Commit | Line | Data |
---|---|---|
6a0979df AG |
1 | /* |
2 | * Copyright (c) 2006 Oracle. All rights reserved. | |
3 | * | |
4 | * This software is available to you under a choice of one of two | |
5 | * licenses. You may choose to be licensed under the terms of the GNU | |
6 | * General Public License (GPL) Version 2, available from the file | |
7 | * COPYING in the main directory of this source tree, or the | |
8 | * OpenIB.org BSD license below: | |
9 | * | |
10 | * Redistribution and use in source and binary forms, with or | |
11 | * without modification, are permitted provided that the following | |
12 | * conditions are met: | |
13 | * | |
14 | * - Redistributions of source code must retain the above | |
15 | * copyright notice, this list of conditions and the following | |
16 | * disclaimer. | |
17 | * | |
18 | * - Redistributions in binary form must reproduce the above | |
19 | * copyright notice, this list of conditions and the following | |
20 | * disclaimer in the documentation and/or other materials | |
21 | * provided with the distribution. | |
22 | * | |
23 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, | |
24 | * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF | |
25 | * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND | |
26 | * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS | |
27 | * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN | |
28 | * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN | |
29 | * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE | |
30 | * SOFTWARE. | |
31 | * | |
32 | */ | |
33 | #include <linux/kernel.h> | |
34 | #include <linux/in.h> | |
35 | #include <linux/device.h> | |
36 | #include <linux/dmapool.h> | |
cb0a6056 | 37 | #include <linux/ratelimit.h> |
6a0979df AG |
38 | |
39 | #include "rds.h" | |
6a0979df AG |
40 | #include "ib.h" |
41 | ||
9c030391 AG |
42 | /* |
43 | * Convert IB-specific error message to RDS error message and call core | |
44 | * completion handler. | |
45 | */ | |
46 | static void rds_ib_send_complete(struct rds_message *rm, | |
47 | int wc_status, | |
48 | void (*complete)(struct rds_message *rm, int status)) | |
6a0979df AG |
49 | { |
50 | int notify_status; | |
51 | ||
52 | switch (wc_status) { | |
53 | case IB_WC_WR_FLUSH_ERR: | |
54 | return; | |
55 | ||
56 | case IB_WC_SUCCESS: | |
57 | notify_status = RDS_RDMA_SUCCESS; | |
58 | break; | |
59 | ||
60 | case IB_WC_REM_ACCESS_ERR: | |
61 | notify_status = RDS_RDMA_REMOTE_ERROR; | |
62 | break; | |
63 | ||
64 | default: | |
65 | notify_status = RDS_RDMA_OTHER_ERROR; | |
66 | break; | |
67 | } | |
9c030391 | 68 | complete(rm, notify_status); |
6a0979df AG |
69 | } |
70 | ||
ff3d7d36 AG |
71 | static void rds_ib_send_unmap_data(struct rds_ib_connection *ic, |
72 | struct rm_data_op *op, | |
73 | int wc_status) | |
6a0979df | 74 | { |
ff3d7d36 AG |
75 | if (op->op_nents) |
76 | ib_dma_unmap_sg(ic->i_cm_id->device, | |
77 | op->op_sg, op->op_nents, | |
78 | DMA_TO_DEVICE); | |
79 | } | |
6a0979df | 80 | |
ff3d7d36 AG |
81 | static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic, |
82 | struct rm_rdma_op *op, | |
83 | int wc_status) | |
84 | { | |
85 | if (op->op_mapped) { | |
86 | ib_dma_unmap_sg(ic->i_cm_id->device, | |
87 | op->op_sg, op->op_nents, | |
88 | op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE); | |
89 | op->op_mapped = 0; | |
90 | } | |
15133f6e | 91 | |
ff3d7d36 AG |
92 | /* If the user asked for a completion notification on this |
93 | * message, we can implement three different semantics: | |
94 | * 1. Notify when we received the ACK on the RDS message | |
95 | * that was queued with the RDMA. This provides reliable | |
96 | * notification of RDMA status at the expense of a one-way | |
97 | * packet delay. | |
98 | * 2. Notify when the IB stack gives us the completion event for | |
99 | * the RDMA operation. | |
100 | * 3. Notify when the IB stack gives us the completion event for | |
101 | * the accompanying RDS messages. | |
102 | * Here, we implement approach #3. To implement approach #2, | |
103 | * we would need to take an event for the rdma WR. To implement #1, | |
104 | * don't call rds_rdma_send_complete at all, and fall back to the notify | |
105 | * handling in the ACK processing code. | |
106 | * | |
107 | * Note: There's no need to explicitly sync any RDMA buffers using | |
108 | * ib_dma_sync_sg_for_cpu - the completion for the RDMA | |
109 | * operation itself unmapped the RDMA buffers, which takes care | |
110 | * of synching. | |
111 | */ | |
112 | rds_ib_send_complete(container_of(op, struct rds_message, rdma), | |
113 | wc_status, rds_rdma_send_complete); | |
6a0979df | 114 | |
ff3d7d36 AG |
115 | if (op->op_write) |
116 | rds_stats_add(s_send_rdma_bytes, op->op_bytes); | |
117 | else | |
118 | rds_stats_add(s_recv_rdma_bytes, op->op_bytes); | |
119 | } | |
6a0979df | 120 | |
ff3d7d36 AG |
121 | static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic, |
122 | struct rm_atomic_op *op, | |
123 | int wc_status) | |
124 | { | |
125 | /* unmap atomic recvbuf */ | |
126 | if (op->op_mapped) { | |
127 | ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1, | |
128 | DMA_FROM_DEVICE); | |
129 | op->op_mapped = 0; | |
6a0979df AG |
130 | } |
131 | ||
ff3d7d36 AG |
132 | rds_ib_send_complete(container_of(op, struct rds_message, atomic), |
133 | wc_status, rds_atomic_send_complete); | |
15133f6e | 134 | |
ff3d7d36 | 135 | if (op->op_type == RDS_ATOMIC_TYPE_CSWP) |
51e2cba8 | 136 | rds_ib_stats_inc(s_ib_atomic_cswp); |
ff3d7d36 | 137 | else |
51e2cba8 | 138 | rds_ib_stats_inc(s_ib_atomic_fadd); |
ff3d7d36 | 139 | } |
15133f6e | 140 | |
ff3d7d36 AG |
141 | /* |
142 | * Unmap the resources associated with a struct send_work. | |
143 | * | |
144 | * Returns the rm for no good reason other than it is unobtainable | |
145 | * other than by switching on wr.opcode, currently, and the caller, | |
146 | * the event handler, needs it. | |
147 | */ | |
148 | static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic, | |
149 | struct rds_ib_send_work *send, | |
150 | int wc_status) | |
151 | { | |
152 | struct rds_message *rm = NULL; | |
153 | ||
154 | /* In the error case, wc.opcode sometimes contains garbage */ | |
155 | switch (send->s_wr.opcode) { | |
156 | case IB_WR_SEND: | |
157 | if (send->s_op) { | |
158 | rm = container_of(send->s_op, struct rds_message, data); | |
159 | rds_ib_send_unmap_data(ic, send->s_op, wc_status); | |
160 | } | |
161 | break; | |
162 | case IB_WR_RDMA_WRITE: | |
163 | case IB_WR_RDMA_READ: | |
164 | if (send->s_op) { | |
165 | rm = container_of(send->s_op, struct rds_message, rdma); | |
166 | rds_ib_send_unmap_rdma(ic, send->s_op, wc_status); | |
167 | } | |
168 | break; | |
169 | case IB_WR_ATOMIC_FETCH_AND_ADD: | |
170 | case IB_WR_ATOMIC_CMP_AND_SWP: | |
171 | if (send->s_op) { | |
172 | rm = container_of(send->s_op, struct rds_message, atomic); | |
173 | rds_ib_send_unmap_atomic(ic, send->s_op, wc_status); | |
174 | } | |
175 | break; | |
176 | default: | |
cb0a6056 | 177 | printk_ratelimited(KERN_NOTICE |
ff3d7d36 AG |
178 | "RDS/IB: %s: unexpected opcode 0x%x in WR!\n", |
179 | __func__, send->s_wr.opcode); | |
180 | break; | |
15133f6e AG |
181 | } |
182 | ||
ff3d7d36 | 183 | send->s_wr.opcode = 0xdead; |
6a0979df | 184 | |
ff3d7d36 | 185 | return rm; |
6a0979df AG |
186 | } |
187 | ||
188 | void rds_ib_send_init_ring(struct rds_ib_connection *ic) | |
189 | { | |
190 | struct rds_ib_send_work *send; | |
191 | u32 i; | |
192 | ||
193 | for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) { | |
194 | struct ib_sge *sge; | |
195 | ||
6a0979df AG |
196 | send->s_op = NULL; |
197 | ||
0c28c045 | 198 | send->s_wr.wr_id = i | RDS_IB_SEND_OP; |
6a0979df | 199 | send->s_wr.sg_list = send->s_sge; |
6a0979df AG |
200 | send->s_wr.ex.imm_data = 0; |
201 | ||
919ced4c | 202 | sge = &send->s_sge[0]; |
6a0979df AG |
203 | sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header)); |
204 | sge->length = sizeof(struct rds_header); | |
e5580242 | 205 | sge->lkey = ic->i_pd->local_dma_lkey; |
919ced4c | 206 | |
e5580242 | 207 | send->s_sge[1].lkey = ic->i_pd->local_dma_lkey; |
6a0979df AG |
208 | } |
209 | } | |
210 | ||
211 | void rds_ib_send_clear_ring(struct rds_ib_connection *ic) | |
212 | { | |
213 | struct rds_ib_send_work *send; | |
214 | u32 i; | |
215 | ||
216 | for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) { | |
ff3d7d36 AG |
217 | if (send->s_op && send->s_wr.opcode != 0xdead) |
218 | rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR); | |
6a0979df AG |
219 | } |
220 | } | |
221 | ||
f046011c ZB |
222 | /* |
223 | * The only fast path caller always has a non-zero nr, so we don't | |
224 | * bother testing nr before performing the atomic sub. | |
225 | */ | |
226 | static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr) | |
227 | { | |
228 | if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) && | |
229 | waitqueue_active(&rds_ib_ring_empty_wait)) | |
230 | wake_up(&rds_ib_ring_empty_wait); | |
231 | BUG_ON(atomic_read(&ic->i_signaled_sends) < 0); | |
232 | } | |
233 | ||
6a0979df AG |
234 | /* |
235 | * The _oldest/_free ring operations here race cleanly with the alloc/unalloc | |
236 | * operations performed in the send path. As the sender allocs and potentially | |
237 | * unallocs the next free entry in the ring it doesn't alter which is | |
238 | * the next to be freed, which is what this is concerned with. | |
239 | */ | |
0c28c045 | 240 | void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc) |
6a0979df | 241 | { |
ff3d7d36 | 242 | struct rds_message *rm = NULL; |
0c28c045 | 243 | struct rds_connection *conn = ic->conn; |
6a0979df AG |
244 | struct rds_ib_send_work *send; |
245 | u32 completed; | |
246 | u32 oldest; | |
247 | u32 i = 0; | |
f046011c | 248 | int nr_sig = 0; |
6a0979df | 249 | |
6a0979df | 250 | |
0c28c045 SS |
251 | rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n", |
252 | (unsigned long long)wc->wr_id, wc->status, | |
253 | ib_wc_status_msg(wc->status), wc->byte_len, | |
254 | be32_to_cpu(wc->ex.imm_data)); | |
255 | rds_ib_stats_inc(s_ib_tx_cq_event); | |
6a0979df | 256 | |
0c28c045 SS |
257 | if (wc->wr_id == RDS_IB_ACK_WR_ID) { |
258 | if (time_after(jiffies, ic->i_ack_queued + HZ / 2)) | |
259 | rds_ib_stats_inc(s_ib_tx_stalled); | |
260 | rds_ib_ack_send_complete(ic); | |
261 | return; | |
262 | } | |
6a0979df | 263 | |
0c28c045 | 264 | oldest = rds_ib_ring_oldest(&ic->i_send_ring); |
6a0979df | 265 | |
0c28c045 SS |
266 | completed = rds_ib_ring_completed(&ic->i_send_ring, |
267 | (wc->wr_id & ~RDS_IB_SEND_OP), | |
268 | oldest); | |
6a0979df | 269 | |
0c28c045 SS |
270 | for (i = 0; i < completed; i++) { |
271 | send = &ic->i_sends[oldest]; | |
272 | if (send->s_wr.send_flags & IB_SEND_SIGNALED) | |
273 | nr_sig++; | |
6a0979df | 274 | |
0c28c045 SS |
275 | rm = rds_ib_send_unmap_op(ic, send, wc->status); |
276 | ||
277 | if (time_after(jiffies, send->s_queued + HZ / 2)) | |
278 | rds_ib_stats_inc(s_ib_tx_stalled); | |
6a0979df | 279 | |
0c28c045 SS |
280 | if (send->s_op) { |
281 | if (send->s_op == rm->m_final_op) { | |
282 | /* If anyone waited for this message to get | |
283 | * flushed out, wake them up now | |
284 | */ | |
285 | rds_message_unmapped(rm); | |
286 | } | |
287 | rds_message_put(rm); | |
288 | send->s_op = NULL; | |
6a0979df AG |
289 | } |
290 | ||
0c28c045 SS |
291 | oldest = (oldest + 1) % ic->i_send_ring.w_nr; |
292 | } | |
6a0979df | 293 | |
0c28c045 SS |
294 | rds_ib_ring_free(&ic->i_send_ring, completed); |
295 | rds_ib_sub_signaled(ic, nr_sig); | |
296 | nr_sig = 0; | |
6a0979df | 297 | |
0c28c045 SS |
298 | if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) || |
299 | test_bit(0, &conn->c_map_queued)) | |
300 | queue_delayed_work(rds_wq, &conn->c_send_w, 0); | |
301 | ||
302 | /* We expect errors as the qp is drained during shutdown */ | |
303 | if (wc->status != IB_WC_SUCCESS && rds_conn_up(conn)) { | |
304 | rds_ib_conn_error(conn, "send completion on %pI4 had status %u (%s), disconnecting and reconnecting\n", | |
305 | &conn->c_faddr, wc->status, | |
306 | ib_wc_status_msg(wc->status)); | |
6a0979df AG |
307 | } |
308 | } | |
309 | ||
310 | /* | |
311 | * This is the main function for allocating credits when sending | |
312 | * messages. | |
313 | * | |
314 | * Conceptually, we have two counters: | |
315 | * - send credits: this tells us how many WRs we're allowed | |
25985edc | 316 | * to submit without overruning the receiver's queue. For |
6a0979df AG |
317 | * each SEND WR we post, we decrement this by one. |
318 | * | |
319 | * - posted credits: this tells us how many WRs we recently | |
320 | * posted to the receive queue. This value is transferred | |
321 | * to the peer as a "credit update" in a RDS header field. | |
322 | * Every time we transmit credits to the peer, we subtract | |
323 | * the amount of transferred credits from this counter. | |
324 | * | |
325 | * It is essential that we avoid situations where both sides have | |
326 | * exhausted their send credits, and are unable to send new credits | |
327 | * to the peer. We achieve this by requiring that we send at least | |
328 | * one credit update to the peer before exhausting our credits. | |
329 | * When new credits arrive, we subtract one credit that is withheld | |
330 | * until we've posted new buffers and are ready to transmit these | |
331 | * credits (see rds_ib_send_add_credits below). | |
332 | * | |
333 | * The RDS send code is essentially single-threaded; rds_send_xmit | |
0f4b1c7e | 334 | * sets RDS_IN_XMIT to ensure exclusive access to the send ring. |
6a0979df AG |
335 | * However, the ACK sending code is independent and can race with |
336 | * message SENDs. | |
337 | * | |
338 | * In the send path, we need to update the counters for send credits | |
339 | * and the counter of posted buffers atomically - when we use the | |
340 | * last available credit, we cannot allow another thread to race us | |
341 | * and grab the posted credits counter. Hence, we have to use a | |
342 | * spinlock to protect the credit counter, or use atomics. | |
343 | * | |
344 | * Spinlocks shared between the send and the receive path are bad, | |
345 | * because they create unnecessary delays. An early implementation | |
346 | * using a spinlock showed a 5% degradation in throughput at some | |
347 | * loads. | |
348 | * | |
349 | * This implementation avoids spinlocks completely, putting both | |
350 | * counters into a single atomic, and updating that atomic using | |
351 | * atomic_add (in the receive path, when receiving fresh credits), | |
352 | * and using atomic_cmpxchg when updating the two counters. | |
353 | */ | |
354 | int rds_ib_send_grab_credits(struct rds_ib_connection *ic, | |
7b70d033 | 355 | u32 wanted, u32 *adv_credits, int need_posted, int max_posted) |
6a0979df AG |
356 | { |
357 | unsigned int avail, posted, got = 0, advertise; | |
358 | long oldval, newval; | |
359 | ||
360 | *adv_credits = 0; | |
361 | if (!ic->i_flowctl) | |
362 | return wanted; | |
363 | ||
364 | try_again: | |
365 | advertise = 0; | |
366 | oldval = newval = atomic_read(&ic->i_credits); | |
367 | posted = IB_GET_POST_CREDITS(oldval); | |
