On Fri, Sep 18, 2020 at 09:20:09PM -0700, Andrew Morton wrote:
> LRU page reclaim always splits the shmem huge page first: I'd prefer not
> to demand that of i915, so check and split compound in shmem_writepage().
Sorry for not checking this earlier, but I don't think this is right.
for (i = 0; i < obj->base.size >> PAGE_SHIFT; i++) {
...
if (!page_mapped(page) && clear_page_dirty_for_io(page)) {
...
ret = mapping->a_ops->writepage(page, &wbc);
so we cleared the dirty bit on the entire hugepage, but then split
it after clearing the dirty bit, so the subpages are now not dirty.
I think we'll lose writes as a result? At least we won't swap pages
out that deserve to be paged out.
>
> - VM_BUG_ON_PAGE(PageCompound(page), page);
> + /*
> + * If /sys/kernel/mm/transparent_hugepage/shmem_enabled is "force",
> + * then drivers/gpu/drm/i915/gem/i915_gem_shmem.c gets huge pages,
> + * and its shmem_writeback() needs them to be split when swapping.
> + */
> + if (PageTransCompound(page))
> + if (split_huge_page(page) < 0)
> + goto redirty;
> +
> BUG_ON(!PageLocked(page));
> mapping = page->mapping;
> index = page->index;
> _
>
