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

TOMOYO Linux Cross Reference
Linux/fs/aio.c

Version: ~ [ linux-5.16-rc1 ] ~ [ linux-5.15.2 ] ~ [ linux-5.14.18 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.79 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.159 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.217 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.255 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.290 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.292 ] ~ [ 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  *      Copyright 2018 Christoph Hellwig.
  9  *
 10  *      See ../COPYING for licensing terms.
 11  */
 12 #define pr_fmt(fmt) "%s: " fmt, __func__
 13 
 14 #include <linux/kernel.h>
 15 #include <linux/init.h>
 16 #include <linux/errno.h>
 17 #include <linux/time.h>
 18 #include <linux/aio_abi.h>
 19 #include <linux/export.h>
 20 #include <linux/syscalls.h>
 21 #include <linux/backing-dev.h>
 22 #include <linux/refcount.h>
 23 #include <linux/uio.h>
 24 
 25 #include <linux/sched/signal.h>
 26 #include <linux/fs.h>
 27 #include <linux/file.h>
 28 #include <linux/mm.h>
 29 #include <linux/mman.h>
 30 #include <linux/mmu_context.h>
 31 #include <linux/percpu.h>
 32 #include <linux/slab.h>
 33 #include <linux/timer.h>
 34 #include <linux/aio.h>
 35 #include <linux/highmem.h>
 36 #include <linux/workqueue.h>
 37 #include <linux/security.h>
 38 #include <linux/eventfd.h>
 39 #include <linux/blkdev.h>
 40 #include <linux/compat.h>
 41 #include <linux/migrate.h>
 42 #include <linux/ramfs.h>
 43 #include <linux/percpu-refcount.h>
 44 #include <linux/mount.h>
 45 #include <linux/pseudo_fs.h>
 46 
 47 #include <asm/kmap_types.h>
 48 #include <linux/uaccess.h>
 49 #include <linux/nospec.h>
 50 
 51 #include "internal.h"
 52 
 53 #define KIOCB_KEY               0
 54 
 55 #define AIO_RING_MAGIC                  0xa10a10a1
 56 #define AIO_RING_COMPAT_FEATURES        1
 57 #define AIO_RING_INCOMPAT_FEATURES      0
 58 struct aio_ring {
 59         unsigned        id;     /* kernel internal index number */
 60         unsigned        nr;     /* number of io_events */
 61         unsigned        head;   /* Written to by userland or under ring_lock
 62                                  * mutex by aio_read_events_ring(). */
 63         unsigned        tail;
 64 
 65         unsigned        magic;
 66         unsigned        compat_features;
 67         unsigned        incompat_features;
 68         unsigned        header_length;  /* size of aio_ring */
 69 
 70 
 71         struct io_event         io_events[0];
 72 }; /* 128 bytes + ring size */
 73 
 74 /*
 75  * Plugging is meant to work with larger batches of IOs. If we don't
 76  * have more than the below, then don't bother setting up a plug.
 77  */
 78 #define AIO_PLUG_THRESHOLD      2
 79 
 80 #define AIO_RING_PAGES  8
 81 
 82 struct kioctx_table {
 83         struct rcu_head         rcu;
 84         unsigned                nr;
 85         struct kioctx __rcu     *table[];
 86 };
 87 
 88 struct kioctx_cpu {
 89         unsigned                reqs_available;
 90 };
 91 
 92 struct ctx_rq_wait {
 93         struct completion comp;
 94         atomic_t count;
 95 };
 96 
 97 struct kioctx {
 98         struct percpu_ref       users;
 99         atomic_t                dead;
100 
101         struct percpu_ref       reqs;
102 
103         unsigned long           user_id;
104 
105         struct __percpu kioctx_cpu *cpu;
106 
107         /*
108          * For percpu reqs_available, number of slots we move to/from global
109          * counter at a time:
110          */
111         unsigned                req_batch;
112         /*
113          * This is what userspace passed to io_setup(), it's not used for
114          * anything but counting against the global max_reqs quota.
115          *
116          * The real limit is nr_events - 1, which will be larger (see
117          * aio_setup_ring())
118          */
119         unsigned                max_reqs;
120 
121         /* Size of ringbuffer, in units of struct io_event */
122         unsigned                nr_events;
123 
124         unsigned long           mmap_base;
125         unsigned long           mmap_size;
126 
127         struct page             **ring_pages;
128         long                    nr_pages;
129 
130         struct rcu_work         free_rwork;     /* see free_ioctx() */
131 
132         /*
133          * signals when all in-flight requests are done
134          */
135         struct ctx_rq_wait      *rq_wait;
136 
137         struct {
138                 /*
139                  * This counts the number of available slots in the ringbuffer,
140                  * so we avoid overflowing it: it's decremented (if positive)
141                  * when allocating a kiocb and incremented when the resulting
142                  * io_event is pulled off the ringbuffer.
143                  *
144                  * We batch accesses to it with a percpu version.
145                  */
146                 atomic_t        reqs_available;
147         } ____cacheline_aligned_in_smp;
148 
149         struct {
150                 spinlock_t      ctx_lock;
151                 struct list_head active_reqs;   /* used for cancellation */
152         } ____cacheline_aligned_in_smp;
153 
154         struct {
155                 struct mutex    ring_lock;
156                 wait_queue_head_t wait;
157         } ____cacheline_aligned_in_smp;
158 
159         struct {
160                 unsigned        tail;
161                 unsigned        completed_events;
162                 spinlock_t      completion_lock;
163         } ____cacheline_aligned_in_smp;
164 
165         struct page             *internal_pages[AIO_RING_PAGES];
166         struct file             *aio_ring_file;
167 
168         unsigned                id;
169 };
170 
171 /*
172  * First field must be the file pointer in all the
173  * iocb unions! See also 'struct kiocb' in <linux/fs.h>
174  */
175 struct fsync_iocb {
176         struct file             *file;
177         struct work_struct      work;
178         bool                    datasync;
179 };
180 
181 struct poll_iocb {
182         struct file             *file;
183         struct wait_queue_head  *head;
184         __poll_t                events;
185         bool                    done;
186         bool                    cancelled;
187         struct wait_queue_entry wait;
188         struct work_struct      work;
189 };
190 
191 /*
192  * NOTE! Each of the iocb union members has the file pointer
193  * as the first entry in their struct definition. So you can
194  * access the file pointer through any of the sub-structs,
195  * or directly as just 'ki_filp' in this struct.
196  */
197 struct aio_kiocb {
198         union {
199                 struct file             *ki_filp;
200                 struct kiocb            rw;
201                 struct fsync_iocb       fsync;
202                 struct poll_iocb        poll;
203         };
204 
205         struct kioctx           *ki_ctx;
206         kiocb_cancel_fn         *ki_cancel;
207 
208         struct io_event         ki_res;
209 
210         struct list_head        ki_list;        /* the aio core uses this
211                                                  * for cancellation */
212         refcount_t              ki_refcnt;
213 
214         /*
215          * If the aio_resfd field of the userspace iocb is not zero,
216          * this is the underlying eventfd context to deliver events to.
217          */
218         struct eventfd_ctx      *ki_eventfd;
219 };
220 
221 /*------ sysctl variables----*/
222 static DEFINE_SPINLOCK(aio_nr_lock);
223 unsigned long aio_nr;           /* current system wide number of aio requests */
224 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/
226 
227 static struct kmem_cache        *kiocb_cachep;
228 static struct kmem_cache        *kioctx_cachep;
229 
230 static struct vfsmount *aio_mnt;
231 
232 static const struct file_operations aio_ring_fops;
233 static const struct address_space_operations aio_ctx_aops;
234 
235 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
236 {
237         struct file *file;
238         struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
239         if (IS_ERR(inode))
240                 return ERR_CAST(inode);
241 
242         inode->i_mapping->a_ops = &aio_ctx_aops;
243         inode->i_mapping->private_data = ctx;
244         inode->i_size = PAGE_SIZE * nr_pages;
245 
246         file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
247                                 O_RDWR, &aio_ring_fops);
248         if (IS_ERR(file))
249                 iput(inode);
250         return file;
251 }
252 
253 static int aio_init_fs_context(struct fs_context *fc)
254 {
255         if (!init_pseudo(fc, AIO_RING_MAGIC))
256                 return -ENOMEM;
257         fc->s_iflags |= SB_I_NOEXEC;
258         return 0;
259 }
260 
261 /* aio_setup
262  *      Creates the slab caches used by the aio routines, panic on
263  *      failure as this is done early during the boot sequence.
264  */
265 static int __init aio_setup(void)
266 {
267         static struct file_system_type aio_fs = {
268                 .name           = "aio",
269                 .init_fs_context = aio_init_fs_context,
270                 .kill_sb        = kill_anon_super,
271         };
272         aio_mnt = kern_mount(&aio_fs);
273         if (IS_ERR(aio_mnt))
274                 panic("Failed to create aio fs mount.");
275 
276         kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
277         kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
278         return 0;
279 }
280 __initcall(aio_setup);
281 
282 static void put_aio_ring_file(struct kioctx *ctx)
283 {
284         struct file *aio_ring_file = ctx->aio_ring_file;
285         struct address_space *i_mapping;
286 
287         if (aio_ring_file) {
288                 truncate_setsize(file_inode(aio_ring_file), 0);
289 
290                 /* Prevent further access to the kioctx from migratepages */
291                 i_mapping = aio_ring_file->f_mapping;
292                 spin_lock(&i_mapping->private_lock);
293                 i_mapping->private_data = NULL;
294                 ctx->aio_ring_file = NULL;
295                 spin_unlock(&i_mapping->private_lock);
296 
297                 fput(aio_ring_file);
298         }
299 }
300 
301 static void aio_free_ring(struct kioctx *ctx)
302 {
303         int i;
304 
305         /* Disconnect the kiotx from the ring file.  This prevents future
306          * accesses to the kioctx from page migration.
