1 /* 1 /*
2 * An async IO implementation for Linux 2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvac 3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 * 4 *
5 * Implements an efficient asynchronous i 5 * Implements an efficient asynchronous io interface.
6 * 6 *
7 * Copyright 2000, 2001, 2002 Red Hat, In 7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 * Copyright 2018 Christoph Hellwig. <<
9 * 8 *
10 * See ../COPYING for licensing terms. 9 * See ../COPYING for licensing terms.
11 */ 10 */
12 #define pr_fmt(fmt) "%s: " fmt, __func__ 11 #define pr_fmt(fmt) "%s: " fmt, __func__
13 12
14 #include <linux/kernel.h> 13 #include <linux/kernel.h>
15 #include <linux/init.h> 14 #include <linux/init.h>
16 #include <linux/errno.h> 15 #include <linux/errno.h>
17 #include <linux/time.h> 16 #include <linux/time.h>
18 #include <linux/aio_abi.h> 17 #include <linux/aio_abi.h>
19 #include <linux/export.h> 18 #include <linux/export.h>
20 #include <linux/syscalls.h> 19 #include <linux/syscalls.h>
21 #include <linux/backing-dev.h> 20 #include <linux/backing-dev.h>
22 #include <linux/refcount.h> <<
23 #include <linux/uio.h> 21 #include <linux/uio.h>
24 22
25 #include <linux/sched/signal.h> 23 #include <linux/sched/signal.h>
26 #include <linux/fs.h> 24 #include <linux/fs.h>
27 #include <linux/file.h> 25 #include <linux/file.h>
28 #include <linux/mm.h> 26 #include <linux/mm.h>
29 #include <linux/mman.h> 27 #include <linux/mman.h>
>> 28 #include <linux/mmu_context.h>
30 #include <linux/percpu.h> 29 #include <linux/percpu.h>
31 #include <linux/slab.h> 30 #include <linux/slab.h>
32 #include <linux/timer.h> 31 #include <linux/timer.h>
33 #include <linux/aio.h> 32 #include <linux/aio.h>
34 #include <linux/highmem.h> 33 #include <linux/highmem.h>
35 #include <linux/workqueue.h> 34 #include <linux/workqueue.h>
36 #include <linux/security.h> 35 #include <linux/security.h>
37 #include <linux/eventfd.h> 36 #include <linux/eventfd.h>
38 #include <linux/blkdev.h> 37 #include <linux/blkdev.h>
39 #include <linux/compat.h> 38 #include <linux/compat.h>
40 #include <linux/migrate.h> 39 #include <linux/migrate.h>
41 #include <linux/ramfs.h> 40 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.h> 41 #include <linux/percpu-refcount.h>
43 #include <linux/mount.h> 42 #include <linux/mount.h>
44 #include <linux/pseudo_fs.h> <<
45 43
>> 44 #include <asm/kmap_types.h>
46 #include <linux/uaccess.h> 45 #include <linux/uaccess.h>
47 #include <linux/nospec.h> <<
48 46
49 #include "internal.h" 47 #include "internal.h"
50 48
51 #define KIOCB_KEY 0 <<
52 <<
53 #define AIO_RING_MAGIC 0xa10a 49 #define AIO_RING_MAGIC 0xa10a10a1
54 #define AIO_RING_COMPAT_FEATURES 1 50 #define AIO_RING_COMPAT_FEATURES 1
55 #define AIO_RING_INCOMPAT_FEATURES 0 51 #define AIO_RING_INCOMPAT_FEATURES 0
56 struct aio_ring { 52 struct aio_ring {
57 unsigned id; /* kernel inte 53 unsigned id; /* kernel internal index number */
58 unsigned nr; /* number of i 54 unsigned nr; /* number of io_events */
59 unsigned head; /* Written to 55 unsigned head; /* Written to by userland or under ring_lock
60 * mutex by ai 56 * mutex by aio_read_events_ring(). */
61 unsigned tail; 57 unsigned tail;
62 58
63 unsigned magic; 59 unsigned magic;
64 unsigned compat_features; 60 unsigned compat_features;
65 unsigned incompat_features; 61 unsigned incompat_features;
66 unsigned header_length; /* siz 62 unsigned header_length; /* size of aio_ring */
67 63
68 64
69 struct io_event io_events[]; !! 65 struct io_event io_events[0];
70 }; /* 128 bytes + ring size */ 66 }; /* 128 bytes + ring size */
71 67
72 /* <<
73 * Plugging is meant to work with larger batch <<
74 * have more than the below, then don't bother <<
75 */ <<
76 #define AIO_PLUG_THRESHOLD 2 <<
77 <<
78 #define AIO_RING_PAGES 8 68 #define AIO_RING_PAGES 8
79 69
80 struct kioctx_table { 70 struct kioctx_table {
81 struct rcu_head rcu; !! 71 struct rcu_head rcu;
82 unsigned nr; !! 72 unsigned nr;
83 struct kioctx __rcu *table[] __cou !! 73 struct kioctx *table[];
84 }; 74 };
85 75
86 struct kioctx_cpu { 76 struct kioctx_cpu {
87 unsigned reqs_available 77 unsigned reqs_available;
88 }; 78 };
89 79
90 struct ctx_rq_wait { 80 struct ctx_rq_wait {
91 struct completion comp; 81 struct completion comp;
92 atomic_t count; 82 atomic_t count;
93 }; 83 };
94 84
95 struct kioctx { 85 struct kioctx {
96 struct percpu_ref users; 86 struct percpu_ref users;
97 atomic_t dead; 87 atomic_t dead;
98 88
99 struct percpu_ref reqs; 89 struct percpu_ref reqs;
100 90
101 unsigned long user_id; 91 unsigned long user_id;
102 92
103 struct kioctx_cpu __percpu *cpu; !! 93 struct __percpu kioctx_cpu *cpu;
104 94
105 /* 95 /*
106 * For percpu reqs_available, number o 96 * For percpu reqs_available, number of slots we move to/from global
107 * counter at a time: 97 * counter at a time:
108 */ 98 */
109 unsigned req_batch; 99 unsigned req_batch;
110 /* 100 /*
111 * This is what userspace passed to io 101 * This is what userspace passed to io_setup(), it's not used for
112 * anything but counting against the g 102 * anything but counting against the global max_reqs quota.
113 * 103 *
114 * The real limit is nr_events - 1, wh 104 * The real limit is nr_events - 1, which will be larger (see
115 * aio_setup_ring()) 105 * aio_setup_ring())
116 */ 106 */
117 unsigned max_reqs; 107 unsigned max_reqs;
118 108
119 /* Size of ringbuffer, in units of str 109 /* Size of ringbuffer, in units of struct io_event */
120 unsigned nr_events; 110 unsigned nr_events;
121 111
122 unsigned long mmap_base; 112 unsigned long mmap_base;
123 unsigned long mmap_size; 113 unsigned long mmap_size;
124 114
125 struct folio **ring_folios; !! 115 struct page **ring_pages;
126 long nr_pages; 116 long nr_pages;
127 117
128 struct rcu_work free_rwork; !! 118 struct work_struct free_work;
129 119
130 /* 120 /*
131 * signals when all in-flight requests 121 * signals when all in-flight requests are done
132 */ 122 */
133 struct ctx_rq_wait *rq_wait; 123 struct ctx_rq_wait *rq_wait;
134 124
135 struct { 125 struct {
136 /* 126 /*
137 * This counts the number of a 127 * This counts the number of available slots in the ringbuffer,
138 * so we avoid overflowing it: 128 * so we avoid overflowing it: it's decremented (if positive)
139 * when allocating a kiocb and 129 * when allocating a kiocb and incremented when the resulting
140 * io_event is pulled off the 130 * io_event is pulled off the ringbuffer.
141 * 131 *
142 * We batch accesses to it wit 132 * We batch accesses to it with a percpu version.
143 */ 133 */
144 atomic_t reqs_available 134 atomic_t reqs_available;
145 } ____cacheline_aligned_in_smp; 135 } ____cacheline_aligned_in_smp;
146 136
147 struct { 137 struct {
148 spinlock_t ctx_lock; 138 spinlock_t ctx_lock;
149 struct list_head active_reqs; 139 struct list_head active_reqs; /* used for cancellation */
150 } ____cacheline_aligned_in_smp; 140 } ____cacheline_aligned_in_smp;
151 141
152 struct { 142 struct {
153 struct mutex ring_lock; 143 struct mutex ring_lock;
154 wait_queue_head_t wait; 144 wait_queue_head_t wait;
155 } ____cacheline_aligned_in_smp; 145 } ____cacheline_aligned_in_smp;
156 146
157 struct { 147 struct {
158 unsigned tail; 148 unsigned tail;
159 unsigned completed_even 149 unsigned completed_events;
160 spinlock_t completion_loc 150 spinlock_t completion_lock;
161 } ____cacheline_aligned_in_smp; 151 } ____cacheline_aligned_in_smp;
162 152
163 struct folio *internal_foli !! 153 struct page *internal_pages[AIO_RING_PAGES];
164 struct file *aio_ring_file 154 struct file *aio_ring_file;
165 155
166 unsigned id; 156 unsigned id;
167 }; 157 };
168 158
169 /* 159 /*
170 * First field must be the file pointer in all !! 160 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
171 * iocb unions! See also 'struct kiocb' in <li !! 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().
172 */ 169 */
173 struct fsync_iocb { !! 170 #define KIOCB_CANCELLED ((void *) (~0ULL))
174 struct file *file; <<
175 struct work_struct work; <<
176 bool datasync; <<
177 struct cred *creds; <<
178 }; <<
179 <<
180 struct poll_iocb { <<
181 struct file *file; <<
182 struct wait_queue_head *head; <<
183 __poll_t events; <<
184 bool cancelled; <<
185 bool work_scheduled <<
186 bool work_need_resc <<
187 struct wait_queue_entry wait; <<
188 struct work_struct work; <<
189 }; <<
190 171
191 /* <<
192 * NOTE! Each of the iocb union members has th <<
193 * as the first entry in their struct definiti <<
194 * access the file pointer through any of the <<
195 * or directly as just 'ki_filp' in this struc <<
196 */ <<
197 struct aio_kiocb { 172 struct aio_kiocb {
198 union { !! 173 struct kiocb common;
199 struct file *ki_fi <<
200 struct kiocb rw; <<
201 struct fsync_iocb fsync; <<
202 struct poll_iocb poll; <<
203 }; <<
204 174
205 struct kioctx *ki_ctx; 175 struct kioctx *ki_ctx;
206 kiocb_cancel_fn *ki_cancel; 176 kiocb_cancel_fn *ki_cancel;
207 177
208 struct io_event ki_res; !! 178 struct iocb __user *ki_user_iocb; /* user's aiocb */
>> 179 __u64 ki_user_data; /* user's data for completion */
209 180
210 struct list_head ki_list; 181 struct list_head ki_list; /* the aio core uses this
211 182 * for cancellation */
212 refcount_t ki_refcnt; <<
213 183
214 /* 184 /*
215 * If the aio_resfd field of the users 185 * If the aio_resfd field of the userspace iocb is not zero,
216 * this is the underlying eventfd cont 186 * this is the underlying eventfd context to deliver events to.
217 */ 187 */
218 struct eventfd_ctx *ki_eventfd; 188 struct eventfd_ctx *ki_eventfd;
219 }; 189 };
220 190
221 /*------ sysctl variables----*/ 191 /*------ sysctl variables----*/
222 static DEFINE_SPINLOCK(aio_nr_lock); 192 static DEFINE_SPINLOCK(aio_nr_lock);
223 static unsigned long aio_nr; /* cur !! 193 unsigned long aio_nr; /* current system wide number of aio requests */
224 static unsigned long aio_max_nr = 0x10000; /* !! 194 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/ 195 /*----end sysctl variables---*/
226 #ifdef CONFIG_SYSCTL <<
227 static struct ctl_table aio_sysctls[] = { <<
228 { <<
229 .procname = "aio-nr", <<
230 .data = &aio_nr, <<
231 .maxlen = sizeof(aio_n <<
232 .mode = 0444, <<
233 .proc_handler = proc_doulong <<
234 }, <<
235 { <<
236 .procname = "aio-max-nr" <<
237 .data = &aio_max_nr, <<
238 .maxlen = sizeof(aio_m <<
239 .mode = 0644, <<
240 .proc_handler = proc_doulong <<
241 }, <<
242 }; <<
243 <<
244 static void __init aio_sysctl_init(void) <<
245 { <<
246 register_sysctl_init("fs", aio_sysctls <<
247 } <<
248 #else <<
249 #define aio_sysctl_init() do { } while (0) <<
250 #endif <<
251 196
252 static struct kmem_cache *kiocb_cachep; 197 static struct kmem_cache *kiocb_cachep;
253 static struct kmem_cache *kioctx_cachep 198 static struct kmem_cache *kioctx_cachep;
254 199
255 static struct vfsmount *aio_mnt; 200 static struct vfsmount *aio_mnt;
256 201
257 static const struct file_operations aio_ring_f 202 static const struct file_operations aio_ring_fops;
258 static const struct address_space_operations a 203 static const struct address_space_operations aio_ctx_aops;
259 204
260 static struct file *aio_private_file(struct ki 205 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
261 { 206 {
>> 207 struct qstr this = QSTR_INIT("[aio]", 5);
262 struct file *file; 208 struct file *file;
>> 209 struct path path;
263 struct inode *inode = alloc_anon_inode 210 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
264 if (IS_ERR(inode)) 211 if (IS_ERR(inode))
265 return ERR_CAST(inode); 212 return ERR_CAST(inode);
266 213
267 inode->i_mapping->a_ops = &aio_ctx_aop 214 inode->i_mapping->a_ops = &aio_ctx_aops;
268 inode->i_mapping->i_private_data = ctx !! 215 inode->i_mapping->private_data = ctx;
269 inode->i_size = PAGE_SIZE * nr_pages; 216 inode->i_size = PAGE_SIZE * nr_pages;
270 217
271 file = alloc_file_pseudo(inode, aio_mn !! 218 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
272 O_RDWR, &aio_r !! 219 if (!path.dentry) {
273 if (IS_ERR(file)) <<
274 iput(inode); 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;
275 return file; 233 return file;
276 } 234 }
277 235
278 static int aio_init_fs_context(struct fs_conte !! 236 static struct dentry *aio_mount(struct file_system_type *fs_type,
>> 237 int flags, const char *dev_name, void *data)
279 { 238 {
280 if (!init_pseudo(fc, AIO_RING_MAGIC)) !! 239 static const struct dentry_operations ops = {
281 return -ENOMEM; !! 240 .d_dname = simple_dname,
282 fc->s_iflags |= SB_I_NOEXEC; !! 241 };
283 return 0; !! 242 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;
284 } 248 }
285 249
286 /* aio_setup 250 /* aio_setup
287 * Creates the slab caches used by the ai 251 * Creates the slab caches used by the aio routines, panic on
288 * failure as this is done early during t 252 * failure as this is done early during the boot sequence.
289 */ 253 */
290 static int __init aio_setup(void) 254 static int __init aio_setup(void)
291 { 255 {
292 static struct file_system_type aio_fs 256 static struct file_system_type aio_fs = {
293 .name = "aio", 257 .name = "aio",
294 .init_fs_context = aio_init_fs !! 258 .mount = aio_mount,
295 .kill_sb = kill_anon_su 259 .kill_sb = kill_anon_super,
296 }; 260 };
297 aio_mnt = kern_mount(&aio_fs); 261 aio_mnt = kern_mount(&aio_fs);
298 if (IS_ERR(aio_mnt)) 262 if (IS_ERR(aio_mnt))
299 panic("Failed to create aio fs 263 panic("Failed to create aio fs mount.");
300 264
301 kiocb_cachep = KMEM_CACHE(aio_kiocb, S 265 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
302 kioctx_cachep = KMEM_CACHE(kioctx,SLAB 266 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
303 aio_sysctl_init(); !! 267
>> 268 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
>> 269
304 return 0; 270 return 0;
305 } 271 }
306 __initcall(aio_setup); 272 __initcall(aio_setup);
307 273
308 static void put_aio_ring_file(struct kioctx *c 274 static void put_aio_ring_file(struct kioctx *ctx)
309 { 275 {
310 struct file *aio_ring_file = ctx->aio_ 276 struct file *aio_ring_file = ctx->aio_ring_file;
311 struct address_space *i_mapping; 277 struct address_space *i_mapping;
312 278
313 if (aio_ring_file) { 279 if (aio_ring_file) {
314 truncate_setsize(file_inode(ai 280 truncate_setsize(file_inode(aio_ring_file), 0);
315 281
316 /* Prevent further access to t 282 /* Prevent further access to the kioctx from migratepages */
317 i_mapping = aio_ring_file->f_m 283 i_mapping = aio_ring_file->f_mapping;
318 spin_lock(&i_mapping->i_privat !! 284 spin_lock(&i_mapping->private_lock);
319 i_mapping->i_private_data = NU !! 285 i_mapping->private_data = NULL;
320 ctx->aio_ring_file = NULL; 286 ctx->aio_ring_file = NULL;
321 spin_unlock(&i_mapping->i_priv !! 287 spin_unlock(&i_mapping->private_lock);
322 288
323 fput(aio_ring_file); 289 fput(aio_ring_file);
324 } 290 }
325 } 291 }
326 292
327 static void aio_free_ring(struct kioctx *ctx) 293 static void aio_free_ring(struct kioctx *ctx)
328 { 294 {
329 int i; 295 int i;
330 296
331 /* Disconnect the kiotx from the ring 297 /* Disconnect the kiotx from the ring file. This prevents future
332 * accesses to the kioctx from page mi 298 * accesses to the kioctx from page migration.
