On 5/25/21 5:33 PM, Jann Horn wrote: > On Tue, May 25, 2021 at 1:40 AM Vlastimil Babka <vbabka@xxxxxxx> wrote: >> In put_cpu_partial, we need a stable cpu, but being preempted is not an issue. >> So, disable migration instead of preemption. > > I wouldn't say "not an issue", more like "you're not making it worse". > > From what I can tell, the following race can already theoretically happen: > > task A: put_cpu_partial() calls preempt_disable() > task A: oldpage = this_cpu_read(s->cpu_slab->partial) > interrupt: kfree() reaches unfreeze_partials() and discards the page > task B (on another CPU): reallocates page as page cache > task A: reads page->pages and page->pobjects, which are actually > halves of the pointer page->lru.prev > task B (on another CPU): frees page > interrupt: allocates page as SLUB page and places it on the percpu partial list > task A: this_cpu_cmpxchg() succeeds Oops, nice find. Thanks. > which would cause page->pages and page->pobjects to end up containing > halves of pointers that would then influence when put_cpu_partial() > happens and show up in root-only sysfs files. Maybe that's acceptable, > I don't know. But there should probably at least be a comment for now > to point out that we're reading union fields of a page that might be > in a completely different state. > > (Someone should probably fix that code sometime and get rid of > page->pobjects entirely, given how inaccurate it is...) I'll try to address it separately later. Probably just target a number of pages, instead of objects, on the list and store the number as part of struct kmem_cache_cpu, not struct page. The inaccuracy leading to potentially long lists is a good reason enough, the race scenario above is another one...
