Version:  2.0.40 2.2.26 2.4.37 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19

Linux/fs/aio.c

  1 /*
  2  *      An async IO implementation for Linux
  3  *      Written by Benjamin LaHaise <bcrl@kvack.org>
  4  *
  5  *      Implements an efficient asynchronous io interface.
  6  *
  7  *      Copyright 2000, 2001, 2002 Red Hat, Inc.  All Rights Reserved.
  8  *
  9  *      See ../COPYING for licensing terms.
 10  */
 11 #define pr_fmt(fmt) "%s: " fmt, __func__
 12 
 13 #include <linux/kernel.h>
 14 #include <linux/init.h>
 15 #include <linux/errno.h>
 16 #include <linux/time.h>
 17 #include <linux/aio_abi.h>
 18 #include <linux/export.h>
 19 #include <linux/syscalls.h>
 20 #include <linux/backing-dev.h>
 21 #include <linux/uio.h>
 22 
 23 #include <linux/sched.h>
 24 #include <linux/fs.h>
 25 #include <linux/file.h>
 26 #include <linux/mm.h>
 27 #include <linux/mman.h>
 28 #include <linux/mmu_context.h>
 29 #include <linux/percpu.h>
 30 #include <linux/slab.h>
 31 #include <linux/timer.h>
 32 #include <linux/aio.h>
 33 #include <linux/highmem.h>
 34 #include <linux/workqueue.h>
 35 #include <linux/security.h>
 36 #include <linux/eventfd.h>
 37 #include <linux/blkdev.h>
 38 #include <linux/compat.h>
 39 #include <linux/migrate.h>
 40 #include <linux/ramfs.h>
 41 #include <linux/percpu-refcount.h>
 42 #include <linux/mount.h>
 43 
 44 #include <asm/kmap_types.h>
 45 #include <asm/uaccess.h>
 46 
 47 #include "internal.h"
 48 
 49 #define AIO_RING_MAGIC                  0xa10a10a1
 50 #define AIO_RING_COMPAT_FEATURES        1
 51 #define AIO_RING_INCOMPAT_FEATURES      0
 52 struct aio_ring {
 53         unsigned        id;     /* kernel internal index number */
 54         unsigned        nr;     /* number of io_events */
 55         unsigned        head;   /* Written to by userland or under ring_lock
 56                                  * mutex by aio_read_events_ring(). */
 57         unsigned        tail;
 58 
 59         unsigned        magic;
 60         unsigned        compat_features;
 61         unsigned        incompat_features;
 62         unsigned        header_length;  /* size of aio_ring */
 63 
 64 
 65         struct io_event         io_events[0];
 66 }; /* 128 bytes + ring size */
 67 
 68 #define AIO_RING_PAGES  8
 69 
 70 struct kioctx_table {
 71         struct rcu_head rcu;
 72         unsigned        nr;
 73         struct kioctx   *table[];
 74 };
 75 
 76 struct kioctx_cpu {
 77         unsigned                reqs_available;
 78 };
 79 
 80 struct kioctx {
 81         struct percpu_ref       users;
 82         atomic_t                dead;
 83 
 84         struct percpu_ref       reqs;
 85 
 86         unsigned long           user_id;
 87 
 88         struct __percpu kioctx_cpu *cpu;
 89 
 90         /*
 91          * For percpu reqs_available, number of slots we move to/from global
 92          * counter at a time:
 93          */
 94         unsigned                req_batch;
 95         /*
 96          * This is what userspace passed to io_setup(), it's not used for
 97          * anything but counting against the global max_reqs quota.
 98          *
 99          * The real limit is nr_events - 1, which will be larger (see
100          * aio_setup_ring())
101          */
102         unsigned                max_reqs;
103 
104         /* Size of ringbuffer, in units of struct io_event */
105         unsigned                nr_events;
106 
107         unsigned long           mmap_base;
108         unsigned long           mmap_size;
109 
110         struct page             **ring_pages;
111         long                    nr_pages;
112 
113         struct work_struct      free_work;
114 
115         /*
116          * signals when all in-flight requests are done
117          */
118         struct completion *requests_done;
119 
120         struct {
121                 /*
122                  * This counts the number of available slots in the ringbuffer,
123                  * so we avoid overflowing it: it's decremented (if positive)
124                  * when allocating a kiocb and incremented when the resulting
125                  * io_event is pulled off the ringbuffer.
126                  *
127                  * We batch accesses to it with a percpu version.
128                  */
129                 atomic_t        reqs_available;
130         } ____cacheline_aligned_in_smp;
131 
132         struct {
133                 spinlock_t      ctx_lock;
134                 struct list_head active_reqs;   /* used for cancellation */
135         } ____cacheline_aligned_in_smp;
136 
137         struct {
138                 struct mutex    ring_lock;
139                 wait_queue_head_t wait;
140         } ____cacheline_aligned_in_smp;
141 
142         struct {
143                 unsigned        tail;
144                 unsigned        completed_events;
145                 spinlock_t      completion_lock;
146         } ____cacheline_aligned_in_smp;
147 
148         struct page             *internal_pages[AIO_RING_PAGES];
149         struct file             *aio_ring_file;
150 
151         unsigned                id;
152 };
153 
154 /*------ sysctl variables----*/
155 static DEFINE_SPINLOCK(aio_nr_lock);
156 unsigned long aio_nr;           /* current system wide number of aio requests */
157 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
158 /*----end sysctl variables---*/
159 
160 static struct kmem_cache        *kiocb_cachep;
161 static struct kmem_cache        *kioctx_cachep;
162 
163 static struct vfsmount *aio_mnt;
164 
165 static const struct file_operations aio_ring_fops;
166 static const struct address_space_operations aio_ctx_aops;
167 
168 /* Backing dev info for aio fs.
169  * -no dirty page accounting or writeback happens
170  */
171 static struct backing_dev_info aio_fs_backing_dev_info = {
172         .name           = "aiofs",
173         .state          = 0,
174         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_MAP_COPY,
175 };
176 
177 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
178 {
179         struct qstr this = QSTR_INIT("[aio]", 5);
180         struct file *file;
181         struct path path;
182         struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
183         if (IS_ERR(inode))
184                 return ERR_CAST(inode);
185 
186         inode->i_mapping->a_ops = &aio_ctx_aops;
187         inode->i_mapping->private_data = ctx;
188         inode->i_mapping->backing_dev_info = &aio_fs_backing_dev_info;
189         inode->i_size = PAGE_SIZE * nr_pages;
190 
191         path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
192         if (!path.dentry) {
193                 iput(inode);
194                 return ERR_PTR(-ENOMEM);
195         }
196         path.mnt = mntget(aio_mnt);
197 
198         d_instantiate(path.dentry, inode);
199         file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
200         if (IS_ERR(file)) {
201                 path_put(&path);
202                 return file;
203         }
204 
205         file->f_flags = O_RDWR;
206         return file;
207 }
208 
209 static struct dentry *aio_mount(struct file_system_type *fs_type,
210                                 int flags, const char *dev_name, void *data)
211 {
212         static const struct dentry_operations ops = {
213                 .d_dname        = simple_dname,
214         };
215         return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
216 }
217 
218 /* aio_setup
219  *      Creates the slab caches used by the aio routines, panic on
220  *      failure as this is done early during the boot sequence.
