1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5 */
6
7 #include <linux/skbuff.h>
8
9 #include "rxe.h"
10 #include "rxe_loc.h"
11 #include "rxe_queue.h"
12 #include "rxe_task.h"
13
14 enum comp_state {
15 COMPST_GET_ACK,
16 COMPST_GET_WQE,
17 COMPST_COMP_WQE,
18 COMPST_COMP_ACK,
19 COMPST_CHECK_PSN,
20 COMPST_CHECK_ACK,
21 COMPST_READ,
22 COMPST_ATOMIC,
23 COMPST_WRITE_SEND,
24 COMPST_UPDATE_COMP,
25 COMPST_ERROR_RETRY,
26 COMPST_RNR_RETRY,
27 COMPST_ERROR,
28 COMPST_EXIT, /* We have an issue, and we want to rerun the completer */
29 COMPST_DONE, /* The completer finished successflly */
30 };
31
32 static char *comp_state_name[] = {
33 [COMPST_GET_ACK] = "GET ACK",
34 [COMPST_GET_WQE] = "GET WQE",
35 [COMPST_COMP_WQE] = "COMP WQE",
36 [COMPST_COMP_ACK] = "COMP ACK",
37 [COMPST_CHECK_PSN] = "CHECK PSN",
38 [COMPST_CHECK_ACK] = "CHECK ACK",
39 [COMPST_READ] = "READ",
40 [COMPST_ATOMIC] = "ATOMIC",
41 [COMPST_WRITE_SEND] = "WRITE/SEND",
42 [COMPST_UPDATE_COMP] = "UPDATE COMP",
43 [COMPST_ERROR_RETRY] = "ERROR RETRY",
44 [COMPST_RNR_RETRY] = "RNR RETRY",
45 [COMPST_ERROR] = "ERROR",
46 [COMPST_EXIT] = "EXIT",
47 [COMPST_DONE] = "DONE",
48 };
49
50 static unsigned long rnrnak_usec[32] = {
51 [IB_RNR_TIMER_655_36] = 655360,
52 [IB_RNR_TIMER_000_01] = 10,
53 [IB_RNR_TIMER_000_02] = 20,
54 [IB_RNR_TIMER_000_03] = 30,
55 [IB_RNR_TIMER_000_04] = 40,
56 [IB_RNR_TIMER_000_06] = 60,
57 [IB_RNR_TIMER_000_08] = 80,
58 [IB_RNR_TIMER_000_12] = 120,
59 [IB_RNR_TIMER_000_16] = 160,
60 [IB_RNR_TIMER_000_24] = 240,
61 [IB_RNR_TIMER_000_32] = 320,
62 [IB_RNR_TIMER_000_48] = 480,
63 [IB_RNR_TIMER_000_64] = 640,
64 [IB_RNR_TIMER_000_96] = 960,
65 [IB_RNR_TIMER_001_28] = 1280,
66 [IB_RNR_TIMER_001_92] = 1920,
67 [IB_RNR_TIMER_002_56] = 2560,
68 [IB_RNR_TIMER_003_84] = 3840,
69 [IB_RNR_TIMER_005_12] = 5120,
70 [IB_RNR_TIMER_007_68] = 7680,
71 [IB_RNR_TIMER_010_24] = 10240,
72 [IB_RNR_TIMER_015_36] = 15360,
73 [IB_RNR_TIMER_020_48] = 20480,
74 [IB_RNR_TIMER_030_72] = 30720,
75 [IB_RNR_TIMER_040_96] = 40960,
76 [IB_RNR_TIMER_061_44] = 61410,
77 [IB_RNR_TIMER_081_92] = 81920,
78 [IB_RNR_TIMER_122_88] = 122880,
79 [IB_RNR_TIMER_163_84] = 163840,
80 [IB_RNR_TIMER_245_76] = 245760,
81 [IB_RNR_TIMER_327_68] = 327680,
82 [IB_RNR_TIMER_491_52] = 491520,
83 };
84
rnrnak_jiffies(u8 timeout)85 static inline unsigned long rnrnak_jiffies(u8 timeout)
86 {
87 return max_t(unsigned long,
88 usecs_to_jiffies(rnrnak_usec[timeout]), 1);
89 }
90
wr_to_wc_opcode(enum ib_wr_opcode opcode)91 static enum ib_wc_opcode wr_to_wc_opcode(enum ib_wr_opcode opcode)
92 {
93 switch (opcode) {
94 case IB_WR_RDMA_WRITE: return IB_WC_RDMA_WRITE;
95 case IB_WR_RDMA_WRITE_WITH_IMM: return IB_WC_RDMA_WRITE;
96 case IB_WR_SEND: return IB_WC_SEND;
97 case IB_WR_SEND_WITH_IMM: return IB_WC_SEND;
98 case IB_WR_RDMA_READ: return IB_WC_RDMA_READ;
99 case IB_WR_ATOMIC_CMP_AND_SWP: return IB_WC_COMP_SWAP;
100 case IB_WR_ATOMIC_FETCH_AND_ADD: return IB_WC_FETCH_ADD;
101 case IB_WR_LSO: return IB_WC_LSO;
102 case IB_WR_SEND_WITH_INV: return IB_WC_SEND;
103 case IB_WR_RDMA_READ_WITH_INV: return IB_WC_RDMA_READ;
104 case IB_WR_LOCAL_INV: return IB_WC_LOCAL_INV;
105 case IB_WR_REG_MR: return IB_WC_REG_MR;
106 case IB_WR_BIND_MW: return IB_WC_BIND_MW;
