Development system update - 20-Jan-2010
dillon at apollo.backplane.com
Wed Jan 20 08:22:52 PST 2010
The development kernel should be stable again relative to the work
I started about a week and a half ago. A great deal of progress has
* Programs reading files via mmap now run optimally. The VM faults
now properly pipeline read-aheads again. Previously the pipelining
was not working properly and access to files via mmap was suboptimal.
* The VM system is now able to pre-fault writable zero-fill pages,
resulting in an approximately 5% improvement in parallel buildworld
times on SMP systems.
Generally speaking the number of VM faults the system takes has
been reduced considerably due to the fixed read-ahead pipelining
and the new pre-faulting code.
* HAMMER and NFS both pass fsx tests again.
* NFS and UFS have been entirely converted over to the kernel generic VM
* The new namecache MP code has been stabilized. vfs.cache_mpsafe
is still turned off by default but will be turned on before the
March release. It should be safe for people to turn on generally.
* A couple of system calls had MP lock mismatches related to the
MP lock work, and have been fixed.
* VKernels were broken due to a long-standing bug that the new VM
fault code revealed. Now fixed (and they should be more stable
now as well).
Those are the main points. The development kernel as of today is in
a very good place.
I am still working on the recovery code for the new REDO features in
HAMMER. This is not finished yet.
<dillon at backplane.com>
More information about the Kernel