On Tue, Jan 30, 2001 at 03:21:38PM -0800, Matt Dillon wrote:
> :vm.swapdev1.total (this is the one that is currently hard to get)
> You can't move swapinfo into the kernel as a sysctl unless you
> solve this problem. Traversing the radix tree is expensive enough
> that the entire system will stall for a short period of time if you run
> the loop in the kernel.
Yes. I did not have that in mind.
> The issue here is that swapinfo tries to break the useage down by
> swap area, whereas the kernel has no real concept of swap areas
> in the allocation map -- it just sees one big contiguous allocation
> map. So the kernel does not track allocations on a per-swap-area
> basis.
My thought was the following: add a swapaccount vnode operation, and
then do (in flushchainbuf):
VOP_SWAPACCOUNT(nbp->b_vp, nbp);
BUF_KERNPROC(nbp);
BUF_STRATEGY(nbp);
Now, I have to define the vop in vm_swap.c, where I can get the area
index from the block number in a clean way and frob my counters.
The same method could be used in swap_pager_putpages().
swap_pager_reserve() would need to construct a buf and act in a similar
way as e.g. swap_pager_strategy(), only that the buffer is not flushed
afterwards, only VOP_SWAPACCOUNT is called. Of course, this can be
done much easier by just passing the block number to VOP_SWAPACCOUNT,
and calling it for each blk returned by swp_pager_getswapspace()
(it costs virtually nothing, so it should be OK).
So, VOP_SWAPACCOUNT would become:
VOP_SWAPACCOUNT(nbp->b_vp, nbp->b_blkno, nbp->bcount);
Or did I get things wrong?
- thomas
To Unsubscribe: send mail to [EMAIL PROTECTED]
with "unsubscribe freebsd-hackers" in the body of the message
