]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/dev/irdma/irdma_verbs.c
Merge llvm-project release/15.x llvmorg-15.0.7-0-g8dfdcc7b7bf6
[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         ether_addr_copy((u8 *)&props->sys_image_guid, IF_LLADDR(iwdev->netdev));
59         props->fw_ver = (u64)irdma_fw_major_ver(&rf->sc_dev) << 32 |
60             irdma_fw_minor_ver(&rf->sc_dev);
61         props->device_cap_flags = IB_DEVICE_MEM_WINDOW |
62             IB_DEVICE_MEM_MGT_EXTENSIONS;
63         props->device_cap_flags |= IB_DEVICE_LOCAL_DMA_LKEY;
64         props->vendor_id = pcidev->vendor;
65         props->vendor_part_id = pcidev->device;
66         props->hw_ver = pcidev->revision;
67         props->page_size_cap = hw_attrs->page_size_cap;
68         props->max_mr_size = hw_attrs->max_mr_size;
69         props->max_qp = rf->max_qp - rf->used_qps;
70         props->max_qp_wr = hw_attrs->max_qp_wr;
71         set_max_sge(props, rf);
72         props->max_cq = rf->max_cq - rf->used_cqs;
73         props->max_cqe = rf->max_cqe - 1;
74         props->max_mr = rf->max_mr - rf->used_mrs;
75         props->max_mw = props->max_mr;
76         props->max_pd = rf->max_pd - rf->used_pds;
77         props->max_sge_rd = hw_attrs->uk_attrs.max_hw_read_sges;
78         props->max_qp_rd_atom = hw_attrs->max_hw_ird;
79         props->max_qp_init_rd_atom = hw_attrs->max_hw_ord;
80         if (rdma_protocol_roce(ibdev, 1)) {
81                 props->device_cap_flags |= IB_DEVICE_RC_RNR_NAK_GEN;
82                 props->max_pkeys = IRDMA_PKEY_TBL_SZ;
83                 props->max_ah = rf->max_ah;
84                 if (hw_attrs->uk_attrs.hw_rev == IRDMA_GEN_2) {
85                         props->max_mcast_grp = rf->max_mcg;
86                         props->max_mcast_qp_attach = IRDMA_MAX_MGS_PER_CTX;
87                         props->max_total_mcast_qp_attach = rf->max_qp * IRDMA_MAX_MGS_PER_CTX;
88                 }
89         }
90         props->max_fast_reg_page_list_len = IRDMA_MAX_PAGES_PER_FMR;
91         if (hw_attrs->uk_attrs.hw_rev >= IRDMA_GEN_2)
92                 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW_TYPE_2B;
93
94         return 0;
95 }
96
97 static int
98 irdma_mmap_legacy(struct irdma_ucontext *ucontext,
99                   struct vm_area_struct *vma)
100 {
101         u64 pfn;
102
103         if (vma->vm_pgoff || vma->vm_end - vma->vm_start != PAGE_SIZE)
104                 return -EINVAL;
105
106         vma->vm_private_data = ucontext;
107         pfn = ((uintptr_t)ucontext->iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET] +
108                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
109
110 #if __FreeBSD_version >= 1400026
111         return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
112                                  pgprot_noncached(vma->vm_page_prot), NULL);
113 #else
114         return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
115                                  pgprot_noncached(vma->vm_page_prot));
116 #endif
117 }
118
119 #if __FreeBSD_version >= 1400026
120 static void
121 irdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
122 {
123         struct irdma_user_mmap_entry *entry = to_irdma_mmap_entry(rdma_entry);
124
125         kfree(entry);
126 }
127
128 struct rdma_user_mmap_entry *
129 irdma_user_mmap_entry_insert(struct irdma_ucontext *ucontext, u64 bar_offset,
130                              enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
131 {
132         struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
133         int ret;
134
135         if (!entry)
136                 return NULL;
137
138         entry->bar_offset = bar_offset;
139         entry->mmap_flag = mmap_flag;
140
141         ret = rdma_user_mmap_entry_insert(&ucontext->ibucontext,
142                                           &entry->rdma_entry, PAGE_SIZE);
143         if (ret) {
144                 kfree(entry);
145                 return NULL;
146         }
147         *mmap_offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
148
149         return &entry->rdma_entry;
150 }
151
152 #else
153 static inline bool
154 find_key_in_mmap_tbl(struct irdma_ucontext *ucontext, u64 key)
155 {
156         struct irdma_user_mmap_entry *entry;
157
158         HASH_FOR_EACH_POSSIBLE(ucontext->mmap_hash_tbl, entry, hlist, key) {
159                 if (entry->pgoff_key == key)
160                         return true;
161         }
162
163         return false;
164 }
165
166 struct irdma_user_mmap_entry *
167 irdma_user_mmap_entry_add_hash(struct irdma_ucontext *ucontext, u64 bar_offset,
168                                enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
169 {
170         struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
171         unsigned long flags;
172         int retry_cnt = 0;
173
174         if (!entry)
175                 return NULL;
176
177         entry->bar_offset = bar_offset;
178         entry->mmap_flag = mmap_flag;
179         entry->ucontext = ucontext;
180         do {
181                 get_random_bytes(&entry->pgoff_key, sizeof(entry->pgoff_key));
182
183                 /* The key is a page offset */
184                 entry->pgoff_key >>= PAGE_SHIFT;
185
186                 /* In the event of a collision in the hash table, retry a new key */
187                 spin_lock_irqsave(&ucontext->mmap_tbl_lock, flags);
188                 if (!find_key_in_mmap_tbl(ucontext, entry->pgoff_key)) {
189                         HASH_ADD(ucontext->mmap_hash_tbl, &entry->hlist, entry->pgoff_key);
190                         spin_unlock_irqrestore(&ucontext->mmap_tbl_lock, flags);
191                         goto hash_add_done;
192                 }
193                 spin_unlock_irqrestore(&ucontext->mmap_tbl_lock, flags);
194         } while (retry_cnt++ < 10);
195
196         irdma_debug(iwdev_to_idev(ucontext->iwdev), IRDMA_DEBUG_VERBS, "mmap table add failed: Cannot find a unique key\n");
197         kfree(entry);
198         return NULL;
199
200 hash_add_done:
201         /* libc mmap uses a byte offset */
202         *mmap_offset = entry->pgoff_key << PAGE_SHIFT;
203
204         return entry;
205 }
206
207 static struct irdma_user_mmap_entry *
208 irdma_find_user_mmap_entry(struct irdma_ucontext *ucontext,
209                            struct vm_area_struct *vma)
210 {
211         struct irdma_user_mmap_entry *entry;
212         unsigned long flags;
213
214         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
215                 return NULL;
216
217         spin_lock_irqsave(&ucontext->mmap_tbl_lock, flags);
218         HASH_FOR_EACH_POSSIBLE(ucontext->mmap_hash_tbl, entry, hlist, vma->vm_pgoff) {
219                 if (entry->pgoff_key == vma->vm_pgoff) {
220                         spin_unlock_irqrestore(&ucontext->mmap_tbl_lock, flags);
221                         return entry;
222                 }
223         }
224
225         spin_unlock_irqrestore(&ucontext->mmap_tbl_lock, flags);
226
227         return NULL;
228 }
229
230 void
231 irdma_user_mmap_entry_del_hash(struct irdma_user_mmap_entry *entry)
232 {
233         struct irdma_ucontext *ucontext;
234         unsigned long flags;
235
236         if (!entry)
237                 return;
238
239         ucontext = entry->ucontext;
240
241         spin_lock_irqsave(&ucontext->mmap_tbl_lock, flags);
242         HASH_DEL(ucontext->mmap_hash_tbl, &entry->hlist);
243         spin_unlock_irqrestore(&ucontext->mmap_tbl_lock, flags);
244
245         kfree(entry);
246 }
247
248 #endif
249 /**
250  * irdma_mmap - user memory map
251  * @context: context created during alloc
252  * @vma: kernel info for user memory map
253  */
254 static int
255 irdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
256 {
257 #if __FreeBSD_version >= 1400026
258         struct rdma_user_mmap_entry *rdma_entry;
259 #endif
260         struct irdma_user_mmap_entry *entry;
261         struct irdma_ucontext *ucontext;
262         u64 pfn;
263         int ret;
264
265         ucontext = to_ucontext(context);
266
267         /* Legacy support for libi40iw with hard-coded mmap key */
268         if (ucontext->legacy_mode)
269                 return irdma_mmap_legacy(ucontext, vma);
270
271 #if __FreeBSD_version >= 1400026
272         rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
273         if (!rdma_entry) {
274                 irdma_debug(iwdev_to_idev(ucontext->iwdev), IRDMA_DEBUG_VERBS,
275                             "pgoff[0x%lx] does not have valid entry\n",
276                             vma->vm_pgoff);
277                 return -EINVAL;
278         }
279
280         entry = to_irdma_mmap_entry(rdma_entry);
281 #else
282         entry = irdma_find_user_mmap_entry(ucontext, vma);
283         if (!entry) {
284                 irdma_debug(iwdev_to_idev(ucontext->iwdev), IRDMA_DEBUG_VERBS,
285                             "pgoff[0x%lx] does not have valid entry\n",
286                             vma->vm_pgoff);
287                 return -EINVAL;
288         }
289 #endif
290         irdma_debug(iwdev_to_idev(ucontext->iwdev), IRDMA_DEBUG_VERBS,
291                     "bar_offset [0x%lx] mmap_flag [%d]\n", entry->bar_offset,
292                     entry->mmap_flag);
293
294         pfn = (entry->bar_offset +
295                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
296
297         switch (entry->mmap_flag) {
298         case IRDMA_MMAP_IO_NC:
299 #if __FreeBSD_version >= 1400026
300                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
301                                         pgprot_noncached(vma->vm_page_prot),
302                                         rdma_entry);
303 #else
304                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
305                                         pgprot_noncached(vma->vm_page_prot));
306 #endif
307                 break;
308         case IRDMA_MMAP_IO_WC:
309 #if __FreeBSD_version >= 1400026
310                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
311                                         pgprot_writecombine(vma->vm_page_prot),
312                                         rdma_entry);
313 #else
314                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
315                                         pgprot_writecombine(vma->vm_page_prot));
316 #endif
317                 break;
318         default:
319                 ret = -EINVAL;
320         }
321
322         if (ret)
323                 irdma_debug(iwdev_to_idev(ucontext->iwdev), IRDMA_DEBUG_VERBS,
324                             "bar_offset [0x%lx] mmap_flag[%d] err[%d]\n",
325                             entry->bar_offset, entry->mmap_flag, ret);
326 #if __FreeBSD_version >= 1400026
327         rdma_user_mmap_entry_put(rdma_entry);
328 #endif
329
330         return ret;
331 }
332
333 /**
334  * irdma_alloc_push_page - allocate a push page for qp
335  * @iwqp: qp pointer
336  */
337 static void
338 irdma_alloc_push_page(struct irdma_qp *iwqp)
339 {
340         struct irdma_cqp_request *cqp_request;
341         struct cqp_cmds_info *cqp_info;
342         struct irdma_device *iwdev = iwqp->iwdev;
343         struct irdma_sc_qp *qp = &iwqp->sc_qp;
344         int status;
345
346         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
347         if (!cqp_request)
348                 return;
349
350         cqp_info = &cqp_request->info;
351         cqp_info->cqp_cmd = IRDMA_OP_MANAGE_PUSH_PAGE;
352         cqp_info->post_sq = 1;
353         cqp_info->in.u.manage_push_page.info.push_idx = 0;
354         cqp_info->in.u.manage_push_page.info.qs_handle =
355             qp->vsi->qos[qp->user_pri].qs_handle;
356         cqp_info->in.u.manage_push_page.info.free_page = 0;
357         cqp_info->in.u.manage_push_page.info.push_page_type = 0;
358         cqp_info->in.u.manage_push_page.cqp = &iwdev->rf->cqp.sc_cqp;
359         cqp_info->in.u.manage_push_page.scratch = (uintptr_t)cqp_request;
360
361         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
362         if (!status && cqp_request->compl_info.op_ret_val <
363             iwdev->rf->sc_dev.hw_attrs.max_hw_device_pages) {
364                 qp->push_idx = cqp_request->compl_info.op_ret_val;
365                 qp->push_offset = 0;
366         }
367
368         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
369 }
370
371 /**
372  * irdma_get_pbl - Retrieve pbl from a list given a virtual
373  * address
374  * @va: user virtual address
375  * @pbl_list: pbl list to search in (QP's or CQ's)
376  */
377 struct irdma_pbl *
378 irdma_get_pbl(unsigned long va,
379               struct list_head *pbl_list)
380 {
381         struct irdma_pbl *iwpbl;
382
383         list_for_each_entry(iwpbl, pbl_list, list) {
384                 if (iwpbl->user_base == va) {
385                         list_del(&iwpbl->list);
386                         iwpbl->on_list = false;
387                         return iwpbl;
388                 }
389         }
390
391         return NULL;
392 }
393
394 /**
395  * irdma_clean_cqes - clean cq entries for qp
396  * @iwqp: qp ptr (user or kernel)
397  * @iwcq: cq ptr
398  */
399 void
400 irdma_clean_cqes(struct irdma_qp *iwqp, struct irdma_cq *iwcq)
401 {
402         struct irdma_cq_uk *ukcq = &iwcq->sc_cq.cq_uk;
403         unsigned long flags;
404
405         spin_lock_irqsave(&iwcq->lock, flags);
406         irdma_uk_clean_cq(&iwqp->sc_qp.qp_uk, ukcq);
407         spin_unlock_irqrestore(&iwcq->lock, flags);
408 }
409
410 static u64 irdma_compute_push_wqe_offset(struct irdma_device *iwdev, u32 page_idx){
411         u64 bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
412
413         if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.hw_rev == IRDMA_GEN_2) {
414                 /* skip over db page */
415                 bar_off += IRDMA_HW_PAGE_SIZE;
416                 /* skip over reserved space */
417                 bar_off += IRDMA_PF_BAR_RSVD;
418         }
419
420         /* push wqe page */
421         bar_off += (u64)page_idx * IRDMA_HW_PAGE_SIZE;
422
423         return bar_off;
424 }
425
426 void
427 irdma_remove_push_mmap_entries(struct irdma_qp *iwqp)
428 {
429         if (iwqp->push_db_mmap_entry) {
430 #if __FreeBSD_version >= 1400026
431                 rdma_user_mmap_entry_remove(iwqp->push_db_mmap_entry);
432 #else
433                 irdma_user_mmap_entry_del_hash(iwqp->push_db_mmap_entry);
434 #endif
435                 iwqp->push_db_mmap_entry = NULL;
436         }
437         if (iwqp->push_wqe_mmap_entry) {
438 #if __FreeBSD_version >= 1400026
439                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
440 #else
441                 irdma_user_mmap_entry_del_hash(iwqp->push_wqe_mmap_entry);
442 #endif
443                 iwqp->push_wqe_mmap_entry = NULL;
444         }
445 }
446
447 static int
448 irdma_setup_push_mmap_entries(struct irdma_ucontext *ucontext,
449                               struct irdma_qp *iwqp,
450                               u64 *push_wqe_mmap_key,
451                               u64 *push_db_mmap_key)
452 {
453         struct irdma_device *iwdev = ucontext->iwdev;
454         u64 bar_off;
455
456         WARN_ON_ONCE(iwdev->rf->sc_dev.hw_attrs.uk_attrs.hw_rev < IRDMA_GEN_2);
457
458         bar_off = irdma_compute_push_wqe_offset(iwdev, iwqp->sc_qp.push_idx);
459
460 #if __FreeBSD_version >= 1400026
461         iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
462                                                                  bar_off, IRDMA_MMAP_IO_WC,
463                                                                  push_wqe_mmap_key);
464 #else
465         iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_add_hash(ucontext, bar_off,
466                                                                    IRDMA_MMAP_IO_WC, push_wqe_mmap_key);
467 #endif
468         if (!iwqp->push_wqe_mmap_entry)
469                 return -ENOMEM;
470
471         /* push doorbell page */
472         bar_off += IRDMA_HW_PAGE_SIZE;
473 #if __FreeBSD_version >= 1400026
474         iwqp->push_db_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
475                                                                 bar_off, IRDMA_MMAP_IO_NC,
476                                                                 push_db_mmap_key);
477 #else
478
479         iwqp->push_db_mmap_entry = irdma_user_mmap_entry_add_hash(ucontext, bar_off,
480                                                                   IRDMA_MMAP_IO_NC, push_db_mmap_key);
481 #endif
482         if (!iwqp->push_db_mmap_entry) {
483 #if __FreeBSD_version >= 1400026
484                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
485 #else
486                 irdma_user_mmap_entry_del_hash(iwqp->push_wqe_mmap_entry);
487 #endif
488                 return -ENOMEM;
489         }
490
491         return 0;
492 }
493
494 /**
495  * irdma_setup_virt_qp - setup for allocation of virtual qp
496  * @iwdev: irdma device
497  * @iwqp: qp ptr
498  * @init_info: initialize info to return
499  */
500 void
501 irdma_setup_virt_qp(struct irdma_device *iwdev,
502                     struct irdma_qp *iwqp,
503                     struct irdma_qp_init_info *init_info)
504 {
505         struct irdma_pbl *iwpbl = iwqp->iwpbl;
506         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
507
508         iwqp->page = qpmr->sq_page;
509         init_info->shadow_area_pa = qpmr->shadow;
510         if (iwpbl->pbl_allocated) {
511                 init_info->virtual_map = true;
512                 init_info->sq_pa = qpmr->sq_pbl.idx;
513                 init_info->rq_pa = qpmr->rq_pbl.idx;
514         } else {
515                 init_info->sq_pa = qpmr->sq_pbl.addr;
516                 init_info->rq_pa = qpmr->rq_pbl.addr;
517         }
518 }
519
520 /**
521  * irdma_setup_umode_qp - setup sq and rq size in user mode qp
522  * @udata: user data
523  * @iwdev: iwarp device
524  * @iwqp: qp ptr (user or kernel)
525  * @info: initialize info to return
526  * @init_attr: Initial QP create attributes
527  */
528 int
529 irdma_setup_umode_qp(struct ib_udata *udata,
530                      struct irdma_device *iwdev,
531                      struct irdma_qp *iwqp,
532                      struct irdma_qp_init_info *info,
533                      struct ib_qp_init_attr *init_attr)
534 {
535         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
536         struct irdma_create_qp_req req = {0};
537         unsigned long flags;
538         int ret;
539
540         ret = ib_copy_from_udata(&req, udata,
541                                  min(sizeof(req), udata->inlen));
542         if (ret) {
543                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
544                             "ib_copy_from_data fail\n");
545                 return ret;
546         }
547
548         iwqp->ctx_info.qp_compl_ctx = req.user_compl_ctx;
549         iwqp->user_mode = 1;
550         if (req.user_wqe_bufs) {
551 #if __FreeBSD_version >= 1400026
552                 struct irdma_ucontext *ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
553 #else
554                 struct irdma_ucontext *ucontext = to_ucontext(iwqp->iwpd->ibpd.uobject->context);
555 #endif
556                 info->qp_uk_init_info.legacy_mode = ucontext->legacy_mode;
557                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
558                 iwqp->iwpbl = irdma_get_pbl((unsigned long)req.user_wqe_bufs,
559                                             &ucontext->qp_reg_mem_list);
560                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
561
562                 if (!iwqp->iwpbl) {
563                         ret = -ENODATA;
564                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
565                                     "no pbl info\n");
566                         return ret;
567                 }
568         }
569
570         if (ukinfo->abi_ver <= 5) {
571                 /**
572                  * For ABI version less than 6 passes raw sq and rq
573                  * quanta in cap.max_send_wr and cap.max_recv_wr.
