xref: /openbmc/linux/drivers/infiniband/hw/mlx5/odp.c (revision 07d9a767)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <rdma/ib_umem.h>
34 #include <rdma/ib_umem_odp.h>
35 #include <linux/kernel.h>
36 
37 #include "mlx5_ib.h"
38 #include "cmd.h"
39 #include "qp.h"
40 
41 #include <linux/mlx5/eq.h>
42 
43 /* Contains the details of a pagefault. */
44 struct mlx5_pagefault {
45 	u32			bytes_committed;
46 	u32			token;
47 	u8			event_subtype;
48 	u8			type;
49 	union {
50 		/* Initiator or send message responder pagefault details. */
51 		struct {
52 			/* Received packet size, only valid for responders. */
53 			u32	packet_size;
54 			/*
55 			 * Number of resource holding WQE, depends on type.
56 			 */
57 			u32	wq_num;
58 			/*
59 			 * WQE index. Refers to either the send queue or
60 			 * receive queue, according to event_subtype.
61 			 */
62 			u16	wqe_index;
63 		} wqe;
64 		/* RDMA responder pagefault details */
65 		struct {
66 			u32	r_key;
67 			/*
68 			 * Received packet size, minimal size page fault
69 			 * resolution required for forward progress.
70 			 */
71 			u32	packet_size;
72 			u32	rdma_op_len;
73 			u64	rdma_va;
74 		} rdma;
75 	};
76 
77 	struct mlx5_ib_pf_eq	*eq;
78 	struct work_struct	work;
79 };
80 
81 #define MAX_PREFETCH_LEN (4*1024*1024U)
82 
83 /* Timeout in ms to wait for an active mmu notifier to complete when handling
84  * a pagefault. */
85 #define MMU_NOTIFIER_TIMEOUT 1000
86 
87 #define MLX5_IMR_MTT_BITS (30 - PAGE_SHIFT)
88 #define MLX5_IMR_MTT_SHIFT (MLX5_IMR_MTT_BITS + PAGE_SHIFT)
89 #define MLX5_IMR_MTT_ENTRIES BIT_ULL(MLX5_IMR_MTT_BITS)
90 #define MLX5_IMR_MTT_SIZE BIT_ULL(MLX5_IMR_MTT_SHIFT)
91 #define MLX5_IMR_MTT_MASK (~(MLX5_IMR_MTT_SIZE - 1))
92 
93 #define MLX5_KSM_PAGE_SHIFT MLX5_IMR_MTT_SHIFT
94 
95 static u64 mlx5_imr_ksm_entries;
96 
97 static void populate_klm(struct mlx5_klm *pklm, size_t idx, size_t nentries,
98 			struct mlx5_ib_mr *imr, int flags)
99 {
100 	struct mlx5_klm *end = pklm + nentries;
101 
102 	if (flags & MLX5_IB_UPD_XLT_ZAP) {
103 		for (; pklm != end; pklm++, idx++) {
104 			pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
105 			pklm->key = cpu_to_be32(imr->dev->null_mkey);
106 			pklm->va = 0;
107 		}
108 		return;
109 	}
110 
111 	/*
112 	 * The locking here is pretty subtle. Ideally the implicit_children
113 	 * xarray would be protected by the umem_mutex, however that is not
114 	 * possible. Instead this uses a weaker update-then-lock pattern:
115 	 *
116 	 *  srcu_read_lock()
117 	 *    xa_store()
118 	 *    mutex_lock(umem_mutex)
119 	 *     mlx5_ib_update_xlt()
120 	 *    mutex_unlock(umem_mutex)
121 	 *    destroy lkey
122 	 *
123 	 * ie any change the xarray must be followed by the locked update_xlt
124 	 * before destroying.
125 	 *
126 	 * The umem_mutex provides the acquire/release semantic needed to make
127 	 * the xa_store() visible to a racing thread. While SRCU is not
128 	 * technically required, using it gives consistent use of the SRCU
129 	 * locking around the xarray.
130 	 */
131 	lockdep_assert_held(&to_ib_umem_odp(imr->umem)->umem_mutex);
132 	lockdep_assert_held(&imr->dev->odp_srcu);
133 
134 	for (; pklm != end; pklm++, idx++) {
135 		struct mlx5_ib_mr *mtt = xa_load(&imr->implicit_children, idx);
136 
137 		pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
138 		if (mtt) {
139 			pklm->key = cpu_to_be32(mtt->ibmr.lkey);
140 			pklm->va = cpu_to_be64(idx * MLX5_IMR_MTT_SIZE);
141 		} else {
142 			pklm->key = cpu_to_be32(imr->dev->null_mkey);
143 			pklm->va = 0;
144 		}
145 	}
146 }
147 
148 static u64 umem_dma_to_mtt(dma_addr_t umem_dma)
149 {
150 	u64 mtt_entry = umem_dma & ODP_DMA_ADDR_MASK;
151 
152 	if (umem_dma & ODP_READ_ALLOWED_BIT)
153 		mtt_entry |= MLX5_IB_MTT_READ;
154 	if (umem_dma & ODP_WRITE_ALLOWED_BIT)
155 		mtt_entry |= MLX5_IB_MTT_WRITE;
156 
157 	return mtt_entry;
158 }
159 
160 static void populate_mtt(__be64 *pas, size_t idx, size_t nentries,
161 			 struct mlx5_ib_mr *mr, int flags)
162 {
163 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
164 	dma_addr_t pa;
165 	size_t i;
166 
167 	if (flags & MLX5_IB_UPD_XLT_ZAP)
168 		return;
169 
170 	for (i = 0; i < nentries; i++) {
171 		pa = odp->dma_list[idx + i];
172 		pas[i] = cpu_to_be64(umem_dma_to_mtt(pa));
173 	}
174 }
175 
176 void mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
177 			   struct mlx5_ib_mr *mr, int flags)
178 {
179 	if (flags & MLX5_IB_UPD_XLT_INDIRECT) {
180 		populate_klm(xlt, idx, nentries, mr, flags);
181 	} else {
182 		populate_mtt(xlt, idx, nentries, mr, flags);
183 	}
184 }
185 
186 static void dma_fence_odp_mr(struct mlx5_ib_mr *mr)
187 {
188 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
189 
190 	/* Ensure mlx5_ib_invalidate_range() will not touch the MR any more */
191 	mutex_lock(&odp->umem_mutex);
192 	if (odp->npages) {
193 		mlx5_mr_cache_invalidate(mr);
194 		ib_umem_odp_unmap_dma_pages(odp, ib_umem_start(odp),
195 					    ib_umem_end(odp));
196 		WARN_ON(odp->npages);
197 	}
198 	odp->private = NULL;
199 	mutex_unlock(&odp->umem_mutex);
200 
201 	if (!mr->cache_ent) {
202 		mlx5_core_destroy_mkey(mr->dev->mdev, &mr->mmkey);
203 		WARN_ON(mr->descs);
204 	}
205 }
206 
207 /*
208  * This must be called after the mr has been removed from implicit_children
209  * and the SRCU synchronized.  NOTE: The MR does not necessarily have to be
210  * empty here, parallel page faults could have raced with the free process and
211  * added pages to it.
212  */
213 static void free_implicit_child_mr(struct mlx5_ib_mr *mr, bool need_imr_xlt)
214 {
215 	struct mlx5_ib_mr *imr = mr->parent;
216 	struct ib_umem_odp *odp_imr = to_ib_umem_odp(imr->umem);
217 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
218 	unsigned long idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT;
219 	int srcu_key;
220 
221 	/* implicit_child_mr's are not allowed to have deferred work */
222 	WARN_ON(atomic_read(&mr->num_deferred_work));
223 
224 	if (need_imr_xlt) {
225 		srcu_key = srcu_read_lock(&mr->dev->odp_srcu);
226 		mutex_lock(&odp_imr->umem_mutex);
227 		mlx5_ib_update_xlt(mr->parent, idx, 1, 0,
228 				   MLX5_IB_UPD_XLT_INDIRECT |
229 				   MLX5_IB_UPD_XLT_ATOMIC);
230 		mutex_unlock(&odp_imr->umem_mutex);
231 		srcu_read_unlock(&mr->dev->odp_srcu, srcu_key);
232 	}
233 
234 	dma_fence_odp_mr(mr);
235 
236 	mr->parent = NULL;
237 	mlx5_mr_cache_free(mr->dev, mr);
238 	ib_umem_odp_release(odp);
239 	if (atomic_dec_and_test(&imr->num_deferred_work))
240 		wake_up(&imr->q_deferred_work);
241 }
242 
243 static void free_implicit_child_mr_work(struct work_struct *work)
244 {
245 	struct mlx5_ib_mr *mr =
246 		container_of(work, struct mlx5_ib_mr, odp_destroy.work);
247 
248 	free_implicit_child_mr(mr, true);
249 }
250 
251 static void free_implicit_child_mr_rcu(struct rcu_head *head)
252 {
253 	struct mlx5_ib_mr *mr =
254 		container_of(head, struct mlx5_ib_mr, odp_destroy.rcu);
255 
256 	/* Freeing a MR is a sleeping operation, so bounce to a work queue */
257 	INIT_WORK(&mr->odp_destroy.work, free_implicit_child_mr_work);
258 	queue_work(system_unbound_wq, &mr->odp_destroy.work);
259 }
260 
261 static void destroy_unused_implicit_child_mr(struct mlx5_ib_mr *mr)
262 {
263 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
264 	unsigned long idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT;
265 	struct mlx5_ib_mr *imr = mr->parent;
266 
267 	xa_lock(&imr->implicit_children);
268 	/*
269 	 * This can race with mlx5_ib_free_implicit_mr(), the first one to
270 	 * reach the xa lock wins the race and destroys the MR.
