Version:  2.0.40 2.2.26 2.4.37 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0

Linux/fs/aio.c

  1 /*
  2  *      An async IO implementation for Linux
  3  *      Written by Benjamin LaHaise <bcrl@kvack.org>
  4  *
  5  *      Implements an efficient asynchronous io interface.
  6  *
  7  *      Copyright 2000, 2001, 2002 Red Hat, Inc.  All Rights Reserved.
  8  *
  9  *      See ../COPYING for licensing terms.
 10  */
 11 #define pr_fmt(fmt) "%s: " fmt, __func__
 12 
 13 #include <linux/kernel.h>
 14 #include <linux/init.h>
 15 #include <linux/errno.h>
 16 #include <linux/time.h>
 17 #include <linux/aio_abi.h>
 18 #include <linux/export.h>
 19 #include <linux/syscalls.h>
 20 #include <linux/backing-dev.h>
 21 #include <linux/uio.h>
 22 
 23 #include <linux/sched.h>
 24 #include <linux/fs.h>
 25 #include <linux/file.h>
 26 #include <linux/mm.h>
 27 #include <linux/mman.h>
 28 #include <linux/mmu_context.h>
 29 #include <linux/percpu.h>
 30 #include <linux/slab.h>
 31 #include <linux/timer.h>
 32 #include <linux/aio.h>
 33 #include <linux/highmem.h>
 34 #include <linux/workqueue.h>
 35 #include <linux/security.h>
 36 #include <linux/eventfd.h>
 37 #include <linux/blkdev.h>
 38 #include <linux/compat.h>
 39 #include <linux/migrate.h>
 40 #include <linux/ramfs.h>
 41 #include <linux/percpu-refcount.h>
 42 #include <linux/mount.h>
 43 
 44 #include <asm/kmap_types.h>
 45 #include <asm/uaccess.h>
 46 
 47 #include "internal.h"
 48 
 49 #define AIO_RING_MAGIC                  0xa10a10a1
 50 #define AIO_RING_COMPAT_FEATURES        1
 51 #define AIO_RING_INCOMPAT_FEATURES      0
 52 struct aio_ring {
 53         unsigned        id;     /* kernel internal index number */
 54         unsigned        nr;     /* number of io_events */
 55         unsigned        head;   /* Written to by userland or under ring_lock
 56                                  * mutex by aio_read_events_ring(). */
 57         unsigned        tail;
 58 
 59         unsigned        magic;
 60         unsigned        compat_features;
 61         unsigned        incompat_features;
 62         unsigned        header_length;  /* size of aio_ring */
 63 
 64 
 65         struct io_event         io_events[0];
 66 }; /* 128 bytes + ring size */
 67 
 68 #define AIO_RING_PAGES  8
 69 
 70 struct kioctx_table {
 71         struct rcu_head rcu;
 72         unsigned        nr;
 73         struct kioctx   *table[];
 74 };
 75 
 76 struct kioctx_cpu {
 77         unsigned                reqs_available;
 78 };
 79 
 80 struct kioctx {
 81         struct percpu_ref       users;
 82         atomic_t                dead;
 83 
 84         struct percpu_ref       reqs;
 85 
 86         unsigned long           user_id;
 87 
 88         struct __percpu kioctx_cpu *cpu;
 89 
 90         /*
 91          * For percpu reqs_available, number of slots we move to/from global
 92          * counter at a time:
 93          */
 94         unsigned                req_batch;
 95         /*
 96          * This is what userspace passed to io_setup(), it's not used for
 97          * anything but counting against the global max_reqs quota.
 98          *
 99          * The real limit is nr_events - 1, which will be larger (see
100          * aio_setup_ring())
101          */
102         unsigned                max_reqs;
103 
104         /* Size of ringbuffer, in units of struct io_event */
105         unsigned                nr_events;
106 
107         unsigned long           mmap_base;
108         unsigned long           mmap_size;
109 
110         struct page             **ring_pages;
111         long                    nr_pages;
112 
113         struct work_struct      free_work;
114 
115         /*
116          * signals when all in-flight requests are done
117          */
118         struct completion *requests_done;
119 
120         struct {
121                 /*
122                  * This counts the number of available slots in the ringbuffer,
123                  * so we avoid overflowing it: it's decremented (if positive)
124                  * when allocating a kiocb and incremented when the resulting
125                  * io_event is pulled off the ringbuffer.
126                  *
127                  * We batch accesses to it with a percpu version.
128                  */
129                 atomic_t        reqs_available;
130         } ____cacheline_aligned_in_smp;
131 
132         struct {
133                 spinlock_t      ctx_lock;
134                 struct list_head active_reqs;   /* used for cancellation */
135         } ____cacheline_aligned_in_smp;
136 
137         struct {
138                 struct mutex    ring_lock;
139                 wait_queue_head_t wait;
140         } ____cacheline_aligned_in_smp;
141 
142         struct {
143                 unsigned        tail;
144                 unsigned        completed_events;
145                 spinlock_t      completion_lock;
146         } ____cacheline_aligned_in_smp;
147 
148         struct page             *internal_pages[AIO_RING_PAGES];
149         struct file             *aio_ring_file;
150 
151         unsigned                id;
152 };
153 
154 /*------ sysctl variables----*/
155 static DEFINE_SPINLOCK(aio_nr_lock);
156 unsigned long aio_nr;           /* current system wide number of aio requests */
157 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
158 /*----end sysctl variables---*/
159 
160 static struct kmem_cache        *kiocb_cachep;
161 static struct kmem_cache        *kioctx_cachep;
162 
163 static struct vfsmount *aio_mnt;
164 
165 static const struct file_operations aio_ring_fops;
166 static const struct address_space_operations aio_ctx_aops;
167 
168 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
169 {
170         struct qstr this = QSTR_INIT("[aio]", 5);
171         struct file *file;
172         struct path path;
173         struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
174         if (IS_ERR(inode))
175                 return ERR_CAST(inode);
176 
177         inode->i_mapping->a_ops = &aio_ctx_aops;
178         inode->i_mapping->private_data = ctx;
179         inode->i_size = PAGE_SIZE * nr_pages;
180 
181         path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
182         if (!path.dentry) {
183                 iput(inode);
184                 return ERR_PTR(-ENOMEM);
185         }
186         path.mnt = mntget(aio_mnt);
187 
188         d_instantiate(path.dentry, inode);
189         file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
190         if (IS_ERR(file)) {
191                 path_put(&path);
192                 return file;
193         }
194 
195         file->f_flags = O_RDWR;
