]> git.proxmox.com Git - ceph.git/blob - ceph/src/seastar/dpdk/drivers/net/mlx4/mlx4_mr.c
import 15.2.0 Octopus source
[ceph.git] / ceph / src / seastar / dpdk / drivers / net / mlx4 / mlx4_mr.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2017 6WIND S.A.
3 * Copyright 2017 Mellanox Technologies, Ltd
4 */
5
6 /**
7 * @file
8 * Memory management functions for mlx4 driver.
9 */
10
11 #include <assert.h>
12 #include <errno.h>
13 #include <inttypes.h>
14 #include <stddef.h>
15 #include <stdint.h>
16 #include <string.h>
17
18 /* Verbs headers do not support -pedantic. */
19 #ifdef PEDANTIC
20 #pragma GCC diagnostic ignored "-Wpedantic"
21 #endif
22 #include <infiniband/verbs.h>
23 #ifdef PEDANTIC
24 #pragma GCC diagnostic error "-Wpedantic"
25 #endif
26
27 #include <rte_branch_prediction.h>
28 #include <rte_common.h>
29 #include <rte_errno.h>
30 #include <rte_malloc.h>
31 #include <rte_memory.h>
32 #include <rte_mempool.h>
33 #include <rte_rwlock.h>
34
35 #include "mlx4_glue.h"
36 #include "mlx4_mr.h"
37 #include "mlx4_rxtx.h"
38 #include "mlx4_utils.h"
39
40 struct mr_find_contig_memsegs_data {
41 uintptr_t addr;
42 uintptr_t start;
43 uintptr_t end;
44 const struct rte_memseg_list *msl;
45 };
46
47 struct mr_update_mp_data {
48 struct rte_eth_dev *dev;
49 struct mlx4_mr_ctrl *mr_ctrl;
50 int ret;
51 };
52
53 /**
54 * Expand B-tree table to a given size. Can't be called with holding
55 * memory_hotplug_lock or priv->mr.rwlock due to rte_realloc().
56 *
57 * @param bt
58 * Pointer to B-tree structure.
59 * @param n
60 * Number of entries for expansion.
61 *
62 * @return
63 * 0 on success, -1 on failure.
64 */
65 static int
66 mr_btree_expand(struct mlx4_mr_btree *bt, int n)
67 {
68 void *mem;
69 int ret = 0;
70
71 if (n <= bt->size)
72 return ret;
73 /*
74 * Downside of directly using rte_realloc() is that SOCKET_ID_ANY is
75 * used inside if there's no room to expand. Because this is a quite
76 * rare case and a part of very slow path, it is very acceptable.
77 * Initially cache_bh[] will be given practically enough space and once
78 * it is expanded, expansion wouldn't be needed again ever.
79 */
80 mem = rte_realloc(bt->table, n * sizeof(struct mlx4_mr_cache), 0);
81 if (mem == NULL) {
82 /* Not an error, B-tree search will be skipped. */
83 WARN("failed to expand MR B-tree (%p) table", (void *)bt);
84 ret = -1;
85 } else {
86 DEBUG("expanded MR B-tree table (size=%u)", n);
87 bt->table = mem;
88 bt->size = n;
89 }
90 return ret;
91 }
92
93 /**
94 * Look up LKey from given B-tree lookup table, store the last index and return
95 * searched LKey.
96 *
97 * @param bt
98 * Pointer to B-tree structure.
99 * @param[out] idx
100 * Pointer to index. Even on search failure, returns index where it stops
101 * searching so that index can be used when inserting a new entry.
102 * @param addr
103 * Search key.
104 *
105 * @return
106 * Searched LKey on success, UINT32_MAX on no match.
107 */
108 static uint32_t
109 mr_btree_lookup(struct mlx4_mr_btree *bt, uint16_t *idx, uintptr_t addr)
110 {
111 struct mlx4_mr_cache *lkp_tbl;
112 uint16_t n;
113 uint16_t base = 0;
114
115 assert(bt != NULL);
116 lkp_tbl = *bt->table;
117 n = bt->len;
118 /* First entry must be NULL for comparison. */
119 assert(bt->len > 0 || (lkp_tbl[0].start == 0 &&
120 lkp_tbl[0].lkey == UINT32_MAX));
121 /* Binary search. */
122 do {
123 register uint16_t delta = n >> 1;
124
125 if (addr < lkp_tbl[base + delta].start) {
126 n = delta;
127 } else {
128 base += delta;
129 n -= delta;
130 }
131 } while (n > 1);
132 assert(addr >= lkp_tbl[base].start);
133 *idx = base;
134 if (addr < lkp_tbl[base].end)
135 return lkp_tbl[base].lkey;
136 /* Not found. */
137 return UINT32_MAX;
138 }
139
140 /**
141 * Insert an entry to B-tree lookup table.
142 *
143 * @param bt
144 * Pointer to B-tree structure.
145 * @param entry
146 * Pointer to new entry to insert.
147 *
148 * @return
149 * 0 on success, -1 on failure.
150 */
151 static int
152 mr_btree_insert(struct mlx4_mr_btree *bt, struct mlx4_mr_cache *entry)
153 {
154 struct mlx4_mr_cache *lkp_tbl;
155 uint16_t idx = 0;
156 size_t shift;
157
158 assert(bt != NULL);
159 assert(bt->len <= bt->size);
160 assert(bt->len > 0);
161 lkp_tbl = *bt->table;
162 /* Find out the slot for insertion. */
163 if (mr_btree_lookup(bt, &idx, entry->start) != UINT32_MAX) {
164 DEBUG("abort insertion to B-tree(%p): already exist at"
165 " idx=%u [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
166 (void *)bt, idx, entry->start, entry->end, entry->lkey);
167 /* Already exist, return. */
168 return 0;
169 }
170 /* If table is full, return error. */
171 if (unlikely(bt->len == bt->size)) {
172 bt->overflow = 1;
173 return -1;
174 }
175 /* Insert entry. */
176 ++idx;
177 shift = (bt->len - idx) * sizeof(struct mlx4_mr_cache);
178 if (shift)
179 memmove(&lkp_tbl[idx + 1], &lkp_tbl[idx], shift);
180 lkp_tbl[idx] = *entry;
181 bt->len++;
182 DEBUG("inserted B-tree(%p)[%u],"
183 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
184 (void *)bt, idx, entry->start, entry->end, entry->lkey);
185 return 0;
186 }
187
188 /**
189 * Initialize B-tree and allocate memory for lookup table.
190 *
191 * @param bt
192 * Pointer to B-tree structure.
193 * @param n
194 * Number of entries to allocate.
195 * @param socket
196 * NUMA socket on which memory must be allocated.
197 *
198 * @return
199 * 0 on success, a negative errno value otherwise and rte_errno is set.
