Hi, On 06/01/2017 18:18, Auger Eric wrote: > Hi Joerg, Robin, > > On 06/01/2017 13:48, Joerg Roedel wrote: >> On Fri, Jan 06, 2017 at 12:46:05PM +0100, Auger Eric wrote: >>> On 06/01/2017 12:00, Joerg Roedel wrote: >> >>>> I think it also makes sense to report the type of the reserved region. >>> >>> What is the best practice in that case? Shall we put the type enum >>> values as strings such as: >>> - direct >>> - nomap >>> - msi >>> >>> and document that in Documentation/ABI/testing/sysfs-kernel-iommu_groups >> >> Yes, a string would be good. An probably 'reserved' is a better name >> than nomap? > the iommu_insert_resv_region() function that builds the group reserved > region list sorts all regions and handles the case where there is an > overlap between regions. Current code does not care about the type of > regions. So in case a NOMAP region overlaps with a direct-mapped region, > what is reported to the user space is the superset and the type depends > on the overlap. This was suggested by Robin at some point to handle > overlaps. > > I guess I should merge regions only in case the types equal? > > I remember that Alex thought that user-space should not care so much > about the type of the regions so I tought it was better for the > user-space to have a minimal view of the regions. > > On the other hand, this issue of merging regions of different types > should not happen often but I prefer to highlight the potential issue.
> > What is your guidance? Please forget the question. From an API point of view It does not make sense that iommu_insert_resv_region() merges regions of a different types since the type field becomes unreliable. I will fix this. Thanks Eric > > Thanks > > Eric >> >> >> Joerg >>
