2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
33 #include <rdma/ib_umem.h>
34 #include <rdma/ib_umem_odp.h>
35 #include <linux/kernel.h>
40 #include <linux/mlx5/eq.h>
42 /* Contains the details of a pagefault. */
43 struct mlx5_pagefault {
49 /* Initiator or send message responder pagefault details. */
51 /* Received packet size, only valid for responders. */
54 * Number of resource holding WQE, depends on type.
58 * WQE index. Refers to either the send queue or
59 * receive queue, according to event_subtype.
63 /* RDMA responder pagefault details */
67 * Received packet size, minimal size page fault
68 * resolution required for forward progress.
76 struct mlx5_ib_pf_eq *eq;
77 struct work_struct work;
80 #define MAX_PREFETCH_LEN (4*1024*1024U)
82 /* Timeout in ms to wait for an active mmu notifier to complete when handling
84 #define MMU_NOTIFIER_TIMEOUT 1000
86 #define MLX5_IMR_MTT_BITS (30 - PAGE_SHIFT)
87 #define MLX5_IMR_MTT_SHIFT (MLX5_IMR_MTT_BITS + PAGE_SHIFT)
88 #define MLX5_IMR_MTT_ENTRIES BIT_ULL(MLX5_IMR_MTT_BITS)
89 #define MLX5_IMR_MTT_SIZE BIT_ULL(MLX5_IMR_MTT_SHIFT)
90 #define MLX5_IMR_MTT_MASK (~(MLX5_IMR_MTT_SIZE - 1))
92 #define MLX5_KSM_PAGE_SHIFT MLX5_IMR_MTT_SHIFT
94 static u64 mlx5_imr_ksm_entries;
96 static void populate_klm(struct mlx5_klm *pklm, size_t idx, size_t nentries,
97 struct mlx5_ib_mr *imr, int flags)
99 struct mlx5_klm *end = pklm + nentries;
101 if (flags & MLX5_IB_UPD_XLT_ZAP) {
102 for (; pklm != end; pklm++, idx++) {
103 pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
104 pklm->key = cpu_to_be32(imr->dev->null_mkey);
111 * The locking here is pretty subtle. Ideally the implicit_children
112 * xarray would be protected by the umem_mutex, however that is not
113 * possible. Instead this uses a weaker update-then-lock pattern:
117 * mutex_lock(umem_mutex)
118 * mlx5_ib_update_xlt()
119 * mutex_unlock(umem_mutex)
122 * ie any change the xarray must be followed by the locked update_xlt
125 * The umem_mutex provides the acquire/release semantic needed to make
126 * the xa_store() visible to a racing thread. While SRCU is not
127 * technically required, using it gives consistent use of the SRCU
128 * locking around the xarray.
130 lockdep_assert_held(&to_ib_umem_odp(imr->umem)->umem_mutex);
131 lockdep_assert_held(&imr->dev->odp_srcu);
133 for (; pklm != end; pklm++, idx++) {
134 struct mlx5_ib_mr *mtt = xa_load(&imr->implicit_children, idx);
136 pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
138 pklm->key = cpu_to_be32(mtt->ibmr.lkey);
139 pklm->va = cpu_to_be64(idx * MLX5_IMR_MTT_SIZE);
141 pklm->key = cpu_to_be32(imr->dev->null_mkey);
147 static u64 umem_dma_to_mtt(dma_addr_t umem_dma)
149 u64 mtt_entry = umem_dma & ODP_DMA_ADDR_MASK;
151 if (umem_dma & ODP_READ_ALLOWED_BIT)
152 mtt_entry |= MLX5_IB_MTT_READ;
153 if (umem_dma & ODP_WRITE_ALLOWED_BIT)
154 mtt_entry |= MLX5_IB_MTT_WRITE;
159 static void populate_mtt(__be64 *pas, size_t idx, size_t nentries,
160 struct mlx5_ib_mr *mr, int flags)
162 struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
166 if (flags & MLX5_IB_UPD_XLT_ZAP)
169 for (i = 0; i < nentries; i++) {
170 pa = odp->dma_list[idx + i];
171 pas[i] = cpu_to_be64(umem_dma_to_mtt(pa));
175 void mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
176 struct mlx5_ib_mr *mr, int flags)
178 if (flags & MLX5_IB_UPD_XLT_INDIRECT) {
179 populate_klm(xlt, idx, nentries, mr, flags);
181 populate_mtt(xlt, idx, nentries, mr, flags);
185 static void dma_fence_odp_mr(struct mlx5_ib_mr *mr)
187 struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
189 /* Ensure mlx5_ib_invalidate_range() will not touch the MR any more */
190 mutex_lock(&odp->umem_mutex);
192 mlx5_mr_cache_invalidate(mr);
193 ib_umem_odp_unmap_dma_pages(odp, ib_umem_start(odp),
195 WARN_ON(odp->npages);
198 mutex_unlock(&odp->umem_mutex);
200 if (!mr->allocated_from_cache) {
201 mlx5_core_destroy_mkey(mr->dev->mdev, &mr->mmkey);
207 * This must be called after the mr has been removed from implicit_children
208 * and the SRCU synchronized. NOTE: The MR does not necessarily have to be
209 * empty here, parallel page faults could have raced with the free process and
212 static void free_implicit_child_mr(struct mlx5_ib_mr *mr, bool need_imr_xlt)
214 struct mlx5_ib_mr *imr = mr->parent;
215 struct ib_umem_odp *odp_imr = to_ib_umem_odp(imr->umem);
216 struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
217 unsigned long idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT;
220 /* implicit_child_mr's are not allowed to have deferred work */
221 WARN_ON(atomic_read(&mr->num_deferred_work));
224 srcu_key = srcu_read_lock(&mr->dev->odp_srcu);
225 mutex_lock(&odp_imr->umem_mutex);
226 mlx5_ib_update_xlt(mr->parent, idx, 1, 0,
227 MLX5_IB_UPD_XLT_INDIRECT |
228 MLX5_IB_UPD_XLT_ATOMIC);
229 mutex_unlock(&odp_imr->umem_mutex);
230 srcu_read_unlock(&mr->dev->odp_srcu, srcu_key);
233 dma_fence_odp_mr(mr);
236 mlx5_mr_cache_free(mr->dev, mr);
237 ib_umem_odp_release(odp);
238 atomic_dec(&imr->num_deferred_work);
241 static void free_implicit_child_mr_work(struct work_struct *work)
243 struct mlx5_ib_mr *mr =
244 container_of(work, struct mlx5_ib_mr, odp_destroy.work);
246 free_implicit_child_mr(mr, true);
249 static void free_implicit_child_mr_rcu(struct rcu_head *head)
251 struct mlx5_ib_mr *mr =
252 container_of(head, struct mlx5_ib_mr, odp_destroy.rcu);
254 /* Freeing a MR is a sleeping operation, so bounce to a work queue */
255 INIT_WORK(&mr->odp_destroy.work, free_implicit_child_mr_work);
256 queue_work(system_unbound_wq, &mr->odp_destroy.work);
259 static void destroy_unused_implicit_child_mr(struct mlx5_ib_mr *mr)
261 struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
262 unsigned long idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT;
263 struct mlx5_ib_mr *imr = mr->parent;
265 xa_lock(&imr->implicit_children);
267 * This can race with mlx5_ib_free_implicit_mr(), the first one to
268 * reach the xa lock wins the race and destroys the MR.
