diff mbox series

[v6,5/9] soc: mellanox: host: Add the common host side Rshim driver

Message ID 1541089538-175682-5-git-send-email-lsun@mellanox.com (mailing list archive)
State New, archived
Headers show
Series [v6,1/9] soc: Add TmFifo driver for Mellanox BlueField Soc | expand

Commit Message

Liming Sun Nov. 1, 2018, 4:25 p.m. UTC
An external host can connect to a Mellanox BlueField SoC via an
interface called Rshim. The Rshim driver provides boot, console,
and networking services over this interface. This commit is
the common driver where the other backend (transport) driver will
use.

Reviewed-by: David Woods <dwoods@mellanox.com>
Signed-off-by: Liming Sun <lsun@mellanox.com>
---
 drivers/soc/mellanox/Kconfig           |    8 +
 drivers/soc/mellanox/Makefile          |    1 +
 drivers/soc/mellanox/host/Makefile     |    2 +
 drivers/soc/mellanox/host/rshim.c      | 2673 ++++++++++++++++++++++++++++++++
 drivers/soc/mellanox/host/rshim.h      |  361 +++++
 drivers/soc/mellanox/host/rshim_regs.h |  152 ++
 6 files changed, 3197 insertions(+)
 create mode 100644 drivers/soc/mellanox/host/Makefile
 create mode 100644 drivers/soc/mellanox/host/rshim.c
 create mode 100644 drivers/soc/mellanox/host/rshim.h
 create mode 100644 drivers/soc/mellanox/host/rshim_regs.h

Comments

Arnd Bergmann Jan. 18, 2019, 4:02 p.m. UTC | #1
On Thu, Nov 1, 2018 at 5:49 PM Liming Sun <lsun@mellanox.com> wrote:
>
> An external host can connect to a Mellanox BlueField SoC via an
> interface called Rshim. The Rshim driver provides boot, console,
> and networking services over this interface. This commit is
> the common driver where the other backend (transport) driver will
> use.
>
> Reviewed-by: David Woods <dwoods@mellanox.com>
> Signed-off-by: Liming Sun <lsun@mellanox.com>

Hi Liming,

I've taken a new look at your patch series for drivers/soc/ now,
thanks for your continued submissions.

This is again just a set of very high-level comments, but I think we
should resolve some of the fundamental questions first.
Incidentally, Vincent Whitchurch has recently posted another
patch series with a very similar intention, but for other hardware
and taking a different approach.

In both cases, the idea is to use virtio based drivers to provide
services from a host machine into another Linux instance running
on an embedded system behind a PCIe slot or similar. Your
Bluefield SoC patches are well written, but are intentionally
kept specific to a particular use case and tied to one piece
of hardware. In contrast, Vincent uses the existing code from
drivers/misc/mic/vop/ that is equally hardware specific, but he
extends it to be applicable to other hardware as well.

It would be good if you could look at each other's approaches
to see where we could take it from here. I think ideally we
should have a common driver framework for doing the same
thing across both of your devices and as well as others.

That would also resolve my main concern about the drivers,
which is the placement in drivers/soc/ for a set of drivers
that are unlike most drivers in that directory not mean for
running on the SoC itself in order drive unusual functionality
on the SoC, but are (at least partially) meant for running on
a host machine to communicate with that SoC over PCIe
or USB.

As an example, your network driver should really be placed
in drivers/net/, though it is unclear to me how it relates
to the existing virtio_net driver. In the case of mic/vop,
the idea is to use virtio_net on the device side, but have
vhost_net or a user space implementation on the host side,
but that is apparently not what you do here. Can you
explain why?

Another high-level question I have is on how your various
drivers relate to one another. This should normally be
explained in the 0/9 email, but I don't seem to have received
such a mail. I see that you have multiple back-end drivers
for the underlying transport, with one of them based on USB.
Have you come up with a way to use the same high-level
driver such as the network link over this USB back-end,
or is this for something else?

      Arnd
Liming Sun Jan. 21, 2019, 7:22 p.m. UTC | #2
Thanks Arnd for the comments. The 0/9 email was sent out just now to
add more details about the design and changes. Please also see my response
below.

- Liming

> -----Original Message-----
> From: Arnd Bergmann <arnd@arndb.de>
> Sent: Friday, January 18, 2019 11:02 AM
> To: Liming Sun <lsun@mellanox.com>
> Cc: Olof Johansson <olof@lixom.net>; David Woods <dwoods@mellanox.com>; Robin Murphy <robin.murphy@arm.com>; arm-soc
> <arm@kernel.org>; DTML <devicetree@vger.kernel.org>; Linux ARM <linux-arm-kernel@lists.infradead.org>; Vincent Whitchurch
> <vincent.whitchurch@axis.com>; linux-pci <linux-pci@vger.kernel.org>; linux-ntb@googlegroups.com; Christoph Hellwig <hch@lst.de>
> Subject: Re: [PATCH v6 5/9] soc: mellanox: host: Add the common host side Rshim driver
> 
> On Thu, Nov 1, 2018 at 5:49 PM Liming Sun <lsun@mellanox.com> wrote:
> >
> > An external host can connect to a Mellanox BlueField SoC via an
> > interface called Rshim. The Rshim driver provides boot, console,
> > and networking services over this interface. This commit is
> > the common driver where the other backend (transport) driver will
> > use.
> >
> > Reviewed-by: David Woods <dwoods@mellanox.com>
> > Signed-off-by: Liming Sun <lsun@mellanox.com>
> 
> Hi Liming,
> 
> I've taken a new look at your patch series for drivers/soc/ now,
> thanks for your continued submissions.
> 
> This is again just a set of very high-level comments, but I think we
> should resolve some of the fundamental questions first.
> Incidentally, Vincent Whitchurch has recently posted another
> patch series with a very similar intention, but for other hardware
> and taking a different approach.
> 
> In both cases, the idea is to use virtio based drivers to provide
> services from a host machine into another Linux instance running
> on an embedded system behind a PCIe slot or similar. Your
> Bluefield SoC patches are well written, but are intentionally
> kept specific to a particular use case and tied to one piece
> of hardware. In contrast, Vincent uses the existing code from
> drivers/misc/mic/vop/ that is equally hardware specific, but he
> extends it to be applicable to other hardware as well.
> 
> It would be good if you could look at each other's approaches
> to see where we could take it from here. I think ideally we
> should have a common driver framework for doing the same
> thing across both of your devices and as well as others.

Yes, I checked drivers/misc/mic/vop and Vincent Whitchurch's patches 
(Virtio-over-PCIe on non-MIC) and related comments. I kind of feel 
that besides the common virtio infrastructure, there seems not much
to be reused in the rest of implementation yet, though they are trying
to do the similar things.  (Feel free to correct me if I misunderstood it.)

I just submitted the patch 0/9 to explain some details of the rshim
component and the driver patches. Could you help take a look?

The rshim driver of BlueField SoC has a few more functionalities 
which are very HW-specific. Some needs driver support from both 
ARM target and the external host, some only needs external host 
driver support.

As for common framework, we used to implement the drivers based on
the remote proc (Documentation/remoteproc.txt), which seems more
close to what we wanted (in my humble opinion). Later due to more 
functionalities to add and the lack of remote proc in old kernels, we 
changed to use virtio framework directly, which seems very helpful and
saved quite some driver work.

> 
> That would also resolve my main concern about the drivers,
> which is the placement in drivers/soc/ for a set of drivers
> that are unlike most drivers in that directory not mean for
> running on the SoC itself in order drive unusual functionality
> on the SoC, but are (at least partially) meant for running on
> a host machine to communicate with that SoC over PCIe
> or USB.
> 
> As an example, your network driver should really be placed
> in drivers/net/, though it is unclear to me how it relates
> to the existing virtio_net driver. In the case of mic/vop,
> the idea is to use virtio_net on the device side, but have
> vhost_net or a user space implementation on the host side,
> but that is apparently not what you do here. Can you
> explain why?

Yes, I actually have the same concerns where the host side
drivers should go.  For now ther're just added for code review
purpose. drivers/soc/ seems not a good place. One thought
is to move the rshim_net, rshim_pcie and rshim_pcie_lf backend
driver to drivers/net/ethernet/Mellanox/rshim/ and move the
rshim common driver to drivers/char as it creates the character
devices?

The device side of this patch uses the virtio_net driver as well. 

The host side is not just for networking, which was mentioned 
in the 0/9 patch. The host side driver manages the whole rshim
component and is called the 'rshim' driver. It includes driver
to access the TmFifo, where virtio_net is used to provide 
networking support. It needs to talk to the common
driver then the USB or PCIe backend driver.  It seems to me that
vhost_net doesn't quite fit this model and might make it 
over-complicated.

> 
> Another high-level question I have is on how your various
> drivers relate to one another. This should normally be
> explained in the 0/9 email, but I don't seem to have received
> such a mail. I see that you have multiple back-end drivers
> for the underlying transport, with one of them based on USB.
> Have you come up with a way to use the same high-level
> driver such as the network link over this USB back-end,
> or is this for something else?

Yes, 0/9 has been sent. Sorry I should have provided since beginning.

The USB (or PCIe) provide the general transport to access the RShim
component, for networking, console, register access, boot service,
etc. So it's not just for network link. The implementation seems very
HW specific, such as providing APIs like rshim_usb_read_rshim()
and rshim_usb_write_rshim(). In PCIe backend it has similar APIs
like rshim_pcie_read(), rshim_pcie_write().

Not very clear about what you meant the " the same high-level driver 
such as the network link over this USB back-end". Do you mean using
any existing network over USB framework or provide some mechanism
to be reused by other network over USB driver?

By the way, the 0/9 has been sent. Could you help take a look whether 
it clarifies a little bit or not?

> 
>       Arnd
Vincent Whitchurch Jan. 22, 2019, 12:20 p.m. UTC | #3
On Fri, Jan 18, 2019 at 05:02:21PM +0100, Arnd Bergmann wrote:
> On Thu, Nov 1, 2018 at 5:49 PM Liming Sun <lsun@mellanox.com> wrote:
> >
> > An external host can connect to a Mellanox BlueField SoC via an
> > interface called Rshim. The Rshim driver provides boot, console,
> > and networking services over this interface. This commit is
> > the common driver where the other backend (transport) driver will
> > use.
> >
> > Reviewed-by: David Woods <dwoods@mellanox.com>
> > Signed-off-by: Liming Sun <lsun@mellanox.com>
> 
> Hi Liming,
> 
> I've taken a new look at your patch series for drivers/soc/ now,
> thanks for your continued submissions.
> 
> This is again just a set of very high-level comments, but I think we
> should resolve some of the fundamental questions first.
> Incidentally, Vincent Whitchurch has recently posted another
> patch series with a very similar intention, but for other hardware
> and taking a different approach.
> 
> In both cases, the idea is to use virtio based drivers to provide
> services from a host machine into another Linux instance running
> on an embedded system behind a PCIe slot or similar. Your
> Bluefield SoC patches are well written, but are intentionally
> kept specific to a particular use case and tied to one piece
> of hardware. In contrast, Vincent uses the existing code from
> drivers/misc/mic/vop/ that is equally hardware specific, but he
> extends it to be applicable to other hardware as well.
> 
> It would be good if you could look at each other's approaches
> to see where we could take it from here. I think ideally we
> should have a common driver framework for doing the same
> thing across both of your devices and as well as others.

As far as I can see the biggest difference is that Rshim appears to
support interfaces which do not have shared memory between the host and
the card, which means that it has to jump through a lot more hoops to
make virtio work.

For example, the card side seems to use normal virtio-net and
virto-console drivers, but the drivers/soc/mellanox/tmfifo.c driver,
also running on the card, appears to have to actually look inside the
virtqueues and shuffle the data over the TmFifo interface, and this
driver has hard-coded support for only network and console, since it
apparently needs to know the details of how the virtio drivers use their
virtqueues (see tmfifo_virtio_rxtx()).

And the host side appears to _also_ run the virtio-net driver and there
the drivers/soc/mellanox/host/rshim_net.c code instead has to look
inside the virtqueues and shuffle the data over the other side of the
TmFifo interface.

So to me this looks very different from a traditional virtio
driver/device setup (which is what mic/vop uses).  I may be missing
something, but I don't quite understand why it's even using virtio in
the first place.
Liming Sun Jan. 22, 2019, 1:36 p.m. UTC | #4
From: Vincent Whitchurch <vincent.whitchurch@axis.com>
Sent: Tuesday, January 22, 2019 7:20 AM
To: Arnd Bergmann
Cc: Liming Sun; Olof Johansson; David Woods; Robin Murphy; arm-soc; DTML; Linux ARM; linux-pci; linux-ntb@googlegroups.com; Christoph Hellwig; virtualization@lists.linux-foundation.org
Subject: Re: [PATCH v6 5/9] soc: mellanox: host: Add the common host side Rshim driver
  

On Fri, Jan 18, 2019 at 05:02:21PM +0100, Arnd Bergmann wrote:
>> On Thu, Nov 1, 2018 at 5:49 PM Liming Sun <lsun@mellanox.com> wrote:
>> >
>> > An external host can connect to a Mellanox BlueField SoC via an
>> > interface called Rshim. The Rshim driver provides boot, console,
>> > and networking services over this interface. This commit is
>> > the common driver where the other backend (transport) driver will
>> > use.
>> >
>> > Reviewed-by: David Woods <dwoods@mellanox.com>
>> > Signed-off-by: Liming Sun <lsun@mellanox.com>
>> 
>> Hi Liming,
>> 
>> I've taken a new look at your patch series for drivers/soc/ now,
>> thanks for your continued submissions.
>> 
>> This is again just a set of very high-level comments, but I think we
>> should resolve some of the fundamental questions first.
>> Incidentally, Vincent Whitchurch has recently posted another
>> patch series with a very similar intention, but for other hardware
>> and taking a different approach.
>> 
>> In both cases, the idea is to use virtio based drivers to provide
>> services from a host machine into another Linux instance running
>> on an embedded system behind a PCIe slot or similar. Your
>> Bluefield SoC patches are well written, but are intentionally
>> kept specific to a particular use case and tied to one piece
>> of hardware. In contrast, Vincent uses the existing code from
>> drivers/misc/mic/vop/ that is equally hardware specific, but he
>> extends it to be applicable to other hardware as well.
>> 
>> It would be good if you could look at each other's approaches
>> to see where we could take it from here. I think ideally we
>> should have a common driver framework for doing the same
>> thing across both of your devices and as well as others.

> As far as I can see the biggest difference is that Rshim appears to
> support interfaces which do not have shared memory between the host and
> the card, which means that it has to jump through a lot more hoops to
> make virtio work.

> For example, the card side seems to use normal virtio-net and
> virto-console drivers, but the drivers/soc/mellanox/tmfifo.c driver,
> also running on the card, appears to have to actually look inside the
> virtqueues and shuffle the data over the TmFifo interface, and this
> driver has hard-coded support for only network and console, since it
> apparently needs to know the details of how the virtio drivers use their
> virtqueues (see tmfifo_virtio_rxtx()).

> And the host side appears to _also_ run the virtio-net driver and there
> the drivers/soc/mellanox/host/rshim_net.c code instead has to look
> inside the virtqueues and shuffle the data over the other side of the
> TmFifo interface.

> So to me this looks very different from a traditional virtio
> driver/device setup (which is what mic/vop uses).  I may be missing
> something, but I don't quite understand why it's even using virtio in
> the first place.

Thanks  Vincent! This appears to be very good summary of this driver
does on the tmfifo part and the difference between  mic/vop. The fifo is
accessed by register instead of shared memory.

The reason to use virtio framework is that it can be easily used to add
more virtual devices as needed without implementing driver details for
each one. For example, the device side supports console and networking
for now over the FIFO. It only needs to implement function 
tmfifo_virtio_rxtx() once to take care of the virtqueues Rx/Tx, which are
shared by all virtual devices. With minimum changes, we could easily add
another device over tmfifo, like a virtio block device, since the queue
handling is already there. 

