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. 8 * Copyright 2018 Christoph Hellwig.
9 * 9 *
10 * See ../COPYING for licensing terms. 10 * See ../COPYING for licensing terms.
11 */ 11 */
12 #define pr_fmt(fmt) "%s: " fmt, __func__ 12 #define pr_fmt(fmt) "%s: " fmt, __func__
13 13
14 #include <linux/kernel.h> 14 #include <linux/kernel.h>
15 #include <linux/init.h> 15 #include <linux/init.h>
16 #include <linux/errno.h> 16 #include <linux/errno.h>
17 #include <linux/time.h> 17 #include <linux/time.h>
18 #include <linux/aio_abi.h> 18 #include <linux/aio_abi.h>
19 #include <linux/export.h> 19 #include <linux/export.h>
20 #include <linux/syscalls.h> 20 #include <linux/syscalls.h>
21 #include <linux/backing-dev.h> 21 #include <linux/backing-dev.h>
22 #include <linux/refcount.h> 22 #include <linux/refcount.h>
23 #include <linux/uio.h> 23 #include <linux/uio.h>
24 24
25 #include <linux/sched/signal.h> 25 #include <linux/sched/signal.h>
26 #include <linux/fs.h> 26 #include <linux/fs.h>
27 #include <linux/file.h> 27 #include <linux/file.h>
28 #include <linux/mm.h> 28 #include <linux/mm.h>
29 #include <linux/mman.h> 29 #include <linux/mman.h>
>> 30 #include <linux/mmu_context.h>
30 #include <linux/percpu.h> 31 #include <linux/percpu.h>
31 #include <linux/slab.h> 32 #include <linux/slab.h>
32 #include <linux/timer.h> 33 #include <linux/timer.h>
33 #include <linux/aio.h> 34 #include <linux/aio.h>
34 #include <linux/highmem.h> 35 #include <linux/highmem.h>
35 #include <linux/workqueue.h> 36 #include <linux/workqueue.h>
36 #include <linux/security.h> 37 #include <linux/security.h>
37 #include <linux/eventfd.h> 38 #include <linux/eventfd.h>
38 #include <linux/blkdev.h> 39 #include <linux/blkdev.h>
39 #include <linux/compat.h> 40 #include <linux/compat.h>
40 #include <linux/migrate.h> 41 #include <linux/migrate.h>
41 #include <linux/ramfs.h> 42 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.h> 43 #include <linux/percpu-refcount.h>
43 #include <linux/mount.h> 44 #include <linux/mount.h>
44 #include <linux/pseudo_fs.h> <<
45 45
>> 46 #include <asm/kmap_types.h>
46 #include <linux/uaccess.h> 47 #include <linux/uaccess.h>
47 #include <linux/nospec.h> 48 #include <linux/nospec.h>
48 49
49 #include "internal.h" 50 #include "internal.h"
50 51
51 #define KIOCB_KEY 0 52 #define KIOCB_KEY 0
52 53
53 #define AIO_RING_MAGIC 0xa10a 54 #define AIO_RING_MAGIC 0xa10a10a1
54 #define AIO_RING_COMPAT_FEATURES 1 55 #define AIO_RING_COMPAT_FEATURES 1
55 #define AIO_RING_INCOMPAT_FEATURES 0 56 #define AIO_RING_INCOMPAT_FEATURES 0
56 struct aio_ring { 57 struct aio_ring {
57 unsigned id; /* kernel inte 58 unsigned id; /* kernel internal index number */
58 unsigned nr; /* number of i 59 unsigned nr; /* number of io_events */
59 unsigned head; /* Written to 60 unsigned head; /* Written to by userland or under ring_lock
60 * mutex by ai 61 * mutex by aio_read_events_ring(). */
61 unsigned tail; 62 unsigned tail;
62 63
63 unsigned magic; 64 unsigned magic;
64 unsigned compat_features; 65 unsigned compat_features;
65 unsigned incompat_features; 66 unsigned incompat_features;
66 unsigned header_length; /* siz 67 unsigned header_length; /* size of aio_ring */
67 68
68 69
69 struct io_event io_events[]; !! 70 struct io_event io_events[0];
70 }; /* 128 bytes + ring size */ 71 }; /* 128 bytes + ring size */
71 72
72 /* 73 /*
73 * Plugging is meant to work with larger batch 74 * Plugging is meant to work with larger batches of IOs. If we don't
74 * have more than the below, then don't bother 75 * have more than the below, then don't bother setting up a plug.
75 */ 76 */
76 #define AIO_PLUG_THRESHOLD 2 77 #define AIO_PLUG_THRESHOLD 2
77 78
78 #define AIO_RING_PAGES 8 79 #define AIO_RING_PAGES 8
79 80
80 struct kioctx_table { 81 struct kioctx_table {
81 struct rcu_head rcu; 82 struct rcu_head rcu;
82 unsigned nr; 83 unsigned nr;
83 struct kioctx __rcu *table[] __cou !! 84 struct kioctx __rcu *table[];
84 }; 85 };
85 86
86 struct kioctx_cpu { 87 struct kioctx_cpu {
87 unsigned reqs_available 88 unsigned reqs_available;
88 }; 89 };
89 90
90 struct ctx_rq_wait { 91 struct ctx_rq_wait {
91 struct completion comp; 92 struct completion comp;
92 atomic_t count; 93 atomic_t count;
93 }; 94 };
94 95
95 struct kioctx { 96 struct kioctx {
96 struct percpu_ref users; 97 struct percpu_ref users;
97 atomic_t dead; 98 atomic_t dead;
98 99
99 struct percpu_ref reqs; 100 struct percpu_ref reqs;
100 101
101 unsigned long user_id; 102 unsigned long user_id;
102 103
103 struct kioctx_cpu __percpu *cpu; !! 104 struct __percpu kioctx_cpu *cpu;
104 105
105 /* 106 /*
106 * For percpu reqs_available, number o 107 * For percpu reqs_available, number of slots we move to/from global
107 * counter at a time: 108 * counter at a time:
108 */ 109 */
109 unsigned req_batch; 110 unsigned req_batch;
110 /* 111 /*
111 * This is what userspace passed to io 112 * This is what userspace passed to io_setup(), it's not used for
112 * anything but counting against the g 113 * anything but counting against the global max_reqs quota.
113 * 114 *
114 * The real limit is nr_events - 1, wh 115 * The real limit is nr_events - 1, which will be larger (see
115 * aio_setup_ring()) 116 * aio_setup_ring())
116 */ 117 */
117 unsigned max_reqs; 118 unsigned max_reqs;
118 119
119 /* Size of ringbuffer, in units of str 120 /* Size of ringbuffer, in units of struct io_event */
120 unsigned nr_events; 121 unsigned nr_events;
121 122
122 unsigned long mmap_base; 123 unsigned long mmap_base;
123 unsigned long mmap_size; 124 unsigned long mmap_size;
124 125
125 struct folio **ring_folios; !! 126 struct page **ring_pages;
126 long nr_pages; 127 long nr_pages;
127 128
128 struct rcu_work free_rwork; 129 struct rcu_work free_rwork; /* see free_ioctx() */
129 130
130 /* 131 /*
131 * signals when all in-flight requests 132 * signals when all in-flight requests are done
132 */ 133 */
133 struct ctx_rq_wait *rq_wait; 134 struct ctx_rq_wait *rq_wait;
134 135
135 struct { 136 struct {
136 /* 137 /*
137 * This counts the number of a 138 * This counts the number of available slots in the ringbuffer,
138 * so we avoid overflowing it: 139 * so we avoid overflowing it: it's decremented (if positive)
139 * when allocating a kiocb and 140 * when allocating a kiocb and incremented when the resulting
140 * io_event is pulled off the 141 * io_event is pulled off the ringbuffer.
141 * 142 *
142 * We batch accesses to it wit 143 * We batch accesses to it with a percpu version.
143 */ 144 */
144 atomic_t reqs_available 145 atomic_t reqs_available;
145 } ____cacheline_aligned_in_smp; 146 } ____cacheline_aligned_in_smp;
146 147
147 struct { 148 struct {
148 spinlock_t ctx_lock; 149 spinlock_t ctx_lock;
149 struct list_head active_reqs; 150 struct list_head active_reqs; /* used for cancellation */
150 } ____cacheline_aligned_in_smp; 151 } ____cacheline_aligned_in_smp;
151 152
152 struct { 153 struct {
153 struct mutex ring_lock; 154 struct mutex ring_lock;
154 wait_queue_head_t wait; 155 wait_queue_head_t wait;
155 } ____cacheline_aligned_in_smp; 156 } ____cacheline_aligned_in_smp;
156 157
157 struct { 158 struct {
158 unsigned tail; 159 unsigned tail;
159 unsigned completed_even 160 unsigned completed_events;
160 spinlock_t completion_loc 161 spinlock_t completion_lock;
161 } ____cacheline_aligned_in_smp; 162 } ____cacheline_aligned_in_smp;
162 163
163 struct folio *internal_foli !! 164 struct page *internal_pages[AIO_RING_PAGES];
164 struct file *aio_ring_file 165 struct file *aio_ring_file;
165 166
166 unsigned id; 167 unsigned id;
167 }; 168 };
168 169
169 /* 170 /*
170 * First field must be the file pointer in all 171 * First field must be the file pointer in all the
171 * iocb unions! See also 'struct kiocb' in <li 172 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
172 */ 173 */
173 struct fsync_iocb { 174 struct fsync_iocb {
174 struct file *file; 175 struct file *file;
175 struct work_struct work; 176 struct work_struct work;
176 bool datasync; 177 bool datasync;
177 struct cred *creds; <<
178 }; 178 };
179 179
180 struct poll_iocb { 180 struct poll_iocb {
181 struct file *file; 181 struct file *file;
182 struct wait_queue_head *head; 182 struct wait_queue_head *head;
183 __poll_t events; 183 __poll_t events;
>> 184 bool done;
184 bool cancelled; 185 bool cancelled;
185 bool work_scheduled <<
186 bool work_need_resc <<
187 struct wait_queue_entry wait; 186 struct wait_queue_entry wait;
188 struct work_struct work; 187 struct work_struct work;
189 }; 188 };
190 189
191 /* 190 /*
192 * NOTE! Each of the iocb union members has th 191 * NOTE! Each of the iocb union members has the file pointer
193 * as the first entry in their struct definiti 192 * as the first entry in their struct definition. So you can
194 * access the file pointer through any of the 193 * access the file pointer through any of the sub-structs,
195 * or directly as just 'ki_filp' in this struc 194 * or directly as just 'ki_filp' in this struct.
196 */ 195 */
197 struct aio_kiocb { 196 struct aio_kiocb {
198 union { 197 union {
199 struct file *ki_fi 198 struct file *ki_filp;
200 struct kiocb rw; 199 struct kiocb rw;
201 struct fsync_iocb fsync; 200 struct fsync_iocb fsync;
202 struct poll_iocb poll; 201 struct poll_iocb poll;
203 }; 202 };
204 203
205 struct kioctx *ki_ctx; 204 struct kioctx *ki_ctx;
206 kiocb_cancel_fn *ki_cancel; 205 kiocb_cancel_fn *ki_cancel;
207 206
208 struct io_event ki_res; 207 struct io_event ki_res;
209 208
210 struct list_head ki_list; 209 struct list_head ki_list; /* the aio core uses this
211 210 * for cancellation */
212 refcount_t ki_refcnt; 211 refcount_t ki_refcnt;
213 212
214 /* 213 /*
215 * If the aio_resfd field of the users 214 * If the aio_resfd field of the userspace iocb is not zero,
216 * this is the underlying eventfd cont 215 * this is the underlying eventfd context to deliver events to.
217 */ 216 */
218 struct eventfd_ctx *ki_eventfd; 217 struct eventfd_ctx *ki_eventfd;
219 }; 218 };
220 219
221 /*------ sysctl variables----*/ 220 /*------ sysctl variables----*/
222 static DEFINE_SPINLOCK(aio_nr_lock); 221 static DEFINE_SPINLOCK(aio_nr_lock);
223 static unsigned long aio_nr; /* cur !! 222 unsigned long aio_nr; /* current system wide number of aio requests */
224 static unsigned long aio_max_nr = 0x10000; /* !! 223 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/ 224 /*----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 <<
251 225
252 static struct kmem_cache *kiocb_cachep; 226 static struct kmem_cache *kiocb_cachep;
253 static struct kmem_cache *kioctx_cachep 227 static struct kmem_cache *kioctx_cachep;
254 228
255 static struct vfsmount *aio_mnt; 229 static struct vfsmount *aio_mnt;
256 230
257 static const struct file_operations aio_ring_f 231 static const struct file_operations aio_ring_fops;
258 static const struct address_space_operations a 232 static const struct address_space_operations aio_ctx_aops;
259 233
260 static struct file *aio_private_file(struct ki 234 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
261 { 235 {
262 struct file *file; 236 struct file *file;
263 struct inode *inode = alloc_anon_inode 237 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
264 if (IS_ERR(inode)) 238 if (IS_ERR(inode))
265 return ERR_CAST(inode); 239 return ERR_CAST(inode);
266 240
267 inode->i_mapping->a_ops = &aio_ctx_aop 241 inode->i_mapping->a_ops = &aio_ctx_aops;
268 inode->i_mapping->i_private_data = ctx !! 242 inode->i_mapping->private_data = ctx;
269 inode->i_size = PAGE_SIZE * nr_pages; 243 inode->i_size = PAGE_SIZE * nr_pages;
270 244
271 file = alloc_file_pseudo(inode, aio_mn 245 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
272 O_RDWR, &aio_r 246 O_RDWR, &aio_ring_fops);
273 if (IS_ERR(file)) 247 if (IS_ERR(file))
274 iput(inode); 248 iput(inode);
275 return file; 249 return file;
276 } 250 }
277 251
278 static int aio_init_fs_context(struct fs_conte !! 252 static struct dentry *aio_mount(struct file_system_type *fs_type,
>> 253 int flags, const char *dev_name, void *data)
279 { 254 {
280 if (!init_pseudo(fc, AIO_RING_MAGIC)) !! 255 struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, NULL,
281 return -ENOMEM; !! 256 AIO_RING_MAGIC);
282 fc->s_iflags |= SB_I_NOEXEC; !! 257
283 return 0; !! 258 if (!IS_ERR(root))
>> 259 root->d_sb->s_iflags |= SB_I_NOEXEC;
>> 260 return root;
284 } 261 }
285 262
286 /* aio_setup 263 /* aio_setup
287 * Creates the slab caches used by the ai 264 * Creates the slab caches used by the aio routines, panic on
288 * failure as this is done early during t 265 * failure as this is done early during the boot sequence.
289 */ 266 */
290 static int __init aio_setup(void) 267 static int __init aio_setup(void)
291 { 268 {
292 static struct file_system_type aio_fs 269 static struct file_system_type aio_fs = {
293 .name = "aio", 270 .name = "aio",
294 .init_fs_context = aio_init_fs !! 271 .mount = aio_mount,
295 .kill_sb = kill_anon_su 272 .kill_sb = kill_anon_super,
296 }; 273 };
297 aio_mnt = kern_mount(&aio_fs); 274 aio_mnt = kern_mount(&aio_fs);
298 if (IS_ERR(aio_mnt)) 275 if (IS_ERR(aio_mnt))
299 panic("Failed to create aio fs 276 panic("Failed to create aio fs mount.");
300 277
301 kiocb_cachep = KMEM_CACHE(aio_kiocb, S 278 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
302 kioctx_cachep = KMEM_CACHE(kioctx,SLAB 279 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
303 aio_sysctl_init(); <<
304 return 0; 280 return 0;
305 } 281 }
306 __initcall(aio_setup); 282 __initcall(aio_setup);
307 283
308 static void put_aio_ring_file(struct kioctx *c 284 static void put_aio_ring_file(struct kioctx *ctx)
309 { 285 {
310 struct file *aio_ring_file = ctx->aio_ 286 struct file *aio_ring_file = ctx->aio_ring_file;
311 struct address_space *i_mapping; 287 struct address_space *i_mapping;
312 288
313 if (aio_ring_file) { 289 if (aio_ring_file) {
314 truncate_setsize(file_inode(ai 290 truncate_setsize(file_inode(aio_ring_file), 0);
315 291
316 /* Prevent further access to t 292 /* Prevent further access to the kioctx from migratepages */
317 i_mapping = aio_ring_file->f_m 293 i_mapping = aio_ring_file->f_mapping;
318 spin_lock(&i_mapping->i_privat !! 294 spin_lock(&i_mapping->private_lock);
319 i_mapping->i_private_data = NU !! 295 i_mapping->private_data = NULL;
320 ctx->aio_ring_file = NULL; 296 ctx->aio_ring_file = NULL;
321 spin_unlock(&i_mapping->i_priv !! 297 spin_unlock(&i_mapping->private_lock);
322 298
323 fput(aio_ring_file); 299 fput(aio_ring_file);
324 } 300 }
325 } 301 }
326 302
327 static void aio_free_ring(struct kioctx *ctx) 303 static void aio_free_ring(struct kioctx *ctx)
328 { 304 {
329 int i; 305 int i;
330 306
331 /* Disconnect the kiotx from the ring 307 /* Disconnect the kiotx from the ring file. This prevents future
332 * accesses to the kioctx from page mi 308 * accesses to the kioctx from page migration.
