xref: /openbmc/linux/drivers/infiniband/hw/mlx5/odp.c (revision 48cc39c3)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <rdma/ib_umem.h>
34 #include <rdma/ib_umem_odp.h>
35 #include <linux/kernel.h>
36 #include <linux/dma-buf.h>
37 #include <linux/dma-resv.h>
38 
39 #include "mlx5_ib.h"
40 #include "cmd.h"
41 #include "qp.h"
42 
43 #include <linux/mlx5/eq.h>
44 
45 /* Contains the details of a pagefault. */
46 struct mlx5_pagefault {
47 	u32			bytes_committed;
48 	u32			token;
49 	u8			event_subtype;
50 	u8			type;
51 	union {
52 		/* Initiator or send message responder pagefault details. */
53 		struct {
54 			/* Received packet size, only valid for responders. */
55 			u32	packet_size;
56 			/*
57 			 * Number of resource holding WQE, depends on type.
58 			 */
59 			u32	wq_num;
60 			/*
61 			 * WQE index. Refers to either the send queue or
62 			 * receive queue, according to event_subtype.
63 			 */
64 			u16	wqe_index;
65 		} wqe;
66 		/* RDMA responder pagefault details */
67 		struct {
68 			u32	r_key;
69 			/*
70 			 * Received packet size, minimal size page fault
71 			 * resolution required for forward progress.
72 			 */
73 			u32	packet_size;
74 			u32	rdma_op_len;
75 			u64	rdma_va;
76 		} rdma;
77 	};
78 
79 	struct mlx5_ib_pf_eq	*eq;
80 	struct work_struct	work;
81 };
82 
83 #define MAX_PREFETCH_LEN (4*1024*1024U)
84 
85 /* Timeout in ms to wait for an active mmu notifier to complete when handling
86  * a pagefault. */
87 #define MMU_NOTIFIER_TIMEOUT 1000
88 
89 #define MLX5_IMR_MTT_BITS (30 - PAGE_SHIFT)
90 #define MLX5_IMR_MTT_SHIFT (MLX5_IMR_MTT_BITS + PAGE_SHIFT)
91 #define MLX5_IMR_MTT_ENTRIES BIT_ULL(MLX5_IMR_MTT_BITS)
92 #define MLX5_IMR_MTT_SIZE BIT_ULL(MLX5_IMR_MTT_SHIFT)
93 #define MLX5_IMR_MTT_MASK (~(MLX5_IMR_MTT_SIZE - 1))
94 
95 #define MLX5_KSM_PAGE_SHIFT MLX5_IMR_MTT_SHIFT
96 
97 static u64 mlx5_imr_ksm_entries;
98 
99 static void populate_klm(struct mlx5_klm *pklm, size_t idx, size_t nentries,
100 			struct mlx5_ib_mr *imr, int flags)
101 {
102 	struct mlx5_klm *end = pklm + nentries;
103 
104 	if (flags & MLX5_IB_UPD_XLT_ZAP) {
105 		for (; pklm != end; pklm++, idx++) {
106 			pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
107 			pklm->key = cpu_to_be32(mr_to_mdev(imr)->null_mkey);
108 			pklm->va = 0;
109 		}
110 		return;
111 	}
112 
113 	/*
114 	 * The locking here is pretty subtle. Ideally the implicit_children
115 	 * xarray would be protected by the umem_mutex, however that is not
116 	 * possible. Instead this uses a weaker update-then-lock pattern:
117 	 *
118 	 *    xa_store()
119 	 *    mutex_lock(umem_mutex)
120 	 *     mlx5_ib_update_xlt()
121 	 *    mutex_unlock(umem_mutex)
122 	 *    destroy lkey
123 	 *
124 	 * ie any change the xarray must be followed by the locked update_xlt
125 	 * before destroying.
126 	 *
127 	 * The umem_mutex provides the acquire/release semantic needed to make
128 	 * the xa_store() visible to a racing thread.
129 	 */
130 	lockdep_assert_held(&to_ib_umem_odp(imr->umem)->umem_mutex);
131 
132 	for (; pklm != end; pklm++, idx++) {
133 		struct mlx5_ib_mr *mtt = xa_load(&imr->implicit_children, idx);
134 
135 		pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
136 		if (mtt) {
137 			pklm->key = cpu_to_be32(mtt->ibmr.lkey);
138 			pklm->va = cpu_to_be64(idx * MLX5_IMR_MTT_SIZE);
139 		} else {
140 			pklm->key = cpu_to_be32(mr_to_mdev(imr)->null_mkey);
141 			pklm->va = 0;
142 		}
143 	}
144 }
145 
146 static u64 umem_dma_to_mtt(dma_addr_t umem_dma)
147 {
148 	u64 mtt_entry = umem_dma & ODP_DMA_ADDR_MASK;
149 
150 	if (umem_dma & ODP_READ_ALLOWED_BIT)
151 		mtt_entry |= MLX5_IB_MTT_READ;
152 	if (umem_dma & ODP_WRITE_ALLOWED_BIT)
153 		mtt_entry |= MLX5_IB_MTT_WRITE;
154 
155 	return mtt_entry;
156 }
157 
158 static void populate_mtt(__be64 *pas, size_t idx, size_t nentries,
159 			 struct mlx5_ib_mr *mr, int flags)
160 {
161 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
162 	dma_addr_t pa;
163 	size_t i;
164 
165 	if (flags & MLX5_IB_UPD_XLT_ZAP)
166 		return;
167 
168 	for (i = 0; i < nentries; i++) {
169 		pa = odp->dma_list[idx + i];
170 		pas[i] = cpu_to_be64(umem_dma_to_mtt(pa));
171 	}
172 }
173 
174 void mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
175 			   struct mlx5_ib_mr *mr, int flags)
176 {
177 	if (flags & MLX5_IB_UPD_XLT_INDIRECT) {
178 		populate_klm(xlt, idx, nentries, mr, flags);
179 	} else {
180 		populate_mtt(xlt, idx, nentries, mr, flags);
181 	}
182 }
183 
184 /*
185  * This must be called after the mr has been removed from implicit_children.
186  * NOTE: The MR does not necessarily have to be
187  * empty here, parallel page faults could have raced with the free process and
188  * added pages to it.
189  */
190 static void free_implicit_child_mr_work(struct work_struct *work)
191 {
192 	struct mlx5_ib_mr *mr =
193 		container_of(work, struct mlx5_ib_mr, odp_destroy.work);
194 	struct mlx5_ib_mr *imr = mr->parent;
195 	struct ib_umem_odp *odp_imr = to_ib_umem_odp(imr->umem);
196 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
197 
198 	mlx5r_deref_wait_odp_mkey(&mr->mmkey);
199 
200 	mutex_lock(&odp_imr->umem_mutex);
201 	mlx5_ib_update_xlt(mr->parent, ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT,
202 			   1, 0,
203 			   MLX5_IB_UPD_XLT_INDIRECT | MLX5_IB_UPD_XLT_ATOMIC);
204 	mutex_unlock(&odp_imr->umem_mutex);
205 	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
206 
207 	mlx5r_deref_odp_mkey(&imr->mmkey);
208 }
209 
210 static void destroy_unused_implicit_child_mr(struct mlx5_ib_mr *mr)
211 {
212 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
213 	unsigned long idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT;
214 	struct mlx5_ib_mr *imr = mr->parent;
215 
216 	if (!refcount_inc_not_zero(&imr->mmkey.usecount))
217 		return;
218 
219 	xa_erase(&imr->implicit_children, idx);
220 
221 	/* Freeing a MR is a sleeping operation, so bounce to a work queue */
222 	INIT_WORK(&mr->odp_destroy.work, free_implicit_child_mr_work);
223 	queue_work(system_unbound_wq, &mr->odp_destroy.work);
224 }
225 
226 static bool mlx5_ib_invalidate_range(struct mmu_interval_notifier *mni,
227 				     const struct mmu_notifier_range *range,
228 				     unsigned long cur_seq)
229 {
230 	struct ib_umem_odp *umem_odp =
231 		container_of(mni, struct ib_umem_odp, notifier);
232 	struct mlx5_ib_mr *mr;
233 	const u64 umr_block_mask = (MLX5_UMR_MTT_ALIGNMENT /
234 				    sizeof(struct mlx5_mtt)) - 1;
235 	u64 idx = 0, blk_start_idx = 0;
236 	u64 invalidations = 0;
237 	unsigned long start;
238 	unsigned long end;
239 	int in_block = 0;
240 	u64 addr;
241 
242 	if (!mmu_notifier_range_blockable(range))
243 		return false;
244 
245 	mutex_lock(&umem_odp->umem_mutex);
246 	mmu_interval_set_seq(mni, cur_seq);
247 	/*
248 	 * If npages is zero then umem_odp->private may not be setup yet. This
249 	 * does not complete until after the first page is mapped for DMA.
