xref: /dpdk/doc/guides/prog_guide/env_abstraction_layer.rst (revision 985130369be32dd68ca104c1ccc86716f6e2bb7b)
1..  SPDX-License-Identifier: BSD-3-Clause
2    Copyright(c) 2010-2014 Intel Corporation.
3
4Environment Abstraction Layer (EAL) Library
5===========================================
6
7The Environment Abstraction Layer (EAL) is responsible for gaining access to low-level resources such as hardware and memory space.
8It provides a generic interface that hides the environment specifics from the applications and libraries.
9It is the responsibility of the initialization routine to decide how to allocate these resources
10(that is, memory space, devices, timers, consoles, and so on).
11
12Typical services expected from the EAL are:
13
14*   DPDK Loading and Launching:
15    The DPDK and its application are linked as a single application and must be loaded by some means.
16
17*   Core Affinity/Assignment Procedures:
18    The EAL provides mechanisms for assigning execution units to specific cores as well as creating execution instances.
19
20*   System Memory Reservation:
21    The EAL facilitates the reservation of different memory zones, for example, physical memory areas for device interactions.
22
23*   Trace and Debug Functions: Logs, dump_stack, panic and so on.
24
25*   Utility Functions: Spinlocks and atomic counters that are not provided in libc.
26
27*   CPU Feature Identification: Determine at runtime if a particular feature, for example, Intel® AVX is supported.
28    Determine if the current CPU supports the feature set that the binary was compiled for.
29
30*   Interrupt Handling: Interfaces to register/unregister callbacks to specific interrupt sources.
31
32*   Alarm Functions: Interfaces to set/remove callbacks to be run at a specific time.
33
34EAL in a Linux-userland Execution Environment
35---------------------------------------------
36
37In a Linux user space environment, the DPDK application runs as a user-space application using the pthread library.
38
39The EAL performs physical memory allocation using mmap() in hugetlbfs (using huge page sizes to increase performance).
40This memory is exposed to DPDK service layers such as the :doc:`mempool_lib`.
41
42At this point, the DPDK services layer will be initialized, then through pthread setaffinity calls,
43each execution unit will be assigned to a specific logical core to run as a user-level thread.
44
45The time reference is provided by the CPU Time-Stamp Counter (TSC) or by the HPET kernel API through a mmap() call.
46
47Initialization and Core Launching
48~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
49
50Part of the initialization is done by the start function of glibc.
51A check is also performed at initialization time to ensure that the micro architecture type chosen in the config file is supported by the CPU.
52Then, the main() function is called. The core initialization and launch is done in rte_eal_init() (see the API documentation).
53It consist of calls to the pthread library (more specifically, pthread_self(), pthread_create(), and pthread_setaffinity_np()).
54
55.. _figure_linux_launch:
56
57.. figure:: img/linuxapp_launch.*
58
59   EAL Initialization in a Linux Application Environment
60
61
62.. note::
63
64    Initialization of objects, such as memory zones, rings, memory pools, lpm tables and hash tables,
65    should be done as part of the overall application initialization on the main lcore.
66    The creation and initialization functions for these objects are not multi-thread safe.
67    However, once initialized, the objects themselves can safely be used in multiple threads simultaneously.
68
69Shutdown and Cleanup
70~~~~~~~~~~~~~~~~~~~~
71
72During the initialization of EAL resources such as hugepage backed memory can be
73allocated by core components.  The memory allocated during ``rte_eal_init()``
74can be released by calling the ``rte_eal_cleanup()`` function. Refer to the
75API documentation for details.
76
77Multi-process Support
78~~~~~~~~~~~~~~~~~~~~~
79
80The Linux EAL allows a multi-process as well as a multi-threaded (pthread) deployment model.
81See chapter :doc:`multi_proc_support` for more details.
82
83Memory Mapping Discovery and Memory Reservation
84~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
85
86The allocation of large contiguous physical memory is done using hugepages.
87The EAL provides an API to reserve named memory zones in this contiguous memory.
88The physical address of the reserved memory for that memory zone is also returned to the user by the memory zone reservation API.
89
90There are two modes in which DPDK memory subsystem can operate: dynamic mode,
91and legacy mode. Both modes are explained below.
92
93.. note::
94
95    Memory reservations done using the APIs provided by rte_malloc
96    are also backed by hugepages unless ``--no-huge`` option is given.
97
98Dynamic Memory Mode
99^^^^^^^^^^^^^^^^^^^
100
101Currently, this mode is only supported on Linux and Windows.
102
103In this mode, usage of hugepages by DPDK application will grow and shrink based
104on application's requests. Any memory allocation through ``rte_malloc()``,
105``rte_memzone_reserve()`` or other methods, can potentially result in more
106hugepages being reserved from the system. Similarly, any memory deallocation can
107potentially result in hugepages being released back to the system.
108
109Memory allocated in this mode is not guaranteed to be IOVA-contiguous. If large
110chunks of IOVA-contiguous are required (with "large" defined as "more than one
111page"), it is recommended to either use VFIO driver for all physical devices (so
112that IOVA and VA addresses can be the same, thereby bypassing physical addresses
113entirely), or use legacy memory mode.
114
115For chunks of memory which must be IOVA-contiguous, it is recommended to use
116``rte_memzone_reserve()`` function with ``RTE_MEMZONE_IOVA_CONTIG`` flag
117specified. This way, memory allocator will ensure that, whatever memory mode is
118in use, either reserved memory will satisfy the requirements, or the allocation
119will fail.
120
121There is no need to preallocate any memory at startup using ``-m`` or
122``--socket-mem`` command-line parameters, however it is still possible to do so,
123in which case preallocate memory will be "pinned" (i.e. will never be released
124by the application back to the system). It will be possible to allocate more
125hugepages, and deallocate those, but any preallocated pages will not be freed.
126If neither ``-m`` nor ``--socket-mem`` were specified, no memory will be
127preallocated, and all memory will be allocated at runtime, as needed.
128
129Another available option to use in dynamic memory mode is
130``--single-file-segments`` command-line option. This option will put pages in
131single files (per memseg list), as opposed to creating a file per page. This is
132normally not needed, but can be useful for use cases like userspace vhost, where
133there is limited number of page file descriptors that can be passed to VirtIO.
134
135If the application (or DPDK-internal code, such as device drivers) wishes to
136receive notifications about newly allocated memory, it is possible to register
137for memory event callbacks via ``rte_mem_event_callback_register()`` function.
138This will call a callback function any time DPDK's memory map has changed.
139
140If the application (or DPDK-internal code, such as device drivers) wishes to be
141notified about memory allocations above specified threshold (and have a chance
142to deny them), allocation validator callbacks are also available via
143``rte_mem_alloc_validator_callback_register()`` function.
144
145A default validator callback is provided by EAL, which can be enabled with a
146``--socket-limit`` command-line option, for a simple way to limit maximum amount
147of memory that can be used by DPDK application.
148
149.. warning::
150    Memory subsystem uses DPDK IPC internally, so memory allocations/callbacks
151    and IPC must not be mixed: it is not safe to allocate/free memory inside
152    memory-related or IPC callbacks, and it is not safe to use IPC inside
153    memory-related callbacks. See chapter
154    :doc:`multi_proc_support` for more details about DPDK IPC.
155
156Legacy Memory Mode
157^^^^^^^^^^^^^^^^^^
158
159This mode is enabled by specifying ``--legacy-mem`` command-line switch to the
160EAL. This switch will have no effect on FreeBSD as FreeBSD only supports
161legacy mode anyway.
162
163This mode mimics historical behavior of EAL. That is, EAL will reserve all
164memory at startup, sort all memory into large IOVA-contiguous chunks, and will
165not allow acquiring or releasing hugepages from the system at runtime.
166
167If neither ``-m`` nor ``--socket-mem`` were specified, the entire available
168hugepage memory will be preallocated.
169
170Hugepage Allocation Matching
171^^^^^^^^^^^^^^^^^^^^^^^^^^^^
172
173This behavior is enabled by specifying the ``--match-allocations`` command-line
174switch to the EAL. This switch is Linux-only and not supported with
175``--legacy-mem`` nor ``--no-huge``.
176
177Some applications using memory event callbacks may require that hugepages be
178freed exactly as they were allocated. These applications may also require
179that any allocation from the malloc heap not span across allocations
180associated with two different memory event callbacks. Hugepage allocation
181matching can be used by these types of applications to satisfy both of these
182requirements. This can result in some increased memory usage which is
183very dependent on the memory allocation patterns of the application.
184
18532-bit Support
186^^^^^^^^^^^^^^
187
188Additional restrictions are present when running in 32-bit mode. In dynamic
189memory mode, by default maximum of 2 gigabytes of VA space will be preallocated,
190and all of it will be on main lcore NUMA node unless ``--socket-mem`` flag is
191used.
192
193In legacy mode, VA space will only be preallocated for segments that were
194requested (plus padding, to keep IOVA-contiguousness).
195
196Maximum Amount of Memory
197^^^^^^^^^^^^^^^^^^^^^^^^
198
199All possible virtual memory space that can ever be used for hugepage mapping in
200a DPDK process is preallocated at startup, thereby placing an upper limit on how
201much memory a DPDK application can have. DPDK memory is stored in segment lists,
202each segment is strictly one physical page. It is possible to change the amount
203of virtual memory being preallocated at startup by editing the following config
204variables:
205
206* ``RTE_MAX_MEMSEG_LISTS`` controls how many segment lists can DPDK have
207* ``RTE_MAX_MEM_MB_PER_LIST`` controls how much megabytes of memory each
208  segment list can address
209* ``RTE_MAX_MEMSEG_PER_LIST`` controls how many segments each segment list
210  can have
211* ``RTE_MAX_MEMSEG_PER_TYPE`` controls how many segments each memory type
212  can have (where "type" is defined as "page size + NUMA node" combination)
213* ``RTE_MAX_MEM_MB_PER_TYPE`` controls how much megabytes of memory each
214  memory type can address
215* ``RTE_MAX_MEM_MB`` places a global maximum on the amount of memory
216  DPDK can reserve
217
218Normally, these options do not need to be changed.
219
220.. note::
221
222    Preallocated virtual memory is not to be confused with preallocated hugepage
223    memory! All DPDK processes preallocate virtual memory at startup. Hugepages
224    can later be mapped into that preallocated VA space (if dynamic memory mode
225    is enabled), and can optionally be mapped into it at startup.
226
227.. _hugepage_mapping:
228
229Hugepage Mapping
230^^^^^^^^^^^^^^^^
231
232Below is an overview of methods used for each OS to obtain hugepages,
233explaining why certain limitations and options exist in EAL.
234See the user guide for a specific OS for configuration details.
235
236FreeBSD uses ``contigmem`` kernel module
237to reserve a fixed number of hugepages at system start,
238which are mapped by EAL at initialization using a specific ``sysctl()``.
239
240Windows EAL allocates hugepages from the OS as needed using Win32 API,
241so available amount depends on the system load.
242It uses ``virt2phys`` kernel module to obtain physical addresses,
243unless running in IOVA-as-VA mode (e.g. forced with ``--iova-mode=va``).
244
245Linux allows to select any combination of the following:
246
247* use files in hugetlbfs (the default)
248  or anonymous mappings (``--in-memory``);
249* map each hugepage from its own file (the default)
250  or map multiple hugepages from one big file (``--single-file-segments``).
251
252Mapping hugepages from files in hugetlbfs is essential for multi-process,
253because secondary processes need to map the same hugepages.
254EAL creates files like ``rtemap_0``
255in directories specified with ``--huge-dir`` option
256(or in the mount point for a specific hugepage size).
257The ``rte`` prefix can be changed using ``--file-prefix``.
258This may be needed for running multiple primary processes
259that share a hugetlbfs mount point.
260Each backing file by default corresponds to one hugepage,
261it is opened and locked for the entire time the hugepage is used.
262This may exhaust the number of open files limit (``NOFILE``).
263See :ref:`segment-file-descriptors` section
264on how the number of open backing file descriptors can be reduced.
265
266In dynamic memory mode, EAL removes a backing hugepage file
267when all pages mapped from it are freed back to the system.
268However, backing files may persist after the application terminates
269in case of a crash or a leak of DPDK memory (e.g. ``rte_free()`` is missing).
270This reduces the number of hugepages available to other processes
271as reported by ``/sys/kernel/mm/hugepages/hugepages-*/free_hugepages``.
272EAL can remove the backing files after opening them for mapping
273if ``--huge-unlink`` is given to avoid polluting hugetlbfs.
274However, since it disables multi-process anyway,
275using anonymous mapping (``--in-memory``) is recommended instead.
276
277:ref:`EAL memory allocator <malloc>` relies on hugepages being zero-filled.
278Hugepages are cleared by the kernel when a file in hugetlbfs or its part
279is mapped for the first time system-wide
280to prevent data leaks from previous users of the same hugepage.
281EAL ensures this behavior by removing existing backing files at startup
282and by recreating them before opening for mapping (as a precaution).
283
284One exception is ``--huge-unlink=never`` mode.
285It is used to speed up EAL initialization, usually on application restart.
286Clearing memory constitutes more than 95% of hugepage mapping time.
287EAL can save it by remapping existing backing files
288with all the data left in the mapped hugepages ("dirty" memory).
289Such segments are marked with ``RTE_MEMSEG_FLAG_DIRTY``.
290Memory allocator detects dirty segments and handles them accordingly,
291in particular, it clears memory requested with ``rte_zmalloc*()``.
292In this mode EAL also does not remove a backing file
293when all pages mapped from it are freed,
294because they are intended to be reusable at restart.
295
296Anonymous mapping does not allow multi-process architecture.
297This mode does not use hugetlbfs
298and thus does not require root permissions for memory management
299(the limit of locked memory amount, ``MEMLOCK``, still applies).
300It is free of filename conflict and leftover file issues.
301If ``memfd_create(2)`` is supported both at build and run time,
302DPDK memory manager can provide file descriptors for memory segments,
303which are required for VirtIO with vhost-user backend.
304This can exhaust the number of open files limit (``NOFILE``)
305despite not creating any visible files.
306See :ref:`segment-file-descriptors` section
307on how the number of open file descriptors used by EAL can be reduced.
308
309.. _segment-file-descriptors:
310
311Segment File Descriptors
312^^^^^^^^^^^^^^^^^^^^^^^^
313
314On Linux, in most cases, EAL will store segment file descriptors in EAL. This
315can become a problem when using smaller page sizes due to underlying limitations
316of ``glibc`` library. For example, Linux API calls such as ``select()`` may not
317work correctly because ``glibc`` does not support more than certain number of
318file descriptors.
319
320There are two possible solutions for this problem. The recommended solution is
321to use ``--single-file-segments`` mode, as that mode will not use a file
322descriptor per each page, and it will keep compatibility with Virtio with
323vhost-user backend. This option is not available when using ``--legacy-mem``
324mode.
325
326Another option is to use bigger page sizes. Since fewer pages are required to
327cover the same memory area, fewer file descriptors will be stored internally
328by EAL.
329
330Hugepage Worker Stacks
331^^^^^^^^^^^^^^^^^^^^^^
332
333When the ``--huge-worker-stack[=size]`` EAL option is specified, worker
334thread stacks are allocated from hugepage memory local to the NUMA node
335of the thread. Worker stack size defaults to system pthread stack size
336if the optional size parameter is not specified.
337
338.. warning::
339    Stacks allocated from hugepage memory are not protected by guard
340    pages. Worker stacks must be sufficiently sized to prevent stack
341    overflow when this option is used.
342
343    As with normal thread stacks, hugepage worker thread stack size is
344    fixed and is not dynamically resized. Therefore, an application that
345    is free of stack page faults under a given load should be safe with
346    hugepage worker thread stacks given the same thread stack size and
347    loading conditions.
348
349Support for Externally Allocated Memory
350~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
351
352It is possible to use externally allocated memory in DPDK. There are two ways in
353which using externally allocated memory can work: the malloc heap API's, and
354manual memory management.
355
356+ Using heap API's for externally allocated memory
357
358Using a set of malloc heap API's is the recommended way to use externally
359allocated memory in DPDK. In this way, support for externally allocated memory
360is implemented through overloading the socket ID - externally allocated heaps
361will have socket ID's that would be considered invalid under normal
362circumstances. Requesting an allocation to take place from a specified
363externally allocated memory is a matter of supplying the correct socket ID to
364DPDK allocator, either directly (e.g. through a call to ``rte_malloc``) or
365indirectly (through data structure-specific allocation API's such as
366``rte_ring_create``). Using these API's also ensures that mapping of externally
367allocated memory for DMA is also performed on any memory segment that is added
368to a DPDK malloc heap.
369
370Since there is no way DPDK can verify whether memory is available or valid, this
371responsibility falls on the shoulders of the user. All multiprocess
372synchronization is also user's responsibility, as well as ensuring  that all
373calls to add/attach/detach/remove memory are done in the correct order. It is
374not required to attach to a memory area in all processes - only attach to memory
375areas as needed.
376
377The expected workflow is as follows:
378
379* Get a pointer to memory area
380* Create a named heap
381* Add memory area(s) to the heap
382    - If IOVA table is not specified, IOVA addresses will be assumed to be
383      unavailable, and DMA mappings will not be performed
384    - Other processes must attach to the memory area before they can use it
385* Get socket ID used for the heap
386* Use normal DPDK allocation procedures, using supplied socket ID
387* If memory area is no longer needed, it can be removed from the heap
388    - Other processes must detach from this memory area before it can be removed
389* If heap is no longer needed, remove it
390    - Socket ID will become invalid and will not be reused
391
392For more information, please refer to ``rte_malloc`` API documentation,
393specifically the ``rte_malloc_heap_*`` family of function calls.
394
395+ Using externally allocated memory without DPDK API's
396
397While using heap API's is the recommended method of using externally allocated
398memory in DPDK, there are certain use cases where the overhead of DPDK heap API
399is undesirable - for example, when manual memory management is performed on an
400externally allocated area. To support use cases where externally allocated
401memory will not be used as part of normal DPDK workflow, there is also another
402set of API's under the ``rte_extmem_*`` namespace.
403
404These API's are (as their name implies) intended to allow registering or
405unregistering externally allocated memory to/from DPDK's internal page table, to
406allow API's like ``rte_mem_virt2memseg`` etc. to work with externally allocated
407memory. Memory added this way will not be available for any regular DPDK
408allocators; DPDK will leave this memory for the user application to manage.
409
410The expected workflow is as follows:
411
412* Get a pointer to memory area
413* Register memory within DPDK
414    - If IOVA table is not specified, IOVA addresses will be assumed to be
415      unavailable
416    - Other processes must attach to the memory area before they can use it
417* Perform DMA mapping with ``rte_dev_dma_map`` if needed
418* Use the memory area in your application
419* If memory area is no longer needed, it can be unregistered
420    - If the area was mapped for DMA, unmapping must be performed before
421      unregistering memory
422    - Other processes must detach from the memory area before it can be
423      unregistered
424
425Since these externally allocated memory areas will not be managed by DPDK, it is
426therefore up to the user application to decide how to use them and what to do
427with them once they're registered.
428
429Per-lcore and Shared Variables
430~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
431
432By default, static variables, memory blocks allocated on the DPDK heap,
433and other types of memory are shared by all DPDK threads.
434
435An application, a DPDK library, or a PMD may opt to keep per-thread state.
436
437Per-thread data can be maintained using either :doc:`lcore variables <lcore_var>`,
438*thread-local storage (TLS)* (see ``rte_per_lcore.h``),
439or a static array of ``RTE_MAX_LCORE`` elements, indexed by ``rte_lcore_id()``.
