blob: 57c9f7c077e60148c24f756f93c151cd650a4c67 [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;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800413 int rc;
414
Jérôme Glisse2916ecc2017-09-08 16:12:06 -0700415 /*
416 * We cannot support the _NO_COPY case here, because copy needs to
417 * happen under the ctx->completion_lock. That does not work with the
418 * migration workflow of MIGRATE_SYNC_NO_COPY.
419 */
420 if (mode == MIGRATE_SYNC_NO_COPY)
421 return -EINVAL;
422
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500423 rc = 0;
424
Matthew Wilcox (Oracle)600f1112023-11-17 21:58:23 +0000425 /* mapping->i_private_lock here protects against the kioctx teardown. */
426 spin_lock(&mapping->i_private_lock);
427 ctx = mapping->i_private_data;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400428 if (!ctx) {
429 rc = -EINVAL;
430 goto out;
431 }
432
433 /* The ring_lock mutex. The prevents aio_read_events() from writing
434 * to the ring's head, and prevents page migration from mucking in
435 * a partially initialized kiotx.
436 */
437 if (!mutex_trylock(&ctx->ring_lock)) {
438 rc = -EAGAIN;
439 goto out;
440 }
441
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400442 idx = src->index;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400443 if (idx < (pgoff_t)ctx->nr_pages) {
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400444 /* Make sure the old folio hasn't already been changed */
Kefeng Wang16594e62024-03-21 21:16:40 +0800445 if (ctx->ring_folios[idx] != src)
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400446 rc = -EAGAIN;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500447 } else
448 rc = -EINVAL;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500449
450 if (rc != 0)
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400451 goto out_unlock;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500452
Gu Zheng36bc08c2013-07-16 17:56:16 +0800453 /* Writeback must be complete */
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400454 BUG_ON(folio_test_writeback(src));
455 folio_get(dst);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800456
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400457 rc = folio_migrate_mapping(mapping, dst, src, 1);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800458 if (rc != MIGRATEPAGE_SUCCESS) {
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400459 folio_put(dst);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400460 goto out_unlock;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800461 }
462
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400463 /* Take completion_lock to prevent other writes to the ring buffer
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400464 * while the old folio is copied to the new. This prevents new
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400465 * events from being lost.
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400466 */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400467 spin_lock_irqsave(&ctx->completion_lock, flags);
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400468 folio_migrate_copy(dst, src);
Kefeng Wang16594e62024-03-21 21:16:40 +0800469 BUG_ON(ctx->ring_folios[idx] != src);
470 ctx->ring_folios[idx] = dst;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400471 spin_unlock_irqrestore(&ctx->completion_lock, flags);
472
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400473 /* The old folio is no longer accessible. */
474 folio_put(src);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400475
476out_unlock:
477 mutex_unlock(&ctx->ring_lock);
478out:
Matthew Wilcox (Oracle)600f1112023-11-17 21:58:23 +0000479 spin_unlock(&mapping->i_private_lock);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800480 return rc;
481}
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400482#else
483#define aio_migrate_folio NULL
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400484#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800485
486static const struct address_space_operations aio_ctx_aops = {
Matthew Wilcox (Oracle)46de8b972022-02-09 20:22:13 +0000487 .dirty_folio = noop_dirty_folio,
Matthew Wilcox (Oracle)36489512022-06-06 10:47:21 -0400488 .migrate_folio = aio_migrate_folio,
Gu Zheng36bc08c2013-07-16 17:56:16 +0800489};
490
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300491static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492{
493 struct aio_ring *ring;
Zach Brown41003a72013-05-07 16:18:25 -0700494 struct mm_struct *mm = current->mm;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900495 unsigned long size, unused;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496 int nr_pages;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800497 int i;
498 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499
500 /* Compensate for the ring buffer's head/tail overlap entry */
501 nr_events += 2; /* 1 is required, 2 for good luck */
502
503 size = sizeof(struct aio_ring);
504 size += sizeof(struct io_event) * nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505
Gu Zheng36bc08c2013-07-16 17:56:16 +0800506 nr_pages = PFN_UP(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 if (nr_pages < 0)
508 return -EINVAL;
509
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400510 file = aio_private_file(ctx, nr_pages);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800511 if (IS_ERR(file)) {
512 ctx->aio_ring_file = NULL;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400513 return -ENOMEM;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800514 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515
Gu Zheng36bc08c2013-07-16 17:56:16 +0800516 ctx->aio_ring_file = file;
517 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
518 / sizeof(struct io_event);
519
Kefeng Wang16594e62024-03-21 21:16:40 +0800520 ctx->ring_folios = ctx->internal_folios;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 if (nr_pages > AIO_RING_PAGES) {
Kefeng Wang16594e62024-03-21 21:16:40 +0800522 ctx->ring_folios = kcalloc(nr_pages, sizeof(struct folio *),
523 GFP_KERNEL);
524 if (!ctx->ring_folios) {
Gu Zhengd1b94322013-12-04 18:19:06 +0800525 put_aio_ring_file(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 return -ENOMEM;
Gu Zhengd1b94322013-12-04 18:19:06 +0800527 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 }
529
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900530 for (i = 0; i < nr_pages; i++) {
Kefeng Wang75a07b52024-03-21 21:16:38 +0800531 struct folio *folio;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900532
Kefeng Wang75a07b52024-03-21 21:16:38 +0800533 folio = __filemap_get_folio(file->f_mapping, i,
534 FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
535 GFP_USER | __GFP_ZERO);
536 if (IS_ERR(folio))
537 break;
538
539 pr_debug("pid(%d) [%d] folio->count=%d\n", current->pid, i,
540 folio_ref_count(folio));
541 folio_end_read(folio, true);
542
Kefeng Wang16594e62024-03-21 21:16:40 +0800543 ctx->ring_folios[i] = folio;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900544 }
545 ctx->nr_pages = i;
546
547 if (unlikely(i != nr_pages)) {
548 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400549 return -ENOMEM;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900550 }
551
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700552 ctx->mmap_size = nr_pages * PAGE_SIZE;
553 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800554
Michel Lespinassed8ed45c2020-06-08 21:33:25 -0700555 if (mmap_write_lock_killable(mm)) {
Michal Hocko013373e2016-05-23 16:25:59 -0700556 ctx->mmap_size = 0;
557 aio_free_ring(ctx);
558 return -EINTR;
559 }
560
Peter Collingbourne45e55302020-08-06 23:23:37 -0700561 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
562 PROT_READ | PROT_WRITE,
Yu-cheng Yu592b5fa2023-06-12 17:10:30 -0700563 MAP_SHARED, 0, 0, &unused, NULL);
Michel Lespinassed8ed45c2020-06-08 21:33:25 -0700564 mmap_write_unlock(mm);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700565 if (IS_ERR((void *)ctx->mmap_base)) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700566 ctx->mmap_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400568 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 }
570
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700571 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400572
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700573 ctx->user_id = ctx->mmap_base;
574 ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575
Kefeng Wang16594e62024-03-21 21:16:40 +0800576 ring = folio_address(ctx->ring_folios[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 ring->nr = nr_events; /* user copy */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400578 ring->id = ~0U;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 ring->head = ring->tail = 0;
580 ring->magic = AIO_RING_MAGIC;
581 ring->compat_features = AIO_RING_COMPAT_FEATURES;
582 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
583 ring->header_length = sizeof(struct aio_ring);
Kefeng Wang16594e62024-03-21 21:16:40 +0800584 flush_dcache_folio(ctx->ring_folios[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585
586 return 0;
587}
588
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
590#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
591#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
592
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100593void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
Kent Overstreet0460fef2013-05-07 16:18:49 -0700594{
Bart Van Assche961ebd12024-03-04 15:57:15 -0800595 struct aio_kiocb *req;
596 struct kioctx *ctx;
Kent Overstreet0460fef2013-05-07 16:18:49 -0700597 unsigned long flags;
598
Bart Van Asscheb820de72024-02-15 12:47:38 -0800599 /*
600 * kiocb didn't come from aio or is neither a read nor a write, hence
601 * ignore it.
602 */
603 if (!(iocb->ki_flags & IOCB_AIO_RW))
604 return;
605
Bart Van Assche961ebd12024-03-04 15:57:15 -0800606 req = container_of(iocb, struct aio_kiocb, rw);
607
Christoph Hellwig75321b52018-04-09 14:57:56 +0200608 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
609 return;
610
Bart Van Assche961ebd12024-03-04 15:57:15 -0800611 ctx = req->ki_ctx;
612
Kent Overstreet0460fef2013-05-07 16:18:49 -0700613 spin_lock_irqsave(&ctx->ctx_lock, flags);
Christoph Hellwig75321b52018-04-09 14:57:56 +0200614 list_add_tail(&req->ki_list, &ctx->active_reqs);
Kent Overstreet0460fef2013-05-07 16:18:49 -0700615 req->ki_cancel = cancel;
Kent Overstreet0460fef2013-05-07 16:18:49 -0700616 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
617}
618EXPORT_SYMBOL(kiocb_set_cancel_fn);
619
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700620/*
621 * free_ioctx() should be RCU delayed to synchronize against the RCU
622 * protected lookup_ioctx() and also needs process context to call
Tejun Heof7298632018-03-14 12:45:15 -0700623 * aio_free_ring(). Use rcu_work.
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700624 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700625static void free_ioctx(struct work_struct *work)
Kent Overstreet36f55882013-05-07 16:18:41 -0700626{
Tejun Heof7298632018-03-14 12:45:15 -0700627 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
628 free_rwork);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700629 pr_debug("freeing %p\n", ctx);
630
631 aio_free_ring(ctx);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000632 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400633 percpu_ref_exit(&ctx->reqs);
634 percpu_ref_exit(&ctx->users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700635 kmem_cache_free(kioctx_cachep, ctx);
636}
637
Kent Overstreete34ecee2013-10-10 19:31:47 -0700638static void free_ioctx_reqs(struct percpu_ref *ref)
639{
640 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
641
Anatol Pomozove02ba722014-04-15 11:31:33 -0700642 /* At this point we know that there are no any in-flight requests */
Jens Axboedc48e562015-04-15 11:17:23 -0600643 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
644 complete(&ctx->rq_wait->comp);
Anatol Pomozove02ba722014-04-15 11:31:33 -0700645
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700646 /* Synchronize against RCU protected table->table[] dereferences */
Tejun Heof7298632018-03-14 12:45:15 -0700647 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
648 queue_rcu_work(system_wq, &ctx->free_rwork);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700649}
650
Kent Overstreet36f55882013-05-07 16:18:41 -0700651/*
652 * When this function runs, the kioctx has been removed from the "hash table"
653 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
654 * now it's safe to cancel any that need to be.
