~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/aio.c

Version: ~ [ linux-5.13-rc1 ] ~ [ linux-5.12.2 ] ~ [ linux-5.11.19 ] ~ [ linux-5.10.35 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.117 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.190 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.232 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.268 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.268 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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/signal.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 <linux/uaccess.h>
 46 
 47 #include "internal.h"
 48 
 49 #define AIO_RING_MAGIC                  0xa10a10a1
 50 #define AIO_RING_COMPAT_FEATURES        1
 51 #define AIO_RING_INCOMPAT_FEATURES      0
 52 struct aio_ring {
 53         unsigned        id;     /* kernel internal index number */
 54         unsigned        nr;     /* number of io_events */
 55         unsigned        head;   /* Written to by userland or under ring_lock
 56                                  * mutex by aio_read_events_ring(). */
 57         unsigned        tail;
 58 
 59         unsigned        magic;
 60         unsigned        compat_features;
 61         unsigned        incompat_features;
 62         unsigned        header_length;  /* size of aio_ring */
 63 
 64 
 65         struct io_event         io_events[0];
 66 }; /* 128 bytes + ring size */
 67 
 68 #define AIO_RING_PAGES  8
 69 
 70 struct kioctx_table {
 71         struct rcu_head rcu;
 72         unsigned        nr;
 73         struct kioctx   *table[];
 74 };
 75 
 76 struct kioctx_cpu {
 77         unsigned                reqs_available;
 78 };
 79 
 80 struct ctx_rq_wait {
 81         struct completion comp;
 82         atomic_t count;
 83 };
 84 
 85 struct kioctx {
 86         struct percpu_ref       users;
 87         atomic_t                dead;
 88 
 89         struct percpu_ref       reqs;
 90 
 91         unsigned long           user_id;
 92 
 93         struct __percpu kioctx_cpu *cpu;
 94 
 95         /*
 96          * For percpu reqs_available, number of slots we move to/from global
 97          * counter at a time:
 98          */
 99         unsigned                req_batch;
100         /*
101          * This is what userspace passed to io_setup(), it's not used for
102          * anything but counting against the global max_reqs quota.
103          *
104          * The real limit is nr_events - 1, which will be larger (see
105          * aio_setup_ring())
106          */
107         unsigned                max_reqs;
108 
109         /* Size of ringbuffer, in units of struct io_event */
110         unsigned                nr_events;
111 
112         unsigned long           mmap_base;
113         unsigned long           mmap_size;
114 
115         struct page             **ring_pages;
116         long                    nr_pages;
117 
118         struct work_struct      free_work;
119 
120         /*
121          * signals when all in-flight requests are done
122          */
123         struct ctx_rq_wait      *rq_wait;
124 
125         struct {
126                 /*
127                  * This counts the number of available slots in the ringbuffer,
128                  * so we avoid overflowing it: it's decremented (if positive)
129                  * when allocating a kiocb and incremented when the resulting
130                  * io_event is pulled off the ringbuffer.
131                  *
132                  * We batch accesses to it with a percpu version.
133                  */
134                 atomic_t        reqs_available;
135         } ____cacheline_aligned_in_smp;
136 
137         struct {
138                 spinlock_t      ctx_lock;
139                 struct list_head active_reqs;   /* used for cancellation */
140         } ____cacheline_aligned_in_smp;
141 
142         struct {
143                 struct mutex    ring_lock;
144                 wait_queue_head_t wait;
145         } ____cacheline_aligned_in_smp;
146 
147         struct {
148                 unsigned        tail;
149                 unsigned        completed_events;
150                 spinlock_t      completion_lock;
151         } ____cacheline_aligned_in_smp;
152 
153         struct page             *internal_pages[AIO_RING_PAGES];
154         struct file             *aio_ring_file;
155 
156         unsigned                id;
157 };
158 
159 /*
160  * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
161  * cancelled or completed (this makes a certain amount of sense because
162  * successful cancellation - io_cancel() - does deliver the completion to
163  * userspace).
164  *
165  * And since most things don't implement kiocb cancellation and we'd really like
166  * kiocb completion to be lockless when possible, we use ki_cancel to
167  * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
168  * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
169  */
170 #define KIOCB_CANCELLED         ((void *) (~0ULL))
171 
172 struct aio_kiocb {
173         struct kiocb            common;
174 
175         struct kioctx           *ki_ctx;
176         kiocb_cancel_fn         *ki_cancel;
177 
178         struct iocb __user      *ki_user_iocb;  /* user's aiocb */
179         __u64                   ki_user_data;   /* user's data for completion */
180 
181         struct list_head        ki_list;        /* the aio core uses this
182                                                  * for cancellation */
183 
184         /*
185          * If the aio_resfd field of the userspace iocb is not zero,
186          * this is the underlying eventfd context to deliver events to.
187          */
188         struct eventfd_ctx      *ki_eventfd;
189 };
190 
191 /*------ sysctl variables----*/
192 static DEFINE_SPINLOCK(aio_nr_lock);
193 unsigned long aio_nr;           /* current system wide number of aio requests */
194 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
195 /*----end sysctl variables---*/
196 
197 static struct kmem_cache        *kiocb_cachep;
198 static struct kmem_cache        *kioctx_cachep;
199 
200 static struct vfsmount *aio_mnt;
201 
202 static const struct file_operations aio_ring_fops;
203 static const struct address_space_operations aio_ctx_aops;
204 
205 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
206 {
207         struct qstr this = QSTR_INIT("[aio]", 5);
208         struct file *file;
209         struct path path;
210         struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
211         if (IS_ERR(inode))
212                 return ERR_CAST(inode);
213 
214         inode->i_mapping->a_ops = &aio_ctx_aops;
215         inode->i_mapping->private_data = ctx;
216         inode->i_size = PAGE_SIZE * nr_pages;
217 
218         path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
219         if (!path.dentry) {
220                 iput(inode);
221                 return ERR_PTR(-ENOMEM);
222         }
223         path.mnt = mntget(aio_mnt);
224 
225         d_instantiate(path.dentry, inode);
226         file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
227         if (IS_ERR(file)) {
228                 path_put(&path);
229                 return file;
230         }
231 
232         file->f_flags = O_RDWR;
233         return file;
234 }
235 
236 static struct dentry *aio_mount(struct file_system_type *fs_type,
237                                 int flags, const char *dev_name, void *data)
238 {
239         static const struct dentry_operations ops = {
240                 .d_dname        = simple_dname,
241         };
242         struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops,
243                                            AIO_RING_MAGIC);
244 
245         if (!IS_ERR(root))
246                 root->d_sb->s_iflags |= SB_I_NOEXEC;
247         return root;
248 }
249 
250 /* aio_setup
251  *      Creates the slab caches used by the aio routines, panic on
252  *      failure as this is done early during the boot sequence.
