Wireless booting and flash RAM fun

After my uncharacteristically non-verbose entry yesterday (Display!), I’m going to post a little more content today. Yesterday I got the LCD display working with my little Python program prototyped on the PC earlier in the week, which got me rather overexcited. The first thing I noticed was that the sound was stuttering playing OGG files … maybe the Python idea was a bit pie-in-the sky. Luckily changing an internal polling interval to be a quarter of a second instead of a tenth fixed that; though I’m now rather doubting that Python is the best idea.

One of the other arguments against Python is the fact it’s not easy for it to talk directly to hardware. In order to program the LCD display I need to poke directly into physical RAM. Currently I’m doing this in a server process that I’ve knocked up in C by modifying Embedded ARM’s sample code. The server listens on port 4000 and with fairly minimal logic sends any incoming data on that port to the LCD device. This has worked well enough to get the Python process running; and has the nice side effect of meaning I can write a simple LCD emulator Windows program (again, in Python) and develop on the PC first. The music player library I’m using — MPD — supports generic control over TCP/IP; so I can pretty much prototype the whole thing on Windows with the sound coming out of the Weebox using this method.

Although this is working OK for now, in the future I’ll want to be able to put various other status messages on the LCD display (e.g. during boot) . A quick skirt around the internet finds lcdmod, a kernel module to control a HD44780-compatible device connected to a PC’s parallel port. This creates a /dev/lcd device which is then easily accessible from shell scripts etc via simple ‘echo “Message” > /dev/lcd’ commands. Of course, my LCD display isn’t attached to the parallel port, nor indeed does the Weebox have a parallel port! One idea I’m toying with is modifying the code to work using the memory- mapped hardware registers, it shouldn’t be too difficult to get working, and could be a good way to reintroduce myself to kernel-space, being my first kernel modifications since about 1996!

As well as playing with the LCD controller, I also managed to get the wireless driver working. Embedded ARM supply their own ‘zd.o’ kernel module, but that’s built for an older kernel, and so moans about invalid kernel versions when I insert it. That kind of thing annoys me, so I wanted to build my own from scratch. This turned out to be fairly straightforward — I grabbed the source from the home page and modified the Makefile to build for 2.4. kernels (it defaults to 2.6.). All built nicely, but there was an unresolved kernel symbol:

zd1211.o: zd1211.o: unresolved symbol do_div64

A bit of hunting around on the Internet and I discovered that do_div64 is implemented in the kernel library only, which doesn’t seem to get linked in. My somewhat cheesey solution was to copy the div64.S file from linux/arch/arm/lib/div64.S into the driver’s folder, and then patch the Makefile to build and link with that file too. (I’ve made a patch for this, which can be installed with patch -p1 < path_to_patch_file from within an unpacked driver folder. You’ll still need to edit the Makefile to edit for your kernel version and location as appropriate.)

With a working wireless connection, I wanted to test the unit booting up without its RJ45 umbilical cord. With this in mind I wanted to get the Weebox files and configuration setup I’d been experimenting with over NFS copied to the flash drive on the unit. My first thoughts were to mount the MTD linux device and copy the files across. So I duly typed ‘mount /dev/mtdblock/1 /mnt’ and expected to be able to modify things in /mnt appropriately. But…problem! The mount command just hung!

Traipsing through the output of dmesg, I spotted:

Using static partition definition
Creating 3 MTD partitions on "NAND 128MiB 3,3V 8-bit":
0x00000000-0x00004000 : "TS-BOOTROM"
mtd: partition "TS-BOOTROM" doesn't end on an erase block -- force read-only
0x00004000-0x07d04000 : "Linux"
mtd: partition "Linux" doesn't start on an erase block boundary -- force

read-only 0x07d04000-0x08000000 : “RedBoot” mtd: partition “RedBoot” doesn’t start on an erase block boundary – force read-only

…oh dear! Have I corrupted my flash RAM somehow? I began to panic, wondering quite what I had done.

Well, luckily I hadn’t blown up my expensive device; I had missed a footnote in the instructions: The TS7250 boards with 128MB NAND flash RAM use the YAFFS2 file system instead of the YAFFS file system. Unbeknownst to me, YAFFS thinks it can read YAFFS2 partitions, but really can’t — giving the cryptic results I was experiencing. When I originally configured the kernel, I used the 7250 configuration. Upon later investigation I realise there’s a 7250-2k configuration too — this is the one I should have used. The ‘2k’ refers to the physical page size of the flash RAM device — the larger flash uses this larger size compared to 512 bytes on smaller devices. It’s important to use YAFFS2 on such a system to prevent too much page wastage, and to improve performance.

At this point I needed to get a new kernel built, with the right version of the flash file system installed. This time round, bored of my overnight builds, I decided to try out the cross-compiler. NFS mounting the Weebox’s root partition from within a Scratchbox session, I was able to compile the kernel in-place, but on my faster 733Mhz x86 linux box. Once all built, I ran ‘make modules_install’ on the Weebox itself to get the relevant kernel modules installed in the right place.

Things to note about building YAFFS and YAFFS2: You can’t have both of them installed! YAFFS2 is backwards-compatible, so it’s an either-or situation with them, which I learned the hard way. If you get duplicate symbols in the final link, make sure you’ve only got one or the other selected. Also to note, if you decide to compile either YAFFS into the kernel directly (instead of as a module) then you must compile MTD support directly into the kernel too. I had initially had MTD support as a module, and YAFFS2 as built-in — you get unresolved symbols for get_mtd_device. Just compile them both in directly.

With an appropriately-built kernel I was then ready to mount the flash RAM image and populate it with all the libraries, executables and configuration files for ALSA, the USB soundcard, wireless drivers, and MPD. At about 7pm last night I was able to boot wirelessly into the weebox and play tunes over the wireless network — hooray! I was very pleased with myself, and when Ness came in I couldn’t help frothing about it all. Bless her, she thinks I’m nuts.

One thing booting up like this told me was that I really need to solve the input issue. It’s all very well having a music playing box, but it’s not exactly user-friendly to have to ssh into it to change tracks! With this in mind this morning I’ve ordered some bits from RS Electronics which will hopefully allow me to detect and decode my remote control handset’s signals. This will mean a bit more kernel fiddling — in order to read the IR detector pulses I need to be able to plug it into an interrupt handler. Luckily, it would seem someone else using the 7200-series board has the same ideas, and the replies to his questions on the 7200 mailing list were extremely useful. One more kernel driver to write though!

Phew, what a long post. And no pictures to break it up either!

Filed under: WeeBox Project
Posted at 10:18:00 BST on 7th July 2006.

About Matt Godbolt

Matt Godbolt is a C++ developer working in Chicago in the finance industry.