lkml.org 
[lkml]   [2014]   [May]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    SubjectRe: [PATCH v2 00/15] PCI/iommu: Fix DMA alias problems
    From
    Date
    On Fri, 2014-05-16 at 07:45 +0800, Andrew Cooks wrote:
    > Hi Alex
    >
    > On Fri, May 16, 2014 at 1:43 AM, Alex Williamson
    > <alex.williamson@redhat.com> wrote:
    > > On Thu, 2014-05-15 at 07:40 +0800, Andrew Cooks wrote:
    > >> Hi Alex
    > >>
    > >> On Fri, May 9, 2014 at 11:28 PM, Alex Williamson
    > >> <alex.williamson@redhat.com> wrote:
    > >> > ....
    > >> >
    > >> > Original description:
    > >> >
    > >> > This series attempts to fix a couple issues we've had outstanding in
    > >> > the PCI/IOMMU code for a while. The first issue is with devices that
    > >> > use the wrong requester ID for DMA transactions. We already have a
    > >> > sort of half-baked attempt to fix this for several Ricoh devices, but
    > >> > the fix only helps them be useful through IOMMU groups, not the
    > >> > general DMA case. There are also several Marvell devices which use
    > >> > use a different wrong requester ID and don't even fit into the DMA
    > >> > source idea. This series creates a DMA alias iterator that will
    > >> > step through each possible alias of a device, allowing IOMMUs to
    > >> > insert mappings for both the device and its aliases.
    > >> >
    > >> > Hand-in-hand with this is our broken pci_find_upstream_pcie_bridge()
    > >> > function, which is known to blowup when it finds itself suddenly at
    > >> > a PCIe device without crossing a PCIe-to-PCI bridge (as identified by
    > >> > the PCIe capability). It also likes to make the invalid assumption
    > >> > that a PCIe device never has its requester ID masked by any usptream
    > >> > bus. We can fix this using the above new DMA alias iterator, since
    > >> > that's effectively what this function was meant to do.
    > >> >
    > >>
    > >> There are two cases where the DMA requester id seems to use the wrong
    > >> slot (as opposed to function) in the patch I attached to
    > >> https://bugzilla.kernel.org/show_bug.cgi?id=42679
    > >> The original bug reports are listed in the patch.
    > >>
    > >> Unfortunately, I wasn't able to get test feedback on those two cases,
    > >> but I'm wondering...
    > >> Did I understand correctly that a slot alias is something that could be needed?
    > >> If so, is it a variation of the PCIe-to-PCI bridge case that's already
    > >> covered or will it require a different approach?
    > >
    > > Wow, I didn't think that kind of broken was possible. Maybe instead of
    > > a bitmap of function aliases we could have a single devfn alias for a
    > > device. That means we'd only be able to support a single alias for a
    > > device, but since I don't think we've seen devices that use more than a
    > > single alias, maybe that's ok.
    >
    > The current patch creates a context entry for the reported device
    > (function 0), plus it's alias (function 1). Is there reason to always
    > add a context entry for the reported devfn and define 'alias' to mean
    > 'one additional devfn'? That will work for all the Marvell cases.

    Right, that's effectively what it would become, probably making use of a
    flag bit to indicate whether the alias devfn is valid. The
    pci_for_each_dma_alias() would just drop the loop over the
    dma_alias_funcs bitmap at replace it with a test of the flag and single
    dma alias devfn. I need to think about whether the IOMMU group code can
    is such a simple replacement.

    I think it makes sense to always include both the actual devfn and, if
    it exists, an alias devfn. Otherwise we'd just end up with more quirks
    to add later.

    > In the testing I did, the Marvell controllers needed context entries
    > for both function 0 and 1. It would be nice if someone could confirm
    > or debunk this. I tested with a 88SE9172 with both ports of the
    > controller in use.

    I think 0 needs to be quirked to 1, but I don't think 1 needs to be
    quirked to 0. Unfortunately intel-iommu doesn't do any of the reference
    counting that amd_iommu does, so if we have 0 aliased to 1 and we unbind
    function 0 from the driver, intel-iommu will also teardown the mapping
    for function 1. It's horrible. That's a problem beyond what I'm trying
    to do here though, it already exists if you have two devices behind a
    PCIe-to-PCI bridge. Thanks,

    Alex



    \
     
     \ /
      Last update: 2014-05-16 06:21    [W:3.106 / U:0.204 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site