Back to home page

LXR

 
 

    


0001 /*
0002  *  An async IO implementation for Linux
0003  *  Written by Benjamin LaHaise <bcrl@kvack.org>
0004  *
0005  *  Implements an efficient asynchronous io interface.
0006  *
0007  *  Copyright 2000, 2001, 2002 Red Hat, Inc.  All Rights Reserved.
0008  *
0009  *  See ../COPYING for licensing terms.
0010  */
0011 #define pr_fmt(fmt) "%s: " fmt, __func__
0012 
0013 #include <linux/kernel.h>
0014 #include <linux/init.h>
0015 #include <linux/errno.h>
0016 #include <linux/time.h>
0017 #include <linux/aio_abi.h>
0018 #include <linux/export.h>
0019 #include <linux/syscalls.h>
0020 #include <linux/backing-dev.h>
0021 #include <linux/uio.h>
0022 
0023 #include <linux/sched.h>
0024 #include <linux/fs.h>
0025 #include <linux/file.h>
0026 #include <linux/mm.h>
0027 #include <linux/mman.h>
0028 #include <linux/mmu_context.h>
0029 #include <linux/percpu.h>
0030 #include <linux/slab.h>
0031 #include <linux/timer.h>
0032 #include <linux/aio.h>
0033 #include <linux/highmem.h>
0034 #include <linux/workqueue.h>
0035 #include <linux/security.h>
0036 #include <linux/eventfd.h>
0037 #include <linux/blkdev.h>
0038 #include <linux/compat.h>
0039 #include <linux/migrate.h>
0040 #include <linux/ramfs.h>
0041 #include <linux/percpu-refcount.h>
0042 #include <linux/mount.h>
0043 
0044 #include <asm/kmap_types.h>
0045 #include <linux/uaccess.h>
0046 
0047 #include "internal.h"
0048 
0049 #define AIO_RING_MAGIC          0xa10a10a1
0050 #define AIO_RING_COMPAT_FEATURES    1
0051 #define AIO_RING_INCOMPAT_FEATURES  0
0052 struct aio_ring {
0053     unsigned    id; /* kernel internal index number */
0054     unsigned    nr; /* number of io_events */
0055     unsigned    head;   /* Written to by userland or under ring_lock
0056                  * mutex by aio_read_events_ring(). */
0057     unsigned    tail;
0058 
0059     unsigned    magic;
0060     unsigned    compat_features;
0061     unsigned    incompat_features;
0062     unsigned    header_length;  /* size of aio_ring */
0063 
0064 
0065     struct io_event     io_events[0];
0066 }; /* 128 bytes + ring size */
0067 
0068 #define AIO_RING_PAGES  8
0069 
0070 struct kioctx_table {
0071     struct rcu_head rcu;
0072     unsigned    nr;
0073     struct kioctx   *table[];
0074 };
0075 
0076 struct kioctx_cpu {
0077     unsigned        reqs_available;
0078 };
0079 
0080 struct ctx_rq_wait {
0081     struct completion comp;
0082     atomic_t count;
0083 };
0084 
0085 struct kioctx {
0086     struct percpu_ref   users;
0087     atomic_t        dead;
0088 
0089     struct percpu_ref   reqs;
0090 
0091     unsigned long       user_id;
0092 
0093     struct __percpu kioctx_cpu *cpu;
0094 
0095     /*
0096      * For percpu reqs_available, number of slots we move to/from global
0097      * counter at a time:
0098      */
0099     unsigned        req_batch;
0100     /*
0101      * This is what userspace passed to io_setup(), it's not used for
0102      * anything but counting against the global max_reqs quota.
0103      *
0104      * The real limit is nr_events - 1, which will be larger (see
0105      * aio_setup_ring())
0106      */
0107     unsigned        max_reqs;
0108 
0109     /* Size of ringbuffer, in units of struct io_event */
0110     unsigned        nr_events;
0111 
0112     unsigned long       mmap_base;
0113     unsigned long       mmap_size;
0114 
0115     struct page     **ring_pages;
0116     long            nr_pages;
0117 
0118     struct work_struct  free_work;
0119 
0120     /*
0121      * signals when all in-flight requests are done
0122      */
0123     struct ctx_rq_wait  *rq_wait;
0124 
0125     struct {
0126         /*
0127          * This counts the number of available slots in the ringbuffer,
0128          * so we avoid overflowing it: it's decremented (if positive)
0129          * when allocating a kiocb and incremented when the resulting
0130          * io_event is pulled off the ringbuffer.
0131          *
0132          * We batch accesses to it with a percpu version.
0133          */
0134         atomic_t    reqs_available;
0135     } ____cacheline_aligned_in_smp;
0136 
0137     struct {
0138         spinlock_t  ctx_lock;
0139         struct list_head active_reqs;   /* used for cancellation */
0140     } ____cacheline_aligned_in_smp;
0141 
0142     struct {
0143         struct mutex    ring_lock;
0144         wait_queue_head_t wait;
0145     } ____cacheline_aligned_in_smp;
0146 
0147     struct {
0148         unsigned    tail;
0149         unsigned    completed_events;
0150         spinlock_t  completion_lock;
0151     } ____cacheline_aligned_in_smp;
0152 
0153     struct page     *internal_pages[AIO_RING_PAGES];
0154     struct file     *aio_ring_file;
0155 
0156     unsigned        id;
0157 };
0158 
0159 /*
0160  * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
0161  * cancelled or completed (this makes a certain amount of sense because
0162  * successful cancellation - io_cancel() - does deliver the completion to
0163  * userspace).
0164  *
0165  * And since most things don't implement kiocb cancellation and we'd really like
0166  * kiocb completion to be lockless when possible, we use ki_cancel to
0167  * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
0168  * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
0169  */
0170 #define KIOCB_CANCELLED     ((void *) (~0ULL))
0171 
0172 struct aio_kiocb {
0173     struct kiocb        common;
0174 
0175     struct kioctx       *ki_ctx;
0176     kiocb_cancel_fn     *ki_cancel;
0177 
0178     struct iocb __user  *ki_user_iocb;  /* user's aiocb */
0179     __u64           ki_user_data;   /* user's data for completion */
0180 
0181     struct list_head    ki_list;    /* the aio core uses this
0182                          * for cancellation */
0183 
0184     /*
0185      * If the aio_resfd field of the userspace iocb is not zero,
0186      * this is the underlying eventfd context to deliver events to.
0187      */
0188     struct eventfd_ctx  *ki_eventfd;
0189 };
0190 
0191 /*------ sysctl variables----*/
0192 static DEFINE_SPINLOCK(aio_nr_lock);
0193 unsigned long aio_nr;       /* current system wide number of aio requests */
0194 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
0195 /*----end sysctl variables---*/
0196 
0197 static struct kmem_cache    *kiocb_cachep;
0198 static struct kmem_cache    *kioctx_cachep;
0199 
0200 static struct vfsmount *aio_mnt;
0201 
0202 static const struct file_operations aio_ring_fops;
0203 static const struct address_space_operations aio_ctx_aops;
0204 
0205 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
0206 {
0207     struct qstr this = QSTR_INIT("[aio]", 5);
0208     struct file *file;
0209     struct path path;
0210     struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
0211     if (IS_ERR(inode))
0212         return ERR_CAST(inode);
0213 
0214     inode->i_mapping->a_ops = &aio_ctx_aops;
0215     inode->i_mapping->private_data = ctx;
0216     inode->i_size = PAGE_SIZE * nr_pages;
0217 
0218     path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
0219     if (!path.dentry) {
0220         iput(inode);
0221         return ERR_PTR(-ENOMEM);
0222     }
0223     path.mnt = mntget(aio_mnt);
0224 
0225     d_instantiate(path.dentry, inode);
0226     file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
0227     if (IS_ERR(file)) {
0228         path_put(&path);
0229         return file;
0230     }
0231 
0232     file->f_flags = O_RDWR;
0233     return file;
0234 }
0235 
0236 static struct dentry *aio_mount(struct file_system_type *fs_type,
0237                 int flags, const char *dev_name, void *data)
0238 {
0239     static const struct dentry_operations ops = {
0240         .d_dname    = simple_dname,
0241     };
0242     struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops,
0243                        AIO_RING_MAGIC);
0244 
0245     if (!IS_ERR(root))
0246         root->d_sb->s_iflags |= SB_I_NOEXEC;
0247     return root;
0248 }
0249 
0250 /* aio_setup
0251  *  Creates the slab caches used by the aio routines, panic on
0252  *  failure as this is done early during the boot sequence.
