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

TOMOYO Linux Cross Reference
Linux/fs/aio.c

Version: ~ [ linux-5.12-rc7 ] ~ [ linux-5.11.13 ] ~ [ linux-5.10.29 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.111 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.186 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.230 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.266 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.266 ] ~ [ 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         struct cred             *creds;
180 };
181 
182 struct poll_iocb {
183         struct file             *file;
184         struct wait_queue_head  *head;
185         __poll_t                events;
186         bool                    done;
187         bool                    cancelled;
188         struct wait_queue_entry wait;
189         struct work_struct      work;
190 };
191 
192 /*
193  * NOTE! Each of the iocb union members has the file pointer
194  * as the first entry in their struct definition. So you can
195  * access the file pointer through any of the sub-structs,
196  * or directly as just 'ki_filp' in this struct.
197  */
198 struct aio_kiocb {
199         union {
200                 struct file             *ki_filp;
201                 struct kiocb            rw;
202                 struct fsync_iocb       fsync;
203                 struct poll_iocb        poll;
204         };
205 
206         struct kioctx           *ki_ctx;
207         kiocb_cancel_fn         *ki_cancel;
208 
209         struct io_event         ki_res;
210 
211         struct list_head        ki_list;        /* the aio core uses this
212                                                  * for cancellation */
213         refcount_t              ki_refcnt;
214 
215         /*
216          * If the aio_resfd field of the userspace iocb is not zero,
217          * this is the underlying eventfd context to deliver events to.
218          */
219         struct eventfd_ctx      *ki_eventfd;
220 };
221 
222 /*------ sysctl variables----*/
223 static DEFINE_SPINLOCK(aio_nr_lock);
224 unsigned long aio_nr;           /* current system wide number of aio requests */
225 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
226 /*----end sysctl variables---*/
227 
228 static struct kmem_cache        *kiocb_cachep;
229 static struct kmem_cache        *kioctx_cachep;
230 
231 static struct vfsmount *aio_mnt;
232 
233 static const struct file_operations aio_ring_fops;
234 static const struct address_space_operations aio_ctx_aops;
235 
236 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
237 {
238         struct file *file;
239         struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
240         if (IS_ERR(inode))
241                 return ERR_CAST(inode);
242 
243         inode->i_mapping->a_ops = &aio_ctx_aops;
244         inode->i_mapping->private_data = ctx;
245         inode->i_size = PAGE_SIZE * nr_pages;
246 
247         file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
248                                 O_RDWR, &aio_ring_fops);
249         if (IS_ERR(file))
250                 iput(inode);
251         return file;
252 }
253 
254 static int aio_init_fs_context(struct fs_context *fc)
255 {
256         if (!init_pseudo(fc, AIO_RING_MAGIC))
257                 return -ENOMEM;
258         fc->s_iflags |= SB_I_NOEXEC;
259         return 0;
260 }
261 
262 /* aio_setup
263  *      Creates the slab caches used by the aio routines, panic on
264  *      failure as this is done early during the boot sequence.
265  */
266 static int __init aio_setup(void)
267 {
268         static struct file_system_type aio_fs = {
269                 .name           = "aio",
270                 .init_fs_context = aio_init_fs_context,
271                 .kill_sb        = kill_anon_super,
272         };
273         aio_mnt = kern_mount(&aio_fs);
274         if (IS_ERR(aio_mnt))
275                 panic("Failed to create aio fs mount.");
276 
277         kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
278         kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
279         return 0;
280 }
281 __initcall(aio_setup);
282 
283 static void put_aio_ring_file(struct kioctx *ctx)
284 {
285         struct file *aio_ring_file = ctx->aio_ring_file;
286         struct address_space *i_mapping;
287 
288         if (aio_ring_file) {
289                 truncate_setsize(file_inode(aio_ring_file), 0);
290 
291                 /* Prevent further access to the kioctx from migratepages */
292                 i_mapping = aio_ring_file->f_mapping;
293                 spin_lock(&i_mapping->private_lock);
294                 i_mapping->private_data = NULL;
295                 ctx->aio_ring_file = NULL;
296                 spin_unlock(&i_mapping->private_lock);
297 
298                 fput(aio_ring_file);
299         }
300 }
301 
302 static void aio_free_ring(struct kioctx *ctx)
303 {
304         int i;
305 
306         /* Disconnect the kiotx from the ring file.  This prevents future
307          * accesses to the kioctx from page migration.
308          */
309         put_aio_ring_file(ctx);
310 
311         for (i = 0; i < ctx->nr_pages; i++) {
312                 struct page *page;
313                 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
314                                 page_count(ctx->ring_pages[i]));
315                 page = ctx->ring_pages[i];
316                 if (!page)
317                         continue;
318                 ctx->ring_pages[i] = NULL;
319                 put_page(page);
320         }
321 
322         if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
323                 kfree(ctx->ring_pages);
324                 ctx->ring_pages = NULL;
325         }
326 }
327 
328 static int aio_ring_mremap(struct vm_area_struct *vma)
329 {
330         struct file *file = vma->vm_file;
331         struct mm_struct *mm = vma->vm_mm;
332         struct kioctx_table *table;
333         int i, res = -EINVAL;
334 
335         spin_lock(&mm->ioctx_lock);
336         rcu_read_lock();
337         table = rcu_dereference(mm->ioctx_table);
338         for (i = 0; i < table->nr; i++) {
339                 struct kioctx *ctx;
340 
341                 ctx = rcu_dereference(table->table[i]);
342                 if (ctx && ctx->aio_ring_file == file) {
343                         if (!atomic_read(&ctx->dead)) {
344                                 ctx->user_id = ctx->mmap_base = vma->vm_start;
345                                 res = 0;
346                         }
347                         break;
348                 }
349         }
350 
351         rcu_read_unlock();
352         spin_unlock(&mm->ioctx_lock);
353         return res;
354 }
355 
356 static const struct vm_operations_struct aio_ring_vm_ops = {
357         .mremap         = aio_ring_mremap,
358 #if IS_ENABLED(CONFIG_MMU)
359         .fault          = filemap_fault,
360         .map_pages      = filemap_map_pages,
361         .page_mkwrite   = filemap_page_mkwrite,
362 #endif
363 };
364 
365 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
366 {
367         vma->vm_flags |= VM_DONTEXPAND;
368         vma->vm_ops = &aio_ring_vm_ops;
369         return 0;
370 }
371 
372 static const struct file_operations aio_ring_fops = {
373         .mmap = aio_ring_mmap,
374 };
375 
376 #if IS_ENABLED(CONFIG_MIGRATION)
377 static int aio_migratepage(struct address_space *mapping, struct page *new,
378                         struct page *old, enum migrate_mode mode)
379 {
380         struct kioctx *ctx;
381         unsigned long flags;
382         pgoff_t idx;
383         int rc;
384 
385         /*
386          * We cannot support the _NO_COPY case here, because copy needs to
387          * happen under the ctx->completion_lock. That does not work with the
388          * migration workflow of MIGRATE_SYNC_NO_COPY.
389          */
390         if (mode == MIGRATE_SYNC_NO_COPY)
391                 return -EINVAL;
392 
393         rc = 0;
394 
395         /* mapping->private_lock here protects against the kioctx teardown.  */
396         spin_lock(&mapping->private_lock);
397         ctx = mapping->private_data;
398         if (!ctx) {
399                 rc = -EINVAL;
400                 goto out;
401         }
402 
403         /* The ring_lock mutex.  The prevents aio_read_events() from writing
404          * to the ring's head, and prevents page migration from mucking in
405          * a partially initialized kiotx.
406          */
407         if (!mutex_trylock(&ctx->ring_lock)) {
408                 rc = -EAGAIN;
409                 goto out;
410         }
411 
412         idx = old->index;
413         if (idx < (pgoff_t)ctx->nr_pages) {
414                 /* Make sure the old page hasn't already been changed */
415                 if (ctx->ring_pages[idx] != old)
416                         rc = -EAGAIN;
417         } else
418                 rc = -EINVAL;
419 
420         if (rc != 0)
421                 goto out_unlock;
422 
423         /* Writeback must be complete */
424         BUG_ON(PageWriteback(old));
425         get_page(new);
426 
427         rc = migrate_page_move_mapping(mapping, new, old, 1);
428         if (rc != MIGRATEPAGE_SUCCESS) {
429                 put_page(new);
430                 goto out_unlock;
431         }
432 
433         /* Take completion_lock to prevent other writes to the ring buffer
434          * while the old page is copied to the new.  This prevents new
435          * events from being lost.
