Skip to contentSkip to conversation

powerpc/powernv: move OPAL call wrapper tracing and interrupt handling to C

PATCH Applied Nicholas Piggin 3 participants 3 messages Feb 26, 2019 – Feb 28, 2019 PowerPC

lore
Share & tools
lore

Nicholas Piggin

author reviewer
#0 lore raw

PATCH 1/1 powerpc/powernv: move OPAL call wrapper tracing and interrupt handling to C

The OPAL call wrapper gets interrupt disabling wrong. It disables interrupts just by clearing MSR[EE], which has two problems:

- It doesn't call into the IRQ tracing subsystem, which means tracing across OPAL calls does not always notice IRQs have been disabled.

- It doesn't go through the IRQ soft-mask code, which causes a minor bug. MSR[EE] can not be restored by saving the MSR then clearing MSR[EE], because a racing interrupt while soft-masked could clear MSR[EE] between the two steps. This can cause MSR[EE] to be incorrectly enabled when the OPAL call returns. Fortunately that should only result in another masked interrupt being taken to disable MSR[EE] again, but it's a bit sloppy.

The existing code also saves MSR to PACA, which is not re-entrant if there is a nested OPAL call from different MSR contexts, which can happen these days with SRESET interrupts on bare metal.

To fix these issues, move the tracing and IRQ handling code to C, and call into asm just for the low level call when everything is ready to go. Save the MSR on stack rather than PACA.

Performance cost is kept to a minimum with a few optimisations:

- The endian switch upon return is combined with the MSR restore, which avoids an expensive context synchronizing operation for LE kernels. This makes up for the additional mtmsrd to enable interrupts with local_irq_enable().

- blr is now used to return from the opal_* functions that are called as C functions, to avoid link stack corruption. This requires a skiboot fix as well to keep the call stack balanced.

A NULL call is more costly after this, (410ns->430ns on POWER9), but OPAL calls are generally not performance critical at this scale.

---
 arch/powerpc/include/asm/asm-prototypes.h     |  10 +-
 arch/powerpc/platforms/powernv/Makefile       |   5 +-
 arch/powerpc/platforms/powernv/opal-call.c    | 283 +++++++++++++++
 .../powerpc/platforms/powernv/opal-wrappers.S | 343 ++----------------
 4 files changed, 328 insertions(+), 313 deletions(-)
 create mode 100644 arch/powerpc/platforms/powernv/opal-call.c
arch/powerpc/include/asm/asm-prototypes.h +4 -6
index 1d911f68a23b..6c67b4bef854 100644
@@ -37,13 +37,11 @@ void kexec_copy_flush(struct kimage *image);
 extern struct static_key hcall_tracepoint_key;
 void __trace_hcall_entry(unsigned long opcode, unsigned long *args);
 void __trace_hcall_exit(long opcode, long retval, unsigned long *retbuf);
-/* OPAL tracing */
-#ifdef CONFIG_JUMP_LABEL
-extern struct static_key opal_tracepoint_key;
-#endif
 
-void __trace_opal_entry(unsigned long opcode, unsigned long *args);
-void __trace_opal_exit(long opcode, unsigned long retval);
+/* OPAL */
+int64_t __opal_call(int64_t a0, int64_t a1, int64_t a2, int64_t a3,
+		    int64_t a4, int64_t a5, int64_t a6, int64_t a7,
+		    int64_t opcode, uint64_t msr);
 
 /* VMX copying */
 int enter_vmx_usercopy(void);
arch/powerpc/platforms/powernv/Makefile +2 -3
index b540ce8eec55..da2e99efbd04 100644
@@ -1,6 +1,6 @@
 # SPDX-License-Identifier: GPL-2.0
-obj-y			+= setup.o opal-wrappers.o opal.o opal-async.o idle.o
-obj-y			+= opal-rtc.o opal-nvram.o opal-lpc.o opal-flash.o
+obj-y			+= setup.o opal-call.o opal-wrappers.o opal.o opal-async.o
+obj-y			+= idle.o opal-rtc.o opal-nvram.o opal-lpc.o opal-flash.o
 obj-y			+= rng.o opal-elog.o opal-dump.o opal-sysparam.o opal-sensor.o
 obj-y			+= opal-msglog.o opal-hmi.o opal-power.o opal-irqchip.o
 obj-y			+= opal-kmsg.o opal-powercap.o opal-psr.o opal-sensor-groups.o
@@ -11,7 +11,6 @@ obj-$(CONFIG_CXL_BASE)	+= pci-cxl.o
 obj-$(CONFIG_EEH)	+= eeh-powernv.o
 obj-$(CONFIG_PPC_SCOM)	+= opal-xscom.o
 obj-$(CONFIG_MEMORY_FAILURE)	+= opal-memory-errors.o
-obj-$(CONFIG_TRACEPOINTS)	+= opal-tracepoints.o
 obj-$(CONFIG_OPAL_PRD)	+= opal-prd.o
 obj-$(CONFIG_PERF_EVENTS) += opal-imc.o
 obj-$(CONFIG_PPC_MEMTRACE)	+= memtrace.o
arch/powerpc/platforms/powernv/opal-call.c new file +283 -0