0253  */
0254 static int __init aio_setup(void)
0255 {
0256     static struct file_system_type aio_fs = {
0257         .name       = "aio",
0258         .mount      = aio_mount,
0259         .kill_sb    = kill_anon_super,
0260     };
0261     aio_mnt = kern_mount(&aio_fs);
0262     if (IS_ERR(aio_mnt))
0263         panic("Failed to create aio fs mount.");
0264 
0265     kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
0266     kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
0267 
0268     pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
0269 
0270     return 0;
0271 }
0272 __initcall(aio_setup);
0273 
0274 static void put_aio_ring_file(struct kioctx *ctx)
0275 {
0276     struct file *aio_ring_file = ctx->aio_ring_file;
0277     struct address_space *i_mapping;
0278 
0279     if (aio_ring_file) {
0280         truncate_setsize(file_inode(aio_ring_file), 0);
0281 
0282         /* Prevent further access to the kioctx from migratepages */
0283         i_mapping = aio_ring_file->f_mapping;
0284         spin_lock(&i_mapping->private_lock);
0285         i_mapping->private_data = NULL;
0286         ctx->aio_ring_file = NULL;
0287         spin_unlock(&i_mapping->private_lock);
0288 
0289         fput(aio_ring_file);
0290     }
0291 }
0292 
0293 static void aio_free_ring(struct kioctx *ctx)
0294 {
0295     int i;
0296 
0297     /* Disconnect the kiotx from the ring file.  This prevents future
0298      * accesses to the kioctx from page migration.
0299      */
0300     put_aio_ring_file(ctx);
0301 
0302     for (i = 0; i < ctx->nr_pages; i++) {
0303         struct page *page;
0304         pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
0305                 page_count(ctx->ring_pages[i]));
0306         page = ctx->ring_pages[i];
0307         if (!page)
0308             continue;
0309         ctx->ring_pages[i] = NULL;
0310         put_page(page);
0311     }
0312 
0313     if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
0314         kfree(ctx->ring_pages);
0315         ctx->ring_pages = NULL;
0316     }
0317 }
0318 
0319 static int aio_ring_mremap(struct vm_area_struct *vma)
0320 {
0321     struct file *file = vma->vm_file;
0322     struct mm_struct *mm = vma->vm_mm;
0323     struct kioctx_table *table;
0324     int i, res = -EINVAL;
0325 
0326     spin_lock(&mm->ioctx_lock);
0327     rcu_read_lock();
0328     table = rcu_dereference(mm->ioctx_table);
0329     for (i = 0; i < table->nr; i++) {
0330         struct kioctx *ctx;
0331 
0332         ctx = table->table[i];
0333         if (ctx && ctx->aio_ring_file == file) {
0334             if (!atomic_read(&ctx->dead)) {
0335                 ctx->user_id = ctx->mmap_base = vma->vm_start;
0336                 res = 0;
0337             }
0338             break;
0339         }
0340     }
0341 
0342     rcu_read_unlock();
0343     spin_unlock(&mm->ioctx_lock);
0344     return res;
0345 }
0346 
0347 static const struct vm_operations_struct aio_ring_vm_ops = {
0348     .mremap     = aio_ring_mremap,
0349 #if IS_ENABLED(CONFIG_MMU)
0350     .fault      = filemap_fault,
0351     .map_pages  = filemap_map_pages,
0352     .page_mkwrite   = filemap_page_mkwrite,
0353 #endif
0354 };
0355 
0356 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
0357 {
0358     vma->vm_flags |= VM_DONTEXPAND;
0359     vma->vm_ops = &aio_ring_vm_ops;
0360     return 0;
0361 }
0362 
0363 static const struct file_operations aio_ring_fops = {
0364     .mmap = aio_ring_mmap,
0365 };
0366 
0367 #if IS_ENABLED(CONFIG_MIGRATION)
0368 static int aio_migratepage(struct address_space *mapping, struct page *new,
0369             struct page *old, enum migrate_mode mode)
0370 {
0371     struct kioctx *ctx;
0372     unsigned long flags;
0373     pgoff_t idx;
0374     int rc;
0375 
0376     rc = 0;
0377 
0378     /* mapping->private_lock here protects against the kioctx teardown.  */
0379     spin_lock(&mapping->private_lock);
0380     ctx = mapping->private_data;
0381     if (!ctx) {
0382         rc = -EINVAL;
0383         goto out;
0384     }
0385 
0386     /* The ring_lock mutex.  The prevents aio_read_events() from writing
0387      * to the ring's head, and prevents page migration from mucking in
0388      * a partially initialized kiotx.
0389      */
0390     if (!mutex_trylock(&ctx->ring_lock)) {
0391         rc = -EAGAIN;
0392         goto out;
0393     }
0394 
0395     idx = old->index;
0396     if (idx < (pgoff_t)ctx->nr_pages) {
0397         /* Make sure the old page hasn't already been changed */
0398         if (ctx->ring_pages[idx] != old)
0399             rc = -EAGAIN;
0400     } else
0401         rc = -EINVAL;
0402 
0403     if (rc != 0)
0404         goto out_unlock;
0405 
0406     /* Writeback must be complete */
0407     BUG_ON(PageWriteback(old));
0408     get_page(new);
0409 
0410     rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
0411     if (rc != MIGRATEPAGE_SUCCESS) {
0412         put_page(new);
0413         goto out_unlock;
0414     }
0415 
0416     /* Take completion_lock to prevent other writes to the ring buffer
0417      * while the old page is copied to the new.  This prevents new
0418      * events from being lost.
