xref: /linux/drivers/infiniband/hw/mlx5/odp.c (revision a44e4f3ab16bc808590763a543a93b6fbf3abcc4)
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 
37 #include "mlx5_ib.h"
38 #include "cmd.h"
39 
40 #include <linux/mlx5/eq.h>
41 
42 /* Contains the details of a pagefault. */
43 struct mlx5_pagefault {
44 	u32			bytes_committed;
45 	u32			token;
46 	u8			event_subtype;
47 	u8			type;
48 	union {
49 		/* Initiator or send message responder pagefault details. */
50 		struct {
51 			/* Received packet size, only valid for responders. */
52 			u32	packet_size;
53 			/*
54 			 * Number of resource holding WQE, depends on type.
55 			 */
56 			u32	wq_num;
57 			/*
58 			 * WQE index. Refers to either the send queue or
59 			 * receive queue, according to event_subtype.
60 			 */
61 			u16	wqe_index;
62 		} wqe;
63 		/* RDMA responder pagefault details */
64 		struct {
65 			u32	r_key;
66 			/*
67 			 * Received packet size, minimal size page fault
68 			 * resolution required for forward progress.
69 			 */
70 			u32	packet_size;
71 			u32	rdma_op_len;
72 			u64	rdma_va;
73 		} rdma;
74 	};
75 
76 	struct mlx5_ib_pf_eq	*eq;
77 	struct work_struct	work;
78 };
79 
80 #define MAX_PREFETCH_LEN (4*1024*1024U)
81 
82 /* Timeout in ms to wait for an active mmu notifier to complete when handling
83  * a pagefault. */
84 #define MMU_NOTIFIER_TIMEOUT 1000
85 
86 #define MLX5_IMR_MTT_BITS (30 - PAGE_SHIFT)
87 #define MLX5_IMR_MTT_SHIFT (MLX5_IMR_MTT_BITS + PAGE_SHIFT)
88 #define MLX5_IMR_MTT_ENTRIES BIT_ULL(MLX5_IMR_MTT_BITS)
89 #define MLX5_IMR_MTT_SIZE BIT_ULL(MLX5_IMR_MTT_SHIFT)
90 #define MLX5_IMR_MTT_MASK (~(MLX5_IMR_MTT_SIZE - 1))
91 
92 #define MLX5_KSM_PAGE_SHIFT MLX5_IMR_MTT_SHIFT
93 
94 static u64 mlx5_imr_ksm_entries;
95 
96 static int check_parent(struct ib_umem_odp *odp,
97 			       struct mlx5_ib_mr *parent)
98 {
99 	struct mlx5_ib_mr *mr = odp->private;
100 
101 	return mr && mr->parent == parent && !odp->dying;
102 }
103 
104 static struct ib_ucontext_per_mm *mr_to_per_mm(struct mlx5_ib_mr *mr)
105 {
106 	if (WARN_ON(!mr || !is_odp_mr(mr)))
107 		return NULL;
108 
109 	return to_ib_umem_odp(mr->umem)->per_mm;
110 }
111 
112 static struct ib_umem_odp *odp_next(struct ib_umem_odp *odp)
113 {
114 	struct mlx5_ib_mr *mr = odp->private, *parent = mr->parent;
115 	struct ib_ucontext_per_mm *per_mm = odp->per_mm;
116 	struct rb_node *rb;
117 
118 	down_read(&per_mm->umem_rwsem);
119 	while (1) {
120 		rb = rb_next(&odp->interval_tree.rb);
121 		if (!rb)
122 			goto not_found;
123 		odp = rb_entry(rb, struct ib_umem_odp, interval_tree.rb);
124 		if (check_parent(odp, parent))
125 			goto end;
126 	}
127 not_found:
128 	odp = NULL;
129 end:
130 	up_read(&per_mm->umem_rwsem);
131 	return odp;
132 }
133 
134 static struct ib_umem_odp *odp_lookup(u64 start, u64 length,
135 				      struct mlx5_ib_mr *parent)
136 {
137 	struct ib_ucontext_per_mm *per_mm = mr_to_per_mm(parent);
138 	struct ib_umem_odp *odp;
139 	struct rb_node *rb;
140 
141 	down_read(&per_mm->umem_rwsem);
142 	odp = rbt_ib_umem_lookup(&per_mm->umem_tree, start, length);
143 	if (!odp)
144 		goto end;
145 
146 	while (1) {
147 		if (check_parent(odp, parent))
148 			goto end;
149 		rb = rb_next(&odp->interval_tree.rb);
150 		if (!rb)
151 			goto not_found;
152 		odp = rb_entry(rb, struct ib_umem_odp, interval_tree.rb);
153 		if (ib_umem_start(odp) > start + length)
154 			goto not_found;
155 	}
156 not_found:
157 	odp = NULL;
158 end:
159 	up_read(&per_mm->umem_rwsem);
160 	return odp;
161 }
162 
163 void mlx5_odp_populate_klm(struct mlx5_klm *pklm, size_t offset,
164 			   size_t nentries, struct mlx5_ib_mr *mr, int flags)
165 {
166 	struct ib_pd *pd = mr->ibmr.pd;
167 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
168 	struct ib_umem_odp *odp;
169 	unsigned long va;
170 	int i;
171 
172 	if (flags & MLX5_IB_UPD_XLT_ZAP) {
173 		for (i = 0; i < nentries; i++, pklm++) {
174 			pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
175 			pklm->key = cpu_to_be32(dev->null_mkey);
176 			pklm->va = 0;
177 		}
178 		return;
179 	}
180 
181 	/*
182 	 * The locking here is pretty subtle. Ideally the implicit children
183 	 * list would be protected by the umem_mutex, however that is not
184 	 * possible. Instead this uses a weaker update-then-lock pattern:
185 	 *
186 	 *  srcu_read_lock()
187 	 *    <change children list>
188 	 *    mutex_lock(umem_mutex)
189 	 *     mlx5_ib_update_xlt()
190 	 *    mutex_unlock(umem_mutex)
191 	 *    destroy lkey
192 	 *
193 	 * ie any change the children list must be followed by the locked
194 	 * update_xlt before destroying.
195 	 *
196 	 * The umem_mutex provides the acquire/release semantic needed to make
197 	 * the children list visible to a racing thread. While SRCU is not
198 	 * technically required, using it gives consistent use of the SRCU
199 	 * locking around the children list.
200 	 */
201 	lockdep_assert_held(&to_ib_umem_odp(mr->umem)->umem_mutex);
202 	lockdep_assert_held(&mr->dev->mr_srcu);
203 
204 	odp = odp_lookup(offset * MLX5_IMR_MTT_SIZE,
205 			 nentries * MLX5_IMR_MTT_SIZE, mr);
206 
207 	for (i = 0; i < nentries; i++, pklm++) {
208 		pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
209 		va = (offset + i) * MLX5_IMR_MTT_SIZE;
210 		if (odp && ib_umem_start(odp) == va) {
211 			struct mlx5_ib_mr *mtt = odp->private;
212 
213 			pklm->key = cpu_to_be32(mtt->ibmr.lkey);
214 			odp = odp_next(odp);
215 		} else {
216 			pklm->key = cpu_to_be32(dev->null_mkey);
217 		}
218 		mlx5_ib_dbg(dev, "[%d] va %lx key %x\n",
219 			    i, va, be32_to_cpu(pklm->key));
220 	}
221 }
222 
223 static void mr_leaf_free_action(struct work_struct *work)
224 {
225 	struct ib_umem_odp *odp = container_of(work, struct ib_umem_odp, work);
226 	int idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT;
227 	struct mlx5_ib_mr *mr = odp->private, *imr = mr->parent;
228 	struct ib_umem_odp *odp_imr = to_ib_umem_odp(imr->umem);
229 	int srcu_key;
230 
231 	mr->parent = NULL;
232 	synchronize_srcu(&mr->dev->mr_srcu);
233 
234 	if (smp_load_acquire(&imr->live)) {
235 		srcu_key = srcu_read_lock(&mr->dev->mr_srcu);
236 		mutex_lock(&odp_imr->umem_mutex);
237 		mlx5_ib_update_xlt(imr, idx, 1, 0,
238 				   MLX5_IB_UPD_XLT_INDIRECT |
239 				   MLX5_IB_UPD_XLT_ATOMIC);
240 		mutex_unlock(&odp_imr->umem_mutex);
241 		srcu_read_unlock(&mr->dev->mr_srcu, srcu_key);
242 	}
243 	ib_umem_odp_release(odp);
244 	mlx5_mr_cache_free(mr->dev, mr);
245 
246 	if (atomic_dec_and_test(&imr->num_leaf_free))
247 		wake_up(&imr->q_leaf_free);
248 }
249 
250 void mlx5_ib_invalidate_range(struct ib_umem_odp *umem_odp, unsigned long start,
251 			      unsigned long end)
252 {
253 	struct mlx5_ib_mr *mr;
254 	const u64 umr_block_mask = (MLX5_UMR_MTT_ALIGNMENT /
255 				    sizeof(struct mlx5_mtt)) - 1;
256 	u64 idx = 0, blk_start_idx = 0;
257 	int in_block = 0;
258 	u64 addr;
259 
260 	if (!umem_odp) {
261 		pr_err("invalidation called on NULL umem or non-ODP umem\n");
262 		return;
263 	}
264 
265 	mr = umem_odp->private;
266 
267 	if (!mr || !mr->ibmr.pd)
268 		return;
269 
270 	start = max_t(u64, ib_umem_start(umem_odp), start);
271 	end = min_t(u64, ib_umem_end(umem_odp), end);
272 
273 	/*
274 	 * Iteration one - zap the HW's MTTs. The notifiers_count ensures that
275 	 * while we are doing the invalidation, no page fault will attempt to
276 	 * overwrite the same MTTs.  Concurent invalidations might race us,
277 	 * but they will write 0s as well, so no difference in the end result.
