Version:  2.0.40 2.2.26 2.4.37 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18

Linux/fs/aio.c

  1 /*
  2  *      An async IO implementation for Linux
  3  *      Written by Benjamin LaHaise <bcrl@kvack.org>
  4  *
  5  *      Implements an efficient asynchronous io interface.
  6  *
  7  *      Copyright 2000, 2001, 2002 Red Hat, Inc.  All Rights Reserved.
  8  *
  9  *      See ../COPYING for licensing terms.
 10  */
 11 #define pr_fmt(fmt) "%s: " fmt, __func__
 12 
 13 #include <linux/kernel.h>
 14 #include <linux/init.h>
 15 #include <linux/errno.h>
 16 #include <linux/time.h>
 17 #include <linux/aio_abi.h>
 18 #include <linux/export.h>
 19 #include <linux/syscalls.h>
 20 #include <linux/backing-dev.h>
 21 #include <linux/uio.h>
 22 
 23 #include <linux/sched.h>
 24 #include <linux/fs.h>
 25 #include <linux/file.h>
 26 #include <linux/mm.h>
 27 #include <linux/mman.h>
 28 #include <linux/mmu_context.h>
 29 #include <linux/percpu.h>
 30 #include <linux/slab.h>
 31 #include <linux/timer.h>
 32 #include <linux/aio.h>
 33 #include <linux/highmem.h>
 34 #include <linux/workqueue.h>
 35 #include <linux/security.h>
 36 #include <linux/eventfd.h>
 37 #include <linux/blkdev.h>
 38 #include <linux/compat.h>
 39 #include <linux/migrate.h>
 40 #include <linux/ramfs.h>
 41 #include <linux/percpu-refcount.h>
 42 #include <linux/mount.h>
 43 
 44 #include <asm/kmap_types.h>
 45 #include <asm/uaccess.h>
 46 
 47 #include "internal.h"
 48 
 49 #define AIO_RING_MAGIC                  0xa10a10a1
 50 #define AIO_RING_COMPAT_FEATURES        1
 51 #define AIO_RING_INCOMPAT_FEATURES      0
 52 struct aio_ring {
 53         unsigned        id;     /* kernel internal index number */
 54         unsigned        nr;     /* number of io_events */
 55         unsigned        head;   /* Written to by userland or under ring_lock
 56                                  * mutex by aio_read_events_ring(). */
 57         unsigned        tail;
 58 
 59         unsigned        magic;
 60         unsigned        compat_features;
 61         unsigned        incompat_features;
 62         unsigned        header_length;  /* size of aio_ring */
 63 
 64 
 65         struct io_event         io_events[0];
 66 }; /* 128 bytes + ring size */
 67 
 68 #define AIO_RING_PAGES  8
 69 
 70 struct kioctx_table {
 71         struct rcu_head rcu;
 72         unsigned        nr;
 73         struct kioctx   *table[];
 74 };
 75 
 76 struct kioctx_cpu {
 77         unsigned                reqs_available;
 78 };
 79 
 80 struct kioctx {
 81         struct percpu_ref       users;
 82         atomic_t                dead;
 83 
 84         struct percpu_ref       reqs;
 85 
 86         unsigned long           user_id;
 87 
 88         struct __percpu kioctx_cpu *cpu;
 89 
 90         /*
 91          * For percpu reqs_available, number of slots we move to/from global
 92          * counter at a time:
 93          */
 94         unsigned                req_batch;
 95         /*
 96          * This is what userspace passed to io_setup(), it's not used for
 97          * anything but counting against the global max_reqs quota.
 98          *
 99          * The real limit is nr_events - 1, which will be larger (see
100          * aio_setup_ring())
101          */
102         unsigned                max_reqs;
103 
104         /* Size of ringbuffer, in units of struct io_event */
105         unsigned                nr_events;
106 
107         unsigned long           mmap_base;
108         unsigned long           mmap_size;
109 
110         struct page             **ring_pages;
111         long                    nr_pages;
112 
113         struct work_struct      free_work;
114 
115         /*
116          * signals when all in-flight requests are done
117          */
118         struct completion *requests_done;
119 
120         struct {
121                 /*
122                  * This counts the number of available slots in the ringbuffer,
123                  * so we avoid overflowing it: it's decremented (if positive)
124                  * when allocating a kiocb and incremented when the resulting
125                  * io_event is pulled off the ringbuffer.
126                  *
127                  * We batch accesses to it with a percpu version.
128                  */
129                 atomic_t        reqs_available;
130         } ____cacheline_aligned_in_smp;
131 
132         struct {
133                 spinlock_t      ctx_lock;
134                 struct list_head active_reqs;   /* used for cancellation */
135         } ____cacheline_aligned_in_smp;
136 
137         struct {
138                 struct mutex    ring_lock;
139                 wait_queue_head_t wait;
140         } ____cacheline_aligned_in_smp;
141 
142         struct {
143                 unsigned        tail;
144                 unsigned        completed_events;
145                 spinlock_t      completion_lock;
146         } ____cacheline_aligned_in_smp;
147 
148         struct page             *internal_pages[AIO_RING_PAGES];
149         struct file             *aio_ring_file;
150 
151         unsigned                id;
152 };
153 
154 /*------ sysctl variables----*/
155 static DEFINE_SPINLOCK(aio_nr_lock);
156 unsigned long aio_nr;           /* current system wide number of aio requests */
157 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
158 /*----end sysctl variables---*/
159 
160 static struct kmem_cache        *kiocb_cachep;
161 static struct kmem_cache        *kioctx_cachep;
162 
163 static struct vfsmount *aio_mnt;
164 
165 static const struct file_operations aio_ring_fops;
166 static const struct address_space_operations aio_ctx_aops;
167 
168 /* Backing dev info for aio fs.
169  * -no dirty page accounting or writeback happens
170  */
171 static struct backing_dev_info aio_fs_backing_dev_info = {
172         .name           = "aiofs",
173         .state          = 0,
174         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_MAP_COPY,
175 };
176 
177 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
178 {
179         struct qstr this = QSTR_INIT("[aio]", 5);
180         struct file *file;
181         struct path path;
182         struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
183         if (IS_ERR(inode))
184                 return ERR_CAST(inode);
185 
186         inode->i_mapping->a_ops = &aio_ctx_aops;
187         inode->i_mapping->private_data = ctx;
188         inode->i_mapping->backing_dev_info = &aio_fs_backing_dev_info;
189         inode->i_size = PAGE_SIZE * nr_pages;
190 
191         path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
192         if (!path.dentry) {
193                 iput(inode);
194                 return ERR_PTR(-ENOMEM);
195         }
196         path.mnt = mntget(aio_mnt);
197 
198         d_instantiate(path.dentry, inode);
199         file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
200         if (IS_ERR(file)) {
201                 path_put(&path);
202                 return file;
203         }
204 
205         file->f_flags = O_RDWR;
206         return file;
207 }
208 
209 static struct dentry *aio_mount(struct file_system_type *fs_type,
210                                 int flags, const char *dev_name, void *data)
211 {
212         static const struct dentry_operations ops = {
213                 .d_dname        = simple_dname,
214         };
215         return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
216 }
217 
218 /* aio_setup
219  *      Creates the slab caches used by the aio routines, panic on
220  *      failure as this is done early during the boot sequence.
