vm: Remove kernel stack swapping support, part 2

[markj]

Summary

After mi_startup() finishes, thread0 becomes the "swapper", whose
responsibility is to swap threads back in on demand. Now that threads
can't be swapped out, there is no use for this thread. Just call
sched_throw() to exit.

[imp]

sys/vm/vm_glue.c

840

Do we need this at all?

[markj]

sys/vm/vm_glue.c

840

No, but fixing up its callers will be a patch of its own. I didn't go that far yet.

[kib]

Does this mean that thread0 is no longer functional? If yes, can we re-purpose it for something/anything instead of throwing?

[markj]

In D46113#1051278, @kib wrote:

Does this mean that thread0 is no longer functional? If yes, can we re-purpose it for something/anything instead of throwing?

Right. Do you have any suggestions there? Hmm, after these changes there is still the vm_daemon(), to handle RACCT_RSS and RLIMIT_RSS. That is probably a good candidate.

[kib]

In D46113#1051367, @markj wrote:

In D46113#1051278, @kib wrote:

Does this mean that thread0 is no longer functional? If yes, can we re-purpose it for something/anything instead of throwing?

Right. Do you have any suggestions there? Hmm, after these changes there is still the vm_daemon(), to handle RACCT_RSS and RLIMIT_RSS. That is probably a good candidate.

This is only for RACCT I think. In the worst case I suggest to just park the thread by blocking it, or doing non-interruptible sleep on local var address.

[markj]

sys/vm/vm_glue.c

840

This is now done in parts 9 and 10.

[markj]

Use tsleep() to park the thread in mi_startup().

[alc]

sys/kern/init_main.c

346

I wouldn't be surprised to find that this results in last being placed on the stack instead of a register. Could __builtin_frame_address(0) be used instead?

[markj]

Apply alc's suggestion.

[kib]

sys/kern/init_main.c

346

There is pause() with the dedicated pause_wchan. Might be loop around pause()?

[markj]

sys/kern/init_main.c

346

With pause(), we have to provide some arbitrary timeout, and a callout will be scheduled, so I slightly prefer the tsleep() call. I don't feel strongly about it and can change it if you prefer.

[kib]

sys/kern/init_main.c

346

We can use pause_wchan at least. But either is fine.