TMS320C6457: PBM resource in NDK library

Hi Takayuki,

Hopefully this page will answer your memory related questions with the NDK: https://processors.wiki.ti.com/index.php/TI-RTOS_Networking_Stack_Memory_Usage

Todd

Hello Todd,

I tried to look into answers for #1 and #2 in this Processor Wiki, I was not able to find them. According to NDK user’s guide( https://www.ti.com/lit/pdf/spru523 ), it seems, user is able to set MMALLOC_MAXSIZE to “65500” Can we have your comments on those inquires, please?

Best regards, Miyazaki

Hello Todd,

I got additional inquiries about NDK API Reference Guide (https://www.ti.com/lit/ug/spru524k/spru524k.pdf) from our customer. I'm not familiar with NDK API, therefore, I just tried to translate their inquiries directly. Sorry.

1. Regarding "_ipcfg.SockBufMinTx" (p173)

Does "The send buffer size" mean “transmitBufSize” ? According to this description, This value is usually about 25% to 50% of the send buffer size. Customer seems to set around 6%. In this case, is there any problem?

2. Regarding each definition in CCS configuration file

Does each definition in CCS configuration file match as those descriptions in NDK? It means, customer would like to confirm the following relationship between them.

【app_p64P.cfg.xml】　→　【spur524】

transmitBufSize　→　_ipcfg.SockTcpTxBufSize

receiveBufSize 　→　_ipcfg.SockTcpRxBufSize

maxReassemblySize 　→　_ipcfg.IpReasmMaxSize

socketBufMinTxSize 　→　_ipcfg.SockBufMinTx

socketBufMinRxSize 　→　ipcfg.SockBufMinRx

3. NDK_send() /NDK_recv()

Customer is using send()/recv() API at this time, should customer use NDK_send()/NDK_recv()? 

4. Regarding PBM

Customer does not use NDK_***() APIs at current project. However, should customer also use PBM tasks in this case?

5. Customer noticed that TaskCreate() was never called in their application software. In this case, customer should also use PBM's buffer and should do Ethernet communication with PBM?

6. In the case of using PBM, send()/recv() are using PBM pool. There is possibility that PBM pool runs out. Something like the following URL's explanation(https://e2e.ti.com/support/processors/f/791/p/905776/3370042#3370042). Is that correct?

It will be appreciated if you are able to share your comments on those inquiries.

Best regards,

Miyazaki

Miyazaki,

Re: #1 - _ipcfg.SockBufMinTx

When the description for _ipcfg.SockBufMinTx is referring to "send buffer size", is means the value as defined by _ipcfg.SockTcpTxBufSize. During startup, this parameter will be set by the Tcp.transmitBufSize defined in the application.cfg script, but might be changed later by a call to CfgAddEntry().

There is no problem setting _ipcfg.SockBufMinTx to 6% of the transmit buffer size. It just means that the socket will be marked as write-able sooner, but with less room for new data. Note that using _ipcfg.SockBufMinTx implies you are using fdSelect() with non-blocking sockets. When using the send() function, it does not use this configuration parameter.

Re: #2 - CCS configuration vs NDK _ipcfg

You have the correct relationship between these parameters. The values you specify in the application.cfg script will be applied to the NDK _ipcfg parameters during the startup phase. When you create a new socket, it will take its values from _ipcfg. Once you have a socket, you can modify its parameters using the setSockOpt() function.

Re: #3 - BSD vs NDK

It is recommended to use the BSD functions (send, recv). But an application is allowed to use both the BSD and the NDK functions (NDK_send, NDK_recv) at the same time. Both APIs must be invoked from an NDK task so that the proper fdSession is active.

It might not be possible to use both APIs from the same source file because the header files might not be incompatible. But it should work if you separate your code into two separate source files. But it would be best to simply use the BSD API alone.

Re: #4 - PBM

The customer should not be using any PBM functions. The only attention to PBM is for configuration of the PBM pool. After that, the PBM buffers are used internally by the NDK and the driver. They should not be directly accessed by
the application.

Re: #5 - TaskCreate

It is highly recommended that an application use TaskCreate() to create a new task from which the network APIs are invoked. This is needed to setup the correct fdSession context. However, it is possible to make BSD and NDK API calls directly from a kernel task, but it requires setting up the proper fdSession context manually. It is not recommended to use this approach.

Re: #6 - PBM pool

Both the BSD and NDK APIs make use of the PBM pool. If the application is not structured correctly, it is possible for the PBM pool to run out and cause the network traffic to stop.

The PBM buffers are used internally for both sending and receiving data. They are used to move data between the driver layer and the application layer, in both directions. The problem can occur on the receiving path. When new data is received, the driver will automatically acquire PBM buffers to store the incoming data, and then pass the PBM buffer up to the NDK stack. The data will sit there until the application calls recv(), at which point the data is transferred out of the PBM buffer into the application buffer. Then the empty PBM buffer is returned to the pool. If the application never calls recv(), or does not call is soon enough, it is possible for the PBM pool to run out. If this happens, and the application attempts to send data, it would fail because there are no PBM buffers available to transfer the outbound data from the application to the driver.

~Ramsey