The host side handles the virtqueues as well in rshim_net.c. It behaves
like a peer-to-peer to the device side while the tmfifo behaves like a
'wire' (transport) to pass data between the host and the device without
worrying about the data details.
diff mbox series

Patch

diff --git a/drivers/soc/mellanox/Kconfig b/drivers/soc/mellanox/Kconfig
index d88efa1..ecd83a4 100644
--- a/drivers/soc/mellanox/Kconfig
+++ b/drivers/soc/mellanox/Kconfig
@@ -16,3 +16,11 @@  config MLNX_BLUEFIELD_TMFIFO
 	  the implementation of a console and network driver.
 
 endif # ARCH_MLNX_BLUEFIELD
+
+config MLNX_BLUEFIELD_HOST
+	tristate "Mellnox BlueField host side drivers"
+	help
+	  If you say yes to this option, then support will be added
+	  for control and communication of Mellanox BlueField SoCs
+	  from an external host via USB or PCI-express.
+
diff --git a/drivers/soc/mellanox/Makefile b/drivers/soc/mellanox/Makefile
index c44c0e2..aaaf2be 100644
--- a/drivers/soc/mellanox/Makefile
+++ b/drivers/soc/mellanox/Makefile
@@ -3,3 +3,4 @@ 
 # Makefile for Mellanox SoC drivers.
 #
 obj-$(CONFIG_MLNX_BLUEFIELD_TMFIFO)	+= tmfifo.o
+obj-$(CONFIG_MLNX_BLUEFIELD_HOST)	+= host/
diff --git a/drivers/soc/mellanox/host/Makefile b/drivers/soc/mellanox/host/Makefile
new file mode 100644
index 0000000..e47842f
--- /dev/null
+++ b/drivers/soc/mellanox/host/Makefile
@@ -0,0 +1,2 @@ 
+obj-m := rshim.o
+
diff --git a/drivers/soc/mellanox/host/rshim.c b/drivers/soc/mellanox/host/rshim.c
new file mode 100644
index 0000000..32f1124
--- /dev/null
+++ b/drivers/soc/mellanox/host/rshim.c
@@ -0,0 +1,2673 @@ 
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * rshim_common.c - Mellanox host-side driver for RShim
+ *
+ * Copyright 2017 Mellanox Technologies. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.	See the GNU General Public License for more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/fs.h>
+#include <linux/errno.h>
+#include <linux/init.h>
+#include <linux/kref.h>
+#include <linux/module.h>
+#include <linux/mutex.h>
+#include <linux/poll.h>
+#include <linux/seq_file.h>
+#include <linux/slab.h>
+#include <linux/version.h>
+#include <linux/uaccess.h>
+#include <linux/ioctl.h>
+#include <linux/termios.h>
+#include <linux/workqueue.h>
+#include <linux/sched.h>
+#include <asm/termbits.h>
+#include <linux/circ_buf.h>
+#include <linux/delay.h>
+#include <linux/virtio_ids.h>
+
+#include "rshim.h"
+
+/* Maximum number of devices controlled by this driver. */
+int rshim_nr_devs = 64;
+module_param(rshim_nr_devs, int, 0444);
+MODULE_PARM_DESC(rshim_nr_devs, "Maximum number of supported devices");
+
+static char *backend_driver = "";
+module_param(backend_driver, charp, 0444);
+MODULE_PARM_DESC(backend_driver, "Rshim backend driver to use");
+
+static int rshim_keepalive_period = 300;
+module_param(rshim_keepalive_period, int, 0644);
+MODULE_PARM_DESC(rshim_keepalive_period, "keepalive period in milliseconds");
+
+#define RSH_KEEPALIVE_MAGIC_NUM 0x5089836482ULL
+
+/* Circular buffer macros. */
+
+#define read_empty(bd, chan) \
+	(CIRC_CNT((bd)->read_fifo[chan].head, \
+		  (bd)->read_fifo[chan].tail, READ_FIFO_SIZE) == 0)
+#define read_full(bd, chan) \
+	(CIRC_SPACE((bd)->read_fifo[chan].head, \
+		    (bd)->read_fifo[chan].tail, READ_FIFO_SIZE) == 0)
+#define read_space(bd, chan) \
+	CIRC_SPACE((bd)->read_fifo[chan].head, \
+		   (bd)->read_fifo[chan].tail, READ_FIFO_SIZE)
+#define read_cnt(bd, chan) \
+	CIRC_CNT((bd)->read_fifo[chan].head, \
+		 (bd)->read_fifo[chan].tail, READ_FIFO_SIZE)
+#define read_cnt_to_end(bd, chan) \
+	CIRC_CNT_TO_END((bd)->read_fifo[chan].head, \
+			(bd)->read_fifo[chan].tail, READ_FIFO_SIZE)
+#define read_data_ptr(bd, chan) \
+	((bd)->read_fifo[chan].data + \
+	 ((bd)->read_fifo[chan].tail & (READ_FIFO_SIZE - 1)))
+#define read_consume_bytes(bd, chan, nbytes) \
+	((bd)->read_fifo[chan].tail = \
+		((bd)->read_fifo[chan].tail + (nbytes)) & \
+		 (READ_FIFO_SIZE - 1))
+#define read_space_to_end(bd, chan) \
+	CIRC_SPACE_TO_END((bd)->read_fifo[chan].head, \
+			  (bd)->read_fifo[chan].tail, READ_FIFO_SIZE)
+#define read_space_offset(bd, chan) \
+	((bd)->read_fifo[chan].head & (READ_FIFO_SIZE - 1))
+#define read_space_ptr(bd, chan) \
+	((bd)->read_fifo[chan].data + read_space_offset(bd, (chan)))
+#define read_add_bytes(bd, chan, nbytes) \
+	((bd)->read_fifo[chan].head = \
+		((bd)->read_fifo[chan].head + (nbytes)) & \
+		 (READ_FIFO_SIZE - 1))
+#define read_reset(bd, chan) \
+	((bd)->read_fifo[chan].head = (bd)->read_fifo[chan].tail = 0)
+
+#define write_empty(bd, chan) \
+	(CIRC_CNT((bd)->write_fifo[chan].head, \
+		  (bd)->write_fifo[chan].tail, WRITE_FIFO_SIZE) == 0)
+#define write_full(bd, chan) \
+	(CIRC_SPACE((bd)->write_fifo[chan].head, \
+		    (bd)->write_fifo[chan].tail, WRITE_FIFO_SIZE) == 0)
+#define write_space(bd, chan) \
+	CIRC_SPACE((bd)->write_fifo[chan].head, \
+		   (bd)->write_fifo[chan].tail, WRITE_FIFO_SIZE)
+#define write_cnt(bd, chan) \
+	CIRC_CNT((bd)->write_fifo[chan].head, \
+		 (bd)->write_fifo[chan].tail, WRITE_FIFO_SIZE)
+#define write_cnt_to_end(bd, chan) \
+	CIRC_CNT_TO_END((bd)->write_fifo[chan].head, \
+			(bd)->write_fifo[chan].tail, WRITE_FIFO_SIZE)
+#define write_data_offset(bd, chan) \
+	((bd)->write_fifo[chan].tail & (WRITE_FIFO_SIZE - 1))
+#define write_data_ptr(bd, chan) \
+	((bd)->write_fifo[chan].data + write_data_offset(bd, (chan)))
+#define write_consume_bytes(bd, chan, nbytes) \
+	((bd)->write_fifo[chan].tail = \
+		 ((bd)->write_fifo[chan].tail + (nbytes)) & \
+		  (WRITE_FIFO_SIZE - 1))
+#define write_space_to_end(bd, chan) \
+	CIRC_SPACE_TO_END((bd)->write_fifo[chan].head, \
+			  (bd)->write_fifo[chan].tail, WRITE_FIFO_SIZE)
+#define write_space_ptr(bd, chan) \
+	((bd)->write_fifo[chan].data + \
+	 ((bd)->write_fifo[chan].head & (WRITE_FIFO_SIZE - 1)))
+#define write_add_bytes(bd, chan, nbytes) \
+	((bd)->write_fifo[chan].head = \
+	 ((bd)->write_fifo[chan].head + (nbytes)) & \
+	  (WRITE_FIFO_SIZE - 1))
+#define write_reset(bd, chan) \
+	((bd)->write_fifo[chan].head = (bd)->write_fifo[chan].tail = 0)
+
+/*
+ * Tile-to-host bits (UART 0 scratchpad).
+ */
+/*
+ * Output write pointer mask.  Note that this is the maximum size; the
+ * write pointer may be smaller if requested by the host.
+ */
+#define CONS_RSHIM_T2H_OUT_WPTR_MASK     0x3FF
+
+/* Tile is done mask. */
+#define CONS_RSHIM_T2H_DONE_MASK         0x400
+
+/*
+ * Input read pointer mask.  Note that this is the maximum size; the read
+ * pointer may be smaller if requested by the host.
+ */
+#define CONS_RSHIM_T2H_IN_RPTR_MASK      0x1FF800
+
+/* Input read pointer shift. */
+#define CONS_RSHIM_T2H_IN_RPTR_SHIFT     11
+
+/* Tile is done mask. */
+#define CONS_RSHIM_T2H_DONE_MASK         0x400
+
+/* Number of words to send as sync-data (calculated by packet MTU). */
+#define TMFIFO_MAX_SYNC_WORDS            (1536 / 8)
+
+/* Terminal characteristics for newly created consoles. */
+static struct ktermios init_console_termios = {
+	.c_iflag = INLCR | ICRNL,
+	.c_oflag = OPOST | ONLCR,
+	.c_cflag = B115200 | HUPCL | CLOCAL | CREAD | CS8,
+	.c_lflag = ISIG | ICANON | ECHOE | ECHOK | ECHOCTL | ECHOKE | IEXTEN,
+	.c_line = 0,
+	.c_cc = INIT_C_CC,
+};
+
+/* Global mutex. */
+static DEFINE_MUTEX(rshim_mutex);
+
+/*
+ * Array of all of the rshim devices.  The high bits of our minor number
+ * index into this table to find the relevant device.
+ */
+struct rshim_backend **rshim_devs;
+
+/*
+ * Work queue. Right now we have one for the whole driver; we might
+ * eventually decide that we need one per device, but we'll see.
+ */
+struct workqueue_struct *rshim_wq;
+EXPORT_SYMBOL(rshim_wq);
+
+/*
+ * Array of pointers to kmalloc'ed strings, holding the path name for
+ * all of the devices we've seen.  If rshim_devs[i] is non-NULL, then
+ * rshim_dev_names[i] is its path name.  If rshim_devs[i] is NULL, then
+ * rshim_dev_names[i] is the name that was last used for that device.
+ * When we see a new device, we look it up in this table; this allows us to
+ * use the same device index we did last time we saw the device.  The
+ * strings within the array persist until the driver is unloaded.
+ */
+char **rshim_dev_names;
+
+/* Name of the sub-device types. */
+char *rshim_dev_minor_names[RSH_DEV_TYPES] = {
+	[RSH_DEV_TYPE_RSHIM] = "rshim",
+	[RSH_DEV_TYPE_BOOT] = "boot",
+	[RSH_DEV_TYPE_CONSOLE] = "console",
+	[RSH_DEV_TYPE_NET] = "net",
+	[RSH_DEV_TYPE_MISC] = "misc",
+};
+
+/* dev_t base index. */
+static dev_t rshim_dev_base;
+
+/* Class structure for our device class. */
+static struct class *rshim_class;
+
+/* Registered services. */
+static struct rshim_service *rshim_svc[RSH_SVC_MAX];
+
+/* FIFO reset. */
+static void rshim_fifo_reset(struct rshim_backend *bd);
+
+/* Global lock / unlock. */
+
+void rshim_lock(void)
+{
+	mutex_lock(&rshim_mutex);
+}
+EXPORT_SYMBOL(rshim_lock);
+
+void rshim_unlock(void)
+{
+	mutex_unlock(&rshim_mutex);
+}
+EXPORT_SYMBOL(rshim_unlock);
+
+/*
+ * Read some bytes from RShim.
+ *
+ * The provided buffer size should be multiple of 8 bytes. If not, the
+ * leftover bytes (which presumably were sent as NUL bytes by the sender)
+ * will be discarded.
+ */
+static ssize_t rshim_read_default(struct rshim_backend *bd, int devtype,
+				char *buf, size_t count)
+{
+	int retval, total = 0, avail = 0;
+	u64 word;
+
+	/* Read is only supported for RShim TMFIFO. */
+	if (devtype != RSH_DEV_TYPE_NET && devtype != RSH_DEV_TYPE_CONSOLE) {
+		pr_err("bad devtype %d\n", devtype);
+		return -EINVAL;
+	}
+	if (bd->is_boot_open)
+		return 0;
+
+	while (total < count) {
+		if (avail == 0) {
+			retval = bd->read_rshim(bd, RSHIM_CHANNEL,
+						RSH_TM_TILE_TO_HOST_STS, &word);
+			if (retval < 0)
+				break;
+			avail = word & RSH_TM_TILE_TO_HOST_STS__COUNT_MASK;
+			if (avail == 0)
+				break;
+		}
+		retval = bd->read_rshim(bd, RSHIM_CHANNEL,
+					RSH_TM_TILE_TO_HOST_DATA, &word);
+		if (retval < 0)
+			break;
+		/*
+		 * Convert it to little endian before sending to RShim. The
+		 * other side should decode it as little endian as well which
+		 * is usually the default case.
+		 */
+		word = le64_to_cpu(word);
+		if (total + sizeof(word) <= count) {
+			*(u64 *)buf = word;
+			buf += sizeof(word);
+			total += sizeof(word);
+		} else {
+			/* Copy the rest data which is less than 8 bytes. */
+			memcpy(buf, &word, count - total);
+			total = count;
+			break;
+		}
+		avail--;
+	}
+
+	return total;
+}
+
+/*
+ * Write some bytes to the RShim backend.
+ *
+ * If count is not multiple of 8-bytes, the data will be padded to 8-byte
+ * aligned which is required by RShim HW.
+ */
+static ssize_t rshim_write_delayed(struct rshim_backend *bd, int devtype,
+				   const char *buf, size_t count)
+{
+	u64 word;
+	char pad_buf[sizeof(u64)] = { 0 };
+	int size_addr, size_mask, data_addr, max_size;
+	int retval, avail = 0, byte_cnt = 0, retry;
+
+	switch (devtype) {
+	case RSH_DEV_TYPE_NET:
+	case RSH_DEV_TYPE_CONSOLE:
+		if (bd->is_boot_open)
+			return count;
+		size_addr = RSH_TM_HOST_TO_TILE_STS;
+		size_mask = RSH_TM_HOST_TO_TILE_STS__COUNT_MASK;
+		data_addr = RSH_TM_HOST_TO_TILE_DATA;
+		retval = bd->read_rshim(bd, RSHIM_CHANNEL,
+					RSH_TM_HOST_TO_TILE_CTL, &word);
+		if (retval < 0) {
+			pr_err("read_rshim error %d\n", retval);
+			return retval;
+		}
+		max_size = (word >> RSH_TM_HOST_TO_TILE_CTL__MAX_ENTRIES_SHIFT)
+			   & RSH_TM_HOST_TO_TILE_CTL__MAX_ENTRIES_RMASK;
+		break;
+
+	case RSH_DEV_TYPE_BOOT:
+		size_addr = RSH_BOOT_FIFO_COUNT;
+		size_mask = RSH_BOOT_FIFO_COUNT__BOOT_FIFO_COUNT_MASK;
+		data_addr = RSH_BOOT_FIFO_DATA;
+		max_size = RSH_BOOT_FIFO_SIZE;
+		break;
+
+	default:
+		pr_err("bad devtype %d\n", devtype);
+		return -EINVAL;
+	}
+
+	while (byte_cnt < count) {
+		/* Check the boot cancel condition. */
+		if (devtype == RSH_DEV_TYPE_BOOT && !bd->boot_work_buf)
+			break;
+
+		/* Add padding if less than 8 bytes left. */
+		if (byte_cnt + sizeof(u64) > count) {
+			memcpy(pad_buf, buf, count - byte_cnt);
+			buf = (const char *)pad_buf;
+		}
+
+		retry = 0;
+		while (avail <= 0) {
+			/* Calculate available space in words. */
+			retval = bd->read_rshim(bd, RSHIM_CHANNEL, size_addr,
+						&word);
+			if (retval < 0) {
+				pr_err("read_rshim error %d\n", retval);
+				break;
+			}
+			avail = max_size - (int)(word & size_mask) - 8;
+			if (avail > 0)
+				break;
+
+			/*
+			 * Retry 100s, or else return failure since the other
+			 * side seems not to be responding.
+			 */
+			if (++retry > 100000)
+				return -ETIMEDOUT;
+			msleep(1);
+		}
+
+		word = *(u64 *)buf;
+		/*
+		 * Convert to little endian before sending to RShim. The
