[message-pool] evict message from coap pending request queue if send queue is empty

[sarveshkumarv3]

If the send queue is empty, currently the messages are not evicted and this could cause received messages to be dropped when too many pending address query notifications. The changes made are to evict the message from head of the coap pending request queue if send queue is empty and new buffer cannot be allocated

[size-report]

Size Report of OpenThread

Merging #9984 into main(0e36799).

name	branch	text	data	bss	total
ot-cli-ftd	main	467112	760	66364	534236
	#9984	467240	760	66364	534364
	+/-	+128	0	0	+128
ot-ncp-ftd	main	436380	760	61576	498716
	#9984	436516	760	61576	498852
	+/-	+136	0	0	+136
libopenthread-ftd.a	main	236152	0	40310	276462
	#9984	236270	0	40310	276580
	+/-	+118	0	0	+118
libopenthread-cli-ftd.a	main	57406	0	8075	65481
	#9984	57406	0	8075	65481
	+/-	0	0	0	0
libopenthread-ncp-ftd.a	main	31857	0	5916	37773
	#9984	31857	0	5916	37773
	+/-	0	0	0	0
ot-cli-mtd	main	364888	760	51220	416868
	#9984	365008	760	51220	416988
	+/-	+120	0	0	+120
ot-ncp-mtd	main	347420	760	46448	394628
	#9984	347564	760	46448	394772
	+/-	+144	0	0	+144
libopenthread-mtd.a	main	158411	0	25182	183593
	#9984	158529	0	25182	183711
	+/-	+118	0	0	+118
libopenthread-cli-mtd.a	main	39787	0	8059	47846
	#9984	39787	0	8059	47846
	+/-	0	0	0	0
libopenthread-ncp-mtd.a	main	24737	0	5916	30653
	#9984	24737	0	5916	30653
	+/-	0	0	0	0
ot-cli-ftd-br	main	536064	768	131076	667908
	#9984	536208	768	131076	668052
	+/-	+144	0	0	+144
libopenthread-ftd-br.a	main	299705	5	104998	404708
	#9984	299823	5	104998	404826
	+/-	+118	0	0	+118
libopenthread-cli-ftd-br.a	main	71069	0	8099	79168
	#9984	71069	0	8099	79168
	+/-	0	0	0	0
ot-rcp	main	62248	564	20604	83416
	#9984	62248	564	20604	83416
	+/-	0	0	0	0
libopenthread-rcp.a	main	9542	0	5052	14594
	#9984	9542	0	5052	14594
	+/-	0	0	0	0
libopenthread-radio.a	main	18821	0	214	19035
	#9984	18821	0	214	19035
	+/-	0	0	0	0

[abtink]

Looks great. Thanks @sarveshkumarv3.
Some suggestions below:

[abtink]

Thinking more this PR and #9949, some aspects to consider:

TMF Eviction Impact: This PR allows any message (regardless of priority) to possibly evict a TMF agent request. Since TMF messages are generally high priority themselves, should we add checks for such evictions (e.g., based on priority)?

Investigating Address Query/Notification Overload: From #9949, this change aims to handle situations where a device receives many address queries, leading to frequent TMF address notifications. It may help to delve deeper into this and understand it better (address the root cause):

• Query Origin: Are address queries coming from a single device/BR or multiple sources?
• Rate Limiting: If queries originate from the same device, are existing query rate-limiting mechanisms functioning correctly? Are the limits appropriate?
• CoAP Acknowledgement: TMF address notification messages are high-priority, confirmable CoAP messages. Upon receiving a CoAP acknowledgement, they should be swiftly removed from the request queue. Why is ack message being delayed, leading to request pile-up? (particularly since device seem to keep receiving query message from the same sender?)

[sarveshkumarv3]

Thinking more this PR and #9949, some aspects to consider:

TMF Eviction Impact: This PR allows any message (regardless of priority) to possibly evict a TMF agent request. Since TMF messages are generally high priority themselves, should we add checks for such evictions (e.g., based on priority)?

thanks for your inputs@abtink!

The eviction ordering is to first evict low priority messages in send queue, high priority messages in send queue and only if that is empty to evict TMF agent request right? Also we evict from the head of the request queue, which means TMF agent messages are consuming lot of messages and one at the head is already in the buffers for a long time (In the log, I see around 26s to go through all the retransmission attempts)?

Please let me know if anything else can be prioritized ahead of TMF agent messages? The other thought is MLE queue or MPL queues. The reassembly queues anyway have a limit of 2s after which it would be discarded?

Investigating Address Query/Notification Overload: From #9949, this change aims to handle situations where a device receives many address queries, leading to frequent TMF address notifications. It may help to delve deeper into this and understand it better (address the root cause):

• Query Origin: Are address queries coming from a single device/BR or multiple sources?
• Rate Limiting: If queries originate from the same device, are existing query rate-limiting mechanisms functioning correctly? Are the limits appropriate?
• CoAP Acknowledgement: TMF address notification messages are high-priority, confirmable CoAP messages. Upon receiving a CoAP acknowledgement, they should be swiftly removed from the request queue. Why is ack message being delayed, leading to request pile-up? (particularly since device seem to keep receiving query message from the same sender?)

Yes, I agree, from the given logs, it is not clear why we see so frequent address queries from the accessory (with no visibility here), but I suspect there is some underlying matter stack related issue, causing multiple accessories to go into a crash-loop (running out of buffers) and hence do not handle the AddressNotify.

The changes here would still help to mitigate the Border Router being able to receive from other accessories (protecting its RX buffers) in case we encounter scenario of misbehaving accessories?

[abtink]

Thanks @sarveshkumarv3 for submitting the PR and all the updates LGTM.

Couple of smaller suggestions below.

On this:

Also we evict from the head of the request queue, which means TMF agent messages are consuming lot of messages and one at the head is already in the buffers for a long time (In the log, I see around 26s to go through all the retransmission attempts)?

I cannot think of any good way to use the priority when deciding to evict from TMF agent queue. So the approach in the PR sounds reaonable.

[sarveshkumarv3]

Thanks @sarveshkumarv3 for submitting the PR and all the updates LGTM.

Couple of smaller suggestions below.

On this:

Also we evict from the head of the request queue, which means TMF agent messages are consuming lot of messages and one at the head is already in the buffers for a long time (In the log, I see around 26s to go through all the retransmission attempts)?

I cannot think of any good way to use the priority when deciding to evict from TMF agent queue. So the approach in the PR sounds reaonable.

thanks for the inputs @abtink, incorporated the changes now

[codecov]

Codecov Report

Attention: Patch coverage is 86.66667% with 2 lines in your changes are missing coverage. Please review.

Project coverage is 82.32%. Comparing base (1fb1d9a) to head (5710ebb).
Report is 5 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #9984       +/-   ##
===========================================
+ Coverage   68.31%   82.32%   +14.01%     
===========================================
  Files         494      549       +55     
  Lines       57933    68748    +10815     
===========================================
+ Hits        39577    56600    +17023     
+ Misses      18356    12148     -6208     

Files	Coverage Δ
src/core/coap/coap.hpp	84.61% <ø> (+60.61%)	⬆️
src/core/common/message.cpp	95.88% <100.00%> (+1.30%)	⬆️
src/core/coap/coap.cpp	81.22% <75.00%> (+35.15%)	⬆️

... and 361 files with indirect coverage changes