[PATCH v2] powerpc: Allow perf_counters to access user memory at interrupt time

[PATCH v2] powerpc: Allow perf_counters to access user memory at interrupt time

[Paul Mackerras]

This provides a mechanism to allow the perf_counters code to access
user memory in a PMU interrupt routine.  Such an access can cause
various kinds of interrupt: SLB miss, MMU hash table miss, segment
table miss, or TLB miss, depending on the processor.  This commit
only deals with the classic/server processors that use an MMU hash
table, not processors that have software-loaded TLBs.

On 64-bit processors, an SLB miss interrupt on a user address will
update the slb_cache and slb_cache_ptr fields in the paca.  This is
OK except in the case where a PMU interrupt occurs in switch_slb,
which also accesses those fields.  To prevent this, we hard-disable
interrupts in switch_slb.  Interrupts are already soft-disabled at
this point, and will get hard-enabled when they get soft-enabled
later.

This also reworks slb_flush_and_rebolt: to avoid hard-disabling twice,
and to make sure that it clears the slb_cache_ptr when called from
other callers than switch_slb, the existing routine is renamed to
__slb_flush_and_rebolt, which is called by switch_slb and the new
version of slb_flush_and_rebolt.

Similarly, switch_stab (used on POWER3 and RS64 processors) gets a
hard_irq_disable() to protect the per-cpu variables used there and
in ste_allocate.

If a MMU hashtable miss interrupt occurs, normally we would call
hash_page to look up the Linux PTE for the address and create a HPTE.
However, hash_page is fairly complex and takes some locks, so to
avoid the possibility of deadlock, we check the preemption count
to see if we are in a (pseudo-)NMI handler, and if so, we don't call
hash_page but instead treat it like a bad access that will get
reported up through the exception table mechanism.  An interrupt
whose handler runs even though the interrupt occurred when
soft-disabled (such as the PMU interrupt) is considered a pseudo-NMI
handler, which should use nmi_enter()/nmi_exit() rather than
irq_enter()/irq_exit().

32-bit processors with an MMU hash table are already able to access
user memory at interrupt time.  Since we don't soft-disable on 32-bit,
we avoid the possibility of reentering hash_page, which runs with
interrupts disabled.

Signed-off-by: Paul Mackerras <paulus@samba.org>
---
Note, this version uses the NMI bit in the preempt count instead of
adding a paca field.

 arch/powerpc/include/asm/paca.h      |    2 +-
 arch/powerpc/kernel/asm-offsets.c    |    2 +
 arch/powerpc/kernel/exceptions-64s.S |   19 +++++++++++++++++
 arch/powerpc/mm/slb.c                |   37 +++++++++++++++++++++++----------
 arch/powerpc/mm/stab.c               |   11 +++++++++-
 5 files changed, 58 insertions(+), 13 deletions(-)

diff --git a/arch/powerpc/include/asm/paca.h b/arch/powerpc/include/asm/paca.h
index c8a3cbf..63f8415 100644
--- a/arch/powerpc/include/asm/paca.h
+++ b/arch/powerpc/include/asm/paca.h
@@ -105,7 +105,7 @@ struct paca_struct {
 	u8 soft_enabled;		/* irq soft-enable flag */
 	u8 hard_enabled;		/* set if irqs are enabled in MSR */
 	u8 io_sync;			/* writel() needs spin_unlock sync */
-	u8 perf_counter_pending;	/* PM interrupt while soft-disabled */
+	u8 perf_counter_pending;	/* perf_counter stuff needs wakeup */
 
 	/* Stuff for accurate time accounting */
 	u64 user_time;			/* accumulated usermode TB ticks */
diff --git a/arch/powerpc/kernel/asm-offsets.c b/arch/powerpc/kernel/asm-offsets.c
index 561b646..197b156 100644
--- a/arch/powerpc/kernel/asm-offsets.c
+++ b/arch/powerpc/kernel/asm-offsets.c
@@ -67,6 +67,8 @@ int main(void)
 	DEFINE(MMCONTEXTID, offsetof(struct mm_struct, context.id));
 #ifdef CONFIG_PPC64
 	DEFINE(AUDITCONTEXT, offsetof(struct task_struct, audit_context));
+	DEFINE(SIGSEGV, SIGSEGV);
+	DEFINE(NMI_MASK, NMI_MASK);
 #else
 	DEFINE(THREAD_INFO, offsetof(struct task_struct, stack));
 #endif /* CONFIG_PPC64 */
diff --git a/arch/powerpc/kernel/exceptions-64s.S b/arch/powerpc/kernel/exceptions-64s.S
index eb89811..8ac85e0 100644
--- a/arch/powerpc/kernel/exceptions-64s.S
+++ b/arch/powerpc/kernel/exceptions-64s.S
@@ -729,6 +729,11 @@ BEGIN_FTR_SECTION
 	bne-	do_ste_alloc		/* If so handle it */
 END_FTR_SECTION_IFCLR(CPU_FTR_SLB)
 
+	clrrdi	r11,r1,THREAD_SHIFT
+	lwz	r0,TI_PREEMPT(r11)	/* If we're in an "NMI" */
+	andis.	r0,r0,NMI_MASK@h	/* (i.e. an irq when soft-disabled) */
+	bne	77f			/* then don't call hash_page now */
+
 	/*
 	 * On iSeries, we soft-disable interrupts here, then
 	 * hard-enable interrupts so that the hash_page code can spin on
@@ -833,6 +838,20 @@ handle_page_fault:
 	bl	.low_hash_fault
 	b	.ret_from_except
 
