]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/dev/irdma/irdma_verbs.c
ssh: Update to OpenSSH 9.3p2
[FreeBSD/FreeBSD.git] / sys / dev / irdma / irdma_verbs.c
1 /*-
2  * SPDX-License-Identifier: GPL-2.0 or Linux-OpenIB
3  *
4  * Copyright (c) 2015 - 2022 Intel Corporation
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenFabrics.org BSD license below:
11  *
12  *   Redistribution and use in source and binary forms, with or
13  *   without modification, are permitted provided that the following
14  *   conditions are met:
15  *
16  *    - Redistributions of source code must retain the above
17  *      copyright notice, this list of conditions and the following
18  *      disclaimer.
19  *
20  *    - Redistributions in binary form must reproduce the above
21  *      copyright notice, this list of conditions and the following
22  *      disclaimer in the documentation and/or other materials
23  *      provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 /*$FreeBSD$*/
35
36 #include "irdma_main.h"
37
38 /**
39  * irdma_query_device - get device attributes
40  * @ibdev: device pointer from stack
41  * @props: returning device attributes
42  * @udata: user data
43  */
44 static int
45 irdma_query_device(struct ib_device *ibdev,
46                    struct ib_device_attr *props,
47                    struct ib_udata *udata)
48 {
49         struct irdma_device *iwdev = to_iwdev(ibdev);
50         struct irdma_pci_f *rf = iwdev->rf;
51         struct pci_dev *pcidev = iwdev->rf->pcidev;
52         struct irdma_hw_attrs *hw_attrs = &rf->sc_dev.hw_attrs;
53
54         if (udata->inlen || udata->outlen)
55                 return -EINVAL;
56
57         memset(props, 0, sizeof(*props));
58         addrconf_addr_eui48((u8 *)&props->sys_image_guid,
59                             if_getlladdr(iwdev->netdev));
60         props->fw_ver = (u64)irdma_fw_major_ver(&rf->sc_dev) << 32 |
61             irdma_fw_minor_ver(&rf->sc_dev);
62         props->device_cap_flags = IB_DEVICE_MEM_WINDOW |
63             IB_DEVICE_MEM_MGT_EXTENSIONS;
64         props->device_cap_flags |= IB_DEVICE_LOCAL_DMA_LKEY;
65         props->vendor_id = pcidev->vendor;
66         props->vendor_part_id = pcidev->device;
67         props->hw_ver = pcidev->revision;
68         props->page_size_cap = hw_attrs->page_size_cap;
69         props->max_mr_size = hw_attrs->max_mr_size;
70         props->max_qp = rf->max_qp - rf->used_qps;
71         props->max_qp_wr = hw_attrs->max_qp_wr;
72         set_max_sge(props, rf);
73         props->max_cq = rf->max_cq - rf->used_cqs;
74         props->max_cqe = rf->max_cqe - 1;
75         props->max_mr = rf->max_mr - rf->used_mrs;
76         props->max_mw = props->max_mr;
77         props->max_pd = rf->max_pd - rf->used_pds;
78         props->max_sge_rd = hw_attrs->uk_attrs.max_hw_read_sges;
79         props->max_qp_rd_atom = hw_attrs->max_hw_ird;
80         props->max_qp_init_rd_atom = hw_attrs->max_hw_ord;
81         if (rdma_protocol_roce(ibdev, 1)) {
82                 props->device_cap_flags |= IB_DEVICE_RC_RNR_NAK_GEN;
83                 props->max_pkeys = IRDMA_PKEY_TBL_SZ;
84                 props->max_ah = rf->max_ah;
85                 if (hw_attrs->uk_attrs.hw_rev == IRDMA_GEN_2) {
86                         props->max_mcast_grp = rf->max_mcg;
87                         props->max_mcast_qp_attach = IRDMA_MAX_MGS_PER_CTX;
88                         props->max_total_mcast_qp_attach = rf->max_qp * IRDMA_MAX_MGS_PER_CTX;
89                 }
90         }
91         props->max_fast_reg_page_list_len = IRDMA_MAX_PAGES_PER_FMR;
92         if (hw_attrs->uk_attrs.hw_rev >= IRDMA_GEN_2)
93                 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW_TYPE_2B;
94
95         return 0;
96 }
97
98 static int
99 irdma_mmap_legacy(struct irdma_ucontext *ucontext,
100                   struct vm_area_struct *vma)
101 {
102         u64 pfn;
103
104         if (vma->vm_pgoff || vma->vm_end - vma->vm_start != PAGE_SIZE)
105                 return -EINVAL;
106
107         vma->vm_private_data = ucontext;
108         pfn = ((uintptr_t)ucontext->iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET] +
109                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
110
111 #if __FreeBSD_version >= 1400026
112         return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
113                                  pgprot_noncached(vma->vm_page_prot), NULL);
114 #else
115         return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
116                                  pgprot_noncached(vma->vm_page_prot));
117 #endif
118 }
119
120 #if __FreeBSD_version >= 1400026
121 static void
122 irdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
123 {
124         struct irdma_user_mmap_entry *entry = to_irdma_mmap_entry(rdma_entry);
125
126         kfree(entry);
127 }
128
129 struct rdma_user_mmap_entry *
130 irdma_user_mmap_entry_insert(struct irdma_ucontext *ucontext, u64 bar_offset,
131                              enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
132 {
133         struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
134         int ret;
135
136         if (!entry)
137                 return NULL;
138
139         entry->bar_offset = bar_offset;
140         entry->mmap_flag = mmap_flag;
141
142         ret = rdma_user_mmap_entry_insert(&ucontext->ibucontext,
143                                           &entry->rdma_entry, PAGE_SIZE);
144         if (ret) {
145                 kfree(entry);
146                 return NULL;
147         }
148         *mmap_offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
149
150         return &entry->rdma_entry;
151 }
152
153 #else
154 static inline bool
155 find_key_in_mmap_tbl(struct irdma_ucontext *ucontext, u64 key)
156 {
157         struct irdma_user_mmap_entry *entry;
158
159         HASH_FOR_EACH_POSSIBLE(ucontext->mmap_hash_tbl, entry, hlist, key) {
160                 if (entry->pgoff_key == key)
161                         return true;
162         }
163
164         return false;
165 }
166
167 struct irdma_user_mmap_entry *
168 irdma_user_mmap_entry_add_hash(struct irdma_ucontext *ucontext, u64 bar_offset,
169                                enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
170 {
171         struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
172         unsigned long flags;
173         int retry_cnt = 0;
174
175         if (!entry)
176                 return NULL;
177
178         entry->bar_offset = bar_offset;
179         entry->mmap_flag = mmap_flag;
180         entry->ucontext = ucontext;
181         do {
182                 get_random_bytes(&entry->pgoff_key, sizeof(entry->pgoff_key));
183
184                 /* The key is a page offset */
185                 entry->pgoff_key >>= PAGE_SHIFT;
186
187                 /* In the event of a collision in the hash table, retry a new key */
188                 spin_lock_irqsave(&ucontext->mmap_tbl_lock, flags);
189                 if (!find_key_in_mmap_tbl(ucontext, entry->pgoff_key)) {
190                         HASH_ADD(ucontext->mmap_hash_tbl, &entry->hlist, entry->pgoff_key);
191                         spin_unlock_irqrestore(&ucontext->mmap_tbl_lock, flags);
192                         goto hash_add_done;
193                 }
194                 spin_unlock_irqrestore(&ucontext->mmap_tbl_lock, flags);
195         } while (retry_cnt++ < 10);
196
197         irdma_debug(&ucontext->iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
198                     "mmap table add failed: Cannot find a unique key\n");
199         kfree(entry);
200         return NULL;
201
202 hash_add_done:
203         /* libc mmap uses a byte offset */
204         *mmap_offset = entry->pgoff_key << PAGE_SHIFT;
205
206         return entry;
207 }
208
209 static struct irdma_user_mmap_entry *
210 irdma_find_user_mmap_entry(struct irdma_ucontext *ucontext,
211                            struct vm_area_struct *vma)
212 {
213         struct irdma_user_mmap_entry *entry;
214         unsigned long flags;
215
216         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
217                 return NULL;
218
219         spin_lock_irqsave(&ucontext->mmap_tbl_lock, flags);
220         HASH_FOR_EACH_POSSIBLE(ucontext->mmap_hash_tbl, entry, hlist, vma->vm_pgoff) {
221                 if (entry->pgoff_key == vma->vm_pgoff) {
222                         spin_unlock_irqrestore(&ucontext->mmap_tbl_lock, flags);
223                         return entry;
224                 }
225         }
226
227         spin_unlock_irqrestore(&ucontext->mmap_tbl_lock, flags);
228
229         return NULL;
230 }
231
232 void
233 irdma_user_mmap_entry_del_hash(struct irdma_user_mmap_entry *entry)
234 {
235         struct irdma_ucontext *ucontext;
236         unsigned long flags;
237
238         if (!entry)
239                 return;
240
241         ucontext = entry->ucontext;
242
243         spin_lock_irqsave(&ucontext->mmap_tbl_lock, flags);
244         HASH_DEL(ucontext->mmap_hash_tbl, &entry->hlist);
245         spin_unlock_irqrestore(&ucontext->mmap_tbl_lock, flags);
246
247         kfree(entry);
248 }
249
250 #endif
251 /**
252  * irdma_mmap - user memory map
253  * @context: context created during alloc
254  * @vma: kernel info for user memory map
255  */
256 static int
257 irdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
258 {
259 #if __FreeBSD_version >= 1400026
260         struct rdma_user_mmap_entry *rdma_entry;
261 #endif
262         struct irdma_user_mmap_entry *entry;
263         struct irdma_ucontext *ucontext;
264         u64 pfn;
265         int ret;
266
267         ucontext = to_ucontext(context);
268
269         /* Legacy support for libi40iw with hard-coded mmap key */
270         if (ucontext->legacy_mode)
271                 return irdma_mmap_legacy(ucontext, vma);
272
273 #if __FreeBSD_version >= 1400026
274         rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
275         if (!rdma_entry) {
276                 irdma_debug(&ucontext->iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
277                             "pgoff[0x%lx] does not have valid entry\n",
278                             vma->vm_pgoff);
279                 return -EINVAL;
280         }
281
282         entry = to_irdma_mmap_entry(rdma_entry);
283 #else
284         entry = irdma_find_user_mmap_entry(ucontext, vma);
285         if (!entry) {
286                 irdma_debug(&ucontext->iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
287                             "pgoff[0x%lx] does not have valid entry\n",
288                             vma->vm_pgoff);
289                 return -EINVAL;
290         }
291 #endif
292         irdma_debug(&ucontext->iwdev->rf->sc_dev,
293                     IRDMA_DEBUG_VERBS, "bar_offset [0x%lx] mmap_flag [%d]\n",
294                     entry->bar_offset, entry->mmap_flag);
295
296         pfn = (entry->bar_offset +
297                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
298
299         switch (entry->mmap_flag) {
300         case IRDMA_MMAP_IO_NC:
301 #if __FreeBSD_version >= 1400026
302                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
303                                         pgprot_noncached(vma->vm_page_prot),
304                                         rdma_entry);
305 #else
306                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
307                                         pgprot_noncached(vma->vm_page_prot));
308 #endif
309                 break;
310         case IRDMA_MMAP_IO_WC:
311 #if __FreeBSD_version >= 1400026
312                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
313                                         pgprot_writecombine(vma->vm_page_prot),
314                                         rdma_entry);
315 #else
316                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
317                                         pgprot_writecombine(vma->vm_page_prot));
318 #endif
319                 break;
320         default:
321                 ret = -EINVAL;
322         }
323
324         if (ret)
325                 irdma_debug(&ucontext->iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
326                             "bar_offset [0x%lx] mmap_flag[%d] err[%d]\n",
327                             entry->bar_offset, entry->mmap_flag, ret);
328 #if __FreeBSD_version >= 1400026
329         rdma_user_mmap_entry_put(rdma_entry);
330 #endif
331
332         return ret;
333 }
334
335 /**
336  * irdma_alloc_push_page - allocate a push page for qp
337  * @iwqp: qp pointer
338  */
339 static void
340 irdma_alloc_push_page(struct irdma_qp *iwqp)
341 {
342         struct irdma_cqp_request *cqp_request;
343         struct cqp_cmds_info *cqp_info;
344         struct irdma_device *iwdev = iwqp->iwdev;
345         struct irdma_sc_qp *qp = &iwqp->sc_qp;
346         int status;
347
348         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
349         if (!cqp_request)
350                 return;
351
352         cqp_info = &cqp_request->info;
353         cqp_info->cqp_cmd = IRDMA_OP_MANAGE_PUSH_PAGE;
354         cqp_info->post_sq = 1;
355         cqp_info->in.u.manage_push_page.info.push_idx = 0;
356         cqp_info->in.u.manage_push_page.info.qs_handle =
357             qp->vsi->qos[qp->user_pri].qs_handle;
358         cqp_info->in.u.manage_push_page.info.free_page = 0;
359         cqp_info->in.u.manage_push_page.info.push_page_type = 0;
360         cqp_info->in.u.manage_push_page.cqp = &iwdev->rf->cqp.sc_cqp;
361         cqp_info->in.u.manage_push_page.scratch = (uintptr_t)cqp_request;
362
363         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
364         if (!status && cqp_request->compl_info.op_ret_val <
365             iwdev->rf->sc_dev.hw_attrs.max_hw_device_pages) {
366                 qp->push_idx = cqp_request->compl_info.op_ret_val;
367                 qp->push_offset = 0;
368         }
369
370         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
371 }
372
373 /**
374  * irdma_get_pbl - Retrieve pbl from a list given a virtual
375  * address
376  * @va: user virtual address
377  * @pbl_list: pbl list to search in (QP's or CQ's)
378  */
379 struct irdma_pbl *
380 irdma_get_pbl(unsigned long va,
381               struct list_head *pbl_list)
382 {
383         struct irdma_pbl *iwpbl;
384
385         list_for_each_entry(iwpbl, pbl_list, list) {
386                 if (iwpbl->user_base == va) {
387                         list_del(&iwpbl->list);
388                         iwpbl->on_list = false;
389                         return iwpbl;
390                 }
391         }
392
393         return NULL;
394 }
395
396 /**
397  * irdma_clean_cqes - clean cq entries for qp
398  * @iwqp: qp ptr (user or kernel)
399  * @iwcq: cq ptr
400  */
401 void
402 irdma_clean_cqes(struct irdma_qp *iwqp, struct irdma_cq *iwcq)
403 {
404         struct irdma_cq_uk *ukcq = &iwcq->sc_cq.cq_uk;
405         unsigned long flags;
406
407         spin_lock_irqsave(&iwcq->lock, flags);
408         irdma_uk_clean_cq(&iwqp->sc_qp.qp_uk, ukcq);
409         spin_unlock_irqrestore(&iwcq->lock, flags);
410 }
411
412 static u64 irdma_compute_push_wqe_offset(struct irdma_device *iwdev, u32 page_idx){
413         u64 bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
414
415         if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.hw_rev == IRDMA_GEN_2) {
416                 /* skip over db page */
417                 bar_off += IRDMA_HW_PAGE_SIZE;
418                 /* skip over reserved space */
419                 bar_off += IRDMA_PF_BAR_RSVD;
420         }
421
422         /* push wqe page */
423         bar_off += (u64)page_idx * IRDMA_HW_PAGE_SIZE;
424
425         return bar_off;
426 }
427
428 void
429 irdma_remove_push_mmap_entries(struct irdma_qp *iwqp)
430 {
431         if (iwqp->push_db_mmap_entry) {
432 #if __FreeBSD_version >= 1400026
433                 rdma_user_mmap_entry_remove(iwqp->push_db_mmap_entry);
434 #else
435                 irdma_user_mmap_entry_del_hash(iwqp->push_db_mmap_entry);
436 #endif
437                 iwqp->push_db_mmap_entry = NULL;
438         }
439         if (iwqp->push_wqe_mmap_entry) {
440 #if __FreeBSD_version >= 1400026
441                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
442 #else
443                 irdma_user_mmap_entry_del_hash(iwqp->push_wqe_mmap_entry);
444 #endif
445                 iwqp->push_wqe_mmap_entry = NULL;
446         }
447 }
448
449 static int
450 irdma_setup_push_mmap_entries(struct irdma_ucontext *ucontext,
451                               struct irdma_qp *iwqp,
452                               u64 *push_wqe_mmap_key,
453                               u64 *push_db_mmap_key)
454 {
455         struct irdma_device *iwdev = ucontext->iwdev;
456         u64 bar_off;
457
458         WARN_ON_ONCE(iwdev->rf->sc_dev.hw_attrs.uk_attrs.hw_rev < IRDMA_GEN_2);
459
460         bar_off = irdma_compute_push_wqe_offset(iwdev, iwqp->sc_qp.push_idx);
461
462 #if __FreeBSD_version >= 1400026
463         iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
464                                                                  bar_off, IRDMA_MMAP_IO_WC,
465                                                                  push_wqe_mmap_key);
466 #else
467         iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_add_hash(ucontext, bar_off,
468                                                                    IRDMA_MMAP_IO_WC,
469                                                                    push_wqe_mmap_key);
470 #endif
471         if (!iwqp->push_wqe_mmap_entry)
472                 return -ENOMEM;
473
474         /* push doorbell page */
475         bar_off += IRDMA_HW_PAGE_SIZE;
476 #if __FreeBSD_version >= 1400026
477         iwqp->push_db_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
478                                                                 bar_off, IRDMA_MMAP_IO_NC,
479                                                                 push_db_mmap_key);
480 #else
481
482         iwqp->push_db_mmap_entry = irdma_user_mmap_entry_add_hash(ucontext, bar_off,
483                                                                   IRDMA_MMAP_IO_NC,
484                                                                   push_db_mmap_key);
485 #endif
486         if (!iwqp->push_db_mmap_entry) {
487 #if __FreeBSD_version >= 1400026
488                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
489 #else
490                 irdma_user_mmap_entry_del_hash(iwqp->push_wqe_mmap_entry);
491 #endif
492                 return -ENOMEM;
493         }
494
495         return 0;
496 }
497
498 /**
499  * irdma_setup_virt_qp - setup for allocation of virtual qp
500  * @iwdev: irdma device
501  * @iwqp: qp ptr
502  * @init_info: initialize info to return
503  */
504 void
505 irdma_setup_virt_qp(struct irdma_device *iwdev,
506                     struct irdma_qp *iwqp,
507                     struct irdma_qp_init_info *init_info)
508 {
509         struct irdma_pbl *iwpbl = iwqp->iwpbl;
510         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
511
512         iwqp->page = qpmr->sq_page;
513         init_info->shadow_area_pa = qpmr->shadow;
514         if (iwpbl->pbl_allocated) {
515                 init_info->virtual_map = true;
516                 init_info->sq_pa = qpmr->sq_pbl.idx;
517                 init_info->rq_pa = qpmr->rq_pbl.idx;
518         } else {
519                 init_info->sq_pa = qpmr->sq_pbl.addr;
520                 init_info->rq_pa = qpmr->rq_pbl.addr;
521         }
522 }
523
524 /**
525  * irdma_setup_umode_qp - setup sq and rq size in user mode qp
526  * @udata: user data
527  * @iwdev: iwarp device
528  * @iwqp: qp ptr (user or kernel)
529  * @info: initialize info to return
530  * @init_attr: Initial QP create attributes
531  */
532 int
533 irdma_setup_umode_qp(struct ib_udata *udata,
534                      struct irdma_device *iwdev,
535                      struct irdma_qp *iwqp,
536                      struct irdma_qp_init_info *info,
537                      struct ib_qp_init_attr *init_attr)
538 {
539 #if __FreeBSD_version >= 1400026
540         struct irdma_ucontext *ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
541 #else
542         struct irdma_ucontext *ucontext = to_ucontext(iwqp->iwpd->ibpd.uobject->context);
543 #endif
544         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
545         struct irdma_create_qp_req req = {0};
546         unsigned long flags;
547         int ret;
548
549         ret = ib_copy_from_udata(&req, udata,
550                                  min(sizeof(req), udata->inlen));
551         if (ret) {
552                 irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
553                             "ib_copy_from_data fail\n");
554                 return ret;
555         }
556
557         iwqp->ctx_info.qp_compl_ctx = req.user_compl_ctx;
558         iwqp->user_mode = 1;
559         if (req.user_wqe_bufs) {
560                 info->qp_uk_init_info.legacy_mode = ucontext->legacy_mode;
561                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
562                 iwqp->iwpbl = irdma_get_pbl((unsigned long)req.user_wqe_bufs,
563                                             &ucontext->qp_reg_mem_list);
564                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
565
566                 if (!iwqp->iwpbl) {
567                         ret = -ENODATA;
568                         irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
569                                     "no pbl info\n");
570                         return ret;
571                 }
572         }
573
574         if (!ucontext->use_raw_attrs) {
575                 /**
576                  * Maintain backward compat with older ABI which passes sq and
577                  * rq depth in quanta in cap.max_send_wr and cap.max_recv_wr.
