Re: [PATCH v9 2/2] PCI: Implement custom llseek for sysfs resource entries

[Valentin Sinitsyn]

On 05.10.2023 16:41, Greg Kroah-Hartman wrote:
> On Mon, Sep 25, 2023 at 11:40:13AM +0300, Valentine Sinitsyn wrote:
> > Since commit 636b21b50152 ("PCI: Revoke mappings like devmem"), mmappable
> > sysfs entries have started to receive their f_mapping from the iomem
> > pseudo filesystem, so that CONFIG_IO_STRICT_DEVMEM is honored in sysfs
> > (and procfs) as well as in /dev/[k]mem.
> >
> > This resulted in a userspace-visible regression:
> >
> > 1. Open a sysfs PCI resource file (eg. /sys/bus/pci/devices/*/resource0)
> > 2. Use lseek(fd, 0, SEEK_END) to determine its size
> >
> > Expected result: a PCI region size is returned.
> > Actual result: 0 is returned.
> >
> > The reason is that PCI resource files residing in sysfs use
> > generic_file_llseek(), which relies on f_mapping->host inode to get the
> > file size. As f_mapping is now redefined, f_mapping->host points to an
> > anonymous zero-sized iomem_inode which has nothing to do with sysfs file
> > in question.
> >
> > Implement a custom llseek method for sysfs PCI resources, which is
> > almost the same as proc_bus_pci_lseek() used for procfs entries.
> >
> > This makes sysfs and procfs entries consistent with regards to seeking,
> > but also introduces userspace-visible changes to seeking PCI resources
> > in sysfs:
> >
> > - SEEK_DATA and SEEK_HOLE are no longer supported;
> > - Seeking past the end of the file is prohibited while previously
> >    offsets up to MAX_NON_LFS were accepted (reading from these offsets
> >    was always invalid).
> >
> > Fixes: 636b21b50152 ("PCI: Revoke mappings like devmem")
> > Cc: stable@xxxxxxxxxxxxxxx
> > Signed-off-by: Valentine Sinitsyn <valesini@xxxxxxxxxxxxxx>
> > Acked-by: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
> > ---
> >   drivers/pci/pci-sysfs.c | 26 +++++++++++++++++++++++++-
> >   1 file changed, 25 insertions(+), 1 deletion(-)
>
> I'll take these now, for 6.7-rc1, but not mark them as fixes or cc:
Thanks, appreciated.

> stable as this is a new functionality, the code has never worked for
> lseek on these files so it's not like anything was broken :)
In fact, lseek() on PCI resource files in sysfs was broken since commit 
636b21b50152. That was the reason why I started to investigate the 
issue: one of our applications stopped working after a kernel update.

I'm not hundred percent sure if it belongs to stable, but it does fix a 
user-visible regression.

Best,
Valentin