333 */ 309 */
334 put_aio_ring_file(ctx); 310 put_aio_ring_file(ctx);
335 311
336 for (i = 0; i < ctx->nr_pages; i++) { 312 for (i = 0; i < ctx->nr_pages; i++) {
337 struct folio *folio = ctx->rin !! 313 struct page *page;
338 !! 314 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
339 if (!folio) !! 315 page_count(ctx->ring_pages[i]));
>> 316 page = ctx->ring_pages[i];
>> 317 if (!page)
340 continue; 318 continue;
341 !! 319 ctx->ring_pages[i] = NULL;
342 pr_debug("pid(%d) [%d] folio-> !! 320 put_page(page);
343 folio_ref_count(folio <<
344 ctx->ring_folios[i] = NULL; <<
345 folio_put(folio); <<
346 } 321 }
347 322
348 if (ctx->ring_folios && ctx->ring_foli !! 323 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
349 kfree(ctx->ring_folios); !! 324 kfree(ctx->ring_pages);
350 ctx->ring_folios = NULL; !! 325 ctx->ring_pages = NULL;
351 } 326 }
352 } 327 }
353 328
354 static int aio_ring_mremap(struct vm_area_stru 329 static int aio_ring_mremap(struct vm_area_struct *vma)
355 { 330 {
356 struct file *file = vma->vm_file; 331 struct file *file = vma->vm_file;
357 struct mm_struct *mm = vma->vm_mm; 332 struct mm_struct *mm = vma->vm_mm;
358 struct kioctx_table *table; 333 struct kioctx_table *table;
359 int i, res = -EINVAL; 334 int i, res = -EINVAL;
360 335
361 spin_lock(&mm->ioctx_lock); 336 spin_lock(&mm->ioctx_lock);
362 rcu_read_lock(); 337 rcu_read_lock();
363 table = rcu_dereference(mm->ioctx_tabl 338 table = rcu_dereference(mm->ioctx_table);
364 if (!table) <<
365 goto out_unlock; <<
366 <<
367 for (i = 0; i < table->nr; i++) { 339 for (i = 0; i < table->nr; i++) {
368 struct kioctx *ctx; 340 struct kioctx *ctx;
369 341
370 ctx = rcu_dereference(table->t 342 ctx = rcu_dereference(table->table[i]);
371 if (ctx && ctx->aio_ring_file 343 if (ctx && ctx->aio_ring_file == file) {
372 if (!atomic_read(&ctx- 344 if (!atomic_read(&ctx->dead)) {
373 ctx->user_id = 345 ctx->user_id = ctx->mmap_base = vma->vm_start;
374 res = 0; 346 res = 0;
375 } 347 }
376 break; 348 break;
377 } 349 }
378 } 350 }
379 351
380 out_unlock: <<
381 rcu_read_unlock(); 352 rcu_read_unlock();
382 spin_unlock(&mm->ioctx_lock); 353 spin_unlock(&mm->ioctx_lock);
383 return res; 354 return res;
384 } 355 }
385 356
386 static const struct vm_operations_struct aio_r 357 static const struct vm_operations_struct aio_ring_vm_ops = {
387 .mremap = aio_ring_mremap, 358 .mremap = aio_ring_mremap,
388 #if IS_ENABLED(CONFIG_MMU) 359 #if IS_ENABLED(CONFIG_MMU)
389 .fault = filemap_fault, 360 .fault = filemap_fault,
390 .map_pages = filemap_map_pages, 361 .map_pages = filemap_map_pages,
391 .page_mkwrite = filemap_page_mkwrite 362 .page_mkwrite = filemap_page_mkwrite,
392 #endif 363 #endif
393 }; 364 };
394 365
395 static int aio_ring_mmap(struct file *file, st 366 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
396 { 367 {
397 vm_flags_set(vma, VM_DONTEXPAND); !! 368 vma->vm_flags |= VM_DONTEXPAND;
398 vma->vm_ops = &aio_ring_vm_ops; 369 vma->vm_ops = &aio_ring_vm_ops;
399 return 0; 370 return 0;
400 } 371 }
401 372
402 static const struct file_operations aio_ring_f 373 static const struct file_operations aio_ring_fops = {
403 .mmap = aio_ring_mmap, 374 .mmap = aio_ring_mmap,
404 }; 375 };
405 376
406 #if IS_ENABLED(CONFIG_MIGRATION) 377 #if IS_ENABLED(CONFIG_MIGRATION)
407 static int aio_migrate_folio(struct address_sp !! 378 static int aio_migratepage(struct address_space *mapping, struct page *new,
408 struct folio *src, enu !! 379 struct page *old, enum migrate_mode mode)
409 { 380 {
410 struct kioctx *ctx; 381 struct kioctx *ctx;
411 unsigned long flags; 382 unsigned long flags;
412 pgoff_t idx; 383 pgoff_t idx;
413 int rc = 0; !! 384 int rc;
>> 385
>> 386 /*
>> 387 * We cannot support the _NO_COPY case here, because copy needs to
>> 388 * happen under the ctx->completion_lock. That does not work with the
>> 389 * migration workflow of MIGRATE_SYNC_NO_COPY.
>> 390 */
>> 391 if (mode == MIGRATE_SYNC_NO_COPY)
>> 392 return -EINVAL;
>> 393
>> 394 rc = 0;
414 395
415 /* mapping->i_private_lock here protec !! 396 /* mapping->private_lock here protects against the kioctx teardown. */
416 spin_lock(&mapping->i_private_lock); !! 397 spin_lock(&mapping->private_lock);
417 ctx = mapping->i_private_data; !! 398 ctx = mapping->private_data;
418 if (!ctx) { 399 if (!ctx) {
419 rc = -EINVAL; 400 rc = -EINVAL;
420 goto out; 401 goto out;
421 } 402 }
422 403
423 /* The ring_lock mutex. The prevents 404 /* The ring_lock mutex. The prevents aio_read_events() from writing
424 * to the ring's head, and prevents pa 405 * to the ring's head, and prevents page migration from mucking in
425 * a partially initialized kiotx. 406 * a partially initialized kiotx.
426 */ 407 */
427 if (!mutex_trylock(&ctx->ring_lock)) { 408 if (!mutex_trylock(&ctx->ring_lock)) {
428 rc = -EAGAIN; 409 rc = -EAGAIN;
429 goto out; 410 goto out;
430 } 411 }
431 412
432 idx = src->index; !! 413 idx = old->index;
433 if (idx < (pgoff_t)ctx->nr_pages) { 414 if (idx < (pgoff_t)ctx->nr_pages) {
434 /* Make sure the old folio has !! 415 /* Make sure the old page hasn't already been changed */
435 if (ctx->ring_folios[idx] != s !! 416 if (ctx->ring_pages[idx] != old)
436 rc = -EAGAIN; 417 rc = -EAGAIN;
437 } else 418 } else
438 rc = -EINVAL; 419 rc = -EINVAL;
439 420
440 if (rc != 0) 421 if (rc != 0)
441 goto out_unlock; 422 goto out_unlock;
442 423
443 /* Writeback must be complete */ 424 /* Writeback must be complete */
444 BUG_ON(folio_test_writeback(src)); !! 425 BUG_ON(PageWriteback(old));
445 folio_get(dst); !! 426 get_page(new);
446 427
447 rc = folio_migrate_mapping(mapping, ds !! 428 rc = migrate_page_move_mapping(mapping, new, old, mode, 1);
448 if (rc != MIGRATEPAGE_SUCCESS) { 429 if (rc != MIGRATEPAGE_SUCCESS) {
449 folio_put(dst); !! 430 put_page(new);
450 goto out_unlock; 431 goto out_unlock;
451 } 432 }
452 433
453 /* Take completion_lock to prevent oth 434 /* Take completion_lock to prevent other writes to the ring buffer
454 * while the old folio is copied to th !! 435 * while the old page is copied to the new. This prevents new
455 * events from being lost. 436 * events from being lost.
456 */ 437 */
457 spin_lock_irqsave(&ctx->completion_loc 438 spin_lock_irqsave(&ctx->completion_lock, flags);
458 folio_copy(dst, src); !! 439 migrate_page_copy(new, old);
459 folio_migrate_flags(dst, src); !! 440 BUG_ON(ctx->ring_pages[idx] != old);
460 BUG_ON(ctx->ring_folios[idx] != src); !! 441 ctx->ring_pages[idx] = new;
461 ctx->ring_folios[idx] = dst; <<
462 spin_unlock_irqrestore(&ctx->completio 442 spin_unlock_irqrestore(&ctx->completion_lock, flags);
463 443
464 /* The old folio is no longer accessib !! 444 /* The old page is no longer accessible. */
465 folio_put(src); !! 445 put_page(old);
466 446
467 out_unlock: 447 out_unlock:
468 mutex_unlock(&ctx->ring_lock); 448 mutex_unlock(&ctx->ring_lock);
469 out: 449 out:
470 spin_unlock(&mapping->i_private_lock); !! 450 spin_unlock(&mapping->private_lock);
471 return rc; 451 return rc;
472 } 452 }
473 #else <<
474 #define aio_migrate_folio NULL <<
475 #endif 453 #endif
476 454
477 static const struct address_space_operations a 455 static const struct address_space_operations aio_ctx_aops = {
478 .dirty_folio = noop_dirty_folio, !! 456 .set_page_dirty = __set_page_dirty_no_writeback,
479 .migrate_folio = aio_migrate_folio, !! 457 #if IS_ENABLED(CONFIG_MIGRATION)
>> 458 .migratepage = aio_migratepage,
>> 459 #endif
480 }; 460 };
481 461
482 static int aio_setup_ring(struct kioctx *ctx, 462 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
483 { 463 {
484 struct aio_ring *ring; 464 struct aio_ring *ring;
485 struct mm_struct *mm = current->mm; 465 struct mm_struct *mm = current->mm;
486 unsigned long size, unused; 466 unsigned long size, unused;
487 int nr_pages; 467 int nr_pages;
488 int i; 468 int i;
489 struct file *file; 469 struct file *file;
490 470
491 /* Compensate for the ring buffer's he 471 /* Compensate for the ring buffer's head/tail overlap entry */
492 nr_events += 2; /* 1 is required, 2 fo 472 nr_events += 2; /* 1 is required, 2 for good luck */
493 473
494 size = sizeof(struct aio_ring); 474 size = sizeof(struct aio_ring);
495 size += sizeof(struct io_event) * nr_e 475 size += sizeof(struct io_event) * nr_events;
496 476
497 nr_pages = PFN_UP(size); 477 nr_pages = PFN_UP(size);
498 if (nr_pages < 0) 478 if (nr_pages < 0)
499 return -EINVAL; 479 return -EINVAL;
500 480
501 file = aio_private_file(ctx, nr_pages) 481 file = aio_private_file(ctx, nr_pages);
502 if (IS_ERR(file)) { 482 if (IS_ERR(file)) {
503 ctx->aio_ring_file = NULL; 483 ctx->aio_ring_file = NULL;
504 return -ENOMEM; 484 return -ENOMEM;
505 } 485 }
506 486
507 ctx->aio_ring_file = file; 487 ctx->aio_ring_file = file;
508 nr_events = (PAGE_SIZE * nr_pages - si 488 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
509 / sizeof(struct io_eve 489 / sizeof(struct io_event);
510 490
511 ctx->ring_folios = ctx->internal_folio !! 491 ctx->ring_pages = ctx->internal_pages;
512 if (nr_pages > AIO_RING_PAGES) { 492 if (nr_pages > AIO_RING_PAGES) {
513 ctx->ring_folios = kcalloc(nr_ !! 493 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
514 GFP !! 494 GFP_KERNEL);
515 if (!ctx->ring_folios) { !! 495 if (!ctx->ring_pages) {
516 put_aio_ring_file(ctx) 496 put_aio_ring_file(ctx);
517 return -ENOMEM; 497 return -ENOMEM;
518 } 498 }
519 } 499 }
520 500
521 for (i = 0; i < nr_pages; i++) { 501 for (i = 0; i < nr_pages; i++) {
522 struct folio *folio; !! 502 struct page *page;
523 !! 503 page = find_or_create_page(file->f_mapping,
524 folio = __filemap_get_folio(fi !! 504 i, GFP_HIGHUSER | __GFP_ZERO);
525 FG !! 505 if (!page)
526 GF <<
527 if (IS_ERR(folio)) <<
528 break; 506 break;
>> 507 pr_debug("pid(%d) page[%d]->count=%d\n",
>> 508 current->pid, i, page_count(page));
>> 509 SetPageUptodate(page);
>> 510 unlock_page(page);
529 511
530 pr_debug("pid(%d) [%d] folio-> !! 512 ctx->ring_pages[i] = page;
531 folio_ref_count(folio <<
532 folio_end_read(folio, true); <<
533 <<
534 ctx->ring_folios[i] = folio; <<
535 } 513 }
536 ctx->nr_pages = i; 514 ctx->nr_pages = i;
537 515
538 if (unlikely(i != nr_pages)) { 516 if (unlikely(i != nr_pages)) {
539 aio_free_ring(ctx); 517 aio_free_ring(ctx);
540 return -ENOMEM; 518 return -ENOMEM;
541 } 519 }
542 520
543 ctx->mmap_size = nr_pages * PAGE_SIZE; 521 ctx->mmap_size = nr_pages * PAGE_SIZE;
544 pr_debug("attempting mmap of %lu bytes 522 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
545 523
546 if (mmap_write_lock_killable(mm)) { !! 524 if (down_write_killable(&mm->mmap_sem)) {
547 ctx->mmap_size = 0; 525 ctx->mmap_size = 0;
548 aio_free_ring(ctx); 526 aio_free_ring(ctx);
549 return -EINTR; 527 return -EINTR;
550 } 528 }
551 529
552 ctx->mmap_base = do_mmap(ctx->aio_ring !! 530 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
553 PROT_READ | P !! 531 PROT_READ | PROT_WRITE,
554 MAP_SHARED, 0 !! 532 MAP_SHARED, 0, &unused, NULL);
555 mmap_write_unlock(mm); !! 533 up_write(&mm->mmap_sem);
556 if (IS_ERR((void *)ctx->mmap_base)) { 534 if (IS_ERR((void *)ctx->mmap_base)) {
557 ctx->mmap_size = 0; 535 ctx->mmap_size = 0;
558 aio_free_ring(ctx); 536 aio_free_ring(ctx);
559 return -ENOMEM; 537 return -ENOMEM;
560 } 538 }
561 539
562 pr_debug("mmap address: 0x%08lx\n", ct 540 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
563 541
564 ctx->user_id = ctx->mmap_base; 542 ctx->user_id = ctx->mmap_base;
565 ctx->nr_events = nr_events; /* trusted 543 ctx->nr_events = nr_events; /* trusted copy */
566 544
567 ring = folio_address(ctx->ring_folios[ !! 545 ring = kmap_atomic(ctx->ring_pages[0]);
568 ring->nr = nr_events; /* user copy * 546 ring->nr = nr_events; /* user copy */
569 ring->id = ~0U; 547 ring->id = ~0U;
570 ring->head = ring->tail = 0; 548 ring->head = ring->tail = 0;
571 ring->magic = AIO_RING_MAGIC; 549 ring->magic = AIO_RING_MAGIC;
572 ring->compat_features = AIO_RING_COMPA 550 ring->compat_features = AIO_RING_COMPAT_FEATURES;
573 ring->incompat_features = AIO_RING_INC 551 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
574 ring->header_length = sizeof(struct ai 552 ring->header_length = sizeof(struct aio_ring);
575 flush_dcache_folio(ctx->ring_folios[0] !! 553 kunmap_atomic(ring);
>> 554 flush_dcache_page(ctx->ring_pages[0]);
576 555
577 return 0; 556 return 0;
578 } 557 }
579 558
580 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / s 559 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
581 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - 560 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
582 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PE 561 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
583 562
584 void kiocb_set_cancel_fn(struct kiocb *iocb, k 563 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
585 { 564 {
586 struct aio_kiocb *req; !! 565 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
587 struct kioctx *ctx; !! 566 struct kioctx *ctx = req->ki_ctx;
588 unsigned long flags; 567 unsigned long flags;
589 568
590 /* <<
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_ 569 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
600 return; 570 return;
601 571
602 ctx = req->ki_ctx; <<
603 <<
604 spin_lock_irqsave(&ctx->ctx_lock, flag 572 spin_lock_irqsave(&ctx->ctx_lock, flags);
605 list_add_tail(&req->ki_list, &ctx->act 573 list_add_tail(&req->ki_list, &ctx->active_reqs);
606 req->ki_cancel = cancel; 574 req->ki_cancel = cancel;
607 spin_unlock_irqrestore(&ctx->ctx_lock, 575 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
608 } 576 }
609 EXPORT_SYMBOL(kiocb_set_cancel_fn); 577 EXPORT_SYMBOL(kiocb_set_cancel_fn);
610 578
611 /* 579 /*
612 * free_ioctx() should be RCU delayed to synch 580 * free_ioctx() should be RCU delayed to synchronize against the RCU
613 * protected lookup_ioctx() and also needs pro 581 * protected lookup_ioctx() and also needs process context to call
614 * aio_free_ring(). Use rcu_work. 582 * aio_free_ring(). Use rcu_work.
