static_call: Properly initialise DEFINE_STATIC_CALL_RET0()

When a static call is updated with __static_call_return0() as target,
arch_static_call_transform() set it to use an optimised set of
instructions which are meant to lay in the same cacheline.

But when initialising a static call with DEFINE_STATIC_CALL_RET0(),
we get a branch to the real __static_call_return0() function instead
of getting the optimised setup:

c00d8120 <__SCT__perf_snapshot_branch_stack>:
c00d8120: 4b ff ff f4 b c00d8114 <__static_call_return0>
c00d8124: 3d 80 c0 0e lis r12,-16370
c00d8128: 81 8c 81 3c lwz r12,-32452(r12)
c00d812c: 7d 89 03 a6 mtctr r12
c00d8130: 4e 80 04 20 bctr
c00d8134: 38 60 00 00 li r3,0
c00d8138: 4e 80 00 20 blr
c00d813c: 00 00 00 00 .long 0x0

Add ARCH_DEFINE_STATIC_CALL_RET0_TRAMP() defined by each architecture
to setup the optimised configuration, and rework
DEFINE_STATIC_CALL_RET0() to call it:

c00d8120 <__SCT__perf_snapshot_branch_stack>:
c00d8120: 48 00 00 14 b c00d8134 <__SCT__perf_snapshot_branch_stack+0x14>
c00d8124: 3d 80 c0 0e lis r12,-16370
c00d8128: 81 8c 81 3c lwz r12,-32452(r12)
c00d812c: 7d 89 03 a6 mtctr r12
c00d8130: 4e 80 04 20 bctr
c00d8134: 38 60 00 00 li r3,0
c00d8138: 4e 80 00 20 blr
c00d813c: 00 00 00 00 .long 0x0

Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Josh Poimboeuf <jpoimboe@redhat.com>
Link: https://lore.kernel.org/r/1e0a61a88f52a460f62a58ffc2a5f847d1f7d9d8.1647253456.git.christophe.leroy@csgroup.eu

powerpc/32: Add support for out-of-line static calls

Add support for out-of-line static calls on PPC32. This change
improve performance of calls to global function pointers by
using direct calls instead of indirect calls.

The trampoline is initialy populated with a 'blr' or branch to target,
followed by an unreachable long jump sequence.

In order to cater with parallele execution, the trampoline needs to
be updated in a way that ensures it remains consistent at all time.
This means we can't use the traditional lis/addi to load r12 with
the target address, otherwise there would be a window during which
the first instruction contains the upper part of the new target
address while the second instruction still contains the lower part of
the old target address. To avoid that the target address is stored
just after the 'bctr' and loaded from there with a single instruction.

Then, depending on the target distance, arch_static_call_transform()
will either replace the first instruction by a direct 'bl <target>' or
'nop' in order to have the trampoline fall through the long jump
sequence.

For the special case of __static_call_return0(), to avoid the risk of
a far branch, a version of it is inlined at the end of the trampoline.

Performancewise the long jump sequence is probably not better than
the indirect calls set by GCC when we don't use static calls, but
such calls are unlikely to be required on powerpc32: With most
configurations the kernel size is far below 32 Mbytes so only
modules may happen to be too far. And even modules are likely to
be close enough as they are allocated below the kernel core and
as close as possible of the kernel text.

static_call selftest is running successfully with this change.

With this patch, __do_irq() has the following sequence to trace
irq entries:

c0004a00 <__SCT__tp_func_irq_entry>:
c0004a00: 48 00 00 e0 b c0004ae0 <__traceiter_irq_entry>
c0004a04: 3d 80 c0 00 lis r12,-16384
c0004a08: 81 8c 4a 1c lwz r12,18972(r12)
c0004a0c: 7d 89 03 a6 mtctr r12
c0004a10: 4e 80 04 20 bctr
c0004a14: 38 60 00 00 li r3,0
c0004a18: 4e 80 00 20 blr
c0004a1c: 00 00 00 00 .long 0x0
...
c0005654 <__do_irq>:
...
c0005664: 7c 7f 1b 78 mr r31,r3
...
c00056a0: 81 22 00 00 lwz r9,0(r2)
c00056a4: 39 29 00 01 addi r9,r9,1
c00056a8: 91 22 00 00 stw r9,0(r2)
c00056ac: 3d 20 c0 af lis r9,-16209
c00056b0: 81 29 74 cc lwz r9,29900(r9)
c00056b4: 2c 09 00 00 cmpwi r9,0
c00056b8: 41 82 00 10 beq c00056c8 <__do_irq+0x74>
c00056bc: 80 69 00 04 lwz r3,4(r9)
c00056c0: 7f e4 fb 78 mr r4,r31
c00056c4: 4b ff f3 3d bl c0004a00 <__SCT__tp_func_irq_entry>

Before this patch, __do_irq() was doing the following to trace irq
entries:

c0005700 <__do_irq>:
...
c0005710: 7c 7e 1b 78 mr r30,r3
...
c000574c: 93 e1 00 0c stw r31,12(r1)
c0005750: 81 22 00 00 lwz r9,0(r2)
c0005754: 39 29 00 01 addi r9,r9,1
c0005758: 91 22 00 00 stw r9,0(r2)
c000575c: 3d 20 c0 af lis r9,-16209
c0005760: 83 e9 f4 cc lwz r31,-2868(r9)
c0005764: 2c 1f 00 00 cmpwi r31,0
c0005768: 41 82 00 24 beq c000578c <__do_irq+0x8c>
c000576c: 81 3f 00 00 lwz r9,0(r31)
c0005770: 80 7f 00 04 lwz r3,4(r31)
c0005774: 7d 29 03 a6 mtctr r9
c0005778: 7f c4 f3 78 mr r4,r30
c000577c: 4e 80 04 21 bctrl
c0005780: 85 3f 00 0c lwzu r9,12(r31)
c0005784: 2c 09 00 00 cmpwi r9,0
c0005788: 40 82 ff e4 bne c000576c <__do_irq+0x6c>

Behind the fact of now using a direct 'bl' instead of a
'load/mtctr/bctr' sequence, we can also see that we get one less
register on the stack.

Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/6ec2a7865ed6a5ec54ab46d026785bafe1d837ea.1630484892.git.christophe.leroy@csgroup.eu