[PATCH 0/3] powerpc: Instrument Hypervisor Calls

[PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Mike Kravetz]

This patch series adds instrumentation to Hypervisor calls.  Code and
ideas were taken from the patch set provided by Christopher Yeoh.

The idea is to put all instrumentation code behind a configuration
option so that it can easily be added/removed.  The instrumentation
code is kept to a minimum in the hopes that it's impact to performance
will not be noticed.  Statistics are made available via files in sysfs.

-- 
Mike

[PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Mike Kravetz]

Here is an updated version of the hcall instrumentation patches.
Unfortunately, I have not been able to spend as much time as I
would like on this (and it seems that will continue).  

This version addresses all comments received except Arnd's issue
with an #ifdef for each function in the assembly file.  I like
Jimi's suggestion for creating macros(which could also minimize
the #ifdefs), but have not got around to implementing this.
Suggestions of others way to accomplish this are welcome.

Statistic files are moved to debugfs, and remain split out by CPU.
After some thought, disable/enable of preemption was added around
the statistic gathering(mostly for timing).

-- 
Mike

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Segher Boessenkool]

> Statistic files are moved to debugfs, and remain split out by CPU.

Having done some if this myself long ago, I really like to see these
per-cpu statistics.  "Not useful" in real life, but for most testcases
they're invaluable.  It would be useful to also add a "total" column
as well, of course, but it can also be done in userland.


Segher

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Mike Kravetz]

On Thu, Jun 22, 2006 at 03:56:09PM -0700, Mike Kravetz wrote:
> This version addresses all comments received except Arnd's issue
> with an #ifdef for each function in the assembly file.

I was thinking of changing the names of all the assembly routines from
plpar_hcall_*() to plpar_hcall_*_asm().  The instrumented version of the
routines would be named plpar_hcall_*_inst().  Then, the header file
would contain definitions such as:

#ifdef CONFIG_HCALL_STATS
#define plpar_hcall_*()		plpar_hcall_*_inst()
.
#else
#define plpar_hcall_*()		plpar_hcall_*_asm()
.
#endif

Is that any better than all the individual #ifdefs in the .S file?  Is it
still too ugly?

I'm open to any suggestions.

Thanks,
-- 
Mike

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Arnd Bergmann]

On Monday 10 July 2006 22:35, Mike Kravetz wrote:
> On Thu, Jun 22, 2006 at 03:56:09PM -0700, Mike Kravetz wrote:
> > This version addresses all comments received except Arnd's issue
> > with an #ifdef for each function in the assembly file.
> 
> I was thinking of changing the names of all the assembly routines from
> plpar_hcall_*() to plpar_hcall_*_asm().  The instrumented version of the
> routines would be named plpar_hcall_*_inst().  Then, the header file
> would contain definitions such as:
> 
> #ifdef CONFIG_HCALL_STATS
> #define plpar_hcall_*()		plpar_hcall_*_inst()
> .
> #else
> #define plpar_hcall_*()		plpar_hcall_*_asm()
> .
> #endif
> 
> Is that any better than all the individual #ifdefs in the .S file?  Is it
> still too ugly?
> 

I guess it's better to have the #ifdef in the header file, but then
again, you could just as well save some source lines doing

#ifndef CONFIG_HCALL_STATS
#define plpar_hcalldef(x) plpar_call_ ## x ## _asm
#else
#define plpar_hcalldef(x) plpar_call_ ## x ## _inst
#endif

#define plpar_call_foo plpar_hcalldef(foo)

	Arnd <><

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Mike Kravetz]

On Mon, Jul 10, 2006 at 01:35:10PM -0700, Mike Kravetz wrote:
> #ifdef CONFIG_HCALL_STATS
> #define plpar_hcall_*()		plpar_hcall_*_inst()

To answer my own question, this won't work because plpar_hcall_*()
routines are exported.  So, marcros can't be used.  Still thinking
about the best way to juggle function names for instrumentation.

-- 
Mike

[PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Mike Kravetz]

Hi Paul,

Here is an updated version of the patch(es) to instrument hcalls.  All
issues from the previous versions have been addressed.  Although, I
haven't really discovered an elegant solution to the assembly routine
name juggling.

In addition, there has been some discussion of these patches on IRC.
Some remaining issues/questions are:
- Exactly how much overhead does the statistic gathering introduce?
- What would be the cost of disabling preemption for more accurate statistics?
- What would be the cost of disabling interrupts for more accurate statistics?
- Should we extend this statistic gathering to RTAS calls?

I would like to get this basic code integrated and then start to address
some of these questions.  The IBM performance group will help in
benchmarking to determine overhead.

Please note that all code is behind a config option (off by default) with
zero impact unless enabled.

