Reply [PATCH] dm-verity: unnecessary data blocks that need not read hash blocks

Reply [PATCH] dm-verity: unnecessary data blocks that need not read hash blocks

[xianrong.zhou(周先荣)]

hey Eric:

On Wed, Dec 11, 2019 at 11:32:40AM +0800, zhou xianrong wrote:
> From: "xianrong.zhou" <xianrong.zhou@transsion.com>
> 
> If check_at_most_once enabled, just like verity work the prefetching 
> work should check for data block bitmap firstly before reading hash 
> block as well. Skip bit-set data blocks from both ends of data block 
> range by testing the validated bitmap. This can reduce the amounts of 
> data blocks which need to read hash blocks.
> 
> Launching 91 apps every 15s and repeat 21 rounds on Android Q.
> In prefetching work we can let only 2602/360312 = 0.72% data blocks 
> really need to read hash blocks.
> 
> But the reduced data blocks range would be enlarged again by 
> dm_verity_prefetch_cluster later.
> 
> Signed-off-by: xianrong.zhou <xianrong.zhou@transsion.com>
> Signed-off-by: yuanjiong.gao <yuanjiong.gao@transsion.com>
> Tested-by: ruxian.feng <ruxian.feng@transsion.com>
> ---
>  drivers/md/dm-verity-target.c | 16 ++++++++++++++++
>  1 file changed, 16 insertions(+)
> 
> diff --git a/drivers/md/dm-verity-target.c 
> b/drivers/md/dm-verity-target.c index 4fb33e7562c5..7b8eb754c0b6 
> 100644
> --- a/drivers/md/dm-verity-target.c
> +++ b/drivers/md/dm-verity-target.c
> @@ -581,6 +581,22 @@ static void verity_prefetch_io(struct work_struct *work)
>  	struct dm_verity *v = pw->v;
>  	int i;
>  
> +	if (v->validated_blocks) {
> +		while (pw->n_blocks) {
> +			if (unlikely(!test_bit(pw->block, v->validated_blocks)))
> +				break;
> +			pw->block++;
> +			pw->n_blocks--;
> +		}
> +		while (pw->n_blocks) {
> +			if (unlikely(!test_bit(pw->block + pw->n_blocks - 1,
> +				v->validated_blocks)))
> +				break;
> +			pw->n_blocks--;
> +		}
> +		if (!pw->n_blocks)
> +			return;
> +	}

This is a good idea, but shouldn't this logic go in verity_submit_prefetch() prior to the struct dm_verity_prefetch_work being allocated?  Then if no prefeching is needed, allocating and scheduling the work object can be skipped.

Eric, Do you mean it is more suitable in dm_bufio_prefetch which is called on different paths even though prefeching is disabled ?

Also note that you're currently leaking the work object with the early return.	

Right! I leaked this. always so. Thanks!!!

- Eric

Re: [PATCH] dm-verity: unnecessary data blocks that need not read hash blocks

[Eric Biggers]

On Mon, Dec 16, 2019 at 02:02:33AM +0000, xianrong.zhou(周先荣) wrote:
> hey Eric:
> 
> On Wed, Dec 11, 2019 at 11:32:40AM +0800, zhou xianrong wrote:
> > From: "xianrong.zhou" <xianrong.zhou@transsion.com>
> > 
> > If check_at_most_once enabled, just like verity work the prefetching 
> > work should check for data block bitmap firstly before reading hash 
> > block as well. Skip bit-set data blocks from both ends of data block 
> > range by testing the validated bitmap. This can reduce the amounts of 
> > data blocks which need to read hash blocks.
> > 
> > Launching 91 apps every 15s and repeat 21 rounds on Android Q.
> > In prefetching work we can let only 2602/360312 = 0.72% data blocks 
> > really need to read hash blocks.
> > 
> > But the reduced data blocks range would be enlarged again by 
> > dm_verity_prefetch_cluster later.
> > 
> > Signed-off-by: xianrong.zhou <xianrong.zhou@transsion.com>
> > Signed-off-by: yuanjiong.gao <yuanjiong.gao@transsion.com>
> > Tested-by: ruxian.feng <ruxian.feng@transsion.com>
> > ---
> >  drivers/md/dm-verity-target.c | 16 ++++++++++++++++
> >  1 file changed, 16 insertions(+)
> > 
> > diff --git a/drivers/md/dm-verity-target.c 
> > b/drivers/md/dm-verity-target.c index 4fb33e7562c5..7b8eb754c0b6 
> > 100644
> > --- a/drivers/md/dm-verity-target.c
> > +++ b/drivers/md/dm-verity-target.c
> > @@ -581,6 +581,22 @@ static void verity_prefetch_io(struct work_struct *work)
> >  	struct dm_verity *v = pw->v;
> >  	int i;
> >  
> > +	if (v->validated_blocks) {
> > +		while (pw->n_blocks) {
> > +			if (unlikely(!test_bit(pw->block, v->validated_blocks)))
> > +				break;
> > +			pw->block++;
> > +			pw->n_blocks--;
> > +		}
> > +		while (pw->n_blocks) {
> > +			if (unlikely(!test_bit(pw->block + pw->n_blocks - 1,
> > +				v->validated_blocks)))
> > +				break;
> > +			pw->n_blocks--;
> > +		}
> > +		if (!pw->n_blocks)
> > +			return;
> > +	}
> 
> This is a good idea, but shouldn't this logic go in verity_submit_prefetch()
> prior to the struct dm_verity_prefetch_work being allocated?  Then if no
> prefeching is needed, allocating and scheduling the work object can be
> skipped.
> 
> Eric, Do you mean it is more suitable in dm_bufio_prefetch which is called on
> different paths even though prefeching is disabled ?
> 

