[syslinux] KRNLINUX? CPIOLINUX?

[Shao Miller]

On 9/5/2011 15:56, H. Peter Anvin wrote:
> On 09/05/2011 08:19 AM, Shao Miller wrote:
>>
>> Also: Do we make this a "cpio FS" driver that uses the block-caching
>> abstraction (and hook the disk I/O to actually do memory access,
>> instead) or do we make this an "initramfs" driver that will never allow
>> for supporting initramfs on a disk, but only in memory? I could see a
>> use for the former, if we every wish to allow "mounting" initramfs files
>> that are not in memory, but it seems a tad more complicated.
>>
>
> The big problem here is of course that the cpio images are extremely
> unlikely to be aligned.

True.  I'm thinking that the hash table you suggested could index nodes 
with a starting block and an offset.  In pcacjr's NTFS work, there's 
lots of computations between blocks, clusters, sectors, MFT records, 
etc.  I guess we just lose out on any meaningful mapping between "files" 
and "blocks" (or even sectors).

...Or maybe a 4-byte block size could work.  Ha!  Imagine an INT 0x13 
hook that reported _that_ to consumers!

The [eventual] Linux x86 kernel format Syslinux payload could hard-wire 
the root FS block device to memory, as we agree.

Now do we walk the whole initramfs with a one-time pass and build the 
hash table, or do we build it as-needed?  With the former, we're slow to 
start.  With the latter, we're slow if we process multiple files.

- Shao Miller