On 12/16/24 20:23, Gerhard Engleder wrote:
@@ -331,8 +331,15 @@ void e1000e_update_mc_addr_list_generic(struct
e1000_hw *hw,
}
/* replace the entire MTA table */
- for (i = hw->mac.mta_reg_count - 1; i >= 0; i--)
+ for (i = hw->mac.mta_reg_count - 1; i >= 0; i--) {
E1000_WRITE_REG_ARRAY(hw, E1000_MTA, i, hw-
>mac.mta_shadow[i]);
+
+ /* do not queue up too many posted writes to prevent increased
+ * latency for other devices on the interconnect
+ */
+ if ((i % 8) == 0 && i != 0)
+ e1e_flush();
I would prefer to avoid adding this code to all devices, particularly
those that don't operate on real-time systems. Implementing this code
will introduce three additional MMIO transactions which will increase
the driver start time in various flows (up, probe, etc.).
Is there a specific reason not to use if
(IS_ENABLED(CONFIG_PREEMPT_RT)) as Andrew initially suggested?
Andrew made two suggestions: IS_ENABLED(CONFIG_PREEMPT_RT) which I used
in the first version after the RFC. And he suggested to check for a
compromise between RT and none RT performance, as some distros might
enable PREEMPT_RT in the future.
Przemek suggested to remove the PREEMPT_RT check as "this change sounds
reasonable also for the standard kernel" after the first version with
IS_ENABLED(CONFIG_PREEMPT_RT).
I used the PREEMPT_RT dependency to limit effects to real-time systems,
to not make none real-time systems slower. But I could also follow the
reasoning of Andrew and Przemek. With that said, I have no problem to
add IS_ENABLED(CONFIG_PREEMPT_RT) again.
Gerhard
I'm also fine with limiting the change to RT kernels.
