blob: 6066f64967b3cce33c43eca027ba0c12e0246649 [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;
Kees Cookdb7fcc82023-09-15 13:14:14 -070083 struct kioctx __rcu *table[] __counted_by(nr);
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
Kefeng Wang16594e62024-03-21 21:16:40 +0800125 struct folio **ring_folios;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700126 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
Kefeng Wang16594e62024-03-21 21:16:40 +0800163 struct folio *internal_folios[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;
Christoph Hellwigbfe40372018-07-16 09:08:20 +0200184 bool cancelled;
Eric Biggers363bee22021-12-08 17:04:54 -0800185 bool work_scheduled;
186 bool work_need_resched;
Christoph Hellwigbfe40372018-07-16 09:08:20 +0200187 struct wait_queue_entry wait;
188 struct work_struct work;
189};
190
Linus Torvalds84c4e1f82019-03-03 14:23:33 -0800191/*
192 * NOTE! Each of the iocb union members has the file pointer
193 * as the first entry in their struct definition. So you can
194 * access the file pointer through any of the sub-structs,
195 * or directly as just 'ki_filp' in this struct.
196 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100197struct aio_kiocb {
Christoph Hellwig54843f82018-05-02 19:57:21 +0200198 union {
Linus Torvalds84c4e1f82019-03-03 14:23:33 -0800199 struct file *ki_filp;
Christoph Hellwig54843f82018-05-02 19:57:21 +0200200 struct kiocb rw;
Christoph Hellwiga3c0d432018-03-27 19:18:57 +0200201 struct fsync_iocb fsync;
Christoph Hellwigbfe40372018-07-16 09:08:20 +0200202 struct poll_iocb poll;
Christoph Hellwig54843f82018-05-02 19:57:21 +0200203 };
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100204
205 struct kioctx *ki_ctx;
206 kiocb_cancel_fn *ki_cancel;
207
Al Viroa9339b72019-03-07 19:43:45 -0500208 struct io_event ki_res;
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100209
210 struct list_head ki_list; /* the aio core uses this
211 * for cancellation */
Christoph Hellwig9018ccc2018-07-24 11:36:37 +0200212 refcount_t ki_refcnt;
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100213
214 /*
215 * If the aio_resfd field of the userspace iocb is not zero,
216 * this is the underlying eventfd context to deliver events to.
217 */
218 struct eventfd_ctx *ki_eventfd;
219};
220
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -0800222static DEFINE_SPINLOCK(aio_nr_lock);
Xiaoming Ni86b12b62022-01-21 22:11:24 -0800223static unsigned long aio_nr; /* current system wide number of aio requests */
224static unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225/*----end sysctl variables---*/
Xiaoming Ni86b12b62022-01-21 22:11:24 -0800226#ifdef CONFIG_SYSCTL
227static struct ctl_table aio_sysctls[] = {
228 {
229 .procname = "aio-nr",
230 .data = &aio_nr,
231 .maxlen = sizeof(aio_nr),
232 .mode = 0444,
233 .proc_handler = proc_doulongvec_minmax,
234 },
235 {
236 .procname = "aio-max-nr",
237 .data = &aio_max_nr,
238 .maxlen = sizeof(aio_max_nr),
239 .mode = 0644,
240 .proc_handler = proc_doulongvec_minmax,
241 },
Xiaoming Ni86b12b62022-01-21 22:11:24 -0800242};
243
244static void __init aio_sysctl_init(void)
245{
246 register_sysctl_init("fs", aio_sysctls);
247}
248#else
249#define aio_sysctl_init() do { } while (0)
250#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
Christoph Lametere18b8902006-12-06 20:33:20 -0800252static struct kmem_cache *kiocb_cachep;
253static struct kmem_cache *kioctx_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400255static struct vfsmount *aio_mnt;
256
257static const struct file_operations aio_ring_fops;
258static const struct address_space_operations aio_ctx_aops;
259
260static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
261{
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400262 struct file *file;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400263 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
Dan Carpenter7f626562013-11-13 10:49:40 +0300264 if (IS_ERR(inode))
265 return ERR_CAST(inode);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400266
267 inode->i_mapping->a_ops = &aio_ctx_aops;
Matthew Wilcox (Oracle)600f1112023-11-17 21:58:23 +0000268 inode->i_mapping->i_private_data = ctx;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400269 inode->i_size = PAGE_SIZE * nr_pages;
270
Al Virod93aa9d2018-06-09 09:40:05 -0400271 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
272 O_RDWR, &aio_ring_fops);
Al Viroc9c554f2018-07-11 14:19:04 -0400273 if (IS_ERR(file))
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400274 iput(inode);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400275 return file;
276}
277
David Howells52db59d2019-03-25 16:38:23 +0000278static int aio_init_fs_context(struct fs_context *fc)
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400279{
David Howells52db59d2019-03-25 16:38:23 +0000280 if (!init_pseudo(fc, AIO_RING_MAGIC))
281 return -ENOMEM;
282 fc->s_iflags |= SB_I_NOEXEC;
283 return 0;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400284}
285
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286/* aio_setup
287 * Creates the slab caches used by the aio routines, panic on
288 * failure as this is done early during the boot sequence.
289 */
290static int __init aio_setup(void)
291{
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400292 static struct file_system_type aio_fs = {
293 .name = "aio",
David Howells52db59d2019-03-25 16:38:23 +0000294 .init_fs_context = aio_init_fs_context,
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400295 .kill_sb = kill_anon_super,
296 };
297 aio_mnt = kern_mount(&aio_fs);
298 if (IS_ERR(aio_mnt))
299 panic("Failed to create aio fs mount.");
300
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100301 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Christoph Lameter0a31bd52007-05-06 14:49:57 -0700302 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Xiaoming Ni86b12b62022-01-21 22:11:24 -0800303 aio_sysctl_init();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 return 0;
305}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700306__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400308static void put_aio_ring_file(struct kioctx *ctx)
309{
310 struct file *aio_ring_file = ctx->aio_ring_file;
Rasmus Villemoesde04e762016-09-15 00:25:03 +0200311 struct address_space *i_mapping;
312
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400313 if (aio_ring_file) {
Al Viro45063092016-12-04 18:24:56 -0500314 truncate_setsize(file_inode(aio_ring_file), 0);
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400315
316 /* Prevent further access to the kioctx from migratepages */
Al Viro45063092016-12-04 18:24:56 -0500317 i_mapping = aio_ring_file->f_mapping;
Matthew Wilcox (Oracle)600f1112023-11-17 21:58:23 +0000318 spin_lock(&i_mapping->i_private_lock);
319 i_mapping->i_private_data = NULL;
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400320 ctx->aio_ring_file = NULL;
Matthew Wilcox (Oracle)600f1112023-11-17 21:58:23 +0000321 spin_unlock(&i_mapping->i_private_lock);
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400322
323 fput(aio_ring_file);
324 }
325}
326
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327static void aio_free_ring(struct kioctx *ctx)
328{
Gu Zheng36bc08c2013-07-16 17:56:16 +0800329 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400331 /* Disconnect the kiotx from the ring file. This prevents future
332 * accesses to the kioctx from page migration.
333 */
334 put_aio_ring_file(ctx);
335
Gu Zheng36bc08c2013-07-16 17:56:16 +0800336 for (i = 0; i < ctx->nr_pages; i++) {
Kefeng Wang16594e62024-03-21 21:16:40 +0800337 struct folio *folio = ctx->ring_folios[i];
Kefeng Wang5ea0a352024-03-21 21:16:39 +0800338
339 if (!folio)
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500340 continue;
Kefeng Wang5ea0a352024-03-21 21:16:39 +0800341
342 pr_debug("pid(%d) [%d] folio->count=%d\n", current->pid, i,
343 folio_ref_count(folio));
Kefeng Wang16594e62024-03-21 21:16:40 +0800344 ctx->ring_folios[i] = NULL;
Kefeng Wang5ea0a352024-03-21 21:16:39 +0800345 folio_put(folio);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800346 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347
Kefeng Wang16594e62024-03-21 21:16:40 +0800348 if (ctx->ring_folios && ctx->ring_folios != ctx->internal_folios) {
349 kfree(ctx->ring_folios);
350 ctx->ring_folios = NULL;
Sasha Levinddb8c452013-11-19 17:33:03 -0500351 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352}
353
Brian Geffon14d07112021-04-29 22:57:48 -0700354static int aio_ring_mremap(struct vm_area_struct *vma)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800355{
Oleg Nesterov5477e702015-09-04 15:48:04 -0700356 struct file *file = vma->vm_file;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400357 struct mm_struct *mm = vma->vm_mm;
358 struct kioctx_table *table;
Al Virob2edffd2015-04-06 17:48:54 -0400359 int i, res = -EINVAL;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400360
361 spin_lock(&mm->ioctx_lock);
362 rcu_read_lock();
363 table = rcu_dereference(mm->ioctx_table);
Seth Jenkins81e9d6f2023-01-31 12:25:55 -0500364 if (!table)
365 goto out_unlock;
366
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400367 for (i = 0; i < table->nr; i++) {
368 struct kioctx *ctx;
369
Tejun Heod0264c02018-03-14 12:10:17 -0700370 ctx = rcu_dereference(table->table[i]);
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400371 if (ctx && ctx->aio_ring_file == file) {
Al Virob2edffd2015-04-06 17:48:54 -0400372 if (!atomic_read(&ctx->dead)) {
373 ctx->user_id = ctx->mmap_base = vma->vm_start;
374 res = 0;
375 }
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400376 break;
377 }
378 }
379
Seth Jenkins81e9d6f2023-01-31 12:25:55 -0500380out_unlock:
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400381 rcu_read_unlock();
382 spin_unlock(&mm->ioctx_lock);
Al Virob2edffd2015-04-06 17:48:54 -0400383 return res;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400384}
385
Oleg Nesterov5477e702015-09-04 15:48:04 -0700386static const struct vm_operations_struct aio_ring_vm_ops = {
387 .mremap = aio_ring_mremap,
388#if IS_ENABLED(CONFIG_MMU)
389 .fault = filemap_fault,
390 .map_pages = filemap_map_pages,
391 .page_mkwrite = filemap_page_mkwrite,
392#endif
393};
394
395static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
396{
Suren Baghdasaryan1c712222023-01-26 11:37:49 -0800397 vm_flags_set(vma, VM_DONTEXPAND);
Oleg Nesterov5477e702015-09-04 15:48:04 -0700398 vma->vm_ops = &aio_ring_vm_ops;
399 return 0;
400}
401
Gu Zheng36bc08c2013-07-16 17:56:16 +0800402static const struct file_operations aio_ring_fops = {
403 .mmap = aio_ring_mmap,
404};
405
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400406#if IS_ENABLED(CONFIG_MIGRATION)
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400407static int aio_migrate_folio(struct address_space *mapping, struct folio *dst,
408 struct folio *src, enum migrate_mode mode)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800409{
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400410 struct kioctx *ctx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800411 unsigned long flags;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400412 pgoff_t idx;
Kefeng Wang90663282024-05-24 13:28:43 +0800413 int rc = 0;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500414
Matthew Wilcox (Oracle)600f1112023-11-17 21:58:23 +0000415 /* mapping->i_private_lock here protects against the kioctx teardown. */
416 spin_lock(&mapping->i_private_lock);
417 ctx = mapping->i_private_data;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400418 if (!ctx) {
419 rc = -EINVAL;
420 goto out;
421 }
422
423 /* The ring_lock mutex. The prevents aio_read_events() from writing
424 * to the ring's head, and prevents page migration from mucking in
425 * a partially initialized kiotx.
426 */
427 if (!mutex_trylock(&ctx->ring_lock)) {
428 rc = -EAGAIN;
429 goto out;
430 }
431
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400432 idx = src->index;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400433 if (idx < (pgoff_t)ctx->nr_pages) {
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400434 /* Make sure the old folio hasn't already been changed */
Kefeng Wang16594e62024-03-21 21:16:40 +0800435 if (ctx->ring_folios[idx] != src)
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400436 rc = -EAGAIN;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500437 } else
438 rc = -EINVAL;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500439
440 if (rc != 0)
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400441 goto out_unlock;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500442
Gu Zheng36bc08c2013-07-16 17:56:16 +0800443 /* Writeback must be complete */
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400444 BUG_ON(folio_test_writeback(src));
445 folio_get(dst);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800446
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400447 rc = folio_migrate_mapping(mapping, dst, src, 1);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800448 if (rc != MIGRATEPAGE_SUCCESS) {
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400449 folio_put(dst);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400450 goto out_unlock;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800451 }
452
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400453 /* Take completion_lock to prevent other writes to the ring buffer
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400454 * while the old folio is copied to the new. This prevents new
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400455 * events from being lost.
