With this patch gb_bootrom_timedout was getting called in interrupt
context and then proceeding to call functions that might block. In
practical terms, this was leading to a kernel BUG at mm/vmalloc.c:1650.
Signed-off-by: Jeffrey Carlyle <jcarlyle@google.com> Acked-by: Johan Hovold <johan@hovoldconsulting.com>
Viresh Kumar [Thu, 5 May 2016 10:03:20 +0000 (15:33 +0530)]
greybus: fw-download: Introduce timeouts for firmware downloads
As per greybus specification, the AP can apply, implementation
dependent, timeouts for:
- The time interval between the Find Firmware Response and the first
Fetch Firmware Request.
- The time interval between a Fetch Firmware Response and the next Fetch
Firmware Request.
- The time interval between a Fetch Firmware Response and the Release
Firmware Request.
- The time interval between the Find Firmware Response and the Release
Firmware Request.
This patch implements those timeouts.
The timeout period for the first three cases is fixed to one-second and
the timeout for the last one is finalized at runtime, dependent on the
total size of the firmware.
There can be two possible paths now, which may race for freeing or
getting the 'struct fw_request'. They are:
- Request handler: initiated from the Module side.
- Timeout handler: initiated on timeout of the programmed timer.
And so a mutex is added to avoid races.
Every caller which needs to access the 'struct fw_request' increments
the reference count, so that the structure doesn't get freed in
parallel. Once the structure is freed and reference is put by all the
users, the structure is freed.
If we timeout while waiting for a request from the Module, the AP frees
the 'struct fw_request', but does *not* free the request-id. This is
done to guarantee that a delayed request from the Module for the expired
id, doesn't get access to a new 'struct fw_request' allocated later with
the same id.
Tested with gbsim by hacking its code to delay the release request and
indefinitely fetch the same section of the firmware package. Both timed
out on the AP side and the 'struct fw_request' is free properly. Further
requests work fine after few are timed out. And rmmod (followed by more
similar testing) works just fine.
Viresh Kumar [Thu, 5 May 2016 10:03:19 +0000 (15:33 +0530)]
greybus: fw-download: Manage firmware requests with kref
This patch updates the fw-download core to manage firmware requests with
kref. This is required for the next patch, which will introduce timeouts
for firmware downloads.
greybus: arche-platform: Rework platform/wd-line state transition logic
This patch simplifies and improves the readability of the internal state
transition logic of arche platform driver state transition logic by:
1. Making a dedicated function to state-transition the platform code.
2. Condense the decision to swtich states down to a single switch
statement instead of four separate if/else clauses.
greybus: arche-platform: Export fn to allow timesync driver to change the state
With the addition of the timesync driver and looking at the hardware
interfaces we have, its clear we need to add a new arche-platform state.
This patch adds ARCHE_PLATFORM_STATE_TIME_SYNC to the arche-platform driver
to facilitate transition to the TIME_SYNC state if-and-only-if the
arche-platform driver is in the ACTIVE state.
This is mainly needed as wake/detect lines are shared between TIMESYNC
operation and basic control functionality of APBs. So during TIMESYNC
we want to make sure that the events on wake/detect lines are
ignored by the arche-platform APB reset logic.
This patch adds one exported function, which can be invoked from
timesync driver code, allowing, switching between
ARCHE_PLATFORM_STATE_TIME_SYNC <=> ARCHE_PLATFORM_STATE_ACTIVE states.
[ bryan.odonoghue@linaro.org: Added mutex, massaged commit text ]
With addition of exported function, required for TIMESYNC operation,
we need more locking mechanism for driver state, so to avoid confusion
rename existing spinlock variable to its appropriate name.
In order to verify TimeSync functionality we require a TimeSync-ping
operation where the AP can command the SVC to initiate a single strobe of
downstream TimeSync slaves, and report the FrameTime at the strobe. Ping
will only be valid after the system has transitioned to a TIMESYNC_ACTIVE
state.
