Merge branch 'eseries' into pxa
[linux-2.6] / 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 "iwch_provider.h"
33 #include "iwch.h"
34 #include "iwch_cm.h"
35 #include "cxio_hal.h"
36
37 #define NO_SUPPORT -1
38
39 static int iwch_build_rdma_send(union t3_wr *wqe, struct ib_send_wr *wr,
40                                 u8 * flit_cnt)
41 {
42         int i;
43         u32 plen;
44
45         switch (wr->opcode) {
46         case IB_WR_SEND:
47         case IB_WR_SEND_WITH_IMM:
48                 if (wr->send_flags & IB_SEND_SOLICITED)
49                         wqe->send.rdmaop = T3_SEND_WITH_SE;
50                 else
51                         wqe->send.rdmaop = T3_SEND;
52                 wqe->send.rem_stag = 0;
53                 break;
54 #if 0                           /* Not currently supported */
55         case TYPE_SEND_INVALIDATE:
56         case TYPE_SEND_INVALIDATE_IMMEDIATE:
57                 wqe->send.rdmaop = T3_SEND_WITH_INV;
58                 wqe->send.rem_stag = cpu_to_be32(wr->wr.rdma.rkey);
59                 break;
60         case TYPE_SEND_SE_INVALIDATE:
61                 wqe->send.rdmaop = T3_SEND_WITH_SE_INV;
62                 wqe->send.rem_stag = cpu_to_be32(wr->wr.rdma.rkey);
63                 break;
64 #endif
65         default:
66                 break;
67         }
68         if (wr->num_sge > T3_MAX_SGE)
69                 return -EINVAL;
70         wqe->send.reserved[0] = 0;
71         wqe->send.reserved[1] = 0;
72         wqe->send.reserved[2] = 0;
73         if (wr->opcode == IB_WR_SEND_WITH_IMM) {
74                 plen = 4;
75                 wqe->send.sgl[0].stag = wr->ex.imm_data;
76                 wqe->send.sgl[0].len = __constant_cpu_to_be32(0);
77                 wqe->send.num_sgle = __constant_cpu_to_be32(0);
78                 *flit_cnt = 5;
79         } else {
80                 plen = 0;
81                 for (i = 0; i < wr->num_sge; i++) {
82                         if ((plen + wr->sg_list[i].length) < plen) {
83                                 return -EMSGSIZE;
84                         }
85                         plen += wr->sg_list[i].length;
86                         wqe->send.sgl[i].stag =
87                             cpu_to_be32(wr->sg_list[i].lkey);
88                         wqe->send.sgl[i].len =
89                             cpu_to_be32(wr->sg_list[i].length);
90                         wqe->send.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
91                 }
92                 wqe->send.num_sgle = cpu_to_be32(wr->num_sge);
93                 *flit_cnt = 4 + ((wr->num_sge) << 1);
94         }
95         wqe->send.plen = cpu_to_be32(plen);
96         return 0;
97 }
98
99 static int iwch_build_rdma_write(union t3_wr *wqe, struct ib_send_wr *wr,
100                                  u8 *flit_cnt)
101 {
102         int i;
103         u32 plen;
104         if (wr->num_sge > T3_MAX_SGE)
105                 return -EINVAL;
106         wqe->write.rdmaop = T3_RDMA_WRITE;
107         wqe->write.reserved[0] = 0;
108         wqe->write.reserved[1] = 0;
109         wqe->write.reserved[2] = 0;
110         wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
111         wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
112
113         if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
114                 plen = 4;
115                 wqe->write.sgl[0].stag = wr->ex.imm_data;
116                 wqe->write.sgl[0].len = __constant_cpu_to_be32(0);
117                 wqe->write.num_sgle = __constant_cpu_to_be32(0);
118                 *flit_cnt = 6;
119         } else {
120                 plen = 0;
121                 for (i = 0; i < wr->num_sge; i++) {
122                         if ((plen + wr->sg_list[i].length) < plen) {
123                                 return -EMSGSIZE;
124                         }
125                         plen += wr->sg_list[i].length;
126                         wqe->write.sgl[i].stag =
127                             cpu_to_be32(wr->sg_list[i].lkey);
128                         wqe->write.sgl[i].len =
129                             cpu_to_be32(wr->sg_list[i].length);
130                         wqe->write.sgl[i].to =
131                             cpu_to_be64(wr->sg_list[i].addr);
132                 }
133                 wqe->write.num_sgle = cpu_to_be32(wr->num_sge);
134                 *flit_cnt = 5 + ((wr->num_sge) << 1);
135         }
136         wqe->write.plen = cpu_to_be32(plen);
137         return 0;
138 }
139
140 static int iwch_build_rdma_read(union t3_wr *wqe, struct ib_send_wr *wr,
141                                 u8 *flit_cnt)
142 {
143         if (wr->num_sge > 1)
144                 return -EINVAL;
145         wqe->read.rdmaop = T3_READ_REQ;
146         wqe->read.reserved[0] = 0;
147         wqe->read.reserved[1] = 0;
148         wqe->read.reserved[2] = 0;
149         wqe->read.rem_stag = cpu_to_be32(wr->wr.rdma.rkey);
150         wqe->read.rem_to = cpu_to_be64(wr->wr.rdma.remote_addr);
151         wqe->read.local_stag = cpu_to_be32(wr->sg_list[0].lkey);
152         wqe->read.local_len = cpu_to_be32(wr->sg_list[0].length);
153         wqe->read.local_to = cpu_to_be64(wr->sg_list[0].addr);
154         *flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
155         return 0;
156 }
157
158 /*
159  * TBD: this is going to be moved to firmware. Missing pdid/qpid check for now.
