GNU Linux-libre 4.19.281-gnu1
[releases.git] / drivers / infiniband / hw / cxgb3 / iwch_qp.c
1 /*
2  * Copyright (c) 2006 Chelsio, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 #include <linux/sched.h>
33 #include <linux/gfp.h>
34 #include "iwch_provider.h"
35 #include "iwch.h"
36 #include "iwch_cm.h"
37 #include "cxio_hal.h"
38 #include "cxio_resource.h"
39
40 #define NO_SUPPORT -1
41
42 static int build_rdma_send(union t3_wr *wqe, const struct ib_send_wr *wr,
43                            u8 *flit_cnt)
44 {
45         int i;
46         u32 plen;
47
48         switch (wr->opcode) {
49         case IB_WR_SEND:
50                 if (wr->send_flags & IB_SEND_SOLICITED)
51                         wqe->send.rdmaop = T3_SEND_WITH_SE;
52                 else
53                         wqe->send.rdmaop = T3_SEND;
54                 wqe->send.rem_stag = 0;
55                 break;
56         case IB_WR_SEND_WITH_INV:
57                 if (wr->send_flags & IB_SEND_SOLICITED)
58                         wqe->send.rdmaop = T3_SEND_WITH_SE_INV;
59                 else
60                         wqe->send.rdmaop = T3_SEND_WITH_INV;
61                 wqe->send.rem_stag = cpu_to_be32(wr->ex.invalidate_rkey);
62                 break;
63         default:
64                 return -EINVAL;
65         }
66         if (wr->num_sge > T3_MAX_SGE)
67                 return -EINVAL;
68         wqe->send.reserved[0] = 0;
69         wqe->send.reserved[1] = 0;
70         wqe->send.reserved[2] = 0;
71         plen = 0;
72         for (i = 0; i < wr->num_sge; i++) {
73                 if ((plen + wr->sg_list[i].length) < plen)
74                         return -EMSGSIZE;
75
76                 plen += wr->sg_list[i].length;
77                 wqe->send.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
78                 wqe->send.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
79                 wqe->send.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
80         }
81         wqe->send.num_sgle = cpu_to_be32(wr->num_sge);
82         *flit_cnt = 4 + ((wr->num_sge) << 1);
83         wqe->send.plen = cpu_to_be32(plen);
84         return 0;
85 }
86
87 static int build_rdma_write(union t3_wr *wqe, const struct ib_send_wr *wr,
88                             u8 *flit_cnt)
89 {
90         int i;
91         u32 plen;
92         if (wr->num_sge > T3_MAX_SGE)
93                 return -EINVAL;
94         wqe->write.rdmaop = T3_RDMA_WRITE;
95         wqe->write.reserved[0] = 0;
96         wqe->write.reserved[1] = 0;
97         wqe->write.reserved[2] = 0;
98         wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
99         wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
100
101         if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
102                 plen = 4;
103                 wqe->write.sgl[0].stag = wr->ex.imm_data;
104                 wqe->write.sgl[0].len = cpu_to_be32(0);
105                 wqe->write.num_sgle = cpu_to_be32(0);
106                 *flit_cnt = 6;
107         } else {
108                 plen = 0;
109                 for (i = 0; i < wr->num_sge; i++) {
110                         if ((plen + wr->sg_list[i].length) < plen) {
111                                 return -EMSGSIZE;
112                         }
113                         plen += wr->sg_list[i].length;
114                         wqe->write.sgl[i].stag =
115                             cpu_to_be32(wr->sg_list[i].lkey);
116                         wqe->write.sgl[i].len =
117                             cpu_to_be32(wr->sg_list[i].length);
118                         wqe->write.sgl[i].to =
119                             cpu_to_be64(wr->sg_list[i].addr);
120                 }
121                 wqe->write.num_sgle = cpu_to_be32(wr->num_sge);
122                 *flit_cnt = 5 + ((wr->num_sge) << 1);
123         }
124         wqe->write.plen = cpu_to_be32(plen);
125         return 0;
126 }
127
128 static int build_rdma_read(union t3_wr *wqe, const struct ib_send_wr *wr,
129                            u8 *flit_cnt)
130 {
131         if (wr->num_sge > 1)
132                 return -EINVAL;
133         wqe->read.rdmaop = T3_READ_REQ;
134         if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
135                 wqe->read.local_inv = 1;
136         else
137                 wqe->read.local_inv = 0;
138         wqe->read.reserved[0] = 0;
139         wqe->read.reserved[1] = 0;
140         wqe->read.rem_stag = cpu_to_be32(rdma_wr(wr)->rkey);
141         wqe->read.rem_to = cpu_to_be64(rdma_wr(wr)->remote_addr);
142         wqe->read.local_stag = cpu_to_be32(wr->sg_list[0].lkey);
143         wqe->read.local_len = cpu_to_be32(wr->sg_list[0].length);
144         wqe->read.local_to = cpu_to_be64(wr->sg_list[0].addr);
145         *flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
146         return 0;
147 }
148
149 static int build_memreg(union t3_wr *wqe, const struct ib_reg_wr *wr,
150                         u8 *flit_cnt, int *wr_cnt, struct t3_wq *wq)
151 {
152         struct iwch_mr *mhp = to_iwch_mr(wr->mr);
153         int i;
154         __be64 *p;
155
156         if (mhp->npages > T3_MAX_FASTREG_DEPTH)
157                 return -EINVAL;
158         *wr_cnt = 1;
159         wqe->fastreg.stag = cpu_to_be32(wr->key);
160         wqe->fastreg.len = cpu_to_be32(mhp->ibmr.length);
161         wqe->fastreg.va_base_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
162         wqe->fastreg.va_base_lo_fbo =
163                                 cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
164         wqe->fastreg.page_type_perms = cpu_to_be32(
165                 V_FR_PAGE_COUNT(mhp->npages) |
166                 V_FR_PAGE_SIZE(ilog2(wr->mr->page_size) - 12) |
167                 V_FR_TYPE(TPT_VATO) |
