blob: 062a5f6a1448c6cff1cd1dc09e84db3f0d59cc27 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 *
5 * Implements an efficient asynchronous io interface.
6 *
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 *
9 * See ../COPYING for licensing terms.
10 */
Kent Overstreetcaf41672013-05-07 16:18:35 -070011#define pr_fmt(fmt) "%s: " fmt, __func__
12
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/kernel.h>
14#include <linux/init.h>
15#include <linux/errno.h>
16#include <linux/time.h>
17#include <linux/aio_abi.h>
Paul Gortmaker630d9c42011-11-16 23:57:37 -050018#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/syscalls.h>
Jens Axboeb9d128f2009-10-29 13:59:26 +010020#include <linux/backing-dev.h>
Badari Pulavarty027445c2006-09-30 23:28:46 -070021#include <linux/uio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070022
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/sched.h>
24#include <linux/fs.h>
25#include <linux/file.h>
26#include <linux/mm.h>
27#include <linux/mman.h>
Michael S. Tsirkin3d2d8272009-09-21 17:03:51 -070028#include <linux/mmu_context.h>
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100029#include <linux/percpu.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#include <linux/slab.h>
31#include <linux/timer.h>
32#include <linux/aio.h>
33#include <linux/highmem.h>
34#include <linux/workqueue.h>
35#include <linux/security.h>
Davide Libenzi9c3060b2007-05-10 22:23:21 -070036#include <linux/eventfd.h>
Jeff Moyercfb1e332009-10-02 18:57:36 -040037#include <linux/blkdev.h>
Jeff Moyer9d85cba2010-05-26 14:44:26 -070038#include <linux/compat.h>
Gu Zheng36bc08c2013-07-16 17:56:16 +080039#include <linux/migrate.h>
40#include <linux/ramfs.h>
Kent Overstreet723be6e2013-05-28 15:14:48 -070041#include <linux/percpu-refcount.h>
Benjamin LaHaise71ad7492013-09-17 10:18:25 -040042#include <linux/mount.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44#include <asm/kmap_types.h>
45#include <asm/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
Al Viro68d70d02013-06-19 15:26:04 +040047#include "internal.h"
48
Kent Overstreet4e179bc2013-05-07 16:18:33 -070049#define AIO_RING_MAGIC 0xa10a10a1
50#define AIO_RING_COMPAT_FEATURES 1
51#define AIO_RING_INCOMPAT_FEATURES 0
52struct aio_ring {
53 unsigned id; /* kernel internal index number */
54 unsigned nr; /* number of io_events */
55 unsigned head;
56 unsigned tail;
57
58 unsigned magic;
59 unsigned compat_features;
60 unsigned incompat_features;
61 unsigned header_length; /* size of aio_ring */
62
63
64 struct io_event io_events[0];
65}; /* 128 bytes + ring size */
66
67#define AIO_RING_PAGES 8
Kent Overstreet4e179bc2013-05-07 16:18:33 -070068
Benjamin LaHaisedb446a02013-07-30 12:54:40 -040069struct kioctx_table {
70 struct rcu_head rcu;
71 unsigned nr;
72 struct kioctx *table[];
73};
74
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100075struct kioctx_cpu {
76 unsigned reqs_available;
77};
78
Kent Overstreet4e179bc2013-05-07 16:18:33 -070079struct kioctx {
Kent Overstreet723be6e2013-05-28 15:14:48 -070080 struct percpu_ref users;
Kent Overstreet36f55882013-05-07 16:18:41 -070081 atomic_t dead;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070082
Kent Overstreete34ecee2013-10-10 19:31:47 -070083 struct percpu_ref reqs;
84
Kent Overstreet4e179bc2013-05-07 16:18:33 -070085 unsigned long user_id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -070086
Kent Overstreete1bdd5f2013-04-26 10:58:39 +100087 struct __percpu kioctx_cpu *cpu;
88
89 /*
90 * For percpu reqs_available, number of slots we move to/from global
91 * counter at a time:
92 */
93 unsigned req_batch;
Kent Overstreet3e845ce2013-05-07 16:18:51 -070094 /*
95 * This is what userspace passed to io_setup(), it's not used for
96 * anything but counting against the global max_reqs quota.
97 *
Kent Overstreet58c85dc2013-05-07 16:18:55 -070098 * The real limit is nr_events - 1, which will be larger (see
Kent Overstreet3e845ce2013-05-07 16:18:51 -070099 * aio_setup_ring())
100 */
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700101 unsigned max_reqs;
102
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700103 /* Size of ringbuffer, in units of struct io_event */
104 unsigned nr_events;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700105
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700106 unsigned long mmap_base;
107 unsigned long mmap_size;
108
109 struct page **ring_pages;
110 long nr_pages;
111
Kent Overstreet723be6e2013-05-28 15:14:48 -0700112 struct work_struct free_work;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700113
114 struct {
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000115 /*
116 * This counts the number of available slots in the ringbuffer,
117 * so we avoid overflowing it: it's decremented (if positive)
118 * when allocating a kiocb and incremented when the resulting
119 * io_event is pulled off the ringbuffer.
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000120 *
121 * We batch accesses to it with a percpu version.
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000122 */
123 atomic_t reqs_available;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700124 } ____cacheline_aligned_in_smp;
125
126 struct {
127 spinlock_t ctx_lock;
128 struct list_head active_reqs; /* used for cancellation */
129 } ____cacheline_aligned_in_smp;
130
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700131 struct {
132 struct mutex ring_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700133 wait_queue_head_t wait;
134 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700135
136 struct {
137 unsigned tail;
138 spinlock_t completion_lock;
Kent Overstreet4e23bca2013-05-07 16:18:56 -0700139 } ____cacheline_aligned_in_smp;
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700140
141 struct page *internal_pages[AIO_RING_PAGES];
Gu Zheng36bc08c2013-07-16 17:56:16 +0800142 struct file *aio_ring_file;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400143
144 unsigned id;
Kent Overstreet4e179bc2013-05-07 16:18:33 -0700145};
146
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -0800148static DEFINE_SPINLOCK(aio_nr_lock);
149unsigned long aio_nr; /* current system wide number of aio requests */
150unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151/*----end sysctl variables---*/
152
Christoph Lametere18b8902006-12-06 20:33:20 -0800153static struct kmem_cache *kiocb_cachep;
154static struct kmem_cache *kioctx_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400156static struct vfsmount *aio_mnt;
157
158static const struct file_operations aio_ring_fops;
159static const struct address_space_operations aio_ctx_aops;
160
161static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
162{
163 struct qstr this = QSTR_INIT("[aio]", 5);
164 struct file *file;
165 struct path path;
166 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
Dan Carpenter7f626562013-11-13 10:49:40 +0300167 if (IS_ERR(inode))
168 return ERR_CAST(inode);
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400169
170 inode->i_mapping->a_ops = &aio_ctx_aops;
171 inode->i_mapping->private_data = ctx;
172 inode->i_size = PAGE_SIZE * nr_pages;
173
174 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
175 if (!path.dentry) {
176 iput(inode);
177 return ERR_PTR(-ENOMEM);
178 }
179 path.mnt = mntget(aio_mnt);
180
181 d_instantiate(path.dentry, inode);
182 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
183 if (IS_ERR(file)) {
184 path_put(&path);
185 return file;
186 }
187
188 file->f_flags = O_RDWR;
189 file->private_data = ctx;
190 return file;
191}
192
193static struct dentry *aio_mount(struct file_system_type *fs_type,
194 int flags, const char *dev_name, void *data)
195{
196 static const struct dentry_operations ops = {
197 .d_dname = simple_dname,
198 };
199 return mount_pseudo(fs_type, "aio:", NULL, &ops, 0xa10a10a1);
200}
201
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202/* aio_setup
203 * Creates the slab caches used by the aio routines, panic on
204 * failure as this is done early during the boot sequence.