270 if (__xa_cmpxchg(&imr->implicit_children, idx, mr, NULL, GFP_ATOMIC) !=
274 atomic_inc(&imr->num_deferred_work);
275 call_srcu(&mr->dev->odp_srcu, &mr->odp_destroy.rcu,
276 free_implicit_child_mr_rcu);
279 xa_unlock(&imr->implicit_children);
282 static bool mlx5_ib_invalidate_range(struct mmu_interval_notifier *mni,
283 const struct mmu_notifier_range *range,
284 unsigned long cur_seq)
286 struct ib_umem_odp *umem_odp =
287 container_of(mni, struct ib_umem_odp, notifier);
288 struct mlx5_ib_mr *mr;
289 const u64 umr_block_mask = (MLX5_UMR_MTT_ALIGNMENT /
290 sizeof(struct mlx5_mtt)) - 1;
291 u64 idx = 0, blk_start_idx = 0;
292 u64 invalidations = 0;
298 if (!mmu_notifier_range_blockable(range))
301 mutex_lock(&umem_odp->umem_mutex);
302 mmu_interval_set_seq(mni, cur_seq);
304 * If npages is zero then umem_odp->private may not be setup yet. This
305 * does not complete until after the first page is mapped for DMA.
307 if (!umem_odp->npages)
309 mr = umem_odp->private;
311 start = max_t(u64, ib_umem_start(umem_odp), range->start);
312 end = min_t(u64, ib_umem_end(umem_odp), range->end);
315 * Iteration one - zap the HW's MTTs. The notifiers_count ensures that
316 * while we are doing the invalidation, no page fault will attempt to
317 * overwrite the same MTTs. Concurent invalidations might race us,
318 * but they will write 0s as well, so no difference in the end result.
320 for (addr = start; addr < end; addr += BIT(umem_odp->page_shift)) {
321 idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
323 * Strive to write the MTTs in chunks, but avoid overwriting
324 * non-existing MTTs. The huristic here can be improved to
325 * estimate the cost of another UMR vs. the cost of bigger
328 if (umem_odp->dma_list[idx] &
329 (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) {
335 /* Count page invalidations */
336 invalidations += idx - blk_start_idx + 1;
338 u64 umr_offset = idx & umr_block_mask;
340 if (in_block && umr_offset == 0) {
341 mlx5_ib_update_xlt(mr, blk_start_idx,
342 idx - blk_start_idx, 0,
343 MLX5_IB_UPD_XLT_ZAP |
344 MLX5_IB_UPD_XLT_ATOMIC);
350 mlx5_ib_update_xlt(mr, blk_start_idx,
351 idx - blk_start_idx + 1, 0,
352 MLX5_IB_UPD_XLT_ZAP |
353 MLX5_IB_UPD_XLT_ATOMIC);
355 mlx5_update_odp_stats(mr, invalidations, invalidations);
358 * We are now sure that the device will not access the
359 * memory. We can safely unmap it, and mark it as dirty if
363 ib_umem_odp_unmap_dma_pages(umem_odp, start, end);
365 if (unlikely(!umem_odp->npages && mr->parent))
366 destroy_unused_implicit_child_mr(mr);
368 mutex_unlock(&umem_odp->umem_mutex);
372 const struct mmu_interval_notifier_ops mlx5_mn_ops = {
373 .invalidate = mlx5_ib_invalidate_range,
376 void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev)
378 struct ib_odp_caps *caps = &dev->odp_caps;
380 memset(caps, 0, sizeof(*caps));
382 if (!MLX5_CAP_GEN(dev->mdev, pg) ||
383 !mlx5_ib_can_use_umr(dev, true, 0))
386 caps->general_caps = IB_ODP_SUPPORT;
388 if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
389 dev->odp_max_size = U64_MAX;
391 dev->odp_max_size = BIT_ULL(MLX5_MAX_UMR_SHIFT + PAGE_SHIFT);
393 if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.send))
394 caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND;
396 if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.srq_receive))
397 caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
399 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.send))
400 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND;
402 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.receive))
403 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV;
405 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.write))
406 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE;
408 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.read))
409 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ;
411 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.atomic))
412 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
414 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.srq_receive))
415 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
417 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.send))
418 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SEND;
420 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.receive))
421 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_RECV;
423 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.write))
424 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_WRITE;
426 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.read))
427 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_READ;
429 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.atomic))
430 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
432 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.srq_receive))
433 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
435 if (MLX5_CAP_GEN(dev->mdev, fixed_buffer_size) &&
436 MLX5_CAP_GEN(dev->mdev, null_mkey) &&
437 MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset) &&
438 !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled))
439 caps->general_caps |= IB_ODP_SUPPORT_IMPLICIT;
442 static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev,
443 struct mlx5_pagefault *pfault,
446 int wq_num = pfault->event_subtype == MLX5_PFAULT_SUBTYPE_WQE ?
