xref: /openbmc/linux/drivers/infiniband/hw/mlx5/odp.c (revision 3d37ef41)
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 	ret = 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 					     struct ib_udata *udata,
482 					     int access_flags)
483 {
484 	struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device);
485 	struct ib_umem_odp *umem_odp;
486 	struct mlx5_ib_mr *imr;
487 	int err;
488 
489 	if (!mlx5_ib_can_load_pas_with_umr(dev,
490 					   MLX5_IMR_MTT_ENTRIES * PAGE_SIZE))
491 		return ERR_PTR(-EOPNOTSUPP);
492 
493 	umem_odp = ib_umem_odp_alloc_implicit(&dev->ib_dev, access_flags);
494 	if (IS_ERR(umem_odp))
495 		return ERR_CAST(umem_odp);
496 
497 	imr = mlx5_mr_cache_alloc(dev, MLX5_IMR_KSM_CACHE_ENTRY, access_flags);
498 	if (IS_ERR(imr)) {
499 		ib_umem_odp_release(umem_odp);
500 		return imr;
501 	}
502 
503 	imr->ibmr.pd = &pd->ibpd;
504 	imr->mmkey.iova = 0;
505 	imr->umem = &umem_odp->umem;
506 	imr->ibmr.lkey = imr->mmkey.key;
507 	imr->ibmr.rkey = imr->mmkey.key;
508 	imr->ibmr.device = &dev->ib_dev;
509 	imr->umem = &umem_odp->umem;
510 	imr->is_odp_implicit = true;
511 	xa_init(&imr->implicit_children);
512 
513 	err = mlx5_ib_update_xlt(imr, 0,
514 				 mlx5_imr_ksm_entries,
515 				 MLX5_KSM_PAGE_SHIFT,
516 				 MLX5_IB_UPD_XLT_INDIRECT |
517 				 MLX5_IB_UPD_XLT_ZAP |
518 				 MLX5_IB_UPD_XLT_ENABLE);
519 	if (err)
520 		goto out_mr;
521 
522 	err = mlx5r_store_odp_mkey(dev, &imr->mmkey);
523 	if (err)
524 		goto out_mr;
525 
526 	mlx5_ib_dbg(dev, "key %x mr %p\n", imr->mmkey.key, imr);
527 	return imr;
528 out_mr:
529 	mlx5_ib_err(dev, "Failed to register MKEY %d\n", err);
530 	mlx5_ib_dereg_mr(&imr->ibmr, NULL);
531 	return ERR_PTR(err);
532 }
533 
534 void mlx5_ib_free_odp_mr(struct mlx5_ib_mr *mr)
535 {
536 	struct mlx5_ib_mr *mtt;
537 	unsigned long idx;
538 
539 	/*
540 	 * If this is an implicit MR it is already invalidated so we can just
541 	 * delete the children mkeys.
542 	 */
543 	xa_for_each(&mr->implicit_children, idx, mtt) {
544 		xa_erase(&mr->implicit_children, idx);
545 		mlx5_ib_dereg_mr(&mtt->ibmr, NULL);
546 	}
547 }
548 
549 #define MLX5_PF_FLAGS_DOWNGRADE BIT(1)
550 #define MLX5_PF_FLAGS_SNAPSHOT BIT(2)
551 #define MLX5_PF_FLAGS_ENABLE BIT(3)
552 static int pagefault_real_mr(struct mlx5_ib_mr *mr, struct ib_umem_odp *odp,
553 			     u64 user_va, size_t bcnt, u32 *bytes_mapped,
554 			     u32 flags)
555 {
556 	int page_shift, ret, np;
557 	bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE;
558 	u64 access_mask;
559 	u64 start_idx;
560 	bool fault = !(flags & MLX5_PF_FLAGS_SNAPSHOT);
561 	u32 xlt_flags = MLX5_IB_UPD_XLT_ATOMIC;
562 
563 	if (flags & MLX5_PF_FLAGS_ENABLE)
564 		xlt_flags |= MLX5_IB_UPD_XLT_ENABLE;
565 
566 	page_shift = odp->page_shift;
567 	start_idx = (user_va - ib_umem_start(odp)) >> page_shift;
568 	access_mask = ODP_READ_ALLOWED_BIT;
569 
570 	if (odp->umem.writable && !downgrade)
571 		access_mask |= ODP_WRITE_ALLOWED_BIT;
572 
573 	np = ib_umem_odp_map_dma_and_lock(odp, user_va, bcnt, access_mask, fault);
574 	if (np < 0)
575 		return np;
576 
577 	/*
578 	 * No need to check whether the MTTs really belong to this MR, since
579 	 * ib_umem_odp_map_dma_and_lock already checks this.
580 	 */
581 	ret = mlx5_ib_update_xlt(mr, start_idx, np, page_shift, xlt_flags);
582 	mutex_unlock(&odp->umem_mutex);
583 
584 	if (ret < 0) {
585 		if (ret != -EAGAIN)
586 			mlx5_ib_err(mr_to_mdev(mr),
587 				    "Failed to update mkey page tables\n");
588 		goto out;
589 	}
590 
591 	if (bytes_mapped) {
592 		u32 new_mappings = (np << page_shift) -
593 			(user_va - round_down(user_va, 1 << page_shift));
594 
595 		*bytes_mapped += min_t(u32, new_mappings, bcnt);
596 	}
597 
598 	return np << (page_shift - PAGE_SHIFT);
599 
600 out:
601 	return ret;
602 }
603 
604 static int pagefault_implicit_mr(struct mlx5_ib_mr *imr,
605 				 struct ib_umem_odp *odp_imr, u64 user_va,
606 				 size_t bcnt, u32 *bytes_mapped, u32 flags)
607 {
608 	unsigned long end_idx = (user_va + bcnt - 1) >> MLX5_IMR_MTT_SHIFT;
609 	unsigned long upd_start_idx = end_idx + 1;
610 	unsigned long upd_len = 0;
611 	unsigned long npages = 0;
612 	int err;
613 	int ret;
614 
615 	if (unlikely(user_va >= mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE ||
616 		     mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE - user_va < bcnt))
617 		return -EFAULT;
618 
619 	/* Fault each child mr that intersects with our interval. */
620 	while (bcnt) {
621 		unsigned long idx = user_va >> MLX5_IMR_MTT_SHIFT;
622 		struct ib_umem_odp *umem_odp;
623 		struct mlx5_ib_mr *mtt;
624 		u64 len;
625 
626 		xa_lock(&imr->implicit_children);
627 		mtt = xa_load(&imr->implicit_children, idx);
628 		if (unlikely(!mtt)) {
629 			xa_unlock(&imr->implicit_children);
630 			mtt = implicit_get_child_mr(imr, idx);
631 			if (IS_ERR(mtt)) {
632 				ret = PTR_ERR(mtt);
633 				goto out;
634 			}
635 			upd_start_idx = min(upd_start_idx, idx);
636 			upd_len = idx - upd_start_idx + 1;
637 		} else {
638 			refcount_inc(&mtt->mmkey.usecount);
639 			xa_unlock(&imr->implicit_children);
640 		}
641 
642 		umem_odp = to_ib_umem_odp(mtt->umem);
643 		len = min_t(u64, user_va + bcnt, ib_umem_end(umem_odp)) -
644 		      user_va;
645 
646 		ret = pagefault_real_mr(mtt, umem_odp, user_va, len,
647 					bytes_mapped, flags);
648 
649 		mlx5r_deref_odp_mkey(&mtt->mmkey);
650 
651 		if (ret < 0)
652 			goto out;
653 		user_va += len;
654 		bcnt -= len;
655 		npages += ret;
656 	}
657 
658 	ret = npages;
659 
660 	/*
661 	 * Any time the implicit_children are changed we must perform an
662 	 * update of the xlt before exiting to ensure the HW and the
663 	 * implicit_children remains synchronized.
