blob: a062d75109cb380aeb819fe42d43644c8be33a91 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 *
5 * Implements an efficient asynchronous io interface.
6 *
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 *
9 * See ../COPYING for licensing terms.
10 */
Kent Overstreetcaf41672013-05-07 16:18:35 -070011#define pr_fmt(fmt) "%s: " fmt, __func__
12
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/kernel.h>
14#include <linux/init.h>
15#include <linux/errno.h>
16#include <linux/time.h>
17#include <linux/aio_abi.h>
Paul Gortmaker630d9c42011-11-16 23:57:37 -050018#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/syscalls.h>
Jens Axboeb9d128f2009-10-29 13:59:26 +010020#include <linux/backing-dev.h>
Badari Pulavarty027445c2006-09-30 23:28:46 -070021#include <linux/uio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070022
Ingo Molnar174cd4b2017-02-02 19:15:33 +010023#include <linux/sched/signal.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/fs.h>
25#include <linux/file.h>
26#include <linux/mm.h>
27#include <linux/mman.h>
Michael S. Tsirkin3d2d8272009-09-21 17:03:51 -070028#include <linux/mmu_context.h>
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100029#include <linux/percpu.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#include <linux/slab.h>
31#include <linux/timer.h>
32#include <linux/aio.h>
33#include <linux/highmem.h>
34#include <linux/workqueue.h>
35#include <linux/security.h>
Davide Libenzi9c3060b2007-05-10 22:23:21 -070036#include <linux/eventfd.h>
Jeff Moyercfb1e332009-10-02 18:57:36 -040037#include <linux/blkdev.h>
Jeff Moyer9d85cba2010-05-26 14:44:26 -070038#include <linux/compat.h>
Gu Zheng36bc08c2013-07-16 17:56:16 +080039#include <linux/migrate.h>
40#include <linux/ramfs.h>
Kent Overstreet723be6e2013-05-28 15:14:48 -070041#include <linux/percpu-refcount.h>
Benjamin LaHaise71ad7492013-09-17 10:18:25 -040042#include <linux/mount.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44#include <asm/kmap_types.h>
Linus Torvalds7c0f6ba2016-12-24 11:46:01 -080045#include <linux/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
Al Viro68d70d02013-06-19 15:26:04 +040047#include "internal.h"
48
Kent Overstreet4e179bc2013-05-07 16:18:33 -070049#define AIO_RING_MAGIC 0xa10a10a1
50#define AIO_RING_COMPAT_FEATURES 1
51#define AIO_RING_INCOMPAT_FEATURES 0
52struct aio_ring {
53 unsigned id; /* kernel internal index number */
54 unsigned nr; /* number of io_events */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -040055 unsigned head; /* Written to by userland or under ring_lock
56 * mutex by aio_read_events_ring(). */
Kent Overstreet4e179bc2013-05-07 16:18:33 -070057 unsigned tail;
58
59 unsigned magic;
60 unsigned compat_features;
61 unsigned incompat_features;
62 unsigned header_length; /* size of aio_ring */
63
64
65 struct io_event io_events[0];
66}; /* 128 bytes + ring size */
67
68#define AIO_RING_PAGES 8
Kent Overstreet4e179bc2013-05-07 16:18:33 -070069
Benjamin LaHaisedb446a02013-07-30 12:54:40 -040070struct kioctx_table {
71 struct rcu_head rcu;
72 unsigned nr;
73 struct kioctx *table[];
74};
75
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100076struct kioctx_cpu {
77 unsigned reqs_available;
78};
79
Jens Axboedc48e562015-04-15 11:17:23 -060080struct ctx_rq_wait {
81 struct completion comp;
82 atomic_t count;
83};
84
Kent Overstreet4e179bc2013-05-07 16:18:33 -070085struct kioctx {
Kent Overstreet723be6e2013-05-28 15:14:48 -070086 struct percpu_ref users;
Kent Overstreet36f55882013-05-07 16:18:41 -070087 atomic_t dead;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070088
Kent Overstreete34ecee2013-10-10 19:31:47 -070089 struct percpu_ref reqs;
90
Kent Overstreet4e179bc2013-05-07 16:18:33 -070091 unsigned long user_id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070092
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100093 struct __percpu kioctx_cpu *cpu;
94
95 /*
96 * For percpu reqs_available, number of slots we move to/from global
97 * counter at a time:
98 */
99 unsigned req_batch;
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700100 /*
101 * This is what userspace passed to io_setup(), it's not used for
102 * anything but counting against the global max_reqs quota.
103 *
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700104 * The real limit is nr_events - 1, which will be larger (see
Kent Overstreet3e845ce2013-05-07 16:18:51 -0700105 * aio_setup_ring())
106 */
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700107 unsigned max_reqs;
108
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700109 /* Size of ringbuffer, in units of struct io_event */
110 unsigned nr_events;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700111
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700112 unsigned long mmap_base;
113 unsigned long mmap_size;
114
115 struct page **ring_pages;
116 long nr_pages;
117
Kent Overstreet723be6e2013-05-28 15:14:48 -0700118 struct work_struct free_work;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700119
Anatol Pomozove02ba722014-04-15 11:31:33 -0700120 /*
121 * signals when all in-flight requests are done
122 */
Jens Axboedc48e562015-04-15 11:17:23 -0600123 struct ctx_rq_wait *rq_wait;
Anatol Pomozove02ba722014-04-15 11:31:33 -0700124
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700125 struct {
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000126 /*
127 * This counts the number of available slots in the ringbuffer,
128 * so we avoid overflowing it: it's decremented (if positive)
129 * when allocating a kiocb and incremented when the resulting
130 * io_event is pulled off the ringbuffer.
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000131 *
132 * We batch accesses to it with a percpu version.
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000133 */
134 atomic_t reqs_available;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700135 } ____cacheline_aligned_in_smp;
136
137 struct {
138 spinlock_t ctx_lock;
139 struct list_head active_reqs; /* used for cancellation */
140 } ____cacheline_aligned_in_smp;
141
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700142 struct {
143 struct mutex ring_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700144 wait_queue_head_t wait;
145 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700146
147 struct {
148 unsigned tail;
Benjamin LaHaised856f322014-08-24 13:14:05 -0400149 unsigned completed_events;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700150 spinlock_t completion_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700151 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700152
153 struct page *internal_pages[AIO_RING_PAGES];
Gu Zheng36bc08c2013-07-16 17:56:16 +0800154 struct file *aio_ring_file;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400155
156 unsigned id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700157};
158
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100159/*
160 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
161 * cancelled or completed (this makes a certain amount of sense because
162 * successful cancellation - io_cancel() - does deliver the completion to
163 * userspace).
164 *
165 * And since most things don't implement kiocb cancellation and we'd really like
166 * kiocb completion to be lockless when possible, we use ki_cancel to
167 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
168 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
169 */
170#define KIOCB_CANCELLED ((void *) (~0ULL))
171
172struct aio_kiocb {
173 struct kiocb common;
174
175 struct kioctx *ki_ctx;
176 kiocb_cancel_fn *ki_cancel;
177
178 struct iocb __user *ki_user_iocb; /* user's aiocb */
179 __u64 ki_user_data; /* user's data for completion */
180
181 struct list_head ki_list; /* the aio core uses this
182 * for cancellation */
183
184 /*
185 * If the aio_resfd field of the userspace iocb is not zero,
186 * this is the underlying eventfd context to deliver events to.
187 */
188 struct eventfd_ctx *ki_eventfd;
189};
190
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -0800192static DEFINE_SPINLOCK(aio_nr_lock);
193unsigned long aio_nr; /* current system wide number of aio requests */
194unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195/*----end sysctl variables---*/
196
Christoph Lametere18b8902006-12-06 20:33:20 -0800197static struct kmem_cache *kiocb_cachep;
198static struct kmem_cache *kioctx_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400200static struct vfsmount *aio_mnt;
201
202static const struct file_operations aio_ring_fops;
203static const struct address_space_operations aio_ctx_aops;
204
205static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
206{
207 struct qstr this = QSTR_INIT("[aio]", 5);
208 struct file *file;
209 struct path path;
210 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
Dan Carpenter7f626562013-11-13 10:49:40 +0300211 if (IS_ERR(inode))
212 return ERR_CAST(inode);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400213
214 inode->i_mapping->a_ops = &aio_ctx_aops;
215 inode->i_mapping->private_data = ctx;
216 inode->i_size = PAGE_SIZE * nr_pages;
217
218 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
219 if (!path.dentry) {
220 iput(inode);
221 return ERR_PTR(-ENOMEM);
222 }
223 path.mnt = mntget(aio_mnt);
224
225 d_instantiate(path.dentry, inode);
226 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
227 if (IS_ERR(file)) {
228 path_put(&path);
229 return file;
230 }
231
232 file->f_flags = O_RDWR;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400233 return file;
234}
235
236static struct dentry *aio_mount(struct file_system_type *fs_type,
237 int flags, const char *dev_name, void *data)
238{
239 static const struct dentry_operations ops = {
240 .d_dname = simple_dname,
241 };
Jann Horn22f6b4d2016-09-16 00:31:22 +0200242 struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops,
243 AIO_RING_MAGIC);
244
245 if (!IS_ERR(root))
246 root->d_sb->s_iflags |= SB_I_NOEXEC;
247 return root;
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400248}
249
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250/* aio_setup
251 * Creates the slab caches used by the aio routines, panic on
252 * failure as this is done early during the boot sequence.