578                  * There is no way to compute the correct value of
579                  * iwqp->max_send_wr/max_recv_wr in the kernel.
580                  */
581                 iwqp->max_send_wr = init_attr->cap.max_send_wr;
582                 iwqp->max_recv_wr = init_attr->cap.max_recv_wr;
583                 ukinfo->sq_size = init_attr->cap.max_send_wr;
584                 ukinfo->rq_size = init_attr->cap.max_recv_wr;
585                 irdma_uk_calc_shift_wq(ukinfo, &ukinfo->sq_shift, &ukinfo->rq_shift);
586         } else {
587                 ret = irdma_uk_calc_depth_shift_sq(ukinfo, &ukinfo->sq_depth,
588                                                    &ukinfo->sq_shift);
589                 if (ret)
590                         return ret;
591
592                 ret = irdma_uk_calc_depth_shift_rq(ukinfo, &ukinfo->rq_depth,
593                                                    &ukinfo->rq_shift);
594                 if (ret)
595                         return ret;
596
597                 iwqp->max_send_wr = (ukinfo->sq_depth - IRDMA_SQ_RSVD) >> ukinfo->sq_shift;
598                 iwqp->max_recv_wr = (ukinfo->rq_depth - IRDMA_RQ_RSVD) >> ukinfo->rq_shift;
599                 ukinfo->sq_size = ukinfo->sq_depth >> ukinfo->sq_shift;
600                 ukinfo->rq_size = ukinfo->rq_depth >> ukinfo->rq_shift;
601         }
602         irdma_setup_virt_qp(iwdev, iwqp, info);
603
604         return 0;
605 }
606
607 /**
608  * irdma_setup_kmode_qp - setup initialization for kernel mode qp
609  * @iwdev: iwarp device
610  * @iwqp: qp ptr (user or kernel)
611  * @info: initialize info to return
612  * @init_attr: Initial QP create attributes
613  */
614 int
615 irdma_setup_kmode_qp(struct irdma_device *iwdev,
616                      struct irdma_qp *iwqp,
617                      struct irdma_qp_init_info *info,
618                      struct ib_qp_init_attr *init_attr)
619 {
620         struct irdma_dma_mem *mem = &iwqp->kqp.dma_mem;
621         u32 size;
622         int status;
623         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
624
625         status = irdma_uk_calc_depth_shift_sq(ukinfo, &ukinfo->sq_depth,
626                                               &ukinfo->sq_shift);
627         if (status)
628                 return status;
629
630         status = irdma_uk_calc_depth_shift_rq(ukinfo, &ukinfo->rq_depth,
631                                               &ukinfo->rq_shift);
632         if (status)
633                 return status;
634
635         iwqp->kqp.sq_wrid_mem =
636             kcalloc(ukinfo->sq_depth, sizeof(*iwqp->kqp.sq_wrid_mem), GFP_KERNEL);
637         if (!iwqp->kqp.sq_wrid_mem)
638                 return -ENOMEM;
639
640         iwqp->kqp.rq_wrid_mem =
641             kcalloc(ukinfo->rq_depth, sizeof(*iwqp->kqp.rq_wrid_mem), GFP_KERNEL);
642         if (!iwqp->kqp.rq_wrid_mem) {
643                 kfree(iwqp->kqp.sq_wrid_mem);
644                 iwqp->kqp.sq_wrid_mem = NULL;
645                 return -ENOMEM;
646         }
647
648         iwqp->kqp.sig_trk_mem = kcalloc(ukinfo->sq_depth, sizeof(u32), GFP_KERNEL);
649         memset(iwqp->kqp.sig_trk_mem, 0, ukinfo->sq_depth * sizeof(u32));
650         if (!iwqp->kqp.sig_trk_mem) {
651                 kfree(iwqp->kqp.sq_wrid_mem);
652                 iwqp->kqp.sq_wrid_mem = NULL;
653                 kfree(iwqp->kqp.rq_wrid_mem);
654                 iwqp->kqp.rq_wrid_mem = NULL;
655                 return -ENOMEM;
656         }
657         ukinfo->sq_sigwrtrk_array = (void *)iwqp->kqp.sig_trk_mem;
658         ukinfo->sq_wrtrk_array = iwqp->kqp.sq_wrid_mem;
659         ukinfo->rq_wrid_array = iwqp->kqp.rq_wrid_mem;
660
661         size = (ukinfo->sq_depth + ukinfo->rq_depth) * IRDMA_QP_WQE_MIN_SIZE;
662         size += (IRDMA_SHADOW_AREA_SIZE << 3);
663
664         mem->size = size;
665         mem->va = irdma_allocate_dma_mem(&iwdev->rf->hw, mem, mem->size,
666                                          256);
667         if (!mem->va) {
668                 kfree(iwqp->kqp.sq_wrid_mem);
669                 iwqp->kqp.sq_wrid_mem = NULL;
670                 kfree(iwqp->kqp.rq_wrid_mem);
671                 iwqp->kqp.rq_wrid_mem = NULL;
672                 return -ENOMEM;
673         }
674
675         ukinfo->sq = mem->va;
676         info->sq_pa = mem->pa;
677         ukinfo->rq = &ukinfo->sq[ukinfo->sq_depth];
678         info->rq_pa = info->sq_pa + (ukinfo->sq_depth * IRDMA_QP_WQE_MIN_SIZE);
679         ukinfo->shadow_area = ukinfo->rq[ukinfo->rq_depth].elem;
680         info->shadow_area_pa = info->rq_pa + (ukinfo->rq_depth * IRDMA_QP_WQE_MIN_SIZE);
681         ukinfo->sq_size = ukinfo->sq_depth >> ukinfo->sq_shift;
682         ukinfo->rq_size = ukinfo->rq_depth >> ukinfo->rq_shift;
683         ukinfo->qp_id = iwqp->ibqp.qp_num;
684
685         iwqp->max_send_wr = (ukinfo->sq_depth - IRDMA_SQ_RSVD) >> ukinfo->sq_shift;
686         iwqp->max_recv_wr = (ukinfo->rq_depth - IRDMA_RQ_RSVD) >> ukinfo->rq_shift;
687         init_attr->cap.max_send_wr = iwqp->max_send_wr;
688         init_attr->cap.max_recv_wr = iwqp->max_recv_wr;
689
690         return 0;
691 }
692
693 int
694 irdma_cqp_create_qp_cmd(struct irdma_qp *iwqp)
695 {
696         struct irdma_pci_f *rf = iwqp->iwdev->rf;
697         struct irdma_cqp_request *cqp_request;
698         struct cqp_cmds_info *cqp_info;
699         struct irdma_create_qp_info *qp_info;
700         int status;
701
702         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
703         if (!cqp_request)
704                 return -ENOMEM;
705
706         cqp_info = &cqp_request->info;
707         qp_info = &cqp_request->info.in.u.qp_create.info;
708         memset(qp_info, 0, sizeof(*qp_info));
709         qp_info->mac_valid = true;
710         qp_info->cq_num_valid = true;
711         qp_info->next_iwarp_state = IRDMA_QP_STATE_IDLE;
712
713         cqp_info->cqp_cmd = IRDMA_OP_QP_CREATE;
714         cqp_info->post_sq = 1;
715         cqp_info->in.u.qp_create.qp = &iwqp->sc_qp;
716         cqp_info->in.u.qp_create.scratch = (uintptr_t)cqp_request;
717         status = irdma_handle_cqp_op(rf, cqp_request);
718         irdma_put_cqp_request(&rf->cqp, cqp_request);
719
720         return status;
721 }
722
723 void
724 irdma_roce_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
725                                    struct irdma_qp_host_ctx_info *ctx_info)
726 {
727         struct irdma_device *iwdev = iwqp->iwdev;
728         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
729         struct irdma_roce_offload_info *roce_info;
730         struct irdma_udp_offload_info *udp_info;
731
732         udp_info = &iwqp->udp_info;
733         udp_info->snd_mss = ib_mtu_enum_to_int(ib_mtu_int_to_enum(iwdev->vsi.mtu));
734         udp_info->cwnd = iwdev->roce_cwnd;
735         udp_info->rexmit_thresh = 2;
736         udp_info->rnr_nak_thresh = 2;
737         udp_info->src_port = 0xc000;
738         udp_info->dst_port = ROCE_V2_UDP_DPORT;
739         roce_info = &iwqp->roce_info;
740         ether_addr_copy(roce_info->mac_addr, if_getlladdr(iwdev->netdev));
741
742         roce_info->rd_en = true;
743         roce_info->wr_rdresp_en = true;
744         roce_info->bind_en = true;
745         roce_info->dcqcn_en = false;
746         roce_info->rtomin = iwdev->roce_rtomin;
747
748         roce_info->ack_credits = iwdev->roce_ackcreds;
749         roce_info->ird_size = dev->hw_attrs.max_hw_ird;
750         roce_info->ord_size = dev->hw_attrs.max_hw_ord;
751
752         if (!iwqp->user_mode) {
753                 roce_info->priv_mode_en = true;
754                 roce_info->fast_reg_en = true;
755                 roce_info->udprivcq_en = true;
756         }
757         roce_info->roce_tver = 0;
758
759         ctx_info->roce_info = &iwqp->roce_info;
760         ctx_info->udp_info = &iwqp->udp_info;
761         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
762 }
763
764 void
765 irdma_iw_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
766                                  struct irdma_qp_host_ctx_info *ctx_info)
767 {
768         struct irdma_device *iwdev = iwqp->iwdev;
769         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
770         struct irdma_iwarp_offload_info *iwarp_info;
771
772         iwarp_info = &iwqp->iwarp_info;
773         ether_addr_copy(iwarp_info->mac_addr, if_getlladdr(iwdev->netdev));
774         iwarp_info->rd_en = true;
775         iwarp_info->wr_rdresp_en = true;
776         iwarp_info->bind_en = true;
777         iwarp_info->ecn_en = true;
778         iwarp_info->rtomin = 5;
779
780         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
781                 iwarp_info->ib_rd_en = true;
782         if (!iwqp->user_mode) {
783                 iwarp_info->priv_mode_en = true;
784                 iwarp_info->fast_reg_en = true;
785         }
786         iwarp_info->ddp_ver = 1;
787         iwarp_info->rdmap_ver = 1;
788
789         ctx_info->iwarp_info = &iwqp->iwarp_info;
790         ctx_info->iwarp_info_valid = true;
791         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
792         ctx_info->iwarp_info_valid = false;
793 }
794
795 int
796 irdma_validate_qp_attrs(struct ib_qp_init_attr *init_attr,
797                         struct irdma_device *iwdev)
798 {
799         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
800         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
801
802         if (init_attr->create_flags)
803                 return -EOPNOTSUPP;
804
805         if (init_attr->cap.max_inline_data > uk_attrs->max_hw_inline ||
806             init_attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags ||
807             init_attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags)
808                 return -EINVAL;
809
810         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
811                 if (init_attr->qp_type != IB_QPT_RC &&
812                     init_attr->qp_type != IB_QPT_UD &&
813                     init_attr->qp_type != IB_QPT_GSI)
814                         return -EOPNOTSUPP;
815         } else {
816                 if (init_attr->qp_type != IB_QPT_RC)
817                         return -EOPNOTSUPP;
818         }
819
820         return 0;
821 }
822
823 void
824 irdma_sched_qp_flush_work(struct irdma_qp *iwqp)
825 {
826         if (iwqp->sc_qp.qp_uk.destroy_pending)
827                 return;
828         irdma_qp_add_ref(&iwqp->ibqp);
829         if (mod_delayed_work(iwqp->iwdev->cleanup_wq, &iwqp->dwork_flush,
830                              msecs_to_jiffies(IRDMA_FLUSH_DELAY_MS)))
831                 irdma_qp_rem_ref(&iwqp->ibqp);
832 }
833
834 void
835 irdma_flush_worker(struct work_struct *work)
836 {
837         struct delayed_work *dwork = to_delayed_work(work);
838         struct irdma_qp *iwqp = container_of(dwork, struct irdma_qp, dwork_flush);
839
840         irdma_generate_flush_completions(iwqp);
841         /* For the add in irdma_sched_qp_flush_work */
842         irdma_qp_rem_ref(&iwqp->ibqp);
843 }
844
845 static int
846 irdma_get_ib_acc_flags(struct irdma_qp *iwqp)
847 {
848         int acc_flags = 0;
849
850         if (rdma_protocol_roce(iwqp->ibqp.device, 1)) {
851                 if (iwqp->roce_info.wr_rdresp_en) {
852                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
853                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
854                 }
855                 if (iwqp->roce_info.rd_en)
856                         acc_flags |= IB_ACCESS_REMOTE_READ;
857                 if (iwqp->roce_info.bind_en)
858                         acc_flags |= IB_ACCESS_MW_BIND;
859         } else {
860                 if (iwqp->iwarp_info.wr_rdresp_en) {
861                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
862                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
863                 }
864                 if (iwqp->iwarp_info.rd_en)
865                         acc_flags |= IB_ACCESS_REMOTE_READ;
866                 if (iwqp->iwarp_info.bind_en)
867                         acc_flags |= IB_ACCESS_MW_BIND;
868         }
869         return acc_flags;
870 }
871
872 /**
873  * irdma_query_qp - query qp attributes
874  * @ibqp: qp pointer
875  * @attr: attributes pointer
876  * @attr_mask: Not used
877  * @init_attr: qp attributes to return
878  */
879 static int
880 irdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
881                int attr_mask, struct ib_qp_init_attr *init_attr)
882 {
883         struct irdma_qp *iwqp = to_iwqp(ibqp);
884         struct irdma_sc_qp *qp = &iwqp->sc_qp;
885
886         memset(attr, 0, sizeof(*attr));
887         memset(init_attr, 0, sizeof(*init_attr));
888
889         attr->qp_state = iwqp->ibqp_state;
890         attr->cur_qp_state = iwqp->ibqp_state;
891         attr->cap.max_send_wr = iwqp->max_send_wr;
892         attr->cap.max_recv_wr = iwqp->max_recv_wr;
893         attr->cap.max_inline_data = qp->qp_uk.max_inline_data;
894         attr->cap.max_send_sge = qp->qp_uk.max_sq_frag_cnt;
895         attr->cap.max_recv_sge = qp->qp_uk.max_rq_frag_cnt;
896         attr->qp_access_flags = irdma_get_ib_acc_flags(iwqp);
897         attr->port_num = 1;
898         if (rdma_protocol_roce(ibqp->device, 1)) {
899                 attr->path_mtu = ib_mtu_int_to_enum(iwqp->udp_info.snd_mss);
900                 attr->qkey = iwqp->roce_info.qkey;
901                 attr->rq_psn = iwqp->udp_info.epsn;
902                 attr->sq_psn = iwqp->udp_info.psn_nxt;
903                 attr->dest_qp_num = iwqp->roce_info.dest_qp;
904                 attr->pkey_index = iwqp->roce_info.p_key;
905                 attr->retry_cnt = iwqp->udp_info.rexmit_thresh;
906                 attr->rnr_retry = iwqp->udp_info.rnr_nak_thresh;
907                 attr->max_rd_atomic = iwqp->roce_info.ord_size;
908                 attr->max_dest_rd_atomic = iwqp->roce_info.ird_size;
909         }
910
911         init_attr->event_handler = iwqp->ibqp.event_handler;
912         init_attr->qp_context = iwqp->ibqp.qp_context;
913         init_attr->send_cq = iwqp->ibqp.send_cq;
914         init_attr->recv_cq = iwqp->ibqp.recv_cq;
915         init_attr->cap = attr->cap;
916
917         return 0;
918 }
919
920 /**
921  * irdma_modify_qp_roce - modify qp request
922  * @ibqp: qp's pointer for modify
923  * @attr: access attributes
924  * @attr_mask: state mask
925  * @udata: user data
926  */
927 int
928 irdma_modify_qp_roce(struct ib_qp *ibqp, struct ib_qp_attr *attr,
929                      int attr_mask, struct ib_udata *udata)
930 {
931 #define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
932 #define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
933         struct irdma_pd *iwpd = to_iwpd(ibqp->pd);
934         struct irdma_qp *iwqp = to_iwqp(ibqp);
935         struct irdma_device *iwdev = iwqp->iwdev;
936         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
937         struct irdma_qp_host_ctx_info *ctx_info;
938         struct irdma_roce_offload_info *roce_info;
939         struct irdma_udp_offload_info *udp_info;
940         struct irdma_modify_qp_info info = {0};
941         struct irdma_modify_qp_resp uresp = {};
942         struct irdma_modify_qp_req ureq;
943         unsigned long flags;
944         u8 issue_modify_qp = 0;
945         int ret = 0;
946
947         ctx_info = &iwqp->ctx_info;
948         roce_info = &iwqp->roce_info;
949         udp_info = &iwqp->udp_info;
950
951         if (udata) {
952                 if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
953                     (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
954                         return -EINVAL;
955         }
956
957         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
958                 return -EOPNOTSUPP;
959
960         if (attr_mask & IB_QP_DEST_QPN)
961                 roce_info->dest_qp = attr->dest_qp_num;
962
963         if (attr_mask & IB_QP_PKEY_INDEX) {
964                 ret = irdma_query_pkey(ibqp->device, 0, attr->pkey_index,
965                                        &roce_info->p_key);
966                 if (ret)
967                         return ret;
968         }
969
970         if (attr_mask & IB_QP_QKEY)
971                 roce_info->qkey = attr->qkey;
972
973         if (attr_mask & IB_QP_PATH_MTU)
974                 udp_info->snd_mss = ib_mtu_enum_to_int(attr->path_mtu);
975
976         if (attr_mask & IB_QP_SQ_PSN) {
977                 udp_info->psn_nxt = attr->sq_psn;
978                 udp_info->lsn = 0xffff;
979                 udp_info->psn_una = attr->sq_psn;
980                 udp_info->psn_max = attr->sq_psn;
981         }
982
983         if (attr_mask & IB_QP_RQ_PSN)
984                 udp_info->epsn = attr->rq_psn;
985
986         if (attr_mask & IB_QP_RNR_RETRY)
987                 udp_info->rnr_nak_thresh = attr->rnr_retry;
988
989         if (attr_mask & IB_QP_RETRY_CNT)
990                 udp_info->rexmit_thresh = attr->retry_cnt;
991
992         ctx_info->roce_info->pd_id = iwpd->sc_pd.pd_id;
993
994         if (attr_mask & IB_QP_AV) {
995                 struct irdma_av *av = &iwqp->roce_ah.av;
996                 u16 vlan_id = VLAN_N_VID;
997                 u32 local_ip[4] = {};
998
999                 memset(&iwqp->roce_ah, 0, sizeof(iwqp->roce_ah));
1000                 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
1001                         udp_info->ttl = attr->ah_attr.grh.hop_limit;
1002                         udp_info->flow_label = attr->ah_attr.grh.flow_label;
1003                         udp_info->tos = attr->ah_attr.grh.traffic_class;
1004
1005                         udp_info->src_port = kc_rdma_get_udp_sport(udp_info->flow_label,
1006                                                                    ibqp->qp_num,