664 	 */
665 out:
666 	if (likely(!upd_len))
667 		return ret;
668 
669 	/*
670 	 * Notice this is not strictly ordered right, the KSM is updated after
671 	 * the implicit_children is updated, so a parallel page fault could
672 	 * see a MR that is not yet visible in the KSM.  This is similar to a
673 	 * parallel page fault seeing a MR that is being concurrently removed
674 	 * from the KSM. Both of these improbable situations are resolved
675 	 * safely by resuming the HW and then taking another page fault. The
676 	 * next pagefault handler will see the new information.
677 	 */
678 	mutex_lock(&odp_imr->umem_mutex);
679 	err = mlx5_ib_update_xlt(imr, upd_start_idx, upd_len, 0,
680 				 MLX5_IB_UPD_XLT_INDIRECT |
681 					 MLX5_IB_UPD_XLT_ATOMIC);
682 	mutex_unlock(&odp_imr->umem_mutex);
683 	if (err) {
684 		mlx5_ib_err(mr_to_mdev(imr), "Failed to update PAS\n");
685 		return err;
686 	}
687 	return ret;
688 }
689 
690 static int pagefault_dmabuf_mr(struct mlx5_ib_mr *mr, size_t bcnt,
691 			       u32 *bytes_mapped, u32 flags)
692 {
693 	struct ib_umem_dmabuf *umem_dmabuf = to_ib_umem_dmabuf(mr->umem);
694 	u32 xlt_flags = 0;
695 	int err;
696 	unsigned int page_size;
697 
698 	if (flags & MLX5_PF_FLAGS_ENABLE)
699 		xlt_flags |= MLX5_IB_UPD_XLT_ENABLE;
700 
701 	dma_resv_lock(umem_dmabuf->attach->dmabuf->resv, NULL);
702 	err = ib_umem_dmabuf_map_pages(umem_dmabuf);
703 	if (err) {
704 		dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
705 		return err;
706 	}
707 
708 	page_size = mlx5_umem_find_best_pgsz(&umem_dmabuf->umem, mkc,
709 					     log_page_size, 0,
710 					     umem_dmabuf->umem.iova);
711 	if (unlikely(page_size < PAGE_SIZE)) {
712 		ib_umem_dmabuf_unmap_pages(umem_dmabuf);
713 		err = -EINVAL;
714 	} else {
715 		err = mlx5_ib_update_mr_pas(mr, xlt_flags);
716 	}
717 	dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
718 
719 	if (err)
720 		return err;
721 
722 	if (bytes_mapped)
723 		*bytes_mapped += bcnt;
724 
725 	return ib_umem_num_pages(mr->umem);
726 }
727 
728 /*
729  * Returns:
730  *  -EFAULT: The io_virt->bcnt is not within the MR, it covers pages that are
731  *           not accessible, or the MR is no longer valid.
732  *  -EAGAIN/-ENOMEM: The operation should be retried
733  *
734  *  -EINVAL/others: General internal malfunction
735  *  >0: Number of pages mapped
736  */
737 static int pagefault_mr(struct mlx5_ib_mr *mr, u64 io_virt, size_t bcnt,
738 			u32 *bytes_mapped, u32 flags)
739 {
740 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
741 
742 	if (unlikely(io_virt < mr->mmkey.iova))
743 		return -EFAULT;
744 
745 	if (mr->umem->is_dmabuf)
746 		return pagefault_dmabuf_mr(mr, bcnt, bytes_mapped, flags);
747 
748 	if (!odp->is_implicit_odp) {
749 		u64 user_va;
750 
751 		if (check_add_overflow(io_virt - mr->mmkey.iova,
752 				       (u64)odp->umem.address, &user_va))
753 			return -EFAULT;
754 		if (unlikely(user_va >= ib_umem_end(odp) ||
755 			     ib_umem_end(odp) - user_va < bcnt))
756 			return -EFAULT;
757 		return pagefault_real_mr(mr, odp, user_va, bcnt, bytes_mapped,
758 					 flags);
759 	}
760 	return pagefault_implicit_mr(mr, odp, io_virt, bcnt, bytes_mapped,
761 				     flags);
762 }
763 
764 int mlx5_ib_init_odp_mr(struct mlx5_ib_mr *mr)
765 {
766 	int ret;
767 
768 	ret = pagefault_real_mr(mr, to_ib_umem_odp(mr->umem), mr->umem->address,
769 				mr->umem->length, NULL,
770 				MLX5_PF_FLAGS_SNAPSHOT | MLX5_PF_FLAGS_ENABLE);
771 	return ret >= 0 ? 0 : ret;
772 }
773 
774 int mlx5_ib_init_dmabuf_mr(struct mlx5_ib_mr *mr)
775 {
776 	int ret;
777 
778 	ret = pagefault_dmabuf_mr(mr, mr->umem->length, NULL,
779 				  MLX5_PF_FLAGS_ENABLE);
780 
781 	return ret >= 0 ? 0 : ret;
782 }
783 
784 struct pf_frame {
785 	struct pf_frame *next;
786 	u32 key;
787 	u64 io_virt;
788 	size_t bcnt;
789 	int depth;
790 };
791 
792 static bool mkey_is_eq(struct mlx5_core_mkey *mmkey, u32 key)
793 {
794 	if (!mmkey)
795 		return false;
796 	if (mmkey->type == MLX5_MKEY_MW)
797 		return mlx5_base_mkey(mmkey->key) == mlx5_base_mkey(key);
798 	return mmkey->key == key;
799 }
800 
801 static int get_indirect_num_descs(struct mlx5_core_mkey *mmkey)
802 {
803 	struct mlx5_ib_mw *mw;
804 	struct mlx5_ib_devx_mr *devx_mr;
805 
806 	if (mmkey->type == MLX5_MKEY_MW) {
807 		mw = container_of(mmkey, struct mlx5_ib_mw, mmkey);
808 		return mw->ndescs;
809 	}
810 
811 	devx_mr = container_of(mmkey, struct mlx5_ib_devx_mr,
812 			       mmkey);
813 	return devx_mr->ndescs;
814 }
815 
816 /*
817  * Handle a single data segment in a page-fault WQE or RDMA region.