574                  */
575                 iwqp->max_send_wr = init_attr->cap.max_send_wr;
576                 iwqp->max_recv_wr = init_attr->cap.max_recv_wr;
577                 ukinfo->sq_size = init_attr->cap.max_send_wr;
578                 ukinfo->rq_size = init_attr->cap.max_recv_wr;
579                 irdma_uk_calc_shift_wq(ukinfo, &ukinfo->sq_shift, &ukinfo->rq_shift);
580         } else {
581                 ret = irdma_uk_calc_depth_shift_sq(ukinfo, &ukinfo->sq_depth,
582                                                    &ukinfo->sq_shift);
583                 if (ret)
584                         return ret;
585
586                 ret = irdma_uk_calc_depth_shift_rq(ukinfo, &ukinfo->rq_depth,
587                                                    &ukinfo->rq_shift);
588                 if (ret)
589                         return ret;
590
591                 iwqp->max_send_wr = (ukinfo->sq_depth - IRDMA_SQ_RSVD) >> ukinfo->sq_shift;
592                 iwqp->max_recv_wr = (ukinfo->rq_depth - IRDMA_RQ_RSVD) >> ukinfo->rq_shift;
593                 ukinfo->sq_size = ukinfo->sq_depth >> ukinfo->sq_shift;
594                 ukinfo->rq_size = ukinfo->rq_depth >> ukinfo->rq_shift;
595         }
596         irdma_setup_virt_qp(iwdev, iwqp, info);
597
598         return 0;
599 }
600
601 /**
602  * irdma_setup_kmode_qp - setup initialization for kernel mode qp
603  * @iwdev: iwarp device
604  * @iwqp: qp ptr (user or kernel)
605  * @info: initialize info to return
606  * @init_attr: Initial QP create attributes
607  */
608 int
609 irdma_setup_kmode_qp(struct irdma_device *iwdev,
610                      struct irdma_qp *iwqp,
611                      struct irdma_qp_init_info *info,
612                      struct ib_qp_init_attr *init_attr)
613 {
614         struct irdma_dma_mem *mem = &iwqp->kqp.dma_mem;
615         u32 size;
616         int status;
617         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
618
619         status = irdma_uk_calc_depth_shift_sq(ukinfo, &ukinfo->sq_depth,
620                                               &ukinfo->sq_shift);
621         if (status)
622                 return status;
623
624         status = irdma_uk_calc_depth_shift_rq(ukinfo, &ukinfo->rq_depth,
625                                               &ukinfo->rq_shift);
626         if (status)
627                 return status;
628
629         iwqp->kqp.sq_wrid_mem =
630             kcalloc(ukinfo->sq_depth, sizeof(*iwqp->kqp.sq_wrid_mem), GFP_KERNEL);
631         if (!iwqp->kqp.sq_wrid_mem)
632                 return -ENOMEM;
633
634         iwqp->kqp.rq_wrid_mem =
635             kcalloc(ukinfo->rq_depth, sizeof(*iwqp->kqp.rq_wrid_mem), GFP_KERNEL);
636         if (!iwqp->kqp.rq_wrid_mem) {
637                 kfree(iwqp->kqp.sq_wrid_mem);
638                 iwqp->kqp.sq_wrid_mem = NULL;
639                 return -ENOMEM;
640         }
641
642         iwqp->kqp.sig_trk_mem = kcalloc(ukinfo->sq_depth, sizeof(u32), GFP_KERNEL);
643         memset(iwqp->kqp.sig_trk_mem, 0, ukinfo->sq_depth * sizeof(u32));
644         if (!iwqp->kqp.sig_trk_mem) {
645                 kfree(iwqp->kqp.sq_wrid_mem);
646                 iwqp->kqp.sq_wrid_mem = NULL;
647                 kfree(iwqp->kqp.rq_wrid_mem);
648                 iwqp->kqp.rq_wrid_mem = NULL;
649                 return -ENOMEM;
650         }
651         ukinfo->sq_sigwrtrk_array = (void *)iwqp->kqp.sig_trk_mem;
652         ukinfo->sq_wrtrk_array = iwqp->kqp.sq_wrid_mem;
653         ukinfo->rq_wrid_array = iwqp->kqp.rq_wrid_mem;
654
655         size = (ukinfo->sq_depth + ukinfo->rq_depth) * IRDMA_QP_WQE_MIN_SIZE;
656         size += (IRDMA_SHADOW_AREA_SIZE << 3);
657
658         mem->size = size;
659         mem->va = irdma_allocate_dma_mem(&iwdev->rf->hw, mem, mem->size,
660                                          256);
661         if (!mem->va) {
662                 kfree(iwqp->kqp.sq_wrid_mem);
663                 iwqp->kqp.sq_wrid_mem = NULL;
664                 kfree(iwqp->kqp.rq_wrid_mem);
665                 iwqp->kqp.rq_wrid_mem = NULL;
666                 return -ENOMEM;
667         }
668
669         ukinfo->sq = mem->va;
670         info->sq_pa = mem->pa;
671         ukinfo->rq = &ukinfo->sq[ukinfo->sq_depth];
672         info->rq_pa = info->sq_pa + (ukinfo->sq_depth * IRDMA_QP_WQE_MIN_SIZE);
673         ukinfo->shadow_area = ukinfo->rq[ukinfo->rq_depth].elem;
674         info->shadow_area_pa = info->rq_pa + (ukinfo->rq_depth * IRDMA_QP_WQE_MIN_SIZE);
675         ukinfo->sq_size = ukinfo->sq_depth >> ukinfo->sq_shift;
676         ukinfo->rq_size = ukinfo->rq_depth >> ukinfo->rq_shift;
677         ukinfo->qp_id = iwqp->ibqp.qp_num;
678
679         iwqp->max_send_wr = (ukinfo->sq_depth - IRDMA_SQ_RSVD) >> ukinfo->sq_shift;
680         iwqp->max_recv_wr = (ukinfo->rq_depth - IRDMA_RQ_RSVD) >> ukinfo->rq_shift;
681         init_attr->cap.max_send_wr = iwqp->max_send_wr;
682         init_attr->cap.max_recv_wr = iwqp->max_recv_wr;
683
684         return 0;
685 }
686
687 int
688 irdma_cqp_create_qp_cmd(struct irdma_qp *iwqp)
689 {
690         struct irdma_pci_f *rf = iwqp->iwdev->rf;
691         struct irdma_cqp_request *cqp_request;
692         struct cqp_cmds_info *cqp_info;
693         struct irdma_create_qp_info *qp_info;
694         int status;
695
696         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
697         if (!cqp_request)
698                 return -ENOMEM;
699
700         cqp_info = &cqp_request->info;
701         qp_info = &cqp_request->info.in.u.qp_create.info;
702         memset(qp_info, 0, sizeof(*qp_info));
703         qp_info->mac_valid = true;
704         qp_info->cq_num_valid = true;
705         qp_info->next_iwarp_state = IRDMA_QP_STATE_IDLE;
706
707         cqp_info->cqp_cmd = IRDMA_OP_QP_CREATE;
708         cqp_info->post_sq = 1;
709         cqp_info->in.u.qp_create.qp = &iwqp->sc_qp;
710         cqp_info->in.u.qp_create.scratch = (uintptr_t)cqp_request;
711         status = irdma_handle_cqp_op(rf, cqp_request);
712         irdma_put_cqp_request(&rf->cqp, cqp_request);
713
714         return status;
715 }
716
717 void
718 irdma_roce_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
719                                    struct irdma_qp_host_ctx_info *ctx_info)
720 {
721         struct irdma_device *iwdev = iwqp->iwdev;
722         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
723         struct irdma_roce_offload_info *roce_info;
724         struct irdma_udp_offload_info *udp_info;
725
726         udp_info = &iwqp->udp_info;
727         udp_info->snd_mss = ib_mtu_enum_to_int(ib_mtu_int_to_enum(iwdev->vsi.mtu));
728         udp_info->cwnd = iwdev->roce_cwnd;
729         udp_info->rexmit_thresh = 2;
730         udp_info->rnr_nak_thresh = 2;
731         udp_info->src_port = 0xc000;
732         udp_info->dst_port = ROCE_V2_UDP_DPORT;
733         roce_info = &iwqp->roce_info;
734         ether_addr_copy(roce_info->mac_addr, IF_LLADDR(iwdev->netdev));
735
736         roce_info->rd_en = true;
737         roce_info->wr_rdresp_en = true;
738         roce_info->bind_en = true;
739         roce_info->dcqcn_en = false;
740         roce_info->rtomin = 5;
741
742         roce_info->ack_credits = iwdev->roce_ackcreds;
743         roce_info->ird_size = dev->hw_attrs.max_hw_ird;
744         roce_info->ord_size = dev->hw_attrs.max_hw_ord;
745
746         if (!iwqp->user_mode) {
747                 roce_info->priv_mode_en = true;
748                 roce_info->fast_reg_en = true;
749                 roce_info->udprivcq_en = true;
750         }
751         roce_info->roce_tver = 0;
752
753         ctx_info->roce_info = &iwqp->roce_info;
754         ctx_info->udp_info = &iwqp->udp_info;
755         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
756 }
757
758 void
759 irdma_iw_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
760                                  struct irdma_qp_host_ctx_info *ctx_info)
761 {
762         struct irdma_device *iwdev = iwqp->iwdev;
763         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
764         struct irdma_iwarp_offload_info *iwarp_info;
765
766         iwarp_info = &iwqp->iwarp_info;
767         ether_addr_copy(iwarp_info->mac_addr, IF_LLADDR(iwdev->netdev));
768         iwarp_info->rd_en = true;
769         iwarp_info->wr_rdresp_en = true;
770         iwarp_info->bind_en = true;
771         iwarp_info->ecn_en = true;
772         iwarp_info->rtomin = 5;
773
774         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
775                 iwarp_info->ib_rd_en = true;
776         if (!iwqp->user_mode) {
777                 iwarp_info->priv_mode_en = true;
778                 iwarp_info->fast_reg_en = true;
779         }
780         iwarp_info->ddp_ver = 1;
781         iwarp_info->rdmap_ver = 1;
782
783         ctx_info->iwarp_info = &iwqp->iwarp_info;
784         ctx_info->iwarp_info_valid = true;
785         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
786         ctx_info->iwarp_info_valid = false;
787 }
788
789 int
790 irdma_validate_qp_attrs(struct ib_qp_init_attr *init_attr,
791                         struct irdma_device *iwdev)
792 {
793         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
794         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
795
796         if (init_attr->create_flags)
797                 return -EOPNOTSUPP;
798
799         if (init_attr->cap.max_inline_data > uk_attrs->max_hw_inline ||
800             init_attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags ||
801             init_attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags)
802                 return -EINVAL;
803
804         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
805                 if (init_attr->qp_type != IB_QPT_RC &&
806                     init_attr->qp_type != IB_QPT_UD &&
807                     init_attr->qp_type != IB_QPT_GSI)
808                         return -EOPNOTSUPP;
809         } else {
810                 if (init_attr->qp_type != IB_QPT_RC)
811                         return -EOPNOTSUPP;
812         }
813
814         return 0;
815 }
816
817 void
818 irdma_sched_qp_flush_work(struct irdma_qp *iwqp)
819 {
820         irdma_qp_add_ref(&iwqp->ibqp);
821         if (mod_delayed_work(iwqp->iwdev->cleanup_wq, &iwqp->dwork_flush,
822                              msecs_to_jiffies(IRDMA_FLUSH_DELAY_MS)))
823                 irdma_qp_rem_ref(&iwqp->ibqp);
824 }
825
826 void
827 irdma_flush_worker(struct work_struct *work)
828 {
829         struct delayed_work *dwork = to_delayed_work(work);
830         struct irdma_qp *iwqp = container_of(dwork, struct irdma_qp, dwork_flush);
831
832         irdma_generate_flush_completions(iwqp);
833         /* For the add in irdma_sched_qp_flush_work */
834         irdma_qp_rem_ref(&iwqp->ibqp);
835 }
836
837 static int
838 irdma_get_ib_acc_flags(struct irdma_qp *iwqp)
839 {
840         int acc_flags = 0;
841
842         if (rdma_protocol_roce(iwqp->ibqp.device, 1)) {
843                 if (iwqp->roce_info.wr_rdresp_en) {
844                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
845                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
846                 }
847                 if (iwqp->roce_info.rd_en)
848                         acc_flags |= IB_ACCESS_REMOTE_READ;
849                 if (iwqp->roce_info.bind_en)
850                         acc_flags |= IB_ACCESS_MW_BIND;
851         } else {
852                 if (iwqp->iwarp_info.wr_rdresp_en) {
853                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
854                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
855                 }
856                 if (iwqp->iwarp_info.rd_en)
857                         acc_flags |= IB_ACCESS_REMOTE_READ;
858                 if (iwqp->iwarp_info.bind_en)
859                         acc_flags |= IB_ACCESS_MW_BIND;
860         }
861         return acc_flags;
862 }
863
864 /**
865  * irdma_query_qp - query qp attributes
866  * @ibqp: qp pointer
867  * @attr: attributes pointer
868  * @attr_mask: Not used
869  * @init_attr: qp attributes to return
870  */
871 static int
872 irdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
873                int attr_mask, struct ib_qp_init_attr *init_attr)
874 {
875         struct irdma_qp *iwqp = to_iwqp(ibqp);
876         struct irdma_sc_qp *qp = &iwqp->sc_qp;
877
878         memset(attr, 0, sizeof(*attr));
879         memset(init_attr, 0, sizeof(*init_attr));
880
881         attr->qp_state = iwqp->ibqp_state;
882         attr->cur_qp_state = iwqp->ibqp_state;
883         attr->cap.max_send_wr = iwqp->max_send_wr;
884         attr->cap.max_recv_wr = iwqp->max_recv_wr;
885         attr->cap.max_inline_data = qp->qp_uk.max_inline_data;
886         attr->cap.max_send_sge = qp->qp_uk.max_sq_frag_cnt;
887         attr->cap.max_recv_sge = qp->qp_uk.max_rq_frag_cnt;
888         attr->qp_access_flags = irdma_get_ib_acc_flags(iwqp);
889         attr->port_num = 1;
890         if (rdma_protocol_roce(ibqp->device, 1)) {
891                 attr->path_mtu = ib_mtu_int_to_enum(iwqp->udp_info.snd_mss);
892                 attr->qkey = iwqp->roce_info.qkey;
893                 attr->rq_psn = iwqp->udp_info.epsn;
894                 attr->sq_psn = iwqp->udp_info.psn_nxt;
895                 attr->dest_qp_num = iwqp->roce_info.dest_qp;
896                 attr->pkey_index = iwqp->roce_info.p_key;
897                 attr->retry_cnt = iwqp->udp_info.rexmit_thresh;
898                 attr->rnr_retry = iwqp->udp_info.rnr_nak_thresh;
899                 attr->max_rd_atomic = iwqp->roce_info.ord_size;
900                 attr->max_dest_rd_atomic = iwqp->roce_info.ird_size;
901         }
902
903         init_attr->event_handler = iwqp->ibqp.event_handler;
904         init_attr->qp_context = iwqp->ibqp.qp_context;
905         init_attr->send_cq = iwqp->ibqp.send_cq;
906         init_attr->recv_cq = iwqp->ibqp.recv_cq;
907         init_attr->cap = attr->cap;
908
909         return 0;
910 }
911
912 /**
913  * irdma_modify_qp_roce - modify qp request
914  * @ibqp: qp's pointer for modify
915  * @attr: access attributes
916  * @attr_mask: state mask
917  * @udata: user data
918  */
919 int
920 irdma_modify_qp_roce(struct ib_qp *ibqp, struct ib_qp_attr *attr,
921                      int attr_mask, struct ib_udata *udata)
922 {
923 #define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
924 #define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
925         struct irdma_pd *iwpd = to_iwpd(ibqp->pd);
926         struct irdma_qp *iwqp = to_iwqp(ibqp);
927         struct irdma_device *iwdev = iwqp->iwdev;
928         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
929         struct irdma_qp_host_ctx_info *ctx_info;
930         struct irdma_roce_offload_info *roce_info;
931         struct irdma_udp_offload_info *udp_info;
932         struct irdma_modify_qp_info info = {0};
933         struct irdma_modify_qp_resp uresp = {};
934         struct irdma_modify_qp_req ureq;
935         unsigned long flags;
936         u8 issue_modify_qp = 0;
937         int ret = 0;
938
939         ctx_info = &iwqp->ctx_info;
940         roce_info = &iwqp->roce_info;
941         udp_info = &iwqp->udp_info;
942
943         if (udata) {
944                 if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
945                     (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
946                         return -EINVAL;
947         }
948
949         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
950                 return -EOPNOTSUPP;
951
952         if (attr_mask & IB_QP_DEST_QPN)
953                 roce_info->dest_qp = attr->dest_qp_num;
954
955         if (attr_mask & IB_QP_PKEY_INDEX) {
956                 ret = irdma_query_pkey(ibqp->device, 0, attr->pkey_index,
957                                        &roce_info->p_key);
958                 if (ret)
959                         return ret;
960         }
961
962         if (attr_mask & IB_QP_QKEY)
963                 roce_info->qkey = attr->qkey;
964
965         if (attr_mask & IB_QP_PATH_MTU)
966                 udp_info->snd_mss = ib_mtu_enum_to_int(attr->path_mtu);
967
968         if (attr_mask & IB_QP_SQ_PSN) {
969                 udp_info->psn_nxt = attr->sq_psn;
970                 udp_info->lsn = 0xffff;
971                 udp_info->psn_una = attr->sq_psn;
972                 udp_info->psn_max = attr->sq_psn;
973         }
974
975         if (attr_mask & IB_QP_RQ_PSN)
976                 udp_info->epsn = attr->rq_psn;
977
978         if (attr_mask & IB_QP_RNR_RETRY)
979                 udp_info->rnr_nak_thresh = attr->rnr_retry;
980
981         if (attr_mask & IB_QP_RETRY_CNT)
982                 udp_info->rexmit_thresh = attr->retry_cnt;
983
984         ctx_info->roce_info->pd_id = iwpd->sc_pd.pd_id;
985
986         if (attr_mask & IB_QP_AV) {
987                 struct irdma_av *av = &iwqp->roce_ah.av;
988                 u16 vlan_id = VLAN_N_VID;
989                 u32 local_ip[4] = {};
990
991                 memset(&iwqp->roce_ah, 0, sizeof(iwqp->roce_ah));
992                 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
993                         udp_info->ttl = attr->ah_attr.grh.hop_limit;
994                         udp_info->flow_label = attr->ah_attr.grh.flow_label;
995                         udp_info->tos = attr->ah_attr.grh.traffic_class;
996
997                         udp_info->src_port = kc_rdma_get_udp_sport(udp_info->flow_label,
998                                                                    ibqp->qp_num,
999                                                                    roce_info->dest_qp);
1000
1001                         irdma_qp_rem_qos(&iwqp->sc_qp);
1002                         dev->ws_remove(iwqp->sc_qp.vsi, ctx_info->user_pri);
1003                         if (iwqp->sc_qp.vsi->dscp_mode)
1004                                 ctx_info->user_pri =
1005                                     iwqp->sc_qp.vsi->dscp_map[irdma_tos2dscp(udp_info->tos)];
1006                         else
1007                                 ctx_info->user_pri = rt_tos2priority(udp_info->tos);
1008                 }
1009                 ret = kc_irdma_set_roce_cm_info(iwqp, attr, &vlan_id);
1010                 if (ret)
1011                         return ret;
1012                 if (dev->ws_add(iwqp->sc_qp.vsi, ctx_info->user_pri))
1013                         return -ENOMEM;
1014                 iwqp->sc_qp.user_pri = ctx_info->user_pri;
1015                 irdma_qp_add_qos(&iwqp->sc_qp);
1016
1017                 if (vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
1018                         vlan_id = 0;
1019                 if (vlan_id < VLAN_N_VID) {
1020                         udp_info->insert_vlan_tag = true;
1021                         udp_info->vlan_tag = vlan_id |
1022                             ctx_info->user_pri << VLAN_PRIO_SHIFT;
1023                 } else {
1024                         udp_info->insert_vlan_tag = false;
1025                 }
1026
1027                 av->attrs = attr->ah_attr;
1028                 rdma_gid2ip((struct sockaddr *)&av->dgid_addr, &attr->ah_attr.grh.dgid);
1029                 if (av->sgid_addr.saddr.sa_family == AF_INET6) {
1030                         __be32 *daddr =
1031                         av->dgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32;
1032                         __be32 *saddr =
1033                         av->sgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32;
1034
1035                         irdma_copy_ip_ntohl(&udp_info->dest_ip_addr[0], daddr);
1036                         irdma_copy_ip_ntohl(&udp_info->local_ipaddr[0], saddr);
1037
1038                         udp_info->ipv4 = false;
1039                         irdma_copy_ip_ntohl(local_ip, daddr);
1040
1041                         udp_info->arp_idx = irdma_arp_table(iwdev->rf, local_ip,
1042                                                             NULL, IRDMA_ARP_RESOLVE);
1043                 } else {
1044                         __be32 saddr = av->sgid_addr.saddr_in.sin_addr.s_addr;
1045                         __be32 daddr = av->dgid_addr.saddr_in.sin_addr.s_addr;
1046
1047                         local_ip[0] = ntohl(daddr);
1048
1049                         udp_info->ipv4 = true;
1050                         udp_info->dest_ip_addr[0] = 0;
1051                         udp_info->dest_ip_addr[1] = 0;
1052                         udp_info->dest_ip_addr[2] = 0;
1053                         udp_info->dest_ip_addr[3] = local_ip[0];
1054
1055                         udp_info->local_ipaddr[0] = 0;
1056                         udp_info->local_ipaddr[1] = 0;
1057                         udp_info->local_ipaddr[2] = 0;
1058                         udp_info->local_ipaddr[3] = ntohl(saddr);
1059                 }
1060                 udp_info->arp_idx =
1061                     irdma_add_arp(iwdev->rf, local_ip,
1062                                   ah_attr_to_dmac(attr->ah_attr));
1063         }
1064
1065         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1066                 if (attr->max_rd_atomic > dev->hw_attrs.max_hw_ord) {
1067                         ibdev_err(&iwdev->ibdev,
1068                                   "rd_atomic = %d, above max_hw_ord=%d\n",
1069                                   attr->max_rd_atomic,
1070                                   dev->hw_attrs.max_hw_ord);
1071                         return -EINVAL;
1072                 }
1073                 if (attr->max_rd_atomic)
1074                         roce_info->ord_size = attr->max_rd_atomic;
1075                 info.ord_valid = true;
1076         }
1077
1078         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1079                 if (attr->max_dest_rd_atomic > dev->hw_attrs.max_hw_ird) {
1080                         ibdev_err(&iwdev->ibdev,
1081                                   "rd_atomic = %d, above max_hw_ird=%d\n",
1082                                   attr->max_rd_atomic,
1083                                   dev->hw_attrs.max_hw_ird);
1084                         return -EINVAL;
1085                 }
1086                 if (attr->max_dest_rd_atomic)
1087                         roce_info->ird_size = attr->max_dest_rd_atomic;
1088         }
1089
1090         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1091                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1092                         roce_info->wr_rdresp_en = true;
1093                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1094                         roce_info->wr_rdresp_en = true;
1095                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1096                         roce_info->rd_en = true;
1097         }
1098
1099         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1100
1101         irdma_debug(dev, IRDMA_DEBUG_VERBS,
1102                     "caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d attr_mask=0x%x\n",
1103                     __builtin_return_address(0), ibqp->qp_num, attr->qp_state, iwqp->ibqp_state,
1104                     iwqp->iwarp_state, attr_mask);
1105
1106         spin_lock_irqsave(&iwqp->lock, flags);
1107         if (attr_mask & IB_QP_STATE) {
1108                 if (!kc_ib_modify_qp_is_ok(iwqp->ibqp_state, attr->qp_state,
1109                                            iwqp->ibqp.qp_type, attr_mask,
1110                                            IB_LINK_LAYER_ETHERNET)) {
1111                         irdma_print("modify_qp invalid for qp_id=%d, old_state=0x%x, new_state=0x%x\n",
1112                                     iwqp->ibqp.qp_num, iwqp->ibqp_state,
1113                                     attr->qp_state);
1114                         ret = -EINVAL;
1115                         goto exit;
1116                 }
1117                 info.curr_iwarp_state = iwqp->iwarp_state;
1118
1119                 switch (attr->qp_state) {
1120                 case IB_QPS_INIT:
1121                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1122                                 ret = -EINVAL;
1123                                 goto exit;
1124                         }
1125
1126                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1127                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1128                                 issue_modify_qp = 1;
1129                         }
1130                         break;
1131                 case IB_QPS_RTR:
1132                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1133                                 ret = -EINVAL;
1134                                 goto exit;
1135                         }
1136                         info.arp_cache_idx_valid = true;
1137                         info.cq_num_valid = true;
1138                         info.next_iwarp_state = IRDMA_QP_STATE_RTR;
1139                         issue_modify_qp = 1;
1140                         break;
1141                 case IB_QPS_RTS:
1142                         if (iwqp->ibqp_state < IB_QPS_RTR ||
1143                             iwqp->ibqp_state == IB_QPS_ERR) {
1144                                 ret = -EINVAL;
1145                                 goto exit;
1146                         }
1147
1148                         info.arp_cache_idx_valid = true;
1149                         info.cq_num_valid = true;
1150                         info.ord_valid = true;
1151                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1152                         issue_modify_qp = 1;
1153                         if (dev->hw_attrs.uk_attrs.hw_rev == IRDMA_GEN_2)
1154                                 iwdev->rf->check_fc(&iwdev->vsi, &iwqp->sc_qp);
1155                         udp_info->cwnd = iwdev->roce_cwnd;
1156                         roce_info->ack_credits = iwdev->roce_ackcreds;
1157                         if (iwdev->push_mode && udata &&
1158                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1159                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1160                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1161                                 irdma_alloc_push_page(iwqp);
1162                                 spin_lock_irqsave(&iwqp->lock, flags);
1163                         }
1164                         break;
1165                 case IB_QPS_SQD:
1166                         if (iwqp->iwarp_state == IRDMA_QP_STATE_SQD)
1167                                 goto exit;
1168
1169                         if (iwqp->iwarp_state != IRDMA_QP_STATE_RTS) {
1170                                 ret = -EINVAL;
1171                                 goto exit;
1172                         }
1173
1174                         info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1175                         issue_modify_qp = 1;
1176                         break;
1177                 case IB_QPS_SQE:
1178                 case IB_QPS_ERR:
1179                 case IB_QPS_RESET:
1180                         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS) {
1181                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1182                                 info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1183                                 irdma_hw_modify_qp(iwdev, iwqp, &info, true);
1184                                 spin_lock_irqsave(&iwqp->lock, flags);
1185                         }
1186
1187                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1188                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1189                                 if (udata && udata->inlen) {
1190                                         if (ib_copy_from_udata(&ureq, udata,
1191                                                                min(sizeof(ureq), udata->inlen)))
1192                                                 return -EINVAL;
1193
1194                                         irdma_flush_wqes(iwqp,
1195                                                          (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1196                                                          (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1197                                                          IRDMA_REFLUSH);
1198                                 }
1199                                 return 0;
1200                         }
1201
1202                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1203                         issue_modify_qp = 1;
1204                         break;
1205                 default:
1206                         ret = -EINVAL;
1207                         goto exit;
1208                 }
1209
1210                 iwqp->ibqp_state = attr->qp_state;
1211         }
1212
1213         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1214         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1215         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1216         spin_unlock_irqrestore(&iwqp->lock, flags);
1217
1218         if (attr_mask & IB_QP_STATE) {
1219                 if (issue_modify_qp) {
1220                         ctx_info->rem_endpoint_idx = udp_info->arp_idx;
1221                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1222                                 return -EINVAL;
1223                         spin_lock_irqsave(&iwqp->lock, flags);
1224                         if (iwqp->iwarp_state == info.curr_iwarp_state) {
1225                                 iwqp->iwarp_state = info.next_iwarp_state;
1226                                 iwqp->ibqp_state = attr->qp_state;
1227                         }
1228                         if (iwqp->ibqp_state > IB_QPS_RTS &&
1229                             !iwqp->flush_issued) {
1230                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1231                                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ |
1232                                                  IRDMA_FLUSH_RQ |
1233                                                  IRDMA_FLUSH_WAIT);
1234                                 iwqp->flush_issued = 1;
1235
1236                         } else {
1237                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1238                         }
1239                 } else {
1240                         iwqp->ibqp_state = attr->qp_state;
1241                 }
1242                 if (udata && udata->outlen && dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1243                         struct irdma_ucontext *ucontext;
1244
1245 #if __FreeBSD_version >= 1400026
1246                         ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
1247 #else
1248                         ucontext = to_ucontext(ibqp->uobject->context);
1249 #endif
1250                         if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1251                             !iwqp->push_wqe_mmap_entry &&
1252                             !irdma_setup_push_mmap_entries(ucontext, iwqp,
1253                                                            &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1254                                 uresp.push_valid = 1;
1255                                 uresp.push_offset = iwqp->sc_qp.push_offset;
1256                         }
1257                         uresp.rd_fence_rate = iwdev->rd_fence_rate;
1258                         ret = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1259                                                                   udata->outlen));
1260                         if (ret) {
1261                                 irdma_remove_push_mmap_entries(iwqp);
1262                                 irdma_debug(iwdev_to_idev(iwdev),
1263                                             IRDMA_DEBUG_VERBS,
1264                                             "copy_to_udata failed\n");
1265                                 return ret;
1266                         }
1267                 }
1268         }
1269
1270         return 0;
1271 exit:
1272         spin_unlock_irqrestore(&iwqp->lock, flags);
1273
1274         return ret;
1275 }
1276
1277 /**
1278  * irdma_modify_qp - modify qp request
1279  * @ibqp: qp's pointer for modify
1280  * @attr: access attributes
1281  * @attr_mask: state mask
1282  * @udata: user data
1283  */
1284 int
1285 irdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
1286                 struct ib_udata *udata)
1287 {
1288 #define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
1289 #define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
1290         struct irdma_qp *iwqp = to_iwqp(ibqp);
1291         struct irdma_device *iwdev = iwqp->iwdev;
1292         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1293         struct irdma_qp_host_ctx_info *ctx_info;
1294         struct irdma_tcp_offload_info *tcp_info;
1295         struct irdma_iwarp_offload_info *offload_info;
1296         struct irdma_modify_qp_info info = {0};
1297         struct irdma_modify_qp_resp uresp = {};
1298         struct irdma_modify_qp_req ureq = {};
1299         u8 issue_modify_qp = 0;
1300         u8 dont_wait = 0;
1301         int err;
1302         unsigned long flags;
1303
1304         if (udata) {
1305                 if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
1306                     (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
1307                         return -EINVAL;
1308         }
1309
1310         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1311                 return -EOPNOTSUPP;
1312
1313         ctx_info = &iwqp->ctx_info;
1314         offload_info = &iwqp->iwarp_info;
1315         tcp_info = &iwqp->tcp_info;
1316         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1317         irdma_debug(dev, IRDMA_DEBUG_VERBS,
1318                     "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",
1319                     __builtin_return_address(0), ibqp->qp_num, attr->qp_state, iwqp->ibqp_state, iwqp->iwarp_state,
1320                     iwqp->last_aeq, iwqp->hw_tcp_state, iwqp->hw_iwarp_state, attr_mask);
1321
1322         spin_lock_irqsave(&iwqp->lock, flags);
1323         if (attr_mask & IB_QP_STATE) {
1324                 info.curr_iwarp_state = iwqp->iwarp_state;
1325                 switch (attr->qp_state) {
1326                 case IB_QPS_INIT:
1327                 case IB_QPS_RTR:
1328                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1329                                 err = -EINVAL;
1330                                 goto exit;
1331                         }
1332
1333                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1334                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1335                                 issue_modify_qp = 1;
1336                         }
1337                         if (iwdev->push_mode && udata &&
1338                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1339                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1340                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1341                                 irdma_alloc_push_page(iwqp);
1342                                 spin_lock_irqsave(&iwqp->lock, flags);
1343                         }
1344                         break;
1345                 case IB_QPS_RTS:
1346                         if (iwqp->iwarp_state > IRDMA_QP_STATE_RTS ||
1347                             !iwqp->cm_id) {
1348                                 err = -EINVAL;
1349                                 goto exit;
1350                         }
1351
1352                         issue_modify_qp = 1;
1353                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
1354                         iwqp->hte_added = 1;
1355                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1356                         info.tcp_ctx_valid = true;
1357                         info.ord_valid = true;
1358                         info.arp_cache_idx_valid = true;
1359                         info.cq_num_valid = true;
1360                         break;
1361                 case IB_QPS_SQD:
1362                         if (iwqp->hw_iwarp_state > IRDMA_QP_STATE_RTS) {
1363                                 err = 0;
1364                                 goto exit;
1365                         }
1366
1367                         if (iwqp->iwarp_state == IRDMA_QP_STATE_CLOSING ||
1368                             iwqp->iwarp_state < IRDMA_QP_STATE_RTS) {
1369                                 err = 0;
1370                                 goto exit;
1371                         }
1372
1373                         if (iwqp->iwarp_state > IRDMA_QP_STATE_CLOSING) {
1374                                 err = -EINVAL;
1375                                 goto exit;
1376                         }
1377
1378                         info.next_iwarp_state = IRDMA_QP_STATE_CLOSING;
1379                         issue_modify_qp = 1;
1380                         break;
1381                 case IB_QPS_SQE:
1382                         if (iwqp->iwarp_state >= IRDMA_QP_STATE_TERMINATE) {
1383                                 err = -EINVAL;
1384                                 goto exit;
1385                         }
1386
1387                         info.next_iwarp_state = IRDMA_QP_STATE_TERMINATE;
1388                         issue_modify_qp = 1;
1389                         break;
1390                 case IB_QPS_ERR:
1391                 case IB_QPS_RESET:
1392                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1393                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1394                                 if (udata && udata->inlen) {
1395                                         if (ib_copy_from_udata(&ureq, udata,
1396                                                                min(sizeof(ureq), udata->inlen)))
1397                                                 return -EINVAL;
1398
1399                                         irdma_flush_wqes(iwqp,
1400                                                          (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1401                                                          (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1402                                                          IRDMA_REFLUSH);
1403                                 }
1404                                 return 0;
1405                         }
1406
1407                         if (iwqp->sc_qp.term_flags) {
1408                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1409                                 irdma_terminate_del_timer(&iwqp->sc_qp);
1410                                 spin_lock_irqsave(&iwqp->lock, flags);
1411                         }
1412                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1413                         if (iwqp->hw_tcp_state > IRDMA_TCP_STATE_CLOSED &&
1414                             iwdev->iw_status &&
1415                             iwqp->hw_tcp_state != IRDMA_TCP_STATE_TIME_WAIT)
1416                                 info.reset_tcp_conn = true;
1417                         else
1418                                 dont_wait = 1;
1419
1420                         issue_modify_qp = 1;
1421                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1422                         break;
1423                 default:
1424                         err = -EINVAL;
1425                         goto exit;
1426                 }
1427
1428                 iwqp->ibqp_state = attr->qp_state;
1429         }
1430         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1431                 ctx_info->iwarp_info_valid = true;
1432                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1433                         offload_info->wr_rdresp_en = true;
1434                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1435                         offload_info->wr_rdresp_en = true;
1436                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1437                         offload_info->rd_en = true;
1438         }
1439
1440         if (ctx_info->iwarp_info_valid) {
1441                 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1442                 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1443                 irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1444         }
1445         spin_unlock_irqrestore(&iwqp->lock, flags);
1446
1447         if (attr_mask & IB_QP_STATE) {
1448                 if (issue_modify_qp) {
1449                         ctx_info->rem_endpoint_idx = tcp_info->arp_idx;
1450                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1451                                 return -EINVAL;
1452                 }
1453
1454                 spin_lock_irqsave(&iwqp->lock, flags);
1455                 if (iwqp->iwarp_state == info.curr_iwarp_state) {
1456                         iwqp->iwarp_state = info.next_iwarp_state;
1457                         iwqp->ibqp_state = attr->qp_state;
1458                 }
1459                 spin_unlock_irqrestore(&iwqp->lock, flags);
1460         }
1461
1462         if (issue_modify_qp && iwqp->ibqp_state > IB_QPS_RTS) {
1463                 if (dont_wait) {
1464                         if (iwqp->hw_tcp_state) {
1465                                 spin_lock_irqsave(&iwqp->lock, flags);
1466                                 iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1467                                 iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1468                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1469                         }