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400456 */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400457 spin_lock_irqsave(&ctx->completion_lock, flags);
Kefeng Wang3f594932024-06-26 16:53:28 +0800458 folio_copy(dst, src);
459 folio_migrate_flags(dst, src);
Kefeng Wang16594e62024-03-21 21:16:40 +0800460 BUG_ON(ctx->ring_folios[idx] != src);
461 ctx->ring_folios[idx] = dst;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400462 spin_unlock_irqrestore(&ctx->completion_lock, flags);
463
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400464 /* The old folio is no longer accessible. */
465 folio_put(src);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400466
467out_unlock:
468 mutex_unlock(&ctx->ring_lock);
469out:
Matthew Wilcox (Oracle)600f1112023-11-17 21:58:23 +0000470 spin_unlock(&mapping->i_private_lock);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800471 return rc;
472}
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400473#else
474#define aio_migrate_folio NULL
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400475#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800476
477static const struct address_space_operations aio_ctx_aops = {
Matthew Wilcox (Oracle)46de8b972022-02-09 20:22:13 +0000478 .dirty_folio = noop_dirty_folio,
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400479 .migrate_folio = aio_migrate_folio,
Gu Zheng36bc08c2013-07-16 17:56:16 +0800480};
481
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300482static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483{
484 struct aio_ring *ring;
Zach Brown41003a72013-05-07 16:18:25 -0700485 struct mm_struct *mm = current->mm;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900486 unsigned long size, unused;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 int nr_pages;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800488 int i;
489 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490
491 /* Compensate for the ring buffer's head/tail overlap entry */
492 nr_events += 2; /* 1 is required, 2 for good luck */
493
494 size = sizeof(struct aio_ring);
495 size += sizeof(struct io_event) * nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496
Gu Zheng36bc08c2013-07-16 17:56:16 +0800497 nr_pages = PFN_UP(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 if (nr_pages < 0)
499 return -EINVAL;
500
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400501 file = aio_private_file(ctx, nr_pages);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800502 if (IS_ERR(file)) {
503 ctx->aio_ring_file = NULL;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400504 return -ENOMEM;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800505 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506
Gu Zheng36bc08c2013-07-16 17:56:16 +0800507 ctx->aio_ring_file = file;
508 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
509 / sizeof(struct io_event);
510
Kefeng Wang16594e62024-03-21 21:16:40 +0800511 ctx->ring_folios = ctx->internal_folios;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 if (nr_pages > AIO_RING_PAGES) {
Kefeng Wang16594e62024-03-21 21:16:40 +0800513 ctx->ring_folios = kcalloc(nr_pages, sizeof(struct folio *),
514 GFP_KERNEL);
515 if (!ctx->ring_folios) {
Gu Zhengd1b94322013-12-04 18:19:06 +0800516 put_aio_ring_file(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 return -ENOMEM;
Gu Zhengd1b94322013-12-04 18:19:06 +0800518 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 }
520
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900521 for (i = 0; i < nr_pages; i++) {
Kefeng Wang75a07b52024-03-21 21:16:38 +0800522 struct folio *folio;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900523
Kefeng Wang75a07b52024-03-21 21:16:38 +0800524 folio = __filemap_get_folio(file->f_mapping, i,
525 FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
526 GFP_USER | __GFP_ZERO);
527 if (IS_ERR(folio))
528 break;
529
530 pr_debug("pid(%d) [%d] folio->count=%d\n", current->pid, i,
531 folio_ref_count(folio));
532 folio_end_read(folio, true);
533
Kefeng Wang16594e62024-03-21 21:16:40 +0800534 ctx->ring_folios[i] = folio;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900535 }
536 ctx->nr_pages = i;
537
538 if (unlikely(i != nr_pages)) {
539 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400540 return -ENOMEM;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900541 }
542
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700543 ctx->mmap_size = nr_pages * PAGE_SIZE;
544 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800545
Michel Lespinassed8ed45c2020-06-08 21:33:25 -0700546 if (mmap_write_lock_killable(mm)) {
Michal Hocko013373e2016-05-23 16:25:59 -0700547 ctx->mmap_size = 0;
548 aio_free_ring(ctx);
549 return -EINTR;
550 }
551
Peter Collingbourne45e55302020-08-06 23:23:37 -0700552 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
553 PROT_READ | PROT_WRITE,
Yu-cheng Yu592b5fa2023-06-12 17:10:30 -0700554 MAP_SHARED, 0, 0, &unused, NULL);
Michel Lespinassed8ed45c2020-06-08 21:33:25 -0700555 mmap_write_unlock(mm);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700556 if (IS_ERR((void *)ctx->mmap_base)) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700557 ctx->mmap_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400559 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560 }
561
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700562 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400563
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700564 ctx->user_id = ctx->mmap_base;
565 ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566
Kefeng Wang16594e62024-03-21 21:16:40 +0800567 ring = folio_address(ctx->ring_folios[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 ring->nr = nr_events; /* user copy */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400569 ring->id = ~0U;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 ring->head = ring->tail = 0;
571 ring->magic = AIO_RING_MAGIC;
572 ring->compat_features = AIO_RING_COMPAT_FEATURES;
573 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
574 ring->header_length = sizeof(struct aio_ring);
Kefeng Wang16594e62024-03-21 21:16:40 +0800575 flush_dcache_folio(ctx->ring_folios[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576
577 return 0;
578}
579
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
581#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
582#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
583
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100584void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
Kent Overstreet0460fef2013-05-07 16:18:49 -0700585{
Bart Van Assche961ebd12024-03-04 15:57:15 -0800586 struct aio_kiocb *req;
587 struct kioctx *ctx;
Kent Overstreet0460fef2013-05-07 16:18:49 -0700588 unsigned long flags;
589
Bart Van Asscheb820de72024-02-15 12:47:38 -0800590 /*
591 * kiocb didn't come from aio or is neither a read nor a write, hence
592 * ignore it.
593 */
594 if (!(iocb->ki_flags & IOCB_AIO_RW))
595 return;
596
Bart Van Assche961ebd12024-03-04 15:57:15 -0800597 req = container_of(iocb, struct aio_kiocb, rw);
598
Christoph Hellwig75321b52018-04-09 14:57:56 +0200599 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
600 return;
601
Bart Van Assche961ebd12024-03-04 15:57:15 -0800602 ctx = req->ki_ctx;
603
Kent Overstreet0460fef2013-05-07 16:18:49 -0700604 spin_lock_irqsave(&ctx->ctx_lock, flags);
Christoph Hellwig75321b52018-04-09 14:57:56 +0200605 list_add_tail(&req->ki_list, &ctx->active_reqs);
Kent Overstreet0460fef2013-05-07 16:18:49 -0700606 req->ki_cancel = cancel;
Kent Overstreet0460fef2013-05-07 16:18:49 -0700607 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
608}
609EXPORT_SYMBOL(kiocb_set_cancel_fn);
610
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700611/*
612 * free_ioctx() should be RCU delayed to synchronize against the RCU
613 * protected lookup_ioctx() and also needs process context to call
Tejun Heof7298632018-03-14 12:45:15 -0700614 * aio_free_ring(). Use rcu_work.
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700615 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700616static void free_ioctx(struct work_struct *work)
Kent Overstreet36f55882013-05-07 16:18:41 -0700617{
Tejun Heof7298632018-03-14 12:45:15 -0700618 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
619 free_rwork);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700620 pr_debug("freeing %p\n", ctx);
621
622 aio_free_ring(ctx);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000623 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400624 percpu_ref_exit(&ctx->reqs);
625 percpu_ref_exit(&ctx->users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700626 kmem_cache_free(kioctx_cachep, ctx);
627}
628
Kent Overstreete34ecee2013-10-10 19:31:47 -0700629static void free_ioctx_reqs(struct percpu_ref *ref)
630{
631 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
632
Anatol Pomozove02ba722014-04-15 11:31:33 -0700633 /* At this point we know that there are no any in-flight requests */
Jens Axboedc48e562015-04-15 11:17:23 -0600634 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
635 complete(&ctx->rq_wait->comp);
Anatol Pomozove02ba722014-04-15 11:31:33 -0700636
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700637 /* Synchronize against RCU protected table->table[] dereferences */
Tejun Heof7298632018-03-14 12:45:15 -0700638 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
639 queue_rcu_work(system_wq, &ctx->free_rwork);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700640}
641
Kent Overstreet36f55882013-05-07 16:18:41 -0700642/*
643 * When this function runs, the kioctx has been removed from the "hash table"
644 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
645 * now it's safe to cancel any that need to be.
646 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700647static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet36f55882013-05-07 16:18:41 -0700648{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700649 struct kioctx *ctx = container_of(ref, struct kioctx, users);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100650 struct aio_kiocb *req;
Kent Overstreet36f55882013-05-07 16:18:41 -0700651
652 spin_lock_irq(&ctx->ctx_lock);
653
654 while (!list_empty(&ctx->active_reqs)) {
655 req = list_first_entry(&ctx->active_reqs,
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100656 struct aio_kiocb, ki_list);
Christoph Hellwig888933f2018-05-23 14:11:02 +0200657 req->ki_cancel(&req->rw);
Al Viro4faa99962018-05-23 22:53:22 -0400658 list_del_init(&req->ki_list);
Kent Overstreet36f55882013-05-07 16:18:41 -0700659 }
660
661 spin_unlock_irq(&ctx->ctx_lock);
662
Kent Overstreete34ecee2013-10-10 19:31:47 -0700663 percpu_ref_kill(&ctx->reqs);
664 percpu_ref_put(&ctx->reqs);
Kent Overstreet36f55882013-05-07 16:18:41 -0700665}
666
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400667static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
668{
669 unsigned i, new_nr;
670 struct kioctx_table *table, *old;
671 struct aio_ring *ring;
672
673 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200674 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400675
676 while (1) {
677 if (table)
678 for (i = 0; i < table->nr; i++)
Tejun Heod0264c02018-03-14 12:10:17 -0700679 if (!rcu_access_pointer(table->table[i])) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400680 ctx->id = i;
Tejun Heod0264c02018-03-14 12:10:17 -0700681 rcu_assign_pointer(table->table[i], ctx);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400682 spin_unlock(&mm->ioctx_lock);
683
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400684 /* While kioctx setup is in progress,
685 * we are protected from page migration
Kefeng Wang16594e62024-03-21 21:16:40 +0800686 * changes ring_folios by ->ring_lock.
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400687 */
Kefeng Wang16594e62024-03-21 21:16:40 +0800688 ring = folio_address(ctx->ring_folios[0]);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400689 ring->id = ctx->id;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400690 return 0;
691 }
692
693 new_nr = (table ? table->nr : 1) * 4;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400694 spin_unlock(&mm->ioctx_lock);
695
Len Baker6446c4f2021-09-19 11:45:39 +0200696 table = kzalloc(struct_size(table, table, new_nr), GFP_KERNEL);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400697 if (!table)
698 return -ENOMEM;
699
700 table->nr = new_nr;
701
702 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200703 old = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400704
705 if (!old) {
706 rcu_assign_pointer(mm->ioctx_table, table);
707 } else if (table->nr > old->nr) {
708 memcpy(table->table, old->table,
709 old->nr * sizeof(struct kioctx *));
710
711 rcu_assign_pointer(mm->ioctx_table, table);
712 kfree_rcu(old, rcu);
713 } else {
714 kfree(table);
715 table = old;
716 }
717 }
718}
719
Kent Overstreete34ecee2013-10-10 19:31:47 -0700720static void aio_nr_sub(unsigned nr)
721{
722 spin_lock(&aio_nr_lock);
723 if (WARN_ON(aio_nr - nr > aio_nr))
724 aio_nr = 0;
725 else
726 aio_nr -= nr;
727 spin_unlock(&aio_nr_lock);
728}
729
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730/* ioctx_alloc
731 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
732 */
733static struct kioctx *ioctx_alloc(unsigned nr_events)
734{
Zach Brown41003a72013-05-07 16:18:25 -0700735 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 struct kioctx *ctx;
Al Viroe23754f2012-03-06 14:33:22 -0500737 int err = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000739 /*
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300740 * Store the original nr_events -- what userspace passed to io_setup(),
741 * for counting against the global limit -- before it changes.
