3. Environment Abstraction Layer

3.1.1. Initialization and Core Launching

Part of the initialization is done by the start function of glibc. A check is also performed at initialization time to ensure that the micro architecture type chosen in the config file is supported by the CPU. Then, the main() function is called. The core initialization and launch is done in rte_eal_init() (see the API documentation). It consist of calls to the pthread library (more specifically, pthread_self(), pthread_create(), and pthread_setaffinity_np()).

Fig. 3.1 EAL Initialization in a Linux Application Environment

3.1.2. Shutdown and Cleanup

During the initialization of EAL resources such as hugepage backed memory can be allocated by core components. The memory allocated during rte_eal_init() can be released by calling the rte_eal_cleanup() function. Refer to the API documentation for details.

3.1.3. Multi-process Support

The Linuxapp EAL allows a multi-process as well as a multi-threaded (pthread) deployment model. See chapter Multi-process Support for more details.

3.1.4. Memory Mapping Discovery and Memory Reservation

The allocation of large contiguous physical memory is done using the hugetlbfs kernel filesystem. The EAL provides an API to reserve named memory zones in this contiguous memory. The physical address of the reserved memory for that memory zone is also returned to the user by the memory zone reservation API.

There are two modes in which DPDK memory subsystem can operate: dynamic mode, and legacy mode. Both modes are explained below.

• Dynamic memory mode

Currently, this mode is only supported on Linux.

In this mode, usage of hugepages by DPDK application will grow and shrink based on application’s requests. Any memory allocation through rte_malloc(), rte_memzone_reserve() or other methods, can potentially result in more hugepages being reserved from the system. Similarly, any memory deallocation can potentially result in hugepages being released back to the system.

Memory allocated in this mode is not guaranteed to be IOVA-contiguous. If large chunks of IOVA-contiguous are required (with “large” defined as “more than one page”), it is recommended to either use VFIO driver for all physical devices (so that IOVA and VA addresses can be the same, thereby bypassing physical addresses entirely), or use legacy memory mode.

For chunks of memory which must be IOVA-contiguous, it is recommended to use rte_memzone_reserve() function with RTE_MEMZONE_IOVA_CONTIG flag specified. This way, memory allocator will ensure that, whatever memory mode is in use, either reserved memory will satisfy the requirements, or the allocation will fail.

There is no need to preallocate any memory at startup using -m or --socket-mem command-line parameters, however it is still possible to do so, in which case preallocate memory will be “pinned” (i.e. will never be released by the application back to the system). It will be possible to allocate more hugepages, and deallocate those, but any preallocated pages will not be freed. If neither -m nor --socket-mem were specified, no memory will be preallocated, and all memory will be allocated at runtime, as needed.

Another available option to use in dynamic memory mode is --single-file-segments command-line option. This option will put pages in single files (per memseg list), as opposed to creating a file per page. This is normally not needed, but can be useful for use cases like userspace vhost, where there is limited number of page file descriptors that can be passed to VirtIO.

If the application (or DPDK-internal code, such as device drivers) wishes to receive notifications about newly allocated memory, it is possible to register for memory event callbacks via rte_mem_event_callback_register() function. This will call a callback function any time DPDK’s memory map has changed.

If the application (or DPDK-internal code, such as device drivers) wishes to be notified about memory allocations above specified threshold (and have a chance to deny them), allocation validator callbacks are also available via rte_mem_alloc_validator_callback_register() function.

• Legacy memory mode

This mode is enabled by specifying --legacy-mem command-line switch to the EAL. This switch will have no effect on FreeBSD as FreeBSD only supports legacy mode anyway.

This mode mimics historical behavior of EAL. That is, EAL will reserve all memory at startup, sort all memory into large IOVA-contiguous chunks, and will not allow acquiring or releasing hugepages from the system at runtime.

If neither -m nor --socket-mem were specified, the entire available hugepage memory will be preallocated.

• 32-bit support

Additional restrictions are present when running in 32-bit mode. In dynamic memory mode, by default maximum of 2 gigabytes of VA space will be preallocated, and all of it will be on master lcore NUMA node unless --socket-mem flag is used.

In legacy mode, VA space will only be preallocated for segments that were requested (plus padding, to keep IOVA-contiguousness).

