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>
38 #define MAX_PREFETCH_LEN (4*1024*1024U)
40 /* Timeout in ms to wait for an active mmu notifier to complete when handling
42 #define MMU_NOTIFIER_TIMEOUT 1000
44 struct workqueue_struct *mlx5_ib_page_fault_wq;
46 void mlx5_ib_invalidate_range(struct ib_umem *umem, unsigned long start,
49 struct mlx5_ib_mr *mr;
50 const u64 umr_block_mask = (MLX5_UMR_MTT_ALIGNMENT / sizeof(u64)) - 1;
51 u64 idx = 0, blk_start_idx = 0;
55 if (!umem || !umem->odp_data) {
56 pr_err("invalidation called on NULL umem or non-ODP umem\n");
60 mr = umem->odp_data->private;
62 if (!mr || !mr->ibmr.pd)
65 start = max_t(u64, ib_umem_start(umem), start);
66 end = min_t(u64, ib_umem_end(umem), end);
69 * Iteration one - zap the HW's MTTs. The notifiers_count ensures that
70 * while we are doing the invalidation, no page fault will attempt to
71 * overwrite the same MTTs. Concurent invalidations might race us,
72 * but they will write 0s as well, so no difference in the end result.
75 for (addr = start; addr < end; addr += (u64)umem->page_size) {
76 idx = (addr - ib_umem_start(umem)) / PAGE_SIZE;
78 * Strive to write the MTTs in chunks, but avoid overwriting
79 * non-existing MTTs. The huristic here can be improved to
80 * estimate the cost of another UMR vs. the cost of bigger
83 if (umem->odp_data->dma_list[idx] &
84 (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) {
90 u64 umr_offset = idx & umr_block_mask;
92 if (in_block && umr_offset == 0) {
93 mlx5_ib_update_mtt(mr, blk_start_idx,
94 idx - blk_start_idx, 1);
100 mlx5_ib_update_mtt(mr, blk_start_idx, idx - blk_start_idx + 1,
104 * We are now sure that the device will not access the
105 * memory. We can safely unmap it, and mark it as dirty if
109 ib_umem_odp_unmap_dma_pages(umem, start, end);
112 void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev)
114 struct ib_odp_caps *caps = &dev->odp_caps;
116 memset(caps, 0, sizeof(*caps));
118 if (!MLX5_CAP_GEN(dev->mdev, pg))
121 caps->general_caps = IB_ODP_SUPPORT;
123 if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.send))
124 caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND;
126 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.send))
127 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND;
129 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.receive))
130 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV;
132 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.write))
133 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE;
135 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.read))
136 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ;
141 static struct mlx5_ib_mr *mlx5_ib_odp_find_mr_lkey(struct mlx5_ib_dev *dev,
144 u32 base_key = mlx5_base_mkey(key);
145 struct mlx5_core_mr *mmr = __mlx5_mr_lookup(dev->mdev, base_key);
146 struct mlx5_ib_mr *mr = container_of(mmr, struct mlx5_ib_mr, mmr);
148 if (!mmr || mmr->key != key || !mr->live)
151 return container_of(mmr, struct mlx5_ib_mr, mmr);
154 static void mlx5_ib_page_fault_resume(struct mlx5_ib_qp *qp,
155 struct mlx5_ib_pfault *pfault,
157 struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.pd->device);
158 int ret = mlx5_core_page_fault_resume(dev->mdev, qp->mqp.qpn,
159 pfault->mpfault.flags,
162 pr_err("Failed to resolve the page fault on QP 0x%x\n",
167 * Handle a single data segment in a page-fault WQE.
169 * Returns number of pages retrieved on success. The caller will continue to
170 * the next data segment.
171 * Can return the following error codes:
172 * -EAGAIN to designate a temporary error. The caller will abort handling the
173 * page fault and resolve it.
