On Mon, 10 Feb 2020 at 21:28, Qian Cai <cai@xxxxxx> wrote: > > On Mon, 2020-02-10 at 11:21 -0800, Matthew Wilcox wrote: > > On Mon, Feb 10, 2020 at 02:20:48PM -0500, Qian Cai wrote: > > > On Mon, 2020-02-10 at 09:25 -0800, Matthew Wilcox wrote: > > > > On Mon, Feb 10, 2020 at 12:00:29PM -0500, Qian Cai wrote: > > > > > @@ -2622,7 +2622,7 @@ void filemap_map_pages(struct vm_fault *vmf, > > > > > if (page->index >= max_idx) > > > > > goto unlock; > > > > > > > > > > - if (file->f_ra.mmap_miss > 0) > > > > > + if (data_race(file->f_ra.mmap_miss > 0)) > > > > > file->f_ra.mmap_miss--; > > > > > > > > How is this safe? Two threads can each see 1, and then both decrement the > > > > in-memory copy, causing it to end up at -1. > > > > > > Well, I meant to say it is safe from *data* races rather than all other races, > > > but it is a good catch for the underflow cases and makes some sense to fix them > > > together (so we don't need to touch the same lines over and over again). > > > > My point is that this is a legitimate warning from the sanitiser. > > The point of your patches should not be to remove all the warnings! > > The KCSAN will assume the write is "atomic" if it is aligned and within word- > size which is the case for "ra->mmap_miss", so I somehow skip auditing the > locking around the concurrent writers, but I got your point. Next time, I'll > spend a bit more time looking. Note: the fact that we assume writes aligned up to word-size are atomic is based on current preferences we were told about. Just because the tool won't complain right now (although a simple config switch will make it complain again), we don't want to forget the writes entirely. If it is a simple write, do the WRITE_ONCE if it makes sense. I, for one, still can't prove if all compilers won't screw up a write due to an omitted WRITE_ONCE somehow. [Yes, for more complex ops like 'var++', turning them into 'WRITE_ONCE(var, var + 1)' isn't as readable, so these are a bit tricky until we get primitives to properly deal with them.]
