1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
11 * Memory-related RTE API.
22 #include <rte_common.h>
23 #include <rte_compat.h>
24 #include <rte_config.h>
26 /* forward declaration for pointers */
27 struct rte_memseg_list;
31 RTE_PGSIZE_4K = 1ULL << 12,
32 RTE_PGSIZE_64K = 1ULL << 16,
33 RTE_PGSIZE_256K = 1ULL << 18,
34 RTE_PGSIZE_2M = 1ULL << 21,
35 RTE_PGSIZE_16M = 1ULL << 24,
36 RTE_PGSIZE_256M = 1ULL << 28,
37 RTE_PGSIZE_512M = 1ULL << 29,
38 RTE_PGSIZE_1G = 1ULL << 30,
39 RTE_PGSIZE_4G = 1ULL << 32,
40 RTE_PGSIZE_16G = 1ULL << 34,
43 #define SOCKET_ID_ANY -1 /**< Any NUMA socket. */
44 #define RTE_CACHE_LINE_MASK (RTE_CACHE_LINE_SIZE-1) /**< Cache line mask. */
46 #define RTE_CACHE_LINE_ROUNDUP(size) \
47 (RTE_CACHE_LINE_SIZE * ((size + RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE))
48 /**< Return the first cache-aligned value greater or equal to size. */
50 /**< Cache line size in terms of log2 */
51 #if RTE_CACHE_LINE_SIZE == 64
52 #define RTE_CACHE_LINE_SIZE_LOG2 6
53 #elif RTE_CACHE_LINE_SIZE == 128
54 #define RTE_CACHE_LINE_SIZE_LOG2 7
56 #error "Unsupported cache line size"
59 #define RTE_CACHE_LINE_MIN_SIZE 64 /**< Minimum Cache line size. */
62 * Force alignment to cache line.
64 #define __rte_cache_aligned __rte_aligned(RTE_CACHE_LINE_SIZE)
67 * Force minimum cache line alignment.
69 #define __rte_cache_min_aligned __rte_aligned(RTE_CACHE_LINE_MIN_SIZE)
71 typedef uint64_t phys_addr_t; /**< Physical address. */
72 #define RTE_BAD_PHYS_ADDR ((phys_addr_t)-1)
74 * IO virtual address type.
75 * When the physical addressing mode (IOVA as PA) is in use,
76 * the translation from an IO virtual address (IOVA) to a physical address
77 * is a direct mapping, i.e. the same value.
78 * Otherwise, in virtual mode (IOVA as VA), an IOMMU may do the translation.
80 typedef uint64_t rte_iova_t;
81 #define RTE_BAD_IOVA ((rte_iova_t)-1)
84 * Physical memory segment descriptor.
86 #define RTE_MEMSEG_FLAG_DO_NOT_FREE (1 << 0)
87 /**< Prevent this segment from being freed back to the OS. */
91 phys_addr_t phys_addr; /**< deprecated - Start physical address. */
92 rte_iova_t iova; /**< Start IO address. */
96 void *addr; /**< Start virtual address. */
97 uint64_t addr_64; /**< Makes sure addr is always 64 bits */
99 size_t len; /**< Length of the segment. */
100 uint64_t hugepage_sz; /**< The pagesize of underlying memory */
101 int32_t socket_id; /**< NUMA socket ID. */
102 uint32_t nchannel; /**< Number of channels. */
103 uint32_t nrank; /**< Number of ranks. */
104 uint32_t flags; /**< Memseg-specific flags */
108 * Lock page in physical memory and prevent from swapping.
111 * The virtual address.
113 * 0 on success, negative on error.
115 int rte_mem_lock_page(const void *virt);
118 * Get physical address of any mapped virtual address in the current process.
119 * It is found by browsing the /proc/self/pagemap special file.
120 * The page must be locked.
123 * The virtual address.
125 * The physical address or RTE_BAD_IOVA on error.
127 phys_addr_t rte_mem_virt2phy(const void *virt);
130 * Get IO virtual address of any mapped virtual address in the current process.
133 * The virtual address.
135 * The IO address or RTE_BAD_IOVA on error.
137 rte_iova_t rte_mem_virt2iova(const void *virt);
140 * Get virtual memory address corresponding to iova address.
142 * @note This function read-locks the memory hotplug subsystem, and thus cannot
143 * be used within memory-related callback functions.