160  */
161 static int iwch_sgl2pbl_map(struct iwch_dev *rhp, struct ib_sge *sg_list,
162                             u32 num_sgle, u32 * pbl_addr, u8 * page_size)
163 {
164         int i;
165         struct iwch_mr *mhp;
166         u32 offset;
167         for (i = 0; i < num_sgle; i++) {
168
169                 mhp = get_mhp(rhp, (sg_list[i].lkey) >> 8);
170                 if (!mhp) {
171                         PDBG("%s %d\n", __func__, __LINE__);
172                         return -EIO;
173                 }
174                 if (!mhp->attr.state) {
175                         PDBG("%s %d\n", __func__, __LINE__);
176                         return -EIO;
177                 }
178                 if (mhp->attr.zbva) {
179                         PDBG("%s %d\n", __func__, __LINE__);
180                         return -EIO;
181                 }
182
183                 if (sg_list[i].addr < mhp->attr.va_fbo) {
184                         PDBG("%s %d\n", __func__, __LINE__);
185                         return -EINVAL;
186                 }
187                 if (sg_list[i].addr + ((u64) sg_list[i].length) <
188                     sg_list[i].addr) {
189                         PDBG("%s %d\n", __func__, __LINE__);
190                         return -EINVAL;
191                 }
192                 if (sg_list[i].addr + ((u64) sg_list[i].length) >
193                     mhp->attr.va_fbo + ((u64) mhp->attr.len)) {
194                         PDBG("%s %d\n", __func__, __LINE__);
195                         return -EINVAL;
196                 }
197                 offset = sg_list[i].addr - mhp->attr.va_fbo;
198                 offset += ((u32) mhp->attr.va_fbo) %
199                           (1UL << (12 + mhp->attr.page_size));
200                 pbl_addr[i] = ((mhp->attr.pbl_addr -
201                                 rhp->rdev.rnic_info.pbl_base) >> 3) +
202                               (offset >> (12 + mhp->attr.page_size));
203                 page_size[i] = mhp->attr.page_size;
204         }
205         return 0;
206 }
207
208 static int iwch_build_rdma_recv(struct iwch_dev *rhp, union t3_wr *wqe,
209                                 struct ib_recv_wr *wr)
210 {
211         int i;
212         if (wr->num_sge > T3_MAX_SGE)
213                 return -EINVAL;
214         wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
215         for (i = 0; i < wr->num_sge; i++) {
216                 wqe->recv.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
217                 wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
218                 wqe->recv.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
219         }
220         for (; i < T3_MAX_SGE; i++) {
221                 wqe->recv.sgl[i].stag = 0;
222                 wqe->recv.sgl[i].len = 0;
223                 wqe->recv.sgl[i].to = 0;
224         }
225         return 0;
226 }
227
228 int iwch_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
229                       struct ib_send_wr **bad_wr)
230 {
231         int err = 0;
232         u8 uninitialized_var(t3_wr_flit_cnt);
233         enum t3_wr_opcode t3_wr_opcode = 0;
234         enum t3_wr_flags t3_wr_flags;
235         struct iwch_qp *qhp;
236         u32 idx;
237         union t3_wr *wqe;
238         u32 num_wrs;
239         unsigned long flag;
240         struct t3_swsq *sqp;
241
242         qhp = to_iwch_qp(ibqp);
243         spin_lock_irqsave(&qhp->lock, flag);
244         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
245                 spin_unlock_irqrestore(&qhp->lock, flag);
246                 return -EINVAL;
247         }
248         num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
249                   qhp->wq.sq_size_log2);
250         if (num_wrs <= 0) {
251                 spin_unlock_irqrestore(&qhp->lock, flag);
252                 return -ENOMEM;
253         }
254         while (wr) {
255                 if (num_wrs == 0) {
256                         err = -ENOMEM;
257                         *bad_wr = wr;
258                         break;
259                 }
260                 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
261                 wqe = (union t3_wr *) (qhp->wq.queue + idx);
262                 t3_wr_flags = 0;
263                 if (wr->send_flags & IB_SEND_SOLICITED)
264                         t3_wr_flags |= T3_SOLICITED_EVENT_FLAG;
265                 if (wr->send_flags & IB_SEND_FENCE)
266                         t3_wr_flags |= T3_READ_FENCE_FLAG;
267                 if (wr->send_flags & IB_SEND_SIGNALED)
268                         t3_wr_flags |= T3_COMPLETION_FLAG;
269                 sqp = qhp->wq.sq +
270                       Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
271                 switch (wr->opcode) {
272                 case IB_WR_SEND:
273                 case IB_WR_SEND_WITH_IMM:
274                         t3_wr_opcode = T3_WR_SEND;
275                         err = iwch_build_rdma_send(wqe, wr, &t3_wr_flit_cnt);
276                         break;
277                 case IB_WR_RDMA_WRITE:
278                 case IB_WR_RDMA_WRITE_WITH_IMM:
279                         t3_wr_opcode = T3_WR_WRITE;
280                         err = iwch_build_rdma_write(wqe, wr, &t3_wr_flit_cnt);
