1.. SPDX-License-Identifier: BSD-3-Clause 2 Copyright(c) 2010-2014 Intel Corporation. 3 4.. _Environment_Abstraction_Layer: 5 6Environment Abstraction Layer 7============================= 8 9The Environment Abstraction Layer (EAL) is responsible for gaining access to low-level resources such as hardware and memory space. 10It provides a generic interface that hides the environment specifics from the applications and libraries. 11It 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). 13 14Typical services expected from the EAL are: 15 16* DPDK Loading and Launching: 17 The DPDK and its application are linked as a single application and must be loaded by some means. 18 19* Core Affinity/Assignment Procedures: 20 The EAL provides mechanisms for assigning execution units to specific cores as well as creating execution instances. 21 22* System Memory Reservation: 23 The EAL facilitates the reservation of different memory zones, for example, physical memory areas for device interactions. 24 25* Trace and Debug Functions: Logs, dump_stack, panic and so on. 26 27* Utility Functions: Spinlocks and atomic counters that are not provided in libc. 28 29* CPU Feature Identification: Determine at runtime if a particular feature, for example, Intel® AVX is supported. 30 Determine if the current CPU supports the feature set that the binary was compiled for. 31 32* Interrupt Handling: Interfaces to register/unregister callbacks to specific interrupt sources. 33 34* Alarm Functions: Interfaces to set/remove callbacks to be run at a specific time. 35 36EAL in a Linux-userland Execution Environment 37--------------------------------------------- 38 39In a Linux user space environment, the DPDK application runs as a user-space application using the pthread library. 40 41The EAL performs physical memory allocation using mmap() in hugetlbfs (using huge page sizes to increase performance). 42This memory is exposed to DPDK service layers such as the :ref:`Mempool Library <Mempool_Library>`. 43 44At this point, the DPDK services layer will be initialized, then through pthread setaffinity calls, 45each execution unit will be assigned to a specific logical core to run as a user-level thread. 46 47The time reference is provided by the CPU Time-Stamp Counter (TSC) or by the HPET kernel API through a mmap() call. 48 49Initialization and Core Launching 50~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 51 52Part of the initialization is done by the start function of glibc. 53A check is also performed at initialization time to ensure that the micro architecture type chosen in the config file is supported by the CPU. 54Then, the main() function is called. The core initialization and launch is done in rte_eal_init() (see the API documentation). 55It consist of calls to the pthread library (more specifically, pthread_self(), pthread_create(), and pthread_setaffinity_np()). 56 57.. _figure_linux_launch: 58 59.. figure:: img/linuxapp_launch.* 60 61 EAL Initialization in a Linux Application Environment 62 63 64.. note:: 65 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. 70 71Shutdown and Cleanup 72~~~~~~~~~~~~~~~~~~~~ 73 74During the initialization of EAL resources such as hugepage backed memory can be 75allocated by core components. The memory allocated during ``rte_eal_init()`` 76can be released by calling the ``rte_eal_cleanup()`` function. Refer to the 77API documentation for details. 78 79Multi-process Support 80~~~~~~~~~~~~~~~~~~~~~ 81 82The Linux EAL allows a multi-process as well as a multi-threaded (pthread) deployment model. 83See chapter 84:ref:`Multi-process Support <Multi-process_Support>` for more details. 85 86Memory Mapping Discovery and Memory Reservation 87~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 88 89The allocation of large contiguous physical memory is done using hugepages. 90The EAL provides an API to reserve named memory zones in this contiguous memory. 91The physical address of the reserved memory for that memory zone is also returned to the user by the memory zone reservation API. 92 93There are two modes in which DPDK memory subsystem can operate: dynamic mode, 94and legacy mode. Both modes are explained below. 95 96.. note:: 97 98 Memory reservations done using the APIs provided by rte_malloc 99 are also backed by hugepages unless ``--no-huge`` option is given. 100 101Dynamic Memory Mode 102^^^^^^^^^^^^^^^^^^^ 103 104Currently, this mode is only supported on Linux and Windows. 105 106In this mode, usage of hugepages by DPDK application will grow and shrink based 107on application's requests. Any memory allocation through ``rte_malloc()``, 108``rte_memzone_reserve()`` or other methods, can potentially result in more 109hugepages being reserved from the system. Similarly, any memory deallocation can 110potentially result in hugepages being released back to the system. 111 112Memory allocated in this mode is not guaranteed to be IOVA-contiguous. If large 113chunks of IOVA-contiguous are required (with "large" defined as "more than one 114page"), it is recommended to either use VFIO driver for all physical devices (so 115that IOVA and VA addresses can be the same, thereby bypassing physical addresses 116entirely), or use legacy memory mode. 117 118For chunks of memory which must be IOVA-contiguous, it is recommended to use 119``rte_memzone_reserve()`` function with ``RTE_MEMZONE_IOVA_CONTIG`` flag 120specified. This way, memory allocator will ensure that, whatever memory mode is 121in use, either reserved memory will satisfy the requirements, or the allocation 122will fail. 123 124There 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, 126in which case preallocate memory will be "pinned" (i.e. will never be released 127by the application back to the system). It will be possible to allocate more 128hugepages, and deallocate those, but any preallocated pages will not be freed. 129If neither ``-m`` nor ``--socket-mem`` were specified, no memory will be 130preallocated, and all memory will be allocated at runtime, as needed. 131 132Another available option to use in dynamic memory mode is 133``--single-file-segments`` command-line option. This option will put pages in 134single files (per memseg list), as opposed to creating a file per page. This is 135normally not needed, but can be useful for use cases like userspace vhost, where 136there is limited number of page file descriptors that can be passed to VirtIO. 