[Intel-gfx] [PATCH 2/2] drm/i915: Consolidate TLB invalidation flow

Andrzej Hajda andrzej.hajda at intel.com
Wed Dec 14 09:02:54 UTC 2022


On 13.12.2022 16:22, Tvrtko Ursulin wrote:
> 
> On 13/12/2022 14:52, Andrzej Hajda wrote:
>> On 13.12.2022 13:39, Tvrtko Ursulin wrote:
>>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>
>>> As the logic for selecting the register and corresponsing values 
>>> grew, the
>>> code become a bit unsightly. Consolidate by storing the required 
>>> values at
>>> engine init time in the engine itself, and by doing so minimise the 
>>> amount
>>> of invariant platform and engine checks during each and every TLB
>>> invalidation.
>>>
>>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>> Cc: Andrzej Hajda <andrzej.hajda at intel.com>
>>> ---
>>> I think this looks nicer, but I don't really want to merge any 
>>> flavour of
>>> consolidation until we get IGT coverage for the issue upstreamed.
>>
>>
>> Yep, the important is to have 1st patch merged.
> 
> Agreed.
> 
> 1)
> Can you send it standalone so it gets CI results and can get merged?

Done.

> 
> 2)
> Could you check internally where did gem_exec_tlb get stuck? >
>>> ---
>>>   drivers/gpu/drm/i915/gt/intel_engine_cs.c    |  81 ++++++++++++
>>>   drivers/gpu/drm/i915/gt/intel_engine_types.h |  16 +++
>>>   drivers/gpu/drm/i915/gt/intel_gt.c           | 126 ++++---------------
>>>   3 files changed, 120 insertions(+), 103 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/i915/gt/intel_engine_cs.c 
>>> b/drivers/gpu/drm/i915/gt/intel_engine_cs.c
>>> index 99c4b866addd..97cdd9853e38 100644
>>> --- a/drivers/gpu/drm/i915/gt/intel_engine_cs.c
>>> +++ b/drivers/gpu/drm/i915/gt/intel_engine_cs.c
>>> @@ -1143,6 +1143,84 @@ static int init_status_page(struct 
>>> intel_engine_cs *engine)
>>>       return ret;
>>>   }
>>> +static bool intel_engine_init_tlb_invalidation(struct 
>>> intel_engine_cs *engine)
>>> +{
>>> +    static const union intel_engine_tlb_inv_reg gen8_regs[] = {
>>> +        [RENDER_CLASS].reg        = GEN8_RTCR,
>>> +        [VIDEO_DECODE_CLASS].reg    = GEN8_M1TCR, /* , GEN8_M2TCR */
>>> +        [VIDEO_ENHANCEMENT_CLASS].reg    = GEN8_VTCR,
>>> +        [COPY_ENGINE_CLASS].reg        = GEN8_BTCR,
>>> +    };
>>> +    static const union intel_engine_tlb_inv_reg gen12_regs[] = {
>>> +        [RENDER_CLASS].reg        = GEN12_GFX_TLB_INV_CR,
>>> +        [VIDEO_DECODE_CLASS].reg    = GEN12_VD_TLB_INV_CR,
>>> +        [VIDEO_ENHANCEMENT_CLASS].reg    = GEN12_VE_TLB_INV_CR,
>>> +        [COPY_ENGINE_CLASS].reg        = GEN12_BLT_TLB_INV_CR,
>>> +        [COMPUTE_CLASS].reg        = GEN12_COMPCTX_TLB_INV_CR,
>>> +    };
>>> +    static const union intel_engine_tlb_inv_reg xehp_regs[] = {
>>> +        [RENDER_CLASS].mcr_reg          = XEHP_GFX_TLB_INV_CR,
>>> +        [VIDEO_DECODE_CLASS].mcr_reg      = XEHP_VD_TLB_INV_CR,
>>> +        [VIDEO_ENHANCEMENT_CLASS].mcr_reg = XEHP_VE_TLB_INV_CR,
>>> +        [COPY_ENGINE_CLASS].mcr_reg      = XEHP_BLT_TLB_INV_CR,
>>> +        [COMPUTE_CLASS].mcr_reg          = XEHP_COMPCTX_TLB_INV_CR,
>>> +    };
>>> +    struct drm_i915_private *i915 = engine->i915;
>>> +    const union intel_engine_tlb_inv_reg *regs;
>>> +    union intel_engine_tlb_inv_reg reg;
>>> +    unsigned int class = engine->class;
>>> +    unsigned int num = 0;
>>> +    u32 val;
>>> +
>>> +    if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) {
>>> +        regs = xehp_regs;
>>> +        num = ARRAY_SIZE(xehp_regs);
>>> +    } else if (GRAPHICS_VER(i915) == 12) {
>>> +        regs = gen12_regs;
>>> +        num = ARRAY_SIZE(gen12_regs);
>>> +    } else if (GRAPHICS_VER(i915) >= 8 && GRAPHICS_VER(i915) <= 11) {
>>
>> "GRAPHICS_VER(i915) <= 11" seems redundant.
> 
> Code movement... Fixes: "patch which added >= 12,50" to clean just that 
> part up. Before it was deliberately made to hit the below warn on so 
> that any new platform added requires a human checking if the registers 
> are still same and compatible. (Instead of assuming future platforms 
> remain compatible - which is IMO too big risk to take.)
> 
>>
>>> +        regs = gen8_regs;
>>> +        num = ARRAY_SIZE(gen8_regs);
>>> +    } else if (GRAPHICS_VER(i915) < 8) {
>>
>> ditto
> 
> How? We want to do nothing before gen 8.

} else {
	return false;
}

?

As I understand there are three 'rules' we want to follow:
1. start from the newest hw, end at the oldest.
2. warn for not yet supported platforms, lets assume it is ver >= 13.
3. cover all cases

Why not use >= consistently, till exhaustion:

if (ver >= 13) {
	warn
	return false;
} else if (ver >= 12.50) {
	...
} else if (ver >= 12) {
	...
} else if (ver >= 8) {
	...
} else {
	return false;
}

It looks cleaner, ordered, no redundant checks, no doubts, hopefuly 
future-proof.

Regards
Andrzej

> 
>>> +        return false;
>>> +    }
>>> +
>>> +    if (drm_WARN_ONCE(&i915->drm, !num,
>>> +              "Platform does not implement TLB invalidation!"))
>>> +        return false;
>>
>> It never happens.
> 
> As above a) it's just code movement and b) it would be safer if it could 
> happen.
> 
>>> +
>>> +    if (drm_WARN_ON_ONCE(&i915->drm,
>>> +                 class >= num ||
>>> +                 (!regs[class].reg.reg &&
>>> +                  !regs[class].mcr_reg.reg)))
>>> +        return false;
>>> +
>>> +    reg = regs[class];
>>> +
>>> +    if (GRAPHICS_VER(i915) == 8 && class == VIDEO_DECODE_CLASS) {
>>> +        reg.reg.reg += 4 * engine->instance; /* GEN8_M2TCR */
>>> +        val = 0;
>>> +    } else {
>>> +        val = engine->instance;
>>> +    }
>>> +
>>> +    val = BIT(val);
>>> +
>>> +    engine->tlb_inv.mcr = regs == xehp_regs;
>>> +    engine->tlb_inv.reg = reg;
>>> +    engine->tlb_inv.done = val;
>>> +
>>> +    if (GRAPHICS_VER(i915) >= 12 &&
>>> +        (engine->class == VIDEO_DECODE_CLASS ||
>>> +         engine->class == VIDEO_ENHANCEMENT_CLASS ||
>>> +         engine->class == COMPUTE_CLASS))
>>> +        engine->tlb_inv.request = _MASKED_BIT_ENABLE(val);
>>> +    else
>>> +        engine->tlb_inv.request = val;
>>> +
>>> +    return true;
>>> +}
>>> +
>>>   static int engine_setup_common(struct intel_engine_cs *engine)
>>>   {
>>>       int err;
>>> @@ -1182,6 +1260,9 @@ static int engine_setup_common(struct 
>>> intel_engine_cs *engine)
>>>       intel_engine_init_whitelist(engine);
>>>       intel_engine_init_ctx_wa(engine);
>>> +    if (intel_engine_init_tlb_invalidation(engine))
>>> +        engine->flags |= I915_ENGINE_HAS_TLB_INVALIDATION;
>>> +
>>>       if (GRAPHICS_VER(engine->i915) >= 12)
>>>           engine->flags |= I915_ENGINE_HAS_RELATIVE_MMIO;
>>> diff --git a/drivers/gpu/drm/i915/gt/intel_engine_types.h 
>>> b/drivers/gpu/drm/i915/gt/intel_engine_types.h
>>> index 4fd54fb8810f..8df4a09a6022 100644
>>> --- a/drivers/gpu/drm/i915/gt/intel_engine_types.h
>>> +++ b/drivers/gpu/drm/i915/gt/intel_engine_types.h
>>> @@ -341,6 +341,19 @@ struct intel_engine_guc_stats {
>>>       u64 start_gt_clk;
>>>   };
>>> +union intel_engine_tlb_inv_reg {
>>> +    i915_reg_t    reg;
>>> +    i915_mcr_reg_t    mcr_reg;
>>> +};
>>> +
>>> +struct intel_engine_tlb_inv
>>> +{
>>> +    bool mcr;
>>> +    union intel_engine_tlb_inv_reg reg;
>>> +    u32 request;
>>> +    u32 done;
>>> +};
>>> +
>>>   struct intel_engine_cs {
>>>       struct drm_i915_private *i915;
>>>       struct intel_gt *gt;
>>> @@ -372,6 +385,8 @@ struct intel_engine_cs {
>>>       u32 context_size;
>>>       u32 mmio_base;
>>> +    struct intel_engine_tlb_inv tlb_inv;
>>> +
>>>       /*
>>>        * Some w/a require forcewake to be held (which prevents RC6) 
>>> while
>>>        * a particular engine is active. If so, we set fw_domain to which
>>> @@ -556,6 +571,7 @@ struct intel_engine_cs {
>>>   #define I915_ENGINE_HAS_EU_PRIORITY    BIT(10)
>>>   #define I915_ENGINE_FIRST_RENDER_COMPUTE BIT(11)
>>>   #define I915_ENGINE_USES_WA_HOLD_CCS_SWITCHOUT BIT(12)
>>> +#define I915_ENGINE_HAS_TLB_INVALIDATION BIT(13)
>>>       unsigned int flags;
>>>       /*
>>> diff --git a/drivers/gpu/drm/i915/gt/intel_gt.c 
>>> b/drivers/gpu/drm/i915/gt/intel_gt.c
>>> index 7eeee5a7cb33..df7afff16fd6 100644
>>> --- a/drivers/gpu/drm/i915/gt/intel_gt.c
>>> +++ b/drivers/gpu/drm/i915/gt/intel_gt.c
>>> @@ -983,36 +983,6 @@ void intel_gt_info_print(const struct 
>>> intel_gt_info *info,
>>>       intel_sseu_dump(&info->sseu, p);
>>>   }
>>> -struct reg_and_bit {
>>> -    union {
>>> -        i915_reg_t reg;
>>> -        i915_mcr_reg_t mcr_reg;
>>> -    };
>>> -    u32 bit;
>>> -};
>>> -
>>> -static struct reg_and_bit
>>> -get_reg_and_bit(const struct intel_engine_cs *engine, const bool gen8,
>>> -        const i915_reg_t *regs, const unsigned int num)
>>> -{
>>> -    const unsigned int class = engine->class;
>>> -    struct reg_and_bit rb = { };
>>> -
>>> -    if (drm_WARN_ON_ONCE(&engine->i915->drm,
>>> -                 class >= num || !regs[class].reg))
>>> -        return rb;
>>> -
>>> -    rb.reg = regs[class];
>>> -    if (gen8 && class == VIDEO_DECODE_CLASS)
>>> -        rb.reg.reg += 4 * engine->instance; /* GEN8_M2TCR */
>>> -    else
>>> -        rb.bit = engine->instance;
>>> -
>>> -    rb.bit = BIT(rb.bit);
>>> -
>>> -    return rb;
>>> -}
>>> -
>>>   /*
>>>    * HW architecture suggest typical invalidation time at 40us,
>>>    * with pessimistic cases up to 100us and a recommendation to
>>> @@ -1026,14 +996,20 @@ get_reg_and_bit(const struct intel_engine_cs 
>>> *engine, const bool gen8,
>>>    * but are now considered MCR registers.  Since they exist within a 
>>> GAM range,
>>>    * the primary instance of the register rolls up the status from 
>>> each unit.
>>>    */
>>> -static int wait_for_invalidate(struct intel_gt *gt, struct 
>>> reg_and_bit rb)
>>> +static int wait_for_invalidate(struct intel_engine_cs *engine)
>>>   {
>>> -    if (GRAPHICS_VER_FULL(gt->i915) >= IP_VER(12, 50))
>>> -        return intel_gt_mcr_wait_for_reg(gt, rb.mcr_reg, rb.bit, 0,
>>> +    if (engine->tlb_inv.mcr)
>>> +        return intel_gt_mcr_wait_for_reg(engine->gt,
>>> +                         engine->tlb_inv.reg.mcr_reg,
>>> +                         engine->tlb_inv.done,
>>> +                         0,
>>>                            TLB_INVAL_TIMEOUT_US,
>>>                            TLB_INVAL_TIMEOUT_MS);
>>>       else
>>> -        return __intel_wait_for_register_fw(gt->uncore, rb.reg, 
>>> rb.bit, 0,
>>> +        return __intel_wait_for_register_fw(engine->gt->uncore,
>>> +                            engine->tlb_inv.reg.reg,
>>> +                            engine->tlb_inv.done,
>>> +                            0,
>>>                               TLB_INVAL_TIMEOUT_US,
>>>                               TLB_INVAL_TIMEOUT_MS,
>>>                               NULL);
>>> @@ -1041,50 +1017,14 @@ static int wait_for_invalidate(struct 
>>> intel_gt *gt, struct reg_and_bit rb)
>>>   static void mmio_invalidate_full(struct intel_gt *gt)
>>>   {
>>> -    static const i915_reg_t gen8_regs[] = {
>>> -        [RENDER_CLASS]            = GEN8_RTCR,
>>> -        [VIDEO_DECODE_CLASS]        = GEN8_M1TCR, /* , GEN8_M2TCR */
>>> -        [VIDEO_ENHANCEMENT_CLASS]    = GEN8_VTCR,
>>> -        [COPY_ENGINE_CLASS]        = GEN8_BTCR,
>>> -    };
>>> -    static const i915_reg_t gen12_regs[] = {
>>> -        [RENDER_CLASS]            = GEN12_GFX_TLB_INV_CR,
>>> -        [VIDEO_DECODE_CLASS]        = GEN12_VD_TLB_INV_CR,
>>> -        [VIDEO_ENHANCEMENT_CLASS]    = GEN12_VE_TLB_INV_CR,
>>> -        [COPY_ENGINE_CLASS]        = GEN12_BLT_TLB_INV_CR,
>>> -        [COMPUTE_CLASS]            = GEN12_COMPCTX_TLB_INV_CR,
>>> -    };
>>> -    static const i915_mcr_reg_t xehp_regs[] = {
>>> -        [RENDER_CLASS]            = XEHP_GFX_TLB_INV_CR,
>>> -        [VIDEO_DECODE_CLASS]        = XEHP_VD_TLB_INV_CR,
>>> -        [VIDEO_ENHANCEMENT_CLASS]    = XEHP_VE_TLB_INV_CR,
>>> -        [COPY_ENGINE_CLASS]        = XEHP_BLT_TLB_INV_CR,
>>> -        [COMPUTE_CLASS]            = XEHP_COMPCTX_TLB_INV_CR,
>>> -    };
>>>       struct drm_i915_private *i915 = gt->i915;
>>>       struct intel_uncore *uncore = gt->uncore;
>>>       struct intel_engine_cs *engine;
>>>       intel_engine_mask_t awake, tmp;
>>>       enum intel_engine_id id;
>>> -    const i915_reg_t *regs;
>>> -    unsigned int num = 0;
>>>       unsigned long flags;
>>> -    if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) {
>>> -        regs = NULL;
>>> -        num = ARRAY_SIZE(xehp_regs);
>>> -    } else if (GRAPHICS_VER(i915) == 12) {
>>> -        regs = gen12_regs;
>>> -        num = ARRAY_SIZE(gen12_regs);
>>> -    } else if (GRAPHICS_VER(i915) >= 8 && GRAPHICS_VER(i915) <= 11) {
>>> -        regs = gen8_regs;
>>> -        num = ARRAY_SIZE(gen8_regs);
>>> -    } else if (GRAPHICS_VER(i915) < 8) {
>>> -        return;
>>> -    }
>>> -
>>> -    if (drm_WARN_ONCE(&i915->drm, !num,
>>> -              "Platform does not implement TLB invalidation!"))
>>> +    if (GRAPHICS_VER(i915) < 8)
>>>           return;
>>>       intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
>>> @@ -1094,33 +1034,22 @@ static void mmio_invalidate_full(struct 
>>> intel_gt *gt)
>>>       awake = 0;
>>>       for_each_engine(engine, gt, id) {
>>> -        struct reg_and_bit rb;
>>> -
>>>           if (!intel_engine_pm_is_awake(engine))