253  */
254 static int __init aio_setup(void)
255 {
256         static struct file_system_type aio_fs = {
257                 .name           = "aio",
258                 .mount          = aio_mount,
259                 .kill_sb        = kill_anon_super,
260         };
261         aio_mnt = kern_mount(&aio_fs);
262         if (IS_ERR(aio_mnt))
263                 panic("Failed to create aio fs mount.");
264 
265         kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
266         kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
267 
268         pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
269 
270         return 0;
271 }
272 __initcall(aio_setup);
273 
274 static void put_aio_ring_file(struct kioctx *ctx)
275 {
276         struct file *aio_ring_file = ctx->aio_ring_file;
277         struct address_space *i_mapping;
278 
279         if (aio_ring_file) {
280                 truncate_setsize(file_inode(aio_ring_file), 0);
281 
282                 /* Prevent further access to the kioctx from migratepages */
283                 i_mapping = aio_ring_file->f_mapping;
284                 spin_lock(&i_mapping->private_lock);
285                 i_mapping->private_data = NULL;
286                 ctx->aio_ring_file = NULL;
287                 spin_unlock(&i_mapping->private_lock);
288 
289                 fput(aio_ring_file);
290         }
291 }
292 
293 static void aio_free_ring(struct kioctx *ctx)
294 {
295         int i;
296 
297         /* Disconnect the kiotx from the ring file.  This prevents future
298          * accesses to the kioctx from page migration.
299          */
300         put_aio_ring_file(ctx);
301 
302         for (i = 0; i < ctx->nr_pages; i++) {
303                 struct page *page;
304                 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
305                                 page_count(ctx->ring_pages[i]));
306                 page = ctx->ring_pages[i];
307                 if (!page)
308                         continue;
309                 ctx->ring_pages[i] = NULL;
310                 put_page(page);
311         }
312 
313         if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
314                 kfree(ctx->ring_pages);
315                 ctx->ring_pages = NULL;
316         }
317 }
318 
319 static int aio_ring_mremap(struct vm_area_struct *vma)
320 {
321         struct file *file = vma->vm_file;
322         struct mm_struct *mm = vma->vm_mm;
323         struct kioctx_table *table;
324         int i, res = -EINVAL;
325 
326         spin_lock(&mm->ioctx_lock);
327         rcu_read_lock();
328         table = rcu_dereference(mm->ioctx_table);
329         for (i = 0; i < table->nr; i++) {
330                 struct kioctx *ctx;
331 
332                 ctx = table->table[i];
333                 if (ctx && ctx->aio_ring_file == file) {
334                         if (!atomic_read(&ctx->dead)) {
335                                 ctx->user_id = ctx->mmap_base = vma->vm_start;
336                                 res = 0;
337                         }
338                         break;
339                 }
340         }
341 
342         rcu_read_unlock();
343         spin_unlock(&mm->ioctx_lock);
344         return res;
345 }
346 
347 static const struct vm_operations_struct aio_ring_vm_ops = {
348         .mremap         = aio_ring_mremap,
349 #if IS_ENABLED(CONFIG_MMU)
350         .fault          = filemap_fault,
351         .map_pages      = filemap_map_pages,
352         .page_mkwrite   = filemap_page_mkwrite,
353 #endif
354 };
355 
356 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
357 {
358         vma->vm_flags |= VM_DONTEXPAND;
359         vma->vm_ops = &aio_ring_vm_ops;
360         return 0;
361 }
362 
363 static const struct file_operations aio_ring_fops = {
364         .mmap = aio_ring_mmap,
365 };
366 
367 #if IS_ENABLED(CONFIG_MIGRATION)
368 static int aio_migratepage(struct address_space *mapping, struct page *new,
369                         struct page *old, enum migrate_mode mode)
370 {
371         struct kioctx *ctx;
372         unsigned long flags;
373         pgoff_t idx;
374         int rc;
375 
376         rc = 0;
377 
378         /* mapping->private_lock here protects against the kioctx teardown.  */
379         spin_lock(&mapping->private_lock);
380         ctx = mapping->private_data;
381         if (!ctx) {
382                 rc = -EINVAL;
383                 goto out;
384         }
385 
386         /* The ring_lock mutex.  The prevents aio_read_events() from writing
387          * to the ring's head, and prevents page migration from mucking in
388          * a partially initialized kiotx.
389          */
390         if (!mutex_trylock(&ctx->ring_lock)) {
391                 rc = -EAGAIN;
392                 goto out;
393         }
394 
395         idx = old->index;
396         if (idx < (pgoff_t)ctx->nr_pages) {
397                 /* Make sure the old page hasn't already been changed */
398                 if (ctx->ring_pages[idx] != old)
399                         rc = -EAGAIN;
400         } else
401                 rc = -EINVAL;
402 
403         if (rc != 0)
404                 goto out_unlock;
405 
406         /* Writeback must be complete */
407         BUG_ON(PageWriteback(old));
408         get_page(new);
409 
410         rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
411         if (rc != MIGRATEPAGE_SUCCESS) {
412                 put_page(new);
413                 goto out_unlock;
414         }
415 
416         /* Take completion_lock to prevent other writes to the ring buffer
417          * while the old page is copied to the new.  This prevents new
418          * events from being lost.
419          */
420         spin_lock_irqsave(&ctx->completion_lock, flags);
421         migrate_page_copy(new, old);
422         BUG_ON(ctx->ring_pages[idx] != old);
423         ctx->ring_pages[idx] = new;
424         spin_unlock_irqrestore(&ctx->completion_lock, flags);
425 
426         /* The old page is no longer accessible. */
427         put_page(old);
428 
429 out_unlock:
430         mutex_unlock(&ctx->ring_lock);
431 out:
432         spin_unlock(&mapping->private_lock);
433         return rc;
434 }
435 #endif
436 
437 static const struct address_space_operations aio_ctx_aops = {
438         .set_page_dirty = __set_page_dirty_no_writeback,
439 #if IS_ENABLED(CONFIG_MIGRATION)
440         .migratepage    = aio_migratepage,
441 #endif
442 };
443 
444 static int aio_setup_ring(struct kioctx *ctx)
445 {
446         struct aio_ring *ring;
447         unsigned nr_events = ctx->max_reqs;
448         struct mm_struct *mm = current->mm;
449         unsigned long size, unused;
450         int nr_pages;
451         int i;
452         struct file *file;
453 
454         /* Compensate for the ring buffer's head/tail overlap entry */
455         nr_events += 2; /* 1 is required, 2 for good luck */
456 
457         size = sizeof(struct aio_ring);
458         size += sizeof(struct io_event) * nr_events;
459 
460         nr_pages = PFN_UP(size);
461         if (nr_pages < 0)
462                 return -EINVAL;
463 
464         file = aio_private_file(ctx, nr_pages);
465         if (IS_ERR(file)) {
466                 ctx->aio_ring_file = NULL;
467                 return -ENOMEM;
468         }
469 
470         ctx->aio_ring_file = file;
471         nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
472                         / sizeof(struct io_event);
473 
474         ctx->ring_pages = ctx->internal_pages;
475         if (nr_pages > AIO_RING_PAGES) {
476                 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
477                                           GFP_KERNEL);
478                 if (!ctx->ring_pages) {
479                         put_aio_ring_file(ctx);
480                         return -ENOMEM;
481                 }
482         }
483 
484         for (i = 0; i < nr_pages; i++) {
485                 struct page *page;
486                 page = find_or_create_page(file->f_mapping,
487                                            i, GFP_HIGHUSER | __GFP_ZERO);
