[1/4] dma-buf: change DMA-buf locking convention

Message ID	20190829142917.13058-2-christian.koenig@amd.com (mailing list archive)
State	New, archived
Headers	show Return-Path: <SRS0=bwQ0=WZ=lists.freedesktop.org=intel-gfx-bounces@kernel.org> DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org 5D9FD21874 From: " =?utf-8?q?Christian_K=C3=B6nig?= " <ckoenig.leichtzumerken@gmail.com> To: daniel@ffwll.ch, dri-devel@lists.freedesktop.org, sumit.semwal@linaro.org, linaro-mm-sig@lists.linaro.org, linux-media@vger.kernel.org, intel-gfx@lists.freedesktop.org Date: Thu, 29 Aug 2019 16:29:14 +0200 Message-Id: <20190829142917.13058-2-christian.koenig@amd.com> In-Reply-To: <20190829142917.13058-1-christian.koenig@amd.com> References: <20190829142917.13058-1-christian.koenig@amd.com> MIME-Version: 1.0 Subject: [Intel-gfx] [PATCH 1/4] dma-buf: change DMA-buf locking convention Precedence: list Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: base64 Errors-To: intel-gfx-bounces@lists.freedesktop.org Sender: "Intel-gfx" <intel-gfx-bounces@lists.freedesktop.org>
Series	[1/4] dma-buf: change DMA-buf locking convention \| expand [1/4] dma-buf: change DMA-buf locking convention [2/4] drm/ttm: use the parent resv for ghost objects v2 [3/4] drm/amdgpu: add independent DMA-buf export v7 [4/4] drm/amdgpu: add independent DMA-buf import v8

Christian König Aug. 29, 2019, 2:29 p.m. UTC

This patch is a stripped down version of the locking changes
necessary to support dynamic DMA-buf handling.

For compatibility we cache the DMA-buf mapping as soon as
exporter/importer disagree on the dynamic handling.

Signed-off-by: Christian König <christian.koenig@amd.com>
---
 drivers/dma-buf/dma-buf.c | 90 ++++++++++++++++++++++++++++++++++++---
 include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
 2 files changed, 133 insertions(+), 8 deletions(-)

Daniel Vetter Sept. 3, 2019, 8:05 a.m. UTC | #1

On Thu, Aug 29, 2019 at 04:29:14PM +0200, Christian König wrote:
> This patch is a stripped down version of the locking changes
> necessary to support dynamic DMA-buf handling.
> 
> For compatibility we cache the DMA-buf mapping as soon as
> exporter/importer disagree on the dynamic handling.
> 
> Signed-off-by: Christian König <christian.koenig@amd.com>
> ---
>  drivers/dma-buf/dma-buf.c | 90 ++++++++++++++++++++++++++++++++++++---
>  include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
>  2 files changed, 133 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> index 433d91d710e4..65052d52602b 100644
> --- a/drivers/dma-buf/dma-buf.c
> +++ b/drivers/dma-buf/dma-buf.c
> @@ -525,6 +525,10 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info)
>  		return ERR_PTR(-EINVAL);
>  	}
>  
> +	if (WARN_ON(exp_info->ops->cache_sgt_mapping &&
> +		    exp_info->ops->dynamic_mapping))
> +		return ERR_PTR(-EINVAL);
> +
>  	if (!try_module_get(exp_info->owner))
>  		return ERR_PTR(-ENOENT);
>  
> @@ -645,10 +649,11 @@ void dma_buf_put(struct dma_buf *dmabuf)
>  EXPORT_SYMBOL_GPL(dma_buf_put);
>  
>  /**
> - * dma_buf_attach - Add the device to dma_buf's attachments list; optionally,
> + * dma_buf_dynamic_attach - Add the device to dma_buf's attachments list; optionally,
>   * calls attach() of dma_buf_ops to allow device-specific attach functionality
> - * @dmabuf:	[in]	buffer to attach device to.
> - * @dev:	[in]	device to be attached.
> + * @dmabuf:		[in]	buffer to attach device to.
> + * @dev:		[in]	device to be attached.
> + * @dynamic_mapping:	[in]	calling convention for map/unmap
>   *
>   * Returns struct dma_buf_attachment pointer for this attachment. Attachments
>   * must be cleaned up by calling dma_buf_detach().
> @@ -662,8 +667,9 @@ EXPORT_SYMBOL_GPL(dma_buf_put);
>   * accessible to @dev, and cannot be moved to a more suitable place. This is
>   * indicated with the error code -EBUSY.
>   */
> -struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> -					  struct device *dev)
> +struct dma_buf_attachment *
> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
> +		       bool dynamic_mapping)
>  {
>  	struct dma_buf_attachment *attach;
>  	int ret;
> @@ -677,6 +683,7 @@ struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>  
>  	attach->dev = dev;
>  	attach->dmabuf = dmabuf;
> +	attach->dynamic_mapping = dynamic_mapping;
>  
>  	mutex_lock(&dmabuf->lock);
>  
> @@ -685,16 +692,64 @@ struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>  		if (ret)
>  			goto err_attach;
>  	}
> +	dma_resv_lock(dmabuf->resv, NULL);
>  	list_add(&attach->node, &dmabuf->attachments);
> +	dma_resv_unlock(dmabuf->resv);
>  
>  	mutex_unlock(&dmabuf->lock);
>  
> +	/* When either the importer or the exporter can't handle dynamic
> +	 * mappings we cache the mapping here to avoid issues with the
> +	 * reservation object lock.
> +	 */
> +	if (dma_buf_attachment_is_dynamic(attach) !=
> +	    dma_buf_is_dynamic(dmabuf)) {
> +		struct sg_table *sgt;
> +
> +		if (dma_buf_is_dynamic(attach->dmabuf))
> +			dma_resv_lock(attach->dmabuf->resv, NULL);
> +
> +		sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);

Now we're back to enforcing DMA_BIDI, which works nicely around the
locking pain, but apparently upsets the arm-soc folks who want to control
this better. That's why your previous version moved the caching into
map/unmap_sg, which resurrected the locking hydra.

I think we're going a bit in circles here, and I don't have a good idea
either :-/
-Daniel

> +		if (!sgt)
> +			sgt = ERR_PTR(-ENOMEM);
> +		if (IS_ERR(sgt)) {
> +			ret = PTR_ERR(sgt);
> +			goto err_unlock;
> +		}
> +		if (dma_buf_is_dynamic(attach->dmabuf))
> +			dma_resv_unlock(attach->dmabuf->resv);
> +		attach->sgt = sgt;
> +		attach->dir = DMA_BIDIRECTIONAL;
> +	}
> +
>  	return attach;
>  
>  err_attach:
>  	kfree(attach);
>  	mutex_unlock(&dmabuf->lock);
>  	return ERR_PTR(ret);
> +
> +err_unlock:
> +	if (dma_buf_is_dynamic(attach->dmabuf))
> +		dma_resv_unlock(attach->dmabuf->resv);
> +
> +	dma_buf_detach(dmabuf, attach);
> +	return ERR_PTR(ret);
> +}
> +EXPORT_SYMBOL_GPL(dma_buf_dynamic_attach);
> +
> +/**
> + * dma_buf_attach - Wrapper for dma_buf_dynamic_attach
> + * @dmabuf:	[in]	buffer to attach device to.
> + * @dev:	[in]	device to be attached.
> + *
> + * Wrapper to call dma_buf_dynamic_attach() for drivers which still use a static
> + * mapping.
> + */
> +struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> +					  struct device *dev)
> +{
> +	return dma_buf_dynamic_attach(dmabuf, dev, false);
>  }
>  EXPORT_SYMBOL_GPL(dma_buf_attach);
>  
> @@ -711,11 +766,20 @@ void dma_buf_detach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach)
>  	if (WARN_ON(!dmabuf || !attach))
>  		return;
>  
> -	if (attach->sgt)
> +	if (attach->sgt) {
> +		if (dma_buf_is_dynamic(attach->dmabuf))
> +			dma_resv_lock(attach->dmabuf->resv, NULL);
> +
>  		dmabuf->ops->unmap_dma_buf(attach, attach->sgt, attach->dir);
>  
> +		if (dma_buf_is_dynamic(attach->dmabuf))
> +			dma_resv_unlock(attach->dmabuf->resv);
> +	}
> +
>  	mutex_lock(&dmabuf->lock);
> +	dma_resv_lock(dmabuf->resv, NULL);
>  	list_del(&attach->node);
> +	dma_resv_unlock(dmabuf->resv);
>  	if (dmabuf->ops->detach)
>  		dmabuf->ops->detach(dmabuf, attach);
>  
> @@ -749,6 +813,9 @@ struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *attach,
>  	if (WARN_ON(!attach || !attach->dmabuf))
>  		return ERR_PTR(-EINVAL);
>  
> +	if (dma_buf_attachment_is_dynamic(attach))
> +		dma_resv_assert_held(attach->dmabuf->resv);
> +
>  	if (attach->sgt) {
>  		/*
>  		 * Two mappings with different directions for the same
> @@ -761,6 +828,9 @@ struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *attach,
>  		return attach->sgt;
>  	}
>  
> +	if (dma_buf_is_dynamic(attach->dmabuf))
> +		dma_resv_assert_held(attach->dmabuf->resv);
> +
>  	sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction);
>  	if (!sg_table)
>  		sg_table = ERR_PTR(-ENOMEM);
> @@ -793,9 +863,15 @@ void dma_buf_unmap_attachment(struct dma_buf_attachment *attach,
>  	if (WARN_ON(!attach || !attach->dmabuf || !sg_table))
>  		return;
>  
> +	if (dma_buf_attachment_is_dynamic(attach))
> +		dma_resv_assert_held(attach->dmabuf->resv);
> +
>  	if (attach->sgt == sg_table)
>  		return;
>  
> +	if (dma_buf_is_dynamic(attach->dmabuf))
> +		dma_resv_assert_held(attach->dmabuf->resv);
> +
>  	attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, direction);
>  }
>  EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment);
> @@ -1219,10 +1295,12 @@ static int dma_buf_debug_show(struct seq_file *s, void *unused)
>  		seq_puts(s, "\tAttached Devices:\n");
>  		attach_count = 0;
>  
> +		dma_resv_lock(buf_obj->resv, NULL);
>  		list_for_each_entry(attach_obj, &buf_obj->attachments, node) {
>  			seq_printf(s, "\t%s\n", dev_name(attach_obj->dev));
>  			attach_count++;
>  		}
> +		dma_resv_unlock(buf_obj->resv);
>  
>  		seq_printf(s, "Total %d devices attached\n\n",
>  				attach_count);
> diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h
> index ec212cb27fdc..a8f8b2b812fd 100644
> --- a/include/linux/dma-buf.h
> +++ b/include/linux/dma-buf.h
> @@ -42,6 +42,17 @@ struct dma_buf_ops {
>  	  */
>  	bool cache_sgt_mapping;
>  
> +	/**
> +	 * @dynamic_mapping:
> +	 *
> +	 * If true the framework makes sure that the map/unmap_dma_buf
> +	 * callbacks are always called with the dma_resv object locked.
> +	 *
> +	 * If false the framework makes ure that the map/unmap_dma_buf
> +	 * callbacks are always called without the dma_resv object locked.
> +	 */
> +	bool dynamic_mapping;
> +
>  	/**
>  	 * @attach:
>  	 *
> @@ -109,6 +120,9 @@ struct dma_buf_ops {
>  	 * any other kind of sharing that the exporter might wish to make
>  	 * available to buffer-users.
>  	 *
> +	 * This is always called with the dmabuf->resv object locked when
> +	 * the dynamic_mapping flag is true.
> +	 *
>  	 * Returns:
>  	 *
>  	 * A &sg_table scatter list of or the backing storage of the DMA buffer,
> @@ -327,6 +341,8 @@ struct dma_buf {
>   * @sgt: cached mapping.
>   * @dir: direction of cached mapping.
>   * @priv: exporter specific attachment data.
> + * @dynamic_mapping: true if dma_buf_map/unmap_attachment() is called with the
> + * dma_resv lock held.
>   *
>   * This structure holds the attachment information between the dma_buf buffer
>   * and its user device(s). The list contains one attachment struct per device
> @@ -343,6 +359,7 @@ struct dma_buf_attachment {
>  	struct list_head node;
>  	struct sg_table *sgt;
>  	enum dma_data_direction dir;
> +	bool dynamic_mapping;
>  	void *priv;
>  };
>  
> @@ -394,10 +411,39 @@ static inline void get_dma_buf(struct dma_buf *dmabuf)
>  	get_file(dmabuf->file);
>  }
>  
> +/**
> + * dma_buf_is_dynamic - check if a DMA-buf uses dynamic mappings.
> + * @dmabuf: the DMA-buf to check
> + *
> + * Returns true if a DMA-buf exporter wants to be called with the dma_resv
> + * locked, false if it doesn't wants to be called with the lock held.
> + */
> +static inline bool dma_buf_is_dynamic(struct dma_buf *dmabuf)
> +{
> +	return dmabuf->ops->dynamic_mapping;
> +}
> +
> +/**
> + * dma_buf_attachment_is_dynamic - check if a DMA-buf attachment uses dynamic
> + * mappinsg
> + * @attach: the DMA-buf attachment to check
> + *
> + * Returns true if a DMA-buf importer wants to call the map/unmap functions with
> + * the dma_resv lock held.
> + */
> +static inline bool
> +dma_buf_attachment_is_dynamic(struct dma_buf_attachment *attach)
> +{
> +	return attach->dynamic_mapping;
> +}
> +
>  struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> -							struct device *dev);
> +					  struct device *dev);
> +struct dma_buf_attachment *
> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
> +		       bool dynamic_mapping);
>  void dma_buf_detach(struct dma_buf *dmabuf,
> -				struct dma_buf_attachment *dmabuf_attach);
> +		    struct dma_buf_attachment *attach);
>  
>  struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info);
>  
> @@ -409,6 +455,7 @@ struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *,
>  					enum dma_data_direction);
>  void dma_buf_unmap_attachment(struct dma_buf_attachment *, struct sg_table *,
>  				enum dma_data_direction);
> +void dma_buf_move_notify(struct dma_buf *dma_buf);
>  int dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
>  			     enum dma_data_direction dir);
>  int dma_buf_end_cpu_access(struct dma_buf *dma_buf,
> -- 
> 2.17.1
>

Christian König Sept. 11, 2019, 10:53 a.m. UTC | #2

Am 03.09.19 um 10:05 schrieb Daniel Vetter:
> On Thu, Aug 29, 2019 at 04:29:14PM +0200, Christian König wrote:
>> This patch is a stripped down version of the locking changes
>> necessary to support dynamic DMA-buf handling.
>>
>> For compatibility we cache the DMA-buf mapping as soon as
>> exporter/importer disagree on the dynamic handling.
>>
>> Signed-off-by: Christian König <christian.koenig@amd.com>
>> ---
>>   drivers/dma-buf/dma-buf.c | 90 ++++++++++++++++++++++++++++++++++++---
>>   include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
>>   2 files changed, 133 insertions(+), 8 deletions(-)
>>
>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
>> index 433d91d710e4..65052d52602b 100644
>> --- a/drivers/dma-buf/dma-buf.c
>> +++ b/drivers/dma-buf/dma-buf.c
>> @@ -525,6 +525,10 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info)
>>   		return ERR_PTR(-EINVAL);
>>   	}
>>   
>> +	if (WARN_ON(exp_info->ops->cache_sgt_mapping &&
>> +		    exp_info->ops->dynamic_mapping))
>> +		return ERR_PTR(-EINVAL);
>> +
>>   	if (!try_module_get(exp_info->owner))
>>   		return ERR_PTR(-ENOENT);
>>   
>> @@ -645,10 +649,11 @@ void dma_buf_put(struct dma_buf *dmabuf)
>>   EXPORT_SYMBOL_GPL(dma_buf_put);
>>   
>>   /**
>> - * dma_buf_attach - Add the device to dma_buf's attachments list; optionally,
>> + * dma_buf_dynamic_attach - Add the device to dma_buf's attachments list; optionally,
>>    * calls attach() of dma_buf_ops to allow device-specific attach functionality
>> - * @dmabuf:	[in]	buffer to attach device to.
>> - * @dev:	[in]	device to be attached.
>> + * @dmabuf:		[in]	buffer to attach device to.
>> + * @dev:		[in]	device to be attached.
>> + * @dynamic_mapping:	[in]	calling convention for map/unmap
>>    *
>>    * Returns struct dma_buf_attachment pointer for this attachment. Attachments
>>    * must be cleaned up by calling dma_buf_detach().
>> @@ -662,8 +667,9 @@ EXPORT_SYMBOL_GPL(dma_buf_put);
>>    * accessible to @dev, and cannot be moved to a more suitable place. This is
>>    * indicated with the error code -EBUSY.
>>    */
>> -struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>> -					  struct device *dev)
>> +struct dma_buf_attachment *
>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
>> +		       bool dynamic_mapping)
>>   {
>>   	struct dma_buf_attachment *attach;
>>   	int ret;
>> @@ -677,6 +683,7 @@ struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>   
>>   	attach->dev = dev;
>>   	attach->dmabuf = dmabuf;
>> +	attach->dynamic_mapping = dynamic_mapping;
>>   
>>   	mutex_lock(&dmabuf->lock);
>>   
>> @@ -685,16 +692,64 @@ struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>   		if (ret)
>>   			goto err_attach;
>>   	}
>> +	dma_resv_lock(dmabuf->resv, NULL);
>>   	list_add(&attach->node, &dmabuf->attachments);
>> +	dma_resv_unlock(dmabuf->resv);
>>   
>>   	mutex_unlock(&dmabuf->lock);
>>   
>> +	/* When either the importer or the exporter can't handle dynamic
>> +	 * mappings we cache the mapping here to avoid issues with the
>> +	 * reservation object lock.
>> +	 */
>> +	if (dma_buf_attachment_is_dynamic(attach) !=
>> +	    dma_buf_is_dynamic(dmabuf)) {
>> +		struct sg_table *sgt;
>> +
>> +		if (dma_buf_is_dynamic(attach->dmabuf))
>> +			dma_resv_lock(attach->dmabuf->resv, NULL);
>> +
>> +		sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);
> Now we're back to enforcing DMA_BIDI, which works nicely around the
> locking pain, but apparently upsets the arm-soc folks who want to control
> this better.

Take another look at dma_buf_map_attachment(), we still try to get the 
caching there for ARM.

What we do here is to bidirectionally map the buffer to avoid the 
locking hydra when importer and exporter disagree on locking.

So the ARM folks can easily avoid that by switching to dynamic locking 
for both.

Regards,
Christian.

> That's why your previous version moved the caching into
> map/unmap_sg, which resurrected the locking hydra.
>
> I think we're going a bit in circles here, and I don't have a good idea
> either :-/
> -Daniel
>
>> +		if (!sgt)
>> +			sgt = ERR_PTR(-ENOMEM);
>> +		if (IS_ERR(sgt)) {
>> +			ret = PTR_ERR(sgt);
>> +			goto err_unlock;
>> +		}
>> +		if (dma_buf_is_dynamic(attach->dmabuf))
>> +			dma_resv_unlock(attach->dmabuf->resv);
>> +		attach->sgt = sgt;
>> +		attach->dir = DMA_BIDIRECTIONAL;
>> +	}
>> +
>>   	return attach;
>>   
>>   err_attach:
>>   	kfree(attach);
>>   	mutex_unlock(&dmabuf->lock);
>>   	return ERR_PTR(ret);
>> +
>> +err_unlock:
>> +	if (dma_buf_is_dynamic(attach->dmabuf))
>> +		dma_resv_unlock(attach->dmabuf->resv);
>> +
>> +	dma_buf_detach(dmabuf, attach);
>> +	return ERR_PTR(ret);
>> +}
>> +EXPORT_SYMBOL_GPL(dma_buf_dynamic_attach);
>> +
>> +/**
>> + * dma_buf_attach - Wrapper for dma_buf_dynamic_attach
>> + * @dmabuf:	[in]	buffer to attach device to.
>> + * @dev:	[in]	device to be attached.
>> + *
>> + * Wrapper to call dma_buf_dynamic_attach() for drivers which still use a static
>> + * mapping.
>> + */
>> +struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>> +					  struct device *dev)
>> +{
>> +	return dma_buf_dynamic_attach(dmabuf, dev, false);
>>   }
>>   EXPORT_SYMBOL_GPL(dma_buf_attach);
>>   
>> @@ -711,11 +766,20 @@ void dma_buf_detach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach)
>>   	if (WARN_ON(!dmabuf || !attach))
>>   		return;
>>   
>> -	if (attach->sgt)
>> +	if (attach->sgt) {
>> +		if (dma_buf_is_dynamic(attach->dmabuf))
>> +			dma_resv_lock(attach->dmabuf->resv, NULL);
>> +
>>   		dmabuf->ops->unmap_dma_buf(attach, attach->sgt, attach->dir);
>>   
>> +		if (dma_buf_is_dynamic(attach->dmabuf))
>> +			dma_resv_unlock(attach->dmabuf->resv);
>> +	}
>> +
>>   	mutex_lock(&dmabuf->lock);
>> +	dma_resv_lock(dmabuf->resv, NULL);
>>   	list_del(&attach->node);
>> +	dma_resv_unlock(dmabuf->resv);
>>   	if (dmabuf->ops->detach)
>>   		dmabuf->ops->detach(dmabuf, attach);
>>   
>> @@ -749,6 +813,9 @@ struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *attach,
>>   	if (WARN_ON(!attach || !attach->dmabuf))
>>   		return ERR_PTR(-EINVAL);
>>   
>> +	if (dma_buf_attachment_is_dynamic(attach))
>> +		dma_resv_assert_held(attach->dmabuf->resv);
>> +
>>   	if (attach->sgt) {
>>   		/*
>>   		 * Two mappings with different directions for the same
>> @@ -761,6 +828,9 @@ struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *attach,
>>   		return attach->sgt;
>>   	}
>>   
>> +	if (dma_buf_is_dynamic(attach->dmabuf))
>> +		dma_resv_assert_held(attach->dmabuf->resv);
>> +
>>   	sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction);
>>   	if (!sg_table)
>>   		sg_table = ERR_PTR(-ENOMEM);
>> @@ -793,9 +863,15 @@ void dma_buf_unmap_attachment(struct dma_buf_attachment *attach,
>>   	if (WARN_ON(!attach || !attach->dmabuf || !sg_table))
>>   		return;
>>   
>> +	if (dma_buf_attachment_is_dynamic(attach))
>> +		dma_resv_assert_held(attach->dmabuf->resv);
>> +
>>   	if (attach->sgt == sg_table)
>>   		return;
>>   
>> +	if (dma_buf_is_dynamic(attach->dmabuf))
>> +		dma_resv_assert_held(attach->dmabuf->resv);
>> +
>>   	attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, direction);
>>   }
>>   EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment);
>> @@ -1219,10 +1295,12 @@ static int dma_buf_debug_show(struct seq_file *s, void *unused)
>>   		seq_puts(s, "\tAttached Devices:\n");
>>   		attach_count = 0;
>>   
>> +		dma_resv_lock(buf_obj->resv, NULL);
>>   		list_for_each_entry(attach_obj, &buf_obj->attachments, node) {
>>   			seq_printf(s, "\t%s\n", dev_name(attach_obj->dev));
>>   			attach_count++;
>>   		}
>> +		dma_resv_unlock(buf_obj->resv);
>>   
>>   		seq_printf(s, "Total %d devices attached\n\n",
>>   				attach_count);
>> diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h
>> index ec212cb27fdc..a8f8b2b812fd 100644
>> --- a/include/linux/dma-buf.h
>> +++ b/include/linux/dma-buf.h
>> @@ -42,6 +42,17 @@ struct dma_buf_ops {
>>   	  */
>>   	bool cache_sgt_mapping;
>>   
>> +	/**
>> +	 * @dynamic_mapping:
>> +	 *
>> +	 * If true the framework makes sure that the map/unmap_dma_buf
>> +	 * callbacks are always called with the dma_resv object locked.
>> +	 *
>> +	 * If false the framework makes ure that the map/unmap_dma_buf
>> +	 * callbacks are always called without the dma_resv object locked.
>> +	 */
>> +	bool dynamic_mapping;
>> +
>>   	/**
>>   	 * @attach:
>>   	 *
>> @@ -109,6 +120,9 @@ struct dma_buf_ops {
>>   	 * any other kind of sharing that the exporter might wish to make
>>   	 * available to buffer-users.
>>   	 *
>> +	 * This is always called with the dmabuf->resv object locked when
>> +	 * the dynamic_mapping flag is true.
>> +	 *
>>   	 * Returns:
>>   	 *
>>   	 * A &sg_table scatter list of or the backing storage of the DMA buffer,
>> @@ -327,6 +341,8 @@ struct dma_buf {
>>    * @sgt: cached mapping.
>>    * @dir: direction of cached mapping.
>>    * @priv: exporter specific attachment data.
>> + * @dynamic_mapping: true if dma_buf_map/unmap_attachment() is called with the
>> + * dma_resv lock held.
>>    *
>>    * This structure holds the attachment information between the dma_buf buffer
>>    * and its user device(s). The list contains one attachment struct per device
>> @@ -343,6 +359,7 @@ struct dma_buf_attachment {
>>   	struct list_head node;
>>   	struct sg_table *sgt;
>>   	enum dma_data_direction dir;
>> +	bool dynamic_mapping;
>>   	void *priv;
>>   };
>>   
>> @@ -394,10 +411,39 @@ static inline void get_dma_buf(struct dma_buf *dmabuf)
>>   	get_file(dmabuf->file);
>>   }
>>   
>> +/**
>> + * dma_buf_is_dynamic - check if a DMA-buf uses dynamic mappings.
>> + * @dmabuf: the DMA-buf to check
>> + *
>> + * Returns true if a DMA-buf exporter wants to be called with the dma_resv
>> + * locked, false if it doesn't wants to be called with the lock held.
>> + */
>> +static inline bool dma_buf_is_dynamic(struct dma_buf *dmabuf)
>> +{
>> +	return dmabuf->ops->dynamic_mapping;
>> +}
>> +
>> +/**
>> + * dma_buf_attachment_is_dynamic - check if a DMA-buf attachment uses dynamic
>> + * mappinsg
>> + * @attach: the DMA-buf attachment to check
>> + *
>> + * Returns true if a DMA-buf importer wants to call the map/unmap functions with
>> + * the dma_resv lock held.
>> + */
>> +static inline bool
>> +dma_buf_attachment_is_dynamic(struct dma_buf_attachment *attach)
>> +{
>> +	return attach->dynamic_mapping;
>> +}
>> +
>>   struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>> -							struct device *dev);
>> +					  struct device *dev);
>> +struct dma_buf_attachment *
>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
>> +		       bool dynamic_mapping);
>>   void dma_buf_detach(struct dma_buf *dmabuf,
>> -				struct dma_buf_attachment *dmabuf_attach);
>> +		    struct dma_buf_attachment *attach);
>>   
>>   struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info);
>>   
>> @@ -409,6 +455,7 @@ struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *,
>>   					enum dma_data_direction);
>>   void dma_buf_unmap_attachment(struct dma_buf_attachment *, struct sg_table *,
>>   				enum dma_data_direction);
>> +void dma_buf_move_notify(struct dma_buf *dma_buf);
>>   int dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
>>   			     enum dma_data_direction dir);
>>   int dma_buf_end_cpu_access(struct dma_buf *dma_buf,
>> -- 
>> 2.17.1
>>

Christian König Sept. 16, 2019, 12:23 p.m. UTC | #3

Ping? Any further comment on this or can't we merge at least the locking 
change?

