aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/infiniband/hw/ipath/ipath_rc.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/infiniband/hw/ipath/ipath_rc.c')
-rw-r--r--drivers/infiniband/hw/ipath/ipath_rc.c920
1 files changed, 533 insertions, 387 deletions
diff --git a/drivers/infiniband/hw/ipath/ipath_rc.c b/drivers/infiniband/hw/ipath/ipath_rc.c
index 5ff20cb04494..b4b88d0b53f5 100644
--- a/drivers/infiniband/hw/ipath/ipath_rc.c
+++ b/drivers/infiniband/hw/ipath/ipath_rc.c
@@ -37,6 +37,19 @@
37/* cut down ridiculously long IB macro names */ 37/* cut down ridiculously long IB macro names */
38#define OP(x) IB_OPCODE_RC_##x 38#define OP(x) IB_OPCODE_RC_##x
39 39
40static u32 restart_sge(struct ipath_sge_state *ss, struct ipath_swqe *wqe,
41 u32 psn, u32 pmtu)
42{
43 u32 len;
44
45 len = ((psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
46 ss->sge = wqe->sg_list[0];
47 ss->sg_list = wqe->sg_list + 1;
48 ss->num_sge = wqe->wr.num_sge;
49 ipath_skip_sge(ss, len);
50 return wqe->length - len;
51}
52
40/** 53/**
41 * ipath_init_restart- initialize the qp->s_sge after a restart 54 * ipath_init_restart- initialize the qp->s_sge after a restart
42 * @qp: the QP who's SGE we're restarting 55 * @qp: the QP who's SGE we're restarting
@@ -47,15 +60,9 @@
47static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe) 60static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
48{ 61{
49 struct ipath_ibdev *dev; 62 struct ipath_ibdev *dev;
50 u32 len;
51 63
52 len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * 64 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn,
53 ib_mtu_enum_to_int(qp->path_mtu); 65 ib_mtu_enum_to_int(qp->path_mtu));
54 qp->s_sge.sge = wqe->sg_list[0];
55 qp->s_sge.sg_list = wqe->sg_list + 1;
56 qp->s_sge.num_sge = wqe->wr.num_sge;
57 ipath_skip_sge(&qp->s_sge, len);
58 qp->s_len = wqe->length - len;
59 dev = to_idev(qp->ibqp.device); 66 dev = to_idev(qp->ibqp.device);
60 spin_lock(&dev->pending_lock); 67 spin_lock(&dev->pending_lock);
61 if (list_empty(&qp->timerwait)) 68 if (list_empty(&qp->timerwait))
@@ -70,107 +77,123 @@ static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
70 * @ohdr: a pointer to the IB header being constructed 77 * @ohdr: a pointer to the IB header being constructed
71 * @pmtu: the path MTU 78 * @pmtu: the path MTU
72 * 79 *
73 * Return bth0 if constructed; otherwise, return 0. 80 * Return 1 if constructed; otherwise, return 0.
81 * Note that we are in the responder's side of the QP context.
74 * Note the QP s_lock must be held. 82 * Note the QP s_lock must be held.
75 */ 83 */
76u32 ipath_make_rc_ack(struct ipath_qp *qp, 84static int ipath_make_rc_ack(struct ipath_qp *qp,
77 struct ipath_other_headers *ohdr, 85 struct ipath_other_headers *ohdr,
78 u32 pmtu) 86 u32 pmtu, u32 *bth0p, u32 *bth2p)
79{ 87{
88 struct ipath_ack_entry *e;
80 u32 hwords; 89 u32 hwords;
81 u32 len; 90 u32 len;
82 u32 bth0; 91 u32 bth0;
92 u32 bth2;
83 93
84 /* header size in 32-bit words LRH+BTH = (8+12)/4. */ 94 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
85 hwords = 5; 95 hwords = 5;
86 96
87 /*
88 * Send a response. Note that we are in the responder's
89 * side of the QP context.
90 */
91 switch (qp->s_ack_state) { 97 switch (qp->s_ack_state) {
92 case OP(RDMA_READ_REQUEST): 98 case OP(RDMA_READ_RESPONSE_LAST):
93 qp->s_cur_sge = &qp->s_rdma_sge; 99 case OP(RDMA_READ_RESPONSE_ONLY):
94 len = qp->s_rdma_len; 100 case OP(ATOMIC_ACKNOWLEDGE):
95 if (len > pmtu) { 101 qp->s_ack_state = OP(ACKNOWLEDGE);
96 len = pmtu; 102 /* FALLTHROUGH */
97 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST); 103 case OP(ACKNOWLEDGE):
98 } else 104 /* Check for no next entry in the queue. */
99 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY); 105 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
100 qp->s_rdma_len -= len; 106 if (qp->s_flags & IPATH_S_ACK_PENDING)
107 goto normal;
108 goto bail;
109 }
110
111 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
112 if (e->opcode == OP(RDMA_READ_REQUEST)) {
113 /* Copy SGE state in case we need to resend */
114 qp->s_ack_rdma_sge = e->rdma_sge;
115 qp->s_cur_sge = &qp->s_ack_rdma_sge;
116 len = e->rdma_sge.sge.sge_length;
117 if (len > pmtu) {
118 len = pmtu;
119 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
120 } else {
121 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
122 if (++qp->s_tail_ack_queue >
123 IPATH_MAX_RDMA_ATOMIC)
124 qp->s_tail_ack_queue = 0;
125 }
126 ohdr->u.aeth = ipath_compute_aeth(qp);
127 hwords++;
128 qp->s_ack_rdma_psn = e->psn;
129 bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
130 } else {
131 /* COMPARE_SWAP or FETCH_ADD */
132 qp->s_cur_sge = NULL;
133 len = 0;
134 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
135 ohdr->u.at.aeth = ipath_compute_aeth(qp);
136 ohdr->u.at.atomic_ack_eth[0] =
137 cpu_to_be32(e->atomic_data >> 32);
138 ohdr->u.at.atomic_ack_eth[1] =
139 cpu_to_be32(e->atomic_data);
140 hwords += sizeof(ohdr->u.at) / sizeof(u32);
141 bth2 = e->psn;
142 if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
143 qp->s_tail_ack_queue = 0;
144 }
101 bth0 = qp->s_ack_state << 24; 145 bth0 = qp->s_ack_state << 24;
102 ohdr->u.aeth = ipath_compute_aeth(qp);
103 hwords++;
104 break; 146 break;
105 147
106 case OP(RDMA_READ_RESPONSE_FIRST): 148 case OP(RDMA_READ_RESPONSE_FIRST):
107 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE); 149 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
108 /* FALLTHROUGH */ 150 /* FALLTHROUGH */
109 case OP(RDMA_READ_RESPONSE_MIDDLE): 151 case OP(RDMA_READ_RESPONSE_MIDDLE):
110 qp->s_cur_sge = &qp->s_rdma_sge; 152 len = qp->s_ack_rdma_sge.sge.sge_length;
111 len = qp->s_rdma_len;
112 if (len > pmtu) 153 if (len > pmtu)
113 len = pmtu; 154 len = pmtu;
114 else { 155 else {
115 ohdr->u.aeth = ipath_compute_aeth(qp); 156 ohdr->u.aeth = ipath_compute_aeth(qp);
116 hwords++; 157 hwords++;
117 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST); 158 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
159 if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
160 qp->s_tail_ack_queue = 0;
118 } 161 }
119 qp->s_rdma_len -= len;
120 bth0 = qp->s_ack_state << 24; 162 bth0 = qp->s_ack_state << 24;
121 break; 163 bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
122
123 case OP(RDMA_READ_RESPONSE_LAST):
124 case OP(RDMA_READ_RESPONSE_ONLY):
125 /*
126 * We have to prevent new requests from changing
127 * the r_sge state while a ipath_verbs_send()
128 * is in progress.
129 */
130 qp->s_ack_state = OP(ACKNOWLEDGE);
131 bth0 = 0;
132 goto bail;
133
134 case OP(COMPARE_SWAP):
135 case OP(FETCH_ADD):
136 qp->s_cur_sge = NULL;
137 len = 0;
138 /*
139 * Set the s_ack_state so the receive interrupt handler
140 * won't try to send an ACK (out of order) until this one
141 * is actually sent.
142 */
143 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
144 bth0 = OP(ATOMIC_ACKNOWLEDGE) << 24;
145 ohdr->u.at.aeth = ipath_compute_aeth(qp);
146 ohdr->u.at.atomic_ack_eth = cpu_to_be64(qp->r_atomic_data);
147 hwords += sizeof(ohdr->u.at) / 4;
148 break; 164 break;
149 165
150 default: 166 default:
151 /* Send a regular ACK. */ 167 normal:
152 qp->s_cur_sge = NULL;
153 len = 0;
154 /* 168 /*
155 * Set the s_ack_state so the receive interrupt handler 169 * Send a regular ACK.
156 * won't try to send an ACK (out of order) until this one 170 * Set the s_ack_state so we wait until after sending
157 * is actually sent. 171 * the ACK before setting s_ack_state to ACKNOWLEDGE
172 * (see above).