0419      */
0420     spin_lock_irqsave(&ctx->completion_lock, flags);
0421     migrate_page_copy(new, old);
0422     BUG_ON(ctx->ring_pages[idx] != old);
0423     ctx->ring_pages[idx] = new;
0424     spin_unlock_irqrestore(&ctx->completion_lock, flags);
0425 
0426     /* The old page is no longer accessible. */
0427     put_page(old);
0428 
0429 out_unlock:
0430     mutex_unlock(&ctx->ring_lock);
0431 out:
0432     spin_unlock(&mapping->private_lock);
0433     return rc;
0434 }
0435 #endif
0436 
0437 static const struct address_space_operations aio_ctx_aops = {
0438     .set_page_dirty = __set_page_dirty_no_writeback,
0439 #if IS_ENABLED(CONFIG_MIGRATION)
0440     .migratepage    = aio_migratepage,
0441 #endif
0442 };
0443 
0444 static int aio_setup_ring(struct kioctx *ctx)
0445 {
0446     struct aio_ring *ring;
0447     unsigned nr_events = ctx->max_reqs;
0448     struct mm_struct *mm = current->mm;
0449     unsigned long size, unused;
0450     int nr_pages;
0451     int i;
0452     struct file *file;
0453 
0454     /* Compensate for the ring buffer's head/tail overlap entry */
0455     nr_events += 2; /* 1 is required, 2 for good luck */
0456 
0457     size = sizeof(struct aio_ring);
0458     size += sizeof(struct io_event) * nr_events;
0459 
0460     nr_pages = PFN_UP(size);
0461     if (nr_pages < 0)
0462         return -EINVAL;
0463 
0464     file = aio_private_file(ctx, nr_pages);
0465     if (IS_ERR(file)) {
0466         ctx->aio_ring_file = NULL;
0467         return -ENOMEM;
0468     }
0469 
0470     ctx->aio_ring_file = file;
0471     nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
0472             / sizeof(struct io_event);
0473 
0474     ctx->ring_pages = ctx->internal_pages;
0475     if (nr_pages > AIO_RING_PAGES) {
0476         ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
0477                       GFP_KERNEL);
0478         if (!ctx->ring_pages) {
0479             put_aio_ring_file(ctx);
0480             return -ENOMEM;
0481         }
0482     }
0483 
0484     for (i = 0; i < nr_pages; i++) {
0485         struct page *page;
0486         page = find_or_create_page(file->f_mapping,
0487                        i, GFP_HIGHUSER | __GFP_ZERO);
0488         if (!page)
0489             break;
0490         pr_debug("pid(%d) page[%d]->count=%d\n",
0491              current->pid, i, page_count(page));
0492         SetPageUptodate(page);
0493         unlock_page(page);
0494 
0495         ctx->ring_pages[i] = page;
0496     }
0497     ctx->nr_pages = i;
0498 
0499     if (unlikely(i != nr_pages)) {
0500         aio_free_ring(ctx);
0501         return -ENOMEM;
0502     }
0503 
0504     ctx->mmap_size = nr_pages * PAGE_SIZE;
0505     pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
0506 
0507     if (down_write_killable(&mm->mmap_sem)) {
0508         ctx->mmap_size = 0;
0509         aio_free_ring(ctx);
0510         return -EINTR;
0511     }
0512 
0513     ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
0514                        PROT_READ | PROT_WRITE,
0515                        MAP_SHARED, 0, &unused);
0516     up_write(&mm->mmap_sem);
0517     if (IS_ERR((void *)ctx->mmap_base)) {
0518         ctx->mmap_size = 0;
0519         aio_free_ring(ctx);
0520         return -ENOMEM;
0521     }
0522 
0523     pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
0524 
0525     ctx->user_id = ctx->mmap_base;
0526     ctx->nr_events = nr_events; /* trusted copy */
0527 
0528     ring = kmap_atomic(ctx->ring_pages[0]);
0529     ring->nr = nr_events;   /* user copy */
0530     ring->id = ~0U;
0531     ring->head = ring->tail = 0;
0532     ring->magic = AIO_RING_MAGIC;
0533     ring->compat_features = AIO_RING_COMPAT_FEATURES;
0534     ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
0535     ring->header_length = sizeof(struct aio_ring);
0536     kunmap_atomic(ring);
0537     flush_dcache_page(ctx->ring_pages[0]);
0538 
0539     return 0;
0540 }
0541 
0542 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
0543 #define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
0544 #define AIO_EVENTS_OFFSET   (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
0545 
0546 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
0547 {
0548     struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
0549     struct kioctx *ctx = req->ki_ctx;
0550     unsigned long flags;
0551 
0552     spin_lock_irqsave(&ctx->ctx_lock, flags);
0553 
0554     if (!req->ki_list.next)
0555         list_add(&req->ki_list, &ctx->active_reqs);
0556 
0557     req->ki_cancel = cancel;
0558 
0559     spin_unlock_irqrestore(&ctx->ctx_lock, flags);
0560 }
0561 EXPORT_SYMBOL(kiocb_set_cancel_fn);
0562 
0563 static int kiocb_cancel(struct aio_kiocb *kiocb)
0564 {
0565     kiocb_cancel_fn *old, *cancel;
0566 
0567     /*
0568      * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
0569      * actually has a cancel function, hence the cmpxchg()
0570      */
0571 
0572     cancel = ACCESS_ONCE(kiocb->ki_cancel);
0573     do {
0574         if (!cancel || cancel == KIOCB_CANCELLED)
0575             return -EINVAL;
0576 
0577         old = cancel;
0578         cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
0579     } while (cancel != old);
0580 
0581     return cancel(&kiocb->common);
0582 }
0583 
0584 static void free_ioctx(struct work_struct *work)
0585 {
0586     struct kioctx *ctx = container_of(work, struct kioctx, free_work);
0587 
0588     pr_debug("freeing %p\n", ctx);
0589 
0590     aio_free_ring(ctx);
0591     free_percpu(ctx->cpu);
0592     percpu_ref_exit(&ctx->reqs);
0593     percpu_ref_exit(&ctx->users);
0594     kmem_cache_free(kioctx_cachep, ctx);
0595 }
0596 
0597 static void free_ioctx_reqs(struct percpu_ref *ref)
0598 {
0599     struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
0600 
0601     /* At this point we know that there are no any in-flight requests */
0602     if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
0603         complete(&ctx->rq_wait->comp);
0604 
0605     INIT_WORK(&ctx->free_work, free_ioctx);
0606     schedule_work(&ctx->free_work);
0607 }
0608 
0609 /*
0610  * When this function runs, the kioctx has been removed from the "hash table"
0611  * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
0612  * now it's safe to cancel any that need to be.
0613  */
0614 static void free_ioctx_users(struct percpu_ref *ref)
0615 {
0616     struct kioctx *ctx = container_of(ref, struct kioctx, users);
0617     struct aio_kiocb *req;
0618 
0619     spin_lock_irq(&ctx->ctx_lock);
0620 
0621     while (!list_empty(&ctx->active_reqs)) {
0622         req = list_first_entry(&ctx->active_reqs,
0623                        struct aio_kiocb, ki_list);
0624 
0625         list_del_init(&req->ki_list);
0626         kiocb_cancel(req);
0627     }
0628 
0629     spin_unlock_irq(&ctx->ctx_lock);
0630 
0631     percpu_ref_kill(&ctx->reqs);
0632     percpu_ref_put(&ctx->reqs);
0633 }
0634 
0635 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
0636 {
0637     unsigned i, new_nr;
0638     struct kioctx_table *table, *old;
0639     struct aio_ring *ring;
0640 
0641     spin_lock(&mm->ioctx_lock);
0642     table = rcu_dereference_raw(mm->ioctx_table);
0643 
0644     while (1) {
0645         if (table)
0646             for (i = 0; i < table->nr; i++)
0647                 if (!table->table[i]) {
0648                     ctx->id = i;
0649                     table->table[i] = ctx;
0650                     spin_unlock(&mm->ioctx_lock);
0651 
0652                     /* While kioctx setup is in progress,
0653                      * we are protected from page migration
0654                      * changes ring_pages by ->ring_lock.
0655                      */
0656                     ring = kmap_atomic(ctx->ring_pages[0]);
0657                     ring->id = ctx->id;
0658                     kunmap_atomic(ring);
0659                     return 0;
0660                 }
0661 
0662         new_nr = (table ? table->nr : 1) * 4;
0663         spin_unlock(&mm->ioctx_lock);
0664 
0665         table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
0666                 new_nr, GFP_KERNEL);
0667         if (!table)
0668             return -ENOMEM;
0669 
0670         table->nr = new_nr;
0671 
0672         spin_lock(&mm->ioctx_lock);
0673         old = rcu_dereference_raw(mm->ioctx_table);
0674 
0675         if (!old) {
0676             rcu_assign_pointer(mm->ioctx_table, table);
0677         } else if (table->nr > old->nr) {
0678             memcpy(table->table, old->table,
0679                    old->nr * sizeof(struct kioctx *));
0680 
0681             rcu_assign_pointer(mm->ioctx_table, table);
0682             kfree_rcu(old, rcu);
0683         } else {
0684             kfree(table);
0685             table = old;
0686         }
0687     }
0688 }
0689 
0690 static void aio_nr_sub(unsigned nr)
0691 {
0692     spin_lock(&aio_nr_lock);
0693     if (WARN_ON(aio_nr - nr > aio_nr))
0694         aio_nr = 0;
0695     else
0696         aio_nr -= nr;
0697     spin_unlock(&aio_nr_lock);
0698 }
0699 
0700 /* ioctx_alloc
0701  *  Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
0702  */
0703 static struct kioctx *ioctx_alloc(unsigned nr_events)
0704 {
0705     struct mm_struct *mm = current->mm;
0706     struct kioctx *ctx;
0707     int err = -ENOMEM;
0708 
0709     /*
0710      * We keep track of the number of available ringbuffer slots, to prevent
0711      * overflow (reqs_available), and we also use percpu counters for this.