Christian.

Am 11.09.19 um 12:53 schrieb Christian König:
> Am 03.09.19 um 10:05 schrieb Daniel Vetter:
>> On Thu, Aug 29, 2019 at 04:29:14PM +0200, Christian König wrote:
>>> This patch is a stripped down version of the locking changes
>>> necessary to support dynamic DMA-buf handling.
>>>
>>> For compatibility we cache the DMA-buf mapping as soon as
>>> exporter/importer disagree on the dynamic handling.
>>>
>>> Signed-off-by: Christian König <christian.koenig@amd.com>
>>> ---
>>>   drivers/dma-buf/dma-buf.c | 90 
>>> ++++++++++++++++++++++++++++++++++++---
>>>   include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
>>>   2 files changed, 133 insertions(+), 8 deletions(-)
>>>
>>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
>>> index 433d91d710e4..65052d52602b 100644
>>> --- a/drivers/dma-buf/dma-buf.c
>>> +++ b/drivers/dma-buf/dma-buf.c
>>> @@ -525,6 +525,10 @@ struct dma_buf *dma_buf_export(const struct 
>>> dma_buf_export_info *exp_info)
>>>           return ERR_PTR(-EINVAL);
>>>       }
>>>   +    if (WARN_ON(exp_info->ops->cache_sgt_mapping &&
>>> +            exp_info->ops->dynamic_mapping))
>>> +        return ERR_PTR(-EINVAL);
>>> +
>>>       if (!try_module_get(exp_info->owner))
>>>           return ERR_PTR(-ENOENT);
>>>   @@ -645,10 +649,11 @@ void dma_buf_put(struct dma_buf *dmabuf)
>>>   EXPORT_SYMBOL_GPL(dma_buf_put);
>>>     /**
>>> - * dma_buf_attach - Add the device to dma_buf's attachments list; 
>>> optionally,
>>> + * dma_buf_dynamic_attach - Add the device to dma_buf's attachments 
>>> list; optionally,
>>>    * calls attach() of dma_buf_ops to allow device-specific attach 
>>> functionality
>>> - * @dmabuf:    [in]    buffer to attach device to.
>>> - * @dev:    [in]    device to be attached.
>>> + * @dmabuf:        [in]    buffer to attach device to.
>>> + * @dev:        [in]    device to be attached.
>>> + * @dynamic_mapping:    [in]    calling convention for map/unmap
>>>    *
>>>    * Returns struct dma_buf_attachment pointer for this attachment. 
>>> Attachments
>>>    * must be cleaned up by calling dma_buf_detach().
>>> @@ -662,8 +667,9 @@ EXPORT_SYMBOL_GPL(dma_buf_put);
>>>    * accessible to @dev, and cannot be moved to a more suitable 
>>> place. This is
>>>    * indicated with the error code -EBUSY.
>>>    */
>>> -struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>> -                      struct device *dev)
>>> +struct dma_buf_attachment *
>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
>>> +               bool dynamic_mapping)
>>>   {
>>>       struct dma_buf_attachment *attach;
>>>       int ret;
>>> @@ -677,6 +683,7 @@ struct dma_buf_attachment *dma_buf_attach(struct 
>>> dma_buf *dmabuf,
>>>         attach->dev = dev;
>>>       attach->dmabuf = dmabuf;
>>> +    attach->dynamic_mapping = dynamic_mapping;
>>>         mutex_lock(&dmabuf->lock);
>>>   @@ -685,16 +692,64 @@ struct dma_buf_attachment 
>>> *dma_buf_attach(struct dma_buf *dmabuf,
>>>           if (ret)
>>>               goto err_attach;
>>>       }
>>> +    dma_resv_lock(dmabuf->resv, NULL);
>>>       list_add(&attach->node, &dmabuf->attachments);
>>> +    dma_resv_unlock(dmabuf->resv);
>>>         mutex_unlock(&dmabuf->lock);
>>>   +    /* When either the importer or the exporter can't handle dynamic
>>> +     * mappings we cache the mapping here to avoid issues with the
>>> +     * reservation object lock.
>>> +     */
>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
>>> +        dma_buf_is_dynamic(dmabuf)) {
>>> +        struct sg_table *sgt;
>>> +
>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
>>> +
>>> +        sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);
>> Now we're back to enforcing DMA_BIDI, which works nicely around the
>> locking pain, but apparently upsets the arm-soc folks who want to 
>> control
>> this better.
>
> Take another look at dma_buf_map_attachment(), we still try to get the 
> caching there for ARM.
>
> What we do here is to bidirectionally map the buffer to avoid the 
> locking hydra when importer and exporter disagree on locking.
>
> So the ARM folks can easily avoid that by switching to dynamic locking 
> for both.
>
> Regards,
> Christian.
>
>> That's why your previous version moved the caching into
>> map/unmap_sg, which resurrected the locking hydra.
>>
>> I think we're going a bit in circles here, and I don't have a good idea
>> either :-/
>> -Daniel
>>
>>> +        if (!sgt)
>>> +            sgt = ERR_PTR(-ENOMEM);
>>> +        if (IS_ERR(sgt)) {
>>> +            ret = PTR_ERR(sgt);
>>> +            goto err_unlock;
>>> +        }
>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>> +            dma_resv_unlock(attach->dmabuf->resv);
>>> +        attach->sgt = sgt;
>>> +        attach->dir = DMA_BIDIRECTIONAL;
>>> +    }
>>> +
>>>       return attach;
>>>     err_attach:
>>>       kfree(attach);
>>>       mutex_unlock(&dmabuf->lock);
>>>       return ERR_PTR(ret);
>>> +
>>> +err_unlock:
>>> +    if (dma_buf_is_dynamic(attach->dmabuf))
>>> +        dma_resv_unlock(attach->dmabuf->resv);
>>> +
>>> +    dma_buf_detach(dmabuf, attach);
>>> +    return ERR_PTR(ret);
>>> +}
>>> +EXPORT_SYMBOL_GPL(dma_buf_dynamic_attach);
>>> +
>>> +/**
>>> + * dma_buf_attach - Wrapper for dma_buf_dynamic_attach
>>> + * @dmabuf:    [in]    buffer to attach device to.
>>> + * @dev:    [in]    device to be attached.
>>> + *
>>> + * Wrapper to call dma_buf_dynamic_attach() for drivers which still 
>>> use a static
>>> + * mapping.
>>> + */
>>> +struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>> +                      struct device *dev)
>>> +{
>>> +    return dma_buf_dynamic_attach(dmabuf, dev, false);
>>>   }
>>>   EXPORT_SYMBOL_GPL(dma_buf_attach);
>>>   @@ -711,11 +766,20 @@ void dma_buf_detach(struct dma_buf *dmabuf, 
>>> struct dma_buf_attachment *attach)
>>>       if (WARN_ON(!dmabuf || !attach))
>>>           return;
>>>   -    if (attach->sgt)
>>> +    if (attach->sgt) {
>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
>>> +
>>>           dmabuf->ops->unmap_dma_buf(attach, attach->sgt, attach->dir);
>>>   +        if (dma_buf_is_dynamic(attach->dmabuf))
>>> +            dma_resv_unlock(attach->dmabuf->resv);
>>> +    }
>>> +
>>>       mutex_lock(&dmabuf->lock);
>>> +    dma_resv_lock(dmabuf->resv, NULL);
>>>       list_del(&attach->node);
>>> +    dma_resv_unlock(dmabuf->resv);
>>>       if (dmabuf->ops->detach)
>>>           dmabuf->ops->detach(dmabuf, attach);
>>>   @@ -749,6 +813,9 @@ struct sg_table *dma_buf_map_attachment(struct 
>>> dma_buf_attachment *attach,
>>>       if (WARN_ON(!attach || !attach->dmabuf))
>>>           return ERR_PTR(-EINVAL);
>>>   +    if (dma_buf_attachment_is_dynamic(attach))
>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>> +
>>>       if (attach->sgt) {
>>>           /*
>>>            * Two mappings with different directions for the same
>>> @@ -761,6 +828,9 @@ struct sg_table *dma_buf_map_attachment(struct 
>>> dma_buf_attachment *attach,
>>>           return attach->sgt;
>>>       }
>>>   +    if (dma_buf_is_dynamic(attach->dmabuf))
>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>> +
>>>       sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction);
>>>       if (!sg_table)
>>>           sg_table = ERR_PTR(-ENOMEM);
>>> @@ -793,9 +863,15 @@ void dma_buf_unmap_attachment(struct 
>>> dma_buf_attachment *attach,
>>>       if (WARN_ON(!attach || !attach->dmabuf || !sg_table))
>>>           return;
>>>   +    if (dma_buf_attachment_is_dynamic(attach))
>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>> +
>>>       if (attach->sgt == sg_table)
>>>           return;
>>>   +    if (dma_buf_is_dynamic(attach->dmabuf))
>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>> +
>>>       attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, direction);
>>>   }
>>>   EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment);
>>> @@ -1219,10 +1295,12 @@ static int dma_buf_debug_show(struct 
>>> seq_file *s, void *unused)
>>>           seq_puts(s, "\tAttached Devices:\n");
>>>           attach_count = 0;
>>>   +        dma_resv_lock(buf_obj->resv, NULL);
>>>           list_for_each_entry(attach_obj, &buf_obj->attachments, 
>>> node) {
>>>               seq_printf(s, "\t%s\n", dev_name(attach_obj->dev));
>>>               attach_count++;
>>>           }
>>> +        dma_resv_unlock(buf_obj->resv);
>>>             seq_printf(s, "Total %d devices attached\n\n",
>>>                   attach_count);
>>> diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h
>>> index ec212cb27fdc..a8f8b2b812fd 100644
>>> --- a/include/linux/dma-buf.h
>>> +++ b/include/linux/dma-buf.h
>>> @@ -42,6 +42,17 @@ struct dma_buf_ops {
>>>         */
>>>       bool cache_sgt_mapping;
>>>   +    /**
>>> +     * @dynamic_mapping:
>>> +     *
>>> +     * If true the framework makes sure that the map/unmap_dma_buf
>>> +     * callbacks are always called with the dma_resv object locked.
>>> +     *
>>> +     * If false the framework makes ure that the map/unmap_dma_buf
>>> +     * callbacks are always called without the dma_resv object locked.
>>> +     */
>>> +    bool dynamic_mapping;
>>> +
>>>       /**
>>>        * @attach:
>>>        *
>>> @@ -109,6 +120,9 @@ struct dma_buf_ops {
>>>        * any other kind of sharing that the exporter might wish to make
>>>        * available to buffer-users.
>>>        *
>>> +     * This is always called with the dmabuf->resv object locked when
>>> +     * the dynamic_mapping flag is true.
>>> +     *
>>>        * Returns:
>>>        *
>>>        * A &sg_table scatter list of or the backing storage of the 
>>> DMA buffer,
>>> @@ -327,6 +341,8 @@ struct dma_buf {
>>>    * @sgt: cached mapping.
>>>    * @dir: direction of cached mapping.
>>>    * @priv: exporter specific attachment data.
>>> + * @dynamic_mapping: true if dma_buf_map/unmap_attachment() is 
>>> called with the
>>> + * dma_resv lock held.
>>>    *
>>>    * This structure holds the attachment information between the 
>>> dma_buf buffer
>>>    * and its user device(s). The list contains one attachment struct 
>>> per device
>>> @@ -343,6 +359,7 @@ struct dma_buf_attachment {
>>>       struct list_head node;
>>>       struct sg_table *sgt;
>>>       enum dma_data_direction dir;
>>> +    bool dynamic_mapping;
>>>       void *priv;
>>>   };
>>>   @@ -394,10 +411,39 @@ static inline void get_dma_buf(struct 
>>> dma_buf *dmabuf)
>>>       get_file(dmabuf->file);
>>>   }
>>>   +/**
>>> + * dma_buf_is_dynamic - check if a DMA-buf uses dynamic mappings.
>>> + * @dmabuf: the DMA-buf to check
>>> + *
>>> + * Returns true if a DMA-buf exporter wants to be called with the 
>>> dma_resv
>>> + * locked, false if it doesn't wants to be called with the lock held.
>>> + */
>>> +static inline bool dma_buf_is_dynamic(struct dma_buf *dmabuf)
>>> +{
>>> +    return dmabuf->ops->dynamic_mapping;
>>> +}
>>> +
>>> +/**
>>> + * dma_buf_attachment_is_dynamic - check if a DMA-buf attachment 
>>> uses dynamic
>>> + * mappinsg
>>> + * @attach: the DMA-buf attachment to check
>>> + *
>>> + * Returns true if a DMA-buf importer wants to call the map/unmap 
>>> functions with
>>> + * the dma_resv lock held.
>>> + */
>>> +static inline bool
>>> +dma_buf_attachment_is_dynamic(struct dma_buf_attachment *attach)
>>> +{
>>> +    return attach->dynamic_mapping;
>>> +}
>>> +
>>>   struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>> -                            struct device *dev);
>>> +                      struct device *dev);
>>> +struct dma_buf_attachment *
>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
>>> +               bool dynamic_mapping);
>>>   void dma_buf_detach(struct dma_buf *dmabuf,
>>> -                struct dma_buf_attachment *dmabuf_attach);
>>> +            struct dma_buf_attachment *attach);
>>>     struct dma_buf *dma_buf_export(const struct dma_buf_export_info 
>>> *exp_info);
>>>   @@ -409,6 +455,7 @@ struct sg_table *dma_buf_map_attachment(struct 
>>> dma_buf_attachment *,
>>>                       enum dma_data_direction);
>>>   void dma_buf_unmap_attachment(struct dma_buf_attachment *, struct 
>>> sg_table *,
>>>                   enum dma_data_direction);
>>> +void dma_buf_move_notify(struct dma_buf *dma_buf);
>>>   int dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
>>>                    enum dma_data_direction dir);
>>>   int dma_buf_end_cpu_access(struct dma_buf *dma_buf,
>>> -- 
>>> 2.17.1
>>>
>

Daniel Vetter Sept. 17, 2019, 12:31 p.m. UTC | #4

On Mon, Sep 16, 2019 at 02:23:13PM +0200, Christian König wrote:
> Ping? Any further comment on this or can't we merge at least the locking
> change?

I was at plumbers ...
> 
> Christian.
> 
> Am 11.09.19 um 12:53 schrieb Christian König:
> > Am 03.09.19 um 10:05 schrieb Daniel Vetter:
> > > On Thu, Aug 29, 2019 at 04:29:14PM +0200, Christian König wrote:
> > > > This patch is a stripped down version of the locking changes
> > > > necessary to support dynamic DMA-buf handling.
> > > > 
> > > > For compatibility we cache the DMA-buf mapping as soon as
> > > > exporter/importer disagree on the dynamic handling.
> > > > 
> > > > Signed-off-by: Christian König <christian.koenig@amd.com>
> > > > ---
> > > >   drivers/dma-buf/dma-buf.c | 90
> > > > ++++++++++++++++++++++++++++++++++++---
> > > >   include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
> > > >   2 files changed, 133 insertions(+), 8 deletions(-)
> > > > 
> > > > diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> > > > index 433d91d710e4..65052d52602b 100644
> > > > --- a/drivers/dma-buf/dma-buf.c
> > > > +++ b/drivers/dma-buf/dma-buf.c
> > > > @@ -525,6 +525,10 @@ struct dma_buf *dma_buf_export(const struct
> > > > dma_buf_export_info *exp_info)
> > > >           return ERR_PTR(-EINVAL);
> > > >       }
> > > >   +    if (WARN_ON(exp_info->ops->cache_sgt_mapping &&
> > > > +            exp_info->ops->dynamic_mapping))
> > > > +        return ERR_PTR(-EINVAL);
> > > > +
> > > >       if (!try_module_get(exp_info->owner))
> > > >           return ERR_PTR(-ENOENT);
> > > >   @@ -645,10 +649,11 @@ void dma_buf_put(struct dma_buf *dmabuf)
> > > >   EXPORT_SYMBOL_GPL(dma_buf_put);
> > > >     /**
> > > > - * dma_buf_attach - Add the device to dma_buf's attachments
> > > > list; optionally,
> > > > + * dma_buf_dynamic_attach - Add the device to dma_buf's
> > > > attachments list; optionally,
> > > >    * calls attach() of dma_buf_ops to allow device-specific
> > > > attach functionality
> > > > - * @dmabuf:    [in]    buffer to attach device to.
> > > > - * @dev:    [in]    device to be attached.
> > > > + * @dmabuf:        [in]    buffer to attach device to.
> > > > + * @dev:        [in]    device to be attached.
> > > > + * @dynamic_mapping:    [in]    calling convention for map/unmap
> > > >    *
> > > >    * Returns struct dma_buf_attachment pointer for this
> > > > attachment. Attachments
> > > >    * must be cleaned up by calling dma_buf_detach().
> > > > @@ -662,8 +667,9 @@ EXPORT_SYMBOL_GPL(dma_buf_put);
> > > >    * accessible to @dev, and cannot be moved to a more suitable
> > > > place. This is
> > > >    * indicated with the error code -EBUSY.
> > > >    */
> > > > -struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> > > > -                      struct device *dev)
> > > > +struct dma_buf_attachment *
> > > > +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
> > > > +               bool dynamic_mapping)
> > > >   {
> > > >       struct dma_buf_attachment *attach;
> > > >       int ret;
> > > > @@ -677,6 +683,7 @@ struct dma_buf_attachment
> > > > *dma_buf_attach(struct dma_buf *dmabuf,
> > > >         attach->dev = dev;
> > > >       attach->dmabuf = dmabuf;
> > > > +    attach->dynamic_mapping = dynamic_mapping;
> > > >         mutex_lock(&dmabuf->lock);
> > > >   @@ -685,16 +692,64 @@ struct dma_buf_attachment
> > > > *dma_buf_attach(struct dma_buf *dmabuf,
> > > >           if (ret)
> > > >               goto err_attach;
> > > >       }
> > > > +    dma_resv_lock(dmabuf->resv, NULL);
> > > >       list_add(&attach->node, &dmabuf->attachments);
> > > > +    dma_resv_unlock(dmabuf->resv);
> > > >         mutex_unlock(&dmabuf->lock);
> > > >   +    /* When either the importer or the exporter can't handle dynamic
> > > > +     * mappings we cache the mapping here to avoid issues with the
> > > > +     * reservation object lock.
> > > > +     */
> > > > +    if (dma_buf_attachment_is_dynamic(attach) !=
> > > > +        dma_buf_is_dynamic(dmabuf)) {
> > > > +        struct sg_table *sgt;
> > > > +
> > > > +        if (dma_buf_is_dynamic(attach->dmabuf))
> > > > +            dma_resv_lock(attach->dmabuf->resv, NULL);
> > > > +
> > > > +        sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);
> > > Now we're back to enforcing DMA_BIDI, which works nicely around the
> > > locking pain, but apparently upsets the arm-soc folks who want to
> > > control
> > > this better.
> > 
> > Take another look at dma_buf_map_attachment(), we still try to get the
> > caching there for ARM.
> > 
> > What we do here is to bidirectionally map the buffer to avoid the
> > locking hydra when importer and exporter disagree on locking.
> > 
> > So the ARM folks can easily avoid that by switching to dynamic locking
> > for both.

So you still break the contract between importer and exporter, except not
for anything that's run in intel-gfx-ci so all is good?

The other issue with "we solve this with caching the mapping": Currently
map/unmap flush (at least on arm, at least on cases where it matters). If
you just return the cached sg, then we don't do the flushing anymore,
which might break importers/exporters in exactly the same way as just
giving them the wrong mapping. There's zero differences between a BIDI,
TO_CPU, or TO_DEVICE mapping, the only places where this matters is for
cache flushing.

So here's something that could actually work:
- We cache the mapping.
- We cache a bidirectional mapping.
- We put the right dma_sg_sync_for_cpu/device calls in place for map/unmap
  to give current importers/exporters the same behaviour they're used to
  now.

