Vladimír Čunát authored 5 years ago and Daniel Salzman committed 4 years ago

Tested briefly, to be sure.

Vladimír Čunát authored 5 years ago and Daniel Salzman committed 4 years ago

(before I forget all about it)

Vladimír Čunát authored 5 years ago and Daniel Salzman committed 4 years ago

Libor was running into that.

Vladimír Čunát authored 5 years ago and Daniel Salzman committed 4 years ago

Some of the conditions were clearly mistakes,
but overall I'm unsure about what to do in some cases.

Vladimír Čunát authored 5 years ago and Daniel Salzman committed 4 years ago

No functional changes inside.  This implements one of the TODOs.

It seems suitable to execute the two phases at different points
in time.  In particular, I didn't like the fact that in knotd
_prepare_alloc() was "missing" between poll() and _alloc_packet().

Vladimír Čunát authored 5 years ago and Daniel Salzman committed 4 years ago

I had looked at this case before but I didn't see the problem;
fortunately Dan pointed me to it again.

When we bail out due to error, we can't just __release the indices
that we haven't processed; they would need to do at least
kxsk_dealloc_umem_frame()

Vladimír Čunát authored 5 years ago and Daniel Salzman committed 4 years ago

We need that in kresd, as overwriting the program would detach
all the AF_XDP sockets and only the last kresd could receive.
It's not perfect, but I suppose we can bear that extraneous message.