While tb_dump_hop() prints out necessary information it is in format
that is quite hard to read from the logs especially when one needs to
follow the path to see that the setup is correct.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Now that the driver can handle every possible tunnel types there is no
point to log everything as info level so turn these to happen at debug
level instead.

While at it remove duplicated tunnel activation log message
(tb_tunnel_activate() calls tb_tunnel_restart() which print the same
message) and add one missing '\n' termination.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

The printing macros do not modify the passed object so make them
const. While there make tb_route() to take const parameter as well.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Two domains (hosts) can be connected through a Thunderbolt cable and in
that case they can start software services such as networking over the
high-speed DMA paths. Now that we have all the basic building blocks in
place to create DMA tunnels over the Thunderbolt fabric we can add this
support to the software connection manager as well.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

In order to detect possible connections to other domains we need to be
able to find out why tb_switch_alloc() fails so make it return ERR_PTR()
instead. This allows the caller to differentiate between errors such as
-ENOMEM which comes from the kernel and for instance -EIO which comes
from the hardware when trying to access the possible switch.

Convert all the current call sites to handle this properly.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

In addition to PCIe and Display Port tunnels it is also possible to
create tunnels that forward DMA traffic from the host interface adapter
(NHI) to a NULL port that is connected to another domain through a
Thunderbolt cable. These tunnels can be used to carry software messages
such as networking packets.

To support this we introduce another tunnel type (TB_TUNNEL_DMA) that
supports paths from NHI to NULL port and back.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Currently ICM has been handling XDomain UUID exchange so there was no
need to have it in the driver yet. However, since now we are going to
add the same capabilities to the software connection manager it needs to
be handled properly.

For this reason modify the driver XDomain protocol handling so that if
the remote domain UUID is not filled in the core will query it first and
only then start the normal property exchange flow.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

We run all XDomain requests during discovery in tb->wq and since it only
runs one work at the time it means that sending back reply to the other
domain may be delayed too much depending whether there is an active
XDomain discovery request running.

To make sure we can send reply to the other domain as soon as possible
run tb_xdp_handle_request() in system workqueue instead. Since the
device can be hot-removed in the middle we need to make sure the domain
structure is still around when the function is run so increase reference
count before we schedule the reply work.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Now that we have capability to discover existing tunnels during driver
load there is no point tearing down tunnels when the driver gets
unloaded. Instead we can just leave them running. If user disconnects
devices while there is no Thunderbolt driver loaded, tunneled protocol
hotplug happens and is handled by the corresponding driver (pciehp in
case of PCIe tunnel, GFX driver in case of DP tunnel).
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Display Port tunnels are somewhat more complex than PCIe tunnels as it
requires 3 tunnels (AUX Rx/Tx and Video). In addition we are not
supposed to create the tunnels immediately when a DP OUT is enumerated.
Instead we need to wait until we get hotplug event to that adapter port
or check if the port has HPD set before tunnels can be established. This
adds Display Port tunneling support to the software connection manager.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

NFC (non flow control) credits is actually 20-bit field so update
tb_port_add_nfc_credits() to handle this properly. This allows us to set
NFC credits for Display Port path in subsequent patches.

Also make sure the function does not update the hardware if the
underlying switch is already unplugged.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

We will be needing these routines to find Display Port adapters as well
so modify them to take port type as the second parameter.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

The only way to expand Thunderbolt topology is through the NULL adapter
ports (typically ports 1, 2, 3 and 4). There is no point handling
Thunderbolt hotplug events on any other port.

Add a helper function (tb_port_is_null()) that can be used to determine
if the port is NULL port, and use it in software connection manager code
when hotplug event is handled.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Currently the software connection manager (tb.c) has only supported
creating a single PCIe tunnel, no PCIe device daisy chaining has been
supported so far. This updates the software connection manager so that
it now can create PCIe tunnels for full chain of six devices.

Because PCIe allows DMA and opens possibility for DMA attacks we change
security level to "user" meaning that PCIe tunneling requires that the
userspace authorizes the devices first. This makes it possible to block
PCIe tunneling completely while still allowing other types of tunnels to
be automatically created.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

In Apple Macs the boot firmware (EFI) connects all devices automatically
when the system is started, before it hands over to the OS. Instead of
ignoring we discover all those PCIe tunnels and record them using our
internal structures, just like we do when a device is connected after
the OS is already up.