200 */
201 int
202 mlx4_mr_btree_init(struct mlx4_mr_btree *bt, int n, int socket)
203 {
204 if (bt == NULL) {
205 rte_errno = EINVAL;
206 return -rte_errno;
207 }
208 memset(bt, 0, sizeof(*bt));
209 bt->table = rte_calloc_socket("B-tree table",
210 n, sizeof(struct mlx4_mr_cache),
211 0, socket);
212 if (bt->table == NULL) {
213 rte_errno = ENOMEM;
214 ERROR("failed to allocate memory for btree cache on socket %d",
215 socket);
216 return -rte_errno;
217 }
218 bt->size = n;
219 /* First entry must be NULL for binary search. */
220 (*bt->table)[bt->len++] = (struct mlx4_mr_cache) {
221 .lkey = UINT32_MAX,
222 };
223 DEBUG("initialized B-tree %p with table %p",
224 (void *)bt, (void *)bt->table);
225 return 0;
226 }
227
228 /**
229 * Free B-tree resources.
230 *
231 * @param bt
232 * Pointer to B-tree structure.
233 */
234 void
235 mlx4_mr_btree_free(struct mlx4_mr_btree *bt)
236 {
237 if (bt == NULL)
238 return;
239 DEBUG("freeing B-tree %p with table %p", (void *)bt, (void *)bt->table);
240 rte_free(bt->table);
241 memset(bt, 0, sizeof(*bt));
242 }
243
244 #ifndef NDEBUG
245 /**
246 * Dump all the entries in a B-tree
247 *
248 * @param bt
249 * Pointer to B-tree structure.
250 */
251 void
252 mlx4_mr_btree_dump(struct mlx4_mr_btree *bt)
253 {
254 int idx;
255 struct mlx4_mr_cache *lkp_tbl;
256
257 if (bt == NULL)
258 return;
259 lkp_tbl = *bt->table;
260 for (idx = 0; idx < bt->len; ++idx) {
261 struct mlx4_mr_cache *entry = &lkp_tbl[idx];
262
263 DEBUG("B-tree(%p)[%u],"
264 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x",
265 (void *)bt, idx, entry->start, entry->end, entry->lkey);
266 }
267 }
268 #endif
269
270 /**
271 * Find virtually contiguous memory chunk in a given MR.
272 *
273 * @param dev
274 * Pointer to MR structure.
275 * @param[out] entry
276 * Pointer to returning MR cache entry. If not found, this will not be
277 * updated.
278 * @param start_idx
279 * Start index of the memseg bitmap.
280 *
281 * @return
282 * Next index to go on lookup.
283 */
284 static int
285 mr_find_next_chunk(struct mlx4_mr *mr, struct mlx4_mr_cache *entry,
286 int base_idx)
287 {
288 uintptr_t start = 0;
289 uintptr_t end = 0;
290 uint32_t idx = 0;
291
292 /* MR for external memory doesn't have memseg list. */
293 if (mr->msl == NULL) {
294 struct ibv_mr *ibv_mr = mr->ibv_mr;
295
296 assert(mr->ms_bmp_n == 1);
297 assert(mr->ms_n == 1);
298 assert(base_idx == 0);
299 /*
300 * Can't search it from memseg list but get it directly from
301 * verbs MR as there's only one chunk.
302 */
303 entry->start = (uintptr_t)ibv_mr->addr;
304 entry->end = (uintptr_t)ibv_mr->addr + mr->ibv_mr->length;
305 entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey);
306 /* Returning 1 ends iteration. */
307 return 1;
308 }
309 for (idx = base_idx; idx < mr->ms_bmp_n; ++idx) {
310 if (rte_bitmap_get(mr->ms_bmp, idx)) {
311 const struct rte_memseg_list *msl;
312 const struct rte_memseg *ms;
313
314 msl = mr->msl;
315 ms = rte_fbarray_get(&msl->memseg_arr,
316 mr->ms_base_idx + idx);
317 assert(msl->page_sz == ms->hugepage_sz);
318 if (!start)
319 start = ms->addr_64;
320 end = ms->addr_64 + ms->hugepage_sz;
321 } else if (start) {
322 /* Passed the end of a fragment. */
323 break;
324 }
325 }
326 if (start) {
327 /* Found one chunk. */
328 entry->start = start;
329 entry->end = end;
330 entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey);
331 }
332 return idx;
333 }
334
335 /**
336 * Insert a MR to the global B-tree cache. It may fail due to low-on-memory.
337 * Then, this entry will have to be searched by mr_lookup_dev_list() in
338 * mlx4_mr_create() on miss.
339 *
340 * @param dev
341 * Pointer to Ethernet device.
342 * @param mr
343 * Pointer to MR to insert.
344 *
345 * @return
346 * 0 on success, -1 on failure.
347 */
348 static int
349 mr_insert_dev_cache(struct rte_eth_dev *dev, struct mlx4_mr *mr)
350 {
351 struct mlx4_priv *priv = dev->data->dev_private;
352 unsigned int n;
353
354 DEBUG("port %u inserting MR(%p) to global cache",
355 dev->data->port_id, (void *)mr);
356 for (n = 0; n < mr->ms_bmp_n; ) {
357 struct mlx4_mr_cache entry;
358
359 memset(&entry, 0, sizeof(entry));
360 /* Find a contiguous chunk and advance the index. */
361 n = mr_find_next_chunk(mr, &entry, n);
362 if (!entry.end)
363 break;
364 if (mr_btree_insert(&priv->mr.cache, &entry) < 0) {
365 /*
366 * Overflowed, but the global table cannot be expanded
367 * because of deadlock.
368 */
369 return -1;
370 }
371 }
372 return 0;
373 }
374
375 /**
376 * Look up address in the original global MR list.
377 *
378 * @param dev
379 * Pointer to Ethernet device.
380 * @param[out] entry
381 * Pointer to returning MR cache entry. If no match, this will not be updated.
382 * @param addr
383 * Search key.
384 *
385 * @return
386 * Found MR on match, NULL otherwise.
387 */
388 static struct mlx4_mr *
389 mr_lookup_dev_list(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
390 uintptr_t addr)
391 {
392 struct mlx4_priv *priv = dev->data->dev_private;
393 struct mlx4_mr *mr;
394
395 /* Iterate all the existing MRs. */
396 LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
397 unsigned int n;
398
399 if (mr->ms_n == 0)
400 continue;
401 for (n = 0; n < mr->ms_bmp_n; ) {
402 struct mlx4_mr_cache ret;
403
404 memset(&ret, 0, sizeof(ret));
405 n = mr_find_next_chunk(mr, &ret, n);
406 if (addr >= ret.start && addr < ret.end) {
407 /* Found. */
408 *entry = ret;
409 return mr;
410 }
411 }
412 }
413 return NULL;
414 }
415
416 /**
417 * Look up address on device.
418 *
419 * @param dev
420 * Pointer to Ethernet device.
