:> it will crunch) at the moment.
:>
:> -Matt
:
:Is there anything I can do?
:
:Cheers,
:-Peter
(CC'ing to current)
Kirk committed the fix -- The sense of a KASSERT() had gotten reversed
accidently.
Everything is hunky dory - except for the fact that I'm not able to
commit any of these things myself. The patches remaining in the current
VFS/BIO set are relatively minor, which means I can shift my attention
over to the VM/INN/MMAP problems.
The INN problem is a stickier issue and the solution may wind up making
-current slightly less efficient in the near term. Basically what we
have to do is to map writeable pages read-only and then take a fault
to detect the clean->dirty transition. We can then synchronously block
in vm_fault (without holding any locks, I might add) when there are too
many dirty pages in the system. The pageout daemon will be able to
run earlier (before it becomes too late).
Also, on the positive side, by accounting dirty pages earlier we can
manage low-memory situations all that more easily, which means that we
can shift the clustering code from the beginning (when queued) to the
end (when I/O is physically initiated). This will have the side effect
of removing a whole cartload of junk from the filesystem critical path
as well as remove a bunch of fields from the vnode structure. It will
also greatly reduce the amount of memory-related blocking that occurs deep
within a call-stack which should help performance on loaded machines
by reducing lock contention.
So if people don't scream at me for the slightly less efficient page
faulting I think we can make progress.
Actually, I don't think the page faulting will be as bad as all that...
the loss of efficiency occurs mostly in shared R+W mmaps, not so much
with BSS extensions because those tend to take a fault anyway because
they start out as zero-fill.
-Matt
Matthew Dillon
<[EMAIL PROTECTED]>
To Unsubscribe: send mail to [EMAIL PROTECTED]
with "unsubscribe freebsd-current" in the body of the message