447 pfault->wqe.wq_num : pfault->token;
448 u32 out[MLX5_ST_SZ_DW(page_fault_resume_out)] = { };
449 u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)] = { };
452 MLX5_SET(page_fault_resume_in, in, opcode, MLX5_CMD_OP_PAGE_FAULT_RESUME);
453 MLX5_SET(page_fault_resume_in, in, page_fault_type, pfault->type);
454 MLX5_SET(page_fault_resume_in, in, token, pfault->token);
455 MLX5_SET(page_fault_resume_in, in, wq_number, wq_num);
456 MLX5_SET(page_fault_resume_in, in, error, !!error);
458 err = mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
460 mlx5_ib_err(dev, "Failed to resolve the page fault on WQ 0x%x err %d\n",
464 static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,
467 struct ib_umem_odp *odp;
468 struct mlx5_ib_mr *mr;
469 struct mlx5_ib_mr *ret;
472 odp = ib_umem_odp_alloc_child(to_ib_umem_odp(imr->umem),
473 idx * MLX5_IMR_MTT_SIZE,
474 MLX5_IMR_MTT_SIZE, &mlx5_mn_ops);
476 return ERR_CAST(odp);
478 ret = mr = mlx5_mr_cache_alloc(imr->dev, MLX5_IMR_MTT_CACHE_ENTRY);
482 mr->ibmr.pd = imr->ibmr.pd;
483 mr->access_flags = imr->access_flags;
484 mr->umem = &odp->umem;
485 mr->ibmr.lkey = mr->mmkey.key;
486 mr->ibmr.rkey = mr->mmkey.key;
487 mr->mmkey.iova = idx * MLX5_IMR_MTT_SIZE;
491 err = mlx5_ib_update_xlt(mr, 0,
492 MLX5_IMR_MTT_ENTRIES,
494 MLX5_IB_UPD_XLT_ZAP |
495 MLX5_IB_UPD_XLT_ENABLE);
502 * Once the store to either xarray completes any error unwind has to
503 * use synchronize_srcu(). Avoid this with xa_reserve()
505 ret = xa_cmpxchg(&imr->implicit_children, idx, NULL, mr,
508 if (xa_is_err(ret)) {
509 ret = ERR_PTR(xa_err(ret));
513 * Another thread beat us to creating the child mr, use
519 mlx5_ib_dbg(imr->dev, "key %x mr %p\n", mr->mmkey.key, mr);
523 mlx5_mr_cache_free(imr->dev, mr);
525 ib_umem_odp_release(odp);
529 struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
530 struct ib_udata *udata,
533 struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device);
534 struct ib_umem_odp *umem_odp;
535 struct mlx5_ib_mr *imr;
538 umem_odp = ib_umem_odp_alloc_implicit(&dev->ib_dev, access_flags);
539 if (IS_ERR(umem_odp))
540 return ERR_CAST(umem_odp);
542 imr = mlx5_mr_cache_alloc(dev, MLX5_IMR_KSM_CACHE_ENTRY);
548 imr->ibmr.pd = &pd->ibpd;
549 imr->access_flags = access_flags;
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 xa_init(&imr->implicit_children);
559 err = mlx5_ib_update_xlt(imr, 0,
560 mlx5_imr_ksm_entries,
562 MLX5_IB_UPD_XLT_INDIRECT |
563 MLX5_IB_UPD_XLT_ZAP |
564 MLX5_IB_UPD_XLT_ENABLE);
568 err = xa_err(xa_store(&dev->odp_mkeys, mlx5_base_mkey(imr->mmkey.key),
569 &imr->mmkey, GFP_KERNEL));
573 mlx5_ib_dbg(dev, "key %x mr %p\n", imr->mmkey.key, imr);
576 mlx5_ib_err(dev, "Failed to register MKEY %d\n", err);
577 mlx5_mr_cache_free(dev, imr);
579 ib_umem_odp_release(umem_odp);
583 void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *imr)
585 struct ib_umem_odp *odp_imr = to_ib_umem_odp(imr->umem);
586 struct mlx5_ib_dev *dev = imr->dev;
587 struct list_head destroy_list;
588 struct mlx5_ib_mr *mtt;
589 struct mlx5_ib_mr *tmp;
592 INIT_LIST_HEAD(&destroy_list);
594 xa_erase(&dev->odp_mkeys, mlx5_base_mkey(imr->mmkey.key));
596 * This stops the SRCU protected page fault path from touching either
597 * the imr or any children. The page fault path can only reach the
598 * children xarray via the imr.
600 synchronize_srcu(&dev->odp_srcu);
602 xa_lock(&imr->implicit_children);
603 xa_for_each (&imr->implicit_children, idx, mtt) {
604 __xa_erase(&imr->implicit_children, idx);
605 list_add(&mtt->odp_destroy.elm, &destroy_list);
607 xa_unlock(&imr->implicit_children);
610 * num_deferred_work can only be incremented inside the odp_srcu, or
611 * under xa_lock while the child is in the xarray. Thus at this point
612 * it is only decreasing, and all work holding it is now on the wq.
614 if (atomic_read(&imr->num_deferred_work)) {
615 flush_workqueue(system_unbound_wq);
616 WARN_ON(atomic_read(&imr->num_deferred_work));
620 * Fence the imr before we destroy the children. This allows us to
621 * skip updating the XLT of the imr during destroy of the child mkey
624 mlx5_mr_cache_invalidate(imr);
626 list_for_each_entry_safe (mtt, tmp, &destroy_list, odp_destroy.elm)
627 free_implicit_child_mr(mtt, false);
629 mlx5_mr_cache_free(dev, imr);
630 ib_umem_odp_release(odp_imr);
634 * mlx5_ib_fence_odp_mr - Stop all access to the ODP MR
637 * On return no parallel threads will be touching this MR and no DMA will be
640 void mlx5_ib_fence_odp_mr(struct mlx5_ib_mr *mr)
642 /* Prevent new page faults and prefetch requests from succeeding */
643 xa_erase(&mr->dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key));
645 /* Wait for all running page-fault handlers to finish. */
646 synchronize_srcu(&mr->dev->odp_srcu);
648 if (atomic_read(&mr->num_deferred_work)) {
649 flush_workqueue(system_unbound_wq);
650 WARN_ON(atomic_read(&mr->num_deferred_work));
653 dma_fence_odp_mr(mr);
656 #define MLX5_PF_FLAGS_DOWNGRADE BIT(1)
657 static int pagefault_real_mr(struct mlx5_ib_mr *mr, struct ib_umem_odp *odp,
658 u64 user_va, size_t bcnt, u32 *bytes_mapped,
661 int page_shift, ret, np;
662 bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE;
663 unsigned long current_seq;
667 page_shift = odp->page_shift;
668 start_idx = (user_va - ib_umem_start(odp)) >> page_shift;
669 access_mask = ODP_READ_ALLOWED_BIT;
671 if (odp->umem.writable && !downgrade)
672 access_mask |= ODP_WRITE_ALLOWED_BIT;
674 current_seq = mmu_interval_read_begin(&odp->notifier);
676 np = ib_umem_odp_map_dma_pages(odp, user_va, bcnt, access_mask,
