This patch modifies scsi_host_template->change_queue_depth so that
it takes an argument indicating why it is being called. This will be
used so that if a LLD needs to do some extra processing when
handling queue fulls or later ramp ups, it can do so.

This is a simple port of the drivers setting a change_queue_depth
callback. In the patch I just have these LLDs adjust the queue depth
if the user was requesting it.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>

[Vasu.Dev: v2
	Also converted pmcraid_change_queue_depth and then verified
all modules compile  using "make allmodconfig" for any new build
warnings on X86_64.

	Updated original description after combing two original
patches from Mike to make this patch git bisectable.]
Signed-off-by: NVasu Dev <vasu.dev@intel.com>
[jejb: fixed up 53c700]
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

This patch contains changes that allow iscsi_session_setup
to allocate private space for LLD's
Signed-off-by: NJayamohan Kallickal <jayamohank@serverengines.com>
Acked-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

bnx2i currently has a check for if a ep is properly bound, so if
iscsi_queuecommand/xmit_task is called while there is no ep
we will not queue IO.

be2iscsi sends IO from queuecommand/xmit_task like how bnx2i does
and needs a similar test. This patch has us just use the suspend_bit
test for this.

When ep_poll has succeeed iscsid will call conn_bind, the LLD will
then call iscsi_conn_bind which will clear the suspend bit.
When ep_disconnect is called (or if there is a conn error) we set
the suspend bit. For the ep_disconnect case I am adding a helper
in this patch that will take the session lock to make sure
iscsi_queuecommand/xmit_task is not running and it will set
the suspend bit.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJayamohan Kallickal <jayamohank@serverengines.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

beiscsi does not need the iscsi scsi cmd processing. It does not
even get this info on the completion path. This adds a function
to just update the sequencing numbers and complete a task.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJayamohan Kallickal <jayamohank@serverengines.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

If we had multiple tasks on the cmd or requeue  lists, and iscsi_tcp
returns a error, the write_space function can still run and queue
iscsi_data_xmit. If it was a legetimate problem and iscsi_conn_failure
was run but we raced and iscsi_data_xmit was run first it could miss
the suspend bit checks, and start trying to send data again and hit
another timeout. A similar problem is present when using cxgb3i.

This has libiscsi check the suspend bit before calling the xmit
task callout, so we at least do not try sending multiple tasks
(one could be sent).
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

If we sent multiple pdus as immediate the target could be
rejecting some and we have just been dropping the rejection
notification. This adds code to handle nop-out rejections,
so if a nop-out was sent as a ping and rejected we do not
mark the connection bad. Instead we just clean up the timers
since we have pdu making a rount trip we know the connection
is good.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

We increment session->cmdsn at the top of iscsi_prep_scsi_cmd_pdu, but
if the prep  ecb or prep bidi or init_task calls fails then we leave the
session->cmdsn incremented. This moves the cmdsn manipulation to the end
of the function when we know it has succeeded.

It also adds a session->cmdsn--; in queuecommand for if a driver like
bnx2i tries to send a a task from that context but it fails. We do not
have to do this in the xmit thread context because that code will retry
the same task if the initial call fails.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

The session lock can be held in the scsi eh thread or the completion
paths run from the net softirq. This disables bhs in iscsi_eh_abort when
taking the session lock.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

Allow the user to control the debug logs in libiscsi. We will now
have a module param for connection, session & error handling.

[Mike Christie - Fixed up to compile on current code and added
missing ISCSI_DBG_EH conversions]
Signed-off-by: NErez Zilber <erezzi.list@gmail.com>
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

If we are sending or receiving data for the task successfully do
not run the scsi eh, because we know the task is making progress.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

This patch just adds some debug statements for the abort
and completion paths.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

If a task did not complete normally due to a TMF, libiscsi will
now complete the task with the state ISCSI_TASK_ABRT_TMF. Drivers
like bnx2i that need to free resources if a command did not complete normally
can then check the task state. If a driver does not need to send
a special command if we have dropped the session then they can check
for ISCSI_TASK_ABRT_SESS_RECOV.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

Instead of having libiscsi check if the offload bit is set, have
it check if the lld created a work queue. I think this is more
clear.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

If the session is failed, but we have not yet fully transitioned
to the recovery stage we were still queueuing IO. The idea is
that for some failures we can recvover at the command level
and still continue to execute other IO. Well, we never have
added the recovery within a command code, so queueing up IO here
just creates the possibility that it might time time out so
this just has us requeue the IO the scsi layer for now.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

bnx2i needs to send a hardware specific cleanup command if
a command has not completed normally (iscsi/scsi response from
target), and the session is still ok (this is the case when we
send a TMF to stop the command).