253 */
254static int __init aio_setup(void)
255{
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400256 static struct file_system_type aio_fs = {
257 .name = "aio",
258 .mount = aio_mount,
259 .kill_sb = kill_anon_super,
260 };
261 aio_mnt = kern_mount(&aio_fs);
262 if (IS_ERR(aio_mnt))
263 panic("Failed to create aio fs mount.");
264
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100265 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Christoph Lameter0a31bd52007-05-06 14:49:57 -0700266 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267
Kent Overstreetcaf41672013-05-07 16:18:35 -0700268 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
270 return 0;
271}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700272__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400274static void put_aio_ring_file(struct kioctx *ctx)
275{
276 struct file *aio_ring_file = ctx->aio_ring_file;
Rasmus Villemoesde04e762016-09-15 00:25:03 +0200277 struct address_space *i_mapping;
278
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400279 if (aio_ring_file) {
Al Viro45063092016-12-04 18:24:56 -0500280 truncate_setsize(file_inode(aio_ring_file), 0);
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400281
282 /* Prevent further access to the kioctx from migratepages */
Al Viro45063092016-12-04 18:24:56 -0500283 i_mapping = aio_ring_file->f_mapping;
Rasmus Villemoesde04e762016-09-15 00:25:03 +0200284 spin_lock(&i_mapping->private_lock);
285 i_mapping->private_data = NULL;
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400286 ctx->aio_ring_file = NULL;
Rasmus Villemoesde04e762016-09-15 00:25:03 +0200287 spin_unlock(&i_mapping->private_lock);
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400288
289 fput(aio_ring_file);
290 }
291}
292
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293static void aio_free_ring(struct kioctx *ctx)
294{
Gu Zheng36bc08c2013-07-16 17:56:16 +0800295 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400297 /* Disconnect the kiotx from the ring file. This prevents future
298 * accesses to the kioctx from page migration.
299 */
300 put_aio_ring_file(ctx);
301
Gu Zheng36bc08c2013-07-16 17:56:16 +0800302 for (i = 0; i < ctx->nr_pages; i++) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500303 struct page *page;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800304 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
305 page_count(ctx->ring_pages[i]));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500306 page = ctx->ring_pages[i];
307 if (!page)
308 continue;
309 ctx->ring_pages[i] = NULL;
310 put_page(page);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800311 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312
Sasha Levinddb8c452013-11-19 17:33:03 -0500313 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700314 kfree(ctx->ring_pages);
Sasha Levinddb8c452013-11-19 17:33:03 -0500315 ctx->ring_pages = NULL;
316 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317}
318
Oleg Nesterov5477e702015-09-04 15:48:04 -0700319static int aio_ring_mremap(struct vm_area_struct *vma)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800320{
Oleg Nesterov5477e702015-09-04 15:48:04 -0700321 struct file *file = vma->vm_file;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400322 struct mm_struct *mm = vma->vm_mm;
323 struct kioctx_table *table;
Al Virob2edffd2015-04-06 17:48:54 -0400324 int i, res = -EINVAL;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400325
326 spin_lock(&mm->ioctx_lock);
327 rcu_read_lock();
328 table = rcu_dereference(mm->ioctx_table);
329 for (i = 0; i < table->nr; i++) {
330 struct kioctx *ctx;
331
332 ctx = table->table[i];
333 if (ctx && ctx->aio_ring_file == file) {
Al Virob2edffd2015-04-06 17:48:54 -0400334 if (!atomic_read(&ctx->dead)) {
335 ctx->user_id = ctx->mmap_base = vma->vm_start;
336 res = 0;
337 }
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400338 break;
339 }
340 }
341
342 rcu_read_unlock();
343 spin_unlock(&mm->ioctx_lock);
Al Virob2edffd2015-04-06 17:48:54 -0400344 return res;
Pavel Emelyanove4a0d3e2014-09-18 19:56:17 +0400345}
346
Oleg Nesterov5477e702015-09-04 15:48:04 -0700347static const struct vm_operations_struct aio_ring_vm_ops = {
348 .mremap = aio_ring_mremap,
349#if IS_ENABLED(CONFIG_MMU)
350 .fault = filemap_fault,
351 .map_pages = filemap_map_pages,
352 .page_mkwrite = filemap_page_mkwrite,
353#endif
354};
355
356static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
357{
358 vma->vm_flags |= VM_DONTEXPAND;
359 vma->vm_ops = &aio_ring_vm_ops;
360 return 0;
361}
362
Gu Zheng36bc08c2013-07-16 17:56:16 +0800363static const struct file_operations aio_ring_fops = {
364 .mmap = aio_ring_mmap,
365};
366
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400367#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800368static int aio_migratepage(struct address_space *mapping, struct page *new,
369 struct page *old, enum migrate_mode mode)
370{
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400371 struct kioctx *ctx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800372 unsigned long flags;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400373 pgoff_t idx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800374 int rc;
375
Jérôme Glisse2916ecc2017-09-08 16:12:06 -0700376 /*
377 * We cannot support the _NO_COPY case here, because copy needs to
378 * happen under the ctx->completion_lock. That does not work with the
379 * migration workflow of MIGRATE_SYNC_NO_COPY.
380 */
381 if (mode == MIGRATE_SYNC_NO_COPY)
382 return -EINVAL;
383
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500384 rc = 0;
385
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400386 /* mapping->private_lock here protects against the kioctx teardown. */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500387 spin_lock(&mapping->private_lock);
388 ctx = mapping->private_data;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400389 if (!ctx) {
390 rc = -EINVAL;
391 goto out;
392 }
393
394 /* The ring_lock mutex. The prevents aio_read_events() from writing
395 * to the ring's head, and prevents page migration from mucking in
396 * a partially initialized kiotx.
397 */
398 if (!mutex_trylock(&ctx->ring_lock)) {
399 rc = -EAGAIN;
400 goto out;
401 }
402
403 idx = old->index;
404 if (idx < (pgoff_t)ctx->nr_pages) {
405 /* Make sure the old page hasn't already been changed */
406 if (ctx->ring_pages[idx] != old)
407 rc = -EAGAIN;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500408 } else
409 rc = -EINVAL;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500410
411 if (rc != 0)
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400412 goto out_unlock;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500413
Gu Zheng36bc08c2013-07-16 17:56:16 +0800414 /* Writeback must be complete */
415 BUG_ON(PageWriteback(old));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500416 get_page(new);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800417
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500418 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800419 if (rc != MIGRATEPAGE_SUCCESS) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500420 put_page(new);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400421 goto out_unlock;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800422 }
423
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400424 /* Take completion_lock to prevent other writes to the ring buffer
425 * while the old page is copied to the new. This prevents new
426 * events from being lost.
