| Message ID | a9daf8d8444d0ebd00bc6d64e336ec49dbb50784.1600254147.git.lorenzo.pieralisi@arm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Fix pci_iounmap() on !CONFIG_GENERIC_IOMAP | expand |
On Wed, Sep 16, 2020 at 12:06:58PM +0100, Lorenzo Pieralisi wrote: > For arches that do not select CONFIG_GENERIC_IOMAP, the current > pci_iounmap() function does nothing causing obvious memory leaks > for mapped regions that are backed by MMIO physical space. > > In order to detect if a mapped pointer is IO vs MMIO, a check must made > available to the pci_iounmap() function so that it can actually detect > whether the pointer has to be unmapped. > > In configurations where CONFIG_HAS_IOPORT_MAP && !CONFIG_GENERIC_IOMAP, > a mapped port is detected using an ioport_map() stub defined in > asm-generic/io.h. > > Use the same logic to implement a stub (ie __pci_ioport_unmap()) that > detects if the passed in pointer in pci_iounmap() is IO vs MMIO to > iounmap conditionally and call it in pci_iounmap() fixing the issue. > > Leave __pci_ioport_unmap() as a NOP for all other config options. > > Reported-by: George Cherian <george.cherian@marvell.com> > Link: https://lore.kernel.org/lkml/20200905024811.74701-1-yangyingliang@huawei.com > Link: https://lore.kernel.org/lkml/20200824132046.3114383-1-george.cherian@marvell.com > Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com> > Cc: Arnd Bergmann <arnd@arndb.de> > Cc: George Cherian <george.cherian@marvell.com> > Cc: Will Deacon <will@kernel.org> > Cc: Bjorn Helgaas <bhelgaas@google.com> > Cc: Catalin Marinas <catalin.marinas@arm.com> > Cc: Yang Yingliang <yangyingliang@huawei.com> > --- > include/asm-generic/io.h | 39 +++++++++++++++++++++++++++------------ > 1 file changed, 27 insertions(+), 12 deletions(-) This works for me. The only question I have is whether pci_iomap.h is better than io.h for __pci_ioport_unmap(). These headers are really confusing. Either way: Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
On Wed, Sep 16, 2020 at 03:51:11PM +0100, Catalin Marinas wrote: > On Wed, Sep 16, 2020 at 12:06:58PM +0100, Lorenzo Pieralisi wrote: > > For arches that do not select CONFIG_GENERIC_IOMAP, the current > > pci_iounmap() function does nothing causing obvious memory leaks > > for mapped regions that are backed by MMIO physical space. > > > > In order to detect if a mapped pointer is IO vs MMIO, a check must made > > available to the pci_iounmap() function so that it can actually detect > > whether the pointer has to be unmapped. > > > > In configurations where CONFIG_HAS_IOPORT_MAP && !CONFIG_GENERIC_IOMAP, > > a mapped port is detected using an ioport_map() stub defined in > > asm-generic/io.h. > > > > Use the same logic to implement a stub (ie __pci_ioport_unmap()) that > > detects if the passed in pointer in pci_iounmap() is IO vs MMIO to > > iounmap conditionally and call it in pci_iounmap() fixing the issue. > > > > Leave __pci_ioport_unmap() as a NOP for all other config options. > > > > Reported-by: George Cherian <george.cherian@marvell.com> > > Link: https://lore.kernel.org/lkml/20200905024811.74701-1-yangyingliang@huawei.com > > Link: https://lore.kernel.org/lkml/20200824132046.3114383-1-george.cherian@marvell.com > > Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com> > > Cc: Arnd Bergmann <arnd@arndb.de> > > Cc: George Cherian <george.cherian@marvell.com> > > Cc: Will Deacon <will@kernel.org> > > Cc: Bjorn Helgaas <bhelgaas@google.com> > > Cc: Catalin Marinas <catalin.marinas@arm.com> > > Cc: Yang Yingliang <yangyingliang@huawei.com> > > --- > > include/asm-generic/io.h | 39 +++++++++++++++++++++++++++------------ > > 1 file changed, 27 insertions(+), 12 deletions(-) > > This works for me. The only question I have is whether pci_iomap.h is > better than io.h for __pci_ioport_unmap(). These headers are really > confusing. Yes