Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Tue, 10 Dec 2019 16:54:36 +0800
From:   Xu Yilum <yilun.xu@intel.com>
To:     Moritz Fischer <mdf@kernel.org>
Cc:     linux-fpga@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] fpga: dfl: support multiple opens on feature device node.
Message-ID: <20191210085435.GA17198@yilunxu-OptiPlex-7050>
References: <1574054441-1568-1-git-send-email-yilun.xu@intel.com>
 <20191210020654.GA7171@archbook>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20191210020654.GA7171@archbook>
User-Agent: Mutt/1.5.24 (2015-08-30)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Mon, Dec 09, 2019 at 06:06:54PM -0800, Moritz Fischer wrote:
> Hi Xu,
> 
> Sorry for the delay.
> 
> On Mon, Nov 18, 2019 at 01:20:41PM +0800, Xu Yilun wrote:
> > Each DFL functional block, e.g. AFU (Accelerated Function Unit) and FME
> > (FPGA Management Engine), could implement more than one function within
> > its region, but current driver only allows one user application to access
> > it by exclusive open on device node. So this is not convenient and
> > flexible for userspace applications, as they have to combine lots of
> > different functions into one single application.
> I'm not entirely sure how I feel about that, wouldn't the right solution
> be to create more devices? If this ends up being 100% passthrough
> basically, VFIO would be a better choice?

The major issue we want to resolve is how to access the same accelerator by
multiple applications. As you know, the accelerator is the dynamic region (AFU)
defined by end user, so we don't know how many sub functions end user will
implement in the one dynamic region (AFU) of FPGA. So driver can't create devices
for logic blocks implemented inside dynamic region.

We don't want 100% passthrough, but only dynamic region (AFU), because for static
region part, we can reuse everything, including hardware logics, kernel drivers and
upper layer stack/lib/app/tools. Use VFIO PCI directly may not be helpful for our case,
and make the device node to support multiple open seems the simplest solution.

> > 
> > This patch removes the limitation here to allow multiple opens to each
> > feature device node for AFU and FME from userspace applications. If user
> > still needs exclusive access to these device node, O_EXCL flag must be
> > issued together with open.
> Wouldn't you wanna default to have exclusive access, and then allow
> multiple opens by means of a management interface, maybe?

I see some examples here which is using O_EXCL flag to indicate exclusive open,
like nvram misc driver. I feel this may be a regular method to follow, I believe
if we can save one userspace interface with this, that will be simpler for users
too.

> > 
> > Signed-off-by: Wu Hao <hao.wu@intel.com>
> > Signed-off-by: Xu Yilun <yilun.xu@intel.com>
> > ---
> >  drivers/fpga/dfl-afu-main.c | 26 +++++++++++++++-----------
> >  drivers/fpga/dfl-fme-main.c | 19 ++++++++++++-------
> >  drivers/fpga/dfl.c          | 15 +++++++++++++--
> >  drivers/fpga/dfl.h          | 35 +++++++++++++++++++++++++++--------
> >  4 files changed, 67 insertions(+), 28 deletions(-)
> > 
> > diff --git a/drivers/fpga/dfl-afu-main.c b/drivers/fpga/dfl-afu-main.c
> > index e4a34dc..c6e0e07 100644
> > --- a/drivers/fpga/dfl-afu-main.c
> > +++ b/drivers/fpga/dfl-afu-main.c
> > @@ -561,14 +561,16 @@ static int afu_open(struct inode *inode, struct file *filp)
> >  	if (WARN_ON(!pdata))
> >  		return -ENODEV;
> >  
> > -	ret = dfl_feature_dev_use_begin(pdata);
> > -	if (ret)
> > -		return ret;
> > -
> > -	dev_dbg(&fdev->dev, "Device File Open\n");
> > -	filp->private_data = fdev;
> > +	mutex_lock(&pdata->lock);
> > +	ret = dfl_feature_dev_use_begin(pdata, filp->f_flags & O_EXCL);
> > +	if (!ret) {
> > +		dev_dbg(&fdev->dev, "Device File Opened %d Times\n",
> > +			dfl_feature_dev_use_count(pdata));
> Could those be trace events instead?

