direct-io: only inc/dec inode->i_dio_count for file systems
authorJens Axboe <axboe@fb.com>
Wed, 15 Apr 2015 23:05:48 +0000 (17:05 -0600)
committerAl Viro <viro@zeniv.linux.org.uk>
Fri, 24 Apr 2015 19:45:28 +0000 (15:45 -0400)
do_blockdev_direct_IO() increments and decrements the inode
->i_dio_count for each IO operation. It does this to protect against
truncate of a file. Block devices don't need this sort of protection.

For a capable multiqueue setup, this atomic int is the only shared
state between applications accessing the device for O_DIRECT, and it
presents a scaling wall for that. In my testing, as much as 30% of
system time is spent incrementing and decrementing this value. A mixed
read/write workload improved from ~2.5M IOPS to ~9.6M IOPS, with
better latencies too. Before:

clat percentiles (usec):
 |  1.00th=[   33],  5.00th=[   34], 10.00th=[   34], 20.00th=[   34],
 | 30.00th=[   34], 40.00th=[   34], 50.00th=[   35], 60.00th=[   35],
 | 70.00th=[   35], 80.00th=[   35], 90.00th=[   37], 95.00th=[   80],
 | 99.00th=[   98], 99.50th=[  151], 99.90th=[  155], 99.95th=[  155],
 | 99.99th=[  165]

After:

clat percentiles (usec):
 |  1.00th=[   95],  5.00th=[  108], 10.00th=[  129], 20.00th=[  149],
 | 30.00th=[  155], 40.00th=[  161], 50.00th=[  167], 60.00th=[  171],
 | 70.00th=[  177], 80.00th=[  185], 90.00th=[  201], 95.00th=[  270],
 | 99.00th=[  390], 99.50th=[  398], 99.90th=[  418], 99.95th=[  422],
 | 99.99th=[  438]

In other setups, Robert Elliott reported seeing good performance
improvements:

https://lkml.org/lkml/2015/4/3/557

The more applications accessing the device, the worse it gets.

Add a new direct-io flags, DIO_SKIP_DIO_COUNT, which tells
do_blockdev_direct_IO() that it need not worry about incrementing
or decrementing the inode i_dio_count for this caller.

Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Elliott, Robert (Server Storage) <elliott@hp.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jens Axboe <axboe@fb.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
fs/block_dev.c
fs/btrfs/inode.c
fs/dax.c
fs/direct-io.c
fs/ext4/indirect.c
fs/ext4/inode.c
fs/inode.c
fs/nfs/direct.c
include/linux/fs.h

index 79b4fa3b391d221b08ffd66a853109901c1cfd3f..c7e4163ede87f3370a235bdb929cf26e5cd487e8 100644 (file)
@@ -152,7 +152,8 @@ blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t offset)
        struct inode *inode = file->f_mapping->host;
 
        return __blockdev_direct_IO(iocb, inode, I_BDEV(inode), iter, offset,
-                                   blkdev_get_block, NULL, NULL, 0);
+                                   blkdev_get_block, NULL, NULL,
+                                   DIO_SKIP_DIO_COUNT);
 }
 
 int __sync_blockdev(struct block_device *bdev, int wait)
index 811576346a92f48508e8538a9e7256c9a6fef635..9b774a0c9ca61c61640219fe7ec5c0c0f8bb428f 100644 (file)
@@ -8129,7 +8129,7 @@ static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
        if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
                return 0;
 
-       atomic_inc(&inode->i_dio_count);
+       inode_dio_begin(inode);
        smp_mb__after_atomic();
 
        /*
@@ -8169,7 +8169,7 @@ static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
                current->journal_info = &outstanding_extents;
        } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
                                     &BTRFS_I(inode)->runtime_flags)) {
-               inode_dio_done(inode);
+               inode_dio_end(inode);
                flags = DIO_LOCKING | DIO_SKIP_HOLES;
                wakeup = false;
        }
@@ -8188,7 +8188,7 @@ static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
        }
 out:
        if (wakeup)
-               inode_dio_done(inode);
+               inode_dio_end(inode);
        if (relock)
                mutex_lock(&inode->i_mutex);
 
index a27846946525ed48b87071743f21ae744dacac36..cdde2b354e47b6925f677f5f66d4586c39ab1a50 100644 (file)
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -209,7 +209,7 @@ ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
        }
 