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400427 */
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400428 spin_lock_irqsave(&ctx->completion_lock, flags);
429 migrate_page_copy(new, old);
430 BUG_ON(ctx->ring_pages[idx] != old);
431 ctx->ring_pages[idx] = new;
432 spin_unlock_irqrestore(&ctx->completion_lock, flags);
433
434 /* The old page is no longer accessible. */
435 put_page(old);
436
437out_unlock:
438 mutex_unlock(&ctx->ring_lock);
439out:
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400440 spin_unlock(&mapping->private_lock);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800441 return rc;
442}
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400443#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800444
445static const struct address_space_operations aio_ctx_aops = {
Gu Zheng835f2522014-11-06 17:46:21 +0800446 .set_page_dirty = __set_page_dirty_no_writeback,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400447#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800448 .migratepage = aio_migratepage,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400449#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800450};
451
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300452static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453{
454 struct aio_ring *ring;
Zach Brown41003a72013-05-07 16:18:25 -0700455 struct mm_struct *mm = current->mm;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900456 unsigned long size, unused;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 int nr_pages;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800458 int i;
459 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460
461 /* Compensate for the ring buffer's head/tail overlap entry */
462 nr_events += 2; /* 1 is required, 2 for good luck */
463
464 size = sizeof(struct aio_ring);
465 size += sizeof(struct io_event) * nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466
Gu Zheng36bc08c2013-07-16 17:56:16 +0800467 nr_pages = PFN_UP(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 if (nr_pages < 0)
469 return -EINVAL;
470
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400471 file = aio_private_file(ctx, nr_pages);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800472 if (IS_ERR(file)) {
473 ctx->aio_ring_file = NULL;
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400474 return -ENOMEM;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800475 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476
Gu Zheng36bc08c2013-07-16 17:56:16 +0800477 ctx->aio_ring_file = file;
478 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
479 / sizeof(struct io_event);
480
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700481 ctx->ring_pages = ctx->internal_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 if (nr_pages > AIO_RING_PAGES) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700483 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
484 GFP_KERNEL);
Gu Zhengd1b94322013-12-04 18:19:06 +0800485 if (!ctx->ring_pages) {
486 put_aio_ring_file(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 return -ENOMEM;
Gu Zhengd1b94322013-12-04 18:19:06 +0800488 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 }
490
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900491 for (i = 0; i < nr_pages; i++) {
492 struct page *page;
Al Viro45063092016-12-04 18:24:56 -0500493 page = find_or_create_page(file->f_mapping,
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900494 i, GFP_HIGHUSER | __GFP_ZERO);
495 if (!page)
496 break;
497 pr_debug("pid(%d) page[%d]->count=%d\n",
498 current->pid, i, page_count(page));
499 SetPageUptodate(page);
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900500 unlock_page(page);
501
502 ctx->ring_pages[i] = page;
503 }
504 ctx->nr_pages = i;
505
506 if (unlikely(i != nr_pages)) {
507 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400508 return -ENOMEM;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900509 }
510
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700511 ctx->mmap_size = nr_pages * PAGE_SIZE;
512 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800513
Michal Hocko013373e2016-05-23 16:25:59 -0700514 if (down_write_killable(&mm->mmap_sem)) {
515 ctx->mmap_size = 0;
516 aio_free_ring(ctx);
517 return -EINTR;
518 }
519
Gu Zheng36bc08c2013-07-16 17:56:16 +0800520 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
521 PROT_READ | PROT_WRITE,
Mike Rapoport897ab3e2017-02-24 14:58:22 -0800522 MAP_SHARED, 0, &unused, NULL);
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900523 up_write(&mm->mmap_sem);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700524 if (IS_ERR((void *)ctx->mmap_base)) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700525 ctx->mmap_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 aio_free_ring(ctx);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400527 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 }
529
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700530 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400531
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700532 ctx->user_id = ctx->mmap_base;
533 ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700535 ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 ring->nr = nr_events; /* user copy */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400537 ring->id = ~0U;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 ring->head = ring->tail = 0;
539 ring->magic = AIO_RING_MAGIC;
540 ring->compat_features = AIO_RING_COMPAT_FEATURES;
541 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
542 ring->header_length = sizeof(struct aio_ring);
Cong Wange8e3c3d2011-11-25 23:14:27 +0800543 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700544 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545
546 return 0;
547}
548
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
550#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
551#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
552
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100553void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
Kent Overstreet0460fef2013-05-07 16:18:49 -0700554{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100555 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
Kent Overstreet0460fef2013-05-07 16:18:49 -0700556 struct kioctx *ctx = req->ki_ctx;
557 unsigned long flags;
558
559 spin_lock_irqsave(&ctx->ctx_lock, flags);
560
561 if (!req->ki_list.next)
562 list_add(&req->ki_list, &ctx->active_reqs);
563
564 req->ki_cancel = cancel;
565
566 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
567}
568EXPORT_SYMBOL(kiocb_set_cancel_fn);
569
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100570static int kiocb_cancel(struct aio_kiocb *kiocb)
Kent Overstreet906b9732013-05-07 16:18:31 -0700571{
Kent Overstreet0460fef2013-05-07 16:18:49 -0700572 kiocb_cancel_fn *old, *cancel;
Kent Overstreet906b9732013-05-07 16:18:31 -0700573
Kent Overstreet0460fef2013-05-07 16:18:49 -0700574 /*
575 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
576 * actually has a cancel function, hence the cmpxchg()
577 */
Kent Overstreet906b9732013-05-07 16:18:31 -0700578
Mark Rutland6aa7de02017-10-23 14:07:29 -0700579 cancel = READ_ONCE(kiocb->ki_cancel);
Kent Overstreet0460fef2013-05-07 16:18:49 -0700580 do {
581 if (!cancel || cancel == KIOCB_CANCELLED)
Kent Overstreet57282d82013-05-13 13:42:52 -0700582 return -EINVAL;
Kent Overstreet906b9732013-05-07 16:18:31 -0700583
Kent Overstreet0460fef2013-05-07 16:18:49 -0700584 old = cancel;
585 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
586 } while (cancel != old);
587
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100588 return cancel(&kiocb->common);
Kent Overstreet906b9732013-05-07 16:18:31 -0700589}
590
Kent Overstreete34ecee2013-10-10 19:31:47 -0700591static void free_ioctx(struct work_struct *work)
Kent Overstreet36f55882013-05-07 16:18:41 -0700592{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700593 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000594
Kent Overstreete34ecee2013-10-10 19:31:47 -0700595 pr_debug("freeing %p\n", ctx);
596
597 aio_free_ring(ctx);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000598 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400599 percpu_ref_exit(&ctx->reqs);
600 percpu_ref_exit(&ctx->users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700601 kmem_cache_free(kioctx_cachep, ctx);
602}
603
Kent Overstreete34ecee2013-10-10 19:31:47 -0700604static void free_ioctx_reqs(struct percpu_ref *ref)
605{
606 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
607
Anatol Pomozove02ba722014-04-15 11:31:33 -0700608 /* At this point we know that there are no any in-flight requests */
Jens Axboedc48e562015-04-15 11:17:23 -0600609 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
610 complete(&ctx->rq_wait->comp);
Anatol Pomozove02ba722014-04-15 11:31:33 -0700611
Kent Overstreete34ecee2013-10-10 19:31:47 -0700612 INIT_WORK(&ctx->free_work, free_ioctx);
613 schedule_work(&ctx->free_work);
614}
615
Kent Overstreet36f55882013-05-07 16:18:41 -0700616/*
617 * When this function runs, the kioctx has been removed from the "hash table"
618 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
619 * now it's safe to cancel any that need to be.
620 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700621static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet36f55882013-05-07 16:18:41 -0700622{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700623 struct kioctx *ctx = container_of(ref, struct kioctx, users);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100624 struct aio_kiocb *req;
Kent Overstreet36f55882013-05-07 16:18:41 -0700625
626 spin_lock_irq(&ctx->ctx_lock);
627
628 while (!list_empty(&ctx->active_reqs)) {
629 req = list_first_entry(&ctx->active_reqs,
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100630 struct aio_kiocb, ki_list);
Kent Overstreet36f55882013-05-07 16:18:41 -0700631
632 list_del_init(&req->ki_list);
Fabian Frederickd52a8f92014-04-22 07:26:58 +0200633 kiocb_cancel(req);
Kent Overstreet36f55882013-05-07 16:18:41 -0700634 }
635
636 spin_unlock_irq(&ctx->ctx_lock);
637
Kent Overstreete34ecee2013-10-10 19:31:47 -0700638 percpu_ref_kill(&ctx->reqs);
639 percpu_ref_put(&ctx->reqs);
Kent Overstreet36f55882013-05-07 16:18:41 -0700640}
641
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400642static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
643{
644 unsigned i, new_nr;
645 struct kioctx_table *table, *old;
646 struct aio_ring *ring;
647
648 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200649 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400650
651 while (1) {
652 if (table)
653 for (i = 0; i < table->nr; i++)
654 if (!table->table[i]) {
655 ctx->id = i;
656 table->table[i] = ctx;
657 spin_unlock(&mm->ioctx_lock);
658
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400659 /* While kioctx setup is in progress,
660 * we are protected from page migration
661 * changes ring_pages by ->ring_lock.
662 */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400663 ring = kmap_atomic(ctx->ring_pages[0]);
664 ring->id = ctx->id;
665 kunmap_atomic(ring);
666 return 0;
667 }
668
669 new_nr = (table ? table->nr : 1) * 4;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400670 spin_unlock(&mm->ioctx_lock);
671
672 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
673 new_nr, GFP_KERNEL);
674 if (!table)
675 return -ENOMEM;
676
677 table->nr = new_nr;
678
679 spin_lock(&mm->ioctx_lock);
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200680 old = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400681
682 if (!old) {
683 rcu_assign_pointer(mm->ioctx_table, table);
684 } else if (table->nr > old->nr) {
685 memcpy(table->table, old->table,
686 old->nr * sizeof(struct kioctx *));
687
688 rcu_assign_pointer(mm->ioctx_table, table);
689 kfree_rcu(old, rcu);
690 } else {
691 kfree(table);
692 table = old;
693 }
694 }
695}
696
Kent Overstreete34ecee2013-10-10 19:31:47 -0700697static void aio_nr_sub(unsigned nr)
698{
699 spin_lock(&aio_nr_lock);
700 if (WARN_ON(aio_nr - nr > aio_nr))
701 aio_nr = 0;
702 else
703 aio_nr -= nr;
704 spin_unlock(&aio_nr_lock);
705}
706
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707/* ioctx_alloc
708 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
709 */
710static struct kioctx *ioctx_alloc(unsigned nr_events)
711{
Zach Brown41003a72013-05-07 16:18:25 -0700712 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 struct kioctx *ctx;
Al Viroe23754f2012-03-06 14:33:22 -0500714 int err = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000716 /*
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300717 * Store the original nr_events -- what userspace passed to io_setup(),
718 * for counting against the global limit -- before it changes.
719 */
720 unsigned int max_reqs = nr_events;
721
722 /*
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000723 * We keep track of the number of available ringbuffer slots, to prevent
724 * overflow (reqs_available), and we also use percpu counters for this.
