[Intel-wired-lan] [iwl-next v2 0/7] ice: managing MSI-X in driver

Michal Swiatkowski michal.swiatkowski at linux.intel.com
Mon Aug 5 04:54:23 UTC 2024


On Fri, Aug 02, 2024 at 02:41:54PM +0200, Przemek Kitszel wrote:
> On 8/1/24 11:31, Michal Swiatkowski wrote:
> > Hi,
> > 
> > It is another try to allow user to manage amount of MSI-X used for each
> > feature in ice. First was via devlink resources API, it wasn't accepted
> > in upstream. Also static MSI-X allocation using devlink resources isn't
> > really user friendly.
> > 
> > This try is using more dynamic way. "Dynamic" across whole kernel when
> > platform supports it and "dynamic" across the driver when not.
> > 
> > To achieve that reuse global devlink parameter pf_msix_max and
> > pf_msix_min. It fits how ice hardware counts MSI-X. In case of ice amount
> > of MSI-X reported on PCI is a whole MSI-X for the card (with MSI-X for
> > VFs also). Having pf_msix_max allow user to statically set how many
> > MSI-X he wants on PF and how many should be reserved for VFs.
> > 
> > pf_msix_min is used to set minimum number of MSI-X with which ice driver
> > should probe correctly.
> > 
> > Meaning of this field in case of dynamic vs static allocation:
> > - on system with dynamic MSI-X allocation support
> >   * alloc pf_msix_min as static, rest will be allocated dynamically
> > - on system without dynamic MSI-X allocation support
> >   * try alloc pf_msix_max as static, minimum acceptable result is
> >   pf_msix_min
> > 
> > As Jesse and Piotr suggested pf_msix_max and pf_msix_min can (an
> > probably should) be stored in NVM. This patchset isn't implementing
> > that.
> > 
> > Dynamic (kernel or driver) way means that splitting MSI-X across the
> > RDMA and eth in case there is a MSI-X shortage isn't correct. Can work
> > when dynamic is only on driver site, but can't when dynamic is on kernel
> > site.
> > 
> > Let's remove this code and move to MSI-X allocation feature by feature.
> > If there is no more MSI-X for a feature, a feature is working with less
> > MSI-X or it is turned off.
> > 
> > There is a regression here. With MSI-X splitting user can run RDMA and
> > eth even on system with not enough MSI-X. Now only eth will work. RDMA
> > can be turned on by changing number of PF queues (lowering) and reprobe
> > RDMA driver.
> > 
> > Example:
> > 72 CPU number, eth, RDMA and flow director (1 MSI-X), 1 MSI-X for OICR
> > on PF, and 1 more for RDMA. Card is using 1 + 72 + 1 + 72 + 1 = 147.
> > 
> > We set pf_msix_min = 2, pf_msix_max = 128
> > 
> > OICR: 1
> > eth: 72
> > RDMA: 128 - 73 = 55
> > flow director: turned off not enough MSI-X
> > 
> > We can change number of queues on pf to 36 and do devlink reinit
> > 
> > OICR: 1
> > eth: 36
> > RDMA: 73
> > flow director: 1
> > 
> > We can also (implemented in "ice: enable_rdma devlink param") turned
> > RDMA off.
> > 
> > OICR: 1
> > eth: 72
> > RDMA: 0 (turned off)
> > flow director: 1
> > 
> > Maybe flow director should have higher priority than RDMA? It needs only
> > 1 MSI-X, so it seems more logic to lower RDMA by one then maxing MSI-X
> > on RDMA and turning off flow director (as default).
> 
> sounds better, less surprising, with only RDMA being affected by this
> series as "regression"
> 

Sounds reasonable

> > 
> > After this changes we have a static base vector for SRIOV (SIOV probably
> > in the feature). Last patch from this series is simplifying managing VF
> > MSI-X code based on static vector.
> > 
> > Now changing queues using ethtool is also changing MSI-X. If there is
> > enough MSI-X it is always one to one. When there is not enough there
> > will be more queues than MSI-X. There is a lack of ability to set how
> > many queues should be used per MSI-X. Maybe we should introduce another
> > ethtool param for it? Sth like queues_per_vector?
> 
> Our 1:1 mapping was too rigid solution (but performant), I like MSI-Xes
> being kept as a detail and [setting of them] decoupled from being
> mandatory on [at least some] flows. Tuning the mapping could be useful,
> esp in heterotelic scenarios (like keeping XDP stuff separate). Could be
> left for the future.
> 
> What happens when user decreases number of MSI-X, queues will just get
> remapped to other?
>

Yes, queue will be remapped

> > 
> > v1 --> v2: [1]
> >   * change permanent MSI-X cmode parameters to driverinit
> >   * remove locking during devlink parameter registration (it is now
> >     locked for whole init/deinit part)
> > 
> > [1] https://lore.kernel.org/netdev/20240213073509.77622-1-michal.swiatkowski@linux.intel.com/
> > 
> > Michal Swiatkowski (7):
> >    ice: devlink PF MSI-X max and min parameter
> >    ice: remove splitting MSI-X between features
> >    ice: get rid of num_lan_msix field
> >    ice, irdma: move interrupts code to irdma
> >    ice: treat dyn_allowed only as suggestion
> >    ice: enable_rdma devlink param
> >    ice: simplify VF MSI-X managing
> > 
> >   drivers/infiniband/hw/irdma/hw.c              |   2 -
> >   drivers/infiniband/hw/irdma/main.c            |  46 ++-
> >   drivers/infiniband/hw/irdma/main.h            |   3 +
> >   .../net/ethernet/intel/ice/devlink/devlink.c  |  75 ++++-
> >   drivers/net/ethernet/intel/ice/ice.h          |  21 +-
> >   drivers/net/ethernet/intel/ice/ice_base.c     |  10 +-
> >   drivers/net/ethernet/intel/ice/ice_ethtool.c  |   8 +-
> >   drivers/net/ethernet/intel/ice/ice_idc.c      |  64 +---
> >   drivers/net/ethernet/intel/ice/ice_irq.c      | 277 ++++++------------
> >   drivers/net/ethernet/intel/ice/ice_irq.h      |  13 +-
> >   drivers/net/ethernet/intel/ice/ice_lib.c      |  36 ++-
> >   drivers/net/ethernet/intel/ice/ice_sriov.c    | 153 +---------
> >   include/linux/net/intel/iidc.h                |   2 +
> >   13 files changed, 287 insertions(+), 423 deletions(-)
> > 
> 


More information about the Intel-wired-lan mailing list