[RFC,00/12] Deal with TM on kernel entry and exit

This is very much a proof of concept and if it isn't clear from the
commit names, still a work in progress.

I believe I have something that works - all the powerpc selftests
pass. I would like to get some eyes on it to a) see if I've missed
anything big and b) some opinions on if it is looking like a net
improvement.

Obviously it is still a bit rough around the edges, I'll have to
convince myself that the SPR code is correct. I don't think the
TM_KERNEL_ENTRY macro needs to check that we came from userspace, if
TM is on then we can probably assume. Maybe a check not in the
fastpath. Some of the BUG_ON()s will probably go.

Background:
Currently TM is dealt with when we need to. That is, when we switch
processes, we'll (if nessesary) reclaim the outgoing process and (if
nessesary) recheckpoint the incoming process. Same with signals, if we
need to deliver a signal, we'll ensure we've reclaimed in order to
have all the information and go from there.
I, along with some others got curious to see what it would look like if
we did the 'opposite'.
At all kernel entry points that won't simply just zoom straight to an
RFID we now check if the thread was transactional and do the reclaim.
Correspondingly do the recheckpoint quite late on exception exit. It
turns out we already had a lot of the code pathes set up on the exit
path as there were things that TM had special cased on exit already.
I wasn't sure it it would lead to more or less complexity and though
I'd have to try it to see. I feel like it was almost a win but SPRs
did add some annoying caveats.

In order to get this past Michael I'm going to prove it performs, or
rather, doesn't slow anything down - workload suggestions welcome.

Thanks,

Cyril Bur (12):
  powerpc/tm: Remove struct thread_info param from tm_reclaim_thread()
  selftests/powerpc: Fix tm.h helpers
  selftests/powerpc: Add tm-signal-drop-transaction TM test
  selftests/powerpc: Use less common thread names
  [WIP] powerpc/tm: Reclaim/recheckpoint on entry/exit
  [WIP] powerpc/tm: Remove dead code from __switch_to_tm()
  [WIP] powerpc/tm: Add TM_KERNEL_ENTRY in more delicate exception
    pathes
  [WIP] powerpc/tm: Fix *unavailable_tm exceptions
  [WIP] powerpc/tm: Tweak signal code to handle new reclaim/recheckpoint
    times
  [WIP] powerpc/tm: Correctly save/restore checkpointed sprs
  [WIP] powerpc/tm: Afterthoughts
  [WIP] selftests/powerpc: Remove incorrect tm-syscall selftest

 arch/powerpc/include/asm/exception-64s.h           |  25 ++++
 arch/powerpc/kernel/entry_64.S                     |  20 ++-
 arch/powerpc/kernel/exceptions-64s.S               |  31 ++++-
 arch/powerpc/kernel/process.c                      | 145 ++++++++++++++++++---
 arch/powerpc/kernel/ptrace.c                       |   9 +-
 arch/powerpc/kernel/signal.c                       |  11 +-
 arch/powerpc/kernel/signal_32.c                    |  16 +--
 arch/powerpc/kernel/signal_64.c                    |  41 ++++--
 arch/powerpc/kernel/traps.c                        |   3 -
 tools/testing/selftests/powerpc/tm/Makefile        |   5 +-
 .../powerpc/tm/tm-signal-drop-transaction.c        |  74 +++++++++++
 .../testing/selftests/powerpc/tm/tm-syscall-asm.S  |  28 ----
 tools/testing/selftests/powerpc/tm/tm-syscall.c    | 106 ---------------
 .../testing/selftests/powerpc/tm/tm-unavailable.c  |   4 +-
 tools/testing/selftests/powerpc/tm/tm.h            |  10 +-
 15 files changed, 319 insertions(+), 209 deletions(-)
 create mode 100644 tools/testing/selftests/powerpc/tm/tm-signal-drop-transaction.c
 delete mode 100644 tools/testing/selftests/powerpc/tm/tm-syscall-asm.S
 delete mode 100644 tools/testing/selftests/powerpc/tm/tm-syscall.c

Hi Cyril,

On 02/19/2018 09:22 PM, Cyril Bur wrote:
> This is very much a proof of concept and if it isn't clear from the
> commit names, still a work in progress.

> I believe I have something that works - all the powerpc selftests
> pass. I would like to get some eyes on it to a) see if I've missed
> anything big and b) some opinions on if it is looking like a net
> improvement.

I started to look at this patchset. The patchset apply cleanly on top of the
current kernel and I started to run some tests.

It seems there is a different behavior when there is a trap (as a illegal
instruction or a 'trap' instruction) inside the transaction.

In this case, the signal handler does not seem to be called, and and the task
segfaults. On current upstream, the signal handler is called and the program
can continue.

4.17 pristine
-------------
	$ ./illegal
	Failure

4.17 plus your patches
----------------------
	$ ./illegal
	[1]    2504 segmentation fault  ./illegal


Here is a minimal example that is able to recreate this behaviour:

	#include <stdio.h>
	#include <signal.h>
	
	int htm(){
		asm goto ("tbegin.  		\n\t"
			  "beq %l[failure]	\n\t"
			  "li 3, 3		\n\t"
			  "trap 		\n\t"
			  "tend. 		\n\t"
			  : : : : failure);
	
		return 0;
	failure:
		printf("Failure\n");
		return 1;
	}
	
	void signal_handler(int signo, siginfo_t *si, void *data) {
		// Do nothing
	}
	
	int main(){
		struct sigaction sa;
	        sa.sa_flags = SA_SIGINFO;
	        sa.sa_sigaction = signal_handler;
	
	        sigaction(SIGTRAP, &sa, NULL);
	        sigaction(SIGILL,  &sa, NULL);
	
		return htm();
	}