]> git.proxmox.com Git - mirror_ubuntu-focal-kernel.git/blob - drivers/infiniband/hw/mlx5/odp.c
projects / mirror_ubuntu-focal-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branches 'acpi-apei', 'acpi-processor', 'acpi-tables', 'acpi-pci' and 'acpi...
[mirror_ubuntu-focal-kernel.git] / drivers / infiniband / hw / mlx5 / odp.c
1 /*
2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 */
32
33 #include <rdma/ib_umem.h>
34 #include <rdma/ib_umem_odp.h>
35 #include <linux/kernel.h>
36
37 #include "mlx5_ib.h"
38 #include "cmd.h"
39
40 #include <linux/mlx5/eq.h>
41
42 /* Contains the details of a pagefault. */
43 struct mlx5_pagefault {
44 u32 bytes_committed;
45 u32 token;
46 u8 event_subtype;
47 u8 type;
48 union {
49 /* Initiator or send message responder pagefault details. */
50 struct {
51 /* Received packet size, only valid for responders. */
52 u32 packet_size;
53 /*
54 * Number of resource holding WQE, depends on type.
55 */
56 u32 wq_num;
57 /*
58 * WQE index. Refers to either the send queue or
59 * receive queue, according to event_subtype.
60 */
61 u16 wqe_index;
62 } wqe;
63 /* RDMA responder pagefault details */
64 struct {
65 u32 r_key;
66 /*
67 * Received packet size, minimal size page fault
68 * resolution required for forward progress.
69 */
70 u32 packet_size;
71 u32 rdma_op_len;
72 u64 rdma_va;
73 } rdma;
74 };
75
76 struct mlx5_ib_pf_eq *eq;
77 struct work_struct work;
78 };
79
80 #define MAX_PREFETCH_LEN (4*1024*1024U)
81
82 /* Timeout in ms to wait for an active mmu notifier to complete when handling
83 * a pagefault. */
84 #define MMU_NOTIFIER_TIMEOUT 1000
85
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))
91
92 #define MLX5_KSM_PAGE_SHIFT MLX5_IMR_MTT_SHIFT
93
94 static u64 mlx5_imr_ksm_entries;
95
96 static int check_parent(struct ib_umem_odp *odp,
97 struct mlx5_ib_mr *parent)
98 {
99 struct mlx5_ib_mr *mr = odp->private;
100
101 return mr && mr->parent == parent && !odp->dying;
102 }
103
104 static struct ib_ucontext_per_mm *mr_to_per_mm(struct mlx5_ib_mr *mr)
105 {
106 if (WARN_ON(!mr || !is_odp_mr(mr)))
107 return NULL;
108
109 return to_ib_umem_odp(mr->umem)->per_mm;
110 }
111
112 static struct ib_umem_odp *odp_next(struct ib_umem_odp *odp)
113 {
114 struct mlx5_ib_mr *mr = odp->private, *parent = mr->parent;
115 struct ib_ucontext_per_mm *per_mm = odp->per_mm;
116 struct rb_node *rb;
117
118 down_read(&per_mm->umem_rwsem);
119 while (1) {
120 rb = rb_next(&odp->interval_tree.rb);
121 if (!rb)
122 goto not_found;
123 odp = rb_entry(rb, struct ib_umem_odp, interval_tree.rb);
124 if (check_parent(odp, parent))
125 goto end;
126 }
127 not_found:
128 odp = NULL;
129 end:
130 up_read(&per_mm->umem_rwsem);
131 return odp;
132 }
133
134 static struct ib_umem_odp *odp_lookup(u64 start, u64 length,
135 struct mlx5_ib_mr *parent)
136 {
137 struct ib_ucontext_per_mm *per_mm = mr_to_per_mm(parent);
138 struct ib_umem_odp *odp;
139 struct rb_node *rb;
140
141 down_read(&per_mm->umem_rwsem);
142 odp = rbt_ib_umem_lookup(&per_mm->umem_tree, start, length);
143 if (!odp)
144 goto end;
145
146 while (1) {
147 if (check_parent(odp, parent))
148 goto end;
149 rb = rb_next(&odp->interval_tree.rb);
150 if (!rb)
151 goto not_found;
152 odp = rb_entry(rb, struct ib_umem_odp, interval_tree.rb);
153 if (ib_umem_start(odp) > start + length)
154 goto not_found;
155 }
156 not_found:
157 odp = NULL;
158 end:
159 up_read(&per_mm->umem_rwsem);
160 return odp;
161 }
162
163 void mlx5_odp_populate_klm(struct mlx5_klm *pklm, size_t offset,
164 size_t nentries, struct mlx5_ib_mr *mr, int flags)
165 {
166 struct ib_pd *pd = mr->ibmr.pd;
167 struct mlx5_ib_dev *dev = to_mdev(pd->device);
168 struct ib_umem_odp *odp;
169 unsigned long va;
170 int i;
171
172 if (flags & MLX5_IB_UPD_XLT_ZAP) {
173 for (i = 0; i < nentries; i++, pklm++) {
174 pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
175 pklm->key = cpu_to_be32(dev->null_mkey);
176 pklm->va = 0;
177 }
178 return;
179 }
180
181 odp = odp_lookup(offset * MLX5_IMR_MTT_SIZE,
182 nentries * MLX5_IMR_MTT_SIZE, mr);
183
184 for (i = 0; i < nentries; i++, pklm++) {
185 pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
186 va = (offset + i) * MLX5_IMR_MTT_SIZE;
187 if (odp && odp->umem.address == va) {
188 struct mlx5_ib_mr *mtt = odp->private;
189
190 pklm->key = cpu_to_be32(mtt->ibmr.lkey);
191 odp = odp_next(odp);
192 } else {
193 pklm->key = cpu_to_be32(dev->null_mkey);
194 }
195 mlx5_ib_dbg(dev, "[%d] va %lx key %x\n",
196 i, va, be32_to_cpu(pklm->key));
197 }
198 }
199
200 static void mr_leaf_free_action(struct work_struct *work)
201 {
202 struct ib_umem_odp *odp = container_of(work, struct ib_umem_odp, work);
203 int idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT;
204 struct mlx5_ib_mr *mr = odp->private, *imr = mr->parent;
205
206 mr->parent = NULL;
207 synchronize_srcu(&mr->dev->mr_srcu);
208
209 ib_umem_release(&odp->umem);
210 if (imr->live)
211 mlx5_ib_update_xlt(imr, idx, 1, 0,
212 MLX5_IB_UPD_XLT_INDIRECT |
213 MLX5_IB_UPD_XLT_ATOMIC);
214 mlx5_mr_cache_free(mr->dev, mr);
215
216 if (atomic_dec_and_test(&imr->num_leaf_free))
217 wake_up(&imr->q_leaf_free);
218 }
219
220 void mlx5_ib_invalidate_range(struct ib_umem_odp *umem_odp, unsigned long start,
221 unsigned long end)
222 {
223 struct mlx5_ib_mr *mr;
224 const u64 umr_block_mask = (MLX5_UMR_MTT_ALIGNMENT /
225 sizeof(struct mlx5_mtt)) - 1;
226 u64 idx = 0, blk_start_idx = 0;
227 int in_block = 0;
228 u64 addr;
229
230 if (!umem_odp) {
231 pr_err("invalidation called on NULL umem or non-ODP umem\n");
232 return;
233 }
234
235 mr = umem_odp->private;
236
237 if (!mr || !mr->ibmr.pd)
238 return;
239
240 start = max_t(u64, ib_umem_start(umem_odp), start);
241 end = min_t(u64, ib_umem_end(umem_odp), end);
242
243 /*
244 * Iteration one - zap the HW's MTTs. The notifiers_count ensures that
245 * while we are doing the invalidation, no page fault will attempt to
246 * overwrite the same MTTs. Concurent invalidations might race us,
247 * but they will write 0s as well, so no difference in the end result.
