Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / 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, 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, 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(wr->wr.rdma.rkey);
99         wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.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, 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(wr->wr.rdma.rkey);
141         wqe->read.rem_to = cpu_to_be64(wr->wr.rdma.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_fastreg(union t3_wr *wqe, struct ib_send_wr *wr,
150                                 u8 *flit_cnt, int *wr_cnt, struct t3_wq *wq)
151 {
152         int i;
153         __be64 *p;
154
155         if (wr->wr.fast_reg.page_list_len > T3_MAX_FASTREG_DEPTH)
156                 return -EINVAL;
157         *wr_cnt = 1;
158         wqe->fastreg.stag = cpu_to_be32(wr->wr.fast_reg.rkey);
159         wqe->fastreg.len = cpu_to_be32(wr->wr.fast_reg.length);
160         wqe->fastreg.va_base_hi = cpu_to_be32(wr->wr.fast_reg.iova_start >> 32);
161         wqe->fastreg.va_base_lo_fbo =
162                                 cpu_to_be32(wr->wr.fast_reg.iova_start & 0xffffffff);
163         wqe->fastreg.page_type_perms = cpu_to_be32(
164                 V_FR_PAGE_COUNT(wr->wr.fast_reg.page_list_len) |
165                 V_FR_PAGE_SIZE(wr->wr.fast_reg.page_shift-12) |
166                 V_FR_TYPE(TPT_VATO) |
167                 V_FR_PERMS(iwch_ib_to_tpt_access(wr->wr.fast_reg.access_flags)));
168         p = &wqe->fastreg.pbl_addrs[0];
169         for (i = 0; i < wr->wr.fast_reg.page_list_len; i++, p++) {
170
171                 /* If we need a 2nd WR, then set it up */
172                 if (i == T3_MAX_FASTREG_FRAG) {
173                         *wr_cnt = 2;
174                         wqe = (union t3_wr *)(wq->queue +
175                                 Q_PTR2IDX((wq->wptr+1), wq->size_log2));
176                         build_fw_riwrh((void *)wqe, T3_WR_FASTREG, 0,
177                                Q_GENBIT(wq->wptr + 1, wq->size_log2),
178                                0, 1 + wr->wr.fast_reg.page_list_len - T3_MAX_FASTREG_FRAG,
179                                T3_EOP);
180
181                         p = &wqe->pbl_frag.pbl_addrs[0];
182                 }
183                 *p = cpu_to_be64((u64)wr->wr.fast_reg.page_list->page_list[i]);
184         }
185         *flit_cnt = 5 + wr->wr.fast_reg.page_list_len;
186         if (*flit_cnt > 15)
187                 *flit_cnt = 15;
188         return 0;
189 }
190
191 static int build_inv_stag(union t3_wr *wqe, struct ib_send_wr *wr,
192                                 u8 *flit_cnt)
193 {
194         wqe->local_inv.stag = cpu_to_be32(wr->ex.invalidate_rkey);
195         wqe->local_inv.reserved = 0;
196         *flit_cnt = sizeof(struct t3_local_inv_wr) >> 3;
197         return 0;
198 }
199
200 static int iwch_sgl2pbl_map(struct iwch_dev *rhp, struct ib_sge *sg_list,
201                             u32 num_sgle, u32 * pbl_addr, u8 * page_size)
202 {
203         int i;
204         struct iwch_mr *mhp;
205         u64 offset;
206         for (i = 0; i < num_sgle; i++) {
207
208                 mhp = get_mhp(rhp, (sg_list[i].lkey) >> 8);
209                 if (!mhp) {
210                         PDBG("%s %d\n", __func__, __LINE__);
211                         return -EIO;
212                 }
213                 if (!mhp->attr.state) {
214                         PDBG("%s %d\n", __func__, __LINE__);
215                         return -EIO;
216                 }
217                 if (mhp->attr.zbva) {
218                         PDBG("%s %d\n", __func__, __LINE__);
219                         return -EIO;
220                 }
221
222                 if (sg_list[i].addr < mhp->attr.va_fbo) {
223                         PDBG("%s %d\n", __func__, __LINE__);
224                         return -EINVAL;
225                 }
226                 if (sg_list[i].addr + ((u64) sg_list[i].length) <
227                     sg_list[i].addr) {
228                         PDBG("%s %d\n", __func__, __LINE__);
229                         return -EINVAL;
230                 }
231                 if (sg_list[i].addr + ((u64) sg_list[i].length) >
232                     mhp->attr.va_fbo + ((u64) mhp->attr.len)) {
233                         PDBG("%s %d\n", __func__, __LINE__);
234                         return -EINVAL;
235                 }
236                 offset = sg_list[i].addr - mhp->attr.va_fbo;
237                 offset += mhp->attr.va_fbo &
238                           ((1UL << (12 + mhp->attr.page_size)) - 1);
239                 pbl_addr[i] = ((mhp->attr.pbl_addr -
240                                 rhp->rdev.rnic_info.pbl_base) >> 3) +
241                               (offset >> (12 + mhp->attr.page_size));
242                 page_size[i] = mhp->attr.page_size;
243         }
244         return 0;
245 }
246
247 static int build_rdma_recv(struct iwch_qp *qhp, union t3_wr *wqe,
248                                 struct ib_recv_wr *wr)
249 {
250         int i, err = 0;
251         u32 pbl_addr[T3_MAX_SGE];
252         u8 page_size[T3_MAX_SGE];
253
254         err = iwch_sgl2pbl_map(qhp->rhp, wr->sg_list, wr->num_sge, pbl_addr,
255                                page_size);
256         if (err)
257                 return err;
258         wqe->recv.pagesz[0] = page_size[0];
259         wqe->recv.pagesz[1] = page_size[1];
260         wqe->recv.pagesz[2] = page_size[2];
261         wqe->recv.pagesz[3] = page_size[3];
