On Thu, Jul 16, 2020 at 03:47:17AM +0100, Matthew Wilcox wrote: > On Thu, Jul 16, 2020 at 11:46:56AM +1000, Dave Chinner wrote: > > And why should we compromise performance on hundreds of millions of > > modern systems to fix an extremely rare race on an extremely rare > > platform that maybe only a hundred people world-wide might still > > use? > > I thought that wasn't the argument here. It was that some future > compiler might choose to do something absolutely awful that no current > compiler does, and that rather than disable the stupid "optimisation", > we'd be glad that we'd already stuffed the source code up so that it > lay within some tortuous reading of the C spec. > > The memory model is just too complicated. Look at the recent exchange > between myself & Dan Williams. I spent literally _hours_ trying to > figure out what rules to follow. > > https://lore.kernel.org/linux-mm/CAPcyv4jgjoLqsV+aHGJwGXbCSwbTnWLmog5-rxD2i31vZ2rDNQ@xxxxxxxxxxxxxx/ > https://lore.kernel.org/linux-mm/CAPcyv4j2+7XiJ9BXQ4mj_XN0N+rCyxch5QkuZ6UsOBsOO1+2Vg@xxxxxxxxxxxxxx/ > > Neither Dan nor I are exactly "new" to Linux kernel development. As Dave > is saying here, having to understand the memory model is too high a bar. > > Hell, I don't know if what we ended up with for v4 is actually correct. > It lokos good to me, but *shrug* > > https://lore.kernel.org/linux-mm/159009507306.847224.8502634072429766747.stgit@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/ Looks like you still got it wrong :-( It needs: diff --git a/drivers/char/mem.c b/drivers/char/mem.c index 934c92dcb9ab..9a95fbe86e15 100644 --- a/drivers/char/mem.c +++ b/drivers/char/mem.c @@ -1029,7 +1029,7 @@ static int devmem_init_inode(void) } /* publish /dev/mem initialized */ - WRITE_ONCE(devmem_inode, inode); + smp_store_release(&devmem_inode, inode); return 0; } It seems one source of confusion is that READ_ONCE() and WRITE_ONCE() don't actually pair with each other, unless no memory barriers are needed at all. Instead, READ_ONCE() pairs with a primitive that has "release" semantics, e.g. smp_store_release() or cmpxchg_release(). But READ_ONCE() is only correct if there's no control flow dependency; if there is, it needs to be upgraded to a primitive with "acquire" semantics, e.g. smp_load_acquire(). The best approach might be to just say that the READ_ONCE() + "release" pairing should be avoided, and we should stick to "acquire" + "release". (And I think Dave may be saying he'd prefer that for ->s_dio_done_wq?) - Eric