1007                                                                    roce_info->dest_qp);
1008
1009                         irdma_qp_rem_qos(&iwqp->sc_qp);
1010                         dev->ws_remove(iwqp->sc_qp.vsi, ctx_info->user_pri);
1011                         if (iwqp->sc_qp.vsi->dscp_mode)
1012                                 ctx_info->user_pri =
1013                                     iwqp->sc_qp.vsi->dscp_map[irdma_tos2dscp(udp_info->tos)];
1014                         else
1015                                 ctx_info->user_pri = rt_tos2priority(udp_info->tos);
1016                 }
1017                 ret = kc_irdma_set_roce_cm_info(iwqp, attr, &vlan_id);
1018                 if (ret)
1019                         return ret;
1020                 if (dev->ws_add(iwqp->sc_qp.vsi, ctx_info->user_pri))
1021                         return -ENOMEM;
1022                 iwqp->sc_qp.user_pri = ctx_info->user_pri;
1023                 irdma_qp_add_qos(&iwqp->sc_qp);
1024
1025                 if (vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
1026                         vlan_id = 0;
1027                 if (vlan_id < VLAN_N_VID) {
1028                         udp_info->insert_vlan_tag = true;
1029                         udp_info->vlan_tag = vlan_id |
1030                             ctx_info->user_pri << VLAN_PRIO_SHIFT;
1031                 } else {
1032                         udp_info->insert_vlan_tag = false;
1033                 }
1034
1035                 av->attrs = attr->ah_attr;
1036                 rdma_gid2ip((struct sockaddr *)&av->dgid_addr, &attr->ah_attr.grh.dgid);
1037                 if (av->net_type == RDMA_NETWORK_IPV6) {
1038                         __be32 *daddr =
1039                         av->dgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32;
1040                         __be32 *saddr =
1041                         av->sgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32;
1042
1043                         irdma_copy_ip_ntohl(&udp_info->dest_ip_addr[0], daddr);
1044                         irdma_copy_ip_ntohl(&udp_info->local_ipaddr[0], saddr);
1045
1046                         udp_info->ipv4 = false;
1047                         irdma_copy_ip_ntohl(local_ip, daddr);
1048                 } else if (av->net_type == RDMA_NETWORK_IPV4) {
1049                         __be32 saddr = av->sgid_addr.saddr_in.sin_addr.s_addr;
1050                         __be32 daddr = av->dgid_addr.saddr_in.sin_addr.s_addr;
1051
1052                         local_ip[0] = ntohl(daddr);
1053
1054                         udp_info->ipv4 = true;
1055                         udp_info->dest_ip_addr[0] = 0;
1056                         udp_info->dest_ip_addr[1] = 0;
1057                         udp_info->dest_ip_addr[2] = 0;
1058                         udp_info->dest_ip_addr[3] = local_ip[0];
1059
1060                         udp_info->local_ipaddr[0] = 0;
1061                         udp_info->local_ipaddr[1] = 0;
1062                         udp_info->local_ipaddr[2] = 0;
1063                         udp_info->local_ipaddr[3] = ntohl(saddr);
1064                 } else {
1065                         return -EINVAL;
1066                 }
1067                 udp_info->arp_idx =
1068                     irdma_add_arp(iwdev->rf, local_ip,
1069                                   ah_attr_to_dmac(attr->ah_attr));
1070         }
1071
1072         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1073                 if (attr->max_rd_atomic > dev->hw_attrs.max_hw_ord) {
1074                         irdma_dev_err(&iwdev->ibdev,
1075                                       "rd_atomic = %d, above max_hw_ord=%d\n",
1076                                       attr->max_rd_atomic,
1077                                       dev->hw_attrs.max_hw_ord);
1078                         return -EINVAL;
1079                 }
1080                 if (attr->max_rd_atomic)
1081                         roce_info->ord_size = attr->max_rd_atomic;
1082                 info.ord_valid = true;
1083         }
1084
1085         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1086                 if (attr->max_dest_rd_atomic > dev->hw_attrs.max_hw_ird) {
1087                         irdma_dev_err(&iwdev->ibdev,
1088                                       "rd_atomic = %d, above max_hw_ird=%d\n",
1089                                       attr->max_rd_atomic,
1090                                       dev->hw_attrs.max_hw_ird);
1091                         return -EINVAL;
1092                 }
1093                 if (attr->max_dest_rd_atomic)
1094                         roce_info->ird_size = attr->max_dest_rd_atomic;
1095         }
1096
1097         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1098                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1099                         roce_info->wr_rdresp_en = true;
1100                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1101                         roce_info->wr_rdresp_en = true;
1102                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1103                         roce_info->rd_en = true;
1104         }
1105
1106         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1107
1108         irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
1109                     "caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d attr_mask=0x%x\n",
1110                     __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1111                     iwqp->ibqp_state, iwqp->iwarp_state, attr_mask);
1112
1113         spin_lock_irqsave(&iwqp->lock, flags);
1114         if (attr_mask & IB_QP_STATE) {
1115                 if (!kc_ib_modify_qp_is_ok(iwqp->ibqp_state, attr->qp_state,
1116                                            iwqp->ibqp.qp_type, attr_mask,
1117                                            IB_LINK_LAYER_ETHERNET)) {
1118                         irdma_dev_warn(&iwdev->ibdev,
1119                                        "modify_qp invalid for qp_id=%d, old_state=0x%x, new_state=0x%x\n",
1120                                        iwqp->ibqp.qp_num, iwqp->ibqp_state,
1121                                        attr->qp_state);
1122                         ret = -EINVAL;
1123                         goto exit;
1124                 }
1125                 info.curr_iwarp_state = iwqp->iwarp_state;
1126
1127                 switch (attr->qp_state) {
1128                 case IB_QPS_INIT:
1129                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1130                                 ret = -EINVAL;
1131                                 goto exit;
1132                         }
1133
1134                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1135                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1136                                 issue_modify_qp = 1;
1137                         }
1138                         break;
1139                 case IB_QPS_RTR:
1140                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1141                                 ret = -EINVAL;
1142                                 goto exit;
1143                         }
1144                         info.arp_cache_idx_valid = true;
1145                         info.cq_num_valid = true;
1146                         info.next_iwarp_state = IRDMA_QP_STATE_RTR;
1147                         issue_modify_qp = 1;
1148                         break;
1149                 case IB_QPS_RTS:
1150                         if (iwqp->ibqp_state < IB_QPS_RTR ||
1151                             iwqp->ibqp_state == IB_QPS_ERR) {
1152                                 ret = -EINVAL;
1153                                 goto exit;
1154                         }
1155
1156                         info.arp_cache_idx_valid = true;
1157                         info.cq_num_valid = true;
1158                         info.ord_valid = true;
1159                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1160                         issue_modify_qp = 1;
1161                         if (dev->hw_attrs.uk_attrs.hw_rev == IRDMA_GEN_2)
1162                                 iwdev->rf->check_fc(&iwdev->vsi, &iwqp->sc_qp);
1163                         udp_info->cwnd = iwdev->roce_cwnd;
1164                         roce_info->ack_credits = iwdev->roce_ackcreds;
1165                         if (iwdev->push_mode && udata &&
1166                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1167                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1168                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1169                                 irdma_alloc_push_page(iwqp);
1170                                 spin_lock_irqsave(&iwqp->lock, flags);
1171                         }
1172                         break;
1173                 case IB_QPS_SQD:
1174                         if (iwqp->iwarp_state == IRDMA_QP_STATE_SQD)
1175                                 goto exit;
1176
1177                         if (iwqp->iwarp_state != IRDMA_QP_STATE_RTS) {
1178                                 ret = -EINVAL;
1179                                 goto exit;
1180                         }
1181
1182                         info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1183                         issue_modify_qp = 1;
1184                         break;
1185                 case IB_QPS_SQE:
1186                 case IB_QPS_ERR:
1187                 case IB_QPS_RESET:
1188                         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS) {
1189                                 if (dev->hw_attrs.uk_attrs.hw_rev <= IRDMA_GEN_2)
1190                                         irdma_cqp_qp_suspend_resume(&iwqp->sc_qp, IRDMA_OP_SUSPEND);
1191                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1192                                 info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1193                                 irdma_hw_modify_qp(iwdev, iwqp, &info, true);
1194                                 spin_lock_irqsave(&iwqp->lock, flags);
1195                         }
1196
1197                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1198                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1199                                 if (udata && udata->inlen) {
1200                                         if (ib_copy_from_udata(&ureq, udata,
1201                                                                min(sizeof(ureq), udata->inlen)))
1202                                                 return -EINVAL;
1203
1204                                         irdma_flush_wqes(iwqp,
1205                                                          (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1206                                                          (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1207                                                          IRDMA_REFLUSH);
1208                                 }
1209                                 return 0;
1210                         }
1211
1212                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1213                         issue_modify_qp = 1;
1214                         break;
1215                 default:
1216                         ret = -EINVAL;
1217                         goto exit;
1218                 }
1219
1220                 iwqp->ibqp_state = attr->qp_state;
1221         }
1222
1223         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1224         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1225         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1226         spin_unlock_irqrestore(&iwqp->lock, flags);
1227
1228         if (attr_mask & IB_QP_STATE) {
1229                 if (issue_modify_qp) {
1230                         ctx_info->rem_endpoint_idx = udp_info->arp_idx;
1231                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1232                                 return -EINVAL;
1233                         spin_lock_irqsave(&iwqp->lock, flags);
1234                         if (iwqp->iwarp_state == info.curr_iwarp_state) {
1235                                 iwqp->iwarp_state = info.next_iwarp_state;
1236                                 iwqp->ibqp_state = attr->qp_state;
1237                         }
1238                         if (iwqp->ibqp_state > IB_QPS_RTS &&
1239                             !iwqp->flush_issued) {
1240                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1241                                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ |
1242                                                  IRDMA_FLUSH_RQ |
1243                                                  IRDMA_FLUSH_WAIT);
1244                                 iwqp->flush_issued = 1;
1245
1246                         } else {
1247                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1248                         }
1249                 } else {
1250                         iwqp->ibqp_state = attr->qp_state;
1251                 }
1252                 if (udata && udata->outlen && dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1253                         struct irdma_ucontext *ucontext;
1254
1255 #if __FreeBSD_version >= 1400026
1256                         ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
1257 #else
1258                         ucontext = to_ucontext(ibqp->uobject->context);
1259 #endif
1260                         if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1261                             !iwqp->push_wqe_mmap_entry &&
1262                             !irdma_setup_push_mmap_entries(ucontext, iwqp,
1263                                                            &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1264                                 uresp.push_valid = 1;
1265                                 uresp.push_offset = iwqp->sc_qp.push_offset;
1266                         }
1267                         uresp.rd_fence_rate = iwdev->rd_fence_rate;
1268                         ret = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1269                                                                   udata->outlen));
1270                         if (ret) {
1271                                 irdma_remove_push_mmap_entries(iwqp);
1272                                 irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
1273                                             "copy_to_udata failed\n");
1274                                 return ret;
1275                         }
1276                 }
1277         }
1278
1279         return 0;
1280 exit:
1281         spin_unlock_irqrestore(&iwqp->lock, flags);
1282
1283         return ret;
1284 }
1285
1286 /**
1287  * irdma_modify_qp - modify qp request
1288  * @ibqp: qp's pointer for modify
1289  * @attr: access attributes
1290  * @attr_mask: state mask
1291  * @udata: user data
1292  */
1293 int
1294 irdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
1295                 struct ib_udata *udata)
1296 {
1297 #define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
1298 #define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
1299         struct irdma_qp *iwqp = to_iwqp(ibqp);
1300         struct irdma_device *iwdev = iwqp->iwdev;
1301         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1302         struct irdma_qp_host_ctx_info *ctx_info;
1303         struct irdma_tcp_offload_info *tcp_info;
1304         struct irdma_iwarp_offload_info *offload_info;
1305         struct irdma_modify_qp_info info = {0};
1306         struct irdma_modify_qp_resp uresp = {};
1307         struct irdma_modify_qp_req ureq = {};
1308         u8 issue_modify_qp = 0;
1309         u8 dont_wait = 0;
1310         int err;
1311         unsigned long flags;
1312
1313         if (udata) {
1314                 if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
1315                     (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
1316                         return -EINVAL;
1317         }
1318
1319         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1320                 return -EOPNOTSUPP;
1321
1322         ctx_info = &iwqp->ctx_info;
1323         offload_info = &iwqp->iwarp_info;
1324         tcp_info = &iwqp->tcp_info;
1325         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1326         irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
1327                     "caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d last_aeq=%d hw_tcp_state=%d hw_iwarp_state=%d attr_mask=0x%x\n",
1328                     __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1329                     iwqp->ibqp_state, iwqp->iwarp_state, iwqp->last_aeq,
1330                     iwqp->hw_tcp_state, iwqp->hw_iwarp_state, attr_mask);
1331
1332         spin_lock_irqsave(&iwqp->lock, flags);
1333         if (attr_mask & IB_QP_STATE) {
1334                 info.curr_iwarp_state = iwqp->iwarp_state;
1335                 switch (attr->qp_state) {
1336                 case IB_QPS_INIT:
1337                 case IB_QPS_RTR:
1338                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1339                                 err = -EINVAL;
1340                                 goto exit;
1341                         }
1342
1343                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1344                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1345                                 issue_modify_qp = 1;
1346                         }
1347                         if (iwdev->push_mode && udata &&
