aio: implement request batching
Jeff Moyer [Fri, 2 Oct 2009 22:57:36 +0000 (18:57 -0400)]
Hi,

Some workloads issue batches of small I/O, and the performance is poor
due to the call to blk_run_address_space for every single iocb.  Nathan
Roberts pointed this out, and suggested that by deferring this call
until all I/Os in the iocb array are submitted to the block layer, we
can realize some impressive performance gains (up to 30% for sequential
4k reads in batches of 16).

Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

fs/aio.c
fs/direct-io.c

index 02a2c93..cf0bef4 100644 (file)
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -32,6 +32,9 @@
 #include <linux/workqueue.h>
 #include <linux/security.h>
 #include <linux/eventfd.h>
+#include <linux/blkdev.h>
+#include <linux/mempool.h>
+#include <linux/hash.h>
 
 #include <asm/kmap_types.h>
 #include <asm/uaccess.h>
@@ -60,6 +63,14 @@ static DECLARE_WORK(fput_work, aio_fput_routine);
 static DEFINE_SPINLOCK(fput_lock);
 static LIST_HEAD(fput_head);
 
+#define AIO_BATCH_HASH_BITS    3 /* allocated on-stack, so don't go crazy */
+#define AIO_BATCH_HASH_SIZE    (1 << AIO_BATCH_HASH_BITS)
+struct aio_batch_entry {
+       struct hlist_node list;
+       struct address_space *mapping;
+};
+mempool_t *abe_pool;
+
 static void aio_kick_handler(struct work_struct *);
 static void aio_queue_work(struct kioctx *);
 
@@ -73,6 +84,8 @@ static int __init aio_setup(void)
        kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 
        aio_wq = create_workqueue("aio");
+       abe_pool = mempool_create_kmalloc_pool(1, sizeof(struct aio_batch_entry));
+       BUG_ON(!abe_pool);
 
        pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
 
@@ -1531,8 +1544,44 @@ static int aio_wake_function(wait_queue_t *wait, unsigned mode,
        return 1;
 }
 
+static void aio_batch_add(struct address_space *mapping,
+                         struct hlist_head *batch_hash)
+{
+       struct aio_batch_entry *abe;
+       struct hlist_node *pos;
+       unsigned bucket;
+
+       bucket = hash_ptr(mapping, AIO_BATCH_HASH_BITS);
+       hlist_for_each_entry(abe, pos, &batch_hash[bucket], list) {
+               if (abe->mapping == mapping)
+                       return;
+       }
+
+       abe = mempool_alloc(abe_pool, GFP_KERNEL);
+       BUG_ON(!igrab(mapping->host));
+       abe->mapping = mapping;
+       hlist_add_head(&abe->list, &batch_hash[bucket]);
+       return;
+}
+
+static void aio_batch_free(struct hlist_head *batch_hash)
+{
+       struct aio_batch_entry *abe;
+       struct hlist_node *pos, *n;
+       int i;
+
+       for (i = 0; i < AIO_BATCH_HASH_SIZE; i++) {
+               hlist_for_each_entry_safe(abe, pos, n, &batch_hash[i], list) {
+                       blk_run_address_space(abe->mapping);
+                       iput(abe->mapping->host);
+                       hlist_del(&abe->list);
+                       mempool_free(abe, abe_pool);
+               }
+       }
+}
+
 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
-                        struct iocb *iocb)
+                        struct iocb *iocb, struct hlist_head *batch_hash)
 {
        struct kiocb *req;
        struct file *file;
@@ -1608,6 +1657,12 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
                        ;
        }
        spin_unlock_irq(&ctx->ctx_lock);
+       if (req->ki_opcode == IOCB_CMD_PREAD ||
+           req->ki_opcode == IOCB_CMD_PREADV ||
+           req->ki_opcode == IOCB_CMD_PWRITE ||
+           req->ki_opcode == IOCB_CMD_PWRITEV)
+               aio_batch_add(file->f_mapping, batch_hash);
+
        aio_put_req(req);       /* drop extra ref to req */
        return 0;
 
@@ -1635,6 +1690,7 @@ SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
        struct kioctx *ctx;
        long ret = 0;
        int i;
+       struct hlist_head batch_hash[AIO_BATCH_HASH_SIZE] = { { 0, }, };
 
        if (unlikely(nr < 0))
                return -EINVAL;
@@ -1666,10 +1722,11 @@ SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
                        break;
                }
 
-               ret = io_submit_one(ctx, user_iocb, &tmp);
+               ret = io_submit_one(ctx, user_iocb, &tmp, batch_hash);
                if (ret)
                        break;
        }
+       aio_batch_free(batch_hash);
 
        put_ioctx(ctx);
        return i ? i : ret;
index c86d35f..3af761c 100644 (file)
@@ -1028,9 +1028,6 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
        if (dio->bio)
                dio_bio_submit(dio);
 
-       /* All IO is now issued, send it on its way */
-       blk_run_address_space(inode->i_mapping);
-
        /*
         * It is possible that, we return short IO due to end of file.
         * In that case, we need to release all the pages we got hold on.
@@ -1057,8 +1054,11 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
            ((rw & READ) || (dio->result == dio->size)))
                ret = -EIOCBQUEUED;
 
-       if (ret != -EIOCBQUEUED)
+       if (ret != -EIOCBQUEUED) {
+               /* All IO is now issued, send it on its way */
+               blk_run_address_space(inode->i_mapping);
                dio_await_completion(dio);
+       }
 
        /*
         * Sync will always be dropping the final ref and completing the