271 	 */
272 	if (__xa_cmpxchg(&imr->implicit_children, idx, mr, NULL, GFP_ATOMIC) !=
273 	    mr)
274 		goto out_unlock;
275 
276 	atomic_inc(&imr->num_deferred_work);
277 	call_srcu(&mr->dev->odp_srcu, &mr->odp_destroy.rcu,
278 		  free_implicit_child_mr_rcu);
279 
280 out_unlock:
281 	xa_unlock(&imr->implicit_children);
282 }
283 
284 static bool mlx5_ib_invalidate_range(struct mmu_interval_notifier *mni,
285 				     const struct mmu_notifier_range *range,
286 				     unsigned long cur_seq)
287 {
288 	struct ib_umem_odp *umem_odp =
289 		container_of(mni, struct ib_umem_odp, notifier);
290 	struct mlx5_ib_mr *mr;
291 	const u64 umr_block_mask = (MLX5_UMR_MTT_ALIGNMENT /
292 				    sizeof(struct mlx5_mtt)) - 1;
293 	u64 idx = 0, blk_start_idx = 0;
294 	u64 invalidations = 0;
295 	unsigned long start;
296 	unsigned long end;
297 	int in_block = 0;
298 	u64 addr;
299 
300 	if (!mmu_notifier_range_blockable(range))
301 		return false;
302 
303 	mutex_lock(&umem_odp->umem_mutex);
304 	mmu_interval_set_seq(mni, cur_seq);
305 	/*
306 	 * If npages is zero then umem_odp->private may not be setup yet. This
307 	 * does not complete until after the first page is mapped for DMA.
308 	 */
309 	if (!umem_odp->npages)
310 		goto out;
311 	mr = umem_odp->private;
312 
313 	start = max_t(u64, ib_umem_start(umem_odp), range->start);
314 	end = min_t(u64, ib_umem_end(umem_odp), range->end);
315 
316 	/*
317 	 * Iteration one - zap the HW's MTTs. The notifiers_count ensures that
318 	 * while we are doing the invalidation, no page fault will attempt to
319 	 * overwrite the same MTTs.  Concurent invalidations might race us,
320 	 * but they will write 0s as well, so no difference in the end result.
321 	 */
322 	for (addr = start; addr < end; addr += BIT(umem_odp->page_shift)) {
323 		idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
324 		/*
325 		 * Strive to write the MTTs in chunks, but avoid overwriting
326 		 * non-existing MTTs. The huristic here can be improved to
327 		 * estimate the cost of another UMR vs. the cost of bigger
328 		 * UMR.
329 		 */
330 		if (umem_odp->dma_list[idx] &
331 		    (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) {
332 			if (!in_block) {
333 				blk_start_idx = idx;
334 				in_block = 1;
335 			}
336 
337 			/* Count page invalidations */
338 			invalidations += idx - blk_start_idx + 1;
339 		} else {
340 			u64 umr_offset = idx & umr_block_mask;
341 
342 			if (in_block && umr_offset == 0) {
343 				mlx5_ib_update_xlt(mr, blk_start_idx,
344 						   idx - blk_start_idx, 0,
345 						   MLX5_IB_UPD_XLT_ZAP |
346 						   MLX5_IB_UPD_XLT_ATOMIC);
347 				in_block = 0;
348 			}
349 		}
350 	}
351 	if (in_block)
352 		mlx5_ib_update_xlt(mr, blk_start_idx,
353 				   idx - blk_start_idx + 1, 0,
354 				   MLX5_IB_UPD_XLT_ZAP |
355 				   MLX5_IB_UPD_XLT_ATOMIC);
356 
357 	mlx5_update_odp_stats(mr, invalidations, invalidations);
358 
359 	/*
360 	 * We are now sure that the device will not access the
361 	 * memory. We can safely unmap it, and mark it as dirty if
362 	 * needed.
363 	 */
364 
365 	ib_umem_odp_unmap_dma_pages(umem_odp, start, end);
366 
367 	if (unlikely(!umem_odp->npages && mr->parent))
368 		destroy_unused_implicit_child_mr(mr);
369 out:
370 	mutex_unlock(&umem_odp->umem_mutex);
371 	return true;
372 }
373 
374 const struct mmu_interval_notifier_ops mlx5_mn_ops = {
375 	.invalidate = mlx5_ib_invalidate_range,
376 };
377 
378 void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev)
379 {
380 	struct ib_odp_caps *caps = &dev->odp_caps;
381 
382 	memset(caps, 0, sizeof(*caps));
383 
384 	if (!MLX5_CAP_GEN(dev->mdev, pg) ||
385 	    !mlx5_ib_can_load_pas_with_umr(dev, 0))
386 		return;
387 
388 	caps->general_caps = IB_ODP_SUPPORT;
389 
390 	if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
391 		dev->odp_max_size = U64_MAX;
392 	else
393 		dev->odp_max_size = BIT_ULL(MLX5_MAX_UMR_SHIFT + PAGE_SHIFT);
394 
395 	if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.send))
396 		caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND;
397 
398 	if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.srq_receive))
399 		caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
400 
401 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.send))
402 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND;
403 
404 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.receive))
405 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV;
406 
407 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.write))
408 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE;
409 
410 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.read))
411 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ;
412 
413 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.atomic))
414 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
415 
416 	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.srq_receive))
417 		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
418 
419 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.send))
420 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SEND;
421 
422 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.receive))
423 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_RECV;
424 
425 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.write))
426 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_WRITE;
427 
428 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.read))
429 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_READ;
430 
431 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.atomic))
432 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
433 
434 	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.srq_receive))
435 		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
436 
437 	if (MLX5_CAP_GEN(dev->mdev, fixed_buffer_size) &&
438 	    MLX5_CAP_GEN(dev->mdev, null_mkey) &&
439 	    MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset) &&
440 	    !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled))
441 		caps->general_caps |= IB_ODP_SUPPORT_IMPLICIT;
442 }
443 
444 static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev,
445 				      struct mlx5_pagefault *pfault,
446 				      int error)
447 {
448 	int wq_num = pfault->event_subtype == MLX5_PFAULT_SUBTYPE_WQE ?
449 		     pfault->wqe.wq_num : pfault->token;
450 	u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)] = {};
451 	int err;
452 
453 	MLX5_SET(page_fault_resume_in, in, opcode, MLX5_CMD_OP_PAGE_FAULT_RESUME);
454 	MLX5_SET(page_fault_resume_in, in, page_fault_type, pfault->type);
455 	MLX5_SET(page_fault_resume_in, in, token, pfault->token);
456 	MLX5_SET(page_fault_resume_in, in, wq_number, wq_num);
457 	MLX5_SET(page_fault_resume_in, in, error, !!error);
458 
459 	err = mlx5_cmd_exec_in(dev->mdev, page_fault_resume, in);
460 	if (err)
461 		mlx5_ib_err(dev, "Failed to resolve the page fault on WQ 0x%x err %d\n",
462 			    wq_num, err);
463 }
464 
465 static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,
466 						unsigned long idx)
467 {
468 	struct ib_umem_odp *odp;
469 	struct mlx5_ib_mr *mr;
470 	struct mlx5_ib_mr *ret;
471 	int err;
472 
473 	odp = ib_umem_odp_alloc_child(to_ib_umem_odp(imr->umem),
474 				      idx * MLX5_IMR_MTT_SIZE,
475 				      MLX5_IMR_MTT_SIZE, &mlx5_mn_ops);
476 	if (IS_ERR(odp))
477 		return ERR_CAST(odp);
478 
479 	ret = mr = mlx5_mr_cache_alloc(imr->dev, MLX5_IMR_MTT_CACHE_ENTRY,
480 				       imr->access_flags);
481 	if (IS_ERR(mr))
482 		goto out_umem;
483 
484 	mr->ibmr.pd = imr->ibmr.pd;
485 	mr->umem = &odp->umem;
486 	mr->ibmr.lkey = mr->mmkey.key;
487 	mr->ibmr.rkey = mr->mmkey.key;
488 	mr->mmkey.iova = idx * MLX5_IMR_MTT_SIZE;
489 	mr->parent = imr;
490 	odp->private = mr;
491 
492 	err = mlx5_ib_update_xlt(mr, 0,
493 				 MLX5_IMR_MTT_ENTRIES,
494 				 PAGE_SHIFT,
495 				 MLX5_IB_UPD_XLT_ZAP |
496 				 MLX5_IB_UPD_XLT_ENABLE);
497 	if (err) {
498 		ret = ERR_PTR(err);
499 		goto out_mr;
500 	}
501 
502 	/*
503 	 * Once the store to either xarray completes any error unwind has to
504 	 * use synchronize_srcu(). Avoid this with xa_reserve()
505 	 */
506 	ret = xa_cmpxchg(&imr->implicit_children, idx, NULL, mr,
507 			 GFP_KERNEL);
508 	if (unlikely(ret)) {
509 		if (xa_is_err(ret)) {
510 			ret = ERR_PTR(xa_err(ret));
511 			goto out_mr;
512 		}
513 		/*
514 		 * Another thread beat us to creating the child mr, use
515 		 * theirs.