168                 V_FR_PERMS(iwch_ib_to_tpt_access(wr->access)));
169         p = &wqe->fastreg.pbl_addrs[0];
170         for (i = 0; i < mhp->npages; i++, p++) {
171
172                 /* If we need a 2nd WR, then set it up */
173                 if (i == T3_MAX_FASTREG_FRAG) {
174                         *wr_cnt = 2;
175                         wqe = (union t3_wr *)(wq->queue +
176                                 Q_PTR2IDX((wq->wptr+1), wq->size_log2));
177                         build_fw_riwrh((void *)wqe, T3_WR_FASTREG, 0,
178                                Q_GENBIT(wq->wptr + 1, wq->size_log2),
179                                0, 1 + mhp->npages - T3_MAX_FASTREG_FRAG,
180                                T3_EOP);
181
182                         p = &wqe->pbl_frag.pbl_addrs[0];
183                 }
184                 *p = cpu_to_be64((u64)mhp->pages[i]);
185         }
186         *flit_cnt = 5 + mhp->npages;
187         if (*flit_cnt > 15)
188                 *flit_cnt = 15;
189         return 0;
190 }
191
192 static int build_inv_stag(union t3_wr *wqe, const struct ib_send_wr *wr,
193                           u8 *flit_cnt)
194 {
195         wqe->local_inv.stag = cpu_to_be32(wr->ex.invalidate_rkey);
196         wqe->local_inv.reserved = 0;
197         *flit_cnt = sizeof(struct t3_local_inv_wr) >> 3;
198         return 0;
199 }
200
201 static int iwch_sgl2pbl_map(struct iwch_dev *rhp, struct ib_sge *sg_list,
202                             u32 num_sgle, u32 * pbl_addr, u8 * page_size)
203 {
204         int i;
205         struct iwch_mr *mhp;
206         u64 offset;
207         for (i = 0; i < num_sgle; i++) {
208
209                 mhp = get_mhp(rhp, (sg_list[i].lkey) >> 8);
210                 if (!mhp) {
211                         pr_debug("%s %d\n", __func__, __LINE__);
212                         return -EIO;
213                 }
214                 if (!mhp->attr.state) {
215                         pr_debug("%s %d\n", __func__, __LINE__);
216                         return -EIO;
217                 }
218                 if (mhp->attr.zbva) {
219                         pr_debug("%s %d\n", __func__, __LINE__);
220                         return -EIO;
221                 }
222
223                 if (sg_list[i].addr < mhp->attr.va_fbo) {
224                         pr_debug("%s %d\n", __func__, __LINE__);
225                         return -EINVAL;
226                 }
227                 if (sg_list[i].addr + ((u64) sg_list[i].length) <
228                     sg_list[i].addr) {
229                         pr_debug("%s %d\n", __func__, __LINE__);
230                         return -EINVAL;
231                 }
232                 if (sg_list[i].addr + ((u64) sg_list[i].length) >
233                     mhp->attr.va_fbo + ((u64) mhp->attr.len)) {
234                         pr_debug("%s %d\n", __func__, __LINE__);
235                         return -EINVAL;
236                 }
237                 offset = sg_list[i].addr - mhp->attr.va_fbo;
238                 offset += mhp->attr.va_fbo &
239                           ((1UL << (12 + mhp->attr.page_size)) - 1);
240                 pbl_addr[i] = ((mhp->attr.pbl_addr -
241                                 rhp->rdev.rnic_info.pbl_base) >> 3) +
242                               (offset >> (12 + mhp->attr.page_size));
243                 page_size[i] = mhp->attr.page_size;
244         }
245         return 0;
246 }
247
248 static int build_rdma_recv(struct iwch_qp *qhp, union t3_wr *wqe,
249                            const struct ib_recv_wr *wr)
250 {
251         int i, err = 0;
252         u32 pbl_addr[T3_MAX_SGE];
253         u8 page_size[T3_MAX_SGE];
254
255         err = iwch_sgl2pbl_map(qhp->rhp, wr->sg_list, wr->num_sge, pbl_addr,
256                                page_size);
257         if (err)
258                 return err;
259         wqe->recv.pagesz[0] = page_size[0];
260         wqe->recv.pagesz[1] = page_size[1];
261         wqe->recv.pagesz[2] = page_size[2];
262         wqe->recv.pagesz[3] = page_size[3];
263         wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
264         for (i = 0; i < wr->num_sge; i++) {
265                 wqe->recv.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
266                 wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
267
268                 /* to in the WQE == the offset into the page */
269                 wqe->recv.sgl[i].to = cpu_to_be64(((u32)wr->sg_list[i].addr) &
270                                 ((1UL << (12 + page_size[i])) - 1));
271
272                 /* pbl_addr is the adapters address in the PBL */
273                 wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_addr[i]);
274         }
275         for (; i < T3_MAX_SGE; i++) {
276                 wqe->recv.sgl[i].stag = 0;
277                 wqe->recv.sgl[i].len = 0;
278                 wqe->recv.sgl[i].to = 0;
279                 wqe->recv.pbl_addr[i] = 0;
280         }
281         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
282                              qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
283         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
284                              qhp->wq.rq_size_log2)].pbl_addr = 0;
285         return 0;
286 }
287
288 static int build_zero_stag_recv(struct iwch_qp *qhp, union t3_wr *wqe,
289                                 const struct ib_recv_wr *wr)
290 {
291         int i;
292         u32 pbl_addr;
293         u32 pbl_offset;
294
295
296         /*
297          * The T3 HW requires the PBL in the HW recv descriptor to reference
298          * a PBL entry.  So we allocate the max needed PBL memory here and pass
299          * it to the uP in the recv WR.  The uP will build the PBL and setup
300          * the HW recv descriptor.
301          */
302         pbl_addr = cxio_hal_pblpool_alloc(&qhp->rhp->rdev, T3_STAG0_PBL_SIZE);
303         if (!pbl_addr)
304                 return -ENOMEM;
305
306         /*
307          * Compute the 8B aligned offset.
308          */
309         pbl_offset = (pbl_addr - qhp->rhp->rdev.rnic_info.pbl_base) >> 3;
310
311         wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
312
313         for (i = 0; i < wr->num_sge; i++) {
314
315                 /*
316                  * Use a 128MB page size. This and an imposed 128MB
317                  * sge length limit allows us to require only a 2-entry HW
318                  * PBL for each SGE.  This restriction is acceptable since
319                  * since it is not possible to allocate 128MB of contiguous
320                  * DMA coherent memory!
321                  */
322                 if (wr->sg_list[i].length > T3_STAG0_MAX_PBE_LEN)
323                         return -EINVAL;
324                 wqe->recv.pagesz[i] = T3_STAG0_PAGE_SHIFT;
325
326                 /*
327                  * T3 restricts a recv to all zero-stag or all non-zero-stag.
328                  */
329                 if (wr->sg_list[i].lkey != 0)
330                         return -EINVAL;
331                 wqe->recv.sgl[i].stag = 0;
332                 wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
333                 wqe->recv.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
334                 wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_offset);
335                 pbl_offset += 2;
336         }
337         for (; i < T3_MAX_SGE; i++) {
338                 wqe->recv.pagesz[i] = 0;
339                 wqe->recv.sgl[i].stag = 0;
340                 wqe->recv.sgl[i].len = 0;
341                 wqe->recv.sgl[i].to = 0;
342                 wqe->recv.pbl_addr[i] = 0;
343         }
344         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
345                              qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
346         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
347                              qhp->wq.rq_size_log2)].pbl_addr = pbl_addr;
348         return 0;
349 }
350
351 int iwch_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
352                    const struct ib_send_wr **bad_wr)
353 {
354         int err = 0;
355         u8 uninitialized_var(t3_wr_flit_cnt);
356         enum t3_wr_opcode t3_wr_opcode = 0;
357         enum t3_wr_flags t3_wr_flags;
358         struct iwch_qp *qhp;
359         u32 idx;
360         union t3_wr *wqe;
361         u32 num_wrs;
362         unsigned long flag;
363         struct t3_swsq *sqp;
364         int wr_cnt = 1;
365
366         qhp = to_iwch_qp(ibqp);
367         spin_lock_irqsave(&qhp->lock, flag);
368         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
369                 spin_unlock_irqrestore(&qhp->lock, flag);
370                 err = -EINVAL;
371                 goto out;
372         }
373         num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
374                   qhp->wq.sq_size_log2);
375         if (num_wrs == 0) {
376                 spin_unlock_irqrestore(&qhp->lock, flag);
377                 err = -ENOMEM;
378                 goto out;
379         }
380         while (wr) {
381                 if (num_wrs == 0) {
382                         err = -ENOMEM;
383                         break;
384                 }
385                 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
386                 wqe = (union t3_wr *) (qhp->wq.queue + idx);
387                 t3_wr_flags = 0;
388                 if (wr->send_flags & IB_SEND_SOLICITED)
389                         t3_wr_flags |= T3_SOLICITED_EVENT_FLAG;
390                 if (wr->send_flags & IB_SEND_SIGNALED)
391                         t3_wr_flags |= T3_COMPLETION_FLAG;
392                 sqp = qhp->wq.sq +
393                       Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
394                 switch (wr->opcode) {
395                 case IB_WR_SEND:
396                 case IB_WR_SEND_WITH_INV:
397                         if (wr->send_flags & IB_SEND_FENCE)
398                                 t3_wr_flags |= T3_READ_FENCE_FLAG;
399                         t3_wr_opcode = T3_WR_SEND;
400                         err = build_rdma_send(wqe, wr, &t3_wr_flit_cnt);
401                         break;
402                 case IB_WR_RDMA_WRITE:
403                 case IB_WR_RDMA_WRITE_WITH_IMM:
404                         t3_wr_opcode = T3_WR_WRITE;
405                         err = build_rdma_write(wqe, wr, &t3_wr_flit_cnt);
406                         break;
407                 case IB_WR_RDMA_READ:
408                 case IB_WR_RDMA_READ_WITH_INV:
409                         t3_wr_opcode = T3_WR_READ;
410                         t3_wr_flags = 0; /* T3 reads are always signaled */
411                         err = build_rdma_read(wqe, wr, &t3_wr_flit_cnt);
412                         if (err)
413                                 break;
414                         sqp->read_len = wqe->read.local_len;
415                         if (!qhp->wq.oldest_read)
416                                 qhp->wq.oldest_read = sqp;
417                         break;
418                 case IB_WR_REG_MR:
419                         t3_wr_opcode = T3_WR_FASTREG;
420                         err = build_memreg(wqe, reg_wr(wr), &t3_wr_flit_cnt,
421                                            &wr_cnt, &qhp->wq);
422                         break;
423                 case IB_WR_LOCAL_INV:
424                         if (wr->send_flags & IB_SEND_FENCE)
425                                 t3_wr_flags |= T3_LOCAL_FENCE_FLAG;
426                         t3_wr_opcode = T3_WR_INV_STAG;
427                         err = build_inv_stag(wqe, wr, &t3_wr_flit_cnt);
428                         break;
429                 default:
430                         pr_debug("%s post of type=%d TBD!\n", __func__,
431                                  wr->opcode);
432                         err = -EINVAL;
433                 }
434                 if (err)
435                         break;
436                 wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
437                 sqp->wr_id = wr->wr_id;
438                 sqp->opcode = wr2opcode(t3_wr_opcode);
439                 sqp->sq_wptr = qhp->wq.sq_wptr;
440                 sqp->complete = 0;
441                 sqp->signaled = (wr->send_flags & IB_SEND_SIGNALED);
442
443                 build_fw_riwrh((void *) wqe, t3_wr_opcode, t3_wr_flags,
444                                Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
445                                0, t3_wr_flit_cnt,
446                                (wr_cnt == 1) ? T3_SOPEOP : T3_SOP);
447                 pr_debug("%s cookie 0x%llx wq idx 0x%x swsq idx %ld opcode %d\n",
448                          __func__, (unsigned long long)wr->wr_id, idx,
449                          Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2),
450                          sqp->opcode);
451                 wr = wr->next;
452                 num_wrs--;
453                 qhp->wq.wptr += wr_cnt;
454                 ++(qhp->wq.sq_wptr);
455         }
456         spin_unlock_irqrestore(&qhp->lock, flag);
457         if (cxio_wq_db_enabled(&qhp->wq))
458                 ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
459
460 out:
461         if (err)
462                 *bad_wr = wr;
463         return err;
464 }
465
466 int iwch_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
467                       const struct ib_recv_wr **bad_wr)
468 {
469         int err = 0;
470         struct iwch_qp *qhp;
471         u32 idx;
472         union t3_wr *wqe;
473         u32 num_wrs;
474         unsigned long flag;
475
476         qhp = to_iwch_qp(ibqp);
477         spin_lock_irqsave(&qhp->lock, flag);
478         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
479                 spin_unlock_irqrestore(&qhp->lock, flag);
480                 err = -EINVAL;
481                 goto out;
482         }
483         num_wrs = Q_FREECNT(qhp->wq.rq_rptr, qhp->wq.rq_wptr,
484                             qhp->wq.rq_size_log2) - 1;
485         if (!wr) {
486                 spin_unlock_irqrestore(&qhp->lock, flag);
487                 err = -ENOMEM;
488                 goto out;
489         }
490         while (wr) {
491                 if (wr->num_sge > T3_MAX_SGE) {
492                         err = -EINVAL;
493                         break;
494                 }
495                 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
496                 wqe = (union t3_wr *) (qhp->wq.queue + idx);
497                 if (num_wrs)
498                         if (wr->sg_list[0].lkey)
499                                 err = build_rdma_recv(qhp, wqe, wr);
500                         else
501                                 err = build_zero_stag_recv(qhp, wqe, wr);
502                 else
503                         err = -ENOMEM;
504
505                 if (err)
506                         break;
507
508                 build_fw_riwrh((void *) wqe, T3_WR_RCV, T3_COMPLETION_FLAG,
509                                Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
510                                0, sizeof(struct t3_receive_wr) >> 3, T3_SOPEOP);
511                 pr_debug("%s cookie 0x%llx idx 0x%x rq_wptr 0x%x rw_rptr 0x%x wqe %p\n",
512                          __func__, (unsigned long long)wr->wr_id,
513                          idx, qhp->wq.rq_wptr, qhp->wq.rq_rptr, wqe);
514                 ++(qhp->wq.rq_wptr);
515                 ++(qhp->wq.wptr);
516                 wr = wr->next;
517                 num_wrs--;
518         }
519         spin_unlock_irqrestore(&qhp->lock, flag);
520         if (cxio_wq_db_enabled(&qhp->wq))
521                 ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
522
523 out:
524         if (err)
525                 *bad_wr = wr;
526         return err;
527 }
528
529 static inline void build_term_codes(struct respQ_msg_t *rsp_msg,
530                                     u8 *layer_type, u8 *ecode)
531 {
532         int status = TPT_ERR_INTERNAL_ERR;
533         int tagged = 0;
534         int opcode = -1;
535         int rqtype = 0;
536         int send_inv = 0;
537
538         if (rsp_msg) {
539                 status = CQE_STATUS(rsp_msg->cqe);
540                 opcode = CQE_OPCODE(rsp_msg->cqe);
541                 rqtype = RQ_TYPE(rsp_msg->cqe);
542                 send_inv = (opcode == T3_SEND_WITH_INV) ||
543                            (opcode == T3_SEND_WITH_SE_INV);
544                 tagged = (opcode == T3_RDMA_WRITE) ||
545                          (rqtype && (opcode == T3_READ_RESP));
546         }
547
548         switch (status) {
549         case TPT_ERR_STAG:
550                 if (send_inv) {
551                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
552                         *ecode = RDMAP_CANT_INV_STAG;
553                 } else {
554                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
555                         *ecode = RDMAP_INV_STAG;
556                 }
557                 break;
558         case TPT_ERR_PDID:
559                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
560                 if ((opcode == T3_SEND_WITH_INV) ||
561                     (opcode == T3_SEND_WITH_SE_INV))
562                         *ecode = RDMAP_CANT_INV_STAG;
563                 else
564                         *ecode = RDMAP_STAG_NOT_ASSOC;
565                 break;
566         case TPT_ERR_QPID:
567                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
568                 *ecode = RDMAP_STAG_NOT_ASSOC;
569                 break;
570         case TPT_ERR_ACCESS:
571                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
572                 *ecode = RDMAP_ACC_VIOL;
573                 break;
574         case TPT_ERR_WRAP:
575                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
576                 *ecode = RDMAP_TO_WRAP;
577                 break;
578         case TPT_ERR_BOUND:
579                 if (tagged) {
580                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
581                         *ecode = DDPT_BASE_BOUNDS;
582                 } else {
583                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
584                         *ecode = RDMAP_BASE_BOUNDS;
585                 }
586                 break;
587         case TPT_ERR_INVALIDATE_SHARED_MR:
588         case TPT_ERR_INVALIDATE_MR_WITH_MW_BOUND:
589                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
590                 *ecode = RDMAP_CANT_INV_STAG;
591                 break;
592         case TPT_ERR_ECC:
593         case TPT_ERR_ECC_PSTAG:
594         case TPT_ERR_INTERNAL_ERR:
595                 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
596                 *ecode = 0;
597                 break;
598         case TPT_ERR_OUT_OF_RQE:
599                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
600                 *ecode = DDPU_INV_MSN_NOBUF;
601                 break;
602         case TPT_ERR_PBL_ADDR_BOUND:
603                 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
604                 *ecode = DDPT_BASE_BOUNDS;
605                 break;
606         case TPT_ERR_CRC:
607                 *layer_type = LAYER_MPA|DDP_LLP;
608                 *ecode = MPA_CRC_ERR;
609                 break;
610         case TPT_ERR_MARKER:
611                 *layer_type = LAYER_MPA|DDP_LLP;
612                 *ecode = MPA_MARKER_ERR;
613                 break;
614         case TPT_ERR_PDU_LEN_ERR:
615                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
616                 *ecode = DDPU_MSG_TOOBIG;
617                 break;
618         case TPT_ERR_DDP_VERSION:
619                 if (tagged) {
620                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
621                         *ecode = DDPT_INV_VERS;
622                 } else {
623                         *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
624                         *ecode = DDPU_INV_VERS;
625                 }
626                 break;
627         case TPT_ERR_RDMA_VERSION:
628                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
629                 *ecode = RDMAP_INV_VERS;
630                 break;
631         case TPT_ERR_OPCODE:
632                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
633                 *ecode = RDMAP_INV_OPCODE;
634                 break;
635         case TPT_ERR_DDP_QUEUE_NUM:
636                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
637                 *ecode = DDPU_INV_QN;
638                 break;
639         case TPT_ERR_MSN:
640         case TPT_ERR_MSN_GAP:
641         case TPT_ERR_MSN_RANGE:
642         case TPT_ERR_IRD_OVERFLOW:
643                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
644                 *ecode = DDPU_INV_MSN_RANGE;
645                 break;
646         case TPT_ERR_TBIT:
647                 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
648                 *ecode = 0;
649                 break;
650         case TPT_ERR_MO:
651                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
652                 *ecode = DDPU_INV_MO;
653                 break;
654         default:
655                 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
656                 *ecode = 0;
657                 break;
658         }
659 }
660
661 int iwch_post_zb_read(struct iwch_ep *ep)
662 {
663         union t3_wr *wqe;
664         struct sk_buff *skb;
665         u8 flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
666
667         pr_debug("%s enter\n", __func__);
668         skb = alloc_skb(40, GFP_KERNEL);
669         if (!skb) {
670                 pr_err("%s cannot send zb_read!!\n", __func__);
671                 return -ENOMEM;
672         }
673         wqe = skb_put_zero(skb, sizeof(struct t3_rdma_read_wr));
674         wqe->read.rdmaop = T3_READ_REQ;
675         wqe->read.reserved[0] = 0;
676         wqe->read.reserved[1] = 0;
677         wqe->read.rem_stag = cpu_to_be32(1);
678         wqe->read.rem_to = cpu_to_be64(1);
679         wqe->read.local_stag = cpu_to_be32(1);
680         wqe->read.local_len = cpu_to_be32(0);
681         wqe->read.local_to = cpu_to_be64(1);
682         wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_READ));
683         wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(ep->hwtid)|
684                                                 V_FW_RIWR_LEN(flit_cnt));
685         skb->priority = CPL_PRIORITY_DATA;
686         return iwch_cxgb3_ofld_send(ep->com.qp->rhp->rdev.t3cdev_p, skb);
687 }
688
689 /*
690  * This posts a TERMINATE with layer=RDMA, type=catastrophic.