221  */
222 static int __init aio_setup(void)
223 {
224         static struct file_system_type aio_fs = {
225                 .name           = "aio",
226                 .mount          = aio_mount,
227                 .kill_sb        = kill_anon_super,
228         };
229         aio_mnt = kern_mount(&aio_fs);
230         if (IS_ERR(aio_mnt))
231                 panic("Failed to create aio fs mount.");
232 
233         if (bdi_init(&aio_fs_backing_dev_info))
234                 panic("Failed to init aio fs backing dev info.");
235 
236         kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
237         kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
238 
239         pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
240 
241         return 0;
242 }
243 __initcall(aio_setup);
244 
245 static void put_aio_ring_file(struct kioctx *ctx)
246 {
247         struct file *aio_ring_file = ctx->aio_ring_file;
248         if (aio_ring_file) {
249                 truncate_setsize(aio_ring_file->f_inode, 0);
250 
251                 /* Prevent further access to the kioctx from migratepages */
252                 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
253                 aio_ring_file->f_inode->i_mapping->private_data = NULL;
254                 ctx->aio_ring_file = NULL;
255                 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
256 
257                 fput(aio_ring_file);
258         }
259 }
260 
261 static void aio_free_ring(struct kioctx *ctx)
262 {
263         int i;
264 
265         /* Disconnect the kiotx from the ring file.  This prevents future
266          * accesses to the kioctx from page migration.
267          */
268         put_aio_ring_file(ctx);
269 
270         for (i = 0; i < ctx->nr_pages; i++) {
271                 struct page *page;
272                 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
273                                 page_count(ctx->ring_pages[i]));
274                 page = ctx->ring_pages[i];
275                 if (!page)
276                         continue;
277                 ctx->ring_pages[i] = NULL;
278                 put_page(page);
279         }
280 
281         if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
282                 kfree(ctx->ring_pages);
283                 ctx->ring_pages = NULL;
284         }
285 }
286 
287 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
288 {
289         vma->vm_ops = &generic_file_vm_ops;
290         return 0;
291 }
292 
293 static const struct file_operations aio_ring_fops = {
294         .mmap = aio_ring_mmap,
295 };
296 
297 #if IS_ENABLED(CONFIG_MIGRATION)
298 static int aio_migratepage(struct address_space *mapping, struct page *new,
299                         struct page *old, enum migrate_mode mode)
300 {
301         struct kioctx *ctx;
302         unsigned long flags;
303         pgoff_t idx;
304         int rc;
305 
306         rc = 0;
307 
308         /* mapping->private_lock here protects against the kioctx teardown.  */
309         spin_lock(&mapping->private_lock);
310         ctx = mapping->private_data;
311         if (!ctx) {
312                 rc = -EINVAL;
313                 goto out;
314         }
315 
316         /* The ring_lock mutex.  The prevents aio_read_events() from writing
317          * to the ring's head, and prevents page migration from mucking in
318          * a partially initialized kiotx.
319          */
320         if (!mutex_trylock(&ctx->ring_lock)) {
321                 rc = -EAGAIN;
322                 goto out;
323         }
324 
325         idx = old->index;
326         if (idx < (pgoff_t)ctx->nr_pages) {
327                 /* Make sure the old page hasn't already been changed */
328                 if (ctx->ring_pages[idx] != old)
329                         rc = -EAGAIN;
330         } else
331                 rc = -EINVAL;
332 
333         if (rc != 0)
334                 goto out_unlock;
335 
336         /* Writeback must be complete */
337         BUG_ON(PageWriteback(old));
338         get_page(new);
339 
340         rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
341         if (rc != MIGRATEPAGE_SUCCESS) {
342                 put_page(new);
343                 goto out_unlock;
344         }
345 
346         /* Take completion_lock to prevent other writes to the ring buffer
347          * while the old page is copied to the new.  This prevents new
348          * events from being lost.
349          */
350         spin_lock_irqsave(&ctx->completion_lock, flags);
351         migrate_page_copy(new, old);
352         BUG_ON(ctx->ring_pages[idx] != old);
353         ctx->ring_pages[idx] = new;
354         spin_unlock_irqrestore(&ctx->completion_lock, flags);
355 
356         /* The old page is no longer accessible. */
357         put_page(old);
358 
359 out_unlock:
360         mutex_unlock(&ctx->ring_lock);
361 out:
362         spin_unlock(&mapping->private_lock);
363         return rc;
364 }
365 #endif
366 
367 static const struct address_space_operations aio_ctx_aops = {
368         .set_page_dirty = __set_page_dirty_no_writeback,
369 #if IS_ENABLED(CONFIG_MIGRATION)
370         .migratepage    = aio_migratepage,
371 #endif
372 };
373 
374 static int aio_setup_ring(struct kioctx *ctx)
375 {
376         struct aio_ring *ring;
377         unsigned nr_events = ctx->max_reqs;
378         struct mm_struct *mm = current->mm;
379         unsigned long size, unused;
380         int nr_pages;
381         int i;
382         struct file *file;
383 
384         /* Compensate for the ring buffer's head/tail overlap entry */
385         nr_events += 2; /* 1 is required, 2 for good luck */
386 
387         size = sizeof(struct aio_ring);
388         size += sizeof(struct io_event) * nr_events;
389 
390         nr_pages = PFN_UP(size);
391         if (nr_pages < 0)
392                 return -EINVAL;
393 
394         file = aio_private_file(ctx, nr_pages);
395         if (IS_ERR(file)) {
396                 ctx->aio_ring_file = NULL;
397                 return -ENOMEM;
398         }
399 
400         ctx->aio_ring_file = file;
401         nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
402                         / sizeof(struct io_event);
403 
404         ctx->ring_pages = ctx->internal_pages;
405         if (nr_pages > AIO_RING_PAGES) {
406                 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
407                                           GFP_KERNEL);
408                 if (!ctx->ring_pages) {
409                         put_aio_ring_file(ctx);
410                         return -ENOMEM;
411                 }
412         }
413 
414         for (i = 0; i < nr_pages; i++) {
415                 struct page *page;
416                 page = find_or_create_page(file->f_inode->i_mapping,
417                                            i, GFP_HIGHUSER | __GFP_ZERO);
418                 if (!page)
419                         break;
420                 pr_debug("pid(%d) page[%d]->count=%d\n",
421                          current->pid, i, page_count(page));
422                 SetPageUptodate(page);
423                 unlock_page(page);
424 
425                 ctx->ring_pages[i] = page;
426         }
427         ctx->nr_pages = i;
428 
429         if (unlikely(i != nr_pages)) {
430                 aio_free_ring(ctx);