281                         break;
282                 case IB_WR_RDMA_READ:
283                         t3_wr_opcode = T3_WR_READ;
284                         t3_wr_flags = 0; /* T3 reads are always signaled */
285                         err = iwch_build_rdma_read(wqe, wr, &t3_wr_flit_cnt);
286                         if (err)
287                                 break;
288                         sqp->read_len = wqe->read.local_len;
289                         if (!qhp->wq.oldest_read)
290                                 qhp->wq.oldest_read = sqp;
291                         break;
292                 default:
293                         PDBG("%s post of type=%d TBD!\n", __func__,
294                              wr->opcode);
295                         err = -EINVAL;
296                 }
297                 if (err) {
298                         *bad_wr = wr;
299                         break;
300                 }
301                 wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
302                 sqp->wr_id = wr->wr_id;
303                 sqp->opcode = wr2opcode(t3_wr_opcode);
304                 sqp->sq_wptr = qhp->wq.sq_wptr;
305                 sqp->complete = 0;
306                 sqp->signaled = (wr->send_flags & IB_SEND_SIGNALED);
307
308                 build_fw_riwrh((void *) wqe, t3_wr_opcode, t3_wr_flags,
309                                Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
310                                0, t3_wr_flit_cnt);
311                 PDBG("%s cookie 0x%llx wq idx 0x%x swsq idx %ld opcode %d\n",
312                      __func__, (unsigned long long) wr->wr_id, idx,
313                      Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2),
314                      sqp->opcode);
315                 wr = wr->next;
316                 num_wrs--;
317                 ++(qhp->wq.wptr);
318                 ++(qhp->wq.sq_wptr);
319         }
320         spin_unlock_irqrestore(&qhp->lock, flag);
321         ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
322         return err;
323 }
324
325 int iwch_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
326                       struct ib_recv_wr **bad_wr)
327 {
328         int err = 0;
329         struct iwch_qp *qhp;
330         u32 idx;
331         union t3_wr *wqe;
332         u32 num_wrs;
333         unsigned long flag;
334
335         qhp = to_iwch_qp(ibqp);
336         spin_lock_irqsave(&qhp->lock, flag);
337         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
338                 spin_unlock_irqrestore(&qhp->lock, flag);
339                 return -EINVAL;
340         }
341         num_wrs = Q_FREECNT(qhp->wq.rq_rptr, qhp->wq.rq_wptr,
342                             qhp->wq.rq_size_log2) - 1;
343         if (!wr) {
344                 spin_unlock_irqrestore(&qhp->lock, flag);
345                 return -EINVAL;
346         }
347         while (wr) {
348                 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
349                 wqe = (union t3_wr *) (qhp->wq.queue + idx);
350                 if (num_wrs)
351                         err = iwch_build_rdma_recv(qhp->rhp, wqe, wr);
352                 else
353                         err = -ENOMEM;
354                 if (err) {
355                         *bad_wr = wr;
356                         break;
357                 }
358                 qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr, qhp->wq.rq_size_log2)] =
359                         wr->wr_id;
360                 build_fw_riwrh((void *) wqe, T3_WR_RCV, T3_COMPLETION_FLAG,
361                                Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
362                                0, sizeof(struct t3_receive_wr) >> 3);
363                 PDBG("%s cookie 0x%llx idx 0x%x rq_wptr 0x%x rw_rptr 0x%x "
364                      "wqe %p \n", __func__, (unsigned long long) wr->wr_id,
365                      idx, qhp->wq.rq_wptr, qhp->wq.rq_rptr, wqe);
366                 ++(qhp->wq.rq_wptr);
367                 ++(qhp->wq.wptr);
368                 wr = wr->next;
369                 num_wrs--;
370         }
371         spin_unlock_irqrestore(&qhp->lock, flag);
372         ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
373         return err;
374 }
375
376 int iwch_bind_mw(struct ib_qp *qp,
377                              struct ib_mw *mw,
378                              struct ib_mw_bind *mw_bind)
379 {
380         struct iwch_dev *rhp;
381         struct iwch_mw *mhp;
382         struct iwch_qp *qhp;
383         union t3_wr *wqe;
384         u32 pbl_addr;
385         u8 page_size;
386         u32 num_wrs;
387         unsigned long flag;
388         struct ib_sge sgl;
389         int err=0;
390         enum t3_wr_flags t3_wr_flags;
391         u32 idx;
392         struct t3_swsq *sqp;
393
394         qhp = to_iwch_qp(qp);
395         mhp = to_iwch_mw(mw);
396         rhp = qhp->rhp;
397
398         spin_lock_irqsave(&qhp->lock, flag);
399         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
400                 spin_unlock_irqrestore(&qhp->lock, flag);
401                 return -EINVAL;
402         }
403         num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
404                             qhp->wq.sq_size_log2);