-- 
Mike

[PATCH 1/3] powerpc: Instrument Hypervisor Calls: merge headers

[Mike Kravetz]

Move all the Hypervisor call definitions to to a single header file.
--
Signed-off-by: Mike Kravetz <kravetz@us.ibm.com>

diff -Naupr linux-2.6.17.4/drivers/net/ibmveth.h linux-2.6.17.4.work/drivers/net/ibmveth.h
--- linux-2.6.17.4/drivers/net/ibmveth.h	2006-07-06 20:02:28.000000000 +0000
+++ linux-2.6.17.4.work/drivers/net/ibmveth.h	2006-07-14 23:05:37.000000000 +0000
@@ -41,16 +41,6 @@
 #define IbmVethMcastRemoveFilter     0x2UL
 #define IbmVethMcastClearFilterTable 0x3UL
 
-/* hcall numbers */
-#define H_VIO_SIGNAL             0x104
-#define H_REGISTER_LOGICAL_LAN   0x114
-#define H_FREE_LOGICAL_LAN       0x118
-#define H_ADD_LOGICAL_LAN_BUFFER 0x11C
-#define H_SEND_LOGICAL_LAN       0x120
-#define H_MULTICAST_CTRL         0x130
-#define H_CHANGE_LOGICAL_LAN_MAC 0x14C
-#define H_FREE_LOGICAL_LAN_BUFFER 0x1D4
-
 /* hcall macros */
 #define h_register_logical_lan(ua, buflst, rxq, fltlst, mac) \
   plpar_hcall_norets(H_REGISTER_LOGICAL_LAN, ua, buflst, rxq, fltlst, mac)
diff -Naupr linux-2.6.17.4/include/asm-powerpc/hvcall.h linux-2.6.17.4.work/include/asm-powerpc/hvcall.h
--- linux-2.6.17.4/include/asm-powerpc/hvcall.h	2006-07-06 20:02:28.000000000 +0000
+++ linux-2.6.17.4.work/include/asm-powerpc/hvcall.h	2006-07-14 23:05:37.000000000 +0000
@@ -155,9 +155,15 @@
 #define H_VIO_SIGNAL		0x104
 #define H_SEND_CRQ		0x108
 #define H_COPY_RDMA		0x110
+#define H_REGISTER_LOGICAL_LAN	0x114
+#define H_FREE_LOGICAL_LAN	0x118
+#define H_ADD_LOGICAL_LAN_BUFFER 0x11C
+#define H_SEND_LOGICAL_LAN	0x120
+#define H_MULTICAST_CTRL	0x130
 #define H_SET_XDABR		0x134
 #define H_STUFF_TCE		0x138
 #define H_PUT_TCE_INDIRECT	0x13C
+#define H_CHANGE_LOGICAL_LAN_MAC 0x14C
 #define H_VTERM_PARTNER_INFO	0x150
 #define H_REGISTER_VTERM	0x154
 #define H_FREE_VTERM		0x158
@@ -187,11 +193,14 @@
 #define H_GET_HCA_INFO          0x1B8
 #define H_GET_PERF_COUNT        0x1BC
 #define H_MANAGE_TRACE          0x1C0
+#define H_FREE_LOGICAL_LAN_BUFFER 0x1D4
 #define H_QUERY_INT_STATE       0x1E4
 #define H_POLL_PENDING		0x1D8
 #define H_JOIN			0x298
 #define H_ENABLE_CRQ		0x2B0
 
+#define MAX_HCALL_OPCODES	(H_ENABLE_CRQ >> 2)
+
 #ifndef __ASSEMBLY__
 
 /* plpar_hcall() -- Generic call interface using above opcodes

[PATCH 2/3] powerpc: Instrument Hypervisor Calls: add wrappers

[Mike Kravetz]

Add wrappers which perform the actual hypervisor call instrumentation.
--
Signed-off-by: Mike Kravetz <kravetz@us.ibm.com>

diff -Naupr linux-2.6.17.4/arch/powerpc/Kconfig.debug linux-2.6.17.4.work/arch/powerpc/Kconfig.debug
--- linux-2.6.17.4/arch/powerpc/Kconfig.debug	2006-07-06 20:02:28.000000000 +0000
+++ linux-2.6.17.4.work/arch/powerpc/Kconfig.debug	2006-07-14 23:28:20.000000000 +0000
@@ -18,6 +18,19 @@ config DEBUG_STACK_USAGE
 
 	  This option will slow down process creation somewhat.
 
+config HCALL_STATS
+	bool "Hypervisor call instrumentation"
+	depends on PPC_PSERIES && DEBUG_FS
+	help
+	  Adds code to keep track of the number of hypervisor calls made and
+	  the amount of time spent in hypervisor calls.  A directory named
+	  hcall_inst is added at the root of the debugfs filesystem.  Within
+	  the hcall_inst directory are files that contain CPU specific call
+	  statistics.
+
+	  This option will add a small amount of overhead to all hypervisor
+	  calls.
+
 config DEBUGGER
 	bool "Enable debugger hooks"
 	depends on DEBUG_KERNEL
diff -Naupr linux-2.6.17.4/arch/powerpc/platforms/pseries/Makefile linux-2.6.17.4.work/arch/powerpc/platforms/pseries/Makefile
--- linux-2.6.17.4/arch/powerpc/platforms/pseries/Makefile	2006-07-06 20:02:28.000000000 +0000
+++ linux-2.6.17.4.work/arch/powerpc/platforms/pseries/Makefile	2006-07-14 23:28:20.000000000 +0000
@@ -9,3 +9,4 @@ obj-$(CONFIG_EEH)	+= eeh.o eeh_cache.o e
 
 obj-$(CONFIG_HVC_CONSOLE)	+= hvconsole.o
 obj-$(CONFIG_HVCS)		+= hvcserver.o
+obj-$(CONFIG_HCALL_STATS)	+= hvCall_inst.o
diff -Naupr linux-2.6.17.4/arch/powerpc/platforms/pseries/hvCall.S linux-2.6.17.4.work/arch/powerpc/platforms/pseries/hvCall.S
--- linux-2.6.17.4/arch/powerpc/platforms/pseries/hvCall.S	2006-07-06 20:02:28.000000000 +0000
+++ linux-2.6.17.4.work/arch/powerpc/platforms/pseries/hvCall.S	2006-07-14 23:28:20.000000000 +0000
@@ -11,7 +11,35 @@
 #include <asm/hvcall.h>
 #include <asm/processor.h>
 #include <asm/ppc_asm.h>
-	
+
+/*
+ * If hcall statistics are desired, all routines are wrapped with code
+ * that does the statistic gathering.
+ */
+#ifndef CONFIG_HCALL_STATS
+#define PLPAR_HCALL		plpar_hcall
+#define PLPAR_HCALL_NORETS	plpar_hcall_norets
+#define PLPAR_HCALL_8ARG_2RET	plpar_hcall_8arg_2ret
+#define PLPAR_HCALL_4OUT	plpar_hcall_4out
+#define PLPAR_HCALL_7ARG_7RET	plpar_hcall_7arg_7ret
+#define PLPAR_HCALL_9ARG_9RET	plpar_hcall_9arg_9ret
+#else
+#define PLPAR_HCALL		plpar_hcall_base
+#define PLPAR_HCALL_NORETS	plpar_hcall_norets_base
+#define PLPAR_HCALL_8ARG_2RET	plpar_hcall_8arg_2ret_base
+#define PLPAR_HCALL_4OUT	plpar_hcall_4out_base
+#define PLPAR_HCALL_7ARG_7RET	plpar_hcall_7arg_7ret_base
+#define PLPAR_HCALL_9ARG_9RET	plpar_hcall_9arg_9ret_base
+
+/*
+ * A special 'indirect' call to a C based wrapper if statistics are desired.
+ * See plpar_hcall_norets_C function header for more details.
+ */
+_GLOBAL(plpar_hcall_norets)
+	b	plpar_hcall_norets_C
+
+#endif
+
 #define STK_PARM(i)     (48 + ((i)-3)*8)
 
 	.text
@@ -25,7 +53,7 @@
 			unsigned long *out2,		R9
 			unsigned long *out3);		R10
  */
-_GLOBAL(plpar_hcall)
+_GLOBAL(PLPAR_HCALL)
 	HMT_MEDIUM
 
 	mfcr	r0
@@ -52,7 +80,7 @@ _GLOBAL(plpar_hcall)
 
 
 /* Simple interface with no output values (other than status) */
-_GLOBAL(plpar_hcall_norets)
+_GLOBAL(PLPAR_HCALL_NORETS)
 	HMT_MEDIUM
 
 	mfcr	r0
@@ -76,7 +104,7 @@ _GLOBAL(plpar_hcall_norets)
 			unsigned long arg8,		112(R1)
 			unsigned long *out1);		120(R1)
  */
-_GLOBAL(plpar_hcall_8arg_2ret)
+_GLOBAL(PLPAR_HCALL_8ARG_2RET)
 	HMT_MEDIUM
 
 	mfcr	r0
@@ -102,7 +130,7 @@ _GLOBAL(plpar_hcall_8arg_2ret)
 		 	unsigned long *out3,		R10
 		 	unsigned long *out4);		112(R1)
  */
-_GLOBAL(plpar_hcall_4out)
+_GLOBAL(PLPAR_HCALL_4OUT)
 	HMT_MEDIUM
 
 	mfcr	r0
@@ -144,7 +172,7 @@ _GLOBAL(plpar_hcall_4out)
 			 unsigned long *out6,		102(R1)
 			 unsigned long *out7);		100(R1)
 */
-_GLOBAL(plpar_hcall_7arg_7ret)
+_GLOBAL(PLPAR_HCALL_7ARG_7RET)
 	HMT_MEDIUM
 
 	mfcr	r0
@@ -193,7 +221,7 @@ _GLOBAL(plpar_hcall_7arg_7ret)
 			 unsigned long *out8,		 94(R1)
 		         unsigned long *out9,            92(R1)
 */
-_GLOBAL(plpar_hcall_9arg_9ret)
+_GLOBAL(PLPAR_HCALL_9ARG_9RET)
 	HMT_MEDIUM
 