615 */ 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(to_rcu_work(work), struct kioctx,
619 free 587 free_rwork);
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 /* Synchronize against RCU protected table->table[] dereferences */
638 INIT_RCU_WORK(&ctx->free_rwork, free_i 606 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
639 queue_rcu_work(system_wq, &ctx->free_r 607 queue_rcu_work(system_wq, &ctx->free_rwork);
640 } 608 }
641 609
642 /* 610 /*
643 * When this function runs, the kioctx has bee 611 * When this function runs, the kioctx has been removed from the "hash table"
644 * and ctx->users has dropped to 0, so we know 612 * 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 613 * now it's safe to cancel any that need to be.
646 */ 614 */
647 static void free_ioctx_users(struct percpu_ref 615 static void free_ioctx_users(struct percpu_ref *ref)
648 { 616 {
649 struct kioctx *ctx = container_of(ref, 617 struct kioctx *ctx = container_of(ref, struct kioctx, users);
650 struct aio_kiocb *req; 618 struct aio_kiocb *req;
651 619
652 spin_lock_irq(&ctx->ctx_lock); 620 spin_lock_irq(&ctx->ctx_lock);
653 621
654 while (!list_empty(&ctx->active_reqs)) 622 while (!list_empty(&ctx->active_reqs)) {
655 req = list_first_entry(&ctx->a 623 req = list_first_entry(&ctx->active_reqs,
656 struct 624 struct aio_kiocb, ki_list);
657 req->ki_cancel(&req->rw); 625 req->ki_cancel(&req->rw);
658 list_del_init(&req->ki_list); 626 list_del_init(&req->ki_list);
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 (!rcu_access_pointer(table->table[i])) {
680 ctx->i 648 ctx->id = i;
681 rcu_as 649 rcu_assign_pointer(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 710 * Store the original nr_events -- what userspace passed to io_setup(),
741 * for counting against the global lim 711 * for counting against the global limit -- before it changes.
742 */ 712 */
743 unsigned int max_reqs = nr_events; 713 unsigned int max_reqs = nr_events;
744 714
745 /* 715 /*
746 * We keep track of the number of avai 716 * We keep track of the number of available ringbuffer slots, to prevent
747 * overflow (reqs_available), and we a 717 * overflow (reqs_available), and we also use percpu counters for this.
748 * 718 *
749 * So since up to half the slots might 719 * So since up to half the slots might be on other cpu's percpu counters
750 * and unavailable, double nr_events s 720 * and unavailable, double nr_events so userspace sees what they
751 * expected: additionally, we move req 721 * expected: additionally, we move req_batch slots to/from percpu
752 * counters at a time, so make sure th 722 * counters at a time, so make sure that isn't 0:
753 */ 723 */
754 nr_events = max(nr_events, num_possibl 724 nr_events = max(nr_events, num_possible_cpus() * 4);
755 nr_events *= 2; 725 nr_events *= 2;
756 726
757 /* Prevent overflows */ 727 /* Prevent overflows */
758 if (nr_events > (0x10000000U / sizeof( 728 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
759 pr_debug("ENOMEM: nr_events to 729 pr_debug("ENOMEM: nr_events too high\n");
760 return ERR_PTR(-EINVAL); 730 return ERR_PTR(-EINVAL);
761 } 731 }
762 732
763 if (!nr_events || (unsigned long)max_r 733 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
764 return ERR_PTR(-EAGAIN); 734 return ERR_PTR(-EAGAIN);
765 735
766 ctx = kmem_cache_zalloc(kioctx_cachep, 736 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
767 if (!ctx) 737 if (!ctx)
768 return ERR_PTR(-ENOMEM); 738 return ERR_PTR(-ENOMEM);
769 739
770 ctx->max_reqs = max_reqs; 740 ctx->max_reqs = max_reqs;
771 741
772 spin_lock_init(&ctx->ctx_lock); 742 spin_lock_init(&ctx->ctx_lock);
773 spin_lock_init(&ctx->completion_lock); 743 spin_lock_init(&ctx->completion_lock);
774 mutex_init(&ctx->ring_lock); 744 mutex_init(&ctx->ring_lock);
775 /* Protect against page migration thro 745 /* Protect against page migration throughout kiotx setup by keeping
776 * the ring_lock mutex held until setu 746 * the ring_lock mutex held until setup is complete. */
777 mutex_lock(&ctx->ring_lock); 747 mutex_lock(&ctx->ring_lock);
778 init_waitqueue_head(&ctx->wait); 748 init_waitqueue_head(&ctx->wait);
779 749
780 INIT_LIST_HEAD(&ctx->active_reqs); 750 INIT_LIST_HEAD(&ctx->active_reqs);
781 751
782 if (percpu_ref_init(&ctx->users, free_ 752 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
783 goto err; 753 goto err;
784 754
785 if (percpu_ref_init(&ctx->reqs, free_i 755 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
786 goto err; 756 goto err;
787 757
788 ctx->cpu = alloc_percpu(struct kioctx_ 758 ctx->cpu = alloc_percpu(struct kioctx_cpu);
789 if (!ctx->cpu) 759 if (!ctx->cpu)
790 goto err; 760 goto err;
791 761
792 err = aio_setup_ring(ctx, nr_events); 762 err = aio_setup_ring(ctx, nr_events);
793 if (err < 0) 763 if (err < 0)
794 goto err; 764 goto err;
795 765
796 atomic_set(&ctx->reqs_available, ctx-> 766 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
797 ctx->req_batch = (ctx->nr_events - 1) 767 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
798 if (ctx->req_batch < 1) 768 if (ctx->req_batch < 1)
799 ctx->req_batch = 1; 769 ctx->req_batch = 1;
800 770
801 /* limit the number of system wide aio 771 /* limit the number of system wide aios */
802 spin_lock(&aio_nr_lock); 772 spin_lock(&aio_nr_lock);
803 if (aio_nr + ctx->max_reqs > aio_max_n 773 if (aio_nr + ctx->max_reqs > aio_max_nr ||
804 aio_nr + ctx->max_reqs < aio_nr) { 774 aio_nr + ctx->max_reqs < aio_nr) {
805 spin_unlock(&aio_nr_lock); 775 spin_unlock(&aio_nr_lock);
806 err = -EAGAIN; 776 err = -EAGAIN;
807 goto err_ctx; 777 goto err_ctx;
808 } 778 }
809 aio_nr += ctx->max_reqs; 779 aio_nr += ctx->max_reqs;
810 spin_unlock(&aio_nr_lock); 780 spin_unlock(&aio_nr_lock);
811 781
812 percpu_ref_get(&ctx->users); /* io_ 782 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
813 percpu_ref_get(&ctx->reqs); /* fre 783 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
814 784
815 err = ioctx_add_table(ctx, mm); 785 err = ioctx_add_table(ctx, mm);
816 if (err) 786 if (err)
817 goto err_cleanup; 787 goto err_cleanup;
818 788
819 /* Release the ring_lock mutex now tha 789 /* Release the ring_lock mutex now that all setup is complete. */
820 mutex_unlock(&ctx->ring_lock); 790 mutex_unlock(&ctx->ring_lock);
821 791
822 pr_debug("allocated ioctx %p[%ld]: mm= 792 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
823 ctx, ctx->user_id, mm, ctx->n 793 ctx, ctx->user_id, mm, ctx->nr_events);
824 return ctx; 794 return ctx;
825 795
826 err_cleanup: 796 err_cleanup:
827 aio_nr_sub(ctx->max_reqs); 797 aio_nr_sub(ctx->max_reqs);
828 err_ctx: 798 err_ctx:
829 atomic_set(&ctx->dead, 1); 799 atomic_set(&ctx->dead, 1);
830 if (ctx->mmap_size) 800 if (ctx->mmap_size)
831 vm_munmap(ctx->mmap_base, ctx- 801 vm_munmap(ctx->mmap_base, ctx->mmap_size);
832 aio_free_ring(ctx); 802 aio_free_ring(ctx);
833 err: 803 err:
834 mutex_unlock(&ctx->ring_lock); 804 mutex_unlock(&ctx->ring_lock);
835 free_percpu(ctx->cpu); 805 free_percpu(ctx->cpu);
836 percpu_ref_exit(&ctx->reqs); 806 percpu_ref_exit(&ctx->reqs);
837 percpu_ref_exit(&ctx->users); 807 percpu_ref_exit(&ctx->users);
838 kmem_cache_free(kioctx_cachep, ctx); 808 kmem_cache_free(kioctx_cachep, ctx);
839 pr_debug("error allocating ioctx %d\n" 809 pr_debug("error allocating ioctx %d\n", err);
840 return ERR_PTR(err); 810 return ERR_PTR(err);
841 } 811 }
842 812
843 /* kill_ioctx 813 /* kill_ioctx
844 * Cancels all outstanding aio requests o 814 * Cancels all outstanding aio requests on an aio context. Used
845 * when the processes owning a context ha 815 * when the processes owning a context have all exited to encourage
846 * the rapid destruction of the kioctx. 816 * the rapid destruction of the kioctx.
847 */ 817 */
848 static int kill_ioctx(struct mm_struct *mm, st 818 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
849 struct ctx_rq_wait *wait 819 struct ctx_rq_wait *wait)
850 { 820 {
851 struct kioctx_table *table; 821 struct kioctx_table *table;
852 822
853 spin_lock(&mm->ioctx_lock); 823 spin_lock(&mm->ioctx_lock);
854 if (atomic_xchg(&ctx->dead, 1)) { 824 if (atomic_xchg(&ctx->dead, 1)) {
855 spin_unlock(&mm->ioctx_lock); 825 spin_unlock(&mm->ioctx_lock);
856 return -EINVAL; 826 return -EINVAL;
857 } 827 }
858 828
859 table = rcu_dereference_raw(mm->ioctx_ 829 table = rcu_dereference_raw(mm->ioctx_table);
860 WARN_ON(ctx != rcu_access_pointer(tabl 830 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
861 RCU_INIT_POINTER(table->table[ctx->id] 831 RCU_INIT_POINTER(table->table[ctx->id], NULL);
862 spin_unlock(&mm->ioctx_lock); 832 spin_unlock(&mm->ioctx_lock);
863 833
864 /* free_ioctx_reqs() will do the neces 834 /* free_ioctx_reqs() will do the necessary RCU synchronization */
865 wake_up_all(&ctx->wait); 835 wake_up_all(&ctx->wait);
866 836
867 /* 837 /*
868 * It'd be more correct to do this in 838 * It'd be more correct to do this in free_ioctx(), after all
869 * the outstanding kiocbs have finishe 839 * the outstanding kiocbs have finished - but by then io_destroy
870 * has already returned, so io_setup() 840 * has already returned, so io_setup() could potentially return
871 * -EAGAIN with no ioctxs actually in 841 * -EAGAIN with no ioctxs actually in use (as far as userspace
872 * could tell). 842 * could tell).
873 */ 843 */
874 aio_nr_sub(ctx->max_reqs); 844 aio_nr_sub(ctx->max_reqs);
875 845
876 if (ctx->mmap_size) 846 if (ctx->mmap_size)
877 vm_munmap(ctx->mmap_base, ctx- 847 vm_munmap(ctx->mmap_base, ctx->mmap_size);
878 848
879 ctx->rq_wait = wait; 849 ctx->rq_wait = wait;
880 percpu_ref_kill(&ctx->users); 850 percpu_ref_kill(&ctx->users);
881 return 0; 851 return 0;
882 } 852 }
883 853
884 /* 854 /*
885 * exit_aio: called when the last user of mm g 855 * 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 856 * no way for any new requests to be submited or any of the io_* syscalls to be
887 * called on the context. 857 * called on the context.
888 * 858 *
889 * There may be outstanding kiocbs, but free_i 859 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
890 * them. 860 * them.
891 */ 861 */
892 void exit_aio(struct mm_struct *mm) 862 void exit_aio(struct mm_struct *mm)
893 { 863 {
894 struct kioctx_table *table = rcu_deref 864 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
895 struct ctx_rq_wait wait; 865 struct ctx_rq_wait wait;
896 int i, skipped; 866 int i, skipped;
897 867
898 if (!table) 868 if (!table)
899 return; 869 return;
900 870
901 atomic_set(&wait.count, table->nr); 871 atomic_set(&wait.count, table->nr);
902 init_completion(&wait.comp); 872 init_completion(&wait.comp);
903 873
904 skipped = 0; 874 skipped = 0;
905 for (i = 0; i < table->nr; ++i) { 875 for (i = 0; i < table->nr; ++i) {
906 struct kioctx *ctx = 876 struct kioctx *ctx =
907 rcu_dereference_protec 877 rcu_dereference_protected(table->table[i], true);
908 878
909 if (!ctx) { 879 if (!ctx) {
910 skipped++; 880 skipped++;
911 continue; 881 continue;
912 } 882 }
913 883
914 /* 884 /*
915 * We don't need to bother wit 885 * We don't need to bother with munmap() here - exit_mmap(mm)
916 * is coming and it'll unmap e 886 * is coming and it'll unmap everything. And we simply can't,
917 * this is not necessarily our 887 * this is not necessarily our ->mm.
918 * Since kill_ioctx() uses non 888 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
919 * that it needs to unmap the 889 * that it needs to unmap the area, just set it to 0.
920 */ 890 */
921 ctx->mmap_size = 0; 891 ctx->mmap_size = 0;
922 kill_ioctx(mm, ctx, &wait); 892 kill_ioctx(mm, ctx, &wait);
923 } 893 }
924 894
925 if (!atomic_sub_and_test(skipped, &wai 895 if (!atomic_sub_and_test(skipped, &wait.count)) {
926 /* Wait until all IO for the c 896 /* Wait until all IO for the context are done. */
927 wait_for_completion(&wait.comp 897 wait_for_completion(&wait.comp);
928 } 898 }
929 899
930 RCU_INIT_POINTER(mm->ioctx_table, NULL 900 RCU_INIT_POINTER(mm->ioctx_table, NULL);
931 kfree(table); 901 kfree(table);
932 } 902 }
933 903
934 static void put_reqs_available(struct kioctx * 904 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
935 { 905 {
936 struct kioctx_cpu *kcpu; 906 struct kioctx_cpu *kcpu;
937 unsigned long flags; 907 unsigned long flags;
938 908
939 local_irq_save(flags); 909 local_irq_save(flags);
940 kcpu = this_cpu_ptr(ctx->cpu); 910 kcpu = this_cpu_ptr(ctx->cpu);
941 kcpu->reqs_available += nr; 911 kcpu->reqs_available += nr;
942 912
943 while (kcpu->reqs_available >= ctx->re 913 while (kcpu->reqs_available >= ctx->req_batch * 2) {
944 kcpu->reqs_available -= ctx->r 914 kcpu->reqs_available -= ctx->req_batch;
945 atomic_add(ctx->req_batch, &ct 915 atomic_add(ctx->req_batch, &ctx->reqs_available);
946 } 916 }
947 917
948 local_irq_restore(flags); 918 local_irq_restore(flags);
949 } 919 }
950 920
951 static bool __get_reqs_available(struct kioctx 921 static bool __get_reqs_available(struct kioctx *ctx)
952 { 922 {
953 struct kioctx_cpu *kcpu; 923 struct kioctx_cpu *kcpu;
954 bool ret = false; 924 bool ret = false;
955 unsigned long flags; 925 unsigned long flags;
956 926
957 local_irq_save(flags); 927 local_irq_save(flags);
958 kcpu = this_cpu_ptr(ctx->cpu); 928 kcpu = this_cpu_ptr(ctx->cpu);
959 if (!kcpu->reqs_available) { 929 if (!kcpu->reqs_available) {
960 int avail = atomic_read(&ctx-> !! 930 int old, avail = atomic_read(&ctx->reqs_available);
961 931
962 do { 932 do {
963 if (avail < ctx->req_b 933 if (avail < ctx->req_batch)
964 goto out; 934 goto out;
965 } while (!atomic_try_cmpxchg(& !! 935
966 & !! 936 old = avail;
>> 937 avail = atomic_cmpxchg(&ctx->reqs_available,
>> 938 avail, avail - ctx->req_batch);
>> 939 } while (avail != old);
967 940
968 kcpu->reqs_available += ctx->r 941 kcpu->reqs_available += ctx->req_batch;
969 } 942 }
970 943
971 ret = true; 944 ret = true;
972 kcpu->reqs_available--; 945 kcpu->reqs_available--;
973 out: 946 out:
974 local_irq_restore(flags); 947 local_irq_restore(flags);
975 return ret; 948 return ret;
976 } 949 }
977 950
978 /* refill_reqs_available 951 /* refill_reqs_available
979 * Updates the reqs_available reference c 952 * Updates the reqs_available reference counts used for tracking the
980 * number of free slots in the completion 953 * number of free slots in the completion ring. This can be called
981 * from aio_complete() (to optimistically 954 * from aio_complete() (to optimistically update reqs_available) or
982 * from aio_get_req() (the we're out of e 955 * from aio_get_req() (the we're out of events case). It must be
983 * called holding ctx->completion_lock. 956 * called holding ctx->completion_lock.
