Support for AXI interconnect targetting the XILINX ZC706

[Misra, Avinash]

Chris,

This is an excellent response and this is information that I certainly wish I had when I started using RTEMS for my Zynq application. 

Tiago, another "gotcha" I want to point out is when launching your target application via Xilinx SDK and using RTEMS with SMP Mode (setting # of processors > 1). When you program and launch your application to CPU0 on the Zynq via Xilinx SDK, Xilinx will place CPU1 in Idle Mode. RTEMS will be stuck in a perpetual wait state waiting for CPU1 to come up before launching. You will need to physical click start on CPU1 in SDK in order for CPU1 to resume execution and allow RTEMS to exit its wait barrier. If you program your image to a SD Card or the onboard flash and use FSBL then you will not need to do this.

Chris -- Does RTEMS have a wiki where we can maybe compile a list of Gotchas/FAQs/Things to be aware of when programming for the Zynq and other platforms? 

Thanks,
Avi
-----Original Message-----
From: devel <devel-bounces at rtems.org> On Behalf Of Chris Johns
Sent: Tuesday, November 5, 2019 6:32 PM
To: Tiago Manuel Da Silva Jorge <tmdasilva at gmv.com>; devel at rtems.org
Subject: Re: Support for AXI interconnect targetting the XILINX ZC706

On 5/11/19 11:43 pm, Tiago Manuel Da Silva Jorge wrote:
> We are working on an interesting project where we are developing 
> applications that should run with RTEMS on ARM and additionally 
> communicate with Programmable Logic (FPGA). We are using the Xilinx 
> Zynq-7000 SoC ZC706 Evaluation Kit, and for communication between its 
> PS (Processing System: ARM Cortex-A9) and PL (Programmable Logic: Artix-7 FPGA) we are planning to use its AXI Interconnect.

Welcome to RTEMS and thank for the introduction. It sounds like an interesting project.

> Hence, the question is: Is there any support (e.g. driver or the like) 
> for this AXI on RTEMS/ARM (to communicate with PL side)?

RTEMS provides all the primitives you need:

1. MMU and cache support
2. Interrupts

After this there is no specific support for the AXI bus and that is mostly due to the wide range of ways you can implement the transfer of data across the AXI bus to and from the PL. The factors that effect the design are the amount of data and the performance needed. An other factor is the experience of the PL design team, if they are new to the Zynq or Xilinx there may be a learning curve here.

For each part of the PL logic you interact with over the AXI bus you will need a set of registers assigned to one of the AXI bus ports. The PL team will do this and set the base address of the port. You need to configure the MMU so you can access that address space and so the PL. To set up the MMU and cache create a file in your application, say mmu.c, and add a table similar to:

---- 8< -----
#define ARM_CP15_TEXT_SECTION BSP_START_TEXT_SECTION

#include <bsp.h>
#include <bsp/start.h>
#include <bsp/arm-gic-irq.h>
#include <bsp/arm-cp15-start.h>
#include <bsp/arm-a9mpcore-start.h>

#ifdef ARMV7_CP15_START_DEFAULT_SECTIONS

BSP_START_DATA_SECTION static const arm_cp15_start_section_config zynq_mmu_config_table[] = {
  ARMV7_CP15_START_DEFAULT_SECTIONS,
  {
    .begin = 0xe0000000U,
    .end   = 0xe0200000U,
    .flags = ARMV7_MMU_DEVICE
  }, {
    .begin = 0xf8000000U,
    .end   = 0xf9000000U,
    .flags = ARMV7_MMU_DEVICE
  }, {
    .begin = 0x40000000U,
    .end   = 0xc0000000U,
    .flags = ARMV7_MMU_DEVICE
  }, {
    .begin = 0x00100000U,
    .end   = 0x00400000U,
    .flags = ARMV7_MMU_DEVICE
  }, {
    .begin = 0xfffc0000u,
    .end   = 0xffffffffu,
    .flags = ARMV7_MMU_DEVICE
  }
};

BSP_START_TEXT_SECTION void zynq_setup_mmu_and_cache(void) {
  uint32_t ctrl = arm_cp15_start_setup_mmu_and_cache(
    ARM_CP15_CTRL_A,
    ARM_CP15_CTRL_AFE | ARM_CP15_CTRL_Z
  );

  arm_cp15_start_setup_translation_table_and_enable_mmu_and_cache(
    ctrl,
    (uint32_t *) bsp_translation_table_base,
    ARM_MMU_DEFAULT_CLIENT_DOMAIN,
    &zynq_mmu_config_table[0],
    RTEMS_ARRAY_SIZE(zynq_mmu_config_table)
  );
}

#endif
---- 8< -----

The function `zynq_setup_mmu_and_cache()` is weak in the BSP and so this function overrides the default.

You can now access the PL over the AXI bus. From here you can use CDMA as a simple but slow method to transfer blocks of data, or the PL designers can bus master the AXI bus and write directly into the main DDR (fast but more complex).
The software would configure the DDR target address via the PL registers. DMA and bus master writes and reads require you manage the cache invalidates and flushes. There are calls in RTEMS to handle this for you.

I hope this helps.

Chris
_______________________________________________
devel mailing list
devel at rtems.org
http://lists.rtems.org/mailman/listinfo/devel