- Import GDB 6.8
Maintain snaphshot entries as btree nodes and not as softlinks
Better on-disk placement to reduce seek times
"Free copies" - cache recently read/written block CRCs, on write check CRC and on match verify that the blocks are the same. Then just create a btree link to the original data. This could also be done in a hammer cleanup pass, effectively de-duplicating the file system.
Memory Management (Kernel)
Overal goal: reduce malloc() use and replace with object cache allocations. This should increase performance.
Make free() nonblocking. At the moment a call to free() might block because it might need to
aquire the BGL and maybe needs to wait for other locks (needs to be checked).
This allows us to use free() with dragonfly matching semantic, i.e. it can be used in critical sections
or with a spinlock held/in an interrupt.
For this to work, we need to:
push locking into vm_map_* routines
make sure no code path from free() might need to block
polish up the objcache to work like bonwick described it, most important: make magazines adjust
their size depending on depot lock contention
implement a real slab allocator (bonwick) and switch the objcache to use it. It could us
per-cpu (not so good, i think) or per-numa-cluster slabs, but in the beginning it for sure is possible
to use one slab for all, as most allocations should be satisfied by the objcache (and be serialized
anyways already by the depot lock)
find and change code paths assuming power-of-2 allocations are power-of-2 allocated
switch malloc() and free() to use the objcache for reasonable sized allocations
use kmem for larger ones
Investigate why the system does not boot within one second. I have a fast box, this should be possible. Somewhere are delays adding up. Maybe use kernel threads for probing/initialisation.
Port pmc (and maybe something like oprofile), so that we can find problematic areas.