| Message ID | 1410534601-11523-1-git-send-email-michel.thierry@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, Sep 12, 2014 at 04:10:01PM +0100, Michel Thierry wrote: > From: Ben Widawsky <benjamin.widawsky@intel.com> > > The simple explanation is, the docs say to do this for GEN8. Perhaps we > want to do this for GEN7 too, I am not certain. > > PDPs are saved and restored with context. Contexts (without execlists) > only exist on the render ring. The docs say that PDPs are not power > context save/restored. I've learned that this actually means something > which SW doesn't care about. So pretend the statement doesn't exist. > For non RCS, nothing changes. > > All this patch now does is change the ordering of LRI vs MI_SET_CONTEXT > for the initialization of the context. I do this because the docs say to > do it, and frankly, I cannot reason why it is necessary. I've thought > about it a lot, and tried, without success, to get a reason from design. > The answer I got more or less says, "gen7 is different than gen8." I've > given up, and am adding this little bit of code to make it in sync with > the docs. > > v2: Completely rewritten commit message that addresses the requests > Ville made for v1 > Only load PDPs for initial context load (Ville) > > v3: Rebased after ppgtt clean-up rules, and apply only for render > ring. This is needed to boot to desktop with full ppgtt in legacy mode > (without execlists). > > Signed-off-by: Ben Widawsky <ben@bwidawsk.net> > Signed-off-by: Michel Thierry <michel.thierry@intel.com> (v3) > --- > drivers/gpu/drm/i915/i915_gem_context.c | 31 +++++++++++++++++++++++++++++-- > 1 file changed, 29 insertions(+), 2 deletions(-) > > diff --git a/drivers/gpu/drm/i915/i915_gem_context.c b/drivers/gpu/drm/i915/i915_gem_context.c > index a5221d8..faebbf3 100644 > --- a/drivers/gpu/drm/i915/i915_gem_context.c > +++ b/drivers/gpu/drm/i915/i915_gem_context.c > @@ -522,6 +522,8 @@ static int do_switch(struct intel_engine_cs *ring, > struct intel_context *from = ring->last_context; > u32 hw_flags = 0; > bool uninitialized = false; > + bool needs_pd_load_rcs = (INTEL_INFO(ring->dev)->gen < 8) && to->ppgtt; > + bool needs_pd_load_xcs = (ring != &dev_priv->ring[RCS]) && to->ppgtt; > int ret, i; > > if (from != NULL && ring == &dev_priv->ring[RCS]) { > @@ -547,7 +549,11 @@ static int do_switch(struct intel_engine_cs *ring, > */ > from = ring->last_context; > > - if (to->ppgtt) { > + if (needs_pd_load_rcs || needs_pd_load_xcs) { > + /* Older GENs and non render rings still want the load first, > + * "PP_DCLV followed by PP_DIR_BASE register through Load > + * Register Immediate commands in Ring Buffer before submitting > + * a context."*/ > ret = to->ppgtt->switch_mm(to->ppgtt, ring); > if (ret) > goto unpin_out; > @@ -577,13 +583,34 @@ static int do_switch(struct intel_engine_cs *ring, > vma->bind_vma(vma, to->legacy_hw_ctx.rcs_state->cache_level, GLOBAL_BIND); > } > > - if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to)) > + if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to)) { > hw_flags |= MI_RESTORE_INHIBIT; > + needs_pd_load_rcs = to->ppgtt && IS_GEN8(ring->dev); > + } > > ret = mi_set_context(ring, to, hw_flags); > if (ret) > goto unpin_out; > > + /* GEN8 does *not* require an explicit reload if the PDPs have been > + * setup, and we do not wish to move them. > + * > + * XXX: If we implemented page directory eviction code, this > + * optimization needs to be removed. > + */ The comment seems a bit misplaced. Would seem more appropriate around where we derive needs_pd_load_rcs. > + if (needs_pd_load_rcs) { > + ret = to->ppgtt->switch_mm(to->ppgtt, ring); Aren't we now loading both before _and_ after on <=gen7 (except when uninitialized or default ctx is used)? Maybe the variables should be called needs_pd_load_pre and needs_pd_load_post or something? The current approach just seems somehow confusing to me. > + /* The hardware context switch is emitted, but we haven't > + * actually changed the state - so it's probably safe to bail > + * here. Still, let the user know something dangerous has > + * happened. > + */ > + if (ret) { > + DRM_ERROR("Failed to change address space on context switch\n"); > + goto unpin_out; > + } > + } > + > for (i = 0; i < MAX_L3_SLICES; i++) { > if (!(to->remap_slice & (1<<i))) > continue; > -- > 2.0.3 > > _______________________________________________ > Intel-gfx mailing list > Intel-gfx@lists.freedesktop.org > http://lists.freedesktop.org/mailman/listinfo/intel-gfx
On Fri, Sep 12, 2014 at 06:35:10PM +0300, Ville Syrjälä wrote: > On Fri, Sep 12, 2014 at 04:10:01PM +0100, Michel Thierry wrote: > > From: Ben Widawsky <benjamin.widawsky@intel.com> > > > > The simple explanation is, the docs say to do this for GEN8. Perhaps we > > want to do this for GEN7 too, I am not certain. > > > > PDPs are saved and restored with context. Contexts (without execlists) > > only exist on the render ring. The docs say that PDPs are not power > > context save/restored. I've learned that this actually means something > > which SW doesn't care about. So pretend the statement doesn't exist. > > For non RCS, nothing changes. Hang on. This is exactly what I was worried about... Which PDPs are saved with context? Why doesn't this mean that we aren't overwitting PDPs in use on other rings when RCS loads a new context? -Chris
On 9/12/2014 4:40 PM, Chris Wilson wrote: > On Fri, Sep 12, 2014 at 06:35:10PM +0300, Ville Syrjälä wrote: >> On Fri, Sep 12, 2014 at 04:10:01PM +0100, Michel Thierry wrote: >>> From: Ben Widawsky <benjamin.widawsky@intel.com> >>> >>> The simple explanation is, the docs say to do this for GEN8. Perhaps we >>> want to do this for GEN7 too, I am not certain. >>> >>> PDPs are saved and restored with context. Contexts (without execlists) >>> only exist on the render ring. The docs say that PDPs are not power >>> context save/restored. I've learned that this actually means something >>> which SW doesn't care about. So pretend the statement doesn't exist. >>> For non RCS, nothing changes. > > Hang on. This is exactly what I was worried about... Which PDPs are > saved with context? Why doesn't this mean that we aren't overwitting > PDPs in use on other rings when RCS loads a new context? > -Chris > I was wondering the same when I read Ben's original comment. Could this be GEN8 specific?
