Repository navigation
fix(netfront): fix queue selection - #125
Conversation
39124dd to
395f4bb
Compare
|
Backport to Waiting: earlier Edera commits on
This is retried when a backport lands on |
??? this is completely wrong. I think the automation broke. looking. EDIT: it did #127 |
* xen/netfront: netfront persistent grants (#119) * xen/netif: add the static grant control messages * xen/netfront: add control ring support * xen/netfront: register static grant pools with the backend * xen/netfront: use static grants on the receive path * xen/netfront: use static grants on the transmit path * fix(netfront): fix queue selection (#125)
Back in the previous repo, a patch edera-dev/linux-kernel-oci#182 added XPS-aware netfront queue placement, but that, as far as I can tell, was deadcode and never worked, because it left
ndo_select_queue = xennet_select_queue, so TX selection was never intelligently placed, so we got zero queue placement benefits from it, and it fact it prevented us falling through to a more intelligent placement mechanism.This is a fixup to that change so queue placement actually falls through correctly to optimized placement, which should help TCP throughput consistency a good deal, since we aren't (incorrectly) relying on random placement anymore.
Details:
netdev_core_pick_tx()callsndo_select_queuewhen a driver has one, and onlynetdev_pick_tx()actually consults XPS.So,
xennet_select_queue()ignored the per-node XPS map installed insetup_netfront(), and TX selection stayed as ye olde randomskb_get_hash() % num_queueson all nodes.Deleting the driver hook causes fallthrough to the existing logic where the sender selects queues with node-local rings, which is what we want. We could use a more complex heuristic here, but this alone is better than what we ship.