        /* Protects against truncate */
-       atomic_inc(&inode->i_dio_count);
+       inode_dio_begin(inode);
 
        retval = dax_io(inode, iter, pos, end, get_block, &bh);
 
@@ -219,7 +219,7 @@ ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
        if ((retval > 0) && end_io)
                end_io(iocb, pos, retval, bh.b_private);
 
-       inode_dio_done(inode);
+       inode_dio_end(inode);
  out:
        return retval;
 }
index c3b560b24a463c50569a2a37c5d04f8d5af92c4d..745d2342651a0b87fcb9dd5b5b50e95b9129e117 100644 (file)
@@ -253,7 +253,9 @@ static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
        if (dio->end_io && dio->result)
                dio->end_io(dio->iocb, offset, transferred, dio->private);
 
-       inode_dio_done(dio->inode);
+       if (!(dio->flags & DIO_SKIP_DIO_COUNT))
+               inode_dio_end(dio->inode);
+
        if (is_async) {
                if (dio->rw & WRITE) {
                        int err;
@@ -1195,7 +1197,8 @@ do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
        /*
         * Will be decremented at I/O completion time.
         */
-       atomic_inc(&inode->i_dio_count);
+       if (!(dio->flags & DIO_SKIP_DIO_COUNT))
+               inode_dio_begin(inode);
 
        retval = 0;
        sdio.blkbits = blkbits;
index 3580629e42d32aadaff81effd01fddb2aaa41484..9588240195090ce4216e87f07379121b4ffe6cfb 100644 (file)
@@ -682,11 +682,11 @@ retry:
                 * via ext4_inode_block_unlocked_dio(). Check inode's state
                 * while holding extra i_dio_count ref.
                 */
-               atomic_inc(&inode->i_dio_count);
+               inode_dio_begin(inode);
                smp_mb();
                if (unlikely(ext4_test_inode_state(inode,
                                                    EXT4_STATE_DIOREAD_LOCK))) {
-                       inode_dio_done(inode);
+                       inode_dio_end(inode);
                        goto locked;
                }
                if (IS_DAX(inode))
@@ -697,7 +697,7 @@ retry:
                                                   inode->i_sb->s_bdev, iter,
                                                   offset, ext4_get_block, NULL,
                                                   NULL, 0);
-               inode_dio_done(inode);
+               inode_dio_end(inode);
        } else {
 locked:
                if (IS_DAX(inode))
index 063052e4aa8ba37e0ab0c3f2479aca097c174c3c..bccec41fb94b732f9ba9911fd542bc095972580d 100644 (file)
@@ -2977,7 +2977,7 @@ static ssize_t ext4_ext_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
         * overwrite DIO as i_dio_count needs to be incremented under i_mutex.
         */
        if (iov_iter_rw(iter) == WRITE)
-               atomic_inc(&inode->i_dio_count);
+               inode_dio_begin(inode);
 
        /* If we do a overwrite dio, i_mutex locking can be released */
        overwrite = *((int *)iocb->private);
@@ -3079,7 +3079,7 @@ static ssize_t ext4_ext_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
 
 retake_lock:
        if (iov_iter_rw(iter) == WRITE)
-               inode_dio_done(inode);
+               inode_dio_end(inode);
        /* take i_mutex locking again if we do a ovewrite dio */
        if (overwrite) {
                up_read(&EXT4_I(inode)->i_data_sem);
index 94886f9fbb0667dccc4ab1d1ca1dab27470375e0..ea37cd17b53f0c98b47e2e626b9ff2a9b6e5699b 100644 (file)
@@ -1945,20 +1945,6 @@ void inode_dio_wait(struct inode *inode)
 }
 EXPORT_SYMBOL(inode_dio_wait);
 
