Commit | Line | Data |
---|---|---|
8cdd312c | 1 | /* |
6cf0a15f | 2 | * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved. |
8cdd312c HE |
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 | ||
7bdf65d4 HE |
33 | #include <rdma/ib_umem.h> |
34 | #include <rdma/ib_umem_odp.h> | |
35 | ||
8cdd312c HE |
36 | #include "mlx5_ib.h" |
37 | ||
eab668a6 HE |
38 | #define MAX_PREFETCH_LEN (4*1024*1024U) |
39 | ||
b4cfe447 HE |
40 | /* Timeout in ms to wait for an active mmu notifier to complete when handling |
41 | * a pagefault. */ | |
42 | #define MMU_NOTIFIER_TIMEOUT 1000 | |
43 | ||
6aec21f6 HE |
44 | struct workqueue_struct *mlx5_ib_page_fault_wq; |
45 | ||
b4cfe447 HE |
46 | void mlx5_ib_invalidate_range(struct ib_umem *umem, unsigned long start, |
47 | unsigned long end) | |
48 | { | |
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; | |
52 | int in_block = 0; | |
53 | u64 addr; | |
54 | ||
55 | if (!umem || !umem->odp_data) { | |
56 | pr_err("invalidation called on NULL umem or non-ODP umem\n"); | |
57 | return; | |
58 | } | |
59 | ||
60 | mr = umem->odp_data->private; | |
61 | ||
62 | if (!mr || !mr->ibmr.pd) | |
63 | return; | |
64 | ||
65 | start = max_t(u64, ib_umem_start(umem), start); | |
66 | end = min_t(u64, ib_umem_end(umem), end); | |
67 | ||
68 | /* | |
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. | |
73 | */ | |
74 | ||
75 | for (addr = start; addr < end; addr += (u64)umem->page_size) { | |
76 | idx = (addr - ib_umem_start(umem)) / PAGE_SIZE; | |
77 | /* | |
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 | |
81 | * UMR. | |
82 | */ | |
83 | if (umem->odp_data->dma_list[idx] & | |
84 | (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) { | |
85 | if (!in_block) { | |
86 | blk_start_idx = idx; | |
87 | in_block = 1; | |
88 | } | |
89 | } else { | |
90 | u64 umr_offset = idx & umr_block_mask; | |
91 | ||
92 | if (in_block && umr_offset == 0) { | |
93 | mlx5_ib_update_mtt(mr, blk_start_idx, | |
94 | idx - blk_start_idx, 1); | |
95 | in_block = 0; | |
96 | } | |
97 | } | |
98 | } | |
99 | if (in_block) | |
100 | mlx5_ib_update_mtt(mr, blk_start_idx, idx - blk_start_idx + 1, | |
101 | 1); | |
102 | ||
103 | /* | |
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 | |
106 | * needed. | |
107 | */ | |
108 | ||
109 | ib_umem_odp_unmap_dma_pages(umem, start, end); | |
110 | } | |
111 | ||
938fe83c | 112 | void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev) |
8cdd312c | 113 | { |
8cdd312c HE |
114 | struct ib_odp_caps *caps = &dev->odp_caps; |
115 | ||
116 | memset(caps, 0, sizeof(*caps)); | |
117 | ||
938fe83c SM |
118 | if (!MLX5_CAP_GEN(dev->mdev, pg)) |
119 | return; | |
8cdd312c | 120 | |
b4cfe447 | 121 | caps->general_caps = IB_ODP_SUPPORT; |
938fe83c SM |
122 | |
123 | if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.send)) | |
124 | caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND; | |
125 | ||
126 | if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.send)) | |
127 | caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND; | |
128 | ||
129 | if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.receive)) | |
130 | caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV; | |
131 | ||
132 | if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.write)) | |
133 | caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE; | |
134 | ||