655 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700656static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet36f55882013-05-07 16:18:41 -0700657{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700658 struct kioctx *ctx = container_of(ref, struct kioctx, users);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100659 struct aio_kiocb *req;
Kent Overstreet36f55882013-05-07 16:18:41 -0700660
661 spin_lock_irq(&ctx->ctx_lock);
662
663 while (!list_empty(&ctx->active_reqs)) {
664 req = list_first_entry(&ctx->active_reqs,
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100665 struct aio_kiocb, ki_list);
Christoph Hellwig888933f2018-05-23 14:11:02 +0200666 req->ki_cancel(&req->rw);
Al Viro4faa99962018-05-23 22:53:22 -0400667 list_del_init(&req->ki_list);
Kent Overstreet36f55882013-05-07 16:18:41 -0700668 }
669
670 spin_unlock_irq(&ctx->ctx_lock);
671
Kent Overstreete34ecee2013-10-10 19:31:47 -0700672 percpu_ref_kill(&ctx->reqs);
673 percpu_ref_put(&ctx->reqs);
Kent Overstreet36f55882013-05-07 16:18:41 -0700674}
675
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400676static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
677{
678 unsigned i, new_nr;
679 struct kioctx_table *table, *old;
680 struct aio_ring *ring;
681
682 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200683 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400684
685 while (1) {
686 if (table)
687 for (i = 0; i < table->nr; i++)
Tejun Heod0264c02018-03-14 12:10:17 -0700688 if (!rcu_access_pointer(table->table[i])) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400689 ctx->id = i;
Tejun Heod0264c02018-03-14 12:10:17 -0700690 rcu_assign_pointer(table->table[i], ctx);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400691 spin_unlock(&mm->ioctx_lock);
692
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400693 /* While kioctx setup is in progress,
694 * we are protected from page migration
Kefeng Wang16594e62024-03-21 21:16:40 +0800695 * changes ring_folios by ->ring_lock.
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400696 */
Kefeng Wang16594e62024-03-21 21:16:40 +0800697 ring = folio_address(ctx->ring_folios[0]);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400698 ring->id = ctx->id;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400699 return 0;
700 }
701
702 new_nr = (table ? table->nr : 1) * 4;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400703 spin_unlock(&mm->ioctx_lock);
704
Len Baker6446c4f2021-09-19 11:45:39 +0200705 table = kzalloc(struct_size(table, table, new_nr), GFP_KERNEL);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400706 if (!table)
707 return -ENOMEM;
708
709 table->nr = new_nr;
710
711 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200712 old = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400713
714 if (!old) {
715 rcu_assign_pointer(mm->ioctx_table, table);
716 } else if (table->nr > old->nr) {
717 memcpy(table->table, old->table,
718 old->nr * sizeof(struct kioctx *));
719
720 rcu_assign_pointer(mm->ioctx_table, table);
721 kfree_rcu(old, rcu);
722 } else {
723 kfree(table);
724 table = old;
725 }
726 }
727}
728
Kent Overstreete34ecee2013-10-10 19:31:47 -0700729static void aio_nr_sub(unsigned nr)
730{
731 spin_lock(&aio_nr_lock);
732 if (WARN_ON(aio_nr - nr > aio_nr))
733 aio_nr = 0;
734 else
735 aio_nr -= nr;
736 spin_unlock(&aio_nr_lock);
737}
738
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739/* ioctx_alloc
740 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
741 */
742static struct kioctx *ioctx_alloc(unsigned nr_events)
743{
Zach Brown41003a72013-05-07 16:18:25 -0700744 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 struct kioctx *ctx;
Al Viroe23754f2012-03-06 14:33:22 -0500746 int err = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000748 /*
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300749 * Store the original nr_events -- what userspace passed to io_setup(),
750 * for counting against the global limit -- before it changes.
751 */
752 unsigned int max_reqs = nr_events;
753
754 /*
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000755 * We keep track of the number of available ringbuffer slots, to prevent
756 * overflow (reqs_available), and we also use percpu counters for this.
757 *
758 * So since up to half the slots might be on other cpu's percpu counters
759 * and unavailable, double nr_events so userspace sees what they
760 * expected: additionally, we move req_batch slots to/from percpu
761 * counters at a time, so make sure that isn't 0:
762 */
763 nr_events = max(nr_events, num_possible_cpus() * 4);
764 nr_events *= 2;
765
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766 /* Prevent overflows */
Al Viro08397ac2015-03-31 11:43:52 -0400767 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 pr_debug("ENOMEM: nr_events too high\n");
769 return ERR_PTR(-EINVAL);
770 }
771
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300772 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773 return ERR_PTR(-EAGAIN);
774
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800775 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 if (!ctx)
777 return ERR_PTR(-ENOMEM);
778
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300779 ctx->max_reqs = max_reqs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400781 spin_lock_init(&ctx->ctx_lock);
782 spin_lock_init(&ctx->completion_lock);
783 mutex_init(&ctx->ring_lock);
784 /* Protect against page migration throughout kiotx setup by keeping
785 * the ring_lock mutex held until setup is complete. */
786 mutex_lock(&ctx->ring_lock);
787 init_waitqueue_head(&ctx->wait);
788
789 INIT_LIST_HEAD(&ctx->active_reqs);
790
Tejun Heo2aad2a82014-09-24 13:31:50 -0400791 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700792 goto err;
793
Tejun Heo2aad2a82014-09-24 13:31:50 -0400794 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700795 goto err;
Kent Overstreet723be6e2013-05-28 15:14:48 -0700796
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000797 ctx->cpu = alloc_percpu(struct kioctx_cpu);
798 if (!ctx->cpu)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700799 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300801 err = aio_setup_ring(ctx, nr_events);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400802 if (err < 0)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700803 goto err;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000804
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000805 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000806 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
Benjamin LaHaise6878ea72013-07-31 10:34:18 -0400807 if (ctx->req_batch < 1)
808 ctx->req_batch = 1;
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000809
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 /* limit the number of system wide aios */
Al Viro9fa1cb32012-03-10 23:14:05 -0500811 spin_lock(&aio_nr_lock);
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300812 if (aio_nr + ctx->max_reqs > aio_max_nr ||
813 aio_nr + ctx->max_reqs < aio_nr) {
Al Viro9fa1cb32012-03-10 23:14:05 -0500814 spin_unlock(&aio_nr_lock);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700815 err = -EAGAIN;
Gu Zhengd1b94322013-12-04 18:19:06 +0800816 goto err_ctx;
Al Viro2dd542b72012-03-10 23:10:35 -0500817 }
818 aio_nr += ctx->max_reqs;
Al Viro9fa1cb32012-03-10 23:14:05 -0500819 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820
Benjamin LaHaise18816862013-12-21 15:49:28 -0500821 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
822 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
Kent Overstreet723be6e2013-05-28 15:14:48 -0700823
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400824 err = ioctx_add_table(ctx, mm);
825 if (err)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700826 goto err_cleanup;
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400827
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400828 /* Release the ring_lock mutex now that all setup is complete. */
829 mutex_unlock(&ctx->ring_lock);
830
Kent Overstreetcaf41672013-05-07 16:18:35 -0700831 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700832 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 return ctx;
834
Kent Overstreete34ecee2013-10-10 19:31:47 -0700835err_cleanup:
836 aio_nr_sub(ctx->max_reqs);
Gu Zhengd1b94322013-12-04 18:19:06 +0800837err_ctx:
Al Virodeeb8522015-04-06 17:57:44 -0400838 atomic_set(&ctx->dead, 1);
839 if (ctx->mmap_size)
840 vm_munmap(ctx->mmap_base, ctx->mmap_size);
Gu Zhengd1b94322013-12-04 18:19:06 +0800841 aio_free_ring(ctx);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700842err:
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400843 mutex_unlock(&ctx->ring_lock);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000844 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400845 percpu_ref_exit(&ctx->reqs);
846 percpu_ref_exit(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreetcaf41672013-05-07 16:18:35 -0700848 pr_debug("error allocating ioctx %d\n", err);
Al Viroe23754f2012-03-06 14:33:22 -0500849 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850}
851
Kent Overstreet36f55882013-05-07 16:18:41 -0700852/* kill_ioctx
853 * Cancels all outstanding aio requests on an aio context. Used
854 * when the processes owning a context have all exited to encourage
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 * the rapid destruction of the kioctx.
856 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400857static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
Jens Axboedc48e562015-04-15 11:17:23 -0600858 struct ctx_rq_wait *wait)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859{
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400860 struct kioctx_table *table;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400861
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400862 spin_lock(&mm->ioctx_lock);
Al Virob2edffd2015-04-06 17:48:54 -0400863 if (atomic_xchg(&ctx->dead, 1)) {
864 spin_unlock(&mm->ioctx_lock);
865 return -EINVAL;
866 }
867
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200868 table = rcu_dereference_raw(mm->ioctx_table);
Tejun Heod0264c02018-03-14 12:10:17 -0700869 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
870 RCU_INIT_POINTER(table->table[ctx->id], NULL);
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400871 spin_unlock(&mm->ioctx_lock);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700872
Tejun Heoa6d7cff2018-03-14 12:10:17 -0700873 /* free_ioctx_reqs() will do the necessary RCU synchronization */
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400874 wake_up_all(&ctx->wait);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700875
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400876 /*
877 * It'd be more correct to do this in free_ioctx(), after all
878 * the outstanding kiocbs have finished - but by then io_destroy
879 * has already returned, so io_setup() could potentially return
880 * -EAGAIN with no ioctxs actually in use (as far as userspace
881 * could tell).
882 */
883 aio_nr_sub(ctx->max_reqs);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400884
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400885 if (ctx->mmap_size)
886 vm_munmap(ctx->mmap_base, ctx->mmap_size);
887
Jens Axboedc48e562015-04-15 11:17:23 -0600888 ctx->rq_wait = wait;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400889 percpu_ref_kill(&ctx->users);
890 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891}
892
Kent Overstreet36f55882013-05-07 16:18:41 -0700893/*
894 * exit_aio: called when the last user of mm goes away. At this point, there is
895 * no way for any new requests to be submited or any of the io_* syscalls to be
896 * called on the context.
897 *
898 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
899 * them.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800901void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902{
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200903 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
Jens Axboedc48e562015-04-15 11:17:23 -0600904 struct ctx_rq_wait wait;
905 int i, skipped;
Jens Axboeabf137d2008-12-09 08:11:22 +0100906
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200907 if (!table)
908 return;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400909
Jens Axboedc48e562015-04-15 11:17:23 -0600910 atomic_set(&wait.count, table->nr);
911 init_completion(&wait.comp);
912
913 skipped = 0;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200914 for (i = 0; i < table->nr; ++i) {
Tejun Heod0264c02018-03-14 12:10:17 -0700915 struct kioctx *ctx =
916 rcu_dereference_protected(table->table[i], true);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400917
Jens Axboedc48e562015-04-15 11:17:23 -0600918 if (!ctx) {
919 skipped++;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200920 continue;
Jens Axboedc48e562015-04-15 11:17:23 -0600921 }
922
Al Viro936af152012-04-20 21:49:41 -0400923 /*
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200924 * We don't need to bother with munmap() here - exit_mmap(mm)
925 * is coming and it'll unmap everything. And we simply can't,
926 * this is not necessarily our ->mm.
927 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
928 * that it needs to unmap the area, just set it to 0.
