1 .. SPDX-License-Identifier: BSD-3-Clause
2 Copyright(c) 2010-2014 Intel Corporation.
4 .. _Environment_Abstraction_Layer:
6 Environment Abstraction Layer
7 =============================
9 The Environment Abstraction Layer (EAL) is responsible for gaining access to low-level resources such as hardware and memory space.
10 It provides a generic interface that hides the environment specifics from the applications and libraries.
11 It is the responsibility of the initialization routine to decide how to allocate these resources
12 (that is, memory space, devices, timers, consoles, and so on).
14 Typical services expected from the EAL are:
16 * DPDK Loading and Launching:
17 The DPDK and its application are linked as a single application and must be loaded by some means.
19 * Core Affinity/Assignment Procedures:
20 The EAL provides mechanisms for assigning execution units to specific cores as well as creating execution instances.
22 * System Memory Reservation:
23 The EAL facilitates the reservation of different memory zones, for example, physical memory areas for device interactions.
25 * Trace and Debug Functions: Logs, dump_stack, panic and so on.
27 * Utility Functions: Spinlocks and atomic counters that are not provided in libc.
29 * CPU Feature Identification: Determine at runtime if a particular feature, for example, IntelĀ® AVX is supported.
30 Determine if the current CPU supports the feature set that the binary was compiled for.
32 * Interrupt Handling: Interfaces to register/unregister callbacks to specific interrupt sources.
34 * Alarm Functions: Interfaces to set/remove callbacks to be run at a specific time.
36 EAL in a Linux-userland Execution Environment
37 ---------------------------------------------
39 In a Linux user space environment, the DPDK application runs as a user-space application using the pthread library.
41 The EAL performs physical memory allocation using mmap() in hugetlbfs (using huge page sizes to increase performance).
42 This memory is exposed to DPDK service layers such as the :ref:`Mempool Library <Mempool_Library>`.
44 At this point, the DPDK services layer will be initialized, then through pthread setaffinity calls,
45 each execution unit will be assigned to a specific logical core to run as a user-level thread.
47 The time reference is provided by the CPU Time-Stamp Counter (TSC) or by the HPET kernel API through a mmap() call.
49 Initialization and Core Launching
50 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
52 Part of the initialization is done by the start function of glibc.
53 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.
54 Then, the main() function is called. The core initialization and launch is done in rte_eal_init() (see the API documentation).
55 It consist of calls to the pthread library (more specifically, pthread_self(), pthread_create(), and pthread_setaffinity_np()).
57 .. _figure_linux_launch:
59 .. figure:: img/linuxapp_launch.*
61 EAL Initialization in a Linux Application Environment
66 Initialization of objects, such as memory zones, rings, memory pools, lpm tables and hash tables,
67 should be done as part of the overall application initialization on the main lcore.
68 The creation and initialization functions for these objects are not multi-thread safe.
69 However, once initialized, the objects themselves can safely be used in multiple threads simultaneously.
74 During the initialization of EAL resources such as hugepage backed memory can be
75 allocated by core components. The memory allocated during ``rte_eal_init()``
76 can be released by calling the ``rte_eal_cleanup()`` function. Refer to the
77 API documentation for details.
82 The Linux EAL allows a multi-process as well as a multi-threaded (pthread) deployment model.
84 :ref:`Multi-process Support <Multi-process_Support>` for more details.
86 Memory Mapping Discovery and Memory Reservation
87 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
89 The allocation of large contiguous physical memory is done using hugepages.
90 The EAL provides an API to reserve named memory zones in this contiguous memory.
91 The physical address of the reserved memory for that memory zone is also returned to the user by the memory zone reservation API.
93 There are two modes in which DPDK memory subsystem can operate: dynamic mode,
94 and legacy mode. Both modes are explained below.
98 Memory reservations done using the APIs provided by rte_malloc
99 are also backed by hugepages unless ``--no-huge`` option is given.
104 Currently, this mode is only supported on Linux and Windows.
106 In this mode, usage of hugepages by DPDK application will grow and shrink based
107 on application's requests. Any memory allocation through ``rte_malloc()``,
108 ``rte_memzone_reserve()`` or other methods, can potentially result in more
109 hugepages being reserved from the system. Similarly, any memory deallocation can
110 potentially result in hugepages being released back to the system.
112 Memory allocated in this mode is not guaranteed to be IOVA-contiguous. If large
113 chunks of IOVA-contiguous are required (with "large" defined as "more than one
114 page"), it is recommended to either use VFIO driver for all physical devices (so
115 that IOVA and VA addresses can be the same, thereby bypassing physical addresses
116 entirely), or use legacy memory mode.
118 For chunks of memory which must be IOVA-contiguous, it is recommended to use
119 ``rte_memzone_reserve()`` function with ``RTE_MEMZONE_IOVA_CONTIG`` flag
120 specified. This way, memory allocator will ensure that, whatever memory mode is
121 in use, either reserved memory will satisfy the requirements, or the allocation
124 There is no need to preallocate any memory at startup using ``-m`` or
125 ``--socket-mem`` command-line parameters, however it is still possible to do so,
126 in which case preallocate memory will be "pinned" (i.e. will never be released
127 by the application back to the system). It will be possible to allocate more
128 hugepages, and deallocate those, but any preallocated pages will not be freed.
129 If neither ``-m`` nor ``--socket-mem`` were specified, no memory will be
130 preallocated, and all memory will be allocated at runtime, as needed.