725 *
726 * So since up to half the slots might be on other cpu's percpu counters
727 * and unavailable, double nr_events so userspace sees what they
728 * expected: additionally, we move req_batch slots to/from percpu
729 * counters at a time, so make sure that isn't 0:
730 */
731 nr_events = max(nr_events, num_possible_cpus() * 4);
732 nr_events *= 2;
733
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734 /* Prevent overflows */
Al Viro08397ac2015-03-31 11:43:52 -0400735 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 pr_debug("ENOMEM: nr_events too high\n");
737 return ERR_PTR(-EINVAL);
738 }
739
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300740 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 return ERR_PTR(-EAGAIN);
742
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800743 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 if (!ctx)
745 return ERR_PTR(-ENOMEM);
746
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300747 ctx->max_reqs = max_reqs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400749 spin_lock_init(&ctx->ctx_lock);
750 spin_lock_init(&ctx->completion_lock);
751 mutex_init(&ctx->ring_lock);
752 /* Protect against page migration throughout kiotx setup by keeping
753 * the ring_lock mutex held until setup is complete. */
754 mutex_lock(&ctx->ring_lock);
755 init_waitqueue_head(&ctx->wait);
756
757 INIT_LIST_HEAD(&ctx->active_reqs);
758
Tejun Heo2aad2a82014-09-24 13:31:50 -0400759 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700760 goto err;
761
Tejun Heo2aad2a82014-09-24 13:31:50 -0400762 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
Kent Overstreete34ecee2013-10-10 19:31:47 -0700763 goto err;
Kent Overstreet723be6e2013-05-28 15:14:48 -0700764
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000765 ctx->cpu = alloc_percpu(struct kioctx_cpu);
766 if (!ctx->cpu)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700767 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300769 err = aio_setup_ring(ctx, nr_events);
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400770 if (err < 0)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700771 goto err;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000772
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000773 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000774 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
Benjamin LaHaise6878ea72013-07-31 10:34:18 -0400775 if (ctx->req_batch < 1)
776 ctx->req_batch = 1;
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000777
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 /* limit the number of system wide aios */
Al Viro9fa1cb32012-03-10 23:14:05 -0500779 spin_lock(&aio_nr_lock);
Mauricio Faria de Oliveira2a8a9862017-07-05 10:53:16 -0300780 if (aio_nr + ctx->max_reqs > aio_max_nr ||
781 aio_nr + ctx->max_reqs < aio_nr) {
Al Viro9fa1cb32012-03-10 23:14:05 -0500782 spin_unlock(&aio_nr_lock);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700783 err = -EAGAIN;
Gu Zhengd1b94322013-12-04 18:19:06 +0800784 goto err_ctx;
Al Viro2dd542b2012-03-10 23:10:35 -0500785 }
786 aio_nr += ctx->max_reqs;
Al Viro9fa1cb32012-03-10 23:14:05 -0500787 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788
Benjamin LaHaise18816862013-12-21 15:49:28 -0500789 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
790 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
Kent Overstreet723be6e2013-05-28 15:14:48 -0700791
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400792 err = ioctx_add_table(ctx, mm);
793 if (err)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700794 goto err_cleanup;
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400795
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400796 /* Release the ring_lock mutex now that all setup is complete. */
797 mutex_unlock(&ctx->ring_lock);
798
Kent Overstreetcaf41672013-05-07 16:18:35 -0700799 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700800 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 return ctx;
802
Kent Overstreete34ecee2013-10-10 19:31:47 -0700803err_cleanup:
804 aio_nr_sub(ctx->max_reqs);
Gu Zhengd1b94322013-12-04 18:19:06 +0800805err_ctx:
Al Virodeeb8522015-04-06 17:57:44 -0400806 atomic_set(&ctx->dead, 1);
807 if (ctx->mmap_size)
808 vm_munmap(ctx->mmap_base, ctx->mmap_size);
Gu Zhengd1b94322013-12-04 18:19:06 +0800809 aio_free_ring(ctx);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700810err:
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -0400811 mutex_unlock(&ctx->ring_lock);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000812 free_percpu(ctx->cpu);
Tejun Heo9a1049d2014-06-28 08:10:14 -0400813 percpu_ref_exit(&ctx->reqs);
814 percpu_ref_exit(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreetcaf41672013-05-07 16:18:35 -0700816 pr_debug("error allocating ioctx %d\n", err);
Al Viroe23754f2012-03-06 14:33:22 -0500817 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818}
819
Kent Overstreet36f55882013-05-07 16:18:41 -0700820/* kill_ioctx
821 * Cancels all outstanding aio requests on an aio context. Used
822 * when the processes owning a context have all exited to encourage
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 * the rapid destruction of the kioctx.
824 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400825static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
Jens Axboedc48e562015-04-15 11:17:23 -0600826 struct ctx_rq_wait *wait)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827{
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400828 struct kioctx_table *table;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400829
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400830 spin_lock(&mm->ioctx_lock);
Al Virob2edffd2015-04-06 17:48:54 -0400831 if (atomic_xchg(&ctx->dead, 1)) {
832 spin_unlock(&mm->ioctx_lock);
833 return -EINVAL;
834 }
835
Oleg Nesterov855ef0d2014-04-30 16:16:36 +0200836 table = rcu_dereference_raw(mm->ioctx_table);
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400837 WARN_ON(ctx != table->table[ctx->id]);
838 table->table[ctx->id] = NULL;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400839 spin_unlock(&mm->ioctx_lock);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700840
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400841 /* percpu_ref_kill() will do the necessary call_rcu() */
842 wake_up_all(&ctx->wait);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700843
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400844 /*
845 * It'd be more correct to do this in free_ioctx(), after all
846 * the outstanding kiocbs have finished - but by then io_destroy
847 * has already returned, so io_setup() could potentially return
848 * -EAGAIN with no ioctxs actually in use (as far as userspace
849 * could tell).
850 */
851 aio_nr_sub(ctx->max_reqs);
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -0400852
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400853 if (ctx->mmap_size)
854 vm_munmap(ctx->mmap_base, ctx->mmap_size);
855
Jens Axboedc48e562015-04-15 11:17:23 -0600856 ctx->rq_wait = wait;
Benjamin LaHaisefa88b6f2014-04-29 12:55:48 -0400857 percpu_ref_kill(&ctx->users);
858 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859}
860
Kent Overstreet36f55882013-05-07 16:18:41 -0700861/*
862 * exit_aio: called when the last user of mm goes away. At this point, there is
863 * no way for any new requests to be submited or any of the io_* syscalls to be
864 * called on the context.
865 *
866 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
867 * them.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800869void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870{
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200871 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
Jens Axboedc48e562015-04-15 11:17:23 -0600872 struct ctx_rq_wait wait;
873 int i, skipped;
Jens Axboeabf137d2008-12-09 08:11:22 +0100874
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200875 if (!table)
876 return;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400877
Jens Axboedc48e562015-04-15 11:17:23 -0600878 atomic_set(&wait.count, table->nr);
879 init_completion(&wait.comp);
880
881 skipped = 0;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200882 for (i = 0; i < table->nr; ++i) {
883 struct kioctx *ctx = table->table[i];
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400884
Jens Axboedc48e562015-04-15 11:17:23 -0600885 if (!ctx) {
886 skipped++;
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200887 continue;
Jens Axboedc48e562015-04-15 11:17:23 -0600888 }
889
Al Viro936af152012-04-20 21:49:41 -0400890 /*
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200891 * We don't need to bother with munmap() here - exit_mmap(mm)
892 * is coming and it'll unmap everything. And we simply can't,
893 * this is not necessarily our ->mm.
894 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
895 * that it needs to unmap the area, just set it to 0.
Al Viro936af152012-04-20 21:49:41 -0400896 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700897 ctx->mmap_size = 0;
Jens Axboedc48e562015-04-15 11:17:23 -0600898 kill_ioctx(mm, ctx, &wait);
899 }
Kent Overstreet36f55882013-05-07 16:18:41 -0700900
Jens Axboedc48e562015-04-15 11:17:23 -0600901 if (!atomic_sub_and_test(skipped, &wait.count)) {
Gu Zheng6098b452014-09-03 17:45:44 +0800902 /* Wait until all IO for the context are done. */
Jens Axboedc48e562015-04-15 11:17:23 -0600903 wait_for_completion(&wait.comp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 }
Oleg Nesterov4b70ac52014-04-30 19:02:48 +0200905
906 RCU_INIT_POINTER(mm->ioctx_table, NULL);
907 kfree(table);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908}
909
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000910static void put_reqs_available(struct kioctx *ctx, unsigned nr)
911{
912 struct kioctx_cpu *kcpu;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400913 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000914
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400915 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400916 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000917 kcpu->reqs_available += nr;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400918
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000919 while (kcpu->reqs_available >= ctx->req_batch * 2) {
920 kcpu->reqs_available -= ctx->req_batch;
921 atomic_add(ctx->req_batch, &ctx->reqs_available);
922 }
923
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400924 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000925}
926
927static bool get_reqs_available(struct kioctx *ctx)
928{
929 struct kioctx_cpu *kcpu;
930 bool ret = false;
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400931 unsigned long flags;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000932
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400933 local_irq_save(flags);
Benjamin LaHaisebe6fb452014-07-22 09:56:56 -0400934 kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000935 if (!kcpu->reqs_available) {
936 int old, avail = atomic_read(&ctx->reqs_available);
937
938 do {
939 if (avail < ctx->req_batch)
940 goto out;
941
942 old = avail;
943 avail = atomic_cmpxchg(&ctx->reqs_available,
944 avail, avail - ctx->req_batch);
945 } while (avail != old);
946
947 kcpu->reqs_available += ctx->req_batch;
948 }
949
950 ret = true;
951 kcpu->reqs_available--;
952out:
Benjamin LaHaise263782c2014-07-14 12:49:26 -0400953 local_irq_restore(flags);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000954 return ret;
955}
956
Benjamin LaHaised856f322014-08-24 13:14:05 -0400957/* refill_reqs_available
958 * Updates the reqs_available reference counts used for tracking the
959 * number of free slots in the completion ring. This can be called
960 * from aio_complete() (to optimistically update reqs_available) or
961 * from aio_get_req() (the we're out of events case). It must be
962 * called holding ctx->completion_lock.