Al Viro936af152012-04-20 21:49:41 -0400929 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700930 ctx->mmap_size = 0;
Jens Axboedc48e562015-04-15 11:17:23 -0600931 kill_ioctx(mm, ctx, &wait);
932 }
Kent Overstreet36f55882013-05-07 16:18:41 -0700933
Jens Axboedc48e562015-04-15 11:17:23 -0600934 if (!atomic_sub_and_test(skipped, &wait.count)) {
Gu Zheng6098b452014-09-03 17:45:44 +0800935 /* Wait until all IO for the context are done. */
Jens Axboedc48e562015-04-15 11:17:23 -0600936 wait_for_completion(&wait.comp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 }
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200938
939 RCU_INIT_POINTER(mm->ioctx_table, NULL);
940 kfree(table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941}
942
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000943static void put_reqs_available(struct kioctx *ctx, unsigned nr)
944{
945 struct kioctx_cpu *kcpu;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400946 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000947
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400948 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400949 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000950 kcpu->reqs_available += nr;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400951
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000952 while (kcpu->reqs_available >= ctx->req_batch * 2) {
953 kcpu->reqs_available -= ctx->req_batch;
954 atomic_add(ctx->req_batch, &ctx->reqs_available);
955 }
956
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400957 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000958}
959
Christoph Hellwig432c7992018-11-19 15:57:42 -0700960static bool __get_reqs_available(struct kioctx *ctx)
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000961{
962 struct kioctx_cpu *kcpu;
963 bool ret = false;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400964 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000965
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400966 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400967 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000968 if (!kcpu->reqs_available) {
Uros Bizjak38ace0d2022-07-14 18:48:51 +0200969 int avail = atomic_read(&ctx->reqs_available);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000970
971 do {
972 if (avail < ctx->req_batch)
973 goto out;
Uros Bizjak38ace0d2022-07-14 18:48:51 +0200974 } while (!atomic_try_cmpxchg(&ctx->reqs_available,
975 &avail, avail - ctx->req_batch));
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000976
977 kcpu->reqs_available += ctx->req_batch;
978 }
979
980 ret = true;
981 kcpu->reqs_available--;
982out:
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400983 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000984 return ret;
985}
986
Benjamin LaHaised856f322014-08-24 13:14:05 -0400987/* refill_reqs_available
988 * Updates the reqs_available reference counts used for tracking the
989 * number of free slots in the completion ring. This can be called
990 * from aio_complete() (to optimistically update reqs_available) or
991 * from aio_get_req() (the we're out of events case). It must be
992 * called holding ctx->completion_lock.
993 */
994static void refill_reqs_available(struct kioctx *ctx, unsigned head,
995 unsigned tail)
996{
997 unsigned events_in_ring, completed;
998
999 /* Clamp head since userland can write to it. */
1000 head %= ctx->nr_events;
1001 if (head <= tail)
1002 events_in_ring = tail - head;
1003 else
1004 events_in_ring = ctx->nr_events - (head - tail);
1005
1006 completed = ctx->completed_events;
1007 if (events_in_ring < completed)
1008 completed -= events_in_ring;
1009 else
1010 completed = 0;
1011
1012 if (!completed)
1013 return;
1014
1015 ctx->completed_events -= completed;
1016 put_reqs_available(ctx, completed);
1017}
1018
1019/* user_refill_reqs_available
1020 * Called to refill reqs_available when aio_get_req() encounters an
1021 * out of space in the completion ring.
1022 */
1023static void user_refill_reqs_available(struct kioctx *ctx)
1024{
1025 spin_lock_irq(&ctx->completion_lock);
1026 if (ctx->completed_events) {
1027 struct aio_ring *ring;
1028 unsigned head;
1029
1030 /* Access of ring->head may race with aio_read_events_ring()
1031 * here, but that's okay since whether we read the old version
1032 * or the new version, and either will be valid. The important
1033 * part is that head cannot pass tail since we prevent
1034 * aio_complete() from updating tail by holding
1035 * ctx->completion_lock. Even if head is invalid, the check
1036 * against ctx->completed_events below will make sure we do the
1037 * safe/right thing.
1038 */
Kefeng Wang16594e62024-03-21 21:16:40 +08001039 ring = folio_address(ctx->ring_folios[0]);
Benjamin LaHaised856f322014-08-24 13:14:05 -04001040 head = ring->head;
Benjamin LaHaised856f322014-08-24 13:14:05 -04001041
1042 refill_reqs_available(ctx, head, ctx->tail);
1043 }
1044
1045 spin_unlock_irq(&ctx->completion_lock);
1046}
1047
Christoph Hellwig432c7992018-11-19 15:57:42 -07001048static bool get_reqs_available(struct kioctx *ctx)
1049{
1050 if (__get_reqs_available(ctx))
1051 return true;
1052 user_refill_reqs_available(ctx);
1053 return __get_reqs_available(ctx);
1054}
1055
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056/* aio_get_req
Kent Overstreet57282d82013-05-13 13:42:52 -07001057 * Allocate a slot for an aio request.
1058 * Returns NULL if no requests are free.
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001059 *
1060 * The refcount is initialized to 2 - one for the async op completion,
1061 * one for the synchronous code that does this.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001063static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001065 struct aio_kiocb *req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001066
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001067 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068 if (unlikely(!req))
Christoph Hellwig432c7992018-11-19 15:57:42 -07001069 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070
Al Virofa0ca2a2019-03-06 18:21:08 -05001071 if (unlikely(!get_reqs_available(ctx))) {
Wei Yongjun6af1c842019-04-04 08:44:05 +00001072 kmem_cache_free(kiocb_cachep, req);
Al Virofa0ca2a2019-03-06 18:21:08 -05001073 return NULL;
1074 }
1075
Kent Overstreete34ecee2013-10-10 19:31:47 -07001076 percpu_ref_get(&ctx->reqs);
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001077 req->ki_ctx = ctx;
Christoph Hellwig75321b52018-04-09 14:57:56 +02001078 INIT_LIST_HEAD(&req->ki_list);
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001079 refcount_set(&req->ki_refcnt, 2);
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001080 req->ki_eventfd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081 return req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082}
1083
Adrian Bunkd5470b52008-04-29 00:58:57 -07001084static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001086 struct aio_ring __user *ring = (void __user *)ctx_id;
Jens Axboeabf137d2008-12-09 08:11:22 +01001087 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -07001088 struct kioctx *ctx, *ret = NULL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001089 struct kioctx_table *table;
1090 unsigned id;
1091
1092 if (get_user(id, &ring->id))
1093 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094
Jens Axboeabf137d2008-12-09 08:11:22 +01001095 rcu_read_lock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001096 table = rcu_dereference(mm->ioctx_table);
Jens Axboeabf137d2008-12-09 08:11:22 +01001097
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001098 if (!table || id >= table->nr)
1099 goto out;
1100
Jeff Moyera538e3f2018-12-11 12:37:49 -05001101 id = array_index_nospec(id, table->nr);
Tejun Heod0264c02018-03-14 12:10:17 -07001102 ctx = rcu_dereference(table->table[id]);
Benjamin LaHaisef30d7042013-08-07 18:23:48 -04001103 if (ctx && ctx->user_id == ctx_id) {
Al Virobaf10562018-05-20 16:46:23 -04001104 if (percpu_ref_tryget_live(&ctx->users))
1105 ret = ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +01001106 }
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001107out:
Jens Axboeabf137d2008-12-09 08:11:22 +01001108 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -07001109 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110}
1111
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001112static inline void iocb_destroy(struct aio_kiocb *iocb)
1113{
Al Viro74259702019-03-06 18:18:31 -05001114 if (iocb->ki_eventfd)
1115 eventfd_ctx_put(iocb->ki_eventfd);
Linus Torvaldsb53119f2019-03-06 20:22:54 -05001116 if (iocb->ki_filp)
1117 fput(iocb->ki_filp);
1118 percpu_ref_put(&iocb->ki_ctx->reqs);
1119 kmem_cache_free(kiocb_cachep, iocb);
1120}
1121
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001122struct aio_waiter {
1123 struct wait_queue_entry w;
1124 size_t min_nr;
1125};
1126
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127/* aio_complete
1128 * Called when the io request on the given iocb is complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 */
Al Viro2bb874c2019-03-07 19:49:55 -05001130static void aio_complete(struct aio_kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131{
1132 struct kioctx *ctx = iocb->ki_ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 struct aio_ring *ring;
Kent Overstreet21b40202013-05-07 16:18:47 -07001134 struct io_event *ev_page, *event;
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001135 unsigned tail, pos, head, avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 /*
Kent Overstreet0460fef2013-05-07 16:18:49 -07001139 * Add a completion event to the ring buffer. Must be done holding
Tang Chen4b30f072013-07-03 15:09:16 -07001140 * ctx->completion_lock to prevent other code from messing with the tail
Kent Overstreet0460fef2013-05-07 16:18:49 -07001141 * pointer since we might be called from irq context.
1142 */
1143 spin_lock_irqsave(&ctx->completion_lock, flags);
1144
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001145 tail = ctx->tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001146 pos = tail + AIO_EVENTS_OFFSET;
1147
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001148 if (++tail >= ctx->nr_events)
Ken Chen4bf69b22005-05-01 08:59:15 -07001149 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150
Kefeng Wang16594e62024-03-21 21:16:40 +08001151 ev_page = folio_address(ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001152 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1153
Al Viroa9339b72019-03-07 19:43:45 -05001154 *event = iocb->ki_res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155
Kefeng Wang16594e62024-03-21 21:16:40 +08001156 flush_dcache_folio(ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001157
Al Viroa9339b72019-03-07 19:43:45 -05001158 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1159 (void __user *)(unsigned long)iocb->ki_res.obj,
1160 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161
1162 /* after flagging the request as done, we
1163 * must never even look at it again
1164 */
1165 smp_wmb(); /* make event visible before updating tail */
1166
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001167 ctx->tail = tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001168
Kefeng Wang16594e62024-03-21 21:16:40 +08001169 ring = folio_address(ctx->ring_folios[0]);
Benjamin LaHaised856f322014-08-24 13:14:05 -04001170 head = ring->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 ring->tail = tail;
Kefeng Wang16594e62024-03-21 21:16:40 +08001172 flush_dcache_folio(ctx->ring_folios[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173
Benjamin LaHaised856f322014-08-24 13:14:05 -04001174 ctx->completed_events++;
1175 if (ctx->completed_events > 1)
1176 refill_reqs_available(ctx, head, tail);
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001177
1178 avail = tail > head
1179 ? tail - head
1180 : tail + ctx->nr_events - head;
Kent Overstreet0460fef2013-05-07 16:18:49 -07001181 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1182
Kent Overstreet21b40202013-05-07 16:18:47 -07001183 pr_debug("added to ring %p at [%u]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001184
1185 /*
1186 * Check if the user asked us to deliver the result through an
1187 * eventfd. The eventfd_signal() function is safe to be called
1188 * from IRQ context.
1189 */
Al Viro74259702019-03-06 18:18:31 -05001190 if (iocb->ki_eventfd)
Christian Brauner36521172023-11-22 13:48:23 +01001191 eventfd_signal(iocb->ki_eventfd);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001192
Quentin Barnes6cb2a212008-03-19 17:00:39 -07001193 /*
1194 * We have to order our ring_info tail store above and test
1195 * of the wait list below outside the wait lock. This is
1196 * like in wake_up_bit() where clearing a bit has to be
1197 * ordered with the unlocked test.
