On Mon 11-01-21 17:20:51, Mike Kravetz wrote:

On 1/10/21 4:40 AM, Muchun Song wrote:

quoted

There is a race between dissolve_free_huge_page() and put_page(),
and the race window is quite small. Theoretically, we should return
-EBUSY when we encounter this race. In fact, we have a chance to
successfully dissolve the page if we do a retry. Because the race
window is quite small. If we seize this opportunity, it is an
optimization for increasing the success rate of dissolving page.

If we free a HugeTLB page from a non-task context, it is deferred
through a workqueue. In this case, we need to flush the work.

The dissolve_free_huge_page() can be called from memory hotplug,
the caller aims to free the HugeTLB page to the buddy allocator
so that the caller can unplug the page successfully.

Signed-off-by: Muchun Song <redacted>
---
 mm/hugetlb.c | 26 +++++++++++++++++++++-----
 1 file changed, 21 insertions(+), 5 deletions(-)

I am unsure about the need for this patch.  The code is OK, there are no
issues with the code.

As mentioned in the commit message, this is an optimization and could
potentially cause a memory offline operation to succeed instead of fail.
However, we are very unlikely to ever exercise this code.  Adding an
optimization that is unlikely to be exercised is certainly questionable.

Memory offline is the only code that could benefit from this optimization.
As someone with more memory offline user experience, what is your opinion
Michal?

I am not a great fun of optimizations without any data to back them up.
I do not see any sign this code has been actually tested and the
condition triggered.

Besides that I have requested to have an explanation of why blocking on
the WQ is safe and that hasn't happened.
-- 
Michal Hocko
SUSE Labs