307          */
308         put_aio_ring_file(ctx);
309 
310         for (i = 0; i < ctx->nr_pages; i++) {
311                 struct page *page;
312                 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
313                                 page_count(ctx->ring_pages[i]));
314                 page = ctx->ring_pages[i];
315                 if (!page)
316                         continue;
317                 ctx->ring_pages[i] = NULL;
318                 put_page(page);
319         }
320 
321         if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
322                 kfree(ctx->ring_pages);
323                 ctx->ring_pages = NULL;
324         }
325 }
326 
327 static int aio_ring_mremap(struct vm_area_struct *vma)
328 {
329         struct file *file = vma->vm_file;
330         struct mm_struct *mm = vma->vm_mm;
331         struct kioctx_table *table;
332         int i, res = -EINVAL;
333 
334         spin_lock(&mm->ioctx_lock);
335         rcu_read_lock();
336         table = rcu_dereference(mm->ioctx_table);
337         for (i = 0; i < table->nr; i++) {
338                 struct kioctx *ctx;
339 
340                 ctx = rcu_dereference(table->table[i]);
341                 if (ctx && ctx->aio_ring_file == file) {
342                         if (!atomic_read(&ctx->dead)) {
343                                 ctx->user_id = ctx->mmap_base = vma->vm_start;
344                                 res = 0;
345                         }
346                         break;
347                 }
348         }
349 
350         rcu_read_unlock();
351         spin_unlock(&mm->ioctx_lock);
352         return res;
353 }
354 
355 static const struct vm_operations_struct aio_ring_vm_ops = {
356         .mremap         = aio_ring_mremap,
357 #if IS_ENABLED(CONFIG_MMU)
358         .fault          = filemap_fault,
359         .map_pages      = filemap_map_pages,
360         .page_mkwrite   = filemap_page_mkwrite,
361 #endif
362 };
363 
364 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
365 {
366         vma->vm_flags |= VM_DONTEXPAND;
367         vma->vm_ops = &aio_ring_vm_ops;
368         return 0;
369 }
370 
371 static const struct file_operations aio_ring_fops = {
372         .mmap = aio_ring_mmap,
373 };
374 
375 #if IS_ENABLED(CONFIG_MIGRATION)
376 static int aio_migratepage(struct address_space *mapping, struct page *new,
377                         struct page *old, enum migrate_mode mode)
378 {
379         struct kioctx *ctx;
380         unsigned long flags;
381         pgoff_t idx;
382         int rc;
383 
384         /*
385          * We cannot support the _NO_COPY case here, because copy needs to
386          * happen under the ctx->completion_lock. That does not work with the
387          * migration workflow of MIGRATE_SYNC_NO_COPY.
388          */
389         if (mode == MIGRATE_SYNC_NO_COPY)
390                 return -EINVAL;
391 
392         rc = 0;
393 
394         /* mapping->private_lock here protects against the kioctx teardown.  */
395         spin_lock(&mapping->private_lock);
396         ctx = mapping->private_data;
397         if (!ctx) {
398                 rc = -EINVAL;
399                 goto out;
400         }
401 
402         /* The ring_lock mutex.  The prevents aio_read_events() from writing
403          * to the ring's head, and prevents page migration from mucking in
404          * a partially initialized kiotx.
405          */
406         if (!mutex_trylock(&ctx->ring_lock)) {
407                 rc = -EAGAIN;
408                 goto out;
409         }
410 
411         idx = old->index;
412         if (idx < (pgoff_t)ctx->nr_pages) {
413                 /* Make sure the old page hasn't already been changed */
414                 if (ctx->ring_pages[idx] != old)
415                         rc = -EAGAIN;
416         } else
417                 rc = -EINVAL;
418 
419         if (rc != 0)
420                 goto out_unlock;
421 
422         /* Writeback must be complete */
423         BUG_ON(PageWriteback(old));
424         get_page(new);
425 
426         rc = migrate_page_move_mapping(mapping, new, old, 1);
427         if (rc != MIGRATEPAGE_SUCCESS) {
428                 put_page(new);
429                 goto out_unlock;
430         }
431 
432         /* Take completion_lock to prevent other writes to the ring buffer
433          * while the old page is copied to the new.  This prevents new
434          * events from being lost.
435          */
436         spin_lock_irqsave(&ctx->completion_lock, flags);
437         migrate_page_copy(new, old);
438         BUG_ON(ctx->ring_pages[idx] != old);
439         ctx->ring_pages[idx] = new;
440         spin_unlock_irqrestore(&ctx->completion_lock, flags);
441 
442         /* The old page is no longer accessible. */
443         put_page(old);
444 
445 out_unlock:
446         mutex_unlock(&ctx->ring_lock);
447 out:
448         spin_unlock(&mapping->private_lock);
449         return rc;
450 }
451 #endif
452 
453 static const struct address_space_operations aio_ctx_aops = {
454         .set_page_dirty = __set_page_dirty_no_writeback,
455 #if IS_ENABLED(CONFIG_MIGRATION)
456         .migratepage    = aio_migratepage,
457 #endif
458 };
459 
460 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
461 {
462         struct aio_ring *ring;
463         struct mm_struct *mm = current->mm;
464         unsigned long size, unused;
465         int nr_pages;
466         int i;
467         struct file *file;
468 
469         /* Compensate for the ring buffer's head/tail overlap entry */
470         nr_events += 2; /* 1 is required, 2 for good luck */
471 
472         size = sizeof(struct aio_ring);
473         size += sizeof(struct io_event) * nr_events;
474 
475         nr_pages = PFN_UP(size);
476         if (nr_pages < 0)
477                 return -EINVAL;
478 
479         file = aio_private_file(ctx, nr_pages);
480         if (IS_ERR(file)) {
481                 ctx->aio_ring_file = NULL;
482                 return -ENOMEM;
483         }
484 
485         ctx->aio_ring_file = file;
486         nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
487                         / sizeof(struct io_event);
488 
489         ctx->ring_pages = ctx->internal_pages;
490         if (nr_pages > AIO_RING_PAGES) {
491                 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
492                                           GFP_KERNEL);
493                 if (!ctx->ring_pages) {
494                         put_aio_ring_file(ctx);
495                         return -ENOMEM;
496                 }
497         }
498 
499         for (i = 0; i < nr_pages; i++) {
500                 struct page *page;
501                 page = find_or_create_page(file->f_mapping,
502                                            i, GFP_HIGHUSER | __GFP_ZERO);
503                 if (!page)
504                         break;
505                 pr_debug("pid(%d) page[%d]->count=%d\n",
506                          current->pid, i, page_count(page));
507                 SetPageUptodate(page);
508                 unlock_page(page);
509 
510                 ctx->ring_pages[i] = page;
511         }
512         ctx->nr_pages = i;
513 
514         if (unlikely(i != nr_pages)) {
515                 aio_free_ring(ctx);
516                 return -ENOMEM;
517         }
518 
519         ctx->mmap_size = nr_pages * PAGE_SIZE;
520         pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
521 
522         if (down_write_killable(&mm->mmap_sem)) {
523                 ctx->mmap_size = 0;
524                 aio_free_ring(ctx);
525                 return -EINTR;
526         }
527 
528         ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
529                                        PROT_READ | PROT_WRITE,
530                                        MAP_SHARED, 0, &unused, NULL);
531         up_write(&mm->mmap_sem);
532         if (IS_ERR((void *)ctx->mmap_base)) {
533                 ctx->mmap_size = 0;
534                 aio_free_ring(ctx);
535                 return -ENOMEM;
536         }
537 
538         pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
539 
540         ctx->user_id = ctx->mmap_base;
541         ctx->nr_events = nr_events; /* trusted copy */
542 
543         ring = kmap_atomic(ctx->ring_pages[0]);
544         ring->nr = nr_events;   /* user copy */
545         ring->id = ~0U;
546         ring->head = ring->tail = 0;
547         ring->magic = AIO_RING_MAGIC;
548         ring->compat_features = AIO_RING_COMPAT_FEATURES;
549         ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
550         ring->header_length = sizeof(struct aio_ring);
551         kunmap_atomic(ring);
552         flush_dcache_page(ctx->ring_pages[0]);
553 
554         return 0;
555 }
556 
557 #define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
558 #define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
559 #define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
560 
561 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
562 {
563         struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
564         struct kioctx *ctx = req->ki_ctx;
565         unsigned long flags;
566 
567         if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
568                 return;
569 
570         spin_lock_irqsave(&ctx->ctx_lock, flags);
571         list_add_tail(&req->ki_list, &ctx->active_reqs);
572         req->ki_cancel = cancel;
573         spin_unlock_irqrestore(&ctx->ctx_lock, flags);
574 }
575 EXPORT_SYMBOL(kiocb_set_cancel_fn);
576 
577 /*
578  * free_ioctx() should be RCU delayed to synchronize against the RCU
579  * protected lookup_ioctx() and also needs process context to call
580  * aio_free_ring().  Use rcu_work.
