xref: /openbmc/linux/drivers/infiniband/hw/mlx5/umr.c (revision 6bfb56e9)
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 
180 	return 0;
181 
182 destroy_qp:
183 	ib_destroy_qp(qp);
184 destroy_cq:
185 	ib_free_cq(cq);
186 destroy_pd:
187 	ib_dealloc_pd(pd);
188 	return ret;
189 }
190 
191 void mlx5r_umr_resource_cleanup(struct mlx5_ib_dev *dev)
192 {
193 	ib_destroy_qp(dev->umrc.qp);
194 	ib_free_cq(dev->umrc.cq);
195 	ib_dealloc_pd(dev->umrc.pd);
196 }
197 
198 static int mlx5r_umr_post_send(struct ib_qp *ibqp, u32 mkey, struct ib_cqe *cqe,
199 			       struct mlx5r_umr_wqe *wqe, bool with_data)
200 {
201 	unsigned int wqe_size =
202 		with_data ? sizeof(struct mlx5r_umr_wqe) :
203 			    sizeof(struct mlx5r_umr_wqe) -
204 				    sizeof(struct mlx5_wqe_data_seg);
205 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
206 	struct mlx5_core_dev *mdev = dev->mdev;
207 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
208 	struct mlx5_wqe_ctrl_seg *ctrl;
209 	union {
210 		struct ib_cqe *ib_cqe;
211 		u64 wr_id;
212 	} id;
213 	void *cur_edge, *seg;
214 	unsigned long flags;
215 	unsigned int idx;
216 	int size, err;
217 
218 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR))
219 		return -EIO;
220 
221 	spin_lock_irqsave(&qp->sq.lock, flags);
222 
223 	err = mlx5r_begin_wqe(qp, &seg, &ctrl, &idx, &size, &cur_edge, 0,
224 			      cpu_to_be32(mkey), false, false);
225 	if (WARN_ON(err))
226 		goto out;
227 
228 	qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
229 
230 	mlx5r_memcpy_send_wqe(&qp->sq, &cur_edge, &seg, &size, wqe, wqe_size);
231 
232 	id.ib_cqe = cqe;
233 	mlx5r_finish_wqe(qp, ctrl, seg, size, cur_edge, idx, id.wr_id, 0,
234 			 MLX5_FENCE_MODE_NONE, MLX5_OPCODE_UMR);
235 
236 	mlx5r_ring_db(qp, 1, ctrl);
237 
238 out:
239 	spin_unlock_irqrestore(&qp->sq.lock, flags);
240 
241 	return err;
242 }
243 
244 static void mlx5r_umr_done(struct ib_cq *cq, struct ib_wc *wc)
245 {
246 	struct mlx5_ib_umr_context *context =
247 		container_of(wc->wr_cqe, struct mlx5_ib_umr_context, cqe);
248 
249 	context->status = wc->status;
250 	complete(&context->done);
251 }
252 
253 static inline void mlx5r_umr_init_context(struct mlx5r_umr_context *context)
254 {
255 	context->cqe.done = mlx5r_umr_done;
256 	init_completion(&context->done);
257 }
258 
259 static int mlx5r_umr_post_send_wait(struct mlx5_ib_dev *dev, u32 mkey,
260 				   struct mlx5r_umr_wqe *wqe, bool with_data)
261 {
262 	struct umr_common *umrc = &dev->umrc;
263 	struct mlx5r_umr_context umr_context;
264 	int err;
265 
266 	err = umr_check_mkey_mask(dev, be64_to_cpu(wqe->ctrl_seg.mkey_mask));
267 	if (WARN_ON(err))
268 		return err;
269 
270 	mlx5r_umr_init_context(&umr_context);
271 
272 	down(&umrc->sem);
273 	err = mlx5r_umr_post_send(umrc->qp, mkey, &umr_context.cqe, wqe,
274 				  with_data);
275 	if (err)
276 		mlx5_ib_warn(dev, "UMR post send failed, err %d\n", err);
277 	else {
278 		wait_for_completion(&umr_context.done);
279 		if (umr_context.status != IB_WC_SUCCESS) {
280 			mlx5_ib_warn(dev, "reg umr failed (%u)\n",
281 				     umr_context.status);
282 			err = -EFAULT;
283 		}
284 	}
285 	up(&umrc->sem);
286 	return err;
287 }
288 
289 /**
290  * mlx5r_umr_revoke_mr - Fence all DMA on the MR
291  * @mr: The MR to fence
292  *
293  * Upon return the NIC will not be doing any DMA to the pages under the MR,
294  * and any DMA in progress will be completed. Failure of this function
295  * indicates the HW has failed catastrophically.
296  */
297 int mlx5r_umr_revoke_mr(struct mlx5_ib_mr *mr)
298 {
299 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
300 	struct mlx5r_umr_wqe wqe = {};
301 
302 	if (dev->mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)
303 		return 0;
304 
305 	wqe.ctrl_seg.mkey_mask |= get_umr_update_pd_mask();
306 	wqe.ctrl_seg.mkey_mask |= get_umr_disable_mr_mask();
307 	wqe.ctrl_seg.flags |= MLX5_UMR_INLINE;
308 
309 	MLX5_SET(mkc, &wqe.mkey_seg, free, 1);
310 	MLX5_SET(mkc, &wqe.mkey_seg, pd, to_mpd(dev->umrc.pd)->pdn);
311 	MLX5_SET(mkc, &wqe.mkey_seg, qpn, 0xffffff);
312 	MLX5_SET(mkc, &wqe.mkey_seg, mkey_7_0,
313 		 mlx5_mkey_variant(mr->mmkey.key));
314 
315 	return mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, false);
316 }
317 
318 static void mlx5r_umr_set_access_flags(struct mlx5_ib_dev *dev,
319 				       struct mlx5_mkey_seg *seg,
320 				       unsigned int access_flags)
321 {
322 	MLX5_SET(mkc, seg, a, !!(access_flags & IB_ACCESS_REMOTE_ATOMIC));
323 	MLX5_SET(mkc, seg, rw, !!(access_flags & IB_ACCESS_REMOTE_WRITE));
324 	MLX5_SET(mkc, seg, rr, !!(access_flags & IB_ACCESS_REMOTE_READ));
325 	MLX5_SET(mkc, seg, lw, !!(access_flags & IB_ACCESS_LOCAL_WRITE));
326 	MLX5_SET(mkc, seg, lr, 1);
327 	MLX5_SET(mkc, seg, relaxed_ordering_write,
328 		 !!(access_flags & IB_ACCESS_RELAXED_ORDERING));
329 	MLX5_SET(mkc, seg, relaxed_ordering_read,
330 		 !!(access_flags & IB_ACCESS_RELAXED_ORDERING));
331 }
332 
333 int mlx5r_umr_rereg_pd_access(struct mlx5_ib_mr *mr, struct ib_pd *pd,
334 			      int access_flags)
335 {
336 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
337 	struct mlx5r_umr_wqe wqe = {};
338 	int err;
339 
340 	wqe.ctrl_seg.mkey_mask = get_umr_update_access_mask(dev);
341 	wqe.ctrl_seg.mkey_mask |= get_umr_update_pd_mask();
342 	wqe.ctrl_seg.flags = MLX5_UMR_CHECK_FREE;
343 	wqe.ctrl_seg.flags |= MLX5_UMR_INLINE;
344 
345 	mlx5r_umr_set_access_flags(dev, &wqe.mkey_seg, access_flags);
346 	MLX5_SET(mkc, &wqe.mkey_seg, pd, to_mpd(pd)->pdn);
347 	MLX5_SET(mkc, &wqe.mkey_seg, qpn, 0xffffff);
348 	MLX5_SET(mkc, &wqe.mkey_seg, mkey_7_0,
349 		 mlx5_mkey_variant(mr->mmkey.key));
350 
351 	err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, false);
352 	if (err)
353 		return err;
354 
355 	mr->access_flags = access_flags;
356 	return 0;
357 }
358 
359 #define MLX5_MAX_UMR_CHUNK                                                     \
360 	((1 << (MLX5_MAX_UMR_SHIFT + 4)) - MLX5_UMR_MTT_ALIGNMENT)
361 #define MLX5_SPARE_UMR_CHUNK 0x10000
362 
363 /*
364  * Allocate a temporary buffer to hold the per-page information to transfer to
365  * HW. For efficiency this should be as large as it can be, but buffer
366  * allocation failure is not allowed, so try smaller sizes.