436          */
437         spin_lock_irqsave(&ctx->completion_lock, flags);
438         migrate_page_copy(new, old);
439         BUG_ON(ctx->ring_pages[idx] != old);
440         ctx->ring_pages[idx] = new;
441         spin_unlock_irqrestore(&ctx->completion_lock, flags);
442 
443         /* The old page is no longer accessible. */
444         put_page(old);
445 
446 out_unlock:
447         mutex_unlock(&ctx->ring_lock);
448 out:
449         spin_unlock(&mapping->private_lock);
450         return rc;
451 }
452 #endif
453 
454 static const struct address_space_operations aio_ctx_aops = {
455         .set_page_dirty = __set_page_dirty_no_writeback,
456 #if IS_ENABLED(CONFIG_MIGRATION)
457         .migratepage    = aio_migratepage,
458 #endif
459 };
460 
461 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
462 {
463         struct aio_ring *ring;
464         struct mm_struct *mm = current->mm;
465         unsigned long size, unused;
466         int nr_pages;
467         int i;
468         struct file *file;
469 
470         /* Compensate for the ring buffer's head/tail overlap entry */
471         nr_events += 2; /* 1 is required, 2 for good luck */
472 
473         size = sizeof(struct aio_ring);
474         size += sizeof(struct io_event) * nr_events;
475 
476         nr_pages = PFN_UP(size);
477         if (nr_pages < 0)
478                 return -EINVAL;
479 
480         file = aio_private_file(ctx, nr_pages);
481         if (IS_ERR(file)) {
482                 ctx->aio_ring_file = NULL;
483                 return -ENOMEM;
484         }
485 
486         ctx->aio_ring_file = file;
487         nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
488                         / sizeof(struct io_event);
489 
490         ctx->ring_pages = ctx->internal_pages;
491         if (nr_pages > AIO_RING_PAGES) {
492                 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
493                                           GFP_KERNEL);
494                 if (!ctx->ring_pages) {
495                         put_aio_ring_file(ctx);
496                         return -ENOMEM;
497                 }
498         }
499 
500         for (i = 0; i < nr_pages; i++) {
501                 struct page *page;
502                 page = find_or_create_page(file->f_mapping,
503                                            i, GFP_HIGHUSER | __GFP_ZERO);
504                 if (!page)
505                         break;
506                 pr_debug("pid(%d) page[%d]->count=%d\n",
507                          current->pid, i, page_count(page));
508                 SetPageUptodate(page);
509                 unlock_page(page);
510 
511                 ctx->ring_pages[i] = page;
512         }
513         ctx->nr_pages = i;
514 
515         if (unlikely(i != nr_pages)) {
516                 aio_free_ring(ctx);
517                 return -ENOMEM;
518         }
519 
520         ctx->mmap_size = nr_pages * PAGE_SIZE;
521         pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
522 
523         if (down_write_killable(&mm->mmap_sem)) {
524                 ctx->mmap_size = 0;
525                 aio_free_ring(ctx);
526                 return -EINTR;
527         }
528 
529         ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
530                                        PROT_READ | PROT_WRITE,
531                                        MAP_SHARED, 0, &unused, NULL);
532         up_write(&mm->mmap_sem);
533         if (IS_ERR((void *)ctx->mmap_base)) {
534                 ctx->mmap_size = 0;
535                 aio_free_ring(ctx);
536                 return -ENOMEM;
537         }
538 
539         pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
540 
541         ctx->user_id = ctx->mmap_base;
542         ctx->nr_events = nr_events; /* trusted copy */
543 
544         ring = kmap_atomic(ctx->ring_pages[0]);
545         ring->nr = nr_events;   /* user copy */
546         ring->id = ~0U;
547         ring->head = ring->tail = 0;
548         ring->magic = AIO_RING_MAGIC;
549         ring->compat_features = AIO_RING_COMPAT_FEATURES;
550         ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
551         ring->header_length = sizeof(struct aio_ring);
552         kunmap_atomic(ring);
553         flush_dcache_page(ctx->ring_pages[0]);
554 
555         return 0;
556 }
557 
558 #define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
559 #define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
560 #define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
561 
562 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
563 {
564         struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
565         struct kioctx *ctx = req->ki_ctx;
566         unsigned long flags;
567 
568         if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
569                 return;
570 
571         spin_lock_irqsave(&ctx->ctx_lock, flags);
572         list_add_tail(&req->ki_list, &ctx->active_reqs);
573         req->ki_cancel = cancel;
574         spin_unlock_irqrestore(&ctx->ctx_lock, flags);
575 }
576 EXPORT_SYMBOL(kiocb_set_cancel_fn);
577 
578 /*
579  * free_ioctx() should be RCU delayed to synchronize against the RCU
580  * protected lookup_ioctx() and also needs process context to call
581  * aio_free_ring().  Use rcu_work.
582  */
583 static void free_ioctx(struct work_struct *work)
584 {
585         struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
586                                           free_rwork);
587         pr_debug("freeing %p\n", ctx);
588 
589         aio_free_ring(ctx);
590         free_percpu(ctx->cpu);
591         percpu_ref_exit(&ctx->reqs);
592         percpu_ref_exit(&ctx->users);
593         kmem_cache_free(kioctx_cachep, ctx);
594 }
595 
596 static void free_ioctx_reqs(struct percpu_ref *ref)
597 {
598         struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
599 
600         /* At this point we know that there are no any in-flight requests */
601         if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
602                 complete(&ctx->rq_wait->comp);
603 
604         /* Synchronize against RCU protected table->table[] dereferences */
605         INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
606         queue_rcu_work(system_wq, &ctx->free_rwork);
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                 req->ki_cancel(&req->rw);
625                 list_del_init(&req->ki_list);
626         }
627 
628         spin_unlock_irq(&ctx->ctx_lock);
629 
630         percpu_ref_kill(&ctx->reqs);
631         percpu_ref_put(&ctx->reqs);
632 }
633 
634 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
635 {
636         unsigned i, new_nr;
637         struct kioctx_table *table, *old;
638         struct aio_ring *ring;
639 
640         spin_lock(&mm->ioctx_lock);
641         table = rcu_dereference_raw(mm->ioctx_table);
642 
643         while (1) {
644                 if (table)
645                         for (i = 0; i < table->nr; i++)
646                                 if (!rcu_access_pointer(table->table[i])) {
647                                         ctx->id = i;
648                                         rcu_assign_pointer(table->table[i], ctx);
649                                         spin_unlock(&mm->ioctx_lock);
650 
651                                         /* While kioctx setup is in progress,
652                                          * we are protected from page migration
653                                          * changes ring_pages by ->ring_lock.
654                                          */
655                                         ring = kmap_atomic(ctx->ring_pages[0]);
656                                         ring->id = ctx->id;
657                                         kunmap_atomic(ring);
658                                         return 0;
659                                 }
660 
661                 new_nr = (table ? table->nr : 1) * 4;
662                 spin_unlock(&mm->ioctx_lock);
663 
664                 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
665                                 new_nr, GFP_KERNEL);
666                 if (!table)
667                         return -ENOMEM;
668 
669                 table->nr = new_nr;
670 
671                 spin_lock(&mm->ioctx_lock);
672                 old = rcu_dereference_raw(mm->ioctx_table);
673 
674                 if (!old) {
675                         rcu_assign_pointer(mm->ioctx_table, table);
676                 } else if (table->nr > old->nr) {
677                         memcpy(table->table, old->table,
678                                old->nr * sizeof(struct kioctx *));
679 
680                         rcu_assign_pointer(mm->ioctx_table, table);
681                         kfree_rcu(old, rcu);
682                 } else {
683                         kfree(table);
684                         table = old;
685                 }
686         }
687 }
688 
689 static void aio_nr_sub(unsigned nr)
690 {
691         spin_lock(&aio_nr_lock);
692         if (WARN_ON(aio_nr - nr > aio_nr))
693                 aio_nr = 0;
694         else
695                 aio_nr -= nr;
696         spin_unlock(&aio_nr_lock);
697 }
698 
699 /* ioctx_alloc
700  *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
701  */
702 static struct kioctx *ioctx_alloc(unsigned nr_events)
703 {
704         struct mm_struct *mm = current->mm;
705         struct kioctx *ctx;
706         int err = -ENOMEM;
707 
708         /*
709          * Store the original nr_events -- what userspace passed to io_setup(),
710          * for counting against the global limit -- before it changes.
