A specific request from the firmware people is the ability to back-off from
sending more asynchronous operations once a specific number of operations
are in-flight.

This patch adds that ability - with a new sysfs parameter
'outstanding_operations_max' which controls the maximum number of
operations that can be outstanding/in-flight at any time.

When outstanding_operations_max contains a non-zero value and asynchronous
operations are being used - we will back-off until the completion counter
is < outstanding_operations_max. Tested in both synchronous and
asynchronous mode and with gb_loopback_connection_exit() interrupting
in-flight operations.
Suggested-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Reviewed-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

On gb_loopback_connection_exit() we should ensure every issued asynchronous
operation completes before exiting connection_exit(). This patch introduces
a waitqueue with a counter which represents the number of incomplete
asynchronous operations. When the counter reaches zero connection_exit()
will complete. At the point which we wait for outstanding operations to
complete the connection-specific loopback thread will have ceased to issue
new operations. Tested with both synchronous and asynchronous operations.
Reviewed-by: Johan Hovold <johan@hovoldconsulting.com>
Suggested-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

If operation allocation fails we should return -ENOMEM in the asynchronous
operation send routine. If we don't return here then the
gb_loopback_async_operation_put() later can dereference a NULL pointer if
the previous gb_operation_create() failed.
Reported-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Reviewed-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

container_of cannot return NULL and the pointer passed to this context uses
reference counter bumped inside a spinlock, so the base pointer will be
valid at this point.
Suggested-by: Johan Hovold <johan@hovoldconstulting.com>
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Reviewed-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Instead of having the loopback attributes in the bundle device,
Add a struct device to the gb_loopback struct and register it on
connection_init, deregister it at connection_exit, and move the
loopback attribute group over to the new device.

Use device_create_with_groups to create sysfs attributes
together with device.
Suggested-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Suggested-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Axel Haslam <ahaslam@baylibre.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

The mask attribute is not used on the driver anymore and can be
removed.
Signed-off-by: Axel Haslam <ahaslam@baylibre.com>
Reviewed-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Fixed in the specification, some values for chipselect count and index
were different in size, just fix that for all reference to chipselect
and move all to one byte size and remove byte order operations.
Signed-off-by: Rui Miguel Silva <rui.silva@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

CPorts 16 and 17 are reserved for CDSI0 and CDSI1 by the ES2 APB, make
sure they won't be allocated dynamically.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Drop redundant __func__ and add missing errno values.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Fix race with user space when registering the interface.

The interface was registered before having been fully initialised,
something which could lead to user space accessing not-yet-initialised
attribute values (e.g. zero vendor and product ids or empty vendor and
product strings).

Note that this is also needed to be able to let attribute visibility
depend on manifest data (e.g. interface unlock).
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Separate manifest parsing, including bundle and connection creation,
from bundle registration.

Note that this is also needed to allow the interface to not be
registered until the manifest has been parsed.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Separate bundle creation and registration.

Note that the bundle connections still needs to be initialised post
registration as protocol drivers create child devices to the bundle.

This will ultimately allow connection structures to be created while
parsing manifests, but the connections to not be enabled until a driver
is bound.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Separate bundle-connection disabling and destruction, and destroy the
connections along with the bundle.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Use dev_err to report duplicate bundle ids when creating a bundle.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Kill the bundle lock, which looked like it protected the interface
bundle lists but really did not as lock-less manipulations were still
made.

No locking for the interface bundle list is in fact needed as bundles
are created along with the interface, and the list is only used to check
for duplicate bundle ids when parsing the manifest and to destroy the
bundles when removing the interface itself.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Remove defensive WARN_ON testing for a NULL-interface when removing an
interface.

Every call site has just dereferenced the interface.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Fix up the last few places where hexadecimal rather than decimal
notation was used for interface, bundle and cport ids.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Create svc device along with host-device and move the svc-connection to
the svc structure.

The svc connection is enabled when registering the host device, but as
the SVC protocol is currently driven by the SVC, we need to defer
registration of the svc device to the connection request handler.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Create the svc-connection as part of the host device, and destroy it in
the host-device destructor.

The svc-connection is enabled when registering the host device, and
disabled when the host device is deregistered.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Initialise the struct device last when creating the host device.

After device_initialize(), or rather dev_set_name(), we must use
put_device to release the host device. Initialising last will allow for
a simpler release callback.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Separate connection creation from enabling.

This will ultimately allow connection structures to be created while
parsing manifests, but the connections to not be enabled until a driver
is bound.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Unbind protocol at connection exit rather than when the connection is
destroyed.

Now a protocol is only bound while a connection is enabled.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Bind protocol at connection init.

This is an intermediate step in separating connection creation and
enabling.

Note that the protocol is currently still unbound when the connection is
destroyed.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Make sure to fail properly when a protocol is missing.