250 	 */
251 	if (!umem_odp->npages)
252 		goto out;
253 	mr = umem_odp->private;
254 
255 	start = max_t(u64, ib_umem_start(umem_odp), range->start);
256 	end = min_t(u64, ib_umem_end(umem_odp), range->end);
257 
258 	/*
259 	 * Iteration one - zap the HW's MTTs. The notifiers_count ensures that
260 	 * while we are doing the invalidation, no page fault will attempt to
261 	 * overwrite the same MTTs.  Concurent invalidations might race us,
262 	 * but they will write 0s as well, so no difference in the end result.
263 	 */
264 	for (addr = start; addr < end; addr += BIT(umem_odp->page_shift)) {
265 		idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
266 		/*
267 		 * Strive to write the MTTs in chunks, but avoid overwriting
268 		 * non-existing MTTs. The huristic here can be improved to
269 		 * estimate the cost of another UMR vs. the cost of bigger
270 		 * UMR.
271 		 */
272 		if (umem_odp->dma_list[idx] &
273 		    (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) {
274 			if (!in_block) {
275 				blk_start_idx = idx;
276 				in_block = 1;
277 			}
278 
279 			/* Count page invalidations */
280 			invalidations += idx - blk_start_idx + 1;
281 		} else {
282 			u64 umr_offset = idx & umr_block_mask;
283 
284 			if (in_block && umr_offset == 0) {
285 				mlx5_ib_update_xlt(mr, blk_start_idx,
286 						   idx - blk_start_idx, 0,
287 						   MLX5_IB_UPD_XLT_ZAP |
288 						   MLX5_IB_UPD_XLT_ATOMIC);
289 				in_block = 0;
290 			}
291 		}
292 	}
293 	if (in_block)
294 		mlx5_ib_update_xlt(mr, blk_start_idx,
295 				   idx - blk_start_idx + 1, 0,
296 				   MLX5_IB_UPD_XLT_ZAP |
297 				   MLX5_IB_UPD_XLT_ATOMIC);
298 
299 	mlx5_update_odp_stats(mr, invalidations, invalidations);
300 
301 	/*
302 	 * We are now sure that the device will not access the
303 	 * memory. We can safely unmap it, and mark it as dirty if
304 	 * needed.
305 	 */
306 
307 	ib_umem_odp_unmap_dma_pages(umem_odp, start, end);
308 
309 	if (unlikely(!umem_odp->npages && mr->parent))
310 		destroy_unused_implicit_child_mr(mr);
311 out:
312 	mutex_unlock(&umem_odp->umem_mutex);
313 	return true;
314 }
315 
316 const struct mmu_interval_notifier_ops mlx5_mn_ops = {
317 	.invalidate = mlx5_ib_invalidate_range,
318 };
319 
320 static void internal_fill_odp_caps(struct mlx5_ib_dev *dev)
321 {
322 	struct ib_odp_caps *caps = &dev->odp_caps;
323 
324 	memset(caps, 0, sizeof(*caps));
325 
326 	if (!MLX5_CAP_GEN(dev->mdev, pg) ||
327 	    !mlx5_ib_can_load_pas_with_umr(dev, 0))
328 		return;
329 
330 	caps->general_caps = IB_ODP_SUPPORT;
331 
332 	if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
333 		dev->odp_max_size = U64_MAX;
334 	else
335 		dev->odp_max_size = BIT_ULL(MLX5_MAX_UMR_SHIFT + PAGE_SHIFT);
336 
337 	if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.send))
338 		caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND;
339 
340 	if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.srq_receive))
341 		caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
342 
343 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.send))
344 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND;
345 
346 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.receive))
347 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV;
348 
349 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.write))
350 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE;
351 
352 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.read))
353 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ;
354 
355 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.atomic))
356 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
357 
358 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.srq_receive))
359 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
360 
361 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.send))
362 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SEND;
363 
364 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.receive))
365 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_RECV;
366 
367 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.write))
368 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_WRITE;
369 
370 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.read))
371 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_READ;
372 
373 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.atomic))
374 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
375 
376 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.srq_receive))
377 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
378 
379 	if (MLX5_CAP_GEN(dev->mdev, fixed_buffer_size) &&
380 	    MLX5_CAP_GEN(dev->mdev, null_mkey) &&
381 	    MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset) &&
382 	    !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled))
383 		caps->general_caps |= IB_ODP_SUPPORT_IMPLICIT;
384 }
385 
386 static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev,
387 				      struct mlx5_pagefault *pfault,
388 				      int error)
389 {
390 	int wq_num = pfault->event_subtype == MLX5_PFAULT_SUBTYPE_WQE ?
391 		     pfault->wqe.wq_num : pfault->token;
392 	u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)] = {};
393 	int err;
394 
395 	MLX5_SET(page_fault_resume_in, in, opcode, MLX5_CMD_OP_PAGE_FAULT_RESUME);
396 	MLX5_SET(page_fault_resume_in, in, page_fault_type, pfault->type);
397 	MLX5_SET(page_fault_resume_in, in, token, pfault->token);
398 	MLX5_SET(page_fault_resume_in, in, wq_number, wq_num);
399 	MLX5_SET(page_fault_resume_in, in, error, !!error);
400 
401 	err = mlx5_cmd_exec_in(dev->mdev, page_fault_resume, in);
402 	if (err)
403 		mlx5_ib_err(dev, "Failed to resolve the page fault on WQ 0x%x err %d\n",
404 			    wq_num, err);
405 }
406 
407 static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,
408 						unsigned long idx)
409 {
410 	struct ib_umem_odp *odp;
411 	struct mlx5_ib_mr *mr;
412 	struct mlx5_ib_mr *ret;
413 	int err;
414 
415 	odp = ib_umem_odp_alloc_child(to_ib_umem_odp(imr->umem),
416 				      idx * MLX5_IMR_MTT_SIZE,
417 				      MLX5_IMR_MTT_SIZE, &mlx5_mn_ops);
418 	if (IS_ERR(odp))
419 		return ERR_CAST(odp);
420 
421 	mr = mlx5_mr_cache_alloc(
422 		mr_to_mdev(imr), MLX5_IMR_MTT_CACHE_ENTRY, imr->access_flags);
423 	if (IS_ERR(mr)) {
424 		ib_umem_odp_release(odp);
425 		return mr;
426 	}
427 
428 	mr->ibmr.pd = imr->ibmr.pd;
429 	mr->ibmr.device = &mr_to_mdev(imr)->ib_dev;
430 	mr->umem = &odp->umem;
431 	mr->ibmr.lkey = mr->mmkey.key;
432 	mr->ibmr.rkey = mr->mmkey.key;
433 	mr->mmkey.iova = idx * MLX5_IMR_MTT_SIZE;
434 	mr->parent = imr;
435 	odp->private = mr;
436 
437 	/*
438 	 * First refcount is owned by the xarray and second refconut
439 	 * is returned to the caller.
440 	 */
441 	refcount_set(&mr->mmkey.usecount, 2);
442 
443 	err = mlx5_ib_update_xlt(mr, 0,
444 				 MLX5_IMR_MTT_ENTRIES,
445 				 PAGE_SHIFT,
446 				 MLX5_IB_UPD_XLT_ZAP |
447 				 MLX5_IB_UPD_XLT_ENABLE);
448 	if (err) {
449 		ret = ERR_PTR(err);
450 		goto out_mr;
451 	}
452 
453 	xa_lock(&imr->implicit_children);
454 	ret = __xa_cmpxchg(&imr->implicit_children, idx, NULL, mr,
455 			   GFP_KERNEL);
456 	if (unlikely(ret)) {
457 		if (xa_is_err(ret)) {
458 			ret = ERR_PTR(xa_err(ret));
459 			goto out_lock;
460 		}
461 		/*
462 		 * Another thread beat us to creating the child mr, use
463 		 * theirs.
