Version:  2.0.40 2.2.26 2.4.37 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8

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 ctx_rq_wait {
 81         struct completion comp;
 82         atomic_t count;
 83 };
 84 
 85 struct kioctx {
 86         struct percpu_ref       users;
 87         atomic_t                dead;
 88 
 89         struct percpu_ref       reqs;
 90 
 91         unsigned long           user_id;
 92 
 93         struct __percpu kioctx_cpu *cpu;
 94 
 95         /*
 96          * For percpu reqs_available, number of slots we move to/from global
 97          * counter at a time:
 98          */
 99         unsigned                req_batch;
100         /*
101          * This is what userspace passed to io_setup(), it's not used for
102          * anything but counting against the global max_reqs quota.
103          *
104          * The real limit is nr_events - 1, which will be larger (see
105          * aio_setup_ring())
106          */
107         unsigned                max_reqs;
108 
109         /* Size of ringbuffer, in units of struct io_event */
110         unsigned                nr_events;
111 
112         unsigned long           mmap_base;
113         unsigned long           mmap_size;
114 
115         struct page             **ring_pages;
116         long                    nr_pages;
117 
118         struct work_struct      free_work;
119 
120         /*
121          * signals when all in-flight requests are done
122          */
123         struct ctx_rq_wait      *rq_wait;
124 
125         struct {
126                 /*
127                  * This counts the number of available slots in the ringbuffer,
128                  * so we avoid overflowing it: it's decremented (if positive)
129                  * when allocating a kiocb and incremented when the resulting
130                  * io_event is pulled off the ringbuffer.
131                  *
132                  * We batch accesses to it with a percpu version.
133                  */
134                 atomic_t        reqs_available;
135         } ____cacheline_aligned_in_smp;
136 
137         struct {
138                 spinlock_t      ctx_lock;
139                 struct list_head active_reqs;   /* used for cancellation */
140         } ____cacheline_aligned_in_smp;
141 
142         struct {
143                 struct mutex    ring_lock;
144                 wait_queue_head_t wait;
145         } ____cacheline_aligned_in_smp;
146 
147         struct {
148                 unsigned        tail;
149                 unsigned        completed_events;
150                 spinlock_t      completion_lock;
151         } ____cacheline_aligned_in_smp;
152 
153         struct page             *internal_pages[AIO_RING_PAGES];
154         struct file             *aio_ring_file;
155 
156         unsigned                id;
157 };
158 
159 /*
160  * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
161  * cancelled or completed (this makes a certain amount of sense because
162  * successful cancellation - io_cancel() - does deliver the completion to
163  * userspace).
164  *
165  * And since most things don't implement kiocb cancellation and we'd really like
166  * kiocb completion to be lockless when possible, we use ki_cancel to
167  * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
168  * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
169  */
170 #define KIOCB_CANCELLED         ((void *) (~0ULL))
171 
172 struct aio_kiocb {
173         struct kiocb            common;
174 
175         struct kioctx           *ki_ctx;
176         kiocb_cancel_fn         *ki_cancel;
177 
178         struct iocb __user      *ki_user_iocb;  /* user's aiocb */
179         __u64                   ki_user_data;   /* user's data for completion */
180 
181         struct list_head        ki_list;        /* the aio core uses this
182                                                  * for cancellation */
183 
184         /*
185          * If the aio_resfd field of the userspace iocb is not zero,
186          * this is the underlying eventfd context to deliver events to.
187          */
188         struct eventfd_ctx      *ki_eventfd;
189 };
190 
191 /*------ sysctl variables----*/
192 static DEFINE_SPINLOCK(aio_nr_lock);
193 unsigned long aio_nr;           /* current system wide number of aio requests */
194 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
195 /*----end sysctl variables---*/
196 
197 static struct kmem_cache        *kiocb_cachep;
198 static struct kmem_cache        *kioctx_cachep;
199 
200 static struct vfsmount *aio_mnt;
201 
202 static const struct file_operations aio_ring_fops;
203 static const struct address_space_operations aio_ctx_aops;
204 
205 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
206 {
207         struct qstr this = QSTR_INIT("[aio]", 5);
208         struct file *file;
209         struct path path;
210         struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
211         if (IS_ERR(inode))
212                 return ERR_CAST(inode);
213 
214         inode->i_mapping->a_ops = &aio_ctx_aops;
215         inode->i_mapping->private_data = ctx;
216         inode->i_size = PAGE_SIZE * nr_pages;
217 
218         path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
219         if (!path.dentry) {
220                 iput(inode);
221                 return ERR_PTR(-ENOMEM);
222         }
223         path.mnt = mntget(aio_mnt);
224 
225         d_instantiate(path.dentry, inode);
226         file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
227         if (IS_ERR(file)) {
228                 path_put(&path);
229                 return file;
230         }
231 
232         file->f_flags = O_RDWR;
233         return file;
234 }
235 
236 static struct dentry *aio_mount(struct file_system_type *fs_type,
237                                 int flags, const char *dev_name, void *data)
238 {
239         static const struct dentry_operations ops = {
240                 .d_dname        = simple_dname,
241         };
242         struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops,
243                                            AIO_RING_MAGIC);
244 
245         if (!IS_ERR(root))
246                 root->d_sb->s_iflags |= SB_I_NOEXEC;
247         return root;
248 }
249 
250 /* aio_setup
251  *      Creates the slab caches used by the aio routines, panic on
252  *      failure as this is done early during the boot sequence.
253  */
254 static int __init aio_setup(void)
255 {
256         static struct file_system_type aio_fs = {
257                 .name           = "aio",
258                 .mount          = aio_mount,
259                 .kill_sb        = kill_anon_super,
260         };
261         aio_mnt = kern_mount(&aio_fs);
262         if (IS_ERR(aio_mnt))
263                 panic("Failed to create aio fs mount.");
264 
265         kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
266         kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
267 
268         pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
269 
270         return 0;
271 }
272 __initcall(aio_setup);
273 
274 static void put_aio_ring_file(struct kioctx *ctx)
275 {
276         struct file *aio_ring_file = ctx->aio_ring_file;
277         if (aio_ring_file) {
278                 truncate_setsize(aio_ring_file->f_inode, 0);
279 
280                 /* Prevent further access to the kioctx from migratepages */
281                 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
282                 aio_ring_file->f_inode->i_mapping->private_data = NULL;
283                 ctx->aio_ring_file = NULL;
284                 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
285 
286                 fput(aio_ring_file);
287         }
288 }
289 
290 static void aio_free_ring(struct kioctx *ctx)
291 {
292         int i;
293 
294         /* Disconnect the kiotx from the ring file.  This prevents future
295          * accesses to the kioctx from page migration.