963 */
964static void refill_reqs_available(struct kioctx *ctx, unsigned head,
965 unsigned tail)
966{
967 unsigned events_in_ring, completed;
968
969 /* Clamp head since userland can write to it. */
970 head %= ctx->nr_events;
971 if (head <= tail)
972 events_in_ring = tail - head;
973 else
974 events_in_ring = ctx->nr_events - (head - tail);
975
976 completed = ctx->completed_events;
977 if (events_in_ring < completed)
978 completed -= events_in_ring;
979 else
980 completed = 0;
981
982 if (!completed)
983 return;
984
985 ctx->completed_events -= completed;
986 put_reqs_available(ctx, completed);
987}
988
989/* user_refill_reqs_available
990 * Called to refill reqs_available when aio_get_req() encounters an
991 * out of space in the completion ring.
992 */
993static void user_refill_reqs_available(struct kioctx *ctx)
994{
995 spin_lock_irq(&ctx->completion_lock);
996 if (ctx->completed_events) {
997 struct aio_ring *ring;
998 unsigned head;
999
1000 /* Access of ring->head may race with aio_read_events_ring()
1001 * here, but that's okay since whether we read the old version
1002 * or the new version, and either will be valid. The important
1003 * part is that head cannot pass tail since we prevent
1004 * aio_complete() from updating tail by holding
1005 * ctx->completion_lock. Even if head is invalid, the check
1006 * against ctx->completed_events below will make sure we do the
1007 * safe/right thing.
1008 */
1009 ring = kmap_atomic(ctx->ring_pages[0]);
1010 head = ring->head;
1011 kunmap_atomic(ring);
1012
1013 refill_reqs_available(ctx, head, ctx->tail);
1014 }
1015
1016 spin_unlock_irq(&ctx->completion_lock);
1017}
1018
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019/* aio_get_req
Kent Overstreet57282d82013-05-13 13:42:52 -07001020 * Allocate a slot for an aio request.
1021 * Returns NULL if no requests are free.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001023static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001025 struct aio_kiocb *req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001026
Benjamin LaHaised856f322014-08-24 13:14:05 -04001027 if (!get_reqs_available(ctx)) {
1028 user_refill_reqs_available(ctx);
1029 if (!get_reqs_available(ctx))
1030 return NULL;
1031 }
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001032
Kent Overstreet0460fef2013-05-07 16:18:49 -07001033 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 if (unlikely(!req))
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001035 goto out_put;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036
Kent Overstreete34ecee2013-10-10 19:31:47 -07001037 percpu_ref_get(&ctx->reqs);
1038
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039 req->ki_ctx = ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 return req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001041out_put:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001042 put_reqs_available(ctx, 1);
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001043 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044}
1045
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001046static void kiocb_free(struct aio_kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001048 if (req->common.ki_filp)
1049 fput(req->common.ki_filp);
Davide Libenzi13389012009-06-30 11:41:11 -07001050 if (req->ki_eventfd != NULL)
1051 eventfd_ctx_put(req->ki_eventfd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 kmem_cache_free(kiocb_cachep, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053}
1054
Adrian Bunkd5470b52008-04-29 00:58:57 -07001055static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001057 struct aio_ring __user *ring = (void __user *)ctx_id;
Jens Axboeabf137d2008-12-09 08:11:22 +01001058 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -07001059 struct kioctx *ctx, *ret = NULL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001060 struct kioctx_table *table;
1061 unsigned id;
1062
1063 if (get_user(id, &ring->id))
1064 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065
Jens Axboeabf137d2008-12-09 08:11:22 +01001066 rcu_read_lock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001067 table = rcu_dereference(mm->ioctx_table);
Jens Axboeabf137d2008-12-09 08:11:22 +01001068
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001069 if (!table || id >= table->nr)
1070 goto out;
1071
1072 ctx = table->table[id];
Benjamin LaHaisef30d7042013-08-07 18:23:48 -04001073 if (ctx && ctx->user_id == ctx_id) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001074 percpu_ref_get(&ctx->users);
1075 ret = ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +01001076 }
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001077out:
Jens Axboeabf137d2008-12-09 08:11:22 +01001078 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -07001079 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080}
1081
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082/* aio_complete
1083 * Called when the io request on the given iocb is complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001085static void aio_complete(struct kiocb *kiocb, long res, long res2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001087 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088 struct kioctx *ctx = iocb->ki_ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 struct aio_ring *ring;
Kent Overstreet21b40202013-05-07 16:18:47 -07001090 struct io_event *ev_page, *event;
Benjamin LaHaised856f322014-08-24 13:14:05 -04001091 unsigned tail, pos, head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093
Jan Kara70fe2f42016-10-30 11:42:04 -05001094 if (kiocb->ki_flags & IOCB_WRITE) {
1095 struct file *file = kiocb->ki_filp;
1096
1097 /*
1098 * Tell lockdep we inherited freeze protection from submission
1099 * thread.
1100 */
Shaohua Lia12f1ae2016-12-13 12:09:56 -08001101 if (S_ISREG(file_inode(file)->i_mode))
1102 __sb_writers_acquired(file_inode(file)->i_sb, SB_FREEZE_WRITE);
Jan Kara70fe2f42016-10-30 11:42:04 -05001103 file_end_write(file);
1104 }
1105
Zach Brown20dcae32005-11-13 16:07:33 -08001106 /*
1107 * Special case handling for sync iocbs:
1108 * - events go directly into the iocb for fast handling
1109 * - the sync task with the iocb in its stack holds the single iocb
1110 * ref, no other paths have a way to get another ref
1111 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001113 BUG_ON(is_sync_kiocb(kiocb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114
Kent Overstreet0460fef2013-05-07 16:18:49 -07001115 if (iocb->ki_list.next) {
1116 unsigned long flags;
1117
1118 spin_lock_irqsave(&ctx->ctx_lock, flags);
1119 list_del(&iocb->ki_list);
1120 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1121 }
Kent Overstreet11599eb2013-05-07 16:18:39 -07001122
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 /*
Kent Overstreet0460fef2013-05-07 16:18:49 -07001124 * Add a completion event to the ring buffer. Must be done holding
Tang Chen4b30f072013-07-03 15:09:16 -07001125 * ctx->completion_lock to prevent other code from messing with the tail
Kent Overstreet0460fef2013-05-07 16:18:49 -07001126 * pointer since we might be called from irq context.
1127 */
1128 spin_lock_irqsave(&ctx->completion_lock, flags);
1129
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001130 tail = ctx->tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001131 pos = tail + AIO_EVENTS_OFFSET;
1132
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001133 if (++tail >= ctx->nr_events)
Ken Chen4bf69b22005-05-01 08:59:15 -07001134 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001136 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001137 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1138
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001139 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 event->data = iocb->ki_user_data;
1141 event->res = res;
1142 event->res2 = res2;
1143
Kent Overstreet21b40202013-05-07 16:18:47 -07001144 kunmap_atomic(ev_page);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001145 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -07001146
1147 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001148 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
Kent Overstreetcaf41672013-05-07 16:18:35 -07001149 res, res2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150
1151 /* after flagging the request as done, we
1152 * must never even look at it again
1153 */
1154 smp_wmb(); /* make event visible before updating tail */
1155
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001156 ctx->tail = tail;
Kent Overstreet21b40202013-05-07 16:18:47 -07001157
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001158 ring = kmap_atomic(ctx->ring_pages[0]);
Benjamin LaHaised856f322014-08-24 13:14:05 -04001159 head = ring->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160 ring->tail = tail;
Cong Wange8e3c3d2011-11-25 23:14:27 +08001161 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001162 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163
Benjamin LaHaised856f322014-08-24 13:14:05 -04001164 ctx->completed_events++;
1165 if (ctx->completed_events > 1)
1166 refill_reqs_available(ctx, head, tail);
Kent Overstreet0460fef2013-05-07 16:18:49 -07001167 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1168
Kent Overstreet21b40202013-05-07 16:18:47 -07001169 pr_debug("added to ring %p at [%u]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001170
1171 /*
1172 * Check if the user asked us to deliver the result through an
1173 * eventfd. The eventfd_signal() function is safe to be called
1174 * from IRQ context.
