Differences

This shows you the differences between two versions of the page.

--- devsummit:201708:iflib [2017/09/08 14:47] – SeanBruno
+++ devsummit:201708:iflib [2017/09/09 13:35] (current) – SeanBruno
@@ Line 5: / Line 5: @@
   * gnn notes that one could just deploy the Bruno-lab to Sentex so that others can use it
   * bhyve has a hardware model that uses an emulated em(4).  This has not been tested.  TBD
+  * em(4) leaking MSI-X vectors on detach
+    * devctl disable em0
+    * em0: detached
+    * pci0:0:25:0: Device leaked MSI vectors
+    * em(4) doesn't appear to be calling pci_release_msi() during detach path
 ixgbe(4) -- https://reviews.freebsd.org/D11727
@@ Line 40: / Line 45: @@
 IFLIB per device queue and descriptor tuning
   * Request to make number of queues and descriptors per device not per driver.  Users may want to configure 1 queue on em0 but use 2 queues on em1.
+    * dev.*.*.override_n([tr]xq|[tr]xd)s is per device, not per driver.  This should work today.
 IFLIB per device MSI-X settings
   * Request to make MSI-X settings per device.
+    * What MSI-X settings?  iflib doesn't have any, but em does have a per-device MSI-X setting (dev.em.0.iflib.disable_msix)
 IFLIB, gtaskqueue and mp_ring design docs?
@@ Line 53: / Line 60: @@
 IFLIB in a NUMA world
   * Is it possible for IFLIB to query and configure queues and tasks in a NUMA friendly way.
+    * It should be only allocating tasks and queues to the socket the device is on today.
+    * Is there a more NUMA friendly thing we should be doing?
   * Allow overrides of the NUMA configuration via loader.conf or other acceptible configurations.
+    * Would a cpuset tunable be enough, or is more fine-grained control desired?
 Not IFLIB but relevant discussion