On Tue, 30 Oct 2018, Mykola Golub wrote: > On Mon, Oct 29, 2018 at 07:35:59PM +0000, Sage Weil wrote: > > > Good analysis. This reminds me of > > 907b6281e99ece3677dd7b012cf4955731db6120, which I think addresses > > this issue. In particular, it waits for applied_sync, which is after the > > write is applied but before the other callbacks fire (which may block on > > pg lock, as you note). So I think this is resolved now in mimic... > > Thanks! I suppose by "it waits for applied_sync" you mean > ea531df216c65a3773a55b5f6c42af20e5004263, which is additionally needed > here. Ooops! Yeah I meant to paste ea531df21. > So, it looks like to fix this on jewel, I need to substitute > `osd->store->apply_transaction` with `osd->store->queue_transaction` > in PG::_scan_snaps and backport ea531df216c65a3773a55b5f6c42af20e5004263. Right. > On luminous ea531df216c65a3773a55b5f6c42af20e5004263 is already > backported, though it looks like does not have much sence right now, until > `apply_transaction` is replaced with `queue_transaction`. > > Do we want PRs for these? Yes, please! > Or would you prefer if > 907b6281e99ece3677dd7b012cf4955731db6120 was backported to luminous > (and jewel)? No, that is a big change that shouldn't be backported. Thanks! sage