174 * -EFAULT when there's an error mapping the requested pages. The caller will
175 * abort the page fault handling and possibly move the QP to an error state.
176 * On other errors the QP should also be closed with an error.
178 static int pagefault_single_data_segment(struct mlx5_ib_qp *qp,
179 struct mlx5_ib_pfault *pfault,
180 u32 key, u64 io_virt, size_t bcnt,
183 struct mlx5_ib_dev *mib_dev = to_mdev(qp->ibqp.pd->device);
185 unsigned int current_seq;
187 int npages = 0, ret = 0;
188 struct mlx5_ib_mr *mr;
189 u64 access_mask = ODP_READ_ALLOWED_BIT;
191 srcu_key = srcu_read_lock(&mib_dev->mr_srcu);
192 mr = mlx5_ib_odp_find_mr_lkey(mib_dev, key);
194 * If we didn't find the MR, it means the MR was closed while we were
195 * handling the ODP event. In this case we return -EFAULT so that the
198 if (!mr || !mr->ibmr.pd) {
199 pr_err("Failed to find relevant mr for lkey=0x%06x, probably the MR was destroyed\n",
204 if (!mr->umem->odp_data) {
205 pr_debug("skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
209 (bcnt - pfault->mpfault.bytes_committed);
212 if (mr->ibmr.pd != qp->ibqp.pd) {
213 pr_err("Page-fault with different PDs for QP and MR.\n");
218 current_seq = ACCESS_ONCE(mr->umem->odp_data->notifiers_seq);
220 * Ensure the sequence number is valid for some time before we call
226 * Avoid branches - this code will perform correctly
227 * in all iterations (in iteration 2 and above,
228 * bytes_committed == 0).
230 io_virt += pfault->mpfault.bytes_committed;
231 bcnt -= pfault->mpfault.bytes_committed;
233 start_idx = (io_virt - (mr->mmr.iova & PAGE_MASK)) >> PAGE_SHIFT;
235 if (mr->umem->writable)
236 access_mask |= ODP_WRITE_ALLOWED_BIT;
237 npages = ib_umem_odp_map_dma_pages(mr->umem, io_virt, bcnt,
238 access_mask, current_seq);
245 mutex_lock(&mr->umem->odp_data->umem_mutex);
246 if (!ib_umem_mmu_notifier_retry(mr->umem, current_seq)) {
248 * No need to check whether the MTTs really belong to
249 * this MR, since ib_umem_odp_map_dma_pages already
252 ret = mlx5_ib_update_mtt(mr, start_idx, npages, 0);
256 mutex_unlock(&mr->umem->odp_data->umem_mutex);
259 pr_err("Failed to update mkey page tables\n");
264 u32 new_mappings = npages * PAGE_SIZE -
265 (io_virt - round_down(io_virt, PAGE_SIZE));
266 *bytes_mapped += min_t(u32, new_mappings, bcnt);
271 if (ret == -EAGAIN) {
272 if (!mr->umem->odp_data->dying) {
273 struct ib_umem_odp *odp_data = mr->umem->odp_data;
274 unsigned long timeout =
275 msecs_to_jiffies(MMU_NOTIFIER_TIMEOUT);
277 if (!wait_for_completion_timeout(
278 &odp_data->notifier_completion,
280 pr_warn("timeout waiting for mmu notifier completion\n");
283 /* The MR is being killed, kill the QP as well. */
287 srcu_read_unlock(&mib_dev->mr_srcu, srcu_key);
288 pfault->mpfault.bytes_committed = 0;
289 return ret ? ret : npages;
293 * Parse a series of data segments for page fault handling.
295 * @qp the QP on which the fault occurred.
296 * @pfault contains page fault information.
297 * @wqe points at the first data segment in the WQE.
298 * @wqe_end points after the end of the WQE.