296          */
297         put_aio_ring_file(ctx);
298 
299         for (i = 0; i < ctx->nr_pages; i++) {
300                 struct page *page;
301                 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
302                                 page_count(ctx->ring_pages[i]));
303                 page = ctx->ring_pages[i];
304                 if (!page)
305                         continue;
306                 ctx->ring_pages[i] = NULL;
307                 put_page(page);
308         }
309 
310         if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
311                 kfree(ctx->ring_pages);
312                 ctx->ring_pages = NULL;
313         }
314 }
315 
316 static int aio_ring_mremap(struct vm_area_struct *vma)
317 {
318         struct file *file = vma->vm_file;
319         struct mm_struct *mm = vma->vm_mm;
320         struct kioctx_table *table;
321         int i, res = -EINVAL;
322 
323         spin_lock(&mm->ioctx_lock);
324         rcu_read_lock();
325         table = rcu_dereference(mm->ioctx_table);
326         for (i = 0; i < table->nr; i++) {
327                 struct kioctx *ctx;
328 
329                 ctx = table->table[i];
330                 if (ctx && ctx->aio_ring_file == file) {
331                         if (!atomic_read(&ctx->dead)) {
332                                 ctx->user_id = ctx->mmap_base = vma->vm_start;
333                                 res = 0;
334                         }
335                         break;
336                 }
337         }
338 
339         rcu_read_unlock();
340         spin_unlock(&mm->ioctx_lock);
341         return res;
342 }
343 
344 static const struct vm_operations_struct aio_ring_vm_ops = {
345         .mremap         = aio_ring_mremap,
346 #if IS_ENABLED(CONFIG_MMU)
347         .fault          = filemap_fault,
348         .map_pages      = filemap_map_pages,
349         .page_mkwrite   = filemap_page_mkwrite,
350 #endif
351 };
352 
353 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
354 {
355         vma->vm_flags |= VM_DONTEXPAND;
356         vma->vm_ops = &aio_ring_vm_ops;
357         return 0;
358 }
359 
360 static const struct file_operations aio_ring_fops = {
361         .mmap = aio_ring_mmap,
362 };
363 
364 #if IS_ENABLED(CONFIG_MIGRATION)
365 static int aio_migratepage(struct address_space *mapping, struct page *new,
366                         struct page *old, enum migrate_mode mode)
367 {
368         struct kioctx *ctx;
369         unsigned long flags;
370         pgoff_t idx;
371         int rc;
372 
373         rc = 0;
374 
375         /* mapping->private_lock here protects against the kioctx teardown.  */
376         spin_lock(&mapping->private_lock);
377         ctx = mapping->private_data;
378         if (!ctx) {
379                 rc = -EINVAL;
380                 goto out;
381         }
382 
383         /* The ring_lock mutex.  The prevents aio_read_events() from writing
384          * to the ring's head, and prevents page migration from mucking in
385          * a partially initialized kiotx.
386          */
387         if (!mutex_trylock(&ctx->ring_lock)) {
388                 rc = -EAGAIN;
389                 goto out;
390         }
391 
392         idx = old->index;
393         if (idx < (pgoff_t)ctx->nr_pages) {
394                 /* Make sure the old page hasn't already been changed */
395                 if (ctx->ring_pages[idx] != old)
396                         rc = -EAGAIN;
397         } else
398                 rc = -EINVAL;
399 
400         if (rc != 0)
401                 goto out_unlock;
402 
403         /* Writeback must be complete */
404         BUG_ON(PageWriteback(old));
405         get_page(new);
406 
407         rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
408         if (rc != MIGRATEPAGE_SUCCESS) {
409                 put_page(new);
410                 goto out_unlock;
411         }
412 
413         /* Take completion_lock to prevent other writes to the ring buffer
414          * while the old page is copied to the new.  This prevents new
415          * events from being lost.
416          */
417         spin_lock_irqsave(&ctx->completion_lock, flags);
418         migrate_page_copy(new, old);
419         BUG_ON(ctx->ring_pages[idx] != old);
420         ctx->ring_pages[idx] = new;
421         spin_unlock_irqrestore(&ctx->completion_lock, flags);
422 
423         /* The old page is no longer accessible. */
424         put_page(old);
425 
426 out_unlock:
427         mutex_unlock(&ctx->ring_lock);
428 out:
429         spin_unlock(&mapping->private_lock);
430         return rc;
431 }
432 #endif
433 
434 static const struct address_space_operations aio_ctx_aops = {
435         .set_page_dirty = __set_page_dirty_no_writeback,
436 #if IS_ENABLED(CONFIG_MIGRATION)
437         .migratepage    = aio_migratepage,
438 #endif
439 };
440 
441 static int aio_setup_ring(struct kioctx *ctx)
442 {
443         struct aio_ring *ring;
444         unsigned nr_events = ctx->max_reqs;
445         struct mm_struct *mm = current->mm;
446         unsigned long size, unused;
447         int nr_pages;
448         int i;
449         struct file *file;
450 
451         /* Compensate for the ring buffer's head/tail overlap entry */
452         nr_events += 2; /* 1 is required, 2 for good luck */
453 
454         size = sizeof(struct aio_ring);
455         size += sizeof(struct io_event) * nr_events;
456 
457         nr_pages = PFN_UP(size);
458         if (nr_pages < 0)
459                 return -EINVAL;
460 
461         file = aio_private_file(ctx, nr_pages);
462         if (IS_ERR(file)) {
463                 ctx->aio_ring_file = NULL;
464                 return -ENOMEM;
465         }
466 
467         ctx->aio_ring_file = file;
468         nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
469                         / sizeof(struct io_event);
470 
471         ctx->ring_pages = ctx->internal_pages;
472         if (nr_pages > AIO_RING_PAGES) {
473                 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
474                                           GFP_KERNEL);
475                 if (!ctx->ring_pages) {
476                         put_aio_ring_file(ctx);
477                         return -ENOMEM;
478                 }
479         }
480 
481         for (i = 0; i < nr_pages; i++) {
482                 struct page *page;
483                 page = find_or_create_page(file->f_inode->i_mapping,
484                                            i, GFP_HIGHUSER | __GFP_ZERO);
485                 if (!page)
486                         break;
487                 pr_debug("pid(%d) page[%d]->count=%d\n",
488                          current->pid, i, page_count(page));
489                 SetPageUptodate(page);
490                 unlock_page(page);
491 
492                 ctx->ring_pages[i] = page;
493         }
494         ctx->nr_pages = i;
495 
496         if (unlikely(i != nr_pages)) {
497                 aio_free_ring(ctx);
498                 return -ENOMEM;
499         }
500 
501         ctx->mmap_size = nr_pages * PAGE_SIZE;
502         pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
503 
504         if (down_write_killable(&mm->mmap_sem)) {
505                 ctx->mmap_size = 0;
506                 aio_free_ring(ctx);
507                 return -EINTR;