262         wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
263         for (i = 0; i < wr->num_sge; i++) {
264                 wqe->recv.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
265                 wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
266
267                 /* to in the WQE == the offset into the page */
268                 wqe->recv.sgl[i].to = cpu_to_be64(((u32)wr->sg_list[i].addr) &
269                                 ((1UL << (12 + page_size[i])) - 1));
270
271                 /* pbl_addr is the adapters address in the PBL */
272                 wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_addr[i]);
273         }
274         for (; i < T3_MAX_SGE; i++) {
275                 wqe->recv.sgl[i].stag = 0;
276                 wqe->recv.sgl[i].len = 0;
277                 wqe->recv.sgl[i].to = 0;
278                 wqe->recv.pbl_addr[i] = 0;
279         }
280         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
281                              qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
282         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
283                              qhp->wq.rq_size_log2)].pbl_addr = 0;
284         return 0;
285 }
286
287 static int build_zero_stag_recv(struct iwch_qp *qhp, union t3_wr *wqe,
288                                 struct ib_recv_wr *wr)
289 {
290         int i;
291         u32 pbl_addr;
292         u32 pbl_offset;
293
294
295         /*
296          * The T3 HW requires the PBL in the HW recv descriptor to reference
297          * a PBL entry.  So we allocate the max needed PBL memory here and pass
298          * it to the uP in the recv WR.  The uP will build the PBL and setup
299          * the HW recv descriptor.
300          */
301         pbl_addr = cxio_hal_pblpool_alloc(&qhp->rhp->rdev, T3_STAG0_PBL_SIZE);
302         if (!pbl_addr)
303                 return -ENOMEM;
304
305         /*
306          * Compute the 8B aligned offset.
307          */
308         pbl_offset = (pbl_addr - qhp->rhp->rdev.rnic_info.pbl_base) >> 3;
309
310         wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
311
312         for (i = 0; i < wr->num_sge; i++) {
313
314                 /*
315                  * Use a 128MB page size. This and an imposed 128MB
316                  * sge length limit allows us to require only a 2-entry HW
317                  * PBL for each SGE.  This restriction is acceptable since
318                  * since it is not possible to allocate 128MB of contiguous
319                  * DMA coherent memory!
320                  */
321                 if (wr->sg_list[i].length > T3_STAG0_MAX_PBE_LEN)
322                         return -EINVAL;
323                 wqe->recv.pagesz[i] = T3_STAG0_PAGE_SHIFT;
324
325                 /*
326                  * T3 restricts a recv to all zero-stag or all non-zero-stag.
327                  */
328                 if (wr->sg_list[i].lkey != 0)
329                         return -EINVAL;
330                 wqe->recv.sgl[i].stag = 0;
331                 wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
332                 wqe->recv.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
333                 wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_offset);
334                 pbl_offset += 2;
335         }
336         for (; i < T3_MAX_SGE; i++) {
337                 wqe->recv.pagesz[i] = 0;
338                 wqe->recv.sgl[i].stag = 0;
339                 wqe->recv.sgl[i].len = 0;
340                 wqe->recv.sgl[i].to = 0;
341                 wqe->recv.pbl_addr[i] = 0;
342         }
343         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
344                              qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
345         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
346                              qhp->wq.rq_size_log2)].pbl_addr = pbl_addr;
347         return 0;
348 }
349
350 int iwch_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
351                       struct ib_send_wr **bad_wr)
352 {
353         int err = 0;
354         u8 uninitialized_var(t3_wr_flit_cnt);
355         enum t3_wr_opcode t3_wr_opcode = 0;
356         enum t3_wr_flags t3_wr_flags;
357         struct iwch_qp *qhp;
358         u32 idx;
359         union t3_wr *wqe;
360         u32 num_wrs;
361         unsigned long flag;
362         struct t3_swsq *sqp;
363         int wr_cnt = 1;
364
365         qhp = to_iwch_qp(ibqp);
366         spin_lock_irqsave(&qhp->lock, flag);
367         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
368                 spin_unlock_irqrestore(&qhp->lock, flag);
369                 err = -EINVAL;
370                 goto out;
371         }
372         num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
373                   qhp->wq.sq_size_log2);
374         if (num_wrs == 0) {
375                 spin_unlock_irqrestore(&qhp->lock, flag);
376                 err = -ENOMEM;
377                 goto out;
378         }
379         while (wr) {
380                 if (num_wrs == 0) {
381                         err = -ENOMEM;
382                         break;
383                 }
384                 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
385                 wqe = (union t3_wr *) (qhp->wq.queue + idx);
386                 t3_wr_flags = 0;
387                 if (wr->send_flags & IB_SEND_SOLICITED)
388                         t3_wr_flags |= T3_SOLICITED_EVENT_FLAG;