221  */
222 static int __init aio_setup(void)
223 {
224         static struct file_system_type aio_fs = {
225                 .name           = "aio",
226                 .mount          = aio_mount,
227                 .kill_sb        = kill_anon_super,
228         };
229         aio_mnt = kern_mount(&aio_fs);
230         if (IS_ERR(aio_mnt))
231                 panic("Failed to create aio fs mount.");
232 
233         if (bdi_init(&aio_fs_backing_dev_info))
234                 panic("Failed to init aio fs backing dev info.");
235 
236         kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
237         kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
238 
239         pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
240 
241         return 0;
242 }
243 __initcall(aio_setup);
244 
245 static void put_aio_ring_file(struct kioctx *ctx)
246 {
247         struct file *aio_ring_file = ctx->aio_ring_file;
248         if (aio_ring_file) {
249                 truncate_setsize(aio_ring_file->f_inode, 0);
250 
251                 /* Prevent further access to the kioctx from migratepages */
252                 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
253                 aio_ring_file->f_inode->i_mapping->private_data = NULL;
254                 ctx->aio_ring_file = NULL;
255                 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
256 
257                 fput(aio_ring_file);
258         }
259 }
260 
261 static void aio_free_ring(struct kioctx *ctx)
262 {
263         int i;
264 
265         /* Disconnect the kiotx from the ring file.  This prevents future
266          * accesses to the kioctx from page migration.
267          */
268         put_aio_ring_file(ctx);
269 
270         for (i = 0; i < ctx->nr_pages; i++) {
271                 struct page *page;
272                 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
273                                 page_count(ctx->ring_pages[i]));
274                 page = ctx->ring_pages[i];
275                 if (!page)
276                         continue;
277                 ctx->ring_pages[i] = NULL;
278                 put_page(page);
279         }
280 
281         if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
282                 kfree(ctx->ring_pages);
283                 ctx->ring_pages = NULL;
284         }
285 }
286 
287 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
288 {
289         vma->vm_flags |= VM_DONTEXPAND;
290         vma->vm_ops = &generic_file_vm_ops;
291         return 0;
292 }
293 
294 static void aio_ring_remap(struct file *file, struct vm_area_struct *vma)
295 {
296         struct mm_struct *mm = vma->vm_mm;
297         struct kioctx_table *table;
298         int i;
299 
300         spin_lock(&mm->ioctx_lock);
301         rcu_read_lock();
302         table = rcu_dereference(mm->ioctx_table);
303         for (i = 0; i < table->nr; i++) {
304                 struct kioctx *ctx;
305 
306                 ctx = table->table[i];
307                 if (ctx && ctx->aio_ring_file == file) {
308                         ctx->user_id = ctx->mmap_base = vma->vm_start;
309                         break;
310                 }
311         }
312 
313         rcu_read_unlock();
314         spin_unlock(&mm->ioctx_lock);
315 }
316 
317 static const struct file_operations aio_ring_fops = {
318         .mmap = aio_ring_mmap,
319         .mremap = aio_ring_remap,
320 };
321 
322 #if IS_ENABLED(CONFIG_MIGRATION)
323 static int aio_migratepage(struct address_space *mapping, struct page *new,
324                         struct page *old, enum migrate_mode mode)
325 {
326         struct kioctx *ctx;
327         unsigned long flags;
328         pgoff_t idx;
329         int rc;
330 
331         rc = 0;
332 
333         /* mapping->private_lock here protects against the kioctx teardown.  */
334         spin_lock(&mapping->private_lock);
335         ctx = mapping->private_data;
336         if (!ctx) {
337                 rc = -EINVAL;
338                 goto out;
339         }
340 
341         /* The ring_lock mutex.  The prevents aio_read_events() from writing
342          * to the ring's head, and prevents page migration from mucking in
343          * a partially initialized kiotx.
344          */
345         if (!mutex_trylock(&ctx->ring_lock)) {
346                 rc = -EAGAIN;
347                 goto out;
348         }
349 
350         idx = old->index;
351         if (idx < (pgoff_t)ctx->nr_pages) {
352                 /* Make sure the old page hasn't already been changed */
353                 if (ctx->ring_pages[idx] != old)
354                         rc = -EAGAIN;
355         } else
356                 rc = -EINVAL;
357 
358         if (rc != 0)
359                 goto out_unlock;
360 
361         /* Writeback must be complete */
362         BUG_ON(PageWriteback(old));
363         get_page(new);
364 
365         rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
366         if (rc != MIGRATEPAGE_SUCCESS) {
367                 put_page(new);
368                 goto out_unlock;
369         }
370 
371         /* Take completion_lock to prevent other writes to the ring buffer
372          * while the old page is copied to the new.  This prevents new
373          * events from being lost.
374          */
375         spin_lock_irqsave(&ctx->completion_lock, flags);
376         migrate_page_copy(new, old);
377         BUG_ON(ctx->ring_pages[idx] != old);
378         ctx->ring_pages[idx] = new;
379         spin_unlock_irqrestore(&ctx->completion_lock, flags);
380 
381         /* The old page is no longer accessible. */
382         put_page(old);
383 
384 out_unlock:
385         mutex_unlock(&ctx->ring_lock);
386 out:
387         spin_unlock(&mapping->private_lock);
388         return rc;
389 }
390 #endif
391 
392 static const struct address_space_operations aio_ctx_aops = {
393         .set_page_dirty = __set_page_dirty_no_writeback,
394 #if IS_ENABLED(CONFIG_MIGRATION)
395         .migratepage    = aio_migratepage,
396 #endif
397 };
398 
399 static int aio_setup_ring(struct kioctx *ctx)
400 {
401         struct aio_ring *ring;
402         unsigned nr_events = ctx->max_reqs;
403         struct mm_struct *mm = current->mm;
404         unsigned long size, unused;
405         int nr_pages;
406         int i;
407         struct file *file;
408 
409         /* Compensate for the ring buffer's head/tail overlap entry */
410         nr_events += 2; /* 1 is required, 2 for good luck */
411 
412         size = sizeof(struct aio_ring);
413         size += sizeof(struct io_event) * nr_events;
414 
415         nr_pages = PFN_UP(size);
416         if (nr_pages < 0)
417                 return -EINVAL;
418 
419         file = aio_private_file(ctx, nr_pages);
420         if (IS_ERR(file)) {
421                 ctx->aio_ring_file = NULL;
422                 return -ENOMEM;
423         }
424 
425         ctx->aio_ring_file = file;
426         nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
427                         / sizeof(struct io_event);
428 
429         ctx->ring_pages = ctx->internal_pages;
430         if (nr_pages > AIO_RING_PAGES) {
431                 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
432                                           GFP_KERNEL);
433                 if (!ctx->ring_pages) {
434                         put_aio_ring_file(ctx);
435                         return -ENOMEM;
436                 }
437         }
438 
439         for (i = 0; i < nr_pages; i++) {
440                 struct page *page;
441                 page = find_or_create_page(file->f_inode->i_mapping,
442                                            i, GFP_HIGHUSER | __GFP_ZERO);
443                 if (!page)
444                         break;
