]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blame - drivers/infiniband/hw/cxgb3/iwch_qp.c
RDMA/cxgb3: Add driver for Chelsio T3 RNIC
[mirror_ubuntu-artful-kernel.git] / drivers / infiniband / hw / cxgb3 / iwch_qp.c
CommitLineData
b038ced7
SW
1/*
2 * Copyright (c) 2006 Chelsio, Inc. All rights reserved.
3 * Copyright (c) 2006 Open Grid Computing, Inc. All rights reserved.
4 *
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
10 *
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
13 * conditions are met:
14 *
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
17 * disclaimer.
18 *
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
23 *
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31 * SOFTWARE.
32 */
33#include "iwch_provider.h"
34#include "iwch.h"
35#include "iwch_cm.h"
36#include "cxio_hal.h"
37
38#define NO_SUPPORT -1
39
40static inline int iwch_build_rdma_send(union t3_wr *wqe, struct ib_send_wr *wr,
41 u8 * flit_cnt)
42{
43 int i;
44 u32 plen;
45
46 switch (wr->opcode) {
47 case IB_WR_SEND:
48 case IB_WR_SEND_WITH_IMM:
49 if (wr->send_flags & IB_SEND_SOLICITED)
50 wqe->send.rdmaop = T3_SEND_WITH_SE;
51 else
52 wqe->send.rdmaop = T3_SEND;
53 wqe->send.rem_stag = 0;
54 break;
55#if 0 /* Not currently supported */
56 case TYPE_SEND_INVALIDATE:
57 case TYPE_SEND_INVALIDATE_IMMEDIATE:
58 wqe->send.rdmaop = T3_SEND_WITH_INV;
59 wqe->send.rem_stag = cpu_to_be32(wr->wr.rdma.rkey);
60 break;
61 case TYPE_SEND_SE_INVALIDATE:
62 wqe->send.rdmaop = T3_SEND_WITH_SE_INV;
63 wqe->send.rem_stag = cpu_to_be32(wr->wr.rdma.rkey);
64 break;
65#endif
66 default:
67 break;
68 }
69 if (wr->num_sge > T3_MAX_SGE)
70 return -EINVAL;
71 wqe->send.reserved[0] = 0;
72 wqe->send.reserved[1] = 0;
73 wqe->send.reserved[2] = 0;
74 if (wr->opcode == IB_WR_SEND_WITH_IMM) {
75 plen = 4;
76 wqe->send.sgl[0].stag = wr->imm_data;
77 wqe->send.sgl[0].len = __constant_cpu_to_be32(0);
78 wqe->send.num_sgle = __constant_cpu_to_be32(0);
79 *flit_cnt = 5;
80 } else {
81 plen = 0;
82 for (i = 0; i < wr->num_sge; i++) {
83 if ((plen + wr->sg_list[i].length) < plen) {
84 return -EMSGSIZE;
85 }
86 plen += wr->sg_list[i].length;
87 wqe->send.sgl[i].stag =
88 cpu_to_be32(wr->sg_list[i].lkey);
89 wqe->send.sgl[i].len =
90 cpu_to_be32(wr->sg_list[i].length);
91 wqe->send.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
92 }
93 wqe->send.num_sgle = cpu_to_be32(wr->num_sge);
94 *flit_cnt = 4 + ((wr->num_sge) << 1);
95 }
96 wqe->send.plen = cpu_to_be32(plen);
97 return 0;
98}
99
100static inline int iwch_build_rdma_write(union t3_wr *wqe, struct ib_send_wr *wr,
101 u8 *flit_cnt)
102{
103 int i;
104 u32 plen;
105 if (wr->num_sge > T3_MAX_SGE)
106 return -EINVAL;
107 wqe->write.rdmaop = T3_RDMA_WRITE;
108 wqe->write.reserved[0] = 0;
109 wqe->write.reserved[1] = 0;
110 wqe->write.reserved[2] = 0;
111 wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
112 wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
113
114 if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
115 plen = 4;
116 wqe->write.sgl[0].stag = wr->imm_data;
117 wqe->write.sgl[0].len = __constant_cpu_to_be32(0);
118 wqe->write.num_sgle = __constant_cpu_to_be32(0);
119 *flit_cnt = 6;
120 } else {
121 plen = 0;
122 for (i = 0; i < wr->num_sge; i++) {
123 if ((plen + wr->sg_list[i].length) < plen) {
124 return -EMSGSIZE;
125 }
126 plen += wr->sg_list[i].length;
127 wqe->write.sgl[i].stag =
128 cpu_to_be32(wr->sg_list[i].lkey);
129 wqe->write.sgl[i].len =
130 cpu_to_be32(wr->sg_list[i].length);
131 wqe->write.sgl[i].to =
132 cpu_to_be64(wr->sg_list[i].addr);
133 }
134 wqe->write.num_sgle = cpu_to_be32(wr->num_sge);
135 *flit_cnt = 5 + ((wr->num_sge) << 1);
136 }
137 wqe->write.plen = cpu_to_be32(plen);
138 return 0;
139}
140
141static inline int iwch_build_rdma_read(union t3_wr *wqe, struct ib_send_wr *wr,
142 u8 *flit_cnt)
143{
144 if (wr->num_sge > 1)
145 return -EINVAL;
146 wqe->read.rdmaop = T3_READ_REQ;
147 wqe->read.reserved[0] = 0;
148 wqe->read.reserved[1] = 0;
149 wqe->read.reserved[2] = 0;
150 wqe->read.rem_stag = cpu_to_be32(wr->wr.rdma.rkey);
151 wqe->read.rem_to = cpu_to_be64(wr->wr.rdma.remote_addr);
152 wqe->read.local_stag = cpu_to_be32(wr->sg_list[0].lkey);
153 wqe->read.local_len = cpu_to_be32(wr->sg_list[0].length);
154 wqe->read.local_to = cpu_to_be64(wr->sg_list[0].addr);
155 *flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
156 return 0;
157}
158
159/*
160 * TBD: this is going to be moved to firmware. Missing pdid/qpid check for now.
161 */
162static inline int iwch_sgl2pbl_map(struct iwch_dev *rhp,
163 struct ib_sge *sg_list, u32 num_sgle,
164 u32 * pbl_addr, u8 * page_size)
165{
166 int i;
167 struct iwch_mr *mhp;
168 u32 offset;
169 for (i = 0; i < num_sgle; i++) {
170
171 mhp = get_mhp(rhp, (sg_list[i].lkey) >> 8);
172 if (!mhp) {
173 PDBG("%s %d\n", __FUNCTION__, __LINE__);
174 return -EIO;
175 }
176 if (!mhp->attr.state) {
177 PDBG("%s %d\n", __FUNCTION__, __LINE__);
178 return -EIO;
179 }
180 if (mhp->attr.zbva) {
181 PDBG("%s %d\n", __FUNCTION__, __LINE__);
182 return -EIO;
183 }
184
185 if (sg_list[i].addr < mhp->attr.va_fbo) {
186 PDBG("%s %d\n", __FUNCTION__, __LINE__);
187 return -EINVAL;
188 }
189 if (sg_list[i].addr + ((u64) sg_list[i].length) <
190 sg_list[i].addr) {
191 PDBG("%s %d\n", __FUNCTION__, __LINE__);
192 return -EINVAL;
193 }
194 if (sg_list[i].addr + ((u64) sg_list[i].length) >
195 mhp->attr.va_fbo + ((u64) mhp->attr.len)) {
196 PDBG("%s %d\n", __FUNCTION__, __LINE__);
197 return -EINVAL;
198 }
199 offset = sg_list[i].addr - mhp->attr.va_fbo;
200 offset += ((u32) mhp->attr.va_fbo) %
201 (1UL << (12 + mhp->attr.page_size));
202 pbl_addr[i] = ((mhp->attr.pbl_addr -
203 rhp->rdev.rnic_info.pbl_base) >> 3) +
204 (offset >> (12 + mhp->attr.page_size));
205 page_size[i] = mhp->attr.page_size;
206 }
207 return 0;
208}
209
210static inline int iwch_build_rdma_recv(struct iwch_dev *rhp,
211 union t3_wr *wqe,
212 struct ib_recv_wr *wr)
213{
214 int i, err = 0;
215 u32 pbl_addr[4];
216 u8 page_size[4];
217 if (wr->num_sge > T3_MAX_SGE)
218 return -EINVAL;
219 err = iwch_sgl2pbl_map(rhp, wr->sg_list, wr->num_sge, pbl_addr,
220 page_size);
221 if (err)
222 return err;
223 wqe->recv.pagesz[0] = page_size[0];
224 wqe->recv.pagesz[1] = page_size[1];
225 wqe->recv.pagesz[2] = page_size[2];
226 wqe->recv.pagesz[3] = page_size[3];
227 wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
228 for (i = 0; i < wr->num_sge; i++) {
229 wqe->recv.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
230 wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
231
232 /* to in the WQE == the offset into the page */
233 wqe->recv.sgl[i].to = cpu_to_be64(((u32) wr->sg_list[i].addr) %
234 (1UL << (12 + page_size[i])));
235
236 /* pbl_addr is the adapters address in the PBL */
237 wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_addr[i]);
238 }
239 for (; i < T3_MAX_SGE; i++) {
240 wqe->recv.sgl[i].stag = 0;
241 wqe->recv.sgl[i].len = 0;
242 wqe->recv.sgl[i].to = 0;
243 wqe->recv.pbl_addr[i] = 0;
244 }
245 return 0;
246}
247
248int iwch_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
249 struct ib_send_wr **bad_wr)
250{
251 int err = 0;
252 u8 t3_wr_flit_cnt;
253 enum t3_wr_opcode t3_wr_opcode = 0;
254 enum t3_wr_flags t3_wr_flags;
255 struct iwch_qp *qhp;
256 u32 idx;
257 union t3_wr *wqe;
258 u32 num_wrs;
259 unsigned long flag;
260 struct t3_swsq *sqp;
261
262 qhp = to_iwch_qp(ibqp);
263 spin_lock_irqsave(&qhp->lock, flag);
264 if (qhp->attr.state > IWCH_QP_STATE_RTS) {
265 spin_unlock_irqrestore(&qhp->lock, flag);
266 return -EINVAL;
267 }
268 num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
269 qhp->wq.sq_size_log2);
270 if (num_wrs <= 0) {
271 spin_unlock_irqrestore(&qhp->lock, flag);
272 return -ENOMEM;
273 }
274 while (wr) {
275 if (num_wrs == 0) {
276 err = -ENOMEM;
277 *bad_wr = wr;
278 break;
279 }
280 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
281 wqe = (union t3_wr *) (qhp->wq.queue + idx);
282 t3_wr_flags = 0;
283 if (wr->send_flags & IB_SEND_SOLICITED)
284 t3_wr_flags |= T3_SOLICITED_EVENT_FLAG;
285 if (wr->send_flags & IB_SEND_FENCE)
286 t3_wr_flags |= T3_READ_FENCE_FLAG;
287 if (wr->send_flags & IB_SEND_SIGNALED)
288 t3_wr_flags |= T3_COMPLETION_FLAG;
289 sqp = qhp->wq.sq +
290 Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
291 switch (wr->opcode) {
292 case IB_WR_SEND:
293 case IB_WR_SEND_WITH_IMM:
294 t3_wr_opcode = T3_WR_SEND;
295 err = iwch_build_rdma_send(wqe, wr, &t3_wr_flit_cnt);
296 break;
297 case IB_WR_RDMA_WRITE:
298 case IB_WR_RDMA_WRITE_WITH_IMM:
299 t3_wr_opcode = T3_WR_WRITE;
300 err = iwch_build_rdma_write(wqe, wr, &t3_wr_flit_cnt);
301 break;
302 case IB_WR_RDMA_READ:
303 t3_wr_opcode = T3_WR_READ;
304 t3_wr_flags = 0; /* T3 reads are always signaled */
305 err = iwch_build_rdma_read(wqe, wr, &t3_wr_flit_cnt);
306 if (err)
307 break;
308 sqp->read_len = wqe->read.local_len;
309 if (!qhp->wq.oldest_read)
310 qhp->wq.oldest_read = sqp;
311 break;
312 default:
313 PDBG("%s post of type=%d TBD!\n", __FUNCTION__,
314 wr->opcode);
315 err = -EINVAL;
316 }
317 if (err) {
318 *bad_wr = wr;
319 break;
320 }
321 wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
322 sqp->wr_id = wr->wr_id;
323 sqp->opcode = wr2opcode(t3_wr_opcode);
324 sqp->sq_wptr = qhp->wq.sq_wptr;
325 sqp->complete = 0;
326 sqp->signaled = (wr->send_flags & IB_SEND_SIGNALED);
327
328 build_fw_riwrh((void *) wqe, t3_wr_opcode, t3_wr_flags,
329 Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
330 0, t3_wr_flit_cnt);
331 PDBG("%s cookie 0x%llx wq idx 0x%x swsq idx %ld opcode %d\n",
332 __FUNCTION__, (unsigned long long) wr->wr_id, idx,
333 Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2),
334 sqp->opcode);
335 wr = wr->next;
336 num_wrs--;
337 ++(qhp->wq.wptr);
338 ++(qhp->wq.sq_wptr);
339 }
340 spin_unlock_irqrestore(&qhp->lock, flag);
341 ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
342 return err;
343}
344
345int iwch_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
346 struct ib_recv_wr **bad_wr)
347{
348 int err = 0;
349 struct iwch_qp *qhp;
350 u32 idx;
351 union t3_wr *wqe;
352 u32 num_wrs;
353 unsigned long flag;
354
355 qhp = to_iwch_qp(ibqp);
356 spin_lock_irqsave(&qhp->lock, flag);
357 if (qhp->attr.state > IWCH_QP_STATE_RTS) {
358 spin_unlock_irqrestore(&qhp->lock, flag);
359 return -EINVAL;
360 }
361 num_wrs = Q_FREECNT(qhp->wq.rq_rptr, qhp->wq.rq_wptr,
362 qhp->wq.rq_size_log2) - 1;
363 if (!wr) {
364 spin_unlock_irqrestore(&qhp->lock, flag);
365 return -EINVAL;
366 }
367 while (wr) {
368 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
369 wqe = (union t3_wr *) (qhp->wq.queue + idx);
370 if (num_wrs)
371 err = iwch_build_rdma_recv(qhp->rhp, wqe, wr);
372 else
373 err = -ENOMEM;
374 if (err) {
375 *bad_wr = wr;
376 break;
377 }
378 qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr, qhp->wq.rq_size_log2)] =
379 wr->wr_id;
380 build_fw_riwrh((void *) wqe, T3_WR_RCV, T3_COMPLETION_FLAG,
381 Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
382 0, sizeof(struct t3_receive_wr) >> 3);
383 PDBG("%s cookie 0x%llx idx 0x%x rq_wptr 0x%x rw_rptr 0x%x "
384 "wqe %p \n", __FUNCTION__, (unsigned long long) wr->wr_id,
385 idx, qhp->wq.rq_wptr, qhp->wq.rq_rptr, wqe);
386 ++(qhp->wq.rq_wptr);
387 ++(qhp->wq.wptr);
388 wr = wr->next;
389 num_wrs--;
390 }
391 spin_unlock_irqrestore(&qhp->lock, flag);
392 ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
393 return err;
394}
395
396int iwch_bind_mw(struct ib_qp *qp,
397 struct ib_mw *mw,
398 struct ib_mw_bind *mw_bind)
399{
400 struct iwch_dev *rhp;
401 struct iwch_mw *mhp;