389                 if (wr->send_flags & IB_SEND_SIGNALED)
390                         t3_wr_flags |= T3_COMPLETION_FLAG;
391                 sqp = qhp->wq.sq +
392                       Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
393                 switch (wr->opcode) {
394                 case IB_WR_SEND:
395                 case IB_WR_SEND_WITH_INV:
396                         if (wr->send_flags & IB_SEND_FENCE)
397                                 t3_wr_flags |= T3_READ_FENCE_FLAG;
398                         t3_wr_opcode = T3_WR_SEND;
399                         err = build_rdma_send(wqe, wr, &t3_wr_flit_cnt);
400                         break;
401                 case IB_WR_RDMA_WRITE:
402                 case IB_WR_RDMA_WRITE_WITH_IMM:
403                         t3_wr_opcode = T3_WR_WRITE;
404                         err = build_rdma_write(wqe, wr, &t3_wr_flit_cnt);
405                         break;
406                 case IB_WR_RDMA_READ:
407                 case IB_WR_RDMA_READ_WITH_INV:
408                         t3_wr_opcode = T3_WR_READ;
409                         t3_wr_flags = 0; /* T3 reads are always signaled */
410                         err = build_rdma_read(wqe, wr, &t3_wr_flit_cnt);
411                         if (err)
412                                 break;
413                         sqp->read_len = wqe->read.local_len;
414                         if (!qhp->wq.oldest_read)
415                                 qhp->wq.oldest_read = sqp;
416                         break;
417                 case IB_WR_FAST_REG_MR:
418                         t3_wr_opcode = T3_WR_FASTREG;
419                         err = build_fastreg(wqe, wr, &t3_wr_flit_cnt,
420                                                  &wr_cnt, &qhp->wq);
421                         break;
422                 case IB_WR_LOCAL_INV:
423                         if (wr->send_flags & IB_SEND_FENCE)
424                                 t3_wr_flags |= T3_LOCAL_FENCE_FLAG;
425                         t3_wr_opcode = T3_WR_INV_STAG;
426                         err = build_inv_stag(wqe, wr, &t3_wr_flit_cnt);
427                         break;
428                 default:
429                         PDBG("%s post of type=%d TBD!\n", __func__,
430                              wr->opcode);
431                         err = -EINVAL;
432                 }
433                 if (err)
434                         break;
435                 wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
436                 sqp->wr_id = wr->wr_id;
437                 sqp->opcode = wr2opcode(t3_wr_opcode);
438                 sqp->sq_wptr = qhp->wq.sq_wptr;
439                 sqp->complete = 0;
440                 sqp->signaled = (wr->send_flags & IB_SEND_SIGNALED);
441
442                 build_fw_riwrh((void *) wqe, t3_wr_opcode, t3_wr_flags,
443                                Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
444                                0, t3_wr_flit_cnt,
445                                (wr_cnt == 1) ? T3_SOPEOP : T3_SOP);
446                 PDBG("%s cookie 0x%llx wq idx 0x%x swsq idx %ld opcode %d\n",
447                      __func__, (unsigned long long) wr->wr_id, idx,
448                      Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2),
449                      sqp->opcode);
450                 wr = wr->next;
451                 num_wrs--;
452                 qhp->wq.wptr += wr_cnt;
453                 ++(qhp->wq.sq_wptr);
454         }
455         spin_unlock_irqrestore(&qhp->lock, flag);
456         if (cxio_wq_db_enabled(&qhp->wq))
457                 ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
458
459 out:
460         if (err)
461                 *bad_wr = wr;
462         return err;
463 }
464
465 int iwch_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
466                       struct ib_recv_wr **bad_wr)
467 {
468         int err = 0;
469         struct iwch_qp *qhp;
470         u32 idx;
471         union t3_wr *wqe;
472         u32 num_wrs;
473         unsigned long flag;
474
475         qhp = to_iwch_qp(ibqp);
476         spin_lock_irqsave(&qhp->lock, flag);
477         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
478                 spin_unlock_irqrestore(&qhp->lock, flag);
479                 err = -EINVAL;
480                 goto out;
481         }
482         num_wrs = Q_FREECNT(qhp->wq.rq_rptr, qhp->wq.rq_wptr,
483                             qhp->wq.rq_size_log2) - 1;
484         if (!wr) {
485                 spin_unlock_irqrestore(&qhp->lock, flag);
486                 err = -ENOMEM;
487                 goto out;
488         }
489         while (wr) {
490                 if (wr->num_sge > T3_MAX_SGE) {
491                         err = -EINVAL;
492                         break;
493                 }
494                 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
495                 wqe = (union t3_wr *) (qhp->wq.queue + idx);
496                 if (num_wrs)
497                         if (wr->sg_list[0].lkey)
498                                 err = build_rdma_recv(qhp, wqe, wr);
499                         else
500                                 err = build_zero_stag_recv(qhp, wqe, wr);
501                 else
502                         err = -ENOMEM;
503
504                 if (err)
505                         break;
506
507                 build_fw_riwrh((void *) wqe, T3_WR_RCV, T3_COMPLETION_FLAG,
508                                Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
