Allow FCP frames to bypass the FCoE receive processing threads and handle
them directly in softirq context, if they are received on the correct CPU.
This preserves the queuing to threads for scaling out receive processing
to multiple CPUs, but allows FCoE-aware multi-queue network drivers that
direct frames to the originating CPUs to handle FCP processing with less
scheduling latency.

Only FCP is handled directly, because libfc makes use of mutexes in ELS
handling routines.

The bulk of this change is just moving the FCoE receive processing out of
the receive thread function, leaving behind just the thread and queue
management.  The interesting bits are in fcoe_rcv()
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

The FC-LS spec. says ELS timeouts should be 2 x R_A_TOV.
The FC-GS spec. says CT timeouts should be 3 x R_A_TOV.

We've been using E_D_TOV for both of those.

Change for all ELS and CT requests except FLOGI, which we
leave at 2 seconds (using E_D_TOV).  One could argue that
R_A_TOV is locally determined until after FLOGI succeeds.

This does change FLOGI for vports which becomes FDISC.
This does not change the REC/SRR timeout which is 2 seconds.
Signed-off-by: NJoe Eykholt <jeykholt@cisco.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

There are cases outside of our control that may result in a transmit
skb being linearized in dev_queue_xmit.  There are a couple of bugs
in libfc/fcoe that can result in a panic at that point.  This patch
contains two fixes to prevent those panics.

1) use fast cloning instead of shared skbs with dev_queue_xmit

dev_queue_xmit doen't want shared skbuffs being passed in, and
__skb_linearize will BUG if the skb is shared.  FCoE is holding an extra
reference around the call to dev_queue_xmit, so that when it returns an
error code indicating the frame has been dropped it can maintain it's
own backlog and retransmit.  Switch to using fast skb cloning for this
instead.

2) don't append compound pages as > PAGE_SIZE skb fragments

fc_fcp_send_data will append pages from a scatterlist to the nr_frags[]
if the netdev supports it.  But, it's using > PAGE_SIZE compound pages
as a single skb_frag.  In the highmem linearize case that page will be
passed to kmap_atomic to get a mapping to copy out of, but
kmap_atomic will only allow access to the first PAGE_SIZE part.
The memcpy will keep going and cause a page fault once is crosses the
first boundary.

If fc_fcp_send_data uses linear buffers from the start, it calls
kmap_atomic one PAGE_SIZE at a time.  That same logic needs to be
applied when setting up skb_frags.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

If the underlying netdev is a VLAN device, make sure the VLAN ID is integrated
into the WWNN/WWPN name generation. Also added/updated the comments to reflect
this change.
Signed-off-by: NYi Zou <yi.zou@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

We are still using netdev->dev_addr to generate lport's WWNN/WWPN even if the
LLD has support for NETDEV_HW_ADDR_T_SAN. Instead, we should just use the
fip->ctl_src_addr, which is the NETDEV_HW_ADDR_T_SAN if LLD supports it or it
is just the netdev->dev_addr if it does not.
Signed-off-by: NYi Zou <yi.zou@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Make sure we are get the SAN MAC address from the real netdev if the input
netdev is a VLAN device.
Signed-off-by: NYi Zou <yi.zou@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

This was fixed before in 7a7f0c7f but it's introduced again recently.
Signed-off-by: NYi Zou <yi.zou@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

There was a locking problem where the fip->lock was held during
the call to update_mac().  The rtnl_lock() must be taken before
the fip->lock, not the other way around.  This fixes that.

Now that fcoe_ctlr_recv_flog() is called only from the response handler
to a FLOGI request, some checking can be eliminated.  Instead of calling
update_mac(), just fill in the granted_mac address for the passed-in
frame (skb).

Eliminate the passed-in source MAC address since it is also in the skb.

Also, in fcoe, call fcoe_set_src_mac() directly instead of going thru
the fip function pointer.  This will generate less code.
Then, since fip isn't needed for LOGO response, use lport as the arg.
Signed-off-by: NJoe Eykholt <jeykholt@cisco.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

This patch adds a check to fail gracefully when the netdevice
is bonded.  Previously, the error was detected but the stack
would continue to load.  This resulted in a partially enabled
fcoe intance and errors when the fcoe instance was destroy.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Remove the two extra function decalartions in fcoe.c.
Signed-off-by: NYi Zou <yi.zou@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Added kernel-doc comment blocks to all structures and functions.

Renamed fc_lport instances rom lp to lport to be inline with our
naming convention.

Renamed all misnamed net_device instances to netdev to be inline
with our naming convention.
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

This is the Open-FCoE implementation of the FC
passthrough support via bsg interface.

Passthrough support is added to both N_Ports and
VN_Ports.
Signed-off-by: NSteve Ma <steve.ma@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Allow a vport specific string to be appended to the port symbolic
name.  The new symbolic name is sent to the name server after it
is set.

