Skip to contentSkip to conversation

compact call trace

PATCH In review Akinobu Mita 6 participants 14 messages Jan 17, 2006

lore
Share & tools
lore
#0 lore raw

cover compact call trace

These patches will:

- break the various custom oops-parsers which people have written themselves.
- use common call trace format on x86-64.
- change offset format from hexadecimal to decimal in print_symbol()
- delete symbolsize in call trace in print_symbol().
- print system_utsname.version in oops so that we can doing a double check that the oops is matching the vmlinux we're looking at.

Example output:

o Currently we get the following call trace
i386: [<f0ad4c51>] kjournald+0x18c/0x207 [jbd]
x86-64: <ffffffffa008ef6c>{:jbd:kjournald+1030}

o After applied these patches.
i386: [<f0ad4c51>] kjournald+396 [jbd]
x86-64: <ffffffffa008ef6c> kjournald+1030 [jbd]

#1 lore raw

PATCH 1/4 makes print_symbol() return int

Use print_symbol() to dump call trace on x86-64.

----

 traps.c |   29 +++++++----------------------
 1 files changed, 7 insertions(+), 22 deletions(-)
arch/x86_64/kernel/traps.c renamed from arch/x86_64/kernel/traps.c.orig +7 -22
@@ -30,6 +30,7 @@
 #include <linux/moduleparam.h>
 #include <linux/nmi.h>
 #include <linux/kprobes.h>
+#include <linux/kallsyms.h> 
 
 #include <asm/system.h>
 #include <asm/uaccess.h>
@@ -92,30 +93,14 @@ static inline void conditional_sti(struc
 
 static int kstack_depth_to_print = 10;
 
-#ifdef CONFIG_KALLSYMS
-#include <linux/kallsyms.h> 
-int printk_address(unsigned long address)
-{ 
-	unsigned long offset = 0, symsize;
-	const char *symname;
-	char *modname;
-	char *delim = ":"; 
-	char namebuf[128];
-
-	symname = kallsyms_lookup(address, &symsize, &offset, &modname, namebuf); 
-	if (!symname) 
-		return printk("[<%016lx>]", address);
-	if (!modname) 
-		modname = delim = ""; 		
-        return printk("<%016lx>{%s%s%s%s%+ld}",
-		      address,delim,modname,delim,symname,offset); 
-} 
-#else
 int printk_address(unsigned long address)
 { 
-	return printk("[<%016lx>]", address);
-} 
-#endif
+	int len;
+
+	len = printk("<%016lx>", address);
+	len += print_symbol(" %s", address);
+	return len;
+}
 
 static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
 					unsigned *usedp, const char **idp)
#2 lore raw

PATCH 2/4 x86-64: Use print_symbol() to dump call trace

In reply to Akinobu Mita

Use print_symbol() to dump call trace on x86-64.

----

 traps.c |   29 +++++++----------------------
 1 files changed, 7 insertions(+), 22 deletions(-)
arch/x86_64/kernel/traps.c renamed from arch/x86_64/kernel/traps.c.orig +7 -22
@@ -30,6 +30,7 @@
 #include <linux/moduleparam.h>
 #include <linux/nmi.h>
 #include <linux/kprobes.h>
+#include <linux/kallsyms.h> 
 
 #include <asm/system.h>
 #include <asm/uaccess.h>
@@ -92,30 +93,14 @@ static inline void conditional_sti(struc
 
 static int kstack_depth_to_print = 10;
 
-#ifdef CONFIG_KALLSYMS
-#include <linux/kallsyms.h> 
-int printk_address(unsigned long address)
-{ 
-	unsigned long offset = 0, symsize;
-	const char *symname;
-	char *modname;
-	char *delim = ":"; 
-	char namebuf[128];
-
-	symname = kallsyms_lookup(address, &symsize, &offset, &modname, namebuf); 
-	if (!symname) 
-		return printk("[<%016lx>]", address);
-	if (!modname) 
-		modname = delim = ""; 		
-        return printk("<%016lx>{%s%s%s%s%+ld}",
-		      address,delim,modname,delim,symname,offset); 
-} 
-#else
 int printk_address(unsigned long address)
 { 
-	return printk("[<%016lx>]", address);
-} 
-#endif
+	int len;
+
+	len = printk("<%016lx>", address);
+	len += print_symbol(" %s", address);
+	return len;
+}
 
 static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
 					unsigned *usedp, const char **idp)
#3 lore raw

PATCH 3/4 compact print_symbol() output

In reply to Akinobu Mita

- remove symbolsize field
- change offset format from hexadecimal to decimal

99.9% of the functions in my vmlinux are smaller than 10000 bytes. Therefore call trace must be more compact.