107 case IB_WR_ATOMIC_WRITE: return IB_WC_ATOMIC_WRITE;
108 case IB_WR_FLUSH: return IB_WC_FLUSH;
109
110 default:
111 return 0xff;
112 }
113 }
114
retransmit_timer(struct timer_list * t)115 void retransmit_timer(struct timer_list *t)
116 {
117 struct rxe_qp *qp = from_timer(qp, t, retrans_timer);
118
119 rxe_dbg_qp(qp, "retransmit timer fired\n");
120
121 if (qp->valid) {
122 qp->comp.timeout = 1;
123 rxe_sched_task(&qp->comp.task);
124 }
125 }
126
rxe_comp_queue_pkt(struct rxe_qp * qp,struct sk_buff * skb)127 void rxe_comp_queue_pkt(struct rxe_qp *qp, struct sk_buff *skb)
128 {
129 int must_sched;
130
131 skb_queue_tail(&qp->resp_pkts, skb);
132
133 must_sched = skb_queue_len(&qp->resp_pkts) > 1;
134 if (must_sched != 0)
135 rxe_counter_inc(SKB_TO_PKT(skb)->rxe, RXE_CNT_COMPLETER_SCHED);
136
137 if (must_sched)
138 rxe_sched_task(&qp->comp.task);
139 else
140 rxe_run_task(&qp->comp.task);
141 }
142
get_wqe(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe ** wqe_p)143 static inline enum comp_state get_wqe(struct rxe_qp *qp,
144 struct rxe_pkt_info *pkt,
145 struct rxe_send_wqe **wqe_p)
146 {
147 struct rxe_send_wqe *wqe;
148
149 /* we come here whether or not we found a response packet to see if
150 * there are any posted WQEs
151 */
152 wqe = queue_head(qp->sq.queue, QUEUE_TYPE_FROM_CLIENT);
153 *wqe_p = wqe;
154
155 /* no WQE or requester has not started it yet */
156 if (!wqe || wqe->state == wqe_state_posted)
157 return pkt ? COMPST_DONE : COMPST_EXIT;
158
159 /* WQE does not require an ack */
160 if (wqe->state == wqe_state_done)
161 return COMPST_COMP_WQE;
162
163 /* WQE caused an error */
164 if (wqe->state == wqe_state_error)
165 return COMPST_ERROR;
166
167 /* we have a WQE, if we also have an ack check its PSN */
168 return pkt ? COMPST_CHECK_PSN : COMPST_EXIT;
169 }
170
reset_retry_counters(struct rxe_qp * qp)171 static inline void reset_retry_counters(struct rxe_qp *qp)
172 {
173 qp->comp.retry_cnt = qp->attr.retry_cnt;
174 qp->comp.rnr_retry = qp->attr.rnr_retry;
175 qp->comp.started_retry = 0;
176 }
177
check_psn(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)178 static inline enum comp_state check_psn(struct rxe_qp *qp,
179 struct rxe_pkt_info *pkt,
180 struct rxe_send_wqe *wqe)
181 {
182 s32 diff;
183
184 /* check to see if response is past the oldest WQE. if it is, complete
185 * send/write or error read/atomic
186 */
187 diff = psn_compare(pkt->psn, wqe->last_psn);
188 if (diff > 0) {
189 if (wqe->state == wqe_state_pending) {
190 if (wqe->mask & WR_ATOMIC_OR_READ_MASK)
191 return COMPST_ERROR_RETRY;
192
193 reset_retry_counters(qp);
194 return COMPST_COMP_WQE;
195 } else {
196 return COMPST_DONE;
197 }
198 }
199
200 /* compare response packet to expected response */
201 diff = psn_compare(pkt->psn, qp->comp.psn);
202 if (diff < 0) {
203 /* response is most likely a retried packet if it matches an
204 * uncompleted WQE go complete it else ignore it
205 */
206 if (pkt->psn == wqe->last_psn)
207 return COMPST_COMP_ACK;
208 else if (pkt->opcode == IB_OPCODE_RC_ACKNOWLEDGE &&
209 (qp->comp.opcode == IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST ||
210 qp->comp.opcode == IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE))
211 return COMPST_CHECK_ACK;
212 else
213 return COMPST_DONE;
214 } else if ((diff > 0) && (wqe->mask & WR_ATOMIC_OR_READ_MASK)) {
215 return COMPST_DONE;
216 } else {
217 return COMPST_CHECK_ACK;
218 }
219 }
220
check_ack(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)221 static inline enum comp_state check_ack(struct rxe_qp *qp,
222 struct rxe_pkt_info *pkt,
223 struct rxe_send_wqe *wqe)
224 {
225 unsigned int mask = pkt->mask;
226 u8 syn;
227 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
228
229 /* Check the sequence only */
230 switch (qp->comp.opcode) {
231 case -1:
232 /* Will catch all *_ONLY cases. */
233 if (!(mask & RXE_START_MASK))
234 return COMPST_ERROR;
235
236 break;
237
238 case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
239 case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
240 /* Check NAK code to handle a remote error */
241 if (pkt->opcode == IB_OPCODE_RC_ACKNOWLEDGE)
242 break;
243
244 if (pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE &&
245 pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST) {
246 /* read retries of partial data may restart from
247 * read response first or response only.
248 */
249 if ((pkt->psn == wqe->first_psn &&
250 pkt->opcode ==
251 IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) ||
252 (wqe->first_psn == wqe->last_psn &&
253 pkt->opcode ==
254 IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY))
255 break;
256
257 return COMPST_ERROR;
258 }
259 break;
260 default:
261 WARN_ON_ONCE(1);
262 }
263
264 /* Check operation validity. */
265 switch (pkt->opcode) {
266 case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
267 case IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST:
268 case IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY:
269 syn = aeth_syn(pkt);
270
271 if ((syn & AETH_TYPE_MASK) != AETH_ACK)
272 return COMPST_ERROR;
273
274 if (wqe->wr.opcode == IB_WR_ATOMIC_WRITE)
275 return COMPST_WRITE_SEND;
276
277 fallthrough;
278 /* (IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE doesn't have an AETH)
279 */
280 case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
281 if (wqe->wr.opcode != IB_WR_RDMA_READ &&
282 wqe->wr.opcode != IB_WR_RDMA_READ_WITH_INV &&
283 wqe->wr.opcode != IB_WR_FLUSH) {
284 wqe->status = IB_WC_FATAL_ERR;
285 return COMPST_ERROR;
286 }
287 reset_retry_counters(qp);
288 return COMPST_READ;
289
290 case IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE:
291 syn = aeth_syn(pkt);
292
293 if ((syn & AETH_TYPE_MASK) != AETH_ACK)
294 return COMPST_ERROR;
295
296 if (wqe->wr.opcode != IB_WR_ATOMIC_CMP_AND_SWP &&
297 wqe->wr.opcode != IB_WR_ATOMIC_FETCH_AND_ADD)
298 return COMPST_ERROR;
299 reset_retry_counters(qp);
300 return COMPST_ATOMIC;
301
302 case IB_OPCODE_RC_ACKNOWLEDGE:
303 syn = aeth_syn(pkt);
304 switch (syn & AETH_TYPE_MASK) {
305 case AETH_ACK:
306 reset_retry_counters(qp);
307 return COMPST_WRITE_SEND;
308
309 case AETH_RNR_NAK:
310 rxe_counter_inc(rxe, RXE_CNT_RCV_RNR);
311 return COMPST_RNR_RETRY;
312
313 case AETH_NAK:
314 switch (syn) {
315 case AETH_NAK_PSN_SEQ_ERROR:
316 /* a nak implicitly acks all packets with psns
317 * before
318 */
319 if (psn_compare(pkt->psn, qp->comp.psn) > 0) {
320 rxe_counter_inc(rxe,
