It is too early to check the flag, with the current code structure. I set OBJ_SWAPPING in swap_pager_alloc_init() right after vm_object_allocate() call.

I can move setting of the flag into vm_object_allocate().

Should we clear OBJ_SWAP as well? Otherwise the checks of object->type that were converted to checks of object->flags need to be audited to make sure they handle dead objects correctly.

Do you think that such places could exist?

swap_pager_dealloc() is called when object refcount == 0, which for non-vnode objects must imply that no one else except our caller has a pointer to the object.

Hmm, I think we can only reach destroyed swap object through the vm_object_list iteration, and then the only such function would be swap_pager_swapoff(). It rechecks OBJ_DEAD after obtaining object lock.

Ok, I will add the cleaning of the flag.

In fact it triggers from e.g. vm_object_get_vnode(). I removed the assert.

I guess pagerops should be annotated __read_mostly.

It is not obvious to me that the object checks in vm_page_reclaim_run() are sufficient, for example. vm_object_terminate() does not necessarily remove all pages (see the OBJ_PG_DTOR flag). This is a contrived example though, OBJ_PG_DTOR has only a single user today. Really my point is that some state is now spread around multiple fields, so we should try to minimize the set of valid combinations.

To be honest, I somewhat dislike OBJ_SWAP. Instead of having a new object type, did you consider using a tuple <obj->type, obj->flags> to select pager methods? Or, is it really painful to check obj->type == OBJT_SWAP || obj->type == OBJT_SWAP_TMPFS inline?

I think they should be marked const and go into .rodata instead.

Plan is to (try to) make OBJT_SWAP_TMPFS pager kldloadable from tmpfs.ko. Hardcoding the type in sys/vm would be undesirable.