ipath_rc.c 51.0 KB
Newer Older
1
/*
J
John Gregor 已提交
2
 * Copyright (c) 2006, 2007 QLogic Corporation. All rights reserved.
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the
 * OpenIB.org BSD license below:
 *
 *     Redistribution and use in source and binary forms, with or
 *     without modification, are permitted provided that the following
 *     conditions are met:
 *
 *      - Redistributions of source code must retain the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer.
 *
 *      - Redistributions in binary form must reproduce the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer in the documentation and/or other materials
 *        provided with the distribution.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */

#include "ipath_verbs.h"
35
#include "ipath_kernel.h"
36 37 38 39

/* cut down ridiculously long IB macro names */
#define OP(x) IB_OPCODE_RC_##x

40 41 42 43 44 45 46 47 48 49 50 51 52
static u32 restart_sge(struct ipath_sge_state *ss, struct ipath_swqe *wqe,
		       u32 psn, u32 pmtu)
{
	u32 len;

	len = ((psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
	ss->sge = wqe->sg_list[0];
	ss->sg_list = wqe->sg_list + 1;
	ss->num_sge = wqe->wr.num_sge;
	ipath_skip_sge(ss, len);
	return wqe->length - len;
}

53 54 55 56 57
/**
 * ipath_init_restart- initialize the qp->s_sge after a restart
 * @qp: the QP who's SGE we're restarting
 * @wqe: the work queue to initialize the QP's SGE from
 *
58
 * The QP s_lock should be held and interrupts disabled.
59 60 61 62 63
 */
static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
{
	struct ipath_ibdev *dev;

64 65
	qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn,
				ib_mtu_enum_to_int(qp->path_mtu));
66 67
	dev = to_idev(qp->ibqp.device);
	spin_lock(&dev->pending_lock);
68
	if (list_empty(&qp->timerwait))
69 70 71 72 73 74 75 76 77 78 79
		list_add_tail(&qp->timerwait,
			      &dev->pending[dev->pending_index]);
	spin_unlock(&dev->pending_lock);
}

/**
 * ipath_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
 * @qp: a pointer to the QP
 * @ohdr: a pointer to the IB header being constructed
 * @pmtu: the path MTU
 *
80 81
 * Return 1 if constructed; otherwise, return 0.
 * Note that we are in the responder's side of the QP context.
82 83
 * Note the QP s_lock must be held.
 */
84 85
static int ipath_make_rc_ack(struct ipath_ibdev *dev, struct ipath_qp *qp,
			     struct ipath_other_headers *ohdr, u32 pmtu)
86
{
87
	struct ipath_ack_entry *e;
88 89 90
	u32 hwords;
	u32 len;
	u32 bth0;
91
	u32 bth2;
92 93 94 95 96

	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
	hwords = 5;

	switch (qp->s_ack_state) {
97 98 99
	case OP(RDMA_READ_RESPONSE_LAST):
	case OP(RDMA_READ_RESPONSE_ONLY):
	case OP(ATOMIC_ACKNOWLEDGE):
100 101 102 103 104 105 106
		/*
		 * We can increment the tail pointer now that the last
		 * response has been sent instead of only being
		 * constructed.
		 */
		if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
			qp->s_tail_ack_queue = 0;
107
		/* FALLTHROUGH */
108
	case OP(SEND_ONLY):
109 110 111 112 113
	case OP(ACKNOWLEDGE):
		/* Check for no next entry in the queue. */
		if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
			if (qp->s_flags & IPATH_S_ACK_PENDING)
				goto normal;
114
			qp->s_ack_state = OP(ACKNOWLEDGE);
115 116 117 118 119 120 121 122 123 124 125 126
			goto bail;
		}

		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
		if (e->opcode == OP(RDMA_READ_REQUEST)) {
			/* Copy SGE state in case we need to resend */
			qp->s_ack_rdma_sge = e->rdma_sge;
			qp->s_cur_sge = &qp->s_ack_rdma_sge;
			len = e->rdma_sge.sge.sge_length;
			if (len > pmtu) {
				len = pmtu;
				qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
127
			} else {
128
				qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
129 130
				e->sent = 1;
			}
131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146
			ohdr->u.aeth = ipath_compute_aeth(qp);
			hwords++;
			qp->s_ack_rdma_psn = e->psn;
			bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
		} else {
			/* COMPARE_SWAP or FETCH_ADD */
			qp->s_cur_sge = NULL;
			len = 0;
			qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
			ohdr->u.at.aeth = ipath_compute_aeth(qp);
			ohdr->u.at.atomic_ack_eth[0] =
				cpu_to_be32(e->atomic_data >> 32);
			ohdr->u.at.atomic_ack_eth[1] =
				cpu_to_be32(e->atomic_data);
			hwords += sizeof(ohdr->u.at) / sizeof(u32);
			bth2 = e->psn;
147
			e->sent = 1;
148
		}
149 150 151 152 153 154 155
		bth0 = qp->s_ack_state << 24;
		break;

	case OP(RDMA_READ_RESPONSE_FIRST):
		qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
		/* FALLTHROUGH */
	case OP(RDMA_READ_RESPONSE_MIDDLE):
156
		len = qp->s_ack_rdma_sge.sge.sge_length;
157 158 159 160 161 162
		if (len > pmtu)
			len = pmtu;
		else {
			ohdr->u.aeth = ipath_compute_aeth(qp);
			hwords++;
			qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
163
			qp->s_ack_queue[qp->s_tail_ack_queue].sent = 1;
164 165
		}
		bth0 = qp->s_ack_state << 24;
166
		bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
167 168 169
		break;

	default:
170
	normal:
171
		/*
172 173 174 175
		 * Send a regular ACK.
		 * Set the s_ack_state so we wait until after sending
		 * the ACK before setting s_ack_state to ACKNOWLEDGE
		 * (see above).
176
		 */
177
		qp->s_ack_state = OP(SEND_ONLY);
178 179
		qp->s_flags &= ~IPATH_S_ACK_PENDING;
		qp->s_cur_sge = NULL;
180
		if (qp->s_nak_state)
181 182 183 184
			ohdr->u.aeth =
				cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
					    (qp->s_nak_state <<
					     IPATH_AETH_CREDIT_SHIFT));
185 186
		else
			ohdr->u.aeth = ipath_compute_aeth(qp);
187
		hwords++;
188 189 190
		len = 0;
		bth0 = OP(ACKNOWLEDGE) << 24;
		bth2 = qp->s_ack_psn & IPATH_PSN_MASK;
191 192 193
	}
	qp->s_hdrwords = hwords;
	qp->s_cur_size = len;
194
	ipath_make_ruc_header(dev, qp, ohdr, bth0, bth2);
195
	return 1;
196

197
bail:
198
	return 0;
199 200 201 202 203 204 205 206
}

/**
 * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
 * @qp: a pointer to the QP
 *
 * Return 1 if constructed; otherwise, return 0.
 */
207
int ipath_make_rc_req(struct ipath_qp *qp)
208 209
{
	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
210
	struct ipath_other_headers *ohdr;
211 212 213 214 215 216
	struct ipath_sge_state *ss;
	struct ipath_swqe *wqe;
	u32 hwords;
	u32 len;
	u32 bth0;
	u32 bth2;
217
	u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
218
	char newreq;
219 220 221 222 223 224 225 226 227 228 229 230
	unsigned long flags;
	int ret = 0;

	ohdr = &qp->s_hdr.u.oth;
	if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
		ohdr = &qp->s_hdr.u.l.oth;

	/*
	 * The lock is needed to synchronize between the sending tasklet,
	 * the receive interrupt handler, and timeout resends.
	 */
	spin_lock_irqsave(&qp->s_lock, flags);
231

232 233 234
	/* Sending responses has higher priority over sending requests. */
	if ((qp->r_head_ack_queue != qp->s_tail_ack_queue ||
	     (qp->s_flags & IPATH_S_ACK_PENDING) ||
235
	     qp->s_ack_state != OP(ACKNOWLEDGE)) &&
236
	    ipath_make_rc_ack(dev, qp, ohdr, pmtu))
237 238
		goto done;

239
	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
240
	    qp->s_rnr_timeout || qp->s_wait_credit)
241
		goto bail;
242

243 244 245 246
	/* Limit the number of packets sent without an ACK. */
	if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT) > 0) {
		qp->s_wait_credit = 1;
		dev->n_rc_stalls++;
247
		goto bail;
248 249
	}

250 251
	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
	hwords = 5;
252
	bth0 = 1 << 22; /* Set M bit */
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268

