xref: /openbmc/linux/drivers/infiniband/hw/mlx5/odp.c (revision 405db98b)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <rdma/ib_umem.h>
34 #include <rdma/ib_umem_odp.h>
35 #include <linux/kernel.h>
36 #include <linux/dma-buf.h>
37 #include <linux/dma-resv.h>
38 
39 #include "mlx5_ib.h"
40 #include "cmd.h"
41 #include "qp.h"
42 
43 #include <linux/mlx5/eq.h>
44 
45 /* Contains the details of a pagefault. */
46 struct mlx5_pagefault {
47 	u32			bytes_committed;
48 	u32			token;
49 	u8			event_subtype;
50 	u8			type;
51 	union {
52 		/* Initiator or send message responder pagefault details. */
53 		struct {
54 			/* Received packet size, only valid for responders. */
55 			u32	packet_size;
56 			/*
57 			 * Number of resource holding WQE, depends on type.
58 			 */
59 			u32	wq_num;
60 			/*
61 			 * WQE index. Refers to either the send queue or
62 			 * receive queue, according to event_subtype.
63 			 */
64 			u16	wqe_index;
65 		} wqe;
66 		/* RDMA responder pagefault details */
67 		struct {
68 			u32	r_key;
69 			/*
70 			 * Received packet size, minimal size page fault
71 			 * resolution required for forward progress.
72 			 */
73 			u32	packet_size;
74 			u32	rdma_op_len;
75 			u64	rdma_va;
76 		} rdma;
77 	};
78 
79 	struct mlx5_ib_pf_eq	*eq;
80 	struct work_struct	work;
81 };
82 
83 #define MAX_PREFETCH_LEN (4*1024*1024U)
84 
85 /* Timeout in ms to wait for an active mmu notifier to complete when handling
86  * a pagefault. */
87 #define MMU_NOTIFIER_TIMEOUT 1000
88 
89 #define MLX5_IMR_MTT_BITS (30 - PAGE_SHIFT)
90 #define MLX5_IMR_MTT_SHIFT (MLX5_IMR_MTT_BITS + PAGE_SHIFT)
91 #define MLX5_IMR_MTT_ENTRIES BIT_ULL(MLX5_IMR_MTT_BITS)
92 #define MLX5_IMR_MTT_SIZE BIT_ULL(MLX5_IMR_MTT_SHIFT)
93 #define MLX5_IMR_MTT_MASK (~(MLX5_IMR_MTT_SIZE - 1))
94 
95 #define MLX5_KSM_PAGE_SHIFT MLX5_IMR_MTT_SHIFT
96 
97 static u64 mlx5_imr_ksm_entries;
98 
99 static void populate_klm(struct mlx5_klm *pklm, size_t idx, size_t nentries,
100 			struct mlx5_ib_mr *imr, int flags)
101 {
102 	struct mlx5_klm *end = pklm + nentries;
103 
104 	if (flags & MLX5_IB_UPD_XLT_ZAP) {
105 		for (; pklm != end; pklm++, idx++) {
106 			pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
107 			pklm->key = cpu_to_be32(mr_to_mdev(imr)->null_mkey);
108 			pklm->va = 0;
109 		}
110 		return;
111 	}
112 
113 	/*
114 	 * The locking here is pretty subtle. Ideally the implicit_children
115 	 * xarray would be protected by the umem_mutex, however that is not
116 	 * possible. Instead this uses a weaker update-then-lock pattern:
117 	 *
118 	 *    xa_store()
119 	 *    mutex_lock(umem_mutex)
120 	 *     mlx5_ib_update_xlt()
121 	 *    mutex_unlock(umem_mutex)
122 	 *    destroy lkey
123 	 *
124 	 * ie any change the xarray must be followed by the locked update_xlt
125 	 * before destroying.
126 	 *
127 	 * The umem_mutex provides the acquire/release semantic needed to make
128 	 * the xa_store() visible to a racing thread.
129 	 */
130 	lockdep_assert_held(&to_ib_umem_odp(imr->umem)->umem_mutex);
131 
132 	for (; pklm != end; pklm++, idx++) {
133 		struct mlx5_ib_mr *mtt = xa_load(&imr->implicit_children, idx);
134 
135 		pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
136 		if (mtt) {
137 			pklm->key = cpu_to_be32(mtt->ibmr.lkey);
138 			pklm->va = cpu_to_be64(idx * MLX5_IMR_MTT_SIZE);
139 		} else {
140 			pklm->key = cpu_to_be32(mr_to_mdev(imr)->null_mkey);
141 			pklm->va = 0;
142 		}
143 	}
144 }
145 
146 static u64 umem_dma_to_mtt(dma_addr_t umem_dma)
147 {
148 	u64 mtt_entry = umem_dma & ODP_DMA_ADDR_MASK;
149 
150 	if (umem_dma & ODP_READ_ALLOWED_BIT)
151 		mtt_entry |= MLX5_IB_MTT_READ;
152 	if (umem_dma & ODP_WRITE_ALLOWED_BIT)
153 		mtt_entry |= MLX5_IB_MTT_WRITE;
154 
155 	return mtt_entry;
156 }
157 
158 static void populate_mtt(__be64 *pas, size_t idx, size_t nentries,
159 			 struct mlx5_ib_mr *mr, int flags)
160 {
161 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
162 	dma_addr_t pa;
163 	size_t i;
164 
165 	if (flags & MLX5_IB_UPD_XLT_ZAP)
166 		return;
167 
168 	for (i = 0; i < nentries; i++) {
169 		pa = odp->dma_list[idx + i];
170 		pas[i] = cpu_to_be64(umem_dma_to_mtt(pa));
171 	}
172 }
173 
174 void mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
175 			   struct mlx5_ib_mr *mr, int flags)
176 {
177 	if (flags & MLX5_IB_UPD_XLT_INDIRECT) {
178 		populate_klm(xlt, idx, nentries, mr, flags);
179 	} else {
180 		populate_mtt(xlt, idx, nentries, mr, flags);
181 	}
182 }
183 
184 /*
185  * This must be called after the mr has been removed from implicit_children.
186  * NOTE: The MR does not necessarily have to be
187  * empty here, parallel page faults could have raced with the free process and
188  * added pages to it.
189  */
190 static void free_implicit_child_mr_work(struct work_struct *work)
191 {
192 	struct mlx5_ib_mr *mr =
193 		container_of(work, struct mlx5_ib_mr, odp_destroy.work);
194 	struct mlx5_ib_mr *imr = mr->parent;
195 	struct ib_umem_odp *odp_imr = to_ib_umem_odp(imr->umem);
196 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
197 
198 	mlx5r_deref_wait_odp_mkey(&mr->mmkey);
199 
200 	mutex_lock(&odp_imr->umem_mutex);
201 	mlx5_ib_update_xlt(mr->parent, ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT,
202 			   1, 0,
203 			   MLX5_IB_UPD_XLT_INDIRECT | MLX5_IB_UPD_XLT_ATOMIC);
204 	mutex_unlock(&odp_imr->umem_mutex);
205 	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
206 
207 	mlx5r_deref_odp_mkey(&imr->mmkey);
208 }
209 
210 static void destroy_unused_implicit_child_mr(struct mlx5_ib_mr *mr)
211 {
212 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
213 	unsigned long idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT;
214 	struct mlx5_ib_mr *imr = mr->parent;
215 
216 	if (!refcount_inc_not_zero(&imr->mmkey.usecount))
217 		return;
218 
219 	xa_erase(&imr->implicit_children, idx);
220 
221 	/* Freeing a MR is a sleeping operation, so bounce to a work queue */
222 	INIT_WORK(&mr->odp_destroy.work, free_implicit_child_mr_work);
223 	queue_work(system_unbound_wq, &mr->odp_destroy.work);
224 }
225 
226 static bool mlx5_ib_invalidate_range(struct mmu_interval_notifier *mni,
227 				     const struct mmu_notifier_range *range,
228 				     unsigned long cur_seq)
229 {
230 	struct ib_umem_odp *umem_odp =
231 		container_of(mni, struct ib_umem_odp, notifier);
232 	struct mlx5_ib_mr *mr;
233 	const u64 umr_block_mask = (MLX5_UMR_MTT_ALIGNMENT /
234 				    sizeof(struct mlx5_mtt)) - 1;
235 	u64 idx = 0, blk_start_idx = 0;
236 	u64 invalidations = 0;
237 	unsigned long start;
238 	unsigned long end;
239 	int in_block = 0;
240 	u64 addr;
241 
242 	if (!mmu_notifier_range_blockable(range))
243 		return false;
244 
245 	mutex_lock(&umem_odp->umem_mutex);
246 	mmu_interval_set_seq(mni, cur_seq);
247 	/*
248 	 * If npages is zero then umem_odp->private may not be setup yet. This
249 	 * does not complete until after the first page is mapped for DMA.
250 	 */
251 	if (!umem_odp->npages)
252 		goto out;
253 	mr = umem_odp->private;
254 
255 	start = max_t(u64, ib_umem_start(umem_odp), range->start);
256 	end = min_t(u64, ib_umem_end(umem_odp), range->end);
257 
258 	/*
259 	 * Iteration one - zap the HW's MTTs. The notifiers_count ensures that
260 	 * while we are doing the invalidation, no page fault will attempt to
261 	 * overwrite the same MTTs.  Concurent invalidations might race us,
262 	 * but they will write 0s as well, so no difference in the end result.
263 	 */
264 	for (addr = start; addr < end; addr += BIT(umem_odp->page_shift)) {
265 		idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
266 		/*
267 		 * Strive to write the MTTs in chunks, but avoid overwriting
268 		 * non-existing MTTs. The huristic here can be improved to
269 		 * estimate the cost of another UMR vs. the cost of bigger
270 		 * UMR.
271 		 */
272 		if (umem_odp->dma_list[idx] &
273 		    (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) {
274 			if (!in_block) {
275 				blk_start_idx = idx;
276 				in_block = 1;
277 			}
278 
279 			/* Count page invalidations */
280 			invalidations += idx - blk_start_idx + 1;
281 		} else {
282 			u64 umr_offset = idx & umr_block_mask;
283 
284 			if (in_block && umr_offset == 0) {
285 				mlx5_ib_update_xlt(mr, blk_start_idx,
286 						   idx - blk_start_idx, 0,
287 						   MLX5_IB_UPD_XLT_ZAP |
288 						   MLX5_IB_UPD_XLT_ATOMIC);
289 				in_block = 0;
290 			}
291 		}
292 	}
293 	if (in_block)
294 		mlx5_ib_update_xlt(mr, blk_start_idx,
295 				   idx - blk_start_idx + 1, 0,
296 				   MLX5_IB_UPD_XLT_ZAP |
297 				   MLX5_IB_UPD_XLT_ATOMIC);
298 
299 	mlx5_update_odp_stats(mr, invalidations, invalidations);
300 
301 	/*
302 	 * We are now sure that the device will not access the
303 	 * memory. We can safely unmap it, and mark it as dirty if
304 	 * needed.
305 	 */
306 
307 	ib_umem_odp_unmap_dma_pages(umem_odp, start, end);
308 
309 	if (unlikely(!umem_odp->npages && mr->parent))
310 		destroy_unused_implicit_child_mr(mr);
311 out:
312 	mutex_unlock(&umem_odp->umem_mutex);
313 	return true;
314 }
315 
316 const struct mmu_interval_notifier_ops mlx5_mn_ops = {
317 	.invalidate = mlx5_ib_invalidate_range,
318 };
319 
320 static void internal_fill_odp_caps(struct mlx5_ib_dev *dev)
321 {
322 	struct ib_odp_caps *caps = &dev->odp_caps;
323 
324 	memset(caps, 0, sizeof(*caps));
325 
326 	if (!MLX5_CAP_GEN(dev->mdev, pg) ||
327 	    !mlx5_ib_can_load_pas_with_umr(dev, 0))
328 		return;
329 
330 	caps->general_caps = IB_ODP_SUPPORT;
331 
332 	if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
333 		dev->odp_max_size = U64_MAX;
334 	else
335 		dev->odp_max_size = BIT_ULL(MLX5_MAX_UMR_SHIFT + PAGE_SHIFT);
336 
337 	if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.send))
338 		caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND;
339 
340 	if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.srq_receive))
341 		caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
342 
343 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.send))
344 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND;
345 
346 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.receive))
347 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV;
348 
349 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.write))
350 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE;
351 
352 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.read))
353 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ;
354 
355 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.atomic))
356 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
357 
358 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.srq_receive))
359 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
360 
361 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.send))
362 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SEND;
363 
364 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.receive))
365 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_RECV;
366 
367 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.write))
368 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_WRITE;
369 
370 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.read))
371 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_READ;
372 
373 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.atomic))
374 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
375 
376 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.srq_receive))
377 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
378 
379 	if (MLX5_CAP_GEN(dev->mdev, fixed_buffer_size) &&
380 	    MLX5_CAP_GEN(dev->mdev, null_mkey) &&
381 	    MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset) &&
382 	    !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled))
383 		caps->general_caps |= IB_ODP_SUPPORT_IMPLICIT;
384 }
385 
386 static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev,
387 				      struct mlx5_pagefault *pfault,
388 				      int error)
389 {
390 	int wq_num = pfault->event_subtype == MLX5_PFAULT_SUBTYPE_WQE ?
391 		     pfault->wqe.wq_num : pfault->token;
392 	u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)] = {};
393 	int err;
394 
395 	MLX5_SET(page_fault_resume_in, in, opcode, MLX5_CMD_OP_PAGE_FAULT_RESUME);
396 	MLX5_SET(page_fault_resume_in, in, page_fault_type, pfault->type);
397 	MLX5_SET(page_fault_resume_in, in, token, pfault->token);
398 	MLX5_SET(page_fault_resume_in, in, wq_number, wq_num);
399 	MLX5_SET(page_fault_resume_in, in, error, !!error);
400 
401 	err = mlx5_cmd_exec_in(dev->mdev, page_fault_resume, in);
402 	if (err)
403 		mlx5_ib_err(dev, "Failed to resolve the page fault on WQ 0x%x err %d\n",
404 			    wq_num, err);
405 }
406 
407 static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,
408 						unsigned long idx)
409 {
410 	struct ib_umem_odp *odp;
411 	struct mlx5_ib_mr *mr;
412 	struct mlx5_ib_mr *ret;
413 	int err;
414 
415 	odp = ib_umem_odp_alloc_child(to_ib_umem_odp(imr->umem),
416 				      idx * MLX5_IMR_MTT_SIZE,
417 				      MLX5_IMR_MTT_SIZE, &mlx5_mn_ops);
418 	if (IS_ERR(odp))
419 		return ERR_CAST(odp);
420 
421 	mr = mlx5_mr_cache_alloc(
422 		mr_to_mdev(imr), MLX5_IMR_MTT_CACHE_ENTRY, imr->access_flags);
423 	if (IS_ERR(mr)) {
424 		ib_umem_odp_release(odp);
425 		return mr;
426 	}
427 
428 	mr->ibmr.pd = imr->ibmr.pd;
429 	mr->ibmr.device = &mr_to_mdev(imr)->ib_dev;
430 	mr->umem = &odp->umem;
431 	mr->ibmr.lkey = mr->mmkey.key;
432 	mr->ibmr.rkey = mr->mmkey.key;
433 	mr->ibmr.iova = idx * MLX5_IMR_MTT_SIZE;
434 	mr->parent = imr;
435 	odp->private = mr;
436 
437 	/*
438 	 * First refcount is owned by the xarray and second refconut
439 	 * is returned to the caller.
440 	 */
441 	refcount_set(&mr->mmkey.usecount, 2);
442 
443 	err = mlx5_ib_update_xlt(mr, 0,
444 				 MLX5_IMR_MTT_ENTRIES,
445 				 PAGE_SHIFT,
446 				 MLX5_IB_UPD_XLT_ZAP |
447 				 MLX5_IB_UPD_XLT_ENABLE);
448 	if (err) {
449 		ret = ERR_PTR(err);
450 		goto out_mr;
451 	}
452 
453 	xa_lock(&imr->implicit_children);
454 	ret = __xa_cmpxchg(&imr->implicit_children, idx, NULL, mr,
455 			   GFP_KERNEL);
456 	if (unlikely(ret)) {
457 		if (xa_is_err(ret)) {
458 			ret = ERR_PTR(xa_err(ret));
459 			goto out_lock;
460 		}
461 		/*
462 		 * Another thread beat us to creating the child mr, use
463 		 * theirs.
464 		 */
465 		refcount_inc(&ret->mmkey.usecount);
466 		goto out_lock;
467 	}
468 	xa_unlock(&imr->implicit_children);
469 
470 	mlx5_ib_dbg(mr_to_mdev(imr), "key %x mr %p\n", mr->mmkey.key, mr);
471 	return mr;
472 
473 out_lock:
474 	xa_unlock(&imr->implicit_children);
475 out_mr:
476 	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
477 	return ret;
478 }
479 
480 struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
481 					     int access_flags)
482 {
483 	struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device);
484 	struct ib_umem_odp *umem_odp;
485 	struct mlx5_ib_mr *imr;
486 	int err;
487 
488 	if (!mlx5_ib_can_load_pas_with_umr(dev,
489 					   MLX5_IMR_MTT_ENTRIES * PAGE_SIZE))
490 		return ERR_PTR(-EOPNOTSUPP);
491 
492 	umem_odp = ib_umem_odp_alloc_implicit(&dev->ib_dev, access_flags);
493 	if (IS_ERR(umem_odp))
494 		return ERR_CAST(umem_odp);
495 
496 	imr = mlx5_mr_cache_alloc(dev, MLX5_IMR_KSM_CACHE_ENTRY, access_flags);
497 	if (IS_ERR(imr)) {
498 		ib_umem_odp_release(umem_odp);
499 		return imr;
500 	}
501 
502 	imr->ibmr.pd = &pd->ibpd;
503 	imr->ibmr.iova = 0;
504 	imr->umem = &umem_odp->umem;
505 	imr->ibmr.lkey = imr->mmkey.key;
506 	imr->ibmr.rkey = imr->mmkey.key;
507 	imr->ibmr.device = &dev->ib_dev;
508 	imr->umem = &umem_odp->umem;
509 	imr->is_odp_implicit = true;
510 	xa_init(&imr->implicit_children);
511 
512 	err = mlx5_ib_update_xlt(imr, 0,
513 				 mlx5_imr_ksm_entries,
514 				 MLX5_KSM_PAGE_SHIFT,
515 				 MLX5_IB_UPD_XLT_INDIRECT |
516 				 MLX5_IB_UPD_XLT_ZAP |
517 				 MLX5_IB_UPD_XLT_ENABLE);
518 	if (err)
519 		goto out_mr;
520 
521 	err = mlx5r_store_odp_mkey(dev, &imr->mmkey);
522 	if (err)
523 		goto out_mr;
524 
525 	mlx5_ib_dbg(dev, "key %x mr %p\n", imr->mmkey.key, imr);
526 	return imr;
527 out_mr:
528 	mlx5_ib_err(dev, "Failed to register MKEY %d\n", err);
529 	mlx5_ib_dereg_mr(&imr->ibmr, NULL);
530 	return ERR_PTR(err);
531 }
532 
533 void mlx5_ib_free_odp_mr(struct mlx5_ib_mr *mr)
534 {
535 	struct mlx5_ib_mr *mtt;
536 	unsigned long idx;
537 
538 	/*
539 	 * If this is an implicit MR it is already invalidated so we can just
540 	 * delete the children mkeys.
541 	 */
542 	xa_for_each(&mr->implicit_children, idx, mtt) {
543 		xa_erase(&mr->implicit_children, idx);
544 		mlx5_ib_dereg_mr(&mtt->ibmr, NULL);
545 	}
546 }
547 
548 #define MLX5_PF_FLAGS_DOWNGRADE BIT(1)
549 #define MLX5_PF_FLAGS_SNAPSHOT BIT(2)
550 #define MLX5_PF_FLAGS_ENABLE BIT(3)
551 static int pagefault_real_mr(struct mlx5_ib_mr *mr, struct ib_umem_odp *odp,
552 			     u64 user_va, size_t bcnt, u32 *bytes_mapped,
553 			     u32 flags)
554 {
555 	int page_shift, ret, np;
556 	bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE;
557 	u64 access_mask;
558 	u64 start_idx;
559 	bool fault = !(flags & MLX5_PF_FLAGS_SNAPSHOT);
560 	u32 xlt_flags = MLX5_IB_UPD_XLT_ATOMIC;
561 
562 	if (flags & MLX5_PF_FLAGS_ENABLE)
563 		xlt_flags |= MLX5_IB_UPD_XLT_ENABLE;
564 
565 	page_shift = odp->page_shift;
566 	start_idx = (user_va - ib_umem_start(odp)) >> page_shift;
567 	access_mask = ODP_READ_ALLOWED_BIT;
568 
569 	if (odp->umem.writable && !downgrade)
570 		access_mask |= ODP_WRITE_ALLOWED_BIT;
571 
572 	np = ib_umem_odp_map_dma_and_lock(odp, user_va, bcnt, access_mask, fault);
573 	if (np < 0)
574 		return np;
575 
576 	/*
577 	 * No need to check whether the MTTs really belong to this MR, since
578 	 * ib_umem_odp_map_dma_and_lock already checks this.
579 	 */
580 	ret = mlx5_ib_update_xlt(mr, start_idx, np, page_shift, xlt_flags);
581 	mutex_unlock(&odp->umem_mutex);
582 
583 	if (ret < 0) {
584 		if (ret != -EAGAIN)
585 			mlx5_ib_err(mr_to_mdev(mr),
586 				    "Failed to update mkey page tables\n");
587 		goto out;
588 	}
589 
590 	if (bytes_mapped) {
591 		u32 new_mappings = (np << page_shift) -
592 			(user_va - round_down(user_va, 1 << page_shift));
593 
594 		*bytes_mapped += min_t(u32, new_mappings, bcnt);
595 	}
596 
597 	return np << (page_shift - PAGE_SHIFT);
598 
599 out:
600 	return ret;
601 }
602 
603 static int pagefault_implicit_mr(struct mlx5_ib_mr *imr,
604 				 struct ib_umem_odp *odp_imr, u64 user_va,
605 				 size_t bcnt, u32 *bytes_mapped, u32 flags)
606 {
607 	unsigned long end_idx = (user_va + bcnt - 1) >> MLX5_IMR_MTT_SHIFT;
608 	unsigned long upd_start_idx = end_idx + 1;
609 	unsigned long upd_len = 0;
610 	unsigned long npages = 0;
611 	int err;
612 	int ret;
613 
614 	if (unlikely(user_va >= mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE ||
615 		     mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE - user_va < bcnt))
616 		return -EFAULT;
617 
618 	/* Fault each child mr that intersects with our interval. */
619 	while (bcnt) {
620 		unsigned long idx = user_va >> MLX5_IMR_MTT_SHIFT;
621 		struct ib_umem_odp *umem_odp;
622 		struct mlx5_ib_mr *mtt;
623 		u64 len;
624 
625 		xa_lock(&imr->implicit_children);
626 		mtt = xa_load(&imr->implicit_children, idx);
627 		if (unlikely(!mtt)) {
628 			xa_unlock(&imr->implicit_children);
629 			mtt = implicit_get_child_mr(imr, idx);
630 			if (IS_ERR(mtt)) {
631 				ret = PTR_ERR(mtt);
632 				goto out;
633 			}
634 			upd_start_idx = min(upd_start_idx, idx);
635 			upd_len = idx - upd_start_idx + 1;
636 		} else {
637 			refcount_inc(&mtt->mmkey.usecount);
638 			xa_unlock(&imr->implicit_children);
639 		}
640 
641 		umem_odp = to_ib_umem_odp(mtt->umem);
642 		len = min_t(u64, user_va + bcnt, ib_umem_end(umem_odp)) -
643 		      user_va;
644 
645 		ret = pagefault_real_mr(mtt, umem_odp, user_va, len,
646 					bytes_mapped, flags);
647 
648 		mlx5r_deref_odp_mkey(&mtt->mmkey);
649 
650 		if (ret < 0)
651 			goto out;
652 		user_va += len;
653 		bcnt -= len;
654 		npages += ret;
655 	}
656 
657 	ret = npages;
658 
659 	/*
660 	 * Any time the implicit_children are changed we must perform an
661 	 * update of the xlt before exiting to ensure the HW and the
662 	 * implicit_children remains synchronized.
663 	 */
664 out:
665 	if (likely(!upd_len))
666 		return ret;
667 
668 	/*
669 	 * Notice this is not strictly ordered right, the KSM is updated after
670 	 * the implicit_children is updated, so a parallel page fault could
671 	 * see a MR that is not yet visible in the KSM.  This is similar to a
672 	 * parallel page fault seeing a MR that is being concurrently removed
673 	 * from the KSM. Both of these improbable situations are resolved
674 	 * safely by resuming the HW and then taking another page fault. The
675 	 * next pagefault handler will see the new information.
676 	 */
677 	mutex_lock(&odp_imr->umem_mutex);
678 	err = mlx5_ib_update_xlt(imr, upd_start_idx, upd_len, 0,
679 				 MLX5_IB_UPD_XLT_INDIRECT |
680 					 MLX5_IB_UPD_XLT_ATOMIC);
681 	mutex_unlock(&odp_imr->umem_mutex);
682 	if (err) {
683 		mlx5_ib_err(mr_to_mdev(imr), "Failed to update PAS\n");
684 		return err;
685 	}
686 	return ret;
687 }
688 
689 static int pagefault_dmabuf_mr(struct mlx5_ib_mr *mr, size_t bcnt,
690 			       u32 *bytes_mapped, u32 flags)
691 {
692 	struct ib_umem_dmabuf *umem_dmabuf = to_ib_umem_dmabuf(mr->umem);
693 	u32 xlt_flags = 0;
694 	int err;
695 	unsigned int page_size;
696 
697 	if (flags & MLX5_PF_FLAGS_ENABLE)
698 		xlt_flags |= MLX5_IB_UPD_XLT_ENABLE;
699 
700 	dma_resv_lock(umem_dmabuf->attach->dmabuf->resv, NULL);
701 	err = ib_umem_dmabuf_map_pages(umem_dmabuf);
702 	if (err) {
703 		dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
704 		return err;
705 	}
706 
707 	page_size = mlx5_umem_find_best_pgsz(&umem_dmabuf->umem, mkc,
708 					     log_page_size, 0,
709 					     umem_dmabuf->umem.iova);
710 	if (unlikely(page_size < PAGE_SIZE)) {
711 		ib_umem_dmabuf_unmap_pages(umem_dmabuf);
712 		err = -EINVAL;
713 	} else {
714 		err = mlx5_ib_update_mr_pas(mr, xlt_flags);
715 	}
716 	dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
717 
718 	if (err)
719 		return err;
720 
721 	if (bytes_mapped)
722 		*bytes_mapped += bcnt;
723 
724 	return ib_umem_num_pages(mr->umem);
725 }
726 
727 /*
728  * Returns:
729  *  -EFAULT: The io_virt->bcnt is not within the MR, it covers pages that are
730  *           not accessible, or the MR is no longer valid.
731  *  -EAGAIN/-ENOMEM: The operation should be retried
732  *
733  *  -EINVAL/others: General internal malfunction
734  *  >0: Number of pages mapped
735  */
736 static int pagefault_mr(struct mlx5_ib_mr *mr, u64 io_virt, size_t bcnt,
737 			u32 *bytes_mapped, u32 flags)
738 {
739 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
740 
741 	if (unlikely(io_virt < mr->ibmr.iova))
742 		return -EFAULT;
743 
744 	if (mr->umem->is_dmabuf)
745 		return pagefault_dmabuf_mr(mr, bcnt, bytes_mapped, flags);
746 
747 	if (!odp->is_implicit_odp) {
748 		u64 user_va;
749 
750 		if (check_add_overflow(io_virt - mr->ibmr.iova,
751 				       (u64)odp->umem.address, &user_va))
752 			return -EFAULT;
753 		if (unlikely(user_va >= ib_umem_end(odp) ||
754 			     ib_umem_end(odp) - user_va < bcnt))
755 			return -EFAULT;
756 		return pagefault_real_mr(mr, odp, user_va, bcnt, bytes_mapped,
757 					 flags);
758 	}
759 	return pagefault_implicit_mr(mr, odp, io_virt, bcnt, bytes_mapped,
760 				     flags);
761 }
762 
763 int mlx5_ib_init_odp_mr(struct mlx5_ib_mr *mr)
764 {
765 	int ret;
766 
767 	ret = pagefault_real_mr(mr, to_ib_umem_odp(mr->umem), mr->umem->address,
768 				mr->umem->length, NULL,
769 				MLX5_PF_FLAGS_SNAPSHOT | MLX5_PF_FLAGS_ENABLE);
770 	return ret >= 0 ? 0 : ret;
771 }
772 
773 int mlx5_ib_init_dmabuf_mr(struct mlx5_ib_mr *mr)
774 {
775 	int ret;
776 
777 	ret = pagefault_dmabuf_mr(mr, mr->umem->length, NULL,
778 				  MLX5_PF_FLAGS_ENABLE);
779 
780 	return ret >= 0 ? 0 : ret;
781 }
782 
783 struct pf_frame {
784 	struct pf_frame *next;
785 	u32 key;
786 	u64 io_virt;
787 	size_t bcnt;
788 	int depth;
789 };
790 
791 static bool mkey_is_eq(struct mlx5_ib_mkey *mmkey, u32 key)
792 {
793 	if (!mmkey)
794 		return false;
795 	if (mmkey->type == MLX5_MKEY_MW)
796 		return mlx5_base_mkey(mmkey->key) == mlx5_base_mkey(key);
797 	return mmkey->key == key;
798 }
799 
800 /*
801  * Handle a single data segment in a page-fault WQE or RDMA region.
802  *
803  * Returns number of OS pages retrieved on success. The caller may continue to
804  * the next data segment.
805  * Can return the following error codes:
806  * -EAGAIN to designate a temporary error. The caller will abort handling the
807  *  page fault and resolve it.
808  * -EFAULT when there's an error mapping the requested pages. The caller will
809  *  abort the page fault handling.
810  */
811 static int pagefault_single_data_segment(struct mlx5_ib_dev *dev,
812 					 struct ib_pd *pd, u32 key,
813 					 u64 io_virt, size_t bcnt,
814 					 u32 *bytes_committed,
815 					 u32 *bytes_mapped)
816 {
817 	int npages = 0, ret, i, outlen, cur_outlen = 0, depth = 0;
818 	struct pf_frame *head = NULL, *frame;
819 	struct mlx5_ib_mkey *mmkey;
820 	struct mlx5_ib_mr *mr;
821 	struct mlx5_klm *pklm;
822 	u32 *out = NULL;
823 	size_t offset;
824 
825 	io_virt += *bytes_committed;
826 	bcnt -= *bytes_committed;
827 
828 next_mr:
829 	xa_lock(&dev->odp_mkeys);
830 	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(key));
831 	if (!mmkey) {
832 		xa_unlock(&dev->odp_mkeys);
833 		mlx5_ib_dbg(
834 			dev,
835 			"skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
836 			key);
837 		if (bytes_mapped)
838 			*bytes_mapped += bcnt;
839 		/*
840 		 * The user could specify a SGL with multiple lkeys and only
841 		 * some of them are ODP. Treat the non-ODP ones as fully
842 		 * faulted.
843 		 */
844 		ret = 0;
845 		goto end;
846 	}
847 	refcount_inc(&mmkey->usecount);
848 	xa_unlock(&dev->odp_mkeys);
849 
850 	if (!mkey_is_eq(mmkey, key)) {
851 		mlx5_ib_dbg(dev, "failed to find mkey %x\n", key);
852 		ret = -EFAULT;
853 		goto end;
854 	}
855 
856 	switch (mmkey->type) {
857 	case MLX5_MKEY_MR:
858 		mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
859 
860 		ret = pagefault_mr(mr, io_virt, bcnt, bytes_mapped, 0);
861 		if (ret < 0)
862 			goto end;
863 
864 		mlx5_update_odp_stats(mr, faults, ret);
865 
866 		npages += ret;
867 		ret = 0;
868 		break;
869 
870 	case MLX5_MKEY_MW:
871 	case MLX5_MKEY_INDIRECT_DEVX:
872 		if (depth >= MLX5_CAP_GEN(dev->mdev, max_indirection)) {
873 			mlx5_ib_dbg(dev, "indirection level exceeded\n");
874 			ret = -EFAULT;
875 			goto end;
876 		}
877 
878 		outlen = MLX5_ST_SZ_BYTES(query_mkey_out) +
879 			sizeof(*pklm) * (mmkey->ndescs - 2);
880 
881 		if (outlen > cur_outlen) {
882 			kfree(out);
883 			out = kzalloc(outlen, GFP_KERNEL);
884 			if (!out) {
885 				ret = -ENOMEM;
886 				goto end;
887 			}
888 			cur_outlen = outlen;
889 		}
890 
891 		pklm = (struct mlx5_klm *)MLX5_ADDR_OF(query_mkey_out, out,
892 						       bsf0_klm0_pas_mtt0_1);
893 
894 		ret = mlx5_core_query_mkey(dev->mdev, mmkey->key, out, outlen);
895 		if (ret)
896 			goto end;
897 
898 		offset = io_virt - MLX5_GET64(query_mkey_out, out,
899 					      memory_key_mkey_entry.start_addr);
900 
901 		for (i = 0; bcnt && i < mmkey->ndescs; i++, pklm++) {
902 			if (offset >= be32_to_cpu(pklm->bcount)) {
903 				offset -= be32_to_cpu(pklm->bcount);
904 				continue;
905 			}
906 
907 			frame = kzalloc(sizeof(*frame), GFP_KERNEL);
908 			if (!frame) {
909 				ret = -ENOMEM;
910 				goto end;
911 			}
912 
913 			frame->key = be32_to_cpu(pklm->key);
914 			frame->io_virt = be64_to_cpu(pklm->va) + offset;
915 			frame->bcnt = min_t(size_t, bcnt,
916 					    be32_to_cpu(pklm->bcount) - offset);
917 			frame->depth = depth + 1;
918 			frame->next = head;
919 			head = frame;
920 
921 			bcnt -= frame->bcnt;
922 			offset = 0;
923 		}
924 		break;
925 
926 	default:
927 		mlx5_ib_dbg(dev, "wrong mkey type %d\n", mmkey->type);
928 		ret = -EFAULT;
929 		goto end;
930 	}
931 
932 	if (head) {
933 		frame = head;
934 		head = frame->next;
935 
936 		key = frame->key;
937 		io_virt = frame->io_virt;
938 		bcnt = frame->bcnt;
939 		depth = frame->depth;
940 		kfree(frame);
941 
942 		mlx5r_deref_odp_mkey(mmkey);
943 		goto next_mr;
944 	}
945 
946 end:
947 	if (mmkey)
948 		mlx5r_deref_odp_mkey(mmkey);
949 	while (head) {
950 		frame = head;
951 		head = frame->next;
952 		kfree(frame);
953 	}
954 	kfree(out);
955 
956 	*bytes_committed = 0;
957 	return ret ? ret : npages;
958 }
959 
960 /*
961  * Parse a series of data segments for page fault handling.
962  *
963  * @dev:  Pointer to mlx5 IB device
964  * @pfault: contains page fault information.
965  * @wqe: points at the first data segment in the WQE.
966  * @wqe_end: points after the end of the WQE.
967  * @bytes_mapped: receives the number of bytes that the function was able to
968  *                map. This allows the caller to decide intelligently whether
969  *                enough memory was mapped to resolve the page fault
970  *                successfully (e.g. enough for the next MTU, or the entire
971  *                WQE).
972  * @total_wqe_bytes: receives the total data size of this WQE in bytes (minus
973  *                   the committed bytes).
974  * @receive_queue: receive WQE end of sg list
975  *
976  * Returns the number of pages loaded if positive, zero for an empty WQE, or a
977  * negative error code.
978  */
979 static int pagefault_data_segments(struct mlx5_ib_dev *dev,
980 				   struct mlx5_pagefault *pfault,
981 				   void *wqe,
982 				   void *wqe_end, u32 *bytes_mapped,
983 				   u32 *total_wqe_bytes, bool receive_queue)
984 {
985 	int ret = 0, npages = 0;
986 	u64 io_virt;
987 	u32 key;
988 	u32 byte_count;
989 	size_t bcnt;
990 	int inline_segment;
991 
992 	if (bytes_mapped)
993 		*bytes_mapped = 0;
994 	if (total_wqe_bytes)
995 		*total_wqe_bytes = 0;
996 
997 	while (wqe < wqe_end) {
998 		struct mlx5_wqe_data_seg *dseg = wqe;
999 
1000 		io_virt = be64_to_cpu(dseg->addr);
1001 		key = be32_to_cpu(dseg->lkey);
1002 		byte_count = be32_to_cpu(dseg->byte_count);
1003 		inline_segment = !!(byte_count &  MLX5_INLINE_SEG);
1004 		bcnt	       = byte_count & ~MLX5_INLINE_SEG;
1005 
1006 		if (inline_segment) {
1007 			bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK;
1008 			wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt,
1009 				     16);
1010 		} else {
1011 			wqe += sizeof(*dseg);
1012 		}
1013 
1014 		/* receive WQE end of sg list. */
1015 		if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY &&
1016 		    io_virt == 0)
1017 			break;
1018 
1019 		if (!inline_segment && total_wqe_bytes) {
1020 			*total_wqe_bytes += bcnt - min_t(size_t, bcnt,
1021 					pfault->bytes_committed);
1022 		}
1023 
1024 		/* A zero length data segment designates a length of 2GB. */
1025 		if (bcnt == 0)
1026 			bcnt = 1U << 31;
1027 
1028 		if (inline_segment || bcnt <= pfault->bytes_committed) {
1029 			pfault->bytes_committed -=
1030 				min_t(size_t, bcnt,
1031 				      pfault->bytes_committed);
1032 			continue;
1033 		}
1034 
1035 		ret = pagefault_single_data_segment(dev, NULL, key,
1036 						    io_virt, bcnt,
1037 						    &pfault->bytes_committed,
1038 						    bytes_mapped);
1039 		if (ret < 0)
1040 			break;
1041 		npages += ret;
1042 	}
1043 
1044 	return ret < 0 ? ret : npages;
1045 }
1046 
1047 /*
1048  * Parse initiator WQE. Advances the wqe pointer to point at the
1049  * scatter-gather list, and set wqe_end to the end of the WQE.
1050  */
1051 static int mlx5_ib_mr_initiator_pfault_handler(
1052 	struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault,
1053 	struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length)
1054 {
1055 	struct mlx5_wqe_ctrl_seg *ctrl = *wqe;
1056 	u16 wqe_index = pfault->wqe.wqe_index;
1057 	struct mlx5_base_av *av;
1058 	unsigned ds, opcode;
1059 	u32 qpn = qp->trans_qp.base.mqp.qpn;
1060 
1061 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
1062 	if (ds * MLX5_WQE_DS_UNITS > wqe_length) {
1063 		mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n",
1064 			    ds, wqe_length);
1065 		return -EFAULT;
1066 	}
1067 
1068 	if (ds == 0) {
1069 		mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n",
1070 			    wqe_index, qpn);
1071 		return -EFAULT;
1072 	}
1073 
1074 	*wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS;
1075 	*wqe += sizeof(*ctrl);
1076 
1077 	opcode = be32_to_cpu(ctrl->opmod_idx_opcode) &
1078 		 MLX5_WQE_CTRL_OPCODE_MASK;
1079 
1080 	if (qp->type == IB_QPT_XRC_INI)
1081 		*wqe += sizeof(struct mlx5_wqe_xrc_seg);
1082 
1083 	if (qp->type == IB_QPT_UD || qp->type == MLX5_IB_QPT_DCI) {
1084 		av = *wqe;
1085 		if (av->dqp_dct & cpu_to_be32(MLX5_EXTENDED_UD_AV))
1086 			*wqe += sizeof(struct mlx5_av);
1087 		else
1088 			*wqe += sizeof(struct mlx5_base_av);
1089 	}
1090 
1091 	switch (opcode) {
1092 	case MLX5_OPCODE_RDMA_WRITE:
1093 	case MLX5_OPCODE_RDMA_WRITE_IMM:
1094 	case MLX5_OPCODE_RDMA_READ:
1095 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1096 		break;
1097 	case MLX5_OPCODE_ATOMIC_CS:
1098 	case MLX5_OPCODE_ATOMIC_FA:
1099 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1100 		*wqe += sizeof(struct mlx5_wqe_atomic_seg);
1101 		break;
1102 	}
1103 
1104 	return 0;
1105 }
1106 
1107 /*
1108  * Parse responder WQE and set wqe_end to the end of the WQE.
1109  */
1110 static int mlx5_ib_mr_responder_pfault_handler_srq(struct mlx5_ib_dev *dev,
1111 						   struct mlx5_ib_srq *srq,
1112 						   void **wqe, void **wqe_end,
1113 						   int wqe_length)
1114 {
1115 	int wqe_size = 1 << srq->msrq.wqe_shift;
1116 
1117 	if (wqe_size > wqe_length) {
1118 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1119 		return -EFAULT;
1120 	}
1121 
1122 	*wqe_end = *wqe + wqe_size;
1123 	*wqe += sizeof(struct mlx5_wqe_srq_next_seg);
1124 
1125 	return 0;
1126 }
1127 
1128 static int mlx5_ib_mr_responder_pfault_handler_rq(struct mlx5_ib_dev *dev,
1129 						  struct mlx5_ib_qp *qp,
1130 						  void *wqe, void **wqe_end,
1131 						  int wqe_length)
1132 {
1133 	struct mlx5_ib_wq *wq = &qp->rq;
1134 	int wqe_size = 1 << wq->wqe_shift;
1135 
1136 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE) {
1137 		mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n");
1138 		return -EFAULT;
1139 	}
1140 
1141 	if (wqe_size > wqe_length) {
1142 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1143 		return -EFAULT;
1144 	}
1145 
1146 	*wqe_end = wqe + wqe_size;
1147 
1148 	return 0;
1149 }
1150 
1151 static inline struct mlx5_core_rsc_common *odp_get_rsc(struct mlx5_ib_dev *dev,
1152 						       u32 wq_num, int pf_type)
1153 {
1154 	struct mlx5_core_rsc_common *common = NULL;
1155 	struct mlx5_core_srq *srq;
1156 
1157 	switch (pf_type) {
1158 	case MLX5_WQE_PF_TYPE_RMP:
1159 		srq = mlx5_cmd_get_srq(dev, wq_num);
1160 		if (srq)
1161 			common = &srq->common;
1162 		break;
1163 	case MLX5_WQE_PF_TYPE_REQ_SEND_OR_WRITE:
1164 	case MLX5_WQE_PF_TYPE_RESP:
1165 	case MLX5_WQE_PF_TYPE_REQ_READ_OR_ATOMIC:
1166 		common = mlx5_core_res_hold(dev, wq_num, MLX5_RES_QP);
1167 		break;
1168 	default:
1169 		break;
1170 	}
1171 
1172 	return common;
1173 }
1174 
1175 static inline struct mlx5_ib_qp *res_to_qp(struct mlx5_core_rsc_common *res)
1176 {
1177 	struct mlx5_core_qp *mqp = (struct mlx5_core_qp *)res;
1178 
1179 	return to_mibqp(mqp);
1180 }
1181 
1182 static inline struct mlx5_ib_srq *res_to_srq(struct mlx5_core_rsc_common *res)
1183 {
1184 	struct mlx5_core_srq *msrq =
1185 		container_of(res, struct mlx5_core_srq, common);
1186 
1187 	return to_mibsrq(msrq);
1188 }
1189 
1190 static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
1191 					  struct mlx5_pagefault *pfault)
1192 {
1193 	bool sq = pfault->type & MLX5_PFAULT_REQUESTOR;
1194 	u16 wqe_index = pfault->wqe.wqe_index;
1195 	void *wqe, *wqe_start = NULL, *wqe_end = NULL;
1196 	u32 bytes_mapped, total_wqe_bytes;
1197 	struct mlx5_core_rsc_common *res;
1198 	int resume_with_error = 1;
1199 	struct mlx5_ib_qp *qp;
1200 	size_t bytes_copied;
1201 	int ret = 0;
1202 
1203 	res = odp_get_rsc(dev, pfault->wqe.wq_num, pfault->type);
1204 	if (!res) {
1205 		mlx5_ib_dbg(dev, "wqe page fault for missing resource %d\n", pfault->wqe.wq_num);
1206 		return;
1207 	}
1208 
1209 	if (res->res != MLX5_RES_QP && res->res != MLX5_RES_SRQ &&
1210 	    res->res != MLX5_RES_XSRQ) {
1211 		mlx5_ib_err(dev, "wqe page fault for unsupported type %d\n",
1212 			    pfault->type);
1213 		goto resolve_page_fault;
1214 	}
1215 
1216 	wqe_start = (void *)__get_free_page(GFP_KERNEL);
1217 	if (!wqe_start) {
1218 		mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n");
1219 		goto resolve_page_fault;
1220 	}
1221 
1222 	wqe = wqe_start;
1223 	qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL;
1224 	if (qp && sq) {
1225 		ret = mlx5_ib_read_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE,
1226 					  &bytes_copied);
1227 		if (ret)
1228 			goto read_user;
1229 		ret = mlx5_ib_mr_initiator_pfault_handler(
1230 			dev, pfault, qp, &wqe, &wqe_end, bytes_copied);
1231 	} else if (qp && !sq) {
1232 		ret = mlx5_ib_read_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE,
1233 					  &bytes_copied);
1234 		if (ret)
1235 			goto read_user;
1236 		ret = mlx5_ib_mr_responder_pfault_handler_rq(
1237 			dev, qp, wqe, &wqe_end, bytes_copied);
1238 	} else if (!qp) {
1239 		struct mlx5_ib_srq *srq = res_to_srq(res);
1240 
1241 		ret = mlx5_ib_read_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE,
1242 					   &bytes_copied);
1243 		if (ret)
1244 			goto read_user;
1245 		ret = mlx5_ib_mr_responder_pfault_handler_srq(
1246 			dev, srq, &wqe, &wqe_end, bytes_copied);
1247 	}
1248 
1249 	if (ret < 0 || wqe >= wqe_end)
1250 		goto resolve_page_fault;
1251 
1252 	ret = pagefault_data_segments(dev, pfault, wqe, wqe_end, &bytes_mapped,
1253 				      &total_wqe_bytes, !sq);
1254 	if (ret == -EAGAIN)
1255 		goto out;
1256 
1257 	if (ret < 0 || total_wqe_bytes > bytes_mapped)
1258 		goto resolve_page_fault;
1259 
1260 out:
1261 	ret = 0;
1262 	resume_with_error = 0;
1263 
1264 read_user:
1265 	if (ret)
1266 		mlx5_ib_err(
1267 			dev,
1268 			"Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n",
1269 			ret, wqe_index, pfault->token);
1270 
1271 resolve_page_fault:
1272 	mlx5_ib_page_fault_resume(dev, pfault, resume_with_error);
1273 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n",
1274 		    pfault->wqe.wq_num, resume_with_error,
1275 		    pfault->type);
1276 	mlx5_core_res_put(res);
1277 	free_page((unsigned long)wqe_start);
1278 }
1279 
1280 static int pages_in_range(u64 address, u32 length)
1281 {
1282 	return (ALIGN(address + length, PAGE_SIZE) -
1283 		(address & PAGE_MASK)) >> PAGE_SHIFT;
1284 }
1285 
1286 static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev,
1287 					   struct mlx5_pagefault *pfault)
1288 {
1289 	u64 address;
1290 	u32 length;
1291 	u32 prefetch_len = pfault->bytes_committed;
1292 	int prefetch_activated = 0;
1293 	u32 rkey = pfault->rdma.r_key;
1294 	int ret;
1295 
1296 	/* The RDMA responder handler handles the page fault in two parts.
1297 	 * First it brings the necessary pages for the current packet
1298 	 * (and uses the pfault context), and then (after resuming the QP)
1299 	 * prefetches more pages. The second operation cannot use the pfault
1300 	 * context and therefore uses the dummy_pfault context allocated on
1301 	 * the stack */
1302 	pfault->rdma.rdma_va += pfault->bytes_committed;
1303 	pfault->rdma.rdma_op_len -= min(pfault->bytes_committed,
1304 					 pfault->rdma.rdma_op_len);
1305 	pfault->bytes_committed = 0;
1306 
1307 	address = pfault->rdma.rdma_va;
1308 	length  = pfault->rdma.rdma_op_len;
1309 
1310 	/* For some operations, the hardware cannot tell the exact message
1311 	 * length, and in those cases it reports zero. Use prefetch
1312 	 * logic. */
1313 	if (length == 0) {
1314 		prefetch_activated = 1;
1315 		length = pfault->rdma.packet_size;
1316 		prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len);
1317 	}
1318 
1319 	ret = pagefault_single_data_segment(dev, NULL, rkey, address, length,
1320 					    &pfault->bytes_committed, NULL);
1321 	if (ret == -EAGAIN) {
1322 		/* We're racing with an invalidation, don't prefetch */
1323 		prefetch_activated = 0;
1324 	} else if (ret < 0 || pages_in_range(address, length) > ret) {
1325 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1326 		if (ret != -ENOENT)
1327 			mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n",
1328 				    ret, pfault->token, pfault->type);
1329 		return;
1330 	}
1331 
1332 	mlx5_ib_page_fault_resume(dev, pfault, 0);
