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>
25 #include <rte_fbarray.h>
27 #define RTE_PGSIZE_4K (1ULL << 12)
28 #define RTE_PGSIZE_64K (1ULL << 16)
29 #define RTE_PGSIZE_256K (1ULL << 18)
30 #define RTE_PGSIZE_2M (1ULL << 21)
31 #define RTE_PGSIZE_16M (1ULL << 24)
32 #define RTE_PGSIZE_256M (1ULL << 28)
33 #define RTE_PGSIZE_512M (1ULL << 29)
34 #define RTE_PGSIZE_1G (1ULL << 30)
35 #define RTE_PGSIZE_4G (1ULL << 32)
36 #define RTE_PGSIZE_16G (1ULL << 34)
38 #define SOCKET_ID_ANY -1 /**< Any NUMA socket. */
41 * Physical memory segment descriptor.
43 #define RTE_MEMSEG_FLAG_DO_NOT_FREE (1 << 0)
44 /**< Prevent this segment from being freed back to the OS. */
46 rte_iova_t iova; /**< Start IO address. */
49 void *addr; /**< Start virtual address. */
50 uint64_t addr_64; /**< Makes sure addr is always 64 bits */
52 size_t len; /**< Length of the segment. */
53 uint64_t hugepage_sz; /**< The pagesize of underlying memory */
54 int32_t socket_id; /**< NUMA socket ID. */
55 uint32_t nchannel; /**< Number of channels. */
56 uint32_t nrank; /**< Number of ranks. */
57 uint32_t flags; /**< Memseg-specific flags */
61 * memseg list is a special case as we need to store a bunch of other data
62 * together with the array itself.
64 struct rte_memseg_list {
68 /**< Base virtual address for this memseg list. */
70 /**< Makes sure addr is always 64-bits */
72 uint64_t page_sz; /**< Page size for all memsegs in this list. */
73 int socket_id; /**< Socket ID for all memsegs in this list. */
74 volatile uint32_t version; /**< version number for multiprocess sync. */
75 size_t len; /**< Length of memory area covered by this memseg list. */
76 unsigned int external; /**< 1 if this list points to external memory */
77 unsigned int heap; /**< 1 if this list points to a heap */
78 struct rte_fbarray memseg_arr;
82 * Lock page in physical memory and prevent from swapping.
85 * The virtual address.
87 * 0 on success, negative on error.
89 int rte_mem_lock_page(const void *virt);
92 * Get physical address of any mapped virtual address in the current process.
93 * It is found by browsing the /proc/self/pagemap special file.
94 * The page must be locked.
97 * The virtual address.
99 * The physical address or RTE_BAD_IOVA on error.
101 phys_addr_t rte_mem_virt2phy(const void *virt);
104 * Get IO virtual address of any mapped virtual address in the current process.
106 * @note This function will not check internal page table. Instead, in IOVA as
107 * PA mode, it will fall back to getting real physical address (which may
108 * not match the expected IOVA, such as what was specified for external
112 * The virtual address.
114 * The IO address or RTE_BAD_IOVA on error.
116 rte_iova_t rte_mem_virt2iova(const void *virt);
119 * Get virtual memory address corresponding to iova address.
121 * @note This function read-locks the memory hotplug subsystem, and thus cannot
122 * be used within memory-related callback functions.
127 * Virtual address corresponding to iova address (or NULL if address does not
128 * exist within DPDK memory map).
132 rte_mem_iova2virt(rte_iova_t iova);
135 * Get memseg to which a particular virtual address belongs.
138 * The virtual address.
140 * The memseg list in which to look up based on ``virt`` address
143 * Memseg pointer on success, or NULL on error.
147 rte_mem_virt2memseg(const void *virt, const struct rte_memseg_list *msl);
150 * Get memseg list corresponding to virtual memory address.
153 * The virtual address.
155 * Memseg list to which this virtual address belongs to.
158 struct rte_memseg_list *
159 rte_mem_virt2memseg_list(const void *virt);
162 * Memseg walk function prototype.
164 * Returning 0 will continue walk
165 * Returning 1 will stop the walk
166 * Returning -1 will stop the walk and report error
168 typedef int (*rte_memseg_walk_t)(const struct rte_memseg_list *msl,
169 const struct rte_memseg *ms, void *arg);
172 * Memseg contig walk function prototype. This will trigger a callback on every
173 * VA-contiguous area starting at memseg ``ms``, so total valid VA space at each
174 * callback call will be [``ms->addr``, ``ms->addr + len``).
176 * Returning 0 will continue walk
177 * Returning 1 will stop the walk
178 * Returning -1 will stop the walk and report error
180 typedef int (*rte_memseg_contig_walk_t)(const struct rte_memseg_list *msl,
181 const struct rte_memseg *ms, size_t len, void *arg);
184 * Memseg list walk function prototype. This will trigger a callback on every
185 * allocated memseg list.
187 * Returning 0 will continue walk
188 * Returning 1 will stop the walk
189 * Returning -1 will stop the walk and report error
191 typedef int (*rte_memseg_list_walk_t)(const struct rte_memseg_list *msl,
195 * Walk list of all memsegs.
197 * @note This function read-locks the memory hotplug subsystem, and thus cannot
198 * be used within memory-related callback functions.
200 * @note This function will also walk through externally allocated segments. It
201 * is up to the user to decide whether to skip through these segments.
206 * Argument passed to iterator
208 * 0 if walked over the entire list
209 * 1 if stopped by the user
210 * -1 if user function reported error
214 rte_memseg_walk(rte_memseg_walk_t func, void *arg);
217 * Walk each VA-contiguous area.
219 * @note This function read-locks the memory hotplug subsystem, and thus cannot
220 * be used within memory-related callback functions.
222 * @note This function will also walk through externally allocated segments. It
223 * is up to the user to decide whether to skip through these segments.
228 * Argument passed to iterator
230 * 0 if walked over the entire list
231 * 1 if stopped by the user
232 * -1 if user function reported error
236 rte_memseg_contig_walk(rte_memseg_contig_walk_t func, void *arg);
239 * Walk each allocated memseg list.
241 * @note This function read-locks the memory hotplug subsystem, and thus cannot
242 * be used within memory-related callback functions.
244 * @note This function will also walk through externally allocated segments. It
245 * is up to the user to decide whether to skip through these segments.
250 * Argument passed to iterator
252 * 0 if walked over the entire list
253 * 1 if stopped by the user
254 * -1 if user function reported error
258 rte_memseg_list_walk(rte_memseg_list_walk_t func, void *arg);
261 * Walk list of all memsegs without performing any locking.
263 * @note This function does not perform any locking, and is only safe to call
264 * from within memory-related callback functions.
269 * Argument passed to iterator
271 * 0 if walked over the entire list
272 * 1 if stopped by the user
273 * -1 if user function reported error
277 rte_memseg_walk_thread_unsafe(rte_memseg_walk_t func, void *arg);
280 * Walk each VA-contiguous area without performing any locking.
282 * @note This function does not perform any locking, and is only safe to call
283 * from within memory-related callback functions.
288 * Argument passed to iterator
290 * 0 if walked over the entire list
291 * 1 if stopped by the user
292 * -1 if user function reported error
296 rte_memseg_contig_walk_thread_unsafe(rte_memseg_contig_walk_t func, void *arg);
299 * Walk each allocated memseg list without performing any locking.
301 * @note This function does not perform any locking, and is only safe to call
302 * from within memory-related callback functions.
307 * Argument passed to iterator
309 * 0 if walked over the entire list
310 * 1 if stopped by the user
311 * -1 if user function reported error
315 rte_memseg_list_walk_thread_unsafe(rte_memseg_list_walk_t func, void *arg);
318 * Return file descriptor associated with a particular memseg (if available).
320 * @note This function read-locks the memory hotplug subsystem, and thus cannot
321 * be used within memory-related callback functions.