984 */ 957 */
985 static void refill_reqs_available(struct kioct 958 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
986 unsigned tai 959 unsigned tail)
987 { 960 {
988 unsigned events_in_ring, completed; 961 unsigned events_in_ring, completed;
989 962
990 /* Clamp head since userland can write 963 /* Clamp head since userland can write to it. */
991 head %= ctx->nr_events; 964 head %= ctx->nr_events;
992 if (head <= tail) 965 if (head <= tail)
993 events_in_ring = tail - head; 966 events_in_ring = tail - head;
994 else 967 else
995 events_in_ring = ctx->nr_event 968 events_in_ring = ctx->nr_events - (head - tail);
996 969
997 completed = ctx->completed_events; 970 completed = ctx->completed_events;
998 if (events_in_ring < completed) 971 if (events_in_ring < completed)
999 completed -= events_in_ring; 972 completed -= events_in_ring;
1000 else 973 else
1001 completed = 0; 974 completed = 0;
1002 975
1003 if (!completed) 976 if (!completed)
1004 return; 977 return;
1005 978
1006 ctx->completed_events -= completed; 979 ctx->completed_events -= completed;
1007 put_reqs_available(ctx, completed); 980 put_reqs_available(ctx, completed);
1008 } 981 }
1009 982
1010 /* user_refill_reqs_available 983 /* user_refill_reqs_available
1011 * Called to refill reqs_available when 984 * Called to refill reqs_available when aio_get_req() encounters an
1012 * out of space in the completion ring. 985 * out of space in the completion ring.
1013 */ 986 */
1014 static void user_refill_reqs_available(struct 987 static void user_refill_reqs_available(struct kioctx *ctx)
1015 { 988 {
1016 spin_lock_irq(&ctx->completion_lock); 989 spin_lock_irq(&ctx->completion_lock);
1017 if (ctx->completed_events) { 990 if (ctx->completed_events) {
1018 struct aio_ring *ring; 991 struct aio_ring *ring;
1019 unsigned head; 992 unsigned head;
1020 993
1021 /* Access of ring->head may r 994 /* Access of ring->head may race with aio_read_events_ring()
1022 * here, but that's okay sinc 995 * here, but that's okay since whether we read the old version
1023 * or the new version, and ei 996 * or the new version, and either will be valid. The important
1024 * part is that head cannot p 997 * part is that head cannot pass tail since we prevent
1025 * aio_complete() from updati 998 * aio_complete() from updating tail by holding
1026 * ctx->completion_lock. Eve 999 * ctx->completion_lock. Even if head is invalid, the check
1027 * against ctx->completed_eve 1000 * against ctx->completed_events below will make sure we do the
1028 * safe/right thing. 1001 * safe/right thing.
1029 */ 1002 */
1030 ring = folio_address(ctx->rin !! 1003 ring = kmap_atomic(ctx->ring_pages[0]);
1031 head = ring->head; 1004 head = ring->head;
>> 1005 kunmap_atomic(ring);
1032 1006
1033 refill_reqs_available(ctx, he 1007 refill_reqs_available(ctx, head, ctx->tail);
1034 } 1008 }
1035 1009
1036 spin_unlock_irq(&ctx->completion_lock 1010 spin_unlock_irq(&ctx->completion_lock);
1037 } 1011 }
1038 1012
1039 static bool get_reqs_available(struct kioctx 1013 static bool get_reqs_available(struct kioctx *ctx)
1040 { 1014 {
1041 if (__get_reqs_available(ctx)) 1015 if (__get_reqs_available(ctx))
1042 return true; 1016 return true;
1043 user_refill_reqs_available(ctx); 1017 user_refill_reqs_available(ctx);
1044 return __get_reqs_available(ctx); 1018 return __get_reqs_available(ctx);
1045 } 1019 }
1046 1020
1047 /* aio_get_req 1021 /* aio_get_req
1048 * Allocate a slot for an aio request. 1022 * Allocate a slot for an aio request.
1049 * Returns NULL if no requests are free. 1023 * Returns NULL if no requests are free.
1050 * 1024 *
1051 * The refcount is initialized to 2 - one for 1025 * The refcount is initialized to 2 - one for the async op completion,
1052 * one for the synchronous code that does thi 1026 * one for the synchronous code that does this.
1053 */ 1027 */
1054 static inline struct aio_kiocb *aio_get_req(s 1028 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1055 { 1029 {
1056 struct aio_kiocb *req; 1030 struct aio_kiocb *req;
1057 1031
1058 req = kmem_cache_alloc(kiocb_cachep, 1032 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1059 if (unlikely(!req)) 1033 if (unlikely(!req))
1060 return NULL; 1034 return NULL;
1061 1035
1062 if (unlikely(!get_reqs_available(ctx) 1036 if (unlikely(!get_reqs_available(ctx))) {
1063 kmem_cache_free(kiocb_cachep, 1037 kmem_cache_free(kiocb_cachep, req);
1064 return NULL; 1038 return NULL;
1065 } 1039 }
1066 1040
1067 percpu_ref_get(&ctx->reqs); 1041 percpu_ref_get(&ctx->reqs);
1068 req->ki_ctx = ctx; 1042 req->ki_ctx = ctx;
1069 INIT_LIST_HEAD(&req->ki_list); 1043 INIT_LIST_HEAD(&req->ki_list);
1070 refcount_set(&req->ki_refcnt, 2); 1044 refcount_set(&req->ki_refcnt, 2);
1071 req->ki_eventfd = NULL; 1045 req->ki_eventfd = NULL;
1072 return req; 1046 return req;
1073 } 1047 }
1074 1048
1075 static struct kioctx *lookup_ioctx(unsigned l 1049 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1076 { 1050 {
1077 struct aio_ring __user *ring = (void 1051 struct aio_ring __user *ring = (void __user *)ctx_id;
1078 struct mm_struct *mm = current->mm; 1052 struct mm_struct *mm = current->mm;
1079 struct kioctx *ctx, *ret = NULL; 1053 struct kioctx *ctx, *ret = NULL;
1080 struct kioctx_table *table; 1054 struct kioctx_table *table;
1081 unsigned id; 1055 unsigned id;
1082 1056
1083 if (get_user(id, &ring->id)) 1057 if (get_user(id, &ring->id))
1084 return NULL; 1058 return NULL;
1085 1059
1086 rcu_read_lock(); 1060 rcu_read_lock();
1087 table = rcu_dereference(mm->ioctx_tab 1061 table = rcu_dereference(mm->ioctx_table);
1088 1062
1089 if (!table || id >= table->nr) 1063 if (!table || id >= table->nr)
1090 goto out; 1064 goto out;
1091 1065
1092 id = array_index_nospec(id, table->nr 1066 id = array_index_nospec(id, table->nr);
1093 ctx = rcu_dereference(table->table[id 1067 ctx = rcu_dereference(table->table[id]);
1094 if (ctx && ctx->user_id == ctx_id) { 1068 if (ctx && ctx->user_id == ctx_id) {
1095 if (percpu_ref_tryget_live(&c 1069 if (percpu_ref_tryget_live(&ctx->users))
1096 ret = ctx; 1070 ret = ctx;
1097 } 1071 }
1098 out: 1072 out:
1099 rcu_read_unlock(); 1073 rcu_read_unlock();
1100 return ret; 1074 return ret;
1101 } 1075 }
1102 1076
1103 static inline void iocb_destroy(struct aio_ki 1077 static inline void iocb_destroy(struct aio_kiocb *iocb)
1104 { 1078 {
1105 if (iocb->ki_eventfd) 1079 if (iocb->ki_eventfd)
1106 eventfd_ctx_put(iocb->ki_even 1080 eventfd_ctx_put(iocb->ki_eventfd);
1107 if (iocb->ki_filp) 1081 if (iocb->ki_filp)
1108 fput(iocb->ki_filp); 1082 fput(iocb->ki_filp);
1109 percpu_ref_put(&iocb->ki_ctx->reqs); 1083 percpu_ref_put(&iocb->ki_ctx->reqs);
1110 kmem_cache_free(kiocb_cachep, iocb); 1084 kmem_cache_free(kiocb_cachep, iocb);
1111 } 1085 }
1112 1086
1113 struct aio_waiter { <<
1114 struct wait_queue_entry w; <<
1115 size_t min_nr; <<
1116 }; <<
1117 <<
1118 /* aio_complete 1087 /* aio_complete
1119 * Called when the io request on the giv 1088 * Called when the io request on the given iocb is complete.
1120 */ 1089 */
1121 static void aio_complete(struct aio_kiocb *io 1090 static void aio_complete(struct aio_kiocb *iocb)
1122 { 1091 {
1123 struct kioctx *ctx = iocb->ki_ctx; 1092 struct kioctx *ctx = iocb->ki_ctx;
1124 struct aio_ring *ring; 1093 struct aio_ring *ring;
1125 struct io_event *ev_page, *event; 1094 struct io_event *ev_page, *event;
1126 unsigned tail, pos, head, avail; !! 1095 unsigned tail, pos, head;
1127 unsigned long flags; 1096 unsigned long flags;
1128 1097
1129 /* 1098 /*
1130 * Add a completion event to the ring 1099 * Add a completion event to the ring buffer. Must be done holding
1131 * ctx->completion_lock to prevent ot 1100 * ctx->completion_lock to prevent other code from messing with the tail
1132 * pointer since we might be called f 1101 * pointer since we might be called from irq context.
1133 */ 1102 */
1134 spin_lock_irqsave(&ctx->completion_lo 1103 spin_lock_irqsave(&ctx->completion_lock, flags);
1135 1104
1136 tail = ctx->tail; 1105 tail = ctx->tail;
1137 pos = tail + AIO_EVENTS_OFFSET; 1106 pos = tail + AIO_EVENTS_OFFSET;
1138 1107
1139 if (++tail >= ctx->nr_events) 1108 if (++tail >= ctx->nr_events)
1140 tail = 0; 1109 tail = 0;
1141 1110
1142 ev_page = folio_address(ctx->ring_fol !! 1111 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1143 event = ev_page + pos % AIO_EVENTS_PE 1112 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1144 1113
1145 *event = iocb->ki_res; 1114 *event = iocb->ki_res;
1146 1115
1147 flush_dcache_folio(ctx->ring_folios[p !! 1116 kunmap_atomic(ev_page);
>> 1117 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1148 1118
1149 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\ 1119 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1150 (void __user *)(unsigned lon 1120 (void __user *)(unsigned long)iocb->ki_res.obj,
1151 iocb->ki_res.data, iocb->ki_ 1121 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1152 1122
1153 /* after flagging the request as done 1123 /* after flagging the request as done, we
1154 * must never even look at it again 1124 * must never even look at it again
1155 */ 1125 */
1156 smp_wmb(); /* make event visible 1126 smp_wmb(); /* make event visible before updating tail */
1157 1127
1158 ctx->tail = tail; 1128 ctx->tail = tail;
1159 1129
1160 ring = folio_address(ctx->ring_folios !! 1130 ring = kmap_atomic(ctx->ring_pages[0]);
1161 head = ring->head; 1131 head = ring->head;
1162 ring->tail = tail; 1132 ring->tail = tail;
1163 flush_dcache_folio(ctx->ring_folios[0 !! 1133 kunmap_atomic(ring);
>> 1134 flush_dcache_page(ctx->ring_pages[0]);
1164 1135
1165 ctx->completed_events++; 1136 ctx->completed_events++;
1166 if (ctx->completed_events > 1) 1137 if (ctx->completed_events > 1)
1167 refill_reqs_available(ctx, he 1138 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 1139 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1173 1140
1174 pr_debug("added to ring %p at [%u]\n" 1141 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1175 1142
1176 /* 1143 /*
1177 * Check if the user asked us to deli 1144 * Check if the user asked us to deliver the result through an
1178 * eventfd. The eventfd_signal() func 1145 * eventfd. The eventfd_signal() function is safe to be called
1179 * from IRQ context. 1146 * from IRQ context.
1180 */ 1147 */
1181 if (iocb->ki_eventfd) 1148 if (iocb->ki_eventfd)
1182 eventfd_signal(iocb->ki_event !! 1149 eventfd_signal(iocb->ki_eventfd, 1);
1183 1150
1184 /* 1151 /*
1185 * We have to order our ring_info tai 1152 * We have to order our ring_info tail store above and test
1186 * of the wait list below outside the 1153 * of the wait list below outside the wait lock. This is
1187 * like in wake_up_bit() where cleari 1154 * like in wake_up_bit() where clearing a bit has to be
1188 * ordered with the unlocked test. 1155 * ordered with the unlocked test.
1189 */ 1156 */
1190 smp_mb(); 1157 smp_mb();
1191 1158
1192 if (waitqueue_active(&ctx->wait)) { !! 1159 if (waitqueue_active(&ctx->wait))
1193 struct aio_waiter *curr, *nex !! 1160 wake_up(&ctx->wait);
1194 unsigned long flags; <<
1195 <<
1196 spin_lock_irqsave(&ctx->wait. <<
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 } 1161 }
1205 1162
1206 static inline void iocb_put(struct aio_kiocb 1163 static inline void iocb_put(struct aio_kiocb *iocb)
1207 { 1164 {
1208 if (refcount_dec_and_test(&iocb->ki_r 1165 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1209 aio_complete(iocb); 1166 aio_complete(iocb);
1210 iocb_destroy(iocb); 1167 iocb_destroy(iocb);
1211 } 1168 }
1212 } 1169 }
1213 1170
1214 /* aio_read_events_ring 1171 /* aio_read_events_ring
1215 * Pull an event off of the ioctx's even 1172 * Pull an event off of the ioctx's event ring. Returns the number of
1216 * events fetched 1173 * events fetched
1217 */ 1174 */
1218 static long aio_read_events_ring(struct kioct 1175 static long aio_read_events_ring(struct kioctx *ctx,
1219 struct io_ev 1176 struct io_event __user *event, long nr)
1220 { 1177 {
1221 struct aio_ring *ring; 1178 struct aio_ring *ring;
1222 unsigned head, tail, pos; 1179 unsigned head, tail, pos;
1223 long ret = 0; 1180 long ret = 0;
1224 int copy_ret; 1181 int copy_ret;
1225 1182
1226 /* 1183 /*
1227 * The mutex can block and wake us up 1184 * The mutex can block and wake us up and that will cause
1228 * wait_event_interruptible_hrtimeout 1185 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1229 * and repeat. This should be rare en 1186 * and repeat. This should be rare enough that it doesn't cause
1230 * peformance issues. See the comment 1187 * peformance issues. See the comment in read_events() for more detail.
1231 */ 1188 */
1232 sched_annotate_sleep(); 1189 sched_annotate_sleep();
1233 mutex_lock(&ctx->ring_lock); 1190 mutex_lock(&ctx->ring_lock);
1234 1191
1235 /* Access to ->ring_folios here is pr !! 1192 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1236 ring = folio_address(ctx->ring_folios !! 1193 ring = kmap_atomic(ctx->ring_pages[0]);
1237 head = ring->head; 1194 head = ring->head;
1238 tail = ring->tail; 1195 tail = ring->tail;
>> 1196 kunmap_atomic(ring);
1239 1197
1240 /* 1198 /*
1241 * Ensure that once we've read the cu 1199 * Ensure that once we've read the current tail pointer, that
1242 * we also see the events that were s 1200 * we also see the events that were stored up to the tail.