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, AIO_RING_MAGIC);
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_flags |= VM_DONTEXPAND;
277         vma->vm_ops = &generic_file_vm_ops;
278         return 0;
279 }
280 
281 static int aio_ring_remap(struct file *file, struct vm_area_struct *vma)
282 {
283         struct mm_struct *mm = vma->vm_mm;
284         struct kioctx_table *table;
285         int i, res = -EINVAL;
286 
287         spin_lock(&mm->ioctx_lock);
288         rcu_read_lock();
289         table = rcu_dereference(mm->ioctx_table);
290         for (i = 0; i < table->nr; i++) {
291                 struct kioctx *ctx;
292 
293                 ctx = table->table[i];
294                 if (ctx && ctx->aio_ring_file == file) {
295                         if (!atomic_read(&ctx->dead)) {
296                                 ctx->user_id = ctx->mmap_base = vma->vm_start;
297                                 res = 0;
298                         }
299                         break;
300                 }
301         }
302 
303         rcu_read_unlock();
304         spin_unlock(&mm->ioctx_lock);
305         return res;
306 }
307 
308 static const struct file_operations aio_ring_fops = {
309         .mmap = aio_ring_mmap,
310         .mremap = aio_ring_remap,
311 };
312 
313 #if IS_ENABLED(CONFIG_MIGRATION)
314 static int aio_migratepage(struct address_space *mapping, struct page *new,
315                         struct page *old, enum migrate_mode mode)
316 {
317         struct kioctx *ctx;
318         unsigned long flags;
319         pgoff_t idx;
320         int rc;
321 
322         rc = 0;
323 
324         /* mapping->private_lock here protects against the kioctx teardown.  */
325         spin_lock(&mapping->private_lock);
326         ctx = mapping->private_data;
327         if (!ctx) {
328                 rc = -EINVAL;
329                 goto out;
330         }
331 
332         /* The ring_lock mutex.  The prevents aio_read_events() from writing
333          * to the ring's head, and prevents page migration from mucking in
334          * a partially initialized kiotx.
335          */
336         if (!mutex_trylock(&ctx->ring_lock)) {
337                 rc = -EAGAIN;
338                 goto out;
339         }
340 
341         idx = old->index;
342         if (idx < (pgoff_t)ctx->nr_pages) {
343                 /* Make sure the old page hasn't already been changed */
344                 if (ctx->ring_pages[idx] != old)
345                         rc = -EAGAIN;
346         } else
347                 rc = -EINVAL;
348 
349         if (rc != 0)
350                 goto out_unlock;
351 
352         /* Writeback must be complete */
353         BUG_ON(PageWriteback(old));
354         get_page(new);
355 
356         rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
357         if (rc != MIGRATEPAGE_SUCCESS) {
358                 put_page(new);
359                 goto out_unlock;
360         }
361 
362         /* Take completion_lock to prevent other writes to the ring buffer
363          * while the old page is copied to the new.  This prevents new
364          * events from being lost.
365          */
366         spin_lock_irqsave(&ctx->completion_lock, flags);
367         migrate_page_copy(new, old);
368         BUG_ON(ctx->ring_pages[idx] != old);
369         ctx->ring_pages[idx] = new;
370         spin_unlock_irqrestore(&ctx->completion_lock, flags);
371 
372         /* The old page is no longer accessible. */
373         put_page(old);
374 
375 out_unlock:
376         mutex_unlock(&ctx->ring_lock);
377 out:
378         spin_unlock(&mapping->private_lock);
379         return rc;
380 }
381 #endif
382 
383 static const struct address_space_operations aio_ctx_aops = {
384         .set_page_dirty = __set_page_dirty_no_writeback,
385 #if IS_ENABLED(CONFIG_MIGRATION)
386         .migratepage    = aio_migratepage,
387 #endif
388 };
389 
390 static int aio_setup_ring(struct kioctx *ctx)
391 {
392         struct aio_ring *ring;
393         unsigned nr_events = ctx->max_reqs;
394         struct mm_struct *mm = current->mm;
395         unsigned long size, unused;
396         int nr_pages;
397         int i;
398         struct file *file;
399 
400         /* Compensate for the ring buffer's head/tail overlap entry */
401         nr_events += 2; /* 1 is required, 2 for good luck */
402 
403         size = sizeof(struct aio_ring);
404         size += sizeof(struct io_event) * nr_events;
405 
406         nr_pages = PFN_UP(size);
407         if (nr_pages < 0)
408                 return -EINVAL;
409 
410         file = aio_private_file(ctx, nr_pages);
411         if (IS_ERR(file)) {
412                 ctx->aio_ring_file = NULL;
413                 return -ENOMEM;
414         }
415 
416         ctx->aio_ring_file = file;
417         nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
418                         / sizeof(struct io_event);
419 
420         ctx->ring_pages = ctx->internal_pages;
421         if (nr_pages > AIO_RING_PAGES) {
422                 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
423                                           GFP_KERNEL);
424                 if (!ctx->ring_pages) {
425                         put_aio_ring_file(ctx);
426                         return -ENOMEM;
427                 }
428         }
429 
430         for (i = 0; i < nr_pages; i++) {
431                 struct page *page;
432                 page = find_or_create_page(file->f_inode->i_mapping,
433                                            i, GFP_HIGHUSER | __GFP_ZERO);
434                 if (!page)
435                         break;
436                 pr_debug("pid(%d) page[%d]->count=%d\n",
437                          current->pid, i, page_count(page));
438                 SetPageUptodate(page);
439                 unlock_page(page);
440 
441                 ctx->ring_pages[i] = page;
442         }
443         ctx->nr_pages = i;
444 
445         if (unlikely(i != nr_pages)) {
446                 aio_free_ring(ctx);
447                 return -ENOMEM;
448         }
449 
450         ctx->mmap_size = nr_pages * PAGE_SIZE;
451         pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
452 
453         down_write(&mm->mmap_sem);
454         ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
455                                        PROT_READ | PROT_WRITE,
456                                        MAP_SHARED, 0, &unused);
457         up_write(&mm->mmap_sem);
458         if (IS_ERR((void *)ctx->mmap_base)) {
459                 ctx->mmap_size = 0;
460                 aio_free_ring(ctx);
461                 return -ENOMEM;
462         }
463 
464         pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
