Alternate scheme for page table page allocation

C. Jayachandran c.jayachandran at gmail.com
Fri Apr 30 11:04:37 UTC 2010


I was looking at a hang I get during 'make -j32 buildworld' on XLR
SMP, and as far as I can see, it is caused by the way page table pages
are allocated when we have >512MB memory.

The attached patch
(http://sites.google.com/site/cjayachandran/files//pmap-page-alloc.patch)
changes the way pages for the PTE/PDE entries are allocated in systems
with >512M. This scheme uses vm_phys_alloc_contig to allocate page
table pages KSEG0 region and  has a UMA zone to cache them. This will
allow us to avoid multi-TLB misses that happen when we access page
table entires in the TLB miss handler. I think a similar approach can
be taken to allocate 8K page-pairs for N32(for >4GB RAM) and N64.

With this patch, 'make -j32 buildworld' consistently works with 32
cpus  on an XLR is booted SMP with 4GB RAM.  With the current page
alloc code, I get a hang about an hour into buildworld.

The patch is based off r206712 - I still have not found a fix for the
crashes I see in versions after that. Please let me know your
comments,  especially if you can think of a better way of doing this.
I can make a version of this patch for HEAD if this is acceptable.

JC.


More information about the freebsd-mips mailing list