+/*
+ * We come here as a result of a DSI at a point where we don't want
+ * to call hash_page, such as when we are accessing memory (possibly
+ * user memory) inside a PMU interrupt that occurred while interrupts
+ * were soft-disabled.  We want to invoke the exception handler for
+ * the access, or panic if there isn't a handler.
+ */
+77:	bl	.save_nvgprs
+	mr	r4,r3
+	addi	r3,r1,STACK_FRAME_OVERHEAD
+	li	r5,SIGSEGV
+	bl	.bad_page_fault
+	b	.ret_from_except
+
 	/* here we have a segment miss */
 do_ste_alloc:
 	bl	.ste_allocate		/* try to insert stab entry */
diff --git a/arch/powerpc/mm/slb.c b/arch/powerpc/mm/slb.c
index 5b7038f..a685652 100644
--- a/arch/powerpc/mm/slb.c
+++ b/arch/powerpc/mm/slb.c
@@ -92,15 +92,13 @@ static inline void create_shadowed_slbe(unsigned long ea, int ssize,
 		     : "memory" );
 }
 
-void slb_flush_and_rebolt(void)
+static void __slb_flush_and_rebolt(void)
 {
 	/* If you change this make sure you change SLB_NUM_BOLTED
 	 * appropriately too. */
 	unsigned long linear_llp, vmalloc_llp, lflags, vflags;
 	unsigned long ksp_esid_data, ksp_vsid_data;
 
-	WARN_ON(!irqs_disabled());
-
 	linear_llp = mmu_psize_defs[mmu_linear_psize].sllp;
 	vmalloc_llp = mmu_psize_defs[mmu_vmalloc_psize].sllp;
 	lflags = SLB_VSID_KERNEL | linear_llp;
@@ -117,12 +115,6 @@ void slb_flush_and_rebolt(void)
 		ksp_vsid_data = get_slb_shadow()->save_area[2].vsid;
 	}
 
-	/*
-	 * We can't take a PMU exception in the following code, so hard
-	 * disable interrupts.
-	 */
-	hard_irq_disable();
-
 	/* We need to do this all in asm, so we're sure we don't touch
 	 * the stack between the slbia and rebolting it. */
 	asm volatile("isync\n"
@@ -139,6 +131,21 @@ void slb_flush_and_rebolt(void)
 		     : "memory");
 }
 
+void slb_flush_and_rebolt(void)
+{
+
+	WARN_ON(!irqs_disabled());
+
+	/*
+	 * We can't take a PMU exception in the following code, so hard
+	 * disable interrupts.
+	 */
+	hard_irq_disable();
+
+	__slb_flush_and_rebolt();
+	get_paca()->slb_cache_ptr = 0;
+}
+
 void slb_vmalloc_update(void)
 {
 	unsigned long vflags;
@@ -180,12 +187,20 @@ static inline int esids_match(unsigned long addr1, unsigned long addr2)
 /* Flush all user entries from the segment table of the current processor. */
 void switch_slb(struct task_struct *tsk, struct mm_struct *mm)
 {
-	unsigned long offset = get_paca()->slb_cache_ptr;
+	unsigned long offset;
 	unsigned long slbie_data = 0;
 	unsigned long pc = KSTK_EIP(tsk);
 	unsigned long stack = KSTK_ESP(tsk);
 	unsigned long unmapped_base;
 
+	/*
+	 * We need interrupts hard-disabled here, not just soft-disabled,
+	 * so that a PMU interrupt can't occur, which might try to access
+	 * user memory (to get a stack trace) and possible cause an SLB miss
+	 * which would update the slb_cache/slb_cache_ptr fields in the PACA.
+	 */
+	hard_irq_disable();
+	offset = get_paca()->slb_cache_ptr;
 	if (!cpu_has_feature(CPU_FTR_NO_SLBIE_B) &&
 	    offset <= SLB_CACHE_ENTRIES) {
 		int i;
@@ -200,7 +215,7 @@ void switch_slb(struct task_struct *tsk, struct mm_struct *mm)
 		}
 		asm volatile("isync" : : : "memory");
 	} else {
-		slb_flush_and_rebolt();
+		__slb_flush_and_rebolt();
 	}
 
 	/* Workaround POWER5 < DD2.1 issue */
diff --git a/arch/powerpc/mm/stab.c b/arch/powerpc/mm/stab.c
index 98cd1dc..ab5fb48 100644
--- a/arch/powerpc/mm/stab.c
+++ b/arch/powerpc/mm/stab.c
@@ -164,7 +164,7 @@ void switch_stab(struct task_struct *tsk, struct mm_struct *mm)
 {
 	struct stab_entry *stab = (struct stab_entry *) get_paca()->stab_addr;
 	struct stab_entry *ste;
-	unsigned long offset = __get_cpu_var(stab_cache_ptr);
+	unsigned long offset;
 	unsigned long pc = KSTK_EIP(tsk);
 	unsigned long stack = KSTK_ESP(tsk);
 	unsigned long unmapped_base;
@@ -172,6 +172,15 @@ void switch_stab(struct task_struct *tsk, struct mm_struct *mm)
 	/* Force previous translations to complete. DRENG */
 	asm volatile("isync" : : : "memory");
 
+	/*
+	 * We need interrupts hard-disabled here, not just soft-disabled,
+	 * so that a PMU interrupt can't occur, which might try to access
+	 * user memory (to get a stack trace) and possible cause an STAB miss
+	 * which would update the stab_cache/stab_cache_ptr per-cpu variables.
+	 */
+	hard_irq_disable();
+
+	offset = __get_cpu_var(stab_cache_ptr);
 	if (offset <= NR_STAB_CACHE_ENTRIES) {
 		int i;
 
-- 
1.5.5.rc3.7.gba13

[PATCH v2] perf_counter: powerpc: Add callchain support

[Paul Mackerras]

This adds support for tracing callchains for powerpc, both 32-bit
and 64-bit, and both in the kernel and userspace, from PMU interrupt
context.

