Re: NET fails

vandys@cisco.com writes:

>You'll find some notes on this under "Threads race for VM" in the
>(recent)
>mail archives. I'm thinking that the real/ultimate fix will be to use
>clock_slot(), and letting the fault resolution drop back out to
>usermode to
>restart the instruction. This will be the most efficient for the
>non-error
>case, which is the one we want to favor. Other fixes would be to scan
>lists
>and such, but that costs extra CPU cycles.

I looked at the code again and I was thinking something
similar. I was going to have lock_slot() return a value
to indicate whether it had to sleep in order to lock the
slot, and if so, vas_fault would simply try the instruction
again. We probably wouldn't want to use clock_slot()
alone, because that would mean the thread would remain
continuously generating faults and looking up the page
while the I/O pages in.
Received on Tue Nov 2 07:51:31 1999