431                 return -ENOMEM;
432         }
433 
434         ctx->mmap_size = nr_pages * PAGE_SIZE;
435         pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
436 
437         down_write(&mm->mmap_sem);
438         ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
439                                        PROT_READ | PROT_WRITE,
440                                        MAP_SHARED, 0, &unused);
441         up_write(&mm->mmap_sem);
442         if (IS_ERR((void *)ctx->mmap_base)) {
443                 ctx->mmap_size = 0;
444                 aio_free_ring(ctx);
445                 return -ENOMEM;
446         }
447 
448         pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
449 
450         ctx->user_id = ctx->mmap_base;
451         ctx->nr_events = nr_events; /* trusted copy */
452 
453         ring = kmap_atomic(ctx->ring_pages[0]);
454         ring->nr = nr_events;   /* user copy */
455         ring->id = ~0U;
456         ring->head = ring->tail = 0;
457         ring->magic = AIO_RING_MAGIC;
458         ring->compat_features = AIO_RING_COMPAT_FEATURES;
459         ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
460         ring->header_length = sizeof(struct aio_ring);
461         kunmap_atomic(ring);
462         flush_dcache_page(ctx->ring_pages[0]);
463 
464         return 0;
465 }
466 
467 #define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
468 #define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
469 #define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
470 
471 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
472 {
473         struct kioctx *ctx = req->ki_ctx;
474         unsigned long flags;
475 
476         spin_lock_irqsave(&ctx->ctx_lock, flags);
477 
478         if (!req->ki_list.next)
479                 list_add(&req->ki_list, &ctx->active_reqs);
480 
481         req->ki_cancel = cancel;
482 
483         spin_unlock_irqrestore(&ctx->ctx_lock, flags);
484 }
485 EXPORT_SYMBOL(kiocb_set_cancel_fn);
486 
487 static int kiocb_cancel(struct kiocb *kiocb)
488 {
489         kiocb_cancel_fn *old, *cancel;
490 
491         /*
492          * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
493          * actually has a cancel function, hence the cmpxchg()
494          */
495 
496         cancel = ACCESS_ONCE(kiocb->ki_cancel);
497         do {
498                 if (!cancel || cancel == KIOCB_CANCELLED)
499                         return -EINVAL;
500 
501                 old = cancel;
502                 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
503         } while (cancel != old);
504 
505         return cancel(kiocb);
506 }
507 
508 static void free_ioctx(struct work_struct *work)
509 {
510         struct kioctx *ctx = container_of(work, struct kioctx, free_work);
511 
512         pr_debug("freeing %p\n", ctx);
513 
514         aio_free_ring(ctx);
515         free_percpu(ctx->cpu);
516         percpu_ref_exit(&ctx->reqs);
517         percpu_ref_exit(&ctx->users);
518         kmem_cache_free(kioctx_cachep, ctx);
519 }
520 
521 static void free_ioctx_reqs(struct percpu_ref *ref)
522 {
523         struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
524 
525         /* At this point we know that there are no any in-flight requests */
526         if (ctx->requests_done)
527                 complete(ctx->requests_done);
528 
529         INIT_WORK(&ctx->free_work, free_ioctx);
530         schedule_work(&ctx->free_work);
531 }
532 
533 /*
534  * When this function runs, the kioctx has been removed from the "hash table"
535  * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
536  * now it's safe to cancel any that need to be.
537  */
538 static void free_ioctx_users(struct percpu_ref *ref)
539 {
540         struct kioctx *ctx = container_of(ref, struct kioctx, users);
541         struct kiocb *req;
542 
543         spin_lock_irq(&ctx->ctx_lock);
544 
545         while (!list_empty(&ctx->active_reqs)) {
546                 req = list_first_entry(&ctx->active_reqs,
547                                        struct kiocb, ki_list);
548 
549                 list_del_init(&req->ki_list);
550                 kiocb_cancel(req);
551         }
552 
553         spin_unlock_irq(&ctx->ctx_lock);
554 
555         percpu_ref_kill(&ctx->reqs);
556         percpu_ref_put(&ctx->reqs);
557 }
558 
559 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
560 {
561         unsigned i, new_nr;
562         struct kioctx_table *table, *old;
563         struct aio_ring *ring;
564 
565         spin_lock(&mm->ioctx_lock);
566         table = rcu_dereference_raw(mm->ioctx_table);
567 
568         while (1) {
569                 if (table)
570                         for (i = 0; i < table->nr; i++)
571                                 if (!table->table[i]) {
572                                         ctx->id = i;
573                                         table->table[i] = ctx;
574                                         spin_unlock(&mm->ioctx_lock);
575 
576                                         /* While kioctx setup is in progress,
577                                          * we are protected from page migration
578                                          * changes ring_pages by ->ring_lock.
579                                          */
580                                         ring = kmap_atomic(ctx->ring_pages[0]);
581                                         ring->id = ctx->id;
582                                         kunmap_atomic(ring);
583                                         return 0;
584                                 }
585 
586                 new_nr = (table ? table->nr : 1) * 4;
587                 spin_unlock(&mm->ioctx_lock);
588 
589                 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
590                                 new_nr, GFP_KERNEL);
591                 if (!table)
592                         return -ENOMEM;
593 
594                 table->nr = new_nr;
595 
596                 spin_lock(&mm->ioctx_lock);
597                 old = rcu_dereference_raw(mm->ioctx_table);
598 
599                 if (!old) {
600                         rcu_assign_pointer(mm->ioctx_table, table);
601                 } else if (table->nr > old->nr) {
602                         memcpy(table->table, old->table,
603                                old->nr * sizeof(struct kioctx *));
604 
605                         rcu_assign_pointer(mm->ioctx_table, table);
606                         kfree_rcu(old, rcu);
607                 } else {
608                         kfree(table);
609                         table = old;
610                 }
611         }
612 }
613 
614 static void aio_nr_sub(unsigned nr)
615 {
616         spin_lock(&aio_nr_lock);
617         if (WARN_ON(aio_nr - nr > aio_nr))
618                 aio_nr = 0;
619         else
620                 aio_nr -= nr;
621         spin_unlock(&aio_nr_lock);
622 }
623 
624 /* ioctx_alloc
625  *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
626  */
627 static struct kioctx *ioctx_alloc(unsigned nr_events)
628 {
629         struct mm_struct *mm = current->mm;
630         struct kioctx *ctx;
631         int err = -ENOMEM;
632 
633         /*
634          * We keep track of the number of available ringbuffer slots, to prevent
635          * overflow (reqs_available), and we also use percpu counters for this.