132 Another available option to use in dynamic memory mode is
133 ``--single-file-segments`` command-line option. This option will put pages in
134 single files (per memseg list), as opposed to creating a file per page. This is
135 normally not needed, but can be useful for use cases like userspace vhost, where
136 there is limited number of page file descriptors that can be passed to VirtIO.
138 If the application (or DPDK-internal code, such as device drivers) wishes to
139 receive notifications about newly allocated memory, it is possible to register
140 for memory event callbacks via ``rte_mem_event_callback_register()`` function.
141 This will call a callback function any time DPDK's memory map has changed.
143 If the application (or DPDK-internal code, such as device drivers) wishes to be
144 notified about memory allocations above specified threshold (and have a chance
145 to deny them), allocation validator callbacks are also available via
146 ``rte_mem_alloc_validator_callback_register()`` function.
148 A default validator callback is provided by EAL, which can be enabled with a
149 ``--socket-limit`` command-line option, for a simple way to limit maximum amount
150 of memory that can be used by DPDK application.
153 Memory subsystem uses DPDK IPC internally, so memory allocations/callbacks
154 and IPC must not be mixed: it is not safe to allocate/free memory inside
155 memory-related or IPC callbacks, and it is not safe to use IPC inside
156 memory-related callbacks. See chapter
157 :ref:`Multi-process Support <Multi-process_Support>` for more details about
163 This mode is enabled by specifying ``--legacy-mem`` command-line switch to the
164 EAL. This switch will have no effect on FreeBSD as FreeBSD only supports
167 This mode mimics historical behavior of EAL. That is, EAL will reserve all
168 memory at startup, sort all memory into large IOVA-contiguous chunks, and will
169 not allow acquiring or releasing hugepages from the system at runtime.
171 If neither ``-m`` nor ``--socket-mem`` were specified, the entire available
172 hugepage memory will be preallocated.
174 Hugepage Allocation Matching
175 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
177 This behavior is enabled by specifying the ``--match-allocations`` command-line
178 switch to the EAL. This switch is Linux-only and not supported with
179 ``--legacy-mem`` nor ``--no-huge``.
181 Some applications using memory event callbacks may require that hugepages be
182 freed exactly as they were allocated. These applications may also require
183 that any allocation from the malloc heap not span across allocations
184 associated with two different memory event callbacks. Hugepage allocation
185 matching can be used by these types of applications to satisfy both of these
186 requirements. This can result in some increased memory usage which is
187 very dependent on the memory allocation patterns of the application.
192 Additional restrictions are present when running in 32-bit mode. In dynamic
193 memory mode, by default maximum of 2 gigabytes of VA space will be preallocated,
194 and all of it will be on main lcore NUMA node unless ``--socket-mem`` flag is
197 In legacy mode, VA space will only be preallocated for segments that were
198 requested (plus padding, to keep IOVA-contiguousness).
200 Maximum Amount of Memory
201 ^^^^^^^^^^^^^^^^^^^^^^^^
203 All possible virtual memory space that can ever be used for hugepage mapping in
204 a DPDK process is preallocated at startup, thereby placing an upper limit on how
205 much memory a DPDK application can have. DPDK memory is stored in segment lists,
206 each segment is strictly one physical page. It is possible to change the amount
207 of virtual memory being preallocated at startup by editing the following config
210 * ``RTE_MAX_MEMSEG_LISTS`` controls how many segment lists can DPDK have
211 * ``RTE_MAX_MEM_MB_PER_LIST`` controls how much megabytes of memory each
212 segment list can address
213 * ``RTE_MAX_MEMSEG_PER_LIST`` controls how many segments each segment list
215 * ``RTE_MAX_MEMSEG_PER_TYPE`` controls how many segments each memory type
216 can have (where "type" is defined as "page size + NUMA node" combination)
217 * ``RTE_MAX_MEM_MB_PER_TYPE`` controls how much megabytes of memory each
218 memory type can address
219 * ``RTE_MAX_MEM_MB`` places a global maximum on the amount of memory
222 Normally, these options do not need to be changed.
226 Preallocated virtual memory is not to be confused with preallocated hugepage
227 memory! All DPDK processes preallocate virtual memory at startup. Hugepages
228 can later be mapped into that preallocated VA space (if dynamic memory mode
229 is enabled), and can optionally be mapped into it at startup.
234 Below is an overview of methods used for each OS to obtain hugepages,
235 explaining why certain limitations and options exist in EAL.
236 See the user guide for a specific OS for configuration details.
238 FreeBSD uses ``contigmem`` kernel module
239 to reserve a fixed number of hugepages at system start,
240 which are mapped by EAL at initialization using a specific ``sysctl()``.
242 Windows EAL allocates hugepages from the OS as needed using Win32 API,
243 so available amount depends on the system load.
244 It uses ``virt2phys`` kernel module to obtain physical addresses,
245 unless running in IOVA-as-VA mode (e.g. forced with ``--iova-mode=va``).
247 Linux allows to select any combination of the following:
249 * use files in hugetlbfs (the default)
250 or anonymous mappings (``--in-memory``);
251 * map each hugepage from its own file (the default)
252 or map multiple hugepages from one big file (``--single-file-segments``).
254 Mapping hugepages from files in hugetlbfs is essential for multi-process,
255 because secondary processes need to map the same hugepages.
256 EAL creates files like ``rtemap_0``
257 in directories specified with ``--huge-dir`` option
258 (or in the mount point for a specific hugepage size).
259 The ``rte`` prefix can be changed using ``--file-prefix``.
260 This may be needed for running multiple primary processes
261 that share a hugetlbfs mount point.
262 Each backing file by default corresponds to one hugepage,
263 it is opened and locked for the entire time the hugepage is used.
264 This may exhaust the number of open files limit (``NOFILE``).
265 See :ref:`segment-file-descriptors` section
266 on how the number of open backing file descriptors can be reduced.
268 In dynamic memory mode, EAL removes a backing hugepage file
269 when all pages mapped from it are freed back to the system.
270 However, backing files may persist after the application terminates
271 in case of a crash or a leak of DPDK memory (e.g. ``rte_free()`` is missing).
272 This reduces the number of hugepages available to other processes
273 as reported by ``/sys/kernel/mm/hugepages/hugepages-*/free_hugepages``.
274 EAL can remove the backing files after opening them for mapping
275 if ``--huge-unlink`` is given to avoid polluting hugetlbfs.
276 However, since it disables multi-process anyway,
277 using anonymous mapping (``--in-memory``) is recommended instead.
279 :ref:`EAL memory allocator <malloc>` relies on hugepages being zero-filled.
280 Hugepages are cleared by the kernel when a file in hugetlbfs or its part
281 is mapped for the first time system-wide
282 to prevent data leaks from previous users of the same hugepage.
283 EAL ensures this behavior by removing existing backing files at startup
284 and by recreating them before opening for mapping (as a precaution).
286 One exception is ``--huge-unlink=never`` mode.
287 It is used to speed up EAL initialization, usually on application restart.
288 Clearing memory constitutes more than 95% of hugepage mapping time.
289 EAL can save it by remapping existing backing files
290 with all the data left in the mapped hugepages ("dirty" memory).
291 Such segments are marked with ``RTE_MEMSEG_FLAG_DIRTY``.
292 Memory allocator detects dirty segments and handles them accordingly,
293 in particular, it clears memory requested with ``rte_zmalloc*()``.
294 In this mode EAL also does not remove a backing file
295 when all pages mapped from it are freed,
296 because they are intended to be reusable at restart.
298 Anonymous mapping does not allow multi-process architecture.
299 This mode does not use hugetlbfs
300 and thus does not require root permissions for memory management
301 (the limit of locked memory amount, ``MEMLOCK``, still applies).
302 It is free of filename conflict and leftover file issues.
303 If ``memfd_create(2)`` is supported both at build and run time,
304 DPDK memory manager can provide file descriptors for memory segments,
305 which are required for VirtIO with vhost-user backend.
306 This can exhaust the number of open files limit (``NOFILE``)
307 despite not creating any visible files.
308 See :ref:`segment-file-descriptors` section
309 on how the number of open file descriptors used by EAL can be reduced.
311 .. _segment-file-descriptors:
313 Segment File Descriptors
314 ^^^^^^^^^^^^^^^^^^^^^^^^
316 On Linux, in most cases, EAL will store segment file descriptors in EAL. This
317 can become a problem when using smaller page sizes due to underlying limitations
318 of ``glibc`` library. For example, Linux API calls such as ``select()`` may not
319 work correctly because ``glibc`` does not support more than certain number of
322 There are two possible solutions for this problem. The recommended solution is
323 to use ``--single-file-segments`` mode, as that mode will not use a file
324 descriptor per each page, and it will keep compatibility with Virtio with
325 vhost-user backend. This option is not available when using ``--legacy-mem``
328 Another option is to use bigger page sizes. Since fewer pages are required to
329 cover the same memory area, fewer file descriptors will be stored internally
332 Support for Externally Allocated Memory
333 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
335 It is possible to use externally allocated memory in DPDK. There are two ways in
336 which using externally allocated memory can work: the malloc heap API's, and
337 manual memory management.
339 + Using heap API's for externally allocated memory
341 Using a set of malloc heap API's is the recommended way to use externally
342 allocated memory in DPDK. In this way, support for externally allocated memory
343 is implemented through overloading the socket ID - externally allocated heaps
344 will have socket ID's that would be considered invalid under normal
345 circumstances. Requesting an allocation to take place from a specified
346 externally allocated memory is a matter of supplying the correct socket ID to
347 DPDK allocator, either directly (e.g. through a call to ``rte_malloc``) or
348 indirectly (through data structure-specific allocation API's such as
349 ``rte_ring_create``). Using these API's also ensures that mapping of externally
350 allocated memory for DMA is also performed on any memory segment that is added
351 to a DPDK malloc heap.
353 Since there is no way DPDK can verify whether memory is available or valid, this
354 responsibility falls on the shoulders of the user. All multiprocess
355 synchronization is also user's responsibility, as well as ensuring that all
356 calls to add/attach/detach/remove memory are done in the correct order. It is
357 not required to attach to a memory area in all processes - only attach to memory
360 The expected workflow is as follows:
362 * Get a pointer to memory area
363 * Create a named heap
364 * Add memory area(s) to the heap
365 - If IOVA table is not specified, IOVA addresses will be assumed to be
366 unavailable, and DMA mappings will not be performed
367 - Other processes must attach to the memory area before they can use it
368 * Get socket ID used for the heap
369 * Use normal DPDK allocation procedures, using supplied socket ID
370 * If memory area is no longer needed, it can be removed from the heap
371 - Other processes must detach from this memory area before it can be removed
372 * If heap is no longer needed, remove it
373 - Socket ID will become invalid and will not be reused
375 For more information, please refer to ``rte_malloc`` API documentation,
376 specifically the ``rte_malloc_heap_*`` family of function calls.
378 + Using externally allocated memory without DPDK API's
380 While using heap API's is the recommended method of using externally allocated
381 memory in DPDK, there are certain use cases where the overhead of DPDK heap API
382 is undesirable - for example, when manual memory management is performed on an
383 externally allocated area. To support use cases where externally allocated
384 memory will not be used as part of normal DPDK workflow, there is also another
385 set of API's under the ``rte_extmem_*`` namespace.
387 These API's are (as their name implies) intended to allow registering or
388 unregistering externally allocated memory to/from DPDK's internal page table, to
389 allow API's like ``rte_mem_virt2memseg`` etc. to work with externally allocated
390 memory. Memory added this way will not be available for any regular DPDK
391 allocators; DPDK will leave this memory for the user application to manage.
393 The expected workflow is as follows:
395 * Get a pointer to memory area
396 * Register memory within DPDK
397 - If IOVA table is not specified, IOVA addresses will be assumed to be
399 - Other processes must attach to the memory area before they can use it
400 * Perform DMA mapping with ``rte_dev_dma_map`` if needed
401 * Use the memory area in your application
402 * If memory area is no longer needed, it can be unregistered
403 - If the area was mapped for DMA, unmapping must be performed before
405 - Other processes must detach from the memory area before it can be
408 Since these externally allocated memory areas will not be managed by DPDK, it is
409 therefore up to the user application to decide how to use them and what to do
410 with them once they're registered.
412 Per-lcore and Shared Variables
413 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
417 lcore refers to a logical execution unit of the processor, sometimes called a hardware *thread*.
419 Shared variables are the default behavior.
420 Per-lcore variables are implemented using *Thread Local Storage* (TLS) to provide per-thread local storage.
425 A logging API is provided by EAL.
426 By default, in a Linux application, logs are sent to syslog and also to the console.
427 However, the log function can be overridden by the user to use a different logging mechanism.
429 Trace and Debug Functions
430 ^^^^^^^^^^^^^^^^^^^^^^^^^
432 There are some debug functions to dump the stack in glibc.
433 The rte_panic() function can voluntarily provoke a SIG_ABORT,
434 which can trigger the generation of a core file, readable by gdb.
436 CPU Feature Identification
437 ~~~~~~~~~~~~~~~~~~~~~~~~~~
439 The EAL can query the CPU at runtime (using the rte_cpu_get_features() function) to determine which CPU features are available.
441 User Space Interrupt Event
442 ~~~~~~~~~~~~~~~~~~~~~~~~~~
444 + User Space Interrupt and Alarm Handling in Host Thread
446 The EAL creates a host thread to poll the UIO device file descriptors to detect the interrupts.
447 Callbacks can be registered or unregistered by the EAL functions for a specific interrupt event
448 and are called in the host thread asynchronously.
449 The EAL also allows timed callbacks to be used in the same way as for NIC interrupts.
453 In DPDK PMD, the only interrupts handled by the dedicated host thread are those for link status change
454 (link up and link down notification) and for sudden device removal.
459 The receive and transmit routines provided by each PMD don't limit themselves to execute in polling thread mode.
460 To ease the idle polling with tiny throughput, it's useful to pause the polling and wait until the wake-up event happens.
461 The RX interrupt is the first choice to be such kind of wake-up event, but probably won't be the only one.
463 EAL provides the event APIs for this event-driven thread mode.
464 Taking Linux as an example, the implementation relies on epoll. Each thread can monitor an epoll instance
465 in which all the wake-up events' file descriptors are added. The event file descriptors are created and mapped to
466 the interrupt vectors according to the UIO/VFIO spec.
467 From FreeBSD's perspective, kqueue is the alternative way, but not implemented yet.
469 EAL initializes the mapping between event file descriptors and interrupt vectors, while each device initializes the mapping
470 between interrupt vectors and queues. In this way, EAL actually is unaware of the interrupt cause on the specific vector.
471 The eth_dev driver takes responsibility to program the latter mapping.
475 Per queue RX interrupt event is only allowed in VFIO which supports multiple MSI-X vector. In UIO, the RX interrupt
476 together with other interrupt causes shares the same vector. In this case, when RX interrupt and LSC(link status change)
477 interrupt are both enabled(intr_conf.lsc == 1 && intr_conf.rxq == 1), only the former is capable.
479 The RX interrupt are controlled/enabled/disabled by ethdev APIs - 'rte_eth_dev_rx_intr_*'. They return failure if the PMD
480 hasn't support them yet. The intr_conf.rxq flag is used to turn on the capability of RX interrupt per device.
482 + Device Removal Event
484 This event is triggered by a device being removed at a bus level. Its
485 underlying resources may have been made unavailable (i.e. PCI mappings
486 unmapped). The PMD must make sure that on such occurrence, the application can
487 still safely use its callbacks.
489 This event can be subscribed to in the same way one would subscribe to a link
490 status change event. The execution context is thus the same, i.e. it is the
491 dedicated interrupt host thread.
493 Considering this, it is likely that an application would want to close a
494 device having emitted a Device Removal Event. In such case, calling
495 ``rte_eth_dev_close()`` can trigger it to unregister its own Device Removal Event
496 callback. Care must be taken not to close the device from the interrupt handler
497 context. It is necessary to reschedule such closing operation.
502 The EAL PCI device block list functionality can be used to mark certain NIC ports as unavailable,
503 so they are ignored by the DPDK.
504 The ports to be blocked are identified using the PCIe* description (Domain:Bus:Device.Function).
509 Locks and atomic operations are per-architecture (i686 and x86_64).
514 IOVA Mode is selected by considering what the current usable Devices on the
515 system require and/or support.
517 On FreeBSD, RTE_IOVA_PA is always the default. On Linux, the IOVA mode is
518 detected based on a 2-step heuristic detailed below.
520 For the first step, EAL asks each bus its requirement in terms of IOVA mode
521 and decides on a preferred IOVA mode.
523 - if all buses report RTE_IOVA_PA, then the preferred IOVA mode is RTE_IOVA_PA,
524 - if all buses report RTE_IOVA_VA, then the preferred IOVA mode is RTE_IOVA_VA,
525 - if all buses report RTE_IOVA_DC, no bus expressed a preference, then the
526 preferred mode is RTE_IOVA_DC,
527 - if the buses disagree (at least one wants RTE_IOVA_PA and at least one wants
528 RTE_IOVA_VA), then the preferred IOVA mode is RTE_IOVA_DC (see below with the
529 check on Physical Addresses availability),
531 If the buses have expressed no preference on which IOVA mode to pick, then a
532 default is selected using the following logic:
534 - if physical addresses are not available, RTE_IOVA_VA mode is used
535 - if /sys/kernel/iommu_groups is not empty, RTE_IOVA_VA mode is used
536 - otherwise, RTE_IOVA_PA mode is used
538 In the case when the buses had disagreed on their preferred IOVA mode, part of
539 the buses won't work because of this decision.
541 The second step checks if the preferred mode complies with the Physical
542 Addresses availability since those are only available to root user in recent
543 kernels. Namely, if the preferred mode is RTE_IOVA_PA but there is no access to
544 Physical Addresses, then EAL init fails early, since later probing of the
545 devices would fail anyway.
549 The RTE_IOVA_VA mode is preferred as the default in most cases for the
552 - All drivers are expected to work in RTE_IOVA_VA mode, irrespective of
553 physical address availability.
554 - By default, the mempool, first asks for IOVA-contiguous memory using
555 ``RTE_MEMZONE_IOVA_CONTIG``. This is slow in RTE_IOVA_PA mode and it may
556 affect the application boot time.
557 - It is easy to enable large amount of IOVA-contiguous memory use cases
558 with IOVA in VA mode.
560 It is expected that all PCI drivers work in both RTE_IOVA_PA and
563 If a PCI driver does not support RTE_IOVA_PA mode, the
564 ``RTE_PCI_DRV_NEED_IOVA_AS_VA`` flag is used to dictate that this PCI
565 driver can only work in RTE_IOVA_VA mode.
567 When the KNI kernel module is detected, RTE_IOVA_PA mode is preferred as a
568 performance penalty is expected in RTE_IOVA_VA mode.
570 IOVA Mode Configuration
571 ~~~~~~~~~~~~~~~~~~~~~~~
573 Auto detection of the IOVA mode, based on probing the bus and IOMMU configuration, may not report
574 the desired addressing mode when virtual devices that are not directly attached to the bus are present.
575 To facilitate forcing the IOVA mode to a specific value the EAL command line option ``--iova-mode`` can
576 be used to select either physical addressing('pa') or virtual addressing('va').
578 .. _max_simd_bitwidth:
584 The EAL provides a single setting to limit the max SIMD bitwidth used by DPDK,
585 which is used in determining the vector path, if any, chosen by a component.
586 The value can be set at runtime by an application using the
587 'rte_vect_set_max_simd_bitwidth(uint16_t bitwidth)' function,
588 which should only be called once at initialization, before EAL init.
589 The value can be overridden by the user using the EAL command-line option '--force-max-simd-bitwidth'.
591 When choosing a vector path, along with checking the CPU feature support,
592 the value of the max SIMD bitwidth must also be checked, and can be retrieved using the
593 'rte_vect_get_max_simd_bitwidth()' function.
594 The value should be compared against the enum values for accepted max SIMD bitwidths:
598 enum rte_vect_max_simd {
599 RTE_VECT_SIMD_DISABLED = 64,
600 RTE_VECT_SIMD_128 = 128,
601 RTE_VECT_SIMD_256 = 256,
602 RTE_VECT_SIMD_512 = 512,
603 RTE_VECT_SIMD_MAX = INT16_MAX + 1,
606 if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512)
607 /* Take AVX-512 vector path */
608 else if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
609 /* Take AVX2 vector path */
612 Memory Segments and Memory Zones (memzone)
613 ------------------------------------------
615 The mapping of physical memory is provided by this feature in the EAL.
616 As physical memory can have gaps, the memory is described in a table of descriptors,
617 and each descriptor (called rte_memseg ) describes a physical page.
619 On top of this, the memzone allocator's role is to reserve contiguous portions of physical memory.
620 These zones are identified by a unique name when the memory is reserved.
622 The rte_memzone descriptors are also located in the configuration structure.
623 This structure is accessed using rte_eal_get_configuration().
624 The lookup (by name) of a memory zone returns a descriptor containing the physical address of the memory zone.
626 Memory zones can be reserved with specific start address alignment by supplying the align parameter
627 (by default, they are aligned to cache line size).
628 The alignment value should be a power of two and not less than the cache line size (64 bytes).
629 Memory zones can also be reserved from either 2 MB or 1 GB hugepages, provided that both are available on the system.
631 Both memsegs and memzones are stored using ``rte_fbarray`` structures. Please
632 refer to *DPDK API Reference* for more information.
638 DPDK usually pins one pthread per core to avoid the overhead of task switching.
639 This allows for significant performance gains, but lacks flexibility and is not always efficient.
641 Power management helps to improve the CPU efficiency by limiting the CPU runtime frequency.
642 However, alternately it is possible to utilize the idle cycles available to take advantage of
643 the full capability of the CPU.
645 By taking advantage of cgroup, the CPU utilization quota can be simply assigned.
646 This gives another way to improve the CPU efficiency, however, there is a prerequisite;
647 DPDK must handle the context switching between multiple pthreads per core.
649 For further flexibility, it is useful to set pthread affinity not only to a CPU but to a CPU set.
651 EAL pthread and lcore Affinity
652 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
654 The term "lcore" refers to an EAL thread, which is really a Linux/FreeBSD pthread.
655 "EAL pthreads" are created and managed by EAL and execute the tasks issued by *remote_launch*.
656 In each EAL pthread, there is a TLS (Thread Local Storage) called *_lcore_id* for unique identification.
657 As EAL pthreads usually bind 1:1 to the physical CPU, the *_lcore_id* is typically equal to the CPU ID.
659 When using multiple pthreads, however, the binding is no longer always 1:1 between an EAL pthread and a specified physical CPU.
660 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.
661 For this reason, there is an EAL long option '--lcores' defined to assign the CPU affinity of lcores.
662 For a specified lcore ID or ID group, the option allows setting the CPU set for that EAL pthread.
665 --lcores='<lcore_set>[@cpu_set][,<lcore_set>[@cpu_set],...]'
667 'lcore_set' and 'cpu_set' can be a single number, range or a group.
669 A number is a "digit([0-9]+)"; a range is "<number>-<number>"; a group is "(<number|range>[,<number|range>,...])".
671 If a '\@cpu_set' value is not supplied, the value of 'cpu_set' will default to the value of 'lcore_set'.
675 For example, "--lcores='1,2@(5-7),(3-5)@(0,2),(0,6),7-8'" which means start 9 EAL thread;
676 lcore 0 runs on cpuset 0x41 (cpu 0,6);
677 lcore 1 runs on cpuset 0x2 (cpu 1);
678 lcore 2 runs on cpuset 0xe0 (cpu 5,6,7);
679 lcore 3,4,5 runs on cpuset 0x5 (cpu 0,2);
680 lcore 6 runs on cpuset 0x41 (cpu 0,6);
681 lcore 7 runs on cpuset 0x80 (cpu 7);
682 lcore 8 runs on cpuset 0x100 (cpu 8).
684 Using this option, for each given lcore ID, the associated CPUs can be assigned.
685 It's also compatible with the pattern of corelist('-l') option.
687 non-EAL pthread support
688 ~~~~~~~~~~~~~~~~~~~~~~~
690 It is possible to use the DPDK execution context with any user pthread (aka. non-EAL pthreads).
691 There are two kinds of non-EAL pthreads:
693 - a registered non-EAL pthread with a valid *_lcore_id* that was successfully assigned by calling ``rte_thread_register()``,
694 - a non registered non-EAL pthread with a LCORE_ID_ANY,
696 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).
698 All these impacts are mentioned in :ref:`known_issue_label` section.
703 There are two public APIs ``rte_thread_set_affinity()`` and ``rte_thread_get_affinity()`` introduced for threads.
704 When they're used in any pthread context, the Thread Local Storage(TLS) will be set/get.
706 Those TLS include *_cpuset* and *_socket_id*:
708 * *_cpuset* stores the CPUs bitmap to which the pthread is affinitized.
710 * *_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.
716 It is possible to create Control Threads using the public API
717 ``rte_ctrl_thread_create()``.
718 Those threads can be used for management/infrastructure tasks and are used
719 internally by DPDK for multi process support and interrupt handling.
721 Those threads will be scheduled on CPUs part of the original process CPU
722 affinity from which the dataplane and service lcores are excluded.
724 For example, on a 8 CPUs system, starting a dpdk application with -l 2,3
725 (dataplane cores), then depending on the affinity configuration which can be
726 controlled with tools like taskset (Linux) or cpuset (FreeBSD),
728 - with no affinity configuration, the Control Threads will end up on
730 - with affinity restricted to 2-4, the Control Threads will end up on
732 - with affinity restricted to 2-3, the Control Threads will end up on
733 CPU 2 (main lcore, which is the default when no CPU is available).
735 .. _known_issue_label:
742 The rte_mempool uses a per-lcore cache inside the mempool.
743 For unregistered non-EAL pthreads, ``rte_lcore_id()`` will not return a valid number.
744 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.
745 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.
749 rte_ring supports multi-producer enqueue and multi-consumer dequeue.
750 However, it is non-preemptive, this has a knock on effect of making rte_mempool non-preemptible.
754 The "non-preemptive" constraint means:
756 - a pthread doing multi-producers enqueues on a given ring must not
757 be preempted by another pthread doing a multi-producer enqueue on
759 - a pthread doing multi-consumers dequeues on a given ring must not
760 be preempted by another pthread doing a multi-consumer dequeue on
763 Bypassing this constraint may cause the 2nd pthread to spin until the 1st one is scheduled again.
764 Moreover, if the 1st pthread is preempted by a context that has an higher priority, it may even cause a dead lock.
766 This means, use cases involving preemptible pthreads should consider using rte_ring carefully.
768 1. It CAN be used for preemptible single-producer and single-consumer use case.
770 2. It CAN be used for non-preemptible multi-producer and preemptible single-consumer use case.
772 3. It CAN be used for preemptible single-producer and non-preemptible multi-consumer use case.
774 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.
776 5. It MUST not be used by multi-producer/consumer pthreads, whose scheduling policies are SCHED_FIFO or SCHED_RR.
