Re: [patch 2.6.20-rc4-git] remove modpost false warnings on ARM

[David Brownell]

On Friday 12 January 2007 8:38 am, Russell King wrote:
> A more correct test would be that found in kallsyms.c:

Good point.  Updated patch appended.

- Dave


===============	CUT HERE
This patch stops "modpost" from issuing erroneous modpost warnings on ARM
builds, which it's been doing simce since maybe last summer.  A canonical
example would be driver method table entries:

  WARNING: <path> - Section mismatch: reference to .exit.text:<name>_remove
	from .data after '$d' (at offset 0x4)

That "$d" symbol is generated by tools conformant with ARM ABI specs; in
this case, it's a relocation in the middle of a "<name>_driver" struct.
The erroneous warnings appear to be issued because "modpost" whitelists
references from "<name>_driver" data into init and exit sections ... but
does NOT whitelist them from "$d" (and can't).

This patch prevents the modpost symbol lookup code from ever returning
those symbols, so it will return a whitelisted symbol instead.

Now to revert various code-bloating "fixes" that got merged because of
this modpost bug....

Signed-off-by: David Brownell <dbrownell@xxxxxxxxxxxxxxxxxxxxx>
---
Likely this patch can be improved on, but there's another issue.
It seems to me that these modpost checks are wrong:

  * Lingering pointers that point into sections modprobe removes are
    *always* unsafe ... including probe() methods marked "__init"
    on hotpluggable busses.  Trivial fix:  use __devinit instead;
    or maybe platform_driver_probe().
  * Lingering pointers that point into sections that aren't removed
    are *never* unsafe ... including this remove() method case, since
    module unloading is configured and the __exit stuff must stay.

Whitelisting the former means not reporting potential oopsing cases;
dangerous.  Whereas even *checking* the latter is a waste of effort.

Index: at91/scripts/mod/modpost.c
===================================================================
--- at91.orig/scripts/mod/modpost.c	2006-12-15 10:08:57.000000000 -0800
+++ at91/scripts/mod/modpost.c	2007-01-12 12:09:29.000000000 -0800
@@ -655,6 +655,30 @@ static Elf_Sym *find_elf_symbol(struct e
 	return NULL;
 }
 
+static inline int is_arm_mapping_symbol(const char *str)
+{
+	return str[0] == '$' && strchr("atd", str[1])
+	       && (str[2] == '\0' || str[2] == '.');
+}
+
+/*
+ * If there's no name there, ignore it; likewise, ignore it if it's
+ * one of the magic symbols emitted used by current ARM tools.
+ *
+ * Otherwise if find_symbols_between() returns those symbols, they'll
+ * fail the whitelist tests and cause lots of false alarms ... fixable
+ * only by shrinking __exit and __init sections into __text, bloating
+ * the kernel (which is especially evil on embedded platforms).
+ */
+static inline int is_valid_name(struct elf_info *elf, Elf_Sym *sym)
+{
+	const char *name = elf->strtab + sym->st_name;
+
+	if (!name || !strlen(name))
+		return 0;
+	return !is_arm_mapping_symbol(name);
+}
+
 /*
  * Find symbols before or equal addr and after addr - in the section sec.
  * If we find two symbols with equal offset prefer one with a valid name.
@@ -683,16 +707,15 @@ static void find_symbols_between(struct 
 		symsec = secstrings + elf->sechdrs[sym->st_shndx].sh_name;
 		if (strcmp(symsec, sec) != 0)
 			continue;
+		if (!is_valid_name(elf, sym))
+			continue;
 		if (sym->st_value <= addr) {
 			if ((addr - sym->st_value) < beforediff) {
 				beforediff = addr - sym->st_value;
 				*before = sym;
 			}
 			else if ((addr - sym->st_value) == beforediff) {
-				/* equal offset, valid name? */
-				const char *name = elf->strtab + sym->st_name;
-				if (name && strlen(name))
-					*before = sym;
+				*before = sym;
 			}
 		}
 		else
@@ -702,10 +725,7 @@ static void find_symbols_between(struct 
 				*after = sym;
 			}
 			else if ((sym->st_value - addr) == afterdiff) {
-				/* equal offset, valid name? */
-				const char *name = elf->strtab + sym->st_name;
-				if (name && strlen(name))
-					*after = sym;
+				*after = sym;
 			}
 		}
 	}
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/