333 */ 299 */
334 put_aio_ring_file(ctx); 300 put_aio_ring_file(ctx);
335 301
336 for (i = 0; i < ctx->nr_pages; i++) { 302 for (i = 0; i < ctx->nr_pages; i++) {
337 struct folio *folio = ctx->rin !! 303 struct page *page;
338 !! 304 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
339 if (!folio) !! 305 page_count(ctx->ring_pages[i]));
>> 306 page = ctx->ring_pages[i];
>> 307 if (!page)
340 continue; 308 continue;
341 !! 309 ctx->ring_pages[i] = NULL;
342 pr_debug("pid(%d) [%d] folio-> !! 310 put_page(page);
343 folio_ref_count(folio <<
344 ctx->ring_folios[i] = NULL; <<
345 folio_put(folio); <<
346 } 311 }
347 312
348 if (ctx->ring_folios && ctx->ring_foli !! 313 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
349 kfree(ctx->ring_folios); !! 314 kfree(ctx->ring_pages);
350 ctx->ring_folios = NULL; !! 315 ctx->ring_pages = NULL;
351 } 316 }
352 } 317 }
353 318
354 static int aio_ring_mremap(struct vm_area_stru 319 static int aio_ring_mremap(struct vm_area_struct *vma)
355 { 320 {
356 struct file *file = vma->vm_file; 321 struct file *file = vma->vm_file;
357 struct mm_struct *mm = vma->vm_mm; 322 struct mm_struct *mm = vma->vm_mm;
358 struct kioctx_table *table; 323 struct kioctx_table *table;
359 int i, res = -EINVAL; 324 int i, res = -EINVAL;
360 325
361 spin_lock(&mm->ioctx_lock); 326 spin_lock(&mm->ioctx_lock);
362 rcu_read_lock(); 327 rcu_read_lock();
363 table = rcu_dereference(mm->ioctx_tabl 328 table = rcu_dereference(mm->ioctx_table);
364 if (!table) <<
365 goto out_unlock; <<
366 <<
367 for (i = 0; i < table->nr; i++) { 329 for (i = 0; i < table->nr; i++) {
368 struct kioctx *ctx; 330 struct kioctx *ctx;
369 331
370 ctx = rcu_dereference(table->t !! 332 ctx = table->table[i];
371 if (ctx && ctx->aio_ring_file 333 if (ctx && ctx->aio_ring_file == file) {
372 if (!atomic_read(&ctx- 334 if (!atomic_read(&ctx->dead)) {
373 ctx->user_id = 335 ctx->user_id = ctx->mmap_base = vma->vm_start;
374 res = 0; 336 res = 0;
375 } 337 }
376 break; 338 break;
377 } 339 }
378 } 340 }
379 341
380 out_unlock: <<
381 rcu_read_unlock(); 342 rcu_read_unlock();
382 spin_unlock(&mm->ioctx_lock); 343 spin_unlock(&mm->ioctx_lock);
383 return res; 344 return res;
384 } 345 }
385 346
386 static const struct vm_operations_struct aio_r 347 static const struct vm_operations_struct aio_ring_vm_ops = {
387 .mremap = aio_ring_mremap, 348 .mremap = aio_ring_mremap,
388 #if IS_ENABLED(CONFIG_MMU) 349 #if IS_ENABLED(CONFIG_MMU)
389 .fault = filemap_fault, 350 .fault = filemap_fault,
390 .map_pages = filemap_map_pages, 351 .map_pages = filemap_map_pages,
391 .page_mkwrite = filemap_page_mkwrite 352 .page_mkwrite = filemap_page_mkwrite,
392 #endif 353 #endif
393 }; 354 };
394 355
395 static int aio_ring_mmap(struct file *file, st 356 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
396 { 357 {
397 vm_flags_set(vma, VM_DONTEXPAND); !! 358 vma->vm_flags |= VM_DONTEXPAND;
398 vma->vm_ops = &aio_ring_vm_ops; 359 vma->vm_ops = &aio_ring_vm_ops;
399 return 0; 360 return 0;
400 } 361 }
401 362
402 static const struct file_operations aio_ring_f 363 static const struct file_operations aio_ring_fops = {
403 .mmap = aio_ring_mmap, 364 .mmap = aio_ring_mmap,
404 }; 365 };
405 366
406 #if IS_ENABLED(CONFIG_MIGRATION) 367 #if IS_ENABLED(CONFIG_MIGRATION)
407 static int aio_migrate_folio(struct address_sp !! 368 static int aio_migratepage(struct address_space *mapping, struct page *new,
408 struct folio *src, enu !! 369 struct page *old, enum migrate_mode mode)
409 { 370 {
410 struct kioctx *ctx; 371 struct kioctx *ctx;
411 unsigned long flags; 372 unsigned long flags;
412 pgoff_t idx; 373 pgoff_t idx;
413 int rc = 0; !! 374 int rc;
414 375
415 /* mapping->i_private_lock here protec !! 376 rc = 0;
416 spin_lock(&mapping->i_private_lock); !! 377
417 ctx = mapping->i_private_data; !! 378 /* mapping->private_lock here protects against the kioctx teardown. */
>> 379 spin_lock(&mapping->private_lock);
>> 380 ctx = mapping->private_data;
418 if (!ctx) { 381 if (!ctx) {
419 rc = -EINVAL; 382 rc = -EINVAL;
420 goto out; 383 goto out;
421 } 384 }
422 385
423 /* The ring_lock mutex. The prevents 386 /* The ring_lock mutex. The prevents aio_read_events() from writing
424 * to the ring's head, and prevents pa 387 * to the ring's head, and prevents page migration from mucking in
425 * a partially initialized kiotx. 388 * a partially initialized kiotx.
426 */ 389 */
427 if (!mutex_trylock(&ctx->ring_lock)) { 390 if (!mutex_trylock(&ctx->ring_lock)) {
428 rc = -EAGAIN; 391 rc = -EAGAIN;
429 goto out; 392 goto out;
430 } 393 }
431 394
432 idx = src->index; !! 395 idx = old->index;
433 if (idx < (pgoff_t)ctx->nr_pages) { 396 if (idx < (pgoff_t)ctx->nr_pages) {
434 /* Make sure the old folio has !! 397 /* Make sure the old page hasn't already been changed */
435 if (ctx->ring_folios[idx] != s !! 398 if (ctx->ring_pages[idx] != old)
436 rc = -EAGAIN; 399 rc = -EAGAIN;
437 } else 400 } else
438 rc = -EINVAL; 401 rc = -EINVAL;
439 402
440 if (rc != 0) 403 if (rc != 0)
441 goto out_unlock; 404 goto out_unlock;
442 405
443 /* Writeback must be complete */ 406 /* Writeback must be complete */
444 BUG_ON(folio_test_writeback(src)); !! 407 BUG_ON(PageWriteback(old));
445 folio_get(dst); !! 408 get_page(new);
446 409
447 rc = folio_migrate_mapping(mapping, ds !! 410 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
448 if (rc != MIGRATEPAGE_SUCCESS) { 411 if (rc != MIGRATEPAGE_SUCCESS) {
449 folio_put(dst); !! 412 put_page(new);
450 goto out_unlock; 413 goto out_unlock;
451 } 414 }
452 415
453 /* Take completion_lock to prevent oth 416 /* Take completion_lock to prevent other writes to the ring buffer
454 * while the old folio is copied to th !! 417 * while the old page is copied to the new. This prevents new
455 * events from being lost. 418 * events from being lost.
456 */ 419 */
457 spin_lock_irqsave(&ctx->completion_loc 420 spin_lock_irqsave(&ctx->completion_lock, flags);
458 folio_copy(dst, src); !! 421 migrate_page_copy(new, old);
459 folio_migrate_flags(dst, src); !! 422 BUG_ON(ctx->ring_pages[idx] != old);
460 BUG_ON(ctx->ring_folios[idx] != src); !! 423 ctx->ring_pages[idx] = new;
461 ctx->ring_folios[idx] = dst; <<
462 spin_unlock_irqrestore(&ctx->completio 424 spin_unlock_irqrestore(&ctx->completion_lock, flags);
463 425
464 /* The old folio is no longer accessib !! 426 /* The old page is no longer accessible. */
465 folio_put(src); !! 427 put_page(old);
466 428
467 out_unlock: 429 out_unlock:
468 mutex_unlock(&ctx->ring_lock); 430 mutex_unlock(&ctx->ring_lock);
469 out: 431 out:
470 spin_unlock(&mapping->i_private_lock); !! 432 spin_unlock(&mapping->private_lock);
471 return rc; 433 return rc;
472 } 434 }
473 #else <<
474 #define aio_migrate_folio NULL <<
475 #endif 435 #endif
476 436
477 static const struct address_space_operations a 437 static const struct address_space_operations aio_ctx_aops = {
478 .dirty_folio = noop_dirty_folio, !! 438 .set_page_dirty = __set_page_dirty_no_writeback,
479 .migrate_folio = aio_migrate_folio, !! 439 #if IS_ENABLED(CONFIG_MIGRATION)
>> 440 .migratepage = aio_migratepage,
>> 441 #endif
480 }; 442 };
481 443
482 static int aio_setup_ring(struct kioctx *ctx, !! 444 static int aio_setup_ring(struct kioctx *ctx)
483 { 445 {
484 struct aio_ring *ring; 446 struct aio_ring *ring;
>> 447 unsigned nr_events = ctx->max_reqs;
485 struct mm_struct *mm = current->mm; 448 struct mm_struct *mm = current->mm;
486 unsigned long size, unused; 449 unsigned long size, unused;
487 int nr_pages; 450 int nr_pages;
488 int i; 451 int i;
489 struct file *file; 452 struct file *file;
490 453
491 /* Compensate for the ring buffer's he 454 /* Compensate for the ring buffer's head/tail overlap entry */
492 nr_events += 2; /* 1 is required, 2 fo 455 nr_events += 2; /* 1 is required, 2 for good luck */
493 456
494 size = sizeof(struct aio_ring); 457 size = sizeof(struct aio_ring);
495 size += sizeof(struct io_event) * nr_e 458 size += sizeof(struct io_event) * nr_events;
496 459
497 nr_pages = PFN_UP(size); 460 nr_pages = PFN_UP(size);
498 if (nr_pages < 0) 461 if (nr_pages < 0)
499 return -EINVAL; 462 return -EINVAL;
500 463
501 file = aio_private_file(ctx, nr_pages) 464 file = aio_private_file(ctx, nr_pages);
502 if (IS_ERR(file)) { 465 if (IS_ERR(file)) {
503 ctx->aio_ring_file = NULL; 466 ctx->aio_ring_file = NULL;
504 return -ENOMEM; 467 return -ENOMEM;
505 } 468 }
506 469
507 ctx->aio_ring_file = file; 470 ctx->aio_ring_file = file;
508 nr_events = (PAGE_SIZE * nr_pages - si 471 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
509 / sizeof(struct io_eve 472 / sizeof(struct io_event);
510 473
511 ctx->ring_folios = ctx->internal_folio !! 474 ctx->ring_pages = ctx->internal_pages;
512 if (nr_pages > AIO_RING_PAGES) { 475 if (nr_pages > AIO_RING_PAGES) {
513 ctx->ring_folios = kcalloc(nr_ !! 476 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
514 GFP !! 477 GFP_KERNEL);
515 if (!ctx->ring_folios) { !! 478 if (!ctx->ring_pages) {
516 put_aio_ring_file(ctx) 479 put_aio_ring_file(ctx);
517 return -ENOMEM; 480 return -ENOMEM;
518 } 481 }
519 } 482 }
520 483
521 for (i = 0; i < nr_pages; i++) { 484 for (i = 0; i < nr_pages; i++) {
522 struct folio *folio; !! 485 struct page *page;
523 !! 486 page = find_or_create_page(file->f_mapping,
524 folio = __filemap_get_folio(fi !! 487 i, GFP_HIGHUSER | __GFP_ZERO);
525 FG !! 488 if (!page)
526 GF <<
527 if (IS_ERR(folio)) <<
528 break; 489 break;
>> 490 pr_debug("pid(%d) page[%d]->count=%d\n",
>> 491 current->pid, i, page_count(page));
>> 492 SetPageUptodate(page);
>> 493 unlock_page(page);
529 494
530 pr_debug("pid(%d) [%d] folio-> !! 495 ctx->ring_pages[i] = page;
531 folio_ref_count(folio <<
532 folio_end_read(folio, true); <<
533 <<
534 ctx->ring_folios[i] = folio; <<
535 } 496 }
536 ctx->nr_pages = i; 497 ctx->nr_pages = i;
537 498
538 if (unlikely(i != nr_pages)) { 499 if (unlikely(i != nr_pages)) {
539 aio_free_ring(ctx); 500 aio_free_ring(ctx);
540 return -ENOMEM; 501 return -ENOMEM;
541 } 502 }
542 503
543 ctx->mmap_size = nr_pages * PAGE_SIZE; 504 ctx->mmap_size = nr_pages * PAGE_SIZE;
544 pr_debug("attempting mmap of %lu bytes 505 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
545 506
546 if (mmap_write_lock_killable(mm)) { !! 507 if (down_write_killable(&mm->mmap_sem)) {
547 ctx->mmap_size = 0; 508 ctx->mmap_size = 0;
548 aio_free_ring(ctx); 509 aio_free_ring(ctx);
549 return -EINTR; 510 return -EINTR;
550 } 511 }
551 512
552 ctx->mmap_base = do_mmap(ctx->aio_ring !! 513 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
553 PROT_READ | P !! 514 PROT_READ | PROT_WRITE,
554 MAP_SHARED, 0 !! 515 MAP_SHARED, 0, &unused, NULL);
555 mmap_write_unlock(mm); !! 516 up_write(&mm->mmap_sem);
556 if (IS_ERR((void *)ctx->mmap_base)) { 517 if (IS_ERR((void *)ctx->mmap_base)) {
557 ctx->mmap_size = 0; 518 ctx->mmap_size = 0;
558 aio_free_ring(ctx); 519 aio_free_ring(ctx);
559 return -ENOMEM; 520 return -ENOMEM;
560 } 521 }
561 522
562 pr_debug("mmap address: 0x%08lx\n", ct 523 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
563 524
564 ctx->user_id = ctx->mmap_base; 525 ctx->user_id = ctx->mmap_base;
565 ctx->nr_events = nr_events; /* trusted 526 ctx->nr_events = nr_events; /* trusted copy */
566 527
567 ring = folio_address(ctx->ring_folios[ !! 528 ring = kmap_atomic(ctx->ring_pages[0]);
568 ring->nr = nr_events; /* user copy * 529 ring->nr = nr_events; /* user copy */
569 ring->id = ~0U; 530 ring->id = ~0U;
570 ring->head = ring->tail = 0; 531 ring->head = ring->tail = 0;
571 ring->magic = AIO_RING_MAGIC; 532 ring->magic = AIO_RING_MAGIC;
572 ring->compat_features = AIO_RING_COMPA 533 ring->compat_features = AIO_RING_COMPAT_FEATURES;
573 ring->incompat_features = AIO_RING_INC 534 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
574 ring->header_length = sizeof(struct ai 535 ring->header_length = sizeof(struct aio_ring);
575 flush_dcache_folio(ctx->ring_folios[0] !! 536 kunmap_atomic(ring);
>> 537 flush_dcache_page(ctx->ring_pages[0]);
576 538
577 return 0; 539 return 0;
578 } 540 }
579 541
580 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / s 542 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
581 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - 543 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
582 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PE 544 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
583 545
584 void kiocb_set_cancel_fn(struct kiocb *iocb, k 546 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
585 { 547 {
586 struct aio_kiocb *req; !! 548 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
587 struct kioctx *ctx; !! 549 struct kioctx *ctx = req->ki_ctx;
588 unsigned long flags; 550 unsigned long flags;
589 551
590 /* !! 552 spin_lock_irqsave(&ctx->ctx_lock, flags);
591 * kiocb didn't come from aio or is ne <<
592 * ignore it. <<
593 */ <<
594 if (!(iocb->ki_flags & IOCB_AIO_RW)) <<
595 return; <<
596 <<
597 req = container_of(iocb, struct aio_ki <<
598 <<
599 if (WARN_ON_ONCE(!list_empty(&req->ki_ <<
600 return; <<
601 553
602 ctx = req->ki_ctx; !! 554 if (!req->ki_list.next)
>> 555 list_add(&req->ki_list, &ctx->active_reqs);
603 556
604 spin_lock_irqsave(&ctx->ctx_lock, flag <<
605 list_add_tail(&req->ki_list, &ctx->act <<
606 req->ki_cancel = cancel; 557 req->ki_cancel = cancel;
>> 558
607 spin_unlock_irqrestore(&ctx->ctx_lock, 559 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
608 } 560 }
609 EXPORT_SYMBOL(kiocb_set_cancel_fn); 561 EXPORT_SYMBOL(kiocb_set_cancel_fn);
610 562
611 /* !! 563 static int kiocb_cancel(struct aio_kiocb *kiocb)
612 * free_ioctx() should be RCU delayed to synch !! 564 {
613 * protected lookup_ioctx() and also needs pro !! 565 kiocb_cancel_fn *old, *cancel;
614 * aio_free_ring(). Use rcu_work. !! 566
615 */ !! 567 /*
>> 568 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
>> 569 * actually has a cancel function, hence the cmpxchg()
>> 570 */
>> 571
>> 572 cancel = ACCESS_ONCE(kiocb->ki_cancel);
>> 573 do {
>> 574 if (!cancel || cancel == KIOCB_CANCELLED)
>> 575 return -EINVAL;
>> 576
>> 577 old = cancel;
>> 578 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
>> 579 } while (cancel != old);
>> 580
>> 581 return cancel(&kiocb->common);
>> 582 }
>> 583
616 static void free_ioctx(struct work_struct *wor 584 static void free_ioctx(struct work_struct *work)
617 { 585 {
618 struct kioctx *ctx = container_of(to_r !! 586 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
619 free !! 587
620 pr_debug("freeing %p\n", ctx); 588 pr_debug("freeing %p\n", ctx);
621 589
622 aio_free_ring(ctx); 590 aio_free_ring(ctx);
623 free_percpu(ctx->cpu); 591 free_percpu(ctx->cpu);
624 percpu_ref_exit(&ctx->reqs); 592 percpu_ref_exit(&ctx->reqs);
625 percpu_ref_exit(&ctx->users); 593 percpu_ref_exit(&ctx->users);
626 kmem_cache_free(kioctx_cachep, ctx); 594 kmem_cache_free(kioctx_cachep, ctx);
627 } 595 }
628 596
629 static void free_ioctx_reqs(struct percpu_ref 597 static void free_ioctx_reqs(struct percpu_ref *ref)
630 { 598 {
631 struct kioctx *ctx = container_of(ref, 599 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
632 600
633 /* At this point we know that there ar 601 /* At this point we know that there are no any in-flight requests */
634 if (ctx->rq_wait && atomic_dec_and_tes 602 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
635 complete(&ctx->rq_wait->comp); 603 complete(&ctx->rq_wait->comp);
636 604
637 /* Synchronize against RCU protected t !! 605 INIT_WORK(&ctx->free_work, free_ioctx);
638 INIT_RCU_WORK(&ctx->free_rwork, free_i !! 606 schedule_work(&ctx->free_work);
639 queue_rcu_work(system_wq, &ctx->free_r <<
640 } 607 }
641 608
642 /* 609 /*
643 * When this function runs, the kioctx has bee 610 * When this function runs, the kioctx has been removed from the "hash table"
644 * and ctx->users has dropped to 0, so we know 611 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
645 * now it's safe to cancel any that need to be 612 * now it's safe to cancel any that need to be.