440These methods allow per-lcore data to be largely internal to the module
441and not directly exposed in its API.
442Another approach is to explicitly handle per-thread aspects in the API
443(e.g., the ports in the eventdev API).
444
445Lcore variables are suitable for small objects that are statically allocated
446at the time of module or application initialization.
447An lcore variable takes on one value for each lcore ID-equipped thread
448(i.e., for both EAL threads and registered non-EAL threads,
449in total ``RTE_MAX_LCORE`` instances).
450The lifetime of lcore variables is independent of the owning threads
451and can, therefore, be initialized before the threads are created.
452
453Variables with thread-local storage are allocated when the thread is created
454and exist until the thread terminates.
455These are applicable for every thread in the process.
456Only very small objects should be allocated in TLS,
457as large TLS objects can significantly slow down thread creation
458and may unnecessarily increase the memory footprint of applications
459that extensively use unregistered threads.
460
461A common but now largely obsolete DPDK pattern is to use a static array
462sized according to the maximum number of lcore ID-equipped threads
463(i.e., with ``RTE_MAX_LCORE`` elements).
464To avoid *false sharing*, each element must be both cache-aligned
465and include an ``RTE_CACHE_GUARD``.
466This extensive use of padding causes internal fragmentation (i.e., unused space)
467and reduces cache hit rates.
468
469For more discussions on per-lcore state,
470refer to the :doc:`lcore variables documentation <lcore_var>`.
471
472Logs
473~~~~
474
475While originally part of EAL, DPDK logging functionality is now provided by the :doc:`log_lib`.
476
477Trace and Debug Functions
478^^^^^^^^^^^^^^^^^^^^^^^^^
479
480There are some debug functions to dump the stack in glibc.
481The rte_panic() function can voluntarily provoke a SIG_ABORT,
482which can trigger the generation of a core file, readable by gdb.
483
484CPU Feature Identification
485~~~~~~~~~~~~~~~~~~~~~~~~~~
486
487The EAL can query the CPU at runtime (using the rte_cpu_get_features() function) to determine which CPU features are available.
488
489User Space Interrupt Event
490~~~~~~~~~~~~~~~~~~~~~~~~~~
491
492+ User Space Interrupt and Alarm Handling in Host Thread
493
494The EAL creates a host thread to poll the UIO device file descriptors to detect the interrupts.
495Callbacks can be registered or unregistered by the EAL functions for a specific interrupt event
496and are called in the host thread asynchronously.
497The EAL also allows timed callbacks to be used in the same way as for NIC interrupts.
498
499.. note::
500
501    In DPDK PMD, the only interrupts handled by the dedicated host thread are those for link status change
502    (link up and link down notification) and for sudden device removal.
503
504
505+ RX Interrupt Event
506
507The receive and transmit routines provided by each PMD don't limit themselves to execute in polling thread mode.
508To ease the idle polling with tiny throughput, it's useful to pause the polling and wait until the wake-up event happens.
509The RX interrupt is the first choice to be such kind of wake-up event, but probably won't be the only one.
510
511EAL provides the event APIs for this event-driven thread mode.
512Taking Linux as an example, the implementation relies on epoll. Each thread can monitor an epoll instance
513in which all the wake-up events' file descriptors are added. The event file descriptors are created and mapped to
514the interrupt vectors according to the UIO/VFIO spec.
515From FreeBSD's perspective, kqueue is the alternative way, but not implemented yet.
516
517EAL initializes the mapping between event file descriptors and interrupt vectors, while each device initializes the mapping
518between interrupt vectors and queues. In this way, EAL actually is unaware of the interrupt cause on the specific vector.
519The eth_dev driver takes responsibility to program the latter mapping.
520
521.. note::
522
523    Per queue RX interrupt event is only allowed in VFIO which supports multiple MSI-X vector. In UIO, the RX interrupt
524    together with other interrupt causes shares the same vector. In this case, when RX interrupt and LSC(link status change)
525    interrupt are both enabled(intr_conf.lsc == 1 && intr_conf.rxq == 1), only the former is capable.
526
527The RX interrupt are controlled/enabled/disabled by ethdev APIs - 'rte_eth_dev_rx_intr_*'. They return failure if the PMD
528hasn't support them yet. The intr_conf.rxq flag is used to turn on the capability of RX interrupt per device.
529
530+ Device Removal Event
531
532This event is triggered by a device being removed at a bus level. Its
533underlying resources may have been made unavailable (i.e. PCI mappings
534unmapped). The PMD must make sure that on such occurrence, the application can
535still safely use its callbacks.
536
537This event can be subscribed to in the same way one would subscribe to a link
538status change event. The execution context is thus the same, i.e. it is the
539dedicated interrupt host thread.
540
541Considering this, it is likely that an application would want to close a
542device having emitted a Device Removal Event. In such case, calling
543``rte_eth_dev_close()`` can trigger it to unregister its own Device Removal Event
544callback. Care must be taken not to close the device from the interrupt handler
545context. It is necessary to reschedule such closing operation.
546
547Block list
548~~~~~~~~~~
549
550The EAL PCI device block list functionality can be used to mark certain NIC ports as unavailable,
551so they are ignored by the DPDK.
552The ports to be blocked are identified using the PCIe* description (Domain:Bus:Device.Function).
553
554Misc Functions
555~~~~~~~~~~~~~~
556
557Locks and atomic operations are per-architecture (i686 and x86_64).
558
559Lock annotations
560~~~~~~~~~~~~~~~~
561
562R/W locks, seq locks and spinlocks have been instrumented to help developers in
563catching issues in DPDK.
564
565This instrumentation relies on
566`clang Thread Safety checks <https://clang.llvm.org/docs/ThreadSafetyAnalysis.html>`_.
567All attributes are prefixed with __rte and are fully described in the clang
568documentation.
569
570Some general comments:
571
572- it is important that lock requirements are expressed at the function
573  declaration level in headers so that other code units can be inspected,
574- when some global lock is necessary to some user-exposed API, it is preferred
575  to expose it via an internal helper rather than expose the global variable,
576- there are a list of known limitations with clang instrumentation, but before
577  waiving checks with ``__rte_no_thread_safety_analysis`` in your code, please
578  discuss it on the mailing list,
579
580The checks are enabled by default for libraries and drivers.
581They can be disabled by setting ``annotate_locks`` to ``false`` in
582the concerned library/driver ``meson.build``.
583
584IOVA Mode Detection
585~~~~~~~~~~~~~~~~~~~
586
587IOVA Mode is selected by considering what the current usable Devices on the
588system require and/or support.
589
590On FreeBSD, RTE_IOVA_PA is always the default. On Linux, the IOVA mode is
591detected based on a 2-step heuristic detailed below.
592
593For the first step, EAL asks each bus its requirement in terms of IOVA mode
594and decides on a preferred IOVA mode.
595
596- if all buses report RTE_IOVA_PA, then the preferred IOVA mode is RTE_IOVA_PA,
597- if all buses report RTE_IOVA_VA, then the preferred IOVA mode is RTE_IOVA_VA,
598- if all buses report RTE_IOVA_DC, no bus expressed a preference, then the
599  preferred mode is RTE_IOVA_DC,
600- if the buses disagree (at least one wants RTE_IOVA_PA and at least one wants
601  RTE_IOVA_VA), then the preferred IOVA mode is RTE_IOVA_DC (see below with the
602  check on Physical Addresses availability),
603
604If the buses have expressed no preference on which IOVA mode to pick, then a
605default is selected using the following logic:
606
607- if physical addresses are not available, RTE_IOVA_VA mode is used
608- if /sys/kernel/iommu_groups is not empty, RTE_IOVA_VA mode is used
609- otherwise, RTE_IOVA_PA mode is used
610
611In the case when the buses had disagreed on their preferred IOVA mode, part of
612the buses won't work because of this decision.
613
614The second step checks if the preferred mode complies with the Physical
615Addresses availability since those are only available to root user in recent
616kernels. Namely, if the preferred mode is RTE_IOVA_PA but there is no access to
617Physical Addresses, then EAL init fails early, since later probing of the
618devices would fail anyway.
619
620.. note::
621
622    The RTE_IOVA_VA mode is preferred as the default in most cases for the
623    following reasons:
624
625    - All drivers are expected to work in RTE_IOVA_VA mode, irrespective of
626      physical address availability.
627    - By default, the mempool, first asks for IOVA-contiguous memory using
628      ``RTE_MEMZONE_IOVA_CONTIG``. This is slow in RTE_IOVA_PA mode and it may
629      affect the application boot time.
630    - It is easy to enable large amount of IOVA-contiguous memory use cases
631      with IOVA in VA mode.
632
633    It is expected that all PCI drivers work in both RTE_IOVA_PA and
634    RTE_IOVA_VA modes.
635
636    If a PCI driver does not support RTE_IOVA_PA mode, the
637    ``RTE_PCI_DRV_NEED_IOVA_AS_VA`` flag is used to dictate that this PCI
638    driver can only work in RTE_IOVA_VA mode.
639
640
641IOVA Mode Configuration
642~~~~~~~~~~~~~~~~~~~~~~~
643
644Auto detection of the IOVA mode, based on probing the bus and IOMMU configuration, may not report
645the desired addressing mode when virtual devices that are not directly attached to the bus are present.
646To facilitate forcing the IOVA mode to a specific value the EAL command line option ``--iova-mode`` can
647be used to select either physical addressing('pa') or virtual addressing('va').
648
649.. _max_simd_bitwidth:
650
651
652Max SIMD bitwidth
653~~~~~~~~~~~~~~~~~
654
655The EAL provides a single setting to limit the max SIMD bitwidth used by DPDK,
656which is used in determining the vector path, if any, chosen by a component.
657The value can be set at runtime by an application using the
658'rte_vect_set_max_simd_bitwidth(uint16_t bitwidth)' function,
659which should only be called once at initialization, before EAL init.
660The value can be overridden by the user using the EAL command-line option '--force-max-simd-bitwidth'.
661
662When choosing a vector path, along with checking the CPU feature support,
663the value of the max SIMD bitwidth must also be checked, and can be retrieved using the
664'rte_vect_get_max_simd_bitwidth()' function.
665The value should be compared against the enum values for accepted max SIMD bitwidths:
666
667.. code-block:: c
668
669   enum rte_vect_max_simd {
670       RTE_VECT_SIMD_DISABLED = 64,
671       RTE_VECT_SIMD_128 = 128,
672       RTE_VECT_SIMD_256 = 256,
673       RTE_VECT_SIMD_512 = 512,
674       RTE_VECT_SIMD_MAX = INT16_MAX + 1,
675   };
676
677    if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512)
678        /* Take AVX-512 vector path */
679    else if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
680        /* Take AVX2 vector path */
681
682
683Memory Segments and Memory Zones (memzone)
684------------------------------------------
685
686The mapping of physical memory is provided by this feature in the EAL.
687As physical memory can have gaps, the memory is described in a table of descriptors,
688and each descriptor (called rte_memseg ) describes a physical page.
689
690On top of this, the memzone allocator's role is to reserve contiguous portions of physical memory.
691These zones are identified by a unique name when the memory is reserved.
692
693The rte_memzone descriptors are also located in the configuration structure.
694This structure is accessed using rte_eal_get_configuration().
695The lookup (by name) of a memory zone returns a descriptor containing the physical address of the memory zone.
696
697Memory zones can be reserved with specific start address alignment by supplying the align parameter
698(by default, they are aligned to cache line size).
699The alignment value should be a power of two and not less than the cache line size (64 bytes).
700Memory zones can also be reserved from either 2 MB or 1 GB hugepages, provided that both are available on the system.
701
702Both memsegs and memzones are stored using ``rte_fbarray`` structures. Please
703refer to *DPDK API Reference* for more information.
704
705
706Multiple pthread
707----------------
708
709DPDK usually pins one pthread per core to avoid the overhead of task switching.
710This allows for significant performance gains, but lacks flexibility and is not always efficient.
711
712Power management helps to improve the CPU efficiency by limiting the CPU runtime frequency.
713However, alternately it is possible to utilize the idle cycles available to take advantage of
714the full capability of the CPU.
715
716By taking advantage of cgroup, the CPU utilization quota can be simply assigned.
717This gives another way to improve the CPU efficiency, however, there is a prerequisite;
718DPDK must handle the context switching between multiple pthreads per core.
719
720For further flexibility, it is useful to set pthread affinity not only to a CPU but to a CPU set.
721
722EAL pthread and lcore Affinity
723~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
724
725The term "lcore" refers to an EAL thread, which is really a Linux/FreeBSD pthread.
726"EAL pthreads"  are created and managed by EAL and execute the tasks issued by *remote_launch*.
727In each EAL pthread, there is a TLS (Thread Local Storage) called *_lcore_id* for unique identification.
728As EAL pthreads usually bind 1:1 to the physical CPU, the *_lcore_id* is typically equal to the CPU ID.
729
730When using multiple pthreads, however, the binding is no longer always 1:1 between an EAL pthread and a specified physical CPU.
731The EAL pthread may have affinity to a CPU set, and as such the *_lcore_id* will not be the same as the CPU ID.
732For this reason, there is an EAL long option '--lcores' defined to assign the CPU affinity of lcores.