1243 */ 1201 */
1244 smp_rmb(); 1202 smp_rmb();
1245 1203
1246 pr_debug("h%u t%u m%u\n", head, tail, 1204 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1247 1205
1248 if (head == tail) 1206 if (head == tail)
1249 goto out; 1207 goto out;
1250 1208
1251 head %= ctx->nr_events; 1209 head %= ctx->nr_events;
1252 tail %= ctx->nr_events; 1210 tail %= ctx->nr_events;
1253 1211
1254 while (ret < nr) { 1212 while (ret < nr) {
1255 long avail; 1213 long avail;
1256 struct io_event *ev; 1214 struct io_event *ev;
1257 struct folio *folio; !! 1215 struct page *page;
1258 1216
1259 avail = (head <= tail ? tail 1217 avail = (head <= tail ? tail : ctx->nr_events) - head;
1260 if (head == tail) 1218 if (head == tail)
1261 break; 1219 break;
1262 1220
1263 pos = head + AIO_EVENTS_OFFSE 1221 pos = head + AIO_EVENTS_OFFSET;
1264 folio = ctx->ring_folios[pos !! 1222 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1265 pos %= AIO_EVENTS_PER_PAGE; 1223 pos %= AIO_EVENTS_PER_PAGE;
1266 1224
1267 avail = min(avail, nr - ret); 1225 avail = min(avail, nr - ret);
1268 avail = min_t(long, avail, AI 1226 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1269 1227
1270 ev = folio_address(folio); !! 1228 ev = kmap(page);
1271 copy_ret = copy_to_user(event 1229 copy_ret = copy_to_user(event + ret, ev + pos,
1272 sizeo 1230 sizeof(*ev) * avail);
>> 1231 kunmap(page);
1273 1232
1274 if (unlikely(copy_ret)) { 1233 if (unlikely(copy_ret)) {
1275 ret = -EFAULT; 1234 ret = -EFAULT;
1276 goto out; 1235 goto out;
1277 } 1236 }
1278 1237
1279 ret += avail; 1238 ret += avail;
1280 head += avail; 1239 head += avail;
1281 head %= ctx->nr_events; 1240 head %= ctx->nr_events;
1282 } 1241 }
1283 1242
1284 ring = folio_address(ctx->ring_folios !! 1243 ring = kmap_atomic(ctx->ring_pages[0]);
1285 ring->head = head; 1244 ring->head = head;
1286 flush_dcache_folio(ctx->ring_folios[0 !! 1245 kunmap_atomic(ring);
>> 1246 flush_dcache_page(ctx->ring_pages[0]);
1287 1247
1288 pr_debug("%li h%u t%u\n", ret, head, 1248 pr_debug("%li h%u t%u\n", ret, head, tail);
1289 out: 1249 out:
1290 mutex_unlock(&ctx->ring_lock); 1250 mutex_unlock(&ctx->ring_lock);
1291 1251
1292 return ret; 1252 return ret;
1293 } 1253 }
1294 1254
1295 static bool aio_read_events(struct kioctx *ct 1255 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1296 struct io_event _ 1256 struct io_event __user *event, long *i)
1297 { 1257 {
1298 long ret = aio_read_events_ring(ctx, 1258 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1299 1259
1300 if (ret > 0) 1260 if (ret > 0)
1301 *i += ret; 1261 *i += ret;
1302 1262
1303 if (unlikely(atomic_read(&ctx->dead)) 1263 if (unlikely(atomic_read(&ctx->dead)))
1304 ret = -EINVAL; 1264 ret = -EINVAL;
1305 1265
1306 if (!*i) 1266 if (!*i)
1307 *i = ret; 1267 *i = ret;
1308 1268
1309 return ret < 0 || *i >= min_nr; 1269 return ret < 0 || *i >= min_nr;
1310 } 1270 }
1311 1271
1312 static long read_events(struct kioctx *ctx, l 1272 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1313 struct io_event __use 1273 struct io_event __user *event,
1314 ktime_t until) 1274 ktime_t until)
1315 { 1275 {
1316 struct hrtimer_sleeper t; !! 1276 long ret = 0;
1317 struct aio_waiter w; <<
1318 long ret = 0, ret2 = 0; <<
1319 1277
1320 /* 1278 /*
1321 * Note that aio_read_events() is bei 1279 * Note that aio_read_events() is being called as the conditional - i.e.
1322 * we're calling it after prepare_to_ 1280 * we're calling it after prepare_to_wait() has set task state to
1323 * TASK_INTERRUPTIBLE. 1281 * TASK_INTERRUPTIBLE.
1324 * 1282 *
1325 * But aio_read_events() can block, a 1283 * But aio_read_events() can block, and if it blocks it's going to flip
1326 * the task state back to TASK_RUNNIN 1284 * the task state back to TASK_RUNNING.
1327 * 1285 *
1328 * This should be ok, provided it doe 1286 * This should be ok, provided it doesn't flip the state back to
1329 * TASK_RUNNING and return 0 too much 1287 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1330 * will only happen if the mutex_lock 1288 * will only happen if the mutex_lock() call blocks, and we then find
1331 * the ringbuffer empty. So in practi 1289 * the ringbuffer empty. So in practice we should be ok, but it's
1332 * something to be aware of when touc 1290 * something to be aware of when touching this code.
1333 */ 1291 */
1334 aio_read_events(ctx, min_nr, nr, even !! 1292 if (until == 0)
1335 if (until == 0 || ret < 0 || ret >= m !! 1293 aio_read_events(ctx, min_nr, nr, event, &ret);
1336 return ret; !! 1294 else
1337 !! 1295 wait_event_interruptible_hrtimeout(ctx->wait,
1338 hrtimer_init_sleeper_on_stack(&t, CLO !! 1296 aio_read_events(ctx, min_nr, nr, event, &ret),
1339 if (until != KTIME_MAX) { !! 1297 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 <<
1362 finish_wait(&ctx->wait, &w.w); <<
1363 hrtimer_cancel(&t.timer); <<
1364 destroy_hrtimer_on_stack(&t.timer); <<
1365 <<
1366 return ret; 1298 return ret;
1367 } 1299 }
1368 1300
1369 /* sys_io_setup: 1301 /* sys_io_setup:
1370 * Create an aio_context capable of rece 1302 * Create an aio_context capable of receiving at least nr_events.
1371 * ctxp must not point to an aio_context 1303 * ctxp must not point to an aio_context that already exists, and
1372 * must be initialized to 0 prior to the 1304 * must be initialized to 0 prior to the call. On successful
1373 * creation of the aio_context, *ctxp is 1305 * creation of the aio_context, *ctxp is filled in with the resulting
1374 * handle. May fail with -EINVAL if *ct 1306 * handle. May fail with -EINVAL if *ctxp is not initialized,
1375 * if the specified nr_events exceeds in 1307 * if the specified nr_events exceeds internal limits. May fail
1376 * with -EAGAIN if the specified nr_even 1308 * with -EAGAIN if the specified nr_events exceeds the user's limit
1377 * of available events. May fail with - 1309 * of available events. May fail with -ENOMEM if insufficient kernel
1378 * resources are available. May fail wi 1310 * resources are available. May fail with -EFAULT if an invalid
1379 * pointer is passed for ctxp. Will fai 1311 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1380 * implemented. 1312 * implemented.
1381 */ 1313 */
1382 SYSCALL_DEFINE2(io_setup, unsigned, nr_events 1314 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1383 { 1315 {
1384 struct kioctx *ioctx = NULL; 1316 struct kioctx *ioctx = NULL;
1385 unsigned long ctx; 1317 unsigned long ctx;
1386 long ret; 1318 long ret;
1387 1319
1388 ret = get_user(ctx, ctxp); 1320 ret = get_user(ctx, ctxp);
1389 if (unlikely(ret)) 1321 if (unlikely(ret))
1390 goto out; 1322 goto out;
1391 1323
1392 ret = -EINVAL; 1324 ret = -EINVAL;
1393 if (unlikely(ctx || nr_events == 0)) 1325 if (unlikely(ctx || nr_events == 0)) {
1394 pr_debug("EINVAL: ctx %lu nr_ 1326 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1395 ctx, nr_events); 1327 ctx, nr_events);
1396 goto out; 1328 goto out;
1397 } 1329 }
1398 1330
1399 ioctx = ioctx_alloc(nr_events); 1331 ioctx = ioctx_alloc(nr_events);
1400 ret = PTR_ERR(ioctx); 1332 ret = PTR_ERR(ioctx);
1401 if (!IS_ERR(ioctx)) { 1333 if (!IS_ERR(ioctx)) {
1402 ret = put_user(ioctx->user_id 1334 ret = put_user(ioctx->user_id, ctxp);
1403 if (ret) 1335 if (ret)
1404 kill_ioctx(current->m 1336 kill_ioctx(current->mm, ioctx, NULL);
1405 percpu_ref_put(&ioctx->users) 1337 percpu_ref_put(&ioctx->users);
1406 } 1338 }
1407 1339
1408 out: 1340 out:
1409 return ret; 1341 return ret;
1410 } 1342 }
1411 1343
1412 #ifdef CONFIG_COMPAT 1344 #ifdef CONFIG_COMPAT
1413 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr 1345 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1414 { 1346 {
1415 struct kioctx *ioctx = NULL; 1347 struct kioctx *ioctx = NULL;
1416 unsigned long ctx; 1348 unsigned long ctx;
1417 long ret; 1349 long ret;
1418 1350
1419 ret = get_user(ctx, ctx32p); 1351 ret = get_user(ctx, ctx32p);
1420 if (unlikely(ret)) 1352 if (unlikely(ret))
1421 goto out; 1353 goto out;
1422 1354
1423 ret = -EINVAL; 1355 ret = -EINVAL;
1424 if (unlikely(ctx || nr_events == 0)) 1356 if (unlikely(ctx || nr_events == 0)) {
1425 pr_debug("EINVAL: ctx %lu nr_ 1357 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1426 ctx, nr_events); 1358 ctx, nr_events);
1427 goto out; 1359 goto out;
1428 } 1360 }
1429 1361
1430 ioctx = ioctx_alloc(nr_events); 1362 ioctx = ioctx_alloc(nr_events);
1431 ret = PTR_ERR(ioctx); 1363 ret = PTR_ERR(ioctx);
1432 if (!IS_ERR(ioctx)) { 1364 if (!IS_ERR(ioctx)) {
1433 /* truncating is ok because i 1365 /* truncating is ok because it's a user address */
1434 ret = put_user((u32)ioctx->us 1366 ret = put_user((u32)ioctx->user_id, ctx32p);
1435 if (ret) 1367 if (ret)
1436 kill_ioctx(current->m 1368 kill_ioctx(current->mm, ioctx, NULL);
1437 percpu_ref_put(&ioctx->users) 1369 percpu_ref_put(&ioctx->users);
1438 } 1370 }
1439 1371
1440 out: 1372 out:
1441 return ret; 1373 return ret;
1442 } 1374 }
1443 #endif 1375 #endif
1444 1376
1445 /* sys_io_destroy: 1377 /* sys_io_destroy:
1446 * Destroy the aio_context specified. M 1378 * Destroy the aio_context specified. May cancel any outstanding
1447 * AIOs and block on completion. Will f 1379 * AIOs and block on completion. Will fail with -ENOSYS if not
1448 * implemented. May fail with -EINVAL i 1380 * implemented. May fail with -EINVAL if the context pointed to
1449 * is invalid. 1381 * is invalid.
1450 */ 1382 */
1451 SYSCALL_DEFINE1(io_destroy, aio_context_t, ct 1383 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1452 { 1384 {
1453 struct kioctx *ioctx = lookup_ioctx(c 1385 struct kioctx *ioctx = lookup_ioctx(ctx);
1454 if (likely(NULL != ioctx)) { 1386 if (likely(NULL != ioctx)) {
1455 struct ctx_rq_wait wait; 1387 struct ctx_rq_wait wait;
1456 int ret; 1388 int ret;
1457 1389
1458 init_completion(&wait.comp); 1390 init_completion(&wait.comp);
1459 atomic_set(&wait.count, 1); 1391 atomic_set(&wait.count, 1);
1460 1392
1461 /* Pass requests_done to kill 1393 /* Pass requests_done to kill_ioctx() where it can be set
1462 * in a thread-safe way. If w 1394 * in a thread-safe way. If we try to set it here then we have
1463 * a race condition if two io 1395 * a race condition if two io_destroy() called simultaneously.
1464 */ 1396 */
1465 ret = kill_ioctx(current->mm, 1397 ret = kill_ioctx(current->mm, ioctx, &wait);
1466 percpu_ref_put(&ioctx->users) 1398 percpu_ref_put(&ioctx->users);
1467 1399
1468 /* Wait until all IO for the 1400 /* Wait until all IO for the context are done. Otherwise kernel
1469 * keep using user-space buff 1401 * keep using user-space buffers even if user thinks the context
1470 * is destroyed. 1402 * is destroyed.
1471 */ 1403 */
1472 if (!ret) 1404 if (!ret)
1473 wait_for_completion(& 1405 wait_for_completion(&wait.comp);
1474 1406
1475 return ret; 1407 return ret;
1476 } 1408 }
1477 pr_debug("EINVAL: invalid context id\ 1409 pr_debug("EINVAL: invalid context id\n");
1478 return -EINVAL; 1410 return -EINVAL;
1479 } 1411 }
1480 1412
1481 static void aio_remove_iocb(struct aio_kiocb 1413 static void aio_remove_iocb(struct aio_kiocb *iocb)
1482 { 1414 {
1483 struct kioctx *ctx = iocb->ki_ctx; 1415 struct kioctx *ctx = iocb->ki_ctx;
1484 unsigned long flags; 1416 unsigned long flags;
1485 1417
1486 spin_lock_irqsave(&ctx->ctx_lock, fla 1418 spin_lock_irqsave(&ctx->ctx_lock, flags);
1487 list_del(&iocb->ki_list); 1419 list_del(&iocb->ki_list);
1488 spin_unlock_irqrestore(&ctx->ctx_lock 1420 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1489 } 1421 }
1490 1422
1491 static void aio_complete_rw(struct kiocb *kio !! 1423 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1492 { 1424 {
1493 struct aio_kiocb *iocb = container_of 1425 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1494 1426
1495 if (!list_empty_careful(&iocb->ki_lis 1427 if (!list_empty_careful(&iocb->ki_list))
1496 aio_remove_iocb(iocb); 1428 aio_remove_iocb(iocb);
1497 1429
1498 if (kiocb->ki_flags & IOCB_WRITE) { 1430 if (kiocb->ki_flags & IOCB_WRITE) {
1499 struct inode *inode = file_in 1431 struct inode *inode = file_inode(kiocb->ki_filp);
1500 1432
>> 1433 /*
>> 1434 * Tell lockdep we inherited freeze protection from submission
>> 1435 * thread.
>> 1436 */
1501 if (S_ISREG(inode->i_mode)) 1437 if (S_ISREG(inode->i_mode))
1502 kiocb_end_write(kiocb !! 1438 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
>> 1439 file_end_write(kiocb->ki_filp);
1503 } 1440 }
1504 1441
1505 iocb->ki_res.res = res; 1442 iocb->ki_res.res = res;
1506 iocb->ki_res.res2 = 0; !! 1443 iocb->ki_res.res2 = res2;
1507 iocb_put(iocb); 1444 iocb_put(iocb);
1508 } 1445 }
1509 1446
1510 static int aio_prep_rw(struct kiocb *req, con !! 1447 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1511 { 1448 {
1512 int ret; 1449 int ret;
1513 1450
1514 req->ki_complete = aio_complete_rw; 1451 req->ki_complete = aio_complete_rw;
1515 req->private = NULL; 1452 req->private = NULL;
1516 req->ki_pos = iocb->aio_offset; 1453 req->ki_pos = iocb->aio_offset;
1517 req->ki_flags = req->ki_filp->f_iocb_ !! 1454 req->ki_flags = iocb_flags(req->ki_filp);
1518 if (iocb->aio_flags & IOCB_FLAG_RESFD 1455 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1519 req->ki_flags |= IOCB_EVENTFD 1456 req->ki_flags |= IOCB_EVENTFD;
>> 1457 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1520 if (iocb->aio_flags & IOCB_FLAG_IOPRI 1458 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1521 /* 1459 /*
1522 * If the IOCB_FLAG_IOPRIO fl 1460 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1523 * aio_reqprio is interpreted 1461 * aio_reqprio is interpreted as an I/O scheduling
1524 * class and priority. 1462 * class and priority.
