4. Environment Abstraction Layer

The Environment Abstraction Layer (EAL) is responsible for gaining access to low-level resources such as hardware and memory space. It provides a generic interface that hides the environment specifics from the applications and libraries. It is the responsibility of the initialization routine to decide how to allocate these resources (that is, memory space, devices, timers, consoles, and so on).

Typical services expected from the EAL are:

  • DPDK Loading and Launching: The DPDK and its application are linked as a single application and must be loaded by some means.
  • Core Affinity/Assignment Procedures: The EAL provides mechanisms for assigning execution units to specific cores as well as creating execution instances.
  • System Memory Reservation: The EAL facilitates the reservation of different memory zones, for example, physical memory areas for device interactions.
  • Trace and Debug Functions: Logs, dump_stack, panic and so on.
  • Utility Functions: Spinlocks and atomic counters that are not provided in libc.
  • CPU Feature Identification: Determine at runtime if a particular feature, for example, Intel® AVX is supported. Determine if the current CPU supports the feature set that the binary was compiled for.
  • Interrupt Handling: Interfaces to register/unregister callbacks to specific interrupt sources.
  • Alarm Functions: Interfaces to set/remove callbacks to be run at a specific time.

4.1. EAL in a Linux-userland Execution Environment

In a Linux user space environment, the DPDK application runs as a user-space application using the pthread library.

The EAL performs physical memory allocation using mmap() in hugetlbfs (using huge page sizes to increase performance). This memory is exposed to DPDK service layers such as the Mempool Library.

At this point, the DPDK services layer will be initialized, then through pthread setaffinity calls, each execution unit will be assigned to a specific logical core to run as a user-level thread.

The time reference is provided by the CPU Time-Stamp Counter (TSC) or by the HPET kernel API through a mmap() call.

4.1.1. 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()).

../_images/linuxapp_launch.svg

Fig. 4.1 EAL Initialization in a Linux Application Environment

Note

Initialization of objects, such as memory zones, rings, memory pools, lpm tables and hash tables, should be done as part of the overall application initialization on the main lcore. The creation and initialization functions for these objects are not multi-thread safe. However, once initialized, the objects themselves can safely be used in multiple threads simultaneously.

4.1.2. 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.

4.1.3. Multi-process Support

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

4.1.4. Memory Mapping Discovery and Memory Reservation

The allocation of large contiguous physical memory is done using hugepages. 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.

Note

Memory reservations done using the APIs provided by rte_malloc are also backed by hugepages unless --no-huge option is given.

4.1.4.1. Dynamic Memory Mode

Currently, this mode is only supported on Linux and Windows.

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.

A default validator callback is provided by EAL, which can be enabled with a --socket-limit command-line option, for a simple way to limit maximum amount of memory that can be used by DPDK application.

Warning

Memory subsystem uses DPDK IPC internally, so memory allocations/callbacks and IPC must not be mixed: it is not safe to allocate/free memory inside memory-related or IPC callbacks, and it is not safe to use IPC inside memory-related callbacks. See chapter Multi-process Support for more details about DPDK IPC.

4.1.4.2. 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.

4.1.4.3. Hugepage Allocation Matching

This behavior is enabled by specifying the --match-allocations command-line switch to the EAL. This switch is Linux-only and not supported with --legacy-mem nor --no-huge.

Some applications using memory event callbacks may require that hugepages be freed exactly as they were allocated. These applications may also require that any allocation from the malloc heap not span across allocations associated with two different memory event callbacks. Hugepage allocation matching can be used by these types of applications to satisfy both of these requirements. This can result in some increased memory usage which is very dependent on the memory allocation patterns of the application.

4.1.4.4. 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 main 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).

4.1.4.5. 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:

  • RTE_MAX_MEMSEG_LISTS controls how many segment lists can DPDK have
  • RTE_MAX_MEM_MB_PER_LIST controls how much megabytes of memory each segment list can address
  • RTE_MAX_MEMSEG_PER_LIST controls how many segments each segment list can have
  • RTE_MAX_MEMSEG_PER_TYPE controls how many segments each memory type can have (where “type” is defined as “page size + NUMA node” combination)
  • RTE_MAX_MEM_MB_PER_TYPE controls how much megabytes of memory each memory type can address
  • 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.

Note

Preallocated virtual memory is not to be confused with preallocated hugepage memory! All DPDK processes preallocate virtual memory at startup. Hugepages can later be mapped into that preallocated VA space (if dynamic memory mode is enabled), and can optionally be mapped into it at startup.

4.1.4.6. Hugepage Mapping

Below is an overview of methods used for each OS to obtain hugepages, explaining why certain limitations and options exist in EAL. See the user guide for a specific OS for configuration details.

FreeBSD uses contigmem kernel module to reserve a fixed number of hugepages at system start, which are mapped by EAL at initialization using a specific sysctl().

Windows EAL allocates hugepages from the OS as needed using Win32 API, so available amount depends on the system load. It uses virt2phys kernel module to obtain physical addresses, unless running in IOVA-as-VA mode (e.g. forced with --iova-mode=va).

Linux allows to select any combination of the following:

  • use files in hugetlbfs (the default) or anonymous mappings (--in-memory);
  • map each hugepage from its own file (the default) or map multiple hugepages from one big file (--single-file-segments).

Mapping hugepages from files in hugetlbfs is essential for multi-process, because secondary processes need to map the same hugepages. EAL creates files like rtemap_0 in directories specified with --huge-dir option (or in the mount point for a specific hugepage size). The rte prefix can be changed using --file-prefix. This may be needed for running multiple primary processes that share a hugetlbfs mount point. Each backing file by default corresponds to one hugepage, it is opened and locked for the entire time the hugepage is used. This may exhaust the number of open files limit (NOFILE). See Segment File Descriptors section on how the number of open backing file descriptors can be reduced.

In dynamic memory mode, EAL removes a backing hugepage file when all pages mapped from it are freed back to the system. However, backing files may persist after the application terminates in case of a crash or a leak of DPDK memory (e.g. rte_free() is missing). This reduces the number of hugepages available to other processes as reported by /sys/kernel/mm/hugepages/hugepages-*/free_hugepages. EAL can remove the backing files after opening them for mapping if --huge-unlink is given to avoid polluting hugetlbfs. However, since it disables multi-process anyway, using anonymous mapping (--in-memory) is recommended instead.

EAL memory allocator relies on hugepages being zero-filled. Hugepages are cleared by the kernel when a file in hugetlbfs or its part is mapped for the first time system-wide to prevent data leaks from previous users of the same hugepage. EAL ensures this behavior by removing existing backing files at startup and by recreating them before opening for mapping (as a precaution).

One exception is --huge-unlink=never mode. It is used to speed up EAL initialization, usually on application restart. Clearing memory constitutes more than 95% of hugepage mapping time. EAL can save it by remapping existing backing files with all the data left in the mapped hugepages (“dirty” memory). Such segments are marked with RTE_MEMSEG_FLAG_DIRTY. Memory allocator detects dirty segments and handles them accordingly, in particular, it clears memory requested with rte_zmalloc*(). In this mode EAL also does not remove a backing file when all pages mapped from it are freed, because they are intended to be reusable at restart.

Anonymous mapping does not allow multi-process architecture. This mode does not use hugetlbfs and thus does not require root permissions for memory management (the limit of locked memory amount, MEMLOCK, still applies). It is free of filename conflict and leftover file issues. If memfd_create(2) is supported both at build and run time, DPDK memory manager can provide file descriptors for memory segments, which are required for VirtIO with vhost-user backend. This can exhaust the number of open files limit (NOFILE) despite not creating any visible files. See Segment File Descriptors section on how the number of open file descriptors used by EAL can be reduced.

4.1.4.7. Segment File Descriptors

On Linux, in most cases, EAL will store segment file descriptors in EAL. This can become a problem when using smaller page sizes due to underlying limitations of glibc library. For example, Linux API calls such as select() may not work correctly because glibc does not support more than certain number of file descriptors.

There are two possible solutions for this problem. The recommended solution is to use --single-file-segments mode, as that mode will not use a file descriptor per each page, and it will keep compatibility with Virtio with vhost-user backend. This option is not available when using --legacy-mem mode.

Another option is to use bigger page sizes. Since fewer pages are required to cover the same memory area, fewer file descriptors will be stored internally by EAL.

4.1.4.8. Hugepage Worker Stacks

When the --huge-worker-stack[=size] EAL option is specified, worker thread stacks are allocated from hugepage memory local to the NUMA node of the thread. Worker stack size defaults to system pthread stack size if the optional size parameter is not specified.

Warning

Stacks allocated from hugepage memory are not protected by guard pages. Worker stacks must be sufficiently sized to prevent stack overflow when this option is used.

As with normal thread stacks, hugepage worker thread stack size is fixed and is not dynamically resized. Therefore, an application that is free of stack page faults under a given load should be safe with hugepage worker thread stacks given the same thread stack size and loading conditions.

4.1.5. Support for Externally Allocated Memory

It is possible to use externally allocated memory in DPDK. There are two ways in which using externally allocated memory can work: the malloc heap API’s, and manual memory management.

  • Using heap API’s for externally allocated memory

Using a set of malloc heap API’s is the recommended way to use externally allocated memory in DPDK. In this way, support for externally allocated memory is implemented through overloading the socket ID - externally allocated heaps will have socket ID’s that would be considered invalid under normal circumstances. Requesting an allocation to take place from a specified externally allocated memory is a matter of supplying the correct socket ID to DPDK allocator, either directly (e.g. through a call to rte_malloc) or indirectly (through data structure-specific allocation API’s such as rte_ring_create). Using these API’s also ensures that mapping of externally allocated memory for DMA is also performed on any memory segment that is added to a DPDK malloc heap.

Since there is no way DPDK can verify whether memory is available or valid, this responsibility falls on the shoulders of the user. All multiprocess synchronization is also user’s responsibility, as well as ensuring that all calls to add/attach/detach/remove memory are done in the correct order. It is not required to attach to a memory area in all processes - only attach to memory areas as needed.

The expected workflow is as follows:

  • Get a pointer to memory area
  • Create a named heap
  • Add memory area(s) to the heap
    • If IOVA table is not specified, IOVA addresses will be assumed to be unavailable, and DMA mappings will not be performed
    • Other processes must attach to the memory area before they can use it
  • Get socket ID used for the heap
  • Use normal DPDK allocation procedures, using supplied socket ID
  • If memory area is no longer needed, it can be removed from the heap
    • Other processes must detach from this memory area before it can be removed
  • If heap is no longer needed, remove it
    • Socket ID will become invalid and will not be reused

For more information, please refer to rte_malloc API documentation, specifically the rte_malloc_heap_* family of function calls.

  • Using externally allocated memory without DPDK API’s

While using heap API’s is the recommended method of using externally allocated memory in DPDK, there are certain use cases where the overhead of DPDK heap API is undesirable - for example, when manual memory management is performed on an externally allocated area. To support use cases where externally allocated memory will not be used as part of normal DPDK workflow, there is also another set of API’s under the rte_extmem_* namespace.

These API’s are (as their name implies) intended to allow registering or unregistering externally allocated memory to/from DPDK’s internal page table, to allow API’s like rte_mem_virt2memseg etc. to work with externally allocated memory. Memory added this way will not be available for any regular DPDK allocators; DPDK will leave this memory for the user application to manage.

The expected workflow is as follows:

  • Get a pointer to memory area
  • Register memory within DPDK
    • If IOVA table is not specified, IOVA addresses will be assumed to be unavailable
    • Other processes must attach to the memory area before they can use it
  • Perform DMA mapping with rte_dev_dma_map if needed
  • Use the memory area in your application
  • If memory area is no longer needed, it can be unregistered
    • If the area was mapped for DMA, unmapping must be performed before unregistering memory
    • Other processes must detach from the memory area before it can be unregistered

Since these externally allocated memory areas will not be managed by DPDK, it is therefore up to the user application to decide how to use them and what to do with them once they’re registered.

4.1.6. Per-lcore and Shared Variables

Note

lcore refers to a logical execution unit of the processor, sometimes called a hardware thread.

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

4.1.7. 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.

4.1.7.1. Trace and Debug Functions

There are some debug functions to dump the stack in glibc. The rte_panic() function can voluntarily provoke a SIG_ABORT, which can trigger the generation of a core file, readable by gdb.

4.1.8. 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.

4.1.9. 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.

Note

In DPDK PMD, the only interrupts handled by the dedicated host thread are those for link status change (link up and link down notification) and for sudden device removal.

  • 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 Linux 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 FreeBSD’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.

Note

Per queue RX interrupt event is only allowed in VFIO which supports multiple MSI-X vector. In UIO, the RX interrupt together with other interrupt causes shares the same vector. In this case, when RX interrupt and LSC(link status change) interrupt are both enabled(intr_conf.lsc == 1 && intr_conf.rxq == 1), only the former is capable.

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.

4.1.10. Block list

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

4.1.11. Misc Functions

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

4.1.12. IOVA Mode Detection

IOVA Mode is selected by considering what the current usable Devices on the system require and/or support.

On FreeBSD, RTE_IOVA_PA is always the default. On Linux, the IOVA mode is detected based on a 2-step heuristic detailed below.

For the first step, EAL asks each bus its requirement in terms of IOVA mode and decides on a preferred IOVA mode.

  • if all buses report RTE_IOVA_PA, then the preferred IOVA mode is RTE_IOVA_PA,
  • if all buses report RTE_IOVA_VA, then the preferred IOVA mode is RTE_IOVA_VA,
  • if all buses report RTE_IOVA_DC, no bus expressed a preference, then the preferred mode is RTE_IOVA_DC,
  • if the buses disagree (at least one wants RTE_IOVA_PA and at least one wants RTE_IOVA_VA), then the preferred IOVA mode is RTE_IOVA_DC (see below with the check on Physical Addresses availability),

If the buses have expressed no preference on which IOVA mode to pick, then a default is selected using the following logic:

  • if physical addresses are not available, RTE_IOVA_VA mode is used
  • if /sys/kernel/iommu_groups is not empty, RTE_IOVA_VA mode is used
  • otherwise, RTE_IOVA_PA mode is used

In the case when the buses had disagreed on their preferred IOVA mode, part of the buses won’t work because of this decision.

The second step checks if the preferred mode complies with the Physical Addresses availability since those are only available to root user in recent kernels. Namely, if the preferred mode is RTE_IOVA_PA but there is no access to Physical Addresses, then EAL init fails early, since later probing of the devices would fail anyway.

Note

The RTE_IOVA_VA mode is preferred as the default in most cases for the following reasons:

  • All drivers are expected to work in RTE_IOVA_VA mode, irrespective of physical address availability.
  • By default, the mempool, first asks for IOVA-contiguous memory using RTE_MEMZONE_IOVA_CONTIG. This is slow in RTE_IOVA_PA mode and it may affect the application boot time.
  • It is easy to enable large amount of IOVA-contiguous memory use cases with IOVA in VA mode.

It is expected that all PCI drivers work in both RTE_IOVA_PA and RTE_IOVA_VA modes.

If a PCI driver does not support RTE_IOVA_PA mode, the RTE_PCI_DRV_NEED_IOVA_AS_VA flag is used to dictate that this PCI driver can only work in RTE_IOVA_VA mode.

When the KNI kernel module is detected, RTE_IOVA_PA mode is preferred as a performance penalty is expected in RTE_IOVA_VA mode.

4.1.13. IOVA Mode Configuration

Auto detection of the IOVA mode, based on probing the bus and IOMMU configuration, may not report the desired addressing mode when virtual devices that are not directly attached to the bus are present. To facilitate forcing the IOVA mode to a specific value the EAL command line option --iova-mode can be used to select either physical addressing(‘pa’) or virtual addressing(‘va’).

4.1.14. Max SIMD bitwidth

The EAL provides a single setting to limit the max SIMD bitwidth used by DPDK, which is used in determining the vector path, if any, chosen by a component. The value can be set at runtime by an application using the ‘rte_vect_set_max_simd_bitwidth(uint16_t bitwidth)’ function, which should only be called once at initialization, before EAL init. The value can be overridden by the user using the EAL command-line option ‘–force-max-simd-bitwidth’.

When choosing a vector path, along with checking the CPU feature support, the value of the max SIMD bitwidth must also be checked, and can be retrieved using the ‘rte_vect_get_max_simd_bitwidth()’ function. The value should be compared against the enum values for accepted max SIMD bitwidths:

enum rte_vect_max_simd {
    RTE_VECT_SIMD_DISABLED = 64,
    RTE_VECT_SIMD_128 = 128,
    RTE_VECT_SIMD_256 = 256,
    RTE_VECT_SIMD_512 = 512,
    RTE_VECT_SIMD_MAX = INT16_MAX + 1,
};

 if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512)
     /* Take AVX-512 vector path */
 else if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
     /* Take AVX2 vector path */

4.2. Memory Segments and Memory Zones (memzone)

The mapping of physical memory is provided by this feature in the EAL. As physical memory can have gaps, the memory is described in a table of descriptors, and each descriptor (called rte_memseg ) describes a physical page.

On top of this, the memzone allocator’s role is to reserve contiguous portions of physical memory. These zones are identified by a unique name when the memory is reserved.

The rte_memzone descriptors are also located in the configuration structure. This structure is accessed using rte_eal_get_configuration(). The lookup (by name) of a memory zone returns a descriptor containing the physical address of the memory zone.

Memory zones can be reserved with specific start address alignment by supplying the align parameter (by default, they are aligned to cache line size). The alignment value should be a power of two and not less than the cache line size (64 bytes). Memory zones can also be reserved from either 2 MB or 1 GB hugepages, provided that both are available on the system.

Both memsegs and memzones are stored using rte_fbarray structures. Please refer to DPDK API Reference for more information.

4.3. Multiple pthread

DPDK usually pins one pthread per core to avoid the overhead of task switching. This allows for significant performance gains, but lacks flexibility and is not always efficient.

Power management helps to improve the CPU efficiency by limiting the CPU runtime frequency. However, alternately it is possible to utilize the idle cycles available to take advantage of the full capability of the CPU.

By taking advantage of cgroup, the CPU utilization quota can be simply assigned. This gives another way to improve the CPU efficiency, however, there is a prerequisite; DPDK must handle the context switching between multiple pthreads per core.

For further flexibility, it is useful to set pthread affinity not only to a CPU but to a CPU set.

4.3.1. 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.

4.3.2. non-EAL pthread support

It is possible to use the DPDK execution context with any user pthread (aka. non-EAL pthreads). There are two kinds of non-EAL pthreads:

  • a registered non-EAL pthread with a valid _lcore_id that was successfully assigned by calling rte_thread_register(),
  • a non registered non-EAL pthread with a LCORE_ID_ANY,

For non registered non-EAL pthread (with a LCORE_ID_ANY _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.

4.3.3. 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.

4.3.4. Control Thread API

It is possible to create Control Threads using the public API rte_ctrl_thread_create(). Those threads can be used for management/infrastructure tasks and are used internally by DPDK for multi process support and interrupt handling.

Those threads will be scheduled on CPUs part of the original process CPU affinity from which the dataplane and service lcores are excluded.

For example, on a 8 CPUs system, starting a dpdk application with -l 2,3 (dataplane cores), then depending on the affinity configuration which can be controlled with tools like taskset (Linux) or cpuset (FreeBSD),

  • with no affinity configuration, the Control Threads will end up on 0-1,4-7 CPUs.
  • with affinity restricted to 2-4, the Control Threads will end up on CPU 4.
  • with affinity restricted to 2-3, the Control Threads will end up on CPU 2 (main lcore, which is the default when no CPU is available).

4.3.5. Known Issues

  • rte_mempool

    The rte_mempool uses a per-lcore cache inside the mempool. For unregistered non-EAL pthreads, rte_lcore_id() will not return a valid number. So for now, when rte_mempool is used with unregistered 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 an unregistered 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-preemptible.

    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 means, use cases involving preemptible pthreads should consider using rte_ring carefully.

    1. It CAN be used for preemptible single-producer and single-consumer use case.
    2. It CAN be used for non-preemptible multi-producer and preemptible single-consumer use case.
    3. It CAN be used for preemptible single-producer and non-preemptible multi-consumer use case.
    4. It MAY be used by preemptible multi-producer and/or preemptible multi-consumer pthreads whose scheduling policy are all SCHED_OTHER(cfs), SCHED_IDLE or SCHED_BATCH. User SHOULD be aware of the performance penalty before using it.
    5. It MUST not be used by multi-producer/consumer pthreads, whose scheduling policies are SCHED_FIFO or SCHED_RR.

    Alternatively, applications can use the lock-free stack mempool handler. When considering this handler, note that:

    • It is currently limited to the aarch64 and x86_64 platforms, because it uses an instruction (16-byte compare-and-swap) that is not yet available on other platforms.
    • It has worse average-case performance than the non-preemptive rte_ring, but software caching (e.g. the mempool cache) can mitigate this by reducing the number of stack accesses.

  • rte_timer

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

  • rte_log

    In unregistered 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 an unregistered non-EAL pthread.

4.3.6. 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

4.4. Malloc

The EAL provides a malloc API to allocate any-sized memory.

The objective of this API is to provide malloc-like functions to allow allocation from hugepage memory and to facilitate application porting. The DPDK API Reference manual describes the available functions.

Typically, these kinds of allocations should not be done in data plane processing because they are slower than pool-based allocation and make use of locks within the allocation and free paths. However, they can be used in configuration code.

Refer to the rte_malloc() function description in the DPDK API Reference manual for more information.

4.4.1. 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.

4.4.2. 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.

4.4.3. Internal Implementation

4.4.3.1. 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.
4.4.3.1.1. 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.
../_images/malloc_heap.svg

Fig. 4.2 Example of a malloc heap and malloc elements within the malloc library

4.4.3.1.2. 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, neighboring 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.
  • dirty - this flag is only meaningful when state is FREE. It indicates that the content of the element is not fully zero-filled. Memory from such blocks must be cleared when requested via rte_zmalloc*(). Dirty elements only appear with --huge-unlink=never.
  • 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.

4.4.3.2. Memory Allocation

On EAL initialization, all preallocated memory segments are setup as part of the malloc heap. This setup involves placing an element header with FREE at the start of each virtually contiguous segment of memory. The FREE element is then added to the free_list for the malloc heap.

This setup also happens whenever memory is allocated at runtime (if supported), in which case newly allocated pages are also added to the heap, merging with any adjacent free segments if there are any.

When an application makes a call to a malloc-like function, the malloc function will first index the lcore_config structure for the calling thread, and determine the NUMA node of that thread. The NUMA node is used to index the array of malloc_heap structures which is passed as a parameter to the heap_alloc() function, along with the requested size, type, alignment and boundary parameters.

The heap_alloc() function will scan the free_list of the heap, and attempt to find a free block suitable for storing data of the requested size, with the requested alignment and boundary constraints.

When a suitable free element has been identified, the pointer to be returned to the user is calculated. The cache-line of memory immediately preceding this pointer is filled with a struct malloc_elem header. Because of alignment and boundary constraints, there could be free space at the start and/or end of the element, resulting in the following behavior:

  1. Check for trailing space. If the trailing space is big enough, i.e. > 128 bytes, then the free element is split. If it is not, then we just ignore it (wasted space).
  2. Check for space at the start of the element. If the space at the start is small, i.e. <=128 bytes, then a pad header is used, and the remaining space is wasted. If, however, the remaining space is greater, then the free element is split.

The advantage of allocating the memory from the end of the existing element is that no adjustment of the free list needs to take place - the existing element on the free list just has its size value adjusted, and the next/previous elements have their “prev”/”next” pointers redirected to the newly created element.

In case when there is not enough memory in the heap to satisfy allocation request, EAL will attempt to allocate more memory from the system (if supported) and, following successful allocation, will retry reserving the memory again. In a multiprocessing scenario, all primary and secondary processes will synchronize their memory maps to ensure that any valid pointer to DPDK memory is guaranteed to be valid at all times in all currently running processes.

Failure to synchronize memory maps in one of the processes will cause allocation to fail, even though some of the processes may have allocated the memory successfully. The memory is not added to the malloc heap unless primary process has ensured that all other processes have mapped this memory successfully.

Any successful allocation event will trigger a callback, for which user applications and other DPDK subsystems can register. Additionally, validation callbacks will be triggered before allocation if the newly allocated memory will exceed threshold set by the user, giving a chance to allow or deny allocation.

Note

Any allocation of new pages has to go through primary process. If the primary process is not active, no memory will be allocated even if it was theoretically possible to do so. This is because primary’s process map acts as an authority on what should or should not be mapped, while each secondary process has its own, local memory map. Secondary processes do not update the shared memory map, they only copy its contents to their local memory map.

4.4.3.3. Freeing Memory

To free an area of memory, the pointer to the start of the data area is passed to the free function. The size of the malloc_elem structure is subtracted from this pointer to get the element header for the block. If this header is of type PAD then the pad length is further subtracted from the pointer to get the proper element header for the entire block.

From this element header, we get pointers to the heap from which the block was allocated and to where it must be freed, as well as the pointer to the previous and next elements. These next and previous elements are then checked to see if they are also FREE and are immediately adjacent to the current one, and if so, they are merged with the current element. This means that we can never have two FREE memory blocks adjacent to one another, as they are always merged into a single block.

If deallocating pages at runtime is supported, and the free element encloses one or more pages, those pages can be deallocated and be removed from the heap. If DPDK was started with command-line parameters for preallocating memory (-m or --socket-mem), then those pages that were allocated at startup will not be deallocated.

Any successful deallocation event will trigger a callback, for which user applications and other DPDK subsystems can register.