	/* Send a request. */
	wqe = get_swqe_ptr(qp, qp->s_cur);
	switch (qp->s_state) {
	default:
		/*
		 * Resend an old request or start a new one.
		 *
		 * We keep track of the current SWQE so that
		 * we don't reset the "furthest progress" state
		 * if we need to back up.
		 */
		newreq = 0;
		if (qp->s_cur == qp->s_tail) {
			/* Check if send work queue is empty. */
			if (qp->s_tail == qp->s_head)
269 270 271 272 273 274 275 276 277 278
				goto bail;
			/*
			 * If a fence is requested, wait for previous
			 * RDMA read and atomic operations to finish.
			 */
			if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
			    qp->s_num_rd_atomic) {
				qp->s_flags |= IPATH_S_FENCE_PENDING;
				goto bail;
			}
279
			wqe->psn = qp->s_next_psn;
280 281 282 283 284 285 286
			newreq = 1;
		}
		/*
		 * Note that we have to be careful not to modify the
		 * original work request since we may need to resend
		 * it.
		 */
R
Ralph Campbell 已提交
287
		len = wqe->length;
288 289 290 291 292 293 294 295
		ss = &qp->s_sge;
		bth2 = 0;
		switch (wqe->wr.opcode) {
		case IB_WR_SEND:
		case IB_WR_SEND_WITH_IMM:
			/* If no credit, return. */
			if (qp->s_lsn != (u32) -1 &&
			    ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
296
				goto bail;
297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319
			wqe->lpsn = wqe->psn;
			if (len > pmtu) {
				wqe->lpsn += (len - 1) / pmtu;
				qp->s_state = OP(SEND_FIRST);
				len = pmtu;
				break;
			}
			if (wqe->wr.opcode == IB_WR_SEND)
				qp->s_state = OP(SEND_ONLY);
			else {
				qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
				/* Immediate data comes after the BTH */
				ohdr->u.imm_data = wqe->wr.imm_data;
				hwords += 1;
			}
			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
				bth0 |= 1 << 23;
			bth2 = 1 << 31;	/* Request ACK. */
			if (++qp->s_cur == qp->s_size)
				qp->s_cur = 0;
			break;

		case IB_WR_RDMA_WRITE:
320
			if (newreq && qp->s_lsn != (u32) -1)
321 322 323 324 325 326
				qp->s_lsn++;
			/* FALLTHROUGH */
		case IB_WR_RDMA_WRITE_WITH_IMM:
			/* If no credit, return. */
			if (qp->s_lsn != (u32) -1 &&
			    ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
327
				goto bail;
328 329 330 331 332
			ohdr->u.rc.reth.vaddr =
				cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
			ohdr->u.rc.reth.rkey =
				cpu_to_be32(wqe->wr.wr.rdma.rkey);
			ohdr->u.rc.reth.length = cpu_to_be32(len);
333
			hwords += sizeof(struct ib_reth) / sizeof(u32);
334 335 336 337 338 339 340 341 342 343 344 345
			wqe->lpsn = wqe->psn;
			if (len > pmtu) {
				wqe->lpsn += (len - 1) / pmtu;
				qp->s_state = OP(RDMA_WRITE_FIRST);
				len = pmtu;
				break;
			}
			if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
				qp->s_state = OP(RDMA_WRITE_ONLY);
			else {
				qp->s_state =
					OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
346
				/* Immediate data comes after RETH */
347 348 349 350 351 352 353 354 355 356 357
				ohdr->u.rc.imm_data = wqe->wr.imm_data;
				hwords += 1;
				if (wqe->wr.send_flags & IB_SEND_SOLICITED)
					bth0 |= 1 << 23;
			}
			bth2 = 1 << 31;	/* Request ACK. */
			if (++qp->s_cur == qp->s_size)
				qp->s_cur = 0;
			break;

		case IB_WR_RDMA_READ:
358 359 360 361
			/*
			 * Don't allow more operations to be started
			 * than the QP limits allow.
			 */
362
			if (newreq) {
363 364 365 366 367 368
				if (qp->s_num_rd_atomic >=
				    qp->s_max_rd_atomic) {
					qp->s_flags |= IPATH_S_RDMAR_PENDING;
					goto bail;
				}
				qp->s_num_rd_atomic++;
369 370
				if (qp->s_lsn != (u32) -1)
					qp->s_lsn++;
371 372 373 374 375 376 377 378
				/*
				 * Adjust s_next_psn to count the
				 * expected number of responses.
				 */
				if (len > pmtu)
					qp->s_next_psn += (len - 1) / pmtu;
				wqe->lpsn = qp->s_next_psn++;
			}
379 380 381 382 383 384 385
			ohdr->u.rc.reth.vaddr =
				cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
			ohdr->u.rc.reth.rkey =
				cpu_to_be32(wqe->wr.wr.rdma.rkey);
			ohdr->u.rc.reth.length = cpu_to_be32(len);
			qp->s_state = OP(RDMA_READ_REQUEST);
			hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
386 387 388 389 390 391 392 393
			ss = NULL;
			len = 0;
			if (++qp->s_cur == qp->s_size)
				qp->s_cur = 0;
			break;

		case IB_WR_ATOMIC_CMP_AND_SWP:
		case IB_WR_ATOMIC_FETCH_AND_ADD:
394 395 396 397
			/*
			 * Don't allow more operations to be started
			 * than the QP limits allow.
			 */
398
			if (newreq) {
399 400 401 402 403 404
				if (qp->s_num_rd_atomic >=
				    qp->s_max_rd_atomic) {
					qp->s_flags |= IPATH_S_RDMAR_PENDING;
					goto bail;
				}
				qp->s_num_rd_atomic++;
405 406
				if (qp->s_lsn != (u32) -1)
					qp->s_lsn++;
407 408
				wqe->lpsn = wqe->psn;
			}
409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427
			if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
				qp->s_state = OP(COMPARE_SWAP);
				ohdr->u.atomic_eth.swap_data = cpu_to_be64(
					wqe->wr.wr.atomic.swap);
				ohdr->u.atomic_eth.compare_data = cpu_to_be64(
					wqe->wr.wr.atomic.compare_add);
			} else {
				qp->s_state = OP(FETCH_ADD);
				ohdr->u.atomic_eth.swap_data = cpu_to_be64(
					wqe->wr.wr.atomic.compare_add);
				ohdr->u.atomic_eth.compare_data = 0;
			}
			ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
				wqe->wr.wr.atomic.remote_addr >> 32);
			ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
				wqe->wr.wr.atomic.remote_addr);
			ohdr->u.atomic_eth.rkey = cpu_to_be32(
				wqe->wr.wr.atomic.rkey);
			hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
428 429
			ss = NULL;
			len = 0;
430 431
			if (++qp->s_cur == qp->s_size)
				qp->s_cur = 0;
432 433 434
			break;

		default:
435
			goto bail;
436
		}
R
Ralph Campbell 已提交
437 438 439 440
		qp->s_sge.sge = wqe->sg_list[0];
		qp->s_sge.sg_list = wqe->sg_list + 1;
		qp->s_sge.num_sge = wqe->wr.num_sge;
		qp->s_len = wqe->length;
441 442 443 444 445
		if (newreq) {
			qp->s_tail++;
			if (qp->s_tail >= qp->s_size)
				qp->s_tail = 0;
		}
R
Ralph Campbell 已提交
446 447 448 449 450
		bth2 |= qp->s_psn & IPATH_PSN_MASK;
		if (wqe->wr.opcode == IB_WR_RDMA_READ)
			qp->s_psn = wqe->lpsn + 1;
		else {
			qp->s_psn++;
451
			if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
R
Ralph Campbell 已提交
452 453
				qp->s_next_psn = qp->s_psn;
		}
454 455 456 457 458
		/*
		 * Put the QP on the pending list so lost ACKs will cause
		 * a retry.  More than one request can be pending so the
		 * QP may already be on the dev->pending list.
		 */
459
		spin_lock(&dev->pending_lock);
460
		if (list_empty(&qp->timerwait))
461 462 463 464 465 466 467
			list_add_tail(&qp->timerwait,
				      &dev->pending[dev->pending_index]);
		spin_unlock(&dev->pending_lock);
		break;