581  */
582 static void free_ioctx(struct work_struct *work)
583 {
584         struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
585                                           free_rwork);
586         pr_debug("freeing %p\n", ctx);
587 
588         aio_free_ring(ctx);
589         free_percpu(ctx->cpu);
590         percpu_ref_exit(&ctx->reqs);
591         percpu_ref_exit(&ctx->users);
592         kmem_cache_free(kioctx_cachep, ctx);
593 }
594 
595 static void free_ioctx_reqs(struct percpu_ref *ref)
596 {
597         struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
598 
599         /* At this point we know that there are no any in-flight requests */
600         if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
601                 complete(&ctx->rq_wait->comp);
602 
603         /* Synchronize against RCU protected table->table[] dereferences */
604         INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
605         queue_rcu_work(system_wq, &ctx->free_rwork);
606 }
607 
608 /*
609  * When this function runs, the kioctx has been removed from the "hash table"
610  * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
611  * now it's safe to cancel any that need to be.
612  */
613 static void free_ioctx_users(struct percpu_ref *ref)
614 {
615         struct kioctx *ctx = container_of(ref, struct kioctx, users);
616         struct aio_kiocb *req;
617 
618         spin_lock_irq(&ctx->ctx_lock);
619 
620         while (!list_empty(&ctx->active_reqs)) {
621                 req = list_first_entry(&ctx->active_reqs,
622                                        struct aio_kiocb, ki_list);
623                 req->ki_cancel(&req->rw);
624                 list_del_init(&req->ki_list);
625         }
626 
627         spin_unlock_irq(&ctx->ctx_lock);
628 
629         percpu_ref_kill(&ctx->reqs);
630         percpu_ref_put(&ctx->reqs);
631 }
632 
633 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
634 {
635         unsigned i, new_nr;
636         struct kioctx_table *table, *old;
637         struct aio_ring *ring;
638 
639         spin_lock(&mm->ioctx_lock);
640         table = rcu_dereference_raw(mm->ioctx_table);
641 
642         while (1) {
643                 if (table)
644                         for (i = 0; i < table->nr; i++)
645                                 if (!rcu_access_pointer(table->table[i])) {
646                                         ctx->id = i;
647                                         rcu_assign_pointer(table->table[i], ctx);
648                                         spin_unlock(&mm->ioctx_lock);
649 
650                                         /* While kioctx setup is in progress,
651                                          * we are protected from page migration
652                                          * changes ring_pages by ->ring_lock.
653                                          */
654                                         ring = kmap_atomic(ctx->ring_pages[0]);
655                                         ring->id = ctx->id;
656                                         kunmap_atomic(ring);
657                                         return 0;
658                                 }
659 
660                 new_nr = (table ? table->nr : 1) * 4;
661                 spin_unlock(&mm->ioctx_lock);
662 
663                 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
664                                 new_nr, GFP_KERNEL);
665                 if (!table)
666                         return -ENOMEM;
667 
668                 table->nr = new_nr;
669 
670                 spin_lock(&mm->ioctx_lock);
671                 old = rcu_dereference_raw(mm->ioctx_table);
672 
673                 if (!old) {
674                         rcu_assign_pointer(mm->ioctx_table, table);
675                 } else if (table->nr > old->nr) {
676                         memcpy(table->table, old->table,
677                                old->nr * sizeof(struct kioctx *));
678 
679                         rcu_assign_pointer(mm->ioctx_table, table);
680                         kfree_rcu(old, rcu);
681                 } else {
682                         kfree(table);
683                         table = old;
684                 }
685         }
686 }
687 
688 static void aio_nr_sub(unsigned nr)
689 {
690         spin_lock(&aio_nr_lock);
691         if (WARN_ON(aio_nr - nr > aio_nr))
692                 aio_nr = 0;
693         else
694                 aio_nr -= nr;
695         spin_unlock(&aio_nr_lock);
696 }
697 
698 /* ioctx_alloc
699  *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
700  */
701 static struct kioctx *ioctx_alloc(unsigned nr_events)
702 {
703         struct mm_struct *mm = current->mm;
704         struct kioctx *ctx;
705         int err = -ENOMEM;
706 
707         /*
708          * Store the original nr_events -- what userspace passed to io_setup(),
709          * for counting against the global limit -- before it changes.
710          */
711         unsigned int max_reqs = nr_events;
712 
713         /*
714          * We keep track of the number of available ringbuffer slots, to prevent
715          * overflow (reqs_available), and we also use percpu counters for this.
716          *
717          * So since up to half the slots might be on other cpu's percpu counters
718          * and unavailable, double nr_events so userspace sees what they
719          * expected: additionally, we move req_batch slots to/from percpu
720          * counters at a time, so make sure that isn't 0:
721          */
722         nr_events = max(nr_events, num_possible_cpus() * 4);
723         nr_events *= 2;
724 
725         /* Prevent overflows */
726         if (nr_events > (0x10000000U / sizeof(struct io_event))) {
727                 pr_debug("ENOMEM: nr_events too high\n");
728                 return ERR_PTR(-EINVAL);
729         }
730 
731         if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
732                 return ERR_PTR(-EAGAIN);
733 
734         ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
735         if (!ctx)
736                 return ERR_PTR(-ENOMEM);
737 
738         ctx->max_reqs = max_reqs;
739 
740         spin_lock_init(&ctx->ctx_lock);
741         spin_lock_init(&ctx->completion_lock);
742         mutex_init(&ctx->ring_lock);
743         /* Protect against page migration throughout kiotx setup by keeping
744          * the ring_lock mutex held until setup is complete. */
745         mutex_lock(&ctx->ring_lock);
746         init_waitqueue_head(&ctx->wait);
747 
748         INIT_LIST_HEAD(&ctx->active_reqs);
749 
750         if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
751                 goto err;
752 
753         if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
754                 goto err;
755 
756         ctx->cpu = alloc_percpu(struct kioctx_cpu);
757         if (!ctx->cpu)
758                 goto err;
759 
760         err = aio_setup_ring(ctx, nr_events);
761         if (err < 0)
762                 goto err;
763 
764         atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
765         ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
766         if (ctx->req_batch < 1)
767                 ctx->req_batch = 1;
768 
769         /* limit the number of system wide aios */
770         spin_lock(&aio_nr_lock);
771         if (aio_nr + ctx->max_reqs > aio_max_nr ||
772             aio_nr + ctx->max_reqs < aio_nr) {
773                 spin_unlock(&aio_nr_lock);
774                 err = -EAGAIN;
775                 goto err_ctx;
776         }
777         aio_nr += ctx->max_reqs;
778         spin_unlock(&aio_nr_lock);
779 
780         percpu_ref_get(&ctx->users);    /* io_setup() will drop this ref */
781         percpu_ref_get(&ctx->reqs);     /* free_ioctx_users() will drop this */
782 
783         err = ioctx_add_table(ctx, mm);
784         if (err)
785                 goto err_cleanup;
786 
787         /* Release the ring_lock mutex now that all setup is complete. */
788         mutex_unlock(&ctx->ring_lock);
789 
790         pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
791                  ctx, ctx->user_id, mm, ctx->nr_events);
792         return ctx;
793 
794 err_cleanup:
795         aio_nr_sub(ctx->max_reqs);
796 err_ctx:
797         atomic_set(&ctx->dead, 1);
798         if (ctx->mmap_size)
799                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
800         aio_free_ring(ctx);
801 err:
802         mutex_unlock(&ctx->ring_lock);
803         free_percpu(ctx->cpu);
804         percpu_ref_exit(&ctx->reqs);
805         percpu_ref_exit(&ctx->users);
806         kmem_cache_free(kioctx_cachep, ctx);
807         pr_debug("error allocating ioctx %d\n", err);
808         return ERR_PTR(err);
809 }
810 
811 /* kill_ioctx
812  *      Cancels all outstanding aio requests on an aio context.  Used
813  *      when the processes owning a context have all exited to encourage
814  *      the rapid destruction of the kioctx.
815  */
816 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
817                       struct ctx_rq_wait *wait)
818 {
819         struct kioctx_table *table;
820 
821         spin_lock(&mm->ioctx_lock);
822         if (atomic_xchg(&ctx->dead, 1)) {
823                 spin_unlock(&mm->ioctx_lock);
824                 return -EINVAL;
825         }
826 
827         table = rcu_dereference_raw(mm->ioctx_table);
828         WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
829         RCU_INIT_POINTER(table->table[ctx->id], NULL);
830         spin_unlock(&mm->ioctx_lock);
831 
832         /* free_ioctx_reqs() will do the necessary RCU synchronization */
833         wake_up_all(&ctx->wait);
834 
835         /*
836          * It'd be more correct to do this in free_ioctx(), after all
837          * the outstanding kiocbs have finished - but by then io_destroy
838          * has already returned, so io_setup() could potentially return
839          * -EAGAIN with no ioctxs actually in use (as far as userspace
840          *  could tell).