405         if ((num_wrs) <= 0) {
406                 spin_unlock_irqrestore(&qhp->lock, flag);
407                 return -ENOMEM;
408         }
409         idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
410         PDBG("%s: idx 0x%0x, mw 0x%p, mw_bind 0x%p\n", __func__, idx,
411              mw, mw_bind);
412         wqe = (union t3_wr *) (qhp->wq.queue + idx);
413
414         t3_wr_flags = 0;
415         if (mw_bind->send_flags & IB_SEND_SIGNALED)
416                 t3_wr_flags = T3_COMPLETION_FLAG;
417
418         sgl.addr = mw_bind->addr;
419         sgl.lkey = mw_bind->mr->lkey;
420         sgl.length = mw_bind->length;
421         wqe->bind.reserved = 0;
422         wqe->bind.type = T3_VA_BASED_TO;
423
424         /* TBD: check perms */
425         wqe->bind.perms = iwch_ib_to_mwbind_access(mw_bind->mw_access_flags);
426         wqe->bind.mr_stag = cpu_to_be32(mw_bind->mr->lkey);
427         wqe->bind.mw_stag = cpu_to_be32(mw->rkey);
428         wqe->bind.mw_len = cpu_to_be32(mw_bind->length);
429         wqe->bind.mw_va = cpu_to_be64(mw_bind->addr);
430         err = iwch_sgl2pbl_map(rhp, &sgl, 1, &pbl_addr, &page_size);
431         if (err) {
432                 spin_unlock_irqrestore(&qhp->lock, flag);
433                 return err;
434         }
435         wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
436         sqp = qhp->wq.sq + Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
437         sqp->wr_id = mw_bind->wr_id;
438         sqp->opcode = T3_BIND_MW;
439         sqp->sq_wptr = qhp->wq.sq_wptr;
440         sqp->complete = 0;
441         sqp->signaled = (mw_bind->send_flags & IB_SEND_SIGNALED);
442         wqe->bind.mr_pbl_addr = cpu_to_be32(pbl_addr);
443         wqe->bind.mr_pagesz = page_size;
444         wqe->flit[T3_SQ_COOKIE_FLIT] = mw_bind->wr_id;
445         build_fw_riwrh((void *)wqe, T3_WR_BIND, t3_wr_flags,
446                        Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2), 0,
447                                 sizeof(struct t3_bind_mw_wr) >> 3);
448         ++(qhp->wq.wptr);
449         ++(qhp->wq.sq_wptr);
450         spin_unlock_irqrestore(&qhp->lock, flag);
451
452         ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
453
454         return err;
455 }
456
457 static inline void build_term_codes(struct respQ_msg_t *rsp_msg,
458                                     u8 *layer_type, u8 *ecode)
459 {
460         int status = TPT_ERR_INTERNAL_ERR;
461         int tagged = 0;
462         int opcode = -1;
463         int rqtype = 0;
464         int send_inv = 0;
465
466         if (rsp_msg) {
467                 status = CQE_STATUS(rsp_msg->cqe);
468                 opcode = CQE_OPCODE(rsp_msg->cqe);
469                 rqtype = RQ_TYPE(rsp_msg->cqe);
470                 send_inv = (opcode == T3_SEND_WITH_INV) ||
471                            (opcode == T3_SEND_WITH_SE_INV);
472                 tagged = (opcode == T3_RDMA_WRITE) ||
473                          (rqtype && (opcode == T3_READ_RESP));
474         }
475
476         switch (status) {
477         case TPT_ERR_STAG:
478                 if (send_inv) {
479                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
480                         *ecode = RDMAP_CANT_INV_STAG;
481                 } else {
482                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
483                         *ecode = RDMAP_INV_STAG;
484                 }
485                 break;
486         case TPT_ERR_PDID:
487                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
488                 if ((opcode == T3_SEND_WITH_INV) ||
489                     (opcode == T3_SEND_WITH_SE_INV))
490                         *ecode = RDMAP_CANT_INV_STAG;
491                 else
492                         *ecode = RDMAP_STAG_NOT_ASSOC;
493                 break;
494         case TPT_ERR_QPID:
495                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
496                 *ecode = RDMAP_STAG_NOT_ASSOC;
497                 break;
498         case TPT_ERR_ACCESS:
499                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
500                 *ecode = RDMAP_ACC_VIOL;
501                 break;
502         case TPT_ERR_WRAP:
503                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
504                 *ecode = RDMAP_TO_WRAP;
505                 break;
506         case TPT_ERR_BOUND:
507                 if (tagged) {
508                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
509                         *ecode = DDPT_BASE_BOUNDS;
510                 } else {
511                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
512                         *ecode = RDMAP_BASE_BOUNDS;
513                 }
514                 break;
515         case TPT_ERR_INVALIDATE_SHARED_MR:
516         case TPT_ERR_INVALIDATE_MR_WITH_MW_BOUND:
517                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
518                 *ecode = RDMAP_CANT_INV_STAG;
519                 break;
520         case TPT_ERR_ECC:
521         case TPT_ERR_ECC_PSTAG:
522         case TPT_ERR_INTERNAL_ERR:
523                 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
524                 *ecode = 0;
525                 break;