323 * @note This returns an internal file descriptor. Performing any operations on
324 * this file descriptor is inherently dangerous, so it should be treated
325 * as read-only for all intents and purposes.
328 * A pointer to memseg for which to get file descriptor.
331 * Valid file descriptor in case of success.
332 * -1 in case of error, with ``rte_errno`` set to the following values:
333 * - EINVAL - ``ms`` pointer was NULL or did not point to a valid memseg
334 * - ENODEV - ``ms`` fd is not available
335 * - ENOENT - ``ms`` is an unused segment
336 * - ENOTSUP - segment fd's are not supported
340 rte_memseg_get_fd(const struct rte_memseg *ms);
343 * Return file descriptor associated with a particular memseg (if available).
345 * @note This function does not perform any locking, and is only safe to call
346 * from within memory-related callback functions.
348 * @note This returns an internal file descriptor. Performing any operations on
349 * this file descriptor is inherently dangerous, so it should be treated
350 * as read-only for all intents and purposes.
353 * A pointer to memseg for which to get file descriptor.
356 * Valid file descriptor in case of success.
357 * -1 in case of error, with ``rte_errno`` set to the following values:
358 * - EINVAL - ``ms`` pointer was NULL or did not point to a valid memseg
359 * - ENODEV - ``ms`` fd is not available
360 * - ENOENT - ``ms`` is an unused segment
361 * - ENOTSUP - segment fd's are not supported
365 rte_memseg_get_fd_thread_unsafe(const struct rte_memseg *ms);
368 * Get offset into segment file descriptor associated with a particular memseg
371 * @note This function read-locks the memory hotplug subsystem, and thus cannot
372 * be used within memory-related callback functions.
375 * A pointer to memseg for which to get file descriptor.
377 * A pointer to offset value where the result will be stored.
380 * Valid file descriptor in case of success.
381 * -1 in case of error, with ``rte_errno`` set to the following values:
382 * - EINVAL - ``ms`` pointer was NULL or did not point to a valid memseg
383 * - EINVAL - ``offset`` pointer was NULL
384 * - ENODEV - ``ms`` fd is not available
385 * - ENOENT - ``ms`` is an unused segment
386 * - ENOTSUP - segment fd's are not supported
390 rte_memseg_get_fd_offset(const struct rte_memseg *ms, size_t *offset);
393 * Get offset into segment file descriptor associated with a particular memseg
396 * @note This function does not perform any locking, and is only safe to call
397 * from within memory-related callback functions.
400 * A pointer to memseg for which to get file descriptor.
402 * A pointer to offset value where the result will be stored.
405 * Valid file descriptor in case of success.
406 * -1 in case of error, with ``rte_errno`` set to the following values:
407 * - EINVAL - ``ms`` pointer was NULL or did not point to a valid memseg
408 * - EINVAL - ``offset`` pointer was NULL
409 * - ENODEV - ``ms`` fd is not available
410 * - ENOENT - ``ms`` is an unused segment
411 * - ENOTSUP - segment fd's are not supported
415 rte_memseg_get_fd_offset_thread_unsafe(const struct rte_memseg *ms,
420 * @b EXPERIMENTAL: this API may change without prior notice
422 * Register external memory chunk with DPDK.
424 * @note Using this API is mutually exclusive with ``rte_malloc`` family of
427 * @note This API will not perform any DMA mapping. It is expected that user
428 * will do that themselves.
430 * @note Before accessing this memory in other processes, it needs to be
431 * attached in each of those processes by calling ``rte_extmem_attach`` in
432 * each other process.
435 * Start of virtual area to register. Must be aligned by ``page_sz``.
437 * Length of virtual area to register. Must be aligned by ``page_sz``.