321 RXE_CNT_RCV_SEQ_ERR);
322 qp->comp.psn = pkt->psn;
323 if (qp->req.wait_psn) {
324 qp->req.wait_psn = 0;
325 rxe_run_task(&qp->req.task);
326 }
327 }
328 return COMPST_ERROR_RETRY;
329
330 case AETH_NAK_INVALID_REQ:
331 wqe->status = IB_WC_REM_INV_REQ_ERR;
332 return COMPST_ERROR;
333
334 case AETH_NAK_REM_ACC_ERR:
335 wqe->status = IB_WC_REM_ACCESS_ERR;
336 return COMPST_ERROR;
337
338 case AETH_NAK_REM_OP_ERR:
339 wqe->status = IB_WC_REM_OP_ERR;
340 return COMPST_ERROR;
341
342 default:
343 rxe_dbg_qp(qp, "unexpected nak %x\n", syn);
344 wqe->status = IB_WC_REM_OP_ERR;
345 return COMPST_ERROR;
346 }
347
348 default:
349 return COMPST_ERROR;
350 }
351 break;
352
353 default:
354 rxe_dbg_qp(qp, "unexpected opcode\n");
355 }
356
357 return COMPST_ERROR;
358 }
359
do_read(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)360 static inline enum comp_state do_read(struct rxe_qp *qp,
361 struct rxe_pkt_info *pkt,
362 struct rxe_send_wqe *wqe)
363 {
364 int ret;
365
366 ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE,
367 &wqe->dma, payload_addr(pkt),
368 payload_size(pkt), RXE_TO_MR_OBJ);
369 if (ret) {
370 wqe->status = IB_WC_LOC_PROT_ERR;
371 return COMPST_ERROR;
372 }
373
374 if (wqe->dma.resid == 0 && (pkt->mask & RXE_END_MASK))
375 return COMPST_COMP_ACK;
376
377 return COMPST_UPDATE_COMP;
378 }
379
do_atomic(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)380 static inline enum comp_state do_atomic(struct rxe_qp *qp,
381 struct rxe_pkt_info *pkt,
382 struct rxe_send_wqe *wqe)
383 {
384 int ret;
385
386 u64 atomic_orig = atmack_orig(pkt);
387
388 ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE,
389 &wqe->dma, &atomic_orig,
390 sizeof(u64), RXE_TO_MR_OBJ);
391 if (ret) {
392 wqe->status = IB_WC_LOC_PROT_ERR;
393 return COMPST_ERROR;
394 }
395
396 return COMPST_COMP_ACK;
397 }
398
make_send_cqe(struct rxe_qp * qp,struct rxe_send_wqe * wqe,struct rxe_cqe * cqe)399 static void make_send_cqe(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
400 struct rxe_cqe *cqe)
401 {
402 struct ib_wc *wc = &cqe->ibwc;
403 struct ib_uverbs_wc *uwc = &cqe->uibwc;
404
405 memset(cqe, 0, sizeof(*cqe));
406
407 if (!qp->is_user) {
408 wc->wr_id = wqe->wr.wr_id;
409 wc->status = wqe->status;
410 wc->qp = &qp->ibqp;
411 } else {
412 uwc->wr_id = wqe->wr.wr_id;
413 uwc->status = wqe->status;
414 uwc->qp_num = qp->ibqp.qp_num;
415 }
416
417 if (wqe->status == IB_WC_SUCCESS) {
418 if (!qp->is_user) {
419 wc->opcode = wr_to_wc_opcode(wqe->wr.opcode);
420 if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
421 wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
422 wc->wc_flags = IB_WC_WITH_IMM;
423 wc->byte_len = wqe->dma.length;
424 } else {
425 uwc->opcode = wr_to_wc_opcode(wqe->wr.opcode);
426 if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
427 wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
428 uwc->wc_flags = IB_WC_WITH_IMM;
429 uwc->byte_len = wqe->dma.length;
430 }
431 }
432 }
433
434 /*
435 * IBA Spec. Section 10.7.3.1 SIGNALED COMPLETIONS
436 * ---------8<---------8<-------------
437 * ...Note that if a completion error occurs, a Work Completion
438 * will always be generated, even if the signaling
439 * indicator requests an Unsignaled Completion.
