I'm trying to restore this uncompleted project.
Original thread introduction by psyke83:
Yesterday I discovered a very interesting source release from Samsung: GT-S5830G_GB_Opensource.zip
Normally, there is a shared Samsung source release for the msm7k range of devices (Ace, Mini, Callisto, Beni, Gio; there is also partial support included for my GT-I5500/Europa, but I adapted the source to properly support my phone). All of these devices usually depend on Samsung's proprietary FSR (Flex Sector Remapper) drivers in order to access the flash memory, which are taken from the stock ROMs, but our dependence on this driver locks us into the 2.6.35 kernel since we don't have access to the driver source.
The new source that I found for the GT-S5830G model, however, appears to contain modifications for the purpose of transitioning from Samsung's proprietary FSR driver to the open source MTD (via msm_nand) driver. If this can work correctly, using this driver would be much preferred over the proprietary Samsung stuff.
Here's the source that I uploaded to github: https://github.com/psyke83/android_k.../tree/purenand
Some observations that I have made, keeping in mind that I'm testing on my GT-I5500:
There are two separate configs for the cooper rev03 (Galaxy Ace): the standard defconfig that uses fsr/rfs, and a "purenand" config that uses the mtd/yaffs2 drivers instead of fsr/rfs. In other words, it's using the open drivers for flash access. Here is the diff: https://gist.github.com/3365123
The dpram driver (Samsung's driver for communication with baseband, used by RIL) is patched to support MTD instead of BML when the proper config is set.
The drivers/mtd/devices/msm_nand.c driver is modified by Samsung, but they applied their patches to an older revision of msm_nand.c from Froyo. Here is the diff when comparing this file vs the Froyo revision, so you can see more clearly the changes: https://gist.github.com/3365161
By default, the msm_nand.c driver causes the kernel to hang on my device (this is true for both this source and the older 2.6.35 Samsung source not based on purenand). I have isolated the hang to the flash_onfi_probe function.
As you can see here, Samsung added code to bypass this function on the Cooper board, and use the secondary detection method only. If I include my board to this ifdef block, it solves the issue with the kernel hanging on my device. I also need to patch some checks in the onenand detection, because the driver explicitly looks for onenand devices with a device_id of 0x40 and num_of_buffers as 0x201, but the chip on the GT-I5500 is different (device_id is 0x50, num_of_buffers is 0x101). This patch solves these problems: https://gist.github.com/3365222
Here is a dmesg log from my device after patching the code: https://gist.github.com/3360727. For comparison purposes, look at the block mapping that the fsr driver reports for my device when using the BML mapping: https://github.com/psyke83/android_d...dConfig.mk#L53
As you can see, the partitions names and order detection is correct for the msm_nand driver, but the address mappings are exactly half of what they are supposed to be (e.g. the first partition, mibib, should range from 0x00000000-0x00180000, but the mtd driver detects the memory range as 0x00000000-0x000C0000.
If I try to perform a "nandump -f /sdcard/cache.img /dev/mtd/mtd13", there are no obvious errors in the dmesg log, but the tool will dump the cache partition until the sd card becomes full (over 300mb in my case, but the real /cache partition is only 25MB), and will then output "nanddump: short write". The resulting dump will be filled with 0xFF when examined with a hex editor (even though I'm sure that the /cache partition is not blank in reality).
These are my findings so far. I'd appreciate any kernel hackers to help me out. If we can crack this problem and get open onenand drivers working, then our devices will no longer be locked to any specific kernel release.
galaxy5 (europa): mtd can read/write partitions (via dd). But CM doesn't seem to work completely. All tests done by psyke83:
Furthermore erikcas managed to mount a partition but its size was reported to 0. Probably a badblock problem -> onenand_read
ace (cooper): mtd can't dump parititons correctly. All tests done by me!
More info in next post.
Last patch: https://gist.github.com/psyke83/6420048
You can apply this again 2.6.35 (purenand branch) or (with minor changes) again 3.0.8/3.0.31
psyke83 and erikcas stopped development and original thread seem abandoned, I hope that here in xda somebody can help me to restore this!