8. Ring Library

The ring allows the management of queues. Instead of having a linked list of infinite size, the rte_ring has the following properties:

FIFO
Maximum size is fixed, the objects are stored in a table
Objects can be pointers or elements of multiple of 4 byte size
Lockless implementation
Multi-consumer or single-consumer dequeue
Multi-producer or single-producer enqueue
Bulk dequeue - Dequeues the specified count of objects if successful; otherwise fails
Bulk enqueue - Enqueues the specified count of objects if successful; otherwise fails
Burst dequeue - Dequeue the maximum available objects if the specified count cannot be fulfilled
Burst enqueue - Enqueue the maximum available objects if the specified count cannot be fulfilled

The advantages of this data structure over a linked list queue are as follows:

Faster; only requires a single 32 bit Compare-And-Swap instruction instead of several pointer size Compare-And-Swap instructions.
Simpler than a full lockless queue.
Adapted to bulk enqueue/dequeue operations. As objects are stored in a table, a dequeue of several objects will not produce as many cache misses as in a linked queue. Also, a bulk dequeue of many objects does not cost more than a dequeue of a simple object.

The disadvantages:

Size is fixed
Having many rings costs more in terms of memory than a linked list queue. An empty ring contains at least N objects.

A simplified representation of a Ring is shown in with consumer and producer head and tail pointers to objects stored in the data structure.

Fig. 8.1 Ring Structure

8.1. References for Ring Implementation in FreeBSD*

The following code was added in FreeBSD 8.0, and is used in some network device drivers (at least in Intel drivers):

bufring.h in FreeBSD

bufring.c in FreeBSD

8.2. Lockless Ring Buffer in Linux*

The following is a link describing the Linux Lockless Ring Buffer Design.

8.3. Additional Features

8.3.1. Name

A ring is identified by a unique name. It is not possible to create two rings with the same name (rte_ring_create() returns NULL if this is attempted).

8.4. Use Cases

Use cases for the Ring library include:

Communication between applications in the DPDK

Used by memory pool allocator

8.5. Anatomy of a Ring Buffer

This section explains how a ring buffer operates. The ring structure is composed of two head and tail couples; one is used by producers and one is used by the consumers. The figures of the following sections refer to them as prod_head, prod_tail, cons_head and cons_tail.

Each figure represents a simplified state of the ring, which is a circular buffer. The content of the function local variables is represented on the top of the figure, and the content of ring structure is represented on the bottom of the figure.

8.5.1. Single Producer Enqueue

This section explains what occurs when a producer adds an object to the ring. In this example, only the producer head and tail (prod_head and prod_tail) are modified, and there is only one producer.

The initial state is to have a prod_head and prod_tail pointing at the same location.

8.5.1.1. Enqueue First Step

First, ring->prod_head and ring->cons_tail are copied in local variables. The prod_next local variable points to the next element of the table, or several elements after in case of bulk enqueue.

If there is not enough room in the ring (this is detected by checking cons_tail), it returns an error.

../_images/ring-enqueue1.svg — Fig. 8.2 Enqueue first step

8.5.1.2. Enqueue Second Step

The second step is to modify ring->prod_head in ring structure to point to the same location as prod_next.

The added object is copied in the ring (obj4).

../_images/ring-enqueue2.svg — Fig. 8.3 Enqueue second step

8.5.1.3. Enqueue Last Step

Once the object is added in the ring, ring->prod_tail in the ring structure is modified to point to the same location as ring->prod_head. The enqueue operation is finished.

../_images/ring-enqueue3.svg — Fig. 8.4 Enqueue last step

8.5.2. Single Consumer Dequeue

This section explains what occurs when a consumer dequeues an object from the ring. In this example, only the consumer head and tail (cons_head and cons_tail) are modified and there is only one consumer.

The initial state is to have a cons_head and cons_tail pointing at the same location.

8.5.2.1. Dequeue First Step

First, ring->cons_head and ring->prod_tail are copied in local variables. The cons_next local variable points to the next element of the table, or several elements after in the case of bulk dequeue.

If there are not enough objects in the ring (this is detected by checking prod_tail), it returns an error.

../_images/ring-dequeue1.svg — Fig. 8.5 Dequeue first step

8.5.2.2. Dequeue Second Step

The second step is to modify ring->cons_head in the ring structure to point to the same location as cons_next.

The dequeued object (obj1) is copied in the pointer given by the user.

../_images/ring-dequeue2.svg — Fig. 8.6 Dequeue second step

8.5.2.3. Dequeue Last Step

Finally, ring->cons_tail in the ring structure is modified to point to the same location as ring->cons_head. The dequeue operation is finished.

../_images/ring-dequeue3.svg — Fig. 8.7 Dequeue last step

8.5.3. Multiple Producers Enqueue

This section explains what occurs when two producers concurrently add an object to the ring. In this example, only the producer head and tail (prod_head and prod_tail) are modified.

The initial state is to have a prod_head and prod_tail pointing at the same location.

8.5.3.1. Multiple Producers Enqueue First Step

On both cores, ring->prod_head and ring->cons_tail are copied in local variables. The prod_next local variable points to the next element of the table, or several elements after in the case of bulk enqueue.

If there is not enough room in the ring (this is detected by checking cons_tail), it returns an error.

../_images/ring-mp-enqueue1.svg — Fig. 8.8 Multiple producer enqueue first step

8.5.3.2. Multiple Producers Enqueue Second Step

The second step is to modify ring->prod_head in the ring structure to point to the same location as prod_next. This operation is done using a Compare And Swap (CAS) instruction, which does the following operations atomically:

If ring->prod_head is different to local variable prod_head, the CAS operation fails, and the code restarts at first step.
Otherwise, ring->prod_head is set to local prod_next, the CAS operation is successful, and processing continues.

In the figure, the operation succeeded on core 1, and step one restarted on core 2.

../_images/ring-mp-enqueue2.svg — Fig. 8.9 Multiple producer enqueue second step

8.5.3.3. Multiple Producers Enqueue Third Step

The CAS operation is retried on core 2 with success.

The core 1 updates one element of the ring(obj4), and the core 2 updates another one (obj5).

../_images/ring-mp-enqueue3.svg — Fig. 8.10 Multiple producer enqueue third step

8.5.3.4. Multiple Producers Enqueue Fourth Step

Each core now wants to update ring->prod_tail. A core can only update it if ring->prod_tail is equal to the prod_head local variable. This is only true on core 1. The operation is finished on core 1.

../_images/ring-mp-enqueue4.svg — Fig. 8.11 Multiple producer enqueue fourth step

8.5.3.5. Multiple Producers Enqueue Last Step

Once ring->prod_tail is updated by core 1, core 2 is allowed to update it too. The operation is also finished on core 2.

../_images/ring-mp-enqueue5.svg — Fig. 8.12 Multiple producer enqueue last step

8.5.4. Modulo 32-bit Indexes

In the preceding figures, the prod_head, prod_tail, cons_head and cons_tail indexes are represented by arrows. In the actual implementation, these values are not between 0 and size(ring)-1 as would be assumed. The indexes are between 0 and 2^32 -1, and we mask their value when we access the object table (the ring itself). 32-bit modulo also implies that operations on indexes (such as, add/subtract) will automatically do 2^32 modulo if the result overflows the 32-bit number range.

The following are two examples that help to explain how indexes are used in a ring.

Note

To simplify the explanation, operations with modulo 16-bit are used instead of modulo 32-bit. In addition, the four indexes are defined as unsigned 16-bit integers, as opposed to unsigned 32-bit integers in the more realistic case.

../_images/ring-modulo1.svg — Fig. 8.13 Modulo 32-bit indexes - Example 1

This ring contains 11000 entries.

../_images/ring-modulo2.svg — Fig. 8.14 Modulo 32-bit indexes - Example 2

This ring contains 12536 entries.

Note

For ease of understanding, we use modulo 65536 operations in the above examples. In real execution cases, this is redundant for low efficiency, but is done automatically when the result overflows.

The code always maintains a distance between producer and consumer between 0 and size(ring)-1. Thanks to this property, we can do subtractions between 2 index values in a modulo-32bit base: that’s why the overflow of the indexes is not a problem.

At any time, entries and free_entries are between 0 and size(ring)-1, even if only the first term of subtraction has overflowed:

uint32_t entries = (prod_tail - cons_head);
uint32_t free_entries = (mask + cons_tail -prod_head);

8.6. Producer/consumer synchronization modes

rte_ring supports different synchronization modes for producers and consumers. These modes can be specified at ring creation/init time via flags parameter. That should help users to configure ring in the most suitable way for his specific usage scenarios. Currently supported modes:

8.6.1. MP/MC (default one)

Multi-producer (/multi-consumer) mode. This is a default enqueue (/dequeue) mode for the ring. In this mode multiple threads can enqueue (/dequeue) objects to (/from) the ring. For ‘classic’ DPDK deployments (with one thread per core) this is usually the most suitable and fastest synchronization mode. As a well known limitation - it can perform quite pure on some overcommitted scenarios.