 	mfcr	r0
diff -Naupr linux-2.6.17.4/arch/powerpc/platforms/pseries/hvCall_inst.c linux-2.6.17.4.work/arch/powerpc/platforms/pseries/hvCall_inst.c
--- linux-2.6.17.4/arch/powerpc/platforms/pseries/hvCall_inst.c	1970-01-01 00:00:00.000000000 +0000
+++ linux-2.6.17.4.work/arch/powerpc/platforms/pseries/hvCall_inst.c	2006-07-14 23:28:44.000000000 +0000
@@ -0,0 +1,207 @@
+/*
+ * Copyright (C) 2006 Mike Kravetz IBM Corporation
+ *
+ * Hypervisor Call Instrumentation
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
+ */
+
+#include <linux/kernel.h>
+#include <linux/percpu.h>
+#include <linux/debugfs.h>
+#include <linux/seq_file.h>
+#include <linux/cpumask.h>
+#include <asm/hvcall.h>
+#include <asm/firmware.h>
+
+DEFINE_PER_CPU(struct hcall_stats[MAX_HCALL_OPCODES+1], hcall_stats);
+
+/*
+ * Common update of the per-CPU/per-hcall statistics
+ */
+static inline void update_stats(unsigned long opcode, unsigned long t_before)
+{
+	unsigned long op_index = opcode >> 2;
+	struct hcall_stats *hs = &__get_cpu_var(hcall_stats[op_index]);
+
+	hs->total_time += (mftb() - t_before);
+	hs->num_calls++;
+}
+
+/*
+ * plpar_hcall wrapper
+ */
+long plpar_hcall(unsigned long opcode,
+                 unsigned long arg1,
+                 unsigned long arg2,
+                 unsigned long arg3,
+                 unsigned long arg4,
+                 unsigned long *out1,
+                 unsigned long *out2,
+                 unsigned long *out3)
+{
+	long rc;
+	unsigned long t_before;
+
+	t_before = mftb();
+	rc = plpar_hcall_base(opcode, arg1, arg2, arg3, arg4, out1, out2, out3);
+
+	update_stats(opcode, t_before);
+	return rc;
+}
+
+/*
+ * A C based wrapper for plpar_hcall_norets
+ * The wrapper for plpar_hcall_norets is a special case because the function
+ * takes a variable number of arguments.  It is almost impossible to write a
+ * wrapper for a function that takes a variable number of arguments in C.
+ * Therefore, there is an assembly routine in hvCall.S that simply branches
+ * to this C wrapper.  This 'indirection' takes care of the variable arguments
+ * issue.  This C wrapper has a fixed maximum number of arguments.
+ */
+long plpar_hcall_norets_C(unsigned long opcode,
+				unsigned long arg1,
+				unsigned long arg2,
+				unsigned long arg3,
+				unsigned long arg4,
+				unsigned long arg5,
+				unsigned long arg6)
+{
+	long rc;
+	unsigned long t_before;
+
+	t_before = mftb();
+	rc = plpar_hcall_norets_base(opcode, arg1, arg2, arg3, arg4, arg5,
+				     arg6);
+
+	update_stats(opcode, t_before);
+	return rc;
+}
+
+/*
+ * plpar_hcall_8arg_2ret wrapper
+ */
+long plpar_hcall_8arg_2ret(unsigned long opcode,
+                           unsigned long arg1,
+                           unsigned long arg2,
+                           unsigned long arg3,
+                           unsigned long arg4,
+                           unsigned long arg5,
+                           unsigned long arg6,
+                           unsigned long arg7,
+                           unsigned long arg8,
+                           unsigned long *out1)
+{
+	long rc;
+	unsigned long t_before;
+
+	t_before = mftb();
+	rc = plpar_hcall_8arg_2ret_base(opcode, arg1, arg2, arg3, arg4, arg5,
+					arg6, arg7, arg8, out1);
+
+	update_stats(opcode, t_before);
+	return rc;
+}
+
+/*
+ * plpar_hcall_4out wrapper
+ */
+long plpar_hcall_4out(unsigned long opcode,
+                      unsigned long arg1,
+                      unsigned long arg2,
+                      unsigned long arg3,
+                      unsigned long arg4,
+                      unsigned long *out1,
+                      unsigned long *out2,
+                      unsigned long *out3,
+                      unsigned long *out4)
+{
+	long rc;
+	unsigned long t_before;
+
+	t_before = mftb();
+	rc = plpar_hcall_4out_base(opcode, arg1, arg2, arg3, arg4, out1,
+				   out2, out3, out4);
+
+	update_stats(opcode, t_before);
+	return rc;
+}
+
+/*
+ * plpar_hcall_7arg_7ret wrapper
+ */
+long plpar_hcall_7arg_7ret(unsigned long opcode,
+                           unsigned long arg1,
+                           unsigned long arg2,
+                           unsigned long arg3,
+                           unsigned long arg4,
+                           unsigned long arg5,
+                           unsigned long arg6,
+                           unsigned long arg7,
+                           unsigned long *out1,
+                           unsigned long *out2,
+                           unsigned long *out3,
+                           unsigned long *out4,
+                           unsigned long *out5,
+                           unsigned long *out6,
+                           unsigned long *out7)
+{
+	long rc;
+	unsigned long t_before;
+
+	t_before = mftb();
+	rc = plpar_hcall_7arg_7ret_base(opcode, arg1, arg2, arg3, arg4, arg5,
+					arg6, arg7, out1, out2, out3, out4,
+					out5, out6, out7);
+
+	update_stats(opcode, t_before);
+	return rc;
+}
+
+/*
+ * plpar_hcall_9arg_9ret wrapper
+ */
+long plpar_hcall_9arg_9ret(unsigned long opcode,
+                           unsigned long arg1,
+                           unsigned long arg2,
+                           unsigned long arg3,
+                           unsigned long arg4,
+                           unsigned long arg5,
+                           unsigned long arg6,
+                           unsigned long arg7,
+                           unsigned long arg8,
+                           unsigned long arg9,
+                           unsigned long *out1,
+                           unsigned long *out2,
+                           unsigned long *out3,
+                           unsigned long *out4,
+                           unsigned long *out5,
+                           unsigned long *out6,
+                           unsigned long *out7,
+                           unsigned long *out8,
+                           unsigned long *out9)
+{
+	long rc;
+	unsigned long t_before;
+
+	t_before = mftb();
+	rc = plpar_hcall_9arg_9ret_base(opcode, arg1, arg2, arg3, arg4, arg5,
+					arg6, arg7, arg8, arg9, out1, out2,
+					out3, out4, out5, out6, out7, out8,
+					out9);
+
+	update_stats(opcode, t_before);
+	return rc;
+}
diff -Naupr linux-2.6.17.4/include/asm-powerpc/hvcall.h linux-2.6.17.4.work/include/asm-powerpc/hvcall.h
--- linux-2.6.17.4/include/asm-powerpc/hvcall.h	2006-07-14 23:27:52.000000000 +0000
+++ linux-2.6.17.4.work/include/asm-powerpc/hvcall.h	2006-07-14 23:28:20.000000000 +0000
@@ -292,6 +292,86 @@ long plpar_hcall_9arg_9ret(unsigned long
 			   unsigned long *out8,
 			   unsigned long *out9);
 