	case OP(RDMA_READ_RESPONSE_FIRST):
		/*
468 469
		 * This case can only happen if a send is restarted.
		 * See ipath_restart_rc().
470 471 472 473 474 475 476
		 */
		ipath_init_restart(qp, wqe);
		/* FALLTHROUGH */
	case OP(SEND_FIRST):
		qp->s_state = OP(SEND_MIDDLE);
		/* FALLTHROUGH */
	case OP(SEND_MIDDLE):
477
		bth2 = qp->s_psn++ & IPATH_PSN_MASK;
478
		if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512
			qp->s_next_psn = qp->s_psn;
		ss = &qp->s_sge;
		len = qp->s_len;
		if (len > pmtu) {
			len = pmtu;
			break;
		}
		if (wqe->wr.opcode == IB_WR_SEND)
			qp->s_state = OP(SEND_LAST);
		else {
			qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
			/* Immediate data comes after the BTH */
			ohdr->u.imm_data = wqe->wr.imm_data;
			hwords += 1;
		}
		if (wqe->wr.send_flags & IB_SEND_SOLICITED)
			bth0 |= 1 << 23;
		bth2 |= 1 << 31;	/* Request ACK. */
		qp->s_cur++;
		if (qp->s_cur >= qp->s_size)
			qp->s_cur = 0;
		break;

	case OP(RDMA_READ_RESPONSE_LAST):
		/*
		 * This case can only happen if a RDMA write is restarted.
		 * See ipath_restart_rc().
		 */
		ipath_init_restart(qp, wqe);
		/* FALLTHROUGH */
	case OP(RDMA_WRITE_FIRST):
		qp->s_state = OP(RDMA_WRITE_MIDDLE);
		/* FALLTHROUGH */
	case OP(RDMA_WRITE_MIDDLE):
513
		bth2 = qp->s_psn++ & IPATH_PSN_MASK;
514
		if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543
			qp->s_next_psn = qp->s_psn;
		ss = &qp->s_sge;
		len = qp->s_len;
		if (len > pmtu) {
			len = pmtu;
			break;
		}
		if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
			qp->s_state = OP(RDMA_WRITE_LAST);
		else {
			qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
			/* Immediate data comes after the BTH */
			ohdr->u.imm_data = wqe->wr.imm_data;
			hwords += 1;
			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
				bth0 |= 1 << 23;
		}
		bth2 |= 1 << 31;	/* Request ACK. */
		qp->s_cur++;
		if (qp->s_cur >= qp->s_size)
			qp->s_cur = 0;
		break;

	case OP(RDMA_READ_RESPONSE_MIDDLE):
		/*
		 * This case can only happen if a RDMA read is restarted.
		 * See ipath_restart_rc().
		 */
		ipath_init_restart(qp, wqe);
544
		len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
545 546 547 548 549 550
		ohdr->u.rc.reth.vaddr =
			cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
		ohdr->u.rc.reth.rkey =
			cpu_to_be32(wqe->wr.wr.rdma.rkey);
		ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
		qp->s_state = OP(RDMA_READ_REQUEST);
551
		hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
552
		bth2 = qp->s_psn++ & IPATH_PSN_MASK;
553
		if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
554 555 556 557 558 559 560 561
			qp->s_next_psn = qp->s_psn;
		ss = NULL;
		len = 0;
		qp->s_cur++;
		if (qp->s_cur == qp->s_size)
			qp->s_cur = 0;
		break;
	}
562 563
	if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0)
		bth2 |= 1 << 31;	/* Request ACK. */
564 565 566 567
	qp->s_len -= len;
	qp->s_hdrwords = hwords;
	qp->s_cur_sge = ss;
	qp->s_cur_size = len;
568
	ipath_make_ruc_header(dev, qp, ohdr, bth0 | (qp->s_state << 24), bth2);
569
done:
570
	ret = 1;
571
bail:
572 573
	spin_unlock_irqrestore(&qp->s_lock, flags);
	return ret;
574 575 576
}

/**
577 578
 * send_rc_ack - Construct an ACK packet and send it
 * @qp: a pointer to the QP
579
 *
580 581
 * This is called from ipath_rc_rcv() and only uses the receive
 * side QP state.
582 583
 * Note that RDMA reads and atomics are handled in the
 * send side QP state and tasklet.
584 585 586 587 588 589
 */
static void send_rc_ack(struct ipath_qp *qp)
{
	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
	u16 lrh0;
	u32 bth0;
590 591
	u32 hwords;
	struct ipath_ib_header hdr;
592
	struct ipath_other_headers *ohdr;
593
	unsigned long flags;
594

595
	/* Don't send ACK or NAK if a RDMA read or atomic is pending. */
596 597 598
	if (qp->r_head_ack_queue != qp->s_tail_ack_queue ||
	    (qp->s_flags & IPATH_S_ACK_PENDING) ||
	    qp->s_ack_state != OP(ACKNOWLEDGE))
599 600
		goto queue_ack;

601
	/* Construct the header. */
602
	ohdr = &hdr.u.oth;
603
	lrh0 = IPATH_LRH_BTH;
604
	/* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
605
	hwords = 6;
606
	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
607 608 609 610
		hwords += ipath_make_grh(dev, &hdr.u.l.grh,
					 &qp->remote_ah_attr.grh,
					 hwords, 0);
		ohdr = &hdr.u.l.oth;
611
		lrh0 = IPATH_LRH_GRH;
612
	}
613
	/* read pkey_index w/o lock (its atomic) */
614
	bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index) |
615
		(OP(ACKNOWLEDGE) << 24) | (1 << 22);
616
	if (qp->r_nak_state)
617
		ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
618
					    (qp->r_nak_state <<
619
					     IPATH_AETH_CREDIT_SHIFT));
620 621
	else
		ohdr->u.aeth = ipath_compute_aeth(qp);
622
	lrh0 |= qp->remote_ah_attr.sl << 4;
623 624 625
	hdr.lrh[0] = cpu_to_be16(lrh0);
	hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
626
	hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
627 628
	ohdr->bth[0] = cpu_to_be32(bth0);
	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
629
	ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
630 631 632 633

	/*
	 * If we can send the ACK, clear the ACK state.
	 */
634
	if (ipath_verbs_send(qp, &hdr, hwords, NULL, 0) == 0) {
635
		dev->n_unicast_xmit++;
636
		goto done;
637
	}
638 639 640 641 642 643 644 645 646 647 648

	/*
	 * We are out of PIO buffers at the moment.
	 * Pass responsibility for sending the ACK to the
	 * send tasklet so that when a PIO buffer becomes
	 * available, the ACK is sent ahead of other outgoing
	 * packets.
	 */
	dev->n_rc_qacks++;

queue_ack:
649
	spin_lock_irqsave(&qp->s_lock, flags);
650 651 652
	qp->s_flags |= IPATH_S_ACK_PENDING;
	qp->s_nak_state = qp->r_nak_state;
	qp->s_ack_psn = qp->r_ack_psn;
653
	spin_unlock_irqrestore(&qp->s_lock, flags);
654 655 656 657 658 659

	/* Call ipath_do_rc_send() in another thread. */
	tasklet_hi_schedule(&qp->s_task);

done:
	return;
660 661
}

662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743
/**
 * reset_psn - reset the QP state to send starting from PSN
 * @qp: the QP
 * @psn: the packet sequence number to restart at
 *
 * This is called from ipath_rc_rcv() to process an incoming RC ACK
 * for the given QP.
 * Called at interrupt level with the QP s_lock held.
 */
static void reset_psn(struct ipath_qp *qp, u32 psn)
{
	u32 n = qp->s_last;
	struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
	u32 opcode;

	qp->s_cur = n;

	/*
	 * If we are starting the request from the beginning,
	 * let the normal send code handle initialization.
	 */
	if (ipath_cmp24(psn, wqe->psn) <= 0) {
		qp->s_state = OP(SEND_LAST);
		goto done;
	}

	/* Find the work request opcode corresponding to the given PSN. */
	opcode = wqe->wr.opcode;
	for (;;) {
		int diff;

		if (++n == qp->s_size)
			n = 0;
		if (n == qp->s_tail)
			break;
		wqe = get_swqe_ptr(qp, n);
		diff = ipath_cmp24(psn, wqe->psn);
		if (diff < 0)
			break;
		qp->s_cur = n;
		/*
		 * If we are starting the request from the beginning,
		 * let the normal send code handle initialization.
		 */
		if (diff == 0) {
			qp->s_state = OP(SEND_LAST);
			goto done;
		}
		opcode = wqe->wr.opcode;
	}

	/*
	 * Set the state to restart in the middle of a request.
	 * Don't change the s_sge, s_cur_sge, or s_cur_size.
	 * See ipath_do_rc_send().
	 */
	switch (opcode) {
	case IB_WR_SEND:
	case IB_WR_SEND_WITH_IMM:
		qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
		break;

	case IB_WR_RDMA_WRITE:
	case IB_WR_RDMA_WRITE_WITH_IMM:
		qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
		break;

	case IB_WR_RDMA_READ:
		qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
		break;

	default:
		/*
		 * This case shouldn't happen since its only
		 * one PSN per req.
		 */
		qp->s_state = OP(SEND_LAST);
	}
done:
	qp->s_psn = psn;
}