636          *
637          * So since up to half the slots might be on other cpu's percpu counters
638          * and unavailable, double nr_events so userspace sees what they
639          * expected: additionally, we move req_batch slots to/from percpu
640          * counters at a time, so make sure that isn't 0:
641          */
642         nr_events = max(nr_events, num_possible_cpus() * 4);
643         nr_events *= 2;
644 
645         /* Prevent overflows */
646         if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
647             (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
648                 pr_debug("ENOMEM: nr_events too high\n");
649                 return ERR_PTR(-EINVAL);
650         }
651 
652         if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
653                 return ERR_PTR(-EAGAIN);
654 
655         ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
656         if (!ctx)
657                 return ERR_PTR(-ENOMEM);
658 
659         ctx->max_reqs = nr_events;
660 
661         spin_lock_init(&ctx->ctx_lock);
662         spin_lock_init(&ctx->completion_lock);
663         mutex_init(&ctx->ring_lock);
664         /* Protect against page migration throughout kiotx setup by keeping
665          * the ring_lock mutex held until setup is complete. */
666         mutex_lock(&ctx->ring_lock);
667         init_waitqueue_head(&ctx->wait);
668 
669         INIT_LIST_HEAD(&ctx->active_reqs);
670 
671         if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
672                 goto err;
673 
674         if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
675                 goto err;
676 
677         ctx->cpu = alloc_percpu(struct kioctx_cpu);
678         if (!ctx->cpu)
679                 goto err;
680 
681         err = aio_setup_ring(ctx);
682         if (err < 0)
683                 goto err;
684 
685         atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
686         ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
687         if (ctx->req_batch < 1)
688                 ctx->req_batch = 1;
689 
690         /* limit the number of system wide aios */
691         spin_lock(&aio_nr_lock);
692         if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
693             aio_nr + nr_events < aio_nr) {
694                 spin_unlock(&aio_nr_lock);
695                 err = -EAGAIN;
696                 goto err_ctx;
697         }
698         aio_nr += ctx->max_reqs;
699         spin_unlock(&aio_nr_lock);
700 
701         percpu_ref_get(&ctx->users);    /* io_setup() will drop this ref */
702         percpu_ref_get(&ctx->reqs);     /* free_ioctx_users() will drop this */
703 
704         err = ioctx_add_table(ctx, mm);
705         if (err)
706                 goto err_cleanup;
707 
708         /* Release the ring_lock mutex now that all setup is complete. */
709         mutex_unlock(&ctx->ring_lock);
710 
711         pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
712                  ctx, ctx->user_id, mm, ctx->nr_events);
713         return ctx;
714 
715 err_cleanup:
716         aio_nr_sub(ctx->max_reqs);
717 err_ctx:
718         aio_free_ring(ctx);
719 err:
720         mutex_unlock(&ctx->ring_lock);
721         free_percpu(ctx->cpu);
722         percpu_ref_exit(&ctx->reqs);
723         percpu_ref_exit(&ctx->users);
724         kmem_cache_free(kioctx_cachep, ctx);
725         pr_debug("error allocating ioctx %d\n", err);
726         return ERR_PTR(err);
727 }
728 
729 /* kill_ioctx
730  *      Cancels all outstanding aio requests on an aio context.  Used
731  *      when the processes owning a context have all exited to encourage
732  *      the rapid destruction of the kioctx.
733  */
734 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
735                 struct completion *requests_done)
736 {
737         struct kioctx_table *table;
738 
739         if (atomic_xchg(&ctx->dead, 1))
740                 return -EINVAL;
741 
742 
743         spin_lock(&mm->ioctx_lock);
744         table = rcu_dereference_raw(mm->ioctx_table);
745         WARN_ON(ctx != table->table[ctx->id]);
746         table->table[ctx->id] = NULL;
747         spin_unlock(&mm->ioctx_lock);
748 
749         /* percpu_ref_kill() will do the necessary call_rcu() */
750         wake_up_all(&ctx->wait);
751 
752         /*
753          * It'd be more correct to do this in free_ioctx(), after all
754          * the outstanding kiocbs have finished - but by then io_destroy
755          * has already returned, so io_setup() could potentially return
756          * -EAGAIN with no ioctxs actually in use (as far as userspace
757          *  could tell).
758          */
759         aio_nr_sub(ctx->max_reqs);
760 
761         if (ctx->mmap_size)
762                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
763 
764         ctx->requests_done = requests_done;
765         percpu_ref_kill(&ctx->users);
766         return 0;
767 }
768 
769 /* wait_on_sync_kiocb:
770  *      Waits on the given sync kiocb to complete.
771  */
772 ssize_t wait_on_sync_kiocb(struct kiocb *req)
773 {
774         while (!req->ki_ctx) {
775                 set_current_state(TASK_UNINTERRUPTIBLE);
776                 if (req->ki_ctx)
777                         break;
778                 io_schedule();
779         }
780         __set_current_state(TASK_RUNNING);
781         return req->ki_user_data;
782 }
783 EXPORT_SYMBOL(wait_on_sync_kiocb);
784 
785 /*
786  * exit_aio: called when the last user of mm goes away.  At this point, there is
787  * no way for any new requests to be submited or any of the io_* syscalls to be
788  * called on the context.
789  *
790  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
791  * them.
792  */
793 void exit_aio(struct mm_struct *mm)
794 {
795         struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
796         int i;
797 
798         if (!table)
799                 return;
800 
801         for (i = 0; i < table->nr; ++i) {
802                 struct kioctx *ctx = table->table[i];
803                 struct completion requests_done =
804                         COMPLETION_INITIALIZER_ONSTACK(requests_done);
805 
806                 if (!ctx)
807                         continue;
808                 /*
809                  * We don't need to bother with munmap() here - exit_mmap(mm)
810                  * is coming and it'll unmap everything. And we simply can't,
811                  * this is not necessarily our ->mm.
812                  * Since kill_ioctx() uses non-zero ->mmap_size as indicator
813                  * that it needs to unmap the area, just set it to 0.
814                  */
815                 ctx->mmap_size = 0;
816                 kill_ioctx(mm, ctx, &requests_done);
817 
818                 /* Wait until all IO for the context are done. */
819                 wait_for_completion(&requests_done);
820         }
821 
822         RCU_INIT_POINTER(mm->ioctx_table, NULL);
823         kfree(table);
824 }
825 
826 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
827 {
828         struct kioctx_cpu *kcpu;
829         unsigned long flags;
830 
831         local_irq_save(flags);
832         kcpu = this_cpu_ptr(ctx->cpu);
833         kcpu->reqs_available += nr;
834 
835         while (kcpu->reqs_available >= ctx->req_batch * 2) {
836                 kcpu->reqs_available -= ctx->req_batch;
837                 atomic_add(ctx->req_batch, &ctx->reqs_available);
838         }
839 
840         local_irq_restore(flags);
841 }
842 
843 static bool get_reqs_available(struct kioctx *ctx)
844 {
845         struct kioctx_cpu *kcpu;
846         bool ret = false;
847         unsigned long flags;
848 
849         local_irq_save(flags);
850         kcpu = this_cpu_ptr(ctx->cpu);
851         if (!kcpu->reqs_available) {
852                 int old, avail = atomic_read(&ctx->reqs_available);
853 
854                 do {
855                         if (avail < ctx->req_batch)
856                                 goto out;
857 
858                         old = avail;
859                         avail = atomic_cmpxchg(&ctx->reqs_available,
860                                                avail, avail - ctx->req_batch);
861                 } while (avail != old);
862 
863                 kcpu->reqs_available += ctx->req_batch;
864         }
865 
866         ret = true;
867         kcpu->reqs_available--;
868 out:
869         local_irq_restore(flags);
870         return ret;
871 }
872 
873 /* refill_reqs_available
874  *      Updates the reqs_available reference counts used for tracking the
875  *      number of free slots in the completion ring.  This can be called
876  *      from aio_complete() (to optimistically update reqs_available) or
877  *      from aio_get_req() (the we're out of events case).  It must be
878  *      called holding ctx->completion_lock.