711          */
712         unsigned int max_reqs = nr_events;
713 
714         /*
715          * We keep track of the number of available ringbuffer slots, to prevent
716          * overflow (reqs_available), and we also use percpu counters for this.
717          *
718          * So since up to half the slots might be on other cpu's percpu counters
719          * and unavailable, double nr_events so userspace sees what they
720          * expected: additionally, we move req_batch slots to/from percpu
721          * counters at a time, so make sure that isn't 0:
722          */
723         nr_events = max(nr_events, num_possible_cpus() * 4);
724         nr_events *= 2;
725 
726         /* Prevent overflows */
727         if (nr_events > (0x10000000U / sizeof(struct io_event))) {
728                 pr_debug("ENOMEM: nr_events too high\n");
729                 return ERR_PTR(-EINVAL);
730         }
731 
732         if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
733                 return ERR_PTR(-EAGAIN);
734 
735         ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
736         if (!ctx)
737                 return ERR_PTR(-ENOMEM);
738 
739         ctx->max_reqs = max_reqs;
740 
741         spin_lock_init(&ctx->ctx_lock);
742         spin_lock_init(&ctx->completion_lock);
743         mutex_init(&ctx->ring_lock);
744         /* Protect against page migration throughout kiotx setup by keeping
745          * the ring_lock mutex held until setup is complete. */
746         mutex_lock(&ctx->ring_lock);
747         init_waitqueue_head(&ctx->wait);
748 
749         INIT_LIST_HEAD(&ctx->active_reqs);
750 
751         if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
752                 goto err;
753 
754         if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
755                 goto err;
756 
757         ctx->cpu = alloc_percpu(struct kioctx_cpu);
758         if (!ctx->cpu)
759                 goto err;
760 
761         err = aio_setup_ring(ctx, nr_events);
762         if (err < 0)
763                 goto err;
764 
765         atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
766         ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
767         if (ctx->req_batch < 1)
768                 ctx->req_batch = 1;
769 
770         /* limit the number of system wide aios */
771         spin_lock(&aio_nr_lock);
772         if (aio_nr + ctx->max_reqs > aio_max_nr ||
773             aio_nr + ctx->max_reqs < aio_nr) {
774                 spin_unlock(&aio_nr_lock);
775                 err = -EAGAIN;
776                 goto err_ctx;
777         }
778         aio_nr += ctx->max_reqs;
779         spin_unlock(&aio_nr_lock);
780 
781         percpu_ref_get(&ctx->users);    /* io_setup() will drop this ref */
782         percpu_ref_get(&ctx->reqs);     /* free_ioctx_users() will drop this */
783 
784         err = ioctx_add_table(ctx, mm);
785         if (err)
786                 goto err_cleanup;
787 
788         /* Release the ring_lock mutex now that all setup is complete. */
789         mutex_unlock(&ctx->ring_lock);
790 
791         pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
792                  ctx, ctx->user_id, mm, ctx->nr_events);
793         return ctx;
794 
795 err_cleanup:
796         aio_nr_sub(ctx->max_reqs);
797 err_ctx:
798         atomic_set(&ctx->dead, 1);
799         if (ctx->mmap_size)
800                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
801         aio_free_ring(ctx);
802 err:
803         mutex_unlock(&ctx->ring_lock);
804         free_percpu(ctx->cpu);
805         percpu_ref_exit(&ctx->reqs);
806         percpu_ref_exit(&ctx->users);
807         kmem_cache_free(kioctx_cachep, ctx);
808         pr_debug("error allocating ioctx %d\n", err);
809         return ERR_PTR(err);
810 }
811 
812 /* kill_ioctx
813  *      Cancels all outstanding aio requests on an aio context.  Used
814  *      when the processes owning a context have all exited to encourage
815  *      the rapid destruction of the kioctx.
816  */
817 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
818                       struct ctx_rq_wait *wait)
819 {
820         struct kioctx_table *table;
821 
822         spin_lock(&mm->ioctx_lock);
823         if (atomic_xchg(&ctx->dead, 1)) {
824                 spin_unlock(&mm->ioctx_lock);
825                 return -EINVAL;
826         }
827 
828         table = rcu_dereference_raw(mm->ioctx_table);
829         WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
830         RCU_INIT_POINTER(table->table[ctx->id], NULL);
831         spin_unlock(&mm->ioctx_lock);
832 
833         /* free_ioctx_reqs() will do the necessary RCU synchronization */
834         wake_up_all(&ctx->wait);
835 
836         /*
837          * It'd be more correct to do this in free_ioctx(), after all
838          * the outstanding kiocbs have finished - but by then io_destroy
839          * has already returned, so io_setup() could potentially return
840          * -EAGAIN with no ioctxs actually in use (as far as userspace
841          *  could tell).
842          */
843         aio_nr_sub(ctx->max_reqs);
844 
845         if (ctx->mmap_size)
846                 vm_munmap(ctx->mmap_base, ctx->mmap_size);
847 
848         ctx->rq_wait = wait;
849         percpu_ref_kill(&ctx->users);
850         return 0;
851 }
852 
853 /*
854  * exit_aio: called when the last user of mm goes away.  At this point, there is
855  * no way for any new requests to be submited or any of the io_* syscalls to be
856  * called on the context.
857  *
858  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
859  * them.
860  */
861 void exit_aio(struct mm_struct *mm)
862 {
863         struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
864         struct ctx_rq_wait wait;
865         int i, skipped;
866 
867         if (!table)
868                 return;
869 
870         atomic_set(&wait.count, table->nr);
871         init_completion(&wait.comp);
872 
873         skipped = 0;
874         for (i = 0; i < table->nr; ++i) {
875                 struct kioctx *ctx =
876                         rcu_dereference_protected(table->table[i], true);
877 
878                 if (!ctx) {
879                         skipped++;
880                         continue;
881                 }
882 
883                 /*
884                  * We don't need to bother with munmap() here - exit_mmap(mm)
885                  * is coming and it'll unmap everything. And we simply can't,
886                  * this is not necessarily our ->mm.
887                  * Since kill_ioctx() uses non-zero ->mmap_size as indicator
888                  * that it needs to unmap the area, just set it to 0.
889                  */
890                 ctx->mmap_size = 0;
891                 kill_ioctx(mm, ctx, &wait);
892         }
893 
894         if (!atomic_sub_and_test(skipped, &wait.count)) {
895                 /* Wait until all IO for the context are done. */
896                 wait_for_completion(&wait.comp);
897         }
898 
899         RCU_INIT_POINTER(mm->ioctx_table, NULL);
900         kfree(table);
901 }
902 
903 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
904 {
905         struct kioctx_cpu *kcpu;
906         unsigned long flags;
907 
908         local_irq_save(flags);
909         kcpu = this_cpu_ptr(ctx->cpu);
910         kcpu->reqs_available += nr;
911 
912         while (kcpu->reqs_available >= ctx->req_batch * 2) {
913                 kcpu->reqs_available -= ctx->req_batch;
914                 atomic_add(ctx->req_batch, &ctx->reqs_available);
915         }
916 
917         local_irq_restore(flags);
918 }
919 
920 static bool __get_reqs_available(struct kioctx *ctx)
921 {
922         struct kioctx_cpu *kcpu;
923         bool ret = false;
924         unsigned long flags;
925 
926         local_irq_save(flags);
927         kcpu = this_cpu_ptr(ctx->cpu);
928         if (!kcpu->reqs_available) {
929                 int old, avail = atomic_read(&ctx->reqs_available);
930 
931                 do {
932                         if (avail < ctx->req_batch)
933                                 goto out;
934 
935                         old = avail;
936                         avail = atomic_cmpxchg(&ctx->reqs_available,
937                                                avail, avail - ctx->req_batch);
938                 } while (avail != old);
939 
940                 kcpu->reqs_available += ctx->req_batch;
941         }
942 
943         ret = true;
944         kcpu->reqs_available--;
945 out:
946         local_irq_restore(flags);
947         return ret;
948 }
949 
950 /* refill_reqs_available
951  *      Updates the reqs_available reference counts used for tracking the
952  *      number of free slots in the completion ring.  This can be called
953  *      from aio_complete() (to optimistically update reqs_available) or
954  *      from aio_get_req() (the we're out of events case).  It must be
955  *      called holding ctx->completion_lock.
