On 2/16/21 6:49 PM, Mike Rapoport wrote:

Hi Vlastimil,

On Tue, Feb 16, 2021 at 05:39:12PM +0100, Vlastimil Babka wrote:

quoted

So, Andrea could you please check if this fixes the original
fast_isolate_around() issue for you? With the VM_BUG_ON not removed, DEBUG_VM
enabled, no changes to struct page initialization...
It relies on pageblock_pfn_to_page as the rest of the compaction code.

Pardon my ignorance of compaction internals, but does this mean that with
your patch we'll never call set_pfnblock_flags_mask() for a pfn in a hole?

No it doesn't mean that kind of guarantee. But we will not call it anymore (if
my patch is correct) from a path which we currently know it's doing that and
triggering the VM_BUG_ON. So that's a targetted fix that matches stable backport
criteria. It doesn't contradict your patch as a way to improve mainline, I still
agree it's best long-term if we initialize the struct pages without such
surprises. But I also agree with Michal that there's a risk of replacing one
corner case with another and thus we shouldn't do that as a stable fix.