137 138If the application (or DPDK-internal code, such as device drivers) wishes to 139receive notifications about newly allocated memory, it is possible to register 140for memory event callbacks via ``rte_mem_event_callback_register()`` function. 141This will call a callback function any time DPDK's memory map has changed. 142 143If the application (or DPDK-internal code, such as device drivers) wishes to be 144notified about memory allocations above specified threshold (and have a chance 145to deny them), allocation validator callbacks are also available via 146``rte_mem_alloc_validator_callback_register()`` function. 147 148A 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 150of memory that can be used by DPDK application. 151 152.. warning:: 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 158 DPDK IPC. 159 160Legacy Memory Mode 161^^^^^^^^^^^^^^^^^^ 162 163This mode is enabled by specifying ``--legacy-mem`` command-line switch to the 164EAL. This switch will have no effect on FreeBSD as FreeBSD only supports 165legacy mode anyway. 166 167This mode mimics historical behavior of EAL. That is, EAL will reserve all 168memory at startup, sort all memory into large IOVA-contiguous chunks, and will 169not allow acquiring or releasing hugepages from the system at runtime. 170 171If neither ``-m`` nor ``--socket-mem`` were specified, the entire available 172hugepage memory will be preallocated. 173 174Hugepage Allocation Matching 175^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 176 177This behavior is enabled by specifying the ``--match-allocations`` command-line 178switch to the EAL. This switch is Linux-only and not supported with 179``--legacy-mem`` nor ``--no-huge``. 180 181Some applications using memory event callbacks may require that hugepages be 182freed exactly as they were allocated. These applications may also require 183that any allocation from the malloc heap not span across allocations 184associated with two different memory event callbacks. Hugepage allocation 185matching can be used by these types of applications to satisfy both of these 186requirements. This can result in some increased memory usage which is 187very dependent on the memory allocation patterns of the application. 188 18932-bit Support 190^^^^^^^^^^^^^^ 191 192Additional restrictions are present when running in 32-bit mode. In dynamic 193memory mode, by default maximum of 2 gigabytes of VA space will be preallocated, 194and all of it will be on main lcore NUMA node unless ``--socket-mem`` flag is 195used. 196 197In legacy mode, VA space will only be preallocated for segments that were 198requested (plus padding, to keep IOVA-contiguousness). 199 200Maximum Amount of Memory 201^^^^^^^^^^^^^^^^^^^^^^^^ 202 203All possible virtual memory space that can ever be used for hugepage mapping in 204a DPDK process is preallocated at startup, thereby placing an upper limit on how 205much memory a DPDK application can have. DPDK memory is stored in segment lists, 206each segment is strictly one physical page. It is possible to change the amount 207of virtual memory being preallocated at startup by editing the following config 208variables: 209 210* ``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 214 can have 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 220 DPDK can reserve 221 222Normally, these options do not need to be changed. 223 224.. note:: 225 226 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. 230 231.. _hugepage_mapping: 232 233Hugepage Mapping 234^^^^^^^^^^^^^^^^ 235 236Below is an overview of methods used for each OS to obtain hugepages, 237explaining why certain limitations and options exist in EAL. 238See the user guide for a specific OS for configuration details. 239 240FreeBSD uses ``contigmem`` kernel module 241to reserve a fixed number of hugepages at system start, 242which are mapped by EAL at initialization using a specific ``sysctl()``. 243 244Windows EAL allocates hugepages from the OS as needed using Win32 API, 245so available amount depends on the system load. 246It uses ``virt2phys`` kernel module to obtain physical addresses, 247unless running in IOVA-as-VA mode (e.g. forced with ``--iova-mode=va``). 248 249Linux allows to select any combination of the following: 250 251* use files in hugetlbfs (the default) 252 or anonymous mappings (``--in-memory``); 253* map each hugepage from its own file (the default) 254 or map multiple hugepages from one big file (``--single-file-segments``). 255 256Mapping hugepages from files in hugetlbfs is essential for multi-process, 257because secondary processes need to map the same hugepages. 258EAL creates files like ``rtemap_0`` 259in directories specified with ``--huge-dir`` option 260(or in the mount point for a specific hugepage size). 261The ``rte`` prefix can be changed using ``--file-prefix``. 262This may be needed for running multiple primary processes 263that share a hugetlbfs mount point. 264Each backing file by default corresponds to one hugepage, 265it is opened and locked for the entire time the hugepage is used. 266This may exhaust the number of open files limit (``NOFILE``). 267See :ref:`segment-file-descriptors` section 268on how the number of open backing file descriptors can be reduced. 269 270In dynamic memory mode, EAL removes a backing hugepage file 271when all pages mapped from it are freed back to the system. 272However, backing files may persist after the application terminates 273in case of a crash or a leak of DPDK memory (e.g. ``rte_free()`` is missing). 274This reduces the number of hugepages available to other processes 275as reported by ``/sys/kernel/mm/hugepages/hugepages-*/free_hugepages``. 276EAL can remove the backing files after opening them for mapping 277if ``--huge-unlink`` is given to avoid polluting hugetlbfs. 278However, since it disables multi-process anyway, 279using anonymous mapping (``--in-memory``) is recommended instead. 280 281:ref:`EAL memory allocator <malloc>` relies on hugepages being zero-filled. 282Hugepages are cleared by the kernel when a file in hugetlbfs or its part 283is mapped for the first time system-wide 284to prevent data leaks from previous users of the same hugepage. 285EAL ensures this behavior by removing existing backing files at startup 286and by recreating them before opening for mapping (as a precaution). 