879  */
880 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
881                                   unsigned tail)
882 {
883         unsigned events_in_ring, completed;
884 
885         /* Clamp head since userland can write to it. */
886         head %= ctx->nr_events;
887         if (head <= tail)
888                 events_in_ring = tail - head;
889         else
890                 events_in_ring = ctx->nr_events - (head - tail);
891 
892         completed = ctx->completed_events;
893         if (events_in_ring < completed)
894                 completed -= events_in_ring;
895         else
896                 completed = 0;
897 
898         if (!completed)
899                 return;
900 
901         ctx->completed_events -= completed;
902         put_reqs_available(ctx, completed);
903 }
904 
905 /* user_refill_reqs_available
906  *      Called to refill reqs_available when aio_get_req() encounters an
907  *      out of space in the completion ring.
908  */
909 static void user_refill_reqs_available(struct kioctx *ctx)
910 {
911         spin_lock_irq(&ctx->completion_lock);
912         if (ctx->completed_events) {
913                 struct aio_ring *ring;
914                 unsigned head;
915 
916                 /* Access of ring->head may race with aio_read_events_ring()
917                  * here, but that's okay since whether we read the old version
918                  * or the new version, and either will be valid.  The important
919                  * part is that head cannot pass tail since we prevent
920                  * aio_complete() from updating tail by holding
921                  * ctx->completion_lock.  Even if head is invalid, the check
922                  * against ctx->completed_events below will make sure we do the
923                  * safe/right thing.
924                  */
925                 ring = kmap_atomic(ctx->ring_pages[0]);
926                 head = ring->head;
927                 kunmap_atomic(ring);
928 
929                 refill_reqs_available(ctx, head, ctx->tail);
930         }
931 
932         spin_unlock_irq(&ctx->completion_lock);
933 }
934 
935 /* aio_get_req
936  *      Allocate a slot for an aio request.
937  * Returns NULL if no requests are free.
938  */
939 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
940 {
941         struct kiocb *req;
942 
943         if (!get_reqs_available(ctx)) {
944                 user_refill_reqs_available(ctx);
945                 if (!get_reqs_available(ctx))
946                         return NULL;
947         }
948 
949         req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
950         if (unlikely(!req))
951                 goto out_put;
952 
953         percpu_ref_get(&ctx->reqs);
954 
955         req->ki_ctx = ctx;
956         return req;
957 out_put:
958         put_reqs_available(ctx, 1);
959         return NULL;
960 }
961 
962 static void kiocb_free(struct kiocb *req)
963 {
964         if (req->ki_filp)
965                 fput(req->ki_filp);
966         if (req->ki_eventfd != NULL)
967                 eventfd_ctx_put(req->ki_eventfd);
968         kmem_cache_free(kiocb_cachep, req);
969 }
970 
971 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
972 {
973         struct aio_ring __user *ring  = (void __user *)ctx_id;
974         struct mm_struct *mm = current->mm;
975         struct kioctx *ctx, *ret = NULL;
976         struct kioctx_table *table;
977         unsigned id;
978 
979         if (get_user(id, &ring->id))
980                 return NULL;
981 
982         rcu_read_lock();
983         table = rcu_dereference(mm->ioctx_table);
984 
985         if (!table || id >= table->nr)
986                 goto out;
987 
988         ctx = table->table[id];
989         if (ctx && ctx->user_id == ctx_id) {
990                 percpu_ref_get(&ctx->users);
991                 ret = ctx;
992         }
993 out:
994         rcu_read_unlock();
995         return ret;
996 }
997 
998 /* aio_complete
999  *      Called when the io request on the given iocb is complete.
1000  */
1001 void aio_complete(struct kiocb *iocb, long res, long res2)
1002 {
1003         struct kioctx   *ctx = iocb->ki_ctx;
1004         struct aio_ring *ring;
1005         struct io_event *ev_page, *event;
1006         unsigned tail, pos, head;
1007         unsigned long   flags;
1008 
1009         /*
1010          * Special case handling for sync iocbs:
1011          *  - events go directly into the iocb for fast handling
1012          *  - the sync task with the iocb in its stack holds the single iocb
1013          *    ref, no other paths have a way to get another ref
1014          *  - the sync task helpfully left a reference to itself in the iocb
1015          */
1016         if (is_sync_kiocb(iocb)) {
1017                 iocb->ki_user_data = res;
1018                 smp_wmb();
1019                 iocb->ki_ctx = ERR_PTR(-EXDEV);
1020                 wake_up_process(iocb->ki_obj.tsk);
1021                 return;
1022         }
1023 
1024         if (iocb->ki_list.next) {
1025                 unsigned long flags;
1026 
1027                 spin_lock_irqsave(&ctx->ctx_lock, flags);
1028                 list_del(&iocb->ki_list);
1029                 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1030         }
1031 
1032         /*
1033          * Add a completion event to the ring buffer. Must be done holding
1034          * ctx->completion_lock to prevent other code from messing with the tail
1035          * pointer since we might be called from irq context.
1036          */
1037         spin_lock_irqsave(&ctx->completion_lock, flags);
1038 
1039         tail = ctx->tail;
1040         pos = tail + AIO_EVENTS_OFFSET;
1041 
1042         if (++tail >= ctx->nr_events)
1043                 tail = 0;
1044 
1045         ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1046         event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1047 
1048         event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1049         event->data = iocb->ki_user_data;
1050         event->res = res;
1051         event->res2 = res2;
1052 
1053         kunmap_atomic(ev_page);
1054         flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1055 
1056         pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1057                  ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1058                  res, res2);
1059 
1060         /* after flagging the request as done, we
1061          * must never even look at it again
1062          */
1063         smp_wmb();      /* make event visible before updating tail */
1064 
1065         ctx->tail = tail;
1066 
1067         ring = kmap_atomic(ctx->ring_pages[0]);
1068         head = ring->head;
1069         ring->tail = tail;
1070         kunmap_atomic(ring);
1071         flush_dcache_page(ctx->ring_pages[0]);
1072 
1073         ctx->completed_events++;
1074         if (ctx->completed_events > 1)
1075                 refill_reqs_available(ctx, head, tail);
1076         spin_unlock_irqrestore(&ctx->completion_lock, flags);
1077 
1078         pr_debug("added to ring %p at [%u]\n", iocb, tail);
1079 
1080         /*
1081          * Check if the user asked us to deliver the result through an
1082          * eventfd. The eventfd_signal() function is safe to be called
1083          * from IRQ context.