646 */ 613 */
647 static void free_ioctx_users(struct percpu_ref 614 static void free_ioctx_users(struct percpu_ref *ref)
648 { 615 {
649 struct kioctx *ctx = container_of(ref, 616 struct kioctx *ctx = container_of(ref, struct kioctx, users);
650 struct aio_kiocb *req; 617 struct aio_kiocb *req;
651 618
652 spin_lock_irq(&ctx->ctx_lock); 619 spin_lock_irq(&ctx->ctx_lock);
653 620
654 while (!list_empty(&ctx->active_reqs)) 621 while (!list_empty(&ctx->active_reqs)) {
655 req = list_first_entry(&ctx->a 622 req = list_first_entry(&ctx->active_reqs,
656 struct 623 struct aio_kiocb, ki_list);
657 req->ki_cancel(&req->rw); !! 624
658 list_del_init(&req->ki_list); 625 list_del_init(&req->ki_list);
>> 626 kiocb_cancel(req);
659 } 627 }
660 628
661 spin_unlock_irq(&ctx->ctx_lock); 629 spin_unlock_irq(&ctx->ctx_lock);
662 630
663 percpu_ref_kill(&ctx->reqs); 631 percpu_ref_kill(&ctx->reqs);
664 percpu_ref_put(&ctx->reqs); 632 percpu_ref_put(&ctx->reqs);
665 } 633 }
666 634
667 static int ioctx_add_table(struct kioctx *ctx, 635 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
668 { 636 {
669 unsigned i, new_nr; 637 unsigned i, new_nr;
670 struct kioctx_table *table, *old; 638 struct kioctx_table *table, *old;
671 struct aio_ring *ring; 639 struct aio_ring *ring;
672 640
673 spin_lock(&mm->ioctx_lock); 641 spin_lock(&mm->ioctx_lock);
674 table = rcu_dereference_raw(mm->ioctx_ 642 table = rcu_dereference_raw(mm->ioctx_table);
675 643
676 while (1) { 644 while (1) {
677 if (table) 645 if (table)
678 for (i = 0; i < table- 646 for (i = 0; i < table->nr; i++)
679 if (!rcu_acces !! 647 if (!table->table[i]) {
680 ctx->i 648 ctx->id = i;
681 rcu_as !! 649 table->table[i] = ctx;
682 spin_u 650 spin_unlock(&mm->ioctx_lock);
683 651
684 /* Whi 652 /* While kioctx setup is in progress,
685 * we 653 * we are protected from page migration
686 * cha !! 654 * changes ring_pages by ->ring_lock.
687 */ 655 */
688 ring = !! 656 ring = kmap_atomic(ctx->ring_pages[0]);
689 ring-> 657 ring->id = ctx->id;
>> 658 kunmap_atomic(ring);
690 return 659 return 0;
691 } 660 }
692 661
693 new_nr = (table ? table->nr : 662 new_nr = (table ? table->nr : 1) * 4;
694 spin_unlock(&mm->ioctx_lock); 663 spin_unlock(&mm->ioctx_lock);
695 664
696 table = kzalloc(struct_size(ta !! 665 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
>> 666 new_nr, GFP_KERNEL);
697 if (!table) 667 if (!table)
698 return -ENOMEM; 668 return -ENOMEM;
699 669
700 table->nr = new_nr; 670 table->nr = new_nr;
701 671
702 spin_lock(&mm->ioctx_lock); 672 spin_lock(&mm->ioctx_lock);
703 old = rcu_dereference_raw(mm-> 673 old = rcu_dereference_raw(mm->ioctx_table);
704 674
705 if (!old) { 675 if (!old) {
706 rcu_assign_pointer(mm- 676 rcu_assign_pointer(mm->ioctx_table, table);
707 } else if (table->nr > old->nr 677 } else if (table->nr > old->nr) {
708 memcpy(table->table, o 678 memcpy(table->table, old->table,
709 old->nr * sizeo 679 old->nr * sizeof(struct kioctx *));
710 680
711 rcu_assign_pointer(mm- 681 rcu_assign_pointer(mm->ioctx_table, table);
712 kfree_rcu(old, rcu); 682 kfree_rcu(old, rcu);
713 } else { 683 } else {
714 kfree(table); 684 kfree(table);
715 table = old; 685 table = old;
716 } 686 }
717 } 687 }
718 } 688 }
719 689
720 static void aio_nr_sub(unsigned nr) 690 static void aio_nr_sub(unsigned nr)
721 { 691 {
722 spin_lock(&aio_nr_lock); 692 spin_lock(&aio_nr_lock);
723 if (WARN_ON(aio_nr - nr > aio_nr)) 693 if (WARN_ON(aio_nr - nr > aio_nr))
724 aio_nr = 0; 694 aio_nr = 0;
725 else 695 else
726 aio_nr -= nr; 696 aio_nr -= nr;
727 spin_unlock(&aio_nr_lock); 697 spin_unlock(&aio_nr_lock);
728 } 698 }
729 699
730 /* ioctx_alloc 700 /* ioctx_alloc
731 * Allocates and initializes an ioctx. R 701 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
732 */ 702 */
733 static struct kioctx *ioctx_alloc(unsigned nr_ 703 static struct kioctx *ioctx_alloc(unsigned nr_events)
734 { 704 {
735 struct mm_struct *mm = current->mm; 705 struct mm_struct *mm = current->mm;
736 struct kioctx *ctx; 706 struct kioctx *ctx;
737 int err = -ENOMEM; 707 int err = -ENOMEM;
738 708
739 /* 709 /*
740 * Store the original nr_events -- wha <<
741 * for counting against the global lim <<
742 */ <<
743 unsigned int max_reqs = nr_events; <<
744 <<
745 /* <<
746 * We keep track of the number of avai 710 * We keep track of the number of available ringbuffer slots, to prevent
747 * overflow (reqs_available), and we a 711 * overflow (reqs_available), and we also use percpu counters for this.
748 * 712 *
749 * So since up to half the slots might 713 * So since up to half the slots might be on other cpu's percpu counters
750 * and unavailable, double nr_events s 714 * and unavailable, double nr_events so userspace sees what they
751 * expected: additionally, we move req 715 * expected: additionally, we move req_batch slots to/from percpu
752 * counters at a time, so make sure th 716 * counters at a time, so make sure that isn't 0:
753 */ 717 */
754 nr_events = max(nr_events, num_possibl 718 nr_events = max(nr_events, num_possible_cpus() * 4);
755 nr_events *= 2; 719 nr_events *= 2;
756 720
757 /* Prevent overflows */ 721 /* Prevent overflows */
758 if (nr_events > (0x10000000U / sizeof( 722 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
759 pr_debug("ENOMEM: nr_events to 723 pr_debug("ENOMEM: nr_events too high\n");
760 return ERR_PTR(-EINVAL); 724 return ERR_PTR(-EINVAL);
761 } 725 }
762 726
763 if (!nr_events || (unsigned long)max_r !! 727 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
764 return ERR_PTR(-EAGAIN); 728 return ERR_PTR(-EAGAIN);
765 729
766 ctx = kmem_cache_zalloc(kioctx_cachep, 730 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
767 if (!ctx) 731 if (!ctx)
768 return ERR_PTR(-ENOMEM); 732 return ERR_PTR(-ENOMEM);
769 733
770 ctx->max_reqs = max_reqs; !! 734 ctx->max_reqs = nr_events;
771 735
772 spin_lock_init(&ctx->ctx_lock); 736 spin_lock_init(&ctx->ctx_lock);
773 spin_lock_init(&ctx->completion_lock); 737 spin_lock_init(&ctx->completion_lock);
774 mutex_init(&ctx->ring_lock); 738 mutex_init(&ctx->ring_lock);
775 /* Protect against page migration thro 739 /* Protect against page migration throughout kiotx setup by keeping
776 * the ring_lock mutex held until setu 740 * the ring_lock mutex held until setup is complete. */
777 mutex_lock(&ctx->ring_lock); 741 mutex_lock(&ctx->ring_lock);
778 init_waitqueue_head(&ctx->wait); 742 init_waitqueue_head(&ctx->wait);
779 743
780 INIT_LIST_HEAD(&ctx->active_reqs); 744 INIT_LIST_HEAD(&ctx->active_reqs);
781 745
782 if (percpu_ref_init(&ctx->users, free_ 746 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
783 goto err; 747 goto err;
784 748
785 if (percpu_ref_init(&ctx->reqs, free_i 749 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
786 goto err; 750 goto err;
787 751
788 ctx->cpu = alloc_percpu(struct kioctx_ 752 ctx->cpu = alloc_percpu(struct kioctx_cpu);
789 if (!ctx->cpu) 753 if (!ctx->cpu)
790 goto err; 754 goto err;
791 755
792 err = aio_setup_ring(ctx, nr_events); !! 756 err = aio_setup_ring(ctx);
793 if (err < 0) 757 if (err < 0)
794 goto err; 758 goto err;
795 759
796 atomic_set(&ctx->reqs_available, ctx-> 760 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
797 ctx->req_batch = (ctx->nr_events - 1) 761 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
798 if (ctx->req_batch < 1) 762 if (ctx->req_batch < 1)
799 ctx->req_batch = 1; 763 ctx->req_batch = 1;
800 764
801 /* limit the number of system wide aio 765 /* limit the number of system wide aios */
802 spin_lock(&aio_nr_lock); 766 spin_lock(&aio_nr_lock);
803 if (aio_nr + ctx->max_reqs > aio_max_n !! 767 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
804 aio_nr + ctx->max_reqs < aio_nr) { !! 768 aio_nr + nr_events < aio_nr) {
805 spin_unlock(&aio_nr_lock); 769 spin_unlock(&aio_nr_lock);
806 err = -EAGAIN; 770 err = -EAGAIN;
807 goto err_ctx; 771 goto err_ctx;
808 } 772 }
809 aio_nr += ctx->max_reqs; 773 aio_nr += ctx->max_reqs;
810 spin_unlock(&aio_nr_lock); 774 spin_unlock(&aio_nr_lock);
811 775
812 percpu_ref_get(&ctx->users); /* io_ 776 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
813 percpu_ref_get(&ctx->reqs); /* fre 777 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
814 778
815 err = ioctx_add_table(ctx, mm); 779 err = ioctx_add_table(ctx, mm);
816 if (err) 780 if (err)
817 goto err_cleanup; 781 goto err_cleanup;
818 782
819 /* Release the ring_lock mutex now tha 783 /* Release the ring_lock mutex now that all setup is complete. */
820 mutex_unlock(&ctx->ring_lock); 784 mutex_unlock(&ctx->ring_lock);
821 785
822 pr_debug("allocated ioctx %p[%ld]: mm= 786 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
823 ctx, ctx->user_id, mm, ctx->n 787 ctx, ctx->user_id, mm, ctx->nr_events);
824 return ctx; 788 return ctx;
825 789
826 err_cleanup: 790 err_cleanup:
827 aio_nr_sub(ctx->max_reqs); 791 aio_nr_sub(ctx->max_reqs);
828 err_ctx: 792 err_ctx:
829 atomic_set(&ctx->dead, 1); 793 atomic_set(&ctx->dead, 1);
830 if (ctx->mmap_size) 794 if (ctx->mmap_size)
831 vm_munmap(ctx->mmap_base, ctx- 795 vm_munmap(ctx->mmap_base, ctx->mmap_size);
832 aio_free_ring(ctx); 796 aio_free_ring(ctx);
833 err: 797 err:
834 mutex_unlock(&ctx->ring_lock); 798 mutex_unlock(&ctx->ring_lock);
835 free_percpu(ctx->cpu); 799 free_percpu(ctx->cpu);
836 percpu_ref_exit(&ctx->reqs); 800 percpu_ref_exit(&ctx->reqs);
837 percpu_ref_exit(&ctx->users); 801 percpu_ref_exit(&ctx->users);
838 kmem_cache_free(kioctx_cachep, ctx); 802 kmem_cache_free(kioctx_cachep, ctx);
839 pr_debug("error allocating ioctx %d\n" 803 pr_debug("error allocating ioctx %d\n", err);
840 return ERR_PTR(err); 804 return ERR_PTR(err);
841 } 805 }
842 806
843 /* kill_ioctx 807 /* kill_ioctx
844 * Cancels all outstanding aio requests o 808 * Cancels all outstanding aio requests on an aio context. Used
845 * when the processes owning a context ha 809 * when the processes owning a context have all exited to encourage
846 * the rapid destruction of the kioctx. 810 * the rapid destruction of the kioctx.
847 */ 811 */
848 static int kill_ioctx(struct mm_struct *mm, st 812 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
849 struct ctx_rq_wait *wait 813 struct ctx_rq_wait *wait)
850 { 814 {
851 struct kioctx_table *table; 815 struct kioctx_table *table;
852 816
853 spin_lock(&mm->ioctx_lock); 817 spin_lock(&mm->ioctx_lock);
854 if (atomic_xchg(&ctx->dead, 1)) { 818 if (atomic_xchg(&ctx->dead, 1)) {
855 spin_unlock(&mm->ioctx_lock); 819 spin_unlock(&mm->ioctx_lock);
856 return -EINVAL; 820 return -EINVAL;
857 } 821 }
858 822
859 table = rcu_dereference_raw(mm->ioctx_ 823 table = rcu_dereference_raw(mm->ioctx_table);
860 WARN_ON(ctx != rcu_access_pointer(tabl !! 824 WARN_ON(ctx != table->table[ctx->id]);
861 RCU_INIT_POINTER(table->table[ctx->id] !! 825 table->table[ctx->id] = NULL;
862 spin_unlock(&mm->ioctx_lock); 826 spin_unlock(&mm->ioctx_lock);
863 827
864 /* free_ioctx_reqs() will do the neces !! 828 /* percpu_ref_kill() will do the necessary call_rcu() */
865 wake_up_all(&ctx->wait); 829 wake_up_all(&ctx->wait);
866 830
867 /* 831 /*
868 * It'd be more correct to do this in 832 * It'd be more correct to do this in free_ioctx(), after all
869 * the outstanding kiocbs have finishe 833 * the outstanding kiocbs have finished - but by then io_destroy
870 * has already returned, so io_setup() 834 * has already returned, so io_setup() could potentially return
871 * -EAGAIN with no ioctxs actually in 835 * -EAGAIN with no ioctxs actually in use (as far as userspace
872 * could tell). 836 * could tell).