158 */ 173 */
159 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST); 174 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
160 bth0 = OP(ACKNOWLEDGE) << 24; 175 qp->s_flags &= ~IPATH_S_ACK_PENDING;
176 qp->s_cur_sge = NULL;
161 if (qp->s_nak_state) 177 if (qp->s_nak_state)
162 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) | 178 ohdr->u.aeth =
163 (qp->s_nak_state << 179 cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
164 IPATH_AETH_CREDIT_SHIFT)); 180 (qp->s_nak_state <<
181 IPATH_AETH_CREDIT_SHIFT));
165 else 182 else
166 ohdr->u.aeth = ipath_compute_aeth(qp); 183 ohdr->u.aeth = ipath_compute_aeth(qp);
167 hwords++; 184 hwords++;
185 len = 0;
186 bth0 = OP(ACKNOWLEDGE) << 24;
187 bth2 = qp->s_ack_psn & IPATH_PSN_MASK;
168 } 188 }
169 qp->s_hdrwords = hwords; 189 qp->s_hdrwords = hwords;
170 qp->s_cur_size = len; 190 qp->s_cur_size = len;
191 *bth0p = bth0;
192 *bth2p = bth2;
193 return 1;
171 194
172bail: 195bail:
173 return bth0; 196 return 0;
174} 197}
175 198
176/** 199/**
@@ -197,9 +220,16 @@ int ipath_make_rc_req(struct ipath_qp *qp,
197 u32 bth2; 220 u32 bth2;
198 char newreq; 221 char newreq;
199 222
223 /* Sending responses has higher priority over sending requests. */
224 if ((qp->r_head_ack_queue != qp->s_tail_ack_queue ||
225 (qp->s_flags & IPATH_S_ACK_PENDING) ||
226 qp->s_ack_state != IB_OPCODE_RC_ACKNOWLEDGE) &&
227 ipath_make_rc_ack(qp, ohdr, pmtu, bth0p, bth2p))
228 goto done;
229
200 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) || 230 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
201 qp->s_rnr_timeout) 231 qp->s_rnr_timeout)
202 goto done; 232 goto bail;
203 233
204 /* Limit the number of packets sent without an ACK. */ 234 /* Limit the number of packets sent without an ACK. */
205 if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT) > 0) { 235 if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT) > 0) {
@@ -210,7 +240,7 @@ int ipath_make_rc_req(struct ipath_qp *qp,
210 list_add_tail(&qp->timerwait, 240 list_add_tail(&qp->timerwait,
211 &dev->pending[dev->pending_index]); 241 &dev->pending[dev->pending_index]);
212 spin_unlock(&dev->pending_lock); 242 spin_unlock(&dev->pending_lock);
213 goto done; 243 goto bail;
214 } 244 }
215 245
216 /* header size in 32-bit words LRH+BTH = (8+12)/4. */ 246 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
@@ -232,7 +262,16 @@ int ipath_make_rc_req(struct ipath_qp *qp,
232 if (qp->s_cur == qp->s_tail) { 262 if (qp->s_cur == qp->s_tail) {
233 /* Check if send work queue is empty. */ 263 /* Check if send work queue is empty. */
234 if (qp->s_tail == qp->s_head) 264 if (qp->s_tail == qp->s_head)
235 goto done; 265 goto bail;
266 /*
267 * If a fence is requested, wait for previous
268 * RDMA read and atomic operations to finish.
269 */
270 if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
271 qp->s_num_rd_atomic) {
272 qp->s_flags |= IPATH_S_FENCE_PENDING;
273 goto bail;
274 }
236 wqe->psn = qp->s_next_psn; 275 wqe->psn = qp->s_next_psn;
237 newreq = 1; 276 newreq = 1;
238 } 277 }
@@ -250,7 +289,7 @@ int ipath_make_rc_req(struct ipath_qp *qp,
250 /* If no credit, return. */ 289 /* If no credit, return. */
251 if (qp->s_lsn != (u32) -1 && 290 if (qp->s_lsn != (u32) -1 &&
252 ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0) 291 ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
253 goto done; 292 goto bail;
254 wqe->lpsn = wqe->psn; 293 wqe->lpsn = wqe->psn;
255 if (len > pmtu) { 294 if (len > pmtu) {
256 wqe->lpsn += (len - 1) / pmtu; 295 wqe->lpsn += (len - 1) / pmtu;
@@ -281,13 +320,13 @@ int ipath_make_rc_req(struct ipath_qp *qp,
281 /* If no credit, return. */ 320 /* If no credit, return. */
282 if (qp->s_lsn != (u32) -1 && 321 if (qp->s_lsn != (u32) -1 &&
283 ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0) 322 ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
284 goto done; 323 goto bail;
285 ohdr->u.rc.reth.vaddr = 324 ohdr->u.rc.reth.vaddr =
286 cpu_to_be64(wqe->wr.wr.rdma.remote_addr); 325 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
287 ohdr->u.rc.reth.rkey = 326 ohdr->u.rc.reth.rkey =
288 cpu_to_be32(wqe->wr.wr.rdma.rkey); 327 cpu_to_be32(wqe->wr.wr.rdma.rkey);
289 ohdr->u.rc.reth.length = cpu_to_be32(len); 328 ohdr->u.rc.reth.length = cpu_to_be32(len);
290 hwords += sizeof(struct ib_reth) / 4; 329 hwords += sizeof(struct ib_reth) / sizeof(u32);
291 wqe->lpsn = wqe->psn; 330 wqe->lpsn = wqe->psn;
292 if (len > pmtu) { 331 if (len > pmtu) {
293 wqe->lpsn += (len - 1) / pmtu; 332 wqe->lpsn += (len - 1) / pmtu;
@@ -312,14 +351,17 @@ int ipath_make_rc_req(struct ipath_qp *qp,
312 break; 351 break;
313 352
314 case IB_WR_RDMA_READ: 353 case IB_WR_RDMA_READ:
315 ohdr->u.rc.reth.vaddr = 354 /*
316 cpu_to_be64(wqe->wr.wr.rdma.remote_addr); 355 * Don't allow more operations to be started
317 ohdr->u.rc.reth.rkey = 356 * than the QP limits allow.
318 cpu_to_be32(wqe->wr.wr.rdma.rkey); 357 */
319 ohdr->u.rc.reth.length = cpu_to_be32(len);
320 qp->s_state = OP(RDMA_READ_REQUEST);
321 hwords += sizeof(ohdr->u.rc.reth) / 4;
322 if (newreq) { 358 if (newreq) {
359 if (qp->s_num_rd_atomic >=
360 qp->s_max_rd_atomic) {
361 qp->s_flags |= IPATH_S_RDMAR_PENDING;
362 goto bail;
363 }
364 qp->s_num_rd_atomic++;
323 if (qp->s_lsn != (u32) -1) 365 if (qp->s_lsn != (u32) -1)
324 qp->s_lsn++; 366 qp->s_lsn++;
325 /* 367 /*
@@ -330,6 +372,13 @@ int ipath_make_rc_req(struct ipath_qp *qp,
330 qp->s_next_psn += (len - 1) / pmtu; 372 qp->s_next_psn += (len - 1) / pmtu;
331 wqe->lpsn = qp->s_next_psn++; 373 wqe->lpsn = qp->s_next_psn++;
332 } 374 }
375 ohdr->u.rc.reth.vaddr =
376 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
377 ohdr->u.rc.reth.rkey =
378 cpu_to_be32(wqe->wr.wr.rdma.rkey);
379 ohdr->u.rc.reth.length = cpu_to_be32(len);
380 qp->s_state = OP(RDMA_READ_REQUEST);
381 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
333 ss = NULL; 382 ss = NULL;
334 len = 0; 383 len = 0;
335 if (++qp->s_cur == qp->s_size) 384 if (++qp->s_cur == qp->s_size)
@@ -338,32 +387,48 @@ int ipath_make_rc_req(struct ipath_qp *qp,
338 387
339 case IB_WR_ATOMIC_CMP_AND_SWP: 388 case IB_WR_ATOMIC_CMP_AND_SWP:
340 case IB_WR_ATOMIC_FETCH_AND_ADD: 389 case IB_WR_ATOMIC_FETCH_AND_ADD:
341 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) 390 /*
342 qp->s_state = OP(COMPARE_SWAP); 391 * Don't allow more operations to be started
343 else 392 * than the QP limits allow.