742 */
743 unsigned int max_reqs = nr_events;
744
745 /*
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000746 * We keep track of the number of available ringbuffer slots, to prevent
747 * overflow (reqs_available), and we also use percpu counters for this.
748 *
749 * So since up to half the slots might be on other cpu's percpu counters
750 * and unavailable, double nr_events so userspace sees what they
751 * expected: additionally, we move req_batch slots to/from percpu
752 * counters at a time, so make sure that isn't 0:
753 */
754 nr_events = max(nr_events, num_possible_cpus() * 4);
755 nr_events *= 2;
756
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 /* Prevent overflows */
Al Viro08397ac2015-03-31 11:43:52 -0400758 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 pr_debug("ENOMEM: nr_events too high\n");
760 return ERR_PTR(-EINVAL);
761 }
762
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300763 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 return ERR_PTR(-EAGAIN);
765
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800766 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767 if (!ctx)
768 return ERR_PTR(-ENOMEM);
769
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300770 ctx->max_reqs = max_reqs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400772 spin_lock_init(&ctx->ctx_lock);
773 spin_lock_init(&ctx->completion_lock);
774 mutex_init(&ctx->ring_lock);
775 /* Protect against page migration throughout kiotx setup by keeping
776 * the ring_lock mutex held until setup is complete. */
777 mutex_lock(&ctx->ring_lock);
778 init_waitqueue_head(&ctx->wait);
779
780 INIT_LIST_HEAD(&ctx->active_reqs);
781
Tejun Heo2aad2a82014-09-24 13:31:50 -0400782 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700783 goto err;
784
Tejun Heo2aad2a82014-09-24 13:31:50 -0400785 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700786 goto err;
Kent Overstreet723be6e2013-05-28 15:14:48 -0700787
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000788 ctx->cpu = alloc_percpu(struct kioctx_cpu);
789 if (!ctx->cpu)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700790 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300792 err = aio_setup_ring(ctx, nr_events);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400793 if (err < 0)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700794 goto err;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000795
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000796 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000797 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
Benjamin LaHaise6878ea72013-07-31 10:34:18 -0400798 if (ctx->req_batch < 1)
799 ctx->req_batch = 1;
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000800
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 /* limit the number of system wide aios */
Al Viro9fa1cb32012-03-10 23:14:05 -0500802 spin_lock(&aio_nr_lock);
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300803 if (aio_nr + ctx->max_reqs > aio_max_nr ||
804 aio_nr + ctx->max_reqs < aio_nr) {
Al Viro9fa1cb32012-03-10 23:14:05 -0500805 spin_unlock(&aio_nr_lock);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700806 err = -EAGAIN;
Gu Zhengd1b94322013-12-04 18:19:06 +0800807 goto err_ctx;
Al Viro2dd542b72012-03-10 23:10:35 -0500808 }
809 aio_nr += ctx->max_reqs;
Al Viro9fa1cb32012-03-10 23:14:05 -0500810 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811
Benjamin LaHaise18816862013-12-21 15:49:28 -0500812 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
813 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
Kent Overstreet723be6e2013-05-28 15:14:48 -0700814
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400815 err = ioctx_add_table(ctx, mm);
816 if (err)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700817 goto err_cleanup;
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400818
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400819 /* Release the ring_lock mutex now that all setup is complete. */
820 mutex_unlock(&ctx->ring_lock);
821
Kent Overstreetcaf41672013-05-07 16:18:35 -0700822 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700823 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 return ctx;
825
Kent Overstreete34ecee2013-10-10 19:31:47 -0700826err_cleanup:
827 aio_nr_sub(ctx->max_reqs);
Gu Zhengd1b94322013-12-04 18:19:06 +0800828err_ctx:
Al Virodeeb8522015-04-06 17:57:44 -0400829 atomic_set(&ctx->dead, 1);
830 if (ctx->mmap_size)
831 vm_munmap(ctx->mmap_base, ctx->mmap_size);
Gu Zhengd1b94322013-12-04 18:19:06 +0800832 aio_free_ring(ctx);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700833err:
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400834 mutex_unlock(&ctx->ring_lock);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000835 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400836 percpu_ref_exit(&ctx->reqs);
837 percpu_ref_exit(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreetcaf41672013-05-07 16:18:35 -0700839 pr_debug("error allocating ioctx %d\n", err);
Al Viroe23754f2012-03-06 14:33:22 -0500840 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841}
842
Kent Overstreet36f55882013-05-07 16:18:41 -0700843/* kill_ioctx
844 * Cancels all outstanding aio requests on an aio context. Used
845 * when the processes owning a context have all exited to encourage
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 * the rapid destruction of the kioctx.
847 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400848static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
Jens Axboedc48e562015-04-15 11:17:23 -0600849 struct ctx_rq_wait *wait)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850{
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400851 struct kioctx_table *table;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400852
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400853 spin_lock(&mm->ioctx_lock);
Al Virob2edffd2015-04-06 17:48:54 -0400854 if (atomic_xchg(&ctx->dead, 1)) {
855 spin_unlock(&mm->ioctx_lock);
856 return -EINVAL;
857 }
858
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200859 table = rcu_dereference_raw(mm->ioctx_table);
Tejun Heod0264c02018-03-14 12:10:17 -0700860 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
861 RCU_INIT_POINTER(table->table[ctx->id], NULL);
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400862 spin_unlock(&mm->ioctx_lock);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700863
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700864 /* free_ioctx_reqs() will do the necessary RCU synchronization */
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400865 wake_up_all(&ctx->wait);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700866
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400867 /*
868 * It'd be more correct to do this in free_ioctx(), after all
869 * the outstanding kiocbs have finished - but by then io_destroy
870 * has already returned, so io_setup() could potentially return
871 * -EAGAIN with no ioctxs actually in use (as far as userspace
872 * could tell).
873 */
874 aio_nr_sub(ctx->max_reqs);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400875
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400876 if (ctx->mmap_size)
877 vm_munmap(ctx->mmap_base, ctx->mmap_size);
878
Jens Axboedc48e562015-04-15 11:17:23 -0600879 ctx->rq_wait = wait;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400880 percpu_ref_kill(&ctx->users);
881 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882}
883
Kent Overstreet36f55882013-05-07 16:18:41 -0700884/*
885 * exit_aio: called when the last user of mm goes away. At this point, there is
886 * no way for any new requests to be submited or any of the io_* syscalls to be
887 * called on the context.
888 *
889 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
890 * them.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800892void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893{
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200894 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
Jens Axboedc48e562015-04-15 11:17:23 -0600895 struct ctx_rq_wait wait;
896 int i, skipped;
Jens Axboeabf137d2008-12-09 08:11:22 +0100897
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200898 if (!table)
899 return;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400900
Jens Axboedc48e562015-04-15 11:17:23 -0600901 atomic_set(&wait.count, table->nr);
902 init_completion(&wait.comp);
903
904 skipped = 0;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200905 for (i = 0; i < table->nr; ++i) {
Tejun Heod0264c02018-03-14 12:10:17 -0700906 struct kioctx *ctx =
907 rcu_dereference_protected(table->table[i], true);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400908
Jens Axboedc48e562015-04-15 11:17:23 -0600909 if (!ctx) {
910 skipped++;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200911 continue;
Jens Axboedc48e562015-04-15 11:17:23 -0600912 }
913
Al Viro936af152012-04-20 21:49:41 -0400914 /*
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200915 * We don't need to bother with munmap() here - exit_mmap(mm)
916 * is coming and it'll unmap everything. And we simply can't,
917 * this is not necessarily our ->mm.
918 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
919 * that it needs to unmap the area, just set it to 0.
Al Viro936af152012-04-20 21:49:41 -0400920 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700921 ctx->mmap_size = 0;
Jens Axboedc48e562015-04-15 11:17:23 -0600922 kill_ioctx(mm, ctx, &wait);
923 }
Kent Overstreet36f55882013-05-07 16:18:41 -0700924
Jens Axboedc48e562015-04-15 11:17:23 -0600925 if (!atomic_sub_and_test(skipped, &wait.count)) {
Gu Zheng6098b452014-09-03 17:45:44 +0800926 /* Wait until all IO for the context are done. */
Jens Axboedc48e562015-04-15 11:17:23 -0600927 wait_for_completion(&wait.comp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 }
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200929
930 RCU_INIT_POINTER(mm->ioctx_table, NULL);
931 kfree(table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932}
933
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000934static void put_reqs_available(struct kioctx *ctx, unsigned nr)
935{
936 struct kioctx_cpu *kcpu;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400937 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000938
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400939 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400940 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000941 kcpu->reqs_available += nr;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400942
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000943 while (kcpu->reqs_available >= ctx->req_batch * 2) {
944 kcpu->reqs_available -= ctx->req_batch;
945 atomic_add(ctx->req_batch, &ctx->reqs_available);
946 }
947
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400948 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000949}
950
Christoph Hellwig432c7992018-11-19 15:57:42 -0700951static bool __get_reqs_available(struct kioctx *ctx)
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000952{
953 struct kioctx_cpu *kcpu;
954 bool ret = false;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400955 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000956
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400957 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400958 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000959 if (!kcpu->reqs_available) {
Uros Bizjak38ace0d2022-07-14 18:48:51 +0200960 int avail = atomic_read(&ctx->reqs_available);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000961
962 do {
963 if (avail < ctx->req_batch)
964 goto out;
Uros Bizjak38ace0d2022-07-14 18:48:51 +0200965 } while (!atomic_try_cmpxchg(&ctx->reqs_available,
966 &avail, avail - ctx->req_batch));
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000967
968 kcpu->reqs_available += ctx->req_batch;
969 }
970
971 ret = true;
972 kcpu->reqs_available--;
973out:
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400974 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000975 return ret;
976}
977
Benjamin LaHaised856f322014-08-24 13:14:05 -0400978/* refill_reqs_available
979 * Updates the reqs_available reference counts used for tracking the
980 * number of free slots in the completion ring. This can be called
981 * from aio_complete() (to optimistically update reqs_available) or
982 * from aio_get_req() (the we're out of events case). It must be
983 * called holding ctx->completion_lock.
984 */
985static void refill_reqs_available(struct kioctx *ctx, unsigned head,
986 unsigned tail)
987{
988 unsigned events_in_ring, completed;
989
990 /* Clamp head since userland can write to it. */
991 head %= ctx->nr_events;
992 if (head <= tail)
993 events_in_ring = tail - head;
994 else
995 events_in_ring = ctx->nr_events - (head - tail);
996
997 completed = ctx->completed_events;
998 if (events_in_ring < completed)
999 completed -= events_in_ring;
1000 else
1001 completed = 0;
1002
1003 if (!completed)
1004 return;
1005
1006 ctx->completed_events -= completed;
1007 put_reqs_available(ctx, completed);
1008}
1009
1010/* user_refill_reqs_available
1011 * Called to refill reqs_available when aio_get_req() encounters an
1012 * out of space in the completion ring.
1013 */
1014static void user_refill_reqs_available(struct kioctx *ctx)
1015{
1016 spin_lock_irq(&ctx->completion_lock);
1017 if (ctx->completed_events) {
1018 struct aio_ring *ring;
1019 unsigned head;
1020
1021 /* Access of ring->head may race with aio_read_events_ring()
1022 * here, but that's okay since whether we read the old version
1023 * or the new version, and either will be valid. The important
1024 * part is that head cannot pass tail since we prevent
1025 * aio_complete() from updating tail by holding
1026 * ctx->completion_lock. Even if head is invalid, the check
1027 * against ctx->completed_events below will make sure we do the
1028 * safe/right thing.
1029 */
Kefeng Wang16594e62024-03-21 21:16:40 +08001030 ring = folio_address(ctx->ring_folios[0]);
Benjamin LaHaised856f322014-08-24 13:14:05 -04001031 head = ring->head;
Benjamin LaHaised856f322014-08-24 13:14:05 -04001032
1033 refill_reqs_available(ctx, head, ctx->tail);
1034 }
1035
1036 spin_unlock_irq(&ctx->completion_lock);
1037}
1038
Christoph Hellwig432c7992018-11-19 15:57:42 -07001039static bool get_reqs_available(struct kioctx *ctx)
1040{
1041 if (__get_reqs_available(ctx))
1042 return true;
1043 user_refill_reqs_available(ctx);
1044 return __get_reqs_available(ctx);
1045}
1046
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047/* aio_get_req
Kent Overstreet57282d82013-05-13 13:42:52 -07001048 * Allocate a slot for an aio request.