1198 */
1199 smp_mb();
1200
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001201 if (waitqueue_active(&ctx->wait)) {
1202 struct aio_waiter *curr, *next;
1203 unsigned long flags;
1204
1205 spin_lock_irqsave(&ctx->wait.lock, flags);
1206 list_for_each_entry_safe(curr, next, &ctx->wait.head, w.entry)
1207 if (avail >= curr->min_nr) {
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001208 wake_up_process(curr->w.private);
Kent Overstreetcaeb4b02024-03-31 17:52:12 -04001209 list_del_init_careful(&curr->w.entry);
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001210 }
1211 spin_unlock_irqrestore(&ctx->wait.lock, flags);
1212 }
Al Viro2bb874c2019-03-07 19:49:55 -05001213}
1214
1215static inline void iocb_put(struct aio_kiocb *iocb)
1216{
1217 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1218 aio_complete(iocb);
1219 iocb_destroy(iocb);
1220 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221}
1222
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001223/* aio_read_events_ring
Kent Overstreeta31ad382013-05-07 16:18:45 -07001224 * Pull an event off of the ioctx's event ring. Returns the number of
1225 * events fetched
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 */
Kent Overstreeta31ad382013-05-07 16:18:45 -07001227static long aio_read_events_ring(struct kioctx *ctx,
1228 struct io_event __user *event, long nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 struct aio_ring *ring;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001231 unsigned head, tail, pos;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001232 long ret = 0;
1233 int copy_ret;
1234
Dave Chinner9c9ce762015-02-03 19:29:05 -05001235 /*
1236 * The mutex can block and wake us up and that will cause
1237 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1238 * and repeat. This should be rare enough that it doesn't cause
1239 * peformance issues. See the comment in read_events() for more detail.
1240 */
1241 sched_annotate_sleep();
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001242 mutex_lock(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243
Kefeng Wang16594e62024-03-21 21:16:40 +08001244 /* Access to ->ring_folios here is protected by ctx->ring_lock. */
1245 ring = folio_address(ctx->ring_folios[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001246 head = ring->head;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001247 tail = ring->tail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248
Jeff Moyer2ff396be2014-09-02 13:17:00 -04001249 /*
1250 * Ensure that once we've read the current tail pointer, that
1251 * we also see the events that were stored up to the tail.
1252 */
1253 smp_rmb();
1254
Kent Overstreet5ffac122013-05-09 15:36:07 -07001255 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001256
Kent Overstreet5ffac122013-05-09 15:36:07 -07001257 if (head == tail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 goto out;
1259
Benjamin LaHaiseedfbbf32014-06-24 13:32:51 -04001260 head %= ctx->nr_events;
1261 tail %= ctx->nr_events;
1262
Kent Overstreeta31ad382013-05-07 16:18:45 -07001263 while (ret < nr) {
1264 long avail;
1265 struct io_event *ev;
Kefeng Wang16594e62024-03-21 21:16:40 +08001266 struct folio *folio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267
Kent Overstreet5ffac122013-05-09 15:36:07 -07001268 avail = (head <= tail ? tail : ctx->nr_events) - head;
1269 if (head == tail)
Kent Overstreeta31ad382013-05-07 16:18:45 -07001270 break;
1271
Kent Overstreeta31ad382013-05-07 16:18:45 -07001272 pos = head + AIO_EVENTS_OFFSET;
Kefeng Wang16594e62024-03-21 21:16:40 +08001273 folio = ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE];
Kent Overstreeta31ad382013-05-07 16:18:45 -07001274 pos %= AIO_EVENTS_PER_PAGE;
1275
Al Virod2988bd42018-05-26 19:13:10 -04001276 avail = min(avail, nr - ret);
1277 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1278
Kefeng Wang16594e62024-03-21 21:16:40 +08001279 ev = folio_address(folio);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001280 copy_ret = copy_to_user(event + ret, ev + pos,
1281 sizeof(*ev) * avail);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001282
1283 if (unlikely(copy_ret)) {
1284 ret = -EFAULT;
1285 goto out;
1286 }
1287
1288 ret += avail;
1289 head += avail;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001290 head %= ctx->nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292
Kefeng Wang16594e62024-03-21 21:16:40 +08001293 ring = folio_address(ctx->ring_folios[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001294 ring->head = head;
Kefeng Wang16594e62024-03-21 21:16:40 +08001295 flush_dcache_folio(ctx->ring_folios[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001296
Kent Overstreet5ffac122013-05-09 15:36:07 -07001297 pr_debug("%li h%u t%u\n", ret, head, tail);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001298out:
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001299 mutex_unlock(&ctx->ring_lock);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001300
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 return ret;
1302}
1303
Kent Overstreeta31ad382013-05-07 16:18:45 -07001304static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1305 struct io_event __user *event, long *i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001307 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308
Kent Overstreeta31ad382013-05-07 16:18:45 -07001309 if (ret > 0)
1310 *i += ret;
1311
1312 if (unlikely(atomic_read(&ctx->dead)))
1313 ret = -EINVAL;
1314
1315 if (!*i)
1316 *i = ret;
1317
1318 return ret < 0 || *i >= min_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319}
1320
Kent Overstreeta31ad382013-05-07 16:18:45 -07001321static long read_events(struct kioctx *ctx, long min_nr, long nr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 struct io_event __user *event,
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07001323 ktime_t until)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324{
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001325 struct hrtimer_sleeper t;
1326 struct aio_waiter w;
1327 long ret = 0, ret2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328
Kent Overstreeta31ad382013-05-07 16:18:45 -07001329 /*
1330 * Note that aio_read_events() is being called as the conditional - i.e.
1331 * we're calling it after prepare_to_wait() has set task state to
1332 * TASK_INTERRUPTIBLE.
1333 *
1334 * But aio_read_events() can block, and if it blocks it's going to flip
1335 * the task state back to TASK_RUNNING.
1336 *
1337 * This should be ok, provided it doesn't flip the state back to
1338 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1339 * will only happen if the mutex_lock() call blocks, and we then find
1340 * the ringbuffer empty. So in practice we should be ok, but it's
1341 * something to be aware of when touching this code.
1342 */
Kent Overstreet71eb6b62023-11-22 18:42:53 -05001343 aio_read_events(ctx, min_nr, nr, event, &ret);
1344 if (until == 0 || ret < 0 || ret >= min_nr)
1345 return ret;
1346
1347 hrtimer_init_sleeper_on_stack(&t, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1348 if (until != KTIME_MAX) {
1349 hrtimer_set_expires_range_ns(&t.timer, until, current->timer_slack_ns);
1350 hrtimer_sleeper_start_expires(&t, HRTIMER_MODE_REL);
1351 }
1352
1353 init_wait(&w.w);
1354
1355 while (1) {
1356 unsigned long nr_got = ret;
1357
1358 w.min_nr = min_nr - ret;
1359
1360 ret2 = prepare_to_wait_event(&ctx->wait, &w.w, TASK_INTERRUPTIBLE);
1361 if (!ret2 && !t.task)
1362 ret2 = -ETIME;
1363
1364 if (aio_read_events(ctx, min_nr, nr, event, &ret) || ret2)
1365 break;
1366
1367 if (nr_got == ret)
1368 schedule();
1369 }
1370
1371 finish_wait(&ctx->wait, &w.w);
1372 hrtimer_cancel(&t.timer);
1373 destroy_hrtimer_on_stack(&t.timer);
1374
Kent Overstreeta31ad382013-05-07 16:18:45 -07001375 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376}
1377
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378/* sys_io_setup:
1379 * Create an aio_context capable of receiving at least nr_events.
1380 * ctxp must not point to an aio_context that already exists, and
1381 * must be initialized to 0 prior to the call. On successful
1382 * creation of the aio_context, *ctxp is filled in with the resulting
1383 * handle. May fail with -EINVAL if *ctxp is not initialized,
1384 * if the specified nr_events exceeds internal limits. May fail
1385 * with -EAGAIN if the specified nr_events exceeds the user's limit
1386 * of available events. May fail with -ENOMEM if insufficient kernel
1387 * resources are available. May fail with -EFAULT if an invalid
1388 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1389 * implemented.
1390 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001391SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392{
1393 struct kioctx *ioctx = NULL;
1394 unsigned long ctx;
1395 long ret;
1396
1397 ret = get_user(ctx, ctxp);
1398 if (unlikely(ret))
1399 goto out;
1400
1401 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001402 if (unlikely(ctx || nr_events == 0)) {
Kinglong Meeacd88d42015-02-04 21:15:59 +08001403 pr_debug("EINVAL: ctx %lu nr_events %u\n",
Zach Brownd55b5fd2005-11-07 00:59:31 -08001404 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405 goto out;
1406 }
1407
1408 ioctx = ioctx_alloc(nr_events);
1409 ret = PTR_ERR(ioctx);
1410 if (!IS_ERR(ioctx)) {
1411 ret = put_user(ioctx->user_id, ctxp);
Al Viroa2e18592012-03-20 16:27:57 -04001412 if (ret)
Anatol Pomozove02ba722014-04-15 11:31:33 -07001413 kill_ioctx(current->mm, ioctx, NULL);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001414 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 }
1416
1417out:
1418 return ret;
1419}
1420
Al Viroc00d2c72016-12-20 07:04:57 -05001421#ifdef CONFIG_COMPAT
1422COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1423{
1424 struct kioctx *ioctx = NULL;
1425 unsigned long ctx;
1426 long ret;
1427
1428 ret = get_user(ctx, ctx32p);
1429 if (unlikely(ret))
1430 goto out;
1431
1432 ret = -EINVAL;
1433 if (unlikely(ctx || nr_events == 0)) {
1434 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1435 ctx, nr_events);
1436 goto out;
1437 }
1438
1439 ioctx = ioctx_alloc(nr_events);
1440 ret = PTR_ERR(ioctx);
1441 if (!IS_ERR(ioctx)) {
1442 /* truncating is ok because it's a user address */
1443 ret = put_user((u32)ioctx->user_id, ctx32p);
1444 if (ret)
1445 kill_ioctx(current->mm, ioctx, NULL);
1446 percpu_ref_put(&ioctx->users);
1447 }
1448
1449out:
1450 return ret;
1451}
1452#endif
1453
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454/* sys_io_destroy:
1455 * Destroy the aio_context specified. May cancel any outstanding
1456 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001457 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458 * is invalid.
1459 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001460SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461{
1462 struct kioctx *ioctx = lookup_ioctx(ctx);
1463 if (likely(NULL != ioctx)) {
Jens Axboedc48e562015-04-15 11:17:23 -06001464 struct ctx_rq_wait wait;
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001465 int ret;
Anatol Pomozove02ba722014-04-15 11:31:33 -07001466
Jens Axboedc48e562015-04-15 11:17:23 -06001467 init_completion(&wait.comp);
1468 atomic_set(&wait.count, 1);
1469
Anatol Pomozove02ba722014-04-15 11:31:33 -07001470 /* Pass requests_done to kill_ioctx() where it can be set
1471 * in a thread-safe way. If we try to set it here then we have
1472 * a race condition if two io_destroy() called simultaneously.
