Re: [PATCH] kernfs: implement custom llseek method to fix userspace regression

[Valentin Sinitsyn]

On 15.08.2023 01:01, Dan Williams wrote:
> Valentine Sinitsyn wrote:
> > Since commit 636b21b50152 ("PCI: Revoke mappings like devmem"),
> > mmapable sysfs binary attributes have started receiving 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: lseek(fd, 0, SEEK_END)
> > now returns zero regardless the real sysfs attribute size which stat()
> > reports. The reason is that kernfs files use generic_file_llseek()
> > implementation, 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
> > attribute or kernfs file representing it. This being said, f_inode
> > remains valid, so stat() which uses it works correctly.
>
> Can you say a bit more about what userspace scenario regressed so that
> others doing backports can make a judgement call on the severity?

We've encountered this regression in the code which used lseek() to 
determine the size of PCI region. It was roughly equivalent to:

#define SYSFS_DEVICE_DIR "/sys/bus/pci/devices/<some id>/"

int fd = open(SYSFS_DEVICE_DIR "/resource0", O_RDWR);
off_t size = lseek(fd, 0, SEEK_END);
assert(size != 0)

Calling lseek() with whence argument other than SEEK_END and non-zero 
offset on this fd returns an error as the kernel considers it as seeking 
past EOF.

I'll add this explanation to v2 commit message.

> > Fixes the regression by implementing a custom llseek fop for kernfs,
> > which uses an attribute's file inode to get the file size,
> > just as stat() does.
> >
> > Fixes: 636b21b50152 ("PCI: Revoke mappings like devmem")
> > Cc: stable@xxxxxxxxxxxxxxx
> > Signed-off-by: Valentine Sinitsyn <valesini@xxxxxxxxxxxxxx>
> > ---
> >   fs/kernfs/file.c | 17 ++++++++++++++++-
> >   1 file changed, 16 insertions(+), 1 deletion(-)
> >
> > diff --git a/fs/kernfs/file.c b/fs/kernfs/file.c
> > index 180906c36f51..6d81e0c981f3 100644
> > --- a/fs/kernfs/file.c
> > +++ b/fs/kernfs/file.c
> > @@ -903,6 +903,21 @@ static __poll_t kernfs_fop_poll(struct file *filp, poll_table *wait)
> >   	return ret;
> >   }
> >   
> > +static loff_t kernfs_fop_llseek(struct file *file, loff_t offset, int whence)
> > +{
> > +	/*
> > +	 * This is almost identical to generic_file_llseek() except it uses
> > +	 * cached inode value instead of f_mapping->host.
> > +	 * The reason is that, for PCI resources in sysfs the latter points to
> > +	 * iomem_inode whose size has nothing to do with the attribute's size.
> > +	 */
> > +	struct inode *inode = file_inode(file);
>
> My only concern is whether there are any scenarios where this is not
> appropriate. I.e. do a bit more work to define a kernfs_ops instance
> specifically for overriding lseek() in this scenario.

Not sure I'm getting you here: do you mean something like this?

struct inode *inode = is_f_mapping_redefined(file) ? file_inode(file) : 
file->f_mapping->host;

My understanding is file->f_inode should always be non-NULL and point to 
the inode corresponding the path of the opened file, so it should be 
safe to call regardless what f_mapping->host is. Do I miss anything?

Best,
Valentin

> > +
> > +	return generic_file_llseek_size(file, offset, whence,
> > +					inode->i_sb->s_maxbytes,
> > +					i_size_read(inode));
> > +}
> > +
> >   static void kernfs_notify_workfn(struct work_struct *work)
> >   {
> >   	struct kernfs_node *kn;
> > @@ -1005,7 +1020,7 @@ EXPORT_SYMBOL_GPL(kernfs_notify);
> >   const struct file_operations kernfs_file_fops = {
> >   	.read_iter	= kernfs_fop_read_iter,
> >   	.write_iter	= kernfs_fop_write_iter,
> > -	.llseek		= generic_file_llseek,
> > +	.llseek		= kernfs_fop_llseek,
> >   	.mmap		= kernfs_fop_mmap,
> >   	.open		= kernfs_fop_open,
> >   	.release	= kernfs_fop_release,
> > --
> > 2.34.1