205 */
206static int __init aio_setup(void)
207{
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400208 static struct file_system_type aio_fs = {
209 .name = "aio",
210 .mount = aio_mount,
211 .kill_sb = kill_anon_super,
212 };
213 aio_mnt = kern_mount(&aio_fs);
214 if (IS_ERR(aio_mnt))
215 panic("Failed to create aio fs mount.");
216
Christoph Lameter0a31bd52007-05-06 14:49:57 -0700217 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
218 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219
Kent Overstreetcaf41672013-05-07 16:18:35 -0700220 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221
222 return 0;
223}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700224__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400226static void put_aio_ring_file(struct kioctx *ctx)
227{
228 struct file *aio_ring_file = ctx->aio_ring_file;
229 if (aio_ring_file) {
230 truncate_setsize(aio_ring_file->f_inode, 0);
231
232 /* Prevent further access to the kioctx from migratepages */
233 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
234 aio_ring_file->f_inode->i_mapping->private_data = NULL;
235 ctx->aio_ring_file = NULL;
236 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
237
238 fput(aio_ring_file);
239 }
240}
241
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242static void aio_free_ring(struct kioctx *ctx)
243{
Gu Zheng36bc08c2013-07-16 17:56:16 +0800244 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245
Gu Zheng36bc08c2013-07-16 17:56:16 +0800246 for (i = 0; i < ctx->nr_pages; i++) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500247 struct page *page;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800248 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
249 page_count(ctx->ring_pages[i]));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500250 page = ctx->ring_pages[i];
251 if (!page)
252 continue;
253 ctx->ring_pages[i] = NULL;
254 put_page(page);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800255 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400257 put_aio_ring_file(ctx);
258
Sasha Levinddb8c452013-11-19 17:33:03 -0500259 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700260 kfree(ctx->ring_pages);
Sasha Levinddb8c452013-11-19 17:33:03 -0500261 ctx->ring_pages = NULL;
262 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263}
264
Gu Zheng36bc08c2013-07-16 17:56:16 +0800265static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
266{
267 vma->vm_ops = &generic_file_vm_ops;
268 return 0;
269}
270
271static const struct file_operations aio_ring_fops = {
272 .mmap = aio_ring_mmap,
273};
274
275static int aio_set_page_dirty(struct page *page)
276{
277 return 0;
278}
279
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400280#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800281static int aio_migratepage(struct address_space *mapping, struct page *new,
282 struct page *old, enum migrate_mode mode)
283{
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400284 struct kioctx *ctx;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800285 unsigned long flags;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800286 int rc;
287
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500288 rc = 0;
289
290 /* Make sure the old page hasn't already been changed */
291 spin_lock(&mapping->private_lock);
292 ctx = mapping->private_data;
293 if (ctx) {
294 pgoff_t idx;
295 spin_lock_irqsave(&ctx->completion_lock, flags);
296 idx = old->index;
297 if (idx < (pgoff_t)ctx->nr_pages) {
298 if (ctx->ring_pages[idx] != old)
299 rc = -EAGAIN;
300 } else
301 rc = -EINVAL;
302 spin_unlock_irqrestore(&ctx->completion_lock, flags);
303 } else
304 rc = -EINVAL;
305 spin_unlock(&mapping->private_lock);
306
307 if (rc != 0)
308 return rc;
309
Gu Zheng36bc08c2013-07-16 17:56:16 +0800310 /* Writeback must be complete */
311 BUG_ON(PageWriteback(old));
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500312 get_page(new);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800313
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500314 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800315 if (rc != MIGRATEPAGE_SUCCESS) {
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500316 put_page(new);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800317 return rc;
318 }
319
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400320 /* We can potentially race against kioctx teardown here. Use the
321 * address_space's private data lock to protect the mapping's
322 * private_data.
323 */
324 spin_lock(&mapping->private_lock);
325 ctx = mapping->private_data;
326 if (ctx) {
327 pgoff_t idx;
328 spin_lock_irqsave(&ctx->completion_lock, flags);
329 migrate_page_copy(new, old);
330 idx = old->index;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500331 if (idx < (pgoff_t)ctx->nr_pages) {
332 /* And only do the move if things haven't changed */
333 if (ctx->ring_pages[idx] == old)
334 ctx->ring_pages[idx] = new;
335 else
336 rc = -EAGAIN;
337 } else
338 rc = -EINVAL;
Benjamin LaHaise5e9ae2e2013-09-26 20:34:51 -0400339 spin_unlock_irqrestore(&ctx->completion_lock, flags);
340 } else
341 rc = -EBUSY;
342 spin_unlock(&mapping->private_lock);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800343
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500344 if (rc == MIGRATEPAGE_SUCCESS)
345 put_page(old);
346 else
347 put_page(new);
348
Gu Zheng36bc08c2013-07-16 17:56:16 +0800349 return rc;
350}
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400351#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800352
353static const struct address_space_operations aio_ctx_aops = {
354 .set_page_dirty = aio_set_page_dirty,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400355#if IS_ENABLED(CONFIG_MIGRATION)
Gu Zheng36bc08c2013-07-16 17:56:16 +0800356 .migratepage = aio_migratepage,
Benjamin LaHaise0c453552013-07-17 09:34:24 -0400357#endif
Gu Zheng36bc08c2013-07-16 17:56:16 +0800358};
359
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360static int aio_setup_ring(struct kioctx *ctx)
361{
362 struct aio_ring *ring;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363 unsigned nr_events = ctx->max_reqs;
Zach Brown41003a72013-05-07 16:18:25 -0700364 struct mm_struct *mm = current->mm;
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900365 unsigned long size, unused;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366 int nr_pages;
Gu Zheng36bc08c2013-07-16 17:56:16 +0800367 int i;
368 struct file *file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369
370 /* Compensate for the ring buffer's head/tail overlap entry */
371 nr_events += 2; /* 1 is required, 2 for good luck */
372
373 size = sizeof(struct aio_ring);
374 size += sizeof(struct io_event) * nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375
Gu Zheng36bc08c2013-07-16 17:56:16 +0800376 nr_pages = PFN_UP(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 if (nr_pages < 0)
378 return -EINVAL;
379
Benjamin LaHaise71ad7492013-09-17 10:18:25 -0400380 file = aio_private_file(ctx, nr_pages);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800381 if (IS_ERR(file)) {
382 ctx->aio_ring_file = NULL;
383 return -EAGAIN;
384 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385
Gu Zheng36bc08c2013-07-16 17:56:16 +0800386 ctx->aio_ring_file = file;
387 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
388 / sizeof(struct io_event);
389
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700390 ctx->ring_pages = ctx->internal_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 if (nr_pages > AIO_RING_PAGES) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700392 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
393 GFP_KERNEL);
Gu Zhengd1b94322013-12-04 18:19:06 +0800394 if (!ctx->ring_pages) {
395 put_aio_ring_file(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 return -ENOMEM;
Gu Zhengd1b94322013-12-04 18:19:06 +0800397 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 }
399
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900400 for (i = 0; i < nr_pages; i++) {
401 struct page *page;
402 page = find_or_create_page(file->f_inode->i_mapping,
403 i, GFP_HIGHUSER | __GFP_ZERO);
404 if (!page)
405 break;