681 mutex_lock(&odp->umem_mutex);
682 if (!mmu_interval_read_retry(&odp->notifier, current_seq)) {
684 * No need to check whether the MTTs really belong to
685 * this MR, since ib_umem_odp_map_dma_pages already
688 ret = mlx5_ib_update_xlt(mr, start_idx, np,
689 page_shift, MLX5_IB_UPD_XLT_ATOMIC);
693 mutex_unlock(&odp->umem_mutex);
698 "Failed to update mkey page tables\n");
703 u32 new_mappings = (np << page_shift) -
704 (user_va - round_down(user_va, 1 << page_shift));
706 *bytes_mapped += min_t(u32, new_mappings, bcnt);
709 return np << (page_shift - PAGE_SHIFT);
715 static int pagefault_implicit_mr(struct mlx5_ib_mr *imr,
716 struct ib_umem_odp *odp_imr, u64 user_va,
717 size_t bcnt, u32 *bytes_mapped, u32 flags)
719 unsigned long end_idx = (user_va + bcnt - 1) >> MLX5_IMR_MTT_SHIFT;
720 unsigned long upd_start_idx = end_idx + 1;
721 unsigned long upd_len = 0;
722 unsigned long npages = 0;
726 if (unlikely(user_va >= mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE ||
727 mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE - user_va < bcnt))
730 /* Fault each child mr that intersects with our interval. */
732 unsigned long idx = user_va >> MLX5_IMR_MTT_SHIFT;
733 struct ib_umem_odp *umem_odp;
734 struct mlx5_ib_mr *mtt;
737 mtt = xa_load(&imr->implicit_children, idx);
738 if (unlikely(!mtt)) {
739 mtt = implicit_get_child_mr(imr, idx);
744 upd_start_idx = min(upd_start_idx, idx);
745 upd_len = idx - upd_start_idx + 1;
748 umem_odp = to_ib_umem_odp(mtt->umem);
749 len = min_t(u64, user_va + bcnt, ib_umem_end(umem_odp)) -
752 ret = pagefault_real_mr(mtt, umem_odp, user_va, len,
753 bytes_mapped, flags);
764 * Any time the implicit_children are changed we must perform an
765 * update of the xlt before exiting to ensure the HW and the
766 * implicit_children remains synchronized.
769 if (likely(!upd_len))
773 * Notice this is not strictly ordered right, the KSM is updated after
774 * the implicit_children is updated, so a parallel page fault could
775 * see a MR that is not yet visible in the KSM. This is similar to a
776 * parallel page fault seeing a MR that is being concurrently removed
777 * from the KSM. Both of these improbable situations are resolved
778 * safely by resuming the HW and then taking another page fault. The
779 * next pagefault handler will see the new information.
781 mutex_lock(&odp_imr->umem_mutex);
782 err = mlx5_ib_update_xlt(imr, upd_start_idx, upd_len, 0,
783 MLX5_IB_UPD_XLT_INDIRECT |
784 MLX5_IB_UPD_XLT_ATOMIC);
785 mutex_unlock(&odp_imr->umem_mutex);
787 mlx5_ib_err(imr->dev, "Failed to update PAS\n");
795 * -EFAULT: The io_virt->bcnt is not within the MR, it covers pages that are
796 * not accessible, or the MR is no longer valid.
797 * -EAGAIN/-ENOMEM: The operation should be retried
799 * -EINVAL/others: General internal malfunction
800 * >0: Number of pages mapped
802 static int pagefault_mr(struct mlx5_ib_mr *mr, u64 io_virt, size_t bcnt,
803 u32 *bytes_mapped, u32 flags)
805 struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
807 if (unlikely(io_virt < mr->mmkey.iova))
810 if (!odp->is_implicit_odp) {
813 if (check_add_overflow(io_virt - mr->mmkey.iova,
814 (u64)odp->umem.address, &user_va))
816 if (unlikely(user_va >= ib_umem_end(odp) ||
817 ib_umem_end(odp) - user_va < bcnt))
819 return pagefault_real_mr(mr, odp, user_va, bcnt, bytes_mapped,
822 return pagefault_implicit_mr(mr, odp, io_virt, bcnt, bytes_mapped,
827 struct pf_frame *next;
834 static bool mkey_is_eq(struct mlx5_core_mkey *mmkey, u32 key)
838 if (mmkey->type == MLX5_MKEY_MW)
839 return mlx5_base_mkey(mmkey->key) == mlx5_base_mkey(key);
840 return mmkey->key == key;
843 static int get_indirect_num_descs(struct mlx5_core_mkey *mmkey)
845 struct mlx5_ib_mw *mw;
846 struct mlx5_ib_devx_mr *devx_mr;
848 if (mmkey->type == MLX5_MKEY_MW) {
849 mw = container_of(mmkey, struct mlx5_ib_mw, mmkey);
853 devx_mr = container_of(mmkey, struct mlx5_ib_devx_mr,
855 return devx_mr->ndescs;
859 * Handle a single data segment in a page-fault WQE or RDMA region.
861 * Returns number of OS pages retrieved on success. The caller may continue to
862 * the next data segment.
863 * Can return the following error codes:
864 * -EAGAIN to designate a temporary error. The caller will abort handling the
865 * page fault and resolve it.
866 * -EFAULT when there's an error mapping the requested pages. The caller will
867 * abort the page fault handling.
869 static int pagefault_single_data_segment(struct mlx5_ib_dev *dev,
870 struct ib_pd *pd, u32 key,
871 u64 io_virt, size_t bcnt,
872 u32 *bytes_committed,
875 int npages = 0, srcu_key, ret, i, outlen, cur_outlen = 0, depth = 0;
876 struct pf_frame *head = NULL, *frame;
877 struct mlx5_core_mkey *mmkey;
878 struct mlx5_ib_mr *mr;
879 struct mlx5_klm *pklm;
884 srcu_key = srcu_read_lock(&dev->odp_srcu);
886 io_virt += *bytes_committed;
887 bcnt -= *bytes_committed;
890 mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(key));
894 "skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
897 *bytes_mapped += bcnt;
899 * The user could specify a SGL with multiple lkeys and only
900 * some of them are ODP. Treat the non-ODP ones as fully
906 if (!mkey_is_eq(mmkey, key)) {
907 mlx5_ib_dbg(dev, "failed to find mkey %x\n", key);
912 switch (mmkey->type) {
914 mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
916 ret = pagefault_mr(mr, io_virt, bcnt, bytes_mapped, 0);
921 * When prefetching a page, page fault is generated
922 * in order to bring the page to the main memory.