And yes the caching we've lifted might have broken something somewhere
already. But generally you only hear about that long time after because
arm vendors roll forward once every few years. Or something like that.
-Daniel

> > 
> > Regards,
> > Christian.
> > 
> > > That's why your previous version moved the caching into
> > > map/unmap_sg, which resurrected the locking hydra.
> > > 
> > > I think we're going a bit in circles here, and I don't have a good idea
> > > either :-/
> > > -Daniel
> > > 
> > > > +        if (!sgt)
> > > > +            sgt = ERR_PTR(-ENOMEM);
> > > > +        if (IS_ERR(sgt)) {
> > > > +            ret = PTR_ERR(sgt);
> > > > +            goto err_unlock;
> > > > +        }
> > > > +        if (dma_buf_is_dynamic(attach->dmabuf))
> > > > +            dma_resv_unlock(attach->dmabuf->resv);
> > > > +        attach->sgt = sgt;
> > > > +        attach->dir = DMA_BIDIRECTIONAL;
> > > > +    }
> > > > +
> > > >       return attach;
> > > >     err_attach:
> > > >       kfree(attach);
> > > >       mutex_unlock(&dmabuf->lock);
> > > >       return ERR_PTR(ret);
> > > > +
> > > > +err_unlock:
> > > > +    if (dma_buf_is_dynamic(attach->dmabuf))
> > > > +        dma_resv_unlock(attach->dmabuf->resv);
> > > > +
> > > > +    dma_buf_detach(dmabuf, attach);
> > > > +    return ERR_PTR(ret);
> > > > +}
> > > > +EXPORT_SYMBOL_GPL(dma_buf_dynamic_attach);
> > > > +
> > > > +/**
> > > > + * dma_buf_attach - Wrapper for dma_buf_dynamic_attach
> > > > + * @dmabuf:    [in]    buffer to attach device to.
> > > > + * @dev:    [in]    device to be attached.
> > > > + *
> > > > + * Wrapper to call dma_buf_dynamic_attach() for drivers which
> > > > still use a static
> > > > + * mapping.
> > > > + */
> > > > +struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> > > > +                      struct device *dev)
> > > > +{
> > > > +    return dma_buf_dynamic_attach(dmabuf, dev, false);
> > > >   }
> > > >   EXPORT_SYMBOL_GPL(dma_buf_attach);
> > > >   @@ -711,11 +766,20 @@ void dma_buf_detach(struct dma_buf
> > > > *dmabuf, struct dma_buf_attachment *attach)
> > > >       if (WARN_ON(!dmabuf || !attach))
> > > >           return;
> > > >   -    if (attach->sgt)
> > > > +    if (attach->sgt) {
> > > > +        if (dma_buf_is_dynamic(attach->dmabuf))
> > > > +            dma_resv_lock(attach->dmabuf->resv, NULL);
> > > > +
> > > >           dmabuf->ops->unmap_dma_buf(attach, attach->sgt, attach->dir);
> > > >   +        if (dma_buf_is_dynamic(attach->dmabuf))
> > > > +            dma_resv_unlock(attach->dmabuf->resv);
> > > > +    }
> > > > +
> > > >       mutex_lock(&dmabuf->lock);
> > > > +    dma_resv_lock(dmabuf->resv, NULL);
> > > >       list_del(&attach->node);
> > > > +    dma_resv_unlock(dmabuf->resv);
> > > >       if (dmabuf->ops->detach)
> > > >           dmabuf->ops->detach(dmabuf, attach);
> > > >   @@ -749,6 +813,9 @@ struct sg_table
> > > > *dma_buf_map_attachment(struct dma_buf_attachment *attach,
> > > >       if (WARN_ON(!attach || !attach->dmabuf))
> > > >           return ERR_PTR(-EINVAL);
> > > >   +    if (dma_buf_attachment_is_dynamic(attach))
> > > > +        dma_resv_assert_held(attach->dmabuf->resv);
> > > > +
> > > >       if (attach->sgt) {
> > > >           /*
> > > >            * Two mappings with different directions for the same
> > > > @@ -761,6 +828,9 @@ struct sg_table
> > > > *dma_buf_map_attachment(struct dma_buf_attachment *attach,
> > > >           return attach->sgt;
> > > >       }
> > > >   +    if (dma_buf_is_dynamic(attach->dmabuf))
> > > > +        dma_resv_assert_held(attach->dmabuf->resv);
> > > > +
> > > >       sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction);
> > > >       if (!sg_table)
> > > >           sg_table = ERR_PTR(-ENOMEM);
> > > > @@ -793,9 +863,15 @@ void dma_buf_unmap_attachment(struct
> > > > dma_buf_attachment *attach,
> > > >       if (WARN_ON(!attach || !attach->dmabuf || !sg_table))
> > > >           return;
> > > >   +    if (dma_buf_attachment_is_dynamic(attach))
> > > > +        dma_resv_assert_held(attach->dmabuf->resv);
> > > > +
> > > >       if (attach->sgt == sg_table)
> > > >           return;
> > > >   +    if (dma_buf_is_dynamic(attach->dmabuf))
> > > > +        dma_resv_assert_held(attach->dmabuf->resv);
> > > > +
> > > >       attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, direction);
> > > >   }
> > > >   EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment);
> > > > @@ -1219,10 +1295,12 @@ static int dma_buf_debug_show(struct
> > > > seq_file *s, void *unused)
> > > >           seq_puts(s, "\tAttached Devices:\n");
> > > >           attach_count = 0;
> > > >   +        dma_resv_lock(buf_obj->resv, NULL);
> > > >           list_for_each_entry(attach_obj, &buf_obj->attachments,
> > > > node) {
> > > >               seq_printf(s, "\t%s\n", dev_name(attach_obj->dev));
> > > >               attach_count++;
> > > >           }
> > > > +        dma_resv_unlock(buf_obj->resv);
> > > >             seq_printf(s, "Total %d devices attached\n\n",
> > > >                   attach_count);
> > > > diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h
> > > > index ec212cb27fdc..a8f8b2b812fd 100644
> > > > --- a/include/linux/dma-buf.h
> > > > +++ b/include/linux/dma-buf.h
> > > > @@ -42,6 +42,17 @@ struct dma_buf_ops {
> > > >         */
> > > >       bool cache_sgt_mapping;
> > > >   +    /**
> > > > +     * @dynamic_mapping:
> > > > +     *
> > > > +     * If true the framework makes sure that the map/unmap_dma_buf
> > > > +     * callbacks are always called with the dma_resv object locked.
> > > > +     *
> > > > +     * If false the framework makes ure that the map/unmap_dma_buf
> > > > +     * callbacks are always called without the dma_resv object locked.
> > > > +     */
> > > > +    bool dynamic_mapping;
> > > > +
> > > >       /**
> > > >        * @attach:
> > > >        *
> > > > @@ -109,6 +120,9 @@ struct dma_buf_ops {
> > > >        * any other kind of sharing that the exporter might wish to make
> > > >        * available to buffer-users.
> > > >        *
> > > > +     * This is always called with the dmabuf->resv object locked when
> > > > +     * the dynamic_mapping flag is true.
> > > > +     *
> > > >        * Returns:
> > > >        *
> > > >        * A &sg_table scatter list of or the backing storage of
> > > > the DMA buffer,
> > > > @@ -327,6 +341,8 @@ struct dma_buf {
> > > >    * @sgt: cached mapping.
> > > >    * @dir: direction of cached mapping.
> > > >    * @priv: exporter specific attachment data.
> > > > + * @dynamic_mapping: true if dma_buf_map/unmap_attachment() is
> > > > called with the
> > > > + * dma_resv lock held.
> > > >    *
> > > >    * This structure holds the attachment information between the
> > > > dma_buf buffer
> > > >    * and its user device(s). The list contains one attachment
> > > > struct per device
> > > > @@ -343,6 +359,7 @@ struct dma_buf_attachment {
> > > >       struct list_head node;
> > > >       struct sg_table *sgt;
> > > >       enum dma_data_direction dir;
> > > > +    bool dynamic_mapping;
> > > >       void *priv;
> > > >   };
> > > >   @@ -394,10 +411,39 @@ static inline void get_dma_buf(struct
> > > > dma_buf *dmabuf)
> > > >       get_file(dmabuf->file);
> > > >   }
> > > >   +/**
> > > > + * dma_buf_is_dynamic - check if a DMA-buf uses dynamic mappings.
> > > > + * @dmabuf: the DMA-buf to check
> > > > + *
> > > > + * Returns true if a DMA-buf exporter wants to be called with
> > > > the dma_resv
> > > > + * locked, false if it doesn't wants to be called with the lock held.
> > > > + */
> > > > +static inline bool dma_buf_is_dynamic(struct dma_buf *dmabuf)
> > > > +{
> > > > +    return dmabuf->ops->dynamic_mapping;
> > > > +}
> > > > +
> > > > +/**
> > > > + * dma_buf_attachment_is_dynamic - check if a DMA-buf
> > > > attachment uses dynamic
> > > > + * mappinsg
> > > > + * @attach: the DMA-buf attachment to check
> > > > + *
> > > > + * Returns true if a DMA-buf importer wants to call the
> > > > map/unmap functions with
> > > > + * the dma_resv lock held.
> > > > + */
> > > > +static inline bool
> > > > +dma_buf_attachment_is_dynamic(struct dma_buf_attachment *attach)
> > > > +{
> > > > +    return attach->dynamic_mapping;
> > > > +}
> > > > +
> > > >   struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> > > > -                            struct device *dev);
> > > > +                      struct device *dev);
> > > > +struct dma_buf_attachment *
> > > > +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
> > > > +               bool dynamic_mapping);
> > > >   void dma_buf_detach(struct dma_buf *dmabuf,
> > > > -                struct dma_buf_attachment *dmabuf_attach);
> > > > +            struct dma_buf_attachment *attach);
> > > >     struct dma_buf *dma_buf_export(const struct
> > > > dma_buf_export_info *exp_info);
> > > >   @@ -409,6 +455,7 @@ struct sg_table
> > > > *dma_buf_map_attachment(struct dma_buf_attachment *,
> > > >                       enum dma_data_direction);
> > > >   void dma_buf_unmap_attachment(struct dma_buf_attachment *,
> > > > struct sg_table *,
> > > >                   enum dma_data_direction);
> > > > +void dma_buf_move_notify(struct dma_buf *dma_buf);
> > > >   int dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
> > > >                    enum dma_data_direction dir);
> > > >   int dma_buf_end_cpu_access(struct dma_buf *dma_buf,
> > > > -- 
> > > > 2.17.1
> > > > 
> > 
>

Christian König Sept. 17, 2019, 12:40 p.m. UTC | #5

Am 17.09.19 um 14:31 schrieb Daniel Vetter:
> On Mon, Sep 16, 2019 at 02:23:13PM +0200, Christian König wrote:
>> Ping? Any further comment on this or can't we merge at least the locking
>> change?
> I was at plumbers ...
>> Christian.
>>
>> Am 11.09.19 um 12:53 schrieb Christian König:
>>> Am 03.09.19 um 10:05 schrieb Daniel Vetter:
>>>> On Thu, Aug 29, 2019 at 04:29:14PM +0200, Christian König wrote:
>>>>> This patch is a stripped down version of the locking changes
>>>>> necessary to support dynamic DMA-buf handling.
>>>>>
>>>>> For compatibility we cache the DMA-buf mapping as soon as
>>>>> exporter/importer disagree on the dynamic handling.
>>>>>
>>>>> Signed-off-by: Christian König <christian.koenig@amd.com>
>>>>> ---
>>>>>    drivers/dma-buf/dma-buf.c | 90
>>>>> ++++++++++++++++++++++++++++++++++++---
>>>>>    include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
>>>>>    2 files changed, 133 insertions(+), 8 deletions(-)
>>>>>
>>>>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
>>>>> index 433d91d710e4..65052d52602b 100644
>>>>> --- a/drivers/dma-buf/dma-buf.c
>>>>> +++ b/drivers/dma-buf/dma-buf.c
>>>>> @@ -525,6 +525,10 @@ struct dma_buf *dma_buf_export(const struct
>>>>> dma_buf_export_info *exp_info)
>>>>>            return ERR_PTR(-EINVAL);
>>>>>        }
>>>>>    +    if (WARN_ON(exp_info->ops->cache_sgt_mapping &&
>>>>> +            exp_info->ops->dynamic_mapping))
>>>>> +        return ERR_PTR(-EINVAL);
>>>>> +
>>>>>        if (!try_module_get(exp_info->owner))
>>>>>            return ERR_PTR(-ENOENT);
>>>>>    @@ -645,10 +649,11 @@ void dma_buf_put(struct dma_buf *dmabuf)
>>>>>    EXPORT_SYMBOL_GPL(dma_buf_put);
>>>>>      /**
>>>>> - * dma_buf_attach - Add the device to dma_buf's attachments
>>>>> list; optionally,
>>>>> + * dma_buf_dynamic_attach - Add the device to dma_buf's
>>>>> attachments list; optionally,
>>>>>     * calls attach() of dma_buf_ops to allow device-specific
>>>>> attach functionality
>>>>> - * @dmabuf:    [in]    buffer to attach device to.
>>>>> - * @dev:    [in]    device to be attached.
>>>>> + * @dmabuf:        [in]    buffer to attach device to.
>>>>> + * @dev:        [in]    device to be attached.
>>>>> + * @dynamic_mapping:    [in]    calling convention for map/unmap
>>>>>     *
>>>>>     * Returns struct dma_buf_attachment pointer for this
>>>>> attachment. Attachments
>>>>>     * must be cleaned up by calling dma_buf_detach().
>>>>> @@ -662,8 +667,9 @@ EXPORT_SYMBOL_GPL(dma_buf_put);
>>>>>     * accessible to @dev, and cannot be moved to a more suitable
>>>>> place. This is
>>>>>     * indicated with the error code -EBUSY.
>>>>>     */
>>>>> -struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>>>> -                      struct device *dev)
>>>>> +struct dma_buf_attachment *
>>>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
>>>>> +               bool dynamic_mapping)
>>>>>    {
>>>>>        struct dma_buf_attachment *attach;
>>>>>        int ret;
>>>>> @@ -677,6 +683,7 @@ struct dma_buf_attachment
>>>>> *dma_buf_attach(struct dma_buf *dmabuf,
>>>>>          attach->dev = dev;
>>>>>        attach->dmabuf = dmabuf;
>>>>> +    attach->dynamic_mapping = dynamic_mapping;
>>>>>          mutex_lock(&dmabuf->lock);
>>>>>    @@ -685,16 +692,64 @@ struct dma_buf_attachment
>>>>> *dma_buf_attach(struct dma_buf *dmabuf,
>>>>>            if (ret)
>>>>>                goto err_attach;
>>>>>        }
>>>>> +    dma_resv_lock(dmabuf->resv, NULL);
>>>>>        list_add(&attach->node, &dmabuf->attachments);
>>>>> +    dma_resv_unlock(dmabuf->resv);
>>>>>          mutex_unlock(&dmabuf->lock);
>>>>>    +    /* When either the importer or the exporter can't handle dynamic
>>>>> +     * mappings we cache the mapping here to avoid issues with the
>>>>> +     * reservation object lock.
>>>>> +     */
>>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
>>>>> +        dma_buf_is_dynamic(dmabuf)) {
>>>>> +        struct sg_table *sgt;
>>>>> +
>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
>>>>> +
>>>>> +        sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);
>>>> Now we're back to enforcing DMA_BIDI, which works nicely around the
>>>> locking pain, but apparently upsets the arm-soc folks who want to
>>>> control
>>>> this better.
>>> Take another look at dma_buf_map_attachment(), we still try to get the
>>> caching there for ARM.
>>>
>>> What we do here is to bidirectionally map the buffer to avoid the
>>> locking hydra when importer and exporter disagree on locking.
>>>
>>> So the ARM folks can easily avoid that by switching to dynamic locking
>>> for both.
> So you still break the contract between importer and exporter, except not
> for anything that's run in intel-gfx-ci so all is good?

No, the contract between importer and exporter stays exactly the same it 
is currently as long as you don't switch to dynamic dma-buf handling.

There is no functional change for the ARM folks here. The only change 
which takes effect is between i915 and amdgpu and that is perfectly 
covered by intel-gfx-ci.

Regards,
Christian.

>
> The other issue with "we solve this with caching the mapping": Currently
> map/unmap flush (at least on arm, at least on cases where it matters). If
> you just return the cached sg, then we don't do the flushing anymore,
> which might break importers/exporters in exactly the same way as just
> giving them the wrong mapping. There's zero differences between a BIDI,
> TO_CPU, or TO_DEVICE mapping, the only places where this matters is for
> cache flushing.
>
> So here's something that could actually work:
> - We cache the mapping.
> - We cache a bidirectional mapping.
> - We put the right dma_sg_sync_for_cpu/device calls in place for map/unmap
>    to give current importers/exporters the same behaviour they're used to
>    now.
>
> And yes the caching we've lifted might have broken something somewhere
> already. But generally you only hear about that long time after because
> arm vendors roll forward once every few years. Or something like that.
> -Daniel
>
>>> Regards,
>>> Christian.
>>>
>>>> That's why your previous version moved the caching into
>>>> map/unmap_sg, which resurrected the locking hydra.
>>>>
>>>> I think we're going a bit in circles here, and I don't have a good idea
>>>> either :-/
>>>> -Daniel
>>>>
>>>>> +        if (!sgt)
>>>>> +            sgt = ERR_PTR(-ENOMEM);
>>>>> +        if (IS_ERR(sgt)) {
>>>>> +            ret = PTR_ERR(sgt);
>>>>> +            goto err_unlock;
>>>>> +        }
>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>> +            dma_resv_unlock(attach->dmabuf->resv);
>>>>> +        attach->sgt = sgt;
>>>>> +        attach->dir = DMA_BIDIRECTIONAL;
>>>>> +    }
>>>>> +
>>>>>        return attach;
>>>>>      err_attach:
>>>>>        kfree(attach);
>>>>>        mutex_unlock(&dmabuf->lock);
>>>>>        return ERR_PTR(ret);
>>>>> +
>>>>> +err_unlock:
>>>>> +    if (dma_buf_is_dynamic(attach->dmabuf))
>>>>> +        dma_resv_unlock(attach->dmabuf->resv);
>>>>> +
>>>>> +    dma_buf_detach(dmabuf, attach);
>>>>> +    return ERR_PTR(ret);
>>>>> +}
>>>>> +EXPORT_SYMBOL_GPL(dma_buf_dynamic_attach);
>>>>> +
>>>>> +/**
>>>>> + * dma_buf_attach - Wrapper for dma_buf_dynamic_attach
>>>>> + * @dmabuf:    [in]    buffer to attach device to.
>>>>> + * @dev:    [in]    device to be attached.
>>>>> + *
>>>>> + * Wrapper to call dma_buf_dynamic_attach() for drivers which
>>>>> still use a static
>>>>> + * mapping.
>>>>> + */
>>>>> +struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>>>> +                      struct device *dev)
>>>>> +{
>>>>> +    return dma_buf_dynamic_attach(dmabuf, dev, false);
>>>>>    }
>>>>>    EXPORT_SYMBOL_GPL(dma_buf_attach);
>>>>>    @@ -711,11 +766,20 @@ void dma_buf_detach(struct dma_buf
>>>>> *dmabuf, struct dma_buf_attachment *attach)
>>>>>        if (WARN_ON(!dmabuf || !attach))
>>>>>            return;
>>>>>    -    if (attach->sgt)
>>>>> +    if (attach->sgt) {
>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
>>>>> +
>>>>>            dmabuf->ops->unmap_dma_buf(attach, attach->sgt, attach->dir);
>>>>>    +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>> +            dma_resv_unlock(attach->dmabuf->resv);
>>>>> +    }
>>>>> +
>>>>>        mutex_lock(&dmabuf->lock);
>>>>> +    dma_resv_lock(dmabuf->resv, NULL);
>>>>>        list_del(&attach->node);
>>>>> +    dma_resv_unlock(dmabuf->resv);
>>>>>        if (dmabuf->ops->detach)
>>>>>            dmabuf->ops->detach(dmabuf, attach);
>>>>>    @@ -749,6 +813,9 @@ struct sg_table
>>>>> *dma_buf_map_attachment(struct dma_buf_attachment *attach,
>>>>>        if (WARN_ON(!attach || !attach->dmabuf))
>>>>>            return ERR_PTR(-EINVAL);
>>>>>    +    if (dma_buf_attachment_is_dynamic(attach))
>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>>>> +
>>>>>        if (attach->sgt) {
>>>>>            /*
>>>>>             * Two mappings with different directions for the same
>>>>> @@ -761,6 +828,9 @@ struct sg_table
>>>>> *dma_buf_map_attachment(struct dma_buf_attachment *attach,
>>>>>            return attach->sgt;
>>>>>        }
>>>>>    +    if (dma_buf_is_dynamic(attach->dmabuf))
>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>>>> +
>>>>>        sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction);
>>>>>        if (!sg_table)
>>>>>            sg_table = ERR_PTR(-ENOMEM);
>>>>> @@ -793,9 +863,15 @@ void dma_buf_unmap_attachment(struct
>>>>> dma_buf_attachment *attach,
>>>>>        if (WARN_ON(!attach || !attach->dmabuf || !sg_table))
>>>>>            return;
>>>>>    +    if (dma_buf_attachment_is_dynamic(attach))
>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>>>> +
>>>>>        if (attach->sgt == sg_table)
>>>>>            return;
>>>>>    +    if (dma_buf_is_dynamic(attach->dmabuf))
>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>>>> +
>>>>>        attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, direction);
>>>>>    }
>>>>>    EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment);
>>>>> @@ -1219,10 +1295,12 @@ static int dma_buf_debug_show(struct
>>>>> seq_file *s, void *unused)
>>>>>            seq_puts(s, "\tAttached Devices:\n");
>>>>>            attach_count = 0;
>>>>>    +        dma_resv_lock(buf_obj->resv, NULL);
>>>>>            list_for_each_entry(attach_obj, &buf_obj->attachments,
>>>>> node) {
>>>>>                seq_printf(s, "\t%s\n", dev_name(attach_obj->dev));
>>>>>                attach_count++;
>>>>>            }
>>>>> +        dma_resv_unlock(buf_obj->resv);
>>>>>              seq_printf(s, "Total %d devices attached\n\n",
>>>>>                    attach_count);
>>>>> diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h
>>>>> index ec212cb27fdc..a8f8b2b812fd 100644
>>>>> --- a/include/linux/dma-buf.h
>>>>> +++ b/include/linux/dma-buf.h
>>>>> @@ -42,6 +42,17 @@ struct dma_buf_ops {
>>>>>          */
>>>>>        bool cache_sgt_mapping;
>>>>>    +    /**
>>>>> +     * @dynamic_mapping:
>>>>> +     *
>>>>> +     * If true the framework makes sure that the map/unmap_dma_buf
>>>>> +     * callbacks are always called with the dma_resv object locked.
>>>>> +     *
>>>>> +     * If false the framework makes ure that the map/unmap_dma_buf
>>>>> +     * callbacks are always called without the dma_resv object locked.
>>>>> +     */
>>>>> +    bool dynamic_mapping;
>>>>> +
>>>>>        /**
>>>>>         * @attach:
>>>>>         *
>>>>> @@ -109,6 +120,9 @@ struct dma_buf_ops {
>>>>>         * any other kind of sharing that the exporter might wish to make
>>>>>         * available to buffer-users.
>>>>>         *
>>>>> +     * This is always called with the dmabuf->resv object locked when
>>>>> +     * the dynamic_mapping flag is true.
>>>>> +     *
>>>>>         * Returns:
>>>>>         *
>>>>>         * A &sg_table scatter list of or the backing storage of
>>>>> the DMA buffer,
>>>>> @@ -327,6 +341,8 @@ struct dma_buf {
>>>>>     * @sgt: cached mapping.
>>>>>     * @dir: direction of cached mapping.
>>>>>     * @priv: exporter specific attachment data.
>>>>> + * @dynamic_mapping: true if dma_buf_map/unmap_attachment() is
>>>>> called with the
>>>>> + * dma_resv lock held.
>>>>>     *
>>>>>     * This structure holds the attachment information between the
>>>>> dma_buf buffer
>>>>>     * and its user device(s). The list contains one attachment
>>>>> struct per device
>>>>> @@ -343,6 +359,7 @@ struct dma_buf_attachment {
>>>>>        struct list_head node;
>>>>>        struct sg_table *sgt;
>>>>>        enum dma_data_direction dir;
>>>>> +    bool dynamic_mapping;
>>>>>        void *priv;
>>>>>    };
>>>>>    @@ -394,10 +411,39 @@ static inline void get_dma_buf(struct
>>>>> dma_buf *dmabuf)
>>>>>        get_file(dmabuf->file);
>>>>>    }
>>>>>    +/**
>>>>> + * dma_buf_is_dynamic - check if a DMA-buf uses dynamic mappings.
>>>>> + * @dmabuf: the DMA-buf to check
>>>>> + *
>>>>> + * Returns true if a DMA-buf exporter wants to be called with
>>>>> the dma_resv
>>>>> + * locked, false if it doesn't wants to be called with the lock held.
>>>>> + */
>>>>> +static inline bool dma_buf_is_dynamic(struct dma_buf *dmabuf)
>>>>> +{
>>>>> +    return dmabuf->ops->dynamic_mapping;
>>>>> +}
>>>>> +
>>>>> +/**
>>>>> + * dma_buf_attachment_is_dynamic - check if a DMA-buf
>>>>> attachment uses dynamic
>>>>> + * mappinsg
>>>>> + * @attach: the DMA-buf attachment to check
>>>>> + *
>>>>> + * Returns true if a DMA-buf importer wants to call the
>>>>> map/unmap functions with
>>>>> + * the dma_resv lock held.
>>>>> + */
>>>>> +static inline bool
>>>>> +dma_buf_attachment_is_dynamic(struct dma_buf_attachment *attach)
>>>>> +{
>>>>> +    return attach->dynamic_mapping;
>>>>> +}
>>>>> +
>>>>>    struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>>>> -                            struct device *dev);
>>>>> +                      struct device *dev);
>>>>> +struct dma_buf_attachment *
>>>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
>>>>> +               bool dynamic_mapping);
>>>>>    void dma_buf_detach(struct dma_buf *dmabuf,
>>>>> -                struct dma_buf_attachment *dmabuf_attach);
>>>>> +            struct dma_buf_attachment *attach);
>>>>>      struct dma_buf *dma_buf_export(const struct
>>>>> dma_buf_export_info *exp_info);
>>>>>    @@ -409,6 +455,7 @@ struct sg_table
>>>>> *dma_buf_map_attachment(struct dma_buf_attachment *,
>>>>>                        enum dma_data_direction);
>>>>>    void dma_buf_unmap_attachment(struct dma_buf_attachment *,
>>>>> struct sg_table *,
>>>>>                    enum dma_data_direction);
>>>>> +void dma_buf_move_notify(struct dma_buf *dma_buf);
>>>>>    int dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
>>>>>                     enum dma_data_direction dir);
>>>>>    int dma_buf_end_cpu_access(struct dma_buf *dma_buf,
>>>>> -- 
>>>>> 2.17.1
>>>>>

