On Sun, Mar 11, 2018 at 10:15 AM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote: > On Sun, Mar 11, 2018 at 4:27 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: >> On Fri, Mar 09, 2018 at 10:55:32PM -0800, Dan Williams wrote: >>> Add a generic facility for awaiting an atomic_t to reach a value of 1. >>> >>> Page reference counts typically need to reach 0 to be considered a >>> free / inactive page. However, ZONE_DEVICE pages allocated via >>> devm_memremap_pages() are never 'onlined', i.e. the put_page() typically >>> done at init time to assign pages to the page allocator is skipped. >>> >>> These pages will have their reference count elevated > 1 by >>> get_user_pages() when they are under DMA. In order to coordinate DMA to >>> these pages vs filesytem operations like hole-punch and truncate the >>> filesystem-dax implementation needs to capture the DMA-idle event i.e. >>> the 2 to 1 count transition). >>> >>> For now, this implementation does not have functional behavior change, >>> follow-on patches will add waiters for these page-idle events. >> >> Argh, no no no.. That whole wait_for_atomic_t thing is a giant >> trainwreck already and now you're making it worse still. >> >> Please have a look here: >> >> https://lkml.kernel.org/r/20171101190644.chwhfpoz3ywxx2m7@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx > > That thread seems to be worried about the object disappearing the > moment it's reference count reaches a target. That isn't the case with > the memmap / struct page objects for ZONE_DEVICE pages. I understand > wait_for_atomic_one() is broken in the general case, but as far as I > can see it works fine specifically for ZONE_DEVICE page busy tracking, > just not generic object lifetime. Ok, that thread is also concerned with cleaning up the wait_for_atomic_* pattern to also do something more idiomatic with wait_event(). I agree that would be better, but I'm running short of time to go refactor this aou for 4.17 inclusion, especially as I expect another couple rounds of review on this more urgent data corruption fix series that depends on this new api. I think the addition of wait_for_atomic_one() makes it clear that we need a way to pass a conditional expression rather than create a variant api for each different condition. Can you help me out with an attempt of your own, or at least point in a direction that you would accept for solving the "Except the current wait_event() doesn't do the whole key part that makes the hash-table 'work'." problem that you highlighted?
