?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
x86/amd-iommu: Add function to complete a tlb flush
[linux/fpc-iii.git] / drivers / infiniband / hw / ipath / ipath_verbs.c
blob9289ab4b0ae86dd0a37da2d7334538d627d090a5
1 /*
2 * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4 *
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
10 *
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
13 * conditions are met:
14 *
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
17 * disclaimer.
18 *
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
23 *
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31 * SOFTWARE.
32 */
33
34 #include <rdma/ib_mad.h>
35 #include <rdma/ib_user_verbs.h>
36 #include <linux/io.h>
37 #include <linux/utsname.h>
38 #include <linux/rculist.h>
39
40 #include "ipath_kernel.h"
41 #include "ipath_verbs.h"
42 #include "ipath_common.h"
43
44 static unsigned int ib_ipath_qp_table_size = 251;
45 module_param_named(qp_table_size, ib_ipath_qp_table_size, uint, S_IRUGO);
46 MODULE_PARM_DESC(qp_table_size, "QP table size");
47
48 unsigned int ib_ipath_lkey_table_size = 12;
49 module_param_named(lkey_table_size, ib_ipath_lkey_table_size, uint,
50 S_IRUGO);
51 MODULE_PARM_DESC(lkey_table_size,
52 "LKEY table size in bits (2^n, 1 <= n <= 23)");
53
54 static unsigned int ib_ipath_max_pds = 0xFFFF;
55 module_param_named(max_pds, ib_ipath_max_pds, uint, S_IWUSR | S_IRUGO);
56 MODULE_PARM_DESC(max_pds,
57 "Maximum number of protection domains to support");
58
59 static unsigned int ib_ipath_max_ahs = 0xFFFF;
60 module_param_named(max_ahs, ib_ipath_max_ahs, uint, S_IWUSR | S_IRUGO);
61 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
62
63 unsigned int ib_ipath_max_cqes = 0x2FFFF;
64 module_param_named(max_cqes, ib_ipath_max_cqes, uint, S_IWUSR | S_IRUGO);
65 MODULE_PARM_DESC(max_cqes,
66 "Maximum number of completion queue entries to support");
67
68 unsigned int ib_ipath_max_cqs = 0x1FFFF;
69 module_param_named(max_cqs, ib_ipath_max_cqs, uint, S_IWUSR | S_IRUGO);
70 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
71
72 unsigned int ib_ipath_max_qp_wrs = 0x3FFF;
73 module_param_named(max_qp_wrs, ib_ipath_max_qp_wrs, uint,
74 S_IWUSR | S_IRUGO);
75 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
76
77 unsigned int ib_ipath_max_qps = 16384;
78 module_param_named(max_qps, ib_ipath_max_qps, uint, S_IWUSR | S_IRUGO);
79 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
80
81 unsigned int ib_ipath_max_sges = 0x60;
82 module_param_named(max_sges, ib_ipath_max_sges, uint, S_IWUSR | S_IRUGO);
83 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
84
85 unsigned int ib_ipath_max_mcast_grps = 16384;
86 module_param_named(max_mcast_grps, ib_ipath_max_mcast_grps, uint,
87 S_IWUSR | S_IRUGO);
88 MODULE_PARM_DESC(max_mcast_grps,
89 "Maximum number of multicast groups to support");
90
91 unsigned int ib_ipath_max_mcast_qp_attached = 16;
92 module_param_named(max_mcast_qp_attached, ib_ipath_max_mcast_qp_attached,
93 uint, S_IWUSR | S_IRUGO);
94 MODULE_PARM_DESC(max_mcast_qp_attached,
95 "Maximum number of attached QPs to support");
96
97 unsigned int ib_ipath_max_srqs = 1024;
98 module_param_named(max_srqs, ib_ipath_max_srqs, uint, S_IWUSR | S_IRUGO);
99 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
100
101 unsigned int ib_ipath_max_srq_sges = 128;
102 module_param_named(max_srq_sges, ib_ipath_max_srq_sges,
103 uint, S_IWUSR | S_IRUGO);
104 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
105
106 unsigned int ib_ipath_max_srq_wrs = 0x1FFFF;
107 module_param_named(max_srq_wrs, ib_ipath_max_srq_wrs,
108 uint, S_IWUSR | S_IRUGO);
109 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
110
111 static unsigned int ib_ipath_disable_sma;
112 module_param_named(disable_sma, ib_ipath_disable_sma, uint, S_IWUSR | S_IRUGO);
113 MODULE_PARM_DESC(disable_sma, "Disable the SMA");
114
115 /*
116 * Note that it is OK to post send work requests in the SQE and ERR
117 * states; ipath_do_send() will process them and generate error
118 * completions as per IB 1.2 C10-96.
119 */
120 const int ib_ipath_state_ops[IB_QPS_ERR + 1] = {
121 [IB_QPS_RESET] = 0,
122 [IB_QPS_INIT] = IPATH_POST_RECV_OK,
123 [IB_QPS_RTR] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
124 [IB_QPS_RTS] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
125 IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK |
126 IPATH_PROCESS_NEXT_SEND_OK,
127 [IB_QPS_SQD] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
128 IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK,
129 [IB_QPS_SQE] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
130 IPATH_POST_SEND_OK | IPATH_FLUSH_SEND,
131 [IB_QPS_ERR] = IPATH_POST_RECV_OK | IPATH_FLUSH_RECV |
132 IPATH_POST_SEND_OK | IPATH_FLUSH_SEND,
133 };
134
135 struct ipath_ucontext {
136 struct ib_ucontext ibucontext;
137 };
138
139 static inline struct ipath_ucontext *to_iucontext(struct ib_ucontext
140 *ibucontext)
141 {
142 return container_of(ibucontext, struct ipath_ucontext, ibucontext);
143 }
144
145 /*
146 * Translate ib_wr_opcode into ib_wc_opcode.
147 */
148 const enum ib_wc_opcode ib_ipath_wc_opcode[] = {
149 [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
150 [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
151 [IB_WR_SEND] = IB_WC_SEND,
152 [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
153 [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
154 [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
155 [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
156 };
157
158 /*
159 * System image GUID.
160 */
161 static __be64 sys_image_guid;
162
163 /**
164 * ipath_copy_sge - copy data to SGE memory
165 * @ss: the SGE state
166 * @data: the data to copy
167 * @length: the length of the data
168 */
169 void ipath_copy_sge(struct ipath_sge_state *ss, void *data, u32 length)
170 {
171 struct ipath_sge *sge = &ss->sge;
172
173 while (length) {
174 u32 len = sge->length;
175
176 if (len > length)
177 len = length;
178 if (len > sge->sge_length)
179 len = sge->sge_length;
180 BUG_ON(len == 0);
181 memcpy(sge->vaddr, data, len);
182 sge->vaddr += len;
183 sge->length -= len;
184 sge->sge_length -= len;
185 if (sge->sge_length == 0) {
186 if (--ss->num_sge)
187 *sge = *ss->sg_list++;
188 } else if (sge->length == 0 && sge->mr != NULL) {
189 if (++sge->n >= IPATH_SEGSZ) {
190 if (++sge->m >= sge->mr->mapsz)
191 break;
192 sge->n = 0;
193 }
194 sge->vaddr =
195 sge->mr->map[sge->m]->segs[sge->n].vaddr;
196 sge->length =
197 sge->mr->map[sge->m]->segs[sge->n].length;
198 }
199 data += len;
200 length -= len;
201 }
202 }
203
204 /**
205 * ipath_skip_sge - skip over SGE memory - XXX almost dup of prev func
206 * @ss: the SGE state
207 * @length: the number of bytes to skip
208 */
209 void ipath_skip_sge(struct ipath_sge_state *ss, u32 length)
210 {
211 struct ipath_sge *sge = &ss->sge;
212
213 while (length) {
214 u32 len = sge->length;
215
216 if (len > length)
217 len = length;
218 if (len > sge->sge_length)
219 len = sge->sge_length;
220 BUG_ON(len == 0);
221 sge->vaddr += len;
222 sge->length -= len;
223 sge->sge_length -= len;
224 if (sge->sge_length == 0) {
225 if (--ss->num_sge)
226 *sge = *ss->sg_list++;
227 } else if (sge->length == 0 && sge->mr != NULL) {
228 if (++sge->n >= IPATH_SEGSZ) {
229 if (++sge->m >= sge->mr->mapsz)
230 break;
231 sge->n = 0;
232 }
233 sge->vaddr =
234 sge->mr->map[sge->m]->segs[sge->n].vaddr;
235 sge->length =
236 sge->mr->map[sge->m]->segs[sge->n].length;
237 }
238 length -= len;
239 }
240 }
241
242 /*
243 * Count the number of DMA descriptors needed to send length bytes of data.
244 * Don't modify the ipath_sge_state to get the count.
245 * Return zero if any of the segments is not aligned.
246 */
247 static u32 ipath_count_sge(struct ipath_sge_state *ss, u32 length)
248 {
249 struct ipath_sge *sg_list = ss->sg_list;
250 struct ipath_sge sge = ss->sge;
251 u8 num_sge = ss->num_sge;
252 u32 ndesc = 1; /* count the header */
253
254 while (length) {
255 u32 len = sge.length;
256
257 if (len > length)
258 len = length;
259 if (len > sge.sge_length)
260 len = sge.sge_length;
261 BUG_ON(len == 0);
262 if (((long) sge.vaddr & (sizeof(u32) - 1)) ||
263 (len != length && (len & (sizeof(u32) - 1)))) {
264 ndesc = 0;
265 break;
266 }
267 ndesc++;
268 sge.vaddr += len;
269 sge.length -= len;
270 sge.sge_length -= len;
271 if (sge.sge_length == 0) {
272 if (--num_sge)
273 sge = *sg_list++;
274 } else if (sge.length == 0 && sge.mr != NULL) {
275 if (++sge.n >= IPATH_SEGSZ) {
276 if (++sge.m >= sge.mr->mapsz)
277 break;
278 sge.n = 0;
279 }
280 sge.vaddr =
281 sge.mr->map[sge.m]->segs[sge.n].vaddr;
282 sge.length =
283 sge.mr->map[sge.m]->segs[sge.n].length;
284 }
285 length -= len;
286 }
287 return ndesc;
288 }
289
290 /*
291 * Copy from the SGEs to the data buffer.