Daniel Vetter Sept. 17, 2019, 1:13 p.m. UTC | #6

On Tue, Sep 17, 2019 at 12:40:51PM +0000, Koenig, Christian wrote:
> Am 17.09.19 um 14:31 schrieb Daniel Vetter:
> > On Mon, Sep 16, 2019 at 02:23:13PM +0200, Christian König wrote:
> >> Ping? Any further comment on this or can't we merge at least the locking
> >> change?
> > I was at plumbers ...
> >> Christian.
> >>
> >> Am 11.09.19 um 12:53 schrieb Christian König:
> >>> Am 03.09.19 um 10:05 schrieb Daniel Vetter:
> >>>> On Thu, Aug 29, 2019 at 04:29:14PM +0200, Christian König wrote:
> >>>>> This patch is a stripped down version of the locking changes
> >>>>> necessary to support dynamic DMA-buf handling.
> >>>>>
> >>>>> For compatibility we cache the DMA-buf mapping as soon as
> >>>>> exporter/importer disagree on the dynamic handling.
> >>>>>
> >>>>> Signed-off-by: Christian König <christian.koenig@amd.com>
> >>>>> ---
> >>>>>    drivers/dma-buf/dma-buf.c | 90
> >>>>> ++++++++++++++++++++++++++++++++++++---
> >>>>>    include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
> >>>>>    2 files changed, 133 insertions(+), 8 deletions(-)
> >>>>>
> >>>>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> >>>>> index 433d91d710e4..65052d52602b 100644
> >>>>> --- a/drivers/dma-buf/dma-buf.c
> >>>>> +++ b/drivers/dma-buf/dma-buf.c
> >>>>> @@ -525,6 +525,10 @@ struct dma_buf *dma_buf_export(const struct
> >>>>> dma_buf_export_info *exp_info)
> >>>>>            return ERR_PTR(-EINVAL);
> >>>>>        }
> >>>>>    +    if (WARN_ON(exp_info->ops->cache_sgt_mapping &&
> >>>>> +            exp_info->ops->dynamic_mapping))
> >>>>> +        return ERR_PTR(-EINVAL);
> >>>>> +
> >>>>>        if (!try_module_get(exp_info->owner))
> >>>>>            return ERR_PTR(-ENOENT);
> >>>>>    @@ -645,10 +649,11 @@ void dma_buf_put(struct dma_buf *dmabuf)
> >>>>>    EXPORT_SYMBOL_GPL(dma_buf_put);
> >>>>>      /**
> >>>>> - * dma_buf_attach - Add the device to dma_buf's attachments
> >>>>> list; optionally,
> >>>>> + * dma_buf_dynamic_attach - Add the device to dma_buf's
> >>>>> attachments list; optionally,
> >>>>>     * calls attach() of dma_buf_ops to allow device-specific
> >>>>> attach functionality
> >>>>> - * @dmabuf:    [in]    buffer to attach device to.
> >>>>> - * @dev:    [in]    device to be attached.
> >>>>> + * @dmabuf:        [in]    buffer to attach device to.
> >>>>> + * @dev:        [in]    device to be attached.
> >>>>> + * @dynamic_mapping:    [in]    calling convention for map/unmap
> >>>>>     *
> >>>>>     * Returns struct dma_buf_attachment pointer for this
> >>>>> attachment. Attachments
> >>>>>     * must be cleaned up by calling dma_buf_detach().
> >>>>> @@ -662,8 +667,9 @@ EXPORT_SYMBOL_GPL(dma_buf_put);
> >>>>>     * accessible to @dev, and cannot be moved to a more suitable
> >>>>> place. This is
> >>>>>     * indicated with the error code -EBUSY.
> >>>>>     */
> >>>>> -struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>> -                      struct device *dev)
> >>>>> +struct dma_buf_attachment *
> >>>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
> >>>>> +               bool dynamic_mapping)
> >>>>>    {
> >>>>>        struct dma_buf_attachment *attach;
> >>>>>        int ret;
> >>>>> @@ -677,6 +683,7 @@ struct dma_buf_attachment
> >>>>> *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>>          attach->dev = dev;
> >>>>>        attach->dmabuf = dmabuf;
> >>>>> +    attach->dynamic_mapping = dynamic_mapping;
> >>>>>          mutex_lock(&dmabuf->lock);
> >>>>>    @@ -685,16 +692,64 @@ struct dma_buf_attachment
> >>>>> *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>>            if (ret)
> >>>>>                goto err_attach;
> >>>>>        }
> >>>>> +    dma_resv_lock(dmabuf->resv, NULL);
> >>>>>        list_add(&attach->node, &dmabuf->attachments);
> >>>>> +    dma_resv_unlock(dmabuf->resv);
> >>>>>          mutex_unlock(&dmabuf->lock);
> >>>>>    +    /* When either the importer or the exporter can't handle dynamic
> >>>>> +     * mappings we cache the mapping here to avoid issues with the
> >>>>> +     * reservation object lock.
> >>>>> +     */
> >>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
> >>>>> +        dma_buf_is_dynamic(dmabuf)) {
> >>>>> +        struct sg_table *sgt;
> >>>>> +
> >>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
> >>>>> +
> >>>>> +        sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);
> >>>> Now we're back to enforcing DMA_BIDI, which works nicely around the
> >>>> locking pain, but apparently upsets the arm-soc folks who want to
> >>>> control
> >>>> this better.
> >>> Take another look at dma_buf_map_attachment(), we still try to get the
> >>> caching there for ARM.
> >>>
> >>> What we do here is to bidirectionally map the buffer to avoid the
> >>> locking hydra when importer and exporter disagree on locking.
> >>>
> >>> So the ARM folks can easily avoid that by switching to dynamic locking
> >>> for both.
> > So you still break the contract between importer and exporter, except not
> > for anything that's run in intel-gfx-ci so all is good?
> 
> No, the contract between importer and exporter stays exactly the same it 
> is currently as long as you don't switch to dynamic dma-buf handling.
> 
> There is no functional change for the ARM folks here. The only change 
> which takes effect is between i915 and amdgpu and that is perfectly 
> covered by intel-gfx-ci.

There's people who want to run amdgpu on ARM? Also, x86 doesn't have cache
flushing in the dma-api, so naturally this isn't any issue for us (we
still have cache flushing in actual hw, but that's a different topic). So
"works on x86" isn't really a great way to justify what we do here I
think.
-Daniel

> 
> Regards,
> Christian.
> 
> >
> > The other issue with "we solve this with caching the mapping": Currently
> > map/unmap flush (at least on arm, at least on cases where it matters). If
> > you just return the cached sg, then we don't do the flushing anymore,
> > which might break importers/exporters in exactly the same way as just
> > giving them the wrong mapping. There's zero differences between a BIDI,
> > TO_CPU, or TO_DEVICE mapping, the only places where this matters is for
> > cache flushing.
> >
> > So here's something that could actually work:
> > - We cache the mapping.
> > - We cache a bidirectional mapping.
> > - We put the right dma_sg_sync_for_cpu/device calls in place for map/unmap
> >    to give current importers/exporters the same behaviour they're used to
> >    now.
> >
> > And yes the caching we've lifted might have broken something somewhere
> > already. But generally you only hear about that long time after because
> > arm vendors roll forward once every few years. Or something like that.
> > -Daniel
> >
> >>> Regards,
> >>> Christian.
> >>>
> >>>> That's why your previous version moved the caching into
> >>>> map/unmap_sg, which resurrected the locking hydra.
> >>>>
> >>>> I think we're going a bit in circles here, and I don't have a good idea
> >>>> either :-/
> >>>> -Daniel
> >>>>
> >>>>> +        if (!sgt)
> >>>>> +            sgt = ERR_PTR(-ENOMEM);
> >>>>> +        if (IS_ERR(sgt)) {
> >>>>> +            ret = PTR_ERR(sgt);
> >>>>> +            goto err_unlock;
> >>>>> +        }
> >>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>> +            dma_resv_unlock(attach->dmabuf->resv);
> >>>>> +        attach->sgt = sgt;
> >>>>> +        attach->dir = DMA_BIDIRECTIONAL;
> >>>>> +    }
> >>>>> +
> >>>>>        return attach;
> >>>>>      err_attach:
> >>>>>        kfree(attach);
> >>>>>        mutex_unlock(&dmabuf->lock);
> >>>>>        return ERR_PTR(ret);
> >>>>> +
> >>>>> +err_unlock:
> >>>>> +    if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>> +        dma_resv_unlock(attach->dmabuf->resv);
> >>>>> +
> >>>>> +    dma_buf_detach(dmabuf, attach);
> >>>>> +    return ERR_PTR(ret);
> >>>>> +}
> >>>>> +EXPORT_SYMBOL_GPL(dma_buf_dynamic_attach);
> >>>>> +
> >>>>> +/**
> >>>>> + * dma_buf_attach - Wrapper for dma_buf_dynamic_attach
> >>>>> + * @dmabuf:    [in]    buffer to attach device to.
> >>>>> + * @dev:    [in]    device to be attached.
> >>>>> + *
> >>>>> + * Wrapper to call dma_buf_dynamic_attach() for drivers which
> >>>>> still use a static
> >>>>> + * mapping.
> >>>>> + */
> >>>>> +struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>> +                      struct device *dev)
> >>>>> +{
> >>>>> +    return dma_buf_dynamic_attach(dmabuf, dev, false);
> >>>>>    }
> >>>>>    EXPORT_SYMBOL_GPL(dma_buf_attach);
> >>>>>    @@ -711,11 +766,20 @@ void dma_buf_detach(struct dma_buf
> >>>>> *dmabuf, struct dma_buf_attachment *attach)
> >>>>>        if (WARN_ON(!dmabuf || !attach))
> >>>>>            return;
> >>>>>    -    if (attach->sgt)
> >>>>> +    if (attach->sgt) {
> >>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
> >>>>> +
> >>>>>            dmabuf->ops->unmap_dma_buf(attach, attach->sgt, attach->dir);
> >>>>>    +        if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>> +            dma_resv_unlock(attach->dmabuf->resv);
> >>>>> +    }
> >>>>> +
> >>>>>        mutex_lock(&dmabuf->lock);
> >>>>> +    dma_resv_lock(dmabuf->resv, NULL);
> >>>>>        list_del(&attach->node);
> >>>>> +    dma_resv_unlock(dmabuf->resv);
> >>>>>        if (dmabuf->ops->detach)
> >>>>>            dmabuf->ops->detach(dmabuf, attach);
> >>>>>    @@ -749,6 +813,9 @@ struct sg_table
> >>>>> *dma_buf_map_attachment(struct dma_buf_attachment *attach,
> >>>>>        if (WARN_ON(!attach || !attach->dmabuf))
> >>>>>            return ERR_PTR(-EINVAL);
> >>>>>    +    if (dma_buf_attachment_is_dynamic(attach))
> >>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
> >>>>> +
> >>>>>        if (attach->sgt) {
> >>>>>            /*
> >>>>>             * Two mappings with different directions for the same
> >>>>> @@ -761,6 +828,9 @@ struct sg_table
> >>>>> *dma_buf_map_attachment(struct dma_buf_attachment *attach,
> >>>>>            return attach->sgt;
> >>>>>        }
> >>>>>    +    if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
> >>>>> +
> >>>>>        sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction);
> >>>>>        if (!sg_table)
> >>>>>            sg_table = ERR_PTR(-ENOMEM);
> >>>>> @@ -793,9 +863,15 @@ void dma_buf_unmap_attachment(struct
> >>>>> dma_buf_attachment *attach,
> >>>>>        if (WARN_ON(!attach || !attach->dmabuf || !sg_table))
> >>>>>            return;
> >>>>>    +    if (dma_buf_attachment_is_dynamic(attach))
> >>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
> >>>>> +
> >>>>>        if (attach->sgt == sg_table)
> >>>>>            return;
> >>>>>    +    if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
> >>>>> +
> >>>>>        attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, direction);
> >>>>>    }
> >>>>>    EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment);
> >>>>> @@ -1219,10 +1295,12 @@ static int dma_buf_debug_show(struct
> >>>>> seq_file *s, void *unused)
> >>>>>            seq_puts(s, "\tAttached Devices:\n");
> >>>>>            attach_count = 0;
> >>>>>    +        dma_resv_lock(buf_obj->resv, NULL);
> >>>>>            list_for_each_entry(attach_obj, &buf_obj->attachments,
> >>>>> node) {
> >>>>>                seq_printf(s, "\t%s\n", dev_name(attach_obj->dev));
> >>>>>                attach_count++;
> >>>>>            }
> >>>>> +        dma_resv_unlock(buf_obj->resv);
> >>>>>              seq_printf(s, "Total %d devices attached\n\n",
> >>>>>                    attach_count);
> >>>>> diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h
> >>>>> index ec212cb27fdc..a8f8b2b812fd 100644
> >>>>> --- a/include/linux/dma-buf.h
> >>>>> +++ b/include/linux/dma-buf.h
> >>>>> @@ -42,6 +42,17 @@ struct dma_buf_ops {
> >>>>>          */
> >>>>>        bool cache_sgt_mapping;
> >>>>>    +    /**
> >>>>> +     * @dynamic_mapping:
> >>>>> +     *
> >>>>> +     * If true the framework makes sure that the map/unmap_dma_buf
> >>>>> +     * callbacks are always called with the dma_resv object locked.
> >>>>> +     *
> >>>>> +     * If false the framework makes ure that the map/unmap_dma_buf
> >>>>> +     * callbacks are always called without the dma_resv object locked.
> >>>>> +     */
> >>>>> +    bool dynamic_mapping;
> >>>>> +
> >>>>>        /**
> >>>>>         * @attach:
> >>>>>         *
> >>>>> @@ -109,6 +120,9 @@ struct dma_buf_ops {
> >>>>>         * any other kind of sharing that the exporter might wish to make
> >>>>>         * available to buffer-users.
> >>>>>         *
> >>>>> +     * This is always called with the dmabuf->resv object locked when
> >>>>> +     * the dynamic_mapping flag is true.
> >>>>> +     *
> >>>>>         * Returns:
> >>>>>         *
> >>>>>         * A &sg_table scatter list of or the backing storage of
> >>>>> the DMA buffer,
> >>>>> @@ -327,6 +341,8 @@ struct dma_buf {
> >>>>>     * @sgt: cached mapping.
> >>>>>     * @dir: direction of cached mapping.
> >>>>>     * @priv: exporter specific attachment data.
> >>>>> + * @dynamic_mapping: true if dma_buf_map/unmap_attachment() is
> >>>>> called with the
> >>>>> + * dma_resv lock held.
> >>>>>     *
> >>>>>     * This structure holds the attachment information between the
> >>>>> dma_buf buffer
> >>>>>     * and its user device(s). The list contains one attachment
> >>>>> struct per device
> >>>>> @@ -343,6 +359,7 @@ struct dma_buf_attachment {
> >>>>>        struct list_head node;
> >>>>>        struct sg_table *sgt;
> >>>>>        enum dma_data_direction dir;
> >>>>> +    bool dynamic_mapping;
> >>>>>        void *priv;
> >>>>>    };
> >>>>>    @@ -394,10 +411,39 @@ static inline void get_dma_buf(struct
> >>>>> dma_buf *dmabuf)
> >>>>>        get_file(dmabuf->file);
> >>>>>    }
> >>>>>    +/**
> >>>>> + * dma_buf_is_dynamic - check if a DMA-buf uses dynamic mappings.
> >>>>> + * @dmabuf: the DMA-buf to check
> >>>>> + *
> >>>>> + * Returns true if a DMA-buf exporter wants to be called with
> >>>>> the dma_resv
> >>>>> + * locked, false if it doesn't wants to be called with the lock held.
> >>>>> + */
> >>>>> +static inline bool dma_buf_is_dynamic(struct dma_buf *dmabuf)
> >>>>> +{
> >>>>> +    return dmabuf->ops->dynamic_mapping;
> >>>>> +}
> >>>>> +
> >>>>> +/**
> >>>>> + * dma_buf_attachment_is_dynamic - check if a DMA-buf
> >>>>> attachment uses dynamic
> >>>>> + * mappinsg
> >>>>> + * @attach: the DMA-buf attachment to check
> >>>>> + *
> >>>>> + * Returns true if a DMA-buf importer wants to call the
> >>>>> map/unmap functions with
> >>>>> + * the dma_resv lock held.
> >>>>> + */
> >>>>> +static inline bool
> >>>>> +dma_buf_attachment_is_dynamic(struct dma_buf_attachment *attach)
> >>>>> +{
> >>>>> +    return attach->dynamic_mapping;
> >>>>> +}
> >>>>> +
> >>>>>    struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>> -                            struct device *dev);
> >>>>> +                      struct device *dev);
> >>>>> +struct dma_buf_attachment *
> >>>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
> >>>>> +               bool dynamic_mapping);
> >>>>>    void dma_buf_detach(struct dma_buf *dmabuf,
> >>>>> -                struct dma_buf_attachment *dmabuf_attach);
> >>>>> +            struct dma_buf_attachment *attach);
> >>>>>      struct dma_buf *dma_buf_export(const struct
> >>>>> dma_buf_export_info *exp_info);
> >>>>>    @@ -409,6 +455,7 @@ struct sg_table
> >>>>> *dma_buf_map_attachment(struct dma_buf_attachment *,
> >>>>>                        enum dma_data_direction);
> >>>>>    void dma_buf_unmap_attachment(struct dma_buf_attachment *,
> >>>>> struct sg_table *,
> >>>>>                    enum dma_data_direction);
> >>>>> +void dma_buf_move_notify(struct dma_buf *dma_buf);
> >>>>>    int dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
> >>>>>                     enum dma_data_direction dir);
> >>>>>    int dma_buf_end_cpu_access(struct dma_buf *dma_buf,
> >>>>> -- 
> >>>>> 2.17.1
> >>>>>
>

Christian König Sept. 17, 2019, 1:24 p.m. UTC | #7

Am 17.09.19 um 15:13 schrieb Daniel Vetter:
> On Tue, Sep 17, 2019 at 12:40:51PM +0000, Koenig, Christian wrote:
>> Am 17.09.19 um 14:31 schrieb Daniel Vetter:
>>> On Mon, Sep 16, 2019 at 02:23:13PM +0200, Christian König wrote:
>>>> Ping? Any further comment on this or can't we merge at least the locking
>>>> change?
>>> I was at plumbers ...
>>>> Christian.
>>>>
>>>> Am 11.09.19 um 12:53 schrieb Christian König:
>>>>> Am 03.09.19 um 10:05 schrieb Daniel Vetter:
>>>>>> On Thu, Aug 29, 2019 at 04:29:14PM +0200, Christian König wrote:
>>>>>>> This patch is a stripped down version of the locking changes
>>>>>>> necessary to support dynamic DMA-buf handling.
>>>>>>>
>>>>>>> For compatibility we cache the DMA-buf mapping as soon as
>>>>>>> exporter/importer disagree on the dynamic handling.
>>>>>>>
>>>>>>> Signed-off-by: Christian König <christian.koenig@amd.com>
>>>>>>> ---
>>>>>>>     drivers/dma-buf/dma-buf.c | 90
>>>>>>> ++++++++++++++++++++++++++++++++++++---
>>>>>>>     include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
>>>>>>>     2 files changed, 133 insertions(+), 8 deletions(-)
>>>>>>>
>>>>>>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
>>>>>>> index 433d91d710e4..65052d52602b 100644
>>>>>>> --- a/drivers/dma-buf/dma-buf.c
>>>>>>> +++ b/drivers/dma-buf/dma-buf.c
>>>>>>> @@ -525,6 +525,10 @@ struct dma_buf *dma_buf_export(const struct
>>>>>>> dma_buf_export_info *exp_info)
>>>>>>>             return ERR_PTR(-EINVAL);
>>>>>>>         }
>>>>>>>     +    if (WARN_ON(exp_info->ops->cache_sgt_mapping &&
>>>>>>> +            exp_info->ops->dynamic_mapping))
>>>>>>> +        return ERR_PTR(-EINVAL);
>>>>>>> +
>>>>>>>         if (!try_module_get(exp_info->owner))
>>>>>>>             return ERR_PTR(-ENOENT);
>>>>>>>     @@ -645,10 +649,11 @@ void dma_buf_put(struct dma_buf *dmabuf)
>>>>>>>     EXPORT_SYMBOL_GPL(dma_buf_put);
>>>>>>>       /**
>>>>>>> - * dma_buf_attach - Add the device to dma_buf's attachments
>>>>>>> list; optionally,
>>>>>>> + * dma_buf_dynamic_attach - Add the device to dma_buf's
>>>>>>> attachments list; optionally,
>>>>>>>      * calls attach() of dma_buf_ops to allow device-specific
>>>>>>> attach functionality
>>>>>>> - * @dmabuf:    [in]    buffer to attach device to.
>>>>>>> - * @dev:    [in]    device to be attached.
>>>>>>> + * @dmabuf:        [in]    buffer to attach device to.
>>>>>>> + * @dev:        [in]    device to be attached.
>>>>>>> + * @dynamic_mapping:    [in]    calling convention for map/unmap
>>>>>>>      *
>>>>>>>      * Returns struct dma_buf_attachment pointer for this
>>>>>>> attachment. Attachments
>>>>>>>      * must be cleaned up by calling dma_buf_detach().
>>>>>>> @@ -662,8 +667,9 @@ EXPORT_SYMBOL_GPL(dma_buf_put);
>>>>>>>      * accessible to @dev, and cannot be moved to a more suitable
>>>>>>> place. This is
>>>>>>>      * indicated with the error code -EBUSY.
>>>>>>>      */
>>>>>>> -struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>>>>>> -                      struct device *dev)
>>>>>>> +struct dma_buf_attachment *
>>>>>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
>>>>>>> +               bool dynamic_mapping)
>>>>>>>     {
>>>>>>>         struct dma_buf_attachment *attach;
>>>>>>>         int ret;
>>>>>>> @@ -677,6 +683,7 @@ struct dma_buf_attachment
>>>>>>> *dma_buf_attach(struct dma_buf *dmabuf,
>>>>>>>           attach->dev = dev;
>>>>>>>         attach->dmabuf = dmabuf;
>>>>>>> +    attach->dynamic_mapping = dynamic_mapping;
>>>>>>>           mutex_lock(&dmabuf->lock);
>>>>>>>     @@ -685,16 +692,64 @@ struct dma_buf_attachment
>>>>>>> *dma_buf_attach(struct dma_buf *dmabuf,
>>>>>>>             if (ret)
>>>>>>>                 goto err_attach;
>>>>>>>         }
>>>>>>> +    dma_resv_lock(dmabuf->resv, NULL);
>>>>>>>         list_add(&attach->node, &dmabuf->attachments);
>>>>>>> +    dma_resv_unlock(dmabuf->resv);
>>>>>>>           mutex_unlock(&dmabuf->lock);
>>>>>>>     +    /* When either the importer or the exporter can't handle dynamic
>>>>>>> +     * mappings we cache the mapping here to avoid issues with the
>>>>>>> +     * reservation object lock.
>>>>>>> +     */
>>>>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
>>>>>>> +        dma_buf_is_dynamic(dmabuf)) {
>>>>>>> +        struct sg_table *sgt;
>>>>>>> +
>>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
>>>>>>> +
>>>>>>> +        sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);
>>>>>> Now we're back to enforcing DMA_BIDI, which works nicely around the
>>>>>> locking pain, but apparently upsets the arm-soc folks who want to
>>>>>> control
>>>>>> this better.
>>>>> Take another look at dma_buf_map_attachment(), we still try to get the
>>>>> caching there for ARM.
>>>>>
>>>>> What we do here is to bidirectionally map the buffer to avoid the
>>>>> locking hydra when importer and exporter disagree on locking.
>>>>>
>>>>> So the ARM folks can easily avoid that by switching to dynamic locking
>>>>> for both.
>>> So you still break the contract between importer and exporter, except not
>>> for anything that's run in intel-gfx-ci so all is good?
>> No, the contract between importer and exporter stays exactly the same it
>> is currently as long as you don't switch to dynamic dma-buf handling.
>>
>> There is no functional change for the ARM folks here. The only change
>> which takes effect is between i915 and amdgpu and that is perfectly
>> covered by intel-gfx-ci.
> There's people who want to run amdgpu on ARM?

