On Sun, 15 Mar 2026 13:01:50 -0700 Jakub Kicinski wrote:
> On Sun, 15 Mar 2026 12:54:51 -0700 Jakub Kicinski wrote:
> > Trying to build a minimal kernel I dropped CONFIG_IRQ_REMAP=y from
> > my config, and (on AMD) that caused IRQ shortage.
> >
> > This seems to crash ice after commit ad61cd9c67ad ("ice: get rid of
> > num_lan_msix field"). Sorry for the lack of line numbers, I also
> > dropped DEBUG_INFO. But I think the problem itself is pretty obvious.
> > The fix less so, short of reverting ad61cd9c67ad. We can't just clamp
> > the queues in ice_vsi_alloc_q_vectors() because AFAICT that would make
> > ethtool -L succeed but driver would have a lower queue count than
> > requested.
>
> Hm, maybe it's not just CONFIG_IRQ_REMAP=y
> Enabling it makes no difference. Let me try to see what state the IRQ
> allocation machinery is in on this kernel. On distro kernel ice gets
> all the IRQs it wants at boot. But it also barfs something RDMA so
> I can't really compare..
I think it's ee13aa1a2c5a ("ice: use netif_get_num_default_rss_queues()")
It clamped the number of allocated queues but I think it meant to only
clamp the default enabled queue count. No idea how y'all gonna get the
extra IRQs later or whether you intended to pack multiple queues per IRQ
so I'll let you figure this out..
Thanks for letting me test crash detection and recovery in NIPA, I guess :D