818  *
819  * Returns number of OS pages retrieved on success. The caller may continue to
820  * the next data segment.
821  * Can return the following error codes:
822  * -EAGAIN to designate a temporary error. The caller will abort handling the
823  *  page fault and resolve it.
824  * -EFAULT when there's an error mapping the requested pages. The caller will
825  *  abort the page fault handling.
826  */
827 static int pagefault_single_data_segment(struct mlx5_ib_dev *dev,
828 					 struct ib_pd *pd, u32 key,
829 					 u64 io_virt, size_t bcnt,
830 					 u32 *bytes_committed,
831 					 u32 *bytes_mapped)
832 {
833 	int npages = 0, ret, i, outlen, cur_outlen = 0, depth = 0;
834 	struct pf_frame *head = NULL, *frame;
835 	struct mlx5_core_mkey *mmkey;
836 	struct mlx5_ib_mr *mr;
837 	struct mlx5_klm *pklm;
838 	u32 *out = NULL;
839 	size_t offset;
840 	int ndescs;
841 
842 	io_virt += *bytes_committed;
843 	bcnt -= *bytes_committed;
844 
845 next_mr:
846 	xa_lock(&dev->odp_mkeys);
847 	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(key));
848 	if (!mmkey) {
849 		xa_unlock(&dev->odp_mkeys);
850 		mlx5_ib_dbg(
851 			dev,
852 			"skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
853 			key);
854 		if (bytes_mapped)
855 			*bytes_mapped += bcnt;
856 		/*
857 		 * The user could specify a SGL with multiple lkeys and only
858 		 * some of them are ODP. Treat the non-ODP ones as fully
859 		 * faulted.
860 		 */
861 		ret = 0;
862 		goto end;
863 	}
864 	refcount_inc(&mmkey->usecount);
865 	xa_unlock(&dev->odp_mkeys);
866 
867 	if (!mkey_is_eq(mmkey, key)) {
868 		mlx5_ib_dbg(dev, "failed to find mkey %x\n", key);
869 		ret = -EFAULT;
870 		goto end;
871 	}
872 
873 	switch (mmkey->type) {
874 	case MLX5_MKEY_MR:
875 		mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
876 
877 		ret = pagefault_mr(mr, io_virt, bcnt, bytes_mapped, 0);
878 		if (ret < 0)
879 			goto end;
880 
881 		mlx5_update_odp_stats(mr, faults, ret);
882 
883 		npages += ret;
884 		ret = 0;
885 		break;
886 
887 	case MLX5_MKEY_MW:
888 	case MLX5_MKEY_INDIRECT_DEVX:
889 		ndescs = get_indirect_num_descs(mmkey);
890 
891 		if (depth >= MLX5_CAP_GEN(dev->mdev, max_indirection)) {
892 			mlx5_ib_dbg(dev, "indirection level exceeded\n");
893 			ret = -EFAULT;
894 			goto end;
895 		}
896 
897 		outlen = MLX5_ST_SZ_BYTES(query_mkey_out) +
898 			sizeof(*pklm) * (ndescs - 2);
899 
900 		if (outlen > cur_outlen) {
901 			kfree(out);
902 			out = kzalloc(outlen, GFP_KERNEL);
903 			if (!out) {
904 				ret = -ENOMEM;
905 				goto end;
906 			}
907 			cur_outlen = outlen;
908 		}
909 
910 		pklm = (struct mlx5_klm *)MLX5_ADDR_OF(query_mkey_out, out,
911 						       bsf0_klm0_pas_mtt0_1);
912 
913 		ret = mlx5_core_query_mkey(dev->mdev, mmkey, out, outlen);
914 		if (ret)
915 			goto end;
916 
917 		offset = io_virt - MLX5_GET64(query_mkey_out, out,
918 					      memory_key_mkey_entry.start_addr);
919 
920 		for (i = 0; bcnt && i < ndescs; i++, pklm++) {
921 			if (offset >= be32_to_cpu(pklm->bcount)) {
922 				offset -= be32_to_cpu(pklm->bcount);
923 				continue;
924 			}
925 
926 			frame = kzalloc(sizeof(*frame), GFP_KERNEL);
927 			if (!frame) {
928 				ret = -ENOMEM;
929 				goto end;
930 			}
931 
932 			frame->key = be32_to_cpu(pklm->key);
933 			frame->io_virt = be64_to_cpu(pklm->va) + offset;
934 			frame->bcnt = min_t(size_t, bcnt,
935 					    be32_to_cpu(pklm->bcount) - offset);
936 			frame->depth = depth + 1;
937 			frame->next = head;
938 			head = frame;
939 
940 			bcnt -= frame->bcnt;
941 			offset = 0;
942 		}
943 		break;
944 
945 	default:
946 		mlx5_ib_dbg(dev, "wrong mkey type %d\n", mmkey->type);
947 		ret = -EFAULT;
948 		goto end;
949 	}
950 
951 	if (head) {
952 		frame = head;
953 		head = frame->next;
954 
955 		key = frame->key;
956 		io_virt = frame->io_virt;
957 		bcnt = frame->bcnt;
958 		depth = frame->depth;
959 		kfree(frame);
960 
961 		mlx5r_deref_odp_mkey(mmkey);
962 		goto next_mr;
963 	}
964 
965 end:
966 	if (mmkey)
967 		mlx5r_deref_odp_mkey(mmkey);
968 	while (head) {
969 		frame = head;
970 		head = frame->next;
971 		kfree(frame);
972 	}
973 	kfree(out);
974 
975 	*bytes_committed = 0;
976 	return ret ? ret : npages;
977 }
978 
979 /*
980  * Parse a series of data segments for page fault handling.
