[PATCH] allow ptrace and /proc/PID/mem to read fixmap pages

[PATCH] allow ptrace and /proc/PID/mem to read fixmap pages

[Roland McGrath]

It's a problem now that ptrace and /proc/PID/mem pretend the vsyscall page
does not exist, and won't let you read it from the inferior.  In fact you
can read it from yourself, but programs like a debugger should not have to
know about such magic addresses.  I think access_process_vm ought to give
you access to all of the process's adddress space.

This patch is i386-specific and should probably be done another way, but
it's what I am using now.  It works and is well-tested. 


Thanks,
Roland


--- stock-2.5.68/kernel/ptrace.c	Sat Apr 19 19:51:21 2003
+++ linux-2.5.68/kernel/ptrace.c	Wed Apr 23 12:39:52 2003
@@ -201,6 +201,25 @@ int access_process_vm(struct task_struct
 	up_read(&mm->mmap_sem);
 	mmput(mm);
 	
+#ifdef FIXADDR_START
+	while (len > 0 && addr < FIXADDR_TOP && addr >= FIXADDR_START) {
+		unsigned long pg = addr & PAGE_MASK;
+		int bytes = (len < pg + PAGE_SIZE - addr
+			     ? len : pg + PAGE_SIZE - addr);
+		pte_t *pte = pte_offset_kernel(pmd_offset(pgd_offset_k(pg),
+							  pg), pg);
+		if (!pte_read(*pte) || (write && !pte_write(*pte)))
+			break;
+		if (write)
+			memcpy((void *)addr, buf, bytes);
+		else
+			memcpy(buf, (const void *)addr, bytes);
+		len -= bytes;
+		buf += bytes;
+		addr += bytes;
+	}
+#endif
+
 	return buf - old_buf;
 }
 

Re: [PATCH] allow ptrace and /proc/PID/mem to read fixmap pages

[Jamie Lokier]

Roland McGrath wrote:
> This patch is i386-specific and should probably be done another way, but
> it's what I am using now.  It works and is well-tested. 

Would it be better to simply check whether _PAGE_USER is set in the pte?

-- Jamie

Re: [PATCH] allow ptrace and /proc/PID/mem to read fixmap pages

[Linus Torvalds]

On Sat, 26 Apr 2003, Roland McGrath wrote:
> 
> This patch is i386-specific and should probably be done another way, but
> it's what I am using now.  It works and is well-tested. 

It looks to me like it will cause NULL pointer dereferences if the user 
asks for something that is inside a pmd isn't mapped.

Maybe that won't happen for the FIXADDR range, but this is just not 
acceptable. We don't do code without error handling in the kernel.

		Linus