+		 * receiving side should call le64_to_cpu() to convert
+		 * it back.
+		 */
+		word = cpu_to_le64(word);
+		retval = bd->write_rshim(bd, RSHIM_CHANNEL, data_addr, word);
+		if (retval < 0) {
+			pr_err("write_rshim error %d\n", retval);
+			break;
+		}
+		buf += sizeof(word);
+		byte_cnt += sizeof(word);
+		avail--;
+	}
+
+	/* Return number shouldn't count the padded bytes. */
+	return (byte_cnt > count) ? count : byte_cnt;
+}
+
+static ssize_t rshim_write_default(struct rshim_backend *bd, int devtype,
+				   const char *buf, size_t count)
+{
+	int retval;
+
+	switch (devtype) {
+	case RSH_DEV_TYPE_NET:
+	case RSH_DEV_TYPE_CONSOLE:
+		if (bd->is_boot_open)
+			return count;
+
+		/* Set the flag so there is only one outstanding request. */
+		bd->spin_flags |= RSH_SFLG_WRITING;
+
+		/* Wake up the worker. */
+		bd->fifo_work_buf = (char *)buf;
+		bd->fifo_work_buf_len = count;
+		bd->fifo_work_devtype = devtype;
+		/*
+		 * Add barrier so the above writes complete before setting the
+		 * has_fifo_work flag.
+		 */
+		wmb();
+		bd->has_fifo_work = 1;
+		queue_delayed_work(rshim_wq, &bd->work, 0);
+		return 0;
+
+	case RSH_DEV_TYPE_BOOT:
+		reinit_completion(&bd->boot_write_complete);
+		bd->boot_work_buf_len = count;
+		bd->boot_work_buf_actual_len = 0;
+		/*
+		 * Add barrier so the above writes complete before setting the
+		 * boot_work_buf pointer since it's checked in other places.
+		 */
+		wmb();
+		bd->boot_work_buf = (char *)buf;
+		queue_delayed_work(rshim_wq, &bd->work, 0);
+
+		mutex_unlock(&bd->mutex);
+		retval = wait_for_completion_interruptible(
+					&bd->boot_write_complete);
+		/* Cancel the request if interrupted. */
+		if (retval)
+			bd->boot_work_buf = NULL;
+
+		mutex_lock(&bd->mutex);
+		return bd->boot_work_buf_actual_len;
+
+	default:
+		pr_err("bad devtype %d\n", devtype);
+		return -EINVAL;
+	}
+}
+
+/* Boot file operations routines */
+
+/*
+ * Wait for boot to complete, if necessary.  Return 0 if the boot is done
+ * and it's safe to continue, an error code if something went wrong.  Note
+ * that this routine must be called with the device mutex held.  If it
+ * returns successfully, the mutex will still be held (although it may have
+ * been dropped and reacquired); if it returns unsuccessfully the mutex
+ * will have been dropped.
+ */
+static int wait_for_boot_done(struct rshim_backend *bd)
+{
+	int retval;
+
+	if (!bd->has_reprobe)
+		return 0;
+
+	if (!bd->has_rshim || bd->is_booting) {
+		while (bd->is_booting) {
+			pr_info("boot write, waiting for re-probe\n");
+			/* We're booting, and the backend isn't ready yet. */
+			mutex_unlock(&bd->mutex);
+			/*
+			 * FIXME: might we want a timeout here, too?  If
+			 * the reprobe takes a very long time, something's
+			 * probably wrong.  Maybe a couple of minutes?
+			 */
+			retval = wait_for_completion_interruptible(
+				&bd->booting_complete);
+			if (retval)
+				return retval;
+			mutex_lock(&bd->mutex);
+		}
+		if (!bd->has_rshim) {
+			mutex_unlock(&bd->mutex);
+			return -ENODEV;
+		}
+	}
+
+	return 0;
+}
+
+static ssize_t rshim_boot_write(struct file *file, const char *user_buffer,
+			      size_t count, loff_t *ppos)
+{
+	struct rshim_backend *bd = file->private_data;
+	int retval = 0, whichbuf = 0;
+	size_t bytes_written = 0, bytes_left;
+
+	/*
+	 * Hardware requires that we send multiples of 8 bytes.  Ideally
+	 * we'd handle the case where we got unaligned writes by
+	 * accumulating the residue somehow, but none of our clients
+	 * typically do this, so we just clip the size to prevent any
+	 * inadvertent errors from causing hardware problems.
+	 */
+	bytes_left = count & (-((size_t)8));
+	if (!bytes_left)
+		return 0;
+
+	mutex_lock(&bd->mutex);
+	if (bd->is_in_boot_write) {
+		mutex_unlock(&bd->mutex);
+		return -EBUSY;
+	}
+
+	retval = wait_for_boot_done(bd);
+	if (retval) {
+		pr_err("boot_write: wait for boot failed, err %d\n", retval);
+		/* wait_for_boot_done already dropped mutex */
+		return retval;
+	}
+
+	/*
+	 * We're going to drop the mutex while we wait for any outstanding
+	 * write to complete; this keeps another thread from getting in here
+	 * while we do that.
+	 */
+	bd->is_in_boot_write = 1;
+
+	while (bytes_left) {
+		size_t buf_bytes = min((size_t)BOOT_BUF_SIZE, bytes_left);
+		char *buf = bd->boot_buf[whichbuf];
+
+		whichbuf ^= 1;
+		if (copy_from_user(buf, user_buffer, buf_bytes)) {
+			retval = -EFAULT;
+			pr_err("boot_write: copy from user failed\n");
+			break;
+		}
+
+		retval = bd->write(bd, RSH_DEV_TYPE_BOOT, buf, buf_bytes);
+		if (retval > 0) {
+			bytes_left -= retval;
+			user_buffer += retval;
+			bytes_written += retval;
+		} else if (retval == 0) {
+			/* Wait for some time instead of busy polling. */
+			msleep_interruptible(1);
+			continue;
+		}
+		if (retval != buf_bytes)
+			break;
+	}
+
+	bd->is_in_boot_write = 0;
+	mutex_unlock(&bd->mutex);
+
+	/*
+	 * Return an error in case the 'count' is not multiple of 8 bytes.
+	 * At this moment, the truncated data has already been sent to
+	 * the BOOT fifo and hopefully it could still boot the chip.
+	 */
+	if (count % 8 != 0)
+		return -EINVAL;
+
+	return bytes_written ? bytes_written : retval;
+}
+
+static int rshim_boot_release(struct inode *inode, struct file *file)
+{
+	struct rshim_backend *bd = file->private_data;
+	struct module *owner;
+	int retval;
+
+	/* Restore the boot mode register. */
+	retval = bd->write_rshim(bd, RSHIM_CHANNEL,
+				 RSH_BOOT_CONTROL,
+				 RSH_BOOT_CONTROL__BOOT_MODE_VAL_EMMC);
+	if (retval)
+		pr_err("couldn't set boot_control, err %d\n", retval);
+
+	mutex_lock(&bd->mutex);
+	bd->is_boot_open = 0;
+	queue_delayed_work(rshim_wq, &bd->work, HZ);
+	mutex_unlock(&bd->mutex);
+
+	rshim_lock();
+	owner = RSHIM_READ_ONCE(bd->owner);
+	kref_put(&bd->kref, bd->destroy);
+	module_put(owner);
+	rshim_unlock();
+
+	return 0;
+}
+
+static const struct file_operations rshim_boot_fops = {
+	.owner = THIS_MODULE,
+	.write = rshim_boot_write,
+	.release = rshim_boot_release,
+};
+
+int rshim_boot_open(struct file *file)
+{
+	int retval;
+	int i;
+	struct rshim_backend *bd = file->private_data;
+#if RSH_RESET_MUTEX
+	unsigned long devs_locked = 0;
+#endif
+
+	file->f_op = &rshim_boot_fops;
+
+#if RSH_RESET_MUTEX
+	/*
+	 * We're going to prevent resets and operations from running in
+	 * parallel with other resets.  Our method for this is to grab
+	 * every device's mutex before doing the reset, and then holding
+	 * onto them until the device we reset is reprobed, or a timeout
+	 * expires; the latter is mostly paranoia.  Anyway, in order to
+	 * find all of the other devices, we're going to need to walk the
+	 * device table, so we need to grab its mutex.  We have to do it
+	 * before we get our own device's mutex for lock ordering reasons.
+	 */
+	rshim_lock();
+#endif
+
+	mutex_lock(&bd->mutex);
+
+	if (bd->is_boot_open) {
+		pr_info("can't boot, boot file already open\n");
+		mutex_unlock(&bd->mutex);
+#if RSH_RESET_MUTEX
+		rshim_unlock();
+#endif
+		return -EBUSY;
+	}
+
+	if (!bd->has_rshim) {
+		mutex_unlock(&bd->mutex);
+#if RSH_RESET_MUTEX
+		rshim_unlock();
+#endif
+		return -ENODEV;
+	}
+
+	pr_info("begin booting\n");
+	reinit_completion(&bd->booting_complete);
+	bd->is_booting = 1;
+
+	/*
+	 * Before we reset the chip, make sure we don't have any
+	 * outstanding writes, and flush the write and read FIFOs. (Note
+	 * that we can't have any outstanding reads, since we kill those
+	 * upon release of the TM FIFO file.)
+	 */
+	if (bd->cancel)
+		bd->cancel(bd, RSH_DEV_TYPE_NET, true);
+	bd->read_buf_bytes = 0;
+	bd->read_buf_pkt_rem = 0;
+	bd->read_buf_pkt_padding = 0;
+	spin_lock_irq(&bd->spinlock);
+	/* FIXME: should we be waiting for WRITING to go off, instead? */
+	bd->spin_flags &= ~RSH_SFLG_WRITING;
+	for (i = 0; i < TMFIFO_MAX_CHAN; i++) {
+		read_reset(bd, i);
+		write_reset(bd, i);
+	}
+	spin_unlock_irq(&bd->spinlock);
+
+	/* Set RShim (external) boot mode. */
+	retval = bd->write_rshim(bd, RSHIM_CHANNEL, RSH_BOOT_CONTROL,
+				 RSH_BOOT_CONTROL__BOOT_MODE_VAL_NONE);
+	if (retval) {
+		pr_err("boot_open: error %d writing boot control\n", retval);
+		bd->is_booting = 0;
+		mutex_unlock(&bd->mutex);
+#if RSH_RESET_MUTEX
+		rshim_unlock();
+#endif
+		return retval;
+	}
+
+#if RSH_RESET_MUTEX
+	/*
+	 * Acquire all of the other devices' mutexes, to keep them from
+	 * doing anything while we're performing the reset.  Also kill
+	 * any outstanding boot urbs; that way we'll restart them, after
+	 * the reset is done, and not report errors to the writers.
+	 */
+	for (i = 0; i < rshim_nr_devs; i++) {
+		if (rshim_devs[i] && rshim_devs[i] != bd) {
+			mutex_lock(&rshim_devs[i]->mutex);
+			devs_locked |= 1UL << i;
+			if (rshim_devs[i]->cancel) {
+				rshim_devs[i]->cancel(rshim_devs[i],
+						    RSH_DEV_TYPE_BOOT, true);
+			}
+		}
+	}
+	reinit_completion(&bd->reset_complete);
+#endif
+
+	bd->is_boot_open = 1;
+
+	/* SW reset. */
+	retval = bd->write_rshim(bd, RSHIM_CHANNEL, RSH_RESET_CONTROL,
+				 RSH_RESET_CONTROL__RESET_CHIP_VAL_KEY);
+
+	/* Reset the TmFifo. */
+	rshim_fifo_reset(bd);
+
+	/*
+	 * Note that occasionally, we get various errors on writing to
+	 * the reset register.  This appears to be caused by the chip
+	 * actually resetting before the response goes out, or perhaps by
+	 * our noticing the device unplug before we've seen the response.
+	 * Either way, the chip _does_ actually reset, so we just ignore
+	 * the error.  Should we ever start getting these errors without
+	 * the chip being reset, we'll have to figure out how to handle
+	 * this more intelligently.  (One potential option is to not reset
+	 * directly, but to set up a down counter to do the reset, but that
+	 * seems kind of kludgy, especially since Tile software might also
+	 * be trying to use the down counter.)
+	 */
+	if (retval && retval != -EPROTO && retval != -ESHUTDOWN &&
+#ifdef RSH_USB_BMC
+	    /*
+	     * The host driver on the BMC sometimes produces EOVERFLOW on
+	     * reset.  It also seems to have seems to have some sort of bug
+	     * which makes it return more bytes than we actually wrote!  In
+	     * that case we're returning EBADE.
+	     */
+	    retval != -EOVERFLOW && retval != -EBADE &&
+#endif
+	    retval != -ETIMEDOUT && retval != -EPIPE) {
+		pr_err("boot_open: error %d writing reset control\n", retval);
+		mutex_unlock(&bd->mutex);
+#if RSH_RESET_MUTEX
+		while (devs_locked) {
+			int i = __builtin_ctzl(devs_locked);
+
+			mutex_unlock(&rshim_devs[i]->mutex);
+			devs_locked &= ~(1UL << i);
+		}
+		rshim_unlock();
+#endif
+		bd->is_boot_open = 0;
+
+		return retval;
+	}
+
+	if (retval)
+		pr_err("boot_open: got error %d on reset write\n", retval);
+
+	mutex_unlock(&bd->mutex);
+
+#if RSH_RESET_MUTEX
+	rshim_unlock();
+	/*
+	 * We wait for reset_complete (signaled by probe), or for an
+	 * interrupt, or a timeout (set to 5s because of no re-probe
+	 * in the PCIe case). Note that we dropped dev->mutex above
+	 * so that probe can run; the BOOT_OPEN flag should keep our device
+	 * from trying to do anything before the device is reprobed.
+	 */
+	retval = wait_for_completion_interruptible_timeout(&bd->reset_complete,
+							   5 * HZ);
+	if (retval == 0)
+		pr_err("timed out waiting for device reprobe after reset\n");
+
+	while (devs_locked) {
+		int i = __builtin_ctz(devs_locked);
+
+		mutex_unlock(&rshim_devs[i]->mutex);
+		devs_locked &= ~(1UL << i);
+	}
+#endif
+
+	return 0;
+}
+
+/* FIFO common file operations routines */
+
+/*
+ * Signal an error on the FIFO, and wake up anyone who might need to know
+ * about it.
+ */
+static void rshim_fifo_err(struct rshim_backend *bd, int err)
+{
+	int i;
+
+	bd->tmfifo_error = err;
+	wake_up_interruptible_all(&bd->write_completed);
+	for (i = 0; i < TMFIFO_MAX_CHAN; i++) {
+		wake_up_interruptible_all(&bd->read_fifo[i].operable);
+		wake_up_interruptible_all(&bd->write_fifo[i].operable);
+	}
+}
+
+/* Drain the read buffer, and start another read/interrupt if needed. */
+static void rshim_fifo_input(struct rshim_backend *bd)
+{
+	union rshim_tmfifo_msg_hdr *hdr;
+	bool rx_avail = false;
+
+	if (bd->is_boot_open)
+		return;
+
+again:
+	while (bd->read_buf_next < bd->read_buf_bytes) {
+		int copysize;
+
+		/*
+		 * If we're at the start of a packet, then extract the
+		 * header, and update our count of bytes remaining in the
+		 * packet.
+		 */
+		if (bd->read_buf_pkt_rem == 0) {
+			/* Make sure header is received. */
+			if (bd->read_buf_next + sizeof(*hdr) >
+				bd->read_buf_bytes)
+				break;
+
