Re: [PATCH v4 00/12] PCI: endpoint: pci-epf-vntb / NTB: epf: Enable per-doorbell bit handling

On Wed, May 13, 2026 at 11:49:11AM +0900, Koichiro Den wrote:
> This series fixes doorbell bit/vector handling for the EPF-based NTB
> pair (ntb_hw_epf <-> pci-epf-*ntb). Its primary goal is to enable safe
> per-db-vector handling in the NTB core and clients (e.g. ntb_transport),
> without changing the on-the-wire doorbell mapping.
> 

FYI, I'm waiting for Acks from NTB maintainers for all the NTB patches.

- Mani

> 
> Background / problem
> ====================
> 
> ntb_hw_epf historically applies an extra offset when ringing peer
> doorbells: the link event uses the first interrupt slot, and doorbells
> start from the third slot (i.e. a second slot is effectively unused).
> pci-epf-vntb carries the matching offset on the EP side as well.
> 
> As long as db_vector_count()/db_vector_mask() are not implemented, this
> mismatch is mostly masked. Doorbell events are effectively treated as
> "can hit any QP" and the off-by-one vector numbering does not surface
> clearly.
> 
> However, once per-vector handling is enabled, the current state becomes
> problematic:
> 
>   - db_valid_mask exposes bits that do not correspond to real doorbells
>     (link/unused slots leak into the mask).
>   - ntb_db_event() is fed with 1-based/shifted vectors, while NTB core
>     expects a 0-based db_vector for doorbells.
>   - On pci-epf-vntb, .peer_db_set() may be called in atomic context, but
>     it directly calls pci_epc_raise_irq(), which can sleep.
> 
> 
> Why NOT fix the root offset?
> ============================
> 
> The natural "root" fix would be to remove the historical extra offset in
> the peer_db_set() doorbell paths for ntb_hw_epf and pci-epf-vntb.
> Unfortunately this would lead to interoperability issues when mixing old
> and new kernel versions (old/new peers). A new side would ring a
> different interrupt slot than what an old peer expects, leading to
> missed or misrouted doorbells, once db_vector_count()/db_vector_mask()
> are implemented.
> 
> Therefore this series intentionally keeps the legacy offset, and instead
> fixes the surrounding pieces so the mapping is documented and handled
> consistently in masks, vector numbering, and per-vector reporting.
> 
> 
> Dependencies
> ============
> 
> This v4 is prepared on top of the current pci/endpoint:
> commit f5aa97125491 ("misc: pci_endpoint_test: remove dead BAR read before
>                       doorbell trigger")
> 
> plus these patches as contextual prerequisites:
> 
>   - [PATCH 0/2] NTB: epf: Fix ntb_hw_epf ISR issues
>     https://lore.kernel.org/ntb/[email protected]/
> 
>   - [PATCH 0/2] PCI: endpoint: pci-epf-{v}ntb: A couple of fixes
>     https://lore.kernel.org/linux-pci/[email protected]/
> 
> (These touch the same code part as this series and would otherwise conflict
> with it. They should land upstream first in my opinion.)
> 
> 
> Practical benefit
> =================
> 
> This also enables ntb_netdev's multi-queue performance scale:
> 
>   [PATCH v3 0/4] net: ntb_netdev: Add Multi-queue support
>   https://lore.kernel.org/netdev/[email protected]/
> 
> Each netdev queue maps to an ntb_transport queue pair. With correct
> db_vector_count() and db_vector_mask() callbacks, ntb_transport can wake
> only the queue pairs covered by the doorbell vector it received. Without
> those callbacks, the NTB core falls back to the full valid doorbell mask,
> so each doorbell interrupt can still wake all queue pairs. See "Testing /
> Results" section in the above link.
> 
> 
> What this series does
> =====================
> 
> - pci-epf-vntb:
> 
>   - Document the legacy offset.
>   - Defer MSI doorbell raises to process context to avoid sleeping in
>     atomic context. This becomes relevant once multiple doorbells are
>     raised concurrently at a high rate.
>   - Report doorbell vectors as 0-based to ntb_db_event().
>   - Reject unusable doorbell counts and guard configfs writes after EPC
>     attach with -EOPNOTSUPP.
>   - Fix db_valid_mask and implement db_vector_count()/db_vector_mask().
> 
> - ntb_hw_epf:
> 
>   - Document the legacy offset in ntb_epf_peer_db_set().
>   - Fix db_valid_mask to cover only real doorbell bits.
>   - Report 0-based db_vector to ntb_db_event() (accounting for the
>     unused slot).
>   - Pass per-vector context as request_irq() dev_id instead of deriving the