248 */
249 mutex_lock(&umem_odp->umem_mutex);
250 for (addr = start; addr < end; addr += BIT(umem_odp->page_shift)) {
251 idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
252 /*
253 * Strive to write the MTTs in chunks, but avoid overwriting
254 * non-existing MTTs. The huristic here can be improved to
255 * estimate the cost of another UMR vs. the cost of bigger
256 * UMR.
257 */
258 if (umem_odp->dma_list[idx] &
259 (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) {
260 if (!in_block) {
261 blk_start_idx = idx;
262 in_block = 1;
263 }
264 } else {
265 u64 umr_offset = idx & umr_block_mask;
266
267 if (in_block && umr_offset == 0) {
268 mlx5_ib_update_xlt(mr, blk_start_idx,
269 idx - blk_start_idx, 0,
270 MLX5_IB_UPD_XLT_ZAP |
271 MLX5_IB_UPD_XLT_ATOMIC);
272 in_block = 0;
273 }
274 }
275 }
276 if (in_block)
277 mlx5_ib_update_xlt(mr, blk_start_idx,
278 idx - blk_start_idx + 1, 0,
279 MLX5_IB_UPD_XLT_ZAP |
280 MLX5_IB_UPD_XLT_ATOMIC);
281 mutex_unlock(&umem_odp->umem_mutex);
282 /*
283 * We are now sure that the device will not access the
284 * memory. We can safely unmap it, and mark it as dirty if
285 * needed.
286 */
287
288 ib_umem_odp_unmap_dma_pages(umem_odp, start, end);
289
290 if (unlikely(!umem_odp->npages && mr->parent &&
291 !umem_odp->dying)) {
292 WRITE_ONCE(umem_odp->dying, 1);
293 atomic_inc(&mr->parent->num_leaf_free);
294 schedule_work(&umem_odp->work);
295 }
296 }
297
298 void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev)
299 {
300 struct ib_odp_caps *caps = &dev->odp_caps;
301
302 memset(caps, 0, sizeof(*caps));
303
304 if (!MLX5_CAP_GEN(dev->mdev, pg) ||
305 !mlx5_ib_can_use_umr(dev, true))
306 return;
307
308 caps->general_caps = IB_ODP_SUPPORT;
309
310 if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
311 dev->odp_max_size = U64_MAX;
312 else
313 dev->odp_max_size = BIT_ULL(MLX5_MAX_UMR_SHIFT + PAGE_SHIFT);
314
315 if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.send))
316 caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND;
317
318 if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.srq_receive))
319 caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
320
321 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.send))
322 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND;
323
324 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.receive))
325 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV;
326
327 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.write))
328 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE;
329
330 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.read))
331 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ;
332
333 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.atomic))
334 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
335
336 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.srq_receive))
337 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
338
339 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.send))
340 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SEND;
341
342 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.receive))
343 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_RECV;
344
345 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.write))
346 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_WRITE;
347
348 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.read))
349 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_READ;
350
351 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.atomic))
352 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
353
354 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.srq_receive))
355 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
356
357 if (MLX5_CAP_GEN(dev->mdev, fixed_buffer_size) &&
358 MLX5_CAP_GEN(dev->mdev, null_mkey) &&
359 MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset) &&
360 !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled))
361 caps->general_caps |= IB_ODP_SUPPORT_IMPLICIT;
362
363 return;
364 }
365
366 static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev,
367 struct mlx5_pagefault *pfault,
368 int error)
369 {
370 int wq_num = pfault->event_subtype == MLX5_PFAULT_SUBTYPE_WQE ?
371 pfault->wqe.wq_num : pfault->token;
372 u32 out[MLX5_ST_SZ_DW(page_fault_resume_out)] = { };
373 u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)] = { };
374 int err;
375
376 MLX5_SET(page_fault_resume_in, in, opcode, MLX5_CMD_OP_PAGE_FAULT_RESUME);
377 MLX5_SET(page_fault_resume_in, in, page_fault_type, pfault->type);
378 MLX5_SET(page_fault_resume_in, in, token, pfault->token);
379 MLX5_SET(page_fault_resume_in, in, wq_number, wq_num);
380 MLX5_SET(page_fault_resume_in, in, error, !!error);
381
382 err = mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
383 if (err)
384 mlx5_ib_err(dev, "Failed to resolve the page fault on WQ 0x%x err %d\n",
385 wq_num, err);
386 }
387
388 static struct mlx5_ib_mr *implicit_mr_alloc(struct ib_pd *pd,
389 struct ib_umem *umem,
390 bool ksm, int access_flags)
391 {
392 struct mlx5_ib_dev *dev = to_mdev(pd->device);
393 struct mlx5_ib_mr *mr;
394 int err;
395
396 mr = mlx5_mr_cache_alloc(dev, ksm ? MLX5_IMR_KSM_CACHE_ENTRY :
397 MLX5_IMR_MTT_CACHE_ENTRY);
398
399 if (IS_ERR(mr))
400 return mr;
401
402 mr->ibmr.pd = pd;
403
404 mr->dev = dev;
405 mr->access_flags = access_flags;
406 mr->mmkey.iova = 0;
407 mr->umem = umem;
408
409 if (ksm) {
410 err = mlx5_ib_update_xlt(mr, 0,
411 mlx5_imr_ksm_entries,
412 MLX5_KSM_PAGE_SHIFT,
413 MLX5_IB_UPD_XLT_INDIRECT |
414 MLX5_IB_UPD_XLT_ZAP |
415 MLX5_IB_UPD_XLT_ENABLE);
416
417 } else {
418 err = mlx5_ib_update_xlt(mr, 0,
419 MLX5_IMR_MTT_ENTRIES,
420 PAGE_SHIFT,
421 MLX5_IB_UPD_XLT_ZAP |
422 MLX5_IB_UPD_XLT_ENABLE |
423 MLX5_IB_UPD_XLT_ATOMIC);
424 }
425
426 if (err)
427 goto fail;
428
429 mr->ibmr.lkey = mr->mmkey.key;
430 mr->ibmr.rkey = mr->mmkey.key;
431
432 mr->live = 1;
433
434 mlx5_ib_dbg(dev, "key %x dev %p mr %p\n",
435 mr->mmkey.key, dev->mdev, mr);
436
437 return mr;
438
439 fail:
440 mlx5_ib_err(dev, "Failed to register MKEY %d\n", err);
441 mlx5_mr_cache_free(dev, mr);
442
443 return ERR_PTR(err);
444 }
445
446 static struct ib_umem_odp *implicit_mr_get_data(struct mlx5_ib_mr *mr,
447 u64 io_virt, size_t bcnt)
448 {
449 struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.pd->device);
450 struct ib_umem_odp *odp, *result = NULL;
451 struct ib_umem_odp *odp_mr = to_ib_umem_odp(mr->umem);
452 u64 addr = io_virt & MLX5_IMR_MTT_MASK;
453 int nentries = 0, start_idx = 0, ret;
454 struct mlx5_ib_mr *mtt;
455
456 mutex_lock(&odp_mr->umem_mutex);
457 odp = odp_lookup(addr, 1, mr);
458
459 mlx5_ib_dbg(dev, "io_virt:%llx bcnt:%zx addr:%llx odp:%p\n",
460 io_virt, bcnt, addr, odp);
461
462 next_mr:
463 if (likely(odp)) {