445                 pr_debug("pid(%d) page[%d]->count=%d\n",
446                          current->pid, i, page_count(page));
447                 SetPageUptodate(page);
448                 unlock_page(page);
449 
450                 ctx->ring_pages[i] = page;
451         }
452         ctx->nr_pages = i;
453 
454         if (unlikely(i != nr_pages)) {
455                 aio_free_ring(ctx);
456                 return -ENOMEM;
457         }
458 
459         ctx->mmap_size = nr_pages * PAGE_SIZE;
460         pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
461 
462         down_write(&mm->mmap_sem);
463         ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
464                                        PROT_READ | PROT_WRITE,
465                                        MAP_SHARED, 0, &unused);
466         up_write(&mm->mmap_sem);
467         if (IS_ERR((void *)ctx->mmap_base)) {
468                 ctx->mmap_size = 0;
469                 aio_free_ring(ctx);
470                 return -ENOMEM;
471         }
472 
473         pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
474 
475         ctx->user_id = ctx->mmap_base;
476         ctx->nr_events = nr_events; /* trusted copy */
477 
478         ring = kmap_atomic(ctx->ring_pages[0]);
479         ring->nr = nr_events;   /* user copy */
480         ring->id = ~0U;
481         ring->head = ring->tail = 0;
482         ring->magic = AIO_RING_MAGIC;
483         ring->compat_features = AIO_RING_COMPAT_FEATURES;
484         ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
485         ring->header_length = sizeof(struct aio_ring);
486         kunmap_atomic(ring);
487         flush_dcache_page(ctx->ring_pages[0]);
488 
489         return 0;
490 }
491 
492 #define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
493 #define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
494 #define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
495 
496 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
497 {
498         struct kioctx *ctx = req->ki_ctx;
499         unsigned long flags;
500 
501         spin_lock_irqsave(&ctx->ctx_lock, flags);
502 
503         if (!req->ki_list.next)
504                 list_add(&req->ki_list, &ctx->active_reqs);
505 
506         req->ki_cancel = cancel;
507 
508         spin_unlock_irqrestore(&ctx->ctx_lock, flags);
509 }
510 EXPORT_SYMBOL(kiocb_set_cancel_fn);
511 
512 static int kiocb_cancel(struct kiocb *kiocb)
513 {
514         kiocb_cancel_fn *old, *cancel;
515 
516         /*
517          * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
518          * actually has a cancel function, hence the cmpxchg()
519          */
520 
521         cancel = ACCESS_ONCE(kiocb->ki_cancel);
522         do {
523                 if (!cancel || cancel == KIOCB_CANCELLED)
524                         return -EINVAL;
525 
526                 old = cancel;
527                 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
528         } while (cancel != old);
529 
530         return cancel(kiocb);
531 }
532 
533 static void free_ioctx(struct work_struct *work)
534 {
535         struct kioctx *ctx = container_of(work, struct kioctx, free_work);
536 
537         pr_debug("freeing %p\n", ctx);
538 
539         aio_free_ring(ctx);
540         free_percpu(ctx->cpu);
541         percpu_ref_exit(&ctx->reqs);
542         percpu_ref_exit(&ctx->users);
543         kmem_cache_free(kioctx_cachep, ctx);
544 }
545 
546 static void free_ioctx_reqs(struct percpu_ref *ref)
547 {
548         struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
549 
550         /* At this point we know that there are no any in-flight requests */
551         if (ctx->requests_done)
552                 complete(ctx->requests_done);
553 
554         INIT_WORK(&ctx->free_work, free_ioctx);
555         schedule_work(&ctx->free_work);
556 }
557 
558 /*
559  * When this function runs, the kioctx has been removed from the "hash table"
560  * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
561  * now it's safe to cancel any that need to be.
562  */
563 static void free_ioctx_users(struct percpu_ref *ref)
564 {
565         struct kioctx *ctx = container_of(ref, struct kioctx, users);
566         struct kiocb *req;
567 
568         spin_lock_irq(&ctx->ctx_lock);
569 
570         while (!list_empty(&ctx->active_reqs)) {
571                 req = list_first_entry(&ctx->active_reqs,
572                                        struct kiocb, ki_list);
573 
574                 list_del_init(&req->ki_list);
575                 kiocb_cancel(req);
576         }
577 
578         spin_unlock_irq(&ctx->ctx_lock);
579 
580         percpu_ref_kill(&ctx->reqs);
581         percpu_ref_put(&ctx->reqs);
582 }
583 
584 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
585 {
586         unsigned i, new_nr;
587         struct kioctx_table *table, *old;
588         struct aio_ring *ring;
589 
590         spin_lock(&mm->ioctx_lock);
591         table = rcu_dereference_raw(mm->ioctx_table);
592 
593         while (1) {
594                 if (table)
595                         for (i = 0; i < table->nr; i++)
596                                 if (!table->table[i]) {
597                                         ctx->id = i;
598                                         table->table[i] = ctx;
599                                         spin_unlock(&mm->ioctx_lock);
600 
601                                         /* While kioctx setup is in progress,
602                                          * we are protected from page migration
603                                          * changes ring_pages by ->ring_lock.
604                                          */
605                                         ring = kmap_atomic(ctx->ring_pages[0]);
606                                         ring->id = ctx->id;
607                                         kunmap_atomic(ring);
608                                         return 0;
609                                 }
610 
611                 new_nr = (table ? table->nr : 1) * 4;
612                 spin_unlock(&mm->ioctx_lock);
613 
614                 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
615                                 new_nr, GFP_KERNEL);
616                 if (!table)
617                         return -ENOMEM;
618 
619                 table->nr = new_nr;
620 
621                 spin_lock(&mm->ioctx_lock);
622                 old = rcu_dereference_raw(mm->ioctx_table);
623 
624                 if (!old) {
625                         rcu_assign_pointer(mm->ioctx_table, table);
626                 } else if (table->nr > old->nr) {
627                         memcpy(table->table, old->table,
628                                old->nr * sizeof(struct kioctx *));
629 
630                         rcu_assign_pointer(mm->ioctx_table, table);
631                         kfree_rcu(old, rcu);
632                 } else {
633                         kfree(table);
634                         table = old;
635                 }
636         }
637 }
638 
639 static void aio_nr_sub(unsigned nr)
640 {
641         spin_lock(&aio_nr_lock);
642         if (WARN_ON(aio_nr - nr > aio_nr))
643                 aio_nr = 0;
644         else
645                 aio_nr -= nr;
646         spin_unlock(&aio_nr_lock);
647 }
648 
649 /* ioctx_alloc
650  *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
651  */
652 static struct kioctx *ioctx_alloc(unsigned nr_events)
653 {
654         struct mm_struct *mm = current->mm;
655         struct kioctx *ctx;
656         int err = -ENOMEM;
657 
658         /*
659          * We keep track of the number of available ringbuffer slots, to prevent
660          * overflow (reqs_available), and we also use percpu counters for this.