367  */
368 static void *mlx5r_umr_alloc_xlt(size_t *nents, size_t ent_size, gfp_t gfp_mask)
369 {
370 	const size_t xlt_chunk_align = MLX5_UMR_MTT_ALIGNMENT / ent_size;
371 	size_t size;
372 	void *res = NULL;
373 
374 	static_assert(PAGE_SIZE % MLX5_UMR_MTT_ALIGNMENT == 0);
375 
376 	/*
377 	 * MLX5_IB_UPD_XLT_ATOMIC doesn't signal an atomic context just that the
378 	 * allocation can't trigger any kind of reclaim.
379 	 */
380 	might_sleep();
381 
382 	gfp_mask |= __GFP_ZERO | __GFP_NORETRY;
383 
384 	/*
385 	 * If the system already has a suitable high order page then just use
386 	 * that, but don't try hard to create one. This max is about 1M, so a
387 	 * free x86 huge page will satisfy it.
388 	 */
389 	size = min_t(size_t, ent_size * ALIGN(*nents, xlt_chunk_align),
390 		     MLX5_MAX_UMR_CHUNK);
391 	*nents = size / ent_size;
392 	res = (void *)__get_free_pages(gfp_mask | __GFP_NOWARN,
393 				       get_order(size));
394 	if (res)
395 		return res;
396 
397 	if (size > MLX5_SPARE_UMR_CHUNK) {
398 		size = MLX5_SPARE_UMR_CHUNK;
399 		*nents = size / ent_size;
400 		res = (void *)__get_free_pages(gfp_mask | __GFP_NOWARN,
401 					       get_order(size));
402 		if (res)
403 			return res;
404 	}
405 
406 	*nents = PAGE_SIZE / ent_size;
407 	res = (void *)__get_free_page(gfp_mask);
408 	if (res)
409 		return res;
410 
411 	mutex_lock(&xlt_emergency_page_mutex);
412 	memset(xlt_emergency_page, 0, PAGE_SIZE);
413 	return xlt_emergency_page;
414 }
415 
416 static void mlx5r_umr_free_xlt(void *xlt, size_t length)
417 {
418 	if (xlt == xlt_emergency_page) {
419 		mutex_unlock(&xlt_emergency_page_mutex);
420 		return;
421 	}
422 
423 	free_pages((unsigned long)xlt, get_order(length));
424 }
425 
426 static void mlx5r_umr_unmap_free_xlt(struct mlx5_ib_dev *dev, void *xlt,
427 				     struct ib_sge *sg)
428 {
429 	struct device *ddev = &dev->mdev->pdev->dev;
430 
431 	dma_unmap_single(ddev, sg->addr, sg->length, DMA_TO_DEVICE);
432 	mlx5r_umr_free_xlt(xlt, sg->length);
433 }
434 
435 /*
436  * Create an XLT buffer ready for submission.
437  */
438 static void *mlx5r_umr_create_xlt(struct mlx5_ib_dev *dev, struct ib_sge *sg,
439 				  size_t nents, size_t ent_size,
440 				  unsigned int flags)
441 {
442 	struct device *ddev = &dev->mdev->pdev->dev;
443 	dma_addr_t dma;
444 	void *xlt;
445 
446 	xlt = mlx5r_umr_alloc_xlt(&nents, ent_size,
447 				 flags & MLX5_IB_UPD_XLT_ATOMIC ? GFP_ATOMIC :
448 								  GFP_KERNEL);
449 	sg->length = nents * ent_size;
450 	dma = dma_map_single(ddev, xlt, sg->length, DMA_TO_DEVICE);
451 	if (dma_mapping_error(ddev, dma)) {
452 		mlx5_ib_err(dev, "unable to map DMA during XLT update.\n");
453 		mlx5r_umr_free_xlt(xlt, sg->length);
454 		return NULL;
455 	}
456 	sg->addr = dma;
457 	sg->lkey = dev->umrc.pd->local_dma_lkey;
458 
459 	return xlt;
460 }
461 
462 static void
463 mlx5r_umr_set_update_xlt_ctrl_seg(struct mlx5_wqe_umr_ctrl_seg *ctrl_seg,
464 				  unsigned int flags, struct ib_sge *sg)
465 {
466 	if (!(flags & MLX5_IB_UPD_XLT_ENABLE))
467 		/* fail if free */
468 		ctrl_seg->flags = MLX5_UMR_CHECK_FREE;
469 	else
470 		/* fail if not free */
471 		ctrl_seg->flags = MLX5_UMR_CHECK_NOT_FREE;
472 	ctrl_seg->xlt_octowords =
473 		cpu_to_be16(mlx5r_umr_get_xlt_octo(sg->length));
474 }
475 
476 static void mlx5r_umr_set_update_xlt_mkey_seg(struct mlx5_ib_dev *dev,
477 					      struct mlx5_mkey_seg *mkey_seg,
478 					      struct mlx5_ib_mr *mr,
479 					      unsigned int page_shift)
480 {
481 	mlx5r_umr_set_access_flags(dev, mkey_seg, mr->access_flags);
482 	MLX5_SET(mkc, mkey_seg, pd, to_mpd(mr->ibmr.pd)->pdn);
483 	MLX5_SET64(mkc, mkey_seg, start_addr, mr->ibmr.iova);
484 	MLX5_SET64(mkc, mkey_seg, len, mr->ibmr.length);
485 	MLX5_SET(mkc, mkey_seg, log_page_size, page_shift);
486 	MLX5_SET(mkc, mkey_seg, qpn, 0xffffff);
487 	MLX5_SET(mkc, mkey_seg, mkey_7_0, mlx5_mkey_variant(mr->mmkey.key));
488 }
489 
490 static void
491 mlx5r_umr_set_update_xlt_data_seg(struct mlx5_wqe_data_seg *data_seg,
492 				  struct ib_sge *sg)
493 {
494 	data_seg->byte_count = cpu_to_be32(sg->length);
495 	data_seg->lkey = cpu_to_be32(sg->lkey);
496 	data_seg->addr = cpu_to_be64(sg->addr);
497 }
498 
499 static void mlx5r_umr_update_offset(struct mlx5_wqe_umr_ctrl_seg *ctrl_seg,
500 				    u64 offset)
501 {
502 	u64 octo_offset = mlx5r_umr_get_xlt_octo(offset);
503 
504 	ctrl_seg->xlt_offset = cpu_to_be16(octo_offset & 0xffff);
505 	ctrl_seg->xlt_offset_47_16 = cpu_to_be32(octo_offset >> 16);
506 	ctrl_seg->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
507 }
508 
509 static void mlx5r_umr_final_update_xlt(struct mlx5_ib_dev *dev,
510 				       struct mlx5r_umr_wqe *wqe,
511 				       struct mlx5_ib_mr *mr, struct ib_sge *sg,
512 				       unsigned int flags)
513 {
514 	bool update_pd_access, update_translation;
515 
516 	if (flags & MLX5_IB_UPD_XLT_ENABLE)
517 		wqe->ctrl_seg.mkey_mask |= get_umr_enable_mr_mask();
518 
519 	update_pd_access = flags & MLX5_IB_UPD_XLT_ENABLE ||
520 			   flags & MLX5_IB_UPD_XLT_PD ||
521 			   flags & MLX5_IB_UPD_XLT_ACCESS;
522 
523 	if (update_pd_access) {
524 		wqe->ctrl_seg.mkey_mask |= get_umr_update_access_mask(dev);
525 		wqe->ctrl_seg.mkey_mask |= get_umr_update_pd_mask();
526 	}
527 
528 	update_translation =
529 		flags & MLX5_IB_UPD_XLT_ENABLE || flags & MLX5_IB_UPD_XLT_ADDR;
530 
531 	if (update_translation) {
532 		wqe->ctrl_seg.mkey_mask |= get_umr_update_translation_mask();
533 		if (!mr->ibmr.length)
534 			MLX5_SET(mkc, &wqe->mkey_seg, length64, 1);
535 	}
536 
537 	wqe->ctrl_seg.xlt_octowords =
538 		cpu_to_be16(mlx5r_umr_get_xlt_octo(sg->length));
539 	wqe->data_seg.byte_count = cpu_to_be32(sg->length);
540 }
541 
542 /*
543  * Send the DMA list to the HW for a normal MR using UMR.
