Andrew, please apply:
The PPC64 interrupt code includes a hook to call when an exception
from the performance monitor unit occurs. However, there's no way of
reserving the hook properly, so if more than one bit of code tries to
use it things will get ugly. Currently oprofile is the only user, but
there are likely to be more in future e.g. perfctr, if and when it
reaches a fit state for merging.
This patch creates functions to reserve and release the performance
monitor hardware (including its interrupt), and makes oprofile use
them. It also creates a new arch/ppc64/kernel/pmc.c, in which we can
put any future helper functions for handling the performance monitor
counters.
Signed-off-by: David Gibson <[email protected]>
Index: working-2.6/arch/ppc64/kernel/pmc.c
===================================================================
--- /dev/null 1970-01-01 00:00:00.000000000 +0000
+++ working-2.6/arch/ppc64/kernel/pmc.c 2005-01-10 16:32:49.733411536 +1100
@@ -0,0 +1,65 @@
+/*
+ * linux/arch/ppc64/kernel/pmc.c
+ *
+ * Copyright (C) 2004 David Gibson, IBM Corporation.
+ *
+ * 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.
+ */
+
+#include <linux/config.h>
+#include <linux/errno.h>
+#include <linux/spinlock.h>
+
+#include <asm/processor.h>
+#include <asm/pmc.h>
+
+/* Ensure exceptions are disabled */
+static void dummy_perf(struct pt_regs *regs)
+{
+ unsigned int mmcr0 = mfspr(SPRN_MMCR0);
+
+ mmcr0 &= ~(MMCR0_PMXE|MMCR0_PMAO);
+ mtspr(SPRN_MMCR0, mmcr0);
+}
+
+static spinlock_t pmc_owner_lock = SPIN_LOCK_UNLOCKED;
+static void *pmc_owner_caller; /* mostly for debugging */
+perf_irq_t perf_irq = dummy_perf;
+
+int reserve_pmc_hardware(perf_irq_t new_perf_irq)
+{
+ int err = -EBUSY;;
+
+ spin_lock(&pmc_owner_lock);
+
+ if (pmc_owner_caller) {
+ printk(KERN_WARNING "reserve_pmc_hardware: "
+ "PMC hardware busy (reserved by caller %p)\n",
+ pmc_owner_caller);
+ goto out;
+ }
+
+ pmc_owner_caller = __builtin_return_address(0);
+ perf_irq = new_perf_irq ? : dummy_perf;
+
+ err = 0;
+
+ out:
+ spin_unlock(&pmc_owner_lock);
+ return err;
+}
+
+void release_pmc_hardware(void)
+{
+ spin_lock(&pmc_owner_lock);
+
+ WARN_ON(! pmc_owner_caller);
+
+ pmc_owner_caller = NULL;
+ perf_irq = dummy_perf;
+
+ spin_unlock(&pmc_owner_lock);
+}
Index: working-2.6/arch/ppc64/kernel/traps.c
===================================================================
--- working-2.6.orig/arch/ppc64/kernel/traps.c 2005-01-10 10:51:31.000000000 +1100
+++ working-2.6/arch/ppc64/kernel/traps.c 2005-01-10 16:33:43.154412536 +1100
@@ -40,6 +40,7 @@
#include <asm/rtas.h>
#include <asm/systemcfg.h>
#include <asm/machdep.h>
+#include <asm/pmc.h>
#ifdef CONFIG_DEBUGGER
int (*__debugger)(struct pt_regs *regs);
@@ -449,18 +450,7 @@
die("Unrecoverable VMX/Altivec Unavailable Exception", regs, SIGABRT);
}
-/* Ensure exceptions are disabled */
-static void dummy_perf(struct pt_regs *regs)
-{
- unsigned int mmcr0 = mfspr(SPRN_MMCR0);
-
- mmcr0 &= ~(MMCR0_PMXE|MMCR0_PMAO);
- mtspr(SPRN_MMCR0, mmcr0);
-}
-
-void (*perf_irq)(struct pt_regs *) = dummy_perf;
-
-EXPORT_SYMBOL(perf_irq);
+extern perf_irq_t perf_irq;
void performance_monitor_exception(struct pt_regs *regs)
{
Index: working-2.6/include/asm-ppc64/pmc.h
===================================================================
--- /dev/null 1970-01-01 00:00:00.000000000 +0000
+++ working-2.6/include/asm-ppc64/pmc.h 2005-01-10 15:24:40.217406672 +1100
@@ -0,0 +1,29 @@
+/*
+ * pmc.h
+ * Copyright (C) 2004 David Gibson, IBM Corporation
+ *
+ * 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
+ */
+#ifndef _PPC64_PMC_H
+#define _PPC64_PMC_H
+
+#include <asm/ptrace.h>
+
+typedef void (*perf_irq_t)(struct pt_regs *);