661          *
662          * So since up to half the slots might be on other cpu's percpu counters
663          * and unavailable, double nr_events so userspace sees what they
664          * expected: additionally, we move req_batch slots to/from percpu
665          * counters at a time, so make sure that isn't 0:
666          */
667         nr_events = max(nr_events, num_possible_cpus() * 4);
668         nr_events *= 2;
669 
670         /* Prevent overflows */
671         if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
672             (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
673                 pr_debug("ENOMEM: nr_events too high\n");
674                 return ERR_PTR(-EINVAL);
675         }
676 
677         if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
678                 return ERR_PTR(-EAGAIN);
679 
680         ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
681         if (!ctx)
682                 return ERR_PTR(-ENOMEM);
683 
684         ctx->max_reqs = nr_events;
685 
686         spin_lock_init(&ctx->ctx_lock);
687         spin_lock_init(&ctx->completion_lock);
688         mutex_init(&ctx->ring_lock);
689         /* Protect against page migration throughout kiotx setup by keeping
690          * the ring_lock mutex held until setup is complete. */
691         mutex_lock(&ctx->ring_lock);
692         init_waitqueue_head(&ctx->wait);
693 
694         INIT_LIST_HEAD(&ctx->active_reqs);
695 
696         if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
697                 goto err;
698 
699         if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
700                 goto err;
701 
702         ctx->cpu = alloc_percpu(struct kioctx_cpu);
703         if (!ctx->cpu)
704                 goto err;
705 
706         err = aio_setup_ring(ctx);
707         if (err < 0)
708                 goto err;
709 
710         atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
711         ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
712         if (ctx->req_batch < 1)
713                 ctx->req_batch = 1;
714 
715         /* limit the number of system wide aios */
716         spin_lock(&aio_nr_lock);
717         if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
718             aio_nr + nr_events < aio_nr) {
719                 spin_unlock(&aio_nr_lock);
720                 err = -EAGAIN;
721                 goto err_ctx;
722         }
723         aio_nr += ctx->max_reqs;
724         spin_unlock(&aio_nr_lock);
725 
726         percpu_ref_get(&ctx->users);    /* io_setup() will drop this ref */
727         percpu_ref_get(&ctx->reqs);     /* free_ioctx_users() will drop this */
728 
729         err = ioctx_add_table(ctx, mm);
730         if (err)
731                 goto err_cleanup;
732 
733         /* Release the ring_lock mutex now that all setup is complete. */
734         mutex_unlock(&ctx->ring_lock);
735 
736         pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
737                  ctx, ctx->user_id, mm, ctx->nr_events);
738         return ctx;
739 
740 err_cleanup:
741         aio_nr_sub(ctx->max_reqs);
742 err_ctx:
743         aio_free_ring(ctx);
744 err:
745         mutex_unlock(&ctx->ring_lock);
746         free_percpu(ctx->cpu);
747         percpu_ref_exit(&ctx->reqs);
748         percpu_ref_exit(&ctx->users);
749         kmem_cache_free(kioctx_cachep, ctx);
750         pr_debug("error allocating ioctx %d\n", err);
751         return ERR_PTR(err);
752 }
753 
754 /* kill_ioctx
755  *      Cancels all outstanding aio requests on an aio context.  Used
756  *      when the processes owning a context have all exited to encourage
757  *      the rapid destruction of the kioctx.
758  */
759 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
760                 struct completion *requests_done)
761 {
762         struct kioctx_table *table;
763 
764         if (atomic_xchg(&ctx->dead, 1))
765                 return -EINVAL;
766 
767 
768         spin_lock(&mm->ioctx_lock);
769         table = rcu_dereference_raw(mm->ioctx_table);
770         WARN_ON(ctx != table->table[ctx->id]);
771         table->table[ctx->id] = NULL;
772         spin_unlock(&mm->ioctx_lock);
773 
774         /* percpu_ref_kill() will do the necessary call_rcu() */
775         wake_up_all(&ctx->wait);
776 
777         /*
778          * It'd be more correct to do this in free_ioctx(), after all
779          * the outstanding kiocbs have finished - but by then io_destroy
780          * has already returned, so io_setup() could potentially return
781          * -EAGAIN with no ioctxs actually in use (as far as userspace
782          *  could tell).
783          */
784         aio_nr_sub(ctx->max_reqs);
785 
786         if (ctx->mmap_size)
787                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
788 
789         ctx->requests_done = requests_done;
790         percpu_ref_kill(&ctx->users);
791         return 0;
792 }
793 
794 /* wait_on_sync_kiocb:
795  *      Waits on the given sync kiocb to complete.
796  */
797 ssize_t wait_on_sync_kiocb(struct kiocb *req)
798 {
799         while (!req->ki_ctx) {
800                 set_current_state(TASK_UNINTERRUPTIBLE);
801                 if (req->ki_ctx)
802                         break;
803                 io_schedule();
804         }
805         __set_current_state(TASK_RUNNING);
806         return req->ki_user_data;
807 }
808 EXPORT_SYMBOL(wait_on_sync_kiocb);
809 
810 /*
811  * exit_aio: called when the last user of mm goes away.  At this point, there is
812  * no way for any new requests to be submited or any of the io_* syscalls to be
813  * called on the context.
814  *
815  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
816  * them.
817  */
818 void exit_aio(struct mm_struct *mm)
819 {
820         struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
821         int i;
822 
823         if (!table)
824                 return;
825 
826         for (i = 0; i < table->nr; ++i) {
827                 struct kioctx *ctx = table->table[i];
828                 struct completion requests_done =
829                         COMPLETION_INITIALIZER_ONSTACK(requests_done);
830 
831                 if (!ctx)
832                         continue;
833                 /*
834                  * We don't need to bother with munmap() here - exit_mmap(mm)
835                  * is coming and it'll unmap everything. And we simply can't,
836                  * this is not necessarily our ->mm.
837                  * Since kill_ioctx() uses non-zero ->mmap_size as indicator
838                  * that it needs to unmap the area, just set it to 0.
839                  */
840                 ctx->mmap_size = 0;
841                 kill_ioctx(mm, ctx, &requests_done);
842 
843                 /* Wait until all IO for the context are done. */
844                 wait_for_completion(&requests_done);
845         }
846 
847         RCU_INIT_POINTER(mm->ioctx_table, NULL);
848         kfree(table);
849 }
850 
851 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
852 {
853         struct kioctx_cpu *kcpu;
854         unsigned long flags;
855 
856         local_irq_save(flags);
857         kcpu = this_cpu_ptr(ctx->cpu);
858         kcpu->reqs_available += nr;
859 
860         while (kcpu->reqs_available >= ctx->req_batch * 2) {
861                 kcpu->reqs_available -= ctx->req_batch;
862                 atomic_add(ctx->req_batch, &ctx->reqs_available);
863         }
864 
865         local_irq_restore(flags);
866 }
867 
868 static bool get_reqs_available(struct kioctx *ctx)
869 {
870         struct kioctx_cpu *kcpu;
871         bool ret = false;
872         unsigned long flags;
873 
874         local_irq_save(flags);
875         kcpu = this_cpu_ptr(ctx->cpu);
876         if (!kcpu->reqs_available) {
877                 int old, avail = atomic_read(&ctx->reqs_available);
878 
879                 do {
880                         if (avail < ctx->req_batch)
881                                 goto out;
882 
883                         old = avail;
884                         avail = atomic_cmpxchg(&ctx->reqs_available,
885                                                avail, avail - ctx->req_batch);
886                 } while (avail != old);
887 
888                 kcpu->reqs_available += ctx->req_batch;
889         }
890 
891         ret = true;
892         kcpu->reqs_available--;
893 out:
894         local_irq_restore(flags);
895         return ret;
896 }
897 
898 /* refill_reqs_available
899  *      Updates the reqs_available reference counts used for tracking the
900  *      number of free slots in the completion ring.  This can be called
901  *      from aio_complete() (to optimistically update reqs_available) or
902  *      from aio_get_req() (the we're out of events case).  It must be
903  *      called holding ctx->completion_lock.
