protocol threads mpsafe mode

Sepherosa Ziehau sepherosa at gmail.com
Mon Sep 22 06:37:01 PDT 2008


On Mon, Sep 22, 2008 at 3:18 AM, Matthew Dillon
<dillon at apollo.backplane.com> wrote:
>    I think we can go with something similar, allowing you to start
>    committing the work without interfering with stability on HEAD.
>    How about a few cleanups though.  For testing purposes we don't
>    really care about another few nanoseconds so:
>
>    * Always start the protocol threads TDF_MPSAFE.  LWKT is MP-SAFE so
>      there is no need to conditionalize it.  The msgport code is also
>      MP-SAFE, so we really only have to worry about the function dispatch.
>
>    * You might as well not bother with two different netmsg_service_loop
>      threads.  Just have one, acquire and release the mp lock based on
>      NETISR_FLAG_MPSAFE and netisr_mpsafe_thread.
>
>
>    Here's what I'm thinking.  This is untested but note that the critical
>    path is handled nicely.

Looks good, this also gives us the ability to change mpsafe mode
dynamicly on a running system, see the new patch.

>
>    --
>
>    netisr_run():   I'm not sure you need any MP tests at all in this
>    code, won't it already be properly held or not held due already having
>    been resolved by the original dispatch?  This routine is only called
>    on re-dispatch.

Nah, it is called on the critical path:
NIC intr mport and sends msg to proto threads, the msg's handler is
ether_input_oncpu() which calls netisr_run().

ether_input_oncpu() has three major functionality:
1) bridge input, which is mpsafe
2) fastforward, which is mpsafe
3) call NETISR handler (netisr_run)

With few changes, I could make major parts of ether_input_oncpu()
mpsafe, thus to turn on MSGF_MPSAFE in the msg sent by NIC intr.  Once
ether_input_oncpu() runs without BGL, then the mpsafe testing in
netisr_run() is necessary.

>
>    netisr_queue(): Would check the netisr flag and set or clear TDF_MPSAFE
>    appropriately.
>
>    schednetisr(): Would check the netisr flag and set or clear TDF_MPSAFE
>    appropriately.

Ah, I forgot these two :), added in the new patch.

New patch is here, please review:
http://leaf.dragonflybsd.org/~sephe/proto.diff2

Best Regards,
sephe

-- 
Live Free or Die





More information about the Kernel mailing list