On 23/06/21(Wed) 23:03, Jonathan Matthew wrote: > On Wed, Jun 23, 2021 at 09:37:10AM +0200, Martin Pieuchot wrote: > > On 16/06/21(Wed) 11:26, Martin Pieuchot wrote: > > > Diff below does two things: > > > > > > - Use atomic operations for incrementing/decrementing references of > > > anonymous objects. This allows us to manipulate them without holding > > > the KERNEL_LOCK(). > > > > > > - Rewrite the loop from uao_swap_off() to only keep a reference to the > > > next item in the list. This is imported from NetBSD and is necessary > > > to introduce locking around uao_pagein(). > > > > > > ok? > > > > Anyone? > > uao_reference_locked() and uao_detach_locked() are prototyped in > uvm_extern.h, so they should be removed here too.
Thanks, I'll do that. > It doesn't look like uao_detach() is safe to call without the > kernel lock; it calls uao_dropswap() for each page, which calls > uao_set_swslot(), which includes a KERNEL_ASSERT_LOCKED(). > Should we keep the KERNEL_ASSERT_LOCKED() in uao_detach()? I prefer to keep the KERNEL_ASSERT_LOCKED() where it is needed and not spread it to all the callers. My current plan is to trade those assert by assertions on the vmobjlock so I don't want to add new ones.