368 | avail = IB_GET_SEND_CREDITS(oldval); | |
369 | ||
11ac1199 | 370 | rdsdebug("wanted=%u credits=%u posted=%u\n", |
6a0979df AG |
371 | wanted, avail, posted); |
372 | ||
373 | /* The last credit must be used to send a credit update. */ | |
374 | if (avail && !posted) | |
375 | avail--; | |
376 | ||
377 | if (avail < wanted) { | |
378 | struct rds_connection *conn = ic->i_cm_id->context; | |
379 | ||
380 | /* Oops, there aren't that many credits left! */ | |
381 | set_bit(RDS_LL_SEND_FULL, &conn->c_flags); | |
382 | got = avail; | |
383 | } else { | |
384 | /* Sometimes you get what you want, lalala. */ | |
385 | got = wanted; | |
386 | } | |
387 | newval -= IB_SET_SEND_CREDITS(got); | |
388 | ||
389 | /* | |
390 | * If need_posted is non-zero, then the caller wants | |
391 | * the posted regardless of whether any send credits are | |
392 | * available. | |
393 | */ | |
394 | if (posted && (got || need_posted)) { | |
7b70d033 | 395 | advertise = min_t(unsigned int, posted, max_posted); |
6a0979df AG |
396 | newval -= IB_SET_POST_CREDITS(advertise); |
397 | } | |
398 | ||
399 | /* Finally bill everything */ | |
400 | if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval) | |
401 | goto try_again; | |
402 | ||
403 | *adv_credits = advertise; | |
404 | return got; | |
405 | } | |
406 | ||
407 | void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits) | |
408 | { | |
409 | struct rds_ib_connection *ic = conn->c_transport_data; | |
410 | ||
411 | if (credits == 0) | |
412 | return; | |
413 | ||
11ac1199 | 414 | rdsdebug("credits=%u current=%u%s\n", |
6a0979df AG |
415 | credits, |
416 | IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)), | |
417 | test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : ""); | |
418 | ||
419 | atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits); | |
420 | if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags)) | |
421 | queue_delayed_work(rds_wq, &conn->c_send_w, 0); | |
422 | ||
423 | WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384); | |
424 | ||
425 | rds_ib_stats_inc(s_ib_rx_credit_updates); | |
426 | } | |
427 | ||
428 | void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted) | |
429 | { | |
430 | struct rds_ib_connection *ic = conn->c_transport_data; | |
431 | ||
432 | if (posted == 0) | |
433 | return; | |
434 | ||
435 | atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits); | |
436 | ||
437 | /* Decide whether to send an update to the peer now. | |
438 | * If we would send a credit update for every single buffer we | |
439 | * post, we would end up with an ACK storm (ACK arrives, | |
440 | * consumes buffer, we refill the ring, send ACK to remote | |
441 | * advertising the newly posted buffer... ad inf) | |
442 | * | |
443 | * Performance pretty much depends on how often we send | |
444 | * credit updates - too frequent updates mean lots of ACKs. | |
445 | * Too infrequent updates, and the peer will run out of | |
446 | * credits and has to throttle. | |
447 | * For the time being, 16 seems to be a good compromise. | |
448 | */ | |
449 | if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16) | |
450 | set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); | |
451 | } | |
452 | ||
f046011c ZB |
453 | static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic, |
454 | struct rds_ib_send_work *send, | |
455 | bool notify) | |
241eef3e AG |
456 | { |
457 | /* | |
458 | * We want to delay signaling completions just enough to get | |
459 | * the batching benefits but not so much that we create dead time | |
460 | * on the wire. | |
461 | */ | |
462 | if (ic->i_unsignaled_wrs-- == 0 || notify) { | |
463 | ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs; | |
464 | send->s_wr.send_flags |= IB_SEND_SIGNALED; | |
f046011c | 465 | return 1; |
241eef3e | 466 | } |
f046011c | 467 | return 0; |
241eef3e AG |
468 | } |
469 | ||
6a0979df AG |
470 | /* |
471 | * This can be called multiple times for a given message. The first time | |
472 | * we see a message we map its scatterlist into the IB device so that | |
473 | * we can provide that mapped address to the IB scatter gather entries | |
474 | * in the IB work requests. We translate the scatterlist into a series | |
475 | * of work requests that fragment the message. These work requests complete | |
476 | * in order so we pass ownership of the message to the completion handler | |
477 | * once we send the final fragment. | |
478 | * | |
479 | * The RDS core uses the c_send_lock to only enter this function once | |
480 | * per connection. This makes sure that the tx ring alloc/unalloc pairs | |
481 | * don't get out of sync and confuse the ring. | |
482 | */ | |
483 | int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm, | |
484 | unsigned int hdr_off, unsigned int sg, unsigned int off) | |
485 | { | |
486 | struct rds_ib_connection *ic = conn->c_transport_data; | |
487 | struct ib_device *dev = ic->i_cm_id->device; | |
488 | struct rds_ib_send_work *send = NULL; | |
489 | struct rds_ib_send_work *first; | |
490 | struct rds_ib_send_work *prev; | |