873 */ 837 */
874 aio_nr_sub(ctx->max_reqs); 838 aio_nr_sub(ctx->max_reqs);
875 839
876 if (ctx->mmap_size) 840 if (ctx->mmap_size)
877 vm_munmap(ctx->mmap_base, ctx- 841 vm_munmap(ctx->mmap_base, ctx->mmap_size);
878 842
879 ctx->rq_wait = wait; 843 ctx->rq_wait = wait;
880 percpu_ref_kill(&ctx->users); 844 percpu_ref_kill(&ctx->users);
881 return 0; 845 return 0;
882 } 846 }
883 847
884 /* 848 /*
885 * exit_aio: called when the last user of mm g 849 * exit_aio: called when the last user of mm goes away. At this point, there is
886 * no way for any new requests to be submited 850 * no way for any new requests to be submited or any of the io_* syscalls to be
887 * called on the context. 851 * called on the context.
888 * 852 *
889 * There may be outstanding kiocbs, but free_i 853 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
890 * them. 854 * them.
891 */ 855 */
892 void exit_aio(struct mm_struct *mm) 856 void exit_aio(struct mm_struct *mm)
893 { 857 {
894 struct kioctx_table *table = rcu_deref 858 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
895 struct ctx_rq_wait wait; 859 struct ctx_rq_wait wait;
896 int i, skipped; 860 int i, skipped;
897 861
898 if (!table) 862 if (!table)
899 return; 863 return;
900 864
901 atomic_set(&wait.count, table->nr); 865 atomic_set(&wait.count, table->nr);
902 init_completion(&wait.comp); 866 init_completion(&wait.comp);
903 867
904 skipped = 0; 868 skipped = 0;
905 for (i = 0; i < table->nr; ++i) { 869 for (i = 0; i < table->nr; ++i) {
906 struct kioctx *ctx = !! 870 struct kioctx *ctx = table->table[i];
907 rcu_dereference_protec <<
908 871
909 if (!ctx) { 872 if (!ctx) {
910 skipped++; 873 skipped++;
911 continue; 874 continue;
912 } 875 }
913 876
914 /* 877 /*
915 * We don't need to bother wit 878 * We don't need to bother with munmap() here - exit_mmap(mm)
916 * is coming and it'll unmap e 879 * is coming and it'll unmap everything. And we simply can't,
917 * this is not necessarily our 880 * this is not necessarily our ->mm.
918 * Since kill_ioctx() uses non 881 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
919 * that it needs to unmap the 882 * that it needs to unmap the area, just set it to 0.
920 */ 883 */
921 ctx->mmap_size = 0; 884 ctx->mmap_size = 0;
922 kill_ioctx(mm, ctx, &wait); 885 kill_ioctx(mm, ctx, &wait);
923 } 886 }
924 887
925 if (!atomic_sub_and_test(skipped, &wai 888 if (!atomic_sub_and_test(skipped, &wait.count)) {
926 /* Wait until all IO for the c 889 /* Wait until all IO for the context are done. */
927 wait_for_completion(&wait.comp 890 wait_for_completion(&wait.comp);
928 } 891 }
929 892
930 RCU_INIT_POINTER(mm->ioctx_table, NULL 893 RCU_INIT_POINTER(mm->ioctx_table, NULL);
931 kfree(table); 894 kfree(table);
932 } 895 }
933 896
934 static void put_reqs_available(struct kioctx * 897 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
935 { 898 {
936 struct kioctx_cpu *kcpu; 899 struct kioctx_cpu *kcpu;
937 unsigned long flags; 900 unsigned long flags;
938 901
939 local_irq_save(flags); 902 local_irq_save(flags);
940 kcpu = this_cpu_ptr(ctx->cpu); 903 kcpu = this_cpu_ptr(ctx->cpu);
941 kcpu->reqs_available += nr; 904 kcpu->reqs_available += nr;
942 905
943 while (kcpu->reqs_available >= ctx->re 906 while (kcpu->reqs_available >= ctx->req_batch * 2) {
944 kcpu->reqs_available -= ctx->r 907 kcpu->reqs_available -= ctx->req_batch;
945 atomic_add(ctx->req_batch, &ct 908 atomic_add(ctx->req_batch, &ctx->reqs_available);
946 } 909 }
947 910
948 local_irq_restore(flags); 911 local_irq_restore(flags);
949 } 912 }
950 913
951 static bool __get_reqs_available(struct kioctx !! 914 static bool get_reqs_available(struct kioctx *ctx)
952 { 915 {
953 struct kioctx_cpu *kcpu; 916 struct kioctx_cpu *kcpu;
954 bool ret = false; 917 bool ret = false;
955 unsigned long flags; 918 unsigned long flags;
956 919
957 local_irq_save(flags); 920 local_irq_save(flags);
958 kcpu = this_cpu_ptr(ctx->cpu); 921 kcpu = this_cpu_ptr(ctx->cpu);
959 if (!kcpu->reqs_available) { 922 if (!kcpu->reqs_available) {
960 int avail = atomic_read(&ctx-> !! 923 int old, avail = atomic_read(&ctx->reqs_available);
961 924
962 do { 925 do {
963 if (avail < ctx->req_b 926 if (avail < ctx->req_batch)
964 goto out; 927 goto out;
965 } while (!atomic_try_cmpxchg(& !! 928
966 & !! 929 old = avail;
>> 930 avail = atomic_cmpxchg(&ctx->reqs_available,
>> 931 avail, avail - ctx->req_batch);
>> 932 } while (avail != old);
967 933
968 kcpu->reqs_available += ctx->r 934 kcpu->reqs_available += ctx->req_batch;
969 } 935 }
970 936
971 ret = true; 937 ret = true;
972 kcpu->reqs_available--; 938 kcpu->reqs_available--;
973 out: 939 out:
974 local_irq_restore(flags); 940 local_irq_restore(flags);
975 return ret; 941 return ret;
976 } 942 }
977 943
978 /* refill_reqs_available 944 /* refill_reqs_available
979 * Updates the reqs_available reference c 945 * Updates the reqs_available reference counts used for tracking the
980 * number of free slots in the completion 946 * number of free slots in the completion ring. This can be called
981 * from aio_complete() (to optimistically 947 * from aio_complete() (to optimistically update reqs_available) or
982 * from aio_get_req() (the we're out of e 948 * from aio_get_req() (the we're out of events case). It must be
983 * called holding ctx->completion_lock. 949 * called holding ctx->completion_lock.
984 */ 950 */
985 static void refill_reqs_available(struct kioct 951 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
986 unsigned tai 952 unsigned tail)
987 { 953 {
988 unsigned events_in_ring, completed; 954 unsigned events_in_ring, completed;
989 955
990 /* Clamp head since userland can write 956 /* Clamp head since userland can write to it. */
991 head %= ctx->nr_events; 957 head %= ctx->nr_events;
992 if (head <= tail) 958 if (head <= tail)
993 events_in_ring = tail - head; 959 events_in_ring = tail - head;
994 else 960 else
995 events_in_ring = ctx->nr_event 961 events_in_ring = ctx->nr_events - (head - tail);
996 962
997 completed = ctx->completed_events; 963 completed = ctx->completed_events;
998 if (events_in_ring < completed) 964 if (events_in_ring < completed)
999 completed -= events_in_ring; 965 completed -= events_in_ring;
1000 else 966 else
1001 completed = 0; 967 completed = 0;
1002 968
1003 if (!completed) 969 if (!completed)
1004 return; 970 return;
1005 971
1006 ctx->completed_events -= completed; 972 ctx->completed_events -= completed;
1007 put_reqs_available(ctx, completed); 973 put_reqs_available(ctx, completed);
1008 } 974 }
1009 975
1010 /* user_refill_reqs_available 976 /* user_refill_reqs_available
1011 * Called to refill reqs_available when 977 * Called to refill reqs_available when aio_get_req() encounters an
1012 * out of space in the completion ring. 978 * out of space in the completion ring.
1013 */ 979 */
1014 static void user_refill_reqs_available(struct 980 static void user_refill_reqs_available(struct kioctx *ctx)
1015 { 981 {
1016 spin_lock_irq(&ctx->completion_lock); 982 spin_lock_irq(&ctx->completion_lock);
1017 if (ctx->completed_events) { 983 if (ctx->completed_events) {
1018 struct aio_ring *ring; 984 struct aio_ring *ring;
1019 unsigned head; 985 unsigned head;
1020 986
1021 /* Access of ring->head may r 987 /* Access of ring->head may race with aio_read_events_ring()
1022 * here, but that's okay sinc 988 * here, but that's okay since whether we read the old version
1023 * or the new version, and ei 989 * or the new version, and either will be valid. The important
1024 * part is that head cannot p 990 * part is that head cannot pass tail since we prevent
1025 * aio_complete() from updati 991 * aio_complete() from updating tail by holding
1026 * ctx->completion_lock. Eve 992 * ctx->completion_lock. Even if head is invalid, the check
1027 * against ctx->completed_eve 993 * against ctx->completed_events below will make sure we do the
1028 * safe/right thing. 994 * safe/right thing.
1029 */ 995 */
1030 ring = folio_address(ctx->rin !! 996 ring = kmap_atomic(ctx->ring_pages[0]);
1031 head = ring->head; 997 head = ring->head;
>> 998 kunmap_atomic(ring);
1032 999
1033 refill_reqs_available(ctx, he 1000 refill_reqs_available(ctx, head, ctx->tail);
1034 } 1001 }
1035 1002
1036 spin_unlock_irq(&ctx->completion_lock 1003 spin_unlock_irq(&ctx->completion_lock);
1037 } 1004 }
1038 1005
1039 static bool get_reqs_available(struct kioctx <<
1040 { <<
1041 if (__get_reqs_available(ctx)) <<
1042 return true; <<
1043 user_refill_reqs_available(ctx); <<
1044 return __get_reqs_available(ctx); <<
1045 } <<
1046 <<
1047 /* aio_get_req 1006 /* aio_get_req
1048 * Allocate a slot for an aio request. 1007 * Allocate a slot for an aio request.
1049 * Returns NULL if no requests are free. 1008 * Returns NULL if no requests are free.
1050 * <<
1051 * The refcount is initialized to 2 - one for <<
1052 * one for the synchronous code that does thi <<
1053 */ 1009 */
1054 static inline struct aio_kiocb *aio_get_req(s 1010 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1055 { 1011 {
1056 struct aio_kiocb *req; 1012 struct aio_kiocb *req;
1057 1013
1058 req = kmem_cache_alloc(kiocb_cachep, !! 1014 if (!get_reqs_available(ctx)) {
1059 if (unlikely(!req)) !! 1015 user_refill_reqs_available(ctx);
1060 return NULL; !! 1016 if (!get_reqs_available(ctx))
1061 !! 1017 return NULL;
1062 if (unlikely(!get_reqs_available(ctx) <<
1063 kmem_cache_free(kiocb_cachep, <<
1064 return NULL; <<
1065 } 1018 }
1066 1019
>> 1020 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
>> 1021 if (unlikely(!req))
>> 1022 goto out_put;
>> 1023
1067 percpu_ref_get(&ctx->reqs); 1024 percpu_ref_get(&ctx->reqs);
>> 1025
1068 req->ki_ctx = ctx; 1026 req->ki_ctx = ctx;
1069 INIT_LIST_HEAD(&req->ki_list); <<
1070 refcount_set(&req->ki_refcnt, 2); <<
1071 req->ki_eventfd = NULL; <<
1072 return req; 1027 return req;
>> 1028 out_put:
>> 1029 put_reqs_available(ctx, 1);
>> 1030 return NULL;
>> 1031 }
>> 1032
>> 1033 static void kiocb_free(struct aio_kiocb *req)
>> 1034 {
>> 1035 if (req->common.ki_filp)
>> 1036 fput(req->common.ki_filp);
>> 1037 if (req->ki_eventfd != NULL)
>> 1038 eventfd_ctx_put(req->ki_eventfd);
>> 1039 kmem_cache_free(kiocb_cachep, req);
1073 } 1040 }
1074 1041
1075 static struct kioctx *lookup_ioctx(unsigned l 1042 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1076 { 1043 {
1077 struct aio_ring __user *ring = (void 1044 struct aio_ring __user *ring = (void __user *)ctx_id;
1078 struct mm_struct *mm = current->mm; 1045 struct mm_struct *mm = current->mm;
1079 struct kioctx *ctx, *ret = NULL; 1046 struct kioctx *ctx, *ret = NULL;
1080 struct kioctx_table *table; 1047 struct kioctx_table *table;
1081 unsigned id; 1048 unsigned id;
1082 1049
1083 if (get_user(id, &ring->id)) 1050 if (get_user(id, &ring->id))
1084 return NULL; 1051 return NULL;
1085 1052
1086 rcu_read_lock(); 1053 rcu_read_lock();
1087 table = rcu_dereference(mm->ioctx_tab 1054 table = rcu_dereference(mm->ioctx_table);
1088 1055
1089 if (!table || id >= table->nr) 1056 if (!table || id >= table->nr)
1090 goto out; 1057 goto out;
1091 1058
1092 id = array_index_nospec(id, table->nr !! 1059 ctx = table->table[id];
1093 ctx = rcu_dereference(table->table[id <<
1094 if (ctx && ctx->user_id == ctx_id) { 1060 if (ctx && ctx->user_id == ctx_id) {
1095 if (percpu_ref_tryget_live(&c !! 1061 percpu_ref_get(&ctx->users);
1096 ret = ctx; !! 1062 ret = ctx;
1097 } 1063 }
1098 out: 1064 out:
1099 rcu_read_unlock(); 1065 rcu_read_unlock();
1100 return ret; 1066 return ret;
1101 } 1067 }
1102 1068
1103 static inline void iocb_destroy(struct aio_ki <<
1104 { <<
1105 if (iocb->ki_eventfd) <<
1106 eventfd_ctx_put(iocb->ki_even <<
1107 if (iocb->ki_filp) <<
1108 fput(iocb->ki_filp); <<
1109 percpu_ref_put(&iocb->ki_ctx->reqs); <<
1110 kmem_cache_free(kiocb_cachep, iocb); <<
1111 } <<
1112 <<
1113 struct aio_waiter { <<
1114 struct wait_queue_entry w; <<
1115 size_t min_nr; <<
1116 }; <<
1117 <<
1118 /* aio_complete 1069 /* aio_complete
1119 * Called when the io request on the giv 1070 * Called when the io request on the given iocb is complete.
1120 */ 1071 */
1121 static void aio_complete(struct aio_kiocb *io !! 1072 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1122 { 1073 {
>> 1074 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1123 struct kioctx *ctx = iocb->ki_ctx; 1075 struct kioctx *ctx = iocb->ki_ctx;
1124 struct aio_ring *ring; 1076 struct aio_ring *ring;
1125 struct io_event *ev_page, *event; 1077 struct io_event *ev_page, *event;
1126 unsigned tail, pos, head, avail; !! 1078 unsigned tail, pos, head;
1127 unsigned long flags; 1079 unsigned long flags;
1128 1080
>> 1081 if (kiocb->ki_flags & IOCB_WRITE) {
>> 1082 struct file *file = kiocb->ki_filp;
>> 1083
>> 1084 /*
>> 1085 * Tell lockdep we inherited freeze protection from submission
>> 1086 * thread.
>> 1087 */
>> 1088 if (S_ISREG(file_inode(file)->i_mode))
>> 1089 __sb_writers_acquired(file_inode(file)->i_sb, SB_FREEZE_WRITE);
>> 1090 file_end_write(file);
>> 1091 }
>> 1092
>> 1093 /*
>> 1094 * Special case handling for sync iocbs:
>> 1095 * - events go directly into the iocb for fast handling
>> 1096 * - the sync task with the iocb in its stack holds the single iocb
>> 1097 * ref, no other paths have a way to get another ref
>> 1098 * - the sync task helpfully left a reference to itself in the iocb
>> 1099 */
>> 1100 BUG_ON(is_sync_kiocb(kiocb));
>> 1101
>> 1102 if (iocb->ki_list.next) {
>> 1103 unsigned long flags;
>> 1104
>> 1105 spin_lock_irqsave(&ctx->ctx_lock, flags);
>> 1106 list_del(&iocb->ki_list);
>> 1107 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
>> 1108 }
>> 1109
1129 /* 1110 /*
1130 * Add a completion event to the ring 1111 * Add a completion event to the ring buffer. Must be done holding
1131 * ctx->completion_lock to prevent ot 1112 * ctx->completion_lock to prevent other code from messing with the tail
1132 * pointer since we might be called f 1113 * pointer since we might be called from irq context.
