Re: [PATCH v2 10/10] drm/msm: Add a way for userspace to allocate GPU iova

From: Dmitry Osipenko <dmitry.osipenko@collabora.com>
To: Rob Clark <robdclark@gmail.com>, dri-devel@lists.freedesktop.org
Cc: freedreno@lists.freedesktop.org, linux-arm-msm@vger.kernel.org,
	Dmitry Baryshkov <dmitry.baryshkov@linaro.org>,
	Rob Clark <robdclark@chromium.org>, Sean Paul <sean@poorly.run>,
	Abhinav Kumar <quic_abhinavk@quicinc.com>,
	David Airlie <airlied@linux.ie>, Daniel Vetter <daniel@ffwll.ch>,
	Akhil P Oommen <quic_akhilpo@quicinc.com>,
	Jonathan Marek <jonathan@marek.ca>,
	Jordan Crouse <jordan@cosmicpenguin.net>,
	Emma Anholt <emma@anholt.net>,
	Dan Carpenter <dan.carpenter@oracle.com>,
	open list <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH v2 10/10] drm/msm: Add a way for userspace to allocate GPU iova
Date: Thu, 31 Mar 2022 21:53:04 +0300	[thread overview]
Message-ID: <b7a0347f-7106-f2af-bc63-40d8bdc2bb02@collabora.com> (raw)
In-Reply-To: <ad97096f-cc90-4f20-0f73-f33e9b275f1a@collabora.com>


On 3/31/22 21:52, Dmitry Osipenko wrote:
> ...
>> +/*
>> + * Get the requested iova but don't pin it.  Fails if the requested iova is
>> + * not available.  Doesn't need a put because iovas are currently valid for
>> + * the life of the object.
>> + *
>> + * Setting an iova of zero will clear the vma.
>> + */
>> +int msm_gem_set_iova(struct drm_gem_object *obj,
>> +		     struct msm_gem_address_space *aspace, uint64_t iova)
>> +{
>> +	int ret = 0;
> 
> nit: No need to initialize the ret
> 
>> +	msm_gem_lock(obj);
>> +	if (!iova) {
>> +		ret = clear_iova(obj, aspace);
>> +	} else {
>> +		struct msm_gem_vma *vma;
>> +		vma = get_vma_locked(obj, aspace, iova, iova + obj->size);
>> +		if (IS_ERR(vma)) {
>> +			ret = PTR_ERR(vma);
>> +		} else if (GEM_WARN_ON(vma->iova != iova)) {
>> +			clear_iova(obj, aspace);
>> +			ret = -ENOSPC;
> 
> The (vma->iova != iova) means that vma is already set, but to a
> different address. Is -ENOSPC really appropriate here? -EBUSY or -EINVAL
> looks more natural to me.
> 
>> +		}
>> +	}
>> +	msm_gem_unlock(obj);
>> +
>> +	return ret;
>> +}
>> +
>>  /*
>>   * Unpin a iova by updating the reference counts. The memory isn't actually
>>   * purged until something else (shrinker, mm_notifier, destroy, etc) decides
>> diff --git a/drivers/gpu/drm/msm/msm_gem.h b/drivers/gpu/drm/msm/msm_gem.h
>> index 38d66e1248b1..efa2e5c19f1e 100644
>> --- a/drivers/gpu/drm/msm/msm_gem.h
>> +++ b/drivers/gpu/drm/msm/msm_gem.h
>> @@ -38,6 +38,12 @@ struct msm_gem_address_space {
>>  
>>  	/* @faults: the number of GPU hangs associated with this address space */
>>  	int faults;
>> +
>> +	/** @va_start: lowest possible address to allocate */
>> +	uint64_t va_start;
>> +
>> +	/** @va_size: the size of the address space (in bytes) */
>> +	uint64_t va_size;
>>  };
>>  
>>  struct msm_gem_address_space *
>> @@ -144,6 +150,8 @@ struct msm_gem_vma *msm_gem_get_vma_locked(struct drm_gem_object *obj,
>>  					   struct msm_gem_address_space *aspace);
>>  int msm_gem_get_iova(struct drm_gem_object *obj,
>>  		struct msm_gem_address_space *aspace, uint64_t *iova);
>> +int msm_gem_set_iova(struct drm_gem_object *obj,
>> +		struct msm_gem_address_space *aspace, uint64_t iova);
>>  int msm_gem_get_and_pin_iova_range(struct drm_gem_object *obj,
>>  		struct msm_gem_address_space *aspace, uint64_t *iova,
>>  		u64 range_start, u64 range_end);
> nit: There is an odd mix of uint64_t and u64 (and alike) in the MSM code
> :) The uint64_t variant shouldn't be used by kernel code in general and
> checkpatch should want about it.

s/want/warn/