Re: [OpenOCD-devel] Improving cmsis-dap performance

[Andreas F. <and...@gm...>]

On Mon, Jul 14, 2014 at 4:47 AM, Angus Gratton <gu...@pr...> wrote:

> Hi Andreas,
>
> Thanks for clarifying a great many things.
>
> I did a bit more reading of the source. I hadn't realised that both
> JTAG and SWD DAP ops go via dap_ops -> cmsis_dap_transport ->
> jtag_interface.swd -> cmsis_dap_swd_driver to generate the USBHID
> commands. I actually thought JTAG/CMSIS-DAP hadn't been implemented in
> openocd yet, my mistake.
>

I'm sorry if I confused you but what I meant was that DAP operation go
through *either* adi_v5_jtag.c, adi_v5_swd.c or adi_v5_cmsis_dap.c. The
JTAG and SWD implementations are only used when another type of adapter
driver is selected, such a ftdi, jlink, vsllink etc. When a CMSIS-DAP
adapter is used, DAP ops from the target *always* go through
adi_v5_cmsis_dap.c (because of "transport select cmsis-dap") and from there
into the jtag_interface->swd which is initialized by the interface
infrastructure (because of "interface cmsis-dap") to point to the
jtag/drivers/cmsis_dap_usb.c swd_driver. The jtag_interface in the same
file is necessary as a container for the swd_driver and for providing an
initialization point and functions to change clock speed. This is the most
problematic part of the OpenOCD internal architecture today, as I see it.

The CMSIS-DAP implementation in OpenOCD still does not support DAP
operations over JTAG, but that's a matter of initialization. The adapter
needs to be told to use JTAG and the structure of the JTAG chain (which is
going to be the messy part). The DAP ops shall not take another route
inside OpenOCD just because the adapter has been told to use JTAG instead
of SWD.


> Now I see cmsis_dap_swd_driver is used here not as a way of
> implementing SWD operations, but just a way of implementating DAP
> operations in a way that the DAP transport can access them.
>
> Does that sound right?
>

I think so, yes.


> Exploring possibilities, I think we could remove the
> cmsis_dap_swd_driver instance and move the read_red/write_reg stuff up
> into the transport. However there's got to be a convenient interface
> between them:
>
> - If we do the actual USB operations as part of the jtag_interface, we
>   need a way for the transport to queue generic CMSIS-DAP commands on
>   the jtag_interface. Maybe a custom command that's just raw CMSIS-DAP
>   bytes? Or a dap_driver to the jtag_interface, similar to swd_driver?
>
> - Maybe it would be possible to do the USB operations on the transport
>   side? This doesn't gel with the description of a transport as
>   relating only to the on-wire signalling not the messaging
>   protcol. To work the jtag_interface would need a way to queue
>   CMSIS-DAP commands on the transport for custom commands,
>   adapter_khz, etc.
>
> - Or alternatively we leave it as it is, but maybe label the
>   cmsis_dap_swd_driver instance more clearly to make it clearer that
>   it's not an SWD-only interface, but just a convenient interface for
>   any DAP register operations?
>
>
> I'd be happy to put time towards exploring any of these though (or
> some fourth choice) if it's helpful?
>
>
Just to provide further food for thought, here are some of the issues that
need to be considered at some point.

The CMSIS-DAP protocol can handle raw JTAG operations (clock this many bits
in and/or out while holding TMS at that level) and we'll at some point want
to use that to enable debugging non-Cortex parts, maybe even on the same
JTAG chain and at the same time as a DAP based part.

It may be necessary to clock out special sequences to switch between JTAG
and SWD, especially to support Dormant operation and multidrop SWD. API for
that was added at the swd_driver layer, not at the dap_ops layer. How do we
handle that if we ditch the swd_driver for CMSIS-DAP?

Struct jtag_interface is currently overloaded with non-JTAG related
functions, such as speed setting functions, reset signals, as a holder of
struct swd_driver and moreover it can only exist in one global instance.
I'd like to minimize struct jtag_interface and keep it (for now) as a pure
JTAG API. The auxiliary functionality should IMO be transferred to a new
"struct adapter" with a separate lifetime. Using operations on that adapter
instance, you could open sessions to access a struct jtag_interface for
JTAG operations, or a struct swd_driver for low-level SWD operations, a
struct dap_ops for high-level DAP operations and so on. New APIs could be
added for new debug protocols or special adapter features. A pure SWD-only
adapter should not need to implement a struct jtag_interface.

In JTAG mode, the CMSIS-DAP adapter needs to be told the structure of the
JTAG chain. In that use case, the TAPs should probably be declared/auto
detected as normal, which means a full JTAG API needs to be implemented
over the DAP_JTAG_Sequence command. TAPs that do not connect to a DAP, such
as old ARMs, CPLDs or whatever that may be in the chain, should be accessed
over the JTAG API but DAP parts should be able to be accessed through the
dap_ops API. Note that there may be several DAP targets on a chain so we
must be able to request several struct dap_ops from the adapter.

/Andreas