439 * Array of page IOVA addresses corresponding to each page in this memory
440 * area. Can be NULL, in which case page IOVA addresses will be set to
443 * Number of elements in the iova_addrs array. Ignored if ``iova_addrs``
446 * Page size of the underlying memory
450 * - -1 in case of error, with rte_errno set to one of the following:
451 * EINVAL - one of the parameters was invalid
452 * EEXIST - memory chunk is already registered
453 * ENOSPC - no more space in internal config to store a new memory chunk
457 rte_extmem_register(void *va_addr, size_t len, rte_iova_t iova_addrs[],
458 unsigned int n_pages, size_t page_sz);
462 * @b EXPERIMENTAL: this API may change without prior notice
464 * Unregister external memory chunk with DPDK.
466 * @note Using this API is mutually exclusive with ``rte_malloc`` family of
469 * @note This API will not perform any DMA unmapping. It is expected that user
470 * will do that themselves.
472 * @note Before calling this function, all other processes must call
473 * ``rte_extmem_detach`` to detach from the memory area.
476 * Start of virtual area to unregister
478 * Length of virtual area to unregister
482 * - -1 in case of error, with rte_errno set to one of the following:
483 * EINVAL - one of the parameters was invalid
484 * ENOENT - memory chunk was not found
488 rte_extmem_unregister(void *va_addr, size_t len);
492 * @b EXPERIMENTAL: this API may change without prior notice
494 * Attach to external memory chunk registered in another process.
496 * @note Using this API is mutually exclusive with ``rte_malloc`` family of
499 * @note This API will not perform any DMA mapping. It is expected that user
500 * will do that themselves.
503 * Start of virtual area to register
505 * Length of virtual area to register
509 * - -1 in case of error, with rte_errno set to one of the following:
510 * EINVAL - one of the parameters was invalid
511 * ENOENT - memory chunk was not found
515 rte_extmem_attach(void *va_addr, size_t len);
519 * @b EXPERIMENTAL: this API may change without prior notice
521 * Detach from external memory chunk registered in another process.
523 * @note Using this API is mutually exclusive with ``rte_malloc`` family of
526 * @note This API will not perform any DMA unmapping. It is expected that user
527 * will do that themselves.
530 * Start of virtual area to unregister
532 * Length of virtual area to unregister
536 * - -1 in case of error, with rte_errno set to one of the following:
537 * EINVAL - one of the parameters was invalid
538 * ENOENT - memory chunk was not found
542 rte_extmem_detach(void *va_addr, size_t len);
545 * Dump the physical memory layout to a file.
547 * @note This function read-locks the memory hotplug subsystem, and thus cannot
548 * be used within memory-related callback functions.
551 * A pointer to a file for output
553 void rte_dump_physmem_layout(FILE *f);
556 * Get the total amount of available physical memory.
558 * @note This function read-locks the memory hotplug subsystem, and thus cannot
559 * be used within memory-related callback functions.
562 * The total amount of available physical memory in bytes.
564 uint64_t rte_eal_get_physmem_size(void);
567 * Get the number of memory channels.
570 * The number of memory channels on the system. The value is 0 if unknown
571 * or not the same on all devices.
573 unsigned rte_memory_get_nchannel(void);
576 * Get the number of memory ranks.
579 * The number of memory ranks on the system. The value is 0 if unknown or
580 * not the same on all devices.
582 unsigned rte_memory_get_nrank(void);
586 * @b EXPERIMENTAL: this API may change without prior notice
588 * Check if all currently allocated memory segments are compliant with
589 * supplied DMA address width.
592 * Address width to check against.
595 int rte_mem_check_dma_mask(uint8_t maskbits);
599 * @b EXPERIMENTAL: this API may change without prior notice
601 * Check if all currently allocated memory segments are compliant with
602 * supplied DMA address width. This function will use
603 * rte_memseg_walk_thread_unsafe instead of rte_memseg_walk implying
604 * memory_hotplug_lock will not be acquired avoiding deadlock during
605 * memory initialization.
607 * This function is just for EAL core memory internal use. Drivers should
608 * use the previous rte_mem_check_dma_mask.
611 * Address width to check against.