148 * Virtual address corresponding to iova address (or NULL if address does not
149 * exist within DPDK memory map).
151 __rte_experimental void *
152 rte_mem_iova2virt(rte_iova_t iova);
155 * Get memseg to which a particular virtual address belongs.
158 * The virtual address.
160 * The memseg list in which to look up based on ``virt`` address
163 * Memseg pointer on success, or NULL on error.
165 __rte_experimental struct rte_memseg *
166 rte_mem_virt2memseg(const void *virt, const struct rte_memseg_list *msl);
169 * Get memseg list corresponding to virtual memory address.
172 * The virtual address.
174 * Memseg list to which this virtual address belongs to.
176 __rte_experimental struct rte_memseg_list *
177 rte_mem_virt2memseg_list(const void *virt);
180 * Memseg walk function prototype.
182 * Returning 0 will continue walk
183 * Returning 1 will stop the walk
184 * Returning -1 will stop the walk and report error
186 typedef int (*rte_memseg_walk_t)(const struct rte_memseg_list *msl,
187 const struct rte_memseg *ms, void *arg);
190 * Memseg contig walk function prototype. This will trigger a callback on every
191 * VA-contiguous are starting at memseg ``ms``, so total valid VA space at each
192 * callback call will be [``ms->addr``, ``ms->addr + len``).
194 * Returning 0 will continue walk
195 * Returning 1 will stop the walk
196 * Returning -1 will stop the walk and report error
198 typedef int (*rte_memseg_contig_walk_t)(const struct rte_memseg_list *msl,
199 const struct rte_memseg *ms, size_t len, void *arg);
202 * Memseg list walk function prototype. This will trigger a callback on every
203 * allocated memseg list.
205 * Returning 0 will continue walk
206 * Returning 1 will stop the walk
207 * Returning -1 will stop the walk and report error
209 typedef int (*rte_memseg_list_walk_t)(const struct rte_memseg_list *msl,
213 * Walk list of all memsegs.
215 * @note This function read-locks the memory hotplug subsystem, and thus cannot
216 * be used within memory-related callback functions.
218 * @note This function will also walk through externally allocated segments. It
219 * is up to the user to decide whether to skip through these segments.
224 * Argument passed to iterator
226 * 0 if walked over the entire list
227 * 1 if stopped by the user
228 * -1 if user function reported error
230 int __rte_experimental
231 rte_memseg_walk(rte_memseg_walk_t func, void *arg);
234 * Walk each VA-contiguous area.
236 * @note This function read-locks the memory hotplug subsystem, and thus cannot
237 * be used within memory-related callback functions.
239 * @note This function will also walk through externally allocated segments. It
240 * is up to the user to decide whether to skip through these segments.
245 * Argument passed to iterator
247 * 0 if walked over the entire list
248 * 1 if stopped by the user
249 * -1 if user function reported error
251 int __rte_experimental
252 rte_memseg_contig_walk(rte_memseg_contig_walk_t func, void *arg);
255 * Walk each allocated memseg list.
257 * @note This function read-locks the memory hotplug subsystem, and thus cannot
258 * be used within memory-related callback functions.
260 * @note This function will also walk through externally allocated segments. It
261 * is up to the user to decide whether to skip through these segments.
266 * Argument passed to iterator
268 * 0 if walked over the entire list
269 * 1 if stopped by the user
270 * -1 if user function reported error
272 int __rte_experimental
273 rte_memseg_list_walk(rte_memseg_list_walk_t func, void *arg);
276 * Walk list of all memsegs without performing any locking.
278 * @note This function does not perform any locking, and is only safe to call
279 * from within memory-related callback functions.
284 * Argument passed to iterator
286 * 0 if walked over the entire list
287 * 1 if stopped by the user
288 * -1 if user function reported error
290 int __rte_experimental
291 rte_memseg_walk_thread_unsafe(rte_memseg_walk_t func, void *arg);
294 * Walk each VA-contiguous area without performing any locking.
296 * @note This function does not perform any locking, and is only safe to call
297 * from within memory-related callback functions.
302 * Argument passed to iterator
304 * 0 if walked over the entire list
305 * 1 if stopped by the user
306 * -1 if user function reported error
308 int __rte_experimental
309 rte_memseg_contig_walk_thread_unsafe(rte_memseg_contig_walk_t func, void *arg);
312 * Walk each allocated memseg list without performing any locking.
314 * @note This function does not perform any locking, and is only safe to call
315 * from within memory-related callback functions.
320 * Argument passed to iterator
322 * 0 if walked over the entire list
323 * 1 if stopped by the user
324 * -1 if user function reported error
326 int __rte_experimental
327 rte_memseg_list_walk_thread_unsafe(rte_memseg_list_walk_t func, void *arg);
330 * Return file descriptor associated with a particular memseg (if available).
332 * @note This function read-locks the memory hotplug subsystem, and thus cannot
333 * be used within memory-related callback functions.
335 * @note This returns an internal file descriptor. Performing any operations on
336 * this file descriptor is inherently dangerous, so it should be treated
337 * as read-only for all intents and purposes.
340 * A pointer to memseg for which to get file descriptor.
343 * Valid file descriptor in case of success.
344 * -1 in case of error, with ``rte_errno`` set to the following values:
345 * - EINVAL - ``ms`` pointer was NULL or did not point to a valid memseg
346 * - ENODEV - ``ms`` fd is not available
347 * - ENOENT - ``ms`` is an unused segment
348 * - ENOTSUP - segment fd's are not supported
350 int __rte_experimental
351 rte_memseg_get_fd(const struct rte_memseg *ms);
354 * Return file descriptor associated with a particular memseg (if available).
356 * @note This function does not perform any locking, and is only safe to call
357 * from within memory-related callback functions.
359 * @note This returns an internal file descriptor. Performing any operations on
360 * this file descriptor is inherently dangerous, so it should be treated
361 * as read-only for all intents and purposes.
364 * A pointer to memseg for which to get file descriptor.
367 * Valid file descriptor in case of success.
368 * -1 in case of error, with ``rte_errno`` set to the following values:
369 * - EINVAL - ``ms`` pointer was NULL or did not point to a valid memseg
370 * - ENODEV - ``ms`` fd is not available
371 * - ENOENT - ``ms`` is an unused segment
372 * - ENOTSUP - segment fd's are not supported
374 int __rte_experimental
375 rte_memseg_get_fd_thread_unsafe(const struct rte_memseg *ms);
378 * Get offset into segment file descriptor associated with a particular memseg
381 * @note This function read-locks the memory hotplug subsystem, and thus cannot
382 * be used within memory-related callback functions.
385 * A pointer to memseg for which to get file descriptor.
387 * A pointer to offset value where the result will be stored.
390 * Valid file descriptor in case of success.
391 * -1 in case of error, with ``rte_errno`` set to the following values:
392 * - EINVAL - ``ms`` pointer was NULL or did not point to a valid memseg
393 * - EINVAL - ``offset`` pointer was NULL
394 * - ENODEV - ``ms`` fd is not available
395 * - ENOENT - ``ms`` is an unused segment
396 * - ENOTSUP - segment fd's are not supported
398 int __rte_experimental
399 rte_memseg_get_fd_offset(const struct rte_memseg *ms, size_t *offset);
402 * Get offset into segment file descriptor associated with a particular memseg
405 * @note This function does not perform any locking, and is only safe to call
406 * from within memory-related callback functions.
409 * A pointer to memseg for which to get file descriptor.
411 * A pointer to offset value where the result will be stored.
414 * Valid file descriptor in case of success.
415 * -1 in case of error, with ``rte_errno`` set to the following values:
416 * - EINVAL - ``ms`` pointer was NULL or did not point to a valid memseg
417 * - EINVAL - ``offset`` pointer was NULL
418 * - ENODEV - ``ms`` fd is not available
419 * - ENOENT - ``ms`` is an unused segment
420 * - ENOTSUP - segment fd's are not supported
422 int __rte_experimental
423 rte_memseg_get_fd_offset_thread_unsafe(const struct rte_memseg *ms,
428 * @b EXPERIMENTAL: this API may change without prior notice
430 * Register external memory chunk with DPDK.
432 * @note Using this API is mutually exclusive with ``rte_malloc`` family of
435 * @note This API will not perform any DMA mapping. It is expected that user
436 * will do that themselves.
438 * @note Before accessing this memory in other processes, it needs to be
439 * attached in each of those processes by calling ``rte_extmem_attach`` in
440 * each other process.
443 * Start of virtual area to register. Must be aligned by ``page_sz``.
445 * Length of virtual area to register. Must be aligned by ``page_sz``.
447 * Array of page IOVA addresses corresponding to each page in this memory
448 * area. Can be NULL, in which case page IOVA addresses will be set to
451 * Number of elements in the iova_addrs array. Ignored if ``iova_addrs``
454 * Page size of the underlying memory
458 * - -1 in case of error, with rte_errno set to one of the following:
459 * EINVAL - one of the parameters was invalid
460 * EEXIST - memory chunk is already registered
461 * ENOSPC - no more space in internal config to store a new memory chunk
463 int __rte_experimental
464 rte_extmem_register(void *va_addr, size_t len, rte_iova_t iova_addrs[],
465 unsigned int n_pages, size_t page_sz);
469 * @b EXPERIMENTAL: this API may change without prior notice
471 * Unregister external memory chunk with DPDK.
473 * @note Using this API is mutually exclusive with ``rte_malloc`` family of
476 * @note This API will not perform any DMA unmapping. It is expected that user
477 * will do that themselves.
479 * @note Before calling this function, all other processes must call
480 * ``rte_extmem_detach`` to detach from the memory area.
483 * Start of virtual area to unregister
485 * Length of virtual area to unregister
489 * - -1 in case of error, with rte_errno set to one of the following:
490 * EINVAL - one of the parameters was invalid
491 * ENOENT - memory chunk was not found
493 int __rte_experimental
494 rte_extmem_unregister(void *va_addr, size_t len);
498 * @b EXPERIMENTAL: this API may change without prior notice
500 * Attach to external memory chunk registered in another process.
502 * @note Using this API is mutually exclusive with ``rte_malloc`` family of
505 * @note This API will not perform any DMA mapping. It is expected that user
506 * will do that themselves.
509 * Start of virtual area to register
511 * Length of virtual area to register
515 * - -1 in case of error, with rte_errno set to one of the following:
516 * EINVAL - one of the parameters was invalid
517 * ENOENT - memory chunk was not found
519 int __rte_experimental
520 rte_extmem_attach(void *va_addr, size_t len);
524 * @b EXPERIMENTAL: this API may change without prior notice
526 * Detach from external memory chunk registered in another process.
528 * @note Using this API is mutually exclusive with ``rte_malloc`` family of
531 * @note This API will not perform any DMA unmapping. It is expected that user
532 * will do that themselves.
535 * Start of virtual area to unregister
537 * Length of virtual area to unregister
541 * - -1 in case of error, with rte_errno set to one of the following:
542 * EINVAL - one of the parameters was invalid
543 * ENOENT - memory chunk was not found
545 int __rte_experimental
546 rte_extmem_detach(void *va_addr, size_t len);
549 * Dump the physical memory layout to a file.
551 * @note This function read-locks the memory hotplug subsystem, and thus cannot
552 * be used within memory-related callback functions.
555 * A pointer to a file for output
557 void rte_dump_physmem_layout(FILE *f);
560 * Get the total amount of available physical memory.
562 * @note This function read-locks the memory hotplug subsystem, and thus cannot
563 * be used within memory-related callback functions.
566 * The total amount of available physical memory in bytes.
568 uint64_t rte_eal_get_physmem_size(void);
571 * Get the number of memory channels.
574 * The number of memory channels on the system. The value is 0 if unknown
575 * or not the same on all devices.
577 unsigned rte_memory_get_nchannel(void);
580 * Get the number of memory ranks.
583 * The number of memory ranks on the system. The value is 0 if unknown or
584 * not the same on all devices.
586 unsigned rte_memory_get_nrank(void);
590 * @b EXPERIMENTAL: this API may change without prior notice
592 * Check if all currently allocated memory segments are compliant with
593 * supplied DMA address width.
596 * Address width to check against.
598 int __rte_experimental rte_mem_check_dma_mask(uint8_t maskbits);
602 * @b EXPERIMENTAL: this API may change without prior notice
604 * Check if all currently allocated memory segments are compliant with
605 * supplied DMA address width. This function will use
606 * rte_memseg_walk_thread_unsafe instead of rte_memseg_walk implying
607 * memory_hotplug_lock will not be acquired avoiding deadlock during
608 * memory initialization.
610 * This function is just for EAL core memory internal use. Drivers should
611 * use the previous rte_mem_check_dma_mask.
614 * Address width to check against.
616 int __rte_experimental rte_mem_check_dma_mask_thread_unsafe(uint8_t maskbits);
620 * @b EXPERIMENTAL: this API may change without prior notice
622 * Set dma mask to use once memory initialization is done. Previous functions
623 * rte_mem_check_dma_mask and rte_mem_check_dma_mask_thread_unsafe can not be
624 * used safely until memory has been initialized.
626 void __rte_experimental rte_mem_set_dma_mask(uint8_t maskbits);
629 * Drivers based on uio will not load unless physical
630 * addresses are obtainable. It is only possible to get
631 * physical addresses when running as a privileged user.
634 * 1 if the system is able to obtain physical addresses.
635 * 0 if using DMA addresses through an IOMMU.
637 int rte_eal_using_phys_addrs(void);
641 * Enum indicating which kind of memory event has happened. Used by callbacks to
642 * distinguish between memory allocations and deallocations.
645 RTE_MEM_EVENT_ALLOC = 0, /**< Allocation event. */
646 RTE_MEM_EVENT_FREE, /**< Deallocation event. */
648 #define RTE_MEM_EVENT_CALLBACK_NAME_LEN 64
649 /**< maximum length of callback name */
652 * Function typedef used to register callbacks for memory events.
654 typedef void (*rte_mem_event_callback_t)(enum rte_mem_event event_type,
655 const void *addr, size_t len, void *arg);
658 * Function used to register callbacks for memory events.
660 * @note callbacks will happen while memory hotplug subsystem is write-locked,
661 * therefore some functions (e.g. `rte_memseg_walk()`) will cause a
662 * deadlock when called from within such callbacks.
664 * @note mem event callbacks not being supported is an expected error condition,
665 * so user code needs to handle this situation. In these cases, return
666 * value will be -1, and rte_errno will be set to ENOTSUP.
669 * Name associated with specified callback to be added to the list.
672 * Callback function pointer.
675 * Argument to pass to the callback.
678 * 0 on successful callback register
679 * -1 on unsuccessful callback register, with rte_errno value indicating
680 * reason for failure.
682 int __rte_experimental
683 rte_mem_event_callback_register(const char *name, rte_mem_event_callback_t clb,
687 * Function used to unregister callbacks for memory events.
690 * Name associated with specified callback to be removed from the list.
693 * Argument to look for among callbacks with specified callback name.
696 * 0 on successful callback unregister
697 * -1 on unsuccessful callback unregister, with rte_errno value indicating
698 * reason for failure.
700 int __rte_experimental
701 rte_mem_event_callback_unregister(const char *name, void *arg);
704 #define RTE_MEM_ALLOC_VALIDATOR_NAME_LEN 64
705 /**< maximum length of alloc validator name */
707 * Function typedef used to register memory allocation validation callbacks.
709 * Returning 0 will allow allocation attempt to continue. Returning -1 will
710 * prevent allocation from succeeding.
712 typedef int (*rte_mem_alloc_validator_t)(int socket_id,
713 size_t cur_limit, size_t new_len);
716 * @brief Register validator callback for memory allocations.
718 * Callbacks registered by this function will be called right before memory
719 * allocator is about to trigger allocation of more pages from the system if
720 * said allocation will bring total memory usage above specified limit on
721 * specified socket. User will be able to cancel pending allocation if callback
724 * @note callbacks will happen while memory hotplug subsystem is write-locked,
725 * therefore some functions (e.g. `rte_memseg_walk()`) will cause a
726 * deadlock when called from within such callbacks.
728 * @note validator callbacks not being supported is an expected error condition,
729 * so user code needs to handle this situation. In these cases, return
730 * value will be -1, and rte_errno will be set to ENOTSUP.
733 * Name associated with specified callback to be added to the list.
736 * Callback function pointer.
739 * Socket ID on which to watch for allocations.
742 * Limit above which to trigger callbacks.
745 * 0 on successful callback register
746 * -1 on unsuccessful callback register, with rte_errno value indicating
747 * reason for failure.
749 int __rte_experimental
750 rte_mem_alloc_validator_register(const char *name,
751 rte_mem_alloc_validator_t clb, int socket_id, size_t limit);
754 * @brief Unregister validator callback for memory allocations.
757 * Name associated with specified callback to be removed from the list.
760 * Socket ID on which to watch for allocations.
763 * 0 on successful callback unregister
764 * -1 on unsuccessful callback unregister, with rte_errno value indicating
765 * reason for failure.
767 int __rte_experimental
768 rte_mem_alloc_validator_unregister(const char *name, int socket_id);
774 #endif /* _RTE_MEMORY_H_ */