So apparently under ridiculous amounts of memory pressure we can get into trouble in do_switch when we try to move the old hw context backing storage object onto the active lists.
With list debugging enabled that usually results in us chasing a poisoned pointer - which means we've hit upon a vma that has been removed from all lrus with list_del (and then deallocated, so it's a real use-after free). Ian Lister has done some great callchain chasing and noticed that we can reenter do_switch: i915_gem_do_execbuffer() i915_switch_context() do_switch() from = ring->last_context; i915_gem_object_pin() i915_gem_object_bind_to_gtt() ret = drm_mm_insert_node_in_range_generic(); // If the above call fails then it will try i915_gem_evict_something() // If that fails it will call i915_gem_evict_everything() ... i915_gem_evict_everything() i915_gpu_idle() i915_switch_context(DEFAULT_CONTEXT) Like with everything else where the shrinker or eviction code can invalidate pointers we need to reload relevant state. Note that there's no need to recheck whether a context switch is still required because: - Doing a switch to the same context is harmless (besides wasting a bit of energy). - This can only happen with the default context. But since that one's pinned we'll never call down into evict_everything under normal circumstances. Note that there's a little driver bringup fun involved namely that we could recourse into do_switch for the initial switch. Atm we're fine since we assign the context pointer only after the call to do_switch at driver load or resume time. And in the gpu reset case we skip the entire setup sequence (which might be a bug on its own, but definitely not this one here). Cc'ing stable since apparently ChromeOS guys are seeing this in the wild (and not just on artificial stress tests), see the reference. Note that in upstream code doesn't calle evict_everything directly from evict_something, that's an extension in this product branch. But we can still hit upon this bug (and apparently we do, see the linked backtraces). I've noticed this while trying to construct a testcase for this bug and utterly failed to provoke it. It looks like we need to driver the system squarly into the lowmem wall and provoke the shrinker to evict the context object by doing the last-ditch evict_everything call. Aside: There's currently no means to get a badly-fragmenting hw context object away from a bad spot in the upstream code. We should fix this by at least adding some code to evict_something to handle hw contexts. References: https://code.google.com/p/chromium/issues/detail?id=248191 Reported-by: Ian Lister <ian.lis...@intel.com> Cc: Ian Lister <ian.lis...@intel.com> Cc: sta...@vger.kernel.org Cc: Ben Widawsky <benjamin.widaw...@intel.com> Cc: Stéphane Marchesin <marc...@chromium.org> Cc: Bloomfield, Jon <jon.bloomfi...@intel.com> Signed-off-by: Daniel Vetter <daniel.vet...@ffwll.ch> --- drivers/gpu/drm/i915/i915_gem_context.c | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/i915/i915_gem_context.c b/drivers/gpu/drm/i915/i915_gem_context.c index 41877045a1a0..2d2877493f61 100644 --- a/drivers/gpu/drm/i915/i915_gem_context.c +++ b/drivers/gpu/drm/i915/i915_gem_context.c @@ -421,11 +421,21 @@ static int do_switch(struct i915_hw_context *to) if (ret) return ret; - /* Clear this page out of any CPU caches for coherent swap-in/out. Note + /* + * Pin can switch back to the default context if we end up calling into + * evict_everything - as a last ditch gtt defrag effort that also + * switches to the default context. Hence we need to reload from here. + */ + from = ring->last_context; + + /* + * Clear this page out of any CPU caches for coherent swap-in/out. Note * that thanks to write = false in this call and us not setting any gpu * write domains when putting a context object onto the active list * (when switching away from it), this won't block. - * XXX: We need a real interface to do this instead of trickery. */ + * + * XXX: We need a real interface to do this instead of trickery. + */ ret = i915_gem_object_set_to_gtt_domain(to->obj, false); if (ret) { i915_gem_object_unpin(to->obj); -- 1.8.4.3 _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx