1// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2/*
3 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5 */
6
7#include <linux/libnvdimm.h>
8
9#include "rxe.h"
10#include "rxe_loc.h"
11
12/* Return a random 8 bit key value that is
13 * different than the last_key. Set last_key to -1
14 * if this is the first key for an MR or MW
15 */
16u8 rxe_get_next_key(u32 last_key)
17{
18 u8 key;
19
20 do {
21 get_random_bytes(buf: &key, len: 1);
22 } while (key == last_key);
23
24 return key;
25}
26
27int mr_check_range(struct rxe_mr *mr, u64 iova, size_t length)
28{
29 switch (mr->ibmr.type) {
30 case IB_MR_TYPE_DMA:
31 return 0;
32
33 case IB_MR_TYPE_USER:
34 case IB_MR_TYPE_MEM_REG:
35 if (iova < mr->ibmr.iova ||
36 iova + length > mr->ibmr.iova + mr->ibmr.length) {
37 rxe_dbg_mr(mr, "iova/length out of range\n");
38 return -EINVAL;
39 }
40 return 0;
41
42 default:
43 rxe_dbg_mr(mr, "mr type not supported\n");
44 return -EINVAL;
45 }
46}
47
48static void rxe_mr_init(int access, struct rxe_mr *mr)
49{
50 u32 key = mr->elem.index << 8 | rxe_get_next_key(last_key: -1);
51
52 /* set ibmr->l/rkey and also copy into private l/rkey
53 * for user MRs these will always be the same
54 * for cases where caller 'owns' the key portion
55 * they may be different until REG_MR WQE is executed.
56 */
57 mr->lkey = mr->ibmr.lkey = key;
58 mr->rkey = mr->ibmr.rkey = key;
59
60 mr->access = access;
61 mr->ibmr.page_size = PAGE_SIZE;
62 mr->page_mask = PAGE_MASK;
63 mr->page_shift = PAGE_SHIFT;
64 mr->state = RXE_MR_STATE_INVALID;
65}
66
67void rxe_mr_init_dma(int access, struct rxe_mr *mr)
68{
69 rxe_mr_init(access, mr);
70
71 mr->state = RXE_MR_STATE_VALID;
72 mr->ibmr.type = IB_MR_TYPE_DMA;
73}
74
75static unsigned long rxe_mr_iova_to_index(struct rxe_mr *mr, u64 iova)
76{
77 return (iova >> mr->page_shift) - (mr->ibmr.iova >> mr->page_shift);
78}
79
80static unsigned long rxe_mr_iova_to_page_offset(struct rxe_mr *mr, u64 iova)
81{
82 return iova & (mr_page_size(mr) - 1);
83}
84
85static bool is_pmem_page(struct page *pg)
86{
87 unsigned long paddr = page_to_phys(pg);
88
89 return REGION_INTERSECTS ==
90 region_intersects(offset: paddr, PAGE_SIZE, IORESOURCE_MEM,
91 desc: IORES_DESC_PERSISTENT_MEMORY);
92}
93
94static int rxe_mr_fill_pages_from_sgt(struct rxe_mr *mr, struct sg_table *sgt)
95{
96 XA_STATE(xas, &mr->page_list, 0);
97 struct sg_page_iter sg_iter;
98 struct page *page;
99 bool persistent = !!(mr->access & IB_ACCESS_FLUSH_PERSISTENT);
100
101 __sg_page_iter_start(piter: &sg_iter, sglist: sgt->sgl, nents: sgt->orig_nents, pgoffset: 0);
102 if (!__sg_page_iter_next(piter: &sg_iter))
103 return 0;
104
105 do {
106 xas_lock(&xas);
107 while (true) {
108 page = sg_page_iter_page(piter: &sg_iter);
109
110 if (persistent && !is_pmem_page(pg: page)) {
111 rxe_dbg_mr(mr, "Page can't be persistent\n");
112 xas_set_err(xas: &xas, err: -EINVAL);
113 break;
114 }
115
116 xas_store(&xas, entry: page);
117 if (xas_error(xas: &xas))
118 break;
119 xas_next(xas: &xas);
120 if (!__sg_page_iter_next(piter: &sg_iter))
121 break;
122 }
123 xas_unlock(&xas);
124 } while (xas_nomem(&xas, GFP_KERNEL));
125
126 return xas_error(xas: &xas);
127}
128
129int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length,
130 int access, struct rxe_mr *mr)
131{
132 struct ib_umem *umem;
133 int err;
134
135 rxe_mr_init(access, mr);
136
137 xa_init(xa: &mr->page_list);
138
139 umem = ib_umem_get(device: &rxe->ib_dev, addr: start, size: length, access);
140 if (IS_ERR(ptr: umem)) {
141 rxe_dbg_mr(mr, "Unable to pin memory region err = %d\n",
142 (int)PTR_ERR(umem));
143 return PTR_ERR(ptr: umem);
144 }
145
146 err = rxe_mr_fill_pages_from_sgt(mr, sgt: &umem->sgt_append.sgt);
147 if (err) {
148 ib_umem_release(umem);
149 return err;
150 }
151
152 mr->umem = umem;
153 mr->ibmr.type = IB_MR_TYPE_USER;
154 mr->state = RXE_MR_STATE_VALID;
155
156 return 0;
157}
158
159static int rxe_mr_alloc(struct rxe_mr *mr, int num_buf)
160{
161 XA_STATE(xas, &mr->page_list, 0);
162 int i = 0;
163 int err;
164
165 xa_init(xa: &mr->page_list);
166
167 do {
168 xas_lock(&xas);
169 while (i != num_buf) {
170 xas_store(&xas, XA_ZERO_ENTRY);
171 if (xas_error(xas: &xas))
172 break;
173 xas_next(xas: &xas);
174 i++;
175 }
176 xas_unlock(&xas);
177 } while (xas_nomem(&xas, GFP_KERNEL));
178
179 err = xas_error(xas: &xas);
180 if (err)
181 return err;
182
183 mr->num_buf = num_buf;
184
185 return 0;
186}
187
188int rxe_mr_init_fast(int max_pages, struct rxe_mr *mr)
189{
190 int err;
191
192 /* always allow remote access for FMRs */
193 rxe_mr_init(access: RXE_ACCESS_REMOTE, mr);
194
195 err = rxe_mr_alloc(mr, num_buf: max_pages);
196 if (err)
197 goto err1;
198
199 mr->state = RXE_MR_STATE_FREE;
200 mr->ibmr.type = IB_MR_TYPE_MEM_REG;
201
202 return 0;
203
204err1:
205 return err;
206}
207
208static int rxe_set_page(struct ib_mr *ibmr, u64 dma_addr)
209{
210 struct rxe_mr *mr = to_rmr(mr: ibmr);
211 struct page *page = ib_virt_dma_to_page(dma_addr);
212 bool persistent = !!(mr->access & IB_ACCESS_FLUSH_PERSISTENT);
213 int err;
214
215 if (persistent && !is_pmem_page(pg: page)) {
216 rxe_dbg_mr(mr, "Page cannot be persistent\n");
217 return -EINVAL;
218 }
219
220 if (unlikely(mr->nbuf == mr->num_buf))
221 return -ENOMEM;
222
223 err = xa_err(entry: xa_store(&mr->page_list, index: mr->nbuf, entry: page, GFP_KERNEL));
224 if (err)
225 return err;
226
227 mr->nbuf++;
228 return 0;
229}
230
231int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sgl,
232 int sg_nents, unsigned int *sg_offset)
233{
234 struct rxe_mr *mr = to_rmr(mr: ibmr);
235 unsigned int page_size = mr_page_size(mr);
236
237 mr->nbuf = 0;
238 mr->page_shift = ilog2(page_size);
239 mr->page_mask = ~((u64)page_size - 1);
240 mr->page_offset = mr->ibmr.iova & (page_size - 1);
241
242 return ib_sg_to_pages(mr: ibmr, sgl, sg_nents, sg_offset, set_page: rxe_set_page);
243}
244
245static int rxe_mr_copy_xarray(struct rxe_mr *mr, u64 iova, void *addr,
246 unsigned int length, enum rxe_mr_copy_dir dir)
247{
248 unsigned int page_offset = rxe_mr_iova_to_page_offset(mr, iova);
249 unsigned long index = rxe_mr_iova_to_index(mr, iova);
250 unsigned int bytes;
251 struct page *page;
252 void *va;
253
254 while (length) {
255 page = xa_load(&mr->page_list, index);
256 if (!page)
257 return -EFAULT;
258
259 bytes = min_t(unsigned int, length,
260 mr_page_size(mr) - page_offset);
261 va = kmap_local_page(page);
262 if (dir == RXE_FROM_MR_OBJ)
263 memcpy(addr, va + page_offset, bytes);
264 else
265 memcpy(va + page_offset, addr, bytes);
266 kunmap_local(va);
267
268 page_offset = 0;
269 addr += bytes;
270 length -= bytes;
271 index++;
272 }
273
274 return 0;
275}
276
277static void rxe_mr_copy_dma(struct rxe_mr *mr, u64 dma_addr, void *addr,
278 unsigned int length, enum rxe_mr_copy_dir dir)
279{
280 unsigned int page_offset = dma_addr & (PAGE_SIZE - 1);
281 unsigned int bytes;
282 struct page *page;
283 u8 *va;
284
285 while (length) {
286 page = ib_virt_dma_to_page(dma_addr);
287 bytes = min_t(unsigned int, length,
288 PAGE_SIZE - page_offset);
289 va = kmap_local_page(page);
290
291 if (dir == RXE_TO_MR_OBJ)
292 memcpy(va + page_offset, addr, bytes);
293 else
294 memcpy(addr, va + page_offset, bytes);
295
296 kunmap_local(va);
297 page_offset = 0;
298 dma_addr += bytes;
299 addr += bytes;
300 length -= bytes;
301 }
302}
303
304int rxe_mr_copy(struct rxe_mr *mr, u64 iova, void *addr,
305 unsigned int length, enum rxe_mr_copy_dir dir)
306{
307 int err;
308
309 if (length == 0)
310 return 0;
311
312 if (WARN_ON(!mr))
313 return -EINVAL;
314
315 if (mr->ibmr.type == IB_MR_TYPE_DMA) {
316 rxe_mr_copy_dma(mr, dma_addr: iova, addr, length, dir);
317 return 0;
318 }
319
320 err = mr_check_range(mr, iova, length);
321 if (unlikely(err)) {
322 rxe_dbg_mr(mr, "iova out of range\n");
323 return err;
324 }
325
326 return rxe_mr_copy_xarray(mr, iova, addr, length, dir);
327}
328
329/* copy data in or out of a wqe, i.e. sg list
330 * under the control of a dma descriptor
331 */
332int copy_data(
333 struct rxe_pd *pd,
334 int access,
335 struct rxe_dma_info *dma,
336 void *addr,
337 int length,
338 enum rxe_mr_copy_dir dir)
339{
340 int bytes;
341 struct rxe_sge *sge = &dma->sge[dma->cur_sge];
342 int offset = dma->sge_offset;
343 int resid = dma->resid;
344 struct rxe_mr *mr = NULL;
345 u64 iova;
346 int err;
347
348 if (length == 0)
349 return 0;
350
351 if (length > resid) {
352 err = -EINVAL;
353 goto err2;
354 }
355
356 if (sge->length && (offset < sge->length)) {
357 mr = lookup_mr(pd, access, key: sge->lkey, type: RXE_LOOKUP_LOCAL);
358 if (!mr) {
359 err = -EINVAL;
360 goto err1;
361 }
362 }
363
364 while (length > 0) {
365 bytes = length;
366
367 if (offset >= sge->length) {
368 if (mr) {
369 rxe_put(mr);
370 mr = NULL;
371 }
372 sge++;
373 dma->cur_sge++;
374 offset = 0;
375
376 if (dma->cur_sge >= dma->num_sge) {
377 err = -ENOSPC;
378 goto err2;
379 }
380
381 if (sge->length) {
382 mr = lookup_mr(pd, access, key: sge->lkey,
383 type: RXE_LOOKUP_LOCAL);
384 if (!mr) {
385 err = -EINVAL;
386 goto err1;
387 }
388 } else {
389 continue;
390 }
391 }
392
393 if (bytes > sge->length - offset)
394 bytes = sge->length - offset;
395
396 if (bytes > 0) {
397 iova = sge->addr + offset;
398 err = rxe_mr_copy(mr, iova, addr, length: bytes, dir);
399 if (err)
400 goto err2;
401
402 offset += bytes;
403 resid -= bytes;
404 length -= bytes;
405 addr += bytes;
406 }
407 }
408
409 dma->sge_offset = offset;
410 dma->resid = resid;
411
412 if (mr)
413 rxe_put(mr);
414
415 return 0;
416
417err2:
418 if (mr)
419 rxe_put(mr);
420err1:
421 return err;
422}
423
424int rxe_flush_pmem_iova(struct rxe_mr *mr, u64 iova, unsigned int length)
425{
426 unsigned int page_offset;
427 unsigned long index;
428 struct page *page;
429 unsigned int bytes;
430 int err;
431 u8 *va;
432
433 /* mr must be valid even if length is zero */
434 if (WARN_ON(!mr))
435 return -EINVAL;
436
437 if (length == 0)
438 return 0;
439
440 if (mr->ibmr.type == IB_MR_TYPE_DMA)
441 return -EFAULT;
442
443 err = mr_check_range(mr, iova, length);
444 if (err)
445 return err;
446
447 while (length > 0) {
448 index = rxe_mr_iova_to_index(mr, iova);
449 page = xa_load(&mr->page_list, index);
450 page_offset = rxe_mr_iova_to_page_offset(mr, iova);
451 if (!page)
452 return -EFAULT;
453 bytes = min_t(unsigned int, length,
454 mr_page_size(mr) - page_offset);
455
456 va = kmap_local_page(page);
457 arch_wb_cache_pmem(addr: va + page_offset, size: bytes);
458 kunmap_local(va);
459
460 length -= bytes;
461 iova += bytes;
462 page_offset = 0;
463 }
464
465 return 0;
466}
467
468/* Guarantee atomicity of atomic operations at the machine level. */
469static DEFINE_SPINLOCK(atomic_ops_lock);
470
471int rxe_mr_do_atomic_op(struct rxe_mr *mr, u64 iova, int opcode,
472 u64 compare, u64 swap_add, u64 *orig_val)
473{
474 unsigned int page_offset;
475 struct page *page;
476 u64 value;
477 u64 *va;
478
479 if (unlikely(mr->state != RXE_MR_STATE_VALID)) {
480 rxe_dbg_mr(mr, "mr not in valid state\n");
481 return RESPST_ERR_RKEY_VIOLATION;
482 }
483
484 if (mr->ibmr.type == IB_MR_TYPE_DMA) {
485 page_offset = iova & (PAGE_SIZE - 1);
486 page = ib_virt_dma_to_page(dma_addr: iova);
487 } else {
488 unsigned long index;
489 int err;
490
491 err = mr_check_range(mr, iova, length: sizeof(value));
492 if (err) {
493 rxe_dbg_mr(mr, "iova out of range\n");
494 return RESPST_ERR_RKEY_VIOLATION;
495 }
496 page_offset = rxe_mr_iova_to_page_offset(mr, iova);
497 index = rxe_mr_iova_to_index(mr, iova);
498 page = xa_load(&mr->page_list, index);
499 if (!page)
500 return RESPST_ERR_RKEY_VIOLATION;
501 }
502
503 if (unlikely(page_offset & 0x7)) {
504 rxe_dbg_mr(mr, "iova not aligned\n");
505 return RESPST_ERR_MISALIGNED_ATOMIC;
506 }
507
508 va = kmap_local_page(page);
509
510 spin_lock_bh(lock: &atomic_ops_lock);
511 value = *orig_val = va[page_offset >> 3];
512
513 if (opcode == IB_OPCODE_RC_COMPARE_SWAP) {
514 if (value == compare)
515 va[page_offset >> 3] = swap_add;
516 } else {
517 value += swap_add;
518 va[page_offset >> 3] = value;
519 }
520 spin_unlock_bh(lock: &atomic_ops_lock);
521
522 kunmap_local(va);
523
524 return 0;
525}
526
527#if defined CONFIG_64BIT
528/* only implemented or called for 64 bit architectures */
529int rxe_mr_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)
530{
531 unsigned int page_offset;
532 struct page *page;
533 u64 *va;
534
535 /* See IBA oA19-28 */
536 if (unlikely(mr->state != RXE_MR_STATE_VALID)) {
537 rxe_dbg_mr(mr, "mr not in valid state\n");
538 return RESPST_ERR_RKEY_VIOLATION;
539 }
540
541 if (mr->ibmr.type == IB_MR_TYPE_DMA) {
542 page_offset = iova & (PAGE_SIZE - 1);
543 page = ib_virt_dma_to_page(dma_addr: iova);
544 } else {
545 unsigned long index;
546 int err;
547
548 /* See IBA oA19-28 */
549 err = mr_check_range(mr, iova, length: sizeof(value));
550 if (unlikely(err)) {
551 rxe_dbg_mr(mr, "iova out of range\n");
552 return RESPST_ERR_RKEY_VIOLATION;
553 }
554 page_offset = rxe_mr_iova_to_page_offset(mr, iova);
555 index = rxe_mr_iova_to_index(mr, iova);
556 page = xa_load(&mr->page_list, index);
557 if (!page)
558 return RESPST_ERR_RKEY_VIOLATION;
559 }
560
561 /* See IBA A19.4.2 */
562 if (unlikely(page_offset & 0x7)) {
563 rxe_dbg_mr(mr, "misaligned address\n");
564 return RESPST_ERR_MISALIGNED_ATOMIC;
565 }
566
567 va = kmap_local_page(page);
568
569 /* Do atomic write after all prior operations have completed */
570 smp_store_release(&va[page_offset >> 3], value);
571
572 kunmap_local(va);
573
574 return 0;
575}
576#else
577int rxe_mr_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)
578{
579 return RESPST_ERR_UNSUPPORTED_OPCODE;
580}
581#endif
582
583int advance_dma_data(struct rxe_dma_info *dma, unsigned int length)
584{
585 struct rxe_sge *sge = &dma->sge[dma->cur_sge];
586 int offset = dma->sge_offset;
587 int resid = dma->resid;
588
589 while (length) {
590 unsigned int bytes;
591
592 if (offset >= sge->length) {
593 sge++;
594 dma->cur_sge++;
595 offset = 0;
596 if (dma->cur_sge >= dma->num_sge)
597 return -ENOSPC;
598 }
599
600 bytes = length;
601
602 if (bytes > sge->length - offset)
603 bytes = sge->length - offset;
604
605 offset += bytes;
606 resid -= bytes;
607 length -= bytes;
608 }
609
610 dma->sge_offset = offset;
611 dma->resid = resid;
612
613 return 0;
614}
615
616struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,
617 enum rxe_mr_lookup_type type)
618{
619 struct rxe_mr *mr;
620 struct rxe_dev *rxe = to_rdev(dev: pd->ibpd.device);
621 int index = key >> 8;
622
623 mr = rxe_pool_get_index(pool: &rxe->mr_pool, index);
624 if (!mr)
625 return NULL;
626
627 if (unlikely((type == RXE_LOOKUP_LOCAL && mr->lkey != key) ||
