If the server application dies and needs to be restarted, all client applications also need to be restarted,
as there is no support in the server application for it to run as a secondary process.
Any client processes that need restarting can be restarted without affecting the server process.
How the Application Works
^^^^^^^^^^^^^^^^^^^^^^^^^
The server process performs the network port and data structure initialization much as the symmetric multi-process application does when run as primary.
One additional enhancement in this sample application is that the server process stores its port configuration data in a memory zone in hugepage shared memory.
This eliminates the need for the client processes to have the portmask parameter passed into them on the command line,
as is done for the symmetric multi-process application, and therefore eliminates mismatched parameters as a potential source of errors.
In the same way that the server process is designed to be run as a primary process instance only,
the client processes are designed to be run as secondary instances only.
They have no code to attempt to create shared memory objects.
Instead, handles to all needed rings and memory pools are obtained via calls to rte_ring_lookup() and rte_mempool_lookup().
The network ports for use by the processes are obtained by loading the network port drivers and probing the PCI bus,
which will, as in the symmetric multi-process example,
automatically get access to the network ports using the settings already configured by the primary/server process.
Once all applications are initialized, the server operates by reading packets from each network port in turn and
distributing those packets to the client queues (software rings, one for each client process) in round-robin order.
On the client side, the packets are read from the rings in as big of bursts as possible, then routed out to a different network port.
The routing used is very simple. All packets received on the first NIC port are transmitted back out on the second port and vice versa.
Similarly, packets are routed between the 3rd and 4th network ports and so on.
The sending of packets is done by writing the packets directly to the network ports; they are not transferred back via the server process.
In both the server and the client processes, outgoing packets are buffered before being sent,
so as to allow the sending of multiple packets in a single burst to improve efficiency.
For example, the client process will buffer packets to send,
until either the buffer is full or until we receive no further packets from the server.
Master-slave Multi-process Example
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The fourth example of DPDK multi-process support demonstrates a master-slave model that
provide the capability of application recovery if a slave process crashes or meets unexpected conditions.
In addition, it also demonstrates the floating process,
which can run among different cores in contrast to the traditional way of binding a process/thread to a specific CPU core,
using the local cache mechanism of mempool structures.
This application performs the same functionality as the L2 Forwarding sample application,
therefore this chapter does not cover that part but describes functionality that is introduced in this multi-process example only.
Please refer to :doc:`l2_forward_real_virtual` for more information.
Unlike previous examples where all processes are started from the command line with input arguments, in this example,
only one process is spawned from the command line and that process creates other processes.
The following section describes this in more detail.
Master-slave Process Models
^^^^^^^^^^^^^^^^^^^^^^^^^^^
The process spawned from the command line is called the *master process* in this document.
A process created by the master is called a *slave process*.
The application has only one master process, but could have multiple slave processes.
Once the master process begins to run, it tries to initialize all the resources such as
memory, CPU cores, driver, ports, and so on, as the other examples do.
Thereafter, it creates slave processes, as shown in the following figure.
.._figure_master_slave_proc:
..figure:: img/master_slave_proc.*
Master-slave Process Workflow
The master process calls the rte_eal_mp_remote_launch() EAL function to launch an application function for each pinned thread through the pipe.
Then, it waits to check if any slave processes have exited.
If so, the process tries to re-initialize the resources that belong to that slave and launch them in the pinned thread entry again.
The following section describes the recovery procedures in more detail.
For each pinned thread in EAL, after reading any data from the pipe, it tries to call the function that the application specified.
In this master specified function, a fork() call creates a slave process that performs the L2 forwarding task.
Then, the function waits until the slave exits, is killed or crashes. Thereafter, it notifies the master of this event and returns.
Finally, the EAL pinned thread waits until the new function is launched.
After discussing the master-slave model, it is necessary to mention another issue, global and static variables.
For multiple-thread cases, all global and static variables have only one copy and they can be accessed by any thread if applicable.
So, they can be used to sync or share data among threads.
In the previous examples, each process has separate global and static variables in memory and are independent of each other.
If it is necessary to share the knowledge, some communication mechanism should be deployed, such as, memzone, ring, shared memory, and so on.
The global or static variables are not a valid approach to share data among processes.
For variables in this example, on the one hand, the slave process inherits all the knowledge of these variables after being created by the master.
On the other hand, other processes cannot know if one or more processes modifies them after slave creation since that
is the nature of a multiple process address space.
But this does not mean that these variables cannot be used to share or sync data; it depends on the use case.