491 | struct ib_send_wr *failed_wr; | |
492 | struct scatterlist *scat; | |
493 | u32 pos; | |
494 | u32 i; | |
495 | u32 work_alloc; | |
da5a06ce | 496 | u32 credit_alloc = 0; |
6a0979df AG |
497 | u32 posted; |
498 | u32 adv_credits = 0; | |
499 | int send_flags = 0; | |
da5a06ce | 500 | int bytes_sent = 0; |
6a0979df AG |
501 | int ret; |
502 | int flow_controlled = 0; | |
f046011c | 503 | int nr_sig = 0; |
6a0979df AG |
504 | |
505 | BUG_ON(off % RDS_FRAG_SIZE); | |
506 | BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header)); | |
507 | ||
2e7b3b99 AG |
508 | /* Do not send cong updates to IB loopback */ |
509 | if (conn->c_loopback | |
510 | && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) { | |
511 | rds_cong_map_updated(conn->c_fcong, ~(u64) 0); | |
6094628b | 512 | scat = &rm->data.op_sg[sg]; |
18fc25c9 VV |
513 | ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length); |
514 | return sizeof(struct rds_header) + ret; | |
2e7b3b99 AG |
515 | } |
516 | ||
6a0979df AG |
517 | /* FIXME we may overallocate here */ |
518 | if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) | |
519 | i = 1; | |
520 | else | |
521 | i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE); | |
522 | ||
523 | work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos); | |
524 | if (work_alloc == 0) { | |
525 | set_bit(RDS_LL_SEND_FULL, &conn->c_flags); | |
526 | rds_ib_stats_inc(s_ib_tx_ring_full); | |
527 | ret = -ENOMEM; | |
528 | goto out; | |
529 | } | |
530 | ||
6a0979df | 531 | if (ic->i_flowctl) { |
7b70d033 | 532 | credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT); |
6a0979df AG |
533 | adv_credits += posted; |
534 | if (credit_alloc < work_alloc) { | |
535 | rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc); | |
536 | work_alloc = credit_alloc; | |
c8de3f10 | 537 | flow_controlled = 1; |
6a0979df AG |
538 | } |
539 | if (work_alloc == 0) { | |
d39e0602 | 540 | set_bit(RDS_LL_SEND_FULL, &conn->c_flags); |
6a0979df AG |
541 | rds_ib_stats_inc(s_ib_tx_throttle); |
542 | ret = -ENOMEM; | |
543 | goto out; | |
544 | } | |
545 | } | |
546 | ||
547 | /* map the message the first time we see it */ | |
ff3d7d36 | 548 | if (!ic->i_data_op) { |
6c7cc6e4 AG |
549 | if (rm->data.op_nents) { |
550 | rm->data.op_count = ib_dma_map_sg(dev, | |
551 | rm->data.op_sg, | |
552 | rm->data.op_nents, | |
553 | DMA_TO_DEVICE); | |
554 | rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count); | |
555 | if (rm->data.op_count == 0) { | |
6a0979df AG |
556 | rds_ib_stats_inc(s_ib_tx_sg_mapping_failure); |
557 | rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); | |
558 | ret = -ENOMEM; /* XXX ? */ | |
559 | goto out; | |
560 | } | |
561 | } else { | |
6c7cc6e4 | 562 | rm->data.op_count = 0; |
6a0979df AG |
563 | } |
564 | ||
6a0979df | 565 | rds_message_addref(rm); |
d655a9fb WW |
566 | rm->data.op_dmasg = 0; |
567 | rm->data.op_dmaoff = 0; | |
ff3d7d36 | 568 | ic->i_data_op = &rm->data; |
6a0979df AG |
569 | |
570 | /* Finalize the header */ | |
571 | if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags)) | |
572 | rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED; | |
573 | if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) | |
574 | rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED; | |
575 | ||
576 | /* If it has a RDMA op, tell the peer we did it. This is | |
577 | * used by the peer to release use-once RDMA MRs. */ | |
f8b3aaf2 | 578 | if (rm->rdma.op_active) { |
6a0979df AG |
579 | struct rds_ext_header_rdma ext_hdr; |
580 | ||
f8b3aaf2 | 581 | ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey); |
6a0979df AG |
582 | rds_message_add_extension(&rm->m_inc.i_hdr, |
583 | RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr)); | |
584 | } | |
585 | if (rm->m_rdma_cookie) { | |
586 | rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr, | |
587 | rds_rdma_cookie_key(rm->m_rdma_cookie), | |
588 | rds_rdma_cookie_offset(rm->m_rdma_cookie)); | |
589 | } | |
590 | ||
591 | /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so | |
592 | * we should not do this unless we have a chance of at least | |
593 | * sticking the header into the send ring. Which is why we | |
594 | * should call rds_ib_ring_alloc first. */ | |
595 | rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic)); | |
596 | rds_message_make_checksum(&rm->m_inc.i_hdr); | |
597 | ||
598 | /* | |
599 | * Update adv_credits since we reset the ACK_REQUIRED bit. | |
600 | */ | |
c8de3f10 AG |
601 | if (ic->i_flowctl) { |
602 | rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits); | |
603 | adv_credits += posted; | |
604 | BUG_ON(adv_credits > 255); | |
605 | } | |
735f61e6 | 606 | } |
6a0979df | 607 | |
6a0979df AG |
608 | /* Sometimes you want to put a fence between an RDMA |
609 | * READ and the following SEND. | |
610 | * We could either do this all the time | |
611 | * or when requested by the user. Right now, we let | |
612 | * the application choose. | |
613 | */ | |
f8b3aaf2 | 614 | if (rm->rdma.op_active && rm->rdma.op_fence) |
6a0979df AG |