488                 if (!page)
489                         break;
490                 pr_debug("pid(%d) page[%d]->count=%d\n",
491                          current->pid, i, page_count(page));
492                 SetPageUptodate(page);
493                 unlock_page(page);
494 
495                 ctx->ring_pages[i] = page;
496         }
497         ctx->nr_pages = i;
498 
499         if (unlikely(i != nr_pages)) {
500                 aio_free_ring(ctx);
501                 return -ENOMEM;
502         }
503 
504         ctx->mmap_size = nr_pages * PAGE_SIZE;
505         pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
506 
507         if (down_write_killable(&mm->mmap_sem)) {
508                 ctx->mmap_size = 0;
509                 aio_free_ring(ctx);
510                 return -EINTR;
511         }
512 
513         ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
514                                        PROT_READ | PROT_WRITE,
515                                        MAP_SHARED, 0, &unused, NULL);
516         up_write(&mm->mmap_sem);
517         if (IS_ERR((void *)ctx->mmap_base)) {
518                 ctx->mmap_size = 0;
519                 aio_free_ring(ctx);
520                 return -ENOMEM;
521         }
522 
523         pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
524 
525         ctx->user_id = ctx->mmap_base;
526         ctx->nr_events = nr_events; /* trusted copy */
527 
528         ring = kmap_atomic(ctx->ring_pages[0]);
529         ring->nr = nr_events;   /* user copy */
530         ring->id = ~0U;
531         ring->head = ring->tail = 0;
532         ring->magic = AIO_RING_MAGIC;
533         ring->compat_features = AIO_RING_COMPAT_FEATURES;
534         ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
535         ring->header_length = sizeof(struct aio_ring);
536         kunmap_atomic(ring);
537         flush_dcache_page(ctx->ring_pages[0]);
538 
539         return 0;
540 }
541 
542 #define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
543 #define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
544 #define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
545 
546 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
547 {
548         struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
549         struct kioctx *ctx = req->ki_ctx;
550         unsigned long flags;
551 
552         spin_lock_irqsave(&ctx->ctx_lock, flags);
553 
554         if (!req->ki_list.next)
555                 list_add(&req->ki_list, &ctx->active_reqs);
556 
557         req->ki_cancel = cancel;
558 
559         spin_unlock_irqrestore(&ctx->ctx_lock, flags);
560 }
561 EXPORT_SYMBOL(kiocb_set_cancel_fn);
562 
563 static int kiocb_cancel(struct aio_kiocb *kiocb)
564 {
565         kiocb_cancel_fn *old, *cancel;
566 
567         /*
568          * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
569          * actually has a cancel function, hence the cmpxchg()
570          */
571 
572         cancel = ACCESS_ONCE(kiocb->ki_cancel);
573         do {
574                 if (!cancel || cancel == KIOCB_CANCELLED)
575                         return -EINVAL;
576 
577                 old = cancel;
578                 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
579         } while (cancel != old);
580 
581         return cancel(&kiocb->common);
582 }
583 
584 static void free_ioctx(struct work_struct *work)
585 {
586         struct kioctx *ctx = container_of(work, struct kioctx, free_work);
587 
588         pr_debug("freeing %p\n", ctx);
589 
590         aio_free_ring(ctx);
591         free_percpu(ctx->cpu);
592         percpu_ref_exit(&ctx->reqs);
593         percpu_ref_exit(&ctx->users);
594         kmem_cache_free(kioctx_cachep, ctx);
595 }
596 
597 static void free_ioctx_reqs(struct percpu_ref *ref)
598 {
599         struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
600 
601         /* At this point we know that there are no any in-flight requests */
602         if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
603                 complete(&ctx->rq_wait->comp);
604 
605         INIT_WORK(&ctx->free_work, free_ioctx);
606         schedule_work(&ctx->free_work);
607 }
608 
609 /*
610  * When this function runs, the kioctx has been removed from the "hash table"
611  * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
612  * now it's safe to cancel any that need to be.
613  */
614 static void free_ioctx_users(struct percpu_ref *ref)
615 {
616         struct kioctx *ctx = container_of(ref, struct kioctx, users);
617         struct aio_kiocb *req;
618 
619         spin_lock_irq(&ctx->ctx_lock);
620 
621         while (!list_empty(&ctx->active_reqs)) {
622                 req = list_first_entry(&ctx->active_reqs,
623                                        struct aio_kiocb, ki_list);
624 
625                 list_del_init(&req->ki_list);
626                 kiocb_cancel(req);
627         }
628 
629         spin_unlock_irq(&ctx->ctx_lock);
630 
631         percpu_ref_kill(&ctx->reqs);
632         percpu_ref_put(&ctx->reqs);
633 }
634 
635 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
636 {
637         unsigned i, new_nr;
638         struct kioctx_table *table, *old;
639         struct aio_ring *ring;
640 
641         spin_lock(&mm->ioctx_lock);
642         table = rcu_dereference_raw(mm->ioctx_table);
643 
644         while (1) {
645                 if (table)
646                         for (i = 0; i < table->nr; i++)
647                                 if (!table->table[i]) {
648                                         ctx->id = i;
649                                         table->table[i] = ctx;
650                                         spin_unlock(&mm->ioctx_lock);
651 
652                                         /* While kioctx setup is in progress,
653                                          * we are protected from page migration
654                                          * changes ring_pages by ->ring_lock.
655                                          */
656                                         ring = kmap_atomic(ctx->ring_pages[0]);
657                                         ring->id = ctx->id;
658                                         kunmap_atomic(ring);
659                                         return 0;
660                                 }
661 
662                 new_nr = (table ? table->nr : 1) * 4;
663                 spin_unlock(&mm->ioctx_lock);
664 
665                 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
666                                 new_nr, GFP_KERNEL);
667                 if (!table)
668                         return -ENOMEM;
669 
670                 table->nr = new_nr;
671 
672                 spin_lock(&mm->ioctx_lock);
673                 old = rcu_dereference_raw(mm->ioctx_table);
674 
675                 if (!old) {
676                         rcu_assign_pointer(mm->ioctx_table, table);
677                 } else if (table->nr > old->nr) {
678                         memcpy(table->table, old->table,
679                                old->nr * sizeof(struct kioctx *));
680 
681                         rcu_assign_pointer(mm->ioctx_table, table);
682                         kfree_rcu(old, rcu);
683                 } else {
684                         kfree(table);
685                         table = old;
686                 }
687         }
688 }
689 
690 static void aio_nr_sub(unsigned nr)
691 {
692         spin_lock(&aio_nr_lock);
693         if (WARN_ON(aio_nr - nr > aio_nr))
694                 aio_nr = 0;
695         else
696                 aio_nr -= nr;
697         spin_unlock(&aio_nr_lock);
698 }
699 
700 /* ioctx_alloc
701  *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
702  */
703 static struct kioctx *ioctx_alloc(unsigned nr_events)
704 {
705         struct mm_struct *mm = current->mm;
706         struct kioctx *ctx;
707         int err = -ENOMEM;
708 
709         /*
710          * We keep track of the number of available ringbuffer slots, to prevent
711          * overflow (reqs_available), and we also use percpu counters for this.
