Re: [Qemu-devel] [PATCH v5 4/4] block: ignore flush requests when storage is clean

[Eric Blake <eblake@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 3169 bytes --]

On 07/04/2016 08:38 AM, Denis V. Lunev wrote:
> From: Evgeny Yakovlev <eyakovlev@virtuozzo.com>
> 
> Some guests (win2008 server for example) do a lot of unnecessary
> flushing when underlying media has not changed. This adds additional
> overhead on host when calling fsync/fdatasync.
> 
> This change introduces a write generation scheme in BlockDriverState.
> Current write generation is checked against last flushed generation to
> avoid unnessesary flushes.
> 
> The problem with excessive flushing was found by a performance test
> which does parallel directory tree creation (from 2 processes).
> Results improved from 0.424 loops/sec to 0.432 loops/sec.
> Each loop creates 10^3 directories with 10 files in each.
> 

> +++ b/block/io.c
> @@ -1294,6 +1294,7 @@ static int coroutine_fn bdrv_aligned_pwritev(BlockDriverState *bs,
>      }
>      bdrv_debug_event(bs, BLKDBG_PWRITEV_DONE);
>  
> +    ++bs->write_gen;

Why pre-increment?  Most code uses post-increment when done as a
statement in isolation.

>      bdrv_set_dirty(bs, start_sector, end_sector - start_sector);
>  
>      if (bs->wr_highest_offset < offset + bytes) {
> @@ -2211,6 +2212,7 @@ int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
>  {
>      int ret;
>      BdrvTrackedRequest req;
> +    int current_gen = bs->write_gen;
>  
>      if (!bs || !bdrv_is_inserted(bs) || bdrv_is_read_only(bs) ||
>          bdrv_is_sg(bs)) {
> @@ -2219,6 +2221,12 @@ int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
>  
>      tracked_request_begin(&req, bs, 0, 0, BDRV_TRACKED_FLUSH);
>  
> +    /* Wait until any previous flushes are completed */
> +    while (bs->flush_started_gen != bs->flushed_gen) {

Should this be an inequality, as in s/!=/</, in case several flushes can
be started in parallel and where the later flush ends up finishing
before the earlier flush?

> +        qemu_co_queue_wait(&bs->flush_queue);
> +    }
> +    bs->flush_started_gen = current_gen;
> +
>      /* Write back all layers by calling one driver function */
>      if (bs->drv->bdrv_co_flush) {
>          ret = bs->drv->bdrv_co_flush(bs);
> @@ -2239,6 +2247,11 @@ int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
>          goto flush_parent;
>      }
>  
> +    /* Check if we really need to flush anything */
> +    if (bs->flushed_gen == current_gen) {

Likewise, if you are tracking generations, should this be s/==/<=/ (am I
getting the direction correct)?

> +++ b/include/block/block_int.h
> @@ -420,6 +420,11 @@ struct BlockDriverState {
>                           note this is a reference count */
>      bool probed;
>  
> +    CoQueue flush_queue;            /* Serializing flush queue */
> +    unsigned int write_gen;         /* Current data generation */
> +    unsigned int flush_started_gen; /* Generation for which flush has started */
> +    unsigned int flushed_gen;       /* Flushed write generation */

Should these be 64-bit integers to avoid risk of overflow after just
2^32 flush attempts?

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 604 bytes --]