Fwd: [PATCH] 2.6.18-rt7: fix more issues with 32-bit cycles_t in latency_trace.c

Hello.

   Ugh, got the linuxppc-dev address wrong in the original mail...

WBR, Sergei

----------  Forwarded Message  ----------

Subject: [PATCH] 2.6.18-rt7: fix more issues with 32-bit cycles_t in latency_trace.c
Date: Monday 13 November 2006 00:23
From: Sergei Shtylyov <redacted>
To: mingo@elte.hu
Cc: linux-kernel@vger.kernel.org, linuxppc-dev@linux-mips.org, dwalker@mvista.com, khilman@mvista.com

In addition to clock wrap check being falsely triggered with 32-bit cycles_t,
as noticed to Kevin Hilman, there's another issue: using %Lx format to print
32-bit values warrants erroneous values on 32-bit machines like ARM and
 PPC32.

Signed-off-by: Sergei Shtylyov <redacted>

---
PPC32 actually has 64-bit timebase counter, so could provide for 64-bit
cycles_t -- maybe it's worth to rewrite get_cycles() to read both lower and
upper registers?

Index: linux-2.6/kernel/latency_trace.c
===================================================================

--- linux-2.6.orig/kernel/latency_trace.c
+++ linux-2.6/kernel/latency_trace.c

@@ -1623,8 +1623,8 @@ check_critical_timing(int cpu, struct cp
 #ifndef CONFIG_CRITICAL_LATENCY_HIST
 	if (!preempt_thresh && preempt_max_latency > delta) {
 		printk("bug: updating %016Lx > %016Lx?\n",
-			preempt_max_latency, delta);
-		printk("  [%016Lx %016Lx %016Lx]\n", T0, T1, T2);
+			(u64)preempt_max_latency, (u64)delta);
+		printk("  [%016Lx %016Lx %016Lx]\n", (u64)T0, (u64)T1, (u64)T2);
 	}
 #endif

-------------------------------------------------------

Hello, I wrote:

In addition to clock wrap check being falsely triggered with 32-bit cycles_t,
as noticed to Kevin Hilman, there's another issue: using %Lx format to print
32-bit values warrants erroneous values on 32-bit machines like ARM and PPC32.

Signed-off-by: Sergei Shtylyov <redacted>

---
PPC32 actually has 64-bit timebase counter, so could provide for 64-bit
cycles_t -- maybe it's worth to rewrite get_cycles() to read both lower and
upper registers?

   This one (called latency-tracer-prink-fix.patch in the broken out patchset) has been erroneously "restored" some months ago. Since preempt_max_latency is of type 'unsigned long' now and all the other vars are of type 'cycle_t' (which aliases 'u64'), only the first cast is actually needed (or rather, the first format specifier needs to be changed to %08lx). Erm, maybe I'll cook a patch...

WBR, Sergei