The following are the possible use cases:
#. The master process starts and initializes a variable and it will never be changed after slave processes created. This case is OK.
#. After the slave processes are created, the master or slave cores need to change a variable, but other processes do not need to know the change.
This case is also OK.
#. After the slave processes are created, the master or a slave needs to change a variable.
In the meantime, one or more other process needs to be aware of the change.
In this case, global and static variables cannot be used to share knowledge. Another communication mechanism is needed.
A simple approach without lock protection can be a heap buffer allocated by rte_malloc or mem zone.
Slave Process Recovery Mechanism
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Before talking about the recovery mechanism, it is necessary to know what is needed before a new slave instance can run if a previous one exited.
When a slave process exits, the system returns all the resources allocated for this process automatically.
However, this does not include the resources that were allocated by the DPDK. All the hardware resources are shared among the processes,
which include memzone, mempool, ring, a heap buffer allocated by the rte_malloc library, and so on.
If the new instance runs and the allocated resource is not returned, either resource allocation failed or the hardware resource is lost forever.
When a slave process runs, it may have dependencies on other processes.
They could have execution sequence orders; they could share the ring to communicate; they could share the same port for reception and forwarding;
they could use lock structures to do exclusive access in some critical path.
What happens to the dependent process(es) if the peer leaves?
The consequence are varied since the dependency cases are complex.
It depends on what the processed had shared.
However, it is necessary to notify the peer(s) if one slave exited.
Then, the peer(s) will be aware of that and wait until the new instance begins to run.
Therefore, to provide the capability to resume the new slave instance if the previous one exited, it is necessary to provide several mechanisms:
#. Keep a resource list for each slave process.
Before a slave process run, the master should prepare a resource list.
After it exits, the master could either delete the allocated resources and create new ones,
or re-initialize those for use by the new instance.
#. Set up a notification mechanism for slave process exit cases. After the specific slave leaves,
the master should be notified and then help to create a new instance.
This mechanism is provided in Section `Master-slave Process Models`_.
#. Use a synchronization mechanism among dependent processes.
The master should have the capability to stop or kill slave processes that have a dependency on the one that has exited.
Then, after the new instance of exited slave process begins to run, the dependency ones could resume or run from the start.
The example sends a STOP command to slave processes dependent on the exited one, then they will exit.
Thereafter, the master creates new instances for the exited slave processes.
The following diagram describes slave process recovery.
.._figure_slave_proc_recov:
..figure:: img/slave_proc_recov.*
Slave Process Recovery Process Flow
Floating Process Support
^^^^^^^^^^^^^^^^^^^^^^^^
When the DPDK application runs, there is always a -c option passed in to indicate the cores that are enabled.
Then, the DPDK creates a thread for each enabled core.
By doing so, it creates a 1:1 mapping between the enabled core and each thread.
The enabled core always has an ID, therefore, each thread has a unique core ID in the DPDK execution environment.
With the ID, each thread can easily access the structures or resources exclusively belonging to it without using function parameter passing.
It can easily use the rte_lcore_id() function to get the value in every function that is called.
For threads/processes not created in that way, either pinned to a core or not, they will not own a unique ID and the
rte_lcore_id() function will not work in the correct way.
However, sometimes these threads/processes still need the unique ID mechanism to do easy access on structures or resources.
For example, the DPDK mempool library provides a local cache mechanism
(refer to :ref:`mempool_local_cache`)
for fast element allocation and freeing.
If using a non-unique ID or a fake one,
a race condition occurs if two or more threads/ processes with the same core ID try to use the local cache.
Therefore, unused core IDs from the passing of parameters with the -c option are used to organize the core ID allocation array.
Once the floating process is spawned, it tries to allocate a unique core ID from the array and release it on exit.
A natural way to spawn a floating process is to use the fork() function and allocate a unique core ID from the unused core ID array.
However, it is necessary to write new code to provide a notification mechanism for slave exit
and make sure the process recovery mechanism can work with it.
To avoid producing redundant code, the Master-Slave process model is still used to spawn floating processes,
then cancel the affinity to specific cores.
Besides that, clear the core ID assigned to the DPDK spawning a thread that has a 1:1 mapping with the core mask.
Thereafter, get a new core ID from the unused core ID allocation array.
Run the Application
^^^^^^^^^^^^^^^^^^^
This example has a command line similar to the L2 Forwarding sample application with a few differences.
To run the application, start one copy of the l2fwd_fork binary in one terminal.
Unlike the L2 Forwarding example,
this example requires at least three cores since the master process will wait and be accountable for slave process recovery.