733For a specified lcore ID or ID group, the option allows setting the CPU set for that EAL pthread.
734
735The format pattern:
736	--lcores='<lcore_set>[@cpu_set][,<lcore_set>[@cpu_set],...]'
737
738'lcore_set' and 'cpu_set' can be a single number, range or a group.
739
740A number is a "digit([0-9]+)"; a range is "<number>-<number>"; a group is "(<number|range>[,<number|range>,...])".
741
742If a '\@cpu_set' value is not supplied, the value of 'cpu_set' will default to the value of 'lcore_set'.
743
744    ::
745
746    	For example, "--lcores='1,2@(5-7),(3-5)@(0,2),(0,6),7-8'" which means start 9 EAL thread;
747    	    lcore 0 runs on cpuset 0x41 (cpu 0,6);
748    	    lcore 1 runs on cpuset 0x2 (cpu 1);
749    	    lcore 2 runs on cpuset 0xe0 (cpu 5,6,7);
750    	    lcore 3,4,5 runs on cpuset 0x5 (cpu 0,2);
751    	    lcore 6 runs on cpuset 0x41 (cpu 0,6);
752    	    lcore 7 runs on cpuset 0x80 (cpu 7);
753    	    lcore 8 runs on cpuset 0x100 (cpu 8).
754
755Using this option, for each given lcore ID, the associated CPUs can be assigned.
756It's also compatible with the pattern of corelist('-l') option.
757
758non-EAL pthread support
759~~~~~~~~~~~~~~~~~~~~~~~
760
761It is possible to use the DPDK execution context with any user pthread (aka. non-EAL pthreads).
762There are two kinds of non-EAL pthreads:
763
764- a registered non-EAL pthread with a valid *_lcore_id* that was successfully assigned by calling ``rte_thread_register()``,
765- a non registered non-EAL pthread with a LCORE_ID_ANY,
766
767For 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).
768
769All these impacts are mentioned in :ref:`known_issue_label` section.
770
771Public Thread API
772~~~~~~~~~~~~~~~~~
773
774There are two public APIs ``rte_thread_set_affinity()`` and ``rte_thread_get_affinity()`` introduced for threads.
775When they're used in any pthread context, the Thread Local Storage(TLS) will be set/get.
776
777Those TLS include *_cpuset* and *_socket_id*:
778
779*	*_cpuset* stores the CPUs bitmap to which the pthread is affinitized.
780
781*	*_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.
782
783
784Control Thread API
785~~~~~~~~~~~~~~~~~~
786
787It is possible to create Control Threads using the public API
788``rte_thread_create_control()``.
789Those threads can be used for management/infrastructure tasks and are used
790internally by DPDK for multi process support and interrupt handling.
791
792Those threads will be scheduled on CPUs part of the original process CPU
793affinity from which the dataplane and service lcores are excluded.
794
795For example, on a 8 CPUs system, starting a dpdk application with -l 2,3
796(dataplane cores), then depending on the affinity configuration which can be
797controlled with tools like taskset (Linux) or cpuset (FreeBSD),
798
799- with no affinity configuration, the Control Threads will end up on
800  0-1,4-7 CPUs.
801- with affinity restricted to 2-4, the Control Threads will end up on
802  CPU 4.
803- with affinity restricted to 2-3, the Control Threads will end up on
804  CPU 2 (main lcore, which is the default when no CPU is available).
805
806.. _known_issue_label:
807
808Known Issues
809~~~~~~~~~~~~
810
811+ rte_mempool
812
813  The rte_mempool uses a per-lcore cache inside the mempool.
814  For unregistered non-EAL pthreads, ``rte_lcore_id()`` will not return a valid number.
815  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.
816  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.
817
818+ rte_ring
819
820  rte_ring supports multi-producer enqueue and multi-consumer dequeue.
821  However, it is non-preemptive, this has a knock on effect of making rte_mempool non-preemptible.
822
823  .. note::
824
825    The "non-preemptive" constraint means:
826
827    - a pthread doing multi-producers enqueues on a given ring must not
828      be preempted by another pthread doing a multi-producer enqueue on
829      the same ring.
830    - a pthread doing multi-consumers dequeues on a given ring must not
831      be preempted by another pthread doing a multi-consumer dequeue on
832      the same ring.
833
834    Bypassing this constraint may cause the 2nd pthread to spin until the 1st one is scheduled again.
835    Moreover, if the 1st pthread is preempted by a context that has an higher priority, it may even cause a dead lock.
836
837  This means, use cases involving preemptible pthreads should consider using rte_ring carefully.
838
839  #. It CAN be used for preemptible single-producer and single-consumer use case.
840
841  #. It CAN be used for non-preemptible multi-producer and preemptible single-consumer use case.
842
843  #. It CAN be used for preemptible single-producer and non-preemptible multi-consumer use case.
844
845  #. 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.
846
847  #. It MUST not be used by multi-producer/consumer pthreads, whose scheduling policies are SCHED_FIFO or SCHED_RR.
848
849  Alternatively, applications can use the lock-free stack mempool handler. When
850  considering this handler, note that:
851
852  - It is currently limited to the aarch64 and x86_64 platforms, because it uses
853    an instruction (16-byte compare-and-swap) that is not yet available on other
854    platforms.
855  - It has worse average-case performance than the non-preemptive rte_ring, but
856    software caching (e.g. the mempool cache) can mitigate this by reducing the
857    number of stack accesses.
858
859+ rte_timer
860
861  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.
862
863+ rte_log
864
865  In unregistered non-EAL pthreads, there is no per thread loglevel and logtype, global loglevels are used.
866
867+ misc
868
869  The debug statistics of rte_ring, rte_mempool and rte_timer are not supported in an unregistered non-EAL pthread.
870
871Signal Safety
872~~~~~~~~~~~~~
873
874  The Posix API defines an async-signal-safe function as one that can be safely
875  called from with a signal handler. Many DPDK functions are non-reentrant and
876  therefore are unsafe to call from a signal handler.
877
878  The kinds of issues that make DPDK functions unsafe can be understood when
879  one considers that much of the code in DPDK uses locks and other shared
880  resources. For example, calling ``rte_mempool_lookup()`` from a signal
881  would deadlock if the signal happened during previous call ``rte_mempool``
882  routines.
883
884  Other functions are not signal safe because they use one or more
885  library routines that are not themselves signal safe.
886  For example, calling ``rte_panic()`` is not safe in a signal handler
887  because it uses ``rte_log()`` and ``rte_log()`` may call ``vfprintf()`` or
888  ``syslog()`` library functions which are not in the list of
889  signal safe functions
890  `Signal-Safety manual page <https://man7.org/linux/man-pages/man7/signal-safety.7.html>`_.
