On Wed, Sep 09, 2020 at 03:06:07PM -0300, Jason Gunthorpe wrote: > On Mon, Sep 07, 2020 at 03:09:12PM +0300, Leon Romanovsky wrote: > > From: Leon Romanovsky <leonro@xxxxxxxxxx> > > > > Changelog: > > v2: > > * Rebased on top of the 524d8ffd07f0 > > * Removed "udata" check in destroy flows > > * Changed ib_free_cq to return early > > * Used Jason's suggestion to implement "RDMA/mlx5: Issue FW command to destroy > > SRQ on reentry" patch. > > v1 > > * Changed returned value in efa_destroy_ah() from EINVAL to EOPNOTSUPP > > * https://lore.kernel.org/lkml/20200830084010.102381-1-leon@xxxxxxxxxx > > v0: > > * https://lore.kernel.org/lkml/20200824103247.1088464-1-leon@xxxxxxxxxx > > > > Hi, > > > > This series restores the ability to fail on destroy commands, due to the > > fact that mlx5_ib DEVX implementation interleaved ib_core objects > > with FW objects without sharing reference counters. > > > > In retrospect, every part of the mlx5_ib flow is correct. > > > > It started from IBTA which was written by HW engineers with HW in mind and > > they allowed to fail in destruction. FW implemented it with symmetrical > > interface like any other command and propagated error back to the kernel, > > which forwarded it to the libibverbs and kernel ULPs. > > > > Libibverbs was designed with IBTA spec in hand putting destroy errors in > > stone. Up till mlx5_ib DEVX, it worked well, because the IB verbs objects > > are counted by the kernel and ib_core ensures that FW destroy will success > > by managing various reference counters on such objects. > > > > The extension of the mlx5 driver changed this flow when allowed DEVX objects > > that are not managed by ib_core to be interleaved with the ones under ib_core > > responsibility. > > > > The drivers that want to implement DEVX flows must ensure that FW/HW > > destroys are performed as early as possible before any other internal > > cleanup. After HW destroys, drivers are not allowed to fail. > > > > This series includes two patches (WQ and "potential race") that will > > require extra work in mlx5_ib, they both theoretical. WQ is not in use > > in DEVX, but is needed to make interface symmetrical to other objects. > > "Potential race" is in ULP flow that ensures that SRQ is destroyed in > > proper order. > > > > Thanks > > > > Leon Romanovsky (9): > > RDMA: Restore ability to fail on PD deallocate > > RDMA: Restore ability to fail on AH destroy > > RDMA/mlx5: Issue FW command to destroy SRQ on reentry > > RDMA: Restore ability to fail on SRQ destroy > > RDMA/core: Delete function indirection for alloc/free kernel CQ > > RDMA: Allow fail of destroy CQ > > RDMA: Change XRCD destroy return value > > RDMA: Restore ability to return error for destroy WQ > > RDMA: Make counters destroy symmetrical > > Thanks, applied to for-next with the changes I noted: Thanks for taking care. LGTM.