1470                         irdma_cm_disconn(iwqp);
1471                 } else {
1472                         int close_timer_started;
1473
1474                         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
1475
1476                         if (iwqp->cm_node) {
1477                                 atomic_inc(&iwqp->cm_node->refcnt);
1478                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1479                                 close_timer_started = atomic_inc_return(&iwqp->close_timer_started);
1480                                 if (iwqp->cm_id && close_timer_started == 1)
1481                                         irdma_schedule_cm_timer(iwqp->cm_node,
1482                                                                 (struct irdma_puda_buf *)iwqp,
1483                                                                 IRDMA_TIMER_TYPE_CLOSE, 1, 0);
1484
1485                                 irdma_rem_ref_cm_node(iwqp->cm_node);
1486                         } else {
1487                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1488                         }
1489                 }
1490         }
1491         if (attr_mask & IB_QP_STATE && udata && udata->outlen &&
1492             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1493                 struct irdma_ucontext *ucontext;
1494
1495 #if __FreeBSD_version >= 1400026
1496                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
1497 #else
1498                 ucontext = to_ucontext(ibqp->uobject->context);
1499 #endif
1500                 if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1501                     !iwqp->push_wqe_mmap_entry &&
1502                     !irdma_setup_push_mmap_entries(ucontext, iwqp,
1503                                                    &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1504                         uresp.push_valid = 1;
1505                         uresp.push_offset = iwqp->sc_qp.push_offset;
1506                 }
1507                 uresp.rd_fence_rate = iwdev->rd_fence_rate;
1508
1509                 err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1510                                                           udata->outlen));
1511                 if (err) {
1512                         irdma_remove_push_mmap_entries(iwqp);
1513                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
1514                                     "copy_to_udata failed\n");
1515                         return err;
1516                 }
1517         }
1518
1519         return 0;
1520 exit:
1521         spin_unlock_irqrestore(&iwqp->lock, flags);
1522
1523         return err;
1524 }
1525
1526 /**
1527  * irdma_cq_free_rsrc - free up resources for cq
1528  * @rf: RDMA PCI function
1529  * @iwcq: cq ptr
1530  */
1531 void
1532 irdma_cq_free_rsrc(struct irdma_pci_f *rf, struct irdma_cq *iwcq)
1533 {
1534         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1535
1536         if (!iwcq->user_mode) {
1537                 irdma_free_dma_mem(rf->sc_dev.hw, &iwcq->kmem);
1538                 irdma_free_dma_mem(rf->sc_dev.hw, &iwcq->kmem_shadow);
1539         }
1540
1541         irdma_free_rsrc(rf, rf->allocated_cqs, cq->cq_uk.cq_id);
1542 }
1543
1544 /**
1545  * irdma_free_cqbuf - worker to free a cq buffer
1546  * @work: provides access to the cq buffer to free
1547  */
1548 static void
1549 irdma_free_cqbuf(struct work_struct *work)
1550 {
1551         struct irdma_cq_buf *cq_buf = container_of(work, struct irdma_cq_buf, work);
1552
1553         irdma_free_dma_mem(cq_buf->hw, &cq_buf->kmem_buf);
1554         kfree(cq_buf);
1555 }
1556
1557 /**
1558  * irdma_process_resize_list - remove resized cq buffers from the resize_list
1559  * @iwcq: cq which owns the resize_list
1560  * @iwdev: irdma device
1561  * @lcqe_buf: the buffer where the last cqe is received
1562  */
1563 int
1564 irdma_process_resize_list(struct irdma_cq *iwcq,
1565                           struct irdma_device *iwdev,
1566                           struct irdma_cq_buf *lcqe_buf)
1567 {
1568         struct list_head *tmp_node, *list_node;
1569         struct irdma_cq_buf *cq_buf;
1570         int cnt = 0;
1571
1572         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
1573                 cq_buf = list_entry(list_node, struct irdma_cq_buf, list);
1574                 if (cq_buf == lcqe_buf)
1575                         return cnt;
1576
1577                 list_del(&cq_buf->list);
1578                 queue_work(iwdev->cleanup_wq, &cq_buf->work);
1579                 cnt++;
1580         }
1581
1582         return cnt;
1583 }
1584
1585 /**
1586  * irdma_resize_cq - resize cq
1587  * @ibcq: cq to be resized
1588  * @entries: desired cq size
1589  * @udata: user data
1590  */
1591 static int
1592 irdma_resize_cq(struct ib_cq *ibcq, int entries,
1593                 struct ib_udata *udata)
1594 {
1595 #define IRDMA_RESIZE_CQ_MIN_REQ_LEN offsetofend(struct irdma_resize_cq_req, user_cq_buffer)
1596         struct irdma_cq *iwcq = to_iwcq(ibcq);
1597         struct irdma_sc_dev *dev = iwcq->sc_cq.dev;
1598         struct irdma_cqp_request *cqp_request;
1599         struct cqp_cmds_info *cqp_info;
1600         struct irdma_modify_cq_info *m_info;
1601         struct irdma_modify_cq_info info = {0};
1602         struct irdma_dma_mem kmem_buf;
1603         struct irdma_cq_mr *cqmr_buf;
1604         struct irdma_pbl *iwpbl_buf;
1605         struct irdma_device *iwdev;
1606         struct irdma_pci_f *rf;
1607         struct irdma_cq_buf *cq_buf = NULL;
1608         unsigned long flags;
1609         int ret;
1610
1611         iwdev = to_iwdev(ibcq->device);
1612         rf = iwdev->rf;
1613
1614         if (!(rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
1615               IRDMA_FEATURE_CQ_RESIZE))
1616                 return -EOPNOTSUPP;
1617
1618         if (udata && udata->inlen < IRDMA_RESIZE_CQ_MIN_REQ_LEN)
1619                 return -EINVAL;
1620
1621         if (entries > rf->max_cqe)
1622                 return -EINVAL;
1623
1624         if (!iwcq->user_mode) {
1625                 entries++;
1626                 if (rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
1627                         entries *= 2;
1628         }
1629
1630         info.cq_size = max(entries, 4);
1631
1632         if (info.cq_size == iwcq->sc_cq.cq_uk.cq_size - 1)
1633                 return 0;
1634
1635         if (udata) {
1636                 struct irdma_resize_cq_req req = {};
1637                 struct irdma_ucontext *ucontext =
1638 #if __FreeBSD_version >= 1400026
1639                 rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
1640 #else
1641                 to_ucontext(ibcq->uobject->context);
1642 #endif
1643
1644                 /* CQ resize not supported with legacy GEN_1 libi40iw */
1645                 if (ucontext->legacy_mode)
1646                         return -EOPNOTSUPP;
1647
1648                 if (ib_copy_from_udata(&req, udata,
1649                                        min(sizeof(req), udata->inlen)))
1650                         return -EINVAL;
1651
1652                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1653                 iwpbl_buf = irdma_get_pbl((unsigned long)req.user_cq_buffer,
1654                                           &ucontext->cq_reg_mem_list);
1655                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1656
1657                 if (!iwpbl_buf)
1658                         return -ENOMEM;
1659
1660                 cqmr_buf = &iwpbl_buf->cq_mr;
1661                 if (iwpbl_buf->pbl_allocated) {
1662                         info.virtual_map = true;
1663                         info.pbl_chunk_size = 1;
1664                         info.first_pm_pbl_idx = cqmr_buf->cq_pbl.idx;
1665                 } else {
1666                         info.cq_pa = cqmr_buf->cq_pbl.addr;
1667                 }
1668         } else {
1669                 /* Kmode CQ resize */
1670                 int rsize;
1671
1672                 rsize = info.cq_size * sizeof(struct irdma_cqe);
1673                 kmem_buf.size = round_up(rsize, 256);
1674                 kmem_buf.va = irdma_allocate_dma_mem(dev->hw, &kmem_buf,
1675                                                      kmem_buf.size, 256);
1676                 if (!kmem_buf.va)
1677                         return -ENOMEM;
1678
1679                 info.cq_base = kmem_buf.va;
1680                 info.cq_pa = kmem_buf.pa;
1681                 cq_buf = kzalloc(sizeof(*cq_buf), GFP_KERNEL);
1682                 if (!cq_buf) {
1683                         ret = -ENOMEM;
1684                         goto error;
1685                 }
1686         }
1687
1688         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
1689         if (!cqp_request) {
1690                 ret = -ENOMEM;
1691                 goto error;
1692         }
1693
1694         info.shadow_read_threshold = iwcq->sc_cq.shadow_read_threshold;
1695         info.cq_resize = true;
1696
1697         cqp_info = &cqp_request->info;
1698         m_info = &cqp_info->in.u.cq_modify.info;
1699         memcpy(m_info, &info, sizeof(*m_info));
1700
1701         cqp_info->cqp_cmd = IRDMA_OP_CQ_MODIFY;
1702         cqp_info->in.u.cq_modify.cq = &iwcq->sc_cq;
1703         cqp_info->in.u.cq_modify.scratch = (uintptr_t)cqp_request;
1704         cqp_info->post_sq = 1;
1705         ret = irdma_handle_cqp_op(rf, cqp_request);
1706         irdma_put_cqp_request(&rf->cqp, cqp_request);
1707         if (ret)
1708                 goto error;
1709
1710         spin_lock_irqsave(&iwcq->lock, flags);
1711         if (cq_buf) {
1712                 cq_buf->kmem_buf = iwcq->kmem;
1713                 cq_buf->hw = dev->hw;
1714                 memcpy(&cq_buf->cq_uk, &iwcq->sc_cq.cq_uk, sizeof(cq_buf->cq_uk));
1715                 INIT_WORK(&cq_buf->work, irdma_free_cqbuf);
1716                 list_add_tail(&cq_buf->list, &iwcq->resize_list);
1717                 iwcq->kmem = kmem_buf;
1718         }
1719
1720         irdma_sc_cq_resize(&iwcq->sc_cq, &info);
1721         ibcq->cqe = info.cq_size - 1;
1722         spin_unlock_irqrestore(&iwcq->lock, flags);
1723
1724         return 0;
1725 error:
1726         if (!udata)
1727                 irdma_free_dma_mem(dev->hw, &kmem_buf);
1728         kfree(cq_buf);
1729
1730         return ret;
1731 }
1732
1733 /**
1734  * irdma_get_mr_access - get hw MR access permissions from IB access flags
1735  * @access: IB access flags
1736  */
1737 static inline u16 irdma_get_mr_access(int access){
1738         u16 hw_access = 0;
1739
1740         hw_access |= (access & IB_ACCESS_LOCAL_WRITE) ?
1741             IRDMA_ACCESS_FLAGS_LOCALWRITE : 0;
1742         hw_access |= (access & IB_ACCESS_REMOTE_WRITE) ?
1743             IRDMA_ACCESS_FLAGS_REMOTEWRITE : 0;
1744         hw_access |= (access & IB_ACCESS_REMOTE_READ) ?
1745             IRDMA_ACCESS_FLAGS_REMOTEREAD : 0;
1746         hw_access |= (access & IB_ACCESS_MW_BIND) ?
1747             IRDMA_ACCESS_FLAGS_BIND_WINDOW : 0;
1748         hw_access |= (access & IB_ZERO_BASED) ?
1749             IRDMA_ACCESS_FLAGS_ZERO_BASED : 0;
1750         hw_access |= IRDMA_ACCESS_FLAGS_LOCALREAD;
1751
1752         return hw_access;
1753 }
1754
1755 /**
1756  * irdma_free_stag - free stag resource
1757  * @iwdev: irdma device
1758  * @stag: stag to free
1759  */
1760 void
1761 irdma_free_stag(struct irdma_device *iwdev, u32 stag)
1762 {
1763         u32 stag_idx;
1764
1765         stag_idx = (stag & iwdev->rf->mr_stagmask) >> IRDMA_CQPSQ_STAG_IDX_S;
1766         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_mrs, stag_idx);
1767 }
1768
1769 /**
1770  * irdma_create_stag - create random stag
1771  * @iwdev: irdma device
1772  */
1773 u32
1774 irdma_create_stag(struct irdma_device *iwdev)
1775 {
1776         u32 stag;
1777         u32 stag_index = 0;
1778         u32 next_stag_index;
1779         u32 driver_key;
1780         u32 random;
1781         u8 consumer_key;
1782         int ret;
1783
1784         get_random_bytes(&random, sizeof(random));
1785         consumer_key = (u8)random;
1786
1787         driver_key = random & ~iwdev->rf->mr_stagmask;
1788         next_stag_index = (random & iwdev->rf->mr_stagmask) >> 8;
1789         next_stag_index %= iwdev->rf->max_mr;
1790
1791         ret = irdma_alloc_rsrc(iwdev->rf, iwdev->rf->allocated_mrs,
1792                                iwdev->rf->max_mr, &stag_index,
1793                                &next_stag_index);
1794         if (ret)
1795                 return 0;
1796         stag = stag_index << IRDMA_CQPSQ_STAG_IDX_S;
1797         stag |= driver_key;
1798         stag += (u32)consumer_key;
1799
1800         return stag;
1801 }
1802
1803 /**
1804  * irdma_check_mem_contiguous - check if pbls stored in arr are contiguous
1805  * @arr: lvl1 pbl array
1806  * @npages: page count
1807  * @pg_size: page size
1808  *
1809  */
1810 static bool
1811 irdma_check_mem_contiguous(u64 *arr, u32 npages, u32 pg_size)
1812 {
1813         u32 pg_idx;
1814
1815         for (pg_idx = 0; pg_idx < npages; pg_idx++) {
1816                 if ((*arr + (pg_size * pg_idx)) != arr[pg_idx])
1817                         return false;
1818         }
1819
1820         return true;
1821 }
1822
1823 /**
1824  * irdma_check_mr_contiguous - check if MR is physically contiguous
1825  * @palloc: pbl allocation struct
1826  * @pg_size: page size
1827  */
1828 static bool
1829 irdma_check_mr_contiguous(struct irdma_pble_alloc *palloc,
1830                           u32 pg_size)
1831 {
1832         struct irdma_pble_level2 *lvl2 = &palloc->level2;
1833         struct irdma_pble_info *leaf = lvl2->leaf;
1834         u64 *arr = NULL;
1835         u64 *start_addr = NULL;
1836         int i;
1837         bool ret;
1838
1839         if (palloc->level == PBLE_LEVEL_1) {
1840                 arr = palloc->level1.addr;
1841                 ret = irdma_check_mem_contiguous(arr, palloc->total_cnt,
1842                                                  pg_size);
1843                 return ret;
1844         }
1845
1846         start_addr = leaf->addr;
1847
1848         for (i = 0; i < lvl2->leaf_cnt; i++, leaf++) {
1849                 arr = leaf->addr;
1850                 if ((*start_addr + (i * pg_size * PBLE_PER_PAGE)) != *arr)
1851                         return false;
1852                 ret = irdma_check_mem_contiguous(arr, leaf->cnt, pg_size);
1853                 if (!ret)
1854                         return false;
1855         }
1856
1857         return true;
1858 }
1859
1860 /**
1861  * irdma_setup_pbles - copy user pg address to pble's
1862  * @rf: RDMA PCI function
1863  * @iwmr: mr pointer for this memory registration
1864  * @use_pbles: flag if to use pble's
1865  * @lvl_1_only: request only level 1 pble if true
1866  */
1867 static int
1868 irdma_setup_pbles(struct irdma_pci_f *rf, struct irdma_mr *iwmr,
1869                   bool use_pbles, bool lvl_1_only)
1870 {
1871         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
1872         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
1873         struct irdma_pble_info *pinfo;
1874         u64 *pbl;
1875         int status;
1876         enum irdma_pble_level level = PBLE_LEVEL_1;
1877
1878         if (use_pbles) {
1879                 status = irdma_get_pble(rf->pble_rsrc, palloc, iwmr->page_cnt,
1880                                         lvl_1_only);
1881                 if (status)
1882                         return status;
1883
1884                 iwpbl->pbl_allocated = true;
1885                 level = palloc->level;
1886                 pinfo = (level == PBLE_LEVEL_1) ? &palloc->level1 :
1887                     palloc->level2.leaf;
1888                 pbl = pinfo->addr;
1889         } else {
1890                 pbl = iwmr->pgaddrmem;
1891         }
1892
1893         irdma_copy_user_pgaddrs(iwmr, pbl, level);
1894
1895         if (use_pbles)
1896                 iwmr->pgaddrmem[0] = *pbl;
1897
1898         return 0;
1899 }
1900
1901 /**
1902  * irdma_handle_q_mem - handle memory for qp and cq
1903  * @iwdev: irdma device
1904  * @req: information for q memory management
1905  * @iwpbl: pble struct
1906  * @use_pbles: flag to use pble
1907  */
1908 static int
1909 irdma_handle_q_mem(struct irdma_device *iwdev,
1910                    struct irdma_mem_reg_req *req,
1911                    struct irdma_pbl *iwpbl, bool use_pbles)
1912 {
1913         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
1914         struct irdma_mr *iwmr = iwpbl->iwmr;
1915         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
1916         struct irdma_cq_mr *cqmr = &iwpbl->cq_mr;
1917         struct irdma_hmc_pble *hmc_p;
1918         u64 *arr = iwmr->pgaddrmem;
1919         u32 pg_size, total;
1920         int err = 0;
1921         bool ret = true;
1922
1923         pg_size = iwmr->page_size;
1924         err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles, true);
1925         if (err)
1926                 return err;
1927
1928         if (use_pbles)
1929                 arr = palloc->level1.addr;
1930
1931         switch (iwmr->type) {
1932         case IRDMA_MEMREG_TYPE_QP:
1933                 total = req->sq_pages + req->rq_pages;
1934                 hmc_p = &qpmr->sq_pbl;
1935                 qpmr->shadow = (dma_addr_t) arr[total];
1936                 if (use_pbles) {
1937                         ret = irdma_check_mem_contiguous(arr, req->sq_pages,
1938                                                          pg_size);
1939                         if (ret)
1940                                 ret = irdma_check_mem_contiguous(&arr[req->sq_pages],
1941                                                                  req->rq_pages,
1942                                                                  pg_size);
1943                 }
1944
1945                 if (!ret) {
1946                         hmc_p->idx = palloc->level1.idx;
1947                         hmc_p = &qpmr->rq_pbl;
1948                         hmc_p->idx = palloc->level1.idx + req->sq_pages;
1949                 } else {
1950                         hmc_p->addr = arr[0];
1951                         hmc_p = &qpmr->rq_pbl;
1952                         hmc_p->addr = arr[req->sq_pages];
1953                 }
1954                 break;
1955         case IRDMA_MEMREG_TYPE_CQ:
1956                 hmc_p = &cqmr->cq_pbl;
1957
1958                 if (!cqmr->split)
1959                         cqmr->shadow = (dma_addr_t) arr[req->cq_pages];
1960
1961                 if (use_pbles)
1962                         ret = irdma_check_mem_contiguous(arr, req->cq_pages,
1963                                                          pg_size);
1964
1965                 if (!ret)
1966                         hmc_p->idx = palloc->level1.idx;
1967                 else
1968                         hmc_p->addr = arr[0];
1969                 break;
1970         default:
1971                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
1972                             "MR type error\n");
1973                 err = -EINVAL;
1974         }
1975
1976         if (use_pbles && ret) {
1977                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
1978                 iwpbl->pbl_allocated = false;
1979         }
1980
1981         return err;
1982 }
1983
1984 /**
1985  * irdma_hw_alloc_mw - create the hw memory window
1986  * @iwdev: irdma device
1987  * @iwmr: pointer to memory window info
1988  */
1989 int
1990 irdma_hw_alloc_mw(struct irdma_device *iwdev, struct irdma_mr *iwmr)
1991 {
1992         struct irdma_mw_alloc_info *info;
1993         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
1994         struct irdma_cqp_request *cqp_request;
1995         struct cqp_cmds_info *cqp_info;
1996         int status;
1997
1998         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
1999         if (!cqp_request)
2000                 return -ENOMEM;
2001
2002         cqp_info = &cqp_request->info;
2003         info = &cqp_info->in.u.mw_alloc.info;
2004         memset(info, 0, sizeof(*info));
2005         if (iwmr->ibmw.type == IB_MW_TYPE_1)
2006                 info->mw_wide = true;
2007
2008         info->page_size = PAGE_SIZE;
2009         info->mw_stag_index = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2010         info->pd_id = iwpd->sc_pd.pd_id;
2011         info->remote_access = true;
2012         cqp_info->cqp_cmd = IRDMA_OP_MW_ALLOC;
2013         cqp_info->post_sq = 1;
2014         cqp_info->in.u.mw_alloc.dev = &iwdev->rf->sc_dev;
2015         cqp_info->in.u.mw_alloc.scratch = (uintptr_t)cqp_request;
2016         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2017         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2018
2019         return status;
2020 }
2021
2022 /**
2023  * irdma_dealloc_mw - Dealloc memory window
2024  * @ibmw: memory window structure.
2025  */
2026 static int
2027 irdma_dealloc_mw(struct ib_mw *ibmw)
2028 {
2029         struct ib_pd *ibpd = ibmw->pd;
2030         struct irdma_pd *iwpd = to_iwpd(ibpd);
2031         struct irdma_mr *iwmr = to_iwmr((struct ib_mr *)ibmw);
2032         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2033         struct irdma_cqp_request *cqp_request;
2034         struct cqp_cmds_info *cqp_info;
2035         struct irdma_dealloc_stag_info *info;
2036
2037         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2038         if (!cqp_request)
2039                 return -ENOMEM;
2040
2041         cqp_info = &cqp_request->info;
2042         info = &cqp_info->in.u.dealloc_stag.info;
2043         memset(info, 0, sizeof(*info));
2044         info->pd_id = iwpd->sc_pd.pd_id;
2045         info->stag_idx = RS_64_1(ibmw->rkey, IRDMA_CQPSQ_STAG_IDX_S);
2046         info->mr = false;
2047         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2048         cqp_info->post_sq = 1;
2049         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2050         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2051         irdma_handle_cqp_op(iwdev->rf, cqp_request);
2052         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2053         irdma_free_stag(iwdev, iwmr->stag);
2054         kfree(iwmr);
2055
2056         return 0;
2057 }
2058
2059 /**
2060  * irdma_hw_alloc_stag - cqp command to allocate stag
2061  * @iwdev: irdma device
2062  * @iwmr: irdma mr pointer
2063  */
2064 int
2065 irdma_hw_alloc_stag(struct irdma_device *iwdev,
2066                     struct irdma_mr *iwmr)
2067 {
2068         struct irdma_allocate_stag_info *info;
2069         struct ib_pd *pd = iwmr->ibmr.pd;
2070         struct irdma_pd *iwpd = to_iwpd(pd);
2071         struct irdma_cqp_request *cqp_request;
2072         struct cqp_cmds_info *cqp_info;
2073         int status;
2074
2075         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2076         if (!cqp_request)
2077                 return -ENOMEM;
2078
2079         cqp_info = &cqp_request->info;
2080         info = &cqp_info->in.u.alloc_stag.info;
2081         memset(info, 0, sizeof(*info));
2082         info->page_size = PAGE_SIZE;
2083         info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2084         info->pd_id = iwpd->sc_pd.pd_id;
2085         info->total_len = iwmr->len;
2086         info->all_memory = (pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) ? true : false;
2087         info->remote_access = true;
2088         cqp_info->cqp_cmd = IRDMA_OP_ALLOC_STAG;
2089         cqp_info->post_sq = 1;
2090         cqp_info->in.u.alloc_stag.dev = &iwdev->rf->sc_dev;
2091         cqp_info->in.u.alloc_stag.scratch = (uintptr_t)cqp_request;
2092         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2093         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2094         if (!status)
2095                 iwmr->is_hwreg = 1;
2096
2097         return status;
2098 }
2099
2100 /**
2101  * irdma_set_page - populate pbl list for fmr
2102  * @ibmr: ib mem to access iwarp mr pointer
2103  * @addr: page dma address fro pbl list
2104  */
2105 static int
2106 irdma_set_page(struct ib_mr *ibmr, u64 addr)
2107 {
2108         struct irdma_mr *iwmr = to_iwmr(ibmr);
2109         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2110         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2111         u64 *pbl;
2112
2113         if (unlikely(iwmr->npages == iwmr->page_cnt))
2114                 return -ENOMEM;
2115
2116         if (palloc->level == PBLE_LEVEL_2) {
2117                 struct irdma_pble_info *palloc_info =
2118                 palloc->level2.leaf + (iwmr->npages >> PBLE_512_SHIFT);
2119
2120                 palloc_info->addr[iwmr->npages & (PBLE_PER_PAGE - 1)] = addr;
2121         } else {
2122                 pbl = palloc->level1.addr;
2123                 pbl[iwmr->npages] = addr;
2124         }
2125
2126         iwmr->npages++;
2127         return 0;
2128 }
2129
2130 /**
2131  * irdma_map_mr_sg - map of sg list for fmr
2132  * @ibmr: ib mem to access iwarp mr pointer
2133  * @sg: scatter gather list
2134  * @sg_nents: number of sg pages
2135  * @sg_offset: scatter gather list for fmr
2136  */
2137 static int
2138 irdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
2139                 int sg_nents, unsigned int *sg_offset)
2140 {
2141         struct irdma_mr *iwmr = to_iwmr(ibmr);
2142
2143         iwmr->npages = 0;
2144
2145         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, irdma_set_page);
2146 }
2147
2148 /**
2149  * irdma_hwreg_mr - send cqp command for memory registration
2150  * @iwdev: irdma device
2151  * @iwmr: irdma mr pointer
2152  * @access: access for MR
2153  */
2154 int
2155 irdma_hwreg_mr(struct irdma_device *iwdev, struct irdma_mr *iwmr,
2156                u16 access)
2157 {
2158         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2159         struct irdma_reg_ns_stag_info *stag_info;
2160         struct ib_pd *pd = iwmr->ibmr.pd;
2161         struct irdma_pd *iwpd = to_iwpd(pd);
2162         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2163         struct irdma_cqp_request *cqp_request;
2164         struct cqp_cmds_info *cqp_info;
2165         int ret;
2166
2167         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2168         if (!cqp_request)
2169                 return -ENOMEM;
2170
2171         cqp_info = &cqp_request->info;
2172         stag_info = &cqp_info->in.u.mr_reg_non_shared.info;
2173         memset(stag_info, 0, sizeof(*stag_info));
2174         stag_info->va = iwpbl->user_base;
2175         stag_info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2176         stag_info->stag_key = (u8)iwmr->stag;
2177         stag_info->total_len = iwmr->len;
2178         stag_info->all_memory = (pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) ? true : false;
2179         stag_info->access_rights = irdma_get_mr_access(access);
2180         stag_info->pd_id = iwpd->sc_pd.pd_id;
2181         if (stag_info->access_rights & IRDMA_ACCESS_FLAGS_ZERO_BASED)
2182                 stag_info->addr_type = IRDMA_ADDR_TYPE_ZERO_BASED;
2183         else
2184                 stag_info->addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2185         stag_info->page_size = iwmr->page_size;
2186
2187         if (iwpbl->pbl_allocated) {
2188                 if (palloc->level == PBLE_LEVEL_1) {
2189                         stag_info->first_pm_pbl_index = palloc->level1.idx;
2190                         stag_info->chunk_size = 1;
2191                 } else {
2192                         stag_info->first_pm_pbl_index = palloc->level2.root.idx;
2193                         stag_info->chunk_size = 3;
2194                 }
2195         } else {
2196                 stag_info->reg_addr_pa = iwmr->pgaddrmem[0];
2197         }
2198
2199         cqp_info->cqp_cmd = IRDMA_OP_MR_REG_NON_SHARED;
2200         cqp_info->post_sq = 1;
2201         cqp_info->in.u.mr_reg_non_shared.dev = &iwdev->rf->sc_dev;
2202         cqp_info->in.u.mr_reg_non_shared.scratch = (uintptr_t)cqp_request;
2203         ret = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2204         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2205
2206         if (!ret)
2207                 iwmr->is_hwreg = 1;
2208
2209         return ret;
2210 }
2211
2212 /**
2213  * irdma_reg_user_mr - Register a user memory region
2214  * @pd: ptr of pd
2215  * @start: virtual start address
2216  * @len: length of mr
2217  * @virt: virtual address
2218  * @access: access of mr
2219  * @udata: user data
2220  */
2221 static struct ib_mr *
2222 irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
2223                   u64 virt, int access,
2224                   struct ib_udata *udata)
2225 {
2226 #define IRDMA_MEM_REG_MIN_REQ_LEN offsetofend(struct irdma_mem_reg_req, sq_pages)
2227         struct irdma_device *iwdev = to_iwdev(pd->device);
2228         struct irdma_ucontext *ucontext;
2229         struct irdma_pble_alloc *palloc;
2230         struct irdma_pbl *iwpbl;
2231         struct irdma_mr *iwmr;
2232         struct ib_umem *region;
2233         struct irdma_mem_reg_req req = {};
2234         u32 total, stag = 0;
2235         u8 shadow_pgcnt = 1;
2236         bool use_pbles = false;
2237         unsigned long flags;
2238         int err = -EINVAL;
2239         int ret;
2240
2241         if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
2242                 return ERR_PTR(-EINVAL);
2243
2244         if (udata->inlen < IRDMA_MEM_REG_MIN_REQ_LEN)
2245                 return ERR_PTR(-EINVAL);
2246
2247         region = ib_umem_get(pd->uobject->context, start, len, access, 0);
2248
2249         if (IS_ERR(region)) {
2250                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
2251                             "Failed to create ib_umem region\n");
2252                 return (struct ib_mr *)region;
2253         }
2254
2255         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen))) {
2256                 ib_umem_release(region);
2257                 return ERR_PTR(-EFAULT);
2258         }
2259
2260         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2261         if (!iwmr) {
2262                 ib_umem_release(region);
2263                 return ERR_PTR(-ENOMEM);
2264         }
2265
2266         iwpbl = &iwmr->iwpbl;
2267         iwpbl->iwmr = iwmr;
2268         iwmr->region = region;
2269         iwmr->ibmr.pd = pd;
2270         iwmr->ibmr.device = pd->device;
2271         iwmr->ibmr.iova = virt;
2272         iwmr->page_size = IRDMA_HW_PAGE_SIZE;
2273         iwmr->page_msk = ~(IRDMA_HW_PAGE_SIZE - 1);
2274
2275         iwmr->len = region->length;
2276         iwpbl->user_base = virt;
2277         palloc = &iwpbl->pble_alloc;
2278         iwmr->type = req.reg_type;
2279         iwmr->page_cnt = irdma_ib_umem_num_dma_blocks(region, iwmr->page_size, virt);
2280
2281         switch (req.reg_type) {
2282         case IRDMA_MEMREG_TYPE_QP:
2283                 total = req.sq_pages + req.rq_pages + shadow_pgcnt;
2284                 if (total > iwmr->page_cnt) {
2285                         err = -EINVAL;
2286                         goto error;
2287                 }
2288                 total = req.sq_pages + req.rq_pages;
2289                 use_pbles = (total > 2);
2290                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2291                 if (err)
2292                         goto error;
2293
2294 #if __FreeBSD_version >= 1400026
2295                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
2296 #else
2297                 ucontext = to_ucontext(pd->uobject->context);
2298 #endif
2299                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2300                 list_add_tail(&iwpbl->list, &ucontext->qp_reg_mem_list);
2301                 iwpbl->on_list = true;
2302                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2303                 break;
2304         case IRDMA_MEMREG_TYPE_CQ:
2305                 if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_CQ_RESIZE)
2306                         shadow_pgcnt = 0;
2307                 total = req.cq_pages + shadow_pgcnt;
2308                 if (total > iwmr->page_cnt) {
2309                         err = -EINVAL;
2310                         goto error;
2311                 }
2312
2313                 use_pbles = (req.cq_pages > 1);
2314                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2315                 if (err)
2316                         goto error;
2317
2318 #if __FreeBSD_version >= 1400026
2319                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
2320 #else
2321                 ucontext = to_ucontext(pd->uobject->context);
2322 #endif
2323                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2324                 list_add_tail(&iwpbl->list, &ucontext->cq_reg_mem_list);
2325                 iwpbl->on_list = true;
2326                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2327                 break;
2328         case IRDMA_MEMREG_TYPE_MEM:
2329                 use_pbles = (iwmr->page_cnt != 1);
2330
2331                 err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles, false);
2332                 if (err)
2333                         goto error;
2334
2335                 if (use_pbles) {
2336                         ret = irdma_check_mr_contiguous(palloc,
2337                                                         iwmr->page_size);
2338                         if (ret) {
2339                                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2340                                 iwpbl->pbl_allocated = false;
2341                         }
2342                 }
2343
2344                 stag = irdma_create_stag(iwdev);
2345                 if (!stag) {
2346                         err = -ENOMEM;
2347                         goto error;
2348                 }
2349
2350                 iwmr->stag = stag;
2351                 iwmr->ibmr.rkey = stag;
2352                 iwmr->ibmr.lkey = stag;
2353                 iwmr->access = access;
2354                 err = irdma_hwreg_mr(iwdev, iwmr, access);
2355                 if (err) {
2356                         irdma_free_stag(iwdev, stag);
2357                         goto error;
2358                 }
2359
2360                 break;
2361         default:
2362                 goto error;
2363         }
2364
2365         iwmr->type = req.reg_type;
2366
2367         return &iwmr->ibmr;
2368
2369 error:
2370         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated)
2371                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2372         ib_umem_release(region);
2373         kfree(iwmr);
2374
2375         return ERR_PTR(err);
2376 }
2377
2378 int
2379 irdma_hwdereg_mr(struct ib_mr *ib_mr)
2380 {
2381         struct irdma_device *iwdev = to_iwdev(ib_mr->device);
2382         struct irdma_mr *iwmr = to_iwmr(ib_mr);
2383         struct irdma_pd *iwpd = to_iwpd(ib_mr->pd);
2384         struct irdma_dealloc_stag_info *info;
2385         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2386         struct irdma_cqp_request *cqp_request;
2387         struct cqp_cmds_info *cqp_info;
2388         int status;
2389
2390         /*
2391          * Skip HW MR de-register when it is already de-registered during an MR re-reregister and the re-registration
2392          * fails
2393          */
2394         if (!iwmr->is_hwreg)
2395                 return 0;
2396
2397         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2398         if (!cqp_request)
2399                 return -ENOMEM;
2400
2401         cqp_info = &cqp_request->info;
2402         info = &cqp_info->in.u.dealloc_stag.info;
2403         memset(info, 0, sizeof(*info));
2404         info->pd_id = iwpd->sc_pd.pd_id;
2405         info->stag_idx = RS_64_1(ib_mr->rkey, IRDMA_CQPSQ_STAG_IDX_S);
2406         info->mr = true;
2407         if (iwpbl->pbl_allocated)
2408                 info->dealloc_pbl = true;
2409
2410         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2411         cqp_info->post_sq = 1;
2412         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2413         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2414         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2415         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2416
2417         if (!status)
2418                 iwmr->is_hwreg = 0;
2419
2420         return status;
2421 }
2422
2423 /*
2424  * irdma_rereg_mr_trans - Re-register a user MR for a change translation. @iwmr: ptr of iwmr @start: virtual start
2425  * address @len: length of mr @virt: virtual address
2426  *
2427  * Re-register a user memory region when a change translation is requested. Re-register a new region while reusing the
2428  * stag from the original registration.
2429  */
2430 struct ib_mr *
2431 irdma_rereg_mr_trans(struct irdma_mr *iwmr, u64 start, u64 len,
2432                      u64 virt, struct ib_udata *udata)
2433 {
2434         struct irdma_device *iwdev = to_iwdev(iwmr->ibmr.device);
2435         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2436         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2437         struct ib_pd *pd = iwmr->ibmr.pd;
2438         struct ib_umem *region;
2439         bool use_pbles;
2440         int err;
2441
2442         region = ib_umem_get(pd->uobject->context, start, len, iwmr->access, 0);
2443
2444         if (IS_ERR(region)) {
2445                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
2446                             "Failed to create ib_umem region\n");
2447                 return (struct ib_mr *)region;
2448         }
2449
2450         iwmr->region = region;
2451         iwmr->ibmr.iova = virt;
2452         iwmr->ibmr.pd = pd;
2453         iwmr->page_size = PAGE_SIZE;
2454
2455         iwmr->len = region->length;
2456         iwpbl->user_base = virt;
2457         iwmr->page_cnt = irdma_ib_umem_num_dma_blocks(region, iwmr->page_size,
2458                                                       virt);
2459
2460         use_pbles = (iwmr->page_cnt != 1);
2461
2462         err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles, false);
2463         if (err)
2464                 goto error;
2465
2466         if (use_pbles) {
2467                 err = irdma_check_mr_contiguous(palloc,
2468                                                 iwmr->page_size);
2469                 if (err) {
2470                         irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2471                         iwpbl->pbl_allocated = false;
2472                 }
2473         }
2474
2475         err = irdma_hwreg_mr(iwdev, iwmr, iwmr->access);
2476         if (err)
2477                 goto error;
2478
2479         return &iwmr->ibmr;
2480
2481 error:
2482         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated) {
2483                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2484                 iwpbl->pbl_allocated = false;
2485         }
2486         ib_umem_release(region);
2487         iwmr->region = NULL;
2488
2489         return ERR_PTR(err);
2490 }
2491
2492 /**
2493  * irdma_reg_phys_mr - register kernel physical memory
2494  * @pd: ibpd pointer
2495  * @addr: physical address of memory to register
2496  * @size: size of memory to register
2497  * @access: Access rights
2498  * @iova_start: start of virtual address for physical buffers
2499  */
2500 struct ib_mr *
2501 irdma_reg_phys_mr(struct ib_pd *pd, u64 addr, u64 size, int access,
2502                   u64 *iova_start)
2503 {
2504         struct irdma_device *iwdev = to_iwdev(pd->device);
2505         struct irdma_pbl *iwpbl;
2506         struct irdma_mr *iwmr;
2507         u32 stag;
2508         int ret;
2509
2510         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2511         if (!iwmr)
2512                 return ERR_PTR(-ENOMEM);
2513
2514         iwmr->ibmr.pd = pd;
2515         iwmr->ibmr.device = pd->device;
2516         iwpbl = &iwmr->iwpbl;
2517         iwpbl->iwmr = iwmr;
2518         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2519         iwpbl->user_base = *iova_start;
2520         stag = irdma_create_stag(iwdev);
2521         if (!stag) {
2522                 ret = -ENOMEM;
2523                 goto err;
2524         }
2525
2526         iwmr->stag = stag;
2527         iwmr->ibmr.iova = *iova_start;
2528         iwmr->ibmr.rkey = stag;
2529         iwmr->ibmr.lkey = stag;
2530         iwmr->page_cnt = 1;
2531         iwmr->pgaddrmem[0] = addr;
2532         iwmr->len = size;
2533         iwmr->page_size = SZ_4K;
2534         ret = irdma_hwreg_mr(iwdev, iwmr, access);
2535         if (ret) {
2536                 irdma_free_stag(iwdev, stag);
2537                 goto err;
2538         }
2539
2540         return &iwmr->ibmr;
2541
2542 err:
2543         kfree(iwmr);
2544
2545         return ERR_PTR(ret);
2546 }
2547
2548 /**
2549  * irdma_get_dma_mr - register physical mem
2550  * @pd: ptr of pd
2551  * @acc: access for memory
2552  */
2553 static struct ib_mr *
2554 irdma_get_dma_mr(struct ib_pd *pd, int acc)
2555 {
2556         u64 kva = 0;
2557
2558         return irdma_reg_phys_mr(pd, 0, 0, acc, &kva);
2559 }
2560
2561 /**
2562  * irdma_del_memlist - Deleting pbl list entries for CQ/QP
2563  * @iwmr: iwmr for IB's user page addresses
2564  * @ucontext: ptr to user context
2565  */
2566 void
2567 irdma_del_memlist(struct irdma_mr *iwmr,
2568                   struct irdma_ucontext *ucontext)
2569 {
2570         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2571         unsigned long flags;
2572
2573         switch (iwmr->type) {
2574         case IRDMA_MEMREG_TYPE_CQ:
2575                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2576                 if (iwpbl->on_list) {
2577                         iwpbl->on_list = false;
2578                         list_del(&iwpbl->list);
2579                 }
2580                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2581                 break;
2582         case IRDMA_MEMREG_TYPE_QP:
2583                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2584                 if (iwpbl->on_list) {
2585                         iwpbl->on_list = false;
2586                         list_del(&iwpbl->list);
2587                 }
2588                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2589                 break;
2590         default:
2591                 break;
2592         }
2593 }
2594
2595 /**
2596  * irdma_copy_sg_list - copy sg list for qp
2597  * @sg_list: copied into sg_list
2598  * @sgl: copy from sgl
2599  * @num_sges: count of sg entries
2600  */
2601 static void
2602 irdma_copy_sg_list(struct irdma_sge *sg_list, struct ib_sge *sgl,
2603                    int num_sges)
2604 {
2605         unsigned int i;
2606
2607         for (i = 0; i < num_sges; i++) {
2608                 sg_list[i].tag_off = sgl[i].addr;
2609                 sg_list[i].len = sgl[i].length;
2610                 sg_list[i].stag = sgl[i].lkey;
2611         }
2612 }
2613
2614 /**
2615  * irdma_post_send -  kernel application wr
2616  * @ibqp: qp ptr for wr
2617  * @ib_wr: work request ptr
2618  * @bad_wr: return of bad wr if err
2619  */
2620 static int
2621 irdma_post_send(struct ib_qp *ibqp,
2622                 const struct ib_send_wr *ib_wr,
2623                 const struct ib_send_wr **bad_wr)
2624 {
2625         struct irdma_qp *iwqp;
2626         struct irdma_qp_uk *ukqp;
2627         struct irdma_sc_dev *dev;
2628         struct irdma_post_sq_info info;
2629         int err = 0;
2630         unsigned long flags;
2631         bool inv_stag;
2632         struct irdma_ah *ah;
2633
2634         iwqp = to_iwqp(ibqp);
2635         ukqp = &iwqp->sc_qp.qp_uk;
2636         dev = &iwqp->iwdev->rf->sc_dev;
2637
2638         spin_lock_irqsave(&iwqp->lock, flags);
2639         while (ib_wr) {
2640                 memset(&info, 0, sizeof(info));
2641                 inv_stag = false;
2642                 info.wr_id = (ib_wr->wr_id);
2643                 if ((ib_wr->send_flags & IB_SEND_SIGNALED) || iwqp->sig_all)
2644                         info.signaled = true;
2645                 if (ib_wr->send_flags & IB_SEND_FENCE)
2646                         info.read_fence = true;
2647                 switch (ib_wr->opcode) {
2648                 case IB_WR_SEND_WITH_IMM:
2649                         if (ukqp->qp_caps & IRDMA_SEND_WITH_IMM) {
2650                                 info.imm_data_valid = true;
2651                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
2652                         } else {
2653                                 err = -EINVAL;
2654                                 break;
2655                         }
2656                         /* fallthrough */
2657                 case IB_WR_SEND:
2658                 case IB_WR_SEND_WITH_INV:
2659                         if (ib_wr->opcode == IB_WR_SEND ||
2660                             ib_wr->opcode == IB_WR_SEND_WITH_IMM) {
2661                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
2662                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL;
2663                                 else
2664                                         info.op_type = IRDMA_OP_TYPE_SEND;
2665                         } else {
2666                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
2667                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL_INV;
2668                                 else
2669                                         info.op_type = IRDMA_OP_TYPE_SEND_INV;
2670                                 info.stag_to_inv = ib_wr->ex.invalidate_rkey;
2671                         }
2672
2673                         info.op.send.num_sges = ib_wr->num_sge;
2674                         info.op.send.sg_list = (struct irdma_sge *)ib_wr->sg_list;
2675                         if (iwqp->ibqp.qp_type == IB_QPT_UD ||
2676                             iwqp->ibqp.qp_type == IB_QPT_GSI) {
2677                                 ah = to_iwah(ud_wr(ib_wr)->ah);
2678                                 info.op.send.ah_id = ah->sc_ah.ah_info.ah_idx;
2679                                 info.op.send.qkey = ud_wr(ib_wr)->remote_qkey;
2680                                 info.op.send.dest_qp = ud_wr(ib_wr)->remote_qpn;
2681                         }
2682
2683                         if (ib_wr->send_flags & IB_SEND_INLINE)
2684                                 err = irdma_uk_inline_send(ukqp, &info, false);
2685                         else
2686                                 err = irdma_uk_send(ukqp, &info, false);
2687                         break;
2688                 case IB_WR_RDMA_WRITE_WITH_IMM:
2689                         if (ukqp->qp_caps & IRDMA_WRITE_WITH_IMM) {
2690                                 info.imm_data_valid = true;
2691                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
2692                         } else {
2693                                 err = -EINVAL;
2694                                 break;
2695                         }
2696                         /* fallthrough */
2697                 case IB_WR_RDMA_WRITE:
2698                         if (ib_wr->send_flags & IB_SEND_SOLICITED)
2699                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE_SOL;
2700                         else
2701                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE;
2702
2703                         info.op.rdma_write.num_lo_sges = ib_wr->num_sge;
2704                         info.op.rdma_write.lo_sg_list = (void *)ib_wr->sg_list;
2705                         info.op.rdma_write.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
2706                         info.op.rdma_write.rem_addr.stag = rdma_wr(ib_wr)->rkey;
2707                         if (ib_wr->send_flags & IB_SEND_INLINE)
2708                                 err = irdma_uk_inline_rdma_write(ukqp, &info, false);
2709                         else
2710                                 err = irdma_uk_rdma_write(ukqp, &info, false);
2711                         break;
2712                 case IB_WR_RDMA_READ_WITH_INV:
2713                         inv_stag = true;
2714                         /* fallthrough */
2715                 case IB_WR_RDMA_READ:
2716                         if (ib_wr->num_sge >
2717                             dev->hw_attrs.uk_attrs.max_hw_read_sges) {
2718                                 err = -EINVAL;
2719                                 break;
2720                         }
2721                         info.op_type = IRDMA_OP_TYPE_RDMA_READ;
2722                         info.op.rdma_read.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
2723                         info.op.rdma_read.rem_addr.stag = rdma_wr(ib_wr)->rkey;
2724                         info.op.rdma_read.lo_sg_list = (void *)ib_wr->sg_list;
2725                         info.op.rdma_read.num_lo_sges = ib_wr->num_sge;
2726                         err = irdma_uk_rdma_read(ukqp, &info, inv_stag, false);
2727                         break;
2728                 case IB_WR_LOCAL_INV:
2729                         info.op_type = IRDMA_OP_TYPE_INV_STAG;
2730                         info.op.inv_local_stag.target_stag = ib_wr->ex.invalidate_rkey;
2731                         err = irdma_uk_stag_local_invalidate(ukqp, &info, true);
2732                         break;
2733                 case IB_WR_REG_MR:{
2734                                 struct irdma_mr *iwmr = to_iwmr(reg_wr(ib_wr)->mr);
2735                                 struct irdma_pble_alloc *palloc = &iwmr->iwpbl.pble_alloc;
2736                                 struct irdma_fast_reg_stag_info stag_info = {0};
2737
2738                                 stag_info.signaled = info.signaled;
2739                                 stag_info.read_fence = info.read_fence;
2740                                 stag_info.access_rights = irdma_get_mr_access(reg_wr(ib_wr)->access);
2741                                 stag_info.stag_key = reg_wr(ib_wr)->key & 0xff;
2742                                 stag_info.stag_idx = reg_wr(ib_wr)->key >> 8;
2743                                 stag_info.page_size = reg_wr(ib_wr)->mr->page_size;
2744                                 stag_info.wr_id = ib_wr->wr_id;
2745                                 stag_info.addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2746                                 stag_info.va = (void *)(uintptr_t)iwmr->ibmr.iova;
2747                                 stag_info.total_len = iwmr->ibmr.length;
2748                                 if (palloc->level == PBLE_LEVEL_2) {
2749                                         stag_info.chunk_size = 3;
2750                                         stag_info.first_pm_pbl_index = palloc->level2.root.idx;
2751                                 } else {
2752                                         stag_info.chunk_size = 1;
2753                                         stag_info.first_pm_pbl_index = palloc->level1.idx;
2754                                 }
2755                                 stag_info.local_fence = ib_wr->send_flags & IB_SEND_FENCE;
2756                                 err = irdma_sc_mr_fast_register(&iwqp->sc_qp, &stag_info,
2757                                                                 true);
2758                                 break;
2759                         }
2760                 default:
2761                         err = -EINVAL;
2762                         irdma_debug(iwdev_to_idev(iwqp->iwdev),
2763                                     IRDMA_DEBUG_VERBS,
2764                                     "upost_send bad opcode = 0x%x\n",
2765                                     ib_wr->opcode);
2766                         break;
2767                 }
2768
2769                 if (err)
2770                         break;
2771                 ib_wr = ib_wr->next;
2772         }
2773
2774         if (!iwqp->flush_issued) {
2775                 if (iwqp->hw_iwarp_state <= IRDMA_QP_STATE_RTS)
2776                         irdma_uk_qp_post_wr(ukqp);
2777                 spin_unlock_irqrestore(&iwqp->lock, flags);
2778         } else {
2779                 spin_unlock_irqrestore(&iwqp->lock, flags);
2780                 irdma_sched_qp_flush_work(iwqp);
2781         }
2782         if (err)
2783                 *bad_wr = ib_wr;
2784
2785         return err;
2786 }
2787
2788 /**
2789  * irdma_post_recv - post receive wr for kernel application
2790  * @ibqp: ib qp pointer
2791  * @ib_wr: work request for receive
2792  * @bad_wr: bad wr caused an error
2793  */
2794 static int
2795 irdma_post_recv(struct ib_qp *ibqp,
2796                 const struct ib_recv_wr *ib_wr,
2797                 const struct ib_recv_wr **bad_wr)
2798 {
2799         struct irdma_qp *iwqp = to_iwqp(ibqp);
2800         struct irdma_qp_uk *ukqp = &iwqp->sc_qp.qp_uk;
2801         struct irdma_post_rq_info post_recv = {0};
2802         struct irdma_sge *sg_list = iwqp->sg_list;
2803         unsigned long flags;
2804         int err = 0;
2805
2806         spin_lock_irqsave(&iwqp->lock, flags);
2807
2808         while (ib_wr) {
2809                 if (ib_wr->num_sge > ukqp->max_rq_frag_cnt) {
2810                         err = -EINVAL;
2811                         goto out;
2812                 }
2813                 post_recv.num_sges = ib_wr->num_sge;
2814                 post_recv.wr_id = ib_wr->wr_id;
2815                 irdma_copy_sg_list(sg_list, ib_wr->sg_list, ib_wr->num_sge);
2816                 post_recv.sg_list = sg_list;
2817                 err = irdma_uk_post_receive(ukqp, &post_recv);
2818                 if (err) {
2819                         irdma_debug(iwdev_to_idev(iwqp->iwdev),
2820                                     IRDMA_DEBUG_VERBS, "post_recv err %d\n",
2821                                     err);
2822                         goto out;
2823                 }
2824
2825                 ib_wr = ib_wr->next;
2826         }
2827
2828 out:
2829         spin_unlock_irqrestore(&iwqp->lock, flags);
2830         if (iwqp->flush_issued)
2831                 irdma_sched_qp_flush_work(iwqp);
2832         if (err)
2833                 *bad_wr = ib_wr;
2834
2835         return err;
2836 }
2837
2838 /**
2839  * irdma_flush_err_to_ib_wc_status - return change flush error code to IB status
2840  * @opcode: iwarp flush code
2841  */
2842 static enum ib_wc_status
2843 irdma_flush_err_to_ib_wc_status(enum irdma_flush_opcode opcode)
2844 {
2845         switch (opcode) {
2846         case FLUSH_PROT_ERR:
2847                 return IB_WC_LOC_PROT_ERR;
2848         case FLUSH_REM_ACCESS_ERR:
2849                 return IB_WC_REM_ACCESS_ERR;
2850         case FLUSH_LOC_QP_OP_ERR:
2851                 return IB_WC_LOC_QP_OP_ERR;
2852         case FLUSH_REM_OP_ERR:
2853                 return IB_WC_REM_OP_ERR;
2854         case FLUSH_LOC_LEN_ERR:
2855                 return IB_WC_LOC_LEN_ERR;
2856         case FLUSH_GENERAL_ERR:
2857                 return IB_WC_WR_FLUSH_ERR;
2858         case FLUSH_MW_BIND_ERR:
2859                 return IB_WC_MW_BIND_ERR;
2860         case FLUSH_REM_INV_REQ_ERR:
2861                 return IB_WC_REM_INV_REQ_ERR;
2862         case FLUSH_RETRY_EXC_ERR:
2863                 return IB_WC_RETRY_EXC_ERR;
2864         case FLUSH_FATAL_ERR:
2865         default:
2866                 return IB_WC_FATAL_ERR;
2867         }
2868 }
2869
2870 static inline void
2871 set_ib_wc_op_sq(struct irdma_cq_poll_info *cq_poll_info,
2872                 struct ib_wc *entry)
2873 {
2874         struct irdma_sc_qp *qp;
2875
2876         switch (cq_poll_info->op_type) {
2877         case IRDMA_OP_TYPE_RDMA_WRITE:
2878         case IRDMA_OP_TYPE_RDMA_WRITE_SOL:
2879                 entry->opcode = IB_WC_RDMA_WRITE;
2880                 break;
2881         case IRDMA_OP_TYPE_RDMA_READ_INV_STAG:
2882         case IRDMA_OP_TYPE_RDMA_READ:
2883                 entry->opcode = IB_WC_RDMA_READ;
2884                 break;
2885         case IRDMA_OP_TYPE_SEND_SOL:
2886         case IRDMA_OP_TYPE_SEND_SOL_INV:
2887         case IRDMA_OP_TYPE_SEND_INV:
2888         case IRDMA_OP_TYPE_SEND:
2889                 entry->opcode = IB_WC_SEND;
2890                 break;
2891         case IRDMA_OP_TYPE_FAST_REG_NSMR:
2892                 entry->opcode = IB_WC_REG_MR;
2893                 break;
2894         case IRDMA_OP_TYPE_INV_STAG:
2895                 entry->opcode = IB_WC_LOCAL_INV;
2896                 break;
2897         default:
2898                 qp = cq_poll_info->qp_handle;
2899                 ibdev_err(irdma_get_ibdev(qp->dev), "Invalid opcode = %d in CQE\n",
2900                           cq_poll_info->op_type);
2901                 entry->status = IB_WC_GENERAL_ERR;
2902         }
2903 }
2904
2905 static inline void
2906 set_ib_wc_op_rq(struct irdma_cq_poll_info *cq_poll_info,
2907                 struct ib_wc *entry, bool send_imm_support)
2908 {
2909         /**
2910          * iWARP does not support sendImm, so the presence of Imm data
2911          * must be WriteImm.