1049 * Returns NULL if no requests are free.
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001050 *
1051 * The refcount is initialized to 2 - one for the async op completion,
1052 * one for the synchronous code that does this.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001054static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001056 struct aio_kiocb *req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001057
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001058 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059 if (unlikely(!req))
Christoph Hellwig432c7992018-11-19 15:57:42 -07001060 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061
Al Virofa0ca2a2019-03-06 18:21:08 -05001062 if (unlikely(!get_reqs_available(ctx))) {
Wei Yongjun6af1c842019-04-04 08:44:05 +00001063 kmem_cache_free(kiocb_cachep, req);
Al Virofa0ca2a2019-03-06 18:21:08 -05001064 return NULL;
1065 }
1066
Kent Overstreete34ecee2013-10-10 19:31:47 -07001067 percpu_ref_get(&ctx->reqs);
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001068 req->ki_ctx = ctx;
Christoph Hellwig75321b52018-04-09 14:57:56 +02001069 INIT_LIST_HEAD(&req->ki_list);
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001070 refcount_set(&req->ki_refcnt, 2);
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001071 req->ki_eventfd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 return req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073}
1074
Adrian Bunkd5470b52008-04-29 00:58:57 -07001075static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001077 struct aio_ring __user *ring = (void __user *)ctx_id;
Jens Axboeabf137d2008-12-09 08:11:22 +01001078 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -07001079 struct kioctx *ctx, *ret = NULL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001080 struct kioctx_table *table;
1081 unsigned id;
1082
1083 if (get_user(id, &ring->id))
1084 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085
Jens Axboeabf137d2008-12-09 08:11:22 +01001086 rcu_read_lock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001087 table = rcu_dereference(mm->ioctx_table);
Jens Axboeabf137d2008-12-09 08:11:22 +01001088
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001089 if (!table || id >= table->nr)
1090 goto out;
1091
Jeff Moyera538e3f2018-12-11 12:37:49 -05001092 id = array_index_nospec(id, table->nr);
Tejun Heod0264c02018-03-14 12:10:17 -07001093 ctx = rcu_dereference(table->table[id]);
Benjamin LaHaisef30d7042013-08-07 18:23:48 -04001094 if (ctx && ctx->user_id == ctx_id) {
Al Virobaf10562018-05-20 16:46:23 -04001095 if (percpu_ref_tryget_live(&ctx->users))
1096 ret = ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +01001097 }
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001098out:
Jens Axboeabf137d2008-12-09 08:11:22 +01001099 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -07001100 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101}
1102
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001103static inline void iocb_destroy(struct aio_kiocb *iocb)
1104{
Al Viro74259702019-03-06 18:18:31 -05001105 if (iocb->ki_eventfd)
1106 eventfd_ctx_put(iocb->ki_eventfd);
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001107 if (iocb->ki_filp)
1108 fput(iocb->ki_filp);
1109 percpu_ref_put(&iocb->ki_ctx->reqs);
1110 kmem_cache_free(kiocb_cachep, iocb);
1111}
1112
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001113struct aio_waiter {
1114 struct wait_queue_entry w;
1115 size_t min_nr;
1116};
1117
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118/* aio_complete
1119 * Called when the io request on the given iocb is complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 */
Al Viro2bb874c2019-03-07 19:49:55 -05001121static void aio_complete(struct aio_kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122{
1123 struct kioctx *ctx = iocb->ki_ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 struct aio_ring *ring;
Kent Overstreet21b40202013-05-07 16:18:47 -07001125 struct io_event *ev_page, *event;
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001126 unsigned tail, pos, head, avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 /*
Kent Overstreet0460fef2013-05-07 16:18:49 -07001130 * Add a completion event to the ring buffer. Must be done holding
Tang Chen4b30f072013-07-03 15:09:16 -07001131 * ctx->completion_lock to prevent other code from messing with the tail
Kent Overstreet0460fef2013-05-07 16:18:49 -07001132 * pointer since we might be called from irq context.
1133 */
1134 spin_lock_irqsave(&ctx->completion_lock, flags);
1135
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001136 tail = ctx->tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001137 pos = tail + AIO_EVENTS_OFFSET;
1138
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001139 if (++tail >= ctx->nr_events)
Ken Chen4bf69b22005-05-01 08:59:15 -07001140 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141
Kefeng Wang16594e62024-03-21 21:16:40 +08001142 ev_page = folio_address(ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001143 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1144
Al Viroa9339b72019-03-07 19:43:45 -05001145 *event = iocb->ki_res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146
Kefeng Wang16594e62024-03-21 21:16:40 +08001147 flush_dcache_folio(ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001148
Al Viroa9339b72019-03-07 19:43:45 -05001149 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1150 (void __user *)(unsigned long)iocb->ki_res.obj,
1151 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152
1153 /* after flagging the request as done, we
1154 * must never even look at it again
1155 */
1156 smp_wmb(); /* make event visible before updating tail */
1157
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001158 ctx->tail = tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001159
Kefeng Wang16594e62024-03-21 21:16:40 +08001160 ring = folio_address(ctx->ring_folios[0]);
Benjamin LaHaised856f322014-08-24 13:14:05 -04001161 head = ring->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 ring->tail = tail;
Kefeng Wang16594e62024-03-21 21:16:40 +08001163 flush_dcache_folio(ctx->ring_folios[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164
Benjamin LaHaised856f322014-08-24 13:14:05 -04001165 ctx->completed_events++;
1166 if (ctx->completed_events > 1)
1167 refill_reqs_available(ctx, head, tail);
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001168
1169 avail = tail > head
1170 ? tail - head
1171 : tail + ctx->nr_events - head;
Kent Overstreet0460fef2013-05-07 16:18:49 -07001172 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1173
Kent Overstreet21b40202013-05-07 16:18:47 -07001174 pr_debug("added to ring %p at [%u]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001175
1176 /*
1177 * Check if the user asked us to deliver the result through an
1178 * eventfd. The eventfd_signal() function is safe to be called
1179 * from IRQ context.
1180 */
Al Viro74259702019-03-06 18:18:31 -05001181 if (iocb->ki_eventfd)
Christian Brauner36521172023-11-22 13:48:23 +01001182 eventfd_signal(iocb->ki_eventfd);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001183
Quentin Barnes6cb2a212008-03-19 17:00:39 -07001184 /*
1185 * We have to order our ring_info tail store above and test
1186 * of the wait list below outside the wait lock. This is
1187 * like in wake_up_bit() where clearing a bit has to be
1188 * ordered with the unlocked test.
1189 */
1190 smp_mb();
1191
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001192 if (waitqueue_active(&ctx->wait)) {
1193 struct aio_waiter *curr, *next;
1194 unsigned long flags;
1195
1196 spin_lock_irqsave(&ctx->wait.lock, flags);
1197 list_for_each_entry_safe(curr, next, &ctx->wait.head, w.entry)
1198 if (avail >= curr->min_nr) {
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001199 wake_up_process(curr->w.private);
Kent Overstreetcaeb4b02024-03-31 17:52:12 -04001200 list_del_init_careful(&curr->w.entry);
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001201 }
1202 spin_unlock_irqrestore(&ctx->wait.lock, flags);
1203 }
Al Viro2bb874c2019-03-07 19:49:55 -05001204}
1205
1206static inline void iocb_put(struct aio_kiocb *iocb)
1207{
1208 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1209 aio_complete(iocb);
1210 iocb_destroy(iocb);
1211 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212}
1213
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001214/* aio_read_events_ring
Kent Overstreeta31ad382013-05-07 16:18:45 -07001215 * Pull an event off of the ioctx's event ring. Returns the number of
1216 * events fetched
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217 */
Kent Overstreeta31ad382013-05-07 16:18:45 -07001218static long aio_read_events_ring(struct kioctx *ctx,
1219 struct io_event __user *event, long nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 struct aio_ring *ring;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001222 unsigned head, tail, pos;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001223 long ret = 0;
1224 int copy_ret;
1225
Dave Chinner9c9ce762015-02-03 19:29:05 -05001226 /*
1227 * The mutex can block and wake us up and that will cause
1228 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1229 * and repeat. This should be rare enough that it doesn't cause
1230 * peformance issues. See the comment in read_events() for more detail.
1231 */
1232 sched_annotate_sleep();
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001233 mutex_lock(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234
Kefeng Wang16594e62024-03-21 21:16:40 +08001235 /* Access to ->ring_folios here is protected by ctx->ring_lock. */
1236 ring = folio_address(ctx->ring_folios[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001237 head = ring->head;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001238 tail = ring->tail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239
Jeff Moyer2ff396be2014-09-02 13:17:00 -04001240 /*
1241 * Ensure that once we've read the current tail pointer, that
1242 * we also see the events that were stored up to the tail.
1243 */
1244 smp_rmb();
1245
Kent Overstreet5ffac122013-05-09 15:36:07 -07001246 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001247
Kent Overstreet5ffac122013-05-09 15:36:07 -07001248 if (head == tail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 goto out;
1250
Benjamin LaHaiseedfbbf32014-06-24 13:32:51 -04001251 head %= ctx->nr_events;
1252 tail %= ctx->nr_events;
1253
Kent Overstreeta31ad382013-05-07 16:18:45 -07001254 while (ret < nr) {
1255 long avail;
1256 struct io_event *ev;
Kefeng Wang16594e62024-03-21 21:16:40 +08001257 struct folio *folio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258
Kent Overstreet5ffac122013-05-09 15:36:07 -07001259 avail = (head <= tail ? tail : ctx->nr_events) - head;
1260 if (head == tail)
Kent Overstreeta31ad382013-05-07 16:18:45 -07001261 break;
1262
Kent Overstreeta31ad382013-05-07 16:18:45 -07001263 pos = head + AIO_EVENTS_OFFSET;
Kefeng Wang16594e62024-03-21 21:16:40 +08001264 folio = ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE];
Kent Overstreeta31ad382013-05-07 16:18:45 -07001265 pos %= AIO_EVENTS_PER_PAGE;
1266
Al Virod2988bd42018-05-26 19:13:10 -04001267 avail = min(avail, nr - ret);
1268 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1269
Kefeng Wang16594e62024-03-21 21:16:40 +08001270 ev = folio_address(folio);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001271 copy_ret = copy_to_user(event + ret, ev + pos,
1272 sizeof(*ev) * avail);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001273
1274 if (unlikely(copy_ret)) {
1275 ret = -EFAULT;
1276 goto out;
1277 }
1278
1279 ret += avail;
1280 head += avail;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001281 head %= ctx->nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283
Kefeng Wang16594e62024-03-21 21:16:40 +08001284 ring = folio_address(ctx->ring_folios[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001285 ring->head = head;
Kefeng Wang16594e62024-03-21 21:16:40 +08001286 flush_dcache_folio(ctx->ring_folios[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001287
Kent Overstreet5ffac122013-05-09 15:36:07 -07001288 pr_debug("%li h%u t%u\n", ret, head, tail);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001289out:
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001290 mutex_unlock(&ctx->ring_lock);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001291
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 return ret;
1293}
1294
Kent Overstreeta31ad382013-05-07 16:18:45 -07001295static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1296 struct io_event __user *event, long *i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001298 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299
Kent Overstreeta31ad382013-05-07 16:18:45 -07001300 if (ret > 0)
1301 *i += ret;
1302
1303 if (unlikely(atomic_read(&ctx->dead)))
1304 ret = -EINVAL;
1305
1306 if (!*i)
1307 *i = ret;
1308
1309 return ret < 0 || *i >= min_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310}
1311
Kent Overstreeta31ad382013-05-07 16:18:45 -07001312static long read_events(struct kioctx *ctx, long min_nr, long nr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313 struct io_event __user *event,
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07001314 ktime_t until)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315{
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001316 struct hrtimer_sleeper t;
1317 struct aio_waiter w;
1318 long ret = 0, ret2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319
Kent Overstreeta31ad382013-05-07 16:18:45 -07001320 /*
1321 * Note that aio_read_events() is being called as the conditional - i.e.