841          */
842         aio_nr_sub(ctx->max_reqs);
843 
844         if (ctx->mmap_size)
845                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
846 
847         ctx->rq_wait = wait;
848         percpu_ref_kill(&ctx->users);
849         return 0;
850 }
851 
852 /*
853  * exit_aio: called when the last user of mm goes away.  At this point, there is
854  * no way for any new requests to be submited or any of the io_* syscalls to be
855  * called on the context.
856  *
857  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
858  * them.
859  */
860 void exit_aio(struct mm_struct *mm)
861 {
862         struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
863         struct ctx_rq_wait wait;
864         int i, skipped;
865 
866         if (!table)
867                 return;
868 
869         atomic_set(&wait.count, table->nr);
870         init_completion(&wait.comp);
871 
872         skipped = 0;
873         for (i = 0; i < table->nr; ++i) {
874                 struct kioctx *ctx =
875                         rcu_dereference_protected(table->table[i], true);
876 
877                 if (!ctx) {
878                         skipped++;
879                         continue;
880                 }
881 
882                 /*
883                  * We don't need to bother with munmap() here - exit_mmap(mm)
884                  * is coming and it'll unmap everything. And we simply can't,
885                  * this is not necessarily our ->mm.
886                  * Since kill_ioctx() uses non-zero ->mmap_size as indicator
887                  * that it needs to unmap the area, just set it to 0.
888                  */
889                 ctx->mmap_size = 0;
890                 kill_ioctx(mm, ctx, &wait);
891         }
892 
893         if (!atomic_sub_and_test(skipped, &wait.count)) {
894                 /* Wait until all IO for the context are done. */
895                 wait_for_completion(&wait.comp);
896         }
897 
898         RCU_INIT_POINTER(mm->ioctx_table, NULL);
899         kfree(table);
900 }
901 
902 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
903 {
904         struct kioctx_cpu *kcpu;
905         unsigned long flags;
906 
907         local_irq_save(flags);
908         kcpu = this_cpu_ptr(ctx->cpu);
909         kcpu->reqs_available += nr;
910 
911         while (kcpu->reqs_available >= ctx->req_batch * 2) {
912                 kcpu->reqs_available -= ctx->req_batch;
913                 atomic_add(ctx->req_batch, &ctx->reqs_available);
914         }
915 
916         local_irq_restore(flags);
917 }
918 
919 static bool __get_reqs_available(struct kioctx *ctx)
920 {
921         struct kioctx_cpu *kcpu;
922         bool ret = false;
923         unsigned long flags;
924 
925         local_irq_save(flags);
926         kcpu = this_cpu_ptr(ctx->cpu);
927         if (!kcpu->reqs_available) {
928                 int old, avail = atomic_read(&ctx->reqs_available);
929 
930                 do {
931                         if (avail < ctx->req_batch)
932                                 goto out;
933 
934                         old = avail;
935                         avail = atomic_cmpxchg(&ctx->reqs_available,
936                                                avail, avail - ctx->req_batch);
937                 } while (avail != old);
938 
939                 kcpu->reqs_available += ctx->req_batch;
940         }
941 
942         ret = true;
943         kcpu->reqs_available--;
944 out:
945         local_irq_restore(flags);
946         return ret;
947 }
948 
949 /* refill_reqs_available
950  *      Updates the reqs_available reference counts used for tracking the
951  *      number of free slots in the completion ring.  This can be called
952  *      from aio_complete() (to optimistically update reqs_available) or
953  *      from aio_get_req() (the we're out of events case).  It must be
954  *      called holding ctx->completion_lock.
955  */
956 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
957                                   unsigned tail)
958 {
959         unsigned events_in_ring, completed;
960 
961         /* Clamp head since userland can write to it. */
962         head %= ctx->nr_events;
963         if (head <= tail)
964                 events_in_ring = tail - head;
965         else
966                 events_in_ring = ctx->nr_events - (head - tail);
967 
968         completed = ctx->completed_events;
969         if (events_in_ring < completed)
970                 completed -= events_in_ring;
971         else
972                 completed = 0;
973 
974         if (!completed)
975                 return;
976 
977         ctx->completed_events -= completed;
978         put_reqs_available(ctx, completed);
979 }
980 
981 /* user_refill_reqs_available
982  *      Called to refill reqs_available when aio_get_req() encounters an
983  *      out of space in the completion ring.
984  */
985 static void user_refill_reqs_available(struct kioctx *ctx)
986 {
987         spin_lock_irq(&ctx->completion_lock);
988         if (ctx->completed_events) {
989                 struct aio_ring *ring;
990                 unsigned head;
991 
992                 /* Access of ring->head may race with aio_read_events_ring()
993                  * here, but that's okay since whether we read the old version
994                  * or the new version, and either will be valid.  The important
995                  * part is that head cannot pass tail since we prevent
996                  * aio_complete() from updating tail by holding
997                  * ctx->completion_lock.  Even if head is invalid, the check
998                  * against ctx->completed_events below will make sure we do the
999                  * safe/right thing.
1000                  */
1001                 ring = kmap_atomic(ctx->ring_pages[0]);
1002                 head = ring->head;
1003                 kunmap_atomic(ring);
1004 
1005                 refill_reqs_available(ctx, head, ctx->tail);
1006         }
1007 
1008         spin_unlock_irq(&ctx->completion_lock);
1009 }
1010 
1011 static bool get_reqs_available(struct kioctx *ctx)
1012 {
1013         if (__get_reqs_available(ctx))
1014                 return true;
1015         user_refill_reqs_available(ctx);
1016         return __get_reqs_available(ctx);
1017 }
1018 
1019 /* aio_get_req
1020  *      Allocate a slot for an aio request.
1021  * Returns NULL if no requests are free.
1022  *
1023  * The refcount is initialized to 2 - one for the async op completion,
1024  * one for the synchronous code that does this.
1025  */
1026 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1027 {
1028         struct aio_kiocb *req;
1029 
1030         req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1031         if (unlikely(!req))
1032                 return NULL;
1033 
1034         if (unlikely(!get_reqs_available(ctx))) {
1035                 kmem_cache_free(kiocb_cachep, req);
1036                 return NULL;
1037         }
1038 
1039         percpu_ref_get(&ctx->reqs);
1040         req->ki_ctx = ctx;
1041         INIT_LIST_HEAD(&req->ki_list);
1042         refcount_set(&req->ki_refcnt, 2);
1043         req->ki_eventfd = NULL;
1044         return req;
1045 }
1046 
1047 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1048 {
1049         struct aio_ring __user *ring  = (void __user *)ctx_id;
1050         struct mm_struct *mm = current->mm;
1051         struct kioctx *ctx, *ret = NULL;
1052         struct kioctx_table *table;
1053         unsigned id;
1054 
1055         if (get_user(id, &ring->id))
1056                 return NULL;
1057 
1058         rcu_read_lock();
1059         table = rcu_dereference(mm->ioctx_table);
1060 
1061         if (!table || id >= table->nr)
1062                 goto out;
1063 
1064         id = array_index_nospec(id, table->nr);
1065         ctx = rcu_dereference(table->table[id]);
1066         if (ctx && ctx->user_id == ctx_id) {
1067                 if (percpu_ref_tryget_live(&ctx->users))
1068                         ret = ctx;
1069         }
1070 out:
1071         rcu_read_unlock();
1072         return ret;
1073 }
1074 
1075 static inline void iocb_destroy(struct aio_kiocb *iocb)
1076 {
1077         if (iocb->ki_eventfd)
1078                 eventfd_ctx_put(iocb->ki_eventfd);
1079         if (iocb->ki_filp)
1080                 fput(iocb->ki_filp);
1081         percpu_ref_put(&iocb->ki_ctx->reqs);
1082         kmem_cache_free(kiocb_cachep, iocb);
1083 }
1084 
1085 /* aio_complete
1086  *      Called when the io request on the given iocb is complete.
1087  */
1088 static void aio_complete(struct aio_kiocb *iocb)
1089 {
1090         struct kioctx   *ctx = iocb->ki_ctx;
1091         struct aio_ring *ring;
1092         struct io_event *ev_page, *event;
1093         unsigned tail, pos, head;
1094         unsigned long   flags;
1095 
1096         /*
1097          * Add a completion event to the ring buffer. Must be done holding
1098          * ctx->completion_lock to prevent other code from messing with the tail
1099          * pointer since we might be called from irq context.
