Re: [PATCH v9 4/4] bus: mhi: Add userspace client interface driver

[Hemant Kumar <hemantk@codeaurora.org>]

Hi Loic,

On 10/26/20 10:34 AM, Loic Poulain wrote:
> Hi Hemant,
> 
> That looks better IMHO, just small comments on locking.
> 
[..]
>     +static ssize_t mhi_uci_write(struct file *file,
>     +                            const char __user *buf,
>     +                            size_t count,
>     +                            loff_t *offp)
>     +{
>     +       struct uci_dev *udev = file->private_data;
>     +       struct mhi_device *mhi_dev = udev->mhi_dev;
>     +       struct device *dev = &mhi_dev->dev;
>     +       struct uci_chan *uchan = udev->uchan;
>     +       size_t bytes_xfered = 0;
>     +       int ret, nr_avail = 0;
>     +
>     +       /* if ul channel is not supported return error */
>     +       if (!buf || !count || !mhi_dev->ul_chan)
>     +               return -EINVAL;
>     +
>     +       dev_dbg(dev, "%s: to xfer: %zu bytes\n", __func__, count);
>     +
>     +       mutex_lock(&uchan->write_lock);
> 
> 
> Maybe mutex_lock_interruptible is more appropriate here (same in read fops).
i agree, will return -EINTR if mutex_lock_interruptible returns < 0.
> 
[..]
>     +static ssize_t mhi_uci_read(struct file *file,
>     +                           char __user *buf,
>     +                           size_t count,
>     +                           loff_t *ppos)
>     +{
>     +       struct uci_dev *udev = file->private_data;
>     +       struct mhi_device *mhi_dev = udev->mhi_dev;
>     +       struct uci_chan *uchan = udev->uchan;
>     +       struct device *dev = &mhi_dev->dev;
>     +       struct uci_buf *ubuf;
>     +       size_t rx_buf_size;
>     +       char *ptr;
>     +       size_t to_copy;
>     +       int ret = 0;
>     +
>     +       /* if dl channel is not supported return error */
>     +       if (!buf || !mhi_dev->dl_chan)
>     +               return -EINVAL;
>     +
>     +       mutex_lock(&uchan->read_lock);
>     +       spin_lock_bh(&uchan->dl_pending_lock);
>     +       /* No data available to read, wait */
>     +       if (!uchan->cur_buf && list_empty(&uchan->dl_pending)) {
>     +               dev_dbg(dev, "No data available to read, waiting\n");
>     +
>     +               spin_unlock_bh(&uchan->dl_pending_lock);
>     +               ret = wait_event_interruptible(uchan->dl_wq,
>     +                                              (!udev->enabled ||
>     +                                           
>       !list_empty(&uchan->dl_pending)));
> 
> 
> If you need to protect dl_pending list against concurent access, you 
> need to do it in wait_event as well. I would suggest to lookg at 
> `wait_event_interruptible_lock_irq` function, that allows to pass a 
> locked spinlock as parameter. That would be safer and prevent this 
> lock/unlock dance.
When using this API difference is, first we take spin_lock_bh() and then 
wait API is using spin_unlock_irq()/spin_lock_irq(). I am getting
"BUG: scheduling while atomic" when i use this way. When i changed 
spin_lock_bh to spin_lock_irq then we got rid of the kernel BUG.

Thanks,
Hemant

-- 
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project