135 | if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.read)) | |
136 | caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ; | |
137 | ||
138 | return; | |
8cdd312c | 139 | } |
6aec21f6 HE |
140 | |
141 | static struct mlx5_ib_mr *mlx5_ib_odp_find_mr_lkey(struct mlx5_ib_dev *dev, | |
142 | u32 key) | |
143 | { | |
144 | u32 base_key = mlx5_base_mkey(key); | |
145 | struct mlx5_core_mr *mmr = __mlx5_mr_lookup(dev->mdev, base_key); | |
b4cfe447 | 146 | struct mlx5_ib_mr *mr = container_of(mmr, struct mlx5_ib_mr, mmr); |
6aec21f6 | 147 | |
b4cfe447 | 148 | if (!mmr || mmr->key != key || !mr->live) |
6aec21f6 HE |
149 | return NULL; |
150 | ||
151 | return container_of(mmr, struct mlx5_ib_mr, mmr); | |
152 | } | |
153 | ||
154 | static void mlx5_ib_page_fault_resume(struct mlx5_ib_qp *qp, | |
155 | struct mlx5_ib_pfault *pfault, | |
156 | int error) { | |
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, | |
160 | error); | |
161 | if (ret) | |
162 | pr_err("Failed to resolve the page fault on QP 0x%x\n", | |
163 | qp->mqp.qpn); | |
164 | } | |
165 | ||
7bdf65d4 HE |
166 | /* |
167 | * Handle a single data segment in a page-fault WQE. | |
168 | * | |
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. | |
177 | */ | |
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, | |
181 | u32 *bytes_mapped) | |
182 | { | |
183 | struct mlx5_ib_dev *mib_dev = to_mdev(qp->ibqp.pd->device); | |
184 | int srcu_key; | |
185 | unsigned int current_seq; | |
186 | u64 start_idx; | |
187 | int npages = 0, ret = 0; | |
188 | struct mlx5_ib_mr *mr; | |
189 | u64 access_mask = ODP_READ_ALLOWED_BIT; | |
190 | ||
191 | srcu_key = srcu_read_lock(&mib_dev->mr_srcu); | |
192 | mr = mlx5_ib_odp_find_mr_lkey(mib_dev, key); | |
193 | /* | |
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 | |
196 | * QP will be closed. | |
197 | */ | |
198 | if (!mr || !mr->ibmr.pd) { | |
199 | pr_err("Failed to find relevant mr for lkey=0x%06x, probably the MR was destroyed\n", | |
200 | key); | |
201 | ret = -EFAULT; | |
202 | goto srcu_unlock; | |
203 | } | |
204 | if (!mr->umem->odp_data) { | |
205 | pr_debug("skipping non ODP MR (lkey=0x%06x) in page fault handler.\n", | |
206 | key); | |
207 | if (bytes_mapped) | |
208 | *bytes_mapped += | |
209 | (bcnt - pfault->mpfault.bytes_committed); | |
210 | goto srcu_unlock; | |
211 | } | |
212 | if (mr->ibmr.pd != qp->ibqp.pd) { | |
213 | pr_err("Page-fault with different PDs for QP and MR.\n"); | |
214 | ret = -EFAULT; | |
215 | goto srcu_unlock; | |
216 | } | |
217 | ||
218 | current_seq = ACCESS_ONCE(mr->umem->odp_data->notifiers_seq); | |
b4cfe447 HE |
219 | /* |
220 | * Ensure the sequence number is valid for some time before we call | |
221 | * gup. | |
222 | */ | |
223 | smp_rmb(); | |
7bdf65d4 HE |
224 | |
225 | /* | |
226 | * Avoid branches - this code will perform correctly | |
227 | * in all iterations (in iteration 2 and above, | |
228 | * bytes_committed == 0). | |
229 | */ | |
230 | io_virt += pfault->mpfault.bytes_committed; | |
231 | bcnt -= pfault->mpfault.bytes_committed; | |
232 | ||
233 | start_idx = (io_virt - (mr->mmr.iova & PAGE_MASK)) >> PAGE_SHIFT; | |
234 | ||
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); | |
239 | if (npages < 0) { | |
240 | ret = npages; | |
241 | goto srcu_unlock; | |
242 | } | |
243 | ||
244 | if (npages > 0) { | |
245 | mutex_lock(&mr->umem->odp_data->umem_mutex); | |
b4cfe447 HE |