At this time it will need to drop the session lock. The problem
with the current code is that fail_all_commands assumes we
will hold the lock the entire time, so it uses list_for_each_entry_safe.
If while bnx2i drops the session lock multiple cmds complete then
list_for_each_entry_safe will not handle this correctly.

This patch removes the running lists and just has us loop over
the cmds array (in later patches we will then replace that
array with a block tag map at the session level). It also fixes
up the completion path so that if the TMF code and the normal recv
path were completing the same command then they both do not try
to do release the refcount taken when the task is queued.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

If we have not got any pdus for recv_timeout seconds, then we will
send a iscsi ping/nop to make sure the target is still around. The
problem is if this is a slow link, and the ping got queued after
the data for a data_out (read), then the transport code could think
the ping has failed when it is just slowly making its way through
the network. This patch has us check if we are making progress while
the nop is outstanding. If we are still reading in data, then we
do not fail the session at that time.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

If we are responding to a nop from the target by sending our nop,
and the session is getting torn down, then iscsi_start_session_recovery
could set the conn stop bits while the recv path is sending the nop
response and we will hit the bug ons in __iscsi_conn_send_pdu.

This has us check the state in __iscsi_conn_send_pdu and fail all
incoming mgmt IO if we are  not logged in and if the pdu is not login
related. It also changes the ordering of the setting of conn stop state
bits so they are set after the session state is set (both are set under
the session lock).
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

This has iscsi_data_in_rsp call iscsi_update_cmdsn when a pdu is
completed like is done for other pdu's that are don.

For libiscsi_tcp, this means that it calls iscsi_update_cmdsn when
it is handling the pdu internally to only transfer data, but if there is
status then it does not need to call it since the completion handling
will do it.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

bnx2i needs to be able to look up mgmt task like login and nop, because
it does some processing of them on the completion path. This exports
iscsi_itt_to_task so it can look up the task.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

If we could not allocate the initiator name or some other id like
the hwaddress or netdev, then userspace could deal with the failure
by just running in a dregraded mode.

Now we want to be able to switch values for the params and we
want some feedback, so this patch will check if a string like
the initiatorname could not be allocated and return an error.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

bnx2i does not have one. It currently preallocates the bdt
when the session is setup.

We probably want to change that to a dma pool, then allocate from
the pool in the alloc pdu. Until then check if there is a alloc
pdu callout.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

Set target can queue limit to the number of preallocated
session tasks we have.

This along with the cxgb3i can_queue patch will fix a throughput
problem where it could only queue one LU worth of data at a time.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

Le lundi 30 mars 2009, Chris Wright a écrit :
> q->queue could be ERR_PTR(-ENOMEM) which will break unwinding
> on error.  Make iscsi_pool_free more defensive.
>

Making the freeing of q->queue dependent on q->pool being set looks
really weird (although it is correct at the moment. But this seems
to be fixable in a much simpler way.

With the benefit that only the error case is slowed down. In both
cases we have a problem if q->queue contains an error value but it's
not -ENOMEM. Apparently this can't happen today, but it doesn't feel
right to assume this will always be true. Maybe it's the right time
to fix this as well.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

If the iscsi eh fires when the current task is a nop, then
the task->sc pointer is null. fail_all_commands could
then try to do task->sc->device and oops. We actually do
not need to access the curr task in this path, because
if it is a cmd task the fail_command call will handle
this and if it is mgmt task then the flush of the mgmt
queues will handle that.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

The api for conn and session failures is akward because
one takes a conn from the lib and one takes a session
from the class. This syncs up the interfaces to use
structs from the lib.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

The qdepth setting was useful when we needed libiscsi to verify
the setting. Now we just need to make sure if older tools
passed in zero then we need to set some default.

