The poll condition should only check response_length, because reads should only be issued if there is data to read. The response_read flag only prevents double writes. The problem was that the write set the response_read to false, enqued a tpm job, and returned. Then application called poll which checked the response_read flag and returned EPOLLIN. Then the application called read, but got nothing. After all that the async_work kicked in. Added also mutex_lock around the poll check to prevent other possible race conditions.
Cc: stable@vger.kernel.org Fixes: 9488585b21bef0df12 ("tpm: add support for partial reads") Reported-by: Mantas Mikulėnas grawity@gmail.com Tested-by: Mantas Mikulėnas grawity@gmail.com Signed-off-by: Tadeusz Struk tadeusz.struk@intel.com --- drivers/char/tpm/tpm-dev-common.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/drivers/char/tpm/tpm-dev-common.c b/drivers/char/tpm/tpm-dev-common.c index 5eecad233ea1..744b0237300a 100644 --- a/drivers/char/tpm/tpm-dev-common.c +++ b/drivers/char/tpm/tpm-dev-common.c @@ -203,12 +203,19 @@ __poll_t tpm_common_poll(struct file *file, poll_table *wait) __poll_t mask = 0;
poll_wait(file, &priv->async_wait, wait); + mutex_lock(&priv->buffer_mutex);
- if (!priv->response_read || priv->response_length) + /* + * The response_length indicates if there is still response + * (or part of it) to be consumed. Partial reads decrease it + * by the number of bytes read, and write resets it the zero. + */ + if (priv->response_length) mask = EPOLLIN | EPOLLRDNORM; else mask = EPOLLOUT | EPOLLWRNORM;
+ mutex_unlock(&priv->buffer_mutex); return mask; }