508         }
509 
510         ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
511                                        PROT_READ | PROT_WRITE,
512                                        MAP_SHARED, 0, &unused);
513         up_write(&mm->mmap_sem);
514         if (IS_ERR((void *)ctx->mmap_base)) {
515                 ctx->mmap_size = 0;
516                 aio_free_ring(ctx);
517                 return -ENOMEM;
518         }
519 
520         pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
521 
522         ctx->user_id = ctx->mmap_base;
523         ctx->nr_events = nr_events; /* trusted copy */
524 
525         ring = kmap_atomic(ctx->ring_pages[0]);
526         ring->nr = nr_events;   /* user copy */
527         ring->id = ~0U;
528         ring->head = ring->tail = 0;
529         ring->magic = AIO_RING_MAGIC;
530         ring->compat_features = AIO_RING_COMPAT_FEATURES;
531         ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
532         ring->header_length = sizeof(struct aio_ring);
533         kunmap_atomic(ring);
534         flush_dcache_page(ctx->ring_pages[0]);
535 
536         return 0;
537 }
538 
539 #define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
540 #define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
541 #define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
542 
543 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
544 {
545         struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
546         struct kioctx *ctx = req->ki_ctx;
547         unsigned long flags;
548 
549         spin_lock_irqsave(&ctx->ctx_lock, flags);
550 
551         if (!req->ki_list.next)
552                 list_add(&req->ki_list, &ctx->active_reqs);
553 
554         req->ki_cancel = cancel;
555 
556         spin_unlock_irqrestore(&ctx->ctx_lock, flags);
557 }
558 EXPORT_SYMBOL(kiocb_set_cancel_fn);
559 
560 static int kiocb_cancel(struct aio_kiocb *kiocb)
561 {
562         kiocb_cancel_fn *old, *cancel;
563 
564         /*
565          * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
566          * actually has a cancel function, hence the cmpxchg()
567          */
568 
569         cancel = ACCESS_ONCE(kiocb->ki_cancel);
570         do {
571                 if (!cancel || cancel == KIOCB_CANCELLED)
572                         return -EINVAL;
573 
574                 old = cancel;
575                 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
576         } while (cancel != old);
577 
578         return cancel(&kiocb->common);
579 }
580 
581 static void free_ioctx(struct work_struct *work)
582 {
583         struct kioctx *ctx = container_of(work, struct kioctx, free_work);
584 
585         pr_debug("freeing %p\n", ctx);
586 
587         aio_free_ring(ctx);
588         free_percpu(ctx->cpu);
589         percpu_ref_exit(&ctx->reqs);
590         percpu_ref_exit(&ctx->users);
591         kmem_cache_free(kioctx_cachep, ctx);
592 }
593 
594 static void free_ioctx_reqs(struct percpu_ref *ref)
595 {
596         struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
597 
598         /* At this point we know that there are no any in-flight requests */
599         if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
600                 complete(&ctx->rq_wait->comp);
601 
602         INIT_WORK(&ctx->free_work, free_ioctx);
603         schedule_work(&ctx->free_work);
604 }
605 
606 /*
607  * When this function runs, the kioctx has been removed from the "hash table"
608  * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
609  * now it's safe to cancel any that need to be.
610  */
611 static void free_ioctx_users(struct percpu_ref *ref)
612 {
613         struct kioctx *ctx = container_of(ref, struct kioctx, users);
614         struct aio_kiocb *req;
615 
616         spin_lock_irq(&ctx->ctx_lock);
617 
618         while (!list_empty(&ctx->active_reqs)) {
619                 req = list_first_entry(&ctx->active_reqs,
620                                        struct aio_kiocb, ki_list);
621 
622                 list_del_init(&req->ki_list);
623                 kiocb_cancel(req);
624         }
625 
626         spin_unlock_irq(&ctx->ctx_lock);
627 
628         percpu_ref_kill(&ctx->reqs);
629         percpu_ref_put(&ctx->reqs);
630 }
631 
632 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
633 {
634         unsigned i, new_nr;
635         struct kioctx_table *table, *old;
636         struct aio_ring *ring;
637 
638         spin_lock(&mm->ioctx_lock);
639         table = rcu_dereference_raw(mm->ioctx_table);
640 
641         while (1) {
642                 if (table)
643                         for (i = 0; i < table->nr; i++)
644                                 if (!table->table[i]) {
645                                         ctx->id = i;
646                                         table->table[i] = ctx;
647                                         spin_unlock(&mm->ioctx_lock);
648 
649                                         /* While kioctx setup is in progress,
650                                          * we are protected from page migration
651                                          * changes ring_pages by ->ring_lock.
652                                          */
653                                         ring = kmap_atomic(ctx->ring_pages[0]);
654                                         ring->id = ctx->id;
655                                         kunmap_atomic(ring);
656                                         return 0;
657                                 }
658 
659                 new_nr = (table ? table->nr : 1) * 4;
660                 spin_unlock(&mm->ioctx_lock);
661 
662                 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
663                                 new_nr, GFP_KERNEL);
664                 if (!table)
665                         return -ENOMEM;
666 
667                 table->nr = new_nr;
668 
669                 spin_lock(&mm->ioctx_lock);
670                 old = rcu_dereference_raw(mm->ioctx_table);
671 
672                 if (!old) {
673                         rcu_assign_pointer(mm->ioctx_table, table);
674                 } else if (table->nr > old->nr) {
675                         memcpy(table->table, old->table,
676                                old->nr * sizeof(struct kioctx *));
677 
678                         rcu_assign_pointer(mm->ioctx_table, table);
679                         kfree_rcu(old, rcu);
680                 } else {
681                         kfree(table);
682                         table = old;
683                 }
684         }
685 }
686 
687 static void aio_nr_sub(unsigned nr)
688 {
689         spin_lock(&aio_nr_lock);
690         if (WARN_ON(aio_nr - nr > aio_nr))
691                 aio_nr = 0;
692         else
693                 aio_nr -= nr;
694         spin_unlock(&aio_nr_lock);
695 }
696 
697 /* ioctx_alloc
698  *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
699  */
700 static struct kioctx *ioctx_alloc(unsigned nr_events)
701 {
702         struct mm_struct *mm = current->mm;
703         struct kioctx *ctx;
704         int err = -ENOMEM;
705 
706         /*
707          * We keep track of the number of available ringbuffer slots, to prevent
708          * overflow (reqs_available), and we also use percpu counters for this.