509                                0, sizeof(struct t3_receive_wr) >> 3, T3_SOPEOP);
510                 PDBG("%s cookie 0x%llx idx 0x%x rq_wptr 0x%x rw_rptr 0x%x "
511                      "wqe %p \n", __func__, (unsigned long long) wr->wr_id,
512                      idx, qhp->wq.rq_wptr, qhp->wq.rq_rptr, wqe);
513                 ++(qhp->wq.rq_wptr);
514                 ++(qhp->wq.wptr);
515                 wr = wr->next;
516                 num_wrs--;
517         }
518         spin_unlock_irqrestore(&qhp->lock, flag);
519         if (cxio_wq_db_enabled(&qhp->wq))
520                 ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
521
522 out:
523         if (err)
524                 *bad_wr = wr;
525         return err;
526 }
527
528 int iwch_bind_mw(struct ib_qp *qp,
529                              struct ib_mw *mw,
530                              struct ib_mw_bind *mw_bind)
531 {
532         struct iwch_dev *rhp;
533         struct iwch_mw *mhp;
534         struct iwch_qp *qhp;
535         union t3_wr *wqe;
536         u32 pbl_addr;
537         u8 page_size;
538         u32 num_wrs;
539         unsigned long flag;
540         struct ib_sge sgl;
541         int err=0;
542         enum t3_wr_flags t3_wr_flags;
543         u32 idx;
544         struct t3_swsq *sqp;
545
546         qhp = to_iwch_qp(qp);
547         mhp = to_iwch_mw(mw);
548         rhp = qhp->rhp;
549
550         spin_lock_irqsave(&qhp->lock, flag);
551         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
552                 spin_unlock_irqrestore(&qhp->lock, flag);
553                 return -EINVAL;
554         }
555         num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
556                             qhp->wq.sq_size_log2);
557         if (num_wrs == 0) {
558                 spin_unlock_irqrestore(&qhp->lock, flag);
559                 return -ENOMEM;
560         }
561         idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
562         PDBG("%s: idx 0x%0x, mw 0x%p, mw_bind 0x%p\n", __func__, idx,
563              mw, mw_bind);
564         wqe = (union t3_wr *) (qhp->wq.queue + idx);
565
566         t3_wr_flags = 0;
567         if (mw_bind->send_flags & IB_SEND_SIGNALED)
568                 t3_wr_flags = T3_COMPLETION_FLAG;
569
570         sgl.addr = mw_bind->bind_info.addr;
571         sgl.lkey = mw_bind->bind_info.mr->lkey;
572         sgl.length = mw_bind->bind_info.length;
573         wqe->bind.reserved = 0;
574         wqe->bind.type = TPT_VATO;
575
576         /* TBD: check perms */
577         wqe->bind.perms = iwch_ib_to_tpt_bind_access(
578                 mw_bind->bind_info.mw_access_flags);
579         wqe->bind.mr_stag = cpu_to_be32(mw_bind->bind_info.mr->lkey);
580         wqe->bind.mw_stag = cpu_to_be32(mw->rkey);
581         wqe->bind.mw_len = cpu_to_be32(mw_bind->bind_info.length);
582         wqe->bind.mw_va = cpu_to_be64(mw_bind->bind_info.addr);
583         err = iwch_sgl2pbl_map(rhp, &sgl, 1, &pbl_addr, &page_size);
584         if (err) {
585                 spin_unlock_irqrestore(&qhp->lock, flag);
586                 return err;
587         }
588         wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
589         sqp = qhp->wq.sq + Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
590         sqp->wr_id = mw_bind->wr_id;
591         sqp->opcode = T3_BIND_MW;
592         sqp->sq_wptr = qhp->wq.sq_wptr;
593         sqp->complete = 0;
594         sqp->signaled = (mw_bind->send_flags & IB_SEND_SIGNALED);
595         wqe->bind.mr_pbl_addr = cpu_to_be32(pbl_addr);
596         wqe->bind.mr_pagesz = page_size;
597         build_fw_riwrh((void *)wqe, T3_WR_BIND, t3_wr_flags,
598                        Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2), 0,
599                        sizeof(struct t3_bind_mw_wr) >> 3, T3_SOPEOP);
600         ++(qhp->wq.wptr);
601         ++(qhp->wq.sq_wptr);
602         spin_unlock_irqrestore(&qhp->lock, flag);
603
604         if (cxio_wq_db_enabled(&qhp->wq))
605                 ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
606
607         return err;
608 }
609
610 static inline void build_term_codes(struct respQ_msg_t *rsp_msg,
611                                     u8 *layer_type, u8 *ecode)
612 {
613         int status = TPT_ERR_INTERNAL_ERR;
614         int tagged = 0;
615         int opcode = -1;
616         int rqtype = 0;
617         int send_inv = 0;
618
619         if (rsp_msg) {
620                 status = CQE_STATUS(rsp_msg->cqe);
621                 opcode = CQE_OPCODE(rsp_msg->cqe);
622                 rqtype = RQ_TYPE(rsp_msg->cqe);
623                 send_inv = (opcode == T3_SEND_WITH_INV) ||
624                            (opcode == T3_SEND_WITH_SE_INV);
625                 tagged = (opcode == T3_RDMA_WRITE) ||
626                          (rqtype && (opcode == T3_READ_RESP));
627         }
628
629         switch (status) {
630         case TPT_ERR_STAG:
631                 if (send_inv) {
632                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
633                         *ecode = RDMAP_CANT_INV_STAG;
634                 } else {
635                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
636                         *ecode = RDMAP_INV_STAG;
637                 }
638                 break;
639         case TPT_ERR_PDID:
640                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
641                 if ((opcode == T3_SEND_WITH_INV) ||
642                     (opcode == T3_SEND_WITH_SE_INV))
643                         *ecode = RDMAP_CANT_INV_STAG;