287 288One exception is ``--huge-unlink=never`` mode. 289It is used to speed up EAL initialization, usually on application restart. 290Clearing memory constitutes more than 95% of hugepage mapping time. 291EAL can save it by remapping existing backing files 292with all the data left in the mapped hugepages ("dirty" memory). 293Such segments are marked with ``RTE_MEMSEG_FLAG_DIRTY``. 294Memory allocator detects dirty segments and handles them accordingly, 295in particular, it clears memory requested with ``rte_zmalloc*()``. 296In this mode EAL also does not remove a backing file 297when all pages mapped from it are freed, 298because they are intended to be reusable at restart. 299 300Anonymous mapping does not allow multi-process architecture. 301This mode does not use hugetlbfs 302and thus does not require root permissions for memory management 303(the limit of locked memory amount, ``MEMLOCK``, still applies). 304It is free of filename conflict and leftover file issues. 305If ``memfd_create(2)`` is supported both at build and run time, 306DPDK memory manager can provide file descriptors for memory segments, 307which are required for VirtIO with vhost-user backend. 308This can exhaust the number of open files limit (``NOFILE``) 309despite not creating any visible files. 310See :ref:`segment-file-descriptors` section 311on how the number of open file descriptors used by EAL can be reduced. 312 313.. _segment-file-descriptors: 314 315Segment File Descriptors 316^^^^^^^^^^^^^^^^^^^^^^^^ 317 318On Linux, in most cases, EAL will store segment file descriptors in EAL. This 319can become a problem when using smaller page sizes due to underlying limitations 320of ``glibc`` library. For example, Linux API calls such as ``select()`` may not 321work correctly because ``glibc`` does not support more than certain number of 322file descriptors. 323 324There are two possible solutions for this problem. The recommended solution is 325to use ``--single-file-segments`` mode, as that mode will not use a file 326descriptor per each page, and it will keep compatibility with Virtio with 327vhost-user backend. This option is not available when using ``--legacy-mem`` 328mode. 329 330Another option is to use bigger page sizes. Since fewer pages are required to 331cover the same memory area, fewer file descriptors will be stored internally 332by EAL. 333 334Hugepage Worker Stacks 335^^^^^^^^^^^^^^^^^^^^^^ 336 337When the ``--huge-worker-stack[=size]`` EAL option is specified, worker 338thread stacks are allocated from hugepage memory local to the NUMA node 339of the thread. Worker stack size defaults to system pthread stack size 340if the optional size parameter is not specified. 341 342.. warning:: 343 Stacks allocated from hugepage memory are not protected by guard 344 pages. Worker stacks must be sufficiently sized to prevent stack 345 overflow when this option is used. 346 347 As with normal thread stacks, hugepage worker thread stack size is 348 fixed and is not dynamically resized. Therefore, an application that 349 is free of stack page faults under a given load should be safe with 350 hugepage worker thread stacks given the same thread stack size and 351 loading conditions. 352 353Support for Externally Allocated Memory 354~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 355 356It is possible to use externally allocated memory in DPDK. There are two ways in 357which using externally allocated memory can work: the malloc heap API's, and 358manual memory management. 359 360+ Using heap API's for externally allocated memory 361 362Using a set of malloc heap API's is the recommended way to use externally 363allocated memory in DPDK. In this way, support for externally allocated memory 364is implemented through overloading the socket ID - externally allocated heaps 365will have socket ID's that would be considered invalid under normal 366circumstances. Requesting an allocation to take place from a specified 367externally allocated memory is a matter of supplying the correct socket ID to 368DPDK allocator, either directly (e.g. through a call to ``rte_malloc``) or 369indirectly (through data structure-specific allocation API's such as 370``rte_ring_create``). Using these API's also ensures that mapping of externally 371allocated memory for DMA is also performed on any memory segment that is added 372to a DPDK malloc heap. 373 374Since there is no way DPDK can verify whether memory is available or valid, this 375responsibility falls on the shoulders of the user. All multiprocess 376synchronization is also user's responsibility, as well as ensuring that all 377calls to add/attach/detach/remove memory are done in the correct order. It is 378not required to attach to a memory area in all processes - only attach to memory 379areas as needed. 380 381The expected workflow is as follows: 382 383* Get a pointer to memory area 384* Create a named heap 385* Add memory area(s) to the heap 386 - If IOVA table is not specified, IOVA addresses will be assumed to be 387 unavailable, and DMA mappings will not be performed 388 - Other processes must attach to the memory area before they can use it 389* Get socket ID used for the heap 390* Use normal DPDK allocation procedures, using supplied socket ID 391* If memory area is no longer needed, it can be removed from the heap 392 - Other processes must detach from this memory area before it can be removed 393* If heap is no longer needed, remove it 394 - Socket ID will become invalid and will not be reused 395 396For more information, please refer to ``rte_malloc`` API documentation, 397specifically the ``rte_malloc_heap_*`` family of function calls. 398 399+ Using externally allocated memory without DPDK API's 400 401While using heap API's is the recommended method of using externally allocated 402memory in DPDK, there are certain use cases where the overhead of DPDK heap API 403is undesirable - for example, when manual memory management is performed on an 404externally allocated area. To support use cases where externally allocated 405memory will not be used as part of normal DPDK workflow, there is also another 406set of API's under the ``rte_extmem_*`` namespace. 407 408These API's are (as their name implies) intended to allow registering or 409unregistering externally allocated memory to/from DPDK's internal page table, to 410allow API's like ``rte_mem_virt2memseg`` etc. to work with externally allocated 411memory. Memory added this way will not be available for any regular DPDK 412allocators; DPDK will leave this memory for the user application to manage. 