I investigated a little to see how we use trace to get this info.
1. config PROFILE_ANNOTATED_BRANCHES in kernel, and add likely()/unlikely()
   for ret judgement. This adds some overhead to system.
2. config CONFIG_PROFILE_ALL_BRANCHES, no code change here. This adds
   great overhead to system.
3. write a dedicated trace event for this count. I think this is a little
   too heavy, we need to add extra file for it, but this count is just
   minor info to the driver.

What's your suggestion for this? Could we just leave the print here for
simplicity.


> > +		filp->private_data = fdev;
> > +	}
> > +	mutex_unlock(&pdata->lock);
> >  
> > -	return 0;
> > +	return ret;
> >  }
> >  
> >  static int afu_release(struct inode *inode, struct file *filp)
> > @@ -581,12 +583,14 @@ static int afu_release(struct inode *inode, struct file *filp)
> >  	pdata = dev_get_platdata(&pdev->dev);
> >  
> >  	mutex_lock(&pdata->lock);
> > -	__port_reset(pdev);
> > -	afu_dma_region_destroy(pdata);
> > -	mutex_unlock(&pdata->lock);
> > -
> >  	dfl_feature_dev_use_end(pdata);
> >  
> > +	if (!dfl_feature_dev_use_count(pdata)) {
> > +		__port_reset(pdev);
> > +		afu_dma_region_destroy(pdata);
> > +	}
> > +	mutex_unlock(&pdata->lock);
> > +
> >  	return 0;
> >  }
> >  
> > diff --git a/drivers/fpga/dfl-fme-main.c b/drivers/fpga/dfl-fme-main.c
> > index 7c930e6..fda8623 100644
> > --- a/drivers/fpga/dfl-fme-main.c
> > +++ b/drivers/fpga/dfl-fme-main.c
> > @@ -600,14 +600,16 @@ static int fme_open(struct inode *inode, struct file *filp)
> >  	if (WARN_ON(!pdata))
> >  		return -ENODEV;
> >  
> > -	ret = dfl_feature_dev_use_begin(pdata);
> > -	if (ret)
> > -		return ret;
> > -
> > -	dev_dbg(&fdev->dev, "Device File Open\n");
> > -	filp->private_data = pdata;
> > +	mutex_lock(&pdata->lock);
> > +	ret = dfl_feature_dev_use_begin(pdata, filp->f_flags & O_EXCL);
> > +	if (!ret) {
> > +		dev_dbg(&fdev->dev, "Device File Opened %d Times\n",
> > +			dfl_feature_dev_use_count(pdata));
> > +		filp->private_data = pdata;
> > +	}
> > +	mutex_unlock(&pdata->lock);
> >  
> > -	return 0;
> > +	return ret;
> >  }
> >  
> >  static int fme_release(struct inode *inode, struct file *filp)
> > @@ -616,7 +618,10 @@ static int fme_release(struct inode *inode, struct file *filp)
> >  	struct platform_device *pdev = pdata->dev;
> >  
> >  	dev_dbg(&pdev->dev, "Device File Release\n");
> > +
> > +	mutex_lock(&pdata->lock);
> >  	dfl_feature_dev_use_end(pdata);
> > +	mutex_unlock(&pdata->lock);
> >  
> >  	return 0;
> >  }
> > diff --git a/drivers/fpga/dfl.c b/drivers/fpga/dfl.c
> > index 96a2b82..9909948 100644
> > --- a/drivers/fpga/dfl.c
> > +++ b/drivers/fpga/dfl.c
> > @@ -1079,6 +1079,7 @@ static int __init dfl_fpga_init(void)
> >   */
> >  int dfl_fpga_cdev_release_port(struct dfl_fpga_cdev *cdev, int port_id)
> >  {
> > +	struct dfl_feature_platform_data *pdata;
> >  	struct platform_device *port_pdev;
> >  	int ret = -ENODEV;
> >  
> > @@ -1093,7 +1094,11 @@ int dfl_fpga_cdev_release_port(struct dfl_fpga_cdev *cdev, int port_id)
> >  		goto put_dev_exit;
> >  	}
> >  
> > -	ret = dfl_feature_dev_use_begin(dev_get_platdata(&port_pdev->dev));
> > +	pdata = dev_get_platdata(&port_pdev->dev);
> > +
> > +	mutex_lock(&pdata->lock);
> > +	ret = dfl_feature_dev_use_begin(pdata, true);