464 		 */
465 		refcount_inc(&ret->mmkey.usecount);
466 		goto out_lock;
467 	}
468 	xa_unlock(&imr->implicit_children);
469 
470 	mlx5_ib_dbg(mr_to_mdev(imr), "key %x mr %p\n", mr->mmkey.key, mr);
471 	return mr;
472 
473 out_lock:
474 	xa_unlock(&imr->implicit_children);
475 out_mr:
476 	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
477 	return ret;
478 }
479 
480 struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
481 					     int access_flags)
482 {
483 	struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device);
484 	struct ib_umem_odp *umem_odp;
485 	struct mlx5_ib_mr *imr;
486 	int err;
487 
488 	if (!mlx5_ib_can_load_pas_with_umr(dev,
489 					   MLX5_IMR_MTT_ENTRIES * PAGE_SIZE))
490 		return ERR_PTR(-EOPNOTSUPP);
491 
492 	umem_odp = ib_umem_odp_alloc_implicit(&dev->ib_dev, access_flags);
493 	if (IS_ERR(umem_odp))
494 		return ERR_CAST(umem_odp);
495 
496 	imr = mlx5_mr_cache_alloc(dev, MLX5_IMR_KSM_CACHE_ENTRY, access_flags);
497 	if (IS_ERR(imr)) {
498 		ib_umem_odp_release(umem_odp);
499 		return imr;
500 	}
501 
502 	imr->ibmr.pd = &pd->ibpd;
503 	imr->mmkey.iova = 0;
504 	imr->umem = &umem_odp->umem;
505 	imr->ibmr.lkey = imr->mmkey.key;
506 	imr->ibmr.rkey = imr->mmkey.key;
507 	imr->ibmr.device = &dev->ib_dev;
508 	imr->umem = &umem_odp->umem;
509 	imr->is_odp_implicit = true;
510 	xa_init(&imr->implicit_children);
511 
512 	err = mlx5_ib_update_xlt(imr, 0,
513 				 mlx5_imr_ksm_entries,
514 				 MLX5_KSM_PAGE_SHIFT,
515 				 MLX5_IB_UPD_XLT_INDIRECT |
516 				 MLX5_IB_UPD_XLT_ZAP |
517 				 MLX5_IB_UPD_XLT_ENABLE);
518 	if (err)
519 		goto out_mr;
520 
521 	err = mlx5r_store_odp_mkey(dev, &imr->mmkey);
522 	if (err)
523 		goto out_mr;
524 
525 	mlx5_ib_dbg(dev, "key %x mr %p\n", imr->mmkey.key, imr);
526 	return imr;
527 out_mr:
528 	mlx5_ib_err(dev, "Failed to register MKEY %d\n", err);
529 	mlx5_ib_dereg_mr(&imr->ibmr, NULL);
530 	return ERR_PTR(err);
531 }
532 
533 void mlx5_ib_free_odp_mr(struct mlx5_ib_mr *mr)
534 {
535 	struct mlx5_ib_mr *mtt;
536 	unsigned long idx;
537 
538 	/*
539 	 * If this is an implicit MR it is already invalidated so we can just
540 	 * delete the children mkeys.
541 	 */
542 	xa_for_each(&mr->implicit_children, idx, mtt) {
543 		xa_erase(&mr->implicit_children, idx);
544 		mlx5_ib_dereg_mr(&mtt->ibmr, NULL);
545 	}
546 }
547 
548 #define MLX5_PF_FLAGS_DOWNGRADE BIT(1)
549 #define MLX5_PF_FLAGS_SNAPSHOT BIT(2)
550 #define MLX5_PF_FLAGS_ENABLE BIT(3)
551 static int pagefault_real_mr(struct mlx5_ib_mr *mr, struct ib_umem_odp *odp,
552 			     u64 user_va, size_t bcnt, u32 *bytes_mapped,
553 			     u32 flags)
554 {
555 	int page_shift, ret, np;
556 	bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE;
557 	u64 access_mask;
558 	u64 start_idx;
559 	bool fault = !(flags & MLX5_PF_FLAGS_SNAPSHOT);
560 	u32 xlt_flags = MLX5_IB_UPD_XLT_ATOMIC;
561 
562 	if (flags & MLX5_PF_FLAGS_ENABLE)
563 		xlt_flags |= MLX5_IB_UPD_XLT_ENABLE;
564 
565 	page_shift = odp->page_shift;
566 	start_idx = (user_va - ib_umem_start(odp)) >> page_shift;
567 	access_mask = ODP_READ_ALLOWED_BIT;
568 
569 	if (odp->umem.writable && !downgrade)
570 		access_mask |= ODP_WRITE_ALLOWED_BIT;
571 
572 	np = ib_umem_odp_map_dma_and_lock(odp, user_va, bcnt, access_mask, fault);
573 	if (np < 0)
574 		return np;
575 
576 	/*
577 	 * No need to check whether the MTTs really belong to this MR, since
578 	 * ib_umem_odp_map_dma_and_lock already checks this.
579 	 */
580 	ret = mlx5_ib_update_xlt(mr, start_idx, np, page_shift, xlt_flags);
581 	mutex_unlock(&odp->umem_mutex);
582 
583 	if (ret < 0) {
584 		if (ret != -EAGAIN)
585 			mlx5_ib_err(mr_to_mdev(mr),
586 				    "Failed to update mkey page tables\n");
587 		goto out;
588 	}
589 
590 	if (bytes_mapped) {
591 		u32 new_mappings = (np << page_shift) -
592 			(user_va - round_down(user_va, 1 << page_shift));
593 
594 		*bytes_mapped += min_t(u32, new_mappings, bcnt);
595 	}
596 
597 	return np << (page_shift - PAGE_SHIFT);
598 
599 out:
600 	return ret;
601 }
602 
603 static int pagefault_implicit_mr(struct mlx5_ib_mr *imr,
604 				 struct ib_umem_odp *odp_imr, u64 user_va,
605 				 size_t bcnt, u32 *bytes_mapped, u32 flags)
606 {
607 	unsigned long end_idx = (user_va + bcnt - 1) >> MLX5_IMR_MTT_SHIFT;
608 	unsigned long upd_start_idx = end_idx + 1;
609 	unsigned long upd_len = 0;
610 	unsigned long npages = 0;
611 	int err;
612 	int ret;
613 
614 	if (unlikely(user_va >= mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE ||
615 		     mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE - user_va < bcnt))
616 		return -EFAULT;
617 
618 	/* Fault each child mr that intersects with our interval. */
619 	while (bcnt) {
620 		unsigned long idx = user_va >> MLX5_IMR_MTT_SHIFT;
621 		struct ib_umem_odp *umem_odp;
622 		struct mlx5_ib_mr *mtt;
623 		u64 len;
624 
625 		xa_lock(&imr->implicit_children);
626 		mtt = xa_load(&imr->implicit_children, idx);
627 		if (unlikely(!mtt)) {
628 			xa_unlock(&imr->implicit_children);
629 			mtt = implicit_get_child_mr(imr, idx);
630 			if (IS_ERR(mtt)) {
631 				ret = PTR_ERR(mtt);
632 				goto out;
633 			}
634 			upd_start_idx = min(upd_start_idx, idx);
635 			upd_len = idx - upd_start_idx + 1;
636 		} else {
637 			refcount_inc(&mtt->mmkey.usecount);
638 			xa_unlock(&imr->implicit_children);
639 		}
640 
641 		umem_odp = to_ib_umem_odp(mtt->umem);
642 		len = min_t(u64, user_va + bcnt, ib_umem_end(umem_odp)) -
643 		      user_va;
644 
645 		ret = pagefault_real_mr(mtt, umem_odp, user_va, len,
646 					bytes_mapped, flags);
647 
648 		mlx5r_deref_odp_mkey(&mtt->mmkey);
649 
650 		if (ret < 0)
651 			goto out;
652 		user_va += len;
653 		bcnt -= len;
654 		npages += ret;
655 	}
656 
657 	ret = npages;
658 
659 	/*
660 	 * Any time the implicit_children are changed we must perform an
661 	 * update of the xlt before exiting to ensure the HW and the
662 	 * implicit_children remains synchronized.
663 	 */
664 out:
665 	if (likely(!upd_len))
666 		return ret;
667 
668 	/*
669 	 * Notice this is not strictly ordered right, the KSM is updated after
670 	 * the implicit_children is updated, so a parallel page fault could
671 	 * see a MR that is not yet visible in the KSM.  This is similar to a
672 	 * parallel page fault seeing a MR that is being concurrently removed
673 	 * from the KSM. Both of these improbable situations are resolved
674 	 * safely by resuming the HW and then taking another page fault. The
675 	 * next pagefault handler will see the new information.
676 	 */
677 	mutex_lock(&odp_imr->umem_mutex);
678 	err = mlx5_ib_update_xlt(imr, upd_start_idx, upd_len, 0,
679 				 MLX5_IB_UPD_XLT_INDIRECT |
680 					 MLX5_IB_UPD_XLT_ATOMIC);
681 	mutex_unlock(&odp_imr->umem_mutex);
682 	if (err) {
683 		mlx5_ib_err(mr_to_mdev(imr), "Failed to update PAS\n");
684 		return err;
685 	}
686 	return ret;
687 }
688 
689 static int pagefault_dmabuf_mr(struct mlx5_ib_mr *mr, size_t bcnt,
690 			       u32 *bytes_mapped, u32 flags)
691 {
692 	struct ib_umem_dmabuf *umem_dmabuf = to_ib_umem_dmabuf(mr->umem);
693 	u32 xlt_flags = 0;
694 	int err;
695 	unsigned int page_size;
696 
697 	if (flags & MLX5_PF_FLAGS_ENABLE)
698 		xlt_flags |= MLX5_IB_UPD_XLT_ENABLE;
699 
700 	dma_resv_lock(umem_dmabuf->attach->dmabuf->resv, NULL);
701 	err = ib_umem_dmabuf_map_pages(umem_dmabuf);
702 	if (err) {
703 		dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
704 		return err;
705 	}
706 
707 	page_size = mlx5_umem_find_best_pgsz(&umem_dmabuf->umem, mkc,
708 					     log_page_size, 0,
709 					     umem_dmabuf->umem.iova);
710 	if (unlikely(page_size < PAGE_SIZE)) {
711 		ib_umem_dmabuf_unmap_pages(umem_dmabuf);
712 		err = -EINVAL;
713 	} else {
714 		err = mlx5_ib_update_mr_pas(mr, xlt_flags);
715 	}
716 	dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
717 
718 	if (err)
719 		return err;
720 
721 	if (bytes_mapped)
722 		*bytes_mapped += bcnt;
723 
724 	return ib_umem_num_pages(mr->umem);
725 }
726 
727 /*
728  * Returns:
729  *  -EFAULT: The io_virt->bcnt is not within the MR, it covers pages that are
730  *           not accessible, or the MR is no longer valid.
731  *  -EAGAIN/-ENOMEM: The operation should be retried
732  *
733  *  -EINVAL/others: General internal malfunction
734  *  >0: Number of pages mapped
735  */
736 static int pagefault_mr(struct mlx5_ib_mr *mr, u64 io_virt, size_t bcnt,
737 			u32 *bytes_mapped, u32 flags)
738 {
739 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
740 
741 	if (unlikely(io_virt < mr->mmkey.iova))
742 		return -EFAULT;
743 
744 	if (mr->umem->is_dmabuf)
745 		return pagefault_dmabuf_mr(mr, bcnt, bytes_mapped, flags);
746 
747 	if (!odp->is_implicit_odp) {
748 		u64 user_va;
749 
750 		if (check_add_overflow(io_virt - mr->mmkey.iova,
751 				       (u64)odp->umem.address, &user_va))
752 			return -EFAULT;
753 		if (unlikely(user_va >= ib_umem_end(odp) ||
754 			     ib_umem_end(odp) - user_va < bcnt))
755 			return -EFAULT;
756 		return pagefault_real_mr(mr, odp, user_va, bcnt, bytes_mapped,
757 					 flags);
758 	}
759 	return pagefault_implicit_mr(mr, odp, io_virt, bcnt, bytes_mapped,
760 				     flags);
761 }
762 
763 int mlx5_ib_init_odp_mr(struct mlx5_ib_mr *mr)
764 {
765 	int ret;
766 
767 	ret = pagefault_real_mr(mr, to_ib_umem_odp(mr->umem), mr->umem->address,
768 				mr->umem->length, NULL,
769 				MLX5_PF_FLAGS_SNAPSHOT | MLX5_PF_FLAGS_ENABLE);
770 	return ret >= 0 ? 0 : ret;
771 }
772 
773 int mlx5_ib_init_dmabuf_mr(struct mlx5_ib_mr *mr)
774 {
775 	int ret;
776 
777 	ret = pagefault_dmabuf_mr(mr, mr->umem->length, NULL,
778 				  MLX5_PF_FLAGS_ENABLE);
779 
780 	return ret >= 0 ? 0 : ret;
781 }
782 
783 struct pf_frame {
784 	struct pf_frame *next;
785 	u32 key;
786 	u64 io_virt;
787 	size_t bcnt;
788 	int depth;
789 };
790 
791 static bool mkey_is_eq(struct mlx5_core_mkey *mmkey, u32 key)
792 {
793 	if (!mmkey)
794 		return false;
795 	if (mmkey->type == MLX5_MKEY_MW)
796 		return mlx5_base_mkey(mmkey->key) == mlx5_base_mkey(key);
797 	return mmkey->key == key;
798 }
799 
800 static int get_indirect_num_descs(struct mlx5_core_mkey *mmkey)
801 {
802 	struct mlx5_ib_mw *mw;
803 	struct mlx5_ib_devx_mr *devx_mr;
804 
805 	if (mmkey->type == MLX5_MKEY_MW) {
806 		mw = container_of(mmkey, struct mlx5_ib_mw, mmkey);
807 		return mw->ndescs;
808 	}
809 
810 	devx_mr = container_of(mmkey, struct mlx5_ib_devx_mr,
811 			       mmkey);
812 	return devx_mr->ndescs;
813 }
814 
815 /*
816  * Handle a single data segment in a page-fault WQE or RDMA region.
817  *
818  * Returns number of OS pages retrieved on success. The caller may continue to
819  * the next data segment.
820  * Can return the following error codes:
821  * -EAGAIN to designate a temporary error. The caller will abort handling the
822  *  page fault and resolve it.
823  * -EFAULT when there's an error mapping the requested pages. The caller will
824  *  abort the page fault handling.
825  */
826 static int pagefault_single_data_segment(struct mlx5_ib_dev *dev,
827 					 struct ib_pd *pd, u32 key,
828 					 u64 io_virt, size_t bcnt,
829 					 u32 *bytes_committed,
830 					 u32 *bytes_mapped)
831 {
832 	int npages = 0, ret, i, outlen, cur_outlen = 0, depth = 0;
833 	struct pf_frame *head = NULL, *frame;
834 	struct mlx5_core_mkey *mmkey;
835 	struct mlx5_ib_mr *mr;
836 	struct mlx5_klm *pklm;
837 	u32 *out = NULL;
838 	size_t offset;
839 	int ndescs;
840 
841 	io_virt += *bytes_committed;
842 	bcnt -= *bytes_committed;
843 
844 next_mr:
845 	xa_lock(&dev->odp_mkeys);
846 	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(key));
847 	if (!mmkey) {
848 		xa_unlock(&dev->odp_mkeys);
849 		mlx5_ib_dbg(
850 			dev,
851 			"skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
852 			key);
853 		if (bytes_mapped)
854 			*bytes_mapped += bcnt;
855 		/*
856 		 * The user could specify a SGL with multiple lkeys and only
857 		 * some of them are ODP. Treat the non-ODP ones as fully
858 		 * faulted.
859 		 */
860 		ret = 0;
861 		goto end;
862 	}
863 	refcount_inc(&mmkey->usecount);
864 	xa_unlock(&dev->odp_mkeys);
865 
866 	if (!mkey_is_eq(mmkey, key)) {
867 		mlx5_ib_dbg(dev, "failed to find mkey %x\n", key);
868 		ret = -EFAULT;
869 		goto end;
870 	}
871 
872 	switch (mmkey->type) {
873 	case MLX5_MKEY_MR:
874 		mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
875 
876 		ret = pagefault_mr(mr, io_virt, bcnt, bytes_mapped, 0);
877 		if (ret < 0)
878 			goto end;
879 
880 		mlx5_update_odp_stats(mr, faults, ret);
881 
882 		npages += ret;
883 		ret = 0;
884 		break;
885 
886 	case MLX5_MKEY_MW:
887 	case MLX5_MKEY_INDIRECT_DEVX:
888 		ndescs = get_indirect_num_descs(mmkey);
889 
890 		if (depth >= MLX5_CAP_GEN(dev->mdev, max_indirection)) {
891 			mlx5_ib_dbg(dev, "indirection level exceeded\n");
892 			ret = -EFAULT;
893 			goto end;
894 		}
895 
896 		outlen = MLX5_ST_SZ_BYTES(query_mkey_out) +
897 			sizeof(*pklm) * (ndescs - 2);
898 
899 		if (outlen > cur_outlen) {
900 			kfree(out);
901 			out = kzalloc(outlen, GFP_KERNEL);
902 			if (!out) {
903 				ret = -ENOMEM;
904 				goto end;
905 			}
906 			cur_outlen = outlen;
907 		}
908 
909 		pklm = (struct mlx5_klm *)MLX5_ADDR_OF(query_mkey_out, out,
910 						       bsf0_klm0_pas_mtt0_1);
911 
912 		ret = mlx5_core_query_mkey(dev->mdev, mmkey, out, outlen);
913 		if (ret)
914 			goto end;
915 
916 		offset = io_virt - MLX5_GET64(query_mkey_out, out,
917 					      memory_key_mkey_entry.start_addr);
918 
919 		for (i = 0; bcnt && i < ndescs; i++, pklm++) {
920 			if (offset >= be32_to_cpu(pklm->bcount)) {
921 				offset -= be32_to_cpu(pklm->bcount);
922 				continue;
923 			}
924 
925 			frame = kzalloc(sizeof(*frame), GFP_KERNEL);
926 			if (!frame) {
927 				ret = -ENOMEM;
928 				goto end;
929 			}
930 
931 			frame->key = be32_to_cpu(pklm->key);
932 			frame->io_virt = be64_to_cpu(pklm->va) + offset;
933 			frame->bcnt = min_t(size_t, bcnt,
934 					    be32_to_cpu(pklm->bcount) - offset);
935 			frame->depth = depth + 1;
936 			frame->next = head;
937 			head = frame;
938 
939 			bcnt -= frame->bcnt;
940 			offset = 0;
941 		}
942 		break;
943 
944 	default:
945 		mlx5_ib_dbg(dev, "wrong mkey type %d\n", mmkey->type);
946 		ret = -EFAULT;
947 		goto end;
948 	}
949 
950 	if (head) {
951 		frame = head;
952 		head = frame->next;
953 
954 		key = frame->key;
955 		io_virt = frame->io_virt;
956 		bcnt = frame->bcnt;
957 		depth = frame->depth;
958 		kfree(frame);
959 
960 		mlx5r_deref_odp_mkey(mmkey);
961 		goto next_mr;
962 	}
963 
964 end:
965 	if (mmkey)
966 		mlx5r_deref_odp_mkey(mmkey);
967 	while (head) {
968 		frame = head;
969 		head = frame->next;
970 		kfree(frame);
971 	}
972 	kfree(out);
973 
974 	*bytes_committed = 0;
975 	return ret ? ret : npages;
976 }
977 
978 /*
979  * Parse a series of data segments for page fault handling.
980  *
981  * @dev:  Pointer to mlx5 IB device
982  * @pfault: contains page fault information.
983  * @wqe: points at the first data segment in the WQE.
984  * @wqe_end: points after the end of the WQE.
985  * @bytes_mapped: receives the number of bytes that the function was able to
986  *                map. This allows the caller to decide intelligently whether
987  *                enough memory was mapped to resolve the page fault
988  *                successfully (e.g. enough for the next MTU, or the entire
989  *                WQE).
990  * @total_wqe_bytes: receives the total data size of this WQE in bytes (minus
991  *                   the committed bytes).
992  * @receive_queue: receive WQE end of sg list
993  *
994  * Returns the number of pages loaded if positive, zero for an empty WQE, or a
995  * negative error code.
996  */
997 static int pagefault_data_segments(struct mlx5_ib_dev *dev,
998 				   struct mlx5_pagefault *pfault,
999 				   void *wqe,
1000 				   void *wqe_end, u32 *bytes_mapped,
1001 				   u32 *total_wqe_bytes, bool receive_queue)
1002 {
1003 	int ret = 0, npages = 0;
1004 	u64 io_virt;
1005 	u32 key;
1006 	u32 byte_count;
1007 	size_t bcnt;
1008 	int inline_segment;
1009 
1010 	if (bytes_mapped)
1011 		*bytes_mapped = 0;
1012 	if (total_wqe_bytes)
1013 		*total_wqe_bytes = 0;
1014 
1015 	while (wqe < wqe_end) {
1016 		struct mlx5_wqe_data_seg *dseg = wqe;
1017 
1018 		io_virt = be64_to_cpu(dseg->addr);
1019 		key = be32_to_cpu(dseg->lkey);
1020 		byte_count = be32_to_cpu(dseg->byte_count);
1021 		inline_segment = !!(byte_count &  MLX5_INLINE_SEG);
1022 		bcnt	       = byte_count & ~MLX5_INLINE_SEG;
1023 
1024 		if (inline_segment) {
1025 			bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK;
1026 			wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt,
1027 				     16);
1028 		} else {
1029 			wqe += sizeof(*dseg);
1030 		}
1031 
1032 		/* receive WQE end of sg list. */
1033 		if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY &&
1034 		    io_virt == 0)
1035 			break;
1036 
1037 		if (!inline_segment && total_wqe_bytes) {
1038 			*total_wqe_bytes += bcnt - min_t(size_t, bcnt,
1039 					pfault->bytes_committed);
1040 		}
1041 
1042 		/* A zero length data segment designates a length of 2GB. */
1043 		if (bcnt == 0)
1044 			bcnt = 1U << 31;
1045 
1046 		if (inline_segment || bcnt <= pfault->bytes_committed) {
1047 			pfault->bytes_committed -=
1048 				min_t(size_t, bcnt,
1049 				      pfault->bytes_committed);
1050 			continue;
1051 		}
1052 
1053 		ret = pagefault_single_data_segment(dev, NULL, key,
1054 						    io_virt, bcnt,
1055 						    &pfault->bytes_committed,
1056 						    bytes_mapped);
1057 		if (ret < 0)
1058 			break;
1059 		npages += ret;
1060 	}
1061 
1062 	return ret < 0 ? ret : npages;
1063 }
1064 
1065 /*
1066  * Parse initiator WQE. Advances the wqe pointer to point at the
1067  * scatter-gather list, and set wqe_end to the end of the WQE.
1068  */
1069 static int mlx5_ib_mr_initiator_pfault_handler(
1070 	struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault,
1071 	struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length)
1072 {
1073 	struct mlx5_wqe_ctrl_seg *ctrl = *wqe;
1074 	u16 wqe_index = pfault->wqe.wqe_index;
1075 	struct mlx5_base_av *av;
1076 	unsigned ds, opcode;
1077 	u32 qpn = qp->trans_qp.base.mqp.qpn;
1078 
1079 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
1080 	if (ds * MLX5_WQE_DS_UNITS > wqe_length) {
1081 		mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n",
1082 			    ds, wqe_length);
1083 		return -EFAULT;
1084 	}
1085 
1086 	if (ds == 0) {
1087 		mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n",
1088 			    wqe_index, qpn);
1089 		return -EFAULT;
1090 	}
1091 
1092 	*wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS;
1093 	*wqe += sizeof(*ctrl);
1094 
1095 	opcode = be32_to_cpu(ctrl->opmod_idx_opcode) &
1096 		 MLX5_WQE_CTRL_OPCODE_MASK;
1097 
1098 	if (qp->type == IB_QPT_XRC_INI)
1099 		*wqe += sizeof(struct mlx5_wqe_xrc_seg);
1100 
1101 	if (qp->type == IB_QPT_UD || qp->type == MLX5_IB_QPT_DCI) {
1102 		av = *wqe;
1103 		if (av->dqp_dct & cpu_to_be32(MLX5_EXTENDED_UD_AV))
1104 			*wqe += sizeof(struct mlx5_av);
1105 		else
1106 			*wqe += sizeof(struct mlx5_base_av);
1107 	}
1108 
1109 	switch (opcode) {
1110 	case MLX5_OPCODE_RDMA_WRITE:
1111 	case MLX5_OPCODE_RDMA_WRITE_IMM:
1112 	case MLX5_OPCODE_RDMA_READ:
1113 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1114 		break;
1115 	case MLX5_OPCODE_ATOMIC_CS:
1116 	case MLX5_OPCODE_ATOMIC_FA:
1117 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1118 		*wqe += sizeof(struct mlx5_wqe_atomic_seg);
1119 		break;
1120 	}
1121 
1122 	return 0;
1123 }
1124 
1125 /*
1126  * Parse responder WQE and set wqe_end to the end of the WQE.