465 
466         ctx->user_id = ctx->mmap_base;
467         ctx->nr_events = nr_events; /* trusted copy */
468 
469         ring = kmap_atomic(ctx->ring_pages[0]);
470         ring->nr = nr_events;   /* user copy */
471         ring->id = ~0U;
472         ring->head = ring->tail = 0;
473         ring->magic = AIO_RING_MAGIC;
474         ring->compat_features = AIO_RING_COMPAT_FEATURES;
475         ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
476         ring->header_length = sizeof(struct aio_ring);
477         kunmap_atomic(ring);
478         flush_dcache_page(ctx->ring_pages[0]);
479 
480         return 0;
481 }
482 
483 #define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
484 #define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
485 #define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
486 
487 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
488 {
489         struct kioctx *ctx = req->ki_ctx;
490         unsigned long flags;
491 
492         spin_lock_irqsave(&ctx->ctx_lock, flags);
493 
494         if (!req->ki_list.next)
495                 list_add(&req->ki_list, &ctx->active_reqs);
496 
497         req->ki_cancel = cancel;
498 
499         spin_unlock_irqrestore(&ctx->ctx_lock, flags);
500 }
501 EXPORT_SYMBOL(kiocb_set_cancel_fn);
502 
503 static int kiocb_cancel(struct kiocb *kiocb)
504 {
505         kiocb_cancel_fn *old, *cancel;
506 
507         /*
508          * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
509          * actually has a cancel function, hence the cmpxchg()
510          */
511 
512         cancel = ACCESS_ONCE(kiocb->ki_cancel);
513         do {
514                 if (!cancel || cancel == KIOCB_CANCELLED)
515                         return -EINVAL;
516 
517                 old = cancel;
518                 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
519         } while (cancel != old);
520 
521         return cancel(kiocb);
522 }
523 
524 static void free_ioctx(struct work_struct *work)
525 {
526         struct kioctx *ctx = container_of(work, struct kioctx, free_work);
527 
528         pr_debug("freeing %p\n", ctx);
529 
530         aio_free_ring(ctx);
531         free_percpu(ctx->cpu);
532         percpu_ref_exit(&ctx->reqs);
533         percpu_ref_exit(&ctx->users);
534         kmem_cache_free(kioctx_cachep, ctx);
535 }
536 
537 static void free_ioctx_reqs(struct percpu_ref *ref)
538 {
539         struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
540 
541         /* At this point we know that there are no any in-flight requests */
542         if (ctx->requests_done)
543                 complete(ctx->requests_done);
544 
545         INIT_WORK(&ctx->free_work, free_ioctx);
546         schedule_work(&ctx->free_work);
547 }
548 
549 /*
550  * When this function runs, the kioctx has been removed from the "hash table"
551  * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
552  * now it's safe to cancel any that need to be.
553  */
554 static void free_ioctx_users(struct percpu_ref *ref)
555 {
556         struct kioctx *ctx = container_of(ref, struct kioctx, users);
557         struct kiocb *req;
558 
559         spin_lock_irq(&ctx->ctx_lock);
560 
561         while (!list_empty(&ctx->active_reqs)) {
562                 req = list_first_entry(&ctx->active_reqs,
563                                        struct kiocb, ki_list);
564 
565                 list_del_init(&req->ki_list);
566                 kiocb_cancel(req);
567         }
568 
569         spin_unlock_irq(&ctx->ctx_lock);
570 
571         percpu_ref_kill(&ctx->reqs);
572         percpu_ref_put(&ctx->reqs);
573 }
574 
575 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
576 {
577         unsigned i, new_nr;
578         struct kioctx_table *table, *old;
579         struct aio_ring *ring;
580 
581         spin_lock(&mm->ioctx_lock);
582         table = rcu_dereference_raw(mm->ioctx_table);
583 
584         while (1) {
585                 if (table)
586                         for (i = 0; i < table->nr; i++)
587                                 if (!table->table[i]) {
588                                         ctx->id = i;
589                                         table->table[i] = ctx;
590                                         spin_unlock(&mm->ioctx_lock);
591 
592                                         /* While kioctx setup is in progress,
593                                          * we are protected from page migration
594                                          * changes ring_pages by ->ring_lock.
595                                          */
596                                         ring = kmap_atomic(ctx->ring_pages[0]);
597                                         ring->id = ctx->id;
598                                         kunmap_atomic(ring);
599                                         return 0;
600                                 }
601 
602                 new_nr = (table ? table->nr : 1) * 4;
603                 spin_unlock(&mm->ioctx_lock);
604 
605                 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
606                                 new_nr, GFP_KERNEL);
607                 if (!table)
608                         return -ENOMEM;
609 
610                 table->nr = new_nr;
611 
612                 spin_lock(&mm->ioctx_lock);
613                 old = rcu_dereference_raw(mm->ioctx_table);
614 
615                 if (!old) {
616                         rcu_assign_pointer(mm->ioctx_table, table);
617                 } else if (table->nr > old->nr) {
618                         memcpy(table->table, old->table,
619                                old->nr * sizeof(struct kioctx *));
620 
621                         rcu_assign_pointer(mm->ioctx_table, table);
622                         kfree_rcu(old, rcu);
623                 } else {
624                         kfree(table);
625                         table = old;
626                 }
627         }
628 }
629 
630 static void aio_nr_sub(unsigned nr)
631 {
632         spin_lock(&aio_nr_lock);
633         if (WARN_ON(aio_nr - nr > aio_nr))
634                 aio_nr = 0;
635         else
636                 aio_nr -= nr;
637         spin_unlock(&aio_nr_lock);
638 }
639 
640 /* ioctx_alloc
641  *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
642  */
643 static struct kioctx *ioctx_alloc(unsigned nr_events)
644 {
645         struct mm_struct *mm = current->mm;
646         struct kioctx *ctx;
647         int err = -ENOMEM;
648 
649         /*
650          * We keep track of the number of available ringbuffer slots, to prevent
651          * overflow (reqs_available), and we also use percpu counters for this.