421 * @param[out] entry
422 * Pointer to returning MR cache entry. If no match, this will not be updated.
423 * @param addr
424 * Search key.
425 *
426 * @return
427 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
428 */
429 static uint32_t
430 mr_lookup_dev(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
431 uintptr_t addr)
432 {
433 struct mlx4_priv *priv = dev->data->dev_private;
434 uint16_t idx;
435 uint32_t lkey = UINT32_MAX;
436 struct mlx4_mr *mr;
437
438 /*
439 * If the global cache has overflowed since it failed to expand the
440 * B-tree table, it can't have all the existing MRs. Then, the address
441 * has to be searched by traversing the original MR list instead, which
442 * is very slow path. Otherwise, the global cache is all inclusive.
443 */
444 if (!unlikely(priv->mr.cache.overflow)) {
445 lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
446 if (lkey != UINT32_MAX)
447 *entry = (*priv->mr.cache.table)[idx];
448 } else {
449 /* Falling back to the slowest path. */
450 mr = mr_lookup_dev_list(dev, entry, addr);
451 if (mr != NULL)
452 lkey = entry->lkey;
453 }
454 assert(lkey == UINT32_MAX || (addr >= entry->start &&
455 addr < entry->end));
456 return lkey;
457 }
458
459 /**
460 * Free MR resources. MR lock must not be held to avoid a deadlock. rte_free()
461 * can raise memory free event and the callback function will spin on the lock.
462 *
463 * @param mr
464 * Pointer to MR to free.
465 */
466 static void
467 mr_free(struct mlx4_mr *mr)
468 {
469 if (mr == NULL)
470 return;
471 DEBUG("freeing MR(%p):", (void *)mr);
472 if (mr->ibv_mr != NULL)
473 claim_zero(mlx4_glue->dereg_mr(mr->ibv_mr));
474 if (mr->ms_bmp != NULL)
475 rte_bitmap_free(mr->ms_bmp);
476 rte_free(mr);
477 }
478
479 /**
480 * Release resources of detached MR having no online entry.
481 *
482 * @param dev
483 * Pointer to Ethernet device.
484 */
485 static void
486 mlx4_mr_garbage_collect(struct rte_eth_dev *dev)
487 {
488 struct mlx4_priv *priv = dev->data->dev_private;
489 struct mlx4_mr *mr_next;
490 struct mlx4_mr_list free_list = LIST_HEAD_INITIALIZER(free_list);
491
492 /* Must be called from the primary process. */
493 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
494 /*
495 * MR can't be freed with holding the lock because rte_free() could call
496 * memory free callback function. This will be a deadlock situation.
497 */
498 rte_rwlock_write_lock(&priv->mr.rwlock);
499 /* Detach the whole free list and release it after unlocking. */
500 free_list = priv->mr.mr_free_list;
501 LIST_INIT(&priv->mr.mr_free_list);
502 rte_rwlock_write_unlock(&priv->mr.rwlock);
503 /* Release resources. */
504 mr_next = LIST_FIRST(&free_list);
505 while (mr_next != NULL) {
506 struct mlx4_mr *mr = mr_next;
507
508 mr_next = LIST_NEXT(mr, mr);
509 mr_free(mr);
510 }
511 }
512
513 /* Called during rte_memseg_contig_walk() by mlx4_mr_create(). */
514 static int
515 mr_find_contig_memsegs_cb(const struct rte_memseg_list *msl,
516 const struct rte_memseg *ms, size_t len, void *arg)
517 {
518 struct mr_find_contig_memsegs_data *data = arg;
519
520 if (data->addr < ms->addr_64 || data->addr >= ms->addr_64 + len)
521 return 0;
522 /* Found, save it and stop walking. */
523 data->start = ms->addr_64;
524 data->end = ms->addr_64 + len;
525 data->msl = msl;
526 return 1;
527 }
528
529 /**
530 * Create a new global Memory Region (MR) for a missing virtual address.
531 * This API should be called on a secondary process, then a request is sent to
532 * the primary process in order to create a MR for the address. As the global MR
533 * list is on the shared memory, following LKey lookup should succeed unless the
534 * request fails.
535 *
536 * @param dev
537 * Pointer to Ethernet device.
538 * @param[out] entry
539 * Pointer to returning MR cache entry, found in the global cache or newly
540 * created. If failed to create one, this will not be updated.
541 * @param addr
542 * Target virtual address to register.
543 *
544 * @return
545 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
546 */
547 static uint32_t
548 mlx4_mr_create_secondary(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
549 uintptr_t addr)
550 {
551 struct mlx4_priv *priv = dev->data->dev_private;
552 int ret;
553
554 DEBUG("port %u requesting MR creation for address (%p)",
555 dev->data->port_id, (void *)addr);
556 ret = mlx4_mp_req_mr_create(dev, addr);
557 if (ret) {
558 DEBUG("port %u fail to request MR creation for address (%p)",
559 dev->data->port_id, (void *)addr);
560 return UINT32_MAX;
561 }
562 rte_rwlock_read_lock(&priv->mr.rwlock);
563 /* Fill in output data. */
564 mr_lookup_dev(dev, entry, addr);
565 /* Lookup can't fail. */
566 assert(entry->lkey != UINT32_MAX);
567 rte_rwlock_read_unlock(&priv->mr.rwlock);
568 DEBUG("port %u MR CREATED by primary process for %p:\n"
569 " [0x%" PRIxPTR ", 0x%" PRIxPTR "), lkey=0x%x",
570 dev->data->port_id, (void *)addr,
571 entry->start, entry->end, entry->lkey);
572 return entry->lkey;
573 }
574
575 /**
576 * Create a new global Memory Region (MR) for a missing virtual address.
577 * Register entire virtually contiguous memory chunk around the address.
578 * This must be called from the primary process.
579 *
580 * @param dev
581 * Pointer to Ethernet device.
582 * @param[out] entry
583 * Pointer to returning MR cache entry, found in the global cache or newly
584 * created. If failed to create one, this will not be updated.
585 * @param addr
586 * Target virtual address to register.
587 *
588 * @return
589 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
590 */
591 uint32_t
592 mlx4_mr_create_primary(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
593 uintptr_t addr)
594 {
595 struct mlx4_priv *priv = dev->data->dev_private;
596 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
597 const struct rte_memseg_list *msl;
598 const struct rte_memseg *ms;
599 struct mlx4_mr *mr = NULL;
600 size_t len;
601 uint32_t ms_n;
602 uint32_t bmp_size;
603 void *bmp_mem;
604 int ms_idx_shift = -1;
605 unsigned int n;
606 struct mr_find_contig_memsegs_data data = {
607 .addr = addr,
608 };
609 struct mr_find_contig_memsegs_data data_re;
610
611 DEBUG("port %u creating a MR using address (%p)",
612 dev->data->port_id, (void *)addr);
613 /*
614 * Release detached MRs if any. This can't be called with holding either
615 * memory_hotplug_lock or priv->mr.rwlock. MRs on the free list have
616 * been detached by the memory free event but it couldn't be released
617 * inside the callback due to deadlock. As a result, releasing resources
618 * is quite opportunistic.
619 */
620 mlx4_mr_garbage_collect(dev);
621 /*
622 * If enabled, find out a contiguous virtual address chunk in use, to
623 * which the given address belongs, in order to register maximum range.
624 * In the best case where mempools are not dynamically recreated and
625 * '--socket-mem' is specified as an EAL option, it is very likely to
626 * have only one MR(LKey) per a socket and per a hugepage-size even
627 * though the system memory is highly fragmented. As the whole memory
628 * chunk will be pinned by kernel, it can't be reused unless entire
629 * chunk is freed from EAL.
630 *
631 * If disabled, just register one memseg (page). Then, memory
632 * consumption will be minimized but it may drop performance if there
633 * are many MRs to lookup on the datapath.
634 */
635 if (!priv->mr_ext_memseg_en) {
636 data.msl = rte_mem_virt2memseg_list((void *)addr);
637 data.start = RTE_ALIGN_FLOOR(addr, data.msl->page_sz);
638 data.end = data.start + data.msl->page_sz;
639 } else if (!rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data)) {
640 WARN("port %u unable to find virtually contiguous"
641 " chunk for address (%p)."
642 " rte_memseg_contig_walk() failed.",
643 dev->data->port_id, (void *)addr);
644 rte_errno = ENXIO;
645 goto err_nolock;
646 }
647 alloc_resources:
648 /* Addresses must be page-aligned. */
649 assert(rte_is_aligned((void *)data.start, data.msl->page_sz));
650 assert(rte_is_aligned((void *)data.end, data.msl->page_sz));
651 msl = data.msl;
652 ms = rte_mem_virt2memseg((void *)data.start, msl);
653 len = data.end - data.start;
654 assert(msl->page_sz == ms->hugepage_sz);
655 /* Number of memsegs in the range. */
656 ms_n = len / msl->page_sz;
657 DEBUG("port %u extending %p to [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
658 " page_sz=0x%" PRIx64 ", ms_n=%u",
659 dev->data->port_id, (void *)addr,
660 data.start, data.end, msl->page_sz, ms_n);
661 /* Size of memory for bitmap. */
662 bmp_size = rte_bitmap_get_memory_footprint(ms_n);
663 mr = rte_zmalloc_socket(NULL,
664 RTE_ALIGN_CEIL(sizeof(*mr),
665 RTE_CACHE_LINE_SIZE) +
666 bmp_size,
667 RTE_CACHE_LINE_SIZE, msl->socket_id);
668 if (mr == NULL) {
669 WARN("port %u unable to allocate memory for a new MR of"
670 " address (%p).",
671 dev->data->port_id, (void *)addr);
672 rte_errno = ENOMEM;
673 goto err_nolock;
674 }
675 mr->msl = msl;
676 /*
677 * Save the index of the first memseg and initialize memseg bitmap. To
678 * see if a memseg of ms_idx in the memseg-list is still valid, check:
679 * rte_bitmap_get(mr->bmp, ms_idx - mr->ms_base_idx)
680 */
681 mr->ms_base_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
682 bmp_mem = RTE_PTR_ALIGN_CEIL(mr + 1, RTE_CACHE_LINE_SIZE);
683 mr->ms_bmp = rte_bitmap_init(ms_n, bmp_mem, bmp_size);
684 if (mr->ms_bmp == NULL) {
685 WARN("port %u unable to initialize bitmap for a new MR of"
686 " address (%p).",
687 dev->data->port_id, (void *)addr);
688 rte_errno = EINVAL;
689 goto err_nolock;
690 }
691 /*
692 * Should recheck whether the extended contiguous chunk is still valid.
693 * Because memory_hotplug_lock can't be held if there's any memory
694 * related calls in a critical path, resource allocation above can't be
695 * locked. If the memory has been changed at this point, try again with
696 * just single page. If not, go on with the big chunk atomically from
697 * here.
698 */
699 rte_rwlock_read_lock(&mcfg->memory_hotplug_lock);
700 data_re = data;
701 if (len > msl->page_sz &&
702 !rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data_re)) {
703 WARN("port %u unable to find virtually contiguous"
704 " chunk for address (%p)."
705 " rte_memseg_contig_walk() failed.",
706 dev->data->port_id, (void *)addr);
707 rte_errno = ENXIO;
708 goto err_memlock;
709 }
710 if (data.start != data_re.start || data.end != data_re.end) {
711 /*
712 * The extended contiguous chunk has been changed. Try again
713 * with single memseg instead.
714 */
715 data.start = RTE_ALIGN_FLOOR(addr, msl->page_sz);
716 data.end = data.start + msl->page_sz;
717 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
718 mr_free(mr);
719 goto alloc_resources;
720 }
721 assert(data.msl == data_re.msl);
722 rte_rwlock_write_lock(&priv->mr.rwlock);
723 /*
724 * Check the address is really missing. If other thread already created
725 * one or it is not found due to overflow, abort and return.
726 */
727 if (mr_lookup_dev(dev, entry, addr) != UINT32_MAX) {
728 /*
729 * Insert to the global cache table. It may fail due to
730 * low-on-memory. Then, this entry will have to be searched
731 * here again.
732 */
733 mr_btree_insert(&priv->mr.cache, entry);
734 DEBUG("port %u found MR for %p on final lookup, abort",
735 dev->data->port_id, (void *)addr);
736 rte_rwlock_write_unlock(&priv->mr.rwlock);
737 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
738 /*
739 * Must be unlocked before calling rte_free() because
740 * mlx4_mr_mem_event_free_cb() can be called inside.
741 */
742 mr_free(mr);
743 return entry->lkey;
744 }
745 /*
746 * Trim start and end addresses for verbs MR. Set bits for registering
747 * memsegs but exclude already registered ones. Bitmap can be
748 * fragmented.
749 */
750 for (n = 0; n < ms_n; ++n) {
751 uintptr_t start;
752 struct mlx4_mr_cache ret;
753
754 memset(&ret, 0, sizeof(ret));
755 start = data_re.start + n * msl->page_sz;
756 /* Exclude memsegs already registered by other MRs. */
757 if (mr_lookup_dev(dev, &ret, start) == UINT32_MAX) {
758 /*
759 * Start from the first unregistered memseg in the
760 * extended range.
761 */
762 if (ms_idx_shift == -1) {
763 mr->ms_base_idx += n;
764 data.start = start;
765 ms_idx_shift = n;
766 }
767 data.end = start + msl->page_sz;
768 rte_bitmap_set(mr->ms_bmp, n - ms_idx_shift);
769 ++mr->ms_n;
770 }
771 }
772 len = data.end - data.start;
773 mr->ms_bmp_n = len / msl->page_sz;
774 assert(ms_idx_shift + mr->ms_bmp_n <= ms_n);
775 /*
776 * Finally create a verbs MR for the memory chunk. ibv_reg_mr() can be
777 * called with holding the memory lock because it doesn't use
778 * mlx4_alloc_buf_extern() which eventually calls rte_malloc_socket()
779 * through mlx4_alloc_verbs_buf().
780 */
781 mr->ibv_mr = mlx4_glue->reg_mr(priv->pd, (void *)data.start, len,
782 IBV_ACCESS_LOCAL_WRITE);
783 if (mr->ibv_mr == NULL) {
784 WARN("port %u fail to create a verbs MR for address (%p)",
785 dev->data->port_id, (void *)addr);
786 rte_errno = EINVAL;
787 goto err_mrlock;
788 }
789 assert((uintptr_t)mr->ibv_mr->addr == data.start);
790 assert(mr->ibv_mr->length == len);
791 LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
792 DEBUG("port %u MR CREATED (%p) for %p:\n"
793 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
794 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
795 dev->data->port_id, (void *)mr, (void *)addr,
796 data.start, data.end, rte_cpu_to_be_32(mr->ibv_mr->lkey),
797 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
798 /* Insert to the global cache table. */
799 mr_insert_dev_cache(dev, mr);
800 /* Fill in output data. */
801 mr_lookup_dev(dev, entry, addr);
802 /* Lookup can't fail. */
803 assert(entry->lkey != UINT32_MAX);
804 rte_rwlock_write_unlock(&priv->mr.rwlock);
805 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
806 return entry->lkey;
807 err_mrlock:
808 rte_rwlock_write_unlock(&priv->mr.rwlock);
809 err_memlock:
810 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
811 err_nolock:
812 /*
813 * In case of error, as this can be called in a datapath, a warning
814 * message per an error is preferable instead. Must be unlocked before
815 * calling rte_free() because mlx4_mr_mem_event_free_cb() can be called
816 * inside.
817 */
818 mr_free(mr);
819 return UINT32_MAX;
820 }
821
822 /**
823 * Create a new global Memory Region (MR) for a missing virtual address.
824 * This can be called from primary and secondary process.
825 *
826 * @param dev
827 * Pointer to Ethernet device.
828 * @param[out] entry
829 * Pointer to returning MR cache entry, found in the global cache or newly
830 * created. If failed to create one, this will not be updated.
831 * @param addr
832 * Target virtual address to register.
833 *
834 * @return
835 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
836 */
837 static uint32_t
838 mlx4_mr_create(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry,
839 uintptr_t addr)
840 {
841 uint32_t ret = 0;
842
843 switch (rte_eal_process_type()) {
844 case RTE_PROC_PRIMARY:
845 ret = mlx4_mr_create_primary(dev, entry, addr);
846 break;
847 case RTE_PROC_SECONDARY:
848 ret = mlx4_mr_create_secondary(dev, entry, addr);
849 break;
850 default:
851 break;
852 }
853 return ret;
854 }
855
856 /**
857 * Rebuild the global B-tree cache of device from the original MR list.
858 *
859 * @param dev
860 * Pointer to Ethernet device.
861 */
862 static void
863 mr_rebuild_dev_cache(struct rte_eth_dev *dev)
864 {
865 struct mlx4_priv *priv = dev->data->dev_private;
866 struct mlx4_mr *mr;
867
868 DEBUG("port %u rebuild dev cache[]", dev->data->port_id);
869 /* Flush cache to rebuild. */
870 priv->mr.cache.len = 1;
871 priv->mr.cache.overflow = 0;
872 /* Iterate all the existing MRs. */
873 LIST_FOREACH(mr, &priv->mr.mr_list, mr)
874 if (mr_insert_dev_cache(dev, mr) < 0)
875 return;
876 }
877
878 /**
879 * Callback for memory free event. Iterate freed memsegs and check whether it
880 * belongs to an existing MR. If found, clear the bit from bitmap of MR. As a
881 * result, the MR would be fragmented. If it becomes empty, the MR will be freed
882 * later by mlx4_mr_garbage_collect().
883 *
884 * The global cache must be rebuilt if there's any change and this event has to
885 * be propagated to dataplane threads to flush the local caches.
886 *
887 * @param dev
888 * Pointer to Ethernet device.
889 * @param addr
890 * Address of freed memory.
891 * @param len
892 * Size of freed memory.
893 */
894 static void
895 mlx4_mr_mem_event_free_cb(struct rte_eth_dev *dev, const void *addr, size_t len)
896 {
897 struct mlx4_priv *priv = dev->data->dev_private;
898 const struct rte_memseg_list *msl;
899 struct mlx4_mr *mr;
900 int ms_n;
901 int i;
902 int rebuild = 0;
903
904 DEBUG("port %u free callback: addr=%p, len=%zu",
905 dev->data->port_id, addr, len);
906 msl = rte_mem_virt2memseg_list(addr);
907 /* addr and len must be page-aligned. */
908 assert((uintptr_t)addr == RTE_ALIGN((uintptr_t)addr, msl->page_sz));
909 assert(len == RTE_ALIGN(len, msl->page_sz));
910 ms_n = len / msl->page_sz;
911 rte_rwlock_write_lock(&priv->mr.rwlock);
912 /* Clear bits of freed memsegs from MR. */
913 for (i = 0; i < ms_n; ++i) {
914 const struct rte_memseg *ms;
915 struct mlx4_mr_cache entry;
916 uintptr_t start;
917 int ms_idx;
918 uint32_t pos;
919
920 /* Find MR having this memseg. */
921 start = (uintptr_t)addr + i * msl->page_sz;
922 mr = mr_lookup_dev_list(dev, &entry, start);
923 if (mr == NULL)
924 continue;
925 assert(mr->msl); /* Can't be external memory. */
926 ms = rte_mem_virt2memseg((void *)start, msl);
927 assert(ms != NULL);
928 assert(msl->page_sz == ms->hugepage_sz);
929 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
930 pos = ms_idx - mr->ms_base_idx;
931 assert(rte_bitmap_get(mr->ms_bmp, pos));
932 assert(pos < mr->ms_bmp_n);
933 DEBUG("port %u MR(%p): clear bitmap[%u] for addr %p",
934 dev->data->port_id, (void *)mr, pos, (void *)start);
935 rte_bitmap_clear(mr->ms_bmp, pos);
936 if (--mr->ms_n == 0) {
937 LIST_REMOVE(mr, mr);
938 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
939 DEBUG("port %u remove MR(%p) from list",
940 dev->data->port_id, (void *)mr);
941 }
942 /*
943 * MR is fragmented or will be freed. the global cache must be
944 * rebuilt.