The first three entries stored for each callchain are the NIP (next
instruction pointer), LR (link register), and the contents of the LR
save area in the second stack frame (the first is ignored because the
ABI convention on powerpc is that functions save their return address
in their caller's stack frame).  Because functions don't have to save
their return address (LR value) and don't have to establish a stack
frame, it's possible for either or both of LR and the second stack
frame's LR save area to have valid return addresses in them.  This
is basically impossible to disambiguate without either reading the
code or looking at auxiliary information such as CFI tables.  Since
we don't want to do that at interrupt time, we store both LR and the
second stack frame's LR save area.

Once we get past the second stack frame, there is no ambiguity; all
return addresses we get are reliable.

For kernel traces, we check whether they are valid kernel instruction
addresses and store zero instead if they are not (rather than
omitting them, which would make it impossible for userspace to know
which was which).  We also store zero instead of the second stack
frame's LR save area value if it is the same as LR.

For kernel traces, we check for interrupt frames, and for user traces,
we check for signal frames.  In each case, since we're starting a new
trace, we store a PERF=5FCONTEXT=5FKERNEL/USER marker so that userspace=

knows that the next three entries are NIP, LR and the second stack fram=
e
for the interrupted context.

We read user memory with =5F=5Fget=5Fuser=5Finatomic.  On 64-bit, if th=
is
PMU interrupt occurred while interrupts are soft-disabled, and
there is no MMU hash table entry for the page, we will get an
-EFAULT return from =5F=5Fget=5Fuser=5Finatomic even if there is a vali=
d
Linux PTE for the page, since hash=5Fpage isn't reentrant.  Thus we
have code here to read the Linux PTE and access the page via the
kernel linear mapping.  Since 64-bit doesn't use (or need) highmem
there is no need to do kmap=5Fatomic.  On 32-bit, we don't do soft
interrupt disabling, so this complication doesn't occur and there
is no need to fall back to reading the Linux PTE, since hash=5Fpage
will get called automatically if necessary.

Note that we cannot get PMU interrupts in the interval during
context switch between switch=5Fmm (which switches the user address
space) and switch=5Fto (which actually changes current to the new
process).  On 64-bit this is because interrupts are hard-disabled
in switch=5Fmm and stay hard-disabled until they are soft-enabled
later, after switch=5Fto has returned.  So there is no possibility
of trying to do a user stack trace when the user address space is
not current's address space.

Signed-off-by: Paul Mackerras <paulus@samba.org>
---
 arch/powerpc/kernel/Makefile         |    2 +-
 arch/powerpc/kernel/perf=5Fcallchain.c |  520 ++++++++++++++++++++++++=
++++++++++
 2 files changed, 521 insertions(+), 1 deletions(-)
 create mode 100644 arch/powerpc/kernel/perf=5Fcallchain.c

