xref: /openbmc/linux/drivers/infiniband/hw/mlx5/odp.c (revision dff03381)
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 "umr.h"
42 #include "qp.h"
43 
44 #include <linux/mlx5/eq.h>
45 
46 /* Contains the details of a pagefault. */
47 struct mlx5_pagefault {
48 	u32			bytes_committed;
49 	u32			token;
50 	u8			event_subtype;
51 	u8			type;
52 	union {
53 		/* Initiator or send message responder pagefault details. */
54 		struct {
55 			/* Received packet size, only valid for responders. */
56 			u32	packet_size;
57 			/*
58 			 * Number of resource holding WQE, depends on type.
59 			 */
60 			u32	wq_num;
61 			/*
62 			 * WQE index. Refers to either the send queue or
63 			 * receive queue, according to event_subtype.
64 			 */
65 			u16	wqe_index;
66 		} wqe;
67 		/* RDMA responder pagefault details */
68 		struct {
69 			u32	r_key;
70 			/*
71 			 * Received packet size, minimal size page fault
72 			 * resolution required for forward progress.
73 			 */
74 			u32	packet_size;
75 			u32	rdma_op_len;
76 			u64	rdma_va;
77 		} rdma;
78 	};
79 
80 	struct mlx5_ib_pf_eq	*eq;
81 	struct work_struct	work;
82 };
83 
84 #define MAX_PREFETCH_LEN (4*1024*1024U)
85 
86 /* Timeout in ms to wait for an active mmu notifier to complete when handling
87  * a pagefault. */
88 #define MMU_NOTIFIER_TIMEOUT 1000
89 
90 #define MLX5_IMR_MTT_BITS (30 - PAGE_SHIFT)
91 #define MLX5_IMR_MTT_SHIFT (MLX5_IMR_MTT_BITS + PAGE_SHIFT)
92 #define MLX5_IMR_MTT_ENTRIES BIT_ULL(MLX5_IMR_MTT_BITS)
93 #define MLX5_IMR_MTT_SIZE BIT_ULL(MLX5_IMR_MTT_SHIFT)
94 #define MLX5_IMR_MTT_MASK (~(MLX5_IMR_MTT_SIZE - 1))
95 
96 #define MLX5_KSM_PAGE_SHIFT MLX5_IMR_MTT_SHIFT
97 
98 static u64 mlx5_imr_ksm_entries;
99 
100 static void populate_klm(struct mlx5_klm *pklm, size_t idx, size_t nentries,
101 			struct mlx5_ib_mr *imr, int flags)
102 {
103 	struct mlx5_klm *end = pklm + nentries;
104 
105 	if (flags & MLX5_IB_UPD_XLT_ZAP) {
106 		for (; pklm != end; pklm++, idx++) {
107 			pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
108 			pklm->key = cpu_to_be32(mr_to_mdev(imr)->null_mkey);
109 			pklm->va = 0;
110 		}
111 		return;
112 	}
113 
114 	/*
115 	 * The locking here is pretty subtle. Ideally the implicit_children
116 	 * xarray would be protected by the umem_mutex, however that is not
117 	 * possible. Instead this uses a weaker update-then-lock pattern:
118 	 *
119 	 *    xa_store()
120 	 *    mutex_lock(umem_mutex)
121 	 *     mlx5r_umr_update_xlt()
122 	 *    mutex_unlock(umem_mutex)
123 	 *    destroy lkey
124 	 *
125 	 * ie any change the xarray must be followed by the locked update_xlt
126 	 * before destroying.
127 	 *
128 	 * The umem_mutex provides the acquire/release semantic needed to make
129 	 * the xa_store() visible to a racing thread.
130 	 */
131 	lockdep_assert_held(&to_ib_umem_odp(imr->umem)->umem_mutex);
132 
133 	for (; pklm != end; pklm++, idx++) {
134 		struct mlx5_ib_mr *mtt = xa_load(&imr->implicit_children, idx);
135 
136 		pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
137 		if (mtt) {
138 			pklm->key = cpu_to_be32(mtt->ibmr.lkey);
139 			pklm->va = cpu_to_be64(idx * MLX5_IMR_MTT_SIZE);
140 		} else {
141 			pklm->key = cpu_to_be32(mr_to_mdev(imr)->null_mkey);
142 			pklm->va = 0;
143 		}
144 	}
145 }
146 
147 static u64 umem_dma_to_mtt(dma_addr_t umem_dma)
148 {
149 	u64 mtt_entry = umem_dma & ODP_DMA_ADDR_MASK;
150 
151 	if (umem_dma & ODP_READ_ALLOWED_BIT)
152 		mtt_entry |= MLX5_IB_MTT_READ;
153 	if (umem_dma & ODP_WRITE_ALLOWED_BIT)
154 		mtt_entry |= MLX5_IB_MTT_WRITE;
155 
156 	return mtt_entry;
157 }
158 
159 static void populate_mtt(__be64 *pas, size_t idx, size_t nentries,
160 			 struct mlx5_ib_mr *mr, int flags)
161 {
162 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
163 	dma_addr_t pa;
164 	size_t i;
165 
166 	if (flags & MLX5_IB_UPD_XLT_ZAP)
167 		return;
168 
169 	for (i = 0; i < nentries; i++) {
170 		pa = odp->dma_list[idx + i];
171 		pas[i] = cpu_to_be64(umem_dma_to_mtt(pa));
172 	}
173 }
174 
175 void mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
176 			   struct mlx5_ib_mr *mr, int flags)
177 {
178 	if (flags & MLX5_IB_UPD_XLT_INDIRECT) {
179 		populate_klm(xlt, idx, nentries, mr, flags);
180 	} else {
181 		populate_mtt(xlt, idx, nentries, mr, flags);
182 	}
183 }
184 
185 /*
186  * This must be called after the mr has been removed from implicit_children.
187  * NOTE: The MR does not necessarily have to be
188  * empty here, parallel page faults could have raced with the free process and
189  * added pages to it.
190  */
191 static void free_implicit_child_mr_work(struct work_struct *work)
192 {
193 	struct mlx5_ib_mr *mr =
194 		container_of(work, struct mlx5_ib_mr, odp_destroy.work);
195 	struct mlx5_ib_mr *imr = mr->parent;
196 	struct ib_umem_odp *odp_imr = to_ib_umem_odp(imr->umem);
197 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
198 
199 	mlx5r_deref_wait_odp_mkey(&mr->mmkey);
200 
201 	mutex_lock(&odp_imr->umem_mutex);
202 	mlx5r_umr_update_xlt(mr->parent,
203 			     ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT, 1, 0,
204 			     MLX5_IB_UPD_XLT_INDIRECT | MLX5_IB_UPD_XLT_ATOMIC);
205 	mutex_unlock(&odp_imr->umem_mutex);
206 	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
207 
208 	mlx5r_deref_odp_mkey(&imr->mmkey);
209 }
210 
211 static void destroy_unused_implicit_child_mr(struct mlx5_ib_mr *mr)
212 {
213 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
214 	unsigned long idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT;
215 	struct mlx5_ib_mr *imr = mr->parent;
216 
217 	if (!refcount_inc_not_zero(&imr->mmkey.usecount))
218 		return;
219 
220 	xa_erase(&imr->implicit_children, idx);
221 
222 	/* Freeing a MR is a sleeping operation, so bounce to a work queue */
223 	INIT_WORK(&mr->odp_destroy.work, free_implicit_child_mr_work);
224 	queue_work(system_unbound_wq, &mr->odp_destroy.work);
225 }
226 
227 static bool mlx5_ib_invalidate_range(struct mmu_interval_notifier *mni,
228 				     const struct mmu_notifier_range *range,
229 				     unsigned long cur_seq)
230 {
231 	struct ib_umem_odp *umem_odp =
232 		container_of(mni, struct ib_umem_odp, notifier);
233 	struct mlx5_ib_mr *mr;
234 	const u64 umr_block_mask = (MLX5_UMR_MTT_ALIGNMENT /
235 				    sizeof(struct mlx5_mtt)) - 1;
236 	u64 idx = 0, blk_start_idx = 0;
237 	u64 invalidations = 0;
238 	unsigned long start;
239 	unsigned long end;
240 	int in_block = 0;
241 	u64 addr;
242 
243 	if (!mmu_notifier_range_blockable(range))
244 		return false;
245 
246 	mutex_lock(&umem_odp->umem_mutex);
247 	mmu_interval_set_seq(mni, cur_seq);
248 	/*
249 	 * If npages is zero then umem_odp->private may not be setup yet. This
250 	 * does not complete until after the first page is mapped for DMA.