1133 */ 1114 */
1134 spin_lock_irqsave(&ctx->completion_lo 1115 spin_lock_irqsave(&ctx->completion_lock, flags);
1135 1116
1136 tail = ctx->tail; 1117 tail = ctx->tail;
1137 pos = tail + AIO_EVENTS_OFFSET; 1118 pos = tail + AIO_EVENTS_OFFSET;
1138 1119
1139 if (++tail >= ctx->nr_events) 1120 if (++tail >= ctx->nr_events)
1140 tail = 0; 1121 tail = 0;
1141 1122
1142 ev_page = folio_address(ctx->ring_fol !! 1123 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1143 event = ev_page + pos % AIO_EVENTS_PE 1124 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1144 1125
1145 *event = iocb->ki_res; !! 1126 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1146 !! 1127 event->data = iocb->ki_user_data;
1147 flush_dcache_folio(ctx->ring_folios[p !! 1128 event->res = res;
1148 !! 1129 event->res2 = res2;
1149 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\ !! 1130
1150 (void __user *)(unsigned lon !! 1131 kunmap_atomic(ev_page);
1151 iocb->ki_res.data, iocb->ki_ !! 1132 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
>> 1133
>> 1134 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
>> 1135 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
>> 1136 res, res2);
1152 1137
1153 /* after flagging the request as done 1138 /* after flagging the request as done, we
1154 * must never even look at it again 1139 * must never even look at it again
1155 */ 1140 */
1156 smp_wmb(); /* make event visible 1141 smp_wmb(); /* make event visible before updating tail */
1157 1142
1158 ctx->tail = tail; 1143 ctx->tail = tail;
1159 1144
1160 ring = folio_address(ctx->ring_folios !! 1145 ring = kmap_atomic(ctx->ring_pages[0]);
1161 head = ring->head; 1146 head = ring->head;
1162 ring->tail = tail; 1147 ring->tail = tail;
1163 flush_dcache_folio(ctx->ring_folios[0 !! 1148 kunmap_atomic(ring);
>> 1149 flush_dcache_page(ctx->ring_pages[0]);
1164 1150
1165 ctx->completed_events++; 1151 ctx->completed_events++;
1166 if (ctx->completed_events > 1) 1152 if (ctx->completed_events > 1)
1167 refill_reqs_available(ctx, he 1153 refill_reqs_available(ctx, head, tail);
1168 <<
1169 avail = tail > head <<
1170 ? tail - head <<
1171 : tail + ctx->nr_events - hea <<
1172 spin_unlock_irqrestore(&ctx->completi 1154 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1173 1155
1174 pr_debug("added to ring %p at [%u]\n" 1156 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1175 1157
1176 /* 1158 /*
1177 * Check if the user asked us to deli 1159 * Check if the user asked us to deliver the result through an
1178 * eventfd. The eventfd_signal() func 1160 * eventfd. The eventfd_signal() function is safe to be called
1179 * from IRQ context. 1161 * from IRQ context.
1180 */ 1162 */
1181 if (iocb->ki_eventfd) !! 1163 if (iocb->ki_eventfd != NULL)
1182 eventfd_signal(iocb->ki_event !! 1164 eventfd_signal(iocb->ki_eventfd, 1);
>> 1165
>> 1166 /* everything turned out well, dispose of the aiocb. */
>> 1167 kiocb_free(iocb);
1183 1168
1184 /* 1169 /*
1185 * We have to order our ring_info tai 1170 * We have to order our ring_info tail store above and test
1186 * of the wait list below outside the 1171 * of the wait list below outside the wait lock. This is
1187 * like in wake_up_bit() where cleari 1172 * like in wake_up_bit() where clearing a bit has to be
1188 * ordered with the unlocked test. 1173 * ordered with the unlocked test.
1189 */ 1174 */
1190 smp_mb(); 1175 smp_mb();
1191 1176
1192 if (waitqueue_active(&ctx->wait)) { !! 1177 if (waitqueue_active(&ctx->wait))
1193 struct aio_waiter *curr, *nex !! 1178 wake_up(&ctx->wait);
1194 unsigned long flags; <<
1195 1179
1196 spin_lock_irqsave(&ctx->wait. !! 1180 percpu_ref_put(&ctx->reqs);
1197 list_for_each_entry_safe(curr <<
1198 if (avail >= curr->mi <<
1199 wake_up_proce <<
1200 list_del_init <<
1201 } <<
1202 spin_unlock_irqrestore(&ctx-> <<
1203 } <<
1204 } <<
1205 <<
1206 static inline void iocb_put(struct aio_kiocb <<
1207 { <<
1208 if (refcount_dec_and_test(&iocb->ki_r <<
1209 aio_complete(iocb); <<
1210 iocb_destroy(iocb); <<
1211 } <<
1212 } 1181 }
1213 1182
1214 /* aio_read_events_ring 1183 /* aio_read_events_ring
1215 * Pull an event off of the ioctx's even 1184 * Pull an event off of the ioctx's event ring. Returns the number of
1216 * events fetched 1185 * events fetched
1217 */ 1186 */
1218 static long aio_read_events_ring(struct kioct 1187 static long aio_read_events_ring(struct kioctx *ctx,
1219 struct io_ev 1188 struct io_event __user *event, long nr)
1220 { 1189 {
1221 struct aio_ring *ring; 1190 struct aio_ring *ring;
1222 unsigned head, tail, pos; 1191 unsigned head, tail, pos;
1223 long ret = 0; 1192 long ret = 0;
1224 int copy_ret; 1193 int copy_ret;
1225 1194
1226 /* 1195 /*
1227 * The mutex can block and wake us up 1196 * The mutex can block and wake us up and that will cause
1228 * wait_event_interruptible_hrtimeout 1197 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1229 * and repeat. This should be rare en 1198 * and repeat. This should be rare enough that it doesn't cause
1230 * peformance issues. See the comment 1199 * peformance issues. See the comment in read_events() for more detail.
1231 */ 1200 */
1232 sched_annotate_sleep(); 1201 sched_annotate_sleep();
1233 mutex_lock(&ctx->ring_lock); 1202 mutex_lock(&ctx->ring_lock);
1234 1203
1235 /* Access to ->ring_folios here is pr !! 1204 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1236 ring = folio_address(ctx->ring_folios !! 1205 ring = kmap_atomic(ctx->ring_pages[0]);
1237 head = ring->head; 1206 head = ring->head;
1238 tail = ring->tail; 1207 tail = ring->tail;
>> 1208 kunmap_atomic(ring);
1239 1209
1240 /* 1210 /*
1241 * Ensure that once we've read the cu 1211 * Ensure that once we've read the current tail pointer, that
1242 * we also see the events that were s 1212 * we also see the events that were stored up to the tail.
1243 */ 1213 */
1244 smp_rmb(); 1214 smp_rmb();
1245 1215
1246 pr_debug("h%u t%u m%u\n", head, tail, 1216 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1247 1217
1248 if (head == tail) 1218 if (head == tail)
1249 goto out; 1219 goto out;
1250 1220
1251 head %= ctx->nr_events; 1221 head %= ctx->nr_events;
1252 tail %= ctx->nr_events; 1222 tail %= ctx->nr_events;
1253 1223
1254 while (ret < nr) { 1224 while (ret < nr) {
1255 long avail; 1225 long avail;
1256 struct io_event *ev; 1226 struct io_event *ev;
1257 struct folio *folio; !! 1227 struct page *page;
1258 1228
1259 avail = (head <= tail ? tail 1229 avail = (head <= tail ? tail : ctx->nr_events) - head;
1260 if (head == tail) 1230 if (head == tail)
1261 break; 1231 break;
1262 1232
>> 1233 avail = min(avail, nr - ret);
>> 1234 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
>> 1235 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
>> 1236
1263 pos = head + AIO_EVENTS_OFFSE 1237 pos = head + AIO_EVENTS_OFFSET;
1264 folio = ctx->ring_folios[pos !! 1238 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1265 pos %= AIO_EVENTS_PER_PAGE; 1239 pos %= AIO_EVENTS_PER_PAGE;
1266 1240
1267 avail = min(avail, nr - ret); !! 1241 ev = kmap(page);
1268 avail = min_t(long, avail, AI <<
1269 <<
1270 ev = folio_address(folio); <<
1271 copy_ret = copy_to_user(event 1242 copy_ret = copy_to_user(event + ret, ev + pos,
1272 sizeo 1243 sizeof(*ev) * avail);
>> 1244 kunmap(page);
1273 1245
1274 if (unlikely(copy_ret)) { 1246 if (unlikely(copy_ret)) {
1275 ret = -EFAULT; 1247 ret = -EFAULT;
1276 goto out; 1248 goto out;
1277 } 1249 }
1278 1250
1279 ret += avail; 1251 ret += avail;
1280 head += avail; 1252 head += avail;
1281 head %= ctx->nr_events; 1253 head %= ctx->nr_events;
1282 } 1254 }
1283 1255
1284 ring = folio_address(ctx->ring_folios !! 1256 ring = kmap_atomic(ctx->ring_pages[0]);
1285 ring->head = head; 1257 ring->head = head;
1286 flush_dcache_folio(ctx->ring_folios[0 !! 1258 kunmap_atomic(ring);
>> 1259 flush_dcache_page(ctx->ring_pages[0]);
1287 1260
1288 pr_debug("%li h%u t%u\n", ret, head, 1261 pr_debug("%li h%u t%u\n", ret, head, tail);
1289 out: 1262 out:
1290 mutex_unlock(&ctx->ring_lock); 1263 mutex_unlock(&ctx->ring_lock);
1291 1264
1292 return ret; 1265 return ret;
1293 } 1266 }
1294 1267
1295 static bool aio_read_events(struct kioctx *ct 1268 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1296 struct io_event _ 1269 struct io_event __user *event, long *i)
1297 { 1270 {
1298 long ret = aio_read_events_ring(ctx, 1271 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1299 1272
1300 if (ret > 0) 1273 if (ret > 0)
1301 *i += ret; 1274 *i += ret;
1302 1275
1303 if (unlikely(atomic_read(&ctx->dead)) 1276 if (unlikely(atomic_read(&ctx->dead)))
1304 ret = -EINVAL; 1277 ret = -EINVAL;
1305 1278
1306 if (!*i) 1279 if (!*i)
1307 *i = ret; 1280 *i = ret;
1308 1281
1309 return ret < 0 || *i >= min_nr; 1282 return ret < 0 || *i >= min_nr;
1310 } 1283 }
1311 1284
1312 static long read_events(struct kioctx *ctx, l 1285 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1313 struct io_event __use 1286 struct io_event __user *event,
1314 ktime_t until) !! 1287 struct timespec __user *timeout)
1315 { 1288 {
1316 struct hrtimer_sleeper t; !! 1289 ktime_t until = KTIME_MAX;
1317 struct aio_waiter w; !! 1290 long ret = 0;
1318 long ret = 0, ret2 = 0; !! 1291
>> 1292 if (timeout) {
>> 1293 struct timespec ts;
>> 1294
>> 1295 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
>> 1296 return -EFAULT;
>> 1297
>> 1298 until = timespec_to_ktime(ts);
>> 1299 }
1319 1300
1320 /* 1301 /*
1321 * Note that aio_read_events() is bei 1302 * Note that aio_read_events() is being called as the conditional - i.e.
1322 * we're calling it after prepare_to_ 1303 * we're calling it after prepare_to_wait() has set task state to
1323 * TASK_INTERRUPTIBLE. 1304 * TASK_INTERRUPTIBLE.
1324 * 1305 *
1325 * But aio_read_events() can block, a 1306 * But aio_read_events() can block, and if it blocks it's going to flip
1326 * the task state back to TASK_RUNNIN 1307 * the task state back to TASK_RUNNING.
1327 * 1308 *
1328 * This should be ok, provided it doe 1309 * This should be ok, provided it doesn't flip the state back to
1329 * TASK_RUNNING and return 0 too much 1310 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1330 * will only happen if the mutex_lock 1311 * will only happen if the mutex_lock() call blocks, and we then find
1331 * the ringbuffer empty. So in practi 1312 * the ringbuffer empty. So in practice we should be ok, but it's
1332 * something to be aware of when touc 1313 * something to be aware of when touching this code.
1333 */ 1314 */
1334 aio_read_events(ctx, min_nr, nr, even !! 1315 if (until == 0)
1335 if (until == 0 || ret < 0 || ret >= m !! 1316 aio_read_events(ctx, min_nr, nr, event, &ret);
1336 return ret; !! 1317 else
1337 !! 1318 wait_event_interruptible_hrtimeout(ctx->wait,
1338 hrtimer_init_sleeper_on_stack(&t, CLO !! 1319 aio_read_events(ctx, min_nr, nr, event, &ret),
1339 if (until != KTIME_MAX) { !! 1320 until);
1340 hrtimer_set_expires_range_ns( <<
1341 hrtimer_sleeper_start_expires <<
1342 } <<
1343 <<
1344 init_wait(&w.w); <<
1345 <<
1346 while (1) { <<
1347 unsigned long nr_got = ret; <<
1348 <<
1349 w.min_nr = min_nr - ret; <<
1350 <<
1351 ret2 = prepare_to_wait_event( <<
1352 if (!ret2 && !t.task) <<
1353 ret2 = -ETIME; <<
1354 <<
1355 if (aio_read_events(ctx, min_ <<
1356 break; <<
1357 <<
1358 if (nr_got == ret) <<
1359 schedule(); <<
1360 } <<
1361 1321
1362 finish_wait(&ctx->wait, &w.w); !! 1322 if (!ret && signal_pending(current))
1363 hrtimer_cancel(&t.timer); !! 1323 ret = -EINTR;
1364 destroy_hrtimer_on_stack(&t.timer); <<
1365 1324
1366 return ret; 1325 return ret;
1367 } 1326 }
1368 1327
1369 /* sys_io_setup: 1328 /* sys_io_setup:
1370 * Create an aio_context capable of rece 1329 * Create an aio_context capable of receiving at least nr_events.
1371 * ctxp must not point to an aio_context 1330 * ctxp must not point to an aio_context that already exists, and
1372 * must be initialized to 0 prior to the 1331 * must be initialized to 0 prior to the call. On successful
1373 * creation of the aio_context, *ctxp is 1332 * creation of the aio_context, *ctxp is filled in with the resulting
1374 * handle. May fail with -EINVAL if *ct 1333 * handle. May fail with -EINVAL if *ctxp is not initialized,
1375 * if the specified nr_events exceeds in 1334 * if the specified nr_events exceeds internal limits. May fail
1376 * with -EAGAIN if the specified nr_even 1335 * with -EAGAIN if the specified nr_events exceeds the user's limit
1377 * of available events. May fail with - 1336 * of available events. May fail with -ENOMEM if insufficient kernel
1378 * resources are available. May fail wi 1337 * resources are available. May fail with -EFAULT if an invalid
1379 * pointer is passed for ctxp. Will fai 1338 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1380 * implemented. 1339 * implemented.
1381 */ 1340 */
1382 SYSCALL_DEFINE2(io_setup, unsigned, nr_events 1341 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1383 { 1342 {
1384 struct kioctx *ioctx = NULL; 1343 struct kioctx *ioctx = NULL;
1385 unsigned long ctx; 1344 unsigned long ctx;
1386 long ret; 1345 long ret;
1387 1346
1388 ret = get_user(ctx, ctxp); 1347 ret = get_user(ctx, ctxp);
1389 if (unlikely(ret)) 1348 if (unlikely(ret))
1390 goto out; 1349 goto out;
1391 1350
1392 ret = -EINVAL; 1351 ret = -EINVAL;
1393 if (unlikely(ctx || nr_events == 0)) 1352 if (unlikely(ctx || nr_events == 0)) {
1394 pr_debug("EINVAL: ctx %lu nr_ 1353 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1395 ctx, nr_events); 1354 ctx, nr_events);
1396 goto out; 1355 goto out;
1397 } 1356 }
1398 1357
1399 ioctx = ioctx_alloc(nr_events); 1358 ioctx = ioctx_alloc(nr_events);
1400 ret = PTR_ERR(ioctx); 1359 ret = PTR_ERR(ioctx);
1401 if (!IS_ERR(ioctx)) { 1360 if (!IS_ERR(ioctx)) {
1402 ret = put_user(ioctx->user_id 1361 ret = put_user(ioctx->user_id, ctxp);
1403 if (ret) 1362 if (ret)
1404 kill_ioctx(current->m 1363 kill_ioctx(current->mm, ioctx, NULL);
1405 percpu_ref_put(&ioctx->users) 1364 percpu_ref_put(&ioctx->users);
1406 } 1365 }
1407 1366
1408 out: 1367 out:
1409 return ret; 1368 return ret;
1410 } 1369 }
1411 1370
1412 #ifdef CONFIG_COMPAT 1371 #ifdef CONFIG_COMPAT
1413 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr 1372 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1414 { 1373 {
1415 struct kioctx *ioctx = NULL; 1374 struct kioctx *ioctx = NULL;
1416 unsigned long ctx; 1375 unsigned long ctx;
1417 long ret; 1376 long ret;
1418 1377
1419 ret = get_user(ctx, ctx32p); 1378 ret = get_user(ctx, ctx32p);
1420 if (unlikely(ret)) 1379 if (unlikely(ret))
1421 goto out; 1380 goto out;
1422 1381
1423 ret = -EINVAL; 1382 ret = -EINVAL;
1424 if (unlikely(ctx || nr_events == 0)) 1383 if (unlikely(ctx || nr_events == 0)) {
1425 pr_debug("EINVAL: ctx %lu nr_ 1384 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1426 ctx, nr_events); 1385 ctx, nr_events);
1427 goto out; 1386 goto out;
1428 } 1387 }
1429 1388
1430 ioctx = ioctx_alloc(nr_events); 1389 ioctx = ioctx_alloc(nr_events);
1431 ret = PTR_ERR(ioctx); 1390 ret = PTR_ERR(ioctx);
1432 if (!IS_ERR(ioctx)) { 1391 if (!IS_ERR(ioctx)) {
1433 /* truncating is ok because i 1392 /* truncating is ok because it's a user address */
1434 ret = put_user((u32)ioctx->us 1393 ret = put_user((u32)ioctx->user_id, ctx32p);
1435 if (ret) 1394 if (ret)
1436 kill_ioctx(current->m 1395 kill_ioctx(current->mm, ioctx, NULL);
1437 percpu_ref_put(&ioctx->users) 1396 percpu_ref_put(&ioctx->users);
1438 } 1397 }
1439 1398
1440 out: 1399 out:
1441 return ret; 1400 return ret;
1442 } 1401 }
1443 #endif 1402 #endif
1444 1403
1445 /* sys_io_destroy: 1404 /* sys_io_destroy:
1446 * Destroy the aio_context specified. M 1405 * Destroy the aio_context specified. May cancel any outstanding
1447 * AIOs and block on completion. Will f 1406 * AIOs and block on completion. Will fail with -ENOSYS if not
1448 * implemented. May fail with -EINVAL i 1407 * implemented. May fail with -EINVAL if the context pointed to
1449 * is invalid. 1408 * is invalid.