923 * In the current flow, page faults are being counted.
925 mlx5_update_odp_stats(mr, faults, ret);
932 case MLX5_MKEY_INDIRECT_DEVX:
933 ndescs = get_indirect_num_descs(mmkey);
935 if (depth >= MLX5_CAP_GEN(dev->mdev, max_indirection)) {
936 mlx5_ib_dbg(dev, "indirection level exceeded\n");
941 outlen = MLX5_ST_SZ_BYTES(query_mkey_out) +
942 sizeof(*pklm) * (ndescs - 2);
944 if (outlen > cur_outlen) {
946 out = kzalloc(outlen, GFP_KERNEL);
954 pklm = (struct mlx5_klm *)MLX5_ADDR_OF(query_mkey_out, out,
955 bsf0_klm0_pas_mtt0_1);
957 ret = mlx5_core_query_mkey(dev->mdev, mmkey, out, outlen);
961 offset = io_virt - MLX5_GET64(query_mkey_out, out,
962 memory_key_mkey_entry.start_addr);
964 for (i = 0; bcnt && i < ndescs; i++, pklm++) {
965 if (offset >= be32_to_cpu(pklm->bcount)) {
966 offset -= be32_to_cpu(pklm->bcount);
970 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
976 frame->key = be32_to_cpu(pklm->key);
977 frame->io_virt = be64_to_cpu(pklm->va) + offset;
978 frame->bcnt = min_t(size_t, bcnt,
979 be32_to_cpu(pklm->bcount) - offset);
980 frame->depth = depth + 1;
990 mlx5_ib_dbg(dev, "wrong mkey type %d\n", mmkey->type);
1000 io_virt = frame->io_virt;
1002 depth = frame->depth;
1016 srcu_read_unlock(&dev->odp_srcu, srcu_key);
1017 *bytes_committed = 0;
1018 return ret ? ret : npages;
1022 * Parse a series of data segments for page fault handling.
1024 * @pfault contains page fault information.
1025 * @wqe points at the first data segment in the WQE.
1026 * @wqe_end points after the end of the WQE.
1027 * @bytes_mapped receives the number of bytes that the function was able to
1028 * map. This allows the caller to decide intelligently whether
1029 * enough memory was mapped to resolve the page fault
1030 * successfully (e.g. enough for the next MTU, or the entire
1032 * @total_wqe_bytes receives the total data size of this WQE in bytes (minus
1033 * the committed bytes).
1035 * Returns the number of pages loaded if positive, zero for an empty WQE, or a
1036 * negative error code.
1038 static int pagefault_data_segments(struct mlx5_ib_dev *dev,
1039 struct mlx5_pagefault *pfault,
1041 void *wqe_end, u32 *bytes_mapped,
1042 u32 *total_wqe_bytes, bool receive_queue)
1044 int ret = 0, npages = 0;
1053 if (total_wqe_bytes)
1054 *total_wqe_bytes = 0;
1056 while (wqe < wqe_end) {
1057 struct mlx5_wqe_data_seg *dseg = wqe;
1059 io_virt = be64_to_cpu(dseg->addr);
1060 key = be32_to_cpu(dseg->lkey);
1061 byte_count = be32_to_cpu(dseg->byte_count);
1062 inline_segment = !!(byte_count & MLX5_INLINE_SEG);
1063 bcnt = byte_count & ~MLX5_INLINE_SEG;
1065 if (inline_segment) {
1066 bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK;
1067 wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt,
1070 wqe += sizeof(*dseg);
1073 /* receive WQE end of sg list. */
1074 if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY &&
1078 if (!inline_segment && total_wqe_bytes) {
1079 *total_wqe_bytes += bcnt - min_t(size_t, bcnt,
1080 pfault->bytes_committed);
1083 /* A zero length data segment designates a length of 2GB. */
1087 if (inline_segment || bcnt <= pfault->bytes_committed) {
1088 pfault->bytes_committed -=
1090 pfault->bytes_committed);
1094 ret = pagefault_single_data_segment(dev, NULL, key,
1096 &pfault->bytes_committed,
1103 return ret < 0 ? ret : npages;
1107 * Parse initiator WQE. Advances the wqe pointer to point at the
1108 * scatter-gather list, and set wqe_end to the end of the WQE.
1110 static int mlx5_ib_mr_initiator_pfault_handler(
1111 struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault,
1112 struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length)
1114 struct mlx5_wqe_ctrl_seg *ctrl = *wqe;
1115 u16 wqe_index = pfault->wqe.wqe_index;
1116 struct mlx5_base_av *av;
1117 unsigned ds, opcode;
1118 u32 qpn = qp->trans_qp.base.mqp.qpn;
1120 ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
1121 if (ds * MLX5_WQE_DS_UNITS > wqe_length) {
1122 mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n",
1128 mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n",
1133 *wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS;
1134 *wqe += sizeof(*ctrl);
1136 opcode = be32_to_cpu(ctrl->opmod_idx_opcode) &
1137 MLX5_WQE_CTRL_OPCODE_MASK;
1139 if (qp->ibqp.qp_type == IB_QPT_XRC_INI)
1140 *wqe += sizeof(struct mlx5_wqe_xrc_seg);
1142 if (qp->ibqp.qp_type == IB_QPT_UD ||
1143 qp->qp_sub_type == MLX5_IB_QPT_DCI) {
1145 if (av->dqp_dct & cpu_to_be32(MLX5_EXTENDED_UD_AV))
1146 *wqe += sizeof(struct mlx5_av);
1148 *wqe += sizeof(struct mlx5_base_av);
1152 case MLX5_OPCODE_RDMA_WRITE:
1153 case MLX5_OPCODE_RDMA_WRITE_IMM:
1154 case MLX5_OPCODE_RDMA_READ:
1155 *wqe += sizeof(struct mlx5_wqe_raddr_seg);
1157 case MLX5_OPCODE_ATOMIC_CS:
1158 case MLX5_OPCODE_ATOMIC_FA:
1159 *wqe += sizeof(struct mlx5_wqe_raddr_seg);
1160 *wqe += sizeof(struct mlx5_wqe_atomic_seg);
1168 * Parse responder WQE and set wqe_end to the end of the WQE.