they are, in total honesty there is much more to do to make them sane, this patch is just a band-aid. I thought about moving this stuff into pci_iomap.h, though that file is included _independently_ from io.h from some arches so I tried to keep everything in io.h to minimize disruption. We can merge this patch - since it is a fix after all - and then I can try to improve the whole pci_iounmap() includes. > Either way: > > Reviewed-by: Catalin Marinas <catalin.marinas@arm.com> Thanks a lot. I'd appreciate a tested-by from the George as he is the one who reported the problem. Lorenzo
diff --git a/include/asm-generic/io.h b/include/asm-generic/io.h index dabf8cb7203b..9ea83d80eb6f 100644 --- a/include/asm-generic/io.h +++ b/include/asm-generic/io.h @@ -911,18 +911,6 @@ static inline void iowrite64_rep(volatile void __iomem *addr, #include <linux/vmalloc.h> #define __io_virt(x) ((void __force *)(x)) -#ifndef CONFIG_GENERIC_IOMAP -struct pci_dev; -extern void __iomem *pci_iomap(struct pci_dev *dev, int bar, unsigned long max); - -#ifndef pci_iounmap -#define pci_iounmap pci_iounmap -static inline void pci_iounmap(struct pci_dev *dev, void __iomem *p) -{ -} -#endif -#endif /* CONFIG_GENERIC_IOMAP */ - /* * Change virtual addresses to physical addresses and vv. * These are pretty trivial @@ -1016,6 +1004,16 @@ static inline void __iomem *ioport_map(unsigned long port, unsigned int nr) port &= IO_SPACE_LIMIT; return (port > MMIO_UPPER_LIMIT) ? NULL : PCI_IOBASE + port; } +#define __pci_ioport_unmap __pci_ioport_unmap +static inline void __pci_ioport_unmap(void __iomem *p) +{ + uintptr_t start = (uintptr_t) PCI_IOBASE; + uintptr_t addr = (uintptr_t) p; + + if (addr >= start && addr < start + IO_SPACE_LIMIT) + return; + iounmap(p); +} #endif #ifndef ioport_unmap @@ -1030,6 +1028,23 @@ extern void ioport_unmap(void __iomem *p); #endif /* CONFIG_GENERIC_IOMAP */ #endif /* CONFIG_HAS_IOPORT_MAP */ +#ifndef CONFIG_GENERIC_IOMAP +struct pci_dev; +extern void __iomem *pci_iomap(struct pci_dev *dev, int bar, unsigned long max); + +#ifndef __pci_ioport_unmap +static inline void __pci_ioport_unmap(void __iomem *p) {} +#endif + +#ifndef pci_iounmap +#define pci_iounmap pci_iounmap +static inline void pci_iounmap(struct pci_dev *dev, void __iomem *p) +{ + __pci_ioport_unmap(p); +} +#endif +#endif /* CONFIG_GENERIC_IOMAP */ + /* * Convert a virtual cached pointer to an uncached pointer */
For arches that do not select CONFIG_GENERIC_IOMAP, the current pci_iounmap() function does nothing causing obvious memory leaks for mapped regions that are backed by MMIO physical space. In order to detect if a mapped pointer is IO vs MMIO, a check must made available to the pci_iounmap() function so that it can actually detect whether the pointer has to be unmapped. In configurations where CONFIG_HAS_IOPORT_MAP && !CONFIG_GENERIC_IOMAP, a mapped port is detected using an ioport_map() stub defined in asm-generic/io.h. Use the same logic to implement a stub (ie __pci_ioport_unmap()) that detects if the passed in pointer in pci_iounmap() is IO vs MMIO to iounmap conditionally and call it in pci_iounmap() fixing the issue. Leave __pci_ioport_unmap() as a NOP for all other config options. Reported-by: George Cherian <george.cherian@marvell.com> Link: https://lore.kernel.org/lkml/20200905024811.74701-1-yangyingliang@huawei.com Link: https://lore.kernel.org/lkml/20200824132046.3114383-1-george.cherian@marvell.com Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com> Cc: Arnd Bergmann <arnd@arndb.de> Cc: George Cherian <george.cherian@marvell.com> Cc: Will Deacon <will@kernel.org> Cc: Bjorn Helgaas <bhelgaas@google.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Yang Yingliang <yangyingliang@huawei.com> --- include/asm-generic/io.h | 39 +++++++++++++++++++++++++++------------ 1 file changed, 27 insertions(+), 12 deletions(-)