1100          */
1101         spin_lock_irqsave(&ctx->completion_lock, flags);
1102 
1103         tail = ctx->tail;
1104         pos = tail + AIO_EVENTS_OFFSET;
1105 
1106         if (++tail >= ctx->nr_events)
1107                 tail = 0;
1108 
1109         ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1110         event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1111 
1112         *event = iocb->ki_res;
1113 
1114         kunmap_atomic(ev_page);
1115         flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1116 
1117         pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1118                  (void __user *)(unsigned long)iocb->ki_res.obj,
1119                  iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1120 
1121         /* after flagging the request as done, we
1122          * must never even look at it again
1123          */
1124         smp_wmb();      /* make event visible before updating tail */
1125 
1126         ctx->tail = tail;
1127 
1128         ring = kmap_atomic(ctx->ring_pages[0]);
1129         head = ring->head;
1130         ring->tail = tail;
1131         kunmap_atomic(ring);
1132         flush_dcache_page(ctx->ring_pages[0]);
1133 
1134         ctx->completed_events++;
1135         if (ctx->completed_events > 1)
1136                 refill_reqs_available(ctx, head, tail);
1137         spin_unlock_irqrestore(&ctx->completion_lock, flags);
1138 
1139         pr_debug("added to ring %p at [%u]\n", iocb, tail);
1140 
1141         /*
1142          * Check if the user asked us to deliver the result through an
1143          * eventfd. The eventfd_signal() function is safe to be called
1144          * from IRQ context.
1145          */
1146         if (iocb->ki_eventfd)
1147                 eventfd_signal(iocb->ki_eventfd, 1);
1148 
1149         /*
1150          * We have to order our ring_info tail store above and test
1151          * of the wait list below outside the wait lock.  This is
1152          * like in wake_up_bit() where clearing a bit has to be
1153          * ordered with the unlocked test.
1154          */
1155         smp_mb();
1156 
1157         if (waitqueue_active(&ctx->wait))
1158                 wake_up(&ctx->wait);
1159 }
1160 
1161 static inline void iocb_put(struct aio_kiocb *iocb)
1162 {
1163         if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1164                 aio_complete(iocb);
1165                 iocb_destroy(iocb);
1166         }
1167 }
1168 
1169 /* aio_read_events_ring
1170  *      Pull an event off of the ioctx's event ring.  Returns the number of
1171  *      events fetched
1172  */
1173 static long aio_read_events_ring(struct kioctx *ctx,
1174                                  struct io_event __user *event, long nr)
1175 {
1176         struct aio_ring *ring;
1177         unsigned head, tail, pos;
1178         long ret = 0;
1179         int copy_ret;
1180 
1181         /*
1182          * The mutex can block and wake us up and that will cause
1183          * wait_event_interruptible_hrtimeout() to schedule without sleeping
1184          * and repeat. This should be rare enough that it doesn't cause
1185          * peformance issues. See the comment in read_events() for more detail.
1186          */
1187         sched_annotate_sleep();
1188         mutex_lock(&ctx->ring_lock);
1189 
1190         /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1191         ring = kmap_atomic(ctx->ring_pages[0]);
1192         head = ring->head;
1193         tail = ring->tail;
1194         kunmap_atomic(ring);
1195 
1196         /*
1197          * Ensure that once we've read the current tail pointer, that
1198          * we also see the events that were stored up to the tail.
1199          */
1200         smp_rmb();
1201 
1202         pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1203 
1204         if (head == tail)
1205                 goto out;
1206 
1207         head %= ctx->nr_events;
1208         tail %= ctx->nr_events;
1209 
1210         while (ret < nr) {
1211                 long avail;
1212                 struct io_event *ev;
1213                 struct page *page;
1214 
1215                 avail = (head <= tail ?  tail : ctx->nr_events) - head;
1216                 if (head == tail)
1217                         break;
1218 
1219                 pos = head + AIO_EVENTS_OFFSET;
1220                 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1221                 pos %= AIO_EVENTS_PER_PAGE;
1222 
1223                 avail = min(avail, nr - ret);
1224                 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1225 
1226                 ev = kmap(page);
1227                 copy_ret = copy_to_user(event + ret, ev + pos,
1228                                         sizeof(*ev) * avail);
1229                 kunmap(page);
1230 
1231                 if (unlikely(copy_ret)) {
1232                         ret = -EFAULT;
1233                         goto out;
1234                 }
1235 
1236                 ret += avail;
1237                 head += avail;
1238                 head %= ctx->nr_events;
1239         }
1240 
1241         ring = kmap_atomic(ctx->ring_pages[0]);
1242         ring->head = head;
1243         kunmap_atomic(ring);
1244         flush_dcache_page(ctx->ring_pages[0]);
1245 
1246         pr_debug("%li  h%u t%u\n", ret, head, tail);
1247 out:
1248         mutex_unlock(&ctx->ring_lock);
1249 
1250         return ret;
1251 }
1252 
1253 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1254                             struct io_event __user *event, long *i)
1255 {
1256         long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1257 
1258         if (ret > 0)
1259                 *i += ret;
1260 
1261         if (unlikely(atomic_read(&ctx->dead)))
1262                 ret = -EINVAL;
1263 
1264         if (!*i)
1265                 *i = ret;
1266 
1267         return ret < 0 || *i >= min_nr;
1268 }
1269 
1270 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1271                         struct io_event __user *event,
1272                         ktime_t until)
1273 {
1274         long ret = 0;
1275 
1276         /*
1277          * Note that aio_read_events() is being called as the conditional - i.e.
1278          * we're calling it after prepare_to_wait() has set task state to
1279          * TASK_INTERRUPTIBLE.
1280          *
1281          * But aio_read_events() can block, and if it blocks it's going to flip
1282          * the task state back to TASK_RUNNING.
1283          *
1284          * This should be ok, provided it doesn't flip the state back to
1285          * TASK_RUNNING and return 0 too much - that causes us to spin. That
1286          * will only happen if the mutex_lock() call blocks, and we then find
1287          * the ringbuffer empty. So in practice we should be ok, but it's
1288          * something to be aware of when touching this code.
1289          */
1290         if (until == 0)
1291                 aio_read_events(ctx, min_nr, nr, event, &ret);
1292         else
1293                 wait_event_interruptible_hrtimeout(ctx->wait,
1294                                 aio_read_events(ctx, min_nr, nr, event, &ret),
1295                                 until);
1296         return ret;
1297 }
1298 
1299 /* sys_io_setup:
1300  *      Create an aio_context capable of receiving at least nr_events.
1301  *      ctxp must not point to an aio_context that already exists, and
1302  *      must be initialized to 0 prior to the call.  On successful
1303  *      creation of the aio_context, *ctxp is filled in with the resulting 
1304  *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1305  *      if the specified nr_events exceeds internal limits.  May fail 
1306  *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1307  *      of available events.  May fail with -ENOMEM if insufficient kernel
1308  *      resources are available.  May fail with -EFAULT if an invalid
1309  *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1310  *      implemented.
1311  */
1312 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1313 {
1314         struct kioctx *ioctx = NULL;
1315         unsigned long ctx;
1316         long ret;
1317 
1318         ret = get_user(ctx, ctxp);
1319         if (unlikely(ret))
1320                 goto out;
1321 
1322         ret = -EINVAL;
1323         if (unlikely(ctx || nr_events == 0)) {
1324                 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1325                          ctx, nr_events);
1326                 goto out;
1327         }
1328 
1329         ioctx = ioctx_alloc(nr_events);
1330         ret = PTR_ERR(ioctx);
1331         if (!IS_ERR(ioctx)) {
1332                 ret = put_user(ioctx->user_id, ctxp);
1333                 if (ret)
1334                         kill_ioctx(current->mm, ioctx, NULL);
1335                 percpu_ref_put(&ioctx->users);
1336         }
1337 
1338 out:
1339         return ret;
1340 }
1341 
1342 #ifdef CONFIG_COMPAT
1343 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1344 {
1345         struct kioctx *ioctx = NULL;
1346         unsigned long ctx;
1347         long ret;
1348 
1349         ret = get_user(ctx, ctx32p);
1350         if (unlikely(ret))
1351                 goto out;
1352 
1353         ret = -EINVAL;
1354         if (unlikely(ctx || nr_events == 0)) {
1355                 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1356                          ctx, nr_events);
1357                 goto out;
1358         }
1359 
1360         ioctx = ioctx_alloc(nr_events);
1361         ret = PTR_ERR(ioctx);
1362         if (!IS_ERR(ioctx)) {
1363                 /* truncating is ok because it's a user address */
1364                 ret = put_user((u32)ioctx->user_id, ctx32p);
1365                 if (ret)
1366                         kill_ioctx(current->mm, ioctx, NULL);
1367                 percpu_ref_put(&ioctx->users);
1368         }
1369 
1370 out:
1371         return ret;
1372 }
1373 #endif
1374 
1375 /* sys_io_destroy:
1376  *      Destroy the aio_context specified.  May cancel any outstanding 
1377  *      AIOs and block on completion.  Will fail with -ENOSYS if not
1378  *      implemented.  May fail with -EINVAL if the context pointed to
1379  *      is invalid.
1380  */
1381 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1382 {
1383         struct kioctx *ioctx = lookup_ioctx(ctx);
1384         if (likely(NULL != ioctx)) {
1385                 struct ctx_rq_wait wait;
1386                 int ret;
1387 
1388                 init_completion(&wait.comp);
1389                 atomic_set(&wait.count, 1);
1390 
1391                 /* Pass requests_done to kill_ioctx() where it can be set
1392                  * in a thread-safe way. If we try to set it here then we have
1393                  * a race condition if two io_destroy() called simultaneously.
1394                  */
1395                 ret = kill_ioctx(current->mm, ioctx, &wait);
1396                 percpu_ref_put(&ioctx->users);
1397 
1398                 /* Wait until all IO for the context are done. Otherwise kernel
1399                  * keep using user-space buffers even if user thinks the context
1400                  * is destroyed.