464 if (nentries)
465 nentries++;
466 } else {
467 odp = ib_alloc_odp_umem(odp_mr, addr,
468 MLX5_IMR_MTT_SIZE);
469 if (IS_ERR(odp)) {
470 mutex_unlock(&odp_mr->umem_mutex);
471 return ERR_CAST(odp);
472 }
473
474 mtt = implicit_mr_alloc(mr->ibmr.pd, &odp->umem, 0,
475 mr->access_flags);
476 if (IS_ERR(mtt)) {
477 mutex_unlock(&odp_mr->umem_mutex);
478 ib_umem_release(&odp->umem);
479 return ERR_CAST(mtt);
480 }
481
482 odp->private = mtt;
483 mtt->umem = &odp->umem;
484 mtt->mmkey.iova = addr;
485 mtt->parent = mr;
486 INIT_WORK(&odp->work, mr_leaf_free_action);
487
488 if (!nentries)
489 start_idx = addr >> MLX5_IMR_MTT_SHIFT;
490 nentries++;
491 }
492
493 /* Return first odp if region not covered by single one */
494 if (likely(!result))
495 result = odp;
496
497 addr += MLX5_IMR_MTT_SIZE;
498 if (unlikely(addr < io_virt + bcnt)) {
499 odp = odp_next(odp);
500 if (odp && odp->umem.address != addr)
501 odp = NULL;
502 goto next_mr;
503 }
504
505 if (unlikely(nentries)) {
506 ret = mlx5_ib_update_xlt(mr, start_idx, nentries, 0,
507 MLX5_IB_UPD_XLT_INDIRECT |
508 MLX5_IB_UPD_XLT_ATOMIC);
509 if (ret) {
510 mlx5_ib_err(dev, "Failed to update PAS\n");
511 result = ERR_PTR(ret);
512 }
513 }
514
515 mutex_unlock(&odp_mr->umem_mutex);
516 return result;
517 }
518
519 struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
520 struct ib_udata *udata,
521 int access_flags)
522 {
523 struct mlx5_ib_mr *imr;
524 struct ib_umem *umem;
525
526 umem = ib_umem_get(udata, 0, 0, access_flags, 0);
527 if (IS_ERR(umem))
528 return ERR_CAST(umem);
529
530 imr = implicit_mr_alloc(&pd->ibpd, umem, 1, access_flags);
531 if (IS_ERR(imr)) {
532 ib_umem_release(umem);
533 return ERR_CAST(imr);
534 }
535
536 imr->umem = umem;
537 init_waitqueue_head(&imr->q_leaf_free);
538 atomic_set(&imr->num_leaf_free, 0);
539 atomic_set(&imr->num_pending_prefetch, 0);
540
541 return imr;
542 }
543
544 static int mr_leaf_free(struct ib_umem_odp *umem_odp, u64 start, u64 end,
545 void *cookie)
546 {
547 struct mlx5_ib_mr *mr = umem_odp->private, *imr = cookie;
548
549 if (mr->parent != imr)
550 return 0;
551
552 ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
553 ib_umem_end(umem_odp));
554
555 if (umem_odp->dying)
556 return 0;
557
558 WRITE_ONCE(umem_odp->dying, 1);
559 atomic_inc(&imr->num_leaf_free);
560 schedule_work(&umem_odp->work);
561
562 return 0;
563 }
564
565 void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *imr)
566 {
567 struct ib_ucontext_per_mm *per_mm = mr_to_per_mm(imr);
568
569 down_read(&per_mm->umem_rwsem);
570 rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, 0, ULLONG_MAX,
571 mr_leaf_free, true, imr);
572 up_read(&per_mm->umem_rwsem);
573
574 wait_event(imr->q_leaf_free, !atomic_read(&imr->num_leaf_free));
575 }
576
577 #define MLX5_PF_FLAGS_PREFETCH BIT(0)
578 #define MLX5_PF_FLAGS_DOWNGRADE BIT(1)
579 static int pagefault_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr,
580 u64 io_virt, size_t bcnt, u32 *bytes_mapped,
581 u32 flags)
582 {
583 int npages = 0, current_seq, page_shift, ret, np;
584 struct ib_umem_odp *odp_mr = to_ib_umem_odp(mr->umem);
585 bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE;
586 bool prefetch = flags & MLX5_PF_FLAGS_PREFETCH;
587 u64 access_mask;
588 u64 start_idx, page_mask;
589 struct ib_umem_odp *odp;
590 size_t size;
591
592 if (!odp_mr->page_list) {
593 odp = implicit_mr_get_data(mr, io_virt, bcnt);
594
595 if (IS_ERR(odp))
596 return PTR_ERR(odp);
597 mr = odp->private;
598 } else {
599 odp = odp_mr;
600 }
601
602 next_mr:
603 size = min_t(size_t, bcnt, ib_umem_end(odp) - io_virt);
604
605 page_shift = odp->page_shift;
606 page_mask = ~(BIT(page_shift) - 1);
607 start_idx = (io_virt - (mr->mmkey.iova & page_mask)) >> page_shift;
608 access_mask = ODP_READ_ALLOWED_BIT;
609
610 if (prefetch && !downgrade && !mr->umem->writable) {
611 /* prefetch with write-access must
612 * be supported by the MR
613 */
614 ret = -EINVAL;
615 goto out;
616 }
617
618 if (mr->umem->writable && !downgrade)
619 access_mask |= ODP_WRITE_ALLOWED_BIT;
620
621 current_seq = READ_ONCE(odp->notifiers_seq);
622 /*
623 * Ensure the sequence number is valid for some time before we call
624 * gup.
625 */
626 smp_rmb();
627
628 ret = ib_umem_odp_map_dma_pages(to_ib_umem_odp(mr->umem), io_virt, size,
629 access_mask, current_seq);
630
631 if (ret < 0)
632 goto out;
633
634 np = ret;
635
636 mutex_lock(&odp->umem_mutex);
637 if (!ib_umem_mmu_notifier_retry(to_ib_umem_odp(mr->umem),
638 current_seq)) {
639 /*
640 * No need to check whether the MTTs really belong to
641 * this MR, since ib_umem_odp_map_dma_pages already
642 * checks this.
643 */
644 ret = mlx5_ib_update_xlt(mr, start_idx, np,
645 page_shift, MLX5_IB_UPD_XLT_ATOMIC);
646 } else {
647 ret = -EAGAIN;
648 }
649 mutex_unlock(&odp->umem_mutex);
650
651 if (ret < 0) {
652 if (ret != -EAGAIN)
653 mlx5_ib_err(dev, "Failed to update mkey page tables\n");
654 goto out;
655 }
656
657 if (bytes_mapped) {
658 u32 new_mappings = (np << page_shift) -
659 (io_virt - round_down(io_virt, 1 << page_shift));
660 *bytes_mapped += min_t(u32, new_mappings, size);
661 }
662
663 npages += np << (page_shift - PAGE_SHIFT);
664 bcnt -= size;
665
666 if (unlikely(bcnt)) {
667 struct ib_umem_odp *next;
668
669 io_virt += size;
670 next = odp_next(odp);
671 if (unlikely(!next || next->umem.address != io_virt)) {
672 mlx5_ib_dbg(dev, "next implicit leaf removed at 0x%llx. got %p\n",
673 io_virt, next);
674 return -EAGAIN;
675 }
676 odp = next;
677 mr = odp->private;
678 goto next_mr;
679 }
680
681 return npages;
682
683 out:
684 if (ret == -EAGAIN) {
685 unsigned long timeout = msecs_to_jiffies(MMU_NOTIFIER_TIMEOUT);
686
687 if (!wait_for_completion_timeout(&odp->notifier_completion,
688 timeout)) {
689 mlx5_ib_warn(
690 dev,
691 "timeout waiting for mmu notifier. seq %d against %d. notifiers_count=%d\n",
692 current_seq, odp->notifiers_seq,
693 odp->notifiers_count);
694 }
695 }
696
697 return ret;
698 }
699
700 struct pf_frame {
701 struct pf_frame *next;
702 u32 key;
703 u64 io_virt;
704 size_t bcnt;
705 int depth;
706 };
707
708 static bool mkey_is_eq(struct mlx5_core_mkey *mmkey, u32 key)
709 {
710 if (!mmkey)
711 return false;
712 if (mmkey->type == MLX5_MKEY_MW)
713 return mlx5_base_mkey(mmkey->key) == mlx5_base_mkey(key);
714 return mmkey->key == key;
715 }
716
717 static int get_indirect_num_descs(struct mlx5_core_mkey *mmkey)
718 {
719 struct mlx5_ib_mw *mw;
720 struct mlx5_ib_devx_mr *devx_mr;
721
722 if (mmkey->type == MLX5_MKEY_MW) {
723 mw = container_of(mmkey, struct mlx5_ib_mw, mmkey);
724 return mw->ndescs;
725 }
726
727 devx_mr = container_of(mmkey, struct mlx5_ib_devx_mr,
728 mmkey);
729 return devx_mr->ndescs;
730 }
731
732 /*
733 * Handle a single data segment in a page-fault WQE or RDMA region.
734 *
735 * Returns number of OS pages retrieved on success. The caller may continue to
736 * the next data segment.
737 * Can return the following error codes:
738 * -EAGAIN to designate a temporary error. The caller will abort handling the
739 * page fault and resolve it.
740 * -EFAULT when there's an error mapping the requested pages. The caller will
741 * abort the page fault handling.
742 */
743 static int pagefault_single_data_segment(struct mlx5_ib_dev *dev,
744 struct ib_pd *pd, u32 key,
745 u64 io_virt, size_t bcnt,
746 u32 *bytes_committed,
747 u32 *bytes_mapped, u32 flags)
748 {
749 int npages = 0, srcu_key, ret, i, outlen, cur_outlen = 0, depth = 0;
750 bool prefetch = flags & MLX5_PF_FLAGS_PREFETCH;
751 struct pf_frame *head = NULL, *frame;
752 struct mlx5_core_mkey *mmkey;
753 struct mlx5_ib_mr *mr;
754 struct mlx5_klm *pklm;
755 u32 *out = NULL;
756 size_t offset;
757 int ndescs;
758
759 srcu_key = srcu_read_lock(&dev->mr_srcu);
760
761 io_virt += *bytes_committed;
762 bcnt -= *bytes_committed;
763
764 next_mr:
765 mmkey = xa_load(&dev->mdev->priv.mkey_table, mlx5_base_mkey(key));
766 if (!mkey_is_eq(mmkey, key)) {
767 mlx5_ib_dbg(dev, "failed to find mkey %x\n", key);
768 ret = -EFAULT;
769 goto srcu_unlock;
770 }
771
772 if (prefetch && mmkey->type != MLX5_MKEY_MR) {
773 mlx5_ib_dbg(dev, "prefetch is allowed only for MR\n");
774 ret = -EINVAL;
775 goto srcu_unlock;
776 }
777
778 switch (mmkey->type) {
779 case MLX5_MKEY_MR:
780 mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
781 if (!mr->live || !mr->ibmr.pd) {
782 mlx5_ib_dbg(dev, "got dead MR\n");
783 ret = -EFAULT;
784 goto srcu_unlock;
785 }
786
787 if (prefetch) {
788 if (!is_odp_mr(mr) ||
789 mr->ibmr.pd != pd) {
790 mlx5_ib_dbg(dev, "Invalid prefetch request: %s\n",
791 is_odp_mr(mr) ? "MR is not ODP" :
792 "PD is not of the MR");
793 ret = -EINVAL;
794 goto srcu_unlock;
795 }
796 }
797
798 if (!is_odp_mr(mr)) {
799 mlx5_ib_dbg(dev, "skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
800 key);
801 if (bytes_mapped)
802 *bytes_mapped += bcnt;
803 ret = 0;
804 goto srcu_unlock;
805 }
806
807 ret = pagefault_mr(dev, mr, io_virt, bcnt, bytes_mapped, flags);
808 if (ret < 0)
809 goto srcu_unlock;
810
811 npages += ret;
812 ret = 0;
813 break;
814
815 case MLX5_MKEY_MW:
816 case MLX5_MKEY_INDIRECT_DEVX:
817 ndescs = get_indirect_num_descs(mmkey);
818
819 if (depth >= MLX5_CAP_GEN(dev->mdev, max_indirection)) {
820 mlx5_ib_dbg(dev, "indirection level exceeded\n");
821 ret = -EFAULT;
822 goto srcu_unlock;
823 }
824
825 outlen = MLX5_ST_SZ_BYTES(query_mkey_out) +
826 sizeof(*pklm) * (ndescs - 2);
827
828 if (outlen > cur_outlen) {
829 kfree(out);
830 out = kzalloc(outlen, GFP_KERNEL);
831 if (!out) {
832 ret = -ENOMEM;
833 goto srcu_unlock;
834 }
835 cur_outlen = outlen;
836 }
837
838 pklm = (struct mlx5_klm *)MLX5_ADDR_OF(query_mkey_out, out,
839 bsf0_klm0_pas_mtt0_1);
840
841 ret = mlx5_core_query_mkey(dev->mdev, mmkey, out, outlen);
842 if (ret)
843 goto srcu_unlock;
844
845 offset = io_virt - MLX5_GET64(query_mkey_out, out,
846 memory_key_mkey_entry.start_addr);
847
848 for (i = 0; bcnt && i < ndescs; i++, pklm++) {
849 if (offset >= be32_to_cpu(pklm->bcount)) {
850 offset -= be32_to_cpu(pklm->bcount);
851 continue;
852 }
853
854 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
855 if (!frame) {
856 ret = -ENOMEM;
857 goto srcu_unlock;
858 }
859
860 frame->key = be32_to_cpu(pklm->key);
861 frame->io_virt = be64_to_cpu(pklm->va) + offset;
862 frame->bcnt = min_t(size_t, bcnt,
863 be32_to_cpu(pklm->bcount) - offset);
864 frame->depth = depth + 1;
865 frame->next = head;
866 head = frame;
867
868 bcnt -= frame->bcnt;
869 offset = 0;
870 }
871 break;
872
873 default:
874 mlx5_ib_dbg(dev, "wrong mkey type %d\n", mmkey->type);
875 ret = -EFAULT;
876 goto srcu_unlock;
877 }
878
879 if (head) {
880 frame = head;
881 head = frame->next;
882
883 key = frame->key;
884 io_virt = frame->io_virt;
885 bcnt = frame->bcnt;
886 depth = frame->depth;
887 kfree(frame);
888
889 goto next_mr;
890 }
891
892 srcu_unlock:
893 while (head) {
894 frame = head;
895 head = frame->next;
896 kfree(frame);
897 }
898 kfree(out);
899
900 srcu_read_unlock(&dev->mr_srcu, srcu_key);
901 *bytes_committed = 0;
902 return ret ? ret : npages;
903 }
904
905 /**
906 * Parse a series of data segments for page fault handling.
907 *
908 * @pfault contains page fault information.
909 * @wqe points at the first data segment in the WQE.
910 * @wqe_end points after the end of the WQE.
911 * @bytes_mapped receives the number of bytes that the function was able to
912 * map. This allows the caller to decide intelligently whether
913 * enough memory was mapped to resolve the page fault
914 * successfully (e.g. enough for the next MTU, or the entire
915 * WQE).
916 * @total_wqe_bytes receives the total data size of this WQE in bytes (minus
917 * the committed bytes).
918 *
919 * Returns the number of pages loaded if positive, zero for an empty WQE, or a
920 * negative error code.