246 | if (!ib_umem_mmu_notifier_retry(mr->umem, current_seq)) { |
247 | /* | |
248 | * No need to check whether the MTTs really belong to | |
249 | * this MR, since ib_umem_odp_map_dma_pages already | |
250 | * checks this. | |
251 | */ | |
252 | ret = mlx5_ib_update_mtt(mr, start_idx, npages, 0); | |
253 | } else { | |
254 | ret = -EAGAIN; | |
255 | } | |
7bdf65d4 HE |
256 | mutex_unlock(&mr->umem->odp_data->umem_mutex); |
257 | if (ret < 0) { | |
b4cfe447 HE |
258 | if (ret != -EAGAIN) |
259 | pr_err("Failed to update mkey page tables\n"); | |
7bdf65d4 HE |
260 | goto srcu_unlock; |
261 | } | |
262 | ||
263 | if (bytes_mapped) { | |
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); | |
267 | } | |
268 | } | |
269 | ||
270 | srcu_unlock: | |
b4cfe447 HE |
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); | |
276 | ||
277 | if (!wait_for_completion_timeout( | |
278 | &odp_data->notifier_completion, | |
279 | timeout)) { | |
280 | pr_warn("timeout waiting for mmu notifier completion\n"); | |
281 | } | |
282 | } else { | |
283 | /* The MR is being killed, kill the QP as well. */ | |
284 | ret = -EFAULT; | |
285 | } | |
286 | } | |
7bdf65d4 HE |
287 | srcu_read_unlock(&mib_dev->mr_srcu, srcu_key); |
288 | pfault->mpfault.bytes_committed = 0; | |
289 | return ret ? ret : npages; | |
290 | } | |
291 | ||
292 | /** | |
293 | * Parse a series of data segments for page fault handling. | |
294 | * | |
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 | |
303 | * WQE). | |
304 | * @total_wqe_bytes receives the total data size of this WQE in bytes (minus | |
305 | * the committed bytes). | |
306 | * | |
307 | * Returns the number of pages loaded if positive, zero for an empty WQE, or a | |
308 | * negative error code. | |
309 | */ | |
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) | |
314 | { | |
315 | int ret = 0, npages = 0; | |
316 | u64 io_virt; | |
317 | u32 key; | |
318 | u32 byte_count; | |
319 | size_t bcnt; | |
320 | int inline_segment; | |
321 | ||
322 | /* Skip SRQ next-WQE segment. */ | |
323 | if (receive_queue && qp->ibqp.srq) | |
324 | wqe += sizeof(struct mlx5_wqe_srq_next_seg); | |
325 | ||
326 | if (bytes_mapped) | |
327 | *bytes_mapped = 0; | |
328 | if (total_wqe_bytes) | |
329 | *total_wqe_bytes = 0; | |
330 | ||
331 | while (wqe < wqe_end) { | |
332 | struct mlx5_wqe_data_seg *dseg = wqe; | |
333 | ||
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; | |
339 | ||
340 | if (inline_segment) { | |
341 | bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK; | |
342 | wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt, | |
343 | 16); | |
344 | } else { | |
345 | wqe += sizeof(*dseg); | |
346 | } | |
347 | ||
348 | /* receive WQE end of sg list. */ | |
349 | if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY && | |
350 | io_virt == 0) | |
351 | break; | |
352 | ||
353 | if (!inline_segment && total_wqe_bytes) { | |
354 | *total_wqe_bytes += bcnt - min_t(size_t, bcnt, | |
355 | pfault->mpfault.bytes_committed); | |
356 | } | |
357 | ||
358 | /* A zero length data segment designates a length of 2GB. */ | |
359 | if (bcnt == 0) | |
360 | bcnt = 1U << 31; | |
361 | ||
362 | if (inline_segment || bcnt <= pfault->mpfault.bytes_committed) { | |
363 | pfault->mpfault.bytes_committed -= | |
364 | min_t(size_t, bcnt, | |
365 | pfault->mpfault.bytes_committed); | |
366 | continue; | |
367 | } | |
368 | ||