516 		 */
517 		goto out_mr;
518 	}
519 
520 	mlx5_ib_dbg(imr->dev, "key %x mr %p\n", mr->mmkey.key, mr);
521 	return mr;
522 
523 out_mr:
524 	mlx5_mr_cache_free(imr->dev, mr);
525 out_umem:
526 	ib_umem_odp_release(odp);
527 	return ret;
528 }
529 
530 struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
531 					     struct ib_udata *udata,
532 					     int access_flags)
533 {
534 	struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device);
535 	struct ib_umem_odp *umem_odp;
536 	struct mlx5_ib_mr *imr;
537 	int err;
538 
539 	umem_odp = ib_umem_odp_alloc_implicit(&dev->ib_dev, access_flags);
540 	if (IS_ERR(umem_odp))
541 		return ERR_CAST(umem_odp);
542 
543 	imr = mlx5_mr_cache_alloc(dev, MLX5_IMR_KSM_CACHE_ENTRY, access_flags);
544 	if (IS_ERR(imr)) {
545 		err = PTR_ERR(imr);
546 		goto out_umem;
547 	}
548 
549 	imr->ibmr.pd = &pd->ibpd;
550 	imr->mmkey.iova = 0;
551 	imr->umem = &umem_odp->umem;
552 	imr->ibmr.lkey = imr->mmkey.key;
553 	imr->ibmr.rkey = imr->mmkey.key;
554 	imr->umem = &umem_odp->umem;
555 	imr->is_odp_implicit = true;
556 	atomic_set(&imr->num_deferred_work, 0);
557 	init_waitqueue_head(&imr->q_deferred_work);
558 	xa_init(&imr->implicit_children);
559 
560 	err = mlx5_ib_update_xlt(imr, 0,
561 				 mlx5_imr_ksm_entries,
562 				 MLX5_KSM_PAGE_SHIFT,
563 				 MLX5_IB_UPD_XLT_INDIRECT |
564 				 MLX5_IB_UPD_XLT_ZAP |
565 				 MLX5_IB_UPD_XLT_ENABLE);
566 	if (err)
567 		goto out_mr;
568 
569 	err = xa_err(xa_store(&dev->odp_mkeys, mlx5_base_mkey(imr->mmkey.key),
570 			      &imr->mmkey, GFP_KERNEL));
571 	if (err)
572 		goto out_mr;
573 
574 	mlx5_ib_dbg(dev, "key %x mr %p\n", imr->mmkey.key, imr);
575 	return imr;
576 out_mr:
577 	mlx5_ib_err(dev, "Failed to register MKEY %d\n", err);
578 	mlx5_mr_cache_free(dev, imr);
579 out_umem:
580 	ib_umem_odp_release(umem_odp);
581 	return ERR_PTR(err);
582 }
583 
584 void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *imr)
585 {
586 	struct ib_umem_odp *odp_imr = to_ib_umem_odp(imr->umem);
587 	struct mlx5_ib_dev *dev = imr->dev;
588 	struct list_head destroy_list;
589 	struct mlx5_ib_mr *mtt;
590 	struct mlx5_ib_mr *tmp;
591 	unsigned long idx;
592 
593 	INIT_LIST_HEAD(&destroy_list);
594 
595 	xa_erase(&dev->odp_mkeys, mlx5_base_mkey(imr->mmkey.key));
596 	/*
597 	 * This stops the SRCU protected page fault path from touching either
598 	 * the imr or any children. The page fault path can only reach the
599 	 * children xarray via the imr.
600 	 */
601 	synchronize_srcu(&dev->odp_srcu);
602 
603 	/*
604 	 * All work on the prefetch list must be completed, xa_erase() prevented
605 	 * new work from being created.
606 	 */
607 	wait_event(imr->q_deferred_work, !atomic_read(&imr->num_deferred_work));
608 
609 	/*
610 	 * At this point it is forbidden for any other thread to enter
611 	 * pagefault_mr() on this imr. It is already forbidden to call
612 	 * pagefault_mr() on an implicit child. Due to this additions to
613 	 * implicit_children are prevented.
614 	 */
615 
616 	/*
617 	 * Block destroy_unused_implicit_child_mr() from incrementing
618 	 * num_deferred_work.
619 	 */
620 	xa_lock(&imr->implicit_children);
621 	xa_for_each (&imr->implicit_children, idx, mtt) {
622 		__xa_erase(&imr->implicit_children, idx);
623 		list_add(&mtt->odp_destroy.elm, &destroy_list);
624 	}
625 	xa_unlock(&imr->implicit_children);
626 
627 	/*
628 	 * Wait for any concurrent destroy_unused_implicit_child_mr() to
629 	 * complete.
630 	 */
631 	wait_event(imr->q_deferred_work, !atomic_read(&imr->num_deferred_work));
632 
633 	/*
634 	 * Fence the imr before we destroy the children. This allows us to
635 	 * skip updating the XLT of the imr during destroy of the child mkey
636 	 * the imr points to.
637 	 */
638 	mlx5_mr_cache_invalidate(imr);
639 
640 	list_for_each_entry_safe (mtt, tmp, &destroy_list, odp_destroy.elm)
641 		free_implicit_child_mr(mtt, false);
642 
643 	mlx5_mr_cache_free(dev, imr);
644 	ib_umem_odp_release(odp_imr);
645 }
646 
647 /**
648  * mlx5_ib_fence_odp_mr - Stop all access to the ODP MR
649  * @mr: to fence
650  *
651  * On return no parallel threads will be touching this MR and no DMA will be
652  * active.
653  */
654 void mlx5_ib_fence_odp_mr(struct mlx5_ib_mr *mr)
655 {
656 	/* Prevent new page faults and prefetch requests from succeeding */
657 	xa_erase(&mr->dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key));
658 
659 	/* Wait for all running page-fault handlers to finish. */
660 	synchronize_srcu(&mr->dev->odp_srcu);
661 
662 	wait_event(mr->q_deferred_work, !atomic_read(&mr->num_deferred_work));
663 
664 	dma_fence_odp_mr(mr);
665 }
666 
667 #define MLX5_PF_FLAGS_DOWNGRADE BIT(1)
668 #define MLX5_PF_FLAGS_SNAPSHOT BIT(2)
669 #define MLX5_PF_FLAGS_ENABLE BIT(3)
670 static int pagefault_real_mr(struct mlx5_ib_mr *mr, struct ib_umem_odp *odp,
671 			     u64 user_va, size_t bcnt, u32 *bytes_mapped,
672 			     u32 flags)
673 {
674 	int page_shift, ret, np;
675 	bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE;
676 	u64 access_mask;
677 	u64 start_idx;
678 	bool fault = !(flags & MLX5_PF_FLAGS_SNAPSHOT);
679 	u32 xlt_flags = MLX5_IB_UPD_XLT_ATOMIC;
680 
681 	if (flags & MLX5_PF_FLAGS_ENABLE)
682 		xlt_flags |= MLX5_IB_UPD_XLT_ENABLE;
683 
684 	page_shift = odp->page_shift;
685 	start_idx = (user_va - ib_umem_start(odp)) >> page_shift;
686 	access_mask = ODP_READ_ALLOWED_BIT;
687 
688 	if (odp->umem.writable && !downgrade)
689 		access_mask |= ODP_WRITE_ALLOWED_BIT;
690 
691 	np = ib_umem_odp_map_dma_and_lock(odp, user_va, bcnt, access_mask, fault);
692 	if (np < 0)
693 		return np;
694 
695 	/*
696 	 * No need to check whether the MTTs really belong to this MR, since
697 	 * ib_umem_odp_map_dma_and_lock already checks this.
