Re: [U-Boot] [RFC] fastboot: sparse image handling and sessionId

Hi Steve,

Sorry for the late reply.

On Fri, Apr 15, 2016 at 09:14:41PM -0700, Steve Rae wrote:

Maxime,

I suspect that the issue fixed in commit c7529db fastboot: sparse: fix block addressing for don't care chunk type

complicated the assumptions which led to the implementation of the "sessionId" in commit 1f8690a sparse: Implement several chunks flashing

The fastboot client will split the sparse images into several chunks if the
image that it tries to flash is bigger than what the device can handle.

In such a case, the bootloader is supposed to retain the last offset to
which it wrote to, so that it can resume the writes at the right offset
when flashing the next chunk.

Retain the last offset we used, and use the session ID to know if we need
it or not.

Testing of this scenario has revealed that "fastboot flash" works as follows:

When the image to download is larger than the "target reported max download size of 40960000 bytes", the fastboot client (running on the host) essentially performs multiple, independent fastboot download/write steps.

As seen in the logs below, each download/write step: (a) starts at the same address (the address of the first block of the partition...) (b) ends at the same address (start address plus number of blocks "written"...) (c) and uses the CHUNK_TYPE_DONT_CARE 'chunk_sz' to manipulate the addresses to achieve (a) and (b)

Here is an example of the first three fastboot download/writes for a large image to "file-system":

• in the first, data is written to 0x24000-0x37848, then there is a

CHUNK_TYPE_DONT_CARE which "moves" the address pointer to 0x424000

• in the second, it starts with a CHUNK_TYPE_DONT_CARE which "moves"

the address from 0x24000 to 0x37848, then writes from 0x37848-0x4b098, then there is a CHUNK_TYPE_DONT_CARE which "moves" the address pointer to 0x424000

• in the third, it starts with a CHUNK_TYPE_DONT_CARE which "moves"

the address from 0x24000 to 0x4b098, then writes from 0x4b098-0x5e8e0, then there is a CHUNK_TYPE_DONT_CARE which "moves" the address pointer to 0x424000

Starting download of 40932412 bytes .......................................................................... downloading of 40932412 bytes finished Flashing sparse image at offset 147456 Flashing Sparse Image

• this is the starting address: 0x24000

++++ (writing to flash)

• this is the CHUNK_TYPE_DONT_CARE (before) address: 0x37848
• this is the CHUNK_TYPE_DONT_CARE (after) address: 0x424000
• this is the ending address: 0x424000

........ wrote 4194304 blocks to 'file-system'

Starting download of 40954208 bytes .......................................................................... downloading of 40954208 bytes finished Flashing sparse image at offset 147456 Flashing Sparse Image

• this is the starting address: 0x24000
• this is the CHUNK_TYPE_DONT_CARE (before) address: 0x24000
• this is the CHUNK_TYPE_DONT_CARE (after) address: 0x37848

++++ (writing to flash)

• this is the CHUNK_TYPE_DONT_CARE (before) address: 0x4b098
• this is the CHUNK_TYPE_DONT_CARE (after) address: 0x424000
• this is the ending address: 0x424000

........ wrote 4194304 blocks to 'file-system'

Starting download of 40955896 bytes .......................................................................... downloading of 40955896 bytes finished Flashing sparse image at offset 147456 Flashing Sparse Image

• this is the starting address: 0x24000
• this is the CHUNK_TYPE_DONT_CARE (before) address: 0x24000
• this is the CHUNK_TYPE_DONT_CARE (after) address: 0x4b098

++++ (writing to flash)

• this is the CHUNK_TYPE_DONT_CARE (before) address: 0x5e8e0
• this is the CHUNK_TYPE_DONT_CARE (after) address: 0x424000
• this is the ending address: 0x424000

........ wrote 4194304 blocks to 'file-system'

[... snip ...]

Therefore, I think that the "sessonId" that is implemented in U-Boot is actually incorrect. There is no need to keep track of the addresses as each "fastboot flash" download/write step is completely independent.

There was at the time, since it wasn't exactly clear that the next chunks was holding DONT_CARE data to move the device pointer properly.

Also, when it comes to the NAND, it will have quite an impact on the flashing time if we have to rescan the whole partition before the area that we want to flash, at each new chunk. Especially when you already have that info from the previous flashing.

And I actually need to remove it in order to get this working again:

•   if (session_id > 0)
  

•           start = last_offset;
  

•   else
  

•           start = storage->start;
  

•   start = storage->start;
  
  
  

Additionally (unfortunately) as with the previous discussion that we had about this code:

#ifdef CONFIG_FASTBOOT_FLASH_MMC_DEV total_blocks += blkcnt; #endif

the NAND implementation requires a method to process "blkcnt" to ensure the blkcnt skips the bad blocks...

I don't really find what's specific about the NAND here. It was supposed to be a generic abstraction to be able to add easily support new storage backends, and supporting several of them at the same time. Adding that kind of code removes that genericity.

You can add a new callback in the passed structure to adjust the block numbers, the implementation for MMC being trivial.

Maxime