628 (type == RXE_LOOKUP_REMOTE && mr->rkey != key) ||
629 mr_pd(mr) != pd || ((access & mr->access) != access) ||
630 mr->state != RXE_MR_STATE_VALID)) {
631 rxe_put(mr);
632 mr = NULL;
633 }
634
635 return mr;
636}
637
638int rxe_invalidate_mr(struct rxe_qp *qp, u32 key)
639{
640 struct rxe_dev *rxe = to_rdev(dev: qp->ibqp.device);
641 struct rxe_mr *mr;
642 int remote;
643 int ret;
644
645 mr = rxe_pool_get_index(pool: &rxe->mr_pool, index: key >> 8);
646 if (!mr) {
647 rxe_dbg_qp(qp, "No MR for key %#x\n", key);
648 ret = -EINVAL;
649 goto err;
650 }
651
652 remote = mr->access & RXE_ACCESS_REMOTE;
653 if (remote ? (key != mr->rkey) : (key != mr->lkey)) {
654 rxe_dbg_mr(mr, "wr key (%#x) doesn't match mr key (%#x)\n",
655 key, (remote ? mr->rkey : mr->lkey));
656 ret = -EINVAL;
657 goto err_drop_ref;
658 }
659
660 if (atomic_read(v: &mr->num_mw) > 0) {
661 rxe_dbg_mr(mr, "Attempt to invalidate an MR while bound to MWs\n");
662 ret = -EINVAL;
663 goto err_drop_ref;
664 }
665
666 if (unlikely(mr->ibmr.type != IB_MR_TYPE_MEM_REG)) {
667 rxe_dbg_mr(mr, "Type (%d) is wrong\n", mr->ibmr.type);
668 ret = -EINVAL;
669 goto err_drop_ref;
670 }
671
672 mr->state = RXE_MR_STATE_FREE;
673 ret = 0;
674
675err_drop_ref:
676 rxe_put(mr);
677err:
678 return ret;
679}
680
681/* user can (re)register fast MR by executing a REG_MR WQE.
682 * user is expected to hold a reference on the ib mr until the
683 * WQE completes.
684 * Once a fast MR is created this is the only way to change the
685 * private keys. It is the responsibility of the user to maintain
686 * the ib mr keys in sync with rxe mr keys.
687 */
688int rxe_reg_fast_mr(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
689{
690 struct rxe_mr *mr = to_rmr(mr: wqe->wr.wr.reg.mr);
691 u32 key = wqe->wr.wr.reg.key;
692 u32 access = wqe->wr.wr.reg.access;
693
694 /* user can only register MR in free state */
695 if (unlikely(mr->state != RXE_MR_STATE_FREE)) {
696 rxe_dbg_mr(mr, "mr->lkey = 0x%x not free\n", mr->lkey);
697 return -EINVAL;
698 }
699
700 /* user can only register mr with qp in same protection domain */
701 if (unlikely(qp->ibqp.pd != mr->ibmr.pd)) {
702 rxe_dbg_mr(mr, "qp->pd and mr->pd don't match\n");
703 return -EINVAL;
704 }
705
706 /* user is only allowed to change key portion of l/rkey */
707 if (unlikely((mr->lkey & ~0xff) != (key & ~0xff))) {
708 rxe_dbg_mr(mr, "key = 0x%x has wrong index mr->lkey = 0x%x\n",
709 key, mr->lkey);
710 return -EINVAL;
711 }
712
713 mr->access = access;
714 mr->lkey = key;
715 mr->rkey = key;
716 mr->ibmr.iova = wqe->wr.wr.reg.mr->iova;
717 mr->state = RXE_MR_STATE_VALID;
718
719 return 0;
720}
721
722void rxe_mr_cleanup(struct rxe_pool_elem *elem)
723{
724 struct rxe_mr *mr = container_of(elem, typeof(*mr), elem);
725
726 rxe_put(mr_pd(mr));
727 ib_umem_release(umem: mr->umem);
728
729 if (mr->ibmr.type != IB_MR_TYPE_DMA)
730 xa_destroy(&mr->page_list);
731}
732

source code of linux/drivers/infiniband/sw/rxe/rxe_mr.c