1473 */
Jens Axboedc48e562015-04-15 11:17:23 -06001474 ret = kill_ioctx(current->mm, ioctx, &wait);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001475 percpu_ref_put(&ioctx->users);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001476
1477 /* Wait until all IO for the context are done. Otherwise kernel
1478 * keep using user-space buffers even if user thinks the context
1479 * is destroyed.
1480 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001481 if (!ret)
Jens Axboedc48e562015-04-15 11:17:23 -06001482 wait_for_completion(&wait.comp);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001483
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001484 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485 }
Kinglong Meeacd88d42015-02-04 21:15:59 +08001486 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 return -EINVAL;
1488}
1489
Al Viro3c96c7f2018-05-28 13:37:43 -04001490static void aio_remove_iocb(struct aio_kiocb *iocb)
1491{
1492 struct kioctx *ctx = iocb->ki_ctx;
1493 unsigned long flags;
1494
1495 spin_lock_irqsave(&ctx->ctx_lock, flags);
1496 list_del(&iocb->ki_list);
1497 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1498}
1499
Jens Axboe6b19b762021-10-21 09:22:35 -06001500static void aio_complete_rw(struct kiocb *kiocb, long res)
Christoph Hellwig54843f82018-05-02 19:57:21 +02001501{
1502 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1503
Al Viro3c96c7f2018-05-28 13:37:43 -04001504 if (!list_empty_careful(&iocb->ki_list))
1505 aio_remove_iocb(iocb);
1506
Christoph Hellwig54843f82018-05-02 19:57:21 +02001507 if (kiocb->ki_flags & IOCB_WRITE) {
1508 struct inode *inode = file_inode(kiocb->ki_filp);
1509
Christoph Hellwig54843f82018-05-02 19:57:21 +02001510 if (S_ISREG(inode->i_mode))
Amir Goldstein8c3cfa82023-08-17 17:13:35 +03001511 kiocb_end_write(kiocb);
Christoph Hellwig54843f82018-05-02 19:57:21 +02001512 }
1513
Al Viro2bb874c2019-03-07 19:49:55 -05001514 iocb->ki_res.res = res;
Jens Axboe6b19b762021-10-21 09:22:35 -06001515 iocb->ki_res.res2 = 0;
Al Viro2bb874c2019-03-07 19:49:55 -05001516 iocb_put(iocb);
Christoph Hellwig54843f82018-05-02 19:57:21 +02001517}
1518
Jens Axboe88a6f182018-11-24 14:46:14 -07001519static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
Christoph Hellwig54843f82018-05-02 19:57:21 +02001520{
1521 int ret;
1522
Christoph Hellwig54843f82018-05-02 19:57:21 +02001523 req->ki_complete = aio_complete_rw;
Mike Marshallec51f8e2019-02-05 14:13:35 -05001524 req->private = NULL;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001525 req->ki_pos = iocb->aio_offset;
Bart Van Asscheb820de72024-02-15 12:47:38 -08001526 req->ki_flags = req->ki_filp->f_iocb_flags | IOCB_AIO_RW;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001527 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1528 req->ki_flags |= IOCB_EVENTFD;
Adam Manzanaresd9a08a92018-05-22 10:52:19 -07001529 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1530 /*
1531 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1532 * aio_reqprio is interpreted as an I/O scheduling
1533 * class and priority.
1534 */
1535 ret = ioprio_check_cap(iocb->aio_reqprio);
1536 if (ret) {
Adam Manzanares9a6d9a62018-06-04 10:59:57 -07001537 pr_debug("aio ioprio check cap error: %d\n", ret);
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001538 return ret;
Adam Manzanaresd9a08a92018-05-22 10:52:19 -07001539 }
1540
1541 req->ki_ioprio = iocb->aio_reqprio;
1542 } else
Damien Le Moal76dc8912018-11-20 10:52:36 +09001543 req->ki_ioprio = get_current_ioprio();
Adam Manzanaresd9a08a92018-05-22 10:52:19 -07001544
Christoph Hellwig54843f82018-05-02 19:57:21 +02001545 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1546 if (unlikely(ret))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001547 return ret;
Christoph Hellwig154989e2018-11-22 16:44:07 +01001548
1549 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1550 return 0;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001551}
1552
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001553static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1554 struct iovec **iovec, bool vectored, bool compat,
1555 struct iov_iter *iter)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001556{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001557 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1558 size_t len = iocb->aio_nbytes;
1559
1560 if (!vectored) {
Jens Axboe9fd78742023-12-04 10:47:50 -07001561 ssize_t ret = import_ubuf(rw, buf, len, iter);
Christoph Hellwig89319d312016-10-30 11:42:03 -05001562 *iovec = NULL;
1563 return ret;
1564 }
Christoph Hellwig89cd35c2020-09-25 06:51:41 +02001565
1566 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001567}
1568
Al Viro9061d142018-05-26 19:11:40 -04001569static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001571 switch (ret) {
1572 case -EIOCBQUEUED:
Al Viro9061d142018-05-26 19:11:40 -04001573 break;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001574 case -ERESTARTSYS:
1575 case -ERESTARTNOINTR:
1576 case -ERESTARTNOHAND:
1577 case -ERESTART_RESTARTBLOCK:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001578 /*
1579 * There's no easy way to restart the syscall since other AIO's
1580 * may be already running. Just fail this IO with EINTR.
1581 */
Christoph Hellwig89319d312016-10-30 11:42:03 -05001582 ret = -EINTR;
Gustavo A. R. Silvadf561f662020-08-23 17:36:59 -05001583 fallthrough;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001584 default:
Jens Axboe6b19b762021-10-21 09:22:35 -06001585 req->ki_complete(req, ret);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001586 }
Christoph Hellwig89319d312016-10-30 11:42:03 -05001587}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588
Al Viro958c13c2019-03-06 18:13:00 -05001589static int aio_read(struct kiocb *req, const struct iocb *iocb,
Jens Axboe88a6f182018-11-24 14:46:14 -07001590 bool vectored, bool compat)
Christoph Hellwig89319d312016-10-30 11:42:03 -05001591{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001592 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1593 struct iov_iter iter;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001594 struct file *file;
Al Viro958c13c2019-03-06 18:13:00 -05001595 int ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001596
Christoph Hellwig54843f82018-05-02 19:57:21 +02001597 ret = aio_prep_rw(req, iocb);
1598 if (ret)
1599 return ret;
1600 file = req->ki_filp;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001601 if (unlikely(!(file->f_mode & FMODE_READ)))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001602 return -EBADF;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001603 if (unlikely(!file->f_op->read_iter))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001604 return -EINVAL;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001605
Al Virode4eda92022-09-15 20:25:47 -04001606 ret = aio_setup_rw(ITER_DEST, iocb, &iovec, vectored, compat, &iter);
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001607 if (ret < 0)
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001608 return ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001609 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1610 if (!ret)
Miklos Szeredi7c98f7c2023-08-28 17:13:18 +02001611 aio_rw_done(req, file->f_op->read_iter(req, &iter));
Christoph Hellwig89319d312016-10-30 11:42:03 -05001612 kfree(iovec);
1613 return ret;
1614}
1615
Al Viro958c13c2019-03-06 18:13:00 -05001616static int aio_write(struct kiocb *req, const struct iocb *iocb,
Jens Axboe88a6f182018-11-24 14:46:14 -07001617 bool vectored, bool compat)
Christoph Hellwig89319d312016-10-30 11:42:03 -05001618{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001619 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1620 struct iov_iter iter;
Christoph Hellwig54843f82018-05-02 19:57:21 +02001621 struct file *file;
Al Viro958c13c2019-03-06 18:13:00 -05001622 int ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001623
Christoph Hellwig54843f82018-05-02 19:57:21 +02001624 ret = aio_prep_rw(req, iocb);
1625 if (ret)
1626 return ret;
1627 file = req->ki_filp;
1628
Christoph Hellwig89319d312016-10-30 11:42:03 -05001629 if (unlikely(!(file->f_mode & FMODE_WRITE)))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001630 return -EBADF;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001631 if (unlikely(!file->f_op->write_iter))
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001632 return -EINVAL;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001633
Al Virode4eda92022-09-15 20:25:47 -04001634 ret = aio_setup_rw(ITER_SOURCE, iocb, &iovec, vectored, compat, &iter);
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001635 if (ret < 0)
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001636 return ret;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001637 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1638 if (!ret) {
Amir Goldstein8c3cfa82023-08-17 17:13:35 +03001639 if (S_ISREG(file_inode(file)->i_mode))
1640 kiocb_start_write(req);
Christoph Hellwig92ce4722018-04-06 09:28:17 +02001641 req->ki_flags |= IOCB_WRITE;
Miklos Szeredi7c98f7c2023-08-28 17:13:18 +02001642 aio_rw_done(req, file->f_op->write_iter(req, &iter));
Christoph Hellwig89319d312016-10-30 11:42:03 -05001643 }
1644 kfree(iovec);
1645 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646}
1647
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001648static void aio_fsync_work(struct work_struct *work)
1649{
Al Viro2bb874c2019-03-07 19:49:55 -05001650 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
Miklos Szeredi530f32f2020-05-14 16:44:24 +02001651 const struct cred *old_cred = override_creds(iocb->fsync.creds);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001652
Al Viro2bb874c2019-03-07 19:49:55 -05001653 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
Miklos Szeredi530f32f2020-05-14 16:44:24 +02001654 revert_creds(old_cred);
1655 put_cred(iocb->fsync.creds);
Al Viro2bb874c2019-03-07 19:49:55 -05001656 iocb_put(iocb);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001657}
1658
Jens Axboe88a6f182018-11-24 14:46:14 -07001659static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1660 bool datasync)
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001661{
1662 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1663 iocb->aio_rw_flags))
1664 return -EINVAL;
Linus Torvaldsa11e1d42018-06-28 09:43:44 -07001665
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001666 if (unlikely(!req->file->f_op->fsync))
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001667 return -EINVAL;
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001668
Miklos Szeredi530f32f2020-05-14 16:44:24 +02001669 req->creds = prepare_creds();
1670 if (!req->creds)
1671 return -ENOMEM;
1672
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001673 req->datasync = datasync;
1674 INIT_WORK(&req->work, aio_fsync_work);
1675 schedule_work(&req->work);
Al Viro9061d142018-05-26 19:11:40 -04001676 return 0;
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02001677}
1678
Jens Axboe01d7a352020-02-03 10:33:42 -07001679static void aio_poll_put_work(struct work_struct *work)
1680{
1681 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1682 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1683
1684 iocb_put(iocb);
1685}
1686
Eric Biggers50252e42021-12-08 17:04:55 -08001687/*
1688 * Safely lock the waitqueue which the request is on, synchronizing with the
1689 * case where the ->poll() provider decides to free its waitqueue early.
1690 *
1691 * Returns true on success, meaning that req->head->lock was locked, req->wait
1692 * is on req->head, and an RCU read lock was taken. Returns false if the
1693 * request was already removed from its waitqueue (which might no longer exist).