278 	 */
279 	mutex_lock(&umem_odp->umem_mutex);
280 	for (addr = start; addr < end; addr += BIT(umem_odp->page_shift)) {
281 		idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
282 		/*
283 		 * Strive to write the MTTs in chunks, but avoid overwriting
284 		 * non-existing MTTs. The huristic here can be improved to
285 		 * estimate the cost of another UMR vs. the cost of bigger
286 		 * UMR.
287 		 */
288 		if (umem_odp->dma_list[idx] &
289 		    (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) {
290 			if (!in_block) {
291 				blk_start_idx = idx;
292 				in_block = 1;
293 			}
294 		} else {
295 			u64 umr_offset = idx & umr_block_mask;
296 
297 			if (in_block && umr_offset == 0) {
298 				mlx5_ib_update_xlt(mr, blk_start_idx,
299 						   idx - blk_start_idx, 0,
300 						   MLX5_IB_UPD_XLT_ZAP |
301 						   MLX5_IB_UPD_XLT_ATOMIC);
302 				in_block = 0;
303 			}
304 		}
305 	}
306 	if (in_block)
307 		mlx5_ib_update_xlt(mr, blk_start_idx,
308 				   idx - blk_start_idx + 1, 0,
309 				   MLX5_IB_UPD_XLT_ZAP |
310 				   MLX5_IB_UPD_XLT_ATOMIC);
311 	/*
312 	 * We are now sure that the device will not access the
313 	 * memory. We can safely unmap it, and mark it as dirty if
314 	 * needed.
315 	 */
316 
317 	ib_umem_odp_unmap_dma_pages(umem_odp, start, end);
318 
319 	if (unlikely(!umem_odp->npages && mr->parent &&
320 		     !umem_odp->dying)) {
321 		WRITE_ONCE(mr->live, 0);
322 		umem_odp->dying = 1;
323 		atomic_inc(&mr->parent->num_leaf_free);
324 		schedule_work(&umem_odp->work);
325 	}
326 	mutex_unlock(&umem_odp->umem_mutex);
327 }
328 
329 void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev)
330 {
331 	struct ib_odp_caps *caps = &dev->odp_caps;
332 
333 	memset(caps, 0, sizeof(*caps));
334 
335 	if (!MLX5_CAP_GEN(dev->mdev, pg) ||
336 	    !mlx5_ib_can_use_umr(dev, true))
337 		return;
338 
339 	caps->general_caps = IB_ODP_SUPPORT;
340 
341 	if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
342 		dev->odp_max_size = U64_MAX;
343 	else
344 		dev->odp_max_size = BIT_ULL(MLX5_MAX_UMR_SHIFT + PAGE_SHIFT);
345 
346 	if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.send))
347 		caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND;
348 
349 	if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.srq_receive))
350 		caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
351 
352 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.send))
353 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND;
354 
355 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.receive))
356 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV;
357 
358 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.write))
359 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE;
360 
361 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.read))
362 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ;
363 
364 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.atomic))
365 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
366 
367 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.srq_receive))
368 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
369 
370 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.send))
371 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SEND;
372 
373 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.receive))
374 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_RECV;
375 
376 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.write))
377 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_WRITE;
378 
379 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.read))
380 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_READ;
381 
382 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.atomic))
383 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
384 
385 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.srq_receive))
386 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
387 
388 	if (MLX5_CAP_GEN(dev->mdev, fixed_buffer_size) &&
389 	    MLX5_CAP_GEN(dev->mdev, null_mkey) &&
390 	    MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset) &&
391 	    !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled))
392 		caps->general_caps |= IB_ODP_SUPPORT_IMPLICIT;
393 
394 	return;
395 }
396 
397 static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev,
398 				      struct mlx5_pagefault *pfault,
399 				      int error)
400 {
401 	int wq_num = pfault->event_subtype == MLX5_PFAULT_SUBTYPE_WQE ?
402 		     pfault->wqe.wq_num : pfault->token;
403 	u32 out[MLX5_ST_SZ_DW(page_fault_resume_out)] = { };
404 	u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)]   = { };
405 	int err;
406 
407 	MLX5_SET(page_fault_resume_in, in, opcode, MLX5_CMD_OP_PAGE_FAULT_RESUME);
408 	MLX5_SET(page_fault_resume_in, in, page_fault_type, pfault->type);
409 	MLX5_SET(page_fault_resume_in, in, token, pfault->token);
410 	MLX5_SET(page_fault_resume_in, in, wq_number, wq_num);
411 	MLX5_SET(page_fault_resume_in, in, error, !!error);
412 
413 	err = mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
414 	if (err)
415 		mlx5_ib_err(dev, "Failed to resolve the page fault on WQ 0x%x err %d\n",
416 			    wq_num, err);
417 }
418 
419 static struct mlx5_ib_mr *implicit_mr_alloc(struct ib_pd *pd,
420 					    struct ib_umem_odp *umem_odp,
421 					    bool ksm, int access_flags)
422 {
423 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
424 	struct mlx5_ib_mr *mr;
425 	int err;
426 
427 	mr = mlx5_mr_cache_alloc(dev, ksm ? MLX5_IMR_KSM_CACHE_ENTRY :
428 					    MLX5_IMR_MTT_CACHE_ENTRY);
429 
430 	if (IS_ERR(mr))
431 		return mr;
432 
433 	mr->ibmr.pd = pd;
434 
435 	mr->dev = dev;
436 	mr->access_flags = access_flags;
437 	mr->mmkey.iova = 0;
438 	mr->umem = &umem_odp->umem;
439 
440 	if (ksm) {
441 		err = mlx5_ib_update_xlt(mr, 0,
442 					 mlx5_imr_ksm_entries,
443 					 MLX5_KSM_PAGE_SHIFT,
444 					 MLX5_IB_UPD_XLT_INDIRECT |
445 					 MLX5_IB_UPD_XLT_ZAP |
446 					 MLX5_IB_UPD_XLT_ENABLE);
447 
448 	} else {
449 		err = mlx5_ib_update_xlt(mr, 0,
450 					 MLX5_IMR_MTT_ENTRIES,
451 					 PAGE_SHIFT,
452 					 MLX5_IB_UPD_XLT_ZAP |
453 					 MLX5_IB_UPD_XLT_ENABLE |
