Differential D2277
Export a list of VM objects via a sysctl. Authored by jhb on Apr 10 2015, 8:22 PM. Tags None Referenced Files
Details
Export a list of VM objects in the system via a sysctl. The list This is a port of the vm_objects kld/program I have had at
Diff Detail
Event TimelineComment Actions I don't know why arc thinks that I copied those two manpages on top of each other. I just added the new page as a new file in git.
Comment Actions
Comment Actions I'll make a tangential remark. Given that vm objects are allocated from a NOFREE zone, it actually strikes me as silly that we keep adding and removing them to and from the global list. Either we decide that iteration over the list is not performance critical and accept the cost of iterating over a lot of dead objects or, perhaps, implement some kind of lazy removal, i.e., during iteration over the list remove dead objects and during allocation insert if necessary.
Comment Actions I also decided to skip OBJ_DEAD objects entirely. They don't really serve any useful purpose to list and this would allow us to consider Alan's proposal in the future without need for any further changes.
Comment Actions I got a LOR when I tried this due to vm_object_list and vm_object. I added the reverse order to WITNESS to tease out where it occurs and this is what I got: lock order reversal:
The problem here being that vm_object_collapse() destroys backing_object while 'object' is locked. I'm not sure how best to rectify this. Even if I use a marker object to walk the list, I believe I still need the vm_object_list mutex to keep each object stable so I can grab a reference, but that alone would trigger the LOR. I assume it wouldn't really be safe to change vm_object_collapse() to bump object's ref count and drop the lock around destroy. OTOH, given that it is a NOFREE zone, I could adopt Alan's suggestion of leaving DEAD objects on the list and only adding objects to the list when they are originally allocated. That would remove the TAILQ_REMOVE from vm_object_destroy() and thus remove the LOR. Comment Actions So you are going to restore the trylock for the commit ? I think that the change to keep even unallocated objects on the list should be separate. Comment Actions John, how do you want to proceed? I see two possibilities. 1. Commit a revised version of this patch that uses trylock, and implement lazy dead object removal later. 2. Wait for the implementation of lazy dead object removal to be committed first, and then update this patch. I'm swamped right now. If you want to take option #2, could one of you implement the lazy dead object removal? Comment Actions Ah, I've just been preempted by other stuff. I already have a patch I haven't yet uploaded (or tested) that changes the list to never remove objects and only add new ones to the list when they are init'd. I can split that out into a separate change and have that go in first before this one. Comment Actions Rebased on newer HEAD after the object list changes.
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||