2.4.25 sync unwind code with 2.6.3

The unwind code in 2.4.25 has drifted away from the 2.6.3 version, including some missing bug fixes. This patch bring 2.4.25 unwind as close to 2.6.3 as possible. Changes are :-

* Move unw_table_entry from one header to another.
* Add Fenghua Yu's change to pt_regs_off. This is not really in 2.4 yet, even though the comment says it is.
* Comment and white space changes.
* Bug fix to alloc_reg_state().
* Replace struct_offset() with offsetof().
* Export unw_* symbols.
* Avoid deadlock on script->lock.

I am not happy with the kludge that avoids the deadlock on script->lock, anybody got a better solution? The kludge assumes that the first field in an ia64 rwsem is a spinlock. I have hit this unwind deadlock a couple of times during processing of asynchronous MCA events, so I want it fixed.

After this patch, the main differences between 2.4 and 2.6 are the handing of the gate page and its DSO data.

Index: 25.4/include/asm-ia64/unwind.h

@@ -93,6 +93,12 @@ struct unw_frame_info {
  * The official API follows below:
  */
 
+struct unw_table_entry {
+	u64 start_offset;
+	u64 end_offset;
+	u64 info_offset;
+};
+
 /*
  * Initialize unwind support.
  */

Index: 25.4/arch/ia64/kernel/unwind_i.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2000, 2002 Hewlett-Packard Co
+ * Copyright (C) 2000, 2002-2003 Hewlett-Packard Co
  *	David Mosberger-Tang <[email protected]>
  *
  * Kernel unwind support.

@@ -45,12 +45,6 @@ struct unw_info_block {
 	/* personality routine and language-specific data follow behind descriptors */
 };
 
-struct unw_table_entry {
-	u64 start_offset;
-	u64 end_offset;
-	u64 info_offset;
-};
-
 struct unw_table {
 	struct unw_table *next;		/* must be first member! */
 	const char *name;

@@ -148,7 +142,7 @@ struct unw_insn {
 };
 
 /*
- * Preserved general static registers (r2-r5) give rise to two script
+ * Preserved general static registers (r4-r7) give rise to two script
  * instructions; everything else yields at most one instruction; at
  * the end of the script, the psp gets popped, accounting for one more
  * instruction.

Index: 25.4/arch/ia64/kernel/unwind.c

Index: 25.4/arch/ia64/kernel/Makefile

@@ -11,7 +11,7 @@ all: kernel.o head.o init_task.o
 
 O_TARGET := kernel.o
 
-export-objs := ia64_ksyms.o
+export-objs := ia64_ksyms.o unwind.o
 
 obj-y := acpi.o entry.o gate.o efi.o efi_stub.o ia64_ksyms.o irq.o irq_ia64.o irq_lsapic.o ivt.o \
 	 machvec.o pal.o process.o perfmon.o ptrace.o sal.o salinfo.o semaphore.o setup.o	 \

Keith> I am not happy with the kludge that avoids the deadlock on Keith> script-> lock, anybody got a better solution? The kludge Keith> assumes script-> that the first field in an ia64 rwsem is a Keith> spinlock. I have hit this unwind deadlock a couple of times Keith> during processing of asynchronous MCA events, so I want it Keith> fixed.

What I did for libunwind is to avoid deadlock by falling back to uncached unwinding if necessary. That is, if a lock is taken, it just creates an unwind script on the local stack.

It would be good to replace the built-in kernel unwinder with something based on libunwind. That would make it much easier to keep the code in sync (and libunwind is much better tested). I have been toying with the idea of doing this for 2.7, but it hasn't been very high priority to me because, for the most part, the kernel unwinder is good enough.

	--david

It was a typo. The code is needed to handle f2-f5, hence the "- 2".

	--david

Good idea, I will do a patch once you push your bk tree.

2.6 is correct, it should be (rval - 2) in both trees.

Ah, and I bet that's the one you mentioned on 2/23 with this link:
http://lia64.bkbits.net:8080/to-linus-2.5/[email protected]
(which no longer points the right place -- rats, bitten again by the fact that BK cset numbers change after the fact :-()

Thanks for straightening me out. I'll apply the following patch to 2.4 so it will be the same as 2.6 in this regard:

=== arch/ia64/kernel/unwind.c 1.30 vs edited ==--- 1.30/arch/ia64/kernel/unwind.c Wed Mar 3 16:51:27 2004
+++ edited/arch/ia64/kernel/unwind.c Wed Mar 3 17:33:27 2004
@@ -1420,7 +1420,7 @@

case UNW_WHERE_FR:
if (rval <= 5)

-			val = unw.preg_index[UNW_REG_F2  + (rval -  1)];
+			val = unw.preg_index[UNW_REG_F2  + (rval -  2)];

else if (rval >= 16 && rval <= 31)
val = unw.preg_index[UNW_REG_F16 + (rval - 16)];
else {

Bjorn> Ah, and I bet that's the one you mentioned on 2/23 with this Bjorn> link: http://lia64.bkbits.net:8080/to-linus-2.5/[email protected]
Bjorn> (which no longer points the right place -- rats, bitten again Bjorn> by the fact that BK cset numbers change after the fact :-()

Argh. Maybe we can use the stable cset numbers, but they're awfully long.

	--david