Messages in this thread Patch in this message |  | | | Subject | Re: Process-shared futexes on hugepages puts the kernel in an infinite loop in 2.6.32.11; is this fixed now? | | From | Peter Zijlstra <> | | Date | Fri, 16 Apr 2010 22:27:48 +0200 |
| |
On Fri, 2010-04-16 at 23:45 +0800, r6144 wrote: > Hello all, > > I'm having an annoying kernel bug regarding huge pages in Fedora 12: > > https://bugzilla.redhat.com/show_bug.cgi?id=552257 > > Basically I want to use huge pages in a multithreaded number crunching > program, which happens to use process-shared semaphores (because fftw > does it). The futex for the semaphore ends up lying on a huge page, and > I then get an endless loop in get_futex_key(), apparently because the > anonymous huge page containing the futex does not have a page->mapping. > A test case is provided in the above link. > > I reported the bug to Fedora bugzilla months ago, but haven't received > any feedback yet.
No, it works much better if you simply mail LKML and CC people who work on the code in question ;-)
> The Fedora kernel is based on 2.6.32.11, and a > cursory glance at the 2.6.34-rc3 source does not yield any relevant > change. > > So, could anyone tell me if the current mainline kernel might act better > in this respect, before I get around to compiling it?
Right, so I had a quick chat with Mel, and it appears MAP_PRIVATE hugetlb pages don't have their page->mapping set.
I guess something like the below might work, but I'd really rather not add hugetlb knowledge to futex.c. Does anybody else have a better idea? Maybe create something similar to an anon_vma for hugetlb pages?
--- kernel/futex.c | 4 ++-- 1 files changed, 2 insertions(+), 2 deletions(-) diff --git a/kernel/futex.c b/kernel/futex.c index e7a35f1..b0f1b2d 100644 --- a/kernel/futex.c +++ b/kernel/futex.c @@ -252,7 +252,7 @@ again: page = compound_head(page); lock_page(page); - if (!page->mapping) { + if (!page->mapping && !PageHuge(page)) { unlock_page(page); put_page(page); goto again; @@ -265,7 +265,7 @@ again: * it's a read-only handle, it's expected that futexes attach to * the object not the particular process. */ - if (PageAnon(page)) { + if (PageAnon(page) || (PageHuge(page) && !page->mapping)) { key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */ key->private.mm = mm; key->private.address = address;
|  |