652          *
653          * So since up to half the slots might be on other cpu's percpu counters
654          * and unavailable, double nr_events so userspace sees what they
655          * expected: additionally, we move req_batch slots to/from percpu
656          * counters at a time, so make sure that isn't 0:
657          */
658         nr_events = max(nr_events, num_possible_cpus() * 4);
659         nr_events *= 2;
660 
661         /* Prevent overflows */
662         if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
663             (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
664                 pr_debug("ENOMEM: nr_events too high\n");
665                 return ERR_PTR(-EINVAL);
666         }
667 
668         if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
669                 return ERR_PTR(-EAGAIN);
670 
671         ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
672         if (!ctx)
673                 return ERR_PTR(-ENOMEM);
674 
675         ctx->max_reqs = nr_events;
676 
677         spin_lock_init(&ctx->ctx_lock);
678         spin_lock_init(&ctx->completion_lock);
679         mutex_init(&ctx->ring_lock);
680         /* Protect against page migration throughout kiotx setup by keeping
681          * the ring_lock mutex held until setup is complete. */
682         mutex_lock(&ctx->ring_lock);
683         init_waitqueue_head(&ctx->wait);
684 
685         INIT_LIST_HEAD(&ctx->active_reqs);
686 
687         if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
688                 goto err;
689 
690         if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
691                 goto err;
692 
693         ctx->cpu = alloc_percpu(struct kioctx_cpu);
694         if (!ctx->cpu)
695                 goto err;
696 
697         err = aio_setup_ring(ctx);
698         if (err < 0)
699                 goto err;
700 
701         atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
702         ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
703         if (ctx->req_batch < 1)
704                 ctx->req_batch = 1;
705 
706         /* limit the number of system wide aios */
707         spin_lock(&aio_nr_lock);
708         if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
709             aio_nr + nr_events < aio_nr) {
710                 spin_unlock(&aio_nr_lock);
711                 err = -EAGAIN;
712                 goto err_ctx;
713         }
714         aio_nr += ctx->max_reqs;
715         spin_unlock(&aio_nr_lock);
716 
717         percpu_ref_get(&ctx->users);    /* io_setup() will drop this ref */
718         percpu_ref_get(&ctx->reqs);     /* free_ioctx_users() will drop this */
719 
720         err = ioctx_add_table(ctx, mm);
721         if (err)
722                 goto err_cleanup;
723 
724         /* Release the ring_lock mutex now that all setup is complete. */
725         mutex_unlock(&ctx->ring_lock);
726 
727         pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
728                  ctx, ctx->user_id, mm, ctx->nr_events);
729         return ctx;
730 
731 err_cleanup:
732         aio_nr_sub(ctx->max_reqs);
733 err_ctx:
734         atomic_set(&ctx->dead, 1);
735         if (ctx->mmap_size)
736                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
737         aio_free_ring(ctx);
738 err:
739         mutex_unlock(&ctx->ring_lock);
740         free_percpu(ctx->cpu);
741         percpu_ref_exit(&ctx->reqs);
742         percpu_ref_exit(&ctx->users);
743         kmem_cache_free(kioctx_cachep, ctx);
744         pr_debug("error allocating ioctx %d\n", err);
745         return ERR_PTR(err);
746 }
747 
748 /* kill_ioctx
749  *      Cancels all outstanding aio requests on an aio context.  Used
750  *      when the processes owning a context have all exited to encourage
751  *      the rapid destruction of the kioctx.
752  */
753 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
754                 struct completion *requests_done)
755 {
756         struct kioctx_table *table;
757 
758         spin_lock(&mm->ioctx_lock);
759         if (atomic_xchg(&ctx->dead, 1)) {
760                 spin_unlock(&mm->ioctx_lock);
761                 return -EINVAL;
762         }
763 
764         table = rcu_dereference_raw(mm->ioctx_table);
765         WARN_ON(ctx != table->table[ctx->id]);
766         table->table[ctx->id] = NULL;
767         spin_unlock(&mm->ioctx_lock);
768 
769         /* percpu_ref_kill() will do the necessary call_rcu() */
770         wake_up_all(&ctx->wait);
771 
772         /*
773          * It'd be more correct to do this in free_ioctx(), after all
774          * the outstanding kiocbs have finished - but by then io_destroy
775          * has already returned, so io_setup() could potentially return
776          * -EAGAIN with no ioctxs actually in use (as far as userspace
777          *  could tell).
778          */
779         aio_nr_sub(ctx->max_reqs);
780 
781         if (ctx->mmap_size)
782                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
783 
784         ctx->requests_done = requests_done;
785         percpu_ref_kill(&ctx->users);
786         return 0;
787 }
788 
789 /* wait_on_sync_kiocb:
790  *      Waits on the given sync kiocb to complete.
791  */
792 ssize_t wait_on_sync_kiocb(struct kiocb *req)
793 {
794         while (!req->ki_ctx) {
795                 set_current_state(TASK_UNINTERRUPTIBLE);
796                 if (req->ki_ctx)
797                         break;
798                 io_schedule();
799         }
800         __set_current_state(TASK_RUNNING);
801         return req->ki_user_data;
802 }
803 EXPORT_SYMBOL(wait_on_sync_kiocb);
804 
805 /*
806  * exit_aio: called when the last user of mm goes away.  At this point, there is
807  * no way for any new requests to be submited or any of the io_* syscalls to be
808  * called on the context.
809  *
810  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
811  * them.
812  */
813 void exit_aio(struct mm_struct *mm)
814 {
815         struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
816         int i;
817 
818         if (!table)
819                 return;
820 
821         for (i = 0; i < table->nr; ++i) {
822                 struct kioctx *ctx = table->table[i];
823                 struct completion requests_done =
824                         COMPLETION_INITIALIZER_ONSTACK(requests_done);
825 
826                 if (!ctx)
827                         continue;
828                 /*
829                  * We don't need to bother with munmap() here - exit_mmap(mm)
830                  * is coming and it'll unmap everything. And we simply can't,
831                  * this is not necessarily our ->mm.
832                  * Since kill_ioctx() uses non-zero ->mmap_size as indicator
833                  * that it needs to unmap the area, just set it to 0.
834                  */
835                 ctx->mmap_size = 0;
836                 kill_ioctx(mm, ctx, &requests_done);
837 
838                 /* Wait until all IO for the context are done. */
839                 wait_for_completion(&requests_done);
840         }
841 
842         RCU_INIT_POINTER(mm->ioctx_table, NULL);
843         kfree(table);
844 }
845 
846 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
847 {
848         struct kioctx_cpu *kcpu;
849         unsigned long flags;
850 
851         local_irq_save(flags);
852         kcpu = this_cpu_ptr(ctx->cpu);
853         kcpu->reqs_available += nr;
854 
855         while (kcpu->reqs_available >= ctx->req_batch * 2) {
856                 kcpu->reqs_available -= ctx->req_batch;
857                 atomic_add(ctx->req_batch, &ctx->reqs_available);
858         }
859 
860         local_irq_restore(flags);
861 }
862 
863 static bool get_reqs_available(struct kioctx *ctx)
864 {
865         struct kioctx_cpu *kcpu;
866         bool ret = false;
867         unsigned long flags;
868 
869         local_irq_save(flags);
870         kcpu = this_cpu_ptr(ctx->cpu);
871         if (!kcpu->reqs_available) {
872                 int old, avail = atomic_read(&ctx->reqs_available);
873 
874                 do {
875                         if (avail < ctx->req_batch)
876                                 goto out;
877 
878                         old = avail;
879                         avail = atomic_cmpxchg(&ctx->reqs_available,
880                                                avail, avail - ctx->req_batch);
881                 } while (avail != old);
882 
883                 kcpu->reqs_available += ctx->req_batch;
884         }
885 
886         ret = true;
887         kcpu->reqs_available--;
888 out:
889         local_irq_restore(flags);
890         return ret;
891 }
892 
893 /* refill_reqs_available
894  *      Updates the reqs_available reference counts used for tracking the
895  *      number of free slots in the completion ring.  This can be called
896  *      from aio_complete() (to optimistically update reqs_available) or
897  *      from aio_get_req() (the we're out of events case).  It must be
898  *      called holding ctx->completion_lock.