921 */
922 static int pagefault_data_segments(struct mlx5_ib_dev *dev,
923 struct mlx5_pagefault *pfault,
924 void *wqe,
925 void *wqe_end, u32 *bytes_mapped,
926 u32 *total_wqe_bytes, bool receive_queue)
927 {
928 int ret = 0, npages = 0;
929 u64 io_virt;
930 u32 key;
931 u32 byte_count;
932 size_t bcnt;
933 int inline_segment;
934
935 if (bytes_mapped)
936 *bytes_mapped = 0;
937 if (total_wqe_bytes)
938 *total_wqe_bytes = 0;
939
940 while (wqe < wqe_end) {
941 struct mlx5_wqe_data_seg *dseg = wqe;
942
943 io_virt = be64_to_cpu(dseg->addr);
944 key = be32_to_cpu(dseg->lkey);
945 byte_count = be32_to_cpu(dseg->byte_count);
946 inline_segment = !!(byte_count & MLX5_INLINE_SEG);
947 bcnt = byte_count & ~MLX5_INLINE_SEG;
948
949 if (inline_segment) {
950 bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK;
951 wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt,
952 16);
953 } else {
954 wqe += sizeof(*dseg);
955 }
956
957 /* receive WQE end of sg list. */
958 if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY &&
959 io_virt == 0)
960 break;
961
962 if (!inline_segment && total_wqe_bytes) {
963 *total_wqe_bytes += bcnt - min_t(size_t, bcnt,
964 pfault->bytes_committed);
965 }
966
967 /* A zero length data segment designates a length of 2GB. */
968 if (bcnt == 0)
969 bcnt = 1U << 31;
970
971 if (inline_segment || bcnt <= pfault->bytes_committed) {
972 pfault->bytes_committed -=
973 min_t(size_t, bcnt,
974 pfault->bytes_committed);
975 continue;
976 }
977
978 ret = pagefault_single_data_segment(dev, NULL, key,
979 io_virt, bcnt,
980 &pfault->bytes_committed,
981 bytes_mapped, 0);
982 if (ret < 0)
983 break;
984 npages += ret;
985 }
986
987 return ret < 0 ? ret : npages;
988 }
989
990 static const u32 mlx5_ib_odp_opcode_cap[] = {
991 [MLX5_OPCODE_SEND] = IB_ODP_SUPPORT_SEND,
992 [MLX5_OPCODE_SEND_IMM] = IB_ODP_SUPPORT_SEND,
993 [MLX5_OPCODE_SEND_INVAL] = IB_ODP_SUPPORT_SEND,
994 [MLX5_OPCODE_RDMA_WRITE] = IB_ODP_SUPPORT_WRITE,
995 [MLX5_OPCODE_RDMA_WRITE_IMM] = IB_ODP_SUPPORT_WRITE,
996 [MLX5_OPCODE_RDMA_READ] = IB_ODP_SUPPORT_READ,
997 [MLX5_OPCODE_ATOMIC_CS] = IB_ODP_SUPPORT_ATOMIC,
998 [MLX5_OPCODE_ATOMIC_FA] = IB_ODP_SUPPORT_ATOMIC,
999 };
1000
1001 /*
1002 * Parse initiator WQE. Advances the wqe pointer to point at the
1003 * scatter-gather list, and set wqe_end to the end of the WQE.
1004 */
1005 static int mlx5_ib_mr_initiator_pfault_handler(
1006 struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault,
1007 struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length)
1008 {
1009 struct mlx5_wqe_ctrl_seg *ctrl = *wqe;
1010 u16 wqe_index = pfault->wqe.wqe_index;
1011 u32 transport_caps;
1012 struct mlx5_base_av *av;
1013 unsigned ds, opcode;
1014 #if defined(DEBUG)
1015 u32 ctrl_wqe_index, ctrl_qpn;
1016 #endif
1017 u32 qpn = qp->trans_qp.base.mqp.qpn;
1018
1019 ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
1020 if (ds * MLX5_WQE_DS_UNITS > wqe_length) {
1021 mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n",
1022 ds, wqe_length);
1023 return -EFAULT;
1024 }
1025
1026 if (ds == 0) {
1027 mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n",
1028 wqe_index, qpn);
1029 return -EFAULT;
1030 }
1031
1032 #if defined(DEBUG)
1033 ctrl_wqe_index = (be32_to_cpu(ctrl->opmod_idx_opcode) &
1034 MLX5_WQE_CTRL_WQE_INDEX_MASK) >>
1035 MLX5_WQE_CTRL_WQE_INDEX_SHIFT;
1036 if (wqe_index != ctrl_wqe_index) {
1037 mlx5_ib_err(dev, "Got WQE with invalid wqe_index. wqe_index=0x%x, qpn=0x%x ctrl->wqe_index=0x%x\n",
1038 wqe_index, qpn,
1039 ctrl_wqe_index);
1040 return -EFAULT;
1041 }
1042
1043 ctrl_qpn = (be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_QPN_MASK) >>
1044 MLX5_WQE_CTRL_QPN_SHIFT;
1045 if (qpn != ctrl_qpn) {
1046 mlx5_ib_err(dev, "Got WQE with incorrect QP number. wqe_index=0x%x, qpn=0x%x ctrl->qpn=0x%x\n",
1047 wqe_index, qpn,
1048 ctrl_qpn);
1049 return -EFAULT;
1050 }
1051 #endif /* DEBUG */
1052
1053 *wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS;
1054 *wqe += sizeof(*ctrl);
1055
1056 opcode = be32_to_cpu(ctrl->opmod_idx_opcode) &
1057 MLX5_WQE_CTRL_OPCODE_MASK;
1058
1059 switch (qp->ibqp.qp_type) {
1060 case IB_QPT_XRC_INI:
1061 *wqe += sizeof(struct mlx5_wqe_xrc_seg);
1062 transport_caps = dev->odp_caps.per_transport_caps.xrc_odp_caps;
1063 break;
1064 case IB_QPT_RC:
1065 transport_caps = dev->odp_caps.per_transport_caps.rc_odp_caps;
1066 break;
1067 case IB_QPT_UD:
1068 transport_caps = dev->odp_caps.per_transport_caps.ud_odp_caps;
1069 break;
1070 default:
1071 mlx5_ib_err(dev, "ODP fault on QP of an unsupported transport 0x%x\n",
1072 qp->ibqp.qp_type);
1073 return -EFAULT;
1074 }
1075
1076 if (unlikely(opcode >= ARRAY_SIZE(mlx5_ib_odp_opcode_cap) ||
1077 !(transport_caps & mlx5_ib_odp_opcode_cap[opcode]))) {
1078 mlx5_ib_err(dev, "ODP fault on QP of an unsupported opcode 0x%x\n",
1079 opcode);
1080 return -EFAULT;
1081 }
1082
1083 if (qp->ibqp.qp_type == IB_QPT_UD) {
1084 av = *wqe;
1085 if (av->dqp_dct & cpu_to_be32(MLX5_EXTENDED_UD_AV))
1086 *wqe += sizeof(struct mlx5_av);
1087 else
1088 *wqe += sizeof(struct mlx5_base_av);
1089 }
1090
1091 switch (opcode) {
1092 case MLX5_OPCODE_RDMA_WRITE:
1093 case MLX5_OPCODE_RDMA_WRITE_IMM:
1094 case MLX5_OPCODE_RDMA_READ:
1095 *wqe += sizeof(struct mlx5_wqe_raddr_seg);
1096 break;
1097 case MLX5_OPCODE_ATOMIC_CS:
1098 case MLX5_OPCODE_ATOMIC_FA:
1099 *wqe += sizeof(struct mlx5_wqe_raddr_seg);
1100 *wqe += sizeof(struct mlx5_wqe_atomic_seg);
1101 break;
1102 }
1103
1104 return 0;
1105 }
1106
1107 /*
1108 * Parse responder WQE and set wqe_end to the end of the WQE.