344 qp->s_state = OP(FETCH_ADD); 393 */
345 ohdr->u.atomic_eth.vaddr = cpu_to_be64(
346 wqe->wr.wr.atomic.remote_addr);
347 ohdr->u.atomic_eth.rkey = cpu_to_be32(
348 wqe->wr.wr.atomic.rkey);
349 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
350 wqe->wr.wr.atomic.swap);
351 ohdr->u.atomic_eth.compare_data = cpu_to_be64(
352 wqe->wr.wr.atomic.compare_add);
353 hwords += sizeof(struct ib_atomic_eth) / 4;
354 if (newreq) { 394 if (newreq) {
395 if (qp->s_num_rd_atomic >=
396 qp->s_max_rd_atomic) {
397 qp->s_flags |= IPATH_S_RDMAR_PENDING;
398 goto bail;
399 }
400 qp->s_num_rd_atomic++;
355 if (qp->s_lsn != (u32) -1) 401 if (qp->s_lsn != (u32) -1)
356 qp->s_lsn++; 402 qp->s_lsn++;
357 wqe->lpsn = wqe->psn; 403 wqe->lpsn = wqe->psn;
358 } 404 }
359 if (++qp->s_cur == qp->s_size) 405 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
360 qp->s_cur = 0; 406 qp->s_state = OP(COMPARE_SWAP);
407 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
408 wqe->wr.wr.atomic.swap);
409 ohdr->u.atomic_eth.compare_data = cpu_to_be64(
410 wqe->wr.wr.atomic.compare_add);
411 } else {
412 qp->s_state = OP(FETCH_ADD);
413 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
414 wqe->wr.wr.atomic.compare_add);
415 ohdr->u.atomic_eth.compare_data = 0;
416 }
417 ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
418 wqe->wr.wr.atomic.remote_addr >> 32);
419 ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
420 wqe->wr.wr.atomic.remote_addr);
421 ohdr->u.atomic_eth.rkey = cpu_to_be32(
422 wqe->wr.wr.atomic.rkey);
423 hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
361 ss = NULL; 424 ss = NULL;
362 len = 0; 425 len = 0;
426 if (++qp->s_cur == qp->s_size)
427 qp->s_cur = 0;
363 break; 428 break;
364 429
365 default: 430 default:
366 goto done; 431 goto bail;
367 } 432 }
368 qp->s_sge.sge = wqe->sg_list[0]; 433 qp->s_sge.sge = wqe->sg_list[0];
369 qp->s_sge.sg_list = wqe->sg_list + 1; 434 qp->s_sge.sg_list = wqe->sg_list + 1;
@@ -379,7 +444,7 @@ int ipath_make_rc_req(struct ipath_qp *qp,
379 qp->s_psn = wqe->lpsn + 1; 444 qp->s_psn = wqe->lpsn + 1;
380 else { 445 else {
381 qp->s_psn++; 446 qp->s_psn++;
382 if ((int)(qp->s_psn - qp->s_next_psn) > 0) 447 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
383 qp->s_next_psn = qp->s_psn; 448 qp->s_next_psn = qp->s_psn;
384 } 449 }
385 /* 450 /*
@@ -406,7 +471,7 @@ int ipath_make_rc_req(struct ipath_qp *qp,
406 /* FALLTHROUGH */ 471 /* FALLTHROUGH */
407 case OP(SEND_MIDDLE): 472 case OP(SEND_MIDDLE):
408 bth2 = qp->s_psn++ & IPATH_PSN_MASK; 473 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
409 if ((int)(qp->s_psn - qp->s_next_psn) > 0) 474 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
410 qp->s_next_psn = qp->s_psn; 475 qp->s_next_psn = qp->s_psn;
411 ss = &qp->s_sge; 476 ss = &qp->s_sge;
412 len = qp->s_len; 477 len = qp->s_len;
@@ -442,7 +507,7 @@ int ipath_make_rc_req(struct ipath_qp *qp,
442 /* FALLTHROUGH */ 507 /* FALLTHROUGH */
443 case OP(RDMA_WRITE_MIDDLE): 508 case OP(RDMA_WRITE_MIDDLE):
444 bth2 = qp->s_psn++ & IPATH_PSN_MASK; 509 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
445 if ((int)(qp->s_psn - qp->s_next_psn) > 0) 510 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
446 qp->s_next_psn = qp->s_psn; 511 qp->s_next_psn = qp->s_psn;
447 ss = &qp->s_sge; 512 ss = &qp->s_sge;
448 len = qp->s_len; 513 len = qp->s_len;
@@ -479,9 +544,9 @@ int ipath_make_rc_req(struct ipath_qp *qp,
479 cpu_to_be32(wqe->wr.wr.rdma.rkey); 544 cpu_to_be32(wqe->wr.wr.rdma.rkey);
480 ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len); 545 ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
481 qp->s_state = OP(RDMA_READ_REQUEST); 546 qp->s_state = OP(RDMA_READ_REQUEST);
482 hwords += sizeof(ohdr->u.rc.reth) / 4; 547 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
483 bth2 = qp->s_psn++ & IPATH_PSN_MASK; 548 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
484 if ((int)(qp->s_psn - qp->s_next_psn) > 0) 549 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
485 qp->s_next_psn = qp->s_psn; 550 qp->s_next_psn = qp->s_psn;
486 ss = NULL; 551 ss = NULL;
487 len = 0; 552 len = 0;
@@ -489,20 +554,6 @@ int ipath_make_rc_req(struct ipath_qp *qp,
489 if (qp->s_cur == qp->s_size) 554 if (qp->s_cur == qp->s_size)
490 qp->s_cur = 0; 555 qp->s_cur = 0;
491 break; 556 break;
492
493 case OP(RDMA_READ_REQUEST):
494 case OP(COMPARE_SWAP):
495 case OP(FETCH_ADD):
496 /*
497 * We shouldn't start anything new until this request is
498 * finished. The ACK will handle rescheduling us. XXX The
499 * number of outstanding ones is negotiated at connection
500 * setup time (see pg. 258,289)? XXX Also, if we support
501 * multiple outstanding requests, we need to check the WQE
502 * IB_SEND_FENCE flag and not send a new request if a RDMA
503 * read or atomic is pending.
504 */
505 goto done;
506 } 557 }
507 if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0) 558 if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0)
508 bth2 |= 1 << 31; /* Request ACK. */ 559 bth2 |= 1 << 31; /* Request ACK. */
@@ -512,9 +563,10 @@ int ipath_make_rc_req(struct ipath_qp *qp,
512 qp->s_cur_size = len; 563 qp->s_cur_size = len;
513 *bth0p = bth0 | (qp->s_state << 24); 564 *bth0p = bth0 | (qp->s_state << 24);
514 *bth2p = bth2; 565 *bth2p = bth2;
566done:
515 return 1; 567 return 1;
516 568
517done: 569bail:
518 return 0; 570 return 0;
519} 571}
520 572
@@ -524,7 +576,8 @@ done:
524 * 576 *
525 * This is called from ipath_rc_rcv() and only uses the receive 577 * This is called from ipath_rc_rcv() and only uses the receive
526 * side QP state. 578 * side QP state.
527 * Note that RDMA reads are handled in the send side QP state and tasklet. 579 * Note that RDMA reads and atomics are handled in the
580 * send side QP state and tasklet.
528 */ 581 */
529static void send_rc_ack(struct ipath_qp *qp) 582static void send_rc_ack(struct ipath_qp *qp)
530{ 583{
@@ -535,6 +588,10 @@ static void send_rc_ack(struct ipath_qp *qp)
535 struct ipath_ib_header hdr; 588 struct ipath_ib_header hdr;
536 struct ipath_other_headers *ohdr; 589 struct ipath_other_headers *ohdr;
537 590
591 /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
592 if (qp->r_head_ack_queue != qp->s_tail_ack_queue)
593 goto queue_ack;
594
538 /* Construct the header. */ 595 /* Construct the header. */
539 ohdr = &hdr.u.oth; 596 ohdr = &hdr.u.oth;
540 lrh0 = IPATH_LRH_BTH; 597 lrh0 = IPATH_LRH_BTH;
@@ -548,19 +605,14 @@ static void send_rc_ack(struct ipath_qp *qp)
548 lrh0 = IPATH_LRH_GRH; 605 lrh0 = IPATH_LRH_GRH;
549 } 606 }
550 /* read pkey_index w/o lock (its atomic) */ 607 /* read pkey_index w/o lock (its atomic) */
551 bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index); 608 bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index) |
609 OP(ACKNOWLEDGE) << 24;
552 if (qp->r_nak_state) 610 if (qp->r_nak_state)
553 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) | 611 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
554 (qp->r_nak_state << 612 (qp->r_nak_state <<
555 IPATH_AETH_CREDIT_SHIFT)); 613 IPATH_AETH_CREDIT_SHIFT));
556 else 614 else
557 ohdr->u.aeth = ipath_compute_aeth(qp); 615 ohdr->u.aeth = ipath_compute_aeth(qp);
558 if (qp->r_ack_state >= OP(COMPARE_SWAP)) {
559 bth0 |= OP(ATOMIC_ACKNOWLEDGE) << 24;
560 ohdr->u.at.atomic_ack_eth = cpu_to_be64(qp->r_atomic_data);
561 hwords += sizeof(ohdr->u.at.atomic_ack_eth) / 4;
562 } else
563 bth0 |= OP(ACKNOWLEDGE) << 24;
564 lrh0 |= qp->remote_ah_attr.sl << 4; 616 lrh0 |= qp->remote_ah_attr.sl << 4;
565 hdr.lrh[0] = cpu_to_be16(lrh0); 617 hdr.lrh[0] = cpu_to_be16(lrh0);
566 hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid); 618 hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
@@ -574,31 +626,31 @@ static void send_rc_ack(struct ipath_qp *qp)
574 * If we can send the ACK, clear the ACK state. 626 * If we can send the ACK, clear the ACK state.