899  */
900 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
901                                   unsigned tail)
902 {
903         unsigned events_in_ring, completed;
904 
905         /* Clamp head since userland can write to it. */
906         head %= ctx->nr_events;
907         if (head <= tail)
908                 events_in_ring = tail - head;
909         else
910                 events_in_ring = ctx->nr_events - (head - tail);
911 
912         completed = ctx->completed_events;
913         if (events_in_ring < completed)
914                 completed -= events_in_ring;
915         else
916                 completed = 0;
917 
918         if (!completed)
919                 return;
920 
921         ctx->completed_events -= completed;
922         put_reqs_available(ctx, completed);
923 }
924 
925 /* user_refill_reqs_available
926  *      Called to refill reqs_available when aio_get_req() encounters an
927  *      out of space in the completion ring.
928  */
929 static void user_refill_reqs_available(struct kioctx *ctx)
930 {
931         spin_lock_irq(&ctx->completion_lock);
932         if (ctx->completed_events) {
933                 struct aio_ring *ring;
934                 unsigned head;
935 
936                 /* Access of ring->head may race with aio_read_events_ring()
937                  * here, but that's okay since whether we read the old version
938                  * or the new version, and either will be valid.  The important
939                  * part is that head cannot pass tail since we prevent
940                  * aio_complete() from updating tail by holding
941                  * ctx->completion_lock.  Even if head is invalid, the check
942                  * against ctx->completed_events below will make sure we do the
943                  * safe/right thing.
944                  */
945                 ring = kmap_atomic(ctx->ring_pages[0]);
946                 head = ring->head;
947                 kunmap_atomic(ring);
948 
949                 refill_reqs_available(ctx, head, ctx->tail);
950         }
951 
952         spin_unlock_irq(&ctx->completion_lock);
953 }
954 
955 /* aio_get_req
956  *      Allocate a slot for an aio request.
957  * Returns NULL if no requests are free.
958  */
959 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
960 {
961         struct kiocb *req;
962 
963         if (!get_reqs_available(ctx)) {
964                 user_refill_reqs_available(ctx);
965                 if (!get_reqs_available(ctx))
966                         return NULL;
967         }
968 
969         req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
970         if (unlikely(!req))
971                 goto out_put;
972 
973         percpu_ref_get(&ctx->reqs);
974 
975         req->ki_ctx = ctx;
976         return req;
977 out_put:
978         put_reqs_available(ctx, 1);
979         return NULL;
980 }
981 
982 static void kiocb_free(struct kiocb *req)
983 {
984         if (req->ki_filp)
985                 fput(req->ki_filp);
986         if (req->ki_eventfd != NULL)
987                 eventfd_ctx_put(req->ki_eventfd);
988         kmem_cache_free(kiocb_cachep, req);
989 }
990 
991 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
992 {
993         struct aio_ring __user *ring  = (void __user *)ctx_id;
994         struct mm_struct *mm = current->mm;
995         struct kioctx *ctx, *ret = NULL;
996         struct kioctx_table *table;
997         unsigned id;
998 
999         if (get_user(id, &ring->id))
1000                 return NULL;
1001 
1002         rcu_read_lock();
1003         table = rcu_dereference(mm->ioctx_table);
1004 
1005         if (!table || id >= table->nr)
1006                 goto out;
1007 
1008         ctx = table->table[id];
1009         if (ctx && ctx->user_id == ctx_id) {
1010                 percpu_ref_get(&ctx->users);
1011                 ret = ctx;
1012         }
1013 out:
1014         rcu_read_unlock();
1015         return ret;
1016 }
1017 
1018 /* aio_complete
1019  *      Called when the io request on the given iocb is complete.
1020  */
1021 void aio_complete(struct kiocb *iocb, long res, long res2)
1022 {
1023         struct kioctx   *ctx = iocb->ki_ctx;
1024         struct aio_ring *ring;
1025         struct io_event *ev_page, *event;
1026         unsigned tail, pos, head;
1027         unsigned long   flags;
1028 
1029         /*
1030          * Special case handling for sync iocbs:
1031          *  - events go directly into the iocb for fast handling
1032          *  - the sync task with the iocb in its stack holds the single iocb
1033          *    ref, no other paths have a way to get another ref
1034          *  - the sync task helpfully left a reference to itself in the iocb
1035          */
1036         if (is_sync_kiocb(iocb)) {
1037                 iocb->ki_user_data = res;
1038                 smp_wmb();
1039                 iocb->ki_ctx = ERR_PTR(-EXDEV);
1040                 wake_up_process(iocb->ki_obj.tsk);
1041                 return;
1042         }
1043 
1044         if (iocb->ki_list.next) {
1045                 unsigned long flags;
1046 
1047                 spin_lock_irqsave(&ctx->ctx_lock, flags);
1048                 list_del(&iocb->ki_list);
1049                 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1050         }
1051 
1052         /*
1053          * Add a completion event to the ring buffer. Must be done holding
1054          * ctx->completion_lock to prevent other code from messing with the tail
1055          * pointer since we might be called from irq context.