891
892  The set of functions that are expected to be async-signal-safe in DPDK
893  is shown in the following table. The functions not otherwise noted
894  are not async-signal-safe.
895
896.. csv-table:: **Signal Safe Functions**
897   :header: "Function"
898   :widths: 32
899
900   rte_dump_stack
901   rte_eal_get_lcore_state
902   rte_eal_get_runtime_dir
903   rte_eal_has_hugepages
904   rte_eal_has_pci
905   rte_eal_lcore_role
906   rte_eal_process_type
907   rte_eal_using_phys_addrs
908   rte_get_hpet_cycles
909   rte_get_hpet_hz
910   rte_get_main_lcore
911   rte_get_next_lcore
912   rte_get_tsc_hz
913   rte_hypervisor_get
914   rte_hypervisor_get_name
915   rte_lcore_count
916   rte_lcore_cpuset
917   rte_lcore_has_role
918   rte_lcore_index
919   rte_lcore_is_enabled
920   rte_lcore_to_cpu_id
921   rte_lcore_to_socket_id
922   rte_log_get_global_level
923   rte_log_get_level
924   rte_memory_get_nchannel
925   rte_memory_get_nrank
926   rte_reciprocal_value
927   rte_reciprocal_value_u64
928   rte_socket_count
929   rte_socket_id
930   rte_socket_id_by_idx
931   rte_strerror
932   rte_strscpy
933   rte_strsplit
934   rte_sys_gettid
935   rte_uuid_compare
936   rte_uuid_is_null
937   rte_uuid_parse
938   rte_uuid_unparse
939
940
941cgroup control
942~~~~~~~~~~~~~~
943
944The 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).
945We expect only 50% of CPU spend on packet IO.
946
947  .. code-block:: console
948
949    mkdir /sys/fs/cgroup/cpu/pkt_io
950    mkdir /sys/fs/cgroup/cpuset/pkt_io
951
952    echo $cpu > /sys/fs/cgroup/cpuset/cpuset.cpus
953
954    echo $t0 > /sys/fs/cgroup/cpu/pkt_io/tasks
955    echo $t0 > /sys/fs/cgroup/cpuset/pkt_io/tasks
956
957    echo $t1 > /sys/fs/cgroup/cpu/pkt_io/tasks
958    echo $t1 > /sys/fs/cgroup/cpuset/pkt_io/tasks
959
960    cd /sys/fs/cgroup/cpu/pkt_io
961    echo 100000 > pkt_io/cpu.cfs_period_us
962    echo  50000 > pkt_io/cpu.cfs_quota_us
963
964.. _malloc:
965
966Malloc
967------
968
969The EAL provides a malloc API to allocate any-sized memory.
970
971The objective of this API is to provide malloc-like functions to allow
972allocation from hugepage memory and to facilitate application porting.
973The *DPDK API Reference* manual describes the available functions.
974
975Typically, these kinds of allocations should not be done in data plane
976processing because they are slower than pool-based allocation and make
977use of locks within the allocation and free paths.
978However, they can be used in configuration code.
979
980Refer to the rte_malloc() function description in the *DPDK API Reference*
981manual for more information.
982
983
984Alignment and NUMA Constraints
985~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
986
987The rte_malloc() takes an align argument that can be used to request a memory
988area that is aligned on a multiple of this value (which must be a power of two).
989
990On systems with NUMA support, a call to the rte_malloc() function will return
991memory that has been allocated on the NUMA socket of the core which made the call.
992A set of APIs is also provided, to allow memory to be explicitly allocated on a
993NUMA socket directly, or by allocated on the NUMA socket where another core is
994located, in the case where the memory is to be used by a logical core other than
995on the one doing the memory allocation.
996
997Use Cases
998~~~~~~~~~
999
1000This API is meant to be used by an application that requires malloc-like
1001functions at initialization time.
1002
1003For allocating/freeing data at runtime, in the fast-path of an application,
1004the memory pool library should be used instead.
1005
1006Internal Implementation
1007~~~~~~~~~~~~~~~~~~~~~~~
1008
1009Data Structures
1010^^^^^^^^^^^^^^^
1011
1012There are two data structure types used internally in the malloc library:
1013
1014*   struct malloc_heap - used to track free space on a per-socket basis
1015
1016*   struct malloc_elem - the basic element of allocation and free-space
1017    tracking inside the library.
1018
1019Structure: malloc_heap
1020""""""""""""""""""""""
1021
1022The malloc_heap structure is used to manage free space on a per-socket basis.
1023Internally, there is one heap structure per NUMA node, which allows us to
1024allocate memory to a thread based on the NUMA node on which this thread runs.
1025While this does not guarantee that the memory will be used on that NUMA node,
1026it is no worse than a scheme where the memory is always allocated on a fixed
1027or random node.
1028
1029The key fields of the heap structure and their function are described below
1030(see also diagram above):
1031
1032*   lock - the lock field is needed to synchronize access to the heap.
1033    Given that the free space in the heap is tracked using a linked list,
1034    we need a lock to prevent two threads manipulating the list at the same time.
1035
1036*   free_head - this points to the first element in the list of free nodes for
1037    this malloc heap.
1038
1039*   first - this points to the first element in the heap.
1040
1041*   last - this points to the last element in the heap.
1042
1043.. _figure_malloc_heap:
1044
1045.. figure:: img/malloc_heap.*
1046
1047   Example of a malloc heap and malloc elements within the malloc library
1048
1049
1050.. _malloc_elem:
1051
1052Structure: malloc_elem
1053""""""""""""""""""""""
1054
1055The malloc_elem structure is used as a generic header structure for various
1056blocks of memory.
1057It is used in two different ways - all shown in the diagram above:
1058
1059#.  As a header on a block of free or allocated memory - normal case
1060
1061#.  As a padding header inside a block of memory
1062
1063The most important fields in the structure and how they are used are described below.
1064
1065Malloc heap is a doubly-linked list, where each element keeps track of its
1066previous and next elements. Due to the fact that hugepage memory can come and
1067go, neighboring malloc elements may not necessarily be adjacent in memory.
1068Also, since a malloc element may span multiple pages, its contents may not
1069necessarily be IOVA-contiguous either - each malloc element is only guaranteed
1070to be virtually contiguous.
1071
1072.. note::
1073
1074    If the usage of a particular field in one of the above three usages is not
1075    described, the field can be assumed to have an undefined value in that
1076    situation, for example, for padding headers only the "state" and "pad"
1077    fields have valid values.
1078
1079*   heap - this pointer is a reference back to the heap structure from which
1080    this block was allocated.
