[PATCH] net: eth-uclass: change state before stop() in eth_halt()

[Marek Vasut]

On 12/5/22 13:06, Niel Fourie wrote:

Hi,

[...]

>>>> How come nobody triggered this problem with regular ethernet in 
>>>> U-Boot ?
>>>>
>>>> If this is isolated to USB gadget ethernet, then please do not hack 
>>>> around this in core networking code, but rather fix the USB ethernet 
>>>> gadget itself. It seems that gadget code should not unregister the 
>>>> gadget in drivers/usb/gadget/ether.c _usb_eth_halt() , at least not 
>>>> fully.
>>>
>>> The reason is simple, the regular ethernet drivers do not get removed 
>>> on stop() like for the gadget ethernet driver, and in their case 
>>> `priv` is still valid.
>>
>> The suggestion for a proper fix is in the last paragraph above -- do 
>> not unregister the usb ethernet gadget device in halt(), keep the 
>> gadget device registered. Then the priv pointer would be valid. (*)
>>
>>> I agree on your point that reworking the ethernet gadget code would 
>>> be preferable, but this would be a much bigger effort (and if I were 
>>> to do it, I would probably introduce even more bugs). I am not 
>>> certain whether this would not also affect the non-DM gadget 
>>> implementation as well, which still contain drivers like ci_udc.c 
>>> which does not appear to have been ported to DM gadget yet? (I only 
>>> see DM USB there...)
>>>
>>> That said, I am not certain whether this is not also bug, as I am not 
>>> certain whether the assumption that `priv` should be available after 
>>> stop() is valid or not.
>>>
>>> The documentation at 
>>> https://source.denx.de/u-boot/u-boot/-/blob/master/doc/develop/driver-model/ethernet.rst?plain=1#L135 states:
>>>
>>> The **stop** function should turn off / disable the hardware and 
>>> place it back in its reset state.  It can be called at any time 
>>> (before any call to the related start() function), so make sure it 
>>> can handle this sort of thing.
>>
>> The ->stop callback is supposed to stop the interface, turn off its 
>> DMA, but NOT deallocate the device (and its associated data) behind it.
>>
>>> In such a complete reset state I am not certain whether the 
>>> assumption that `priv` should exist is still valid, at least not 
>>> without another call to dev_get_uclass_priv() and revalidating it first?
>>
>> In case a device is probe()d, its private data are also allocated and 
>> available, so yes, 'priv' pointer should still be valid.
> 
> Granted, the usb gadget driver implementation is problematic, and it 
> definitely belongs on the TODO list.
> 
> That being said, priv not being valid at this stage is not a new 
> problem, as validation for it after stop() was explicitly added in 
> commit c3211708 ("net: eth-uclass: Fix for DM USB ethernet support") for 
> this reason 4 years ago. Fortunately in that case, it just fixed a null 
> pointer de-reference, not corruption of freed memory.

I believe the aforementioned patch is already also wrong, since the priv 
pointer would be invalid for gadget ethernet driver got removed in the 
stop callback.

>>> Lastly, this assumption that priv is still valid is rather new and it 
>>> was introduced here:
>>>
>>> https://source.denx.de/u-boot/u-boot/-/commit/fa795f452541ce07b33be603de36cac3c5d7dfcf
>>
>> I disagree, the device private data are valid during the entire 
>> lifespan of the device. That assumption has been baked into the driver 
>> model itself and far predates that commit.
>>
>> In case of a usb ethernet, the lifespan of the device starts with the 
>> 'bind' command which triggers the ->bind callback, and first use which 
>> triggers its ->probe callback. The lifespan ends with 'unbind' command 
>> or OS boot, which triggers ->remove callback and ->unbind callbacks.
>>
>>> This commit appears to be a workaround for drivers which cannot deal 
>>> with stop() being called at any time as required in the above quoted 
>>> documentation.
>>
>> This commit prevents network device ->start() from being called 
>> multiple times, which is a valid precaution, as calling start while 
>> the interface is already up would interfere with existing connection 
>> (e.g. the netconsole as mentioned in the commit message). That does 
>> not seem to be a workaround to me.
> 
> Ah yes, the commit message indeed states that some drivers (like 
> fec_mxc) cannot deal with multiple start() calls, not multiple stop() 
> calls. My mistake, good point.
> 
> But preventing multiple start()s getting called is not what the code 
> change in this commit is actually doing. It is instead inhibiting 
> calling stop() on devices for which priv is not valid or for which 
> priv->running is false. Therefore it is avoiding calling stop() on 
> un-started devices.

