aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/infiniband/hw/ipath/ipath_rc.c
diff options
context:
space:
mode:
authorRalph Campbell <ralph.campbell@qlogic.com>2007-03-15 17:44:51 -0400
committerRoland Dreier <rolandd@cisco.com>2007-04-18 23:20:55 -0400
commit3859e39d75b72f35f7d38c618fbbacb39a440c22 (patch)
tree51d57723574395b54914c08260b9d0a8467a91b1 /drivers/infiniband/hw/ipath/ipath_rc.c
parent7b21d26ddad6912bf345e8e88a51a5ce98a036ad (diff)
IB/ipath: Support larger IB_QP_MAX_DEST_RD_ATOMIC and IB_QP_MAX_QP_RD_ATOMIC
This patch adds support for multiple RDMA reads and atomics to be sent before an ACK is required to be seen by the requester. Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com> Signed-off-by: Roland Dreier <rolandd@cisco.com>
Diffstat (limited to 'drivers/infiniband/hw/ipath/ipath_rc.c')
-rw-r--r--drivers/infiniband/hw/ipath/ipath_rc.c800
1 files changed, 454 insertions, 346 deletions
diff --git a/drivers/infiniband/hw/ipath/ipath_rc.c b/drivers/infiniband/hw/ipath/ipath_rc.c
index 5ff20cb0449..c9c3d7cd292 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;
@@ -479,7 +544,7 @@ 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 ((int)(qp->s_psn - qp->s_next_psn) > 0)
485 qp->s_next_psn = qp->s_psn; 550 qp->s_next_psn = qp->s_psn;
@@ -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/**
@@ -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) ||
@@ -824,12 +872,23 @@ 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;
@@ -1003,6 +1062,7 @@ static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1003 u32 psn, u32 hdrsize, u32 pmtu, 1062 u32 psn, u32 hdrsize, u32 pmtu,
1004 int header_in_data) 1063 int header_in_data)
1005{ 1064{
1065 struct ipath_swqe *wqe;
1006 unsigned long flags; 1066 unsigned long flags;
1007 struct ib_wc wc; 1067 struct ib_wc wc;
1008 int diff; 1068 int diff;
@@ -1032,6 +1092,10 @@ static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1032 goto ack_done; 1092 goto ack_done;
1033 } 1093 }
1034 1094
1095 if (unlikely(qp->s_last == qp->s_tail))
1096 goto ack_done;
1097 wqe = get_swqe_ptr(qp, qp->s_last);
1098
1035 switch (opcode) { 1099 switch (opcode) {
1036 case OP(ACKNOWLEDGE): 1100 case OP(ACKNOWLEDGE):
1037 case OP(ATOMIC_ACKNOWLEDGE): 1101 case OP(ATOMIC_ACKNOWLEDGE):
@@ -1042,38 +1106,49 @@ static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1042 aeth = be32_to_cpu(((__be32 *) data)[0]); 1106 aeth = be32_to_cpu(((__be32 *) data)[0]);
1043 data += sizeof(__be32); 1107 data += sizeof(__be32);
1044 } 1108 }
1045 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) 1109 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1046 *(u64 *) qp->s_sge.sge.vaddr = *(u64 *) data; 1110 u64 val;
1111
1112 if (!header_in_data) {
1113 __be32 *p = ohdr->u.at.atomic_ack_eth;
1114
1115 val = ((u64) be32_to_cpu(p[0]) << 32) |
1116 be32_to_cpu(p[1]);
1117 } else
1118 val = be64_to_cpu(((__be64 *) data)[0]);
1119 *(u64 *) wqe->sg_list[0].vaddr = val;
1120 }
1047 if (!do_rc_ack(qp, aeth, psn, opcode) || 1121 if (!do_rc_ack(qp, aeth, psn, opcode) ||
1048 opcode != OP(RDMA_READ_RESPONSE_FIRST)) 1122 opcode != OP(RDMA_READ_RESPONSE_FIRST))
1049 goto ack_done; 1123 goto ack_done;
1050 hdrsize += 4; 1124 hdrsize += 4;
1125 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1126 goto ack_done;
1051 /* 1127 /*
1052 * do_rc_ack() has already checked the PSN so skip 1128 * If this is a response to a resent RDMA read, we
1053 * the sequence check. 1129 * have to be careful to copy the data to the right
1130 * location.