644                 else
645                         *ecode = RDMAP_STAG_NOT_ASSOC;
646                 break;
647         case TPT_ERR_QPID:
648                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
649                 *ecode = RDMAP_STAG_NOT_ASSOC;
650                 break;
651         case TPT_ERR_ACCESS:
652                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
653                 *ecode = RDMAP_ACC_VIOL;
654                 break;
655         case TPT_ERR_WRAP:
656                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
657                 *ecode = RDMAP_TO_WRAP;
658                 break;
659         case TPT_ERR_BOUND:
660                 if (tagged) {
661                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
662                         *ecode = DDPT_BASE_BOUNDS;
663                 } else {
664                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
665                         *ecode = RDMAP_BASE_BOUNDS;
666                 }
667                 break;
668         case TPT_ERR_INVALIDATE_SHARED_MR:
669         case TPT_ERR_INVALIDATE_MR_WITH_MW_BOUND:
670                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
671                 *ecode = RDMAP_CANT_INV_STAG;
672                 break;
673         case TPT_ERR_ECC:
674         case TPT_ERR_ECC_PSTAG:
675         case TPT_ERR_INTERNAL_ERR:
676                 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
677                 *ecode = 0;
678                 break;
679         case TPT_ERR_OUT_OF_RQE:
680                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
681                 *ecode = DDPU_INV_MSN_NOBUF;
682                 break;
683         case TPT_ERR_PBL_ADDR_BOUND:
684                 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
685                 *ecode = DDPT_BASE_BOUNDS;
686                 break;
687         case TPT_ERR_CRC:
688                 *layer_type = LAYER_MPA|DDP_LLP;
689                 *ecode = MPA_CRC_ERR;
690                 break;
691         case TPT_ERR_MARKER:
692                 *layer_type = LAYER_MPA|DDP_LLP;
693                 *ecode = MPA_MARKER_ERR;
694                 break;
695         case TPT_ERR_PDU_LEN_ERR:
696                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
697                 *ecode = DDPU_MSG_TOOBIG;
698                 break;
699         case TPT_ERR_DDP_VERSION:
700                 if (tagged) {
701                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
702                         *ecode = DDPT_INV_VERS;
703                 } else {
704                         *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
705                         *ecode = DDPU_INV_VERS;
706                 }
707                 break;
708         case TPT_ERR_RDMA_VERSION:
709                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
710                 *ecode = RDMAP_INV_VERS;
711                 break;
712         case TPT_ERR_OPCODE:
713                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
714                 *ecode = RDMAP_INV_OPCODE;
715                 break;
716         case TPT_ERR_DDP_QUEUE_NUM:
717                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
718                 *ecode = DDPU_INV_QN;
719                 break;
720         case TPT_ERR_MSN:
721         case TPT_ERR_MSN_GAP:
722         case TPT_ERR_MSN_RANGE:
723         case TPT_ERR_IRD_OVERFLOW:
724                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
725                 *ecode = DDPU_INV_MSN_RANGE;
726                 break;
727         case TPT_ERR_TBIT:
728                 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
729                 *ecode = 0;
730                 break;
731         case TPT_ERR_MO:
732                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
733                 *ecode = DDPU_INV_MO;
734                 break;
735         default:
736                 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
737                 *ecode = 0;
738                 break;
739         }
740 }
741
742 int iwch_post_zb_read(struct iwch_ep *ep)
743 {
744         union t3_wr *wqe;
745         struct sk_buff *skb;
746         u8 flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
747
748         PDBG("%s enter\n", __func__);
749         skb = alloc_skb(40, GFP_KERNEL);
750         if (!skb) {
751                 printk(KERN_ERR "%s cannot send zb_read!!\n", __func__);
752                 return -ENOMEM;
753         }
754         wqe = (union t3_wr *)skb_put(skb, sizeof(struct t3_rdma_read_wr));
755         memset(wqe, 0, sizeof(struct t3_rdma_read_wr));
756         wqe->read.rdmaop = T3_READ_REQ;
757         wqe->read.reserved[0] = 0;
758         wqe->read.reserved[1] = 0;
759         wqe->read.rem_stag = cpu_to_be32(1);
760         wqe->read.rem_to = cpu_to_be64(1);
761         wqe->read.local_stag = cpu_to_be32(1);
762         wqe->read.local_len = cpu_to_be32(0);
763         wqe->read.local_to = cpu_to_be64(1);
764         wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_READ));
765         wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(ep->hwtid)|
766                                                 V_FW_RIWR_LEN(flit_cnt));
767         skb->priority = CPL_PRIORITY_DATA;
768         return iwch_cxgb3_ofld_send(ep->com.qp->rhp->rdev.t3cdev_p, skb);
769 }
770
771 /*
772  * This posts a TERMINATE with layer=RDMA, type=catastrophic.
