Hello, I'm not patchbombing further due the size of the patchset (71 patches now). git clone git://git.kernel.org/pub/scm/linux/kernel/git/andrea/aa.git git clone --reference linux-2.6 git://git.kernel.org/pub/scm/linux/kernel/git/andrea/aa.git http://www.kernel.org/pub/linux/kernel/people/andrea/patches/v2.6/2.6.34-rc3/transparent_hugepage-19/ http://www.kernel.org/pub/linux/kernel/people/andrea/patches/v2.6/2.6.34-rc3/transparent_hugepage-19.gz Differences from #18: 1) backout mainline anon-vma changes, I had one bugcheck trigger that seems could be caused by bugs in the anon-vma changes. Those bugs are much more noticeable and severe with transparent hugepage support enabled. I tried the patches available but named wouldn't start anymore with futex returning sigbus. Note the commentary added here about it: http://www.kernel.org/pub/linux/kernel/people/andrea/patches/v2.6/2.6.34-rc3/transparent_hugepage-19/split_huge_page-anon_vma 2) don't compact already compcat transparent hugepages (thanks to Avi, Mel and Johannes for the help). Avi you may want to re-test the sort workload with kernel build in parallel, memory compaction won't split hugepages anymore. http://www.kernel.org/pub/linux/kernel/people/andrea/patches/v2.6/2.6.34-rc3/transparent_hugepage-19/transhuge-isolate_migratepages 3) include AnonHugePages into AnonPages in /proc/meminfo as suggested by Avi http://www.kernel.org/pub/linux/kernel/people/andrea/patches/v2.6/2.6.34-rc3/transparent_hugepage-19/transparent_hugepage_vmstat 4) fix some issue in move_pages syscall (using a suggestion from Christoph to have follow_page split them internally, adding FOLL_SPLIT). Otherwise only the second run of move_pages would succeed. This is one of the special user of gup that like futex is doing more than just DMA or obtaining physical address of the page to setup a secondary MMU on the virtual memory. http://www.kernel.org/pub/linux/kernel/people/andrea/patches/v2.6/2.6.34-rc3/transparent_hugepage-19/pmd_trans_huge_migrate Let's see if this makes it rock solid..... I loaded it in enough places that we'll know in a couple of days. I recommend to use this tree for the benchmarking. Thanks, Andrea -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>