251 	 */
252 	if (!umem_odp->npages)
253 		goto out;
254 	mr = umem_odp->private;
255 
256 	start = max_t(u64, ib_umem_start(umem_odp), range->start);
257 	end = min_t(u64, ib_umem_end(umem_odp), range->end);
258 
259 	/*
260 	 * Iteration one - zap the HW's MTTs. The notifiers_count ensures that
261 	 * while we are doing the invalidation, no page fault will attempt to
262 	 * overwrite the same MTTs.  Concurent invalidations might race us,
263 	 * but they will write 0s as well, so no difference in the end result.
264 	 */
265 	for (addr = start; addr < end; addr += BIT(umem_odp->page_shift)) {
266 		idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
267 		/*
268 		 * Strive to write the MTTs in chunks, but avoid overwriting
269 		 * non-existing MTTs. The huristic here can be improved to
270 		 * estimate the cost of another UMR vs. the cost of bigger
271 		 * UMR.
272 		 */
273 		if (umem_odp->dma_list[idx] &
274 		    (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) {
275 			if (!in_block) {
276 				blk_start_idx = idx;
277 				in_block = 1;
278 			}
279 
280 			/* Count page invalidations */
281 			invalidations += idx - blk_start_idx + 1;
282 		} else {
283 			u64 umr_offset = idx & umr_block_mask;
284 
285 			if (in_block && umr_offset == 0) {
286 				mlx5r_umr_update_xlt(mr, blk_start_idx,
287 						     idx - blk_start_idx, 0,
288 						     MLX5_IB_UPD_XLT_ZAP |
289 						     MLX5_IB_UPD_XLT_ATOMIC);
290 				in_block = 0;
291 			}
292 		}
293 	}
294 	if (in_block)
295 		mlx5r_umr_update_xlt(mr, blk_start_idx,
296 				     idx - blk_start_idx + 1, 0,
297 				     MLX5_IB_UPD_XLT_ZAP |
298 				     MLX5_IB_UPD_XLT_ATOMIC);
299 
300 	mlx5_update_odp_stats(mr, invalidations, invalidations);
301 
302 	/*
303 	 * We are now sure that the device will not access the
304 	 * memory. We can safely unmap it, and mark it as dirty if
305 	 * needed.
306 	 */
307 
308 	ib_umem_odp_unmap_dma_pages(umem_odp, start, end);
309 
310 	if (unlikely(!umem_odp->npages && mr->parent))
311 		destroy_unused_implicit_child_mr(mr);
312 out:
313 	mutex_unlock(&umem_odp->umem_mutex);
314 	return true;
315 }
316 
317 const struct mmu_interval_notifier_ops mlx5_mn_ops = {
318 	.invalidate = mlx5_ib_invalidate_range,
319 };
320 
321 static void internal_fill_odp_caps(struct mlx5_ib_dev *dev)
322 {
323 	struct ib_odp_caps *caps = &dev->odp_caps;
324 
325 	memset(caps, 0, sizeof(*caps));
326 
327 	if (!MLX5_CAP_GEN(dev->mdev, pg) || !mlx5r_umr_can_load_pas(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 mlx5_ib_dev *dev = mr_to_mdev(imr);
411 	struct ib_umem_odp *odp;
412 	struct mlx5_ib_mr *mr;
413 	struct mlx5_ib_mr *ret;
414 	int err;
415 
416 	odp = ib_umem_odp_alloc_child(to_ib_umem_odp(imr->umem),
417 				      idx * MLX5_IMR_MTT_SIZE,
418 				      MLX5_IMR_MTT_SIZE, &mlx5_mn_ops);
419 	if (IS_ERR(odp))
420 		return ERR_CAST(odp);
421 
422 	mr = mlx5_mr_cache_alloc(dev, &dev->cache.ent[MLX5_IMR_MTT_CACHE_ENTRY],
423 				 imr->access_flags);
424 	if (IS_ERR(mr)) {
425 		ib_umem_odp_release(odp);
426 		return mr;
427 	}
428 
429 	mr->access_flags = imr->access_flags;
430 	mr->ibmr.pd = imr->ibmr.pd;
431 	mr->ibmr.device = &mr_to_mdev(imr)->ib_dev;
432 	mr->umem = &odp->umem;
433 	mr->ibmr.lkey = mr->mmkey.key;
434 	mr->ibmr.rkey = mr->mmkey.key;
435 	mr->ibmr.iova = idx * MLX5_IMR_MTT_SIZE;
436 	mr->parent = imr;
437 	odp->private = mr;
438 
439 	/*
440 	 * First refcount is owned by the xarray and second refconut
441 	 * is returned to the caller.
442 	 */
443 	refcount_set(&mr->mmkey.usecount, 2);
444 
445 	err = mlx5r_umr_update_xlt(mr, 0,
446 				   MLX5_IMR_MTT_ENTRIES,
447 				   PAGE_SHIFT,
448 				   MLX5_IB_UPD_XLT_ZAP |
449 				   MLX5_IB_UPD_XLT_ENABLE);
450 	if (err) {
451 		ret = ERR_PTR(err);
452 		goto out_mr;
453 	}
454 
455 	xa_lock(&imr->implicit_children);
456 	ret = __xa_cmpxchg(&imr->implicit_children, idx, NULL, mr,
457 			   GFP_KERNEL);
458 	if (unlikely(ret)) {
459 		if (xa_is_err(ret)) {
460 			ret = ERR_PTR(xa_err(ret));
461 			goto out_lock;
462 		}
463 		/*
464 		 * Another thread beat us to creating the child mr, use
465 		 * theirs.
466 		 */
467 		refcount_inc(&ret->mmkey.usecount);
468 		goto out_lock;
469 	}
470 	xa_unlock(&imr->implicit_children);
471 
472 	mlx5_ib_dbg(mr_to_mdev(imr), "key %x mr %p\n", mr->mmkey.key, mr);
473 	return mr;
474 
475 out_lock:
476 	xa_unlock(&imr->implicit_children);
477 out_mr:
478 	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
479 	return ret;
480 }
481 
482 struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
483 					     int access_flags)
484 {
485 	struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device);
486 	struct ib_umem_odp *umem_odp;
487 	struct mlx5_ib_mr *imr;
488 	int err;
489 
490 	if (!mlx5r_umr_can_load_pas(dev, MLX5_IMR_MTT_ENTRIES * PAGE_SIZE))
491 		return ERR_PTR(-EOPNOTSUPP);
492 
493 	umem_odp = ib_umem_odp_alloc_implicit(&dev->ib_dev, access_flags);
494 	if (IS_ERR(umem_odp))
495 		return ERR_CAST(umem_odp);
496 
497 	imr = mlx5_mr_cache_alloc(dev,
498 				  &dev->cache.ent[MLX5_IMR_KSM_CACHE_ENTRY],
499 				  access_flags);
500 	if (IS_ERR(imr)) {
501 		ib_umem_odp_release(umem_odp);
502 		return imr;
503 	}
504 
505 	imr->access_flags = access_flags;
506 	imr->ibmr.pd = &pd->ibpd;
507 	imr->ibmr.iova = 0;
508 	imr->umem = &umem_odp->umem;
509 	imr->ibmr.lkey = imr->mmkey.key;
510 	imr->ibmr.rkey = imr->mmkey.key;
511 	imr->ibmr.device = &dev->ib_dev;
512 	imr->is_odp_implicit = true;
513 	xa_init(&imr->implicit_children);
514 
515 	err = mlx5r_umr_update_xlt(imr, 0,
516 				   mlx5_imr_ksm_entries,
517 				   MLX5_KSM_PAGE_SHIFT,
518 				   MLX5_IB_UPD_XLT_INDIRECT |
519 				   MLX5_IB_UPD_XLT_ZAP |
520 				   MLX5_IB_UPD_XLT_ENABLE);
521 	if (err)
522 		goto out_mr;
523 
524 	err = mlx5r_store_odp_mkey(dev, &imr->mmkey);
525 	if (err)
526 		goto out_mr;
527 
528 	mlx5_ib_dbg(dev, "key %x mr %p\n", imr->mmkey.key, imr);
529 	return imr;
530 out_mr:
531 	mlx5_ib_err(dev, "Failed to register MKEY %d\n", err);
532 	mlx5_ib_dereg_mr(&imr->ibmr, NULL);
533 	return ERR_PTR(err);
534 }
535 
536 void mlx5_ib_free_odp_mr(struct mlx5_ib_mr *mr)
537 {
538 	struct mlx5_ib_mr *mtt;
539 	unsigned long idx;
540 
541 	/*
542 	 * If this is an implicit MR it is already invalidated so we can just
543 	 * delete the children mkeys.
