JFFS2 mount fail & crash on 2.4.26 kernel

[Jamie Lokier]

Another amazing coincidence...

My work is using JFFS2 on 2.4.26-uc0 kernels too.  The devices have
been in the field for about 5 years and are still very much in use -
we're still shipping new ones, and I'm still doing some development
work on those devices - and still write patches occasionally for those
ancient kernels.  Binary patches sometimes to fix bugs without
reflashing the kernel.  No joke :-)

Sure, 2.4.26 was old even 5 years ago, but it wasn't old when the
upstream chip supplier produced their SDK, from which the shipped SDK
was a "stable" branch presumably, and it wasn't economic to follow the
2.6 cycle, as the CPU architecture wasn't well supported in 2.6 even
1 year ago (no-MMU ARM).

Back to JFFS2 and 90% full.

In our case, we found that on a 1MB JFFS2 filesytem, with 64k erase
blocks, 90% is simply too much.  All sorts of problems ensue, like the
GC thread running for 5 minutes or more occasionally, writes sometimes
being very slow (but sometimes quick), can't predict when "disk full"
will occur.  Sometimes they linger near full for a long time without
problems, and then suddenly it's a problem.  We've had many devices
crash (without a panic message) and go into weird bootup states where
they don't boot properly and their watchdogs get stuck.  I suspect a
kernel bug is responsible for triggering the weird bootup state.

Limiting fullness to 60% appears to have solved all those problems.

Amol, the amount you can get away with will depend on erase block size
and filesystem size.

-- Jamie