+			pr_debug("next hdr %d\n", bd->read_buf_next);
+
+			hdr = (union rshim_tmfifo_msg_hdr *)
+				&bd->read_buf[bd->read_buf_next];
+
+			bd->read_buf_pkt_rem = ntohs(hdr->len) + sizeof(*hdr);
+			bd->read_buf_pkt_padding =
+				(8 - (bd->read_buf_pkt_rem & 7)) & 7;
+			if (hdr->type == VIRTIO_ID_NET)
+				bd->rx_chan = TMFIFO_NET_CHAN;
+			else if (hdr->type == VIRTIO_ID_CONSOLE) {
+				bd->rx_chan = TMFIFO_CONS_CHAN;
+				/* Strip off the message header for console. */
+				bd->read_buf_next += sizeof(*hdr);
+				bd->read_buf_pkt_rem -= sizeof(*hdr);
+				if (bd->read_buf_pkt_rem == 0)
+					continue;
+			} else {
+				pr_debug("bad type %d, drop it", hdr->type);
+				bd->read_buf_pkt_rem = 0;
+				bd->read_buf_pkt_padding = 0;
+				bd->read_buf_next = bd->read_buf_bytes;
+				break;
+			}
+
+			pr_debug("drain: hdr, nxt %d rem %d chn %d\n",
+			      bd->read_buf_next, bd->read_buf_pkt_rem,
+			      bd->rx_chan);
+			bd->drop = 0;
+		}
+
+		if (bd->rx_chan == TMFIFO_CONS_CHAN &&
+		    !(bd->spin_flags & RSH_SFLG_CONS_OPEN)) {
+			/*
+			 * If data is coming in for a closed console
+			 * channel, we want to just throw it away.
+			 * Resetting the channel every time through this
+			 * loop is a relatively cheap way to do that.  Note
+			 * that this works because the read buffer is no
+			 * larger than the read FIFO; thus, we know that if
+			 * we reset it here, we will always be able to
+			 * drain the read buffer of any console data, and
+			 * will then launch another read.
+			 */
+			read_reset(bd, TMFIFO_CONS_CHAN);
+			bd->drop = 1;
+		} else if (bd->rx_chan == TMFIFO_NET_CHAN && bd->net == NULL) {
+			/* Drop if networking is not enabled. */
+			read_reset(bd, TMFIFO_NET_CHAN);
+			bd->drop = 1;
+		}
+
+		copysize = min(bd->read_buf_pkt_rem,
+			       bd->read_buf_bytes - bd->read_buf_next);
+		copysize = min(copysize,
+			       read_space_to_end(bd, bd->rx_chan));
+
+		pr_debug("drain: copysize %d, head %d, tail %d, remaining %d\n",
+			 copysize, bd->read_fifo[bd->rx_chan].head,
+			 bd->read_fifo[bd->rx_chan].tail,
+			 bd->read_buf_pkt_rem);
+
+		if (copysize == 0) {
+			/*
+			 * We have data, but no space to put it in, so
+			 * we're done.
+			 */
+			pr_debug("drain: no more space in channel %d\n",
+				 bd->rx_chan);
+			break;
+		}
+
+		if (!bd->drop) {
+			memcpy(read_space_ptr(bd, bd->rx_chan),
+			       &bd->read_buf[bd->read_buf_next],
+			       copysize);
+			read_add_bytes(bd, bd->rx_chan, copysize);
+		}
+
+		bd->read_buf_next += copysize;
+		bd->read_buf_pkt_rem -= copysize;
+
+		wake_up_interruptible_all(&bd->read_fifo[
+				      bd->rx_chan].operable);
+		pr_debug("woke up readable chan %d\n", bd->rx_chan);
+
+		if (bd->read_buf_pkt_rem <= 0) {
+			bd->read_buf_next = bd->read_buf_next +
+				bd->read_buf_pkt_padding;
+			rx_avail = true;
+		}
+	}
+
+	/*
+	 * We've processed all of the data we can, so now we decide if we
+	 * need to launch another I/O.  If there's still data in the read
+	 * buffer, or if we're already reading, don't launch any new
+	 * operations.  If an interrupt just completed, and said there was
+	 * data, or the last time we did a read we got some data, then do
+	 * another read.  Otherwise, do an interrupt.
+	 */
+	if (bd->read_buf_next < bd->read_buf_bytes ||
+	    (bd->spin_flags & RSH_SFLG_READING)) {
+		/* We're doing nothing. */
+		pr_debug("fifo_input: no new read: %s\n",
+			 (bd->read_buf_next < bd->read_buf_bytes) ?
+			 "have data" : "already reading");
+	} else {
+		int len;
+
+		/* Process it if more data is received. */
+		len = bd->read(bd, RSH_DEV_TYPE_NET, (char *)bd->read_buf,
+			      READ_BUF_SIZE);
+		if (len > 0) {
+			bd->read_buf_bytes = len;
+			bd->read_buf_next = 0;
+			goto again;
+		}
+	}
+
+	if (rx_avail) {
+		if (bd->rx_chan == TMFIFO_NET_CHAN) {
+			struct rshim_service *svc;
+
+			/*
+			 * Protect rshim_svc with RCU lock. See comments in
+			 * rshim_register_service() / rshim_register_service()
+			 */
+			rcu_read_lock();
+			svc = rcu_dereference(rshim_svc[RSH_SVC_NET]);
+			if (svc != NULL)
+				(*svc->rx_notify)(bd);
+			rcu_read_unlock();
+		}
+	}
+}
+
+ssize_t rshim_fifo_read(struct rshim_backend *bd, char *buffer,
+		      size_t count, int chan, bool nonblock,
+		      bool to_user)
+{
+	size_t rd_cnt = 0;
+
+	mutex_lock(&bd->mutex);
+
+	while (count) {
+		size_t readsize;
+		int pass1;
+		int pass2;
+
+		pr_debug("fifo_read, top of loop, remaining count %zd\n",
+			 count);
+
+		/*
+		 * We check this each time through the loop since the
+		 * device could get disconnected while we're waiting for
+		 * more data in the read FIFO.
+		 */
+		if (!bd->has_tm) {
+			mutex_unlock(&bd->mutex);
+			pr_debug("fifo_read: returning %zd/ENODEV\n", rd_cnt);
+			return rd_cnt ? rd_cnt : -ENODEV;
+		}
+
+		if (bd->tmfifo_error) {
+			mutex_unlock(&bd->mutex);
+			pr_debug("fifo_read: returning %zd/%d\n", rd_cnt,
+			      bd->tmfifo_error);
+			return rd_cnt ? rd_cnt : bd->tmfifo_error;
+		}
+
+		if (read_empty(bd, chan)) {
+			pr_debug("fifo_read: fifo empty\n");
+			if (rd_cnt || nonblock) {
+				if (rd_cnt == 0) {
+					spin_lock_irq(&bd->spinlock);
+					rshim_fifo_input(bd);
+					spin_unlock_irq(&bd->spinlock);
+				}
+				mutex_unlock(&bd->mutex);
+				pr_debug("fifo_read: returning %zd/EAGAIN\n",
+				      rd_cnt);
+				return rd_cnt ? rd_cnt : -EAGAIN;
+			}
+
+			mutex_unlock(&bd->mutex);
+
+			pr_debug("fifo_read: waiting for readable chan %d\n",
+				 chan);
+			if (wait_event_interruptible(
+					bd->read_fifo[chan].operable,
+					    !read_empty(bd, chan))) {
+				pr_debug("fifo_read: returning ERESTARTSYS\n");
+				return to_user ? -EINTR : -ERESTARTSYS;
+			}
+
+			mutex_lock(&bd->mutex);
+
+			/*
+			 * Since we dropped the mutex, we must make
+			 * sure our interface is still there before
+			 * we do anything else.
+			 */
+			continue;
+		}
+
+		/*
+		 * Figure out how many bytes we will transfer on this pass.
+		 */
+		spin_lock_irq(&bd->spinlock);
+
+		readsize = min(count, (size_t)read_cnt(bd, chan));
+
+		pass1 = min(readsize, (size_t)read_cnt_to_end(bd, chan));
+		pass2 = readsize - pass1;
+
+		spin_unlock_irq(&bd->spinlock);
+
+		pr_debug("fifo_read: readsize %zd, head %d, tail %d\n",
+			 readsize, bd->read_fifo[chan].head,
+			 bd->read_fifo[chan].tail);
+
+		if (!to_user) {
+			memcpy(buffer, read_data_ptr(bd, chan), pass1);
+			if (pass2) {
+				memcpy(buffer + pass1,
+				       bd->read_fifo[chan].data, pass2);
+			}
+		} else {
+			if (copy_to_user(buffer, read_data_ptr(bd, chan),
+				pass1) || (pass2 && copy_to_user(buffer + pass1,
+				bd->read_fifo[chan].data, pass2))) {
+				mutex_unlock(&bd->mutex);
+				pr_debug("fifo_read: returns %zd/EFAULT\n",
+					 rd_cnt);
+				return rd_cnt ? rd_cnt : -EFAULT;
+			}
+		}
+
+		spin_lock_irq(&bd->spinlock);
+
+		read_consume_bytes(bd, chan, readsize);
+
+		/*
+		 * We consumed some bytes, so let's see if we can process
+		 * any more incoming data.
+		 */
+		rshim_fifo_input(bd);
+
+		spin_unlock_irq(&bd->spinlock);
+
+		count -= readsize;
+		buffer += readsize;
+		rd_cnt += readsize;
+		pr_debug("fifo_read: transferred %zd bytes\n", readsize);
+	}
+
+	mutex_unlock(&bd->mutex);
+
+	pr_debug("fifo_read: returning %zd\n", rd_cnt);
+	return rd_cnt;
+}
+EXPORT_SYMBOL(rshim_fifo_read);
+
+static void rshim_fifo_output(struct rshim_backend *bd)
+{
+	int writesize, write_buf_next = 0;
+	int write_avail = WRITE_BUF_SIZE - write_buf_next;
+	int numchan = TMFIFO_MAX_CHAN;
+	int chan, chan_offset;
+
+	/* If we're already writing, we have nowhere to put data. */
+	if (bd->spin_flags & RSH_SFLG_WRITING)
+		return;
+
+	/* Walk through all the channels, sending as much data as possible. */
+	for (chan_offset = 0; chan_offset < numchan; chan_offset++) {
+		/*
+		 * Pick the current channel if not done, otherwise round-robin
+		 * to the next channel.
+		 */
+		if (bd->write_buf_pkt_rem > 0)
+			chan = bd->tx_chan;
+		else {
+			u16 cur_len;
+			union rshim_tmfifo_msg_hdr *hdr = &bd->msg_hdr;
+
+			chan = bd->tx_chan = (bd->tx_chan + 1) % numchan;
+			cur_len = write_cnt(bd, chan);
+
+			/*
+			 * Set up message header for console data which is byte
+			 * stream. Network packets already have the message
+			 * header included.
+			 */
+			if (chan == TMFIFO_CONS_CHAN) {
+				if (cur_len == 0)
+					continue;
+				hdr->data = 0;
+				hdr->type = VIRTIO_ID_CONSOLE;
+				hdr->len = htons(cur_len);
+			} else {
+				int pass1;
+
+				if (cur_len <
+					sizeof(union rshim_tmfifo_msg_hdr))
+					continue;
+
+				pass1 = write_cnt_to_end(bd, chan);
+				if (pass1 >= sizeof(*hdr)) {
+					hdr = (union rshim_tmfifo_msg_hdr *)
+						write_data_ptr(bd, chan);
+				} else {
+					memcpy(hdr, write_data_ptr(bd, chan),
+					       pass1);
+					memcpy((u8 *)hdr + pass1,
+					       bd->write_fifo[chan].data,
+					       sizeof(*hdr) - pass1);
+				}
+			}
+
+			bd->write_buf_pkt_rem = ntohs(hdr->len) + sizeof(*hdr);
+		}
+
+		/* Send out the packet header for the console data. */
+		if (chan == TMFIFO_CONS_CHAN &&
+		    bd->write_buf_pkt_rem > ntohs(bd->msg_hdr.len)) {
+			union rshim_tmfifo_msg_hdr *hdr = &bd->msg_hdr;
+			int left = bd->write_buf_pkt_rem - ntohs(hdr->len);
+			u8 *pos = (u8 *)hdr + sizeof(*hdr) - left;
+
+			writesize = min(write_avail, left);
+			memcpy(&bd->write_buf[write_buf_next], pos, writesize);
+			write_buf_next += writesize;
+			bd->write_buf_pkt_rem -= writesize;
+			write_avail -= writesize;
+
+			/*
+			 * Don't continue if no more space for the header.
+			 * It'll be picked up next time.
+			 */
+			if (left != writesize)
+				break;
+		}
+
+		writesize = min(write_avail, (int)write_cnt(bd, chan));
+		writesize = min(writesize, bd->write_buf_pkt_rem);
+
+		/*
+		 * The write size should be aligned to 8 bytes unless for the
+		 * last block, which will be padded at the end.
+		 */
+		if (bd->write_buf_pkt_rem != writesize)
+			writesize &= -8;
+
+		if (writesize > 0) {
+			int pass1;
+			int pass2;
+
+			pass1 = min(writesize,
+				    (int)write_cnt_to_end(bd, chan));
+			pass2 = writesize - pass1;
+
+			pr_debug("fifo_outproc: chan %d, writesize %d, next %d,"
+				 " head %d, tail %d\n",
+				 chan, writesize, write_buf_next,
+				 bd->write_fifo[chan].head,
+				 bd->write_fifo[chan].tail);
+
+			memcpy(&bd->write_buf[write_buf_next],
+			       write_data_ptr(bd, chan), pass1);
+			memcpy(&bd->write_buf[write_buf_next + pass1],
+			       bd->write_fifo[chan].data, pass2);
+
+			write_consume_bytes(bd, chan, writesize);
+			write_buf_next += writesize;
+			bd->write_buf_pkt_rem -= writesize;
+			/* Add padding at the end. */
+			if (bd->write_buf_pkt_rem == 0)
+				write_buf_next = (write_buf_next + 7) & -8;
+			write_avail = WRITE_BUF_SIZE - write_buf_next;
+
+			wake_up_interruptible_all(
+				&bd->write_fifo[chan].operable);
+			pr_debug("woke up writable chan %d\n", chan);
+		}
+	}
+
+	/* Drop the data if it is still booting. */
+	if (bd->is_boot_open)
+		return;
+
+	/* If we actually put anything in the buffer, send it. */
+	if (write_buf_next) {
+		bd->write(bd, RSH_DEV_TYPE_NET, (char *)bd->write_buf,
+			  write_buf_next);
+	}
+}
+
+int rshim_fifo_alloc(struct rshim_backend *bd)
+{
+	int i, allocfail = 0;
+
+	for (i = 0; i < TMFIFO_MAX_CHAN; i++) {
+		if (!bd->read_fifo[i].data)
+			bd->read_fifo[i].data =
+				kmalloc(READ_FIFO_SIZE, GFP_KERNEL);
+		allocfail |= bd->read_fifo[i].data == 0;
+
+		if (!bd->write_fifo[i].data)
+			bd->write_fifo[i].data =
+				kmalloc(WRITE_FIFO_SIZE, GFP_KERNEL);
+		allocfail |= bd->write_fifo[i].data == 0;
+	}
+
+	return allocfail;
+}
+EXPORT_SYMBOL(rshim_fifo_alloc);
+
+static void rshim_fifo_reset(struct rshim_backend *bd)
+{
+	int i;
+
+	bd->read_buf_bytes = 0;
+	bd->read_buf_pkt_rem = 0;
+	bd->read_buf_next = 0;
+	bd->read_buf_pkt_padding = 0;
+	bd->write_buf_pkt_rem = 0;
+	bd->rx_chan = bd->tx_chan = 0;
+
+	spin_lock_irq(&bd->spinlock);
+	bd->spin_flags &= ~(RSH_SFLG_WRITING |
+			    RSH_SFLG_READING);
+	for (i = 0; i < TMFIFO_MAX_CHAN; i++) {
+		read_reset(bd, i);
+		write_reset(bd, i);
+	}
+	spin_unlock_irq(&bd->spinlock);
+}
+
+void rshim_fifo_free(struct rshim_backend *bd)
+{
+	int i;
+
+	for (i = 0; i < TMFIFO_MAX_CHAN; i++) {
+		kfree(bd->read_fifo[i].data);
+		bd->read_fifo[i].data = NULL;
+		kfree(bd->write_fifo[i].data);
+		bd->write_fifo[i].data = NULL;
+	}
+
+	rshim_fifo_reset(bd);
+
+	bd->has_tm = 0;
+}
+EXPORT_SYMBOL(rshim_fifo_free);
+
+ssize_t rshim_fifo_write(struct rshim_backend *bd, const char *buffer,
+		       size_t count, int chan, bool nonblock,
+		       bool from_user)
+{
+	size_t wr_cnt = 0;