544 	 */
545 	xa_for_each(&mr->implicit_children, idx, mtt) {
546 		xa_erase(&mr->implicit_children, idx);
547 		mlx5_ib_dereg_mr(&mtt->ibmr, NULL);
548 	}
549 }
550 
551 #define MLX5_PF_FLAGS_DOWNGRADE BIT(1)
552 #define MLX5_PF_FLAGS_SNAPSHOT BIT(2)
553 #define MLX5_PF_FLAGS_ENABLE BIT(3)
554 static int pagefault_real_mr(struct mlx5_ib_mr *mr, struct ib_umem_odp *odp,
555 			     u64 user_va, size_t bcnt, u32 *bytes_mapped,
556 			     u32 flags)
557 {
558 	int page_shift, ret, np;
559 	bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE;
560 	u64 access_mask;
561 	u64 start_idx;
562 	bool fault = !(flags & MLX5_PF_FLAGS_SNAPSHOT);
563 	u32 xlt_flags = MLX5_IB_UPD_XLT_ATOMIC;
564 
565 	if (flags & MLX5_PF_FLAGS_ENABLE)
566 		xlt_flags |= MLX5_IB_UPD_XLT_ENABLE;
567 
568 	page_shift = odp->page_shift;
569 	start_idx = (user_va - ib_umem_start(odp)) >> page_shift;
570 	access_mask = ODP_READ_ALLOWED_BIT;
571 
572 	if (odp->umem.writable && !downgrade)
573 		access_mask |= ODP_WRITE_ALLOWED_BIT;
574 
575 	np = ib_umem_odp_map_dma_and_lock(odp, user_va, bcnt, access_mask, fault);
576 	if (np < 0)
577 		return np;
578 
579 	/*
580 	 * No need to check whether the MTTs really belong to this MR, since
581 	 * ib_umem_odp_map_dma_and_lock already checks this.
582 	 */
583 	ret = mlx5r_umr_update_xlt(mr, start_idx, np, page_shift, xlt_flags);
584 	mutex_unlock(&odp->umem_mutex);
585 
586 	if (ret < 0) {
587 		if (ret != -EAGAIN)
588 			mlx5_ib_err(mr_to_mdev(mr),
589 				    "Failed to update mkey page tables\n");
590 		goto out;
591 	}
592 
593 	if (bytes_mapped) {
594 		u32 new_mappings = (np << page_shift) -
595 			(user_va - round_down(user_va, 1 << page_shift));
596 
597 		*bytes_mapped += min_t(u32, new_mappings, bcnt);
598 	}
599 
600 	return np << (page_shift - PAGE_SHIFT);
601 
602 out:
603 	return ret;
604 }
605 
606 static int pagefault_implicit_mr(struct mlx5_ib_mr *imr,
607 				 struct ib_umem_odp *odp_imr, u64 user_va,
608 				 size_t bcnt, u32 *bytes_mapped, u32 flags)
609 {
610 	unsigned long end_idx = (user_va + bcnt - 1) >> MLX5_IMR_MTT_SHIFT;
611 	unsigned long upd_start_idx = end_idx + 1;
612 	unsigned long upd_len = 0;
613 	unsigned long npages = 0;
614 	int err;
615 	int ret;
616 
617 	if (unlikely(user_va >= mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE ||
618 		     mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE - user_va < bcnt))
619 		return -EFAULT;
620 
621 	/* Fault each child mr that intersects with our interval. */
622 	while (bcnt) {
623 		unsigned long idx = user_va >> MLX5_IMR_MTT_SHIFT;
624 		struct ib_umem_odp *umem_odp;
625 		struct mlx5_ib_mr *mtt;
626 		u64 len;
627 
628 		xa_lock(&imr->implicit_children);
629 		mtt = xa_load(&imr->implicit_children, idx);
630 		if (unlikely(!mtt)) {
631 			xa_unlock(&imr->implicit_children);
632 			mtt = implicit_get_child_mr(imr, idx);
633 			if (IS_ERR(mtt)) {
634 				ret = PTR_ERR(mtt);
635 				goto out;
636 			}
637 			upd_start_idx = min(upd_start_idx, idx);
638 			upd_len = idx - upd_start_idx + 1;
639 		} else {
640 			refcount_inc(&mtt->mmkey.usecount);
641 			xa_unlock(&imr->implicit_children);
642 		}
643 
644 		umem_odp = to_ib_umem_odp(mtt->umem);
645 		len = min_t(u64, user_va + bcnt, ib_umem_end(umem_odp)) -
646 		      user_va;
647 
648 		ret = pagefault_real_mr(mtt, umem_odp, user_va, len,
649 					bytes_mapped, flags);
650 
651 		mlx5r_deref_odp_mkey(&mtt->mmkey);
652 
653 		if (ret < 0)
654 			goto out;
655 		user_va += len;
656 		bcnt -= len;
657 		npages += ret;
658 	}
659 
660 	ret = npages;
661 
662 	/*
663 	 * Any time the implicit_children are changed we must perform an
664 	 * update of the xlt before exiting to ensure the HW and the
665 	 * implicit_children remains synchronized.
666 	 */
667 out:
668 	if (likely(!upd_len))
669 		return ret;
670 
671 	/*
672 	 * Notice this is not strictly ordered right, the KSM is updated after
673 	 * the implicit_children is updated, so a parallel page fault could
674 	 * see a MR that is not yet visible in the KSM.  This is similar to a
675 	 * parallel page fault seeing a MR that is being concurrently removed
676 	 * from the KSM. Both of these improbable situations are resolved
677 	 * safely by resuming the HW and then taking another page fault. The
678 	 * next pagefault handler will see the new information.
679 	 */
680 	mutex_lock(&odp_imr->umem_mutex);
681 	err = mlx5r_umr_update_xlt(imr, upd_start_idx, upd_len, 0,
682 				   MLX5_IB_UPD_XLT_INDIRECT |
683 					  MLX5_IB_UPD_XLT_ATOMIC);
684 	mutex_unlock(&odp_imr->umem_mutex);
685 	if (err) {
686 		mlx5_ib_err(mr_to_mdev(imr), "Failed to update PAS\n");
687 		return err;
688 	}
689 	return ret;
690 }
691 
692 static int pagefault_dmabuf_mr(struct mlx5_ib_mr *mr, size_t bcnt,
693 			       u32 *bytes_mapped, u32 flags)
694 {
695 	struct ib_umem_dmabuf *umem_dmabuf = to_ib_umem_dmabuf(mr->umem);
696 	u32 xlt_flags = 0;
697 	int err;
698 	unsigned int page_size;
699 
700 	if (flags & MLX5_PF_FLAGS_ENABLE)
701 		xlt_flags |= MLX5_IB_UPD_XLT_ENABLE;
702 
703 	dma_resv_lock(umem_dmabuf->attach->dmabuf->resv, NULL);
704 	err = ib_umem_dmabuf_map_pages(umem_dmabuf);
705 	if (err) {
706 		dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
707 		return err;
708 	}
709 
710 	page_size = mlx5_umem_find_best_pgsz(&umem_dmabuf->umem, mkc,
711 					     log_page_size, 0,
712 					     umem_dmabuf->umem.iova);
713 	if (unlikely(page_size < PAGE_SIZE)) {
714 		ib_umem_dmabuf_unmap_pages(umem_dmabuf);
715 		err = -EINVAL;
716 	} else {
717 		err = mlx5r_umr_update_mr_pas(mr, xlt_flags);
718 	}
719 	dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
720 
721 	if (err)
722 		return err;
723 
724 	if (bytes_mapped)
725 		*bytes_mapped += bcnt;
726 
727 	return ib_umem_num_pages(mr->umem);
728 }
729 
730 /*
731  * Returns:
732  *  -EFAULT: The io_virt->bcnt is not within the MR, it covers pages that are
733  *           not accessible, or the MR is no longer valid.
734  *  -EAGAIN/-ENOMEM: The operation should be retried
735  *
736  *  -EINVAL/others: General internal malfunction
737  *  >0: Number of pages mapped
738  */
739 static int pagefault_mr(struct mlx5_ib_mr *mr, u64 io_virt, size_t bcnt,
740 			u32 *bytes_mapped, u32 flags)
741 {
742 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
743 
744 	if (unlikely(io_virt < mr->ibmr.iova))
745 		return -EFAULT;
746 
747 	if (mr->umem->is_dmabuf)
748 		return pagefault_dmabuf_mr(mr, bcnt, bytes_mapped, flags);
749 
750 	if (!odp->is_implicit_odp) {
751 		u64 user_va;
752 
753 		if (check_add_overflow(io_virt - mr->ibmr.iova,
754 				       (u64)odp->umem.address, &user_va))
755 			return -EFAULT;
756 		if (unlikely(user_va >= ib_umem_end(odp) ||
757 			     ib_umem_end(odp) - user_va < bcnt))
758 			return -EFAULT;
759 		return pagefault_real_mr(mr, odp, user_va, bcnt, bytes_mapped,
760 					 flags);
761 	}
762 	return pagefault_implicit_mr(mr, odp, io_virt, bcnt, bytes_mapped,
763 				     flags);
764 }
765 
766 int mlx5_ib_init_odp_mr(struct mlx5_ib_mr *mr)
767 {
768 	int ret;
769 
770 	ret = pagefault_real_mr(mr, to_ib_umem_odp(mr->umem), mr->umem->address,
771 				mr->umem->length, NULL,
772 				MLX5_PF_FLAGS_SNAPSHOT | MLX5_PF_FLAGS_ENABLE);
773 	return ret >= 0 ? 0 : ret;
774 }
775 
776 int mlx5_ib_init_dmabuf_mr(struct mlx5_ib_mr *mr)
777 {
778 	int ret;
779 
780 	ret = pagefault_dmabuf_mr(mr, mr->umem->length, NULL,
781 				  MLX5_PF_FLAGS_ENABLE);
782 
783 	return ret >= 0 ? 0 : ret;
784 }
785 
786 struct pf_frame {
787 	struct pf_frame *next;
788 	u32 key;
789 	u64 io_virt;
790 	size_t bcnt;
791 	int depth;
792 };
793 
794 static bool mkey_is_eq(struct mlx5_ib_mkey *mmkey, u32 key)
795 {
796 	if (!mmkey)
797 		return false;
798 	if (mmkey->type == MLX5_MKEY_MW)
799 		return mlx5_base_mkey(mmkey->key) == mlx5_base_mkey(key);
800 	return mmkey->key == key;
801 }
802 
803 /*
804  * Handle a single data segment in a page-fault WQE or RDMA region.
805  *
806  * Returns number of OS pages retrieved on success. The caller may continue to
807  * the next data segment.
808  * Can return the following error codes:
809  * -EAGAIN to designate a temporary error. The caller will abort handling the
810  *  page fault and resolve it.
811  * -EFAULT when there's an error mapping the requested pages. The caller will
812  *  abort the page fault handling.
813  */
814 static int pagefault_single_data_segment(struct mlx5_ib_dev *dev,
815 					 struct ib_pd *pd, u32 key,
816 					 u64 io_virt, size_t bcnt,
817 					 u32 *bytes_committed,
818 					 u32 *bytes_mapped)
819 {
820 	int npages = 0, ret, i, outlen, cur_outlen = 0, depth = 0;
821 	struct pf_frame *head = NULL, *frame;
822 	struct mlx5_ib_mkey *mmkey;
823 	struct mlx5_ib_mr *mr;
824 	struct mlx5_klm *pklm;
825 	u32 *out = NULL;
826 	size_t offset;
827 
828 	io_virt += *bytes_committed;
829 	bcnt -= *bytes_committed;
830 
831 next_mr:
832 	xa_lock(&dev->odp_mkeys);
833 	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(key));
834 	if (!mmkey) {
835 		xa_unlock(&dev->odp_mkeys);
836 		mlx5_ib_dbg(
837 			dev,
838 			"skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
839 			key);
840 		if (bytes_mapped)
841 			*bytes_mapped += bcnt;
842 		/*
843 		 * The user could specify a SGL with multiple lkeys and only
844 		 * some of them are ODP. Treat the non-ODP ones as fully
845 		 * faulted.
846 		 */
847 		ret = 0;
848 		goto end;
849 	}
850 	refcount_inc(&mmkey->usecount);
851 	xa_unlock(&dev->odp_mkeys);
852 
853 	if (!mkey_is_eq(mmkey, key)) {
854 		mlx5_ib_dbg(dev, "failed to find mkey %x\n", key);
855 		ret = -EFAULT;
856 		goto end;
857 	}
858 
859 	switch (mmkey->type) {
860 	case MLX5_MKEY_MR:
861 		mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
862 
863 		ret = pagefault_mr(mr, io_virt, bcnt, bytes_mapped, 0);
864 		if (ret < 0)
865 			goto end;
866 
867 		mlx5_update_odp_stats(mr, faults, ret);
868 
869 		npages += ret;
870 		ret = 0;
871 		break;
872 
873 	case MLX5_MKEY_MW:
874 	case MLX5_MKEY_INDIRECT_DEVX:
875 		if (depth >= MLX5_CAP_GEN(dev->mdev, max_indirection)) {
876 			mlx5_ib_dbg(dev, "indirection level exceeded\n");
877 			ret = -EFAULT;
878 			goto end;
879 		}
880 
881 		outlen = MLX5_ST_SZ_BYTES(query_mkey_out) +
882 			sizeof(*pklm) * (mmkey->ndescs - 2);
883 
884 		if (outlen > cur_outlen) {
885 			kfree(out);
886 			out = kzalloc(outlen, GFP_KERNEL);
887 			if (!out) {
888 				ret = -ENOMEM;
889 				goto end;
890 			}
891 			cur_outlen = outlen;
892 		}
893 
894 		pklm = (struct mlx5_klm *)MLX5_ADDR_OF(query_mkey_out, out,
895 						       bsf0_klm0_pas_mtt0_1);
896 
897 		ret = mlx5_core_query_mkey(dev->mdev, mmkey->key, out, outlen);
898 		if (ret)
899 			goto end;
900 
901 		offset = io_virt - MLX5_GET64(query_mkey_out, out,
902 					      memory_key_mkey_entry.start_addr);
903 
904 		for (i = 0; bcnt && i < mmkey->ndescs; i++, pklm++) {
905 			if (offset >= be32_to_cpu(pklm->bcount)) {
906 				offset -= be32_to_cpu(pklm->bcount);
907 				continue;
908 			}
909 
910 			frame = kzalloc(sizeof(*frame), GFP_KERNEL);
911 			if (!frame) {
912 				ret = -ENOMEM;
913 				goto end;
914 			}
915 
916 			frame->key = be32_to_cpu(pklm->key);
917 			frame->io_virt = be64_to_cpu(pklm->va) + offset;
918 			frame->bcnt = min_t(size_t, bcnt,
919 					    be32_to_cpu(pklm->bcount) - offset);
920 			frame->depth = depth + 1;
921 			frame->next = head;
922 			head = frame;
923 
924 			bcnt -= frame->bcnt;
925 			offset = 0;
926 		}
927 		break;
928 
929 	default:
930 		mlx5_ib_dbg(dev, "wrong mkey type %d\n", mmkey->type);
931 		ret = -EFAULT;
932 		goto end;
933 	}
934 
935 	if (head) {
936 		frame = head;
937 		head = frame->next;
938 
939 		key = frame->key;
940 		io_virt = frame->io_virt;
941 		bcnt = frame->bcnt;
942 		depth = frame->depth;
943 		kfree(frame);
944 
945 		mlx5r_deref_odp_mkey(mmkey);
946 		goto next_mr;
947 	}
948 
949 end:
950 	if (mmkey)
951 		mlx5r_deref_odp_mkey(mmkey);
952 	while (head) {
953 		frame = head;
954 		head = frame->next;
955 		kfree(frame);
956 	}
957 	kfree(out);
958 
959 	*bytes_committed = 0;
960 	return ret ? ret : npages;
961 }
962 
963 /*
964  * Parse a series of data segments for page fault handling.
965  *
966  * @dev:  Pointer to mlx5 IB device
967  * @pfault: contains page fault information.
968  * @wqe: points at the first data segment in the WQE.
969  * @wqe_end: points after the end of the WQE.
970  * @bytes_mapped: receives the number of bytes that the function was able to
971  *                map. This allows the caller to decide intelligently whether
972  *                enough memory was mapped to resolve the page fault
973  *                successfully (e.g. enough for the next MTU, or the entire
974  *                WQE).
975  * @total_wqe_bytes: receives the total data size of this WQE in bytes (minus
976  *                   the committed bytes).
977  * @receive_queue: receive WQE end of sg list
978  *
979  * Returns the number of pages loaded if positive, zero for an empty WQE, or a
980  * negative error code.
981  */
982 static int pagefault_data_segments(struct mlx5_ib_dev *dev,
983 				   struct mlx5_pagefault *pfault,
984 				   void *wqe,
985 				   void *wqe_end, u32 *bytes_mapped,
986 				   u32 *total_wqe_bytes, bool receive_queue)
987 {
988 	int ret = 0, npages = 0;
989 	u64 io_virt;
990 	u32 key;
991 	u32 byte_count;
992 	size_t bcnt;
993 	int inline_segment;
994 
995 	if (bytes_mapped)
996 		*bytes_mapped = 0;
997 	if (total_wqe_bytes)
998 		*total_wqe_bytes = 0;
999 
1000 	while (wqe < wqe_end) {
1001 		struct mlx5_wqe_data_seg *dseg = wqe;
1002 
1003 		io_virt = be64_to_cpu(dseg->addr);
1004 		key = be32_to_cpu(dseg->lkey);
1005 		byte_count = be32_to_cpu(dseg->byte_count);
1006 		inline_segment = !!(byte_count &  MLX5_INLINE_SEG);
1007 		bcnt	       = byte_count & ~MLX5_INLINE_SEG;
1008 
1009 		if (inline_segment) {
1010 			bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK;
1011 			wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt,
1012 				     16);
1013 		} else {
1014 			wqe += sizeof(*dseg);
1015 		}
1016 
1017 		/* receive WQE end of sg list. */
1018 		if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY &&
1019 		    io_virt == 0)
1020 			break;
1021 
1022 		if (!inline_segment && total_wqe_bytes) {
1023 			*total_wqe_bytes += bcnt - min_t(size_t, bcnt,
1024 					pfault->bytes_committed);
1025 		}
1026 
1027 		/* A zero length data segment designates a length of 2GB. */
1028 		if (bcnt == 0)
1029 			bcnt = 1U << 31;
1030 
1031 		if (inline_segment || bcnt <= pfault->bytes_committed) {
1032 			pfault->bytes_committed -=
1033 				min_t(size_t, bcnt,
1034 				      pfault->bytes_committed);
1035 			continue;
1036 		}
1037 
1038 		ret = pagefault_single_data_segment(dev, NULL, key,
1039 						    io_virt, bcnt,
1040 						    &pfault->bytes_committed,
1041 						    bytes_mapped);
1042 		if (ret < 0)
1043 			break;
1044 		npages += ret;
1045 	}
1046 
1047 	return ret < 0 ? ret : npages;
1048 }
1049 
1050 /*
1051  * Parse initiator WQE. Advances the wqe pointer to point at the
1052  * scatter-gather list, and set wqe_end to the end of the WQE.
1053  */
1054 static int mlx5_ib_mr_initiator_pfault_handler(
1055 	struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault,
1056 	struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length)
1057 {
1058 	struct mlx5_wqe_ctrl_seg *ctrl = *wqe;
1059 	u16 wqe_index = pfault->wqe.wqe_index;
1060 	struct mlx5_base_av *av;
1061 	unsigned ds, opcode;
1062 	u32 qpn = qp->trans_qp.base.mqp.qpn;
1063 
1064 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
1065 	if (ds * MLX5_WQE_DS_UNITS > wqe_length) {
1066 		mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n",
1067 			    ds, wqe_length);
1068 		return -EFAULT;
1069 	}
1070 
1071 	if (ds == 0) {
1072 		mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n",
1073 			    wqe_index, qpn);
1074 		return -EFAULT;
1075 	}
1076 
1077 	*wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS;
1078 	*wqe += sizeof(*ctrl);
1079 
1080 	opcode = be32_to_cpu(ctrl->opmod_idx_opcode) &
1081 		 MLX5_WQE_CTRL_OPCODE_MASK;
1082 
1083 	if (qp->type == IB_QPT_XRC_INI)
1084 		*wqe += sizeof(struct mlx5_wqe_xrc_seg);
1085 
1086 	if (qp->type == IB_QPT_UD || qp->type == MLX5_IB_QPT_DCI) {
1087 		av = *wqe;
1088 		if (av->dqp_dct & cpu_to_be32(MLX5_EXTENDED_UD_AV))
1089 			*wqe += sizeof(struct mlx5_av);
1090 		else
1091 			*wqe += sizeof(struct mlx5_base_av);
1092 	}
1093 
1094 	switch (opcode) {
1095 	case MLX5_OPCODE_RDMA_WRITE:
1096 	case MLX5_OPCODE_RDMA_WRITE_IMM:
1097 	case MLX5_OPCODE_RDMA_READ:
1098 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1099 		break;
1100 	case MLX5_OPCODE_ATOMIC_CS:
1101 	case MLX5_OPCODE_ATOMIC_FA:
1102 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1103 		*wqe += sizeof(struct mlx5_wqe_atomic_seg);
1104 		break;
1105 	}
1106 
1107 	return 0;
1108 }
1109 
1110 /*
1111  * Parse responder WQE and set wqe_end to the end of the WQE.
1112  */
1113 static int mlx5_ib_mr_responder_pfault_handler_srq(struct mlx5_ib_dev *dev,
1114 						   struct mlx5_ib_srq *srq,
1115 						   void **wqe, void **wqe_end,
1116 						   int wqe_length)
1117 {
1118 	int wqe_size = 1 << srq->msrq.wqe_shift;
1119 
1120 	if (wqe_size > wqe_length) {
1121 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1122 		return -EFAULT;
1123 	}
1124 
1125 	*wqe_end = *wqe + wqe_size;
1126 	*wqe += sizeof(struct mlx5_wqe_srq_next_seg);
1127 
1128 	return 0;
1129 }
1130 
1131 static int mlx5_ib_mr_responder_pfault_handler_rq(struct mlx5_ib_dev *dev,
1132 						  struct mlx5_ib_qp *qp,
1133 						  void *wqe, void **wqe_end,
1134 						  int wqe_length)
1135 {
1136 	struct mlx5_ib_wq *wq = &qp->rq;
1137 	int wqe_size = 1 << wq->wqe_shift;
1138 
1139 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE) {
1140 		mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n");
1141 		return -EFAULT;
1142 	}
1143 
1144 	if (wqe_size > wqe_length) {
1145 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1146 		return -EFAULT;
1147 	}
1148 
1149 	*wqe_end = wqe + wqe_size;
1150 
1151 	return 0;
1152 }
1153 
1154 static inline struct mlx5_core_rsc_common *odp_get_rsc(struct mlx5_ib_dev *dev,
1155 						       u32 wq_num, int pf_type)
1156 {
1157 	struct mlx5_core_rsc_common *common = NULL;
1158 	struct mlx5_core_srq *srq;
1159 
1160 	switch (pf_type) {
1161 	case MLX5_WQE_PF_TYPE_RMP:
1162 		srq = mlx5_cmd_get_srq(dev, wq_num);
1163 		if (srq)
1164 			common = &srq->common;
1165 		break;
1166 	case MLX5_WQE_PF_TYPE_REQ_SEND_OR_WRITE:
1167 	case MLX5_WQE_PF_TYPE_RESP:
1168 	case MLX5_WQE_PF_TYPE_REQ_READ_OR_ATOMIC:
1169 		common = mlx5_core_res_hold(dev, wq_num, MLX5_RES_QP);
1170 		break;
1171 	default:
1172 		break;
1173 	}
1174 
1175 	return common;
1176 }
1177 
1178 static inline struct mlx5_ib_qp *res_to_qp(struct mlx5_core_rsc_common *res)
1179 {
1180 	struct mlx5_core_qp *mqp = (struct mlx5_core_qp *)res;
1181 
1182 	return to_mibqp(mqp);
1183 }
1184 
1185 static inline struct mlx5_ib_srq *res_to_srq(struct mlx5_core_rsc_common *res)
1186 {
1187 	struct mlx5_core_srq *msrq =
1188 		container_of(res, struct mlx5_core_srq, common);
1189 
1190 	return to_mibsrq(msrq);
1191 }
1192 
1193 static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
1194 					  struct mlx5_pagefault *pfault)
1195 {
1196 	bool sq = pfault->type & MLX5_PFAULT_REQUESTOR;
1197 	u16 wqe_index = pfault->wqe.wqe_index;
1198 	void *wqe, *wqe_start = NULL, *wqe_end = NULL;
1199 	u32 bytes_mapped, total_wqe_bytes;
1200 	struct mlx5_core_rsc_common *res;
1201 	int resume_with_error = 1;
1202 	struct mlx5_ib_qp *qp;
1203 	size_t bytes_copied;
1204 	int ret = 0;
1205 
1206 	res = odp_get_rsc(dev, pfault->wqe.wq_num, pfault->type);
1207 	if (!res) {
1208 		mlx5_ib_dbg(dev, "wqe page fault for missing resource %d\n", pfault->wqe.wq_num);
1209 		return;
1210 	}
1211 
1212 	if (res->res != MLX5_RES_QP && res->res != MLX5_RES_SRQ &&
1213 	    res->res != MLX5_RES_XSRQ) {
1214 		mlx5_ib_err(dev, "wqe page fault for unsupported type %d\n",
1215 			    pfault->type);
1216 		goto resolve_page_fault;
1217 	}
1218 
1219 	wqe_start = (void *)__get_free_page(GFP_KERNEL);
1220 	if (!wqe_start) {
1221 		mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n");
1222 		goto resolve_page_fault;
1223 	}
1224 
1225 	wqe = wqe_start;
1226 	qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL;
1227 	if (qp && sq) {
1228 		ret = mlx5_ib_read_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE,
1229 					  &bytes_copied);
1230 		if (ret)
1231 			goto read_user;
1232 		ret = mlx5_ib_mr_initiator_pfault_handler(
1233 			dev, pfault, qp, &wqe, &wqe_end, bytes_copied);
1234 	} else if (qp && !sq) {
1235 		ret = mlx5_ib_read_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE,
1236 					  &bytes_copied);
1237 		if (ret)
1238 			goto read_user;
1239 		ret = mlx5_ib_mr_responder_pfault_handler_rq(
1240 			dev, qp, wqe, &wqe_end, bytes_copied);
1241 	} else if (!qp) {
1242 		struct mlx5_ib_srq *srq = res_to_srq(res);
1243 
1244 		ret = mlx5_ib_read_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE,
1245 					   &bytes_copied);
1246 		if (ret)
1247 			goto read_user;
1248 		ret = mlx5_ib_mr_responder_pfault_handler_srq(
1249 			dev, srq, &wqe, &wqe_end, bytes_copied);
1250 	}
1251 
1252 	if (ret < 0 || wqe >= wqe_end)
1253 		goto resolve_page_fault;
1254 
1255 	ret = pagefault_data_segments(dev, pfault, wqe, wqe_end, &bytes_mapped,
1256 				      &total_wqe_bytes, !sq);
1257 	if (ret == -EAGAIN)
1258 		goto out;
1259 
1260 	if (ret < 0 || total_wqe_bytes > bytes_mapped)
1261 		goto resolve_page_fault;
1262 
1263 out:
1264 	ret = 0;
1265 	resume_with_error = 0;
1266 
1267 read_user:
1268 	if (ret)
1269 		mlx5_ib_err(
1270 			dev,
1271 			"Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n",
1272 			ret, wqe_index, pfault->token);
1273 
1274 resolve_page_fault:
1275 	mlx5_ib_page_fault_resume(dev, pfault, resume_with_error);
1276 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n",
1277 		    pfault->wqe.wq_num, resume_with_error,
1278 		    pfault->type);
1279 	mlx5_core_res_put(res);
1280 	free_page((unsigned long)wqe_start);
1281 }
1282 
1283 static int pages_in_range(u64 address, u32 length)
1284 {
1285 	return (ALIGN(address + length, PAGE_SIZE) -
1286 		(address & PAGE_MASK)) >> PAGE_SHIFT;
1287 }
1288 
1289 static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev,
1290 					   struct mlx5_pagefault *pfault)
1291 {
1292 	u64 address;
1293 	u32 length;
1294 	u32 prefetch_len = pfault->bytes_committed;
1295 	int prefetch_activated = 0;
1296 	u32 rkey = pfault->rdma.r_key;
1297 	int ret;
1298 
1299 	/* The RDMA responder handler handles the page fault in two parts.
1300 	 * First it brings the necessary pages for the current packet
1301 	 * (and uses the pfault context), and then (after resuming the QP)
1302 	 * prefetches more pages. The second operation cannot use the pfault
1303 	 * context and therefore uses the dummy_pfault context allocated on
1304 	 * the stack */
1305 	pfault->rdma.rdma_va += pfault->bytes_committed;
1306 	pfault->rdma.rdma_op_len -= min(pfault->bytes_committed,
1307 					 pfault->rdma.rdma_op_len);
1308 	pfault->bytes_committed = 0;
1309 
1310 	address = pfault->rdma.rdma_va;
1311 	length  = pfault->rdma.rdma_op_len;
1312 
1313 	/* For some operations, the hardware cannot tell the exact message
1314 	 * length, and in those cases it reports zero. Use prefetch
1315 	 * logic. */
1316 	if (length == 0) {
1317 		prefetch_activated = 1;
1318 		length = pfault->rdma.packet_size;
1319 		prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len);
1320 	}
1321 
1322 	ret = pagefault_single_data_segment(dev, NULL, rkey, address, length,
1323 					    &pfault->bytes_committed, NULL);
1324 	if (ret == -EAGAIN) {
1325 		/* We're racing with an invalidation, don't prefetch */
1326 		prefetch_activated = 0;
1327 	} else if (ret < 0 || pages_in_range(address, length) > ret) {
1328 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1329 		if (ret != -ENOENT)
1330 			mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n",
1331 				    ret, pfault->token, pfault->type);