+
+/* For hcall instrumentation.  One structure per-hcall, per-CPU */
+struct hcall_stats {
+	unsigned long	num_calls;	/* number of calls (on this CPU) */
+	unsigned long	total_time;	/* total cputime (PURR) of calls. */
+};
+
+/* If Hypervisor call instrumentation is enabled, the assembly routine
+ * names are changed from 'plpar_hcall*' to 'plpar_hcall*_base' and
+ * 'plpar_hcall*' routines become instrumented wrappers.  The following
+ * are declarations for the renamed 'plpar_hcall*_base' routines.
+ */
+long plpar_hcall_base (unsigned long opcode,
+		       unsigned long arg1,
+		       unsigned long arg2,
+		       unsigned long arg3,
+		       unsigned long arg4,
+		       unsigned long *out1,
+		       unsigned long *out2,
+		       unsigned long *out3);
+
+long plpar_hcall_norets_base(unsigned long opcode, ...);
+
+long plpar_hcall_8arg_2ret_base(unsigned long opcode,
+				unsigned long arg1,
+				unsigned long arg2,
+				unsigned long arg3,
+				unsigned long arg4,
+				unsigned long arg5,
+				unsigned long arg6,
+				unsigned long arg7,
+				unsigned long arg8,
+				unsigned long *out1);
+
+long plpar_hcall_4out_base(unsigned long opcode,
+			   unsigned long arg1,
+			   unsigned long arg2,
+			   unsigned long arg3,
+			   unsigned long arg4,
+			   unsigned long *out1,
+			   unsigned long *out2,
+			   unsigned long *out3,
+			   unsigned long *out4);
+
+long plpar_hcall_7arg_7ret_base(unsigned long opcode,
+				unsigned long arg1,
+				unsigned long arg2,
+				unsigned long arg3,
+				unsigned long arg4,
+				unsigned long arg5,
+				unsigned long arg6,
+				unsigned long arg7,
+				unsigned long *out1,
+				unsigned long *out2,
+				unsigned long *out3,
+				unsigned long *out4,
+				unsigned long *out5,
+				unsigned long *out6,
+				unsigned long *out7);
+
+long plpar_hcall_9arg_9ret_base(unsigned long opcode,
+				unsigned long arg1,
+				unsigned long arg2,
+				unsigned long arg3,
+				unsigned long arg4,
+				unsigned long arg5,
+				unsigned long arg6,
+				unsigned long arg7,
+				unsigned long arg8,
+				unsigned long arg9,
+				unsigned long *out1,
+				unsigned long *out2,
+				unsigned long *out3,
+				unsigned long *out4,
+				unsigned long *out5,
+				unsigned long *out6,
+				unsigned long *out7,
+				unsigned long *out8,
+				unsigned long *out9);
+
 #endif /* __ASSEMBLY__ */
 #endif /* __KERNEL__ */
 #endif /* _ASM_POWERPC_HVCALL_H */

[PATCH 3/3] powerpc: Instrument Hypervisor Calls: add debugfs files

[Mike Kravetz]

Make statistics available via files in debugfs.
--
Signed-off-by: Mike Kravetz <kravetz@us.ibm.com>

diff -Naupr linux-2.6.17.4/arch/powerpc/platforms/pseries/hvCall_inst.c linux-2.6.17.4.work/arch/powerpc/platforms/pseries/hvCall_inst.c
--- linux-2.6.17.4/arch/powerpc/platforms/pseries/hvCall_inst.c	2006-07-14 23:32:01.000000000 +0000
+++ linux-2.6.17.4.work/arch/powerpc/platforms/pseries/hvCall_inst.c	2006-07-14 23:32:54.000000000 +0000
@@ -205,3 +205,94 @@ long plpar_hcall_9arg_9ret(unsigned long
 	update_stats(opcode, t_before);
 	return rc;
 }
+
+/*
+ * Routines for displaying the statistics in debugfs
+ */
+static void *hc_start(struct seq_file *m, loff_t *pos)
+{
+	if ((int)*pos < MAX_HCALL_OPCODES)
+		return (void *)(unsigned long)(*pos + 1);
+
+	return NULL;
+}
+
+static void *hc_next(struct seq_file *m, void *p, loff_t * pos)
+{
+	++*pos;
+
+	return hc_start(m, pos);
+}
+
+static void hc_stop(struct seq_file *m, void *p)
+{
+}
+
+static int hc_show(struct seq_file *m, void *p)
+{
+	unsigned long h_num = (unsigned long)p;
+	struct hcall_stats *hs = (struct hcall_stats *)m->private;
+
+	if (hs[h_num].num_calls)
+		seq_printf(m, "%lu %lu %lu\n", h_num<<2, hs[h_num].num_calls,
+			   hs[h_num].total_time);
+
+	return 0;
+}
+
+static struct seq_operations hcall_inst_seq_ops = {
+        .start = hc_start,
+        .next  = hc_next,
+        .stop  = hc_stop,
+        .show  = hc_show
+};
+
+static int hcall_inst_seq_open(struct inode *inode, struct file *file)
+{
+	int rc;
+	struct seq_file *seq;
+
+	rc = seq_open(file, &hcall_inst_seq_ops);
+	seq = file->private_data;
+	seq->private = file->f_dentry->d_inode->u.generic_ip;
+
+	return rc;
+}
+
+static struct file_operations hcall_inst_seq_fops = {
+	.open = hcall_inst_seq_open,
+	.read = seq_read,
+	.llseek = seq_lseek,
+	.release = seq_release,
+};
+
+#define	HCALL_ROOT_DIR		"hcall_inst"
+#define CPU_NAME_BUF_SIZE	32
+
+static int __init hcall_inst_init(void)
+{
+	struct dentry *hcall_root;
+	struct dentry *hcall_file;
+	char cpu_name_buf[CPU_NAME_BUF_SIZE];
+	int cpu;
+
+	if (!firmware_has_feature(FW_FEATURE_LPAR))
+		return 0;
+
+	hcall_root = debugfs_create_dir(HCALL_ROOT_DIR, NULL);
+	if (!hcall_root)
+		return -ENOMEM;
+
+	for_each_cpu(cpu) {
+		snprintf(cpu_name_buf, CPU_NAME_BUF_SIZE, "cpu%d", cpu);
+		hcall_file = debugfs_create_file(cpu_name_buf, S_IRUGO,
+						 hcall_root,
+						 per_cpu(hcall_stats, cpu),
+						 &hcall_inst_seq_fops);
+		if (!hcall_file)
+			return -ENOMEM;
+	}
+
+	return 0;
+}
+__initcall(hcall_inst_init);