1401                  */
1402                 if (!ret)
1403                         wait_for_completion(&wait.comp);
1404 
1405                 return ret;
1406         }
1407         pr_debug("EINVAL: invalid context id\n");
1408         return -EINVAL;
1409 }
1410 
1411 static void aio_remove_iocb(struct aio_kiocb *iocb)
1412 {
1413         struct kioctx *ctx = iocb->ki_ctx;
1414         unsigned long flags;
1415 
1416         spin_lock_irqsave(&ctx->ctx_lock, flags);
1417         list_del(&iocb->ki_list);
1418         spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1419 }
1420 
1421 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1422 {
1423         struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1424 
1425         if (!list_empty_careful(&iocb->ki_list))
1426                 aio_remove_iocb(iocb);
1427 
1428         if (kiocb->ki_flags & IOCB_WRITE) {
1429                 struct inode *inode = file_inode(kiocb->ki_filp);
1430 
1431                 /*
1432                  * Tell lockdep we inherited freeze protection from submission
1433                  * thread.
1434                  */
1435                 if (S_ISREG(inode->i_mode))
1436                         __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1437                 file_end_write(kiocb->ki_filp);
1438         }
1439 
1440         iocb->ki_res.res = res;
1441         iocb->ki_res.res2 = res2;
1442         iocb_put(iocb);
1443 }
1444 
1445 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1446 {
1447         int ret;
1448 
1449         req->ki_complete = aio_complete_rw;
1450         req->private = NULL;
1451         req->ki_pos = iocb->aio_offset;
1452         req->ki_flags = iocb_flags(req->ki_filp);
1453         if (iocb->aio_flags & IOCB_FLAG_RESFD)
1454                 req->ki_flags |= IOCB_EVENTFD;
1455         req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1456         if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1457                 /*
1458                  * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1459                  * aio_reqprio is interpreted as an I/O scheduling
1460                  * class and priority.
1461                  */
1462                 ret = ioprio_check_cap(iocb->aio_reqprio);
1463                 if (ret) {
1464                         pr_debug("aio ioprio check cap error: %d\n", ret);
1465                         return ret;
1466                 }
1467 
1468                 req->ki_ioprio = iocb->aio_reqprio;
1469         } else
1470                 req->ki_ioprio = get_current_ioprio();
1471 
1472         ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1473         if (unlikely(ret))
1474                 return ret;
1475 
1476         req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1477         return 0;
1478 }
1479 
1480 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1481                 struct iovec **iovec, bool vectored, bool compat,
1482                 struct iov_iter *iter)
1483 {
1484         void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1485         size_t len = iocb->aio_nbytes;
1486 
1487         if (!vectored) {
1488                 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1489                 *iovec = NULL;
1490                 return ret;
1491         }
1492 #ifdef CONFIG_COMPAT
1493         if (compat)
1494                 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1495                                 iter);
1496 #endif
1497         return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1498 }
1499 
1500 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1501 {
1502         switch (ret) {
1503         case -EIOCBQUEUED:
1504                 break;
1505         case -ERESTARTSYS:
1506         case -ERESTARTNOINTR:
1507         case -ERESTARTNOHAND:
1508         case -ERESTART_RESTARTBLOCK:
1509                 /*
1510                  * There's no easy way to restart the syscall since other AIO's
1511                  * may be already running. Just fail this IO with EINTR.
1512                  */
1513                 ret = -EINTR;
1514                 /*FALLTHRU*/
1515         default:
1516                 req->ki_complete(req, ret, 0);
1517         }
1518 }
1519 
1520 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1521                         bool vectored, bool compat)
1522 {
1523         struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1524         struct iov_iter iter;
1525         struct file *file;
1526         int ret;
1527 
1528         ret = aio_prep_rw(req, iocb);
1529         if (ret)
1530                 return ret;
1531         file = req->ki_filp;
1532         if (unlikely(!(file->f_mode & FMODE_READ)))
1533                 return -EBADF;
1534         ret = -EINVAL;
1535         if (unlikely(!file->f_op->read_iter))
1536                 return -EINVAL;
1537 
1538         ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1539         if (ret < 0)
1540                 return ret;
1541         ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1542         if (!ret)
1543                 aio_rw_done(req, call_read_iter(file, req, &iter));
1544         kfree(iovec);
1545         return ret;
1546 }
1547 
1548 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1549                          bool vectored, bool compat)
1550 {
1551         struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1552         struct iov_iter iter;
1553         struct file *file;
1554         int ret;
1555 
1556         ret = aio_prep_rw(req, iocb);
1557         if (ret)
1558                 return ret;
1559         file = req->ki_filp;
1560 
1561         if (unlikely(!(file->f_mode & FMODE_WRITE)))
1562                 return -EBADF;
1563         if (unlikely(!file->f_op->write_iter))
1564                 return -EINVAL;
1565 
1566         ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1567         if (ret < 0)
1568                 return ret;
1569         ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1570         if (!ret) {
1571                 /*
1572                  * Open-code file_start_write here to grab freeze protection,
1573                  * which will be released by another thread in
1574                  * aio_complete_rw().  Fool lockdep by telling it the lock got
1575                  * released so that it doesn't complain about the held lock when
1576                  * we return to userspace.
1577                  */
1578                 if (S_ISREG(file_inode(file)->i_mode)) {
1579                         __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true);
1580                         __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1581                 }
1582                 req->ki_flags |= IOCB_WRITE;
1583                 aio_rw_done(req, call_write_iter(file, req, &iter));
1584         }
1585         kfree(iovec);
1586         return ret;
1587 }
1588 
1589 static void aio_fsync_work(struct work_struct *work)
1590 {
1591         struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1592 
1593         iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1594         iocb_put(iocb);
1595 }
1596 
1597 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1598                      bool datasync)
1599 {
1600         if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1601                         iocb->aio_rw_flags))
1602                 return -EINVAL;
1603 
1604         if (unlikely(!req->file->f_op->fsync))
1605                 return -EINVAL;
1606 
1607         req->datasync = datasync;
1608         INIT_WORK(&req->work, aio_fsync_work);
1609         schedule_work(&req->work);
1610         return 0;
1611 }
1612 
1613 static void aio_poll_put_work(struct work_struct *work)
1614 {
1615         struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1616         struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1617 
1618         iocb_put(iocb);
1619 }
1620 
1621 static void aio_poll_complete_work(struct work_struct *work)
1622 {
1623         struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1624         struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1625         struct poll_table_struct pt = { ._key = req->events };
1626         struct kioctx *ctx = iocb->ki_ctx;
1627         __poll_t mask = 0;
1628 
1629         if (!READ_ONCE(req->cancelled))
1630                 mask = vfs_poll(req->file, &pt) & req->events;
1631 
1632         /*
1633          * Note that ->ki_cancel callers also delete iocb from active_reqs after
1634          * calling ->ki_cancel.  We need the ctx_lock roundtrip here to
1635          * synchronize with them.  In the cancellation case the list_del_init
1636          * itself is not actually needed, but harmless so we keep it in to
1637          * avoid further branches in the fast path.
1638          */
1639         spin_lock_irq(&ctx->ctx_lock);
1640         if (!mask && !READ_ONCE(req->cancelled)) {
1641                 add_wait_queue(req->head, &req->wait);
1642                 spin_unlock_irq(&ctx->ctx_lock);
1643                 return;
1644         }
1645         list_del_init(&iocb->ki_list);
1646         iocb->ki_res.res = mangle_poll(mask);
1647         req->done = true;
1648         spin_unlock_irq(&ctx->ctx_lock);
1649 
1650         iocb_put(iocb);
1651 }
1652 
1653 /* assumes we are called with irqs disabled */
1654 static int aio_poll_cancel(struct kiocb *iocb)
1655 {
1656         struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1657         struct poll_iocb *req = &aiocb->poll;
1658 
1659         spin_lock(&req->head->lock);
1660         WRITE_ONCE(req->cancelled, true);
1661         if (!list_empty(&req->wait.entry)) {
1662                 list_del_init(&req->wait.entry);
1663                 schedule_work(&aiocb->poll.work);
1664         }
1665         spin_unlock(&req->head->lock);
1666 
1667         return 0;
1668 }
1669 
1670 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1671                 void *key)
1672 {
1673         struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1674         struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1675         __poll_t mask = key_to_poll(key);
1676         unsigned long flags;
1677 
1678         /* for instances that support it check for an event match first: */
1679         if (mask && !(mask & req->events))
1680                 return 0;
1681 
1682         list_del_init(&req->wait.entry);
1683 
1684         if (mask && spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1685                 struct kioctx *ctx = iocb->ki_ctx;
1686 
1687                 /*
1688                  * Try to complete the iocb inline if we can. Use
1689                  * irqsave/irqrestore because not all filesystems (e.g. fuse)
1690                  * call this function with IRQs disabled and because IRQs
1691                  * have to be disabled before ctx_lock is obtained.
