> -----Original Message----- > From: Jan Beulich [mailto:jbeul...@suse.com] > Sent: 02 August 2018 09:19 > To: Paul Durrant <paul.durr...@citrix.com> > Cc: Julien Grall <julien.gr...@arm.com>; Andrew Cooper > <andrew.coop...@citrix.com>; George Dunlap > <george.dun...@citrix.com>; Ian Jackson <ian.jack...@citrix.com>; Wei Liu > <wei.l...@citrix.com>; Stefano Stabellini <sstabell...@kernel.org>; xen- > devel <xen-devel@lists.xenproject.org>; Konrad Rzeszutek Wilk > <konrad.w...@oracle.com>; Tim (Xen.org) <t...@xen.org> > Subject: RE: [PATCH v4 01/15] re-work commit 3e06b989 "IOMMU: make > page table population preemptible"... > > >>> On 02.08.18 at 10:04, <paul.durr...@citrix.com> wrote: > >> From: Paul Durrant > >> Sent: 02 August 2018 09:03 > >> > From: Jan Beulich [mailto:jbeul...@suse.com] > >> > Sent: 02 August 2018 08:20 > >> > >>> On 01.08.18 at 15:40, <paul.durr...@citrix.com> wrote: > >> > > ...to simplify the implementation and turn need_iommu back into a > >> > boolean. > >> > > > >> > > As noted in [1] the tri-state nature of need_iommu after commit > >> 3e06b989 > >> > is > >> > > confusing, as is the implementation of pre-emption using relmem_list. > >> > > > >> > > This patch instead uses a simple count of pages already populated > stored > >> in > >> > > the x86 variant of struct arch_iommu and skips over that number of > >> pages > >> > > if arch_iommu_populate_page_table() is re-invoked after pre- > emption. > >> > > >> > Well, yes, I would have used that model in said commit if it was > >> > reliable, > >> > but it isn't: What if the list of pages changed between two (re- > )invocations? > >> > >> Is that really going to happen? This is the result of a domctl, which is a > tools- > >> only hypercall right? > > > > Oh, I see what you mean... the guest could do something like a > > decrease_reservation... I was overlooking that setting up the iommu is > > happening while the guest is live. Would it be reasonable to > domain_pause() > > for safety then? > > I'm hesitant to see domains (or vcpus) paused other than when absolutely > necessary. If everyone else thinks this is a good idea here, I think I won't > object, but please don't forget that any pausing for perhaps an extended > period of time may cause the guest to misbehave subsequently. >
Yes, true, but I also wonder how safe it is to empty the page_list of a running guest. I guess it may be the best way but I think having a dedicate page_list in the iommu struct to host pages that have already been mapped and then transfer them back at the end would be cleaner and should allow need_iommu to stay boolean. Also, given the expense of the operation to set up the mappings, I guess it may also be a good idea to leave them alone until domain destruction once the set-up has been done rather than removing them when the last device is de-assigned. Paul > Jan > _______________________________________________ Xen-devel mailing list Xen-devel@lists.xenproject.org https://lists.xenproject.org/mailman/listinfo/xen-devel
