On Tue, 2016-08-02 at 10:18 +0200, Michal Hocko wrote: > On Tue 02-08-16 10:06:12, Oliver Neukum wrote: > > On Mon, 2016-08-01 at 10:20 -0400, Tejun Heo wrote: > > > Hello, > > > > > > If any real IO depends on those devices then this is not sufficient and > > > > they need some form of guarantee for progress (aka mempool). > > > > > > Oliver, Alan, what do you think? If USB itself can't operate without > > > allocating memory during transactions, whatever USB storage drivers > > > > It cannot. The IO must be described to the hardware with a data > > structure in memory. > > > > > are doing isn't all that meaningful. Can we proceed with the > > > workqueue patches? Also, it could be that the only thing GFP_NOIO and > > > GFP_ATOMIC are doing is increasing the chance of IO failures under > > > memory pressure. Maybe it'd be a good idea to reconsider the > > > approach? > > > > We had actual deadlocks with GFP_KERNEL. It seems to me that the SCSI > > layer can deal with IO that cannot be completed due to a lack of memory > > at least somewhat, but a deadlock within a driver would obviously be > > deadly. So I don't think that mempools would remove the need for > > GFP_NOIO as there are places in usbcore we cannot enter the page > > laundering path from. They are an additional need. > > OK, I guess there is some misunderstanding here. I believe that Tejun > wasn't arguing to drop GFP_NOIO. It might be really needed for the dead > lock avoidance. No question about that. The whole point is that > WQ_RECLAIM might be completely pointless because a rescuer wouldn't help > much if the work item would do GFP_NOIO and get stuck in the page > allocator.
But that can be a problem only if the items on the work queue are actually run and without WQ_MEM_RECLAIM that guarantee cannot be made. We can deal with failures of memory allocation. But the requests must actually terminate. Regards Oliver