In the active state each TimeSync slave will graph a single incoming SVC
strobe as a ping and will store its frame time. The AP will then gather
each last-event FrameTime for presentation to user-space in the AP and/or
further automation based on the reported FrameTimes at the SVC ping.
This patch adds the SVC ping command definition to greybus_protocols.h.
Its necessary to establish an initial state on the wake-detect lines before
enabling timesync on APB/GPB in order to ensure all GPIO edge-transitions are
correctly interpreted by the receiving processor.
This patch adds the operations defined in the Greybus specification to
greybus_protocols.h, this involves adding the
SVC_TIMESYNC_WAKE_PINS_ACQUIRE and SVC_TIMESYNC_WAKE_PINS_RELEASE commands
and moving the 'strobe_mask' parameter from 'struct
gb_svc_timesync_enable_request' to 'struct
gb_svc_timesync_wd_pins_acquire_request' since the communication of the
strobe_mask will be communicated before preparing any of the bridges to
receive the TimeSync pulses.
A separate patch to the greybus specification describes the transition
between the wake sub-state and the timesync sub-state.
Mitchell Tasman [Wed, 4 May 2016 21:30:23 +0000 (17:30 -0400)]
greybus: svc: reconfig APBridgeA-Switch link to handle required load
SW-4894, SW-4389, and share a common root cause, namely that
the power-on reset configuration of the APBridgeA-Switch link of PWM
Gear 1, 1 Lane, Slow Auto, is insufficient to handle some required
traffic loads, such as 3 audio streams plus boot-over-UniPro or 4 audio
streams.
The correct long-term solution is to implement a UniPro Power Mode
Manager as in that considers the demands placed on the network,
and adjusts power modes accordingly.
The present commit implements a short-term, brute-force hack to allow
continued system testing:
- Upon receiving an SVC HELLO request, schedule deferred work to
reconfigure the APB1-Switch link to PWM G2, 1 lane, Slow Auto
- When the Camera driver transitions a White Camera module from active to
inactive, return the APB1-Switch link to PWM G2, 1 lane, Slow Auto
The Camera driver already steps up the APBridgeA-Camera link speed while a
camera module is active, which affords sufficient margin for simultaneous
audio and hotplug activity, and the Camera driver already steps down the
link speed thereafter: the change made by the present patch is simply to
tweak the stepped-down power mode to match the new baseline configuration.
Viresh Kumar [Tue, 3 May 2016 10:43:12 +0000 (16:13 +0530)]
greybus: bootrom: Implement timeouts to detect Module failures
Its possible that the Module may fail to download the next stage
firmware, or to jump into it and boot into the new personality.
We have already seen examples of both of these cases on EVT 1.5.
This patch implements timeouts in the bootrom bundle driver, which now
expects the next request from the Module to be received at the AP within
1 second of the previous request/response. The time interval can be
increased later if required.
The timeouts are added between:
- AP_READY and FIRMWARE_SIZE operations
- FIRMWARE_SIZE and GET_FIRMWARE operations
- Two GET_FIRMWARE operations
- GET_FIRMWARE and READY_TO_BOOT operations
- READY_TO_BOOT operation and the call to the ->disconnect() event of
the bootrom bundle (once the new hotplug request is received).
The timeout for the last case is kept at 5 seconds right now (random
value), as it may take a bit longer.
Because 'bootrom->fw' can be accessed simultaneously (from timeout
handler and incoming requests) and one of them can potentially free the
'->fw' structure, a mutex is also added to take care of such races while
accessing 'bootrom->fw' structure.
Also note that the '!bootrom->fw' check is moved to free_firmware()
routine.
Tested on EVT 1.5, by faking errors on certain requests, so that the
bootrom doesn't send any more requests. Normal case is working just
fine for audio and GP modules.
Modify sequence of register_module & unregister_module in bundle
driver. This would affect the uevent generated for above user
space. Accordingly, we need to modify snd_soc_xxx sequence in
register_module() in codec driver.