1525 */ 1463 */
1526 ret = ioprio_check_cap(iocb-> 1464 ret = ioprio_check_cap(iocb->aio_reqprio);
1527 if (ret) { 1465 if (ret) {
1528 pr_debug("aio ioprio 1466 pr_debug("aio ioprio check cap error: %d\n", ret);
1529 return ret; 1467 return ret;
1530 } 1468 }
1531 1469
1532 req->ki_ioprio = iocb->aio_re 1470 req->ki_ioprio = iocb->aio_reqprio;
1533 } else 1471 } else
1534 req->ki_ioprio = get_current_ 1472 req->ki_ioprio = get_current_ioprio();
1535 1473
1536 ret = kiocb_set_rw_flags(req, iocb->a !! 1474 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1537 if (unlikely(ret)) 1475 if (unlikely(ret))
1538 return ret; 1476 return ret;
1539 1477
1540 req->ki_flags &= ~IOCB_HIPRI; /* no o 1478 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1541 return 0; 1479 return 0;
1542 } 1480 }
1543 1481
1544 static ssize_t aio_setup_rw(int rw, const str !! 1482 static int aio_setup_rw(int rw, const struct iocb *iocb, struct iovec **iovec,
1545 struct iovec **iovec, bool ve !! 1483 bool vectored, bool compat, struct iov_iter *iter)
1546 struct iov_iter *iter) <<
1547 { 1484 {
1548 void __user *buf = (void __user *)(ui 1485 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1549 size_t len = iocb->aio_nbytes; 1486 size_t len = iocb->aio_nbytes;
1550 1487
1551 if (!vectored) { 1488 if (!vectored) {
1552 ssize_t ret = import_ubuf(rw, !! 1489 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1553 *iovec = NULL; 1490 *iovec = NULL;
1554 return ret; 1491 return ret;
1555 } 1492 }
1556 !! 1493 #ifdef CONFIG_COMPAT
1557 return __import_iovec(rw, buf, len, U !! 1494 if (compat)
>> 1495 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
>> 1496 iter);
>> 1497 #endif
>> 1498 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1558 } 1499 }
1559 1500
1560 static inline void aio_rw_done(struct kiocb * 1501 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1561 { 1502 {
1562 switch (ret) { 1503 switch (ret) {
1563 case -EIOCBQUEUED: 1504 case -EIOCBQUEUED:
1564 break; 1505 break;
1565 case -ERESTARTSYS: 1506 case -ERESTARTSYS:
1566 case -ERESTARTNOINTR: 1507 case -ERESTARTNOINTR:
1567 case -ERESTARTNOHAND: 1508 case -ERESTARTNOHAND:
1568 case -ERESTART_RESTARTBLOCK: 1509 case -ERESTART_RESTARTBLOCK:
1569 /* 1510 /*
1570 * There's no easy way to res 1511 * There's no easy way to restart the syscall since other AIO's
1571 * may be already running. Ju 1512 * may be already running. Just fail this IO with EINTR.
1572 */ 1513 */
1573 ret = -EINTR; 1514 ret = -EINTR;
1574 fallthrough; !! 1515 /*FALLTHRU*/
1575 default: 1516 default:
1576 req->ki_complete(req, ret); !! 1517 req->ki_complete(req, ret, 0);
1577 } 1518 }
1578 } 1519 }
1579 1520
1580 static int aio_read(struct kiocb *req, const 1521 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1581 bool vectored, bool c 1522 bool vectored, bool compat)
1582 { 1523 {
1583 struct iovec inline_vecs[UIO_FASTIOV] 1524 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1584 struct iov_iter iter; 1525 struct iov_iter iter;
1585 struct file *file; 1526 struct file *file;
1586 int ret; 1527 int ret;
1587 1528
1588 ret = aio_prep_rw(req, iocb, READ); !! 1529 ret = aio_prep_rw(req, iocb);
1589 if (ret) 1530 if (ret)
1590 return ret; 1531 return ret;
1591 file = req->ki_filp; 1532 file = req->ki_filp;
1592 if (unlikely(!(file->f_mode & FMODE_R 1533 if (unlikely(!(file->f_mode & FMODE_READ)))
1593 return -EBADF; 1534 return -EBADF;
>> 1535 ret = -EINVAL;
1594 if (unlikely(!file->f_op->read_iter)) 1536 if (unlikely(!file->f_op->read_iter))
1595 return -EINVAL; 1537 return -EINVAL;
1596 1538
1597 ret = aio_setup_rw(ITER_DEST, iocb, & !! 1539 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1598 if (ret < 0) !! 1540 if (ret)
1599 return ret; 1541 return ret;
1600 ret = rw_verify_area(READ, file, &req 1542 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1601 if (!ret) 1543 if (!ret)
1602 aio_rw_done(req, file->f_op-> !! 1544 aio_rw_done(req, call_read_iter(file, req, &iter));
1603 kfree(iovec); 1545 kfree(iovec);
1604 return ret; 1546 return ret;
1605 } 1547 }
1606 1548
1607 static int aio_write(struct kiocb *req, const 1549 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1608 bool vectored, bool 1550 bool vectored, bool compat)
1609 { 1551 {
1610 struct iovec inline_vecs[UIO_FASTIOV] 1552 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1611 struct iov_iter iter; 1553 struct iov_iter iter;
1612 struct file *file; 1554 struct file *file;
1613 int ret; 1555 int ret;
1614 1556
1615 ret = aio_prep_rw(req, iocb, WRITE); !! 1557 ret = aio_prep_rw(req, iocb);
1616 if (ret) 1558 if (ret)
1617 return ret; 1559 return ret;
1618 file = req->ki_filp; 1560 file = req->ki_filp;
1619 1561
1620 if (unlikely(!(file->f_mode & FMODE_W 1562 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1621 return -EBADF; 1563 return -EBADF;
1622 if (unlikely(!file->f_op->write_iter) 1564 if (unlikely(!file->f_op->write_iter))
1623 return -EINVAL; 1565 return -EINVAL;
1624 1566
1625 ret = aio_setup_rw(ITER_SOURCE, iocb, !! 1567 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1626 if (ret < 0) !! 1568 if (ret)
1627 return ret; 1569 return ret;
1628 ret = rw_verify_area(WRITE, file, &re 1570 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1629 if (!ret) { 1571 if (!ret) {
1630 if (S_ISREG(file_inode(file)- !! 1572 /*
1631 kiocb_start_write(req !! 1573 * Open-code file_start_write here to grab freeze protection,
>> 1574 * which will be released by another thread in
>> 1575 * aio_complete_rw(). Fool lockdep by telling it the lock got
>> 1576 * released so that it doesn't complain about the held lock when
>> 1577 * we return to userspace.
>> 1578 */
>> 1579 if (S_ISREG(file_inode(file)->i_mode)) {
>> 1580 __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true);
>> 1581 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
>> 1582 }
1632 req->ki_flags |= IOCB_WRITE; 1583 req->ki_flags |= IOCB_WRITE;
1633 aio_rw_done(req, file->f_op-> !! 1584 aio_rw_done(req, call_write_iter(file, req, &iter));
1634 } 1585 }
1635 kfree(iovec); 1586 kfree(iovec);
1636 return ret; 1587 return ret;
1637 } 1588 }
1638 1589
1639 static void aio_fsync_work(struct work_struct 1590 static void aio_fsync_work(struct work_struct *work)
1640 { 1591 {
1641 struct aio_kiocb *iocb = container_of 1592 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1642 const struct cred *old_cred = overrid <<
1643 1593
1644 iocb->ki_res.res = vfs_fsync(iocb->fs 1594 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1645 revert_creds(old_cred); <<
1646 put_cred(iocb->fsync.creds); <<
1647 iocb_put(iocb); 1595 iocb_put(iocb);
1648 } 1596 }
1649 1597
1650 static int aio_fsync(struct fsync_iocb *req, 1598 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1651 bool datasync) 1599 bool datasync)
1652 { 1600 {
1653 if (unlikely(iocb->aio_buf || iocb->a 1601 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1654 iocb->aio_rw_flags)) 1602 iocb->aio_rw_flags))
1655 return -EINVAL; 1603 return -EINVAL;
1656 1604
1657 if (unlikely(!req->file->f_op->fsync) 1605 if (unlikely(!req->file->f_op->fsync))
1658 return -EINVAL; 1606 return -EINVAL;
1659 1607
1660 req->creds = prepare_creds(); <<
1661 if (!req->creds) <<
1662 return -ENOMEM; <<
1663 <<
1664 req->datasync = datasync; 1608 req->datasync = datasync;
1665 INIT_WORK(&req->work, aio_fsync_work) 1609 INIT_WORK(&req->work, aio_fsync_work);
1666 schedule_work(&req->work); 1610 schedule_work(&req->work);
1667 return 0; 1611 return 0;
1668 } 1612 }
1669 1613
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 1614 static void aio_poll_complete_work(struct work_struct *work)
1724 { 1615 {
1725 struct poll_iocb *req = container_of( 1616 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1726 struct aio_kiocb *iocb = container_of 1617 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1727 struct poll_table_struct pt = { ._key 1618 struct poll_table_struct pt = { ._key = req->events };
1728 struct kioctx *ctx = iocb->ki_ctx; 1619 struct kioctx *ctx = iocb->ki_ctx;
1729 __poll_t mask = 0; 1620 __poll_t mask = 0;
1730 1621
1731 if (!READ_ONCE(req->cancelled)) 1622 if (!READ_ONCE(req->cancelled))
1732 mask = vfs_poll(req->file, &p 1623 mask = vfs_poll(req->file, &pt) & req->events;
1733 1624
1734 /* 1625 /*
1735 * Note that ->ki_cancel callers also 1626 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1736 * calling ->ki_cancel. We need the 1627 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1737 * synchronize with them. In the can 1628 * synchronize with them. In the cancellation case the list_del_init
1738 * itself is not actually needed, but 1629 * itself is not actually needed, but harmless so we keep it in to
1739 * avoid further branches in the fast 1630 * avoid further branches in the fast path.
1740 */ 1631 */
1741 spin_lock_irq(&ctx->ctx_lock); 1632 spin_lock_irq(&ctx->ctx_lock);
1742 if (poll_iocb_lock_wq(req)) { !! 1633 if (!mask && !READ_ONCE(req->cancelled)) {
1743 if (!mask && !READ_ONCE(req-> !! 1634 add_wait_queue(req->head, &req->wait);
1744 /* !! 1635 spin_unlock_irq(&ctx->ctx_lock);
1745 * The request isn't !! 1636 return;
1746 * Reschedule complet !! 1637 }
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); 1638 list_del_init(&iocb->ki_list);
1762 iocb->ki_res.res = mangle_poll(mask); 1639 iocb->ki_res.res = mangle_poll(mask);
>> 1640 req->done = true;
1763 spin_unlock_irq(&ctx->ctx_lock); 1641 spin_unlock_irq(&ctx->ctx_lock);
1764 1642
1765 iocb_put(iocb); 1643 iocb_put(iocb);
1766 } 1644 }
1767 1645
1768 /* assumes we are called with irqs disabled * 1646 /* assumes we are called with irqs disabled */
1769 static int aio_poll_cancel(struct kiocb *iocb 1647 static int aio_poll_cancel(struct kiocb *iocb)
1770 { 1648 {
1771 struct aio_kiocb *aiocb = container_o 1649 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1772 struct poll_iocb *req = &aiocb->poll; 1650 struct poll_iocb *req = &aiocb->poll;
1773 1651
1774 if (poll_iocb_lock_wq(req)) { !! 1652 spin_lock(&req->head->lock);
1775 WRITE_ONCE(req->cancelled, tr !! 1653 WRITE_ONCE(req->cancelled, true);
1776 if (!req->work_scheduled) { !! 1654 if (!list_empty(&req->wait.entry)) {
1777 schedule_work(&aiocb- !! 1655 list_del_init(&req->wait.entry);
1778 req->work_scheduled = !! 1656 schedule_work(&aiocb->poll.work);
1779 } !! 1657 }
1780 poll_iocb_unlock_wq(req); !! 1658 spin_unlock(&req->head->lock);
1781 } /* else, the request was force-canc <<
1782 1659
1783 return 0; 1660 return 0;
1784 } 1661 }
1785 1662
1786 static int aio_poll_wake(struct wait_queue_en 1663 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1787 void *key) 1664 void *key)
1788 { 1665 {
1789 struct poll_iocb *req = container_of( 1666 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1790 struct aio_kiocb *iocb = container_of 1667 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1791 __poll_t mask = key_to_poll(key); 1668 __poll_t mask = key_to_poll(key);
1792 unsigned long flags; 1669 unsigned long flags;
1793 1670
1794 /* for instances that support it chec 1671 /* for instances that support it check for an event match first: */
1795 if (mask && !(mask & req->events)) 1672 if (mask && !(mask & req->events))
1796 return 0; 1673 return 0;
1797 1674
1798 /* !! 1675 list_del_init(&req->wait.entry);
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 1676
1815 list_del_init(&req->wait.entr !! 1677 if (mask && spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
>> 1678 /*
>> 1679 * Try to complete the iocb inline if we can. Use
>> 1680 * irqsave/irqrestore because not all filesystems (e.g. fuse)
>> 1681 * call this function with IRQs disabled and because IRQs
>> 1682 * have to be disabled before ctx_lock is obtained.