778 Alternatively, applications can use the lock-free stack mempool handler. When
779 considering this handler, note that:
781 - It is currently limited to the aarch64 and x86_64 platforms, because it uses
782 an instruction (16-byte compare-and-swap) that is not yet available on other
784 - It has worse average-case performance than the non-preemptive rte_ring, but
785 software caching (e.g. the mempool cache) can mitigate this by reducing the
786 number of stack accesses.
790 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.
794 In unregistered non-EAL pthreads, there is no per thread loglevel and logtype, global loglevels are used.
798 The debug statistics of rte_ring, rte_mempool and rte_timer are not supported in an unregistered non-EAL pthread.
803 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).
804 We expect only 50% of CPU spend on packet IO.
806 .. code-block:: console
808 mkdir /sys/fs/cgroup/cpu/pkt_io
809 mkdir /sys/fs/cgroup/cpuset/pkt_io
811 echo $cpu > /sys/fs/cgroup/cpuset/cpuset.cpus
813 echo $t0 > /sys/fs/cgroup/cpu/pkt_io/tasks
814 echo $t0 > /sys/fs/cgroup/cpuset/pkt_io/tasks
816 echo $t1 > /sys/fs/cgroup/cpu/pkt_io/tasks
817 echo $t1 > /sys/fs/cgroup/cpuset/pkt_io/tasks
819 cd /sys/fs/cgroup/cpu/pkt_io
820 echo 100000 > pkt_io/cpu.cfs_period_us
821 echo 50000 > pkt_io/cpu.cfs_quota_us
828 The EAL provides a malloc API to allocate any-sized memory.
830 The objective of this API is to provide malloc-like functions to allow
831 allocation from hugepage memory and to facilitate application porting.
832 The *DPDK API Reference* manual describes the available functions.
834 Typically, these kinds of allocations should not be done in data plane
835 processing because they are slower than pool-based allocation and make
836 use of locks within the allocation and free paths.
837 However, they can be used in configuration code.
839 Refer to the rte_malloc() function description in the *DPDK API Reference*
840 manual for more information.
843 Alignment and NUMA Constraints
844 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
846 The rte_malloc() takes an align argument that can be used to request a memory
847 area that is aligned on a multiple of this value (which must be a power of two).
849 On systems with NUMA support, a call to the rte_malloc() function will return
850 memory that has been allocated on the NUMA socket of the core which made the call.
851 A set of APIs is also provided, to allow memory to be explicitly allocated on a
852 NUMA socket directly, or by allocated on the NUMA socket where another core is
853 located, in the case where the memory is to be used by a logical core other than
854 on the one doing the memory allocation.
859 This API is meant to be used by an application that requires malloc-like
860 functions at initialization time.
862 For allocating/freeing data at runtime, in the fast-path of an application,
863 the memory pool library should be used instead.
865 Internal Implementation
866 ~~~~~~~~~~~~~~~~~~~~~~~
871 There are two data structure types used internally in the malloc library:
873 * struct malloc_heap - used to track free space on a per-socket basis
875 * struct malloc_elem - the basic element of allocation and free-space
876 tracking inside the library.
878 Structure: malloc_heap
879 """"""""""""""""""""""
881 The malloc_heap structure is used to manage free space on a per-socket basis.
882 Internally, there is one heap structure per NUMA node, which allows us to
883 allocate memory to a thread based on the NUMA node on which this thread runs.
884 While this does not guarantee that the memory will be used on that NUMA node,
885 it is no worse than a scheme where the memory is always allocated on a fixed
888 The key fields of the heap structure and their function are described below
889 (see also diagram above):
891 * lock - the lock field is needed to synchronize access to the heap.
892 Given that the free space in the heap is tracked using a linked list,
893 we need a lock to prevent two threads manipulating the list at the same time.
895 * free_head - this points to the first element in the list of free nodes for
898 * first - this points to the first element in the heap.
900 * last - this points to the last element in the heap.
902 .. _figure_malloc_heap:
904 .. figure:: img/malloc_heap.*
906 Example of a malloc heap and malloc elements within the malloc library
911 Structure: malloc_elem
912 """"""""""""""""""""""
914 The malloc_elem structure is used as a generic header structure for various
916 It is used in two different ways - all shown in the diagram above:
918 #. As a header on a block of free or allocated memory - normal case
920 #. As a padding header inside a block of memory
922 The most important fields in the structure and how they are used are described below.
924 Malloc heap is a doubly-linked list, where each element keeps track of its
925 previous and next elements. Due to the fact that hugepage memory can come and
926 go, neighboring malloc elements may not necessarily be adjacent in memory.
927 Also, since a malloc element may span multiple pages, its contents may not
928 necessarily be IOVA-contiguous either - each malloc element is only guaranteed
929 to be virtually contiguous.
933 If the usage of a particular field in one of the above three usages is not
934 described, the field can be assumed to have an undefined value in that
935 situation, for example, for padding headers only the "state" and "pad"
936 fields have valid values.
938 * heap - this pointer is a reference back to the heap structure from which
939 this block was allocated.
940 It is used for normal memory blocks when they are being freed, to add the
941 newly-freed block to the heap's free-list.
943 * prev - this pointer points to previous header element/block in memory. When
944 freeing a block, this pointer is used to reference the previous block to
945 check if that block is also free. If so, and the two blocks are immediately
946 adjacent to each other, then the two free blocks are merged to form a single
949 * next - this pointer points to next header element/block in memory. When
950 freeing a block, this pointer is used to reference the next block to check
951 if that block is also free. If so, and the two blocks are immediately
952 adjacent to each other, then the two free blocks are merged to form a single
955 * free_list - this is a structure pointing to previous and next elements in
956 this heap's free list.
