illumos gate

See bug 9601 for the original fix.
After that, we got some new code via bug 12933:

+       /* OCP cards case: if a mezz is removed the ethernet port is at
+        * disabled state in PRTGEN_CNF register. Additional NVM read is
+        * needed in order to check if we are dealing with OCP card.
+        * Those cards have 4 PFs at minimum, so using PRTGEN_CNF for counting
+        * physical ports results in wrong partition id calculation and thus
+        * not supporting WoL.
+        */
+       if (hw->mac.type == I40E_MAC_X722) {
+               if (i40e_acquire_nvm(hw, I40E_RESOURCE_READ) == I40E_SUCCESS) {
+                       status = i40e_aq_read_nvm(hw, I40E_SR_EMP_MODULE_PTR,
+                                                 2 * I40E_SR_OCP_CFG_WORD0,
+                                                 sizeof(ocp_cfg_word0),
+                                                 &ocp_cfg_word0, TRUE, NULL);
+                       if (status == I40E_SUCCESS &&
+                           (ocp_cfg_word0 & I40E_SR_OCP_ENABLED))
+                               hw->num_ports = 4;
+                       i40e_release_nvm(hw);
+               }
+       }
+

The assertion here that PF is at least 4 is clearly not the case as shown
in the above bug, and on my device 0x8086:0x37d2

num_functions is 2, so we end up with:

3951                 hw->num_partitions = num_functions / hw->num_ports;              

num_partitions as zero, and we go boom.
What's perplexing here is that the upstream code has the same code: it
forces OCP X722 to have 4 ports.

The above comment is talking about the PRTGEN_CNF register, but we
no longer use that, preferring the status one, as described in bug 9601.

I don't have a spec with any X722 or OCP details, but it seems like we might
have a chance at fixing this just by disabling that new code again.