709          *
710          * So since up to half the slots might be on other cpu's percpu counters
711          * and unavailable, double nr_events so userspace sees what they
712          * expected: additionally, we move req_batch slots to/from percpu
713          * counters at a time, so make sure that isn't 0:
714          */
715         nr_events = max(nr_events, num_possible_cpus() * 4);
716         nr_events *= 2;
717 
718         /* Prevent overflows */
719         if (nr_events > (0x10000000U / sizeof(struct io_event))) {
720                 pr_debug("ENOMEM: nr_events too high\n");
721                 return ERR_PTR(-EINVAL);
722         }
723 
724         if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
725                 return ERR_PTR(-EAGAIN);
726 
727         ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
728         if (!ctx)
729                 return ERR_PTR(-ENOMEM);
730 
731         ctx->max_reqs = nr_events;
732 
733         spin_lock_init(&ctx->ctx_lock);
734         spin_lock_init(&ctx->completion_lock);
735         mutex_init(&ctx->ring_lock);
736         /* Protect against page migration throughout kiotx setup by keeping
737          * the ring_lock mutex held until setup is complete. */
738         mutex_lock(&ctx->ring_lock);
739         init_waitqueue_head(&ctx->wait);
740 
741         INIT_LIST_HEAD(&ctx->active_reqs);
742 
743         if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
744                 goto err;
745 
746         if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
747                 goto err;
748 
749         ctx->cpu = alloc_percpu(struct kioctx_cpu);
750         if (!ctx->cpu)
751                 goto err;
752 
753         err = aio_setup_ring(ctx);
754         if (err < 0)
755                 goto err;
756 
757         atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
758         ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
759         if (ctx->req_batch < 1)
760                 ctx->req_batch = 1;
761 
762         /* limit the number of system wide aios */
763         spin_lock(&aio_nr_lock);
764         if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
765             aio_nr + nr_events < aio_nr) {
766                 spin_unlock(&aio_nr_lock);
767                 err = -EAGAIN;
768                 goto err_ctx;
769         }
770         aio_nr += ctx->max_reqs;
771         spin_unlock(&aio_nr_lock);
772 
773         percpu_ref_get(&ctx->users);    /* io_setup() will drop this ref */
774         percpu_ref_get(&ctx->reqs);     /* free_ioctx_users() will drop this */
775 
776         err = ioctx_add_table(ctx, mm);
777         if (err)
778                 goto err_cleanup;
779 
780         /* Release the ring_lock mutex now that all setup is complete. */
781         mutex_unlock(&ctx->ring_lock);
782 
783         pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
784                  ctx, ctx->user_id, mm, ctx->nr_events);
785         return ctx;
786 
787 err_cleanup:
788         aio_nr_sub(ctx->max_reqs);
789 err_ctx:
790         atomic_set(&ctx->dead, 1);
791         if (ctx->mmap_size)
792                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
793         aio_free_ring(ctx);
794 err:
795         mutex_unlock(&ctx->ring_lock);
796         free_percpu(ctx->cpu);
797         percpu_ref_exit(&ctx->reqs);
798         percpu_ref_exit(&ctx->users);
799         kmem_cache_free(kioctx_cachep, ctx);
800         pr_debug("error allocating ioctx %d\n", err);
801         return ERR_PTR(err);
802 }
803 
804 /* kill_ioctx
805  *      Cancels all outstanding aio requests on an aio context.  Used
806  *      when the processes owning a context have all exited to encourage
807  *      the rapid destruction of the kioctx.
808  */
809 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
810                       struct ctx_rq_wait *wait)
811 {
812         struct kioctx_table *table;
813 
814         spin_lock(&mm->ioctx_lock);
815         if (atomic_xchg(&ctx->dead, 1)) {
816                 spin_unlock(&mm->ioctx_lock);
817                 return -EINVAL;
818         }
819 
820         table = rcu_dereference_raw(mm->ioctx_table);
821         WARN_ON(ctx != table->table[ctx->id]);
822         table->table[ctx->id] = NULL;
823         spin_unlock(&mm->ioctx_lock);
824 
825         /* percpu_ref_kill() will do the necessary call_rcu() */
826         wake_up_all(&ctx->wait);
827 
828         /*
829          * It'd be more correct to do this in free_ioctx(), after all
830          * the outstanding kiocbs have finished - but by then io_destroy
831          * has already returned, so io_setup() could potentially return
832          * -EAGAIN with no ioctxs actually in use (as far as userspace
833          *  could tell).
834          */
835         aio_nr_sub(ctx->max_reqs);
836 
837         if (ctx->mmap_size)
838                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
839 
840         ctx->rq_wait = wait;
841         percpu_ref_kill(&ctx->users);
842         return 0;
843 }
844 
845 /*
846  * exit_aio: called when the last user of mm goes away.  At this point, there is
847  * no way for any new requests to be submited or any of the io_* syscalls to be
848  * called on the context.
849  *
850  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
851  * them.
852  */
853 void exit_aio(struct mm_struct *mm)
854 {
855         struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
856         struct ctx_rq_wait wait;
857         int i, skipped;
858 
859         if (!table)
860                 return;
861 
862         atomic_set(&wait.count, table->nr);
863         init_completion(&wait.comp);
864 
865         skipped = 0;
866         for (i = 0; i < table->nr; ++i) {
867                 struct kioctx *ctx = table->table[i];
868 
869                 if (!ctx) {
870                         skipped++;
871                         continue;
872                 }
873 
874                 /*
875                  * We don't need to bother with munmap() here - exit_mmap(mm)
876                  * is coming and it'll unmap everything. And we simply can't,
877                  * this is not necessarily our ->mm.
878                  * Since kill_ioctx() uses non-zero ->mmap_size as indicator
879                  * that it needs to unmap the area, just set it to 0.
880                  */
881                 ctx->mmap_size = 0;
882                 kill_ioctx(mm, ctx, &wait);
883         }
884 
885         if (!atomic_sub_and_test(skipped, &wait.count)) {
886                 /* Wait until all IO for the context are done. */
887                 wait_for_completion(&wait.comp);
888         }
889 
890         RCU_INIT_POINTER(mm->ioctx_table, NULL);
891         kfree(table);
892 }
893 
894 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
895 {
896         struct kioctx_cpu *kcpu;
897         unsigned long flags;
898 
899         local_irq_save(flags);
900         kcpu = this_cpu_ptr(ctx->cpu);
901         kcpu->reqs_available += nr;
902 
903         while (kcpu->reqs_available >= ctx->req_batch * 2) {
904                 kcpu->reqs_available -= ctx->req_batch;
905                 atomic_add(ctx->req_batch, &ctx->reqs_available);
906         }
907 
908         local_irq_restore(flags);
909 }
910 
911 static bool get_reqs_available(struct kioctx *ctx)
912 {
913         struct kioctx_cpu *kcpu;
914         bool ret = false;
915         unsigned long flags;
916 
917         local_irq_save(flags);
918         kcpu = this_cpu_ptr(ctx->cpu);
919         if (!kcpu->reqs_available) {
920                 int old, avail = atomic_read(&ctx->reqs_available);
921 
922                 do {
923                         if (avail < ctx->req_batch)
924                                 goto out;
925 
926                         old = avail;
927                         avail = atomic_cmpxchg(&ctx->reqs_available,
928                                                avail, avail - ctx->req_batch);
929                 } while (avail != old);
930 
931                 kcpu->reqs_available += ctx->req_batch;
932         }
933 
934         ret = true;
935         kcpu->reqs_available--;
936 out:
937         local_irq_restore(flags);
938         return ret;
939 }
940 
941 /* refill_reqs_available
942  *      Updates the reqs_available reference counts used for tracking the
943  *      number of free slots in the completion ring.  This can be called
944  *      from aio_complete() (to optimistically update reqs_available) or
945  *      from aio_get_req() (the we're out of events case).  It must be
946  *      called holding ctx->completion_lock.
