On Mon, 2008-09-22 at 10:59 -0700, Jesse Barnes wrote: > On Wednesday, August 27, 2008 5:22 pm Nick Piggin wrote: > > > However, when other DRM drivers get around to doing memory management, > > > I'm sure they'll also be interested in an ioremap_wc that doesn't eat > > > ipi costs. For us, the ipis for flushing were eating over 10% of CPU > > > time. If your patch series cuts that cost, we could drop this piece at > > > that point. > > > > It can cut the cost quite significantly on normal vmap/vunmap loads I > > tested. Whether it will work as well on your workload, I don't know > > but I would have liked to find out. I raised this issue quite a while > > back, so I'm disappointed it had not been tried... > > I think Eric has code with the vmap changes now? > > Given our discussions at KS/Plumbers would you be ok with acking these > patches? Or do you want a repost so you can check out the vmap stuff? > > After talking a bit more about it, I think we agreed that the ioctl interface > is actually a better approach then trying to shoehorn this stuff into system > calls, so aside from the vmap code (which could be done in 2.6.29 or whenever > the vmap stuff lands) I think this patchset is pretty close to what we want > in drm-next now... I've been trying to get a good test of the vmap changes. Unfortunately, it looks like things have changed in ioremap in the intervening time between when I last tested and now. I was taking 2.6.27-rc5-mm1 (since that was the only git tree I found with Nick's changes in it, unfortunately) and merging our stuff into there then reverting the kmap_atomic_prot_pfn bit. However, we've moved from 10% cost in unmapping due to IPIing to a >30% cost in mapping due to change_page_attr. This is regardless of whether I do ioremap or ioremap_wc. The physical range is covered by a WC MTRR, and the X Server's mapping the memory using the _wc resource. In the DRM, I just tried moving every ioremap to _wc, and it's the same. The call chain looks like i915_gem_pwrite_ioctl (60%) ioremap_wc (39%) ioremap_nocache (39%) ioremap_caller (39%) ioremap_change_attr (36%) _set_memory_uc (36%) change_page_attr_set_clr (36%) vm_unmap_aliases (31%) Maybe if I go back and generate a tree of just Nick's changes and try to merge them forward I can get a good test. However, it looks to me like we've got some serious brokenness in ioremap and attribute handling coming up. -- Eric Anholt eric@xxxxxxxxxx eric.anholt@xxxxxxxxx
Attachment:
signature.asc
Description: This is a digitally signed message part