904  */
905 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
906                                   unsigned tail)
907 {
908         unsigned events_in_ring, completed;
909 
910         /* Clamp head since userland can write to it. */
911         head %= ctx->nr_events;
912         if (head <= tail)
913                 events_in_ring = tail - head;
914         else
915                 events_in_ring = ctx->nr_events - (head - tail);
916 
917         completed = ctx->completed_events;
918         if (events_in_ring < completed)
919                 completed -= events_in_ring;
920         else
921                 completed = 0;
922 
923         if (!completed)
924                 return;
925 
926         ctx->completed_events -= completed;
927         put_reqs_available(ctx, completed);
928 }
929 
930 /* user_refill_reqs_available
931  *      Called to refill reqs_available when aio_get_req() encounters an
932  *      out of space in the completion ring.
933  */
934 static void user_refill_reqs_available(struct kioctx *ctx)
935 {
936         spin_lock_irq(&ctx->completion_lock);
937         if (ctx->completed_events) {
938                 struct aio_ring *ring;
939                 unsigned head;
940 
941                 /* Access of ring->head may race with aio_read_events_ring()
942                  * here, but that's okay since whether we read the old version
943                  * or the new version, and either will be valid.  The important
944                  * part is that head cannot pass tail since we prevent
945                  * aio_complete() from updating tail by holding
946                  * ctx->completion_lock.  Even if head is invalid, the check
947                  * against ctx->completed_events below will make sure we do the
948                  * safe/right thing.
949                  */
950                 ring = kmap_atomic(ctx->ring_pages[0]);
951                 head = ring->head;
952                 kunmap_atomic(ring);
953 
954                 refill_reqs_available(ctx, head, ctx->tail);
955         }
956 
957         spin_unlock_irq(&ctx->completion_lock);
958 }
959 
960 /* aio_get_req
961  *      Allocate a slot for an aio request.
962  * Returns NULL if no requests are free.
963  */
964 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
965 {
966         struct kiocb *req;
967 
968         if (!get_reqs_available(ctx)) {
969                 user_refill_reqs_available(ctx);
970                 if (!get_reqs_available(ctx))
971                         return NULL;
972         }
973 
974         req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
975         if (unlikely(!req))
976                 goto out_put;
977 
978         percpu_ref_get(&ctx->reqs);
979 
980         req->ki_ctx = ctx;
981         return req;
982 out_put:
983         put_reqs_available(ctx, 1);
984         return NULL;
985 }
986 
987 static void kiocb_free(struct kiocb *req)
988 {
989         if (req->ki_filp)
990                 fput(req->ki_filp);
991         if (req->ki_eventfd != NULL)
992                 eventfd_ctx_put(req->ki_eventfd);
993         kmem_cache_free(kiocb_cachep, req);
994 }
995 
996 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
997 {
998         struct aio_ring __user *ring  = (void __user *)ctx_id;
999         struct mm_struct *mm = current->mm;
1000         struct kioctx *ctx, *ret = NULL;
1001         struct kioctx_table *table;
1002         unsigned id;
1003 
1004         if (get_user(id, &ring->id))
1005                 return NULL;
1006 
1007         rcu_read_lock();
1008         table = rcu_dereference(mm->ioctx_table);
1009 
1010         if (!table || id >= table->nr)
1011                 goto out;
1012 
1013         ctx = table->table[id];
1014         if (ctx && ctx->user_id == ctx_id) {
1015                 percpu_ref_get(&ctx->users);
1016                 ret = ctx;
1017         }
1018 out:
1019         rcu_read_unlock();
1020         return ret;
1021 }
1022 
1023 /* aio_complete
1024  *      Called when the io request on the given iocb is complete.
1025  */
1026 void aio_complete(struct kiocb *iocb, long res, long res2)
1027 {
1028         struct kioctx   *ctx = iocb->ki_ctx;
1029         struct aio_ring *ring;
1030         struct io_event *ev_page, *event;
1031         unsigned tail, pos, head;
1032         unsigned long   flags;
1033 
1034         /*
1035          * Special case handling for sync iocbs:
1036          *  - events go directly into the iocb for fast handling
1037          *  - the sync task with the iocb in its stack holds the single iocb
1038          *    ref, no other paths have a way to get another ref
1039          *  - the sync task helpfully left a reference to itself in the iocb
1040          */
1041         if (is_sync_kiocb(iocb)) {
1042                 iocb->ki_user_data = res;
1043                 smp_wmb();
1044                 iocb->ki_ctx = ERR_PTR(-EXDEV);
1045                 wake_up_process(iocb->ki_obj.tsk);
1046                 return;
1047         }
1048 
1049         if (iocb->ki_list.next) {
1050                 unsigned long flags;
1051 
1052                 spin_lock_irqsave(&ctx->ctx_lock, flags);
1053                 list_del(&iocb->ki_list);
1054                 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1055         }
1056 
1057         /*
1058          * Add a completion event to the ring buffer. Must be done holding
1059          * ctx->completion_lock to prevent other code from messing with the tail
1060          * pointer since we might be called from irq context.
1061          */
1062         spin_lock_irqsave(&ctx->completion_lock, flags);
1063 
1064         tail = ctx->tail;
1065         pos = tail + AIO_EVENTS_OFFSET;
1066 
1067         if (++tail >= ctx->nr_events)
1068                 tail = 0;
1069 
1070         ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1071         event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1072 
1073         event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1074         event->data = iocb->ki_user_data;
1075         event->res = res;
1076         event->res2 = res2;
1077 
1078         kunmap_atomic(ev_page);
1079         flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1080 
1081         pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1082                  ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1083                  res, res2);
1084 
1085         /* after flagging the request as done, we
1086          * must never even look at it again
1087          */
1088         smp_wmb();      /* make event visible before updating tail */
1089 
1090         ctx->tail = tail;
1091 
1092         ring = kmap_atomic(ctx->ring_pages[0]);
1093         head = ring->head;
1094         ring->tail = tail;
1095         kunmap_atomic(ring);
1096         flush_dcache_page(ctx->ring_pages[0]);
1097 
1098         ctx->completed_events++;
1099         if (ctx->completed_events > 1)
1100                 refill_reqs_available(ctx, head, tail);
1101         spin_unlock_irqrestore(&ctx->completion_lock, flags);
1102 
1103         pr_debug("added to ring %p at [%u]\n", iocb, tail);
1104 
1105         /*
1106          * Check if the user asked us to deliver the result through an
1107          * eventfd. The eventfd_signal() function is safe to be called
1108          * from IRQ context.