1127  */
1128 static int mlx5_ib_mr_responder_pfault_handler_srq(struct mlx5_ib_dev *dev,
1129 						   struct mlx5_ib_srq *srq,
1130 						   void **wqe, void **wqe_end,
1131 						   int wqe_length)
1132 {
1133 	int wqe_size = 1 << srq->msrq.wqe_shift;
1134 
1135 	if (wqe_size > wqe_length) {
1136 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1137 		return -EFAULT;
1138 	}
1139 
1140 	*wqe_end = *wqe + wqe_size;
1141 	*wqe += sizeof(struct mlx5_wqe_srq_next_seg);
1142 
1143 	return 0;
1144 }
1145 
1146 static int mlx5_ib_mr_responder_pfault_handler_rq(struct mlx5_ib_dev *dev,
1147 						  struct mlx5_ib_qp *qp,
1148 						  void *wqe, void **wqe_end,
1149 						  int wqe_length)
1150 {
1151 	struct mlx5_ib_wq *wq = &qp->rq;
1152 	int wqe_size = 1 << wq->wqe_shift;
1153 
1154 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE) {
1155 		mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n");
1156 		return -EFAULT;
1157 	}
1158 
1159 	if (wqe_size > wqe_length) {
1160 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1161 		return -EFAULT;
1162 	}
1163 
1164 	*wqe_end = wqe + wqe_size;
1165 
1166 	return 0;
1167 }
1168 
1169 static inline struct mlx5_core_rsc_common *odp_get_rsc(struct mlx5_ib_dev *dev,
1170 						       u32 wq_num, int pf_type)
1171 {
1172 	struct mlx5_core_rsc_common *common = NULL;
1173 	struct mlx5_core_srq *srq;
1174 
1175 	switch (pf_type) {
1176 	case MLX5_WQE_PF_TYPE_RMP:
1177 		srq = mlx5_cmd_get_srq(dev, wq_num);
1178 		if (srq)
1179 			common = &srq->common;
1180 		break;
1181 	case MLX5_WQE_PF_TYPE_REQ_SEND_OR_WRITE:
1182 	case MLX5_WQE_PF_TYPE_RESP:
1183 	case MLX5_WQE_PF_TYPE_REQ_READ_OR_ATOMIC:
1184 		common = mlx5_core_res_hold(dev, wq_num, MLX5_RES_QP);
1185 		break;
1186 	default:
1187 		break;
1188 	}
1189 
1190 	return common;
1191 }
1192 
1193 static inline struct mlx5_ib_qp *res_to_qp(struct mlx5_core_rsc_common *res)
1194 {
1195 	struct mlx5_core_qp *mqp = (struct mlx5_core_qp *)res;
1196 
1197 	return to_mibqp(mqp);
1198 }
1199 
1200 static inline struct mlx5_ib_srq *res_to_srq(struct mlx5_core_rsc_common *res)
1201 {
1202 	struct mlx5_core_srq *msrq =
1203 		container_of(res, struct mlx5_core_srq, common);
1204 
1205 	return to_mibsrq(msrq);
1206 }
1207 
1208 static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
1209 					  struct mlx5_pagefault *pfault)
1210 {
1211 	bool sq = pfault->type & MLX5_PFAULT_REQUESTOR;
1212 	u16 wqe_index = pfault->wqe.wqe_index;
1213 	void *wqe, *wqe_start = NULL, *wqe_end = NULL;
1214 	u32 bytes_mapped, total_wqe_bytes;
1215 	struct mlx5_core_rsc_common *res;
1216 	int resume_with_error = 1;
1217 	struct mlx5_ib_qp *qp;
1218 	size_t bytes_copied;
1219 	int ret = 0;
1220 
1221 	res = odp_get_rsc(dev, pfault->wqe.wq_num, pfault->type);
1222 	if (!res) {
1223 		mlx5_ib_dbg(dev, "wqe page fault for missing resource %d\n", pfault->wqe.wq_num);
1224 		return;
1225 	}
1226 
1227 	if (res->res != MLX5_RES_QP && res->res != MLX5_RES_SRQ &&
1228 	    res->res != MLX5_RES_XSRQ) {
1229 		mlx5_ib_err(dev, "wqe page fault for unsupported type %d\n",
1230 			    pfault->type);
1231 		goto resolve_page_fault;
1232 	}
1233 
1234 	wqe_start = (void *)__get_free_page(GFP_KERNEL);
1235 	if (!wqe_start) {
1236 		mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n");
1237 		goto resolve_page_fault;
1238 	}
1239 
1240 	wqe = wqe_start;
1241 	qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL;
1242 	if (qp && sq) {
1243 		ret = mlx5_ib_read_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE,
1244 					  &bytes_copied);
1245 		if (ret)
1246 			goto read_user;
1247 		ret = mlx5_ib_mr_initiator_pfault_handler(
1248 			dev, pfault, qp, &wqe, &wqe_end, bytes_copied);
1249 	} else if (qp && !sq) {
1250 		ret = mlx5_ib_read_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE,
1251 					  &bytes_copied);
1252 		if (ret)
1253 			goto read_user;
1254 		ret = mlx5_ib_mr_responder_pfault_handler_rq(
1255 			dev, qp, wqe, &wqe_end, bytes_copied);
1256 	} else if (!qp) {
1257 		struct mlx5_ib_srq *srq = res_to_srq(res);
1258 
1259 		ret = mlx5_ib_read_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE,
1260 					   &bytes_copied);
1261 		if (ret)
1262 			goto read_user;
1263 		ret = mlx5_ib_mr_responder_pfault_handler_srq(
1264 			dev, srq, &wqe, &wqe_end, bytes_copied);
1265 	}
1266 
1267 	if (ret < 0 || wqe >= wqe_end)
1268 		goto resolve_page_fault;
1269 
1270 	ret = pagefault_data_segments(dev, pfault, wqe, wqe_end, &bytes_mapped,
1271 				      &total_wqe_bytes, !sq);
1272 	if (ret == -EAGAIN)
1273 		goto out;
1274 
1275 	if (ret < 0 || total_wqe_bytes > bytes_mapped)
1276 		goto resolve_page_fault;
1277 
1278 out:
1279 	ret = 0;
1280 	resume_with_error = 0;
1281 
1282 read_user:
1283 	if (ret)
1284 		mlx5_ib_err(
1285 			dev,
1286 			"Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n",
1287 			ret, wqe_index, pfault->token);
1288 
1289 resolve_page_fault:
1290 	mlx5_ib_page_fault_resume(dev, pfault, resume_with_error);
1291 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n",
1292 		    pfault->wqe.wq_num, resume_with_error,
1293 		    pfault->type);
1294 	mlx5_core_res_put(res);
1295 	free_page((unsigned long)wqe_start);
1296 }
1297 
1298 static int pages_in_range(u64 address, u32 length)
1299 {
1300 	return (ALIGN(address + length, PAGE_SIZE) -
1301 		(address & PAGE_MASK)) >> PAGE_SHIFT;
1302 }
1303 
1304 static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev,
1305 					   struct mlx5_pagefault *pfault)
1306 {
1307 	u64 address;
1308 	u32 length;
1309 	u32 prefetch_len = pfault->bytes_committed;
1310 	int prefetch_activated = 0;
1311 	u32 rkey = pfault->rdma.r_key;
1312 	int ret;
1313 
1314 	/* The RDMA responder handler handles the page fault in two parts.
