Intel10G driver: multi-iovec and jumbo frames #161
Conversation
|
Great! Requests: Could you add more detail to the pull request so that it can serve as release notes? e.g. title "Intel10G: Added support for multi-iovec packets and jumbo frames" and text saying if any configuration is needed etc? Reading the code, the main change I see is adding support for stop()ing VFs. That is not mentioned in the pull request. Can you split that out? |
|
is it better now? |
|
Looks great! How about the Travis-CI failing test? |
|
It would also be nice to have the detailed description as the commit message so that we see it in "Git log". It seems like a bother to have to duplicate the text for both Commit and Pull Request. I wonder how people usually deal with this. |
This handles some issues on "bigger than trivial" packets. - sets the maximum frame size (register MAXFRS) to 9000+18 bytes. It could be nice to use a configurable option, but it's a chip-wide setting, so it's likely better to simply set to the max. - fixes the too-many-releasing bug by incrementing the reference counter of the packet for each descriptor in the transmit queue. This makes it possible to send a multi-iovec packet without coalescing it first. - allows the NIC to split packets that are bigger than a single buffer (currently 4K). Both `:can_receive()` and `:receive()` have tight loops to check for an EOP descriptor after some ready-but-partial descriptors. This could be manually unrolled (just 3 iterations with current defaults) if benchmarks prove it to be a performance regression. - implements the `:stop()` method for the app, in the VF case, it calls the PF if it's the last VF. This is necessary to allow changing configurations. (seem unrelated, but some settings (frame size in this case) persist even across process invocations, so it's impossible to properly test it without `:stop()`)
|
One more request: Could you update the title to say that these changes relate to the Intel 10G driver? |
javierguerragiraldez
changed the title
|
just extended the commit message. also fixed the failing Travis CI (it was failing at sending the 'skip' result, duh!), but now it failed on the packet filter app. :-( |
Intel10G driver: multi-iovec and jumbo frames
This handles some issues on "bigger than trivial" packets.
:can_receive()and:receive()have tight loops to check for an EOP descriptor after some ready-but-partial descriptors. This could be manually unrolled (just 3 iterations with current defaults) if benchmarks prove it to be a performance regression.:stop()method for the app, in the VF case, it calls the PF if it's the last VF. This is necessary to allow changing configurations. (seem unrelated, but some settings (frame size in this case) persist even across process invocations, so it's impossible to properly test it without:stop())