744 745 746 747 748 749
/**
 * ipath_restart_rc - back up requester to resend the last un-ACKed request
 * @qp: the QP to restart
 * @psn: packet sequence number for the request
 * @wc: the work completion request
 *
750
 * The QP s_lock should be held and interrupts disabled.
751 752 753 754 755 756 757 758 759 760 761 762
 */
void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
{
	struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
	struct ipath_ibdev *dev;

	if (qp->s_retry == 0) {
		wc->wr_id = wqe->wr.wr_id;
		wc->status = IB_WC_RETRY_EXC_ERR;
		wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
		wc->vendor_err = 0;
		wc->byte_len = 0;
763
		wc->qp = &qp->ibqp;
764
		wc->imm_data = 0;
765
		wc->src_qp = qp->remote_qpn;
766
		wc->wc_flags = 0;
767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782
		wc->pkey_index = 0;
		wc->slid = qp->remote_ah_attr.dlid;
		wc->sl = qp->remote_ah_attr.sl;
		wc->dlid_path_bits = 0;
		wc->port_num = 0;
		ipath_sqerror_qp(qp, wc);
		goto bail;
	}
	qp->s_retry--;

	/*
	 * Remove the QP from the timeout queue.
	 * Note: it may already have been removed by ipath_ib_timer().
	 */
	dev = to_idev(qp->ibqp.device);
	spin_lock(&dev->pending_lock);
783 784
	if (!list_empty(&qp->timerwait))
		list_del_init(&qp->timerwait);
785 786 787 788 789
	spin_unlock(&dev->pending_lock);

	if (wqe->wr.opcode == IB_WR_RDMA_READ)
		dev->n_rc_resends++;
	else
790
		dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
791

792
	reset_psn(qp, psn);
793 794 795 796 797 798
	tasklet_hi_schedule(&qp->s_task);

bail:
	return;
}

799 800 801 802 803 804 805 806 807
static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
{
	if (qp->s_wait_credit) {
		qp->s_wait_credit = 0;
		tasklet_hi_schedule(&qp->s_task);
	}
	qp->s_last_psn = psn;
}

808 809 810 811 812 813
/**
 * do_rc_ack - process an incoming RC ACK
 * @qp: the QP the ACK came in on
 * @psn: the packet sequence number of the ACK
 * @opcode: the opcode of the request that resulted in the ACK
 *
814
 * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
815
 * for the given QP.
816
 * Called at interrupt level with the QP s_lock held and interrupts disabled.
817 818
 * Returns 1 if OK, 0 if current operation should be aborted (NAK).
 */
819 820
static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode,
		     u64 val)
821 822 823 824 825
{
	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
	struct ib_wc wc;
	struct ipath_swqe *wqe;
	int ret = 0;
R
Ralph Campbell 已提交
826
	u32 ack_psn;
827
	int diff;
828 829 830 831 832 833 834 835

	/*
	 * Remove the QP from the timeout queue (or RNR timeout queue).
	 * If ipath_ib_timer() has already removed it,
	 * it's OK since we hold the QP s_lock and ipath_restart_rc()
	 * just won't find anything to restart if we ACK everything.
	 */
	spin_lock(&dev->pending_lock);
836 837
	if (!list_empty(&qp->timerwait))
		list_del_init(&qp->timerwait);
838 839 840 841 842 843 844 845
	spin_unlock(&dev->pending_lock);

	/*
	 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
	 * requests and implicitly NAK RDMA read and atomic requests issued
	 * before the NAK'ed request.  The MSN won't include the NAK'ed
	 * request but will include an ACK'ed request(s).
	 */
R
Ralph Campbell 已提交
846 847 848
	ack_psn = psn;
	if (aeth >> 29)
		ack_psn--;
849 850 851 852 853 854
	wqe = get_swqe_ptr(qp, qp->s_last);

	/*
	 * The MSN might be for a later WQE than the PSN indicates so
	 * only complete WQEs that the PSN finishes.
	 */
855 856 857 858 859 860 861 862 863 864 865 866 867
	while ((diff = ipath_cmp24(ack_psn, wqe->lpsn)) >= 0) {
		/*
		 * RDMA_READ_RESPONSE_ONLY is a special case since
		 * we want to generate completion events for everything
		 * before the RDMA read, copy the data, then generate
		 * the completion for the read.
		 */
		if (wqe->wr.opcode == IB_WR_RDMA_READ &&
		    opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
		    diff == 0) {
			ret = 1;
			goto bail;
		}
868 869 870 871 872 873 874 875 876 877
		/*
		 * If this request is a RDMA read or atomic, and the ACK is
		 * for a later operation, this ACK NAKs the RDMA read or
		 * atomic.  In other words, only a RDMA_READ_LAST or ONLY
		 * can ACK a RDMA read and likewise for atomic ops.  Note
		 * that the NAK case can only happen if relaxed ordering is
		 * used and requests are sent after an RDMA read or atomic
		 * is sent but before the response is received.
		 */
		if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
878
		     (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
879 880
		    ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
		      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
881
		     (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
882 883 884 885
			/*
			 * The last valid PSN seen is the previous
			 * request's.
			 */
886
			update_last_psn(qp, wqe->psn - 1);
887 888 889 890 891 892 893 894
			/* Retry this request. */
			ipath_restart_rc(qp, wqe->psn, &wc);
			/*
			 * No need to process the ACK/NAK since we are
			 * restarting an earlier request.
			 */
			goto bail;
		}
895 896 897
		if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
			*(u64 *) wqe->sg_list[0].vaddr = val;
898 899 900 901 902 903 904 905 906 907 908 909 910 911 912
		if (qp->s_num_rd_atomic &&
		    (wqe->wr.opcode == IB_WR_RDMA_READ ||
		     wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
		     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
			qp->s_num_rd_atomic--;
			/* Restart sending task if fence is complete */
			if ((qp->s_flags & IPATH_S_FENCE_PENDING) &&
			    !qp->s_num_rd_atomic) {
				qp->s_flags &= ~IPATH_S_FENCE_PENDING;
				tasklet_hi_schedule(&qp->s_task);
			} else if (qp->s_flags & IPATH_S_RDMAR_PENDING) {
				qp->s_flags &= ~IPATH_S_RDMAR_PENDING;
				tasklet_hi_schedule(&qp->s_task);
			}
		}
913
		/* Post a send completion queue entry if requested. */
914
		if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
915 916 917 918 919 920
		    (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
			wc.wr_id = wqe->wr.wr_id;
			wc.status = IB_WC_SUCCESS;
			wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
			wc.vendor_err = 0;
			wc.byte_len = wqe->length;
921
			wc.imm_data = 0;
922
			wc.qp = &qp->ibqp;
923
			wc.src_qp = qp->remote_qpn;
924
			wc.wc_flags = 0;
925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940
			wc.pkey_index = 0;
			wc.slid = qp->remote_ah_attr.dlid;
			wc.sl = qp->remote_ah_attr.sl;
			wc.dlid_path_bits = 0;
			wc.port_num = 0;
			ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
		}
		qp->s_retry = qp->s_retry_cnt;
		/*
		 * If we are completing a request which is in the process of
		 * being resent, we can stop resending it since we know the
		 * responder has already seen it.
		 */
		if (qp->s_last == qp->s_cur) {
			if (++qp->s_cur >= qp->s_size)
				qp->s_cur = 0;
941 942 943
			qp->s_last = qp->s_cur;
			if (qp->s_last == qp->s_tail)
				break;
944 945 946
			wqe = get_swqe_ptr(qp, qp->s_cur);
			qp->s_state = OP(SEND_LAST);
			qp->s_psn = wqe->psn;
947 948 949 950 951 952
		} else {
			if (++qp->s_last >= qp->s_size)
				qp->s_last = 0;
			if (qp->s_last == qp->s_tail)
				break;
			wqe = get_swqe_ptr(qp, qp->s_last);
953 954 955 956 957 958 959 960 961 962 963 964
		}
	}

	switch (aeth >> 29) {
	case 0:		/* ACK */
		dev->n_rc_acks++;
		/* If this is a partial ACK, reset the retransmit timer. */
		if (qp->s_last != qp->s_tail) {
			spin_lock(&dev->pending_lock);
			list_add_tail(&qp->timerwait,
				      &dev->pending[dev->pending_index]);
			spin_unlock(&dev->pending_lock);
965 966 967 968 969 970 971 972 973 974 975 976
			/*
			 * If we get a partial ACK for a resent operation,
			 * we can stop resending the earlier packets and
			 * continue with the next packet the receiver wants.
			 */
			if (ipath_cmp24(qp->s_psn, psn) <= 0) {
				reset_psn(qp, psn + 1);
				tasklet_hi_schedule(&qp->s_task);
			}
		} else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
			qp->s_state = OP(SEND_LAST);
			qp->s_psn = psn + 1;
977 978 979 980
		}
		ipath_get_credit(qp, aeth);
		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
		qp->s_retry = qp->s_retry_cnt;
981
		update_last_psn(qp, psn);
982 983 984 985 986
		ret = 1;
		goto bail;

	case 1:		/* RNR NAK */
		dev->n_rnr_naks++;
987 988
		if (qp->s_last == qp->s_tail)
			goto bail;
989 990 991 992 993 994 995
		if (qp->s_rnr_retry == 0) {
			wc.status = IB_WC_RNR_RETRY_EXC_ERR;
			goto class_b;
		}
		if (qp->s_rnr_retry_cnt < 7)
			qp->s_rnr_retry--;

