Version:  2.0.40 2.2.26 2.4.37 2.6.39 3.0 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15

Linux/fs/aio.c

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

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