615 | send_flags = IB_SEND_FENCE; |
616 | ||
da5a06ce AG |
617 | /* Each frag gets a header. Msgs may be 0 bytes */ |
618 | send = &ic->i_sends[pos]; | |
619 | first = send; | |
620 | prev = NULL; | |
d655a9fb | 621 | scat = &ic->i_data_op->op_sg[rm->data.op_dmasg]; |
da5a06ce AG |
622 | i = 0; |
623 | do { | |
624 | unsigned int len = 0; | |
625 | ||
626 | /* Set up the header */ | |
627 | send->s_wr.send_flags = send_flags; | |
628 | send->s_wr.opcode = IB_WR_SEND; | |
629 | send->s_wr.num_sge = 1; | |
630 | send->s_wr.next = NULL; | |
631 | send->s_queued = jiffies; | |
632 | send->s_op = NULL; | |
6a0979df | 633 | |
da5a06ce AG |
634 | send->s_sge[0].addr = ic->i_send_hdrs_dma |
635 | + (pos * sizeof(struct rds_header)); | |
636 | send->s_sge[0].length = sizeof(struct rds_header); | |
637 | ||
638 | memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header)); | |
6a0979df | 639 | |
da5a06ce AG |
640 | /* Set up the data, if present */ |
641 | if (i < work_alloc | |
6c7cc6e4 | 642 | && scat != &rm->data.op_sg[rm->data.op_count]) { |
d655a9fb WW |
643 | len = min(RDS_FRAG_SIZE, |
644 | ib_sg_dma_len(dev, scat) - rm->data.op_dmaoff); | |
da5a06ce | 645 | send->s_wr.num_sge = 2; |
6a0979df | 646 | |
d655a9fb WW |
647 | send->s_sge[1].addr = ib_sg_dma_address(dev, scat); |
648 | send->s_sge[1].addr += rm->data.op_dmaoff; | |
da5a06ce | 649 | send->s_sge[1].length = len; |
6a0979df | 650 | |
da5a06ce | 651 | bytes_sent += len; |
d655a9fb WW |
652 | rm->data.op_dmaoff += len; |
653 | if (rm->data.op_dmaoff == ib_sg_dma_len(dev, scat)) { | |
da5a06ce | 654 | scat++; |
d655a9fb WW |
655 | rm->data.op_dmasg++; |
656 | rm->data.op_dmaoff = 0; | |
da5a06ce AG |
657 | } |
658 | } | |
6a0979df | 659 | |
241eef3e | 660 | rds_ib_set_wr_signal_state(ic, send, 0); |
6a0979df | 661 | |
6a0979df AG |
662 | /* |
663 | * Always signal the last one if we're stopping due to flow control. | |
664 | */ | |
c8de3f10 | 665 | if (ic->i_flowctl && flow_controlled && i == (work_alloc-1)) |
6a0979df AG |
666 | send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED; |
667 | ||
f046011c ZB |
668 | if (send->s_wr.send_flags & IB_SEND_SIGNALED) |
669 | nr_sig++; | |
670 | ||
6a0979df AG |
671 | rdsdebug("send %p wr %p num_sge %u next %p\n", send, |
672 | &send->s_wr, send->s_wr.num_sge, send->s_wr.next); | |
673 | ||
c8de3f10 | 674 | if (ic->i_flowctl && adv_credits) { |
6a0979df AG |
675 | struct rds_header *hdr = &ic->i_send_hdrs[pos]; |
676 | ||
677 | /* add credit and redo the header checksum */ | |
678 | hdr->h_credit = adv_credits; | |
679 | rds_message_make_checksum(hdr); | |
680 | adv_credits = 0; | |
681 | rds_ib_stats_inc(s_ib_tx_credit_updates); | |
682 | } | |
683 | ||
684 | if (prev) | |
685 | prev->s_wr.next = &send->s_wr; | |
686 | prev = send; | |
687 | ||
688 | pos = (pos + 1) % ic->i_send_ring.w_nr; | |
da5a06ce AG |
689 | send = &ic->i_sends[pos]; |
690 | i++; | |
691 | ||
692 | } while (i < work_alloc | |
6c7cc6e4 | 693 | && scat != &rm->data.op_sg[rm->data.op_count]); |
6a0979df AG |
694 | |
695 | /* Account the RDS header in the number of bytes we sent, but just once. | |
696 | * The caller has no concept of fragmentation. */ | |
697 | if (hdr_off == 0) | |
da5a06ce | 698 | bytes_sent += sizeof(struct rds_header); |
6a0979df AG |
699 | |
700 | /* if we finished the message then send completion owns it */ | |
6c7cc6e4 | 701 | if (scat == &rm->data.op_sg[rm->data.op_count]) { |
ff3d7d36 | 702 | prev->s_op = ic->i_data_op; |
241eef3e | 703 | prev->s_wr.send_flags |= IB_SEND_SOLICITED; |
3049147c | 704 | if (!(prev->s_wr.send_flags & IB_SEND_SIGNALED)) { |
705 | ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs; | |
706 | prev->s_wr.send_flags |= IB_SEND_SIGNALED; | |
707 | nr_sig++; | |
708 | } | |
ff3d7d36 | 709 | ic->i_data_op = NULL; |
6a0979df AG |
710 | } |
711 | ||
da5a06ce | 712 | /* Put back wrs & credits we didn't use */ |
6a0979df AG |
713 | if (i < work_alloc) { |
714 | rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i); | |
715 | work_alloc = i; | |
716 | } | |
717 | if (ic->i_flowctl && i < credit_alloc) | |
718 | rds_ib_send_add_credits(conn, credit_alloc - i); | |
719 | ||
f046011c ZB |
720 | if (nr_sig) |
721 | atomic_add(nr_sig, &ic->i_signaled_sends); | |
722 | ||
6a0979df AG |
723 | /* XXX need to worry about failed_wr and partial sends. */ |
724 | failed_wr = &first->s_wr; | |
725 | ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr); | |
726 | rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic, | |
727 | first, &first->s_wr, ret, failed_wr); | |
728 | BUG_ON(failed_wr != &first->s_wr); | |
729 | if (ret) { | |
730 | printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 " | |
731 | "returned %d\n", &conn->c_faddr, ret); | |
732 | rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); | |
f046011c | 733 | rds_ib_sub_signaled(ic, nr_sig); |
ff3d7d36 AG |
734 | if (prev->s_op) { |
735 | ic->i_data_op = prev->s_op; | |
736 | prev->s_op = NULL; | |
6a0979df | 737 | } |
735f61e6 AG |
738 | |
739 | rds_ib_conn_error(ic->conn, "ib_post_send failed\n"); | |
6a0979df AG |
740 | goto out; |
741 | } | |
742 | ||
da5a06ce | 743 | ret = bytes_sent; |
6a0979df AG |
744 | out: |
745 | BUG_ON(adv_credits); | |
746 | return ret; | |
747 | } | |
748 | ||
15133f6e AG |
749 | /* |
750 | * Issue atomic operation. | |
751 | * A simplified version of the rdma case, we always map 1 SG, and | |
752 | * only 8 bytes, for the return value from the atomic operation. | |
753 | */ | |
ff3d7d36 | 754 | int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op) |
15133f6e AG |
755 | { |
756 | struct rds_ib_connection *ic = conn->c_transport_data; | |
757 | struct rds_ib_send_work *send = NULL; | |
758 | struct ib_send_wr *failed_wr; | |
759 | struct rds_ib_device *rds_ibdev; | |
760 | u32 pos; | |
761 | u32 work_alloc; | |
762 | int ret; | |
f046011c | 763 | int nr_sig = 0; |
15133f6e AG |
764 | |
765 | rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client); | |
766 | ||
767 | work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos); | |
768 | if (work_alloc != 1) { | |
769 | rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); | |
770 | rds_ib_stats_inc(s_ib_tx_ring_full); | |
771 | ret = -ENOMEM; | |
772 | goto out; | |
773 | } | |
774 | ||
775 | /* address of send request in ring */ | |
776 | send = &ic->i_sends[pos]; | |
777 | send->s_queued = jiffies; | |
778 | ||
779 | if (op->op_type == RDS_ATOMIC_TYPE_CSWP) { | |
e622f2f4 CH |
780 | send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP; |
781 | send->s_atomic_wr.compare_add = op->op_m_cswp.compare; | |
782 | send->s_atomic_wr.swap = op->op_m_cswp.swap; | |
783 | send->s_atomic_wr.compare_add_mask = op->op_m_cswp.compare_mask; | |
784 | send->s_atomic_wr.swap_mask = op->op_m_cswp.swap_mask; | |
15133f6e | 785 | } else { /* FADD */ |
e622f2f4 CH |
786 | send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD; |
787 | send->s_atomic_wr.compare_add = op->op_m_fadd.add; | |
788 | send->s_atomic_wr.swap = 0; | |
789 | send->s_atomic_wr.compare_add_mask = op->op_m_fadd.nocarry_mask; | |
790 | send->s_atomic_wr.swap_mask = 0; | |
15133f6e | 791 | } |
f046011c | 792 | nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify); |
e622f2f4 CH |
793 | send->s_atomic_wr.wr.num_sge = 1; |
794 | send->s_atomic_wr.wr.next = NULL; | |
795 | send->s_atomic_wr.remote_addr = op->op_remote_addr; | |
796 | send->s_atomic_wr.rkey = op->op_rkey; | |
1cc2228c CM |
797 | send->s_op = op; |
798 | rds_message_addref(container_of(send->s_op, struct rds_message, atomic)); | |
15133f6e AG |
799 | |
800 | /* map 8 byte retval buffer to the device */ | |
801 | ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE); | |
802 | rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret); | |
803 | if (ret != 1) { | |
804 | rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); | |
805 | rds_ib_stats_inc(s_ib_tx_sg_mapping_failure); | |
806 | ret = -ENOMEM; /* XXX ? */ | |
807 | goto out; | |
808 | } | |
809 | ||
810 | /* Convert our struct scatterlist to struct ib_sge */ | |
811 | send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg); | |
812 | send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg); | |
e5580242 | 813 | send->s_sge[0].lkey = ic->i_pd->local_dma_lkey; |
15133f6e AG |
814 | |
815 | rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr, | |
816 | send->s_sge[0].addr, send->s_sge[0].length); | |
817 | ||
f046011c ZB |
818 | if (nr_sig) |
819 | atomic_add(nr_sig, &ic->i_signaled_sends); | |
820 | ||
e622f2f4 CH |
821 | failed_wr = &send->s_atomic_wr.wr; |
822 | ret = ib_post_send(ic->i_cm_id->qp, &send->s_atomic_wr.wr, &failed_wr); | |
15133f6e | 823 | rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic, |
e622f2f4 CH |
824 | send, &send->s_atomic_wr, ret, failed_wr); |
825 | BUG_ON(failed_wr != &send->s_atomic_wr.wr); | |
15133f6e AG |
826 | if (ret) { |
827 | printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 " | |
828 | "returned %d\n", &conn->c_faddr, ret); | |
829 | rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); | |
f046011c | 830 | rds_ib_sub_signaled(ic, nr_sig); |
15133f6e AG |
831 | goto out; |
832 | } | |
833 | ||
e622f2f4 | 834 | if (unlikely(failed_wr != &send->s_atomic_wr.wr)) { |
15133f6e | 835 | printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret); |
e622f2f4 | 836 | BUG_ON(failed_wr != &send->s_atomic_wr.wr); |
15133f6e AG |
837 | } |
838 | ||
839 | out: | |
840 | return ret; | |
841 | } | |
842 | ||
f8b3aaf2 | 843 | int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op) |
6a0979df AG |
844 | { |
845 | struct rds_ib_connection *ic = conn->c_transport_data; | |
846 | struct rds_ib_send_work *send = NULL; | |
847 | struct rds_ib_send_work *first; | |
848 | struct rds_ib_send_work *prev; | |
849 | struct ib_send_wr *failed_wr; | |
6a0979df AG |
850 | struct scatterlist *scat; |
851 | unsigned long len; | |
f8b3aaf2 | 852 | u64 remote_addr = op->op_remote_addr; |
89bf9d41 | 853 | u32 max_sge = ic->rds_ibdev->max_sge; |
6a0979df AG |
854 | u32 pos; |
855 | u32 work_alloc; | |
856 | u32 i; | |
857 | u32 j; | |
858 | int sent; | |
859 | int ret; | |
860 | int num_sge; | |
f046011c | 861 | int nr_sig = 0; |
6a0979df | 862 | |
ff3d7d36 | 863 | /* map the op the first time we see it */ |
f8b3aaf2 AG |
864 | if (!op->op_mapped) { |
865 | op->op_count = ib_dma_map_sg(ic->i_cm_id->device, | |
866 | op->op_sg, op->op_nents, (op->op_write) ? | |
867 | DMA_TO_DEVICE : DMA_FROM_DEVICE); | |
868 | rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count); | |
869 | if (op->op_count == 0) { | |
6a0979df AG |
870 | rds_ib_stats_inc(s_ib_tx_sg_mapping_failure); |
871 | ret = -ENOMEM; /* XXX ? */ | |
872 | goto out; | |
873 | } | |
874 | ||
f8b3aaf2 | 875 | op->op_mapped = 1; |
6a0979df AG |
876 | } |
877 | ||
878 | /* | |
879 | * Instead of knowing how to return a partial rdma read/write we insist that there | |
880 | * be enough work requests to send the entire message. | |
881 | */ | |
89bf9d41 | 882 | i = ceil(op->op_count, max_sge); |
6a0979df AG |
883 | |
884 | work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos); | |
885 | if (work_alloc != i) { | |
886 | rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); | |
887 | rds_ib_stats_inc(s_ib_tx_ring_full); | |
888 | ret = -ENOMEM; | |
889 | goto out; | |
890 | } | |
891 | ||
892 | send = &ic->i_sends[pos]; | |
893 | first = send; | |
894 | prev = NULL; | |
f8b3aaf2 | 895 | scat = &op->op_sg[0]; |
6a0979df | 896 | sent = 0; |
f8b3aaf2 | 897 | num_sge = op->op_count; |
6a0979df | 898 | |
f8b3aaf2 | 899 | for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) { |
6a0979df AG |
900 | send->s_wr.send_flags = 0; |
901 | send->s_queued = jiffies; | |
1cc2228c | 902 | send->s_op = NULL; |
241eef3e | 903 | |
f046011c | 904 | nr_sig += rds_ib_set_wr_signal_state(ic, send, op->op_notify); |
6a0979df | 905 | |
f8b3aaf2 | 906 | send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ; |
e622f2f4 CH |
907 | send->s_rdma_wr.remote_addr = remote_addr; |
908 | send->s_rdma_wr.rkey = op->op_rkey; | |
6a0979df | 909 | |
89bf9d41 | 910 | if (num_sge > max_sge) { |
e622f2f4 | 911 | send->s_rdma_wr.wr.num_sge = max_sge; |
89bf9d41 | 912 | num_sge -= max_sge; |
6a0979df | 913 | } else { |
e622f2f4 | 914 | send->s_rdma_wr.wr.num_sge = num_sge; |
6a0979df AG |
915 | } |
916 | ||
e622f2f4 | 917 | send->s_rdma_wr.wr.next = NULL; |
6a0979df AG |
918 | |
919 | if (prev) | |
e622f2f4 | 920 | prev->s_rdma_wr.wr.next = &send->s_rdma_wr.wr; |
6a0979df | 921 | |
e622f2f4 CH |
922 | for (j = 0; j < send->s_rdma_wr.wr.num_sge && |
923 | scat != &op->op_sg[op->op_count]; j++) { | |
6a0979df AG |
924 | len = ib_sg_dma_len(ic->i_cm_id->device, scat); |
925 | send->s_sge[j].addr = | |
926 | ib_sg_dma_address(ic->i_cm_id->device, scat); | |
927 | send->s_sge[j].length = len; | |
e5580242 | 928 | send->s_sge[j].lkey = ic->i_pd->local_dma_lkey; |
6a0979df AG |
929 | |
930 | sent += len; | |
931 | rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr); | |
932 | ||
933 | remote_addr += len; | |
934 | scat++; | |
935 | } | |
936 | ||
937 | rdsdebug("send %p wr %p num_sge %u next %p\n", send, | |
e622f2f4 CH |
938 | &send->s_rdma_wr.wr, |
939 | send->s_rdma_wr.wr.num_sge, | |
940 | send->s_rdma_wr.wr.next); | |
6a0979df AG |
941 | |
942 | prev = send; | |
943 | if (++send == &ic->i_sends[ic->i_send_ring.w_nr]) | |
944 | send = ic->i_sends; | |
945 | } | |
946 | ||
1cc2228c CM |
947 | /* give a reference to the last op */ |
948 | if (scat == &op->op_sg[op->op_count]) { | |
949 | prev->s_op = op; | |
950 | rds_message_addref(container_of(op, struct rds_message, rdma)); | |
951 | } | |
952 | ||
6a0979df AG |
953 | if (i < work_alloc) { |
954 | rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i); | |
955 | work_alloc = i; | |
956 | } | |
957 | ||
f046011c ZB |
958 | if (nr_sig) |
959 | atomic_add(nr_sig, &ic->i_signaled_sends); | |
960 | ||
e622f2f4 CH |
961 | failed_wr = &first->s_rdma_wr.wr; |
962 | ret = ib_post_send(ic->i_cm_id->qp, &first->s_rdma_wr.wr, &failed_wr); | |
6a0979df | 963 | rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic, |
e622f2f4 CH |
964 | first, &first->s_rdma_wr.wr, ret, failed_wr); |
965 | BUG_ON(failed_wr != &first->s_rdma_wr.wr); | |
6a0979df AG |
966 | if (ret) { |
967 | printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 " | |
968 | "returned %d\n", &conn->c_faddr, ret); | |
969 | rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); | |
f046011c | 970 | rds_ib_sub_signaled(ic, nr_sig); |
6a0979df AG |
971 | goto out; |
972 | } | |
973 | ||
e622f2f4 | 974 | if (unlikely(failed_wr != &first->s_rdma_wr.wr)) { |
6a0979df | 975 | printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret); |
e622f2f4 | 976 | BUG_ON(failed_wr != &first->s_rdma_wr.wr); |
6a0979df AG |
977 | } |
978 | ||
979 | ||
980 | out: | |
981 | return ret; | |
982 | } | |
983 | ||
984 | void rds_ib_xmit_complete(struct rds_connection *conn) | |
985 | { | |
986 | struct rds_ib_connection *ic = conn->c_transport_data; | |
987 | ||
988 | /* We may have a pending ACK or window update we were unable | |
989 | * to send previously (due to flow control). Try again. */ | |
990 | rds_ib_attempt_ack(ic); | |
991 | } |