struct {
unsigned tail;
+ unsigned completed_events;
spinlock_t completion_lock;
} ____cacheline_aligned_in_smp;
static const struct file_operations aio_ring_fops;
static const struct address_space_operations aio_ctx_aops;
+/* Backing dev info for aio fs.
+ * -no dirty page accounting or writeback happens
+ */
+static struct backing_dev_info aio_fs_backing_dev_info = {
+ .name = "aiofs",
+ .state = 0,
+ .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_MAP_COPY,
+};
+
static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
{
struct qstr this = QSTR_INIT("[aio]", 5);
inode->i_mapping->a_ops = &aio_ctx_aops;
inode->i_mapping->private_data = ctx;
+ inode->i_mapping->backing_dev_info = &aio_fs_backing_dev_info;
inode->i_size = PAGE_SIZE * nr_pages;
path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
if (IS_ERR(aio_mnt))
panic("Failed to create aio fs mount.");
+ if (bdi_init(&aio_fs_backing_dev_info))
+ panic("Failed to init aio fs backing dev info.");
+
kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
.mmap = aio_ring_mmap,
};
-static int aio_set_page_dirty(struct page *page)
-{
- return 0;
-}
-
#if IS_ENABLED(CONFIG_MIGRATION)
static int aio_migratepage(struct address_space *mapping, struct page *new,
struct page *old, enum migrate_mode mode)
#endif
static const struct address_space_operations aio_ctx_aops = {
- .set_page_dirty = aio_set_page_dirty,
+ .set_page_dirty = __set_page_dirty_no_writeback,
#if IS_ENABLED(CONFIG_MIGRATION)
.migratepage = aio_migratepage,
#endif
pr_debug("pid(%d) page[%d]->count=%d\n",
current->pid, i, page_count(page));
SetPageUptodate(page);
- SetPageDirty(page);
unlock_page(page);
ctx->ring_pages[i] = page;
unsigned i = 0;
while (1) {
+ struct completion requests_done =
+ COMPLETION_INITIALIZER_ONSTACK(requests_done);
+
rcu_read_lock();
table = rcu_dereference(mm->ioctx_table);
*/
ctx->mmap_size = 0;
- kill_ioctx(mm, ctx, NULL);
+ kill_ioctx(mm, ctx, &requests_done);
+
+ /* Wait until all IO for the context are done. */
+ wait_for_completion(&requests_done);
}
}
return ret;
}
+/* refill_reqs_available
+ * Updates the reqs_available reference counts used for tracking the
+ * number of free slots in the completion ring. This can be called
+ * from aio_complete() (to optimistically update reqs_available) or
+ * from aio_get_req() (the we're out of events case). It must be
+ * called holding ctx->completion_lock.
+ */
+static void refill_reqs_available(struct kioctx *ctx, unsigned head,
+ unsigned tail)
+{
+ unsigned events_in_ring, completed;
+
+ /* Clamp head since userland can write to it. */
+ head %= ctx->nr_events;
+ if (head <= tail)
+ events_in_ring = tail - head;
+ else
+ events_in_ring = ctx->nr_events - (head - tail);
+
+ completed = ctx->completed_events;
+ if (events_in_ring < completed)
+ completed -= events_in_ring;
+ else
+ completed = 0;
+
+ if (!completed)
+ return;
+
+ ctx->completed_events -= completed;
+ put_reqs_available(ctx, completed);
+}
+
+/* user_refill_reqs_available
+ * Called to refill reqs_available when aio_get_req() encounters an
+ * out of space in the completion ring.
+ */
+static void user_refill_reqs_available(struct kioctx *ctx)
+{
+ spin_lock_irq(&ctx->completion_lock);
+ if (ctx->completed_events) {
+ struct aio_ring *ring;
+ unsigned head;
+
+ /* Access of ring->head may race with aio_read_events_ring()
+ * here, but that's okay since whether we read the old version
+ * or the new version, and either will be valid. The important
+ * part is that head cannot pass tail since we prevent
+ * aio_complete() from updating tail by holding
+ * ctx->completion_lock. Even if head is invalid, the check
+ * against ctx->completed_events below will make sure we do the
+ * safe/right thing.
+ */
+ ring = kmap_atomic(ctx->ring_pages[0]);
+ head = ring->head;
+ kunmap_atomic(ring);
+
+ refill_reqs_available(ctx, head, ctx->tail);
+ }
+
+ spin_unlock_irq(&ctx->completion_lock);
+}
+
/* aio_get_req
* Allocate a slot for an aio request.
* Returns NULL if no requests are free.
{
struct kiocb *req;
- if (!get_reqs_available(ctx))
- return NULL;
+ if (!get_reqs_available(ctx)) {
+ user_refill_reqs_available(ctx);
+ if (!get_reqs_available(ctx))
+ return NULL;
+ }
req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
if (unlikely(!req))
struct kioctx *ctx = iocb->ki_ctx;
struct aio_ring *ring;
struct io_event *ev_page, *event;
+ unsigned tail, pos, head;
unsigned long flags;
- unsigned tail, pos;
/*
* Special case handling for sync iocbs:
ctx->tail = tail;
ring = kmap_atomic(ctx->ring_pages[0]);
+ head = ring->head;
ring->tail = tail;
kunmap_atomic(ring);
flush_dcache_page(ctx->ring_pages[0]);
+ ctx->completed_events++;
+ if (ctx->completed_events > 1)
+ refill_reqs_available(ctx, head, tail);
spin_unlock_irqrestore(&ctx->completion_lock, flags);
pr_debug("added to ring %p at [%u]\n", iocb, tail);
/* everything turned out well, dispose of the aiocb. */
kiocb_free(iocb);
- put_reqs_available(ctx, 1);
/*
* We have to order our ring_info tail store above and test
tail = ring->tail;
kunmap_atomic(ring);
+ /*
+ * Ensure that once we've read the current tail pointer, that
+ * we also see the events that were stored up to the tail.
+ */
+ smp_rmb();
+
pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
if (head == tail)