> From: Jacob Pan <jacob.jun.pan@xxxxxxxxxxxxxxx> > Sent: Wednesday, March 8, 2023 1:42 AM > > Hi Jason, > > On Fri, 3 Mar 2023 09:15:45 -0400, Jason Gunthorpe <jgg@xxxxxxxxxx> > wrote: > > > On Fri, Mar 03, 2023 at 05:57:41PM +0800, Baolu Lu wrote: > > > On 2023/3/3 17:32, Jean-Philippe Brucker wrote: > > > > > I suppose the common thing is reserving some kind of special > > > > > PASIDs. > > > > Are you planning to use RID_PASID != 0 in VT-d? Otherwise we could > > > > just communicate min_pasid from the IOMMU driver the same way we > do > > > > max_pasid. > > > > > > > > Otherwise I guess re-introduce a lighter ioasid_alloc() that the IOMMU > > > > driver calls to reserve PASID0/RID_PASID. > > > > > > Yes. We probably will use a non-zero RID_PASID in the future. An > > > interface to reserve (or allocate) a PASID from iommu_global_pasid_ida > > > should work then. > > > > Just allowing the driver to store XA_ZERO_ENTRY would be fine > > > So we provide APIs for both? > 1. alloc a global PASID, returned by this API > 2. try to reserve a global PASID given by the driver, i.e. > xa_cmpxchg(&iommu_global_pasid_ida.xa, 2, NULL, > XA_ZERO_ENTRY, > GFP_KERNEL); > seems #1 is sufficient. > No need for both. There will be on-demand allocation on this global space so a reservation interface doesn't make sense.