712          *
713          * So since up to half the slots might be on other cpu's percpu counters
714          * and unavailable, double nr_events so userspace sees what they
715          * expected: additionally, we move req_batch slots to/from percpu
716          * counters at a time, so make sure that isn't 0:
717          */
718         nr_events = max(nr_events, num_possible_cpus() * 4);
719         nr_events *= 2;
720 
721         /* Prevent overflows */
722         if (nr_events > (0x10000000U / sizeof(struct io_event))) {
723                 pr_debug("ENOMEM: nr_events too high\n");
724                 return ERR_PTR(-EINVAL);
725         }
726 
727         if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
728                 return ERR_PTR(-EAGAIN);
729 
730         ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
731         if (!ctx)
732                 return ERR_PTR(-ENOMEM);
733 
734         ctx->max_reqs = nr_events;
735 
736         spin_lock_init(&ctx->ctx_lock);
737         spin_lock_init(&ctx->completion_lock);
738         mutex_init(&ctx->ring_lock);
739         /* Protect against page migration throughout kiotx setup by keeping
740          * the ring_lock mutex held until setup is complete. */
741         mutex_lock(&ctx->ring_lock);
742         init_waitqueue_head(&ctx->wait);
743 
744         INIT_LIST_HEAD(&ctx->active_reqs);
745 
746         if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
747                 goto err;
748 
749         if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
750                 goto err;
751 
752         ctx->cpu = alloc_percpu(struct kioctx_cpu);
753         if (!ctx->cpu)
754                 goto err;
755 
756         err = aio_setup_ring(ctx);
757         if (err < 0)
758                 goto err;
759 
760         atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
761         ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
762         if (ctx->req_batch < 1)
763                 ctx->req_batch = 1;
764 
765         /* limit the number of system wide aios */
766         spin_lock(&aio_nr_lock);
767         if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
768             aio_nr + nr_events < aio_nr) {
769                 spin_unlock(&aio_nr_lock);
770                 err = -EAGAIN;
771                 goto err_ctx;
772         }
773         aio_nr += ctx->max_reqs;
774         spin_unlock(&aio_nr_lock);
775 
776         percpu_ref_get(&ctx->users);    /* io_setup() will drop this ref */
777         percpu_ref_get(&ctx->reqs);     /* free_ioctx_users() will drop this */
778 
779         err = ioctx_add_table(ctx, mm);
780         if (err)
781                 goto err_cleanup;
782 
783         /* Release the ring_lock mutex now that all setup is complete. */
784         mutex_unlock(&ctx->ring_lock);
785 
786         pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
787                  ctx, ctx->user_id, mm, ctx->nr_events);
788         return ctx;
789 
790 err_cleanup:
791         aio_nr_sub(ctx->max_reqs);
792 err_ctx:
793         atomic_set(&ctx->dead, 1);
794         if (ctx->mmap_size)
795                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
796         aio_free_ring(ctx);
797 err:
798         mutex_unlock(&ctx->ring_lock);
799         free_percpu(ctx->cpu);
800         percpu_ref_exit(&ctx->reqs);
801         percpu_ref_exit(&ctx->users);
802         kmem_cache_free(kioctx_cachep, ctx);
803         pr_debug("error allocating ioctx %d\n", err);
804         return ERR_PTR(err);
805 }
806 
807 /* kill_ioctx
808  *      Cancels all outstanding aio requests on an aio context.  Used
809  *      when the processes owning a context have all exited to encourage
810  *      the rapid destruction of the kioctx.
811  */
812 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
813                       struct ctx_rq_wait *wait)
814 {
815         struct kioctx_table *table;
816 
817         spin_lock(&mm->ioctx_lock);
818         if (atomic_xchg(&ctx->dead, 1)) {
819                 spin_unlock(&mm->ioctx_lock);
820                 return -EINVAL;
821         }
822 
823         table = rcu_dereference_raw(mm->ioctx_table);
824         WARN_ON(ctx != table->table[ctx->id]);
825         table->table[ctx->id] = NULL;
826         spin_unlock(&mm->ioctx_lock);
827 
828         /* percpu_ref_kill() will do the necessary call_rcu() */
829         wake_up_all(&ctx->wait);
830 
831         /*
832          * It'd be more correct to do this in free_ioctx(), after all
833          * the outstanding kiocbs have finished - but by then io_destroy
834          * has already returned, so io_setup() could potentially return
835          * -EAGAIN with no ioctxs actually in use (as far as userspace
836          *  could tell).
837          */
838         aio_nr_sub(ctx->max_reqs);
839 
840         if (ctx->mmap_size)
841                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
842 
843         ctx->rq_wait = wait;
844         percpu_ref_kill(&ctx->users);
845         return 0;
846 }
847 
848 /*
849  * exit_aio: called when the last user of mm goes away.  At this point, there is
850  * no way for any new requests to be submited or any of the io_* syscalls to be
851  * called on the context.
852  *
853  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
854  * them.
855  */
856 void exit_aio(struct mm_struct *mm)
857 {
858         struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
859         struct ctx_rq_wait wait;
860         int i, skipped;
861 
862         if (!table)
863                 return;
864 
865         atomic_set(&wait.count, table->nr);
866         init_completion(&wait.comp);
867 
868         skipped = 0;
869         for (i = 0; i < table->nr; ++i) {
870                 struct kioctx *ctx = table->table[i];
871 
872                 if (!ctx) {
873                         skipped++;
874                         continue;
875                 }
876 
877                 /*
878                  * We don't need to bother with munmap() here - exit_mmap(mm)
879                  * is coming and it'll unmap everything. And we simply can't,
880                  * this is not necessarily our ->mm.
881                  * Since kill_ioctx() uses non-zero ->mmap_size as indicator
882                  * that it needs to unmap the area, just set it to 0.
883                  */
884                 ctx->mmap_size = 0;
885                 kill_ioctx(mm, ctx, &wait);
886         }
887 
888         if (!atomic_sub_and_test(skipped, &wait.count)) {
889                 /* Wait until all IO for the context are done. */
890                 wait_for_completion(&wait.comp);
891         }
892 
893         RCU_INIT_POINTER(mm->ioctx_table, NULL);
894         kfree(table);
895 }
896 
897 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
898 {
899         struct kioctx_cpu *kcpu;
900         unsigned long flags;
901 
902         local_irq_save(flags);
903         kcpu = this_cpu_ptr(ctx->cpu);
904         kcpu->reqs_available += nr;
905 
906         while (kcpu->reqs_available >= ctx->req_batch * 2) {
907                 kcpu->reqs_available -= ctx->req_batch;
908                 atomic_add(ctx->req_batch, &ctx->reqs_available);
909         }
910 
911         local_irq_restore(flags);
912 }
913 
914 static bool get_reqs_available(struct kioctx *ctx)
915 {
916         struct kioctx_cpu *kcpu;
917         bool ret = false;
918         unsigned long flags;
919 
920         local_irq_save(flags);
921         kcpu = this_cpu_ptr(ctx->cpu);
922         if (!kcpu->reqs_available) {
923                 int old, avail = atomic_read(&ctx->reqs_available);
924 
925                 do {
926                         if (avail < ctx->req_batch)
927                                 goto out;
928 
929                         old = avail;
930                         avail = atomic_cmpxchg(&ctx->reqs_available,
931                                                avail, avail - ctx->req_batch);
932                 } while (avail != old);
933 
934                 kcpu->reqs_available += ctx->req_batch;
935         }
936 
937         ret = true;
938         kcpu->reqs_available--;
939 out:
940         local_irq_restore(flags);
941         return ret;
942 }
943 
944 /* refill_reqs_available
945  *      Updates the reqs_available reference counts used for tracking the
946  *      number of free slots in the completion ring.  This can be called
947  *      from aio_complete() (to optimistically update reqs_available) or
948  *      from aio_get_req() (the we're out of events case).  It must be
949  *      called holding ctx->completion_lock.
950  */
951 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
952                                   unsigned tail)
953 {
954         unsigned events_in_ring, completed;
955 
956         /* Clamp head since userland can write to it. */
957         head %= ctx->nr_events;
958         if (head <= tail)
959                 events_in_ring = tail - head;
960         else
961                 events_in_ring = ctx->nr_events - (head - tail);
962 
963         completed = ctx->completed_events;
964         if (events_in_ring < completed)
965                 completed -= events_in_ring;
966         else
967                 completed = 0;
968 
969         if (!completed)
970                 return;
971 
972         ctx->completed_events -= completed;
973         put_reqs_available(ctx, completed);
974 }
975 
976 /* user_refill_reqs_available
977  *      Called to refill reqs_available when aio_get_req() encounters an
978  *      out of space in the completion ring.