+
+	mutex_lock(&bd->mutex);
+
+	while (count) {
+		size_t writesize;
+		int pass1;
+		int pass2;
+
+		pr_debug("fifo_write, top of loop, remaining count %zd\n",
+			 count);
+
+		/*
+		 * We check this each time through the loop since the
+		 * device could get disconnected while we're waiting for
+		 * more space in the write buffer.
+		 */
+		if (!bd->has_tm) {
+			mutex_unlock(&bd->mutex);
+			pr_debug("fifo_write: returning %zd/ENODEV\n", wr_cnt);
+			return wr_cnt ? wr_cnt : -ENODEV;
+		}
+
+		if (bd->tmfifo_error) {
+			mutex_unlock(&bd->mutex);
+			pr_debug("fifo_write: returning %zd/%d\n", wr_cnt,
+				 bd->tmfifo_error);
+			return wr_cnt ? wr_cnt : bd->tmfifo_error;
+		}
+
+		if (write_full(bd, chan)) {
+			pr_debug("fifo_write: fifo full\n");
+			if (nonblock) {
+				mutex_unlock(&bd->mutex);
+				pr_debug("fifo_write: returning %zd/EAGAIN\n",
+					 wr_cnt);
+				return wr_cnt ? wr_cnt : -EAGAIN;
+			}
+
+			mutex_unlock(&bd->mutex);
+			pr_debug("fifo_write: waiting for writable chan %d\n",
+				 chan);
+			if (wait_event_interruptible(
+				     bd->write_fifo[chan].operable,
+					     !write_full(bd, chan))) {
+				pr_debug("fifo_write: returning %zd/ERESTARTSYS\n",
+					 wr_cnt);
+				return wr_cnt ? wr_cnt : -ERESTARTSYS;
+			}
+			mutex_lock(&bd->mutex);
+			/*
+			 * Since we dropped the mutex, we must make
+			 * sure our interface is still there before
+			 * we do anything else.
+			 */
+			continue;
+		}
+
+		spin_lock_irq(&bd->spinlock);
+
+		writesize = min(count, (size_t)write_space(bd, chan));
+		pass1 = min(writesize, (size_t)write_space_to_end(bd, chan));
+		pass2 = writesize - pass1;
+
+		spin_unlock_irq(&bd->spinlock);
+
+		pr_debug("fifo_write: writesize %zd, head %d, tail %d\n",
+			 writesize, bd->write_fifo[chan].head,
+			 bd->write_fifo[chan].tail);
+
+		if (!from_user) {
+			memcpy(write_space_ptr(bd, chan), buffer, pass1);
+			if (pass2) {
+				memcpy(bd->write_fifo[chan].data,
+				       buffer + pass1, pass2);
+			}
+		} else {
+			if (copy_from_user(write_space_ptr(bd, chan), buffer,
+				pass1) || (pass2 &&
+				copy_from_user(bd->write_fifo[chan].data,
+						buffer + pass1, pass2))) {
+				mutex_unlock(&bd->mutex);
+				pr_debug("fifo_write: returns %zd/EFAULT\n",
+					 wr_cnt);
+				return wr_cnt ? wr_cnt : -EFAULT;
+			}
+		}
+
+		spin_lock_irq(&bd->spinlock);
+
+		write_add_bytes(bd, chan, writesize);
+
+		/* We have some new bytes, let's see if we can write any. */
+		rshim_fifo_output(bd);
+
+		spin_unlock_irq(&bd->spinlock);
+
+		count -= writesize;
+		buffer += writesize;
+		wr_cnt += writesize;
+		pr_debug("fifo_write: transferred %zd bytes this pass\n",
+			 writesize);
+	}
+
+	mutex_unlock(&bd->mutex);
+
+	pr_debug("fifo_write: returning %zd\n", wr_cnt);
+	return wr_cnt;
+}
+EXPORT_SYMBOL(rshim_fifo_write);
+
+static int rshim_fifo_fsync(struct file *file, loff_t start, loff_t end,
+			    int datasync, int chan)
+{
+	struct rshim_backend *bd = file->private_data;
+
+	mutex_lock(&bd->mutex);
+
+	/*
+	 * To ensure that all of our data has actually made it to the
+	 * device, we first wait until the channel is empty, then we wait
+	 * until there is no outstanding write urb.
+	 */
+	while (!write_empty(bd, chan))
+		if (wait_event_interruptible(bd->write_fifo[chan].operable,
+					     write_empty(bd, chan))) {
+			mutex_unlock(&bd->mutex);
+			return -ERESTARTSYS;
+		}
+
+	while (bd->spin_flags & RSH_SFLG_WRITING)
+		if (wait_event_interruptible(bd->write_completed,
+					     !(bd->spin_flags &
+					       RSH_SFLG_WRITING))) {
+			mutex_unlock(&bd->mutex);
+			return -ERESTARTSYS;
+		}
+
+	mutex_unlock(&bd->mutex);
+
+	return 0;
+}
+
+static unsigned int rshim_fifo_poll(struct file *file, poll_table *wait,
+				  int chan)
+{
+	struct rshim_backend *bd = file->private_data;
+	unsigned int retval = 0;
+
+	mutex_lock(&bd->mutex);
+
+	poll_wait(file, &bd->read_fifo[chan].operable, wait);
+	poll_wait(file, &bd->write_fifo[chan].operable, wait);
+
+	spin_lock_irq(&bd->spinlock);
+
+	if (!read_empty(bd, chan))
+		retval |= POLLIN | POLLRDNORM;
+	if (!write_full(bd, chan))
+		retval |= POLLOUT | POLLWRNORM;
+	/*
+	 * We don't report POLLERR on the console so that it doesn't get
+	 * automatically disconnected when it fails, and so that you can
+	 * connect to it in the error state before rebooting the target.
+	 * This is inconsistent, but being consistent turns out to be very
+	 * annoying.  If someone tries to actually type on it, they'll
+	 * get an error.
+	 */
+	if (bd->tmfifo_error && chan != TMFIFO_CONS_CHAN)
+		retval |= POLLERR;
+	spin_unlock_irq(&bd->spinlock);
+
+	mutex_unlock(&bd->mutex);
+
+	pr_debug("poll chan %d file %p returns 0x%x\n", chan, file, retval);
+
+	return retval;
+}
+
+
+static int rshim_fifo_release(struct inode *inode, struct file *file,
+			      int chan)
+{
+	struct rshim_backend *bd = file->private_data;
+	struct module *owner;
+
+	mutex_lock(&bd->mutex);
+
+	if (chan == TMFIFO_CONS_CHAN) {
+		/*
+		 * If we aren't the last console file, nothing to do but
+		 * fix the reference count.
+		 */
+		bd->console_opens--;
+		if (bd->console_opens) {
+			mutex_unlock(&bd->mutex);
+			return 0;
+		}
+
+		/*
+		 * We've told the host to stop using the TM FIFO console,
+		 * but there may be a lag before it does.  Unless we
+		 * continue to read data from the console stream, the host
+		 * may spin forever waiting for the console to be drained
+		 * and not realize that it's time to stop using it.
+		 * Clearing the CONS_OPEN spin flag will discard any future
+		 * incoming console data, but if our input buffers are full
+		 * now, we might not be even reading from the hardware
+		 * FIFO.  To avoid problems, clear the buffers and call the
+		 * drainer so that it knows there's space.
+		 */
+		spin_lock_irq(&bd->spinlock);
+
+		bd->spin_flags &= ~RSH_SFLG_CONS_OPEN;
+
+		read_reset(bd, TMFIFO_CONS_CHAN);
+		write_reset(bd, TMFIFO_CONS_CHAN);
+
+		if (bd->has_tm)
+			rshim_fifo_input(bd);
+
+		spin_unlock_irq(&bd->spinlock);
+	}
+
+	if (chan == TMFIFO_CONS_CHAN)
+		bd->is_cons_open = 0;
+	else
+		bd->is_tm_open = 0;
+
+	if (!bd->is_tm_open && !bd->is_cons_open) {
+		if (bd->cancel)
+			bd->cancel(bd, RSH_DEV_TYPE_NET, false);
+
+		spin_lock_irq(&bd->spinlock);
+		bd->spin_flags &= ~RSH_SFLG_READING;
+		spin_unlock_irq(&bd->spinlock);
+	}
+
+	mutex_unlock(&bd->mutex);
+
+	rshim_lock();
+	owner = RSHIM_READ_ONCE(bd->owner);
+	kref_put(&bd->kref, bd->destroy);
+	module_put(owner);
+	rshim_unlock();
+
+	return 0;
+}
+
+/* TMFIFO file operations routines */
+
+static ssize_t rshim_tmfifo_read(struct file *file, char *user_buffer,
+				   size_t count, loff_t *ppos)
+{
+	struct rshim_backend *bd = file->private_data;
+
+	return rshim_fifo_read(bd, user_buffer, count, TMFIFO_NET_CHAN,
+			     file->f_flags & O_NONBLOCK, true);
+}
+
+static ssize_t rshim_tmfifo_write(struct file *file, const char *user_buffer,
+				size_t count, loff_t *ppos)
+{
+	struct rshim_backend *bd = file->private_data;
+
+	return rshim_fifo_write(bd, user_buffer, count, TMFIFO_NET_CHAN,
+			      file->f_flags & O_NONBLOCK, true);
+}
+
+static int rshim_tmfifo_fsync(struct file *file, loff_t start,
+			      loff_t end, int datasync)
+{
+	return rshim_fifo_fsync(file, start, end, datasync, TMFIFO_NET_CHAN);
+}
+
+static unsigned int rshim_tmfifo_poll(struct file *file, poll_table *wait)
+{
+	return rshim_fifo_poll(file, wait, TMFIFO_NET_CHAN);
+}
+
+static int rshim_tmfifo_release(struct inode *inode, struct file *file)
+{
+	return rshim_fifo_release(inode, file, TMFIFO_NET_CHAN);
+}
+
+static const struct file_operations rshim_tmfifo_fops = {
+	.owner = THIS_MODULE,
+	.read = rshim_tmfifo_read,
+	.write = rshim_tmfifo_write,
+	.fsync = rshim_tmfifo_fsync,
+	.poll = rshim_tmfifo_poll,
+	.release = rshim_tmfifo_release,
+};
+
+static int rshim_tmfifo_open(struct file *file)
+{
+	struct rshim_backend *bd = file->private_data;
+
+	file->f_op = &rshim_tmfifo_fops;
+
+	mutex_lock(&bd->mutex);
+
+	if (bd->is_tm_open) {
+		pr_debug("tmfifo_open: file already open\n");
+		mutex_unlock(&bd->mutex);
+		return -EBUSY;
+	}
+
+	bd->is_tm_open = 1;
+
+	spin_lock_irq(&bd->spinlock);
+
+	/* Call the drainer to do an initial read, if needed. */
+	rshim_fifo_input(bd);
+
+	spin_unlock_irq(&bd->spinlock);
+
+	mutex_unlock(&bd->mutex);
+
+	return 0;
+}
+
+/* Console file operations routines */
+
+static void rshim_work_handler(struct work_struct *work)
+{
+	struct rshim_backend *bd = container_of((struct delayed_work *) work,
+					      struct rshim_backend, work);
+
+	mutex_lock(&bd->mutex);
+
+	if (bd->keepalive && bd->has_rshim) {
+		bd->write_rshim(bd, RSHIM_CHANNEL, RSH_SCRATCHPAD1,
+				RSH_KEEPALIVE_MAGIC_NUM);
+		bd->keepalive = 0;
+	}
+
+	if (bd->boot_work_buf != NULL) {
+		bd->boot_work_buf_actual_len = rshim_write_delayed(bd,
+							RSH_DEV_TYPE_BOOT,
+							bd->boot_work_buf,
+							bd->boot_work_buf_len);
+		bd->boot_work_buf = NULL;
+		complete_all(&bd->boot_write_complete);
+	}
+
+	if (bd->is_boot_open) {
+		mutex_unlock(&bd->mutex);
+		return;
+	}
+
+	if (bd->has_fifo_work) {
+		int len;
+
+		len = rshim_write_delayed(bd, bd->fifo_work_devtype,
+					  bd->fifo_work_buf,
+					  bd->fifo_work_buf_len);
+		bd->has_fifo_work = 0;
+
+		spin_lock(&bd->spinlock);
+		bd->spin_flags &= ~RSH_SFLG_WRITING;
+		if (len == bd->fifo_work_buf_len) {
+			wake_up_interruptible_all(&bd->write_completed);
+			rshim_notify(bd, RSH_EVENT_FIFO_OUTPUT, 0);
+		} else {
+			pr_err("fifo_write: completed abnormally.\n");
+			rshim_notify(bd, RSH_EVENT_FIFO_ERR, -1);
+		}
+		spin_unlock(&bd->spinlock);
+	}
+
+	if (bd->has_cons_work) {
+		spin_lock_irq(&bd->spinlock);
+
+		/* FIFO output. */
+		rshim_fifo_output(bd);
+
+		/* FIFO input. */
+		rshim_fifo_input(bd);
+
+		spin_unlock_irq(&bd->spinlock);
+
+		bd->has_cons_work = 0;
+	}
+
+	if (!bd->has_reprobe && bd->is_cons_open) {
+		bd->has_cons_work = 1;
+		mod_timer(&bd->timer, jiffies + HZ / 10);
+	}
+
+	mutex_unlock(&bd->mutex);
+}
+
+static ssize_t rshim_console_read(struct file *file, char *user_buffer,
+				    size_t count, loff_t *ppos)
+{
+	struct rshim_backend *bd = file->private_data;
+
+	return rshim_fifo_read(bd, user_buffer, count, TMFIFO_CONS_CHAN,
+			     file->f_flags & O_NONBLOCK, true);
+}
+
+static ssize_t rshim_console_write(struct file *file, const char *user_buffer,
+				 size_t count, loff_t *ppos)
+{
+	struct rshim_backend *bd = file->private_data;
+
+	return rshim_fifo_write(bd, user_buffer, count, TMFIFO_CONS_CHAN,
+			      file->f_flags & O_NONBLOCK, true);
+}
+
+static int rshim_console_fsync(struct file *file, loff_t start,
+			       loff_t end, int datasync)
+{
+	return rshim_fifo_fsync(file, start, end, datasync, TMFIFO_CONS_CHAN);
+}
+
+static long rshim_console_unlocked_ioctl(struct file *file, unsigned int
+				       cmd, unsigned long arg)
+{
+	struct rshim_backend *bd = file->private_data;
+	int retval = 0;
+
+	mutex_lock(&bd->mutex);
+
+	switch (cmd) {
+	case TCGETS: {
+#ifdef TCGETS2
+		if (kernel_termios_to_user_termios_1(
+			(struct termios __user *)arg, &bd->cons_termios))
+#else
+		if (kernel_termios_to_user_termios(
+			(struct termios __user *)arg, &bd->cons_termios))
+#endif
+			retval = -EFAULT;
+		break;
+	}
+
+	case TCSETS:
+	case TCSETSW:
+	case TCSETSF: {
+#ifdef TCGETS2
+		if (user_termios_to_kernel_termios_1(
+			&bd->cons_termios, (struct termios __user *)arg))
+#else
+		if (user_termios_to_kernel_termios(
+			&bd->cons_termios, (struct termios __user *)arg))
+#endif
+			retval = -EFAULT;
+		break;
+	}
+
+	default:
+		retval = -EINVAL;
+		break;
+	}
+
+	mutex_unlock(&bd->mutex);
+
+	return retval;
+}
+
+static unsigned int rshim_console_poll(struct file *file, poll_table *wait)
+{
+	return rshim_fifo_poll(file, wait, TMFIFO_CONS_CHAN);
+}
+
+static int rshim_console_release(struct inode *inode, struct file *file)
+{
+	return rshim_fifo_release(inode, file, TMFIFO_CONS_CHAN);
+}
+
+static const struct file_operations rshim_console_fops = {
+	.owner = THIS_MODULE,
+	.read = rshim_console_read,
+	.write = rshim_console_write,
+	.fsync = rshim_console_fsync,
+	.unlocked_ioctl = rshim_console_unlocked_ioctl,
+	.poll = rshim_console_poll,
+	.release = rshim_console_release,
+};
+
+static int rshim_console_open(struct file *file)
+{
+	struct rshim_backend *bd = file->private_data;
+
+	file->f_op = &rshim_console_fops;
+
+	mutex_lock(&bd->mutex);
+
+	if (bd->is_cons_open) {
+		/*
+		 * The console is already open.  This is OK, but it means
+		 * there's no work to do other than updating the reference