diff --git a/arch/powerpc/kernel/Makefile b/arch/powerpc/kernel/Makefil=
e
index b73396b..9619285 100644
--- a/arch/powerpc/kernel/Makefile
+++ b/arch/powerpc/kernel/Makefile
@@ -97,7 +97,7 @@ obj64-$(CONFIG=5FAUDIT)=09=09+=3D compat=5Faudit.o
=20
 obj-$(CONFIG=5FDYNAMIC=5FFTRACE)=09+=3D ftrace.o
 obj-$(CONFIG=5FFUNCTION=5FGRAPH=5FTRACER)=09+=3D ftrace.o
-obj-$(CONFIG=5FPPC=5FPERF=5FCTRS)=09+=3D perf=5Fcounter.o
+obj-$(CONFIG=5FPPC=5FPERF=5FCTRS)=09+=3D perf=5Fcounter.o perf=5Fcallc=
hain.o
 obj64-$(CONFIG=5FPPC=5FPERF=5FCTRS)=09+=3D power4-pmu.o ppc970-pmu.o p=
ower5-pmu.o \
 =09=09=09=09   power5+-pmu.o power6-pmu.o power7-pmu.o
 obj32-$(CONFIG=5FPPC=5FPERF=5FCTRS)=09+=3D mpc7450-pmu.o
diff --git a/arch/powerpc/kernel/perf=5Fcallchain.c b/arch/powerpc/kern=
el/perf=5Fcallchain.c
new file mode 100644
index 0000000..ed13777
--- /dev/null
+++ b/arch/powerpc/kernel/perf=5Fcallchain.c
@@ -0,0 +1,520 @@
+/*
+ * Performance counter callchain support - powerpc architecture code
+ *
+ * Copyright =A9 2009 Paul Mackerras, 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/kernel.h>
+#include <linux/sched.h>
+#include <linux/perf=5Fcounter.h>
+#include <linux/percpu.h>
+#include <linux/uaccess.h>
+#include <linux/mm.h>
+#include <asm/ptrace.h>
+#include <asm/pgtable.h>
+#include <asm/sigcontext.h>
+#include <asm/ucontext.h>
+#include <asm/vdso.h>
+#ifdef CONFIG=5FPPC64
+#include "ppc32.h"
+#endif
+
+/*
+ * Store another value in a callchain=5Fentry.
+ */
+static inline void callchain=5Fstore(struct perf=5Fcallchain=5Fentry *=
entry, u64 ip)
+{
+=09unsigned int nr =3D entry->nr;
+
+=09if (nr < PERF=5FMAX=5FSTACK=5FDEPTH) {
+=09=09entry->ip[nr] =3D ip;
+=09=09entry->nr =3D nr + 1;
+=09}
+}
+
+/*
+ * Is sp valid as the address of the next kernel stack frame after pre=
v=5Fsp=3F
+ * The next frame may be in a different stack area but should not go
+ * back down in the same stack area.
+ */
+static int valid=5Fnext=5Fsp(unsigned long sp, unsigned long prev=5Fsp=
)
+{
+=09if (sp & 0xf)
+=09=09return 0;=09=09/* must be 16-byte aligned */
+=09if (!validate=5Fsp(sp, current, STACK=5FFRAME=5FOVERHEAD))
+=09=09return 0;
+=09if (sp >=3D prev=5Fsp + STACK=5FFRAME=5FOVERHEAD)
+=09=09return 1;
+=09/*
+=09 * sp could decrease when we jump off an interrupt stack
+=09 * back to the regular process stack.
+=09 */
+=09if ((sp & ~(THREAD=5FSIZE - 1)) !=3D (prev=5Fsp & ~(THREAD=5FSIZE -=
 1)))
+=09=09return 1;
+=09return 0;
+}
+
+static void perf=5Fcallchain=5Fkernel(struct pt=5Fregs *regs,
+=09=09=09=09  struct perf=5Fcallchain=5Fentry *entry)
+{
+=09unsigned long sp, next=5Fsp;
+=09unsigned long next=5Fip;
+=09unsigned long lr;
+=09long level =3D 0;
+=09unsigned long *fp;
+
+=09lr =3D regs->link;
+=09sp =3D regs->gpr[1];
+=09callchain=5Fstore(entry, PERF=5FCONTEXT=5FKERNEL);
+=09callchain=5Fstore(entry, regs->nip);
+
+=09if (!validate=5Fsp(sp, current, STACK=5FFRAME=5FOVERHEAD))
+=09=09return;
+
+=09for (;;) {
+=09=09fp =3D (unsigned long *) sp;
+=09=09next=5Fsp =3D fp[0];
+
+=09=09if (next=5Fsp =3D=3D sp + STACK=5FINT=5FFRAME=5FSIZE &&
+=09=09    fp[STACK=5FFRAME=5FMARKER] =3D=3D STACK=5FFRAME=5FREGS=5FMAR=
KER) {
+=09=09=09/*
+=09=09=09 * This looks like an interrupt frame for an
+=09=09=09 * interrupt that occurred in the kernel
+=09=09=09 */
+=09=09=09regs =3D (struct pt=5Fregs *)(sp + STACK=5FFRAME=5FOVERHEAD);=

+=09=09=09next=5Fip =3D regs->nip;
+=09=09=09lr =3D regs->link;
+=09=09=09level =3D 0;
+=09=09=09callchain=5Fstore(entry, PERF=5FCONTEXT=5FKERNEL);
+
+=09=09} else {
+=09=09=09if (level =3D=3D 0)
+=09=09=09=09next=5Fip =3D lr;
+=09=09=09else
+=09=09=09=09next=5Fip =3D fp[STACK=5FFRAME=5FLR=5FSAVE];
+
+=09=09=09/*
+=09=09=09 * We can't tell which of the first two addresses
+=09=09=09 * we get are valid, but we can filter out the
+=09=09=09 * obviously bogus ones here.  We replace them
+=09=09=09 * with 0 rather than removing them entirely so
+=09=09=09 * that userspace can tell which is which.
+=09=09=09 */
+=09=09=09if ((level =3D=3D 1 && next=5Fip =3D=3D lr) ||
+=09=09=09    (level <=3D 1 && !kernel=5Ftext=5Faddress(next=5Fip)))
+=09=09=09=09next=5Fip =3D 0;
+
+=09=09=09++level;
+=09=09}
+
+=09=09callchain=5Fstore(entry, next=5Fip);
+=09=09if (!valid=5Fnext=5Fsp(next=5Fsp, sp))
+=09=09=09return;
+=09=09sp =3D next=5Fsp;
+=09}
+}
+
+#ifdef CONFIG=5FPPC64
+/*
+ * On 64-bit we don't want to invoke hash=5Fpage on user addresses fro=
m
+ * interrupt context, so if the access faults, we read the page tables=