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Arnd Bergmann]

On Saturday 15 July 2006 01:37, Mike Kravetz wrote:
> In addition, there has been some discussion of these patches on IRC.
> Some remaining issues/questions are:
> - Exactly how much overhead does the statistic gathering introduce?
> - What would be the cost of disabling preemption for more accurate statistics?
> - What would be the cost of disabling interrupts for more accurate statistics?
> - Should we extend this statistic gathering to RTAS calls?

What happened to the question whether to use PURR values for also measuring
cycles spent executing the hcall as opposed to cycles that passed before
the hcall returns. Did that turn out not giving extra information after all
or was there a different reason to drop that idea?

	Arnd <><

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Mike Kravetz]

On Sat, Jul 15, 2006 at 02:00:02AM +0200, Arnd Bergmann wrote:
> What happened to the question whether to use PURR values for also measuring
> cycles spent executing the hcall as opposed to cycles that passed before
> the hcall returns. Did that turn out not giving extra information after all
> or was there a different reason to drop that idea?

Oops, forgot that as an additional issue/question.  In this patch, I went
back to mftb() as 'wall time' made more sense for the group wanting this
functionality.  It is easy to switch, or collect both (I think).  Since I
started with mftb went to PURR and then back to mftb, it certainly does
look like an agument to try and collect both. :)

-- 
Mike

Re: [PATCH 2/3] powerpc: Instrument Hypervisor Calls: add wrappers

[Nathan Lynch]

Mike Kravetz wrote:
> +static inline void update_stats(unsigned long opcode, unsigned long t_before)
> +{
> +	unsigned long op_index = opcode >> 2;
> +	struct hcall_stats *hs = &__get_cpu_var(hcall_stats[op_index]);
> +
> +	hs->total_time += (mftb() - t_before);
> +	hs->num_calls++;
> +}
> +
> +/*
> + * plpar_hcall wrapper
> + */
> +long plpar_hcall(unsigned long opcode,
> +                 unsigned long arg1,
> +                 unsigned long arg2,
> +                 unsigned long arg3,
> +                 unsigned long arg4,
> +                 unsigned long *out1,
> +                 unsigned long *out2,
> +                 unsigned long *out3)
> +{
> +	long rc;
> +	unsigned long t_before;
> +
> +	t_before = mftb();
> +	rc = plpar_hcall_base(opcode, arg1, arg2, arg3, arg4, out1, out2, out3);
> +
> +	update_stats(opcode, t_before);
> +	return rc;
> +}

Hmm, isn't it possible that the stats could be corrupted if we process
an interrupt/softirq which does an hcall while the stats are being
updated?  Maybe it's not a show-stopper, but it seems to me that
accuracy could suffer under some workloads.

Re: [PATCH 2/3] powerpc: Instrument Hypervisor Calls: add wrappers

[Mike Kravetz]

On Fri, Jul 14, 2006 at 07:15:45PM -0500, Nathan Lynch wrote:
> Hmm, isn't it possible that the stats could be corrupted if we process
> an interrupt/softirq which does an hcall while the stats are being
> updated?  Maybe it's not a show-stopper, but it seems to me that
> accuracy could suffer under some workloads.

Yes, they can become inaccurate.  Is this what you mean by 'corrupted' or
you thinking of something worse?

There is going to be a trade off between accuracy and the overhead of the
data gathering.  What really needs to happen is some good benchmarking to
measure the costs of different approaches.  I'm pretty sure I can get
someone to provide such data.

In this patch, I have minimized the overhead in data collection at the
cost of accuracy.