1084          */
1085         if (iocb->ki_eventfd != NULL)
1086                 eventfd_signal(iocb->ki_eventfd, 1);
1087 
1088         /* everything turned out well, dispose of the aiocb. */
1089         kiocb_free(iocb);
1090 
1091         /*
1092          * We have to order our ring_info tail store above and test
1093          * of the wait list below outside the wait lock.  This is
1094          * like in wake_up_bit() where clearing a bit has to be
1095          * ordered with the unlocked test.
1096          */
1097         smp_mb();
1098 
1099         if (waitqueue_active(&ctx->wait))
1100                 wake_up(&ctx->wait);
1101 
1102         percpu_ref_put(&ctx->reqs);
1103 }
1104 EXPORT_SYMBOL(aio_complete);
1105 
1106 /* aio_read_events_ring
1107  *      Pull an event off of the ioctx's event ring.  Returns the number of
1108  *      events fetched
1109  */
1110 static long aio_read_events_ring(struct kioctx *ctx,
1111                                  struct io_event __user *event, long nr)
1112 {
1113         struct aio_ring *ring;
1114         unsigned head, tail, pos;
1115         long ret = 0;
1116         int copy_ret;
1117 
1118         mutex_lock(&ctx->ring_lock);
1119 
1120         /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1121         ring = kmap_atomic(ctx->ring_pages[0]);
1122         head = ring->head;
1123         tail = ring->tail;
1124         kunmap_atomic(ring);
1125 
1126         /*
1127          * Ensure that once we've read the current tail pointer, that
1128          * we also see the events that were stored up to the tail.
1129          */
1130         smp_rmb();
1131 
1132         pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1133 
1134         if (head == tail)
1135                 goto out;
1136 
1137         head %= ctx->nr_events;
1138         tail %= ctx->nr_events;
1139 
1140         while (ret < nr) {
1141                 long avail;
1142                 struct io_event *ev;
1143                 struct page *page;
1144 
1145                 avail = (head <= tail ?  tail : ctx->nr_events) - head;
1146                 if (head == tail)
1147                         break;
1148 
1149                 avail = min(avail, nr - ret);
1150                 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1151                             ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1152 
1153                 pos = head + AIO_EVENTS_OFFSET;
1154                 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1155                 pos %= AIO_EVENTS_PER_PAGE;
1156 
1157                 ev = kmap(page);
1158                 copy_ret = copy_to_user(event + ret, ev + pos,
1159                                         sizeof(*ev) * avail);
1160                 kunmap(page);
1161 
1162                 if (unlikely(copy_ret)) {
1163                         ret = -EFAULT;
1164                         goto out;
1165                 }
1166 
1167                 ret += avail;
1168                 head += avail;
1169                 head %= ctx->nr_events;
1170         }
1171 
1172         ring = kmap_atomic(ctx->ring_pages[0]);
1173         ring->head = head;
1174         kunmap_atomic(ring);
1175         flush_dcache_page(ctx->ring_pages[0]);
1176 
1177         pr_debug("%li  h%u t%u\n", ret, head, tail);
1178 out:
1179         mutex_unlock(&ctx->ring_lock);
1180 
1181         return ret;
1182 }
1183 
1184 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1185                             struct io_event __user *event, long *i)
1186 {
1187         long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1188 
1189         if (ret > 0)
1190                 *i += ret;
1191 
1192         if (unlikely(atomic_read(&ctx->dead)))
1193                 ret = -EINVAL;
1194 
1195         if (!*i)
1196                 *i = ret;
1197 
1198         return ret < 0 || *i >= min_nr;
1199 }
1200 
1201 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1202                         struct io_event __user *event,
1203                         struct timespec __user *timeout)
1204 {
1205         ktime_t until = { .tv64 = KTIME_MAX };
1206         long ret = 0;
1207 
1208         if (timeout) {
1209                 struct timespec ts;
1210 
1211                 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1212                         return -EFAULT;
1213 
1214                 until = timespec_to_ktime(ts);
1215         }
1216 
1217         /*
1218          * Note that aio_read_events() is being called as the conditional - i.e.
1219          * we're calling it after prepare_to_wait() has set task state to
1220          * TASK_INTERRUPTIBLE.
1221          *
1222          * But aio_read_events() can block, and if it blocks it's going to flip
1223          * the task state back to TASK_RUNNING.
1224          *
1225          * This should be ok, provided it doesn't flip the state back to
1226          * TASK_RUNNING and return 0 too much - that causes us to spin. That
1227          * will only happen if the mutex_lock() call blocks, and we then find
1228          * the ringbuffer empty. So in practice we should be ok, but it's
1229          * something to be aware of when touching this code.
1230          */
1231         wait_event_interruptible_hrtimeout(ctx->wait,
1232                         aio_read_events(ctx, min_nr, nr, event, &ret), until);
1233 
1234         if (!ret && signal_pending(current))
1235                 ret = -EINTR;
1236 
1237         return ret;
1238 }
1239 
1240 /* sys_io_setup:
1241  *      Create an aio_context capable of receiving at least nr_events.
1242  *      ctxp must not point to an aio_context that already exists, and
1243  *      must be initialized to 0 prior to the call.  On successful
1244  *      creation of the aio_context, *ctxp is filled in with the resulting 
1245  *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1246  *      if the specified nr_events exceeds internal limits.  May fail 
1247  *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1248  *      of available events.  May fail with -ENOMEM if insufficient kernel
1249  *      resources are available.  May fail with -EFAULT if an invalid
1250  *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1251  *      implemented.
1252  */
1253 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1254 {
1255         struct kioctx *ioctx = NULL;
1256         unsigned long ctx;
1257         long ret;
1258 
1259         ret = get_user(ctx, ctxp);
1260         if (unlikely(ret))
1261                 goto out;
1262 
1263         ret = -EINVAL;
1264         if (unlikely(ctx || nr_events == 0)) {
1265                 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1266                          ctx, nr_events);
1267                 goto out;
1268         }
1269 
1270         ioctx = ioctx_alloc(nr_events);
1271         ret = PTR_ERR(ioctx);
1272         if (!IS_ERR(ioctx)) {
1273                 ret = put_user(ioctx->user_id, ctxp);
1274                 if (ret)
1275                         kill_ioctx(current->mm, ioctx, NULL);
1276                 percpu_ref_put(&ioctx->users);
1277         }
1278 
1279 out:
1280         return ret;
1281 }
1282 
1283 /* sys_io_destroy:
1284  *      Destroy the aio_context specified.  May cancel any outstanding 
1285  *      AIOs and block on completion.  Will fail with -ENOSYS if not
1286  *      implemented.  May fail with -EINVAL if the context pointed to
1287  *      is invalid.
1288  */
1289 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1290 {
1291         struct kioctx *ioctx = lookup_ioctx(ctx);
1292         if (likely(NULL != ioctx)) {
1293                 struct completion requests_done =
1294                         COMPLETION_INITIALIZER_ONSTACK(requests_done);
1295                 int ret;
1296 
1297                 /* Pass requests_done to kill_ioctx() where it can be set
1298                  * in a thread-safe way. If we try to set it here then we have
1299                  * a race condition if two io_destroy() called simultaneously.