691  */
692 int iwch_post_terminate(struct iwch_qp *qhp, struct respQ_msg_t *rsp_msg)
693 {
694         union t3_wr *wqe;
695         struct terminate_message *term;
696         struct sk_buff *skb;
697
698         pr_debug("%s %d\n", __func__, __LINE__);
699         skb = alloc_skb(40, GFP_ATOMIC);
700         if (!skb) {
701                 pr_err("%s cannot send TERMINATE!\n", __func__);
702                 return -ENOMEM;
703         }
704         wqe = skb_put_zero(skb, 40);
705         wqe->send.rdmaop = T3_TERMINATE;
706
707         /* immediate data length */
708         wqe->send.plen = htonl(4);
709
710         /* immediate data starts here. */
711         term = (struct terminate_message *)wqe->send.sgl;
712         build_term_codes(rsp_msg, &term->layer_etype, &term->ecode);
713         wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_SEND) |
714                          V_FW_RIWR_FLAGS(T3_COMPLETION_FLAG | T3_NOTIFY_FLAG));
715         wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(qhp->ep->hwtid));
716         skb->priority = CPL_PRIORITY_DATA;
717         return iwch_cxgb3_ofld_send(qhp->rhp->rdev.t3cdev_p, skb);
718 }
719
720 /*
721  * Assumes qhp lock is held.
722  */
723 static void __flush_qp(struct iwch_qp *qhp, struct iwch_cq *rchp,
724                                 struct iwch_cq *schp)
725         __releases(&qhp->lock)
726         __acquires(&qhp->lock)
727 {
728         int count;
729         int flushed;
730
731         lockdep_assert_held(&qhp->lock);
732
733         pr_debug("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
734         /* take a ref on the qhp since we must release the lock */
735         atomic_inc(&qhp->refcnt);
736         spin_unlock(&qhp->lock);
737
738         /* locking hierarchy: cq lock first, then qp lock. */
739         spin_lock(&rchp->lock);
740         spin_lock(&qhp->lock);
741         cxio_flush_hw_cq(&rchp->cq);
742         cxio_count_rcqes(&rchp->cq, &qhp->wq, &count);
743         flushed = cxio_flush_rq(&qhp->wq, &rchp->cq, count);
744         spin_unlock(&qhp->lock);
745         spin_unlock(&rchp->lock);
746         if (flushed) {
747                 spin_lock(&rchp->comp_handler_lock);
748                 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
749                 spin_unlock(&rchp->comp_handler_lock);
750         }
751
752         /* locking hierarchy: cq lock first, then qp lock. */
753         spin_lock(&schp->lock);
754         spin_lock(&qhp->lock);
755         cxio_flush_hw_cq(&schp->cq);
756         cxio_count_scqes(&schp->cq, &qhp->wq, &count);
757         flushed = cxio_flush_sq(&qhp->wq, &schp->cq, count);
758         spin_unlock(&qhp->lock);
759         spin_unlock(&schp->lock);
760         if (flushed) {
761                 spin_lock(&schp->comp_handler_lock);
762                 (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
763                 spin_unlock(&schp->comp_handler_lock);
764         }
765
766         /* deref */
767         if (atomic_dec_and_test(&qhp->refcnt))
768                 wake_up(&qhp->wait);
769
770         spin_lock(&qhp->lock);
771 }
772
773 static void flush_qp(struct iwch_qp *qhp)
774 {
775         struct iwch_cq *rchp, *schp;
776
777         rchp = get_chp(qhp->rhp, qhp->attr.rcq);
778         schp = get_chp(qhp->rhp, qhp->attr.scq);
779
780         if (qhp->ibqp.uobject) {
781                 cxio_set_wq_in_error(&qhp->wq);
782                 cxio_set_cq_in_error(&rchp->cq);
783                 spin_lock(&rchp->comp_handler_lock);
784                 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
785                 spin_unlock(&rchp->comp_handler_lock);
786                 if (schp != rchp) {
787                         cxio_set_cq_in_error(&schp->cq);
788                         spin_lock(&schp->comp_handler_lock);
789                         (*schp->ibcq.comp_handler)(&schp->ibcq,
790                                                    schp->ibcq.cq_context);
791                         spin_unlock(&schp->comp_handler_lock);
792                 }
793                 return;
794         }
795         __flush_qp(qhp, rchp, schp);
796 }
797
798
799 /*
800  * Return count of RECV WRs posted
801  */
802 u16 iwch_rqes_posted(struct iwch_qp *qhp)
803 {
804         union t3_wr *wqe = qhp->wq.queue;
805         u16 count = 0;
806
807         while (count < USHRT_MAX && fw_riwrh_opcode((struct fw_riwrh *)wqe) == T3_WR_RCV) {
808                 count++;
809                 wqe++;
810         }
811         pr_debug("%s qhp %p count %u\n", __func__, qhp, count);
812         return count;
813 }
814
815 static int rdma_init(struct iwch_dev *rhp, struct iwch_qp *qhp,
816                                 enum iwch_qp_attr_mask mask,
817                                 struct iwch_qp_attributes *attrs)
818 {
819         struct t3_rdma_init_attr init_attr;
820         int ret;
821
822         init_attr.tid = qhp->ep->hwtid;
823         init_attr.qpid = qhp->wq.qpid;
824         init_attr.pdid = qhp->attr.pd;
825         init_attr.scqid = qhp->attr.scq;
826         init_attr.rcqid = qhp->attr.rcq;
827         init_attr.rq_addr = qhp->wq.rq_addr;
828         init_attr.rq_size = 1 << qhp->wq.rq_size_log2;
829         init_attr.mpaattrs = uP_RI_MPA_IETF_ENABLE |
830                 qhp->attr.mpa_attr.recv_marker_enabled |
831                 (qhp->attr.mpa_attr.xmit_marker_enabled << 1) |
