Re: [PATCH v5 RESEND 1/5] lib/pci_iomap.c: fix cleanup bugs in pci_iounmap()

[Philipp Stanner]

On Tue, 2024-01-23 at 12:46 -0600, Bjorn Helgaas wrote:
> On Thu, Jan 11, 2024 at 09:55:36AM +0100, Philipp Stanner wrote:
> > pci_iounmap() in lib/pci_iomap.c is supposed to check whether an
> > address
> > is within ioport-range IF the config specifies that ioports exist.
> > If
> > so, the port should be unmapped with ioport_unmap(). If not, it's a
> > generic MMIO address that has to be passed to iounmap().
> > 
> > The bugs are:
> >   1. ioport_unmap() is missing entirely, so this function will
> > never
> >      actually unmap a port.
> 
> The preceding comment suggests that in this default implementation,
> the ioport does not need unmapping, and it wasn't something it was
> supposed to do but just failed to do:
> 
>  * NOTE! This default implementation assumes that if the architecture
>  * support ioport mapping (HAS_IOPORT_MAP), the ioport mapping will
>  * be fixed to the range [ PCI_IOBASE, PCI_IOBASE+IO_SPACE_LIMIT [,
>  * and does not need unmapping with 'ioport_unmap()'.
>  *
>  * If you have different rules for your architecture, you need to
>  * implement your own pci_iounmap() that knows the rules for where
>  * and how IO vs MEM get mapped.
> 
> Almost all ioport_unmap() implementations are empty, so in most cases
> it's a no-op (parisc is an exception).

That sounds correct.

> 
> I'm happy to add the ioport_unmap() even just for symmetry, but if we
> do, I think we should update or remove that comment.

Yes, I think it's the right way: either all architectures should
provide ioport_unmap(), empty or not, or all should use a centralized
PCI function

I can remove the wrong statement.

> 
> >   2. the #ifdef for the ioport-ranges accidentally also guards
> >      iounmap(), potentially compiling an empty function. This would
> >      cause the mapping to be leaked.
> > 
> > Implement the missing call to ioport_unmap().
> > 
> > Move the guard so that iounmap() will always be part of the
> > function.
> 
> I think we should fix this bug in a separate patch because the
> ioport_unmap() is much more subtle and doesn't need to be complicated
> with this fix.

If we agree that one is a bug and the other isn't, then ACK, we should
probably split it.

> 
> > CC: <stable@xxxxxxxxxxxxxxx> # v5.15+
> > Fixes: 316e8d79a095 ("pci_iounmap'2: Electric Boogaloo: try to make
> > sense of it all")
> > Reported-by: Danilo Krummrich <dakr@xxxxxxxxxx>
> 
> Is there a URL we can include for Danilo's report?  I found
> https://lore.kernel.org/all/a6ef92ae-0747-435b-822d-d0229da4683c@xxxxxxxxxx/
> ,
> but I'm not sure that's the right part of the conversation.

He pointed out it's a bug in an offlist conversation with me. The link
you provided is his only public statement about the topic.
The Reported-by served more acknowledging the contribution than issue-
tracking


P.

> 
> > Suggested-by: Arnd Bergmann <arnd@xxxxxxxxxx>
> > Signed-off-by: Philipp Stanner <pstanner@xxxxxxxxxx>
> > Reviewed-by: Arnd Bergmann <arnd@xxxxxxxx>
> > ---
> >  lib/pci_iomap.c | 6 ++++--
> >  1 file changed, 4 insertions(+), 2 deletions(-)
> > 
> > diff --git a/lib/pci_iomap.c b/lib/pci_iomap.c
> > index ce39ce9f3526..6e144b017c48 100644
> > --- a/lib/pci_iomap.c
> > +++ b/lib/pci_iomap.c
> > @@ -168,10 +168,12 @@ void pci_iounmap(struct pci_dev *dev, void
> > __iomem *p)
> >         uintptr_t start = (uintptr_t) PCI_IOBASE;
> >         uintptr_t addr = (uintptr_t) p;
> >  
> > -       if (addr >= start && addr < start + IO_SPACE_LIMIT)
> > +       if (addr >= start && addr < start + IO_SPACE_LIMIT) {
> > +               ioport_unmap(p);
> >                 return;
> > -       iounmap(p);
> > +       }
> >  #endif
> > +       iounmap(p);
> >  }
> >  EXPORT_SYMBOL(pci_iounmap);
> >  
> > -- 
> > 2.43.0
> > 
>