292 */
293 static void ipath_copy_from_sge(void *data, struct ipath_sge_state *ss,
294 u32 length)
295 {
296 struct ipath_sge *sge = &ss->sge;
297
298 while (length) {
299 u32 len = sge->length;
300
301 if (len > length)
302 len = length;
303 if (len > sge->sge_length)
304 len = sge->sge_length;
305 BUG_ON(len == 0);
306 memcpy(data, sge->vaddr, len);
307 sge->vaddr += len;
308 sge->length -= len;
309 sge->sge_length -= len;
310 if (sge->sge_length == 0) {
311 if (--ss->num_sge)
312 *sge = *ss->sg_list++;
313 } else if (sge->length == 0 && sge->mr != NULL) {
314 if (++sge->n >= IPATH_SEGSZ) {
315 if (++sge->m >= sge->mr->mapsz)
316 break;
317 sge->n = 0;
318 }
319 sge->vaddr =
320 sge->mr->map[sge->m]->segs[sge->n].vaddr;
321 sge->length =
322 sge->mr->map[sge->m]->segs[sge->n].length;
323 }
324 data += len;
325 length -= len;
326 }
327 }
328
329 /**
330 * ipath_post_one_send - post one RC, UC, or UD send work request
331 * @qp: the QP to post on
332 * @wr: the work request to send
333 */
334 static int ipath_post_one_send(struct ipath_qp *qp, struct ib_send_wr *wr)
335 {
336 struct ipath_swqe *wqe;
337 u32 next;
338 int i;
339 int j;
340 int acc;
341 int ret;
342 unsigned long flags;
343 struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
344
345 spin_lock_irqsave(&qp->s_lock, flags);
346
347 if (qp->ibqp.qp_type != IB_QPT_SMI &&
348 !(dd->ipath_flags & IPATH_LINKACTIVE)) {
349 ret = -ENETDOWN;
350 goto bail;
351 }
352
353 /* Check that state is OK to post send. */
354 if (unlikely(!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK)))
355 goto bail_inval;
356
357 /* IB spec says that num_sge == 0 is OK. */
358 if (wr->num_sge > qp->s_max_sge)
359 goto bail_inval;
360
361 /*
362 * Don't allow RDMA reads or atomic operations on UC or
363 * undefined operations.
364 * Make sure buffer is large enough to hold the result for atomics.
365 */
366 if (qp->ibqp.qp_type == IB_QPT_UC) {
367 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ)
368 goto bail_inval;
369 } else if (qp->ibqp.qp_type == IB_QPT_UD) {
370 /* Check UD opcode */
371 if (wr->opcode != IB_WR_SEND &&
372 wr->opcode != IB_WR_SEND_WITH_IMM)
373 goto bail_inval;
374 /* Check UD destination address PD */
375 if (qp->ibqp.pd != wr->wr.ud.ah->pd)
376 goto bail_inval;
377 } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD)
378 goto bail_inval;
379 else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
380 (wr->num_sge == 0 ||
381 wr->sg_list[0].length < sizeof(u64) ||
382 wr->sg_list[0].addr & (sizeof(u64) - 1)))
383 goto bail_inval;
384 else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic)
385 goto bail_inval;
386
387 next = qp->s_head + 1;
388 if (next >= qp->s_size)
389 next = 0;
390 if (next == qp->s_last) {
391 ret = -ENOMEM;
392 goto bail;
393 }
394
395 wqe = get_swqe_ptr(qp, qp->s_head);
396 wqe->wr = *wr;
397 wqe->length = 0;
398 if (wr->num_sge) {
399 acc = wr->opcode >= IB_WR_RDMA_READ ?
400 IB_ACCESS_LOCAL_WRITE : 0;
401 for (i = 0, j = 0; i < wr->num_sge; i++) {
402 u32 length = wr->sg_list[i].length;
403 int ok;
404
405 if (length == 0)
406 continue;
407 ok = ipath_lkey_ok(qp, &wqe->sg_list[j],
408 &wr->sg_list[i], acc);
409 if (!ok)
410 goto bail_inval;
411 wqe->length += length;
412 j++;
413 }
414 wqe->wr.num_sge = j;
415 }
416 if (qp->ibqp.qp_type == IB_QPT_UC ||
417 qp->ibqp.qp_type == IB_QPT_RC) {
418 if (wqe->length > 0x80000000U)
419 goto bail_inval;
420 } else if (wqe->length > to_idev(qp->ibqp.device)->dd->ipath_ibmtu)
421 goto bail_inval;
422 wqe->ssn = qp->s_ssn++;
423 qp->s_head = next;
424
425 ret = 0;
426 goto bail;
427
428 bail_inval:
429 ret = -EINVAL;
430 bail:
431 spin_unlock_irqrestore(&qp->s_lock, flags);
432 return ret;
433 }
434
435 /**
436 * ipath_post_send - post a send on a QP
437 * @ibqp: the QP to post the send on
438 * @wr: the list of work requests to post
439 * @bad_wr: the first bad WR is put here
440 *
441 * This may be called from interrupt context.
442 */
443 static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
444 struct ib_send_wr **bad_wr)
445 {
446 struct ipath_qp *qp = to_iqp(ibqp);
447 int err = 0;
448
449 for (; wr; wr = wr->next) {
450 err = ipath_post_one_send(qp, wr);
451 if (err) {
452 *bad_wr = wr;
453 goto bail;
454 }
455 }
456
457 /* Try to do the send work in the caller's context. */
458 ipath_do_send((unsigned long) qp);
459
460 bail:
461 return err;
462 }
463
464 /**
465 * ipath_post_receive - post a receive on a QP
466 * @ibqp: the QP to post the receive on
467 * @wr: the WR to post
468 * @bad_wr: the first bad WR is put here
469 *
470 * This may be called from interrupt context.
471 */
472 static int ipath_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
473 struct ib_recv_wr **bad_wr)
474 {
475 struct ipath_qp *qp = to_iqp(ibqp);
476 struct ipath_rwq *wq = qp->r_rq.wq;
477 unsigned long flags;
478 int ret;
479
480 /* Check that state is OK to post receive. */
481 if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_RECV_OK) || !wq) {
482 *bad_wr = wr;
483 ret = -EINVAL;
484 goto bail;
485 }
486
487 for (; wr; wr = wr->next) {
488 struct ipath_rwqe *wqe;
489 u32 next;
490 int i;
491
492 if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
493 *bad_wr = wr;
494 ret = -EINVAL;
495 goto bail;
496 }
497
498 spin_lock_irqsave(&qp->r_rq.lock, flags);
499 next = wq->head + 1;
500 if (next >= qp->r_rq.size)
501 next = 0;
502 if (next == wq->tail) {
503 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
504 *bad_wr = wr;
505 ret = -ENOMEM;
506 goto bail;
507 }
508
509 wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
510 wqe->wr_id = wr->wr_id;
511 wqe->num_sge = wr->num_sge;
512 for (i = 0; i < wr->num_sge; i++)
513 wqe->sg_list[i] = wr->sg_list[i];
514 /* Make sure queue entry is written before the head index. */
515 smp_wmb();
516 wq->head = next;
517 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
518 }
519 ret = 0;
520
521 bail:
522 return ret;
523 }
524
525 /**
526 * ipath_qp_rcv - processing an incoming packet on a QP
527 * @dev: the device the packet came on
528 * @hdr: the packet header
529 * @has_grh: true if the packet has a GRH
530 * @data: the packet data
531 * @tlen: the packet length
532 * @qp: the QP the packet came on
533 *
534 * This is called from ipath_ib_rcv() to process an incoming packet
535 * for the given QP.
536 * Called at interrupt level.
537 */
538 static void ipath_qp_rcv(struct ipath_ibdev *dev,
539 struct ipath_ib_header *hdr, int has_grh,
540 void *data, u32 tlen, struct ipath_qp *qp)
541 {
542 /* Check for valid receive state. */
543 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
544 dev->n_pkt_drops++;
545 return;
546 }
547
548 switch (qp->ibqp.qp_type) {
549 case IB_QPT_SMI:
550 case IB_QPT_GSI:
551 if (ib_ipath_disable_sma)
552 break;
553 /* FALLTHROUGH */
554 case IB_QPT_UD:
555 ipath_ud_rcv(dev, hdr, has_grh, data, tlen, qp);
556 break;
557
558 case IB_QPT_RC:
559 ipath_rc_rcv(dev, hdr, has_grh, data, tlen, qp);
560 break;
561
562 case IB_QPT_UC:
563 ipath_uc_rcv(dev, hdr, has_grh, data, tlen, qp);
564 break;
565
566 default:
567 break;
568 }
569 }
570
571 /**
572 * ipath_ib_rcv - process an incoming packet
573 * @arg: the device pointer
574 * @rhdr: the header of the packet
575 * @data: the packet data
576 * @tlen: the packet length
577 *
578 * This is called from ipath_kreceive() to process an incoming packet at
579 * interrupt level. Tlen is the length of the header + data + CRC in bytes.
580 */
581 void ipath_ib_rcv(struct ipath_ibdev *dev, void *rhdr, void *data,
582 u32 tlen)
583 {
584 struct ipath_ib_header *hdr = rhdr;
585 struct ipath_other_headers *ohdr;
586 struct ipath_qp *qp;
587 u32 qp_num;
588 int lnh;
589 u8 opcode;
590 u16 lid;
591
592 if (unlikely(dev == NULL))
593 goto bail;
594
595 if (unlikely(tlen < 24)) { /* LRH+BTH+CRC */
596 dev->rcv_errors++;
597 goto bail;
598 }
599
600 /* Check for a valid destination LID (see ch. 7.11.1). */
601 lid = be16_to_cpu(hdr->lrh[1]);
602 if (lid < IPATH_MULTICAST_LID_BASE) {
603 lid &= ~((1 << dev->dd->ipath_lmc) - 1);
604 if (unlikely(lid != dev->dd->ipath_lid)) {
605 dev->rcv_errors++;
606 goto bail;
607 }
608 }
609
610 /* Check for GRH */
611 lnh = be16_to_cpu(hdr->lrh[0]) & 3;
612 if (lnh == IPATH_LRH_BTH)
613 ohdr = &hdr->u.oth;
614 else if (lnh == IPATH_LRH_GRH)
615 ohdr = &hdr->u.l.oth;
616 else {
617 dev->rcv_errors++;
618 goto bail;
619 }
620
621 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
622 dev->opstats[opcode].n_bytes += tlen;
623 dev->opstats[opcode].n_packets++;
624
625 /* Get the destination QP number. */
626 qp_num = be32_to_cpu(ohdr->bth[1]) & IPATH_QPN_MASK;
627 if (qp_num == IPATH_MULTICAST_QPN) {
628 struct ipath_mcast *mcast;
629 struct ipath_mcast_qp *p;
630
631 if (lnh != IPATH_LRH_GRH) {
632 dev->n_pkt_drops++;
633 goto bail;
634 }
635 mcast = ipath_mcast_find(&hdr->u.l.grh.dgid);
636 if (mcast == NULL) {
637 dev->n_pkt_drops++;
638 goto bail;
639 }
640 dev->n_multicast_rcv++;
641 list_for_each_entry_rcu(p, &mcast->qp_list, list)
642 ipath_qp_rcv(dev, hdr, 1, data, tlen, p->qp);
643 /*
644 * Notify ipath_multicast_detach() if it is waiting for us
645 * to finish.