406 pr_debug("pid(%d) page[%d]->count=%d\n",
407 current->pid, i, page_count(page));
408 SetPageUptodate(page);
409 SetPageDirty(page);
410 unlock_page(page);
411
412 ctx->ring_pages[i] = page;
413 }
414 ctx->nr_pages = i;
415
416 if (unlikely(i != nr_pages)) {
417 aio_free_ring(ctx);
418 return -EAGAIN;
419 }
420
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700421 ctx->mmap_size = nr_pages * PAGE_SIZE;
422 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800423
Zach Brown41003a72013-05-07 16:18:25 -0700424 down_write(&mm->mmap_sem);
Gu Zheng36bc08c2013-07-16 17:56:16 +0800425 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
426 PROT_READ | PROT_WRITE,
Linus Torvalds3dc9acb2013-12-20 05:11:12 +0900427 MAP_SHARED, 0, &unused);
428 up_write(&mm->mmap_sem);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700429 if (IS_ERR((void *)ctx->mmap_base)) {
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700430 ctx->mmap_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431 aio_free_ring(ctx);
432 return -EAGAIN;
433 }
434
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700435 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
Benjamin LaHaised6c355c2013-09-09 11:57:59 -0400436
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700437 ctx->user_id = ctx->mmap_base;
438 ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700440 ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441 ring->nr = nr_events; /* user copy */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400442 ring->id = ~0U;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 ring->head = ring->tail = 0;
444 ring->magic = AIO_RING_MAGIC;
445 ring->compat_features = AIO_RING_COMPAT_FEATURES;
446 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
447 ring->header_length = sizeof(struct aio_ring);
Cong Wange8e3c3d2011-11-25 23:14:27 +0800448 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700449 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
451 return 0;
452}
453
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
455#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
456#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
457
Kent Overstreet0460fef2013-05-07 16:18:49 -0700458void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
459{
460 struct kioctx *ctx = req->ki_ctx;
461 unsigned long flags;
462
463 spin_lock_irqsave(&ctx->ctx_lock, flags);
464
465 if (!req->ki_list.next)
466 list_add(&req->ki_list, &ctx->active_reqs);
467
468 req->ki_cancel = cancel;
469
470 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
471}
472EXPORT_SYMBOL(kiocb_set_cancel_fn);
473
Kent Overstreetbec68faa2013-05-13 14:45:08 -0700474static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
Kent Overstreet906b9732013-05-07 16:18:31 -0700475{
Kent Overstreet0460fef2013-05-07 16:18:49 -0700476 kiocb_cancel_fn *old, *cancel;
Kent Overstreet906b9732013-05-07 16:18:31 -0700477
Kent Overstreet0460fef2013-05-07 16:18:49 -0700478 /*
479 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
480 * actually has a cancel function, hence the cmpxchg()
481 */
Kent Overstreet906b9732013-05-07 16:18:31 -0700482
Kent Overstreet0460fef2013-05-07 16:18:49 -0700483 cancel = ACCESS_ONCE(kiocb->ki_cancel);
484 do {
485 if (!cancel || cancel == KIOCB_CANCELLED)
Kent Overstreet57282d82013-05-13 13:42:52 -0700486 return -EINVAL;
Kent Overstreet906b9732013-05-07 16:18:31 -0700487
Kent Overstreet0460fef2013-05-07 16:18:49 -0700488 old = cancel;
489 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
490 } while (cancel != old);
491
Kent Overstreet57282d82013-05-13 13:42:52 -0700492 return cancel(kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -0700493}
494
Kent Overstreete34ecee2013-10-10 19:31:47 -0700495static void free_ioctx(struct work_struct *work)
Kent Overstreet36f55882013-05-07 16:18:41 -0700496{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700497 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000498
Kent Overstreete34ecee2013-10-10 19:31:47 -0700499 pr_debug("freeing %p\n", ctx);
500
501 aio_free_ring(ctx);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000502 free_percpu(ctx->cpu);
Kent Overstreet36f55882013-05-07 16:18:41 -0700503 kmem_cache_free(kioctx_cachep, ctx);
504}
505
Kent Overstreete34ecee2013-10-10 19:31:47 -0700506static void free_ioctx_reqs(struct percpu_ref *ref)
507{
508 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
509
510 INIT_WORK(&ctx->free_work, free_ioctx);
511 schedule_work(&ctx->free_work);
512}
513
Kent Overstreet36f55882013-05-07 16:18:41 -0700514/*
515 * When this function runs, the kioctx has been removed from the "hash table"
516 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
517 * now it's safe to cancel any that need to be.
518 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700519static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet36f55882013-05-07 16:18:41 -0700520{
Kent Overstreete34ecee2013-10-10 19:31:47 -0700521 struct kioctx *ctx = container_of(ref, struct kioctx, users);
Kent Overstreet36f55882013-05-07 16:18:41 -0700522 struct kiocb *req;
523
524 spin_lock_irq(&ctx->ctx_lock);
525
526 while (!list_empty(&ctx->active_reqs)) {
527 req = list_first_entry(&ctx->active_reqs,
528 struct kiocb, ki_list);
529
530 list_del_init(&req->ki_list);
Kent Overstreetbec68faa2013-05-13 14:45:08 -0700531 kiocb_cancel(ctx, req);
Kent Overstreet36f55882013-05-07 16:18:41 -0700532 }
533
534 spin_unlock_irq(&ctx->ctx_lock);
535
Kent Overstreete34ecee2013-10-10 19:31:47 -0700536 percpu_ref_kill(&ctx->reqs);
537 percpu_ref_put(&ctx->reqs);
Kent Overstreet36f55882013-05-07 16:18:41 -0700538}
539
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400540static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
541{
542 unsigned i, new_nr;
543 struct kioctx_table *table, *old;
544 struct aio_ring *ring;
545
546 spin_lock(&mm->ioctx_lock);
Artem Savkovd9b2c872013-09-09 12:29:35 -0400547 rcu_read_lock();
Benjamin LaHaise77d30b12013-08-30 11:12:50 -0400548 table = rcu_dereference(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400549
550 while (1) {
551 if (table)
552 for (i = 0; i < table->nr; i++)
553 if (!table->table[i]) {
554 ctx->id = i;
555 table->table[i] = ctx;
Artem Savkovd9b2c872013-09-09 12:29:35 -0400556 rcu_read_unlock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400557 spin_unlock(&mm->ioctx_lock);
558
559 ring = kmap_atomic(ctx->ring_pages[0]);
560 ring->id = ctx->id;
561 kunmap_atomic(ring);
562 return 0;
563 }
564
565 new_nr = (table ? table->nr : 1) * 4;
566
Artem Savkovd9b2c872013-09-09 12:29:35 -0400567 rcu_read_unlock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400568 spin_unlock(&mm->ioctx_lock);
569
570 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
571 new_nr, GFP_KERNEL);
572 if (!table)
573 return -ENOMEM;
574
575 table->nr = new_nr;
576
577 spin_lock(&mm->ioctx_lock);
Artem Savkovd9b2c872013-09-09 12:29:35 -0400578 rcu_read_lock();
Benjamin LaHaise77d30b12013-08-30 11:12:50 -0400579 old = rcu_dereference(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400580
581 if (!old) {
582 rcu_assign_pointer(mm->ioctx_table, table);
583 } else if (table->nr > old->nr) {
584 memcpy(table->table, old->table,
585 old->nr * sizeof(struct kioctx *));
586
587 rcu_assign_pointer(mm->ioctx_table, table);
588 kfree_rcu(old, rcu);
589 } else {
590 kfree(table);
591 table = old;
592 }
593 }
594}
595
Kent Overstreete34ecee2013-10-10 19:31:47 -0700596static void aio_nr_sub(unsigned nr)
597{
598 spin_lock(&aio_nr_lock);
599 if (WARN_ON(aio_nr - nr > aio_nr))
600 aio_nr = 0;
601 else
602 aio_nr -= nr;
603 spin_unlock(&aio_nr_lock);
604}
605
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606/* ioctx_alloc
607 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
608 */
609static struct kioctx *ioctx_alloc(unsigned nr_events)
610{
Zach Brown41003a72013-05-07 16:18:25 -0700611 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 struct kioctx *ctx;
Al Viroe23754f2012-03-06 14:33:22 -0500613 int err = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000615 /*
616 * We keep track of the number of available ringbuffer slots, to prevent
617 * overflow (reqs_available), and we also use percpu counters for this.
