Re: [PATCH] asm-generic: uaccess: fix up local access_ok() usage

[Arnd Bergmann <arnd@arndb.de>]

On Saturday 13 June 2009, Mike Frysinger wrote:
> There's no reason that I can see to use the short __access_ok() form
> directly when the access_ok() is clearer in intent and for more people,
> expands to the same C code (i.e. always specify the first field -- access
> type).  Not all no-mmu systems lack memory protection, so the read/write
> could feasibly be checked.

Ah, I didn't consider this. I checked all the architectures and could not
find a case where access_ok actually evaluates the the first argument, so
I chose the slightly terser variant. I also don't let you override
access_ok() at this moment, which means that you don't have a choice
to use the generic uaccess.h and still differentiate between read and
write accesses.

What I really got wrong was the prototype for __access_ok(), as you
showed in your follow-up. I only tested this with the microblaze
patch that overrides __access_ok() with an architecture specific
version that gets this part right.

Would this simpler patch help you as well?

--- a/include/asm-generic/uaccess.h
+++ b/include/asm-generic/uaccess.h
@@ -37,14 +37,14 @@ static inline void set_fs(mm_segment_t fs)
 #define VERIFY_READ	0
 #define VERIFY_WRITE	1
 
-#define access_ok(type, addr, size) __access_ok((unsigned long)(addr),(size))
+#define access_ok(type, addr, size) __access_ok((addr), (size))
 
 /*
  * The architecture should really override this if possible, at least
  * doing a check on the get_fs()
  */
 #ifndef __access_ok
-static inline int __access_ok(unsigned long addr, unsigned long size)
+static inline int __access_ok(void __user *ptr, unsigned long size)
 {
 	return 1;
 }

It may not be clearer in intent, but it's what the majority (by a small
margin) of architecture do anyway.

> Also, the strnlen_user() function was missing a access_ok() check on the
> pointer given.  We've had cases on Blackfin systems where test cases
> caused kernel crashes here because userspace passed up a NULL/-1 pointer
> and the kernel gladly attempted to run strlen() on it.

Right, well spotted. I'll take this fix as a separate patch, ok?

	Arnd <><