832                 (qhp->attr.mpa_attr.crc_enabled << 2);
833
834         init_attr.qpcaps = uP_RI_QP_RDMA_READ_ENABLE |
835                            uP_RI_QP_RDMA_WRITE_ENABLE |
836                            uP_RI_QP_BIND_ENABLE;
837         if (!qhp->ibqp.uobject)
838                 init_attr.qpcaps |= uP_RI_QP_STAG0_ENABLE |
839                                     uP_RI_QP_FAST_REGISTER_ENABLE;
840
841         init_attr.tcp_emss = qhp->ep->emss;
842         init_attr.ord = qhp->attr.max_ord;
843         init_attr.ird = qhp->attr.max_ird;
844         init_attr.qp_dma_addr = qhp->wq.dma_addr;
845         init_attr.qp_dma_size = (1UL << qhp->wq.size_log2);
846         init_attr.rqe_count = iwch_rqes_posted(qhp);
847         init_attr.flags = qhp->attr.mpa_attr.initiator ? MPA_INITIATOR : 0;
848         init_attr.chan = qhp->ep->l2t->smt_idx;
849         if (peer2peer) {
850                 init_attr.rtr_type = RTR_READ;
851                 if (init_attr.ord == 0 && qhp->attr.mpa_attr.initiator)
852                         init_attr.ord = 1;
853                 if (init_attr.ird == 0 && !qhp->attr.mpa_attr.initiator)
854                         init_attr.ird = 1;
855         } else
856                 init_attr.rtr_type = 0;
857         init_attr.irs = qhp->ep->rcv_seq;
858         pr_debug("%s init_attr.rq_addr 0x%x init_attr.rq_size = %d flags 0x%x qpcaps 0x%x\n",
859                  __func__,
860                  init_attr.rq_addr, init_attr.rq_size,
861                  init_attr.flags, init_attr.qpcaps);
862         ret = cxio_rdma_init(&rhp->rdev, &init_attr);
863         pr_debug("%s ret %d\n", __func__, ret);
864         return ret;
865 }
866
867 int iwch_modify_qp(struct iwch_dev *rhp, struct iwch_qp *qhp,
868                                 enum iwch_qp_attr_mask mask,
869                                 struct iwch_qp_attributes *attrs,
870                                 int internal)
871 {
872         int ret = 0;
873         struct iwch_qp_attributes newattr = qhp->attr;
874         unsigned long flag;
875         int disconnect = 0;
876         int terminate = 0;
877         int abort = 0;
878         int free = 0;
879         struct iwch_ep *ep = NULL;
880
881         pr_debug("%s qhp %p qpid 0x%x ep %p state %d -> %d\n", __func__,
882                  qhp, qhp->wq.qpid, qhp->ep, qhp->attr.state,
883                  (mask & IWCH_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
884
885         spin_lock_irqsave(&qhp->lock, flag);
886
887         /* Process attr changes if in IDLE */
888         if (mask & IWCH_QP_ATTR_VALID_MODIFY) {
889                 if (qhp->attr.state != IWCH_QP_STATE_IDLE) {
890                         ret = -EIO;
891                         goto out;
892                 }
893                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_READ)
894                         newattr.enable_rdma_read = attrs->enable_rdma_read;
895                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_WRITE)
896                         newattr.enable_rdma_write = attrs->enable_rdma_write;
897                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_BIND)
898                         newattr.enable_bind = attrs->enable_bind;
899                 if (mask & IWCH_QP_ATTR_MAX_ORD) {
900                         if (attrs->max_ord >
901                             rhp->attr.max_rdma_read_qp_depth) {
902                                 ret = -EINVAL;
903                                 goto out;
904                         }
905                         newattr.max_ord = attrs->max_ord;
906                 }
907                 if (mask & IWCH_QP_ATTR_MAX_IRD) {
908                         if (attrs->max_ird >
909                             rhp->attr.max_rdma_reads_per_qp) {
910                                 ret = -EINVAL;
911                                 goto out;
912                         }
913                         newattr.max_ird = attrs->max_ird;
914                 }
915                 qhp->attr = newattr;
916         }
917
918         if (!(mask & IWCH_QP_ATTR_NEXT_STATE))
919                 goto out;
920         if (qhp->attr.state == attrs->next_state)
921                 goto out;
922
923         switch (qhp->attr.state) {
924         case IWCH_QP_STATE_IDLE:
925                 switch (attrs->next_state) {
926                 case IWCH_QP_STATE_RTS:
927                         if (!(mask & IWCH_QP_ATTR_LLP_STREAM_HANDLE)) {
928                                 ret = -EINVAL;
929                                 goto out;
930                         }
931                         if (!(mask & IWCH_QP_ATTR_MPA_ATTR)) {
932                                 ret = -EINVAL;
933                                 goto out;
934                         }
935                         qhp->attr.mpa_attr = attrs->mpa_attr;
936                         qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
937                         qhp->ep = qhp->attr.llp_stream_handle;
938                         qhp->attr.state = IWCH_QP_STATE_RTS;
939
940                         /*
941                          * Ref the endpoint here and deref when we
942                          * disassociate the endpoint from the QP.  This
943                          * happens in CLOSING->IDLE transition or *->ERROR
944                          * transition.