This prevents the connection from being created, which is fine as we
currently never bind protocols post creation.

This is an intermediate step in moving protocol binding to
connection_init.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Remove gb_bundle_bind_protocols() that was used to iterate over all
registered bundles and bind protocols to them should a protocol become
available post bundle creation.

The protocol abstraction as a generic construct is going away in favour
of class drivers. Connections will be setup when a class driver is
probed, and driver modules can be loaded on-demand by user space based
on uevents and modalias.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Remove obsolete function prototype that was left after a recent code
relocation.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Make sure to print an error message when aborting hotplug processing due
to failure to clear the interface boot status.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

9445c54c ('greybus/loopback: drop bus aggregate calculation') removed the
aggregation of data in-kernel but instead of dropping the reset of
aggregate stastics, converted that reset into a second reset of the
connection-level stats. While this doesn't result in anything bad it's
also definitely a dumb thing to be doing, so, drop it now.

Also ensure we reset the bridge-specific tracking variables at least once.
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Currently the loopback code allows a delay between operations specified in
milliseconds. Having added asynchronous bi-directional support to loopback
its obvious that the delay value would be far more useful specified in
microseconds than milliseconds. So, this patch makes the necessary
conversion.
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

A particular ask from the firmware people for some time now has been the
ability to drive multiple outstanding bi-directional operations from
loopback to loopback Interfaces. This patch implments that change.

The approach taken is to make a call to gb_operation_send() and have
loopback capture the completion callback itself, with a parallel timer to
timeout completion callbacks that take too long. The calling thread will
issue each gb_operation_send() as fast as it can within the constraints of
thread-safety.

In order to support this addition the following new sysfs entries are
created on a per-connection basis.

- async
  Zero indicates loopback should use the traditional synchronous model
  i.e. gb_operation_request_send_sync().

  Non-zero indicates loopback should use the new asynchronous model i.e.
  gb_operation_send()

- requests_completed
  This value indicates the number of requests successfully completed.

- requests_timedout
  This value indicates the number of requests which timed out.

- timeout
  The number of microseconds to give an individual asynchronous request
  before timing that request out.

- timeout_min
  Read-only attribute informs user-space of the minimum allowed timeout.

- timeout_max
  Read-only attribute informs user-space of the maximum allowed timeout.

Note requests_completed + requests_timedout should always equal
iteration_max, once iteration_count == iteration_max. Also, at this time we
support either synchronous or asynchronous operations in one set of
transactions.
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

This patch converts the cross-thread mutex used to synchronize threads with
respect to each other to a spinlock. This is done to enable taking of locks
in the following patches while in atomic context. A small re-order of
locking in connection setup/tear-down is done to minimize the amount of
time spent in spinlock_irqsave().
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

To clearly specify the base for printed values, prefix hexadecimal
values with 0x.
Suggested-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Because the width of our fields is already known, we can use %0Nx (for
hex) to print N bytes and %u (for unsigned decimal), instead of using %h
and %hh, which isn't that readable.

This patch makes following changes:
- s/%hx/%04x
- s/%04hx/%04x
- s/%hhx/%02x
- s/%02hhx/%02x
- s/%hhu/%u
- s/%hu/%u
- s/%x/%02x for u8 value (only at a single place)
Suggested-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

-EIO fits better here, rather than -EINVAL as the arguments to the
routine itself are valid, just that we failed while doing unipro
transfers.
Suggested-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

The order of entries in sysfs-bus-greybus file doesn't match the order
files/directories in sysfs on a real board. More specifically, N-svc
comes at last and ap_interface_id comes before endo_id within the svc.

Fix that.
Reviewed-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Add the bundle id to bundle uevents.

This is needed to identify bundles that are being removed (e.g. at
hot-unplug).
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Add the interface id to interface and bundle uevents.

This is needed to identify interfaces that are being removed (e.g. at
hot-unplug).
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Add the bus id to all greybus uevents.

This is needed to identify devices that are being removed (e.g. at
hot-unplug).
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

Make sure to flush the workqueue from hotplug and hotunplug events when
the svc connection is tore down.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>

At some point we had a statement in a Jira work item to pull out 'bus
level' data from greybus and to have messages to different interfaces be
synchronized with respect to each other. Synchronizing threads with respect
to each other is slow and it turns out we can get the same 'bus level'
stastics by making the user-space test application smarter.

That's great news for the in-kernel code since it means we can cut out a
whole lot of code to-do with calculating 'bus level' aggregate data and we
can stop forcing threads to hit a rendezvous before sending out another
loopback operation.

So this patch drops bus level aggregates in favour of doing that in
user-space. It subtracts a lot of code and cycles that in practice nobody
cares about anyway.
Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>