On Thu, Jan 09, 2014 at 10:56:44AM -0500, Alan Stern wrote: > On Thu, 9 Jan 2014, David Laight wrote: > > > From: Alan Stern > > > On Thu, 9 Jan 2014, David Laight wrote: > > > > > > > > From: Sarah Sharp > > > > > Currently prepare_ring() returns -ENOMEM if the urb won't fit into a > > > > > single ring segment. usb_sg_wait() treats this error as a temporary > > > > > condition and will keep retrying until something else goes wrong. > > > > > > > > Even if it were a temporary condition then an immediate retry > > > > is rather pointless. > > > > Retry after a slight delay might make some sense. > > > > > > There is a yield(). It could be changed to something longer. > > > > Will yield allow low priority processes to run? > > No (depending to some extent on the scheduler). > > > > > Of course, if this is the pagedaemon trying to write out > > > > pages in order to generate some free kernel memory all is lost > > > > as soon as the usb subsystem need to allocate memory in order > > > > to perform a transfer. > > > > > > That never happens, because all memory allocation requests on the > > > storage I/O pathway are done using GFP_NOIO. > > > > Picking hairs - not the ones in the xhci driver if, for example, > > it has to extend the rings. > > If that is true, it is a bug in xhci-hcd.
The xHCI driver passes down GFP_ATOMIC to the URB queueing mechanisms, since they're called while holding the xhci spinlock, and that will be used in ring expansion. I believe that's the correct memory flag. Sarah Sharp -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html