On Wed, Feb 18, 2004 at 02:45:02PM +0100, Lars Marowsky-Bree wrote: > On 2004-02-18T13:11:04,> > Would this not be an option for a _carefully written_ user-space tester > too? I think we have to see if it's needed. If you are careful you can perform io from userland without requiring any kernel memory allocations (io. use raw io and open the device in advance). Whether this is true for the typical test instructions I don't know, you would hope so given they don't carry any data. If we do need something like PF_MEMALLOC, then I think we'll have to export a flag that specifically means 'userland program that is on/near the io path'. Rather than stealing a kernel internal flag. There will always be a race between erroring an io and testing failed paths that may have recovered. This was the 'window of stupidity' that I was trying so hard to reduce in the kernel test paths. By moving testing to userland we have hugely widened this window - but I still think it will be ok. How many paths all fail and recover within a couple of seconds of each other ? We talked about holding onto ios when there were no valid paths at Ottawa, and I thought we all agreed that it just shouldn't be done, because of the memory exhaustion deadlock reasons that Heinz is fretting about. - Joe