When fast provisioning is enabled, Provisioner shall not
ignore messages from the nodes whose addresses are not in
the provisioning database. Because other nodes which are
not provisioned by the Primary Provisioner will send node
address messages to the Primary Provisioner.
When using fast provisioning, the Provisioner functionality
will be enabled. Unicast addresses within the pre-allocated
range will be used for provisioning nodes. And during the
address assignment, the allocated unicast address will be
checked that if it's duplicated with other nodes addresses
and the Provisioner's own addresses.
So before starting using fast provisioning, we need to update
the Provisioner address.
Since the behavior of sending segmented messages has been
changed properly, the calculation of timeout value which
will be used when sending an acknowledged message by a
client model also needs to be updated.
Add mutex to protect some variables of the mesh segmented
messages.
Currently the timeout handler of mesh will be executed in
the btc task, while the received mesh messages will be
handled in the btu task. In case some variables are set to
NULL when handled in the btu task, meanwhile these variables
are also accessed in the timeout handler, so we add mutex to
protect these variables and related operations.
The publication context is checked for NULL in bt_mesh_model_publish()
however it was dereferenced before that. Move the assignment to
ctx.send_rel to the same place where other ctx members are set.
Split mesh_util.h into mesh_byteorder.h, mesh_compiler.h,
mesh_ffs.h and mesh_util.h based on the classification of
Zephyr, which will make further porting more clear.
The Friend queue uses the message SeqAuth to determine whether the
message is already in the queue. To facilitate this, the SeqAuth is
passed around as a pointer throughout the transport modules. In the
bt_mesh_ctl_send functions, this parameter is also exposed in the API,
but the internal usage is inconsistent and buggy. Also, no one actually
uses this parameter.
- Removes seq_auth param from bt_mesh_ctl_send, instead passing NULL
directly to the friend module, to enforce its addition to the queue.
- Makes the seq_auth pointer const throughout the friend module.
Changes the behavior of the message cache to optimize for cache
capacity. Previously, the message cache's primary function was to avoid
decrypting messages multiple times, although the cache's main function
in the spec is to avoid message rebroadcasting. Optimizing for minimal
decryption causes us to fill the network cache faster, which in turn
causes more cache misses, potentially outweighing the advantage.
Now stores src + seq in message cache instead of field hash value. This
cuts cache size in two, while including more of the sequence number than
before.
Adds messages to the cache only after the packet is successfully
decrypted. This reduces noise in the cache, and ensures that no
invalid deobfuscations are added.
Additionally, this fixes a bug where multiple calls to net_decrypt with
the same packet failed, as the message cache found its own entry from
the previous call.
Implements several changes to the transport layer segmented tx to
improve group message performance:
- Moves retransmit counter to tx context instead of per packet. As every
unacked packet is sent every retransmit, the retransmit counters would
be the same in each segment. This makes it easier to control progress.
- Delays the scheduling of the retransmit until the completion of the
last segment by adding a seg_pending counter. This is essentially the
same as the old behavior, except that the old behavior might retrigger
the sending before all segments are finished if the advertising is
slow.
- Allows the group transmits to stop as soon as all retransmits have
been exhausted, instead of timing out waiting for acks that won't
come. This allows group tx to finish without error.
- Fixes a bug where a failed TX would block IV update.
- Cancels any pending transmissions of acked segments.
- Reduces log level for several common group tx scenarios that aren't
erronous.
Backport of https://github.com/apache/mynewt-nimble/pull/724
Mesh spec 1.0.1 changes proxy disabling behavior to only affect the
relaying from proxy nodes. Previously, disabling proxy would shut down
all proxy and node activity.
Tweaks from the original commit:
- Removed redundant call to bt_mesh_adv_update() in gatt_proxy_set()
- Removed invalid ref to 4.2.11.1 in node_identity_set()
---
According to Mesh Profile Spec 1.0.1, Section 4.2.11:
"If the Proxy feature is disabled, a GATT client device can connect
over GATT to that node for configuration and control. Messages from
the GATT bearer are not relayed to the advertising bearer."
Moreover some notes have been removed from the spec compared to
version 1.0:
Mesh Profile Spec 1.0, Section 4.2.11:
"Upon transition from GATT Proxy state 0x01 to GATT Proxy state 0x00
the GATT Bearer Server shall disconnect all GATT Bearer Clients."
"The Configuration Client should turn off the Proxy state as the last
step in the configuration process."
Mesh Profile Spec 1.0, Section 4.2.11.1:
"When the GATT Proxy state is set to 0x00, the Node Identity state
for all subnets shall be set to 0x00 and shall not be changed."
Until now the choice of reliable sending (segmented messages with
acks) was implicitly dependent on the size of the payload. Add a new
member to the bt_mesh_model_pub to force using segment acks even when
the payload would fit a single unsegmented message.
When PB-GATT support has been enabled the provisioning code "borrows"
the buffer from the proxy code. However, the way that initialization
was happening the proxy buffers were initialized only after
provisioning initialization, resulting in a corrupted buffer with
buf->data pointing to NULL. Reorder the initialization calls so that
proxy is done first and provisioning only after it.
Allow models to skip a periodic publish interval by returning an error
from the publish update callback.
Previously, an error return from publish update would cancel periodic
publishing. This can't be recovered from, and as such, no valid model
implementation could return an error from this callback, and there was
no way to skip a periodic publish.
