The branch main has been updated by kbowling: URL: https://cgit.FreeBSD.org/src/commit/?id=2cf580c694f6f392531a63f01c3fb89c0244f89a
commit 2cf580c694f6f392531a63f01c3fb89c0244f89a Author: Kevin Bowling <[email protected]> AuthorDate: 2026-08-06 11:03:28 +0000 Commit: Kevin Bowling <[email protected]> CommitDate: 2026-08-07 07:24:57 +0000 iflib: Permit SR-IOV configuration on a down interface Drivers which remap PF queues need a stop/mutate/restart transaction only when the interface has live queues. Permit their IOV initialization callback while the interface is administratively down and leave it down afterward. This restores the standard boot-time iovctl.conf workflow for igb and lets other opt-in drivers configure VFs before netif brings the PF up. MFC after: 1 week --- share/man/man4/em.4 | 8 +++++--- sys/dev/e1000/if_igb_iov.c | 4 ++++ sys/net/iflib.c | 22 +++++++++++----------- 3 files changed, 20 insertions(+), 14 deletions(-) diff --git a/share/man/man4/em.4 b/share/man/man4/em.4 index 328e1e77b18b..22898e894b00 100644 --- a/share/man/man4/em.4 +++ b/share/man/man4/em.4 @@ -32,7 +32,7 @@ .\" .\" * Other names and brands may be claimed as the property of others. .\" -.Dd July 29, 2026 +.Dd August 7, 2026 .Dt EM 4 .Os .Sh NAME @@ -188,8 +188,10 @@ The supported SR-IOV configuration uses one PF transmit and receive queue and requires MSI-X. I350 requires this layout; 82576 hardware can combine VMDq with per-pool RSS, but that mode is not implemented by this driver. -The PF must be administratively up and running when VFs are created and must -remain up for VF mailbox and datapath service. +The PF may be administratively down when VFs are created and remains down +afterward. +VF mailbox and datapath service begin when the PF is brought up, and the PF +must remain up to provide those services. Creating or destroying VFs temporarily stops and restarts a running PF. Set the following per-device .Xr iflib 4 diff --git a/sys/dev/e1000/if_igb_iov.c b/sys/dev/e1000/if_igb_iov.c index 7b6c18690db5..dc4b23397431 100644 --- a/sys/dev/e1000/if_igb_iov.c +++ b/sys/dev/e1000/if_igb_iov.c @@ -2003,6 +2003,10 @@ igb_if_iov_init(if_ctx_t ctx, u16 num_vfs, const nvlist_t *config) sc = iflib_get_softc(ctx); (void)config; + /* + * This callback may run while the PF is down. Record the software + * layout here; igb_iov_initialize() programs it during interface init. + */ atomic_store_rel_32(&sc->iov_teardown, 0); error = igb_iov_validate(sc, num_vfs); if (error != 0) diff --git a/sys/net/iflib.c b/sys/net/iflib.c index 63455fb46d70..da26926c8b17 100644 --- a/sys/net/iflib.c +++ b/sys/net/iflib.c @@ -5621,6 +5621,7 @@ iflib_device_iov_init_restart(device_t dev, uint16_t num_vfs, { if_ctx_t ctx; if_t ifp; + bool restart, running; int error; ctx = device_get_softc(dev); @@ -5629,19 +5630,18 @@ iflib_device_iov_init_restart(device_t dev, uint16_t num_vfs, CTX_LOCK(ctx); /* * Drivers which change the PF queue layout need the complete iflib - * stop/init sequence around their IOV callback. Keep that transition - * within one context-lock critical section. + * stop/init sequence around their IOV callback when the interface is + * active. An administratively-down interface has no live queues to + * quiesce, and must remain down after the new layout is installed. + * Keep the transition within one context-lock critical section. */ - if ((if_getflags(ifp) & IFF_UP) == 0 || - (if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0) { - error = ENETDOWN; - goto out; - } - - iflib_stop(ctx); + restart = (if_getflags(ifp) & IFF_UP) != 0; + running = (if_getdrvflags(ifp) & IFF_DRV_RUNNING) != 0; + if (restart || running) + iflib_stop(ctx); error = IFDI_IOV_INIT(ctx, num_vfs, params); - iflib_init_locked(ctx); -out: + if (restart) + iflib_init_locked(ctx); CTX_UNLOCK(ctx); return (error); }