957 It is only used in normal memory blocks; on ``malloc()`` to find a suitable
958 free block to allocate and on ``free()`` to add the newly freed element to
961 * state - This field can have one of three values: ``FREE``, ``BUSY`` or
963 The former two are to indicate the allocation state of a normal memory block
964 and the latter is to indicate that the element structure is a dummy structure
965 at the end of the start-of-block padding, i.e. where the start of the data
966 within a block is not at the start of the block itself, due to alignment
968 In that case, the pad header is used to locate the actual malloc element
969 header for the block.
971 * dirty - this flag is only meaningful when ``state`` is ``FREE``.
972 It indicates that the content of the element is not fully zero-filled.
973 Memory from such blocks must be cleared when requested via ``rte_zmalloc*()``.
974 Dirty elements only appear with ``--huge-unlink=never``.
976 * pad - this holds the length of the padding present at the start of the block.
977 In the case of a normal block header, it is added to the address of the end
978 of the header to give the address of the start of the data area, i.e. the
979 value passed back to the application on a malloc.
980 Within a dummy header inside the padding, this same value is stored, and is
981 subtracted from the address of the dummy header to yield the address of the
984 * size - the size of the data block, including the header itself.
989 On EAL initialization, all preallocated memory segments are setup as part of the
990 malloc heap. This setup involves placing an :ref:`element header<malloc_elem>`
991 with ``FREE`` at the start of each virtually contiguous segment of memory.
992 The ``FREE`` element is then added to the ``free_list`` for the malloc heap.
994 This setup also happens whenever memory is allocated at runtime (if supported),
995 in which case newly allocated pages are also added to the heap, merging with any
996 adjacent free segments if there are any.
998 When an application makes a call to a malloc-like function, the malloc function
999 will first index the ``lcore_config`` structure for the calling thread, and
1000 determine the NUMA node of that thread.
1001 The NUMA node is used to index the array of ``malloc_heap`` structures which is
1002 passed as a parameter to the ``heap_alloc()`` function, along with the
1003 requested size, type, alignment and boundary parameters.
1005 The ``heap_alloc()`` function will scan the free_list of the heap, and attempt
1006 to find a free block suitable for storing data of the requested size, with the
1007 requested alignment and boundary constraints.
1009 When a suitable free element has been identified, the pointer to be returned
1010 to the user is calculated.
1011 The cache-line of memory immediately preceding this pointer is filled with a
1012 struct malloc_elem header.
1013 Because of alignment and boundary constraints, there could be free space at
1014 the start and/or end of the element, resulting in the following behavior:
1016 #. Check for trailing space.
1017 If the trailing space is big enough, i.e. > 128 bytes, then the free element
1019 If it is not, then we just ignore it (wasted space).
1021 #. Check for space at the start of the element.
1022 If the space at the start is small, i.e. <=128 bytes, then a pad header is
1023 used, and the remaining space is wasted.
1024 If, however, the remaining space is greater, then the free element is split.
1026 The advantage of allocating the memory from the end of the existing element is
1027 that no adjustment of the free list needs to take place - the existing element
1028 on the free list just has its size value adjusted, and the next/previous elements
1029 have their "prev"/"next" pointers redirected to the newly created element.
1031 In case when there is not enough memory in the heap to satisfy allocation
1032 request, EAL will attempt to allocate more memory from the system (if supported)
1033 and, following successful allocation, will retry reserving the memory again. In
1034 a multiprocessing scenario, all primary and secondary processes will synchronize
1035 their memory maps to ensure that any valid pointer to DPDK memory is guaranteed
1036 to be valid at all times in all currently running processes.
1038 Failure to synchronize memory maps in one of the processes will cause allocation
1039 to fail, even though some of the processes may have allocated the memory
1040 successfully. The memory is not added to the malloc heap unless primary process
1041 has ensured that all other processes have mapped this memory successfully.
1043 Any successful allocation event will trigger a callback, for which user
1044 applications and other DPDK subsystems can register. Additionally, validation
1045 callbacks will be triggered before allocation if the newly allocated memory will
1046 exceed threshold set by the user, giving a chance to allow or deny allocation.
1050 Any allocation of new pages has to go through primary process. If the
1051 primary process is not active, no memory will be allocated even if it was
1052 theoretically possible to do so. This is because primary's process map acts
1053 as an authority on what should or should not be mapped, while each secondary
1054 process has its own, local memory map. Secondary processes do not update the
1055 shared memory map, they only copy its contents to their local memory map.
1060 To free an area of memory, the pointer to the start of the data area is passed
1061 to the free function.
1062 The size of the ``malloc_elem`` structure is subtracted from this pointer to get
1063 the element header for the block.
1064 If this header is of type ``PAD`` then the pad length is further subtracted from
1065 the pointer to get the proper element header for the entire block.
1067 From this element header, we get pointers to the heap from which the block was
1068 allocated and to where it must be freed, as well as the pointer to the previous
1069 and next elements. These next and previous elements are then checked to see if
1070 they are also ``FREE`` and are immediately adjacent to the current one, and if
1071 so, they are merged with the current element. This means that we can never have
1072 two ``FREE`` memory blocks adjacent to one another, as they are always merged
1073 into a single block.
1075 If deallocating pages at runtime is supported, and the free element encloses
1076 one or more pages, those pages can be deallocated and be removed from the heap.
1077 If DPDK was started with command-line parameters for preallocating memory
1078 (``-m`` or ``--socket-mem``), then those pages that were allocated at startup
1079 will not be deallocated.
1081 Any successful deallocation event will trigger a callback, for which user
1082 applications and other DPDK subsystems can register.