440 * ---------8<---------8<-------------
441 */
do_complete(struct rxe_qp * qp,struct rxe_send_wqe * wqe)442 static void do_complete(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
443 {
444 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
445 struct rxe_cqe cqe;
446 bool post;
447
448 /* do we need to post a completion */
449 post = ((qp->sq_sig_type == IB_SIGNAL_ALL_WR) ||
450 (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
451 wqe->status != IB_WC_SUCCESS);
452
453 if (post)
454 make_send_cqe(qp, wqe, &cqe);
455
456 queue_advance_consumer(qp->sq.queue, QUEUE_TYPE_FROM_CLIENT);
457
458 if (post)
459 rxe_cq_post(qp->scq, &cqe, 0);
460
461 if (wqe->wr.opcode == IB_WR_SEND ||
462 wqe->wr.opcode == IB_WR_SEND_WITH_IMM ||
463 wqe->wr.opcode == IB_WR_SEND_WITH_INV)
464 rxe_counter_inc(rxe, RXE_CNT_RDMA_SEND);
465
466 /*
467 * we completed something so let req run again
468 * if it is trying to fence
469 */
470 if (qp->req.wait_fence) {
471 qp->req.wait_fence = 0;
472 rxe_run_task(&qp->req.task);
473 }
474 }
475
complete_ack(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)476 static inline enum comp_state complete_ack(struct rxe_qp *qp,
477 struct rxe_pkt_info *pkt,
478 struct rxe_send_wqe *wqe)
479 {
480 if (wqe->has_rd_atomic) {
481 wqe->has_rd_atomic = 0;
482 atomic_inc(&qp->req.rd_atomic);
483 if (qp->req.need_rd_atomic) {
484 qp->comp.timeout_retry = 0;
485 qp->req.need_rd_atomic = 0;
486 rxe_run_task(&qp->req.task);
487 }
488 }
489
490 if (unlikely(qp->req.state == QP_STATE_DRAIN)) {
491 /* state_lock used by requester & completer */
492 spin_lock_bh(&qp->state_lock);
493 if ((qp->req.state == QP_STATE_DRAIN) &&
494 (qp->comp.psn == qp->req.psn)) {
495 qp->req.state = QP_STATE_DRAINED;
496 spin_unlock_bh(&qp->state_lock);
497
498 if (qp->ibqp.event_handler) {
499 struct ib_event ev;
500
501 ev.device = qp->ibqp.device;
502 ev.element.qp = &qp->ibqp;
503 ev.event = IB_EVENT_SQ_DRAINED;
504 qp->ibqp.event_handler(&ev,
505 qp->ibqp.qp_context);
506 }
507 } else {
508 spin_unlock_bh(&qp->state_lock);
509 }
510 }
511
512 do_complete(qp, wqe);
513
514 if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
515 return COMPST_UPDATE_COMP;
516 else
517 return COMPST_DONE;
518 }
519
complete_wqe(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)520 static inline enum comp_state complete_wqe(struct rxe_qp *qp,
521 struct rxe_pkt_info *pkt,
522 struct rxe_send_wqe *wqe)
523 {
524 if (pkt && wqe->state == wqe_state_pending) {
525 if (psn_compare(wqe->last_psn, qp->comp.psn) >= 0) {
526 qp->comp.psn = (wqe->last_psn + 1) & BTH_PSN_MASK;
527 qp->comp.opcode = -1;
528 }
529
530 if (qp->req.wait_psn) {
531 qp->req.wait_psn = 0;
532 rxe_sched_task(&qp->req.task);
533 }
534 }
535
536 do_complete(qp, wqe);
537
538 return COMPST_GET_WQE;
539 }
540
rxe_drain_resp_pkts(struct rxe_qp * qp,bool notify)541 static void rxe_drain_resp_pkts(struct rxe_qp *qp, bool notify)
542 {
543 struct sk_buff *skb;
544 struct rxe_send_wqe *wqe;