+ * to find which page (if any) is mapped and access it directly.
+ */
+static int read=5Fuser=5Fstack=5Fslow(void =5F=5Fuser *ptr, void *ret,=
 int nb)
+{
+=09pgd=5Ft *pgdir;
+=09pte=5Ft *ptep, pte;
+=09int pagesize;
+=09unsigned long addr =3D (unsigned long) ptr;
+=09unsigned long offset;
+=09unsigned long pfn;
+=09void *kaddr;
+
+=09pgdir =3D current->mm->pgd;
+=09if (!pgdir)
+=09=09return -EFAULT;
+
+=09pagesize =3D get=5Fslice=5Fpsize(current->mm, addr);
+
+=09/* align address to page boundary */
+=09offset =3D addr & ((1ul << mmu=5Fpsize=5Fdefs[pagesize].shift) - 1)=
;
+=09addr -=3D offset;
+
+=09if (HPAGE=5FSHIFT && mmu=5Fhuge=5Fpsizes[pagesize])
+=09=09ptep =3D huge=5Fpte=5Foffset(current->mm, addr);
+=09else
+=09=09ptep =3D find=5Flinux=5Fpte(pgdir, addr);
+
+=09if (ptep =3D=3D NULL)
+=09=09return -EFAULT;
+=09pte =3D *ptep;
+=09if (!pte=5Fpresent(pte) || !(pte=5Fval(pte) & =5FPAGE=5FUSER))
+=09=09return -EFAULT;
+=09pfn =3D pte=5Fpfn(pte);
+=09if (!page=5Fis=5Fram(pfn))
+=09=09return -EFAULT;
+
+=09/* no highmem to worry about here */
+=09kaddr =3D pfn=5Fto=5Fkaddr(pfn);
+=09memcpy(ret, kaddr + offset, nb);
+=09return 0;
+}
+
+static int read=5Fuser=5Fstack=5F64(unsigned long =5F=5Fuser *ptr, uns=
igned long *ret)
+{
+=09if ((unsigned long)ptr > TASK=5FSIZE - sizeof(unsigned long) ||
+=09    ((unsigned long)ptr & 7))
+=09=09return -EFAULT;
+
+=09if (!=5F=5Fget=5Fuser=5Finatomic(*ret, ptr))
+=09=09return 0;
+
+=09return read=5Fuser=5Fstack=5Fslow(ptr, ret, 8);
+}
+
+static int read=5Fuser=5Fstack=5F32(unsigned int =5F=5Fuser *ptr, unsi=
gned int *ret)
+{
+=09if ((unsigned long)ptr > TASK=5FSIZE - sizeof(unsigned int) ||
+=09    ((unsigned long)ptr & 3))
+=09=09return -EFAULT;
+
+=09if (!=5F=5Fget=5Fuser=5Finatomic(*ret, ptr))
+=09=09return 0;
+
+=09return read=5Fuser=5Fstack=5Fslow(ptr, ret, 4);
+}
+
+static inline int valid=5Fuser=5Fsp(unsigned long sp, int is=5F64)
+{
+=09if (!sp || (sp & 7) || sp > (is=5F64 =3F TASK=5FSIZE : 0x100000000U=
L) - 32)
+=09=09return 0;
+=09return 1;
+}
+
+/*
+ * 64-bit user processes use the same stack frame for RT and non-RT si=
gnals.
+ */
+struct signal=5Fframe=5F64 {
+=09char=09=09dummy[=5F=5FSIGNAL=5FFRAMESIZE];
+=09struct ucontext=09uc;
+=09unsigned long=09unused[2];
+=09unsigned int=09tramp[6];
+=09struct siginfo=09*pinfo;
+=09void=09=09*puc;
+=09struct siginfo=09info;
+=09char=09=09abigap[288];
+};
+
+static int is=5Fsigreturn=5F64=5Faddress(unsigned long nip, unsigned l=
ong fp)
+{
+=09if (nip =3D=3D fp + offsetof(struct signal=5Fframe=5F64, tramp))
+=09=09return 1;
+=09if (vdso64=5Frt=5Fsigtramp && current->mm->context.vdso=5Fbase &&
+=09    nip =3D=3D current->mm->context.vdso=5Fbase + vdso64=5Frt=5Fsig=
tramp)
+=09=09return 1;
+=09return 0;
+}
+
+/*
+ * Do some sanity checking on the signal frame pointed to by sp.
+ * We check the pinfo and puc pointers in the frame.
+ */
+static int sane=5Fsignal=5F64=5Fframe(unsigned long sp)
+{
+=09struct signal=5Fframe=5F64 =5F=5Fuser *sf;
+=09unsigned long pinfo, puc;
+
+=09sf =3D (struct signal=5Fframe=5F64 =5F=5Fuser *) sp;
+=09if (read=5Fuser=5Fstack=5F64((unsigned long =5F=5Fuser *) &sf->pinf=
o, &pinfo) ||
+=09    read=5Fuser=5Fstack=5F64((unsigned long =5F=5Fuser *) &sf->puc,=
 &puc))
+=09=09return 0;
+=09return pinfo =3D=3D (unsigned long) &sf->info &&
+=09=09puc =3D=3D (unsigned long) &sf->uc;
+}
+
+static void perf=5Fcallchain=5Fuser=5F64(struct pt=5Fregs *regs,
+=09=09=09=09   struct perf=5Fcallchain=5Fentry *entry)
+{
+=09unsigned long sp, next=5Fsp;
+=09unsigned long next=5Fip;
+=09unsigned long lr;
+=09long level =3D 0;
+=09struct signal=5Fframe=5F64 =5F=5Fuser *sigframe;
+=09unsigned long =5F=5Fuser *fp, *uregs;
+
+=09next=5Fip =3D regs->nip;
+=09lr =3D regs->link;
+=09sp =3D regs->gpr[1];
+=09callchain=5Fstore(entry, PERF=5FCONTEXT=5FUSER);
+=09callchain=5Fstore(entry, next=5Fip);
+
+=09for (;;) {
+=09=09fp =3D (unsigned long =5F=5Fuser *) sp;
+=09=09if (!valid=5Fuser=5Fsp(sp, 1) || read=5Fuser=5Fstack=5F64(fp, &n=
ext=5Fsp))
+=09=09=09return;
+=09=09if (level > 0 && read=5Fuser=5Fstack=5F64(&fp[2], &next=5Fip))
+=09=09=09return;
+
+=09=09/*
+=09=09 * Note: the next=5Fsp - sp >=3D signal frame size check
+=09=09 * is true when next=5Fsp < sp, which can happen when
+=09=09 * transitioning from an alternate signal stack to the
+=09=09 * normal stack.
+=09=09 */
+=09=09if (next=5Fsp - sp >=3D sizeof(struct signal=5Fframe=5F64) &&
+=09=09    (is=5Fsigreturn=5F64=5Faddress(next=5Fip, sp) ||
+=09=09     (level <=3D 1 && is=5Fsigreturn=5F64=5Faddress(lr, sp))) &&=

