On Mon, Jul 06, 2020 at 01:51:37PM -0700, Jacob Pan wrote: > Hi Jean, > Thanks for the feedback, please see replies inline. > > On Mon, 6 Jul 2020 12:30:41 +0200 > Jean-Philippe Brucker <jean-phili...@linaro.org> wrote: > > > Hi Jacob, > > > > On Thu, Jul 02, 2020 at 06:48:25AM -0700, Jacob Pan wrote: > > > Hi Jean, > > > > > > Just realized I should send this to your Linaro account instead of > > > ARM. So Hi again :) > > > > > > On Wed, 1 Jul 2020 23:29:16 -0700 > > > Jacob Pan <jacob.jun....@linux.intel.com> wrote: > > > > > > > Hi Jean, > > > > > > > > Have a question for you on whether we can have a fixed token type > > > > for ioasid_set. > > > > > > > > Currently, ioasid_set has an arbitrary token. For VT-d vSVA > > > > usage, we choose mm as ioasid_set token to identify PASIDs within > > > > a guest. We have multiple in-kernel users of PASIDs such as VFIO, > > > > KVM, and VDCM. When an IOASID set is created, there is not a good > > > > way to communicate about the token choices. So we have to let > > > > VDCM and KVM *assume* mm is used as token, then retrieve > > > > ioasid_set based on the token. > > > > > > > > This assumption of "mm as token" is not a reliable SW > > > > architecture. > > > > I don't see this as a problem. The token type is tied to the IOASID > > set, so users that pass those IOASID sets to ioasid_find() can safely > > assume that the returned pointer is an mm_struct. That said I'm not > > opposed to consolidating the API with explicit types, it could > > definitely be more elegant. > > > In our use case, we need token to ioasid_set lookup. e.g. VFIO creates > an ioasid_set with a token. KVM instance needs to know the > ioasid_set based on its mm. So ioasid_find() does not know the set > until it finds it based on the token and token type.
Right, I mixed up "token" with the private data associated with each IOASID. I had mostly forgotten about the principles you introduced in the "IOASID extensions for guest SVA" series, sorry for the confusion. > > > > So > > > > we are thinking if we can have an explicit ioasid_set token type > > > > where mm is used. After all, PASID and mm are closely related. > > > > > > > > The code change might be the following: > > > > 1. add a flag to indicate token type when ioasid_set is allocated, > > > > e.g. IOASID_SET_TYPE_MM > > > > IOASID_SET_TYPE_ANY > > > > 2. other users of the ioasid_set can query if an mm token exists > > > > based on the flag IOASID_SET_TYPE_MM, then retrieve the > > > > ioasid_set. > > > > > > > > Existing ioasid_set user can still use arbitrary token under the > > > > flag IOASID_SET_TYPE_ANY > > > > > > > > Would this be an issue for ARM usage? > > > > In my current implementation of auxiliary domains for Arm SMMU (which > > might never be useful enough to go upstream) I don't even use a token > > for the private IOASID set. However I still think we should leave the > > option to use a type different than mm_struct as token for some > > IOASID sets because device drivers (e.g. AMD kfd) may also want to > > dip into the IOASID space and use their own token type. > > > > For the moment, though, we could actually specialize the IOASID API to > > only take an mm_struct as token. > That would be fine with VT-d. We can use init_mm for host PASID set, > process mm for VM set. I'm not fond of using init_mm for the host PASID set. Does it need a token at all? > > > For example the functions exported > > by the IOASID lib would be: > > > > ioasid_t ioasid_alloc_mm(set, min, max, struct mm_struct *mm) > what is the set argument used for? In my view, ioasid_set and mm are > 1:1 mapped. Please disregard this, it was replacing the void* argument of ioasid_alloc() with an mm_struct. > struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota) > > > I was thinking we still have APIs for IOASID set alloc/free since we > want to embed ioasid_set info in driver data structures for ownership > check. > > > struct mm_struct *ioasid_find_mm(set, ioasid) > I don't get why we need ioasid to find the set token. If we put > mm_struct* inside ioasid_set, then we can get the token from the set > directly. Same here, this was replacing the void* returned by ioasid_find() with an mm_struct. > > > ... > > > > And ioasid_alloc(), ioasid_find(), etc would be internal to ioasid.c > > and deal with IOASID_SET_TYPE_MM (or even be removed entirely for > > now). New users that need different token types could then introduce > > their own IOASID_SET_TYPE_* and use the lower-level functions. > > > I will keep that in mind in my next set. I think it would be much > easier to explain with code. > > My takeaway is that we have a high-level agreement to have explicit mm > as token in IOASID APIs. I think we can work out the details with > patches. Yes I think it would be easier to discuss with code. Thanks, Jean _______________________________________________ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu
