-----------------------
Quiescent State can be defined as "any point in the thread execution where the
-thread does not hold a reference to shared memory". It is up to the application
-to determine its quiescent state.
+thread does not hold a reference to shared memory". It is the responsibility of
+the application to determine its quiescent state.
Let us consider the following diagram:
However, the writer does not need to wait for reader thread 3 to enter
quiescent state. Reader thread 3 was not accessing D1 when the delete
-operation happened. So, reader thread 1 will not have a reference to the
+operation happened. So, reader thread 3 will not have a reference to the
deleted entry.
It can be noted that, the critical sections for D2 is a quiescent state
It is important to make sure that this library keeps the overhead of
identifying the end of grace period and subsequent freeing of memory,
-to a minimum. The following explains how grace period and critical
+to a minimum. The following paras explain how grace period and critical
section affect this overhead.
The writer has to poll the readers to identify the end of grace period.
reporting) in the readers.
Hence, we need the characteristics of a small grace period and large critical
-section. This library addresses this by allowing the writer to do
-other work without having to block until the readers report their quiescent
-state.
+section. This library addresses these characteristics by allowing the writer
+to do other work without having to block until the readers report their
+quiescent state.
RCU in DPDK
-----------
-For DPDK applications, the start and end of a ``while(1)`` loop (where no
+For DPDK applications, the beginning and end of a ``while(1)`` loop (where no
references to shared data structures are kept) act as perfect quiescent
states. This will combine all the shared data structure accesses into a
single, large critical section which helps keep the overhead on the
DPDK supports a pipeline model of packet processing and service cores.
In these use cases, a given data structure may not be used by all the
-workers in the application. The writer does not have to wait for all
-the workers to report their quiescent state. To provide the required
-flexibility, this library has a concept of a QS variable. The application
-can create one QS variable per data structure to help it track the
-end of grace period for each data structure. This helps keep the grace
+workers in the application. The writer has to wait only for the workers that
+use the data structure to report their quiescent state. To provide the required
+flexibility, this library has a concept of a QS variable. If required, the
+application can create one QS variable per data structure to help it track the
+end of grace period for each data structure. This helps keep the length of grace
period to a minimum.
How to use this library
Applications can call ``rte_rcu_qsbr_get_memsize()`` to calculate the size
of memory to allocate. This API takes a maximum number of reader threads,
-using this variable, as a parameter. Currently, a maximum of 1024 threads
-are supported.
+using this variable, as a parameter.
Further, the application can initialize a QS variable using the API
``rte_rcu_qsbr_init()``.
The separation of triggering the reporting from querying the status provides
the writer threads flexibility to do useful work instead of blocking for the
reader threads to enter the quiescent state or go offline. This reduces the
-memory accesses due to continuous polling for the status.
+memory accesses due to continuous polling for the status. But, since the
+resource is freed at a later time, the token and the reference to the deleted
+resource need to be stored for later queries.
The ``rte_rcu_qsbr_synchronize()`` API combines the functionality of
``rte_rcu_qsbr_start()`` and blocking ``rte_rcu_qsbr_check()`` into a single
API. This API triggers the reader threads to report their quiescent state and
polls till all the readers enter the quiescent state or go offline. This API
does not allow the writer to do useful work while waiting and introduces
-additional memory accesses due to continuous polling.
+additional memory accesses due to continuous polling. However, the application
+does not have to store the token or the reference to the deleted resource. The
+resource can be freed immediately after ``rte_rcu_qsbr_synchronize()`` API
+returns.
The reader thread must call ``rte_rcu_qsbr_thread_offline()`` and
``rte_rcu_qsbr_thread_unregister()`` APIs to remove itself from reporting its
quiescent state query and update the state accordingly.
The ``rte_rcu_qsbr_lock()`` and ``rte_rcu_qsbr_unlock()`` are empty functions.
-However, when ``CONFIG_RTE_LIBRTE_RCU_DEBUG`` is enabled, these APIs aid
-in debugging issues. One can mark the access to shared data structures on the
-reader side using these APIs. The ``rte_rcu_qsbr_quiescent()`` will check if
-all the locks are unlocked.
+However, these APIs can aid in debugging issues. One can mark the access to
+shared data structures on the reader side using these APIs. The
+``rte_rcu_qsbr_quiescent()`` will check if all the locks are unlocked.
+
+Resource reclamation framework for DPDK
+---------------------------------------
+
+Lock-free algorithms place additional burden of resource reclamation on
+the application. When a writer deletes an entry from a data structure, the writer:
+
+#. Has to start the grace period
+#. Has to store a reference to the deleted resources in a FIFO
+#. Should check if the readers have completed a grace period and free the resources.
+
+There are several APIs provided to help with this process. The writer
+can create a FIFO to store the references to deleted resources using ``rte_rcu_qsbr_dq_create()``.
+The resources can be enqueued to this FIFO using ``rte_rcu_qsbr_dq_enqueue()``.
+If the FIFO is full, ``rte_rcu_qsbr_dq_enqueue`` will reclaim the resources before enqueuing. It will also reclaim resources on regular basis to keep the FIFO from growing too large. If the writer runs out of resources, the writer can call ``rte_rcu_qsbr_dq_reclaim`` API to reclaim resources. ``rte_rcu_qsbr_dq_delete`` is provided to reclaim any remaining resources and free the FIFO while shutting down.
+
+However, if this resource reclamation process were to be integrated in lock-free data structure libraries, it
+hides this complexity from the application and makes it easier for the application to adopt lock-free algorithms. The following paragraphs discuss how the reclamation process can be integrated in DPDK libraries.
+
+In any DPDK application, the resource reclamation process using QSBR can be split into 4 parts:
+
+#. Initialization
+#. Quiescent State Reporting
+#. Reclaiming Resources
+#. Shutdown
+
+The design proposed here assigns different parts of this process to client libraries and applications. The term 'client library' refers to lock-free data structure libraries such at rte_hash, rte_lpm etc. in DPDK or similar libraries outside of DPDK. The term 'application' refers to the packet processing application that makes use of DPDK such as L3 Forwarding example application, OVS, VPP etc..
+
+The application has to handle 'Initialization' and 'Quiescent State Reporting'. So,
+
+* the application has to create the RCU variable and register the reader threads to report their quiescent state.
+* the application has to register the same RCU variable with the client library.
+* reader threads in the application have to report the quiescent state. This allows for the application to control the length of the critical section/how frequently the application wants to report the quiescent state.
+
+The client library will handle 'Reclaiming Resources' part of the process. The
+client libraries will make use of the writer thread context to execute the memory
+reclamation algorithm. So,
+
+* client library should provide an API to register a RCU variable that it will use. It should call ``rte_rcu_qsbr_dq_create()`` to create the FIFO to store the references to deleted entries.
+* client library should use ``rte_rcu_qsbr_dq_enqueue`` to enqueue the deleted resources on the FIFO and start the grace period.
+* if the library runs out of resources while adding entries, it should call ``rte_rcu_qsbr_dq_reclaim`` to reclaim the resources and try the resource allocation again.
+
+The 'Shutdown' process needs to be shared between the application and the
+client library.
+
+* the application should make sure that the reader threads are not using the shared data structure, unregister the reader threads from the QSBR variable before calling the client library's shutdown function.
+
+* client library should call ``rte_rcu_qsbr_dq_delete`` to reclaim any remaining resources and free the FIFO.
+
+Integrating the resource reclamation with client libraries removes the burden from
+the application and makes it easy to use lock-free algorithms.
+
+This design has several advantages over currently known methods.
+
+#. Application does not need a dedicated thread to reclaim resources. Memory
+ reclamation happens as part of the writer thread with little impact on
+ performance.
+#. The client library has better control over the resources. For example: the client
+ library can attempt to reclaim when it has run out of resources.
git.droids-corp.org / dpdk.git / blobdiff