+
+int reserve_pmc_hardware(perf_irq_t new_perf_irq);
+void release_pmc_hardware(void);
+
+#endif /* _PPC64_PMC_H */
Index: working-2.6/arch/ppc64/kernel/Makefile
===================================================================
--- working-2.6.orig/arch/ppc64/kernel/Makefile 2005-01-10 10:51:31.000000000 +1100
+++ working-2.6/arch/ppc64/kernel/Makefile 2005-01-10 15:24:40.218406520 +1100
@@ -11,7 +11,7 @@
udbg.o binfmt_elf32.o sys_ppc32.o ioctl32.o \
ptrace32.o signal32.o rtc.o init_task.o \
lmb.o cputable.o cpu_setup_power4.o idle_power4.o \
- iommu.o sysfs.o
+ iommu.o sysfs.o pmc.o
obj-$(CONFIG_PPC_OF) += of_device.o
--
David Gibson | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist. NOT _the_ _other_ _way_
| _around_!
http://www.ozlabs.org/people/dgibson
On Tue, Jan 11, 2005 at 05:01:27AM +1100, David Gibson wrote:
> This patch creates functions to reserve and release the performance
> monitor hardware (including its interrupt), and makes oprofile use
> them.
I don't see where you make oprofile use the functions? op_model_*
changes aren't included in the patch.
> +int reserve_pmc_hardware(perf_irq_t new_perf_irq)
> +{
> + int err = -EBUSY;;
Keeping an extra semicolon around in case you need one in a hurry? :)
> + spin_lock(&pmc_owner_lock);
> +
> + if (pmc_owner_caller) {
> + printk(KERN_WARNING "reserve_pmc_hardware: "
> + "PMC hardware busy (reserved by caller %p)\n",
> + pmc_owner_caller);
> + goto out;
> + }
> +
> + pmc_owner_caller = __builtin_return_address(0);
> + perf_irq = new_perf_irq ? : dummy_perf;
> +
> + err = 0;
Maybe I'm the only one with such an opinion, but I find it more readable
to set the error code in the error case (if section above) instead of
defaulting to error and clearing it before returning. :)
> + pmc_owner_caller = NULL;
> + perf_irq = dummy_perf;
> +
> + spin_unlock(&pmc_owner_lock);
Current oprofile code has an implicit mb(); after restoring perf_irq. I
think the implied lwsync in spin_unlock is sufficient, but I wanted to
mention it.
How do you expect the function to be used, will there really be users
reserving the hardware without registering the interrupt handler? If
there are no such users then it could be nice to reserve using the
handler instead of the return address.
-Olof
On Mon, Jan 10, 2005 at 04:23:40PM -0600, Olof Johansson wrote:
> On Tue, Jan 11, 2005 at 05:01:27AM +1100, David Gibson wrote:
>
> > This patch creates functions to reserve and release the performance
> > monitor hardware (including its interrupt), and makes oprofile use
> > them.
>
> I don't see where you make oprofile use the functions? op_model_*
> changes aren't included in the patch.
Ah, bugger. I could have sworn I made the changes, wonder where I
managed to drop them.
> > +int reserve_pmc_hardware(perf_irq_t new_perf_irq)
> > +{
> > + int err = -EBUSY;;
>
> Keeping an extra semicolon around in case you need one in a hurry? :)
Oh, dear, I clearly wasn't having a good day.
> > + spin_lock(&pmc_owner_lock);
> > +
> > + if (pmc_owner_caller) {
> > + printk(KERN_WARNING "reserve_pmc_hardware: "
> > + "PMC hardware busy (reserved by caller %p)\n",
> > + pmc_owner_caller);
> > + goto out;
> > + }
> > +
> > + pmc_owner_caller = __builtin_return_address(0);
> > + perf_irq = new_perf_irq ? : dummy_perf;
> > +
> > + err = 0;
>
> Maybe I'm the only one with such an opinion, but I find it more readable
> to set the error code in the error case (if section above) instead of
> defaulting to error and clearing it before returning. :)
Actually, I think I do to, but I've been experimenting with this
style, since it seems to be rather common in the kernel. Anyway,
revised below.
