On Fri, Sep 17, 2021 at 09:54:21PM -0700, Alan Previn wrote:
This series adds gem_pxp tests for the new PXP subsystem currently being reviewed at https://patchwork.freedesktop.org/series/90504/. This series currently includes 4 groups of tests addressing the features and restrictions described by Daniele's series :
- test i915 interfaces for allocation of protected bo's and contexts and enforcement of UAPI rule disallowing the modification of parameters after it's been created.
- verify PXP subsystem protected sessions generate encrypted content on protected output buffers and decrypt protected inputs buffers.
- verify i915 PXP auto-teardown succeeds on suspend-resume cycles and gem-exec of stale protected assets fail. Ensure protected-contexts adhere to stricter invalidation enforcement upon teardown event.
- Ensure that display plane decryption works as expected with protected buffers.
NOTE: This series is on the tenth revision. All R-v-b's have been received except UAPI patch which will be dropped at merge time (after kernel patches gets merged and igt's drm UAPI gets sync'd).
Changes from prior rev1 to now: v11: - When detecting hw support, retry pxp context creation multiple times a timeout as per HW SLA. - initialize bo or ctx handles to zero before calling creation ioctl wrapper.
Acked-by: Rodrigo Vivi rodrigo.vivi@intel.com
v10: - In patch #2, reuse existing gem_create_ext wrapper. - In patch #10, kernel side changed the debugfs file name (but no difference in behavior / usage). - Removed patch #14 from Rev9 as decision on kernel side was to drop the usage of RESET_STATS IOCTL to track invalidated pxp contexts. v9: - Remove patch #2 from rev7 as it was duplicating an existing ioctl wrapper helper - Fix the false-negative warnings when triggering auto-suspend-resume (remove checking if we are suspending after the system has already resumed). v8: - Nothing - mistaken detection from patchwork v7: - In prior rev, Patches #11->13 was testing expected results from calling gem_execbuf with stale pxp-context, pxp-buffer or combinations of them (including an opt-out usage). All of them used a single suspend-resume power state cycles to trigger the PXP teardown event. These patches have been combined into patch #14 that continues to carry the prior rev Rvb. - In its place, the new patches of #11->#13 do the identical set of tests as before (results from gem_execbuf with various combinations of stale pxp context and buffer), but this time using a debugfs file handle that triggers the same code path taken when the HW triggers the pxp teardown. That said, the code is nearly identical as v6 but I did not keep the Rvb's. - In patch #15, RESET_STAT now reports invalidated / banned pxp contexts via the existing batch_active's lost count. v6: - Addressed rev5 review comments for patch #1, #7, #14 and #17. - For #17, I'm using Rodrigo's Rv-b because offline discussions concluded that we couldn't use those test sequences with HDCP and so it was removed it. - Added Rv-b into all patches that received it. - Modified the test requirement from a list of device ids to checking if runtime PXP interface succeeds due to kernel's build config dependency. v5: - Addressed all rev4 review comments. No changes to overall flow and logic compared to the last rev. v4: - Addressed all rev3 review comments. NOTE: that all test cases and code logic are the same but a decent amount of refactoring has occured due to address v3 comments to break out subtests into separate functions while combining certain checks into the same function to reduce test time by minimizing number of suspend-resume power cycles. v3: - Addressed all rev2 review comments. - In line with one of the rev2 comments, a thorough fixup of all line-breaks in function calls was made for a more consistent styling. - Rebased on igt upstream repo and updated to latest kernel UAPI that added GEM_CREATE_EXT. v2: - Addressed all rev1 review comments except these: 1.Chris Wilson : "...have the caller do 1-3 once for its protected context. Call it something like intel_bb_enable_pxp(), intel_bb_set_pxp if it should be reversible.". - This couldn't be implemented because [1] HW needs different instruction sequences for enabling/disabling PXP depending on the engine class and [2] the pair of "pxp-enable" and "pxp- disable" instructions need to be contained within the same batch that is dispatched to the hardware. That said, implementing internal intel_batchbuffer funtionality for this would conflict with how rendercopy_gen9 uses batch buffer memory by repositioing the pointer and consuming unused portions of the batch buffer as 3d state offsets that batchbuffer has no visibility.
- Added these additional subtests: 1. verify that buffer sharing works across testing pxp context. 2. verify teardown bans contexts via DRM_IOCTL_I915_GET_RESET_STAT. 3. verify display plane decryption of protected buffers.Alan Previn (14): Add PXP UAPI support in i915_drm.h Add basic PXP testing of buffer and context alloc Perform a regular 3d copy as a control checkpoint Add PXP attribute support in batchbuffer and buffer_ops libs Add MI_SET_APPID instruction definition Enable protected session cmd in gen12_render_copyfunc Add subtest to copy raw source to protected dest Add test where both src and dest are protected Verify PXP teardown occurred through suspend-resume Verify execbuf fails with stale PXP context after teardown Verify execbuf fails with stale PXP buffer after teardown Verify execbuf ok with stale PXP buf in opt-out use Verify execution behavior with stale PXP assets through suspend-resume Verify protected surfaces are dma buffer sharable
Karthik B S (1): tests/i915_pxp: CRC validation for display tests.
lib/i915/i915_drm_local.h | 102 +++ lib/intel_batchbuffer.c | 23 +- lib/intel_batchbuffer.h | 31 + lib/intel_bufops.h | 15 + lib/intel_reg.h | 8 + lib/rendercopy_gen9.c | 57 ++ tests/i915/gem_pxp.c | 1283 +++++++++++++++++++++++++++++++++++++ tests/meson.build | 1 + 8 files changed, 1519 insertions(+), 1 deletion(-) create mode 100644 tests/i915/gem_pxp.c
-- 2.25.1