>     device vector from Linux IRQ numbers.
>   - Keep db_val as a bitmask and fix db_read/db_clear semantics
>     accordingly.
>   - Implement db_vector_count()/db_vector_mask().
> 
> 
> Compatibility
> =============
> 
> By keeping the legacy offset intact, this series aims to remain
> compatible across mixed kernel versions. The observable changes are
> limited to correct mask/vector reporting and safer execution context
> handling.
> 
> Patches 1-7 (PCI Endpoint) and 8-12 (NTB) are independent and can be
> applied separately through the respective trees. They are sent together so
> the two sides of the same legacy layout do not drift apart, and the
> resulting code stays easier to follow.
> 
> The plan is still to take the whole series through the PCI EP tree once the
> remaining NTB acks are collected. See:
> https://lore.kernel.org/linux-pci/rnzsnp5de4qf5w7smebkmqekpuaqckltx73rj6ha3q2nrby5yp@7hsgvdzvjkp6/
> 
> 
> Note
> ====
> 
> Most v3->v4 changes are responses to Sashiko feedback:
> https://sashiko.dev/#/patchset/20260323031544.2598111-1-den%40valinux.co.jp
> 
> Sashiko also flagged a broader driver-unbind lifetime concern around the
> vNTB PCI driver. That is real, but it predates this series and needs a
> proper .remove() path, including NTB unregistration and work cancellation.
> I am keeping it as follow-up lifecycle work:
> 
> https://lore.kernel.org/linux-pci/[email protected]/
> 
> This v4 only clears the deferred doorbell state added by this series around
> EPC init/cleanup.
> 
> 
> ---
> Changelog
> =========
> 
> Changes since v3:
>   - Rebased on the dependencies listed above.
>   - For Sashiko feedback on state added by this series: clear
>     peer_db_pending around work enable/disable, mask peer_db_set() with
>     db_valid_mask, and use __ffs64().
>   - For pre-existing issues made more relevant by per-vector semantics:
>     stop deriving ntb_hw_epf vectors from Linux IRQ numbers, pass
>     per-vector request_irq() context, ignore the reserved slot, and
>     validate vectors before BIT_ULL()/ntb_db_event().
>   - Added adjacent hardening: pci-epf-vntb db_count range/configfs guards,
>     bounded db_vector masks, ntb_hw_epf mw_count precheck, and helper guard
>     cleanups.
>   - Dropped Reviewed-by tags from patches that changed to non-trivial extent.
> 
> Changes since v2:
>   - No functional changes.
>   - Rebased onto current pci/endpoint
>     e022f0c72c7f ("selftests: pci_endpoint: Skip reserved BARs").
>     * Patch 2 needed a trivial context-only adjustment while rebasing, due
>       to commit d799984233a5 ("PCI: endpoint: pci-epf-vntb: Stop
>       cmd_handler work in epf_ntb_epc_cleanup").
>   - Picked up additional Reviewed-by tags from Frank.
>   - Fixed the incorrect v2 series title.
> 
> Changes since v1:
>   - Addressed feedback from Dave (add a source code comment, introduce
>     enum to eliminate magic numbers)
>   - Updated source code comment in Patch 2.
>   - No functional changes, so retained Reviewed-by tags by Frank and Dave.
>     Thank you both for the review.
> 
> v3: https://lore.kernel.org/linux-pci/[email protected]/
> v2: https://lore.kernel.org/linux-pci/[email protected]/
> v1: https://lore.kernel.org/linux-pci/[email protected]/
> 
> 
> Best regards,
> Koichiro
> 
> 
> Koichiro Den (12):
>   PCI: endpoint: pci-epf-vntb: Document legacy MSI doorbell offset
>   PCI: endpoint: pci-epf-vntb: Defer pci_epc_raise_irq() out of atomic
>     context
>   PCI: endpoint: pci-epf-vntb: Report 0-based doorbell vector via
>     ntb_db_event()
>   PCI: endpoint: pci-epf-vntb: Reject unusable doorbell counts
>   PCI: endpoint: pci-epf-vntb: Guard configfs writes after EPC attach
>   PCI: endpoint: pci-epf-vntb: Exclude reserved slots from db_valid_mask
>   PCI: endpoint: pci-epf-vntb: Implement db_vector_count/mask for
>     doorbells
>   NTB: epf: Document legacy doorbell slot offset in
>     ntb_epf_peer_db_set()
>   NTB: epf: Make db_valid_mask cover only real doorbell bits
>   NTB: epf: Report 0-based doorbell vector via ntb_db_event()
>   NTB: epf: Fix doorbell bitmask and IRQ vector handling
>   NTB: epf: Implement db_vector_count/mask for doorbells
> 
>  drivers/ntb/hw/epf/ntb_hw_epf.c               | 133 ++++++++---
>  drivers/pci/endpoint/functions/pci-epf-vntb.c | 210 ++++++++++++++++--
>  2 files changed, 296 insertions(+), 47 deletions(-)
> 
> -- 
> 2.51.0
> 

-- 
மணிவண்ணன் சதாசிவம்