>> 1683 */
1816 list_del(&iocb->ki_list); 1684 list_del(&iocb->ki_list);
1817 iocb->ki_res.res = mangle_pol 1685 iocb->ki_res.res = mangle_poll(mask);
1818 if (iocb->ki_eventfd && !even !! 1686 req->done = true;
1819 iocb = NULL; !! 1687 spin_unlock_irqrestore(&iocb->ki_ctx->ctx_lock, flags);
1820 INIT_WORK(&req->work, !! 1688 iocb_put(iocb);
1821 schedule_work(&req->w <<
1822 } <<
1823 spin_unlock_irqrestore(&ctx-> <<
1824 if (iocb) <<
1825 iocb_put(iocb); <<
1826 } else { 1689 } else {
1827 /* !! 1690 schedule_work(&req->work);
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 } 1691 }
1865 return 1; 1692 return 1;
1866 } 1693 }
1867 1694
1868 struct aio_poll_table { 1695 struct aio_poll_table {
1869 struct poll_table_struct pt; 1696 struct poll_table_struct pt;
1870 struct aio_kiocb *iocb 1697 struct aio_kiocb *iocb;
1871 bool queue <<
1872 int error 1698 int error;
1873 }; 1699 };
1874 1700
1875 static void 1701 static void
1876 aio_poll_queue_proc(struct file *file, struct 1702 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1877 struct poll_table_struct *p) 1703 struct poll_table_struct *p)
1878 { 1704 {
1879 struct aio_poll_table *pt = container 1705 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1880 1706
1881 /* multiple wait queues per file are 1707 /* multiple wait queues per file are not supported */
1882 if (unlikely(pt->queued)) { !! 1708 if (unlikely(pt->iocb->poll.head)) {
1883 pt->error = -EINVAL; 1709 pt->error = -EINVAL;
1884 return; 1710 return;
1885 } 1711 }
1886 1712
1887 pt->queued = true; <<
1888 pt->error = 0; 1713 pt->error = 0;
1889 pt->iocb->poll.head = head; 1714 pt->iocb->poll.head = head;
1890 add_wait_queue(head, &pt->iocb->poll. 1715 add_wait_queue(head, &pt->iocb->poll.wait);
1891 } 1716 }
1892 1717
1893 static int aio_poll(struct aio_kiocb *aiocb, 1718 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1894 { 1719 {
1895 struct kioctx *ctx = aiocb->ki_ctx; 1720 struct kioctx *ctx = aiocb->ki_ctx;
1896 struct poll_iocb *req = &aiocb->poll; 1721 struct poll_iocb *req = &aiocb->poll;
1897 struct aio_poll_table apt; 1722 struct aio_poll_table apt;
1898 bool cancel = false; 1723 bool cancel = false;
1899 __poll_t mask; 1724 __poll_t mask;
1900 1725
1901 /* reject any unknown events outside 1726 /* reject any unknown events outside the normal event mask. */
1902 if ((u16)iocb->aio_buf != iocb->aio_b 1727 if ((u16)iocb->aio_buf != iocb->aio_buf)
1903 return -EINVAL; 1728 return -EINVAL;
1904 /* reject fields that are not defined 1729 /* reject fields that are not defined for poll */
1905 if (iocb->aio_offset || iocb->aio_nby 1730 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1906 return -EINVAL; 1731 return -EINVAL;
1907 1732
1908 INIT_WORK(&req->work, aio_poll_comple 1733 INIT_WORK(&req->work, aio_poll_complete_work);
1909 req->events = demangle_poll(iocb->aio 1734 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1910 1735
1911 req->head = NULL; 1736 req->head = NULL;
>> 1737 req->done = false;
1912 req->cancelled = false; 1738 req->cancelled = false;
1913 req->work_scheduled = false; <<
1914 req->work_need_resched = false; <<
1915 1739
1916 apt.pt._qproc = aio_poll_queue_proc; 1740 apt.pt._qproc = aio_poll_queue_proc;
1917 apt.pt._key = req->events; 1741 apt.pt._key = req->events;
1918 apt.iocb = aiocb; 1742 apt.iocb = aiocb;
1919 apt.queued = false; <<
1920 apt.error = -EINVAL; /* same as no su 1743 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1921 1744
1922 /* initialized the list so that we ca 1745 /* initialized the list so that we can do list_empty checks */
1923 INIT_LIST_HEAD(&req->wait.entry); 1746 INIT_LIST_HEAD(&req->wait.entry);
1924 init_waitqueue_func_entry(&req->wait, 1747 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1925 1748
1926 mask = vfs_poll(req->file, &apt.pt) & 1749 mask = vfs_poll(req->file, &apt.pt) & req->events;
1927 spin_lock_irq(&ctx->ctx_lock); 1750 spin_lock_irq(&ctx->ctx_lock);
1928 if (likely(apt.queued)) { !! 1751 if (likely(req->head)) {
1929 bool on_queue = poll_iocb_loc !! 1752 spin_lock(&req->head->lock);
1930 !! 1753 if (unlikely(list_empty(&req->wait.entry))) {
1931 if (!on_queue || req->work_sc !! 1754 if (apt.error)
1932 /* <<
1933 * aio_poll_wake() al <<
1934 * completion work, o <<
1935 */ <<
1936 if (apt.error) /* uns <<
1937 cancel = true 1755 cancel = true;
1938 apt.error = 0; 1756 apt.error = 0;
1939 mask = 0; 1757 mask = 0;
1940 } 1758 }
1941 if (mask || apt.error) { 1759 if (mask || apt.error) {
1942 /* Steal to complete <<
1943 list_del_init(&req->w 1760 list_del_init(&req->wait.entry);
1944 } else if (cancel) { 1761 } else if (cancel) {
1945 /* Cancel if possible <<
1946 WRITE_ONCE(req->cance 1762 WRITE_ONCE(req->cancelled, true);
1947 } else if (on_queue) { !! 1763 } else if (!req->done) { /* actually waiting for an event */
1948 /* <<
1949 * Actually waiting f <<
1950 * active_reqs so tha <<
1951 */ <<
1952 list_add_tail(&aiocb- 1764 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1953 aiocb->ki_cancel = ai 1765 aiocb->ki_cancel = aio_poll_cancel;
1954 } 1766 }
1955 if (on_queue) !! 1767 spin_unlock(&req->head->lock);
1956 poll_iocb_unlock_wq(r <<
1957 } 1768 }
1958 if (mask) { /* no async, we'd stolen 1769 if (mask) { /* no async, we'd stolen it */
1959 aiocb->ki_res.res = mangle_po 1770 aiocb->ki_res.res = mangle_poll(mask);
1960 apt.error = 0; 1771 apt.error = 0;
1961 } 1772 }
1962 spin_unlock_irq(&ctx->ctx_lock); 1773 spin_unlock_irq(&ctx->ctx_lock);
1963 if (mask) 1774 if (mask)
1964 iocb_put(aiocb); 1775 iocb_put(aiocb);
1965 return apt.error; 1776 return apt.error;
1966 } 1777 }
1967 1778
1968 static int __io_submit_one(struct kioctx *ctx 1779 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1969 struct iocb __user 1780 struct iocb __user *user_iocb, struct aio_kiocb *req,
1970 bool compat) 1781 bool compat)
1971 { 1782 {
1972 req->ki_filp = fget(iocb->aio_fildes) 1783 req->ki_filp = fget(iocb->aio_fildes);
1973 if (unlikely(!req->ki_filp)) 1784 if (unlikely(!req->ki_filp))
1974 return -EBADF; 1785 return -EBADF;
1975 1786
1976 if (iocb->aio_flags & IOCB_FLAG_RESFD 1787 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1977 struct eventfd_ctx *eventfd; 1788 struct eventfd_ctx *eventfd;
1978 /* 1789 /*
1979 * If the IOCB_FLAG_RESFD fla 1790 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1980 * instance of the file* now. 1791 * instance of the file* now. The file descriptor must be
1981 * an eventfd() fd, and will 1792 * an eventfd() fd, and will be signaled for each completed
1982 * event using the eventfd_si 1793 * event using the eventfd_signal() function.
1983 */ 1794 */
1984 eventfd = eventfd_ctx_fdget(i 1795 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1985 if (IS_ERR(eventfd)) 1796 if (IS_ERR(eventfd))
1986 return PTR_ERR(eventf 1797 return PTR_ERR(eventfd);
1987 1798
1988 req->ki_eventfd = eventfd; 1799 req->ki_eventfd = eventfd;
1989 } 1800 }
1990 1801
1991 if (unlikely(put_user(KIOCB_KEY, &use 1802 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1992 pr_debug("EFAULT: aio_key\n") 1803 pr_debug("EFAULT: aio_key\n");
1993 return -EFAULT; 1804 return -EFAULT;
1994 } 1805 }
1995 1806
1996 req->ki_res.obj = (u64)(unsigned long 1807 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1997 req->ki_res.data = iocb->aio_data; 1808 req->ki_res.data = iocb->aio_data;
1998 req->ki_res.res = 0; 1809 req->ki_res.res = 0;
1999 req->ki_res.res2 = 0; 1810 req->ki_res.res2 = 0;
2000 1811
2001 switch (iocb->aio_lio_opcode) { 1812 switch (iocb->aio_lio_opcode) {
2002 case IOCB_CMD_PREAD: 1813 case IOCB_CMD_PREAD:
2003 return aio_read(&req->rw, ioc 1814 return aio_read(&req->rw, iocb, false, compat);
2004 case IOCB_CMD_PWRITE: 1815 case IOCB_CMD_PWRITE:
2005 return aio_write(&req->rw, io 1816 return aio_write(&req->rw, iocb, false, compat);
2006 case IOCB_CMD_PREADV: 1817 case IOCB_CMD_PREADV:
2007 return aio_read(&req->rw, ioc 1818 return aio_read(&req->rw, iocb, true, compat);
2008 case IOCB_CMD_PWRITEV: 1819 case IOCB_CMD_PWRITEV:
2009 return aio_write(&req->rw, io 1820 return aio_write(&req->rw, iocb, true, compat);
2010 case IOCB_CMD_FSYNC: 1821 case IOCB_CMD_FSYNC:
2011 return aio_fsync(&req->fsync, 1822 return aio_fsync(&req->fsync, iocb, false);
2012 case IOCB_CMD_FDSYNC: 1823 case IOCB_CMD_FDSYNC:
2013 return aio_fsync(&req->fsync, 1824 return aio_fsync(&req->fsync, iocb, true);
2014 case IOCB_CMD_POLL: 1825 case IOCB_CMD_POLL:
2015 return aio_poll(req, iocb); 1826 return aio_poll(req, iocb);
2016 default: 1827 default:
2017 pr_debug("invalid aio operati 1828 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
2018 return -EINVAL; 1829 return -EINVAL;
2019 } 1830 }
2020 } 1831 }
2021 1832
2022 static int io_submit_one(struct kioctx *ctx, 1833 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
2023 bool compat) 1834 bool compat)
2024 { 1835 {
2025 struct aio_kiocb *req; 1836 struct aio_kiocb *req;
2026 struct iocb iocb; 1837 struct iocb iocb;
2027 int err; 1838 int err;
2028 1839
2029 if (unlikely(copy_from_user(&iocb, us 1840 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
2030 return -EFAULT; 1841 return -EFAULT;
2031 1842
2032 /* enforce forwards compatibility on 1843 /* enforce forwards compatibility on users */
2033 if (unlikely(iocb.aio_reserved2)) { 1844 if (unlikely(iocb.aio_reserved2)) {
2034 pr_debug("EINVAL: reserve fie 1845 pr_debug("EINVAL: reserve field set\n");
2035 return -EINVAL; 1846 return -EINVAL;
2036 } 1847 }
2037 1848
2038 /* prevent overflows */ 1849 /* prevent overflows */
2039 if (unlikely( 1850 if (unlikely(
2040 (iocb.aio_buf != (unsigned long)i 1851 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2041 (iocb.aio_nbytes != (size_t)iocb. 1852 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2042 ((ssize_t)iocb.aio_nbytes < 0) 1853 ((ssize_t)iocb.aio_nbytes < 0)
2043 )) { 1854 )) {
2044 pr_debug("EINVAL: overflow ch 1855 pr_debug("EINVAL: overflow check\n");
2045 return -EINVAL; 1856 return -EINVAL;
2046 } 1857 }
2047 1858
2048 req = aio_get_req(ctx); 1859 req = aio_get_req(ctx);
2049 if (unlikely(!req)) 1860 if (unlikely(!req))
2050 return -EAGAIN; 1861 return -EAGAIN;
2051 1862
2052 err = __io_submit_one(ctx, &iocb, use 1863 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2053 1864
2054 /* Done with the synchronous referenc 1865 /* Done with the synchronous reference */
2055 iocb_put(req); 1866 iocb_put(req);
2056 1867
2057 /* 1868 /*
2058 * If err is 0, we'd either done aio_ 1869 * If err is 0, we'd either done aio_complete() ourselves or have
2059 * arranged for that to be done async 1870 * arranged for that to be done asynchronously. Anything non-zero
2060 * means that we need to destroy req 1871 * means that we need to destroy req ourselves.
2061 */ 1872 */
2062 if (unlikely(err)) { 1873 if (unlikely(err)) {
2063 iocb_destroy(req); 1874 iocb_destroy(req);
2064 put_reqs_available(ctx, 1); 1875 put_reqs_available(ctx, 1);
2065 } 1876 }
2066 return err; 1877 return err;
2067 } 1878 }
2068 1879
2069 /* sys_io_submit: 1880 /* sys_io_submit:
2070 * Queue the nr iocbs pointed to by iocb 1881 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2071 * the number of iocbs queued. May retu 1882 * the number of iocbs queued. May return -EINVAL if the aio_context
2072 * specified by ctx_id is invalid, if nr 1883 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2073 * *iocbpp[0] is not properly initialize 1884 * *iocbpp[0] is not properly initialized, if the operation specified
2074 * is invalid for the file descriptor in 1885 * is invalid for the file descriptor in the iocb. May fail with
2075 * -EFAULT if any of the data structures 1886 * -EFAULT if any of the data structures point to invalid data. May
2076 * fail with -EBADF if the file descript 1887 * fail with -EBADF if the file descriptor specified in the first
2077 * iocb is invalid. May fail with -EAGA 1888 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2078 * are available to queue any iocbs. Wi 1889 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2079 * fail with -ENOSYS if not implemented. 1890 * fail with -ENOSYS if not implemented.
2080 */ 1891 */
2081 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx 1892 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2082 struct iocb __user * __user * 1893 struct iocb __user * __user *, iocbpp)
2083 { 1894 {
2084 struct kioctx *ctx; 1895 struct kioctx *ctx;
2085 long ret = 0; 1896 long ret = 0;
2086 int i = 0; 1897 int i = 0;
2087 struct blk_plug plug; 1898 struct blk_plug plug;
2088 1899
2089 if (unlikely(nr < 0)) 1900 if (unlikely(nr < 0))
2090 return -EINVAL; 1901 return -EINVAL;
2091 1902
2092 ctx = lookup_ioctx(ctx_id); 1903 ctx = lookup_ioctx(ctx_id);
2093 if (unlikely(!ctx)) { 1904 if (unlikely(!ctx)) {
2094 pr_debug("EINVAL: invalid con 1905 pr_debug("EINVAL: invalid context id\n");
2095 return -EINVAL; 1906 return -EINVAL;
2096 } 1907 }
2097 1908
2098 if (nr > ctx->nr_events) 1909 if (nr > ctx->nr_events)
2099 nr = ctx->nr_events; 1910 nr = ctx->nr_events;
2100 1911
2101 if (nr > AIO_PLUG_THRESHOLD) 1912 if (nr > AIO_PLUG_THRESHOLD)
2102 blk_start_plug(&plug); 1913 blk_start_plug(&plug);
2103 for (i = 0; i < nr; i++) { 1914 for (i = 0; i < nr; i++) {
2104 struct iocb __user *user_iocb 1915 struct iocb __user *user_iocb;
2105 1916
2106 if (unlikely(get_user(user_io 1917 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2107 ret = -EFAULT; 1918 ret = -EFAULT;
2108 break; 1919 break;
2109 } 1920 }
2110 1921
2111 ret = io_submit_one(ctx, user 1922 ret = io_submit_one(ctx, user_iocb, false);
2112 if (ret) 1923 if (ret)
2113 break; 1924 break;
2114 } 1925 }
2115 if (nr > AIO_PLUG_THRESHOLD) 1926 if (nr > AIO_PLUG_THRESHOLD)
2116 blk_finish_plug(&plug); 1927 blk_finish_plug(&plug);
2117 1928
2118 percpu_ref_put(&ctx->users); 1929 percpu_ref_put(&ctx->users);
2119 return i ? i : ret; 1930 return i ? i : ret;
2120 } 1931 }
2121 1932
2122 #ifdef CONFIG_COMPAT 1933 #ifdef CONFIG_COMPAT
2123 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_ 1934 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2124 int, nr, compat_uptr_t 1935 int, nr, compat_uptr_t __user *, iocbpp)
2125 { 1936 {
2126 struct kioctx *ctx; 1937 struct kioctx *ctx;
2127 long ret = 0; 1938 long ret = 0;
2128 int i = 0; 1939 int i = 0;
2129 struct blk_plug plug; 1940 struct blk_plug plug;
2130 1941
2131 if (unlikely(nr < 0)) 1942 if (unlikely(nr < 0))
2132 return -EINVAL; 1943 return -EINVAL;
2133 1944
2134 ctx = lookup_ioctx(ctx_id); 1945 ctx = lookup_ioctx(ctx_id);
2135 if (unlikely(!ctx)) { 1946 if (unlikely(!ctx)) {
2136 pr_debug("EINVAL: invalid con 1947 pr_debug("EINVAL: invalid context id\n");
2137 return -EINVAL; 1948 return -EINVAL;
2138 } 1949 }
2139 1950
2140 if (nr > ctx->nr_events) 1951 if (nr > ctx->nr_events)
2141 nr = ctx->nr_events; 1952 nr = ctx->nr_events;
2142 1953
2143 if (nr > AIO_PLUG_THRESHOLD) 1954 if (nr > AIO_PLUG_THRESHOLD)
2144 blk_start_plug(&plug); 1955 blk_start_plug(&plug);
2145 for (i = 0; i < nr; i++) { 1956 for (i = 0; i < nr; i++) {
2146 compat_uptr_t user_iocb; 1957 compat_uptr_t user_iocb;
2147 1958
2148 if (unlikely(get_user(user_io 1959 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2149 ret = -EFAULT; 1960 ret = -EFAULT;
2150 break; 1961 break;
2151 } 1962 }
2152 1963
2153 ret = io_submit_one(ctx, comp 1964 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2154 if (ret) 1965 if (ret)
2155 break; 1966 break;
2156 } 1967 }
2157 if (nr > AIO_PLUG_THRESHOLD) 1968 if (nr > AIO_PLUG_THRESHOLD)
2158 blk_finish_plug(&plug); 1969 blk_finish_plug(&plug);
2159 1970
2160 percpu_ref_put(&ctx->users); 1971 percpu_ref_put(&ctx->users);
2161 return i ? i : ret; 1972 return i ? i : ret;
2162 } 1973 }
2163 #endif 1974 #endif
2164 1975
2165 /* sys_io_cancel: 1976 /* sys_io_cancel:
2166 * Attempts to cancel an iocb previously 1977 * Attempts to cancel an iocb previously passed to io_submit. If
2167 * the operation is successfully cancell 1978 * the operation is successfully cancelled, the resulting event is
2168 * copied into the memory pointed to by 1979 * copied into the memory pointed to by result without being placed
2169 * into the completion queue and 0 is re 1980 * into the completion queue and 0 is returned. May fail with
2170 * -EFAULT if any of the data structures 1981 * -EFAULT if any of the data structures pointed to are invalid.