1450 */ 1409 */
1451 SYSCALL_DEFINE1(io_destroy, aio_context_t, ct 1410 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1452 { 1411 {
1453 struct kioctx *ioctx = lookup_ioctx(c 1412 struct kioctx *ioctx = lookup_ioctx(ctx);
1454 if (likely(NULL != ioctx)) { 1413 if (likely(NULL != ioctx)) {
1455 struct ctx_rq_wait wait; 1414 struct ctx_rq_wait wait;
1456 int ret; 1415 int ret;
1457 1416
1458 init_completion(&wait.comp); 1417 init_completion(&wait.comp);
1459 atomic_set(&wait.count, 1); 1418 atomic_set(&wait.count, 1);
1460 1419
1461 /* Pass requests_done to kill 1420 /* Pass requests_done to kill_ioctx() where it can be set
1462 * in a thread-safe way. If w 1421 * in a thread-safe way. If we try to set it here then we have
1463 * a race condition if two io 1422 * a race condition if two io_destroy() called simultaneously.
1464 */ 1423 */
1465 ret = kill_ioctx(current->mm, 1424 ret = kill_ioctx(current->mm, ioctx, &wait);
1466 percpu_ref_put(&ioctx->users) 1425 percpu_ref_put(&ioctx->users);
1467 1426
1468 /* Wait until all IO for the 1427 /* Wait until all IO for the context are done. Otherwise kernel
1469 * keep using user-space buff 1428 * keep using user-space buffers even if user thinks the context
1470 * is destroyed. 1429 * is destroyed.
1471 */ 1430 */
1472 if (!ret) 1431 if (!ret)
1473 wait_for_completion(& 1432 wait_for_completion(&wait.comp);
1474 1433
1475 return ret; 1434 return ret;
1476 } 1435 }
1477 pr_debug("EINVAL: invalid context id\ 1436 pr_debug("EINVAL: invalid context id\n");
1478 return -EINVAL; 1437 return -EINVAL;
1479 } 1438 }
1480 1439
1481 static void aio_remove_iocb(struct aio_kiocb !! 1440 static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec,
1482 { !! 1441 bool vectored, bool compat, struct iov_iter *iter)
1483 struct kioctx *ctx = iocb->ki_ctx; <<
1484 unsigned long flags; <<
1485 <<
1486 spin_lock_irqsave(&ctx->ctx_lock, fla <<
1487 list_del(&iocb->ki_list); <<
1488 spin_unlock_irqrestore(&ctx->ctx_lock <<
1489 } <<
1490 <<
1491 static void aio_complete_rw(struct kiocb *kio <<
1492 { <<
1493 struct aio_kiocb *iocb = container_of <<
1494 <<
1495 if (!list_empty_careful(&iocb->ki_lis <<
1496 aio_remove_iocb(iocb); <<
1497 <<
1498 if (kiocb->ki_flags & IOCB_WRITE) { <<
1499 struct inode *inode = file_in <<
1500 <<
1501 if (S_ISREG(inode->i_mode)) <<
1502 kiocb_end_write(kiocb <<
1503 } <<
1504 <<
1505 iocb->ki_res.res = res; <<
1506 iocb->ki_res.res2 = 0; <<
1507 iocb_put(iocb); <<
1508 } <<
1509 <<
1510 static int aio_prep_rw(struct kiocb *req, con <<
1511 { <<
1512 int ret; <<
1513 <<
1514 req->ki_complete = aio_complete_rw; <<
1515 req->private = NULL; <<
1516 req->ki_pos = iocb->aio_offset; <<
1517 req->ki_flags = req->ki_filp->f_iocb_ <<
1518 if (iocb->aio_flags & IOCB_FLAG_RESFD <<
1519 req->ki_flags |= IOCB_EVENTFD <<
1520 if (iocb->aio_flags & IOCB_FLAG_IOPRI <<
1521 /* <<
1522 * If the IOCB_FLAG_IOPRIO fl <<
1523 * aio_reqprio is interpreted <<
1524 * class and priority. <<
1525 */ <<
1526 ret = ioprio_check_cap(iocb-> <<
1527 if (ret) { <<
1528 pr_debug("aio ioprio <<
1529 return ret; <<
1530 } <<
1531 <<
1532 req->ki_ioprio = iocb->aio_re <<
1533 } else <<
1534 req->ki_ioprio = get_current_ <<
1535 <<
1536 ret = kiocb_set_rw_flags(req, iocb->a <<
1537 if (unlikely(ret)) <<
1538 return ret; <<
1539 <<
1540 req->ki_flags &= ~IOCB_HIPRI; /* no o <<
1541 return 0; <<
1542 } <<
1543 <<
1544 static ssize_t aio_setup_rw(int rw, const str <<
1545 struct iovec **iovec, bool ve <<
1546 struct iov_iter *iter) <<
1547 { 1442 {
1548 void __user *buf = (void __user *)(ui 1443 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1549 size_t len = iocb->aio_nbytes; 1444 size_t len = iocb->aio_nbytes;
1550 1445
1551 if (!vectored) { 1446 if (!vectored) {
1552 ssize_t ret = import_ubuf(rw, !! 1447 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1553 *iovec = NULL; 1448 *iovec = NULL;
1554 return ret; 1449 return ret;
1555 } 1450 }
1556 !! 1451 #ifdef CONFIG_COMPAT
1557 return __import_iovec(rw, buf, len, U !! 1452 if (compat)
>> 1453 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
>> 1454 iter);
>> 1455 #endif
>> 1456 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1558 } 1457 }
1559 1458
1560 static inline void aio_rw_done(struct kiocb * !! 1459 static inline ssize_t aio_ret(struct kiocb *req, ssize_t ret)
1561 { 1460 {
1562 switch (ret) { 1461 switch (ret) {
1563 case -EIOCBQUEUED: 1462 case -EIOCBQUEUED:
1564 break; !! 1463 return ret;
1565 case -ERESTARTSYS: 1464 case -ERESTARTSYS:
1566 case -ERESTARTNOINTR: 1465 case -ERESTARTNOINTR:
1567 case -ERESTARTNOHAND: 1466 case -ERESTARTNOHAND:
1568 case -ERESTART_RESTARTBLOCK: 1467 case -ERESTART_RESTARTBLOCK:
1569 /* 1468 /*
1570 * There's no easy way to res 1469 * There's no easy way to restart the syscall since other AIO's
1571 * may be already running. Ju 1470 * may be already running. Just fail this IO with EINTR.
1572 */ 1471 */
1573 ret = -EINTR; 1472 ret = -EINTR;
1574 fallthrough; !! 1473 /*FALLTHRU*/
1575 default: 1474 default:
1576 req->ki_complete(req, ret); !! 1475 aio_complete(req, ret, 0);
>> 1476 return 0;
1577 } 1477 }
1578 } 1478 }
1579 1479
1580 static int aio_read(struct kiocb *req, const !! 1480 static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
1581 bool vectored, bool c !! 1481 bool compat)
1582 { 1482 {
>> 1483 struct file *file = req->ki_filp;
1583 struct iovec inline_vecs[UIO_FASTIOV] 1484 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1584 struct iov_iter iter; 1485 struct iov_iter iter;
1585 struct file *file; !! 1486 ssize_t ret;
1586 int ret; <<
1587 1487
1588 ret = aio_prep_rw(req, iocb, READ); <<
1589 if (ret) <<
1590 return ret; <<
1591 file = req->ki_filp; <<
1592 if (unlikely(!(file->f_mode & FMODE_R 1488 if (unlikely(!(file->f_mode & FMODE_READ)))
1593 return -EBADF; 1489 return -EBADF;
1594 if (unlikely(!file->f_op->read_iter)) 1490 if (unlikely(!file->f_op->read_iter))
1595 return -EINVAL; 1491 return -EINVAL;
1596 1492
1597 ret = aio_setup_rw(ITER_DEST, iocb, & !! 1493 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1598 if (ret < 0) !! 1494 if (ret)
1599 return ret; 1495 return ret;
1600 ret = rw_verify_area(READ, file, &req 1496 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1601 if (!ret) 1497 if (!ret)
1602 aio_rw_done(req, file->f_op-> !! 1498 ret = aio_ret(req, call_read_iter(file, req, &iter));
1603 kfree(iovec); 1499 kfree(iovec);
1604 return ret; 1500 return ret;
1605 } 1501 }
1606 1502
1607 static int aio_write(struct kiocb *req, const !! 1503 static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored,
1608 bool vectored, bool !! 1504 bool compat)
1609 { 1505 {
>> 1506 struct file *file = req->ki_filp;
1610 struct iovec inline_vecs[UIO_FASTIOV] 1507 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1611 struct iov_iter iter; 1508 struct iov_iter iter;
1612 struct file *file; !! 1509 ssize_t ret;
1613 int ret; <<
1614 <<
1615 ret = aio_prep_rw(req, iocb, WRITE); <<
1616 if (ret) <<
1617 return ret; <<
1618 file = req->ki_filp; <<
1619 1510
1620 if (unlikely(!(file->f_mode & FMODE_W 1511 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1621 return -EBADF; 1512 return -EBADF;
1622 if (unlikely(!file->f_op->write_iter) 1513 if (unlikely(!file->f_op->write_iter))
1623 return -EINVAL; 1514 return -EINVAL;
1624 1515
1625 ret = aio_setup_rw(ITER_SOURCE, iocb, !! 1516 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1626 if (ret < 0) !! 1517 if (ret)
1627 return ret; 1518 return ret;
1628 ret = rw_verify_area(WRITE, file, &re 1519 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1629 if (!ret) { 1520 if (!ret) {
1630 if (S_ISREG(file_inode(file)- <<
1631 kiocb_start_write(req <<
1632 req->ki_flags |= IOCB_WRITE; 1521 req->ki_flags |= IOCB_WRITE;
1633 aio_rw_done(req, file->f_op-> !! 1522 file_start_write(file);
>> 1523 ret = aio_ret(req, call_write_iter(file, req, &iter));
>> 1524 /*
>> 1525 * We release freeze protection in aio_complete(). Fool lockdep
>> 1526 * by telling it the lock got released so that it doesn't
>> 1527 * complain about held lock when we return to userspace.
>> 1528 */
>> 1529 if (S_ISREG(file_inode(file)->i_mode))
>> 1530 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1634 } 1531 }
1635 kfree(iovec); 1532 kfree(iovec);
1636 return ret; 1533 return ret;
1637 } 1534 }
1638 1535
1639 static void aio_fsync_work(struct work_struct !! 1536 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1640 { !! 1537 struct iocb *iocb, bool compat)
1641 struct aio_kiocb *iocb = container_of <<
1642 const struct cred *old_cred = overrid <<
1643 <<
1644 iocb->ki_res.res = vfs_fsync(iocb->fs <<
1645 revert_creds(old_cred); <<
1646 put_cred(iocb->fsync.creds); <<
1647 iocb_put(iocb); <<
1648 } <<
1649 <<
1650 static int aio_fsync(struct fsync_iocb *req, <<
1651 bool datasync) <<
1652 { 1538 {
1653 if (unlikely(iocb->aio_buf || iocb->a !! 1539 struct aio_kiocb *req;
1654 iocb->aio_rw_flags)) !! 1540 struct file *file;
1655 return -EINVAL; !! 1541 ssize_t ret;
1656 1542
1657 if (unlikely(!req->file->f_op->fsync) !! 1543 /* enforce forwards compatibility on users */
>> 1544 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
>> 1545 pr_debug("EINVAL: reserve field set\n");
1658 return -EINVAL; 1546 return -EINVAL;
1659 <<
1660 req->creds = prepare_creds(); <<
1661 if (!req->creds) <<
1662 return -ENOMEM; <<
1663 <<
1664 req->datasync = datasync; <<
1665 INIT_WORK(&req->work, aio_fsync_work) <<
1666 schedule_work(&req->work); <<
1667 return 0; <<
1668 } <<
1669 <<
1670 static void aio_poll_put_work(struct work_str <<
1671 { <<
1672 struct poll_iocb *req = container_of( <<
1673 struct aio_kiocb *iocb = container_of <<
1674 <<
1675 iocb_put(iocb); <<
1676 } <<
1677 <<
1678 /* <<
1679 * Safely lock the waitqueue which the reques <<
1680 * case where the ->poll() provider decides t <<
1681 * <<
1682 * Returns true on success, meaning that req- <<
1683 * is on req->head, and an RCU read lock was <<
1684 * request was already removed from its waitq <<
1685 */ <<
1686 static bool poll_iocb_lock_wq(struct poll_ioc <<
1687 { <<
1688 wait_queue_head_t *head; <<
1689 <<
1690 /* <<
1691 * While we hold the waitqueue lock a <<
1692 * wake_up_pollfree() will wait for u <<
1693 * lock in the first place can race w <<
1694 * <<
1695 * We solve this as eventpoll does: b <<
1696 * all users of wake_up_pollfree() wi <<
1697 * we enter rcu_read_lock() and see t <<
1698 * non-NULL, we can then lock it with <<
1699 * under us, then check whether the r <<
1700 * <<
1701 * Keep holding rcu_read_lock() as lo <<
1702 * case the caller deletes the entry <<
1703 * In that case, only RCU prevents th <<
1704 */ <<
1705 rcu_read_lock(); <<
1706 head = smp_load_acquire(&req->head); <<
1707 if (head) { <<
1708 spin_lock(&head->lock); <<
1709 if (!list_empty(&req->wait.en <<
1710 return true; <<
1711 spin_unlock(&head->lock); <<
1712 } <<
1713 rcu_read_unlock(); <<
1714 return false; <<
1715 } <<
1716 <<
1717 static void poll_iocb_unlock_wq(struct poll_i <<
1718 { <<
1719 spin_unlock(&req->head->lock); <<
1720 rcu_read_unlock(); <<
1721 } <<
1722 <<
1723 static void aio_poll_complete_work(struct wor <<
1724 { <<
1725 struct poll_iocb *req = container_of( <<
1726 struct aio_kiocb *iocb = container_of <<
1727 struct poll_table_struct pt = { ._key <<
1728 struct kioctx *ctx = iocb->ki_ctx; <<
1729 __poll_t mask = 0; <<
1730 <<
1731 if (!READ_ONCE(req->cancelled)) <<
1732 mask = vfs_poll(req->file, &p <<
1733 <<
1734 /* <<
1735 * Note that ->ki_cancel callers also <<
1736 * calling ->ki_cancel. We need the <<
1737 * synchronize with them. In the can <<
1738 * itself is not actually needed, but <<
1739 * avoid further branches in the fast <<
1740 */ <<
1741 spin_lock_irq(&ctx->ctx_lock); <<
1742 if (poll_iocb_lock_wq(req)) { <<
1743 if (!mask && !READ_ONCE(req-> <<
1744 /* <<
1745 * The request isn't <<
1746 * Reschedule complet <<
1747 */ <<
1748 if (req->work_need_re <<
1749 schedule_work <<
1750 req->work_nee <<
1751 } else { <<
1752 req->work_sch <<
1753 } <<
1754 poll_iocb_unlock_wq(r <<
1755 spin_unlock_irq(&ctx- <<
1756 return; <<
1757 } <<
1758 list_del_init(&req->wait.entr <<
1759 poll_iocb_unlock_wq(req); <<
1760 } /* else, POLLFREE has freed the wai <<
1761 list_del_init(&iocb->ki_list); <<
1762 iocb->ki_res.res = mangle_poll(mask); <<
1763 spin_unlock_irq(&ctx->ctx_lock); <<
1764 <<
1765 iocb_put(iocb); <<
1766 } <<
1767 <<
1768 /* assumes we are called with irqs disabled * <<
1769 static int aio_poll_cancel(struct kiocb *iocb <<
1770 { <<
1771 struct aio_kiocb *aiocb = container_o <<
1772 struct poll_iocb *req = &aiocb->poll; <<
1773 <<
1774 if (poll_iocb_lock_wq(req)) { <<