544  * Dmabuf MR is handled in a similar way, except that the MLX5_IB_UPD_XLT_ZAP
545  * flag may be used.
546  */
547 int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
548 {
549 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
550 	struct device *ddev = &dev->mdev->pdev->dev;
551 	struct mlx5r_umr_wqe wqe = {};
552 	struct ib_block_iter biter;
553 	struct mlx5_mtt *cur_mtt;
554 	size_t orig_sg_length;
555 	struct mlx5_mtt *mtt;
556 	size_t final_size;
557 	struct ib_sge sg;
558 	u64 offset = 0;
559 	int err = 0;
560 
561 	if (WARN_ON(mr->umem->is_odp))
562 		return -EINVAL;
563 
564 	mtt = mlx5r_umr_create_xlt(
565 		dev, &sg, ib_umem_num_dma_blocks(mr->umem, 1 << mr->page_shift),
566 		sizeof(*mtt), flags);
567 	if (!mtt)
568 		return -ENOMEM;
569 
570 	orig_sg_length = sg.length;
571 
572 	mlx5r_umr_set_update_xlt_ctrl_seg(&wqe.ctrl_seg, flags, &sg);
573 	mlx5r_umr_set_update_xlt_mkey_seg(dev, &wqe.mkey_seg, mr,
574 					  mr->page_shift);
575 	mlx5r_umr_set_update_xlt_data_seg(&wqe.data_seg, &sg);
576 
577 	cur_mtt = mtt;
578 	rdma_for_each_block(mr->umem->sgt_append.sgt.sgl, &biter,
579 			    mr->umem->sgt_append.sgt.nents,
580 			    BIT(mr->page_shift)) {
581 		if (cur_mtt == (void *)mtt + sg.length) {
582 			dma_sync_single_for_device(ddev, sg.addr, sg.length,
583 						   DMA_TO_DEVICE);
584 
585 			err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe,
586 						       true);
587 			if (err)
588 				goto err;
589 			dma_sync_single_for_cpu(ddev, sg.addr, sg.length,
590 						DMA_TO_DEVICE);
591 			offset += sg.length;
592 			mlx5r_umr_update_offset(&wqe.ctrl_seg, offset);
593 
594 			cur_mtt = mtt;
595 		}
596 
597 		cur_mtt->ptag =
598 			cpu_to_be64(rdma_block_iter_dma_address(&biter) |
599 				    MLX5_IB_MTT_PRESENT);
600 
601 		if (mr->umem->is_dmabuf && (flags & MLX5_IB_UPD_XLT_ZAP))
602 			cur_mtt->ptag = 0;
603 
604 		cur_mtt++;
605 	}
606 
607 	final_size = (void *)cur_mtt - (void *)mtt;
608 	sg.length = ALIGN(final_size, MLX5_UMR_MTT_ALIGNMENT);
609 	memset(cur_mtt, 0, sg.length - final_size);
610 	mlx5r_umr_final_update_xlt(dev, &wqe, mr, &sg, flags);
611 
612 	dma_sync_single_for_device(ddev, sg.addr, sg.length, DMA_TO_DEVICE);
613 	err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, true);
614 
615 err:
616 	sg.length = orig_sg_length;
617 	mlx5r_umr_unmap_free_xlt(dev, mtt, &sg);
618 	return err;
619 }
620 
621 static bool umr_can_use_indirect_mkey(struct mlx5_ib_dev *dev)
622 {
623 	return !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled);
624 }
625 
626 int mlx5r_umr_update_xlt(struct mlx5_ib_mr *mr, u64 idx, int npages,