369 | ret = pagefault_single_data_segment(qp, pfault, key, io_virt, | |
370 | bcnt, bytes_mapped); | |
371 | if (ret < 0) | |
372 | break; | |
373 | npages += ret; | |
374 | } | |
375 | ||
376 | return ret < 0 ? ret : npages; | |
377 | } | |
378 | ||
379 | /* | |
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. | |
382 | */ | |
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) | |
386 | { | |
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; | |
390 | unsigned ds, opcode; | |
391 | #if defined(DEBUG) | |
392 | u32 ctrl_wqe_index, ctrl_qpn; | |
393 | #endif | |
394 | ||
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", | |
398 | ds, wqe_length); | |
399 | return -EFAULT; | |
400 | } | |
401 | ||
402 | if (ds == 0) { | |
403 | mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n", | |
404 | wqe_index, qp->mqp.qpn); | |
405 | return -EFAULT; | |
406 | } | |
407 | ||
408 | #if defined(DEBUG) | |
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, | |
415 | ctrl_wqe_index); | |
416 | return -EFAULT; | |
417 | } | |
418 | ||
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, | |
424 | ctrl_qpn); | |
425 | return -EFAULT; | |
426 | } | |
427 | #endif /* DEBUG */ | |
428 | ||
429 | *wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS; | |
430 | *wqe += sizeof(*ctrl); | |
431 | ||
432 | opcode = be32_to_cpu(ctrl->opmod_idx_opcode) & | |
433 | MLX5_WQE_CTRL_OPCODE_MASK; | |
434 | switch (qp->ibqp.qp_type) { | |
435 | case IB_QPT_RC: | |
436 | switch (opcode) { | |
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; | |
443 | break; | |
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); | |
450 | break; | |
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); | |
456 | break; | |
457 | default: | |
458 | goto invalid_transport_or_opcode; | |
459 | } | |
460 | break; | |
461 | case IB_QPT_UD: | |
462 | switch (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); | |
469 | break; | |
470 | default: | |
471 | goto invalid_transport_or_opcode; | |
472 | } | |
473 | break; | |
474 | default: | |
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); | |
478 | return -EFAULT; | |
479 | } | |
480 | ||
481 | return 0; | |
482 | } | |
483 | ||
484 | /* | |
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. | |
487 | */ | |
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) | |
491 | { | |
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; | |
495 | ||
496 | if (qp->ibqp.srq) { | |
497 | mlx5_ib_err(dev, "ODP fault on SRQ is not supported\n"); | |
498 | return -EFAULT; | |
499 | } | |
500 | ||
501 | if (qp->wq_sig) { | |
502 | mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n"); | |
503 | return -EFAULT; | |
504 | } | |
505 | ||
506 | if (wqe_size > wqe_length) { | |
507 | mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n"); | |
508 | return -EFAULT; | |
509 | } | |
510 | ||
511 | switch (qp->ibqp.qp_type) { | |
512 | case IB_QPT_RC: | |
513 | if (!(dev->odp_caps.per_transport_caps.rc_odp_caps & | |
514 | IB_ODP_SUPPORT_RECV)) | |
515 | goto invalid_transport_or_opcode; | |
516 | break; | |
517 | default: | |
518 | invalid_transport_or_opcode: | |
519 | mlx5_ib_err(dev, "ODP fault on QP of an unsupported transport. transport: 0x%x\n", | |
520 | qp->ibqp.qp_type); | |
521 | return -EFAULT; | |
522 | } | |
523 | ||
524 | *wqe_end = *wqe + wqe_size; | |
525 | ||
526 | return 0; | |
527 | } | |
528 | ||