Sure there are, we even recently fixed some bugs for this.

But as far as I know there is no one currently which is affect by this 
change on ARM with amdgpu.

> Also, x86 doesn't have cache
> flushing in the dma-api, so naturally this isn't any issue for us (we
> still have cache flushing in actual hw, but that's a different topic). So
> "works on x86" isn't really a great way to justify what we do here I
> think.

Well it is the exact same caching we previously had as well, so there is 
absolutely no functional change here except that we now explicitly note 
that amdgpu always needs bidirectional mappings.

I agree that we should get rid of this caching as soon as possible, but 
we should not fix things which where broken before.

On the other hand adding dma_sg_sync_for_cpu/device sounds like 
something we could easily add separately to the caching if you think 
that this will help.

Christian.

> -Daniel
>
>> Regards,
>> Christian.
>>
>>> The other issue with "we solve this with caching the mapping": Currently
>>> map/unmap flush (at least on arm, at least on cases where it matters). If
>>> you just return the cached sg, then we don't do the flushing anymore,
>>> which might break importers/exporters in exactly the same way as just
>>> giving them the wrong mapping. There's zero differences between a BIDI,
>>> TO_CPU, or TO_DEVICE mapping, the only places where this matters is for
>>> cache flushing.
>>>
>>> So here's something that could actually work:
>>> - We cache the mapping.
>>> - We cache a bidirectional mapping.
>>> - We put the right dma_sg_sync_for_cpu/device calls in place for map/unmap
>>>     to give current importers/exporters the same behaviour they're used to
>>>     now.
>>>
>>> And yes the caching we've lifted might have broken something somewhere
>>> already. But generally you only hear about that long time after because
>>> arm vendors roll forward once every few years. Or something like that.
>>> -Daniel
>>>
>>>>> Regards,
>>>>> Christian.
>>>>>
>>>>>> That's why your previous version moved the caching into
>>>>>> map/unmap_sg, which resurrected the locking hydra.
>>>>>>
>>>>>> I think we're going a bit in circles here, and I don't have a good idea
>>>>>> either :-/
>>>>>> -Daniel
>>>>>>
>>>>>>> +        if (!sgt)
>>>>>>> +            sgt = ERR_PTR(-ENOMEM);
>>>>>>> +        if (IS_ERR(sgt)) {
>>>>>>> +            ret = PTR_ERR(sgt);
>>>>>>> +            goto err_unlock;
>>>>>>> +        }
>>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>> +            dma_resv_unlock(attach->dmabuf->resv);
>>>>>>> +        attach->sgt = sgt;
>>>>>>> +        attach->dir = DMA_BIDIRECTIONAL;
>>>>>>> +    }
>>>>>>> +
>>>>>>>         return attach;
>>>>>>>       err_attach:
>>>>>>>         kfree(attach);
>>>>>>>         mutex_unlock(&dmabuf->lock);
>>>>>>>         return ERR_PTR(ret);
>>>>>>> +
>>>>>>> +err_unlock:
>>>>>>> +    if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>> +        dma_resv_unlock(attach->dmabuf->resv);
>>>>>>> +
>>>>>>> +    dma_buf_detach(dmabuf, attach);
>>>>>>> +    return ERR_PTR(ret);
>>>>>>> +}
>>>>>>> +EXPORT_SYMBOL_GPL(dma_buf_dynamic_attach);
>>>>>>> +
>>>>>>> +/**
>>>>>>> + * dma_buf_attach - Wrapper for dma_buf_dynamic_attach
>>>>>>> + * @dmabuf:    [in]    buffer to attach device to.
>>>>>>> + * @dev:    [in]    device to be attached.
>>>>>>> + *
>>>>>>> + * Wrapper to call dma_buf_dynamic_attach() for drivers which
>>>>>>> still use a static
>>>>>>> + * mapping.
>>>>>>> + */
>>>>>>> +struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>>>>>> +                      struct device *dev)
>>>>>>> +{
>>>>>>> +    return dma_buf_dynamic_attach(dmabuf, dev, false);
>>>>>>>     }
>>>>>>>     EXPORT_SYMBOL_GPL(dma_buf_attach);
>>>>>>>     @@ -711,11 +766,20 @@ void dma_buf_detach(struct dma_buf
>>>>>>> *dmabuf, struct dma_buf_attachment *attach)
>>>>>>>         if (WARN_ON(!dmabuf || !attach))
>>>>>>>             return;
>>>>>>>     -    if (attach->sgt)
>>>>>>> +    if (attach->sgt) {
>>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
>>>>>>> +
>>>>>>>             dmabuf->ops->unmap_dma_buf(attach, attach->sgt, attach->dir);
>>>>>>>     +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>> +            dma_resv_unlock(attach->dmabuf->resv);
>>>>>>> +    }
>>>>>>> +
>>>>>>>         mutex_lock(&dmabuf->lock);
>>>>>>> +    dma_resv_lock(dmabuf->resv, NULL);
>>>>>>>         list_del(&attach->node);
>>>>>>> +    dma_resv_unlock(dmabuf->resv);
>>>>>>>         if (dmabuf->ops->detach)
>>>>>>>             dmabuf->ops->detach(dmabuf, attach);
>>>>>>>     @@ -749,6 +813,9 @@ struct sg_table
>>>>>>> *dma_buf_map_attachment(struct dma_buf_attachment *attach,
>>>>>>>         if (WARN_ON(!attach || !attach->dmabuf))
>>>>>>>             return ERR_PTR(-EINVAL);
>>>>>>>     +    if (dma_buf_attachment_is_dynamic(attach))
>>>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>>>>>> +
>>>>>>>         if (attach->sgt) {
>>>>>>>             /*
>>>>>>>              * Two mappings with different directions for the same
>>>>>>> @@ -761,6 +828,9 @@ struct sg_table
>>>>>>> *dma_buf_map_attachment(struct dma_buf_attachment *attach,
>>>>>>>             return attach->sgt;
>>>>>>>         }
>>>>>>>     +    if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>>>>>> +
>>>>>>>         sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction);
>>>>>>>         if (!sg_table)
>>>>>>>             sg_table = ERR_PTR(-ENOMEM);
>>>>>>> @@ -793,9 +863,15 @@ void dma_buf_unmap_attachment(struct
>>>>>>> dma_buf_attachment *attach,
>>>>>>>         if (WARN_ON(!attach || !attach->dmabuf || !sg_table))
>>>>>>>             return;
>>>>>>>     +    if (dma_buf_attachment_is_dynamic(attach))
>>>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>>>>>> +
>>>>>>>         if (attach->sgt == sg_table)
>>>>>>>             return;
>>>>>>>     +    if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
>>>>>>> +
>>>>>>>         attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, direction);
>>>>>>>     }
>>>>>>>     EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment);
>>>>>>> @@ -1219,10 +1295,12 @@ static int dma_buf_debug_show(struct
>>>>>>> seq_file *s, void *unused)
>>>>>>>             seq_puts(s, "\tAttached Devices:\n");
>>>>>>>             attach_count = 0;
>>>>>>>     +        dma_resv_lock(buf_obj->resv, NULL);
>>>>>>>             list_for_each_entry(attach_obj, &buf_obj->attachments,
>>>>>>> node) {
>>>>>>>                 seq_printf(s, "\t%s\n", dev_name(attach_obj->dev));
>>>>>>>                 attach_count++;
>>>>>>>             }
>>>>>>> +        dma_resv_unlock(buf_obj->resv);
>>>>>>>               seq_printf(s, "Total %d devices attached\n\n",
>>>>>>>                     attach_count);
>>>>>>> diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h
>>>>>>> index ec212cb27fdc..a8f8b2b812fd 100644
>>>>>>> --- a/include/linux/dma-buf.h
>>>>>>> +++ b/include/linux/dma-buf.h
>>>>>>> @@ -42,6 +42,17 @@ struct dma_buf_ops {
>>>>>>>           */
>>>>>>>         bool cache_sgt_mapping;
>>>>>>>     +    /**
>>>>>>> +     * @dynamic_mapping:
>>>>>>> +     *
>>>>>>> +     * If true the framework makes sure that the map/unmap_dma_buf
>>>>>>> +     * callbacks are always called with the dma_resv object locked.
>>>>>>> +     *
>>>>>>> +     * If false the framework makes ure that the map/unmap_dma_buf
>>>>>>> +     * callbacks are always called without the dma_resv object locked.
>>>>>>> +     */
>>>>>>> +    bool dynamic_mapping;
>>>>>>> +
>>>>>>>         /**
>>>>>>>          * @attach:
>>>>>>>          *
>>>>>>> @@ -109,6 +120,9 @@ struct dma_buf_ops {
>>>>>>>          * any other kind of sharing that the exporter might wish to make
>>>>>>>          * available to buffer-users.
>>>>>>>          *
>>>>>>> +     * This is always called with the dmabuf->resv object locked when
>>>>>>> +     * the dynamic_mapping flag is true.
>>>>>>> +     *
>>>>>>>          * Returns:
>>>>>>>          *
>>>>>>>          * A &sg_table scatter list of or the backing storage of
>>>>>>> the DMA buffer,
>>>>>>> @@ -327,6 +341,8 @@ struct dma_buf {
>>>>>>>      * @sgt: cached mapping.
>>>>>>>      * @dir: direction of cached mapping.
>>>>>>>      * @priv: exporter specific attachment data.
>>>>>>> + * @dynamic_mapping: true if dma_buf_map/unmap_attachment() is
>>>>>>> called with the
>>>>>>> + * dma_resv lock held.
>>>>>>>      *
>>>>>>>      * This structure holds the attachment information between the
>>>>>>> dma_buf buffer
>>>>>>>      * and its user device(s). The list contains one attachment
>>>>>>> struct per device
>>>>>>> @@ -343,6 +359,7 @@ struct dma_buf_attachment {
>>>>>>>         struct list_head node;
>>>>>>>         struct sg_table *sgt;
>>>>>>>         enum dma_data_direction dir;
>>>>>>> +    bool dynamic_mapping;
>>>>>>>         void *priv;
>>>>>>>     };
>>>>>>>     @@ -394,10 +411,39 @@ static inline void get_dma_buf(struct
>>>>>>> dma_buf *dmabuf)
>>>>>>>         get_file(dmabuf->file);
>>>>>>>     }
>>>>>>>     +/**
>>>>>>> + * dma_buf_is_dynamic - check if a DMA-buf uses dynamic mappings.
>>>>>>> + * @dmabuf: the DMA-buf to check
>>>>>>> + *
>>>>>>> + * Returns true if a DMA-buf exporter wants to be called with
>>>>>>> the dma_resv
>>>>>>> + * locked, false if it doesn't wants to be called with the lock held.
>>>>>>> + */
>>>>>>> +static inline bool dma_buf_is_dynamic(struct dma_buf *dmabuf)
>>>>>>> +{
>>>>>>> +    return dmabuf->ops->dynamic_mapping;
>>>>>>> +}
>>>>>>> +
>>>>>>> +/**
>>>>>>> + * dma_buf_attachment_is_dynamic - check if a DMA-buf
>>>>>>> attachment uses dynamic
>>>>>>> + * mappinsg
>>>>>>> + * @attach: the DMA-buf attachment to check
>>>>>>> + *
>>>>>>> + * Returns true if a DMA-buf importer wants to call the
>>>>>>> map/unmap functions with
>>>>>>> + * the dma_resv lock held.
>>>>>>> + */
>>>>>>> +static inline bool
>>>>>>> +dma_buf_attachment_is_dynamic(struct dma_buf_attachment *attach)
>>>>>>> +{
>>>>>>> +    return attach->dynamic_mapping;
>>>>>>> +}
>>>>>>> +
>>>>>>>     struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>>>>>> -                            struct device *dev);
>>>>>>> +                      struct device *dev);
>>>>>>> +struct dma_buf_attachment *
>>>>>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
>>>>>>> +               bool dynamic_mapping);
>>>>>>>     void dma_buf_detach(struct dma_buf *dmabuf,
>>>>>>> -                struct dma_buf_attachment *dmabuf_attach);
>>>>>>> +            struct dma_buf_attachment *attach);
>>>>>>>       struct dma_buf *dma_buf_export(const struct
>>>>>>> dma_buf_export_info *exp_info);
>>>>>>>     @@ -409,6 +455,7 @@ struct sg_table
>>>>>>> *dma_buf_map_attachment(struct dma_buf_attachment *,
>>>>>>>                         enum dma_data_direction);
>>>>>>>     void dma_buf_unmap_attachment(struct dma_buf_attachment *,
>>>>>>> struct sg_table *,
>>>>>>>                     enum dma_data_direction);
>>>>>>> +void dma_buf_move_notify(struct dma_buf *dma_buf);
>>>>>>>     int dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
>>>>>>>                      enum dma_data_direction dir);
>>>>>>>     int dma_buf_end_cpu_access(struct dma_buf *dma_buf,
>>>>>>> -- 
>>>>>>> 2.17.1
>>>>>>>

Daniel Vetter Sept. 17, 2019, 1:45 p.m. UTC | #8

On Tue, Sep 17, 2019 at 01:24:10PM +0000, Koenig, Christian wrote:
> Am 17.09.19 um 15:13 schrieb Daniel Vetter:
> > On Tue, Sep 17, 2019 at 12:40:51PM +0000, Koenig, Christian wrote:
> >> Am 17.09.19 um 14:31 schrieb Daniel Vetter:
> >>> On Mon, Sep 16, 2019 at 02:23:13PM +0200, Christian König wrote:
> >>>> Ping? Any further comment on this or can't we merge at least the locking
> >>>> change?
> >>> I was at plumbers ...
> >>>> Christian.
> >>>>
> >>>> Am 11.09.19 um 12:53 schrieb Christian König:
> >>>>> Am 03.09.19 um 10:05 schrieb Daniel Vetter:
> >>>>>> On Thu, Aug 29, 2019 at 04:29:14PM +0200, Christian König wrote:
> >>>>>>> This patch is a stripped down version of the locking changes
> >>>>>>> necessary to support dynamic DMA-buf handling.
> >>>>>>>
> >>>>>>> For compatibility we cache the DMA-buf mapping as soon as
> >>>>>>> exporter/importer disagree on the dynamic handling.
> >>>>>>>
> >>>>>>> Signed-off-by: Christian König <christian.koenig@amd.com>
> >>>>>>> ---
> >>>>>>>     drivers/dma-buf/dma-buf.c | 90
> >>>>>>> ++++++++++++++++++++++++++++++++++++---
> >>>>>>>     include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
> >>>>>>>     2 files changed, 133 insertions(+), 8 deletions(-)
> >>>>>>>
> >>>>>>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> >>>>>>> index 433d91d710e4..65052d52602b 100644
> >>>>>>> --- a/drivers/dma-buf/dma-buf.c
> >>>>>>> +++ b/drivers/dma-buf/dma-buf.c
> >>>>>>> @@ -525,6 +525,10 @@ struct dma_buf *dma_buf_export(const struct
> >>>>>>> dma_buf_export_info *exp_info)
> >>>>>>>             return ERR_PTR(-EINVAL);
> >>>>>>>         }
> >>>>>>>     +    if (WARN_ON(exp_info->ops->cache_sgt_mapping &&
> >>>>>>> +            exp_info->ops->dynamic_mapping))
> >>>>>>> +        return ERR_PTR(-EINVAL);
> >>>>>>> +
> >>>>>>>         if (!try_module_get(exp_info->owner))
> >>>>>>>             return ERR_PTR(-ENOENT);
> >>>>>>>     @@ -645,10 +649,11 @@ void dma_buf_put(struct dma_buf *dmabuf)
> >>>>>>>     EXPORT_SYMBOL_GPL(dma_buf_put);
> >>>>>>>       /**
> >>>>>>> - * dma_buf_attach - Add the device to dma_buf's attachments
> >>>>>>> list; optionally,
> >>>>>>> + * dma_buf_dynamic_attach - Add the device to dma_buf's
> >>>>>>> attachments list; optionally,
> >>>>>>>      * calls attach() of dma_buf_ops to allow device-specific
> >>>>>>> attach functionality
> >>>>>>> - * @dmabuf:    [in]    buffer to attach device to.
> >>>>>>> - * @dev:    [in]    device to be attached.
> >>>>>>> + * @dmabuf:        [in]    buffer to attach device to.
> >>>>>>> + * @dev:        [in]    device to be attached.
> >>>>>>> + * @dynamic_mapping:    [in]    calling convention for map/unmap
> >>>>>>>      *
> >>>>>>>      * Returns struct dma_buf_attachment pointer for this
> >>>>>>> attachment. Attachments
> >>>>>>>      * must be cleaned up by calling dma_buf_detach().
> >>>>>>> @@ -662,8 +667,9 @@ EXPORT_SYMBOL_GPL(dma_buf_put);
> >>>>>>>      * accessible to @dev, and cannot be moved to a more suitable
> >>>>>>> place. This is
> >>>>>>>      * indicated with the error code -EBUSY.
> >>>>>>>      */
> >>>>>>> -struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>>>> -                      struct device *dev)
> >>>>>>> +struct dma_buf_attachment *
> >>>>>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
> >>>>>>> +               bool dynamic_mapping)
> >>>>>>>     {
> >>>>>>>         struct dma_buf_attachment *attach;
> >>>>>>>         int ret;
> >>>>>>> @@ -677,6 +683,7 @@ struct dma_buf_attachment
> >>>>>>> *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>>>>           attach->dev = dev;
> >>>>>>>         attach->dmabuf = dmabuf;
> >>>>>>> +    attach->dynamic_mapping = dynamic_mapping;
> >>>>>>>           mutex_lock(&dmabuf->lock);
> >>>>>>>     @@ -685,16 +692,64 @@ struct dma_buf_attachment
> >>>>>>> *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>>>>             if (ret)
> >>>>>>>                 goto err_attach;
> >>>>>>>         }
> >>>>>>> +    dma_resv_lock(dmabuf->resv, NULL);
> >>>>>>>         list_add(&attach->node, &dmabuf->attachments);
> >>>>>>> +    dma_resv_unlock(dmabuf->resv);
> >>>>>>>           mutex_unlock(&dmabuf->lock);
> >>>>>>>     +    /* When either the importer or the exporter can't handle dynamic
> >>>>>>> +     * mappings we cache the mapping here to avoid issues with the
> >>>>>>> +     * reservation object lock.
> >>>>>>> +     */
> >>>>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
> >>>>>>> +        dma_buf_is_dynamic(dmabuf)) {
> >>>>>>> +        struct sg_table *sgt;
> >>>>>>> +
> >>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
> >>>>>>> +
> >>>>>>> +        sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);
> >>>>>> Now we're back to enforcing DMA_BIDI, which works nicely around the
> >>>>>> locking pain, but apparently upsets the arm-soc folks who want to
> >>>>>> control
> >>>>>> this better.
> >>>>> Take another look at dma_buf_map_attachment(), we still try to get the
> >>>>> caching there for ARM.
> >>>>>
> >>>>> What we do here is to bidirectionally map the buffer to avoid the
> >>>>> locking hydra when importer and exporter disagree on locking.
> >>>>>
> >>>>> So the ARM folks can easily avoid that by switching to dynamic locking
> >>>>> for both.
> >>> So you still break the contract between importer and exporter, except not
> >>> for anything that's run in intel-gfx-ci so all is good?
> >> No, the contract between importer and exporter stays exactly the same it
> >> is currently as long as you don't switch to dynamic dma-buf handling.
> >>
> >> There is no functional change for the ARM folks here. The only change
> >> which takes effect is between i915 and amdgpu and that is perfectly
> >> covered by intel-gfx-ci.
> > There's people who want to run amdgpu on ARM?
> 
> Sure there are, we even recently fixed some bugs for this.
> 
> But as far as I know there is no one currently which is affect by this 
> change on ARM with amdgpu.

But don't you break them with this now?

amdgpu will soon set the dynamic flag on exports, which forces the caching
at create time (to avoid the locking fun), which will then result in a
EBUSY at map_attachment time because we have a cached mapping, but it's
the wrong type.

> > Also, x86 doesn't have cache
> > flushing in the dma-api, so naturally this isn't any issue for us (we
> > still have cache flushing in actual hw, but that's a different topic). So
> > "works on x86" isn't really a great way to justify what we do here I
> > think.
> 
> Well it is the exact same caching we previously had as well, so there is 
> absolutely no functional change here except that we now explicitly note 
> that amdgpu always needs bidirectional mappings.
> 
> I agree that we should get rid of this caching as soon as possible, but 
> we should not fix things which where broken before.
> 
> On the other hand adding dma_sg_sync_for_cpu/device sounds like 
> something we could easily add separately to the caching if you think 
> that this will help.

The current code maybe lacks some cache flushes, but we already require a
fixed direction per attachment. So I guess not a real problem, probably.