283 lines not on this page · raw message

arch/powerpc/platforms/powernv/opal-wrappers.S +39 -304

356 lines not on this page · raw message

Signature · 1 line
2.18.0
#1 lore raw
Quoted 29 lines · Nicholas Piggin

The OPAL call wrapper gets interrupt disabling wrong. It disables interrupts just by clearing MSR[EE], which has two problems:

- It doesn't call into the IRQ tracing subsystem, which means tracing across OPAL calls does not always notice IRQs have been disabled.

- It doesn't go through the IRQ soft-mask code, which causes a minor bug. MSR[EE] can not be restored by saving the MSR then clearing MSR[EE], because a racing interrupt while soft-masked could clear MSR[EE] between the two steps. This can cause MSR[EE] to be incorrectly enabled when the OPAL call returns. Fortunately that should only result in another masked interrupt being taken to disable MSR[EE] again, but it's a bit sloppy.

The existing code also saves MSR to PACA, which is not re-entrant if there is a nested OPAL call from different MSR contexts, which can happen these days with SRESET interrupts on bare metal.

To fix these issues, move the tracing and IRQ handling code to C, and call into asm just for the low level call when everything is ready to go. Save the MSR on stack rather than PACA.

Performance cost is kept to a minimum with a few optimisations:

- The endian switch upon return is combined with the MSR restore, which avoids an expensive context synchronizing operation for LE kernels. This makes up for the additional mtmsrd to enable interrupts with local_irq_enable().

Quoted 6 lines

- blr is now used to return from the opal_* functions that are called as C functions, to avoid link stack corruption. This requires a skiboot fix as well to keep the call stack balanced.

A NULL call is more costly after this, (410ns->430ns on POWER9), but OPAL calls are generally not performance critical at this scale.

At least with this patch we can start to better measure things, and that's a big plus. Also, When I get to start worrying about <1ms OPAL calls, I think we can revisit optimising this path :)

Signature · 2 lines
Stewart Smith
OPAL Architect, IBM.
#2 lore raw
In reply to Nicholas Piggin
Quoted 29 lines · Nicholas Piggin

The OPAL call wrapper gets interrupt disabling wrong. It disables interrupts just by clearing MSR[EE], which has two problems:

- It doesn't call into the IRQ tracing subsystem, which means tracing across OPAL calls does not always notice IRQs have been disabled.

- It doesn't go through the IRQ soft-mask code, which causes a minor bug. MSR[EE] can not be restored by saving the MSR then clearing MSR[EE], because a racing interrupt while soft-masked could clear MSR[EE] between the two steps. This can cause MSR[EE] to be incorrectly enabled when the OPAL call returns. Fortunately that should only result in another masked interrupt being taken to disable MSR[EE] again, but it's a bit sloppy.

The existing code also saves MSR to PACA, which is not re-entrant if there is a nested OPAL call from different MSR contexts, which can happen these days with SRESET interrupts on bare metal.

To fix these issues, move the tracing and IRQ handling code to C, and call into asm just for the low level call when everything is ready to go. Save the MSR on stack rather than PACA.

Performance cost is kept to a minimum with a few optimisations:

- The endian switch upon return is combined with the MSR restore, which avoids an expensive context synchronizing operation for LE kernels. This makes up for the additional mtmsrd to enable interrupts with local_irq_enable().

Quoted 8 lines

- blr is now used to return from the opal_* functions that are called as C functions, to avoid link stack corruption. This requires a skiboot fix as well to keep the call stack balanced.

A NULL call is more costly after this, (410ns->430ns on POWER9), but OPAL calls are generally not performance critical at this scale.

Applied to powerpc next, thanks.

https://git.kernel.org/powerpc/c/75d9fc7fd94eb43cdf0bec04499a27ce

cheers