1109 */
1110 static int mlx5_ib_mr_responder_pfault_handler_srq(struct mlx5_ib_dev *dev,
1111 struct mlx5_ib_srq *srq,
1112 void **wqe, void **wqe_end,
1113 int wqe_length)
1114 {
1115 int wqe_size = 1 << srq->msrq.wqe_shift;
1116
1117 if (wqe_size > wqe_length) {
1118 mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1119 return -EFAULT;
1120 }
1121
1122 *wqe_end = *wqe + wqe_size;
1123 *wqe += sizeof(struct mlx5_wqe_srq_next_seg);
1124
1125 return 0;
1126 }
1127
1128 static int mlx5_ib_mr_responder_pfault_handler_rq(struct mlx5_ib_dev *dev,
1129 struct mlx5_ib_qp *qp,
1130 void *wqe, void **wqe_end,
1131 int wqe_length)
1132 {
1133 struct mlx5_ib_wq *wq = &qp->rq;
1134 int wqe_size = 1 << wq->wqe_shift;
1135
1136 if (qp->wq_sig) {
1137 mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n");
1138 return -EFAULT;
1139 }
1140
1141 if (wqe_size > wqe_length) {
1142 mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
1143 return -EFAULT;
1144 }
1145
1146 switch (qp->ibqp.qp_type) {
1147 case IB_QPT_RC:
1148 if (!(dev->odp_caps.per_transport_caps.rc_odp_caps &
1149 IB_ODP_SUPPORT_RECV))
1150 goto invalid_transport_or_opcode;
1151 break;
1152 default:
1153 invalid_transport_or_opcode:
1154 mlx5_ib_err(dev, "ODP fault on QP of an unsupported transport. transport: 0x%x\n",
1155 qp->ibqp.qp_type);
1156 return -EFAULT;
1157 }
1158
1159 *wqe_end = wqe + wqe_size;
1160
1161 return 0;
1162 }
1163
1164 static inline struct mlx5_core_rsc_common *odp_get_rsc(struct mlx5_ib_dev *dev,
1165 u32 wq_num, int pf_type)
1166 {
1167 struct mlx5_core_rsc_common *common = NULL;
1168 struct mlx5_core_srq *srq;
1169
1170 switch (pf_type) {
1171 case MLX5_WQE_PF_TYPE_RMP:
1172 srq = mlx5_cmd_get_srq(dev, wq_num);
1173 if (srq)
1174 common = &srq->common;
1175 break;
1176 case MLX5_WQE_PF_TYPE_REQ_SEND_OR_WRITE:
1177 case MLX5_WQE_PF_TYPE_RESP:
1178 case MLX5_WQE_PF_TYPE_REQ_READ_OR_ATOMIC:
1179 common = mlx5_core_res_hold(dev->mdev, wq_num, MLX5_RES_QP);
1180 break;
1181 default:
1182 break;
1183 }
1184
1185 return common;
1186 }
1187
1188 static inline struct mlx5_ib_qp *res_to_qp(struct mlx5_core_rsc_common *res)
1189 {
1190 struct mlx5_core_qp *mqp = (struct mlx5_core_qp *)res;
1191
1192 return to_mibqp(mqp);
1193 }
1194
1195 static inline struct mlx5_ib_srq *res_to_srq(struct mlx5_core_rsc_common *res)
1196 {
1197 struct mlx5_core_srq *msrq =
1198 container_of(res, struct mlx5_core_srq, common);
1199
1200 return to_mibsrq(msrq);
1201 }
1202
1203 static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
1204 struct mlx5_pagefault *pfault)
1205 {
1206 bool sq = pfault->type & MLX5_PFAULT_REQUESTOR;
1207 u16 wqe_index = pfault->wqe.wqe_index;
1208 void *wqe = NULL, *wqe_end = NULL;
1209 u32 bytes_mapped, total_wqe_bytes;
1210 struct mlx5_core_rsc_common *res;
1211 int resume_with_error = 1;
1212 struct mlx5_ib_qp *qp;
1213 size_t bytes_copied;
1214 int ret = 0;
1215
1216 res = odp_get_rsc(dev, pfault->wqe.wq_num, pfault->type);
1217 if (!res) {
1218 mlx5_ib_dbg(dev, "wqe page fault for missing resource %d\n", pfault->wqe.wq_num);
1219 return;
1220 }
1221
1222 if (res->res != MLX5_RES_QP && res->res != MLX5_RES_SRQ &&
1223 res->res != MLX5_RES_XSRQ) {
1224 mlx5_ib_err(dev, "wqe page fault for unsupported type %d\n",
1225 pfault->type);
1226 goto resolve_page_fault;
1227 }
1228
1229 wqe = (void *)__get_free_page(GFP_KERNEL);
1230 if (!wqe) {
1231 mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n");
1232 goto resolve_page_fault;
1233 }
1234
1235 qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL;
1236 if (qp && sq) {
1237 ret = mlx5_ib_read_user_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE,
1238 &bytes_copied);
1239 if (ret)
1240 goto read_user;
1241 ret = mlx5_ib_mr_initiator_pfault_handler(
1242 dev, pfault, qp, &wqe, &wqe_end, bytes_copied);
1243 } else if (qp && !sq) {
1244 ret = mlx5_ib_read_user_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE,
1245 &bytes_copied);
1246 if (ret)
1247 goto read_user;
1248 ret = mlx5_ib_mr_responder_pfault_handler_rq(
1249 dev, qp, wqe, &wqe_end, bytes_copied);
1250 } else if (!qp) {
1251 struct mlx5_ib_srq *srq = res_to_srq(res);
1252
1253 ret = mlx5_ib_read_user_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE,
1254 &bytes_copied);
1255 if (ret)
1256 goto read_user;
1257 ret = mlx5_ib_mr_responder_pfault_handler_srq(
1258 dev, srq, &wqe, &wqe_end, bytes_copied);
1259 }
1260
1261 if (ret < 0 || wqe >= wqe_end)
1262 goto resolve_page_fault;
1263
1264 ret = pagefault_data_segments(dev, pfault, wqe, wqe_end, &bytes_mapped,
1265 &total_wqe_bytes, !sq);
1266 if (ret == -EAGAIN)
1267 goto out;
1268
1269 if (ret < 0 || total_wqe_bytes > bytes_mapped)
1270 goto resolve_page_fault;
1271
1272 out:
1273 ret = 0;
1274 resume_with_error = 0;
1275
1276 read_user:
1277 if (ret)
1278 mlx5_ib_err(
1279 dev,
1280 "Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n",
1281 ret, wqe_index, pfault->token);
1282
1283 resolve_page_fault:
1284 mlx5_ib_page_fault_resume(dev, pfault, resume_with_error);
1285 mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n",
1286 pfault->wqe.wq_num, resume_with_error,
1287 pfault->type);
1288 mlx5_core_res_put(res);
1289 free_page((unsigned long)wqe);
1290 }
1291
1292 static int pages_in_range(u64 address, u32 length)
1293 {
1294 return (ALIGN(address + length, PAGE_SIZE) -
1295 (address & PAGE_MASK)) >> PAGE_SHIFT;
1296 }
1297
1298 static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev,
1299 struct mlx5_pagefault *pfault)
1300 {
1301 u64 address;
1302 u32 length;
1303 u32 prefetch_len = pfault->bytes_committed;
1304 int prefetch_activated = 0;
1305 u32 rkey = pfault->rdma.r_key;
1306 int ret;
1307
1308 /* The RDMA responder handler handles the page fault in two parts.