+		 * count.
+		 */
+		bd->console_opens++;
+		mutex_unlock(&bd->mutex);
+		return 0;
+	}
+
+	bd->is_cons_open = 1;
+
+	spin_lock_irq(&bd->spinlock);
+
+	bd->spin_flags |= RSH_SFLG_CONS_OPEN;
+
+	spin_unlock_irq(&bd->spinlock);
+
+	if (!bd->has_cons_work) {
+		bd->has_cons_work = 1;
+		queue_delayed_work(rshim_wq, &bd->work, HZ / 10);
+	}
+
+	bd->console_opens++;
+	mutex_unlock(&bd->mutex);
+
+	return 0;
+}
+
+static int rshim_boot_done(struct rshim_backend *bd)
+{
+	if (bd->has_rshim && bd->has_tm) {
+		/* Clear any previous errors. */
+		bd->tmfifo_error = 0;
+
+		/*
+		 * If someone might be waiting for the device to come up,
+		 * tell them it's ready.
+		 */
+		if (bd->is_booting) {
+			bd->is_booting = 0;
+
+			pr_debug("signaling booting complete\n");
+			complete_all(&bd->booting_complete);
+#if RSH_RESET_MUTEX
+			complete_all(&bd->reset_complete);
+#endif
+		};
+
+		/* If the console device is open, start the worker. */
+		if (bd->is_cons_open && !bd->has_cons_work) {
+			bd->has_cons_work = 1;
+			pr_debug("probe: console_work submitted\n");
+			queue_delayed_work(rshim_wq, &bd->work, 0);
+		}
+
+		/* Tell the user this device is now attached. */
+		pr_info("%s now attached\n", rshim_dev_names[bd->dev_index]);
+	}
+
+	return 0;
+}
+
+/* Rshim file operations routines */
+
+static ssize_t rshim_rshim_read(struct file *file, char *user_buffer,
+			      size_t count, loff_t *ppos)
+{
+	struct rshim_backend *bd;
+	int retval = 0;
+	u64 buf;
+
+	/* rshim registers are all 8-byte aligned. */
+	if (count != 8 || (*ppos & 7) != 0)
+		return -EINVAL;
+
+	bd = file->private_data;
+
+	mutex_lock(&bd->mutex);
+	retval = bd->read_rshim(bd,
+				(*ppos >> 16) & 0xF, /* channel # */
+				*ppos & 0xFFFF,	 /* addr */
+				&buf);
+	mutex_unlock(&bd->mutex);
+
+	/* If the read was successful, copy the data to userspace */
+	if (!retval && copy_to_user(user_buffer, &buf, count))
+		return -EFAULT;
+
+	return retval ? retval : count;
+}
+
+static ssize_t rshim_rshim_write(struct file *file, const char *user_buffer,
+			       size_t count, loff_t *ppos)
+{
+	struct rshim_backend *bd;
+	int retval = 0;
+	u64 buf;
+
+	/* rshim registers are all 8-byte aligned. */
+	if (count != 8 || (*ppos & 7) != 0)
+		return -EINVAL;
+
+	/* Copy the data from userspace */
+	if (copy_from_user(&buf, user_buffer, count))
+		return -EFAULT;
+
+	bd = file->private_data;
+
+	mutex_lock(&bd->mutex);
+	retval = bd->write_rshim(bd,
+				 (*ppos >> 16) & 0xF, /* channel # */
+				 *ppos & 0xFFFF, /* addr */
+				 buf);
+	mutex_unlock(&bd->mutex);
+
+	return retval ? retval : count;
+}
+
+static int rshim_rshim_release(struct inode *inode, struct file *file)
+{
+	struct rshim_backend *bd = file->private_data;
+	struct module *owner;
+
+	rshim_lock();
+	owner = RSHIM_READ_ONCE(bd->owner);
+	kref_put(&bd->kref, bd->destroy);
+	module_put(owner);
+	rshim_unlock();
+
+	return 0;
+}
+
+static const struct file_operations rshim_rshim_fops = {
+	.owner = THIS_MODULE,
+	.read = rshim_rshim_read,
+	.write = rshim_rshim_write,
+	.release = rshim_rshim_release,
+	.llseek = default_llseek,
+};
+
+static int rshim_rshim_open(struct file *file)
+{
+	file->f_op = &rshim_rshim_fops;
+
+	return 0;
+}
+
+/* Misc file operations routines */
+
+static int
+rshim_misc_seq_show(struct seq_file *s, void *token)
+{
+	struct rshim_backend *bd = s->private;
+	int retval;
+	u64 value;
+
+	/* Boot mode. */
+	retval = bd->read_rshim(bd, RSHIM_CHANNEL, RSH_BOOT_CONTROL,
+				&value);
+	if (retval) {
+		pr_err("couldn't read rshim register\n");
+		return retval;
+	}
+	seq_printf(s, "BOOT_MODE %lld\n",
+		   value & RSH_BOOT_CONTROL__BOOT_MODE_MASK);
+
+	/* SW reset flag is always 0. */
+	seq_printf(s, "SW_RESET  %d\n", 0);
+
+	/* Display the driver name. */
+	seq_printf(s, "DRV_NAME  %s\n", bd->owner->name);
+
+	return 0;
+}
+
+static ssize_t rshim_misc_write(struct file *file, const char *user_buffer,
+				size_t count, loff_t *ppos)
+{
+	struct rshim_backend *bd;
+	int retval = 0, value;
+	char buf[64], key[32];
+
+	if (*ppos != 0 || count >= sizeof(buf))
+		return -EINVAL;
+
+	/* Copy the data from userspace */
+	if (copy_from_user(buf, user_buffer, count))
+		return -EFAULT;
+
+	if (sscanf(buf, "%s %x", key, &value) != 2)
+		return -EINVAL;
+
+	bd = ((struct seq_file *)file->private_data)->private;
+
+	if (strcmp(key, "BOOT_MODE") == 0) {
+		retval = bd->write_rshim(bd, RSHIM_CHANNEL, RSH_BOOT_CONTROL,
+				 value & RSH_BOOT_CONTROL__BOOT_MODE_MASK);
+	} else if (strcmp(key, "SW_RESET") == 0) {
+		if (value) {
+			if (!bd->has_reprobe) {
+				/* Detach, which shouldn't hold bd->mutex. */
+				rshim_notify(bd, RSH_EVENT_DETACH, 0);
+
+				mutex_lock(&bd->mutex);
+				/* Reset the TmFifo. */
+				rshim_fifo_reset(bd);
+				mutex_unlock(&bd->mutex);
+			}
+
+			retval = bd->write_rshim(bd, RSHIM_CHANNEL,
+					RSH_RESET_CONTROL,
+					RSH_RESET_CONTROL__RESET_CHIP_VAL_KEY);
+
+			if (!bd->has_reprobe) {
+				/* Attach. */
+				msleep_interruptible(1000);
+				mutex_lock(&bd->mutex);
+				rshim_notify(bd, RSH_EVENT_ATTACH, 0);
+				mutex_unlock(&bd->mutex);
+			}
+		}
+	} else
+		return -EINVAL;
+
+	return retval ? retval : count;
+}
+
+static int rshim_misc_release(struct inode *inode, struct file *file)
+{
+	struct rshim_backend *bd;
+	struct module *owner;
+	int retval;
+
+	/*
+	 * Note that since this got turned into a seq file by
+	 * rshim_misc_open(), our device pointer isn't in the usual spot
+	 * (the file's private data); that's used by the seq file
+	 * subsystem.
+	 */
+	bd = ((struct seq_file *)file->private_data)->private;
+
+	retval = single_release(inode, file);
+	if (retval)
+		return retval;
+
+	rshim_lock();
+	owner = RSHIM_READ_ONCE(bd->owner);
+	kref_put(&bd->kref, bd->destroy);
+	module_put(owner);
+	rshim_unlock();
+
+	return 0;
+}
+
+static const struct file_operations rshim_misc_fops = {
+	.owner = THIS_MODULE,
+	.read = seq_read,
+	.llseek = seq_lseek,
+	.write = rshim_misc_write,
+	.release = rshim_misc_release,
+};
+
+static int rshim_misc_open(struct file *file)
+{
+	struct rshim_backend *bd = file->private_data;
+	int retval;
+
+	/*
+	 * If file->private_data is non-NULL, seq_open (called by
+	 * single_open) thinks it's already a seq_file struct, and
+	 * scribbles over it!  Very bad.
+	 */
+	file->private_data = NULL;
+
+	file->f_op = &rshim_misc_fops;
+	retval = single_open(file, rshim_misc_seq_show, bd);
+
+	return retval;
+}
+
+/* Common file operations routines */
+
+static int rshim_open(struct inode *inode, struct file *file)
+{
+	struct rshim_backend *bd;
+	int subminor = iminor(inode);
+	int retval;
+
+	rshim_lock();
+
+	bd = rshim_devs[subminor / RSH_DEV_TYPES];
+	if (!bd) {
+		rshim_unlock();
+		return -ENODEV;
+	}
+
+	/* Add a reference to the owner. */
+	if (!try_module_get(bd->owner)) {
+		rshim_unlock();
+		return -ENODEV;
+	}
+
+	/* Increment our usage count for the device. */
+	kref_get(&bd->kref);
+
+	rshim_unlock();
+
+	file->private_data = bd;
+
+	switch (subminor % RSH_DEV_TYPES) {
+	case RSH_DEV_TYPE_BOOT:
+		retval = rshim_boot_open(file);
+		break;
+
+	case RSH_DEV_TYPE_RSHIM:
+		retval = rshim_rshim_open(file);
+		break;
+
+	case RSH_DEV_TYPE_CONSOLE:
+		retval = rshim_console_open(file);
+		break;
+
+	case RSH_DEV_TYPE_NET:
+		retval = rshim_tmfifo_open(file);
+		break;
+
+	case RSH_DEV_TYPE_MISC:
+		retval = rshim_misc_open(file);
+		break;
+
+	default:
+		retval = -ENODEV;
+		break;
+	}
+
+	/* If the minor open failed, drop the usage count. */
+	if (retval < 0) {
+		struct module *owner;
+
+		rshim_lock();
+		owner = RSHIM_READ_ONCE(bd->owner);
+		kref_put(&bd->kref, bd->destroy);
+		module_put(owner);
+		rshim_unlock();
+	}
+
+	return retval;
+}
+
+static const struct file_operations rshim_fops = {
+	.owner = THIS_MODULE,
+	.open =	rshim_open,
+};
+
+int rshim_tmfifo_sync(struct rshim_backend *bd)
+{
+	u64 word;
+	int i, retval, max_size, avail;
+	union rshim_tmfifo_msg_hdr hdr;
+
+	/* Get FIFO max size. */
+	retval = bd->read_rshim(bd, RSHIM_CHANNEL,
+				RSH_TM_HOST_TO_TILE_CTL, &word);
+	if (retval < 0) {
+		pr_err("read_rshim error %d\n", retval);
+		return retval;
+	}
+	max_size = (word >> RSH_TM_HOST_TO_TILE_CTL__MAX_ENTRIES_SHIFT)
+		   & RSH_TM_HOST_TO_TILE_CTL__MAX_ENTRIES_RMASK;
+
+	/* Calculate available size. */
+	retval = bd->read_rshim(bd, RSHIM_CHANNEL, RSH_TM_HOST_TO_TILE_STS,
+				&word);
+	if (retval < 0) {
+		pr_err("read_rshim error %d\n", retval);
+		return retval;
+	}
+	avail = max_size - (int)(word & RSH_TM_HOST_TO_TILE_STS__COUNT_MASK);
+
+	if (avail > TMFIFO_MAX_SYNC_WORDS)
+		avail = TMFIFO_MAX_SYNC_WORDS;
+
+	hdr.type = VIRTIO_ID_NET;
+	hdr.len = 0;
+	for (i = 0; i < avail; i++) {
+		retval = bd->write_rshim(bd, RSHIM_CHANNEL,
+					 RSH_TM_HOST_TO_TILE_STS, hdr.data);
+		if (retval < 0)
+			break;
+	}
+
+	return 0;
+}
+
+int rshim_notify(struct rshim_backend *bd, int event, int code)
+{
+	int i, rc = 0;
+	struct rshim_service *svc;
+
+	switch (event) {
+	case RSH_EVENT_FIFO_INPUT:
+		rshim_fifo_input(bd);
+		break;
+
+	case RSH_EVENT_FIFO_OUTPUT:
+		rshim_fifo_output(bd);
+		break;
+
+	case RSH_EVENT_FIFO_ERR:
+		rshim_fifo_err(bd, code);
+		break;
+
+	case RSH_EVENT_ATTACH:
+		rshim_boot_done(bd);
+
+		/* Sync-up the tmfifo if reprobe is not supported. */
+		if (!bd->has_reprobe && bd->has_rshim)
+			rshim_tmfifo_sync(bd);
+
+		rcu_read_lock();
+		for (i = 0; i < RSH_SVC_MAX; i++) {
+			svc = rcu_dereference(rshim_svc[i]);
+			if (svc != NULL && svc->create != NULL) {
+				rc = (*svc->create)(bd);
+				if (rc == -EEXIST)
+					rc = 0;
+				else if (rc) {
+					pr_err("Failed to attach svc %d\n", i);
+					break;
+				}
+			}
+		}
+		rcu_read_unlock();
+
+		spin_lock_irq(&bd->spinlock);
+		rshim_fifo_input(bd);
+		spin_unlock_irq(&bd->spinlock);
+		break;
+
+	case RSH_EVENT_DETACH:
+		for (i = 0; i < RSH_SVC_MAX; i++) {
+			/*
+			 * The svc->delete() could call into Linux kernel and
+			 * potentially trigger synchronize_rcu(). So it should
+			 * be outside of the rcu_read_lock(). Instead, a ref
+			 * counter is used here to avoid race condition between
+			 * svc deletion such as caused by kernel module unload.
+			 */
+			rcu_read_lock();
+			svc = rcu_dereference(rshim_svc[i]);
+			if (svc != NULL)
+				atomic_inc(&svc->ref);
+			rcu_read_unlock();
+
+			if (svc != NULL) {
+				(*svc->delete)(bd);
+				atomic_dec(&svc->ref);
+			}
+		}
+		bd->dev = NULL;
+		break;
+	}
+
+	return rc;
+}
+EXPORT_SYMBOL(rshim_notify);
+
+static int rshim_find_index(char *dev_name)
+{
+	int i, dev_index = -1;
+
+	/* First look for a match with a previous device name. */
+	for (i = 0; i < rshim_nr_devs; i++)
+		if (rshim_dev_names[i] &&
+		    !strcmp(dev_name, rshim_dev_names[i])) {
+			pr_debug("found match with previous at index %d\n", i);
+			dev_index = i;
+			break;
+		}
+
+	/* Then look for a never-used slot. */
+	if (dev_index < 0) {
+		for (i = 0; i < rshim_nr_devs; i++)
+			if (!rshim_dev_names[i]) {
+				pr_debug("found never-used slot %d\n", i);
+				dev_index = i;
+				break;
+			}
+	}
+
+	/* Finally look for a currently-unused slot. */
+	if (dev_index < 0) {
+		for (i = 0; i < rshim_nr_devs; i++)
+			if (!rshim_devs[i]) {
+				pr_debug("found unused slot %d\n", i);
+				dev_index = i;
+				break;
+			}
+	}
+
+	return dev_index;
+}
+
+struct rshim_backend *rshim_find(char *dev_name)
+{
+	int dev_index = rshim_find_index(dev_name);
+
+	/* If none of that worked, we fail. */
+	if (dev_index < 0) {
+		pr_err("couldn't find slot for new device %s\n", dev_name);
+		return NULL;
+	}
+
+	return rshim_devs[dev_index];
+}
+EXPORT_SYMBOL(rshim_find);
+
+/* House-keeping timer. */
+static void rshim_timer_func(struct timer_list *arg)
+{
+	struct rshim_backend *bd =
+	  container_of(arg, struct rshim_backend, timer);
+
+	u32 period = msecs_to_jiffies(rshim_keepalive_period);
+
+	if (bd->has_cons_work)
+		queue_delayed_work(rshim_wq, &bd->work, 0);
+
+	/* Request keepalive update and restart the ~300ms timer. */
+	if (time_after(jiffies, (unsigned long)bd->last_keepalive + period)) {
+		bd->keepalive = 1;
+		bd->last_keepalive = jiffies;
+		queue_delayed_work(rshim_wq, &bd->work, 0);
+	}
+	mod_timer(&bd->timer, jiffies + period);
+}
+
+static ssize_t rshim_path_show(struct device *cdev,
+			       struct device_attribute *attr, char *buf)
+{
+	struct rshim_backend *bd = dev_get_drvdata(cdev);
+
+	if (bd == NULL)
+		return -ENODEV;
+	return snprintf(buf, PAGE_SIZE, "%s\n",
+			rshim_dev_names[bd->dev_index]);
+}
+
+static DEVICE_ATTR(rshim_path, 0444, rshim_path_show, NULL);
+
+static void
+rshim_load_modules(struct work_struct *work)
