Version:  2.0.40 2.2.26 2.4.37 3.0 3.1 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

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

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