979  */
980 static void user_refill_reqs_available(struct kioctx *ctx)
981 {
982         spin_lock_irq(&ctx->completion_lock);
983         if (ctx->completed_events) {
984                 struct aio_ring *ring;
985                 unsigned head;
986 
987                 /* Access of ring->head may race with aio_read_events_ring()
988                  * here, but that's okay since whether we read the old version
989                  * or the new version, and either will be valid.  The important
990                  * part is that head cannot pass tail since we prevent
991                  * aio_complete() from updating tail by holding
992                  * ctx->completion_lock.  Even if head is invalid, the check
993                  * against ctx->completed_events below will make sure we do the
994                  * safe/right thing.
995                  */
996                 ring = kmap_atomic(ctx->ring_pages[0]);
997                 head = ring->head;
998                 kunmap_atomic(ring);
999 
1000                 refill_reqs_available(ctx, head, ctx->tail);
1001         }
1002 
1003         spin_unlock_irq(&ctx->completion_lock);
1004 }
1005 
1006 /* aio_get_req
1007  *      Allocate a slot for an aio request.
1008  * Returns NULL if no requests are free.
1009  */
1010 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1011 {
1012         struct aio_kiocb *req;
1013 
1014         if (!get_reqs_available(ctx)) {
1015                 user_refill_reqs_available(ctx);
1016                 if (!get_reqs_available(ctx))
1017                         return NULL;
1018         }
1019 
1020         req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1021         if (unlikely(!req))
1022                 goto out_put;
1023 
1024         percpu_ref_get(&ctx->reqs);
1025 
1026         req->ki_ctx = ctx;
1027         return req;
1028 out_put:
1029         put_reqs_available(ctx, 1);
1030         return NULL;
1031 }
1032 
1033 static void kiocb_free(struct aio_kiocb *req)
1034 {
1035         if (req->common.ki_filp)
1036                 fput(req->common.ki_filp);
1037         if (req->ki_eventfd != NULL)
1038                 eventfd_ctx_put(req->ki_eventfd);
1039         kmem_cache_free(kiocb_cachep, req);
1040 }
1041 
1042 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1043 {
1044         struct aio_ring __user *ring  = (void __user *)ctx_id;
1045         struct mm_struct *mm = current->mm;
1046         struct kioctx *ctx, *ret = NULL;
1047         struct kioctx_table *table;
1048         unsigned id;
1049 
1050         if (get_user(id, &ring->id))
1051                 return NULL;
1052 
1053         rcu_read_lock();
1054         table = rcu_dereference(mm->ioctx_table);
1055 
1056         if (!table || id >= table->nr)
1057                 goto out;
1058 
1059         ctx = table->table[id];
1060         if (ctx && ctx->user_id == ctx_id) {
1061                 percpu_ref_get(&ctx->users);
1062                 ret = ctx;
1063         }
1064 out:
1065         rcu_read_unlock();
1066         return ret;
1067 }
1068 
1069 /* aio_complete
1070  *      Called when the io request on the given iocb is complete.
1071  */
1072 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1073 {
1074         struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1075         struct kioctx   *ctx = iocb->ki_ctx;
1076         struct aio_ring *ring;
1077         struct io_event *ev_page, *event;
1078         unsigned tail, pos, head;
1079         unsigned long   flags;
1080 
1081         if (kiocb->ki_flags & IOCB_WRITE) {
1082                 struct file *file = kiocb->ki_filp;
1083 
1084                 /*
1085                  * Tell lockdep we inherited freeze protection from submission
1086                  * thread.
1087                  */
1088                 if (S_ISREG(file_inode(file)->i_mode))
1089                         __sb_writers_acquired(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1090                 file_end_write(file);
1091         }
1092 
1093         /*
1094          * Special case handling for sync iocbs:
1095          *  - events go directly into the iocb for fast handling
1096          *  - the sync task with the iocb in its stack holds the single iocb
1097          *    ref, no other paths have a way to get another ref
1098          *  - the sync task helpfully left a reference to itself in the iocb
1099          */
1100         BUG_ON(is_sync_kiocb(kiocb));
1101 
1102         if (iocb->ki_list.next) {
1103                 unsigned long flags;
1104 
1105                 spin_lock_irqsave(&ctx->ctx_lock, flags);
1106                 list_del(&iocb->ki_list);
1107                 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1108         }
1109 
1110         /*
1111          * Add a completion event to the ring buffer. Must be done holding
1112          * ctx->completion_lock to prevent other code from messing with the tail
1113          * pointer since we might be called from irq context.
1114          */
1115         spin_lock_irqsave(&ctx->completion_lock, flags);
1116 
1117         tail = ctx->tail;
1118         pos = tail + AIO_EVENTS_OFFSET;
1119 
1120         if (++tail >= ctx->nr_events)
1121                 tail = 0;
1122 
1123         ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1124         event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1125 
1126         event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1127         event->data = iocb->ki_user_data;
1128         event->res = res;
1129         event->res2 = res2;
1130 
1131         kunmap_atomic(ev_page);
1132         flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1133 
1134         pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1135                  ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1136                  res, res2);
1137 
1138         /* after flagging the request as done, we
1139          * must never even look at it again
1140          */
1141         smp_wmb();      /* make event visible before updating tail */
1142 
1143         ctx->tail = tail;
1144 
1145         ring = kmap_atomic(ctx->ring_pages[0]);
1146         head = ring->head;
1147         ring->tail = tail;
1148         kunmap_atomic(ring);
1149         flush_dcache_page(ctx->ring_pages[0]);
1150 
1151         ctx->completed_events++;
1152         if (ctx->completed_events > 1)
1153                 refill_reqs_available(ctx, head, tail);
1154         spin_unlock_irqrestore(&ctx->completion_lock, flags);
1155 
1156         pr_debug("added to ring %p at [%u]\n", iocb, tail);
1157 
1158         /*
1159          * Check if the user asked us to deliver the result through an
1160          * eventfd. The eventfd_signal() function is safe to be called
1161          * from IRQ context.
1162          */
1163         if (iocb->ki_eventfd != NULL)
1164                 eventfd_signal(iocb->ki_eventfd, 1);
1165 
1166         /* everything turned out well, dispose of the aiocb. */
1167         kiocb_free(iocb);
1168 
1169         /*
1170          * We have to order our ring_info tail store above and test
1171          * of the wait list below outside the wait lock.  This is
1172          * like in wake_up_bit() where clearing a bit has to be
1173          * ordered with the unlocked test.
1174          */
1175         smp_mb();
1176 
1177         if (waitqueue_active(&ctx->wait))
1178                 wake_up(&ctx->wait);
1179 
1180         percpu_ref_put(&ctx->reqs);
1181 }
1182 
1183 /* aio_read_events_ring
1184  *      Pull an event off of the ioctx's event ring.  Returns the number of
1185  *      events fetched
1186  */
1187 static long aio_read_events_ring(struct kioctx *ctx,
1188                                  struct io_event __user *event, long nr)
1189 {
1190         struct aio_ring *ring;
1191         unsigned head, tail, pos;
1192         long ret = 0;
1193         int copy_ret;
1194 
1195         /*
1196          * The mutex can block and wake us up and that will cause
1197          * wait_event_interruptible_hrtimeout() to schedule without sleeping
1198          * and repeat. This should be rare enough that it doesn't cause
1199          * peformance issues. See the comment in read_events() for more detail.
