tcp: tcp_probe buffer overflow and incorrect return value

[ Upstream commit: 8d390efd903485923419584275fd0c2aa4c94183 ]

tcp_probe has a bounds-checking bug that causes many programs (less,
python) to crash reading /proc/net/tcp_probe. When it outputs a log
line to the reader, it only checks if that line alone will fit in the
reader's buffer, rather than that line and all the previous lines it
has already written.

tcpprobe_read also returns the wrong value if copy_to_user fails--it
just passes on the return value of copy_to_user (number of bytes not
copied), which makes a failure look like a success.

This patch fixes the buffer overflow and sets the return value to
-EFAULT if copy_to_user fails.

Patch is against latest net-2.6; tested briefly and seems to fix the
crashes in less and python.

Signed-off-by: Tom Quetchenbach <virtualphtn@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>

diff --git a/net/ipv4/tcp_probe.c b/net/ipv4/tcp_probe.c
index 87dd5bff315fcede15cb1bd88f072a48d4df4ca9..a79a547464a5a21303f75483fef1d1ec63abefa5 100644 (file)
--- a/net/ipv4/tcp_probe.c
+++ b/net/ipv4/tcp_probe.c
@@ -190,19 +190,18 @@ static ssize_t tcpprobe_read(struct file *file, char __user *buf,
 
                width = tcpprobe_sprint(tbuf, sizeof(tbuf));
 
-               if (width < len)
+               if (cnt + width < len)
                        tcp_probe.tail = (tcp_probe.tail + 1) % bufsize;
 
                spin_unlock_bh(&tcp_probe.lock);
 
                /* if record greater than space available
                   return partial buffer (so far) */
-               if (width >= len)
+               if (cnt + width >= len)
                        break;
 
-               error = copy_to_user(buf + cnt, tbuf, width);
-               if (error)
-                       break;
+               if (copy_to_user(buf + cnt, tbuf, width))
+                       return -EFAULT;
                cnt += width;
        }