618 *
619 * So since up to half the slots might be on other cpu's percpu counters
620 * and unavailable, double nr_events so userspace sees what they
621 * expected: additionally, we move req_batch slots to/from percpu
622 * counters at a time, so make sure that isn't 0:
623 */
624 nr_events = max(nr_events, num_possible_cpus() * 4);
625 nr_events *= 2;
626
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 /* Prevent overflows */
628 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
629 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
630 pr_debug("ENOMEM: nr_events too high\n");
631 return ERR_PTR(-EINVAL);
632 }
633
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400634 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 return ERR_PTR(-EAGAIN);
636
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800637 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 if (!ctx)
639 return ERR_PTR(-ENOMEM);
640
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 ctx->max_reqs = nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642
Kent Overstreete34ecee2013-10-10 19:31:47 -0700643 if (percpu_ref_init(&ctx->users, free_ioctx_users))
644 goto err;
645
646 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
647 goto err;
Kent Overstreet723be6e2013-05-28 15:14:48 -0700648
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 spin_lock_init(&ctx->ctx_lock);
Kent Overstreet0460fef2013-05-07 16:18:49 -0700650 spin_lock_init(&ctx->completion_lock);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700651 mutex_init(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 init_waitqueue_head(&ctx->wait);
653
654 INIT_LIST_HEAD(&ctx->active_reqs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000656 ctx->cpu = alloc_percpu(struct kioctx_cpu);
657 if (!ctx->cpu)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700658 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000660 if (aio_setup_ring(ctx) < 0)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700661 goto err;
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000662
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000663 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000664 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
Benjamin LaHaise6878ea72013-07-31 10:34:18 -0400665 if (ctx->req_batch < 1)
666 ctx->req_batch = 1;
Kent Overstreet34e83fc2013-04-26 10:58:39 +1000667
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 /* limit the number of system wide aios */
Al Viro9fa1cb32012-03-10 23:14:05 -0500669 spin_lock(&aio_nr_lock);
Benjamin LaHaise4cd81c32013-07-30 12:06:37 -0400670 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
Al Viro2dd542b2012-03-10 23:10:35 -0500671 aio_nr + nr_events < aio_nr) {
Al Viro9fa1cb32012-03-10 23:14:05 -0500672 spin_unlock(&aio_nr_lock);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700673 err = -EAGAIN;
Gu Zhengd1b94322013-12-04 18:19:06 +0800674 goto err_ctx;
Al Viro2dd542b2012-03-10 23:10:35 -0500675 }
676 aio_nr += ctx->max_reqs;
Al Viro9fa1cb32012-03-10 23:14:05 -0500677 spin_unlock(&aio_nr_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678
Benjamin LaHaise18816862013-12-21 15:49:28 -0500679 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
680 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
Kent Overstreet723be6e2013-05-28 15:14:48 -0700681
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400682 err = ioctx_add_table(ctx, mm);
683 if (err)
Kent Overstreete34ecee2013-10-10 19:31:47 -0700684 goto err_cleanup;
Benjamin LaHaiseda903822013-08-05 13:21:43 -0400685
Kent Overstreetcaf41672013-05-07 16:18:35 -0700686 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700687 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 return ctx;
689
Kent Overstreete34ecee2013-10-10 19:31:47 -0700690err_cleanup:
691 aio_nr_sub(ctx->max_reqs);
Gu Zhengd1b94322013-12-04 18:19:06 +0800692err_ctx:
693 aio_free_ring(ctx);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700694err:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000695 free_percpu(ctx->cpu);
Kent Overstreete34ecee2013-10-10 19:31:47 -0700696 free_percpu(ctx->reqs.pcpu_count);
Kent Overstreet723be6e2013-05-28 15:14:48 -0700697 free_percpu(ctx->users.pcpu_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreetcaf41672013-05-07 16:18:35 -0700699 pr_debug("error allocating ioctx %d\n", err);
Al Viroe23754f2012-03-06 14:33:22 -0500700 return ERR_PTR(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701}
702
Kent Overstreet36f55882013-05-07 16:18:41 -0700703/* kill_ioctx
704 * Cancels all outstanding aio requests on an aio context. Used
705 * when the processes owning a context have all exited to encourage
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 * the rapid destruction of the kioctx.
707 */
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400708static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709{
Kent Overstreet36f55882013-05-07 16:18:41 -0700710 if (!atomic_xchg(&ctx->dead, 1)) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400711 struct kioctx_table *table;
712
713 spin_lock(&mm->ioctx_lock);
Artem Savkovd9b2c872013-09-09 12:29:35 -0400714 rcu_read_lock();
Benjamin LaHaise77d30b12013-08-30 11:12:50 -0400715 table = rcu_dereference(mm->ioctx_table);
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400716
717 WARN_ON(ctx != table->table[ctx->id]);
718 table->table[ctx->id] = NULL;
Artem Savkovd9b2c872013-09-09 12:29:35 -0400719 rcu_read_unlock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400720 spin_unlock(&mm->ioctx_lock);
721
Kent Overstreet723be6e2013-05-28 15:14:48 -0700722 /* percpu_ref_kill() will do the necessary call_rcu() */
723 wake_up_all(&ctx->wait);
Al Viro06af1212012-03-20 16:26:24 -0400724
Kent Overstreet36f55882013-05-07 16:18:41 -0700725 /*
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700726 * It'd be more correct to do this in free_ioctx(), after all
727 * the outstanding kiocbs have finished - but by then io_destroy
728 * has already returned, so io_setup() could potentially return
729 * -EAGAIN with no ioctxs actually in use (as far as userspace
730 * could tell).
Kent Overstreet36f55882013-05-07 16:18:41 -0700731 */
Kent Overstreete34ecee2013-10-10 19:31:47 -0700732 aio_nr_sub(ctx->max_reqs);
Kent Overstreet4fcc7122013-06-12 14:04:59 -0700733
734 if (ctx->mmap_size)
735 vm_munmap(ctx->mmap_base, ctx->mmap_size);
736
Kent Overstreet723be6e2013-05-28 15:14:48 -0700737 percpu_ref_kill(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739}
740
741/* wait_on_sync_kiocb:
742 * Waits on the given sync kiocb to complete.
743 */
Kent Overstreet57282d82013-05-13 13:42:52 -0700744ssize_t wait_on_sync_kiocb(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745{
Kent Overstreet57282d82013-05-13 13:42:52 -0700746 while (!req->ki_ctx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 set_current_state(TASK_UNINTERRUPTIBLE);
Kent Overstreet57282d82013-05-13 13:42:52 -0700748 if (req->ki_ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 break;
Jeff Moyer41d10da2007-10-16 23:27:20 -0700750 io_schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 }
752 __set_current_state(TASK_RUNNING);
Kent Overstreet57282d82013-05-13 13:42:52 -0700753 return req->ki_user_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700755EXPORT_SYMBOL(wait_on_sync_kiocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756
Kent Overstreet36f55882013-05-07 16:18:41 -0700757/*
758 * exit_aio: called when the last user of mm goes away. At this point, there is
759 * no way for any new requests to be submited or any of the io_* syscalls to be
760 * called on the context.
