mbox series

[RESEND,0/2] tty: serial: add and use a managed variant of uart_add_one_port()

Message ID 20221229161948.594102-1-brgl@bgdev.pl
Headers
Series tty: serial: add and use a managed variant of uart_add_one_port() |

Message

Bartosz Golaszewski Dec. 29, 2022, 4:19 p.m. UTC 
  From: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>

Resending rebased on top of v6.2-rc1

--

This series adds a managed variant of uart_add_one_port() and uses it in the
qcom-geni-serial driver.

I've been asked by Greg to send it separately and he didn't seem to be
impressed by the proposition of adding devres interfaces to the tty layer
in general. I can only assume it has something to do with the ongoing
discussion about the supposed danger of using devres interfaces in conjunction
with exporting character devices to user-space.

The bug in question can be triggered by opening a device file, unbinding the
driver that exported it and then calling any of the system calls on the
associated file descriptor.

After some testing I noticed that many subsystems are indeed either crashing
or deadlocking in the above situation. I've sent patches that attempt to fix
the GPIO and I2C subsystems[1][2]. Neither of these issues have anything to
do with devres and all to do with the fact that certain resources are freed
on driver unbind and others need to live for as long as the character device
exists. More details on that in the cover letters and commit messages in the
links.

I'd like to point out that the serial code is immune to this issue as before
every operation, the serial core takes the port lock and checks the uart
state. If the device no longer exists (when the uart port is removed, the
pointer to uart_port inside uart_state is to NULL), it gracefully returns
-ENODEV to user-space.

Please consider applying the patches in the series as devres is the easiest
way to lessen the burden on driver developers when dealing with complex error
paths and resource leaks. The general rule for devres is: if it can be freed
in .remove() then it can be managed by devres, which is the case for this new
helper.

Bart

[1] https://lkml.org/lkml/2022/12/8/826
[2] https://lkml.org/lkml/2022/12/5/414

Bartosz Golaszewski (2):
  tty: serial: provide devm_uart_add_one_port()
  tty: serial: qcom-geni-serial: use devres for uart port management

 .../driver-api/driver-model/devres.rst        |  3 ++
 drivers/tty/serial/qcom_geni_serial.c         |  8 +---
 drivers/tty/serial/serial_core.c              | 48 +++++++++++++++++++
 include/linux/serial_core.h                   |  6 +++
 4 files changed, 58 insertions(+), 7 deletions(-)

Comments

Greg KH Jan. 19, 2023, 2:10 p.m. UTC | #1 
On Thu, Dec 29, 2022 at 05:19:46PM +0100, Bartosz Golaszewski wrote:
> From: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
> 
> Resending rebased on top of v6.2-rc1
> 
> --
> 
> This series adds a managed variant of uart_add_one_port() and uses it in the
> qcom-geni-serial driver.
> 
> I've been asked by Greg to send it separately and he didn't seem to be
> impressed by the proposition of adding devres interfaces to the tty layer
> in general. I can only assume it has something to do with the ongoing
> discussion about the supposed danger of using devres interfaces in conjunction
> with exporting character devices to user-space.

That is correct.

> The bug in question can be triggered by opening a device file, unbinding the
> driver that exported it and then calling any of the system calls on the
> associated file descriptor.
> 
> After some testing I noticed that many subsystems are indeed either crashing
> or deadlocking in the above situation. I've sent patches that attempt to fix
> the GPIO and I2C subsystems[1][2]. Neither of these issues have anything to
> do with devres and all to do with the fact that certain resources are freed
> on driver unbind and others need to live for as long as the character device
> exists. More details on that in the cover letters and commit messages in the
> links.
> 
> I'd like to point out that the serial code is immune to this issue as before
> every operation, the serial core takes the port lock and checks the uart
> state. If the device no longer exists (when the uart port is removed, the
> pointer to uart_port inside uart_state is to NULL), it gracefully returns
> -ENODEV to user-space.
> 
> Please consider applying the patches in the series as devres is the easiest
> way to lessen the burden on driver developers when dealing with complex error
> paths and resource leaks. The general rule for devres is: if it can be freed
> in .remove() then it can be managed by devres, which is the case for this new
> helper.

Overall you are adding more code to the kernel than removing, so how is
this a win?  Perhaps if other drivers were converted over to this new
function then I would be more inclined to be able to accept it.

But as-is, with only one user, it's a non-starter, sorry.

thanks,

greg k-h