Re: [PATCH] arm64: PCI: fix memleak when calling pci_iomap/unmap()

From: Lorenzo Pieralisi
Date: Wed Sep 09 2020 - 13:38:17 EST


On Wed, Sep 09, 2020 at 02:54:01PM +0100, Catalin Marinas wrote:
> On Wed, Sep 09, 2020 at 12:36:13PM +0100, Lorenzo Pieralisi wrote:
> > On Mon, Sep 07, 2020 at 12:21:19PM +0100, Catalin Marinas wrote:
> > > On Mon, Sep 07, 2020 at 10:51:21AM +0000, George Cherian wrote:
> > > > Catalin Marinas <catalin.marinas@xxxxxxx> wrote:
> > > > > On Sat, Sep 05, 2020 at 10:48:11AM +0800, Yang Yingliang wrote:
> > > > > > diff --git a/arch/arm64/kernel/pci.c b/arch/arm64/kernel/pci.c index
> > > > > > 1006ed2d7c604..ddfa1c53def48 100644
> > > > > > --- a/arch/arm64/kernel/pci.c
> > > > > > +++ b/arch/arm64/kernel/pci.c
> > > > > > @@ -217,4 +217,9 @@ void pcibios_remove_bus(struct pci_bus *bus)
> > > > > > acpi_pci_remove_bus(bus);
> > > > > > }
> > > > > >
> > > > > > +void pci_iounmap(struct pci_dev *dev, void __iomem *addr) {
> > > > > > + iounmap(addr);
> > > > > > +}
> > > > > > +EXPORT_SYMBOL(pci_iounmap);
> > > > >
> > > > > So, what's wrong with the generic pci_iounmap() implementation?
> > > > > Shouldn't it call iounmap() already?
> > > >
> > > > Since ARM64 selects CONFIG_GENERIC_PCI_IOMAP and not
> > > > CONFIG_GENERIC_IOMAP, the pci_iounmap function is reduced to a NULL
> > > > function. Due to this, even the managed release variants or even the explicit
> > > > pci_iounmap calls doesn't really remove the mappings leading to leak.
> > >
> > > Ah, I missed the fact that pci_iounmap() depends on a different
> > > config option.
> > >
> > > > https://lkml.org/lkml/2020/8/20/28
> > >
> > > So is this going to be fixed in the generic code? That would be my
> > > preference.
> > >
> > > A problem with the iounmap() in the proposed patch is that the region
> > > may have been an I/O port, so we could end up unmapping the I/O space.
> >
> > It boils down to finding a way to match a VA to a BAR resource so that
> > we can mirror on pci_iounmap() what's done in pci_iomap_range() (ie
> > check BAR resource flags to define how/if to unmap them), that would do
> > as a generic pci_iounmap() implementation.
>
> In the !CONFIG_GENERIC_IOMAP case (arm64), for IORESOURCE_IO,
> pci_iomap_range() calls __pci_ioport_map() which, with the default
> ioport_map(), it ends up with a simple PCI_IOBASE + (port &
> IO_SPACE_LIMIT).
>
> pci_iounmap() could check whether the pointer is in the PCI_IOBASE -
> PCI_IOBASE+IO_SPACE_LIMIT range before calling ioremap(), unless the
> arch code re-defined ioport_map. Something like below (not even
> compiled):

I gave it some thought - with the current state of affairs (which is not
ideal - this *_IOMAP stuff is ways too complex) it is likely to be the
safest/only way we can have this in generic code, short of implementing
what I mentioned (but that implies keeping track of BAR VA mappings)
or cleaning up this nest of defines.

Lorenzo

> diff --git a/include/asm-generic/io.h b/include/asm-generic/io.h
> index dabf8cb7203b..fada420c9cd6 100644
> --- a/include/asm-generic/io.h
> +++ b/include/asm-generic/io.h
> @@ -919,6 +919,11 @@ extern void __iomem *pci_iomap(struct pci_dev *dev, int bar, unsigned long max);
> #define pci_iounmap pci_iounmap
> static inline void pci_iounmap(struct pci_dev *dev, void __iomem *p)
> {
> +#ifndef ARCH_HAS_IOPORT_MAP
> + if (p >= PCI_IOBASE && p < PCI_IOBASE + IO_SPACE_LIMIT)
> + return;
> + iounmap(p);
> +#endif
> }
> #endif
> #endif /* CONFIG_GENERIC_IOMAP */
> @@ -1009,7 +1014,9 @@ static inline void __iomem *ioremap_uc(phys_addr_t offset, size_t size)
>
> #ifdef CONFIG_HAS_IOPORT_MAP
> #ifndef CONFIG_GENERIC_IOMAP
> -#ifndef ioport_map
> +#ifdef ioport_map
> +#define ARCH_HAS_IOPORT_MAP
> +#else
> #define ioport_map ioport_map
> static inline void __iomem *ioport_map(unsigned long port, unsigned int nr)
> {
>
> --
> Catalin