1081    It is used for normal memory blocks when they are being freed, to add the
1082    newly-freed block to the heap's free-list.
1083
1084*   prev - this pointer points to previous header element/block in memory. When
1085    freeing a block, this pointer is used to reference the previous block to
1086    check if that block is also free. If so, and the two blocks are immediately
1087    adjacent to each other, then the two free blocks are merged to form a single
1088    larger block.
1089
1090*   next - this pointer points to next header element/block in memory. When
1091    freeing a block, this pointer is used to reference the next block to check
1092    if that block is also free. If so, and the two blocks are immediately
1093    adjacent to each other, then the two free blocks are merged to form a single
1094    larger block.
1095
1096*   free_list - this is a structure pointing to previous and next elements in
1097    this heap's free list.
1098    It is only used in normal memory blocks; on ``malloc()`` to find a suitable
1099    free block to allocate and on ``free()`` to add the newly freed element to
1100    the free-list.
1101
1102*   state - This field can have one of three values: ``FREE``, ``BUSY`` or
1103    ``PAD``.
1104    The former two are to indicate the allocation state of a normal memory block
1105    and the latter is to indicate that the element structure is a dummy structure
1106    at the end of the start-of-block padding, i.e. where the start of the data
1107    within a block is not at the start of the block itself, due to alignment
1108    constraints.
1109    In that case, the pad header is used to locate the actual malloc element
1110    header for the block.
1111
1112*   dirty - this flag is only meaningful when ``state`` is ``FREE``.
1113    It indicates that the content of the element is not fully zero-filled.
1114    Memory from such blocks must be cleared when requested via ``rte_zmalloc*()``.
1115    Dirty elements only appear with ``--huge-unlink=never``.
1116
1117*   pad - this holds the length of the padding present at the start of the block.
1118    In the case of a normal block header, it is added to the address of the end
1119    of the header to give the address of the start of the data area, i.e. the
1120    value passed back to the application on a malloc.
1121    Within a dummy header inside the padding, this same value is stored, and is
1122    subtracted from the address of the dummy header to yield the address of the
1123    actual block header.
1124
1125*   size - the size of the data block, including the header itself.
1126
1127Memory Allocation
1128^^^^^^^^^^^^^^^^^
1129
1130On EAL initialization, all preallocated memory segments are setup as part of the
1131malloc heap. This setup involves placing an :ref:`element header<malloc_elem>`
1132with ``FREE`` at the start of each virtually contiguous segment of memory.
1133The ``FREE`` element is then added to the ``free_list`` for the malloc heap.
1134
1135This setup also happens whenever memory is allocated at runtime (if supported),
1136in which case newly allocated pages are also added to the heap, merging with any
1137adjacent free segments if there are any.
1138
1139When an application makes a call to a malloc-like function, the malloc function
1140will first index the ``lcore_config`` structure for the calling thread, and
1141determine the NUMA node of that thread.
1142The NUMA node is used to index the array of ``malloc_heap`` structures which is
1143passed as a parameter to the ``heap_alloc()`` function, along with the
1144requested size, type, alignment and boundary parameters.
1145
1146The ``heap_alloc()`` function will scan the free_list of the heap, and attempt
1147to find a free block suitable for storing data of the requested size, with the
1148requested alignment and boundary constraints.
1149
1150When a suitable free element has been identified, the pointer to be returned
1151to the user is calculated.
1152The cache-line of memory immediately preceding this pointer is filled with a
1153struct malloc_elem header.
1154Because of alignment and boundary constraints, there could be free space at
1155the start and/or end of the element, resulting in the following behavior:
1156
1157#. Check for trailing space.
1158   If the trailing space is big enough, i.e. > 128 bytes, then the free element
1159   is split.
1160   If it is not, then we just ignore it (wasted space).
1161
1162#. Check for space at the start of the element.
1163   If the space at the start is small, i.e. <=128 bytes, then a pad header is
1164   used, and the remaining space is wasted.
1165   If, however, the remaining space is greater, then the free element is split.
1166
1167The advantage of allocating the memory from the end of the existing element is
1168that no adjustment of the free list needs to take place - the existing element
1169on the free list just has its size value adjusted, and the next/previous elements
1170have their "prev"/"next" pointers redirected to the newly created element.
1171
1172In case when there is not enough memory in the heap to satisfy allocation
1173request, EAL will attempt to allocate more memory from the system (if supported)
1174and, following successful allocation, will retry reserving the memory again. In
1175a multiprocessing scenario, all primary and secondary processes will synchronize
1176their memory maps to ensure that any valid pointer to DPDK memory is guaranteed
1177to be valid at all times in all currently running processes.
1178
1179Failure to synchronize memory maps in one of the processes will cause allocation
1180to fail, even though some of the processes may have allocated the memory
1181successfully. The memory is not added to the malloc heap unless primary process
1182has ensured that all other processes have mapped this memory successfully.
1183
1184Any successful allocation event will trigger a callback, for which user
1185applications and other DPDK subsystems can register. Additionally, validation
1186callbacks will be triggered before allocation if the newly allocated memory will
1187exceed threshold set by the user, giving a chance to allow or deny allocation.
1188
1189.. note::
1190
1191    Any allocation of new pages has to go through primary process. If the
1192    primary process is not active, no memory will be allocated even if it was
1193    theoretically possible to do so. This is because primary's process map acts
1194    as an authority on what should or should not be mapped, while each secondary
1195    process has its own, local memory map. Secondary processes do not update the
1196    shared memory map, they only copy its contents to their local memory map.
1197
1198Freeing Memory
1199^^^^^^^^^^^^^^
1200
1201To free an area of memory, the pointer to the start of the data area is passed
1202to the free function.
1203The size of the ``malloc_elem`` structure is subtracted from this pointer to get
1204the element header for the block.
1205If this header is of type ``PAD`` then the pad length is further subtracted from
1206the pointer to get the proper element header for the entire block.
1207
1208From this element header, we get pointers to the heap from which the block was
1209allocated and to where it must be freed, as well as the pointer to the previous
1210and next elements. These next and previous elements are then checked to see if
1211they are also ``FREE`` and are immediately adjacent to the current one, and if
1212so, they are merged with the current element. This means that we can never have
1213two ``FREE`` memory blocks adjacent to one another, as they are always merged
1214into a single block.
1215
1216If deallocating pages at runtime is supported, and the free element encloses
1217one or more pages, those pages can be deallocated and be removed from the heap.
1218If DPDK was started with command-line parameters for preallocating memory
1219(``-m`` or ``--socket-mem``), then those pages that were allocated at startup
1220will not be deallocated.
1221
1222Any successful deallocation event will trigger a callback, for which user
1223applications and other DPDK subsystems can register.
1224