Re: [PATCH 1/1] iommu/arm-smmu-v3: add support for BBML

[Will Deacon <will@kernel.org>]

On Fri, Jan 22, 2021 at 09:21:20PM +0800, Leizhen (ThunderTown) wrote:
> On 2021/1/22 21:00, Robin Murphy wrote:
> > On 2021-01-22 12:51, Will Deacon wrote:
> >> On Thu, Nov 26, 2020 at 11:42:30AM +0800, Zhen Lei wrote:
> >>>   -        tblp = ptep - ARM_LPAE_LVL_IDX(iova, lvl, data);
> >>> -        if (__arm_lpae_unmap(data, NULL, iova, sz, lvl, tblp) != sz) {
> >>> -            WARN_ON(1);
> >>> -            return -EINVAL;
> >>> +        switch (cfg->bbml) {
> >>> +        case 0:
> >>> +            /*
> >>> +             * We need to unmap and free the old table before
> >>> +             * overwriting it with a block entry.
> >>> +             */
> >>> +            tblp = ptep - ARM_LPAE_LVL_IDX(iova, lvl, data);
> >>> +            if (__arm_lpae_unmap(data, NULL, iova, sz, lvl, tblp) != sz) {
> >>> +                WARN_ON(1);
> >>> +                return -EINVAL;
> >>> +            }
> >>> +            break;
> >>> +        case 1:
> >>> +            __arm_lpae_init_pte(data, paddr, prot, lvl, ptep, ARM_LPAE_PTE_nT);
> >>> +
> >>> +            io_pgtable_tlb_flush_walk(iop, iova, sz, ARM_LPAE_GRANULE(data));
> >>> +            tblp = iopte_deref(pte, data);
> >>> +            __arm_lpae_free_pgtable(data, lvl + 1, tblp);
> >>> +            break;
> >>> +        case 2:
> >>> +            __arm_lpae_init_pte(data, paddr, prot, lvl, ptep, 0);
> >>> +
> >>> +            io_pgtable_tlb_flush_walk(iop, iova, sz, ARM_LPAE_GRANULE(data));
> >>> +            tblp = iopte_deref(pte, data);
> >>> +            __arm_lpae_free_pgtable(data, lvl + 1, tblp);
> >>> +            return 0;
> >>
> >> Sorry, but I really don't understand what you're trying to do here. The old
> >> code uses BBM for the table -> block path so we don't need anything extra
> >> here. The dodgy case is when we unmap part of a block, and end up installing
> >> a table via arm_lpae_split_blk_unmap(). We can't use BBM there because there
> >> could be ongoing DMA to parts of the block mapping that we want to remain in
> >> place.
> >>
> >> Are you seeing a problem in practice?
> > 
> > Right, I was under the assumption that we could ignore BBML because we
> > should never have a legitimate reason to split blocks. I'm certainly not
> > keen on piling any more complexity into split_blk_unmap, because the
> > IOMMU API clearly doesn't have a well-defined behaviour for that case
> > anyway - some other drivers will just unmap the entire block, and IIRC
> > there was a hint somewhere in VFIO that it might actually expect that
> > behaviour.
> 
> I'm going home. I'll answer you two tomorrow.

It can wait until Monday! Have a good weekend :)

Will