xref: /openbmc/linux/drivers/infiniband/hw/mlx5/umr.c (revision 0e407915)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /* Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. */
3 
4 #include <rdma/ib_umem_odp.h>
5 #include "mlx5_ib.h"
6 #include "umr.h"
7 #include "wr.h"
8 
9 /*
10  * We can't use an array for xlt_emergency_page because dma_map_single doesn't
11  * work on kernel modules memory
12  */
13 void *xlt_emergency_page;
14 static DEFINE_MUTEX(xlt_emergency_page_mutex);
15 
16 static __be64 get_umr_enable_mr_mask(void)
17 {
18 	u64 result;
19 
20 	result = MLX5_MKEY_MASK_KEY |
21 		 MLX5_MKEY_MASK_FREE;
22 
23 	return cpu_to_be64(result);
24 }
25 
26 static __be64 get_umr_disable_mr_mask(void)
27 {
28 	u64 result;
29 
30 	result = MLX5_MKEY_MASK_FREE;
31 
32 	return cpu_to_be64(result);
33 }
34 
35 static __be64 get_umr_update_translation_mask(void)
36 {
37 	u64 result;
38 
39 	result = MLX5_MKEY_MASK_LEN |
40 		 MLX5_MKEY_MASK_PAGE_SIZE |
41 		 MLX5_MKEY_MASK_START_ADDR;
42 
43 	return cpu_to_be64(result);
44 }
45 
46 static __be64 get_umr_update_access_mask(struct mlx5_ib_dev *dev)
47 {
48 	u64 result;
49 
50 	result = MLX5_MKEY_MASK_LR |
51 		 MLX5_MKEY_MASK_LW |
52 		 MLX5_MKEY_MASK_RR |
53 		 MLX5_MKEY_MASK_RW;
54 
55 	if (MLX5_CAP_GEN(dev->mdev, atomic))
56 		result |= MLX5_MKEY_MASK_A;
57 
58 	if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write_umr))
59 		result |= MLX5_MKEY_MASK_RELAXED_ORDERING_WRITE;
60 
61 	if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read_umr))
62 		result |= MLX5_MKEY_MASK_RELAXED_ORDERING_READ;
63 
64 	return cpu_to_be64(result);
65 }
66 
67 static __be64 get_umr_update_pd_mask(void)
68 {
69 	u64 result;
70 
71 	result = MLX5_MKEY_MASK_PD;
72 
73 	return cpu_to_be64(result);
74 }
75 
76 static int umr_check_mkey_mask(struct mlx5_ib_dev *dev, u64 mask)
77 {
78 	if (mask & MLX5_MKEY_MASK_PAGE_SIZE &&
79 	    MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
80 		return -EPERM;
81 
82 	if (mask & MLX5_MKEY_MASK_A &&
83 	    MLX5_CAP_GEN(dev->mdev, umr_modify_atomic_disabled))
84 		return -EPERM;
85 
86 	if (mask & MLX5_MKEY_MASK_RELAXED_ORDERING_WRITE &&
87 	    !MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write_umr))
88 		return -EPERM;
89 
90 	if (mask & MLX5_MKEY_MASK_RELAXED_ORDERING_READ &&
91 	    !MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read_umr))
92 		return -EPERM;
93 
94 	return 0;
95 }
96 
97 enum {
98 	MAX_UMR_WR = 128,
99 };
100 
101 static int mlx5r_umr_qp_rst2rts(struct mlx5_ib_dev *dev, struct ib_qp *qp)
102 {
103 	struct ib_qp_attr attr = {};
104 	int ret;
105 
106 	attr.qp_state = IB_QPS_INIT;
107 	attr.port_num = 1;
108 	ret = ib_modify_qp(qp, &attr,
109 			   IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT);
110 	if (ret) {
111 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
112 		return ret;
113 	}
114 
115 	memset(&attr, 0, sizeof(attr));
116 	attr.qp_state = IB_QPS_RTR;
117 
118 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
119 	if (ret) {
120 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
121 		return ret;
122 	}
123 
124 	memset(&attr, 0, sizeof(attr));
125 	attr.qp_state = IB_QPS_RTS;
126 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
127 	if (ret) {
128 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
129 		return ret;
130 	}
131 
132 	return 0;
133 }
134 
135 int mlx5r_umr_resource_init(struct mlx5_ib_dev *dev)
136 {
137 	struct ib_qp_init_attr init_attr = {};
138 	struct ib_pd *pd;
139 	struct ib_cq *cq;
140 	struct ib_qp *qp;
141 	int ret;
142 
143 	pd = ib_alloc_pd(&dev->ib_dev, 0);
144 	if (IS_ERR(pd)) {
145 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
146 		return PTR_ERR(pd);
147 	}
148 
149 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
150 	if (IS_ERR(cq)) {
151 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
152 		ret = PTR_ERR(cq);
153 		goto destroy_pd;
154 	}
155 
156 	init_attr.send_cq = cq;
157 	init_attr.recv_cq = cq;
158 	init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
159 	init_attr.cap.max_send_wr = MAX_UMR_WR;
160 	init_attr.cap.max_send_sge = 1;
161 	init_attr.qp_type = MLX5_IB_QPT_REG_UMR;
162 	init_attr.port_num = 1;
163 	qp = ib_create_qp(pd, &init_attr);
164 	if (IS_ERR(qp)) {
165 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
166 		ret = PTR_ERR(qp);
167 		goto destroy_cq;
168 	}
169 
170 	ret = mlx5r_umr_qp_rst2rts(dev, qp);
171 	if (ret)
172 		goto destroy_qp;
173 
174 	dev->umrc.qp = qp;
175 	dev->umrc.cq = cq;
176 	dev->umrc.pd = pd;
177 
178 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
179 	mutex_init(&dev->umrc.lock);
180 	dev->umrc.state = MLX5_UMR_STATE_ACTIVE;
181 
182 	return 0;
183 
184 destroy_qp:
185 	ib_destroy_qp(qp);
186 destroy_cq:
187 	ib_free_cq(cq);
188 destroy_pd:
189 	ib_dealloc_pd(pd);
190 	return ret;
191 }
192 
193 void mlx5r_umr_resource_cleanup(struct mlx5_ib_dev *dev)
194 {
195 	if (dev->umrc.state == MLX5_UMR_STATE_UNINIT)
196 		return;
197 	ib_destroy_qp(dev->umrc.qp);
198 	ib_free_cq(dev->umrc.cq);
199 	ib_dealloc_pd(dev->umrc.pd);
200 }
201 
202 static int mlx5r_umr_recover(struct mlx5_ib_dev *dev)
203 {
204 	struct umr_common *umrc = &dev->umrc;
205 	struct ib_qp_attr attr;
206 	int err;
207 
208 	attr.qp_state = IB_QPS_RESET;
209 	err = ib_modify_qp(umrc->qp, &attr, IB_QP_STATE);
210 	if (err) {
211 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
212 		goto err;
213 	}
214 
215 	err = mlx5r_umr_qp_rst2rts(dev, umrc->qp);
216 	if (err)
217 		goto err;
218 
219 	umrc->state = MLX5_UMR_STATE_ACTIVE;
220 	return 0;
221 
222 err:
223 	umrc->state = MLX5_UMR_STATE_ERR;
224 	return err;
225 }
226 
227 static int mlx5r_umr_post_send(struct ib_qp *ibqp, u32 mkey, struct ib_cqe *cqe,
228 			       struct mlx5r_umr_wqe *wqe, bool with_data)
229 {
230 	unsigned int wqe_size =
231 		with_data ? sizeof(struct mlx5r_umr_wqe) :
232 			    sizeof(struct mlx5r_umr_wqe) -
233 				    sizeof(struct mlx5_wqe_data_seg);
234 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
235 	struct mlx5_core_dev *mdev = dev->mdev;
236 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
237 	struct mlx5_wqe_ctrl_seg *ctrl;
238 	union {
239 		struct ib_cqe *ib_cqe;
240 		u64 wr_id;
241 	} id;
242 	void *cur_edge, *seg;
243 	unsigned long flags;
244 	unsigned int idx;
245 	int size, err;
246 
247 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR))
248 		return -EIO;
249 
250 	spin_lock_irqsave(&qp->sq.lock, flags);
251 
252 	err = mlx5r_begin_wqe(qp, &seg, &ctrl, &idx, &size, &cur_edge, 0,
253 			      cpu_to_be32(mkey), false, false);
254 	if (WARN_ON(err))
255 		goto out;
256 
257 	qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
258 
259 	mlx5r_memcpy_send_wqe(&qp->sq, &cur_edge, &seg, &size, wqe, wqe_size);
260 
261 	id.ib_cqe = cqe;
262 	mlx5r_finish_wqe(qp, ctrl, seg, size, cur_edge, idx, id.wr_id, 0,
263 			 MLX5_FENCE_MODE_INITIATOR_SMALL, MLX5_OPCODE_UMR);
264 
265 	mlx5r_ring_db(qp, 1, ctrl);
266 
267 out:
268 	spin_unlock_irqrestore(&qp->sq.lock, flags);
269 
270 	return err;
271 }
272 
273 static void mlx5r_umr_done(struct ib_cq *cq, struct ib_wc *wc)
274 {
275 	struct mlx5_ib_umr_context *context =
276 		container_of(wc->wr_cqe, struct mlx5_ib_umr_context, cqe);
277 
278 	context->status = wc->status;
279 	complete(&context->done);
280 }
281 
282 static inline void mlx5r_umr_init_context(struct mlx5r_umr_context *context)
283 {
284 	context->cqe.done = mlx5r_umr_done;
285 	init_completion(&context->done);
286 }
287 
288 static int mlx5r_umr_post_send_wait(struct mlx5_ib_dev *dev, u32 mkey,
289 				   struct mlx5r_umr_wqe *wqe, bool with_data)
290 {
291 	struct umr_common *umrc = &dev->umrc;
292 	struct mlx5r_umr_context umr_context;
293 	int err;
294 
295 	err = umr_check_mkey_mask(dev, be64_to_cpu(wqe->ctrl_seg.mkey_mask));
296 	if (WARN_ON(err))
297 		return err;
298 
299 	mlx5r_umr_init_context(&umr_context);
300 
301 	down(&umrc->sem);
302 	while (true) {
303 		mutex_lock(&umrc->lock);
304 		if (umrc->state == MLX5_UMR_STATE_ERR) {
305 			mutex_unlock(&umrc->lock);
306 			err = -EFAULT;
307 			break;
308 		}
309 
310 		if (umrc->state == MLX5_UMR_STATE_RECOVER) {
311 			mutex_unlock(&umrc->lock);
312 			usleep_range(3000, 5000);
313 			continue;
314 		}
315 
316 		err = mlx5r_umr_post_send(umrc->qp, mkey, &umr_context.cqe, wqe,
317 					  with_data);
318 		mutex_unlock(&umrc->lock);
319 		if (err) {
320 			mlx5_ib_warn(dev, "UMR post send failed, err %d\n",
321 				     err);
322 			break;
323 		}
324 
325 		wait_for_completion(&umr_context.done);
326 
327 		if (umr_context.status == IB_WC_SUCCESS)
328 			break;
329 
330 		if (umr_context.status == IB_WC_WR_FLUSH_ERR)
331 			continue;
332 
333 		WARN_ON_ONCE(1);
334 		mlx5_ib_warn(dev,
335 			"reg umr failed (%u). Trying to recover and resubmit the flushed WQEs\n",
336 			umr_context.status);
337 		mutex_lock(&umrc->lock);
338 		err = mlx5r_umr_recover(dev);
339 		mutex_unlock(&umrc->lock);
340 		if (err)
341 			mlx5_ib_warn(dev, "couldn't recover UMR, err %d\n",
342 				     err);
343 		err = -EFAULT;
344 		break;
345 	}
346 	up(&umrc->sem);
347 	return err;
348 }
349 
350 /**
351  * mlx5r_umr_revoke_mr - Fence all DMA on the MR
352  * @mr: The MR to fence
353  *
354  * Upon return the NIC will not be doing any DMA to the pages under the MR,
355  * and any DMA in progress will be completed. Failure of this function
356  * indicates the HW has failed catastrophically.
357  */
358 int mlx5r_umr_revoke_mr(struct mlx5_ib_mr *mr)
359 {
360 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
361 	struct mlx5r_umr_wqe wqe = {};
362 
363 	if (dev->mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)
364 		return 0;
365 
366 	wqe.ctrl_seg.mkey_mask |= get_umr_update_pd_mask();
367 	wqe.ctrl_seg.mkey_mask |= get_umr_disable_mr_mask();
368 	wqe.ctrl_seg.flags |= MLX5_UMR_INLINE;
369 
370 	MLX5_SET(mkc, &wqe.mkey_seg, free, 1);
371 	MLX5_SET(mkc, &wqe.mkey_seg, pd, to_mpd(dev->umrc.pd)->pdn);
372 	MLX5_SET(mkc, &wqe.mkey_seg, qpn, 0xffffff);
373 	MLX5_SET(mkc, &wqe.mkey_seg, mkey_7_0,
374 		 mlx5_mkey_variant(mr->mmkey.key));
375 
376 	return mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, false);
377 }
378 
379 static void mlx5r_umr_set_access_flags(struct mlx5_ib_dev *dev,
380 				       struct mlx5_mkey_seg *seg,
381 				       unsigned int access_flags)
382 {
383 	MLX5_SET(mkc, seg, a, !!(access_flags & IB_ACCESS_REMOTE_ATOMIC));
384 	MLX5_SET(mkc, seg, rw, !!(access_flags & IB_ACCESS_REMOTE_WRITE));
385 	MLX5_SET(mkc, seg, rr, !!(access_flags & IB_ACCESS_REMOTE_READ));
386 	MLX5_SET(mkc, seg, lw, !!(access_flags & IB_ACCESS_LOCAL_WRITE));
387 	MLX5_SET(mkc, seg, lr, 1);
388 	MLX5_SET(mkc, seg, relaxed_ordering_write,
389 		 !!(access_flags & IB_ACCESS_RELAXED_ORDERING));
390 	MLX5_SET(mkc, seg, relaxed_ordering_read,
391 		 !!(access_flags & IB_ACCESS_RELAXED_ORDERING));
392 }
393 
394 int mlx5r_umr_rereg_pd_access(struct mlx5_ib_mr *mr, struct ib_pd *pd,
395 			      int access_flags)
396 {
397 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
398 	struct mlx5r_umr_wqe wqe = {};
399 	int err;
400 
401 	wqe.ctrl_seg.mkey_mask = get_umr_update_access_mask(dev);
402 	wqe.ctrl_seg.mkey_mask |= get_umr_update_pd_mask();
403 	wqe.ctrl_seg.flags = MLX5_UMR_CHECK_FREE;
404 	wqe.ctrl_seg.flags |= MLX5_UMR_INLINE;
405 
406 	mlx5r_umr_set_access_flags(dev, &wqe.mkey_seg, access_flags);
407 	MLX5_SET(mkc, &wqe.mkey_seg, pd, to_mpd(pd)->pdn);
408 	MLX5_SET(mkc, &wqe.mkey_seg, qpn, 0xffffff);
409 	MLX5_SET(mkc, &wqe.mkey_seg, mkey_7_0,
410 		 mlx5_mkey_variant(mr->mmkey.key));
411 
412 	err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, false);
413 	if (err)
414 		return err;
415 
416 	mr->access_flags = access_flags;
417 	return 0;
418 }
419 
420 #define MLX5_MAX_UMR_CHUNK                                                     \
421 	((1 << (MLX5_MAX_UMR_SHIFT + 4)) - MLX5_UMR_MTT_ALIGNMENT)
422 #define MLX5_SPARE_UMR_CHUNK 0x10000
423 
424 /*
425  * Allocate a temporary buffer to hold the per-page information to transfer to
426  * HW. For efficiency this should be as large as it can be, but buffer
427  * allocation failure is not allowed, so try smaller sizes.
428  */
429 static void *mlx5r_umr_alloc_xlt(size_t *nents, size_t ent_size, gfp_t gfp_mask)
430 {
431 	const size_t xlt_chunk_align = MLX5_UMR_MTT_ALIGNMENT / ent_size;
432 	size_t size;
433 	void *res = NULL;
434 
435 	static_assert(PAGE_SIZE % MLX5_UMR_MTT_ALIGNMENT == 0);
436 
437 	/*
438 	 * MLX5_IB_UPD_XLT_ATOMIC doesn't signal an atomic context just that the
439 	 * allocation can't trigger any kind of reclaim.
440 	 */
441 	might_sleep();
442 
443 	gfp_mask |= __GFP_ZERO | __GFP_NORETRY;
444 
445 	/*
446 	 * If the system already has a suitable high order page then just use
447 	 * that, but don't try hard to create one. This max is about 1M, so a
448 	 * free x86 huge page will satisfy it.
449 	 */
450 	size = min_t(size_t, ent_size * ALIGN(*nents, xlt_chunk_align),
451 		     MLX5_MAX_UMR_CHUNK);
452 	*nents = size / ent_size;
453 	res = (void *)__get_free_pages(gfp_mask | __GFP_NOWARN,
454 				       get_order(size));
455 	if (res)
456 		return res;
457 
458 	if (size > MLX5_SPARE_UMR_CHUNK) {
459 		size = MLX5_SPARE_UMR_CHUNK;
460 		*nents = size / ent_size;
461 		res = (void *)__get_free_pages(gfp_mask | __GFP_NOWARN,
462 					       get_order(size));
463 		if (res)
464 			return res;
465 	}
466 
467 	*nents = PAGE_SIZE / ent_size;
468 	res = (void *)__get_free_page(gfp_mask);
469 	if (res)
470 		return res;
471 
472 	mutex_lock(&xlt_emergency_page_mutex);
473 	memset(xlt_emergency_page, 0, PAGE_SIZE);
474 	return xlt_emergency_page;
475 }
476 
477 static void mlx5r_umr_free_xlt(void *xlt, size_t length)
478 {
479 	if (xlt == xlt_emergency_page) {
480 		mutex_unlock(&xlt_emergency_page_mutex);
481 		return;
482 	}
483 
484 	free_pages((unsigned long)xlt, get_order(length));
485 }
486 
487 static void mlx5r_umr_unmap_free_xlt(struct mlx5_ib_dev *dev, void *xlt,
488 				     struct ib_sge *sg)
489 {
490 	struct device *ddev = &dev->mdev->pdev->dev;
491 
492 	dma_unmap_single(ddev, sg->addr, sg->length, DMA_TO_DEVICE);
493 	mlx5r_umr_free_xlt(xlt, sg->length);
494 }
495 
496 /*
497  * Create an XLT buffer ready for submission.
498  */
499 static void *mlx5r_umr_create_xlt(struct mlx5_ib_dev *dev, struct ib_sge *sg,
500 				  size_t nents, size_t ent_size,
501 				  unsigned int flags)
502 {
503 	struct device *ddev = &dev->mdev->pdev->dev;
504 	dma_addr_t dma;
505 	void *xlt;
506 
507 	xlt = mlx5r_umr_alloc_xlt(&nents, ent_size,
508 				 flags & MLX5_IB_UPD_XLT_ATOMIC ? GFP_ATOMIC :
509 								  GFP_KERNEL);
510 	sg->length = nents * ent_size;
511 	dma = dma_map_single(ddev, xlt, sg->length, DMA_TO_DEVICE);
512 	if (dma_mapping_error(ddev, dma)) {
513 		mlx5_ib_err(dev, "unable to map DMA during XLT update.\n");
514 		mlx5r_umr_free_xlt(xlt, sg->length);
515 		return NULL;
516 	}
517 	sg->addr = dma;
518 	sg->lkey = dev->umrc.pd->local_dma_lkey;
519 
520 	return xlt;
521 }
522 
523 static void
524 mlx5r_umr_set_update_xlt_ctrl_seg(struct mlx5_wqe_umr_ctrl_seg *ctrl_seg,
525 				  unsigned int flags, struct ib_sge *sg)
526 {
527 	if (!(flags & MLX5_IB_UPD_XLT_ENABLE))
528 		/* fail if free */
529 		ctrl_seg->flags = MLX5_UMR_CHECK_FREE;
530 	else
531 		/* fail if not free */
532 		ctrl_seg->flags = MLX5_UMR_CHECK_NOT_FREE;
533 	ctrl_seg->xlt_octowords =
534 		cpu_to_be16(mlx5r_umr_get_xlt_octo(sg->length));
535 }
536 
537 static void mlx5r_umr_set_update_xlt_mkey_seg(struct mlx5_ib_dev *dev,
538 					      struct mlx5_mkey_seg *mkey_seg,
539 					      struct mlx5_ib_mr *mr,
540 					      unsigned int page_shift)
541 {
542 	mlx5r_umr_set_access_flags(dev, mkey_seg, mr->access_flags);
543 	MLX5_SET(mkc, mkey_seg, pd, to_mpd(mr->ibmr.pd)->pdn);
544 	MLX5_SET64(mkc, mkey_seg, start_addr, mr->ibmr.iova);
545 	MLX5_SET64(mkc, mkey_seg, len, mr->ibmr.length);
546 	MLX5_SET(mkc, mkey_seg, log_page_size, page_shift);
547 	MLX5_SET(mkc, mkey_seg, qpn, 0xffffff);
548 	MLX5_SET(mkc, mkey_seg, mkey_7_0, mlx5_mkey_variant(mr->mmkey.key));
549 }
550 
551 static void
552 mlx5r_umr_set_update_xlt_data_seg(struct mlx5_wqe_data_seg *data_seg,
553 				  struct ib_sge *sg)
554 {
555 	data_seg->byte_count = cpu_to_be32(sg->length);
556 	data_seg->lkey = cpu_to_be32(sg->lkey);
557 	data_seg->addr = cpu_to_be64(sg->addr);
558 }
559 
560 static void mlx5r_umr_update_offset(struct mlx5_wqe_umr_ctrl_seg *ctrl_seg,
561 				    u64 offset)
562 {
563 	u64 octo_offset = mlx5r_umr_get_xlt_octo(offset);
564 
565 	ctrl_seg->xlt_offset = cpu_to_be16(octo_offset & 0xffff);
566 	ctrl_seg->xlt_offset_47_16 = cpu_to_be32(octo_offset >> 16);
567 	ctrl_seg->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
568 }
569 
570 static void mlx5r_umr_final_update_xlt(struct mlx5_ib_dev *dev,
571 				       struct mlx5r_umr_wqe *wqe,
572 				       struct mlx5_ib_mr *mr, struct ib_sge *sg,
573 				       unsigned int flags)
574 {
575 	bool update_pd_access, update_translation;
576 
577 	if (flags & MLX5_IB_UPD_XLT_ENABLE)
578 		wqe->ctrl_seg.mkey_mask |= get_umr_enable_mr_mask();
579 
580 	update_pd_access = flags & MLX5_IB_UPD_XLT_ENABLE ||
581 			   flags & MLX5_IB_UPD_XLT_PD ||
582 			   flags & MLX5_IB_UPD_XLT_ACCESS;
583 
584 	if (update_pd_access) {
585 		wqe->ctrl_seg.mkey_mask |= get_umr_update_access_mask(dev);
586 		wqe->ctrl_seg.mkey_mask |= get_umr_update_pd_mask();
587 	}
588 
589 	update_translation =
590 		flags & MLX5_IB_UPD_XLT_ENABLE || flags & MLX5_IB_UPD_XLT_ADDR;
591 
592 	if (update_translation) {
593 		wqe->ctrl_seg.mkey_mask |= get_umr_update_translation_mask();
594 		if (!mr->ibmr.length)
595 			MLX5_SET(mkc, &wqe->mkey_seg, length64, 1);
596 	}
597 
598 	wqe->ctrl_seg.xlt_octowords =
599 		cpu_to_be16(mlx5r_umr_get_xlt_octo(sg->length));
600 	wqe->data_seg.byte_count = cpu_to_be32(sg->length);
601 }
602 
603 /*
604  * Send the DMA list to the HW for a normal MR using UMR.