+=09=09    sane=5Fsignal=5F64=5Fframe(sp)) {
+=09=09=09/*
+=09=09=09 * This looks like an signal frame
+=09=09=09 */
+=09=09=09sigframe =3D (struct signal=5Fframe=5F64 =5F=5Fuser *) sp;
+=09=09=09uregs =3D sigframe->uc.uc=5Fmcontext.gp=5Fregs;
+=09=09=09if (read=5Fuser=5Fstack=5F64(&uregs[PT=5FNIP], &next=5Fip) ||=

+=09=09=09    read=5Fuser=5Fstack=5F64(&uregs[PT=5FLNK], &lr) ||
+=09=09=09    read=5Fuser=5Fstack=5F64(&uregs[PT=5FR1], &sp))
+=09=09=09=09return;
+=09=09=09level =3D 0;
+=09=09=09callchain=5Fstore(entry, PERF=5FCONTEXT=5FUSER);
+=09=09=09callchain=5Fstore(entry, next=5Fip);
+=09=09=09continue;
+=09=09}
+
+=09=09if (level =3D=3D 0)
+=09=09=09next=5Fip =3D lr;
+=09=09callchain=5Fstore(entry, next=5Fip);
+=09=09++level;
+=09=09sp =3D next=5Fsp;
+=09}
+}
+
+static inline int current=5Fis=5F64bit(void)
+{
+=09/*
+=09 * We can't use test=5Fthread=5Fflag() here because we may be on an=

+=09 * interrupt stack, and the thread flags don't get copied over
+=09 * from the thread=5Finfo on the main stack to the interrupt stack.=