1333 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n",
1334 		    pfault->token, pfault->type,
1335 		    prefetch_activated);
1336 
1337 	/* At this point, there might be a new pagefault already arriving in
1338 	 * the eq, switch to the dummy pagefault for the rest of the
1339 	 * processing. We're still OK with the objects being alive as the
1340 	 * work-queue is being fenced. */
1341 
1342 	if (prefetch_activated) {
1343 		u32 bytes_committed = 0;
1344 
1345 		ret = pagefault_single_data_segment(dev, NULL, rkey, address,
1346 						    prefetch_len,
1347 						    &bytes_committed, NULL);
1348 		if (ret < 0 && ret != -EAGAIN) {
1349 			mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n",
1350 				    ret, pfault->token, address, prefetch_len);
1351 		}
1352 	}
1353 }
1354 
1355 static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault)
1356 {
1357 	u8 event_subtype = pfault->event_subtype;
1358 
1359 	switch (event_subtype) {
1360 	case MLX5_PFAULT_SUBTYPE_WQE:
1361 		mlx5_ib_mr_wqe_pfault_handler(dev, pfault);
1362 		break;
1363 	case MLX5_PFAULT_SUBTYPE_RDMA:
1364 		mlx5_ib_mr_rdma_pfault_handler(dev, pfault);
1365 		break;
1366 	default:
1367 		mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n",
1368 			    event_subtype);
1369 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1370 	}
1371 }
1372 
1373 static void mlx5_ib_eqe_pf_action(struct work_struct *work)
1374 {
1375 	struct mlx5_pagefault *pfault = container_of(work,
1376 						     struct mlx5_pagefault,
1377 						     work);
1378 	struct mlx5_ib_pf_eq *eq = pfault->eq;
1379 
1380 	mlx5_ib_pfault(eq->dev, pfault);
1381 	mempool_free(pfault, eq->pool);
1382 }
1383 
1384 static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq)
1385 {
1386 	struct mlx5_eqe_page_fault *pf_eqe;
1387 	struct mlx5_pagefault *pfault;
1388 	struct mlx5_eqe *eqe;
1389 	int cc = 0;
1390 
1391 	while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) {
1392 		pfault = mempool_alloc(eq->pool, GFP_ATOMIC);
1393 		if (!pfault) {
1394 			schedule_work(&eq->work);
1395 			break;
1396 		}
1397 
1398 		pf_eqe = &eqe->data.page_fault;
1399 		pfault->event_subtype = eqe->sub_type;
1400 		pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);
1401 
1402 		mlx5_ib_dbg(eq->dev,
1403 			    "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
1404 			    eqe->sub_type, pfault->bytes_committed);
1405 
1406 		switch (eqe->sub_type) {
1407 		case MLX5_PFAULT_SUBTYPE_RDMA:
1408 			/* RDMA based event */
1409 			pfault->type =
1410 				be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
1411 			pfault->token =
1412 				be32_to_cpu(pf_eqe->rdma.pftype_token) &
1413 				MLX5_24BIT_MASK;
1414 			pfault->rdma.r_key =
1415 				be32_to_cpu(pf_eqe->rdma.r_key);
1416 			pfault->rdma.packet_size =
1417 				be16_to_cpu(pf_eqe->rdma.packet_length);
1418 			pfault->rdma.rdma_op_len =
1419 				be32_to_cpu(pf_eqe->rdma.rdma_op_len);
1420 			pfault->rdma.rdma_va =
1421 				be64_to_cpu(pf_eqe->rdma.rdma_va);
1422 			mlx5_ib_dbg(eq->dev,
1423 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
1424 				    pfault->type, pfault->token,
1425 				    pfault->rdma.r_key);
1426 			mlx5_ib_dbg(eq->dev,
1427 				    "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
1428 				    pfault->rdma.rdma_op_len,
1429 				    pfault->rdma.rdma_va);
1430 			break;
1431 
1432 		case MLX5_PFAULT_SUBTYPE_WQE:
1433 			/* WQE based event */
1434 			pfault->type =
1435 				(be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7;
1436 			pfault->token =
1437 				be32_to_cpu(pf_eqe->wqe.token);
1438 			pfault->wqe.wq_num =
1439 				be32_to_cpu(pf_eqe->wqe.pftype_wq) &
1440 				MLX5_24BIT_MASK;
1441 			pfault->wqe.wqe_index =
1442 				be16_to_cpu(pf_eqe->wqe.wqe_index);
1443 			pfault->wqe.packet_size =
1444 				be16_to_cpu(pf_eqe->wqe.packet_length);
1445 			mlx5_ib_dbg(eq->dev,
1446 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
1447 				    pfault->type, pfault->token,
1448 				    pfault->wqe.wq_num,
1449 				    pfault->wqe.wqe_index);
1450 			break;
1451 
1452 		default:
1453 			mlx5_ib_warn(eq->dev,
1454 				     "Unsupported page fault event sub-type: 0x%02hhx\n",
1455 				     eqe->sub_type);
1456 			/* Unsupported page faults should still be
1457 			 * resolved by the page fault handler
1458 			 */
1459 		}
1460 
1461 		pfault->eq = eq;
1462 		INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action);
1463 		queue_work(eq->wq, &pfault->work);
1464 
1465 		cc = mlx5_eq_update_cc(eq->core, ++cc);
1466 	}
1467 
1468 	mlx5_eq_update_ci(eq->core, cc, 1);
1469 }
1470 
1471 static int mlx5_ib_eq_pf_int(struct notifier_block *nb, unsigned long type,
1472 			     void *data)
1473 {
1474 	struct mlx5_ib_pf_eq *eq =
1475 		container_of(nb, struct mlx5_ib_pf_eq, irq_nb);
1476 	unsigned long flags;
1477 
1478 	if (spin_trylock_irqsave(&eq->lock, flags)) {
1479 		mlx5_ib_eq_pf_process(eq);
1480 		spin_unlock_irqrestore(&eq->lock, flags);
1481 	} else {
1482 		schedule_work(&eq->work);
1483 	}
1484 
1485 	return IRQ_HANDLED;
1486 }
1487 
1488 /* mempool_refill() was proposed but unfortunately wasn't accepted
1489  * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
1490  * Cheap workaround.
1491  */
1492 static void mempool_refill(mempool_t *pool)
1493 {
1494 	while (pool->curr_nr < pool->min_nr)
1495 		mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
1496 }
1497 
1498 static void mlx5_ib_eq_pf_action(struct work_struct *work)
1499 {
1500 	struct mlx5_ib_pf_eq *eq =
1501 		container_of(work, struct mlx5_ib_pf_eq, work);
1502 
1503 	mempool_refill(eq->pool);
1504 
1505 	spin_lock_irq(&eq->lock);
1506 	mlx5_ib_eq_pf_process(eq);
1507 	spin_unlock_irq(&eq->lock);
1508 }
1509 
1510 enum {
1511 	MLX5_IB_NUM_PF_EQE	= 0x1000,
1512 	MLX5_IB_NUM_PF_DRAIN	= 64,
1513 };
1514 
1515 int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1516 {
1517 	struct mlx5_eq_param param = {};
1518 	int err = 0;
1519 
1520 	mutex_lock(&dev->odp_eq_mutex);
1521 	if (eq->core)
1522 		goto unlock;
1523 	INIT_WORK(&eq->work, mlx5_ib_eq_pf_action);
1524 	spin_lock_init(&eq->lock);
1525 	eq->dev = dev;
1526 
1527 	eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN,
1528 					       sizeof(struct mlx5_pagefault));
1529 	if (!eq->pool) {
1530 		err = -ENOMEM;
1531 		goto unlock;
1532 	}
1533 
1534 	eq->wq = alloc_workqueue("mlx5_ib_page_fault",
1535 				 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM,
1536 				 MLX5_NUM_CMD_EQE);
1537 	if (!eq->wq) {
1538 		err = -ENOMEM;
1539 		goto err_mempool;
1540 	}
1541 
1542 	eq->irq_nb.notifier_call = mlx5_ib_eq_pf_int;
1543 	param = (struct mlx5_eq_param) {
1544 		.irq_index = MLX5_IRQ_EQ_CTRL,
1545 		.nent = MLX5_IB_NUM_PF_EQE,
1546 	};
1547 	param.mask[0] = 1ull << MLX5_EVENT_TYPE_PAGE_FAULT;
1548 	if (!zalloc_cpumask_var(&param.affinity, GFP_KERNEL)) {
1549 		err = -ENOMEM;
1550 		goto err_wq;
1551 	}
1552 	eq->core = mlx5_eq_create_generic(dev->mdev, &param);
1553 	free_cpumask_var(param.affinity);
1554 	if (IS_ERR(eq->core)) {
1555 		err = PTR_ERR(eq->core);
1556 		goto err_wq;
1557 	}
1558 	err = mlx5_eq_enable(dev->mdev, eq->core, &eq->irq_nb);
1559 	if (err) {
1560 		mlx5_ib_err(dev, "failed to enable odp EQ %d\n", err);
1561 		goto err_eq;
1562 	}
1563 
1564 	mutex_unlock(&dev->odp_eq_mutex);
1565 	return 0;
