Re: [PATCH rdma-next 00/15] RDMA: Support HW requiring relaxed ordering and add Unordered access flag
Michael Gur <[email protected]>
| Newsgroups | org.kernel.vger.linux-rdma,org.kernel.vger.netdev |
|---|---|
| Message-ID | <[email protected]> |
On 8/10/2026 4:05 PM, Leon Romanovsky wrote: > External email: Use caution opening links or attachments > > > On Sun, Jul 26, 2026 at 12:29:14PM +0300, Michael Gur wrote: >> This series prepares the mlx5 stack for hardware that restricts strong- >> ordered memory accesses, i.e. HW whose fabric doesn't guarantee memory >> operation ordering. The HW may choose to reject creation of an mkey that >> doesn't request relaxed ordering, so every mkey the driver creates - >> internal and user-facing - has to be handled accordingly. >> >> The previously added 'ConnectX-10 NVLink-C2C' PCI device is such a device, >> as will be any future PCIe Gen7 UIO devices. >> >> The series does four things: >> >> 1. Allow optional access flags on all memory registration >> The uverbs ABI defines the optional access flags range as best-effort: >> drivers must accept any flag in this range and treat support as >> opportunistic, never failing registration. Several registration paths >> (DM, DEVX UMEM, dmabuf ioctl, reg_mr ioctl) did not allow this range >> and would reject such requests. These are fixed to allow newly added >> optional access flag added later in the series and any future optional >> access flag. >> >> 2. Groundwork and internal mkeys >> Rename the relaxed-ordering IFC fields to the new naming, add the new >> ordering capability and mkey-context bits, and consolidate RO-setting >> to a single helper in mlx5_core so all consumers share one cap-gated >> implementation. Then enable relaxed ordering on the internal mkeys that >> don't require strong ordering (resource dump, vfio live-migration data, >> ODP null mkey, memory window mkey), so they keep working on strict-RO >> HW. >> >> 3. New Unordered access flag and enforcement >> Add IB_ACCESS_UNORDERED, an optional access flag that additionally >> allows read-after-write reordering, and wire up mlx5 support for it. >> >> 4. Enforce no-strong-ordering in user MR registration on HW that requires >> it. >> >> Michael Gur (15): >> RDMA/mlx5: Allow optional access flags in DM registration >> RDMA/mlx5: Allow optional access flags in DEVX UMEM registration >> RDMA/core: Allow optional access flags in dmabuf reg ioctl >> RDMA/core: Allow optional access flags in reg mr ioctl >> mlx5: Rename IFC bits of relaxed ordering fields >> net/mlx5: Add IFC bits for new ordering caps >> mlx5: Move RO setting helper to core and consolidate mlx5 consumers >> net/mlx5: Enable relaxed ordering on resource dump mkey >> vfio/mlx5: Enable relaxed ordering on the live migration data mkey >> RDMA/mlx5: Enable relaxed ordering on ODP null mkey >> RDMA/mlx5: Enable relaxed ordering on Memory Window mkey on strict-RO >> HW >> RDMA/mlx5: Converge UMR access-flag cap checks >> RDMA/uverbs: Add new Unordered MR access flag >> RDMA/mlx5: Support new Unordered access flag >> RDMA/mlx5: Enforce relaxed ordering when HW requires it >> >> drivers/infiniband/core/umem.c | 2 +- >> drivers/infiniband/core/uverbs_std_types_mr.c | 4 +- >> drivers/infiniband/hw/mlx5/devx.c | 2 +- >> drivers/infiniband/hw/mlx5/main.c | 8 +- >> drivers/infiniband/hw/mlx5/mlx5_ib.h | 4 +- >> drivers/infiniband/hw/mlx5/mr.c | 79 ++++++++++--------- >> drivers/infiniband/hw/mlx5/odp.c | 1 + >> drivers/infiniband/hw/mlx5/umr.c | 18 ++--- >> drivers/infiniband/hw/mlx5/umr.h | 45 ++++++----- >> .../mellanox/mlx5/core/diag/rsc_dump.c | 1 + >> drivers/net/ethernet/mellanox/mlx5/core/en.h | 1 - >> .../ethernet/mellanox/mlx5/core/en/params.c | 2 +- >> .../ethernet/mellanox/mlx5/core/en_common.c | 13 +-- >> .../net/ethernet/mellanox/mlx5/core/en_main.c | 4 +- >> drivers/net/ethernet/mellanox/mlx5/core/mr.c | 21 +++++ >> drivers/vfio/pci/mlx5/cmd.c | 7 +- >> include/linux/mlx5/driver.h | 1 + >> include/linux/mlx5/mlx5_ifc.h | 24 ++++-- >> include/rdma/ib_verbs.h | 1 + >> include/uapi/rdma/ib_user_ioctl_verbs.h | 1 + >> 20 files changed, 134 insertions(+), 105 deletions(-) > Sashiko is not happy about this series. > > Thanks Sashiko has 4 comments on this series: 2 comments (on patch 14 and 15) concerning data-direct flows not handling the Unordered access flag or RO enforcement: These are irrelevant as we don't have HW that will support both caps. That is currently an invalid configuration. Comment on patch 3: suggesting we need a fix that is actually added in patch 4. Comment on patch 15: Sashiko misreading the meaning of the cap. 'mkc_order_write_after_write_ro_only' means we must set RO in the write_after_write field of every mkey, regardless of actual write access flags at registration. Thanks >> -- >> 2.52.0 >>