947  */
948 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
949                                   unsigned tail)
950 {
951         unsigned events_in_ring, completed;
952 
953         /* Clamp head since userland can write to it. */
954         head %= ctx->nr_events;
955         if (head <= tail)
956                 events_in_ring = tail - head;
957         else
958                 events_in_ring = ctx->nr_events - (head - tail);
959 
960         completed = ctx->completed_events;
961         if (events_in_ring < completed)
962                 completed -= events_in_ring;
963         else
964                 completed = 0;
965 
966         if (!completed)
967                 return;
968 
969         ctx->completed_events -= completed;
970         put_reqs_available(ctx, completed);
971 }
972 
973 /* user_refill_reqs_available
974  *      Called to refill reqs_available when aio_get_req() encounters an
975  *      out of space in the completion ring.
976  */
977 static void user_refill_reqs_available(struct kioctx *ctx)
978 {
979         spin_lock_irq(&ctx->completion_lock);
980         if (ctx->completed_events) {
981                 struct aio_ring *ring;
982                 unsigned head;
983 
984                 /* Access of ring->head may race with aio_read_events_ring()
985                  * here, but that's okay since whether we read the old version
986                  * or the new version, and either will be valid.  The important
987                  * part is that head cannot pass tail since we prevent
988                  * aio_complete() from updating tail by holding
989                  * ctx->completion_lock.  Even if head is invalid, the check
990                  * against ctx->completed_events below will make sure we do the
991                  * safe/right thing.
992                  */
993                 ring = kmap_atomic(ctx->ring_pages[0]);
994                 head = ring->head;
995                 kunmap_atomic(ring);
996 
997                 refill_reqs_available(ctx, head, ctx->tail);
998         }
999 
1000         spin_unlock_irq(&ctx->completion_lock);
1001 }
1002 
1003 /* aio_get_req
1004  *      Allocate a slot for an aio request.
1005  * Returns NULL if no requests are free.
1006  */
1007 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1008 {
1009         struct aio_kiocb *req;
1010 
1011         if (!get_reqs_available(ctx)) {
1012                 user_refill_reqs_available(ctx);
1013                 if (!get_reqs_available(ctx))
1014                         return NULL;
1015         }
1016 
1017         req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1018         if (unlikely(!req))
1019                 goto out_put;
1020 
1021         percpu_ref_get(&ctx->reqs);
1022 
1023         req->ki_ctx = ctx;
1024         return req;
1025 out_put:
1026         put_reqs_available(ctx, 1);
1027         return NULL;
1028 }
1029 
1030 static void kiocb_free(struct aio_kiocb *req)
1031 {
1032         if (req->common.ki_filp)
1033                 fput(req->common.ki_filp);
1034         if (req->ki_eventfd != NULL)
1035                 eventfd_ctx_put(req->ki_eventfd);
1036         kmem_cache_free(kiocb_cachep, req);
1037 }
1038 
1039 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1040 {
1041         struct aio_ring __user *ring  = (void __user *)ctx_id;
1042         struct mm_struct *mm = current->mm;
1043         struct kioctx *ctx, *ret = NULL;
1044         struct kioctx_table *table;
1045         unsigned id;
1046 
1047         if (get_user(id, &ring->id))
1048                 return NULL;
1049 
1050         rcu_read_lock();
1051         table = rcu_dereference(mm->ioctx_table);
1052 
1053         if (!table || id >= table->nr)
1054                 goto out;
1055 
1056         ctx = table->table[id];
1057         if (ctx && ctx->user_id == ctx_id) {
1058                 percpu_ref_get(&ctx->users);
1059                 ret = ctx;
1060         }
1061 out:
1062         rcu_read_unlock();
1063         return ret;
1064 }
1065 
1066 /* aio_complete
1067  *      Called when the io request on the given iocb is complete.
1068  */
1069 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1070 {
1071         struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1072         struct kioctx   *ctx = iocb->ki_ctx;
1073         struct aio_ring *ring;
1074         struct io_event *ev_page, *event;
1075         unsigned tail, pos, head;
1076         unsigned long   flags;
1077 
1078         /*
1079          * Special case handling for sync iocbs:
1080          *  - events go directly into the iocb for fast handling
1081          *  - the sync task with the iocb in its stack holds the single iocb
1082          *    ref, no other paths have a way to get another ref
1083          *  - the sync task helpfully left a reference to itself in the iocb
1084          */
1085         BUG_ON(is_sync_kiocb(kiocb));
1086 
1087         if (iocb->ki_list.next) {
1088                 unsigned long flags;
1089 
1090                 spin_lock_irqsave(&ctx->ctx_lock, flags);
1091                 list_del(&iocb->ki_list);
1092                 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1093         }
1094 
1095         /*
1096          * Add a completion event to the ring buffer. Must be done holding
1097          * ctx->completion_lock to prevent other code from messing with the tail
1098          * pointer since we might be called from irq context.
1099          */
1100         spin_lock_irqsave(&ctx->completion_lock, flags);
1101 
1102         tail = ctx->tail;
1103         pos = tail + AIO_EVENTS_OFFSET;
1104 
1105         if (++tail >= ctx->nr_events)
1106                 tail = 0;
1107 
1108         ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1109         event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1110 
1111         event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1112         event->data = iocb->ki_user_data;
1113         event->res = res;
1114         event->res2 = res2;
1115 
1116         kunmap_atomic(ev_page);
1117         flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1118 
1119         pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1120                  ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1121                  res, res2);
1122 
1123         /* after flagging the request as done, we
1124          * must never even look at it again
1125          */
1126         smp_wmb();      /* make event visible before updating tail */
1127 
1128         ctx->tail = tail;
1129 
1130         ring = kmap_atomic(ctx->ring_pages[0]);
1131         head = ring->head;
1132         ring->tail = tail;
1133         kunmap_atomic(ring);
1134         flush_dcache_page(ctx->ring_pages[0]);
1135 
1136         ctx->completed_events++;
1137         if (ctx->completed_events > 1)
1138                 refill_reqs_available(ctx, head, tail);
1139         spin_unlock_irqrestore(&ctx->completion_lock, flags);
1140 
1141         pr_debug("added to ring %p at [%u]\n", iocb, tail);
1142 
1143         /*
1144          * Check if the user asked us to deliver the result through an
1145          * eventfd. The eventfd_signal() function is safe to be called
1146          * from IRQ context.