Vaibhav Agarwal [Wed, 4 May 2016 10:59:21 +0000 (16:29 +0530)]
greybus: audio: Remove redundant lock protection & is_connected field
Each module maintains connected status & a lock to protect it.
Using codec->lock we can safely serialize ASoC specific callbacks
(in response to mixer_ctl update or dai_ops) and gb module
disconnect. Thus is_connected field can be removed.
Vaibhav Agarwal [Wed, 4 May 2016 10:59:19 +0000 (16:29 +0530)]
greybus: audio:gb_manager: Use proper locking around kobject_xxx
read/write_lock_irqsave mechanism was used to protect modules
list & kobject_xxx() in gb_audio_manager. Since kobject_xxx calls
can sleep spin_lock variants can't be used there. So use rw_sem
for protecting modules_list.
All firmware packages on the Modules or Interfaces are now managed by a
special Firmware Management Protocol. The Interface Manifest shall
at least contain the Firmware Management Bundle and a Firmware
Management Protocol CPort within it.
The bundle may contain additional CPorts based on the extra
functionality required to manage firmware packages.
For example, this is how the Firmware Management Bundle of the Interface
Manifest may look like:
This patch adds the basic firmware-management bundle driver, which just
creates a firmware-management connection. Support for individual
protocols will be added separately.
greybus: audio: Reorder gb_deactivate sequence to avoid protocol error
gb_activate_tx/rx is triggered from _prepare() & gb_deactivate
from shutdown(). This may cause protocol error in case shutdown
executes without _prepare due to some hw_params failure.
Also, reorganise _prepare & _shutdown calls to make it more
readable & cleaner.
David Lin [Sat, 23 Apr 2016 02:03:43 +0000 (19:03 -0700)]
greybus: svc: free pwrmon_rails memory upon exit
For every time SVC instance is created, memories for storing the rail IDs
are allocated, however, they are not freed when the SVC is destroyed.
This patch fixes the memory leak by freeing the memory when debugfs for
SVC is no longer needed.
Testing Done:
- Check pwrmon debugfs after turning on and off SVC
Signed-off-by: David Lin <dtwlin@google.com> Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
David Lin [Sat, 23 Apr 2016 02:03:42 +0000 (19:03 -0700)]
greybus: svc: clean up gb_svc struct for pwrmon
The power rail names and counts are unnecessarily stored in the gb_svc
structure once the SVC created, this causes waste of memory usage. This
patch removes rail names and rail counts storage from th gb_svc
structure.
Testing Done:
- Validated the readings from /d/greybus/1-svc/pwrmon/*
Signed-off-by: David Lin <dtwlin@google.com> Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
Johan Hovold [Sat, 23 Apr 2016 16:47:31 +0000 (18:47 +0200)]
greybus: svc: implement interface mailbox event
Implement the new interface mailbox-event operation.
The event is sent by the SVC under certain conditions when an interface
updates its mailbox value. Specifically, this event will be used to
implement the new mode-switch functionality.
Upon reception the AP verifies that the interface is known and that the
mailbox has the expected MAILBOX_GREYBUS value. If so, the interface is
disabled before being re-enabled (re-enumerated).
Note that during mode-switch, the interface will typically already be in
a disabled state when the mailbox is written (with the ES3 bootrom being
the notable exception).
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com> Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
Johan Hovold [Sat, 23 Apr 2016 16:47:30 +0000 (18:47 +0200)]
greybus: svc: implement module inserted and removed operations
Implement the new module inserted and removed operations.
The SVC sends these after detecting a module insertion or removal, and
in the former case after having determined the module geometry (i.e.
position and size).
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com> Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
Johan Hovold [Sat, 23 Apr 2016 16:47:28 +0000 (18:47 +0200)]
greybus: svc: add stub interface-activate function
Add message structures (based on the current spec) for the SVC Interface
Activate operation, and a stub function that always return the Greybus
interface type.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com> Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
Implement controlled module removal through a new module attribute
"eject".
