On Thu, Jan 24, 2019 at 12:48 AM Sascha Hauer <s.hauer@xxxxxxxxxxxxxx> wrote:
>
> On Tue, Jan 22, 2019 at 05:13:37PM -0800, Andrey Smirnov wrote:
> > On 64-bit platforms /dev/mem exceeds the size supported by loff_t and
> > needs special treatment within the rest of FS API. Specifically
> > lseek() needs to be modified to make sure it does the right thing.
> >
> > Prievious attempt at fixing this issue by using IS_ERR_VALUE()
> >
> > e10efc5080 ("fs: fix memory access via /dev/mem for MIPS64")
> >
> > doesn't really work 100% on 64-bit platforms, becuase it still leaves
> > out a number of perfectly valid offsets (e.g. "md 0xffffffffffffff00"
> > doesn't work) . Moreso it breaks lseek() on 32-bit platforms, since
> > IS_ERR_VALUE will retrurn true for any offset that is >= (unsigned
> > long) -MAX_ERRNO.
> >
> > In order to fix this issue on both 32 and 64 bit platforms, introduce
> > DEVFS_UNBOUNDED flag that cdevs can use to denote that they span all
> > 64-bit address space and effectively have not limits. To propagate
> > that info to FS layer, add "unbounded" boolean to FILE. As a last step
> > modify lseek() to be aware of that field and do the right checks in
> > that case.
> >
> > Note, that since loff_t has no problem covering all of address space
> > on 32-bit platforms, DEVFS_UNBOUNDED is defined to expand into 0 and
> > not be settable there.
> >
> > Signed-off-by: Andrey Smirnov <andrew.smirnov@xxxxxxxxx>
> > @@ -422,20 +422,41 @@ loff_t lseek(int fildes, loff_t offset, int whence)
>
> lseek takes a signed loff_t argument. We store the file size in an also
> signed variable. An lseek to the end of a file covering the whole 64bit
> address space (0xffffffffffffffff) will always return an error as
> (loff_t)-1 is both the position and the error code.
>
> I think instead of casting loff_t to an unsigned type whenever we find
> it convenient and then still not having enough bits for storing the
> filesize of 0x10000000000000000 we should rather face the fact that our
> maximum filesize is only half of the 64bit address space.
>
> To put it differently we should think about creating a second /dev/mem
> for the upper half of the 64bit address space.
>
OK, I was trying to preserve as much backwards compatibility as
possible in v1. If creating a separate device is on the table, I'd be
more that happy to make that change in v2, since it'll make a number
of codepaths simpler and saner.
Thanks,
Andrey Smirnov
_______________________________________________
barebox mailing list
barebox@xxxxxxxxxxxxxxxxxxx
http://lists.infradead.org/mailman/listinfo/barebox