2912          */
2913         if (!send_imm_support) {
2914                 entry->opcode = cq_poll_info->imm_valid ?
2915                     IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
2916                 return;
2917         }
2918         switch (cq_poll_info->op_type) {
2919         case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
2920         case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
2921                 entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2922                 break;
2923         default:
2924                 entry->opcode = IB_WC_RECV;
2925         }
2926 }
2927
2928 /**
2929  * irdma_process_cqe - process cqe info
2930  * @entry: processed cqe
2931  * @cq_poll_info: cqe info
2932  */
2933 static void
2934 irdma_process_cqe(struct ib_wc *entry,
2935                   struct irdma_cq_poll_info *cq_poll_info)
2936 {
2937         struct irdma_sc_qp *qp;
2938
2939         entry->wc_flags = 0;
2940         entry->pkey_index = 0;
2941         entry->wr_id = cq_poll_info->wr_id;
2942
2943         qp = cq_poll_info->qp_handle;
2944         entry->qp = qp->qp_uk.back_qp;
2945
2946         if (cq_poll_info->error) {
2947                 entry->status = (cq_poll_info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ?
2948                     irdma_flush_err_to_ib_wc_status(cq_poll_info->minor_err) : IB_WC_GENERAL_ERR;
2949
2950                 entry->vendor_err = cq_poll_info->major_err << 16 |
2951                     cq_poll_info->minor_err;
2952         } else {
2953                 entry->status = IB_WC_SUCCESS;
2954                 if (cq_poll_info->imm_valid) {
2955                         entry->ex.imm_data = htonl(cq_poll_info->imm_data);
2956                         entry->wc_flags |= IB_WC_WITH_IMM;
2957                 }
2958                 if (cq_poll_info->ud_smac_valid) {
2959                         ether_addr_copy(entry->smac, cq_poll_info->ud_smac);
2960                         entry->wc_flags |= IB_WC_WITH_SMAC;
2961                 }
2962
2963                 if (cq_poll_info->ud_vlan_valid) {
2964                         u16 vlan = cq_poll_info->ud_vlan & EVL_VLID_MASK;
2965
2966                         entry->sl = cq_poll_info->ud_vlan >> VLAN_PRIO_SHIFT;
2967                         if (vlan) {
2968                                 entry->vlan_id = vlan;
2969                                 entry->wc_flags |= IB_WC_WITH_VLAN;
2970                         }
2971                 } else {
2972                         entry->sl = 0;
2973                 }
2974         }
2975
2976         if (cq_poll_info->q_type == IRDMA_CQE_QTYPE_SQ) {
2977                 set_ib_wc_op_sq(cq_poll_info, entry);
2978         } else {
2979                 set_ib_wc_op_rq(cq_poll_info, entry,
2980                                 qp->qp_uk.qp_caps & IRDMA_SEND_WITH_IMM ?
2981                                 true : false);
2982                 if (qp->qp_uk.qp_type != IRDMA_QP_TYPE_ROCE_UD &&
2983                     cq_poll_info->stag_invalid_set) {
2984                         entry->ex.invalidate_rkey = cq_poll_info->inv_stag;
2985                         entry->wc_flags |= IB_WC_WITH_INVALIDATE;
2986                 }
2987         }
2988
2989         if (qp->qp_uk.qp_type == IRDMA_QP_TYPE_ROCE_UD) {
2990                 entry->src_qp = cq_poll_info->ud_src_qpn;
2991                 entry->slid = 0;
2992                 entry->wc_flags |=
2993                     (IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE);
2994                 entry->network_hdr_type = cq_poll_info->ipv4 ?
2995                     RDMA_NETWORK_IPV4 :
2996                     RDMA_NETWORK_IPV6;
2997         } else {
2998                 entry->src_qp = cq_poll_info->qp_id;
2999         }
3000
3001         entry->byte_len = cq_poll_info->bytes_xfered;
3002 }
3003
3004 /**
3005  * irdma_poll_one - poll one entry of the CQ
3006  * @ukcq: ukcq to poll
3007  * @cur_cqe: current CQE info to be filled in
3008  * @entry: ibv_wc object to be filled for non-extended CQ or NULL for extended CQ
3009  *
3010  * Returns the internal irdma device error code or 0 on success
3011  */
3012 static inline int
3013 irdma_poll_one(struct irdma_cq_uk *ukcq,
3014                struct irdma_cq_poll_info *cur_cqe,
3015                struct ib_wc *entry)
3016 {
3017         int ret = irdma_uk_cq_poll_cmpl(ukcq, cur_cqe);
3018
3019         if (ret)
3020                 return ret;
3021
3022         irdma_process_cqe(entry, cur_cqe);
3023
3024         return 0;
3025 }
3026
3027 /**
3028  * __irdma_poll_cq - poll cq for completion (kernel apps)
3029  * @iwcq: cq to poll
3030  * @num_entries: number of entries to poll
3031  * @entry: wr of a completed entry
3032  */
3033 static int
3034 __irdma_poll_cq(struct irdma_cq *iwcq, int num_entries, struct ib_wc *entry)
3035 {
3036         struct list_head *tmp_node, *list_node;
3037         struct irdma_cq_buf *last_buf = NULL;
3038         struct irdma_cq_poll_info *cur_cqe = &iwcq->cur_cqe;
3039         struct irdma_cq_buf *cq_buf;
3040         int ret;
3041         struct irdma_device *iwdev;
3042         struct irdma_cq_uk *ukcq;
3043         bool cq_new_cqe = false;
3044         int resized_bufs = 0;
3045         int npolled = 0;
3046
3047         iwdev = to_iwdev(iwcq->ibcq.device);
3048         ukcq = &iwcq->sc_cq.cq_uk;
3049
3050         /* go through the list of previously resized CQ buffers */
3051         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
3052                 cq_buf = container_of(list_node, struct irdma_cq_buf, list);
3053                 while (npolled < num_entries) {
3054                         ret = irdma_poll_one(&cq_buf->cq_uk, cur_cqe, entry + npolled);
3055                         if (!ret) {
3056                                 ++npolled;
3057                                 cq_new_cqe = true;
3058                                 continue;
3059                         }
3060                         if (ret == -ENOENT)
3061                                 break;
3062                         /* QP using the CQ is destroyed. Skip reporting this CQE */
3063                         if (ret == -EFAULT) {
3064                                 cq_new_cqe = true;
3065                                 continue;
3066                         }
3067                         goto error;
3068                 }
3069
3070                 /* save the resized CQ buffer which received the last cqe */
3071                 if (cq_new_cqe)
3072                         last_buf = cq_buf;
3073                 cq_new_cqe = false;
3074         }
3075
3076         /* check the current CQ for new cqes */
3077         while (npolled < num_entries) {
3078                 ret = irdma_poll_one(ukcq, cur_cqe, entry + npolled);
3079                 if (ret == -ENOENT) {
3080                         ret = irdma_generated_cmpls(iwcq, cur_cqe);
3081                         if (!ret)
3082                                 irdma_process_cqe(entry + npolled, cur_cqe);
3083                 }
3084                 if (!ret) {
3085                         ++npolled;
3086                         cq_new_cqe = true;
3087                         continue;
3088                 }
3089
3090                 if (ret == -ENOENT)
3091                         break;
3092                 /* QP using the CQ is destroyed. Skip reporting this CQE */
3093                 if (ret == -EFAULT) {
3094                         cq_new_cqe = true;
3095                         continue;
3096                 }
3097                 goto error;
3098         }
3099
3100         if (cq_new_cqe)
3101                 /* all previous CQ resizes are complete */
3102                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, NULL);
3103         else if (last_buf)
3104                 /* only CQ resizes up to the last_buf are complete */
3105                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, last_buf);
3106         if (resized_bufs)
3107                 /* report to the HW the number of complete CQ resizes */
3108                 irdma_uk_cq_set_resized_cnt(ukcq, resized_bufs);
3109
3110         return npolled;
3111 error:
3112         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
3113                     "%s: Error polling CQ, irdma_err: %d\n", __func__, ret);
3114
3115         return ret;
3116 }
3117
3118 /**
3119  * irdma_poll_cq - poll cq for completion (kernel apps)
3120  * @ibcq: cq to poll
3121  * @num_entries: number of entries to poll
3122  * @entry: wr of a completed entry
3123  */
3124 static int
3125 irdma_poll_cq(struct ib_cq *ibcq, int num_entries,
3126               struct ib_wc *entry)
3127 {
3128         struct irdma_cq *iwcq;
3129         unsigned long flags;
3130         int ret;
3131
3132         iwcq = to_iwcq(ibcq);
3133
3134         spin_lock_irqsave(&iwcq->lock, flags);
3135         ret = __irdma_poll_cq(iwcq, num_entries, entry);
3136         spin_unlock_irqrestore(&iwcq->lock, flags);
3137
3138         return ret;
3139 }
3140
3141 /**
3142  * irdma_req_notify_cq - arm cq kernel application
3143  * @ibcq: cq to arm
3144  * @notify_flags: notofication flags
3145  */
3146 static int
3147 irdma_req_notify_cq(struct ib_cq *ibcq,
3148                     enum ib_cq_notify_flags notify_flags)
3149 {
3150         struct irdma_cq *iwcq;
3151         struct irdma_cq_uk *ukcq;
3152         unsigned long flags;
3153         enum irdma_cmpl_notify cq_notify = IRDMA_CQ_COMPL_EVENT;
3154         bool promo_event = false;
3155         int ret = 0;
3156
3157         iwcq = to_iwcq(ibcq);
3158         ukcq = &iwcq->sc_cq.cq_uk;
3159
3160         spin_lock_irqsave(&iwcq->lock, flags);
3161         if (notify_flags == IB_CQ_SOLICITED) {
3162                 cq_notify = IRDMA_CQ_COMPL_SOLICITED;
3163         } else {
3164                 if (iwcq->last_notify == IRDMA_CQ_COMPL_SOLICITED)
3165                         promo_event = true;
3166         }
3167
3168         if (!atomic_cmpxchg(&iwcq->armed, 0, 1) || promo_event) {
3169                 iwcq->last_notify = cq_notify;
3170                 irdma_uk_cq_request_notification(ukcq, cq_notify);
3171         }
3172
3173         if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3174             (!irdma_cq_empty(iwcq) || !list_empty(&iwcq->cmpl_generated)))
3175                 ret = 1;
3176         spin_unlock_irqrestore(&iwcq->lock, flags);
3177
3178         return ret;
3179 }
3180
3181 /**
3182  * mcast_list_add -  Add a new mcast item to list
3183  * @rf: RDMA PCI function
3184  * @new_elem: pointer to element to add
3185  */
3186 static void
3187 mcast_list_add(struct irdma_pci_f *rf,
3188                struct mc_table_list *new_elem)
3189 {
3190         list_add(&new_elem->list, &rf->mc_qht_list.list);
3191 }
3192
3193 /**
3194  * mcast_list_del - Remove an mcast item from list
3195  * @mc_qht_elem: pointer to mcast table list element
3196  */
3197 static void
3198 mcast_list_del(struct mc_table_list *mc_qht_elem)
3199 {
3200         if (mc_qht_elem)
3201                 list_del(&mc_qht_elem->list);
3202 }
3203
3204 /**
3205  * mcast_list_lookup_ip - Search mcast list for address
3206  * @rf: RDMA PCI function
3207  * @ip_mcast: pointer to mcast IP address
3208  */
3209 static struct mc_table_list *
3210 mcast_list_lookup_ip(struct irdma_pci_f *rf,
3211                      u32 *ip_mcast)
3212 {
3213         struct mc_table_list *mc_qht_el;
3214         struct list_head *pos, *q;
3215
3216         list_for_each_safe(pos, q, &rf->mc_qht_list.list) {
3217                 mc_qht_el = list_entry(pos, struct mc_table_list, list);
3218                 if (!memcmp(mc_qht_el->mc_info.dest_ip, ip_mcast,
3219                             sizeof(mc_qht_el->mc_info.dest_ip)))
3220                         return mc_qht_el;
3221         }
3222
3223         return NULL;
3224 }
3225
3226 /**
3227  * irdma_mcast_cqp_op - perform a mcast cqp operation
3228  * @iwdev: irdma device
3229  * @mc_grp_ctx: mcast group info
3230  * @op: operation
3231  *
3232  * returns error status
3233  */
3234 static int
3235 irdma_mcast_cqp_op(struct irdma_device *iwdev,
3236                    struct irdma_mcast_grp_info *mc_grp_ctx, u8 op)
3237 {
3238         struct cqp_cmds_info *cqp_info;
3239         struct irdma_cqp_request *cqp_request;
3240         int status;
3241
3242         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3243         if (!cqp_request)
3244                 return -ENOMEM;
3245
3246         cqp_request->info.in.u.mc_create.info = *mc_grp_ctx;
3247         cqp_info = &cqp_request->info;
3248         cqp_info->cqp_cmd = op;
3249         cqp_info->post_sq = 1;
3250         cqp_info->in.u.mc_create.scratch = (uintptr_t)cqp_request;
3251         cqp_info->in.u.mc_create.cqp = &iwdev->rf->cqp.sc_cqp;
3252         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3253         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3254
3255         return status;
3256 }
3257
3258 /**
3259  * irdma_attach_mcast - attach a qp to a multicast group
3260  * @ibqp: ptr to qp
3261  * @ibgid: pointer to global ID
3262  * @lid: local ID
3263  *
3264  * returns error status
3265  */
3266 static int
3267 irdma_attach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3268 {
3269         struct irdma_qp *iwqp = to_iwqp(ibqp);
3270         struct irdma_device *iwdev = iwqp->iwdev;
3271         struct irdma_pci_f *rf = iwdev->rf;
3272         struct mc_table_list *mc_qht_elem;
3273         struct irdma_mcast_grp_ctx_entry_info mcg_info = {0};
3274         unsigned long flags;
3275         u32 ip_addr[4] = {0};
3276         u32 mgn;
3277         u32 no_mgs;
3278         int ret = 0;
3279         bool ipv4;
3280         u16 vlan_id;
3281         union {
3282                 struct sockaddr saddr;
3283                 struct sockaddr_in saddr_in;
3284                 struct sockaddr_in6 saddr_in6;
3285         } sgid_addr;
3286         unsigned char dmac[ETH_ALEN];
3287
3288         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3289
3290         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid)) {
3291                 irdma_copy_ip_ntohl(ip_addr,
3292                                     sgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32);
3293                 irdma_netdev_vlan_ipv6(ip_addr, &vlan_id, NULL);
3294                 ipv4 = false;
3295                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
3296                             "qp_id=%d, IP6address=%pI6\n", ibqp->qp_num,
3297                             ip_addr);
3298                 irdma_mcast_mac_v6(ip_addr, dmac);
3299         } else {
3300                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3301                 ipv4 = true;
3302                 vlan_id = irdma_get_vlan_ipv4(ip_addr);
3303                 irdma_mcast_mac_v4(ip_addr, dmac);
3304                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
3305                             "qp_id=%d, IP4address=%pI4, MAC=%pM\n",
3306                             ibqp->qp_num, ip_addr, dmac);
3307         }
3308
3309         spin_lock_irqsave(&rf->qh_list_lock, flags);
3310         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3311         if (!mc_qht_elem) {
3312                 struct irdma_dma_mem *dma_mem_mc;
3313
3314                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3315                 mc_qht_elem = kzalloc(sizeof(*mc_qht_elem), GFP_KERNEL);
3316                 if (!mc_qht_elem)
3317                         return -ENOMEM;
3318
3319                 mc_qht_elem->mc_info.ipv4_valid = ipv4;
3320                 memcpy(mc_qht_elem->mc_info.dest_ip, ip_addr,
3321                        sizeof(mc_qht_elem->mc_info.dest_ip));
3322                 ret = irdma_alloc_rsrc(rf, rf->allocated_mcgs, rf->max_mcg,
3323                                        &mgn, &rf->next_mcg);
3324                 if (ret) {
3325                         kfree(mc_qht_elem);
3326                         return -ENOMEM;
3327                 }
3328
3329                 mc_qht_elem->mc_info.mgn = mgn;
3330                 dma_mem_mc = &mc_qht_elem->mc_grp_ctx.dma_mem_mc;
3331                 dma_mem_mc->size = sizeof(u64)* IRDMA_MAX_MGS_PER_CTX;
3332                 dma_mem_mc->va = irdma_allocate_dma_mem(&rf->hw, dma_mem_mc,
3333                                                         dma_mem_mc->size,
3334                                                         IRDMA_HW_PAGE_SIZE);
3335                 if (!dma_mem_mc->va) {
3336                         irdma_free_rsrc(rf, rf->allocated_mcgs, mgn);
3337                         kfree(mc_qht_elem);
3338                         return -ENOMEM;
3339                 }
3340