996
		/* The last valid PSN is the previous PSN. */
997
		update_last_psn(qp, psn - 1);
998

999 1000 1001 1002 1003
		if (wqe->wr.opcode == IB_WR_RDMA_READ)
			dev->n_rc_resends++;
		else
			dev->n_rc_resends +=
				(qp->s_psn - psn) & IPATH_PSN_MASK;
1004

1005
		reset_psn(qp, psn);
1006 1007

		qp->s_rnr_timeout =
1008 1009
			ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
					   IPATH_AETH_CREDIT_MASK];
1010 1011 1012 1013
		ipath_insert_rnr_queue(qp);
		goto bail;

	case 3:		/* NAK */
1014 1015 1016 1017
		if (qp->s_last == qp->s_tail)
			goto bail;
		/* The last valid PSN is the previous PSN. */
		update_last_psn(qp, psn - 1);
1018 1019
		switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
			IPATH_AETH_CREDIT_MASK) {
1020 1021 1022
		case 0:	/* PSN sequence error */
			dev->n_seq_naks++;
			/*
1023
			 * Back up to the responder's expected PSN.
1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048
			 * Note that we might get a NAK in the middle of an
			 * RDMA READ response which terminates the RDMA
			 * READ.
			 */
			ipath_restart_rc(qp, psn, &wc);
			break;

		case 1:	/* Invalid Request */
			wc.status = IB_WC_REM_INV_REQ_ERR;
			dev->n_other_naks++;
			goto class_b;

		case 2:	/* Remote Access Error */
			wc.status = IB_WC_REM_ACCESS_ERR;
			dev->n_other_naks++;
			goto class_b;

		case 3:	/* Remote Operation Error */
			wc.status = IB_WC_REM_OP_ERR;
			dev->n_other_naks++;
		class_b:
			wc.wr_id = wqe->wr.wr_id;
			wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
			wc.vendor_err = 0;
			wc.byte_len = 0;
1049
			wc.qp = &qp->ibqp;
1050
			wc.imm_data = 0;
1051
			wc.src_qp = qp->remote_qpn;
1052
			wc.wc_flags = 0;
1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102
			wc.pkey_index = 0;
			wc.slid = qp->remote_ah_attr.dlid;
			wc.sl = qp->remote_ah_attr.sl;
			wc.dlid_path_bits = 0;
			wc.port_num = 0;
			ipath_sqerror_qp(qp, &wc);
			break;

		default:
			/* Ignore other reserved NAK error codes */
			goto reserved;
		}
		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
		goto bail;

	default:		/* 2: reserved */
	reserved:
		/* Ignore reserved NAK codes. */
		goto bail;
	}

bail:
	return ret;
}

/**
 * ipath_rc_rcv_resp - process an incoming RC response packet
 * @dev: the device this packet came in on
 * @ohdr: the other headers for this packet
 * @data: the packet data
 * @tlen: the packet length
 * @qp: the QP for this packet
 * @opcode: the opcode for this packet
 * @psn: the packet sequence number for this packet
 * @hdrsize: the header length
 * @pmtu: the path MTU
 * @header_in_data: true if part of the header data is in the data buffer
 *
 * This is called from ipath_rc_rcv() to process an incoming RC response
 * packet for the given QP.
 * Called at interrupt level.
 */
static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
				     struct ipath_other_headers *ohdr,
				     void *data, u32 tlen,
				     struct ipath_qp *qp,
				     u32 opcode,
				     u32 psn, u32 hdrsize, u32 pmtu,
				     int header_in_data)
{
1103
	struct ipath_swqe *wqe;
1104 1105 1106 1107 1108
	unsigned long flags;
	struct ib_wc wc;
	int diff;
	u32 pad;
	u32 aeth;
1109
	u64 val;
1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133

	spin_lock_irqsave(&qp->s_lock, flags);

	/* Ignore invalid responses. */
	if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
		goto ack_done;

	/* Ignore duplicate responses. */
	diff = ipath_cmp24(psn, qp->s_last_psn);
	if (unlikely(diff <= 0)) {
		/* Update credits for "ghost" ACKs */
		if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
			if (!header_in_data)
				aeth = be32_to_cpu(ohdr->u.aeth);
			else {
				aeth = be32_to_cpu(((__be32 *) data)[0]);
				data += sizeof(__be32);
			}
			if ((aeth >> 29) == 0)
				ipath_get_credit(qp, aeth);
		}
		goto ack_done;
	}

1134 1135 1136 1137
	if (unlikely(qp->s_last == qp->s_tail))
		goto ack_done;
	wqe = get_swqe_ptr(qp, qp->s_last);

1138 1139 1140 1141 1142 1143 1144 1145 1146 1147
	switch (opcode) {
	case OP(ACKNOWLEDGE):
	case OP(ATOMIC_ACKNOWLEDGE):
	case OP(RDMA_READ_RESPONSE_FIRST):
		if (!header_in_data)
			aeth = be32_to_cpu(ohdr->u.aeth);
		else {
			aeth = be32_to_cpu(((__be32 *) data)[0]);
			data += sizeof(__be32);
		}
1148 1149 1150 1151 1152 1153 1154 1155
		if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
			if (!header_in_data) {
				__be32 *p = ohdr->u.at.atomic_ack_eth;

				val = ((u64) be32_to_cpu(p[0]) << 32) |
					be32_to_cpu(p[1]);
			} else
				val = be64_to_cpu(((__be64 *) data)[0]);
1156 1157 1158
		} else
			val = 0;
		if (!do_rc_ack(qp, aeth, psn, opcode, val) ||
1159 1160 1161
		    opcode != OP(RDMA_READ_RESPONSE_FIRST))
			goto ack_done;
		hdrsize += 4;
1162
		wqe = get_swqe_ptr(qp, qp->s_last);
1163
		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1164
			goto ack_op_err;
1165
		/*
1166 1167 1168
		 * If this is a response to a resent RDMA read, we
		 * have to be careful to copy the data to the right
		 * location.
1169
		 */
1170 1171 1172
		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
						  wqe, psn, pmtu);
		goto read_middle;
1173 1174 1175 1176 1177

	case OP(RDMA_READ_RESPONSE_MIDDLE):
		/* no AETH, no ACK */
		if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
			dev->n_rdma_seq++;
1178
			ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1179 1180
			goto ack_done;
		}
1181
		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1182
			goto ack_op_err;
1183
	read_middle:
1184
		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1185
			goto ack_len_err;
1186
		if (unlikely(pmtu >= qp->s_rdma_read_len))
1187
			goto ack_len_err;
1188

1189 1190 1191 1192 1193 1194 1195
		/* We got a response so update the timeout. */
		spin_lock(&dev->pending_lock);
		if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
			list_move_tail(&qp->timerwait,
				       &dev->pending[dev->pending_index]);
		spin_unlock(&dev->pending_lock);
		/*
1196 1197
		 * Update the RDMA receive state but do the copy w/o
		 * holding the locks and blocking interrupts.
1198
		 */
1199
		qp->s_rdma_read_len -= pmtu;
1200
		update_last_psn(qp, psn);
1201
		spin_unlock_irqrestore(&qp->s_lock, flags);
1202
		ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1203
		goto bail;
1204

1205
	case OP(RDMA_READ_RESPONSE_ONLY):
1206 1207 1208 1209 1210
		if (!header_in_data)
			aeth = be32_to_cpu(ohdr->u.aeth);
		else
			aeth = be32_to_cpu(((__be32 *) data)[0]);
		if (!do_rc_ack(qp, aeth, psn, opcode, 0))
1211
			goto ack_done;
1212 1213 1214 1215 1216 1217 1218 1219 1220
		/* Get the number of bytes the message was padded by. */
		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
		/*
		 * Check that the data size is >= 0 && <= pmtu.
		 * Remember to account for the AETH header (4) and
		 * ICRC (4).
		 */
		if (unlikely(tlen < (hdrsize + pad + 8)))
			goto ack_len_err;
1221 1222 1223 1224 1225
		/*
		 * If this is a response to a resent RDMA read, we
		 * have to be careful to copy the data to the right
		 * location.
		 */
1226
		wqe = get_swqe_ptr(qp, qp->s_last);
1227 1228 1229 1230
		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
						  wqe, psn, pmtu);
		goto read_last;