1147          */
1148         if (iocb->ki_eventfd != NULL)
1149                 eventfd_signal(iocb->ki_eventfd, 1);
1150 
1151         /* everything turned out well, dispose of the aiocb. */
1152         kiocb_free(iocb);
1153 
1154         /*
1155          * We have to order our ring_info tail store above and test
1156          * of the wait list below outside the wait lock.  This is
1157          * like in wake_up_bit() where clearing a bit has to be
1158          * ordered with the unlocked test.
1159          */
1160         smp_mb();
1161 
1162         if (waitqueue_active(&ctx->wait))
1163                 wake_up(&ctx->wait);
1164 
1165         percpu_ref_put(&ctx->reqs);
1166 }
1167 
1168 /* aio_read_events_ring
1169  *      Pull an event off of the ioctx's event ring.  Returns the number of
1170  *      events fetched
1171  */
1172 static long aio_read_events_ring(struct kioctx *ctx,
1173                                  struct io_event __user *event, long nr)
1174 {
1175         struct aio_ring *ring;
1176         unsigned head, tail, pos;
1177         long ret = 0;
1178         int copy_ret;
1179 
1180         /*
1181          * The mutex can block and wake us up and that will cause
1182          * wait_event_interruptible_hrtimeout() to schedule without sleeping
1183          * and repeat. This should be rare enough that it doesn't cause
1184          * peformance issues. See the comment in read_events() for more detail.
1185          */
1186         sched_annotate_sleep();
1187         mutex_lock(&ctx->ring_lock);
1188 
1189         /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1190         ring = kmap_atomic(ctx->ring_pages[0]);
1191         head = ring->head;
1192         tail = ring->tail;
1193         kunmap_atomic(ring);
1194 
1195         /*
1196          * Ensure that once we've read the current tail pointer, that
1197          * we also see the events that were stored up to the tail.
1198          */
1199         smp_rmb();
1200 
1201         pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1202 
1203         if (head == tail)
1204                 goto out;
1205 
1206         head %= ctx->nr_events;
1207         tail %= ctx->nr_events;
1208 
1209         while (ret < nr) {
1210                 long avail;
1211                 struct io_event *ev;
1212                 struct page *page;
1213 
1214                 avail = (head <= tail ?  tail : ctx->nr_events) - head;
1215                 if (head == tail)
1216                         break;
1217 
1218                 avail = min(avail, nr - ret);
1219                 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1220                             ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1221 
1222                 pos = head + AIO_EVENTS_OFFSET;
1223                 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1224                 pos %= AIO_EVENTS_PER_PAGE;
1225 
1226                 ev = kmap(page);
1227                 copy_ret = copy_to_user(event + ret, ev + pos,
1228                                         sizeof(*ev) * avail);
1229                 kunmap(page);
1230 
1231                 if (unlikely(copy_ret)) {
1232                         ret = -EFAULT;
1233                         goto out;
1234                 }
1235 
1236                 ret += avail;
1237                 head += avail;
1238                 head %= ctx->nr_events;
1239         }
1240 
1241         ring = kmap_atomic(ctx->ring_pages[0]);
1242         ring->head = head;
1243         kunmap_atomic(ring);
1244         flush_dcache_page(ctx->ring_pages[0]);
1245 
1246         pr_debug("%li  h%u t%u\n", ret, head, tail);
1247 out:
1248         mutex_unlock(&ctx->ring_lock);
1249 
1250         return ret;
1251 }
1252 
1253 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1254                             struct io_event __user *event, long *i)
1255 {
1256         long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1257 
1258         if (ret > 0)
1259                 *i += ret;
1260 
1261         if (unlikely(atomic_read(&ctx->dead)))
1262                 ret = -EINVAL;
1263 
1264         if (!*i)
1265                 *i = ret;
1266 
1267         return ret < 0 || *i >= min_nr;
1268 }
1269 
1270 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1271                         struct io_event __user *event,
1272                         struct timespec __user *timeout)
1273 {
1274         ktime_t until = { .tv64 = KTIME_MAX };
1275         long ret = 0;
1276 
1277         if (timeout) {
1278                 struct timespec ts;
1279 
1280                 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1281                         return -EFAULT;
1282 
1283                 until = timespec_to_ktime(ts);
1284         }
1285 
1286         /*
1287          * Note that aio_read_events() is being called as the conditional - i.e.
1288          * we're calling it after prepare_to_wait() has set task state to
1289          * TASK_INTERRUPTIBLE.
1290          *
1291          * But aio_read_events() can block, and if it blocks it's going to flip
1292          * the task state back to TASK_RUNNING.
1293          *
1294          * This should be ok, provided it doesn't flip the state back to
1295          * TASK_RUNNING and return 0 too much - that causes us to spin. That
1296          * will only happen if the mutex_lock() call blocks, and we then find
1297          * the ringbuffer empty. So in practice we should be ok, but it's
1298          * something to be aware of when touching this code.
1299          */
1300         if (until.tv64 == 0)
1301                 aio_read_events(ctx, min_nr, nr, event, &ret);
1302         else
1303                 wait_event_interruptible_hrtimeout(ctx->wait,
1304                                 aio_read_events(ctx, min_nr, nr, event, &ret),
1305                                 until);
1306 
1307         if (!ret && signal_pending(current))
1308                 ret = -EINTR;
1309 
1310         return ret;
1311 }
1312 
1313 /* sys_io_setup:
1314  *      Create an aio_context capable of receiving at least nr_events.
1315  *      ctxp must not point to an aio_context that already exists, and
1316  *      must be initialized to 0 prior to the call.  On successful
1317  *      creation of the aio_context, *ctxp is filled in with the resulting 
1318  *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1319  *      if the specified nr_events exceeds internal limits.  May fail 
1320  *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1321  *      of available events.  May fail with -ENOMEM if insufficient kernel
1322  *      resources are available.  May fail with -EFAULT if an invalid
1323  *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1324  *      implemented.