1300                  */
1301                 ret = kill_ioctx(current->mm, ioctx, &requests_done);
1302                 percpu_ref_put(&ioctx->users);
1303 
1304                 /* Wait until all IO for the context are done. Otherwise kernel
1305                  * keep using user-space buffers even if user thinks the context
1306                  * is destroyed.
1307                  */
1308                 if (!ret)
1309                         wait_for_completion(&requests_done);
1310 
1311                 return ret;
1312         }
1313         pr_debug("EINVAL: io_destroy: invalid context id\n");
1314         return -EINVAL;
1315 }
1316 
1317 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1318                             unsigned long, loff_t);
1319 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1320 
1321 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1322                                      int rw, char __user *buf,
1323                                      unsigned long *nr_segs,
1324                                      struct iovec **iovec,
1325                                      bool compat)
1326 {
1327         ssize_t ret;
1328 
1329         *nr_segs = kiocb->ki_nbytes;
1330 
1331 #ifdef CONFIG_COMPAT
1332         if (compat)
1333                 ret = compat_rw_copy_check_uvector(rw,
1334                                 (struct compat_iovec __user *)buf,
1335                                 *nr_segs, UIO_FASTIOV, *iovec, iovec);
1336         else
1337 #endif
1338                 ret = rw_copy_check_uvector(rw,
1339                                 (struct iovec __user *)buf,
1340                                 *nr_segs, UIO_FASTIOV, *iovec, iovec);
1341         if (ret < 0)
1342                 return ret;
1343 
1344         /* ki_nbytes now reflect bytes instead of segs */
1345         kiocb->ki_nbytes = ret;
1346         return 0;
1347 }
1348 
1349 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1350                                        int rw, char __user *buf,
1351                                        unsigned long *nr_segs,
1352                                        struct iovec *iovec)
1353 {
1354         if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1355                 return -EFAULT;
1356 
1357         iovec->iov_base = buf;
1358         iovec->iov_len = kiocb->ki_nbytes;
1359         *nr_segs = 1;
1360         return 0;
1361 }
1362 
1363 /*
1364  * aio_run_iocb:
1365  *      Performs the initial checks and io submission.
1366  */
1367 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1368                             char __user *buf, bool compat)
1369 {
1370         struct file *file = req->ki_filp;
1371         ssize_t ret;
1372         unsigned long nr_segs;
1373         int rw;
1374         fmode_t mode;
1375         aio_rw_op *rw_op;
1376         rw_iter_op *iter_op;
1377         struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1378         struct iov_iter iter;
1379 
1380         switch (opcode) {
1381         case IOCB_CMD_PREAD:
1382         case IOCB_CMD_PREADV:
1383                 mode    = FMODE_READ;
1384                 rw      = READ;
1385                 rw_op   = file->f_op->aio_read;
1386                 iter_op = file->f_op->read_iter;
1387                 goto rw_common;
1388 
1389         case IOCB_CMD_PWRITE:
1390         case IOCB_CMD_PWRITEV:
1391                 mode    = FMODE_WRITE;
1392                 rw      = WRITE;
1393                 rw_op   = file->f_op->aio_write;
1394                 iter_op = file->f_op->write_iter;
1395                 goto rw_common;
1396 rw_common:
1397                 if (unlikely(!(file->f_mode & mode)))
1398                         return -EBADF;
1399 
1400                 if (!rw_op && !iter_op)
1401                         return -EINVAL;
1402 
1403                 ret = (opcode == IOCB_CMD_PREADV ||
1404                        opcode == IOCB_CMD_PWRITEV)
1405                         ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1406                                                 &iovec, compat)
1407                         : aio_setup_single_vector(req, rw, buf, &nr_segs,
1408                                                   iovec);
1409                 if (!ret)
1410                         ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1411                 if (ret < 0) {
1412                         if (iovec != inline_vecs)
1413                                 kfree(iovec);
1414                         return ret;
1415                 }
1416 
1417                 req->ki_nbytes = ret;
1418 
1419                 /* XXX: move/kill - rw_verify_area()? */
1420                 /* This matches the pread()/pwrite() logic */
1421                 if (req->ki_pos < 0) {
1422                         ret = -EINVAL;
1423                         break;
1424                 }
1425 
1426                 if (rw == WRITE)
1427                         file_start_write(file);
1428 
1429                 if (iter_op) {
1430                         iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes);
1431                         ret = iter_op(req, &iter);
1432                 } else {
1433                         ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1434                 }
1435 
1436                 if (rw == WRITE)
1437                         file_end_write(file);
1438                 break;
1439 
1440         case IOCB_CMD_FDSYNC:
1441                 if (!file->f_op->aio_fsync)
1442                         return -EINVAL;
1443 
1444                 ret = file->f_op->aio_fsync(req, 1);
1445                 break;
1446 
1447         case IOCB_CMD_FSYNC:
1448                 if (!file->f_op->aio_fsync)
1449                         return -EINVAL;
1450 
1451                 ret = file->f_op->aio_fsync(req, 0);
1452                 break;
1453 
1454         default:
1455                 pr_debug("EINVAL: no operation provided\n");
1456                 return -EINVAL;
1457         }
1458 
1459         if (iovec != inline_vecs)
1460                 kfree(iovec);
1461 
1462         if (ret != -EIOCBQUEUED) {
1463                 /*
1464                  * There's no easy way to restart the syscall since other AIO's
1465                  * may be already running. Just fail this IO with EINTR.
1466                  */
1467                 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1468                              ret == -ERESTARTNOHAND ||
1469                              ret == -ERESTART_RESTARTBLOCK))
1470                         ret = -EINTR;
1471                 aio_complete(req, ret, 0);
1472         }
1473 
1474         return 0;
1475 }
1476 
1477 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1478                          struct iocb *iocb, bool compat)
1479 {
1480         struct kiocb *req;
1481         ssize_t ret;
1482 
1483         /* enforce forwards compatibility on users */
1484         if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1485                 pr_debug("EINVAL: reserve field set\n");
1486                 return -EINVAL;
1487         }
1488 
1489         /* prevent overflows */
1490         if (unlikely(
1491             (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1492             (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1493             ((ssize_t)iocb->aio_nbytes < 0)
1494            )) {
1495                 pr_debug("EINVAL: io_submit: overflow check\n");
1496                 return -EINVAL;
1497         }
1498 
1499         req = aio_get_req(ctx);
1500         if (unlikely(!req))
1501                 return -EAGAIN;
1502 
1503         req->ki_filp = fget(iocb->aio_fildes);
1504         if (unlikely(!req->ki_filp)) {
1505                 ret = -EBADF;
1506                 goto out_put_req;
1507         }
1508 
1509         if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1510                 /*
1511                  * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1512                  * instance of the file* now. The file descriptor must be
1513                  * an eventfd() fd, and will be signaled for each completed
1514                  * event using the eventfd_signal() function.