1348                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1349                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1350                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1351                                 irdma_alloc_push_page(iwqp);
1352                                 spin_lock_irqsave(&iwqp->lock, flags);
1353                         }
1354                         break;
1355                 case IB_QPS_RTS:
1356                         if (iwqp->iwarp_state > IRDMA_QP_STATE_RTS ||
1357                             !iwqp->cm_id) {
1358                                 err = -EINVAL;
1359                                 goto exit;
1360                         }
1361
1362                         issue_modify_qp = 1;
1363                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
1364                         iwqp->hte_added = 1;
1365                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1366                         info.tcp_ctx_valid = true;
1367                         info.ord_valid = true;
1368                         info.arp_cache_idx_valid = true;
1369                         info.cq_num_valid = true;
1370                         break;
1371                 case IB_QPS_SQD:
1372                         if (iwqp->hw_iwarp_state > IRDMA_QP_STATE_RTS) {
1373                                 err = 0;
1374                                 goto exit;
1375                         }
1376
1377                         if (iwqp->iwarp_state == IRDMA_QP_STATE_CLOSING ||
1378                             iwqp->iwarp_state < IRDMA_QP_STATE_RTS) {
1379                                 err = 0;
1380                                 goto exit;
1381                         }
1382
1383                         if (iwqp->iwarp_state > IRDMA_QP_STATE_CLOSING) {
1384                                 err = -EINVAL;
1385                                 goto exit;
1386                         }
1387
1388                         info.next_iwarp_state = IRDMA_QP_STATE_CLOSING;
1389                         issue_modify_qp = 1;
1390                         break;
1391                 case IB_QPS_SQE:
1392                         if (iwqp->iwarp_state >= IRDMA_QP_STATE_TERMINATE) {
1393                                 err = -EINVAL;
1394                                 goto exit;
1395                         }
1396
1397                         info.next_iwarp_state = IRDMA_QP_STATE_TERMINATE;
1398                         issue_modify_qp = 1;
1399                         break;
1400                 case IB_QPS_ERR:
1401                 case IB_QPS_RESET:
1402                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1403                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1404                                 if (udata && udata->inlen) {
1405                                         if (ib_copy_from_udata(&ureq, udata,
1406                                                                min(sizeof(ureq), udata->inlen)))
1407                                                 return -EINVAL;
1408
1409                                         irdma_flush_wqes(iwqp,
1410                                                          (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1411                                                          (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1412                                                          IRDMA_REFLUSH);
1413                                 }
1414                                 return 0;
1415                         }
1416
1417                         if (iwqp->sc_qp.term_flags) {
1418                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1419                                 irdma_terminate_del_timer(&iwqp->sc_qp);
1420                                 spin_lock_irqsave(&iwqp->lock, flags);
1421                         }
1422                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1423                         if (iwqp->hw_tcp_state > IRDMA_TCP_STATE_CLOSED &&
1424                             iwdev->iw_status &&
1425                             iwqp->hw_tcp_state != IRDMA_TCP_STATE_TIME_WAIT)
1426                                 info.reset_tcp_conn = true;
1427                         else
1428                                 dont_wait = 1;
1429
1430                         issue_modify_qp = 1;
1431                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1432                         break;
1433                 default:
1434                         err = -EINVAL;
1435                         goto exit;
1436                 }
1437
1438                 iwqp->ibqp_state = attr->qp_state;
1439         }
1440         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1441                 ctx_info->iwarp_info_valid = true;
1442                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1443                         offload_info->wr_rdresp_en = true;
1444                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1445                         offload_info->wr_rdresp_en = true;
1446                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1447                         offload_info->rd_en = true;
1448         }
1449
1450         if (ctx_info->iwarp_info_valid) {
1451                 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1452                 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1453                 irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1454         }
1455         spin_unlock_irqrestore(&iwqp->lock, flags);
1456
1457         if (attr_mask & IB_QP_STATE) {
1458                 if (issue_modify_qp) {
1459                         ctx_info->rem_endpoint_idx = tcp_info->arp_idx;
1460                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1461                                 return -EINVAL;
1462                 }
1463
1464                 spin_lock_irqsave(&iwqp->lock, flags);
1465                 if (iwqp->iwarp_state == info.curr_iwarp_state) {
1466                         iwqp->iwarp_state = info.next_iwarp_state;
1467                         iwqp->ibqp_state = attr->qp_state;
1468                 }
1469                 spin_unlock_irqrestore(&iwqp->lock, flags);
1470         }
1471
1472         if (issue_modify_qp && iwqp->ibqp_state > IB_QPS_RTS) {
1473                 if (dont_wait) {
1474                         if (iwqp->hw_tcp_state) {
1475                                 spin_lock_irqsave(&iwqp->lock, flags);
1476                                 iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1477                                 iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1478                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1479                         }
1480                         irdma_cm_disconn(iwqp);
1481                 } else {
1482                         int close_timer_started;
1483
1484                         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
1485
1486                         if (iwqp->cm_node) {
1487                                 atomic_inc(&iwqp->cm_node->refcnt);
1488                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1489                                 close_timer_started = atomic_inc_return(&iwqp->close_timer_started);
1490                                 if (iwqp->cm_id && close_timer_started == 1)
1491                                         irdma_schedule_cm_timer(iwqp->cm_node,
1492                                                                 (struct irdma_puda_buf *)iwqp,
1493                                                                 IRDMA_TIMER_TYPE_CLOSE, 1, 0);
1494
1495                                 irdma_rem_ref_cm_node(iwqp->cm_node);
1496                         } else {
1497                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1498                         }
1499                 }
1500         }
1501         if (attr_mask & IB_QP_STATE && udata && udata->outlen &&
1502             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1503                 struct irdma_ucontext *ucontext;
1504
1505 #if __FreeBSD_version >= 1400026
1506                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
1507 #else
1508                 ucontext = to_ucontext(ibqp->uobject->context);
1509 #endif
1510                 if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1511                     !iwqp->push_wqe_mmap_entry &&
1512                     !irdma_setup_push_mmap_entries(ucontext, iwqp,
1513                                                    &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1514                         uresp.push_valid = 1;
1515                         uresp.push_offset = iwqp->sc_qp.push_offset;
1516                 }
1517                 uresp.rd_fence_rate = iwdev->rd_fence_rate;
1518
1519                 err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1520                                                           udata->outlen));
1521                 if (err) {
1522                         irdma_remove_push_mmap_entries(iwqp);
1523                         irdma_debug(&iwdev->rf->sc_dev,
1524                                     IRDMA_DEBUG_VERBS, "copy_to_udata failed\n");
1525                         return err;
1526                 }
1527         }
1528
1529         return 0;
1530 exit:
1531         spin_unlock_irqrestore(&iwqp->lock, flags);
1532
1533         return err;
1534 }
1535
1536 /**
1537  * irdma_cq_free_rsrc - free up resources for cq
1538  * @rf: RDMA PCI function
1539  * @iwcq: cq ptr
1540  */
1541 void
1542 irdma_cq_free_rsrc(struct irdma_pci_f *rf, struct irdma_cq *iwcq)
1543 {
1544         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1545
1546         if (!iwcq->user_mode) {
1547                 irdma_free_dma_mem(rf->sc_dev.hw, &iwcq->kmem);
1548                 irdma_free_dma_mem(rf->sc_dev.hw, &iwcq->kmem_shadow);
1549         }
1550
1551         irdma_free_rsrc(rf, rf->allocated_cqs, cq->cq_uk.cq_id);
1552 }
1553
1554 /**
1555  * irdma_free_cqbuf - worker to free a cq buffer
1556  * @work: provides access to the cq buffer to free
1557  */
1558 static void
1559 irdma_free_cqbuf(struct work_struct *work)
1560 {
1561         struct irdma_cq_buf *cq_buf = container_of(work, struct irdma_cq_buf, work);
1562
1563         irdma_free_dma_mem(cq_buf->hw, &cq_buf->kmem_buf);
1564         kfree(cq_buf);
1565 }
1566
1567 /**
1568  * irdma_process_resize_list - remove resized cq buffers from the resize_list
1569  * @iwcq: cq which owns the resize_list
1570  * @iwdev: irdma device
1571  * @lcqe_buf: the buffer where the last cqe is received
1572  */
1573 int
1574 irdma_process_resize_list(struct irdma_cq *iwcq,
1575                           struct irdma_device *iwdev,
1576                           struct irdma_cq_buf *lcqe_buf)
1577 {
1578         struct list_head *tmp_node, *list_node;
1579         struct irdma_cq_buf *cq_buf;
1580         int cnt = 0;
1581
1582         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
1583                 cq_buf = list_entry(list_node, struct irdma_cq_buf, list);
1584                 if (cq_buf == lcqe_buf)
1585                         return cnt;
1586
1587                 list_del(&cq_buf->list);
1588                 queue_work(iwdev->cleanup_wq, &cq_buf->work);
1589                 cnt++;
1590         }
1591
1592         return cnt;
1593 }
1594
1595 /**
1596  * irdma_resize_cq - resize cq
1597  * @ibcq: cq to be resized
1598  * @entries: desired cq size
1599  * @udata: user data
1600  */
1601 static int
1602 irdma_resize_cq(struct ib_cq *ibcq, int entries,
1603                 struct ib_udata *udata)
1604 {
1605 #define IRDMA_RESIZE_CQ_MIN_REQ_LEN offsetofend(struct irdma_resize_cq_req, user_cq_buffer)
1606         struct irdma_cq *iwcq = to_iwcq(ibcq);
1607         struct irdma_sc_dev *dev = iwcq->sc_cq.dev;
1608         struct irdma_cqp_request *cqp_request;
1609         struct cqp_cmds_info *cqp_info;
1610         struct irdma_modify_cq_info *m_info;
1611         struct irdma_modify_cq_info info = {0};
1612         struct irdma_dma_mem kmem_buf;
1613         struct irdma_cq_mr *cqmr_buf;
1614         struct irdma_pbl *iwpbl_buf;
1615         struct irdma_device *iwdev;
1616         struct irdma_pci_f *rf;
1617         struct irdma_cq_buf *cq_buf = NULL;
1618         unsigned long flags;
1619         int ret;
1620
1621         iwdev = to_iwdev(ibcq->device);
1622         rf = iwdev->rf;
1623
1624         if (!(rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
1625               IRDMA_FEATURE_CQ_RESIZE))
1626                 return -EOPNOTSUPP;
1627
1628         if (udata && udata->inlen < IRDMA_RESIZE_CQ_MIN_REQ_LEN)
1629                 return -EINVAL;
1630
1631         if (entries > rf->max_cqe)
1632                 return -EINVAL;
1633
1634         if (!iwcq->user_mode) {
1635                 entries++;
1636                 if (rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
1637                         entries *= 2;
1638         }
1639
1640         info.cq_size = max(entries, 4);
1641
1642         if (info.cq_size == iwcq->sc_cq.cq_uk.cq_size - 1)
1643                 return 0;
1644
1645         if (udata) {
1646                 struct irdma_resize_cq_req req = {};
1647                 struct irdma_ucontext *ucontext =
1648 #if __FreeBSD_version >= 1400026
1649                 rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
1650 #else
1651                 to_ucontext(ibcq->uobject->context);
1652 #endif
1653
1654                 /* CQ resize not supported with legacy GEN_1 libi40iw */
1655                 if (ucontext->legacy_mode)
1656                         return -EOPNOTSUPP;
1657
1658                 if (ib_copy_from_udata(&req, udata,
1659                                        min(sizeof(req), udata->inlen)))
1660                         return -EINVAL;
1661
1662                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1663                 iwpbl_buf = irdma_get_pbl((unsigned long)req.user_cq_buffer,
1664                                           &ucontext->cq_reg_mem_list);
1665                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1666
1667                 if (!iwpbl_buf)
1668                         return -ENOMEM;
1669
1670                 cqmr_buf = &iwpbl_buf->cq_mr;
1671                 if (iwpbl_buf->pbl_allocated) {
1672                         info.virtual_map = true;
1673                         info.pbl_chunk_size = 1;
1674                         info.first_pm_pbl_idx = cqmr_buf->cq_pbl.idx;
1675                 } else {
1676                         info.cq_pa = cqmr_buf->cq_pbl.addr;
1677                 }
1678         } else {
1679                 /* Kmode CQ resize */
1680                 int rsize;
1681
1682                 rsize = info.cq_size * sizeof(struct irdma_cqe);
1683                 kmem_buf.size = round_up(rsize, 256);
1684                 kmem_buf.va = irdma_allocate_dma_mem(dev->hw, &kmem_buf,
1685                                                      kmem_buf.size, 256);
1686                 if (!kmem_buf.va)
1687                         return -ENOMEM;
1688
1689                 info.cq_base = kmem_buf.va;
1690                 info.cq_pa = kmem_buf.pa;
1691                 cq_buf = kzalloc(sizeof(*cq_buf), GFP_KERNEL);
1692                 if (!cq_buf) {
1693                         ret = -ENOMEM;
1694                         goto error;
1695                 }
1696         }
1697
1698         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
1699         if (!cqp_request) {
1700                 ret = -ENOMEM;
1701                 goto error;
1702         }
1703
1704         info.shadow_read_threshold = iwcq->sc_cq.shadow_read_threshold;
1705         info.cq_resize = true;
1706
1707         cqp_info = &cqp_request->info;
1708         m_info = &cqp_info->in.u.cq_modify.info;
1709         memcpy(m_info, &info, sizeof(*m_info));
1710
1711         cqp_info->cqp_cmd = IRDMA_OP_CQ_MODIFY;
1712         cqp_info->in.u.cq_modify.cq = &iwcq->sc_cq;
1713         cqp_info->in.u.cq_modify.scratch = (uintptr_t)cqp_request;
1714         cqp_info->post_sq = 1;
1715         ret = irdma_handle_cqp_op(rf, cqp_request);
1716         irdma_put_cqp_request(&rf->cqp, cqp_request);
1717         if (ret)
1718                 goto error;
1719
1720         spin_lock_irqsave(&iwcq->lock, flags);
1721         if (cq_buf) {
1722                 cq_buf->kmem_buf = iwcq->kmem;
1723                 cq_buf->hw = dev->hw;
1724                 memcpy(&cq_buf->cq_uk, &iwcq->sc_cq.cq_uk, sizeof(cq_buf->cq_uk));
1725                 INIT_WORK(&cq_buf->work, irdma_free_cqbuf);
1726                 list_add_tail(&cq_buf->list, &iwcq->resize_list);
1727                 iwcq->kmem = kmem_buf;
1728         }
1729
1730         irdma_sc_cq_resize(&iwcq->sc_cq, &info);
1731         ibcq->cqe = info.cq_size - 1;
1732         spin_unlock_irqrestore(&iwcq->lock, flags);
1733
1734         return 0;
1735 error:
1736         if (!udata)
1737                 irdma_free_dma_mem(dev->hw, &kmem_buf);
1738         kfree(cq_buf);
1739
1740         return ret;
1741 }
1742
1743 /**
1744  * irdma_get_mr_access - get hw MR access permissions from IB access flags
1745  * @access: IB access flags
1746  */
1747 static inline u16 irdma_get_mr_access(int access){
1748         u16 hw_access = 0;
1749
1750         hw_access |= (access & IB_ACCESS_LOCAL_WRITE) ?
1751             IRDMA_ACCESS_FLAGS_LOCALWRITE : 0;
1752         hw_access |= (access & IB_ACCESS_REMOTE_WRITE) ?
1753             IRDMA_ACCESS_FLAGS_REMOTEWRITE : 0;
1754         hw_access |= (access & IB_ACCESS_REMOTE_READ) ?
1755             IRDMA_ACCESS_FLAGS_REMOTEREAD : 0;
1756         hw_access |= (access & IB_ACCESS_MW_BIND) ?
1757             IRDMA_ACCESS_FLAGS_BIND_WINDOW : 0;
1758         hw_access |= (access & IB_ZERO_BASED) ?
1759             IRDMA_ACCESS_FLAGS_ZERO_BASED : 0;
1760         hw_access |= IRDMA_ACCESS_FLAGS_LOCALREAD;
1761
1762         return hw_access;
1763 }
1764
1765 /**
1766  * irdma_free_stag - free stag resource
1767  * @iwdev: irdma device
1768  * @stag: stag to free
1769  */
1770 void
1771 irdma_free_stag(struct irdma_device *iwdev, u32 stag)
1772 {
1773         u32 stag_idx;
1774
1775         stag_idx = (stag & iwdev->rf->mr_stagmask) >> IRDMA_CQPSQ_STAG_IDX_S;
1776         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_mrs, stag_idx);
1777 }
1778
1779 /**
1780  * irdma_create_stag - create random stag
1781  * @iwdev: irdma device
1782  */
1783 u32
1784 irdma_create_stag(struct irdma_device *iwdev)
1785 {
1786         u32 stag;
1787         u32 stag_index = 0;
1788         u32 next_stag_index;
1789         u32 driver_key;
1790         u32 random;
1791         u8 consumer_key;
1792         int ret;
1793
1794         get_random_bytes(&random, sizeof(random));
1795         consumer_key = (u8)random;
1796
1797         driver_key = random & ~iwdev->rf->mr_stagmask;
1798         next_stag_index = (random & iwdev->rf->mr_stagmask) >> 8;
1799         next_stag_index %= iwdev->rf->max_mr;
1800
1801         ret = irdma_alloc_rsrc(iwdev->rf, iwdev->rf->allocated_mrs,
1802                                iwdev->rf->max_mr, &stag_index,
1803                                &next_stag_index);
1804         if (ret)
1805                 return 0;
1806         stag = stag_index << IRDMA_CQPSQ_STAG_IDX_S;
1807         stag |= driver_key;
1808         stag += (u32)consumer_key;
1809
1810         return stag;
1811 }
1812
1813 /**
1814  * irdma_check_mem_contiguous - check if pbls stored in arr are contiguous
1815  * @arr: lvl1 pbl array
1816  * @npages: page count
1817  * @pg_size: page size
1818  *
1819  */
1820 static bool
1821 irdma_check_mem_contiguous(u64 *arr, u32 npages, u32 pg_size)
1822 {
1823         u32 pg_idx;
1824
1825         for (pg_idx = 0; pg_idx < npages; pg_idx++) {
1826                 if ((*arr + (pg_size * pg_idx)) != arr[pg_idx])
1827                         return false;
1828         }
1829
1830         return true;
1831 }
1832
1833 /**
1834  * irdma_check_mr_contiguous - check if MR is physically contiguous
1835  * @palloc: pbl allocation struct
1836  * @pg_size: page size
1837  */
1838 static bool
1839 irdma_check_mr_contiguous(struct irdma_pble_alloc *palloc,
1840                           u32 pg_size)
1841 {
1842         struct irdma_pble_level2 *lvl2 = &palloc->level2;
1843         struct irdma_pble_info *leaf = lvl2->leaf;
1844         u64 *arr = NULL;
1845         u64 *start_addr = NULL;
1846         int i;
1847         bool ret;
1848
1849         if (palloc->level == PBLE_LEVEL_1) {
1850                 arr = palloc->level1.addr;
1851                 ret = irdma_check_mem_contiguous(arr, palloc->total_cnt,
1852                                                  pg_size);
1853                 return ret;
1854         }
1855
1856         start_addr = leaf->addr;
1857
1858         for (i = 0; i < lvl2->leaf_cnt; i++, leaf++) {
1859                 arr = leaf->addr;
1860                 if ((*start_addr + (i * pg_size * PBLE_PER_PAGE)) != *arr)
1861                         return false;
1862                 ret = irdma_check_mem_contiguous(arr, leaf->cnt, pg_size);
1863                 if (!ret)
1864                         return false;
1865         }
1866
1867         return true;
1868 }
1869
1870 /**
1871  * irdma_setup_pbles - copy user pg address to pble's
1872  * @rf: RDMA PCI function
1873  * @iwmr: mr pointer for this memory registration
1874  * @lvl: requested pble levels
1875  */
1876 static int
1877 irdma_setup_pbles(struct irdma_pci_f *rf, struct irdma_mr *iwmr,
1878                   u8 lvl)
1879 {
1880         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
1881         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
1882         struct irdma_pble_info *pinfo;
1883         u64 *pbl;
1884         int status;
1885         enum irdma_pble_level level = PBLE_LEVEL_1;
1886
1887         if (lvl) {
1888                 status = irdma_get_pble(rf->pble_rsrc, palloc, iwmr->page_cnt,
1889                                         lvl);
1890                 if (status)
1891                         return status;
1892
1893                 iwpbl->pbl_allocated = true;
1894                 level = palloc->level;
1895                 pinfo = (level == PBLE_LEVEL_1) ? &palloc->level1 :
1896                     palloc->level2.leaf;
1897                 pbl = pinfo->addr;
1898         } else {
1899                 pbl = iwmr->pgaddrmem;
1900         }
1901
1902         irdma_copy_user_pgaddrs(iwmr, pbl, level);
1903
1904         if (lvl)
1905                 iwmr->pgaddrmem[0] = *pbl;
1906
1907         return 0;
1908 }
1909
1910 /**
1911  * irdma_handle_q_mem - handle memory for qp and cq
1912  * @iwdev: irdma device
1913  * @req: information for q memory management
1914  * @iwpbl: pble struct
1915  * @lvl: pble level mask
1916  */
1917 static int
1918 irdma_handle_q_mem(struct irdma_device *iwdev,
1919                    struct irdma_mem_reg_req *req,
1920                    struct irdma_pbl *iwpbl, u8 lvl)
1921 {
1922         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
1923         struct irdma_mr *iwmr = iwpbl->iwmr;
1924         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
1925         struct irdma_cq_mr *cqmr = &iwpbl->cq_mr;
1926         struct irdma_hmc_pble *hmc_p;
1927         u64 *arr = iwmr->pgaddrmem;
1928         u32 pg_size, total;
1929         int err = 0;
1930         bool ret = true;
1931
1932         pg_size = iwmr->page_size;
1933         err = irdma_setup_pbles(iwdev->rf, iwmr, lvl);
1934         if (err)
1935                 return err;
1936
1937         if (lvl)
1938                 arr = palloc->level1.addr;
1939
1940         switch (iwmr->type) {
1941         case IRDMA_MEMREG_TYPE_QP:
1942                 total = req->sq_pages + req->rq_pages;
1943                 hmc_p = &qpmr->sq_pbl;
1944                 qpmr->shadow = (dma_addr_t) arr[total];
1945                 if (lvl) {
1946                         ret = irdma_check_mem_contiguous(arr, req->sq_pages,
1947                                                          pg_size);
1948                         if (ret)
1949                                 ret = irdma_check_mem_contiguous(&arr[req->sq_pages],
1950                                                                  req->rq_pages,
1951                                                                  pg_size);
1952                 }
1953
1954                 if (!ret) {
1955                         hmc_p->idx = palloc->level1.idx;
1956                         hmc_p = &qpmr->rq_pbl;
1957                         hmc_p->idx = palloc->level1.idx + req->sq_pages;
1958                 } else {
1959                         hmc_p->addr = arr[0];
1960                         hmc_p = &qpmr->rq_pbl;
1961                         hmc_p->addr = arr[req->sq_pages];
1962                 }
1963                 break;
1964         case IRDMA_MEMREG_TYPE_CQ:
1965                 hmc_p = &cqmr->cq_pbl;
1966
1967                 if (!cqmr->split)
1968                         cqmr->shadow = (dma_addr_t) arr[req->cq_pages];
1969
1970                 if (lvl)
1971                         ret = irdma_check_mem_contiguous(arr, req->cq_pages,
1972                                                          pg_size);
1973
1974                 if (!ret)
1975                         hmc_p->idx = palloc->level1.idx;
1976                 else
1977                         hmc_p->addr = arr[0];
1978                 break;
1979         default:
1980                 irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS, "MR type error\n");
1981                 err = -EINVAL;
1982         }
1983
1984         if (lvl && ret) {
1985                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
1986                 iwpbl->pbl_allocated = false;
1987         }
1988
1989         return err;
1990 }
1991
1992 /**
1993  * irdma_hw_alloc_mw - create the hw memory window
1994  * @iwdev: irdma device
1995  * @iwmr: pointer to memory window info
1996  */
1997 int
1998 irdma_hw_alloc_mw(struct irdma_device *iwdev, struct irdma_mr *iwmr)
1999 {
2000         struct irdma_mw_alloc_info *info;
2001         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2002         struct irdma_cqp_request *cqp_request;
2003         struct cqp_cmds_info *cqp_info;
2004         int status;
2005
2006         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2007         if (!cqp_request)
2008                 return -ENOMEM;
2009
2010         cqp_info = &cqp_request->info;
2011         info = &cqp_info->in.u.mw_alloc.info;
2012         memset(info, 0, sizeof(*info));
2013         if (iwmr->ibmw.type == IB_MW_TYPE_1)
2014                 info->mw_wide = true;
2015
2016         info->page_size = PAGE_SIZE;
2017         info->mw_stag_index = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2018         info->pd_id = iwpd->sc_pd.pd_id;
2019         info->remote_access = true;
2020         cqp_info->cqp_cmd = IRDMA_OP_MW_ALLOC;
2021         cqp_info->post_sq = 1;
2022         cqp_info->in.u.mw_alloc.dev = &iwdev->rf->sc_dev;
2023         cqp_info->in.u.mw_alloc.scratch = (uintptr_t)cqp_request;
2024         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2025         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2026
2027         return status;
2028 }
2029
2030 /**
2031  * irdma_dealloc_mw - Dealloc memory window
2032  * @ibmw: memory window structure.
2033  */
2034 static int
2035 irdma_dealloc_mw(struct ib_mw *ibmw)
2036 {
2037         struct ib_pd *ibpd = ibmw->pd;
2038         struct irdma_pd *iwpd = to_iwpd(ibpd);
2039         struct irdma_mr *iwmr = to_iwmr((struct ib_mr *)ibmw);
2040         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2041         struct irdma_cqp_request *cqp_request;
2042         struct cqp_cmds_info *cqp_info;
2043         struct irdma_dealloc_stag_info *info;
2044
2045         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2046         if (!cqp_request)
2047                 return -ENOMEM;
2048
2049         cqp_info = &cqp_request->info;
2050         info = &cqp_info->in.u.dealloc_stag.info;
2051         memset(info, 0, sizeof(*info));
2052         info->pd_id = iwpd->sc_pd.pd_id;
2053         info->stag_idx = RS_64_1(ibmw->rkey, IRDMA_CQPSQ_STAG_IDX_S);
2054         info->mr = false;
2055         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2056         cqp_info->post_sq = 1;
2057         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2058         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2059         irdma_handle_cqp_op(iwdev->rf, cqp_request);
2060         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2061         irdma_free_stag(iwdev, iwmr->stag);
2062         kfree(iwmr);
2063
2064         return 0;
2065 }
2066
2067 /**
2068  * irdma_hw_alloc_stag - cqp command to allocate stag
2069  * @iwdev: irdma device
2070  * @iwmr: irdma mr pointer
2071  */
2072 int
2073 irdma_hw_alloc_stag(struct irdma_device *iwdev,
2074                     struct irdma_mr *iwmr)
2075 {
2076         struct irdma_allocate_stag_info *info;
2077         struct ib_pd *pd = iwmr->ibmr.pd;
2078         struct irdma_pd *iwpd = to_iwpd(pd);
2079         struct irdma_cqp_request *cqp_request;
2080         struct cqp_cmds_info *cqp_info;
2081         int status;
2082
2083         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2084         if (!cqp_request)
2085                 return -ENOMEM;
2086
2087         cqp_info = &cqp_request->info;
2088         info = &cqp_info->in.u.alloc_stag.info;
2089         memset(info, 0, sizeof(*info));
2090         info->page_size = PAGE_SIZE;
2091         info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2092         info->pd_id = iwpd->sc_pd.pd_id;
2093         info->total_len = iwmr->len;
2094         info->all_memory = (pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) ? true : false;
2095         info->remote_access = true;
2096         cqp_info->cqp_cmd = IRDMA_OP_ALLOC_STAG;
2097         cqp_info->post_sq = 1;
2098         cqp_info->in.u.alloc_stag.dev = &iwdev->rf->sc_dev;
2099         cqp_info->in.u.alloc_stag.scratch = (uintptr_t)cqp_request;
2100         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2101         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2102         if (!status)
2103                 iwmr->is_hwreg = 1;
2104
2105         return status;
2106 }
2107
2108 /**
2109  * irdma_set_page - populate pbl list for fmr
2110  * @ibmr: ib mem to access iwarp mr pointer
2111  * @addr: page dma address fro pbl list
2112  */
2113 static int
2114 irdma_set_page(struct ib_mr *ibmr, u64 addr)
2115 {
2116         struct irdma_mr *iwmr = to_iwmr(ibmr);
2117         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2118         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2119         u64 *pbl;
2120
2121         if (unlikely(iwmr->npages == iwmr->page_cnt))
2122                 return -ENOMEM;
2123
2124         if (palloc->level == PBLE_LEVEL_2) {
2125                 struct irdma_pble_info *palloc_info =
2126                 palloc->level2.leaf + (iwmr->npages >> PBLE_512_SHIFT);
2127
2128                 palloc_info->addr[iwmr->npages & (PBLE_PER_PAGE - 1)] = addr;
2129         } else {
2130                 pbl = palloc->level1.addr;
2131                 pbl[iwmr->npages] = addr;
2132         }
2133
2134         iwmr->npages++;
2135         return 0;
2136 }
2137
2138 /**
2139  * irdma_map_mr_sg - map of sg list for fmr
2140  * @ibmr: ib mem to access iwarp mr pointer
2141  * @sg: scatter gather list
2142  * @sg_nents: number of sg pages
2143  * @sg_offset: scatter gather list for fmr
2144  */
2145 static int
2146 irdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
2147                 int sg_nents, unsigned int *sg_offset)
2148 {
2149         struct irdma_mr *iwmr = to_iwmr(ibmr);
2150
2151         iwmr->npages = 0;
2152
2153         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, irdma_set_page);
2154 }
2155
2156 /**
2157  * irdma_hwreg_mr - send cqp command for memory registration
2158  * @iwdev: irdma device
2159  * @iwmr: irdma mr pointer
2160  * @access: access for MR
2161  */
2162 int
2163 irdma_hwreg_mr(struct irdma_device *iwdev, struct irdma_mr *iwmr,
2164                u16 access)
2165 {
2166         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2167         struct irdma_reg_ns_stag_info *stag_info;
2168         struct ib_pd *pd = iwmr->ibmr.pd;
2169         struct irdma_pd *iwpd = to_iwpd(pd);
2170         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2171         struct irdma_cqp_request *cqp_request;
2172         struct cqp_cmds_info *cqp_info;
2173         int ret;
2174
2175         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2176         if (!cqp_request)
2177                 return -ENOMEM;
2178
2179         cqp_info = &cqp_request->info;
2180         stag_info = &cqp_info->in.u.mr_reg_non_shared.info;
2181         memset(stag_info, 0, sizeof(*stag_info));
2182         stag_info->va = iwpbl->user_base;
2183         stag_info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2184         stag_info->stag_key = (u8)iwmr->stag;
2185         stag_info->total_len = iwmr->len;
2186         stag_info->all_memory = (pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) ? true : false;
2187         stag_info->access_rights = irdma_get_mr_access(access);
2188         stag_info->pd_id = iwpd->sc_pd.pd_id;
2189         if (stag_info->access_rights & IRDMA_ACCESS_FLAGS_ZERO_BASED)
2190                 stag_info->addr_type = IRDMA_ADDR_TYPE_ZERO_BASED;
2191         else
2192                 stag_info->addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2193         stag_info->page_size = iwmr->page_size;
2194
2195         if (iwpbl->pbl_allocated) {
2196                 if (palloc->level == PBLE_LEVEL_1) {
2197                         stag_info->first_pm_pbl_index = palloc->level1.idx;
2198                         stag_info->chunk_size = 1;
2199                 } else {
2200                         stag_info->first_pm_pbl_index = palloc->level2.root.idx;
2201                         stag_info->chunk_size = 3;
2202                 }
2203         } else {
2204                 stag_info->reg_addr_pa = iwmr->pgaddrmem[0];
2205         }
2206
2207         cqp_info->cqp_cmd = IRDMA_OP_MR_REG_NON_SHARED;
2208         cqp_info->post_sq = 1;
2209         cqp_info->in.u.mr_reg_non_shared.dev = &iwdev->rf->sc_dev;
2210         cqp_info->in.u.mr_reg_non_shared.scratch = (uintptr_t)cqp_request;
2211         ret = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2212         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2213
2214         if (!ret)
2215                 iwmr->is_hwreg = 1;
2216
2217         return ret;
2218 }
2219
2220 /**
2221  * irdma_reg_user_mr - Register a user memory region
2222  * @pd: ptr of pd
2223  * @start: virtual start address
2224  * @len: length of mr
2225  * @virt: virtual address
2226  * @access: access of mr
2227  * @udata: user data
2228  */
2229 static struct ib_mr *
2230 irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
2231                   u64 virt, int access,
2232                   struct ib_udata *udata)
2233 {
2234 #define IRDMA_MEM_REG_MIN_REQ_LEN offsetofend(struct irdma_mem_reg_req, sq_pages)
2235         struct irdma_device *iwdev = to_iwdev(pd->device);
2236         struct irdma_ucontext *ucontext;
2237         struct irdma_pble_alloc *palloc;
2238         struct irdma_pbl *iwpbl;
2239         struct irdma_mr *iwmr;
2240         struct ib_umem *region;
2241         struct irdma_mem_reg_req req = {};
2242         u32 total, stag = 0;
2243         u8 shadow_pgcnt = 1;
2244         unsigned long flags;
2245         int err = -EINVAL;
2246         u8 lvl;
2247         int ret;
2248
2249         if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
2250                 return ERR_PTR(-EINVAL);
2251
2252         if (udata->inlen < IRDMA_MEM_REG_MIN_REQ_LEN)
2253                 return ERR_PTR(-EINVAL);
2254
2255         region = ib_umem_get(pd->uobject->context, start, len, access, 0);
2256
2257         if (IS_ERR(region)) {
2258                 irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
2259                             "Failed to create ib_umem region\n");
2260                 return (struct ib_mr *)region;
2261         }
2262
2263         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen))) {
2264                 ib_umem_release(region);
2265                 return ERR_PTR(-EFAULT);
2266         }
2267
2268         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2269         if (!iwmr) {
2270                 ib_umem_release(region);
2271                 return ERR_PTR(-ENOMEM);
2272         }
2273
2274         iwpbl = &iwmr->iwpbl;
2275         iwpbl->iwmr = iwmr;
2276         iwmr->region = region;
2277         iwmr->ibmr.pd = pd;
2278         iwmr->ibmr.device = pd->device;
2279         iwmr->ibmr.iova = virt;
2280         iwmr->page_size = IRDMA_HW_PAGE_SIZE;
2281         iwmr->page_msk = ~(IRDMA_HW_PAGE_SIZE - 1);
2282
2283         iwmr->len = region->length;
2284         iwpbl->user_base = virt;
2285         palloc = &iwpbl->pble_alloc;
2286         iwmr->type = req.reg_type;
2287         iwmr->page_cnt = irdma_ib_umem_num_dma_blocks(region, iwmr->page_size, virt);
2288
2289         switch (req.reg_type) {
2290         case IRDMA_MEMREG_TYPE_QP:
2291                 total = req.sq_pages + req.rq_pages + shadow_pgcnt;
2292                 if (total > iwmr->page_cnt) {
2293                         err = -EINVAL;
2294                         goto error;
2295                 }
2296                 total = req.sq_pages + req.rq_pages;
2297                 lvl = total > 2 ? PBLE_LEVEL_1 : PBLE_LEVEL_0;
2298                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, lvl);
2299                 if (err)
2300                         goto error;
2301
2302 #if __FreeBSD_version >= 1400026
2303                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
2304 #else
2305                 ucontext = to_ucontext(pd->uobject->context);
2306 #endif
2307                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2308                 list_add_tail(&iwpbl->list, &ucontext->qp_reg_mem_list);
2309                 iwpbl->on_list = true;
2310                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2311                 break;
2312         case IRDMA_MEMREG_TYPE_CQ:
2313                 if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_CQ_RESIZE)
2314                         shadow_pgcnt = 0;
2315                 total = req.cq_pages + shadow_pgcnt;
2316                 if (total > iwmr->page_cnt) {
2317                         err = -EINVAL;
2318                         goto error;
2319                 }
2320
2321                 lvl = req.cq_pages > 1 ? PBLE_LEVEL_1 : PBLE_LEVEL_0;
2322                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, lvl);
2323                 if (err)
2324                         goto error;
2325
2326 #if __FreeBSD_version >= 1400026
2327                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
2328 #else
2329                 ucontext = to_ucontext(pd->uobject->context);
2330 #endif
2331                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2332                 list_add_tail(&iwpbl->list, &ucontext->cq_reg_mem_list);
2333                 iwpbl->on_list = true;
2334                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2335                 break;
2336         case IRDMA_MEMREG_TYPE_MEM:
2337                 lvl = iwmr->page_cnt != 1 ? PBLE_LEVEL_1 | PBLE_LEVEL_2 : PBLE_LEVEL_0;
2338                 err = irdma_setup_pbles(iwdev->rf, iwmr, lvl);
2339                 if (err)
2340                         goto error;
2341
2342                 if (lvl) {
2343                         ret = irdma_check_mr_contiguous(palloc,
2344                                                         iwmr->page_size);
2345                         if (ret) {
2346                                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2347                                 iwpbl->pbl_allocated = false;
2348                         }
2349                 }
2350
2351                 stag = irdma_create_stag(iwdev);
2352                 if (!stag) {
2353                         err = -ENOMEM;
2354                         goto error;
2355                 }
2356
2357                 iwmr->stag = stag;
2358                 iwmr->ibmr.rkey = stag;
2359                 iwmr->ibmr.lkey = stag;
2360                 iwmr->access = access;
2361                 err = irdma_hwreg_mr(iwdev, iwmr, access);
2362                 if (err) {
2363                         irdma_free_stag(iwdev, stag);
2364                         goto error;
2365                 }
2366
2367                 break;
2368         default:
2369                 goto error;
2370         }
2371
2372         iwmr->type = req.reg_type;
2373
2374         return &iwmr->ibmr;
2375
2376 error:
2377         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated)
2378                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2379         ib_umem_release(region);
2380         kfree(iwmr);
2381
2382         return ERR_PTR(err);
2383 }
2384
2385 int
2386 irdma_hwdereg_mr(struct ib_mr *ib_mr)
2387 {
2388         struct irdma_device *iwdev = to_iwdev(ib_mr->device);
2389         struct irdma_mr *iwmr = to_iwmr(ib_mr);
2390         struct irdma_pd *iwpd = to_iwpd(ib_mr->pd);
2391         struct irdma_dealloc_stag_info *info;
2392         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2393         struct irdma_cqp_request *cqp_request;
2394         struct cqp_cmds_info *cqp_info;
2395         int status;
2396
2397         /*
2398          * Skip HW MR de-register when it is already de-registered during an MR re-reregister and the re-registration
2399          * fails
2400          */
2401         if (!iwmr->is_hwreg)
2402                 return 0;
2403
2404         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2405         if (!cqp_request)
2406                 return -ENOMEM;
2407
2408         cqp_info = &cqp_request->info;
2409         info = &cqp_info->in.u.dealloc_stag.info;
2410         memset(info, 0, sizeof(*info));
2411         info->pd_id = iwpd->sc_pd.pd_id;
2412         info->stag_idx = RS_64_1(ib_mr->rkey, IRDMA_CQPSQ_STAG_IDX_S);
2413         info->mr = true;
2414         if (iwpbl->pbl_allocated)
2415                 info->dealloc_pbl = true;
2416
2417         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2418         cqp_info->post_sq = 1;
2419         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2420         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2421         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2422         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2423
2424         if (!status)
2425                 iwmr->is_hwreg = 0;
2426
2427         return status;
2428 }
2429
2430 /*
2431  * irdma_rereg_mr_trans - Re-register a user MR for a change translation. @iwmr: ptr of iwmr @start: virtual start
2432  * address @len: length of mr @virt: virtual address
2433  *
2434  * Re-register a user memory region when a change translation is requested. Re-register a new region while reusing the
2435  * stag from the original registration.
2436  */
2437 struct ib_mr *
2438 irdma_rereg_mr_trans(struct irdma_mr *iwmr, u64 start, u64 len,
2439                      u64 virt, struct ib_udata *udata)
2440 {
2441         struct irdma_device *iwdev = to_iwdev(iwmr->ibmr.device);
2442         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2443         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2444         struct ib_pd *pd = iwmr->ibmr.pd;
2445         struct ib_umem *region;
2446         u8 lvl;
2447         int err;
2448
2449         region = ib_umem_get(pd->uobject->context, start, len, iwmr->access, 0);
2450
2451         if (IS_ERR(region)) {
2452                 irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
2453                             "Failed to create ib_umem region\n");
2454                 return (struct ib_mr *)region;
2455         }
2456
2457         iwmr->region = region;
2458         iwmr->ibmr.iova = virt;
2459         iwmr->ibmr.pd = pd;
2460         iwmr->page_size = PAGE_SIZE;
2461
2462         iwmr->len = region->length;
2463         iwpbl->user_base = virt;
2464         iwmr->page_cnt = irdma_ib_umem_num_dma_blocks(region, iwmr->page_size,
2465                                                       virt);
2466
2467         lvl = iwmr->page_cnt != 1 ? PBLE_LEVEL_1 | PBLE_LEVEL_2 : PBLE_LEVEL_0;
2468
2469         err = irdma_setup_pbles(iwdev->rf, iwmr, lvl);
2470         if (err)
2471                 goto error;
2472
2473         if (lvl) {
2474                 err = irdma_check_mr_contiguous(palloc,
2475                                                 iwmr->page_size);
2476                 if (err) {
2477                         irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2478                         iwpbl->pbl_allocated = false;
2479                 }
2480         }
2481
2482         err = irdma_hwreg_mr(iwdev, iwmr, iwmr->access);
2483         if (err)
2484                 goto error;
2485
2486         return &iwmr->ibmr;
2487
2488 error:
2489         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated) {
2490                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2491                 iwpbl->pbl_allocated = false;
2492         }
2493         ib_umem_release(region);
2494         iwmr->region = NULL;
2495
2496         return ERR_PTR(err);
2497 }
2498
2499 /**
2500  * irdma_reg_phys_mr - register kernel physical memory
2501  * @pd: ibpd pointer
2502  * @addr: physical address of memory to register
2503  * @size: size of memory to register
2504  * @access: Access rights
2505  * @iova_start: start of virtual address for physical buffers
2506  */
2507 struct ib_mr *
2508 irdma_reg_phys_mr(struct ib_pd *pd, u64 addr, u64 size, int access,
2509                   u64 *iova_start)
2510 {
2511         struct irdma_device *iwdev = to_iwdev(pd->device);
2512         struct irdma_pbl *iwpbl;
2513         struct irdma_mr *iwmr;
2514         u32 stag;
2515         int ret;
2516
2517         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2518         if (!iwmr)
2519                 return ERR_PTR(-ENOMEM);
2520
2521         iwmr->ibmr.pd = pd;
2522         iwmr->ibmr.device = pd->device;
2523         iwpbl = &iwmr->iwpbl;
2524         iwpbl->iwmr = iwmr;
2525         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2526         iwpbl->user_base = *iova_start;
2527         stag = irdma_create_stag(iwdev);
2528         if (!stag) {
2529                 ret = -ENOMEM;
2530                 goto err;
2531         }
2532
2533         iwmr->stag = stag;
2534         iwmr->ibmr.iova = *iova_start;
2535         iwmr->ibmr.rkey = stag;
2536         iwmr->ibmr.lkey = stag;
2537         iwmr->page_cnt = 1;
2538         iwmr->pgaddrmem[0] = addr;
2539         iwmr->len = size;
2540         iwmr->page_size = SZ_4K;
2541         ret = irdma_hwreg_mr(iwdev, iwmr, access);
2542         if (ret) {
2543                 irdma_free_stag(iwdev, stag);
2544                 goto err;
2545         }
2546
2547         return &iwmr->ibmr;
2548
2549 err:
2550         kfree(iwmr);
2551
2552         return ERR_PTR(ret);
2553 }
2554
2555 /**
2556  * irdma_get_dma_mr - register physical mem
2557  * @pd: ptr of pd
2558  * @acc: access for memory
2559  */
2560 static struct ib_mr *
2561 irdma_get_dma_mr(struct ib_pd *pd, int acc)
2562 {
2563         u64 kva = 0;
2564
2565         return irdma_reg_phys_mr(pd, 0, 0, acc, &kva);
2566 }
2567
2568 /**
2569  * irdma_del_memlist - Deleting pbl list entries for CQ/QP
2570  * @iwmr: iwmr for IB's user page addresses
2571  * @ucontext: ptr to user context
2572  */
2573 void
2574 irdma_del_memlist(struct irdma_mr *iwmr,
2575                   struct irdma_ucontext *ucontext)
2576 {
2577         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2578         unsigned long flags;
2579
2580         switch (iwmr->type) {
2581         case IRDMA_MEMREG_TYPE_CQ:
2582                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2583                 if (iwpbl->on_list) {
2584                         iwpbl->on_list = false;
2585                         list_del(&iwpbl->list);
2586                 }
2587                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2588                 break;
2589         case IRDMA_MEMREG_TYPE_QP:
2590                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2591                 if (iwpbl->on_list) {
2592                         iwpbl->on_list = false;
2593                         list_del(&iwpbl->list);
2594                 }
2595                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2596                 break;
2597         default:
2598                 break;
2599         }
2600 }
2601
2602 /**
2603  * irdma_copy_sg_list - copy sg list for qp
2604  * @sg_list: copied into sg_list
2605  * @sgl: copy from sgl
2606  * @num_sges: count of sg entries
2607  */
2608 static void
2609 irdma_copy_sg_list(struct irdma_sge *sg_list, struct ib_sge *sgl,
2610                    int num_sges)
2611 {
2612         unsigned int i;
2613
2614         for (i = 0; i < num_sges; i++) {
2615                 sg_list[i].tag_off = sgl[i].addr;
2616                 sg_list[i].len = sgl[i].length;
2617                 sg_list[i].stag = sgl[i].lkey;
2618         }
2619 }
2620
2621 /**
2622  * irdma_post_send -  kernel application wr
2623  * @ibqp: qp ptr for wr
2624  * @ib_wr: work request ptr
2625  * @bad_wr: return of bad wr if err
2626  */
2627 static int
2628 irdma_post_send(struct ib_qp *ibqp,
2629                 const struct ib_send_wr *ib_wr,
2630                 const struct ib_send_wr **bad_wr)
2631 {
2632         struct irdma_qp *iwqp;
2633         struct irdma_qp_uk *ukqp;
2634         struct irdma_sc_dev *dev;
2635         struct irdma_post_sq_info info;
2636         int err = 0;
2637         unsigned long flags;
2638         bool inv_stag;
2639         struct irdma_ah *ah;
2640
2641         iwqp = to_iwqp(ibqp);
2642         ukqp = &iwqp->sc_qp.qp_uk;
2643         dev = &iwqp->iwdev->rf->sc_dev;
2644
2645         spin_lock_irqsave(&iwqp->lock, flags);
2646         while (ib_wr) {
2647                 memset(&info, 0, sizeof(info));
2648                 inv_stag = false;
2649                 info.wr_id = (ib_wr->wr_id);
2650                 if ((ib_wr->send_flags & IB_SEND_SIGNALED) || iwqp->sig_all)
2651                         info.signaled = true;
2652                 if (ib_wr->send_flags & IB_SEND_FENCE)
2653                         info.read_fence = true;
2654                 switch (ib_wr->opcode) {
2655                 case IB_WR_SEND_WITH_IMM:
2656                         if (ukqp->qp_caps & IRDMA_SEND_WITH_IMM) {
2657                                 info.imm_data_valid = true;
2658                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
2659                         } else {
2660                                 err = -EINVAL;
2661                                 break;
2662                         }
2663                         /* fallthrough */
2664                 case IB_WR_SEND:
2665                 case IB_WR_SEND_WITH_INV:
2666                         if (ib_wr->opcode == IB_WR_SEND ||
2667                             ib_wr->opcode == IB_WR_SEND_WITH_IMM) {
2668                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
2669                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL;
2670                                 else
2671                                         info.op_type = IRDMA_OP_TYPE_SEND;
2672                         } else {
2673                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
2674                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL_INV;
2675                                 else
2676                                         info.op_type = IRDMA_OP_TYPE_SEND_INV;
2677                                 info.stag_to_inv = ib_wr->ex.invalidate_rkey;
2678                         }
2679
2680                         info.op.send.num_sges = ib_wr->num_sge;
2681                         info.op.send.sg_list = (struct irdma_sge *)ib_wr->sg_list;
2682                         if (iwqp->ibqp.qp_type == IB_QPT_UD ||
2683                             iwqp->ibqp.qp_type == IB_QPT_GSI) {
2684                                 ah = to_iwah(ud_wr(ib_wr)->ah);
2685                                 info.op.send.ah_id = ah->sc_ah.ah_info.ah_idx;
2686                                 info.op.send.qkey = ud_wr(ib_wr)->remote_qkey;
2687                                 info.op.send.dest_qp = ud_wr(ib_wr)->remote_qpn;
2688                         }
2689
2690                         if (ib_wr->send_flags & IB_SEND_INLINE)
2691                                 err = irdma_uk_inline_send(ukqp, &info, false);
2692                         else
2693                                 err = irdma_uk_send(ukqp, &info, false);
2694                         break;
2695                 case IB_WR_RDMA_WRITE_WITH_IMM:
2696                         if (ukqp->qp_caps & IRDMA_WRITE_WITH_IMM) {
2697                                 info.imm_data_valid = true;
2698                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
2699                         } else {
2700                                 err = -EINVAL;
2701                                 break;
2702                         }
2703                         /* fallthrough */
2704                 case IB_WR_RDMA_WRITE:
2705                         if (ib_wr->send_flags & IB_SEND_SOLICITED)
2706                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE_SOL;
2707                         else
2708                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE;
2709
2710                         info.op.rdma_write.num_lo_sges = ib_wr->num_sge;
2711                         info.op.rdma_write.lo_sg_list = (void *)ib_wr->sg_list;
2712                         info.op.rdma_write.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
2713                         info.op.rdma_write.rem_addr.stag = rdma_wr(ib_wr)->rkey;
2714                         if (ib_wr->send_flags & IB_SEND_INLINE)
2715                                 err = irdma_uk_inline_rdma_write(ukqp, &info, false);
2716                         else
2717                                 err = irdma_uk_rdma_write(ukqp, &info, false);
2718                         break;
2719                 case IB_WR_RDMA_READ_WITH_INV:
2720                         inv_stag = true;
2721                         /* fallthrough */
2722                 case IB_WR_RDMA_READ:
2723                         if (ib_wr->num_sge >
2724                             dev->hw_attrs.uk_attrs.max_hw_read_sges) {
2725                                 err = -EINVAL;
2726                                 break;
2727                         }
2728                         info.op_type = IRDMA_OP_TYPE_RDMA_READ;
2729                         info.op.rdma_read.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
2730                         info.op.rdma_read.rem_addr.stag = rdma_wr(ib_wr)->rkey;
2731                         info.op.rdma_read.lo_sg_list = (void *)ib_wr->sg_list;
2732                         info.op.rdma_read.num_lo_sges = ib_wr->num_sge;
2733                         err = irdma_uk_rdma_read(ukqp, &info, inv_stag, false);
2734                         break;
2735                 case IB_WR_LOCAL_INV:
2736                         info.op_type = IRDMA_OP_TYPE_INV_STAG;
2737                         info.local_fence = info.read_fence;
2738                         info.op.inv_local_stag.target_stag = ib_wr->ex.invalidate_rkey;
2739                         err = irdma_uk_stag_local_invalidate(ukqp, &info, true);
2740                         break;
2741                 case IB_WR_REG_MR:{
2742                                 struct irdma_mr *iwmr = to_iwmr(reg_wr(ib_wr)->mr);
2743                                 struct irdma_pble_alloc *palloc = &iwmr->iwpbl.pble_alloc;
2744                                 struct irdma_fast_reg_stag_info stag_info = {0};
2745
2746                                 stag_info.signaled = info.signaled;
2747                                 stag_info.read_fence = info.read_fence;
2748                                 stag_info.access_rights = irdma_get_mr_access(reg_wr(ib_wr)->access);
2749                                 stag_info.stag_key = reg_wr(ib_wr)->key & 0xff;
2750                                 stag_info.stag_idx = reg_wr(ib_wr)->key >> 8;
2751                                 stag_info.page_size = reg_wr(ib_wr)->mr->page_size;
2752                                 stag_info.wr_id = ib_wr->wr_id;
2753                                 stag_info.addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2754                                 stag_info.va = (void *)(uintptr_t)iwmr->ibmr.iova;
2755                                 stag_info.total_len = iwmr->ibmr.length;
2756                                 if (palloc->level == PBLE_LEVEL_2) {
2757                                         stag_info.chunk_size = 3;
2758                                         stag_info.first_pm_pbl_index = palloc->level2.root.idx;
2759                                 } else {
2760                                         stag_info.chunk_size = 1;
2761                                         stag_info.first_pm_pbl_index = palloc->level1.idx;
2762                                 }
2763                                 stag_info.local_fence = ib_wr->send_flags & IB_SEND_FENCE;
2764                                 err = irdma_sc_mr_fast_register(&iwqp->sc_qp, &stag_info,
2765                                                                 true);
2766                                 break;
2767                         }
2768                 default:
2769                         err = -EINVAL;
2770                         irdma_debug(&iwqp->iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
2771                                     "upost_send bad opcode = 0x%x\n",
2772                                     ib_wr->opcode);
2773                         break;
2774                 }
2775
2776                 if (err)
2777                         break;
2778                 ib_wr = ib_wr->next;
2779         }
2780
2781         if (!iwqp->flush_issued) {
2782                 if (iwqp->hw_iwarp_state <= IRDMA_QP_STATE_RTS)
2783                         irdma_uk_qp_post_wr(ukqp);
2784                 spin_unlock_irqrestore(&iwqp->lock, flags);
2785         } else {
2786                 spin_unlock_irqrestore(&iwqp->lock, flags);
2787                 irdma_sched_qp_flush_work(iwqp);
2788         }
2789
2790         if (err)
2791                 *bad_wr = ib_wr;
2792
2793         return err;
2794 }
2795
2796 /**
2797  * irdma_post_recv - post receive wr for kernel application
2798  * @ibqp: ib qp pointer
2799  * @ib_wr: work request for receive
2800  * @bad_wr: bad wr caused an error
2801  */
2802 static int
2803 irdma_post_recv(struct ib_qp *ibqp,
2804                 const struct ib_recv_wr *ib_wr,
2805                 const struct ib_recv_wr **bad_wr)
2806 {
2807         struct irdma_qp *iwqp = to_iwqp(ibqp);
2808         struct irdma_qp_uk *ukqp = &iwqp->sc_qp.qp_uk;
2809         struct irdma_post_rq_info post_recv = {0};
2810         struct irdma_sge *sg_list = iwqp->sg_list;
2811         unsigned long flags;
2812         int err = 0;
2813
2814         spin_lock_irqsave(&iwqp->lock, flags);
2815
2816         while (ib_wr) {
2817                 if (ib_wr->num_sge > ukqp->max_rq_frag_cnt) {
2818                         err = -EINVAL;
2819                         goto out;
2820                 }
2821                 post_recv.num_sges = ib_wr->num_sge;
2822                 post_recv.wr_id = ib_wr->wr_id;
2823                 irdma_copy_sg_list(sg_list, ib_wr->sg_list, ib_wr->num_sge);
2824                 post_recv.sg_list = sg_list;
2825                 err = irdma_uk_post_receive(ukqp, &post_recv);
2826                 if (err) {
2827                         irdma_debug(&iwqp->iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
2828                                     "post_recv err %d\n",
2829                                     err);
2830                         goto out;
2831                 }
2832
2833                 ib_wr = ib_wr->next;
2834         }
2835
2836 out:
2837         spin_unlock_irqrestore(&iwqp->lock, flags);
2838         if (iwqp->flush_issued)
2839                 irdma_sched_qp_flush_work(iwqp);
2840
2841         if (err)
2842                 *bad_wr = ib_wr;
2843
2844         return err;
2845 }
2846
2847 /**
2848  * irdma_flush_err_to_ib_wc_status - return change flush error code to IB status
2849  * @opcode: iwarp flush code
2850  */
2851 static enum ib_wc_status
2852 irdma_flush_err_to_ib_wc_status(enum irdma_flush_opcode opcode)
2853 {
2854         switch (opcode) {
2855         case FLUSH_PROT_ERR:
2856                 return IB_WC_LOC_PROT_ERR;
2857         case FLUSH_REM_ACCESS_ERR:
2858                 return IB_WC_REM_ACCESS_ERR;
2859         case FLUSH_LOC_QP_OP_ERR:
2860                 return IB_WC_LOC_QP_OP_ERR;
2861         case FLUSH_REM_OP_ERR:
2862                 return IB_WC_REM_OP_ERR;
2863         case FLUSH_LOC_LEN_ERR:
2864                 return IB_WC_LOC_LEN_ERR;
2865         case FLUSH_GENERAL_ERR:
2866                 return IB_WC_WR_FLUSH_ERR;
2867         case FLUSH_MW_BIND_ERR:
2868                 return IB_WC_MW_BIND_ERR;
2869         case FLUSH_REM_INV_REQ_ERR:
2870                 return IB_WC_REM_INV_REQ_ERR;
2871         case FLUSH_RETRY_EXC_ERR:
2872                 return IB_WC_RETRY_EXC_ERR;
2873         case FLUSH_FATAL_ERR:
2874         default:
2875                 return IB_WC_FATAL_ERR;
2876         }
2877 }
2878
2879 static inline void
2880 set_ib_wc_op_sq(struct irdma_cq_poll_info *cq_poll_info,
2881                 struct ib_wc *entry)
2882 {
2883         struct irdma_sc_qp *qp;
2884
2885         switch (cq_poll_info->op_type) {
2886         case IRDMA_OP_TYPE_RDMA_WRITE:
2887         case IRDMA_OP_TYPE_RDMA_WRITE_SOL:
2888                 entry->opcode = IB_WC_RDMA_WRITE;
2889                 break;
2890         case IRDMA_OP_TYPE_RDMA_READ_INV_STAG:
2891         case IRDMA_OP_TYPE_RDMA_READ:
2892                 entry->opcode = IB_WC_RDMA_READ;
2893                 break;
2894         case IRDMA_OP_TYPE_SEND_SOL:
2895         case IRDMA_OP_TYPE_SEND_SOL_INV:
2896         case IRDMA_OP_TYPE_SEND_INV:
2897         case IRDMA_OP_TYPE_SEND:
2898                 entry->opcode = IB_WC_SEND;
2899                 break;
2900         case IRDMA_OP_TYPE_FAST_REG_NSMR:
2901                 entry->opcode = IB_WC_REG_MR;
2902                 break;
2903         case IRDMA_OP_TYPE_INV_STAG:
2904                 entry->opcode = IB_WC_LOCAL_INV;
2905                 break;
2906         default:
2907                 qp = cq_poll_info->qp_handle;
2908                 irdma_dev_err(to_ibdev(qp->dev), "Invalid opcode = %d in CQE\n",
2909                               cq_poll_info->op_type);
2910                 entry->status = IB_WC_GENERAL_ERR;
2911         }
2912 }
2913
2914 static inline void
2915 set_ib_wc_op_rq(struct irdma_cq_poll_info *cq_poll_info,
2916                 struct ib_wc *entry, bool send_imm_support)
2917 {
2918         /**
2919          * iWARP does not support sendImm, so the presence of Imm data
2920          * must be WriteImm.