• Maximum amount of memory

All possible virtual memory space that can ever be used for hugepage mapping in a DPDK process is preallocated at startup, thereby placing an upper limit on how much memory a DPDK application can have. DPDK memory is stored in segment lists, each segment is strictly one physical page. It is possible to change the amount of virtual memory being preallocated at startup by editing the following config variables:

• CONFIG_RTE_MAX_MEMSEG_LISTS controls how many segment lists can DPDK have
• CONFIG_RTE_MAX_MEM_MB_PER_LIST controls how much megabytes of memory each segment list can address
• CONFIG_RTE_MAX_MEMSEG_PER_LIST controls how many segments each segment can have
• CONFIG_RTE_MAX_MEMSEG_PER_TYPE controls how many segments each memory type can have (where “type” is defined as “page size + NUMA node” combination)
• CONFIG_RTE_MAX_MEM_MB_PER_TYPE controls how much megabytes of memory each memory type can address
• CONFIG_RTE_MAX_MEM_MB places a global maximum on the amount of memory DPDK can reserve

Normally, these options do not need to be changed.

3.1.5. PCI Access

The EAL uses the /sys/bus/pci utilities provided by the kernel to scan the content on the PCI bus. To access PCI memory, a kernel module called uio_pci_generic provides a /dev/uioX device file and resource files in /sys that can be mmap’d to obtain access to PCI address space from the application. The DPDK-specific igb_uio module can also be used for this. Both drivers use the uio kernel feature (userland driver).

3.1.6. Per-lcore and Shared Variables

Shared variables are the default behavior. Per-lcore variables are implemented using Thread Local Storage (TLS) to provide per-thread local storage.

3.1.7. Logs

A logging API is provided by EAL. By default, in a Linux application, logs are sent to syslog and also to the console. However, the log function can be overridden by the user to use a different logging mechanism.

3.1.8. CPU Feature Identification

The EAL can query the CPU at runtime (using the rte_cpu_get_features() function) to determine which CPU features are available.

3.1.9. User Space Interrupt Event

• User Space Interrupt and Alarm Handling in Host Thread

The EAL creates a host thread to poll the UIO device file descriptors to detect the interrupts. Callbacks can be registered or unregistered by the EAL functions for a specific interrupt event and are called in the host thread asynchronously. The EAL also allows timed callbacks to be used in the same way as for NIC interrupts.

• RX Interrupt Event

The receive and transmit routines provided by each PMD don’t limit themselves to execute in polling thread mode. To ease the idle polling with tiny throughput, it’s useful to pause the polling and wait until the wake-up event happens. The RX interrupt is the first choice to be such kind of wake-up event, but probably won’t be the only one.

EAL provides the event APIs for this event-driven thread mode. Taking linuxapp as an example, the implementation relies on epoll. Each thread can monitor an epoll instance in which all the wake-up events’ file descriptors are added. The event file descriptors are created and mapped to the interrupt vectors according to the UIO/VFIO spec. From bsdapp’s perspective, kqueue is the alternative way, but not implemented yet.

EAL initializes the mapping between event file descriptors and interrupt vectors, while each device initializes the mapping between interrupt vectors and queues. In this way, EAL actually is unaware of the interrupt cause on the specific vector. The eth_dev driver takes responsibility to program the latter mapping.

The RX interrupt are controlled/enabled/disabled by ethdev APIs - ‘rte_eth_dev_rx_intr_*’. They return failure if the PMD hasn’t support them yet. The intr_conf.rxq flag is used to turn on the capability of RX interrupt per device.

• Device Removal Event

This event is triggered by a device being removed at a bus level. Its underlying resources may have been made unavailable (i.e. PCI mappings unmapped). The PMD must make sure that on such occurrence, the application can still safely use its callbacks.

This event can be subscribed to in the same way one would subscribe to a link status change event. The execution context is thus the same, i.e. it is the dedicated interrupt host thread.

Considering this, it is likely that an application would want to close a device having emitted a Device Removal Event. In such case, calling rte_eth_dev_close() can trigger it to unregister its own Device Removal Event callback. Care must be taken not to close the device from the interrupt handler context. It is necessary to reschedule such closing operation.

3.1.10. Blacklisting

The EAL PCI device blacklist functionality can be used to mark certain NIC ports as blacklisted, so they are ignored by the DPDK. The ports to be blacklisted are identified using the PCIe* description (Domain:Bus:Device.Function).

3.1.11. Misc Functions

Locks and atomic operations are per-architecture (i686 and x86_64).

3.3.1. EAL pthread and lcore Affinity

The term “lcore” refers to an EAL thread, which is really a Linux/FreeBSD pthread. “EAL pthreads” are created and managed by EAL and execute the tasks issued by remote_launch. In each EAL pthread, there is a TLS (Thread Local Storage) called _lcore_id for unique identification. As EAL pthreads usually bind 1:1 to the physical CPU, the _lcore_id is typically equal to the CPU ID.

When using multiple pthreads, however, the binding is no longer always 1:1 between an EAL pthread and a specified physical CPU. The EAL pthread may have affinity to a CPU set, and as such the _lcore_id will not be the same as the CPU ID. For this reason, there is an EAL long option ‘–lcores’ defined to assign the CPU affinity of lcores. For a specified lcore ID or ID group, the option allows setting the CPU set for that EAL pthread.

The format pattern:

–lcores=’<lcore_set>[@cpu_set][,<lcore_set>[@cpu_set],...]’

‘lcore_set’ and ‘cpu_set’ can be a single number, range or a group.

A number is a “digit([0-9]+)”; a range is “<number>-<number>”; a group is “(<number|range>[,<number|range>,...])”.

If a ‘@cpu_set’ value is not supplied, the value of ‘cpu_set’ will default to the value of ‘lcore_set’.

For example, "--lcores='1,2@(5-7),(3-5)@(0,2),(0,6),7-8'" which means start 9 EAL thread;
    lcore 0 runs on cpuset 0x41 (cpu 0,6);
    lcore 1 runs on cpuset 0x2 (cpu 1);
    lcore 2 runs on cpuset 0xe0 (cpu 5,6,7);
    lcore 3,4,5 runs on cpuset 0x5 (cpu 0,2);
    lcore 6 runs on cpuset 0x41 (cpu 0,6);
    lcore 7 runs on cpuset 0x80 (cpu 7);
    lcore 8 runs on cpuset 0x100 (cpu 8).

Using this option, for each given lcore ID, the associated CPUs can be assigned. It’s also compatible with the pattern of corelist(‘-l’) option.

3.3.2. non-EAL pthread support

It is possible to use the DPDK execution context with any user pthread (aka. Non-EAL pthreads). In a non-EAL pthread, the _lcore_id is always LCORE_ID_ANY which identifies that it is not an EAL thread with a valid, unique, _lcore_id. Some libraries will use an alternative unique ID (e.g. TID), some will not be impacted at all, and some will work but with limitations (e.g. timer and mempool libraries).

All these impacts are mentioned in Known Issues section.

3.3.3. Public Thread API

There are two public APIs rte_thread_set_affinity() and rte_thread_get_affinity() introduced for threads. When they’re used in any pthread context, the Thread Local Storage(TLS) will be set/get.

Those TLS include _cpuset and _socket_id:

• _cpuset stores the CPUs bitmap to which the pthread is affinitized.
• _socket_id stores the NUMA node of the CPU set. If the CPUs in CPU set belong to different NUMA node, the _socket_id will be set to SOCKET_ID_ANY.

3.3.4. Known Issues

• rte_mempool

  The rte_mempool uses a per-lcore cache inside the mempool. For non-EAL pthreads, rte_lcore_id() will not return a valid number. So for now, when rte_mempool is used with non-EAL pthreads, the put/get operations will bypass the default mempool cache and there is a performance penalty because of this bypass. Only user-owned external caches can be used in a non-EAL context in conjunction with rte_mempool_generic_put() and rte_mempool_generic_get() that accept an explicit cache parameter.

• rte_ring

  rte_ring supports multi-producer enqueue and multi-consumer dequeue. However, it is non-preemptive, this has a knock on effect of making rte_mempool non-preemptable.

  Note

  The “non-preemptive” constraint means:

  • a pthread doing multi-producers enqueues on a given ring must not be preempted by another pthread doing a multi-producer enqueue on the same ring.
  • a pthread doing multi-consumers dequeues on a given ring must not be preempted by another pthread doing a multi-consumer dequeue on the same ring.

  Bypassing this constraint may cause the 2nd pthread to spin until the 1st one is scheduled again. Moreover, if the 1st pthread is preempted by a context that has an higher priority, it may even cause a dead lock.

  This does not mean it cannot be used, simply, there is a need to narrow down the situation when it is used by multi-pthread on the same core.

  1. It CAN be used for any single-producer or single-consumer situation.
  2. It MAY be used by multi-producer/consumer pthread whose scheduling policy are all SCHED_OTHER(cfs). User SHOULD be aware of the performance penalty before using it.
  3. It MUST not be used by multi-producer/consumer pthreads, whose scheduling policies are SCHED_FIFO or SCHED_RR.

• rte_timer

  Running rte_timer_manage() on a non-EAL pthread is not allowed. However, resetting/stopping the timer from a non-EAL pthread is allowed.

• rte_log

  In non-EAL pthreads, there is no per thread loglevel and logtype, global loglevels are used.

• misc

  The debug statistics of rte_ring, rte_mempool and rte_timer are not supported in a non-EAL pthread.

3.3.5. cgroup control

The following is a simple example of cgroup control usage, there are two pthreads(t0 and t1) doing packet I/O on the same core ($CPU). We expect only 50% of CPU spend on packet IO.

mkdir /sys/fs/cgroup/cpu/pkt_io
mkdir /sys/fs/cgroup/cpuset/pkt_io

echo $cpu > /sys/fs/cgroup/cpuset/cpuset.cpus

echo $t0 > /sys/fs/cgroup/cpu/pkt_io/tasks
echo $t0 > /sys/fs/cgroup/cpuset/pkt_io/tasks

echo $t1 > /sys/fs/cgroup/cpu/pkt_io/tasks
echo $t1 > /sys/fs/cgroup/cpuset/pkt_io/tasks

cd /sys/fs/cgroup/cpu/pkt_io
echo 100000 > pkt_io/cpu.cfs_period_us
echo  50000 > pkt_io/cpu.cfs_quota_us

3.4.1. Cookies

When CONFIG_RTE_MALLOC_DEBUG is enabled, the allocated memory contains overwrite protection fields to help identify buffer overflows.

3.4.2. Alignment and NUMA Constraints

The rte_malloc() takes an align argument that can be used to request a memory area that is aligned on a multiple of this value (which must be a power of two).

On systems with NUMA support, a call to the rte_malloc() function will return memory that has been allocated on the NUMA socket of the core which made the call. A set of APIs is also provided, to allow memory to be explicitly allocated on a NUMA socket directly, or by allocated on the NUMA socket where another core is located, in the case where the memory is to be used by a logical core other than on the one doing the memory allocation.

3.4.3. Use Cases

This API is meant to be used by an application that requires malloc-like functions at initialization time.

For allocating/freeing data at runtime, in the fast-path of an application, the memory pool library should be used instead.

3.4.4. Internal Implementation

3.4.4.1. Data Structures

There are two data structure types used internally in the malloc library:

• struct malloc_heap - used to track free space on a per-socket basis
• struct malloc_elem - the basic element of allocation and free-space tracking inside the library.

3.4.4.1.1. Structure: malloc_heap

The malloc_heap structure is used to manage free space on a per-socket basis. Internally, there is one heap structure per NUMA node, which allows us to allocate memory to a thread based on the NUMA node on which this thread runs. While this does not guarantee that the memory will be used on that NUMA node, it is no worse than a scheme where the memory is always allocated on a fixed or random node.

The key fields of the heap structure and their function are described below (see also diagram above):

• lock - the lock field is needed to synchronize access to the heap. Given that the free space in the heap is tracked using a linked list, we need a lock to prevent two threads manipulating the list at the same time.
• free_head - this points to the first element in the list of free nodes for this malloc heap.
• first - this points to the first element in the heap.
• last - this points to the last element in the heap.

Fig. 3.2 Example of a malloc heap and malloc elements within the malloc library

3.4.4.1.2. Structure: malloc_elem

The malloc_elem structure is used as a generic header structure for various blocks of memory. It is used in two different ways - all shown in the diagram above:

1. As a header on a block of free or allocated memory - normal case
2. As a padding header inside a block of memory

The most important fields in the structure and how they are used are described below.

Malloc heap is a doubly-linked list, where each element keeps track of its previous and next elements. Due to the fact that hugepage memory can come and go, neighbouring malloc elements may not necessarily be adjacent in memory. Also, since a malloc element may span multiple pages, its contents may not necessarily be IOVA-contiguous either - each malloc element is only guaranteed to be virtually contiguous.

Note

If the usage of a particular field in one of the above three usages is not described, the field can be assumed to have an undefined value in that situation, for example, for padding headers only the “state” and “pad” fields have valid values.

• heap - this pointer is a reference back to the heap structure from which this block was allocated. It is used for normal memory blocks when they are being freed, to add the newly-freed block to the heap’s free-list.
• prev - this pointer points to previous header element/block in memory. When freeing a block, this pointer is used to reference the previous block to check if that block is also free. If so, and the two blocks are immediately adjacent to each other, then the two free blocks are merged to form a single larger block.
• next - this pointer points to next header element/block in memory. When freeing a block, this pointer is used to reference the next block to check if that block is also free. If so, and the two blocks are immediately adjacent to each other, then the two free blocks are merged to form a single larger block.
• free_list - this is a structure pointing to previous and next elements in this heap’s free list. It is only used in normal memory blocks; on malloc() to find a suitable free block to allocate and on free() to add the newly freed element to the free-list.
• state - This field can have one of three values: FREE, BUSY or PAD. The former two are to indicate the allocation state of a normal memory block and the latter is to indicate that the element structure is a dummy structure at the end of the start-of-block padding, i.e. where the start of the data within a block is not at the start of the block itself, due to alignment constraints. In that case, the pad header is used to locate the actual malloc element header for the block.
• pad - this holds the length of the padding present at the start of the block. In the case of a normal block header, it is added to the address of the end of the header to give the address of the start of the data area, i.e. the value passed back to the application on a malloc. Within a dummy header inside the padding, this same value is stored, and is subtracted from the address of the dummy header to yield the address of the actual block header.
• size - the size of the data block, including the header itself.