0712      *
0713      * So since up to half the slots might be on other cpu's percpu counters
0714      * and unavailable, double nr_events so userspace sees what they
0715      * expected: additionally, we move req_batch slots to/from percpu
0716      * counters at a time, so make sure that isn't 0:
0717      */
0718     nr_events = max(nr_events, num_possible_cpus() * 4);
0719     nr_events *= 2;
0720 
0721     /* Prevent overflows */
0722     if (nr_events > (0x10000000U / sizeof(struct io_event))) {
0723         pr_debug("ENOMEM: nr_events too high\n");
0724         return ERR_PTR(-EINVAL);
0725     }
0726 
0727     if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
0728         return ERR_PTR(-EAGAIN);
0729 
0730     ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
0731     if (!ctx)
0732         return ERR_PTR(-ENOMEM);
0733 
0734     ctx->max_reqs = nr_events;
0735 
0736     spin_lock_init(&ctx->ctx_lock);
0737     spin_lock_init(&ctx->completion_lock);
0738     mutex_init(&ctx->ring_lock);
0739     /* Protect against page migration throughout kiotx setup by keeping
0740      * the ring_lock mutex held until setup is complete. */
0741     mutex_lock(&ctx->ring_lock);
0742     init_waitqueue_head(&ctx->wait);
0743 
0744     INIT_LIST_HEAD(&ctx->active_reqs);
0745 
0746     if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
0747         goto err;
0748 
0749     if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
0750         goto err;
0751 
0752     ctx->cpu = alloc_percpu(struct kioctx_cpu);
0753     if (!ctx->cpu)
0754         goto err;
0755 
0756     err = aio_setup_ring(ctx);
0757     if (err < 0)
0758         goto err;
0759 
0760     atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
0761     ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
0762     if (ctx->req_batch < 1)
0763         ctx->req_batch = 1;
0764 
0765     /* limit the number of system wide aios */
0766     spin_lock(&aio_nr_lock);
0767     if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
0768         aio_nr + nr_events < aio_nr) {
0769         spin_unlock(&aio_nr_lock);
0770         err = -EAGAIN;
0771         goto err_ctx;
0772     }
0773     aio_nr += ctx->max_reqs;
0774     spin_unlock(&aio_nr_lock);
0775 
0776     percpu_ref_get(&ctx->users);    /* io_setup() will drop this ref */
0777     percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
0778 
0779     err = ioctx_add_table(ctx, mm);
0780     if (err)
0781         goto err_cleanup;
0782 
0783     /* Release the ring_lock mutex now that all setup is complete. */
0784     mutex_unlock(&ctx->ring_lock);
0785 
0786     pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
0787          ctx, ctx->user_id, mm, ctx->nr_events);
0788     return ctx;
0789 
0790 err_cleanup:
0791     aio_nr_sub(ctx->max_reqs);
0792 err_ctx:
0793     atomic_set(&ctx->dead, 1);
0794     if (ctx->mmap_size)
0795         vm_munmap(ctx->mmap_base, ctx->mmap_size);
0796     aio_free_ring(ctx);
0797 err:
0798     mutex_unlock(&ctx->ring_lock);
0799     free_percpu(ctx->cpu);
0800     percpu_ref_exit(&ctx->reqs);
0801     percpu_ref_exit(&ctx->users);
0802     kmem_cache_free(kioctx_cachep, ctx);
0803     pr_debug("error allocating ioctx %d\n", err);
0804     return ERR_PTR(err);
0805 }
0806 
0807 /* kill_ioctx
0808  *  Cancels all outstanding aio requests on an aio context.  Used
0809  *  when the processes owning a context have all exited to encourage
0810  *  the rapid destruction of the kioctx.
0811  */
0812 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
0813               struct ctx_rq_wait *wait)
0814 {
0815     struct kioctx_table *table;
0816 
0817     spin_lock(&mm->ioctx_lock);
0818     if (atomic_xchg(&ctx->dead, 1)) {
0819         spin_unlock(&mm->ioctx_lock);
0820         return -EINVAL;
0821     }
0822 
0823     table = rcu_dereference_raw(mm->ioctx_table);
0824     WARN_ON(ctx != table->table[ctx->id]);
0825     table->table[ctx->id] = NULL;
0826     spin_unlock(&mm->ioctx_lock);
0827 
0828     /* percpu_ref_kill() will do the necessary call_rcu() */
0829     wake_up_all(&ctx->wait);
0830 
0831     /*
0832      * It'd be more correct to do this in free_ioctx(), after all
0833      * the outstanding kiocbs have finished - but by then io_destroy
0834      * has already returned, so io_setup() could potentially return
0835      * -EAGAIN with no ioctxs actually in use (as far as userspace
0836      *  could tell).
0837      */
0838     aio_nr_sub(ctx->max_reqs);
0839 
0840     if (ctx->mmap_size)
0841         vm_munmap(ctx->mmap_base, ctx->mmap_size);
0842 
0843     ctx->rq_wait = wait;
0844     percpu_ref_kill(&ctx->users);
0845     return 0;
0846 }
0847 
0848 /*
0849  * exit_aio: called when the last user of mm goes away.  At this point, there is
0850  * no way for any new requests to be submited or any of the io_* syscalls to be
0851  * called on the context.
0852  *
0853  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
0854  * them.
0855  */
0856 void exit_aio(struct mm_struct *mm)
0857 {
0858     struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
0859     struct ctx_rq_wait wait;
0860     int i, skipped;
0861 
0862     if (!table)
0863         return;
0864 
0865     atomic_set(&wait.count, table->nr);
0866     init_completion(&wait.comp);
0867 
0868     skipped = 0;
0869     for (i = 0; i < table->nr; ++i) {
0870         struct kioctx *ctx = table->table[i];
0871 
0872         if (!ctx) {
0873             skipped++;
0874             continue;
0875         }
0876 
0877         /*
0878          * We don't need to bother with munmap() here - exit_mmap(mm)
0879          * is coming and it'll unmap everything. And we simply can't,
0880          * this is not necessarily our ->mm.
0881          * Since kill_ioctx() uses non-zero ->mmap_size as indicator
0882          * that it needs to unmap the area, just set it to 0.