526         case TPT_ERR_OUT_OF_RQE:
527                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
528                 *ecode = DDPU_INV_MSN_NOBUF;
529                 break;
530         case TPT_ERR_PBL_ADDR_BOUND:
531                 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
532                 *ecode = DDPT_BASE_BOUNDS;
533                 break;
534         case TPT_ERR_CRC:
535                 *layer_type = LAYER_MPA|DDP_LLP;
536                 *ecode = MPA_CRC_ERR;
537                 break;
538         case TPT_ERR_MARKER:
539                 *layer_type = LAYER_MPA|DDP_LLP;
540                 *ecode = MPA_MARKER_ERR;
541                 break;
542         case TPT_ERR_PDU_LEN_ERR:
543                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
544                 *ecode = DDPU_MSG_TOOBIG;
545                 break;
546         case TPT_ERR_DDP_VERSION:
547                 if (tagged) {
548                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
549                         *ecode = DDPT_INV_VERS;
550                 } else {
551                         *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
552                         *ecode = DDPU_INV_VERS;
553                 }
554                 break;
555         case TPT_ERR_RDMA_VERSION:
556                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
557                 *ecode = RDMAP_INV_VERS;
558                 break;
559         case TPT_ERR_OPCODE:
560                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
561                 *ecode = RDMAP_INV_OPCODE;
562                 break;
563         case TPT_ERR_DDP_QUEUE_NUM:
564                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
565                 *ecode = DDPU_INV_QN;
566                 break;
567         case TPT_ERR_MSN:
568         case TPT_ERR_MSN_GAP:
569         case TPT_ERR_MSN_RANGE:
570         case TPT_ERR_IRD_OVERFLOW:
571                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
572                 *ecode = DDPU_INV_MSN_RANGE;
573                 break;
574         case TPT_ERR_TBIT:
575                 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
576                 *ecode = 0;
577                 break;
578         case TPT_ERR_MO:
579                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
580                 *ecode = DDPU_INV_MO;
581                 break;
582         default:
583                 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
584                 *ecode = 0;
585                 break;
586         }
587 }
588
589 int iwch_post_zb_read(struct iwch_qp *qhp)
590 {
591         union t3_wr *wqe;
592         struct sk_buff *skb;
593         u8 flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
594
595         PDBG("%s enter\n", __func__);
596         skb = alloc_skb(40, GFP_KERNEL);
597         if (!skb) {
598                 printk(KERN_ERR "%s cannot send zb_read!!\n", __func__);
599                 return -ENOMEM;
600         }
601         wqe = (union t3_wr *)skb_put(skb, sizeof(struct t3_rdma_read_wr));
602         memset(wqe, 0, sizeof(struct t3_rdma_read_wr));
603         wqe->read.rdmaop = T3_READ_REQ;
604         wqe->read.reserved[0] = 0;
605         wqe->read.reserved[1] = 0;
606         wqe->read.reserved[2] = 0;
607         wqe->read.rem_stag = cpu_to_be32(1);
608         wqe->read.rem_to = cpu_to_be64(1);
609         wqe->read.local_stag = cpu_to_be32(1);
610         wqe->read.local_len = cpu_to_be32(0);
611         wqe->read.local_to = cpu_to_be64(1);
612         wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_READ));
613         wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(qhp->ep->hwtid)|
614                                                 V_FW_RIWR_LEN(flit_cnt));
615         skb->priority = CPL_PRIORITY_DATA;
616         return cxgb3_ofld_send(qhp->rhp->rdev.t3cdev_p, skb);
617 }
618
619 /*
620  * This posts a TERMINATE with layer=RDMA, type=catastrophic.
621  */
622 int iwch_post_terminate(struct iwch_qp *qhp, struct respQ_msg_t *rsp_msg)
623 {
624         union t3_wr *wqe;
625         struct terminate_message *term;
626         struct sk_buff *skb;
627
628         PDBG("%s %d\n", __func__, __LINE__);
629         skb = alloc_skb(40, GFP_ATOMIC);
630         if (!skb) {
631                 printk(KERN_ERR "%s cannot send TERMINATE!\n", __func__);
632                 return -ENOMEM;
633         }
634         wqe = (union t3_wr *)skb_put(skb, 40);
635         memset(wqe, 0, 40);
636         wqe->send.rdmaop = T3_TERMINATE;
637
638         /* immediate data length */
639         wqe->send.plen = htonl(4);
640
641         /* immediate data starts here. */
642         term = (struct terminate_message *)wqe->send.sgl;
643         build_term_codes(rsp_msg, &term->layer_etype, &term->ecode);
644         wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_SEND) |
645                          V_FW_RIWR_FLAGS(T3_COMPLETION_FLAG | T3_NOTIFY_FLAG));
646         wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(qhp->ep->hwtid));
647         skb->priority = CPL_PRIORITY_DATA;
648         return cxgb3_ofld_send(qhp->rhp->rdev.t3cdev_p, skb);
649 }
650
651 /*
652  * Assumes qhp lock is held.
653  */
654 static void __flush_qp(struct iwch_qp *qhp, unsigned long *flag)
655 {
656         struct iwch_cq *rchp, *schp;
657         int count;
658         int flushed;
659
660         rchp = get_chp(qhp->rhp, qhp->attr.rcq);
661         schp = get_chp(qhp->rhp, qhp->attr.scq);
662
663         PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
664         /* take a ref on the qhp since we must release the lock */
665         atomic_inc(&qhp->refcnt);
666         spin_unlock_irqrestore(&qhp->lock, *flag);
667
668         /* locking heirarchy: cq lock first, then qp lock. */
669         spin_lock_irqsave(&rchp->lock, *flag);
670         spin_lock(&qhp->lock);
671         cxio_flush_hw_cq(&rchp->cq);
672         cxio_count_rcqes(&rchp->cq, &qhp->wq, &count);
673         flushed = cxio_flush_rq(&qhp->wq, &rchp->cq, count);
674         spin_unlock(&qhp->lock);
675         spin_unlock_irqrestore(&rchp->lock, *flag);
676         if (flushed)
677                 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
678
679         /* locking heirarchy: cq lock first, then qp lock. */
680         spin_lock_irqsave(&schp->lock, *flag);
681         spin_lock(&qhp->lock);
682         cxio_flush_hw_cq(&schp->cq);
683         cxio_count_scqes(&schp->cq, &qhp->wq, &count);
684         flushed = cxio_flush_sq(&qhp->wq, &schp->cq, count);
685         spin_unlock(&qhp->lock);
686         spin_unlock_irqrestore(&schp->lock, *flag);
687         if (flushed)
688                 (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
689
690         /* deref */
691         if (atomic_dec_and_test(&qhp->refcnt))
692                 wake_up(&qhp->wait);
693
694         spin_lock_irqsave(&qhp->lock, *flag);
695 }
696
697 static void flush_qp(struct iwch_qp *qhp, unsigned long *flag)
698 {
699         if (qhp->ibqp.uobject)
700                 cxio_set_wq_in_error(&qhp->wq);
701         else
702                 __flush_qp(qhp, flag);
703 }
704
705
706 /*
707  * Return count of RECV WRs posted
708  */
709 u16 iwch_rqes_posted(struct iwch_qp *qhp)
710 {
711         union t3_wr *wqe = qhp->wq.queue;
712         u16 count = 0;
713         while ((count+1) != 0 && fw_riwrh_opcode((struct fw_riwrh *)wqe) == T3_WR_RCV) {
714                 count++;
715                 wqe++;
716         }
717         PDBG("%s qhp %p count %u\n", __func__, qhp, count);
718         return count;
719 }
720
721 static int rdma_init(struct iwch_dev *rhp, struct iwch_qp *qhp,
722                                 enum iwch_qp_attr_mask mask,
723                                 struct iwch_qp_attributes *attrs)
724 {
725         struct t3_rdma_init_attr init_attr;
726         int ret;
727
728         init_attr.tid = qhp->ep->hwtid;
729         init_attr.qpid = qhp->wq.qpid;
730         init_attr.pdid = qhp->attr.pd;
731         init_attr.scqid = qhp->attr.scq;
732         init_attr.rcqid = qhp->attr.rcq;
733         init_attr.rq_addr = qhp->wq.rq_addr;
734         init_attr.rq_size = 1 << qhp->wq.rq_size_log2;
735         init_attr.mpaattrs = uP_RI_MPA_IETF_ENABLE |
736                 qhp->attr.mpa_attr.recv_marker_enabled |
737                 (qhp->attr.mpa_attr.xmit_marker_enabled << 1) |
738                 (qhp->attr.mpa_attr.crc_enabled << 2);
739
740         /*
741          * XXX - The IWCM doesn't quite handle getting these
742          * attrs set before going into RTS.  For now, just turn
743          * them on always...
744          */
745 #if 0
746         init_attr.qpcaps = qhp->attr.enableRdmaRead |
747                 (qhp->attr.enableRdmaWrite << 1) |
748                 (qhp->attr.enableBind << 2) |
749                 (qhp->attr.enable_stag0_fastreg << 3) |
750                 (qhp->attr.enable_stag0_fastreg << 4);
751 #else
752         init_attr.qpcaps = 0x1f;
753 #endif
754         init_attr.tcp_emss = qhp->ep->emss;
755         init_attr.ord = qhp->attr.max_ord;
756         init_attr.ird = qhp->attr.max_ird;
757         init_attr.qp_dma_addr = qhp->wq.dma_addr;
758         init_attr.qp_dma_size = (1UL << qhp->wq.size_log2);
759         init_attr.rqe_count = iwch_rqes_posted(qhp);
760         init_attr.flags = qhp->attr.mpa_attr.initiator ? MPA_INITIATOR : 0;
761         init_attr.flags |= capable(CAP_NET_BIND_SERVICE) ? PRIV_QP : 0;
762         if (peer2peer) {
763                 init_attr.rtr_type = RTR_READ;
764                 if (init_attr.ord == 0 && qhp->attr.mpa_attr.initiator)
765                         init_attr.ord = 1;
766                 if (init_attr.ird == 0 && !qhp->attr.mpa_attr.initiator)
767                         init_attr.ird = 1;
768         } else
769                 init_attr.rtr_type = 0;
770         init_attr.irs = qhp->ep->rcv_seq;
771         PDBG("%s init_attr.rq_addr 0x%x init_attr.rq_size = %d "
772              "flags 0x%x qpcaps 0x%x\n", __func__,
773              init_attr.rq_addr, init_attr.rq_size,
774              init_attr.flags, init_attr.qpcaps);
775         ret = cxio_rdma_init(&rhp->rdev, &init_attr);
776         PDBG("%s ret %d\n", __func__, ret);
777         return ret;
778 }
779
780 int iwch_modify_qp(struct iwch_dev *rhp, struct iwch_qp *qhp,
781                                 enum iwch_qp_attr_mask mask,
782                                 struct iwch_qp_attributes *attrs,
783                                 int internal)
784 {
785         int ret = 0;
786         struct iwch_qp_attributes newattr = qhp->attr;