545 struct rxe_queue *q = qp->sq.queue;
546
547 while ((skb = skb_dequeue(&qp->resp_pkts))) {
548 rxe_put(qp);
549 kfree_skb(skb);
550 ib_device_put(qp->ibqp.device);
551 }
552
553 while ((wqe = queue_head(q, q->type))) {
554 if (notify) {
555 wqe->status = IB_WC_WR_FLUSH_ERR;
556 do_complete(qp, wqe);
557 } else {
558 queue_advance_consumer(q, q->type);
559 }
560 }
561 }
562
free_pkt(struct rxe_pkt_info * pkt)563 static void free_pkt(struct rxe_pkt_info *pkt)
564 {
565 struct sk_buff *skb = PKT_TO_SKB(pkt);
566 struct rxe_qp *qp = pkt->qp;
567 struct ib_device *dev = qp->ibqp.device;
568
569 kfree_skb(skb);
570 rxe_put(qp);
571 ib_device_put(dev);
572 }
573
rxe_completer(void * arg)574 int rxe_completer(void *arg)
575 {
576 struct rxe_qp *qp = (struct rxe_qp *)arg;
577 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
578 struct rxe_send_wqe *wqe = NULL;
579 struct sk_buff *skb = NULL;
580 struct rxe_pkt_info *pkt = NULL;
581 enum comp_state state;
582 int ret;
583
584 if (!rxe_get(qp))
585 return -EAGAIN;
586
587 if (!qp->valid || qp->comp.state == QP_STATE_ERROR ||
588 qp->comp.state == QP_STATE_RESET) {
589 rxe_drain_resp_pkts(qp, qp->valid &&
590 qp->comp.state == QP_STATE_ERROR);
591 goto exit;
592 }
593
594 if (qp->comp.timeout) {
595 qp->comp.timeout_retry = 1;
596 qp->comp.timeout = 0;
597 } else {
598 qp->comp.timeout_retry = 0;
599 }
600
601 if (qp->req.need_retry)
602 goto exit;
603
604 state = COMPST_GET_ACK;
605
606 while (1) {
607 rxe_dbg_qp(qp, "state = %s\n", comp_state_name[state]);
608 switch (state) {
609 case COMPST_GET_ACK:
610 skb = skb_dequeue(&qp->resp_pkts);
611 if (skb) {
612 pkt = SKB_TO_PKT(skb);
613 qp->comp.timeout_retry = 0;
614 }
615 state = COMPST_GET_WQE;
616 break;
617
618 case COMPST_GET_WQE:
619 state = get_wqe(qp, pkt, &wqe);
620 break;
621
622 case COMPST_CHECK_PSN:
623 state = check_psn(qp, pkt, wqe);
624 break;
625
626 case COMPST_CHECK_ACK:
627 state = check_ack(qp, pkt, wqe);
628 break;
629
630 case COMPST_READ:
631 state = do_read(qp, pkt, wqe);
632 break;
633
634 case COMPST_ATOMIC:
635 state = do_atomic(qp, pkt, wqe);
636 break;
637
638 case COMPST_WRITE_SEND:
639 if (wqe->state == wqe_state_pending &&
640 wqe->last_psn == pkt->psn)
641 state = COMPST_COMP_ACK;
642 else
643 state = COMPST_UPDATE_COMP;
644 break;
645
646 case COMPST_COMP_ACK:
647 state = complete_ack(qp, pkt, wqe);
648 break;
649
650 case COMPST_COMP_WQE:
651 state = complete_wqe(qp, pkt, wqe);
652 break;
653
654 case COMPST_UPDATE_COMP:
655 if (pkt->mask & RXE_END_MASK)
656 qp->comp.opcode = -1;
657 else
658 qp->comp.opcode = pkt->opcode;
659
660 if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
661 qp->comp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
662
663 if (qp->req.wait_psn) {
664 qp->req.wait_psn = 0;
665 rxe_sched_task(&qp->req.task);
666 }
667
668 state = COMPST_DONE;
669 break;
670
671 case COMPST_DONE:
672 goto done;
673
674 case COMPST_EXIT:
675 if (qp->comp.timeout_retry && wqe) {
676 state = COMPST_ERROR_RETRY;
