[PATCH][COMPAT] fix for net/compat.c

[Stephen Rothwell]

Hi Linus, Dave,

This is basically a patch from Randolph Chung who tells me that when
a syscall is done from the kernel, you cannot pass user mode pointers
to it on some architectures.  So we need to copy the sock_filter
array into kernel space before we pass it to the real system call.

His original patch (which does the same thing) has been tested on
parisc64-linux and this patch has been cross compiled for ppc64-linux.

Please apply.

-- 
Cheers,
Stephen Rothwell                    [e-mail]
http://www.canb.auug.org.au/~sfr/

diff -ruN 2.5.66-033114/net/compat.c 2.5.66-033114-netfix/net/compat.c
--- 2.5.66-033114/net/compat.c	2003-03-25 12:08:29.000000000 +1100
+++ 2.5.66-033114-netfix/net/compat.c	2003-03-31 16:08:02.000000000 +1000
@@ -496,6 +496,7 @@
 	struct sock_fprog kfprog;
 	mm_segment_t old_fs;
 	compat_uptr_t uptr;
+	unsigned int fsize;
 	int ret;
 
 	if (!access_ok(VERIFY_READ, fprog32, sizeof(*fprog32)) ||
@@ -503,15 +504,14 @@
 	    __get_user(uptr, &fprog32->filter))
 		return -EFAULT;
 
-	kfprog.filter = compat_ptr(uptr);
-	/*
-	 * Since struct sock_filter is architecure independent,
-	 * we can just do the access_ok check and pass the
-	 * same pointer to the real syscall.
-	 */
-	if (!access_ok(VERIFY_READ, kfprog.filter,
-			kfprog.len * sizeof(struct sock_filter)))
+	fsize = kfprog.len * sizeof(struct sock_filter);
+	kfprog.filter = (struct sock_filter *)kmalloc(fsize, GFP_KERNEL);
+	if (kfprog.filter == NULL)
+		return -ENOMEM;
+	if (copy_from_user(kfprog.filter, compat_ptr(uptr), fsize)) {
+		kfree(kfprog.filter);
 		return -EFAULT;
+	}
 
 	old_fs = get_fs();
 	set_fs(KERNEL_DS);
@@ -519,6 +519,7 @@
 			     (char *)&kfprog, sizeof(kfprog));
 	set_fs(old_fs);
 
+	kfree(kfprog.filter);
 	return ret;
 }