956  */
957 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
958                                   unsigned tail)
959 {
960         unsigned events_in_ring, completed;
961 
962         /* Clamp head since userland can write to it. */
963         head %= ctx->nr_events;
964         if (head <= tail)
965                 events_in_ring = tail - head;
966         else
967                 events_in_ring = ctx->nr_events - (head - tail);
968 
969         completed = ctx->completed_events;
970         if (events_in_ring < completed)
971                 completed -= events_in_ring;
972         else
973                 completed = 0;
974 
975         if (!completed)
976                 return;
977 
978         ctx->completed_events -= completed;
979         put_reqs_available(ctx, completed);
980 }
981 
982 /* user_refill_reqs_available
983  *      Called to refill reqs_available when aio_get_req() encounters an
984  *      out of space in the completion ring.
985  */
986 static void user_refill_reqs_available(struct kioctx *ctx)
987 {
988         spin_lock_irq(&ctx->completion_lock);
989         if (ctx->completed_events) {
990                 struct aio_ring *ring;
991                 unsigned head;
992 
993                 /* Access of ring->head may race with aio_read_events_ring()
994                  * here, but that's okay since whether we read the old version
995                  * or the new version, and either will be valid.  The important
996                  * part is that head cannot pass tail since we prevent
997                  * aio_complete() from updating tail by holding
998                  * ctx->completion_lock.  Even if head is invalid, the check
999                  * against ctx->completed_events below will make sure we do the
1000                  * safe/right thing.
1001                  */
1002                 ring = kmap_atomic(ctx->ring_pages[0]);
1003                 head = ring->head;
1004                 kunmap_atomic(ring);
1005 
1006                 refill_reqs_available(ctx, head, ctx->tail);
1007         }
1008 
1009         spin_unlock_irq(&ctx->completion_lock);
1010 }
1011 
1012 static bool get_reqs_available(struct kioctx *ctx)
1013 {
1014         if (__get_reqs_available(ctx))
1015                 return true;
1016         user_refill_reqs_available(ctx);
1017         return __get_reqs_available(ctx);
1018 }
1019 
1020 /* aio_get_req
1021  *      Allocate a slot for an aio request.
1022  * Returns NULL if no requests are free.
1023  *
1024  * The refcount is initialized to 2 - one for the async op completion,
1025  * one for the synchronous code that does this.
1026  */
1027 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1028 {
1029         struct aio_kiocb *req;
1030 
1031         req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1032         if (unlikely(!req))
1033                 return NULL;
1034 
1035         if (unlikely(!get_reqs_available(ctx))) {
1036                 kmem_cache_free(kiocb_cachep, req);
1037                 return NULL;
1038         }
1039 
1040         percpu_ref_get(&ctx->reqs);
1041         req->ki_ctx = ctx;
1042         INIT_LIST_HEAD(&req->ki_list);
1043         refcount_set(&req->ki_refcnt, 2);
1044         req->ki_eventfd = NULL;
1045         return req;
1046 }
1047 
1048 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1049 {
1050         struct aio_ring __user *ring  = (void __user *)ctx_id;
1051         struct mm_struct *mm = current->mm;
1052         struct kioctx *ctx, *ret = NULL;
1053         struct kioctx_table *table;
1054         unsigned id;
1055 
1056         if (get_user(id, &ring->id))
1057                 return NULL;
1058 
1059         rcu_read_lock();
1060         table = rcu_dereference(mm->ioctx_table);
1061 
1062         if (!table || id >= table->nr)
1063                 goto out;
1064 
1065         id = array_index_nospec(id, table->nr);
1066         ctx = rcu_dereference(table->table[id]);
1067         if (ctx && ctx->user_id == ctx_id) {
1068                 if (percpu_ref_tryget_live(&ctx->users))
1069                         ret = ctx;
1070         }
1071 out:
1072         rcu_read_unlock();
1073         return ret;
1074 }
1075 
1076 static inline void iocb_destroy(struct aio_kiocb *iocb)
1077 {
1078         if (iocb->ki_eventfd)
1079                 eventfd_ctx_put(iocb->ki_eventfd);
1080         if (iocb->ki_filp)
1081                 fput(iocb->ki_filp);
1082         percpu_ref_put(&iocb->ki_ctx->reqs);
1083         kmem_cache_free(kiocb_cachep, iocb);
1084 }
1085 
1086 /* aio_complete
1087  *      Called when the io request on the given iocb is complete.
1088  */
1089 static void aio_complete(struct aio_kiocb *iocb)
1090 {
1091         struct kioctx   *ctx = iocb->ki_ctx;
1092         struct aio_ring *ring;
1093         struct io_event *ev_page, *event;
1094         unsigned tail, pos, head;
1095         unsigned long   flags;
1096 
1097         /*
1098          * Add a completion event to the ring buffer. Must be done holding
1099          * ctx->completion_lock to prevent other code from messing with the tail
1100          * pointer since we might be called from irq context.
1101          */
1102         spin_lock_irqsave(&ctx->completion_lock, flags);
1103 
1104         tail = ctx->tail;
1105         pos = tail + AIO_EVENTS_OFFSET;
1106 
1107         if (++tail >= ctx->nr_events)
1108                 tail = 0;
1109 
1110         ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1111         event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1112 
1113         *event = iocb->ki_res;
1114 
1115         kunmap_atomic(ev_page);
1116         flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1117 
1118         pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1119                  (void __user *)(unsigned long)iocb->ki_res.obj,
1120                  iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1121 
1122         /* after flagging the request as done, we
1123          * must never even look at it again
1124          */
1125         smp_wmb();      /* make event visible before updating tail */
1126 
1127         ctx->tail = tail;
1128 
1129         ring = kmap_atomic(ctx->ring_pages[0]);
1130         head = ring->head;
1131         ring->tail = tail;
1132         kunmap_atomic(ring);
1133         flush_dcache_page(ctx->ring_pages[0]);
1134 
1135         ctx->completed_events++;
1136         if (ctx->completed_events > 1)
1137                 refill_reqs_available(ctx, head, tail);
1138         spin_unlock_irqrestore(&ctx->completion_lock, flags);
1139 
1140         pr_debug("added to ring %p at [%u]\n", iocb, tail);
1141 
1142         /*
1143          * Check if the user asked us to deliver the result through an
1144          * eventfd. The eventfd_signal() function is safe to be called
1145          * from IRQ context.
1146          */
1147         if (iocb->ki_eventfd)
1148                 eventfd_signal(iocb->ki_eventfd, 1);
1149 
1150         /*
1151          * We have to order our ring_info tail store above and test
1152          * of the wait list below outside the wait lock.  This is
1153          * like in wake_up_bit() where clearing a bit has to be
1154          * ordered with the unlocked test.
1155          */
1156         smp_mb();
1157 
1158         if (waitqueue_active(&ctx->wait))
1159                 wake_up(&ctx->wait);
1160 }
1161 
1162 static inline void iocb_put(struct aio_kiocb *iocb)
1163 {
1164         if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1165                 aio_complete(iocb);
1166                 iocb_destroy(iocb);
1167         }
1168 }
1169 
1170 /* aio_read_events_ring
1171  *      Pull an event off of the ioctx's event ring.  Returns the number of
1172  *      events fetched
1173  */
1174 static long aio_read_events_ring(struct kioctx *ctx,
1175                                  struct io_event __user *event, long nr)
1176 {
1177         struct aio_ring *ring;
1178         unsigned head, tail, pos;
1179         long ret = 0;
1180         int copy_ret;
1181 
1182         /*
1183          * The mutex can block and wake us up and that will cause
1184          * wait_event_interruptible_hrtimeout() to schedule without sleeping
1185          * and repeat. This should be rare enough that it doesn't cause
1186          * peformance issues. See the comment in read_events() for more detail.
1187          */
1188         sched_annotate_sleep();
1189         mutex_lock(&ctx->ring_lock);
1190 
1191         /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1192         ring = kmap_atomic(ctx->ring_pages[0]);
1193         head = ring->head;
1194         tail = ring->tail;
1195         kunmap_atomic(ring);
1196 
1197         /*
1198          * Ensure that once we've read the current tail pointer, that
1199          * we also see the events that were stored up to the tail.