2921          */
2922         if (!send_imm_support) {
2923                 entry->opcode = cq_poll_info->imm_valid ?
2924                     IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
2925                 return;
2926         }
2927         switch (cq_poll_info->op_type) {
2928         case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
2929         case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
2930                 entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2931                 break;
2932         default:
2933                 entry->opcode = IB_WC_RECV;
2934         }
2935 }
2936
2937 /**
2938  * irdma_process_cqe - process cqe info
2939  * @entry: processed cqe
2940  * @cq_poll_info: cqe info
2941  */
2942 static void
2943 irdma_process_cqe(struct ib_wc *entry,
2944                   struct irdma_cq_poll_info *cq_poll_info)
2945 {
2946         struct irdma_sc_qp *qp;
2947
2948         entry->wc_flags = 0;
2949         entry->pkey_index = 0;
2950         entry->wr_id = cq_poll_info->wr_id;
2951
2952         qp = cq_poll_info->qp_handle;
2953         entry->qp = qp->qp_uk.back_qp;
2954
2955         if (cq_poll_info->error) {
2956                 entry->status = (cq_poll_info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ?
2957                     irdma_flush_err_to_ib_wc_status(cq_poll_info->minor_err) : IB_WC_GENERAL_ERR;
2958
2959                 entry->vendor_err = cq_poll_info->major_err << 16 |
2960                     cq_poll_info->minor_err;
2961         } else {
2962                 entry->status = IB_WC_SUCCESS;
2963                 if (cq_poll_info->imm_valid) {
2964                         entry->ex.imm_data = htonl(cq_poll_info->imm_data);
2965                         entry->wc_flags |= IB_WC_WITH_IMM;
2966                 }
2967                 if (cq_poll_info->ud_smac_valid) {
2968                         ether_addr_copy(entry->smac, cq_poll_info->ud_smac);
2969                         entry->wc_flags |= IB_WC_WITH_SMAC;
2970                 }
2971
2972                 if (cq_poll_info->ud_vlan_valid) {
2973                         u16 vlan = cq_poll_info->ud_vlan & EVL_VLID_MASK;
2974
2975                         entry->sl = cq_poll_info->ud_vlan >> VLAN_PRIO_SHIFT;
2976                         if (vlan) {
2977                                 entry->vlan_id = vlan;
2978                                 entry->wc_flags |= IB_WC_WITH_VLAN;
2979                         }
2980                 } else {
2981                         entry->sl = 0;
2982                 }
2983         }
2984
2985         if (cq_poll_info->q_type == IRDMA_CQE_QTYPE_SQ) {
2986                 set_ib_wc_op_sq(cq_poll_info, entry);
2987         } else {
2988                 set_ib_wc_op_rq(cq_poll_info, entry,
2989                                 qp->qp_uk.qp_caps & IRDMA_SEND_WITH_IMM ?
2990                                 true : false);
2991                 if (qp->qp_uk.qp_type != IRDMA_QP_TYPE_ROCE_UD &&
2992                     cq_poll_info->stag_invalid_set) {
2993                         entry->ex.invalidate_rkey = cq_poll_info->inv_stag;
2994                         entry->wc_flags |= IB_WC_WITH_INVALIDATE;
2995                 }
2996         }
2997
2998         if (qp->qp_uk.qp_type == IRDMA_QP_TYPE_ROCE_UD) {
2999                 entry->src_qp = cq_poll_info->ud_src_qpn;
3000                 entry->slid = 0;
3001                 entry->wc_flags |=
3002                     (IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE);
3003                 entry->network_hdr_type = cq_poll_info->ipv4 ?
3004                     RDMA_NETWORK_IPV4 :
3005                     RDMA_NETWORK_IPV6;
3006         } else {
3007                 entry->src_qp = cq_poll_info->qp_id;
3008         }
3009
3010         entry->byte_len = cq_poll_info->bytes_xfered;
3011 }
3012
3013 /**
3014  * irdma_poll_one - poll one entry of the CQ
3015  * @ukcq: ukcq to poll
3016  * @cur_cqe: current CQE info to be filled in
3017  * @entry: ibv_wc object to be filled for non-extended CQ or NULL for extended CQ
3018  *
3019  * Returns the internal irdma device error code or 0 on success
3020  */
3021 static inline int
3022 irdma_poll_one(struct irdma_cq_uk *ukcq,
3023                struct irdma_cq_poll_info *cur_cqe,
3024                struct ib_wc *entry)
3025 {
3026         int ret = irdma_uk_cq_poll_cmpl(ukcq, cur_cqe);
3027
3028         if (ret)
3029                 return ret;
3030
3031         irdma_process_cqe(entry, cur_cqe);
3032
3033         return 0;
3034 }
3035
3036 /**
3037  * __irdma_poll_cq - poll cq for completion (kernel apps)
3038  * @iwcq: cq to poll
3039  * @num_entries: number of entries to poll
3040  * @entry: wr of a completed entry
3041  */
3042 static int
3043 __irdma_poll_cq(struct irdma_cq *iwcq, int num_entries, struct ib_wc *entry)
3044 {
3045         struct list_head *tmp_node, *list_node;
3046         struct irdma_cq_buf *last_buf = NULL;
3047         struct irdma_cq_poll_info *cur_cqe = &iwcq->cur_cqe;
3048         struct irdma_cq_buf *cq_buf;
3049         int ret;
3050         struct irdma_device *iwdev;
3051         struct irdma_cq_uk *ukcq;
3052         bool cq_new_cqe = false;
3053         int resized_bufs = 0;
3054         int npolled = 0;
3055
3056         iwdev = to_iwdev(iwcq->ibcq.device);
3057         ukcq = &iwcq->sc_cq.cq_uk;
3058
3059         /* go through the list of previously resized CQ buffers */
3060         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
3061                 cq_buf = container_of(list_node, struct irdma_cq_buf, list);
3062                 while (npolled < num_entries) {
3063                         ret = irdma_poll_one(&cq_buf->cq_uk, cur_cqe, entry + npolled);
3064                         if (!ret) {
3065                                 ++npolled;
3066                                 cq_new_cqe = true;
3067                                 continue;
3068                         }
3069                         if (ret == -ENOENT)
3070                                 break;
3071                         /* QP using the CQ is destroyed. Skip reporting this CQE */
3072                         if (ret == -EFAULT) {
3073                                 cq_new_cqe = true;
3074                                 continue;
3075                         }
3076                         goto error;
3077                 }
3078
3079                 /* save the resized CQ buffer which received the last cqe */
3080                 if (cq_new_cqe)
3081                         last_buf = cq_buf;
3082                 cq_new_cqe = false;
3083         }
3084
3085         /* check the current CQ for new cqes */
3086         while (npolled < num_entries) {
3087                 ret = irdma_poll_one(ukcq, cur_cqe, entry + npolled);
3088                 if (ret == -ENOENT) {
3089                         ret = irdma_generated_cmpls(iwcq, cur_cqe);
3090                         if (!ret)
3091                                 irdma_process_cqe(entry + npolled, cur_cqe);
3092                 }
3093                 if (!ret) {
3094                         ++npolled;
3095                         cq_new_cqe = true;
3096                         continue;
3097                 }
3098
3099                 if (ret == -ENOENT)
3100                         break;
3101                 /* QP using the CQ is destroyed. Skip reporting this CQE */
3102                 if (ret == -EFAULT) {
3103                         cq_new_cqe = true;
3104                         continue;
3105                 }
3106                 goto error;
3107         }
3108
3109         if (cq_new_cqe)
3110                 /* all previous CQ resizes are complete */
3111                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, NULL);
3112         else if (last_buf)
3113                 /* only CQ resizes up to the last_buf are complete */
3114                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, last_buf);
3115         if (resized_bufs)
3116                 /* report to the HW the number of complete CQ resizes */
3117                 irdma_uk_cq_set_resized_cnt(ukcq, resized_bufs);
3118
3119         return npolled;
3120 error:
3121         irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
3122                     "%s: Error polling CQ, irdma_err: %d\n",
3123                     __func__, ret);
3124
3125         return ret;
3126 }
3127
3128 /**
3129  * irdma_poll_cq - poll cq for completion (kernel apps)
3130  * @ibcq: cq to poll
3131  * @num_entries: number of entries to poll
3132  * @entry: wr of a completed entry
3133  */
3134 static int
3135 irdma_poll_cq(struct ib_cq *ibcq, int num_entries,
3136               struct ib_wc *entry)
3137 {
3138         struct irdma_cq *iwcq;
3139         unsigned long flags;
3140         int ret;
3141
3142         iwcq = to_iwcq(ibcq);
3143
3144         spin_lock_irqsave(&iwcq->lock, flags);
3145         ret = __irdma_poll_cq(iwcq, num_entries, entry);
3146         spin_unlock_irqrestore(&iwcq->lock, flags);
3147
3148         return ret;
3149 }
3150
3151 /**
3152  * irdma_req_notify_cq - arm cq kernel application
3153  * @ibcq: cq to arm
3154  * @notify_flags: notofication flags
3155  */
3156 static int
3157 irdma_req_notify_cq(struct ib_cq *ibcq,
3158                     enum ib_cq_notify_flags notify_flags)
3159 {
3160         struct irdma_cq *iwcq;
3161         struct irdma_cq_uk *ukcq;
3162         unsigned long flags;
3163         enum irdma_cmpl_notify cq_notify = IRDMA_CQ_COMPL_EVENT;
3164         bool promo_event = false;
3165         int ret = 0;
3166
3167         iwcq = to_iwcq(ibcq);
3168         ukcq = &iwcq->sc_cq.cq_uk;
3169
3170         spin_lock_irqsave(&iwcq->lock, flags);
3171         if (notify_flags == IB_CQ_SOLICITED) {
3172                 cq_notify = IRDMA_CQ_COMPL_SOLICITED;
3173         } else {
3174                 if (iwcq->last_notify == IRDMA_CQ_COMPL_SOLICITED)
3175                         promo_event = true;
3176         }
3177
3178         if (!atomic_cmpxchg(&iwcq->armed, 0, 1) || promo_event) {
3179                 iwcq->last_notify = cq_notify;
3180                 irdma_uk_cq_request_notification(ukcq, cq_notify);
3181         }
3182
3183         if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3184             (!irdma_cq_empty(iwcq) || !list_empty(&iwcq->cmpl_generated)))
3185                 ret = 1;
3186         spin_unlock_irqrestore(&iwcq->lock, flags);
3187
3188         return ret;
3189 }
3190
3191 /**
3192  * mcast_list_add -  Add a new mcast item to list
3193  * @rf: RDMA PCI function
3194  * @new_elem: pointer to element to add
3195  */
3196 static void
3197 mcast_list_add(struct irdma_pci_f *rf,
3198                struct mc_table_list *new_elem)
3199 {
3200         list_add(&new_elem->list, &rf->mc_qht_list.list);
3201 }
3202
3203 /**
3204  * mcast_list_del - Remove an mcast item from list
3205  * @mc_qht_elem: pointer to mcast table list element
3206  */
3207 static void
3208 mcast_list_del(struct mc_table_list *mc_qht_elem)
3209 {
3210         if (mc_qht_elem)
3211                 list_del(&mc_qht_elem->list);
3212 }
3213
3214 /**
3215  * mcast_list_lookup_ip - Search mcast list for address
3216  * @rf: RDMA PCI function
3217  * @ip_mcast: pointer to mcast IP address
3218  */
3219 static struct mc_table_list *
3220 mcast_list_lookup_ip(struct irdma_pci_f *rf,
3221                      u32 *ip_mcast)
3222 {
3223         struct mc_table_list *mc_qht_el;
3224         struct list_head *pos, *q;
3225
3226         list_for_each_safe(pos, q, &rf->mc_qht_list.list) {
3227                 mc_qht_el = list_entry(pos, struct mc_table_list, list);
3228                 if (!memcmp(mc_qht_el->mc_info.dest_ip, ip_mcast,
3229                             sizeof(mc_qht_el->mc_info.dest_ip)))
3230                         return mc_qht_el;
3231         }
3232
3233         return NULL;
3234 }
3235
3236 /**
3237  * irdma_mcast_cqp_op - perform a mcast cqp operation
3238  * @iwdev: irdma device
3239  * @mc_grp_ctx: mcast group info
3240  * @op: operation
3241  *
3242  * returns error status
3243  */
3244 static int
3245 irdma_mcast_cqp_op(struct irdma_device *iwdev,
3246                    struct irdma_mcast_grp_info *mc_grp_ctx, u8 op)
3247 {
3248         struct cqp_cmds_info *cqp_info;
3249         struct irdma_cqp_request *cqp_request;
3250         int status;
3251
3252         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3253         if (!cqp_request)
3254                 return -ENOMEM;
3255
3256         cqp_request->info.in.u.mc_create.info = *mc_grp_ctx;
3257         cqp_info = &cqp_request->info;
3258         cqp_info->cqp_cmd = op;
3259         cqp_info->post_sq = 1;
3260         cqp_info->in.u.mc_create.scratch = (uintptr_t)cqp_request;
3261         cqp_info->in.u.mc_create.cqp = &iwdev->rf->cqp.sc_cqp;
3262         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3263         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3264
3265         return status;
3266 }
3267
3268 /**
3269  * irdma_attach_mcast - attach a qp to a multicast group
3270  * @ibqp: ptr to qp
3271  * @ibgid: pointer to global ID
3272  * @lid: local ID
3273  *
3274  * returns error status
3275  */
3276 static int
3277 irdma_attach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3278 {
3279         struct irdma_qp *iwqp = to_iwqp(ibqp);
3280         struct irdma_device *iwdev = iwqp->iwdev;
3281         struct irdma_pci_f *rf = iwdev->rf;
3282         struct mc_table_list *mc_qht_elem;
3283         struct irdma_mcast_grp_ctx_entry_info mcg_info = {0};
3284         unsigned long flags;
3285         u32 ip_addr[4] = {0};
3286         u32 mgn;
3287         u32 no_mgs;
3288         int ret = 0;
3289         bool ipv4;
3290         u16 vlan_id;
3291         union {
3292                 struct sockaddr saddr;
3293                 struct sockaddr_in saddr_in;
3294                 struct sockaddr_in6 saddr_in6;
3295         } sgid_addr;
3296         unsigned char dmac[ETH_ALEN];
3297
3298         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3299
3300         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid)) {
3301                 irdma_copy_ip_ntohl(ip_addr,
3302                                     sgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32);
3303                 irdma_netdev_vlan_ipv6(ip_addr, &vlan_id, NULL);
3304                 ipv4 = false;
3305                 irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
3306                             "qp_id=%d, IP6address=%pI6\n",
3307                             ibqp->qp_num,
3308                             ip_addr);
3309                 irdma_mcast_mac_v6(ip_addr, dmac);
3310         } else {
3311                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3312                 ipv4 = true;
3313                 vlan_id = irdma_get_vlan_ipv4(ip_addr);
3314                 irdma_mcast_mac_v4(ip_addr, dmac);
3315                 irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS,
3316                             "qp_id=%d, IP4address=%pI4, MAC=%pM\n",
3317                             ibqp->qp_num, ip_addr, dmac);
3318         }
3319
3320         spin_lock_irqsave(&rf->qh_list_lock, flags);
3321         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3322         if (!mc_qht_elem) {
3323                 struct irdma_dma_mem *dma_mem_mc;
3324
3325                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3326                 mc_qht_elem = kzalloc(sizeof(*mc_qht_elem), GFP_KERNEL);
3327                 if (!mc_qht_elem)
3328                         return -ENOMEM;
3329
3330                 mc_qht_elem->mc_info.ipv4_valid = ipv4;
3331                 memcpy(mc_qht_elem->mc_info.dest_ip, ip_addr,
3332                        sizeof(mc_qht_elem->mc_info.dest_ip));
3333                 ret = irdma_alloc_rsrc(rf, rf->allocated_mcgs, rf->max_mcg,
3334                                        &mgn, &rf->next_mcg);
3335                 if (ret) {
3336                         kfree(mc_qht_elem);
3337                         return -ENOMEM;
3338                 }
3339
3340                 mc_qht_elem->mc_info.mgn = mgn;
3341                 dma_mem_mc = &mc_qht_elem->mc_grp_ctx.dma_mem_mc;
3342                 dma_mem_mc->size = sizeof(u64)* IRDMA_MAX_MGS_PER_CTX;
3343                 dma_mem_mc->va = irdma_allocate_dma_mem(&rf->hw, dma_mem_mc,
3344                                                         dma_mem_mc->size,
3345                                                         IRDMA_HW_PAGE_SIZE);