981  *
982  * @dev:  Pointer to mlx5 IB device
983  * @pfault: contains page fault information.
984  * @wqe: points at the first data segment in the WQE.
985  * @wqe_end: points after the end of the WQE.
986  * @bytes_mapped: receives the number of bytes that the function was able to
987  *                map. This allows the caller to decide intelligently whether
988  *                enough memory was mapped to resolve the page fault
989  *                successfully (e.g. enough for the next MTU, or the entire
990  *                WQE).
991  * @total_wqe_bytes: receives the total data size of this WQE in bytes (minus
992  *                   the committed bytes).
993  * @receive_queue: receive WQE end of sg list
994  *
995  * Returns the number of pages loaded if positive, zero for an empty WQE, or a
996  * negative error code.
997  */
998 static int pagefault_data_segments(struct mlx5_ib_dev *dev,
999 				   struct mlx5_pagefault *pfault,
1000 				   void *wqe,
1001 				   void *wqe_end, u32 *bytes_mapped,
1002 				   u32 *total_wqe_bytes, bool receive_queue)
1003 {
1004 	int ret = 0, npages = 0;
1005 	u64 io_virt;
1006 	u32 key;
1007 	u32 byte_count;
1008 	size_t bcnt;
1009 	int inline_segment;
1010 
1011 	if (bytes_mapped)
1012 		*bytes_mapped = 0;
1013 	if (total_wqe_bytes)
1014 		*total_wqe_bytes = 0;
1015 
1016 	while (wqe < wqe_end) {
1017 		struct mlx5_wqe_data_seg *dseg = wqe;
1018 
1019 		io_virt = be64_to_cpu(dseg->addr);
1020 		key = be32_to_cpu(dseg->lkey);
1021 		byte_count = be32_to_cpu(dseg->byte_count);
1022 		inline_segment = !!(byte_count &  MLX5_INLINE_SEG);
1023 		bcnt	       = byte_count & ~MLX5_INLINE_SEG;
1024 
1025 		if (inline_segment) {
1026 			bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK;
1027 			wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt,
1028 				     16);
1029 		} else {
1030 			wqe += sizeof(*dseg);
1031 		}
1032 
1033 		/* receive WQE end of sg list. */
1034 		if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY &&
1035 		    io_virt == 0)
1036 			break;
1037 
1038 		if (!inline_segment && total_wqe_bytes) {
1039 			*total_wqe_bytes += bcnt - min_t(size_t, bcnt,
1040 					pfault->bytes_committed);
1041 		}
1042 
1043 		/* A zero length data segment designates a length of 2GB. */
1044 		if (bcnt == 0)
1045 			bcnt = 1U << 31;
1046 
1047 		if (inline_segment || bcnt <= pfault->bytes_committed) {
1048 			pfault->bytes_committed -=
1049 				min_t(size_t, bcnt,
1050 				      pfault->bytes_committed);
1051 			continue;
1052 		}
1053 
1054 		ret = pagefault_single_data_segment(dev, NULL, key,
1055 						    io_virt, bcnt,
1056 						    &pfault->bytes_committed,
1057 						    bytes_mapped);
1058 		if (ret < 0)
1059 			break;
1060 		npages += ret;
1061 	}
1062 
1063 	return ret < 0 ? ret : npages;
1064 }
1065 
1066 /*
1067  * Parse initiator WQE. Advances the wqe pointer to point at the
1068  * scatter-gather list, and set wqe_end to the end of the WQE.
1069  */
1070 static int mlx5_ib_mr_initiator_pfault_handler(
1071 	struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault,
1072 	struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length)
1073 {
1074 	struct mlx5_wqe_ctrl_seg *ctrl = *wqe;
1075 	u16 wqe_index = pfault->wqe.wqe_index;
1076 	struct mlx5_base_av *av;
1077 	unsigned ds, opcode;
1078 	u32 qpn = qp->trans_qp.base.mqp.qpn;
1079 
1080 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
1081 	if (ds * MLX5_WQE_DS_UNITS > wqe_length) {
1082 		mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n",
1083 			    ds, wqe_length);
1084 		return -EFAULT;
1085 	}
1086 
1087 	if (ds == 0) {
1088 		mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n",
1089 			    wqe_index, qpn);
1090 		return -EFAULT;
1091 	}
1092 
1093 	*wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS;
1094 	*wqe += sizeof(*ctrl);
1095 
1096 	opcode = be32_to_cpu(ctrl->opmod_idx_opcode) &
1097 		 MLX5_WQE_CTRL_OPCODE_MASK;
1098 
1099 	if (qp->ibqp.qp_type == IB_QPT_XRC_INI)
1100 		*wqe += sizeof(struct mlx5_wqe_xrc_seg);
1101 
1102 	if (qp->type == IB_QPT_UD || qp->type == MLX5_IB_QPT_DCI) {
1103 		av = *wqe;
1104 		if (av->dqp_dct & cpu_to_be32(MLX5_EXTENDED_UD_AV))
1105 			*wqe += sizeof(struct mlx5_av);
1106 		else
1107 			*wqe += sizeof(struct mlx5_base_av);
1108 	}
1109 
1110 	switch (opcode) {
1111 	case MLX5_OPCODE_RDMA_WRITE:
1112 	case MLX5_OPCODE_RDMA_WRITE_IMM:
1113 	case MLX5_OPCODE_RDMA_READ:
1114 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1115 		break;
1116 	case MLX5_OPCODE_ATOMIC_CS:
1117 	case MLX5_OPCODE_ATOMIC_FA:
1118 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1119 		*wqe += sizeof(struct mlx5_wqe_atomic_seg);
1120 		break;
1121 	}
1122 
1123 	return 0;
1124 }
1125 
1126 /*
1127  * Parse responder WQE and set wqe_end to the end of the WQE.