677 break;
678 }
679
680 /* re reset the timeout counter if
681 * (1) QP is type RC
682 * (2) the QP is alive
683 * (3) there is a packet sent by the requester that
684 * might be acked (we still might get spurious
685 * timeouts but try to keep them as few as possible)
686 * (4) the timeout parameter is set
687 */
688 if ((qp_type(qp) == IB_QPT_RC) &&
689 (qp->req.state == QP_STATE_READY) &&
690 (psn_compare(qp->req.psn, qp->comp.psn) > 0) &&
691 qp->qp_timeout_jiffies)
692 mod_timer(&qp->retrans_timer,
693 jiffies + qp->qp_timeout_jiffies);
694 goto exit;
695
696 case COMPST_ERROR_RETRY:
697 /* we come here if the retry timer fired and we did
698 * not receive a response packet. try to retry the send
699 * queue if that makes sense and the limits have not
700 * been exceeded. remember that some timeouts are
701 * spurious since we do not reset the timer but kick
702 * it down the road or let it expire
703 */
704
705 /* there is nothing to retry in this case */
706 if (!wqe || (wqe->state == wqe_state_posted))
707 goto exit;
708
709 /* if we've started a retry, don't start another
710 * retry sequence, unless this is a timeout.
711 */
712 if (qp->comp.started_retry &&
713 !qp->comp.timeout_retry)
714 goto done;
715
716 if (qp->comp.retry_cnt > 0) {
717 if (qp->comp.retry_cnt != 7)
718 qp->comp.retry_cnt--;
719
720 /* no point in retrying if we have already
721 * seen the last ack that the requester could
722 * have caused
723 */
724 if (psn_compare(qp->req.psn,
725 qp->comp.psn) > 0) {
726 /* tell the requester to retry the
727 * send queue next time around
728 */
729 rxe_counter_inc(rxe,
730 RXE_CNT_COMP_RETRY);
731 qp->req.need_retry = 1;
732 qp->comp.started_retry = 1;
733 rxe_run_task(&qp->req.task);
734 }
735 goto done;
736
737 } else {
738 rxe_counter_inc(rxe, RXE_CNT_RETRY_EXCEEDED);
739 wqe->status = IB_WC_RETRY_EXC_ERR;
740 state = COMPST_ERROR;
741 }
742 break;
743
744 case COMPST_RNR_RETRY:
745 /* we come here if we received an RNR NAK */
746 if (qp->comp.rnr_retry > 0) {
747 if (qp->comp.rnr_retry != 7)
748 qp->comp.rnr_retry--;
749
750 /* don't start a retry flow until the
751 * rnr timer has fired
752 */
753 qp->req.wait_for_rnr_timer = 1;
754 rxe_dbg_qp(qp, "set rnr nak timer\n");
755 mod_timer(&qp->rnr_nak_timer,
756 jiffies + rnrnak_jiffies(aeth_syn(pkt)
757 & ~AETH_TYPE_MASK));
758 goto exit;
759 } else {
760 rxe_counter_inc(rxe,
761 RXE_CNT_RNR_RETRY_EXCEEDED);
762 wqe->status = IB_WC_RNR_RETRY_EXC_ERR;
763 state = COMPST_ERROR;
764 }
765 break;
766
767 case COMPST_ERROR:
768 WARN_ON_ONCE(wqe->status == IB_WC_SUCCESS);
769 do_complete(qp, wqe);
770 rxe_qp_error(qp);
771 goto exit;
772 }
773 }
774
775 /* A non-zero return value will cause rxe_do_task to
776 * exit its loop and end the tasklet. A zero return
777 * will continue looping and return to rxe_completer
778 */
779 done:
780 ret = 0;
781 goto out;
782 exit:
783 ret = -EAGAIN;
784 out:
785 if (pkt)
786 free_pkt(pkt);
787 rxe_put(qp);
788
789 return ret;
790 }
791