From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail.kernel.org ([198.145.29.136]:40695 "EHLO mail.kernel.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S965230AbcCPAsW (ORCPT ); Tue, 15 Mar 2016 20:48:22 -0400 Date: Tue, 15 Mar 2016 19:48:17 -0500 From: Bjorn Helgaas To: David Woodhouse Cc: Bjorn Helgaas , "Lawrynowicz, Jacek" , Joerg Roedel , linux-pci@vger.kernel.org, iommu@lists.linux-foundation.org Subject: Re: [PATCH v4 3/6] PCI: Add support for multiple DMA aliases Message-ID: <20160316004817.GG19974@localhost> References: <20160224193926.7585.10833.stgit@bhelgaas-glaptop2.roam.corp.google.com> <20160224194406.7585.17447.stgit@bhelgaas-glaptop2.roam.corp.google.com> <20160225143841.GA8726@localhost> <1457995420.78634.63.camel@infradead.org> MIME-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 In-Reply-To: <1457995420.78634.63.camel@infradead.org> Sender: linux-pci-owner@vger.kernel.org List-ID: On Mon, Mar 14, 2016 at 10:43:40PM +0000, David Woodhouse wrote: > On Thu, 2016-02-25 at 08:38 -0600, Bjorn Helgaas wrote: > > > > >  /* > > > - * Look for aliases to or from the given device for exisiting groups.  The > > > - * dma_alias_devfn only supports aliases on the same bus, therefore the search > > > + * Look for aliases to or from the given device for existing groups. DMA > > > + * aliases are only supported on the same bus, therefore the search > > > > I'm trying to reconcile this statement that "DMA aliases are only > > supported on the same bus" (which was there even before this patch) > > with the fact that pci_for_each_dma_alias() does not have that > > limitation. > > Doesn't it? You can still only set a DMA alias on the same bus with > pci_add_dma_alias(), can't you? I guess it's true that PCI_DEV_FLAGS_DMA_ALIAS_DEVFN and the proposed pci_add_dma_alias() only add aliases on the same bus. I was thinking about a scenario like this: 00:00.0 PCIe-to-PCI bridge to [bus 01] 01:01.0 conventional PCI device where I think 01:00.0 is a DMA alias for 01:01.0 because the bridge takes ownership of DMA transactions from 01:01.0 and assigns a Requester ID of 01:00.0 (secondary bus number, device 0, function 0). > > >   * space is quite small (especially since we're really only looking at pcie > > >   * device, and therefore only expect multiple slots on the root complex or > > >   * downstream switch ports).  It's conceivable though that a pair of > > > @@ -686,11 +692,8 @@ static struct iommu_group *get_pci_alias_group(struct pci_dev *pdev, > > >                       continue; > > >   > > >               /* We alias them or they alias us */ > > > -             if (((pdev->dev_flags & PCI_DEV_FLAGS_DMA_ALIAS_DEVFN) && > > > -                  pdev->dma_alias_devfn == tmp->devfn) || > > > -                 ((tmp->dev_flags & PCI_DEV_FLAGS_DMA_ALIAS_DEVFN) && > > > -                  tmp->dma_alias_devfn == pdev->devfn)) { > > > - > > > +             if (dma_alias_is_enabled(pdev, tmp->devfn) || > > > +                 dma_alias_is_enabled(tmp, pdev->devfn)) { > > >                       group = get_pci_alias_group(tmp, devfns); > > > > We basically have this: > > > >   for_each_pci_dev(tmp) { > >     if () > >       group = get_pci_alias_group(); > >       ... > >   } > > Strictly, that's: > >  for_each_pci_dev(tmp) { >    if (pdev is an alias of tmp || tmp is an alias of pdev) >      group = get_pci_alias_group(); >    ... >  } OK. > > I'm trying to figure out why we don't do something like the following > > instead: > > > >   callback(struct pci_dev *pdev, u16 alias, void *opaque) > >   { > >     struct iommu_group *group; > > > >     group = get_pci_alias_group(); > >     if (group) > >       return group; > > > >     return 0; > >   } > > > >   pci_for_each_dma_alias(pdev, callback, ...); > > And this would be equivalent to > >  for_each_pci_dev(tmp) { >    if (tmp is an alias of pdev) >      group = get_pci_alias_group(); >    ... >  } > > The "is an alias of" property is not commutative. Perhaps it should be. > But that's hard because in some cases the alias doesn't even *exist* as > a real PCI device. It's just that you appear to get DMA transactions > from a given source-id. Right. In my example above, 01:00.0 is not a PCI device; it's only a Requester ID that is fabricated by the bridge when it forwards DMA transactions upstream. I think I'm confused because I don't really understand IOMMU groups. Let me explain what I think they are and you can correct me when I go wrong. The iommu_group_alloc() comment says "The IOMMU group represents the minimum granularity of the IOMMU." So I suppose the IOMMU cannot distinguish between devices in a group. All the devices in the group use the same set of DMA mappings. Granting device A DMA access to a buffer grants the same access to all other members of A's IOMMU group. That would mean my question was fundamentally backwards. In get_pci_alias_group(A), we're not trying to figure out what all the aliases of A are, which is what pci_for_each_dma_alias() does. Instead, we're trying to figure out which IOMMU group A belongs to. But I still don't quite understand how aliases fit into this. Let's go back to my example and assume we've already put 00:00.0 and 01:01.0 in IOMMU groups: 00:00.0 PCIe-to-PCI bridge to [bus 01] # in IOMMU group G0 01:01.0 conventional PCI device # in IOMMU group G1 I assume these devices are in different IOMMU groups because if the bridge generated its own DMA, it would use Requester ID 00:00.0, which is distinct from the 01:00.0 it would use when forwarding DMAs from its secondary side. What happens when we add 01:02.0? I think 01:01.0 and 01:02.0 should both end up in IOMMU group G1 because the IOMMU will see only Requester ID 01:00.0, so it can't distinguish them. When we add 01:02.0, the ops->add_device() ... ops->device_group() path calls pci_device_group(01:02.0): pci_device_group(01:02.0) pci_for_each_dma_alias(01:02.0, get_pci_alias_or_group) get_pci_alias_or_group(01:02.0, 01:02.0) # callback return 0 # 01:02.0 group not set yet get_pci_alias_or_group(00:00.0, 01:00.0) # callback return 1 # 00:00.0 is in G0 It seems like we'll assign 01:02.0 to group G0, when I think it should be in G1. Where did I go wrong? Bjorn