1322 * we're calling it after prepare_to_wait() has set task state to
1323 * TASK_INTERRUPTIBLE.
1324 *
1325 * But aio_read_events() can block, and if it blocks it's going to flip
1326 * the task state back to TASK_RUNNING.
1327 *
1328 * This should be ok, provided it doesn't flip the state back to
1329 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1330 * will only happen if the mutex_lock() call blocks, and we then find
1331 * the ringbuffer empty. So in practice we should be ok, but it's
1332 * something to be aware of when touching this code.
1333 */
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001334 aio_read_events(ctx, min_nr, nr, event, &ret);
1335 if (until == 0 || ret < 0 || ret >= min_nr)
1336 return ret;
1337
1338 hrtimer_init_sleeper_on_stack(&t, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1339 if (until != KTIME_MAX) {
1340 hrtimer_set_expires_range_ns(&t.timer, until, current->timer_slack_ns);
1341 hrtimer_sleeper_start_expires(&t, HRTIMER_MODE_REL);
1342 }
1343
1344 init_wait(&w.w);
1345
1346 while (1) {
1347 unsigned long nr_got = ret;
1348
1349 w.min_nr = min_nr - ret;
1350
1351 ret2 = prepare_to_wait_event(&ctx->wait, &w.w, TASK_INTERRUPTIBLE);
1352 if (!ret2 && !t.task)
1353 ret2 = -ETIME;
1354
1355 if (aio_read_events(ctx, min_nr, nr, event, &ret) || ret2)
1356 break;
1357
1358 if (nr_got == ret)
1359 schedule();
1360 }
1361
1362 finish_wait(&ctx->wait, &w.w);
1363 hrtimer_cancel(&t.timer);
1364 destroy_hrtimer_on_stack(&t.timer);
1365
Kent Overstreeta31ad382013-05-07 16:18:45 -07001366 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367}
1368
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369/* sys_io_setup:
1370 * Create an aio_context capable of receiving at least nr_events.
1371 * ctxp must not point to an aio_context that already exists, and
1372 * must be initialized to 0 prior to the call. On successful
1373 * creation of the aio_context, *ctxp is filled in with the resulting
1374 * handle. May fail with -EINVAL if *ctxp is not initialized,
1375 * if the specified nr_events exceeds internal limits. May fail
1376 * with -EAGAIN if the specified nr_events exceeds the user's limit
1377 * of available events. May fail with -ENOMEM if insufficient kernel
1378 * resources are available. May fail with -EFAULT if an invalid
1379 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1380 * implemented.
1381 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001382SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383{
1384 struct kioctx *ioctx = NULL;
1385 unsigned long ctx;
1386 long ret;
1387
1388 ret = get_user(ctx, ctxp);
1389 if (unlikely(ret))
1390 goto out;
1391
1392 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001393 if (unlikely(ctx || nr_events == 0)) {
Kinglong Meeacd88d42015-02-04 21:15:59 +08001394 pr_debug("EINVAL: ctx %lu nr_events %u\n",
Zach Brownd55b5fd2005-11-07 00:59:31 -08001395 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 goto out;
1397 }
1398
1399 ioctx = ioctx_alloc(nr_events);
1400 ret = PTR_ERR(ioctx);
1401 if (!IS_ERR(ioctx)) {
1402 ret = put_user(ioctx->user_id, ctxp);
Al Viroa2e18592012-03-20 16:27:57 -04001403 if (ret)
Anatol Pomozove02ba722014-04-15 11:31:33 -07001404 kill_ioctx(current->mm, ioctx, NULL);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001405 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 }
1407
1408out:
1409 return ret;
1410}
1411
Al Viroc00d2c72016-12-20 07:04:57 -05001412#ifdef CONFIG_COMPAT
1413COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1414{
1415 struct kioctx *ioctx = NULL;
1416 unsigned long ctx;
1417 long ret;
1418
1419 ret = get_user(ctx, ctx32p);
1420 if (unlikely(ret))
1421 goto out;
1422
1423 ret = -EINVAL;
1424 if (unlikely(ctx || nr_events == 0)) {
1425 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1426 ctx, nr_events);
1427 goto out;
1428 }
1429
1430 ioctx = ioctx_alloc(nr_events);
1431 ret = PTR_ERR(ioctx);
1432 if (!IS_ERR(ioctx)) {
1433 /* truncating is ok because it's a user address */
1434 ret = put_user((u32)ioctx->user_id, ctx32p);
1435 if (ret)
1436 kill_ioctx(current->mm, ioctx, NULL);
1437 percpu_ref_put(&ioctx->users);
1438 }
1439
1440out:
1441 return ret;
1442}
1443#endif
1444
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445/* sys_io_destroy:
1446 * Destroy the aio_context specified. May cancel any outstanding
1447 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001448 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449 * is invalid.
1450 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001451SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452{
1453 struct kioctx *ioctx = lookup_ioctx(ctx);
1454 if (likely(NULL != ioctx)) {
Jens Axboedc48e562015-04-15 11:17:23 -06001455 struct ctx_rq_wait wait;
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001456 int ret;
Anatol Pomozove02ba722014-04-15 11:31:33 -07001457
Jens Axboedc48e562015-04-15 11:17:23 -06001458 init_completion(&wait.comp);
1459 atomic_set(&wait.count, 1);
1460
Anatol Pomozove02ba722014-04-15 11:31:33 -07001461 /* Pass requests_done to kill_ioctx() where it can be set
1462 * in a thread-safe way. If we try to set it here then we have
1463 * a race condition if two io_destroy() called simultaneously.
1464 */
Jens Axboedc48e562015-04-15 11:17:23 -06001465 ret = kill_ioctx(current->mm, ioctx, &wait);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001466 percpu_ref_put(&ioctx->users);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001467
1468 /* Wait until all IO for the context are done. Otherwise kernel
1469 * keep using user-space buffers even if user thinks the context
1470 * is destroyed.
1471 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001472 if (!ret)
Jens Axboedc48e562015-04-15 11:17:23 -06001473 wait_for_completion(&wait.comp);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001474
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001475 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476 }
Kinglong Meeacd88d42015-02-04 21:15:59 +08001477 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 return -EINVAL;
1479}
1480
Al Viro3c96c7f2018-05-28 13:37:43 -04001481static void aio_remove_iocb(struct aio_kiocb *iocb)
1482{
1483 struct kioctx *ctx = iocb->ki_ctx;
1484 unsigned long flags;
1485
1486 spin_lock_irqsave(&ctx->ctx_lock, flags);
1487 list_del(&iocb->ki_list);
1488 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1489}
1490
Jens Axboe6b19b762021-10-21 09:22:35 -06001491static void aio_complete_rw(struct kiocb *kiocb, long res)
Christoph Hellwig54843f82018-05-02 19:57:21 +02001492{
1493 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1494
Al Viro3c96c7f2018-05-28 13:37:43 -04001495 if (!list_empty_careful(&iocb->ki_list))
1496 aio_remove_iocb(iocb);
1497
Christoph Hellwig54843f82018-05-02 19:57:21 +02001498 if (kiocb->ki_flags & IOCB_WRITE) {
1499 struct inode *inode = file_inode(kiocb->ki_filp);
1500
Christoph Hellwig54843f82018-05-02 19:57:21 +02001501 if (S_ISREG(inode->i_mode))
Amir Goldstein8c3cfa82023-08-17 17:13:35 +03001502 kiocb_end_write(kiocb);
Christoph Hellwig54843f82018-05-02 19:57:21 +02001503 }
1504
Al Viro2bb874c2019-03-07 19:49:55 -05001505 iocb->ki_res.res = res;
Jens Axboe6b19b762021-10-21 09:22:35 -06001506 iocb->ki_res.res2 = 0;
Al Viro2bb874c2019-03-07 19:49:55 -05001507 iocb_put(iocb);
Christoph Hellwig54843f82018-05-02 19:57:21 +02001508}
1509
Prasad Singamsettyc34fc6f2024-06-20 12:53:52 +00001510static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb, int rw_type)
Christoph Hellwig54843f82018-05-02 19:57:21 +02001511{
1512 int ret;
1513
Christoph Hellwig54843f82018-05-02 19:57:21 +02001514 req->ki_complete = aio_complete_rw;
Mike Marshallec51f8e2019-02-05 14:13:35 -05001515 req->private = NULL;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001516 req->ki_pos = iocb->aio_offset;
Bart Van Asscheb820de72024-02-15 12:47:38 -08001517 req->ki_flags = req->ki_filp->f_iocb_flags | IOCB_AIO_RW;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001518 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1519 req->ki_flags |= IOCB_EVENTFD;
Adam Manzanaresd9a08a92018-05-22 10:52:19 -07001520 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1521 /*
1522 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1523 * aio_reqprio is interpreted as an I/O scheduling
1524 * class and priority.
1525 */
1526 ret = ioprio_check_cap(iocb->aio_reqprio);
1527 if (ret) {
Adam Manzanares9a6d9a62018-06-04 10:59:57 -07001528 pr_debug("aio ioprio check cap error: %d\n", ret);
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001529 return ret;
Adam Manzanaresd9a08a92018-05-22 10:52:19 -07001530 }
1531
1532 req->ki_ioprio = iocb->aio_reqprio;
1533 } else
Damien Le Moal76dc8912018-11-20 10:52:36 +09001534 req->ki_ioprio = get_current_ioprio();
Adam Manzanaresd9a08a92018-05-22 10:52:19 -07001535
Prasad Singamsettyc34fc6f2024-06-20 12:53:52 +00001536 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags, rw_type);
Christoph Hellwig54843f82018-05-02 19:57:21 +02001537 if (unlikely(ret))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001538 return ret;
Christoph Hellwig154989e2018-11-22 16:44:07 +01001539
1540 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1541 return 0;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001542}
1543
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001544static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1545 struct iovec **iovec, bool vectored, bool compat,
1546 struct iov_iter *iter)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001547{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001548 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1549 size_t len = iocb->aio_nbytes;
1550
1551 if (!vectored) {
Jens Axboe9fd78742023-12-04 10:47:50 -07001552 ssize_t ret = import_ubuf(rw, buf, len, iter);
Christoph Hellwig89319d312016-10-30 11:42:03 -05001553 *iovec = NULL;
1554 return ret;
1555 }
Christoph Hellwig89cd35c2020-09-25 06:51:41 +02001556
1557 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001558}
1559
Al Viro9061d142018-05-26 19:11:40 -04001560static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001562 switch (ret) {
1563 case -EIOCBQUEUED:
Al Viro9061d142018-05-26 19:11:40 -04001564 break;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001565 case -ERESTARTSYS:
1566 case -ERESTARTNOINTR:
1567 case -ERESTARTNOHAND:
1568 case -ERESTART_RESTARTBLOCK:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001569 /*
1570 * There's no easy way to restart the syscall since other AIO's
1571 * may be already running. Just fail this IO with EINTR.