605  * Dmabuf MR is handled in a similar way, except that the MLX5_IB_UPD_XLT_ZAP
606  * flag may be used.
607  */
608 int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
609 {
610 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
611 	struct device *ddev = &dev->mdev->pdev->dev;
612 	struct mlx5r_umr_wqe wqe = {};
613 	struct ib_block_iter biter;
614 	struct mlx5_mtt *cur_mtt;
615 	size_t orig_sg_length;
616 	struct mlx5_mtt *mtt;
617 	size_t final_size;
618 	struct ib_sge sg;
619 	u64 offset = 0;
620 	int err = 0;
621 
622 	if (WARN_ON(mr->umem->is_odp))
623 		return -EINVAL;
624 
625 	mtt = mlx5r_umr_create_xlt(
626 		dev, &sg, ib_umem_num_dma_blocks(mr->umem, 1 << mr->page_shift),
627 		sizeof(*mtt), flags);
628 	if (!mtt)
629 		return -ENOMEM;
630 
631 	orig_sg_length = sg.length;
632 
633 	mlx5r_umr_set_update_xlt_ctrl_seg(&wqe.ctrl_seg, flags, &sg);
634 	mlx5r_umr_set_update_xlt_mkey_seg(dev, &wqe.mkey_seg, mr,
635 					  mr->page_shift);
636 	mlx5r_umr_set_update_xlt_data_seg(&wqe.data_seg, &sg);
637 
638 	cur_mtt = mtt;
639 	rdma_for_each_block(mr->umem->sgt_append.sgt.sgl, &biter,
640 			    mr->umem->sgt_append.sgt.nents,
641 			    BIT(mr->page_shift)) {
642 		if (cur_mtt == (void *)mtt + sg.length) {
643 			dma_sync_single_for_device(ddev, sg.addr, sg.length,
644 						   DMA_TO_DEVICE);
645 
646 			err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe,
647 						       true);
648 			if (err)
649 				goto err;
650 			dma_sync_single_for_cpu(ddev, sg.addr, sg.length,
651 						DMA_TO_DEVICE);
652 			offset += sg.length;
653 			mlx5r_umr_update_offset(&wqe.ctrl_seg, offset);
654 
655 			cur_mtt = mtt;
656 		}
657 
658 		cur_mtt->ptag =
659 			cpu_to_be64(rdma_block_iter_dma_address(&biter) |
660 				    MLX5_IB_MTT_PRESENT);
661 
662 		if (mr->umem->is_dmabuf && (flags & MLX5_IB_UPD_XLT_ZAP))
663 			cur_mtt->ptag = 0;
664 
665 		cur_mtt++;
666 	}
667 
668 	final_size = (void *)cur_mtt - (void *)mtt;
669 	sg.length = ALIGN(final_size, MLX5_UMR_MTT_ALIGNMENT);
670 	memset(cur_mtt, 0, sg.length - final_size);
671 	mlx5r_umr_final_update_xlt(dev, &wqe, mr, &sg, flags);
672 
673 	dma_sync_single_for_device(ddev, sg.addr, sg.length, DMA_TO_DEVICE);
674 	err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, true);
675 
676 err:
677 	sg.length = orig_sg_length;
678 	mlx5r_umr_unmap_free_xlt(dev, mtt, &sg);
679 	return err;
680 }
681 
682 static bool umr_can_use_indirect_mkey(struct mlx5_ib_dev *dev)
683 {
684 	return !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled);
685 }
686 
687 int mlx5r_umr_update_xlt(struct mlx5_ib_mr *mr, u64 idx, int npages,
688 			 int page_shift, int flags)
689 {
690 	int desc_size = (flags & MLX5_IB_UPD_XLT_INDIRECT)
691 			       ? sizeof(struct mlx5_klm)
692 			       : sizeof(struct mlx5_mtt);
693 	const int page_align = MLX5_UMR_MTT_ALIGNMENT / desc_size;
694 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
695 	struct device *ddev = &dev->mdev->pdev->dev;
696 	const int page_mask = page_align - 1;
697 	struct mlx5r_umr_wqe wqe = {};
698 	size_t pages_mapped = 0;
699 	size_t pages_to_map = 0;
700 	size_t size_to_map = 0;
701 	size_t orig_sg_length;
702 	size_t pages_iter;
703 	struct ib_sge sg;
704 	int err = 0;
705 	void *xlt;
706 
707 	if ((flags & MLX5_IB_UPD_XLT_INDIRECT) &&
708 	    !umr_can_use_indirect_mkey(dev))
709 		return -EPERM;
710 
711 	if (WARN_ON(!mr->umem->is_odp))
712 		return -EINVAL;
713 
714 	/* UMR copies MTTs in units of MLX5_UMR_MTT_ALIGNMENT bytes,
715 	 * so we need to align the offset and length accordingly
716 	 */
717 	if (idx & page_mask) {
718 		npages += idx & page_mask;
719 		idx &= ~page_mask;
720 	}
721 	pages_to_map = ALIGN(npages, page_align);
722 
723 	xlt = mlx5r_umr_create_xlt(dev, &sg, npages, desc_size, flags);
724 	if (!xlt)
725 		return -ENOMEM;
726 
727 	pages_iter = sg.length / desc_size;
728 	orig_sg_length = sg.length;
729 
730 	if (!(flags & MLX5_IB_UPD_XLT_INDIRECT)) {
731 		struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
732 		size_t max_pages = ib_umem_odp_num_pages(odp) - idx;
733 
734 		pages_to_map = min_t(size_t, pages_to_map, max_pages);
735 	}
736 
737 	mlx5r_umr_set_update_xlt_ctrl_seg(&wqe.ctrl_seg, flags, &sg);
738 	mlx5r_umr_set_update_xlt_mkey_seg(dev, &wqe.mkey_seg, mr, page_shift);
739 	mlx5r_umr_set_update_xlt_data_seg(&wqe.data_seg, &sg);
740 
741 	for (pages_mapped = 0;
742 	     pages_mapped < pages_to_map && !err;
743 	     pages_mapped += pages_iter, idx += pages_iter) {
744 		npages = min_t(int, pages_iter, pages_to_map - pages_mapped);
745 		size_to_map = npages * desc_size;
746 		dma_sync_single_for_cpu(ddev, sg.addr, sg.length,
747 					DMA_TO_DEVICE);
748 		mlx5_odp_populate_xlt(xlt, idx, npages, mr, flags);
749 		dma_sync_single_for_device(ddev, sg.addr, sg.length,
750 					   DMA_TO_DEVICE);
751 		sg.length = ALIGN(size_to_map, MLX5_UMR_MTT_ALIGNMENT);
752 
753 		if (pages_mapped + pages_iter >= pages_to_map)
754 			mlx5r_umr_final_update_xlt(dev, &wqe, mr, &sg, flags);
755 		mlx5r_umr_update_offset(&wqe.ctrl_seg, idx * desc_size);
756 		err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, true);
757 	}
758 	sg.length = orig_sg_length;
759 	mlx5r_umr_unmap_free_xlt(dev, xlt, &sg);
760 	return err;
761 }
762