299 * @bytes_mapped receives the number of bytes that the function was able to
300 * map. This allows the caller to decide intelligently whether
301 * enough memory was mapped to resolve the page fault
302 * successfully (e.g. enough for the next MTU, or the entire
304 * @total_wqe_bytes receives the total data size of this WQE in bytes (minus
305 * the committed bytes).
307 * Returns the number of pages loaded if positive, zero for an empty WQE, or a
308 * negative error code.
310 static int pagefault_data_segments(struct mlx5_ib_qp *qp,
311 struct mlx5_ib_pfault *pfault, void *wqe,
312 void *wqe_end, u32 *bytes_mapped,
313 u32 *total_wqe_bytes, int receive_queue)
315 int ret = 0, npages = 0;
322 /* Skip SRQ next-WQE segment. */
323 if (receive_queue && qp->ibqp.srq)
324 wqe += sizeof(struct mlx5_wqe_srq_next_seg);
329 *total_wqe_bytes = 0;
331 while (wqe < wqe_end) {
332 struct mlx5_wqe_data_seg *dseg = wqe;
334 io_virt = be64_to_cpu(dseg->addr);
335 key = be32_to_cpu(dseg->lkey);
336 byte_count = be32_to_cpu(dseg->byte_count);
337 inline_segment = !!(byte_count & MLX5_INLINE_SEG);
338 bcnt = byte_count & ~MLX5_INLINE_SEG;
340 if (inline_segment) {
341 bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK;
342 wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt,
345 wqe += sizeof(*dseg);
348 /* receive WQE end of sg list. */
349 if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY &&
353 if (!inline_segment && total_wqe_bytes) {
354 *total_wqe_bytes += bcnt - min_t(size_t, bcnt,
355 pfault->mpfault.bytes_committed);
358 /* A zero length data segment designates a length of 2GB. */
362 if (inline_segment || bcnt <= pfault->mpfault.bytes_committed) {
363 pfault->mpfault.bytes_committed -=
365 pfault->mpfault.bytes_committed);
369 ret = pagefault_single_data_segment(qp, pfault, key, io_virt,
376 return ret < 0 ? ret : npages;
380 * Parse initiator WQE. Advances the wqe pointer to point at the
381 * scatter-gather list, and set wqe_end to the end of the WQE.
383 static int mlx5_ib_mr_initiator_pfault_handler(
384 struct mlx5_ib_qp *qp, struct mlx5_ib_pfault *pfault,
385 void **wqe, void **wqe_end, int wqe_length)
387 struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.pd->device);
388 struct mlx5_wqe_ctrl_seg *ctrl = *wqe;
389 u16 wqe_index = pfault->mpfault.wqe.wqe_index;
392 u32 ctrl_wqe_index, ctrl_qpn;
395 ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
396 if (ds * MLX5_WQE_DS_UNITS > wqe_length) {
397 mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n",
403 mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n",
404 wqe_index, qp->mqp.qpn);
409 ctrl_wqe_index = (be32_to_cpu(ctrl->opmod_idx_opcode) &
410 MLX5_WQE_CTRL_WQE_INDEX_MASK) >>
411 MLX5_WQE_CTRL_WQE_INDEX_SHIFT;
412 if (wqe_index != ctrl_wqe_index) {
413 mlx5_ib_err(dev, "Got WQE with invalid wqe_index. wqe_index=0x%x, qpn=0x%x ctrl->wqe_index=0x%x\n",
414 wqe_index, qp->mqp.qpn,
419 ctrl_qpn = (be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_QPN_MASK) >>
420 MLX5_WQE_CTRL_QPN_SHIFT;
421 if (qp->mqp.qpn != ctrl_qpn) {
422 mlx5_ib_err(dev, "Got WQE with incorrect QP number. wqe_index=0x%x, qpn=0x%x ctrl->qpn=0x%x\n",
423 wqe_index, qp->mqp.qpn,
429 *wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS;
430 *wqe += sizeof(*ctrl);
432 opcode = be32_to_cpu(ctrl->opmod_idx_opcode) &
433 MLX5_WQE_CTRL_OPCODE_MASK;
434 switch (qp->ibqp.qp_type) {
437 case MLX5_OPCODE_SEND:
438 case MLX5_OPCODE_SEND_IMM:
439 case MLX5_OPCODE_SEND_INVAL:
440 if (!(dev->odp_caps.per_transport_caps.rc_odp_caps &
441 IB_ODP_SUPPORT_SEND))
442 goto invalid_transport_or_opcode;
444 case MLX5_OPCODE_RDMA_WRITE:
445 case MLX5_OPCODE_RDMA_WRITE_IMM:
446 if (!(dev->odp_caps.per_transport_caps.rc_odp_caps &
447 IB_ODP_SUPPORT_WRITE))
448 goto invalid_transport_or_opcode;
449 *wqe += sizeof(struct mlx5_wqe_raddr_seg);
451 case MLX5_OPCODE_RDMA_READ:
452 if (!(dev->odp_caps.per_transport_caps.rc_odp_caps &
453 IB_ODP_SUPPORT_READ))
454 goto invalid_transport_or_opcode;
455 *wqe += sizeof(struct mlx5_wqe_raddr_seg);
458 goto invalid_transport_or_opcode;
463 case MLX5_OPCODE_SEND:
464 case MLX5_OPCODE_SEND_IMM:
465 if (!(dev->odp_caps.per_transport_caps.ud_odp_caps &
466 IB_ODP_SUPPORT_SEND))
467 goto invalid_transport_or_opcode;
468 *wqe += sizeof(struct mlx5_wqe_datagram_seg);
471 goto invalid_transport_or_opcode;
475 invalid_transport_or_opcode:
476 mlx5_ib_err(dev, "ODP fault on QP of an unsupported opcode or transport. transport: 0x%x opcode: 0x%x.\n",
477 qp->ibqp.qp_type, opcode);
485 * Parse responder WQE. Advances the wqe pointer to point at the
486 * scatter-gather list, and set wqe_end to the end of the WQE.
488 static int mlx5_ib_mr_responder_pfault_handler(
489 struct mlx5_ib_qp *qp, struct mlx5_ib_pfault *pfault,
490 void **wqe, void **wqe_end, int wqe_length)
492 struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.pd->device);
493 struct mlx5_ib_wq *wq = &qp->rq;
494 int wqe_size = 1 << wq->wqe_shift;
497 mlx5_ib_err(dev, "ODP fault on SRQ is not supported\n");
502 mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n");
506 if (wqe_size > wqe_length) {
507 mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
511 switch (qp->ibqp.qp_type) {
513 if (!(dev->odp_caps.per_transport_caps.rc_odp_caps &
514 IB_ODP_SUPPORT_RECV))
515 goto invalid_transport_or_opcode;
518 invalid_transport_or_opcode:
519 mlx5_ib_err(dev, "ODP fault on QP of an unsupported transport. transport: 0x%x\n",
524 *wqe_end = *wqe + wqe_size;
529 static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_qp *qp,
530 struct mlx5_ib_pfault *pfault)
532 struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.pd->device);
535 u32 bytes_mapped, total_wqe_bytes;
537 int resume_with_error = 0;
538 u16 wqe_index = pfault->mpfault.wqe.wqe_index;
539 int requestor = pfault->mpfault.flags & MLX5_PFAULT_REQUESTOR;
541 buffer = (char *)__get_free_page(GFP_KERNEL);
543 mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n");
544 resume_with_error = 1;
545 goto resolve_page_fault;
548 ret = mlx5_ib_read_user_wqe(qp, requestor, wqe_index, buffer,
551 mlx5_ib_err(dev, "Failed reading a WQE following page fault, error=%x, wqe_index=%x, qpn=%x\n",
552 -ret, wqe_index, qp->mqp.qpn);
553 resume_with_error = 1;
554 goto resolve_page_fault;