1175 */
Davide Libenzi87c3a862009-03-18 17:04:19 -07001176 if (iocb->ki_eventfd != NULL)
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001177 eventfd_signal(iocb->ki_eventfd, 1);
1178
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 /* everything turned out well, dispose of the aiocb. */
Kent Overstreet57282d82013-05-13 13:42:52 -07001180 kiocb_free(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181
Quentin Barnes6cb2a212008-03-19 17:00:39 -07001182 /*
1183 * We have to order our ring_info tail store above and test
1184 * of the wait list below outside the wait lock. This is
1185 * like in wake_up_bit() where clearing a bit has to be
1186 * ordered with the unlocked test.
1187 */
1188 smp_mb();
1189
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190 if (waitqueue_active(&ctx->wait))
1191 wake_up(&ctx->wait);
1192
Kent Overstreete34ecee2013-10-10 19:31:47 -07001193 percpu_ref_put(&ctx->reqs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194}
1195
Gu Zheng2be4e7d2014-07-23 18:03:53 +08001196/* aio_read_events_ring
Kent Overstreeta31ad382013-05-07 16:18:45 -07001197 * Pull an event off of the ioctx's event ring. Returns the number of
1198 * events fetched
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 */
Kent Overstreeta31ad382013-05-07 16:18:45 -07001200static long aio_read_events_ring(struct kioctx *ctx,
1201 struct io_event __user *event, long nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203 struct aio_ring *ring;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001204 unsigned head, tail, pos;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001205 long ret = 0;
1206 int copy_ret;
1207
Dave Chinner9c9ce762015-02-03 19:29:05 -05001208 /*
1209 * The mutex can block and wake us up and that will cause
1210 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1211 * and repeat. This should be rare enough that it doesn't cause
1212 * peformance issues. See the comment in read_events() for more detail.
1213 */
1214 sched_annotate_sleep();
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001215 mutex_lock(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216
Benjamin LaHaisefa8a53c2014-03-28 10:14:45 -04001217 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001218 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001219 head = ring->head;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001220 tail = ring->tail;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001221 kunmap_atomic(ring);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222
Jeff Moyer2ff396be2014-09-02 13:17:00 -04001223 /*
1224 * Ensure that once we've read the current tail pointer, that
1225 * we also see the events that were stored up to the tail.
1226 */
1227 smp_rmb();
1228
Kent Overstreet5ffac122013-05-09 15:36:07 -07001229 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001230
Kent Overstreet5ffac122013-05-09 15:36:07 -07001231 if (head == tail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 goto out;
1233
Benjamin LaHaiseedfbbf32014-06-24 13:32:51 -04001234 head %= ctx->nr_events;
1235 tail %= ctx->nr_events;
1236
Kent Overstreeta31ad382013-05-07 16:18:45 -07001237 while (ret < nr) {
1238 long avail;
1239 struct io_event *ev;
1240 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241
Kent Overstreet5ffac122013-05-09 15:36:07 -07001242 avail = (head <= tail ? tail : ctx->nr_events) - head;
1243 if (head == tail)
Kent Overstreeta31ad382013-05-07 16:18:45 -07001244 break;
1245
1246 avail = min(avail, nr - ret);
1247 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1248 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1249
1250 pos = head + AIO_EVENTS_OFFSET;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001251 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
Kent Overstreeta31ad382013-05-07 16:18:45 -07001252 pos %= AIO_EVENTS_PER_PAGE;
1253
1254 ev = kmap(page);
1255 copy_ret = copy_to_user(event + ret, ev + pos,
1256 sizeof(*ev) * avail);
1257 kunmap(page);
1258
1259 if (unlikely(copy_ret)) {
1260 ret = -EFAULT;
1261 goto out;
1262 }
1263
1264 ret += avail;
1265 head += avail;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001266 head %= ctx->nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001269 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001270 ring->head = head;
Zhao Hongjiang91d80a82013-04-26 11:03:53 +08001271 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001272 flush_dcache_page(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001273
Kent Overstreet5ffac122013-05-09 15:36:07 -07001274 pr_debug("%li h%u t%u\n", ret, head, tail);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001275out:
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001276 mutex_unlock(&ctx->ring_lock);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001277
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 return ret;
1279}
1280
Kent Overstreeta31ad382013-05-07 16:18:45 -07001281static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1282 struct io_event __user *event, long *i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001284 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285
Kent Overstreeta31ad382013-05-07 16:18:45 -07001286 if (ret > 0)
1287 *i += ret;
1288
1289 if (unlikely(atomic_read(&ctx->dead)))
1290 ret = -EINVAL;
1291
1292 if (!*i)
1293 *i = ret;
1294
1295 return ret < 0 || *i >= min_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296}
1297
Kent Overstreeta31ad382013-05-07 16:18:45 -07001298static long read_events(struct kioctx *ctx, long min_nr, long nr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299 struct io_event __user *event,
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07001300 ktime_t until)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001302 long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303
Kent Overstreeta31ad382013-05-07 16:18:45 -07001304 /*
1305 * Note that aio_read_events() is being called as the conditional - i.e.
1306 * we're calling it after prepare_to_wait() has set task state to
1307 * TASK_INTERRUPTIBLE.
1308 *
1309 * But aio_read_events() can block, and if it blocks it's going to flip
1310 * the task state back to TASK_RUNNING.
1311 *
1312 * This should be ok, provided it doesn't flip the state back to
1313 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1314 * will only happen if the mutex_lock() call blocks, and we then find
1315 * the ringbuffer empty. So in practice we should be ok, but it's
1316 * something to be aware of when touching this code.
1317 */
Thomas Gleixner2456e852016-12-25 11:38:40 +01001318 if (until == 0)
Fam Zheng5f785de2014-11-06 20:44:36 +08001319 aio_read_events(ctx, min_nr, nr, event, &ret);
1320 else
1321 wait_event_interruptible_hrtimeout(ctx->wait,
1322 aio_read_events(ctx, min_nr, nr, event, &ret),
1323 until);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324
Kent Overstreeta31ad382013-05-07 16:18:45 -07001325 if (!ret && signal_pending(current))
1326 ret = -EINTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327
Kent Overstreeta31ad382013-05-07 16:18:45 -07001328 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329}
1330
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331/* sys_io_setup:
1332 * Create an aio_context capable of receiving at least nr_events.
1333 * ctxp must not point to an aio_context that already exists, and
1334 * must be initialized to 0 prior to the call. On successful
1335 * creation of the aio_context, *ctxp is filled in with the resulting
1336 * handle. May fail with -EINVAL if *ctxp is not initialized,
1337 * if the specified nr_events exceeds internal limits. May fail
1338 * with -EAGAIN if the specified nr_events exceeds the user's limit
1339 * of available events. May fail with -ENOMEM if insufficient kernel
1340 * resources are available. May fail with -EFAULT if an invalid
1341 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1342 * implemented.
1343 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001344SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345{
1346 struct kioctx *ioctx = NULL;
1347 unsigned long ctx;
1348 long ret;
1349
1350 ret = get_user(ctx, ctxp);
1351 if (unlikely(ret))
1352 goto out;
1353
1354 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001355 if (unlikely(ctx || nr_events == 0)) {
Kinglong Meeacd88d42015-02-04 21:15:59 +08001356 pr_debug("EINVAL: ctx %lu nr_events %u\n",
Zach Brownd55b5fd2005-11-07 00:59:31 -08001357 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 goto out;
1359 }
1360
1361 ioctx = ioctx_alloc(nr_events);
1362 ret = PTR_ERR(ioctx);
1363 if (!IS_ERR(ioctx)) {
1364 ret = put_user(ioctx->user_id, ctxp);
Al Viroa2e18592012-03-20 16:27:57 -04001365 if (ret)
Anatol Pomozove02ba722014-04-15 11:31:33 -07001366 kill_ioctx(current->mm, ioctx, NULL);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001367 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 }
1369
1370out:
1371 return ret;
1372}
1373
Al Viroc00d2c72016-12-20 07:04:57 -05001374#ifdef CONFIG_COMPAT
1375COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1376{
1377 struct kioctx *ioctx = NULL;
1378 unsigned long ctx;
1379 long ret;
1380
1381 ret = get_user(ctx, ctx32p);
1382 if (unlikely(ret))
1383 goto out;
1384
1385 ret = -EINVAL;
1386 if (unlikely(ctx || nr_events == 0)) {
1387 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1388 ctx, nr_events);
1389 goto out;
1390 }
1391
1392 ioctx = ioctx_alloc(nr_events);
1393 ret = PTR_ERR(ioctx);
1394 if (!IS_ERR(ioctx)) {
1395 /* truncating is ok because it's a user address */
1396 ret = put_user((u32)ioctx->user_id, ctx32p);
1397 if (ret)
1398 kill_ioctx(current->mm, ioctx, NULL);
1399 percpu_ref_put(&ioctx->users);
1400 }
1401
1402out:
1403 return ret;
1404}
1405#endif
1406
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407/* sys_io_destroy:
1408 * Destroy the aio_context specified. May cancel any outstanding
1409 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001410 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 * is invalid.
