[PATCH v3] tpm: add support for partial reads

[Jarkko Sakkinen]

On Fri, Nov 16, 2018 at 01:31:58AM +0200, Jarkko Sakkinen wrote:
> On Thu, Nov 15, 2018 at 12:42:22PM -0800, Tadeusz Struk wrote:
> > Currently to read a response from the TPM device an application needs
> > provide big enough buffer for the whole response and read it in one go.
> > The application doesn't know how big the response it beforehand so it
> > always needs to maintain a 4K buffer and read the max (4K).
> > In case if the user of the TSS library doesn't provide big enough
> > buffer the TCTI spec says that the library should set the required
> > size and return TSS2_TCTI_RC_INSUFFICIENT_BUFFER error code so that the
> > application could allocate a bigger buffer and call receive again.
> > To make it possible in the TSS library, this requires being able to do
> > partial reads from the driver.
> > The library would read the 10 bytes header first to get the actual size
> > of the response from the header, and then read the rest of the response.
> > 
> > This patch adds support for partial reads, i.e. the user can read the
> > response in one or multiple reads, until the whole response is consumed.
> > The user can also read only part of the response and ignore
> > the rest by issuing a new write to send a new command.
> > 
> > Signed-off-by: Tadeusz Struk <tadeusz.struk at intel.com>
> > ---
> > The usecase is implemented in this TSS commit:
> > https://github.com/tpm2-software/tpm2-tss/commit/ce982f67a67dc08e24683d30b05800648d8a264c
> > 
> > Changes in v3:
> >  - Remove link to usecase implemented in TSS out of the commit message.
> >  - Update the conddition in tpm_common_poll() to take into account
> >    the partial_data also.
> > 
> > Changes in v2:
> >  - Allow writes after only partial response is consumed to maintain
> >    backwords compatibility.
> > ---
> >  drivers/char/tpm/tpm-dev-common.c |   41 ++++++++++++++++++++++++++++---------
> >  drivers/char/tpm/tpm-dev.h        |    2 ++
> >  2 files changed, 33 insertions(+), 10 deletions(-)
> > 
> > diff --git a/drivers/char/tpm/tpm-dev-common.c b/drivers/char/tpm/tpm-dev-common.c
> > index 99b5133a9d05..5d43b0c28565 100644
> > --- a/drivers/char/tpm/tpm-dev-common.c
> > +++ b/drivers/char/tpm/tpm-dev-common.c
> > @@ -64,6 +64,7 @@ static void tpm_timeout_work(struct work_struct *work)
> >  
> >  	mutex_lock(&priv->buffer_mutex);
> >  	priv->data_pending = 0;
> > +	priv->partial_data = 0;
> >  	memset(priv->data_buffer, 0, sizeof(priv->data_buffer));
> >  	mutex_unlock(&priv->buffer_mutex);
> >  	wake_up_interruptible(&priv->async_wait);
> > @@ -90,22 +91,39 @@ ssize_t tpm_common_read(struct file *file, char __user *buf,
> >  	ssize_t ret_size = 0;
> >  	int rc;
> >  
> > -	del_singleshot_timer_sync(&priv->user_read_timer);
> > -	flush_work(&priv->timeout_work);
> >  	mutex_lock(&priv->buffer_mutex);
> > +	if (priv->data_pending || priv->partial_data) {
> > +		if (*off == 0)
> > +			priv->partial_data = priv->data_pending;
> > +
> > +		ret_size = min_t(ssize_t, size, priv->partial_data);
> > +		if (ret_size <= 0) {
> > +			ret_size = 0;
> > +			priv->data_pending = 0;
> > +			priv->partial_data = 0;
> > +			goto out;
> > +		}
> >  
> > -	if (priv->data_pending) {
> > -		ret_size = min_t(ssize_t, size, priv->data_pending);
> > -		if (ret_size > 0) {
> > -			rc = copy_to_user(buf, priv->data_buffer, ret_size);
> > -			memset(priv->data_buffer, 0, priv->data_pending);
> > -			if (rc)
> > -				ret_size = -EFAULT;
> > +		rc = copy_to_user(buf, priv->data_buffer + *off, ret_size);
> > +		if (rc) {
> > +			memset(priv->data_buffer, 0, TPM_BUFSIZE);
> > +			priv->partial_data = 0;
> > +			ret_size = -EFAULT;
> > +		} else {
> > +			memset(priv->data_buffer + *off, 0, ret_size);
> > +			priv->partial_data -= ret_size;
> > +			*off += ret_size;
> 
> You could drop these memset() calls and also one from
> tpm_timeout_work(). The call could be done once in the beginning of
> tpm_common_write() instead of having three different call sites.
> 
> Replacing two memsets with one could be a prepending commit.
> 
> >  		}
> >  
> >  		priv->data_pending = 0;
> >  	}
> >  
> > +out:
> > +	if (!priv->partial_data) {
> > +		*off = 0;
> > +		del_singleshot_timer_sync(&priv->user_read_timer);
> > +		flush_work(&priv->timeout_work);
> > +	}
> >  	mutex_unlock(&priv->buffer_mutex);
> >  	return ret_size;
> >  }
> > @@ -150,6 +168,9 @@ ssize_t tpm_common_write(struct file *file, const char __user *buf,
> >  		goto out;
> >  	}
> >  
> > +	priv->partial_data = 0;
> > +	*off = 0;
> > +
> >  	/*
> >  	 * If in nonblocking mode schedule an async job to send
> >  	 * the command return the size.
> > @@ -184,7 +205,7 @@ __poll_t tpm_common_poll(struct file *file, poll_table *wait)
> >  
> >  	poll_wait(file, &priv->async_wait, wait);
> >  
> > -	if (priv->data_pending)
> > +	if (priv->data_pending || priv->partial_data)
> >  		mask = EPOLLIN | EPOLLRDNORM;
> >  	else
> >  		mask = EPOLLOUT | EPOLLWRNORM;
> > diff --git a/drivers/char/tpm/tpm-dev.h b/drivers/char/tpm/tpm-dev.h
> > index a126b575cb8c..a2ca6a7a06f1 100644
> > --- a/drivers/char/tpm/tpm-dev.h
> > +++ b/drivers/char/tpm/tpm-dev.h
> > @@ -11,6 +11,8 @@ struct file_priv {
> >  
> >  	/* Holds the amount of data passed or an error code from async op */
> >  	ssize_t data_pending;
> > +	/* For partial reads, holds the reminder of a response */
> > +	ssize_t partial_data;
> 
> Naming becomes a mess and the comment for data_pending variable is
> incorrect as it is also used for synchronous operation.
> 
> Maybe add a prepending commit to rename it as
> 
> 	/* Holds the resul of the tpm_transmit() last call. */
> 	ssize_t transmit_result;
> 
> That is at least clear and obvious on what it contains.
> 
> The comment for partial_data is incorrect as the variable does not
> contain any data.
> 
> We could use declare:
> 
> 	/* Holds the count how much of the response is still unread. */
> 	size_t response_pending;
> 
> Observe another remark from your commit: there is no reaso to ssize_t as
> the type as the value should never be a negative number.

In order to correct amazing amount of typos in the last sentence I meant
to say that there is no reason to use ssize_t as response_pending can
hold only positive values.

/Jarkko