+{
+	request_module("rshim_net");
+}
+
+static DECLARE_DELAYED_WORK(rshim_load_modules_work, rshim_load_modules);
+
+/* Check whether backend is allowed to register or not. */
+static int rshim_access_check(struct rshim_backend *bd)
+{
+	int i, retval;
+	u64 value;
+
+	/* Write value 0 to RSH_SCRATCHPAD1. */
+	retval = bd->write_rshim(bd, RSHIM_CHANNEL, RSH_SCRATCHPAD1, 0);
+	if (retval < 0)
+		return -ENODEV;
+
+	/*
+	 * Poll RSH_SCRATCHPAD1 up to one second to check whether it's reset to
+	 * the keepalive magic value, which indicates another backend driver has
+	 * already attached to this target.
+	 */
+	for (i = 0; i < 10; i++) {
+		retval = bd->read_rshim(bd, RSHIM_CHANNEL, RSH_SCRATCHPAD1,
+					&value);
+		if (retval < 0)
+			return -ENODEV;
+
+		if (value == RSH_KEEPALIVE_MAGIC_NUM) {
+			pr_info("another backend already attached.\n");
+			return -EEXIST;
+		}
+
+		msleep(100);
+	}
+
+	return 0;
+}
+
+int rshim_register(struct rshim_backend *bd)
+{
+	int i, retval, dev_index;
+
+	if (bd->registered)
+		return 0;
+
+	if (backend_driver[0] && strcmp(backend_driver, bd->owner->name))
+		return -EACCES;
+
+	dev_index = rshim_find_index(bd->dev_name);
+	if (dev_index < 0)
+		return -ENODEV;
+
+	if (!bd->read_rshim || !bd->write_rshim) {
+		pr_err("read_rshim/write_rshim missing\n");
+		return -EINVAL;
+	}
+
+	retval = rshim_access_check(bd);
+	if (retval)
+		return retval;
+
+	if (!bd->write)
+		bd->write = rshim_write_default;
+	if (!bd->read)
+		bd->read = rshim_read_default;
+
+	kref_init(&bd->kref);
+	spin_lock_init(&bd->spinlock);
+#if RSH_RESET_MUTEX
+	init_completion(&bd->reset_complete);
+#endif
+	for (i = 0; i < TMFIFO_MAX_CHAN; i++) {
+		init_waitqueue_head(&bd->read_fifo[i].operable);
+		init_waitqueue_head(&bd->write_fifo[i].operable);
+	}
+
+	init_waitqueue_head(&bd->write_completed);
+	init_completion(&bd->booting_complete);
+	init_completion(&bd->boot_write_complete);
+	memcpy(&bd->cons_termios, &init_console_termios,
+	       sizeof(init_console_termios));
+	INIT_DELAYED_WORK(&bd->work, rshim_work_handler);
+
+	bd->dev_index = dev_index;
+	if (rshim_dev_names[dev_index] != bd->dev_name) {
+		kfree(rshim_dev_names[dev_index]);
+		rshim_dev_names[dev_index] = bd->dev_name;
+	}
+	rshim_devs[dev_index] = bd;
+
+	for (i = 0; i < RSH_DEV_TYPES; i++) {
+		struct device *cl_dev;
+		int err;
+		char devbuf[32];
+
+		cdev_init(&bd->cdevs[i], &rshim_fops);
+		bd->cdevs[i].owner = THIS_MODULE;
+		/*
+		 * FIXME: is this addition really legal, or should
+		 * we be using MKDEV?
+		 */
+		err = cdev_add(&bd->cdevs[i],
+			       rshim_dev_base +
+			       bd->dev_index * RSH_DEV_TYPES + i,
+			       1);
+		/*
+		 * We complain if this fails, but we don't return
+		 * an error; it really shouldn't happen, and it's
+		 * hard to go un-do the rest of the adds.
+		 */
+		if (err)
+			pr_err("rsh%d: couldn't add minor %d\n", dev_index, i);
+
+		cl_dev = device_create(rshim_class, NULL, rshim_dev_base +
+				       bd->dev_index * RSH_DEV_TYPES + i, NULL,
+				       "rshim%d!%s",
+				       bd->dev_index, rshim_dev_minor_names[i]);
+		if (IS_ERR(cl_dev)) {
+			pr_err("rsh%d: couldn't add dev %s, err %ld\n",
+			       dev_index,
+			       format_dev_t(devbuf, rshim_dev_base + dev_index *
+					    RSH_DEV_TYPES + i),
+			       PTR_ERR(cl_dev));
+		} else {
+			pr_debug("added class dev %s\n",
+				 format_dev_t(devbuf, rshim_dev_base +
+					      bd->dev_index *
+					      RSH_DEV_TYPES + i));
+		}
+
+		dev_set_drvdata(cl_dev, bd);
+		if (device_create_file(cl_dev, &dev_attr_rshim_path))
+			pr_err("could not create rshim_path file in sysfs\n");
+	}
+
+	for (i = 0; i < 2; i++) {
+		bd->boot_buf[i] = kmalloc(BOOT_BUF_SIZE, GFP_KERNEL);
+		if (!bd->boot_buf[i]) {
+			if (i == 1) {
+				kfree(bd->boot_buf[0]);
+				bd->boot_buf[0] = NULL;
+			}
+		}
+	}
+
+	timer_setup(&bd->timer, rshim_timer_func, 0);
+
+	bd->registered = 1;
+
+	/* Start the keepalive timer. */
+	bd->last_keepalive = jiffies;
+	mod_timer(&bd->timer, jiffies + 1);
+
+	schedule_delayed_work(&rshim_load_modules_work, 3 * HZ);
+
+	return 0;
+}
+EXPORT_SYMBOL(rshim_register);
+
+void rshim_deregister(struct rshim_backend *bd)
+{
+	int i;
+
+	if (!bd->registered)
+		return;
+
+	/* Stop the timer. */
+	del_timer_sync(&bd->timer);
+
+	for (i = 0; i < 2; i++)
+		kfree(bd->boot_buf[i]);
+
+	for (i = 0; i < RSH_DEV_TYPES; i++) {
+		cdev_del(&bd->cdevs[i]);
+		device_destroy(rshim_class,
+			       rshim_dev_base + bd->dev_index *
+			       RSH_DEV_TYPES + i);
+	}
+
+	rshim_devs[bd->dev_index] = NULL;
+	bd->registered = 0;
+}
+EXPORT_SYMBOL(rshim_deregister);
+
+int rshim_register_service(struct rshim_service *service)
+{
+	int i, retval = 0;
+	struct rshim_service *svc;
+
+	rshim_lock();
+
+	atomic_set(&service->ref, 0);
+
+	BUG_ON(service->type >= RSH_SVC_MAX);
+
+	if (!rshim_svc[service->type]) {
+		svc = kmalloc(sizeof(*svc), GFP_KERNEL);
+		if (svc) {
+			memcpy(svc, service, sizeof(*svc));
+			/*
+			 * Add memory barrir to make sure 'svc' is ready
+			 * before switching the pointer.
+			 */
+			smp_mb();
+
+			/*
+			 * rshim_svc[] is protected by RCU. References to it
+			 * should have rcu_read_lock() / rcu_dereference() /
+			 * rcu_read_lock().
+			 */
+			rcu_assign_pointer(rshim_svc[service->type], svc);
+
+			/* Attach the service to all backends. */
+			for (i = 0; i < rshim_nr_devs; i++) {
+				if (rshim_devs[i] != NULL) {
+					retval = svc->create(rshim_devs[i]);
+					if (retval && retval != -EEXIST)
+						break;
+				}
+			}
+		} else
+			retval = -ENOMEM;
+	} else
+		retval = -EEXIST;
+
+	rshim_unlock();
+
+	/* Deregister / cleanup the service in case of failures. */
+	if (retval && retval != -EEXIST)
+		rshim_deregister_service(service);
+
+	return retval;
+}
+EXPORT_SYMBOL(rshim_register_service);
+
+void rshim_deregister_service(struct rshim_service *service)
+{
+	int i;
+	struct rshim_service *svc = NULL;
+
+	BUG_ON(service->type >= RSH_SVC_MAX);
+
+	/*
+	 * Use synchronize_rcu() to make sure no more outstanding
+	 * references to the 'svc' pointer before releasing it.
+	 *
+	 * The reason to use RCU is that the rshim_svc pointer will be
+	 * accessed in rshim_notify() which could be called in interrupt
+	 * context and not suitable for mutex lock.
+	 */
+	rshim_lock();
+	if (rshim_svc[service->type]) {
+		svc = rshim_svc[service->type];
+
+		/* Delete the service from all backends. */
+		for (i = 0; i < rshim_nr_devs; i++)
+			if (rshim_devs[i] != NULL)
+				svc->delete(rshim_devs[i]);
+
+		rcu_assign_pointer(rshim_svc[service->type], NULL);
+	}
+	rshim_unlock();
+	if (svc != NULL) {
+		synchronize_rcu();
+
+		/* Make sure no more references to the svc pointer. */
+		while (atomic_read(&svc->ref) != 0)
+			msleep(100);
+		kfree(svc);
+	}
+}
+EXPORT_SYMBOL(rshim_deregister_service);
+
+static int __init rshim_init(void)
+{
+	int result, class_registered = 0;
+
+	/* Register our device class. */
+	rshim_class = class_create(THIS_MODULE, "rsh");
+	if (IS_ERR(rshim_class)) {
+		result = PTR_ERR(rshim_class);
+		goto error;
+	}
+	class_registered = 1;
+
+	/* Allocate major/minor numbers. */
+	result = alloc_chrdev_region(&rshim_dev_base, 0,
+				     rshim_nr_devs * RSH_DEV_TYPES,
+				     "rsh");
+	if (result < 0) {
+		pr_err("can't get rshim major\n");
+		goto error;
+	}
+
+	rshim_dev_names = kzalloc(rshim_nr_devs *
+				    sizeof(rshim_dev_names[0]), GFP_KERNEL);
+	rshim_devs = kcalloc(rshim_nr_devs, sizeof(rshim_devs[0]),
+			       GFP_KERNEL);
+
+	if (!rshim_dev_names || !rshim_devs) {
+		result = -ENOMEM;
+		goto error;
+	}
+
+	rshim_wq = create_workqueue("rshim");
+	if (!rshim_wq) {
+		result = -ENOMEM;
+		goto error;
+	}
+
+	return 0;
+
+error:
+	if (rshim_dev_base)
+		unregister_chrdev_region(rshim_dev_base,
+				 rshim_nr_devs * RSH_DEV_TYPES);
+	if (class_registered)
+		class_destroy(rshim_class);
+	kfree(rshim_dev_names);
+	kfree(rshim_devs);
+
+	return result;
+}
+
+static void __exit rshim_exit(void)
+{
+	int i;
+
+	flush_delayed_work(&rshim_load_modules_work);
+
+	/* Free the major/minor numbers. */
+	unregister_chrdev_region(rshim_dev_base,
+				 rshim_nr_devs * RSH_DEV_TYPES);
+
+	/* Destroy our device class. */
+	class_destroy(rshim_class);
+
+	/* Destroy our work queue. */
+	destroy_workqueue(rshim_wq);
+
+	for (i = 0; i < RSH_SVC_MAX; i++)
+		kfree(rshim_svc[i]);
+
+	for (i = 0; i < rshim_nr_devs; i++)
+		kfree(rshim_dev_names[i]);
+
+	kfree(rshim_dev_names);
+	kfree(rshim_devs);
+}
+
+module_init(rshim_init);
+module_exit(rshim_exit);
+
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("Mellanox Technologies");
+MODULE_VERSION("0.12");
diff --git a/drivers/soc/mellanox/host/rshim.h b/drivers/soc/mellanox/host/rshim.h
new file mode 100644
index 0000000..3ac3410
--- /dev/null
+++ b/drivers/soc/mellanox/host/rshim.h
@@ -0,0 +1,361 @@ 
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright 2017 Mellanox Technologies. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _RSHIM_H
+#define _RSHIM_H
+
+#include <linux/kernel.h>
+#include <linux/kref.h>
+#include <linux/module.h>
+#include <linux/mutex.h>
+#include <linux/termios.h>
+#include <linux/workqueue.h>
+#include <linux/device.h>
+#include <linux/cdev.h>
+
+#include "rshim_regs.h"
+
+/* ACCESS_ONCE() wrapper. */
+#define RSHIM_READ_ONCE(x)	READ_ONCE(x)
+
+/*
+ * This forces only one reset to occur at a time.  Once we've gotten
+ * more experience with this mode we'll probably remove the #define.
+ */
+#define RSH_RESET_MUTEX		1
+
+/* Spin flag values. */
+#define RSH_SFLG_READING	0x1  /* read is active. */
+#define RSH_SFLG_WRITING	0x2  /* write_urb is active. */
+#define RSH_SFLG_CONS_OPEN	0x4  /* console stream is open. */
+
+/*
+ * Buffer/FIFO sizes.  Note that the FIFO sizes must be powers of 2; also,
+ * the read and write buffers must be no larger than the corresponding
+ * FIFOs.
+ */
+#define READ_BUF_SIZE		2048
+#define WRITE_BUF_SIZE		2048
+#define READ_FIFO_SIZE		(4 * 1024)
+#define WRITE_FIFO_SIZE		(4 * 1024)
+#define BOOT_BUF_SIZE		(16 * 1024)
+
+/* Sub-device types. */
+enum {
+	RSH_DEV_TYPE_RSHIM,
+	RSH_DEV_TYPE_BOOT,
+	RSH_DEV_TYPE_CONSOLE,
+	RSH_DEV_TYPE_NET,
+	RSH_DEV_TYPE_MISC,
+	RSH_DEV_TYPES
+};
+
+/* Event types used in rshim_notify(). */
+enum {
+	RSH_EVENT_FIFO_INPUT,		/* fifo ready for input */
+	RSH_EVENT_FIFO_OUTPUT,		/* fifo ready for output */
+	RSH_EVENT_FIFO_ERR,		/* fifo error */
+	RSH_EVENT_ATTACH,		/* backend attaching */
+	RSH_EVENT_DETACH,		/* backend detaching */
+};
+
+/* RShim service types. */
+enum {
+	RSH_SVC_NET,			/* networking service */
+	RSH_SVC_MAX
+};
+
+/* TMFIFO message header. */
+union rshim_tmfifo_msg_hdr {
+	struct {
+		u8 type;		/* message type */
+		__be16 len;		/* payload length */
+		u8 unused[5];		/* reserved, set to 0 */
+	} __packed;
+	u64 data;
+};
+
+/* TMFIFO demux channels. */
+enum {
+	TMFIFO_CONS_CHAN,	/* Console */
+	TMFIFO_NET_CHAN,	/* Network */
+	TMFIFO_MAX_CHAN		/* Number of channels */
+};
+
+/* Various rshim definitions. */
+#define RSH_INT_VEC0_RTC__SWINT3_MASK 0x8
+
+#define RSH_BYTE_ACC_READ_TRIGGER 0x50000000
+#define RSH_BYTE_ACC_SIZE 0x10000000
+#define RSH_BYTE_ACC_PENDING 0x20000000
+
+
+#define BOOT_CHANNEL        RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_BOOT
+#define RSHIM_CHANNEL       RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_RSHIM
+#define UART0_CHANNEL       RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_UART0
+#define UART1_CHANNEL       RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_UART1
+
+#define RSH_BOOT_FIFO_SIZE   512
+
+/* FIFO structure. */
+struct rshim_fifo {
+	unsigned char *data;
+	unsigned int head;
+	unsigned int tail;
+	wait_queue_head_t operable;
+};
+
+/* RShim backend. */
+struct rshim_backend {
+	/* Device name. */
+	char *dev_name;
+
+	/* Backend owner. */
+	struct module *owner;
+
+	/* Pointer to the backend device. */
+	struct device *dev;
+
+	/* Pointer to the net device. */
+	void *net;
+
+	/* House-keeping Timer. */
+	struct timer_list timer;
+
+	/* Character device structure for each device. */
+	struct cdev cdevs[RSH_DEV_TYPES];
+
+	/*
+	 * The reference count for this structure.  This is incremented by
+	 * each open, and by the probe routine (thus, one reference for
+	 * each of the two interfaces).  It's decremented on each release,
+	 * and on each disconnect.
+	 */
+	struct kref kref;
+
+	/* State flags. */
+	u32 is_booting : 1;        /* Waiting for device to come back. */
+	u32 is_boot_open : 1;      /* Boot device is open. */
+	u32 is_tm_open : 1;        /* TM FIFO device is open. */