1054 */ 1131 */
1055 goto rdma_read; 1132 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1133 wqe, psn, pmtu);
1134 goto read_middle;
1056 1135
1057 case OP(RDMA_READ_RESPONSE_MIDDLE): 1136 case OP(RDMA_READ_RESPONSE_MIDDLE):
1058 /* no AETH, no ACK */ 1137 /* no AETH, no ACK */
1059 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) { 1138 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1060 dev->n_rdma_seq++; 1139 dev->n_rdma_seq++;
1061 if (qp->s_last != qp->s_tail) 1140 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; 1141 goto ack_done;
1064 } 1142 }
1065 rdma_read: 1143 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1066 if (unlikely(qp->s_state != OP(RDMA_READ_REQUEST)))
1067 goto ack_done; 1144 goto ack_done;
1145 read_middle:
1068 if (unlikely(tlen != (hdrsize + pmtu + 4))) 1146 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1069 goto ack_done; 1147 goto ack_done;
1070 if (unlikely(pmtu >= qp->s_len)) 1148 if (unlikely(pmtu >= qp->s_rdma_read_len))
1071 goto ack_done; 1149 goto ack_done;
1150
1072 /* We got a response so update the timeout. */ 1151 /* 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); 1152 spin_lock(&dev->pending_lock);
1078 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait)) 1153 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1079 list_move_tail(&qp->timerwait, 1154 list_move_tail(&qp->timerwait,
@@ -1082,27 +1157,41 @@ static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1082 /* 1157 /*
1083 * Update the RDMA receive state but do the copy w/o 1158 * Update the RDMA receive state but do the copy w/o
1084 * holding the locks and blocking interrupts. 1159 * 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 */ 1160 */
1088 qp->s_len -= pmtu; 1161 qp->s_rdma_read_len -= pmtu;
1089 update_last_psn(qp, psn); 1162 update_last_psn(qp, psn);
1090 spin_unlock_irqrestore(&qp->s_lock, flags); 1163 spin_unlock_irqrestore(&qp->s_lock, flags);
1091 ipath_copy_sge(&qp->s_sge, data, pmtu); 1164 ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1092 goto bail; 1165 goto bail;
1093 1166
1167 case OP(RDMA_READ_RESPONSE_ONLY):
1168 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1169 dev->n_rdma_seq++;
1170 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1171 goto ack_done;
1172 }
1173 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1174 goto ack_done;
1175 /*
1176 * If this is a response to a resent RDMA read, we
1177 * have to be careful to copy the data to the right
1178 * location.
1179 * XXX should check PSN and wqe opcode first.
1180 */
1181 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1182 wqe, psn, pmtu);
1183 goto read_last;
1184
1094 case OP(RDMA_READ_RESPONSE_LAST): 1185 case OP(RDMA_READ_RESPONSE_LAST):
1095 /* ACKs READ req. */ 1186 /* ACKs READ req. */
1096 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) { 1187 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1097 dev->n_rdma_seq++; 1188 dev->n_rdma_seq++;
1098 if (qp->s_last != qp->s_tail) 1189 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; 1190 goto ack_done;
1101 } 1191 }
1102 /* FALLTHROUGH */ 1192 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1103 case OP(RDMA_READ_RESPONSE_ONLY):
1104 if (unlikely(qp->s_state != OP(RDMA_READ_REQUEST)))
1105 goto ack_done; 1193 goto ack_done;
1194 read_last:
1106 /* 1195 /*
1107 * Get the number of bytes the message was padded by. 1196 * Get the number of bytes the message was padded by.