698 	 */
699 	ret = mlx5_ib_update_xlt(mr, start_idx, np, page_shift, xlt_flags);
700 	mutex_unlock(&odp->umem_mutex);
701 
702 	if (ret < 0) {
703 		if (ret != -EAGAIN)
704 			mlx5_ib_err(mr->dev,
705 				    "Failed to update mkey page tables\n");
706 		goto out;
707 	}
708 
709 	if (bytes_mapped) {
710 		u32 new_mappings = (np << page_shift) -
711 			(user_va - round_down(user_va, 1 << page_shift));
712 
713 		*bytes_mapped += min_t(u32, new_mappings, bcnt);
714 	}
715 
716 	return np << (page_shift - PAGE_SHIFT);
717 
718 out:
719 	return ret;
720 }
721 
722 static int pagefault_implicit_mr(struct mlx5_ib_mr *imr,
723 				 struct ib_umem_odp *odp_imr, u64 user_va,
724 				 size_t bcnt, u32 *bytes_mapped, u32 flags)
725 {
726 	unsigned long end_idx = (user_va + bcnt - 1) >> MLX5_IMR_MTT_SHIFT;
727 	unsigned long upd_start_idx = end_idx + 1;
728 	unsigned long upd_len = 0;
729 	unsigned long npages = 0;
730 	int err;
731 	int ret;
732 
733 	if (unlikely(user_va >= mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE ||
734 		     mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE - user_va < bcnt))
735 		return -EFAULT;
736 
737 	/* Fault each child mr that intersects with our interval. */
738 	while (bcnt) {
739 		unsigned long idx = user_va >> MLX5_IMR_MTT_SHIFT;
740 		struct ib_umem_odp *umem_odp;
741 		struct mlx5_ib_mr *mtt;
742 		u64 len;
743 
744 		mtt = xa_load(&imr->implicit_children, idx);
745 		if (unlikely(!mtt)) {
746 			mtt = implicit_get_child_mr(imr, idx);
747 			if (IS_ERR(mtt)) {
748 				ret = PTR_ERR(mtt);
749 				goto out;
750 			}
751 			upd_start_idx = min(upd_start_idx, idx);
752 			upd_len = idx - upd_start_idx + 1;
753 		}
754 
755 		umem_odp = to_ib_umem_odp(mtt->umem);
756 		len = min_t(u64, user_va + bcnt, ib_umem_end(umem_odp)) -
757 		      user_va;
758 
759 		ret = pagefault_real_mr(mtt, umem_odp, user_va, len,
760 					bytes_mapped, flags);
761 		if (ret < 0)
762 			goto out;
763 		user_va += len;
764 		bcnt -= len;
765 		npages += ret;
766 	}
767 
768 	ret = npages;
769 
770 	/*
771 	 * Any time the implicit_children are changed we must perform an
772 	 * update of the xlt before exiting to ensure the HW and the
773 	 * implicit_children remains synchronized.
774 	 */
775 out:
776 	if (likely(!upd_len))
777 		return ret;
778 
779 	/*
780 	 * Notice this is not strictly ordered right, the KSM is updated after
781 	 * the implicit_children is updated, so a parallel page fault could
782 	 * see a MR that is not yet visible in the KSM.  This is similar to a
783 	 * parallel page fault seeing a MR that is being concurrently removed
784 	 * from the KSM. Both of these improbable situations are resolved
785 	 * safely by resuming the HW and then taking another page fault. The
786 	 * next pagefault handler will see the new information.
787 	 */
788 	mutex_lock(&odp_imr->umem_mutex);
789 	err = mlx5_ib_update_xlt(imr, upd_start_idx, upd_len, 0,
790 				 MLX5_IB_UPD_XLT_INDIRECT |
791 					 MLX5_IB_UPD_XLT_ATOMIC);
792 	mutex_unlock(&odp_imr->umem_mutex);
793 	if (err) {
794 		mlx5_ib_err(imr->dev, "Failed to update PAS\n");
795 		return err;
796 	}
797 	return ret;
798 }
799 
800 /*
801  * Returns:
802  *  -EFAULT: The io_virt->bcnt is not within the MR, it covers pages that are
803  *           not accessible, or the MR is no longer valid.
804  *  -EAGAIN/-ENOMEM: The operation should be retried
805  *
806  *  -EINVAL/others: General internal malfunction
807  *  >0: Number of pages mapped
808  */
809 static int pagefault_mr(struct mlx5_ib_mr *mr, u64 io_virt, size_t bcnt,
810 			u32 *bytes_mapped, u32 flags)
811 {
812 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
813 
814 	lockdep_assert_held(&mr->dev->odp_srcu);
815 	if (unlikely(io_virt < mr->mmkey.iova))
816 		return -EFAULT;
817 
818 	if (!odp->is_implicit_odp) {
819 		u64 user_va;
820 
821 		if (check_add_overflow(io_virt - mr->mmkey.iova,
822 				       (u64)odp->umem.address, &user_va))
823 			return -EFAULT;
824 		if (unlikely(user_va >= ib_umem_end(odp) ||
825 			     ib_umem_end(odp) - user_va < bcnt))
826 			return -EFAULT;
827 		return pagefault_real_mr(mr, odp, user_va, bcnt, bytes_mapped,
828 					 flags);
829 	}
830 	return pagefault_implicit_mr(mr, odp, io_virt, bcnt, bytes_mapped,
831 				     flags);
832 }
833 
834 int mlx5_ib_init_odp_mr(struct mlx5_ib_mr *mr, bool enable)
835 {
836 	u32 flags = MLX5_PF_FLAGS_SNAPSHOT;
837 	int ret;
838 
839 	if (enable)
840 		flags |= MLX5_PF_FLAGS_ENABLE;
841 
842 	ret = pagefault_real_mr(mr, to_ib_umem_odp(mr->umem),
843 				mr->umem->address, mr->umem->length, NULL,
844 				flags);
845 	return ret >= 0 ? 0 : ret;
846 }
847 
848 struct pf_frame {
849 	struct pf_frame *next;
850 	u32 key;
851 	u64 io_virt;
852 	size_t bcnt;
853 	int depth;
854 };
855 
856 static bool mkey_is_eq(struct mlx5_core_mkey *mmkey, u32 key)
857 {
858 	if (!mmkey)
859 		return false;
860 	if (mmkey->type == MLX5_MKEY_MW)
861 		return mlx5_base_mkey(mmkey->key) == mlx5_base_mkey(key);
862 	return mmkey->key == key;
863 }
864 
865 static int get_indirect_num_descs(struct mlx5_core_mkey *mmkey)
866 {
867 	struct mlx5_ib_mw *mw;
868 	struct mlx5_ib_devx_mr *devx_mr;
869 
870 	if (mmkey->type == MLX5_MKEY_MW) {
871 		mw = container_of(mmkey, struct mlx5_ib_mw, mmkey);
872 		return mw->ndescs;
873 	}
874 
875 	devx_mr = container_of(mmkey, struct mlx5_ib_devx_mr,
876 			       mmkey);
877 	return devx_mr->ndescs;
878 }
879 
880 /*
881  * Handle a single data segment in a page-fault WQE or RDMA region.
882  *
883  * Returns number of OS pages retrieved on success. The caller may continue to
884  * the next data segment.
885  * Can return the following error codes:
886  * -EAGAIN to designate a temporary error. The caller will abort handling the
887  *  page fault and resolve it.
888  * -EFAULT when there's an error mapping the requested pages. The caller will
889  *  abort the page fault handling.
890  */
891 static int pagefault_single_data_segment(struct mlx5_ib_dev *dev,
892 					 struct ib_pd *pd, u32 key,
893 					 u64 io_virt, size_t bcnt,
894 					 u32 *bytes_committed,
895 					 u32 *bytes_mapped)
896 {
897 	int npages = 0, srcu_key, ret, i, outlen, cur_outlen = 0, depth = 0;
898 	struct pf_frame *head = NULL, *frame;
899 	struct mlx5_core_mkey *mmkey;
900 	struct mlx5_ib_mr *mr;
901 	struct mlx5_klm *pklm;
902 	u32 *out = NULL;
903 	size_t offset;
904 	int ndescs;
905 
906 	srcu_key = srcu_read_lock(&dev->odp_srcu);
907 
908 	io_virt += *bytes_committed;
909 	bcnt -= *bytes_committed;
910 
911 next_mr:
912 	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(key));
913 	if (!mmkey) {
914 		mlx5_ib_dbg(
915 			dev,
916 			"skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
917 			key);
918 		if (bytes_mapped)
919 			*bytes_mapped += bcnt;
920 		/*
921 		 * The user could specify a SGL with multiple lkeys and only
922 		 * some of them are ODP. Treat the non-ODP ones as fully
923 		 * faulted.
924 		 */
925 		ret = 0;
926 		goto srcu_unlock;
927 	}
928 	if (!mkey_is_eq(mmkey, key)) {
929 		mlx5_ib_dbg(dev, "failed to find mkey %x\n", key);
930 		ret = -EFAULT;
931 		goto srcu_unlock;
932 	}
933 
934 	switch (mmkey->type) {
935 	case MLX5_MKEY_MR:
936 		mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
937 
938 		ret = pagefault_mr(mr, io_virt, bcnt, bytes_mapped, 0);
939 		if (ret < 0)
940 			goto srcu_unlock;
941 
942 		mlx5_update_odp_stats(mr, faults, ret);
943 
944 		npages += ret;
945 		ret = 0;
946 		break;
947 
948 	case MLX5_MKEY_MW:
949 	case MLX5_MKEY_INDIRECT_DEVX:
950 		ndescs = get_indirect_num_descs(mmkey);
951 
952 		if (depth >= MLX5_CAP_GEN(dev->mdev, max_indirection)) {
953 			mlx5_ib_dbg(dev, "indirection level exceeded\n");
954 			ret = -EFAULT;
955 			goto srcu_unlock;
956 		}
957 
958 		outlen = MLX5_ST_SZ_BYTES(query_mkey_out) +
959 			sizeof(*pklm) * (ndescs - 2);
960 
961 		if (outlen > cur_outlen) {
962 			kfree(out);
963 			out = kzalloc(outlen, GFP_KERNEL);
964 			if (!out) {
965 				ret = -ENOMEM;
966 				goto srcu_unlock;
967 			}
968 			cur_outlen = outlen;
969 		}
970 
971 		pklm = (struct mlx5_klm *)MLX5_ADDR_OF(query_mkey_out, out,
972 						       bsf0_klm0_pas_mtt0_1);
973 
974 		ret = mlx5_core_query_mkey(dev->mdev, mmkey, out, outlen);
975 		if (ret)
976 			goto srcu_unlock;
977 
978 		offset = io_virt - MLX5_GET64(query_mkey_out, out,
979 					      memory_key_mkey_entry.start_addr);
980 
981 		for (i = 0; bcnt && i < ndescs; i++, pklm++) {
982 			if (offset >= be32_to_cpu(pklm->bcount)) {
983 				offset -= be32_to_cpu(pklm->bcount);
984 				continue;
985 			}
986 
987 			frame = kzalloc(sizeof(*frame), GFP_KERNEL);
988 			if (!frame) {
989 				ret = -ENOMEM;
990 				goto srcu_unlock;
991 			}
992 
993 			frame->key = be32_to_cpu(pklm->key);
994 			frame->io_virt = be64_to_cpu(pklm->va) + offset;
995 			frame->bcnt = min_t(size_t, bcnt,
996 					    be32_to_cpu(pklm->bcount) - offset);
997 			frame->depth = depth + 1;
998 			frame->next = head;
999 			head = frame;
1000 
1001 			bcnt -= frame->bcnt;
1002 			offset = 0;
1003 		}
1004 		break;
1005 
1006 	default:
1007 		mlx5_ib_dbg(dev, "wrong mkey type %d\n", mmkey->type);
1008 		ret = -EFAULT;
1009 		goto srcu_unlock;
1010 	}
1011 
1012 	if (head) {
1013 		frame = head;
1014 		head = frame->next;
1015 
1016 		key = frame->key;
1017 		io_virt = frame->io_virt;
1018 		bcnt = frame->bcnt;
1019 		depth = frame->depth;
1020 		kfree(frame);
1021 
1022 		goto next_mr;
1023 	}
1024 
1025 srcu_unlock:
1026 	while (head) {
1027 		frame = head;
1028 		head = frame->next;
1029 		kfree(frame);
1030 	}
1031 	kfree(out);
1032 
1033 	srcu_read_unlock(&dev->odp_srcu, srcu_key);
1034 	*bytes_committed = 0;
1035 	return ret ? ret : npages;
1036 }
1037 
1038 /**
1039  * Parse a series of data segments for page fault handling.
1040  *
1041  * @pfault contains page fault information.
1042  * @wqe points at the first data segment in the WQE.
1043  * @wqe_end points after the end of the WQE.
1044  * @bytes_mapped receives the number of bytes that the function was able to
1045  *               map. This allows the caller to decide intelligently whether
1046  *               enough memory was mapped to resolve the page fault
1047  *               successfully (e.g. enough for the next MTU, or the entire
1048  *               WQE).
1049  * @total_wqe_bytes receives the total data size of this WQE in bytes (minus
1050  *                  the committed bytes).
1051  *
1052  * Returns the number of pages loaded if positive, zero for an empty WQE, or a
1053  * negative error code.
1054  */
1055 static int pagefault_data_segments(struct mlx5_ib_dev *dev,
1056 				   struct mlx5_pagefault *pfault,
1057 				   void *wqe,
1058 				   void *wqe_end, u32 *bytes_mapped,
1059 				   u32 *total_wqe_bytes, bool receive_queue)
1060 {
1061 	int ret = 0, npages = 0;
1062 	u64 io_virt;
1063 	u32 key;
1064 	u32 byte_count;
1065 	size_t bcnt;
1066 	int inline_segment;
1067 
1068 	if (bytes_mapped)
1069 		*bytes_mapped = 0;
1070 	if (total_wqe_bytes)
1071 		*total_wqe_bytes = 0;
1072 
1073 	while (wqe < wqe_end) {
1074 		struct mlx5_wqe_data_seg *dseg = wqe;
1075 
1076 		io_virt = be64_to_cpu(dseg->addr);
1077 		key = be32_to_cpu(dseg->lkey);
1078 		byte_count = be32_to_cpu(dseg->byte_count);
1079 		inline_segment = !!(byte_count &  MLX5_INLINE_SEG);
1080 		bcnt	       = byte_count & ~MLX5_INLINE_SEG;
1081 
1082 		if (inline_segment) {
1083 			bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK;
1084 			wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt,
1085 				     16);
1086 		} else {
1087 			wqe += sizeof(*dseg);
1088 		}
1089 
1090 		/* receive WQE end of sg list. */
1091 		if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY &&
1092 		    io_virt == 0)
1093 			break;
1094 
1095 		if (!inline_segment && total_wqe_bytes) {
1096 			*total_wqe_bytes += bcnt - min_t(size_t, bcnt,
1097 					pfault->bytes_committed);
1098 		}
1099 
1100 		/* A zero length data segment designates a length of 2GB. */
1101 		if (bcnt == 0)
1102 			bcnt = 1U << 31;
1103 
1104 		if (inline_segment || bcnt <= pfault->bytes_committed) {
1105 			pfault->bytes_committed -=
1106 				min_t(size_t, bcnt,
1107 				      pfault->bytes_committed);
1108 			continue;
1109 		}
1110 
1111 		ret = pagefault_single_data_segment(dev, NULL, key,
1112 						    io_virt, bcnt,
1113 						    &pfault->bytes_committed,
1114 						    bytes_mapped);
1115 		if (ret < 0)
1116 			break;
1117 		npages += ret;
1118 	}
1119 
1120 	return ret < 0 ? ret : npages;
1121 }
1122 
1123 /*
1124  * Parse initiator WQE. Advances the wqe pointer to point at the
1125  * scatter-gather list, and set wqe_end to the end of the WQE.
1126  */
1127 static int mlx5_ib_mr_initiator_pfault_handler(
1128 	struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault,
1129 	struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length)
1130 {
1131 	struct mlx5_wqe_ctrl_seg *ctrl = *wqe;
1132 	u16 wqe_index = pfault->wqe.wqe_index;
1133 	struct mlx5_base_av *av;
1134 	unsigned ds, opcode;
1135 	u32 qpn = qp->trans_qp.base.mqp.qpn;
1136 
1137 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
1138 	if (ds * MLX5_WQE_DS_UNITS > wqe_length) {
1139 		mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n",
1140 			    ds, wqe_length);
1141 		return -EFAULT;
1142 	}
1143 
1144 	if (ds == 0) {
1145 		mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n",
1146 			    wqe_index, qpn);
1147 		return -EFAULT;
1148 	}
1149 
1150 	*wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS;
1151 	*wqe += sizeof(*ctrl);
1152 
1153 	opcode = be32_to_cpu(ctrl->opmod_idx_opcode) &
1154 		 MLX5_WQE_CTRL_OPCODE_MASK;
1155 
1156 	if (qp->ibqp.qp_type == IB_QPT_XRC_INI)
1157 		*wqe += sizeof(struct mlx5_wqe_xrc_seg);
1158 
1159 	if (qp->type == IB_QPT_UD || qp->type == MLX5_IB_QPT_DCI) {
1160 		av = *wqe;
1161 		if (av->dqp_dct & cpu_to_be32(MLX5_EXTENDED_UD_AV))
1162 			*wqe += sizeof(struct mlx5_av);
1163 		else
1164 			*wqe += sizeof(struct mlx5_base_av);
1165 	}
1166 
1167 	switch (opcode) {
1168 	case MLX5_OPCODE_RDMA_WRITE:
1169 	case MLX5_OPCODE_RDMA_WRITE_IMM:
1170 	case MLX5_OPCODE_RDMA_READ:
1171 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1172 		break;
1173 	case MLX5_OPCODE_ATOMIC_CS:
1174 	case MLX5_OPCODE_ATOMIC_FA:
1175 		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
1176 		*wqe += sizeof(struct mlx5_wqe_atomic_seg);
1177 		break;
1178 	}
1179 
1180 	return 0;
1181 }
1182 
1183 /*
1184  * Parse responder WQE and set wqe_end to the end of the WQE.
