[U-Boot] dfu: dfu and UBI Volumes

[Benoît Thébaudeau]

Hi Pantelis,

On Tuesday, May 28, 2013 6:43:06 PM, Pantelis Antoniou wrote:
> Hi Benoît
> 
> On May 28, 2013, at 7:31 PM, Benoît Thébaudeau wrote:
> 
> > Dear Pantelis Antoniou,
> > 
> > On Tuesday, May 28, 2013 5:05:12 PM, Pantelis Antoniou wrote:
> >> Hi Tom,
> >> 
> >> On May 28, 2013, at 6:01 PM, Tom Rini wrote:
> >> 
> >>> On Tue, May 28, 2013 at 07:50:46AM +0200, Wolfgang Denk wrote:
> >>>> Dear Tom,
> >>>> 
> >>>> In message <20130527233735.GZ17119 at bill-the-cat> you wrote:
> >>>>> 
> >>>>>> Where exactly is this 8 MB limit coming into play?
> >>>>> 
> >>>>> In buffering the data.  We cannot write a chunk of a file to a
> >>>>> filesystem and then append to it, we don't have the API today.
> >>>> 
> >>>> Sorry, I still don't get it. Assuming I have a GiB of RAM, why can I
> >>>> not load a 256 MiB file to RAM, and then write it to a file system?
> >>>> 
> >>>> I have definitely dealt with images and files bigger than 8 MiB in
> >>>> thepast, so I really don't see where any buffer problem could be.
> >>> 
> >>> I thought I might not have been clear about where this limit comes from,
> >>> after I sent the email.  The problem we have, and this is only for
> >>> writing to a filesystem (_not_ writing of a filesystem) is that we do
> >>> not have the API for appending to files, only create/overwrite.  So we
> >>> must read the whole file into memory, and then write it out.  The DFU
> >>> protocol doesn't have (I would swear anyhow) a part where it says "I'm
> >>> about to send you a blob of X bytes", so we cannot know at the start how
> >>> much data is coming our way.
> >>> 
> >>> Today we "solve" this with a statically defined
> >>> CONFIG_SYS_DFU_MAX_FILE_SIZE.  Looking at things again, I think this is
> >>> buggy right now in that we need to also whack DFU_DATA_BUF_SIZE to also
> >>> be that same value.  Going forward, we may be able to switch this to
> >>> (and both of these are off the top of my head) a getenv to see how much
> >>> space to malloc, or just making it a malloc and adding some compile-time
> >>> check to ensure that the malloc area is at least as big as
> >>> CONFIG_SYS_DFU_MAX_FILE_SIZE.
> >>> 
> >> 
> >> Correct, the DFU protocol doesn't have a method to inform you before hand
> >> about the size of the transfer about to happen.
> >> 
> >> The only possible solution I see at this point is to have an environment
> >> variable, i.e. dfubuf that controls the size of the buffer.
> >> 
> >> Upon start of a dfu transfer we can allocate the buffer, and do our
> >> thing.
> > 
> > I don't know the details of the DFU implementation in U-Boot, but the
> > specification leaves the choice between programming the firmware on-the-fly
> > during the download, and later during the manifestation phase (or a mix of
> > both). Hence, there is not need for a global firmware buffer if U-Boot goes
> > for
> > the on-the-fly programming strategy. The only buffer constraint would be
> > wTransferSize (chosen by U-Boot for the control endpoint) in that case. See
> > "7. Manifestation Phase" on page 26 here:
> > http://www.usb.org/developers/devclass_docs/DFU_1.1.pdf
> > 
> 
> The problem is not DFU TBH, it's that since we don't have an option to append
> to a file, we have to have the whole file transferred in RAM and written in
> one go. The raw medium dfu methods in u-boot don't have a problem.
> 
> > Of course this can't yet apply to writing files on file systems since the
> > current API in U-Boot misses the append feature, but this could be applied
> > to
> > program raw memory partitions, including UBI images.
> > 
> 
> It already happens for raw memory partitions, it's the UBI images being
> discussed.

But what does appending to a file has to do with programming a UBI image, which
is a memory partition containing a whole file system? This is what I don't get
in this discussion. Is it because of a restriction of the DFU API in U-Boot?

Best regards,
Benoît