+	u32 is_cons_open : 1;      /* Console device is open. */
+	u32 is_in_boot_write : 1;  /* A thread is in boot_write(). */
+	u32 has_cons_work : 1;     /* Console worker thread running. */
+	u32 has_debug : 1;         /* Debug enabled for this device. */
+	u32 has_tm : 1;            /* TM FIFO found. */
+	u32 has_rshim : 1;         /* RSHIM found. */
+	u32 has_fifo_work : 1;     /* FIFO output to be done in worker. */
+	u32 has_reprobe : 1;       /* Reprobe support after SW reset. */
+	u32 drop : 1;              /* Drop the rest of the packet. */
+	u32 registered : 1;        /* Backend has been registered. */
+	u32 keepalive : 1;         /* A flag to update keepalive. */
+
+	/* Jiffies of last keepalive. */
+	u64 last_keepalive;
+
+	/* State flag bits from RSH_SFLG_xxx (see above). */
+	int spin_flags;
+
+	/* Total bytes in the read buffer. */
+	int read_buf_bytes;
+	/* Offset of next unread byte in the read buffer. */
+	int read_buf_next;
+	/* Bytes left in the current packet, or 0 if no current packet. */
+	int read_buf_pkt_rem;
+	/* Padded bytes in the read buffer. */
+	int read_buf_pkt_padding;
+
+	/* Bytes left in the current packet pending to write. */
+	int write_buf_pkt_rem;
+
+	/* Current message header. */
+	union rshim_tmfifo_msg_hdr msg_hdr;
+
+	/* Read FIFOs. */
+	struct rshim_fifo read_fifo[TMFIFO_MAX_CHAN];
+
+	/* Write FIFOs. */
+	struct rshim_fifo write_fifo[TMFIFO_MAX_CHAN];
+
+	/* Read buffer.  This is a DMA'able buffer. */
+	unsigned char *read_buf;
+	dma_addr_t read_buf_dma;
+
+	/* Write buffer.  This is a DMA'able buffer. */
+	unsigned char *write_buf;
+	dma_addr_t write_buf_dma;
+
+	/* Current Tx FIFO channel. */
+	int tx_chan;
+
+	/* Current Rx FIFO channel. */
+	int rx_chan;
+
+	/* First error encountered during read or write. */
+	int tmfifo_error;
+
+	/* Buffers used for boot writes.  Allocated at startup. */
+	char *boot_buf[2];
+
+	/*
+	 * This mutex is used to prevent the interface pointers and the
+	 * device pointer from disappearing while a driver entry point
+	 * is using them.  It's held throughout a read or write operation
+	 * (at least the parts of those operations which depend upon those
+	 * pointers) and is also held whenever those pointers are modified.
+	 * It also protects state flags, and booting_complete.
+	 */
+	struct mutex mutex;
+
+	/* We'll signal completion on this when FLG_BOOTING is turned off. */
+	struct completion booting_complete;
+
+#ifdef RSH_RESET_MUTEX
+	/* Signaled when a device is disconnected. */
+	struct completion reset_complete;
+#endif
+
+	/*
+	 * This wait queue supports fsync; it's woken up whenever an
+	 * outstanding USB write URB is done.  This will need to be more
+	 * complex if we start doing write double-buffering.
+	 */
+	wait_queue_head_t write_completed;
+
+	/* State for our outstanding boot write. */
+	struct completion boot_write_complete;
+
+	/*
+	 * This spinlock is used to protect items which must be updated by
+	 * URB completion handlers, since those can't sleep.  This includes
+	 * the read and write buffer pointers, as well as spin_flags.
+	 */
+	spinlock_t spinlock;
+
+	/* Current termios settings for the console. */
+	struct ktermios cons_termios;
+
+	/* Work queue entry. */
+	struct delayed_work	work;
+
+	/* Pending boot & fifo request for the worker. */
+	u8 *boot_work_buf;
+	u32 boot_work_buf_len;
+	u32 boot_work_buf_actual_len;
+	u8 *fifo_work_buf;
+	u32 fifo_work_buf_len;
+	int fifo_work_devtype;
+
+	/* Number of open console files. */
+	long console_opens;
+
+	/*
+	 * Our index in rshim_devs, which is also the high bits of our
+	 * minor number.
+	 */
+	int dev_index;
+
+	/* APIs provided by backend. */
+
+	/* API to write bulk data to RShim via the backend. */
+	ssize_t (*write)(struct rshim_backend *bd, int devtype,
+			 const char *buf, size_t count);
+
+	/* API to read bulk data from RShim via the backend. */
+	ssize_t (*read)(struct rshim_backend *bd, int devtype,
+			char *buf, size_t count);
+
+	/* API to cancel a read / write request (optional). */
+	void (*cancel)(struct rshim_backend *bd, int devtype, bool is_write);
+
+	/* API to destroy the backend. */
+	void (*destroy)(struct kref *kref);
+
+	/* API to read 8 bytes from RShim. */
+	int (*read_rshim)(struct rshim_backend *bd, int chan, int addr,
+			  u64 *value);
+
+	/* API to write 8 bytes to RShim. */
+	int (*write_rshim)(struct rshim_backend *bd, int chan, int addr,
+			   u64 value);
+};
+
+/* RShim service. */
+struct rshim_service {
+	/* Service type RSH_SVC_xxx. */
+	int type;
+
+	/* Reference number. */
+	atomic_t ref;
+
+	/* Create service. */
+	int (*create)(struct rshim_backend *bd);
+
+	/* Delete service. */
+	int (*delete)(struct rshim_backend *bd);
+
+	/* Notify service Rx is ready. */
+	void (*rx_notify)(struct rshim_backend *bd);
+};
+
+/* Global variables. */
+
+/* Global array to store RShim devices and names. */
+extern struct workqueue_struct *rshim_wq;
+
+/* Common APIs. */
+
+/* Register/unregister backend. */
+int rshim_register(struct rshim_backend *bd);
+void rshim_deregister(struct rshim_backend *bd);
+
+/* Register / deregister service. */
+int rshim_register_service(struct rshim_service *service);
+void rshim_deregister_service(struct rshim_service *service);
+
+/* Find backend by name. */
+struct rshim_backend *rshim_find(char *dev_name);
+
+/* RShim global lock. */
+void rshim_lock(void);
+void rshim_unlock(void);
+
+/* Event notification. */
+int rshim_notify(struct rshim_backend *bd, int event, int code);
+
+/*
+ * FIFO APIs.
+ *
+ * FIFO is demuxed into two channels, one for network interface
+ * (TMFIFO_NET_CHAN), one for console (TMFIFO_CONS_CHAN).
+ */
+
+/* Write / read some bytes to / from the FIFO via the backend. */
+ssize_t rshim_fifo_read(struct rshim_backend *bd, char *buffer,
+		      size_t count, int chan, bool nonblock,
+		      bool to_user);
+ssize_t rshim_fifo_write(struct rshim_backend *bd, const char *buffer,
+		       size_t count, int chan, bool nonblock,
+		       bool from_user);
+
+/* Alloc/free the FIFO. */
+int rshim_fifo_alloc(struct rshim_backend *bd);
+void rshim_fifo_free(struct rshim_backend *bd);
+
+/* Console APIs. */
+
+/* Enable early console. */
+int rshim_cons_early_enable(struct rshim_backend *bd);
+
+#endif /* _RSHIM_H */
diff --git a/drivers/soc/mellanox/host/rshim_regs.h b/drivers/soc/mellanox/host/rshim_regs.h
new file mode 100644
index 0000000..b14df716
--- /dev/null
+++ b/drivers/soc/mellanox/host/rshim_regs.h
@@ -0,0 +1,152 @@ 
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (c) 2018, Mellanox Technologies. All rights reserved.
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ */
+#ifndef __RSHIM_REGS_H__
+#define __RSHIM_REGS_H__
+
+#ifdef __ASSEMBLER__
+#define _64bit(x) x
+#else /* __ASSEMBLER__ */
+#define _64bit(x) x ## ULL
+#endif /* __ASSEMBLER */
+
+#include <linux/types.h>
+
+#define RSH_BOOT_FIFO_DATA 0x408
+
+#define RSH_BOOT_FIFO_COUNT 0x488
+#define RSH_BOOT_FIFO_COUNT__LENGTH 0x0001
+#define RSH_BOOT_FIFO_COUNT__BOOT_FIFO_COUNT_SHIFT 0
+#define RSH_BOOT_FIFO_COUNT__BOOT_FIFO_COUNT_WIDTH 10
+#define RSH_BOOT_FIFO_COUNT__BOOT_FIFO_COUNT_RESET_VAL 0
+#define RSH_BOOT_FIFO_COUNT__BOOT_FIFO_COUNT_RMASK 0x3ff
+#define RSH_BOOT_FIFO_COUNT__BOOT_FIFO_COUNT_MASK  0x3ff
+
+#define RSH_BOOT_CONTROL 0x528
+#define RSH_BOOT_CONTROL__LENGTH 0x0001
+#define RSH_BOOT_CONTROL__BOOT_MODE_SHIFT 0
+#define RSH_BOOT_CONTROL__BOOT_MODE_WIDTH 2
+#define RSH_BOOT_CONTROL__BOOT_MODE_RESET_VAL 0
+#define RSH_BOOT_CONTROL__BOOT_MODE_RMASK 0x3
+#define RSH_BOOT_CONTROL__BOOT_MODE_MASK  0x3
+#define RSH_BOOT_CONTROL__BOOT_MODE_VAL_NONE 0x0
+#define RSH_BOOT_CONTROL__BOOT_MODE_VAL_EMMC 0x1
+#define RSH_BOOT_CONTROL__BOOT_MODE_VAL_EMMC_LEGACY 0x3
+
+#define RSH_RESET_CONTROL 0x500
+#define RSH_RESET_CONTROL__LENGTH 0x0001
+#define RSH_RESET_CONTROL__RESET_CHIP_SHIFT 0
+#define RSH_RESET_CONTROL__RESET_CHIP_WIDTH 32
+#define RSH_RESET_CONTROL__RESET_CHIP_RESET_VAL 0
+#define RSH_RESET_CONTROL__RESET_CHIP_RMASK 0xffffffff
+#define RSH_RESET_CONTROL__RESET_CHIP_MASK  0xffffffff
+#define RSH_RESET_CONTROL__RESET_CHIP_VAL_KEY 0xca710001
+#define RSH_RESET_CONTROL__DISABLE_SHIFT 32
+#define RSH_RESET_CONTROL__DISABLE_WIDTH 1
+#define RSH_RESET_CONTROL__DISABLE_RESET_VAL 0
+#define RSH_RESET_CONTROL__DISABLE_RMASK 0x1
+#define RSH_RESET_CONTROL__DISABLE_MASK  _64bit(0x100000000)
+#define RSH_RESET_CONTROL__REQ_PND_SHIFT 33
+#define RSH_RESET_CONTROL__REQ_PND_WIDTH 1
+#define RSH_RESET_CONTROL__REQ_PND_RESET_VAL 0
+#define RSH_RESET_CONTROL__REQ_PND_RMASK 0x1
+#define RSH_RESET_CONTROL__REQ_PND_MASK  _64bit(0x200000000)
+
+#define RSH_SCRATCHPAD1 0xc20
+
+#define RSH_SCRATCH_BUF_CTL 0x600
+
+#define RSH_SCRATCH_BUF_DAT 0x610
+
+#define RSH_SEMAPHORE0 0x28
+
+#define RSH_SCRATCHPAD 0x20
+
+#define RSH_TM_HOST_TO_TILE_CTL 0xa30
+#define RSH_TM_HOST_TO_TILE_CTL__LENGTH 0x0001
+#define RSH_TM_HOST_TO_TILE_CTL__LWM_SHIFT 0
+#define RSH_TM_HOST_TO_TILE_CTL__LWM_WIDTH 8
+#define RSH_TM_HOST_TO_TILE_CTL__LWM_RESET_VAL 128
+#define RSH_TM_HOST_TO_TILE_CTL__LWM_RMASK 0xff
+#define RSH_TM_HOST_TO_TILE_CTL__LWM_MASK  0xff
+#define RSH_TM_HOST_TO_TILE_CTL__HWM_SHIFT 8
+#define RSH_TM_HOST_TO_TILE_CTL__HWM_WIDTH 8
+#define RSH_TM_HOST_TO_TILE_CTL__HWM_RESET_VAL 128
+#define RSH_TM_HOST_TO_TILE_CTL__HWM_RMASK 0xff
+#define RSH_TM_HOST_TO_TILE_CTL__HWM_MASK  0xff00
+#define RSH_TM_HOST_TO_TILE_CTL__MAX_ENTRIES_SHIFT 32
+#define RSH_TM_HOST_TO_TILE_CTL__MAX_ENTRIES_WIDTH 9
+#define RSH_TM_HOST_TO_TILE_CTL__MAX_ENTRIES_RESET_VAL 256
+#define RSH_TM_HOST_TO_TILE_CTL__MAX_ENTRIES_RMASK 0x1ff
+#define RSH_TM_HOST_TO_TILE_CTL__MAX_ENTRIES_MASK  _64bit(0x1ff00000000)
+
+#define RSH_TM_HOST_TO_TILE_STS 0xa28
+#define RSH_TM_HOST_TO_TILE_STS__LENGTH 0x0001
+#define RSH_TM_HOST_TO_TILE_STS__COUNT_SHIFT 0
+#define RSH_TM_HOST_TO_TILE_STS__COUNT_WIDTH 9
+#define RSH_TM_HOST_TO_TILE_STS__COUNT_RESET_VAL 0
+#define RSH_TM_HOST_TO_TILE_STS__COUNT_RMASK 0x1ff
+#define RSH_TM_HOST_TO_TILE_STS__COUNT_MASK  0x1ff
+
+#define RSH_TM_TILE_TO_HOST_STS 0xa48
+#define RSH_TM_TILE_TO_HOST_STS__LENGTH 0x0001
+#define RSH_TM_TILE_TO_HOST_STS__COUNT_SHIFT 0
+#define RSH_TM_TILE_TO_HOST_STS__COUNT_WIDTH 9
+#define RSH_TM_TILE_TO_HOST_STS__COUNT_RESET_VAL 0
+#define RSH_TM_TILE_TO_HOST_STS__COUNT_RMASK 0x1ff
+#define RSH_TM_TILE_TO_HOST_STS__COUNT_MASK  0x1ff
+
+#define RSH_TM_HOST_TO_TILE_DATA 0xa20
+
+#define RSH_TM_TILE_TO_HOST_DATA 0xa40
+
+#define RSH_MMIO_ADDRESS_SPACE__LENGTH 0x10000000000
+#define RSH_MMIO_ADDRESS_SPACE__STRIDE 0x8
+#define RSH_MMIO_ADDRESS_SPACE__OFFSET_SHIFT 0
+#define RSH_MMIO_ADDRESS_SPACE__OFFSET_WIDTH 16
+#define RSH_MMIO_ADDRESS_SPACE__OFFSET_RESET_VAL 0
+#define RSH_MMIO_ADDRESS_SPACE__OFFSET_RMASK 0xffff
+#define RSH_MMIO_ADDRESS_SPACE__OFFSET_MASK  0xffff
+#define RSH_MMIO_ADDRESS_SPACE__PROT_SHIFT 16
+#define RSH_MMIO_ADDRESS_SPACE__PROT_WIDTH 3
+#define RSH_MMIO_ADDRESS_SPACE__PROT_RESET_VAL 0
+#define RSH_MMIO_ADDRESS_SPACE__PROT_RMASK 0x7
+#define RSH_MMIO_ADDRESS_SPACE__PROT_MASK  0x70000
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_SHIFT 23
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_WIDTH 4
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_RESET_VAL 0
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_RMASK 0xf
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_MASK  0x7800000
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_BOOT 0x0
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_RSHIM 0x1
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_UART0 0x2
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_UART1 0x3
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_DIAG_UART 0x4
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_TYU 0x5
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_TYU_EXT1 0x6
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_TYU_EXT2 0x7
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_TYU_EXT3 0x8
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_TIMER 0x9
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_USB 0xa
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_GPIO 0xb
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_MMC 0xc
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_TIMER_EXT 0xd
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_WDOG_NS 0xe
+#define RSH_MMIO_ADDRESS_SPACE__CHANNEL_VAL_WDOG_SEC 0xf
+
+#define RSH_SWINT 0x318
+
+#define RSH_BYTE_ACC_CTL 0x490
+
+#define RSH_BYTE_ACC_WDAT 0x498
+
+#define RSH_BYTE_ACC_RDAT 0x4a0
+
+#define RSH_BYTE_ACC_ADDR 0x4a8
+
+#endif /* !defined(__RSHIM_REGS_H__) */