402 struct iwch_qp *qhp;
403 union t3_wr *wqe;
404 u32 pbl_addr;
405 u8 page_size;
406 u32 num_wrs;
407 unsigned long flag;
408 struct ib_sge sgl;
409 int err=0;
410 enum t3_wr_flags t3_wr_flags;
411 u32 idx;
412 struct t3_swsq *sqp;
413
414 qhp = to_iwch_qp(qp);
415 mhp = to_iwch_mw(mw);
416 rhp = qhp->rhp;
417
418 spin_lock_irqsave(&qhp->lock, flag);
419 if (qhp->attr.state > IWCH_QP_STATE_RTS) {
420 spin_unlock_irqrestore(&qhp->lock, flag);
421 return -EINVAL;
422 }
423 num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
424 qhp->wq.sq_size_log2);
425 if ((num_wrs) <= 0) {
426 spin_unlock_irqrestore(&qhp->lock, flag);
427 return -ENOMEM;
428 }
429 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
430 PDBG("%s: idx 0x%0x, mw 0x%p, mw_bind 0x%p\n", __FUNCTION__, idx,
431 mw, mw_bind);
432 wqe = (union t3_wr *) (qhp->wq.queue + idx);
433
434 t3_wr_flags = 0;
435 if (mw_bind->send_flags & IB_SEND_SIGNALED)
436 t3_wr_flags = T3_COMPLETION_FLAG;
437
438 sgl.addr = mw_bind->addr;
439 sgl.lkey = mw_bind->mr->lkey;
440 sgl.length = mw_bind->length;
441 wqe->bind.reserved = 0;
442 wqe->bind.type = T3_VA_BASED_TO;
443
444 /* TBD: check perms */
445 wqe->bind.perms = iwch_convert_access(mw_bind->mw_access_flags);
446 wqe->bind.mr_stag = cpu_to_be32(mw_bind->mr->lkey);
447 wqe->bind.mw_stag = cpu_to_be32(mw->rkey);
448 wqe->bind.mw_len = cpu_to_be32(mw_bind->length);
449 wqe->bind.mw_va = cpu_to_be64(mw_bind->addr);
450 err = iwch_sgl2pbl_map(rhp, &sgl, 1, &pbl_addr, &page_size);
451 if (err) {
452 spin_unlock_irqrestore(&qhp->lock, flag);
453 return err;
454 }
455 wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
456 sqp = qhp->wq.sq + Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
457 sqp->wr_id = mw_bind->wr_id;
458 sqp->opcode = T3_BIND_MW;
459 sqp->sq_wptr = qhp->wq.sq_wptr;
460 sqp->complete = 0;
461 sqp->signaled = (mw_bind->send_flags & IB_SEND_SIGNALED);
462 wqe->bind.mr_pbl_addr = cpu_to_be32(pbl_addr);
463 wqe->bind.mr_pagesz = page_size;
464 wqe->flit[T3_SQ_COOKIE_FLIT] = mw_bind->wr_id;
465 build_fw_riwrh((void *)wqe, T3_WR_BIND, t3_wr_flags,
466 Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2), 0,
467 sizeof(struct t3_bind_mw_wr) >> 3);
468 ++(qhp->wq.wptr);
469 ++(qhp->wq.sq_wptr);
470 spin_unlock_irqrestore(&qhp->lock, flag);
471
472 ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
473
474 return err;
475}
476
477static inline void build_term_codes(int t3err, u8 *layer_type, u8 *ecode,
478 int tagged)
479{
480 switch (t3err) {
481 case TPT_ERR_STAG:
482 if (tagged == 1) {
483 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
484 *ecode = DDPT_INV_STAG;
485 } else if (tagged == 2) {
486 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
487 *ecode = RDMAP_INV_STAG;
488 }
489 break;
490 case TPT_ERR_PDID:
491 case TPT_ERR_QPID:
492 case TPT_ERR_ACCESS:
493 if (tagged == 1) {
494 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
495 *ecode = DDPT_STAG_NOT_ASSOC;
496 } else if (tagged == 2) {
497 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
498 *ecode = RDMAP_STAG_NOT_ASSOC;
499 }
500 break;
501 case TPT_ERR_WRAP:
502 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
503 *ecode = RDMAP_TO_WRAP;
504 break;
505 case TPT_ERR_BOUND:
506 if (tagged == 1) {
507 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
508 *ecode = DDPT_BASE_BOUNDS;
509 } else if (tagged == 2) {
510 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
511 *ecode = RDMAP_BASE_BOUNDS;
512 } else {
513 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
