Re: [PATCH 2/3] serial: 8250_mxpcie: fail the probe when a port cannot be registered

在 2026/8/18 21:19, Andy Shevchenko 写道:
> On Tue, Aug 18, 2026 at 05:39:17PM +0800, Linmao Li wrote:
>> When serial8250_register_8250_port() fails the loop only breaks out and
>> the probe still returns success.  The entries it did not reach keep the
>> zero devm_kzalloc() left there, and the entry that failed keeps a
>> negative error code.
>>
>> mxpcie8250_remove() then feeds all of them to
>> serial8250_unregister_port(), which checks neither: line 0 unregisters a
>> port this driver does not own, and the negative line indexes
>> serial8250_ports[] out of bounds.
>>
>> Fail the probe instead, unregistering the ports registered so far.
> This is behavioural difference. Before it was "register as much ports as we
> can", now "don't register even a single one". If one of those UARTs worked in
> the previous case for whatever reason, this patch will break that.
You are right, and 8250_pci.c - which this driver was split out of - is
explicit about it: pciserial_init_ports() breaks out of the loop on the
first failure, but returns the ports it did register and records how many
in priv->nr, which is all pciserial_detach_ports() then walks.

So what the split lost is that counter, not the "keep what works"
behaviour. mxpcie8250_remove() walks all num_ports entries, so it hands
serial8250_unregister_port() a line 0 for every entry the loop never
reached - a port this driver does not own - and a negative line for the
one that failed.

v2 keeps the partial registration and restores the counter only.


Thanks,
Linmao

>