1566 err_eq:
1567 	mlx5_eq_destroy_generic(dev->mdev, eq->core);
1568 err_wq:
1569 	eq->core = NULL;
1570 	destroy_workqueue(eq->wq);
1571 err_mempool:
1572 	mempool_destroy(eq->pool);
1573 unlock:
1574 	mutex_unlock(&dev->odp_eq_mutex);
1575 	return err;
1576 }
1577 
1578 static int
1579 mlx5_ib_odp_destroy_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1580 {
1581 	int err;
1582 
1583 	if (!eq->core)
1584 		return 0;
1585 	mlx5_eq_disable(dev->mdev, eq->core, &eq->irq_nb);
1586 	err = mlx5_eq_destroy_generic(dev->mdev, eq->core);
1587 	cancel_work_sync(&eq->work);
1588 	destroy_workqueue(eq->wq);
1589 	mempool_destroy(eq->pool);
1590 
1591 	return err;
1592 }
1593 
1594 void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
1595 {
1596 	if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1597 		return;
1598 
1599 	switch (ent->order - 2) {
1600 	case MLX5_IMR_MTT_CACHE_ENTRY:
1601 		ent->page = PAGE_SHIFT;
1602 		ent->xlt = MLX5_IMR_MTT_ENTRIES *
1603 			   sizeof(struct mlx5_mtt) /
1604 			   MLX5_IB_UMR_OCTOWORD;
1605 		ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
1606 		ent->limit = 0;
1607 		break;
1608 
1609 	case MLX5_IMR_KSM_CACHE_ENTRY:
1610 		ent->page = MLX5_KSM_PAGE_SHIFT;
1611 		ent->xlt = mlx5_imr_ksm_entries *
1612 			   sizeof(struct mlx5_klm) /
1613 			   MLX5_IB_UMR_OCTOWORD;
1614 		ent->access_mode = MLX5_MKC_ACCESS_MODE_KSM;
1615 		ent->limit = 0;
1616 		break;
1617 	}
1618 }
1619 
1620 static const struct ib_device_ops mlx5_ib_dev_odp_ops = {
1621 	.advise_mr = mlx5_ib_advise_mr,
1622 };
1623 
1624 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
1625 {
1626 	int ret = 0;
1627 
1628 	internal_fill_odp_caps(dev);
1629 
1630 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1631 		return ret;
1632 
1633 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_odp_ops);
1634 
1635 	if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) {
1636 		ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey);
1637 		if (ret) {
1638 			mlx5_ib_err(dev, "Error getting null_mkey %d\n", ret);
1639 			return ret;
1640 		}
1641 	}
1642 
1643 	mutex_init(&dev->odp_eq_mutex);
1644 	return ret;
1645 }
1646 
1647 void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev)
1648 {
1649 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1650 		return;
1651 
1652 	mlx5_ib_odp_destroy_eq(dev, &dev->odp_pf_eq);
1653 }
1654 
1655 int mlx5_ib_odp_init(void)
1656 {
1657 	mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) -
1658 				       MLX5_IMR_MTT_BITS);
1659 
1660 	return 0;
1661 }
1662 
1663 struct prefetch_mr_work {
1664 	struct work_struct work;
1665 	u32 pf_flags;
1666 	u32 num_sge;
1667 	struct {
1668 		u64 io_virt;
1669 		struct mlx5_ib_mr *mr;
1670 		size_t length;
1671 	} frags[];
1672 };
1673 
1674 static void destroy_prefetch_work(struct prefetch_mr_work *work)
1675 {
1676 	u32 i;
1677 
1678 	for (i = 0; i < work->num_sge; ++i)
1679 		mlx5r_deref_odp_mkey(&work->frags[i].mr->mmkey);
1680 
1681 	kvfree(work);
1682 }
1683 
1684 static struct mlx5_ib_mr *
1685 get_prefetchable_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
1686 		    u32 lkey)
1687 {
1688 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1689 	struct mlx5_ib_mr *mr = NULL;
1690 	struct mlx5_ib_mkey *mmkey;
1691 
1692 	xa_lock(&dev->odp_mkeys);
1693 	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(lkey));
1694 	if (!mmkey || mmkey->key != lkey) {
1695 		mr = ERR_PTR(-ENOENT);
1696 		goto end;
1697 	}
1698 	if (mmkey->type != MLX5_MKEY_MR) {
1699 		mr = ERR_PTR(-EINVAL);
1700 		goto end;
1701 	}
1702 
1703 	mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
1704 
1705 	if (mr->ibmr.pd != pd) {
1706 		mr = ERR_PTR(-EPERM);
1707 		goto end;
1708 	}
1709 
1710 	/* prefetch with write-access must be supported by the MR */
1711 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
1712 	    !mr->umem->writable) {
1713 		mr = ERR_PTR(-EPERM);
1714 		goto end;
1715 	}
1716 
1717 	refcount_inc(&mmkey->usecount);
1718 end:
1719 	xa_unlock(&dev->odp_mkeys);
1720 	return mr;
1721 }
1722 
1723 static void mlx5_ib_prefetch_mr_work(struct work_struct *w)
1724 {
1725 	struct prefetch_mr_work *work =
1726 		container_of(w, struct prefetch_mr_work, work);
1727 	u32 bytes_mapped = 0;
1728 	int ret;
1729 	u32 i;
1730 
1731 	/* We rely on IB/core that work is executed if we have num_sge != 0 only. */
1732 	WARN_ON(!work->num_sge);
1733 	for (i = 0; i < work->num_sge; ++i) {
1734 		ret = pagefault_mr(work->frags[i].mr, work->frags[i].io_virt,
1735 				   work->frags[i].length, &bytes_mapped,
1736 				   work->pf_flags);
1737 		if (ret <= 0)
1738 			continue;
1739 		mlx5_update_odp_stats(work->frags[i].mr, prefetch, ret);
1740 	}
1741 
1742 	destroy_prefetch_work(work);
1743 }
1744 
1745 static int init_prefetch_work(struct ib_pd *pd,
1746 			       enum ib_uverbs_advise_mr_advice advice,
1747 			       u32 pf_flags, struct prefetch_mr_work *work,
1748 			       struct ib_sge *sg_list, u32 num_sge)
1749 {
1750 	u32 i;
1751 
1752 	INIT_WORK(&work->work, mlx5_ib_prefetch_mr_work);
1753 	work->pf_flags = pf_flags;
1754 
1755 	for (i = 0; i < num_sge; ++i) {
1756 		struct mlx5_ib_mr *mr;
1757 
1758 		mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1759 		if (IS_ERR(mr)) {
1760 			work->num_sge = i;
1761 			return PTR_ERR(mr);
1762 		}
1763 		work->frags[i].io_virt = sg_list[i].addr;
1764 		work->frags[i].length = sg_list[i].length;
1765 		work->frags[i].mr = mr;
1766 	}
1767 	work->num_sge = num_sge;
1768 	return 0;
1769 }
1770 
1771 static int mlx5_ib_prefetch_sg_list(struct ib_pd *pd,
1772 				    enum ib_uverbs_advise_mr_advice advice,
1773 				    u32 pf_flags, struct ib_sge *sg_list,
1774 				    u32 num_sge)
1775 {
1776 	u32 bytes_mapped = 0;
1777 	int ret = 0;
1778 	u32 i;
1779 
1780 	for (i = 0; i < num_sge; ++i) {
1781 		struct mlx5_ib_mr *mr;
1782 
1783 		mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1784 		if (IS_ERR(mr))
1785 			return PTR_ERR(mr);
1786 		ret = pagefault_mr(mr, sg_list[i].addr, sg_list[i].length,
1787 				   &bytes_mapped, pf_flags);
1788 		if (ret < 0) {
1789 			mlx5r_deref_odp_mkey(&mr->mmkey);
1790 			return ret;
1791 		}
1792 		mlx5_update_odp_stats(mr, prefetch, ret);
1793 		mlx5r_deref_odp_mkey(&mr->mmkey);
1794 	}
1795 
1796 	return 0;
1797 }
1798 
1799 int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
1800 			       enum ib_uverbs_advise_mr_advice advice,
1801 			       u32 flags, struct ib_sge *sg_list, u32 num_sge)
1802 {
1803 	u32 pf_flags = 0;
1804 	struct prefetch_mr_work *work;
1805 	int rc;
1806 
1807 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH)
1808 		pf_flags |= MLX5_PF_FLAGS_DOWNGRADE;
1809 
1810 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT)
1811 		pf_flags |= MLX5_PF_FLAGS_SNAPSHOT;
1812 
1813 	if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH)
1814 		return mlx5_ib_prefetch_sg_list(pd, advice, pf_flags, sg_list,
1815 						num_sge);
1816 
1817 	work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL);
1818 	if (!work)
1819 		return -ENOMEM;
1820 
1821 	rc = init_prefetch_work(pd, advice, pf_flags, work, sg_list, num_sge);
1822 	if (rc) {
1823 		destroy_prefetch_work(work);
1824 		return rc;
1825 	}
1826 	queue_work(system_unbound_wq, &work->work);
1827 	return 0;
1828 }
1829