[Intel-gfx] [PATCH] drm/i915: avoid concurrent writes to aux_inv
Yang, Fei
fei.yang at intel.com
Fri Mar 18 18:12:11 UTC 2022
>> static u32 *gen12_emit_aux_table_inv(const i915_reg_t inv_reg, u32 *cs)
>> {
>> *cs++ = MI_LOAD_REGISTER_IMM(1);
>> @@ -296,7 +272,7 @@ int gen12_emit_flush_xcs(struct i915_request *rq, u32 mode)
>> if (!HAS_FLAT_CCS(rq->engine->i915)) {
>> aux_inv = rq->engine->mask & ~BIT(BCS0);
>> if (aux_inv)
>> - cmd += 2 * hweight32(aux_inv) + 2;
>> + cmd += 4;
>> }
>> }
>>
>> @@ -329,14 +305,16 @@ int gen12_emit_flush_xcs(struct i915_request *rq, u32 mode)
>> *cs++ = 0; /* value */
>>
>> if (aux_inv) { /* hsdes: 1809175790 */
>> - struct intel_engine_cs *engine;
>> - unsigned int tmp;
>> -
>> - *cs++ = MI_LOAD_REGISTER_IMM(hweight32(aux_inv));
>> - for_each_engine_masked(engine, rq->engine->gt, aux_inv, tmp) {
>> - *cs++ = i915_mmio_reg_offset(aux_inv_reg(engine));
>> - *cs++ = AUX_INV;
>> + *cs++ = MI_LOAD_REGISTER_IMM(1) | MI_LRI_MMIO_REMAP_EN;
>
> Cool, I didn't know this exists. First Bspec link I found did not mention these register, but second did.
> That one however (29545) has a worrying "removed by" tag which seems to point to a story suggesting the
> remapping table might be gone on machines with flat ccs?! Could you double check please?
The variable aux_inv is set only if (!HAS_FLAT_CCS(rq->engine->i915)).
>> + if (rq->engine->class == VIDEO_DECODE_CLASS) {
>> + *cs++ = i915_mmio_reg_offset(GEN12_VD0_AUX_NV);
>> + } else if (rq->engine->class == VIDEO_ENHANCEMENT_CLASS) {
>> + *cs++ = i915_mmio_reg_offset(GEN12_VE0_AUX_NV);
>> + } else {
>> + GEM_BUG_ON("unknown aux_inv reg\n");
>> + *cs++ = i915_mmio_reg_offset(INVALID_MMIO_REG);
>
> I'd consider not emitting the LRI if we don't know what to put in unless there is some hidden point to do it?
That's true. I was following the original logic flow here. I think it would be better to check for engine class before setting the variable aux_inv.
>
>> }
>> + *cs++ = AUX_INV;
>> *cs++ = MI_NOOP;
>> }
>>
>> diff --git a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
>> b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
>> index d112ffd56418..2d150eec5c65 100644
>> --- a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
>> +++ b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
>> @@ -144,6 +144,7 @@
>> #define MI_LOAD_REGISTER_IMM(x) MI_INSTR(0x22, 2*(x)-1)
>> /* Gen11+. addr = base + (ctx_restore ? offset & GENMASK(12,2) : offset) */
>> #define MI_LRI_LRM_CS_MMIO REG_BIT(19)
>> +#define MI_LRI_MMIO_REMAP_EN (1 << 17)
>> #define MI_LRI_FORCE_POSTED (1<<12)
>
> Passing observation - three bits, three flavours of expressing them, sigh...
Haha, REG_BIT(17) it is. The other one causes a CHECK:SPACING, but don't want to touch that in this patch.
>> #define MI_LOAD_REGISTER_IMM_MAX_REGS (126)
>> #define MI_STORE_REGISTER_MEM MI_INSTR(0x24, 1)
More information about the Intel-gfx
mailing list