1325  */
1326 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1327 {
1328         struct kioctx *ioctx = NULL;
1329         unsigned long ctx;
1330         long ret;
1331 
1332         ret = get_user(ctx, ctxp);
1333         if (unlikely(ret))
1334                 goto out;
1335 
1336         ret = -EINVAL;
1337         if (unlikely(ctx || nr_events == 0)) {
1338                 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1339                          ctx, nr_events);
1340                 goto out;
1341         }
1342 
1343         ioctx = ioctx_alloc(nr_events);
1344         ret = PTR_ERR(ioctx);
1345         if (!IS_ERR(ioctx)) {
1346                 ret = put_user(ioctx->user_id, ctxp);
1347                 if (ret)
1348                         kill_ioctx(current->mm, ioctx, NULL);
1349                 percpu_ref_put(&ioctx->users);
1350         }
1351 
1352 out:
1353         return ret;
1354 }
1355 
1356 /* sys_io_destroy:
1357  *      Destroy the aio_context specified.  May cancel any outstanding 
1358  *      AIOs and block on completion.  Will fail with -ENOSYS if not
1359  *      implemented.  May fail with -EINVAL if the context pointed to
1360  *      is invalid.
1361  */
1362 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1363 {
1364         struct kioctx *ioctx = lookup_ioctx(ctx);
1365         if (likely(NULL != ioctx)) {
1366                 struct ctx_rq_wait wait;
1367                 int ret;
1368 
1369                 init_completion(&wait.comp);
1370                 atomic_set(&wait.count, 1);
1371 
1372                 /* Pass requests_done to kill_ioctx() where it can be set
1373                  * in a thread-safe way. If we try to set it here then we have
1374                  * a race condition if two io_destroy() called simultaneously.
1375                  */
1376                 ret = kill_ioctx(current->mm, ioctx, &wait);
1377                 percpu_ref_put(&ioctx->users);
1378 
1379                 /* Wait until all IO for the context are done. Otherwise kernel
1380                  * keep using user-space buffers even if user thinks the context
1381                  * is destroyed.
1382                  */
1383                 if (!ret)
1384                         wait_for_completion(&wait.comp);
1385 
1386                 return ret;
1387         }
1388         pr_debug("EINVAL: invalid context id\n");
1389         return -EINVAL;
1390 }
1391 
1392 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1393 
1394 static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1395                                  struct iovec **iovec,
1396                                  bool compat,
1397                                  struct iov_iter *iter)
1398 {
1399 #ifdef CONFIG_COMPAT
1400         if (compat)
1401                 return compat_import_iovec(rw,
1402                                 (struct compat_iovec __user *)buf,
1403                                 len, UIO_FASTIOV, iovec, iter);
1404 #endif
1405         return import_iovec(rw, (struct iovec __user *)buf,
1406                                 len, UIO_FASTIOV, iovec, iter);
1407 }
1408 
1409 /*
1410  * aio_run_iocb:
1411  *      Performs the initial checks and io submission.
1412  */
1413 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1414                             char __user *buf, size_t len, bool compat)
1415 {
1416         struct file *file = req->ki_filp;
1417         ssize_t ret;
1418         int rw;
1419         fmode_t mode;
1420         rw_iter_op *iter_op;
1421         struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1422         struct iov_iter iter;
1423 
1424         switch (opcode) {
1425         case IOCB_CMD_PREAD:
1426         case IOCB_CMD_PREADV:
1427                 mode    = FMODE_READ;
1428                 rw      = READ;
1429                 iter_op = file->f_op->read_iter;
1430                 goto rw_common;
1431 
1432         case IOCB_CMD_PWRITE:
1433         case IOCB_CMD_PWRITEV:
1434                 mode    = FMODE_WRITE;
1435                 rw      = WRITE;
1436                 iter_op = file->f_op->write_iter;
1437                 goto rw_common;
1438 rw_common:
1439                 if (unlikely(!(file->f_mode & mode)))
1440                         return -EBADF;
1441 
1442                 if (!iter_op)
1443                         return -EINVAL;
1444 
1445                 if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
1446                         ret = aio_setup_vectored_rw(rw, buf, len,
1447                                                 &iovec, compat, &iter);
1448                 else {
1449                         ret = import_single_range(rw, buf, len, iovec, &iter);
1450                         iovec = NULL;
1451                 }
1452                 if (!ret)
1453                         ret = rw_verify_area(rw, file, &req->ki_pos,
1454                                              iov_iter_count(&iter));
1455                 if (ret < 0) {
1456                         kfree(iovec);
1457                         return ret;
1458                 }
1459 
1460                 if (rw == WRITE)
1461                         file_start_write(file);
1462 
1463                 ret = iter_op(req, &iter);
1464 
1465                 if (rw == WRITE)
1466                         file_end_write(file);
1467                 kfree(iovec);
1468                 break;
1469 
1470         case IOCB_CMD_FDSYNC:
1471                 if (!file->f_op->aio_fsync)
1472                         return -EINVAL;
1473 
1474                 ret = file->f_op->aio_fsync(req, 1);
1475                 break;
1476 
1477         case IOCB_CMD_FSYNC:
1478                 if (!file->f_op->aio_fsync)
1479                         return -EINVAL;
1480 
1481                 ret = file->f_op->aio_fsync(req, 0);
1482                 break;
1483 
1484         default:
1485                 pr_debug("EINVAL: no operation provided\n");
1486                 return -EINVAL;
1487         }
1488 
1489         if (ret != -EIOCBQUEUED) {
1490                 /*
1491                  * There's no easy way to restart the syscall since other AIO's
1492                  * may be already running. Just fail this IO with EINTR.
1493                  */
1494                 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1495                              ret == -ERESTARTNOHAND ||
1496                              ret == -ERESTART_RESTARTBLOCK))
1497                         ret = -EINTR;
1498                 aio_complete(req, ret, 0);
1499         }
1500 
1501         return 0;
1502 }
1503 
1504 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1505                          struct iocb *iocb, bool compat)
1506 {
1507         struct aio_kiocb *req;
1508         ssize_t ret;
1509 
1510         /* enforce forwards compatibility on users */
1511         if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1512                 pr_debug("EINVAL: reserve field set\n");
1513                 return -EINVAL;
1514         }
1515 
1516         /* prevent overflows */
1517         if (unlikely(
1518             (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1519             (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1520             ((ssize_t)iocb->aio_nbytes < 0)
1521            )) {
1522                 pr_debug("EINVAL: overflow check\n");
1523                 return -EINVAL;
1524         }
1525 
1526         req = aio_get_req(ctx);
1527         if (unlikely(!req))
1528                 return -EAGAIN;
1529 
1530         req->common.ki_filp = fget(iocb->aio_fildes);
1531         if (unlikely(!req->common.ki_filp)) {
1532                 ret = -EBADF;
1533                 goto out_put_req;
1534         }
1535         req->common.ki_pos = iocb->aio_offset;
1536         req->common.ki_complete = aio_complete;
1537         req->common.ki_flags = iocb_flags(req->common.ki_filp);
1538 
1539         if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1540                 /*
1541                  * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1542                  * instance of the file* now. The file descriptor must be
1543                  * an eventfd() fd, and will be signaled for each completed
1544                  * event using the eventfd_signal() function.