1200          */
1201         sched_annotate_sleep();
1202         mutex_lock(&ctx->ring_lock);
1203 
1204         /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1205         ring = kmap_atomic(ctx->ring_pages[0]);
1206         head = ring->head;
1207         tail = ring->tail;
1208         kunmap_atomic(ring);
1209 
1210         /*
1211          * Ensure that once we've read the current tail pointer, that
1212          * we also see the events that were stored up to the tail.
1213          */
1214         smp_rmb();
1215 
1216         pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1217 
1218         if (head == tail)
1219                 goto out;
1220 
1221         head %= ctx->nr_events;
1222         tail %= ctx->nr_events;
1223 
1224         while (ret < nr) {
1225                 long avail;
1226                 struct io_event *ev;
1227                 struct page *page;
1228 
1229                 avail = (head <= tail ?  tail : ctx->nr_events) - head;
1230                 if (head == tail)
1231                         break;
1232 
1233                 avail = min(avail, nr - ret);
1234                 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1235                             ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1236 
1237                 pos = head + AIO_EVENTS_OFFSET;
1238                 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1239                 pos %= AIO_EVENTS_PER_PAGE;
1240 
1241                 ev = kmap(page);
1242                 copy_ret = copy_to_user(event + ret, ev + pos,
1243                                         sizeof(*ev) * avail);
1244                 kunmap(page);
1245 
1246                 if (unlikely(copy_ret)) {
1247                         ret = -EFAULT;
1248                         goto out;
1249                 }
1250 
1251                 ret += avail;
1252                 head += avail;
1253                 head %= ctx->nr_events;
1254         }
1255 
1256         ring = kmap_atomic(ctx->ring_pages[0]);
1257         ring->head = head;
1258         kunmap_atomic(ring);
1259         flush_dcache_page(ctx->ring_pages[0]);
1260 
1261         pr_debug("%li  h%u t%u\n", ret, head, tail);
1262 out:
1263         mutex_unlock(&ctx->ring_lock);
1264 
1265         return ret;
1266 }
1267 
1268 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1269                             struct io_event __user *event, long *i)
1270 {
1271         long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1272 
1273         if (ret > 0)
1274                 *i += ret;
1275 
1276         if (unlikely(atomic_read(&ctx->dead)))
1277                 ret = -EINVAL;
1278 
1279         if (!*i)
1280                 *i = ret;
1281 
1282         return ret < 0 || *i >= min_nr;
1283 }
1284 
1285 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1286                         struct io_event __user *event,
1287                         struct timespec __user *timeout)
1288 {
1289         ktime_t until = KTIME_MAX;
1290         long ret = 0;
1291 
1292         if (timeout) {
1293                 struct timespec ts;
1294 
1295                 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1296                         return -EFAULT;
1297 
1298                 until = timespec_to_ktime(ts);
1299         }
1300 
1301         /*
1302          * Note that aio_read_events() is being called as the conditional - i.e.
1303          * we're calling it after prepare_to_wait() has set task state to
1304          * TASK_INTERRUPTIBLE.
1305          *
1306          * But aio_read_events() can block, and if it blocks it's going to flip
1307          * the task state back to TASK_RUNNING.
1308          *
1309          * This should be ok, provided it doesn't flip the state back to
1310          * TASK_RUNNING and return 0 too much - that causes us to spin. That
1311          * will only happen if the mutex_lock() call blocks, and we then find
1312          * the ringbuffer empty. So in practice we should be ok, but it's
1313          * something to be aware of when touching this code.
1314          */
1315         if (until == 0)
1316                 aio_read_events(ctx, min_nr, nr, event, &ret);
1317         else
1318                 wait_event_interruptible_hrtimeout(ctx->wait,
1319                                 aio_read_events(ctx, min_nr, nr, event, &ret),
1320                                 until);
1321 
1322         if (!ret && signal_pending(current))
1323                 ret = -EINTR;
1324 
1325         return ret;
1326 }
1327 
1328 /* sys_io_setup:
1329  *      Create an aio_context capable of receiving at least nr_events.
1330  *      ctxp must not point to an aio_context that already exists, and
1331  *      must be initialized to 0 prior to the call.  On successful
1332  *      creation of the aio_context, *ctxp is filled in with the resulting 
1333  *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1334  *      if the specified nr_events exceeds internal limits.  May fail 
1335  *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1336  *      of available events.  May fail with -ENOMEM if insufficient kernel
1337  *      resources are available.  May fail with -EFAULT if an invalid
1338  *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1339  *      implemented.
1340  */
1341 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1342 {
1343         struct kioctx *ioctx = NULL;
1344         unsigned long ctx;
1345         long ret;
1346 
1347         ret = get_user(ctx, ctxp);
1348         if (unlikely(ret))
1349                 goto out;
1350 
1351         ret = -EINVAL;
1352         if (unlikely(ctx || nr_events == 0)) {
1353                 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1354                          ctx, nr_events);
1355                 goto out;
1356         }
1357 
1358         ioctx = ioctx_alloc(nr_events);
1359         ret = PTR_ERR(ioctx);
1360         if (!IS_ERR(ioctx)) {
1361                 ret = put_user(ioctx->user_id, ctxp);
1362                 if (ret)
1363                         kill_ioctx(current->mm, ioctx, NULL);
1364                 percpu_ref_put(&ioctx->users);
1365         }
1366 
1367 out:
1368         return ret;
1369 }
1370 
1371 #ifdef CONFIG_COMPAT
1372 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1373 {
1374         struct kioctx *ioctx = NULL;
1375         unsigned long ctx;
1376         long ret;
1377 
1378         ret = get_user(ctx, ctx32p);
1379         if (unlikely(ret))
1380                 goto out;
1381 
1382         ret = -EINVAL;
1383         if (unlikely(ctx || nr_events == 0)) {
1384                 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1385                          ctx, nr_events);
1386                 goto out;
1387         }
1388 
1389         ioctx = ioctx_alloc(nr_events);
1390         ret = PTR_ERR(ioctx);
1391         if (!IS_ERR(ioctx)) {
1392                 /* truncating is ok because it's a user address */
1393                 ret = put_user((u32)ioctx->user_id, ctx32p);
1394                 if (ret)
1395                         kill_ioctx(current->mm, ioctx, NULL);
1396                 percpu_ref_put(&ioctx->users);
1397         }
1398 
1399 out:
1400         return ret;
1401 }
1402 #endif
1403 
1404 /* sys_io_destroy:
1405  *      Destroy the aio_context specified.  May cancel any outstanding 
1406  *      AIOs and block on completion.  Will fail with -ENOSYS if not
1407  *      implemented.  May fail with -EINVAL if the context pointed to
1408  *      is invalid.
1409  */
1410 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1411 {
1412         struct kioctx *ioctx = lookup_ioctx(ctx);
1413         if (likely(NULL != ioctx)) {
1414                 struct ctx_rq_wait wait;
1415                 int ret;
1416 
1417                 init_completion(&wait.comp);
1418                 atomic_set(&wait.count, 1);
1419 
1420                 /* Pass requests_done to kill_ioctx() where it can be set
1421                  * in a thread-safe way. If we try to set it here then we have
1422                  * a race condition if two io_destroy() called simultaneously.