454 					 MLX5_IB_UPD_XLT_ATOMIC);
455 	}
456 
457 	if (err)
458 		goto fail;
459 
460 	mr->ibmr.lkey = mr->mmkey.key;
461 	mr->ibmr.rkey = mr->mmkey.key;
462 
463 	mlx5_ib_dbg(dev, "key %x dev %p mr %p\n",
464 		    mr->mmkey.key, dev->mdev, mr);
465 
466 	return mr;
467 
468 fail:
469 	mlx5_ib_err(dev, "Failed to register MKEY %d\n", err);
470 	mlx5_mr_cache_free(dev, mr);
471 
472 	return ERR_PTR(err);
473 }
474 
475 static struct ib_umem_odp *implicit_mr_get_data(struct mlx5_ib_mr *mr,
476 						u64 io_virt, size_t bcnt)
477 {
478 	struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.pd->device);
479 	struct ib_umem_odp *odp, *result = NULL;
480 	struct ib_umem_odp *odp_mr = to_ib_umem_odp(mr->umem);
481 	u64 addr = io_virt & MLX5_IMR_MTT_MASK;
482 	int nentries = 0, start_idx = 0, ret;
483 	struct mlx5_ib_mr *mtt;
484 
485 	mutex_lock(&odp_mr->umem_mutex);
486 	odp = odp_lookup(addr, 1, mr);
487 
488 	mlx5_ib_dbg(dev, "io_virt:%llx bcnt:%zx addr:%llx odp:%p\n",
489 		    io_virt, bcnt, addr, odp);
490 
491 next_mr:
492 	if (likely(odp)) {
493 		if (nentries)
494 			nentries++;
495 	} else {
496 		odp = ib_umem_odp_alloc_child(odp_mr, addr, MLX5_IMR_MTT_SIZE);
497 		if (IS_ERR(odp)) {
498 			mutex_unlock(&odp_mr->umem_mutex);
499 			return ERR_CAST(odp);
500 		}
501 
502 		mtt = implicit_mr_alloc(mr->ibmr.pd, odp, 0,
503 					mr->access_flags);
504 		if (IS_ERR(mtt)) {
505 			mutex_unlock(&odp_mr->umem_mutex);
506 			ib_umem_odp_release(odp);
507 			return ERR_CAST(mtt);
508 		}
509 
510 		odp->private = mtt;
511 		mtt->umem = &odp->umem;
512 		mtt->mmkey.iova = addr;
513 		mtt->parent = mr;
514 		INIT_WORK(&odp->work, mr_leaf_free_action);
515 
516 		smp_store_release(&mtt->live, 1);
517 
518 		if (!nentries)
519 			start_idx = addr >> MLX5_IMR_MTT_SHIFT;
520 		nentries++;
521 	}
522 
523 	/* Return first odp if region not covered by single one */
524 	if (likely(!result))
525 		result = odp;
526 
527 	addr += MLX5_IMR_MTT_SIZE;
528 	if (unlikely(addr < io_virt + bcnt)) {
529 		odp = odp_next(odp);
530 		if (odp && ib_umem_start(odp) != addr)
531 			odp = NULL;
532 		goto next_mr;
533 	}
534 
535 	if (unlikely(nentries)) {
536 		ret = mlx5_ib_update_xlt(mr, start_idx, nentries, 0,
537 					 MLX5_IB_UPD_XLT_INDIRECT |
538 					 MLX5_IB_UPD_XLT_ATOMIC);
539 		if (ret) {
540 			mlx5_ib_err(dev, "Failed to update PAS\n");
541 			result = ERR_PTR(ret);
542 		}
543 	}
544 
545 	mutex_unlock(&odp_mr->umem_mutex);
546 	return result;
547 }
548 
549 struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
550 					     struct ib_udata *udata,
551 					     int access_flags)
552 {
553 	struct mlx5_ib_mr *imr;
554 	struct ib_umem_odp *umem_odp;
555 
556 	umem_odp = ib_umem_odp_alloc_implicit(udata, access_flags);
557 	if (IS_ERR(umem_odp))
558 		return ERR_CAST(umem_odp);
559 
560 	imr = implicit_mr_alloc(&pd->ibpd, umem_odp, 1, access_flags);
561 	if (IS_ERR(imr)) {
562 		ib_umem_odp_release(umem_odp);
563 		return ERR_CAST(imr);
564 	}
565 
566 	imr->umem = &umem_odp->umem;
567 	init_waitqueue_head(&imr->q_leaf_free);
568 	atomic_set(&imr->num_leaf_free, 0);
569 	atomic_set(&imr->num_pending_prefetch, 0);
570 	smp_store_release(&imr->live, 1);
571 
572 	return imr;
573 }
574 
575 void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *imr)
576 {
577 	struct ib_ucontext_per_mm *per_mm = mr_to_per_mm(imr);
578 	struct rb_node *node;
579 
580 	down_read(&per_mm->umem_rwsem);
581 	for (node = rb_first_cached(&per_mm->umem_tree); node;
582 	     node = rb_next(node)) {
583 		struct ib_umem_odp *umem_odp =
584 			rb_entry(node, struct ib_umem_odp, interval_tree.rb);
585 		struct mlx5_ib_mr *mr = umem_odp->private;
586 
587 		if (mr->parent != imr)
588 			continue;
589 
590 		mutex_lock(&umem_odp->umem_mutex);
591 		ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
592 					    ib_umem_end(umem_odp));
593 
594 		if (umem_odp->dying) {
595 			mutex_unlock(&umem_odp->umem_mutex);
596 			continue;
597 		}
598 
599 		umem_odp->dying = 1;
600 		atomic_inc(&imr->num_leaf_free);
601 		schedule_work(&umem_odp->work);
602 		mutex_unlock(&umem_odp->umem_mutex);
603 	}
604 	up_read(&per_mm->umem_rwsem);
605 
606 	wait_event(imr->q_leaf_free, !atomic_read(&imr->num_leaf_free));
607 }
608 
609 #define MLX5_PF_FLAGS_PREFETCH  BIT(0)
610 #define MLX5_PF_FLAGS_DOWNGRADE BIT(1)
611 static int pagefault_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr,
612 			u64 io_virt, size_t bcnt, u32 *bytes_mapped,
613 			u32 flags)
614 {
615 	int npages = 0, current_seq, page_shift, ret, np;
616 	struct ib_umem_odp *odp_mr = to_ib_umem_odp(mr->umem);
617 	bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE;
618 	bool prefetch = flags & MLX5_PF_FLAGS_PREFETCH;
619 	u64 access_mask;
620 	u64 start_idx, page_mask;
621 	struct ib_umem_odp *odp;
622 	size_t size;
623 
624 	if (odp_mr->is_implicit_odp) {
625 		odp = implicit_mr_get_data(mr, io_virt, bcnt);
626 
627 		if (IS_ERR(odp))
628 			return PTR_ERR(odp);
629 		mr = odp->private;
630 	} else {
631 		odp = odp_mr;
632 	}
633 
634 next_mr:
635 	size = min_t(size_t, bcnt, ib_umem_end(odp) - io_virt);
636 
637 	page_shift = odp->page_shift;
638 	page_mask = ~(BIT(page_shift) - 1);
639 	start_idx = (io_virt - (mr->mmkey.iova & page_mask)) >> page_shift;
640 	access_mask = ODP_READ_ALLOWED_BIT;
641 
642 	if (prefetch && !downgrade && !odp->umem.writable) {
643 		/* prefetch with write-access must
644 		 * be supported by the MR
645 		 */
646 		ret = -EINVAL;
647 		goto out;
648 	}
649 
650 	if (odp->umem.writable && !downgrade)
651 		access_mask |= ODP_WRITE_ALLOWED_BIT;
652 
653 	current_seq = READ_ONCE(odp->notifiers_seq);
654 	/*
655 	 * Ensure the sequence number is valid for some time before we call
656 	 * gup.
657 	 */
658 	smp_rmb();
659 
660 	ret = ib_umem_odp_map_dma_pages(odp, io_virt, size, access_mask,
661 					current_seq);
662 
663 	if (ret < 0)
664 		goto out;
665 
666 	np = ret;
667 
668 	mutex_lock(&odp->umem_mutex);
669 	if (!ib_umem_mmu_notifier_retry(odp, current_seq)) {
670 		/*
671 		 * No need to check whether the MTTs really belong to
672 		 * this MR, since ib_umem_odp_map_dma_pages already
673 		 * checks this.