1170 static int mlx5_ib_mr_responder_pfault_handler_srq(struct mlx5_ib_dev *dev,
1171 struct mlx5_ib_srq *srq,
1172 void **wqe, void **wqe_end,
1175 int wqe_size = 1 << srq->msrq.wqe_shift;
1177 if (wqe_size > wqe_length) {
1178 mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1182 *wqe_end = *wqe + wqe_size;
1183 *wqe += sizeof(struct mlx5_wqe_srq_next_seg);
1188 static int mlx5_ib_mr_responder_pfault_handler_rq(struct mlx5_ib_dev *dev,
1189 struct mlx5_ib_qp *qp,
1190 void *wqe, void **wqe_end,
1193 struct mlx5_ib_wq *wq = &qp->rq;
1194 int wqe_size = 1 << wq->wqe_shift;
1197 mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n");
1201 if (wqe_size > wqe_length) {
1202 mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1206 *wqe_end = wqe + wqe_size;
1211 static inline struct mlx5_core_rsc_common *odp_get_rsc(struct mlx5_ib_dev *dev,
1212 u32 wq_num, int pf_type)
1214 struct mlx5_core_rsc_common *common = NULL;
1215 struct mlx5_core_srq *srq;
1218 case MLX5_WQE_PF_TYPE_RMP:
1219 srq = mlx5_cmd_get_srq(dev, wq_num);
1221 common = &srq->common;
1223 case MLX5_WQE_PF_TYPE_REQ_SEND_OR_WRITE:
1224 case MLX5_WQE_PF_TYPE_RESP:
1225 case MLX5_WQE_PF_TYPE_REQ_READ_OR_ATOMIC:
1226 common = mlx5_core_res_hold(dev->mdev, wq_num, MLX5_RES_QP);
1235 static inline struct mlx5_ib_qp *res_to_qp(struct mlx5_core_rsc_common *res)
1237 struct mlx5_core_qp *mqp = (struct mlx5_core_qp *)res;
1239 return to_mibqp(mqp);
1242 static inline struct mlx5_ib_srq *res_to_srq(struct mlx5_core_rsc_common *res)
1244 struct mlx5_core_srq *msrq =
1245 container_of(res, struct mlx5_core_srq, common);
1247 return to_mibsrq(msrq);
1250 static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
1251 struct mlx5_pagefault *pfault)
1253 bool sq = pfault->type & MLX5_PFAULT_REQUESTOR;
1254 u16 wqe_index = pfault->wqe.wqe_index;
1255 void *wqe, *wqe_start = NULL, *wqe_end = NULL;
1256 u32 bytes_mapped, total_wqe_bytes;
1257 struct mlx5_core_rsc_common *res;
1258 int resume_with_error = 1;
1259 struct mlx5_ib_qp *qp;
1260 size_t bytes_copied;
1263 res = odp_get_rsc(dev, pfault->wqe.wq_num, pfault->type);
1265 mlx5_ib_dbg(dev, "wqe page fault for missing resource %d\n", pfault->wqe.wq_num);
1269 if (res->res != MLX5_RES_QP && res->res != MLX5_RES_SRQ &&
1270 res->res != MLX5_RES_XSRQ) {
1271 mlx5_ib_err(dev, "wqe page fault for unsupported type %d\n",
1273 goto resolve_page_fault;
1276 wqe_start = (void *)__get_free_page(GFP_KERNEL);
1278 mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n");
1279 goto resolve_page_fault;
1283 qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL;
1285 ret = mlx5_ib_read_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE,
1289 ret = mlx5_ib_mr_initiator_pfault_handler(
1290 dev, pfault, qp, &wqe, &wqe_end, bytes_copied);
1291 } else if (qp && !sq) {
1292 ret = mlx5_ib_read_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE,
1296 ret = mlx5_ib_mr_responder_pfault_handler_rq(
1297 dev, qp, wqe, &wqe_end, bytes_copied);
1299 struct mlx5_ib_srq *srq = res_to_srq(res);
1301 ret = mlx5_ib_read_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE,
1305 ret = mlx5_ib_mr_responder_pfault_handler_srq(
1306 dev, srq, &wqe, &wqe_end, bytes_copied);
1309 if (ret < 0 || wqe >= wqe_end)
1310 goto resolve_page_fault;
1312 ret = pagefault_data_segments(dev, pfault, wqe, wqe_end, &bytes_mapped,
1313 &total_wqe_bytes, !sq);
1317 if (ret < 0 || total_wqe_bytes > bytes_mapped)
1318 goto resolve_page_fault;
1322 resume_with_error = 0;
1328 "Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n",
1329 ret, wqe_index, pfault->token);
1332 mlx5_ib_page_fault_resume(dev, pfault, resume_with_error);
1333 mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n",
1334 pfault->wqe.wq_num, resume_with_error,
1336 mlx5_core_res_put(res);
1337 free_page((unsigned long)wqe_start);
1340 static int pages_in_range(u64 address, u32 length)
1342 return (ALIGN(address + length, PAGE_SIZE) -
1343 (address & PAGE_MASK)) >> PAGE_SHIFT;
1346 static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev,
1347 struct mlx5_pagefault *pfault)
1351 u32 prefetch_len = pfault->bytes_committed;
1352 int prefetch_activated = 0;
1353 u32 rkey = pfault->rdma.r_key;
1356 /* The RDMA responder handler handles the page fault in two parts.