773  */
774 int iwch_post_terminate(struct iwch_qp *qhp, struct respQ_msg_t *rsp_msg)
775 {
776         union t3_wr *wqe;
777         struct terminate_message *term;
778         struct sk_buff *skb;
779
780         PDBG("%s %d\n", __func__, __LINE__);
781         skb = alloc_skb(40, GFP_ATOMIC);
782         if (!skb) {
783                 printk(KERN_ERR "%s cannot send TERMINATE!\n", __func__);
784                 return -ENOMEM;
785         }
786         wqe = (union t3_wr *)skb_put(skb, 40);
787         memset(wqe, 0, 40);
788         wqe->send.rdmaop = T3_TERMINATE;
789
790         /* immediate data length */
791         wqe->send.plen = htonl(4);
792
793         /* immediate data starts here. */
794         term = (struct terminate_message *)wqe->send.sgl;
795         build_term_codes(rsp_msg, &term->layer_etype, &term->ecode);
796         wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_SEND) |
797                          V_FW_RIWR_FLAGS(T3_COMPLETION_FLAG | T3_NOTIFY_FLAG));
798         wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(qhp->ep->hwtid));
799         skb->priority = CPL_PRIORITY_DATA;
800         return iwch_cxgb3_ofld_send(qhp->rhp->rdev.t3cdev_p, skb);
801 }
802
803 /*
804  * Assumes qhp lock is held.
805  */
806 static void __flush_qp(struct iwch_qp *qhp, struct iwch_cq *rchp,
807                                 struct iwch_cq *schp)
808 {
809         int count;
810         int flushed;
811
812
813         PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
814         /* take a ref on the qhp since we must release the lock */
815         atomic_inc(&qhp->refcnt);
816         spin_unlock(&qhp->lock);
817
818         /* locking hierarchy: cq lock first, then qp lock. */
819         spin_lock(&rchp->lock);
820         spin_lock(&qhp->lock);
821         cxio_flush_hw_cq(&rchp->cq);
822         cxio_count_rcqes(&rchp->cq, &qhp->wq, &count);
823         flushed = cxio_flush_rq(&qhp->wq, &rchp->cq, count);
824         spin_unlock(&qhp->lock);
825         spin_unlock(&rchp->lock);
826         if (flushed) {
827                 spin_lock(&rchp->comp_handler_lock);
828                 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
829                 spin_unlock(&rchp->comp_handler_lock);
830         }
831
832         /* locking hierarchy: cq lock first, then qp lock. */
833         spin_lock(&schp->lock);
834         spin_lock(&qhp->lock);
835         cxio_flush_hw_cq(&schp->cq);
836         cxio_count_scqes(&schp->cq, &qhp->wq, &count);
837         flushed = cxio_flush_sq(&qhp->wq, &schp->cq, count);
838         spin_unlock(&qhp->lock);
839         spin_unlock(&schp->lock);
840         if (flushed) {
841                 spin_lock(&schp->comp_handler_lock);
842                 (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
843                 spin_unlock(&schp->comp_handler_lock);
844         }
845
846         /* deref */
847         if (atomic_dec_and_test(&qhp->refcnt))
848                 wake_up(&qhp->wait);
849
850         spin_lock(&qhp->lock);
851 }
852
853 static void flush_qp(struct iwch_qp *qhp)
854 {
855         struct iwch_cq *rchp, *schp;
856
857         rchp = get_chp(qhp->rhp, qhp->attr.rcq);
858         schp = get_chp(qhp->rhp, qhp->attr.scq);
859
860         if (qhp->ibqp.uobject) {
861                 cxio_set_wq_in_error(&qhp->wq);
862                 cxio_set_cq_in_error(&rchp->cq);
863                 spin_lock(&rchp->comp_handler_lock);
864                 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
865                 spin_unlock(&rchp->comp_handler_lock);
866                 if (schp != rchp) {
867                         cxio_set_cq_in_error(&schp->cq);
868                         spin_lock(&schp->comp_handler_lock);
869                         (*schp->ibcq.comp_handler)(&schp->ibcq,
870                                                    schp->ibcq.cq_context);
871                         spin_unlock(&schp->comp_handler_lock);
872                 }
873                 return;
874         }
875         __flush_qp(qhp, rchp, schp);
876 }
877
878
879 /*
880  * Return count of RECV WRs posted
881  */
882 u16 iwch_rqes_posted(struct iwch_qp *qhp)
883 {
884         union t3_wr *wqe = qhp->wq.queue;
885         u16 count = 0;
886
887         while (count < USHRT_MAX && fw_riwrh_opcode((struct fw_riwrh *)wqe) == T3_WR_RCV) {
888                 count++;
889                 wqe++;
890         }
891         PDBG("%s qhp %p count %u\n", __func__, qhp, count);
892         return count;
893 }
894
895 static int rdma_init(struct iwch_dev *rhp, struct iwch_qp *qhp,
896                                 enum iwch_qp_attr_mask mask,
897                                 struct iwch_qp_attributes *attrs)
898 {
899         struct t3_rdma_init_attr init_attr;
900         int ret;
901
902         init_attr.tid = qhp->ep->hwtid;
903         init_attr.qpid = qhp->wq.qpid;
904         init_attr.pdid = qhp->attr.pd;
905         init_attr.scqid = qhp->attr.scq;
906         init_attr.rcqid = qhp->attr.rcq;
907         init_attr.rq_addr = qhp->wq.rq_addr;
908         init_attr.rq_size = 1 << qhp->wq.rq_size_log2;
909         init_attr.mpaattrs = uP_RI_MPA_IETF_ENABLE |
910                 qhp->attr.mpa_attr.recv_marker_enabled |
911                 (qhp->attr.mpa_attr.xmit_marker_enabled << 1) |