575 */ 627 */
576 if (ipath_verbs_send(dev->dd, hwords, (u32 *) &hdr, 0, NULL) == 0) { 628 if (ipath_verbs_send(dev->dd, hwords, (u32 *) &hdr, 0, NULL) == 0) {
577 qp->r_ack_state = OP(ACKNOWLEDGE);
578 dev->n_unicast_xmit++; 629 dev->n_unicast_xmit++;
579 } else { 630 goto done;
580 /*
581 * We are out of PIO buffers at the moment.
582 * Pass responsibility for sending the ACK to the
583 * send tasklet so that when a PIO buffer becomes
584 * available, the ACK is sent ahead of other outgoing
585 * packets.
586 */
587 dev->n_rc_qacks++;
588 spin_lock_irq(&qp->s_lock);
589 /* Don't coalesce if a RDMA read or atomic is pending. */
590 if (qp->s_ack_state == OP(ACKNOWLEDGE) ||
591 qp->s_ack_state < OP(RDMA_READ_REQUEST)) {
592 qp->s_ack_state = qp->r_ack_state;
593 qp->s_nak_state = qp->r_nak_state;
594 qp->s_ack_psn = qp->r_ack_psn;
595 qp->r_ack_state = OP(ACKNOWLEDGE);
596 }
597 spin_unlock_irq(&qp->s_lock);
598
599 /* Call ipath_do_rc_send() in another thread. */
600 tasklet_hi_schedule(&qp->s_task);
601 } 631 }
632
633 /*
634 * We are out of PIO buffers at the moment.
635 * Pass responsibility for sending the ACK to the
636 * send tasklet so that when a PIO buffer becomes
637 * available, the ACK is sent ahead of other outgoing
638 * packets.
639 */
640 dev->n_rc_qacks++;
641
642queue_ack:
643 spin_lock_irq(&qp->s_lock);
644 qp->s_flags |= IPATH_S_ACK_PENDING;
645 qp->s_nak_state = qp->r_nak_state;
646 qp->s_ack_psn = qp->r_ack_psn;
647 spin_unlock_irq(&qp->s_lock);
648
649 /* Call ipath_do_rc_send() in another thread. */
650 tasklet_hi_schedule(&qp->s_task);
651
652done:
653 return;
602} 654}
603 655
604/** 656/**
@@ -727,7 +779,7 @@ void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
727 if (wqe->wr.opcode == IB_WR_RDMA_READ) 779 if (wqe->wr.opcode == IB_WR_RDMA_READ)
728 dev->n_rc_resends++; 780 dev->n_rc_resends++;
729 else 781 else
730 dev->n_rc_resends += (int)qp->s_psn - (int)psn; 782 dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
731 783
732 reset_psn(qp, psn); 784 reset_psn(qp, psn);
733 tasklet_hi_schedule(&qp->s_task); 785 tasklet_hi_schedule(&qp->s_task);
@@ -775,10 +827,6 @@ static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
775 list_del_init(&qp->timerwait); 827 list_del_init(&qp->timerwait);
776 spin_unlock(&dev->pending_lock); 828 spin_unlock(&dev->pending_lock);
777 829
778 /* Nothing is pending to ACK/NAK. */
779 if (unlikely(qp->s_last == qp->s_tail))
780 goto bail;
781
782 /* 830 /*
783 * Note that NAKs implicitly ACK outstanding SEND and RDMA write 831 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
784 * requests and implicitly NAK RDMA read and atomic requests issued 832 * requests and implicitly NAK RDMA read and atomic requests issued
@@ -806,7 +854,7 @@ static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
806 */ 854 */
807 if ((wqe->wr.opcode == IB_WR_RDMA_READ && 855 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
808 (opcode != OP(RDMA_READ_RESPONSE_LAST) || 856 (opcode != OP(RDMA_READ_RESPONSE_LAST) ||
809 ipath_cmp24(ack_psn, wqe->lpsn) != 0)) || 857 ipath_cmp24(ack_psn, wqe->lpsn) != 0)) ||
810 ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 858 ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
811 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) && 859 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
812 (opcode != OP(ATOMIC_ACKNOWLEDGE) || 860 (opcode != OP(ATOMIC_ACKNOWLEDGE) ||
@@ -824,20 +872,33 @@ static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
824 */ 872 */
825 goto bail; 873 goto bail;
826 } 874 }
827 if (wqe->wr.opcode == IB_WR_RDMA_READ || 875 if (qp->s_num_rd_atomic &&
828 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 876 (wqe->wr.opcode == IB_WR_RDMA_READ ||
829 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) 877 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
830 tasklet_hi_schedule(&qp->s_task); 878 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
879 qp->s_num_rd_atomic--;
880 /* Restart sending task if fence is complete */
881 if ((qp->s_flags & IPATH_S_FENCE_PENDING) &&
882 !qp->s_num_rd_atomic) {
883 qp->s_flags &= ~IPATH_S_FENCE_PENDING;
884 tasklet_hi_schedule(&qp->s_task);
885 } else if (qp->s_flags & IPATH_S_RDMAR_PENDING) {
886 qp->s_flags &= ~IPATH_S_RDMAR_PENDING;
887 tasklet_hi_schedule(&qp->s_task);
888 }
889 }
831 /* Post a send completion queue entry if requested. */ 890 /* Post a send completion queue entry if requested. */
832 if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &qp->s_flags) || 891 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
833 (wqe->wr.send_flags & IB_SEND_SIGNALED)) { 892 (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
834 wc.wr_id = wqe->wr.wr_id; 893 wc.wr_id = wqe->wr.wr_id;
835 wc.status = IB_WC_SUCCESS; 894 wc.status = IB_WC_SUCCESS;
836 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode]; 895 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
837 wc.vendor_err = 0; 896 wc.vendor_err = 0;
838 wc.byte_len = wqe->length; 897 wc.byte_len = wqe->length;
898 wc.imm_data = 0;
839 wc.qp = &qp->ibqp; 899 wc.qp = &qp->ibqp;
840 wc.src_qp = qp->remote_qpn; 900 wc.src_qp = qp->remote_qpn;
901 wc.wc_flags = 0;
841 wc.pkey_index = 0; 902 wc.pkey_index = 0;
842 wc.slid = qp->remote_ah_attr.dlid; 903 wc.slid = qp->remote_ah_attr.dlid;
843 wc.sl = qp->remote_ah_attr.sl; 904 wc.sl = qp->remote_ah_attr.sl;
@@ -854,15 +915,19 @@ static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
854 if (qp->s_last == qp->s_cur) { 915 if (qp->s_last == qp->s_cur) {
855 if (++qp->s_cur >= qp->s_size) 916 if (++qp->s_cur >= qp->s_size)
856 qp->s_cur = 0; 917 qp->s_cur = 0;
918 qp->s_last = qp->s_cur;
919 if (qp->s_last == qp->s_tail)
920 break;
857 wqe = get_swqe_ptr(qp, qp->s_cur); 921 wqe = get_swqe_ptr(qp, qp->s_cur);
858 qp->s_state = OP(SEND_LAST); 922 qp->s_state = OP(SEND_LAST);
859 qp->s_psn = wqe->psn; 923 qp->s_psn = wqe->psn;
924 } else {
925 if (++qp->s_last >= qp->s_size)
926 qp->s_last = 0;
927 if (qp->s_last == qp->s_tail)
928 break;
929 wqe = get_swqe_ptr(qp, qp->s_last);
860 } 930 }
861 if (++qp->s_last >= qp->s_size)
862 qp->s_last = 0;
863 wqe = get_swqe_ptr(qp, qp->s_last);
864 if (qp->s_last == qp->s_tail)
865 break;
866 } 931 }
867 932
868 switch (aeth >> 29) { 933 switch (aeth >> 29) {
@@ -874,6 +939,18 @@ static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
874 list_add_tail(&qp->timerwait, 939 list_add_tail(&qp->timerwait,
875 &dev->pending[dev->pending_index]); 940 &dev->pending[dev->pending_index]);
876 spin_unlock(&dev->pending_lock); 941 spin_unlock(&dev->pending_lock);
942 /*
943 * If we get a partial ACK for a resent operation,
944 * we can stop resending the earlier packets and
945 * continue with the next packet the receiver wants.
946 */
947 if (ipath_cmp24(qp->s_psn, psn) <= 0) {
948 reset_psn(qp, psn + 1);
949 tasklet_hi_schedule(&qp->s_task);
950 }
951 } else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
952 qp->s_state = OP(SEND_LAST);
953 qp->s_psn = psn + 1;
877 } 954 }
878 ipath_get_credit(qp, aeth); 955 ipath_get_credit(qp, aeth);
879 qp->s_rnr_retry = qp->s_rnr_retry_cnt; 956 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
@@ -884,22 +961,23 @@ static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
884 961
885 case 1: /* RNR NAK */ 962 case 1: /* RNR NAK */
886 dev->n_rnr_naks++; 963 dev->n_rnr_naks++;
964 if (qp->s_last == qp->s_tail)
965 goto bail;
887 if (qp->s_rnr_retry == 0) { 966 if (qp->s_rnr_retry == 0) {
888 if (qp->s_last == qp->s_tail)
889 goto bail;
890
891 wc.status = IB_WC_RNR_RETRY_EXC_ERR; 967 wc.status = IB_WC_RNR_RETRY_EXC_ERR;
892 goto class_b; 968 goto class_b;
893 } 969 }
894 if (qp->s_rnr_retry_cnt < 7) 970 if (qp->s_rnr_retry_cnt < 7)
895 qp->s_rnr_retry--; 971 qp->s_rnr_retry--;
896 if (qp->s_last == qp->s_tail)
897 goto bail;
898 972
899 /* The last valid PSN is the previous PSN. */ 973 /* The last valid PSN is the previous PSN. */
900 update_last_psn(qp, psn - 1); 974 update_last_psn(qp, psn - 1);
901 975
902 dev->n_rc_resends += (int)qp->s_psn - (int)psn; 976 if (wqe->wr.opcode == IB_WR_RDMA_READ)
977 dev->n_rc_resends++;
978 else
979 dev->n_rc_resends +=
980 (qp->s_psn - psn) & IPATH_PSN_MASK;
903 981
904 reset_psn(qp, psn); 982 reset_psn(qp, psn);
905 983
@@ -910,26 +988,20 @@ static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
910 goto bail; 988 goto bail;
911 989
912 case 3: /* NAK */ 990 case 3: /* NAK */
913 /* The last valid PSN seen is the previous request's. */ 991 if (qp->s_last == qp->s_tail)
914 if (qp->s_last != qp->s_tail) 992 goto bail;
915 update_last_psn(qp, wqe->psn - 1); 993 /* The last valid PSN is the previous PSN. */
994 update_last_psn(qp, psn - 1);
916 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) & 995 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
917 IPATH_AETH_CREDIT_MASK) { 996 IPATH_AETH_CREDIT_MASK) {
918 case 0: /* PSN sequence error */ 997 case 0: /* PSN sequence error */
919 dev->n_seq_naks++; 998 dev->n_seq_naks++;
920 /* 999 /*
921 * Back up to the responder's expected PSN. XXX 1000 * Back up to the responder's expected PSN.
922 * Note that we might get a NAK in the middle of an 1001 * Note that we might get a NAK in the middle of an
923 * RDMA READ response which terminates the RDMA 1002 * RDMA READ response which terminates the RDMA
924 * READ. 1003 * READ.
925 */ 1004 */
926 if (qp->s_last == qp->s_tail)
927 break;
928
929 if (ipath_cmp24(psn, wqe->psn) < 0)
930 break;
931
932 /* Retry the request. */
933 ipath_restart_rc(qp, psn, &wc); 1005 ipath_restart_rc(qp, psn, &wc);
934 break; 1006 break;
935 1007
@@ -1003,6 +1075,7 @@ static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1003 u32 psn, u32 hdrsize, u32 pmtu, 1075 u32 psn, u32 hdrsize, u32 pmtu,
1004 int header_in_data) 1076 int header_in_data)
1005{ 1077{
1078 struct ipath_swqe *wqe;
1006 unsigned long flags; 1079 unsigned long flags;
1007 struct ib_wc wc; 1080 struct ib_wc wc;
1008 int diff; 1081 int diff;
@@ -1032,6 +1105,10 @@ static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1032 goto ack_done; 1105 goto ack_done;
1033 } 1106 }
1034 1107
1108 if (unlikely(qp->s_last == qp->s_tail))
1109 goto ack_done;
1110 wqe = get_swqe_ptr(qp, qp->s_last);
1111
1035 switch (opcode) { 1112 switch (opcode) {
1036 case OP(ACKNOWLEDGE): 1113 case OP(ACKNOWLEDGE):
1037 case OP(ATOMIC_ACKNOWLEDGE): 1114 case OP(ATOMIC_ACKNOWLEDGE):
@@ -1042,38 +1119,49 @@ static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1042 aeth = be32_to_cpu(((__be32 *) data)[0]); 1119 aeth = be32_to_cpu(((__be32 *) data)[0]);
1043 data += sizeof(__be32); 1120 data += sizeof(__be32);
1044 } 1121 }
1045 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) 1122 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1046 *(u64 *) qp->s_sge.sge.vaddr = *(u64 *) data; 1123 u64 val;
1124
1125 if (!header_in_data) {
1126 __be32 *p = ohdr->u.at.atomic_ack_eth;
1127
1128 val = ((u64) be32_to_cpu(p[0]) << 32) |
1129 be32_to_cpu(p[1]);
1130 } else
1131 val = be64_to_cpu(((__be64 *) data)[0]);
1132 *(u64 *) wqe->sg_list[0].vaddr = val;
1133 }
1047 if (!do_rc_ack(qp, aeth, psn, opcode) || 1134 if (!do_rc_ack(qp, aeth, psn, opcode) ||
1048 opcode != OP(RDMA_READ_RESPONSE_FIRST)) 1135 opcode != OP(RDMA_READ_RESPONSE_FIRST))
1049 goto ack_done; 1136 goto ack_done;
1050 hdrsize += 4; 1137 hdrsize += 4;
1138 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1139 goto ack_op_err;
1051 /* 1140 /*
1052 * do_rc_ack() has already checked the PSN so skip 1141 * If this is a response to a resent RDMA read, we
1053 * the sequence check. 1142 * have to be careful to copy the data to the right
1143 * location.
1054 */ 1144 */
1055 goto rdma_read; 1145 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1146 wqe, psn, pmtu);
1147 goto read_middle;
1056 1148
1057 case OP(RDMA_READ_RESPONSE_MIDDLE): 1149 case OP(RDMA_READ_RESPONSE_MIDDLE):
1058 /* no AETH, no ACK */ 1150 /* no AETH, no ACK */
1059 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) { 1151 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1060 dev->n_rdma_seq++; 1152 dev->n_rdma_seq++;
1061 if (qp->s_last != qp->s_tail) 1153 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1062 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1063 goto ack_done; 1154 goto ack_done;
1064 } 1155 }
1065 rdma_read: 1156 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1066 if (unlikely(qp->s_state != OP(RDMA_READ_REQUEST))) 1157 goto ack_op_err;
1067 goto ack_done; 1158 read_middle:
1068 if (unlikely(tlen != (hdrsize + pmtu + 4))) 1159 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1069 goto ack_done; 1160 goto ack_len_err;
1070 if (unlikely(pmtu >= qp->s_len)) 1161 if (unlikely(pmtu >= qp->s_rdma_read_len))
1071 goto ack_done; 1162 goto ack_len_err;
1163
1072 /* We got a response so update the timeout. */ 1164 /* We got a response so update the timeout. */
1073 if (unlikely(qp->s_last == qp->s_tail ||
1074 get_swqe_ptr(qp, qp->s_last)->wr.opcode !=
1075 IB_WR_RDMA_READ))
1076 goto ack_done;
1077 spin_lock(&dev->pending_lock); 1165 spin_lock(&dev->pending_lock);
1078 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait)) 1166 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1079 list_move_tail(&qp->timerwait, 1167 list_move_tail(&qp->timerwait,
@@ -1082,67 +1170,97 @@ static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1082 /* 1170 /*
1083 * Update the RDMA receive state but do the copy w/o 1171 * Update the RDMA receive state but do the copy w/o
1084 * holding the locks and blocking interrupts. 1172 * holding the locks and blocking interrupts.
1085 * XXX Yet another place that affects relaxed RDMA order
1086 * since we don't want s_sge modified.
1087 */ 1173 */
1088 qp->s_len -= pmtu; 1174 qp->s_rdma_read_len -= pmtu;
1089 update_last_psn(qp, psn); 1175 update_last_psn(qp, psn);
1090 spin_unlock_irqrestore(&qp->s_lock, flags); 1176 spin_unlock_irqrestore(&qp->s_lock, flags);
1091 ipath_copy_sge(&qp->s_sge, data, pmtu); 1177 ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1092 goto bail; 1178 goto bail;
1093 1179
1094 case OP(RDMA_READ_RESPONSE_LAST): 1180 case OP(RDMA_READ_RESPONSE_ONLY):
1095 /* ACKs READ req. */
1096 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) { 1181 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1097 dev->n_rdma_seq++; 1182 dev->n_rdma_seq++;
1098 if (qp->s_last != qp->s_tail) 1183 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1099 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1100 goto ack_done; 1184 goto ack_done;
1101 } 1185 }
1102 /* FALLTHROUGH */ 1186 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1103 case OP(RDMA_READ_RESPONSE_ONLY): 1187 goto ack_op_err;
1104 if (unlikely(qp->s_state != OP(RDMA_READ_REQUEST))) 1188 /* Get the number of bytes the message was padded by. */
1105 goto ack_done; 1189 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1190 /*
1191 * Check that the data size is >= 0 && <= pmtu.
1192 * Remember to account for the AETH header (4) and
1193 * ICRC (4).
1194 */
1195 if (unlikely(tlen < (hdrsize + pad + 8)))
1196 goto ack_len_err;
1106 /* 1197 /*
1107 * Get the number of bytes the message was padded by. 1198 * If this is a response to a resent RDMA read, we
1199 * have to be careful to copy the data to the right
1200 * location.
1108 */ 1201 */
1202 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1203 wqe, psn, pmtu);
1204 goto read_last;
1205
1206 case OP(RDMA_READ_RESPONSE_LAST):
1207 /* ACKs READ req. */
1208 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1209 dev->n_rdma_seq++;
1210 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1211 goto ack_done;
1212 }
1213 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1214 goto ack_op_err;
1215 /* Get the number of bytes the message was padded by. */
1109 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3; 1216 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1110 /* 1217 /*
1111 * Check that the data size is >= 1 && <= pmtu. 1218 * Check that the data size is >= 1 && <= pmtu.
1112 * Remember to account for the AETH header (4) and 1219 * Remember to account for the AETH header (4) and
1113 * ICRC (4). 1220 * ICRC (4).
1114 */ 1221 */
1115 if (unlikely(tlen <= (hdrsize + pad + 8))) { 1222 if (unlikely(tlen <= (hdrsize + pad + 8)))
1116 /* XXX Need to generate an error CQ entry. */ 1223 goto ack_len_err;
1117 goto ack_done; 1224 read_last:
1118 }
1119 tlen -= hdrsize + pad + 8; 1225 tlen -= hdrsize + pad + 8;
1120 if (unlikely(tlen != qp->s_len)) { 1226 if (unlikely(tlen != qp->s_rdma_read_len))
1121 /* XXX Need to generate an error CQ entry. */ 1227 goto ack_len_err;
1122 goto ack_done;
1123 }
1124 if (!header_in_data) 1228 if (!header_in_data)
1125 aeth = be32_to_cpu(ohdr->u.aeth); 1229 aeth = be32_to_cpu(ohdr->u.aeth);
1126 else { 1230 else {
1127 aeth = be32_to_cpu(((__be32 *) data)[0]); 1231 aeth = be32_to_cpu(((__be32 *) data)[0]);
1128 data += sizeof(__be32); 1232 data += sizeof(__be32);
1129 } 1233 }
1130 ipath_copy_sge(&qp->s_sge, data, tlen); 1234 ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1131 if (do_rc_ack(qp, aeth, psn, OP(RDMA_READ_RESPONSE_LAST))) { 1235 (void) do_rc_ack(qp, aeth, psn, OP(RDMA_READ_RESPONSE_LAST));
1132 /*
1133 * Change the state so we contimue
1134 * processing new requests and wake up the
1135 * tasklet if there are posted sends.
1136 */
1137 qp->s_state = OP(SEND_LAST);
1138 if (qp->s_tail != qp->s_head)
1139 tasklet_hi_schedule(&qp->s_task);
1140 }
1141 goto ack_done; 1236 goto ack_done;
1142 } 1237 }
1143 1238
1144ack_done: 1239ack_done:
1145 spin_unlock_irqrestore(&qp->s_lock, flags); 1240 spin_unlock_irqrestore(&qp->s_lock, flags);
1241 goto bail;
1242
1243ack_op_err:
1244 wc.status = IB_WC_LOC_QP_OP_ERR;
1245 goto ack_err;
1246
1247ack_len_err:
1248 wc.status = IB_WC_LOC_LEN_ERR;
1249ack_err:
1250 wc.wr_id = wqe->wr.wr_id;
1251 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1252 wc.vendor_err = 0;
1253 wc.byte_len = 0;
1254 wc.imm_data = 0;
1255 wc.qp = &qp->ibqp;
1256 wc.src_qp = qp->remote_qpn;
1257 wc.wc_flags = 0;
1258 wc.pkey_index = 0;
1259 wc.slid = qp->remote_ah_attr.dlid;
1260 wc.sl = qp->remote_ah_attr.sl;
1261 wc.dlid_path_bits = 0;
1262 wc.port_num = 0;
1263 ipath_sqerror_qp(qp, &wc);
1146bail: 1264bail:
1147 return; 1265 return;
1148} 1266}
@@ -1162,7 +1280,7 @@ bail:
1162 * incoming RC packet for the given QP. 1280 * incoming RC packet for the given QP.
1163 * Called at interrupt level. 1281 * Called at interrupt level.
1164 * Return 1 if no more processing is needed; otherwise return 0 to 1282 * Return 1 if no more processing is needed; otherwise return 0 to
1165 * schedule a response to be sent and the s_lock unlocked. 1283 * schedule a response to be sent.
1166 */ 1284 */
1167static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev, 1285static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1168 struct ipath_other_headers *ohdr, 1286 struct ipath_other_headers *ohdr,
@@ -1173,25 +1291,23 @@ static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1173 int diff, 1291 int diff,
1174 int header_in_data) 1292 int header_in_data)
1175{ 1293{
1176 struct ib_reth *reth; 1294 struct ipath_ack_entry *e;
1295 u8 i, prev;
1296 int old_req;
1177 1297
1178 if (diff > 0) { 1298 if (diff > 0) {
1179 /* 1299 /*
1180 * Packet sequence error. 1300 * Packet sequence error.
1181 * A NAK will ACK earlier sends and RDMA writes. 1301 * A NAK will ACK earlier sends and RDMA writes.
1182 * Don't queue the NAK if a RDMA read, atomic, or 1302 * Don't queue the NAK if we already sent one.
1183 * NAK is pending though.
1184 */ 1303 */
1185 if (qp->s_ack_state != OP(ACKNOWLEDGE) || 1304 if (!qp->r_nak_state) {
1186 qp->r_nak_state != 0)
1187 goto done;
1188 if (qp->r_ack_state < OP(COMPARE_SWAP)) {
1189 qp->r_ack_state = OP(SEND_ONLY);
1190 qp->r_nak_state = IB_NAK_PSN_ERROR; 1305 qp->r_nak_state = IB_NAK_PSN_ERROR;
1191 /* Use the expected PSN. */ 1306 /* Use the expected PSN. */
1192 qp->r_ack_psn = qp->r_psn; 1307 qp->r_ack_psn = qp->r_psn;
1308 goto send_ack;
1193 } 1309 }
1194 goto send_ack; 1310 goto done;
1195 } 1311 }
1196 1312
1197 /* 1313 /*
@@ -1204,8 +1320,46 @@ static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1204 * can coalesce an outstanding duplicate ACK. We have to 1320 * can coalesce an outstanding duplicate ACK. We have to
1205 * send the earliest so that RDMA reads can be restarted at 1321 * send the earliest so that RDMA reads can be restarted at
1206 * the requester's expected PSN. 1322 * the requester's expected PSN.
1323 *
1324 * First, find where this duplicate PSN falls within the
1325 * ACKs previously sent.
1207 */ 1326 */
1208 if (opcode == OP(RDMA_READ_REQUEST)) { 1327 psn &= IPATH_PSN_MASK;
1328 e = NULL;
1329 old_req = 1;
1330 spin_lock_irq(&qp->s_lock);
1331 for (i = qp->r_head_ack_queue; ; i = prev) {
1332 if (i == qp->s_tail_ack_queue)
1333 old_req = 0;
1334 if (i)
1335 prev = i - 1;
1336 else
1337 prev = IPATH_MAX_RDMA_ATOMIC;
1338 if (prev == qp->r_head_ack_queue) {
1339 e = NULL;
1340 break;
1341 }
1342 e = &qp->s_ack_queue[prev];
1343 if (!e->opcode) {
1344 e = NULL;
1345 break;
1346 }
1347 if (ipath_cmp24(psn, e->psn) >= 0)
1348 break;
1349 }
1350 switch (opcode) {
1351 case OP(RDMA_READ_REQUEST): {
1352 struct ib_reth *reth;
1353 u32 offset;
1354 u32 len;
1355
1356 /*
1357 * If we didn't find the RDMA read request in the ack queue,
1358 * or the send tasklet is already backed up to send an
1359 * earlier entry, we can ignore this request.
1360 */
1361 if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
1362 goto unlock_done;
1209 /* RETH comes after BTH */ 1363 /* RETH comes after BTH */
1210 if (!header_in_data) 1364 if (!header_in_data)
1211 reth = &ohdr->u.rc.reth; 1365 reth = &ohdr->u.rc.reth;
@@ -1214,88 +1368,87 @@ static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1214 data += sizeof(*reth); 1368 data += sizeof(*reth);
1215 } 1369 }
1216 /* 1370 /*
1217 * If we receive a duplicate RDMA request, it means the 1371 * Address range must be a subset of the original
1218 * requester saw a sequence error and needs to restart 1372 * request and start on pmtu boundaries.
1219 * from an earlier point. We can abort the current 1373 * We reuse the old ack_queue slot since the requester
1220 * RDMA read send in that case. 1374 * should not back up and request an earlier PSN for the
1375 * same request.
1221 */ 1376 */
1222 spin_lock_irq(&qp->s_lock); 1377 offset = ((psn - e->psn) & IPATH_PSN_MASK) *
1223 if (qp->s_ack_state != OP(ACKNOWLEDGE) && 1378 ib_mtu_enum_to_int(qp->path_mtu);
1224 (qp->s_hdrwords || ipath_cmp24(psn, qp->s_ack_psn) >= 0)) { 1379 len = be32_to_cpu(reth->length);
1225 /* 1380 if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
1226 * We are already sending earlier requested data. 1381 goto unlock_done;
1227 * Don't abort it to send later out of sequence data. 1382 if (len != 0) {
1228 */
1229 spin_unlock_irq(&qp->s_lock);
1230 goto done;
1231 }
1232 qp->s_rdma_len = be32_to_cpu(reth->length);
1233 if (qp->s_rdma_len != 0) {
1234 u32 rkey = be32_to_cpu(reth->rkey); 1383 u32 rkey = be32_to_cpu(reth->rkey);
1235 u64 vaddr = be64_to_cpu(reth->vaddr); 1384 u64 vaddr = be64_to_cpu(reth->vaddr);
1236 int ok; 1385 int ok;
1237 1386
1238 /* 1387 ok = ipath_rkey_ok(qp, &e->rdma_sge,
1239 * Address range must be a subset of the original 1388 len, vaddr, rkey,
1240 * request and start on pmtu boundaries.
1241 */
1242 ok = ipath_rkey_ok(qp, &qp->s_rdma_sge,
1243 qp->s_rdma_len, vaddr, rkey,
1244 IB_ACCESS_REMOTE_READ); 1389 IB_ACCESS_REMOTE_READ);
1245 if (unlikely(!ok)) { 1390 if (unlikely(!ok))
1246 spin_unlock_irq(&qp->s_lock); 1391 goto unlock_done;
1247 goto done;
1248 }
1249 } else { 1392 } else {
1250 qp->s_rdma_sge.sg_list = NULL; 1393 e->rdma_sge.sg_list = NULL;
1251 qp->s_rdma_sge.num_sge = 0; 1394 e->rdma_sge.num_sge = 0;
1252 qp->s_rdma_sge.sge.mr = NULL; 1395 e->rdma_sge.sge.mr = NULL;
1253 qp->s_rdma_sge.sge.vaddr = NULL; 1396 e->rdma_sge.sge.vaddr = NULL;
1254 qp->s_rdma_sge.sge.length = 0; 1397 e->rdma_sge.sge.length = 0;
1255 qp->s_rdma_sge.sge.sge_length = 0; 1398 e->rdma_sge.sge.sge_length = 0;
1256 } 1399 }
1257 qp->s_ack_state = opcode; 1400 e->psn = psn;
1258 qp->s_ack_psn = psn; 1401 qp->s_ack_state = OP(ACKNOWLEDGE);
1259 spin_unlock_irq(&qp->s_lock); 1402 qp->s_tail_ack_queue = prev;
1260 tasklet_hi_schedule(&qp->s_task); 1403 break;
1261 goto send_ack;
1262 } 1404 }
1263 1405
1264 /*
1265 * A pending RDMA read will ACK anything before it so
1266 * ignore earlier duplicate requests.
1267 */
1268 if (qp->s_ack_state != OP(ACKNOWLEDGE))
1269 goto done;
1270
1271 /*
1272 * If an ACK is pending, don't replace the pending ACK
1273 * with an earlier one since the later one will ACK the earlier.
1274 * Also, if we already have a pending atomic, send it.
1275 */
1276 if (qp->r_ack_state != OP(ACKNOWLEDGE) &&
1277 (ipath_cmp24(psn, qp->r_ack_psn) <= 0 ||
1278 qp->r_ack_state >= OP(COMPARE_SWAP)))
1279 goto send_ack;
1280 switch (opcode) {
1281 case OP(COMPARE_SWAP): 1406 case OP(COMPARE_SWAP):
1282 case OP(FETCH_ADD): 1407 case OP(FETCH_ADD): {
1283 /* 1408 /*
1284 * Check for the PSN of the last atomic operation 1409 * If we didn't find the atomic request in the ack queue
1285 * performed and resend the result if found. 1410 * or the send tasklet is already backed up to send an
1411 * earlier entry, we can ignore this request.
1286 */ 1412 */
1287 if ((psn & IPATH_PSN_MASK) != qp->r_atomic_psn) 1413 if (!e || e->opcode != (u8) opcode || old_req)
1288 goto done; 1414 goto unlock_done;
1415 qp->s_ack_state = OP(ACKNOWLEDGE);
1416 qp->s_tail_ack_queue = prev;
1417 break;
1418 }
1419
1420 default:
1421 if (old_req)
1422 goto unlock_done;
1423 /*
1424 * Resend the most recent ACK if this request is
1425 * after all the previous RDMA reads and atomics.
1426 */
1427 if (i == qp->r_head_ack_queue) {
1428 spin_unlock_irq(&qp->s_lock);
1429 qp->r_nak_state = 0;
1430 qp->r_ack_psn = qp->r_psn - 1;
1431 goto send_ack;
1432 }
1433 /*
1434 * Resend the RDMA read or atomic op which
1435 * ACKs this duplicate request.
1436 */
1437 qp->s_ack_state = OP(ACKNOWLEDGE);
1438 qp->s_tail_ack_queue = i;
1289 break; 1439 break;
1290 } 1440 }
1291 qp->r_ack_state = opcode;
1292 qp->r_nak_state = 0; 1441 qp->r_nak_state = 0;
1293 qp->r_ack_psn = psn; 1442 spin_unlock_irq(&qp->s_lock);
1294send_ack: 1443 tasklet_hi_schedule(&qp->s_task);
1295 return 0;
1296 1444
1445unlock_done:
1446 spin_unlock_irq(&qp->s_lock);
1297done: 1447done:
1298 return 1; 1448 return 1;
1449
1450send_ack:
1451 return 0;
1299} 1452}
1300 1453
1301static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err) 1454static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
@@ -1391,15 +1544,7 @@ void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1391 opcode == OP(SEND_LAST_WITH_IMMEDIATE)) 1544 opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1392 break; 1545 break;
1393 nack_inv: 1546 nack_inv:
1394 /*
1395 * A NAK will ACK earlier sends and RDMA writes.
1396 * Don't queue the NAK if a RDMA read, atomic, or NAK
1397 * is pending though.
1398 */
1399 if (qp->r_ack_state >= OP(COMPARE_SWAP))
1400 goto send_ack;
1401 ipath_rc_error(qp, IB_WC_REM_INV_REQ_ERR); 1547 ipath_rc_error(qp, IB_WC_REM_INV_REQ_ERR);
1402 qp->r_ack_state = OP(SEND_ONLY);
1403 qp->r_nak_state = IB_NAK_INVALID_REQUEST; 1548 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1404 qp->r_ack_psn = qp->r_psn; 1549 qp->r_ack_psn = qp->r_psn;
1405 goto send_ack; 1550 goto send_ack;
@@ -1441,9 +1586,8 @@ void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1441 * Don't queue the NAK if a RDMA read or atomic 1586 * Don't queue the NAK if a RDMA read or atomic
1442 * is pending though. 1587 * is pending though.
1443 */ 1588 */
1444 if (qp->r_ack_state >= OP(COMPARE_SWAP)) 1589 if (qp->r_nak_state)
1445 goto send_ack; 1590 goto done;
1446 qp->r_ack_state = OP(SEND_ONLY);
1447 qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer; 1591 qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1448 qp->r_ack_psn = qp->r_psn; 1592 qp->r_ack_psn = qp->r_psn;
1449 goto send_ack; 1593 goto send_ack;
@@ -1567,7 +1711,19 @@ void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1567 goto rnr_nak; 1711 goto rnr_nak;
1568 goto send_last_imm; 1712 goto send_last_imm;
1569 1713
1570 case OP(RDMA_READ_REQUEST): 1714 case OP(RDMA_READ_REQUEST): {
1715 struct ipath_ack_entry *e;
1716 u32 len;
1717 u8 next;
1718
1719 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
1720 goto nack_acc;
1721 next = qp->r_head_ack_queue + 1;
1722 if (next > IPATH_MAX_RDMA_ATOMIC)
1723 next = 0;
1724 if (unlikely(next == qp->s_tail_ack_queue))
1725 goto nack_inv;
1726 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1571 /* RETH comes after BTH */ 1727 /* RETH comes after BTH */
1572 if (!header_in_data) 1728 if (!header_in_data)
1573 reth = &ohdr->u.rc.reth; 1729 reth = &ohdr->u.rc.reth;
@@ -1575,72 +1731,75 @@ void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1575 reth = (struct ib_reth *)data; 1731 reth = (struct ib_reth *)data;
1576 data += sizeof(*reth); 1732 data += sizeof(*reth);
1577 } 1733 }
1578 if (unlikely(!(qp->qp_access_flags & 1734 len = be32_to_cpu(reth->length);
1579 IB_ACCESS_REMOTE_READ))) 1735 if (len) {
1580 goto nack_acc;
1581 spin_lock_irq(&qp->s_lock);
1582 qp->s_rdma_len = be32_to_cpu(reth->length);
1583 if (qp->s_rdma_len != 0) {
1584 u32 rkey = be32_to_cpu(reth->rkey); 1736 u32 rkey = be32_to_cpu(reth->rkey);
1585 u64 vaddr = be64_to_cpu(reth->vaddr); 1737 u64 vaddr = be64_to_cpu(reth->vaddr);
1586 int ok; 1738 int ok;
1587 1739
1588 /* Check rkey & NAK */ 1740 /* Check rkey & NAK */
1589 ok = ipath_rkey_ok(qp, &qp->s_rdma_sge, 1741 ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1590 qp->s_rdma_len, vaddr, rkey, 1742 rkey, IB_ACCESS_REMOTE_READ);
1591 IB_ACCESS_REMOTE_READ); 1743 if (unlikely(!ok))
1592 if (unlikely(!ok)) {
1593 spin_unlock_irq(&qp->s_lock);
1594 goto nack_acc; 1744 goto nack_acc;
1595 }
1596 /* 1745 /*
1597 * Update the next expected PSN. We add 1 later 1746 * Update the next expected PSN. We add 1 later
1598 * below, so only add the remainder here. 1747 * below, so only add the remainder here.
1599 */ 1748 */
1600 if (qp->s_rdma_len > pmtu) 1749 if (len > pmtu)
1601 qp->r_psn += (qp->s_rdma_len - 1) / pmtu; 1750 qp->r_psn += (len - 1) / pmtu;
1602 } else { 1751 } else {
1603 qp->s_rdma_sge.sg_list = NULL; 1752 e->rdma_sge.sg_list = NULL;
1604 qp->s_rdma_sge.num_sge = 0; 1753 e->rdma_sge.num_sge = 0;
1605 qp->s_rdma_sge.sge.mr = NULL; 1754 e->rdma_sge.sge.mr = NULL;
1606 qp->s_rdma_sge.sge.vaddr = NULL; 1755 e->rdma_sge.sge.vaddr = NULL;
1607 qp->s_rdma_sge.sge.length = 0; 1756 e->rdma_sge.sge.length = 0;
1608 qp->s_rdma_sge.sge.sge_length = 0; 1757 e->rdma_sge.sge.sge_length = 0;
1609 } 1758 }
1759 e->opcode = opcode;
1760 e->psn = psn;
1610 /* 1761 /*
1611 * We need to increment the MSN here instead of when we 1762 * We need to increment the MSN here instead of when we
1612 * finish sending the result since a duplicate request would 1763 * finish sending the result since a duplicate request would
1613 * increment it more than once. 1764 * increment it more than once.
1614 */ 1765 */
1615 qp->r_msn++; 1766 qp->r_msn++;
1616
1617 qp->s_ack_state = opcode;
1618 qp->s_ack_psn = psn;
1619 spin_unlock_irq(&qp->s_lock);
1620
1621 qp->r_psn++; 1767 qp->r_psn++;
1622 qp->r_state = opcode; 1768 qp->r_state = opcode;
1623 qp->r_nak_state = 0; 1769 qp->r_nak_state = 0;
1770 barrier();
1771 qp->r_head_ack_queue = next;
1624 1772
1625 /* Call ipath_do_rc_send() in another thread. */ 1773 /* Call ipath_do_rc_send() in another thread. */
1626 tasklet_hi_schedule(&qp->s_task); 1774 tasklet_hi_schedule(&qp->s_task);
1627 1775
1628 goto done; 1776 goto done;
1777 }
1629 1778
1630 case OP(COMPARE_SWAP): 1779 case OP(COMPARE_SWAP):
1631 case OP(FETCH_ADD): { 1780 case OP(FETCH_ADD): {
1632 struct ib_atomic_eth *ateth; 1781 struct ib_atomic_eth *ateth;
1782 struct ipath_ack_entry *e;
1633 u64 vaddr; 1783 u64 vaddr;
1784 atomic64_t *maddr;
1634 u64 sdata; 1785 u64 sdata;
1635 u32 rkey; 1786 u32 rkey;
1787 u8 next;
1636 1788
1789 if (unlikely(!(qp->qp_access_flags &
1790 IB_ACCESS_REMOTE_ATOMIC)))
1791 goto nack_acc;
1792 next = qp->r_head_ack_queue + 1;
1793 if (next > IPATH_MAX_RDMA_ATOMIC)
1794 next = 0;
1795 if (unlikely(next == qp->s_tail_ack_queue))
1796 goto nack_inv;
1637 if (!header_in_data) 1797 if (!header_in_data)
1638 ateth = &ohdr->u.atomic_eth; 1798 ateth = &ohdr->u.atomic_eth;
1639 else { 1799 else
1640 ateth = (struct ib_atomic_eth *)data; 1800 ateth = (struct ib_atomic_eth *)data;
1641 data += sizeof(*ateth); 1801 vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1642 } 1802 be32_to_cpu(ateth->vaddr[1]);
1643 vaddr = be64_to_cpu(ateth->vaddr);
1644 if (unlikely(vaddr & (sizeof(u64) - 1))) 1803 if (unlikely(vaddr & (sizeof(u64) - 1)))
1645 goto nack_inv; 1804 goto nack_inv;
1646 rkey = be32_to_cpu(ateth->rkey); 1805 rkey = be32_to_cpu(ateth->rkey);
@@ -1649,63 +1808,50 @@ void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1649 sizeof(u64), vaddr, rkey, 1808 sizeof(u64), vaddr, rkey,
1650 IB_ACCESS_REMOTE_ATOMIC))) 1809 IB_ACCESS_REMOTE_ATOMIC)))
1651 goto nack_acc; 1810 goto nack_acc;
1652 if (unlikely(!(qp->qp_access_flags &
1653 IB_ACCESS_REMOTE_ATOMIC)))
1654 goto nack_acc;
1655 /* Perform atomic OP and save result. */ 1811 /* Perform atomic OP and save result. */
1812 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1656 sdata = be64_to_cpu(ateth->swap_data); 1813 sdata = be64_to_cpu(ateth->swap_data);
1657 spin_lock_irq(&dev->pending_lock); 1814 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1658 qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr; 1815 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1659 if (opcode == OP(FETCH_ADD)) 1816 (u64) atomic64_add_return(sdata, maddr) - sdata :
1660 *(u64 *) qp->r_sge.sge.vaddr = 1817 (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1661 qp->r_atomic_data + sdata; 1818 be64_to_cpu(ateth->compare_data),
1662 else if (qp->r_atomic_data == 1819 sdata);
1663 be64_to_cpu(ateth->compare_data)) 1820 e->opcode = opcode;
1664 *(u64 *) qp->r_sge.sge.vaddr = sdata; 1821 e->psn = psn & IPATH_PSN_MASK;
1665 spin_unlock_irq(&dev->pending_lock);
1666 qp->r_msn++; 1822 qp->r_msn++;
1667 qp->r_atomic_psn = psn & IPATH_PSN_MASK; 1823 qp->r_psn++;
1668 psn |= 1 << 31; 1824 qp->r_state = opcode;
1669 break; 1825 qp->r_nak_state = 0;
1826 barrier();
1827 qp->r_head_ack_queue = next;
1828
1829 /* Call ipath_do_rc_send() in another thread. */
1830 tasklet_hi_schedule(&qp->s_task);
1831
1832 goto done;
1670 } 1833 }
1671 1834
1672 default: 1835 default:
1673 /* Drop packet for unknown opcodes. */ 1836 /* NAK unknown opcodes. */
1674 goto done; 1837 goto nack_inv;
1675 } 1838 }
1676 qp->r_psn++; 1839 qp->r_psn++;
1677 qp->r_state = opcode; 1840 qp->r_state = opcode;
1841 qp->r_ack_psn = psn;
1678 qp->r_nak_state = 0; 1842 qp->r_nak_state = 0;
1679 /* Send an ACK if requested or required. */ 1843 /* Send an ACK if requested or required. */
1680 if (psn & (1 << 31)) { 1844 if (psn & (1 << 31))
1681 /*
1682 * Coalesce ACKs unless there is a RDMA READ or
1683 * ATOMIC pending.
1684 */
1685 if (qp->r_ack_state < OP(COMPARE_SWAP)) {
1686 qp->r_ack_state = opcode;
1687 qp->r_ack_psn = psn;
1688 }
1689 goto send_ack; 1845 goto send_ack;
1690 }
1691 goto done; 1846 goto done;
1692 1847
1693nack_acc: 1848nack_acc:
1694 /* 1849 ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
1695 * A NAK will ACK earlier sends and RDMA writes. 1850 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1696 * Don't queue the NAK if a RDMA read, atomic, or NAK 1851 qp->r_ack_psn = qp->r_psn;
1697 * is pending though. 1852
1698 */
1699 if (qp->r_ack_state < OP(COMPARE_SWAP)) {
1700 ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
1701 qp->r_ack_state = OP(RDMA_WRITE_ONLY);
1702 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1703 qp->r_ack_psn = qp->r_psn;
1704 }
1705send_ack: 1853send_ack:
1706 /* Send ACK right away unless the send tasklet has a pending ACK. */ 1854 send_rc_ack(qp);
1707 if (qp->s_ack_state == OP(ACKNOWLEDGE))
1708 send_rc_ack(qp);
1709 1855
1710done: 1856done:
1711 return; 1857 return;
generated by cgit v1.2.2 (git 2.25.0) at 2025-01-12 04:02:01 -0500