413 414The expected workflow is as follows: 415 416* Get a pointer to memory area 417* Register memory within DPDK 418 - If IOVA table is not specified, IOVA addresses will be assumed to be 419 unavailable 420 - Other processes must attach to the memory area before they can use it 421* Perform DMA mapping with ``rte_dev_dma_map`` if needed 422* Use the memory area in your application 423* If memory area is no longer needed, it can be unregistered 424 - If the area was mapped for DMA, unmapping must be performed before 425 unregistering memory 426 - Other processes must detach from the memory area before it can be 427 unregistered 428 429Since these externally allocated memory areas will not be managed by DPDK, it is 430therefore up to the user application to decide how to use them and what to do 431with them once they're registered. 432 433Per-lcore and Shared Variables 434~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 435 436.. note:: 437 438 lcore refers to a logical execution unit of the processor, sometimes called a hardware *thread*. 439 440Shared variables are the default behavior. 441Per-lcore variables are implemented using *Thread Local Storage* (TLS) to provide per-thread local storage. 442 443Logs 444~~~~ 445 446A logging API is provided by EAL. 447By default, in a Linux application, logs are sent to syslog and also to the console. 448However, the log function can be overridden by the user to use a different logging mechanism. 449 450Trace and Debug Functions 451^^^^^^^^^^^^^^^^^^^^^^^^^ 452 453There are some debug functions to dump the stack in glibc. 454The rte_panic() function can voluntarily provoke a SIG_ABORT, 455which can trigger the generation of a core file, readable by gdb. 456 457CPU Feature Identification 458~~~~~~~~~~~~~~~~~~~~~~~~~~ 459 460The EAL can query the CPU at runtime (using the rte_cpu_get_features() function) to determine which CPU features are available. 461 462User Space Interrupt Event 463~~~~~~~~~~~~~~~~~~~~~~~~~~ 464 465+ User Space Interrupt and Alarm Handling in Host Thread 466 467The EAL creates a host thread to poll the UIO device file descriptors to detect the interrupts. 468Callbacks can be registered or unregistered by the EAL functions for a specific interrupt event 469and are called in the host thread asynchronously. 470The EAL also allows timed callbacks to be used in the same way as for NIC interrupts. 471 472.. note:: 473 474 In DPDK PMD, the only interrupts handled by the dedicated host thread are those for link status change 475 (link up and link down notification) and for sudden device removal. 476 477 478+ RX Interrupt Event 479 480The receive and transmit routines provided by each PMD don't limit themselves to execute in polling thread mode. 481To ease the idle polling with tiny throughput, it's useful to pause the polling and wait until the wake-up event happens. 482The RX interrupt is the first choice to be such kind of wake-up event, but probably won't be the only one. 483 484EAL provides the event APIs for this event-driven thread mode. 485Taking Linux as an example, the implementation relies on epoll. Each thread can monitor an epoll instance 486in which all the wake-up events' file descriptors are added. The event file descriptors are created and mapped to 487the interrupt vectors according to the UIO/VFIO spec. 488From FreeBSD's perspective, kqueue is the alternative way, but not implemented yet. 489 490EAL initializes the mapping between event file descriptors and interrupt vectors, while each device initializes the mapping 491between interrupt vectors and queues. In this way, EAL actually is unaware of the interrupt cause on the specific vector. 492The eth_dev driver takes responsibility to program the latter mapping. 493 494.. note:: 495 496 Per queue RX interrupt event is only allowed in VFIO which supports multiple MSI-X vector. In UIO, the RX interrupt 497 together with other interrupt causes shares the same vector. In this case, when RX interrupt and LSC(link status change) 498 interrupt are both enabled(intr_conf.lsc == 1 && intr_conf.rxq == 1), only the former is capable. 499 500The RX interrupt are controlled/enabled/disabled by ethdev APIs - 'rte_eth_dev_rx_intr_*'. They return failure if the PMD 501hasn't support them yet. The intr_conf.rxq flag is used to turn on the capability of RX interrupt per device. 502 503+ Device Removal Event 504 505This event is triggered by a device being removed at a bus level. Its 506underlying resources may have been made unavailable (i.e. PCI mappings 507unmapped). The PMD must make sure that on such occurrence, the application can 508still safely use its callbacks. 509 510This event can be subscribed to in the same way one would subscribe to a link 511status change event. The execution context is thus the same, i.e. it is the 512dedicated interrupt host thread. 513 514Considering this, it is likely that an application would want to close a 515device having emitted a Device Removal Event. In such case, calling 516``rte_eth_dev_close()`` can trigger it to unregister its own Device Removal Event 517callback. Care must be taken not to close the device from the interrupt handler 518context. It is necessary to reschedule such closing operation. 519 520Block list 521~~~~~~~~~~ 522 523The EAL PCI device block list functionality can be used to mark certain NIC ports as unavailable, 524so they are ignored by the DPDK. 525The ports to be blocked are identified using the PCIe* description (Domain:Bus:Device.Function). 526 527Misc Functions 528~~~~~~~~~~~~~~ 529 530Locks and atomic operations are per-architecture (i686 and x86_64). 531 532IOVA Mode Detection 533~~~~~~~~~~~~~~~~~~~ 534 535IOVA Mode is selected by considering what the current usable Devices on the 536system require and/or support. 537 538On FreeBSD, RTE_IOVA_PA is always the default. On Linux, the IOVA mode is 539detected based on a 2-step heuristic detailed below. 540 541For the first step, EAL asks each bus its requirement in terms of IOVA mode 542and decides on a preferred IOVA mode. 543 544- if all buses report RTE_IOVA_PA, then the preferred IOVA mode is RTE_IOVA_PA, 545- if all buses report RTE_IOVA_VA, then the preferred IOVA mode is RTE_IOVA_VA, 546- if all buses report RTE_IOVA_DC, no bus expressed a preference, then the 547 preferred mode is RTE_IOVA_DC, 548- if the buses disagree (at least one wants RTE_IOVA_PA and at least one wants 549 RTE_IOVA_VA), then the preferred IOVA mode is RTE_IOVA_DC (see below with the 550 check on Physical Addresses availability), 551 552If the buses have expressed no preference on which IOVA mode to pick, then a 553default is selected using the following logic: 554 555- if physical addresses are not available, RTE_IOVA_VA mode is used 556- if /sys/kernel/iommu_groups is not empty, RTE_IOVA_VA mode is used 557- otherwise, RTE_IOVA_PA mode is used 558 559In the case when the buses had disagreed on their preferred IOVA mode, part of 560the buses won't work because of this decision. 561 562The second step checks if the preferred mode complies with the Physical 563Addresses availability since those are only available to root user in recent 564kernels. Namely, if the preferred mode is RTE_IOVA_PA but there is no access to 565Physical Addresses, then EAL init fails early, since later probing of the 566devices would fail anyway. 567 568.. note:: 569 570 The RTE_IOVA_VA mode is preferred as the default in most cases for the 571 following reasons: 572 573 - All drivers are expected to work in RTE_IOVA_VA mode, irrespective of 574 physical address availability. 575 - By default, the mempool, first asks for IOVA-contiguous memory using 576 ``RTE_MEMZONE_IOVA_CONTIG``. This is slow in RTE_IOVA_PA mode and it may 577 affect the application boot time. 578 - It is easy to enable large amount of IOVA-contiguous memory use cases 579 with IOVA in VA mode. 580 581 It is expected that all PCI drivers work in both RTE_IOVA_PA and 582 RTE_IOVA_VA modes. 583 584 If a PCI driver does not support RTE_IOVA_PA mode, the 585 ``RTE_PCI_DRV_NEED_IOVA_AS_VA`` flag is used to dictate that this PCI 586 driver can only work in RTE_IOVA_VA mode. 587 588 When the KNI kernel module is detected, RTE_IOVA_PA mode is preferred as a 589 performance penalty is expected in RTE_IOVA_VA mode. 590 591IOVA Mode Configuration 592~~~~~~~~~~~~~~~~~~~~~~~ 593 594Auto detection of the IOVA mode, based on probing the bus and IOMMU configuration, may not report 595the desired addressing mode when virtual devices that are not directly attached to the bus are present. 596To facilitate forcing the IOVA mode to a specific value the EAL command line option ``--iova-mode`` can 597be used to select either physical addressing('pa') or virtual addressing('va'). 598 599.. _max_simd_bitwidth: 600 601 602Max SIMD bitwidth 603~~~~~~~~~~~~~~~~~ 604 605The EAL provides a single setting to limit the max SIMD bitwidth used by DPDK, 606which is used in determining the vector path, if any, chosen by a component. 607The value can be set at runtime by an application using the 608'rte_vect_set_max_simd_bitwidth(uint16_t bitwidth)' function, 609which should only be called once at initialization, before EAL init. 610The value can be overridden by the user using the EAL command-line option '--force-max-simd-bitwidth'. 611 612When choosing a vector path, along with checking the CPU feature support, 613the value of the max SIMD bitwidth must also be checked, and can be retrieved using the 614'rte_vect_get_max_simd_bitwidth()' function. 615The value should be compared against the enum values for accepted max SIMD bitwidths: 616 617.. code-block:: c 618 619 enum rte_vect_max_simd { 620 RTE_VECT_SIMD_DISABLED = 64, 621 RTE_VECT_SIMD_128 = 128, 622 RTE_VECT_SIMD_256 = 256, 623 RTE_VECT_SIMD_512 = 512, 624 RTE_VECT_SIMD_MAX = INT16_MAX + 1, 625 }; 626 627 if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512) 628 /* Take AVX-512 vector path */ 629 else if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256) 630 /* Take AVX2 vector path */ 631 632 633Memory Segments and Memory Zones (memzone) 634------------------------------------------ 635 636The mapping of physical memory is provided by this feature in the EAL. 637As physical memory can have gaps, the memory is described in a table of descriptors, 638and each descriptor (called rte_memseg ) describes a physical page. 639 640On top of this, the memzone allocator's role is to reserve contiguous portions of physical memory. 641These zones are identified by a unique name when the memory is reserved. 642 643The rte_memzone descriptors are also located in the configuration structure. 644This structure is accessed using rte_eal_get_configuration(). 645The lookup (by name) of a memory zone returns a descriptor containing the physical address of the memory zone. 646 647Memory zones can be reserved with specific start address alignment by supplying the align parameter 648(by default, they are aligned to cache line size). 649The alignment value should be a power of two and not less than the cache line size (64 bytes). 650Memory zones can also be reserved from either 2 MB or 1 GB hugepages, provided that both are available on the system. 651 652Both memsegs and memzones are stored using ``rte_fbarray`` structures. Please 653refer to *DPDK API Reference* for more information. 654 655 656Multiple pthread 657---------------- 658 659DPDK usually pins one pthread per core to avoid the overhead of task switching. 660This allows for significant performance gains, but lacks flexibility and is not always efficient. 661 662Power management helps to improve the CPU efficiency by limiting the CPU runtime frequency. 663However, alternately it is possible to utilize the idle cycles available to take advantage of 664the full capability of the CPU. 665 666By taking advantage of cgroup, the CPU utilization quota can be simply assigned. 667This gives another way to improve the CPU efficiency, however, there is a prerequisite; 668DPDK must handle the context switching between multiple pthreads per core. 669 670For further flexibility, it is useful to set pthread affinity not only to a CPU but to a CPU set. 671 672EAL pthread and lcore Affinity 673~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 674 675The term "lcore" refers to an EAL thread, which is really a Linux/FreeBSD pthread. 676"EAL pthreads" are created and managed by EAL and execute the tasks issued by *remote_launch*. 677In each EAL pthread, there is a TLS (Thread Local Storage) called *_lcore_id* for unique identification. 678As EAL pthreads usually bind 1:1 to the physical CPU, the *_lcore_id* is typically equal to the CPU ID. 679 680When using multiple pthreads, however, the binding is no longer always 1:1 between an EAL pthread and a specified physical CPU. 681The EAL pthread may have affinity to a CPU set, and as such the *_lcore_id* will not be the same as the CPU ID. 682For this reason, there is an EAL long option '--lcores' defined to assign the CPU affinity of lcores. 683For a specified lcore ID or ID group, the option allows setting the CPU set for that EAL pthread. 684 685The format pattern: 686 --lcores='<lcore_set>[@cpu_set][,<lcore_set>[@cpu_set],...]' 687 688'lcore_set' and 'cpu_set' can be a single number, range or a group. 689 690A number is a "digit([0-9]+)"; a range is "<number>-<number>"; a group is "(<number|range>[,<number|range>,...])". 691 692If a '\@cpu_set' value is not supplied, the value of 'cpu_set' will default to the value of 'lcore_set'. 693 694 :: 695 696 For example, "--lcores='1,2@(5-7),(3-5)@(0,2),(0,6),7-8'" which means start 9 EAL thread; 697 lcore 0 runs on cpuset 0x41 (cpu 0,6); 698 lcore 1 runs on cpuset 0x2 (cpu 1); 699 lcore 2 runs on cpuset 0xe0 (cpu 5,6,7); 700 lcore 3,4,5 runs on cpuset 0x5 (cpu 0,2); 701 lcore 6 runs on cpuset 0x41 (cpu 0,6); 702 lcore 7 runs on cpuset 0x80 (cpu 7); 703 lcore 8 runs on cpuset 0x100 (cpu 8). 704 705Using this option, for each given lcore ID, the associated CPUs can be assigned. 706It's also compatible with the pattern of corelist('-l') option. 707 708non-EAL pthread support 709~~~~~~~~~~~~~~~~~~~~~~~ 710 711It is possible to use the DPDK execution context with any user pthread (aka. non-EAL pthreads). 712There are two kinds of non-EAL pthreads: 713 714- a registered non-EAL pthread with a valid *_lcore_id* that was successfully assigned by calling ``rte_thread_register()``, 715- a non registered non-EAL pthread with a LCORE_ID_ANY, 716 717For 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). 718 719All these impacts are mentioned in :ref:`known_issue_label` section. 720 721Public Thread API 722~~~~~~~~~~~~~~~~~ 723 724There are two public APIs ``rte_thread_set_affinity()`` and ``rte_thread_get_affinity()`` introduced for threads. 725When they're used in any pthread context, the Thread Local Storage(TLS) will be set/get. 726 727Those TLS include *_cpuset* and *_socket_id*: 728 729* *_cpuset* stores the CPUs bitmap to which the pthread is affinitized. 730 731* *_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. 732 733 734Control Thread API 735~~~~~~~~~~~~~~~~~~ 736 737It is possible to create Control Threads using the public API 738``rte_ctrl_thread_create()``. 739Those threads can be used for management/infrastructure tasks and are used 740internally by DPDK for multi process support and interrupt handling. 741 742Those threads will be scheduled on CPUs part of the original process CPU 743affinity from which the dataplane and service lcores are excluded. 744 745For example, on a 8 CPUs system, starting a dpdk application with -l 2,3 746(dataplane cores), then depending on the affinity configuration which can be 747controlled with tools like taskset (Linux) or cpuset (FreeBSD), 748 749- with no affinity configuration, the Control Threads will end up on 750 0-1,4-7 CPUs. 751- with affinity restricted to 2-4, the Control Threads will end up on 752 CPU 4. 753- with affinity restricted to 2-3, the Control Threads will end up on 754 CPU 2 (main lcore, which is the default when no CPU is available). 755 756.. _known_issue_label: 757 758Known Issues 759~~~~~~~~~~~~ 760 761+ rte_mempool 762 763 The rte_mempool uses a per-lcore cache inside the mempool. 764 For unregistered non-EAL pthreads, ``rte_lcore_id()`` will not return a valid number. 765 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. 766 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. 767 768+ rte_ring 769 770 rte_ring supports multi-producer enqueue and multi-consumer dequeue. 771 However, it is non-preemptive, this has a knock on effect of making rte_mempool non-preemptible. 772 773 .. note:: 774 775 The "non-preemptive" constraint means: 776 777 - a pthread doing multi-producers enqueues on a given ring must not 778 be preempted by another pthread doing a multi-producer enqueue on 779 the same ring. 780 - a pthread doing multi-consumers dequeues on a given ring must not 781 be preempted by another pthread doing a multi-consumer dequeue on 782 the same ring. 783 784 Bypassing this constraint may cause the 2nd pthread to spin until the 1st one is scheduled again. 785 Moreover, if the 1st pthread is preempted by a context that has an higher priority, it may even cause a dead lock. 786 787 This means, use cases involving preemptible pthreads should consider using rte_ring carefully. 788 789 1. It CAN be used for preemptible single-producer and single-consumer use case. 790 791 2. It CAN be used for non-preemptible multi-producer and preemptible single-consumer use case. 792 793 3. It CAN be used for preemptible single-producer and non-preemptible multi-consumer use case. 794 795 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. 796 797 5. It MUST not be used by multi-producer/consumer pthreads, whose scheduling policies are SCHED_FIFO or SCHED_RR. 798 799 Alternatively, applications can use the lock-free stack mempool handler. When 800 considering this handler, note that: 801 802 - It is currently limited to the aarch64 and x86_64 platforms, because it uses 803 an instruction (16-byte compare-and-swap) that is not yet available on other 804 platforms. 805 - It has worse average-case performance than the non-preemptive rte_ring, but 806 software caching (e.g. the mempool cache) can mitigate this by reducing the 807 number of stack accesses. 808 809+ rte_timer 810 811 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. 812 813+ rte_log 814 815 In unregistered non-EAL pthreads, there is no per thread loglevel and logtype, global loglevels are used. 816 817+ misc 818 819 The debug statistics of rte_ring, rte_mempool and rte_timer are not supported in an unregistered non-EAL pthread. 820 821cgroup control 822~~~~~~~~~~~~~~ 823 824The 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). 825We expect only 50% of CPU spend on packet IO. 826 827 .. code-block:: console 828 829 mkdir /sys/fs/cgroup/cpu/pkt_io 830 mkdir /sys/fs/cgroup/cpuset/pkt_io 831 832 echo $cpu > /sys/fs/cgroup/cpuset/cpuset.cpus 833 834 echo $t0 > /sys/fs/cgroup/cpu/pkt_io/tasks 835 echo $t0 > /sys/fs/cgroup/cpuset/pkt_io/tasks 836 837 echo $t1 > /sys/fs/cgroup/cpu/pkt_io/tasks 838 echo $t1 > /sys/fs/cgroup/cpuset/pkt_io/tasks 839 840 cd /sys/fs/cgroup/cpu/pkt_io 841 echo 100000 > pkt_io/cpu.cfs_period_us 842 echo 50000 > pkt_io/cpu.cfs_quota_us 843 844.. _malloc: 845 846Malloc 847------ 848 849The EAL provides a malloc API to allocate any-sized memory. 850 851The objective of this API is to provide malloc-like functions to allow 852allocation from hugepage memory and to facilitate application porting. 853The *DPDK API Reference* manual describes the available functions. 854 855Typically, these kinds of allocations should not be done in data plane 856processing because they are slower than pool-based allocation and make 857use of locks within the allocation and free paths. 858However, they can be used in configuration code. 859 860Refer to the rte_malloc() function description in the *DPDK API Reference* 861manual for more information. 862 863 864Alignment and NUMA Constraints 865~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 866 867The rte_malloc() takes an align argument that can be used to request a memory 868area that is aligned on a multiple of this value (which must be a power of two). 869 870On systems with NUMA support, a call to the rte_malloc() function will return 871memory that has been allocated on the NUMA socket of the core which made the call. 872A set of APIs is also provided, to allow memory to be explicitly allocated on a 873NUMA socket directly, or by allocated on the NUMA socket where another core is 874located, in the case where the memory is to be used by a logical core other than 875on the one doing the memory allocation. 876 877Use Cases 878~~~~~~~~~ 879 880This API is meant to be used by an application that requires malloc-like 881functions at initialization time. 882 883For allocating/freeing data at runtime, in the fast-path of an application, 884the memory pool library should be used instead. 885 886Internal Implementation 887~~~~~~~~~~~~~~~~~~~~~~~ 888 889Data Structures 890^^^^^^^^^^^^^^^ 891 892There are two data structure types used internally in the malloc library: 893 894* struct malloc_heap - used to track free space on a per-socket basis 895 896* struct malloc_elem - the basic element of allocation and free-space 897 tracking inside the library. 898 899Structure: malloc_heap 900"""""""""""""""""""""" 901 902The malloc_heap structure is used to manage free space on a per-socket basis. 903Internally, there is one heap structure per NUMA node, which allows us to 904allocate memory to a thread based on the NUMA node on which this thread runs. 905While this does not guarantee that the memory will be used on that NUMA node, 906it is no worse than a scheme where the memory is always allocated on a fixed 907or random node. 908 909The key fields of the heap structure and their function are described below 910(see also diagram above): 911 912* lock - the lock field is needed to synchronize access to the heap. 913 Given that the free space in the heap is tracked using a linked list, 914 we need a lock to prevent two threads manipulating the list at the same time. 915 916* free_head - this points to the first element in the list of free nodes for 917 this malloc heap. 918 919* first - this points to the first element in the heap. 920 921* last - this points to the last element in the heap. 922 923.. _figure_malloc_heap: 924 925.. figure:: img/malloc_heap.* 926 927 Example of a malloc heap and malloc elements within the malloc library 928 929 930.. _malloc_elem: 931 932Structure: malloc_elem 933"""""""""""""""""""""" 934 935The malloc_elem structure is used as a generic header structure for various 936blocks of memory. 937It is used in two different ways - all shown in the diagram above: 938 939#. As a header on a block of free or allocated memory - normal case 940 941#. As a padding header inside a block of memory 942 943The most important fields in the structure and how they are used are described below. 944 945Malloc heap is a doubly-linked list, where each element keeps track of its 946previous and next elements. Due to the fact that hugepage memory can come and 947go, neighboring malloc elements may not necessarily be adjacent in memory. 948Also, since a malloc element may span multiple pages, its contents may not 949necessarily be IOVA-contiguous either - each malloc element is only guaranteed 950to be virtually contiguous. 951 952.. note:: 953 954 If the usage of a particular field in one of the above three usages is not 955 described, the field can be assumed to have an undefined value in that 956 situation, for example, for padding headers only the "state" and "pad" 957 fields have valid values. 958 959* heap - this pointer is a reference back to the heap structure from which 960 this block was allocated. 961 It is used for normal memory blocks when they are being freed, to add the 962 newly-freed block to the heap's free-list. 963 964* prev - this pointer points to previous header element/block in memory. When 965 freeing a block, this pointer is used to reference the previous block to 966 check if that block is also free. If so, and the two blocks are immediately 967 adjacent to each other, then the two free blocks are merged to form a single 968 larger block. 969 970* next - this pointer points to next header element/block in memory. When 971 freeing a block, this pointer is used to reference the next block to check 972 if that block is also free. If so, and the two blocks are immediately 973 adjacent to each other, then the two free blocks are merged to form a single 974 larger block. 975 976* free_list - this is a structure pointing to previous and next elements in 977 this heap's free list. 978 It is only used in normal memory blocks; on ``malloc()`` to find a suitable 979 free block to allocate and on ``free()`` to add the newly freed element to 980 the free-list. 981 982* state - This field can have one of three values: ``FREE``, ``BUSY`` or 983 ``PAD``. 984 The former two are to indicate the allocation state of a normal memory block 985 and the latter is to indicate that the element structure is a dummy structure 986 at the end of the start-of-block padding, i.e. where the start of the data 987 within a block is not at the start of the block itself, due to alignment 988 constraints. 989 In that case, the pad header is used to locate the actual malloc element 990 header for the block. 