But with your patches I think we now fail with EBUSY. Not exactly nice ...
-Daniel

> 
> Christian.
> 
> > -Daniel
> >
> >> Regards,
> >> Christian.
> >>
> >>> The other issue with "we solve this with caching the mapping": Currently
> >>> map/unmap flush (at least on arm, at least on cases where it matters). If
> >>> you just return the cached sg, then we don't do the flushing anymore,
> >>> which might break importers/exporters in exactly the same way as just
> >>> giving them the wrong mapping. There's zero differences between a BIDI,
> >>> TO_CPU, or TO_DEVICE mapping, the only places where this matters is for
> >>> cache flushing.
> >>>
> >>> So here's something that could actually work:
> >>> - We cache the mapping.
> >>> - We cache a bidirectional mapping.
> >>> - We put the right dma_sg_sync_for_cpu/device calls in place for map/unmap
> >>>     to give current importers/exporters the same behaviour they're used to
> >>>     now.
> >>>
> >>> And yes the caching we've lifted might have broken something somewhere
> >>> already. But generally you only hear about that long time after because
> >>> arm vendors roll forward once every few years. Or something like that.
> >>> -Daniel
> >>>
> >>>>> Regards,
> >>>>> Christian.
> >>>>>
> >>>>>> That's why your previous version moved the caching into
> >>>>>> map/unmap_sg, which resurrected the locking hydra.
> >>>>>>
> >>>>>> I think we're going a bit in circles here, and I don't have a good idea
> >>>>>> either :-/
> >>>>>> -Daniel
> >>>>>>
> >>>>>>> +        if (!sgt)
> >>>>>>> +            sgt = ERR_PTR(-ENOMEM);
> >>>>>>> +        if (IS_ERR(sgt)) {
> >>>>>>> +            ret = PTR_ERR(sgt);
> >>>>>>> +            goto err_unlock;
> >>>>>>> +        }
> >>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>>>> +            dma_resv_unlock(attach->dmabuf->resv);
> >>>>>>> +        attach->sgt = sgt;
> >>>>>>> +        attach->dir = DMA_BIDIRECTIONAL;
> >>>>>>> +    }
> >>>>>>> +
> >>>>>>>         return attach;
> >>>>>>>       err_attach:
> >>>>>>>         kfree(attach);
> >>>>>>>         mutex_unlock(&dmabuf->lock);
> >>>>>>>         return ERR_PTR(ret);
> >>>>>>> +
> >>>>>>> +err_unlock:
> >>>>>>> +    if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>>>> +        dma_resv_unlock(attach->dmabuf->resv);
> >>>>>>> +
> >>>>>>> +    dma_buf_detach(dmabuf, attach);
> >>>>>>> +    return ERR_PTR(ret);
> >>>>>>> +}
> >>>>>>> +EXPORT_SYMBOL_GPL(dma_buf_dynamic_attach);
> >>>>>>> +
> >>>>>>> +/**
> >>>>>>> + * dma_buf_attach - Wrapper for dma_buf_dynamic_attach
> >>>>>>> + * @dmabuf:    [in]    buffer to attach device to.
> >>>>>>> + * @dev:    [in]    device to be attached.
> >>>>>>> + *
> >>>>>>> + * Wrapper to call dma_buf_dynamic_attach() for drivers which
> >>>>>>> still use a static
> >>>>>>> + * mapping.
> >>>>>>> + */
> >>>>>>> +struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>>>> +                      struct device *dev)
> >>>>>>> +{
> >>>>>>> +    return dma_buf_dynamic_attach(dmabuf, dev, false);
> >>>>>>>     }
> >>>>>>>     EXPORT_SYMBOL_GPL(dma_buf_attach);
> >>>>>>>     @@ -711,11 +766,20 @@ void dma_buf_detach(struct dma_buf
> >>>>>>> *dmabuf, struct dma_buf_attachment *attach)
> >>>>>>>         if (WARN_ON(!dmabuf || !attach))
> >>>>>>>             return;
> >>>>>>>     -    if (attach->sgt)
> >>>>>>> +    if (attach->sgt) {
> >>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
> >>>>>>> +
> >>>>>>>             dmabuf->ops->unmap_dma_buf(attach, attach->sgt, attach->dir);
> >>>>>>>     +        if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>>>> +            dma_resv_unlock(attach->dmabuf->resv);
> >>>>>>> +    }
> >>>>>>> +
> >>>>>>>         mutex_lock(&dmabuf->lock);
> >>>>>>> +    dma_resv_lock(dmabuf->resv, NULL);
> >>>>>>>         list_del(&attach->node);
> >>>>>>> +    dma_resv_unlock(dmabuf->resv);
> >>>>>>>         if (dmabuf->ops->detach)
> >>>>>>>             dmabuf->ops->detach(dmabuf, attach);
> >>>>>>>     @@ -749,6 +813,9 @@ struct sg_table
> >>>>>>> *dma_buf_map_attachment(struct dma_buf_attachment *attach,
> >>>>>>>         if (WARN_ON(!attach || !attach->dmabuf))
> >>>>>>>             return ERR_PTR(-EINVAL);
> >>>>>>>     +    if (dma_buf_attachment_is_dynamic(attach))
> >>>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
> >>>>>>> +
> >>>>>>>         if (attach->sgt) {
> >>>>>>>             /*
> >>>>>>>              * Two mappings with different directions for the same
> >>>>>>> @@ -761,6 +828,9 @@ struct sg_table
> >>>>>>> *dma_buf_map_attachment(struct dma_buf_attachment *attach,
> >>>>>>>             return attach->sgt;
> >>>>>>>         }
> >>>>>>>     +    if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
> >>>>>>> +
> >>>>>>>         sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction);
> >>>>>>>         if (!sg_table)
> >>>>>>>             sg_table = ERR_PTR(-ENOMEM);
> >>>>>>> @@ -793,9 +863,15 @@ void dma_buf_unmap_attachment(struct
> >>>>>>> dma_buf_attachment *attach,
> >>>>>>>         if (WARN_ON(!attach || !attach->dmabuf || !sg_table))
> >>>>>>>             return;
> >>>>>>>     +    if (dma_buf_attachment_is_dynamic(attach))
> >>>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
> >>>>>>> +
> >>>>>>>         if (attach->sgt == sg_table)
> >>>>>>>             return;
> >>>>>>>     +    if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>>>> +        dma_resv_assert_held(attach->dmabuf->resv);
> >>>>>>> +
> >>>>>>>         attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, direction);
> >>>>>>>     }
> >>>>>>>     EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment);
> >>>>>>> @@ -1219,10 +1295,12 @@ static int dma_buf_debug_show(struct
> >>>>>>> seq_file *s, void *unused)
> >>>>>>>             seq_puts(s, "\tAttached Devices:\n");
> >>>>>>>             attach_count = 0;
> >>>>>>>     +        dma_resv_lock(buf_obj->resv, NULL);
> >>>>>>>             list_for_each_entry(attach_obj, &buf_obj->attachments,
> >>>>>>> node) {
> >>>>>>>                 seq_printf(s, "\t%s\n", dev_name(attach_obj->dev));
> >>>>>>>                 attach_count++;
> >>>>>>>             }
> >>>>>>> +        dma_resv_unlock(buf_obj->resv);
> >>>>>>>               seq_printf(s, "Total %d devices attached\n\n",
> >>>>>>>                     attach_count);
> >>>>>>> diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h
> >>>>>>> index ec212cb27fdc..a8f8b2b812fd 100644
> >>>>>>> --- a/include/linux/dma-buf.h
> >>>>>>> +++ b/include/linux/dma-buf.h
> >>>>>>> @@ -42,6 +42,17 @@ struct dma_buf_ops {
> >>>>>>>           */
> >>>>>>>         bool cache_sgt_mapping;
> >>>>>>>     +    /**
> >>>>>>> +     * @dynamic_mapping:
> >>>>>>> +     *
> >>>>>>> +     * If true the framework makes sure that the map/unmap_dma_buf
> >>>>>>> +     * callbacks are always called with the dma_resv object locked.
> >>>>>>> +     *
> >>>>>>> +     * If false the framework makes ure that the map/unmap_dma_buf
> >>>>>>> +     * callbacks are always called without the dma_resv object locked.
> >>>>>>> +     */
> >>>>>>> +    bool dynamic_mapping;
> >>>>>>> +
> >>>>>>>         /**
> >>>>>>>          * @attach:
> >>>>>>>          *
> >>>>>>> @@ -109,6 +120,9 @@ struct dma_buf_ops {
> >>>>>>>          * any other kind of sharing that the exporter might wish to make
> >>>>>>>          * available to buffer-users.
> >>>>>>>          *
> >>>>>>> +     * This is always called with the dmabuf->resv object locked when
> >>>>>>> +     * the dynamic_mapping flag is true.
> >>>>>>> +     *
> >>>>>>>          * Returns:
> >>>>>>>          *
> >>>>>>>          * A &sg_table scatter list of or the backing storage of
> >>>>>>> the DMA buffer,
> >>>>>>> @@ -327,6 +341,8 @@ struct dma_buf {
> >>>>>>>      * @sgt: cached mapping.
> >>>>>>>      * @dir: direction of cached mapping.
> >>>>>>>      * @priv: exporter specific attachment data.
> >>>>>>> + * @dynamic_mapping: true if dma_buf_map/unmap_attachment() is
> >>>>>>> called with the
> >>>>>>> + * dma_resv lock held.
> >>>>>>>      *
> >>>>>>>      * This structure holds the attachment information between the
> >>>>>>> dma_buf buffer
> >>>>>>>      * and its user device(s). The list contains one attachment
> >>>>>>> struct per device
> >>>>>>> @@ -343,6 +359,7 @@ struct dma_buf_attachment {
> >>>>>>>         struct list_head node;
> >>>>>>>         struct sg_table *sgt;
> >>>>>>>         enum dma_data_direction dir;
> >>>>>>> +    bool dynamic_mapping;
> >>>>>>>         void *priv;
> >>>>>>>     };
> >>>>>>>     @@ -394,10 +411,39 @@ static inline void get_dma_buf(struct
> >>>>>>> dma_buf *dmabuf)
> >>>>>>>         get_file(dmabuf->file);
> >>>>>>>     }
> >>>>>>>     +/**
> >>>>>>> + * dma_buf_is_dynamic - check if a DMA-buf uses dynamic mappings.
> >>>>>>> + * @dmabuf: the DMA-buf to check
> >>>>>>> + *
> >>>>>>> + * Returns true if a DMA-buf exporter wants to be called with
> >>>>>>> the dma_resv
> >>>>>>> + * locked, false if it doesn't wants to be called with the lock held.
> >>>>>>> + */
> >>>>>>> +static inline bool dma_buf_is_dynamic(struct dma_buf *dmabuf)
> >>>>>>> +{
> >>>>>>> +    return dmabuf->ops->dynamic_mapping;
> >>>>>>> +}
> >>>>>>> +
> >>>>>>> +/**
> >>>>>>> + * dma_buf_attachment_is_dynamic - check if a DMA-buf
> >>>>>>> attachment uses dynamic
> >>>>>>> + * mappinsg
> >>>>>>> + * @attach: the DMA-buf attachment to check
> >>>>>>> + *
> >>>>>>> + * Returns true if a DMA-buf importer wants to call the
> >>>>>>> map/unmap functions with
> >>>>>>> + * the dma_resv lock held.
> >>>>>>> + */
> >>>>>>> +static inline bool
> >>>>>>> +dma_buf_attachment_is_dynamic(struct dma_buf_attachment *attach)
> >>>>>>> +{
> >>>>>>> +    return attach->dynamic_mapping;
> >>>>>>> +}
> >>>>>>> +
> >>>>>>>     struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>>>> -                            struct device *dev);
> >>>>>>> +                      struct device *dev);
> >>>>>>> +struct dma_buf_attachment *
> >>>>>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
> >>>>>>> +               bool dynamic_mapping);
> >>>>>>>     void dma_buf_detach(struct dma_buf *dmabuf,
> >>>>>>> -                struct dma_buf_attachment *dmabuf_attach);
> >>>>>>> +            struct dma_buf_attachment *attach);
> >>>>>>>       struct dma_buf *dma_buf_export(const struct
> >>>>>>> dma_buf_export_info *exp_info);
> >>>>>>>     @@ -409,6 +455,7 @@ struct sg_table
> >>>>>>> *dma_buf_map_attachment(struct dma_buf_attachment *,
> >>>>>>>                         enum dma_data_direction);
> >>>>>>>     void dma_buf_unmap_attachment(struct dma_buf_attachment *,
> >>>>>>> struct sg_table *,
> >>>>>>>                     enum dma_data_direction);
> >>>>>>> +void dma_buf_move_notify(struct dma_buf *dma_buf);
> >>>>>>>     int dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
> >>>>>>>                      enum dma_data_direction dir);
> >>>>>>>     int dma_buf_end_cpu_access(struct dma_buf *dma_buf,
> >>>>>>> -- 
> >>>>>>> 2.17.1
> >>>>>>>
>

Christian König Sept. 17, 2019, 2:47 p.m. UTC | #9

Am 17.09.19 um 15:45 schrieb Daniel Vetter:
> On Tue, Sep 17, 2019 at 01:24:10PM +0000, Koenig, Christian wrote:
>> Am 17.09.19 um 15:13 schrieb Daniel Vetter:
>>> On Tue, Sep 17, 2019 at 12:40:51PM +0000, Koenig, Christian wrote:
>>>> Am 17.09.19 um 14:31 schrieb Daniel Vetter:
>>>>> On Mon, Sep 16, 2019 at 02:23:13PM +0200, Christian König wrote:
>>>>>> Ping? Any further comment on this or can't we merge at least the locking
>>>>>> change?
>>>>> I was at plumbers ...
>>>>>> Christian.
>>>>>>
>>>>>> Am 11.09.19 um 12:53 schrieb Christian König:
>>>>>>> Am 03.09.19 um 10:05 schrieb Daniel Vetter:
>>>>>>>> On Thu, Aug 29, 2019 at 04:29:14PM +0200, Christian König wrote:
>>>>>>>>> This patch is a stripped down version of the locking changes
>>>>>>>>> necessary to support dynamic DMA-buf handling.
>>>>>>>>>
>>>>>>>>> For compatibility we cache the DMA-buf mapping as soon as
>>>>>>>>> exporter/importer disagree on the dynamic handling.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Christian König <christian.koenig@amd.com>
>>>>>>>>> ---
>>>>>>>>>      drivers/dma-buf/dma-buf.c | 90
>>>>>>>>> ++++++++++++++++++++++++++++++++++++---
>>>>>>>>>      include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
>>>>>>>>>      2 files changed, 133 insertions(+), 8 deletions(-)
>>>>>>>>>
>>>>>>>>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
>>>>>>>>> index 433d91d710e4..65052d52602b 100644
>>>>>>>>> --- a/drivers/dma-buf/dma-buf.c
>>>>>>>>> +++ b/drivers/dma-buf/dma-buf.c
>>>>>>>>> @@ -525,6 +525,10 @@ struct dma_buf *dma_buf_export(const struct
>>>>>>>>> dma_buf_export_info *exp_info)
>>>>>>>>>              return ERR_PTR(-EINVAL);
>>>>>>>>>          }
>>>>>>>>>      +    if (WARN_ON(exp_info->ops->cache_sgt_mapping &&
>>>>>>>>> +            exp_info->ops->dynamic_mapping))
>>>>>>>>> +        return ERR_PTR(-EINVAL);
>>>>>>>>> +
>>>>>>>>>          if (!try_module_get(exp_info->owner))
>>>>>>>>>              return ERR_PTR(-ENOENT);
>>>>>>>>>      @@ -645,10 +649,11 @@ void dma_buf_put(struct dma_buf *dmabuf)
>>>>>>>>>      EXPORT_SYMBOL_GPL(dma_buf_put);
>>>>>>>>>        /**
>>>>>>>>> - * dma_buf_attach - Add the device to dma_buf's attachments
>>>>>>>>> list; optionally,
>>>>>>>>> + * dma_buf_dynamic_attach - Add the device to dma_buf's
>>>>>>>>> attachments list; optionally,
>>>>>>>>>       * calls attach() of dma_buf_ops to allow device-specific
>>>>>>>>> attach functionality
>>>>>>>>> - * @dmabuf:    [in]    buffer to attach device to.
>>>>>>>>> - * @dev:    [in]    device to be attached.
>>>>>>>>> + * @dmabuf:        [in]    buffer to attach device to.
>>>>>>>>> + * @dev:        [in]    device to be attached.
>>>>>>>>> + * @dynamic_mapping:    [in]    calling convention for map/unmap
>>>>>>>>>       *
>>>>>>>>>       * Returns struct dma_buf_attachment pointer for this
>>>>>>>>> attachment. Attachments
>>>>>>>>>       * must be cleaned up by calling dma_buf_detach().
>>>>>>>>> @@ -662,8 +667,9 @@ EXPORT_SYMBOL_GPL(dma_buf_put);
>>>>>>>>>       * accessible to @dev, and cannot be moved to a more suitable
>>>>>>>>> place. This is
>>>>>>>>>       * indicated with the error code -EBUSY.
>>>>>>>>>       */
>>>>>>>>> -struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>>>>>>>>> -                      struct device *dev)
>>>>>>>>> +struct dma_buf_attachment *
>>>>>>>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
>>>>>>>>> +               bool dynamic_mapping)
>>>>>>>>>      {
>>>>>>>>>          struct dma_buf_attachment *attach;
>>>>>>>>>          int ret;
>>>>>>>>> @@ -677,6 +683,7 @@ struct dma_buf_attachment
>>>>>>>>> *dma_buf_attach(struct dma_buf *dmabuf,
>>>>>>>>>            attach->dev = dev;
>>>>>>>>>          attach->dmabuf = dmabuf;
>>>>>>>>> +    attach->dynamic_mapping = dynamic_mapping;
>>>>>>>>>            mutex_lock(&dmabuf->lock);
>>>>>>>>>      @@ -685,16 +692,64 @@ struct dma_buf_attachment
>>>>>>>>> *dma_buf_attach(struct dma_buf *dmabuf,
>>>>>>>>>              if (ret)
>>>>>>>>>                  goto err_attach;
>>>>>>>>>          }
>>>>>>>>> +    dma_resv_lock(dmabuf->resv, NULL);
>>>>>>>>>          list_add(&attach->node, &dmabuf->attachments);
>>>>>>>>> +    dma_resv_unlock(dmabuf->resv);
>>>>>>>>>            mutex_unlock(&dmabuf->lock);
>>>>>>>>>      +    /* When either the importer or the exporter can't handle dynamic
>>>>>>>>> +     * mappings we cache the mapping here to avoid issues with the
>>>>>>>>> +     * reservation object lock.
>>>>>>>>> +     */
>>>>>>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
>>>>>>>>> +        dma_buf_is_dynamic(dmabuf)) {
>>>>>>>>> +        struct sg_table *sgt;
>>>>>>>>> +
>>>>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
>>>>>>>>> +
>>>>>>>>> +        sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);
>>>>>>>> Now we're back to enforcing DMA_BIDI, which works nicely around the
>>>>>>>> locking pain, but apparently upsets the arm-soc folks who want to
>>>>>>>> control
>>>>>>>> this better.
>>>>>>> Take another look at dma_buf_map_attachment(), we still try to get the
>>>>>>> caching there for ARM.
>>>>>>>
>>>>>>> What we do here is to bidirectionally map the buffer to avoid the
>>>>>>> locking hydra when importer and exporter disagree on locking.
>>>>>>>
>>>>>>> So the ARM folks can easily avoid that by switching to dynamic locking
>>>>>>> for both.
>>>>> So you still break the contract between importer and exporter, except not
>>>>> for anything that's run in intel-gfx-ci so all is good?
>>>> No, the contract between importer and exporter stays exactly the same it
>>>> is currently as long as you don't switch to dynamic dma-buf handling.
>>>>
>>>> There is no functional change for the ARM folks here. The only change
>>>> which takes effect is between i915 and amdgpu and that is perfectly
>>>> covered by intel-gfx-ci.
>>> There's people who want to run amdgpu on ARM?
>> Sure there are, we even recently fixed some bugs for this.
>>
>> But as far as I know there is no one currently which is affect by this
>> change on ARM with amdgpu.
> But don't you break them with this now?

No, we see the bidirectional attachment as compatible with the other ones.

> amdgpu will soon set the dynamic flag on exports, which forces the caching
> at create time (to avoid the locking fun), which will then result in a
> EBUSY at map_attachment time because we have a cached mapping, but it's
> the wrong type.

See the check in dma_buf_map_attachment():

     if (attach->dir != direction && attach->dir != DMA_BIDIRECTIONAL)
         return ERR_PTR(-EBUSY);


Regards,
Christian.

>
>>> Also, x86 doesn't have cache
>>> flushing in the dma-api, so naturally this isn't any issue for us (we
>>> still have cache flushing in actual hw, but that's a different topic). So
>>> "works on x86" isn't really a great way to justify what we do here I
>>> think.
>> Well it is the exact same caching we previously had as well, so there is
>> absolutely no functional change here except that we now explicitly note
>> that amdgpu always needs bidirectional mappings.
>>
>> I agree that we should get rid of this caching as soon as possible, but
>> we should not fix things which where broken before.
>>
>> On the other hand adding dma_sg_sync_for_cpu/device sounds like
>> something we could easily add separately to the caching if you think
>> that this will help.
> The current code maybe lacks some cache flushes, but we already require a
> fixed direction per attachment. So I guess not a real problem, probably.
>
> But with your patches I think we now fail with EBUSY. Not exactly nice ...
> -Daniel
>
>> Christian.
>>
>>> -Daniel
>>>
>>>> Regards,
>>>> Christian.
>>>>
>>>>> The other issue with "we solve this with caching the mapping": Currently
>>>>> map/unmap flush (at least on arm, at least on cases where it matters). If
>>>>> you just return the cached sg, then we don't do the flushing anymore,
>>>>> which might break importers/exporters in exactly the same way as just
>>>>> giving them the wrong mapping. There's zero differences between a BIDI,
>>>>> TO_CPU, or TO_DEVICE mapping, the only places where this matters is for
>>>>> cache flushing.
>>>>>
>>>>> So here's something that could actually work:
>>>>> - We cache the mapping.
>>>>> - We cache a bidirectional mapping.
>>>>> - We put the right dma_sg_sync_for_cpu/device calls in place for map/unmap
>>>>>      to give current importers/exporters the same behaviour they're used to
>>>>>      now.
>>>>>
>>>>> And yes the caching we've lifted might have broken something somewhere
>>>>> already. But generally you only hear about that long time after because
>>>>> arm vendors roll forward once every few years. Or something like that.
>>>>> -Daniel
>>>>>
>>>>>>> Regards,
>>>>>>> Christian.
>>>>>>>
>>>>>>>> That's why your previous version moved the caching into
>>>>>>>> map/unmap_sg, which resurrected the locking hydra.
>>>>>>>>
>>>>>>>> I think we're going a bit in circles here, and I don't have a good idea
>>>>>>>> either :-/
>>>>>>>> -Daniel
>>>>>>>>
>>>>>>>>