This currently messes with libhbalinux, which is looking for
the fcoe "fcoe <ver> over <ethX>" string and expects whatever
comes after the "over" to be a network interface name only.

Adds an EXPORT_SYMBOL to libfc for fc_frame_alloc_fill, which is
needed to allow fcoe to allocate a frame of variable length for
the RSPN request.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Register the fc_host symbolic name as the symbolic node name
with the fabric name server.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Add NPIV vport create and destroy handlers and register them with the
FC transport.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Right now it's exactly the same as the physical port template,
and there is no way to create a port on anything other than the
netdev.  When the vport_create entry point gets hooked up it will
create lports on top of vport devices, which will use this.

Rename scsi_transport_fcoe_sw to fcoe_transport_template to be more
clear with naming now that there are two templates.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

The FIP code in libfcoe needed several changes to support NPIV

1) dst_src_addr needs to be managed per-n_port-ID for FPMA fabrics with NPIV
   enabled.  Managing the MAC address is now handled in fcoe, with some slight
   changes to update_mac() and a new get_src_addr() function pointer.

2) The libfc elsct_send() hook is used to setup FCoE specific response
   handlers for FIP encapsulated ELS exchanges.  This lets the FCoE specific
   handling know which VN_Port the exchange is for, and doesn't require
   tracking OX_IDs.  It might be possible to roll back to the full FIP frame
   in these, but for now I've just stashed the contents of the MAC address
   descriptor in the skb context block for later use.  Also, because
   fcoe_elsct_send() just passes control on to fc_elsct_send(), all transmits
   still come through the normal frame_send() path.

3) The NPIV changes added a mutex hold in the keep alive sending, the lport
   mutex is protecting the vport list.  We can't take a mutex from a timer,
   so move the FIP keep alive logic to the link work struct.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

I'd like to keep basic initialization together with allocation, which means
this can't just be a tail-call to scsi_host_alloc.

This is needed to create a generic libfc host allocation routine for NPIV
VN_Ports, which will share the exchange ID space (through sharing exchange
manager structures) with the parent lport.  In order to clone the exchange
manager list when the lport is allocated, the list head must be initialized
earlier.

Also, update fnic to use the libfc_host_alloc so that later changes do not break
it. (contribution by Joe Eykholt)
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NJoe Eykholt <jeykholt@cisco.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

The cmd_per_lun value is used by scsi-ml as fall back lowest
queue_depth value but in case of libfc cmd_per_lun is set to
same value as max queue_depth = 32.

So this patch reduces cmd_per_lun value to 3 and configures
each lun with default max queue_depth 32 in fc_slave_alloc.
Signed-off-by: NVasu Dev <vasu.dev@intel.com>
Acked-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Calls ndo_fcoe_enabled() of the associated netdev upon creating the FCoE
instance to make sure LLD has all necessary resources allocated and setup
properly before passing FCoE traffic. Similarly, calls ndo_fcoe_disable()
upon destroying the FCoE instance on the associated netdev to allow the LLD
to release all allocated resources for FCoE.
Signed-off-by: NYi Zou <yi.zou@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Add a define of FCOE_MTU as 2158 bytes and use FCOE_MTU when the LLD is found
to support NETIF_F_FCOE_MTU. The lport->mfs is then calculated out of the
2158 FCOE_MTU. Otherwise, we stick with the netdev->mtu, i.e., LAN MTU. Also,
change the notification on NETDEV_CHANGEMTU event to bypass changing mfs when
LAN MTU is changed if NETIF_F_FCOE_MTU is supported.
Signed-off-by: NYi Zou <yi.zou@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

When doing echo ethX > /sys..../destroy I am getting
errors when the tear down succeeds. It looks like the
reason for this is because the rc var is not getting set
when the destruction works. This just sets it to zero.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Remove the redundant checking of netdev->netdev_ops as it will never be NULL.
Signed-off-by: NYi Zou <yi.zou@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

This fixes one cause of an occational problem when unloading
libfc where the exchange manager pool doesn't have all items freed.

The existing WARN_ON(mp->total_exches <= 0) isn't hit.
However, note that total_exches is decremented when the
exchange is completed, and it can be held with a refcnt
for a while after that.

I'm not sure what the offending exchange is, but I suspect
it is an incoming request, because outgoing state machines
should be all stopped at this point.

Note that although receive is stopped before the exchange
manager is freed, there could still be active threads
handling received frames.

This patch flushes the queues by allocating a new skb
and sending it through, and have the thread handle
this new skb specially.  This is similar to the way the work
queues are flushed now by putting work items in them and waiting
until they make it through the queue.

An skb->destructor function is used to inform us of
the completion of the flush, and the fr_dev() is left
NULL to indicate to fcoe_percpu_receive_thread() that
the skb should be just freed.  There's already a check
for the lp being NULL which prints a message.
We skip printing the message if the destructor is for flushing.
Signed-off-by: NJoe Eykholt <jeykholt@cisco.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

This just cuts down on the number of locks we're dealing with, and
eliminates the need to take another lock in the netdev notifier.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Fixes reference counting on fcoe_instance and net_device, and adds
NETDEV_UNREGISTER notifier handling so that you can unload network drivers.
FCoE no longer increments the module use count for the network driver.