1109          */
1110         if (iocb->ki_eventfd != NULL)
1111                 eventfd_signal(iocb->ki_eventfd, 1);
1112 
1113         /* everything turned out well, dispose of the aiocb. */
1114         kiocb_free(iocb);
1115 
1116         /*
1117          * We have to order our ring_info tail store above and test
1118          * of the wait list below outside the wait lock.  This is
1119          * like in wake_up_bit() where clearing a bit has to be
1120          * ordered with the unlocked test.
1121          */
1122         smp_mb();
1123 
1124         if (waitqueue_active(&ctx->wait))
1125                 wake_up(&ctx->wait);
1126 
1127         percpu_ref_put(&ctx->reqs);
1128 }
1129 EXPORT_SYMBOL(aio_complete);
1130 
1131 /* aio_read_events_ring
1132  *      Pull an event off of the ioctx's event ring.  Returns the number of
1133  *      events fetched
1134  */
1135 static long aio_read_events_ring(struct kioctx *ctx,
1136                                  struct io_event __user *event, long nr)
1137 {
1138         struct aio_ring *ring;
1139         unsigned head, tail, pos;
1140         long ret = 0;
1141         int copy_ret;
1142 
1143         /*
1144          * The mutex can block and wake us up and that will cause
1145          * wait_event_interruptible_hrtimeout() to schedule without sleeping
1146          * and repeat. This should be rare enough that it doesn't cause
1147          * peformance issues. See the comment in read_events() for more detail.
1148          */
1149         sched_annotate_sleep();
1150         mutex_lock(&ctx->ring_lock);
1151 
1152         /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1153         ring = kmap_atomic(ctx->ring_pages[0]);
1154         head = ring->head;
1155         tail = ring->tail;
1156         kunmap_atomic(ring);
1157 
1158         /*
1159          * Ensure that once we've read the current tail pointer, that
1160          * we also see the events that were stored up to the tail.
1161          */
1162         smp_rmb();
1163 
1164         pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1165 
1166         if (head == tail)
1167                 goto out;
1168 
1169         head %= ctx->nr_events;
1170         tail %= ctx->nr_events;
1171 
1172         while (ret < nr) {
1173                 long avail;
1174                 struct io_event *ev;
1175                 struct page *page;
1176 
1177                 avail = (head <= tail ?  tail : ctx->nr_events) - head;
1178                 if (head == tail)
1179                         break;
1180 
1181                 avail = min(avail, nr - ret);
1182                 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1183                             ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1184 
1185                 pos = head + AIO_EVENTS_OFFSET;
1186                 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1187                 pos %= AIO_EVENTS_PER_PAGE;
1188 
1189                 ev = kmap(page);
1190                 copy_ret = copy_to_user(event + ret, ev + pos,
1191                                         sizeof(*ev) * avail);
1192                 kunmap(page);
1193 
1194                 if (unlikely(copy_ret)) {
1195                         ret = -EFAULT;
1196                         goto out;
1197                 }
1198 
1199                 ret += avail;
1200                 head += avail;
1201                 head %= ctx->nr_events;
1202         }
1203 
1204         ring = kmap_atomic(ctx->ring_pages[0]);
1205         ring->head = head;
1206         kunmap_atomic(ring);
1207         flush_dcache_page(ctx->ring_pages[0]);
1208 
1209         pr_debug("%li  h%u t%u\n", ret, head, tail);
1210 out:
1211         mutex_unlock(&ctx->ring_lock);
1212 
1213         return ret;
1214 }
1215 
1216 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1217                             struct io_event __user *event, long *i)
1218 {
1219         long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1220 
1221         if (ret > 0)
1222                 *i += ret;
1223 
1224         if (unlikely(atomic_read(&ctx->dead)))
1225                 ret = -EINVAL;
1226 
1227         if (!*i)
1228                 *i = ret;
1229 
1230         return ret < 0 || *i >= min_nr;
1231 }
1232 
1233 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1234                         struct io_event __user *event,
1235                         struct timespec __user *timeout)
1236 {
1237         ktime_t until = { .tv64 = KTIME_MAX };
1238         long ret = 0;
1239 
1240         if (timeout) {
1241                 struct timespec ts;
1242 
1243                 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1244                         return -EFAULT;
1245 
1246                 until = timespec_to_ktime(ts);
1247         }
1248 
1249         /*
1250          * Note that aio_read_events() is being called as the conditional - i.e.
1251          * we're calling it after prepare_to_wait() has set task state to
1252          * TASK_INTERRUPTIBLE.
1253          *
1254          * But aio_read_events() can block, and if it blocks it's going to flip
1255          * the task state back to TASK_RUNNING.
1256          *
1257          * This should be ok, provided it doesn't flip the state back to
1258          * TASK_RUNNING and return 0 too much - that causes us to spin. That
1259          * will only happen if the mutex_lock() call blocks, and we then find
1260          * the ringbuffer empty. So in practice we should be ok, but it's
1261          * something to be aware of when touching this code.
1262          */
1263         if (until.tv64 == 0)
1264                 aio_read_events(ctx, min_nr, nr, event, &ret);
1265         else
1266                 wait_event_interruptible_hrtimeout(ctx->wait,
1267                                 aio_read_events(ctx, min_nr, nr, event, &ret),
1268                                 until);
1269 
1270         if (!ret && signal_pending(current))
1271                 ret = -EINTR;
1272 
1273         return ret;
1274 }
1275 
1276 /* sys_io_setup:
1277  *      Create an aio_context capable of receiving at least nr_events.
1278  *      ctxp must not point to an aio_context that already exists, and
1279  *      must be initialized to 0 prior to the call.  On successful
1280  *      creation of the aio_context, *ctxp is filled in with the resulting 
1281  *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1282  *      if the specified nr_events exceeds internal limits.  May fail 
1283  *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1284  *      of available events.  May fail with -ENOMEM if insufficient kernel
1285  *      resources are available.  May fail with -EFAULT if an invalid
1286  *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1287  *      implemented.
1288  */
1289 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1290 {
1291         struct kioctx *ioctx = NULL;
1292         unsigned long ctx;
1293         long ret;
1294 
1295         ret = get_user(ctx, ctxp);
1296         if (unlikely(ret))
1297                 goto out;
1298 
1299         ret = -EINVAL;
1300         if (unlikely(ctx || nr_events == 0)) {
1301                 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1302                          ctx, nr_events);
1303                 goto out;
1304         }
1305 
1306         ioctx = ioctx_alloc(nr_events);
1307         ret = PTR_ERR(ioctx);
1308         if (!IS_ERR(ioctx)) {
1309                 ret = put_user(ioctx->user_id, ctxp);
1310                 if (ret)
1311                         kill_ioctx(current->mm, ioctx, NULL);
1312                 percpu_ref_put(&ioctx->users);
1313         }
1314 
1315 out:
1316         return ret;
1317 }
1318 
1319 /* sys_io_destroy:
1320  *      Destroy the aio_context specified.  May cancel any outstanding 
1321  *      AIOs and block on completion.  Will fail with -ENOSYS if not
1322  *      implemented.  May fail with -EINVAL if the context pointed to
1323  *      is invalid.
