[Qemu-devel] [PATCH] block/io: optimize bdrv_co_pwritev for small requests

[Qemu-devel] [PATCH] block/io: optimize bdrv_co_pwritev for small requests

[Peter Lieven]

in a read-modify-write cycle a small request might cause
head and tail to fall into the same alignment. Currently
QEMU reads the same block twice in this case which is
not necessary.

Signed-off-by: Peter Lieven <pl@kamp.de>
---
 block/io.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/block/io.c b/block/io.c
index 60a6bd8..fa40121 100644
--- a/block/io.c
+++ b/block/io.c
@@ -1430,6 +1430,18 @@ int coroutine_fn bdrv_co_pwritev(BlockDriverState *bs,
 
         bytes += offset & (align - 1);
         offset = offset & ~(align - 1);
+
+        /* if head and tail fall into the same alignment
+         * we can omit the second read as it would read
+         * the same block again */
+        if ((offset + bytes) & (align - 1) &&
+            offset / align == (offset + bytes) / align) {
+            size_t tail_offs;
+            tail_offs = (offset + bytes) & (align - 1);
+            qemu_iovec_add(&local_qiov, head_buf + tail_offs,
+                           align - tail_offs);
+            bytes += align - tail_offs;
+        }
     }
 
     if ((offset + bytes) & (align - 1)) {
-- 
1.9.1

Re: [Qemu-devel] [PATCH] block/io: optimize bdrv_co_pwritev for small requests

[Paolo Bonzini]

On 24/05/2016 15:39, Peter Lieven wrote:
>          bytes += offset & (align - 1);
>          offset = offset & ~(align - 1);

Because the low bits have been masked away from offset and added to bytes,

> +
> +        /* if head and tail fall into the same alignment
> +         * we can omit the second read as it would read
> +         * the same block again */
> +        if ((offset + bytes) & (align - 1) &&

... the first part is just "bytes & (align - 1)"...

> +            offset / align == (offset + bytes) / align) {

... and the second part is just "bytes < align" (you can distribute
division over addition because offset / align has no reminder, and
simplify to "0 == bytes / align").

Putting it together, it becomes "bytes > 0 && bytes < align", or even
"bytes < align".

Thanks,

Paolo

> +            size_t tail_offs;
> +            tail_offs = (offset + bytes) & (align - 1);
> +            qemu_iovec_add(&local_qiov, head_buf + tail_offs,
> +                           align - tail_offs);
> +            bytes += align - tail_offs;
> +        }

Re: [Qemu-devel] [PATCH] block/io: optimize bdrv_co_pwritev for small requests

[Peter Lieven]

Am 24.05.2016 um 15:59 schrieb Paolo Bonzini:
>
> On 24/05/2016 15:39, Peter Lieven wrote:
>>           bytes += offset & (align - 1);
>>           offset = offset & ~(align - 1);
> Because the low bits have been masked away from offset and added to bytes,
>
>> +
>> +        /* if head and tail fall into the same alignment
>> +         * we can omit the second read as it would read
>> +         * the same block again */
>> +        if ((offset + bytes) & (align - 1) &&
> ... the first part is just "bytes & (align - 1)"...
>
>> +            offset / align == (offset + bytes) / align) {
> ... and the second part is just "bytes < align" (you can distribute
> division over addition because offset / align has no reminder, and
> simplify to "0 == bytes / align").
>
> Putting it together, it becomes "bytes > 0 && bytes < align", or even
> "bytes < align".

Oh, thanks, and the if block also too complicated. If I am right it should
collapse to:

         if (bytes < align) {
             qemu_iovec_add(&local_qiov, head_buf + bytes,
                            align - bytes);
             bytes = align;
         }

Right?

Thanks,
Peter

Re: [Qemu-devel] [PATCH] block/io: optimize bdrv_co_pwritev for small requests

[Paolo Bonzini]

On 24/05/2016 16:07, Peter Lieven wrote:
> 
> Oh, thanks, and the if block also too complicated. If I am right it should
> collapse to:
> 
>         if (bytes < align) {
>             qemu_iovec_add(&local_qiov, head_buf + bytes,
>                            align - bytes);
>             bytes = align;
>         }
> 
> Right?

Yes, that should work.  But add a comment because it's much more
mysterious than your v1 patch. :)

Even just

	/* We have read the tail already if the request is smaller
	 * than one aligned block.
	 */

Thanks,

Paolo

Re: [Qemu-devel] [PATCH] block/io: optimize bdrv_co_pwritev for small requests

[Kevin Wolf]

Am 2.05.2016 um 16:07 hat Peter Lieven geschrieben:
> Am 24.05.2016 um 15:59 schrieb Paolo Bonzini:
> >
> >On 24/05/2016 15:39, Peter Lieven wrote:
> >>          bytes += offset & (align - 1);
> >>          offset = offset & ~(align - 1);
> >Because the low bits have been masked away from offset and added to bytes,
> >
> >>+
> >>+        /* if head and tail fall into the same alignment
> >>+         * we can omit the second read as it would read
> >>+         * the same block again */
> >>+        if ((offset + bytes) & (align - 1) &&
> >... the first part is just "bytes & (align - 1)"...
> >
> >>+            offset / align == (offset + bytes) / align) {
> >... and the second part is just "bytes < align" (you can distribute
> >division over addition because offset / align has no reminder, and
> >simplify to "0 == bytes / align").
> >
> >Putting it together, it becomes "bytes > 0 && bytes < align", or even
> >"bytes < align".
> 
> Oh, thanks, and the if block also too complicated. If I am right it should
> collapse to:
> 
>         if (bytes < align) {
>             qemu_iovec_add(&local_qiov, head_buf + bytes,
>                            align - bytes);
>             bytes = align;
>         }
> 
> Right?

Looks good to me.

Another mostly unrelated thing I just noticed while looking at this
code: Should we assert(is_power_of_2(align)) somewhere?

Kevin