514 *ecode = DDPU_MSG_TOOBIG;
515 }
516 break;
517 case TPT_ERR_INVALIDATE_SHARED_MR:
518 case TPT_ERR_INVALIDATE_MR_WITH_MW_BOUND:
519 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
520 *ecode = RDMAP_CANT_INV_STAG;
521 break;
522 case TPT_ERR_ECC:
523 case TPT_ERR_ECC_PSTAG:
524 case TPT_ERR_INTERNAL_ERR:
525 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
526 *ecode = 0;
527 break;
528 case TPT_ERR_OUT_OF_RQE:
529 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
530 *ecode = DDPU_INV_MSN_NOBUF;
531 break;
532 case TPT_ERR_PBL_ADDR_BOUND:
533 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
534 *ecode = DDPT_BASE_BOUNDS;
535 break;
536 case TPT_ERR_CRC:
537 *layer_type = LAYER_MPA|DDP_LLP;
538 *ecode = MPA_CRC_ERR;
539 break;
540 case TPT_ERR_MARKER:
541 *layer_type = LAYER_MPA|DDP_LLP;
542 *ecode = MPA_MARKER_ERR;
543 break;
544 case TPT_ERR_PDU_LEN_ERR:
545 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
546 *ecode = DDPU_MSG_TOOBIG;
547 break;
548 case TPT_ERR_DDP_VERSION:
549 if (tagged) {
550 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
551 *ecode = DDPT_INV_VERS;
552 } else {
553 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
554 *ecode = DDPU_INV_VERS;
555 }
556 break;
557 case TPT_ERR_RDMA_VERSION:
558 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
559 *ecode = RDMAP_INV_VERS;
560 break;
561 case TPT_ERR_OPCODE:
562 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
563 *ecode = RDMAP_INV_OPCODE;
564 break;
565 case TPT_ERR_DDP_QUEUE_NUM:
566 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
567 *ecode = DDPU_INV_QN;
568 break;
569 case TPT_ERR_MSN:
570 case TPT_ERR_MSN_GAP:
571 case TPT_ERR_MSN_RANGE:
572 case TPT_ERR_IRD_OVERFLOW:
573 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
574 *ecode = DDPU_INV_MSN_RANGE;
575 break;
576 case TPT_ERR_TBIT:
577 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
578 *ecode = 0;
579 break;
580 case TPT_ERR_MO:
581 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
582 *ecode = DDPU_INV_MO;
583 break;
584 default:
585 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
586 *ecode = 0;
587 break;
588 }
589}
590
591/*
592 * This posts a TERMINATE with layer=RDMA, type=catastrophic.
593 */
594int iwch_post_terminate(struct iwch_qp *qhp, struct respQ_msg_t *rsp_msg)
595{
596 union t3_wr *wqe;
597 struct terminate_message *term;
598 int status;
599 int tagged = 0;
600 struct sk_buff *skb;
601
602 PDBG("%s %d\n", __FUNCTION__, __LINE__);
603 skb = alloc_skb(40, GFP_ATOMIC);
604 if (!skb) {
605 printk(KERN_ERR "%s cannot send TERMINATE!\n", __FUNCTION__);
606 return -ENOMEM;
607 }
608 wqe = (union t3_wr *)skb_put(skb, 40);
609 memset(wqe, 0, 40);
610 wqe->send.rdmaop = T3_TERMINATE;
611
612 /* immediate data length */
613 wqe->send.plen = htonl(4);
614
615 /* immediate data starts here. */
616 term = (struct terminate_message *)wqe->send.sgl;
617 if (rsp_msg) {
618 status = CQE_STATUS(rsp_msg->cqe);
619 if (CQE_OPCODE(rsp_msg->cqe) == T3_RDMA_WRITE)
620 tagged = 1;
621 if ((CQE_OPCODE(rsp_msg->cqe) == T3_READ_REQ) ||
622 (CQE_OPCODE(rsp_msg->cqe) == T3_READ_RESP))
623 tagged = 2;
624 } else {
625 status = TPT_ERR_INTERNAL_ERR;
626 }
627 build_term_codes(status, &term->layer_etype, &term->ecode, tagged);
628 build_fw_riwrh((void *)wqe, T3_WR_SEND,
629 T3_COMPLETION_FLAG | T3_NOTIFY_FLAG, 1,
630 qhp->ep->hwtid, 5);
631 skb->priority = CPL_PRIORITY_DATA;
632 return cxgb3_ofld_send(qhp->rhp->rdev.t3cdev_p, skb);
633}
634
635/*
636 * Assumes qhp lock is held.
