002218 struct timespec64 t;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002219 bool interrupted;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002220 int ret;
2221
Arnd Bergmann9afc5ee2018-07-13 12:52:28 +02002222 if (timeout && get_old_timespec32(&t, timeout))
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002223 return -EFAULT;
2224
2225 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2226 return -EFAULT;
2227
Guillem Jover97eba802019-08-21 05:38:20 +02002228 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
Deepa Dinamanided653c2018-09-19 21:41:04 -07002229 if (ret)
2230 return ret;
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002231
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002232 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002233
2234 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002235 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002236 if (interrupted && !ret)
Deepa Dinamani854a6ed2018-09-19 21:41:05 -07002237 ret = -ERESTARTNOHAND;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002238
2239 return ret;
Al Viroc00d2c72016-12-20 07:04:57 -05002240}
Deepa Dinamani7a353972018-09-19 21:41:08 -07002241
2242#endif
2243
2244COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2245 compat_aio_context_t, ctx_id,
2246 compat_long_t, min_nr,
2247 compat_long_t, nr,
2248 struct io_event __user *, events,
2249 struct __kernel_timespec __user *, timeout,
2250 const struct __compat_aio_sigset __user *, usig)
2251{
Guillem Jover97eba802019-08-21 05:38:20 +02002252 struct __compat_aio_sigset ksig = { 0, };
Deepa Dinamani7a353972018-09-19 21:41:08 -07002253 struct timespec64 t;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002254 bool interrupted;
Deepa Dinamani7a353972018-09-19 21:41:08 -07002255 int ret;
2256
2257 if (timeout && get_timespec64(&t, timeout))
2258 return -EFAULT;
2259
2260 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2261 return -EFAULT;
2262
Guillem Jover97eba802019-08-21 05:38:20 +02002263 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
Deepa Dinamani7a353972018-09-19 21:41:08 -07002264 if (ret)
2265 return ret;
2266
2267 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002268
2269 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002270 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002271 if (interrupted && !ret)
Deepa Dinamani7a353972018-09-19 21:41:08 -07002272 ret = -ERESTARTNOHAND;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273
2274 return ret;
2275}
2276#endif