1185  */
1186 static int mlx5_ib_mr_responder_pfault_handler_srq(struct mlx5_ib_dev *dev,
1187 						   struct mlx5_ib_srq *srq,
1188 						   void **wqe, void **wqe_end,
1189 						   int wqe_length)
1190 {
1191 	int wqe_size = 1 << srq->msrq.wqe_shift;
1192 
1193 	if (wqe_size > wqe_length) {
1194 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1195 		return -EFAULT;
1196 	}
1197 
1198 	*wqe_end = *wqe + wqe_size;
1199 	*wqe += sizeof(struct mlx5_wqe_srq_next_seg);
1200 
1201 	return 0;
1202 }
1203 
1204 static int mlx5_ib_mr_responder_pfault_handler_rq(struct mlx5_ib_dev *dev,
1205 						  struct mlx5_ib_qp *qp,
1206 						  void *wqe, void **wqe_end,
1207 						  int wqe_length)
1208 {
1209 	struct mlx5_ib_wq *wq = &qp->rq;
1210 	int wqe_size = 1 << wq->wqe_shift;
1211 
1212 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE) {
1213 		mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n");
1214 		return -EFAULT;
1215 	}
1216 
1217 	if (wqe_size > wqe_length) {
1218 		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1219 		return -EFAULT;
1220 	}
1221 
1222 	*wqe_end = wqe + wqe_size;
1223 
1224 	return 0;
1225 }
1226 
1227 static inline struct mlx5_core_rsc_common *odp_get_rsc(struct mlx5_ib_dev *dev,
1228 						       u32 wq_num, int pf_type)
1229 {
1230 	struct mlx5_core_rsc_common *common = NULL;
1231 	struct mlx5_core_srq *srq;
1232 
1233 	switch (pf_type) {
1234 	case MLX5_WQE_PF_TYPE_RMP:
1235 		srq = mlx5_cmd_get_srq(dev, wq_num);
1236 		if (srq)
1237 			common = &srq->common;
1238 		break;
1239 	case MLX5_WQE_PF_TYPE_REQ_SEND_OR_WRITE:
1240 	case MLX5_WQE_PF_TYPE_RESP:
1241 	case MLX5_WQE_PF_TYPE_REQ_READ_OR_ATOMIC:
1242 		common = mlx5_core_res_hold(dev, wq_num, MLX5_RES_QP);
1243 		break;
1244 	default:
1245 		break;
1246 	}
1247 
1248 	return common;
1249 }
1250 
1251 static inline struct mlx5_ib_qp *res_to_qp(struct mlx5_core_rsc_common *res)
1252 {
1253 	struct mlx5_core_qp *mqp = (struct mlx5_core_qp *)res;
1254 
1255 	return to_mibqp(mqp);
1256 }
1257 
1258 static inline struct mlx5_ib_srq *res_to_srq(struct mlx5_core_rsc_common *res)
1259 {
1260 	struct mlx5_core_srq *msrq =
1261 		container_of(res, struct mlx5_core_srq, common);
1262 
1263 	return to_mibsrq(msrq);
1264 }
1265 
1266 static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
1267 					  struct mlx5_pagefault *pfault)
1268 {
1269 	bool sq = pfault->type & MLX5_PFAULT_REQUESTOR;
1270 	u16 wqe_index = pfault->wqe.wqe_index;
1271 	void *wqe, *wqe_start = NULL, *wqe_end = NULL;
1272 	u32 bytes_mapped, total_wqe_bytes;
1273 	struct mlx5_core_rsc_common *res;
1274 	int resume_with_error = 1;
1275 	struct mlx5_ib_qp *qp;
1276 	size_t bytes_copied;
1277 	int ret = 0;
1278 
1279 	res = odp_get_rsc(dev, pfault->wqe.wq_num, pfault->type);
1280 	if (!res) {
1281 		mlx5_ib_dbg(dev, "wqe page fault for missing resource %d\n", pfault->wqe.wq_num);
1282 		return;
1283 	}
1284 
1285 	if (res->res != MLX5_RES_QP && res->res != MLX5_RES_SRQ &&
1286 	    res->res != MLX5_RES_XSRQ) {
1287 		mlx5_ib_err(dev, "wqe page fault for unsupported type %d\n",
1288 			    pfault->type);
1289 		goto resolve_page_fault;
1290 	}
1291 
1292 	wqe_start = (void *)__get_free_page(GFP_KERNEL);
1293 	if (!wqe_start) {
1294 		mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n");
1295 		goto resolve_page_fault;
1296 	}
1297 
1298 	wqe = wqe_start;
1299 	qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL;
1300 	if (qp && sq) {
1301 		ret = mlx5_ib_read_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE,
1302 					  &bytes_copied);
1303 		if (ret)
1304 			goto read_user;
1305 		ret = mlx5_ib_mr_initiator_pfault_handler(
1306 			dev, pfault, qp, &wqe, &wqe_end, bytes_copied);
1307 	} else if (qp && !sq) {
1308 		ret = mlx5_ib_read_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE,
1309 					  &bytes_copied);
1310 		if (ret)
1311 			goto read_user;
1312 		ret = mlx5_ib_mr_responder_pfault_handler_rq(
1313 			dev, qp, wqe, &wqe_end, bytes_copied);
1314 	} else if (!qp) {
1315 		struct mlx5_ib_srq *srq = res_to_srq(res);
1316 
1317 		ret = mlx5_ib_read_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE,
1318 					   &bytes_copied);
1319 		if (ret)
1320 			goto read_user;
1321 		ret = mlx5_ib_mr_responder_pfault_handler_srq(
1322 			dev, srq, &wqe, &wqe_end, bytes_copied);
1323 	}
1324 
1325 	if (ret < 0 || wqe >= wqe_end)
1326 		goto resolve_page_fault;
1327 
1328 	ret = pagefault_data_segments(dev, pfault, wqe, wqe_end, &bytes_mapped,
1329 				      &total_wqe_bytes, !sq);
1330 	if (ret == -EAGAIN)
1331 		goto out;
1332 
1333 	if (ret < 0 || total_wqe_bytes > bytes_mapped)
1334 		goto resolve_page_fault;
1335 
1336 out:
1337 	ret = 0;
1338 	resume_with_error = 0;
1339 
1340 read_user:
1341 	if (ret)
1342 		mlx5_ib_err(
1343 			dev,
1344 			"Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n",
1345 			ret, wqe_index, pfault->token);
1346 
1347 resolve_page_fault:
1348 	mlx5_ib_page_fault_resume(dev, pfault, resume_with_error);
1349 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n",
1350 		    pfault->wqe.wq_num, resume_with_error,
1351 		    pfault->type);
1352 	mlx5_core_res_put(res);
1353 	free_page((unsigned long)wqe_start);
1354 }
1355 
1356 static int pages_in_range(u64 address, u32 length)
1357 {
1358 	return (ALIGN(address + length, PAGE_SIZE) -
1359 		(address & PAGE_MASK)) >> PAGE_SHIFT;
1360 }
1361 
1362 static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev,
1363 					   struct mlx5_pagefault *pfault)
1364 {
1365 	u64 address;
1366 	u32 length;
1367 	u32 prefetch_len = pfault->bytes_committed;
1368 	int prefetch_activated = 0;
1369 	u32 rkey = pfault->rdma.r_key;
1370 	int ret;
1371 
1372 	/* The RDMA responder handler handles the page fault in two parts.