1324  */
1325 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1326 {
1327         struct kioctx *ioctx = lookup_ioctx(ctx);
1328         if (likely(NULL != ioctx)) {
1329                 struct completion requests_done =
1330                         COMPLETION_INITIALIZER_ONSTACK(requests_done);
1331                 int ret;
1332 
1333                 /* Pass requests_done to kill_ioctx() where it can be set
1334                  * in a thread-safe way. If we try to set it here then we have
1335                  * a race condition if two io_destroy() called simultaneously.
1336                  */
1337                 ret = kill_ioctx(current->mm, ioctx, &requests_done);
1338                 percpu_ref_put(&ioctx->users);
1339 
1340                 /* Wait until all IO for the context are done. Otherwise kernel
1341                  * keep using user-space buffers even if user thinks the context
1342                  * is destroyed.
1343                  */
1344                 if (!ret)
1345                         wait_for_completion(&requests_done);
1346 
1347                 return ret;
1348         }
1349         pr_debug("EINVAL: io_destroy: invalid context id\n");
1350         return -EINVAL;
1351 }
1352 
1353 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1354                             unsigned long, loff_t);
1355 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1356 
1357 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1358                                      int rw, char __user *buf,
1359                                      unsigned long *nr_segs,
1360                                      struct iovec **iovec,
1361                                      bool compat)
1362 {
1363         ssize_t ret;
1364 
1365         *nr_segs = kiocb->ki_nbytes;
1366 
1367 #ifdef CONFIG_COMPAT
1368         if (compat)
1369                 ret = compat_rw_copy_check_uvector(rw,
1370                                 (struct compat_iovec __user *)buf,
1371                                 *nr_segs, UIO_FASTIOV, *iovec, iovec);
1372         else
1373 #endif
1374                 ret = rw_copy_check_uvector(rw,
1375                                 (struct iovec __user *)buf,
1376                                 *nr_segs, UIO_FASTIOV, *iovec, iovec);
1377         if (ret < 0)
1378                 return ret;
1379 
1380         /* ki_nbytes now reflect bytes instead of segs */
1381         kiocb->ki_nbytes = ret;
1382         return 0;
1383 }
1384 
1385 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1386                                        int rw, char __user *buf,
1387                                        unsigned long *nr_segs,
1388                                        struct iovec *iovec)
1389 {
1390         if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1391                 return -EFAULT;
1392 
1393         iovec->iov_base = buf;
1394         iovec->iov_len = kiocb->ki_nbytes;
1395         *nr_segs = 1;
1396         return 0;
1397 }
1398 
1399 /*
1400  * aio_run_iocb:
1401  *      Performs the initial checks and io submission.
1402  */
1403 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1404                             char __user *buf, bool compat)
1405 {
1406         struct file *file = req->ki_filp;
1407         ssize_t ret;
1408         unsigned long nr_segs;
1409         int rw;
1410         fmode_t mode;
1411         aio_rw_op *rw_op;
1412         rw_iter_op *iter_op;
1413         struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1414         struct iov_iter iter;
1415 
1416         switch (opcode) {
1417         case IOCB_CMD_PREAD:
1418         case IOCB_CMD_PREADV:
1419                 mode    = FMODE_READ;
1420                 rw      = READ;
1421                 rw_op   = file->f_op->aio_read;
1422                 iter_op = file->f_op->read_iter;
1423                 goto rw_common;
1424 
1425         case IOCB_CMD_PWRITE:
1426         case IOCB_CMD_PWRITEV:
1427                 mode    = FMODE_WRITE;
1428                 rw      = WRITE;
1429                 rw_op   = file->f_op->aio_write;
1430                 iter_op = file->f_op->write_iter;
1431                 goto rw_common;
1432 rw_common:
1433                 if (unlikely(!(file->f_mode & mode)))
1434                         return -EBADF;
1435 
1436                 if (!rw_op && !iter_op)
1437                         return -EINVAL;
1438 
1439                 ret = (opcode == IOCB_CMD_PREADV ||
1440                        opcode == IOCB_CMD_PWRITEV)
1441                         ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1442                                                 &iovec, compat)
1443                         : aio_setup_single_vector(req, rw, buf, &nr_segs,
1444                                                   iovec);
1445                 if (!ret)
1446                         ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1447                 if (ret < 0) {
1448                         if (iovec != inline_vecs)
1449                                 kfree(iovec);
1450                         return ret;
1451                 }
1452 
1453                 req->ki_nbytes = ret;
1454 
1455                 /* XXX: move/kill - rw_verify_area()? */
1456                 /* This matches the pread()/pwrite() logic */
1457                 if (req->ki_pos < 0) {
1458                         ret = -EINVAL;
1459                         break;
1460                 }
1461 
1462                 if (rw == WRITE)
1463                         file_start_write(file);
1464 
1465                 if (iter_op) {
1466                         iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes);
1467                         ret = iter_op(req, &iter);
1468                 } else {
1469                         ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1470                 }
1471 
1472                 if (rw == WRITE)
1473                         file_end_write(file);
1474                 break;
1475 
1476         case IOCB_CMD_FDSYNC:
1477                 if (!file->f_op->aio_fsync)
1478                         return -EINVAL;
1479 
1480                 ret = file->f_op->aio_fsync(req, 1);
1481                 break;
1482 
1483         case IOCB_CMD_FSYNC:
1484                 if (!file->f_op->aio_fsync)
1485                         return -EINVAL;
1486 
1487                 ret = file->f_op->aio_fsync(req, 0);
1488                 break;
1489 
1490         default:
1491                 pr_debug("EINVAL: no operation provided\n");
1492                 return -EINVAL;
1493         }
1494 
1495         if (iovec != inline_vecs)
1496                 kfree(iovec);
1497 
1498         if (ret != -EIOCBQUEUED) {
1499                 /*
1500                  * There's no easy way to restart the syscall since other AIO's
1501                  * may be already running. Just fail this IO with EINTR.
1502                  */
1503                 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1504                              ret == -ERESTARTNOHAND ||
1505                              ret == -ERESTART_RESTARTBLOCK))
1506                         ret = -EINTR;
1507                 aio_complete(req, ret, 0);
1508         }
1509 
1510         return 0;
1511 }
1512 
1513 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1514                          struct iocb *iocb, bool compat)
1515 {
1516         struct kiocb *req;
1517         ssize_t ret;
1518 
1519         /* enforce forwards compatibility on users */
1520         if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1521                 pr_debug("EINVAL: reserve field set\n");
1522                 return -EINVAL;
1523         }
1524 
1525         /* prevent overflows */
1526         if (unlikely(
1527             (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1528             (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1529             ((ssize_t)iocb->aio_nbytes < 0)
1530            )) {
1531                 pr_debug("EINVAL: io_submit: overflow check\n");
1532                 return -EINVAL;
1533         }
1534 
1535         req = aio_get_req(ctx);
1536         if (unlikely(!req))
1537                 return -EAGAIN;
1538 
1539         req->ki_filp = fget(iocb->aio_fildes);
1540         if (unlikely(!req->ki_filp)) {
1541                 ret = -EBADF;
1542                 goto out_put_req;
1543         }
1544 
1545         if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1546                 /*
1547                  * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1548                  * instance of the file* now. The file descriptor must be
1549                  * an eventfd() fd, and will be signaled for each completed
1550                  * event using the eventfd_signal() function.