945 */
946 rebuild = 1;
947 }
948 if (rebuild) {
949 mr_rebuild_dev_cache(dev);
950 /*
951 * Flush local caches by propagating invalidation across cores.
952 * rte_smp_wmb() is enough to synchronize this event. If one of
953 * freed memsegs is seen by other core, that means the memseg
954 * has been allocated by allocator, which will come after this
955 * free call. Therefore, this store instruction (incrementing
956 * generation below) will be guaranteed to be seen by other core
957 * before the core sees the newly allocated memory.
958 */
959 ++priv->mr.dev_gen;
960 DEBUG("broadcasting local cache flush, gen=%d",
961 priv->mr.dev_gen);
962 rte_smp_wmb();
963 }
964 rte_rwlock_write_unlock(&priv->mr.rwlock);
965 #ifndef NDEBUG
966 if (rebuild)
967 mlx4_mr_dump_dev(dev);
968 #endif
969 }
970
971 /**
972 * Callback for memory event.
973 *
974 * @param event_type
975 * Memory event type.
976 * @param addr
977 * Address of memory.
978 * @param len
979 * Size of memory.
980 */
981 void
982 mlx4_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
983 size_t len, void *arg __rte_unused)
984 {
985 struct mlx4_priv *priv;
986 struct mlx4_dev_list *dev_list = &mlx4_shared_data->mem_event_cb_list;
987
988 /* Must be called from the primary process. */
989 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
990 switch (event_type) {
991 case RTE_MEM_EVENT_FREE:
992 rte_rwlock_read_lock(&mlx4_shared_data->mem_event_rwlock);
993 /* Iterate all the existing mlx4 devices. */
994 LIST_FOREACH(priv, dev_list, mem_event_cb)
995 mlx4_mr_mem_event_free_cb(ETH_DEV(priv), addr, len);
996 rte_rwlock_read_unlock(&mlx4_shared_data->mem_event_rwlock);
997 break;
998 case RTE_MEM_EVENT_ALLOC:
999 default:
1000 break;
1001 }
1002 }
1003
1004 /**
1005 * Look up address in the global MR cache table. If not found, create a new MR.
1006 * Insert the found/created entry to local bottom-half cache table.
1007 *
1008 * @param dev
1009 * Pointer to Ethernet device.
1010 * @param mr_ctrl
1011 * Pointer to per-queue MR control structure.
1012 * @param[out] entry
1013 * Pointer to returning MR cache entry, found in the global cache or newly
1014 * created. If failed to create one, this is not written.
1015 * @param addr
1016 * Search key.
1017 *
1018 * @return
1019 * Searched LKey on success, UINT32_MAX on no match.
1020 */
1021 static uint32_t
1022 mlx4_mr_lookup_dev(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1023 struct mlx4_mr_cache *entry, uintptr_t addr)
1024 {
1025 struct mlx4_priv *priv = dev->data->dev_private;
1026 struct mlx4_mr_btree *bt = &mr_ctrl->cache_bh;
1027 uint16_t idx;
1028 uint32_t lkey;
1029
1030 /* If local cache table is full, try to double it. */
1031 if (unlikely(bt->len == bt->size))
1032 mr_btree_expand(bt, bt->size << 1);
1033 /* Look up in the global cache. */
1034 rte_rwlock_read_lock(&priv->mr.rwlock);
1035 lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
1036 if (lkey != UINT32_MAX) {
1037 /* Found. */
1038 *entry = (*priv->mr.cache.table)[idx];
1039 rte_rwlock_read_unlock(&priv->mr.rwlock);
1040 /*
1041 * Update local cache. Even if it fails, return the found entry
1042 * to update top-half cache. Next time, this entry will be found
1043 * in the global cache.
1044 */
1045 mr_btree_insert(bt, entry);
1046 return lkey;
1047 }
1048 rte_rwlock_read_unlock(&priv->mr.rwlock);
1049 /* First time to see the address? Create a new MR. */
1050 lkey = mlx4_mr_create(dev, entry, addr);
1051 /*
1052 * Update the local cache if successfully created a new global MR. Even
1053 * if failed to create one, there's no action to take in this datapath
1054 * code. As returning LKey is invalid, this will eventually make HW
1055 * fail.
1056 */
1057 if (lkey != UINT32_MAX)
1058 mr_btree_insert(bt, entry);
1059 return lkey;
1060 }
1061
1062 /**
1063 * Bottom-half of LKey search on datapath. Firstly search in cache_bh[] and if
1064 * misses, search in the global MR cache table and update the new entry to
1065 * per-queue local caches.
1066 *
1067 * @param dev
1068 * Pointer to Ethernet device.
1069 * @param mr_ctrl
1070 * Pointer to per-queue MR control structure.
1071 * @param addr
1072 * Search key.
1073 *
1074 * @return
1075 * Searched LKey on success, UINT32_MAX on no match.
1076 */
1077 static uint32_t
1078 mlx4_mr_addr2mr_bh(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1079 uintptr_t addr)
1080 {
1081 uint32_t lkey;
1082 uint16_t bh_idx = 0;
1083 /* Victim in top-half cache to replace with new entry. */
1084 struct mlx4_mr_cache *repl = &mr_ctrl->cache[mr_ctrl->head];
1085
1086 /* Binary-search MR translation table. */
1087 lkey = mr_btree_lookup(&mr_ctrl->cache_bh, &bh_idx, addr);
1088 /* Update top-half cache. */
1089 if (likely(lkey != UINT32_MAX)) {
1090 *repl = (*mr_ctrl->cache_bh.table)[bh_idx];
1091 } else {
1092 /*
1093 * If missed in local lookup table, search in the global cache
1094 * and local cache_bh[] will be updated inside if possible.
1095 * Top-half cache entry will also be updated.
1096 */
1097 lkey = mlx4_mr_lookup_dev(dev, mr_ctrl, repl, addr);
1098 if (unlikely(lkey == UINT32_MAX))
1099 return UINT32_MAX;
1100 }
1101 /* Update the most recently used entry. */
1102 mr_ctrl->mru = mr_ctrl->head;
1103 /* Point to the next victim, the oldest. */
1104 mr_ctrl->head = (mr_ctrl->head + 1) % MLX4_MR_CACHE_N;
1105 return lkey;
1106 }
1107
1108 /**
1109 * Bottom-half of LKey search on Rx.
1110 *
1111 * @param rxq
1112 * Pointer to Rx queue structure.
1113 * @param addr
1114 * Search key.
1115 *
1116 * @return
1117 * Searched LKey on success, UINT32_MAX on no match.
1118 */
1119 uint32_t
1120 mlx4_rx_addr2mr_bh(struct rxq *rxq, uintptr_t addr)
1121 {
1122 struct mlx4_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
1123 struct mlx4_priv *priv = rxq->priv;
1124
1125 return mlx4_mr_addr2mr_bh(ETH_DEV(priv), mr_ctrl, addr);
1126 }
1127
1128 /**
1129 * Bottom-half of LKey search on Tx.