1694 */
1695static bool poll_iocb_lock_wq(struct poll_iocb *req)
1696{
1697 wait_queue_head_t *head;
1698
1699 /*
1700 * While we hold the waitqueue lock and the waitqueue is nonempty,
1701 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1702 * lock in the first place can race with the waitqueue being freed.
1703 *
1704 * We solve this as eventpoll does: by taking advantage of the fact that
1705 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1706 * we enter rcu_read_lock() and see that the pointer to the queue is
1707 * non-NULL, we can then lock it without the memory being freed out from
1708 * under us, then check whether the request is still on the queue.
1709 *
1710 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1711 * case the caller deletes the entry from the queue, leaving it empty.
1712 * In that case, only RCU prevents the queue memory from being freed.
1713 */
1714 rcu_read_lock();
1715 head = smp_load_acquire(&req->head);
1716 if (head) {
1717 spin_lock(&head->lock);
1718 if (!list_empty(&req->wait.entry))
1719 return true;
1720 spin_unlock(&head->lock);
1721 }
1722 rcu_read_unlock();
1723 return false;
1724}
1725
1726static void poll_iocb_unlock_wq(struct poll_iocb *req)
1727{
1728 spin_unlock(&req->head->lock);
1729 rcu_read_unlock();
1730}
1731
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001732static void aio_poll_complete_work(struct work_struct *work)
1733{
1734 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1735 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1736 struct poll_table_struct pt = { ._key = req->events };
1737 struct kioctx *ctx = iocb->ki_ctx;
1738 __poll_t mask = 0;
1739
1740 if (!READ_ONCE(req->cancelled))
1741 mask = vfs_poll(req->file, &pt) & req->events;
1742
1743 /*
1744 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1745 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1746 * synchronize with them. In the cancellation case the list_del_init
1747 * itself is not actually needed, but harmless so we keep it in to
1748 * avoid further branches in the fast path.
1749 */
1750 spin_lock_irq(&ctx->ctx_lock);
Eric Biggers50252e42021-12-08 17:04:55 -08001751 if (poll_iocb_lock_wq(req)) {
1752 if (!mask && !READ_ONCE(req->cancelled)) {
1753 /*
1754 * The request isn't actually ready to be completed yet.
1755 * Reschedule completion if another wakeup came in.
1756 */
1757 if (req->work_need_resched) {
1758 schedule_work(&req->work);
1759 req->work_need_resched = false;
1760 } else {
1761 req->work_scheduled = false;
1762 }
1763 poll_iocb_unlock_wq(req);
1764 spin_unlock_irq(&ctx->ctx_lock);
1765 return;
Eric Biggers363bee22021-12-08 17:04:54 -08001766 }
Eric Biggers50252e42021-12-08 17:04:55 -08001767 list_del_init(&req->wait.entry);
1768 poll_iocb_unlock_wq(req);
1769 } /* else, POLLFREE has freed the waitqueue, so we must complete */
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001770 list_del_init(&iocb->ki_list);
Al Viroaf5c72b2019-03-07 21:45:41 -05001771 iocb->ki_res.res = mangle_poll(mask);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001772 spin_unlock_irq(&ctx->ctx_lock);
1773
Al Viroaf5c72b2019-03-07 21:45:41 -05001774 iocb_put(iocb);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001775}
1776
1777/* assumes we are called with irqs disabled */
1778static int aio_poll_cancel(struct kiocb *iocb)
1779{
1780 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1781 struct poll_iocb *req = &aiocb->poll;
1782
Eric Biggers50252e42021-12-08 17:04:55 -08001783 if (poll_iocb_lock_wq(req)) {
1784 WRITE_ONCE(req->cancelled, true);
1785 if (!req->work_scheduled) {
1786 schedule_work(&aiocb->poll.work);
1787 req->work_scheduled = true;
1788 }
1789 poll_iocb_unlock_wq(req);
1790 } /* else, the request was force-cancelled by POLLFREE already */
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001791
1792 return 0;
1793}
1794
1795static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1796 void *key)
1797{
1798 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
Christoph Hellwige8693bc2018-07-16 12:25:17 +02001799 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001800 __poll_t mask = key_to_poll(key);
Bart Van Assched3d6a182019-02-08 16:59:49 -08001801 unsigned long flags;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001802
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001803 /* for instances that support it check for an event match first: */
Al Viroaf5c72b2019-03-07 21:45:41 -05001804 if (mask && !(mask & req->events))
1805 return 0;
Christoph Hellwige8693bc2018-07-16 12:25:17 +02001806
Eric Biggers363bee22021-12-08 17:04:54 -08001807 /*
1808 * Complete the request inline if possible. This requires that three
1809 * conditions be met:
1810 * 1. An event mask must have been passed. If a plain wakeup was done
1811 * instead, then mask == 0 and we have to call vfs_poll() to get
1812 * the events, so inline completion isn't possible.
1813 * 2. The completion work must not have already been scheduled.
1814 * 3. ctx_lock must not be busy. We have to use trylock because we
1815 * already hold the waitqueue lock, so this inverts the normal
1816 * locking order. Use irqsave/irqrestore because not all
1817 * filesystems (e.g. fuse) call this function with IRQs disabled,
1818 * yet IRQs have to be disabled before ctx_lock is obtained.
1819 */
1820 if (mask && !req->work_scheduled &&
1821 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
Jens Axboe01d7a352020-02-03 10:33:42 -07001822 struct kioctx *ctx = iocb->ki_ctx;
1823
Eric Biggers363bee22021-12-08 17:04:54 -08001824 list_del_init(&req->wait.entry);
Al Viroaf5c72b2019-03-07 21:45:41 -05001825 list_del(&iocb->ki_list);
1826 iocb->ki_res.res = mangle_poll(mask);
Xie Yongji4b374982021-09-13 19:19:28 +08001827 if (iocb->ki_eventfd && !eventfd_signal_allowed()) {
Jens Axboe01d7a352020-02-03 10:33:42 -07001828 iocb = NULL;
1829 INIT_WORK(&req->work, aio_poll_put_work);
1830 schedule_work(&req->work);
1831 }
1832 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1833 if (iocb)
1834 iocb_put(iocb);
Al Viroaf5c72b2019-03-07 21:45:41 -05001835 } else {
Eric Biggers363bee22021-12-08 17:04:54 -08001836 /*
1837 * Schedule the completion work if needed. If it was already
1838 * scheduled, record that another wakeup came in.
1839 *
1840 * Don't remove the request from the waitqueue here, as it might
1841 * not actually be complete yet (we won't know until vfs_poll()
Eric Biggers50252e42021-12-08 17:04:55 -08001842 * is called), and we must not miss any wakeups. POLLFREE is an
1843 * exception to this; see below.
Eric Biggers363bee22021-12-08 17:04:54 -08001844 */
1845 if (req->work_scheduled) {
1846 req->work_need_resched = true;
1847 } else {
1848 schedule_work(&req->work);
1849 req->work_scheduled = true;
1850 }
Eric Biggers50252e42021-12-08 17:04:55 -08001851
1852 /*
1853 * If the waitqueue is being freed early but we can't complete
1854 * the request inline, we have to tear down the request as best
1855 * we can. That means immediately removing the request from its
1856 * waitqueue and preventing all further accesses to the
1857 * waitqueue via the request. We also need to schedule the
1858 * completion work (done above). Also mark the request as
1859 * cancelled, to potentially skip an unneeded call to ->poll().
1860 */
1861 if (mask & POLLFREE) {
1862 WRITE_ONCE(req->cancelled, true);
1863 list_del_init(&req->wait.entry);
1864
1865 /*
1866 * Careful: this *must* be the last step, since as soon
1867 * as req->head is NULL'ed out, the request can be
1868 * completed and freed, since aio_poll_complete_work()
1869 * will no longer need to take the waitqueue lock.
1870 */
1871 smp_store_release(&req->head, NULL);
1872 }
Christoph Hellwige8693bc2018-07-16 12:25:17 +02001873 }
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001874 return 1;
1875}
1876
1877struct aio_poll_table {
1878 struct poll_table_struct pt;
1879 struct aio_kiocb *iocb;
Eric Biggers50252e42021-12-08 17:04:55 -08001880 bool queued;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001881 int error;
1882};
1883
1884static void
1885aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1886 struct poll_table_struct *p)
1887{
1888 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1889
1890 /* multiple wait queues per file are not supported */
Eric Biggers50252e42021-12-08 17:04:55 -08001891 if (unlikely(pt->queued)) {
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001892 pt->error = -EINVAL;
1893 return;
1894 }
1895
Eric Biggers50252e42021-12-08 17:04:55 -08001896 pt->queued = true;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001897 pt->error = 0;
1898 pt->iocb->poll.head = head;
1899 add_wait_queue(head, &pt->iocb->poll.wait);
1900}
1901
Al Viro958c13c2019-03-06 18:13:00 -05001902static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001903{
1904 struct kioctx *ctx = aiocb->ki_ctx;
1905 struct poll_iocb *req = &aiocb->poll;
1906 struct aio_poll_table apt;
Al Viroaf5c72b2019-03-07 21:45:41 -05001907 bool cancel = false;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001908 __poll_t mask;
1909
1910 /* reject any unknown events outside the normal event mask. */
1911 if ((u16)iocb->aio_buf != iocb->aio_buf)
1912 return -EINVAL;
1913 /* reject fields that are not defined for poll */
1914 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1915 return -EINVAL;
1916
1917 INIT_WORK(&req->work, aio_poll_complete_work);
1918 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001919
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001920 req->head = NULL;
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001921 req->cancelled = false;
Eric Biggers363bee22021-12-08 17:04:54 -08001922 req->work_scheduled = false;
1923 req->work_need_resched = false;
Jens Axboe2bc4ca92018-12-04 09:44:49 -07001924
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001925 apt.pt._qproc = aio_poll_queue_proc;
1926 apt.pt._key = req->events;
1927 apt.iocb = aiocb;
Eric Biggers50252e42021-12-08 17:04:55 -08001928 apt.queued = false;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001929 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1930
1931 /* initialized the list so that we can do list_empty checks */
1932 INIT_LIST_HEAD(&req->wait.entry);
1933 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1934
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001935 mask = vfs_poll(req->file, &apt.pt) & req->events;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001936 spin_lock_irq(&ctx->ctx_lock);
Eric Biggers50252e42021-12-08 17:04:55 -08001937 if (likely(apt.queued)) {
1938 bool on_queue = poll_iocb_lock_wq(req);
1939
1940 if (!on_queue || req->work_scheduled) {
Eric Biggers363bee22021-12-08 17:04:54 -08001941 /*
1942 * aio_poll_wake() already either scheduled the async
1943 * completion work, or completed the request inline.
1944 */
1945 if (apt.error) /* unsupported case: multiple queues */
Al Viroaf5c72b2019-03-07 21:45:41 -05001946 cancel = true;
1947 apt.error = 0;
1948 mask = 0;
1949 }
1950 if (mask || apt.error) {
Eric Biggers363bee22021-12-08 17:04:54 -08001951 /* Steal to complete synchronously. */
Al Viroaf5c72b2019-03-07 21:45:41 -05001952 list_del_init(&req->wait.entry);
1953 } else if (cancel) {
Eric Biggers363bee22021-12-08 17:04:54 -08001954 /* Cancel if possible (may be too late though). */
Al Viroaf5c72b2019-03-07 21:45:41 -05001955 WRITE_ONCE(req->cancelled, true);
Eric Biggers50252e42021-12-08 17:04:55 -08001956 } else if (on_queue) {
Eric Biggers363bee22021-12-08 17:04:54 -08001957 /*
1958 * Actually waiting for an event, so add the request to
1959 * active_reqs so that it can be cancelled if needed.