1412 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001413SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414{
1415 struct kioctx *ioctx = lookup_ioctx(ctx);
1416 if (likely(NULL != ioctx)) {
Jens Axboedc48e562015-04-15 11:17:23 -06001417 struct ctx_rq_wait wait;
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001418 int ret;
Anatol Pomozove02ba722014-04-15 11:31:33 -07001419
Jens Axboedc48e562015-04-15 11:17:23 -06001420 init_completion(&wait.comp);
1421 atomic_set(&wait.count, 1);
1422
Anatol Pomozove02ba722014-04-15 11:31:33 -07001423 /* Pass requests_done to kill_ioctx() where it can be set
1424 * in a thread-safe way. If we try to set it here then we have
1425 * a race condition if two io_destroy() called simultaneously.
1426 */
Jens Axboedc48e562015-04-15 11:17:23 -06001427 ret = kill_ioctx(current->mm, ioctx, &wait);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001428 percpu_ref_put(&ioctx->users);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001429
1430 /* Wait until all IO for the context are done. Otherwise kernel
1431 * keep using user-space buffers even if user thinks the context
1432 * is destroyed.
1433 */
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001434 if (!ret)
Jens Axboedc48e562015-04-15 11:17:23 -06001435 wait_for_completion(&wait.comp);
Anatol Pomozove02ba722014-04-15 11:31:33 -07001436
Benjamin LaHaisefb2d4482014-04-29 12:45:17 -04001437 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 }
Kinglong Meeacd88d42015-02-04 21:15:59 +08001439 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 return -EINVAL;
1441}
1442
Christoph Hellwig89319d312016-10-30 11:42:03 -05001443static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec,
1444 bool vectored, bool compat, struct iov_iter *iter)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001445{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001446 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1447 size_t len = iocb->aio_nbytes;
1448
1449 if (!vectored) {
1450 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1451 *iovec = NULL;
1452 return ret;
1453 }
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001454#ifdef CONFIG_COMPAT
1455 if (compat)
Christoph Hellwig89319d312016-10-30 11:42:03 -05001456 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1457 iter);
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001458#endif
Christoph Hellwig89319d312016-10-30 11:42:03 -05001459 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001460}
1461
Christoph Hellwig89319d312016-10-30 11:42:03 -05001462static inline ssize_t aio_ret(struct kiocb *req, ssize_t ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463{
Christoph Hellwig89319d312016-10-30 11:42:03 -05001464 switch (ret) {
1465 case -EIOCBQUEUED:
1466 return ret;
1467 case -ERESTARTSYS:
1468 case -ERESTARTNOINTR:
1469 case -ERESTARTNOHAND:
1470 case -ERESTART_RESTARTBLOCK:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001471 /*
1472 * There's no easy way to restart the syscall since other AIO's
1473 * may be already running. Just fail this IO with EINTR.
1474 */
Christoph Hellwig89319d312016-10-30 11:42:03 -05001475 ret = -EINTR;
1476 /*FALLTHRU*/
1477 default:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001478 aio_complete(req, ret, 0);
Christoph Hellwig89319d312016-10-30 11:42:03 -05001479 return 0;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001480 }
Christoph Hellwig89319d312016-10-30 11:42:03 -05001481}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482
Christoph Hellwig89319d312016-10-30 11:42:03 -05001483static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
1484 bool compat)
1485{
1486 struct file *file = req->ki_filp;
1487 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1488 struct iov_iter iter;
1489 ssize_t ret;
1490
1491 if (unlikely(!(file->f_mode & FMODE_READ)))
1492 return -EBADF;
1493 if (unlikely(!file->f_op->read_iter))
1494 return -EINVAL;
1495
1496 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1497 if (ret)
1498 return ret;
1499 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1500 if (!ret)
Miklos Szeredibb7462b2017-02-20 16:51:23 +01001501 ret = aio_ret(req, call_read_iter(file, req, &iter));
Christoph Hellwig89319d312016-10-30 11:42:03 -05001502 kfree(iovec);
1503 return ret;
1504}
1505
1506static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored,
1507 bool compat)
1508{
1509 struct file *file = req->ki_filp;
1510 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1511 struct iov_iter iter;
1512 ssize_t ret;
1513
1514 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1515 return -EBADF;
1516 if (unlikely(!file->f_op->write_iter))
1517 return -EINVAL;
1518
1519 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1520 if (ret)
1521 return ret;
1522 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1523 if (!ret) {
Jan Kara70fe2f42016-10-30 11:42:04 -05001524 req->ki_flags |= IOCB_WRITE;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001525 file_start_write(file);
Miklos Szeredibb7462b2017-02-20 16:51:23 +01001526 ret = aio_ret(req, call_write_iter(file, req, &iter));
Jan Kara70fe2f42016-10-30 11:42:04 -05001527 /*
1528 * We release freeze protection in aio_complete(). Fool lockdep
1529 * by telling it the lock got released so that it doesn't
1530 * complain about held lock when we return to userspace.
1531 */
Shaohua Lia12f1ae2016-12-13 12:09:56 -08001532 if (S_ISREG(file_inode(file)->i_mode))
1533 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
Christoph Hellwig89319d312016-10-30 11:42:03 -05001534 }
1535 kfree(iovec);
1536 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537}
1538
Adrian Bunkd5470b52008-04-29 00:58:57 -07001539static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001540 struct iocb *iocb, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001542 struct aio_kiocb *req;
Christoph Hellwig89319d312016-10-30 11:42:03 -05001543 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 ssize_t ret;
1545
1546 /* enforce forwards compatibility on users */
Goldwyn Rodrigues9830f4b2017-06-20 07:05:42 -05001547 if (unlikely(iocb->aio_reserved2)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001548 pr_debug("EINVAL: reserve field set\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549 return -EINVAL;
1550 }
1551
1552 /* prevent overflows */
1553 if (unlikely(
1554 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1555 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1556 ((ssize_t)iocb->aio_nbytes < 0)
1557 )) {
Kinglong Meeacd88d42015-02-04 21:15:59 +08001558 pr_debug("EINVAL: overflow check\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559 return -EINVAL;
1560 }
1561
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001562 req = aio_get_req(ctx);
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001563 if (unlikely(!req))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564 return -EAGAIN;
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001565
Christoph Hellwig89319d312016-10-30 11:42:03 -05001566 req->common.ki_filp = file = fget(iocb->aio_fildes);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001567 if (unlikely(!req->common.ki_filp)) {
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001568 ret = -EBADF;
1569 goto out_put_req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 }
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001571 req->common.ki_pos = iocb->aio_offset;
1572 req->common.ki_complete = aio_complete;
Al Viro2ba48ce2015-04-09 13:52:01 -04001573 req->common.ki_flags = iocb_flags(req->common.ki_filp);
Jens Axboe45d06cf2017-06-27 11:01:22 -06001574 req->common.ki_hint = file_write_hint(file);
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001575
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001576 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1577 /*
1578 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1579 * instance of the file* now. The file descriptor must be
1580 * an eventfd() fd, and will be signaled for each completed
1581 * event using the eventfd_signal() function.
1582 */
Davide Libenzi13389012009-06-30 11:41:11 -07001583 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
Hirofumi Nakagawa801678c2008-04-29 01:03:09 -07001584 if (IS_ERR(req->ki_eventfd)) {
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001585 ret = PTR_ERR(req->ki_eventfd);
Davide Libenzi87c3a862009-03-18 17:04:19 -07001586 req->ki_eventfd = NULL;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001587 goto out_put_req;
1588 }
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001589
1590 req->common.ki_flags |= IOCB_EVENTFD;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001591 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592
Goldwyn Rodrigues9830f4b2017-06-20 07:05:42 -05001593 ret = kiocb_set_rw_flags(&req->common, iocb->aio_rw_flags);
1594 if (unlikely(ret)) {
1595 pr_debug("EINVAL: aio_rw_flags\n");
1596 goto out_put_req;
1597 }
1598
Kent Overstreet8a660892013-05-07 16:19:10 -07001599 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600 if (unlikely(ret)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001601 pr_debug("EFAULT: aio_key\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 goto out_put_req;
1603 }
1604
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001605 req->ki_user_iocb = user_iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 req->ki_user_data = iocb->aio_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607
Christoph Hellwig89319d312016-10-30 11:42:03 -05001608 get_file(file);
1609 switch (iocb->aio_lio_opcode) {
1610 case IOCB_CMD_PREAD:
1611 ret = aio_read(&req->common, iocb, false, compat);
1612 break;
1613 case IOCB_CMD_PWRITE:
1614 ret = aio_write(&req->common, iocb, false, compat);
1615 break;
1616 case IOCB_CMD_PREADV:
1617 ret = aio_read(&req->common, iocb, true, compat);
1618 break;
1619 case IOCB_CMD_PWRITEV:
1620 ret = aio_write(&req->common, iocb, true, compat);
1621 break;
1622 default:
1623 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1624 ret = -EINVAL;
1625 break;
1626 }
1627 fput(file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628
Christoph Hellwig89319d312016-10-30 11:42:03 -05001629 if (ret && ret != -EIOCBQUEUED)
1630 goto out_put_req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632out_put_req:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001633 put_reqs_available(ctx, 1);
Kent Overstreete34ecee2013-10-10 19:31:47 -07001634 percpu_ref_put(&ctx->reqs);
Kent Overstreet57282d82013-05-13 13:42:52 -07001635 kiocb_free(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636 return ret;
1637}
1638
Al Viroc00d2c72016-12-20 07:04:57 -05001639static long do_io_submit(aio_context_t ctx_id, long nr,
1640 struct iocb __user *__user *iocbpp, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641{
1642 struct kioctx *ctx;
1643 long ret = 0;
Jeff Moyer080d6762011-11-02 13:40:10 -07001644 int i = 0;
Shaohua Li9f5b9422010-07-01 07:55:01 +02001645 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646
1647 if (unlikely(nr < 0))
1648 return -EINVAL;
1649
Jeff Moyer75e1c702010-09-10 14:16:00 -07001650 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1651 nr = LONG_MAX/sizeof(*iocbpp);
1652
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1654 return -EFAULT;
1655
1656 ctx = lookup_ioctx(ctx_id);
1657 if (unlikely(!ctx)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001658 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 return -EINVAL;
1660 }
1661
Shaohua Li9f5b9422010-07-01 07:55:01 +02001662 blk_start_plug(&plug);
1663
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664 /*
1665 * AKPM: should this return a partial result if some of the IOs were
1666 * successfully submitted?