787         unsigned long flag;
788         int disconnect = 0;
789         int terminate = 0;
790         int abort = 0;
791         int free = 0;
792         struct iwch_ep *ep = NULL;
793
794         PDBG("%s qhp %p qpid 0x%x ep %p state %d -> %d\n", __func__,
795              qhp, qhp->wq.qpid, qhp->ep, qhp->attr.state,
796              (mask & IWCH_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
797
798         spin_lock_irqsave(&qhp->lock, flag);
799
800         /* Process attr changes if in IDLE */
801         if (mask & IWCH_QP_ATTR_VALID_MODIFY) {
802                 if (qhp->attr.state != IWCH_QP_STATE_IDLE) {
803                         ret = -EIO;
804                         goto out;
805                 }
806                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_READ)
807                         newattr.enable_rdma_read = attrs->enable_rdma_read;
808                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_WRITE)
809                         newattr.enable_rdma_write = attrs->enable_rdma_write;
810                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_BIND)
811                         newattr.enable_bind = attrs->enable_bind;
812                 if (mask & IWCH_QP_ATTR_MAX_ORD) {
813                         if (attrs->max_ord >
814                             rhp->attr.max_rdma_read_qp_depth) {
815                                 ret = -EINVAL;
816                                 goto out;
817                         }
818                         newattr.max_ord = attrs->max_ord;
819                 }
820                 if (mask & IWCH_QP_ATTR_MAX_IRD) {
821                         if (attrs->max_ird >
822                             rhp->attr.max_rdma_reads_per_qp) {
823                                 ret = -EINVAL;
824                                 goto out;
825                         }
826                         newattr.max_ird = attrs->max_ird;
827                 }
828                 qhp->attr = newattr;
829         }
830
831         if (!(mask & IWCH_QP_ATTR_NEXT_STATE))
832                 goto out;
833         if (qhp->attr.state == attrs->next_state)
834                 goto out;
835
836         switch (qhp->attr.state) {
837         case IWCH_QP_STATE_IDLE:
838                 switch (attrs->next_state) {
839                 case IWCH_QP_STATE_RTS:
840                         if (!(mask & IWCH_QP_ATTR_LLP_STREAM_HANDLE)) {
841                                 ret = -EINVAL;
842                                 goto out;
843                         }
844                         if (!(mask & IWCH_QP_ATTR_MPA_ATTR)) {
845                                 ret = -EINVAL;
846                                 goto out;
847                         }
848                         qhp->attr.mpa_attr = attrs->mpa_attr;
849                         qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
850                         qhp->ep = qhp->attr.llp_stream_handle;
851                         qhp->attr.state = IWCH_QP_STATE_RTS;
852
853                         /*
854                          * Ref the endpoint here and deref when we
855                          * disassociate the endpoint from the QP.  This
856                          * happens in CLOSING->IDLE transition or *->ERROR
857                          * transition.
858                          */
859                         get_ep(&qhp->ep->com);
860                         spin_unlock_irqrestore(&qhp->lock, flag);
861                         ret = rdma_init(rhp, qhp, mask, attrs);
862                         spin_lock_irqsave(&qhp->lock, flag);
863                         if (ret)
864                                 goto err;
865                         break;
866                 case IWCH_QP_STATE_ERROR:
867                         qhp->attr.state = IWCH_QP_STATE_ERROR;
868                         flush_qp(qhp, &flag);
869                         break;
870                 default:
871                         ret = -EINVAL;
872                         goto out;
873                 }
874                 break;
875         case IWCH_QP_STATE_RTS:
876                 switch (attrs->next_state) {
877                 case IWCH_QP_STATE_CLOSING:
878                         BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
879                         qhp->attr.state = IWCH_QP_STATE_CLOSING;
880                         if (!internal) {
881                                 abort=0;
882                                 disconnect = 1;
883                                 ep = qhp->ep;
884                                 get_ep(&ep->com);
885                         }
886                         break;
887                 case IWCH_QP_STATE_TERMINATE:
888                         qhp->attr.state = IWCH_QP_STATE_TERMINATE;
889                         if (qhp->ibqp.uobject)
890                                 cxio_set_wq_in_error(&qhp->wq);
891                         if (!internal)
892                                 terminate = 1;
893                         break;
894                 case IWCH_QP_STATE_ERROR:
895                         qhp->attr.state = IWCH_QP_STATE_ERROR;
896                         if (!internal) {
897                                 abort=1;
898                                 disconnect = 1;
899                                 ep = qhp->ep;
900                                 get_ep(&ep->com);