1108 */ 1197 */
@@ -1117,7 +1206,7 @@ static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1117 goto ack_done; 1206 goto ack_done;
1118 } 1207 }
1119 tlen -= hdrsize + pad + 8; 1208 tlen -= hdrsize + pad + 8;
1120 if (unlikely(tlen != qp->s_len)) { 1209 if (unlikely(tlen != qp->s_rdma_read_len)) {
1121 /* XXX Need to generate an error CQ entry. */ 1210 /* XXX Need to generate an error CQ entry. */
1122 goto ack_done; 1211 goto ack_done;
1123 } 1212 }
@@ -1127,17 +1216,8 @@ static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1127 aeth = be32_to_cpu(((__be32 *) data)[0]); 1216 aeth = be32_to_cpu(((__be32 *) data)[0]);
1128 data += sizeof(__be32); 1217 data += sizeof(__be32);
1129 } 1218 }
1130 ipath_copy_sge(&qp->s_sge, data, tlen); 1219 ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1131 if (do_rc_ack(qp, aeth, psn, OP(RDMA_READ_RESPONSE_LAST))) { 1220 (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; 1221 goto ack_done;
1142 } 1222 }
1143 1223
@@ -1162,7 +1242,7 @@ bail:
1162 * incoming RC packet for the given QP. 1242 * incoming RC packet for the given QP.
1163 * Called at interrupt level. 1243 * Called at interrupt level.
1164 * Return 1 if no more processing is needed; otherwise return 0 to 1244 * Return 1 if no more processing is needed; otherwise return 0 to
1165 * schedule a response to be sent and the s_lock unlocked. 1245 * schedule a response to be sent.
1166 */ 1246 */
1167static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev, 1247static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1168 struct ipath_other_headers *ohdr, 1248 struct ipath_other_headers *ohdr,
@@ -1173,25 +1253,23 @@ static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1173 int diff, 1253 int diff,
1174 int header_in_data) 1254 int header_in_data)
1175{ 1255{
1176 struct ib_reth *reth; 1256 struct ipath_ack_entry *e;
1257 u8 i, prev;
1258 int old_req;
1177 1259
1178 if (diff > 0) { 1260 if (diff > 0) {
1179 /* 1261 /*
1180 * Packet sequence error. 1262 * Packet sequence error.
1181 * A NAK will ACK earlier sends and RDMA writes. 1263 * A NAK will ACK earlier sends and RDMA writes.
1182 * Don't queue the NAK if a RDMA read, atomic, or 1264 * Don't queue the NAK if we already sent one.
1183 * NAK is pending though.
1184 */ 1265 */
1185 if (qp->s_ack_state != OP(ACKNOWLEDGE) || 1266 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; 1267 qp->r_nak_state = IB_NAK_PSN_ERROR;
1191 /* Use the expected PSN. */ 1268 /* Use the expected PSN. */
1192 qp->r_ack_psn = qp->r_psn; 1269 qp->r_ack_psn = qp->r_psn;
1270 goto send_ack;
1193 } 1271 }
1194 goto send_ack; 1272 goto done;
1195 } 1273 }
1196 1274
1197 /* 1275 /*
@@ -1204,8 +1282,46 @@ static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1204 * can coalesce an outstanding duplicate ACK. We have to 1282 * can coalesce an outstanding duplicate ACK. We have to
1205 * send the earliest so that RDMA reads can be restarted at 1283 * send the earliest so that RDMA reads can be restarted at
1206 * the requester's expected PSN. 1284 * the requester's expected PSN.
1285 *
1286 * First, find where this duplicate PSN falls within the
1287 * ACKs previously sent.
1207 */ 1288 */
1208 if (opcode == OP(RDMA_READ_REQUEST)) { 1289 psn &= IPATH_PSN_MASK;
1290 e = NULL;
1291 old_req = 1;
1292 spin_lock_irq(&qp->s_lock);
1293 for (i = qp->r_head_ack_queue; ; i = prev) {
1294 if (i == qp->s_tail_ack_queue)
1295 old_req = 0;
1296 if (i)
1297 prev = i - 1;
1298 else
1299 prev = IPATH_MAX_RDMA_ATOMIC;
1300 if (prev == qp->r_head_ack_queue) {
1301 e = NULL;
1302 break;
1303 }
1304 e = &qp->s_ack_queue[prev];
1305 if (!e->opcode) {
1306 e = NULL;
1307 break;
1308 }
1309 if (ipath_cmp24(psn, e->psn) >= 0)
1310 break;
1311 }
1312 switch (opcode) {
1313 case OP(RDMA_READ_REQUEST): {
1314 struct ib_reth *reth;
1315 u32 offset;
1316 u32 len;
1317
1318 /*
1319 * If we didn't find the RDMA read request in the ack queue,
1320 * or the send tasklet is already backed up to send an
1321 * earlier entry, we can ignore this request.