912                 (qhp->attr.mpa_attr.crc_enabled << 2);
913
914         init_attr.qpcaps = uP_RI_QP_RDMA_READ_ENABLE |
915                            uP_RI_QP_RDMA_WRITE_ENABLE |
916                            uP_RI_QP_BIND_ENABLE;
917         if (!qhp->ibqp.uobject)
918                 init_attr.qpcaps |= uP_RI_QP_STAG0_ENABLE |
919                                     uP_RI_QP_FAST_REGISTER_ENABLE;
920
921         init_attr.tcp_emss = qhp->ep->emss;
922         init_attr.ord = qhp->attr.max_ord;
923         init_attr.ird = qhp->attr.max_ird;
924         init_attr.qp_dma_addr = qhp->wq.dma_addr;
925         init_attr.qp_dma_size = (1UL << qhp->wq.size_log2);
926         init_attr.rqe_count = iwch_rqes_posted(qhp);
927         init_attr.flags = qhp->attr.mpa_attr.initiator ? MPA_INITIATOR : 0;
928         init_attr.chan = qhp->ep->l2t->smt_idx;
929         if (peer2peer) {
930                 init_attr.rtr_type = RTR_READ;
931                 if (init_attr.ord == 0 && qhp->attr.mpa_attr.initiator)
932                         init_attr.ord = 1;
933                 if (init_attr.ird == 0 && !qhp->attr.mpa_attr.initiator)
934                         init_attr.ird = 1;
935         } else
936                 init_attr.rtr_type = 0;
937         init_attr.irs = qhp->ep->rcv_seq;
938         PDBG("%s init_attr.rq_addr 0x%x init_attr.rq_size = %d "
939              "flags 0x%x qpcaps 0x%x\n", __func__,
940              init_attr.rq_addr, init_attr.rq_size,
941              init_attr.flags, init_attr.qpcaps);
942         ret = cxio_rdma_init(&rhp->rdev, &init_attr);
943         PDBG("%s ret %d\n", __func__, ret);
944         return ret;
945 }
946
947 int iwch_modify_qp(struct iwch_dev *rhp, struct iwch_qp *qhp,
948                                 enum iwch_qp_attr_mask mask,
949                                 struct iwch_qp_attributes *attrs,
950                                 int internal)
951 {
952         int ret = 0;
953         struct iwch_qp_attributes newattr = qhp->attr;
954         unsigned long flag;
955         int disconnect = 0;
956         int terminate = 0;
957         int abort = 0;
958         int free = 0;
959         struct iwch_ep *ep = NULL;
960
961         PDBG("%s qhp %p qpid 0x%x ep %p state %d -> %d\n", __func__,
962              qhp, qhp->wq.qpid, qhp->ep, qhp->attr.state,
963              (mask & IWCH_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
964
965         spin_lock_irqsave(&qhp->lock, flag);
966
967         /* Process attr changes if in IDLE */
968         if (mask & IWCH_QP_ATTR_VALID_MODIFY) {
969                 if (qhp->attr.state != IWCH_QP_STATE_IDLE) {
970                         ret = -EIO;
971                         goto out;
972                 }
973                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_READ)
974                         newattr.enable_rdma_read = attrs->enable_rdma_read;
975                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_WRITE)
976                         newattr.enable_rdma_write = attrs->enable_rdma_write;
977                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_BIND)
978                         newattr.enable_bind = attrs->enable_bind;
979                 if (mask & IWCH_QP_ATTR_MAX_ORD) {
980                         if (attrs->max_ord >
981                             rhp->attr.max_rdma_read_qp_depth) {
982                                 ret = -EINVAL;
983                                 goto out;
984                         }
985                         newattr.max_ord = attrs->max_ord;
986                 }
987                 if (mask & IWCH_QP_ATTR_MAX_IRD) {
988                         if (attrs->max_ird >
989                             rhp->attr.max_rdma_reads_per_qp) {
990                                 ret = -EINVAL;
991                                 goto out;
992                         }
993                         newattr.max_ird = attrs->max_ird;
994                 }
995                 qhp->attr = newattr;
996         }
997
998         if (!(mask & IWCH_QP_ATTR_NEXT_STATE))
999                 goto out;
1000         if (qhp->attr.state == attrs->next_state)
1001                 goto out;
1002
1003         switch (qhp->attr.state) {
1004         case IWCH_QP_STATE_IDLE:
1005                 switch (attrs->next_state) {
1006                 case IWCH_QP_STATE_RTS:
1007                         if (!(mask & IWCH_QP_ATTR_LLP_STREAM_HANDLE)) {
1008                                 ret = -EINVAL;
1009                                 goto out;
1010                         }
1011                         if (!(mask & IWCH_QP_ATTR_MPA_ATTR)) {
1012                                 ret = -EINVAL;
1013                                 goto out;
1014                         }
1015                         qhp->attr.mpa_attr = attrs->mpa_attr;
1016                         qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1017                         qhp->ep = qhp->attr.llp_stream_handle;
1018                         qhp->attr.state = IWCH_QP_STATE_RTS;
1019
1020                         /*
1021                          * Ref the endpoint here and deref when we
1022                          * disassociate the endpoint from the QP.  This
1023                          * happens in CLOSING->IDLE transition or *->ERROR
1024                          * transition.