1130 *
1131 * @param txq
1132 * Pointer to Tx queue structure.
1133 * @param addr
1134 * Search key.
1135 *
1136 * @return
1137 * Searched LKey on success, UINT32_MAX on no match.
1138 */
1139 static uint32_t
1140 mlx4_tx_addr2mr_bh(struct txq *txq, uintptr_t addr)
1141 {
1142 struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1143 struct mlx4_priv *priv = txq->priv;
1144
1145 return mlx4_mr_addr2mr_bh(ETH_DEV(priv), mr_ctrl, addr);
1146 }
1147
1148 /**
1149 * Bottom-half of LKey search on Tx. If it can't be searched in the memseg
1150 * list, register the mempool of the mbuf as externally allocated memory.
1151 *
1152 * @param txq
1153 * Pointer to Tx queue structure.
1154 * @param mb
1155 * Pointer to mbuf.
1156 *
1157 * @return
1158 * Searched LKey on success, UINT32_MAX on no match.
1159 */
1160 uint32_t
1161 mlx4_tx_mb2mr_bh(struct txq *txq, struct rte_mbuf *mb)
1162 {
1163 uintptr_t addr = (uintptr_t)mb->buf_addr;
1164 uint32_t lkey;
1165
1166 lkey = mlx4_tx_addr2mr_bh(txq, addr);
1167 if (lkey == UINT32_MAX && rte_errno == ENXIO) {
1168 /* Mempool may have externally allocated memory. */
1169 return mlx4_tx_update_ext_mp(txq, addr, mlx4_mb2mp(mb));
1170 }
1171 return lkey;
1172 }
1173
1174 /**
1175 * Flush all of the local cache entries.
1176 *
1177 * @param mr_ctrl
1178 * Pointer to per-queue MR control structure.
1179 */
1180 void
1181 mlx4_mr_flush_local_cache(struct mlx4_mr_ctrl *mr_ctrl)
1182 {
1183 /* Reset the most-recently-used index. */
1184 mr_ctrl->mru = 0;
1185 /* Reset the linear search array. */
1186 mr_ctrl->head = 0;
1187 memset(mr_ctrl->cache, 0, sizeof(mr_ctrl->cache));
1188 /* Reset the B-tree table. */
1189 mr_ctrl->cache_bh.len = 1;
1190 mr_ctrl->cache_bh.overflow = 0;
1191 /* Update the generation number. */
1192 mr_ctrl->cur_gen = *mr_ctrl->dev_gen_ptr;
1193 DEBUG("mr_ctrl(%p): flushed, cur_gen=%d",
1194 (void *)mr_ctrl, mr_ctrl->cur_gen);
1195 }
1196
1197 /**
1198 * Called during rte_mempool_mem_iter() by mlx4_mr_update_ext_mp().
1199 *
1200 * Externally allocated chunk is registered and a MR is created for the chunk.
1201 * The MR object is added to the global list. If memseg list of a MR object
1202 * (mr->msl) is null, the MR object can be regarded as externally allocated
1203 * memory.
1204 *
1205 * Once external memory is registered, it should be static. If the memory is
1206 * freed and the virtual address range has different physical memory mapped
1207 * again, it may cause crash on device due to the wrong translation entry. PMD
1208 * can't track the free event of the external memory for now.
1209 */
1210 static void
1211 mlx4_mr_update_ext_mp_cb(struct rte_mempool *mp, void *opaque,
1212 struct rte_mempool_memhdr *memhdr,
1213 unsigned mem_idx __rte_unused)
1214 {
1215 struct mr_update_mp_data *data = opaque;
1216 struct rte_eth_dev *dev = data->dev;
1217 struct mlx4_priv *priv = dev->data->dev_private;
1218 struct mlx4_mr_ctrl *mr_ctrl = data->mr_ctrl;
1219 struct mlx4_mr *mr = NULL;
1220 uintptr_t addr = (uintptr_t)memhdr->addr;
1221 size_t len = memhdr->len;
1222 struct mlx4_mr_cache entry;
1223 uint32_t lkey;
1224
1225 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
1226 /* If already registered, it should return. */
1227 rte_rwlock_read_lock(&priv->mr.rwlock);
1228 lkey = mr_lookup_dev(dev, &entry, addr);
1229 rte_rwlock_read_unlock(&priv->mr.rwlock);
1230 if (lkey != UINT32_MAX)
1231 return;
1232 mr = rte_zmalloc_socket(NULL,
1233 RTE_ALIGN_CEIL(sizeof(*mr),
1234 RTE_CACHE_LINE_SIZE),
1235 RTE_CACHE_LINE_SIZE, mp->socket_id);
1236 if (mr == NULL) {
1237 WARN("port %u unable to allocate memory for a new MR of"
1238 " mempool (%s).",
1239 dev->data->port_id, mp->name);
1240 data->ret = -1;
1241 return;
1242 }
1243 DEBUG("port %u register MR for chunk #%d of mempool (%s)",
1244 dev->data->port_id, mem_idx, mp->name);
1245 mr->ibv_mr = mlx4_glue->reg_mr(priv->pd, (void *)addr, len,
1246 IBV_ACCESS_LOCAL_WRITE);
1247 if (mr->ibv_mr == NULL) {
1248 WARN("port %u fail to create a verbs MR for address (%p)",
1249 dev->data->port_id, (void *)addr);
1250 rte_free(mr);
1251 data->ret = -1;
1252 return;
1253 }
1254 mr->msl = NULL; /* Mark it is external memory. */
1255 mr->ms_bmp = NULL;
1256 mr->ms_n = 1;
1257 mr->ms_bmp_n = 1;
1258 rte_rwlock_write_lock(&priv->mr.rwlock);
1259 LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
1260 DEBUG("port %u MR CREATED (%p) for external memory %p:\n"
1261 " [0x%" PRIxPTR ", 0x%" PRIxPTR "),"
1262 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
1263 dev->data->port_id, (void *)mr, (void *)addr,
1264 addr, addr + len, rte_cpu_to_be_32(mr->ibv_mr->lkey),
1265 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
1266 /* Insert to the global cache table. */
1267 mr_insert_dev_cache(dev, mr);
1268 rte_rwlock_write_unlock(&priv->mr.rwlock);
1269 /* Insert to the local cache table */
1270 mlx4_mr_addr2mr_bh(dev, mr_ctrl, addr);
1271 }
1272
1273 /**
1274 * Register MR for entire memory chunks in a Mempool having externally allocated
1275 * memory and fill in local cache.
1276 *
1277 * @param dev
1278 * Pointer to Ethernet device.
1279 * @param mr_ctrl
1280 * Pointer to per-queue MR control structure.
1281 * @param mp
1282 * Pointer to registering Mempool.
1283 *
1284 * @return
1285 * 0 on success, -1 on failure.