1056          */
1057         spin_lock_irqsave(&ctx->completion_lock, flags);
1058 
1059         tail = ctx->tail;
1060         pos = tail + AIO_EVENTS_OFFSET;
1061 
1062         if (++tail >= ctx->nr_events)
1063                 tail = 0;
1064 
1065         ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1066         event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1067 
1068         event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1069         event->data = iocb->ki_user_data;
1070         event->res = res;
1071         event->res2 = res2;
1072 
1073         kunmap_atomic(ev_page);
1074         flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1075 
1076         pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1077                  ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1078                  res, res2);
1079 
1080         /* after flagging the request as done, we
1081          * must never even look at it again
1082          */
1083         smp_wmb();      /* make event visible before updating tail */
1084 
1085         ctx->tail = tail;
1086 
1087         ring = kmap_atomic(ctx->ring_pages[0]);
1088         head = ring->head;
1089         ring->tail = tail;
1090         kunmap_atomic(ring);
1091         flush_dcache_page(ctx->ring_pages[0]);
1092 
1093         ctx->completed_events++;
1094         if (ctx->completed_events > 1)
1095                 refill_reqs_available(ctx, head, tail);
1096         spin_unlock_irqrestore(&ctx->completion_lock, flags);
1097 
1098         pr_debug("added to ring %p at [%u]\n", iocb, tail);
1099 
1100         /*
1101          * Check if the user asked us to deliver the result through an
1102          * eventfd. The eventfd_signal() function is safe to be called
1103          * from IRQ context.
1104          */
1105         if (iocb->ki_eventfd != NULL)
1106                 eventfd_signal(iocb->ki_eventfd, 1);
1107 
1108         /* everything turned out well, dispose of the aiocb. */
1109         kiocb_free(iocb);
1110 
1111         /*
1112          * We have to order our ring_info tail store above and test
1113          * of the wait list below outside the wait lock.  This is
1114          * like in wake_up_bit() where clearing a bit has to be
1115          * ordered with the unlocked test.
1116          */
1117         smp_mb();
1118 
1119         if (waitqueue_active(&ctx->wait))
1120                 wake_up(&ctx->wait);
1121 
1122         percpu_ref_put(&ctx->reqs);
1123 }
1124 EXPORT_SYMBOL(aio_complete);
1125 
1126 /* aio_read_events_ring
1127  *      Pull an event off of the ioctx's event ring.  Returns the number of
1128  *      events fetched
1129  */
1130 static long aio_read_events_ring(struct kioctx *ctx,
1131                                  struct io_event __user *event, long nr)
1132 {
1133         struct aio_ring *ring;
1134         unsigned head, tail, pos;
1135         long ret = 0;
1136         int copy_ret;
1137 
1138         /*
1139          * The mutex can block and wake us up and that will cause
1140          * wait_event_interruptible_hrtimeout() to schedule without sleeping
1141          * and repeat. This should be rare enough that it doesn't cause
1142          * peformance issues. See the comment in read_events() for more detail.
1143          */
1144         sched_annotate_sleep();
1145         mutex_lock(&ctx->ring_lock);
1146 
1147         /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1148         ring = kmap_atomic(ctx->ring_pages[0]);
1149         head = ring->head;
1150         tail = ring->tail;
1151         kunmap_atomic(ring);
1152 
1153         /*
1154          * Ensure that once we've read the current tail pointer, that
1155          * we also see the events that were stored up to the tail.
1156          */
1157         smp_rmb();
1158 
1159         pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1160 
1161         if (head == tail)
1162                 goto out;
1163 
1164         head %= ctx->nr_events;
1165         tail %= ctx->nr_events;
1166 
1167         while (ret < nr) {
1168                 long avail;
1169                 struct io_event *ev;
1170                 struct page *page;
1171 
1172                 avail = (head <= tail ?  tail : ctx->nr_events) - head;
1173                 if (head == tail)
1174                         break;
1175 
1176                 avail = min(avail, nr - ret);
1177                 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1178                             ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1179 
1180                 pos = head + AIO_EVENTS_OFFSET;
1181                 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1182                 pos %= AIO_EVENTS_PER_PAGE;
1183 
1184                 ev = kmap(page);
1185                 copy_ret = copy_to_user(event + ret, ev + pos,
1186                                         sizeof(*ev) * avail);
1187                 kunmap(page);
1188 
1189                 if (unlikely(copy_ret)) {
1190                         ret = -EFAULT;
1191                         goto out;
1192                 }
1193 
1194                 ret += avail;
1195                 head += avail;
1196                 head %= ctx->nr_events;
1197         }
1198 
1199         ring = kmap_atomic(ctx->ring_pages[0]);
1200         ring->head = head;
1201         kunmap_atomic(ring);
1202         flush_dcache_page(ctx->ring_pages[0]);
1203 
1204         pr_debug("%li  h%u t%u\n", ret, head, tail);
1205 out:
1206         mutex_unlock(&ctx->ring_lock);
1207 
1208         return ret;
1209 }
1210 
1211 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1212                             struct io_event __user *event, long *i)
1213 {
1214         long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1215 
1216         if (ret > 0)
1217                 *i += ret;
1218 
1219         if (unlikely(atomic_read(&ctx->dead)))
1220                 ret = -EINVAL;
1221 
1222         if (!*i)
1223                 *i = ret;
1224 
1225         return ret < 0 || *i >= min_nr;
1226 }
1227 
1228 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1229                         struct io_event __user *event,
1230                         struct timespec __user *timeout)
1231 {
1232         ktime_t until = { .tv64 = KTIME_MAX };
1233         long ret = 0;
1234 
1235         if (timeout) {
1236                 struct timespec ts;
1237 
1238                 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1239                         return -EFAULT;
1240 
1241                 until = timespec_to_ktime(ts);
1242         }
1243 
1244         /*
1245          * Note that aio_read_events() is being called as the conditional - i.e.
1246          * we're calling it after prepare_to_wait() has set task state to
1247          * TASK_INTERRUPTIBLE.
1248          *
1249          * But aio_read_events() can block, and if it blocks it's going to flip
1250          * the task state back to TASK_RUNNING.
1251          *
1252          * This should be ok, provided it doesn't flip the state back to
1253          * TASK_RUNNING and return 0 too much - that causes us to spin. That
1254          * will only happen if the mutex_lock() call blocks, and we then find
1255          * the ringbuffer empty. So in practice we should be ok, but it's
1256          * something to be aware of when touching this code.
1257          */
1258         if (until.tv64 == 0)
1259                 aio_read_events(ctx, min_nr, nr, event, &ret);
1260         else
1261                 wait_event_interruptible_hrtimeout(ctx->wait,
1262                                 aio_read_events(ctx, min_nr, nr, event, &ret),
1263                                 until);
1264 
1265         if (!ret && signal_pending(current))
1266                 ret = -EINTR;
1267 
1268         return ret;
1269 }
1270 
1271 /* sys_io_setup:
1272  *      Create an aio_context capable of receiving at least nr_events.