1423                  */
1424                 ret = kill_ioctx(current->mm, ioctx, &wait);
1425                 percpu_ref_put(&ioctx->users);
1426 
1427                 /* Wait until all IO for the context are done. Otherwise kernel
1428                  * keep using user-space buffers even if user thinks the context
1429                  * is destroyed.
1430                  */
1431                 if (!ret)
1432                         wait_for_completion(&wait.comp);
1433 
1434                 return ret;
1435         }
1436         pr_debug("EINVAL: invalid context id\n");
1437         return -EINVAL;
1438 }
1439 
1440 static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec,
1441                 bool vectored, bool compat, struct iov_iter *iter)
1442 {
1443         void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1444         size_t len = iocb->aio_nbytes;
1445 
1446         if (!vectored) {
1447                 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1448                 *iovec = NULL;
1449                 return ret;
1450         }
1451 #ifdef CONFIG_COMPAT
1452         if (compat)
1453                 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1454                                 iter);
1455 #endif
1456         return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1457 }
1458 
1459 static inline ssize_t aio_ret(struct kiocb *req, ssize_t ret)
1460 {
1461         switch (ret) {
1462         case -EIOCBQUEUED:
1463                 return ret;
1464         case -ERESTARTSYS:
1465         case -ERESTARTNOINTR:
1466         case -ERESTARTNOHAND:
1467         case -ERESTART_RESTARTBLOCK:
1468                 /*
1469                  * There's no easy way to restart the syscall since other AIO's
1470                  * may be already running. Just fail this IO with EINTR.
1471                  */
1472                 ret = -EINTR;
1473                 /*FALLTHRU*/
1474         default:
1475                 aio_complete(req, ret, 0);
1476                 return 0;
1477         }
1478 }
1479 
1480 static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
1481                 bool compat)
1482 {
1483         struct file *file = req->ki_filp;
1484         struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1485         struct iov_iter iter;
1486         ssize_t ret;
1487 
1488         if (unlikely(!(file->f_mode & FMODE_READ)))
1489                 return -EBADF;
1490         if (unlikely(!file->f_op->read_iter))
1491                 return -EINVAL;
1492 
1493         ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1494         if (ret)
1495                 return ret;
1496         ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1497         if (!ret)
1498                 ret = aio_ret(req, call_read_iter(file, req, &iter));
1499         kfree(iovec);
1500         return ret;
1501 }
1502 
1503 static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored,
1504                 bool compat)
1505 {
1506         struct file *file = req->ki_filp;
1507         struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1508         struct iov_iter iter;
1509         ssize_t ret;
1510 
1511         if (unlikely(!(file->f_mode & FMODE_WRITE)))
1512                 return -EBADF;
1513         if (unlikely(!file->f_op->write_iter))
1514                 return -EINVAL;
1515 
1516         ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1517         if (ret)
1518                 return ret;
1519         ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1520         if (!ret) {
1521                 req->ki_flags |= IOCB_WRITE;
1522                 file_start_write(file);
1523                 ret = aio_ret(req, call_write_iter(file, req, &iter));
1524                 /*
1525                  * We release freeze protection in aio_complete().  Fool lockdep
1526                  * by telling it the lock got released so that it doesn't
1527                  * complain about held lock when we return to userspace.
1528                  */
1529                 if (S_ISREG(file_inode(file)->i_mode))
1530                         __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1531         }
1532         kfree(iovec);
1533         return ret;
1534 }
1535 
1536 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1537                          struct iocb *iocb, bool compat)
1538 {
1539         struct aio_kiocb *req;
1540         struct file *file;
1541         ssize_t ret;
1542 
1543         /* enforce forwards compatibility on users */
1544         if (unlikely(iocb->aio_reserved2)) {
1545                 pr_debug("EINVAL: reserve field set\n");
1546                 return -EINVAL;
1547         }
1548 
1549         /* prevent overflows */
1550         if (unlikely(
1551             (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1552             (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1553             ((ssize_t)iocb->aio_nbytes < 0)
1554            )) {
1555                 pr_debug("EINVAL: overflow check\n");
1556                 return -EINVAL;
1557         }
1558 
1559         req = aio_get_req(ctx);
1560         if (unlikely(!req))
1561                 return -EAGAIN;
1562 
1563         req->common.ki_filp = file = fget(iocb->aio_fildes);
1564         if (unlikely(!req->common.ki_filp)) {
1565                 ret = -EBADF;
1566                 goto out_put_req;
1567         }
1568         req->common.ki_pos = iocb->aio_offset;
1569         req->common.ki_complete = aio_complete;
1570         req->common.ki_flags = iocb_flags(req->common.ki_filp);
1571         req->common.ki_hint = file_write_hint(file);
1572 
1573         if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1574                 /*
1575                  * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1576                  * instance of the file* now. The file descriptor must be
1577                  * an eventfd() fd, and will be signaled for each completed
1578                  * event using the eventfd_signal() function.
1579                  */
1580                 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1581                 if (IS_ERR(req->ki_eventfd)) {
1582                         ret = PTR_ERR(req->ki_eventfd);
1583                         req->ki_eventfd = NULL;
1584                         goto out_put_req;
1585                 }
1586 
1587                 req->common.ki_flags |= IOCB_EVENTFD;
1588         }
1589 
1590         ret = kiocb_set_rw_flags(&req->common, iocb->aio_rw_flags);
1591         if (unlikely(ret)) {
1592                 pr_debug("EINVAL: aio_rw_flags\n");
1593                 goto out_put_req;
1594         }
1595 
1596         if ((req->common.ki_flags & IOCB_NOWAIT) &&
1597                         !(req->common.ki_flags & IOCB_DIRECT)) {
1598                 ret = -EOPNOTSUPP;
1599                 goto out_put_req;
1600         }
1601 
1602         ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1603         if (unlikely(ret)) {
1604                 pr_debug("EFAULT: aio_key\n");
1605                 goto out_put_req;
1606         }
1607 
1608         req->ki_user_iocb = user_iocb;
1609         req->ki_user_data = iocb->aio_data;
1610 
1611         get_file(file);
1612         switch (iocb->aio_lio_opcode) {
1613         case IOCB_CMD_PREAD:
1614                 ret = aio_read(&req->common, iocb, false, compat);
1615                 break;
1616         case IOCB_CMD_PWRITE:
1617                 ret = aio_write(&req->common, iocb, false, compat);
1618                 break;
1619         case IOCB_CMD_PREADV:
1620                 ret = aio_read(&req->common, iocb, true, compat);
1621                 break;
1622         case IOCB_CMD_PWRITEV:
1623                 ret = aio_write(&req->common, iocb, true, compat);
1624                 break;
1625         default:
1626                 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1627                 ret = -EINVAL;
1628                 break;
1629         }
1630         fput(file);
1631 
1632         if (ret && ret != -EIOCBQUEUED)
1633                 goto out_put_req;
1634         return 0;
1635 out_put_req:
1636         put_reqs_available(ctx, 1);
1637         percpu_ref_put(&ctx->reqs);
1638         kiocb_free(req);
1639         return ret;
1640 }
1641 
1642 static long do_io_submit(aio_context_t ctx_id, long nr,
1643                           struct iocb __user *__user *iocbpp, bool compat)
1644 {
1645         struct kioctx *ctx;
1646         long ret = 0;
1647         int i = 0;
1648         struct blk_plug plug;
1649 
1650         if (unlikely(nr < 0))
1651                 return -EINVAL;
1652 
1653         if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1654                 nr = LONG_MAX/sizeof(*iocbpp);
1655 
1656         if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1657                 return -EFAULT;
1658 
1659         ctx = lookup_ioctx(ctx_id);
1660         if (unlikely(!ctx)) {
1661                 pr_debug("EINVAL: invalid context id\n");
1662                 return -EINVAL;
1663         }
1664 
1665         blk_start_plug(&plug);
1666 
1667         /*
1668          * AKPM: should this return a partial result if some of the IOs were
1669          * successfully submitted?
