[Intel-gfx] [PATCH v4] drm/i915: Add IOCTL Param to control data port coherency.

[Lis, Tomasz]

On 2018-07-09 18:28, Tvrtko Ursulin wrote:
>
> On 09/07/2018 14:20, Tomasz Lis wrote:
>> The patch adds a parameter to control the data port coherency 
>> functionality
>> on a per-context level. When the IOCTL is called, a command to switch 
>> data
>> port coherency state is added to the ordered list. All prior requests 
>> are
>> executed on old coherency settings, and all exec requests after the 
>> IOCTL
>> will use new settings.
>>
>> Rationale:
>>
>> The OpenCL driver develpers requested a functionality to control cache
>> coherency at data port level. Keeping the coherency at that level is 
>> disabled
>> by default due to its performance costs. OpenCL driver is planning to
>> enable it for a small subset of submissions, when such functionality is
>> required. Below are answers to basic question explaining background
>> of the functionality and reasoning for the proposed implementation:
>>
>> 1. Why do we need a coherency enable/disable switch for memory that 
>> is shared
>> between CPU and GEN (GPU)?
>>
>> Memory coherency between CPU and GEN, while being a great feature 
>> that enables
>> CL_MEM_SVM_FINE_GRAIN_BUFFER OCL capability on Intel GEN 
>> architecture, adds
>> overhead related to tracking (snooping) memory inside different cache 
>> units
>> (L1$, L2$, L3$, LLC$, etc.). At the same time, minority of modern OCL
>> applications actually use CL_MEM_SVM_FINE_GRAIN_BUFFER (and hence 
>> require
>> memory coherency between CPU and GPU). The goal of coherency 
>> enable/disable
>> switch is to remove overhead of memory coherency when memory 
>> coherency is not
>> needed.
>>
>> 2. Why do we need a global coherency switch?
>>
>> In order to support I/O commands from within EUs (Execution Units), 
>> Intel GEN
>> ISA (GEN Instruction Set Assembly) contains dedicated "send" 
>> instructions.
>> These send instructions provide several addressing models. One of these
>> addressing models (named "stateless") provides most flexible I/O 
>> using plain
>> virtual addresses (as opposed to buffer_handle+offset models). This 
>> "stateless"
>> model is similar to regular memory load/store operations available on 
>> typical
>> CPUs. Since this model provides I/O using arbitrary virtual 
>> addresses, it
>> enables algorithmic designs that are based on pointer-to-pointer 
>> (e.g. buffer
>> of pointers) concepts. For instance, it allows creating tree-like data
>> structures such as:
>>                     ________________
>>                    |      NODE1     |
>>                    | uint64_t data  |
>>                    +----------------|
>>                    | NODE*  |  NODE*|
>>                    +--------+-------+
>>                      /              \
>>     ________________/                \________________
>>    |      NODE2     |                |      NODE3     |
>>    | uint64_t data  |                | uint64_t data  |
>>    +----------------|                +----------------|
>>    | NODE*  |  NODE*|                | NODE*  |  NODE*|
>>    +--------+-------+                +--------+-------+
>>
>> Please note that pointers inside such structures can point to memory 
>> locations
>> in different OCL allocations  - e.g. NODE1 and NODE2 can reside in 
>> one OCL
>> allocation while NODE3 resides in a completely separate OCL allocation.
>> Additionally, such pointers can be shared with CPU (i.e. using SVM - 
>> Shared
>> Virtual Memory feature). Using pointers from different allocations 
>> doesn't
>> affect the stateless addressing model which even allows scattered 
>> reading from
>> different allocations at the same time (i.e. by utilizing SIMD-nature 
>> of send
>> instructions).
>>
>> When it comes to coherency programming, send instructions in 
>> stateless model
>> can be encoded (at ISA level) to either use or disable coherency. 
>> However, for
>> generic OCL applications (such as example with tree-like data 
>> structure), OCL
>> compiler is not able to determine origin of memory pointed to by an 
>> arbitrary
>> pointer - i.e. is not able to track given pointer back to a specific
>> allocation. As such, it's not able to decide whether coherency is 
>> needed or not
>> for specific pointer (or for specific I/O instruction). As a result, 
>> compiler
>> encodes all stateless sends as coherent (doing otherwise would lead to
>> functional issues resulting from data corruption). Please note that 
>> it would be
>> possible to workaround this (e.g. based on allocations map and 
>> pointer bounds
>> checking prior to each I/O instruction) but the performance cost of such
>> workaround would be many times greater than the cost of keeping 
>> coherency
>> always enabled. As such, enabling/disabling memory coherency at GEN 
>> ISA level
>> is not feasible and alternative method is needed.
>>
>> Such alternative solution is to have a global coherency switch that 
>> allows
>> disabling coherency for single (though entire) GPU submission. This is
>> beneficial because this way we:
>> * can enable (and pay for) coherency only in submissions that 
>> actually need
>> coherency (submissions that use CL_MEM_SVM_FINE_GRAIN_BUFFER resources)
>> * don't care about coherency at GEN ISA granularity (no performance 
>> impact)
>>
>> 3. Will coherency switch be used frequently?
>>
>> There are scenarios that will require frequent toggling of the coherency
>> switch.
>> E.g. an application has two OCL compute kernels: kern_master and 
>> kern_worker.
>> kern_master uses, concurrently with CPU, some fine grain SVM resources
>> (CL_MEM_SVM_FINE_GRAIN_BUFFER). These resources contain descriptors of
>> computational work that needs to be executed. kern_master analyzes 
>> incoming
>> work descriptors and populates a plain OCL buffer (non-fine-grain) 
>> with payload
>> for kern_worker. Once kern_master is done, kern_worker kicks-in and 
>> processes
>> the payload that kern_master produced. These two kernels work in a 
>> loop, one
>> after another. Since only kern_master requires coherency, kern_worker 
>> should
>> not be forced to pay for it. This means that we need to have the 
>> ability to
>> toggle coherency switch on or off per each GPU submission:
>> (ENABLE COHERENCY) kern_master -> (DISABLE COHERENCY)kern_worker -> 
>> (ENABLE
>> COHERENCY) kern_master -> (DISABLE COHERENCY)kern_worker -> ...
>>
>> v2: Fixed compilation warning.
>> v3: Refactored the patch to add IOCTL instead of exec flag.
>> v4: Renamed and documented the API flag. Used strict values.
>>      Removed redundant GEM_WARN_ON()s. Improved to coding standard.
>>      Introduced a macro for checking whether hardware supports the 
>> feature.
>>
>> Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
>> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>> Cc: Chris Wilson <chris at chris-wilson.co.uk>
>> Cc: Michal Winiarski <michal.winiarski at intel.com>
>>
>> Bspec: 11419
>> Signed-off-by: Tomasz Lis <tomasz.lis at intel.com>
>> ---
>>   drivers/gpu/drm/i915/i915_drv.h         |  1 +
>>   drivers/gpu/drm/i915/i915_gem_context.c | 41 
>> +++++++++++++++++++++++++++
>>   drivers/gpu/drm/i915/i915_gem_context.h |  6 ++++
>>   drivers/gpu/drm/i915/intel_lrc.c        | 49 
>> +++++++++++++++++++++++++++++++++
>>   drivers/gpu/drm/i915/intel_lrc.h        |  4 +++
>>   include/uapi/drm/i915_drm.h             |  6 ++++
>>   6 files changed, 107 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/i915/i915_drv.h 
>> b/drivers/gpu/drm/i915/i915_drv.h
>> index 09ab124..7d4bbd5 100644
>> --- a/drivers/gpu/drm/i915/i915_drv.h
>> +++ b/drivers/gpu/drm/i915/i915_drv.h
>> @@ -2524,6 +2524,7 @@ intel_info(const struct drm_i915_private 
>> *dev_priv)
>>   #define HAS_EDRAM(dev_priv)    (!!((dev_priv)->edram_cap & 
>> EDRAM_ENABLED))
>>   #define HAS_WT(dev_priv)    ((IS_HASWELL(dev_priv) || \
>>                    IS_BROADWELL(dev_priv)) && HAS_EDRAM(dev_priv))
>> +#define HAS_DATA_PORT_COHERENCY(dev_priv) (INTEL_GEN(dev_priv) >= 9)
>>     #define HWS_NEEDS_PHYSICAL(dev_priv) 
>> ((dev_priv)->info.hws_needs_physical)
>>   diff --git a/drivers/gpu/drm/i915/i915_gem_context.c 
>> b/drivers/gpu/drm/i915/i915_gem_context.c
>> index b10770c..6db352e 100644
>> --- a/drivers/gpu/drm/i915/i915_gem_context.c
>> +++ b/drivers/gpu/drm/i915/i915_gem_context.c
>> @@ -711,6 +711,26 @@ static bool client_is_banned(struct 
>> drm_i915_file_private *file_priv)
>>       return atomic_read(&file_priv->ban_score) >= 
>> I915_CLIENT_SCORE_BANNED;
>>   }
>>   +static int i915_gem_context_set_data_port_coherent(struct 
>> i915_gem_context *ctx)
>> +{
>> +    int ret;
>> +
>> +    ret = intel_lr_context_modify_data_port_coherency(ctx, true);
>> +    if (!ret)
>> +        __set_bit(CONTEXT_DATA_PORT_COHERENT, &ctx->flags);
>> +    return ret;
>> +}
>> +
>> +static int i915_gem_context_clear_data_port_coherent(struct 
>> i915_gem_context *ctx)
>> +{
>> +    int ret;
>> +
>> +    ret = intel_lr_context_modify_data_port_coherency(ctx, false);
>> +    if (!ret)
>> +        __clear_bit(CONTEXT_DATA_PORT_COHERENT, &ctx->flags);
>> +    return ret;
>
> Is there a good reason you allow userspace to keep emitting unlimited 
> number of commands which actually do not change the status? If not 
> please consider gating the command emission with 
> test_and_set_bit/test_and_clear_bit. Hm.. apart even with that they 
> could keep toggling ad infinitum with no real work in between. Has it 
> been considered to only save the desired state in set param and then 
> emit it, if needed, before next execbuf? Minor thing in any case, just 
> curious since I wasn't following the threads.
>
(discussed further in separate thread)
>> +}
>> +
>>   int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
>>                     struct drm_file *file)
>>   {
>> @@ -784,6 +804,7 @@ int i915_gem_context_destroy_ioctl(struct 
>> drm_device *dev, void *data,
>>   int i915_gem_context_getparam_ioctl(struct drm_device *dev, void 
>> *data,
>>                       struct drm_file *file)
>>   {
>> +    struct drm_i915_private *dev_priv = to_i915(dev);
>
> Feel free to use the local for the other existing to_i915(dev) call 
> sites in here.
>
> Also use i915 for the local name. Unless I915_READ/WRITE is used i915 
> is preferred nowadays.
Will do.
>
>>       struct drm_i915_file_private *file_priv = file->driver_priv;
>>       struct drm_i915_gem_context_param *args = data;
>>       struct i915_gem_context *ctx;
>> @@ -818,6 +839,12 @@ int i915_gem_context_getparam_ioctl(struct 
>> drm_device *dev, void *data,
>>       case I915_CONTEXT_PARAM_PRIORITY:
>>           args->value = ctx->sched.priority;
>>           break;
>> +    case I915_CONTEXT_PARAM_DATA_PORT_COHERENCY:
>> +        if (!HAS_DATA_PORT_COHERENCY(dev_priv))
>> +            ret = -ENODEV;
>> +        else
>> +            args->value = i915_gem_context_is_data_port_coherent(ctx);
>> +        break;
>>       default:
>>           ret = -EINVAL;
>>           break;
>> @@ -830,6 +857,7 @@ int i915_gem_context_getparam_ioctl(struct 
>> drm_device *dev, void *data,
>>   int i915_gem_context_setparam_ioctl(struct drm_device *dev, void 
>> *data,
>>                       struct drm_file *file)
>>   {
>> +    struct drm_i915_private *dev_priv = to_i915(dev);
>
> As with get_param.
Ack.
>
>>       struct drm_i915_file_private *file_priv = file->driver_priv;
>>       struct drm_i915_gem_context_param *args = data;
>>       struct i915_gem_context *ctx;
>> @@ -893,6 +921,19 @@ int i915_gem_context_setparam_ioctl(struct 
>> drm_device *dev, void *data,
>>           }
>>           break;
>>   +    case I915_CONTEXT_PARAM_DATA_PORT_COHERENCY:
>> +        if (args->size)
>> +            ret = -EINVAL;
>> +        else if (!HAS_DATA_PORT_COHERENCY(dev_priv))
>> +            ret = -ENODEV;
>> +        else if (args->value == 1)
>> +            ret = i915_gem_context_set_data_port_coherent(ctx);
>> +        else if (args->value == 0)
>> +            ret = i915_gem_context_clear_data_port_coherent(ctx);
>> +        else
>> +            ret = -EINVAL;
>> +        break;
>> +
>>       default:
>>           ret = -EINVAL;
>>           break;
>> diff --git a/drivers/gpu/drm/i915/i915_gem_context.h 
>> b/drivers/gpu/drm/i915/i915_gem_context.h
>> index b116e49..e8ccb70 100644
>> --- a/drivers/gpu/drm/i915/i915_gem_context.h
>> +++ b/drivers/gpu/drm/i915/i915_gem_context.h
>> @@ -126,6 +126,7 @@ struct i915_gem_context {
>>   #define CONTEXT_BANNABLE        3
>>   #define CONTEXT_BANNED            4
>>   #define CONTEXT_FORCE_SINGLE_SUBMISSION    5
>> +#define CONTEXT_DATA_PORT_COHERENT    6
>>         /**
>>        * @hw_id: - unique identifier for the context
>> @@ -257,6 +258,11 @@ static inline void 
>> i915_gem_context_set_force_single_submission(struct i915_gem_
>>       __set_bit(CONTEXT_FORCE_SINGLE_SUBMISSION, &ctx->flags);
>>   }
>>   +static inline bool i915_gem_context_is_data_port_coherent(struct 
>> i915_gem_context *ctx)
>> +{
>> +    return test_bit(CONTEXT_DATA_PORT_COHERENT, &ctx->flags);
>> +}
>> +
>>   static inline bool i915_gem_context_is_default(const struct 
>> i915_gem_context *c)
>>   {
>>       return c->user_handle == DEFAULT_CONTEXT_HANDLE;
>> diff --git a/drivers/gpu/drm/i915/intel_lrc.c 
>> b/drivers/gpu/drm/i915/intel_lrc.c
>> index ab89dab..1f037e3 100644
>> --- a/drivers/gpu/drm/i915/intel_lrc.c
>> +++ b/drivers/gpu/drm/i915/intel_lrc.c
>> @@ -259,6 +259,55 @@ intel_lr_context_descriptor_update(struct 
>> i915_gem_context *ctx,
>>       ce->lrc_desc = desc;
>>   }
>>   +static int emit_set_data_port_coherency(struct i915_request *req, 
>> bool enable)
>
> After much disagreement we ended up with rq as the consistent naming 
> for requests.
:)
ok.
>
>> +{
>> +    u32 *cs;
>> +    i915_reg_t reg;
>> +
>> +    GEM_BUG_ON(req->engine->class != RENDER_CLASS);
>> +    GEM_BUG_ON(INTEL_GEN(req->i915) < 9);
>> +
>> +    cs = intel_ring_begin(req, 4);
>> +    if (IS_ERR(cs))
>> +        return PTR_ERR(cs);
>> +
>> +    if (INTEL_GEN(req->i915) >= 10)
>> +        reg = CNL_HDC_CHICKEN0;
>> +    else
>> +        reg = HDC_CHICKEN0;
>> +
>> +    *cs++ = MI_LOAD_REGISTER_IMM(1);
>> +    *cs++ = i915_mmio_reg_offset(reg);
>> +    /* Enabling coherency means disabling the bit which forces it 
>> off */
>> +    if (enable)
>> +        *cs++ = _MASKED_BIT_DISABLE(HDC_FORCE_NON_COHERENT);
>> +    else
>> +        *cs++ = _MASKED_BIT_ENABLE(HDC_FORCE_NON_COHERENT);
>> +    *cs++ = MI_NOOP;
>> +
>> +    intel_ring_advance(req, cs);
>> +
>> +    return 0;
>> +}
>> +
>> +int
>> +intel_lr_context_modify_data_port_coherency(struct i915_gem_context 
>> *ctx,
>> +                        bool enable)
>> +{
>> +    struct i915_request *req;
>
> rq as above.
ack
>
>> +    int ret;
>> +
>> +    req = i915_request_alloc(ctx->i915->engine[RCS], ctx);
>> +    if (IS_ERR(req))
>> +        return PTR_ERR(req);
>> +
>> +    ret = emit_set_data_port_coherency(req, enable);
>> +
>> +    i915_request_add(req);
>> +
>> +    return ret;
>> +}
>> +
>>   static struct i915_priolist *
>>   lookup_priolist(struct intel_engine_cs *engine, int prio)
>>   {
>> diff --git a/drivers/gpu/drm/i915/intel_lrc.h 
>> b/drivers/gpu/drm/i915/intel_lrc.h
>> index 1593194..f6965ae 100644
>> --- a/drivers/gpu/drm/i915/intel_lrc.h
>> +++ b/drivers/gpu/drm/i915/intel_lrc.h
>> @@ -104,4 +104,8 @@ struct i915_gem_context;
>>     void intel_lr_context_resume(struct drm_i915_private *dev_priv);
>>   +int
>> +intel_lr_context_modify_data_port_coherency(struct i915_gem_context 
>> *ctx,
>> +                        bool enable);
>> +
>>   #endif /* _INTEL_LRC_H_ */
>> diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
>> index 7f5634c..e677bea 100644
>> --- a/include/uapi/drm/i915_drm.h
>> +++ b/include/uapi/drm/i915_drm.h
>> @@ -1456,6 +1456,12 @@ struct drm_i915_gem_context_param {
>>   #define   I915_CONTEXT_MAX_USER_PRIORITY    1023 /* inclusive */
>>   #define   I915_CONTEXT_DEFAULT_PRIORITY        0
>>   #define   I915_CONTEXT_MIN_USER_PRIORITY    -1023 /* inclusive */
>> +/*
>> + * When data port level coherency is enabled, the GPU will update 
>> memory
>> + * buffers shared with CPU, by forcing internal cache units to send 
>> memory
>> + * writes to real RAM faster. Keeping such coherency has performance 
>> cost.
>
> Is this comment correct? Is it actually sending memory writes to 
> _RAM_, or just the coherency mode enabled, even if only targetting CPU 
> or shared cache, which adds a cost?
I'm not sure whether there are further coherency modes to choose how 
"deep" coherency goes. The use case of OCL Team is to see gradual 
changes in the buffers on CPU side while the execution progresses. Write 
to RAM is needed to achieve that. And that limits performance by using 
RAM bandwidth.
>
> s/Keeping such coherency has performance cost./Enabling data port 
> coherency has a performance cost./ ? Or "can have a performance cost"?
I would prefer "Enabling data port coherency has a performance cost.". 
There likely are workloads with unmeasureable performance impact, but in 
real world using more memory writes will always slow down something.
>
>> + */
>> +#define I915_CONTEXT_PARAM_DATA_PORT_COHERENCY    0x7
>>       __u64 value;
>>   };
>>
>
> Since I understand this design has been approved already on the high 
> level, and as you can see I only had some minor comments to add, I can 
> say that the patch in principle looks okay to me.
Great; will produce a v5 soon.
-Tomasz
>
> Regards,
>
> Tvrtko