0883          */
0884         ctx->mmap_size = 0;
0885         kill_ioctx(mm, ctx, &wait);
0886     }
0887 
0888     if (!atomic_sub_and_test(skipped, &wait.count)) {
0889         /* Wait until all IO for the context are done. */
0890         wait_for_completion(&wait.comp);
0891     }
0892 
0893     RCU_INIT_POINTER(mm->ioctx_table, NULL);
0894     kfree(table);
0895 }
0896 
0897 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
0898 {
0899     struct kioctx_cpu *kcpu;
0900     unsigned long flags;
0901 
0902     local_irq_save(flags);
0903     kcpu = this_cpu_ptr(ctx->cpu);
0904     kcpu->reqs_available += nr;
0905 
0906     while (kcpu->reqs_available >= ctx->req_batch * 2) {
0907         kcpu->reqs_available -= ctx->req_batch;
0908         atomic_add(ctx->req_batch, &ctx->reqs_available);
0909     }
0910 
0911     local_irq_restore(flags);
0912 }
0913 
0914 static bool get_reqs_available(struct kioctx *ctx)
0915 {
0916     struct kioctx_cpu *kcpu;
0917     bool ret = false;
0918     unsigned long flags;
0919 
0920     local_irq_save(flags);
0921     kcpu = this_cpu_ptr(ctx->cpu);
0922     if (!kcpu->reqs_available) {
0923         int old, avail = atomic_read(&ctx->reqs_available);
0924 
0925         do {
0926             if (avail < ctx->req_batch)
0927                 goto out;
0928 
0929             old = avail;
0930             avail = atomic_cmpxchg(&ctx->reqs_available,
0931                            avail, avail - ctx->req_batch);
0932         } while (avail != old);
0933 
0934         kcpu->reqs_available += ctx->req_batch;
0935     }
0936 
0937     ret = true;
0938     kcpu->reqs_available--;
0939 out:
0940     local_irq_restore(flags);
0941     return ret;
0942 }
0943 
0944 /* refill_reqs_available
0945  *  Updates the reqs_available reference counts used for tracking the
0946  *  number of free slots in the completion ring.  This can be called
0947  *  from aio_complete() (to optimistically update reqs_available) or
0948  *  from aio_get_req() (the we're out of events case).  It must be
0949  *  called holding ctx->completion_lock.
0950  */
0951 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
0952                                   unsigned tail)
0953 {
0954     unsigned events_in_ring, completed;
0955 
0956     /* Clamp head since userland can write to it. */
0957     head %= ctx->nr_events;
0958     if (head <= tail)
0959         events_in_ring = tail - head;
0960     else
0961         events_in_ring = ctx->nr_events - (head - tail);
0962 
0963     completed = ctx->completed_events;
0964     if (events_in_ring < completed)
0965         completed -= events_in_ring;
0966     else
0967         completed = 0;
0968 
0969     if (!completed)
0970         return;
0971 
0972     ctx->completed_events -= completed;
0973     put_reqs_available(ctx, completed);
0974 }
0975 
0976 /* user_refill_reqs_available
0977  *  Called to refill reqs_available when aio_get_req() encounters an
0978  *  out of space in the completion ring.
0979  */
0980 static void user_refill_reqs_available(struct kioctx *ctx)
0981 {
0982     spin_lock_irq(&ctx->completion_lock);
0983     if (ctx->completed_events) {
0984         struct aio_ring *ring;
0985         unsigned head;
0986 
0987         /* Access of ring->head may race with aio_read_events_ring()
0988          * here, but that's okay since whether we read the old version
0989          * or the new version, and either will be valid.  The important
0990          * part is that head cannot pass tail since we prevent
0991          * aio_complete() from updating tail by holding
0992          * ctx->completion_lock.  Even if head is invalid, the check
0993          * against ctx->completed_events below will make sure we do the
0994          * safe/right thing.
0995          */
0996         ring = kmap_atomic(ctx->ring_pages[0]);
0997         head = ring->head;
0998         kunmap_atomic(ring);
0999 
1000         refill_reqs_available(ctx, head, ctx->tail);
1001     }
1002 
1003     spin_unlock_irq(&ctx->completion_lock);
1004 }
1005 
1006 /* aio_get_req
1007  *  Allocate a slot for an aio request.
1008  * Returns NULL if no requests are free.
1009  */
1010 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1011 {
1012     struct aio_kiocb *req;
1013 
1014     if (!get_reqs_available(ctx)) {
1015         user_refill_reqs_available(ctx);
1016         if (!get_reqs_available(ctx))
1017             return NULL;
1018     }
1019 
1020     req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1021     if (unlikely(!req))
1022         goto out_put;
1023 
1024     percpu_ref_get(&ctx->reqs);
1025 
1026     req->ki_ctx = ctx;
1027     return req;
1028 out_put:
1029     put_reqs_available(ctx, 1);
1030     return NULL;
1031 }
1032 
1033 static void kiocb_free(struct aio_kiocb *req)
1034 {
1035     if (req->common.ki_filp)
1036         fput(req->common.ki_filp);
1037     if (req->ki_eventfd != NULL)
1038         eventfd_ctx_put(req->ki_eventfd);
1039     kmem_cache_free(kiocb_cachep, req);
1040 }
1041 
1042 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1043 {
1044     struct aio_ring __user *ring  = (void __user *)ctx_id;
1045     struct mm_struct *mm = current->mm;
1046     struct kioctx *ctx, *ret = NULL;
1047     struct kioctx_table *table;
1048     unsigned id;
1049 
1050     if (get_user(id, &ring->id))
1051         return NULL;
1052 
1053     rcu_read_lock();
1054     table = rcu_dereference(mm->ioctx_table);
1055 
1056     if (!table || id >= table->nr)
1057         goto out;
1058 
1059     ctx = table->table[id];
1060     if (ctx && ctx->user_id == ctx_id) {
1061         percpu_ref_get(&ctx->users);
1062         ret = ctx;
1063     }
1064 out:
1065     rcu_read_unlock();
1066     return ret;
1067 }
1068 
1069 /* aio_complete
1070  *  Called when the io request on the given iocb is complete.
1071  */
1072 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1073 {
1074     struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1075     struct kioctx   *ctx = iocb->ki_ctx;
1076     struct aio_ring *ring;
1077     struct io_event *ev_page, *event;
1078     unsigned tail, pos, head;
1079     unsigned long   flags;
1080 
1081     if (kiocb->ki_flags & IOCB_WRITE) {
1082         struct file *file = kiocb->ki_filp;
1083 
1084         /*
1085          * Tell lockdep we inherited freeze protection from submission
1086          * thread.
1087          */
1088         if (S_ISREG(file_inode(file)->i_mode))
1089             __sb_writers_acquired(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1090         file_end_write(file);
1091     }
1092 
1093     /*
1094      * Special case handling for sync iocbs:
1095      *  - events go directly into the iocb for fast handling
1096      *  - the sync task with the iocb in its stack holds the single iocb
1097      *    ref, no other paths have a way to get another ref
1098      *  - the sync task helpfully left a reference to itself in the iocb
1099      */
1100     BUG_ON(is_sync_kiocb(kiocb));
1101 
1102     if (iocb->ki_list.next) {
1103         unsigned long flags;
1104 
1105         spin_lock_irqsave(&ctx->ctx_lock, flags);
1106         list_del(&iocb->ki_list);
1107         spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1108     }
1109 
1110     /*
1111      * Add a completion event to the ring buffer. Must be done holding
1112      * ctx->completion_lock to prevent other code from messing with the tail
1113      * pointer since we might be called from irq context.
1114      */
1115     spin_lock_irqsave(&ctx->completion_lock, flags);
1116 
1117     tail = ctx->tail;
1118     pos = tail + AIO_EVENTS_OFFSET;
1119 
1120     if (++tail >= ctx->nr_events)
1121         tail = 0;
1122 
1123     ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1124     event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1125 
1126     event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1127     event->data = iocb->ki_user_data;
1128     event->res = res;
1129     event->res2 = res2;
1130 
1131     kunmap_atomic(ev_page);
1132     flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1133 
1134     pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1135          ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1136          res, res2);
1137 
1138     /* after flagging the request as done, we
1139      * must never even look at it again
1140      */
1141     smp_wmb();  /* make event visible before updating tail */
1142 
1143     ctx->tail = tail;
1144 
1145     ring = kmap_atomic(ctx->ring_pages[0]);
1146     head = ring->head;
1147     ring->tail = tail;
1148     kunmap_atomic(ring);
1149     flush_dcache_page(ctx->ring_pages[0]);
1150 
1151     ctx->completed_events++;
1152     if (ctx->completed_events > 1)
1153         refill_reqs_available(ctx, head, tail);
1154     spin_unlock_irqrestore(&ctx->completion_lock, flags);
1155 
1156     pr_debug("added to ring %p at [%u]\n", iocb, tail);
1157 
1158     /*
1159      * Check if the user asked us to deliver the result through an
1160      * eventfd. The eventfd_signal() function is safe to be called
1161      * from IRQ context.
