| Message ID | 1449844453-33974-1-git-send-email-michel.thierry@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, Dec 11, 2015 at 02:34:13PM +0000, Michel Thierry wrote: > We detected if objects should be moved to the lower parts when 48-bit > support flag was not set, but not the other way around. > > This handles the case in which an object was allocated in the 32-bit > address range, but it has been marked as safe to move above it, which > theoretically would help to keep the lower addresses available for > objects which really need to be there. > > Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> > Signed-off-by: Michel Thierry <michel.thierry@intel.com> No. This is not lazy. When we run out of low space, we evict. Until then don't cause extra work for no reason. -Chris
On Fri, Dec 11, 2015 at 02:49:52PM +0000, Chris Wilson wrote: > On Fri, Dec 11, 2015 at 02:34:13PM +0000, Michel Thierry wrote: > > We detected if objects should be moved to the lower parts when 48-bit > > support flag was not set, but not the other way around. > > > > This handles the case in which an object was allocated in the 32-bit > > address range, but it has been marked as safe to move above it, which > > theoretically would help to keep the lower addresses available for > > objects which really need to be there. > > > > Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> > > Signed-off-by: Michel Thierry <michel.thierry@intel.com> > > No. This is not lazy. When we run out of low space, we evict. Until then > don't cause extra work for no reason. Yeah, this stuff should just work. First the eviction code should kick stuff out, and if we totally deadlock then we'll retry with everything placed nicely. Long-term objects should segregate (assuming you're not mixing them up badly in the userspace cache). How did this come up? I think there's a more in-depth story to be shared here, with some perf data to illustrate it ... -Daniel
On 12/11/2015 6:57 PM, Daniel Vetter wrote: > On Fri, Dec 11, 2015 at 02:49:52PM +0000, Chris Wilson wrote: >> On Fri, Dec 11, 2015 at 02:34:13PM +0000, Michel Thierry wrote: >>> We detected if objects should be moved to the lower parts when 48-bit >>> support flag was not set, but not the other way around. >>> >>> This handles the case in which an object was allocated in the 32-bit >>> address range, but it has been marked as safe to move above it, which >>> theoretically would help to keep the lower addresses available for >>> objects which really need to be there. >>> >>> Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> >>> Signed-off-by: Michel Thierry <michel.thierry@intel.com> >> >> No. This is not lazy. When we run out of low space, we evict. Until then >> don't cause extra work for no reason. > > Yeah, this stuff should just work. First the eviction code should kick > stuff out, and if we totally deadlock then we'll retry with everything > placed nicely. Long-term objects should segregate (assuming you're not > mixing them up badly in the userspace cache). > > How did this come up? I think there's a more in-depth story to be shared > here, with some perf data to illustrate it ... Hi, It came from some local testing; Daniele saw bo's with the support flag enabled staying in the 32-bit range (the test was changing the flag between submissions). If there's no space constraints, re-enabling the flag won't relocate them, only when evict is required as you said... and that's not really an issue. Sorry for the noise. -Michel
On Mon, Dec 14, 2015 at 10:18:02AM +0000, Michel Thierry wrote: > On 12/11/2015 6:57 PM, Daniel Vetter wrote: > >On Fri, Dec 11, 2015 at 02:49:52PM +0000, Chris Wilson wrote: > >>On Fri, Dec 11, 2015 at 02:34:13PM +0000, Michel Thierry wrote: > >>>We detected if objects should be moved to the lower parts when 48-bit > >>>support flag was not set, but not the other way around. > >>> > >>>This handles the case in which an object was allocated in the 32-bit > >>>address range, but it has been marked as safe to move above it, which > >>>theoretically would help to keep the lower addresses available for > >>>objects which really need to be there. > >>> > >>>Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> > >>>Signed-off-by: Michel Thierry <michel.thierry@intel.com> > >> > >>No. This is not lazy. When we run out of low space, we evict. Until then > >>don't cause extra work for no reason. > > > >Yeah, this stuff should just work. First the eviction code should kick > >stuff out, and if we totally deadlock then we'll retry with everything > >placed nicely. Long-term objects should segregate (assuming you're not > >mixing them up badly in the userspace cache). > > > >How did this come up? I think there's a more in-depth story to be shared > >here, with some perf data to illustrate it ... > > Hi, > > It came from some local testing; Daniele saw bo's with the support flag > enabled staying in the 32-bit range (the test was changing the flag between > submissions). > > If there's no space constraints, re-enabling the flag won't relocate them, > only when evict is required as you said... and that's not really an issue. Hm, it might make sense for validation to catch userspace bugs where it mislabels bos wrt 48bit. But then unit tests (at least with mesa/piglit) destroy the gpu context (well entire executable) for every test, so false sharing like this shouldn't ever happen and we should be able to catch issues reliably. So I think we don't need some special debugfs tunable to kick the kernel into validation mode for this case. And if you absolutely want it we already have the evict interface to forcefully throw out everything and provoke fresh bindings. Not cheap, but should still work for at least validating the logic. > Sorry for the noise. No worries, just figured there's a bigger story worth sharing here. Validation noticing this makes sense. -Daniel
diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c index 8df5b96..a83916e 100644 --- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c +++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c @@ -696,6 +696,11 @@ eb_vma_misplaced(struct i915_vma *vma) (vma->node.start + vma->node.size - 1) >> 32) return true; + /* keep the lower addresses free of unnecessary objects */ + if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) && + !((vma->node.start + vma->node.size - 1) >> 32)) + return true; + return false; }
We detected if objects should be moved to the lower parts when 48-bit support flag was not set, but not the other way around. This handles the case in which an object was allocated in the 32-bit address range, but it has been marked as safe to move above it, which theoretically would help to keep the lower addresses available for objects which really need to be there. Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Signed-off-by: Michel Thierry <michel.thierry@intel.com> --- drivers/gpu/drm/i915/i915_gem_execbuffer.c | 5 +++++ 1 file changed, 5 insertions(+)