1128  */
1129 static int mlx5_ib_mr_responder_pfault_handler_srq(struct mlx5_ib_dev *dev,
1130 						   struct mlx5_ib_srq *srq,
1131 						   void **wqe, void **wqe_end,
1132 						   int wqe_length)
1133 {
1134 	int wqe_size = 1 << srq->msrq.wqe_shift;
1135 
1136 	if (wqe_size > wqe_length) {
1137 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1138 		return -EFAULT;
1139 	}
1140 
1141 	*wqe_end = *wqe + wqe_size;
1142 	*wqe += sizeof(struct mlx5_wqe_srq_next_seg);
1143 
1144 	return 0;
1145 }
1146 
1147 static int mlx5_ib_mr_responder_pfault_handler_rq(struct mlx5_ib_dev *dev,
1148 						  struct mlx5_ib_qp *qp,
1149 						  void *wqe, void **wqe_end,
1150 						  int wqe_length)
1151 {
1152 	struct mlx5_ib_wq *wq = &qp->rq;
1153 	int wqe_size = 1 << wq->wqe_shift;
1154 
1155 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE) {
1156 		mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n");
1157 		return -EFAULT;
1158 	}
1159 
1160 	if (wqe_size > wqe_length) {
1161 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1162 		return -EFAULT;
1163 	}
1164 
1165 	*wqe_end = wqe + wqe_size;
1166 
1167 	return 0;
1168 }
1169 
1170 static inline struct mlx5_core_rsc_common *odp_get_rsc(struct mlx5_ib_dev *dev,
1171 						       u32 wq_num, int pf_type)
1172 {
1173 	struct mlx5_core_rsc_common *common = NULL;
1174 	struct mlx5_core_srq *srq;
1175 
1176 	switch (pf_type) {
1177 	case MLX5_WQE_PF_TYPE_RMP:
1178 		srq = mlx5_cmd_get_srq(dev, wq_num);
1179 		if (srq)
1180 			common = &srq->common;
1181 		break;
1182 	case MLX5_WQE_PF_TYPE_REQ_SEND_OR_WRITE:
1183 	case MLX5_WQE_PF_TYPE_RESP:
1184 	case MLX5_WQE_PF_TYPE_REQ_READ_OR_ATOMIC:
1185 		common = mlx5_core_res_hold(dev, wq_num, MLX5_RES_QP);
1186 		break;
1187 	default:
1188 		break;
1189 	}
1190 
1191 	return common;
1192 }
1193 
1194 static inline struct mlx5_ib_qp *res_to_qp(struct mlx5_core_rsc_common *res)
1195 {
1196 	struct mlx5_core_qp *mqp = (struct mlx5_core_qp *)res;
1197 
1198 	return to_mibqp(mqp);
1199 }
1200 
1201 static inline struct mlx5_ib_srq *res_to_srq(struct mlx5_core_rsc_common *res)
1202 {
1203 	struct mlx5_core_srq *msrq =
1204 		container_of(res, struct mlx5_core_srq, common);
1205 
1206 	return to_mibsrq(msrq);
1207 }
1208 
1209 static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
1210 					  struct mlx5_pagefault *pfault)
1211 {
1212 	bool sq = pfault->type & MLX5_PFAULT_REQUESTOR;
1213 	u16 wqe_index = pfault->wqe.wqe_index;
1214 	void *wqe, *wqe_start = NULL, *wqe_end = NULL;
1215 	u32 bytes_mapped, total_wqe_bytes;
1216 	struct mlx5_core_rsc_common *res;
1217 	int resume_with_error = 1;
1218 	struct mlx5_ib_qp *qp;
1219 	size_t bytes_copied;
1220 	int ret = 0;
1221 
1222 	res = odp_get_rsc(dev, pfault->wqe.wq_num, pfault->type);
1223 	if (!res) {
1224 		mlx5_ib_dbg(dev, "wqe page fault for missing resource %d\n", pfault->wqe.wq_num);
1225 		return;
1226 	}
1227 
1228 	if (res->res != MLX5_RES_QP && res->res != MLX5_RES_SRQ &&
1229 	    res->res != MLX5_RES_XSRQ) {
1230 		mlx5_ib_err(dev, "wqe page fault for unsupported type %d\n",
1231 			    pfault->type);
1232 		goto resolve_page_fault;
1233 	}
1234 
1235 	wqe_start = (void *)__get_free_page(GFP_KERNEL);
1236 	if (!wqe_start) {
1237 		mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n");
1238 		goto resolve_page_fault;
1239 	}
1240 
1241 	wqe = wqe_start;
1242 	qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL;
1243 	if (qp && sq) {
1244 		ret = mlx5_ib_read_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE,
1245 					  &bytes_copied);
1246 		if (ret)
1247 			goto read_user;
1248 		ret = mlx5_ib_mr_initiator_pfault_handler(
1249 			dev, pfault, qp, &wqe, &wqe_end, bytes_copied);
1250 	} else if (qp && !sq) {
1251 		ret = mlx5_ib_read_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE,
1252 					  &bytes_copied);
1253 		if (ret)
1254 			goto read_user;
1255 		ret = mlx5_ib_mr_responder_pfault_handler_rq(
1256 			dev, qp, wqe, &wqe_end, bytes_copied);
1257 	} else if (!qp) {
1258 		struct mlx5_ib_srq *srq = res_to_srq(res);
1259 
1260 		ret = mlx5_ib_read_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE,
1261 					   &bytes_copied);
1262 		if (ret)
1263 			goto read_user;
1264 		ret = mlx5_ib_mr_responder_pfault_handler_srq(
1265 			dev, srq, &wqe, &wqe_end, bytes_copied);
1266 	}
1267 
1268 	if (ret < 0 || wqe >= wqe_end)
1269 		goto resolve_page_fault;
1270 
1271 	ret = pagefault_data_segments(dev, pfault, wqe, wqe_end, &bytes_mapped,
1272 				      &total_wqe_bytes, !sq);
1273 	if (ret == -EAGAIN)
1274 		goto out;
1275 
1276 	if (ret < 0 || total_wqe_bytes > bytes_mapped)
1277 		goto resolve_page_fault;
1278 
1279 out:
1280 	ret = 0;
1281 	resume_with_error = 0;
1282 
1283 read_user:
1284 	if (ret)
1285 		mlx5_ib_err(
1286 			dev,
1287 			"Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n",
1288 			ret, wqe_index, pfault->token);
1289 
1290 resolve_page_fault:
1291 	mlx5_ib_page_fault_resume(dev, pfault, resume_with_error);
1292 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n",
1293 		    pfault->wqe.wq_num, resume_with_error,
1294 		    pfault->type);
1295 	mlx5_core_res_put(res);
1296 	free_page((unsigned long)wqe_start);
1297 }
1298 
1299 static int pages_in_range(u64 address, u32 length)
1300 {
1301 	return (ALIGN(address + length, PAGE_SIZE) -
1302 		(address & PAGE_MASK)) >> PAGE_SHIFT;
1303 }
1304 
1305 static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev,
1306 					   struct mlx5_pagefault *pfault)
1307 {
1308 	u64 address;
1309 	u32 length;
1310 	u32 prefetch_len = pfault->bytes_committed;
1311 	int prefetch_activated = 0;
1312 	u32 rkey = pfault->rdma.r_key;
1313 	int ret;
1314 
1315 	/* The RDMA responder handler handles the page fault in two parts.