No, I'm talking about verity_submit_prefetch().  verity_submit_prefetch()
allocates and schedules a work object, which executes verity_prefetch_io().
If all data blocks in the I/O request were already validated, there's no need to
allocate and schedule the prefetch work.

- Eric

Re: [dm-devel] [PATCH] dm-verity: unnecessary data blocks that need not read hash blocks

[Eric Biggers]

On Mon, Dec 16, 2019 at 10:50:26AM -0800, Eric Biggers wrote:
> On Mon, Dec 16, 2019 at 02:02:33AM +0000, xianrong.zhou(周先荣) wrote:
> > hey Eric:
> > 
> > On Wed, Dec 11, 2019 at 11:32:40AM +0800, zhou xianrong wrote:
> > > From: "xianrong.zhou" <xianrong.zhou@transsion.com>
> > > 
> > > If check_at_most_once enabled, just like verity work the prefetching 
> > > work should check for data block bitmap firstly before reading hash 
> > > block as well. Skip bit-set data blocks from both ends of data block 
> > > range by testing the validated bitmap. This can reduce the amounts of 
> > > data blocks which need to read hash blocks.
> > > 
> > > Launching 91 apps every 15s and repeat 21 rounds on Android Q.
> > > In prefetching work we can let only 2602/360312 = 0.72% data blocks 
> > > really need to read hash blocks.
> > > 
> > > But the reduced data blocks range would be enlarged again by 
> > > dm_verity_prefetch_cluster later.
> > > 
> > > Signed-off-by: xianrong.zhou <xianrong.zhou@transsion.com>
> > > Signed-off-by: yuanjiong.gao <yuanjiong.gao@transsion.com>
> > > Tested-by: ruxian.feng <ruxian.feng@transsion.com>
> > > ---
> > >  drivers/md/dm-verity-target.c | 16 ++++++++++++++++
> > >  1 file changed, 16 insertions(+)
> > > 
> > > diff --git a/drivers/md/dm-verity-target.c 
> > > b/drivers/md/dm-verity-target.c index 4fb33e7562c5..7b8eb754c0b6 
> > > 100644
> > > --- a/drivers/md/dm-verity-target.c
> > > +++ b/drivers/md/dm-verity-target.c
> > > @@ -581,6 +581,22 @@ static void verity_prefetch_io(struct work_struct *work)
> > >  	struct dm_verity *v = pw->v;
> > >  	int i;
> > >  
> > > +	if (v->validated_blocks) {
> > > +		while (pw->n_blocks) {
> > > +			if (unlikely(!test_bit(pw->block, v->validated_blocks)))
> > > +				break;
> > > +			pw->block++;
> > > +			pw->n_blocks--;
> > > +		}
> > > +		while (pw->n_blocks) {
> > > +			if (unlikely(!test_bit(pw->block + pw->n_blocks - 1,
> > > +				v->validated_blocks)))
> > > +				break;
> > > +			pw->n_blocks--;
> > > +		}
> > > +		if (!pw->n_blocks)
> > > +			return;
> > > +	}
> > 
> > This is a good idea, but shouldn't this logic go in verity_submit_prefetch()
> > prior to the struct dm_verity_prefetch_work being allocated?  Then if no
> > prefeching is needed, allocating and scheduling the work object can be
> > skipped.
> > 
> > Eric, Do you mean it is more suitable in dm_bufio_prefetch which is called on
> > different paths even though prefeching is disabled ?
> > 
> 
> No, I'm talking about verity_submit_prefetch().  verity_submit_prefetch()
> allocates and schedules a work object, which executes verity_prefetch_io().
> If all data blocks in the I/O request were already validated, there's no need to
> allocate and schedule the prefetch work.
> 
> - Eric
> 

Are you planning to send a new version of this patch?

- Eric