"David S. Miller" wrote:
>
> Date: Mon, 30 Oct 2000 17:24:24 -0500
> From: Tom Leete <tleete@xxxxxxxxxxxx>
>
> This fixes tests of a socket buffer done without holding the
> lock. tcp_data_wait() and wait_for_tcp_memory() both had
> unguarded refs in their sleep conditionals.
>
> These are not buggy at all, see the discussion which took place here
> over the past few days.
I'll post traces privately. It was my lockups that got Rick
interested.
> Look, if the sleep condition test "races" due to not holding the lock,
> the schedule() just returns because if the sleep condidion test passes
> then by definition this means we were also woken up, see?
Would you explain what is accomplished by toggling the lock
every time through? What breaks by not doing so?
> BTW, while we're on the topic of people not understanding the
> networking locking and proposing bogus patches, does anyone know who
> sent the bogon IP tunneling locking "fixes" to Linus behind my back?
Not me, but see below.
> They were crap too, and I had to revert them in test10-pre7. It's
> another case of people just not understanding how the code works and
> thus that it is correct without any changes.
If it's perfect why can't I use it without locking up the
machine? BTW, I'm currently running my patch. I can now
flood ping 100000 packets in either direction. With
unmodified tcp that is a red switcher (reliably hard locks
all i/o including the serial console).
> Please send such fixes to me, and I'll set you straight with a
> description as to why your change is unnecessary :-)
Perhaps that's why somebody wants to bypass you.
Tom Leete
|