1162      */
1163     if (iocb->ki_eventfd != NULL)
1164         eventfd_signal(iocb->ki_eventfd, 1);
1165 
1166     /* everything turned out well, dispose of the aiocb. */
1167     kiocb_free(iocb);
1168 
1169     /*
1170      * We have to order our ring_info tail store above and test
1171      * of the wait list below outside the wait lock.  This is
1172      * like in wake_up_bit() where clearing a bit has to be
1173      * ordered with the unlocked test.
1174      */
1175     smp_mb();
1176 
1177     if (waitqueue_active(&ctx->wait))
1178         wake_up(&ctx->wait);
1179 
1180     percpu_ref_put(&ctx->reqs);
1181 }
1182 
1183 /* aio_read_events_ring
1184  *  Pull an event off of the ioctx's event ring.  Returns the number of
1185  *  events fetched
1186  */
1187 static long aio_read_events_ring(struct kioctx *ctx,
1188                  struct io_event __user *event, long nr)
1189 {
1190     struct aio_ring *ring;
1191     unsigned head, tail, pos;
1192     long ret = 0;
1193     int copy_ret;
1194 
1195     /*
1196      * The mutex can block and wake us up and that will cause
1197      * wait_event_interruptible_hrtimeout() to schedule without sleeping
1198      * and repeat. This should be rare enough that it doesn't cause
1199      * peformance issues. See the comment in read_events() for more detail.
1200      */
1201     sched_annotate_sleep();
1202     mutex_lock(&ctx->ring_lock);
1203 
1204     /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1205     ring = kmap_atomic(ctx->ring_pages[0]);
1206     head = ring->head;
1207     tail = ring->tail;
1208     kunmap_atomic(ring);
1209 
1210     /*
1211      * Ensure that once we've read the current tail pointer, that
1212      * we also see the events that were stored up to the tail.
1213      */
1214     smp_rmb();
1215 
1216     pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1217 
1218     if (head == tail)
1219         goto out;
1220 
1221     head %= ctx->nr_events;
1222     tail %= ctx->nr_events;
1223 
1224     while (ret < nr) {
1225         long avail;
1226         struct io_event *ev;
1227         struct page *page;
1228 
1229         avail = (head <= tail ?  tail : ctx->nr_events) - head;
1230         if (head == tail)
1231             break;
1232 
1233         avail = min(avail, nr - ret);
1234         avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1235                 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1236 
1237         pos = head + AIO_EVENTS_OFFSET;
1238         page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1239         pos %= AIO_EVENTS_PER_PAGE;
1240 
1241         ev = kmap(page);
1242         copy_ret = copy_to_user(event + ret, ev + pos,
1243                     sizeof(*ev) * avail);
1244         kunmap(page);
1245 
1246         if (unlikely(copy_ret)) {
1247             ret = -EFAULT;
1248             goto out;
1249         }
1250 
1251         ret += avail;
1252         head += avail;
1253         head %= ctx->nr_events;
1254     }
1255 
1256     ring = kmap_atomic(ctx->ring_pages[0]);
1257     ring->head = head;
1258     kunmap_atomic(ring);
1259     flush_dcache_page(ctx->ring_pages[0]);
1260 
1261     pr_debug("%li  h%u t%u\n", ret, head, tail);
1262 out:
1263     mutex_unlock(&ctx->ring_lock);
1264 
1265     return ret;
1266 }
1267 
1268 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1269                 struct io_event __user *event, long *i)
1270 {
1271     long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1272 
1273     if (ret > 0)
1274         *i += ret;
1275 
1276     if (unlikely(atomic_read(&ctx->dead)))
1277         ret = -EINVAL;
1278 
1279     if (!*i)
1280         *i = ret;
1281 
1282     return ret < 0 || *i >= min_nr;
1283 }
1284 
1285 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1286             struct io_event __user *event,
1287             struct timespec __user *timeout)
1288 {
1289     ktime_t until = KTIME_MAX;
1290     long ret = 0;
1291 
1292     if (timeout) {
1293         struct timespec ts;
1294 
1295         if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1296             return -EFAULT;
1297 
1298         until = timespec_to_ktime(ts);
1299     }
1300 
1301     /*
1302      * Note that aio_read_events() is being called as the conditional - i.e.
1303      * we're calling it after prepare_to_wait() has set task state to
1304      * TASK_INTERRUPTIBLE.
1305      *
1306      * But aio_read_events() can block, and if it blocks it's going to flip
1307      * the task state back to TASK_RUNNING.
1308      *
1309      * This should be ok, provided it doesn't flip the state back to
1310      * TASK_RUNNING and return 0 too much - that causes us to spin. That
1311      * will only happen if the mutex_lock() call blocks, and we then find
1312      * the ringbuffer empty. So in practice we should be ok, but it's
1313      * something to be aware of when touching this code.
1314      */
1315     if (until == 0)
1316         aio_read_events(ctx, min_nr, nr, event, &ret);
1317     else
1318         wait_event_interruptible_hrtimeout(ctx->wait,
1319                 aio_read_events(ctx, min_nr, nr, event, &ret),
1320                 until);
1321 
1322     if (!ret && signal_pending(current))
1323         ret = -EINTR;
1324 
1325     return ret;
1326 }
1327 
1328 /* sys_io_setup:
1329  *  Create an aio_context capable of receiving at least nr_events.
1330  *  ctxp must not point to an aio_context that already exists, and
1331  *  must be initialized to 0 prior to the call.  On successful
1332  *  creation of the aio_context, *ctxp is filled in with the resulting 
1333  *  handle.  May fail with -EINVAL if *ctxp is not initialized,
1334  *  if the specified nr_events exceeds internal limits.  May fail 
1335  *  with -EAGAIN if the specified nr_events exceeds the user's limit 
1336  *  of available events.  May fail with -ENOMEM if insufficient kernel
1337  *  resources are available.  May fail with -EFAULT if an invalid
1338  *  pointer is passed for ctxp.  Will fail with -ENOSYS if not
1339  *  implemented.
1340  */
1341 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1342 {
1343     struct kioctx *ioctx = NULL;
1344     unsigned long ctx;
1345     long ret;
1346 
1347     ret = get_user(ctx, ctxp);
1348     if (unlikely(ret))
1349         goto out;
1350 
1351     ret = -EINVAL;
1352     if (unlikely(ctx || nr_events == 0)) {
1353         pr_debug("EINVAL: ctx %lu nr_events %u\n",
1354                  ctx, nr_events);
1355         goto out;
1356     }
1357 
1358     ioctx = ioctx_alloc(nr_events);
1359     ret = PTR_ERR(ioctx);
1360     if (!IS_ERR(ioctx)) {
1361         ret = put_user(ioctx->user_id, ctxp);
1362         if (ret)
1363             kill_ioctx(current->mm, ioctx, NULL);
1364         percpu_ref_put(&ioctx->users);
1365     }
1366 
1367 out:
1368     return ret;
1369 }
1370 
1371 #ifdef CONFIG_COMPAT
1372 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1373 {
1374     struct kioctx *ioctx = NULL;
1375     unsigned long ctx;
1376     long ret;
1377 
1378     ret = get_user(ctx, ctx32p);
1379     if (unlikely(ret))
1380         goto out;
1381 
1382     ret = -EINVAL;
1383     if (unlikely(ctx || nr_events == 0)) {
1384         pr_debug("EINVAL: ctx %lu nr_events %u\n",
1385                  ctx, nr_events);
1386         goto out;
1387     }
1388 
1389     ioctx = ioctx_alloc(nr_events);
1390     ret = PTR_ERR(ioctx);
1391     if (!IS_ERR(ioctx)) {
1392         /* truncating is ok because it's a user address */
1393         ret = put_user((u32)ioctx->user_id, ctx32p);
1394         if (ret)
1395             kill_ioctx(current->mm, ioctx, NULL);
1396         percpu_ref_put(&ioctx->users);
1397     }
1398 
1399 out:
1400     return ret;
1401 }
1402 #endif
1403 
1404 /* sys_io_destroy:
1405  *  Destroy the aio_context specified.  May cancel any outstanding 
1406  *  AIOs and block on completion.  Will fail with -ENOSYS if not
1407  *  implemented.  May fail with -EINVAL if the context pointed to
1408  *  is invalid.