1332 		return;
1333 	}
1334 
1335 	mlx5_ib_page_fault_resume(dev, pfault, 0);
1336 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n",
1337 		    pfault->token, pfault->type,
1338 		    prefetch_activated);
1339 
1340 	/* At this point, there might be a new pagefault already arriving in
1341 	 * the eq, switch to the dummy pagefault for the rest of the
1342 	 * processing. We're still OK with the objects being alive as the
1343 	 * work-queue is being fenced. */
1344 
1345 	if (prefetch_activated) {
1346 		u32 bytes_committed = 0;
1347 
1348 		ret = pagefault_single_data_segment(dev, NULL, rkey, address,
1349 						    prefetch_len,
1350 						    &bytes_committed, NULL);
1351 		if (ret < 0 && ret != -EAGAIN) {
1352 			mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n",
1353 				    ret, pfault->token, address, prefetch_len);
1354 		}
1355 	}
1356 }
1357 
1358 static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault)
1359 {
1360 	u8 event_subtype = pfault->event_subtype;
1361 
1362 	switch (event_subtype) {
1363 	case MLX5_PFAULT_SUBTYPE_WQE:
1364 		mlx5_ib_mr_wqe_pfault_handler(dev, pfault);
1365 		break;
1366 	case MLX5_PFAULT_SUBTYPE_RDMA:
1367 		mlx5_ib_mr_rdma_pfault_handler(dev, pfault);
1368 		break;
1369 	default:
1370 		mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n",
1371 			    event_subtype);
1372 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1373 	}
1374 }
1375 
1376 static void mlx5_ib_eqe_pf_action(struct work_struct *work)
1377 {
1378 	struct mlx5_pagefault *pfault = container_of(work,
1379 						     struct mlx5_pagefault,
1380 						     work);
1381 	struct mlx5_ib_pf_eq *eq = pfault->eq;
1382 
1383 	mlx5_ib_pfault(eq->dev, pfault);
1384 	mempool_free(pfault, eq->pool);
1385 }
1386 
1387 static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq)
1388 {
1389 	struct mlx5_eqe_page_fault *pf_eqe;
1390 	struct mlx5_pagefault *pfault;
1391 	struct mlx5_eqe *eqe;
1392 	int cc = 0;
1393 
1394 	while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) {
1395 		pfault = mempool_alloc(eq->pool, GFP_ATOMIC);
1396 		if (!pfault) {
1397 			schedule_work(&eq->work);
1398 			break;
1399 		}
1400 
1401 		pf_eqe = &eqe->data.page_fault;
1402 		pfault->event_subtype = eqe->sub_type;
1403 		pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);
1404 
1405 		mlx5_ib_dbg(eq->dev,
1406 			    "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
1407 			    eqe->sub_type, pfault->bytes_committed);
1408 
1409 		switch (eqe->sub_type) {
1410 		case MLX5_PFAULT_SUBTYPE_RDMA:
1411 			/* RDMA based event */
1412 			pfault->type =
1413 				be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
1414 			pfault->token =
1415 				be32_to_cpu(pf_eqe->rdma.pftype_token) &
1416 				MLX5_24BIT_MASK;
1417 			pfault->rdma.r_key =
1418 				be32_to_cpu(pf_eqe->rdma.r_key);
1419 			pfault->rdma.packet_size =
1420 				be16_to_cpu(pf_eqe->rdma.packet_length);
1421 			pfault->rdma.rdma_op_len =
1422 				be32_to_cpu(pf_eqe->rdma.rdma_op_len);
1423 			pfault->rdma.rdma_va =
1424 				be64_to_cpu(pf_eqe->rdma.rdma_va);
1425 			mlx5_ib_dbg(eq->dev,
1426 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
1427 				    pfault->type, pfault->token,
1428 				    pfault->rdma.r_key);
1429 			mlx5_ib_dbg(eq->dev,
1430 				    "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
1431 				    pfault->rdma.rdma_op_len,
1432 				    pfault->rdma.rdma_va);
1433 			break;
1434 
1435 		case MLX5_PFAULT_SUBTYPE_WQE:
1436 			/* WQE based event */
1437 			pfault->type =
1438 				(be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7;
1439 			pfault->token =
1440 				be32_to_cpu(pf_eqe->wqe.token);
1441 			pfault->wqe.wq_num =
1442 				be32_to_cpu(pf_eqe->wqe.pftype_wq) &
1443 				MLX5_24BIT_MASK;
1444 			pfault->wqe.wqe_index =
1445 				be16_to_cpu(pf_eqe->wqe.wqe_index);
1446 			pfault->wqe.packet_size =
1447 				be16_to_cpu(pf_eqe->wqe.packet_length);
1448 			mlx5_ib_dbg(eq->dev,
1449 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
1450 				    pfault->type, pfault->token,
1451 				    pfault->wqe.wq_num,
1452 				    pfault->wqe.wqe_index);
1453 			break;
1454 
1455 		default:
1456 			mlx5_ib_warn(eq->dev,
1457 				     "Unsupported page fault event sub-type: 0x%02hhx\n",
1458 				     eqe->sub_type);
1459 			/* Unsupported page faults should still be
1460 			 * resolved by the page fault handler
1461 			 */
1462 		}
1463 
1464 		pfault->eq = eq;
1465 		INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action);
1466 		queue_work(eq->wq, &pfault->work);
1467 
1468 		cc = mlx5_eq_update_cc(eq->core, ++cc);
1469 	}
1470 
1471 	mlx5_eq_update_ci(eq->core, cc, 1);
1472 }
1473 
1474 static int mlx5_ib_eq_pf_int(struct notifier_block *nb, unsigned long type,
1475 			     void *data)
1476 {
1477 	struct mlx5_ib_pf_eq *eq =
1478 		container_of(nb, struct mlx5_ib_pf_eq, irq_nb);
1479 	unsigned long flags;
1480 
1481 	if (spin_trylock_irqsave(&eq->lock, flags)) {
1482 		mlx5_ib_eq_pf_process(eq);
1483 		spin_unlock_irqrestore(&eq->lock, flags);
1484 	} else {
1485 		schedule_work(&eq->work);
1486 	}
1487 
1488 	return IRQ_HANDLED;
1489 }
1490 
1491 /* mempool_refill() was proposed but unfortunately wasn't accepted
1492  * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
1493  * Cheap workaround.
1494  */
1495 static void mempool_refill(mempool_t *pool)
1496 {
1497 	while (pool->curr_nr < pool->min_nr)
1498 		mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
1499 }
1500 
1501 static void mlx5_ib_eq_pf_action(struct work_struct *work)
1502 {
1503 	struct mlx5_ib_pf_eq *eq =
1504 		container_of(work, struct mlx5_ib_pf_eq, work);
1505 
1506 	mempool_refill(eq->pool);
1507 
1508 	spin_lock_irq(&eq->lock);
1509 	mlx5_ib_eq_pf_process(eq);
1510 	spin_unlock_irq(&eq->lock);
1511 }
1512 
1513 enum {
1514 	MLX5_IB_NUM_PF_EQE	= 0x1000,
1515 	MLX5_IB_NUM_PF_DRAIN	= 64,
1516 };
1517 
1518 int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1519 {
1520 	struct mlx5_eq_param param = {};
1521 	int err = 0;
1522 
1523 	mutex_lock(&dev->odp_eq_mutex);
1524 	if (eq->core)
1525 		goto unlock;
1526 	INIT_WORK(&eq->work, mlx5_ib_eq_pf_action);
1527 	spin_lock_init(&eq->lock);
1528 	eq->dev = dev;
1529 
1530 	eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN,
1531 					       sizeof(struct mlx5_pagefault));
1532 	if (!eq->pool) {
1533 		err = -ENOMEM;
1534 		goto unlock;
1535 	}
1536 
1537 	eq->wq = alloc_workqueue("mlx5_ib_page_fault",
1538 				 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM,
1539 				 MLX5_NUM_CMD_EQE);
1540 	if (!eq->wq) {
1541 		err = -ENOMEM;
1542 		goto err_mempool;
1543 	}
1544 
1545 	eq->irq_nb.notifier_call = mlx5_ib_eq_pf_int;
1546 	param = (struct mlx5_eq_param) {
1547 		.nent = MLX5_IB_NUM_PF_EQE,
1548 	};
1549 	param.mask[0] = 1ull << MLX5_EVENT_TYPE_PAGE_FAULT;
1550 	eq->core = mlx5_eq_create_generic(dev->mdev, &param);
1551 	if (IS_ERR(eq->core)) {
1552 		err = PTR_ERR(eq->core);
1553 		goto err_wq;
1554 	}
1555 	err = mlx5_eq_enable(dev->mdev, eq->core, &eq->irq_nb);
1556 	if (err) {
1557 		mlx5_ib_err(dev, "failed to enable odp EQ %d\n", err);
1558 		goto err_eq;
1559 	}
1560 
1561 	mutex_unlock(&dev->odp_eq_mutex);