So this patch just has us use the sht->cmd_per_lun or if
for LLD does a host per session then we can set it on per
host basis.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

We were using the shost work queue which ended up being
a little akward since all iscsi hosts need a thread for
scanning, but only drivers hooked into libiscsi need
a workqueue for transmitting. So this patch moves the
xmit workqueue to the lib.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

There is no need to cap the queue depth in the modules. We set
this in userspace and can do that there. For performance testing
with ram based targets, this is helpful since we can have very
high queue depths.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

This makes the logging a compile time option and replaces
the scsi_debug macro with session and connection ones
that print out a driver model id prefix.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

Memory freeing in iscsi_pool_free() looks wrong to me. Either q->pool
can be NULL and this should be tested before dereferencing it, or it
can't be NULL and it shouldn't be tested at all. As far as I can see,
the only case where q->pool is NULL is on early error in
iscsi_pool_init(). One possible way to fix the bug is thus to not
call iscsi_pool_free() in this case (nothing needs to be freed anyway)
and then we can get rid of the q->pool check.
Signed-off-by: NJean Delvare <jdelvare@suse.de>
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

Yanling Qi from LSI found the root cause of the panic, below is his
analysis:

Problem description: the open iscsi driver installs eh_timed_out handler
to the
blank_transport_template of the scsi middle level that causes panic of
timed
out command of other host

Here are the details

Iscsi Session creation

During iscsi session creation time, the iscsi_tcp_session_create() of
iscsi_tpc.c will create a scsi-host for the session. See the statement
marked
with the label A. The statement B replaces the shost->transportt point
with a
local struct variable.

static struct iscsi_cls_session *
iscsi_tcp_session_create(struct iscsi_endpoint *ep, uint16_t cmds_max,
                         uint16_t qdepth, uint32_t initial_cmdsn,
                         uint32_t *hostno)
{
        struct iscsi_cls_session *cls_session;
        struct iscsi_session *session;
        struct Scsi_Host *shost;
        int cmd_i;
        if (ep) {
                printk(KERN_ERR "iscsi_tcp: invalid ep %p.\n", ep);
                return NULL;
        }

A        shost = iscsi_host_alloc(&iscsi_sht, 0, qdepth);

        if (!shost)

                return NULL;

B         shost->transportt = iscsi_tcp_scsi_transport;

        shost->max_lun = iscsi_max_lun;

Please note the scsi host is allocated by invoking isccsi_host_alloc()
in
libiscsi.c

Polluting the middle level blank_transport_template in
iscsi_host_alloc() of
libiscsi.c

The iscsi_host_alloc() invokes the middle level function
scsi_host_alloc() in
hosts.c for allocating a scsi_host. Then the statement marked with C
assigns
the iscsi_eh_cmd_timed_out handler to the eh_timed_out callback
function.

struct Scsi_Host *iscsi_host_alloc(struct scsi_host_template *sht,

                                   int dd_data_size, uint16_t qdepth)

{
        struct Scsi_Host *shost;
        struct iscsi_host *ihost;
        shost = scsi_host_alloc(sht, sizeof(struct iscsi_host) +
dd_data_size);
        if (!shost)
                return NULL;

 C      shost->transportt->eh_timed_out = iscsi_eh_cmd_timed_out;

Please note the shost->transport is the middle level
blank_transport_template
as shown in the code segment below. We see two problems here. 1.
iscsi_eh_cmd_timed_out is installed to the blank_transport_template that
will
cause some body else problem. 2. iscsi_eh_cmd_timed_out will never be
invoked
when iscsi command gets timeout because the statement B resets the
pointer.

Middle level blank_transport_template

In the middle level function scsi_host_alloc() of hosts.c, the middle
level
assigns a blank_transport_template for those hosts not implementing its
transport layer. All HBAs without supporting a specific scsi_transport
will
share the middle level blank_transport_template. Please see the
statement D

struct Scsi_Host *scsi_host_alloc(struct scsi_host_template *sht, int
privsize)

{
        struct Scsi_Host *shost;
        gfp_t gfp_mask = GFP_KERNEL;
        int rval;
        if (sht->unchecked_isa_dma && privsize)
                gfp_mask |= __GFP_DMA;

         shost = kzalloc(sizeof(struct Scsi_Host) + privsize, gfp_mask);
        if (!shost)
                return NULL;

        shost->host_lock = &shost->default_lock;

        spin_lock_init(shost->host_lock);

        shost->shost_state = SHOST_CREATED;

        INIT_LIST_HEAD(&shost->__devices);

        INIT_LIST_HEAD(&shost->__targets);

        INIT_LIST_HEAD(&shost->eh_cmd_q);

        INIT_LIST_HEAD(&shost->starved_list);

        init_waitqueue_head(&shost->host_wait);

        mutex_init(&shost->scan_mutex);

        shost->host_no = scsi_host_next_hn++; /* XXX(hch): still racy */

        shost->dma_channel = 0xff;

        /* These three are default values which can be overridden */

        shost->max_channel = 0;

        shost->max_id = 8;

        shost->max_lun = 8;

        /* Give each shost a default transportt */

 D       shost->transportt = &blank_transport_template;