1692                  */
1693                 list_del(&iocb->ki_list);
1694                 iocb->ki_res.res = mangle_poll(mask);
1695                 req->done = true;
1696                 if (iocb->ki_eventfd && eventfd_signal_count()) {
1697                         iocb = NULL;
1698                         INIT_WORK(&req->work, aio_poll_put_work);
1699                         schedule_work(&req->work);
1700                 }
1701                 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1702                 if (iocb)
1703                         iocb_put(iocb);
1704         } else {
1705                 schedule_work(&req->work);
1706         }
1707         return 1;
1708 }
1709 
1710 struct aio_poll_table {
1711         struct poll_table_struct        pt;
1712         struct aio_kiocb                *iocb;
1713         int                             error;
1714 };
1715 
1716 static void
1717 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1718                 struct poll_table_struct *p)
1719 {
1720         struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1721 
1722         /* multiple wait queues per file are not supported */
1723         if (unlikely(pt->iocb->poll.head)) {
1724                 pt->error = -EINVAL;
1725                 return;
1726         }
1727 
1728         pt->error = 0;
1729         pt->iocb->poll.head = head;
1730         add_wait_queue(head, &pt->iocb->poll.wait);
1731 }
1732 
1733 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1734 {
1735         struct kioctx *ctx = aiocb->ki_ctx;
1736         struct poll_iocb *req = &aiocb->poll;
1737         struct aio_poll_table apt;
1738         bool cancel = false;
1739         __poll_t mask;
1740 
1741         /* reject any unknown events outside the normal event mask. */
1742         if ((u16)iocb->aio_buf != iocb->aio_buf)
1743                 return -EINVAL;
1744         /* reject fields that are not defined for poll */
1745         if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1746                 return -EINVAL;
1747 
1748         INIT_WORK(&req->work, aio_poll_complete_work);
1749         req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1750 
1751         req->head = NULL;
1752         req->done = false;
1753         req->cancelled = false;
1754 
1755         apt.pt._qproc = aio_poll_queue_proc;
1756         apt.pt._key = req->events;
1757         apt.iocb = aiocb;
1758         apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1759 
1760         /* initialized the list so that we can do list_empty checks */
1761         INIT_LIST_HEAD(&req->wait.entry);
1762         init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1763 
1764         mask = vfs_poll(req->file, &apt.pt) & req->events;
1765         spin_lock_irq(&ctx->ctx_lock);
1766         if (likely(req->head)) {
1767                 spin_lock(&req->head->lock);
1768                 if (unlikely(list_empty(&req->wait.entry))) {
1769                         if (apt.error)
1770                                 cancel = true;
1771                         apt.error = 0;
1772                         mask = 0;
1773                 }
1774                 if (mask || apt.error) {
1775                         list_del_init(&req->wait.entry);
1776                 } else if (cancel) {
1777                         WRITE_ONCE(req->cancelled, true);
1778                 } else if (!req->done) { /* actually waiting for an event */
1779                         list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1780                         aiocb->ki_cancel = aio_poll_cancel;
1781                 }
1782                 spin_unlock(&req->head->lock);
1783         }
1784         if (mask) { /* no async, we'd stolen it */
1785                 aiocb->ki_res.res = mangle_poll(mask);
1786                 apt.error = 0;
1787         }
1788         spin_unlock_irq(&ctx->ctx_lock);
1789         if (mask)
1790                 iocb_put(aiocb);
1791         return apt.error;
1792 }
1793 
1794 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1795                            struct iocb __user *user_iocb, struct aio_kiocb *req,
1796                            bool compat)
1797 {
1798         req->ki_filp = fget(iocb->aio_fildes);
1799         if (unlikely(!req->ki_filp))
1800                 return -EBADF;
1801 
1802         if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1803                 struct eventfd_ctx *eventfd;
1804                 /*
1805                  * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1806                  * instance of the file* now. The file descriptor must be
1807                  * an eventfd() fd, and will be signaled for each completed
1808                  * event using the eventfd_signal() function.
1809                  */
1810                 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1811                 if (IS_ERR(eventfd))
1812                         return PTR_ERR(eventfd);
1813 
1814                 req->ki_eventfd = eventfd;
1815         }
1816 
1817         if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1818                 pr_debug("EFAULT: aio_key\n");
1819                 return -EFAULT;
1820         }
1821 
1822         req->ki_res.obj = (u64)(unsigned long)user_iocb;
1823         req->ki_res.data = iocb->aio_data;
1824         req->ki_res.res = 0;
1825         req->ki_res.res2 = 0;
1826 
1827         switch (iocb->aio_lio_opcode) {
1828         case IOCB_CMD_PREAD:
1829                 return aio_read(&req->rw, iocb, false, compat);
1830         case IOCB_CMD_PWRITE:
1831                 return aio_write(&req->rw, iocb, false, compat);
1832         case IOCB_CMD_PREADV:
1833                 return aio_read(&req->rw, iocb, true, compat);
1834         case IOCB_CMD_PWRITEV:
1835                 return aio_write(&req->rw, iocb, true, compat);
1836         case IOCB_CMD_FSYNC:
1837                 return aio_fsync(&req->fsync, iocb, false);
1838         case IOCB_CMD_FDSYNC:
1839                 return aio_fsync(&req->fsync, iocb, true);
1840         case IOCB_CMD_POLL:
1841                 return aio_poll(req, iocb);
1842         default:
1843                 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1844                 return -EINVAL;
1845         }
1846 }
1847 
1848 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1849                          bool compat)
1850 {
1851         struct aio_kiocb *req;
1852         struct iocb iocb;
1853         int err;
1854 
1855         if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1856                 return -EFAULT;
1857 
1858         /* enforce forwards compatibility on users */
1859         if (unlikely(iocb.aio_reserved2)) {
1860                 pr_debug("EINVAL: reserve field set\n");
1861                 return -EINVAL;
1862         }
1863 
1864         /* prevent overflows */
1865         if (unlikely(
1866             (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1867             (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1868             ((ssize_t)iocb.aio_nbytes < 0)
1869            )) {
1870                 pr_debug("EINVAL: overflow check\n");
1871                 return -EINVAL;
1872         }
1873 
1874         req = aio_get_req(ctx);
1875         if (unlikely(!req))
1876                 return -EAGAIN;
1877 
1878         err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
1879 
1880         /* Done with the synchronous reference */
1881         iocb_put(req);
1882 
1883         /*
1884          * If err is 0, we'd either done aio_complete() ourselves or have
1885          * arranged for that to be done asynchronously.  Anything non-zero
1886          * means that we need to destroy req ourselves.
1887          */
1888         if (unlikely(err)) {
1889                 iocb_destroy(req);
1890                 put_reqs_available(ctx, 1);
1891         }
1892         return err;
1893 }
1894 
1895 /* sys_io_submit:
1896  *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1897  *      the number of iocbs queued.  May return -EINVAL if the aio_context
1898  *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1899  *      *iocbpp[0] is not properly initialized, if the operation specified
1900  *      is invalid for the file descriptor in the iocb.  May fail with
1901  *      -EFAULT if any of the data structures point to invalid data.  May
1902  *      fail with -EBADF if the file descriptor specified in the first
1903  *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1904  *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1905  *      fail with -ENOSYS if not implemented.
1906  */
1907 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1908                 struct iocb __user * __user *, iocbpp)
1909 {
1910         struct kioctx *ctx;
1911         long ret = 0;
1912         int i = 0;
1913         struct blk_plug plug;
1914 
1915         if (unlikely(nr < 0))
1916                 return -EINVAL;
1917 
1918         ctx = lookup_ioctx(ctx_id);
1919         if (unlikely(!ctx)) {
1920                 pr_debug("EINVAL: invalid context id\n");
1921                 return -EINVAL;
1922         }
1923 
1924         if (nr > ctx->nr_events)
1925                 nr = ctx->nr_events;
1926 
1927         if (nr > AIO_PLUG_THRESHOLD)
1928                 blk_start_plug(&plug);
1929         for (i = 0; i < nr; i++) {
1930                 struct iocb __user *user_iocb;
1931 
1932                 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1933                         ret = -EFAULT;
1934                         break;
1935                 }
1936 
1937                 ret = io_submit_one(ctx, user_iocb, false);
1938                 if (ret)
1939                         break;
1940         }
1941         if (nr > AIO_PLUG_THRESHOLD)
1942                 blk_finish_plug(&plug);
1943 
1944         percpu_ref_put(&ctx->users);
1945         return i ? i : ret;
1946 }
1947 
1948 #ifdef CONFIG_COMPAT
1949 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1950                        int, nr, compat_uptr_t __user *, iocbpp)
1951 {
1952         struct kioctx *ctx;
1953         long ret = 0;
1954         int i = 0;
1955         struct blk_plug plug;
1956 
1957         if (unlikely(nr < 0))
1958                 return -EINVAL;
1959 
1960         ctx = lookup_ioctx(ctx_id);
1961         if (unlikely(!ctx)) {
1962                 pr_debug("EINVAL: invalid context id\n");
1963                 return -EINVAL;
1964         }
1965 
1966         if (nr > ctx->nr_events)
1967                 nr = ctx->nr_events;
1968 
1969         if (nr > AIO_PLUG_THRESHOLD)
1970                 blk_start_plug(&plug);
1971         for (i = 0; i < nr; i++) {
1972                 compat_uptr_t user_iocb;
1973 
1974                 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1975                         ret = -EFAULT;
1976                         break;
1977                 }
1978 
1979                 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
1980                 if (ret)
1981                         break;
1982         }
1983         if (nr > AIO_PLUG_THRESHOLD)
1984                 blk_finish_plug(&plug);
1985 
1986         percpu_ref_put(&ctx->users);
1987         return i ? i : ret;
1988 }
1989 #endif
1990 
1991 /* sys_io_cancel:
1992  *      Attempts to cancel an iocb previously passed to io_submit.  If
1993  *      the operation is successfully cancelled, the resulting event is
1994  *      copied into the memory pointed to by result without being placed
1995  *      into the completion queue and 0 is returned.  May fail with
1996  *      -EFAULT if any of the data structures pointed to are invalid.