>>>               continue;
>>> -        if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) {
>>> -            u32 val = BIT(engine->instance);
>>> +        if (drm_WARN_ON_ONCE(&i915->drm,
>>> +                     !(engine->flags & 
>>> I915_ENGINE_HAS_TLB_INVALIDATION)))
>>> +            continue;
>>
>> Hmm, can this flag change dynamically? If not why not put it in 
>> initialization phase.
> 
> Could do. In that case I couldn't have any asserts that engine->tlb_inv 
> is valid but perhaps that is okay. Could just fail engine probe from 
> there if register table is incomplete. Sounds better indeed, I'll change 
> it.
> 
>>
>>> -            if (engine->class == VIDEO_DECODE_CLASS ||
>>> -                engine->class == VIDEO_ENHANCEMENT_CLASS ||
>>> -                engine->class == COMPUTE_CLASS)
>>> -                val = _MASKED_BIT_ENABLE(val);
>>> +        if (engine->tlb_inv.mcr)
>>>               intel_gt_mcr_multicast_write_fw(gt,
>>> -                            xehp_regs[engine->class],
>>> -                            val);
>>> -        } else {
>>> -            rb = get_reg_and_bit(engine, regs == gen8_regs, regs, num);
>>> -            if (!i915_mmio_reg_offset(rb.reg))
>>> -                continue;
>>> -
>>> -            if (GRAPHICS_VER(i915) == 12 && (engine->class == 
>>> VIDEO_DECODE_CLASS ||
>>> -                engine->class == VIDEO_ENHANCEMENT_CLASS ||
>>> -                engine->class == COMPUTE_CLASS))
>>> -                rb.bit = _MASKED_BIT_ENABLE(rb.bit);
>>> -
>>> -            intel_uncore_write_fw(uncore, rb.reg, rb.bit);
>>> -        }
>>> +                            engine->tlb_inv.reg.mcr_reg,
>>> +                            engine->tlb_inv.request);
>>> +        else
>>> +            intel_uncore_write_fw(uncore,
>>> +                          engine->tlb_inv.reg.reg,
>>> +                          engine->tlb_inv.request);
>>> +
>>>           awake |= engine->mask;
>>>       }
>>> @@ -1139,16 +1068,7 @@ static void mmio_invalidate_full(struct 
>>> intel_gt *gt)
>>>       intel_gt_mcr_unlock(gt, flags);
>>>       for_each_engine_masked(engine, gt, awake, tmp) {
>>> -        struct reg_and_bit rb;
>>> -
>>> -        if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) {
>>> -            rb.mcr_reg = xehp_regs[engine->class];
>>> -            rb.bit = BIT(engine->instance);
>>> -        } else {
>>> -            rb = get_reg_and_bit(engine, regs == gen8_regs, regs, num);
>>> -        }
>>> -
>>> -        if (wait_for_invalidate(gt, rb))
>>> +        if (wait_for_invalidate(engine))
>>>               drm_err_ratelimited(&gt->i915->drm,
>>>                           "%s TLB invalidation did not complete in 
>>> %ums!\n",
>>>                           engine->name, TLB_INVAL_TIMEOUT_MS);
>>
>> With minor comments addressed:
>>
>> Reviewed-by: Andrzej Hajda <andrzej.hajda at intel.com>
> 
> I'll send v2 at some point, please stand by.
> 
> Regards,
> 
> Tvrtko



More information about the Intel-gfx mailing list