1316 	 * First it brings the necessary pages for the current packet
1317 	 * (and uses the pfault context), and then (after resuming the QP)
1318 	 * prefetches more pages. The second operation cannot use the pfault
1319 	 * context and therefore uses the dummy_pfault context allocated on
1320 	 * the stack */
1321 	pfault->rdma.rdma_va += pfault->bytes_committed;
1322 	pfault->rdma.rdma_op_len -= min(pfault->bytes_committed,
1323 					 pfault->rdma.rdma_op_len);
1324 	pfault->bytes_committed = 0;
1325 
1326 	address = pfault->rdma.rdma_va;
1327 	length  = pfault->rdma.rdma_op_len;
1328 
1329 	/* For some operations, the hardware cannot tell the exact message
1330 	 * length, and in those cases it reports zero. Use prefetch
1331 	 * logic. */
1332 	if (length == 0) {
1333 		prefetch_activated = 1;
1334 		length = pfault->rdma.packet_size;
1335 		prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len);
1336 	}
1337 
1338 	ret = pagefault_single_data_segment(dev, NULL, rkey, address, length,
1339 					    &pfault->bytes_committed, NULL);
1340 	if (ret == -EAGAIN) {
1341 		/* We're racing with an invalidation, don't prefetch */
1342 		prefetch_activated = 0;
1343 	} else if (ret < 0 || pages_in_range(address, length) > ret) {
1344 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1345 		if (ret != -ENOENT)
1346 			mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n",
1347 				    ret, pfault->token, pfault->type);
1348 		return;
1349 	}
1350 
1351 	mlx5_ib_page_fault_resume(dev, pfault, 0);
1352 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n",
1353 		    pfault->token, pfault->type,
1354 		    prefetch_activated);
1355 
1356 	/* At this point, there might be a new pagefault already arriving in
1357 	 * the eq, switch to the dummy pagefault for the rest of the
1358 	 * processing. We're still OK with the objects being alive as the
1359 	 * work-queue is being fenced. */
1360 
1361 	if (prefetch_activated) {
1362 		u32 bytes_committed = 0;
1363 
1364 		ret = pagefault_single_data_segment(dev, NULL, rkey, address,
1365 						    prefetch_len,
1366 						    &bytes_committed, NULL);
1367 		if (ret < 0 && ret != -EAGAIN) {
1368 			mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n",
1369 				    ret, pfault->token, address, prefetch_len);
1370 		}
1371 	}
1372 }
1373 
1374 static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault)
1375 {
1376 	u8 event_subtype = pfault->event_subtype;
1377 
1378 	switch (event_subtype) {
1379 	case MLX5_PFAULT_SUBTYPE_WQE:
1380 		mlx5_ib_mr_wqe_pfault_handler(dev, pfault);
1381 		break;
1382 	case MLX5_PFAULT_SUBTYPE_RDMA:
1383 		mlx5_ib_mr_rdma_pfault_handler(dev, pfault);
1384 		break;
1385 	default:
1386 		mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n",
1387 			    event_subtype);
1388 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1389 	}
1390 }
1391 
1392 static void mlx5_ib_eqe_pf_action(struct work_struct *work)
1393 {
1394 	struct mlx5_pagefault *pfault = container_of(work,
1395 						     struct mlx5_pagefault,
1396 						     work);
1397 	struct mlx5_ib_pf_eq *eq = pfault->eq;
1398 
1399 	mlx5_ib_pfault(eq->dev, pfault);
1400 	mempool_free(pfault, eq->pool);
1401 }
1402 
1403 static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq)
1404 {
1405 	struct mlx5_eqe_page_fault *pf_eqe;
1406 	struct mlx5_pagefault *pfault;
1407 	struct mlx5_eqe *eqe;
1408 	int cc = 0;
1409 
1410 	while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) {
1411 		pfault = mempool_alloc(eq->pool, GFP_ATOMIC);
1412 		if (!pfault) {
1413 			schedule_work(&eq->work);
1414 			break;
1415 		}
1416 
1417 		pf_eqe = &eqe->data.page_fault;
1418 		pfault->event_subtype = eqe->sub_type;
1419 		pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);
1420 
1421 		mlx5_ib_dbg(eq->dev,
1422 			    "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
1423 			    eqe->sub_type, pfault->bytes_committed);
1424 
1425 		switch (eqe->sub_type) {
1426 		case MLX5_PFAULT_SUBTYPE_RDMA:
1427 			/* RDMA based event */
1428 			pfault->type =
1429 				be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
1430 			pfault->token =
1431 				be32_to_cpu(pf_eqe->rdma.pftype_token) &
1432 				MLX5_24BIT_MASK;
1433 			pfault->rdma.r_key =
1434 				be32_to_cpu(pf_eqe->rdma.r_key);
1435 			pfault->rdma.packet_size =
1436 				be16_to_cpu(pf_eqe->rdma.packet_length);
1437 			pfault->rdma.rdma_op_len =
1438 				be32_to_cpu(pf_eqe->rdma.rdma_op_len);
1439 			pfault->rdma.rdma_va =
1440 				be64_to_cpu(pf_eqe->rdma.rdma_va);
1441 			mlx5_ib_dbg(eq->dev,
1442 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
1443 				    pfault->type, pfault->token,
1444 				    pfault->rdma.r_key);
1445 			mlx5_ib_dbg(eq->dev,
1446 				    "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
1447 				    pfault->rdma.rdma_op_len,
1448 				    pfault->rdma.rdma_va);
1449 			break;
1450 
1451 		case MLX5_PFAULT_SUBTYPE_WQE:
1452 			/* WQE based event */
1453 			pfault->type =
1454 				(be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7;
1455 			pfault->token =
1456 				be32_to_cpu(pf_eqe->wqe.token);
1457 			pfault->wqe.wq_num =
1458 				be32_to_cpu(pf_eqe->wqe.pftype_wq) &
1459 				MLX5_24BIT_MASK;
1460 			pfault->wqe.wqe_index =
1461 				be16_to_cpu(pf_eqe->wqe.wqe_index);
1462 			pfault->wqe.packet_size =
1463 				be16_to_cpu(pf_eqe->wqe.packet_length);
1464 			mlx5_ib_dbg(eq->dev,
1465 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
1466 				    pfault->type, pfault->token,
1467 				    pfault->wqe.wq_num,
1468 				    pfault->wqe.wqe_index);
1469 			break;
1470 
1471 		default:
1472 			mlx5_ib_warn(eq->dev,
1473 				     "Unsupported page fault event sub-type: 0x%02hhx\n",
1474 				     eqe->sub_type);
1475 			/* Unsupported page faults should still be
1476 			 * resolved by the page fault handler
1477 			 */
1478 		}
1479 
1480 		pfault->eq = eq;
1481 		INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action);
1482 		queue_work(eq->wq, &pfault->work);
1483 
1484 		cc = mlx5_eq_update_cc(eq->core, ++cc);
1485 	}
1486 
1487 	mlx5_eq_update_ci(eq->core, cc, 1);
1488 }
1489 
1490 static int mlx5_ib_eq_pf_int(struct notifier_block *nb, unsigned long type,
1491 			     void *data)
1492 {
1493 	struct mlx5_ib_pf_eq *eq =
1494 		container_of(nb, struct mlx5_ib_pf_eq, irq_nb);
1495 	unsigned long flags;
1496 
1497 	if (spin_trylock_irqsave(&eq->lock, flags)) {
1498 		mlx5_ib_eq_pf_process(eq);
1499 		spin_unlock_irqrestore(&eq->lock, flags);
1500 	} else {
1501 		schedule_work(&eq->work);
1502 	}
1503 
1504 	return IRQ_HANDLED;
1505 }
1506 
1507 /* mempool_refill() was proposed but unfortunately wasn't accepted
1508  * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
1509  * Cheap workaround.