1545                  */
1546                 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1547                 if (IS_ERR(req->ki_eventfd)) {
1548                         ret = PTR_ERR(req->ki_eventfd);
1549                         req->ki_eventfd = NULL;
1550                         goto out_put_req;
1551                 }
1552 
1553                 req->common.ki_flags |= IOCB_EVENTFD;
1554         }
1555 
1556         ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1557         if (unlikely(ret)) {
1558                 pr_debug("EFAULT: aio_key\n");
1559                 goto out_put_req;
1560         }
1561 
1562         req->ki_user_iocb = user_iocb;
1563         req->ki_user_data = iocb->aio_data;
1564 
1565         ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
1566                            (char __user *)(unsigned long)iocb->aio_buf,
1567                            iocb->aio_nbytes,
1568                            compat);
1569         if (ret)
1570                 goto out_put_req;
1571 
1572         return 0;
1573 out_put_req:
1574         put_reqs_available(ctx, 1);
1575         percpu_ref_put(&ctx->reqs);
1576         kiocb_free(req);
1577         return ret;
1578 }
1579 
1580 long do_io_submit(aio_context_t ctx_id, long nr,
1581                   struct iocb __user *__user *iocbpp, bool compat)
1582 {
1583         struct kioctx *ctx;
1584         long ret = 0;
1585         int i = 0;
1586         struct blk_plug plug;
1587 
1588         if (unlikely(nr < 0))
1589                 return -EINVAL;
1590 
1591         if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1592                 nr = LONG_MAX/sizeof(*iocbpp);
1593 
1594         if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1595                 return -EFAULT;
1596 
1597         ctx = lookup_ioctx(ctx_id);
1598         if (unlikely(!ctx)) {
1599                 pr_debug("EINVAL: invalid context id\n");
1600                 return -EINVAL;
1601         }
1602 
1603         blk_start_plug(&plug);
1604 
1605         /*
1606          * AKPM: should this return a partial result if some of the IOs were
1607          * successfully submitted?
1608          */
1609         for (i=0; i<nr; i++) {
1610                 struct iocb __user *user_iocb;
1611                 struct iocb tmp;
1612 
1613                 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1614                         ret = -EFAULT;
1615                         break;
1616                 }
1617 
1618                 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1619                         ret = -EFAULT;
1620                         break;
1621                 }
1622 
1623                 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1624                 if (ret)
1625                         break;
1626         }
1627         blk_finish_plug(&plug);
1628 
1629         percpu_ref_put(&ctx->users);
1630         return i ? i : ret;
1631 }
1632 
1633 /* sys_io_submit:
1634  *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1635  *      the number of iocbs queued.  May return -EINVAL if the aio_context
1636  *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1637  *      *iocbpp[0] is not properly initialized, if the operation specified
1638  *      is invalid for the file descriptor in the iocb.  May fail with
1639  *      -EFAULT if any of the data structures point to invalid data.  May
1640  *      fail with -EBADF if the file descriptor specified in the first
1641  *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1642  *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1643  *      fail with -ENOSYS if not implemented.
1644  */
1645 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1646                 struct iocb __user * __user *, iocbpp)
1647 {
1648         return do_io_submit(ctx_id, nr, iocbpp, 0);
1649 }
1650 
1651 /* lookup_kiocb
1652  *      Finds a given iocb for cancellation.
1653  */
1654 static struct aio_kiocb *
1655 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1656 {
1657         struct aio_kiocb *kiocb;
1658 
1659         assert_spin_locked(&ctx->ctx_lock);
1660 
1661         if (key != KIOCB_KEY)
1662                 return NULL;
1663 
1664         /* TODO: use a hash or array, this sucks. */
1665         list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1666                 if (kiocb->ki_user_iocb == iocb)
1667                         return kiocb;
1668         }
1669         return NULL;
1670 }
1671 
1672 /* sys_io_cancel:
1673  *      Attempts to cancel an iocb previously passed to io_submit.  If
1674  *      the operation is successfully cancelled, the resulting event is
1675  *      copied into the memory pointed to by result without being placed
1676  *      into the completion queue and 0 is returned.  May fail with
1677  *      -EFAULT if any of the data structures pointed to are invalid.
1678  *      May fail with -EINVAL if aio_context specified by ctx_id is
1679  *      invalid.  May fail with -EAGAIN if the iocb specified was not
1680  *      cancelled.  Will fail with -ENOSYS if not implemented.
1681  */
1682 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1683                 struct io_event __user *, result)
1684 {
1685         struct kioctx *ctx;
1686         struct aio_kiocb *kiocb;
1687         u32 key;
1688         int ret;
1689 
1690         ret = get_user(key, &iocb->aio_key);
1691         if (unlikely(ret))
1692                 return -EFAULT;
1693 
1694         ctx = lookup_ioctx(ctx_id);
1695         if (unlikely(!ctx))
1696                 return -EINVAL;
1697 
1698         spin_lock_irq(&ctx->ctx_lock);
1699 
1700         kiocb = lookup_kiocb(ctx, iocb, key);
1701         if (kiocb)
1702                 ret = kiocb_cancel(kiocb);
1703         else
1704                 ret = -EINVAL;
1705 
1706         spin_unlock_irq(&ctx->ctx_lock);
1707 
1708         if (!ret) {
1709                 /*
1710                  * The result argument is no longer used - the io_event is
1711                  * always delivered via the ring buffer. -EINPROGRESS indicates
1712                  * cancellation is progress:
1713                  */
1714                 ret = -EINPROGRESS;
1715         }
1716 
1717         percpu_ref_put(&ctx->users);
1718 
1719         return ret;
1720 }
1721 
1722 /* io_getevents:
1723  *      Attempts to read at least min_nr events and up to nr events from
1724  *      the completion queue for the aio_context specified by ctx_id. If
1725  *      it succeeds, the number of read events is returned. May fail with
1726  *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1727  *      out of range, if timeout is out of range.  May fail with -EFAULT
1728  *      if any of the memory specified is invalid.  May return 0 or
1729  *      < min_nr if the timeout specified by timeout has elapsed
1730  *      before sufficient events are available, where timeout == NULL
1731  *      specifies an infinite timeout. Note that the timeout pointed to by
1732  *      timeout is relative.  Will fail with -ENOSYS if not implemented.
1733  */
1734 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1735                 long, min_nr,
1736                 long, nr,
1737                 struct io_event __user *, events,
1738                 struct timespec __user *, timeout)
1739 {
1740         struct kioctx *ioctx = lookup_ioctx(ctx_id);
1741         long ret = -EINVAL;
1742 
1743         if (likely(ioctx)) {
1744                 if (likely(min_nr <= nr && min_nr >= 0))
1745                         ret = read_events(ioctx, min_nr, nr, events, timeout);
1746                 percpu_ref_put(&ioctx->users);
1747         }
1748         return ret;
1749 }
1750 

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