Why we see panic at iscsi_eh_cmd_timed_out()

The mpp virtual HBA doesn’t have a specific scsi_transport. Therefore,
the
blank_transport_template will be assigned to the virtual host of the MPP
virtual HBA by SCSI middle level. Please note that the statement C has
assigned
iscsi-transport eh_timedout handler to the blank_transport_template.
When a mpp
virtual command gets timedout, the iscsi_eh_cmd_timed_out() will be
invoked to
handle mpp virtual command timeout from the middle level
scsi_times_out()
function of the scsi_error.c.

enum blk_eh_timer_return scsi_times_out(struct request *req)

{

        struct scsi_cmnd *scmd = req->special;

        enum blk_eh_timer_return (*eh_timed_out)(struct scsi_cmnd *);

        enum blk_eh_timer_return rtn = BLK_EH_NOT_HANDLED;

        scsi_log_completion(scmd, TIMEOUT_ERROR);

        if (scmd->device->host->transportt->eh_timed_out)

 E               eh_timed_out =
scmd->device->host->transportt->eh_timed_out;

        else if (scmd->device->host->hostt->eh_timed_out)

                eh_timed_out = scmd->device->host->hostt->eh_timed_out;

        else

                eh_timed_out = NULL;

        if (eh_timed_out) {

                rtn = eh_timed_out(scmd);

It is very easy to understand why we get panic in the
iscsi_eh_cmd_timed_out().
A scsi_cmnd from a no-iscsi device definitely can not resolve out a
session and
session->lock. The panic can be happed anywhere during the differencing.

static enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd
*scmd)

{

        struct iscsi_cls_session *cls_session;

        struct iscsi_session *session;

        struct iscsi_conn *conn;

        enum blk_eh_timer_return rc = BLK_EH_NOT_HANDLED;

        cls_session = starget_to_session(scsi_target(scmd->device));

        session = cls_session->dd_data;

        debug_scsi("scsi cmd %p timedout\n", scmd);

        spin_lock(&session->lock);

This patch fixes the problem by moving the setting of the
iscsi_eh_cmd_timed_out to iscsi_add_host, which is after the LLDs
have set their transport template to shost->transportt.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

I am not sure what happened. It looks like we have always leaked
the q->queue that is allocated from the kfifo_init call. nab finally
noticed that we were leaking and this patch fixes it by adding a
kfree call to iscsi_pool_free. kfifo_free is not used per kfifo_init's
instructions to use kfree.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

Mgmt setup used to not fail so we did not have to check
the return value. Now with cxgb3i it can so this has us
pass up a error.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

We do not need to allocate a itt for data_out, so this
passes the opcode to the alloc_pdu callout.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

bnx2i and cxgb3i need to encode LLD info in the itt so that
the firmware/hardware can process the pdu. This patch allows
the LLDs to encode info in the task->hdr->itt that they
setup in the alloc_pdu callout (any resources that are allocated
can be freed with the pdu in the cleanup_task callout). If
the LLD encodes info in the itt they should implement a
parse_pdu_itt callout. If parse_pdu_itt is not implemented
libiscsi will do the right thing for the LLD.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

This modifies the login buffer allocation to use __get_free_pages.
It will allow drivers that want to send this data with zero copy
operations to easily line things up on page boundaries.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

cxgb3i offloads data transfers. It does not offload the entire scsi/iscsi
procssing like qla4xxx and it does not offload the iscsi sequence
processing like how bnx2i does. cxgb3i relies on iscsi_tcp for the
seqeunce handling so this changes how we transfer unsolicitied data by
adding a common r2t struct and helpers.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

This regression was added in 2.6.27, when the mtask and ctask were
merged into the the common task struct. The patch applies to
scsi-rc-fixes, but also applies to 2.6.27 with some offsets.

The problem is that __iscsi_conn_send_pdu assumes that userspace was
not sending nops with the format it is checking for in the "if" below.
It turns out that older userspace tools are. This patch moves the
setting of the internal ping_task tracker (it tracks libiscsi current
outstanding nop) to iscsi_send_nopout which is only used by kernel callers.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

We must be using the bh spin locking functions in
iscsi_eh_device_reset becuase the session lock interacts with
a thread and softirq.

This patch also fixes up a bogus comment and check in fail_command,
because no one drops the lock (bnx2i did but it is not going
upstream yet and there were other refcount changes for that).
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

Some wires got crossed on some patches and I messed up in the code
below when rebuilding a patch. We want to be checking if flag
equaled the value indicating if we killing the session due to
final logout or if we just trying to relogin.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>