----

 kallsyms.c |    7 +++----
 1 files changed, 3 insertions(+), 4 deletions(-)
kernel/kallsyms.c renamed from kernel/kallsyms.c.orig +3 -4
@@ -237,7 +237,7 @@ int __print_symbol(const char *fmt, unsi
 	const char *name;
 	unsigned long offset, size;
 	char namebuf[KSYM_NAME_LEN+1];
-	char buffer[sizeof("%s+%#lx/%#lx [%s]") + KSYM_NAME_LEN +
+	char buffer[sizeof("%s+%ld [%s]") + KSYM_NAME_LEN +
 		    2*(BITS_PER_LONG*3/10) + MODULE_NAME_LEN + 1];
 
 	name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
@@ -246,10 +246,9 @@ int __print_symbol(const char *fmt, unsi
 		sprintf(buffer, "0x%lx", address);
 	else {
 		if (modname)
-			sprintf(buffer, "%s+%#lx/%#lx [%s]", name, offset,
-				size, modname);
+			sprintf(buffer, "%s+%ld [%s]", name, offset, modname);
 		else
-			sprintf(buffer, "%s+%#lx/%#lx", name, offset, size);
+			sprintf(buffer, "%s+%ld", name, offset);
 	}
 	return printk(fmt, buffer);
 }
#4 lore raw

Re: [PATCH 3/4] compact print_symbol() output

Quoted 2 lines · Akinobu Mita (on Tue, 17 Jan 2006 19:15:55 +0900) wrote:

- remove symbolsize field
- change offset format from hexadecimal to decimal

That is silly. Almost every binutils tool prints offsets in hex, including objdump and gdb. Printing the trace offset in decimal just makes more work for users to convert back to decimal to match up with all the other tools.

#5 lore raw

Re: [PATCH 3/4] compact print_symbol() output

Quoted 9 lines · Keith Owens
Quoted 2 lines · Akinobu Mita (on Tue, 17 Jan 2006 19:15:55 +0900) wrote:

- remove symbolsize field
- change offset format from hexadecimal to decimal

That is silly. Almost every binutils tool prints offsets in hex, including objdump and gdb. Printing the trace offset in decimal just makes more work for users to convert back to decimal to match up with all the other tools.

Agreed.
Also, hex output is shorter and often more natural for this type of data.

--
Jesper Juhl <[email protected]>
Don't top-post http://www.catb.org/~esr/jargon/html/T/top-post.html
Plain text mails only, please http://www.expita.com/nomime.html

#6 lore raw

Re: [PATCH 3/4] compact print_symbol() output

Quoted 5 lines · Jesper Juhl
Quoted 4 lines · Keith Owens
Quoted 2 lines · Akinobu Mita (on Tue, 17 Jan 2006 19:15:55 +0900) wrote:

- remove symbolsize field
- change offset format from hexadecimal to decimal

Quoted 8 lines
Quoted 5 lines

That is silly. Almost every binutils tool prints offsets in hex, including objdump and gdb. Printing the trace offset in decimal just makes more work for users to convert back to decimal to match up with all the other tools.

Agreed.
Also, hex output is shorter and often more natural for this type of data.

In my vmlinux, 99.9% of the functions are smaller than 10000 bytes.

10000 == 0x2710
So. strlen("10000") == 5 < strlen("0x2710") == 6
Therefore call trace must be more compact.

#7 lore raw

Re: [PATCH 3/4] compact print_symbol() output

Quoted 15 lines · Akinobu Mita (on Tue, 17 Jan 2006 19:58:27 +0900) wrote:
Quoted 5 lines · Jesper Juhl
Quoted 4 lines · Keith Owens
Quoted 2 lines · Akinobu Mita (on Tue, 17 Jan 2006 19:15:55 +0900) wrote:

- remove symbolsize field
- change offset format from hexadecimal to decimal

Quoted 8 lines
Quoted 5 lines

That is silly. Almost every binutils tool prints offsets in hex, including objdump and gdb. Printing the trace offset in decimal just makes more work for users to convert back to decimal to match up with all the other tools.

Agreed.
Also, hex output is shorter and often more natural for this type of data.

Quoted 5 lines

In my vmlinux, 99.9% of the functions are smaller than 10000 bytes.

10000 == 0x2710
So. strlen("10000") == 5 < strlen("0x2710") == 6
Therefore call trace must be more compact.

Which is irrelevant when your change makes more work for everybody who reads the trace.

#8 lore raw

Re: [PATCH 3/4] compact print_symbol() output

In reply to Keith Owens
Quoted 8 lines · Keith Owens
Quoted 2 lines · Akinobu Mita (on Tue, 17 Jan 2006 19:15:55 +0900) wrote:

- remove symbolsize field
- change offset format from hexadecimal to decimal

That is silly. Almost every binutils tool prints offsets in hex, including objdump and gdb. Printing the trace offset in decimal just makes more work for users to convert back to decimal to match up with all the other tools.