3341                 mc_qht_elem->mc_grp_ctx.mg_id = (u16)mgn;
3342                 memcpy(mc_qht_elem->mc_grp_ctx.dest_ip_addr, ip_addr,
3343                        sizeof(mc_qht_elem->mc_grp_ctx.dest_ip_addr));
3344                 mc_qht_elem->mc_grp_ctx.ipv4_valid = ipv4;
3345                 mc_qht_elem->mc_grp_ctx.vlan_id = vlan_id;
3346                 if (vlan_id < VLAN_N_VID)
3347                         mc_qht_elem->mc_grp_ctx.vlan_valid = true;
3348                 mc_qht_elem->mc_grp_ctx.hmc_fcn_id = iwdev->rf->sc_dev.hmc_fn_id;
3349                 mc_qht_elem->mc_grp_ctx.qs_handle =
3350                     iwqp->sc_qp.vsi->qos[iwqp->sc_qp.user_pri].qs_handle;
3351                 ether_addr_copy(mc_qht_elem->mc_grp_ctx.dest_mac_addr, dmac);
3352
3353                 spin_lock_irqsave(&rf->qh_list_lock, flags);
3354                 mcast_list_add(rf, mc_qht_elem);
3355         } else {
3356                 if (mc_qht_elem->mc_grp_ctx.no_of_mgs ==
3357                     IRDMA_MAX_MGS_PER_CTX) {
3358                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3359                         return -ENOMEM;
3360                 }
3361         }
3362
3363         mcg_info.qp_id = iwqp->ibqp.qp_num;
3364         no_mgs = mc_qht_elem->mc_grp_ctx.no_of_mgs;
3365         irdma_sc_add_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3366         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3367
3368         /* Only if there is a change do we need to modify or create */
3369         if (!no_mgs) {
3370                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3371                                          IRDMA_OP_MC_CREATE);
3372         } else if (no_mgs != mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3373                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3374                                          IRDMA_OP_MC_MODIFY);
3375         } else {
3376                 return 0;
3377         }
3378
3379         if (ret)
3380                 goto error;
3381
3382         return 0;
3383
3384 error:
3385         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3386         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3387                 mcast_list_del(mc_qht_elem);
3388                 irdma_free_dma_mem(&rf->hw,
3389                                    &mc_qht_elem->mc_grp_ctx.dma_mem_mc);
3390                 irdma_free_rsrc(rf, rf->allocated_mcgs,
3391                                 mc_qht_elem->mc_grp_ctx.mg_id);
3392                 kfree(mc_qht_elem);
3393         }
3394
3395         return ret;
3396 }
3397
3398 /**
3399  * irdma_detach_mcast - detach a qp from a multicast group
3400  * @ibqp: ptr to qp
3401  * @ibgid: pointer to global ID
3402  * @lid: local ID
3403  *
3404  * returns error status
3405  */
3406 static int
3407 irdma_detach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3408 {
3409         struct irdma_qp *iwqp = to_iwqp(ibqp);
3410         struct irdma_device *iwdev = iwqp->iwdev;
3411         struct irdma_pci_f *rf = iwdev->rf;
3412         u32 ip_addr[4] = {0};
3413         struct mc_table_list *mc_qht_elem;
3414         struct irdma_mcast_grp_ctx_entry_info mcg_info = {0};
3415         int ret;
3416         unsigned long flags;
3417         union {
3418                 struct sockaddr saddr;
3419                 struct sockaddr_in saddr_in;
3420                 struct sockaddr_in6 saddr_in6;
3421         } sgid_addr;
3422
3423         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3424         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid))
3425                 irdma_copy_ip_ntohl(ip_addr,
3426                                     sgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32);
3427         else
3428                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3429
3430         spin_lock_irqsave(&rf->qh_list_lock, flags);
3431         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3432         if (!mc_qht_elem) {
3433                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3434                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
3435                             "address not found MCG\n");
3436                 return 0;
3437         }
3438
3439         mcg_info.qp_id = iwqp->ibqp.qp_num;
3440         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3441         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3442                 mcast_list_del(mc_qht_elem);
3443                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3444                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3445                                          IRDMA_OP_MC_DESTROY);
3446                 if (ret) {
3447                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
3448                                     "failed MC_DESTROY MCG\n");
3449                         spin_lock_irqsave(&rf->qh_list_lock, flags);
3450                         mcast_list_add(rf, mc_qht_elem);
3451                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3452                         return -EAGAIN;
3453                 }
3454
3455                 irdma_free_dma_mem(&rf->hw,
3456                                    &mc_qht_elem->mc_grp_ctx.dma_mem_mc);
3457                 irdma_free_rsrc(rf, rf->allocated_mcgs,
3458                                 mc_qht_elem->mc_grp_ctx.mg_id);
3459                 kfree(mc_qht_elem);
3460         } else {
3461                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3462                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3463                                          IRDMA_OP_MC_MODIFY);
3464                 if (ret) {
3465                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
3466                                     "failed Modify MCG\n");
3467                         return ret;
3468                 }
3469         }
3470
3471         return 0;
3472 }
3473
3474 /**
3475  * irdma_query_ah - Query address handle
3476  * @ibah: pointer to address handle
3477  * @ah_attr: address handle attributes
3478  */
3479 static int
3480 irdma_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
3481 {
3482         struct irdma_ah *ah = to_iwah(ibah);
3483
3484         memset(ah_attr, 0, sizeof(*ah_attr));
3485         if (ah->av.attrs.ah_flags & IB_AH_GRH) {
3486                 ah_attr->ah_flags = IB_AH_GRH;
3487                 ah_attr->grh.flow_label = ah->sc_ah.ah_info.flow_label;
3488                 ah_attr->grh.traffic_class = ah->sc_ah.ah_info.tc_tos;
3489                 ah_attr->grh.hop_limit = ah->sc_ah.ah_info.hop_ttl;
3490                 ah_attr->grh.sgid_index = ah->sgid_index;
3491                 ah_attr->grh.sgid_index = ah->sgid_index;
3492                 memcpy(&ah_attr->grh.dgid, &ah->dgid,
3493                        sizeof(ah_attr->grh.dgid));
3494         }
3495
3496         return 0;
3497 }
3498
3499 static __be64 irdma_mac_to_guid(struct ifnet *ndev){
3500         const unsigned char *mac = IF_LLADDR(ndev);
3501         __be64 guid;
3502         unsigned char *dst = (unsigned char *)&guid;
3503
3504         dst[0] = mac[0] ^ 2;
3505         dst[1] = mac[1];
3506         dst[2] = mac[2];
3507         dst[3] = 0xff;
3508         dst[4] = 0xfe;
3509         dst[5] = mac[3];
3510         dst[6] = mac[4];
3511         dst[7] = mac[5];
3512
3513         return guid;
3514 }
3515
3516 static struct ifnet *
3517 irdma_get_netdev(struct ib_device *ibdev, u8 port_num)
3518 {
3519         struct irdma_device *iwdev = to_iwdev(ibdev);
3520
3521         if (iwdev->netdev) {
3522                 dev_hold(iwdev->netdev);
3523                 return iwdev->netdev;
3524         }
3525
3526         return NULL;
3527 }
3528
3529 static void
3530 irdma_set_device_ops(struct ib_device *ibdev)
3531 {
3532         struct ib_device *dev_ops = ibdev;
3533
3534 #if __FreeBSD_version >= 1400000
3535         dev_ops->ops.driver_id = RDMA_DRIVER_I40IW;
3536         dev_ops->ops.size_ib_ah = IRDMA_SET_RDMA_OBJ_SIZE(ib_ah, irdma_ah, ibah);
3537         dev_ops->ops.size_ib_cq = IRDMA_SET_RDMA_OBJ_SIZE(ib_cq, irdma_cq, ibcq);
3538         dev_ops->ops.size_ib_pd = IRDMA_SET_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd);
3539         dev_ops->ops.size_ib_ucontext = IRDMA_SET_RDMA_OBJ_SIZE(ib_ucontext,
3540                                                                 irdma_ucontext,
3541                                                                 ibucontext);
3542
3543 #endif                          /* __FreeBSD_version >= 1400000 */
3544         dev_ops->alloc_hw_stats = irdma_alloc_hw_stats;
3545         dev_ops->alloc_mr = irdma_alloc_mr;
3546         dev_ops->alloc_mw = irdma_alloc_mw;
3547         dev_ops->alloc_pd = irdma_alloc_pd;
3548         dev_ops->alloc_ucontext = irdma_alloc_ucontext;
3549         dev_ops->create_cq = irdma_create_cq;
3550         dev_ops->create_qp = irdma_create_qp;
3551         dev_ops->dealloc_mw = irdma_dealloc_mw;
3552         dev_ops->dealloc_pd = irdma_dealloc_pd;
3553         dev_ops->dealloc_ucontext = irdma_dealloc_ucontext;
3554         dev_ops->dereg_mr = irdma_dereg_mr;
3555         dev_ops->destroy_cq = irdma_destroy_cq;
3556         dev_ops->destroy_qp = irdma_destroy_qp;
3557         dev_ops->disassociate_ucontext = irdma_disassociate_ucontext;
3558         dev_ops->get_dev_fw_str = irdma_get_dev_fw_str;
3559         dev_ops->get_dma_mr = irdma_get_dma_mr;
3560         dev_ops->get_hw_stats = irdma_get_hw_stats;
3561         dev_ops->get_netdev = irdma_get_netdev;
3562         dev_ops->map_mr_sg = irdma_map_mr_sg;
3563         dev_ops->mmap = irdma_mmap;
3564 #if __FreeBSD_version >= 1400026
3565         dev_ops->mmap_free = irdma_mmap_free;
3566 #endif
3567         dev_ops->poll_cq = irdma_poll_cq;
3568         dev_ops->post_recv = irdma_post_recv;
3569         dev_ops->post_send = irdma_post_send;
3570         dev_ops->query_device = irdma_query_device;
3571         dev_ops->query_port = irdma_query_port;
3572         dev_ops->modify_port = irdma_modify_port;
3573         dev_ops->query_qp = irdma_query_qp;
3574         dev_ops->reg_user_mr = irdma_reg_user_mr;
3575         dev_ops->rereg_user_mr = irdma_rereg_user_mr;
3576         dev_ops->req_notify_cq = irdma_req_notify_cq;
3577         dev_ops->resize_cq = irdma_resize_cq;
3578 }
3579
3580 static void
3581 irdma_set_device_mcast_ops(struct ib_device *ibdev)
3582 {
3583         struct ib_device *dev_ops = ibdev;
3584         dev_ops->attach_mcast = irdma_attach_mcast;
3585         dev_ops->detach_mcast = irdma_detach_mcast;
3586 }
3587
3588 static void
3589 irdma_set_device_roce_ops(struct ib_device *ibdev)
3590 {
3591         struct ib_device *dev_ops = ibdev;
3592         dev_ops->create_ah = irdma_create_ah;
3593         dev_ops->destroy_ah = irdma_destroy_ah;
3594         dev_ops->get_link_layer = irdma_get_link_layer;
3595         dev_ops->get_port_immutable = irdma_roce_port_immutable;
3596         dev_ops->modify_qp = irdma_modify_qp_roce;
3597         dev_ops->query_ah = irdma_query_ah;
3598         dev_ops->query_gid = irdma_query_gid_roce;
3599         dev_ops->query_pkey = irdma_query_pkey;
3600         ibdev->add_gid = irdma_add_gid;
3601         ibdev->del_gid = irdma_del_gid;
3602 }
3603
3604 static void
3605 irdma_set_device_iw_ops(struct ib_device *ibdev)
3606 {
3607         struct ib_device *dev_ops = ibdev;
3608
3609         ibdev->uverbs_cmd_mask |=
3610             (1ull << IB_USER_VERBS_CMD_CREATE_AH) |
3611             (1ull << IB_USER_VERBS_CMD_DESTROY_AH);
3612
3613         dev_ops->create_ah = irdma_create_ah_stub;
3614         dev_ops->destroy_ah = irdma_destroy_ah_stub;
3615         dev_ops->get_port_immutable = irdma_iw_port_immutable;
3616         dev_ops->modify_qp = irdma_modify_qp;
3617         dev_ops->query_gid = irdma_query_gid;
3618         dev_ops->query_pkey = irdma_iw_query_pkey;
3619 }
3620
3621 static inline void
3622 irdma_set_device_gen1_ops(struct ib_device *ibdev)
3623 {
3624 }
3625
3626 /**
3627  * irdma_init_roce_device - initialization of roce rdma device
3628  * @iwdev: irdma device
3629  */
3630 static void
3631 irdma_init_roce_device(struct irdma_device *iwdev)
3632 {
3633         kc_set_roce_uverbs_cmd_mask(iwdev);
3634         iwdev->ibdev.node_type = RDMA_NODE_IB_CA;
3635         iwdev->ibdev.node_guid = irdma_mac_to_guid(iwdev->netdev);
3636         irdma_set_device_roce_ops(&iwdev->ibdev);
3637         if (iwdev->rf->rdma_ver == IRDMA_GEN_2)
3638                 irdma_set_device_mcast_ops(&iwdev->ibdev);
3639 }
3640
3641 /**
3642  * irdma_init_iw_device - initialization of iwarp rdma device
3643  * @iwdev: irdma device
3644  */
3645 static int
3646 irdma_init_iw_device(struct irdma_device *iwdev)
3647 {
3648         struct ifnet *netdev = iwdev->netdev;
3649
3650         iwdev->ibdev.node_type = RDMA_NODE_RNIC;
3651         ether_addr_copy((u8 *)&iwdev->ibdev.node_guid, IF_LLADDR(netdev));
3652         iwdev->ibdev.iwcm = kzalloc(sizeof(*iwdev->ibdev.iwcm), GFP_KERNEL);
3653         if (!iwdev->ibdev.iwcm)
3654                 return -ENOMEM;
3655
3656         iwdev->ibdev.iwcm->add_ref = irdma_qp_add_ref;
3657         iwdev->ibdev.iwcm->rem_ref = irdma_qp_rem_ref;
3658         iwdev->ibdev.iwcm->get_qp = irdma_get_qp;
3659         iwdev->ibdev.iwcm->connect = irdma_connect;
3660         iwdev->ibdev.iwcm->accept = irdma_accept;
3661         iwdev->ibdev.iwcm->reject = irdma_reject;
3662         iwdev->ibdev.iwcm->create_listen = irdma_create_listen;
3663         iwdev->ibdev.iwcm->destroy_listen = irdma_destroy_listen;
3664         memcpy(iwdev->ibdev.iwcm->ifname, if_name(netdev),
3665                sizeof(iwdev->ibdev.iwcm->ifname));
3666         irdma_set_device_iw_ops(&iwdev->ibdev);
3667
3668         return 0;
3669 }
3670
3671 /**
3672  * irdma_init_rdma_device - initialization of rdma device
3673  * @iwdev: irdma device
3674  */
3675 static int
3676 irdma_init_rdma_device(struct irdma_device *iwdev)
3677 {
3678         struct pci_dev *pcidev = iwdev->rf->pcidev;
3679         int ret;
3680
3681         iwdev->ibdev.owner = THIS_MODULE;
3682         iwdev->ibdev.uverbs_abi_ver = IRDMA_ABI_VER;
3683         kc_set_rdma_uverbs_cmd_mask(iwdev);
3684
3685         if (iwdev->roce_mode) {
3686                 irdma_init_roce_device(iwdev);
3687         } else {
3688                 ret = irdma_init_iw_device(iwdev);
3689                 if (ret)
3690                         return ret;
3691         }
3692
3693         iwdev->ibdev.phys_port_cnt = 1;
3694         iwdev->ibdev.num_comp_vectors = iwdev->rf->ceqs_count;
3695         iwdev->ibdev.dev.parent = iwdev->rf->dev_ctx.dev;
3696         set_ibdev_dma_device(iwdev->ibdev, &pcidev->dev);
3697         irdma_set_device_ops(&iwdev->ibdev);
3698         if (iwdev->rf->rdma_ver == IRDMA_GEN_1)
3699                 irdma_set_device_gen1_ops(&iwdev->ibdev);
3700
3701         return 0;
3702 }
3703
3704 /**
3705  * irdma_port_ibevent - indicate port event
3706  * @iwdev: irdma device
3707  */
3708 void
3709 irdma_port_ibevent(struct irdma_device *iwdev)
3710 {
3711         struct ib_event event;
3712
3713         event.device = &iwdev->ibdev;
3714         event.element.port_num = 1;
3715         event.event =
3716             iwdev->iw_status ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
3717         ib_dispatch_event(&event);
3718 }
3719
3720 /**
3721  * irdma_ib_unregister_device - unregister rdma device from IB
3722  * core
3723  * @iwdev: irdma device
3724  */
3725 void
3726 irdma_ib_unregister_device(struct irdma_device *iwdev)
3727 {
3728         iwdev->iw_status = 0;
3729         irdma_port_ibevent(iwdev);
3730         ib_unregister_device(&iwdev->ibdev);
3731         dev_put(iwdev->netdev);
3732         kfree(iwdev->ibdev.iwcm);
3733         iwdev->ibdev.iwcm = NULL;
3734 }
3735
3736 /**
3737  * irdma_ib_register_device - register irdma device to IB core
3738  * @iwdev: irdma device
3739  */
3740 int
3741 irdma_ib_register_device(struct irdma_device *iwdev)
3742 {
3743         int ret;
3744
3745         ret = irdma_init_rdma_device(iwdev);
3746         if (ret)
3747                 return ret;
3748
3749         dev_hold(iwdev->netdev);
3750         sprintf(iwdev->ibdev.name, "irdma-%s", if_name(iwdev->netdev));
3751         ret = ib_register_device(&iwdev->ibdev, NULL);
3752         if (ret)
3753                 goto error;
3754
3755         iwdev->iw_status = 1;
3756         irdma_port_ibevent(iwdev);
3757
3758         return 0;
3759
3760 error:
3761         kfree(iwdev->ibdev.iwcm);
3762         iwdev->ibdev.iwcm = NULL;
3763         irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS, "Register RDMA device fail\n");
3764
3765         return ret;
3766 }