1775 WRITE_ONCE(req->cancelled, tr <<
1776 if (!req->work_scheduled) { <<
1777 schedule_work(&aiocb- <<
1778 req->work_scheduled = <<
1779 } <<
1780 poll_iocb_unlock_wq(req); <<
1781 } /* else, the request was force-canc <<
1782 <<
1783 return 0; <<
1784 } <<
1785 <<
1786 static int aio_poll_wake(struct wait_queue_en <<
1787 void *key) <<
1788 { <<
1789 struct poll_iocb *req = container_of( <<
1790 struct aio_kiocb *iocb = container_of <<
1791 __poll_t mask = key_to_poll(key); <<
1792 unsigned long flags; <<
1793 <<
1794 /* for instances that support it chec <<
1795 if (mask && !(mask & req->events)) <<
1796 return 0; <<
1797 <<
1798 /* <<
1799 * Complete the request inline if pos <<
1800 * conditions be met: <<
1801 * 1. An event mask must have been <<
1802 * instead, then mask == 0 and w <<
1803 * the events, so inline complet <<
1804 * 2. The completion work must not <<
1805 * 3. ctx_lock must not be busy. W <<
1806 * already hold the waitqueue lo <<
1807 * locking order. Use irqsave/i <<
1808 * filesystems (e.g. fuse) call <<
1809 * yet IRQs have to be disabled <<
1810 */ <<
1811 if (mask && !req->work_scheduled && <<
1812 spin_trylock_irqsave(&iocb->ki_ct <<
1813 struct kioctx *ctx = iocb->ki <<
1814 <<
1815 list_del_init(&req->wait.entr <<
1816 list_del(&iocb->ki_list); <<
1817 iocb->ki_res.res = mangle_pol <<
1818 if (iocb->ki_eventfd && !even <<
1819 iocb = NULL; <<
1820 INIT_WORK(&req->work, <<
1821 schedule_work(&req->w <<
1822 } <<
1823 spin_unlock_irqrestore(&ctx-> <<
1824 if (iocb) <<
1825 iocb_put(iocb); <<
1826 } else { <<
1827 /* <<
1828 * Schedule the completion wo <<
1829 * scheduled, record that ano <<
1830 * <<
1831 * Don't remove the request f <<
1832 * not actually be complete y <<
1833 * is called), and we must no <<
1834 * exception to this; see bel <<
1835 */ <<
1836 if (req->work_scheduled) { <<
1837 req->work_need_resche <<
1838 } else { <<
1839 schedule_work(&req->w <<
1840 req->work_scheduled = <<
1841 } <<
1842 <<
1843 /* <<
1844 * If the waitqueue is being <<
1845 * the request inline, we hav <<
1846 * we can. That means immedi <<
1847 * waitqueue and preventing a <<
1848 * waitqueue via the request. <<
1849 * completion work (done abov <<
1850 * cancelled, to potentially <<
1851 */ <<
1852 if (mask & POLLFREE) { <<
1853 WRITE_ONCE(req->cance <<
1854 list_del_init(&req->w <<
1855 <<
1856 /* <<
1857 * Careful: this *mus <<
1858 * as req->head is NU <<
1859 * completed and free <<
1860 * will no longer nee <<
1861 */ <<
1862 smp_store_release(&re <<
1863 } <<
1864 } <<
1865 return 1; <<
1866 } <<
1867 <<
1868 struct aio_poll_table { <<
1869 struct poll_table_struct pt; <<
1870 struct aio_kiocb *iocb <<
1871 bool queue <<
1872 int error <<
1873 }; <<
1874 <<
1875 static void <<
1876 aio_poll_queue_proc(struct file *file, struct <<
1877 struct poll_table_struct *p) <<
1878 { <<
1879 struct aio_poll_table *pt = container <<
1880 <<
1881 /* multiple wait queues per file are <<
1882 if (unlikely(pt->queued)) { <<
1883 pt->error = -EINVAL; <<
1884 return; <<
1885 } 1547 }
1886 1548
1887 pt->queued = true; !! 1549 /* prevent overflows */
1888 pt->error = 0; !! 1550 if (unlikely(
1889 pt->iocb->poll.head = head; !! 1551 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1890 add_wait_queue(head, &pt->iocb->poll. !! 1552 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1891 } !! 1553 ((ssize_t)iocb->aio_nbytes < 0)
1892 !! 1554 )) {
1893 static int aio_poll(struct aio_kiocb *aiocb, !! 1555 pr_debug("EINVAL: overflow check\n");
1894 { <<
1895 struct kioctx *ctx = aiocb->ki_ctx; <<
1896 struct poll_iocb *req = &aiocb->poll; <<
1897 struct aio_poll_table apt; <<
1898 bool cancel = false; <<
1899 __poll_t mask; <<
1900 <<
1901 /* reject any unknown events outside <<
1902 if ((u16)iocb->aio_buf != iocb->aio_b <<
1903 return -EINVAL; <<
1904 /* reject fields that are not defined <<
1905 if (iocb->aio_offset || iocb->aio_nby <<
1906 return -EINVAL; 1556 return -EINVAL;
1907 <<
1908 INIT_WORK(&req->work, aio_poll_comple <<
1909 req->events = demangle_poll(iocb->aio <<
1910 <<
1911 req->head = NULL; <<
1912 req->cancelled = false; <<
1913 req->work_scheduled = false; <<
1914 req->work_need_resched = false; <<
1915 <<
1916 apt.pt._qproc = aio_poll_queue_proc; <<
1917 apt.pt._key = req->events; <<
1918 apt.iocb = aiocb; <<
1919 apt.queued = false; <<
1920 apt.error = -EINVAL; /* same as no su <<
1921 <<
1922 /* initialized the list so that we ca <<
1923 INIT_LIST_HEAD(&req->wait.entry); <<
1924 init_waitqueue_func_entry(&req->wait, <<
1925 <<
1926 mask = vfs_poll(req->file, &apt.pt) & <<
1927 spin_lock_irq(&ctx->ctx_lock); <<
1928 if (likely(apt.queued)) { <<
1929 bool on_queue = poll_iocb_loc <<
1930 <<
1931 if (!on_queue || req->work_sc <<
1932 /* <<
1933 * aio_poll_wake() al <<
1934 * completion work, o <<
1935 */ <<
1936 if (apt.error) /* uns <<
1937 cancel = true <<
1938 apt.error = 0; <<
1939 mask = 0; <<
1940 } <<
1941 if (mask || apt.error) { <<
1942 /* Steal to complete <<
1943 list_del_init(&req->w <<
1944 } else if (cancel) { <<
1945 /* Cancel if possible <<
1946 WRITE_ONCE(req->cance <<
1947 } else if (on_queue) { <<
1948 /* <<
1949 * Actually waiting f <<
1950 * active_reqs so tha <<
1951 */ <<
1952 list_add_tail(&aiocb- <<
1953 aiocb->ki_cancel = ai <<
1954 } <<
1955 if (on_queue) <<
1956 poll_iocb_unlock_wq(r <<
1957 } 1557 }
1958 if (mask) { /* no async, we'd stolen <<
1959 aiocb->ki_res.res = mangle_po <<
1960 apt.error = 0; <<
1961 } <<
1962 spin_unlock_irq(&ctx->ctx_lock); <<
1963 if (mask) <<
1964 iocb_put(aiocb); <<
1965 return apt.error; <<
1966 } <<
1967 1558
1968 static int __io_submit_one(struct kioctx *ctx !! 1559 req = aio_get_req(ctx);
1969 struct iocb __user !! 1560 if (unlikely(!req))
1970 bool compat) !! 1561 return -EAGAIN;
1971 { !! 1562
1972 req->ki_filp = fget(iocb->aio_fildes) !! 1563 req->common.ki_filp = file = fget(iocb->aio_fildes);
1973 if (unlikely(!req->ki_filp)) !! 1564 if (unlikely(!req->common.ki_filp)) {
1974 return -EBADF; !! 1565 ret = -EBADF;
>> 1566 goto out_put_req;
>> 1567 }
>> 1568 req->common.ki_pos = iocb->aio_offset;
>> 1569 req->common.ki_complete = aio_complete;
>> 1570 req->common.ki_flags = iocb_flags(req->common.ki_filp);
1975 1571
1976 if (iocb->aio_flags & IOCB_FLAG_RESFD 1572 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1977 struct eventfd_ctx *eventfd; <<
1978 /* 1573 /*
1979 * If the IOCB_FLAG_RESFD fla 1574 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1980 * instance of the file* now. 1575 * instance of the file* now. The file descriptor must be
1981 * an eventfd() fd, and will 1576 * an eventfd() fd, and will be signaled for each completed
1982 * event using the eventfd_si 1577 * event using the eventfd_signal() function.
1983 */ 1578 */
1984 eventfd = eventfd_ctx_fdget(i !! 1579 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1985 if (IS_ERR(eventfd)) !! 1580 if (IS_ERR(req->ki_eventfd)) {
1986 return PTR_ERR(eventf !! 1581 ret = PTR_ERR(req->ki_eventfd);
>> 1582 req->ki_eventfd = NULL;
>> 1583 goto out_put_req;
>> 1584 }
1987 1585
1988 req->ki_eventfd = eventfd; !! 1586 req->common.ki_flags |= IOCB_EVENTFD;
1989 } 1587 }
1990 1588
1991 if (unlikely(put_user(KIOCB_KEY, &use !! 1589 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
>> 1590 if (unlikely(ret)) {
1992 pr_debug("EFAULT: aio_key\n") 1591 pr_debug("EFAULT: aio_key\n");
1993 return -EFAULT; !! 1592 goto out_put_req;
1994 } 1593 }
1995 1594
1996 req->ki_res.obj = (u64)(unsigned long !! 1595 req->ki_user_iocb = user_iocb;
1997 req->ki_res.data = iocb->aio_data; !! 1596 req->ki_user_data = iocb->aio_data;
1998 req->ki_res.res = 0; <<
1999 req->ki_res.res2 = 0; <<
2000 1597
>> 1598 get_file(file);
2001 switch (iocb->aio_lio_opcode) { 1599 switch (iocb->aio_lio_opcode) {
2002 case IOCB_CMD_PREAD: 1600 case IOCB_CMD_PREAD:
2003 return aio_read(&req->rw, ioc !! 1601 ret = aio_read(&req->common, iocb, false, compat);
>> 1602 break;
2004 case IOCB_CMD_PWRITE: 1603 case IOCB_CMD_PWRITE:
2005 return aio_write(&req->rw, io !! 1604 ret = aio_write(&req->common, iocb, false, compat);
>> 1605 break;
2006 case IOCB_CMD_PREADV: 1606 case IOCB_CMD_PREADV:
2007 return aio_read(&req->rw, ioc !! 1607 ret = aio_read(&req->common, iocb, true, compat);
>> 1608 break;
2008 case IOCB_CMD_PWRITEV: 1609 case IOCB_CMD_PWRITEV:
2009 return aio_write(&req->rw, io !! 1610 ret = aio_write(&req->common, iocb, true, compat);
2010 case IOCB_CMD_FSYNC: !! 1611 break;
2011 return aio_fsync(&req->fsync, <<
2012 case IOCB_CMD_FDSYNC: <<
2013 return aio_fsync(&req->fsync, <<
2014 case IOCB_CMD_POLL: <<
2015 return aio_poll(req, iocb); <<
2016 default: 1612 default:
2017 pr_debug("invalid aio operati 1613 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
2018 return -EINVAL; !! 1614 ret = -EINVAL;
>> 1615 break;
2019 } 1616 }
>> 1617 fput(file);
>> 1618
>> 1619 if (ret && ret != -EIOCBQUEUED)
>> 1620 goto out_put_req;
>> 1621 return 0;
>> 1622 out_put_req:
>> 1623 put_reqs_available(ctx, 1);
>> 1624 percpu_ref_put(&ctx->reqs);
>> 1625 kiocb_free(req);
>> 1626 return ret;
2020 } 1627 }
2021 1628
2022 static int io_submit_one(struct kioctx *ctx, !! 1629 static long do_io_submit(aio_context_t ctx_id, long nr,
2023 bool compat) !! 1630 struct iocb __user *__user *iocbpp, bool compat)
2024 { 1631 {
2025 struct aio_kiocb *req; !! 1632 struct kioctx *ctx;
2026 struct iocb iocb; !! 1633 long ret = 0;
2027 int err; !! 1634 int i = 0;
>> 1635 struct blk_plug plug;
>> 1636
>> 1637 if (unlikely(nr < 0))
>> 1638 return -EINVAL;
>> 1639
>> 1640 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
>> 1641 nr = LONG_MAX/sizeof(*iocbpp);
2028 1642
2029 if (unlikely(copy_from_user(&iocb, us !! 1643 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
2030 return -EFAULT; 1644 return -EFAULT;
2031 1645
2032 /* enforce forwards compatibility on !! 1646 ctx = lookup_ioctx(ctx_id);
2033 if (unlikely(iocb.aio_reserved2)) { !! 1647 if (unlikely(!ctx)) {
2034 pr_debug("EINVAL: reserve fie !! 1648 pr_debug("EINVAL: invalid context id\n");
2035 return -EINVAL; 1649 return -EINVAL;
2036 } 1650 }
2037 1651
2038 /* prevent overflows */ !! 1652 blk_start_plug(&plug);
2039 if (unlikely( <<
2040 (iocb.aio_buf != (unsigned long)i <<
2041 (iocb.aio_nbytes != (size_t)iocb. <<
2042 ((ssize_t)iocb.aio_nbytes < 0) <<
2043 )) { <<
2044 pr_debug("EINVAL: overflow ch <<
2045 return -EINVAL; <<
2046 } <<
2047 1653
2048 req = aio_get_req(ctx); !! 1654 /*
2049 if (unlikely(!req)) !! 1655 * AKPM: should this return a partial result if some of the IOs were
2050 return -EAGAIN; !! 1656 * successfully submitted?
>> 1657 */
>> 1658 for (i=0; i<nr; i++) {
>> 1659 struct iocb __user *user_iocb;
>> 1660 struct iocb tmp;
2051 1661
2052 err = __io_submit_one(ctx, &iocb, use !! 1662 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
>> 1663 ret = -EFAULT;
>> 1664 break;
>> 1665 }
2053 1666
2054 /* Done with the synchronous referenc !! 1667 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
2055 iocb_put(req); !! 1668 ret = -EFAULT;
>> 1669 break;
>> 1670 }
2056 1671
2057 /* !! 1672 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
2058 * If err is 0, we'd either done aio_ !! 1673 if (ret)
2059 * arranged for that to be done async !! 1674 break;
2060 * means that we need to destroy req <<
2061 */ <<
2062 if (unlikely(err)) { <<
2063 iocb_destroy(req); <<
2064 put_reqs_available(ctx, 1); <<
2065 } 1675 }
2066 return err; !! 1676 blk_finish_plug(&plug);
>> 1677
>> 1678 percpu_ref_put(&ctx->users);
>> 1679 return i ? i : ret;
2067 } 1680 }
2068 1681
2069 /* sys_io_submit: 1682 /* sys_io_submit:
2070 * Queue the nr iocbs pointed to by iocb 1683 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2071 * the number of iocbs queued. May retu 1684 * the number of iocbs queued. May return -EINVAL if the aio_context
2072 * specified by ctx_id is invalid, if nr 1685 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2073 * *iocbpp[0] is not properly initialize 1686 * *iocbpp[0] is not properly initialized, if the operation specified
2074 * is invalid for the file descriptor in 1687 * is invalid for the file descriptor in the iocb. May fail with
2075 * -EFAULT if any of the data structures 1688 * -EFAULT if any of the data structures point to invalid data. May
2076 * fail with -EBADF if the file descript 1689 * fail with -EBADF if the file descriptor specified in the first
2077 * iocb is invalid. May fail with -EAGA 1690 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2078 * are available to queue any iocbs. Wi 1691 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2079 * fail with -ENOSYS if not implemented. 1692 * fail with -ENOSYS if not implemented.