1231 1232 1233 1234
	case OP(RDMA_READ_RESPONSE_LAST):
		/* ACKs READ req. */
		if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
			dev->n_rdma_seq++;
1235
			ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1236 1237
			goto ack_done;
		}
1238
		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1239 1240
			goto ack_op_err;
		/* Get the number of bytes the message was padded by. */
1241 1242 1243 1244 1245 1246
		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
		/*
		 * Check that the data size is >= 1 && <= pmtu.
		 * Remember to account for the AETH header (4) and
		 * ICRC (4).
		 */
1247 1248 1249
		if (unlikely(tlen <= (hdrsize + pad + 8)))
			goto ack_len_err;
	read_last:
1250
		tlen -= hdrsize + pad + 8;
1251 1252
		if (unlikely(tlen != qp->s_rdma_read_len))
			goto ack_len_err;
1253 1254 1255 1256 1257 1258
		if (!header_in_data)
			aeth = be32_to_cpu(ohdr->u.aeth);
		else {
			aeth = be32_to_cpu(((__be32 *) data)[0]);
			data += sizeof(__be32);
		}
1259
		ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1260 1261
		(void) do_rc_ack(qp, aeth, psn,
				 OP(RDMA_READ_RESPONSE_LAST), 0);
1262 1263 1264 1265 1266
		goto ack_done;
	}

ack_done:
	spin_unlock_irqrestore(&qp->s_lock, flags);
1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289
	goto bail;

ack_op_err:
	wc.status = IB_WC_LOC_QP_OP_ERR;
	goto ack_err;

ack_len_err:
	wc.status = IB_WC_LOC_LEN_ERR;
ack_err:
	wc.wr_id = wqe->wr.wr_id;
	wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
	wc.vendor_err = 0;
	wc.byte_len = 0;
	wc.imm_data = 0;
	wc.qp = &qp->ibqp;
	wc.src_qp = qp->remote_qpn;
	wc.wc_flags = 0;
	wc.pkey_index = 0;
	wc.slid = qp->remote_ah_attr.dlid;
	wc.sl = qp->remote_ah_attr.sl;
	wc.dlid_path_bits = 0;
	wc.port_num = 0;
	ipath_sqerror_qp(qp, &wc);
1290
	spin_unlock_irqrestore(&qp->s_lock, flags);
1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309
bail:
	return;
}

/**
 * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
 * @dev: the device this packet came in on
 * @ohdr: the other headers for this packet
 * @data: the packet data
 * @qp: the QP for this packet
 * @opcode: the opcode for this packet
 * @psn: the packet sequence number for this packet
 * @diff: the difference between the PSN and the expected PSN
 * @header_in_data: true if part of the header data is in the data buffer
 *
 * This is called from ipath_rc_rcv() to process an unexpected
 * incoming RC packet for the given QP.
 * Called at interrupt level.
 * Return 1 if no more processing is needed; otherwise return 0 to
1310
 * schedule a response to be sent.
1311 1312 1313 1314 1315 1316 1317 1318 1319 1320
 */
static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
				     struct ipath_other_headers *ohdr,
				     void *data,
				     struct ipath_qp *qp,
				     u32 opcode,
				     u32 psn,
				     int diff,
				     int header_in_data)
{
1321 1322 1323
	struct ipath_ack_entry *e;
	u8 i, prev;
	int old_req;
1324
	unsigned long flags;
1325 1326 1327 1328 1329

	if (diff > 0) {
		/*
		 * Packet sequence error.
		 * A NAK will ACK earlier sends and RDMA writes.
1330
		 * Don't queue the NAK if we already sent one.
1331
		 */
1332
		if (!qp->r_nak_state) {
1333 1334 1335
			qp->r_nak_state = IB_NAK_PSN_ERROR;
			/* Use the expected PSN. */
			qp->r_ack_psn = qp->r_psn;
1336
			goto send_ack;
1337
		}
1338
		goto done;
1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350
	}

	/*
	 * Handle a duplicate request.  Don't re-execute SEND, RDMA
	 * write or atomic op.  Don't NAK errors, just silently drop
	 * the duplicate request.  Note that r_sge, r_len, and
	 * r_rcv_len may be in use so don't modify them.
	 *
	 * We are supposed to ACK the earliest duplicate PSN but we
	 * can coalesce an outstanding duplicate ACK.  We have to
	 * send the earliest so that RDMA reads can be restarted at
	 * the requester's expected PSN.
1351 1352 1353
	 *
	 * First, find where this duplicate PSN falls within the
	 * ACKs previously sent.
1354
	 */
1355 1356 1357
	psn &= IPATH_PSN_MASK;
	e = NULL;
	old_req = 1;
1358
	spin_lock_irqsave(&qp->s_lock, flags);
1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374
	for (i = qp->r_head_ack_queue; ; i = prev) {
		if (i == qp->s_tail_ack_queue)
			old_req = 0;
		if (i)
			prev = i - 1;
		else
			prev = IPATH_MAX_RDMA_ATOMIC;
		if (prev == qp->r_head_ack_queue) {
			e = NULL;
			break;
		}
		e = &qp->s_ack_queue[prev];
		if (!e->opcode) {
			e = NULL;
			break;
		}
1375 1376 1377
		if (ipath_cmp24(psn, e->psn) >= 0) {
			if (prev == qp->s_tail_ack_queue)
				old_req = 0;
1378
			break;
1379
		}
1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393
	}
	switch (opcode) {
	case OP(RDMA_READ_REQUEST): {
		struct ib_reth *reth;
		u32 offset;
		u32 len;

		/*
		 * If we didn't find the RDMA read request in the ack queue,
		 * or the send tasklet is already backed up to send an
		 * earlier entry, we can ignore this request.
		 */
		if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
			goto unlock_done;
1394 1395 1396 1397 1398 1399 1400
		/* RETH comes after BTH */
		if (!header_in_data)
			reth = &ohdr->u.rc.reth;
		else {
			reth = (struct ib_reth *)data;
			data += sizeof(*reth);
		}
1401
		/*
1402 1403 1404 1405 1406
		 * Address range must be a subset of the original
		 * request and start on pmtu boundaries.
		 * We reuse the old ack_queue slot since the requester
		 * should not back up and request an earlier PSN for the
		 * same request.
1407
		 */
1408 1409 1410 1411 1412 1413
		offset = ((psn - e->psn) & IPATH_PSN_MASK) *
			ib_mtu_enum_to_int(qp->path_mtu);
		len = be32_to_cpu(reth->length);
		if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
			goto unlock_done;
		if (len != 0) {
1414 1415 1416 1417
			u32 rkey = be32_to_cpu(reth->rkey);
			u64 vaddr = be64_to_cpu(reth->vaddr);
			int ok;

1418 1419
			ok = ipath_rkey_ok(qp, &e->rdma_sge,
					   len, vaddr, rkey,
1420
					   IB_ACCESS_REMOTE_READ);
1421 1422
			if (unlikely(!ok))
				goto unlock_done;
1423
		} else {
1424 1425 1426 1427 1428 1429
			e->rdma_sge.sg_list = NULL;
			e->rdma_sge.num_sge = 0;
			e->rdma_sge.sge.mr = NULL;
			e->rdma_sge.sge.vaddr = NULL;
			e->rdma_sge.sge.length = 0;
			e->rdma_sge.sge.sge_length = 0;
1430
		}
1431 1432 1433 1434
		e->psn = psn;
		qp->s_ack_state = OP(ACKNOWLEDGE);
		qp->s_tail_ack_queue = prev;
		break;
1435 1436
	}