-- 
Mike

Re: [PATCH 2/3] powerpc: Instrument Hypervisor Calls: add wrappers

[Nathan Lynch]

Mike Kravetz wrote:
> On Fri, Jul 14, 2006 at 07:15:45PM -0500, Nathan Lynch wrote:
> > Hmm, isn't it possible that the stats could be corrupted if we process
> > an interrupt/softirq which does an hcall while the stats are being
> > updated?  Maybe it's not a show-stopper, but it seems to me that
> > accuracy could suffer under some workloads.
> 
> Yes, they can become inaccurate.  Is this what you mean by 'corrupted' or
> you thinking of something worse?

Yes, "corrupted" was perhaps too strong, but we could lose the
accounting for hcalls from interrupts that we take at exactly the
wrong time.  I guess it's a pretty tiny risk.

I just wanted to make sure this had been considered, but if you're
comfortable with the overhead/accuracy tradeoff, then okay.

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Anton Blanchard]

 
> What happened to the question whether to use PURR values for also measuring
> cycles spent executing the hcall as opposed to cycles that passed before
> the hcall returns. Did that turn out not giving extra information after all
> or was there a different reason to drop that idea?

You have to be careful with PURR since it may be context switched
between partitions.

Anton

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Arnd Bergmann]

On Saturday 15 July 2006 17:30, Anton Blanchard wrote:
> > What happened to the question whether to use PURR values for also measuring
> > cycles spent executing the hcall as opposed to cycles that passed before
> > the hcall returns. Did that turn out not giving extra information after all
> > or was there a different reason to drop that idea?
> 
> You have to be careful with PURR since it may be context switched
> between partitions.
> 

Not sure I follow you. I would expect the PURR value to be restored after
a context switch, even if we continue on a different physical CPU.

The idea behind monitoring both PURR and timebase is that the difference
between the two tells you how long the partition was suspended during
the hcall.

	Arnd <><

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Anton Blanchard]

 
> Not sure I follow you. I would expect the PURR value to be restored after
> a context switch, even if we continue on a different physical CPU.
> 
> The idea behind monitoring both PURR and timebase is that the difference
> between the two tells you how long the partition was suspended during
> the hcall.

Sounds good, last time I looked at the patch I thought it was gathering
the PURR only. That on its own would make for some confusing results.

Anton

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Olof Johansson]

On Sat, Jul 15, 2006 at 02:00:02AM +0200, Arnd Bergmann wrote:

> What happened to the question whether to use PURR values for also measuring
> cycles spent executing the hcall as opposed to cycles that passed before
> the hcall returns. Did that turn out not giving extra information after all
> or was there a different reason to drop that idea?

Just so people don't forget: this can't be done on all processors. For
example, PPC970 and POWER4 don't implement the PURR SPR. And it doesn't
make sense to use H_PURR to get the software emulated ones there. Not
really an issue on POWER4 since they don't do shared processor LPAR, but
on JS21 I think they might do?


-Olof

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Luke Browning]

[-- Attachment #1: Type: text/plain, Size: 910 bytes --]

linuxppc-dev-bounces+lukebrowning=us.ibm.com@ozlabs.org wrote on 
07/16/2006 12:53:54 AM:

> On Sat, Jul 15, 2006 at 02:00:02AM +0200, Arnd Bergmann wrote:
> 
> > What happened to the question whether to use PURR values for also 
measuring
> > cycles spent executing the hcall as opposed to cycles that passed 
before
> > the hcall returns. Did that turn out not giving extra information 
after all
> > or was there a different reason to drop that idea?
> 
> Just so people don't forget: this can't be done on all processors. For
> example, PPC970 and POWER4 don't implement the PURR SPR. And it doesn't
> make sense to use H_PURR to get the software emulated ones there. Not
> really an issue on POWER4 since they don't do shared processor LPAR, but
> on JS21 I think they might do?
> 
Isn't there someway to do a platform specific processor overlay, where you 

get PURR if it exists and TB otherwise.

Luke

[-- Attachment #2: Type: text/html, Size: 1129 bytes --]

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Luke Browning]

[-- Attachment #1: Type: text/plain, Size: 986 bytes --]

linuxppc-dev-bounces+lukebrowning=us.ibm.com@ozlabs.org wrote on 
07/15/2006 07:07:58 PM:

> 
> > Not sure I follow you. I would expect the PURR value to be restored 
after
> > a context switch, even if we continue on a different physical CPU.
> > 
> > The idea behind monitoring both PURR and timebase is that the 
difference
> > between the two tells you how long the partition was suspended during
> > the hcall.
> 
> Sounds good, last time I looked at the patch I thought it was gathering
> the PURR only. That on its own would make for some confusing results.
>

It would be more efficient to have a separate trace log for PHYP 
dispatches as 
there are many more hypervisor calls than PHYP dispatches.  I believe PHYP 
provides
a trace records for dispatches, which could be seperately written.  Perf 
tools can 
assemble the information from multiple records.  It would be nice to have 
processor
specific time function that reads either PURR or TB based on the platform.

Luke


[-- Attachment #2: Type: text/html, Size: 1376 bytes --]

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Olof Johansson]

On Sun, Jul 16, 2006 at 08:53:21PM -0200, Luke Browning wrote:

> > Just so people don't forget: this can't be done on all processors. For
> > example, PPC970 and POWER4 don't implement the PURR SPR. And it doesn't
> > make sense to use H_PURR to get the software emulated ones there. Not
> > really an issue on POWER4 since they don't do shared processor LPAR, but
> > on JS21 I think they might do?
> > 
> Isn't there someway to do a platform specific processor overlay, where you 
> get PURR if it exists and TB otherwise.