637 */
638static void __flush_qp(struct iwch_qp *qhp, unsigned long *flag)
639{
640 struct iwch_cq *rchp, *schp;
641 int count;
642
643 rchp = get_chp(qhp->rhp, qhp->attr.rcq);
644 schp = get_chp(qhp->rhp, qhp->attr.scq);
645
646 PDBG("%s qhp %p rchp %p schp %p\n", __FUNCTION__, qhp, rchp, schp);
647 /* take a ref on the qhp since we must release the lock */
648 atomic_inc(&qhp->refcnt);
649 spin_unlock_irqrestore(&qhp->lock, *flag);
650
651 /* locking heirarchy: cq lock first, then qp lock. */
652 spin_lock_irqsave(&rchp->lock, *flag);
653 spin_lock(&qhp->lock);
654 cxio_flush_hw_cq(&rchp->cq);
655 cxio_count_rcqes(&rchp->cq, &qhp->wq, &count);
656 cxio_flush_rq(&qhp->wq, &rchp->cq, count);
657 spin_unlock(&qhp->lock);
658 spin_unlock_irqrestore(&rchp->lock, *flag);
659
660 /* locking heirarchy: cq lock first, then qp lock. */
661 spin_lock_irqsave(&schp->lock, *flag);
662 spin_lock(&qhp->lock);
663 cxio_flush_hw_cq(&schp->cq);
664 cxio_count_scqes(&schp->cq, &qhp->wq, &count);
665 cxio_flush_sq(&qhp->wq, &schp->cq, count);
666 spin_unlock(&qhp->lock);
667 spin_unlock_irqrestore(&schp->lock, *flag);
668
669 /* deref */
670 if (atomic_dec_and_test(&qhp->refcnt))
671 wake_up(&qhp->wait);
672
673 spin_lock_irqsave(&qhp->lock, *flag);
674}
675
676static inline void flush_qp(struct iwch_qp *qhp, unsigned long *flag)
677{
678 if (t3b_device(qhp->rhp))
679 cxio_set_wq_in_error(&qhp->wq);
680 else
681 __flush_qp(qhp, flag);
682}
683
684
685/*
686 * Return non zero if at least one RECV was pre-posted.
687 */
688static inline int rqes_posted(struct iwch_qp *qhp)
689{
690 return fw_riwrh_opcode((struct fw_riwrh *)qhp->wq.queue) == T3_WR_RCV;
691}
692
693static int rdma_init(struct iwch_dev *rhp, struct iwch_qp *qhp,
694 enum iwch_qp_attr_mask mask,
695 struct iwch_qp_attributes *attrs)
696{
697 struct t3_rdma_init_attr init_attr;
698 int ret;
699
700 init_attr.tid = qhp->ep->hwtid;
701 init_attr.qpid = qhp->wq.qpid;
702 init_attr.pdid = qhp->attr.pd;
703 init_attr.scqid = qhp->attr.scq;
704 init_attr.rcqid = qhp->attr.rcq;
705 init_attr.rq_addr = qhp->wq.rq_addr;
706 init_attr.rq_size = 1 << qhp->wq.rq_size_log2;
707 init_attr.mpaattrs = uP_RI_MPA_IETF_ENABLE |
708 qhp->attr.mpa_attr.recv_marker_enabled |
709 (qhp->attr.mpa_attr.xmit_marker_enabled << 1) |
710 (qhp->attr.mpa_attr.crc_enabled << 2);
711
712 /*
713 * XXX - The IWCM doesn't quite handle getting these
714 * attrs set before going into RTS. For now, just turn
715 * them on always...
