Subject: Re: [PATCH v4 19/23] drivers/fsi: Add GPIO based FSI master
From: Benjamin Herrenschmidt <benh@au1.ibm.com>
Reply-To: benh@au1.ibm.com
To: Christopher Bostic <cbostic@linux.vnet.ibm.com>,
        Joel Stanley <joel@jms.id.au>
Cc: Rob Herring <robh+dt@kernel.org>, Mark Rutland <mark.rutland@arm.com>,
        Russell King <linux@armlinux.org.uk>, rostedt@goodmis.org,
        mingo@redhat.com, Greg KH <gregkh@linuxfoundation.org>,
        devicetree@vger.kernel.org, linux-arm-kernel@lists.infradead.org,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Andrew Jeffery <andrew@aj.id.au>,
        Alistair Popple <alistair@popple.id.au>,
        "Edward A . James" <eajames@us.ibm.com>, Jeremy Kerr <jk@ozlabs.org>
Date: Mon, 10 Apr 2017 07:53:55 +1000
In-Reply-To: <1491774092.4166.195.camel@kernel.crashing.org>
References: <20170329174340.89109-1-cbostic@linux.vnet.ibm.com>
         <20170329174340.89109-20-cbostic@linux.vnet.ibm.com>
         <CACPK8XfOSYPidHF6=psC4mP-tkRyWTu0psdcK6EwA2E6FF53Uw@mail.gmail.com>
         <0e1bcf3a-e8d7-9f50-bdf7-2a1e7466665b@linux.vnet.ibm.com>
         <1490907014.3177.207.camel@kernel.crashing.org>
         <93b21624-11fc-b71b-aa78-6cb4371c87ae@linux.vnet.ibm.com>
         <1491344360.4166.68.camel@kernel.crashing.org>
         <5344146a-f450-7959-4688-a83ee022574b@linux.vnet.ibm.com>
         <1491774092.4166.195.camel@kernel.crashing.org>
Organization: IBM Australia
Content-Type: text/plain; charset="UTF-8"
Mime-Version: 1.0
Content-Transfer-Encoding: 8bit
Message-Id: <1491774835.4166.197.camel@au1.ibm.com>
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1504
Lines: 37

On Mon, 2017-04-10 at 07:41 +1000, Benjamin Herrenschmidt wrote:
> On Sun, 2017-04-09 at 16:22 -0500, Christopher Bostic wrote:
> > A 3 microsecond delay is required, however, to prevent occasional
> > issues 
> > during heavy FSI bus load stress testing.
> > A 1 nanosecond delay using ndelay(1) had been specified prior to
> > this 
> > but after looking more closely at real time performance it turned
> > out to 
> > actually be roughly 1-2 microseconds.   This appears to be the
> > minimum 
> > resolution using the delay() linux libraries on the
> > AST2400/2500.   
> > Given this, increasing delay to 3 microseconds doesn't impact 
> > performance much considering I can now remove the sample input
> > delay 
> > based on your recommendations to re-order the two clock delays.
> 
> This is huge delays. We should consider a AST2xxx specific variant of
> the backend that uses nops or similar lab-calibrated constructs
> instead. Otherwise we are stuck in the kHz range, this is a >200Mhz
> bus
> :)
> 
> I don't understand why 3us delay would thus be necessary.
> 
> Where about did you observe issues ? Could it be that you don't wait
> long enough in the transitions from write to read ?

FYI. pdbg in userspace operates without any  delays in practice, the
overhead between the various load/store instructions seems sufficient.

The only delay that's needed is when going through the FSI2PIB (to do
SCOMs) where it seems like back-2-back accesses can be problematic.

Cheers,
Ben