1357 * First it brings the necessary pages for the current packet
1358 * (and uses the pfault context), and then (after resuming the QP)
1359 * prefetches more pages. The second operation cannot use the pfault
1360 * context and therefore uses the dummy_pfault context allocated on
1362 pfault->rdma.rdma_va += pfault->bytes_committed;
1363 pfault->rdma.rdma_op_len -= min(pfault->bytes_committed,
1364 pfault->rdma.rdma_op_len);
1365 pfault->bytes_committed = 0;
1367 address = pfault->rdma.rdma_va;
1368 length = pfault->rdma.rdma_op_len;
1370 /* For some operations, the hardware cannot tell the exact message
1371 * length, and in those cases it reports zero. Use prefetch
1374 prefetch_activated = 1;
1375 length = pfault->rdma.packet_size;
1376 prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len);
1379 ret = pagefault_single_data_segment(dev, NULL, rkey, address, length,
1380 &pfault->bytes_committed, NULL);
1381 if (ret == -EAGAIN) {
1382 /* We're racing with an invalidation, don't prefetch */
1383 prefetch_activated = 0;
1384 } else if (ret < 0 || pages_in_range(address, length) > ret) {
1385 mlx5_ib_page_fault_resume(dev, pfault, 1);
1387 mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n",
1388 ret, pfault->token, pfault->type);
1392 mlx5_ib_page_fault_resume(dev, pfault, 0);
1393 mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n",
1394 pfault->token, pfault->type,
1395 prefetch_activated);
1397 /* At this point, there might be a new pagefault already arriving in
1398 * the eq, switch to the dummy pagefault for the rest of the
1399 * processing. We're still OK with the objects being alive as the
1400 * work-queue is being fenced. */
1402 if (prefetch_activated) {
1403 u32 bytes_committed = 0;
1405 ret = pagefault_single_data_segment(dev, NULL, rkey, address,
1407 &bytes_committed, NULL);
1408 if (ret < 0 && ret != -EAGAIN) {
1409 mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n",
1410 ret, pfault->token, address, prefetch_len);
1415 static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault)
1417 u8 event_subtype = pfault->event_subtype;
1419 switch (event_subtype) {
1420 case MLX5_PFAULT_SUBTYPE_WQE:
1421 mlx5_ib_mr_wqe_pfault_handler(dev, pfault);
1423 case MLX5_PFAULT_SUBTYPE_RDMA:
1424 mlx5_ib_mr_rdma_pfault_handler(dev, pfault);
1427 mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n",
1429 mlx5_ib_page_fault_resume(dev, pfault, 1);
1433 static void mlx5_ib_eqe_pf_action(struct work_struct *work)
1435 struct mlx5_pagefault *pfault = container_of(work,
1436 struct mlx5_pagefault,
1438 struct mlx5_ib_pf_eq *eq = pfault->eq;
1440 mlx5_ib_pfault(eq->dev, pfault);
1441 mempool_free(pfault, eq->pool);
1444 static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq)
1446 struct mlx5_eqe_page_fault *pf_eqe;
1447 struct mlx5_pagefault *pfault;
1448 struct mlx5_eqe *eqe;
1451 while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) {
1452 pfault = mempool_alloc(eq->pool, GFP_ATOMIC);
1454 schedule_work(&eq->work);
1458 pf_eqe = &eqe->data.page_fault;
1459 pfault->event_subtype = eqe->sub_type;
1460 pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);
1462 mlx5_ib_dbg(eq->dev,
1463 "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
1464 eqe->sub_type, pfault->bytes_committed);
1466 switch (eqe->sub_type) {
1467 case MLX5_PFAULT_SUBTYPE_RDMA:
1468 /* RDMA based event */
1470 be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
1472 be32_to_cpu(pf_eqe->rdma.pftype_token) &
1474 pfault->rdma.r_key =
1475 be32_to_cpu(pf_eqe->rdma.r_key);
1476 pfault->rdma.packet_size =
1477 be16_to_cpu(pf_eqe->rdma.packet_length);
1478 pfault->rdma.rdma_op_len =
1479 be32_to_cpu(pf_eqe->rdma.rdma_op_len);
1480 pfault->rdma.rdma_va =
1481 be64_to_cpu(pf_eqe->rdma.rdma_va);
1482 mlx5_ib_dbg(eq->dev,
1483 "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
1484 pfault->type, pfault->token,
1485 pfault->rdma.r_key);
1486 mlx5_ib_dbg(eq->dev,
1487 "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
1488 pfault->rdma.rdma_op_len,
1489 pfault->rdma.rdma_va);
1492 case MLX5_PFAULT_SUBTYPE_WQE:
1493 /* WQE based event */
1495 (be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7;
1497 be32_to_cpu(pf_eqe->wqe.token);
1498 pfault->wqe.wq_num =
1499 be32_to_cpu(pf_eqe->wqe.pftype_wq) &
1501 pfault->wqe.wqe_index =
1502 be16_to_cpu(pf_eqe->wqe.wqe_index);
1503 pfault->wqe.packet_size =
1504 be16_to_cpu(pf_eqe->wqe.packet_length);
1505 mlx5_ib_dbg(eq->dev,
1506 "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
1507 pfault->type, pfault->token,
1509 pfault->wqe.wqe_index);
1513 mlx5_ib_warn(eq->dev,
1514 "Unsupported page fault event sub-type: 0x%02hhx\n",
1516 /* Unsupported page faults should still be
1517 * resolved by the page fault handler
1522 INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action);
1523 queue_work(eq->wq, &pfault->work);
1525 cc = mlx5_eq_update_cc(eq->core, ++cc);
1528 mlx5_eq_update_ci(eq->core, cc, 1);
1531 static int mlx5_ib_eq_pf_int(struct notifier_block *nb, unsigned long type,
1534 struct mlx5_ib_pf_eq *eq =
1535 container_of(nb, struct mlx5_ib_pf_eq, irq_nb);
1536 unsigned long flags;
1538 if (spin_trylock_irqsave(&eq->lock, flags)) {
1539 mlx5_ib_eq_pf_process(eq);
1540 spin_unlock_irqrestore(&eq->lock, flags);
1542 schedule_work(&eq->work);
1548 /* mempool_refill() was proposed but unfortunately wasn't accepted
1549 * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
1552 static void mempool_refill(mempool_t *pool)
1554 while (pool->curr_nr < pool->min_nr)
1555 mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
1558 static void mlx5_ib_eq_pf_action(struct work_struct *work)
1560 struct mlx5_ib_pf_eq *eq =
1561 container_of(work, struct mlx5_ib_pf_eq, work);
1563 mempool_refill(eq->pool);
1565 spin_lock_irq(&eq->lock);
1566 mlx5_ib_eq_pf_process(eq);
1567 spin_unlock_irq(&eq->lock);
1571 MLX5_IB_NUM_PF_EQE = 0x1000,
1572 MLX5_IB_NUM_PF_DRAIN = 64,
