Re: [PATCH] misc: fastrpc: copy to user only for non-DMA-BUF heap buffers

[jeyr@codeaurora.org]

On 2021-09-23 23:13, Greg KH wrote:
> On Thu, Sep 23, 2021 at 03:54:15PM +0530, jeyr@codeaurora.org wrote:
>> On 2021-09-23 14:19, Greg KH wrote:
>> > On Thu, Sep 23, 2021 at 02:07:52PM +0530, Jeya R wrote:
>> > > fastrpc_put_args is copying all the output buffers to user. For large
>> > > number of output context buffers, this might cause performance
>> > > degradation. Copying is not needed for DMA-BUF heap buffers.
>> >
>> > What does "performance degradation" really mean?
>> 
>> Unnecessary copying for large number of buffers would cause some
>> additional time which would get added to overall fastrpc call cost.
>> 
>> >
>> > >
>> > > Signed-off-by: Jeya R <jeyr@codeaurora.org>
>> > > ---
>> > >  drivers/misc/fastrpc.c | 18 ++++++++++--------
>> > >  1 file changed, 10 insertions(+), 8 deletions(-)
>> > >
>> > > diff --git a/drivers/misc/fastrpc.c b/drivers/misc/fastrpc.c
>> > > index beda610..536eabf 100644
>> > > --- a/drivers/misc/fastrpc.c
>> > > +++ b/drivers/misc/fastrpc.c
>> > > @@ -890,15 +890,17 @@ static int fastrpc_put_args(struct
>> > > fastrpc_invoke_ctx *ctx,
>> > >  	inbufs = REMOTE_SCALARS_INBUFS(ctx->sc);
>> > >
>> > >  	for (i = inbufs; i < ctx->nbufs; ++i) {
>> > > -		void *src = (void *)(uintptr_t)rpra[i].pv;
>> > > -		void *dst = (void *)(uintptr_t)ctx->args[i].ptr;
>> > > -		u64 len = rpra[i].len;
>> > > +		if (!ctx->maps[i]) {
>> > > +			void *src = (void *)(uintptr_t)rpra[i].pv;
>> > > +			void *dst = (void *)(uintptr_t)ctx->args[i].ptr;
>> >
>> > uintptr_t is not a kernel variable type.  Please use the real kernel
>> > type for this as you are touching these lines.
>> >
>> 
>> Sure, thanks for pointing this. Will update this in the next patch.
>> 
>> > > +			u64 len = rpra[i].len;
>> > >
>> > > -		if (!kernel) {
>> > > -			if (copy_to_user((void __user *)dst, src, len))
>> > > -				return -EFAULT;
>> > > -		} else {
>> > > -			memcpy(dst, src, len);
>> > > +			if (!kernel) {
>> > > +				if (copy_to_user((void __user *)dst, src, len))
>> > > +					return -EFAULT;
>> > > +			} else {
>> > > +				memcpy(dst, src, len);
>> > > +			}
>> >
>> > So you were copying buffers that didn't need to be copied?  So you are
>> > now doing less work?  Or is this fixing a bug where you were copying
>> > things that you should not have been copying?
>> >
>> > What commit does this fix?  Does this need to go to the stable kernel
>> > trees?
>> >
>> 
>> Yes, not all buffer needs to be copied. This change would avoid 
>> unnecessary
>> copying of buffers. Not adding fix tag as it's not exactly fixing any 
>> bug.
>> This should go to stable kernel trees.
> 
> If it's not fixing a bug, why should it go into the stable trees?

This is not a bug fix, it can be considered as an enhancement and it can 
go in
to new release.