I would like to revisit Qu's proposal to not charge btrfs extent_buffer allocations to the user's cgroup. https://lore.kernel.org/linux-mm/b5fef5372ae454a7b6da4f2f75c427aeab6a07d6.1727498749.git.wqu@xxxxxxxx/ I believe it is detrimental to randomly account these global pages to the cgroup using them, basically at random. A bit more justification and explanation in the patches themselves. Three meta-considerations/questions: 1. Which tree should this go through, assuming it is acceptable? For now, I have based it off btrfs/for-next as that is what I am used to doing, but I am happy to re-send it based off the appropriate mm branch. 2. Christoph wrote the first patch as-is in his suggestion to Qu. I am happy to replace it with his authorship/s-o-b, I just didn't want to do that without asking. For now, I put his "Suggested-by". 3. The previous suggestion also requested "proper" documentation. I don't know what that entails in this case, and was unable to find corresponding documentation for filemap_add_folio() in the code or in Documentation/. Please let me know what I should be doing there, as well. Boris Burkov (3): mm/filemap: add filemap_add_folio_nocharge() btrfs: use filemap_add_folio_nocharge() for extent_buffers mm: add vmstat for cgroup uncharged pages fs/btrfs/extent_io.c | 4 ++-- include/linux/mmzone.h | 3 +++ include/linux/pagemap.h | 2 ++ mm/filemap.c | 41 +++++++++++++++++++++++++++++++++++------ mm/vmstat.c | 3 +++ 5 files changed, 45 insertions(+), 8 deletions(-) -- 2.50.1