1322 */
1323 if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
1324 goto unlock_done;
1209 /* RETH comes after BTH */ 1325 /* RETH comes after BTH */
1210 if (!header_in_data) 1326 if (!header_in_data)
1211 reth = &ohdr->u.rc.reth; 1327 reth = &ohdr->u.rc.reth;
@@ -1214,88 +1330,87 @@ static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1214 data += sizeof(*reth); 1330 data += sizeof(*reth);
1215 } 1331 }
1216 /* 1332 /*
1217 * If we receive a duplicate RDMA request, it means the 1333 * Address range must be a subset of the original
1218 * requester saw a sequence error and needs to restart 1334 * request and start on pmtu boundaries.
1219 * from an earlier point. We can abort the current 1335 * We reuse the old ack_queue slot since the requester
1220 * RDMA read send in that case. 1336 * should not back up and request an earlier PSN for the
1337 * same request.
1221 */ 1338 */
1222 spin_lock_irq(&qp->s_lock); 1339 offset = ((psn - e->psn) & IPATH_PSN_MASK) *
1223 if (qp->s_ack_state != OP(ACKNOWLEDGE) && 1340 ib_mtu_enum_to_int(qp->path_mtu);
1224 (qp->s_hdrwords || ipath_cmp24(psn, qp->s_ack_psn) >= 0)) { 1341 len = be32_to_cpu(reth->length);
1225 /* 1342 if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
1226 * We are already sending earlier requested data. 1343 goto unlock_done;
1227 * Don't abort it to send later out of sequence data. 1344 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); 1345 u32 rkey = be32_to_cpu(reth->rkey);
1235 u64 vaddr = be64_to_cpu(reth->vaddr); 1346 u64 vaddr = be64_to_cpu(reth->vaddr);
1236 int ok; 1347 int ok;
1237 1348
1238 /* 1349 ok = ipath_rkey_ok(qp, &e->rdma_sge,
1239 * Address range must be a subset of the original 1350 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); 1351 IB_ACCESS_REMOTE_READ);
1245 if (unlikely(!ok)) { 1352 if (unlikely(!ok))
1246 spin_unlock_irq(&qp->s_lock); 1353 goto unlock_done;
1247 goto done;
1248 }
1249 } else { 1354 } else {
1250 qp->s_rdma_sge.sg_list = NULL; 1355 e->rdma_sge.sg_list = NULL;
1251 qp->s_rdma_sge.num_sge = 0; 1356 e->rdma_sge.num_sge = 0;
1252 qp->s_rdma_sge.sge.mr = NULL; 1357 e->rdma_sge.sge.mr = NULL;
1253 qp->s_rdma_sge.sge.vaddr = NULL; 1358 e->rdma_sge.sge.vaddr = NULL;
1254 qp->s_rdma_sge.sge.length = 0; 1359 e->rdma_sge.sge.length = 0;
1255 qp->s_rdma_sge.sge.sge_length = 0; 1360 e->rdma_sge.sge.sge_length = 0;
1256 } 1361 }
1257 qp->s_ack_state = opcode; 1362 e->psn = psn;
1258 qp->s_ack_psn = psn; 1363 qp->s_ack_state = OP(ACKNOWLEDGE);
1259 spin_unlock_irq(&qp->s_lock); 1364 qp->s_tail_ack_queue = prev;
1260 tasklet_hi_schedule(&qp->s_task); 1365 break;
1261 goto send_ack;
1262 } 1366 }
1263 1367
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): 1368 case OP(COMPARE_SWAP):
1282 case OP(FETCH_ADD): 1369 case OP(FETCH_ADD): {
1283 /* 1370 /*
1284 * Check for the PSN of the last atomic operation 1371 * If we didn't find the atomic request in the ack queue
1285 * performed and resend the result if found. 1372 * or the send tasklet is already backed up to send an
1373 * earlier entry, we can ignore this request.
1286 */ 1374 */
1287 if ((psn & IPATH_PSN_MASK) != qp->r_atomic_psn) 1375 if (!e || e->opcode != (u8) opcode || old_req)
1288 goto done; 1376 goto unlock_done;
1377 qp->s_ack_state = OP(ACKNOWLEDGE);
1378 qp->s_tail_ack_queue = prev;
1379 break;
1380 }
1381
1382 default:
1383 if (old_req)
1384 goto unlock_done;
1385 /*
1386 * Resend the most recent ACK if this request is
1387 * after all the previous RDMA reads and atomics.
1388 */
1389 if (i == qp->r_head_ack_queue) {
1390 spin_unlock_irq(&qp->s_lock);
1391 qp->r_nak_state = 0;
1392 qp->r_ack_psn = qp->r_psn - 1;
1393 goto send_ack;
1394 }
1395 /*
1396 * Resend the RDMA read or atomic op which
1397 * ACKs this duplicate request.
1398 */
1399 qp->s_ack_state = OP(ACKNOWLEDGE);
1400 qp->s_tail_ack_queue = i;
1289 break; 1401 break;
1290 } 1402 }
1291 qp->r_ack_state = opcode;
1292 qp->r_nak_state = 0; 1403 qp->r_nak_state = 0;
1293 qp->r_ack_psn = psn; 1404 spin_unlock_irq(&qp->s_lock);
1294send_ack: 1405 tasklet_hi_schedule(&qp->s_task);
1295 return 0;
1296 1406
1407unlock_done:
1408 spin_unlock_irq(&qp->s_lock);
1297done: 1409done:
1298 return 1; 1410 return 1;
1411
1412send_ack:
1413 return 0;
1299} 1414}
1300 1415
1301static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err) 1416static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
@@ -1391,15 +1506,7 @@ void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1391 opcode == OP(SEND_LAST_WITH_IMMEDIATE)) 1506 opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1392 break; 1507 break;
1393 nack_inv: 1508 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); 1509 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; 1510 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1404 qp->r_ack_psn = qp->r_psn; 1511 qp->r_ack_psn = qp->r_psn;
1405 goto send_ack; 1512 goto send_ack;
@@ -1441,9 +1548,8 @@ 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 1548 * Don't queue the NAK if a RDMA read or atomic
1442 * is pending though. 1549 * is pending though.
1443 */ 1550 */
1444 if (qp->r_ack_state >= OP(COMPARE_SWAP)) 1551 if (qp->r_nak_state)
1445 goto send_ack; 1552 goto done;
1446 qp->r_ack_state = OP(SEND_ONLY);
1447 qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer; 1553 qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1448 qp->r_ack_psn = qp->r_psn; 1554 qp->r_ack_psn = qp->r_psn;
1449 goto send_ack; 1555 goto send_ack;
@@ -1567,7 +1673,19 @@ void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1567 goto rnr_nak; 1673 goto rnr_nak;
1568 goto send_last_imm; 1674 goto send_last_imm;
1569 1675
1570 case OP(RDMA_READ_REQUEST): 1676 case OP(RDMA_READ_REQUEST): {
1677 struct ipath_ack_entry *e;
1678 u32 len;
1679 u8 next;
1680
1681 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
1682 goto nack_acc;
1683 next = qp->r_head_ack_queue + 1;
1684 if (next > IPATH_MAX_RDMA_ATOMIC)
1685 next = 0;
1686 if (unlikely(next == qp->s_tail_ack_queue))
1687 goto nack_inv;
1688 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1571 /* RETH comes after BTH */ 1689 /* RETH comes after BTH */
1572 if (!header_in_data) 1690 if (!header_in_data)
1573 reth = &ohdr->u.rc.reth; 1691 reth = &ohdr->u.rc.reth;
@@ -1575,72 +1693,75 @@ void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1575 reth = (struct ib_reth *)data; 1693 reth = (struct ib_reth *)data;
1576 data += sizeof(*reth); 1694 data += sizeof(*reth);
1577 } 1695 }
1578 if (unlikely(!(qp->qp_access_flags & 1696 len = be32_to_cpu(reth->length);
1579 IB_ACCESS_REMOTE_READ))) 1697 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); 1698 u32 rkey = be32_to_cpu(reth->rkey);
1585 u64 vaddr = be64_to_cpu(reth->vaddr); 1699 u64 vaddr = be64_to_cpu(reth->vaddr);
1586 int ok; 1700 int ok;
1587 1701
1588 /* Check rkey & NAK */ 1702 /* Check rkey & NAK */
1589 ok = ipath_rkey_ok(qp, &qp->s_rdma_sge, 1703 ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1590 qp->s_rdma_len, vaddr, rkey, 1704 rkey, IB_ACCESS_REMOTE_READ);
1591 IB_ACCESS_REMOTE_READ); 1705 if (unlikely(!ok))
1592 if (unlikely(!ok)) {
1593 spin_unlock_irq(&qp->s_lock);
1594 goto nack_acc; 1706 goto nack_acc;
1595 }
1596 /* 1707 /*
1597 * Update the next expected PSN. We add 1 later 1708 * Update the next expected PSN. We add 1 later
1598 * below, so only add the remainder here. 1709 * below, so only add the remainder here.
1599 */ 1710 */
1600 if (qp->s_rdma_len > pmtu) 1711 if (len > pmtu)
1601 qp->r_psn += (qp->s_rdma_len - 1) / pmtu; 1712 qp->r_psn += (len - 1) / pmtu;
1602 } else { 1713 } else {
1603 qp->s_rdma_sge.sg_list = NULL; 1714 e->rdma_sge.sg_list = NULL;
1604 qp->s_rdma_sge.num_sge = 0; 1715 e->rdma_sge.num_sge = 0;
1605 qp->s_rdma_sge.sge.mr = NULL; 1716 e->rdma_sge.sge.mr = NULL;
1606 qp->s_rdma_sge.sge.vaddr = NULL; 1717 e->rdma_sge.sge.vaddr = NULL;
1607 qp->s_rdma_sge.sge.length = 0; 1718 e->rdma_sge.sge.length = 0;
1608 qp->s_rdma_sge.sge.sge_length = 0; 1719 e->rdma_sge.sge.sge_length = 0;
1609 } 1720 }
1721 e->opcode = opcode;
1722 e->psn = psn;
1610 /* 1723 /*
1611 * We need to increment the MSN here instead of when we 1724 * We need to increment the MSN here instead of when we
1612 * finish sending the result since a duplicate request would 1725 * finish sending the result since a duplicate request would
1613 * increment it more than once. 1726 * increment it more than once.
1614 */ 1727 */
1615 qp->r_msn++; 1728 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++; 1729 qp->r_psn++;
1622 qp->r_state = opcode; 1730 qp->r_state = opcode;
1623 qp->r_nak_state = 0; 1731 qp->r_nak_state = 0;
1732 barrier();
1733 qp->r_head_ack_queue = next;
1624 1734
1625 /* Call ipath_do_rc_send() in another thread. */ 1735 /* Call ipath_do_rc_send() in another thread. */
1626 tasklet_hi_schedule(&qp->s_task); 1736 tasklet_hi_schedule(&qp->s_task);
1627 1737
1628 goto done; 1738 goto done;
1739 }
1629 1740
1630 case OP(COMPARE_SWAP): 1741 case OP(COMPARE_SWAP):
1631 case OP(FETCH_ADD): { 1742 case OP(FETCH_ADD): {
1632 struct ib_atomic_eth *ateth; 1743 struct ib_atomic_eth *ateth;
1744 struct ipath_ack_entry *e;
1633 u64 vaddr; 1745 u64 vaddr;
1746 atomic64_t *maddr;
1634 u64 sdata; 1747 u64 sdata;
1635 u32 rkey; 1748 u32 rkey;
1749 u8 next;
1636 1750
1751 if (unlikely(!(qp->qp_access_flags &
1752 IB_ACCESS_REMOTE_ATOMIC)))
1753 goto nack_acc;
1754 next = qp->r_head_ack_queue + 1;
1755 if (next > IPATH_MAX_RDMA_ATOMIC)
1756 next = 0;
1757 if (unlikely(next == qp->s_tail_ack_queue))
1758 goto nack_inv;
1637 if (!header_in_data) 1759 if (!header_in_data)
1638 ateth = &ohdr->u.atomic_eth; 1760 ateth = &ohdr->u.atomic_eth;
1639 else { 1761 else
1640 ateth = (struct ib_atomic_eth *)data; 1762 ateth = (struct ib_atomic_eth *)data;
1641 data += sizeof(*ateth); 1763 vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1642 } 1764 be32_to_cpu(ateth->vaddr[1]);
1643 vaddr = be64_to_cpu(ateth->vaddr);
1644 if (unlikely(vaddr & (sizeof(u64) - 1))) 1765 if (unlikely(vaddr & (sizeof(u64) - 1)))
1645 goto nack_inv; 1766 goto nack_inv;
1646 rkey = be32_to_cpu(ateth->rkey); 1767 rkey = be32_to_cpu(ateth->rkey);
@@ -1649,63 +1770,50 @@ void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1649 sizeof(u64), vaddr, rkey, 1770 sizeof(u64), vaddr, rkey,
1650 IB_ACCESS_REMOTE_ATOMIC))) 1771 IB_ACCESS_REMOTE_ATOMIC)))
1651 goto nack_acc; 1772 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. */ 1773 /* Perform atomic OP and save result. */
1774 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1656 sdata = be64_to_cpu(ateth->swap_data); 1775 sdata = be64_to_cpu(ateth->swap_data);
1657 spin_lock_irq(&dev->pending_lock); 1776 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1658 qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr; 1777 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1659 if (opcode == OP(FETCH_ADD)) 1778 (u64) atomic64_add_return(sdata, maddr) - sdata :
1660 *(u64 *) qp->r_sge.sge.vaddr = 1779 (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1661 qp->r_atomic_data + sdata; 1780 be64_to_cpu(ateth->compare_data),
1662 else if (qp->r_atomic_data == 1781 sdata);
1663 be64_to_cpu(ateth->compare_data)) 1782 e->opcode = opcode;
1664 *(u64 *) qp->r_sge.sge.vaddr = sdata; 1783 e->psn = psn & IPATH_PSN_MASK;
1665 spin_unlock_irq(&dev->pending_lock);
1666 qp->r_msn++; 1784 qp->r_msn++;
1667 qp->r_atomic_psn = psn & IPATH_PSN_MASK; 1785 qp->r_psn++;
1668 psn |= 1 << 31; 1786 qp->r_state = opcode;
1669 break; 1787 qp->r_nak_state = 0;
1788 barrier();
1789 qp->r_head_ack_queue = next;
1790
1791 /* Call ipath_do_rc_send() in another thread. */
1792 tasklet_hi_schedule(&qp->s_task);
1793
1794 goto done;
1670 } 1795 }
1671 1796
1672 default: 1797 default:
1673 /* Drop packet for unknown opcodes. */ 1798 /* NAK unknown opcodes. */
1674 goto done; 1799 goto nack_inv;
1675 } 1800 }
1676 qp->r_psn++; 1801 qp->r_psn++;
1677 qp->r_state = opcode; 1802 qp->r_state = opcode;
1803 qp->r_ack_psn = psn;
1678 qp->r_nak_state = 0; 1804 qp->r_nak_state = 0;
1679 /* Send an ACK if requested or required. */ 1805 /* Send an ACK if requested or required. */
1680 if (psn & (1 << 31)) { 1806 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; 1807 goto send_ack;
1690 }
1691 goto done; 1808 goto done;
1692 1809
1693nack_acc: 1810nack_acc:
1694 /* 1811 ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
1695 * A NAK will ACK earlier sends and RDMA writes. 1812 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1696 * Don't queue the NAK if a RDMA read, atomic, or NAK 1813 qp->r_ack_psn = qp->r_psn;
1697 * is pending though. 1814
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: 1815send_ack:
1706 /* Send ACK right away unless the send tasklet has a pending ACK. */ 1816 send_rc_ack(qp);
1707 if (qp->s_ack_state == OP(ACKNOWLEDGE))
1708 send_rc_ack(qp);
1709 1817
1710done: 1818done:
1711 return; 1819 return;
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-22 20:07:28 -0500