When a non-zero argument is written to the attribute, all interfaces of
the module are disabled (e.g. bundles are deregistered) and deactivated
(e.g. powered off) before instructing the SVC to physically eject the
module.
Note that the module device is not deregistered until the SVC has
reported the physical removal of all of its interfaces.
A new interface mutex is added to enforce interface state-change
serialisation.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com> Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
Johan Hovold [Sat, 23 Apr 2016 16:47:24 +0000 (18:47 +0200)]
greybus: core: add module abstraction
Add Greybus module abstraction that will be used to implement controlled
module removal (eject) and represent module geometry.
Greybus module devices correspond to physical modules and have one or
more interfaces. Modules have an id that is identical to the id of their
primary interface, which in turn is the interface with lowest numbered
id. The module name is constructed from the bus and module id:
<bus_id>-<module_id>
Interfaces, bundles, and control devices are consequently renamed as
As before, interface ids (and therefore in a sense now also module ids)
correspond to physical interface positions on the frame.
Modules have the following attributes:
module_id
num_interfaces
where module_id is the id of the module and num_interface the number of
interfaces the module has.
Note that until SVC module-size detection has been implemented, all
interfaces are considered to be part of 1x2 modules. Specifically, the
two interfaces of a 2x2 module will be presented as two 1x2 modules for
now.
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com> Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
The attribute entries have been kept mostly sorted within each device
type. Let's move the three more-recently added interface attributes
that were not.
Johan Hovold [Sat, 23 Apr 2016 16:47:20 +0000 (18:47 +0200)]
greybus: svc: use a common prefix for debugfs functions
Use the common gb_svc functions also for the recently added svc
functions. Having a common prefix clearly signals where the code
resides, something which improves readability and helps during
debugging (e.g. stack traces).
Note that all functions in svc.c except for these three use the common
prefix with the exception of the pwr_debugfs callbacks (that still use
*a* common prefix) and the attribute accessors (than can not have a
common prefix due to some macro magic).
Signed-off-by: Johan Hovold <johan@hovoldconsulting.com> Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
greybus: arche-ctrl: Don't expose driver internals to arche-platform driver
We have chosen the *ugly* way of registering two platform drivers from
the module_init() of only one of them, so that we can avoid having two
separate modules for them.
But we should still be doing this in a sane way. There is no need to
expose internals of arche-ctrl to arche-platform, like PM-ops, probe,
resume, id-table, etc. Just expose an init and a exit callback.
greybus: audio: acquire wakelock during active playback
use pm_stay_awake & pm_relax to avoid suspend sequence during
active playback
testing Done:
Music Playback ongoing
$ cat /sys/devices/soc.0/qcom,ara-codec.82/power/wakeup_active
1
Music Playback stopped
$ cat /sys/devices/soc.0/qcom,ara-codec.82/power/wakeup_active
0
Tested-by: David Lin <dtwlin@google.com> Reviewed-by: David Lin <dtwlin@google.com> Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Akash Choudhari <akashtc@google.com> Signed-off-by: Vaibhav Agarwal <vaibhav.agarwal@linaro.org>
Johan Hovold [Thu, 21 Apr 2016 09:43:36 +0000 (11:43 +0200)]
greybus: svc: keep error messages uniform
All SVC error messages, except for a few recently added ones, place the
errno last after a colon (:). Let's at least try to be consistent
within the svc code.
Note that this format also allows for more concise messages without risk
for ambiguity.
greybus: es2: don't protest when getting -ENOENT/ESHUTDOWN USB errors
-ENOENT or -ESHUTDOWN happens when the urbs are being killed from
->disconnect() callback. Don't complain to userspace about this, as the
user will see this on es2 module removal.
Tested by removing the gb-es2.ko module on both EVT1.5 and qemu with
gbsim. The driver doesn't throw anymore errors like:
"urb cport in error -2 (dropped)".
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Tested-by: David Lin <dtwlin@google.com> Reviewed-by: Johan Hovold <johan@hovoldconsulting.com> Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
greybus: es2: disable cport-in after remove greybus host device
This allows greybus to do operations before being removed.