1960 */
Al Viroaf5c72b2019-03-07 21:45:41 -05001961 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1962 aiocb->ki_cancel = aio_poll_cancel;
1963 }
Eric Biggers50252e42021-12-08 17:04:55 -08001964 if (on_queue)
1965 poll_iocb_unlock_wq(req);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001966 }
Al Viroaf5c72b2019-03-07 21:45:41 -05001967 if (mask) { /* no async, we'd stolen it */
1968 aiocb->ki_res.res = mangle_poll(mask);
1969 apt.error = 0;
1970 }
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001971 spin_unlock_irq(&ctx->ctx_lock);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001972 if (mask)
Al Viroaf5c72b2019-03-07 21:45:41 -05001973 iocb_put(aiocb);
1974 return apt.error;
Christoph Hellwigbfe40372018-07-16 09:08:20 +02001975}
1976
Jens Axboe88a6f182018-11-24 14:46:14 -07001977static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
Al Viro7316b492019-03-06 18:24:51 -05001978 struct iocb __user *user_iocb, struct aio_kiocb *req,
1979 bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980{
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001981 req->ki_filp = fget(iocb->aio_fildes);
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001982 if (unlikely(!req->ki_filp))
Al Viro7316b492019-03-06 18:24:51 -05001983 return -EBADF;
Linus Torvalds84c4e1f82019-03-03 14:23:33 -08001984
Jens Axboe88a6f182018-11-24 14:46:14 -07001985 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
Al Viro74259702019-03-06 18:18:31 -05001986 struct eventfd_ctx *eventfd;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001987 /*
1988 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1989 * instance of the file* now. The file descriptor must be
1990 * an eventfd() fd, and will be signaled for each completed
1991 * event using the eventfd_signal() function.
1992 */
Al Viro74259702019-03-06 18:18:31 -05001993 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
Al Viro7316b492019-03-06 18:24:51 -05001994 if (IS_ERR(eventfd))
Dan Carpenter18bfb9c2019-04-03 09:22:35 +03001995 return PTR_ERR(eventfd);
Al Viro7316b492019-03-06 18:24:51 -05001996
Al Viro74259702019-03-06 18:18:31 -05001997 req->ki_eventfd = eventfd;
Goldwyn Rodrigues9830f4b2017-06-20 07:05:42 -05001998 }
1999
Al Viro7316b492019-03-06 18:24:51 -05002000 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07002001 pr_debug("EFAULT: aio_key\n");
Al Viro7316b492019-03-06 18:24:51 -05002002 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 }
2004
Al Viroa9339b72019-03-07 19:43:45 -05002005 req->ki_res.obj = (u64)(unsigned long)user_iocb;
2006 req->ki_res.data = iocb->aio_data;
2007 req->ki_res.res = 0;
2008 req->ki_res.res2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009
Jens Axboe88a6f182018-11-24 14:46:14 -07002010 switch (iocb->aio_lio_opcode) {
Christoph Hellwig89319d312016-10-30 11:42:03 -05002011 case IOCB_CMD_PREAD:
Al Viro7316b492019-03-06 18:24:51 -05002012 return aio_read(&req->rw, iocb, false, compat);
Christoph Hellwig89319d312016-10-30 11:42:03 -05002013 case IOCB_CMD_PWRITE:
Al Viro7316b492019-03-06 18:24:51 -05002014 return aio_write(&req->rw, iocb, false, compat);
Christoph Hellwig89319d312016-10-30 11:42:03 -05002015 case IOCB_CMD_PREADV:
Al Viro7316b492019-03-06 18:24:51 -05002016 return aio_read(&req->rw, iocb, true, compat);
Christoph Hellwig89319d312016-10-30 11:42:03 -05002017 case IOCB_CMD_PWRITEV:
Al Viro7316b492019-03-06 18:24:51 -05002018 return aio_write(&req->rw, iocb, true, compat);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02002019 case IOCB_CMD_FSYNC:
Al Viro7316b492019-03-06 18:24:51 -05002020 return aio_fsync(&req->fsync, iocb, false);
Christoph Hellwiga3c0d432018-03-27 19:18:57 +02002021 case IOCB_CMD_FDSYNC:
Al Viro7316b492019-03-06 18:24:51 -05002022 return aio_fsync(&req->fsync, iocb, true);
Christoph Hellwigbfe40372018-07-16 09:08:20 +02002023 case IOCB_CMD_POLL:
Al Viro7316b492019-03-06 18:24:51 -05002024 return aio_poll(req, iocb);
Christoph Hellwig89319d312016-10-30 11:42:03 -05002025 default:
Jens Axboe88a6f182018-11-24 14:46:14 -07002026 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
Al Viro7316b492019-03-06 18:24:51 -05002027 return -EINVAL;
Christoph Hellwig89319d312016-10-30 11:42:03 -05002028 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029}
2030
Jens Axboe88a6f182018-11-24 14:46:14 -07002031static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
2032 bool compat)
2033{
Al Viro7316b492019-03-06 18:24:51 -05002034 struct aio_kiocb *req;
Jens Axboe88a6f182018-11-24 14:46:14 -07002035 struct iocb iocb;
Al Viro7316b492019-03-06 18:24:51 -05002036 int err;
Jens Axboe88a6f182018-11-24 14:46:14 -07002037
2038 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
2039 return -EFAULT;
2040
Al Viro7316b492019-03-06 18:24:51 -05002041 /* enforce forwards compatibility on users */
2042 if (unlikely(iocb.aio_reserved2)) {
2043 pr_debug("EINVAL: reserve field set\n");
2044 return -EINVAL;
2045 }
2046
2047 /* prevent overflows */
2048 if (unlikely(
2049 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2050 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2051 ((ssize_t)iocb.aio_nbytes < 0)
2052 )) {
2053 pr_debug("EINVAL: overflow check\n");
2054 return -EINVAL;
2055 }
2056
2057 req = aio_get_req(ctx);
2058 if (unlikely(!req))
2059 return -EAGAIN;
2060
2061 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2062
2063 /* Done with the synchronous reference */
2064 iocb_put(req);
2065
2066 /*
2067 * If err is 0, we'd either done aio_complete() ourselves or have
2068 * arranged for that to be done asynchronously. Anything non-zero
2069 * means that we need to destroy req ourselves.
2070 */
2071 if (unlikely(err)) {
2072 iocb_destroy(req);
2073 put_reqs_available(ctx, 1);
2074 }
2075 return err;
Jens Axboe88a6f182018-11-24 14:46:14 -07002076}
2077
Jeff Moyer9d85cba72010-05-26 14:44:26 -07002078/* sys_io_submit:
2079 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2080 * the number of iocbs queued. May return -EINVAL if the aio_context
2081 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2082 * *iocbpp[0] is not properly initialized, if the operation specified
2083 * is invalid for the file descriptor in the iocb. May fail with
2084 * -EFAULT if any of the data structures point to invalid data. May
2085 * fail with -EBADF if the file descriptor specified in the first
2086 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2087 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2088 * fail with -ENOSYS if not implemented.
2089 */
2090SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2091 struct iocb __user * __user *, iocbpp)
2092{
Al Viro67ba0492018-05-26 20:10:07 -04002093 struct kioctx *ctx;
2094 long ret = 0;
2095 int i = 0;
2096 struct blk_plug plug;
2097
2098 if (unlikely(nr < 0))
2099 return -EINVAL;
2100
Al Viro67ba0492018-05-26 20:10:07 -04002101 ctx = lookup_ioctx(ctx_id);
2102 if (unlikely(!ctx)) {
2103 pr_debug("EINVAL: invalid context id\n");
2104 return -EINVAL;
2105 }
2106
Al Viro1da92772018-05-26 20:10:07 -04002107 if (nr > ctx->nr_events)
2108 nr = ctx->nr_events;
2109
Jens Axboea79d40e2018-12-04 09:45:32 -07002110 if (nr > AIO_PLUG_THRESHOLD)
2111 blk_start_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04002112 for (i = 0; i < nr; i++) {
2113 struct iocb __user *user_iocb;
2114
2115 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2116 ret = -EFAULT;
2117 break;
2118 }
2119
2120 ret = io_submit_one(ctx, user_iocb, false);
2121 if (ret)
2122 break;
2123 }
Jens Axboea79d40e2018-12-04 09:45:32 -07002124 if (nr > AIO_PLUG_THRESHOLD)
2125 blk_finish_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04002126
2127 percpu_ref_put(&ctx->users);
2128 return i ? i : ret;
Jeff Moyer9d85cba72010-05-26 14:44:26 -07002129}
2130
Al Viroc00d2c72016-12-20 07:04:57 -05002131#ifdef CONFIG_COMPAT
Al Viroc00d2c72016-12-20 07:04:57 -05002132COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
Al Viro67ba0492018-05-26 20:10:07 -04002133 int, nr, compat_uptr_t __user *, iocbpp)
Al Viroc00d2c72016-12-20 07:04:57 -05002134{
Al Viro67ba0492018-05-26 20:10:07 -04002135 struct kioctx *ctx;
2136 long ret = 0;
2137 int i = 0;
2138 struct blk_plug plug;
Al Viroc00d2c72016-12-20 07:04:57 -05002139
2140 if (unlikely(nr < 0))
2141 return -EINVAL;
2142
Al Viro67ba0492018-05-26 20:10:07 -04002143 ctx = lookup_ioctx(ctx_id);
2144 if (unlikely(!ctx)) {
2145 pr_debug("EINVAL: invalid context id\n");
2146 return -EINVAL;
2147 }
2148
Al Viro1da92772018-05-26 20:10:07 -04002149 if (nr > ctx->nr_events)
2150 nr = ctx->nr_events;
2151
Jens Axboea79d40e2018-12-04 09:45:32 -07002152 if (nr > AIO_PLUG_THRESHOLD)
2153 blk_start_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04002154 for (i = 0; i < nr; i++) {
2155 compat_uptr_t user_iocb;
2156
2157 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2158 ret = -EFAULT;
2159 break;
2160 }
2161
2162 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2163 if (ret)
2164 break;
2165 }
Jens Axboea79d40e2018-12-04 09:45:32 -07002166 if (nr > AIO_PLUG_THRESHOLD)
2167 blk_finish_plug(&plug);
Al Viro67ba0492018-05-26 20:10:07 -04002168
2169 percpu_ref_put(&ctx->users);
2170 return i ? i : ret;
Al Viroc00d2c72016-12-20 07:04:57 -05002171}
2172#endif
2173
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174/* sys_io_cancel:
Bart Van Assche28468cb2024-03-04 10:29:44 -08002175 * Attempts to cancel an iocb previously passed to io_submit. If
2176 * the operation is successfully cancelled, the resulting event is
2177 * copied into the memory pointed to by result without being placed
2178 * into the completion queue and 0 is returned. May fail with
2179 * -EFAULT if any of the data structures pointed to are invalid.