By doing this we can properly tear down tunnels when devices are
disconnected. Also this allows us to resume the existing tunnels after
system suspend/resume cycle.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

State of the connected devices and tunnel configuration is not known
during resume. For example some paths may not be complete anymore if the
user has unplugged the related devices. So instead of marking all paths
as inactive we go ahead and deactivate them explicitly before we restart
them.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Now that we can allocate hop IDs per port on a path, we can take
advantage of this and create tunnels covering longer paths than just
between two adjacent switches. PCIe actually does not need this as it
is typically a daisy chain between two adjacent switches but this way we
do not need to hard-code creation of the tunnel.

While there add name to struct tb_path to make debugging easier, and
update kernel-doc comments.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

We need to be able to walk from one port to another when we are creating
paths where there are multiple switches between two ports. For this
reason introduce a new function tb_next_port_on_path().
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
Reviewed-by: Lukas Wunner <lukas@wunner.de>

Currently the driver only assigns remote port for the primary port if in
case of dual link. This makes things such as walking from one port to
another more complex than necessary because the code needs to change
from secondary to primary port if the path that is established is
created using secondary links.

In order to always assign both remote pointers we need to prevent the
scanning code from following the secondary link. Failing to do that
might cause problems as the same switch may be enumerated twice (or
removed in case of unplug). Handle that properly by introducing a new
function tb_port_has_remote() that returns true only for the primary
port. We also update tb_is_upstream_port() to support both dual link
ports, make it take const port pointer and move it below
tb_upstream_port() to keep similar functions close.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Each port has a separate path configuration space that is used for
finding the next hop (switch) in the path. HopID is an index to this
configuration space. HopIDs 0 - 7 are reserved by the protocol.

In order to get next available HopID for each direction we provide two
pairs of helper functions that can be used to allocate and release
HopIDs for a given port.

While there remove obsolete TODO comment.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

To be able to tunnel non-PCIe traffic, separate tunnel functionality
into generic and PCIe specific parts. Rename struct tb_pci_tunnel to
tb_tunnel, and make it hold an array of paths instead of just two.
Update all the tunneling functions to take this structure as parameter.

We also move tb_pci_port_active() to switch.c (and rename it) where we
will be keeping all port and switch related functions.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

In order to tunnel non-PCIe traffic as well rename tunnel_pci.[ch] to
tunnel.[ch] to reflect this fact. No functional changes.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

The adapter specific capability either is there or not if the port does
not hold an adapter. Instead of always finding it on-demand we read the
offset just once when the port is initialized.

While there we update the struct port documentation to follow kernel-doc
format.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

We need to wait until all buffers have been drained before the path can
be considered disabled. Do this for every hop in a path.

This adds another bit field to struct tb_regs_hop even if we are trying
to get rid of them but we can clean them up another day.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Thunderbolt 2 devices and beyond link controller needs to be notified
when a switch is going to be suspended by setting bit 31 in LC_SX_CTRL
register. Add this functionality to the software connection manager.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Thunderbolt 2 devices and beyond need to have additional bits set in
link controller specific registers. This includes two bits in LC_SX_CTRL
that tell the link controller which lane is connected and whether it is
upstream facing or not.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

We will be adding more link controller functionality in subsequent
patches and it does not make sense to keep all that in switch.c, so
separate LC functionality into its own file.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Light Ridge has an issue where reading the next capability pointer
location in port config space the read data is not cleared. It is fine
to read capabilities each after another so only thing we need to do is
to make sure we issue dummy read after tb_port_find_cap() is finished to
avoid the issue in next read.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Light Ridge and Eagle Ridge both need to have TMU access enabled before
port space can be fully accessed so make sure it happens on those. This
allows us to get rid of the offset quirk in tb_port_find_cap().
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Maximum depth in Thunderbolt topology is 6 so make sure it is not
possible to allocate switches that exceed the depth limit.

While at it update tb_switch_alloc() to use upper/lower_32_bits()
following tb_switch_alloc_safe_mode().
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

switch_lock was introduced because it allowed serialization of device
authorization requests from userspace without need to take the big
domain lock (tb->lock). This was fine because device authorization with
ICM is just one command that is sent to the firmware. Now that we start
to handle all tunneling in the driver switch_lock is not enough because
we need to walk over the topology to establish paths.