Tested by removing the gb-es2.ko module on both EVT1.5 and qemu with
gbsim. The driver removes the greybus host device successfully before
disabling the urbs. The errors are still coming ("urb cport in error -2
(dropped)"), but with an error value -2 (ENOENT) instead of -108. And
that happens while the urbs are disabled, not while doing greybus
operations.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Acked-by: David Lin <dtwlin@google.com> Reviewed-by: Johan Hovold <johan@hovoldconsulting.com> Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
greybus: es2: Allow proper release of greybus host device
The usb core disables the Interface prior to calling ->disconnect() for
the usb driver. That disallows the es2 driver to remove the greybus host
device and every entity below it in a clean way, as the greybus core may
want to do few operations over the usb connection before getting
removed.
And so we see bunch of errors while we remove the es2 module, like:
"urb cport in error -108 (dropped)"
The usb core has a special per-driver flag, 'soft_unbind', for such usb
drivers. If this flag is set by a driver, the usb core doesn't disable
the Interface prior to calling ->disconnect().
Set that flag for es2.
Tested by removing the gb-es2.ko module on both EVT1.5 and qemu with
gbsim. The interface isn't disabled by the core and the driver is still
able initiate greybus operations over the Interface. This can be
properly tested only after the next patch which removes the greybus host
device before disabling the urbs.
greybus: audio-codec: Staticize few routines to fix build warnings
This fixes below warnings ..
greybus/audio_codec.c:20:32: warning: symbol 'find_data' was not declared. Should it be static?
greybus/audio_codec.c:955:6: warning: symbol 'gbaudio_codec_cleanup' was not declared. Should it be static?
Johan Hovold [Wed, 13 Apr 2016 17:19:08 +0000 (19:19 +0200)]
greybus: svc: keep interfaces registered during mode switch
Keep a detected interface registered until it is physically removed.
Specifically, do not re-register an interface that is switching mode.
Note that this also allows us to get rid of some nasty hacks from core.
The Ara VID/PID bootrom hack for ES2 will continue to work, but is now
mostly confined to the bootrom driver.
Johan Hovold [Wed, 13 Apr 2016 17:19:04 +0000 (19:19 +0200)]
greybus: interface: move vendor and product strings to control device
The control device is an abstraction of the control connection over
which a greybus manifest is retrieved. As interfaces switch modes (e.g.
after boot-over-unipro) they expose new manifests, which can contain
different vendor and product strings.
Eventually control devices will be deregistered and recreated after an
interface mode switch, while the interface itself remains registered.
Note that only interfaces of type greybus will have control devices.
Specifically, dummy interfaces will not.
Johan Hovold [Wed, 13 Apr 2016 17:18:58 +0000 (19:18 +0200)]
greybus: Documentation/sysfs: rename the svc eject attribute
The svc eject attribute was added as an interim solution and is still
used to implement a form of forced ejection.
This will soon be superseded by the module eject attribute, which will
provide an interface for clean eject. We may keep the forced-eject
mechanism around indefinitely, albeit possibly with a different name
(e.g. forced_intf_eject). Either way, update the example tree to reflect
the actual name, intf_eject, which currently used for this svc attribute.
Johan Hovold [Wed, 13 Apr 2016 17:18:55 +0000 (19:18 +0200)]
greybus: manifest: fix illegal free in error path
The manifest-parsing code could end up leaving the interface
vendor_string set to an error pointer that we'd eventually try to free
when destroying the interface.
When transfer speed is too slow (less than 17Khz) the operation can take
longer than the default greybus timeout. Because of this we need to use
the request_send_sync_timeout and calculate the correct timeout for each
operation.
Signed-off-by: Rui Miguel Silva <rui.silva@linaro.org> Tested-by: Philip Yang <philipy@bsquare.com> Tested-by: Axel Haslam <ahaslam@baylibre.com> Signed-off-by: Greg Kroah-Hartman <gregkh@google.com>
projects / mirror_ubuntu-bionic-kernel.git / log