1409  */
1410 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1411 {
1412     struct kioctx *ioctx = lookup_ioctx(ctx);
1413     if (likely(NULL != ioctx)) {
1414         struct ctx_rq_wait wait;
1415         int ret;
1416 
1417         init_completion(&wait.comp);
1418         atomic_set(&wait.count, 1);
1419 
1420         /* Pass requests_done to kill_ioctx() where it can be set
1421          * in a thread-safe way. If we try to set it here then we have
1422          * a race condition if two io_destroy() called simultaneously.
1423          */
1424         ret = kill_ioctx(current->mm, ioctx, &wait);
1425         percpu_ref_put(&ioctx->users);
1426 
1427         /* Wait until all IO for the context are done. Otherwise kernel
1428          * keep using user-space buffers even if user thinks the context
1429          * is destroyed.
1430          */
1431         if (!ret)
1432             wait_for_completion(&wait.comp);
1433 
1434         return ret;
1435     }
1436     pr_debug("EINVAL: invalid context id\n");
1437     return -EINVAL;
1438 }
1439 
1440 static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec,
1441         bool vectored, bool compat, struct iov_iter *iter)
1442 {
1443     void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1444     size_t len = iocb->aio_nbytes;
1445 
1446     if (!vectored) {
1447         ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1448         *iovec = NULL;
1449         return ret;
1450     }
1451 #ifdef CONFIG_COMPAT
1452     if (compat)
1453         return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1454                 iter);
1455 #endif
1456     return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1457 }
1458 
1459 static inline ssize_t aio_ret(struct kiocb *req, ssize_t ret)
1460 {
1461     switch (ret) {
1462     case -EIOCBQUEUED:
1463         return ret;
1464     case -ERESTARTSYS:
1465     case -ERESTARTNOINTR:
1466     case -ERESTARTNOHAND:
1467     case -ERESTART_RESTARTBLOCK:
1468         /*
1469          * There's no easy way to restart the syscall since other AIO's
1470          * may be already running. Just fail this IO with EINTR.
1471          */
1472         ret = -EINTR;
1473         /*FALLTHRU*/
1474     default:
1475         aio_complete(req, ret, 0);
1476         return 0;
1477     }
1478 }
1479 
1480 static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
1481         bool compat)
1482 {
1483     struct file *file = req->ki_filp;
1484     struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1485     struct iov_iter iter;
1486     ssize_t ret;
1487 
1488     if (unlikely(!(file->f_mode & FMODE_READ)))
1489         return -EBADF;
1490     if (unlikely(!file->f_op->read_iter))
1491         return -EINVAL;
1492 
1493     ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1494     if (ret)
1495         return ret;
1496     ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1497     if (!ret)
1498         ret = aio_ret(req, file->f_op->read_iter(req, &iter));
1499     kfree(iovec);
1500     return ret;
1501 }
1502 
1503 static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored,
1504         bool compat)
1505 {
1506     struct file *file = req->ki_filp;
1507     struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1508     struct iov_iter iter;
1509     ssize_t ret;
1510 
1511     if (unlikely(!(file->f_mode & FMODE_WRITE)))
1512         return -EBADF;
1513     if (unlikely(!file->f_op->write_iter))
1514         return -EINVAL;
1515 
1516     ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1517     if (ret)
1518         return ret;
1519     ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1520     if (!ret) {
1521         req->ki_flags |= IOCB_WRITE;
1522         file_start_write(file);
1523         ret = aio_ret(req, file->f_op->write_iter(req, &iter));
1524         /*
1525          * We release freeze protection in aio_complete().  Fool lockdep
1526          * by telling it the lock got released so that it doesn't
1527          * complain about held lock when we return to userspace.
1528          */
1529         if (S_ISREG(file_inode(file)->i_mode))
1530             __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1531     }
1532     kfree(iovec);
1533     return ret;
1534 }
1535 
1536 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1537              struct iocb *iocb, bool compat)
1538 {
1539     struct aio_kiocb *req;
1540     struct file *file;
1541     ssize_t ret;
1542 
1543     /* enforce forwards compatibility on users */
1544     if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1545         pr_debug("EINVAL: reserve field set\n");
1546         return -EINVAL;
1547     }
1548 
1549     /* prevent overflows */
1550     if (unlikely(
1551         (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1552         (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1553         ((ssize_t)iocb->aio_nbytes < 0)
1554        )) {
1555         pr_debug("EINVAL: overflow check\n");
1556         return -EINVAL;
1557     }
1558 
1559     req = aio_get_req(ctx);
1560     if (unlikely(!req))
1561         return -EAGAIN;
1562 
1563     req->common.ki_filp = file = fget(iocb->aio_fildes);
1564     if (unlikely(!req->common.ki_filp)) {
1565         ret = -EBADF;
1566         goto out_put_req;
1567     }
1568     req->common.ki_pos = iocb->aio_offset;
1569     req->common.ki_complete = aio_complete;
1570     req->common.ki_flags = iocb_flags(req->common.ki_filp);
1571 
1572     if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1573         /*
1574          * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1575          * instance of the file* now. The file descriptor must be
1576          * an eventfd() fd, and will be signaled for each completed
1577          * event using the eventfd_signal() function.
1578          */
1579         req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1580         if (IS_ERR(req->ki_eventfd)) {
1581             ret = PTR_ERR(req->ki_eventfd);
1582             req->ki_eventfd = NULL;
1583             goto out_put_req;
1584         }
1585 
1586         req->common.ki_flags |= IOCB_EVENTFD;
1587     }
1588 
1589     ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1590     if (unlikely(ret)) {
1591         pr_debug("EFAULT: aio_key\n");
1592         goto out_put_req;
1593     }
1594 
1595     req->ki_user_iocb = user_iocb;
1596     req->ki_user_data = iocb->aio_data;
1597 
1598     get_file(file);
1599     switch (iocb->aio_lio_opcode) {
1600     case IOCB_CMD_PREAD:
1601         ret = aio_read(&req->common, iocb, false, compat);
1602         break;
1603     case IOCB_CMD_PWRITE:
1604         ret = aio_write(&req->common, iocb, false, compat);
1605         break;
1606     case IOCB_CMD_PREADV:
1607         ret = aio_read(&req->common, iocb, true, compat);
1608         break;
1609     case IOCB_CMD_PWRITEV:
1610         ret = aio_write(&req->common, iocb, true, compat);
1611         break;
1612     default:
1613         pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1614         ret = -EINVAL;
1615         break;
1616     }
1617     fput(file);
1618 
1619     if (ret && ret != -EIOCBQUEUED)
1620         goto out_put_req;
1621     return 0;
1622 out_put_req:
1623     put_reqs_available(ctx, 1);
1624     percpu_ref_put(&ctx->reqs);
1625     kiocb_free(req);
1626     return ret;
1627 }
1628 
1629 static long do_io_submit(aio_context_t ctx_id, long nr,
1630               struct iocb __user *__user *iocbpp, bool compat)
1631 {
1632     struct kioctx *ctx;
1633     long ret = 0;
1634     int i = 0;
1635     struct blk_plug plug;
1636 
1637     if (unlikely(nr < 0))
1638         return -EINVAL;
1639 
1640     if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1641         nr = LONG_MAX/sizeof(*iocbpp);
1642 
1643     if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1644         return -EFAULT;
1645 
1646     ctx = lookup_ioctx(ctx_id);
1647     if (unlikely(!ctx)) {
1648         pr_debug("EINVAL: invalid context id\n");
1649         return -EINVAL;
1650     }
1651 
1652     blk_start_plug(&plug);
1653 
1654     /*
1655      * AKPM: should this return a partial result if some of the IOs were
1656      * successfully submitted?