1273  *      ctxp must not point to an aio_context that already exists, and
1274  *      must be initialized to 0 prior to the call.  On successful
1275  *      creation of the aio_context, *ctxp is filled in with the resulting 
1276  *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1277  *      if the specified nr_events exceeds internal limits.  May fail 
1278  *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1279  *      of available events.  May fail with -ENOMEM if insufficient kernel
1280  *      resources are available.  May fail with -EFAULT if an invalid
1281  *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1282  *      implemented.
1283  */
1284 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1285 {
1286         struct kioctx *ioctx = NULL;
1287         unsigned long ctx;
1288         long ret;
1289 
1290         ret = get_user(ctx, ctxp);
1291         if (unlikely(ret))
1292                 goto out;
1293 
1294         ret = -EINVAL;
1295         if (unlikely(ctx || nr_events == 0)) {
1296                 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1297                          ctx, nr_events);
1298                 goto out;
1299         }
1300 
1301         ioctx = ioctx_alloc(nr_events);
1302         ret = PTR_ERR(ioctx);
1303         if (!IS_ERR(ioctx)) {
1304                 ret = put_user(ioctx->user_id, ctxp);
1305                 if (ret)
1306                         kill_ioctx(current->mm, ioctx, NULL);
1307                 percpu_ref_put(&ioctx->users);
1308         }
1309 
1310 out:
1311         return ret;
1312 }
1313 
1314 /* sys_io_destroy:
1315  *      Destroy the aio_context specified.  May cancel any outstanding 
1316  *      AIOs and block on completion.  Will fail with -ENOSYS if not
1317  *      implemented.  May fail with -EINVAL if the context pointed to
1318  *      is invalid.
1319  */
1320 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1321 {
1322         struct kioctx *ioctx = lookup_ioctx(ctx);
1323         if (likely(NULL != ioctx)) {
1324                 struct completion requests_done =
1325                         COMPLETION_INITIALIZER_ONSTACK(requests_done);
1326                 int ret;
1327 
1328                 /* Pass requests_done to kill_ioctx() where it can be set
1329                  * in a thread-safe way. If we try to set it here then we have
1330                  * a race condition if two io_destroy() called simultaneously.
1331                  */
1332                 ret = kill_ioctx(current->mm, ioctx, &requests_done);
1333                 percpu_ref_put(&ioctx->users);
1334 
1335                 /* Wait until all IO for the context are done. Otherwise kernel
1336                  * keep using user-space buffers even if user thinks the context
1337                  * is destroyed.
1338                  */
1339                 if (!ret)
1340                         wait_for_completion(&requests_done);
1341 
1342                 return ret;
1343         }
1344         pr_debug("EINVAL: invalid context id\n");
1345         return -EINVAL;
1346 }
1347 
1348 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1349                             unsigned long, loff_t);
1350 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1351 
1352 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1353                                      int rw, char __user *buf,
1354                                      unsigned long *nr_segs,
1355                                      struct iovec **iovec,
1356                                      bool compat)
1357 {
1358         ssize_t ret;
1359 
1360         *nr_segs = kiocb->ki_nbytes;
1361 
1362 #ifdef CONFIG_COMPAT
1363         if (compat)
1364                 ret = compat_rw_copy_check_uvector(rw,
1365                                 (struct compat_iovec __user *)buf,
1366                                 *nr_segs, UIO_FASTIOV, *iovec, iovec);
1367         else
1368 #endif
1369                 ret = rw_copy_check_uvector(rw,
1370                                 (struct iovec __user *)buf,
1371                                 *nr_segs, UIO_FASTIOV, *iovec, iovec);
1372         if (ret < 0)
1373                 return ret;
1374 
1375         /* ki_nbytes now reflect bytes instead of segs */
1376         kiocb->ki_nbytes = ret;
1377         return 0;
1378 }
1379 
1380 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1381                                        int rw, char __user *buf,
1382                                        unsigned long *nr_segs,
1383                                        struct iovec *iovec)
1384 {
1385         if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1386                 return -EFAULT;
1387 
1388         iovec->iov_base = buf;
1389         iovec->iov_len = kiocb->ki_nbytes;
1390         *nr_segs = 1;
1391         return 0;
1392 }
1393 
1394 /*
1395  * aio_run_iocb:
1396  *      Performs the initial checks and io submission.
1397  */
1398 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1399                             char __user *buf, bool compat)
1400 {
1401         struct file *file = req->ki_filp;
1402         ssize_t ret;
1403         unsigned long nr_segs;
1404         int rw;
1405         fmode_t mode;
1406         aio_rw_op *rw_op;
1407         rw_iter_op *iter_op;
1408         struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1409         struct iov_iter iter;
1410 
1411         switch (opcode) {
1412         case IOCB_CMD_PREAD:
1413         case IOCB_CMD_PREADV:
1414                 mode    = FMODE_READ;
1415                 rw      = READ;
1416                 rw_op   = file->f_op->aio_read;
1417                 iter_op = file->f_op->read_iter;
1418                 goto rw_common;
1419 
1420         case IOCB_CMD_PWRITE:
1421         case IOCB_CMD_PWRITEV:
1422                 mode    = FMODE_WRITE;
1423                 rw      = WRITE;
1424                 rw_op   = file->f_op->aio_write;
1425                 iter_op = file->f_op->write_iter;
1426                 goto rw_common;
1427 rw_common:
1428                 if (unlikely(!(file->f_mode & mode)))
1429                         return -EBADF;
1430 
1431                 if (!rw_op && !iter_op)
1432                         return -EINVAL;
1433 
1434                 ret = (opcode == IOCB_CMD_PREADV ||
1435                        opcode == IOCB_CMD_PWRITEV)
1436                         ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1437                                                 &iovec, compat)
1438                         : aio_setup_single_vector(req, rw, buf, &nr_segs,
1439                                                   iovec);
1440                 if (!ret)
1441                         ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1442                 if (ret < 0) {
1443                         if (iovec != inline_vecs)
1444                                 kfree(iovec);
1445                         return ret;
1446                 }
1447 
1448                 req->ki_nbytes = ret;
1449 
1450                 /* XXX: move/kill - rw_verify_area()? */
1451                 /* This matches the pread()/pwrite() logic */
1452                 if (req->ki_pos < 0) {
1453                         ret = -EINVAL;
1454                         break;
1455                 }
1456 
1457                 if (rw == WRITE)
1458                         file_start_write(file);
1459 
1460                 if (iter_op) {
1461                         iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes);
1462                         ret = iter_op(req, &iter);
1463                 } else {
1464                         ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1465                 }
1466 
1467                 if (rw == WRITE)
1468                         file_end_write(file);
1469                 break;
1470 
1471         case IOCB_CMD_FDSYNC:
1472                 if (!file->f_op->aio_fsync)
1473                         return -EINVAL;
1474 
1475                 ret = file->f_op->aio_fsync(req, 1);
1476                 break;
1477 
1478         case IOCB_CMD_FSYNC:
1479                 if (!file->f_op->aio_fsync)
1480                         return -EINVAL;
1481 
1482                 ret = file->f_op->aio_fsync(req, 0);
1483                 break;
1484 
1485         default:
1486                 pr_debug("EINVAL: no operation provided\n");
1487                 return -EINVAL;
1488         }
1489 
1490         if (iovec != inline_vecs)
1491                 kfree(iovec);
1492 
1493         if (ret != -EIOCBQUEUED) {
1494                 /*
1495                  * There's no easy way to restart the syscall since other AIO's
1496                  * may be already running. Just fail this IO with EINTR.
