Date:   Mon, 23 May 2022 15:28:09 +0200
From:   Luca Ceresoli <luca@...aceresoli.net>
To:     Saravana Kannan <saravanak@...gle.com>
Cc:     Lorenzo Pieralisi <lorenzo.pieralisi@....com>,
        Rob Herring <robh@...nel.org>, PCI <linux-pci@...r.kernel.org>,
        Dan Carpenter <dan.carpenter@...cle.com>,
        linux-omap <linux-omap@...r.kernel.org>,
        linux-arm-kernel <linux-arm-kernel@...ts.infradead.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Kishon Vijay Abraham I <kishon@...com>,
        Krzysztof Wilczyński <kw@...ux.com>,
        Bjorn Helgaas <bhelgaas@...gle.com>,
        Sekhar Nori <nsekhar@...com>,
        Android Kernel Team <kernel-team@...roid.com>
Subject: Re: [PATCH 1/2] PCI: dra7xx: Fix link removal on probe error

Hi Saravana,

On 19/05/22 22:25, Saravana Kannan wrote:
> On Tue, May 17, 2022 at 12:32 AM Luca Ceresoli <luca@...aceresoli.net> wrote:
>>
>> Hi Saravana,
>>
>> On 14/05/22 05:46, Saravana Kannan wrote:
>>> On Thu, May 12, 2022 at 7:07 AM Luca Ceresoli <luca@...aceresoli.net> wrote:
>>>>
>>>> Hi Lorenzo,
>>>>
>>>> On 11/05/22 18:41, Lorenzo Pieralisi wrote:
>>>>> On Sat, Jan 15, 2022 at 10:02:00AM -0600, Rob Herring wrote:
>>>>>> +Saravana
>>>>>>
>>>>>> On Tue, Jan 11, 2022 at 4:35 AM Luca Ceresoli <luca@...aceresoli.net> wrote:
>>>>>>>
>>>>>>> Hi Rob,
>>>>>>>
>>>>>>> On 16/12/21 10:08, Luca Ceresoli wrote:
>>>>>>>> Hi Rob,
>>>>>>>>
>>>>>>>> thanks for the quick feedback!
>>>>>>>>
>>>>>>>> On 14/12/21 23:42, Rob Herring wrote:
>>>>>>>>> On Tue, Dec 14, 2021 at 4:15 PM Luca Ceresoli <luca@...aceresoli.net> wrote:
>>>>>>>>>>
>>>>>>>>>> If a devm_phy_get() calls fails with phy_count==N (N > 0), then N links
>>>>>>>>>> have already been added by device_link_add() and won't be deleted by
>>>>>>>>>> device_link_del() because the code calls 'return' and not 'goto err_link'.
>>>>>>>>>>
>>>>>>>>>> Fix in a very simple way by doing all the devm_phy_get() calls before all
>>>>>>>>>> the device_link_add() calls.
>>>>>>>>>>
>>>>>>>>>> Fixes: 7a4db656a635 ("PCI: dra7xx: Create functional dependency between PCIe and PHY")
>>>>>>>>>> Signed-off-by: Luca Ceresoli <luca@...aceresoli.net>
>>>>>>>>>> ---
>>>>>>>>>>  drivers/pci/controller/dwc/pci-dra7xx.c | 2 ++
>>>>>>>>>>  1 file changed, 2 insertions(+)
>>>>>>>>>>
>>>>>>>>>> diff --git a/drivers/pci/controller/dwc/pci-dra7xx.c b/drivers/pci/controller/dwc/pci-dra7xx.c
>>>>>>>>>> index f7f1490e7beb..2ccc53869e13 100644
>>>>>>>>>> --- a/drivers/pci/controller/dwc/pci-dra7xx.c
>>>>>>>>>> +++ b/drivers/pci/controller/dwc/pci-dra7xx.c
>>>>>>>>>> @@ -757,7 +757,9 @@ static int dra7xx_pcie_probe(struct platform_device *pdev)
>>>>>>>>>>                 phy[i] = devm_phy_get(dev, name);
>>>>>>>>>>                 if (IS_ERR(phy[i]))
>>>>>>>>>>                         return PTR_ERR(phy[i]);
>>>>>>>>>> +       }
>>>>>>>>>>
>>>>>>>>>> +       for (i = 0; i < phy_count; i++) {
>>>>>>>>>>                 link[i] = device_link_add(dev, &phy[i]->dev, DL_FLAG_STATELESS);
>>>>>>>>>
>>>>>>>>> I think this should happen automatically now with fw_devlink being
>>>>>>>>> enabled by default. Can you try?
>>>>>>>>
>>>>>>>> Do you mean removal should be done automatically? I think they are not
>>>>>>>> due to the DL_FLAG_STATELESS flag.
>>>>>>>
>>>>>>> I would love to have feedback because, as said, I think my patch is
>>>>>>> correct, but if I'm wrong (which might well be) I have to drop patch 1
>>>>>>> and rewrite patch 2 in a slightly more complex form.
>>>>>>
>>>>>> I mean that why do you need explicit dependency tracking here when
>>>>>> dependencies on a PHY should happen automatically now. IOW, what is
>>>>>> special about this driver and dependency?
>>>>>
>>>>> Any update on this patch ? I think patch 2 can be merged, please
>>>>> let me know if this one can be dropped.
>>>>
>>>> Thanks for the feedback! You would say yes, you can merge patch 2,
>>>> except it probably does not even apply as it is written in a way that is
>>>> based on the changes in patch 1.
>>>>
>>>> I could rewrite patch 2 to not depend on patch 1 of course, but it
>>>> wouldn't make code simpler, perhaps more complex. And moreover the
>>>> hardware that I used to have access to has phy_count==1 so I could never
>>>> test the failing case, and sadly now I have no access to that hardware.
>>>
>>> Hi Luca,
>>>
>>> The fw_devlink code to create device links from consumers to "phys"
>>> suppliers is pretty well exercised. Most/all Android devices running
>>> 5.10+ kernels (including Pixel 6) use fw_devlink=on to be able to boot
>>> properly.
>>>
>>> So I'd be pretty confident in deleting the device_link_add/del() code
>>> in drivers/pci/controller/dwc/pci-dra7xx.c. The device links should
>>> already be there before the probe is even called.
>>>
>>> Also, if you want to check if the device links (even the 1 phy one you
>>> have) are being created, you can look at /sys/class/devlink to see the
>>> list of all device links that are currently present. You can delete
>>> the code and then use this to check too.
>>
>> Thank you for your feedback. Unfortunately as I said I have no access to
>> the hardware, and won't have anymore. I don't think it is a good idea to
>> send a patch that I cannot test on real hardware, especially since it is
>> for a generic hardware that thus might affect others. But I would be
>> glad to review any such patch that might be sent, FWIW.
> 
> Just to make sure I'm on the same page. I thought you at least had a
> device where phy_count = 1. But looks like you are saying you don't?

I used to have access to a hardware with phy_count = 1 on a former job,
but I don't have it anymore and won't have it since I left that job
position.

> If all you want to check is "phys" have device links created for them
> for whatever random DT device that has a "phys" property, then I can
> test and confirm that for you on whatever platform I have. But if you
> want a test specifically for the device that corresponds to the driver
> you were fixing, then I can't. Let me know.

Honestly, I'm afraid I don't have much time to invest in trying to
recollect all the details and motivations for this patchset. Likely I
spotted this by code inspection while debugging other issues (I had a
non-working PCIe device, but it was not the host fault). If you think
there is little value in these patches, I'm OK in dropping them.

-- 
Luca

Powered by blists - more mailing lists