1562 	return 0;
1563 err_eq:
1564 	mlx5_eq_destroy_generic(dev->mdev, eq->core);
1565 err_wq:
1566 	eq->core = NULL;
1567 	destroy_workqueue(eq->wq);
1568 err_mempool:
1569 	mempool_destroy(eq->pool);
1570 unlock:
1571 	mutex_unlock(&dev->odp_eq_mutex);
1572 	return err;
1573 }
1574 
1575 static int
1576 mlx5_ib_odp_destroy_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1577 {
1578 	int err;
1579 
1580 	if (!eq->core)
1581 		return 0;
1582 	mlx5_eq_disable(dev->mdev, eq->core, &eq->irq_nb);
1583 	err = mlx5_eq_destroy_generic(dev->mdev, eq->core);
1584 	cancel_work_sync(&eq->work);
1585 	destroy_workqueue(eq->wq);
1586 	mempool_destroy(eq->pool);
1587 
1588 	return err;
1589 }
1590 
1591 void mlx5_odp_init_mkey_cache_entry(struct mlx5_cache_ent *ent)
1592 {
1593 	if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1594 		return;
1595 
1596 	switch (ent->order - 2) {
1597 	case MLX5_IMR_MTT_CACHE_ENTRY:
1598 		ent->page = PAGE_SHIFT;
1599 		ent->ndescs = MLX5_IMR_MTT_ENTRIES;
1600 		ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
1601 		ent->limit = 0;
1602 		break;
1603 
1604 	case MLX5_IMR_KSM_CACHE_ENTRY:
1605 		ent->page = MLX5_KSM_PAGE_SHIFT;
1606 		ent->ndescs = mlx5_imr_ksm_entries;
1607 		ent->access_mode = MLX5_MKC_ACCESS_MODE_KSM;
1608 		ent->limit = 0;
1609 		break;
1610 	}
1611 }
1612 
1613 static const struct ib_device_ops mlx5_ib_dev_odp_ops = {
1614 	.advise_mr = mlx5_ib_advise_mr,
1615 };
1616 
1617 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
1618 {
1619 	int ret = 0;
1620 
1621 	internal_fill_odp_caps(dev);
1622 
1623 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1624 		return ret;
1625 
1626 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_odp_ops);
1627 
1628 	if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) {
1629 		ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey);
1630 		if (ret) {
1631 			mlx5_ib_err(dev, "Error getting null_mkey %d\n", ret);
1632 			return ret;
1633 		}
1634 	}
1635 
1636 	mutex_init(&dev->odp_eq_mutex);
1637 	return ret;
1638 }
1639 
1640 void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev)
1641 {
1642 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1643 		return;
1644 
1645 	mlx5_ib_odp_destroy_eq(dev, &dev->odp_pf_eq);
1646 }
1647 
1648 int mlx5_ib_odp_init(void)
1649 {
1650 	mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) -
1651 				       MLX5_IMR_MTT_BITS);
1652 
1653 	return 0;
1654 }
1655 
1656 struct prefetch_mr_work {
1657 	struct work_struct work;
1658 	u32 pf_flags;
1659 	u32 num_sge;
1660 	struct {
1661 		u64 io_virt;
1662 		struct mlx5_ib_mr *mr;
1663 		size_t length;
1664 	} frags[];
1665 };
1666 
1667 static void destroy_prefetch_work(struct prefetch_mr_work *work)
1668 {
1669 	u32 i;
1670 
1671 	for (i = 0; i < work->num_sge; ++i)
1672 		mlx5r_deref_odp_mkey(&work->frags[i].mr->mmkey);
1673 
1674 	kvfree(work);
1675 }
1676 
1677 static struct mlx5_ib_mr *
1678 get_prefetchable_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
1679 		    u32 lkey)
1680 {
1681 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1682 	struct mlx5_ib_mr *mr = NULL;
1683 	struct mlx5_ib_mkey *mmkey;
1684 
1685 	xa_lock(&dev->odp_mkeys);
1686 	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(lkey));
1687 	if (!mmkey || mmkey->key != lkey) {
1688 		mr = ERR_PTR(-ENOENT);
1689 		goto end;
1690 	}
1691 	if (mmkey->type != MLX5_MKEY_MR) {
1692 		mr = ERR_PTR(-EINVAL);
1693 		goto end;
1694 	}
1695 
1696 	mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
1697 
1698 	if (mr->ibmr.pd != pd) {
1699 		mr = ERR_PTR(-EPERM);
1700 		goto end;
1701 	}
1702 
1703 	/* prefetch with write-access must be supported by the MR */
1704 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
1705 	    !mr->umem->writable) {
1706 		mr = ERR_PTR(-EPERM);
1707 		goto end;
1708 	}
1709 
1710 	refcount_inc(&mmkey->usecount);
1711 end:
1712 	xa_unlock(&dev->odp_mkeys);
1713 	return mr;
1714 }
1715 
1716 static void mlx5_ib_prefetch_mr_work(struct work_struct *w)
1717 {
1718 	struct prefetch_mr_work *work =
1719 		container_of(w, struct prefetch_mr_work, work);
1720 	u32 bytes_mapped = 0;
1721 	int ret;
1722 	u32 i;
1723 
1724 	/* We rely on IB/core that work is executed if we have num_sge != 0 only. */
1725 	WARN_ON(!work->num_sge);
1726 	for (i = 0; i < work->num_sge; ++i) {
1727 		ret = pagefault_mr(work->frags[i].mr, work->frags[i].io_virt,
1728 				   work->frags[i].length, &bytes_mapped,
1729 				   work->pf_flags);
1730 		if (ret <= 0)
1731 			continue;
1732 		mlx5_update_odp_stats(work->frags[i].mr, prefetch, ret);
1733 	}
1734 
1735 	destroy_prefetch_work(work);
1736 }
1737 
1738 static int init_prefetch_work(struct ib_pd *pd,
1739 			       enum ib_uverbs_advise_mr_advice advice,
1740 			       u32 pf_flags, struct prefetch_mr_work *work,
1741 			       struct ib_sge *sg_list, u32 num_sge)
1742 {
1743 	u32 i;
1744 
1745 	INIT_WORK(&work->work, mlx5_ib_prefetch_mr_work);
1746 	work->pf_flags = pf_flags;
1747 
1748 	for (i = 0; i < num_sge; ++i) {
1749 		struct mlx5_ib_mr *mr;
1750 
1751 		mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1752 		if (IS_ERR(mr)) {
1753 			work->num_sge = i;
1754 			return PTR_ERR(mr);
1755 		}
1756 		work->frags[i].io_virt = sg_list[i].addr;
1757 		work->frags[i].length = sg_list[i].length;
1758 		work->frags[i].mr = mr;
1759 	}
1760 	work->num_sge = num_sge;
1761 	return 0;
1762 }
1763 
1764 static int mlx5_ib_prefetch_sg_list(struct ib_pd *pd,
1765 				    enum ib_uverbs_advise_mr_advice advice,
1766 				    u32 pf_flags, struct ib_sge *sg_list,
1767 				    u32 num_sge)
1768 {
1769 	u32 bytes_mapped = 0;
1770 	int ret = 0;
1771 	u32 i;
1772 
1773 	for (i = 0; i < num_sge; ++i) {
1774 		struct mlx5_ib_mr *mr;
1775 
1776 		mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1777 		if (IS_ERR(mr))
1778 			return PTR_ERR(mr);
1779 		ret = pagefault_mr(mr, sg_list[i].addr, sg_list[i].length,
1780 				   &bytes_mapped, pf_flags);
1781 		if (ret < 0) {
1782 			mlx5r_deref_odp_mkey(&mr->mmkey);
1783 			return ret;
1784 		}
1785 		mlx5_update_odp_stats(mr, prefetch, ret);
1786 		mlx5r_deref_odp_mkey(&mr->mmkey);
1787 	}
1788 
1789 	return 0;
1790 }
1791 
1792 int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
1793 			       enum ib_uverbs_advise_mr_advice advice,
1794 			       u32 flags, struct ib_sge *sg_list, u32 num_sge)
1795 {
1796 	u32 pf_flags = 0;
1797 	struct prefetch_mr_work *work;
1798 	int rc;
1799 
1800 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH)
1801 		pf_flags |= MLX5_PF_FLAGS_DOWNGRADE;
1802 
1803 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT)
1804 		pf_flags |= MLX5_PF_FLAGS_SNAPSHOT;
1805 
1806 	if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH)
1807 		return mlx5_ib_prefetch_sg_list(pd, advice, pf_flags, sg_list,
1808 						num_sge);
1809 
1810 	work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL);
1811 	if (!work)
1812 		return -ENOMEM;
1813 
1814 	rc = init_prefetch_work(pd, advice, pf_flags, work, sg_list, num_sge);
1815 	if (rc) {
1816 		destroy_prefetch_work(work);
1817 		return rc;
1818 	}
1819 	queue_work(system_unbound_wq, &work->work);
1820 	return 0;
1821 }
1822