1572 */
Christoph Hellwig89319d312016-10-30 11:42:03 -05001573 ret = -EINTR;
Gustavo A. R. Silvadf561f662020-08-23 17:36:59 -05001574 fallthrough;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001575 default:
Jens Axboe6b19b762021-10-21 09:22:35 -06001576 req->ki_complete(req, ret);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001577 }
Christoph Hellwig89319d312016-10-30 11:42:03 -05001578}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579
Al Viro958c13c2019-03-06 18:13:00 -05001580static int aio_read(struct kiocb *req, const struct iocb *iocb,
Jens Axboe88a6f182018-11-24 14:46:14 -07001581 bool vectored, bool compat)
Christoph Hellwig89319d312016-10-30 11:42:03 -05001582{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001583 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1584 struct iov_iter iter;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001585 struct file *file;
Al Viro958c13c2019-03-06 18:13:00 -05001586 int ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001587
Prasad Singamsettyc34fc6f2024-06-20 12:53:52 +00001588 ret = aio_prep_rw(req, iocb, READ);
Christoph Hellwig54843f82018-05-02 19:57:21 +02001589 if (ret)
1590 return ret;
1591 file = req->ki_filp;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001592 if (unlikely(!(file->f_mode & FMODE_READ)))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001593 return -EBADF;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001594 if (unlikely(!file->f_op->read_iter))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001595 return -EINVAL;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001596
Al Virode4eda92022-09-15 20:25:47 -04001597 ret = aio_setup_rw(ITER_DEST, iocb, &iovec, vectored, compat, &iter);
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001598 if (ret < 0)
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001599 return ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001600 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1601 if (!ret)
Miklos Szeredi7c98f7c2023-08-28 17:13:18 +02001602 aio_rw_done(req, file->f_op->read_iter(req, &iter));
Christoph Hellwig89319d312016-10-30 11:42:03 -05001603 kfree(iovec);
1604 return ret;
1605}
1606
Al Viro958c13c2019-03-06 18:13:00 -05001607static int aio_write(struct kiocb *req, const struct iocb *iocb,
Jens Axboe88a6f182018-11-24 14:46:14 -07001608 bool vectored, bool compat)
Christoph Hellwig89319d312016-10-30 11:42:03 -05001609{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001610 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1611 struct iov_iter iter;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001612 struct file *file;
Al Viro958c13c2019-03-06 18:13:00 -05001613 int ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001614
Prasad Singamsettyc34fc6f2024-06-20 12:53:52 +00001615 ret = aio_prep_rw(req, iocb, WRITE);
Christoph Hellwig54843f82018-05-02 19:57:21 +02001616 if (ret)
1617 return ret;
1618 file = req->ki_filp;
1619
Christoph Hellwig89319d312016-10-30 11:42:03 -05001620 if (unlikely(!(file->f_mode & FMODE_WRITE)))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001621 return -EBADF;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001622 if (unlikely(!file->f_op->write_iter))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001623 return -EINVAL;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001624
Al Virode4eda92022-09-15 20:25:47 -04001625 ret = aio_setup_rw(ITER_SOURCE, iocb, &iovec, vectored, compat, &iter);
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001626 if (ret < 0)
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001627 return ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001628 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1629 if (!ret) {
Amir Goldstein8c3cfa82023-08-17 17:13:35 +03001630 if (S_ISREG(file_inode(file)->i_mode))
1631 kiocb_start_write(req);
Christoph Hellwig92ce4722018-04-06 09:28:17 +02001632 req->ki_flags |= IOCB_WRITE;
Miklos Szeredi7c98f7c2023-08-28 17:13:18 +02001633 aio_rw_done(req, file->f_op->write_iter(req, &iter));
Christoph Hellwig89319d312016-10-30 11:42:03 -05001634 }
1635 kfree(iovec);
1636 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637}
1638
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001639static void aio_fsync_work(struct work_struct *work)
1640{
Al Viro2bb874c2019-03-07 19:49:55 -05001641 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
Miklos Szeredi530f32f2020-05-14 16:44:24 +02001642 const struct cred *old_cred = override_creds(iocb->fsync.creds);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001643
Al Viro2bb874c2019-03-07 19:49:55 -05001644 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
Miklos Szeredi530f32f2020-05-14 16:44:24 +02001645 revert_creds(old_cred);
1646 put_cred(iocb->fsync.creds);
Al Viro2bb874c2019-03-07 19:49:55 -05001647 iocb_put(iocb);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001648}
1649
Jens Axboe88a6f182018-11-24 14:46:14 -07001650static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1651 bool datasync)
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001652{
1653 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1654 iocb->aio_rw_flags))
1655 return -EINVAL;
Linus Torvaldsa11e1d42018-06-28 09:43:44 -07001656
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001657 if (unlikely(!req->file->f_op->fsync))
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001658 return -EINVAL;
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001659
Miklos Szeredi530f32f2020-05-14 16:44:24 +02001660 req->creds = prepare_creds();
1661 if (!req->creds)
1662 return -ENOMEM;
1663
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001664 req->datasync = datasync;
1665 INIT_WORK(&req->work, aio_fsync_work);
1666 schedule_work(&req->work);
Al Viro9061d142018-05-26 19:11:40 -04001667 return 0;
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001668}
1669
Jens Axboe01d7a352020-02-03 10:33:42 -07001670static void aio_poll_put_work(struct work_struct *work)
1671{
1672 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1673 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1674
1675 iocb_put(iocb);
1676}
1677
Eric Biggers50252e42021-12-08 17:04:55 -08001678/*
1679 * Safely lock the waitqueue which the request is on, synchronizing with the
1680 * case where the ->poll() provider decides to free its waitqueue early.
1681 *
1682 * Returns true on success, meaning that req->head->lock was locked, req->wait
1683 * is on req->head, and an RCU read lock was taken. Returns false if the
1684 * request was already removed from its waitqueue (which might no longer exist).
1685 */
1686static bool poll_iocb_lock_wq(struct poll_iocb *req)
1687{
1688 wait_queue_head_t *head;
1689
1690 /*
1691 * While we hold the waitqueue lock and the waitqueue is nonempty,
1692 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1693 * lock in the first place can race with the waitqueue being freed.
1694 *
1695 * We solve this as eventpoll does: by taking advantage of the fact that
1696 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1697 * we enter rcu_read_lock() and see that the pointer to the queue is
1698 * non-NULL, we can then lock it without the memory being freed out from
1699 * under us, then check whether the request is still on the queue.
1700 *
1701 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1702 * case the caller deletes the entry from the queue, leaving it empty.
1703 * In that case, only RCU prevents the queue memory from being freed.
1704 */
1705 rcu_read_lock();
1706 head = smp_load_acquire(&req->head);
1707 if (head) {
1708 spin_lock(&head->lock);
1709 if (!list_empty(&req->wait.entry))
1710 return true;
1711 spin_unlock(&head->lock);
1712 }
1713 rcu_read_unlock();
1714 return false;
1715}
1716
1717static void poll_iocb_unlock_wq(struct poll_iocb *req)
1718{
1719 spin_unlock(&req->head->lock);
1720 rcu_read_unlock();
1721}
1722
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001723static void aio_poll_complete_work(struct work_struct *work)
1724{
1725 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1726 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1727 struct poll_table_struct pt = { ._key = req->events };
1728 struct kioctx *ctx = iocb->ki_ctx;
1729 __poll_t mask = 0;
1730
1731 if (!READ_ONCE(req->cancelled))
1732 mask = vfs_poll(req->file, &pt) & req->events;
1733
1734 /*
1735 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1736 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1737 * synchronize with them. In the cancellation case the list_del_init
1738 * itself is not actually needed, but harmless so we keep it in to
1739 * avoid further branches in the fast path.
1740 */
1741 spin_lock_irq(&ctx->ctx_lock);
Eric Biggers50252e42021-12-08 17:04:55 -08001742 if (poll_iocb_lock_wq(req)) {
1743 if (!mask && !READ_ONCE(req->cancelled)) {
1744 /*
1745 * The request isn't actually ready to be completed yet.
1746 * Reschedule completion if another wakeup came in.
1747 */
1748 if (req->work_need_resched) {
1749 schedule_work(&req->work);
1750 req->work_need_resched = false;
1751 } else {
1752 req->work_scheduled = false;
1753 }
1754 poll_iocb_unlock_wq(req);
1755 spin_unlock_irq(&ctx->ctx_lock);
1756 return;
Eric Biggers363bee22021-12-08 17:04:54 -08001757 }
Eric Biggers50252e42021-12-08 17:04:55 -08001758 list_del_init(&req->wait.entry);
1759 poll_iocb_unlock_wq(req);
1760 } /* else, POLLFREE has freed the waitqueue, so we must complete */
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001761 list_del_init(&iocb->ki_list);
Al Viroaf5c72b2019-03-07 21:45:41 -05001762 iocb->ki_res.res = mangle_poll(mask);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001763 spin_unlock_irq(&ctx->ctx_lock);
1764
Al Viroaf5c72b2019-03-07 21:45:41 -05001765 iocb_put(iocb);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001766}
1767
1768/* assumes we are called with irqs disabled */
1769static int aio_poll_cancel(struct kiocb *iocb)
1770{
1771 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1772 struct poll_iocb *req = &aiocb->poll;
1773
Eric Biggers50252e42021-12-08 17:04:55 -08001774 if (poll_iocb_lock_wq(req)) {
1775 WRITE_ONCE(req->cancelled, true);
1776 if (!req->work_scheduled) {
1777 schedule_work(&aiocb->poll.work);
1778 req->work_scheduled = true;
1779 }
1780 poll_iocb_unlock_wq(req);
1781 } /* else, the request was force-cancelled by POLLFREE already */
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001782
1783 return 0;
1784}
1785
1786static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1787 void *key)
1788{
1789 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
Christoph Hellwige8693bc2018-07-16 12:25:17 +02001790 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001791 __poll_t mask = key_to_poll(key);
Bart Van Assched3d6a182019-02-08 16:59:49 -08001792 unsigned long flags;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001793
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001794 /* for instances that support it check for an event match first: */
Al Viroaf5c72b2019-03-07 21:45:41 -05001795 if (mask && !(mask & req->events))
1796 return 0;
Christoph Hellwige8693bc2018-07-16 12:25:17 +02001797
Eric Biggers363bee22021-12-08 17:04:54 -08001798 /*
1799 * Complete the request inline if possible. This requires that three
1800 * conditions be met:
1801 * 1. An event mask must have been passed. If a plain wakeup was done
1802 * instead, then mask == 0 and we have to call vfs_poll() to get
1803 * the events, so inline completion isn't possible.
1804 * 2. The completion work must not have already been scheduled.
1805 * 3. ctx_lock must not be busy. We have to use trylock because we
1806 * already hold the waitqueue lock, so this inverts the normal
1807 * locking order. Use irqsave/irqrestore because not all
1808 * filesystems (e.g. fuse) call this function with IRQs disabled,
1809 * yet IRQs have to be disabled before ctx_lock is obtained.
1810 */
1811 if (mask && !req->work_scheduled &&
1812 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
Jens Axboe01d7a352020-02-03 10:33:42 -07001813 struct kioctx *ctx = iocb->ki_ctx;
1814
Eric Biggers363bee22021-12-08 17:04:54 -08001815 list_del_init(&req->wait.entry);
Al Viroaf5c72b2019-03-07 21:45:41 -05001816 list_del(&iocb->ki_list);
1817 iocb->ki_res.res = mangle_poll(mask);
Xie Yongji4b374982021-09-13 19:19:28 +08001818 if (iocb->ki_eventfd && !eventfd_signal_allowed()) {
Jens Axboe01d7a352020-02-03 10:33:42 -07001819 iocb = NULL;
1820 INIT_WORK(&req->work, aio_poll_put_work);
1821 schedule_work(&req->work);
1822 }
1823 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1824 if (iocb)
1825 iocb_put(iocb);
Al Viroaf5c72b2019-03-07 21:45:41 -05001826 } else {
Eric Biggers363bee22021-12-08 17:04:54 -08001827 /*
1828 * Schedule the completion work if needed. If it was already
1829 * scheduled, record that another wakeup came in.
1830 *
1831 * Don't remove the request from the waitqueue here, as it might
1832 * not actually be complete yet (we won't know until vfs_poll()
Eric Biggers50252e42021-12-08 17:04:55 -08001833 * is called), and we must not miss any wakeups. POLLFREE is an
1834 * exception to this; see below.
Eric Biggers363bee22021-12-08 17:04:54 -08001835 */
1836 if (req->work_scheduled) {
1837 req->work_need_resched = true;
1838 } else {
1839 schedule_work(&req->work);
1840 req->work_scheduled = true;
1841 }
Eric Biggers50252e42021-12-08 17:04:55 -08001842
1843 /*
1844 * If the waitqueue is being freed early but we can't complete
1845 * the request inline, we have to tear down the request as best
1846 * we can. That means immediately removing the request from its
1847 * waitqueue and preventing all further accesses to the
1848 * waitqueue via the request. We also need to schedule the
1849 * completion work (done above). Also mark the request as
1850 * cancelled, to potentially skip an unneeded call to ->poll().
1851 */
1852 if (mask & POLLFREE) {
1853 WRITE_ONCE(req->cancelled, true);
1854 list_del_init(&req->wait.entry);
1855
1856 /*
1857 * Careful: this *must* be the last step, since as soon
1858 * as req->head is NULL'ed out, the request can be
1859 * completed and freed, since aio_poll_complete_work()
1860 * will no longer need to take the waitqueue lock.