> > +	mutex_unlock(&pdata->lock);
> >  	if (ret)
> >  		goto put_dev_exit;
> >  
> > @@ -1120,6 +1125,7 @@ EXPORT_SYMBOL_GPL(dfl_fpga_cdev_release_port);
> >   */
> >  int dfl_fpga_cdev_assign_port(struct dfl_fpga_cdev *cdev, int port_id)
> >  {
> > +	struct dfl_feature_platform_data *pdata;
> >  	struct platform_device *port_pdev;
> >  	int ret = -ENODEV;
> >  
> > @@ -1138,7 +1144,12 @@ int dfl_fpga_cdev_assign_port(struct dfl_fpga_cdev *cdev, int port_id)
> >  	if (ret)
> >  		goto put_dev_exit;
> >  
> > -	dfl_feature_dev_use_end(dev_get_platdata(&port_pdev->dev));
> > +	pdata = dev_get_platdata(&port_pdev->dev);
> > +
> > +	mutex_lock(&pdata->lock);
> > +	dfl_feature_dev_use_end(pdata);
> > +	mutex_unlock(&pdata->lock);
> > +
> >  	cdev->released_port_num--;
> >  put_dev_exit:
> >  	put_device(&port_pdev->dev);
> > diff --git a/drivers/fpga/dfl.h b/drivers/fpga/dfl.h
> > index 9f0e656..4a9a33c 100644
> > --- a/drivers/fpga/dfl.h
> > +++ b/drivers/fpga/dfl.h
> > @@ -205,8 +205,6 @@ struct dfl_feature {
> >  	const struct dfl_feature_ops *ops;
> >  };
> >  
> > -#define DEV_STATUS_IN_USE	0
> > -
> >  #define FEATURE_DEV_ID_UNUSED	(-1)
> >  
> >  /**
> > @@ -219,8 +217,9 @@ struct dfl_feature {
> >   * @dfl_cdev: ptr to container device.
> >   * @id: id used for this feature device.
> >   * @disable_count: count for port disable.
> > + * @excl_open: set on feature device exclusive open.
> > + * @open_count: count for feature device open.
> >   * @num: number for sub features.
> > - * @dev_status: dev status (e.g. DEV_STATUS_IN_USE).
> >   * @private: ptr to feature dev private data.
> >   * @features: sub features of this feature dev.
> >   */
> > @@ -232,26 +231,46 @@ struct dfl_feature_platform_data {
> >  	struct dfl_fpga_cdev *dfl_cdev;
> >  	int id;
> >  	unsigned int disable_count;
> > -	unsigned long dev_status;
> > +	bool excl_open;
> > +	int open_count;
> >  	void *private;
> >  	int num;
> >  	struct dfl_feature features[0];
> >  };
> >  
> >  static inline
> 
> > -int dfl_feature_dev_use_begin(struct dfl_feature_platform_data *pdata)
> > +int dfl_feature_dev_use_begin(struct dfl_feature_platform_data *pdata,
> > +			      bool excl)
> >  {
> > -	/* Test and set IN_USE flags to ensure file is exclusively used */
> > -	if (test_and_set_bit_lock(DEV_STATUS_IN_USE, &pdata->dev_status))
> > +	if (pdata->excl_open)
> >  		return -EBUSY;
> >  
> > +	if (excl) {
> > +		if (pdata->open_count)
> > +			return -EBUSY;
> > +
> > +		pdata->excl_open = true;
> > +	}
> > +	pdata->open_count++;
> > +
> >  	return 0;
> >  }
> >  
> >  static inline
> >  void dfl_feature_dev_use_end(struct dfl_feature_platform_data *pdata)
> >  {
> > -	clear_bit_unlock(DEV_STATUS_IN_USE, &pdata->dev_status);
> > +	pdata->excl_open = false;
> > +
> > +	if (WARN_ON(pdata->open_count <= 0))
> > +		return;
> > +
> > +	pdata->open_count--;
> > +}
> > +
> > +static inline
> Feel free to drop the inline here. Your compiler will figure that out by
> iteself.

Yes I'll change it.

> > +int dfl_feature_dev_use_count(struct dfl_feature_platform_data *pdata)
> > +{
> > +	return pdata->open_count;
> >  }
> >  
> >  static inline
> > -- 
> > 2.7.4
> > 
> 
> Again, sorry for the delay,
> 
> Moritz

Thanks for take your time out :)

Yilun