1286 */
1287 static uint32_t
1288 mlx4_mr_update_ext_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1289 struct rte_mempool *mp)
1290 {
1291 struct mr_update_mp_data data = {
1292 .dev = dev,
1293 .mr_ctrl = mr_ctrl,
1294 .ret = 0,
1295 };
1296
1297 rte_mempool_mem_iter(mp, mlx4_mr_update_ext_mp_cb, &data);
1298 return data.ret;
1299 }
1300
1301 /**
1302 * Register MR entire memory chunks in a Mempool having externally allocated
1303 * memory and search LKey of the address to return.
1304 *
1305 * @param dev
1306 * Pointer to Ethernet device.
1307 * @param addr
1308 * Search key.
1309 * @param mp
1310 * Pointer to registering Mempool where addr belongs.
1311 *
1312 * @return
1313 * LKey for address on success, UINT32_MAX on failure.
1314 */
1315 uint32_t
1316 mlx4_tx_update_ext_mp(struct txq *txq, uintptr_t addr, struct rte_mempool *mp)
1317 {
1318 struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
1319 struct mlx4_priv *priv = txq->priv;
1320
1321 if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
1322 WARN("port %u using address (%p) from unregistered mempool"
1323 " having externally allocated memory"
1324 " in secondary process, please create mempool"
1325 " prior to rte_eth_dev_start()",
1326 PORT_ID(priv), (void *)addr);
1327 return UINT32_MAX;
1328 }
1329 mlx4_mr_update_ext_mp(ETH_DEV(priv), mr_ctrl, mp);
1330 return mlx4_tx_addr2mr_bh(txq, addr);
1331 }
1332
1333 /* Called during rte_mempool_mem_iter() by mlx4_mr_update_mp(). */
1334 static void
1335 mlx4_mr_update_mp_cb(struct rte_mempool *mp __rte_unused, void *opaque,
1336 struct rte_mempool_memhdr *memhdr,
1337 unsigned mem_idx __rte_unused)
1338 {
1339 struct mr_update_mp_data *data = opaque;
1340 uint32_t lkey;
1341
1342 /* Stop iteration if failed in the previous walk. */
1343 if (data->ret < 0)
1344 return;
1345 /* Register address of the chunk and update local caches. */
1346 lkey = mlx4_mr_addr2mr_bh(data->dev, data->mr_ctrl,
1347 (uintptr_t)memhdr->addr);
1348 if (lkey == UINT32_MAX)
1349 data->ret = -1;
1350 }
1351
1352 /**
1353 * Register entire memory chunks in a Mempool.
1354 *
1355 * @param dev
1356 * Pointer to Ethernet device.
1357 * @param mr_ctrl
1358 * Pointer to per-queue MR control structure.
1359 * @param mp
1360 * Pointer to registering Mempool.
1361 *
1362 * @return
1363 * 0 on success, -1 on failure.
1364 */
1365 int
1366 mlx4_mr_update_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl,
1367 struct rte_mempool *mp)
1368 {
1369 struct mr_update_mp_data data = {
1370 .dev = dev,
1371 .mr_ctrl = mr_ctrl,
1372 .ret = 0,
1373 };
1374
1375 rte_mempool_mem_iter(mp, mlx4_mr_update_mp_cb, &data);
1376 if (data.ret < 0 && rte_errno == ENXIO) {
1377 /* Mempool may have externally allocated memory. */
1378 return mlx4_mr_update_ext_mp(dev, mr_ctrl, mp);
1379 }
1380 return data.ret;
1381 }
1382
1383 #ifndef NDEBUG
1384 /**
1385 * Dump all the created MRs and the global cache entries.
1386 *
1387 * @param dev
1388 * Pointer to Ethernet device.
1389 */
1390 void
1391 mlx4_mr_dump_dev(struct rte_eth_dev *dev)
1392 {
1393 struct mlx4_priv *priv = dev->data->dev_private;
1394 struct mlx4_mr *mr;
1395 int mr_n = 0;
1396 int chunk_n = 0;
1397
1398 rte_rwlock_read_lock(&priv->mr.rwlock);
1399 /* Iterate all the existing MRs. */
1400 LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
1401 unsigned int n;
1402
1403 DEBUG("port %u MR[%u], LKey = 0x%x, ms_n = %u, ms_bmp_n = %u",
1404 dev->data->port_id, mr_n++,
1405 rte_cpu_to_be_32(mr->ibv_mr->lkey),
1406 mr->ms_n, mr->ms_bmp_n);
1407 if (mr->ms_n == 0)
1408 continue;
1409 for (n = 0; n < mr->ms_bmp_n; ) {
1410 struct mlx4_mr_cache ret;
1411
1412 memset(&ret, 0, sizeof(ret));
1413 n = mr_find_next_chunk(mr, &ret, n);
1414 if (!ret.end)
1415 break;
1416 DEBUG(" chunk[%u], [0x%" PRIxPTR ", 0x%" PRIxPTR ")",
1417 chunk_n++, ret.start, ret.end);
1418 }
1419 }
1420 DEBUG("port %u dumping global cache", dev->data->port_id);
1421 mlx4_mr_btree_dump(&priv->mr.cache);
1422 rte_rwlock_read_unlock(&priv->mr.rwlock);
1423 }
1424 #endif
1425
1426 /**
1427 * Release all the created MRs and resources. Remove device from memory callback
1428 * list.
1429 *
1430 * @param dev
1431 * Pointer to Ethernet device.
1432 */
1433 void
1434 mlx4_mr_release(struct rte_eth_dev *dev)
1435 {
1436 struct mlx4_priv *priv = dev->data->dev_private;
1437 struct mlx4_mr *mr_next;
1438
1439 /* Remove from memory callback device list. */
1440 rte_rwlock_write_lock(&mlx4_shared_data->mem_event_rwlock);
1441 LIST_REMOVE(priv, mem_event_cb);
1442 rte_rwlock_write_unlock(&mlx4_shared_data->mem_event_rwlock);
1443 #ifndef NDEBUG
1444 mlx4_mr_dump_dev(dev);
1445 #endif
1446 rte_rwlock_write_lock(&priv->mr.rwlock);
1447 /* Detach from MR list and move to free list. */
1448 mr_next = LIST_FIRST(&priv->mr.mr_list);
1449 while (mr_next != NULL) {
1450 struct mlx4_mr *mr = mr_next;
1451
1452 mr_next = LIST_NEXT(mr, mr);
1453 LIST_REMOVE(mr, mr);
1454 LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
1455 }
1456 LIST_INIT(&priv->mr.mr_list);
1457 /* Free global cache. */
1458 mlx4_mr_btree_free(&priv->mr.cache);
1459 rte_rwlock_write_unlock(&priv->mr.rwlock);
1460 /* Free all remaining MRs. */
1461 mlx4_mr_garbage_collect(dev);
1462 }