Hi Marcel, On Tue, Jul 28, 2020 at 12:22 AM Marcel Holtmann <marcel@xxxxxxxxxxxx> wrote: > > Hi Luiz, > > > Bluetooth Core Specification v5.2, Vol. 3, Part A, section 1.4, table > > 1.1: > > > > 'Start Fragments always either begin with the first octet of the Basic > > L2CAP header of a PDU or they have a length of zero (see [Vol 2] Part > > B, Section 6.6.2).' > > > > This text has been changed recently as it previously stated: > > > > 'Start Fragments always begin with the Basic L2CAP header of a PDU.' > > > > Apparently this was changed by the following errata: > > > > https://www.bluetooth.org/tse/errata_view.cfm?errata_id=10216 > > > > In past this has not been a problem but it seems new controllers are > > apparently doing it as it has been reported in Zephyr: > > > > https://github.com/zephyrproject-rtos/zephyr/issues/26900 > > > > Signed-off-by: Luiz Augusto von Dentz <luiz.von.dentz@xxxxxxxxx> > > --- > > net/bluetooth/l2cap_core.c | 104 +++++++++++++++++++++++++++++-------- > > 1 file changed, 83 insertions(+), 21 deletions(-) > > > > diff --git a/net/bluetooth/l2cap_core.c b/net/bluetooth/l2cap_core.c > > index ade83e224567..193bea314222 100644 > > --- a/net/bluetooth/l2cap_core.c > > +++ b/net/bluetooth/l2cap_core.c > > @@ -8269,6 +8269,63 @@ static void l2cap_security_cfm(struct hci_conn *hcon, u8 status, u8 encrypt) > > mutex_unlock(&conn->chan_lock); > > } > > > > +/* Append fragment into frame respecting the maximum len of rx_skb */ > > +static int l2cap_recv_frag(struct l2cap_conn *conn, struct sk_buff *skb, > > + u16 len) > > +{ > > + if (!conn->rx_skb) { > > + /* Allocate skb for the complete frame (with header) */ > > + conn->rx_skb = bt_skb_alloc(len, GFP_KERNEL); > > + if (!conn->rx_skb) > > + return -ENOMEM; > > + /* Init rx_len */ > > + conn->rx_len = len; > > + } > > + > > + /* Copy as much as the rx_skb can hold */ > > + len = min_t(u16, len, skb->len); > > + skb_copy_from_linear_data(skb, skb_put(conn->rx_skb, len), len); > > + skb_pull(skb, len); > > + conn->rx_len -= len; > > + > > + return len; > > +} > > + > > +static int l2cap_recv_header(struct l2cap_conn *conn, struct sk_buff *skb) > > +{ > > + struct l2cap_hdr *hdr; > > + struct sk_buff *rx_skb; > > + int len; > > + > > + /* Append just enough to complete the header */ > > + len = l2cap_recv_frag(conn, skb, L2CAP_HDR_SIZE - conn->rx_skb->len); > > + > > + /* If header could not be read just continue */ > > + if (len < 0 || conn->rx_skb->len < L2CAP_HDR_SIZE) > > + return len; > > + > > + rx_skb = conn->rx_skb; > > + conn->rx_skb = NULL; > > + > > + hdr = (struct l2cap_hdr *) rx_skb->data; > > so I think it is pointless to insist on getting the complete header. We really just need the first 2 octets. > > struct l2cap_hdr { > __le16 len; > __le16 cid; > } __packed; Indeed, I've totally forgotten about the cid so I will change this to not use L2CAP_HDR_SIZE but 2 instead. > Once we have received at least 2 octets, we can get_unaligned_le16(rx_skb->data) and then just continue. Sure, I was trying to figure out if there is any way to grow the the rx_skb since I will be just allocating 2 bytes for it if the header is not available, we could perhaps take a different approach and always allocate based on the conn->mtu that way we don't have to wait the length to received to allocate a second skb and copy over the length into it, obviously that would only be done if length was fragmented. > Regards > > Marcel > -- Luiz Augusto von Dentz
