On Tue 20-09-16 10:37:04, Mike Kravetz wrote: > On 09/20/2016 08:53 AM, Gerald Schaefer wrote: > > dissolve_free_huge_pages() will either run into the VM_BUG_ON() or a > > list corruption and addressing exception when trying to set a memory > > block offline that is part (but not the first part) of a gigantic > > hugetlb page with a size > memory block size. > > > > When no other smaller hugepage sizes are present, the VM_BUG_ON() will > > trigger directly. In the other case we will run into an addressing > > exception later, because dissolve_free_huge_page() will not use the head > > page of the compound hugetlb page which will result in a NULL hstate > > from page_hstate(). list_del() would also not work well on a tail page. > > > > To fix this, first remove the VM_BUG_ON() because it is wrong, and then > > use the compound head page in dissolve_free_huge_page(). > > > > However, this all assumes that it is the desired behaviour to remove > > a (gigantic) unused hugetlb page from the pool, just because a small > > (in relation to the hugepage size) memory block is going offline. Not > > sure if this is the right thing, and it doesn't look very consistent > > given that in this scenario it is _not_ possible to migrate > > such a (gigantic) hugepage if it is in use. OTOH, has_unmovable_pages() > > will return false in both cases, i.e. the memory block will be reported > > as removable, no matter if the hugepage that it is part of is unused or > > in use. > > > > This patch is assuming that it would be OK to remove the hugepage, > > i.e. memory offline beats pre-allocated unused (gigantic) hugepages. > > > > Any thoughts? > > Cc'ed Rui Teng and Dave Hansen as they were discussing the issue in > this thread: > https://lkml.org/lkml/2016/9/13/146 > > Their approach (I believe) would be to fail the offline operation in > this case. However, I could argue that failing the operation, or > dissolving the unused huge page containing the area to be offlined is > the right thing to do. I am sorry I have noticed this thread only now. I was arguing about this in the original thread. I would be rather reluctant to free gigantic page just because somebody wants to offline a small part of it because setup is really expensive and a lost page would be really hard to get back. I would even question the per page block offlining itself. Why would anybody want to offline few blocks rather than the whole node? What is the usecase here? -- Michal Hocko SUSE Labs -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>