2171 * May fail with -EINVAL if aio_context 1982 * May fail with -EINVAL if aio_context specified by ctx_id is
2172 * invalid. May fail with -EAGAIN if th 1983 * invalid. May fail with -EAGAIN if the iocb specified was not
2173 * cancelled. Will fail with -ENOSYS if 1984 * cancelled. Will fail with -ENOSYS if not implemented.
2174 */ 1985 */
2175 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx 1986 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2176 struct io_event __user *, res 1987 struct io_event __user *, result)
2177 { 1988 {
2178 struct kioctx *ctx; 1989 struct kioctx *ctx;
2179 struct aio_kiocb *kiocb; 1990 struct aio_kiocb *kiocb;
2180 int ret = -EINVAL; 1991 int ret = -EINVAL;
2181 u32 key; 1992 u32 key;
2182 u64 obj = (u64)(unsigned long)iocb; 1993 u64 obj = (u64)(unsigned long)iocb;
2183 1994
2184 if (unlikely(get_user(key, &iocb->aio 1995 if (unlikely(get_user(key, &iocb->aio_key)))
2185 return -EFAULT; 1996 return -EFAULT;
2186 if (unlikely(key != KIOCB_KEY)) 1997 if (unlikely(key != KIOCB_KEY))
2187 return -EINVAL; 1998 return -EINVAL;
2188 1999
2189 ctx = lookup_ioctx(ctx_id); 2000 ctx = lookup_ioctx(ctx_id);
2190 if (unlikely(!ctx)) 2001 if (unlikely(!ctx))
2191 return -EINVAL; 2002 return -EINVAL;
2192 2003
2193 spin_lock_irq(&ctx->ctx_lock); 2004 spin_lock_irq(&ctx->ctx_lock);
2194 /* TODO: use a hash or array, this su 2005 /* TODO: use a hash or array, this sucks. */
2195 list_for_each_entry(kiocb, &ctx->acti 2006 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2196 if (kiocb->ki_res.obj == obj) 2007 if (kiocb->ki_res.obj == obj) {
2197 ret = kiocb->ki_cance 2008 ret = kiocb->ki_cancel(&kiocb->rw);
2198 list_del_init(&kiocb- 2009 list_del_init(&kiocb->ki_list);
2199 break; 2010 break;
2200 } 2011 }
2201 } 2012 }
2202 spin_unlock_irq(&ctx->ctx_lock); 2013 spin_unlock_irq(&ctx->ctx_lock);
2203 2014
2204 if (!ret) { 2015 if (!ret) {
2205 /* 2016 /*
2206 * The result argument is no 2017 * The result argument is no longer used - the io_event is
2207 * always delivered via the r 2018 * always delivered via the ring buffer. -EINPROGRESS indicates
2208 * cancellation is progress: 2019 * cancellation is progress:
2209 */ 2020 */
2210 ret = -EINPROGRESS; 2021 ret = -EINPROGRESS;
2211 } 2022 }
2212 2023
2213 percpu_ref_put(&ctx->users); 2024 percpu_ref_put(&ctx->users);
2214 2025
2215 return ret; 2026 return ret;
2216 } 2027 }
2217 2028
2218 static long do_io_getevents(aio_context_t ctx 2029 static long do_io_getevents(aio_context_t ctx_id,
2219 long min_nr, 2030 long min_nr,
2220 long nr, 2031 long nr,
2221 struct io_event __user *event 2032 struct io_event __user *events,
2222 struct timespec64 *ts) 2033 struct timespec64 *ts)
2223 { 2034 {
2224 ktime_t until = ts ? timespec64_to_kt 2035 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2225 struct kioctx *ioctx = lookup_ioctx(c 2036 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2226 long ret = -EINVAL; 2037 long ret = -EINVAL;
2227 2038
2228 if (likely(ioctx)) { 2039 if (likely(ioctx)) {
2229 if (likely(min_nr <= nr && mi 2040 if (likely(min_nr <= nr && min_nr >= 0))
2230 ret = read_events(ioc 2041 ret = read_events(ioctx, min_nr, nr, events, until);
2231 percpu_ref_put(&ioctx->users) 2042 percpu_ref_put(&ioctx->users);
2232 } 2043 }
2233 2044
2234 return ret; 2045 return ret;
2235 } 2046 }
2236 2047
2237 /* io_getevents: 2048 /* io_getevents:
2238 * Attempts to read at least min_nr even 2049 * Attempts to read at least min_nr events and up to nr events from
2239 * the completion queue for the aio_cont 2050 * the completion queue for the aio_context specified by ctx_id. If
2240 * it succeeds, the number of read event 2051 * it succeeds, the number of read events is returned. May fail with
2241 * -EINVAL if ctx_id is invalid, if min_ 2052 * -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 2053 * out of range, if timeout is out of range. May fail with -EFAULT
2243 * if any of the memory specified is inv 2054 * if any of the memory specified is invalid. May return 0 or
2244 * < min_nr if the timeout specified by 2055 * < min_nr if the timeout specified by timeout has elapsed
2245 * before sufficient events are availabl 2056 * before sufficient events are available, where timeout == NULL
2246 * specifies an infinite timeout. Note t 2057 * specifies an infinite timeout. Note that the timeout pointed to by
2247 * timeout is relative. Will fail with 2058 * timeout is relative. Will fail with -ENOSYS if not implemented.
2248 */ 2059 */
2249 #ifdef CONFIG_64BIT !! 2060 #if !defined(CONFIG_64BIT_TIME) || defined(CONFIG_64BIT)
2250 2061
2251 SYSCALL_DEFINE5(io_getevents, aio_context_t, 2062 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2252 long, min_nr, 2063 long, min_nr,
2253 long, nr, 2064 long, nr,
2254 struct io_event __user *, eve 2065 struct io_event __user *, events,
2255 struct __kernel_timespec __us 2066 struct __kernel_timespec __user *, timeout)
2256 { 2067 {
2257 struct timespec64 ts; 2068 struct timespec64 ts;
2258 int ret; 2069 int ret;
2259 2070
2260 if (timeout && unlikely(get_timespec6 2071 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2261 return -EFAULT; 2072 return -EFAULT;
2262 2073
2263 ret = do_io_getevents(ctx_id, min_nr, 2074 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2264 if (!ret && signal_pending(current)) 2075 if (!ret && signal_pending(current))
2265 ret = -EINTR; 2076 ret = -EINTR;
2266 return ret; 2077 return ret;
2267 } 2078 }
2268 2079
2269 #endif 2080 #endif
2270 2081
2271 struct __aio_sigset { 2082 struct __aio_sigset {
2272 const sigset_t __user *sigmask; 2083 const sigset_t __user *sigmask;
2273 size_t sigsetsize; 2084 size_t sigsetsize;
2274 }; 2085 };
2275 2086
2276 SYSCALL_DEFINE6(io_pgetevents, 2087 SYSCALL_DEFINE6(io_pgetevents,
2277 aio_context_t, ctx_id, 2088 aio_context_t, ctx_id,
2278 long, min_nr, 2089 long, min_nr,
2279 long, nr, 2090 long, nr,
2280 struct io_event __user *, eve 2091 struct io_event __user *, events,
2281 struct __kernel_timespec __us 2092 struct __kernel_timespec __user *, timeout,
2282 const struct __aio_sigset __u 2093 const struct __aio_sigset __user *, usig)
2283 { 2094 {
2284 struct __aio_sigset ksig = { NULL 2095 struct __aio_sigset ksig = { NULL, };
>> 2096 sigset_t ksigmask, sigsaved;
2285 struct timespec64 ts; 2097 struct timespec64 ts;
2286 bool interrupted; 2098 bool interrupted;
2287 int ret; 2099 int ret;
2288 2100
2289 if (timeout && unlikely(get_timespec6 2101 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2290 return -EFAULT; 2102 return -EFAULT;
2291 2103
2292 if (usig && copy_from_user(&ksig, usi 2104 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2293 return -EFAULT; 2105 return -EFAULT;
2294 2106
2295 ret = set_user_sigmask(ksig.sigmask, !! 2107 ret = set_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2296 if (ret) 2108 if (ret)
2297 return ret; 2109 return ret;
2298 2110
2299 ret = do_io_getevents(ctx_id, min_nr, 2111 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2300 2112
2301 interrupted = signal_pending(current) 2113 interrupted = signal_pending(current);
2302 restore_saved_sigmask_unless(interrup !! 2114 restore_user_sigmask(ksig.sigmask, &sigsaved, interrupted);
2303 if (interrupted && !ret) 2115 if (interrupted && !ret)
2304 ret = -ERESTARTNOHAND; 2116 ret = -ERESTARTNOHAND;
2305 2117
2306 return ret; 2118 return ret;
2307 } 2119 }
2308 2120
2309 #if defined(CONFIG_COMPAT_32BIT_TIME) && !def 2121 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2310 2122
2311 SYSCALL_DEFINE6(io_pgetevents_time32, 2123 SYSCALL_DEFINE6(io_pgetevents_time32,
2312 aio_context_t, ctx_id, 2124 aio_context_t, ctx_id,
2313 long, min_nr, 2125 long, min_nr,
2314 long, nr, 2126 long, nr,
2315 struct io_event __user *, eve 2127 struct io_event __user *, events,
2316 struct old_timespec32 __user 2128 struct old_timespec32 __user *, timeout,
2317 const struct __aio_sigset __u 2129 const struct __aio_sigset __user *, usig)
2318 { 2130 {
2319 struct __aio_sigset ksig = { NULL 2131 struct __aio_sigset ksig = { NULL, };
>> 2132 sigset_t ksigmask, sigsaved;
2320 struct timespec64 ts; 2133 struct timespec64 ts;
2321 bool interrupted; 2134 bool interrupted;
2322 int ret; 2135 int ret;
2323 2136
2324 if (timeout && unlikely(get_old_times 2137 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2325 return -EFAULT; 2138 return -EFAULT;
2326 2139
2327 if (usig && copy_from_user(&ksig, usi 2140 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2328 return -EFAULT; 2141 return -EFAULT;
2329 2142
2330 2143
2331 ret = set_user_sigmask(ksig.sigmask, !! 2144 ret = set_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2332 if (ret) 2145 if (ret)
2333 return ret; 2146 return ret;
2334 2147
2335 ret = do_io_getevents(ctx_id, min_nr, 2148 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2336 2149
2337 interrupted = signal_pending(current) 2150 interrupted = signal_pending(current);
2338 restore_saved_sigmask_unless(interrup !! 2151 restore_user_sigmask(ksig.sigmask, &sigsaved, interrupted);
2339 if (interrupted && !ret) 2152 if (interrupted && !ret)
2340 ret = -ERESTARTNOHAND; 2153 ret = -ERESTARTNOHAND;
2341 2154
2342 return ret; 2155 return ret;
2343 } 2156 }
2344 2157
2345 #endif 2158 #endif
2346 2159
2347 #if defined(CONFIG_COMPAT_32BIT_TIME) 2160 #if defined(CONFIG_COMPAT_32BIT_TIME)
2348 2161
2349 SYSCALL_DEFINE5(io_getevents_time32, __u32, c 2162 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2350 __s32, min_nr, 2163 __s32, min_nr,
2351 __s32, nr, 2164 __s32, nr,
2352 struct io_event __user *, eve 2165 struct io_event __user *, events,
2353 struct old_timespec32 __user 2166 struct old_timespec32 __user *, timeout)
2354 { 2167 {
2355 struct timespec64 t; 2168 struct timespec64 t;
2356 int ret; 2169 int ret;
2357 2170
2358 if (timeout && get_old_timespec32(&t, 2171 if (timeout && get_old_timespec32(&t, timeout))
2359 return -EFAULT; 2172 return -EFAULT;
2360 2173
2361 ret = do_io_getevents(ctx_id, min_nr, 2174 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2362 if (!ret && signal_pending(current)) 2175 if (!ret && signal_pending(current))
2363 ret = -EINTR; 2176 ret = -EINTR;
2364 return ret; 2177 return ret;
2365 } 2178 }
2366 2179
2367 #endif 2180 #endif
2368 2181
2369 #ifdef CONFIG_COMPAT 2182 #ifdef CONFIG_COMPAT
2370 2183
2371 struct __compat_aio_sigset { 2184 struct __compat_aio_sigset {
2372 compat_uptr_t sigmask; !! 2185 compat_sigset_t __user *sigmask;
2373 compat_size_t sigsetsize; 2186 compat_size_t sigsetsize;
2374 }; 2187 };
2375 2188
2376 #if defined(CONFIG_COMPAT_32BIT_TIME) 2189 #if defined(CONFIG_COMPAT_32BIT_TIME)
2377 2190
2378 COMPAT_SYSCALL_DEFINE6(io_pgetevents, 2191 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2379 compat_aio_context_t, ctx_id, 2192 compat_aio_context_t, ctx_id,
2380 compat_long_t, min_nr, 2193 compat_long_t, min_nr,
2381 compat_long_t, nr, 2194 compat_long_t, nr,
2382 struct io_event __user *, eve 2195 struct io_event __user *, events,
2383 struct old_timespec32 __user 2196 struct old_timespec32 __user *, timeout,
2384 const struct __compat_aio_sig 2197 const struct __compat_aio_sigset __user *, usig)
2385 { 2198 {
2386 struct __compat_aio_sigset ksig = { 0 !! 2199 struct __compat_aio_sigset ksig = { NULL, };
>> 2200 sigset_t ksigmask, sigsaved;
2387 struct timespec64 t; 2201 struct timespec64 t;
2388 bool interrupted; 2202 bool interrupted;
2389 int ret; 2203 int ret;
2390 2204
2391 if (timeout && get_old_timespec32(&t, 2205 if (timeout && get_old_timespec32(&t, timeout))
2392 return -EFAULT; 2206 return -EFAULT;
2393 2207
2394 if (usig && copy_from_user(&ksig, usi 2208 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2395 return -EFAULT; 2209 return -EFAULT;
2396 2210
2397 ret = set_compat_user_sigmask(compat_ !! 2211 ret = set_compat_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2398 if (ret) 2212 if (ret)
2399 return ret; 2213 return ret;
2400 2214
2401 ret = do_io_getevents(ctx_id, min_nr, 2215 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2402 2216
2403 interrupted = signal_pending(current) 2217 interrupted = signal_pending(current);
2404 restore_saved_sigmask_unless(interrup !! 2218 restore_user_sigmask(ksig.sigmask, &sigsaved, interrupted);
2405 if (interrupted && !ret) 2219 if (interrupted && !ret)
2406 ret = -ERESTARTNOHAND; 2220 ret = -ERESTARTNOHAND;
2407 2221
2408 return ret; 2222 return ret;
2409 } 2223 }
2410 2224
2411 #endif 2225 #endif
2412 2226
2413 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64, 2227 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2414 compat_aio_context_t, ctx_id, 2228 compat_aio_context_t, ctx_id,
2415 compat_long_t, min_nr, 2229 compat_long_t, min_nr,
2416 compat_long_t, nr, 2230 compat_long_t, nr,
2417 struct io_event __user *, eve 2231 struct io_event __user *, events,
2418 struct __kernel_timespec __us 2232 struct __kernel_timespec __user *, timeout,
2419 const struct __compat_aio_sig 2233 const struct __compat_aio_sigset __user *, usig)
2420 { 2234 {
2421 struct __compat_aio_sigset ksig = { 0 !! 2235 struct __compat_aio_sigset ksig = { NULL, };
>> 2236 sigset_t ksigmask, sigsaved;
2422 struct timespec64 t; 2237 struct timespec64 t;
2423 bool interrupted; 2238 bool interrupted;
2424 int ret; 2239 int ret;
2425 2240
2426 if (timeout && get_timespec64(&t, tim 2241 if (timeout && get_timespec64(&t, timeout))
2427 return -EFAULT; 2242 return -EFAULT;
2428 2243
2429 if (usig && copy_from_user(&ksig, usi 2244 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2430 return -EFAULT; 2245 return -EFAULT;
2431 2246
2432 ret = set_compat_user_sigmask(compat_ !! 2247 ret = set_compat_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2433 if (ret) 2248 if (ret)
2434 return ret; 2249 return ret;
2435 2250
2436 ret = do_io_getevents(ctx_id, min_nr, 2251 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2437 2252
2438 interrupted = signal_pending(current) 2253 interrupted = signal_pending(current);
2439 restore_saved_sigmask_unless(interrup !! 2254 restore_user_sigmask(ksig.sigmask, &sigsaved, interrupted);
2440 if (interrupted && !ret) 2255 if (interrupted && !ret)
2441 ret = -ERESTARTNOHAND; 2256 ret = -ERESTARTNOHAND;
2442 2257
2443 return ret; 2258 return ret;
2444 } 2259 }
2445 #endif 2260 #endif
2446 2261
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.