646 */
647 if (atomic_dec_return(&mcast->refcount) <= 1)
648 wake_up(&mcast->wait);
649 } else {
650 qp = ipath_lookup_qpn(&dev->qp_table, qp_num);
651 if (qp) {
652 dev->n_unicast_rcv++;
653 ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
654 tlen, qp);
655 /*
656 * Notify ipath_destroy_qp() if it is waiting
657 * for us to finish.
658 */
659 if (atomic_dec_and_test(&qp->refcount))
660 wake_up(&qp->wait);
661 } else
662 dev->n_pkt_drops++;
663 }
664
665 bail:;
666 }
667
668 /**
669 * ipath_ib_timer - verbs timer
670 * @arg: the device pointer
671 *
672 * This is called from ipath_do_rcv_timer() at interrupt level to check for
673 * QPs which need retransmits and to collect performance numbers.
674 */
675 static void ipath_ib_timer(struct ipath_ibdev *dev)
676 {
677 struct ipath_qp *resend = NULL;
678 struct ipath_qp *rnr = NULL;
679 struct list_head *last;
680 struct ipath_qp *qp;
681 unsigned long flags;
682
683 if (dev == NULL)
684 return;
685
686 spin_lock_irqsave(&dev->pending_lock, flags);
687 /* Start filling the next pending queue. */
688 if (++dev->pending_index >= ARRAY_SIZE(dev->pending))
689 dev->pending_index = 0;
690 /* Save any requests still in the new queue, they have timed out. */
691 last = &dev->pending[dev->pending_index];
692 while (!list_empty(last)) {
693 qp = list_entry(last->next, struct ipath_qp, timerwait);
694 list_del_init(&qp->timerwait);
695 qp->timer_next = resend;
696 resend = qp;
697 atomic_inc(&qp->refcount);
698 }
699 last = &dev->rnrwait;
700 if (!list_empty(last)) {
701 qp = list_entry(last->next, struct ipath_qp, timerwait);
702 if (--qp->s_rnr_timeout == 0) {
703 do {
704 list_del_init(&qp->timerwait);
705 qp->timer_next = rnr;
706 rnr = qp;
707 atomic_inc(&qp->refcount);
708 if (list_empty(last))
709 break;
710 qp = list_entry(last->next, struct ipath_qp,
711 timerwait);
712 } while (qp->s_rnr_timeout == 0);
713 }
714 }
715 /*
716 * We should only be in the started state if pma_sample_start != 0
717 */
718 if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_STARTED &&
719 --dev->pma_sample_start == 0) {
720 dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_RUNNING;
721 ipath_snapshot_counters(dev->dd, &dev->ipath_sword,
722 &dev->ipath_rword,
723 &dev->ipath_spkts,
724 &dev->ipath_rpkts,
725 &dev->ipath_xmit_wait);
726 }
727 if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_RUNNING) {
728 if (dev->pma_sample_interval == 0) {
729 u64 ta, tb, tc, td, te;
730
731 dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_DONE;
732 ipath_snapshot_counters(dev->dd, &ta, &tb,
733 &tc, &td, &te);
734
735 dev->ipath_sword = ta - dev->ipath_sword;
736 dev->ipath_rword = tb - dev->ipath_rword;
737 dev->ipath_spkts = tc - dev->ipath_spkts;
738 dev->ipath_rpkts = td - dev->ipath_rpkts;
739 dev->ipath_xmit_wait = te - dev->ipath_xmit_wait;
740 }
741 else
742 dev->pma_sample_interval--;
743 }
744 spin_unlock_irqrestore(&dev->pending_lock, flags);
745
746 /* XXX What if timer fires again while this is running? */
747 while (resend != NULL) {
748 qp = resend;
749 resend = qp->timer_next;
750
751 spin_lock_irqsave(&qp->s_lock, flags);
752 if (qp->s_last != qp->s_tail &&
753 ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) {
754 dev->n_timeouts++;
755 ipath_restart_rc(qp, qp->s_last_psn + 1);
756 }
757 spin_unlock_irqrestore(&qp->s_lock, flags);
758
759 /* Notify ipath_destroy_qp() if it is waiting. */
760 if (atomic_dec_and_test(&qp->refcount))
761 wake_up(&qp->wait);
762 }
763 while (rnr != NULL) {
764 qp = rnr;
765 rnr = qp->timer_next;
766
767 spin_lock_irqsave(&qp->s_lock, flags);
768 if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)
769 ipath_schedule_send(qp);
770 spin_unlock_irqrestore(&qp->s_lock, flags);
771
772 /* Notify ipath_destroy_qp() if it is waiting. */
773 if (atomic_dec_and_test(&qp->refcount))
774 wake_up(&qp->wait);
775 }
776 }
777
778 static void update_sge(struct ipath_sge_state *ss, u32 length)
779 {
780 struct ipath_sge *sge = &ss->sge;
781
782 sge->vaddr += length;
783 sge->length -= length;
784 sge->sge_length -= length;
785 if (sge->sge_length == 0) {
786 if (--ss->num_sge)
787 *sge = *ss->sg_list++;
788 } else if (sge->length == 0 && sge->mr != NULL) {
789 if (++sge->n >= IPATH_SEGSZ) {
790 if (++sge->m >= sge->mr->mapsz)
791 return;
792 sge->n = 0;
793 }
794 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
795 sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
796 }
797 }
798
799 #ifdef __LITTLE_ENDIAN
800 static inline u32 get_upper_bits(u32 data, u32 shift)
801 {
802 return data >> shift;
803 }
804
805 static inline u32 set_upper_bits(u32 data, u32 shift)
806 {
807 return data << shift;
808 }
809
810 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
811 {
812 data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
813 data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
814 return data;
815 }
816 #else
817 static inline u32 get_upper_bits(u32 data, u32 shift)
818 {
819 return data << shift;
820 }
821
822 static inline u32 set_upper_bits(u32 data, u32 shift)
823 {
824 return data >> shift;
825 }
826
827 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
828 {
829 data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
830 data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
831 return data;
832 }
833 #endif
834
835 static void copy_io(u32 __iomem *piobuf, struct ipath_sge_state *ss,
836 u32 length, unsigned flush_wc)
837 {
838 u32 extra = 0;
839 u32 data = 0;
840 u32 last;
841
842 while (1) {
843 u32 len = ss->sge.length;
844 u32 off;
845
846 if (len > length)
847 len = length;
848 if (len > ss->sge.sge_length)
849 len = ss->sge.sge_length;
850 BUG_ON(len == 0);
851 /* If the source address is not aligned, try to align it. */
852 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
853 if (off) {
854 u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
855 ~(sizeof(u32) - 1));
856 u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
857 u32 y;
858
859 y = sizeof(u32) - off;
860 if (len > y)
861 len = y;
862 if (len + extra >= sizeof(u32)) {
863 data |= set_upper_bits(v, extra *
864 BITS_PER_BYTE);
865 len = sizeof(u32) - extra;
866 if (len == length) {
867 last = data;
868 break;
869 }
870 __raw_writel(data, piobuf);
871 piobuf++;
872 extra = 0;
873 data = 0;
874 } else {
875 /* Clear unused upper bytes */
876 data |= clear_upper_bytes(v, len, extra);
877 if (len == length) {
878 last = data;
879 break;
880 }
881 extra += len;
882 }
883 } else if (extra) {
884 /* Source address is aligned. */
885 u32 *addr = (u32 *) ss->sge.vaddr;
886 int shift = extra * BITS_PER_BYTE;
887 int ushift = 32 - shift;
888 u32 l = len;
889
890 while (l >= sizeof(u32)) {
891 u32 v = *addr;
892
893 data |= set_upper_bits(v, shift);
894 __raw_writel(data, piobuf);
895 data = get_upper_bits(v, ushift);
896 piobuf++;
897 addr++;
898 l -= sizeof(u32);
899 }
900 /*
901 * We still have 'extra' number of bytes leftover.
902 */
903 if (l) {
904 u32 v = *addr;
905
906 if (l + extra >= sizeof(u32)) {
907 data |= set_upper_bits(v, shift);
908 len -= l + extra - sizeof(u32);
909 if (len == length) {
910 last = data;
911 break;
912 }
913 __raw_writel(data, piobuf);
914 piobuf++;
915 extra = 0;
916 data = 0;
917 } else {
918 /* Clear unused upper bytes */
919 data |= clear_upper_bytes(v, l,
920 extra);
921 if (len == length) {
922 last = data;
923 break;
924 }
925 extra += l;
926 }
927 } else if (len == length) {
928 last = data;
929 break;
930 }
931 } else if (len == length) {
932 u32 w;
933
934 /*
935 * Need to round up for the last dword in the
936 * packet.
937 */
938 w = (len + 3) >> 2;
939 __iowrite32_copy(piobuf, ss->sge.vaddr, w - 1);
940 piobuf += w - 1;
941 last = ((u32 *) ss->sge.vaddr)[w - 1];
942 break;
943 } else {
944 u32 w = len >> 2;
945
946 __iowrite32_copy(piobuf, ss->sge.vaddr, w);
947 piobuf += w;
948
949 extra = len & (sizeof(u32) - 1);
950 if (extra) {
951 u32 v = ((u32 *) ss->sge.vaddr)[w];
952
953 /* Clear unused upper bytes */
954 data = clear_upper_bytes(v, extra, 0);
955 }
956 }
957 update_sge(ss, len);
958 length -= len;
959 }
960 /* Update address before sending packet. */
961 update_sge(ss, length);
962 if (flush_wc) {
963 /* must flush early everything before trigger word */
964 ipath_flush_wc();
965 __raw_writel(last, piobuf);
966 /* be sure trigger word is written */
967 ipath_flush_wc();
968 } else
969 __raw_writel(last, piobuf);
970 }
971
972 /*
973 * Convert IB rate to delay multiplier.
