Did you actually observed a case when the the VOP was called with vp->v_mount == NULL ?

You do not need a nested block there.

} else if (vp->v_iflag & VI_OWEINACT) {
   vdefer_inactive(vp);
} else {
   vdropl(vp);
}

It is strange that you do not check VV_NOSYNC there. Once the flag is set, it is not cleared.

And I do not like VOP_ISLOCKED() there, as before. It allows the rare and then esp. hard to note race where vnode' pages could be left unsynced for at least one syncer run. I do not see why not use trylock.

This is the part that I dislike most from the whole patch. It blows struct vnode by 16 bytes, which scales to additional tens of MBs of kernel memory for typical machine.

Why don't you reused v_actfreelist ?

The internal fragmentation in the slab means there's more space here before we actually consume more physical memory. It is a step function on PAGE_SIZE / sizeof(vnode). So we shouldn't gratuitously blow it out and there are ways we could make it smaller fairly reasonably but every pointer doesn't translate to increased physical memory.

We are looking at multi-page slabs for some of these larger zones to control fragmentation better but unless you make ridiculous allocation sizes you don't get a lot of benefit.

Yes, stress2 managed to panic the patchset without this. I don't remember the backtrace.

ok

The check in the filter is just a copy-paste of what msync was doing in the loop. I can add VV_NOSYNC no problem.

Similarly, VOP_ISLOCKED was there. I think it's rather iffy and I'm more than happy to just remove it.

I should have noted this. According to my conversations with jeff there was unused space in the slab.

With the D22997 the active list is retired and instead global vnode linkage is introduced (making 3 linkages in total: all allocated managed LRU, assigned to a mount point, dirty list on a mount point)

As for sizeof itself, I plan to shrink the struct later. v_tag is almost unused (the only consumer can be easily patched away), flag fields are bigger than they need to be and fit in the hole at the beginning of the struct. Note there are several ufs-isms in the vnode perhaps should be moved to something else so that othe filesystems don't have to pay for it.

It would be very helpful to add multi-page slabs. When I was writing my book some years ago I did some tests and found that nearly 10% of wired memory in the kernel was wasted space due to bad packing of slabsAt that time vnodes were the worst offenders followed by processes. I note that Solaris supports multi-page slabs and has much less wasted kernel memory.