559 ret = mlx5_ib_mr_initiator_pfault_handler(qp, pfault, &wqe,
562 ret = mlx5_ib_mr_responder_pfault_handler(qp, pfault, &wqe,
565 resume_with_error = 1;
566 goto resolve_page_fault;
569 if (wqe >= wqe_end) {
570 mlx5_ib_err(dev, "ODP fault on invalid WQE.\n");
571 resume_with_error = 1;
572 goto resolve_page_fault;
575 ret = pagefault_data_segments(qp, pfault, wqe, wqe_end, &bytes_mapped,
576 &total_wqe_bytes, !requestor);
577 if (ret == -EAGAIN) {
578 goto resolve_page_fault;
579 } else if (ret < 0 || total_wqe_bytes > bytes_mapped) {
580 mlx5_ib_err(dev, "Error getting user pages for page fault. Error: 0x%x\n",
582 resume_with_error = 1;
583 goto resolve_page_fault;
587 mlx5_ib_page_fault_resume(qp, pfault, resume_with_error);
588 mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, flags: 0x%x\n",
589 qp->mqp.qpn, resume_with_error, pfault->mpfault.flags);
591 free_page((unsigned long)buffer);
594 static int pages_in_range(u64 address, u32 length)
596 return (ALIGN(address + length, PAGE_SIZE) -
597 (address & PAGE_MASK)) >> PAGE_SHIFT;
600 static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_qp *qp,
601 struct mlx5_ib_pfault *pfault)
603 struct mlx5_pagefault *mpfault = &pfault->mpfault;
606 u32 prefetch_len = mpfault->bytes_committed;
607 int prefetch_activated = 0;
608 u32 rkey = mpfault->rdma.r_key;
611 /* The RDMA responder handler handles the page fault in two parts.
612 * First it brings the necessary pages for the current packet
613 * (and uses the pfault context), and then (after resuming the QP)
614 * prefetches more pages. The second operation cannot use the pfault
615 * context and therefore uses the dummy_pfault context allocated on
617 struct mlx5_ib_pfault dummy_pfault = {};
619 dummy_pfault.mpfault.bytes_committed = 0;
621 mpfault->rdma.rdma_va += mpfault->bytes_committed;
622 mpfault->rdma.rdma_op_len -= min(mpfault->bytes_committed,
623 mpfault->rdma.rdma_op_len);
624 mpfault->bytes_committed = 0;
626 address = mpfault->rdma.rdma_va;
627 length = mpfault->rdma.rdma_op_len;
629 /* For some operations, the hardware cannot tell the exact message
630 * length, and in those cases it reports zero. Use prefetch
633 prefetch_activated = 1;
634 length = mpfault->rdma.packet_size;
635 prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len);
638 ret = pagefault_single_data_segment(qp, pfault, rkey, address, length,
640 if (ret == -EAGAIN) {
641 /* We're racing with an invalidation, don't prefetch */
642 prefetch_activated = 0;
643 } else if (ret < 0 || pages_in_range(address, length) > ret) {
644 mlx5_ib_page_fault_resume(qp, pfault, 1);
648 mlx5_ib_page_fault_resume(qp, pfault, 0);
650 /* At this point, there might be a new pagefault already arriving in
651 * the eq, switch to the dummy pagefault for the rest of the
652 * processing. We're still OK with the objects being alive as the
653 * work-queue is being fenced. */
655 if (prefetch_activated) {
656 ret = pagefault_single_data_segment(qp, &dummy_pfault, rkey,
661 pr_warn("Prefetch failed (ret = %d, prefetch_activated = %d) for QPN %d, address: 0x%.16llx, length = 0x%.16x\n",
662 ret, prefetch_activated,