1200          */
1201         smp_rmb();
1202 
1203         pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1204 
1205         if (head == tail)
1206                 goto out;
1207 
1208         head %= ctx->nr_events;
1209         tail %= ctx->nr_events;
1210 
1211         while (ret < nr) {
1212                 long avail;
1213                 struct io_event *ev;
1214                 struct page *page;
1215 
1216                 avail = (head <= tail ?  tail : ctx->nr_events) - head;
1217                 if (head == tail)
1218                         break;
1219 
1220                 pos = head + AIO_EVENTS_OFFSET;
1221                 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1222                 pos %= AIO_EVENTS_PER_PAGE;
1223 
1224                 avail = min(avail, nr - ret);
1225                 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1226 
1227                 ev = kmap(page);
1228                 copy_ret = copy_to_user(event + ret, ev + pos,
1229                                         sizeof(*ev) * avail);
1230                 kunmap(page);
1231 
1232                 if (unlikely(copy_ret)) {
1233                         ret = -EFAULT;
1234                         goto out;
1235                 }
1236 
1237                 ret += avail;
1238                 head += avail;
1239                 head %= ctx->nr_events;
1240         }
1241 
1242         ring = kmap_atomic(ctx->ring_pages[0]);
1243         ring->head = head;
1244         kunmap_atomic(ring);
1245         flush_dcache_page(ctx->ring_pages[0]);
1246 
1247         pr_debug("%li  h%u t%u\n", ret, head, tail);
1248 out:
1249         mutex_unlock(&ctx->ring_lock);
1250 
1251         return ret;
1252 }
1253 
1254 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1255                             struct io_event __user *event, long *i)
1256 {
1257         long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1258 
1259         if (ret > 0)
1260                 *i += ret;
1261 
1262         if (unlikely(atomic_read(&ctx->dead)))
1263                 ret = -EINVAL;
1264 
1265         if (!*i)
1266                 *i = ret;
1267 
1268         return ret < 0 || *i >= min_nr;
1269 }
1270 
1271 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1272                         struct io_event __user *event,
1273                         ktime_t until)
1274 {
1275         long ret = 0;
1276 
1277         /*
1278          * Note that aio_read_events() is being called as the conditional - i.e.
1279          * we're calling it after prepare_to_wait() has set task state to
1280          * TASK_INTERRUPTIBLE.
1281          *
1282          * But aio_read_events() can block, and if it blocks it's going to flip
1283          * the task state back to TASK_RUNNING.
1284          *
1285          * This should be ok, provided it doesn't flip the state back to
1286          * TASK_RUNNING and return 0 too much - that causes us to spin. That
1287          * will only happen if the mutex_lock() call blocks, and we then find
1288          * the ringbuffer empty. So in practice we should be ok, but it's
1289          * something to be aware of when touching this code.
1290          */
1291         if (until == 0)
1292                 aio_read_events(ctx, min_nr, nr, event, &ret);
1293         else
1294                 wait_event_interruptible_hrtimeout(ctx->wait,
1295                                 aio_read_events(ctx, min_nr, nr, event, &ret),
1296                                 until);
1297         return ret;
1298 }
1299 
1300 /* sys_io_setup:
1301  *      Create an aio_context capable of receiving at least nr_events.
1302  *      ctxp must not point to an aio_context that already exists, and
1303  *      must be initialized to 0 prior to the call.  On successful
1304  *      creation of the aio_context, *ctxp is filled in with the resulting 
1305  *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1306  *      if the specified nr_events exceeds internal limits.  May fail 
1307  *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1308  *      of available events.  May fail with -ENOMEM if insufficient kernel
1309  *      resources are available.  May fail with -EFAULT if an invalid
1310  *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1311  *      implemented.
1312  */
1313 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1314 {
1315         struct kioctx *ioctx = NULL;
1316         unsigned long ctx;
1317         long ret;
1318 
1319         ret = get_user(ctx, ctxp);
1320         if (unlikely(ret))
1321                 goto out;
1322 
1323         ret = -EINVAL;
1324         if (unlikely(ctx || nr_events == 0)) {
1325                 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1326                          ctx, nr_events);
1327                 goto out;
1328         }
1329 
1330         ioctx = ioctx_alloc(nr_events);
1331         ret = PTR_ERR(ioctx);
1332         if (!IS_ERR(ioctx)) {
1333                 ret = put_user(ioctx->user_id, ctxp);
1334                 if (ret)
1335                         kill_ioctx(current->mm, ioctx, NULL);
1336                 percpu_ref_put(&ioctx->users);
1337         }
1338 
1339 out:
1340         return ret;
1341 }
1342 
1343 #ifdef CONFIG_COMPAT
1344 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1345 {
1346         struct kioctx *ioctx = NULL;
1347         unsigned long ctx;
1348         long ret;
1349 
1350         ret = get_user(ctx, ctx32p);
1351         if (unlikely(ret))
1352                 goto out;
1353 
1354         ret = -EINVAL;
1355         if (unlikely(ctx || nr_events == 0)) {
1356                 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1357                          ctx, nr_events);
1358                 goto out;
1359         }
1360 
1361         ioctx = ioctx_alloc(nr_events);
1362         ret = PTR_ERR(ioctx);
1363         if (!IS_ERR(ioctx)) {
1364                 /* truncating is ok because it's a user address */
1365                 ret = put_user((u32)ioctx->user_id, ctx32p);
1366                 if (ret)
1367                         kill_ioctx(current->mm, ioctx, NULL);
1368                 percpu_ref_put(&ioctx->users);
1369         }
1370 
1371 out:
1372         return ret;
1373 }
1374 #endif
1375 
1376 /* sys_io_destroy:
1377  *      Destroy the aio_context specified.  May cancel any outstanding 
1378  *      AIOs and block on completion.  Will fail with -ENOSYS if not
1379  *      implemented.  May fail with -EINVAL if the context pointed to
1380  *      is invalid.
1381  */
1382 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1383 {
1384         struct kioctx *ioctx = lookup_ioctx(ctx);
1385         if (likely(NULL != ioctx)) {
1386                 struct ctx_rq_wait wait;
1387                 int ret;
1388 
1389                 init_completion(&wait.comp);
1390                 atomic_set(&wait.count, 1);
1391 
1392                 /* Pass requests_done to kill_ioctx() where it can be set
1393                  * in a thread-safe way. If we try to set it here then we have
1394                  * a race condition if two io_destroy() called simultaneously.
1395                  */
1396                 ret = kill_ioctx(current->mm, ioctx, &wait);
1397                 percpu_ref_put(&ioctx->users);
1398 
1399                 /* Wait until all IO for the context are done. Otherwise kernel
1400                  * keep using user-space buffers even if user thinks the context
1401                  * is destroyed.
1402                  */
1403                 if (!ret)
1404                         wait_for_completion(&wait.comp);
1405 
1406                 return ret;
1407         }
1408         pr_debug("EINVAL: invalid context id\n");
1409         return -EINVAL;
1410 }
1411 
1412 static void aio_remove_iocb(struct aio_kiocb *iocb)
1413 {
1414         struct kioctx *ctx = iocb->ki_ctx;
1415         unsigned long flags;
1416 
1417         spin_lock_irqsave(&ctx->ctx_lock, flags);
1418         list_del(&iocb->ki_list);
1419         spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1420 }
1421 
1422 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1423 {
1424         struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1425 
1426         if (!list_empty_careful(&iocb->ki_list))
1427                 aio_remove_iocb(iocb);
1428 
1429         if (kiocb->ki_flags & IOCB_WRITE) {
1430                 struct inode *inode = file_inode(kiocb->ki_filp);
1431 
1432                 /*
1433                  * Tell lockdep we inherited freeze protection from submission
1434                  * thread.
1435                  */
1436                 if (S_ISREG(inode->i_mode))
1437                         __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1438                 file_end_write(kiocb->ki_filp);
1439         }
1440 
1441         iocb->ki_res.res = res;
1442         iocb->ki_res.res2 = res2;
1443         iocb_put(iocb);
1444 }
1445 
1446 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1447 {
1448         int ret;
1449 
1450         req->ki_complete = aio_complete_rw;
1451         req->private = NULL;
1452         req->ki_pos = iocb->aio_offset;
1453         req->ki_flags = iocb_flags(req->ki_filp);
1454         if (iocb->aio_flags & IOCB_FLAG_RESFD)
1455                 req->ki_flags |= IOCB_EVENTFD;
1456         req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1457         if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1458                 /*
1459                  * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1460                  * aio_reqprio is interpreted as an I/O scheduling
1461                  * class and priority.