+=09 */
+=09return !test=5Fti=5Fthread=5Fflag(task=5Fthread=5Finfo(current), TI=
F=5F32BIT);
+}
+
+#else  /* CONFIG=5FPPC64 */
+/*
+ * On 32-bit we just access the address and let hash=5Fpage create a
+ * HPTE if necessary, so there is no need to fall back to reading
+ * the page tables.  Since this is called at interrupt level,
+ * do=5Fpage=5Ffault() won't treat a DSI as a page fault.
+ */
+static int read=5Fuser=5Fstack=5F32(unsigned int =5F=5Fuser *ptr, unsi=
gned int *ret)
+{
+=09if ((unsigned long)ptr > TASK=5FSIZE - sizeof(unsigned int) ||
+=09    ((unsigned long)ptr & 3))
+=09=09return -EFAULT;
+
+=09return =5F=5Fget=5Fuser=5Finatomic(*ret, ptr);
+}
+
+static inline void perf=5Fcallchain=5Fuser=5F64(struct pt=5Fregs *regs=
,
+=09=09=09=09=09  struct perf=5Fcallchain=5Fentry *entry)
+{
+}
+
+static inline int current=5Fis=5F64bit(void)
+{
+=09return 0;
+}
+
+static inline int valid=5Fuser=5Fsp(unsigned long sp, int is=5F64)
+{
+=09if (!sp || (sp & 7) || sp > TASK=5FSIZE - 32)
+=09=09return 0;
+=09return 1;
+}
+
+#define =5F=5FSIGNAL=5FFRAMESIZE32=09=5F=5FSIGNAL=5FFRAMESIZE
+#define sigcontext32=09=09sigcontext
+#define mcontext32=09=09mcontext
+#define ucontext32=09=09ucontext
+#define compat=5Fsiginfo=5Ft=09struct siginfo
+
+#endif /* CONFIG=5FPPC64 */
+
+/*
+ * Layout for non-RT signal frames
+ */
+struct signal=5Fframe=5F32 {
+=09char=09=09=09dummy[=5F=5FSIGNAL=5FFRAMESIZE32];
+=09struct sigcontext32=09sctx;
+=09struct mcontext32=09mctx;
+=09int=09=09=09abigap[56];
+};
+
+/*
+ * Layout for RT signal frames
+ */
+struct rt=5Fsignal=5Fframe=5F32 {
+=09char=09=09=09dummy[=5F=5FSIGNAL=5FFRAMESIZE32 + 16];
+=09compat=5Fsiginfo=5Ft=09info;
+=09struct ucontext32=09uc;
+=09int=09=09=09abigap[56];
+};
+
+static int is=5Fsigreturn=5F32=5Faddress(unsigned int nip, unsigned in=
t fp)
+{
+=09if (nip =3D=3D fp + offsetof(struct signal=5Fframe=5F32, mctx.mc=5F=
pad))
+=09=09return 1;
+=09if (vdso32=5Fsigtramp && current->mm->context.vdso=5Fbase &&
+=09    nip =3D=3D current->mm->context.vdso=5Fbase + vdso32=5Fsigtramp=
)
+=09=09return 1;
+=09return 0;
+}
+
+static int is=5Frt=5Fsigreturn=5F32=5Faddress(unsigned int nip, unsign=
ed int fp)
+{
+=09if (nip =3D=3D fp + offsetof(struct rt=5Fsignal=5Fframe=5F32,
+=09=09=09=09 uc.uc=5Fmcontext.mc=5Fpad))
+=09=09return 1;
+=09if (vdso32=5Frt=5Fsigtramp && current->mm->context.vdso=5Fbase &&
+=09    nip =3D=3D current->mm->context.vdso=5Fbase + vdso32=5Frt=5Fsig=
tramp)
+=09=09return 1;
+=09return 0;
+}
+
+static int sane=5Fsignal=5F32=5Fframe(unsigned int sp)
+{
+=09struct signal=5Fframe=5F32 =5F=5Fuser *sf;
+=09unsigned int regs;
+
+=09sf =3D (struct signal=5Fframe=5F32 =5F=5Fuser *) (unsigned long) sp=
;
+=09if (read=5Fuser=5Fstack=5F32((unsigned int =5F=5Fuser *) &sf->sctx.=
regs, &regs))
+=09=09return 0;
+=09return regs =3D=3D (unsigned long) &sf->mctx;
+}
+
+static int sane=5Frt=5Fsignal=5F32=5Fframe(unsigned int sp)
+{
+=09struct rt=5Fsignal=5Fframe=5F32 =5F=5Fuser *sf;
+=09unsigned int regs;
+
+=09sf =3D (struct rt=5Fsignal=5Fframe=5F32 =5F=5Fuser *) (unsigned lon=
g) sp;
+=09if (read=5Fuser=5Fstack=5F32((unsigned int =5F=5Fuser *) &sf->uc.uc=
=5Fregs, &regs))
+=09=09return 0;
+=09return regs =3D=3D (unsigned long) &sf->uc.uc=5Fmcontext;
+}
+
+static unsigned int =5F=5Fuser *signal=5Fframe=5F32=5Fregs(unsigned in=
t sp,
+=09=09=09=09unsigned int next=5Fsp, unsigned int next=5Fip)
+{
+=09struct mcontext32 =5F=5Fuser *mctx =3D NULL;
+=09struct signal=5Fframe=5F32 =5F=5Fuser *sf;
+=09struct rt=5Fsignal=5Fframe=5F32 =5F=5Fuser *rt=5Fsf;
+
+=09/*
+=09 * Note: the next=5Fsp - sp >=3D signal frame size check
+=09 * is true when next=5Fsp < sp, for example, when
+=09 * transitioning from an alternate signal stack to the
+=09 * normal stack.
+=09 */
+=09if (next=5Fsp - sp >=3D sizeof(struct signal=5Fframe=5F32) &&
+=09    is=5Fsigreturn=5F32=5Faddress(next=5Fip, sp) &&
+=09    sane=5Fsignal=5F32=5Fframe(sp)) {
+=09=09sf =3D (struct signal=5Fframe=5F32 =5F=5Fuser *) (unsigned long)=
 sp;
+=09=09mctx =3D &sf->mctx;
+=09}
+
+=09if (!mctx && next=5Fsp - sp >=3D sizeof(struct rt=5Fsignal=5Fframe=5F=
32) &&
+=09    is=5Frt=5Fsigreturn=5F32=5Faddress(next=5Fip, sp) &&
+=09    sane=5Frt=5Fsignal=5F32=5Fframe(sp)) {
+=09=09rt=5Fsf =3D (struct rt=5Fsignal=5Fframe=5F32 =5F=5Fuser *) (unsi=
gned long) sp;
+=09=09mctx =3D &rt=5Fsf->uc.uc=5Fmcontext;
+=09}
+
+=09if (!mctx)
+=09=09return NULL;
+=09return mctx->mc=5Fgregs;
+}
+
+static void perf=5Fcallchain=5Fuser=5F32(struct pt=5Fregs *regs,
+=09=09=09=09   struct perf=5Fcallchain=5Fentry *entry)
+{
+=09unsigned int sp, next=5Fsp;
+=09unsigned int next=5Fip;
+=09unsigned int lr;
+=09long level =3D 0;
+=09unsigned int =5F=5Fuser *fp, *uregs;
+
+=09next=5Fip =3D regs->nip;
+=09lr =3D regs->link;
+=09sp =3D regs->gpr[1];
+=09callchain=5Fstore(entry, PERF=5FCONTEXT=5FUSER);
+=09callchain=5Fstore(entry, next=5Fip);
+
+=09while (entry->nr < PERF=5FMAX=5FSTACK=5FDEPTH) {
+=09=09fp =3D (unsigned int =5F=5Fuser *) (unsigned long) sp;
+=09=09if (!valid=5Fuser=5Fsp(sp, 0) || read=5Fuser=5Fstack=5F32(fp, &n=
ext=5Fsp))
+=09=09=09return;
+=09=09if (level > 0 && read=5Fuser=5Fstack=5F32(&fp[1], &next=5Fip))
+=09=09=09return;
+
+=09=09uregs =3D signal=5Fframe=5F32=5Fregs(sp, next=5Fsp, next=5Fip);
+=09=09if (!uregs && level <=3D 1)
+=09=09=09uregs =3D signal=5Fframe=5F32=5Fregs(sp, next=5Fsp, lr);
+=09=09if (uregs) {
+=09=09=09/*
+=09=09=09 * This looks like an signal frame, so restart
+=09=09=09 * the stack trace with the values in it.
+=09=09=09 */
+=09=09=09if (read=5Fuser=5Fstack=5F32(&uregs[PT=5FNIP], &next=5Fip) ||=

