Hi, I have tried to look into Linus's suggestions to reduce slab memcg accounting overhead [1] [2]. The reorganized hooks are in Patch 1 and it definitely seems like nice cleanup on its own. In Patch 2 I have tried to move them to mm/memcontrol.c to reduce calls to memcg code. I hoped to see better performance, but probably didn't. Patch 3 introduces the suggested kmem_cache_charge() API and Patch 4 tries to use it for the testcase in [1] but it's unfinished due to my lack of VFS knowledge. I haven't done much benchmarking yet, just in a guest VM on my desktop for the test case from [1]. Applying patches 1+2 might have improved it slightly, but could be noise. With 3+4 the memcg overhead is gone as expected (the charging never happens) but due to the unfinished state I don't know yet if the separation might hurt cases where the open() actually succeeds. Anyway thought I would share already so others can play with it and see if it's a good direction to pursue (with patches 3+4). I think Patch 1 should be good to apply in any case (after more testing, and review), not yet sure about Patch 2. [1] https://lore.kernel.org/all/CAHk-=whYOOdM7jWy5jdrAm8LxcgCMFyk2bt8fYYvZzM4U-zAQA@xxxxxxxxxxxxxx/ [2] https://lore.kernel.org/all/CAHk-=whw936qzDLBQdUz-He5WK_0fRSWwKAjtbVsMGfX70Nf_Q@xxxxxxxxxxxxxx/ Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx> --- Vlastimil Babka (4): mm, slab: move memcg charging to post-alloc hook mm, slab: move slab_memcg hooks to mm/memcontrol.c mm, slab: introduce kmem_cache_charge() UNFINISHED mm, fs: use kmem_cache_charge() in path_openat() fs/file_table.c | 9 +- fs/internal.h | 1 + fs/namei.c | 4 +- include/linux/slab.h | 10 +++ mm/memcontrol.c | 90 ++++++++++++++++++++ mm/slab.h | 10 +++ mm/slub.c | 231 +++++++++++++++------------------------------------ 7 files changed, 188 insertions(+), 167 deletions(-) --- base-commit: 6613476e225e090cc9aad49be7fa504e290dd33d change-id: 20240229-slab-memcg-ae6b3789c924 Best regards, -- Vlastimil Babka <vbabka@xxxxxxx>