614 int rte_mem_check_dma_mask_thread_unsafe(uint8_t maskbits);
618 * @b EXPERIMENTAL: this API may change without prior notice
620 * Set dma mask to use once memory initialization is done. Previous functions
621 * rte_mem_check_dma_mask and rte_mem_check_dma_mask_thread_unsafe can not be
622 * used safely until memory has been initialized.
625 void rte_mem_set_dma_mask(uint8_t maskbits);
628 * Drivers based on uio will not load unless physical
629 * addresses are obtainable. It is only possible to get
630 * physical addresses when running as a privileged user.
633 * 1 if the system is able to obtain physical addresses.
634 * 0 if using DMA addresses through an IOMMU.
636 int rte_eal_using_phys_addrs(void);
640 * Enum indicating which kind of memory event has happened. Used by callbacks to
641 * distinguish between memory allocations and deallocations.
644 RTE_MEM_EVENT_ALLOC = 0, /**< Allocation event. */
645 RTE_MEM_EVENT_FREE, /**< Deallocation event. */
647 #define RTE_MEM_EVENT_CALLBACK_NAME_LEN 64
648 /**< maximum length of callback name */
651 * Function typedef used to register callbacks for memory events.
653 typedef void (*rte_mem_event_callback_t)(enum rte_mem_event event_type,
654 const void *addr, size_t len, void *arg);
657 * Function used to register callbacks for memory events.
659 * @note callbacks will happen while memory hotplug subsystem is write-locked,
660 * therefore some functions (e.g. `rte_memseg_walk()`) will cause a
661 * deadlock when called from within such callbacks.
663 * @note mem event callbacks not being supported is an expected error condition,
664 * so user code needs to handle this situation. In these cases, return
665 * value will be -1, and rte_errno will be set to ENOTSUP.
668 * Name associated with specified callback to be added to the list.
671 * Callback function pointer.
674 * Argument to pass to the callback.
677 * 0 on successful callback register
678 * -1 on unsuccessful callback register, with rte_errno value indicating
679 * reason for failure.
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.
702 rte_mem_event_callback_unregister(const char *name, void *arg);
705 #define RTE_MEM_ALLOC_VALIDATOR_NAME_LEN 64
706 /**< maximum length of alloc validator name */
708 * Function typedef used to register memory allocation validation callbacks.
710 * Returning 0 will allow allocation attempt to continue. Returning -1 will
711 * prevent allocation from succeeding.
713 typedef int (*rte_mem_alloc_validator_t)(int socket_id,
714 size_t cur_limit, size_t new_len);
717 * @brief Register validator callback for memory allocations.
719 * Callbacks registered by this function will be called right before memory
720 * allocator is about to trigger allocation of more pages from the system if
721 * said allocation will bring total memory usage above specified limit on
722 * specified socket. User will be able to cancel pending allocation if callback
725 * @note callbacks will happen while memory hotplug subsystem is write-locked,
726 * therefore some functions (e.g. `rte_memseg_walk()`) will cause a
727 * deadlock when called from within such callbacks.
729 * @note validator callbacks not being supported is an expected error condition,
730 * so user code needs to handle this situation. In these cases, return
731 * value will be -1, and rte_errno will be set to ENOTSUP.
734 * Name associated with specified callback to be added to the list.
737 * Callback function pointer.
740 * Socket ID on which to watch for allocations.
743 * Limit above which to trigger callbacks.
746 * 0 on successful callback register
747 * -1 on unsuccessful callback register, with rte_errno value indicating
748 * reason for failure.
752 rte_mem_alloc_validator_register(const char *name,
753 rte_mem_alloc_validator_t clb, int socket_id, size_t limit);
756 * @brief Unregister validator callback for memory allocations.
759 * Name associated with specified callback to be removed from the list.
762 * Socket ID on which to watch for allocations.
765 * 0 on successful callback unregister
766 * -1 on unsuccessful callback unregister, with rte_errno value indicating
767 * reason for failure.
771 rte_mem_alloc_validator_unregister(const char *name, int socket_id);
777 #endif /* _RTE_MEMORY_H_ */