On Mon, Jun 12, 2017 at 10:34:43AM +0800, Peter Xu wrote: > On Sun, Jun 11, 2017 at 08:10:15PM +0800, David Gibson wrote: > > On Sun, Jun 11, 2017 at 01:09:26PM +0300, Michael S. Tsirkin wrote: > > > On Fri, Jun 09, 2017 at 09:58:47AM +0800, Peter Xu wrote: > > > > > > The problem is that when I was fixing the problem that vhost had > > > > > > with > > > > > > PT (a764040, "exec: abstract address_space_do_translate()"), I did > > > > > > broke the IOTLB translation a bit (it was using page masks). IMHO we > > > > > > need to fix it first for correctness (patch 1/2). > > > > > > > > > > > > For patch 3, if we can have Jason's patch to allow dynamic > > > > > > iommu_platform switching, that'll be the best, then I can rewrite > > > > > > patch 3 with the switching logic rather than caching anything. But > > > > > > IMHO that can be separated from patch 1/2 if you like. > > > > > > > > > > > > Or do you have better suggestion on how should we fix it? > > > > > > > > > > > > Thanks, > > > > > > > > > > Can we drop masks completely and replace with length? I think we > > > > > should do that instead of trying to fix masks. > > > > > > > > Do you mean to modify IOMMUTLBEntry.addr_mask into length? > > > > > > I think it's better than alternatives. > > > > > > > Again, I am not sure this is good... At least we need to get ack from > > > > David since spapr should be the initial user of it, and possibly also > > > > Alex since vfio should be assuming that (IIUC both in QEMU and kernel) > > > > addr_mask is page masks rather than arbirary length. > > > > > > > > (CC Alex) > > > > > > > > Thanks, > > > > > > Callbacks that need powers of two can easily split up the range. > > > > I think I missed part of the thread. What's the original use case for > > non-power-of-two IOTLB entries? It certainly won't happen on Power. > > Currently address_space_get_iotlb_entry() didn't really follow the > rule, addr_mask can be arbitary length. This series tried to fix it, > while Michael was questioning about whether we should really fix that > at all. > > Michael, > > Even if for performance's sake, I should still think we should fix it. > Let's consider a most simple worst case: we have a single page mapped > with IOVA range (2M page): > > [0x0, 0x200000) > > And if guest access IOVA using the following patern: > > 0x1fffff, 0x1ffffe, 0x1ffffd, ... > > Then now we'll get this: > > - request 0x1fffff, cache miss, will get iotlb [0x1fffff, 0x200000) > - request 0x1ffffe, cache miss, will get iotlb [0x1ffffe, 0x200000) > - request 0x1ffffd, cache miss, will get iotlb [0x1ffffd, 0x200000) > - ...
We pass an offset too, do we not? So callee can figure out the region starts at 0x0 and avoid 2nd and 3rd misses. > We'll all cache miss along the way until we access 0x0. While if we > are with page mask, we'll get: > > - request 0x1fffff, cache miss, will get iotlb [0x0, 0x200000) > - request 0x1ffffe, cache hit > - request 0x1ffffd, cache hit > - ... > > We'll only miss at the first IO. I think we should send as much info as we can. There should be a way to find full region start and length. > -- > Peter Xu