1315 	 * First it brings the necessary pages for the current packet
1316 	 * (and uses the pfault context), and then (after resuming the QP)
1317 	 * prefetches more pages. The second operation cannot use the pfault
1318 	 * context and therefore uses the dummy_pfault context allocated on
1319 	 * the stack */
1320 	pfault->rdma.rdma_va += pfault->bytes_committed;
1321 	pfault->rdma.rdma_op_len -= min(pfault->bytes_committed,
1322 					 pfault->rdma.rdma_op_len);
1323 	pfault->bytes_committed = 0;
1324 
1325 	address = pfault->rdma.rdma_va;
1326 	length  = pfault->rdma.rdma_op_len;
1327 
1328 	/* For some operations, the hardware cannot tell the exact message
1329 	 * length, and in those cases it reports zero. Use prefetch
1330 	 * logic. */
1331 	if (length == 0) {
1332 		prefetch_activated = 1;
1333 		length = pfault->rdma.packet_size;
1334 		prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len);
1335 	}
1336 
1337 	ret = pagefault_single_data_segment(dev, NULL, rkey, address, length,
1338 					    &pfault->bytes_committed, NULL);
1339 	if (ret == -EAGAIN) {
1340 		/* We're racing with an invalidation, don't prefetch */
1341 		prefetch_activated = 0;
1342 	} else if (ret < 0 || pages_in_range(address, length) > ret) {
1343 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1344 		if (ret != -ENOENT)
1345 			mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n",
1346 				    ret, pfault->token, pfault->type);
1347 		return;
1348 	}
1349 
1350 	mlx5_ib_page_fault_resume(dev, pfault, 0);
1351 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n",
1352 		    pfault->token, pfault->type,
1353 		    prefetch_activated);
1354 
1355 	/* At this point, there might be a new pagefault already arriving in
1356 	 * the eq, switch to the dummy pagefault for the rest of the
1357 	 * processing. We're still OK with the objects being alive as the
1358 	 * work-queue is being fenced. */
1359 
1360 	if (prefetch_activated) {
1361 		u32 bytes_committed = 0;
1362 
1363 		ret = pagefault_single_data_segment(dev, NULL, rkey, address,
1364 						    prefetch_len,
1365 						    &bytes_committed, NULL);
1366 		if (ret < 0 && ret != -EAGAIN) {
1367 			mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n",
1368 				    ret, pfault->token, address, prefetch_len);
1369 		}
1370 	}
1371 }
1372 
1373 static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault)
1374 {
1375 	u8 event_subtype = pfault->event_subtype;
1376 
1377 	switch (event_subtype) {
1378 	case MLX5_PFAULT_SUBTYPE_WQE:
1379 		mlx5_ib_mr_wqe_pfault_handler(dev, pfault);
1380 		break;
1381 	case MLX5_PFAULT_SUBTYPE_RDMA:
1382 		mlx5_ib_mr_rdma_pfault_handler(dev, pfault);
1383 		break;
1384 	default:
1385 		mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n",
1386 			    event_subtype);
1387 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1388 	}
1389 }
1390 
1391 static void mlx5_ib_eqe_pf_action(struct work_struct *work)
1392 {
1393 	struct mlx5_pagefault *pfault = container_of(work,
1394 						     struct mlx5_pagefault,
1395 						     work);
1396 	struct mlx5_ib_pf_eq *eq = pfault->eq;
1397 
1398 	mlx5_ib_pfault(eq->dev, pfault);
1399 	mempool_free(pfault, eq->pool);
1400 }
1401 
1402 static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq)
1403 {
1404 	struct mlx5_eqe_page_fault *pf_eqe;
1405 	struct mlx5_pagefault *pfault;
1406 	struct mlx5_eqe *eqe;
1407 	int cc = 0;
1408 
1409 	while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) {
1410 		pfault = mempool_alloc(eq->pool, GFP_ATOMIC);
1411 		if (!pfault) {
1412 			schedule_work(&eq->work);
1413 			break;
1414 		}
1415 
1416 		pf_eqe = &eqe->data.page_fault;
1417 		pfault->event_subtype = eqe->sub_type;
1418 		pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);
1419 
1420 		mlx5_ib_dbg(eq->dev,
1421 			    "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
1422 			    eqe->sub_type, pfault->bytes_committed);
1423 
1424 		switch (eqe->sub_type) {
1425 		case MLX5_PFAULT_SUBTYPE_RDMA:
1426 			/* RDMA based event */
1427 			pfault->type =
1428 				be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
1429 			pfault->token =
1430 				be32_to_cpu(pf_eqe->rdma.pftype_token) &
1431 				MLX5_24BIT_MASK;
1432 			pfault->rdma.r_key =
1433 				be32_to_cpu(pf_eqe->rdma.r_key);
1434 			pfault->rdma.packet_size =
1435 				be16_to_cpu(pf_eqe->rdma.packet_length);
1436 			pfault->rdma.rdma_op_len =
1437 				be32_to_cpu(pf_eqe->rdma.rdma_op_len);
1438 			pfault->rdma.rdma_va =
1439 				be64_to_cpu(pf_eqe->rdma.rdma_va);
1440 			mlx5_ib_dbg(eq->dev,
1441 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
1442 				    pfault->type, pfault->token,
1443 				    pfault->rdma.r_key);
1444 			mlx5_ib_dbg(eq->dev,
1445 				    "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
1446 				    pfault->rdma.rdma_op_len,
1447 				    pfault->rdma.rdma_va);
1448 			break;
1449 
1450 		case MLX5_PFAULT_SUBTYPE_WQE:
1451 			/* WQE based event */
1452 			pfault->type =
1453 				(be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7;
1454 			pfault->token =
1455 				be32_to_cpu(pf_eqe->wqe.token);
1456 			pfault->wqe.wq_num =
1457 				be32_to_cpu(pf_eqe->wqe.pftype_wq) &
1458 				MLX5_24BIT_MASK;
1459 			pfault->wqe.wqe_index =
1460 				be16_to_cpu(pf_eqe->wqe.wqe_index);
1461 			pfault->wqe.packet_size =
1462 				be16_to_cpu(pf_eqe->wqe.packet_length);
1463 			mlx5_ib_dbg(eq->dev,
1464 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
1465 				    pfault->type, pfault->token,
1466 				    pfault->wqe.wq_num,
1467 				    pfault->wqe.wqe_index);
1468 			break;
1469 
1470 		default:
1471 			mlx5_ib_warn(eq->dev,
1472 				     "Unsupported page fault event sub-type: 0x%02hhx\n",
1473 				     eqe->sub_type);
1474 			/* Unsupported page faults should still be
1475 			 * resolved by the page fault handler
1476 			 */
1477 		}
1478 
1479 		pfault->eq = eq;
1480 		INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action);
1481 		queue_work(eq->wq, &pfault->work);
1482 
1483 		cc = mlx5_eq_update_cc(eq->core, ++cc);
1484 	}
1485 
1486 	mlx5_eq_update_ci(eq->core, cc, 1);
1487 }
1488 
1489 static int mlx5_ib_eq_pf_int(struct notifier_block *nb, unsigned long type,
1490 			     void *data)
1491 {
1492 	struct mlx5_ib_pf_eq *eq =
1493 		container_of(nb, struct mlx5_ib_pf_eq, irq_nb);
1494 	unsigned long flags;
1495 
1496 	if (spin_trylock_irqsave(&eq->lock, flags)) {
1497 		mlx5_ib_eq_pf_process(eq);
1498 		spin_unlock_irqrestore(&eq->lock, flags);
1499 	} else {
1500 		schedule_work(&eq->work);
1501 	}
1502 
1503 	return IRQ_HANDLED;
1504 }
1505 
1506 /* mempool_refill() was proposed but unfortunately wasn't accepted
1507  * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
1508  * Cheap workaround.
1509  */
1510 static void mempool_refill(mempool_t *pool)
1511 {
1512 	while (pool->curr_nr < pool->min_nr)
1513 		mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
1514 }
1515 
1516 static void mlx5_ib_eq_pf_action(struct work_struct *work)
1517 {
1518 	struct mlx5_ib_pf_eq *eq =
1519 		container_of(work, struct mlx5_ib_pf_eq, work);
1520 
1521 	mempool_refill(eq->pool);
1522 
1523 	spin_lock_irq(&eq->lock);
1524 	mlx5_ib_eq_pf_process(eq);
1525 	spin_unlock_irq(&eq->lock);
1526 }
1527 
1528 enum {
1529 	MLX5_IB_NUM_PF_EQE	= 0x1000,
1530 	MLX5_IB_NUM_PF_DRAIN	= 64,
1531 };
1532 
1533 int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1534 {
1535 	struct mlx5_eq_param param = {};
1536 	int err = 0;
1537 
1538 	mutex_lock(&dev->odp_eq_mutex);
1539 	if (eq->core)
1540 		goto unlock;
1541 	INIT_WORK(&eq->work, mlx5_ib_eq_pf_action);
1542 	spin_lock_init(&eq->lock);
1543 	eq->dev = dev;
1544 
1545 	eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN,
1546 					       sizeof(struct mlx5_pagefault));
1547 	if (!eq->pool) {
1548 		err = -ENOMEM;
1549 		goto unlock;
1550 	}
1551 
1552 	eq->wq = alloc_workqueue("mlx5_ib_page_fault",
1553 				 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM,