1551                  */
1552                 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1553                 if (IS_ERR(req->ki_eventfd)) {
1554                         ret = PTR_ERR(req->ki_eventfd);
1555                         req->ki_eventfd = NULL;
1556                         goto out_put_req;
1557                 }
1558         }
1559 
1560         ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1561         if (unlikely(ret)) {
1562                 pr_debug("EFAULT: aio_key\n");
1563                 goto out_put_req;
1564         }
1565 
1566         req->ki_obj.user = user_iocb;
1567         req->ki_user_data = iocb->aio_data;
1568         req->ki_pos = iocb->aio_offset;
1569         req->ki_nbytes = iocb->aio_nbytes;
1570 
1571         ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1572                            (char __user *)(unsigned long)iocb->aio_buf,
1573                            compat);
1574         if (ret)
1575                 goto out_put_req;
1576 
1577         return 0;
1578 out_put_req:
1579         put_reqs_available(ctx, 1);
1580         percpu_ref_put(&ctx->reqs);
1581         kiocb_free(req);
1582         return ret;
1583 }
1584 
1585 long do_io_submit(aio_context_t ctx_id, long nr,
1586                   struct iocb __user *__user *iocbpp, bool compat)
1587 {
1588         struct kioctx *ctx;
1589         long ret = 0;
1590         int i = 0;
1591         struct blk_plug plug;
1592 
1593         if (unlikely(nr < 0))
1594                 return -EINVAL;
1595 
1596         if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1597                 nr = LONG_MAX/sizeof(*iocbpp);
1598 
1599         if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1600                 return -EFAULT;
1601 
1602         ctx = lookup_ioctx(ctx_id);
1603         if (unlikely(!ctx)) {
1604                 pr_debug("EINVAL: invalid context id\n");
1605                 return -EINVAL;
1606         }
1607 
1608         blk_start_plug(&plug);
1609 
1610         /*
1611          * AKPM: should this return a partial result if some of the IOs were
1612          * successfully submitted?
1613          */
1614         for (i=0; i<nr; i++) {
1615                 struct iocb __user *user_iocb;
1616                 struct iocb tmp;
1617 
1618                 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1619                         ret = -EFAULT;
1620                         break;
1621                 }
1622 
1623                 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1624                         ret = -EFAULT;
1625                         break;
1626                 }
1627 
1628                 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1629                 if (ret)
1630                         break;
1631         }
1632         blk_finish_plug(&plug);
1633 
1634         percpu_ref_put(&ctx->users);
1635         return i ? i : ret;
1636 }
1637 
1638 /* sys_io_submit:
1639  *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1640  *      the number of iocbs queued.  May return -EINVAL if the aio_context
1641  *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1642  *      *iocbpp[0] is not properly initialized, if the operation specified
1643  *      is invalid for the file descriptor in the iocb.  May fail with
1644  *      -EFAULT if any of the data structures point to invalid data.  May
1645  *      fail with -EBADF if the file descriptor specified in the first
1646  *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1647  *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1648  *      fail with -ENOSYS if not implemented.
1649  */
1650 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1651                 struct iocb __user * __user *, iocbpp)
1652 {
1653         return do_io_submit(ctx_id, nr, iocbpp, 0);
1654 }
1655 
1656 /* lookup_kiocb
1657  *      Finds a given iocb for cancellation.
1658  */
1659 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1660                                   u32 key)
1661 {
1662         struct list_head *pos;
1663 
1664         assert_spin_locked(&ctx->ctx_lock);
1665 
1666         if (key != KIOCB_KEY)
1667                 return NULL;
1668 
1669         /* TODO: use a hash or array, this sucks. */
1670         list_for_each(pos, &ctx->active_reqs) {
1671                 struct kiocb *kiocb = list_kiocb(pos);
1672                 if (kiocb->ki_obj.user == iocb)
1673                         return kiocb;
1674         }
1675         return NULL;
1676 }
1677 
1678 /* sys_io_cancel:
1679  *      Attempts to cancel an iocb previously passed to io_submit.  If
1680  *      the operation is successfully cancelled, the resulting event is
1681  *      copied into the memory pointed to by result without being placed
1682  *      into the completion queue and 0 is returned.  May fail with
1683  *      -EFAULT if any of the data structures pointed to are invalid.
1684  *      May fail with -EINVAL if aio_context specified by ctx_id is
1685  *      invalid.  May fail with -EAGAIN if the iocb specified was not
1686  *      cancelled.  Will fail with -ENOSYS if not implemented.
1687  */
1688 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1689                 struct io_event __user *, result)
1690 {
1691         struct kioctx *ctx;
1692         struct kiocb *kiocb;
1693         u32 key;
1694         int ret;
1695 
1696         ret = get_user(key, &iocb->aio_key);
1697         if (unlikely(ret))
1698                 return -EFAULT;
1699 
1700         ctx = lookup_ioctx(ctx_id);
1701         if (unlikely(!ctx))
1702                 return -EINVAL;
1703 
1704         spin_lock_irq(&ctx->ctx_lock);
1705 
1706         kiocb = lookup_kiocb(ctx, iocb, key);
1707         if (kiocb)
1708                 ret = kiocb_cancel(kiocb);
1709         else
1710                 ret = -EINVAL;
1711 
1712         spin_unlock_irq(&ctx->ctx_lock);
1713 
1714         if (!ret) {
1715                 /*
1716                  * The result argument is no longer used - the io_event is
1717                  * always delivered via the ring buffer. -EINPROGRESS indicates
1718                  * cancellation is progress:
1719                  */
1720                 ret = -EINPROGRESS;
1721         }
1722 
1723         percpu_ref_put(&ctx->users);
1724 
1725         return ret;
1726 }
1727 
1728 /* io_getevents:
1729  *      Attempts to read at least min_nr events and up to nr events from
1730  *      the completion queue for the aio_context specified by ctx_id. If
1731  *      it succeeds, the number of read events is returned. May fail with
1732  *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1733  *      out of range, if timeout is out of range.  May fail with -EFAULT
1734  *      if any of the memory specified is invalid.  May return 0 or
1735  *      < min_nr if the timeout specified by timeout has elapsed
1736  *      before sufficient events are available, where timeout == NULL
1737  *      specifies an infinite timeout. Note that the timeout pointed to by
1738  *      timeout is relative.  Will fail with -ENOSYS if not implemented.
1739  */
1740 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1741                 long, min_nr,
1742                 long, nr,
1743                 struct io_event __user *, events,
1744                 struct timespec __user *, timeout)
1745 {
1746         struct kioctx *ioctx = lookup_ioctx(ctx_id);
1747         long ret = -EINVAL;
1748 
1749         if (likely(ioctx)) {
1750                 if (likely(min_nr <= nr && min_nr >= 0))
1751                         ret = read_events(ioctx, min_nr, nr, events, timeout);
1752                 percpu_ref_put(&ioctx->users);
1753         }
1754         return ret;
1755 }
1756 

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