674 		 */
675 		ret = mlx5_ib_update_xlt(mr, start_idx, np,
676 					 page_shift, MLX5_IB_UPD_XLT_ATOMIC);
677 	} else {
678 		ret = -EAGAIN;
679 	}
680 	mutex_unlock(&odp->umem_mutex);
681 
682 	if (ret < 0) {
683 		if (ret != -EAGAIN)
684 			mlx5_ib_err(dev, "Failed to update mkey page tables\n");
685 		goto out;
686 	}
687 
688 	if (bytes_mapped) {
689 		u32 new_mappings = (np << page_shift) -
690 			(io_virt - round_down(io_virt, 1 << page_shift));
691 		*bytes_mapped += min_t(u32, new_mappings, size);
692 	}
693 
694 	npages += np << (page_shift - PAGE_SHIFT);
695 	bcnt -= size;
696 
697 	if (unlikely(bcnt)) {
698 		struct ib_umem_odp *next;
699 
700 		io_virt += size;
701 		next = odp_next(odp);
702 		if (unlikely(!next || ib_umem_start(next) != io_virt)) {
703 			mlx5_ib_dbg(dev, "next implicit leaf removed at 0x%llx. got %p\n",
704 				    io_virt, next);
705 			return -EAGAIN;
706 		}
707 		odp = next;
708 		mr = odp->private;
709 		goto next_mr;
710 	}
711 
712 	return npages;
713 
714 out:
715 	if (ret == -EAGAIN) {
716 		unsigned long timeout = msecs_to_jiffies(MMU_NOTIFIER_TIMEOUT);
717 
718 		if (!wait_for_completion_timeout(&odp->notifier_completion,
719 						 timeout)) {
720 			mlx5_ib_warn(
721 				dev,
722 				"timeout waiting for mmu notifier. seq %d against %d. notifiers_count=%d\n",
723 				current_seq, odp->notifiers_seq,
724 				odp->notifiers_count);
725 		}
726 	}
727 
728 	return ret;
729 }
730 
731 struct pf_frame {
732 	struct pf_frame *next;
733 	u32 key;
734 	u64 io_virt;
735 	size_t bcnt;
736 	int depth;
737 };
738 
739 static bool mkey_is_eq(struct mlx5_core_mkey *mmkey, u32 key)
740 {
741 	if (!mmkey)
742 		return false;
743 	if (mmkey->type == MLX5_MKEY_MW)
744 		return mlx5_base_mkey(mmkey->key) == mlx5_base_mkey(key);
745 	return mmkey->key == key;
746 }
747 
748 static int get_indirect_num_descs(struct mlx5_core_mkey *mmkey)
749 {
750 	struct mlx5_ib_mw *mw;
751 	struct mlx5_ib_devx_mr *devx_mr;
752 
753 	if (mmkey->type == MLX5_MKEY_MW) {
754 		mw = container_of(mmkey, struct mlx5_ib_mw, mmkey);
755 		return mw->ndescs;
756 	}
757 
758 	devx_mr = container_of(mmkey, struct mlx5_ib_devx_mr,
759 			       mmkey);
760 	return devx_mr->ndescs;
761 }
762 
763 /*
764  * Handle a single data segment in a page-fault WQE or RDMA region.
765  *
766  * Returns number of OS pages retrieved on success. The caller may continue to
767  * the next data segment.
768  * Can return the following error codes:
769  * -EAGAIN to designate a temporary error. The caller will abort handling the
770  *  page fault and resolve it.
771  * -EFAULT when there's an error mapping the requested pages. The caller will
772  *  abort the page fault handling.
773  */
774 static int pagefault_single_data_segment(struct mlx5_ib_dev *dev,
775 					 struct ib_pd *pd, u32 key,
776 					 u64 io_virt, size_t bcnt,
777 					 u32 *bytes_committed,
778 					 u32 *bytes_mapped, u32 flags)
779 {
780 	int npages = 0, srcu_key, ret, i, outlen, cur_outlen = 0, depth = 0;
781 	bool prefetch = flags & MLX5_PF_FLAGS_PREFETCH;
782 	struct pf_frame *head = NULL, *frame;
783 	struct mlx5_core_mkey *mmkey;
784 	struct mlx5_ib_mr *mr;
785 	struct mlx5_klm *pklm;
786 	u32 *out = NULL;
787 	size_t offset;
788 	int ndescs;
789 
790 	srcu_key = srcu_read_lock(&dev->mr_srcu);
791 
792 	io_virt += *bytes_committed;
793 	bcnt -= *bytes_committed;
794 
795 next_mr:
796 	mmkey = xa_load(&dev->mdev->priv.mkey_table, mlx5_base_mkey(key));
797 	if (!mkey_is_eq(mmkey, key)) {
798 		mlx5_ib_dbg(dev, "failed to find mkey %x\n", key);
799 		ret = -EFAULT;
800 		goto srcu_unlock;
801 	}
802 
803 	if (prefetch && mmkey->type != MLX5_MKEY_MR) {
804 		mlx5_ib_dbg(dev, "prefetch is allowed only for MR\n");
805 		ret = -EINVAL;
806 		goto srcu_unlock;
807 	}
808 
809 	switch (mmkey->type) {
810 	case MLX5_MKEY_MR:
811 		mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
812 		if (!smp_load_acquire(&mr->live) || !mr->ibmr.pd) {
813 			mlx5_ib_dbg(dev, "got dead MR\n");
814 			ret = -EFAULT;
815 			goto srcu_unlock;
816 		}
817 
818 		if (prefetch) {
819 			if (!is_odp_mr(mr) ||
820 			    mr->ibmr.pd != pd) {
821 				mlx5_ib_dbg(dev, "Invalid prefetch request: %s\n",
822 					    is_odp_mr(mr) ?  "MR is not ODP" :
823 					    "PD is not of the MR");
824 				ret = -EINVAL;
825 				goto srcu_unlock;
826 			}
827 		}
828 
829 		if (!is_odp_mr(mr)) {
830 			mlx5_ib_dbg(dev, "skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
831 				    key);
832 			if (bytes_mapped)
833 				*bytes_mapped += bcnt;
834 			ret = 0;
835 			goto srcu_unlock;
836 		}
837 
838 		ret = pagefault_mr(dev, mr, io_virt, bcnt, bytes_mapped, flags);
839 		if (ret < 0)
840 			goto srcu_unlock;
841 
842 		npages += ret;
843 		ret = 0;
844 		break;
845 
846 	case MLX5_MKEY_MW:
847 	case MLX5_MKEY_INDIRECT_DEVX:
848 		ndescs = get_indirect_num_descs(mmkey);
849 
850 		if (depth >= MLX5_CAP_GEN(dev->mdev, max_indirection)) {
851 			mlx5_ib_dbg(dev, "indirection level exceeded\n");
852 			ret = -EFAULT;
853 			goto srcu_unlock;
854 		}
855 
856 		outlen = MLX5_ST_SZ_BYTES(query_mkey_out) +
857 			sizeof(*pklm) * (ndescs - 2);
858 
859 		if (outlen > cur_outlen) {
860 			kfree(out);
861 			out = kzalloc(outlen, GFP_KERNEL);
862 			if (!out) {
863 				ret = -ENOMEM;
864 				goto srcu_unlock;
865 			}
866 			cur_outlen = outlen;
867 		}
868 
869 		pklm = (struct mlx5_klm *)MLX5_ADDR_OF(query_mkey_out, out,
870 						       bsf0_klm0_pas_mtt0_1);
871 
872 		ret = mlx5_core_query_mkey(dev->mdev, mmkey, out, outlen);
873 		if (ret)
874 			goto srcu_unlock;
875 
876 		offset = io_virt - MLX5_GET64(query_mkey_out, out,
877 					      memory_key_mkey_entry.start_addr);
878 
879 		for (i = 0; bcnt && i < ndescs; i++, pklm++) {
880 			if (offset >= be32_to_cpu(pklm->bcount)) {
881 				offset -= be32_to_cpu(pklm->bcount);
882 				continue;
883 			}
884 
885 			frame = kzalloc(sizeof(*frame), GFP_KERNEL);
886 			if (!frame) {
887 				ret = -ENOMEM;
888 				goto srcu_unlock;
889 			}
890 
891 			frame->key = be32_to_cpu(pklm->key);
892 			frame->io_virt = be64_to_cpu(pklm->va) + offset;
893 			frame->bcnt = min_t(size_t, bcnt,
894 					    be32_to_cpu(pklm->bcount) - offset);
895 			frame->depth = depth + 1;
896 			frame->next = head;
897 			head = frame;
898 
899 			bcnt -= frame->bcnt;
900 			offset = 0;
901 		}
902 		break;
903 
904 	default:
905 		mlx5_ib_dbg(dev, "wrong mkey type %d\n", mmkey->type);
906 		ret = -EFAULT;
907 		goto srcu_unlock;
908 	}
909 
910 	if (head) {
911 		frame = head;
912 		head = frame->next;
913 
914 		key = frame->key;
915 		io_virt = frame->io_virt;
916 		bcnt = frame->bcnt;
917 		depth = frame->depth;
918 		kfree(frame);
919 
920 		goto next_mr;
921 	}
922 
923 srcu_unlock:
924 	while (head) {
925 		frame = head;
926 		head = frame->next;
927 		kfree(frame);
928 	}
929 	kfree(out);
930 
931 	srcu_read_unlock(&dev->mr_srcu, srcu_key);
932 	*bytes_committed = 0;
933 	return ret ? ret : npages;
934 }
935 
936 /**
937  * Parse a series of data segments for page fault handling.
938  *
939  * @pfault contains page fault information.
940  * @wqe points at the first data segment in the WQE.
941  * @wqe_end points after the end of the WQE.
942  * @bytes_mapped receives the number of bytes that the function was able to
943  *               map. This allows the caller to decide intelligently whether
944  *               enough memory was mapped to resolve the page fault
945  *               successfully (e.g. enough for the next MTU, or the entire
946  *               WQE).
947  * @total_wqe_bytes receives the total data size of this WQE in bytes (minus
948  *                  the committed bytes).
949  *
950  * Returns the number of pages loaded if positive, zero for an empty WQE, or a
951  * negative error code.
952  */
953 static int pagefault_data_segments(struct mlx5_ib_dev *dev,
954 				   struct mlx5_pagefault *pfault,
955 				   void *wqe,
956 				   void *wqe_end, u32 *bytes_mapped,
957 				   u32 *total_wqe_bytes, bool receive_queue)
958 {
959 	int ret = 0, npages = 0;
960 	u64 io_virt;
961 	u32 key;
962 	u32 byte_count;
963 	size_t bcnt;
964 	int inline_segment;
965 
966 	if (bytes_mapped)
967 		*bytes_mapped = 0;
968 	if (total_wqe_bytes)
969 		*total_wqe_bytes = 0;
970 
971 	while (wqe < wqe_end) {
972 		struct mlx5_wqe_data_seg *dseg = wqe;
973 
974 		io_virt = be64_to_cpu(dseg->addr);
975 		key = be32_to_cpu(dseg->lkey);
976 		byte_count = be32_to_cpu(dseg->byte_count);
977 		inline_segment = !!(byte_count &  MLX5_INLINE_SEG);
978 		bcnt	       = byte_count & ~MLX5_INLINE_SEG;
979 
980 		if (inline_segment) {
981 			bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK;
982 			wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt,
983 				     16);
984 		} else {
985 			wqe += sizeof(*dseg);
986 		}
987 
988 		/* receive WQE end of sg list. */
989 		if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY &&
990 		    io_virt == 0)
991 			break;
992 
993 		if (!inline_segment && total_wqe_bytes) {
994 			*total_wqe_bytes += bcnt - min_t(size_t, bcnt,
995 					pfault->bytes_committed);
996 		}
997 
998 		/* A zero length data segment designates a length of 2GB. */
999 		if (bcnt == 0)
1000 			bcnt = 1U << 31;
1001 
1002 		if (inline_segment || bcnt <= pfault->bytes_committed) {
1003 			pfault->bytes_committed -=
1004 				min_t(size_t, bcnt,
1005 				      pfault->bytes_committed);
1006 			continue;
1007 		}
1008 
1009 		ret = pagefault_single_data_segment(dev, NULL, key,
1010 						    io_virt, bcnt,
1011 						    &pfault->bytes_committed,
1012 						    bytes_mapped, 0);
1013 		if (ret < 0)
1014 			break;
1015 		npages += ret;
1016 	}
1017 
1018 	return ret < 0 ? ret : npages;
1019 }
1020 
1021 /*
1022  * Parse initiator WQE. Advances the wqe pointer to point at the
1023  * scatter-gather list, and set wqe_end to the end of the WQE.
1024  */
1025 static int mlx5_ib_mr_initiator_pfault_handler(
1026 	struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault,
1027 	struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length)
1028 {
1029 	struct mlx5_wqe_ctrl_seg *ctrl = *wqe;
1030 	u16 wqe_index = pfault->wqe.wqe_index;
1031 	struct mlx5_base_av *av;
1032 	unsigned ds, opcode;
1033 	u32 qpn = qp->trans_qp.base.mqp.qpn;
1034 
1035 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
1036 	if (ds * MLX5_WQE_DS_UNITS > wqe_length) {
1037 		mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n",
1038 			    ds, wqe_length);
1039 		return -EFAULT;
1040 	}
1041 
1042 	if (ds == 0) {
1043 		mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n",
1044 			    wqe_index, qpn);
1045 		return -EFAULT;
1046 	}
1047 
1048 	*wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS;
1049 	*wqe += sizeof(*ctrl);
1050 
1051 	opcode = be32_to_cpu(ctrl->opmod_idx_opcode) &
1052 		 MLX5_WQE_CTRL_OPCODE_MASK;
1053 
1054 	if (qp->ibqp.qp_type == IB_QPT_XRC_INI)
1055 		*wqe += sizeof(struct mlx5_wqe_xrc_seg);
1056 
1057 	if (qp->ibqp.qp_type == IB_QPT_UD ||
1058 	    qp->qp_sub_type == MLX5_IB_QPT_DCI) {
1059 		av = *wqe;
1060 		if (av->dqp_dct & cpu_to_be32(MLX5_EXTENDED_UD_AV))
1061 			*wqe += sizeof(struct mlx5_av);
1062 		else
1063 			*wqe += sizeof(struct mlx5_base_av);
1064 	}
1065 
1066 	switch (opcode) {
1067 	case MLX5_OPCODE_RDMA_WRITE:
1068 	case MLX5_OPCODE_RDMA_WRITE_IMM:
1069 	case MLX5_OPCODE_RDMA_READ:
1070 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1071 		break;
1072 	case MLX5_OPCODE_ATOMIC_CS:
1073 	case MLX5_OPCODE_ATOMIC_FA:
1074 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1075 		*wqe += sizeof(struct mlx5_wqe_atomic_seg);
1076 		break;
1077 	}
1078 
1079 	return 0;
1080 }
1081 
1082 /*
1083  * Parse responder WQE and set wqe_end to the end of the WQE.
1084  */
1085 static int mlx5_ib_mr_responder_pfault_handler_srq(struct mlx5_ib_dev *dev,
1086 						   struct mlx5_ib_srq *srq,
1087 						   void **wqe, void **wqe_end,
1088 						   int wqe_length)
1089 {
1090 	int wqe_size = 1 << srq->msrq.wqe_shift;
1091 
1092 	if (wqe_size > wqe_length) {
1093 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1094 		return -EFAULT;
1095 	}
1096 
1097 	*wqe_end = *wqe + wqe_size;
1098 	*wqe += sizeof(struct mlx5_wqe_srq_next_seg);
1099 
1100 	return 0;
1101 }
1102 
1103 static int mlx5_ib_mr_responder_pfault_handler_rq(struct mlx5_ib_dev *dev,
1104 						  struct mlx5_ib_qp *qp,
1105 						  void *wqe, void **wqe_end,
1106 						  int wqe_length)
1107 {
1108 	struct mlx5_ib_wq *wq = &qp->rq;
1109 	int wqe_size = 1 << wq->wqe_shift;
1110 
1111 	if (qp->wq_sig) {
1112 		mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n");
1113 		return -EFAULT;
1114 	}
1115 
1116 	if (wqe_size > wqe_length) {
1117 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1118 		return -EFAULT;
1119 	}
1120 
1121 	*wqe_end = wqe + wqe_size;
1122 
1123 	return 0;
1124 }
1125 
1126 static inline struct mlx5_core_rsc_common *odp_get_rsc(struct mlx5_ib_dev *dev,
1127 						       u32 wq_num, int pf_type)
1128 {
1129 	struct mlx5_core_rsc_common *common = NULL;
1130 	struct mlx5_core_srq *srq;
1131 
1132 	switch (pf_type) {
1133 	case MLX5_WQE_PF_TYPE_RMP:
1134 		srq = mlx5_cmd_get_srq(dev, wq_num);
1135 		if (srq)
1136 			common = &srq->common;
1137 		break;
1138 	case MLX5_WQE_PF_TYPE_REQ_SEND_OR_WRITE:
1139 	case MLX5_WQE_PF_TYPE_RESP:
1140 	case MLX5_WQE_PF_TYPE_REQ_READ_OR_ATOMIC:
1141 		common = mlx5_core_res_hold(dev->mdev, wq_num, MLX5_RES_QP);
1142 		break;
1143 	default:
1144 		break;
1145 	}
1146 
1147 	return common;
1148 }
1149 
1150 static inline struct mlx5_ib_qp *res_to_qp(struct mlx5_core_rsc_common *res)
1151 {
1152 	struct mlx5_core_qp *mqp = (struct mlx5_core_qp *)res;
1153 
1154 	return to_mibqp(mqp);
1155 }
1156 
1157 static inline struct mlx5_ib_srq *res_to_srq(struct mlx5_core_rsc_common *res)
1158 {
1159 	struct mlx5_core_srq *msrq =
1160 		container_of(res, struct mlx5_core_srq, common);
1161 
1162 	return to_mibsrq(msrq);
1163 }
1164 
1165 static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
1166 					  struct mlx5_pagefault *pfault)
1167 {
1168 	bool sq = pfault->type & MLX5_PFAULT_REQUESTOR;
1169 	u16 wqe_index = pfault->wqe.wqe_index;
1170 	void *wqe, *wqe_start = NULL, *wqe_end = NULL;
1171 	u32 bytes_mapped, total_wqe_bytes;
1172 	struct mlx5_core_rsc_common *res;
1173 	int resume_with_error = 1;
1174 	struct mlx5_ib_qp *qp;
1175 	size_t bytes_copied;
1176 	int ret = 0;
1177 
1178 	res = odp_get_rsc(dev, pfault->wqe.wq_num, pfault->type);
1179 	if (!res) {
1180 		mlx5_ib_dbg(dev, "wqe page fault for missing resource %d\n", pfault->wqe.wq_num);
1181 		return;
1182 	}
1183 
1184 	if (res->res != MLX5_RES_QP && res->res != MLX5_RES_SRQ &&
1185 	    res->res != MLX5_RES_XSRQ) {
1186 		mlx5_ib_err(dev, "wqe page fault for unsupported type %d\n",
1187 			    pfault->type);
1188 		goto resolve_page_fault;
1189 	}
1190 
1191 	wqe_start = (void *)__get_free_page(GFP_KERNEL);
1192 	if (!wqe_start) {
1193 		mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n");
1194 		goto resolve_page_fault;
1195 	}
1196 
1197 	wqe = wqe_start;
1198 	qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL;
1199 	if (qp && sq) {
1200 		ret = mlx5_ib_read_user_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE,
1201 					       &bytes_copied);
1202 		if (ret)
1203 			goto read_user;
1204 		ret = mlx5_ib_mr_initiator_pfault_handler(
1205 			dev, pfault, qp, &wqe, &wqe_end, bytes_copied);
1206 	} else if (qp && !sq) {
1207 		ret = mlx5_ib_read_user_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE,
1208 					       &bytes_copied);
1209 		if (ret)
1210 			goto read_user;
1211 		ret = mlx5_ib_mr_responder_pfault_handler_rq(
1212 			dev, qp, wqe, &wqe_end, bytes_copied);
1213 	} else if (!qp) {
1214 		struct mlx5_ib_srq *srq = res_to_srq(res);
1215 
1216 		ret = mlx5_ib_read_user_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE,
1217 						&bytes_copied);
1218 		if (ret)
1219 			goto read_user;
1220 		ret = mlx5_ib_mr_responder_pfault_handler_srq(
1221 			dev, srq, &wqe, &wqe_end, bytes_copied);
1222 	}
1223 
1224 	if (ret < 0 || wqe >= wqe_end)
1225 		goto resolve_page_fault;
1226 
1227 	ret = pagefault_data_segments(dev, pfault, wqe, wqe_end, &bytes_mapped,
1228 				      &total_wqe_bytes, !sq);
1229 	if (ret == -EAGAIN)
1230 		goto out;
1231 
1232 	if (ret < 0 || total_wqe_bytes > bytes_mapped)
1233 		goto resolve_page_fault;
1234 
1235 out:
1236 	ret = 0;
1237 	resume_with_error = 0;
1238 
1239 read_user:
1240 	if (ret)
1241 		mlx5_ib_err(
1242 			dev,
1243 			"Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n",
1244 			ret, wqe_index, pfault->token);
1245 
1246 resolve_page_fault:
1247 	mlx5_ib_page_fault_resume(dev, pfault, resume_with_error);
1248 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n",
1249 		    pfault->wqe.wq_num, resume_with_error,
1250 		    pfault->type);
1251 	mlx5_core_res_put(res);
1252 	free_page((unsigned long)wqe_start);
1253 }
1254 
1255 static int pages_in_range(u64 address, u32 length)
1256 {
1257 	return (ALIGN(address + length, PAGE_SIZE) -
1258 		(address & PAGE_MASK)) >> PAGE_SHIFT;
1259 }
1260 
1261 static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev,
1262 					   struct mlx5_pagefault *pfault)
1263 {
1264 	u64 address;
1265 	u32 length;
1266 	u32 prefetch_len = pfault->bytes_committed;
1267 	int prefetch_activated = 0;
1268 	u32 rkey = pfault->rdma.r_key;
1269 	int ret;
1270 
1271 	/* The RDMA responder handler handles the page fault in two parts.
1272 	 * First it brings the necessary pages for the current packet
1273 	 * (and uses the pfault context), and then (after resuming the QP)
1274 	 * prefetches more pages. The second operation cannot use the pfault
1275 	 * context and therefore uses the dummy_pfault context allocated on
1276 	 * the stack */
1277 	pfault->rdma.rdma_va += pfault->bytes_committed;
1278 	pfault->rdma.rdma_op_len -= min(pfault->bytes_committed,
1279 					 pfault->rdma.rdma_op_len);
1280 	pfault->bytes_committed = 0;
1281 
1282 	address = pfault->rdma.rdma_va;
1283 	length  = pfault->rdma.rdma_op_len;
1284 
1285 	/* For some operations, the hardware cannot tell the exact message
1286 	 * length, and in those cases it reports zero. Use prefetch
1287 	 * logic. */
1288 	if (length == 0) {
1289 		prefetch_activated = 1;
1290 		length = pfault->rdma.packet_size;
1291 		prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len);
1292 	}
1293 
1294 	ret = pagefault_single_data_segment(dev, NULL, rkey, address, length,
1295 					    &pfault->bytes_committed, NULL,
1296 					    0);
1297 	if (ret == -EAGAIN) {
1298 		/* We're racing with an invalidation, don't prefetch */
1299 		prefetch_activated = 0;
1300 	} else if (ret < 0 || pages_in_range(address, length) > ret) {
1301 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1302 		if (ret != -ENOENT)
