There is a race condition in iscsi_tcp.c that may cause it to forget
that it received a R2T from the target. This race may cause a data-out
command (such as a write) to lock up. The race occurs here:
static int
iscsi_send_unsol_pdu(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
{
struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
int rc;
if (tcp_ctask->xmstate & XMSTATE_UNS_HDR) {
BUG_ON(!ctask->unsol_count);
tcp_ctask->xmstate &= ~XMSTATE_UNS_HDR; <---- RACE
...
static int
iscsi_r2t_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
{
...
tcp_ctask->xmstate |= XMSTATE_SOL_HDR_INIT; <---- RACE
...
While iscsi_xmitworker() (called from scsi_queue_work()) is preparing to
send unsolicited data, iscsi_tcp_data_recv() (called from
tcp_read_sock()) interrupts it upon receipt of a R2T from the target.
Both contexts do read-modify-write of tcp_ctask->xmstate. Usually, gcc
on x86 will make &= and |= atomic on UP (not guaranteed of course), but
in this case iscsi_send_unsol_pdu() reads the value of xmstate before
clearing the bit, which causes gcc to read xmstate into a CPU register,
test it, clear the bit, and then store it back to memory. If the recv
interrupt happens during this sequence, then the XMSTATE_SOL_HDR_INIT
bit set by the recv interrupt will be lost, and the R2T will be
forgotten.
The patch below (against 2.6.24-rc1) converts accesses of xmstate to use
set_bit, clear_bit, and test_bit instead of |= and &=. I have tested
this patch and verified that it fixes the problem. Another possible
approach would be to hold a lock during most of the rx/tx setup and
post-processing, and drop the lock only for the actual rx/tx.
Signed-off-by: Tony Battersby <tonyb@cybernetics.com>
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
505f76b306