1497                  */
1498                 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1499                              ret == -ERESTARTNOHAND ||
1500                              ret == -ERESTART_RESTARTBLOCK))
1501                         ret = -EINTR;
1502                 aio_complete(req, ret, 0);
1503         }
1504 
1505         return 0;
1506 }
1507 
1508 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1509                          struct iocb *iocb, bool compat)
1510 {
1511         struct kiocb *req;
1512         ssize_t ret;
1513 
1514         /* enforce forwards compatibility on users */
1515         if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1516                 pr_debug("EINVAL: reserve field set\n");
1517                 return -EINVAL;
1518         }
1519 
1520         /* prevent overflows */
1521         if (unlikely(
1522             (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1523             (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1524             ((ssize_t)iocb->aio_nbytes < 0)
1525            )) {
1526                 pr_debug("EINVAL: overflow check\n");
1527                 return -EINVAL;
1528         }
1529 
1530         req = aio_get_req(ctx);
1531         if (unlikely(!req))
1532                 return -EAGAIN;
1533 
1534         req->ki_filp = fget(iocb->aio_fildes);
1535         if (unlikely(!req->ki_filp)) {
1536                 ret = -EBADF;
1537                 goto out_put_req;
1538         }
1539 
1540         if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1541                 /*
1542                  * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1543                  * instance of the file* now. The file descriptor must be
1544                  * an eventfd() fd, and will be signaled for each completed
1545                  * event using the eventfd_signal() function.
1546                  */
1547                 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1548                 if (IS_ERR(req->ki_eventfd)) {
1549                         ret = PTR_ERR(req->ki_eventfd);
1550                         req->ki_eventfd = NULL;
1551                         goto out_put_req;
1552                 }
1553         }
1554 
1555         ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1556         if (unlikely(ret)) {
1557                 pr_debug("EFAULT: aio_key\n");
1558                 goto out_put_req;
1559         }
1560 
1561         req->ki_obj.user = user_iocb;
1562         req->ki_user_data = iocb->aio_data;
1563         req->ki_pos = iocb->aio_offset;
1564         req->ki_nbytes = iocb->aio_nbytes;
1565 
1566         ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1567                            (char __user *)(unsigned long)iocb->aio_buf,
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 kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1655                                   u32 key)
1656 {
1657         struct list_head *pos;
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(pos, &ctx->active_reqs) {
1666                 struct kiocb *kiocb = list_kiocb(pos);
1667                 if (kiocb->ki_obj.user == iocb)
1668                         return kiocb;
1669         }
1670         return NULL;
1671 }
1672 
1673 /* sys_io_cancel:
1674  *      Attempts to cancel an iocb previously passed to io_submit.  If
1675  *      the operation is successfully cancelled, the resulting event is
1676  *      copied into the memory pointed to by result without being placed
1677  *      into the completion queue and 0 is returned.  May fail with
1678  *      -EFAULT if any of the data structures pointed to are invalid.
1679  *      May fail with -EINVAL if aio_context specified by ctx_id is
1680  *      invalid.  May fail with -EAGAIN if the iocb specified was not
1681  *      cancelled.  Will fail with -ENOSYS if not implemented.
1682  */
1683 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1684                 struct io_event __user *, result)
1685 {
1686         struct kioctx *ctx;
1687         struct kiocb *kiocb;
1688         u32 key;
1689         int ret;
1690 
1691         ret = get_user(key, &iocb->aio_key);
1692         if (unlikely(ret))
1693                 return -EFAULT;
1694 
1695         ctx = lookup_ioctx(ctx_id);
1696         if (unlikely(!ctx))
1697                 return -EINVAL;
1698 
1699         spin_lock_irq(&ctx->ctx_lock);
1700 
1701         kiocb = lookup_kiocb(ctx, iocb, key);
1702         if (kiocb)
1703                 ret = kiocb_cancel(kiocb);
1704         else
1705                 ret = -EINVAL;
1706 
1707         spin_unlock_irq(&ctx->ctx_lock);
1708 
1709         if (!ret) {
1710                 /*
1711                  * The result argument is no longer used - the io_event is
1712                  * always delivered via the ring buffer. -EINPROGRESS indicates
1713                  * cancellation is progress:
1714                  */
1715                 ret = -EINPROGRESS;
1716         }
1717 
1718         percpu_ref_put(&ctx->users);
1719 
1720         return ret;
1721 }
1722 
1723 /* io_getevents:
1724  *      Attempts to read at least min_nr events and up to nr events from
1725  *      the completion queue for the aio_context specified by ctx_id. If
1726  *      it succeeds, the number of read events is returned. May fail with
1727  *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1728  *      out of range, if timeout is out of range.  May fail with -EFAULT
1729  *      if any of the memory specified is invalid.  May return 0 or
1730  *      < min_nr if the timeout specified by timeout has elapsed
1731  *      before sufficient events are available, where timeout == NULL
1732  *      specifies an infinite timeout. Note that the timeout pointed to by
1733  *      timeout is relative.  Will fail with -ENOSYS if not implemented.
1734  */
1735 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1736                 long, min_nr,
1737                 long, nr,
1738                 struct io_event __user *, events,
1739                 struct timespec __user *, timeout)
1740 {
1741         struct kioctx *ioctx = lookup_ioctx(ctx_id);
1742         long ret = -EINVAL;
1743 
1744         if (likely(ioctx)) {
1745                 if (likely(min_nr <= nr && min_nr >= 0))
1746                         ret = read_events(ioctx, min_nr, nr, events, timeout);
1747                 percpu_ref_put(&ioctx->users);
1748         }
1749         return ret;
1750 }
1751 

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