Re: [PATCH 4/4] trace, powerpc: Implement raw syscall tracepoints on PowerPC

Excerpts from Michael Ellerman's message of Thu May 13 22:09:31 +1000 2010:

On Thu, 2010-05-13 at 17:43 +1000, Ian Munsie wrote:

quoted

From: Ian Munsie <redacted>

Hi Ian,

Just a few comments ..

quoted

This patch implements the raw syscall tracepoints on PowerPC required
for ftrace syscalls.

OK. It also adds a bunch of code under CONFIG_FTRACE_*, so does it
implement raw syscall tracepoints _and_ hook them up to ftrace?

Yes, that's correct. CONFIG_FTRACE_SYSCALLS depends solely on, and is
the primary consumer of, HAVE_SYSCALL_TRACEPOINTS. It makes little sense
to me to provide the raw syscall tracepoints without exporting the
syscall table for ftrace syscalls to use - otherwise they would be
available to select in make config, but broken.

quoted

To minimise reworking existing code, I slightly re-ordered the thread
info flags such that the new TIF_SYSCALL_TRACEPOINT bit would still fit
within the 16 bits of the andi instruction's UI field.

Which andi instruction? That could use a bit more explaining.

The ones under /arch/powerpc/kernel/entry_{32,64}.S using andi to
and the _TIF_SYSCALL_T_OR_A with the thread flags to see if system call
tracing is enabled.
 
For instance, from entry_64.S:
	ld	r10,TI_FLAGS(r11)
	andi.	r11,r10,_TIF_SYSCALL_T_OR_A   <-- that one
	bne-	syscall_dotrace
.......
syscall_dotrace:
	bl	.save_nvgprs
	addi	r3,r1,STACK_FRAME_OVERHEAD
	bl	.do_syscall_trace_enter

And similarly elsewhere in the same file:
	ld	r9,TI_FLAGS(r12)
	li	r11,-_LAST_ERRNO
	andi.	r0,r9,(_TIF_SYSCALL_T_OR_A|_TIF_SINGLESTEP|_TIF_USER_WORK_MASK|_TIF_PERSYSCALL_MASK)
	bne-	syscall_exit_work
.......
syscall_exit_work:
.......
	bl	.save_nvgprs
	addi	r3,r1,STACK_FRAME_OVERHEAD
	bl	.do_syscall_trace_leave

entry_32.S contains very similar assembly to the above.

The alternative to renumbering the thread flags would be to rework the
assembly to use and. instead of andi. to avoid having to squeeze that
flag into 16 bits.

quoted

In the case of 64bit PowerPC, arch_syscall_addr and
arch_syscall_match_sym_name are overridden to allow ftrace syscalls to
work given the unusual system call table structure and symbol names that
start with a period.

Not unusual, just different (ie. better) than x86 ;)

Fair enough ;)

quoted

diff --git a/arch/powerpc/include/asm/syscall.h b/arch/powerpc/include/asm/syscall.h
index 23913e9..4098105 100644
--- a/arch/powerpc/include/asm/syscall.h
+++ b/arch/powerpc/include/asm/syscall.h

@@ -15,6 +15,15 @@
 
 #include <linux/sched.h>
 
+/* ftrace syscalls requires exporting the sys_call_table */
+#ifdef CONFIG_FTRACE_SYSCALLS
+#ifdef CONFIG_PPC64
+extern const unsigned long long *sys_call_table;

I'm not sure why this is ULL ? UL and ULL are both 64 bits (on 64bit),
and it would save you this ifdef block and a cast in
arch_syscall_addr().