1462                  */
1463                 ret = ioprio_check_cap(iocb->aio_reqprio);
1464                 if (ret) {
1465                         pr_debug("aio ioprio check cap error: %d\n", ret);
1466                         return ret;
1467                 }
1468 
1469                 req->ki_ioprio = iocb->aio_reqprio;
1470         } else
1471                 req->ki_ioprio = get_current_ioprio();
1472 
1473         ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1474         if (unlikely(ret))
1475                 return ret;
1476 
1477         req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1478         return 0;
1479 }
1480 
1481 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1482                 struct iovec **iovec, bool vectored, bool compat,
1483                 struct iov_iter *iter)
1484 {
1485         void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1486         size_t len = iocb->aio_nbytes;
1487 
1488         if (!vectored) {
1489                 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1490                 *iovec = NULL;
1491                 return ret;
1492         }
1493 #ifdef CONFIG_COMPAT
1494         if (compat)
1495                 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1496                                 iter);
1497 #endif
1498         return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1499 }
1500 
1501 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1502 {
1503         switch (ret) {
1504         case -EIOCBQUEUED:
1505                 break;
1506         case -ERESTARTSYS:
1507         case -ERESTARTNOINTR:
1508         case -ERESTARTNOHAND:
1509         case -ERESTART_RESTARTBLOCK:
1510                 /*
1511                  * There's no easy way to restart the syscall since other AIO's
1512                  * may be already running. Just fail this IO with EINTR.
1513                  */
1514                 ret = -EINTR;
1515                 /*FALLTHRU*/
1516         default:
1517                 req->ki_complete(req, ret, 0);
1518         }
1519 }
1520 
1521 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1522                         bool vectored, bool compat)
1523 {
1524         struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1525         struct iov_iter iter;
1526         struct file *file;
1527         int ret;
1528 
1529         ret = aio_prep_rw(req, iocb);
1530         if (ret)
1531                 return ret;
1532         file = req->ki_filp;
1533         if (unlikely(!(file->f_mode & FMODE_READ)))
1534                 return -EBADF;
1535         ret = -EINVAL;
1536         if (unlikely(!file->f_op->read_iter))
1537                 return -EINVAL;
1538 
1539         ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1540         if (ret < 0)
1541                 return ret;
1542         ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1543         if (!ret)
1544                 aio_rw_done(req, call_read_iter(file, req, &iter));
1545         kfree(iovec);
1546         return ret;
1547 }
1548 
1549 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1550                          bool vectored, bool compat)
1551 {
1552         struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1553         struct iov_iter iter;
1554         struct file *file;
1555         int ret;
1556 
1557         ret = aio_prep_rw(req, iocb);
1558         if (ret)
1559                 return ret;
1560         file = req->ki_filp;
1561 
1562         if (unlikely(!(file->f_mode & FMODE_WRITE)))
1563                 return -EBADF;
1564         if (unlikely(!file->f_op->write_iter))
1565                 return -EINVAL;
1566 
1567         ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1568         if (ret < 0)
1569                 return ret;
1570         ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1571         if (!ret) {
1572                 /*
1573                  * Open-code file_start_write here to grab freeze protection,
1574                  * which will be released by another thread in
1575                  * aio_complete_rw().  Fool lockdep by telling it the lock got
1576                  * released so that it doesn't complain about the held lock when
1577                  * we return to userspace.
1578                  */
1579                 if (S_ISREG(file_inode(file)->i_mode)) {
1580                         __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true);
1581                         __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1582                 }
1583                 req->ki_flags |= IOCB_WRITE;
1584                 aio_rw_done(req, call_write_iter(file, req, &iter));
1585         }
1586         kfree(iovec);
1587         return ret;
1588 }
1589 
1590 static void aio_fsync_work(struct work_struct *work)
1591 {
1592         struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1593         const struct cred *old_cred = override_creds(iocb->fsync.creds);
1594 
1595         iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1596         revert_creds(old_cred);
1597         put_cred(iocb->fsync.creds);
1598         iocb_put(iocb);
1599 }
1600 
1601 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1602                      bool datasync)
1603 {
1604         if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1605                         iocb->aio_rw_flags))
1606                 return -EINVAL;
1607 
1608         if (unlikely(!req->file->f_op->fsync))
1609                 return -EINVAL;
1610 
1611         req->creds = prepare_creds();
1612         if (!req->creds)
1613                 return -ENOMEM;
1614 
1615         req->datasync = datasync;
1616         INIT_WORK(&req->work, aio_fsync_work);
1617         schedule_work(&req->work);
1618         return 0;
1619 }
1620 
1621 static void aio_poll_put_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 
1626         iocb_put(iocb);
1627 }
1628 
1629 static void aio_poll_complete_work(struct work_struct *work)
1630 {
1631         struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1632         struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1633         struct poll_table_struct pt = { ._key = req->events };
1634         struct kioctx *ctx = iocb->ki_ctx;
1635         __poll_t mask = 0;
1636 
1637         if (!READ_ONCE(req->cancelled))
1638                 mask = vfs_poll(req->file, &pt) & req->events;
1639 
1640         /*
1641          * Note that ->ki_cancel callers also delete iocb from active_reqs after
1642          * calling ->ki_cancel.  We need the ctx_lock roundtrip here to
1643          * synchronize with them.  In the cancellation case the list_del_init
1644          * itself is not actually needed, but harmless so we keep it in to
1645          * avoid further branches in the fast path.
1646          */
1647         spin_lock_irq(&ctx->ctx_lock);
1648         if (!mask && !READ_ONCE(req->cancelled)) {
1649                 add_wait_queue(req->head, &req->wait);
1650                 spin_unlock_irq(&ctx->ctx_lock);
1651                 return;
1652         }
1653         list_del_init(&iocb->ki_list);
1654         iocb->ki_res.res = mangle_poll(mask);
1655         req->done = true;
1656         spin_unlock_irq(&ctx->ctx_lock);
1657 
1658         iocb_put(iocb);
1659 }
1660 
1661 /* assumes we are called with irqs disabled */
1662 static int aio_poll_cancel(struct kiocb *iocb)
1663 {
1664         struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1665         struct poll_iocb *req = &aiocb->poll;
1666 
1667         spin_lock(&req->head->lock);
1668         WRITE_ONCE(req->cancelled, true);
1669         if (!list_empty(&req->wait.entry)) {
1670                 list_del_init(&req->wait.entry);
1671                 schedule_work(&aiocb->poll.work);
1672         }
1673         spin_unlock(&req->head->lock);
1674 
1675         return 0;
1676 }
1677 
1678 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1679                 void *key)
1680 {
1681         struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1682         struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1683         __poll_t mask = key_to_poll(key);
1684         unsigned long flags;
1685 
1686         /* for instances that support it check for an event match first: */
1687         if (mask && !(mask & req->events))
1688                 return 0;
1689 
1690         list_del_init(&req->wait.entry);
1691 
1692         if (mask && spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1693                 struct kioctx *ctx = iocb->ki_ctx;
1694 
1695                 /*
1696                  * Try to complete the iocb inline if we can. Use
1697                  * irqsave/irqrestore because not all filesystems (e.g. fuse)
1698                  * call this function with IRQs disabled and because IRQs
1699                  * have to be disabled before ctx_lock is obtained.
