Re: [dpdk-dev] [PATCH] vhost: Fix `struct file' leakage in `eventfd_link'

["Xie, Huawei" <huawei.xie@intel.com>]

On 3/23/2015 11:27 PM, Pavel Boldin wrote:


On Mon, Mar 23, 2015 at 5:16 PM, Xie, Huawei <huawei.xie@intel.com<mailto:huawei.xie@intel.com>> wrote:
On 3/23/2015 10:52 PM, Pavel Boldin wrote:


On Mon, Mar 23, 2015 at 4:41 PM, Xie, Huawei <huawei.xie@intel.com<mailto:huawei.xie@intel.com><mailto:huawei.xie@intel.com<mailto:huawei.xie@intel.com>>> wrote:
On 3/23/2015 10:37 PM, Pavel Boldin wrote:


On Mon, Mar 23, 2015 at 4:21 PM, Xie, Huawei <huawei.xie@intel.com<mailto:huawei.xie@intel.com><mailto:huawei.xie@intel.com<mailto:huawei.xie@intel.com>><mailto:huawei.xie@intel.com<mailto:huawei.xie@intel.com><mailto:huawei.xie@intel.com<mailto:huawei.xie@intel.com>>>> wrote:
On 3/23/2015 8:54 PM, Pavel Boldin wrote:
> Due to increased `struct file's reference counter subsequent call
> to `filp_close' does not free the `struct file'. Prepend `fput' call
> to decrease the reference counter.
>
> Signed-off-by: Pavel Boldin <pboldin@mirantis.com<mailto:pboldin@mirantis.com><mailto:pboldin@mirantis.com<mailto:pboldin@mirantis.com>><mailto:pboldin@mirantis.com<mailto:pboldin@mirantis.com><mailto:pboldin@mirantis.com<mailto:pboldin@mirantis.com>>>>
> ---
>  lib/librte_vhost/eventfd_link/eventfd_link.c | 1 +
>  1 file changed, 1 insertion(+)
>
> diff --git a/lib/librte_vhost/eventfd_link/eventfd_link.c b/lib/librte_vhost/eventfd_link/eventfd_link.c
> index 7755dd6..62c45c8 100644
> --- a/lib/librte_vhost/eventfd_link/eventfd_link.c
> +++ b/lib/librte_vhost/eventfd_link/eventfd_link.c
> @@ -117,6 +117,7 @@ eventfd_link_ioctl(struct file *f, unsigned int ioctl, unsigned long arg)
>                * Release the existing eventfd in the source process
>                */
>               spin_lock(&files->file_lock);
> +             fput(file);
Could we just call atomic_long_dec here?

We can but I don't like breaking encapsulation (which is broken anyway by the code). So, there is a special method and we should use it in my opinion.
it is increased by atomic_long_inc_not_zero so why don't we use the symmetric function?
The code with `atomic_long_inc_not_zero' call is a copy-paste of the `fget' function. If we want to make it clear we should make a separate function and name it so: `fget_from_files'.

I don't understand why there is a (exact?) copy&paste of fget there. :).
Maybe you could post a patchset, first replace the copy/paste with fget and then this patch. It will looks much clearer.
The code of this module received little to none review and requires some love at the moment.

I wanted to refactor the module completely but Thomas said it is not going to go into the 2.0. So I decided to make a simple one-line fix.
Another isse is do we really need a src fd here? Could we just allocate a unsed fd in the kernel and installed it with the target eventfd.

If you are interested this [0] is the latest version of the refactoring patch.

I can provide you with an application that checks that there is indeed no leakage and ensures that the `eventfd' moving works. It is being used in our builds as a test [1]. The code is "heredoc"ed in [2]

[0] http://dpdk.org/dev/patchwork/patch/4113/
[1] https://review.fuel-infra.org/#/c/4639/
[2] https://review.fuel-infra.org/#/c/4639/3/tests/runtests.sh

Pavel

Second thing is: another thread of the same processor can call the `sys_close' on the `fd' and this will dereference counter so `fput' will correctly free the `struct file'. Using `atomic_long_dec' will leak a `struct file' and print a KERN_ERR message by `filp_close'.

So, the common thing is to use appropriate functions and don't reinvent the wheel.

Pavel



Pavel

>               filp_close(file, files);
>               fdt = files_fdtable(files);
>               fdt->fd[eventfd_copy.source_fd] = NULL;