663 qp->ibqp.qp_num, address, prefetch_len);
668 void mlx5_ib_mr_pfault_handler(struct mlx5_ib_qp *qp,
669 struct mlx5_ib_pfault *pfault)
671 u8 event_subtype = pfault->mpfault.event_subtype;
673 switch (event_subtype) {
674 case MLX5_PFAULT_SUBTYPE_WQE:
675 mlx5_ib_mr_wqe_pfault_handler(qp, pfault);
677 case MLX5_PFAULT_SUBTYPE_RDMA:
678 mlx5_ib_mr_rdma_pfault_handler(qp, pfault);
681 pr_warn("Invalid page fault event subtype: 0x%x\n",
683 mlx5_ib_page_fault_resume(qp, pfault, 1);
688 static void mlx5_ib_qp_pfault_action(struct work_struct *work)
690 struct mlx5_ib_pfault *pfault = container_of(work,
691 struct mlx5_ib_pfault,
693 enum mlx5_ib_pagefault_context context =
694 mlx5_ib_get_pagefault_context(&pfault->mpfault);
695 struct mlx5_ib_qp *qp = container_of(pfault, struct mlx5_ib_qp,
696 pagefaults[context]);
697 mlx5_ib_mr_pfault_handler(qp, pfault);
700 void mlx5_ib_qp_disable_pagefaults(struct mlx5_ib_qp *qp)
704 spin_lock_irqsave(&qp->disable_page_faults_lock, flags);
705 qp->disable_page_faults = 1;
706 spin_unlock_irqrestore(&qp->disable_page_faults_lock, flags);
709 * Note that at this point, we are guarenteed that no more
710 * work queue elements will be posted to the work queue with
711 * the QP we are closing.
713 flush_workqueue(mlx5_ib_page_fault_wq);
716 void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp)
720 spin_lock_irqsave(&qp->disable_page_faults_lock, flags);
721 qp->disable_page_faults = 0;
722 spin_unlock_irqrestore(&qp->disable_page_faults_lock, flags);
725 static void mlx5_ib_pfault_handler(struct mlx5_core_qp *qp,
726 struct mlx5_pagefault *pfault)
729 * Note that we will only get one fault event per QP per context
730 * (responder/initiator, read/write), until we resolve the page fault
731 * with the mlx5_ib_page_fault_resume command. Since this function is
732 * called from within the work element, there is no risk of missing
735 struct mlx5_ib_qp *mibqp = to_mibqp(qp);
736 enum mlx5_ib_pagefault_context context =
737 mlx5_ib_get_pagefault_context(pfault);
738 struct mlx5_ib_pfault *qp_pfault = &mibqp->pagefaults[context];
740 qp_pfault->mpfault = *pfault;
742 /* No need to stop interrupts here since we are in an interrupt */
743 spin_lock(&mibqp->disable_page_faults_lock);
744 if (!mibqp->disable_page_faults)
745 queue_work(mlx5_ib_page_fault_wq, &qp_pfault->work);
746 spin_unlock(&mibqp->disable_page_faults_lock);
749 void mlx5_ib_odp_create_qp(struct mlx5_ib_qp *qp)
753 qp->disable_page_faults = 1;
754 spin_lock_init(&qp->disable_page_faults_lock);
756 qp->mqp.pfault_handler = mlx5_ib_pfault_handler;
758 for (i = 0; i < MLX5_IB_PAGEFAULT_CONTEXTS; ++i)
759 INIT_WORK(&qp->pagefaults[i].work, mlx5_ib_qp_pfault_action);
762 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev)
766 ret = init_srcu_struct(&ibdev->mr_srcu);
773 void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *ibdev)
775 cleanup_srcu_struct(&ibdev->mr_srcu);
778 int __init mlx5_ib_odp_init(void)
780 mlx5_ib_page_fault_wq =
781 create_singlethread_workqueue("mlx5_ib_page_faults");
782 if (!mlx5_ib_page_fault_wq)
788 void mlx5_ib_odp_cleanup(void)
790 destroy_workqueue(mlx5_ib_page_fault_wq);