1576 mlx5_ib_create_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1578 struct mlx5_eq_param param = {};
1581 INIT_WORK(&eq->work, mlx5_ib_eq_pf_action);
1582 spin_lock_init(&eq->lock);
1585 eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN,
1586 sizeof(struct mlx5_pagefault));
1590 eq->wq = alloc_workqueue("mlx5_ib_page_fault",
1591 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM,
1598 eq->irq_nb.notifier_call = mlx5_ib_eq_pf_int;
1599 param = (struct mlx5_eq_param) {
1601 .nent = MLX5_IB_NUM_PF_EQE,
1603 param.mask[0] = 1ull << MLX5_EVENT_TYPE_PAGE_FAULT;
1604 eq->core = mlx5_eq_create_generic(dev->mdev, ¶m);
1605 if (IS_ERR(eq->core)) {
1606 err = PTR_ERR(eq->core);
1609 err = mlx5_eq_enable(dev->mdev, eq->core, &eq->irq_nb);
1611 mlx5_ib_err(dev, "failed to enable odp EQ %d\n", err);
1617 mlx5_eq_destroy_generic(dev->mdev, eq->core);
1619 destroy_workqueue(eq->wq);
1621 mempool_destroy(eq->pool);
1626 mlx5_ib_destroy_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1630 mlx5_eq_disable(dev->mdev, eq->core, &eq->irq_nb);
1631 err = mlx5_eq_destroy_generic(dev->mdev, eq->core);
1632 cancel_work_sync(&eq->work);
1633 destroy_workqueue(eq->wq);
1634 mempool_destroy(eq->pool);
1639 void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
1641 if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1644 switch (ent->order - 2) {
1645 case MLX5_IMR_MTT_CACHE_ENTRY:
1646 ent->page = PAGE_SHIFT;
1647 ent->xlt = MLX5_IMR_MTT_ENTRIES *
1648 sizeof(struct mlx5_mtt) /
1649 MLX5_IB_UMR_OCTOWORD;
1650 ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
1654 case MLX5_IMR_KSM_CACHE_ENTRY:
1655 ent->page = MLX5_KSM_PAGE_SHIFT;
1656 ent->xlt = mlx5_imr_ksm_entries *
1657 sizeof(struct mlx5_klm) /
1658 MLX5_IB_UMR_OCTOWORD;
1659 ent->access_mode = MLX5_MKC_ACCESS_MODE_KSM;
1665 static const struct ib_device_ops mlx5_ib_dev_odp_ops = {
1666 .advise_mr = mlx5_ib_advise_mr,
1669 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
1673 if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1676 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_odp_ops);
1678 if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) {
1679 ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey);
1681 mlx5_ib_err(dev, "Error getting null_mkey %d\n", ret);
1686 ret = mlx5_ib_create_pf_eq(dev, &dev->odp_pf_eq);
1691 void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev)
1693 if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1696 mlx5_ib_destroy_pf_eq(dev, &dev->odp_pf_eq);
1699 int mlx5_ib_odp_init(void)
1701 mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) -
1707 struct prefetch_mr_work {
1708 struct work_struct work;
1713 struct mlx5_ib_mr *mr;
1718 static void destroy_prefetch_work(struct prefetch_mr_work *work)
1722 for (i = 0; i < work->num_sge; ++i)
1723 atomic_dec(&work->frags[i].mr->num_deferred_work);
1727 static struct mlx5_ib_mr *
1728 get_prefetchable_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
1731 struct mlx5_ib_dev *dev = to_mdev(pd->device);
1732 struct mlx5_core_mkey *mmkey;
1733 struct ib_umem_odp *odp;
1734 struct mlx5_ib_mr *mr;
1736 lockdep_assert_held(&dev->odp_srcu);
1738 mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(lkey));
1739 if (!mmkey || mmkey->key != lkey || mmkey->type != MLX5_MKEY_MR)
1742 mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
1744 if (mr->ibmr.pd != pd)
1747 odp = to_ib_umem_odp(mr->umem);
1749 /* prefetch with write-access must be supported by the MR */
1750 if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
1751 !odp->umem.writable)
1757 static void mlx5_ib_prefetch_mr_work(struct work_struct *w)
1759 struct prefetch_mr_work *work =
1760 container_of(w, struct prefetch_mr_work, work);
1761 u32 bytes_mapped = 0;
1764 for (i = 0; i < work->num_sge; ++i)
1765 pagefault_mr(work->frags[i].mr, work->frags[i].io_virt,
1766 work->frags[i].length, &bytes_mapped,
1769 destroy_prefetch_work(work);
1772 static bool init_prefetch_work(struct ib_pd *pd,
1773 enum ib_uverbs_advise_mr_advice advice,
1774 u32 pf_flags, struct prefetch_mr_work *work,
1775 struct ib_sge *sg_list, u32 num_sge)
1779 INIT_WORK(&work->work, mlx5_ib_prefetch_mr_work);
1780 work->pf_flags = pf_flags;
1782 for (i = 0; i < num_sge; ++i) {
1783 work->frags[i].io_virt = sg_list[i].addr;
1784 work->frags[i].length = sg_list[i].length;
1786 get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1787 if (!work->frags[i].mr) {
1788 work->num_sge = i - 1;
1790 destroy_prefetch_work(work);
1794 /* Keep the MR pointer will valid outside the SRCU */
1795 atomic_inc(&work->frags[i].mr->num_deferred_work);
1797 work->num_sge = num_sge;
1801 static int mlx5_ib_prefetch_sg_list(struct ib_pd *pd,
1802 enum ib_uverbs_advise_mr_advice advice,
1803 u32 pf_flags, struct ib_sge *sg_list,
1806 struct mlx5_ib_dev *dev = to_mdev(pd->device);
1807 u32 bytes_mapped = 0;
1812 srcu_key = srcu_read_lock(&dev->odp_srcu);
1813 for (i = 0; i < num_sge; ++i) {
1814 struct mlx5_ib_mr *mr;
1816 mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey);
1821 ret = pagefault_mr(mr, sg_list[i].addr, sg_list[i].length,
1822 &bytes_mapped, pf_flags);
1829 srcu_read_unlock(&dev->odp_srcu, srcu_key);
1833 int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
1834 enum ib_uverbs_advise_mr_advice advice,
1835 u32 flags, struct ib_sge *sg_list, u32 num_sge)
1837 struct mlx5_ib_dev *dev = to_mdev(pd->device);
1839 struct prefetch_mr_work *work;
1842 if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH)
1843 pf_flags |= MLX5_PF_FLAGS_DOWNGRADE;
1845 if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH)
1846 return mlx5_ib_prefetch_sg_list(pd, advice, pf_flags, sg_list,
1849 work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL);
1853 srcu_key = srcu_read_lock(&dev->odp_srcu);
1854 if (!init_prefetch_work(pd, advice, pf_flags, work, sg_list, num_sge)) {
1855 srcu_read_unlock(&dev->odp_srcu, srcu_key);
1858 queue_work(system_unbound_wq, &work->work);
1859 srcu_read_unlock(&dev->odp_srcu, srcu_key);