Good point - I was just following the format from
/arch/powerpc/kernel/systbl.S, which uses pairs of .llong on PPC64 and
single .long on PPC32. Does the assembler treat .llong different from
.long on 64bit?

quoted

+#else /* !CONFIG_PPC64 */
+extern const unsigned long *sys_call_table;
+#endif /* CONFIG_PPC64 */
+#endif /* CONFIG_FTRACE_SYSCALLS */
+
 static inline long syscall_get_nr(struct task_struct *task,
                   struct pt_regs *regs)
 {

diff --git a/arch/powerpc/include/asm/thread_info.h b/arch/powerpc/include/asm/thread_info.h
index aa9d383..e7a8af2 100644
--- a/arch/powerpc/include/asm/thread_info.h
+++ b/arch/powerpc/include/asm/thread_info.h

@@ -110,7 +110,8 @@ static inline struct thread_info *current_thread_info(void)
 #define TIF_NOERROR        12    /* Force successful syscall return */
 #define TIF_NOTIFY_RESUME    13    /* callback before returning to user */
 #define TIF_FREEZE        14    /* Freezing for suspend */
-#define TIF_RUNLATCH        15    /* Is the runlatch enabled? */
+#define TIF_SYSCALL_TRACEPOINT    15    /* syscall tracepoint instrumentation */
+#define TIF_RUNLATCH        16    /* Is the runlatch enabled? */

I don't grok why this is good or safe, not that it isn't but please tell
me why it is :)

Ok - now I could be wrong on this, but AFAICT the flags are only used
internally within the kernel by name and not exported to userspace (ie,
not part of the kernel ABI).  That specific flag is only set and cleared
within /arch/powerpc/kernel/process.c so recompiling the kernel should
be sufficient to change those instances of TIF_RUNLATCH from 15 to 16.

quoted

diff --git a/arch/powerpc/kernel/Makefile b/arch/powerpc/kernel/Makefile
index 8773263..9c404bb 100644
--- a/arch/powerpc/kernel/Makefile
+++ b/arch/powerpc/kernel/Makefile

@@ -98,6 +98,7 @@ obj64-$(CONFIG_AUDIT)        += compat_audit.o
 
 obj-$(CONFIG_DYNAMIC_FTRACE)    += ftrace.o
 obj-$(CONFIG_FUNCTION_GRAPH_TRACER)    += ftrace.o
+obj-$(CONFIG_FTRACE_SYSCALLS)    += ftrace.o

You're following the existing pattern there, but it's a little odd.
Seems like those three config options should really all depend on
something common and that should trigger the build of ftrace.c

There isn't anything in the arch specific ftrace.c that depends purely on
ftrace. It's all stuff that depends on the above specific ftrace options
that have some arch specific implementation.

quoted

diff --git a/arch/powerpc/kernel/ftrace.c b/arch/powerpc/kernel/ftrace.c
index ce1f3e4..b34171e 100644
--- a/arch/powerpc/kernel/ftrace.c
+++ b/arch/powerpc/kernel/ftrace.c

@@ -22,6 +22,7 @@
 #include <asm/cacheflush.h>
 #include <asm/code-patching.h>
 #include <asm/ftrace.h>
+#include <asm/syscall.h>
 

 #ifdef CONFIG_DYNAMIC_FTRACE

@@ -600,3 +601,15 @@ void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr)
     }
 }
 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
+
+#if defined(CONFIG_FTRACE_SYSCALLS) && defined(CONFIG_PPC64)

Does 32-bit just work using the existing routines? Or do we not support
it on 32-bit (though that's not what your Kconfig change said).

It should work with the existing routines - I still need to test this on
feugo to make sure, but following from /arch/powerpc/kernel/systbl.S it
seems that the symbol names should match the function names and the
system call table is a trivial lookup table, both of which will work
with the generic implementation.

quoted

+unsigned long __init arch_syscall_addr(int nr)
+{
+    return (unsigned long)sys_call_table[nr*2];

That's the cast I was referring to.

Ok, I'll change that in v2.

quoted

+}
+
+inline bool arch_syscall_match_sym_name(const char *sym, const char *name)
+{
+    return (!strcmp(sym + 4, name + 3));

So the +4 is to skip ".sys" and the +3 is to skip "sys" ? That could use
a comment IMHO :)

Yeah that's right, I'll add a comment there to clarify that.

cheers

Thanks for the feedback,
-Ian