761 *
762 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
763 * them.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800765void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400767 struct kioctx_table *table;
Jens Axboeabf137d2008-12-09 08:11:22 +0100768 struct kioctx *ctx;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400769 unsigned i = 0;
Jens Axboeabf137d2008-12-09 08:11:22 +0100770
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400771 while (1) {
772 rcu_read_lock();
773 table = rcu_dereference(mm->ioctx_table);
774
775 do {
776 if (!table || i >= table->nr) {
777 rcu_read_unlock();
778 rcu_assign_pointer(mm->ioctx_table, NULL);
779 if (table)
780 kfree(table);
781 return;
782 }
783
784 ctx = table->table[i++];
785 } while (!ctx);
786
787 rcu_read_unlock();
788
Al Viro936af152012-04-20 21:49:41 -0400789 /*
790 * We don't need to bother with munmap() here -
791 * exit_mmap(mm) is coming and it'll unmap everything.
792 * Since aio_free_ring() uses non-zero ->mmap_size
793 * as indicator that it needs to unmap the area,
794 * just set it to 0; aio_free_ring() is the only
795 * place that uses ->mmap_size, so it's safe.
Al Viro936af152012-04-20 21:49:41 -0400796 */
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700797 ctx->mmap_size = 0;
Kent Overstreet36f55882013-05-07 16:18:41 -0700798
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400799 kill_ioctx(mm, ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 }
801}
802
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000803static void put_reqs_available(struct kioctx *ctx, unsigned nr)
804{
805 struct kioctx_cpu *kcpu;
806
807 preempt_disable();
808 kcpu = this_cpu_ptr(ctx->cpu);
809
810 kcpu->reqs_available += nr;
811 while (kcpu->reqs_available >= ctx->req_batch * 2) {
812 kcpu->reqs_available -= ctx->req_batch;
813 atomic_add(ctx->req_batch, &ctx->reqs_available);
814 }
815
816 preempt_enable();
817}
818
819static bool get_reqs_available(struct kioctx *ctx)
820{
821 struct kioctx_cpu *kcpu;
822 bool ret = false;
823
824 preempt_disable();
825 kcpu = this_cpu_ptr(ctx->cpu);
826
827 if (!kcpu->reqs_available) {
828 int old, avail = atomic_read(&ctx->reqs_available);
829
830 do {
831 if (avail < ctx->req_batch)
832 goto out;
833
834 old = avail;
835 avail = atomic_cmpxchg(&ctx->reqs_available,
836 avail, avail - ctx->req_batch);
837 } while (avail != old);
838
839 kcpu->reqs_available += ctx->req_batch;
840 }
841
842 ret = true;
843 kcpu->reqs_available--;
844out:
845 preempt_enable();
846 return ret;
847}
848
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849/* aio_get_req
Kent Overstreet57282d82013-05-13 13:42:52 -0700850 * Allocate a slot for an aio request.
851 * Returns NULL if no requests are free.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 */
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700853static inline struct kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854{
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700855 struct kiocb *req;
856
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000857 if (!get_reqs_available(ctx))
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700858 return NULL;
859
Kent Overstreet0460fef2013-05-07 16:18:49 -0700860 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 if (unlikely(!req))
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700862 goto out_put;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863
Kent Overstreete34ecee2013-10-10 19:31:47 -0700864 percpu_ref_get(&ctx->reqs);
865
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 req->ki_ctx = ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 return req;
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700868out_put:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +1000869 put_reqs_available(ctx, 1);
Kent Overstreeta1c8eae2013-05-07 16:18:53 -0700870 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871}
872
Kent Overstreet11599eb2013-05-07 16:18:39 -0700873static void kiocb_free(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874{
Kent Overstreet1d98ebf2013-05-07 16:18:37 -0700875 if (req->ki_filp)
876 fput(req->ki_filp);
Davide Libenzi13389012009-06-30 11:41:11 -0700877 if (req->ki_eventfd != NULL)
878 eventfd_ctx_put(req->ki_eventfd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 kmem_cache_free(kiocb_cachep, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880}
881
Adrian Bunkd5470b52008-04-29 00:58:57 -0700882static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883{
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400884 struct aio_ring __user *ring = (void __user *)ctx_id;
Jens Axboeabf137d2008-12-09 08:11:22 +0100885 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -0700886 struct kioctx *ctx, *ret = NULL;
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400887 struct kioctx_table *table;
888 unsigned id;
889
890 if (get_user(id, &ring->id))
891 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892
Jens Axboeabf137d2008-12-09 08:11:22 +0100893 rcu_read_lock();
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400894 table = rcu_dereference(mm->ioctx_table);
Jens Axboeabf137d2008-12-09 08:11:22 +0100895
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400896 if (!table || id >= table->nr)
897 goto out;
898
899 ctx = table->table[id];
Benjamin LaHaisef30d7042013-08-07 18:23:48 -0400900 if (ctx && ctx->user_id == ctx_id) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400901 percpu_ref_get(&ctx->users);
902 ret = ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +0100903 }
Benjamin LaHaisedb446a02013-07-30 12:54:40 -0400904out:
Jens Axboeabf137d2008-12-09 08:11:22 +0100905 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -0700906 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907}
908
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909/* aio_complete
910 * Called when the io request on the given iocb is complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 */
Kent Overstreet2d684492013-05-07 16:18:29 -0700912void aio_complete(struct kiocb *iocb, long res, long res2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913{
914 struct kioctx *ctx = iocb->ki_ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 struct aio_ring *ring;
Kent Overstreet21b40202013-05-07 16:18:47 -0700916 struct io_event *ev_page, *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 unsigned long flags;
Kent Overstreet21b40202013-05-07 16:18:47 -0700918 unsigned tail, pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919
Zach Brown20dcae32005-11-13 16:07:33 -0800920 /*
921 * Special case handling for sync iocbs:
922 * - events go directly into the iocb for fast handling
923 * - the sync task with the iocb in its stack holds the single iocb
924 * ref, no other paths have a way to get another ref
925 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 */
927 if (is_sync_kiocb(iocb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 iocb->ki_user_data = res;
Kent Overstreet57282d82013-05-13 13:42:52 -0700929 smp_wmb();
930 iocb->ki_ctx = ERR_PTR(-EXDEV);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 wake_up_process(iocb->ki_obj.tsk);
Kent Overstreet2d684492013-05-07 16:18:29 -0700932 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 }
934
Kent Overstreet0460fef2013-05-07 16:18:49 -0700935 if (iocb->ki_list.next) {
936 unsigned long flags;
937
938 spin_lock_irqsave(&ctx->ctx_lock, flags);
939 list_del(&iocb->ki_list);
940 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
941 }
Kent Overstreet11599eb2013-05-07 16:18:39 -0700942
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 /*
Kent Overstreet0460fef2013-05-07 16:18:49 -0700944 * Add a completion event to the ring buffer. Must be done holding
Tang Chen4b30f072013-07-03 15:09:16 -0700945 * ctx->completion_lock to prevent other code from messing with the tail
Kent Overstreet0460fef2013-05-07 16:18:49 -0700946 * pointer since we might be called from irq context.