1437
	case OP(COMPARE_SWAP):
1438
	case OP(FETCH_ADD): {
1439
		/*
1440 1441 1442
		 * If we didn't find the atomic request in the ack queue
		 * or the send tasklet is already backed up to send an
		 * earlier entry, we can ignore this request.
1443
		 */
1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458
		if (!e || e->opcode != (u8) opcode || old_req)
			goto unlock_done;
		qp->s_ack_state = OP(ACKNOWLEDGE);
		qp->s_tail_ack_queue = prev;
		break;
	}

	default:
		if (old_req)
			goto unlock_done;
		/*
		 * Resend the most recent ACK if this request is
		 * after all the previous RDMA reads and atomics.
		 */
		if (i == qp->r_head_ack_queue) {
1459
			spin_unlock_irqrestore(&qp->s_lock, flags);
1460 1461 1462 1463
			qp->r_nak_state = 0;
			qp->r_ack_psn = qp->r_psn - 1;
			goto send_ack;
		}
1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476
		/*
		 * Try to send a simple ACK to work around a Mellanox bug
		 * which doesn't accept a RDMA read response or atomic
		 * response as an ACK for earlier SENDs or RDMA writes.
		 */
		if (qp->r_head_ack_queue == qp->s_tail_ack_queue &&
		    !(qp->s_flags & IPATH_S_ACK_PENDING) &&
		    qp->s_ack_state == OP(ACKNOWLEDGE)) {
			spin_unlock_irqrestore(&qp->s_lock, flags);
			qp->r_nak_state = 0;
			qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
			goto send_ack;
		}
1477 1478 1479 1480 1481 1482
		/*
		 * Resend the RDMA read or atomic op which
		 * ACKs this duplicate request.
		 */
		qp->s_ack_state = OP(ACKNOWLEDGE);
		qp->s_tail_ack_queue = i;
1483 1484
		break;
	}
1485
	qp->r_nak_state = 0;
1486
	tasklet_hi_schedule(&qp->s_task);
1487

1488
unlock_done:
1489
	spin_unlock_irqrestore(&qp->s_lock, flags);
1490 1491
done:
	return 1;
1492 1493 1494

send_ack:
	return 0;
1495 1496
}

1497 1498
static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
{
1499 1500 1501
	unsigned long flags;

	spin_lock_irqsave(&qp->s_lock, flags);
1502 1503
	qp->state = IB_QPS_ERR;
	ipath_error_qp(qp, err);
1504
	spin_unlock_irqrestore(&qp->s_lock, flags);
1505 1506
}

1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522
static inline void ipath_update_ack_queue(struct ipath_qp *qp, unsigned n)
{
	unsigned long flags;
	unsigned next;

	next = n + 1;
	if (next > IPATH_MAX_RDMA_ATOMIC)
		next = 0;
	spin_lock_irqsave(&qp->s_lock, flags);
	if (n == qp->s_tail_ack_queue) {
		qp->s_tail_ack_queue = next;
		qp->s_ack_state = OP(ACKNOWLEDGE);
	}
	spin_unlock_irqrestore(&qp->s_lock, flags);
}

1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549
/**
 * ipath_rc_rcv - process an incoming RC packet
 * @dev: the device this packet came in on
 * @hdr: the header of this packet
 * @has_grh: true if the header has a GRH
 * @data: the packet data
 * @tlen: the packet length
 * @qp: the QP for this packet
 *
 * This is called from ipath_qp_rcv() to process an incoming RC packet
 * for the given QP.
 * Called at interrupt level.
 */
void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
		  int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
{
	struct ipath_other_headers *ohdr;
	u32 opcode;
	u32 hdrsize;
	u32 psn;
	u32 pad;
	struct ib_wc wc;
	u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
	int diff;
	struct ib_reth *reth;
	int header_in_data;

1550 1551 1552 1553
	/* Validate the SLID. See Ch. 9.6.1.5 */
	if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
		goto done;

1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567
	/* Check for GRH */
	if (!has_grh) {
		ohdr = &hdr->u.oth;
		hdrsize = 8 + 12;	/* LRH + BTH */
		psn = be32_to_cpu(ohdr->bth[2]);
		header_in_data = 0;
	} else {
		ohdr = &hdr->u.l.oth;
		hdrsize = 8 + 40 + 12;	/* LRH + GRH + BTH */
		/*
		 * The header with GRH is 60 bytes and the core driver sets
		 * the eager header buffer size to 56 bytes so the last 4
		 * bytes of the BTH header (PSN) is in the data buffer.
		 */
1568
		header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581
		if (header_in_data) {
			psn = be32_to_cpu(((__be32 *) data)[0]);
			data += sizeof(__be32);
		} else
			psn = be32_to_cpu(ohdr->bth[2]);
	}

	/*
	 * Process responses (ACKs) before anything else.  Note that the
	 * packet sequence number will be for something in the send work
	 * queue rather than the expected receive packet sequence number.
	 * In other words, this QP is the requester.
	 */
1582
	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1583 1584 1585 1586
	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
		ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
				  hdrsize, pmtu, header_in_data);
1587
		goto done;
1588 1589 1590 1591 1592 1593 1594 1595
	}

	/* Compute 24 bits worth of difference. */
	diff = ipath_cmp24(psn, qp->r_psn);
	if (unlikely(diff)) {
		if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
				       psn, diff, header_in_data))
			goto done;
1596
		goto send_ack;
1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607
	}

	/* Check for opcode sequence errors. */
	switch (qp->r_state) {
	case OP(SEND_FIRST):
	case OP(SEND_MIDDLE):
		if (opcode == OP(SEND_MIDDLE) ||
		    opcode == OP(SEND_LAST) ||
		    opcode == OP(SEND_LAST_WITH_IMMEDIATE))
			break;
	nack_inv:
1608
		ipath_rc_error(qp, IB_WC_REM_INV_REQ_ERR);
1609 1610 1611
		qp->r_nak_state = IB_NAK_INVALID_REQUEST;
		qp->r_ack_psn = qp->r_psn;
		goto send_ack;
1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628

	case OP(RDMA_WRITE_FIRST):
	case OP(RDMA_WRITE_MIDDLE):
		if (opcode == OP(RDMA_WRITE_MIDDLE) ||
		    opcode == OP(RDMA_WRITE_LAST) ||
		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
			break;
		goto nack_inv;

	default:
		if (opcode == OP(SEND_MIDDLE) ||
		    opcode == OP(SEND_LAST) ||
		    opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
		    opcode == OP(RDMA_WRITE_MIDDLE) ||
		    opcode == OP(RDMA_WRITE_LAST) ||
		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
			goto nack_inv;
1629 1630 1631 1632 1633
		/*
		 * Note that it is up to the requester to not send a new
		 * RDMA read or atomic operation before receiving an ACK
		 * for the previous operation.
		 */
1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649
		break;
	}

	wc.imm_data = 0;
	wc.wc_flags = 0;

	/* OK, process the packet. */
	switch (opcode) {
	case OP(SEND_FIRST):
		if (!ipath_get_rwqe(qp, 0)) {
		rnr_nak:
			/*
			 * A RNR NAK will ACK earlier sends and RDMA writes.
			 * Don't queue the NAK if a RDMA read or atomic
			 * is pending though.
			 */
1650 1651
			if (qp->r_nak_state)
				goto done;
1652 1653 1654
			qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
			qp->r_ack_psn = qp->r_psn;
			goto send_ack;
1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710
		}
		qp->r_rcv_len = 0;
		/* FALLTHROUGH */
	case OP(SEND_MIDDLE):
	case OP(RDMA_WRITE_MIDDLE):
	send_middle:
		/* Check for invalid length PMTU or posted rwqe len. */
		if (unlikely(tlen != (hdrsize + pmtu + 4)))
			goto nack_inv;
		qp->r_rcv_len += pmtu;
		if (unlikely(qp->r_rcv_len > qp->r_len))
			goto nack_inv;
		ipath_copy_sge(&qp->r_sge, data, pmtu);
		break;

	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
		/* consume RWQE */
		if (!ipath_get_rwqe(qp, 1))
			goto rnr_nak;
		goto send_last_imm;

	case OP(SEND_ONLY):
	case OP(SEND_ONLY_WITH_IMMEDIATE):
		if (!ipath_get_rwqe(qp, 0))
			goto rnr_nak;
		qp->r_rcv_len = 0;
		if (opcode == OP(SEND_ONLY))
			goto send_last;
		/* FALLTHROUGH */
	case OP(SEND_LAST_WITH_IMMEDIATE):
	send_last_imm:
		if (header_in_data) {
			wc.imm_data = *(__be32 *) data;
			data += sizeof(__be32);
		} else {
			/* Immediate data comes after BTH */
			wc.imm_data = ohdr->u.imm_data;
		}
		hdrsize += 4;
		wc.wc_flags = IB_WC_WITH_IMM;
		/* FALLTHROUGH */
	case OP(SEND_LAST):
	case OP(RDMA_WRITE_LAST):
	send_last:
		/* Get the number of bytes the message was padded by. */
		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
		/* Check for invalid length. */
		/* XXX LAST len should be >= 1 */
		if (unlikely(tlen < (hdrsize + pad + 4)))
			goto nack_inv;
		/* Don't count the CRC. */
		tlen -= (hdrsize + pad + 4);
		wc.byte_len = tlen + qp->r_rcv_len;
		if (unlikely(wc.byte_len > qp->r_len))
			goto nack_inv;
		ipath_copy_sge(&qp->r_sge, data, tlen);
1711
		qp->r_msn++;
1712
		if (!qp->r_wrid_valid)
1713
			break;
1714
		qp->r_wrid_valid = 0;
1715 1716 1717 1718
		wc.wr_id = qp->r_wr_id;
		wc.status = IB_WC_SUCCESS;
		wc.opcode = IB_WC_RECV;
		wc.vendor_err = 0;
1719
		wc.qp = &qp->ibqp;
1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734
		wc.src_qp = qp->remote_qpn;
		wc.pkey_index = 0;
		wc.slid = qp->remote_ah_attr.dlid;
		wc.sl = qp->remote_ah_attr.sl;
		wc.dlid_path_bits = 0;
		wc.port_num = 0;
		/* Signal completion event if the solicited bit is set. */
		ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
			       (ohdr->bth[0] &
				__constant_cpu_to_be32(1 << 23)) != 0);
		break;

	case OP(RDMA_WRITE_FIRST):
	case OP(RDMA_WRITE_ONLY):
	case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1735 1736 1737
		if (unlikely(!(qp->qp_access_flags &
			       IB_ACCESS_REMOTE_WRITE)))
			goto nack_inv;
1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754
		/* consume RWQE */
		/* RETH comes after BTH */
		if (!header_in_data)
			reth = &ohdr->u.rc.reth;
		else {
			reth = (struct ib_reth *)data;
			data += sizeof(*reth);
		}
		hdrsize += sizeof(*reth);
		qp->r_len = be32_to_cpu(reth->length);
		qp->r_rcv_len = 0;
		if (qp->r_len != 0) {
			u32 rkey = be32_to_cpu(reth->rkey);
			u64 vaddr = be64_to_cpu(reth->vaddr);
			int ok;

			/* Check rkey & NAK */
1755
			ok = ipath_rkey_ok(qp, &qp->r_sge,
1756 1757
					   qp->r_len, vaddr, rkey,
					   IB_ACCESS_REMOTE_WRITE);
1758 1759
			if (unlikely(!ok))
				goto nack_acc;
1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774
		} else {
			qp->r_sge.sg_list = NULL;
			qp->r_sge.sge.mr = NULL;
			qp->r_sge.sge.vaddr = NULL;
			qp->r_sge.sge.length = 0;
			qp->r_sge.sge.sge_length = 0;
		}
		if (opcode == OP(RDMA_WRITE_FIRST))
			goto send_middle;
		else if (opcode == OP(RDMA_WRITE_ONLY))
			goto send_last;
		if (!ipath_get_rwqe(qp, 1))
			goto rnr_nak;
		goto send_last_imm;

1775 1776 1777 1778 1779
	case OP(RDMA_READ_REQUEST): {
		struct ipath_ack_entry *e;
		u32 len;
		u8 next;

1780 1781 1782
		if (unlikely(!(qp->qp_access_flags &
			       IB_ACCESS_REMOTE_READ)))
			goto nack_inv;
1783 1784 1785
		next = qp->r_head_ack_queue + 1;
		if (next > IPATH_MAX_RDMA_ATOMIC)
			next = 0;
1786 1787 1788 1789 1790
		if (unlikely(next == qp->s_tail_ack_queue)) {
			if (!qp->s_ack_queue[next].sent)
				goto nack_inv;
			ipath_update_ack_queue(qp, next);
		}
1791
		e = &qp->s_ack_queue[qp->r_head_ack_queue];
1792 1793 1794 1795 1796 1797 1798
		/* RETH comes after BTH */
		if (!header_in_data)
			reth = &ohdr->u.rc.reth;
		else {
			reth = (struct ib_reth *)data;
			data += sizeof(*reth);
		}
1799 1800
		len = be32_to_cpu(reth->length);
		if (len) {
1801 1802 1803 1804 1805
			u32 rkey = be32_to_cpu(reth->rkey);
			u64 vaddr = be64_to_cpu(reth->vaddr);
			int ok;

			/* Check rkey & NAK */
1806 1807 1808
			ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
					   rkey, IB_ACCESS_REMOTE_READ);
			if (unlikely(!ok))
1809 1810 1811 1812 1813
				goto nack_acc;
			/*
			 * Update the next expected PSN.  We add 1 later
			 * below, so only add the remainder here.
			 */
1814 1815
			if (len > pmtu)
				qp->r_psn += (len - 1) / pmtu;
1816
		} else {
1817 1818 1819 1820 1821 1822
			e->rdma_sge.sg_list = NULL;
			e->rdma_sge.num_sge = 0;
			e->rdma_sge.sge.mr = NULL;
			e->rdma_sge.sge.vaddr = NULL;
			e->rdma_sge.sge.length = 0;
			e->rdma_sge.sge.sge_length = 0;
1823
		}
1824
		e->opcode = opcode;
1825
		e->sent = 0;
1826
		e->psn = psn;
1827 1828 1829 1830 1831
		/*
		 * We need to increment the MSN here instead of when we
		 * finish sending the result since a duplicate request would
		 * increment it more than once.
		 */
1832
		qp->r_msn++;
1833 1834
		qp->r_psn++;
		qp->r_state = opcode;
1835
		qp->r_nak_state = 0;
1836 1837
		barrier();
		qp->r_head_ack_queue = next;
1838 1839 1840 1841 1842

		/* Call ipath_do_rc_send() in another thread. */
		tasklet_hi_schedule(&qp->s_task);

		goto done;
1843
	}
1844 1845 1846 1847

	case OP(COMPARE_SWAP):
	case OP(FETCH_ADD): {
		struct ib_atomic_eth *ateth;
1848
		struct ipath_ack_entry *e;
1849
		u64 vaddr;
1850
		atomic64_t *maddr;
1851 1852
		u64 sdata;
		u32 rkey;
1853
		u8 next;
1854

1855 1856
		if (unlikely(!(qp->qp_access_flags &
			       IB_ACCESS_REMOTE_ATOMIC)))
1857
			goto nack_inv;
1858 1859 1860
		next = qp->r_head_ack_queue + 1;
		if (next > IPATH_MAX_RDMA_ATOMIC)
			next = 0;
1861 1862 1863 1864 1865
		if (unlikely(next == qp->s_tail_ack_queue)) {
			if (!qp->s_ack_queue[next].sent)
				goto nack_inv;
			ipath_update_ack_queue(qp, next);
		}
1866 1867
		if (!header_in_data)
			ateth = &ohdr->u.atomic_eth;
1868
		else
1869
			ateth = (struct ib_atomic_eth *)data;
1870 1871
		vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
			be32_to_cpu(ateth->vaddr[1]);
1872 1873 1874 1875
		if (unlikely(vaddr & (sizeof(u64) - 1)))
			goto nack_inv;
		rkey = be32_to_cpu(ateth->rkey);
		/* Check rkey & NAK */
1876
		if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
1877 1878 1879 1880
					    sizeof(u64), vaddr, rkey,
					    IB_ACCESS_REMOTE_ATOMIC)))
			goto nack_acc;
		/* Perform atomic OP and save result. */
1881
		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1882
		sdata = be64_to_cpu(ateth->swap_data);
1883 1884 1885 1886 1887 1888 1889
		e = &qp->s_ack_queue[qp->r_head_ack_queue];
		e->atomic_data = (opcode == OP(FETCH_ADD)) ?
			(u64) atomic64_add_return(sdata, maddr) - sdata :
			(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
				      be64_to_cpu(ateth->compare_data),
				      sdata);
		e->opcode = opcode;
1890
		e->sent = 0;
1891
		e->psn = psn & IPATH_PSN_MASK;
1892
		qp->r_msn++;
1893 1894 1895 1896 1897 1898 1899 1900 1901 1902
		qp->r_psn++;
		qp->r_state = opcode;
		qp->r_nak_state = 0;
		barrier();
		qp->r_head_ack_queue = next;

		/* Call ipath_do_rc_send() in another thread. */
		tasklet_hi_schedule(&qp->s_task);

		goto done;
1903 1904 1905
	}

	default:
1906 1907
		/* NAK unknown opcodes. */
		goto nack_inv;
1908 1909 1910
	}
	qp->r_psn++;
	qp->r_state = opcode;
1911
	qp->r_ack_psn = psn;
1912
	qp->r_nak_state = 0;
1913
	/* Send an ACK if requested or required. */
1914
	if (psn & (1 << 31))
1915 1916
		goto send_ack;
	goto done;
1917

1918
nack_acc:
1919 1920 1921 1922
	ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
	qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
	qp->r_ack_psn = qp->r_psn;

1923
send_ack:
1924
	send_rc_ack(qp);
1925

1926
done:
1927 1928
	return;
}