1861 */
1862 smp_store_release(&req->head, NULL);
1863 }
Christoph Hellwige8693bc2018-07-16 12:25:17 +02001864 }
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001865 return 1;
1866}
1867
1868struct aio_poll_table {
1869 struct poll_table_struct pt;
1870 struct aio_kiocb *iocb;
Eric Biggers50252e42021-12-08 17:04:55 -08001871 bool queued;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001872 int error;
1873};
1874
1875static void
1876aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1877 struct poll_table_struct *p)
1878{
1879 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1880
1881 /* multiple wait queues per file are not supported */
Eric Biggers50252e42021-12-08 17:04:55 -08001882 if (unlikely(pt->queued)) {
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001883 pt->error = -EINVAL;
1884 return;
1885 }
1886
Eric Biggers50252e42021-12-08 17:04:55 -08001887 pt->queued = true;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001888 pt->error = 0;
1889 pt->iocb->poll.head = head;
1890 add_wait_queue(head, &pt->iocb->poll.wait);
1891}
1892
Al Viro958c13c2019-03-06 18:13:00 -05001893static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001894{
1895 struct kioctx *ctx = aiocb->ki_ctx;
1896 struct poll_iocb *req = &aiocb->poll;
1897 struct aio_poll_table apt;
Al Viroaf5c72b2019-03-07 21:45:41 -05001898 bool cancel = false;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001899 __poll_t mask;
1900
1901 /* reject any unknown events outside the normal event mask. */
1902 if ((u16)iocb->aio_buf != iocb->aio_buf)
1903 return -EINVAL;
1904 /* reject fields that are not defined for poll */
1905 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1906 return -EINVAL;
1907
1908 INIT_WORK(&req->work, aio_poll_complete_work);
1909 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001910
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001911 req->head = NULL;
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001912 req->cancelled = false;
Eric Biggers363bee22021-12-08 17:04:54 -08001913 req->work_scheduled = false;
1914 req->work_need_resched = false;
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001915
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001916 apt.pt._qproc = aio_poll_queue_proc;
1917 apt.pt._key = req->events;
1918 apt.iocb = aiocb;
Eric Biggers50252e42021-12-08 17:04:55 -08001919 apt.queued = false;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001920 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1921
1922 /* initialized the list so that we can do list_empty checks */
1923 INIT_LIST_HEAD(&req->wait.entry);
1924 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1925
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001926 mask = vfs_poll(req->file, &apt.pt) & req->events;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001927 spin_lock_irq(&ctx->ctx_lock);
Eric Biggers50252e42021-12-08 17:04:55 -08001928 if (likely(apt.queued)) {
1929 bool on_queue = poll_iocb_lock_wq(req);
1930
1931 if (!on_queue || req->work_scheduled) {
Eric Biggers363bee22021-12-08 17:04:54 -08001932 /*
1933 * aio_poll_wake() already either scheduled the async
1934 * completion work, or completed the request inline.
1935 */
1936 if (apt.error) /* unsupported case: multiple queues */
Al Viroaf5c72b2019-03-07 21:45:41 -05001937 cancel = true;
1938 apt.error = 0;
1939 mask = 0;
1940 }
1941 if (mask || apt.error) {
Eric Biggers363bee22021-12-08 17:04:54 -08001942 /* Steal to complete synchronously. */
Al Viroaf5c72b2019-03-07 21:45:41 -05001943 list_del_init(&req->wait.entry);
1944 } else if (cancel) {
Eric Biggers363bee22021-12-08 17:04:54 -08001945 /* Cancel if possible (may be too late though). */
Al Viroaf5c72b2019-03-07 21:45:41 -05001946 WRITE_ONCE(req->cancelled, true);
Eric Biggers50252e42021-12-08 17:04:55 -08001947 } else if (on_queue) {
Eric Biggers363bee22021-12-08 17:04:54 -08001948 /*
1949 * Actually waiting for an event, so add the request to
1950 * active_reqs so that it can be cancelled if needed.
1951 */
Al Viroaf5c72b2019-03-07 21:45:41 -05001952 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1953 aiocb->ki_cancel = aio_poll_cancel;
1954 }
Eric Biggers50252e42021-12-08 17:04:55 -08001955 if (on_queue)
1956 poll_iocb_unlock_wq(req);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001957 }
Al Viroaf5c72b2019-03-07 21:45:41 -05001958 if (mask) { /* no async, we'd stolen it */
1959 aiocb->ki_res.res = mangle_poll(mask);
1960 apt.error = 0;
1961 }
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001962 spin_unlock_irq(&ctx->ctx_lock);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001963 if (mask)
Al Viroaf5c72b2019-03-07 21:45:41 -05001964 iocb_put(aiocb);
1965 return apt.error;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001966}
1967
Jens Axboe88a6f182018-11-24 14:46:14 -07001968static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
Al Viro7316b492019-03-06 18:24:51 -05001969 struct iocb __user *user_iocb, struct aio_kiocb *req,
1970 bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971{
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001972 req->ki_filp = fget(iocb->aio_fildes);
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001973 if (unlikely(!req->ki_filp))
Al Viro7316b492019-03-06 18:24:51 -05001974 return -EBADF;
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001975
Jens Axboe88a6f182018-11-24 14:46:14 -07001976 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
Al Viro74259702019-03-06 18:18:31 -05001977 struct eventfd_ctx *eventfd;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001978 /*
1979 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1980 * instance of the file* now. The file descriptor must be
1981 * an eventfd() fd, and will be signaled for each completed
1982 * event using the eventfd_signal() function.
1983 */
Al Viro74259702019-03-06 18:18:31 -05001984 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
Al Viro7316b492019-03-06 18:24:51 -05001985 if (IS_ERR(eventfd))
Dan Carpenter18bfb9c2019-04-03 09:22:35 +03001986 return PTR_ERR(eventfd);
Al Viro7316b492019-03-06 18:24:51 -05001987
Al Viro74259702019-03-06 18:18:31 -05001988 req->ki_eventfd = eventfd;
Goldwyn Rodrigues9830f4b2017-06-20 07:05:42 -05001989 }
1990
Al Viro7316b492019-03-06 18:24:51 -05001991 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001992 pr_debug("EFAULT: aio_key\n");
Al Viro7316b492019-03-06 18:24:51 -05001993 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 }
1995
Al Viroa9339b72019-03-07 19:43:45 -05001996 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1997 req->ki_res.data = iocb->aio_data;
1998 req->ki_res.res = 0;
1999 req->ki_res.res2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000
Jens Axboe88a6f182018-11-24 14:46:14 -07002001 switch (iocb->aio_lio_opcode) {
Christoph Hellwig89319d312016-10-30 11:42:03 -05002002 case IOCB_CMD_PREAD:
Al Viro7316b492019-03-06 18:24:51 -05002003 return aio_read(&req->rw, iocb, false, compat);
Christoph Hellwig89319d312016-10-30 11:42:03 -05002004 case IOCB_CMD_PWRITE:
Al Viro7316b492019-03-06 18:24:51 -05002005 return aio_write(&req->rw, iocb, false, compat);
Christoph Hellwig89319d312016-10-30 11:42:03 -05002006 case IOCB_CMD_PREADV:
Al Viro7316b492019-03-06 18:24:51 -05002007 return aio_read(&req->rw, iocb, true, compat);
Christoph Hellwig89319d312016-10-30 11:42:03 -05002008 case IOCB_CMD_PWRITEV:
Al Viro7316b492019-03-06 18:24:51 -05002009 return aio_write(&req->rw, iocb, true, compat);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02002010 case IOCB_CMD_FSYNC:
Al Viro7316b492019-03-06 18:24:51 -05002011 return aio_fsync(&req->fsync, iocb, false);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02002012 case IOCB_CMD_FDSYNC:
Al Viro7316b492019-03-06 18:24:51 -05002013 return aio_fsync(&req->fsync, iocb, true);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02002014 case IOCB_CMD_POLL:
Al Viro7316b492019-03-06 18:24:51 -05002015 return aio_poll(req, iocb);
Christoph Hellwig89319d312016-10-30 11:42:03 -05002016 default:
Jens Axboe88a6f182018-11-24 14:46:14 -07002017 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
Al Viro7316b492019-03-06 18:24:51 -05002018 return -EINVAL;
Christoph Hellwig89319d312016-10-30 11:42:03 -05002019 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020}
2021
Jens Axboe88a6f182018-11-24 14:46:14 -07002022static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
2023 bool compat)
2024{
Al Viro7316b492019-03-06 18:24:51 -05002025 struct aio_kiocb *req;
Jens Axboe88a6f182018-11-24 14:46:14 -07002026 struct iocb iocb;
Al Viro7316b492019-03-06 18:24:51 -05002027 int err;
Jens Axboe88a6f182018-11-24 14:46:14 -07002028
2029 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
2030 return -EFAULT;
2031
Al Viro7316b492019-03-06 18:24:51 -05002032 /* enforce forwards compatibility on users */
2033 if (unlikely(iocb.aio_reserved2)) {
2034 pr_debug("EINVAL: reserve field set\n");
2035 return -EINVAL;
2036 }
2037
2038 /* prevent overflows */
2039 if (unlikely(
2040 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2041 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2042 ((ssize_t)iocb.aio_nbytes < 0)
2043 )) {
2044 pr_debug("EINVAL: overflow check\n");
2045 return -EINVAL;
2046 }
2047
2048 req = aio_get_req(ctx);
2049 if (unlikely(!req))
2050 return -EAGAIN;
2051
2052 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2053
2054 /* Done with the synchronous reference */
2055 iocb_put(req);
2056
2057 /*
2058 * If err is 0, we'd either done aio_complete() ourselves or have
2059 * arranged for that to be done asynchronously. Anything non-zero
2060 * means that we need to destroy req ourselves.
2061 */
2062 if (unlikely(err)) {
2063 iocb_destroy(req);
2064 put_reqs_available(ctx, 1);
2065 }
2066 return err;
Jens Axboe88a6f182018-11-24 14:46:14 -07002067}
2068
Jeff Moyer9d85cba72010-05-26 14:44:26 -07002069/* sys_io_submit:
2070 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2071 * the number of iocbs queued. May return -EINVAL if the aio_context
2072 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2073 * *iocbpp[0] is not properly initialized, if the operation specified
2074 * is invalid for the file descriptor in the iocb. May fail with
2075 * -EFAULT if any of the data structures point to invalid data. May
2076 * fail with -EBADF if the file descriptor specified in the first
2077 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2078 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2079 * fail with -ENOSYS if not implemented.
2080 */
2081SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2082 struct iocb __user * __user *, iocbpp)
2083{
Al Viro67ba0492018-05-26 20:10:07 -04002084 struct kioctx *ctx;
2085 long ret = 0;
2086 int i = 0;
2087 struct blk_plug plug;
2088
2089 if (unlikely(nr < 0))
2090 return -EINVAL;
2091
Al Viro67ba0492018-05-26 20:10:07 -04002092 ctx = lookup_ioctx(ctx_id);
2093 if (unlikely(!ctx)) {
2094 pr_debug("EINVAL: invalid context id\n");
2095 return -EINVAL;
2096 }
2097
Al Viro1da92772018-05-26 20:10:07 -04002098 if (nr > ctx->nr_events)
2099 nr = ctx->nr_events;
2100
Jens Axboea79d40e2018-12-04 09:45:32 -07002101 if (nr > AIO_PLUG_THRESHOLD)
2102 blk_start_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04002103 for (i = 0; i < nr; i++) {
2104 struct iocb __user *user_iocb;
2105
2106 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2107 ret = -EFAULT;
2108 break;
2109 }
2110
2111 ret = io_submit_one(ctx, user_iocb, false);
2112 if (ret)
2113 break;
2114 }
Jens Axboea79d40e2018-12-04 09:45:32 -07002115 if (nr > AIO_PLUG_THRESHOLD)
2116 blk_finish_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04002117
2118 percpu_ref_put(&ctx->users);
2119 return i ? i : ret;
Jeff Moyer9d85cba72010-05-26 14:44:26 -07002120}
2121
Al Viroc00d2c72016-12-20 07:04:57 -05002122#ifdef CONFIG_COMPAT
Al Viroc00d2c72016-12-20 07:04:57 -05002123COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
Al Viro67ba0492018-05-26 20:10:07 -04002124 int, nr, compat_uptr_t __user *, iocbpp)
Al Viroc00d2c72016-12-20 07:04:57 -05002125{
Al Viro67ba0492018-05-26 20:10:07 -04002126 struct kioctx *ctx;
2127 long ret = 0;
2128 int i = 0;
2129 struct blk_plug plug;
Al Viroc00d2c72016-12-20 07:04:57 -05002130
2131 if (unlikely(nr < 0))
2132 return -EINVAL;
2133
Al Viro67ba0492018-05-26 20:10:07 -04002134 ctx = lookup_ioctx(ctx_id);
2135 if (unlikely(!ctx)) {
2136 pr_debug("EINVAL: invalid context id\n");
2137 return -EINVAL;
2138 }
2139
Al Viro1da92772018-05-26 20:10:07 -04002140 if (nr > ctx->nr_events)
2141 nr = ctx->nr_events;
2142
Jens Axboea79d40e2018-12-04 09:45:32 -07002143 if (nr > AIO_PLUG_THRESHOLD)
2144 blk_start_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04002145 for (i = 0; i < nr; i++) {
2146 compat_uptr_t user_iocb;
2147
2148 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2149 ret = -EFAULT;
2150 break;
2151 }
2152
2153 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2154 if (ret)
2155 break;
2156 }
Jens Axboea79d40e2018-12-04 09:45:32 -07002157 if (nr > AIO_PLUG_THRESHOLD)
2158 blk_finish_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04002159
2160 percpu_ref_put(&ctx->users);
2161 return i ? i : ret;
Al Viroc00d2c72016-12-20 07:04:57 -05002162}
2163#endif
2164
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165/* sys_io_cancel:
Bart Van Assche28468cb2024-03-04 10:29:44 -08002166 * Attempts to cancel an iocb previously passed to io_submit. If
2167 * the operation is successfully cancelled, the resulting event is
2168 * copied into the memory pointed to by result without being placed
2169 * into the completion queue and 0 is returned. May fail with
2170 * -EFAULT if any of the data structures pointed to are invalid.