8.6.2. SP/SC

Single-producer (/single-consumer) mode. In this mode only one thread at a time is allowed to enqueue (/dequeue) objects to (/from) the ring.

8.6.3. MP_RTS/MC_RTS

Multi-producer (/multi-consumer) with Relaxed Tail Sync (RTS) mode. The main difference from the original MP/MC algorithm is that tail value is increased not by every thread that finished enqueue/dequeue, but only by the last one. That allows threads to avoid spinning on ring tail value, leaving actual tail value change to the last thread at a given instance. That technique helps to avoid the Lock-Waiter-Preemption (LWP) problem on tail update and improves average enqueue/dequeue times on overcommitted systems. To achieve that RTS requires 2 64-bit CAS for each enqueue(/dequeue) operation: one for head update, second for tail update. In comparison the original MP/MC algorithm requires one 32-bit CAS for head update and waiting/spinning on tail value.

8.6.4. MP_HTS/MC_HTS

Multi-producer (/multi-consumer) with Head/Tail Sync (HTS) mode. In that mode enqueue/dequeue operation is fully serialized: at any given moment only one enqueue/dequeue operation can proceed. This is achieved by allowing a thread to proceed with changing head.value only when head.value == tail.value. Both head and tail values are updated atomically (as one 64-bit value). To achieve that 64-bit CAS is used by head update routine. That technique also avoids the Lock-Waiter-Preemption (LWP) problem on tail update and helps to improve ring enqueue/dequeue behavior in overcommitted scenarios. Another advantage of fully serialized producer/consumer - it provides the ability to implement MT safe peek API for rte_ring.

8.7. Ring Peek API

For ring with serialized producer/consumer (HTS sync mode) it is possible to split public enqueue/dequeue API into two phases:

enqueue/dequeue start
enqueue/dequeue finish

That allows user to inspect objects in the ring without removing them from it (aka MT safe peek) and reserve space for the objects in the ring before actual enqueue. Note that this API is available only for two sync modes:

Single Producer/Single Consumer (SP/SC)
Multi-producer/Multi-consumer with Head/Tail Sync (HTS)

It is a user responsibility to create/init ring with appropriate sync modes selected. As an example of usage:

/* read 1 elem from the ring: */
uint32_t n = rte_ring_dequeue_bulk_start(ring, &obj, 1, NULL);
if (n != 0) {
    /* examine object */
    if (object_examine(obj) == KEEP)
        /* decided to keep it in the ring. */
        rte_ring_dequeue_finish(ring, 0);
    else
        /* decided to remove it from the ring. */
        rte_ring_dequeue_finish(ring, n);
}

Note that between _start_ and _finish_ none other thread can proceed with enqueue(/dequeue) operation till _finish_ completes.

8.8. Ring Peek Zero Copy API

Along with the advantages of the peek APIs, zero copy APIs provide the ability to copy the data to the ring memory directly without the need for temporary storage (for ex: array of mbufs on the stack).

These APIs make it possible to split public enqueue/dequeue API into 3 phases:

enqueue/dequeue start
copy data to/from the ring
enqueue/dequeue finish

Note that this API is available only for two sync modes:

Single Producer/Single Consumer (SP/SC)
Multi-producer/Multi-consumer with Head/Tail Sync (HTS)

It is a user responsibility to create/init ring with appropriate sync modes. Following is an example of usage:

/* Reserve space on the ring */
n = rte_ring_enqueue_zc_burst_start(r, 32, &zcd, NULL);
/* Pkt I/O core polls packets from the NIC */
if (n != 0) {
    nb_rx = rte_eth_rx_burst(portid, queueid, zcd->ptr1, zcd->n1);
    if (nb_rx == zcd->n1 && n != zcd->n1)
        nb_rx += rte_eth_rx_burst(portid, queueid, zcd->ptr2,
                                                    n - zcd->n1);
    /* Provide packets to the packet processing cores */
    rte_ring_enqueue_zc_finish(r, nb_rx);
}

Note that between _start_ and _finish_ no other thread can proceed with enqueue(/dequeue) operation till _finish_ completes.

8.9. References

bufring.h in FreeBSD (version 8)

bufring.c in FreeBSD (version 8)

Linux Lockless Ring Buffer Design