lib_ethernet and ET_LOAD_STORE

[CousinItt]

Hi all,

I'm having problems with random ET_LOAD_STORE faults in lib_ethernet, the 3 Aug 18 rev (3.4.0rc2 I think) from Github. Having searched for similar problems here, I've tried tinkering with the build options for the library.

Dropping the optimization level to 2 in the following line seems to fix it.

Code: Select all

MODULE_XCC_FLAGS = $(XCC_FLAGS) -g -O3 -mno-dual-issue

The following line I've left untouched.

Code: Select all

XCC_FLAGS_mii_master.xc = $(XCC_FLAGS) -O3 -fschedule -g0 -mno-dual-issue

Is anyone aware of any problems in making this change?

By the way I'm fairly sure this isn't a hardware problem. It's an XMOS eXplorerKIT and I've checked the supply voltages.

[edit] Also I'm still using xTIMEcomposer 14.3.3 on Windows.

Thanks

[akp]

What xMII mode are you using? RGMII, MII real time, or MII lite?

[CousinItt]

The code for the ethernet tile follows the example in AN00199 fairly closely, using RGMII.

BTW the example uses the -Os switch for the application code on that tile (similar to below), whereas I'm using -O2.

Code: Select all

void tile_9_tasks
(
   server ethernet_rx_if i_rx[n_eth],
   server ethernet_tx_if i_tx[n_eth],
   static const unsigned n_eth,
   server eth_cfg_proxy_if prif
)
{
   ethernet_cfg_if i_cfg[NUM_CFG_CLIENTS];
   streaming chan c_rgmii_cfg;
   smi_if i_smi;

   par
   {
      rgmii_ethernet_mac(i_rx, n_eth, i_tx, n_eth, null, null, c_rgmii_cfg, rgmii_ports, ETHERNET_DISABLE_SHAPER);
      smi(i_smi, p_smi_mdio, p_smi_mdc);
      [[combine]]
      par
      {
         rgmii_ethernet_mac_config(i_cfg, NUM_CFG_CLIENTS, c_rgmii_cfg);
         ar8035_phy_driver(i_smi, i_cfg[CFG_TO_PHY_DRIVER]);
         eth_config_proxy(i_cfg[CFG_TO_IP], prif);
      }
   }
}

[akp]

I can't help but wonder if it's in your ethernet config proxy? Or is that proven elsewhere? I suppose it would be interesting to try running without SMI (and force the link up in software). You have lots of the same interface stuff running on the same core.

[CousinItt]

Hi akp,

I don't think it's the proxy. It's fairly innocuous and inactive once initial configuration is complete. Code is included below. The only reason it's there is to work around a problem with xc. See https://www.xcore.com/viewtopic.php?p=36862

I don't know enough about SMI to replace it at the moment, and I'm not sure it's worth the investment if the -O2 switch works. Please see the code and app note for the AN00199 example. This has a very similar setup, apart from using direct core assignment for the combinable rgmii_ethernet_mac_config and ar8035_phy_driver in main, whereas I use [[combine]] with a nested par and include the combinable proxy with them.

Any other thoughts? Thanks for your input so far.

Code: Select all

[[combinable]]
void eth_config_proxy
(
   client ethernet_cfg_if cfg_if,
   server eth_cfg_proxy_if prif
)
{
   // just echo particular requests...
   while(1)
   {
      select
      {
         case prif.set_macaddr(size_t ifnum, uint8_t mac_address[MACADDR_NUM_BYTES]):

            uint8_t macad[MACADDR_NUM_BYTES];

            for(unsigned i; i < MACADDR_NUM_BYTES; i++)
            {
               macad[i] = mac_address[i];
            }

            cfg_if.set_macaddr(ifnum, macad);
            break;

         case prif.add_macaddr_filter(size_t client_num, int is_hp, ethernet_macaddr_filter_t entry)
               -> ethernet_macaddr_filter_result_t result:
            result = cfg_if.add_macaddr_filter(client_num, is_hp, entry);
            break;

         case prif.add_ethertype_filter(size_t client_num, uint16_t ethertype):
            cfg_if.add_ethertype_filter(client_num, ethertype);
            break;
      }
   }
}

[akp]

Yeah I suppose I didn't read your post, I thought you were still having problems with the exception at -O2. There shouldn't be any big trouble using -O2 I wouldn't think. The high speed stuff is all assembly anyway.

[CousinItt]

OK thanks. When I have a bit of time I might also check the effect of using the original switches for the library but adding the no-dual-issue switch for the non-library code on that tile.