3346                 if (!dma_mem_mc->va) {
3347                         irdma_free_rsrc(rf, rf->allocated_mcgs, mgn);
3348                         kfree(mc_qht_elem);
3349                         return -ENOMEM;
3350                 }
3351
3352                 mc_qht_elem->mc_grp_ctx.mg_id = (u16)mgn;
3353                 memcpy(mc_qht_elem->mc_grp_ctx.dest_ip_addr, ip_addr,
3354                        sizeof(mc_qht_elem->mc_grp_ctx.dest_ip_addr));
3355                 mc_qht_elem->mc_grp_ctx.ipv4_valid = ipv4;
3356                 mc_qht_elem->mc_grp_ctx.vlan_id = vlan_id;
3357                 if (vlan_id < VLAN_N_VID)
3358                         mc_qht_elem->mc_grp_ctx.vlan_valid = true;
3359                 mc_qht_elem->mc_grp_ctx.hmc_fcn_id = iwdev->rf->sc_dev.hmc_fn_id;
3360                 mc_qht_elem->mc_grp_ctx.qs_handle =
3361                     iwqp->sc_qp.vsi->qos[iwqp->sc_qp.user_pri].qs_handle;
3362                 ether_addr_copy(mc_qht_elem->mc_grp_ctx.dest_mac_addr, dmac);
3363
3364                 spin_lock_irqsave(&rf->qh_list_lock, flags);
3365                 mcast_list_add(rf, mc_qht_elem);
3366         } else {
3367                 if (mc_qht_elem->mc_grp_ctx.no_of_mgs ==
3368                     IRDMA_MAX_MGS_PER_CTX) {
3369                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3370                         return -ENOMEM;
3371                 }
3372         }
3373
3374         mcg_info.qp_id = iwqp->ibqp.qp_num;
3375         no_mgs = mc_qht_elem->mc_grp_ctx.no_of_mgs;
3376         irdma_sc_add_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3377         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3378
3379         /* Only if there is a change do we need to modify or create */
3380         if (!no_mgs) {
3381                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3382                                          IRDMA_OP_MC_CREATE);
3383         } else if (no_mgs != mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3384                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3385                                          IRDMA_OP_MC_MODIFY);
3386         } else {
3387                 return 0;
3388         }
3389
3390         if (ret)
3391                 goto error;
3392
3393         return 0;
3394
3395 error:
3396         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3397         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3398                 mcast_list_del(mc_qht_elem);
3399                 irdma_free_dma_mem(&rf->hw,
3400                                    &mc_qht_elem->mc_grp_ctx.dma_mem_mc);
3401                 irdma_free_rsrc(rf, rf->allocated_mcgs,
3402                                 mc_qht_elem->mc_grp_ctx.mg_id);
3403                 kfree(mc_qht_elem);
3404         }
3405
3406         return ret;
3407 }
3408
3409 /**
3410  * irdma_detach_mcast - detach a qp from a multicast group
3411  * @ibqp: ptr to qp
3412  * @ibgid: pointer to global ID
3413  * @lid: local ID
3414  *
3415  * returns error status
3416  */
3417 static int
3418 irdma_detach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3419 {
3420         struct irdma_qp *iwqp = to_iwqp(ibqp);
3421         struct irdma_device *iwdev = iwqp->iwdev;
3422         struct irdma_pci_f *rf = iwdev->rf;
3423         u32 ip_addr[4] = {0};
3424         struct mc_table_list *mc_qht_elem;
3425         struct irdma_mcast_grp_ctx_entry_info mcg_info = {0};
3426         int ret;
3427         unsigned long flags;
3428         union {
3429                 struct sockaddr saddr;
3430                 struct sockaddr_in saddr_in;
3431                 struct sockaddr_in6 saddr_in6;
3432         } sgid_addr;
3433
3434         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3435         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid))
3436                 irdma_copy_ip_ntohl(ip_addr,
3437                                     sgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32);
3438         else
3439                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3440
3441         spin_lock_irqsave(&rf->qh_list_lock, flags);
3442         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3443         if (!mc_qht_elem) {
3444                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3445                 irdma_debug(&iwdev->rf->sc_dev,
3446                             IRDMA_DEBUG_VERBS, "address not found MCG\n");
3447                 return 0;
3448         }
3449
3450         mcg_info.qp_id = iwqp->ibqp.qp_num;
3451         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3452         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3453                 mcast_list_del(mc_qht_elem);
3454                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3455                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3456                                          IRDMA_OP_MC_DESTROY);
3457                 if (ret) {
3458                         irdma_debug(&iwdev->rf->sc_dev,
3459                                     IRDMA_DEBUG_VERBS, "failed MC_DESTROY MCG\n");
3460                         spin_lock_irqsave(&rf->qh_list_lock, flags);
3461                         mcast_list_add(rf, mc_qht_elem);
3462                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3463                         return -EAGAIN;
3464                 }
3465
3466                 irdma_free_dma_mem(&rf->hw,
3467                                    &mc_qht_elem->mc_grp_ctx.dma_mem_mc);
3468                 irdma_free_rsrc(rf, rf->allocated_mcgs,
3469                                 mc_qht_elem->mc_grp_ctx.mg_id);
3470                 kfree(mc_qht_elem);
3471         } else {
3472                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3473                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3474                                          IRDMA_OP_MC_MODIFY);
3475                 if (ret) {
3476                         irdma_debug(&iwdev->rf->sc_dev,
3477                                     IRDMA_DEBUG_VERBS, "failed Modify MCG\n");
3478                         return ret;
3479                 }
3480         }
3481
3482         return 0;
3483 }
3484
3485 /**
3486  * irdma_query_ah - Query address handle
3487  * @ibah: pointer to address handle
3488  * @ah_attr: address handle attributes
3489  */
3490 static int
3491 irdma_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
3492 {
3493         struct irdma_ah *ah = to_iwah(ibah);
3494
3495         memset(ah_attr, 0, sizeof(*ah_attr));
3496         if (ah->av.attrs.ah_flags & IB_AH_GRH) {
3497                 ah_attr->ah_flags = IB_AH_GRH;
3498                 ah_attr->grh.flow_label = ah->sc_ah.ah_info.flow_label;
3499                 ah_attr->grh.traffic_class = ah->sc_ah.ah_info.tc_tos;
3500                 ah_attr->grh.hop_limit = ah->sc_ah.ah_info.hop_ttl;
3501                 ah_attr->grh.sgid_index = ah->sgid_index;
3502                 ah_attr->grh.sgid_index = ah->sgid_index;
3503                 memcpy(&ah_attr->grh.dgid, &ah->dgid,
3504                        sizeof(ah_attr->grh.dgid));
3505         }
3506
3507         return 0;
3508 }
3509
3510 static if_t
3511 irdma_get_netdev(struct ib_device *ibdev, u8 port_num)
3512 {
3513         struct irdma_device *iwdev = to_iwdev(ibdev);
3514
3515         if (iwdev->netdev) {
3516                 dev_hold(iwdev->netdev);
3517                 return iwdev->netdev;
3518         }
3519
3520         return NULL;
3521 }
3522
3523 static void
3524 irdma_set_device_ops(struct ib_device *ibdev)
3525 {
3526         struct ib_device *dev_ops = ibdev;
3527
3528 #if __FreeBSD_version >= 1400000
3529         dev_ops->ops.driver_id = RDMA_DRIVER_I40IW;
3530         dev_ops->ops.size_ib_ah = IRDMA_SET_RDMA_OBJ_SIZE(ib_ah, irdma_ah, ibah);
3531         dev_ops->ops.size_ib_cq = IRDMA_SET_RDMA_OBJ_SIZE(ib_cq, irdma_cq, ibcq);
3532         dev_ops->ops.size_ib_pd = IRDMA_SET_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd);
3533         dev_ops->ops.size_ib_ucontext = IRDMA_SET_RDMA_OBJ_SIZE(ib_ucontext,
3534                                                                 irdma_ucontext,
3535                                                                 ibucontext);
3536
3537 #endif                          /* __FreeBSD_version >= 1400000 */
3538         dev_ops->alloc_hw_stats = irdma_alloc_hw_stats;
3539         dev_ops->alloc_mr = irdma_alloc_mr;
3540         dev_ops->alloc_mw = irdma_alloc_mw;
3541         dev_ops->alloc_pd = irdma_alloc_pd;
3542         dev_ops->alloc_ucontext = irdma_alloc_ucontext;
3543         dev_ops->create_cq = irdma_create_cq;
3544         dev_ops->create_qp = irdma_create_qp;
3545         dev_ops->dealloc_mw = irdma_dealloc_mw;
3546         dev_ops->dealloc_pd = irdma_dealloc_pd;
3547         dev_ops->dealloc_ucontext = irdma_dealloc_ucontext;
3548         dev_ops->dereg_mr = irdma_dereg_mr;
3549         dev_ops->destroy_cq = irdma_destroy_cq;
3550         dev_ops->destroy_qp = irdma_destroy_qp;
3551         dev_ops->disassociate_ucontext = irdma_disassociate_ucontext;
3552         dev_ops->get_dev_fw_str = irdma_get_dev_fw_str;
3553         dev_ops->get_dma_mr = irdma_get_dma_mr;
3554         dev_ops->get_hw_stats = irdma_get_hw_stats;
3555         dev_ops->get_netdev = irdma_get_netdev;
3556         dev_ops->map_mr_sg = irdma_map_mr_sg;
3557         dev_ops->mmap = irdma_mmap;
3558 #if __FreeBSD_version >= 1400026
3559         dev_ops->mmap_free = irdma_mmap_free;
3560 #endif
3561         dev_ops->poll_cq = irdma_poll_cq;
3562         dev_ops->post_recv = irdma_post_recv;
3563         dev_ops->post_send = irdma_post_send;
3564         dev_ops->query_device = irdma_query_device;
3565         dev_ops->query_port = irdma_query_port;
3566         dev_ops->modify_port = irdma_modify_port;
3567         dev_ops->query_qp = irdma_query_qp;
3568         dev_ops->reg_user_mr = irdma_reg_user_mr;
3569         dev_ops->rereg_user_mr = irdma_rereg_user_mr;
3570         dev_ops->req_notify_cq = irdma_req_notify_cq;
3571         dev_ops->resize_cq = irdma_resize_cq;
3572 }
3573
3574 static void
3575 irdma_set_device_mcast_ops(struct ib_device *ibdev)
3576 {
3577         struct ib_device *dev_ops = ibdev;
3578         dev_ops->attach_mcast = irdma_attach_mcast;
3579         dev_ops->detach_mcast = irdma_detach_mcast;
3580 }
3581
3582 static void
3583 irdma_set_device_roce_ops(struct ib_device *ibdev)
3584 {
3585         struct ib_device *dev_ops = ibdev;
3586         dev_ops->create_ah = irdma_create_ah;
3587         dev_ops->destroy_ah = irdma_destroy_ah;
3588         dev_ops->get_link_layer = irdma_get_link_layer;
3589         dev_ops->get_port_immutable = irdma_roce_port_immutable;
3590         dev_ops->modify_qp = irdma_modify_qp_roce;
3591         dev_ops->query_ah = irdma_query_ah;
3592         dev_ops->query_gid = irdma_query_gid_roce;
3593         dev_ops->query_pkey = irdma_query_pkey;
3594         ibdev->add_gid = irdma_add_gid;
3595         ibdev->del_gid = irdma_del_gid;
3596 }
3597
3598 static void
3599 irdma_set_device_iw_ops(struct ib_device *ibdev)
3600 {
3601         struct ib_device *dev_ops = ibdev;
3602
3603         ibdev->uverbs_cmd_mask |=
3604             (1ull << IB_USER_VERBS_CMD_CREATE_AH) |
3605             (1ull << IB_USER_VERBS_CMD_DESTROY_AH);
3606
3607         dev_ops->create_ah = irdma_create_ah_stub;
3608         dev_ops->destroy_ah = irdma_destroy_ah_stub;
3609         dev_ops->get_port_immutable = irdma_iw_port_immutable;
3610         dev_ops->modify_qp = irdma_modify_qp;
3611         dev_ops->query_gid = irdma_query_gid;
3612         dev_ops->query_pkey = irdma_iw_query_pkey;
3613 }
3614
3615 static inline void
3616 irdma_set_device_gen1_ops(struct ib_device *ibdev)
3617 {
3618 }
3619
3620 /**
3621  * irdma_init_roce_device - initialization of roce rdma device
3622  * @iwdev: irdma device
3623  */
3624 static void
3625 irdma_init_roce_device(struct irdma_device *iwdev)
3626 {
3627         kc_set_roce_uverbs_cmd_mask(iwdev);
3628         iwdev->ibdev.node_type = RDMA_NODE_IB_CA;
3629         addrconf_addr_eui48((u8 *)&iwdev->ibdev.node_guid,
3630                             if_getlladdr(iwdev->netdev));
3631         irdma_set_device_roce_ops(&iwdev->ibdev);
3632         if (iwdev->rf->rdma_ver == IRDMA_GEN_2)
3633                 irdma_set_device_mcast_ops(&iwdev->ibdev);
3634 }
3635
3636 /**
3637  * irdma_init_iw_device - initialization of iwarp rdma device
3638  * @iwdev: irdma device
3639  */
3640 static int
3641 irdma_init_iw_device(struct irdma_device *iwdev)
3642 {
3643         if_t netdev = iwdev->netdev;
3644
3645         iwdev->ibdev.node_type = RDMA_NODE_RNIC;
3646         addrconf_addr_eui48((u8 *)&iwdev->ibdev.node_guid,
3647                             if_getlladdr(netdev));
3648         iwdev->ibdev.iwcm = kzalloc(sizeof(*iwdev->ibdev.iwcm), GFP_KERNEL);
3649         if (!iwdev->ibdev.iwcm)
3650                 return -ENOMEM;
3651
3652         iwdev->ibdev.iwcm->add_ref = irdma_qp_add_ref;
3653         iwdev->ibdev.iwcm->rem_ref = irdma_qp_rem_ref;
3654         iwdev->ibdev.iwcm->get_qp = irdma_get_qp;
3655         iwdev->ibdev.iwcm->connect = irdma_connect;
3656         iwdev->ibdev.iwcm->accept = irdma_accept;
3657         iwdev->ibdev.iwcm->reject = irdma_reject;
3658         iwdev->ibdev.iwcm->create_listen = irdma_create_listen;
3659         iwdev->ibdev.iwcm->destroy_listen = irdma_destroy_listen;
3660         memcpy(iwdev->ibdev.iwcm->ifname, if_name(netdev),
3661                sizeof(iwdev->ibdev.iwcm->ifname));
3662         irdma_set_device_iw_ops(&iwdev->ibdev);
3663
3664         return 0;
3665 }
3666
3667 /**
3668  * irdma_init_rdma_device - initialization of rdma device
3669  * @iwdev: irdma device
3670  */
3671 static int
3672 irdma_init_rdma_device(struct irdma_device *iwdev)
3673 {
3674         int ret;
3675
3676         iwdev->ibdev.owner = THIS_MODULE;
3677         iwdev->ibdev.uverbs_abi_ver = IRDMA_ABI_VER;
3678         kc_set_rdma_uverbs_cmd_mask(iwdev);
3679
3680         if (iwdev->roce_mode) {
3681                 irdma_init_roce_device(iwdev);
3682         } else {
3683                 ret = irdma_init_iw_device(iwdev);
3684                 if (ret)
3685                         return ret;
3686         }
3687
3688         iwdev->ibdev.phys_port_cnt = 1;
3689         iwdev->ibdev.num_comp_vectors = iwdev->rf->ceqs_count;
3690         iwdev->ibdev.dev.parent = iwdev->rf->dev_ctx.dev;
3691         set_ibdev_dma_device(iwdev->ibdev, &iwdev->rf->pcidev->dev);
3692         irdma_set_device_ops(&iwdev->ibdev);
3693         if (iwdev->rf->rdma_ver == IRDMA_GEN_1)
3694                 irdma_set_device_gen1_ops(&iwdev->ibdev);
3695
3696         return 0;
3697 }
3698
3699 /**
3700  * irdma_port_ibevent - indicate port event
3701  * @iwdev: irdma device
3702  */
3703 void
3704 irdma_port_ibevent(struct irdma_device *iwdev)
3705 {
3706         struct ib_event event;
3707
3708         event.device = &iwdev->ibdev;
3709         event.element.port_num = 1;
3710         event.event =
3711             iwdev->iw_status ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
3712         ib_dispatch_event(&event);
3713 }
3714
3715 /**
3716  * irdma_ib_unregister_device - unregister rdma device from IB
3717  * core
3718  * @iwdev: irdma device
3719  */
3720 void
3721 irdma_ib_unregister_device(struct irdma_device *iwdev)
3722 {
3723         iwdev->iw_status = 0;
3724         irdma_port_ibevent(iwdev);
3725         ib_unregister_device(&iwdev->ibdev);
3726         dev_put(iwdev->netdev);
3727         kfree(iwdev->ibdev.iwcm);
3728         iwdev->ibdev.iwcm = NULL;
3729 }
3730
3731 /**
3732  * irdma_ib_register_device - register irdma device to IB core
3733  * @iwdev: irdma device
3734  */
3735 int
3736 irdma_ib_register_device(struct irdma_device *iwdev)
3737 {
3738         int ret;
3739
3740         ret = irdma_init_rdma_device(iwdev);
3741         if (ret)
3742                 return ret;
3743
3744         dev_hold(iwdev->netdev);
3745         sprintf(iwdev->ibdev.name, "irdma-%s", if_name(iwdev->netdev));
3746         ret = ib_register_device(&iwdev->ibdev, NULL);
3747         if (ret)
3748                 goto error;
3749
3750         iwdev->iw_status = 1;
3751         irdma_port_ibevent(iwdev);
3752
3753         return 0;
3754
3755 error:
3756         kfree(iwdev->ibdev.iwcm);
3757         iwdev->ibdev.iwcm = NULL;
3758         irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS, "Register RDMA device fail\n");
3759
3760         return ret;
3761 }