529 | static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_qp *qp, | |
530 | struct mlx5_ib_pfault *pfault) | |
531 | { | |
532 | struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.pd->device); | |
533 | int ret; | |
534 | void *wqe, *wqe_end; | |
535 | u32 bytes_mapped, total_wqe_bytes; | |
536 | char *buffer = NULL; | |
537 | int resume_with_error = 0; | |
538 | u16 wqe_index = pfault->mpfault.wqe.wqe_index; | |
539 | int requestor = pfault->mpfault.flags & MLX5_PFAULT_REQUESTOR; | |
540 | ||
541 | buffer = (char *)__get_free_page(GFP_KERNEL); | |
542 | if (!buffer) { | |
543 | mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n"); | |
544 | resume_with_error = 1; | |
545 | goto resolve_page_fault; | |
546 | } | |
547 | ||
548 | ret = mlx5_ib_read_user_wqe(qp, requestor, wqe_index, buffer, | |
549 | PAGE_SIZE); | |
550 | if (ret < 0) { | |
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; | |
555 | } | |
556 | ||
557 | wqe = buffer; | |
558 | if (requestor) | |
559 | ret = mlx5_ib_mr_initiator_pfault_handler(qp, pfault, &wqe, | |
560 | &wqe_end, ret); | |
561 | else | |
562 | ret = mlx5_ib_mr_responder_pfault_handler(qp, pfault, &wqe, | |
563 | &wqe_end, ret); | |
564 | if (ret < 0) { | |
565 | resume_with_error = 1; | |
566 | goto resolve_page_fault; | |
567 | } | |
568 | ||
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; | |
573 | } | |
574 | ||
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", | |
581 | -ret); | |
582 | resume_with_error = 1; | |
583 | goto resolve_page_fault; | |
584 | } | |
585 | ||
586 | 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); | |
590 | ||
591 | free_page((unsigned long)buffer); | |
592 | } | |
593 | ||
eab668a6 HE |
594 | static int pages_in_range(u64 address, u32 length) |
595 | { | |
596 | return (ALIGN(address + length, PAGE_SIZE) - | |
597 | (address & PAGE_MASK)) >> PAGE_SHIFT; | |
598 | } | |
599 | ||
600 | static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_qp *qp, | |
601 | struct mlx5_ib_pfault *pfault) | |
602 | { | |
603 | struct mlx5_pagefault *mpfault = &pfault->mpfault; | |
604 | u64 address; | |
605 | u32 length; | |
606 | u32 prefetch_len = mpfault->bytes_committed; | |
607 | int prefetch_activated = 0; | |
608 | u32 rkey = mpfault->rdma.r_key; | |
609 | int ret; | |
610 | ||
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 | |
616 | * the stack */ | |
617 | struct mlx5_ib_pfault dummy_pfault = {}; | |
618 | ||
619 | dummy_pfault.mpfault.bytes_committed = 0; | |
620 | ||
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; | |
625 | ||
626 | address = mpfault->rdma.rdma_va; | |
627 | length = mpfault->rdma.rdma_op_len; | |
628 | ||
629 | /* For some operations, the hardware cannot tell the exact message | |
630 | * length, and in those cases it reports zero. Use prefetch | |
631 | * logic. */ | |
632 | if (length == 0) { | |
633 | prefetch_activated = 1; | |
634 | length = mpfault->rdma.packet_size; | |
635 | prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len); | |
636 | } | |
637 | ||
638 | ret = pagefault_single_data_segment(qp, pfault, rkey, address, length, | |
639 | NULL); | |
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); | |
645 | return; | |
646 | } | |
647 | ||
648 | mlx5_ib_page_fault_resume(qp, pfault, 0); | |
649 | ||
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. */ | |
654 | ||
655 | if (prefetch_activated) { | |