Daniel Vetter Sept. 17, 2019, 2:56 p.m. UTC | #10

On Tue, Sep 17, 2019 at 4:47 PM Koenig, Christian
<Christian.Koenig@amd.com> wrote:
>
> Am 17.09.19 um 15:45 schrieb Daniel Vetter:
> > On Tue, Sep 17, 2019 at 01:24:10PM +0000, Koenig, Christian wrote:
> >> Am 17.09.19 um 15:13 schrieb Daniel Vetter:
> >>> On Tue, Sep 17, 2019 at 12:40:51PM +0000, Koenig, Christian wrote:
> >>>> Am 17.09.19 um 14:31 schrieb Daniel Vetter:
> >>>>> On Mon, Sep 16, 2019 at 02:23:13PM +0200, Christian König wrote:
> >>>>>> Ping? Any further comment on this or can't we merge at least the locking
> >>>>>> change?
> >>>>> I was at plumbers ...
> >>>>>> Christian.
> >>>>>>
> >>>>>> Am 11.09.19 um 12:53 schrieb Christian König:
> >>>>>>> Am 03.09.19 um 10:05 schrieb Daniel Vetter:
> >>>>>>>> On Thu, Aug 29, 2019 at 04:29:14PM +0200, Christian König wrote:
> >>>>>>>>> This patch is a stripped down version of the locking changes
> >>>>>>>>> necessary to support dynamic DMA-buf handling.
> >>>>>>>>>
> >>>>>>>>> For compatibility we cache the DMA-buf mapping as soon as
> >>>>>>>>> exporter/importer disagree on the dynamic handling.
> >>>>>>>>>
> >>>>>>>>> Signed-off-by: Christian König <christian.koenig@amd.com>
> >>>>>>>>> ---
> >>>>>>>>>      drivers/dma-buf/dma-buf.c | 90
> >>>>>>>>> ++++++++++++++++++++++++++++++++++++---
> >>>>>>>>>      include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
> >>>>>>>>>      2 files changed, 133 insertions(+), 8 deletions(-)
> >>>>>>>>>
> >>>>>>>>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> >>>>>>>>> index 433d91d710e4..65052d52602b 100644
> >>>>>>>>> --- a/drivers/dma-buf/dma-buf.c
> >>>>>>>>> +++ b/drivers/dma-buf/dma-buf.c
> >>>>>>>>> @@ -525,6 +525,10 @@ struct dma_buf *dma_buf_export(const struct
> >>>>>>>>> dma_buf_export_info *exp_info)
> >>>>>>>>>              return ERR_PTR(-EINVAL);
> >>>>>>>>>          }
> >>>>>>>>>      +    if (WARN_ON(exp_info->ops->cache_sgt_mapping &&
> >>>>>>>>> +            exp_info->ops->dynamic_mapping))
> >>>>>>>>> +        return ERR_PTR(-EINVAL);
> >>>>>>>>> +
> >>>>>>>>>          if (!try_module_get(exp_info->owner))
> >>>>>>>>>              return ERR_PTR(-ENOENT);
> >>>>>>>>>      @@ -645,10 +649,11 @@ void dma_buf_put(struct dma_buf *dmabuf)
> >>>>>>>>>      EXPORT_SYMBOL_GPL(dma_buf_put);
> >>>>>>>>>        /**
> >>>>>>>>> - * dma_buf_attach - Add the device to dma_buf's attachments
> >>>>>>>>> list; optionally,
> >>>>>>>>> + * dma_buf_dynamic_attach - Add the device to dma_buf's
> >>>>>>>>> attachments list; optionally,
> >>>>>>>>>       * calls attach() of dma_buf_ops to allow device-specific
> >>>>>>>>> attach functionality
> >>>>>>>>> - * @dmabuf:    [in]    buffer to attach device to.
> >>>>>>>>> - * @dev:    [in]    device to be attached.
> >>>>>>>>> + * @dmabuf:        [in]    buffer to attach device to.
> >>>>>>>>> + * @dev:        [in]    device to be attached.
> >>>>>>>>> + * @dynamic_mapping:    [in]    calling convention for map/unmap
> >>>>>>>>>       *
> >>>>>>>>>       * Returns struct dma_buf_attachment pointer for this
> >>>>>>>>> attachment. Attachments
> >>>>>>>>>       * must be cleaned up by calling dma_buf_detach().
> >>>>>>>>> @@ -662,8 +667,9 @@ EXPORT_SYMBOL_GPL(dma_buf_put);
> >>>>>>>>>       * accessible to @dev, and cannot be moved to a more suitable
> >>>>>>>>> place. This is
> >>>>>>>>>       * indicated with the error code -EBUSY.
> >>>>>>>>>       */
> >>>>>>>>> -struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>>>>>> -                      struct device *dev)
> >>>>>>>>> +struct dma_buf_attachment *
> >>>>>>>>> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
> >>>>>>>>> +               bool dynamic_mapping)
> >>>>>>>>>      {
> >>>>>>>>>          struct dma_buf_attachment *attach;
> >>>>>>>>>          int ret;
> >>>>>>>>> @@ -677,6 +683,7 @@ struct dma_buf_attachment
> >>>>>>>>> *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>>>>>>            attach->dev = dev;
> >>>>>>>>>          attach->dmabuf = dmabuf;
> >>>>>>>>> +    attach->dynamic_mapping = dynamic_mapping;
> >>>>>>>>>            mutex_lock(&dmabuf->lock);
> >>>>>>>>>      @@ -685,16 +692,64 @@ struct dma_buf_attachment
> >>>>>>>>> *dma_buf_attach(struct dma_buf *dmabuf,
> >>>>>>>>>              if (ret)
> >>>>>>>>>                  goto err_attach;
> >>>>>>>>>          }
> >>>>>>>>> +    dma_resv_lock(dmabuf->resv, NULL);
> >>>>>>>>>          list_add(&attach->node, &dmabuf->attachments);
> >>>>>>>>> +    dma_resv_unlock(dmabuf->resv);
> >>>>>>>>>            mutex_unlock(&dmabuf->lock);
> >>>>>>>>>      +    /* When either the importer or the exporter can't handle dynamic
> >>>>>>>>> +     * mappings we cache the mapping here to avoid issues with the
> >>>>>>>>> +     * reservation object lock.
> >>>>>>>>> +     */
> >>>>>>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
> >>>>>>>>> +        dma_buf_is_dynamic(dmabuf)) {
> >>>>>>>>> +        struct sg_table *sgt;
> >>>>>>>>> +
> >>>>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>>>>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
> >>>>>>>>> +
> >>>>>>>>> +        sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);
> >>>>>>>> Now we're back to enforcing DMA_BIDI, which works nicely around the
> >>>>>>>> locking pain, but apparently upsets the arm-soc folks who want to
> >>>>>>>> control
> >>>>>>>> this better.
> >>>>>>> Take another look at dma_buf_map_attachment(), we still try to get the
> >>>>>>> caching there for ARM.
> >>>>>>>
> >>>>>>> What we do here is to bidirectionally map the buffer to avoid the
> >>>>>>> locking hydra when importer and exporter disagree on locking.
> >>>>>>>
> >>>>>>> So the ARM folks can easily avoid that by switching to dynamic locking
> >>>>>>> for both.
> >>>>> So you still break the contract between importer and exporter, except not
> >>>>> for anything that's run in intel-gfx-ci so all is good?
> >>>> No, the contract between importer and exporter stays exactly the same it
> >>>> is currently as long as you don't switch to dynamic dma-buf handling.
> >>>>
> >>>> There is no functional change for the ARM folks here. The only change
> >>>> which takes effect is between i915 and amdgpu and that is perfectly
> >>>> covered by intel-gfx-ci.
> >>> There's people who want to run amdgpu on ARM?
> >> Sure there are, we even recently fixed some bugs for this.
> >>
> >> But as far as I know there is no one currently which is affect by this
> >> change on ARM with amdgpu.
> > But don't you break them with this now?
>
> No, we see the bidirectional attachment as compatible with the other ones.
>
> > amdgpu will soon set the dynamic flag on exports, which forces the caching
> > at create time (to avoid the locking fun), which will then result in a
> > EBUSY at map_attachment time because we have a cached mapping, but it's
> > the wrong type.
>
> See the check in dma_buf_map_attachment():
>
>      if (attach->dir != direction && attach->dir != DMA_BIDIRECTIONAL)
>          return ERR_PTR(-EBUSY);

Hm, I misread this. So yeah should work, +/- the issue that we might
not flush enough. But I guess that can be fixed whenever, it's not
like dma-api semantics are a great fit for us. Maybe a fixme comment
would be useful here ... I'll look at this tomorrow or so because atm
brain is slow, I'm down with the usual post-conference cold it seems
:-/
-Daniel

>
> Regards,
> Christian.
>
> >
> >>> Also, x86 doesn't have cache
> >>> flushing in the dma-api, so naturally this isn't any issue for us (we
> >>> still have cache flushing in actual hw, but that's a different topic). So
> >>> "works on x86" isn't really a great way to justify what we do here I
> >>> think.
> >> Well it is the exact same caching we previously had as well, so there is
> >> absolutely no functional change here except that we now explicitly note
> >> that amdgpu always needs bidirectional mappings.
> >>
> >> I agree that we should get rid of this caching as soon as possible, but
> >> we should not fix things which where broken before.
> >>
> >> On the other hand adding dma_sg_sync_for_cpu/device sounds like
> >> something we could easily add separately to the caching if you think
> >> that this will help.
> > The current code maybe lacks some cache flushes, but we already require a
> > fixed direction per attachment. So I guess not a real problem, probably.
> >
> > But with your patches I think we now fail with EBUSY. Not exactly nice ...
> > -Daniel
> >
> >> Christian.
> >>
> >>> -Daniel
> >>>
> >>>> Regards,
> >>>> Christian.
> >>>>
> >>>>> The other issue with "we solve this with caching the mapping": Currently
> >>>>> map/unmap flush (at least on arm, at least on cases where it matters). If
> >>>>> you just return the cached sg, then we don't do the flushing anymore,
> >>>>> which might break importers/exporters in exactly the same way as just
> >>>>> giving them the wrong mapping. There's zero differences between a BIDI,
> >>>>> TO_CPU, or TO_DEVICE mapping, the only places where this matters is for
> >>>>> cache flushing.
> >>>>>
> >>>>> So here's something that could actually work:
> >>>>> - We cache the mapping.
> >>>>> - We cache a bidirectional mapping.
> >>>>> - We put the right dma_sg_sync_for_cpu/device calls in place for map/unmap
> >>>>>      to give current importers/exporters the same behaviour they're used to
> >>>>>      now.
> >>>>>
> >>>>> And yes the caching we've lifted might have broken something somewhere
> >>>>> already. But generally you only hear about that long time after because
> >>>>> arm vendors roll forward once every few years. Or something like that.
> >>>>> -Daniel
> >>>>>
> >>>>>>> Regards,
> >>>>>>> Christian.
> >>>>>>>
> >>>>>>>> That's why your previous version moved the caching into
> >>>>>>>> map/unmap_sg, which resurrected the locking hydra.
> >>>>>>>>
> >>>>>>>> I think we're going a bit in circles here, and I don't have a good idea
> >>>>>>>> either :-/
> >>>>>>>> -Daniel
> >>>>>>>>
> >>>>>>>>
>

Christian König Sept. 24, 2019, 9:51 a.m. UTC | #11

Am 17.09.19 um 16:56 schrieb Daniel Vetter:
> [SNIP]
>>>>>>>>>>>       +    /* When either the importer or the exporter can't handle dynamic
>>>>>>>>>>> +     * mappings we cache the mapping here to avoid issues with the
>>>>>>>>>>> +     * reservation object lock.
>>>>>>>>>>> +     */
>>>>>>>>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
>>>>>>>>>>> +        dma_buf_is_dynamic(dmabuf)) {
>>>>>>>>>>> +        struct sg_table *sgt;
>>>>>>>>>>> +
>>>>>>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>>>>>> +            dma_resv_lock(attach->dmabuf->resv, NULL);
>>>>>>>>>>> +
>>>>>>>>>>> +        sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);
>>>>>>>>>> Now we're back to enforcing DMA_BIDI, which works nicely around the
>>>>>>>>>> locking pain, but apparently upsets the arm-soc folks who want to
>>>>>>>>>> control
>>>>>>>>>> this better.
>>>>>>>>> Take another look at dma_buf_map_attachment(), we still try to get the
>>>>>>>>> caching there for ARM.
>>>>>>>>>
>>>>>>>>> What we do here is to bidirectionally map the buffer to avoid the
>>>>>>>>> locking hydra when importer and exporter disagree on locking.
>>>>>>>>>
>>>>>>>>> So the ARM folks can easily avoid that by switching to dynamic locking
>>>>>>>>> for both.
>>>>>>> So you still break the contract between importer and exporter, except not
>>>>>>> for anything that's run in intel-gfx-ci so all is good?
>>>>>> No, the contract between importer and exporter stays exactly the same it
>>>>>> is currently as long as you don't switch to dynamic dma-buf handling.
>>>>>>
>>>>>> There is no functional change for the ARM folks here. The only change
>>>>>> which takes effect is between i915 and amdgpu and that is perfectly
>>>>>> covered by intel-gfx-ci.
>>>>> There's people who want to run amdgpu on ARM?
>>>> Sure there are, we even recently fixed some bugs for this.
>>>>
>>>> But as far as I know there is no one currently which is affect by this
>>>> change on ARM with amdgpu.
>>> But don't you break them with this now?
>> No, we see the bidirectional attachment as compatible with the other ones.
>>
>>> amdgpu will soon set the dynamic flag on exports, which forces the caching
>>> at create time (to avoid the locking fun), which will then result in a
>>> EBUSY at map_attachment time because we have a cached mapping, but it's
>>> the wrong type.
>> See the check in dma_buf_map_attachment():
>>
>>       if (attach->dir != direction && attach->dir != DMA_BIDIRECTIONAL)
>>           return ERR_PTR(-EBUSY);
> Hm, I misread this. So yeah should work, +/- the issue that we might
> not flush enough. But I guess that can be fixed whenever, it's not
> like dma-api semantics are a great fit for us. Maybe a fixme comment
> would be useful here ... I'll look at this tomorrow or so because atm
> brain is slow, I'm down with the usual post-conference cold it seems
> :-/

Hope your are feeling better now, adding a comment is of course not a 
problem.

With that fixed can I get an reviewed-by or at least and acked-by?

I want to land at least some parts of those changes now.

Regards,
Christian.

> -Daniel
>

Christian König Oct. 2, 2019, 8:37 a.m. UTC | #12

Hi Daniel,

once more a ping on this. Any more comments or can we get it comitted?

Thanks,
Christian.

Am 24.09.19 um 11:50 schrieb Christian König:
> Am 17.09.19 um 16:56 schrieb Daniel Vetter:
>> [SNIP]
>>>>>>>>>>>>       +    /* When either the importer or the exporter 
>>>>>>>>>>>> can't handle dynamic
>>>>>>>>>>>> +     * mappings we cache the mapping here to avoid issues 
>>>>>>>>>>>> with the
>>>>>>>>>>>> +     * reservation object lock.
>>>>>>>>>>>> +     */
>>>>>>>>>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
>>>>>>>>>>>> +        dma_buf_is_dynamic(dmabuf)) {
>>>>>>>>>>>> +        struct sg_table *sgt;
>>>>>>>>>>>> +
>>>>>>>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>>>>>>> + dma_resv_lock(attach->dmabuf->resv, NULL);
>>>>>>>>>>>> +
>>>>>>>>>>>> +        sgt = dmabuf->ops->map_dma_buf(attach, 
>>>>>>>>>>>> DMA_BIDIRECTIONAL);
>>>>>>>>>>> Now we're back to enforcing DMA_BIDI, which works nicely 
>>>>>>>>>>> around the
>>>>>>>>>>> locking pain, but apparently upsets the arm-soc folks who 
>>>>>>>>>>> want to
>>>>>>>>>>> control
>>>>>>>>>>> this better.
>>>>>>>>>> Take another look at dma_buf_map_attachment(), we still try 
>>>>>>>>>> to get the
>>>>>>>>>> caching there for ARM.
>>>>>>>>>>
>>>>>>>>>> What we do here is to bidirectionally map the buffer to avoid 
>>>>>>>>>> the
>>>>>>>>>> locking hydra when importer and exporter disagree on locking.
>>>>>>>>>>
>>>>>>>>>> So the ARM folks can easily avoid that by switching to 
>>>>>>>>>> dynamic locking
>>>>>>>>>> for both.
>>>>>>>> So you still break the contract between importer and exporter, 
>>>>>>>> except not
>>>>>>>> for anything that's run in intel-gfx-ci so all is good?
>>>>>>> No, the contract between importer and exporter stays exactly the 
>>>>>>> same it
>>>>>>> is currently as long as you don't switch to dynamic dma-buf 
>>>>>>> handling.
>>>>>>>
>>>>>>> There is no functional change for the ARM folks here. The only 
>>>>>>> change
>>>>>>> which takes effect is between i915 and amdgpu and that is perfectly
>>>>>>> covered by intel-gfx-ci.
>>>>>> There's people who want to run amdgpu on ARM?
>>>>> Sure there are, we even recently fixed some bugs for this.
>>>>>
>>>>> But as far as I know there is no one currently which is affect by 
>>>>> this
>>>>> change on ARM with amdgpu.
>>>> But don't you break them with this now?
>>> No, we see the bidirectional attachment as compatible with the other 
>>> ones.
>>>
>>>> amdgpu will soon set the dynamic flag on exports, which forces the 
>>>> caching
>>>> at create time (to avoid the locking fun), which will then result in a
>>>> EBUSY at map_attachment time because we have a cached mapping, but 
>>>> it's
>>>> the wrong type.
>>> See the check in dma_buf_map_attachment():
>>>
>>>       if (attach->dir != direction && attach->dir != DMA_BIDIRECTIONAL)
>>>           return ERR_PTR(-EBUSY);
>> Hm, I misread this. So yeah should work, +/- the issue that we might
>> not flush enough. But I guess that can be fixed whenever, it's not
>> like dma-api semantics are a great fit for us. Maybe a fixme comment
>> would be useful here ... I'll look at this tomorrow or so because atm
>> brain is slow, I'm down with the usual post-conference cold it seems
>> :-/
>
> Hope your are feeling better now, adding a comment is of course not a 
> problem.
>
> With that fixed can I get an reviewed-by or at least and acked-by?
>
> I want to land at least some parts of those changes now.
>
> Regards,
> Christian.
>
>> -Daniel
>>
>

Daniel Vetter Oct. 8, 2019, 8:55 a.m. UTC | #13

On Thu, Aug 29, 2019 at 04:29:14PM +0200, Christian König wrote:
> This patch is a stripped down version of the locking changes
> necessary to support dynamic DMA-buf handling.
> 
> For compatibility we cache the DMA-buf mapping as soon as
> exporter/importer disagree on the dynamic handling.


Needs huge summary of all the discussions and options we've considered
here. Copypasta totally fine with me if you don't feel like typing up the
entire sage anew, but imo really the more the better. Iirc we've had 3-4
dead-ends with different attempts at caching or locking rules.
 
> Signed-off-by: Christian König <christian.koenig@amd.com>
> ---
>  drivers/dma-buf/dma-buf.c | 90 ++++++++++++++++++++++++++++++++++++---
>  include/linux/dma-buf.h   | 51 +++++++++++++++++++++-
>  2 files changed, 133 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> index 433d91d710e4..65052d52602b 100644
> --- a/drivers/dma-buf/dma-buf.c
> +++ b/drivers/dma-buf/dma-buf.c
> @@ -525,6 +525,10 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info)
>  		return ERR_PTR(-EINVAL);
>  	}
>  
> +	if (WARN_ON(exp_info->ops->cache_sgt_mapping &&
> +		    exp_info->ops->dynamic_mapping))
> +		return ERR_PTR(-EINVAL);
> +
>  	if (!try_module_get(exp_info->owner))
>  		return ERR_PTR(-ENOENT);
>  
> @@ -645,10 +649,11 @@ void dma_buf_put(struct dma_buf *dmabuf)
>  EXPORT_SYMBOL_GPL(dma_buf_put);
>  
>  /**
> - * dma_buf_attach - Add the device to dma_buf's attachments list; optionally,
> + * dma_buf_dynamic_attach - Add the device to dma_buf's attachments list; optionally,
>   * calls attach() of dma_buf_ops to allow device-specific attach functionality
> - * @dmabuf:	[in]	buffer to attach device to.
> - * @dev:	[in]	device to be attached.
> + * @dmabuf:		[in]	buffer to attach device to.
> + * @dev:		[in]	device to be attached.
> + * @dynamic_mapping:	[in]	calling convention for map/unmap
>   *
>   * Returns struct dma_buf_attachment pointer for this attachment. Attachments
>   * must be cleaned up by calling dma_buf_detach().
> @@ -662,8 +667,9 @@ EXPORT_SYMBOL_GPL(dma_buf_put);
>   * accessible to @dev, and cannot be moved to a more suitable place. This is
>   * indicated with the error code -EBUSY.
>   */
> -struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> -					  struct device *dev)
> +struct dma_buf_attachment *
> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
> +		       bool dynamic_mapping)
>  {
>  	struct dma_buf_attachment *attach;
>  	int ret;
> @@ -677,6 +683,7 @@ struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>  
>  	attach->dev = dev;
>  	attach->dmabuf = dmabuf;
> +	attach->dynamic_mapping = dynamic_mapping;
>  
>  	mutex_lock(&dmabuf->lock);
>  
> @@ -685,16 +692,64 @@ struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
>  		if (ret)
>  			goto err_attach;
>  	}
> +	dma_resv_lock(dmabuf->resv, NULL);
>  	list_add(&attach->node, &dmabuf->attachments);
> +	dma_resv_unlock(dmabuf->resv);

Still inconsistent locking here with dma_resv_lock vs. dma_buf->lock. I'd
completely switch over dmabuf->attachments to dma_resv_lock and remove
dmabuf->lock for that use case (and just keep it for the vmap counter, at
least for now, but that should be moved over too imo in a later patch
perhaps).

Might also be a patch collision, but set_name is not switched over.

Also please update the kerneldoc for ->attachments and ->node to mention
which lock protects them.

>  
>  	mutex_unlock(&dmabuf->lock);
>  
> +	/* When either the importer or the exporter can't handle dynamic
> +	 * mappings we cache the mapping here to avoid issues with the
> +	 * reservation object lock.
> +	 */
> +	if (dma_buf_attachment_is_dynamic(attach) !=
> +	    dma_buf_is_dynamic(dmabuf)) {
> +		struct sg_table *sgt;
> +
> +		if (dma_buf_is_dynamic(attach->dmabuf))
> +			dma_resv_lock(attach->dmabuf->resv, NULL);
> +
> +		sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL);
> +		if (!sgt)
> +			sgt = ERR_PTR(-ENOMEM);
> +		if (IS_ERR(sgt)) {
> +			ret = PTR_ERR(sgt);
> +			goto err_unlock;
> +		}
> +		if (dma_buf_is_dynamic(attach->dmabuf))
> +			dma_resv_unlock(attach->dmabuf->resv);
> +		attach->sgt = sgt;
> +		attach->dir = DMA_BIDIRECTIONAL;
> +	}
> +
>  	return attach;
>  
>  err_attach:
>  	kfree(attach);
>  	mutex_unlock(&dmabuf->lock);
>  	return ERR_PTR(ret);
> +
> +err_unlock:
> +	if (dma_buf_is_dynamic(attach->dmabuf))
> +		dma_resv_unlock(attach->dmabuf->resv);
> +
> +	dma_buf_detach(dmabuf, attach);
> +	return ERR_PTR(ret);
> +}
> +EXPORT_SYMBOL_GPL(dma_buf_dynamic_attach);
> +
> +/**
> + * dma_buf_attach - Wrapper for dma_buf_dynamic_attach
> + * @dmabuf:	[in]	buffer to attach device to.
> + * @dev:	[in]	device to be attached.
> + *
> + * Wrapper to call dma_buf_dynamic_attach() for drivers which still use a static
> + * mapping.
> + */
> +struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> +					  struct device *dev)
> +{
> +	return dma_buf_dynamic_attach(dmabuf, dev, false);
>  }
>  EXPORT_SYMBOL_GPL(dma_buf_attach);
>  
> @@ -711,11 +766,20 @@ void dma_buf_detach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach)
>  	if (WARN_ON(!dmabuf || !attach))
>  		return;
>  
> -	if (attach->sgt)
> +	if (attach->sgt) {
> +		if (dma_buf_is_dynamic(attach->dmabuf))
> +			dma_resv_lock(attach->dmabuf->resv, NULL);
> +
>  		dmabuf->ops->unmap_dma_buf(attach, attach->sgt, attach->dir);
>  
> +		if (dma_buf_is_dynamic(attach->dmabuf))
> +			dma_resv_unlock(attach->dmabuf->resv);
> +	}
> +
>  	mutex_lock(&dmabuf->lock);
> +	dma_resv_lock(dmabuf->resv, NULL);
>  	list_del(&attach->node);
> +	dma_resv_unlock(dmabuf->resv);
>  	if (dmabuf->ops->detach)
>  		dmabuf->ops->detach(dmabuf, attach);
>  
> @@ -749,6 +813,9 @@ struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *attach,
>  	if (WARN_ON(!attach || !attach->dmabuf))
>  		return ERR_PTR(-EINVAL);
>  
> +	if (dma_buf_attachment_is_dynamic(attach))
> +		dma_resv_assert_held(attach->dmabuf->resv);
> +
>  	if (attach->sgt) {
>  		/*
>  		 * Two mappings with different directions for the same
> @@ -761,6 +828,9 @@ struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *attach,
>  		return attach->sgt;
>  	}
>  
> +	if (dma_buf_is_dynamic(attach->dmabuf))
> +		dma_resv_assert_held(attach->dmabuf->resv);
> +
>  	sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction);
>  	if (!sg_table)
>  		sg_table = ERR_PTR(-ENOMEM);
> @@ -793,9 +863,15 @@ void dma_buf_unmap_attachment(struct dma_buf_attachment *attach,
>  	if (WARN_ON(!attach || !attach->dmabuf || !sg_table))
>  		return;
>  
> +	if (dma_buf_attachment_is_dynamic(attach))
> +		dma_resv_assert_held(attach->dmabuf->resv);
> +
>  	if (attach->sgt == sg_table)
>  		return;
>  
> +	if (dma_buf_is_dynamic(attach->dmabuf))
> +		dma_resv_assert_held(attach->dmabuf->resv);
> +
>  	attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, direction);
>  }
>  EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment);
> @@ -1219,10 +1295,12 @@ static int dma_buf_debug_show(struct seq_file *s, void *unused)
>  		seq_puts(s, "\tAttached Devices:\n");
>  		attach_count = 0;
>  
> +		dma_resv_lock(buf_obj->resv, NULL);
>  		list_for_each_entry(attach_obj, &buf_obj->attachments, node) {
>  			seq_printf(s, "\t%s\n", dev_name(attach_obj->dev));
>  			attach_count++;
>  		}
> +		dma_resv_unlock(buf_obj->resv);
>  
>  		seq_printf(s, "Total %d devices attached\n\n",
>  				attach_count);
> diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h
> index ec212cb27fdc..a8f8b2b812fd 100644
> --- a/include/linux/dma-buf.h
> +++ b/include/linux/dma-buf.h
> @@ -42,6 +42,17 @@ struct dma_buf_ops {
>  	  */
>  	bool cache_sgt_mapping;
>  
> +	/**
> +	 * @dynamic_mapping:
> +	 *
> +	 * If true the framework makes sure that the map/unmap_dma_buf
> +	 * callbacks are always called with the dma_resv object locked.
> +	 *
> +	 * If false the framework makes ure that the map/unmap_dma_buf
> +	 * callbacks are always called without the dma_resv object locked.
> +	 */

Please add a note that this is an cache_sgt_mapping are exlusive.

