Date:   Wed, 13 Mar 2019 13:59:16 +0100
From:   Stanislaw Gruszka <sgruszka@redhat.com>
To:     Lorenzo Bianconi <lorenzo.bianconi@redhat.com>
Cc:     linux-wireless@vger.kernel.org, Felix Fietkau <nbd@nbd.name>
Subject: Re: [RFC 11/12] mt76usb: allocate urb and sg as linear data
Message-ID: <20190313125916.GC2003@redhat.com>
References: <1552403166-3821-1-git-send-email-sgruszka@redhat.com>
 <1552403166-3821-12-git-send-email-sgruszka@redhat.com>
 <20190312163431.GF2611@localhost.localdomain>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20190312163431.GF2611@localhost.localdomain>
User-Agent: Mutt/1.5.21 (2010-09-15)
Sender: linux-wireless-owner@vger.kernel.org
Precedence: bulk

On Tue, Mar 12, 2019 at 05:34:32PM +0100, Lorenzo Bianconi wrote:
> > Alloc sg table at the end of urb structure. This will increase
> > cache usage.
> > 
> 
> I am curious, have you observed any performance improvement doing so?

It's hard to measure that. Stressing net transfer with about 175 Mbit/s
by netperf (what is max I can get) only took about 15% of CPU for me
and net performance fluctuate so CPU pref stats vary too.

> > -	if (dev->usb.sg_en) {
> > -		urb->sg = devm_kcalloc(dev->dev, MT_SG_MAX_SIZE,
> > -				       sizeof(urb->sg), GFP_KERNEL);
FTR: Here is bug it should be sizeof(*urb->sg).

> > -		if (!urb->sg)
> > -			return -ENOMEM;
> > -	}
> > +	usb_init_urb(e->urb);
> > +
> > +	if (dev->usb.sg_en)
> > +		e->urb->sg = (struct scatterlist *)((u8 *)e->urb + sizeof(struct urb));
> 
> You can avoid u8 cast doing:
> (struct scatterlist *)(e->urb + 1)
Cool trick!

Stanislaw