I wonder if such a patch is needed to fix it ?

diff --git a/net/eth-uclass.c b/net/eth-uclass.c
index f41da4b37b3..e7b5d3943e8 100644
--- a/net/eth-uclass.c
+++ b/net/eth-uclass.c
@@ -298,7 +298,7 @@ int eth_init(void)
                 if (current) {
                         debug("Trying %s\n", current->name);

-                       if (device_active(current)) {
+                       if (!eth_is_active(current)) {

>>> I would consider adding a workaround to a workaround in this case to 
>>> be the lesser evil, as tracking down this bug in the first place was 
>>> like looking for a needle in a haystack. This change would at least 
>>> save everybody else from strange crashes in particular configurations 
>>> without any negative impact. But this is fortunately not my decision. :)
>>
>> Commit fa795f45254 ("net: eth-uclass: avoid running start() twice 
>> without stop()") is as far as I can tell unrelated to this change and 
>> does not seem to me like a workaround.
> 
> It does introduce writing to priv after stop() is called without 
> (re-)validating priv first, though.
> 
> The code previously first called stop(), then called 
> dev_get_uclass_priv() to get priv and then validated priv before writing 
> to it, which avoids exactly the problem of priv no longer being valid.
> 
> As mentioned above, this was explicitly addressed by commit c3211708 
> ("net: eth-uclass: Fix for DM USB ethernet support").

I'm afraid this is just piling workarounds on top of workarounds.

>> It does however show that this patch introduces a bug -- this patch 
>> changes the order in which priv->state = ETH_STATE_PASSIVE; is 
>> assigned from _after_ the ->stop callback to _before_ the -> stop 
>> callback. This breaks drivers/net/ldpaa_eth/ldpaa_eth.c which checks 
>> the priv->state in its ->stop callback, either on its own in non-DM 
>> case, or in eth_is_active() implementation in DM case. With this 
>> patch, the interface would never be stopped in the ->stop callback, 
>> because the condition (net_dev->state == ETH_STATE_PASSIVE) test in 
>> the ldpaa stop callback implementation would always be true.
>>
> 
> In drivers/net/ldpaa_eth/ldpaa_eth.c:ldpaa_eth_stop(), priv is of type
> struct ldpaa_eth_priv*, defined in drivers/net/ldpaa_eth/ldpaa_eth.h and 
> is accessed using dev_get_priv().
> 
> In net/eth-uclass.c:eht_halt(), priv is of type struct eth_device_priv* 
> and defined in the same .c file, and is accessed using 
> dev_get_uclass_priv(). As the structure is local to this file, nothing 
> outside of this file should have any knowledge of its contents, and 
> changing of the order of the calls should only impact this file.

As you already found out yourself in subsequent email:

  655 static void ldpaa_eth_stop(struct udevice *dev)
  656 {
  657         struct ldpaa_eth_priv *priv = dev_get_priv(dev);
...
  666 #ifdef CONFIG_DM_ETH
  667         if (!eth_is_active(dev))
...

The eth_is_active():

350 int eth_is_active(struct udevice *dev)
351 {
352         struct eth_device_priv *priv;
353
354         if (!dev || !device_active(dev))
355                 return 0;
356
357         priv = dev_get_uclass_priv(dev);
358         return priv->state == ETH_STATE_ACTIVE;
359 }

does access dev_get_uclass_priv(dev)->priv .

> I sincerely hope that these two are not interfering with each other, 
> otherwise we have much bigger problems...
> 
>> The proper fix is in the usb ethernet gadget code, see (*) above, 
>> let's not pile workarounds onto already difficult to maintain code.
>>
>> [...]
> Agreed. The problem is unfortunately just that I don't see that 
> happening simply to fix this bug.

Please, do not pile workarounds on top of workarounds, there is too many 
of those already as you found out above. Each one just slaps something 
on, which later on breaks and causes pain a few years down the line. 
This is not manageable in the long run.

The proper fix I believe is (*) above and it should not be hard to 
implement that one.