Skip to contentSkip to conversation

Add fptest for powerpc

PATCH In review Heikki Lindholm 2 participants 4 messages Jun 10, 2006 – Jun 11, 2006

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Add fptest for powerpc

This adds powerpc version of fptest.h. It just stores and loads the raw integer into/from FPU registers without converting to proper IEEE form. Converting would have added unneccessary complexity. As a side note, the ppc barfs on the test with "1000 != 10". :I

-- hl

include/asm-powerpc/fptest.h +84 -1
diff -Nru xenomai.orig/include/asm-powerpc/fptest.h xenomai/include/asm-powerpc/fptest.h
@@ -10,11 +10,94 @@
 
 static inline void fp_regs_set(unsigned val)
 {
+	uint64_t fpval = val;
+	__asm__ __volatile__("lfd	0, %0\n"
+		"	fmr	1, 0\n"
+		"	fmr	2, 0\n"
+		"	fmr	3, 0\n"
+		"	fmr	4, 0\n"
+		"	fmr	5, 0\n"
+		"	fmr	6, 0\n"
+		"	fmr	7, 0\n"
+		"	fmr	8, 0\n"
+		"	fmr	9, 0\n"
+		"	fmr	10, 0\n"
+		"	fmr	11, 0\n"
+		"	fmr	12, 0\n"
+		"	fmr	13, 0\n"
+		"	fmr	14, 0\n"
+		"	fmr	15, 0\n"
+		"	fmr	16, 0\n"
+		"	fmr	17, 0\n"
+		"	fmr	18, 0\n"
+		"	fmr	19, 0\n"
+		"	fmr	20, 0\n"
+		"	fmr	21, 0\n"
+		"	fmr	22, 0\n"
+		"	fmr	23, 0\n"
+		"	fmr	24, 0\n"
+		"	fmr	25, 0\n"
+		"	fmr	26, 0\n"
+		"	fmr	27, 0\n"
+		"	fmr	28, 0\n"
+		"	fmr	29, 0\n"
+		"	fmr	30, 0\n"
+		"	fmr	31, 0\n"
+		: 
+		: "m" (fpval));
 }
 
+#define FPTEST_REGVAL(n) {						\
+	uint64_t t;							\
+	__asm__ __volatile__("	stfd	" #n ", %0" : "=m" (t));	\
+	e[n] = (unsigned)t;						\
+	}
+
 static inline int fp_regs_check(unsigned val)
 {
-    return 0;
+	unsigned i, failed = 0;
+	uint32_t e[32];
+
+	FPTEST_REGVAL(0);
+	FPTEST_REGVAL(1);
+	FPTEST_REGVAL(2);
+	FPTEST_REGVAL(3);
+	FPTEST_REGVAL(4);
+	FPTEST_REGVAL(5);
+	FPTEST_REGVAL(6);
+	FPTEST_REGVAL(7);
+	FPTEST_REGVAL(8);
+	FPTEST_REGVAL(9);
+	FPTEST_REGVAL(10);
+	FPTEST_REGVAL(11);
+	FPTEST_REGVAL(12);
+	FPTEST_REGVAL(13);
+	FPTEST_REGVAL(14);
+	FPTEST_REGVAL(15);
+	FPTEST_REGVAL(16);
+	FPTEST_REGVAL(17);
+	FPTEST_REGVAL(18);
+	FPTEST_REGVAL(19);
+	FPTEST_REGVAL(20);
+	FPTEST_REGVAL(21);
+	FPTEST_REGVAL(22);
+	FPTEST_REGVAL(23);
+	FPTEST_REGVAL(24);
+	FPTEST_REGVAL(25);
+	FPTEST_REGVAL(26);
+	FPTEST_REGVAL(27);
+	FPTEST_REGVAL(28);
+	FPTEST_REGVAL(29);
+	FPTEST_REGVAL(30);
+	FPTEST_REGVAL(31);
+
+	for (i = 0; i < 32; i++)
+		if (e[i] != val) {
+			printk("r%d: %u != %u\n", i, e[i], val);
+			failed = 1;
+		}
+
+	return failed;
 }
 
 #endif /* FPTEST_H */
#1 lore raw
Quoted 4 lines · Heikki Lindholm

This adds powerpc version of fptest.h. It just stores and loads the raw integer into/from FPU registers without converting to proper IEEE form. Converting would have added unneccessary complexity. As a side note, the ppc barfs on the test with "1000 != 10". :I

Ok, this test tool is far from being user-friendly :-)

Just to be sure, I have tested the latest revision on x86 and it still runs fine, so I guess there really is a bug in power pc fpu switching code.

Now, if we decode a bit, the integer passed to fp_regs_set and fp_regs_check are 1000 * iterations + task_index, so 1000 is task 0, iteration 1
10 is task 10, iteration 0.
It means that task 10 after returning from its first context switch, is getting the context of task 0 before its second context switch.

Now the thing to try is to reduce the number of threads involved, since by default there are 23 threads, so you should add a printk for each context switch in the kernel driver (and now that I think about it this printk should be activable with an ioctl), and try and reduce the list of threads by removing threads which do not appear between the serie of switches from task 0 to task 10. The correspondance between the task numbers and their type is not easy, but /proc/xenomai/sched and switch --help should help...

Task 0 is the "idle" taks, an ordinary Linux task
Task 10 is of type "rtup_ufpp" here, which means that it is a real-time task in primary mode using the FPU in primary mode.

Signature · 1 line
					    Gilles Chanteperdrix.
#2 lore raw

Gilles Chanteperdrix kirjoitti:

Quoted 28 lines
Quoted 4 lines · Heikki Lindholm

This adds powerpc version of fptest.h. It just stores and loads the raw integer into/from FPU registers without converting to proper IEEE form. Converting would have added unneccessary complexity. As a side note, the ppc barfs on the test with "1000 != 10". :I

Ok, this test tool is far from being user-friendly :-)

Just to be sure, I have tested the latest revision on x86 and it still runs fine, so I guess there really is a bug in power pc fpu switching code.

Now, if we decode a bit, the integer passed to fp_regs_set and
fp_regs_check are 1000 * iterations + task_index, so
1000 is task 0, iteration 1
10 is task 10, iteration 0.
It means that task 10 after returning from its first context switch, is getting the context of task 0 before its second context switch.

Now the thing to try is to reduce the number of threads involved, since by default there are 23 threads, so you should add a printk for each context switch in the kernel driver (and now that I think about it this printk should be activable with an ioctl), and try and reduce the list of threads by removing threads which do not appear between the serie of switches from task 0 to task 10. The correspondance between the task numbers and their type is not easy, but /proc/xenomai/sched and switch --help should help...

Quoted 3 lines

Task 0 is the "idle" taks, an ordinary Linux task
Task 10 is of type "rtup_ufpp" here, which means that it is a real-time task in primary mode using the FPU in primary mode.

I probably don't have time for debugging for some days now, but if any of the code goes through sigreturn path, it might be a kernel bug that's fixed in 2.6.16.

-- hl

#3 lore raw
Quoted 7 lines · Heikki Lindholm
Quoted 3 lines

Task 0 is the "idle" taks, an ordinary Linux task
Task 10 is of type "rtup_ufpp" here, which means that it is a real-time task in primary mode using the FPU in primary mode.

I probably don't have time for debugging for some days now, but if any of the code goes through sigreturn path, it might be a kernel bug that's fixed in 2.6.16.

Anyway, the patch is applied, thanks.

Signature · 1 line
					    Gilles Chanteperdrix.