Sure, it's what we have cpu feature bits for. It's not a difficult
problem to solve, I was just reminding of the lack of SPRN_PURR on some
processors that can run hypervisors.


-Olof

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Luke Browning]

[-- Attachment #1: Type: text/plain, Size: 602 bytes --]

linuxppc-dev-bounces+lukebrowning=us.ibm.com@ozlabs.org wrote on 
07/15/2006 12:30:54 PM:

> 
> > What happened to the question whether to use PURR values for also 
measuring
> > cycles spent executing the hcall as opposed to cycles that passed 
before
> > the hcall returns. Did that turn out not giving extra information 
after all
> > or was there a different reason to drop that idea?
> 
> You have to be careful with PURR since it may be context switched
> between partitions.
> 

That shouldn't be an issue as the PURR is supposed to be virtualized by 
PHYP at 
the virtual processor level.

Luke

[-- Attachment #2: Type: text/html, Size: 828 bytes --]

[PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Mike Kravetz]

A small update from the last version.  By popular demand, both
wall time (mftb) and cpu cycles(PURR) are collected for each call.
It is interesting to see these two values side by side in the
output files.

-- 
Mike

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Paul Mackerras]

Mike Kravetz writes:

> A small update from the last version.  By popular demand, both
> wall time (mftb) and cpu cycles(PURR) are collected for each call.
> It is interesting to see these two values side by side in the
> output files.

Did you see the patch from Anton that I posted a week or so ago, which
reduces the number of plpar_hcall_* functions?  I'd rather the
instrumentation stuff was based on that.

My other comment is this: wouldn't it actually turn out simpler if we
read the timebase (and PURR, if we really want to do that too) in the
assembly code that implements plpar_hcall_*?

Paul.

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Mike Kravetz]

On Tue, Jul 18, 2006 at 01:47:23PM -0700, Mike Kravetz wrote:
> A small update from the last version.  By popular demand, both
> wall time (mftb) and cpu cycles(PURR) are collected for each call.
> It is interesting to see these two values side by side in the
> output files.

Arnd asked on IRC, so I thought others may be interested in a
sample output file.  Remember, there is one file per CPU. All
values are decimal.  First number is hcall # (in decimal), second
is number of calls, third is total time (mftb) and fourth is cpu
time (PURR).  Note that 'total times' are still in cycles and
can be converted to another format in user space.

[elm3b125]/sys/kernel/debug/hcall_inst >cat cpu0
4 140651 21538230 14456409
8 130064 11855094 9459178
24 14259 1496316 986702
32 1809 354584 315399
84 103 27196 17763
88 1046 108106 109322
100 2596 251229 205154
104 1 1040 1061
108 2208 320491 242708
116 2669 1030894 713272
220 2 532 510
224 130662 47508588184 169383891
228 5 2689902 4622
260 3018 261893 230753

-- 
Mike

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Olof Johansson]

On Tue, Jul 18, 2006 at 02:29:13PM -0700, Mike Kravetz wrote:

> 88 1046 108106 109322
[...]
> 104 1 1040 1061

Heh. Shouldn't PURR/TB be <= 1 at all times? :-) I bet this is because
they're not sampled at exactly the same time, etc.

Maybe recoding it to assembly so you can do mftb and mfspr right after
each other, and do the arithmetics/stores later could help. Inline asm
might be sufficient.


-Olof

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Mike Kravetz]

On Wed, Jul 19, 2006 at 07:00:22AM +1000, Paul Mackerras wrote:
> Mike Kravetz writes:
> > A small update from the last version.  By popular demand, both
> > wall time (mftb) and cpu cycles(PURR) are collected for each call.
> > It is interesting to see these two values side by side in the
> > output files.
> 
> Did you see the patch from Anton that I posted a week or so ago, which
> reduces the number of plpar_hcall_* functions?  I'd rather the
> instrumentation stuff was based on that.

Agreed.  I missed the patch from last week, but saw the one from Anton
that was posted today.  Future instrumentation patches will assume this
change.

> My other comment is this: wouldn't it actually turn out simpler if we
> read the timebase (and PURR, if we really want to do that too) in the
> assembly code that implements plpar_hcall_*?

I'll give that a try.  My assembly skills are a bit rusty, so it may
take a little longer to produce something.

-- 
Mike

Re: [PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Mike Kravetz]

On Tue, Jul 18, 2006 at 08:36:18PM -0700, Mike Kravetz wrote:
> > My other comment is this: wouldn't it actually turn out simpler if we
> > read the timebase (and PURR, if we really want to do that too) in the
> > assembly code that implements plpar_hcall_*?
> 
> I'll give that a try.  My assembly skills are a bit rusty, so it may
> take a little longer to produce something.

Getting the timebase and PURR in assembly is easy enough.  So, I thought
about updating the per-cpu statistics while in there.  But, I couldn't
find any other assembly code that does this.  Sure there is asm code that
updated fields in the PACA, but none that messes with paca.data_offset
(that I could find).

Should the statistic updates also be done in the assembly routines?  Or,
how about a call out to a C routine for this?

-- 
Mike

[PATCH 0/3] powerpc: Instrument Hypervisor Calls

[Mike Kravetz]

Here is a version of the instrumentation patches built on top of
Anton's hcall cleanup patch.  In addition, I have gathered the
timebase and PURR snapshots in assembly code.  However, it still
performs a call out to C code that updates the data structures.

I'd appreciate some comments on the assembly code as it has been
a looooong time since I've written any.

-- 
Mike