945                          */
946                         get_ep(&qhp->ep->com);
947                         spin_unlock_irqrestore(&qhp->lock, flag);
948                         ret = rdma_init(rhp, qhp, mask, attrs);
949                         spin_lock_irqsave(&qhp->lock, flag);
950                         if (ret)
951                                 goto err;
952                         break;
953                 case IWCH_QP_STATE_ERROR:
954                         qhp->attr.state = IWCH_QP_STATE_ERROR;
955                         flush_qp(qhp);
956                         break;
957                 default:
958                         ret = -EINVAL;
959                         goto out;
960                 }
961                 break;
962         case IWCH_QP_STATE_RTS:
963                 switch (attrs->next_state) {
964                 case IWCH_QP_STATE_CLOSING:
965                         BUG_ON(kref_read(&qhp->ep->com.kref) < 2);
966                         qhp->attr.state = IWCH_QP_STATE_CLOSING;
967                         if (!internal) {
968                                 abort=0;
969                                 disconnect = 1;
970                                 ep = qhp->ep;
971                                 get_ep(&ep->com);
972                         }
973                         break;
974                 case IWCH_QP_STATE_TERMINATE:
975                         qhp->attr.state = IWCH_QP_STATE_TERMINATE;
976                         if (qhp->ibqp.uobject)
977                                 cxio_set_wq_in_error(&qhp->wq);
978                         if (!internal)
979                                 terminate = 1;
980                         break;
981                 case IWCH_QP_STATE_ERROR:
982                         qhp->attr.state = IWCH_QP_STATE_ERROR;
983                         if (!internal) {
984                                 abort=1;
985                                 disconnect = 1;
986                                 ep = qhp->ep;
987                                 get_ep(&ep->com);
988                         }
989                         goto err;
990                         break;
991                 default:
992                         ret = -EINVAL;
993                         goto out;
994                 }
995                 break;
996         case IWCH_QP_STATE_CLOSING:
997                 if (!internal) {
998                         ret = -EINVAL;
999                         goto out;
1000                 }
1001                 switch (attrs->next_state) {
1002                         case IWCH_QP_STATE_IDLE:
1003                                 flush_qp(qhp);
1004                                 qhp->attr.state = IWCH_QP_STATE_IDLE;
1005                                 qhp->attr.llp_stream_handle = NULL;
1006                                 put_ep(&qhp->ep->com);
1007                                 qhp->ep = NULL;
1008                                 wake_up(&qhp->wait);
1009                                 break;
1010                         case IWCH_QP_STATE_ERROR:
1011                                 goto err;
1012                         default:
1013                                 ret = -EINVAL;
1014                                 goto err;
1015                 }
1016                 break;
1017         case IWCH_QP_STATE_ERROR:
1018                 if (attrs->next_state != IWCH_QP_STATE_IDLE) {
1019                         ret = -EINVAL;
1020                         goto out;
1021                 }
1022
1023                 if (!Q_EMPTY(qhp->wq.sq_rptr, qhp->wq.sq_wptr) ||
1024                     !Q_EMPTY(qhp->wq.rq_rptr, qhp->wq.rq_wptr)) {
1025                         ret = -EINVAL;
1026                         goto out;
1027                 }
1028                 qhp->attr.state = IWCH_QP_STATE_IDLE;
1029                 break;
1030         case IWCH_QP_STATE_TERMINATE:
1031                 if (!internal) {
1032                         ret = -EINVAL;
1033                         goto out;
1034                 }
1035                 goto err;
1036                 break;
1037         default:
1038                 pr_err("%s in a bad state %d\n", __func__, qhp->attr.state);
1039                 ret = -EINVAL;
1040                 goto err;
1041                 break;
1042         }
1043         goto out;
1044 err:
1045         pr_debug("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1046                  qhp->wq.qpid);
1047
1048         /* disassociate the LLP connection */
1049         qhp->attr.llp_stream_handle = NULL;
1050         ep = qhp->ep;
1051         qhp->ep = NULL;
1052         qhp->attr.state = IWCH_QP_STATE_ERROR;
1053         free=1;
1054         wake_up(&qhp->wait);
1055         BUG_ON(!ep);
1056         flush_qp(qhp);
1057 out:
1058         spin_unlock_irqrestore(&qhp->lock, flag);
1059
1060         if (terminate)
1061                 iwch_post_terminate(qhp, NULL);
1062
1063         /*
1064          * If disconnect is 1, then we need to initiate a disconnect
1065          * on the EP.  This can be a normal close (RTS->CLOSING) or
1066          * an abnormal close (RTS/CLOSING->ERROR).
1067          */
1068         if (disconnect) {
1069                 iwch_ep_disconnect(ep, abort, GFP_KERNEL);
1070                 put_ep(&ep->com);
1071         }
1072
1073         /*
1074          * If free is 1, then we've disassociated the EP from the QP
1075          * and we need to dereference the EP.
1076          */
1077         if (free)
1078                 put_ep(&ep->com);
1079
1080         pr_debug("%s exit state %d\n", __func__, qhp->attr.state);
1081         return ret;
1082 }