947 */
948 spin_lock_irqsave(&ctx->completion_lock, flags);
949
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700950 tail = ctx->tail;
Kent Overstreet21b40202013-05-07 16:18:47 -0700951 pos = tail + AIO_EVENTS_OFFSET;
952
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700953 if (++tail >= ctx->nr_events)
Ken Chen4bf69b22005-05-01 08:59:15 -0700954 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700956 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -0700957 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
958
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
960 event->data = iocb->ki_user_data;
961 event->res = res;
962 event->res2 = res2;
963
Kent Overstreet21b40202013-05-07 16:18:47 -0700964 kunmap_atomic(ev_page);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700965 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet21b40202013-05-07 16:18:47 -0700966
967 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
Kent Overstreetcaf41672013-05-07 16:18:35 -0700968 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
969 res, res2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970
971 /* after flagging the request as done, we
972 * must never even look at it again
973 */
974 smp_wmb(); /* make event visible before updating tail */
975
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700976 ctx->tail = tail;
Kent Overstreet21b40202013-05-07 16:18:47 -0700977
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700978 ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979 ring->tail = tail;
Cong Wange8e3c3d2011-11-25 23:14:27 +0800980 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -0700981 flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982
Kent Overstreet0460fef2013-05-07 16:18:49 -0700983 spin_unlock_irqrestore(&ctx->completion_lock, flags);
984
Kent Overstreet21b40202013-05-07 16:18:47 -0700985 pr_debug("added to ring %p at [%u]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -0700986
987 /*
988 * Check if the user asked us to deliver the result through an
989 * eventfd. The eventfd_signal() function is safe to be called
990 * from IRQ context.
991 */
Davide Libenzi87c3a862009-03-18 17:04:19 -0700992 if (iocb->ki_eventfd != NULL)
Davide Libenzi8d1c98b2008-04-10 21:29:19 -0700993 eventfd_signal(iocb->ki_eventfd, 1);
994
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995 /* everything turned out well, dispose of the aiocb. */
Kent Overstreet57282d82013-05-13 13:42:52 -0700996 kiocb_free(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997
Quentin Barnes6cb2a212008-03-19 17:00:39 -0700998 /*
999 * We have to order our ring_info tail store above and test
1000 * of the wait list below outside the wait lock. This is
1001 * like in wake_up_bit() where clearing a bit has to be
1002 * ordered with the unlocked test.
1003 */
1004 smp_mb();
1005
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006 if (waitqueue_active(&ctx->wait))
1007 wake_up(&ctx->wait);
1008
Kent Overstreete34ecee2013-10-10 19:31:47 -07001009 percpu_ref_put(&ctx->reqs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010}
H Hartley Sweeten385773e2009-09-22 16:43:53 -07001011EXPORT_SYMBOL(aio_complete);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012
Kent Overstreeta31ad382013-05-07 16:18:45 -07001013/* aio_read_events
1014 * Pull an event off of the ioctx's event ring. Returns the number of
1015 * events fetched
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016 */
Kent Overstreeta31ad382013-05-07 16:18:45 -07001017static long aio_read_events_ring(struct kioctx *ctx,
1018 struct io_event __user *event, long nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020 struct aio_ring *ring;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001021 unsigned head, tail, pos;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001022 long ret = 0;
1023 int copy_ret;
1024
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001025 mutex_lock(&ctx->ring_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001027 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001028 head = ring->head;
Kent Overstreet5ffac122013-05-09 15:36:07 -07001029 tail = ring->tail;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001030 kunmap_atomic(ring);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031
Kent Overstreet5ffac122013-05-09 15:36:07 -07001032 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001033
Kent Overstreet5ffac122013-05-09 15:36:07 -07001034 if (head == tail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035 goto out;
1036
Kent Overstreeta31ad382013-05-07 16:18:45 -07001037 while (ret < nr) {
1038 long avail;
1039 struct io_event *ev;
1040 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041
Kent Overstreet5ffac122013-05-09 15:36:07 -07001042 avail = (head <= tail ? tail : ctx->nr_events) - head;
1043 if (head == tail)
Kent Overstreeta31ad382013-05-07 16:18:45 -07001044 break;
1045
1046 avail = min(avail, nr - ret);
1047 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1048 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1049
1050 pos = head + AIO_EVENTS_OFFSET;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001051 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
Kent Overstreeta31ad382013-05-07 16:18:45 -07001052 pos %= AIO_EVENTS_PER_PAGE;
1053
1054 ev = kmap(page);
1055 copy_ret = copy_to_user(event + ret, ev + pos,
1056 sizeof(*ev) * avail);
1057 kunmap(page);
1058
1059 if (unlikely(copy_ret)) {
1060 ret = -EFAULT;
1061 goto out;
1062 }
1063
1064 ret += avail;
1065 head += avail;
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001066 head %= ctx->nr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001069 ring = kmap_atomic(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001070 ring->head = head;
Zhao Hongjiang91d80a82013-04-26 11:03:53 +08001071 kunmap_atomic(ring);
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001072 flush_dcache_page(ctx->ring_pages[0]);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001073
Kent Overstreet5ffac122013-05-09 15:36:07 -07001074 pr_debug("%li h%u t%u\n", ret, head, tail);
Kent Overstreet3e845ce2013-05-07 16:18:51 -07001075
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001076 put_reqs_available(ctx, ret);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001077out:
Kent Overstreet58c85dc2013-05-07 16:18:55 -07001078 mutex_unlock(&ctx->ring_lock);
Kent Overstreeta31ad382013-05-07 16:18:45 -07001079
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 return ret;
1081}
1082
Kent Overstreeta31ad382013-05-07 16:18:45 -07001083static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1084 struct io_event __user *event, long *i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001086 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087
Kent Overstreeta31ad382013-05-07 16:18:45 -07001088 if (ret > 0)
1089 *i += ret;
1090
1091 if (unlikely(atomic_read(&ctx->dead)))
1092 ret = -EINVAL;
1093
1094 if (!*i)
1095 *i = ret;
1096
1097 return ret < 0 || *i >= min_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098}
1099
Kent Overstreeta31ad382013-05-07 16:18:45 -07001100static long read_events(struct kioctx *ctx, long min_nr, long nr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 struct io_event __user *event,
1102 struct timespec __user *timeout)
1103{
Kent Overstreeta31ad382013-05-07 16:18:45 -07001104 ktime_t until = { .tv64 = KTIME_MAX };
1105 long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 if (timeout) {
1108 struct timespec ts;
Kent Overstreeta31ad382013-05-07 16:18:45 -07001109
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
Kent Overstreeta31ad382013-05-07 16:18:45 -07001111 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112
Kent Overstreeta31ad382013-05-07 16:18:45 -07001113 until = timespec_to_ktime(ts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114 }
1115
Kent Overstreeta31ad382013-05-07 16:18:45 -07001116 /*
1117 * Note that aio_read_events() is being called as the conditional - i.e.
1118 * we're calling it after prepare_to_wait() has set task state to
1119 * TASK_INTERRUPTIBLE.
1120 *
1121 * But aio_read_events() can block, and if it blocks it's going to flip
1122 * the task state back to TASK_RUNNING.
1123 *
1124 * This should be ok, provided it doesn't flip the state back to
1125 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1126 * will only happen if the mutex_lock() call blocks, and we then find
1127 * the ringbuffer empty. So in practice we should be ok, but it's
1128 * something to be aware of when touching this code.
1129 */
1130 wait_event_interruptible_hrtimeout(ctx->wait,
1131 aio_read_events(ctx, min_nr, nr, event, &ret), until);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132
Kent Overstreeta31ad382013-05-07 16:18:45 -07001133 if (!ret && signal_pending(current))
1134 ret = -EINTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135
Kent Overstreeta31ad382013-05-07 16:18:45 -07001136 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137}
1138
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139/* sys_io_setup:
1140 * Create an aio_context capable of receiving at least nr_events.
1141 * ctxp must not point to an aio_context that already exists, and
1142 * must be initialized to 0 prior to the call. On successful
1143 * creation of the aio_context, *ctxp is filled in with the resulting
1144 * handle. May fail with -EINVAL if *ctxp is not initialized,
1145 * if the specified nr_events exceeds internal limits. May fail
1146 * with -EAGAIN if the specified nr_events exceeds the user's limit
1147 * of available events. May fail with -ENOMEM if insufficient kernel
1148 * resources are available. May fail with -EFAULT if an invalid
1149 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1150 * implemented.