1309 * First it brings the necessary pages for the current packet
1310 * (and uses the pfault context), and then (after resuming the QP)
1311 * prefetches more pages. The second operation cannot use the pfault
1312 * context and therefore uses the dummy_pfault context allocated on
1313 * the stack */
1314 pfault->rdma.rdma_va += pfault->bytes_committed;
1315 pfault->rdma.rdma_op_len -= min(pfault->bytes_committed,
1316 pfault->rdma.rdma_op_len);
1317 pfault->bytes_committed = 0;
1318
1319 address = pfault->rdma.rdma_va;
1320 length = pfault->rdma.rdma_op_len;
1321
1322 /* For some operations, the hardware cannot tell the exact message
1323 * length, and in those cases it reports zero. Use prefetch
1324 * logic. */
1325 if (length == 0) {
1326 prefetch_activated = 1;
1327 length = pfault->rdma.packet_size;
1328 prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len);
1329 }
1330
1331 ret = pagefault_single_data_segment(dev, NULL, rkey, address, length,
1332 &pfault->bytes_committed, NULL,
1333 0);
1334 if (ret == -EAGAIN) {
1335 /* We're racing with an invalidation, don't prefetch */
1336 prefetch_activated = 0;
1337 } else if (ret < 0 || pages_in_range(address, length) > ret) {
1338 mlx5_ib_page_fault_resume(dev, pfault, 1);
1339 if (ret != -ENOENT)
1340 mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n",
1341 ret, pfault->token, pfault->type);
1342 return;
1343 }
1344
1345 mlx5_ib_page_fault_resume(dev, pfault, 0);
1346 mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n",
1347 pfault->token, pfault->type,
1348 prefetch_activated);
1349
1350 /* At this point, there might be a new pagefault already arriving in
1351 * the eq, switch to the dummy pagefault for the rest of the
1352 * processing. We're still OK with the objects being alive as the
1353 * work-queue is being fenced. */
1354
1355 if (prefetch_activated) {
1356 u32 bytes_committed = 0;
1357
1358 ret = pagefault_single_data_segment(dev, NULL, rkey, address,
1359 prefetch_len,
1360 &bytes_committed, NULL,
1361 0);
1362 if (ret < 0 && ret != -EAGAIN) {
1363 mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n",
1364 ret, pfault->token, address, prefetch_len);
1365 }
1366 }
1367 }
1368
1369 static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault)
1370 {
1371 u8 event_subtype = pfault->event_subtype;
1372
1373 switch (event_subtype) {
1374 case MLX5_PFAULT_SUBTYPE_WQE:
1375 mlx5_ib_mr_wqe_pfault_handler(dev, pfault);
1376 break;
1377 case MLX5_PFAULT_SUBTYPE_RDMA:
1378 mlx5_ib_mr_rdma_pfault_handler(dev, pfault);
1379 break;
1380 default:
1381 mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n",
1382 event_subtype);
1383 mlx5_ib_page_fault_resume(dev, pfault, 1);
1384 }
1385 }
1386
1387 static void mlx5_ib_eqe_pf_action(struct work_struct *work)
1388 {
1389 struct mlx5_pagefault *pfault = container_of(work,
1390 struct mlx5_pagefault,
1391 work);
1392 struct mlx5_ib_pf_eq *eq = pfault->eq;
1393
1394 mlx5_ib_pfault(eq->dev, pfault);
1395 mempool_free(pfault, eq->pool);
1396 }
1397
1398 static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq)
1399 {
1400 struct mlx5_eqe_page_fault *pf_eqe;
1401 struct mlx5_pagefault *pfault;
1402 struct mlx5_eqe *eqe;
1403 int cc = 0;
1404
1405 while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) {
1406 pfault = mempool_alloc(eq->pool, GFP_ATOMIC);
1407 if (!pfault) {
1408 schedule_work(&eq->work);
1409 break;
1410 }
1411
1412 pf_eqe = &eqe->data.page_fault;
1413 pfault->event_subtype = eqe->sub_type;
1414 pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);
1415
1416 mlx5_ib_dbg(eq->dev,
1417 "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
1418 eqe->sub_type, pfault->bytes_committed);
1419
1420 switch (eqe->sub_type) {
1421 case MLX5_PFAULT_SUBTYPE_RDMA:
1422 /* RDMA based event */
1423 pfault->type =
1424 be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
1425 pfault->token =
1426 be32_to_cpu(pf_eqe->rdma.pftype_token) &
1427 MLX5_24BIT_MASK;
1428 pfault->rdma.r_key =
1429 be32_to_cpu(pf_eqe->rdma.r_key);
1430 pfault->rdma.packet_size =
1431 be16_to_cpu(pf_eqe->rdma.packet_length);
1432 pfault->rdma.rdma_op_len =
1433 be32_to_cpu(pf_eqe->rdma.rdma_op_len);
1434 pfault->rdma.rdma_va =
1435 be64_to_cpu(pf_eqe->rdma.rdma_va);
1436 mlx5_ib_dbg(eq->dev,
1437 "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
1438 pfault->type, pfault->token,
1439 pfault->rdma.r_key);
1440 mlx5_ib_dbg(eq->dev,
1441 "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
1442 pfault->rdma.rdma_op_len,
1443 pfault->rdma.rdma_va);
1444 break;
1445
1446 case MLX5_PFAULT_SUBTYPE_WQE:
1447 /* WQE based event */
1448 pfault->type =
1449 (be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7;
1450 pfault->token =
1451 be32_to_cpu(pf_eqe->wqe.token);
1452 pfault->wqe.wq_num =
1453 be32_to_cpu(pf_eqe->wqe.pftype_wq) &
1454 MLX5_24BIT_MASK;
1455 pfault->wqe.wqe_index =
1456 be16_to_cpu(pf_eqe->wqe.wqe_index);
1457 pfault->wqe.packet_size =
1458 be16_to_cpu(pf_eqe->wqe.packet_length);
1459 mlx5_ib_dbg(eq->dev,
1460 "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
1461 pfault->type, pfault->token,
1462 pfault->wqe.wq_num,
1463 pfault->wqe.wqe_index);
1464 break;
1465
1466 default:
1467 mlx5_ib_warn(eq->dev,
1468 "Unsupported page fault event sub-type: 0x%02hhx\n",
1469 eqe->sub_type);
1470 /* Unsupported page faults should still be
1471 * resolved by the page fault handler
1472 */
1473 }
1474
1475 pfault->eq = eq;
1476 INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action);
1477 queue_work(eq->wq, &pfault->work);
1478
1479 cc = mlx5_eq_update_cc(eq->core, ++cc);
1480 }
1481
1482 mlx5_eq_update_ci(eq->core, cc, 1);
1483 }
1484
1485 static int mlx5_ib_eq_pf_int(struct notifier_block *nb, unsigned long type,
1486 void *data)
1487 {
1488 struct mlx5_ib_pf_eq *eq =
1489 container_of(nb, struct mlx5_ib_pf_eq, irq_nb);
1490 unsigned long flags;
1491
1492 if (spin_trylock_irqsave(&eq->lock, flags)) {
1493 mlx5_ib_eq_pf_process(eq);
1494 spin_unlock_irqrestore(&eq->lock, flags);
1495 } else {
1496 schedule_work(&eq->work);
1497 }
1498
1499 return IRQ_HANDLED;
1500 }
1501
1502 /* mempool_refill() was proposed but unfortunately wasn't accepted
1503 * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
1504 * Cheap workaround.
1505 */
1506 static void mempool_refill(mempool_t *pool)
1507 {
1508 while (pool->curr_nr < pool->min_nr)
1509 mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
1510 }
1511
1512 static void mlx5_ib_eq_pf_action(struct work_struct *work)
1513 {
1514 struct mlx5_ib_pf_eq *eq =
1515 container_of(work, struct mlx5_ib_pf_eq, work);
1516
1517 mempool_refill(eq->pool);
1518
1519 spin_lock_irq(&eq->lock);
1520 mlx5_ib_eq_pf_process(eq);
1521 spin_unlock_irq(&eq->lock);
1522 }
1523
1524 enum {
1525 MLX5_IB_NUM_PF_EQE = 0x1000,
1526 MLX5_IB_NUM_PF_DRAIN = 64,
1527 };
1528
1529 static int
1530 mlx5_ib_create_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1531 {
1532 struct mlx5_eq_param param = {};
1533 int err;
1534
1535 INIT_WORK(&eq->work, mlx5_ib_eq_pf_action);
1536 spin_lock_init(&eq->lock);
1537 eq->dev = dev;
1538
1539 eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN,
1540 sizeof(struct mlx5_pagefault));
1541 if (!eq->pool)
1542 return -ENOMEM;