1510  */
1511 static void mempool_refill(mempool_t *pool)
1512 {
1513 	while (pool->curr_nr < pool->min_nr)
1514 		mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
1515 }
1516 
1517 static void mlx5_ib_eq_pf_action(struct work_struct *work)
1518 {
1519 	struct mlx5_ib_pf_eq *eq =
1520 		container_of(work, struct mlx5_ib_pf_eq, work);
1521 
1522 	mempool_refill(eq->pool);
1523 
1524 	spin_lock_irq(&eq->lock);
1525 	mlx5_ib_eq_pf_process(eq);
1526 	spin_unlock_irq(&eq->lock);
1527 }
1528 
1529 enum {
1530 	MLX5_IB_NUM_PF_EQE	= 0x1000,
1531 	MLX5_IB_NUM_PF_DRAIN	= 64,
1532 };
1533 
1534 int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1535 {
1536 	struct mlx5_eq_param param = {};
1537 	int err = 0;
1538 
1539 	mutex_lock(&dev->odp_eq_mutex);
1540 	if (eq->core)
1541 		goto unlock;
1542 	INIT_WORK(&eq->work, mlx5_ib_eq_pf_action);
1543 	spin_lock_init(&eq->lock);
1544 	eq->dev = dev;
1545 
1546 	eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN,
1547 					       sizeof(struct mlx5_pagefault));
1548 	if (!eq->pool) {
1549 		err = -ENOMEM;
1550 		goto unlock;
1551 	}
1552 
1553 	eq->wq = alloc_workqueue("mlx5_ib_page_fault",
1554 				 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM,
1555 				 MLX5_NUM_CMD_EQE);
1556 	if (!eq->wq) {
1557 		err = -ENOMEM;
1558 		goto err_mempool;
1559 	}
1560 
1561 	eq->irq_nb.notifier_call = mlx5_ib_eq_pf_int;
1562 	param = (struct mlx5_eq_param){
1563 		.irq_index = 0,
1564 		.nent = MLX5_IB_NUM_PF_EQE,
1565 	};
1566 	param.mask[0] = 1ull << MLX5_EVENT_TYPE_PAGE_FAULT;
1567 	eq->core = mlx5_eq_create_generic(dev->mdev, &param);
1568 	if (IS_ERR(eq->core)) {
1569 		err = PTR_ERR(eq->core);
1570 		goto err_wq;
1571 	}
1572 	err = mlx5_eq_enable(dev->mdev, eq->core, &eq->irq_nb);
1573 	if (err) {
1574 		mlx5_ib_err(dev, "failed to enable odp EQ %d\n", err);
1575 		goto err_eq;
1576 	}
1577 
1578 	mutex_unlock(&dev->odp_eq_mutex);
1579 	return 0;
1580 err_eq:
1581 	mlx5_eq_destroy_generic(dev->mdev, eq->core);
1582 err_wq:
1583 	eq->core = NULL;
1584 	destroy_workqueue(eq->wq);
1585 err_mempool:
1586 	mempool_destroy(eq->pool);
1587 unlock:
1588 	mutex_unlock(&dev->odp_eq_mutex);
1589 	return err;
1590 }
1591 
1592 static int
1593 mlx5_ib_odp_destroy_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1594 {
1595 	int err;
1596 
1597 	if (!eq->core)
1598 		return 0;
1599 	mlx5_eq_disable(dev->mdev, eq->core, &eq->irq_nb);
1600 	err = mlx5_eq_destroy_generic(dev->mdev, eq->core);
1601 	cancel_work_sync(&eq->work);
1602 	destroy_workqueue(eq->wq);
1603 	mempool_destroy(eq->pool);
1604 
1605 	return err;
1606 }
1607 
1608 void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
1609 {
1610 	if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1611 		return;
1612 
1613 	switch (ent->order - 2) {
1614 	case MLX5_IMR_MTT_CACHE_ENTRY:
1615 		ent->page = PAGE_SHIFT;
1616 		ent->xlt = MLX5_IMR_MTT_ENTRIES *
1617 			   sizeof(struct mlx5_mtt) /
1618 			   MLX5_IB_UMR_OCTOWORD;
1619 		ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
1620 		ent->limit = 0;
1621 		break;
1622 
1623 	case MLX5_IMR_KSM_CACHE_ENTRY:
1624 		ent->page = MLX5_KSM_PAGE_SHIFT;
1625 		ent->xlt = mlx5_imr_ksm_entries *
1626 			   sizeof(struct mlx5_klm) /
1627 			   MLX5_IB_UMR_OCTOWORD;
1628 		ent->access_mode = MLX5_MKC_ACCESS_MODE_KSM;
1629 		ent->limit = 0;
1630 		break;
1631 	}
1632 }
1633 
1634 static const struct ib_device_ops mlx5_ib_dev_odp_ops = {
1635 	.advise_mr = mlx5_ib_advise_mr,
1636 };
1637 
1638 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
1639 {
1640 	int ret = 0;
1641 
1642 	internal_fill_odp_caps(dev);
1643 
1644 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1645 		return ret;
1646 
1647 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_odp_ops);
1648 
1649 	if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) {
1650 		ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey);
1651 		if (ret) {
1652 			mlx5_ib_err(dev, "Error getting null_mkey %d\n", ret);
1653 			return ret;
1654 		}
1655 	}
1656 
1657 	mutex_init(&dev->odp_eq_mutex);
1658 	return ret;
1659 }
1660 
1661 void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev)
1662 {
1663 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1664 		return;
1665 
1666 	mlx5_ib_odp_destroy_eq(dev, &dev->odp_pf_eq);
1667 }
1668 
1669 int mlx5_ib_odp_init(void)
1670 {
1671 	mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) -
1672 				       MLX5_IMR_MTT_BITS);
1673 
1674 	return 0;
1675 }
1676 
1677 struct prefetch_mr_work {