1151 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001152SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153{
1154 struct kioctx *ioctx = NULL;
1155 unsigned long ctx;
1156 long ret;
1157
1158 ret = get_user(ctx, ctxp);
1159 if (unlikely(ret))
1160 goto out;
1161
1162 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001163 if (unlikely(ctx || nr_events == 0)) {
1164 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1165 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 goto out;
1167 }
1168
1169 ioctx = ioctx_alloc(nr_events);
1170 ret = PTR_ERR(ioctx);
1171 if (!IS_ERR(ioctx)) {
1172 ret = put_user(ioctx->user_id, ctxp);
Al Viroa2e18592012-03-20 16:27:57 -04001173 if (ret)
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001174 kill_ioctx(current->mm, ioctx);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001175 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 }
1177
1178out:
1179 return ret;
1180}
1181
1182/* sys_io_destroy:
1183 * Destroy the aio_context specified. May cancel any outstanding
1184 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001185 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 * is invalid.
1187 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001188SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189{
1190 struct kioctx *ioctx = lookup_ioctx(ctx);
1191 if (likely(NULL != ioctx)) {
Benjamin LaHaisedb446a02013-07-30 12:54:40 -04001192 kill_ioctx(current->mm, ioctx);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001193 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 return 0;
1195 }
1196 pr_debug("EINVAL: io_destroy: invalid context id\n");
1197 return -EINVAL;
1198}
1199
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001200typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1201 unsigned long, loff_t);
1202
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001203static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1204 int rw, char __user *buf,
1205 unsigned long *nr_segs,
1206 struct iovec **iovec,
1207 bool compat)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001208{
1209 ssize_t ret;
1210
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001211 *nr_segs = kiocb->ki_nbytes;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001212
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001213#ifdef CONFIG_COMPAT
1214 if (compat)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001215 ret = compat_rw_copy_check_uvector(rw,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001216 (struct compat_iovec __user *)buf,
1217 *nr_segs, 1, *iovec, iovec);
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001218 else
1219#endif
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001220 ret = rw_copy_check_uvector(rw,
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001221 (struct iovec __user *)buf,
1222 *nr_segs, 1, *iovec, iovec);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001223 if (ret < 0)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001224 return ret;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001225
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001226 /* ki_nbytes now reflect bytes instead of segs */
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001227 kiocb->ki_nbytes = ret;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001228 return 0;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001229}
1230
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001231static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1232 int rw, char __user *buf,
1233 unsigned long *nr_segs,
1234 struct iovec *iovec)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001235{
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001236 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001237 return -EFAULT;
Linus Torvaldsa70b52e2012-05-21 16:06:20 -07001238
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001239 iovec->iov_base = buf;
1240 iovec->iov_len = kiocb->ki_nbytes;
1241 *nr_segs = 1;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001242 return 0;
1243}
1244
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245/*
1246 * aio_setup_iocb:
1247 * Performs the initial checks and aio retry method
1248 * setup for the kiocb at the time of io submission.
1249 */
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001250static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1251 char __user *buf, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252{
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001253 struct file *file = req->ki_filp;
1254 ssize_t ret;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001255 unsigned long nr_segs;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001256 int rw;
1257 fmode_t mode;
1258 aio_rw_op *rw_op;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001259 struct iovec inline_vec, *iovec = &inline_vec;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001261 switch (opcode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262 case IOCB_CMD_PREAD:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001263 case IOCB_CMD_PREADV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001264 mode = FMODE_READ;
1265 rw = READ;
1266 rw_op = file->f_op->aio_read;
1267 goto rw_common;
1268
1269 case IOCB_CMD_PWRITE:
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001270 case IOCB_CMD_PWRITEV:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001271 mode = FMODE_WRITE;
1272 rw = WRITE;
1273 rw_op = file->f_op->aio_write;
1274 goto rw_common;
1275rw_common:
1276 if (unlikely(!(file->f_mode & mode)))
1277 return -EBADF;
1278
1279 if (!rw_op)
1280 return -EINVAL;
1281
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001282 ret = (opcode == IOCB_CMD_PREADV ||
1283 opcode == IOCB_CMD_PWRITEV)
1284 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1285 &iovec, compat)
1286 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1287 iovec);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001288 if (ret)
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001289 return ret;
1290
1291 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001292 if (ret < 0) {
1293 if (iovec != &inline_vec)
1294 kfree(iovec);
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001295 return ret;
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001296 }
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001297
1298 req->ki_nbytes = ret;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001299
Kent Overstreet73a70752013-05-09 15:03:42 -07001300 /* XXX: move/kill - rw_verify_area()? */
1301 /* This matches the pread()/pwrite() logic */
1302 if (req->ki_pos < 0) {
1303 ret = -EINVAL;
1304 break;
1305 }
1306
1307 if (rw == WRITE)
1308 file_start_write(file);
1309
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001310 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
Kent Overstreet73a70752013-05-09 15:03:42 -07001311
1312 if (rw == WRITE)
1313 file_end_write(file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001315
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316 case IOCB_CMD_FDSYNC:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001317 if (!file->f_op->aio_fsync)
1318 return -EINVAL;
1319
1320 ret = file->f_op->aio_fsync(req, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001322
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 case IOCB_CMD_FSYNC:
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001324 if (!file->f_op->aio_fsync)
1325 return -EINVAL;
1326
1327 ret = file->f_op->aio_fsync(req, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 break;
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001329
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 default:
Kent Overstreetcaf41672013-05-07 16:18:35 -07001331 pr_debug("EINVAL: no operation provided\n");
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001332 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 }
1334
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001335 if (iovec != &inline_vec)
1336 kfree(iovec);
1337
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001338 if (ret != -EIOCBQUEUED) {
1339 /*
1340 * There's no easy way to restart the syscall since other AIO's
1341 * may be already running. Just fail this IO with EINTR.
1342 */
1343 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1344 ret == -ERESTARTNOHAND ||
1345 ret == -ERESTART_RESTARTBLOCK))
1346 ret = -EINTR;
1347 aio_complete(req, ret, 0);
1348 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349
1350 return 0;
1351}
1352
Adrian Bunkd5470b52008-04-29 00:58:57 -07001353static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001354 struct iocb *iocb, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355{
1356 struct kiocb *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 ssize_t ret;
1358
1359 /* enforce forwards compatibility on users */
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001360 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001361 pr_debug("EINVAL: reserve field set\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 return -EINVAL;
1363 }
1364
1365 /* prevent overflows */
1366 if (unlikely(
1367 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1368 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1369 ((ssize_t)iocb->aio_nbytes < 0)
1370 )) {
1371 pr_debug("EINVAL: io_submit: overflow check\n");
1372 return -EINVAL;
1373 }
1374
Kent Overstreet41ef4eb2013-05-07 16:19:11 -07001375 req = aio_get_req(ctx);
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001376 if (unlikely(!req))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 return -EAGAIN;
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001378
1379 req->ki_filp = fget(iocb->aio_fildes);
1380 if (unlikely(!req->ki_filp)) {
1381 ret = -EBADF;
1382 goto out_put_req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 }
Kent Overstreet1d98ebf2013-05-07 16:18:37 -07001384
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001385 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1386 /*
1387 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1388 * instance of the file* now. The file descriptor must be
1389 * an eventfd() fd, and will be signaled for each completed
1390 * event using the eventfd_signal() function.