1373 	 * First it brings the necessary pages for the current packet
1374 	 * (and uses the pfault context), and then (after resuming the QP)
1375 	 * prefetches more pages. The second operation cannot use the pfault
1376 	 * context and therefore uses the dummy_pfault context allocated on
1377 	 * the stack */
1378 	pfault->rdma.rdma_va += pfault->bytes_committed;
1379 	pfault->rdma.rdma_op_len -= min(pfault->bytes_committed,
1380 					 pfault->rdma.rdma_op_len);
1381 	pfault->bytes_committed = 0;
1382 
1383 	address = pfault->rdma.rdma_va;
1384 	length  = pfault->rdma.rdma_op_len;
1385 
1386 	/* For some operations, the hardware cannot tell the exact message
1387 	 * length, and in those cases it reports zero. Use prefetch
1388 	 * logic. */
1389 	if (length == 0) {
1390 		prefetch_activated = 1;
1391 		length = pfault->rdma.packet_size;
1392 		prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len);
1393 	}
1394 
1395 	ret = pagefault_single_data_segment(dev, NULL, rkey, address, length,
1396 					    &pfault->bytes_committed, NULL);
1397 	if (ret == -EAGAIN) {
1398 		/* We're racing with an invalidation, don't prefetch */
1399 		prefetch_activated = 0;
1400 	} else if (ret < 0 || pages_in_range(address, length) > ret) {
1401 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1402 		if (ret != -ENOENT)
1403 			mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n",
1404 				    ret, pfault->token, pfault->type);
1405 		return;
1406 	}
1407 
1408 	mlx5_ib_page_fault_resume(dev, pfault, 0);
1409 	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n",
1410 		    pfault->token, pfault->type,
1411 		    prefetch_activated);
1412 
1413 	/* At this point, there might be a new pagefault already arriving in
1414 	 * the eq, switch to the dummy pagefault for the rest of the
1415 	 * processing. We're still OK with the objects being alive as the
1416 	 * work-queue is being fenced. */
1417 
1418 	if (prefetch_activated) {
1419 		u32 bytes_committed = 0;
1420 
1421 		ret = pagefault_single_data_segment(dev, NULL, rkey, address,
1422 						    prefetch_len,
1423 						    &bytes_committed, NULL);
1424 		if (ret < 0 && ret != -EAGAIN) {
1425 			mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n",
1426 				    ret, pfault->token, address, prefetch_len);
1427 		}
1428 	}
1429 }
1430 
1431 static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault)
1432 {
1433 	u8 event_subtype = pfault->event_subtype;
1434 
1435 	switch (event_subtype) {
1436 	case MLX5_PFAULT_SUBTYPE_WQE:
1437 		mlx5_ib_mr_wqe_pfault_handler(dev, pfault);
1438 		break;
1439 	case MLX5_PFAULT_SUBTYPE_RDMA:
1440 		mlx5_ib_mr_rdma_pfault_handler(dev, pfault);
1441 		break;
1442 	default:
1443 		mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n",
1444 			    event_subtype);
1445 		mlx5_ib_page_fault_resume(dev, pfault, 1);
1446 	}
1447 }
1448 
1449 static void mlx5_ib_eqe_pf_action(struct work_struct *work)
1450 {
1451 	struct mlx5_pagefault *pfault = container_of(work,
1452 						     struct mlx5_pagefault,
1453 						     work);
1454 	struct mlx5_ib_pf_eq *eq = pfault->eq;
1455 
1456 	mlx5_ib_pfault(eq->dev, pfault);
1457 	mempool_free(pfault, eq->pool);
1458 }
1459 
1460 static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq)
1461 {
1462 	struct mlx5_eqe_page_fault *pf_eqe;
1463 	struct mlx5_pagefault *pfault;
1464 	struct mlx5_eqe *eqe;
1465 	int cc = 0;
1466 
1467 	while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) {
1468 		pfault = mempool_alloc(eq->pool, GFP_ATOMIC);
1469 		if (!pfault) {
1470 			schedule_work(&eq->work);
1471 			break;
1472 		}
1473 
1474 		pf_eqe = &eqe->data.page_fault;
1475 		pfault->event_subtype = eqe->sub_type;
1476 		pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);
1477 
1478 		mlx5_ib_dbg(eq->dev,
1479 			    "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
1480 			    eqe->sub_type, pfault->bytes_committed);
1481 
1482 		switch (eqe->sub_type) {
1483 		case MLX5_PFAULT_SUBTYPE_RDMA:
1484 			/* RDMA based event */
1485 			pfault->type =
1486 				be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
1487 			pfault->token =
1488 				be32_to_cpu(pf_eqe->rdma.pftype_token) &
1489 				MLX5_24BIT_MASK;
1490 			pfault->rdma.r_key =
1491 				be32_to_cpu(pf_eqe->rdma.r_key);
1492 			pfault->rdma.packet_size =
1493 				be16_to_cpu(pf_eqe->rdma.packet_length);
1494 			pfault->rdma.rdma_op_len =
1495 				be32_to_cpu(pf_eqe->rdma.rdma_op_len);
1496 			pfault->rdma.rdma_va =
1497 				be64_to_cpu(pf_eqe->rdma.rdma_va);
1498 			mlx5_ib_dbg(eq->dev,
1499 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
1500 				    pfault->type, pfault->token,
1501 				    pfault->rdma.r_key);
1502 			mlx5_ib_dbg(eq->dev,
1503 				    "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
1504 				    pfault->rdma.rdma_op_len,
1505 				    pfault->rdma.rdma_va);
1506 			break;
1507 
1508 		case MLX5_PFAULT_SUBTYPE_WQE:
1509 			/* WQE based event */
1510 			pfault->type =
1511 				(be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7;
1512 			pfault->token =
1513 				be32_to_cpu(pf_eqe->wqe.token);
1514 			pfault->wqe.wq_num =
1515 				be32_to_cpu(pf_eqe->wqe.pftype_wq) &
1516 				MLX5_24BIT_MASK;
1517 			pfault->wqe.wqe_index =
1518 				be16_to_cpu(pf_eqe->wqe.wqe_index);
1519 			pfault->wqe.packet_size =
1520 				be16_to_cpu(pf_eqe->wqe.packet_length);
1521 			mlx5_ib_dbg(eq->dev,
1522 				    "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
1523 				    pfault->type, pfault->token,
1524 				    pfault->wqe.wq_num,
1525 				    pfault->wqe.wqe_index);
1526 			break;
1527 
1528 		default:
1529 			mlx5_ib_warn(eq->dev,
1530 				     "Unsupported page fault event sub-type: 0x%02hhx\n",
1531 				     eqe->sub_type);
1532 			/* Unsupported page faults should still be
1533 			 * resolved by the page fault handler
1534 			 */
1535 		}
1536 
1537 		pfault->eq = eq;
1538 		INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action);
1539 		queue_work(eq->wq, &pfault->work);
1540 
1541 		cc = mlx5_eq_update_cc(eq->core, ++cc);
1542 	}
1543 
1544 	mlx5_eq_update_ci(eq->core, cc, 1);
1545 }
1546 
1547 static int mlx5_ib_eq_pf_int(struct notifier_block *nb, unsigned long type,
1548 			     void *data)
1549 {
1550 	struct mlx5_ib_pf_eq *eq =
1551 		container_of(nb, struct mlx5_ib_pf_eq, irq_nb);
1552 	unsigned long flags;
1553 
1554 	if (spin_trylock_irqsave(&eq->lock, flags)) {
1555 		mlx5_ib_eq_pf_process(eq);
1556 		spin_unlock_irqrestore(&eq->lock, flags);
1557 	} else {
1558 		schedule_work(&eq->work);
1559 	}
1560 
1561 	return IRQ_HANDLED;
1562 }
1563 
1564 /* mempool_refill() was proposed but unfortunately wasn't accepted
1565  * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
1566  * Cheap workaround.
1567  */
1568 static void mempool_refill(mempool_t *pool)
1569 {
1570 	while (pool->curr_nr < pool->min_nr)
1571 		mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
1572 }
1573 
1574 static void mlx5_ib_eq_pf_action(struct work_struct *work)
1575 {
1576 	struct mlx5_ib_pf_eq *eq =
1577 		container_of(work, struct mlx5_ib_pf_eq, work);
1578 
1579 	mempool_refill(eq->pool);
1580 
1581 	spin_lock_irq(&eq->lock);
1582 	mlx5_ib_eq_pf_process(eq);
1583 	spin_unlock_irq(&eq->lock);
1584 }
1585 
1586 enum {
1587 	MLX5_IB_NUM_PF_EQE	= 0x1000,
1588 	MLX5_IB_NUM_PF_DRAIN	= 64,
1589 };
1590 
1591 static int
1592 mlx5_ib_create_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1593 {
1594 	struct mlx5_eq_param param = {};
1595 	int err;
1596 
1597 	INIT_WORK(&eq->work, mlx5_ib_eq_pf_action);
1598 	spin_lock_init(&eq->lock);
1599 	eq->dev = dev;
1600 
1601 	eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN,
1602 					       sizeof(struct mlx5_pagefault));
1603 	if (!eq->pool)
1604 		return -ENOMEM;
1605 
1606 	eq->wq = alloc_workqueue("mlx5_ib_page_fault",
1607 				 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM,
1608 				 MLX5_NUM_CMD_EQE);
1609 	if (!eq->wq) {
1610 		err = -ENOMEM;
1611 		goto err_mempool;
1612 	}
1613 
1614 	eq->irq_nb.notifier_call = mlx5_ib_eq_pf_int;
1615 	param = (struct mlx5_eq_param) {
1616 		.irq_index = 0,
1617 		.nent = MLX5_IB_NUM_PF_EQE,