974 */
975 unsigned ipath_ib_rate_to_mult(enum ib_rate rate)
976 {
977 switch (rate) {
978 case IB_RATE_2_5_GBPS: return 8;
979 case IB_RATE_5_GBPS: return 4;
980 case IB_RATE_10_GBPS: return 2;
981 case IB_RATE_20_GBPS: return 1;
982 default: return 0;
983 }
984 }
985
986 /*
987 * Convert delay multiplier to IB rate
988 */
989 static enum ib_rate ipath_mult_to_ib_rate(unsigned mult)
990 {
991 switch (mult) {
992 case 8: return IB_RATE_2_5_GBPS;
993 case 4: return IB_RATE_5_GBPS;
994 case 2: return IB_RATE_10_GBPS;
995 case 1: return IB_RATE_20_GBPS;
996 default: return IB_RATE_PORT_CURRENT;
997 }
998 }
999
1000 static inline struct ipath_verbs_txreq *get_txreq(struct ipath_ibdev *dev)
1001 {
1002 struct ipath_verbs_txreq *tx = NULL;
1003 unsigned long flags;
1004
1005 spin_lock_irqsave(&dev->pending_lock, flags);
1006 if (!list_empty(&dev->txreq_free)) {
1007 struct list_head *l = dev->txreq_free.next;
1008
1009 list_del(l);
1010 tx = list_entry(l, struct ipath_verbs_txreq, txreq.list);
1011 }
1012 spin_unlock_irqrestore(&dev->pending_lock, flags);
1013 return tx;
1014 }
1015
1016 static inline void put_txreq(struct ipath_ibdev *dev,
1017 struct ipath_verbs_txreq *tx)
1018 {
1019 unsigned long flags;
1020
1021 spin_lock_irqsave(&dev->pending_lock, flags);
1022 list_add(&tx->txreq.list, &dev->txreq_free);
1023 spin_unlock_irqrestore(&dev->pending_lock, flags);
1024 }
1025
1026 static void sdma_complete(void *cookie, int status)
1027 {
1028 struct ipath_verbs_txreq *tx = cookie;
1029 struct ipath_qp *qp = tx->qp;
1030 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1031 unsigned long flags;
1032 enum ib_wc_status ibs = status == IPATH_SDMA_TXREQ_S_OK ?
1033 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR;
1034
1035 if (atomic_dec_and_test(&qp->s_dma_busy)) {
1036 spin_lock_irqsave(&qp->s_lock, flags);
1037 if (tx->wqe)
1038 ipath_send_complete(qp, tx->wqe, ibs);
1039 if ((ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND &&
1040 qp->s_last != qp->s_head) ||
1041 (qp->s_flags & IPATH_S_WAIT_DMA))
1042 ipath_schedule_send(qp);
1043 spin_unlock_irqrestore(&qp->s_lock, flags);
1044 wake_up(&qp->wait_dma);
1045 } else if (tx->wqe) {
1046 spin_lock_irqsave(&qp->s_lock, flags);
1047 ipath_send_complete(qp, tx->wqe, ibs);
1048 spin_unlock_irqrestore(&qp->s_lock, flags);
1049 }
1050
1051 if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEBUF)
1052 kfree(tx->txreq.map_addr);
1053 put_txreq(dev, tx);
1054
1055 if (atomic_dec_and_test(&qp->refcount))
1056 wake_up(&qp->wait);
1057 }
1058
1059 static void decrement_dma_busy(struct ipath_qp *qp)
1060 {
1061 unsigned long flags;
1062
1063 if (atomic_dec_and_test(&qp->s_dma_busy)) {
1064 spin_lock_irqsave(&qp->s_lock, flags);
1065 if ((ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND &&
1066 qp->s_last != qp->s_head) ||
1067 (qp->s_flags & IPATH_S_WAIT_DMA))
1068 ipath_schedule_send(qp);
1069 spin_unlock_irqrestore(&qp->s_lock, flags);
1070 wake_up(&qp->wait_dma);
1071 }
1072 }
1073
1074 /*
1075 * Compute the number of clock cycles of delay before sending the next packet.
1076 * The multipliers reflect the number of clocks for the fastest rate so
1077 * one tick at 4xDDR is 8 ticks at 1xSDR.
1078 * If the destination port will take longer to receive a packet than
1079 * the outgoing link can send it, we need to delay sending the next packet
1080 * by the difference in time it takes the receiver to receive and the sender
1081 * to send this packet.
1082 * Note that this delay is always correct for UC and RC but not always
1083 * optimal for UD. For UD, the destination HCA can be different for each
1084 * packet, in which case, we could send packets to a different destination
1085 * while "waiting" for the delay. The overhead for doing this without
1086 * HW support is more than just paying the cost of delaying some packets
1087 * unnecessarily.
1088 */
1089 static inline unsigned ipath_pkt_delay(u32 plen, u8 snd_mult, u8 rcv_mult)
1090 {
1091 return (rcv_mult > snd_mult) ?
1092 (plen * (rcv_mult - snd_mult) + 1) >> 1 : 0;
1093 }
1094
1095 static int ipath_verbs_send_dma(struct ipath_qp *qp,
1096 struct ipath_ib_header *hdr, u32 hdrwords,
1097 struct ipath_sge_state *ss, u32 len,
1098 u32 plen, u32 dwords)
1099 {
1100 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1101 struct ipath_devdata *dd = dev->dd;
1102 struct ipath_verbs_txreq *tx;
1103 u32 *piobuf;
1104 u32 control;
1105 u32 ndesc;
1106 int ret;
1107
1108 tx = qp->s_tx;
1109 if (tx) {
1110 qp->s_tx = NULL;
1111 /* resend previously constructed packet */
1112 atomic_inc(&qp->s_dma_busy);
1113 ret = ipath_sdma_verbs_send(dd, tx->ss, tx->len, tx);
1114 if (ret) {
1115 qp->s_tx = tx;
1116 decrement_dma_busy(qp);
1117 }
1118 goto bail;
1119 }
1120
1121 tx = get_txreq(dev);
1122 if (!tx) {
1123 ret = -EBUSY;
1124 goto bail;
1125 }
1126
1127 /*
1128 * Get the saved delay count we computed for the previous packet
1129 * and save the delay count for this packet to be used next time
1130 * we get here.
1131 */
1132 control = qp->s_pkt_delay;
1133 qp->s_pkt_delay = ipath_pkt_delay(plen, dd->delay_mult, qp->s_dmult);
1134
1135 tx->qp = qp;
1136 atomic_inc(&qp->refcount);
1137 tx->wqe = qp->s_wqe;
1138 tx->txreq.callback = sdma_complete;
1139 tx->txreq.callback_cookie = tx;
1140 tx->txreq.flags = IPATH_SDMA_TXREQ_F_HEADTOHOST |
1141 IPATH_SDMA_TXREQ_F_INTREQ | IPATH_SDMA_TXREQ_F_FREEDESC;
1142 if (plen + 1 >= IPATH_SMALLBUF_DWORDS)
1143 tx->txreq.flags |= IPATH_SDMA_TXREQ_F_USELARGEBUF;
1144
1145 /* VL15 packets bypass credit check */
1146 if ((be16_to_cpu(hdr->lrh[0]) >> 12) == 15) {
1147 control |= 1ULL << 31;
1148 tx->txreq.flags |= IPATH_SDMA_TXREQ_F_VL15;
1149 }
1150
1151 if (len) {
1152 /*
1153 * Don't try to DMA if it takes more descriptors than
1154 * the queue holds.
1155 */
1156 ndesc = ipath_count_sge(ss, len);
1157 if (ndesc >= dd->ipath_sdma_descq_cnt)
1158 ndesc = 0;
1159 } else
1160 ndesc = 1;
1161 if (ndesc) {
1162 tx->hdr.pbc[0] = cpu_to_le32(plen);
1163 tx->hdr.pbc[1] = cpu_to_le32(control);
1164 memcpy(&tx->hdr.hdr, hdr, hdrwords << 2);
1165 tx->txreq.sg_count = ndesc;
1166 tx->map_len = (hdrwords + 2) << 2;
1167 tx->txreq.map_addr = &tx->hdr;
1168 atomic_inc(&qp->s_dma_busy);
1169 ret = ipath_sdma_verbs_send(dd, ss, dwords, tx);
1170 if (ret) {
1171 /* save ss and length in dwords */
1172 tx->ss = ss;
1173 tx->len = dwords;
1174 qp->s_tx = tx;
1175 decrement_dma_busy(qp);
1176 }
1177 goto bail;
1178 }
1179
1180 /* Allocate a buffer and copy the header and payload to it. */
1181 tx->map_len = (plen + 1) << 2;
1182 piobuf = kmalloc(tx->map_len, GFP_ATOMIC);
1183 if (unlikely(piobuf == NULL)) {
1184 ret = -EBUSY;
1185 goto err_tx;
1186 }
1187 tx->txreq.map_addr = piobuf;
1188 tx->txreq.flags |= IPATH_SDMA_TXREQ_F_FREEBUF;
1189 tx->txreq.sg_count = 1;
1190
1191 *piobuf++ = (__force u32) cpu_to_le32(plen);
1192 *piobuf++ = (__force u32) cpu_to_le32(control);
1193 memcpy(piobuf, hdr, hdrwords << 2);
1194 ipath_copy_from_sge(piobuf + hdrwords, ss, len);
1195
1196 atomic_inc(&qp->s_dma_busy);
1197 ret = ipath_sdma_verbs_send(dd, NULL, 0, tx);
1198 /*
1199 * If we couldn't queue the DMA request, save the info
1200 * and try again later rather than destroying the
1201 * buffer and undoing the side effects of the copy.
1202 */
1203 if (ret) {
1204 tx->ss = NULL;
1205 tx->len = 0;
1206 qp->s_tx = tx;
1207 decrement_dma_busy(qp);
1208 }
1209 dev->n_unaligned++;
1210 goto bail;
1211
1212 err_tx:
1213 if (atomic_dec_and_test(&qp->refcount))
1214 wake_up(&qp->wait);
1215 put_txreq(dev, tx);
1216 bail:
1217 return ret;
1218 }
1219
1220 static int ipath_verbs_send_pio(struct ipath_qp *qp,
1221 struct ipath_ib_header *ibhdr, u32 hdrwords,
1222 struct ipath_sge_state *ss, u32 len,
1223 u32 plen, u32 dwords)
1224 {
1225 struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
1226 u32 *hdr = (u32 *) ibhdr;
1227 u32 __iomem *piobuf;
1228 unsigned flush_wc;
1229 u32 control;
1230 int ret;
1231 unsigned long flags;
1232
1233 piobuf = ipath_getpiobuf(dd, plen, NULL);
1234 if (unlikely(piobuf == NULL)) {
1235 ret = -EBUSY;
1236 goto bail;
1237 }
1238
1239 /*
1240 * Get the saved delay count we computed for the previous packet
1241 * and save the delay count for this packet to be used next time
1242 * we get here.
1243 */
1244 control = qp->s_pkt_delay;
1245 qp->s_pkt_delay = ipath_pkt_delay(plen, dd->delay_mult, qp->s_dmult);
1246
1247 /* VL15 packets bypass credit check */
1248 if ((be16_to_cpu(ibhdr->lrh[0]) >> 12) == 15)
1249 control |= 1ULL << 31;
1250
1251 /*
1252 * Write the length to the control qword plus any needed flags.
1253 * We have to flush after the PBC for correctness on some cpus
1254 * or WC buffer can be written out of order.
1255 */
1256 writeq(((u64) control << 32) | plen, piobuf);
1257 piobuf += 2;
1258
1259 flush_wc = dd->ipath_flags & IPATH_PIO_FLUSH_WC;
1260 if (len == 0) {
1261 /*
1262 * If there is just the header portion, must flush before
1263 * writing last word of header for correctness, and after
1264 * the last header word (trigger word).
1265 */
1266 if (flush_wc) {
1267 ipath_flush_wc();
1268 __iowrite32_copy(piobuf, hdr, hdrwords - 1);
1269 ipath_flush_wc();
1270 __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
1271 ipath_flush_wc();
1272 } else
1273 __iowrite32_copy(piobuf, hdr, hdrwords);
1274 goto done;
1275 }
1276
1277 if (flush_wc)
1278 ipath_flush_wc();
1279 __iowrite32_copy(piobuf, hdr, hdrwords);
1280 piobuf += hdrwords;
1281
1282 /* The common case is aligned and contained in one segment. */
1283 if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
1284 !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
1285 u32 *addr = (u32 *) ss->sge.vaddr;
1286
1287 /* Update address before sending packet. */
1288 update_sge(ss, len);
1289 if (flush_wc) {
1290 __iowrite32_copy(piobuf, addr, dwords - 1);
1291 /* must flush early everything before trigger word */
1292 ipath_flush_wc();
1293 __raw_writel(addr[dwords - 1], piobuf + dwords - 1);
1294 /* be sure trigger word is written */
1295 ipath_flush_wc();
1296 } else
1297 __iowrite32_copy(piobuf, addr, dwords);
1298 goto done;
1299 }
1300 copy_io(piobuf, ss, len, flush_wc);
1301 done:
1302 if (qp->s_wqe) {
1303 spin_lock_irqsave(&qp->s_lock, flags);
1304 ipath_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1305 spin_unlock_irqrestore(&qp->s_lock, flags);
1306 }
1307 ret = 0;
1308 bail:
1309 return ret;
1310 }
1311
1312 /**
1313 * ipath_verbs_send - send a packet
1314 * @qp: the QP to send on
1315 * @hdr: the packet header
1316 * @hdrwords: the number of 32-bit words in the header
1317 * @ss: the SGE to send
1318 * @len: the length of the packet in bytes
1319 */
1320 int ipath_verbs_send(struct ipath_qp *qp, struct ipath_ib_header *hdr,
1321 u32 hdrwords, struct ipath_sge_state *ss, u32 len)
1322 {
1323 struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
1324 u32 plen;
1325 int ret;
1326 u32 dwords = (len + 3) >> 2;
1327
1328 /*
1329 * Calculate the send buffer trigger address.
1330 * The +1 counts for the pbc control dword following the pbc length.
1331 */
1332 plen = hdrwords + dwords + 1;
1333
1334 /*
1335 * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1336 * can defer SDMA restart until link goes ACTIVE without
1337 * worrying about just how we got there.
1338 */
1339 if (qp->ibqp.qp_type == IB_QPT_SMI ||
1340 !(dd->ipath_flags & IPATH_HAS_SEND_DMA))
1341 ret = ipath_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1342 plen, dwords);
1343 else
1344 ret = ipath_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1345 plen, dwords);
1346
1347 return ret;
1348 }
1349
1350 int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords,
1351 u64 *rwords, u64 *spkts, u64 *rpkts,
1352 u64 *xmit_wait)
1353 {
1354 int ret;
1355
1356 if (!(dd->ipath_flags & IPATH_INITTED)) {
1357 /* no hardware, freeze, etc. */
1358 ret = -EINVAL;
1359 goto bail;
1360 }
1361 *swords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
1362 *rwords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
1363 *spkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
1364 *rpkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
1365 *xmit_wait = ipath_snap_cntr(dd, dd->ipath_cregs->cr_sendstallcnt);
1366
1367 ret = 0;
1368
1369 bail:
1370 return ret;
1371 }
1372
1373 /**
1374 * ipath_get_counters - get various chip counters
1375 * @dd: the infinipath device
1376 * @cntrs: counters are placed here
1377 *
1378 * Return the counters needed by recv_pma_get_portcounters().
1379 */
1380 int ipath_get_counters(struct ipath_devdata *dd,
1381 struct ipath_verbs_counters *cntrs)
1382 {
1383 struct ipath_cregs const *crp = dd->ipath_cregs;
1384 int ret;
1385
1386 if (!(dd->ipath_flags & IPATH_INITTED)) {
1387 /* no hardware, freeze, etc. */
1388 ret = -EINVAL;
1389 goto bail;
1390 }
1391 cntrs->symbol_error_counter =
1392 ipath_snap_cntr(dd, crp->cr_ibsymbolerrcnt);
1393 cntrs->link_error_recovery_counter =
1394 ipath_snap_cntr(dd, crp->cr_iblinkerrrecovcnt);
1395 /*
1396 * The link downed counter counts when the other side downs the
1397 * connection. We add in the number of times we downed the link
1398 * due to local link integrity errors to compensate.
1399 */
1400 cntrs->link_downed_counter =
1401 ipath_snap_cntr(dd, crp->cr_iblinkdowncnt);
1402 cntrs->port_rcv_errors =
1403 ipath_snap_cntr(dd, crp->cr_rxdroppktcnt) +
1404 ipath_snap_cntr(dd, crp->cr_rcvovflcnt) +
1405 ipath_snap_cntr(dd, crp->cr_portovflcnt) +
1406 ipath_snap_cntr(dd, crp->cr_err_rlencnt) +
1407 ipath_snap_cntr(dd, crp->cr_invalidrlencnt) +
1408 ipath_snap_cntr(dd, crp->cr_errlinkcnt) +
1409 ipath_snap_cntr(dd, crp->cr_erricrccnt) +
1410 ipath_snap_cntr(dd, crp->cr_errvcrccnt) +
1411 ipath_snap_cntr(dd, crp->cr_errlpcrccnt) +
1412 ipath_snap_cntr(dd, crp->cr_badformatcnt) +
1413 dd->ipath_rxfc_unsupvl_errs;
1414 if (crp->cr_rxotherlocalphyerrcnt)
1415 cntrs->port_rcv_errors +=
1416 ipath_snap_cntr(dd, crp->cr_rxotherlocalphyerrcnt);
1417 if (crp->cr_rxvlerrcnt)
1418 cntrs->port_rcv_errors +=
1419 ipath_snap_cntr(dd, crp->cr_rxvlerrcnt);
1420 cntrs->port_rcv_remphys_errors =
1421 ipath_snap_cntr(dd, crp->cr_rcvebpcnt);
1422 cntrs->port_xmit_discards = ipath_snap_cntr(dd, crp->cr_unsupvlcnt);
1423 cntrs->port_xmit_data = ipath_snap_cntr(dd, crp->cr_wordsendcnt);
1424 cntrs->port_rcv_data = ipath_snap_cntr(dd, crp->cr_wordrcvcnt);
1425 cntrs->port_xmit_packets = ipath_snap_cntr(dd, crp->cr_pktsendcnt);
1426 cntrs->port_rcv_packets = ipath_snap_cntr(dd, crp->cr_pktrcvcnt);
1427 cntrs->local_link_integrity_errors =
1428 crp->cr_locallinkintegrityerrcnt ?
1429 ipath_snap_cntr(dd, crp->cr_locallinkintegrityerrcnt) :
1430 ((dd->ipath_flags & IPATH_GPIO_ERRINTRS) ?
1431 dd->ipath_lli_errs : dd->ipath_lli_errors);
1432 cntrs->excessive_buffer_overrun_errors =
1433 crp->cr_excessbufferovflcnt ?
1434 ipath_snap_cntr(dd, crp->cr_excessbufferovflcnt) :
1435 dd->ipath_overrun_thresh_errs;
1436 cntrs->vl15_dropped = crp->cr_vl15droppedpktcnt ?
1437 ipath_snap_cntr(dd, crp->cr_vl15droppedpktcnt) : 0;
1438
1439 ret = 0;
1440
1441 bail:
1442 return ret;
1443 }
1444
1445 /**
1446 * ipath_ib_piobufavail - callback when a PIO buffer is available
1447 * @arg: the device pointer
1448 *
1449 * This is called from ipath_intr() at interrupt level when a PIO buffer is
1450 * available after ipath_verbs_send() returned an error that no buffers were
1451 * available. Return 1 if we consumed all the PIO buffers and we still have
1452 * QPs waiting for buffers (for now, just restart the send tasklet and
1453 * return zero).
1454 */
1455 int ipath_ib_piobufavail(struct ipath_ibdev *dev)
1456 {
1457 struct list_head *list;
1458 struct ipath_qp *qplist;
1459 struct ipath_qp *qp;
1460 unsigned long flags;
1461
1462 if (dev == NULL)
1463 goto bail;
1464
1465 list = &dev->piowait;
1466 qplist = NULL;
1467
1468 spin_lock_irqsave(&dev->pending_lock, flags);
1469 while (!list_empty(list)) {
1470 qp = list_entry(list->next, struct ipath_qp, piowait);
1471 list_del_init(&qp->piowait);
1472 qp->pio_next = qplist;
1473 qplist = qp;
1474 atomic_inc(&qp->refcount);
1475 }
1476 spin_unlock_irqrestore(&dev->pending_lock, flags);
1477
1478 while (qplist != NULL) {
1479 qp = qplist;
1480 qplist = qp->pio_next;
1481
1482 spin_lock_irqsave(&qp->s_lock, flags);
1483 if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)
1484 ipath_schedule_send(qp);
1485 spin_unlock_irqrestore(&qp->s_lock, flags);
1486
1487 /* Notify ipath_destroy_qp() if it is waiting. */
1488 if (atomic_dec_and_test(&qp->refcount))
1489 wake_up(&qp->wait);
1490 }
1491
1492 bail:
1493 return 0;
1494 }
1495
1496 static int ipath_query_device(struct ib_device *ibdev,
1497 struct ib_device_attr *props)
1498 {
1499 struct ipath_ibdev *dev = to_idev(ibdev);
1500
1501 memset(props, 0, sizeof(*props));
1502
1503 props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1504 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1505 IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1506 IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1507 props->page_size_cap = PAGE_SIZE;
1508 props->vendor_id =
1509 IPATH_SRC_OUI_1 << 16 | IPATH_SRC_OUI_2 << 8 | IPATH_SRC_OUI_3;
1510 props->vendor_part_id = dev->dd->ipath_deviceid;
1511 props->hw_ver = dev->dd->ipath_pcirev;
1512
1513 props->sys_image_guid = dev->sys_image_guid;
1514
1515 props->max_mr_size = ~0ull;
1516 props->max_qp = ib_ipath_max_qps;
1517 props->max_qp_wr = ib_ipath_max_qp_wrs;
1518 props->max_sge = ib_ipath_max_sges;
1519 props->max_cq = ib_ipath_max_cqs;
1520 props->max_ah = ib_ipath_max_ahs;
1521 props->max_cqe = ib_ipath_max_cqes;
1522 props->max_mr = dev->lk_table.max;
1523 props->max_fmr = dev->lk_table.max;
1524 props->max_map_per_fmr = 32767;
1525 props->max_pd = ib_ipath_max_pds;
1526 props->max_qp_rd_atom = IPATH_MAX_RDMA_ATOMIC;
1527 props->max_qp_init_rd_atom = 255;
1528 /* props->max_res_rd_atom */
1529 props->max_srq = ib_ipath_max_srqs;
1530 props->max_srq_wr = ib_ipath_max_srq_wrs;
1531 props->max_srq_sge = ib_ipath_max_srq_sges;
1532 /* props->local_ca_ack_delay */
1533 props->atomic_cap = IB_ATOMIC_GLOB;
1534 props->max_pkeys = ipath_get_npkeys(dev->dd);
1535 props->max_mcast_grp = ib_ipath_max_mcast_grps;
1536 props->max_mcast_qp_attach = ib_ipath_max_mcast_qp_attached;
1537 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1538 props->max_mcast_grp;
1539
1540 return 0;
1541 }
1542
1543 const u8 ipath_cvt_physportstate[32] = {
1544 [INFINIPATH_IBCS_LT_STATE_DISABLED] = IB_PHYSPORTSTATE_DISABLED,
1545 [INFINIPATH_IBCS_LT_STATE_LINKUP] = IB_PHYSPORTSTATE_LINKUP,
1546 [INFINIPATH_IBCS_LT_STATE_POLLACTIVE] = IB_PHYSPORTSTATE_POLL,
1547 [INFINIPATH_IBCS_LT_STATE_POLLQUIET] = IB_PHYSPORTSTATE_POLL,
1548 [INFINIPATH_IBCS_LT_STATE_SLEEPDELAY] = IB_PHYSPORTSTATE_SLEEP,
1549 [INFINIPATH_IBCS_LT_STATE_SLEEPQUIET] = IB_PHYSPORTSTATE_SLEEP,
1550 [INFINIPATH_IBCS_LT_STATE_CFGDEBOUNCE] =
1551 IB_PHYSPORTSTATE_CFG_TRAIN,
1552 [INFINIPATH_IBCS_LT_STATE_CFGRCVFCFG] =
1553 IB_PHYSPORTSTATE_CFG_TRAIN,
1554 [INFINIPATH_IBCS_LT_STATE_CFGWAITRMT] =
1555 IB_PHYSPORTSTATE_CFG_TRAIN,
1556 [INFINIPATH_IBCS_LT_STATE_CFGIDLE] = IB_PHYSPORTSTATE_CFG_TRAIN,
1557 [INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN] =
1558 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1559 [INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT] =
1560 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1561 [INFINIPATH_IBCS_LT_STATE_RECOVERIDLE] =
1562 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1563 [0x10] = IB_PHYSPORTSTATE_CFG_TRAIN,
1564 [0x11] = IB_PHYSPORTSTATE_CFG_TRAIN,
1565 [0x12] = IB_PHYSPORTSTATE_CFG_TRAIN,
1566 [0x13] = IB_PHYSPORTSTATE_CFG_TRAIN,
1567 [0x14] = IB_PHYSPORTSTATE_CFG_TRAIN,
1568 [0x15] = IB_PHYSPORTSTATE_CFG_TRAIN,
1569 [0x16] = IB_PHYSPORTSTATE_CFG_TRAIN,
1570 [0x17] = IB_PHYSPORTSTATE_CFG_TRAIN
1571 };
1572
1573 u32 ipath_get_cr_errpkey(struct ipath_devdata *dd)
1574 {
1575 return ipath_read_creg32(dd, dd->ipath_cregs->cr_errpkey);
1576 }
1577
1578 static int ipath_query_port(struct ib_device *ibdev,
1579 u8 port, struct ib_port_attr *props)
1580 {
1581 struct ipath_ibdev *dev = to_idev(ibdev);
1582 struct ipath_devdata *dd = dev->dd;
1583 enum ib_mtu mtu;
1584 u16 lid = dd->ipath_lid;
1585 u64 ibcstat;
1586
1587 memset(props, 0, sizeof(*props));
1588 props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1589 props->lmc = dd->ipath_lmc;
1590 props->sm_lid = dev->sm_lid;
1591 props->sm_sl = dev->sm_sl;
1592 ibcstat = dd->ipath_lastibcstat;
1593 /* map LinkState to IB portinfo values. */
1594 props->state = ipath_ib_linkstate(dd, ibcstat) + 1;
1595
1596 /* See phys_state_show() */
1597 props->phys_state = /* MEA: assumes shift == 0 */
1598 ipath_cvt_physportstate[dd->ipath_lastibcstat &
1599 dd->ibcs_lts_mask];
1600 props->port_cap_flags = dev->port_cap_flags;
1601 props->gid_tbl_len = 1;
1602 props->max_msg_sz = 0x80000000;
1603 props->pkey_tbl_len = ipath_get_npkeys(dd);
1604 props->bad_pkey_cntr = ipath_get_cr_errpkey(dd) -
1605 dev->z_pkey_violations;
1606 props->qkey_viol_cntr = dev->qkey_violations;
1607 props->active_width = dd->ipath_link_width_active;
1608 /* See rate_show() */
1609 props->active_speed = dd->ipath_link_speed_active;
1610 props->max_vl_num = 1; /* VLCap = VL0 */
1611 props->init_type_reply = 0;
1612
1613 props->max_mtu = ipath_mtu4096 ? IB_MTU_4096 : IB_MTU_2048;
1614 switch (dd->ipath_ibmtu) {
1615 case 4096:
1616 mtu = IB_MTU_4096;
1617 break;
1618 case 2048:
1619 mtu = IB_MTU_2048;
1620 break;
1621 case 1024:
1622 mtu = IB_MTU_1024;
1623 break;
1624 case 512:
1625 mtu = IB_MTU_512;
1626 break;
1627 case 256:
1628 mtu = IB_MTU_256;
1629 break;
1630 default:
1631 mtu = IB_MTU_2048;
1632 }
1633 props->active_mtu = mtu;
1634 props->subnet_timeout = dev->subnet_timeout;
1635
1636 return 0;
1637 }
1638
1639 static int ipath_modify_device(struct ib_device *device,
1640 int device_modify_mask,
1641 struct ib_device_modify *device_modify)
1642 {
1643 int ret;
1644
1645 if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1646 IB_DEVICE_MODIFY_NODE_DESC)) {
1647 ret = -EOPNOTSUPP;
1648 goto bail;
1649 }
1650
1651 if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC)
1652 memcpy(device->node_desc, device_modify->node_desc, 64);
1653
1654 if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
1655 to_idev(device)->sys_image_guid =
1656 cpu_to_be64(device_modify->sys_image_guid);
1657
1658 ret = 0;
1659
1660 bail:
1661 return ret;
1662 }
1663
1664 static int ipath_modify_port(struct ib_device *ibdev,
1665 u8 port, int port_modify_mask,
1666 struct ib_port_modify *props)
1667 {
1668 struct ipath_ibdev *dev = to_idev(ibdev);
1669
1670 dev->port_cap_flags |= props->set_port_cap_mask;
1671 dev->port_cap_flags &= ~props->clr_port_cap_mask;
1672 if (port_modify_mask & IB_PORT_SHUTDOWN)
1673 ipath_set_linkstate(dev->dd, IPATH_IB_LINKDOWN);
1674 if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1675 dev->qkey_violations = 0;
1676 return 0;
1677 }
1678
1679 static int ipath_query_gid(struct ib_device *ibdev, u8 port,
1680 int index, union ib_gid *gid)
1681 {
1682 struct ipath_ibdev *dev = to_idev(ibdev);
1683 int ret;
1684
1685 if (index >= 1) {
1686 ret = -EINVAL;
1687 goto bail;
1688 }
1689 gid->global.subnet_prefix = dev->gid_prefix;
1690 gid->global.interface_id = dev->dd->ipath_guid;
1691
1692 ret = 0;
1693
1694 bail:
1695 return ret;
1696 }
1697
1698 static struct ib_pd *ipath_alloc_pd(struct ib_device *ibdev,
1699 struct ib_ucontext *context,
1700 struct ib_udata *udata)
1701 {
1702 struct ipath_ibdev *dev = to_idev(ibdev);
1703 struct ipath_pd *pd;
1704 struct ib_pd *ret;
1705
1706 /*
1707 * This is actually totally arbitrary. Some correctness tests
1708 * assume there's a maximum number of PDs that can be allocated.
1709 * We don't actually have this limit, but we fail the test if
1710 * we allow allocations of more than we report for this value.
1711 */
1712
1713 pd = kmalloc(sizeof *pd, GFP_KERNEL);
1714 if (!pd) {
1715 ret = ERR_PTR(-ENOMEM);
1716 goto bail;
1717 }
1718
1719 spin_lock(&dev->n_pds_lock);
1720 if (dev->n_pds_allocated == ib_ipath_max_pds) {
1721 spin_unlock(&dev->n_pds_lock);
1722 kfree(pd);
1723 ret = ERR_PTR(-ENOMEM);
1724 goto bail;
1725 }
1726
1727 dev->n_pds_allocated++;
1728 spin_unlock(&dev->n_pds_lock);
1729
1730 /* ib_alloc_pd() will initialize pd->ibpd. */
1731 pd->user = udata != NULL;
1732
1733 ret = &pd->ibpd;
1734
1735 bail:
1736 return ret;
1737 }
1738
1739 static int ipath_dealloc_pd(struct ib_pd *ibpd)
1740 {
1741 struct ipath_pd *pd = to_ipd(ibpd);
1742 struct ipath_ibdev *dev = to_idev(ibpd->device);
1743
1744 spin_lock(&dev->n_pds_lock);
1745 dev->n_pds_allocated--;
1746 spin_unlock(&dev->n_pds_lock);
1747
1748 kfree(pd);
1749
1750 return 0;
1751 }
1752
1753 /**
1754 * ipath_create_ah - create an address handle
1755 * @pd: the protection domain
1756 * @ah_attr: the attributes of the AH
1757 *
1758 * This may be called from interrupt context.
1759 */
1760 static struct ib_ah *ipath_create_ah(struct ib_pd *pd,
1761 struct ib_ah_attr *ah_attr)
1762 {
1763 struct ipath_ah *ah;
1764 struct ib_ah *ret;
1765 struct ipath_ibdev *dev = to_idev(pd->device);
1766 unsigned long flags;
1767
1768 /* A multicast address requires a GRH (see ch. 8.4.1). */
1769 if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
1770 ah_attr->dlid != IPATH_PERMISSIVE_LID &&
1771 !(ah_attr->ah_flags & IB_AH_GRH)) {
1772 ret = ERR_PTR(-EINVAL);
1773 goto bail;
1774 }
1775
1776 if (ah_attr->dlid == 0) {
1777 ret = ERR_PTR(-EINVAL);
1778 goto bail;
1779 }
1780
1781 if (ah_attr->port_num < 1 ||
1782 ah_attr->port_num > pd->device->phys_port_cnt) {
1783 ret = ERR_PTR(-EINVAL);
1784 goto bail;
1785 }
1786
1787 ah = kmalloc(sizeof *ah, GFP_ATOMIC);
1788 if (!ah) {
1789 ret = ERR_PTR(-ENOMEM);
1790 goto bail;
1791 }
1792
1793 spin_lock_irqsave(&dev->n_ahs_lock, flags);
1794 if (dev->n_ahs_allocated == ib_ipath_max_ahs) {
1795 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1796 kfree(ah);
1797 ret = ERR_PTR(-ENOMEM);
1798 goto bail;
1799 }
1800
1801 dev->n_ahs_allocated++;
1802 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1803
1804 /* ib_create_ah() will initialize ah->ibah. */
1805 ah->attr = *ah_attr;
1806 ah->attr.static_rate = ipath_ib_rate_to_mult(ah_attr->static_rate);
1807
1808 ret = &ah->ibah;
1809
1810 bail:
1811 return ret;
1812 }
1813
1814 /**
1815 * ipath_destroy_ah - destroy an address handle
1816 * @ibah: the AH to destroy
1817 *
1818 * This may be called from interrupt context.
1819 */
1820 static int ipath_destroy_ah(struct ib_ah *ibah)
1821 {
1822 struct ipath_ibdev *dev = to_idev(ibah->device);
1823 struct ipath_ah *ah = to_iah(ibah);
1824 unsigned long flags;
1825
1826 spin_lock_irqsave(&dev->n_ahs_lock, flags);
1827 dev->n_ahs_allocated--;
1828 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1829
1830 kfree(ah);
1831
1832 return 0;
1833 }
1834
1835 static int ipath_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1836 {
1837 struct ipath_ah *ah = to_iah(ibah);
1838
1839 *ah_attr = ah->attr;
1840 ah_attr->static_rate = ipath_mult_to_ib_rate(ah->attr.static_rate);
1841
1842 return 0;
1843 }
1844
1845 /**
1846 * ipath_get_npkeys - return the size of the PKEY table for port 0
1847 * @dd: the infinipath device
1848 */
1849 unsigned ipath_get_npkeys(struct ipath_devdata *dd)
1850 {
1851 return ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys);
1852 }
1853
1854 /**
1855 * ipath_get_pkey - return the indexed PKEY from the port PKEY table
1856 * @dd: the infinipath device
1857 * @index: the PKEY index
1858 */
1859 unsigned ipath_get_pkey(struct ipath_devdata *dd, unsigned index)
1860 {
1861 unsigned ret;
1862
1863 /* always a kernel port, no locking needed */
1864 if (index >= ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys))
1865 ret = 0;
1866 else
1867 ret = dd->ipath_pd[0]->port_pkeys[index];
1868
1869 return ret;
1870 }
1871
1872 static int ipath_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1873 u16 *pkey)
1874 {
1875 struct ipath_ibdev *dev = to_idev(ibdev);
1876 int ret;
1877
1878 if (index >= ipath_get_npkeys(dev->dd)) {
1879 ret = -EINVAL;
1880 goto bail;
1881 }
1882
1883 *pkey = ipath_get_pkey(dev->dd, index);
1884 ret = 0;
1885
1886 bail:
1887 return ret;
1888 }
1889
1890 /**
1891 * ipath_alloc_ucontext - allocate a ucontest
1892 * @ibdev: the infiniband device
1893 * @udata: not used by the InfiniPath driver
1894 */
1895
1896 static struct ib_ucontext *ipath_alloc_ucontext(struct ib_device *ibdev,
1897 struct ib_udata *udata)
1898 {
1899 struct ipath_ucontext *context;
1900 struct ib_ucontext *ret;
1901
1902 context = kmalloc(sizeof *context, GFP_KERNEL);
1903 if (!context) {
1904 ret = ERR_PTR(-ENOMEM);
1905 goto bail;
1906 }
1907
1908 ret = &context->ibucontext;
1909
1910 bail:
1911 return ret;
1912 }
1913
1914 static int ipath_dealloc_ucontext(struct ib_ucontext *context)
1915 {
1916 kfree(to_iucontext(context));
1917 return 0;
1918 }
1919
1920 static int ipath_verbs_register_sysfs(struct ib_device *dev);
1921
1922 static void __verbs_timer(unsigned long arg)
1923 {
1924 struct ipath_devdata *dd = (struct ipath_devdata *) arg;
1925
1926 /* Handle verbs layer timeouts. */
1927 ipath_ib_timer(dd->verbs_dev);
1928
1929 mod_timer(&dd->verbs_timer, jiffies + 1);
1930 }
1931
1932 static int enable_timer(struct ipath_devdata *dd)
1933 {
1934 /*
1935 * Early chips had a design flaw where the chip and kernel idea
1936 * of the tail register don't always agree, and therefore we won't
1937 * get an interrupt on the next packet received.
1938 * If the board supports per packet receive interrupts, use it.
1939 * Otherwise, the timer function periodically checks for packets
1940 * to cover this case.
1941 * Either way, the timer is needed for verbs layer related
1942 * processing.
1943 */
1944 if (dd->ipath_flags & IPATH_GPIO_INTR) {
1945 ipath_write_kreg(dd, dd->ipath_kregs->kr_debugportselect,
1946 0x2074076542310ULL);
1947 /* Enable GPIO bit 2 interrupt */
1948 dd->ipath_gpio_mask |= (u64) (1 << IPATH_GPIO_PORT0_BIT);
1949 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1950 dd->ipath_gpio_mask);
1951 }
1952
1953 init_timer(&dd->verbs_timer);
1954 dd->verbs_timer.function = __verbs_timer;
1955 dd->verbs_timer.data = (unsigned long)dd;
1956 dd->verbs_timer.expires = jiffies + 1;
1957 add_timer(&dd->verbs_timer);
1958
1959 return 0;
1960 }
1961
1962 static int disable_timer(struct ipath_devdata *dd)
1963 {
1964 /* Disable GPIO bit 2 interrupt */
1965 if (dd->ipath_flags & IPATH_GPIO_INTR) {
1966 /* Disable GPIO bit 2 interrupt */
1967 dd->ipath_gpio_mask &= ~((u64) (1 << IPATH_GPIO_PORT0_BIT));
1968 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1969 dd->ipath_gpio_mask);
1970 /*
1971 * We might want to undo changes to debugportselect,
1972 * but how?
1973 */
1974 }
1975
1976 del_timer_sync(&dd->verbs_timer);
1977
1978 return 0;
1979 }
1980
1981 /**
1982 * ipath_register_ib_device - register our device with the infiniband core
1983 * @dd: the device data structure
1984 * Return the allocated ipath_ibdev pointer or NULL on error.
1985 */
1986 int ipath_register_ib_device(struct ipath_devdata *dd)
1987 {
1988 struct ipath_verbs_counters cntrs;
1989 struct ipath_ibdev *idev;
1990 struct ib_device *dev;
1991 struct ipath_verbs_txreq *tx;
1992 unsigned i;
1993 int ret;
1994
1995 idev = (struct ipath_ibdev *)ib_alloc_device(sizeof *idev);
1996 if (idev == NULL) {
1997 ret = -ENOMEM;
1998 goto bail;
1999 }
2000
2001 dev = &idev->ibdev;
2002
2003 if (dd->ipath_sdma_descq_cnt) {
2004 tx = kmalloc(dd->ipath_sdma_descq_cnt * sizeof *tx,
2005 GFP_KERNEL);
2006 if (tx == NULL) {
2007 ret = -ENOMEM;
2008 goto err_tx;
2009 }
2010 } else
2011 tx = NULL;
2012 idev->txreq_bufs = tx;
2013
2014 /* Only need to initialize non-zero fields. */
2015 spin_lock_init(&idev->n_pds_lock);
2016 spin_lock_init(&idev->n_ahs_lock);
2017 spin_lock_init(&idev->n_cqs_lock);
2018 spin_lock_init(&idev->n_qps_lock);
2019 spin_lock_init(&idev->n_srqs_lock);
2020 spin_lock_init(&idev->n_mcast_grps_lock);
2021
2022 spin_lock_init(&idev->qp_table.lock);
2023 spin_lock_init(&idev->lk_table.lock);
2024 idev->sm_lid = __constant_be16_to_cpu(IB_LID_PERMISSIVE);
2025 /* Set the prefix to the default value (see ch. 4.1.1) */
2026 idev->gid_prefix = __constant_cpu_to_be64(0xfe80000000000000ULL);
2027
2028 ret = ipath_init_qp_table(idev, ib_ipath_qp_table_size);
2029 if (ret)
2030 goto err_qp;
2031
2032 /*
2033 * The top ib_ipath_lkey_table_size bits are used to index the
2034 * table. The lower 8 bits can be owned by the user (copied from
2035 * the LKEY). The remaining bits act as a generation number or tag.
2036 */
2037 idev->lk_table.max = 1 << ib_ipath_lkey_table_size;
2038 idev->lk_table.table = kzalloc(idev->lk_table.max *
2039 sizeof(*idev->lk_table.table),
2040 GFP_KERNEL);
2041 if (idev->lk_table.table == NULL) {
2042 ret = -ENOMEM;
2043 goto err_lk;
2044 }
2045 INIT_LIST_HEAD(&idev->pending_mmaps);
2046 spin_lock_init(&idev->pending_lock);
2047 idev->mmap_offset = PAGE_SIZE;
2048 spin_lock_init(&idev->mmap_offset_lock);
2049 INIT_LIST_HEAD(&idev->pending[0]);
2050 INIT_LIST_HEAD(&idev->pending[1]);
2051 INIT_LIST_HEAD(&idev->pending[2]);
2052 INIT_LIST_HEAD(&idev->piowait);
2053 INIT_LIST_HEAD(&idev->rnrwait);
2054 INIT_LIST_HEAD(&idev->txreq_free);
2055 idev->pending_index = 0;
2056 idev->port_cap_flags =
2057 IB_PORT_SYS_IMAGE_GUID_SUP | IB_PORT_CLIENT_REG_SUP;
2058 if (dd->ipath_flags & IPATH_HAS_LINK_LATENCY)
2059 idev->port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
2060 idev->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
2061 idev->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
2062 idev->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
2063 idev->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
2064 idev->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
2065
2066 /* Snapshot current HW counters to "clear" them. */
2067 ipath_get_counters(dd, &cntrs);
2068 idev->z_symbol_error_counter = cntrs.symbol_error_counter;
2069 idev->z_link_error_recovery_counter =
2070 cntrs.link_error_recovery_counter;
2071 idev->z_link_downed_counter = cntrs.link_downed_counter;
2072 idev->z_port_rcv_errors = cntrs.port_rcv_errors;
2073 idev->z_port_rcv_remphys_errors =
2074 cntrs.port_rcv_remphys_errors;
2075 idev->z_port_xmit_discards = cntrs.port_xmit_discards;
2076 idev->z_port_xmit_data = cntrs.port_xmit_data;
2077 idev->z_port_rcv_data = cntrs.port_rcv_data;
2078 idev->z_port_xmit_packets = cntrs.port_xmit_packets;
2079 idev->z_port_rcv_packets = cntrs.port_rcv_packets;
2080 idev->z_local_link_integrity_errors =
2081 cntrs.local_link_integrity_errors;
2082 idev->z_excessive_buffer_overrun_errors =
2083 cntrs.excessive_buffer_overrun_errors;
2084 idev->z_vl15_dropped = cntrs.vl15_dropped;
2085
2086 for (i = 0; i < dd->ipath_sdma_descq_cnt; i++, tx++)
2087 list_add(&tx->txreq.list, &idev->txreq_free);
2088
2089 /*
2090 * The system image GUID is supposed to be the same for all
2091 * IB HCAs in a single system but since there can be other
2092 * device types in the system, we can't be sure this is unique.
2093 */
2094 if (!sys_image_guid)
2095 sys_image_guid = dd->ipath_guid;
2096 idev->sys_image_guid = sys_image_guid;
2097 idev->ib_unit = dd->ipath_unit;
2098 idev->dd = dd;
2099
2100 strlcpy(dev->name, "ipath%d", IB_DEVICE_NAME_MAX);
2101 dev->owner = THIS_MODULE;
2102 dev->node_guid = dd->ipath_guid;
2103 dev->uverbs_abi_ver = IPATH_UVERBS_ABI_VERSION;
2104 dev->uverbs_cmd_mask =
2105 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
2106 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
2107 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
2108 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
2109 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
2110 (1ull << IB_USER_VERBS_CMD_CREATE_AH) |
2111 (1ull << IB_USER_VERBS_CMD_DESTROY_AH) |
2112 (1ull << IB_USER_VERBS_CMD_QUERY_AH) |
2113 (1ull << IB_USER_VERBS_CMD_REG_MR) |
2114 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
2115 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2116 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
2117 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
2118 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
2119 (1ull << IB_USER_VERBS_CMD_POLL_CQ) |
2120 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) |
2121 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
2122 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
2123 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
2124 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
2125 (1ull << IB_USER_VERBS_CMD_POST_SEND) |
2126 (1ull << IB_USER_VERBS_CMD_POST_RECV) |
2127 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
2128 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST) |
2129 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
2130 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
2131 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
2132 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) |
2133 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
2134 dev->node_type = RDMA_NODE_IB_CA;
2135 dev->phys_port_cnt = 1;
2136 dev->num_comp_vectors = 1;
2137 dev->dma_device = &dd->pcidev->dev;
2138 dev->query_device = ipath_query_device;
2139 dev->modify_device = ipath_modify_device;
2140 dev->query_port = ipath_query_port;
2141 dev->modify_port = ipath_modify_port;
2142 dev->query_pkey = ipath_query_pkey;
2143 dev->query_gid = ipath_query_gid;
2144 dev->alloc_ucontext = ipath_alloc_ucontext;
2145 dev->dealloc_ucontext = ipath_dealloc_ucontext;
2146 dev->alloc_pd = ipath_alloc_pd;
2147 dev->dealloc_pd = ipath_dealloc_pd;
2148 dev->create_ah = ipath_create_ah;
2149 dev->destroy_ah = ipath_destroy_ah;
2150 dev->query_ah = ipath_query_ah;
2151 dev->create_srq = ipath_create_srq;
2152 dev->modify_srq = ipath_modify_srq;
2153 dev->query_srq = ipath_query_srq;
2154 dev->destroy_srq = ipath_destroy_srq;
2155 dev->create_qp = ipath_create_qp;
2156 dev->modify_qp = ipath_modify_qp;
2157 dev->query_qp = ipath_query_qp;
2158 dev->destroy_qp = ipath_destroy_qp;
2159 dev->post_send = ipath_post_send;
2160 dev->post_recv = ipath_post_receive;
2161 dev->post_srq_recv = ipath_post_srq_receive;
2162 dev->create_cq = ipath_create_cq;
2163 dev->destroy_cq = ipath_destroy_cq;
2164 dev->resize_cq = ipath_resize_cq;
2165 dev->poll_cq = ipath_poll_cq;
2166 dev->req_notify_cq = ipath_req_notify_cq;
2167 dev->get_dma_mr = ipath_get_dma_mr;
2168 dev->reg_phys_mr = ipath_reg_phys_mr;
2169 dev->reg_user_mr = ipath_reg_user_mr;
2170 dev->dereg_mr = ipath_dereg_mr;
2171 dev->alloc_fmr = ipath_alloc_fmr;
2172 dev->map_phys_fmr = ipath_map_phys_fmr;
2173 dev->unmap_fmr = ipath_unmap_fmr;
2174 dev->dealloc_fmr = ipath_dealloc_fmr;
2175 dev->attach_mcast = ipath_multicast_attach;
2176 dev->detach_mcast = ipath_multicast_detach;
2177 dev->process_mad = ipath_process_mad;
2178 dev->mmap = ipath_mmap;
2179 dev->dma_ops = &ipath_dma_mapping_ops;
2180
2181 snprintf(dev->node_desc, sizeof(dev->node_desc),
2182 IPATH_IDSTR " %s", init_utsname()->nodename);
2183
2184 ret = ib_register_device(dev);
2185 if (ret)
2186 goto err_reg;
2187
2188 if (ipath_verbs_register_sysfs(dev))
2189 goto err_class;
2190
2191 enable_timer(dd);
2192
2193 goto bail;
2194
2195 err_class:
2196 ib_unregister_device(dev);
2197 err_reg:
2198 kfree(idev->lk_table.table);
2199 err_lk:
2200 kfree(idev->qp_table.table);
2201 err_qp:
2202 kfree(idev->txreq_bufs);
2203 err_tx:
2204 ib_dealloc_device(dev);
2205 ipath_dev_err(dd, "cannot register verbs: %d!\n", -ret);
2206 idev = NULL;
2207
2208 bail:
2209 dd->verbs_dev = idev;
2210 return ret;
2211 }
2212
2213 void ipath_unregister_ib_device(struct ipath_ibdev *dev)
2214 {
2215 struct ib_device *ibdev = &dev->ibdev;
2216 u32 qps_inuse;
2217
2218 ib_unregister_device(ibdev);
2219
2220 disable_timer(dev->dd);
2221
2222 if (!list_empty(&dev->pending[0]) ||
2223 !list_empty(&dev->pending[1]) ||
2224 !list_empty(&dev->pending[2]))
2225 ipath_dev_err(dev->dd, "pending list not empty!\n");
2226 if (!list_empty(&dev->piowait))
2227 ipath_dev_err(dev->dd, "piowait list not empty!\n");
2228 if (!list_empty(&dev->rnrwait))
2229 ipath_dev_err(dev->dd, "rnrwait list not empty!\n");
2230 if (!ipath_mcast_tree_empty())
2231 ipath_dev_err(dev->dd, "multicast table memory leak!\n");
2232 /*
2233 * Note that ipath_unregister_ib_device() can be called before all
2234 * the QPs are destroyed!
2235 */
2236 qps_inuse = ipath_free_all_qps(&dev->qp_table);
2237 if (qps_inuse)
2238 ipath_dev_err(dev->dd, "QP memory leak! %u still in use\n",
2239 qps_inuse);
2240 kfree(dev->qp_table.table);
2241 kfree(dev->lk_table.table);
2242 kfree(dev->txreq_bufs);
2243 ib_dealloc_device(ibdev);
2244 }
2245
2246 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2247 char *buf)
2248 {
2249 struct ipath_ibdev *dev =
2250 container_of(device, struct ipath_ibdev, ibdev.dev);
2251
2252 return sprintf(buf, "%x\n", dev->dd->ipath_pcirev);
2253 }
2254
2255 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2256 char *buf)
2257 {
2258 struct ipath_ibdev *dev =
2259 container_of(device, struct ipath_ibdev, ibdev.dev);
2260 int ret;
2261
2262 ret = dev->dd->ipath_f_get_boardname(dev->dd, buf, 128);
2263 if (ret < 0)
2264 goto bail;
2265 strcat(buf, "\n");
2266 ret = strlen(buf);
2267
2268 bail:
2269 return ret;
2270 }
2271
2272 static ssize_t show_stats(struct device *device, struct device_attribute *attr,
2273 char *buf)
2274 {
2275 struct ipath_ibdev *dev =
2276 container_of(device, struct ipath_ibdev, ibdev.dev);
2277 int i;
2278 int len;
2279
2280 len = sprintf(buf,
2281 "RC resends %d\n"
2282 "RC no QACK %d\n"
2283 "RC ACKs %d\n"
2284 "RC SEQ NAKs %d\n"
2285 "RC RDMA seq %d\n"
2286 "RC RNR NAKs %d\n"
2287 "RC OTH NAKs %d\n"
2288 "RC timeouts %d\n"
2289 "RC RDMA dup %d\n"
2290 "piobuf wait %d\n"
2291 "unaligned %d\n"
2292 "PKT drops %d\n"
2293 "WQE errs %d\n",
2294 dev->n_rc_resends, dev->n_rc_qacks, dev->n_rc_acks,
2295 dev->n_seq_naks, dev->n_rdma_seq, dev->n_rnr_naks,
2296 dev->n_other_naks, dev->n_timeouts,
2297 dev->n_rdma_dup_busy, dev->n_piowait, dev->n_unaligned,
2298 dev->n_pkt_drops, dev->n_wqe_errs);
2299 for (i = 0; i < ARRAY_SIZE(dev->opstats); i++) {
2300 const struct ipath_opcode_stats *si = &dev->opstats[i];
2301
2302 if (!si->n_packets && !si->n_bytes)
2303 continue;
2304 len += sprintf(buf + len, "%02x %llu/%llu\n", i,
2305 (unsigned long long) si->n_packets,
2306 (unsigned long long) si->n_bytes);
2307 }
2308 return len;
2309 }
2310
2311 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
2312 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
2313 static DEVICE_ATTR(board_id, S_IRUGO, show_hca, NULL);
2314 static DEVICE_ATTR(stats, S_IRUGO, show_stats, NULL);
2315
2316 static struct device_attribute *ipath_class_attributes[] = {
2317 &dev_attr_hw_rev,
2318 &dev_attr_hca_type,
2319 &dev_attr_board_id,
2320 &dev_attr_stats
2321 };
2322
2323 static int ipath_verbs_register_sysfs(struct ib_device *dev)
2324 {
2325 int i;
2326 int ret;
2327
2328 for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i)
2329 if (device_create_file(&dev->dev,
2330 ipath_class_attributes[i])) {
2331 ret = 1;
2332 goto bail;
2333 }
2334
2335 ret = 0;
2336
2337 bail:
2338 return ret;
2339 }
Linux with added FPC-III support