Currently call trace on x86-64 prints offset in decimal. Do you think it is better to print offset in hex on x86-64 too? But Andi Kleen said he likes compact call trace in the reply to my first version of these patches.

And do you have any objection to remove symbolsize output in print_symbol()? I can't find useful usage about symbolsize in print_symbol() except to do a double check that the oops is matching the vmlinux we're looking at. (so I made 4/4) Do you know any useful usage about symbolsize?

#9 lore raw

Re: [PATCH 3/4] compact print_symbol() output

Quoted 6 lines · Akinobu Mita

And do you have any objection to remove symbolsize output in print_symbol()? I can't find useful usage about symbolsize in print_symbol() except to do a double check that the oops is matching the vmlinux we're looking at. (so I made 4/4) Do you know any useful usage about symbolsize?

I've often found symbolsize useful. Not when looking at an oops from my own machine. But when looking at an oops posted on LKML, from someone who most likely has a different .config and different compiler, different optimization and different inlining from mine. symbolsize is a good clue as to how close their kernel is to the one I've got built on my machine, how likely guesses I make based on mine will apply to theirs, and whereabouts in the function that it oopsed.

Hugh

#10 lore raw

Re: [PATCH 3/4] compact print_symbol() output

Quoted 8 lines · Hugh Dickins

I've often found symbolsize useful. Not when looking at an oops from my own machine. But when looking at an oops posted on LKML, from someone who most likely has a different .config and different compiler, different optimization and different inlining from mine. symbolsize is a good clue as to how close their kernel is to the one I've got built on my machine, how likely guesses I make based on mine will apply to theirs, and whereabouts in the function that it oopsed.

Yes that is why I want it too.

-Andi

#11 lore raw

PATCH 4/4 i386: print system_utsname.version in oops

In reply to Akinobu Mita

print system_utsname.version in i386 oops to make it possible doing a double check that the oops is matching the vmlinux we're looking at.

for example:
(2.6.15-git12) --> (2.6.15-git12 #1 SMP Tue Jan 17 13:59:15 JST 2006)

----

 traps.c |    5 +++--
 1 files changed, 3 insertions(+), 2 deletions(-)
arch/i386/kernel/traps.c renamed from arch/i386/kernel/traps.c.orig +3 -2
@@ -239,9 +239,10 @@ void show_registers(struct pt_regs *regs
 	}
 	print_modules();
 	printk(KERN_EMERG "CPU:    %d\nEIP:    %04x:[<%08lx>]    %s VLI\n"
-			"EFLAGS: %08lx   (%s) \n",
+			"EFLAGS: %08lx   (%s %s) \n",
 		smp_processor_id(), 0xffff & regs->xcs, regs->eip,
-		print_tainted(), regs->eflags, system_utsname.release);
+		print_tainted(), regs->eflags, system_utsname.release,
+		system_utsname.version);
 	print_symbol(KERN_EMERG "EIP is at %s\n", regs->eip);
 	printk(KERN_EMERG "eax: %08lx   ebx: %08lx   ecx: %08lx   edx: %08lx\n",
 		regs->eax, regs->ebx, regs->ecx, regs->edx);
#12 lore raw
In reply to Akinobu Mita
Quoted 8 lines · Akinobu Mita

These patches will:

- break the various custom oops-parsers which people have written themselves.
- use common call trace format on x86-64.
- change offset format from hexadecimal to decimal in print_symbol()
- delete symbolsize in call trace in print_symbol().
- print system_utsname.version in oops so that we can doing a double check that the oops is matching the vmlinux we're looking at.

at least then make the kallsyms lookup mark up the string somehow (say by putting a * in front of it) if the EIP is outside the size of the symbol; so that we can spot garbage better.

#13 lore raw
Quoted 11 lines · Arjan van de Ven (on Tue, 17 Jan 2006 11:31:27 +0100) wrote:
Quoted 8 lines · Akinobu Mita

These patches will:

- break the various custom oops-parsers which people have written themselves.
- use common call trace format on x86-64.
- change offset format from hexadecimal to decimal in print_symbol()
- delete symbolsize in call trace in print_symbol().
- print system_utsname.version in oops so that we can doing a double check that the oops is matching the vmlinux we're looking at.

Quoted 3 lines

at least then make the kallsyms lookup mark up the string somehow (say by putting a * in front of it) if the EIP is outside the size of the symbol; so that we can spot garbage better.

There is no such thing as "outside the size of the symbol". kallsyms does not save symbol size, it is calculated as the difference between this symbol and the next one in the table. By definition, every size is correct - as long as all symbols are in the table.

Some symbols are omitted from kallsyms, which makes the calculation of "symbol size" wrong for some symbols, those symbols appear bigger than they really are. Printing the symbol size was done years ago as a hint to the reader, it told them what the kernel thought the symbol size was and gave them something to check against. That was when there were very few symbols in the kernel, so most size calculations were wrong. Nowadays it is probably irrelevant.