On an NETDEV_UNREGISTER event, destroying the FCoE instance is deferred to a
workqueue context to avoid RTNL deadlocks.

Based in part by an earlier patch from John Fastabend

John's patch description:
Currently, the netdev module ref count is not decremented with module_put()
when the module is unloaded while fcoe instances are present. To fix this
removed reference count on netdev module completely and added functionality to
netdev event handling for NETDEV_UNREGISTER events.

This allows fcoe to remove devices cleanly when the netdev module is unloaded
so we no longer need to hold a reference count for the netdev module.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

We only want the FCoE create and destroy routines to deal with top level
N_Ports, the VN_Ports are tracked on the vport list (see scsi_transport_fc).
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Rather than rely on the hostlist_lock to be held while creating exchange
managers, serialize fcoe instance creation and destruction with a mutex.
This will allow the hostlist addition to be moved out of fcoe_if_create(),
which will simplify NPIV support.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

fcoe_netdev_config() is called during initialization of a libfc instance.
Much of what was there only needs to be done once for each net_device.
The same goes for the corresponding cleanup.

The FIP controller initialization is moved to interface creation time.
Otherwise it will keep getting re-initialized for every VN_Port once NPIV is
enabled.

fcoe_if_destroy() has some reordering to deal with the changes.  Receives are
not stopped until after fcoe_interface_put() is called, but transmits must be
stopped before.  So there is some care to stop libfc transmits and the
transmit backlog timer, then call fcoe_interface_put which will stop receives
and cleanup the FIP controller, then the receive queues can be cleaned and the
port freed.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

Up to this point the fcoe_instance structure was simply kzalloc/kfreed.  This
patch introduces create and destroy functions as well as kref based reference
counting.  The create function will grow as the initialization code is moved
there.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

The priv pointer is no longer needed, and once NPIV is enabled
fcoe_interface:fc_lport becomes a one-to-many relationship.

Remove the single pointer.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

The offload EM pointer is only used when setting up a new libfc instance, but
as it's designed to be shared among NPIV VN_Ports it should be tracked in
fcoe_interface.

With the host-list changed to track fcoe_interfaces as well, this is needed
before we can remove the priv pointer from that structure (which is only there
to help in the transition, and stops making sense once NPIV is enabled).
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

There is only one FIP state per net_device, so the FIP controller needs to be
moved from the per-SCSI-host fcoe_port to the per-net_device fcoe_interface
structure.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

The packet handlers need to be tracked in fcoe_interface so there is only one
set per net_device.  When NPIV is enabled there will be multiple SCSI hosts
and multiple fcoe_port structures on a single net_device.

The packet handlers match by ethertype and netdev.  If the same handler gets
registered on a single netdev multiple times, the receive function will be
called multiple times for each frame.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

The network interface needs to be shared between all NPIV VN_Ports, therefor
it should be tracked in the fcoe_interface and not for each SCSI host in
fcoe_port.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

In preparation for NPIV support, I'm splitting the fcoe instance structure
into two to remove the assumptions about it being 1:1 with the net_device.
There will now be two structures, one which is 1:1 with the underlying
net_device and one which is allocated per virtual SCSI/FC host.

fcoe_softc is renamed to fcoe_port for the per Scsi_Host FCoE private data.

Later patches with start moving shared stuff from fcoe_port to fcoe_interface
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

By passing in the parent device instead of assuming the netdev is what
should be used, fcoe_if_create becomes usable for NPIV vports as well.
You still need a netdev, because that's how FCoE works.  Also removed some
duplicate checks from fcoe_if_create that are already in fcoe_create.

fcoe_if_destroy needs to take an lport as it's only argument, not a netdev.
That removes the 1:1 netdev:lport assumption from the destroy path.
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

The hostlist and the hostlist_lock were initialized both in
the delcaration and in fcoe_init().  Remove the unneeded code.
Signed-off-by: NJoe Eykholt <jeykholt@cisco.com>
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

fcoe_if_init() can fail, but it's return value wasn't checked
Signed-off-by: NChris Leech <christopher.leech@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>

This adds fcoe_ddp_min as a module parameter for fcoe module to:
/sys/module/fcoe/parameters/ddp_min

It is observed that for some hardware, particularly Intel 82599, there is too
much overhead in setting up context for direct data placement (DDP) read when
the requested read I/O size is small. This is added as a module parameter for
performance tuning and is set as 0 by default and user can change this based
on their own hardware.
Signed-off-by: NYi Zou <yi.zou@intel.com>
Signed-off-by: NRobert Love <robert.w.love@intel.com>
Signed-off-by: NJames Bottomley <James.Bottomley@suse.de>