991 992* dirty - this flag is only meaningful when ``state`` is ``FREE``. 993 It indicates that the content of the element is not fully zero-filled. 994 Memory from such blocks must be cleared when requested via ``rte_zmalloc*()``. 995 Dirty elements only appear with ``--huge-unlink=never``. 996 997* pad - this holds the length of the padding present at the start of the block. 998 In the case of a normal block header, it is added to the address of the end 999 of the header to give the address of the start of the data area, i.e. the 1000 value passed back to the application on a malloc. 1001 Within a dummy header inside the padding, this same value is stored, and is 1002 subtracted from the address of the dummy header to yield the address of the 1003 actual block header. 1004 1005* size - the size of the data block, including the header itself. 1006 1007Memory Allocation 1008^^^^^^^^^^^^^^^^^ 1009 1010On EAL initialization, all preallocated memory segments are setup as part of the 1011malloc heap. This setup involves placing an :ref:`element header<malloc_elem>` 1012with ``FREE`` at the start of each virtually contiguous segment of memory. 1013The ``FREE`` element is then added to the ``free_list`` for the malloc heap. 1014 1015This setup also happens whenever memory is allocated at runtime (if supported), 1016in which case newly allocated pages are also added to the heap, merging with any 1017adjacent free segments if there are any. 1018 1019When an application makes a call to a malloc-like function, the malloc function 1020will first index the ``lcore_config`` structure for the calling thread, and 1021determine the NUMA node of that thread. 1022The NUMA node is used to index the array of ``malloc_heap`` structures which is 1023passed as a parameter to the ``heap_alloc()`` function, along with the 1024requested size, type, alignment and boundary parameters. 1025 1026The ``heap_alloc()`` function will scan the free_list of the heap, and attempt 1027to find a free block suitable for storing data of the requested size, with the 1028requested alignment and boundary constraints. 1029 1030When a suitable free element has been identified, the pointer to be returned 1031to the user is calculated. 1032The cache-line of memory immediately preceding this pointer is filled with a 1033struct malloc_elem header. 1034Because of alignment and boundary constraints, there could be free space at 1035the start and/or end of the element, resulting in the following behavior: 1036 1037#. Check for trailing space. 1038 If the trailing space is big enough, i.e. > 128 bytes, then the free element 1039 is split. 1040 If it is not, then we just ignore it (wasted space). 1041 1042#. Check for space at the start of the element. 1043 If the space at the start is small, i.e. <=128 bytes, then a pad header is 1044 used, and the remaining space is wasted. 1045 If, however, the remaining space is greater, then the free element is split. 1046 1047The advantage of allocating the memory from the end of the existing element is 1048that no adjustment of the free list needs to take place - the existing element 1049on the free list just has its size value adjusted, and the next/previous elements 1050have their "prev"/"next" pointers redirected to the newly created element. 1051 1052In case when there is not enough memory in the heap to satisfy allocation 1053request, EAL will attempt to allocate more memory from the system (if supported) 1054and, following successful allocation, will retry reserving the memory again. In 1055a multiprocessing scenario, all primary and secondary processes will synchronize 1056their memory maps to ensure that any valid pointer to DPDK memory is guaranteed 1057to be valid at all times in all currently running processes. 1058 1059Failure to synchronize memory maps in one of the processes will cause allocation 1060to fail, even though some of the processes may have allocated the memory 1061successfully. The memory is not added to the malloc heap unless primary process 1062has ensured that all other processes have mapped this memory successfully. 1063 1064Any successful allocation event will trigger a callback, for which user 1065applications and other DPDK subsystems can register. Additionally, validation 1066callbacks will be triggered before allocation if the newly allocated memory will 1067exceed threshold set by the user, giving a chance to allow or deny allocation. 1068 1069.. note:: 1070 1071 Any allocation of new pages has to go through primary process. If the 1072 primary process is not active, no memory will be allocated even if it was 1073 theoretically possible to do so. This is because primary's process map acts 1074 as an authority on what should or should not be mapped, while each secondary 1075 process has its own, local memory map. Secondary processes do not update the 1076 shared memory map, they only copy its contents to their local memory map. 1077 1078Freeing Memory 1079^^^^^^^^^^^^^^ 1080 1081To free an area of memory, the pointer to the start of the data area is passed 1082to the free function. 1083The size of the ``malloc_elem`` structure is subtracted from this pointer to get 1084the element header for the block. 1085If this header is of type ``PAD`` then the pad length is further subtracted from 1086the pointer to get the proper element header for the entire block. 1087 1088From this element header, we get pointers to the heap from which the block was 1089allocated and to where it must be freed, as well as the pointer to the previous 1090and next elements. These next and previous elements are then checked to see if 1091they are also ``FREE`` and are immediately adjacent to the current one, and if 1092so, they are merged with the current element. This means that we can never have 1093two ``FREE`` memory blocks adjacent to one another, as they are always merged 1094into a single block. 1095 1096If deallocating pages at runtime is supported, and the free element encloses 1097one or more pages, those pages can be deallocated and be removed from the heap. 1098If DPDK was started with command-line parameters for preallocating memory 1099(``-m`` or ``--socket-mem``), then those pages that were allocated at startup 1100will not be deallocated. 1101 1102Any successful deallocation event will trigger a callback, for which user 1103applications and other DPDK subsystems can register. 1104