2180 * May fail with -EINVAL if aio_context specified by ctx_id is
2181 * invalid. May fail with -EAGAIN if the iocb specified was not
2182 * cancelled. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 */
Heiko Carstens002c8972009-01-14 14:14:18 +01002184SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2185 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187 struct kioctx *ctx;
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01002188 struct aio_kiocb *kiocb;
Christoph Hellwig888933f2018-05-23 14:11:02 +02002189 int ret = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190 u32 key;
Al Viroa9339b72019-03-07 19:43:45 -05002191 u64 obj = (u64)(unsigned long)iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192
Christoph Hellwigf3a27522018-03-30 11:19:25 +02002193 if (unlikely(get_user(key, &iocb->aio_key)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 return -EFAULT;
Christoph Hellwigf3a27522018-03-30 11:19:25 +02002195 if (unlikely(key != KIOCB_KEY))
2196 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197
2198 ctx = lookup_ioctx(ctx_id);
2199 if (unlikely(!ctx))
2200 return -EINVAL;
2201
2202 spin_lock_irq(&ctx->ctx_lock);
Al Viro833f4152019-03-11 19:00:36 -04002203 /* TODO: use a hash or array, this sucks. */
2204 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
Al Viroa9339b72019-03-07 19:43:45 -05002205 if (kiocb->ki_res.obj == obj) {
Al Viro833f4152019-03-11 19:00:36 -04002206 ret = kiocb->ki_cancel(&kiocb->rw);
2207 list_del_init(&kiocb->ki_list);
2208 break;
2209 }
Christoph Hellwig888933f2018-05-23 14:11:02 +02002210 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 spin_unlock_irq(&ctx->ctx_lock);
2212
Bart Van Assche28468cb2024-03-04 10:29:44 -08002213 if (!ret) {
2214 /*
2215 * The result argument is no longer used - the io_event is
2216 * always delivered via the ring buffer. -EINPROGRESS indicates
2217 * cancellation is progress:
2218 */
2219 ret = -EINPROGRESS;
2220 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221
Kent Overstreet723be6e2013-05-28 15:14:48 -07002222 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223
2224 return ret;
2225}
2226
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002227static long do_io_getevents(aio_context_t ctx_id,
2228 long min_nr,
2229 long nr,
2230 struct io_event __user *events,
2231 struct timespec64 *ts)
2232{
2233 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2234 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2235 long ret = -EINVAL;
2236
2237 if (likely(ioctx)) {
2238 if (likely(min_nr <= nr && min_nr >= 0))
2239 ret = read_events(ioctx, min_nr, nr, events, until);
2240 percpu_ref_put(&ioctx->users);
2241 }
2242
2243 return ret;
2244}
2245
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246/* io_getevents:
2247 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07002248 * the completion queue for the aio_context specified by ctx_id. If
2249 * it succeeds, the number of read events is returned. May fail with
2250 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2251 * out of range, if timeout is out of range. May fail with -EFAULT
2252 * if any of the memory specified is invalid. May return 0 or
2253 * < min_nr if the timeout specified by timeout has elapsed
2254 * before sufficient events are available, where timeout == NULL
2255 * specifies an infinite timeout. Note that the timeout pointed to by
Jeff Moyer69008072013-05-24 15:55:24 -07002256 * timeout is relative. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 */
Arnd Bergmann3ca47e92019-04-23 17:43:50 +02002258#ifdef CONFIG_64BIT
Deepa Dinamani7a353972018-09-19 21:41:08 -07002259
Heiko Carstens002c8972009-01-14 14:14:18 +01002260SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2261 long, min_nr,
2262 long, nr,
2263 struct io_event __user *, events,
Deepa Dinamani7a353972018-09-19 21:41:08 -07002264 struct __kernel_timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265{
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002266 struct timespec64 ts;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002267 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002269 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2270 return -EFAULT;
2271
2272 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2273 if (!ret && signal_pending(current))
2274 ret = -EINTR;
2275 return ret;
2276}
2277
Deepa Dinamani7a353972018-09-19 21:41:08 -07002278#endif
2279
Christoph Hellwig9ba546c2018-07-11 15:48:46 +02002280struct __aio_sigset {
2281 const sigset_t __user *sigmask;
2282 size_t sigsetsize;
2283};
2284
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002285SYSCALL_DEFINE6(io_pgetevents,
2286 aio_context_t, ctx_id,
2287 long, min_nr,
2288 long, nr,
2289 struct io_event __user *, events,
Deepa Dinamani7a353972018-09-19 21:41:08 -07002290 struct __kernel_timespec __user *, timeout,
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002291 const struct __aio_sigset __user *, usig)
2292{
2293 struct __aio_sigset ksig = { NULL, };
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002294 struct timespec64 ts;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002295 bool interrupted;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002296 int ret;
2297
2298 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2299 return -EFAULT;
2300
2301 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2302 return -EFAULT;
2303
Oleg Nesterovb7724342019-07-16 16:29:53 -07002304 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
Deepa Dinamani7a353972018-09-19 21:41:08 -07002305 if (ret)
2306 return ret;
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002307
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002308 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002309
2310 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002311 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002312 if (interrupted && !ret)
Deepa Dinamani7a353972018-09-19 21:41:08 -07002313 ret = -ERESTARTNOHAND;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002314
2315 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316}
Al Viroc00d2c72016-12-20 07:04:57 -05002317
Deepa Dinamani7a353972018-09-19 21:41:08 -07002318#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2319
2320SYSCALL_DEFINE6(io_pgetevents_time32,
2321 aio_context_t, ctx_id,
2322 long, min_nr,
2323 long, nr,
2324 struct io_event __user *, events,
2325 struct old_timespec32 __user *, timeout,
2326 const struct __aio_sigset __user *, usig)
2327{
2328 struct __aio_sigset ksig = { NULL, };
Deepa Dinamani7a353972018-09-19 21:41:08 -07002329 struct timespec64 ts;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002330 bool interrupted;
Deepa Dinamani7a353972018-09-19 21:41:08 -07002331 int ret;
2332
2333 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2334 return -EFAULT;
2335
2336 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2337 return -EFAULT;
2338
Deepa Dinamanided653c2018-09-19 21:41:04 -07002339
Oleg Nesterovb7724342019-07-16 16:29:53 -07002340 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
Deepa Dinamanided653c2018-09-19 21:41:04 -07002341 if (ret)
2342 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343
2344 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002345
2346 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002347 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002348 if (interrupted && !ret)
Deepa Dinamani854a6ed2018-09-19 21:41:05 -07002349 ret = -ERESTARTNOHAND;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350
2351 return ret;
2352}
Al Viroc00d2c72016-12-20 07:04:57 -05002353
Deepa Dinamani7a353972018-09-19 21:41:08 -07002354#endif
2355
2356#if defined(CONFIG_COMPAT_32BIT_TIME)
2357
Arnd Bergmann8dabe722019-01-07 00:33:08 +01002358SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2359 __s32, min_nr,
2360 __s32, nr,
2361 struct io_event __user *, events,
2362 struct old_timespec32 __user *, timeout)
Al Viroc00d2c72016-12-20 07:04:57 -05002363{
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002364 struct timespec64 t;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002365 int ret;
Al Viroc00d2c72016-12-20 07:04:57 -05002366
Arnd Bergmann9afc5ee2018-07-13 12:52:28 +02002367 if (timeout && get_old_timespec32(&t, timeout))
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002368 return -EFAULT;
2369
2370 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2371 if (!ret && signal_pending(current))
2372 ret = -EINTR;
2373 return ret;
2374}
2375
Deepa Dinamani7a353972018-09-19 21:41:08 -07002376#endif
2377
2378#ifdef CONFIG_COMPAT
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002379
2380struct __compat_aio_sigset {
Guillem Jover97eba802019-08-21 05:38:20 +02002381 compat_uptr_t sigmask;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002382 compat_size_t sigsetsize;
2383};
2384
Deepa Dinamani7a353972018-09-19 21:41:08 -07002385#if defined(CONFIG_COMPAT_32BIT_TIME)
2386
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002387COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2388 compat_aio_context_t, ctx_id,
2389 compat_long_t, min_nr,
2390 compat_long_t, nr,
2391 struct io_event __user *, events,
Arnd Bergmann9afc5ee2018-07-13 12:52:28 +02002392 struct old_timespec32 __user *, timeout,
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002393 const struct __compat_aio_sigset __user *, usig)
2394{
Guillem Jover97eba802019-08-21 05:38:20 +02002395 struct __compat_aio_sigset ksig = { 0, };
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002396 struct timespec64 t;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002397 bool interrupted;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002398 int ret;
2399
Arnd Bergmann9afc5ee2018-07-13 12:52:28 +02002400 if (timeout && get_old_timespec32(&t, timeout))
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002401 return -EFAULT;
2402
2403 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2404 return -EFAULT;
2405
Guillem Jover97eba802019-08-21 05:38:20 +02002406 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
Deepa Dinamanided653c2018-09-19 21:41:04 -07002407 if (ret)
2408 return ret;
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07002409
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002410 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002411
2412 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002413 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002414 if (interrupted && !ret)
Deepa Dinamani854a6ed2018-09-19 21:41:05 -07002415 ret = -ERESTARTNOHAND;
Christoph Hellwig7a074e92018-05-02 19:51:00 +02002416
2417 return ret;
Al Viroc00d2c72016-12-20 07:04:57 -05002418}
Deepa Dinamani7a353972018-09-19 21:41:08 -07002419
2420#endif
2421
2422COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2423 compat_aio_context_t, ctx_id,
2424 compat_long_t, min_nr,
2425 compat_long_t, nr,
2426 struct io_event __user *, events,
2427 struct __kernel_timespec __user *, timeout,
2428 const struct __compat_aio_sigset __user *, usig)
2429{
Guillem Jover97eba802019-08-21 05:38:20 +02002430 struct __compat_aio_sigset ksig = { 0, };
Deepa Dinamani7a353972018-09-19 21:41:08 -07002431 struct timespec64 t;
Oleg Nesterov97abc882019-06-28 12:06:50 -07002432 bool interrupted;
Deepa Dinamani7a353972018-09-19 21:41:08 -07002433 int ret;
2434
2435 if (timeout && get_timespec64(&t, timeout))
2436 return -EFAULT;
2437
2438 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2439 return -EFAULT;
2440
Guillem Jover97eba802019-08-21 05:38:20 +02002441 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
Deepa Dinamani7a353972018-09-19 21:41:08 -07002442 if (ret)
2443 return ret;
2444
2445 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002446
2447 interrupted = signal_pending(current);
Oleg Nesterovb7724342019-07-16 16:29:53 -07002448 restore_saved_sigmask_unless(interrupted);
Oleg Nesterov97abc882019-06-28 12:06:50 -07002449 if (interrupted && !ret)
Deepa Dinamani7a353972018-09-19 21:41:08 -07002450 ret = -ERESTARTNOHAND;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451
2452 return ret;
2453}
2454#endif