901                         }
902                         goto err;
903                         break;
904                 default:
905                         ret = -EINVAL;
906                         goto out;
907                 }
908                 break;
909         case IWCH_QP_STATE_CLOSING:
910                 if (!internal) {
911                         ret = -EINVAL;
912                         goto out;
913                 }
914                 switch (attrs->next_state) {
915                         case IWCH_QP_STATE_IDLE:
916                                 flush_qp(qhp, &flag);
917                                 qhp->attr.state = IWCH_QP_STATE_IDLE;
918                                 qhp->attr.llp_stream_handle = NULL;
919                                 put_ep(&qhp->ep->com);
920                                 qhp->ep = NULL;
921                                 wake_up(&qhp->wait);
922                                 break;
923                         case IWCH_QP_STATE_ERROR:
924                                 goto err;
925                         default:
926                                 ret = -EINVAL;
927                                 goto err;
928                 }
929                 break;
930         case IWCH_QP_STATE_ERROR:
931                 if (attrs->next_state != IWCH_QP_STATE_IDLE) {
932                         ret = -EINVAL;
933                         goto out;
934                 }
935
936                 if (!Q_EMPTY(qhp->wq.sq_rptr, qhp->wq.sq_wptr) ||
937                     !Q_EMPTY(qhp->wq.rq_rptr, qhp->wq.rq_wptr)) {
938                         ret = -EINVAL;
939                         goto out;
940                 }
941                 qhp->attr.state = IWCH_QP_STATE_IDLE;
942                 memset(&qhp->attr, 0, sizeof(qhp->attr));
943                 break;
944         case IWCH_QP_STATE_TERMINATE:
945                 if (!internal) {
946                         ret = -EINVAL;
947                         goto out;
948                 }
949                 goto err;
950                 break;
951         default:
952                 printk(KERN_ERR "%s in a bad state %d\n",
953                        __func__, qhp->attr.state);
954                 ret = -EINVAL;
955                 goto err;
956                 break;
957         }
958         goto out;
959 err:
960         PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
961              qhp->wq.qpid);
962
963         /* disassociate the LLP connection */
964         qhp->attr.llp_stream_handle = NULL;
965         ep = qhp->ep;
966         qhp->ep = NULL;
967         qhp->attr.state = IWCH_QP_STATE_ERROR;
968         free=1;
969         wake_up(&qhp->wait);
970         BUG_ON(!ep);
971         flush_qp(qhp, &flag);
972 out:
973         spin_unlock_irqrestore(&qhp->lock, flag);
974
975         if (terminate)
976                 iwch_post_terminate(qhp, NULL);
977
978         /*
979          * If disconnect is 1, then we need to initiate a disconnect
980          * on the EP.  This can be a normal close (RTS->CLOSING) or
981          * an abnormal close (RTS/CLOSING->ERROR).
982          */
983         if (disconnect) {
984                 iwch_ep_disconnect(ep, abort, GFP_KERNEL);
985                 put_ep(&ep->com);
986         }
987
988         /*
989          * If free is 1, then we've disassociated the EP from the QP
990          * and we need to dereference the EP.
991          */
992         if (free)
993                 put_ep(&ep->com);
994
995         PDBG("%s exit state %d\n", __func__, qhp->attr.state);
996         return ret;
997 }
998
999 static int quiesce_qp(struct iwch_qp *qhp)
1000 {
1001         spin_lock_irq(&qhp->lock);
1002         iwch_quiesce_tid(qhp->ep);
1003         qhp->flags |= QP_QUIESCED;
1004         spin_unlock_irq(&qhp->lock);
1005         return 0;
1006 }
1007
1008 static int resume_qp(struct iwch_qp *qhp)
1009 {
1010         spin_lock_irq(&qhp->lock);
1011         iwch_resume_tid(qhp->ep);
1012         qhp->flags &= ~QP_QUIESCED;
1013         spin_unlock_irq(&qhp->lock);
1014         return 0;
1015 }
1016
1017 int iwch_quiesce_qps(struct iwch_cq *chp)
1018 {
1019         int i;
1020         struct iwch_qp *qhp;
1021
1022         for (i=0; i < T3_MAX_NUM_QP; i++) {
1023                 qhp = get_qhp(chp->rhp, i);
1024                 if (!qhp)
1025                         continue;
1026                 if ((qhp->attr.rcq == chp->cq.cqid) && !qp_quiesced(qhp)) {
1027                         quiesce_qp(qhp);
1028                         continue;
1029                 }
1030                 if ((qhp->attr.scq == chp->cq.cqid) && !qp_quiesced(qhp))
1031                         quiesce_qp(qhp);
1032         }
1033         return 0;
1034 }
1035
1036 int iwch_resume_qps(struct iwch_cq *chp)
1037 {
1038         int i;
1039         struct iwch_qp *qhp;
1040
1041         for (i=0; i < T3_MAX_NUM_QP; i++) {
1042                 qhp = get_qhp(chp->rhp, i);
1043                 if (!qhp)
1044                         continue;
1045                 if ((qhp->attr.rcq == chp->cq.cqid) && qp_quiesced(qhp)) {
1046                         resume_qp(qhp);
1047                         continue;
1048                 }
1049                 if ((qhp->attr.scq == chp->cq.cqid) && qp_quiesced(qhp))
1050                         resume_qp(qhp);
1051         }
1052         return 0;
1053 }