1025                          */
1026                         get_ep(&qhp->ep->com);
1027                         spin_unlock_irqrestore(&qhp->lock, flag);
1028                         ret = rdma_init(rhp, qhp, mask, attrs);
1029                         spin_lock_irqsave(&qhp->lock, flag);
1030                         if (ret)
1031                                 goto err;
1032                         break;
1033                 case IWCH_QP_STATE_ERROR:
1034                         qhp->attr.state = IWCH_QP_STATE_ERROR;
1035                         flush_qp(qhp);
1036                         break;
1037                 default:
1038                         ret = -EINVAL;
1039                         goto out;
1040                 }
1041                 break;
1042         case IWCH_QP_STATE_RTS:
1043                 switch (attrs->next_state) {
1044                 case IWCH_QP_STATE_CLOSING:
1045                         BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1046                         qhp->attr.state = IWCH_QP_STATE_CLOSING;
1047                         if (!internal) {
1048                                 abort=0;
1049                                 disconnect = 1;
1050                                 ep = qhp->ep;
1051                                 get_ep(&ep->com);
1052                         }
1053                         break;
1054                 case IWCH_QP_STATE_TERMINATE:
1055                         qhp->attr.state = IWCH_QP_STATE_TERMINATE;
1056                         if (qhp->ibqp.uobject)
1057                                 cxio_set_wq_in_error(&qhp->wq);
1058                         if (!internal)
1059                                 terminate = 1;
1060                         break;
1061                 case IWCH_QP_STATE_ERROR:
1062                         qhp->attr.state = IWCH_QP_STATE_ERROR;
1063                         if (!internal) {
1064                                 abort=1;
1065                                 disconnect = 1;
1066                                 ep = qhp->ep;
1067                                 get_ep(&ep->com);
1068                         }
1069                         goto err;
1070                         break;
1071                 default:
1072                         ret = -EINVAL;
1073                         goto out;
1074                 }
1075                 break;
1076         case IWCH_QP_STATE_CLOSING:
1077                 if (!internal) {
1078                         ret = -EINVAL;
1079                         goto out;
1080                 }
1081                 switch (attrs->next_state) {
1082                         case IWCH_QP_STATE_IDLE:
1083                                 flush_qp(qhp);
1084                                 qhp->attr.state = IWCH_QP_STATE_IDLE;
1085                                 qhp->attr.llp_stream_handle = NULL;
1086                                 put_ep(&qhp->ep->com);
1087                                 qhp->ep = NULL;
1088                                 wake_up(&qhp->wait);
1089                                 break;
1090                         case IWCH_QP_STATE_ERROR:
1091                                 goto err;
1092                         default:
1093                                 ret = -EINVAL;
1094                                 goto err;
1095                 }
1096                 break;
1097         case IWCH_QP_STATE_ERROR:
1098                 if (attrs->next_state != IWCH_QP_STATE_IDLE) {
1099                         ret = -EINVAL;
1100                         goto out;
1101                 }
1102
1103                 if (!Q_EMPTY(qhp->wq.sq_rptr, qhp->wq.sq_wptr) ||
1104                     !Q_EMPTY(qhp->wq.rq_rptr, qhp->wq.rq_wptr)) {
1105                         ret = -EINVAL;
1106                         goto out;
1107                 }
1108                 qhp->attr.state = IWCH_QP_STATE_IDLE;
1109                 break;
1110         case IWCH_QP_STATE_TERMINATE:
1111                 if (!internal) {
1112                         ret = -EINVAL;
1113                         goto out;
1114                 }
1115                 goto err;
1116                 break;
1117         default:
1118                 printk(KERN_ERR "%s in a bad state %d\n",
1119                        __func__, qhp->attr.state);
1120                 ret = -EINVAL;
1121                 goto err;
1122                 break;
1123         }
1124         goto out;
1125 err:
1126         PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1127              qhp->wq.qpid);
1128
1129         /* disassociate the LLP connection */
1130         qhp->attr.llp_stream_handle = NULL;
1131         ep = qhp->ep;
1132         qhp->ep = NULL;
1133         qhp->attr.state = IWCH_QP_STATE_ERROR;
1134         free=1;
1135         wake_up(&qhp->wait);
1136         BUG_ON(!ep);
1137         flush_qp(qhp);
1138 out:
1139         spin_unlock_irqrestore(&qhp->lock, flag);
1140
1141         if (terminate)
1142                 iwch_post_terminate(qhp, NULL);
1143
1144         /*
1145          * If disconnect is 1, then we need to initiate a disconnect
1146          * on the EP.  This can be a normal close (RTS->CLOSING) or
1147          * an abnormal close (RTS/CLOSING->ERROR).
1148          */
1149         if (disconnect) {
1150                 iwch_ep_disconnect(ep, abort, GFP_KERNEL);
1151                 put_ep(&ep->com);
1152         }
1153
1154         /*
1155          * If free is 1, then we've disassociated the EP from the QP
1156          * and we need to dereference the EP.
1157          */
1158         if (free)
1159                 put_ep(&ep->com);
1160
1161         PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1162         return ret;
1163 }