1543
1544 eq->wq = alloc_workqueue("mlx5_ib_page_fault",
1545 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM,
1546 MLX5_NUM_CMD_EQE);
1547 if (!eq->wq) {
1548 err = -ENOMEM;
1549 goto err_mempool;
1550 }
1551
1552 eq->irq_nb.notifier_call = mlx5_ib_eq_pf_int;
1553 param = (struct mlx5_eq_param) {
1554 .irq_index = 0,
1555 .nent = MLX5_IB_NUM_PF_EQE,
1556 };
1557 param.mask[0] = 1ull << MLX5_EVENT_TYPE_PAGE_FAULT;
1558 eq->core = mlx5_eq_create_generic(dev->mdev, &param);
1559 if (IS_ERR(eq->core)) {
1560 err = PTR_ERR(eq->core);
1561 goto err_wq;
1562 }
1563 err = mlx5_eq_enable(dev->mdev, eq->core, &eq->irq_nb);
1564 if (err) {
1565 mlx5_ib_err(dev, "failed to enable odp EQ %d\n", err);
1566 goto err_eq;
1567 }
1568
1569 return 0;
1570 err_eq:
1571 mlx5_eq_destroy_generic(dev->mdev, eq->core);
1572 err_wq:
1573 destroy_workqueue(eq->wq);
1574 err_mempool:
1575 mempool_destroy(eq->pool);
1576 return err;
1577 }
1578
1579 static int
1580 mlx5_ib_destroy_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
1581 {
1582 int err;
1583
1584 mlx5_eq_disable(dev->mdev, eq->core, &eq->irq_nb);
1585 err = mlx5_eq_destroy_generic(dev->mdev, eq->core);
1586 cancel_work_sync(&eq->work);
1587 destroy_workqueue(eq->wq);
1588 mempool_destroy(eq->pool);
1589
1590 return err;
1591 }
1592
1593 void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
1594 {
1595 if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1596 return;
1597
1598 switch (ent->order - 2) {
1599 case MLX5_IMR_MTT_CACHE_ENTRY:
1600 ent->page = PAGE_SHIFT;
1601 ent->xlt = MLX5_IMR_MTT_ENTRIES *
1602 sizeof(struct mlx5_mtt) /
1603 MLX5_IB_UMR_OCTOWORD;
1604 ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
1605 ent->limit = 0;
1606 break;
1607
1608 case MLX5_IMR_KSM_CACHE_ENTRY:
1609 ent->page = MLX5_KSM_PAGE_SHIFT;
1610 ent->xlt = mlx5_imr_ksm_entries *
1611 sizeof(struct mlx5_klm) /
1612 MLX5_IB_UMR_OCTOWORD;
1613 ent->access_mode = MLX5_MKC_ACCESS_MODE_KSM;
1614 ent->limit = 0;
1615 break;
1616 }
1617 }
1618
1619 static const struct ib_device_ops mlx5_ib_dev_odp_ops = {
1620 .advise_mr = mlx5_ib_advise_mr,
1621 };
1622
1623 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
1624 {
1625 int ret = 0;
1626
1627 if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1628 return ret;
1629
1630 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_odp_ops);
1631
1632 if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) {
1633 ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey);
1634 if (ret) {
1635 mlx5_ib_err(dev, "Error getting null_mkey %d\n", ret);
1636 return ret;
1637 }
1638 }
1639
1640 ret = mlx5_ib_create_pf_eq(dev, &dev->odp_pf_eq);
1641
1642 return ret;
1643 }
1644
1645 void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev)
1646 {
1647 if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
1648 return;
1649
1650 mlx5_ib_destroy_pf_eq(dev, &dev->odp_pf_eq);
1651 }
1652
1653 int mlx5_ib_odp_init(void)
1654 {
1655 mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) -
1656 MLX5_IMR_MTT_BITS);
1657
1658 return 0;
1659 }
1660
1661 struct prefetch_mr_work {
1662 struct work_struct work;
1663 struct ib_pd *pd;
1664 u32 pf_flags;
1665 u32 num_sge;
1666 struct ib_sge sg_list[0];
1667 };
1668
1669 static void num_pending_prefetch_dec(struct mlx5_ib_dev *dev,
1670 struct ib_sge *sg_list, u32 num_sge,
1671 u32 from)
1672 {
1673 u32 i;
1674 int srcu_key;
1675
1676 srcu_key = srcu_read_lock(&dev->mr_srcu);
1677
1678 for (i = from; i < num_sge; ++i) {
1679 struct mlx5_core_mkey *mmkey;
1680 struct mlx5_ib_mr *mr;
1681
1682 mmkey = xa_load(&dev->mdev->priv.mkey_table,
1683 mlx5_base_mkey(sg_list[i].lkey));
1684 mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
1685 atomic_dec(&mr->num_pending_prefetch);
1686 }
1687
1688 srcu_read_unlock(&dev->mr_srcu, srcu_key);
1689 }
1690
1691 static bool num_pending_prefetch_inc(struct ib_pd *pd,
1692 struct ib_sge *sg_list, u32 num_sge)
1693 {
1694 struct mlx5_ib_dev *dev = to_mdev(pd->device);
1695 bool ret = true;
1696 u32 i;
1697
1698 for (i = 0; i < num_sge; ++i) {
1699 struct mlx5_core_mkey *mmkey;
1700 struct mlx5_ib_mr *mr;
1701
1702 mmkey = xa_load(&dev->mdev->priv.mkey_table,
1703 mlx5_base_mkey(sg_list[i].lkey));
1704 if (!mmkey || mmkey->key != sg_list[i].lkey) {
1705 ret = false;
1706 break;
1707 }
1708
1709 if (mmkey->type != MLX5_MKEY_MR) {
1710 ret = false;
1711 break;
1712 }
1713
1714 mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
1715
1716 if (mr->ibmr.pd != pd) {
1717 ret = false;
1718 break;
1719 }
1720
1721 if (!mr->live) {
1722 ret = false;
1723 break;
1724 }
1725
1726 atomic_inc(&mr->num_pending_prefetch);
1727 }
1728
1729 if (!ret)
1730 num_pending_prefetch_dec(dev, sg_list, i, 0);
1731
1732 return ret;
1733 }
1734
1735 static int mlx5_ib_prefetch_sg_list(struct ib_pd *pd, u32 pf_flags,
1736 struct ib_sge *sg_list, u32 num_sge)
1737 {
1738 u32 i;
1739 int ret = 0;
1740 struct mlx5_ib_dev *dev = to_mdev(pd->device);
1741
1742 for (i = 0; i < num_sge; ++i) {
1743 struct ib_sge *sg = &sg_list[i];
1744 int bytes_committed = 0;
1745
1746 ret = pagefault_single_data_segment(dev, pd, sg->lkey, sg->addr,
1747 sg->length,
1748 &bytes_committed, NULL,
1749 pf_flags);
1750 if (ret < 0)
1751 break;
1752 }
1753
1754 return ret < 0 ? ret : 0;
1755 }
1756
1757 static void mlx5_ib_prefetch_mr_work(struct work_struct *work)
1758 {
1759 struct prefetch_mr_work *w =
1760 container_of(work, struct prefetch_mr_work, work);
1761
1762 if (ib_device_try_get(w->pd->device)) {
1763 mlx5_ib_prefetch_sg_list(w->pd, w->pf_flags, w->sg_list,
1764 w->num_sge);
1765 ib_device_put(w->pd->device);
1766 }
1767
1768 num_pending_prefetch_dec(to_mdev(w->pd->device), w->sg_list,
1769 w->num_sge, 0);
1770 kvfree(w);
1771 }
1772
1773 int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
1774 enum ib_uverbs_advise_mr_advice advice,
1775 u32 flags, struct ib_sge *sg_list, u32 num_sge)
1776 {
1777 struct mlx5_ib_dev *dev = to_mdev(pd->device);
1778 u32 pf_flags = MLX5_PF_FLAGS_PREFETCH;
1779 struct prefetch_mr_work *work;
1780 bool valid_req;
1781 int srcu_key;
1782
1783 if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH)
1784 pf_flags |= MLX5_PF_FLAGS_DOWNGRADE;
1785
1786 if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH)
1787 return mlx5_ib_prefetch_sg_list(pd, pf_flags, sg_list,
1788 num_sge);
1789
1790 work = kvzalloc(struct_size(work, sg_list, num_sge), GFP_KERNEL);
1791 if (!work)
1792 return -ENOMEM;
1793
1794 memcpy(work->sg_list, sg_list, num_sge * sizeof(struct ib_sge));
1795
1796 /* It is guaranteed that the pd when work is executed is the pd when
1797 * work was queued since pd can't be destroyed while it holds MRs and
1798 * destroying a MR leads to flushing the workquque
1799 */
1800 work->pd = pd;
1801 work->pf_flags = pf_flags;
1802 work->num_sge = num_sge;
1803
1804 INIT_WORK(&work->work, mlx5_ib_prefetch_mr_work);
1805
1806 srcu_key = srcu_read_lock(&dev->mr_srcu);
1807
1808 valid_req = num_pending_prefetch_inc(pd, sg_list, num_sge);
1809 if (valid_req)
1810 queue_work(system_unbound_wq, &work->work);
1811 else
1812 kvfree(work);
1813
1814 srcu_read_unlock(&dev->mr_srcu, srcu_key);
1815
1816 return valid_req ? 0 : -EINVAL;
1817 }
Ubuntu Focal Linux kernel mirror