1667 */
1668 for (i=0; i<nr; i++) {
1669 struct iocb __user *user_iocb;
1670 struct iocb tmp;
1671
1672 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1673 ret = -EFAULT;
1674 break;
1675 }
1676
1677 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1678 ret = -EFAULT;
1679 break;
1680 }
1681
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001682 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683 if (ret)
1684 break;
1685 }
Shaohua Li9f5b9422010-07-01 07:55:01 +02001686 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687
Kent Overstreet723be6e2013-05-28 15:14:48 -07001688 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 return i ? i : ret;
1690}
1691
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001692/* sys_io_submit:
1693 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1694 * the number of iocbs queued. May return -EINVAL if the aio_context
1695 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1696 * *iocbpp[0] is not properly initialized, if the operation specified
1697 * is invalid for the file descriptor in the iocb. May fail with
1698 * -EFAULT if any of the data structures point to invalid data. May
1699 * fail with -EBADF if the file descriptor specified in the first
1700 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1701 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1702 * fail with -ENOSYS if not implemented.
1703 */
1704SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1705 struct iocb __user * __user *, iocbpp)
1706{
1707 return do_io_submit(ctx_id, nr, iocbpp, 0);
1708}
1709
Al Viroc00d2c72016-12-20 07:04:57 -05001710#ifdef CONFIG_COMPAT
1711static inline long
1712copy_iocb(long nr, u32 __user *ptr32, struct iocb __user * __user *ptr64)
1713{
1714 compat_uptr_t uptr;
1715 int i;
1716
1717 for (i = 0; i < nr; ++i) {
1718 if (get_user(uptr, ptr32 + i))
1719 return -EFAULT;
1720 if (put_user(compat_ptr(uptr), ptr64 + i))
1721 return -EFAULT;
1722 }
1723 return 0;
1724}
1725
1726#define MAX_AIO_SUBMITS (PAGE_SIZE/sizeof(struct iocb *))
1727
1728COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1729 int, nr, u32 __user *, iocb)
1730{
1731 struct iocb __user * __user *iocb64;
1732 long ret;
1733
1734 if (unlikely(nr < 0))
1735 return -EINVAL;
1736
1737 if (nr > MAX_AIO_SUBMITS)
1738 nr = MAX_AIO_SUBMITS;
1739
1740 iocb64 = compat_alloc_user_space(nr * sizeof(*iocb64));
1741 ret = copy_iocb(nr, iocb, iocb64);
1742 if (!ret)
1743 ret = do_io_submit(ctx_id, nr, iocb64, 1);
1744 return ret;
1745}
1746#endif
1747
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748/* lookup_kiocb
1749 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001751static struct aio_kiocb *
1752lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753{
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001754 struct aio_kiocb *kiocb;
Zach Brownd00689a2005-11-13 16:07:34 -08001755
1756 assert_spin_locked(&ctx->ctx_lock);
1757
Kent Overstreet8a660892013-05-07 16:19:10 -07001758 if (key != KIOCB_KEY)
1759 return NULL;
1760
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 /* TODO: use a hash or array, this sucks. */
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001762 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1763 if (kiocb->ki_user_iocb == iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764 return kiocb;
1765 }
1766 return NULL;
1767}
1768
1769/* sys_io_cancel:
1770 * Attempts to cancel an iocb previously passed to io_submit. If
1771 * the operation is successfully cancelled, the resulting event is
1772 * copied into the memory pointed to by result without being placed
1773 * into the completion queue and 0 is returned. May fail with
1774 * -EFAULT if any of the data structures pointed to are invalid.
1775 * May fail with -EINVAL if aio_context specified by ctx_id is
1776 * invalid. May fail with -EAGAIN if the iocb specified was not
1777 * cancelled. Will fail with -ENOSYS if not implemented.
1778 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001779SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1780 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782 struct kioctx *ctx;
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001783 struct aio_kiocb *kiocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 u32 key;
1785 int ret;
1786
1787 ret = get_user(key, &iocb->aio_key);
1788 if (unlikely(ret))
1789 return -EFAULT;
1790
1791 ctx = lookup_ioctx(ctx_id);
1792 if (unlikely(!ctx))
1793 return -EINVAL;
1794
1795 spin_lock_irq(&ctx->ctx_lock);
Kent Overstreet906b9732013-05-07 16:18:31 -07001796
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 kiocb = lookup_kiocb(ctx, iocb, key);
Kent Overstreet906b9732013-05-07 16:18:31 -07001798 if (kiocb)
Fabian Frederickd52a8f92014-04-22 07:26:58 +02001799 ret = kiocb_cancel(kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -07001800 else
1801 ret = -EINVAL;
1802
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 spin_unlock_irq(&ctx->ctx_lock);
1804
Kent Overstreet906b9732013-05-07 16:18:31 -07001805 if (!ret) {
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001806 /*
1807 * The result argument is no longer used - the io_event is
1808 * always delivered via the ring buffer. -EINPROGRESS indicates
1809 * cancellation is progress:
Kent Overstreet906b9732013-05-07 16:18:31 -07001810 */
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001811 ret = -EINPROGRESS;
Kent Overstreet906b9732013-05-07 16:18:31 -07001812 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813
Kent Overstreet723be6e2013-05-28 15:14:48 -07001814 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815
1816 return ret;
1817}
1818
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07001819static long do_io_getevents(aio_context_t ctx_id,
1820 long min_nr,
1821 long nr,
1822 struct io_event __user *events,
1823 struct timespec64 *ts)
1824{
1825 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
1826 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1827 long ret = -EINVAL;
1828
1829 if (likely(ioctx)) {
1830 if (likely(min_nr <= nr && min_nr >= 0))
1831 ret = read_events(ioctx, min_nr, nr, events, until);
1832 percpu_ref_put(&ioctx->users);
1833 }
1834
1835 return ret;
1836}
1837
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838/* io_getevents:
1839 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001840 * the completion queue for the aio_context specified by ctx_id. If
1841 * it succeeds, the number of read events is returned. May fail with
1842 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1843 * out of range, if timeout is out of range. May fail with -EFAULT
1844 * if any of the memory specified is invalid. May return 0 or
1845 * < min_nr if the timeout specified by timeout has elapsed
1846 * before sufficient events are available, where timeout == NULL
1847 * specifies an infinite timeout. Note that the timeout pointed to by
Jeff Moyer69008072013-05-24 15:55:24 -07001848 * timeout is relative. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001850SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1851 long, min_nr,
1852 long, nr,
1853 struct io_event __user *, events,
1854 struct timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855{
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07001856 struct timespec64 ts;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07001858 if (timeout) {
1859 if (unlikely(get_timespec64(&ts, timeout)))
1860 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861 }
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07001862
1863 return do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864}
Al Viroc00d2c72016-12-20 07:04:57 -05001865
1866#ifdef CONFIG_COMPAT
1867COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id,
1868 compat_long_t, min_nr,
1869 compat_long_t, nr,
1870 struct io_event __user *, events,
1871 struct compat_timespec __user *, timeout)
1872{
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07001873 struct timespec64 t;
Al Viroc00d2c72016-12-20 07:04:57 -05001874
1875 if (timeout) {
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07001876 if (compat_get_timespec64(&t, timeout))
Al Viroc00d2c72016-12-20 07:04:57 -05001877 return -EFAULT;
1878
Al Viroc00d2c72016-12-20 07:04:57 -05001879 }
Deepa Dinamanifa2e62a2017-08-04 21:12:32 -07001880
1881 return do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
Al Viroc00d2c72016-12-20 07:04:57 -05001882}
1883#endif