716 */
717#if 0
718 init_attr.qpcaps = qhp->attr.enableRdmaRead |
719 (qhp->attr.enableRdmaWrite << 1) |
720 (qhp->attr.enableBind << 2) |
721 (qhp->attr.enable_stag0_fastreg << 3) |
722 (qhp->attr.enable_stag0_fastreg << 4);
723#else
724 init_attr.qpcaps = 0x1f;
725#endif
726 init_attr.tcp_emss = qhp->ep->emss;
727 init_attr.ord = qhp->attr.max_ord;
728 init_attr.ird = qhp->attr.max_ird;
729 init_attr.qp_dma_addr = qhp->wq.dma_addr;
730 init_attr.qp_dma_size = (1UL << qhp->wq.size_log2);
731 init_attr.flags = rqes_posted(qhp) ? RECVS_POSTED : 0;
732 PDBG("%s init_attr.rq_addr 0x%x init_attr.rq_size = %d "
733 "flags 0x%x qpcaps 0x%x\n", __FUNCTION__,
734 init_attr.rq_addr, init_attr.rq_size,
735 init_attr.flags, init_attr.qpcaps);
736 ret = cxio_rdma_init(&rhp->rdev, &init_attr);
737 PDBG("%s ret %d\n", __FUNCTION__, ret);
738 return ret;
739}
740
741int iwch_modify_qp(struct iwch_dev *rhp, struct iwch_qp *qhp,
742 enum iwch_qp_attr_mask mask,
743 struct iwch_qp_attributes *attrs,
744 int internal)
745{
746 int ret = 0;
747 struct iwch_qp_attributes newattr = qhp->attr;
748 unsigned long flag;
749 int disconnect = 0;
750 int terminate = 0;
751 int abort = 0;
752 int free = 0;
753 struct iwch_ep *ep = NULL;
754
755 PDBG("%s qhp %p qpid 0x%x ep %p state %d -> %d\n", __FUNCTION__,
756 qhp, qhp->wq.qpid, qhp->ep, qhp->attr.state,
757 (mask & IWCH_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
758
759 spin_lock_irqsave(&qhp->lock, flag);
760
761 /* Process attr changes if in IDLE */
762 if (mask & IWCH_QP_ATTR_VALID_MODIFY) {
763 if (qhp->attr.state != IWCH_QP_STATE_IDLE) {
764 ret = -EIO;
765 goto out;
766 }
767 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_READ)
768 newattr.enable_rdma_read = attrs->enable_rdma_read;
769 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_WRITE)
770 newattr.enable_rdma_write = attrs->enable_rdma_write;
771 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_BIND)
772 newattr.enable_bind = attrs->enable_bind;
773 if (mask & IWCH_QP_ATTR_MAX_ORD) {
774 if (attrs->max_ord >
775 rhp->attr.max_rdma_read_qp_depth) {
776 ret = -EINVAL;
777 goto out;
778 }
779 newattr.max_ord = attrs->max_ord;
780 }
781 if (mask & IWCH_QP_ATTR_MAX_IRD) {
782 if (attrs->max_ird >
783 rhp->attr.max_rdma_reads_per_qp) {
784 ret = -EINVAL;
785 goto out;
786 }
787 newattr.max_ird = attrs->max_ird;
788 }
789 qhp->attr = newattr;
790 }
791
792 if (!(mask & IWCH_QP_ATTR_NEXT_STATE))
793 goto out;
794 if (qhp->attr.state == attrs->next_state)
795 goto out;
796
797 switch (qhp->attr.state) {
798 case IWCH_QP_STATE_IDLE:
799 switch (attrs->next_state) {
800 case IWCH_QP_STATE_RTS:
801 if (!(mask & IWCH_QP_ATTR_LLP_STREAM_HANDLE)) {
802 ret = -EINVAL;
803 goto out;
804 }
805 if (!(mask & IWCH_QP_ATTR_MPA_ATTR)) {
806 ret = -EINVAL;
807 goto out;
808 }
809 qhp->attr.mpa_attr = attrs->mpa_attr;
810 qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
811 qhp->ep = qhp->attr.llp_stream_handle;
812 qhp->attr.state = IWCH_QP_STATE_RTS;
813
814 /*
815 * Ref the endpoint here and deref when we
816 * disassociate the endpoint from the QP. This
817 * happens in CLOSING->IDLE transition or *->ERROR
818 * transition.
819 */
820 get_ep(&qhp->ep->com);
821 spin_unlock_irqrestore(&qhp->lock, flag);
822 ret = rdma_init(rhp, qhp, mask, attrs);
823 spin_lock_irqsave(&qhp->lock, flag);
824 if (ret)
825 goto err;
826 break;
827 case IWCH_QP_STATE_ERROR:
828 qhp->attr.state = IWCH_QP_STATE_ERROR;
829 flush_qp(qhp, &flag);
830 break;
831 default:
832 ret = -EINVAL;
833 goto out;
834 }
835 break;
836 case IWCH_QP_STATE_RTS:
837 switch (attrs->next_state) {
838 case IWCH_QP_STATE_CLOSING:
839 BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