2080 */ 1693 */
2081 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx 1694 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2082 struct iocb __user * __user * 1695 struct iocb __user * __user *, iocbpp)
2083 { 1696 {
2084 struct kioctx *ctx; !! 1697 return do_io_submit(ctx_id, nr, iocbpp, 0);
2085 long ret = 0; !! 1698 }
2086 int i = 0; <<
2087 struct blk_plug plug; <<
2088 1699
2089 if (unlikely(nr < 0)) !! 1700 #ifdef CONFIG_COMPAT
2090 return -EINVAL; !! 1701 static inline long
>> 1702 copy_iocb(long nr, u32 __user *ptr32, struct iocb __user * __user *ptr64)
>> 1703 {
>> 1704 compat_uptr_t uptr;
>> 1705 int i;
2091 1706
2092 ctx = lookup_ioctx(ctx_id); !! 1707 for (i = 0; i < nr; ++i) {
2093 if (unlikely(!ctx)) { !! 1708 if (get_user(uptr, ptr32 + i))
2094 pr_debug("EINVAL: invalid con !! 1709 return -EFAULT;
2095 return -EINVAL; !! 1710 if (put_user(compat_ptr(uptr), ptr64 + i))
>> 1711 return -EFAULT;
2096 } 1712 }
2097 !! 1713 return 0;
2098 if (nr > ctx->nr_events) <<
2099 nr = ctx->nr_events; <<
2100 <<
2101 if (nr > AIO_PLUG_THRESHOLD) <<
2102 blk_start_plug(&plug); <<
2103 for (i = 0; i < nr; i++) { <<
2104 struct iocb __user *user_iocb <<
2105 <<
2106 if (unlikely(get_user(user_io <<
2107 ret = -EFAULT; <<
2108 break; <<
2109 } <<
2110 <<
2111 ret = io_submit_one(ctx, user <<
2112 if (ret) <<
2113 break; <<
2114 } <<
2115 if (nr > AIO_PLUG_THRESHOLD) <<
2116 blk_finish_plug(&plug); <<
2117 <<
2118 percpu_ref_put(&ctx->users); <<
2119 return i ? i : ret; <<
2120 } 1714 }
2121 1715
2122 #ifdef CONFIG_COMPAT !! 1716 #define MAX_AIO_SUBMITS (PAGE_SIZE/sizeof(struct iocb *))
>> 1717
2123 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_ 1718 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2124 int, nr, compat_uptr_t !! 1719 int, nr, u32 __user *, iocb)
2125 { 1720 {
2126 struct kioctx *ctx; !! 1721 struct iocb __user * __user *iocb64;
2127 long ret = 0; !! 1722 long ret;
2128 int i = 0; <<
2129 struct blk_plug plug; <<
2130 1723
2131 if (unlikely(nr < 0)) 1724 if (unlikely(nr < 0))
2132 return -EINVAL; 1725 return -EINVAL;
2133 1726
2134 ctx = lookup_ioctx(ctx_id); !! 1727 if (nr > MAX_AIO_SUBMITS)
2135 if (unlikely(!ctx)) { !! 1728 nr = MAX_AIO_SUBMITS;
2136 pr_debug("EINVAL: invalid con <<
2137 return -EINVAL; <<
2138 } <<
2139 1729
2140 if (nr > ctx->nr_events) !! 1730 iocb64 = compat_alloc_user_space(nr * sizeof(*iocb64));
2141 nr = ctx->nr_events; !! 1731 ret = copy_iocb(nr, iocb, iocb64);
>> 1732 if (!ret)
>> 1733 ret = do_io_submit(ctx_id, nr, iocb64, 1);
>> 1734 return ret;
>> 1735 }
>> 1736 #endif
2142 1737
2143 if (nr > AIO_PLUG_THRESHOLD) !! 1738 /* lookup_kiocb
2144 blk_start_plug(&plug); !! 1739 * Finds a given iocb for cancellation.
2145 for (i = 0; i < nr; i++) { !! 1740 */
2146 compat_uptr_t user_iocb; !! 1741 static struct aio_kiocb *
>> 1742 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
>> 1743 {
>> 1744 struct aio_kiocb *kiocb;
2147 1745
2148 if (unlikely(get_user(user_io !! 1746 assert_spin_locked(&ctx->ctx_lock);
2149 ret = -EFAULT; <<
2150 break; <<
2151 } <<
2152 1747
2153 ret = io_submit_one(ctx, comp !! 1748 if (key != KIOCB_KEY)
2154 if (ret) !! 1749 return NULL;
2155 break; <<
2156 } <<
2157 if (nr > AIO_PLUG_THRESHOLD) <<
2158 blk_finish_plug(&plug); <<
2159 1750
2160 percpu_ref_put(&ctx->users); !! 1751 /* TODO: use a hash or array, this sucks. */
2161 return i ? i : ret; !! 1752 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
>> 1753 if (kiocb->ki_user_iocb == iocb)
>> 1754 return kiocb;
>> 1755 }
>> 1756 return NULL;
2162 } 1757 }
2163 #endif <<
2164 1758
2165 /* sys_io_cancel: 1759 /* sys_io_cancel:
2166 * Attempts to cancel an iocb previously 1760 * Attempts to cancel an iocb previously passed to io_submit. If
2167 * the operation is successfully cancell 1761 * the operation is successfully cancelled, the resulting event is
2168 * copied into the memory pointed to by 1762 * copied into the memory pointed to by result without being placed
2169 * into the completion queue and 0 is re 1763 * into the completion queue and 0 is returned. May fail with
2170 * -EFAULT if any of the data structures 1764 * -EFAULT if any of the data structures pointed to are invalid.
2171 * May fail with -EINVAL if aio_context 1765 * May fail with -EINVAL if aio_context specified by ctx_id is
2172 * invalid. May fail with -EAGAIN if th 1766 * invalid. May fail with -EAGAIN if the iocb specified was not
2173 * cancelled. Will fail with -ENOSYS if 1767 * cancelled. Will fail with -ENOSYS if not implemented.
2174 */ 1768 */
2175 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx 1769 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2176 struct io_event __user *, res 1770 struct io_event __user *, result)
2177 { 1771 {
2178 struct kioctx *ctx; 1772 struct kioctx *ctx;
2179 struct aio_kiocb *kiocb; 1773 struct aio_kiocb *kiocb;
2180 int ret = -EINVAL; <<
2181 u32 key; 1774 u32 key;
2182 u64 obj = (u64)(unsigned long)iocb; !! 1775 int ret;
2183 1776
2184 if (unlikely(get_user(key, &iocb->aio !! 1777 ret = get_user(key, &iocb->aio_key);
>> 1778 if (unlikely(ret))
2185 return -EFAULT; 1779 return -EFAULT;
2186 if (unlikely(key != KIOCB_KEY)) <<
2187 return -EINVAL; <<
2188 1780
2189 ctx = lookup_ioctx(ctx_id); 1781 ctx = lookup_ioctx(ctx_id);
2190 if (unlikely(!ctx)) 1782 if (unlikely(!ctx))
2191 return -EINVAL; 1783 return -EINVAL;
2192 1784
2193 spin_lock_irq(&ctx->ctx_lock); 1785 spin_lock_irq(&ctx->ctx_lock);
2194 /* TODO: use a hash or array, this su !! 1786
2195 list_for_each_entry(kiocb, &ctx->acti !! 1787 kiocb = lookup_kiocb(ctx, iocb, key);
2196 if (kiocb->ki_res.obj == obj) !! 1788 if (kiocb)
2197 ret = kiocb->ki_cance !! 1789 ret = kiocb_cancel(kiocb);
2198 list_del_init(&kiocb- !! 1790 else
2199 break; !! 1791 ret = -EINVAL;
2200 } !! 1792
2201 } <<
2202 spin_unlock_irq(&ctx->ctx_lock); 1793 spin_unlock_irq(&ctx->ctx_lock);
2203 1794
2204 if (!ret) { 1795 if (!ret) {
2205 /* 1796 /*
2206 * The result argument is no 1797 * The result argument is no longer used - the io_event is
2207 * always delivered via the r 1798 * always delivered via the ring buffer. -EINPROGRESS indicates
2208 * cancellation is progress: 1799 * cancellation is progress:
2209 */ 1800 */
2210 ret = -EINPROGRESS; 1801 ret = -EINPROGRESS;
2211 } 1802 }
2212 1803
2213 percpu_ref_put(&ctx->users); 1804 percpu_ref_put(&ctx->users);
2214 1805
2215 return ret; 1806 return ret;
2216 } 1807 }
2217 1808
2218 static long do_io_getevents(aio_context_t ctx <<
2219 long min_nr, <<
2220 long nr, <<
2221 struct io_event __user *event <<
2222 struct timespec64 *ts) <<
2223 { <<
2224 ktime_t until = ts ? timespec64_to_kt <<
2225 struct kioctx *ioctx = lookup_ioctx(c <<
2226 long ret = -EINVAL; <<
2227 <<
2228 if (likely(ioctx)) { <<
2229 if (likely(min_nr <= nr && mi <<
2230 ret = read_events(ioc <<
2231 percpu_ref_put(&ioctx->users) <<
2232 } <<
2233 <<
2234 return ret; <<
2235 } <<
2236 <<
2237 /* io_getevents: 1809 /* io_getevents:
2238 * Attempts to read at least min_nr even 1810 * Attempts to read at least min_nr events and up to nr events from
2239 * the completion queue for the aio_cont 1811 * the completion queue for the aio_context specified by ctx_id. If
2240 * it succeeds, the number of read event 1812 * it succeeds, the number of read events is returned. May fail with
2241 * -EINVAL if ctx_id is invalid, if min_ 1813 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2242 * out of range, if timeout is out of ra 1814 * out of range, if timeout is out of range. May fail with -EFAULT
2243 * if any of the memory specified is inv 1815 * if any of the memory specified is invalid. May return 0 or
2244 * < min_nr if the timeout specified by 1816 * < min_nr if the timeout specified by timeout has elapsed
2245 * before sufficient events are availabl 1817 * before sufficient events are available, where timeout == NULL
2246 * specifies an infinite timeout. Note t 1818 * specifies an infinite timeout. Note that the timeout pointed to by
2247 * timeout is relative. Will fail with 1819 * timeout is relative. Will fail with -ENOSYS if not implemented.
2248 */ 1820 */
2249 #ifdef CONFIG_64BIT <<
2250 <<
2251 SYSCALL_DEFINE5(io_getevents, aio_context_t, 1821 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2252 long, min_nr, 1822 long, min_nr,
2253 long, nr, 1823 long, nr,
2254 struct io_event __user *, eve 1824 struct io_event __user *, events,
2255 struct __kernel_timespec __us !! 1825 struct timespec __user *, timeout)
2256 { <<
2257 struct timespec64 ts; <<
2258 int ret; <<
2259 <<
2260 if (timeout && unlikely(get_timespec6 <<
2261 return -EFAULT; <<
2262 <<
2263 ret = do_io_getevents(ctx_id, min_nr, <<
2264 if (!ret && signal_pending(current)) <<
2265 ret = -EINTR; <<
2266 return ret; <<
2267 } <<
2268 <<
2269 #endif <<
2270 <<
2271 struct __aio_sigset { <<
2272 const sigset_t __user *sigmask; <<
2273 size_t sigsetsize; <<
2274 }; <<
2275 <<
2276 SYSCALL_DEFINE6(io_pgetevents, <<
2277 aio_context_t, ctx_id, <<
2278 long, min_nr, <<
2279 long, nr, <<
2280 struct io_event __user *, eve <<
2281 struct __kernel_timespec __us <<
2282 const struct __aio_sigset __u <<
2283 { <<
2284 struct __aio_sigset ksig = { NULL <<
2285 struct timespec64 ts; <<
2286 bool interrupted; <<
2287 int ret; <<
2288 <<
2289 if (timeout && unlikely(get_timespec6 <<
2290 return -EFAULT; <<
2291 <<
2292 if (usig && copy_from_user(&ksig, usi <<
2293 return -EFAULT; <<
2294 <<
2295 ret = set_user_sigmask(ksig.sigmask, <<
2296 if (ret) <<
2297 return ret; <<
2298 <<
2299 ret = do_io_getevents(ctx_id, min_nr, <<
2300 <<
2301 interrupted = signal_pending(current) <<
2302 restore_saved_sigmask_unless(interrup <<
2303 if (interrupted && !ret) <<
2304 ret = -ERESTARTNOHAND; <<
2305 <<
2306 return ret; <<
2307 } <<
2308 <<
2309 #if defined(CONFIG_COMPAT_32BIT_TIME) && !def <<
2310 <<
2311 SYSCALL_DEFINE6(io_pgetevents_time32, <<
2312 aio_context_t, ctx_id, <<
2313 long, min_nr, <<
2314 long, nr, <<
2315 struct io_event __user *, eve <<
2316 struct old_timespec32 __user <<
2317 const struct __aio_sigset __u <<
2318 { 1826 {
2319 struct __aio_sigset ksig = { NULL !! 1827 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2320 struct timespec64 ts; !! 1828 long ret = -EINVAL;
2321 bool interrupted; <<
2322 int ret; <<
2323 <<
2324 if (timeout && unlikely(get_old_times <<
2325 return -EFAULT; <<
2326 <<
2327 if (usig && copy_from_user(&ksig, usi <<
2328 return -EFAULT; <<
2329 <<
2330 <<
2331 ret = set_user_sigmask(ksig.sigmask, <<
2332 if (ret) <<
2333 return ret; <<
2334 <<
2335 ret = do_io_getevents(ctx_id, min_nr, <<
2336 <<
2337 interrupted = signal_pending(current) <<
2338 restore_saved_sigmask_unless(interrup <<
2339 if (interrupted && !ret) <<
2340 ret = -ERESTARTNOHAND; <<
2341 <<
2342 return ret; <<
2343 } <<
2344 <<
2345 #endif <<
2346 <<
2347 #if defined(CONFIG_COMPAT_32BIT_TIME) <<
2348 <<
2349 SYSCALL_DEFINE5(io_getevents_time32, __u32, c <<
2350 __s32, min_nr, <<
2351 __s32, nr, <<
2352 struct io_event __user *, eve <<
2353 struct old_timespec32 __user <<
2354 { <<
2355 struct timespec64 t; <<
2356 int ret; <<
2357 <<
2358 if (timeout && get_old_timespec32(&t, <<
2359 return -EFAULT; <<
2360 1829
2361 ret = do_io_getevents(ctx_id, min_nr, !! 1830 if (likely(ioctx)) {
2362 if (!ret && signal_pending(current)) !! 1831 if (likely(min_nr <= nr && min_nr >= 0))
2363 ret = -EINTR; !! 1832 ret = read_events(ioctx, min_nr, nr, events, timeout);
>> 1833 percpu_ref_put(&ioctx->users);
>> 1834 }
2364 return ret; 1835 return ret;
2365 } 1836 }
2366 1837
2367 #endif <<
2368 <<
2369 #ifdef CONFIG_COMPAT 1838 #ifdef CONFIG_COMPAT
2370 !! 1839 COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id,
2371 struct __compat_aio_sigset { !! 1840 compat_long_t, min_nr,
2372 compat_uptr_t sigmask; !! 1841 compat_long_t, nr,
2373 compat_size_t sigsetsize; !! 1842 struct io_event __user *, events,
2374 }; !! 1843 struct compat_timespec __user *, timeout)
2375 !! 1844 {
2376 #if defined(CONFIG_COMPAT_32BIT_TIME) !! 1845 struct timespec t;
2377 !! 1846 struct timespec __user *ut = NULL;
2378 COMPAT_SYSCALL_DEFINE6(io_pgetevents, !! 1847
2379 compat_aio_context_t, ctx_id, !! 1848 if (timeout) {
2380 compat_long_t, min_nr, !! 1849 if (compat_get_timespec(&t, timeout))
2381 compat_long_t, nr, !! 1850 return -EFAULT;
2382 struct io_event __user *, eve !! 1851
2383 struct old_timespec32 __user !! 1852 ut = compat_alloc_user_space(sizeof(*ut));
2384 const struct __compat_aio_sig !! 1853 if (copy_to_user(ut, &t, sizeof(t)))
2385 { !! 1854 return -EFAULT;
2386 struct __compat_aio_sigset ksig = { 0 !! 1855 }
2387 struct timespec64 t; !! 1856 return sys_io_getevents(ctx_id, min_nr, nr, events, ut);
2388 bool interrupted; <<
2389 int ret; <<
2390 <<
2391 if (timeout && get_old_timespec32(&t, <<
2392 return -EFAULT; <<
2393 <<
2394 if (usig && copy_from_user(&ksig, usi <<
2395 return -EFAULT; <<
2396 <<
2397 ret = set_compat_user_sigmask(compat_ <<
2398 if (ret) <<
2399 return ret; <<
2400 <<
2401 ret = do_io_getevents(ctx_id, min_nr, <<
2402 <<
2403 interrupted = signal_pending(current) <<
2404 restore_saved_sigmask_unless(interrup <<
2405 if (interrupted && !ret) <<
2406 ret = -ERESTARTNOHAND; <<
2407 <<
2408 return ret; <<
2409 } <<
2410 <<
2411 #endif <<
2412 <<
2413 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64, <<
2414 compat_aio_context_t, ctx_id, <<
2415 compat_long_t, min_nr, <<
2416 compat_long_t, nr, <<
2417 struct io_event __user *, eve <<
2418 struct __kernel_timespec __us <<
2419 const struct __compat_aio_sig <<
2420 { <<
2421 struct __compat_aio_sigset ksig = { 0 <<
2422 struct timespec64 t; <<
2423 bool interrupted; <<
2424 int ret; <<
2425 <<
2426 if (timeout && get_timespec64(&t, tim <<
2427 return -EFAULT; <<
2428 <<
2429 if (usig && copy_from_user(&ksig, usi <<
2430 return -EFAULT; <<
2431 <<
2432 ret = set_compat_user_sigmask(compat_ <<
2433 if (ret) <<
2434 return ret; <<
2435 <<
2436 ret = do_io_getevents(ctx_id, min_nr, <<
2437 <<
2438 interrupted = signal_pending(current) <<
2439 restore_saved_sigmask_unless(interrup <<
2440 if (interrupted && !ret) <<
2441 ret = -ERESTARTNOHAND; <<
2442 <<
2443 return ret; <<
2444 } 1857 }
2445 #endif 1858 #endif
2446 1859
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.