1303 			mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n",
1304 				    ret, pfault->token, pfault->type);
1305 		return;
1306 	}
1307 
1308 	mlx5_ib_page_fault_resume(dev, pfault, 0);
1309 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n",
1310 		    pfault->token, pfault->type,
1311 		    prefetch_activated);
1312 
1313 	/* At this point, there might be a new pagefault already arriving in
1314 	 * the eq, switch to the dummy pagefault for the rest of the
1315 	 * processing. We're still OK with the objects being alive as the
1316 	 * work-queue is being fenced. */
1317 
1318 	if (prefetch_activated) {
1319 		u32 bytes_committed = 0;
1320 
1321 		ret = pagefault_single_data_segment(dev, NULL, rkey, address,
1322 						    prefetch_len,
1323 						    &bytes_committed, NULL,
1324 						    0);
1325 		if (ret < 0 && ret != -EAGAIN) {
1326 			mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n",
1327 				    ret, pfault->token, address, prefetch_len);
1328 		}
1329 	}
1330 }
1331 
1332 static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault)
1333 {
1334 	u8 event_subtype = pfault->event_subtype;
1335 
1336 	switch (event_subtype) {
1337 	case MLX5_PFAULT_SUBTYPE_WQE:
1338 		mlx5_ib_mr_wqe_pfault_handler(dev, pfault);
1339 		break;
1340 	case MLX5_PFAULT_SUBTYPE_RDMA:
1341 		mlx5_ib_mr_rdma_pfault_handler(dev, pfault);
1342 		break;
1343 	default:
1344 		mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n",
1345 			    event_subtype);
1346 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1347 	}
1348 }
1349 
1350 static void mlx5_ib_eqe_pf_action(struct work_struct *work)
1351 {
1352 	struct mlx5_pagefault *pfault = container_of(work,
1353 						     struct mlx5_pagefault,
1354 						     work);
1355 	struct mlx5_ib_pf_eq *eq = pfault->eq;
1356 
1357 	mlx5_ib_pfault(eq->dev, pfault);
1358 	mempool_free(pfault, eq->pool);
1359 }
1360 
1361 static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq)
1362 {
1363 	struct mlx5_eqe_page_fault *pf_eqe;
1364 	struct mlx5_pagefault *pfault;
1365 	struct mlx5_eqe *eqe;
1366 	int cc = 0;
1367 
1368 	while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) {
1369 		pfault = mempool_alloc(eq->pool, GFP_ATOMIC);
1370 		if (!pfault) {
1371 			schedule_work(&eq->work);
1372 			break;
1373 		}
1374 
1375 		pf_eqe = &eqe->data.page_fault;
1376 		pfault->event_subtype = eqe->sub_type;
1377 		pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);
1378 
1379 		mlx5_ib_dbg(eq->dev,
1380 			    "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
1381 			    eqe->sub_type, pfault->bytes_committed);
1382 
1383 		switch (eqe->sub_type) {
1384 		case MLX5_PFAULT_SUBTYPE_RDMA:
1385 			/* RDMA based event */
1386 			pfault->type =
1387 				be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
1388 			pfault->token =
1389 				be32_to_cpu(pf_eqe->rdma.pftype_token) &
1390 				MLX5_24BIT_MASK;
1391 			pfault->rdma.r_key =
1392 				be32_to_cpu(pf_eqe->rdma.r_key);
1393 			pfault->rdma.packet_size =
1394 				be16_to_cpu(pf_eqe->rdma.packet_length);
1395 			pfault->rdma.rdma_op_len =
1396 				be32_to_cpu(pf_eqe->rdma.rdma_op_len);
1397 			pfault->rdma.rdma_va =
1398 				be64_to_cpu(pf_eqe->rdma.rdma_va);
1399 			mlx5_ib_dbg(eq->dev,
1400 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
1401 				    pfault->type, pfault->token,
1402 				    pfault->rdma.r_key);
1403 			mlx5_ib_dbg(eq->dev,
1404 				    "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
1405 				    pfault->rdma.rdma_op_len,
1406 				    pfault->rdma.rdma_va);
1407 			break;
1408 
1409 		case MLX5_PFAULT_SUBTYPE_WQE:
1410 			/* WQE based event */
1411 			pfault->type =
1412 				(be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7;
1413 			pfault->token =
1414 				be32_to_cpu(pf_eqe->wqe.token);
1415 			pfault->wqe.wq_num =
1416 				be32_to_cpu(pf_eqe->wqe.pftype_wq) &
1417 				MLX5_24BIT_MASK;
1418 			pfault->wqe.wqe_index =
1419 				be16_to_cpu(pf_eqe->wqe.wqe_index);
1420 			pfault->wqe.packet_size =
1421 				be16_to_cpu(pf_eqe->wqe.packet_length);
1422 			mlx5_ib_dbg(eq->dev,
1423 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
1424 				    pfault->type, pfault->token,
1425 				    pfault->wqe.wq_num,
1426 				    pfault->wqe.wqe_index);
1427 			break;
1428 
1429 		default:
1430 			mlx5_ib_warn(eq->dev,
1431 				     "Unsupported page fault event sub-type: 0x%02hhx\n",
1432 				     eqe->sub_type);
1433 			/* Unsupported page faults should still be
1434 			 * resolved by the page fault handler
1435 			 */
1436 		}
1437 
1438 		pfault->eq = eq;
1439 		INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action);
1440 		queue_work(eq->wq, &pfault->work);
1441 
1442 		cc = mlx5_eq_update_cc(eq->core, ++cc);
1443 	}
1444 
1445 	mlx5_eq_update_ci(eq->core, cc, 1);
1446 }
1447 
1448 static int mlx5_ib_eq_pf_int(struct notifier_block *nb, unsigned long type,
1449 			     void *data)
1450 {
1451 	struct mlx5_ib_pf_eq *eq =
1452 		container_of(nb, struct mlx5_ib_pf_eq, irq_nb);
1453 	unsigned long flags;
1454 
1455 	if (spin_trylock_irqsave(&eq->lock, flags)) {
1456 		mlx5_ib_eq_pf_process(eq);
1457 		spin_unlock_irqrestore(&eq->lock, flags);
1458 	} else {
1459 		schedule_work(&eq->work);
1460 	}
1461 
1462 	return IRQ_HANDLED;
1463 }
1464 
1465 /* mempool_refill() was proposed but unfortunately wasn't accepted
1466  * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
1467  * Cheap workaround.
1468  */
1469 static void mempool_refill(mempool_t *pool)
1470 {
1471 	while (pool->curr_nr < pool->min_nr)
1472 		mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
1473 }
1474 
1475 static void mlx5_ib_eq_pf_action(struct work_struct *work)
1476 {
1477 	struct mlx5_ib_pf_eq *eq =
1478 		container_of(work, struct mlx5_ib_pf_eq, work);
1479 
1480 	mempool_refill(eq->pool);
1481 
1482 	spin_lock_irq(&eq->lock);
1483 	mlx5_ib_eq_pf_process(eq);
1484 	spin_unlock_irq(&eq->lock);
1485 }
1486 
1487 enum {
1488 	MLX5_IB_NUM_PF_EQE	= 0x1000,
1489 	MLX5_IB_NUM_PF_DRAIN	= 64,
1490 };
1491 
1492 static int
1493 mlx5_ib_create_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1494 {
1495 	struct mlx5_eq_param param = {};
1496 	int err;
1497 
1498 	INIT_WORK(&eq->work, mlx5_ib_eq_pf_action);
1499 	spin_lock_init(&eq->lock);
1500 	eq->dev = dev;
1501 
1502 	eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN,
1503 					       sizeof(struct mlx5_pagefault));
1504 	if (!eq->pool)
1505 		return -ENOMEM;
1506 
1507 	eq->wq = alloc_workqueue("mlx5_ib_page_fault",
1508 				 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM,
1509 				 MLX5_NUM_CMD_EQE);
1510 	if (!eq->wq) {
1511 		err = -ENOMEM;
1512 		goto err_mempool;
1513 	}
1514 
1515 	eq->irq_nb.notifier_call = mlx5_ib_eq_pf_int;
1516 	param = (struct mlx5_eq_param) {
1517 		.irq_index = 0,
1518 		.nent = MLX5_IB_NUM_PF_EQE,
1519 	};
1520 	param.mask[0] = 1ull << MLX5_EVENT_TYPE_PAGE_FAULT;
1521 	eq->core = mlx5_eq_create_generic(dev->mdev, &param);
1522 	if (IS_ERR(eq->core)) {
1523 		err = PTR_ERR(eq->core);
1524 		goto err_wq;
1525 	}
1526 	err = mlx5_eq_enable(dev->mdev, eq->core, &eq->irq_nb);