656 | ret = pagefault_single_data_segment(qp, &dummy_pfault, rkey, | |
657 | address, | |
658 | prefetch_len, | |
659 | NULL); | |
660 | if (ret < 0) { | |
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); | |
664 | } | |
665 | } | |
666 | } | |
667 | ||
6aec21f6 HE |
668 | void mlx5_ib_mr_pfault_handler(struct mlx5_ib_qp *qp, |
669 | struct mlx5_ib_pfault *pfault) | |
670 | { | |
671 | u8 event_subtype = pfault->mpfault.event_subtype; | |
672 | ||
673 | switch (event_subtype) { | |
7bdf65d4 HE |
674 | case MLX5_PFAULT_SUBTYPE_WQE: |
675 | mlx5_ib_mr_wqe_pfault_handler(qp, pfault); | |
676 | break; | |
eab668a6 HE |
677 | case MLX5_PFAULT_SUBTYPE_RDMA: |
678 | mlx5_ib_mr_rdma_pfault_handler(qp, pfault); | |
679 | break; | |
6aec21f6 HE |
680 | default: |
681 | pr_warn("Invalid page fault event subtype: 0x%x\n", | |
682 | event_subtype); | |
683 | mlx5_ib_page_fault_resume(qp, pfault, 1); | |
684 | break; | |
685 | } | |
686 | } | |
687 | ||
688 | static void mlx5_ib_qp_pfault_action(struct work_struct *work) | |
689 | { | |
690 | struct mlx5_ib_pfault *pfault = container_of(work, | |
691 | struct mlx5_ib_pfault, | |
692 | work); | |
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); | |
698 | } | |
699 | ||
700 | void mlx5_ib_qp_disable_pagefaults(struct mlx5_ib_qp *qp) | |
701 | { | |
702 | unsigned long flags; | |
703 | ||
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); | |
707 | ||
708 | /* | |
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. | |
712 | */ | |
713 | flush_workqueue(mlx5_ib_page_fault_wq); | |
714 | } | |
715 | ||
716 | void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp) | |
717 | { | |
718 | unsigned long flags; | |
719 | ||
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); | |
723 | } | |
724 | ||
725 | static void mlx5_ib_pfault_handler(struct mlx5_core_qp *qp, | |
726 | struct mlx5_pagefault *pfault) | |
727 | { | |
728 | /* | |
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 | |
733 | * events. | |
734 | */ | |
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]; | |
739 | ||
740 | qp_pfault->mpfault = *pfault; | |
741 | ||
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); | |
747 | } | |
748 | ||
749 | void mlx5_ib_odp_create_qp(struct mlx5_ib_qp *qp) | |
750 | { | |
751 | int i; | |
752 | ||
753 | qp->disable_page_faults = 1; | |
754 | spin_lock_init(&qp->disable_page_faults_lock); | |
755 | ||
756 | qp->mqp.pfault_handler = mlx5_ib_pfault_handler; | |
757 | ||
758 | for (i = 0; i < MLX5_IB_PAGEFAULT_CONTEXTS; ++i) | |
759 | INIT_WORK(&qp->pagefaults[i].work, mlx5_ib_qp_pfault_action); | |
760 | } | |
761 | ||
762 | int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev) | |
763 | { | |
764 | int ret; | |
765 | ||
766 | ret = init_srcu_struct(&ibdev->mr_srcu); | |
767 | if (ret) | |
768 | return ret; | |
769 | ||
770 | return 0; | |
771 | } | |
772 | ||
773 | void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *ibdev) | |
774 | { | |
775 | cleanup_srcu_struct(&ibdev->mr_srcu); | |
776 | } | |
777 | ||
778 | int __init mlx5_ib_odp_init(void) | |
779 | { | |
780 | mlx5_ib_page_fault_wq = | |
781 | create_singlethread_workqueue("mlx5_ib_page_faults"); | |
782 | if (!mlx5_ib_page_fault_wq) | |
783 | return -ENOMEM; | |
784 | ||
785 | return 0; | |
786 | } | |
787 | ||
788 | void mlx5_ib_odp_cleanup(void) | |
789 | { | |
790 | destroy_workqueue(mlx5_ib_page_fault_wq); | |
791 | } |