840 qhp->attr.state = IWCH_QP_STATE_CLOSING;
841 if (!internal) {
842 abort=0;
843 disconnect = 1;
844 ep = qhp->ep;
845 }
846 break;
847 case IWCH_QP_STATE_TERMINATE:
848 qhp->attr.state = IWCH_QP_STATE_TERMINATE;
849 if (!internal)
850 terminate = 1;
851 break;
852 case IWCH_QP_STATE_ERROR:
853 qhp->attr.state = IWCH_QP_STATE_ERROR;
854 if (!internal) {
855 abort=1;
856 disconnect = 1;
857 ep = qhp->ep;
858 }
859 goto err;
860 break;
861 default:
862 ret = -EINVAL;
863 goto out;
864 }
865 break;
866 case IWCH_QP_STATE_CLOSING:
867 if (!internal) {
868 ret = -EINVAL;
869 goto out;
870 }
871 switch (attrs->next_state) {
872 case IWCH_QP_STATE_IDLE:
873 qhp->attr.state = IWCH_QP_STATE_IDLE;
874 qhp->attr.llp_stream_handle = NULL;
875 put_ep(&qhp->ep->com);
876 qhp->ep = NULL;
877 wake_up(&qhp->wait);
878 break;
879 case IWCH_QP_STATE_ERROR:
880 goto err;
881 default:
882 ret = -EINVAL;
883 goto err;
884 }
885 break;
886 case IWCH_QP_STATE_ERROR:
887 if (attrs->next_state != IWCH_QP_STATE_IDLE) {
888 ret = -EINVAL;
889 goto out;
890 }
891
892 if (!Q_EMPTY(qhp->wq.sq_rptr, qhp->wq.sq_wptr) ||
893 !Q_EMPTY(qhp->wq.rq_rptr, qhp->wq.rq_wptr)) {
894 ret = -EINVAL;
895 goto out;
896 }
897 qhp->attr.state = IWCH_QP_STATE_IDLE;
898 memset(&qhp->attr, 0, sizeof(qhp->attr));
899 break;
900 case IWCH_QP_STATE_TERMINATE:
901 if (!internal) {
902 ret = -EINVAL;
903 goto out;
904 }
905 goto err;
906 break;
907 default:
908 printk(KERN_ERR "%s in a bad state %d\n",
909 __FUNCTION__, qhp->attr.state);
910 ret = -EINVAL;
911 goto err;
912 break;
913 }
914 goto out;
915err:
916 PDBG("%s disassociating ep %p qpid 0x%x\n", __FUNCTION__, qhp->ep,
917 qhp->wq.qpid);
918
919 /* disassociate the LLP connection */
920 qhp->attr.llp_stream_handle = NULL;
921 ep = qhp->ep;
922 qhp->ep = NULL;
923 qhp->attr.state = IWCH_QP_STATE_ERROR;
924 free=1;
925 wake_up(&qhp->wait);
926 BUG_ON(!ep);
927 flush_qp(qhp, &flag);
928out:
929 spin_unlock_irqrestore(&qhp->lock, flag);
930
931 if (terminate)
932 iwch_post_terminate(qhp, NULL);
933
934 /*
935 * If disconnect is 1, then we need to initiate a disconnect
936 * on the EP. This can be a normal close (RTS->CLOSING) or
937 * an abnormal close (RTS/CLOSING->ERROR).
938 */
939 if (disconnect)
940 iwch_ep_disconnect(ep, abort, GFP_KERNEL);
941
942 /*
943 * If free is 1, then we've disassociated the EP from the QP
944 * and we need to dereference the EP.
945 */
946 if (free)
947 put_ep(&ep->com);
948
949 PDBG("%s exit state %d\n", __FUNCTION__, qhp->attr.state);
950 return ret;
951}
952
953static int quiesce_qp(struct iwch_qp *qhp)
954{
955 spin_lock_irq(&qhp->lock);
956 iwch_quiesce_tid(qhp->ep);
957 qhp->flags |= QP_QUIESCED;
958 spin_unlock_irq(&qhp->lock);
959 return 0;
960}
961
962static int resume_qp(struct iwch_qp *qhp)
963{
964 spin_lock_irq(&qhp->lock);
965 iwch_resume_tid(qhp->ep);
966 qhp->flags &= ~QP_QUIESCED;
967 spin_unlock_irq(&qhp->lock);
968 return 0;
969}
970
971int iwch_quiesce_qps(struct iwch_cq *chp)
972{
973 int i;
974 struct iwch_qp *qhp;
975
976 for (i=0; i < T3_MAX_NUM_QP; i++) {
977 qhp = get_qhp(chp->rhp, i);
978 if (!qhp)
979 continue;
980 if ((qhp->attr.rcq == chp->cq.cqid) && !qp_quiesced(qhp)) {
981 quiesce_qp(qhp);
982 continue;
983 }
984 if ((qhp->attr.scq == chp->cq.cqid) && !qp_quiesced(qhp))
985 quiesce_qp(qhp);
986 }
987 return 0;
988}
989
990int iwch_resume_qps(struct iwch_cq *chp)
991{
992 int i;
993 struct iwch_qp *qhp;
994
995 for (i=0; i < T3_MAX_NUM_QP; i++) {
996 qhp = get_qhp(chp->rhp, i);
997 if (!qhp)
998 continue;
999 if ((qhp->attr.rcq == chp->cq.cqid) && qp_quiesced(qhp)) {
1000 resume_qp(qhp);
1001 continue;
1002 }
1003 if ((qhp->attr.scq == chp->cq.cqid) && qp_quiesced(qhp))
1004 resume_qp(qhp);
1005 }
1006 return 0;
1007}