1657      */
1658     for (i=0; i<nr; i++) {
1659         struct iocb __user *user_iocb;
1660         struct iocb tmp;
1661 
1662         if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1663             ret = -EFAULT;
1664             break;
1665         }
1666 
1667         if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1668             ret = -EFAULT;
1669             break;
1670         }
1671 
1672         ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1673         if (ret)
1674             break;
1675     }
1676     blk_finish_plug(&plug);
1677 
1678     percpu_ref_put(&ctx->users);
1679     return i ? i : ret;
1680 }
1681 
1682 /* sys_io_submit:
1683  *  Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1684  *  the number of iocbs queued.  May return -EINVAL if the aio_context
1685  *  specified by ctx_id is invalid, if nr is < 0, if the iocb at
1686  *  *iocbpp[0] is not properly initialized, if the operation specified
1687  *  is invalid for the file descriptor in the iocb.  May fail with
1688  *  -EFAULT if any of the data structures point to invalid data.  May
1689  *  fail with -EBADF if the file descriptor specified in the first
1690  *  iocb is invalid.  May fail with -EAGAIN if insufficient resources
1691  *  are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1692  *  fail with -ENOSYS if not implemented.
1693  */
1694 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1695         struct iocb __user * __user *, iocbpp)
1696 {
1697     return do_io_submit(ctx_id, nr, iocbpp, 0);
1698 }
1699 
1700 #ifdef CONFIG_COMPAT
1701 static inline long
1702 copy_iocb(long nr, u32 __user *ptr32, struct iocb __user * __user *ptr64)
1703 {
1704     compat_uptr_t uptr;
1705     int i;
1706 
1707     for (i = 0; i < nr; ++i) {
1708         if (get_user(uptr, ptr32 + i))
1709             return -EFAULT;
1710         if (put_user(compat_ptr(uptr), ptr64 + i))
1711             return -EFAULT;
1712     }
1713     return 0;
1714 }
1715 
1716 #define MAX_AIO_SUBMITS     (PAGE_SIZE/sizeof(struct iocb *))
1717 
1718 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1719                int, nr, u32 __user *, iocb)
1720 {
1721     struct iocb __user * __user *iocb64;
1722     long ret;
1723 
1724     if (unlikely(nr < 0))
1725         return -EINVAL;
1726 
1727     if (nr > MAX_AIO_SUBMITS)
1728         nr = MAX_AIO_SUBMITS;
1729 
1730     iocb64 = compat_alloc_user_space(nr * sizeof(*iocb64));
1731     ret = copy_iocb(nr, iocb, iocb64);
1732     if (!ret)
1733         ret = do_io_submit(ctx_id, nr, iocb64, 1);
1734     return ret;
1735 }
1736 #endif
1737 
1738 /* lookup_kiocb
1739  *  Finds a given iocb for cancellation.
1740  */
1741 static struct aio_kiocb *
1742 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1743 {
1744     struct aio_kiocb *kiocb;
1745 
1746     assert_spin_locked(&ctx->ctx_lock);
1747 
1748     if (key != KIOCB_KEY)
1749         return NULL;
1750 
1751     /* TODO: use a hash or array, this sucks. */
1752     list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1753         if (kiocb->ki_user_iocb == iocb)
1754             return kiocb;
1755     }
1756     return NULL;
1757 }
1758 
1759 /* sys_io_cancel:
1760  *  Attempts to cancel an iocb previously passed to io_submit.  If
1761  *  the operation is successfully cancelled, the resulting event is
1762  *  copied into the memory pointed to by result without being placed
1763  *  into the completion queue and 0 is returned.  May fail with
1764  *  -EFAULT if any of the data structures pointed to are invalid.
1765  *  May fail with -EINVAL if aio_context specified by ctx_id is
1766  *  invalid.  May fail with -EAGAIN if the iocb specified was not
1767  *  cancelled.  Will fail with -ENOSYS if not implemented.
1768  */
1769 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1770         struct io_event __user *, result)
1771 {
1772     struct kioctx *ctx;
1773     struct aio_kiocb *kiocb;
1774     u32 key;
1775     int ret;
1776 
1777     ret = get_user(key, &iocb->aio_key);
1778     if (unlikely(ret))
1779         return -EFAULT;
1780 
1781     ctx = lookup_ioctx(ctx_id);
1782     if (unlikely(!ctx))
1783         return -EINVAL;
1784 
1785     spin_lock_irq(&ctx->ctx_lock);
1786 
1787     kiocb = lookup_kiocb(ctx, iocb, key);
1788     if (kiocb)
1789         ret = kiocb_cancel(kiocb);
1790     else
1791         ret = -EINVAL;
1792 
1793     spin_unlock_irq(&ctx->ctx_lock);
1794 
1795     if (!ret) {
1796         /*
1797          * The result argument is no longer used - the io_event is
1798          * always delivered via the ring buffer. -EINPROGRESS indicates
1799          * cancellation is progress:
1800          */
1801         ret = -EINPROGRESS;
1802     }
1803 
1804     percpu_ref_put(&ctx->users);
1805 
1806     return ret;
1807 }
1808 
1809 /* io_getevents:
1810  *  Attempts to read at least min_nr events and up to nr events from
1811  *  the completion queue for the aio_context specified by ctx_id. If
1812  *  it succeeds, the number of read events is returned. May fail with
1813  *  -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1814  *  out of range, if timeout is out of range.  May fail with -EFAULT
1815  *  if any of the memory specified is invalid.  May return 0 or
1816  *  < min_nr if the timeout specified by timeout has elapsed
1817  *  before sufficient events are available, where timeout == NULL
1818  *  specifies an infinite timeout. Note that the timeout pointed to by
1819  *  timeout is relative.  Will fail with -ENOSYS if not implemented.
1820  */
1821 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1822         long, min_nr,
1823         long, nr,
1824         struct io_event __user *, events,
1825         struct timespec __user *, timeout)
1826 {
1827     struct kioctx *ioctx = lookup_ioctx(ctx_id);
1828     long ret = -EINVAL;
1829 
1830     if (likely(ioctx)) {
1831         if (likely(min_nr <= nr && min_nr >= 0))
1832             ret = read_events(ioctx, min_nr, nr, events, timeout);
1833         percpu_ref_put(&ioctx->users);
1834     }
1835     return ret;
1836 }
1837 
1838 #ifdef CONFIG_COMPAT
1839 COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id,
1840                compat_long_t, min_nr,
1841                compat_long_t, nr,
1842                struct io_event __user *, events,
1843                struct compat_timespec __user *, timeout)
1844 {
1845     struct timespec t;
1846     struct timespec __user *ut = NULL;
1847 
1848     if (timeout) {
1849         if (compat_get_timespec(&t, timeout))
1850             return -EFAULT;
1851 
1852         ut = compat_alloc_user_space(sizeof(*ut));
1853         if (copy_to_user(ut, &t, sizeof(t)))
1854             return -EFAULT;
1855     }
1856     return sys_io_getevents(ctx_id, min_nr, nr, events, ut);
1857 }
1858 #endif