1700                  */
1701                 list_del(&iocb->ki_list);
1702                 iocb->ki_res.res = mangle_poll(mask);
1703                 req->done = true;
1704                 if (iocb->ki_eventfd && eventfd_signal_count()) {
1705                         iocb = NULL;
1706                         INIT_WORK(&req->work, aio_poll_put_work);
1707                         schedule_work(&req->work);
1708                 }
1709                 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1710                 if (iocb)
1711                         iocb_put(iocb);
1712         } else {
1713                 schedule_work(&req->work);
1714         }
1715         return 1;
1716 }
1717 
1718 struct aio_poll_table {
1719         struct poll_table_struct        pt;
1720         struct aio_kiocb                *iocb;
1721         int                             error;
1722 };
1723 
1724 static void
1725 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1726                 struct poll_table_struct *p)
1727 {
1728         struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1729 
1730         /* multiple wait queues per file are not supported */
1731         if (unlikely(pt->iocb->poll.head)) {
1732                 pt->error = -EINVAL;
1733                 return;
1734         }
1735 
1736         pt->error = 0;
1737         pt->iocb->poll.head = head;
1738         add_wait_queue(head, &pt->iocb->poll.wait);
1739 }
1740 
1741 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1742 {
1743         struct kioctx *ctx = aiocb->ki_ctx;
1744         struct poll_iocb *req = &aiocb->poll;
1745         struct aio_poll_table apt;
1746         bool cancel = false;
1747         __poll_t mask;
1748 
1749         /* reject any unknown events outside the normal event mask. */
1750         if ((u16)iocb->aio_buf != iocb->aio_buf)
1751                 return -EINVAL;
1752         /* reject fields that are not defined for poll */
1753         if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1754                 return -EINVAL;
1755 
1756         INIT_WORK(&req->work, aio_poll_complete_work);
1757         req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1758 
1759         req->head = NULL;
1760         req->done = false;
1761         req->cancelled = false;
1762 
1763         apt.pt._qproc = aio_poll_queue_proc;
1764         apt.pt._key = req->events;
1765         apt.iocb = aiocb;
1766         apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1767 
1768         /* initialized the list so that we can do list_empty checks */
1769         INIT_LIST_HEAD(&req->wait.entry);
1770         init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1771 
1772         mask = vfs_poll(req->file, &apt.pt) & req->events;
1773         spin_lock_irq(&ctx->ctx_lock);
1774         if (likely(req->head)) {
1775                 spin_lock(&req->head->lock);
1776                 if (unlikely(list_empty(&req->wait.entry))) {
1777                         if (apt.error)
1778                                 cancel = true;
1779                         apt.error = 0;
1780                         mask = 0;
1781                 }
1782                 if (mask || apt.error) {
1783                         list_del_init(&req->wait.entry);
1784                 } else if (cancel) {
1785                         WRITE_ONCE(req->cancelled, true);
1786                 } else if (!req->done) { /* actually waiting for an event */
1787                         list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1788                         aiocb->ki_cancel = aio_poll_cancel;
1789                 }
1790                 spin_unlock(&req->head->lock);
1791         }
1792         if (mask) { /* no async, we'd stolen it */
1793                 aiocb->ki_res.res = mangle_poll(mask);
1794                 apt.error = 0;
1795         }
1796         spin_unlock_irq(&ctx->ctx_lock);
1797         if (mask)
1798                 iocb_put(aiocb);
1799         return apt.error;
1800 }
1801 
1802 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1803                            struct iocb __user *user_iocb, struct aio_kiocb *req,
1804                            bool compat)
1805 {
1806         req->ki_filp = fget(iocb->aio_fildes);
1807         if (unlikely(!req->ki_filp))
1808                 return -EBADF;
1809 
1810         if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1811                 struct eventfd_ctx *eventfd;
1812                 /*
1813                  * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1814                  * instance of the file* now. The file descriptor must be
1815                  * an eventfd() fd, and will be signaled for each completed
1816                  * event using the eventfd_signal() function.
1817                  */
1818                 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1819                 if (IS_ERR(eventfd))
1820                         return PTR_ERR(eventfd);
1821 
1822                 req->ki_eventfd = eventfd;
1823         }
1824 
1825         if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1826                 pr_debug("EFAULT: aio_key\n");
1827                 return -EFAULT;
1828         }
1829 
1830         req->ki_res.obj = (u64)(unsigned long)user_iocb;
1831         req->ki_res.data = iocb->aio_data;
1832         req->ki_res.res = 0;
1833         req->ki_res.res2 = 0;
1834 
1835         switch (iocb->aio_lio_opcode) {
1836         case IOCB_CMD_PREAD:
1837                 return aio_read(&req->rw, iocb, false, compat);
1838         case IOCB_CMD_PWRITE:
1839                 return aio_write(&req->rw, iocb, false, compat);
1840         case IOCB_CMD_PREADV:
1841                 return aio_read(&req->rw, iocb, true, compat);
1842         case IOCB_CMD_PWRITEV:
1843                 return aio_write(&req->rw, iocb, true, compat);
1844         case IOCB_CMD_FSYNC:
1845                 return aio_fsync(&req->fsync, iocb, false);
1846         case IOCB_CMD_FDSYNC:
1847                 return aio_fsync(&req->fsync, iocb, true);
1848         case IOCB_CMD_POLL:
1849                 return aio_poll(req, iocb);
1850         default:
1851                 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1852                 return -EINVAL;
1853         }
1854 }
1855 
1856 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1857                          bool compat)
1858 {
1859         struct aio_kiocb *req;
1860         struct iocb iocb;
1861         int err;
1862 
1863         if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1864                 return -EFAULT;
1865 
1866         /* enforce forwards compatibility on users */
1867         if (unlikely(iocb.aio_reserved2)) {
1868                 pr_debug("EINVAL: reserve field set\n");
1869                 return -EINVAL;
1870         }
1871 
1872         /* prevent overflows */
1873         if (unlikely(
1874             (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1875             (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1876             ((ssize_t)iocb.aio_nbytes < 0)
1877            )) {
1878                 pr_debug("EINVAL: overflow check\n");
1879                 return -EINVAL;
1880         }
1881 
1882         req = aio_get_req(ctx);
1883         if (unlikely(!req))
1884                 return -EAGAIN;
1885 
1886         err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
1887 
1888         /* Done with the synchronous reference */
1889         iocb_put(req);
1890 
1891         /*
1892          * If err is 0, we'd either done aio_complete() ourselves or have
1893          * arranged for that to be done asynchronously.  Anything non-zero
1894          * means that we need to destroy req ourselves.
1895          */
1896         if (unlikely(err)) {
1897                 iocb_destroy(req);
1898                 put_reqs_available(ctx, 1);
1899         }
1900         return err;
1901 }
1902 
1903 /* sys_io_submit:
1904  *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1905  *      the number of iocbs queued.  May return -EINVAL if the aio_context
1906  *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1907  *      *iocbpp[0] is not properly initialized, if the operation specified
1908  *      is invalid for the file descriptor in the iocb.  May fail with
1909  *      -EFAULT if any of the data structures point to invalid data.  May
1910  *      fail with -EBADF if the file descriptor specified in the first
1911  *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1912  *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1913  *      fail with -ENOSYS if not implemented.
1914  */
1915 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1916                 struct iocb __user * __user *, iocbpp)
1917 {
1918         struct kioctx *ctx;
1919         long ret = 0;
1920         int i = 0;
1921         struct blk_plug plug;
1922 
1923         if (unlikely(nr < 0))
1924                 return -EINVAL;
1925 
1926         ctx = lookup_ioctx(ctx_id);
1927         if (unlikely(!ctx)) {
1928                 pr_debug("EINVAL: invalid context id\n");
1929                 return -EINVAL;
1930         }
1931 
1932         if (nr > ctx->nr_events)
1933                 nr = ctx->nr_events;
1934 
1935         if (nr > AIO_PLUG_THRESHOLD)
1936                 blk_start_plug(&plug);
1937         for (i = 0; i < nr; i++) {
1938                 struct iocb __user *user_iocb;
1939 
1940                 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1941                         ret = -EFAULT;
1942                         break;
1943                 }
1944 
1945                 ret = io_submit_one(ctx, user_iocb, false);
1946                 if (ret)
1947                         break;
1948         }
1949         if (nr > AIO_PLUG_THRESHOLD)
1950                 blk_finish_plug(&plug);
1951 
1952         percpu_ref_put(&ctx->users);
1953         return i ? i : ret;
1954 }
1955 
1956 #ifdef CONFIG_COMPAT
1957 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1958                        int, nr, compat_uptr_t __user *, iocbpp)
1959 {
1960         struct kioctx *ctx;
1961         long ret = 0;
1962         int i = 0;
1963         struct blk_plug plug;
1964 
1965         if (unlikely(nr < 0))
1966                 return -EINVAL;
1967 
1968         ctx = lookup_ioctx(ctx_id);
1969         if (unlikely(!ctx)) {
1970                 pr_debug("EINVAL: invalid context id\n");
1971                 return -EINVAL;
1972         }
1973 
1974         if (nr > ctx->nr_events)
1975                 nr = ctx->nr_events;
1976 
1977         if (nr > AIO_PLUG_THRESHOLD)
1978                 blk_start_plug(&plug);
1979         for (i = 0; i < nr; i++) {
1980                 compat_uptr_t user_iocb;
1981 
1982                 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1983                         ret = -EFAULT;
1984                         break;
1985                 }
1986 
1987                 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
1988                 if (ret)
1989                         break;
1990         }
1991         if (nr > AIO_PLUG_THRESHOLD)
1992                 blk_finish_plug(&plug);
1993 
1994         percpu_ref_put(&ctx->users);
1995         return i ? i : ret;
1996 }
1997 #endif
1998 
1999 /* sys_io_cancel:
2000  *      Attempts to cancel an iocb previously passed to io_submit.  If
2001  *      the operation is successfully cancelled, the resulting event is
2002  *      copied into the memory pointed to by result without being placed
2003  *      into the completion queue and 0 is returned.  May fail with
2004  *      -EFAULT if any of the data structures pointed to are invalid.
