JFFS2 alloc_sem contention

[David Wuertele]

I have a 45MB JFFS2 partition that is usually about 70% full.  My application
developers are sad that sometimes a small write can take up to 30 seconds.

I have increased the priority of the GC thread so that the chances of a write
having to do garbage collection are small.

I have increased the value of resv_blocks_write and resv_blocks_gctrigger so
that when GC does happen, it has enough free blocks to work with, maximizing
efficiency.

I have run KFT to discover exactly where the block is happening.  Here is the
story:

1.  garbage collection runs
2.  a big file gets written, eg "cp /path/to/mybigfile /path/to/jffs2"
3.  the big file cp enters jffs2_reserve_space and grabs the alloc_sem
4.  a small file gets written, eg "cp /path/to/smallfile /path/to/jffs2"
5.  the small file cp enters jffs2_reserve_space and blocks on alloc_sem
6.  the big file takes up to 30s to finish writing, then releases alloc_sem
7.  the small file cp completes with a huge latency

I know the pid of the process that does the small writes.  I would like to give
it priority over the big writes.  I would be satisfied if I could have the big
write return from sys_write having only written a part of the buffer.

I have tried to use ioprio to increase the small writer's priority, but
apparently JFFS2 does not check ioprio.

Any advice?

Thanks,
Dave