drm/i915: Suppress switch_mm emission between the same aliasing_ppgtt
authorChris Wilson <chris@chris-wilson.co.uk>
Sat, 12 Aug 2017 15:27:24 +0000 (16:27 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Sat, 12 Aug 2017 15:50:12 +0000 (16:50 +0100)
When switching between contexts using the aliasing_ppgtt, the VM is
shared. We don't need to reload the PD registers unless they are dirty.

Martin Peres reported an issue that looks like corruption between
Haswell context switches, bisecting to commit f9326be5f1d3 ("drm/i915:
Rearrange switch_context to load the aliasing ppgtt on first use").
Switching between the same mm (the aliasing_ppgtt is used for all
contexts in this case) should be a nop, but appears to trigger some
side-effects in the context switch. However, as we know the switch
is redundant in this case, we can skip it and continue to ignore the
issue until somebody feels strong enough to investigate full-ppgtt on
gen7 again!

Except.. Martin was using full-ppgtt which is not supported as it
doesn't work correctly yet. So whilst the bisect did yield valuable
information about the failures, the fix should not have any user impact
under default settings, with the exception of a slightly lower
throughput on xcs as the VM would always be reloaded.

v2: Also remember to set the legacy_active_context following the switch
on xcs (commit e8a9c58fcd9a ("drm/i915: Unify active context tracking
between legacy/execlists/guc"))

Fixes: f9326be5f1d3 ("drm/i915: Rearrange switch_context to load the aliasing ppgtt on first use")
Fixes: e8a9c58fcd9a ("drm/i915: Unify active context tracking between legacy/execlists/guc")
Reported-by: Martin Peres <martin.peres@linux.intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Martin Peres <martin.peres@linux.intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170812152724.6883-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem_context.c

index ed91ac8ca832c3f030b4280755f10efdf62e0650..86ac74a8a5b2693ceb6f935bf0a20a7c1b0b1c72 100644 (file)
@@ -741,19 +741,19 @@ static inline bool skip_rcs_switch(struct i915_hw_ppgtt *ppgtt,
 }
 
 static bool
-needs_pd_load_pre(struct i915_hw_ppgtt *ppgtt,
-                 struct intel_engine_cs *engine,
-                 struct i915_gem_context *to)
+needs_pd_load_pre(struct i915_hw_ppgtt *ppgtt, struct intel_engine_cs *engine)
 {
+       struct i915_gem_context *from = engine->legacy_active_context;
+
        if (!ppgtt)
                return false;
 
        /* Always load the ppgtt on first use */
-       if (!engine->legacy_active_context)
+       if (!from)
                return true;
 
        /* Same context without new entries, skip */
-       if (engine->legacy_active_context == to &&
+       if ((!from->ppgtt || from->ppgtt == ppgtt) &&
            !(intel_engine_flag(engine) & ppgtt->pd_dirty_rings))
                return false;
 
@@ -797,7 +797,7 @@ static int do_rcs_switch(struct drm_i915_gem_request *req)
        if (skip_rcs_switch(ppgtt, engine, to))
                return 0;
 
-       if (needs_pd_load_pre(ppgtt, engine, to)) {
+       if (needs_pd_load_pre(ppgtt, engine)) {
                /* 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
@@ -894,7 +894,7 @@ int i915_switch_context(struct drm_i915_gem_request *req)
                struct i915_hw_ppgtt *ppgtt =
                        to->ppgtt ?: req->i915->mm.aliasing_ppgtt;
 
-               if (needs_pd_load_pre(ppgtt, engine, to)) {
+               if (needs_pd_load_pre(ppgtt, engine)) {
                        int ret;
 
                        trace_switch_mm(engine, to);
@@ -905,6 +905,7 @@ int i915_switch_context(struct drm_i915_gem_request *req)
                        ppgtt->pd_dirty_rings &= ~intel_engine_flag(engine);
                }
 
+               engine->legacy_active_context = to;
                return 0;
        }