Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Sun, 23 Aug 2020 04:01:06 -0400
From:   "Michael S. Tsirkin" <mst@redhat.com>
To:     Bjorn Helgaas <helgaas@kernel.org>
Cc:     George Cherian <george.cherian@marvell.com>,
        linux-kernel@vger.kernel.org, linux-arch@vger.kernel.org,
        linux-pci@vger.kernel.org, bhelgaas@google.com, arnd@arndb.de
Subject: Re: [PATCHv2] PCI: Add pci_iounmap
Message-ID: <20200823040008-mutt-send-email-mst@kernel.org>
References: <20200820050306.2015009-1-george.cherian@marvell.com>
 <20200820215549.GA1569713@bjorn-Precision-5520>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20200820215549.GA1569713@bjorn-Precision-5520>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Aug 20, 2020 at 04:55:49PM -0500, Bjorn Helgaas wrote:
> [+cc Michael, author of 66eab4df288a ("lib: add GENERIC_PCI_IOMAP")]
> 
> On Thu, Aug 20, 2020 at 10:33:06AM +0530, George Cherian wrote:
> > In case if any architecture selects CONFIG_GENERIC_PCI_IOMAP and not
> > CONFIG_GENERIC_IOMAP, then the pci_iounmap function is reduced to a NULL
> > function. Due to this the managed release variants or even the explicit
> > pci_iounmap calls doesn't really remove the mappings.
> > 
> > This issue is seen on an arm64 based system. arm64 by default selects
> > only CONFIG_GENERIC_PCI_IOMAP and not CONFIG_GENERIC_IOMAP from this
> > 'commit cb61f6769b88 ("ARM64: use GENERIC_PCI_IOMAP")'
> > 
> > Simple bind/unbind test of any pci driver using pcim_iomap/pci_iomap,
> > would lead to the following error message after long hour tests
> > 
> > "allocation failed: out of vmalloc space - use vmalloc=<size> to
> > increase size."
> > 
> > Signed-off-by: George Cherian <george.cherian@marvell.com>
> > ---
> > * Changes from v1
> > 	- Fix the 0-day compilation error.
> > 	- Mark the lib/iomap pci_iounmap call as weak incase 
> > 	  if any architecture have there own implementation.
> > 
> >  include/asm-generic/io.h |  4 ++++
> >  lib/pci_iomap.c          | 10 ++++++++++
> >  2 files changed, 14 insertions(+)
> > 
> > diff --git a/include/asm-generic/io.h b/include/asm-generic/io.h
> > index dabf8cb7203b..5986b37226b7 100644
> > --- a/include/asm-generic/io.h
> > +++ b/include/asm-generic/io.h
> > @@ -915,12 +915,16 @@ static inline void iowrite64_rep(volatile void __iomem *addr,
> >  struct pci_dev;
> >  extern void __iomem *pci_iomap(struct pci_dev *dev, int bar, unsigned long max);
> >  
> > +#ifdef CONFIG_GENERIC_PCI_IOMAP
> > +extern void pci_iounmap(struct pci_dev *dev, void __iomem *p);
> > +#else
> >  #ifndef pci_iounmap
> >  #define pci_iounmap pci_iounmap
> >  static inline void pci_iounmap(struct pci_dev *dev, void __iomem *p)
> >  {
> >  }
> >  #endif
> > +#endif /* CONFIG_GENERIC_PCI_IOMAP */
> >  #endif /* CONFIG_GENERIC_IOMAP */
> >  
> >  /*
> > diff --git a/lib/pci_iomap.c b/lib/pci_iomap.c
> > index 2d3eb1cb73b8..ecd1eb3f6c25 100644
> > --- a/lib/pci_iomap.c
> > +++ b/lib/pci_iomap.c
> > @@ -134,4 +134,14 @@ void __iomem *pci_iomap_wc(struct pci_dev *dev, int bar, unsigned long maxlen)
> >  	return pci_iomap_wc_range(dev, bar, 0, maxlen);
> >  }
> >  EXPORT_SYMBOL_GPL(pci_iomap_wc);
> > +
> > +#ifndef CONFIG_GENERIC_IOMAP
> > +#define pci_iounmap pci_iounmap
> > +void __weak pci_iounmap(struct pci_dev *dev, void __iomem *addr);
> > +void __weak pci_iounmap(struct pci_dev *dev, void __iomem *addr)
> > +{
> > +	iounmap(addr);
> > +}
> > +EXPORT_SYMBOL(pci_iounmap);
> > +#endif
> 
> I completely agree that this looks like a leak that needs to be fixed.
> 
> But my head hurts after trying to understand pci_iomap() and
> pci_iounmap().  I hate to add even more #ifdefs here.  Can't we
> somehow rationalize this and put pci_iounmap() next to pci_iomap()?
> 
> 66eab4df288a ("lib: add GENERIC_PCI_IOMAP") moved pci_iomap() from
> lib/iomap.c to lib/pci_iomap.c, but left pci_iounmap() in lib/iomap.c.
> There must be some good reason why they're separated, but I don't know
> what it is.

My recollection is vaguely that map code was more or less
the same across architectures, but unmap was often different,
so I only moved map into the library.

> >  #endif /* CONFIG_PCI */
> > -- 
> > 2.25.1
> >