1391 */
Davide Libenzi13389012009-06-30 11:41:11 -07001392 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
Hirofumi Nakagawa801678c2008-04-29 01:03:09 -07001393 if (IS_ERR(req->ki_eventfd)) {
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001394 ret = PTR_ERR(req->ki_eventfd);
Davide Libenzi87c3a862009-03-18 17:04:19 -07001395 req->ki_eventfd = NULL;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001396 goto out_put_req;
1397 }
1398 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399
Kent Overstreet8a660892013-05-07 16:19:10 -07001400 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401 if (unlikely(ret)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001402 pr_debug("EFAULT: aio_key\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403 goto out_put_req;
1404 }
1405
1406 req->ki_obj.user = user_iocb;
1407 req->ki_user_data = iocb->aio_data;
1408 req->ki_pos = iocb->aio_offset;
Kent Overstreet73a70752013-05-09 15:03:42 -07001409 req->ki_nbytes = iocb->aio_nbytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410
Kent Overstreet8bc92af2013-02-25 16:36:27 -08001411 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1412 (char __user *)(unsigned long)iocb->aio_buf,
1413 compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 if (ret)
1415 goto out_put_req;
1416
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418out_put_req:
Kent Overstreete1bdd5f2013-04-26 10:58:39 +10001419 put_reqs_available(ctx, 1);
Kent Overstreete34ecee2013-10-10 19:31:47 -07001420 percpu_ref_put(&ctx->reqs);
Kent Overstreet57282d82013-05-13 13:42:52 -07001421 kiocb_free(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 return ret;
1423}
1424
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001425long do_io_submit(aio_context_t ctx_id, long nr,
1426 struct iocb __user *__user *iocbpp, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427{
1428 struct kioctx *ctx;
1429 long ret = 0;
Jeff Moyer080d6762011-11-02 13:40:10 -07001430 int i = 0;
Shaohua Li9f5b9422010-07-01 07:55:01 +02001431 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432
1433 if (unlikely(nr < 0))
1434 return -EINVAL;
1435
Jeff Moyer75e1c702010-09-10 14:16:00 -07001436 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1437 nr = LONG_MAX/sizeof(*iocbpp);
1438
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1440 return -EFAULT;
1441
1442 ctx = lookup_ioctx(ctx_id);
1443 if (unlikely(!ctx)) {
Kent Overstreetcaf41672013-05-07 16:18:35 -07001444 pr_debug("EINVAL: invalid context id\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445 return -EINVAL;
1446 }
1447
Shaohua Li9f5b9422010-07-01 07:55:01 +02001448 blk_start_plug(&plug);
1449
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 /*
1451 * AKPM: should this return a partial result if some of the IOs were
1452 * successfully submitted?
1453 */
1454 for (i=0; i<nr; i++) {
1455 struct iocb __user *user_iocb;
1456 struct iocb tmp;
1457
1458 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1459 ret = -EFAULT;
1460 break;
1461 }
1462
1463 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1464 ret = -EFAULT;
1465 break;
1466 }
1467
Kent Overstreeta1c8eae2013-05-07 16:18:53 -07001468 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 if (ret)
1470 break;
1471 }
Shaohua Li9f5b9422010-07-01 07:55:01 +02001472 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473
Kent Overstreet723be6e2013-05-28 15:14:48 -07001474 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 return i ? i : ret;
1476}
1477
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001478/* sys_io_submit:
1479 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1480 * the number of iocbs queued. May return -EINVAL if the aio_context
1481 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1482 * *iocbpp[0] is not properly initialized, if the operation specified
1483 * is invalid for the file descriptor in the iocb. May fail with
1484 * -EFAULT if any of the data structures point to invalid data. May
1485 * fail with -EBADF if the file descriptor specified in the first
1486 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1487 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1488 * fail with -ENOSYS if not implemented.
1489 */
1490SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1491 struct iocb __user * __user *, iocbpp)
1492{
1493 return do_io_submit(ctx_id, nr, iocbpp, 0);
1494}
1495
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496/* lookup_kiocb
1497 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001499static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1500 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501{
1502 struct list_head *pos;
Zach Brownd00689a2005-11-13 16:07:34 -08001503
1504 assert_spin_locked(&ctx->ctx_lock);
1505
Kent Overstreet8a660892013-05-07 16:19:10 -07001506 if (key != KIOCB_KEY)
1507 return NULL;
1508
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 /* TODO: use a hash or array, this sucks. */
1510 list_for_each(pos, &ctx->active_reqs) {
1511 struct kiocb *kiocb = list_kiocb(pos);
Kent Overstreet8a660892013-05-07 16:19:10 -07001512 if (kiocb->ki_obj.user == iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513 return kiocb;
1514 }
1515 return NULL;
1516}
1517
1518/* sys_io_cancel:
1519 * Attempts to cancel an iocb previously passed to io_submit. If
1520 * the operation is successfully cancelled, the resulting event is
1521 * copied into the memory pointed to by result without being placed
1522 * into the completion queue and 0 is returned. May fail with
1523 * -EFAULT if any of the data structures pointed to are invalid.
1524 * May fail with -EINVAL if aio_context specified by ctx_id is
1525 * invalid. May fail with -EAGAIN if the iocb specified was not
1526 * cancelled. Will fail with -ENOSYS if not implemented.
1527 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001528SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1529 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 struct kioctx *ctx;
1532 struct kiocb *kiocb;
1533 u32 key;
1534 int ret;
1535
1536 ret = get_user(key, &iocb->aio_key);
1537 if (unlikely(ret))
1538 return -EFAULT;
1539
1540 ctx = lookup_ioctx(ctx_id);
1541 if (unlikely(!ctx))
1542 return -EINVAL;
1543
1544 spin_lock_irq(&ctx->ctx_lock);
Kent Overstreet906b9732013-05-07 16:18:31 -07001545
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546 kiocb = lookup_kiocb(ctx, iocb, key);
Kent Overstreet906b9732013-05-07 16:18:31 -07001547 if (kiocb)
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001548 ret = kiocb_cancel(ctx, kiocb);
Kent Overstreet906b9732013-05-07 16:18:31 -07001549 else
1550 ret = -EINVAL;
1551
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552 spin_unlock_irq(&ctx->ctx_lock);
1553
Kent Overstreet906b9732013-05-07 16:18:31 -07001554 if (!ret) {
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001555 /*
1556 * The result argument is no longer used - the io_event is
1557 * always delivered via the ring buffer. -EINPROGRESS indicates
1558 * cancellation is progress:
Kent Overstreet906b9732013-05-07 16:18:31 -07001559 */
Kent Overstreetbec68faa2013-05-13 14:45:08 -07001560 ret = -EINPROGRESS;
Kent Overstreet906b9732013-05-07 16:18:31 -07001561 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562
Kent Overstreet723be6e2013-05-28 15:14:48 -07001563 percpu_ref_put(&ctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564
1565 return ret;
1566}
1567
1568/* io_getevents:
1569 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001570 * the completion queue for the aio_context specified by ctx_id. If
1571 * it succeeds, the number of read events is returned. May fail with
1572 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1573 * out of range, if timeout is out of range. May fail with -EFAULT
1574 * if any of the memory specified is invalid. May return 0 or
1575 * < min_nr if the timeout specified by timeout has elapsed
1576 * before sufficient events are available, where timeout == NULL
1577 * specifies an infinite timeout. Note that the timeout pointed to by
Jeff Moyer69008072013-05-24 15:55:24 -07001578 * timeout is relative. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001580SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1581 long, min_nr,
1582 long, nr,
1583 struct io_event __user *, events,
1584 struct timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585{
1586 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1587 long ret = -EINVAL;
1588
1589 if (likely(ioctx)) {
Namhyung Kim2e410252011-01-12 17:01:08 -08001590 if (likely(min_nr <= nr && min_nr >= 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591 ret = read_events(ioctx, min_nr, nr, events, timeout);
Kent Overstreet723be6e2013-05-28 15:14:48 -07001592 percpu_ref_put(&ioctx->users);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594 return ret;
1595}