> > + pmc_owner_caller = NULL;
> > + perf_irq = dummy_perf;
> > +
> > + spin_unlock(&pmc_owner_lock);
>
> Current oprofile code has an implicit mb(); after restoring perf_irq. I
> think the implied lwsync in spin_unlock is sufficient, but I wanted to
> mention it.
Yes, I did think about that, and figured the barrier in the
spin_unlock() should be sufficient.
> How do you expect the function to be used, will there really be users
> reserving the hardware without registering the interrupt handler?
I think it's entirely plausible that there could be. It would seem a
bit yucky for a user that wasn't using interrupts to have to define
their own copy of the dummy_perf() routine.
> If
> there are no such users then it could be nice to reserve using the
> handler instead of the return address.
Well, bear in mind that from the semantics point of view it's only the
non-nullness of the return address that matters, so essentially it's
just a flag. The rest of the return address is just there for
debugging convenience.
Anyway, patch with the abovementioned stupidities removed follows.
Andrew, please apply:
The PPC64 interrupt code includes a hook to call when an exception
from the performance monitor unit occurs. However, there's no way of
reserving the hook properly, so if more than one bit of code tries to
use it things will get ugly. Currently oprofile is the only user, but
there are likely to be more in future e.g. perfctr, if and when it
reaches a fit state for merging.
This patch creates functions to reserve and release the performance
monitor hardware (including its interrupt), and makes oprofile use
them. It also creates a new arch/ppc64/kernel/pmc.c, in which we can
put any future helper functions for handling the performance monitor
counters.
Signed-off-by: David Gibson <[email protected]>
Index: working-2.6/arch/ppc64/kernel/pmc.c
===================================================================
--- /dev/null 1970-01-01 00:00:00.000000000 +0000
+++ working-2.6/arch/ppc64/kernel/pmc.c 2005-01-11 10:37:52.001422584 +1100
@@ -0,0 +1,64 @@
+/*
+ * linux/arch/ppc64/kernel/pmc.c
+ *
+ * Copyright (C) 2004 David Gibson, IBM Corporation.
+ *
+ * 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.
+ */
+
+#include <linux/config.h>
+#include <linux/errno.h>
+#include <linux/spinlock.h>
+
+#include <asm/processor.h>
+#include <asm/pmc.h>
+
+/* Ensure exceptions are disabled */
+static void dummy_perf(struct pt_regs *regs)
+{
+ unsigned int mmcr0 = mfspr(SPRN_MMCR0);
+
+ mmcr0 &= ~(MMCR0_PMXE|MMCR0_PMAO);
+ mtspr(SPRN_MMCR0, mmcr0);
+}
+
+static spinlock_t pmc_owner_lock = SPIN_LOCK_UNLOCKED;
+static void *pmc_owner_caller; /* mostly for debugging */
+perf_irq_t perf_irq = dummy_perf;
+
+int reserve_pmc_hardware(perf_irq_t new_perf_irq)
+{
+ int err = 0;
+
+ spin_lock(&pmc_owner_lock);
+
+ if (pmc_owner_caller) {
+ printk(KERN_WARNING "reserve_pmc_hardware: "
+ "PMC hardware busy (reserved by caller %p)\n",
+ pmc_owner_caller);
+ err = -EBUSY;
+ goto out;
+ }
+
+ pmc_owner_caller = __builtin_return_address(0);
+ perf_irq = new_perf_irq ? : dummy_perf;
+
+ out:
+ spin_unlock(&pmc_owner_lock);
+ return err;
+}
+
+void release_pmc_hardware(void)
+{
+ spin_lock(&pmc_owner_lock);
+
+ WARN_ON(! pmc_owner_caller);
+
+ pmc_owner_caller = NULL;
+ perf_irq = dummy_perf;
+
+ spin_unlock(&pmc_owner_lock);
+}
Index: working-2.6/arch/ppc64/kernel/traps.c
===================================================================
