On 11/4/20 1:33 PM, Mike Christie wrote:
> On 9/18/20 4:09 PM, Lee Duncan wrote:
>> From: Lee Duncan <lduncan@xxxxxxxx>
>>
>> iSCSI NOPs are sometimes "lost", mistakenly sent to the
>> user-land iscsid daemon instead of handled in the kernel,
>> as they should be, resulting in a message from the daemon like:
>>
>>> iscsid: Got nop in, but kernel supports nop handling.
>>
>> This can occur because of the new forward- and back-locks,
>> and the fact that an iSCSI NOP response can occur before
>> processing of the NOP send is complete. This can result
>> in "conn->ping_task" being NULL in iscsi_nop_out_rsp(),
>> when the pointer is actually in the process of being set.
>>
>> To work around this, we add a new state to the "ping_task"
>> pointer. In addition to NULL (not assigned) and a pointer
>> (assigned), we add the state "being set", which is signaled
>> with an INVALID pointer (using "-1").
>>
>> Signed-off-by: Lee Duncan <lduncan@xxxxxxxx>
>> ---
>> drivers/scsi/libiscsi.c | 11 ++++++++++-
>> include/scsi/libiscsi.h | 3 +++
>> 2 files changed, 13 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/scsi/libiscsi.c b/drivers/scsi/libiscsi.c
>> index 1e9c3171fa9f..5eb064787ee2 100644
>> --- a/drivers/scsi/libiscsi.c
>> +++ b/drivers/scsi/libiscsi.c
>> @@ -738,6 +738,9 @@ __iscsi_conn_send_pdu(struct iscsi_conn *conn,
>> struct iscsi_hdr *hdr,
>> task->conn->session->age);
>> }
>> + if (unlikely(READ_ONCE(conn->ping_task) == INVALID_SCSI_TASK))
>> + WRITE_ONCE(conn->ping_task, task);
>> +
>> if (!ihost->workq) {
>> if (iscsi_prep_mgmt_task(conn, task))
>> goto free_task;
>> @@ -941,6 +944,11 @@ static int iscsi_send_nopout(struct iscsi_conn
>> *conn, struct iscsi_nopin *rhdr)
>> struct iscsi_nopout hdr;
>> struct iscsi_task *task;
>> + if (!rhdr) {
>> + if (READ_ONCE(conn->ping_task))
>> + return -EINVAL;
>> + WRITE_ONCE(conn->ping_task, INVALID_SCSI_TASK);
>> + }
>> if (!rhdr && conn->ping_task)
>> return -EINVAL;
>> @@ -957,11 +965,12 @@ static int iscsi_send_nopout(struct iscsi_conn
>> *conn, struct iscsi_nopin *rhdr)
>> task = __iscsi_conn_send_pdu(conn, (struct iscsi_hdr *)&hdr,
>> NULL, 0);
>> if (!task) {
>> + if (!rhdr)
>> + WRITE_ONCE(conn->ping_task, NULL);
>
> I don't think you need this. If __iscsi_conn_send_pdu returns NULL, it
> will have done __iscsi_put_task and done this already.
Not an issue, as you already replied.
>
>> iscsi_conn_printk(KERN_ERR, conn, "Could not send nopout\n");
>> return -EIO;
>> } else if (!rhdr) {
>> /* only track our nops */
>> - conn->ping_task = task;
>> conn->last_ping = jiffies;
>> }
>
> Why in the send path do we always use the READ_ONCE/WRITE_ONCE, but in
> the completion path like in iscsi_complete_task we don't.
>
The answer is that I was only modifying the code that needed changing
for this bug. My first pass did not use READ_ONCE() or WRITE_ONCE(), but
Hannes suggested the change.
Now that I think about it more, the memory barrier stuff would make
sense only if all the access to that field are protected.
I will resubmit V2 of the patch.
--
Lee Duncan