For this reason drop switch_lock from the driver completely in favour of
big domain lock.

There is one complication, though. If userspace is waiting for the lock
in tb_switch_set_authorized(), it keeps the device_del() from removing
the sysfs attribute because it waits for active users to release the
attribute first which leads into following splat:

    INFO: task kworker/u8:3:73 blocked for more than 61 seconds.
          Tainted: G        W         5.1.0-rc1+ #244
    "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
    kworker/u8:3    D12976    73      2 0x80000000
    Workqueue: thunderbolt0 tb_handle_hotplug [thunderbolt]
    Call Trace:
     ? __schedule+0x2e5/0x740
     ? _raw_spin_lock_irqsave+0x12/0x40
     ? prepare_to_wait_event+0xc5/0x160
     schedule+0x2d/0x80
     __kernfs_remove.part.17+0x183/0x1f0
     ? finish_wait+0x80/0x80
     kernfs_remove_by_name_ns+0x4a/0x90
     remove_files.isra.1+0x2b/0x60
     sysfs_remove_group+0x38/0x80
     sysfs_remove_groups+0x24/0x40
     device_remove_attrs+0x3d/0x70
     device_del+0x14c/0x360
     device_unregister+0x15/0x50
     tb_switch_remove+0x9e/0x1d0 [thunderbolt]
     tb_handle_hotplug+0x119/0x5a0 [thunderbolt]
     ? process_one_work+0x1b7/0x420
     process_one_work+0x1b7/0x420
     worker_thread+0x37/0x380
     ? _raw_spin_unlock_irqrestore+0xf/0x30
     ? process_one_work+0x420/0x420
     kthread+0x118/0x130
     ? kthread_create_on_node+0x60/0x60
     ret_from_fork+0x35/0x40

We deal this by following what network stack did for some of their
attributes and use mutex_trylock() with restart_syscall(). This makes
userspace release the attribute allowing sysfs attribute removal to
progress before the write is restarted and eventually fail when the
attribute is removed.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

If switch is already disconnected there is no point sending it commands
and waiting for timeout. Instead in that case return error immediately.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

tb_switch_find_by_route() does the same already so use it instead and
remove duplicated get_switch_at_route().
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
Reviewed-by: Lukas Wunner <lukas@wunner.de>

This field is not used anywhere so remove it.
Reported-by: Lukas Wunner <lukas@wunner.de>
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

The XDomain protocol messages may start as soon as Thunderbolt control
channel is started. This means that if the other host starts sending
ThunderboltIP packets early enough they will be passed to the network
driver which then gets confused because its resume hook is not called
yet.

Fix this by unregistering the ThunderboltIP protocol handler when
suspending and registering it back on resume.
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
Acked-by: David S. Miller <davem@davemloft.net>

uuid in add_switch is allocted via kmemdup which can fail. The patch
logs the error and cleans up the allocated memory for switch.
Signed-off-by: Aditya Pakki <pakki001@umn.edu>
Reviewed-by: Mukesh Ojha <mojha@codeaurora.org>
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

No check is enforced for the return value of kzalloc,
which may lead to NULL-pointer dereference.

The patch fixes this issue.
Signed-off-by: Kangjie Lu <kjlu@umn.edu>
Reviewed-by: Mukesh Ojha <mojha@codeaurora.org>
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

kmemdup can fail and return a NULL pointer. The patch modifies the
signature of tb_xdp_schedule_request and passes the failure error upstream.
Signed-off-by: Aditya Pakki <pakki001@umn.edu>
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

In enumerate_services, ida_simple_get on failure can return an error and
leaks memory. The patch ensures that the dev_set_name is set on non
failure cases, and releases memory during failure.
Signed-off-by: Aditya Pakki <pakki001@umn.edu>
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

Memory allocated via kmemdup might fail and return a NULL pointer.
This patch adds a check on the return value of kmemdup and passes the
error upstream.
Signed-off-by: Aditya Pakki <pakki001@umn.edu>
Reviewed-by: Mukesh Ojha <mojha@codeaurora.org>
Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>