--- working-2.6.orig/arch/ppc64/kernel/traps.c 2005-01-11 10:36:44.555424864 +1100
+++ working-2.6/arch/ppc64/kernel/traps.c 2005-01-11 10:36:46.969324088 +1100
@@ -40,6 +40,7 @@
#include <asm/rtas.h>
#include <asm/systemcfg.h>
#include <asm/machdep.h>
+#include <asm/pmc.h>
#ifdef CONFIG_DEBUGGER
int (*__debugger)(struct pt_regs *regs);
@@ -449,18 +450,7 @@
die("Unrecoverable VMX/Altivec Unavailable Exception", regs, SIGABRT);
}
-/* Ensure exceptions are disabled */
-static void dummy_perf(struct pt_regs *regs)
-{
- unsigned int mmcr0 = mfspr(SPRN_MMCR0);
-
- mmcr0 &= ~(MMCR0_PMXE|MMCR0_PMAO);
- mtspr(SPRN_MMCR0, mmcr0);
-}
-
-void (*perf_irq)(struct pt_regs *) = dummy_perf;
-
-EXPORT_SYMBOL(perf_irq);
+extern perf_irq_t perf_irq;
void performance_monitor_exception(struct pt_regs *regs)
{
Index: working-2.6/include/asm-ppc64/pmc.h
===================================================================
--- /dev/null 1970-01-01 00:00:00.000000000 +0000
+++ working-2.6/include/asm-ppc64/pmc.h 2005-01-11 10:36:46.970323936 +1100
@@ -0,0 +1,29 @@
+/*
+ * pmc.h
+ * Copyright (C) 2004 David Gibson, IBM Corporation
+ *
+ * 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
+ */
+#ifndef _PPC64_PMC_H
+#define _PPC64_PMC_H
+
+#include <asm/ptrace.h>
+
+typedef void (*perf_irq_t)(struct pt_regs *);
+
+int reserve_pmc_hardware(perf_irq_t new_perf_irq);
+void release_pmc_hardware(void);
+
+#endif /* _PPC64_PMC_H */
Index: working-2.6/arch/ppc64/kernel/Makefile
===================================================================
--- working-2.6.orig/arch/ppc64/kernel/Makefile 2005-01-11 10:36:44.555424864 +1100
+++ working-2.6/arch/ppc64/kernel/Makefile 2005-01-11 10:36:46.970323936 +1100
@@ -11,7 +11,7 @@
udbg.o binfmt_elf32.o sys_ppc32.o ioctl32.o \
ptrace32.o signal32.o rtc.o init_task.o \
lmb.o cputable.o cpu_setup_power4.o idle_power4.o \
- iommu.o sysfs.o
+ iommu.o sysfs.o pmc.o
obj-$(CONFIG_PPC_OF) += of_device.o
Index: working-2.6/arch/ppc64/oprofile/common.c
===================================================================
--- working-2.6.orig/arch/ppc64/oprofile/common.c 2005-01-06 10:47:48.000000000 +1100
+++ working-2.6/arch/ppc64/oprofile/common.c 2005-01-11 10:42:26.788317488 +1100
@@ -15,6 +15,7 @@
#include <linux/errno.h>
#include <asm/ptrace.h>
#include <asm/system.h>
+#include <asm/pmc.h>
#include "op_impl.h"
@@ -22,9 +23,6 @@
extern struct op_ppc64_model op_model_power4;
static struct op_ppc64_model *model;
-extern void (*perf_irq)(struct pt_regs *);
-static void (*save_perf_irq)(struct pt_regs *);
-
static struct op_counter_config ctr[OP_MAX_COUNTER];
static struct op_system_config sys;
@@ -35,11 +33,12 @@
static int op_ppc64_setup(void)
{
- /* Install our interrupt handler into the existing hook. */
- save_perf_irq = perf_irq;
- perf_irq = op_handle_interrupt;
+ int err;
- mb();
+ /* Grab the hardware */
+ err = reserve_pmc_hardware(op_handle_interrupt);
+ if (err)
+ return err;
/* Pre-compute the values to stuff in the hardware registers. */
model->reg_setup(ctr, &sys, model->num_counters);
@@ -52,10 +51,7 @@
static void op_ppc64_shutdown(void)
{
- mb();
-
- /* Remove our interrupt handler. We may be removing this module. */
- perf_irq = save_perf_irq;
+ release_pmc_hardware();
}
static void op_ppc64_cpu_start(void *dummy)
--
David Gibson | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist. NOT _the_ _other_ _way_
| _around_!
http://www.ozlabs.org/people/dgibson