[V1,net,1/3] net: ena: Prevent reset after device destruction

The reset work is scheduled by the timer routine whenever it
detects that a device reset is required (e.g. when a keep_alive signal
is missing).
When releasing device resources in ena_destroy() the driver cancels the
scheduling of the timer routine without destroying the reset
work explicitly.

This creates the following bug:
    The driver is suspended and the ena_suspend() function is called
	-> This function calls ena_destroy() to free the net device
	   resources
	    -> The driver waits for the timer routine to finish
	    its execution and then cancels it, thus preventing from it
	    to be called again.

    If, in its final execution, the timer routine schedules a reset,
    the reset routine might be called after the device resources are
    freed, which might cause a kernel panic.

By changing the reset routine so that it cannot run simultaneously with
the destruction routine, we allow the reset routine read the device's
state accurately.
This is achieved by checking whether ENA_FLAG_TRIGGER_RESET flag is set
before resetting the device and making both the destruction function and
the flag check are under rtnl lock.
The ENA_FLAG_TRIGGER_RESET is cleared at the end of the destruction
routine. Also surround the flag check with 'likely' because
we expect that the reset routine would be called only when
ENA_FLAG_TRIGGER_RESET flag is set.

This patch also removes the destruction of the timer and reset services
from ena_remove() since the timer is destroyed by the destruction
routine and the reset work is handled by this patch.

Fixes: 8c5c7abdeb2d ("net: ena: add power management ops to the ENA driver")
Signed-off-by: Shay Agroskin <shayagr@amazon.com>
---
 drivers/net/ethernet/amazon/ena/ena_netdev.c | 17 ++++++-----------
 1 file changed, 6 insertions(+), 11 deletions(-)

On Wed, 12 Aug 2020 13:10:57 +0300 Shay Agroskin wrote:
> This patch also removes the destruction of the timer and reset services
> from ena_remove() since the timer is destroyed by the destruction
> routine and the reset work is handled by this patch.

You'd still have a use after free if the work runs after the device is
removed. I think cancel_work_sync() gotta stay.

Jakub Kicinski <kuba@kernel.org> writes:

> On Wed, 12 Aug 2020 13:10:57 +0300 Shay Agroskin wrote:
>> This patch also removes the destruction of the timer and reset 
>> services
>> from ena_remove() since the timer is destroyed by the 
>> destruction
>> routine and the reset work is handled by this patch.
>
> You'd still have a use after free if the work runs after the 
> device is
> removed. I think cancel_work_sync() gotta stay.

Hi, thank you for reviewing the patch. Short answer: I verified 
that the ENA_FLAG_TRIGGER_RESET flag cannot be set after 
ena_destroy_device() finishes its execution.

Long answer:
The ena_destroy_device() function is called with rtnl_lock() held, 
so it cannot run in parallel with the reset function. Also the 
destroy function clears the bit ENA_FLAG_TRIGGER_RESET without 
which the reset function just exits without doing anything.

A problem can then only happen when some routine sets the 
ENA_FLAG_TRIGGER_RESET bit before the reset function is executed, 
the following describes all functions from which this bit can be 
set:

- check_* functions: these function are called from the timer 
  routine which is destroyed in ena_destroy_device(), so by the 
  time the rtnl_lock() released, the bit is cleared

- napi related functions (io_poll, xdp_io_poll, validate_rx_req_id 
  etc.): the napi is de-registered in ena_destroy_device(), so 
  none of these functions is called after destroying the device.

- xmit functions (ena_xmit_common, ena_tx_timeout): the device is 
  brought down and all its RX/TX resources are freed before 
  releasing the lock.

These are all the occurrences I found. Without this bit set, the 
reset function would fail the 'if' check in this patch, and exit 
without doing anything. Destroying the reset function explicitly 
won't help since by the time we do it, the function can already be 
executed.

On Thu, 13 Aug 2020 15:51:46 +0300 Shay Agroskin wrote:
> Long answer:
> The ena_destroy_device() function is called with rtnl_lock() held, 
> so it cannot run in parallel with the reset function. Also the 
> destroy function clears the bit ENA_FLAG_TRIGGER_RESET without 
> which the reset function just exits without doing anything.
> 
> A problem can then only happen when some routine sets the 
> ENA_FLAG_TRIGGER_RESET bit before the reset function is executed, 
> the following describes all functions from which this bit can be 
> set:

ena_fw_reset_device() runs from a workqueue, it can be preempted right
before it tries to take the rtnl_lock. Then after arbitrarily long
delay it will start again, take the lock, and dereference
adapter->flags. But adapter could have been long freed at this point.

Unless you flush a workqueue or cancel_work_sync() you can never be
sure it's not scheduled. And I can only see a flush when module is
unloaded now.

Jakub Kicinski <kuba@kernel.org> writes:

> On Thu, 13 Aug 2020 15:51:46 +0300 Shay Agroskin wrote:
>> Long answer:
>> The ena_destroy_device() function is called with rtnl_lock() 
>> held, 
>> so it cannot run in parallel with the reset function. Also the 
>> destroy function clears the bit ENA_FLAG_TRIGGER_RESET without 
>> which the reset function just exits without doing anything.
>> 
>> A problem can then only happen when some routine sets the 
>> ENA_FLAG_TRIGGER_RESET bit before the reset function is 
>> executed, 
>> the following describes all functions from which this bit can 
>> be 
>> set:
>
> ena_fw_reset_device() runs from a workqueue, it can be preempted 
> right
> before it tries to take the rtnl_lock. Then after arbitrarily 
> long
> delay it will start again, take the lock, and dereference
> adapter->flags. But adapter could have been long freed at this 
> point.

Missed that the check for the 'flags' field also requires that 
netdev_priv field (adapter variable) would be allocated. Thank you 
for pointing that out, this indeed needs to be fixed. I'll add 
reset work destruction in next patchset.

Thank you for reviewing it
>
> Unless you flush a workqueue or cancel_work_sync() you can never 
> be
> sure it's not scheduled. And I can only see a flush when module 
> is
> unloaded now.