On Fri, Sep 12, 2014 at 05:56:30PM +0100, Michel Thierry wrote: > On 9/12/2014 4:40 PM, Chris Wilson wrote: > > On Fri, Sep 12, 2014 at 06:35:10PM +0300, Ville Syrjälä wrote: > >> On Fri, Sep 12, 2014 at 04:10:01PM +0100, Michel Thierry wrote: > >>> From: Ben Widawsky <benjamin.widawsky@intel.com> > >>> > >>> The simple explanation is, the docs say to do this for GEN8. Perhaps we > >>> want to do this for GEN7 too, I am not certain. > >>> > >>> PDPs are saved and restored with context. Contexts (without execlists) > >>> only exist on the render ring. The docs say that PDPs are not power > >>> context save/restored. I've learned that this actually means something > >>> which SW doesn't care about. So pretend the statement doesn't exist. > >>> For non RCS, nothing changes. > > > > Hang on. This is exactly what I was worried about... Which PDPs are > > saved with context? Why doesn't this mean that we aren't overwitting > > PDPs in use on other rings when RCS loads a new context? > > -Chris > > > > I was wondering the same when I read Ben's original comment. Could this > be GEN8 specific? Yes. For earlier hardware it's stated that the RCS PP_DIR registers are saved in the power context when in ring buffer mode. For gen8 it's very unclear from the docs. If you look at the context image it shows the RCS PDPs in the execlist context. But if it's saving the execlist context even in ring buffer mode (which the earlier platforms didn't do) then it means that our GEN8_CXT_TOTAL_SIZE is wrong since it doesn't include the execlist context. I don't think anyone did for gen8 what I did for gen6/7 and actually look at the context after the hardware had saved it. That's how I figured out which parts actually get saved. As far as the other engines go, I think they must still save their PDPs in some power context since the execlist context only contains the RCS PDPs. Well, assuming the execlist context is really where this stuff is saved for RCS.
diff --git a/drivers/gpu/drm/i915/i915_gem_context.c b/drivers/gpu/drm/i915/i915_gem_context.c index a5221d8..faebbf3 100644 --- a/drivers/gpu/drm/i915/i915_gem_context.c +++ b/drivers/gpu/drm/i915/i915_gem_context.c @@ -522,6 +522,8 @@ static int do_switch(struct intel_engine_cs *ring, struct intel_context *from = ring->last_context; u32 hw_flags = 0; bool uninitialized = false; + bool needs_pd_load_rcs = (INTEL_INFO(ring->dev)->gen < 8) && to->ppgtt; + bool needs_pd_load_xcs = (ring != &dev_priv->ring[RCS]) && to->ppgtt; int ret, i; if (from != NULL && ring == &dev_priv->ring[RCS]) { @@ -547,7 +549,11 @@ static int do_switch(struct intel_engine_cs *ring, */ from = ring->last_context; - if (to->ppgtt) { + if (needs_pd_load_rcs || needs_pd_load_xcs) { + /* Older GENs and non render rings still want the load first, + * "PP_DCLV followed by PP_DIR_BASE register through Load + * Register Immediate commands in Ring Buffer before submitting + * a context."*/ ret = to->ppgtt->switch_mm(to->ppgtt, ring); if (ret) goto unpin_out; @@ -577,13 +583,34 @@ static int do_switch(struct intel_engine_cs *ring, vma->bind_vma(vma, to->legacy_hw_ctx.rcs_state->cache_level, GLOBAL_BIND); } - if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to)) + if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to)) { hw_flags |= MI_RESTORE_INHIBIT; + needs_pd_load_rcs = to->ppgtt && IS_GEN8(ring->dev); + } ret = mi_set_context(ring, to, hw_flags); if (ret) goto unpin_out; + /* GEN8 does *not* require an explicit reload if the PDPs have been + * setup, and we do not wish to move them. + * + * XXX: If we implemented page directory eviction code, this + * optimization needs to be removed. + */ + if (needs_pd_load_rcs) { + ret = to->ppgtt->switch_mm(to->ppgtt, ring); + /* The hardware context switch is emitted, but we haven't + * actually changed the state - so it's probably safe to bail + * here. Still, let the user know something dangerous has + * happened. + */ + if (ret) { + DRM_ERROR("Failed to change address space on context switch\n"); + goto unpin_out; + } + } + for (i = 0; i < MAX_L3_SLICES; i++) { if (!(to->remap_slice & (1<<i))) continue;