Instead of waiting for TIMEDWAIT event when disconnecting and blocking,
start the disconnect process when TIMEDWAIT event is received
and only perform the actual rdma_disconnect upon disconnect request.
Note that rdma_disconnect can not be called from atomic context, so need
to execute it from workqueue.
Signed-off-by: Yan Burman <yanb@mellanox.com>
git-svn-id: http://svn.code.sf.net/p/scst/svn/branches/iser@5317 d57e44dd-8a1f-0410-8b47-8ef2f437770f
If rdma_accept fails, use the cleanup mechanism we already have
for disconnect, instead of trying to reproduce the same cleanup.
From upper layer it really does not matter if rdma_accept failed
or we received disconnect immediately after rdma_accept succeeded.
Signed-off-by: Yan Burman <yanb@mellanox.com>
git-svn-id: http://svn.code.sf.net/p/scst/svn/branches/iser@5262 d57e44dd-8a1f-0410-8b47-8ef2f437770f
Add iser character device handling for accepting and handling connections
received through RDMA transport.
Add isert_listener device to the poll() loop and handle incoming
connection requests.
Differentiate between iser and non iser connections
Validate RDMAExtension field and reject it if found in iscsi login request.
Also, disable immediate data and first burst for iSER since it is not supported yet
Signed-off-by: Yan Burman <yanb@mellanox.com>
git-svn-id: http://svn.code.sf.net/p/scst/svn/branches/iser@5235 d57e44dd-8a1f-0410-8b47-8ef2f437770f
Replace iscsi-tcp specific calls with transport API calls
in order to be able to override them with iser implementation specifics.
Make iscsi-tcp specific debug print into a general print
Refactor conn close code to work with isert
Only allocate RX data in NOP for iscsi-tcp
In case of iser, the data is already received
Signed-off-by: Yan Burman <yanb@mellanox.com>
git-svn-id: http://svn.code.sf.net/p/scst/svn/branches/iser@5230 d57e44dd-8a1f-0410-8b47-8ef2f437770f
May be not completed, because this kernel has too many internal issues
reported after enabling "Kernel hacking" config options
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5163 d57e44dd-8a1f-0410-8b47-8ef2f437770f
The following procfs-related changes have been included in kernel 3.10:
- The PDE() macro has been removed from <linux/proc_fs.h>.
- The create_proc_entry() function has been removed.
- The proc_info field has been removed from struct scsi_host_template.
Unbreak the procfs build for kernel versions >= 3.10
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5148 d57e44dd-8a1f-0410-8b47-8ef2f437770f
Avoid that sparse complains about a local symbol not having been declared
static (scst_pool_dep_map and scst_conn_dep_map). Also avoid that sparse
reports these two variables as unused in the procfs build.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5110 d57e44dd-8a1f-0410-8b47-8ef2f437770f
the timeout value. The problem is that if commands were started at
00:00:05 and 00:00:10 and the timeout value was 15 seconds and the
first command completed ok, then conn_rsp_timer_fn would have run
at 00:00:20 and rescheduled the next conn_rsp_timer_fn run for 15
secs from that time (00:00:35). But, we would have wanted to check on
the second command at 15 secs after its write_start time which is
00:00:25.
This patch has us reschedule the next conn_rsp_timer_fn run for
the req's write_start time + its timeout.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5093 d57e44dd-8a1f-0410-8b47-8ef2f437770f
1. If we have a nop in flight then we do not really need to send
more to test if the other side is still there. Either that nop
will timeout and conn will be dropped or it will complete
and last_rcv_time will be updated.
2. We are sending nop-ins every nop_in_interval seconds, but if
we got a data half way in to the window, then when conn_nop_in_delayed_work_fn
is run it was scheduling the next check to be nop_in_interval more
seconds. With the current settings this could end up with us waiting
59 secs before sending a nop and about 130 secs for it to timeout, when
at 61 it should be timing out and getting cleaned up already. This patch
has us take into account when the last time we got data to try and
reschedule the next nop check closer to nop_in_interval from that time.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5092 d57e44dd-8a1f-0410-8b47-8ef2f437770f
when a command/data is in the process of being sent. The snd
buffer space goes to zero and sendpage returns -EAGAIN. We then
wait for space to open. If space never opens then the cmd rsp timer
fires and the session/connection is dropped.
During that time we could try to send a nop as a ping, but it
is stuck behind that other command that got EAGAIN and we are
waiting for space to open to send them. We would expect the
nop in timer to fire before the cmd rsp timer, but it cannot
becuase the nop is sitting in the internal iscsi-scst queue.
This patch just has us start the timer ont he nop in right
away.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5091 d57e44dd-8a1f-0410-8b47-8ef2f437770f
This patch does not change any functionality other than the message
printed when it takes more than one second before the completion is
signalled.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5086 d57e44dd-8a1f-0410-8b47-8ef2f437770f
That patch contains the following changes:
* Move the code for creating a "close_sess" sysfs attribute from iscsi-scst to
the SCST core such that it becomes available for other target drivers. This
does not change the functionality of iscsi-scst.
* Add code in ib_srpt and qla2x00t to allow a session to be forcibly closed from
user space.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@4852 d57e44dd-8a1f-0410-8b47-8ef2f437770f
Yan Burman
Vladislav Bolkhovitin
Bart Van Assche