1515                  */
1516                 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1517                 if (IS_ERR(req->ki_eventfd)) {
1518                         ret = PTR_ERR(req->ki_eventfd);
1519                         req->ki_eventfd = NULL;
1520                         goto out_put_req;
1521                 }
1522         }
1523 
1524         ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1525         if (unlikely(ret)) {
1526                 pr_debug("EFAULT: aio_key\n");
1527                 goto out_put_req;
1528         }
1529 
1530         req->ki_obj.user = user_iocb;
1531         req->ki_user_data = iocb->aio_data;
1532         req->ki_pos = iocb->aio_offset;
1533         req->ki_nbytes = iocb->aio_nbytes;
1534 
1535         ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1536                            (char __user *)(unsigned long)iocb->aio_buf,
1537                            compat);
1538         if (ret)
1539                 goto out_put_req;
1540 
1541         return 0;
1542 out_put_req:
1543         put_reqs_available(ctx, 1);
1544         percpu_ref_put(&ctx->reqs);
1545         kiocb_free(req);
1546         return ret;
1547 }
1548 
1549 long do_io_submit(aio_context_t ctx_id, long nr,
1550                   struct iocb __user *__user *iocbpp, bool compat)
1551 {
1552         struct kioctx *ctx;
1553         long ret = 0;
1554         int i = 0;
1555         struct blk_plug plug;
1556 
1557         if (unlikely(nr < 0))
1558                 return -EINVAL;
1559 
1560         if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1561                 nr = LONG_MAX/sizeof(*iocbpp);
1562 
1563         if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1564                 return -EFAULT;
1565 
1566         ctx = lookup_ioctx(ctx_id);
1567         if (unlikely(!ctx)) {
1568                 pr_debug("EINVAL: invalid context id\n");
1569                 return -EINVAL;
1570         }
1571 
1572         blk_start_plug(&plug);
1573 
1574         /*
1575          * AKPM: should this return a partial result if some of the IOs were
1576          * successfully submitted?
1577          */
1578         for (i=0; i<nr; i++) {
1579                 struct iocb __user *user_iocb;
1580                 struct iocb tmp;
1581 
1582                 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1583                         ret = -EFAULT;
1584                         break;
1585                 }
1586 
1587                 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1588                         ret = -EFAULT;
1589                         break;
1590                 }
1591 
1592                 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1593                 if (ret)
1594                         break;
1595         }
1596         blk_finish_plug(&plug);
1597 
1598         percpu_ref_put(&ctx->users);
1599         return i ? i : ret;
1600 }
1601 
1602 /* sys_io_submit:
1603  *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1604  *      the number of iocbs queued.  May return -EINVAL if the aio_context
1605  *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1606  *      *iocbpp[0] is not properly initialized, if the operation specified
1607  *      is invalid for the file descriptor in the iocb.  May fail with
1608  *      -EFAULT if any of the data structures point to invalid data.  May
1609  *      fail with -EBADF if the file descriptor specified in the first
1610  *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1611  *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1612  *      fail with -ENOSYS if not implemented.
1613  */
1614 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1615                 struct iocb __user * __user *, iocbpp)
1616 {
1617         return do_io_submit(ctx_id, nr, iocbpp, 0);
1618 }
1619 
1620 /* lookup_kiocb
1621  *      Finds a given iocb for cancellation.
1622  */
1623 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1624                                   u32 key)
1625 {
1626         struct list_head *pos;
1627 
1628         assert_spin_locked(&ctx->ctx_lock);
1629 
1630         if (key != KIOCB_KEY)
1631                 return NULL;
1632 
1633         /* TODO: use a hash or array, this sucks. */
1634         list_for_each(pos, &ctx->active_reqs) {
1635                 struct kiocb *kiocb = list_kiocb(pos);
1636                 if (kiocb->ki_obj.user == iocb)
1637                         return kiocb;
1638         }
1639         return NULL;
1640 }
1641 
1642 /* sys_io_cancel:
1643  *      Attempts to cancel an iocb previously passed to io_submit.  If
1644  *      the operation is successfully cancelled, the resulting event is
1645  *      copied into the memory pointed to by result without being placed
1646  *      into the completion queue and 0 is returned.  May fail with
1647  *      -EFAULT if any of the data structures pointed to are invalid.
1648  *      May fail with -EINVAL if aio_context specified by ctx_id is
1649  *      invalid.  May fail with -EAGAIN if the iocb specified was not
1650  *      cancelled.  Will fail with -ENOSYS if not implemented.
1651  */
1652 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1653                 struct io_event __user *, result)
1654 {
1655         struct kioctx *ctx;
1656         struct kiocb *kiocb;
1657         u32 key;
1658         int ret;
1659 
1660         ret = get_user(key, &iocb->aio_key);
1661         if (unlikely(ret))
1662                 return -EFAULT;
1663 
1664         ctx = lookup_ioctx(ctx_id);
1665         if (unlikely(!ctx))
1666                 return -EINVAL;
1667 
1668         spin_lock_irq(&ctx->ctx_lock);
1669 
1670         kiocb = lookup_kiocb(ctx, iocb, key);
1671         if (kiocb)
1672                 ret = kiocb_cancel(kiocb);
1673         else
1674                 ret = -EINVAL;
1675 
1676         spin_unlock_irq(&ctx->ctx_lock);
1677 
1678         if (!ret) {
1679                 /*
1680                  * The result argument is no longer used - the io_event is
1681                  * always delivered via the ring buffer. -EINPROGRESS indicates
1682                  * cancellation is progress:
1683                  */
1684                 ret = -EINPROGRESS;
1685         }
1686 
1687         percpu_ref_put(&ctx->users);
1688 
1689         return ret;
1690 }
1691 
1692 /* io_getevents:
1693  *      Attempts to read at least min_nr events and up to nr events from
1694  *      the completion queue for the aio_context specified by ctx_id. If
1695  *      it succeeds, the number of read events is returned. May fail with
1696  *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1697  *      out of range, if timeout is out of range.  May fail with -EFAULT
1698  *      if any of the memory specified is invalid.  May return 0 or
1699  *      < min_nr if the timeout specified by timeout has elapsed
1700  *      before sufficient events are available, where timeout == NULL
1701  *      specifies an infinite timeout. Note that the timeout pointed to by
1702  *      timeout is relative.  Will fail with -ENOSYS if not implemented.
1703  */
1704 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1705                 long, min_nr,
1706                 long, nr,
1707                 struct io_event __user *, events,
1708                 struct timespec __user *, timeout)
1709 {
1710         struct kioctx *ioctx = lookup_ioctx(ctx_id);
1711         long ret = -EINVAL;
1712 
1713         if (likely(ioctx)) {
1714                 if (likely(min_nr <= nr && min_nr >= 0))
1715                         ret = read_events(ioctx, min_nr, nr, events, timeout);
1716                 percpu_ref_put(&ioctx->users);
1717         }
1718         return ret;
1719 }
1720 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us