Alvaro Herrera <alvherre@alvh.no-ip.org> writes:
> All of them have in common that the slotno being passed ($3 below) is in
> SLRU_PAGE_READ_IN_PROGRESS state ... could it be a problem with lock
> reordering? Maybe somebody is trying to read in a page, and somebody
> else steals the buffer from under them. Not sure how likely is that.
It's even more interesting than that: in all three cases,
SlruSelectLRUPage has selected a "least recently used" page that is
still in READ_IN_PROGRESS state (ie, we haven't finished faulting it in)
and is recursively calling SimpleLruReadPage to wait for that condition
to terminate.
Apparently, Jim's setup could desperately do with a larger SLRU arena
for pg_subtrans, because this is supposed to be a never-happen path ---
if you can't finish loading a page before you need its slot for
something else, you are thrashing with a capital T.
I suppose there's a bug in this path, but I'm darned if I can see what
it is. There are a number of obvious inefficiencies, but those
shouldn't be important given that this isn't supposed to happen much.
But how's it getting to the Assert failure?
regards, tom lane