1554 				 MLX5_NUM_CMD_EQE);
1555 	if (!eq->wq) {
1556 		err = -ENOMEM;
1557 		goto err_mempool;
1558 	}
1559 
1560 	eq->irq_nb.notifier_call = mlx5_ib_eq_pf_int;
1561 	param = (struct mlx5_eq_param) {
1562 		.nent = MLX5_IB_NUM_PF_EQE,
1563 	};
1564 	param.mask[0] = 1ull << MLX5_EVENT_TYPE_PAGE_FAULT;
1565 	if (!zalloc_cpumask_var(&param.affinity, GFP_KERNEL)) {
1566 		err = -ENOMEM;
1567 		goto err_wq;
1568 	}
1569 	eq->core = mlx5_eq_create_generic(dev->mdev, &param);
1570 	free_cpumask_var(param.affinity);
1571 	if (IS_ERR(eq->core)) {
1572 		err = PTR_ERR(eq->core);
1573 		goto err_wq;
1574 	}
1575 	err = mlx5_eq_enable(dev->mdev, eq->core, &eq->irq_nb);
1576 	if (err) {
1577 		mlx5_ib_err(dev, "failed to enable odp EQ %d\n", err);
1578 		goto err_eq;
1579 	}
1580 
1581 	mutex_unlock(&dev->odp_eq_mutex);
1582 	return 0;
1583 err_eq:
1584 	mlx5_eq_destroy_generic(dev->mdev, eq->core);
1585 err_wq:
1586 	eq->core = NULL;
1587 	destroy_workqueue(eq->wq);
1588 err_mempool:
1589 	mempool_destroy(eq->pool);
1590 unlock:
1591 	mutex_unlock(&dev->odp_eq_mutex);
1592 	return err;
1593 }
1594 
1595 static int
1596 mlx5_ib_odp_destroy_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1597 {
1598 	int err;
1599 
1600 	if (!eq->core)
1601 		return 0;
1602 	mlx5_eq_disable(dev->mdev, eq->core, &eq->irq_nb);
1603 	err = mlx5_eq_destroy_generic(dev->mdev, eq->core);
1604 	cancel_work_sync(&eq->work);
1605 	destroy_workqueue(eq->wq);
1606 	mempool_destroy(eq->pool);
1607 
1608 	return err;
1609 }
1610 
1611 void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
1612 {
1613 	if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1614 		return;
1615 
1616 	switch (ent->order - 2) {
1617 	case MLX5_IMR_MTT_CACHE_ENTRY:
1618 		ent->page = PAGE_SHIFT;
1619 		ent->xlt = MLX5_IMR_MTT_ENTRIES *
1620 			   sizeof(struct mlx5_mtt) /
1621 			   MLX5_IB_UMR_OCTOWORD;
1622 		ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
1623 		ent->limit = 0;
1624 		break;
1625 
1626 	case MLX5_IMR_KSM_CACHE_ENTRY:
1627 		ent->page = MLX5_KSM_PAGE_SHIFT;
1628 		ent->xlt = mlx5_imr_ksm_entries *
1629 			   sizeof(struct mlx5_klm) /
1630 			   MLX5_IB_UMR_OCTOWORD;
1631 		ent->access_mode = MLX5_MKC_ACCESS_MODE_KSM;
1632 		ent->limit = 0;
1633 		break;
1634 	}
1635 }
1636 
1637 static const struct ib_device_ops mlx5_ib_dev_odp_ops = {
1638 	.advise_mr = mlx5_ib_advise_mr,
1639 };
1640 
1641 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
1642 {
1643 	int ret = 0;
1644 
1645 	internal_fill_odp_caps(dev);
1646 
1647 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1648 		return ret;
1649 
1650 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_odp_ops);
1651 
1652 	if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) {
1653 		ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey);
1654 		if (ret) {
1655 			mlx5_ib_err(dev, "Error getting null_mkey %d\n", ret);
1656 			return ret;
1657 		}
1658 	}
1659 
1660 	mutex_init(&dev->odp_eq_mutex);
1661 	return ret;
1662 }
1663 
1664 void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev)
1665 {
1666 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1667 		return;
1668 
1669 	mlx5_ib_odp_destroy_eq(dev, &dev->odp_pf_eq);
1670 }
1671 
1672 int mlx5_ib_odp_init(void)
1673 {
1674 	mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) -
1675 				       MLX5_IMR_MTT_BITS);
1676 
1677 	return 0;
1678 }
1679 
1680 struct prefetch_mr_work {
1681 	struct work_struct work;
1682 	u32 pf_flags;
1683 	u32 num_sge;
1684 	struct {
1685 		u64 io_virt;
1686 		struct mlx5_ib_mr *mr;
1687 		size_t length;
1688 	} frags[];
1689 };
1690 
1691 static void destroy_prefetch_work(struct prefetch_mr_work *work)
1692 {
1693 	u32 i;
1694 
1695 	for (i = 0; i < work->num_sge; ++i)
1696 		mlx5r_deref_odp_mkey(&work->frags[i].mr->mmkey);
1697 
1698 	kvfree(work);
1699 }
1700 
1701 static struct mlx5_ib_mr *
1702 get_prefetchable_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
1703 		    u32 lkey)
1704 {
1705 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1706 	struct mlx5_core_mkey *mmkey;
1707 	struct mlx5_ib_mr *mr = NULL;
1708 
1709 	xa_lock(&dev->odp_mkeys);
1710 	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(lkey));
1711 	if (!mmkey || mmkey->key != lkey || mmkey->type != MLX5_MKEY_MR)
1712 		goto end;
1713 
1714 	mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
1715 
1716 	if (mr->ibmr.pd != pd) {
1717 		mr = NULL;
1718 		goto end;
1719 	}
1720 
1721 	/* prefetch with write-access must be supported by the MR */
1722 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
1723 	    !mr->umem->writable) {
1724 		mr = NULL;
1725 		goto end;
1726 	}
1727 
1728 	refcount_inc(&mmkey->usecount);
1729 end:
1730 	xa_unlock(&dev->odp_mkeys);
1731 	return mr;
1732 }
1733 
1734 static void mlx5_ib_prefetch_mr_work(struct work_struct *w)
1735 {
1736 	struct prefetch_mr_work *work =
1737 		container_of(w, struct prefetch_mr_work, work);
1738 	u32 bytes_mapped = 0;
1739 	int ret;
1740 	u32 i;
1741 
1742 	/* We rely on IB/core that work is executed if we have num_sge != 0 only. */
1743 	WARN_ON(!work->num_sge);
1744 	for (i = 0; i < work->num_sge; ++i) {
1745 		ret = pagefault_mr(work->frags[i].mr, work->frags[i].io_virt,
1746 				   work->frags[i].length, &bytes_mapped,
1747 				   work->pf_flags);
1748 		if (ret <= 0)
1749 			continue;
1750 		mlx5_update_odp_stats(work->frags[i].mr, prefetch, ret);
1751 	}
1752 
1753 	destroy_prefetch_work(work);
1754 }
1755 
1756 static bool init_prefetch_work(struct ib_pd *pd,
1757 			       enum ib_uverbs_advise_mr_advice advice,
1758 			       u32 pf_flags, struct prefetch_mr_work *work,
1759 			       struct ib_sge *sg_list, u32 num_sge)
1760 {
1761 	u32 i;
1762 
1763 	INIT_WORK(&work->work, mlx5_ib_prefetch_mr_work);
1764 	work->pf_flags = pf_flags;
1765 
1766 	for (i = 0; i < num_sge; ++i) {
1767 		work->frags[i].io_virt = sg_list[i].addr;
1768 		work->frags[i].length = sg_list[i].length;
1769 		work->frags[i].mr =
1770 			get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1771 		if (!work->frags[i].mr) {
1772 			work->num_sge = i;
1773 			return false;
1774 		}
1775 	}
1776 	work->num_sge = num_sge;
1777 	return true;
1778 }
1779 
1780 static int mlx5_ib_prefetch_sg_list(struct ib_pd *pd,
1781 				    enum ib_uverbs_advise_mr_advice advice,
1782 				    u32 pf_flags, struct ib_sge *sg_list,
1783 				    u32 num_sge)
1784 {
1785 	u32 bytes_mapped = 0;
1786 	int ret = 0;
1787 	u32 i;
1788 
1789 	for (i = 0; i < num_sge; ++i) {
1790 		struct mlx5_ib_mr *mr;
1791 
1792 		mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1793 		if (!mr)
1794 			return -ENOENT;
1795 		ret = pagefault_mr(mr, sg_list[i].addr, sg_list[i].length,
1796 				   &bytes_mapped, pf_flags);
1797 		if (ret < 0) {
1798 			mlx5r_deref_odp_mkey(&mr->mmkey);
1799 			return ret;
1800 		}
1801 		mlx5_update_odp_stats(mr, prefetch, ret);
1802 		mlx5r_deref_odp_mkey(&mr->mmkey);
1803 	}
1804 
1805 	return 0;
1806 }
1807 
1808 int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
1809 			       enum ib_uverbs_advise_mr_advice advice,
1810 			       u32 flags, struct ib_sge *sg_list, u32 num_sge)
1811 {
1812 	u32 pf_flags = 0;
1813 	struct prefetch_mr_work *work;
1814 
1815 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH)
1816 		pf_flags |= MLX5_PF_FLAGS_DOWNGRADE;
1817 
1818 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT)
1819 		pf_flags |= MLX5_PF_FLAGS_SNAPSHOT;
1820 
1821 	if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH)
1822 		return mlx5_ib_prefetch_sg_list(pd, advice, pf_flags, sg_list,
1823 						num_sge);
1824 
1825 	work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL);
1826 	if (!work)
1827 		return -ENOMEM;
1828 
1829 	if (!init_prefetch_work(pd, advice, pf_flags, work, sg_list, num_sge)) {
1830 		destroy_prefetch_work(work);
1831 		return -EINVAL;
1832 	}
1833 	queue_work(system_unbound_wq, &work->work);
1834 	return 0;
1835 }
1836