@@ -14,6 +14,9 @@
/* Skylake: MOCS is now an index into an array of 62 different caching
* configurations programmed by the kernel.
*/
+
+/* TC=PTE, LeCC=PTE, LRUM=3, L3CC=WB */
+#define SKL_MOCS_PTE_PTE (3 << 1)
/* TC=LLC/eLLC, LeCC=WB, LRUM=3, L3CC=WB */
#define SKL_MOCS_WB (2 << 1)
/* TC=LLC/eLLC, LeCC=PTE, LRUM=3, L3CC=WB */
@@ -26,6 +29,9 @@ brw_mocs_get_control_state(const struct brw_context *brw,
switch (brw->gen) {
default:
case 9:
+ if (brw->intelScreen->mocs_version > 1)
+ return SKL_MOCS_PTE_PTE;
+
return type == INTEL_MOCS_PTE ? SKL_MOCS_PTE : SKL_MOCS_WB;
case 8:
return type == INTEL_MOCS_PTE ? BDW_MOCS_PTE : BDW_MOCS_WB;
Copying the kernel commit message: Starting with GEN9, Memory Object Control State (MOCS) becomes an index into a table as opposed to the direct programming within the command. The table has 62 usable entries (ie 6 bits can represent all settings), and each buffer type may use one of these 62 entries to describe cacheability type, and age (and some other less useful fields). Because we hadn't dealt with MOCS settings like this, we didn't think ahead too well and have ended up with a mess for GEN9 (and soon GEN10) platform. The plan for for future platforms is that the ideal MOCS settings will be determined, defined, and written in the public PRMs. After this point, the i915.ko will absorb these settings and sometime afterwards flip the alpha switch. All driver releases without the final MOCS table must be considered alpha. Here on, userspace can assume the MOCS table is definitively done. There will be some reserved entries for 'oh shit' scenarios. This avoids versioning the MOCS table which leaves somewhat of a mess in userspace trying to handle arbitrarily many MOCS versions. But we do have a mess on GEN9. In the beginning, the MOCS table entries were pre-populated by the hardware based on estimations made prior to tapeout and we could just use that. Subsequently much performance tuning was done to determine optimal settings that the i915 driver should load on top of the hardware defaults. That was posted last as v6 of the original per-engine MOCS settings: https://patchwork.freedesktop.org/patch/53237/. Since the MOCS table is not context saved/restored, it isn't feasible to let userspace upload its own MOCS table. After a good amount of debate, it was decided that we'd utilize only the minimal set of entires in mesa anyway, and so we took only those entries for our MOCS entries. Now we've come to the realization that indeed there are other MOCS entries which are more optimal for various buffer types and workloads. The problem is that the meaning of the indices is ABI (we assume index 0 is the uncached entry, and that there are only 3 entries total). What this patch [simply] aims to do is expose a parameter to inform userspace which "version" of the table was loaded by i915. Upon sufficient data, new entries can be added, and the version can be bumped. For example, from my original mesa mocs branch: commit c9b0481bce24af032386701de0266eb5bc24e988 Author: Ben Widawsky <benjamin.widawsky@intel.com> Date: Fri Apr 8 10:21:16 2016 -0700 i965: Use PTE mocs Signed-off-by: Ben Widawsky <benjamin.widawsky@intel.com> tl;dr: A versioned MOCS table will allow userspace to be aware of new and potentially interesting cacheability settings. Next GEN platforms will not be considered production worthy until the MOCS table is finalized. Ben Widawsky (1): drm/i915: Version the MOCS settings drivers/gpu/drm/i915/i915_drv.c | 3 +++ drivers/gpu/drm/i915/i915_drv.h | 2 ++ drivers/gpu/drm/i915/i915_pci.c | 13 +++++++++---- include/uapi/drm/i915_drm.h | 8 ++++++++ 4 files changed, 22 insertions(+), 4 deletions(-) Ben Widawsky (2): intel: Merge latest i915 uapi intel: Make driver aware of MOCS table version src/intel/drm/i915_drm.h | 8 ++++++++ src/intel/vulkan/anv_device.c | 12 ++++++++++++ src/intel/vulkan/anv_private.h | 2 ++ src/mesa/drivers/dri/i915/intel_context.c | 7 ++++++- src/mesa/drivers/dri/i965/intel_screen.c | 14 ++++++++++++++ src/mesa/drivers/dri/i965/intel_screen.h | 2 ++ 6 files changed, 44 insertions(+), 1 deletion(-)