1618 	};
1619 	param.mask[0] = 1ull << MLX5_EVENT_TYPE_PAGE_FAULT;
1620 	eq->core = mlx5_eq_create_generic(dev->mdev, &param);
1621 	if (IS_ERR(eq->core)) {
1622 		err = PTR_ERR(eq->core);
1623 		goto err_wq;
1624 	}
1625 	err = mlx5_eq_enable(dev->mdev, eq->core, &eq->irq_nb);
1626 	if (err) {
1627 		mlx5_ib_err(dev, "failed to enable odp EQ %d\n", err);
1628 		goto err_eq;
1629 	}
1630 
1631 	return 0;
1632 err_eq:
1633 	mlx5_eq_destroy_generic(dev->mdev, eq->core);
1634 err_wq:
1635 	destroy_workqueue(eq->wq);
1636 err_mempool:
1637 	mempool_destroy(eq->pool);
1638 	return err;
1639 }
1640 
1641 static int
1642 mlx5_ib_destroy_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1643 {
1644 	int err;
1645 
1646 	mlx5_eq_disable(dev->mdev, eq->core, &eq->irq_nb);
1647 	err = mlx5_eq_destroy_generic(dev->mdev, eq->core);
1648 	cancel_work_sync(&eq->work);
1649 	destroy_workqueue(eq->wq);
1650 	mempool_destroy(eq->pool);
1651 
1652 	return err;
1653 }
1654 
1655 void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
1656 {
1657 	if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1658 		return;
1659 
1660 	switch (ent->order - 2) {
1661 	case MLX5_IMR_MTT_CACHE_ENTRY:
1662 		ent->page = PAGE_SHIFT;
1663 		ent->xlt = MLX5_IMR_MTT_ENTRIES *
1664 			   sizeof(struct mlx5_mtt) /
1665 			   MLX5_IB_UMR_OCTOWORD;
1666 		ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
1667 		ent->limit = 0;
1668 		break;
1669 
1670 	case MLX5_IMR_KSM_CACHE_ENTRY:
1671 		ent->page = MLX5_KSM_PAGE_SHIFT;
1672 		ent->xlt = mlx5_imr_ksm_entries *
1673 			   sizeof(struct mlx5_klm) /
1674 			   MLX5_IB_UMR_OCTOWORD;
1675 		ent->access_mode = MLX5_MKC_ACCESS_MODE_KSM;
1676 		ent->limit = 0;
1677 		break;
1678 	}
1679 }
1680 
1681 static const struct ib_device_ops mlx5_ib_dev_odp_ops = {
1682 	.advise_mr = mlx5_ib_advise_mr,
1683 };
1684 
1685 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
1686 {
1687 	int ret = 0;
1688 
1689 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1690 		return ret;
1691 
1692 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_odp_ops);
1693 
1694 	if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) {
1695 		ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey);
1696 		if (ret) {
1697 			mlx5_ib_err(dev, "Error getting null_mkey %d\n", ret);
1698 			return ret;
1699 		}
1700 	}
1701 
1702 	ret = mlx5_ib_create_pf_eq(dev, &dev->odp_pf_eq);
1703 
1704 	return ret;
1705 }
1706 
1707 void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev)
1708 {
1709 	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1710 		return;
1711 
1712 	mlx5_ib_destroy_pf_eq(dev, &dev->odp_pf_eq);
1713 }
1714 
1715 int mlx5_ib_odp_init(void)
1716 {
1717 	mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) -
1718 				       MLX5_IMR_MTT_BITS);
1719 
1720 	return 0;
1721 }
1722 
1723 struct prefetch_mr_work {
1724 	struct work_struct work;
1725 	u32 pf_flags;
1726 	u32 num_sge;
1727 	struct {
1728 		u64 io_virt;
1729 		struct mlx5_ib_mr *mr;
1730 		size_t length;
1731 	} frags[];
1732 };
1733 
1734 static void destroy_prefetch_work(struct prefetch_mr_work *work)
1735 {
1736 	u32 i;
1737 
1738 	for (i = 0; i < work->num_sge; ++i)
1739 		if (atomic_dec_and_test(&work->frags[i].mr->num_deferred_work))
1740 			wake_up(&work->frags[i].mr->q_deferred_work);
1741 	kvfree(work);
1742 }
1743 
1744 static struct mlx5_ib_mr *
1745 get_prefetchable_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
1746 		    u32 lkey)
1747 {
1748 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1749 	struct mlx5_core_mkey *mmkey;
1750 	struct ib_umem_odp *odp;
1751 	struct mlx5_ib_mr *mr;
1752 
1753 	lockdep_assert_held(&dev->odp_srcu);
1754 
1755 	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(lkey));
1756 	if (!mmkey || mmkey->key != lkey || mmkey->type != MLX5_MKEY_MR)
1757 		return NULL;
1758 
1759 	mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
1760 
1761 	if (mr->ibmr.pd != pd)
1762 		return NULL;
1763 
1764 	odp = to_ib_umem_odp(mr->umem);
1765 
1766 	/* prefetch with write-access must be supported by the MR */
1767 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
1768 	    !odp->umem.writable)
1769 		return NULL;
1770 
1771 	return mr;
1772 }
1773 
1774 static void mlx5_ib_prefetch_mr_work(struct work_struct *w)
1775 {
1776 	struct prefetch_mr_work *work =
1777 		container_of(w, struct prefetch_mr_work, work);
1778 	struct mlx5_ib_dev *dev;
1779 	u32 bytes_mapped = 0;
1780 	int srcu_key;
1781 	int ret;
1782 	u32 i;
1783 
1784 	/* We rely on IB/core that work is executed if we have num_sge != 0 only. */
1785 	WARN_ON(!work->num_sge);
1786 	dev = work->frags[0].mr->dev;
1787 	/* SRCU should be held when calling to mlx5_odp_populate_xlt() */
1788 	srcu_key = srcu_read_lock(&dev->odp_srcu);
1789 	for (i = 0; i < work->num_sge; ++i) {
1790 		ret = pagefault_mr(work->frags[i].mr, work->frags[i].io_virt,
1791 				   work->frags[i].length, &bytes_mapped,
1792 				   work->pf_flags);
1793 		if (ret <= 0)
1794 			continue;
1795 		mlx5_update_odp_stats(work->frags[i].mr, prefetch, ret);
1796 	}
1797 	srcu_read_unlock(&dev->odp_srcu, srcu_key);
1798 
1799 	destroy_prefetch_work(work);
1800 }
1801 
1802 static bool init_prefetch_work(struct ib_pd *pd,
1803 			       enum ib_uverbs_advise_mr_advice advice,
1804 			       u32 pf_flags, struct prefetch_mr_work *work,
1805 			       struct ib_sge *sg_list, u32 num_sge)
1806 {
1807 	u32 i;
1808 
1809 	INIT_WORK(&work->work, mlx5_ib_prefetch_mr_work);
1810 	work->pf_flags = pf_flags;
1811 
1812 	for (i = 0; i < num_sge; ++i) {
1813 		work->frags[i].io_virt = sg_list[i].addr;
1814 		work->frags[i].length = sg_list[i].length;
1815 		work->frags[i].mr =
1816 			get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1817 		if (!work->frags[i].mr) {
1818 			work->num_sge = i;
1819 			return false;
1820 		}
1821 
1822 		/* Keep the MR pointer will valid outside the SRCU */
1823 		atomic_inc(&work->frags[i].mr->num_deferred_work);
1824 	}
1825 	work->num_sge = num_sge;
1826 	return true;
1827 }
1828 
1829 static int mlx5_ib_prefetch_sg_list(struct ib_pd *pd,
1830 				    enum ib_uverbs_advise_mr_advice advice,
1831 				    u32 pf_flags, struct ib_sge *sg_list,
1832 				    u32 num_sge)
1833 {
1834 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1835 	u32 bytes_mapped = 0;
1836 	int srcu_key;
1837 	int ret = 0;
1838 	u32 i;
1839 
1840 	srcu_key = srcu_read_lock(&dev->odp_srcu);
1841 	for (i = 0; i < num_sge; ++i) {
1842 		struct mlx5_ib_mr *mr;
1843 
1844 		mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1845 		if (!mr) {
1846 			ret = -ENOENT;
1847 			goto out;
1848 		}
1849 		ret = pagefault_mr(mr, sg_list[i].addr, sg_list[i].length,
1850 				   &bytes_mapped, pf_flags);
1851 		if (ret < 0)
1852 			goto out;
1853 		mlx5_update_odp_stats(mr, prefetch, ret);
1854 	}
1855 	ret = 0;
1856 
1857 out:
1858 	srcu_read_unlock(&dev->odp_srcu, srcu_key);
1859 	return ret;
1860 }
1861 
1862 int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
1863 			       enum ib_uverbs_advise_mr_advice advice,
1864 			       u32 flags, struct ib_sge *sg_list, u32 num_sge)
1865 {
1866 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1867 	u32 pf_flags = 0;
1868 	struct prefetch_mr_work *work;
1869 	int srcu_key;
1870 
1871 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH)
1872 		pf_flags |= MLX5_PF_FLAGS_DOWNGRADE;
1873 
1874 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT)
1875 		pf_flags |= MLX5_PF_FLAGS_SNAPSHOT;
1876 
1877 	if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH)
1878 		return mlx5_ib_prefetch_sg_list(pd, advice, pf_flags, sg_list,
1879 						num_sge);
1880 
1881 	work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL);
1882 	if (!work)
1883 		return -ENOMEM;
1884 
1885 	srcu_key = srcu_read_lock(&dev->odp_srcu);
1886 	if (!init_prefetch_work(pd, advice, pf_flags, work, sg_list, num_sge)) {
1887 		srcu_read_unlock(&dev->odp_srcu, srcu_key);
1888 		destroy_prefetch_work(work);
1889 		return -EINVAL;
1890 	}
1891 	queue_work(system_unbound_wq, &work->work);
1892 	srcu_read_unlock(&dev->odp_srcu, srcu_key);
1893 	return 0;
1894 }
1895