From: Dave Chinner <dchinner@xxxxxxxxxx>
If we are doing direct IO writes with datasync semantics, we often
have to flush metadata changes along with the data write. However,
if we are overwriting existing data, there are no metadata changes
that we need to flush. In this case, optimising the IO by using
FUA write makes sense.
We know from the IOMAP_F_DIRTY flag as to whether a specific inode
requires a metadata flush - this is currently used by DAX to ensure
extent modification as stable in page fault operations. For direct
IO writes, we can use it to determine if we need to flush metadata
or not once the data is on disk.
Hence if we have been returned a mapped extent that is not new and
the IO mapping is not dirty, then we can use a FUA write to provide
datasync semantics. This allows us to short-cut the
generic_write_sync() call in IO completion and hence avoid
unnecessary operations. This makes pure direct IO data write
behaviour identical to the way block devices use REQ_FUA to provide
datasync semantics.
On a FUA enabled device, a synchronous direct IO write workload
(sequential 4k overwrites in 32MB file) had the following results:
# xfs_io -fd -c "pwrite -V 1 -D 0 32m" /mnt/scratch/boo
kernel time write()s write iops Write b/w
------ ---- -------- ---------- ---------
(no dsync) 4s 2173/s 2173 8.5MB/s
vanilla 22s 370/s 750 1.4MB/s
patched 19s 420/s 420 1.6MB/s
The patched code clearly doesn't send cache flushes anymore, but
instead uses FUA (confirmed via blktrace), and performance improves
a bit as a result. However, the benefits will be higher on workloads
that mix O_DSYNC overwrites with other write IO as we won't be
flushing the entire device cache on every DSYNC overwrite IO
anymore.
Signed-Off-By: Dave Chinner <dchinner@xxxxxxxxxx>
Reviewed-by: Christoph Hellwig <hch@xxxxxx>
---
fs/iomap.c | 48 +++++++++++++++++++++++++++++++++++++++++++-----
1 file changed, 43 insertions(+), 5 deletions(-)
diff --git a/fs/iomap.c b/fs/iomap.c
index 1f59c2d9ade6..62f1f8256da2 100644
--- a/fs/iomap.c
+++ b/fs/iomap.c
@@ -685,6 +685,7 @@ EXPORT_SYMBOL_GPL(iomap_seek_data);
* Private flags for iomap_dio, must not overlap with the public ones in
* iomap.h:
*/
+#define IOMAP_DIO_WRITE_FUA (1 << 28)
#define IOMAP_DIO_NEED_SYNC (1 << 29)
#define IOMAP_DIO_WRITE (1 << 30)
#define IOMAP_DIO_DIRTY (1 << 31)
@@ -860,6 +861,7 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
struct iov_iter iter;
struct bio *bio;
bool need_zeroout = false;
+ bool use_fua = false;
int nr_pages, ret;
size_t copied = 0;
@@ -883,8 +885,20 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
case IOMAP_MAPPED:
if (iomap->flags & IOMAP_F_SHARED)
dio->flags |= IOMAP_DIO_COW;
- if (iomap->flags & IOMAP_F_NEW)
+ if (iomap->flags & IOMAP_F_NEW) {
need_zeroout = true;
+ } else {
+ /*
+ * Use a FUA write if we need datasync semantics, this
+ * is a pure data IO that doesn't require any metadata
+ * updates and the underlying device supports FUA. This
+ * allows us to avoid cache flushes on IO completion.
+ */
+ if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
+ (dio->flags & IOMAP_DIO_WRITE_FUA) &&
+ blk_queue_fua(bdev_get_queue(iomap->bdev)))
+ use_fua = true;
+ }
break;
default:
WARN_ON_ONCE(1);
@@ -932,10 +946,14 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
n = bio->bi_iter.bi_size;
if (dio->flags & IOMAP_DIO_WRITE) {
- bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
+ bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
+ if (use_fua)
+ bio->bi_opf |= REQ_FUA;
+ else
+ dio->flags &= ~IOMAP_DIO_WRITE_FUA;
task_io_account_write(n);
} else {
- bio_set_op_attrs(bio, REQ_OP_READ, 0);
+ bio->bi_opf = REQ_OP_READ;
if (dio->flags & IOMAP_DIO_DIRTY)
bio_set_pages_dirty(bio);
}
@@ -965,7 +983,12 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
/*
* iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
- * is being issued as AIO or not.
+ * is being issued as AIO or not. This allows us to optimise pure data writes
+ * to use REQ_FUA rather than requiring generic_write_sync() to issue a
+ * REQ_FLUSH post write. This is slightly tricky because a single request here
+ * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
+ * may be pure data writes. In that case, we still need to do a full data sync
+ * completion.
*/
ssize_t
iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
@@ -1012,8 +1035,16 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
dio->flags |= IOMAP_DIO_DIRTY;
} else {
dio->flags |= IOMAP_DIO_WRITE;
- if (iocb->ki_flags & IOCB_DSYNC)
+ if (iocb->ki_flags & IOCB_DSYNC) {
dio->flags |= IOMAP_DIO_NEED_SYNC;
+ /*
+ * We optimistically try using FUA for this IO. Any
+ * non-FUA write that occurs will clear this flag, hence
+ * we know before completion whether a cache flush is
+ * necessary.
+ */
+ dio->flags |= IOMAP_DIO_WRITE_FUA;
+ }
flags |= IOMAP_WRITE;
}
@@ -1070,6 +1101,13 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
if (ret < 0)
iomap_dio_set_error(dio, ret);
+ /*
+ * If all the writes we issued were FUA, we don't need to flush the
+ * cache on IO completion. Clear the sync flag for this case.
+ */
+ if (dio->flags & IOMAP_DIO_WRITE_FUA)
+ dio->flags &= ~IOMAP_DIO_NEED_SYNC;
+
if (!atomic_dec_and_test(&dio->ref)) {
if (!is_sync_kiocb(iocb))
return -EIOCBQUEUED;
--
2.16.1