-/*
- * inode_dio_done - signal finish of a direct I/O requests
- * @inode: inode the direct I/O happens on
- *
- * This is called once we've finished processing a direct I/O request,
- * and is used to wake up callers waiting for direct I/O to be quiesced.
- */
-void inode_dio_done(struct inode *inode)
-{
-       if (atomic_dec_and_test(&inode->i_dio_count))
-               wake_up_bit(&inode->i_state, __I_DIO_WAKEUP);
-}
-EXPORT_SYMBOL(inode_dio_done);
-
 /*
  * inode_set_flags - atomically set some inode flags
  *
index ed0e6031be88eb2d291f7105a4112560c39caf2b..b2cbc3a6cdd9dd2702c97872f8caa14038dbe88a 100644 (file)
@@ -386,7 +386,7 @@ static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
        if (write)
                nfs_zap_mapping(inode, inode->i_mapping);
 
-       inode_dio_done(inode);
+       inode_dio_end(inode);
 
        if (dreq->iocb) {
                long res = (long) dreq->error;
@@ -486,7 +486,7 @@ static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
                             &nfs_direct_read_completion_ops);
        get_dreq(dreq);
        desc.pg_dreq = dreq;
-       atomic_inc(&inode->i_dio_count);
+       inode_dio_begin(inode);
 
        while (iov_iter_count(iter)) {
                struct page **pagevec;
@@ -538,7 +538,7 @@ static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
         * generic layer handle the completion.
         */
        if (requested_bytes == 0) {
-               inode_dio_done(inode);
+               inode_dio_end(inode);
                nfs_direct_req_release(dreq);
                return result < 0 ? result : -EIO;
        }
@@ -872,7 +872,7 @@ static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
                              &nfs_direct_write_completion_ops);
        desc.pg_dreq = dreq;
        get_dreq(dreq);
-       atomic_inc(&inode->i_dio_count);
+       inode_dio_begin(inode);
 
        NFS_I(inode)->write_io += iov_iter_count(iter);
        while (iov_iter_count(iter)) {
@@ -928,7 +928,7 @@ static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
         * generic layer handle the completion.
         */
        if (requested_bytes == 0) {
-               inode_dio_done(inode);
+               inode_dio_end(inode);
                nfs_direct_req_release(dreq);
                return result < 0 ? result : -EIO;
        }
index b1d7db28c13c90b80e2e02954e753f31cb5b0f27..9055eefa92c74049b718729e1c54be47e2767ed6 100644 (file)
@@ -2635,6 +2635,9 @@ enum {
 
        /* filesystem can handle aio writes beyond i_size */
        DIO_ASYNC_EXTEND = 0x04,
+
+       /* inode/fs/bdev does not need truncate protection */
+       DIO_SKIP_DIO_COUNT = 0x08,
 };
 
 void dio_end_io(struct bio *bio, int error);
@@ -2657,7 +2660,31 @@ static inline ssize_t blockdev_direct_IO(struct kiocb *iocb,
 #endif
 
 void inode_dio_wait(struct inode *inode);
-void inode_dio_done(struct inode *inode);
+
+/*
+ * inode_dio_begin - signal start of a direct I/O requests
+ * @inode: inode the direct I/O happens on
+ *
+ * This is called once we've finished processing a direct I/O request,
+ * and is used to wake up callers waiting for direct I/O to be quiesced.
+ */
+static inline void inode_dio_begin(struct inode *inode)
+{
+       atomic_inc(&inode->i_dio_count);
+}
+
+/*
+ * inode_dio_end - signal finish of a direct I/O requests
+ * @inode: inode the direct I/O happens on
+ *
+ * This is called once we've finished processing a direct I/O request,
+ * and is used to wake up callers waiting for direct I/O to be quiesced.
+ */
+static inline void inode_dio_end(struct inode *inode)
+{
+       if (atomic_dec_and_test(&inode->i_dio_count))
+               wake_up_bit(&inode->i_state, __I_DIO_WAKEUP);
+}
 
 extern void inode_set_flags(struct inode *inode, unsigned int flags,
                            unsigned int mask);