Re: [RFC PATCH 4/4] tpm: Require O_EXCL for exclusive /dev/tpm access

[Jarkko Sakkinen <jarkko@kernel.org>]

On Tue, Sep 02, 2025 at 06:27:17PM +0100, Jonathan McDowell wrote:
> From: Jonathan McDowell <noodles@meta.com>
> 
> Given that /dev/tpm has not had exclusive access to the TPM since the
> existence of the kernel resource broker and other internal users, stop
> defaulted to exclusive access to the first client that opens the device.
> Continue to support exclusive access, but only with the use of the
> O_EXCL flag on device open.
> 
> Signed-off-by: Jonathan McDowell <noodles@meta.com>
> ---
>  drivers/char/tpm/tpm-dev.c | 25 +++++++++++++++++++------
>  drivers/char/tpm/tpm-dev.h |  1 +
>  2 files changed, 20 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/char/tpm/tpm-dev.c b/drivers/char/tpm/tpm-dev.c
> index 80c4b3f3ad18..8921bbb541c1 100644
> --- a/drivers/char/tpm/tpm-dev.c
> +++ b/drivers/char/tpm/tpm-dev.c
> @@ -19,15 +19,21 @@ static int tpm_open(struct inode *inode, struct file *file)
>  {
>  	struct tpm_chip *chip;
>  	struct file_priv *priv;
> +	int rc;
>  
>  	chip = container_of(inode->i_cdev, struct tpm_chip, cdev);
>  
>  	/*
> -	 * Only one client is allowed to have /dev/tpm0 open at a time, so we
> -	 * treat it as a write lock. The shared /dev/tpmrm0 is treated as a
> -	 * read lock.
> +	 * If a client uses the O_EXCL flag then it expects to be the only TPM
> +	 * user, so we treat it as a write lock. Otherwise we do as /dev/tpmrm
> +	 * and use a read lock.
>  	 */
> -	if (!down_write_trylock(&chip->open_lock)) {
> +	if (file->f_flags & O_EXCL)
> +		rc = down_write_trylock(&chip->open_lock);
> +	else
> +		rc = down_read_trylock(&chip->open_lock);
> +
> +	if (!rc) {
>  		dev_dbg(&chip->dev, "Another process owns this TPM\n");
>  		return -EBUSY;
>  	}
> @@ -35,13 +41,17 @@ static int tpm_open(struct inode *inode, struct file *file)
>  	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
>  	if (priv == NULL)
>  		goto out;
> +	priv->exclusive = (file->f_flags & O_EXCL);
>  
>  	tpm_common_open(file, chip, priv, NULL);
>  
>  	return 0;
>  
>   out:
> -	up_write(&chip->open_lock);
> +	if (file->f_flags & O_EXCL)
> +		up_write(&chip->open_lock);
> +	else
> +		up_read(&chip->open_lock);
>  	return -ENOMEM;
>  }
>  
> @@ -53,7 +63,10 @@ static int tpm_release(struct inode *inode, struct file *file)
>  	struct file_priv *priv = file->private_data;
>  
>  	tpm_common_release(file, priv);
> -	up_write(&priv->chip->open_lock);
> +	if (priv->exclusive)
> +		up_write(&priv->chip->open_lock);
> +	else
> +		up_read(&priv->chip->open_lock);
>  	kfree(priv);
>  
>  	return 0;
> diff --git a/drivers/char/tpm/tpm-dev.h b/drivers/char/tpm/tpm-dev.h
> index f3742bcc73e3..0ad8504c73e4 100644
> --- a/drivers/char/tpm/tpm-dev.h
> +++ b/drivers/char/tpm/tpm-dev.h
> @@ -17,6 +17,7 @@ struct file_priv {
>  	ssize_t response_length;
>  	bool response_read;
>  	bool command_enqueued;
> +	bool exclusive;
>  
>  	u8 data_buffer[TPM_BUFSIZE];
>  };
> -- 
> 2.51.0
> 

I'll hold with testing to +1 version but overall patch set looks good.

BR, Jarkko