2171 * May fail with -EINVAL if aio_context specified by ctx_id is
2172 * invalid. May fail with -EAGAIN if the iocb specified was not
2173 * cancelled. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 */
Heiko Carstens002c8972009-01-14 14:14:18 +01002175SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2176 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 struct kioctx *ctx;
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01002179 struct aio_kiocb *kiocb;
Christoph Hellwig888933f2018-05-23 14:11:02 +02002180 int ret = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181 u32 key;
Al Viroa9339b72019-03-07 19:43:45 -05002182 u64 obj = (u64)(unsigned long)iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183
Christoph Hellwigf3a27522018-03-30 11:19:25 +02002184 if (unlikely(get_user(key, &iocb->aio_key)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 return -EFAULT;
Christoph Hellwigf3a27522018-03-30 11:19:25 +02002186 if (unlikely(key != KIOCB_KEY))
2187 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188
2189 ctx = lookup_ioctx(ctx_id);
2190 if (unlikely(!ctx))
2191 return -EINVAL;
2192
2193 spin_lock_irq(&ctx->ctx_lock);
Al Viro833f4152019-03-11 19:00:36 -04002194 /* TODO: use a hash or array, this sucks. */
2195 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
Al Viroa9339b72019-03-07 19:43:45 -05002196 if (kiocb->ki_res.obj == obj) {
Al Viro833f4152019-03-11 19:00:36 -04002197 ret = kiocb->ki_cancel(&kiocb->rw);
2198 list_del_init(&kiocb->ki_list);
2199 break;
2200 }
Christoph Hellwig888933f2018-05-23 14:11:02 +02002201 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 spin_unlock_irq(&ctx->ctx_lock);
2203
Bart Van Assche28468cb2024-03-04 10:29:44 -08002204 if (!ret) {
2205 /*
2206 * The result argument is no longer used - the io_event is
2207 * always delivered via the ring buffer. -EINPROGRESS indicates
2208 * cancellation is progress:
2209 */
2210 ret = -EINPROGRESS;
2211 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212
Kent Overstreet723be6e2013-05-28 15:14:48 -07002213 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214
2215 return ret;
2216}
2217
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002218static long do_io_getevents(aio_context_t ctx_id,
2219 long min_nr,
2220 long nr,
2221 struct io_event __user *events,
2222 struct timespec64 *ts)
2223{
2224 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2225 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2226 long ret = -EINVAL;
2227
2228 if (likely(ioctx)) {
2229 if (likely(min_nr <= nr && min_nr >= 0))
2230 ret = read_events(ioctx, min_nr, nr, events, until);
2231 percpu_ref_put(&ioctx->users);
2232 }
2233
2234 return ret;
2235}
2236
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237/* io_getevents:
2238 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07002239 * the completion queue for the aio_context specified by ctx_id. If
2240 * it succeeds, the number of read events is returned. May fail with
2241 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2242 * out of range, if timeout is out of range. May fail with -EFAULT
2243 * if any of the memory specified is invalid. May return 0 or
2244 * < min_nr if the timeout specified by timeout has elapsed
2245 * before sufficient events are available, where timeout == NULL
2246 * specifies an infinite timeout. Note that the timeout pointed to by
Jeff Moyer69008072013-05-24 15:55:24 -07002247 * timeout is relative. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248 */
Arnd Bergmann3ca47e92019-04-23 17:43:50 +02002249#ifdef CONFIG_64BIT
Deepa Dinamani7a353972018-09-19 21:41:08 -07002250
Heiko Carstens002c8972009-01-14 14:14:18 +01002251SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2252 long, min_nr,
2253 long, nr,
2254 struct io_event __user *, events,
Deepa Dinamani7a353972018-09-19 21:41:08 -07002255 struct __kernel_timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256{
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002257 struct timespec64 ts;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002258 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002260 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2261 return -EFAULT;
2262
2263 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2264 if (!ret && signal_pending(current))
2265 ret = -EINTR;
2266 return ret;
2267}
2268
Deepa Dinamani7a353972018-09-19 21:41:08 -07002269#endif
2270
Christoph Hellwig9ba546c2018-07-11 15:48:46 +02002271struct __aio_sigset {
2272 const sigset_t __user *sigmask;
2273 size_t sigsetsize;
2274};
2275
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002276SYSCALL_DEFINE6(io_pgetevents,
2277 aio_context_t, ctx_id,
2278 long, min_nr,
2279 long, nr,
2280 struct io_event __user *, events,
Deepa Dinamani7a353972018-09-19 21:41:08 -07002281 struct __kernel_timespec __user *, timeout,
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002282 const struct __aio_sigset __user *, usig)
2283{
2284 struct __aio_sigset ksig = { NULL, };
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002285 struct timespec64 ts;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002286 bool interrupted;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002287 int ret;
2288
2289 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2290 return -EFAULT;
2291
2292 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2293 return -EFAULT;
2294
Oleg Nesterovb7724342019-07-16 16:29:53 -07002295 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
Deepa Dinamani7a353972018-09-19 21:41:08 -07002296 if (ret)
2297 return ret;
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002298
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002299 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002300
2301 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002302 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002303 if (interrupted && !ret)
Deepa Dinamani7a353972018-09-19 21:41:08 -07002304 ret = -ERESTARTNOHAND;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002305
2306 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307}
Al Viroc00d2c72016-12-20 07:04:57 -05002308
Deepa Dinamani7a353972018-09-19 21:41:08 -07002309#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2310
2311SYSCALL_DEFINE6(io_pgetevents_time32,
2312 aio_context_t, ctx_id,
2313 long, min_nr,
2314 long, nr,
2315 struct io_event __user *, events,
2316 struct old_timespec32 __user *, timeout,
2317 const struct __aio_sigset __user *, usig)
2318{
2319 struct __aio_sigset ksig = { NULL, };
Deepa Dinamani7a353972018-09-19 21:41:08 -07002320 struct timespec64 ts;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002321 bool interrupted;
Deepa Dinamani7a353972018-09-19 21:41:08 -07002322 int ret;
2323
2324 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2325 return -EFAULT;
2326
2327 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2328 return -EFAULT;
2329
Deepa Dinamanided653c2018-09-19 21:41:04 -07002330
Oleg Nesterovb7724342019-07-16 16:29:53 -07002331 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
Deepa Dinamanided653c2018-09-19 21:41:04 -07002332 if (ret)
2333 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334
2335 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002336
2337 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002338 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002339 if (interrupted && !ret)
Deepa Dinamani854a6ed2018-09-19 21:41:05 -07002340 ret = -ERESTARTNOHAND;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341
2342 return ret;
2343}
Al Viroc00d2c72016-12-20 07:04:57 -05002344
Deepa Dinamani7a353972018-09-19 21:41:08 -07002345#endif
2346
2347#if defined(CONFIG_COMPAT_32BIT_TIME)
2348
Arnd Bergmann8dabe722019-01-07 00:33:08 +01002349SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2350 __s32, min_nr,
2351 __s32, nr,
2352 struct io_event __user *, events,
2353 struct old_timespec32 __user *, timeout)
Al Viroc00d2c72016-12-20 07:04:57 -05002354{
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002355 struct timespec64 t;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002356 int ret;
Al Viroc00d2c72016-12-20 07:04:57 -05002357
Arnd Bergmann9afc5ee2018-07-13 12:52:28 +02002358 if (timeout && get_old_timespec32(&t, timeout))
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002359 return -EFAULT;
2360
2361 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2362 if (!ret && signal_pending(current))
2363 ret = -EINTR;
2364 return ret;
2365}
2366
Deepa Dinamani7a353972018-09-19 21:41:08 -07002367#endif
2368
2369#ifdef CONFIG_COMPAT
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002370
2371struct __compat_aio_sigset {
Guillem Jover97eba802019-08-21 05:38:20 +02002372 compat_uptr_t sigmask;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002373 compat_size_t sigsetsize;
2374};
2375
Deepa Dinamani7a353972018-09-19 21:41:08 -07002376#if defined(CONFIG_COMPAT_32BIT_TIME)
2377
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002378COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2379 compat_aio_context_t, ctx_id,
2380 compat_long_t, min_nr,
2381 compat_long_t, nr,
2382 struct io_event __user *, events,
Arnd Bergmann9afc5ee2018-07-13 12:52:28 +02002383 struct old_timespec32 __user *, timeout,
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002384 const struct __compat_aio_sigset __user *, usig)
2385{
Guillem Jover97eba802019-08-21 05:38:20 +02002386 struct __compat_aio_sigset ksig = { 0, };
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002387 struct timespec64 t;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002388 bool interrupted;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002389 int ret;
2390
Arnd Bergmann9afc5ee2018-07-13 12:52:28 +02002391 if (timeout && get_old_timespec32(&t, timeout))
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002392 return -EFAULT;
2393
2394 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2395 return -EFAULT;
2396
Guillem Jover97eba802019-08-21 05:38:20 +02002397 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
Deepa Dinamanided653c2018-09-19 21:41:04 -07002398 if (ret)
2399 return ret;
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002400
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002401 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002402
2403 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002404 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002405 if (interrupted && !ret)
Deepa Dinamani854a6ed2018-09-19 21:41:05 -07002406 ret = -ERESTARTNOHAND;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002407
2408 return ret;
Al Viroc00d2c72016-12-20 07:04:57 -05002409}
Deepa Dinamani7a353972018-09-19 21:41:08 -07002410
2411#endif
2412
2413COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2414 compat_aio_context_t, ctx_id,
2415 compat_long_t, min_nr,
2416 compat_long_t, nr,
2417 struct io_event __user *, events,
2418 struct __kernel_timespec __user *, timeout,
2419 const struct __compat_aio_sigset __user *, usig)
2420{
Guillem Jover97eba802019-08-21 05:38:20 +02002421 struct __compat_aio_sigset ksig = { 0, };
Deepa Dinamani7a353972018-09-19 21:41:08 -07002422 struct timespec64 t;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002423 bool interrupted;
Deepa Dinamani7a353972018-09-19 21:41:08 -07002424 int ret;
2425
2426 if (timeout && get_timespec64(&t, timeout))
2427 return -EFAULT;
2428
2429 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2430 return -EFAULT;
2431
Guillem Jover97eba802019-08-21 05:38:20 +02002432 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
Deepa Dinamani7a353972018-09-19 21:41:08 -07002433 if (ret)
2434 return ret;
2435
2436 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002437
2438 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002439 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002440 if (interrupted && !ret)
Deepa Dinamani7a353972018-09-19 21:41:08 -07002441 ret = -ERESTARTNOHAND;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442
2443 return ret;
2444}
2445#endif