> +	bool dynamic_mapping;
> +
>  	/**
>  	 * @attach:
>  	 *
> @@ -109,6 +120,9 @@ struct dma_buf_ops {
>  	 * any other kind of sharing that the exporter might wish to make
>  	 * available to buffer-users.
>  	 *
> +	 * This is always called with the dmabuf->resv object locked when
> +	 * the dynamic_mapping flag is true.
> +	 *
>  	 * Returns:
>  	 *
>  	 * A &sg_table scatter list of or the backing storage of the DMA buffer,
> @@ -327,6 +341,8 @@ struct dma_buf {
>   * @sgt: cached mapping.
>   * @dir: direction of cached mapping.
>   * @priv: exporter specific attachment data.
> + * @dynamic_mapping: true if dma_buf_map/unmap_attachment() is called with the
> + * dma_resv lock held.
>   *
>   * This structure holds the attachment information between the dma_buf buffer
>   * and its user device(s). The list contains one attachment struct per device
> @@ -343,6 +359,7 @@ struct dma_buf_attachment {
>  	struct list_head node;
>  	struct sg_table *sgt;
>  	enum dma_data_direction dir;
> +	bool dynamic_mapping;
>  	void *priv;
>  };
>  
> @@ -394,10 +411,39 @@ static inline void get_dma_buf(struct dma_buf *dmabuf)
>  	get_file(dmabuf->file);
>  }
>  
> +/**
> + * dma_buf_is_dynamic - check if a DMA-buf uses dynamic mappings.
> + * @dmabuf: the DMA-buf to check
> + *
> + * Returns true if a DMA-buf exporter wants to be called with the dma_resv
> + * locked, false if it doesn't wants to be called with the lock held.
> + */
> +static inline bool dma_buf_is_dynamic(struct dma_buf *dmabuf)
> +{
> +	return dmabuf->ops->dynamic_mapping;
> +}
> +
> +/**
> + * dma_buf_attachment_is_dynamic - check if a DMA-buf attachment uses dynamic
> + * mappinsg
> + * @attach: the DMA-buf attachment to check
> + *
> + * Returns true if a DMA-buf importer wants to call the map/unmap functions with
> + * the dma_resv lock held.
> + */
> +static inline bool
> +dma_buf_attachment_is_dynamic(struct dma_buf_attachment *attach)
> +{
> +	return attach->dynamic_mapping;
> +}
> +
>  struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
> -							struct device *dev);
> +					  struct device *dev);
> +struct dma_buf_attachment *
> +dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
> +		       bool dynamic_mapping);
>  void dma_buf_detach(struct dma_buf *dmabuf,
> -				struct dma_buf_attachment *dmabuf_attach);
> +		    struct dma_buf_attachment *attach);
>  
>  struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info);
>  
> @@ -409,6 +455,7 @@ struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *,
>  					enum dma_data_direction);
>  void dma_buf_unmap_attachment(struct dma_buf_attachment *, struct sg_table *,
>  				enum dma_data_direction);
> +void dma_buf_move_notify(struct dma_buf *dma_buf);
>  int dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
>  			     enum dma_data_direction dir);
>  int dma_buf_end_cpu_access(struct dma_buf *dma_buf,

Aside from the nits here I think this is ok, count me convinced on
principle.
-Daniel

Daniel Vetter Oct. 8, 2019, 8:55 a.m. UTC | #14

On Wed, Oct 02, 2019 at 08:37:50AM +0000, Koenig, Christian wrote:
> Hi Daniel,
> 
> once more a ping on this. Any more comments or can we get it comitted?

Sorry got a bit smashed past weeks, but should be resurrected now back
from xdc.
-Daniel
> 
> Thanks,
> Christian.
> 
> Am 24.09.19 um 11:50 schrieb Christian König:
> > Am 17.09.19 um 16:56 schrieb Daniel Vetter:
> >> [SNIP]
> >>>>>>>>>>>>       +    /* When either the importer or the exporter 
> >>>>>>>>>>>> can't handle dynamic
> >>>>>>>>>>>> +     * mappings we cache the mapping here to avoid issues 
> >>>>>>>>>>>> with the
> >>>>>>>>>>>> +     * reservation object lock.
> >>>>>>>>>>>> +     */
> >>>>>>>>>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
> >>>>>>>>>>>> +        dma_buf_is_dynamic(dmabuf)) {
> >>>>>>>>>>>> +        struct sg_table *sgt;
> >>>>>>>>>>>> +
> >>>>>>>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>>>>>>>>> + dma_resv_lock(attach->dmabuf->resv, NULL);
> >>>>>>>>>>>> +
> >>>>>>>>>>>> +        sgt = dmabuf->ops->map_dma_buf(attach, 
> >>>>>>>>>>>> DMA_BIDIRECTIONAL);
> >>>>>>>>>>> Now we're back to enforcing DMA_BIDI, which works nicely 
> >>>>>>>>>>> around the
> >>>>>>>>>>> locking pain, but apparently upsets the arm-soc folks who 
> >>>>>>>>>>> want to
> >>>>>>>>>>> control
> >>>>>>>>>>> this better.
> >>>>>>>>>> Take another look at dma_buf_map_attachment(), we still try 
> >>>>>>>>>> to get the
> >>>>>>>>>> caching there for ARM.
> >>>>>>>>>>
> >>>>>>>>>> What we do here is to bidirectionally map the buffer to avoid 
> >>>>>>>>>> the
> >>>>>>>>>> locking hydra when importer and exporter disagree on locking.
> >>>>>>>>>>
> >>>>>>>>>> So the ARM folks can easily avoid that by switching to 
> >>>>>>>>>> dynamic locking
> >>>>>>>>>> for both.
> >>>>>>>> So you still break the contract between importer and exporter, 
> >>>>>>>> except not
> >>>>>>>> for anything that's run in intel-gfx-ci so all is good?
> >>>>>>> No, the contract between importer and exporter stays exactly the 
> >>>>>>> same it
> >>>>>>> is currently as long as you don't switch to dynamic dma-buf 
> >>>>>>> handling.
> >>>>>>>
> >>>>>>> There is no functional change for the ARM folks here. The only 
> >>>>>>> change
> >>>>>>> which takes effect is between i915 and amdgpu and that is perfectly
> >>>>>>> covered by intel-gfx-ci.
> >>>>>> There's people who want to run amdgpu on ARM?
> >>>>> Sure there are, we even recently fixed some bugs for this.
> >>>>>
> >>>>> But as far as I know there is no one currently which is affect by 
> >>>>> this
> >>>>> change on ARM with amdgpu.
> >>>> But don't you break them with this now?
> >>> No, we see the bidirectional attachment as compatible with the other 
> >>> ones.
> >>>
> >>>> amdgpu will soon set the dynamic flag on exports, which forces the 
> >>>> caching
> >>>> at create time (to avoid the locking fun), which will then result in a
> >>>> EBUSY at map_attachment time because we have a cached mapping, but 
> >>>> it's
> >>>> the wrong type.
> >>> See the check in dma_buf_map_attachment():
> >>>
> >>>       if (attach->dir != direction && attach->dir != DMA_BIDIRECTIONAL)
> >>>           return ERR_PTR(-EBUSY);
> >> Hm, I misread this. So yeah should work, +/- the issue that we might
> >> not flush enough. But I guess that can be fixed whenever, it's not
> >> like dma-api semantics are a great fit for us. Maybe a fixme comment
> >> would be useful here ... I'll look at this tomorrow or so because atm
> >> brain is slow, I'm down with the usual post-conference cold it seems
> >> :-/
> >
> > Hope your are feeling better now, adding a comment is of course not a 
> > problem.
> >
> > With that fixed can I get an reviewed-by or at least and acked-by?
> >
> > I want to land at least some parts of those changes now.
> >
> > Regards,
> > Christian.
> >
> >> -Daniel
> >>
> >
>

Christian König Oct. 16, 2019, 1:46 p.m. UTC | #15

Am 08.10.19 um 10:55 schrieb Daniel Vetter:
> On Wed, Oct 02, 2019 at 08:37:50AM +0000, Koenig, Christian wrote:
>> Hi Daniel,
>>
>> once more a ping on this. Any more comments or can we get it comitted?
> Sorry got a bit smashed past weeks, but should be resurrected now back
> from xdc.

And any more thoughts on this? I mean we are blocked for month on this 
now :(

Thanks,
Christian.

> -Daniel
>> Thanks,
>> Christian.
>>
>> Am 24.09.19 um 11:50 schrieb Christian König:
>>> Am 17.09.19 um 16:56 schrieb Daniel Vetter:
>>>> [SNIP]
>>>>>>>>>>>>>>        +    /* When either the importer or the exporter
>>>>>>>>>>>>>> can't handle dynamic
>>>>>>>>>>>>>> +     * mappings we cache the mapping here to avoid issues
>>>>>>>>>>>>>> with the
>>>>>>>>>>>>>> +     * reservation object lock.
>>>>>>>>>>>>>> +     */
>>>>>>>>>>>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
>>>>>>>>>>>>>> +        dma_buf_is_dynamic(dmabuf)) {
>>>>>>>>>>>>>> +        struct sg_table *sgt;
>>>>>>>>>>>>>> +
>>>>>>>>>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>>>>>>>>> + dma_resv_lock(attach->dmabuf->resv, NULL);
>>>>>>>>>>>>>> +
>>>>>>>>>>>>>> +        sgt = dmabuf->ops->map_dma_buf(attach,
>>>>>>>>>>>>>> DMA_BIDIRECTIONAL);
>>>>>>>>>>>>> Now we're back to enforcing DMA_BIDI, which works nicely
>>>>>>>>>>>>> around the
>>>>>>>>>>>>> locking pain, but apparently upsets the arm-soc folks who
>>>>>>>>>>>>> want to
>>>>>>>>>>>>> control
>>>>>>>>>>>>> this better.
>>>>>>>>>>>> Take another look at dma_buf_map_attachment(), we still try
>>>>>>>>>>>> to get the
>>>>>>>>>>>> caching there for ARM.
>>>>>>>>>>>>
>>>>>>>>>>>> What we do here is to bidirectionally map the buffer to avoid
>>>>>>>>>>>> the
>>>>>>>>>>>> locking hydra when importer and exporter disagree on locking.
>>>>>>>>>>>>
>>>>>>>>>>>> So the ARM folks can easily avoid that by switching to
>>>>>>>>>>>> dynamic locking
>>>>>>>>>>>> for both.
>>>>>>>>>> So you still break the contract between importer and exporter,
>>>>>>>>>> except not
>>>>>>>>>> for anything that's run in intel-gfx-ci so all is good?
>>>>>>>>> No, the contract between importer and exporter stays exactly the
>>>>>>>>> same it
>>>>>>>>> is currently as long as you don't switch to dynamic dma-buf
>>>>>>>>> handling.
>>>>>>>>>
>>>>>>>>> There is no functional change for the ARM folks here. The only
>>>>>>>>> change
>>>>>>>>> which takes effect is between i915 and amdgpu and that is perfectly
>>>>>>>>> covered by intel-gfx-ci.
>>>>>>>> There's people who want to run amdgpu on ARM?
>>>>>>> Sure there are, we even recently fixed some bugs for this.
>>>>>>>
>>>>>>> But as far as I know there is no one currently which is affect by
>>>>>>> this
>>>>>>> change on ARM with amdgpu.
>>>>>> But don't you break them with this now?
>>>>> No, we see the bidirectional attachment as compatible with the other
>>>>> ones.
>>>>>
>>>>>> amdgpu will soon set the dynamic flag on exports, which forces the
>>>>>> caching
>>>>>> at create time (to avoid the locking fun), which will then result in a
>>>>>> EBUSY at map_attachment time because we have a cached mapping, but
>>>>>> it's
>>>>>> the wrong type.
>>>>> See the check in dma_buf_map_attachment():
>>>>>
>>>>>        if (attach->dir != direction && attach->dir != DMA_BIDIRECTIONAL)
>>>>>            return ERR_PTR(-EBUSY);
>>>> Hm, I misread this. So yeah should work, +/- the issue that we might
>>>> not flush enough. But I guess that can be fixed whenever, it's not
>>>> like dma-api semantics are a great fit for us. Maybe a fixme comment
>>>> would be useful here ... I'll look at this tomorrow or so because atm
>>>> brain is slow, I'm down with the usual post-conference cold it seems
>>>> :-/
>>> Hope your are feeling better now, adding a comment is of course not a
>>> problem.
>>>
>>> With that fixed can I get an reviewed-by or at least and acked-by?
>>>
>>> I want to land at least some parts of those changes now.
>>>
>>> Regards,
>>> Christian.
>>>
>>>> -Daniel
>>>>

Daniel Vetter Oct. 16, 2019, 2:23 p.m. UTC | #16

On Wed, Oct 16, 2019 at 3:46 PM Koenig, Christian
<Christian.Koenig@amd.com> wrote:
>
> Am 08.10.19 um 10:55 schrieb Daniel Vetter:
> > On Wed, Oct 02, 2019 at 08:37:50AM +0000, Koenig, Christian wrote:
> >> Hi Daniel,
> >>
> >> once more a ping on this. Any more comments or can we get it comitted?
> > Sorry got a bit smashed past weeks, but should be resurrected now back
> > from xdc.
>
> And any more thoughts on this? I mean we are blocked for month on this
> now :(

I replied to both 1 and 2 in this series on 8th Oct. You even replied
to me in the thread on patch 2 ... but not here (I top posted since
this detour here just me being confused).
-Daniel

>
> Thanks,
> Christian.
>
> > -Daniel
> >> Thanks,
> >> Christian.
> >>
> >> Am 24.09.19 um 11:50 schrieb Christian König:
> >>> Am 17.09.19 um 16:56 schrieb Daniel Vetter:
> >>>> [SNIP]
> >>>>>>>>>>>>>>        +    /* When either the importer or the exporter
> >>>>>>>>>>>>>> can't handle dynamic
> >>>>>>>>>>>>>> +     * mappings we cache the mapping here to avoid issues
> >>>>>>>>>>>>>> with the
> >>>>>>>>>>>>>> +     * reservation object lock.
> >>>>>>>>>>>>>> +     */
> >>>>>>>>>>>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
> >>>>>>>>>>>>>> +        dma_buf_is_dynamic(dmabuf)) {
> >>>>>>>>>>>>>> +        struct sg_table *sgt;
> >>>>>>>>>>>>>> +
> >>>>>>>>>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
> >>>>>>>>>>>>>> + dma_resv_lock(attach->dmabuf->resv, NULL);
> >>>>>>>>>>>>>> +
> >>>>>>>>>>>>>> +        sgt = dmabuf->ops->map_dma_buf(attach,
> >>>>>>>>>>>>>> DMA_BIDIRECTIONAL);
> >>>>>>>>>>>>> Now we're back to enforcing DMA_BIDI, which works nicely
> >>>>>>>>>>>>> around the
> >>>>>>>>>>>>> locking pain, but apparently upsets the arm-soc folks who
> >>>>>>>>>>>>> want to
> >>>>>>>>>>>>> control
> >>>>>>>>>>>>> this better.
> >>>>>>>>>>>> Take another look at dma_buf_map_attachment(), we still try
> >>>>>>>>>>>> to get the
> >>>>>>>>>>>> caching there for ARM.
> >>>>>>>>>>>>
> >>>>>>>>>>>> What we do here is to bidirectionally map the buffer to avoid
> >>>>>>>>>>>> the
> >>>>>>>>>>>> locking hydra when importer and exporter disagree on locking.
> >>>>>>>>>>>>
> >>>>>>>>>>>> So the ARM folks can easily avoid that by switching to
> >>>>>>>>>>>> dynamic locking
> >>>>>>>>>>>> for both.
> >>>>>>>>>> So you still break the contract between importer and exporter,
> >>>>>>>>>> except not
> >>>>>>>>>> for anything that's run in intel-gfx-ci so all is good?
> >>>>>>>>> No, the contract between importer and exporter stays exactly the
> >>>>>>>>> same it
> >>>>>>>>> is currently as long as you don't switch to dynamic dma-buf
> >>>>>>>>> handling.
> >>>>>>>>>
> >>>>>>>>> There is no functional change for the ARM folks here. The only
> >>>>>>>>> change
> >>>>>>>>> which takes effect is between i915 and amdgpu and that is perfectly
> >>>>>>>>> covered by intel-gfx-ci.
> >>>>>>>> There's people who want to run amdgpu on ARM?
> >>>>>>> Sure there are, we even recently fixed some bugs for this.
> >>>>>>>
> >>>>>>> But as far as I know there is no one currently which is affect by
> >>>>>>> this
> >>>>>>> change on ARM with amdgpu.
> >>>>>> But don't you break them with this now?
> >>>>> No, we see the bidirectional attachment as compatible with the other
> >>>>> ones.
> >>>>>
> >>>>>> amdgpu will soon set the dynamic flag on exports, which forces the
> >>>>>> caching
> >>>>>> at create time (to avoid the locking fun), which will then result in a
> >>>>>> EBUSY at map_attachment time because we have a cached mapping, but
> >>>>>> it's
> >>>>>> the wrong type.
> >>>>> See the check in dma_buf_map_attachment():
> >>>>>
> >>>>>        if (attach->dir != direction && attach->dir != DMA_BIDIRECTIONAL)
> >>>>>            return ERR_PTR(-EBUSY);
> >>>> Hm, I misread this. So yeah should work, +/- the issue that we might
> >>>> not flush enough. But I guess that can be fixed whenever, it's not
> >>>> like dma-api semantics are a great fit for us. Maybe a fixme comment
> >>>> would be useful here ... I'll look at this tomorrow or so because atm
> >>>> brain is slow, I'm down with the usual post-conference cold it seems
> >>>> :-/
> >>> Hope your are feeling better now, adding a comment is of course not a
> >>> problem.
> >>>
> >>> With that fixed can I get an reviewed-by or at least and acked-by?
> >>>
> >>> I want to land at least some parts of those changes now.
> >>>
> >>> Regards,
> >>> Christian.
> >>>
> >>>> -Daniel
> >>>>
>

Christian König Oct. 17, 2019, 9:04 a.m. UTC | #17

Am 16.10.19 um 16:23 schrieb Daniel Vetter:
> On Wed, Oct 16, 2019 at 3:46 PM Koenig, Christian
> <Christian.Koenig@amd.com> wrote:
>> Am 08.10.19 um 10:55 schrieb Daniel Vetter:
>>> On Wed, Oct 02, 2019 at 08:37:50AM +0000, Koenig, Christian wrote:
>>>> Hi Daniel,
>>>>
>>>> once more a ping on this. Any more comments or can we get it comitted?
>>> Sorry got a bit smashed past weeks, but should be resurrected now back
>>> from xdc.
>> And any more thoughts on this? I mean we are blocked for month on this
>> now :(
> I replied to both 1 and 2 in this series on 8th Oct. You even replied
> to me in the thread on patch 2 ... but not here (I top posted since
> this detour here just me being confused).

Ok, in this case its my fault. I totally missed your reply on 1 and 
thought that the reply on 2 was actually for a different thread.

I'm going to submit the TTM changes separately, cause that is actually a 
bug fix for a completely different issue which just happens to surface 
because we change the locking.

Thanks,
Christian.

> -Daniel
>
>> Thanks,
>> Christian.
>>
>>> -Daniel
>>>> Thanks,
>>>> Christian.
>>>>
>>>> Am 24.09.19 um 11:50 schrieb Christian König:
>>>>> Am 17.09.19 um 16:56 schrieb Daniel Vetter:
>>>>>> [SNIP]
>>>>>>>>>>>>>>>>         +    /* When either the importer or the exporter
>>>>>>>>>>>>>>>> can't handle dynamic
>>>>>>>>>>>>>>>> +     * mappings we cache the mapping here to avoid issues
>>>>>>>>>>>>>>>> with the
>>>>>>>>>>>>>>>> +     * reservation object lock.
>>>>>>>>>>>>>>>> +     */
>>>>>>>>>>>>>>>> +    if (dma_buf_attachment_is_dynamic(attach) !=
>>>>>>>>>>>>>>>> +        dma_buf_is_dynamic(dmabuf)) {
>>>>>>>>>>>>>>>> +        struct sg_table *sgt;
>>>>>>>>>>>>>>>> +
>>>>>>>>>>>>>>>> +        if (dma_buf_is_dynamic(attach->dmabuf))
>>>>>>>>>>>>>>>> + dma_resv_lock(attach->dmabuf->resv, NULL);
>>>>>>>>>>>>>>>> +
>>>>>>>>>>>>>>>> +        sgt = dmabuf->ops->map_dma_buf(attach,
>>>>>>>>>>>>>>>> DMA_BIDIRECTIONAL);
>>>>>>>>>>>>>>> Now we're back to enforcing DMA_BIDI, which works nicely
>>>>>>>>>>>>>>> around the
>>>>>>>>>>>>>>> locking pain, but apparently upsets the arm-soc folks who
>>>>>>>>>>>>>>> want to
>>>>>>>>>>>>>>> control
>>>>>>>>>>>>>>> this better.
>>>>>>>>>>>>>> Take another look at dma_buf_map_attachment(), we still try
>>>>>>>>>>>>>> to get the
>>>>>>>>>>>>>> caching there for ARM.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> What we do here is to bidirectionally map the buffer to avoid
>>>>>>>>>>>>>> the
>>>>>>>>>>>>>> locking hydra when importer and exporter disagree on locking.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> So the ARM folks can easily avoid that by switching to
>>>>>>>>>>>>>> dynamic locking
>>>>>>>>>>>>>> for both.
>>>>>>>>>>>> So you still break the contract between importer and exporter,
>>>>>>>>>>>> except not
>>>>>>>>>>>> for anything that's run in intel-gfx-ci so all is good?
>>>>>>>>>>> No, the contract between importer and exporter stays exactly the
>>>>>>>>>>> same it
>>>>>>>>>>> is currently as long as you don't switch to dynamic dma-buf
>>>>>>>>>>> handling.
>>>>>>>>>>>
>>>>>>>>>>> There is no functional change for the ARM folks here. The only
>>>>>>>>>>> change
>>>>>>>>>>> which takes effect is between i915 and amdgpu and that is perfectly
>>>>>>>>>>> covered by intel-gfx-ci.
>>>>>>>>>> There's people who want to run amdgpu on ARM?
>>>>>>>>> Sure there are, we even recently fixed some bugs for this.
>>>>>>>>>
>>>>>>>>> But as far as I know there is no one currently which is affect by
>>>>>>>>> this
>>>>>>>>> change on ARM with amdgpu.
>>>>>>>> But don't you break them with this now?
>>>>>>> No, we see the bidirectional attachment as compatible with the other
>>>>>>> ones.
>>>>>>>
>>>>>>>> amdgpu will soon set the dynamic flag on exports, which forces the
>>>>>>>> caching
>>>>>>>> at create time (to avoid the locking fun), which will then result in a
>>>>>>>> EBUSY at map_attachment time because we have a cached mapping, but
>>>>>>>> it's
>>>>>>>> the wrong type.
>>>>>>> See the check in dma_buf_map_attachment():
>>>>>>>
>>>>>>>         if (attach->dir != direction && attach->dir != DMA_BIDIRECTIONAL)
>>>>>>>             return ERR_PTR(-EBUSY);
>>>>>> Hm, I misread this. So yeah should work, +/- the issue that we might
>>>>>> not flush enough. But I guess that can be fixed whenever, it's not
>>>>>> like dma-api semantics are a great fit for us. Maybe a fixme comment
>>>>>> would be useful here ... I'll look at this tomorrow or so because atm
>>>>>> brain is slow, I'm down with the usual post-conference cold it seems
>>>>>> :-/
>>>>> Hope your are feeling better now, adding a comment is of course not a
>>>>> problem.
>>>>>
>>>>> With that fixed can I get an reviewed-by or at least and acked-by?
>>>>>
>>>>> I want to land at least some parts of those changes now.
>>>>>
>>>>> Regards,
>>>>> Christian.
>>>>>
>>>>>> -Daniel
>>>>>>
>

[1/4] dma-buf: change DMA-buf locking convention

Commit Message

Comments

Patch