1997  *      May fail with -EINVAL if aio_context specified by ctx_id is
1998  *      invalid.  May fail with -EAGAIN if the iocb specified was not
1999  *      cancelled.  Will fail with -ENOSYS if not implemented.
2000  */
2001 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2002                 struct io_event __user *, result)
2003 {
2004         struct kioctx *ctx;
2005         struct aio_kiocb *kiocb;
2006         int ret = -EINVAL;
2007         u32 key;
2008         u64 obj = (u64)(unsigned long)iocb;
2009 
2010         if (unlikely(get_user(key, &iocb->aio_key)))
2011                 return -EFAULT;
2012         if (unlikely(key != KIOCB_KEY))
2013                 return -EINVAL;
2014 
2015         ctx = lookup_ioctx(ctx_id);
2016         if (unlikely(!ctx))
2017                 return -EINVAL;
2018 
2019         spin_lock_irq(&ctx->ctx_lock);
2020         /* TODO: use a hash or array, this sucks. */
2021         list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2022                 if (kiocb->ki_res.obj == obj) {
2023                         ret = kiocb->ki_cancel(&kiocb->rw);
2024                         list_del_init(&kiocb->ki_list);
2025                         break;
2026                 }
2027         }
2028         spin_unlock_irq(&ctx->ctx_lock);
2029 
2030         if (!ret) {
2031                 /*
2032                  * The result argument is no longer used - the io_event is
2033                  * always delivered via the ring buffer. -EINPROGRESS indicates
2034                  * cancellation is progress:
2035                  */
2036                 ret = -EINPROGRESS;
2037         }
2038 
2039         percpu_ref_put(&ctx->users);
2040 
2041         return ret;
2042 }
2043 
2044 static long do_io_getevents(aio_context_t ctx_id,
2045                 long min_nr,
2046                 long nr,
2047                 struct io_event __user *events,
2048                 struct timespec64 *ts)
2049 {
2050         ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2051         struct kioctx *ioctx = lookup_ioctx(ctx_id);
2052         long ret = -EINVAL;
2053 
2054         if (likely(ioctx)) {
2055                 if (likely(min_nr <= nr && min_nr >= 0))
2056                         ret = read_events(ioctx, min_nr, nr, events, until);
2057                 percpu_ref_put(&ioctx->users);
2058         }
2059 
2060         return ret;
2061 }
2062 
2063 /* io_getevents:
2064  *      Attempts to read at least min_nr events and up to nr events from
2065  *      the completion queue for the aio_context specified by ctx_id. If
2066  *      it succeeds, the number of read events is returned. May fail with
2067  *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2068  *      out of range, if timeout is out of range.  May fail with -EFAULT
2069  *      if any of the memory specified is invalid.  May return 0 or
2070  *      < min_nr if the timeout specified by timeout has elapsed
2071  *      before sufficient events are available, where timeout == NULL
2072  *      specifies an infinite timeout. Note that the timeout pointed to by
2073  *      timeout is relative.  Will fail with -ENOSYS if not implemented.
2074  */
2075 #ifdef CONFIG_64BIT
2076 
2077 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2078                 long, min_nr,
2079                 long, nr,
2080                 struct io_event __user *, events,
2081                 struct __kernel_timespec __user *, timeout)
2082 {
2083         struct timespec64       ts;
2084         int                     ret;
2085 
2086         if (timeout && unlikely(get_timespec64(&ts, timeout)))
2087                 return -EFAULT;
2088 
2089         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2090         if (!ret && signal_pending(current))
2091                 ret = -EINTR;
2092         return ret;
2093 }
2094 
2095 #endif
2096 
2097 struct __aio_sigset {
2098         const sigset_t __user   *sigmask;
2099         size_t          sigsetsize;
2100 };
2101 
2102 SYSCALL_DEFINE6(io_pgetevents,
2103                 aio_context_t, ctx_id,
2104                 long, min_nr,
2105                 long, nr,
2106                 struct io_event __user *, events,
2107                 struct __kernel_timespec __user *, timeout,
2108                 const struct __aio_sigset __user *, usig)
2109 {
2110         struct __aio_sigset     ksig = { NULL, };
2111         struct timespec64       ts;
2112         bool interrupted;
2113         int ret;
2114 
2115         if (timeout && unlikely(get_timespec64(&ts, timeout)))
2116                 return -EFAULT;
2117 
2118         if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2119                 return -EFAULT;
2120 
2121         ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2122         if (ret)
2123                 return ret;
2124 
2125         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2126 
2127         interrupted = signal_pending(current);
2128         restore_saved_sigmask_unless(interrupted);
2129         if (interrupted && !ret)
2130                 ret = -ERESTARTNOHAND;
2131 
2132         return ret;
2133 }
2134 
2135 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2136 
2137 SYSCALL_DEFINE6(io_pgetevents_time32,
2138                 aio_context_t, ctx_id,
2139                 long, min_nr,
2140                 long, nr,
2141                 struct io_event __user *, events,
2142                 struct old_timespec32 __user *, timeout,
2143                 const struct __aio_sigset __user *, usig)
2144 {
2145         struct __aio_sigset     ksig = { NULL, };
2146         struct timespec64       ts;
2147         bool interrupted;
2148         int ret;
2149 
2150         if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2151                 return -EFAULT;
2152 
2153         if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2154                 return -EFAULT;
2155 
2156 
2157         ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2158         if (ret)
2159                 return ret;
2160 
2161         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2162 
2163         interrupted = signal_pending(current);
2164         restore_saved_sigmask_unless(interrupted);
2165         if (interrupted && !ret)
2166                 ret = -ERESTARTNOHAND;
2167 
2168         return ret;
2169 }
2170 
2171 #endif
2172 
2173 #if defined(CONFIG_COMPAT_32BIT_TIME)
2174 
2175 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2176                 __s32, min_nr,
2177                 __s32, nr,
2178                 struct io_event __user *, events,
2179                 struct old_timespec32 __user *, timeout)
2180 {
2181         struct timespec64 t;
2182         int ret;
2183 
2184         if (timeout && get_old_timespec32(&t, timeout))
2185                 return -EFAULT;
2186 
2187         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2188         if (!ret && signal_pending(current))
2189                 ret = -EINTR;
2190         return ret;
2191 }
2192 
2193 #endif
2194 
2195 #ifdef CONFIG_COMPAT
2196 
2197 struct __compat_aio_sigset {
2198         compat_uptr_t           sigmask;
2199         compat_size_t           sigsetsize;
2200 };
2201 
2202 #if defined(CONFIG_COMPAT_32BIT_TIME)
2203 
2204 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2205                 compat_aio_context_t, ctx_id,
2206                 compat_long_t, min_nr,
2207                 compat_long_t, nr,
2208                 struct io_event __user *, events,
2209                 struct old_timespec32 __user *, timeout,
2210                 const struct __compat_aio_sigset __user *, usig)
2211 {
2212         struct __compat_aio_sigset ksig = { 0, };
2213         struct timespec64 t;
2214         bool interrupted;
2215         int ret;
2216 
2217         if (timeout && get_old_timespec32(&t, timeout))
2218                 return -EFAULT;
2219 
2220         if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2221                 return -EFAULT;
2222 
2223         ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2224         if (ret)
2225                 return ret;
2226 
2227         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2228 
2229         interrupted = signal_pending(current);
2230         restore_saved_sigmask_unless(interrupted);
2231         if (interrupted && !ret)
2232                 ret = -ERESTARTNOHAND;
2233 
2234         return ret;
2235 }
2236 
2237 #endif
2238 
2239 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2240                 compat_aio_context_t, ctx_id,
2241                 compat_long_t, min_nr,
2242                 compat_long_t, nr,
2243                 struct io_event __user *, events,
2244                 struct __kernel_timespec __user *, timeout,
2245                 const struct __compat_aio_sigset __user *, usig)
2246 {
2247         struct __compat_aio_sigset ksig = { 0, };
2248         struct timespec64 t;
2249         bool interrupted;
2250         int ret;
2251 
2252         if (timeout && get_timespec64(&t, timeout))
2253                 return -EFAULT;
2254 
2255         if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2256                 return -EFAULT;
2257 
2258         ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2259         if (ret)
2260                 return ret;
2261 
2262         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2263 
2264         interrupted = signal_pending(current);
2265         restore_saved_sigmask_unless(interrupted);
2266         if (interrupted && !ret)
2267                 ret = -ERESTARTNOHAND;
2268 
2269         return ret;
2270 }
2271 #endif
2272 

~ [ 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