1527 	if (err) {
1528 		mlx5_ib_err(dev, "failed to enable odp EQ %d\n", err);
1529 		goto err_eq;
1530 	}
1531 
1532 	return 0;
1533 err_eq:
1534 	mlx5_eq_destroy_generic(dev->mdev, eq->core);
1535 err_wq:
1536 	destroy_workqueue(eq->wq);
1537 err_mempool:
1538 	mempool_destroy(eq->pool);
1539 	return err;
1540 }
1541 
1542 static int
1543 mlx5_ib_destroy_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1544 {
1545 	int err;
1546 
1547 	mlx5_eq_disable(dev->mdev, eq->core, &eq->irq_nb);
1548 	err = mlx5_eq_destroy_generic(dev->mdev, eq->core);
1549 	cancel_work_sync(&eq->work);
1550 	destroy_workqueue(eq->wq);
1551 	mempool_destroy(eq->pool);
1552 
1553 	return err;
1554 }
1555 
1556 void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
1557 {
1558 	if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1559 		return;
1560 
1561 	switch (ent->order - 2) {
1562 	case MLX5_IMR_MTT_CACHE_ENTRY:
1563 		ent->page = PAGE_SHIFT;
1564 		ent->xlt = MLX5_IMR_MTT_ENTRIES *
1565 			   sizeof(struct mlx5_mtt) /
1566 			   MLX5_IB_UMR_OCTOWORD;
1567 		ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
1568 		ent->limit = 0;
1569 		break;
1570 
1571 	case MLX5_IMR_KSM_CACHE_ENTRY:
1572 		ent->page = MLX5_KSM_PAGE_SHIFT;
1573 		ent->xlt = mlx5_imr_ksm_entries *
1574 			   sizeof(struct mlx5_klm) /
1575 			   MLX5_IB_UMR_OCTOWORD;
1576 		ent->access_mode = MLX5_MKC_ACCESS_MODE_KSM;
1577 		ent->limit = 0;
1578 		break;
1579 	}
1580 }
1581 
1582 static const struct ib_device_ops mlx5_ib_dev_odp_ops = {
1583 	.advise_mr = mlx5_ib_advise_mr,
1584 	.invalidate_range = mlx5_ib_invalidate_range,
1585 };
1586 
1587 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
1588 {
1589 	int ret = 0;
1590 
1591 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1592 		return ret;
1593 
1594 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_odp_ops);
1595 
1596 	if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) {
1597 		ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey);
1598 		if (ret) {
1599 			mlx5_ib_err(dev, "Error getting null_mkey %d\n", ret);
1600 			return ret;
1601 		}
1602 	}
1603 
1604 	ret = mlx5_ib_create_pf_eq(dev, &dev->odp_pf_eq);
1605 
1606 	return ret;
1607 }
1608 
1609 void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev)
1610 {
1611 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1612 		return;
1613 
1614 	mlx5_ib_destroy_pf_eq(dev, &dev->odp_pf_eq);
1615 }
1616 
1617 int mlx5_ib_odp_init(void)
1618 {
1619 	mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) -
1620 				       MLX5_IMR_MTT_BITS);
1621 
1622 	return 0;
1623 }
1624 
1625 struct prefetch_mr_work {
1626 	struct work_struct work;
1627 	struct ib_pd *pd;
1628 	u32 pf_flags;
1629 	u32 num_sge;
1630 	struct ib_sge sg_list[0];
1631 };
1632 
1633 static void num_pending_prefetch_dec(struct mlx5_ib_dev *dev,
1634 				     struct ib_sge *sg_list, u32 num_sge,
1635 				     u32 from)
1636 {
1637 	u32 i;
1638 	int srcu_key;
1639 
1640 	srcu_key = srcu_read_lock(&dev->mr_srcu);
1641 
1642 	for (i = from; i < num_sge; ++i) {
1643 		struct mlx5_core_mkey *mmkey;
1644 		struct mlx5_ib_mr *mr;
1645 
1646 		mmkey = xa_load(&dev->mdev->priv.mkey_table,
1647 				mlx5_base_mkey(sg_list[i].lkey));
1648 		mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
1649 		atomic_dec(&mr->num_pending_prefetch);
1650 	}
1651 
1652 	srcu_read_unlock(&dev->mr_srcu, srcu_key);
1653 }
1654 
1655 static bool num_pending_prefetch_inc(struct ib_pd *pd,
1656 				     struct ib_sge *sg_list, u32 num_sge)
1657 {
1658 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1659 	bool ret = true;
1660 	u32 i;
1661 
1662 	for (i = 0; i < num_sge; ++i) {
1663 		struct mlx5_core_mkey *mmkey;
1664 		struct mlx5_ib_mr *mr;
1665 
1666 		mmkey = xa_load(&dev->mdev->priv.mkey_table,
1667 				mlx5_base_mkey(sg_list[i].lkey));
1668 		if (!mmkey || mmkey->key != sg_list[i].lkey) {
1669 			ret = false;
1670 			break;
1671 		}
1672 
1673 		if (mmkey->type != MLX5_MKEY_MR) {
1674 			ret = false;
1675 			break;
1676 		}
1677 
1678 		mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
1679 
1680 		if (!smp_load_acquire(&mr->live)) {
1681 			ret = false;
1682 			break;
1683 		}
1684 
1685 		if (mr->ibmr.pd != pd) {
1686 			ret = false;
1687 			break;
1688 		}
1689 
1690 		atomic_inc(&mr->num_pending_prefetch);
1691 	}
1692 
1693 	if (!ret)
1694 		num_pending_prefetch_dec(dev, sg_list, i, 0);
1695 
1696 	return ret;
1697 }
1698 
1699 static int mlx5_ib_prefetch_sg_list(struct ib_pd *pd, u32 pf_flags,
1700 				    struct ib_sge *sg_list, u32 num_sge)
1701 {
1702 	u32 i;
1703 	int ret = 0;
1704 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1705 
1706 	for (i = 0; i < num_sge; ++i) {
1707 		struct ib_sge *sg = &sg_list[i];
1708 		int bytes_committed = 0;
1709 
1710 		ret = pagefault_single_data_segment(dev, pd, sg->lkey, sg->addr,
1711 						    sg->length,
1712 						    &bytes_committed, NULL,
1713 						    pf_flags);
1714 		if (ret < 0)
1715 			break;
1716 	}
1717 
1718 	return ret < 0 ? ret : 0;
1719 }
1720 
1721 static void mlx5_ib_prefetch_mr_work(struct work_struct *work)
1722 {
1723 	struct prefetch_mr_work *w =
1724 		container_of(work, struct prefetch_mr_work, work);
1725 
1726 	if (ib_device_try_get(w->pd->device)) {
1727 		mlx5_ib_prefetch_sg_list(w->pd, w->pf_flags, w->sg_list,
1728 					 w->num_sge);
1729 		ib_device_put(w->pd->device);
1730 	}
1731 
1732 	num_pending_prefetch_dec(to_mdev(w->pd->device), w->sg_list,
1733 				 w->num_sge, 0);
1734 	kvfree(w);
1735 }
1736 
1737 int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
1738 			       enum ib_uverbs_advise_mr_advice advice,
1739 			       u32 flags, struct ib_sge *sg_list, u32 num_sge)
1740 {
1741 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1742 	u32 pf_flags = MLX5_PF_FLAGS_PREFETCH;
1743 	struct prefetch_mr_work *work;
1744 	bool valid_req;
1745 	int srcu_key;
1746 
1747 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH)
1748 		pf_flags |= MLX5_PF_FLAGS_DOWNGRADE;
1749 
1750 	if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH)
1751 		return mlx5_ib_prefetch_sg_list(pd, pf_flags, sg_list,
1752 						num_sge);
1753 
1754 	work = kvzalloc(struct_size(work, sg_list, num_sge), GFP_KERNEL);
1755 	if (!work)
1756 		return -ENOMEM;
1757 
1758 	memcpy(work->sg_list, sg_list, num_sge * sizeof(struct ib_sge));
1759 
1760 	/* It is guaranteed that the pd when work is executed is the pd when
1761 	 * work was queued since pd can't be destroyed while it holds MRs and
1762 	 * destroying a MR leads to flushing the workquque
1763 	 */
1764 	work->pd = pd;
1765 	work->pf_flags = pf_flags;
1766 	work->num_sge = num_sge;
1767 
1768 	INIT_WORK(&work->work, mlx5_ib_prefetch_mr_work);
1769 
1770 	srcu_key = srcu_read_lock(&dev->mr_srcu);
1771 
1772 	valid_req = num_pending_prefetch_inc(pd, sg_list, num_sge);
1773 	if (valid_req)
1774 		queue_work(system_unbound_wq, &work->work);
1775 	else
1776 		kvfree(work);
1777 
1778 	srcu_read_unlock(&dev->mr_srcu, srcu_key);
1779 
1780 	return valid_req ? 0 : -EINVAL;
1781 }
1782