From: Rafael Barbalho <rafael.barbalho@xxxxxxxxx> This particular memory leak, if I can call it that, shows itself when i915 is in true PPGTT mode and you share a buffer object to another hardware context using flink. In the failure case harware context A creates an object & does some rendering to it, in turn mapping to its PPGTT, flinks it and then shares the object with hardware context B. Hardware context B then does some rendering operation with the share object, adding a VMA to its PPGTT address space, but it's eventually exited by the user. Because i915 doesn't clean-up an object's VMAs until the object is destroyed all the PPGTT memory allocations for hardware context B will still be kept a live because an object is still referencing it. When you repeat this sharing and re-using multiple times the system eventually runs out of memory because of all these PPGTT memory allocations for old contexts are still hanging around but will actually never be used again. I am also not seeing the shrinker coming in to reap this object because it is active in another hardware contexts. This naive attempt at fixing the issue is to clean-up the PPGTT entries when the context is destroyed. Patch 1 is what I used to help me track the issue and see the VM leak, patch 2 is the naive fix. Cc: Daniel Vetter <daniel@xxxxxxxx> Cc: Jon Bloomfield <jon.bloomfield@xxxxxxxxx> Rafael Barbalho (2): drm/i915: Export active PPGTTs in debugfs drm/i915: Clean-up PPGTT on context destruction drivers/gpu/drm/i915/i915_debugfs.c | 41 +++++++++++++++++++++---------- drivers/gpu/drm/i915/i915_gem.c | 7 +++--- drivers/gpu/drm/i915/i915_gem_context.c | 2 +- drivers/gpu/drm/i915/i915_gem_gtt.c | 43 ++++++++++++++++++++++++++++++++- drivers/gpu/drm/i915/i915_gem_gtt.h | 7 ++++++ 5 files changed, 82 insertions(+), 18 deletions(-) -- 2.3.0 _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx http://lists.freedesktop.org/mailman/listinfo/intel-gfx