1678 	struct work_struct work;
1679 	u32 pf_flags;
1680 	u32 num_sge;
1681 	struct {
1682 		u64 io_virt;
1683 		struct mlx5_ib_mr *mr;
1684 		size_t length;
1685 	} frags[];
1686 };
1687 
1688 static void destroy_prefetch_work(struct prefetch_mr_work *work)
1689 {
1690 	u32 i;
1691 
1692 	for (i = 0; i < work->num_sge; ++i)
1693 		mlx5r_deref_odp_mkey(&work->frags[i].mr->mmkey);
1694 
1695 	kvfree(work);
1696 }
1697 
1698 static struct mlx5_ib_mr *
1699 get_prefetchable_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
1700 		    u32 lkey)
1701 {
1702 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1703 	struct mlx5_core_mkey *mmkey;
1704 	struct mlx5_ib_mr *mr = NULL;
1705 
1706 	xa_lock(&dev->odp_mkeys);
1707 	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(lkey));
1708 	if (!mmkey || mmkey->key != lkey || mmkey->type != MLX5_MKEY_MR)
1709 		goto end;
1710 
1711 	mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
1712 
1713 	if (mr->ibmr.pd != pd) {
1714 		mr = NULL;
1715 		goto end;
1716 	}
1717 
1718 	/* prefetch with write-access must be supported by the MR */
1719 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
1720 	    !mr->umem->writable) {
1721 		mr = NULL;
1722 		goto end;
1723 	}
1724 
1725 	refcount_inc(&mmkey->usecount);
1726 end:
1727 	xa_unlock(&dev->odp_mkeys);
1728 	return mr;
1729 }
1730 
1731 static void mlx5_ib_prefetch_mr_work(struct work_struct *w)
1732 {
1733 	struct prefetch_mr_work *work =
1734 		container_of(w, struct prefetch_mr_work, work);
1735 	u32 bytes_mapped = 0;
1736 	int ret;
1737 	u32 i;
1738 
1739 	/* We rely on IB/core that work is executed if we have num_sge != 0 only. */
1740 	WARN_ON(!work->num_sge);
1741 	for (i = 0; i < work->num_sge; ++i) {
1742 		ret = pagefault_mr(work->frags[i].mr, work->frags[i].io_virt,
1743 				   work->frags[i].length, &bytes_mapped,
1744 				   work->pf_flags);
1745 		if (ret <= 0)
1746 			continue;
1747 		mlx5_update_odp_stats(work->frags[i].mr, prefetch, ret);
1748 	}
1749 
1750 	destroy_prefetch_work(work);
1751 }
1752 
1753 static bool init_prefetch_work(struct ib_pd *pd,
1754 			       enum ib_uverbs_advise_mr_advice advice,
1755 			       u32 pf_flags, struct prefetch_mr_work *work,
1756 			       struct ib_sge *sg_list, u32 num_sge)
1757 {
1758 	u32 i;
1759 
1760 	INIT_WORK(&work->work, mlx5_ib_prefetch_mr_work);
1761 	work->pf_flags = pf_flags;
1762 
1763 	for (i = 0; i < num_sge; ++i) {
1764 		work->frags[i].io_virt = sg_list[i].addr;
1765 		work->frags[i].length = sg_list[i].length;
1766 		work->frags[i].mr =
1767 			get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1768 		if (!work->frags[i].mr) {
1769 			work->num_sge = i;
1770 			return false;
1771 		}
1772 	}
1773 	work->num_sge = num_sge;
1774 	return true;
1775 }
1776 
1777 static int mlx5_ib_prefetch_sg_list(struct ib_pd *pd,
1778 				    enum ib_uverbs_advise_mr_advice advice,
1779 				    u32 pf_flags, struct ib_sge *sg_list,
1780 				    u32 num_sge)
1781 {
1782 	u32 bytes_mapped = 0;
1783 	int ret = 0;
1784 	u32 i;
1785 
1786 	for (i = 0; i < num_sge; ++i) {
1787 		struct mlx5_ib_mr *mr;
1788 
1789 		mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1790 		if (!mr)
1791 			return -ENOENT;
1792 		ret = pagefault_mr(mr, sg_list[i].addr, sg_list[i].length,
1793 				   &bytes_mapped, pf_flags);
1794 		if (ret < 0) {
1795 			mlx5r_deref_odp_mkey(&mr->mmkey);
1796 			return ret;
1797 		}
1798 		mlx5_update_odp_stats(mr, prefetch, ret);
1799 		mlx5r_deref_odp_mkey(&mr->mmkey);
1800 	}
1801 
1802 	return 0;
1803 }
1804 
1805 int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
1806 			       enum ib_uverbs_advise_mr_advice advice,
1807 			       u32 flags, struct ib_sge *sg_list, u32 num_sge)
1808 {
1809 	u32 pf_flags = 0;
1810 	struct prefetch_mr_work *work;
1811 
1812 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH)
1813 		pf_flags |= MLX5_PF_FLAGS_DOWNGRADE;
1814 
1815 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT)
1816 		pf_flags |= MLX5_PF_FLAGS_SNAPSHOT;
1817 
1818 	if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH)
1819 		return mlx5_ib_prefetch_sg_list(pd, advice, pf_flags, sg_list,
1820 						num_sge);
1821 
1822 	work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL);
1823 	if (!work)
1824 		return -ENOMEM;
1825 
1826 	if (!init_prefetch_work(pd, advice, pf_flags, work, sg_list, num_sge)) {
1827 		destroy_prefetch_work(work);
1828 		return -EINVAL;
1829 	}
1830 	queue_work(system_unbound_wq, &work->work);
1831 	return 0;
1832 }
1833