2005  *      May fail with -EINVAL if aio_context specified by ctx_id is
2006  *      invalid.  May fail with -EAGAIN if the iocb specified was not
2007  *      cancelled.  Will fail with -ENOSYS if not implemented.
2008  */
2009 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2010                 struct io_event __user *, result)
2011 {
2012         struct kioctx *ctx;
2013         struct aio_kiocb *kiocb;
2014         int ret = -EINVAL;
2015         u32 key;
2016         u64 obj = (u64)(unsigned long)iocb;
2017 
2018         if (unlikely(get_user(key, &iocb->aio_key)))
2019                 return -EFAULT;
2020         if (unlikely(key != KIOCB_KEY))
2021                 return -EINVAL;
2022 
2023         ctx = lookup_ioctx(ctx_id);
2024         if (unlikely(!ctx))
2025                 return -EINVAL;
2026 
2027         spin_lock_irq(&ctx->ctx_lock);
2028         /* TODO: use a hash or array, this sucks. */
2029         list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2030                 if (kiocb->ki_res.obj == obj) {
2031                         ret = kiocb->ki_cancel(&kiocb->rw);
2032                         list_del_init(&kiocb->ki_list);
2033                         break;
2034                 }
2035         }
2036         spin_unlock_irq(&ctx->ctx_lock);
2037 
2038         if (!ret) {
2039                 /*
2040                  * The result argument is no longer used - the io_event is
2041                  * always delivered via the ring buffer. -EINPROGRESS indicates
2042                  * cancellation is progress:
2043                  */
2044                 ret = -EINPROGRESS;
2045         }
2046 
2047         percpu_ref_put(&ctx->users);
2048 
2049         return ret;
2050 }
2051 
2052 static long do_io_getevents(aio_context_t ctx_id,
2053                 long min_nr,
2054                 long nr,
2055                 struct io_event __user *events,
2056                 struct timespec64 *ts)
2057 {
2058         ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2059         struct kioctx *ioctx = lookup_ioctx(ctx_id);
2060         long ret = -EINVAL;
2061 
2062         if (likely(ioctx)) {
2063                 if (likely(min_nr <= nr && min_nr >= 0))
2064                         ret = read_events(ioctx, min_nr, nr, events, until);
2065                 percpu_ref_put(&ioctx->users);
2066         }
2067 
2068         return ret;
2069 }
2070 
2071 /* io_getevents:
2072  *      Attempts to read at least min_nr events and up to nr events from
2073  *      the completion queue for the aio_context specified by ctx_id. If
2074  *      it succeeds, the number of read events is returned. May fail with
2075  *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2076  *      out of range, if timeout is out of range.  May fail with -EFAULT
2077  *      if any of the memory specified is invalid.  May return 0 or
2078  *      < min_nr if the timeout specified by timeout has elapsed
2079  *      before sufficient events are available, where timeout == NULL
2080  *      specifies an infinite timeout. Note that the timeout pointed to by
2081  *      timeout is relative.  Will fail with -ENOSYS if not implemented.
2082  */
2083 #ifdef CONFIG_64BIT
2084 
2085 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2086                 long, min_nr,
2087                 long, nr,
2088                 struct io_event __user *, events,
2089                 struct __kernel_timespec __user *, timeout)
2090 {
2091         struct timespec64       ts;
2092         int                     ret;
2093 
2094         if (timeout && unlikely(get_timespec64(&ts, timeout)))
2095                 return -EFAULT;
2096 
2097         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2098         if (!ret && signal_pending(current))
2099                 ret = -EINTR;
2100         return ret;
2101 }
2102 
2103 #endif
2104 
2105 struct __aio_sigset {
2106         const sigset_t __user   *sigmask;
2107         size_t          sigsetsize;
2108 };
2109 
2110 SYSCALL_DEFINE6(io_pgetevents,
2111                 aio_context_t, ctx_id,
2112                 long, min_nr,
2113                 long, nr,
2114                 struct io_event __user *, events,
2115                 struct __kernel_timespec __user *, timeout,
2116                 const struct __aio_sigset __user *, usig)
2117 {
2118         struct __aio_sigset     ksig = { NULL, };
2119         struct timespec64       ts;
2120         bool interrupted;
2121         int ret;
2122 
2123         if (timeout && unlikely(get_timespec64(&ts, timeout)))
2124                 return -EFAULT;
2125 
2126         if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2127                 return -EFAULT;
2128 
2129         ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2130         if (ret)
2131                 return ret;
2132 
2133         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2134 
2135         interrupted = signal_pending(current);
2136         restore_saved_sigmask_unless(interrupted);
2137         if (interrupted && !ret)
2138                 ret = -ERESTARTNOHAND;
2139 
2140         return ret;
2141 }
2142 
2143 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2144 
2145 SYSCALL_DEFINE6(io_pgetevents_time32,
2146                 aio_context_t, ctx_id,
2147                 long, min_nr,
2148                 long, nr,
2149                 struct io_event __user *, events,
2150                 struct old_timespec32 __user *, timeout,
2151                 const struct __aio_sigset __user *, usig)
2152 {
2153         struct __aio_sigset     ksig = { NULL, };
2154         struct timespec64       ts;
2155         bool interrupted;
2156         int ret;
2157 
2158         if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2159                 return -EFAULT;
2160 
2161         if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2162                 return -EFAULT;
2163 
2164 
2165         ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2166         if (ret)
2167                 return ret;
2168 
2169         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2170 
2171         interrupted = signal_pending(current);
2172         restore_saved_sigmask_unless(interrupted);
2173         if (interrupted && !ret)
2174                 ret = -ERESTARTNOHAND;
2175 
2176         return ret;
2177 }
2178 
2179 #endif
2180 
2181 #if defined(CONFIG_COMPAT_32BIT_TIME)
2182 
2183 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2184                 __s32, min_nr,
2185                 __s32, nr,
2186                 struct io_event __user *, events,
2187                 struct old_timespec32 __user *, timeout)
2188 {
2189         struct timespec64 t;
2190         int ret;
2191 
2192         if (timeout && get_old_timespec32(&t, timeout))
2193                 return -EFAULT;
2194 
2195         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2196         if (!ret && signal_pending(current))
2197                 ret = -EINTR;
2198         return ret;
2199 }
2200 
2201 #endif
2202 
2203 #ifdef CONFIG_COMPAT
2204 
2205 struct __compat_aio_sigset {
2206         compat_uptr_t           sigmask;
2207         compat_size_t           sigsetsize;
2208 };
2209 
2210 #if defined(CONFIG_COMPAT_32BIT_TIME)
2211 
2212 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2213                 compat_aio_context_t, ctx_id,
2214                 compat_long_t, min_nr,
2215                 compat_long_t, nr,
2216                 struct io_event __user *, events,
2217                 struct old_timespec32 __user *, timeout,
2218                 const struct __compat_aio_sigset __user *, usig)
2219 {
2220         struct __compat_aio_sigset ksig = { 0, };
2221         struct timespec64 t;
2222         bool interrupted;
2223         int ret;
2224 
2225         if (timeout && get_old_timespec32(&t, timeout))
2226                 return -EFAULT;
2227 
2228         if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2229                 return -EFAULT;
2230 
2231         ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2232         if (ret)
2233                 return ret;
2234 
2235         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2236 
2237         interrupted = signal_pending(current);
2238         restore_saved_sigmask_unless(interrupted);
2239         if (interrupted && !ret)
2240                 ret = -ERESTARTNOHAND;
2241 
2242         return ret;
2243 }
2244 
2245 #endif
2246 
2247 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2248                 compat_aio_context_t, ctx_id,
2249                 compat_long_t, min_nr,
2250                 compat_long_t, nr,
2251                 struct io_event __user *, events,
2252                 struct __kernel_timespec __user *, timeout,
2253                 const struct __compat_aio_sigset __user *, usig)
2254 {
2255         struct __compat_aio_sigset ksig = { 0, };
2256         struct timespec64 t;
2257         bool interrupted;
2258         int ret;
2259 
2260         if (timeout && get_timespec64(&t, timeout))
2261                 return -EFAULT;
2262 
2263         if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2264                 return -EFAULT;
2265 
2266         ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2267         if (ret)
2268                 return ret;
2269 
2270         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2271 
2272         interrupted = signal_pending(current);
2273         restore_saved_sigmask_unless(interrupted);
2274         if (interrupted && !ret)
2275                 ret = -ERESTARTNOHAND;
2276 
2277         return ret;
2278 }
2279 #endif
2280 

~ [ 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