Re: About the "__refdata" tag in pci-keystone.c

From: Rob Herring <robh@kernel.org>
To: Bjorn Helgaas <helgaas@kernel.org>
Cc: "Sergio M. Iglesias" <sergio@lony.xyz>,
	Bjorn Helgaas <bhelgaas@google.com>,
	Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>,
	Krzysztof Wilczynski <kw@linux.com>,
	PCI <linux-pci@vger.kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: Re: About the "__refdata" tag in pci-keystone.c
Date: Wed, 29 Sep 2021 08:23:57 -0500	[thread overview]
Message-ID: <CAL_JsqLjpe94EGx2fmVmcoXniRMnQ16+RKcYATub4cm0TPF0gw@mail.gmail.com> (raw)
In-Reply-To: <20210927153128.GA646260@bhelgaas>

On Mon, Sep 27, 2021 at 10:31 AM Bjorn Helgaas <helgaas@kernel.org> wrote:
>
> On Mon, Sep 27, 2021 at 12:33:21PM +0200, Sergio M. Iglesias wrote:
> > Hello!
> >
> > I have checked the "__refdata" tag that appears in the file
> > "drivers/pci/controller/dwc/pci-keystone.c" and it is needed. The tag has
> > been there since the creation of the file on commit 6e0832fa432e and
> > nothing has changed since that would make it redundant.
> >
> > The reason it is needed is because the struct references "ks_pcie_probe",
> > which is a function tagged as "__init", so the compiler will most likely
> > complain about the "__refdata" being removed.
> >
> > Should I send a patch to add a comment explaining why it is a necessary
> > tag as recommended in "include/linux/init.h"?
> > > [...] so optimally document why the __ref is needed and why it's OK).
>
> Thanks a lot for looking into this.
>
> I'm not yet convinced that either the __init or the __refdata is
> necessary.  If there is a reason, it would not be "to silence a
> compiler complaint"; it would be something like "the keystone platform
> is different from all the other platforms because the other platforms
> support X but keystone does not."
>
> Also, there are a couple other .probe() functions that are marked
> __init:
>
>   $ git grep "static int .*_pci.*_probe" drivers/pci | grep __init
>   drivers/pci/controller/dwc/pci-keystone.c:static int __init ks_pcie_probe(struct platform_device *pdev)
>   drivers/pci/controller/dwc/pci-layerscape-ep.c:static int __init ls_pcie_ep_probe(struct platform_device *pdev)
>   drivers/pci/controller/mobiveil/pcie-layerscape-gen4.c:static int __init ls_pcie_g4_probe(struct platform_device *pdev)
>   drivers/pci/controller/pci-ixp4xx.c:static int __init ixp4xx_pci_probe(struct platform_device *pdev)
>
> and their platform_driver structs are not marked __refdata:
>
>   $ git grep "static struct platform_driver" drivers/pci | grep __refdata
>   drivers/pci/controller/dwc/pci-keystone.c:static struct platform_driver ks_pcie_driver __refdata = {
>
> I think this should all be more consistent: either all these __init
> and __refdata annotations should be removed, or they should be used by
> many more drivers.

__init should definitely be removed. There's no guarantee that probe
completes before init memory is freed even for built-in drivers.

Rob