+=09=09=09    read=5Fuser=5Fstack=5F32(&uregs[PT=5FLNK], &lr) ||
+=09=09=09    read=5Fuser=5Fstack=5F32(&uregs[PT=5FR1], &sp))
+=09=09=09=09return;
+=09=09=09level =3D 0;
+=09=09=09callchain=5Fstore(entry, PERF=5FCONTEXT=5FUSER);
+=09=09=09callchain=5Fstore(entry, next=5Fip);
+=09=09=09continue;
+=09=09}
+
+=09=09if (level =3D=3D 0)
+=09=09=09next=5Fip =3D lr;
+=09=09callchain=5Fstore(entry, next=5Fip);
+=09=09++level;
+=09=09sp =3D next=5Fsp;
+=09}
+}
+
+/*
+ * Since we can't get PMU interrupts inside a PMU interrupt handler,
+ * we don't need separate irq and nmi entries here.
+ */
+static DEFINE=5FPER=5FCPU(struct perf=5Fcallchain=5Fentry, callchain);=

+
+struct perf=5Fcallchain=5Fentry *perf=5Fcallchain(struct pt=5Fregs *re=
gs)
+{
+=09struct perf=5Fcallchain=5Fentry *entry =3D &=5F=5Fget=5Fcpu=5Fvar(c=
allchain);
+
+=09entry->nr =3D 0;
+
+=09if (current->pid =3D=3D 0)=09=09/* idle task=3F */
+=09=09return entry;
+
+=09if (!user=5Fmode(regs)) {
+=09=09perf=5Fcallchain=5Fkernel(regs, entry);
+=09=09if (current->mm)
+=09=09=09regs =3D task=5Fpt=5Fregs(current);
+=09=09else
+=09=09=09regs =3D NULL;
+=09}
+
+=09if (regs) {
+=09=09if (current=5Fis=5F64bit())
+=09=09=09perf=5Fcallchain=5Fuser=5F64(regs, entry);
+=09=09else
+=09=09=09perf=5Fcallchain=5Fuser=5F32(regs, entry);
+=09}
+
+=09return entry;
+}
--=20
1.5.5.rc3.7.gba13

Re: [PATCH v2] powerpc: Allow perf_counters to access user memory at interrupt time

[Benjamin Herrenschmidt]

On Thu, 2009-08-06 at 14:57 +1000, Paul Mackerras wrote:
> This provides a mechanism to allow the perf_counters code to access
> user memory in a PMU interrupt routine.  Such an access can cause
> various kinds of interrupt: SLB miss, MMU hash table miss, segment
> table miss, or TLB miss, depending on the processor.  This commit
> only deals with the classic/server processors that use an MMU hash
> table, not processors that have software-loaded TLBs.

 .../...

> Signed-off-by: Paul Mackerras <paulus@samba.org>

Acked-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>

As discussed in the lab, you should also do a pre-req patch to pgtable.h
that changes ppc32 with 64-bit PTE without CONFIG_SMP to use the same
path as SMP to order the stores to the two halves of the PTEs though.

Cheers,
Ben.

Re: [PATCH v2] perf_counter: powerpc: Add callchain support

[Benjamin Herrenschmidt]

On Thu, 2009-08-06 at 14:58 +1000, Paul Mackerras wrote:

> +
> +#else  /* CONFIG_PPC64 */
> +/*
> + * On 32-bit we just access the address and let hash_page create a
> + * HPTE if necessary, so there is no need to fall back to reading
> + * the page tables.  Since this is called at interrupt level,
> + * do_page_fault() won't treat a DSI as a page fault.
> + */

Minor nit here... The comment makes it think there's only hash based
32-bit processors :-) In fact, there's a little issue with non-hash ones
here, which is that they rely on
do_page_fault->handle_mm_fault->ptep_set_access_flags to set
_PAGE_ACCESSED, and the TLB miss handlers are going to fault if that's
not set.

Not a big deal, but it does mean that if you have stack pages that
aren't young, they will fail to backtrace (though that's probably
unlikely unless you spend a lot of time very deep down a huge call
chain).

> +static int read_user_stack_32(unsigned int __user *ptr, unsigned int *ret)
> +{
> +	if ((unsigned long)ptr > TASK_SIZE - sizeof(unsigned int) ||
> +	    ((unsigned long)ptr & 3))
> +		return -EFAULT;
> +
> +	return __get_user_inatomic(*ret, ptr);
> +}
> +
> +static inline void perf_callchain_user_64(struct pt_regs *regs,
> +					  struct perf_callchain_entry *entry)
> +{
> +}
> +
> +static inline int current_is_64bit(void)
> +{
> +	return 0;
> +}
> +
> +static inline int valid_user_sp(unsigned long sp, int is_64)
> +{
> +	if (!sp || (sp & 7) || sp > TASK_SIZE - 32)

I know the above is right but I would still have preferred () around
TASK_SIZE - 32 :-) In fact, || has lower precedence than & (I checked !)
so in theory if you really wanted to get rid of braces, you could have
written

	if (!sp || sp & 7 || sp > TASK_SIZE - 32)

But heh, that sucks :-)

> +struct signal_frame_32 {
> +	char			dummy[__SIGNAL_FRAMESIZE32];
> +	struct sigcontext32	sctx;
> +	struct mcontext32	mctx;
> +	int			abigap[56];
> +};
> +
> +/*
> + * Layout for RT signal frames
> + */
> +struct rt_signal_frame_32 {
> +	char			dummy[__SIGNAL_FRAMESIZE32 + 16];
> +	compat_siginfo_t	info;
> +	struct ucontext32	uc;
> +	int			abigap[56];
> +};

Should we put those somewhere shared ? They are almost the same
as the ones in signal_32.c apart from the initial gap... oh well, no big
deal if you want to keep them here for now.
 
Overall looks fine and I suppose it also works but I may have missed
something subtle.

Cheers,
Ben.