On Tue, Feb 18, 2020 at 11:25 AM Mina Almasry [off-list ref] wrote:

On Tue, Feb 18, 2020 at 11:14 AM Mike Kravetz [off-list ref] wrote:

quoted

On 2/18/20 10:35 AM, Mina Almasry wrote:

quoted

On Tue, Feb 18, 2020 at 6:21 AM Qian Cai [off-list ref] wrote:

quoted

On Tue, 2020-02-11 at 15:19 -0800, Andrew Morton wrote:

quoted

On Tue, 11 Feb 2020 13:31:20 -0800 Mina Almasry [off-list ref] wrote:

[ 7933.806377][T14355] ------------[ cut here ]------------
[ 7933.806541][T14355] kernel BUG at mm/hugetlb.c:490!
VM_BUG_ON(t - f <= 1);
[ 7933.806562][T14355] Oops: Exception in kernel mode, sig: 5 [#1]

<snip>

quoted

Hi Qian,

Yes this VM_BUG_ON was added by a patch in the series ("hugetlb:
disable region_add file_region coalescing") so it's definitely related
to the series. I'm taking a look at why this VM_BUG_ON fires. Can you
confirm you reproduce this by running hugemmap06 from the ltp on a
powerpc machine? Can I maybe have your config?

Thanks!

Hi Mina,

Looking at the region_chg code again, we do a

        resv->adds_in_progress += *out_regions_needed;

and then potentially drop the lock to allocate the needed entries.  Could
anopther thread (only adding reservation for a single page) then come in
and notice that there are not enough entries in the cache and hit the
VM_BUG_ON()?

Maybe. Also I'm thinking the code thinks actual_regions_needed >=
in_regions_needed, but that doesn't seem like a guarantee. I think
this call sequence with the same t->f range would violate that:

region_chg (regions_needed=1)
region_chg (regions_needed=1)
region_add (fills in the range)
region_add (in_regions_needed = 1, actual_regions_needed = 0, so
assumptions in the code break).

Luckily it seems the ltp readily reproduces this, so I'm working on
reproducing it. I should have a fix soon, at least if I can reproduce
it as well.

I had a bit of trouble reproducing this but I got it just now.

Makes sense I've never run into this even though others can readily
reproduce it. I happen to run my kernels on a pretty beefy 36 core
machine and in that setup things seem to execute fast and there is
never a queue of pending file_region inserts into the resv_map. Once I
limited qemu to only use 2 cores I ran into the issue right away.
Looking into a fix now.