627 			 int page_shift, int flags)
628 {
629 	int desc_size = (flags & MLX5_IB_UPD_XLT_INDIRECT)
630 			       ? sizeof(struct mlx5_klm)
631 			       : sizeof(struct mlx5_mtt);
632 	const int page_align = MLX5_UMR_MTT_ALIGNMENT / desc_size;
633 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
634 	struct device *ddev = &dev->mdev->pdev->dev;
635 	const int page_mask = page_align - 1;
636 	struct mlx5r_umr_wqe wqe = {};
637 	size_t pages_mapped = 0;
638 	size_t pages_to_map = 0;
639 	size_t size_to_map = 0;
640 	size_t orig_sg_length;
641 	size_t pages_iter;
642 	struct ib_sge sg;
643 	int err = 0;
644 	void *xlt;
645 
646 	if ((flags & MLX5_IB_UPD_XLT_INDIRECT) &&
647 	    !umr_can_use_indirect_mkey(dev))
648 		return -EPERM;
649 
650 	if (WARN_ON(!mr->umem->is_odp))
651 		return -EINVAL;
652 
653 	/* UMR copies MTTs in units of MLX5_UMR_MTT_ALIGNMENT bytes,
654 	 * so we need to align the offset and length accordingly
655 	 */
656 	if (idx & page_mask) {
657 		npages += idx & page_mask;
658 		idx &= ~page_mask;
659 	}
660 	pages_to_map = ALIGN(npages, page_align);
661 
662 	xlt = mlx5r_umr_create_xlt(dev, &sg, npages, desc_size, flags);
663 	if (!xlt)
664 		return -ENOMEM;
665 
666 	pages_iter = sg.length / desc_size;
667 	orig_sg_length = sg.length;
668 
669 	if (!(flags & MLX5_IB_UPD_XLT_INDIRECT)) {
670 		struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
671 		size_t max_pages = ib_umem_odp_num_pages(odp) - idx;
672 
673 		pages_to_map = min_t(size_t, pages_to_map, max_pages);
674 	}
675 
676 	mlx5r_umr_set_update_xlt_ctrl_seg(&wqe.ctrl_seg, flags, &sg);
677 	mlx5r_umr_set_update_xlt_mkey_seg(dev, &wqe.mkey_seg, mr, page_shift);
678 	mlx5r_umr_set_update_xlt_data_seg(&wqe.data_seg, &sg);
679 
680 	for (pages_mapped = 0;
681 	     pages_mapped < pages_to_map && !err;
682 	     pages_mapped += pages_iter, idx += pages_iter) {
683 		npages = min_t(int, pages_iter, pages_to_map - pages_mapped);
684 		size_to_map = npages * desc_size;
685 		dma_sync_single_for_cpu(ddev, sg.addr, sg.length,
686 					DMA_TO_DEVICE);
687 		mlx5_odp_populate_xlt(xlt, idx, npages, mr, flags);
688 		dma_sync_single_for_device(ddev, sg.addr, sg.length,
689 					   DMA_TO_DEVICE);
690 		sg.length = ALIGN(size_to_map, MLX5_UMR_MTT_ALIGNMENT);
691 
692 		if (pages_mapped + pages_iter >= pages_to_map)
693 			mlx5r_umr_final_update_xlt(dev, &wqe, mr, &sg, flags);
694 		mlx5r_umr_update_offset(&wqe.ctrl_seg, idx * desc_size);
695 		err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, true);
696 	}
697 	sg.length = orig_sg_length;
698 	mlx5r_umr_unmap_free_xlt(dev, xlt, &sg);
699 	return err;
700 }
701