1670          */
1671         for (i=0; i<nr; i++) {
1672                 struct iocb __user *user_iocb;
1673                 struct iocb tmp;
1674 
1675                 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1676                         ret = -EFAULT;
1677                         break;
1678                 }
1679 
1680                 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1681                         ret = -EFAULT;
1682                         break;
1683                 }
1684 
1685                 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1686                 if (ret)
1687                         break;
1688         }
1689         blk_finish_plug(&plug);
1690 
1691         percpu_ref_put(&ctx->users);
1692         return i ? i : ret;
1693 }
1694 
1695 /* sys_io_submit:
1696  *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1697  *      the number of iocbs queued.  May return -EINVAL if the aio_context
1698  *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1699  *      *iocbpp[0] is not properly initialized, if the operation specified
1700  *      is invalid for the file descriptor in the iocb.  May fail with
1701  *      -EFAULT if any of the data structures point to invalid data.  May
1702  *      fail with -EBADF if the file descriptor specified in the first
1703  *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1704  *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1705  *      fail with -ENOSYS if not implemented.
1706  */
1707 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1708                 struct iocb __user * __user *, iocbpp)
1709 {
1710         return do_io_submit(ctx_id, nr, iocbpp, 0);
1711 }
1712 
1713 #ifdef CONFIG_COMPAT
1714 static inline long
1715 copy_iocb(long nr, u32 __user *ptr32, struct iocb __user * __user *ptr64)
1716 {
1717         compat_uptr_t uptr;
1718         int i;
1719 
1720         for (i = 0; i < nr; ++i) {
1721                 if (get_user(uptr, ptr32 + i))
1722                         return -EFAULT;
1723                 if (put_user(compat_ptr(uptr), ptr64 + i))
1724                         return -EFAULT;
1725         }
1726         return 0;
1727 }
1728 
1729 #define MAX_AIO_SUBMITS         (PAGE_SIZE/sizeof(struct iocb *))
1730 
1731 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1732                        int, nr, u32 __user *, iocb)
1733 {
1734         struct iocb __user * __user *iocb64;
1735         long ret;
1736 
1737         if (unlikely(nr < 0))
1738                 return -EINVAL;
1739 
1740         if (nr > MAX_AIO_SUBMITS)
1741                 nr = MAX_AIO_SUBMITS;
1742 
1743         iocb64 = compat_alloc_user_space(nr * sizeof(*iocb64));
1744         ret = copy_iocb(nr, iocb, iocb64);
1745         if (!ret)
1746                 ret = do_io_submit(ctx_id, nr, iocb64, 1);
1747         return ret;
1748 }
1749 #endif
1750 
1751 /* lookup_kiocb
1752  *      Finds a given iocb for cancellation.
1753  */
1754 static struct aio_kiocb *
1755 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1756 {
1757         struct aio_kiocb *kiocb;
1758 
1759         assert_spin_locked(&ctx->ctx_lock);
1760 
1761         if (key != KIOCB_KEY)
1762                 return NULL;
1763 
1764         /* TODO: use a hash or array, this sucks. */
1765         list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1766                 if (kiocb->ki_user_iocb == iocb)
1767                         return kiocb;
1768         }
1769         return NULL;
1770 }
1771 
1772 /* sys_io_cancel:
1773  *      Attempts to cancel an iocb previously passed to io_submit.  If
1774  *      the operation is successfully cancelled, the resulting event is
1775  *      copied into the memory pointed to by result without being placed
1776  *      into the completion queue and 0 is returned.  May fail with
1777  *      -EFAULT if any of the data structures pointed to are invalid.
1778  *      May fail with -EINVAL if aio_context specified by ctx_id is
1779  *      invalid.  May fail with -EAGAIN if the iocb specified was not
1780  *      cancelled.  Will fail with -ENOSYS if not implemented.
1781  */
1782 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1783                 struct io_event __user *, result)
1784 {
1785         struct kioctx *ctx;
1786         struct aio_kiocb *kiocb;
1787         u32 key;
1788         int ret;
1789 
1790         ret = get_user(key, &iocb->aio_key);
1791         if (unlikely(ret))
1792                 return -EFAULT;
1793 
1794         ctx = lookup_ioctx(ctx_id);
1795         if (unlikely(!ctx))
1796                 return -EINVAL;
1797 
1798         spin_lock_irq(&ctx->ctx_lock);
1799 
1800         kiocb = lookup_kiocb(ctx, iocb, key);
1801         if (kiocb)
1802                 ret = kiocb_cancel(kiocb);
1803         else
1804                 ret = -EINVAL;
1805 
1806         spin_unlock_irq(&ctx->ctx_lock);
1807 
1808         if (!ret) {
1809                 /*
1810                  * The result argument is no longer used - the io_event is
1811                  * always delivered via the ring buffer. -EINPROGRESS indicates
1812                  * cancellation is progress:
1813                  */
1814                 ret = -EINPROGRESS;
1815         }
1816 
1817         percpu_ref_put(&ctx->users);
1818 
1819         return ret;
1820 }
1821 
1822 /* io_getevents:
1823  *      Attempts to read at least min_nr events and up to nr events from
1824  *      the completion queue for the aio_context specified by ctx_id. If
1825  *      it succeeds, the number of read events is returned. May fail with
1826  *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1827  *      out of range, if timeout is out of range.  May fail with -EFAULT
1828  *      if any of the memory specified is invalid.  May return 0 or
1829  *      < min_nr if the timeout specified by timeout has elapsed
1830  *      before sufficient events are available, where timeout == NULL
1831  *      specifies an infinite timeout. Note that the timeout pointed to by
1832  *      timeout is relative.  Will fail with -ENOSYS if not implemented.
1833  */
1834 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1835                 long, min_nr,
1836                 long, nr,
1837                 struct io_event __user *, events,
1838                 struct timespec __user *, timeout)
1839 {
1840         struct kioctx *ioctx = lookup_ioctx(ctx_id);
1841         long ret = -EINVAL;
1842 
1843         if (likely(ioctx)) {
1844                 if (likely(min_nr <= nr && min_nr >= 0))
1845                         ret = read_events(ioctx, min_nr, nr, events, timeout);
1846                 percpu_ref_put(&ioctx->users);
1847         }
1848         return ret;
1849 }
1850 
1851 #ifdef CONFIG_COMPAT
1852 COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id,
1853                        compat_long_t, min_nr,
1854                        compat_long_t, nr,
1855                        struct io_event __user *, events,
1856                        struct compat_timespec __user *, timeout)
1857 {
1858         struct timespec t;
1859         struct timespec __user *ut = NULL;
1860 
1861         if (timeout) {
1862                 if (compat_get_timespec(&t, timeout))
1863                         return -EFAULT;
1864 
1865                 ut = compat_alloc_user_space(sizeof(*ut));
1866                 if (copy_to_user(ut, &t, sizeof(t)))
1867                         return -EFAULT;
1868         }
1869         return sys_io_getevents(ctx_id, min_nr, nr, events, ut);
1870 }
1871 #endif
1872 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp