]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - drivers/infiniband/hw/qib/qib_verbs.c
IB/qib: Fix QP RCU sparse warnings
[~shefty/rdma-dev.git] / drivers / infiniband / hw / qib / qib_verbs.c
1 /*
2  * Copyright (c) 2012 Intel Corporation.  All rights reserved.
3  * Copyright (c) 2006 - 2012 QLogic Corporation. All rights reserved.
4  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34
35 #include <rdma/ib_mad.h>
36 #include <rdma/ib_user_verbs.h>
37 #include <linux/io.h>
38 #include <linux/module.h>
39 #include <linux/utsname.h>
40 #include <linux/rculist.h>
41 #include <linux/mm.h>
42 #include <linux/random.h>
43
44 #include "qib.h"
45 #include "qib_common.h"
46
47 static unsigned int ib_qib_qp_table_size = 256;
48 module_param_named(qp_table_size, ib_qib_qp_table_size, uint, S_IRUGO);
49 MODULE_PARM_DESC(qp_table_size, "QP table size");
50
51 unsigned int ib_qib_lkey_table_size = 16;
52 module_param_named(lkey_table_size, ib_qib_lkey_table_size, uint,
53                    S_IRUGO);
54 MODULE_PARM_DESC(lkey_table_size,
55                  "LKEY table size in bits (2^n, 1 <= n <= 23)");
56
57 static unsigned int ib_qib_max_pds = 0xFFFF;
58 module_param_named(max_pds, ib_qib_max_pds, uint, S_IRUGO);
59 MODULE_PARM_DESC(max_pds,
60                  "Maximum number of protection domains to support");
61
62 static unsigned int ib_qib_max_ahs = 0xFFFF;
63 module_param_named(max_ahs, ib_qib_max_ahs, uint, S_IRUGO);
64 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
65
66 unsigned int ib_qib_max_cqes = 0x2FFFF;
67 module_param_named(max_cqes, ib_qib_max_cqes, uint, S_IRUGO);
68 MODULE_PARM_DESC(max_cqes,
69                  "Maximum number of completion queue entries to support");
70
71 unsigned int ib_qib_max_cqs = 0x1FFFF;
72 module_param_named(max_cqs, ib_qib_max_cqs, uint, S_IRUGO);
73 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
74
75 unsigned int ib_qib_max_qp_wrs = 0x3FFF;
76 module_param_named(max_qp_wrs, ib_qib_max_qp_wrs, uint, S_IRUGO);
77 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
78
79 unsigned int ib_qib_max_qps = 16384;
80 module_param_named(max_qps, ib_qib_max_qps, uint, S_IRUGO);
81 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
82
83 unsigned int ib_qib_max_sges = 0x60;
84 module_param_named(max_sges, ib_qib_max_sges, uint, S_IRUGO);
85 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
86
87 unsigned int ib_qib_max_mcast_grps = 16384;
88 module_param_named(max_mcast_grps, ib_qib_max_mcast_grps, uint, S_IRUGO);
89 MODULE_PARM_DESC(max_mcast_grps,
90                  "Maximum number of multicast groups to support");
91
92 unsigned int ib_qib_max_mcast_qp_attached = 16;
93 module_param_named(max_mcast_qp_attached, ib_qib_max_mcast_qp_attached,
94                    uint, S_IRUGO);
95 MODULE_PARM_DESC(max_mcast_qp_attached,
96                  "Maximum number of attached QPs to support");
97
98 unsigned int ib_qib_max_srqs = 1024;
99 module_param_named(max_srqs, ib_qib_max_srqs, uint, S_IRUGO);
100 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
101
102 unsigned int ib_qib_max_srq_sges = 128;
103 module_param_named(max_srq_sges, ib_qib_max_srq_sges, uint, S_IRUGO);
104 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
105
106 unsigned int ib_qib_max_srq_wrs = 0x1FFFF;
107 module_param_named(max_srq_wrs, ib_qib_max_srq_wrs, uint, S_IRUGO);
108 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
109
110 static unsigned int ib_qib_disable_sma;
111 module_param_named(disable_sma, ib_qib_disable_sma, uint, S_IWUSR | S_IRUGO);
112 MODULE_PARM_DESC(disable_sma, "Disable the SMA");
113
114 /*
115  * Note that it is OK to post send work requests in the SQE and ERR
116  * states; qib_do_send() will process them and generate error
117  * completions as per IB 1.2 C10-96.
118  */
119 const int ib_qib_state_ops[IB_QPS_ERR + 1] = {
120         [IB_QPS_RESET] = 0,
121         [IB_QPS_INIT] = QIB_POST_RECV_OK,
122         [IB_QPS_RTR] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK,
123         [IB_QPS_RTS] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK |
124             QIB_POST_SEND_OK | QIB_PROCESS_SEND_OK |
125             QIB_PROCESS_NEXT_SEND_OK,
126         [IB_QPS_SQD] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK |
127             QIB_POST_SEND_OK | QIB_PROCESS_SEND_OK,
128         [IB_QPS_SQE] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK |
129             QIB_POST_SEND_OK | QIB_FLUSH_SEND,
130         [IB_QPS_ERR] = QIB_POST_RECV_OK | QIB_FLUSH_RECV |
131             QIB_POST_SEND_OK | QIB_FLUSH_SEND,
132 };
133
134 struct qib_ucontext {
135         struct ib_ucontext ibucontext;
136 };
137
138 static inline struct qib_ucontext *to_iucontext(struct ib_ucontext
139                                                   *ibucontext)
140 {
141         return container_of(ibucontext, struct qib_ucontext, ibucontext);
142 }
143
144 /*
145  * Translate ib_wr_opcode into ib_wc_opcode.
146  */
147 const enum ib_wc_opcode ib_qib_wc_opcode[] = {
148         [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
149         [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
150         [IB_WR_SEND] = IB_WC_SEND,
151         [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
152         [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
153         [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
154         [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
155 };
156
157 /*
158  * System image GUID.
159  */
160 __be64 ib_qib_sys_image_guid;
161
162 /**
163  * qib_copy_sge - copy data to SGE memory
164  * @ss: the SGE state
165  * @data: the data to copy
166  * @length: the length of the data
167  */
168 void qib_copy_sge(struct qib_sge_state *ss, void *data, u32 length, int release)
169 {
170         struct qib_sge *sge = &ss->sge;
171
172         while (length) {
173                 u32 len = sge->length;
174
175                 if (len > length)
176                         len = length;
177                 if (len > sge->sge_length)
178                         len = sge->sge_length;
179                 BUG_ON(len == 0);
180                 memcpy(sge->vaddr, data, len);
181                 sge->vaddr += len;
182                 sge->length -= len;
183                 sge->sge_length -= len;
184                 if (sge->sge_length == 0) {
185                         if (release)
186                                 qib_put_mr(sge->mr);
187                         if (--ss->num_sge)
188                                 *sge = *ss->sg_list++;
189                 } else if (sge->length == 0 && sge->mr->lkey) {
190                         if (++sge->n >= QIB_SEGSZ) {
191                                 if (++sge->m >= sge->mr->mapsz)
192                                         break;
193                                 sge->n = 0;
194                         }
195                         sge->vaddr =
196                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
197                         sge->length =
198                                 sge->mr->map[sge->m]->segs[sge->n].length;
199                 }
200                 data += len;
201                 length -= len;
202         }
203 }
204
205 /**
206  * qib_skip_sge - skip over SGE memory - XXX almost dup of prev func
207  * @ss: the SGE state
208  * @length: the number of bytes to skip
209  */
210 void qib_skip_sge(struct qib_sge_state *ss, u32 length, int release)
211 {
212         struct qib_sge *sge = &ss->sge;
213
214         while (length) {
215                 u32 len = sge->length;
216
217                 if (len > length)
218                         len = length;
219                 if (len > sge->sge_length)
220                         len = sge->sge_length;
221                 BUG_ON(len == 0);
222                 sge->vaddr += len;
223                 sge->length -= len;
224                 sge->sge_length -= len;
225                 if (sge->sge_length == 0) {
226                         if (release)
227                                 qib_put_mr(sge->mr);
228                         if (--ss->num_sge)
229                                 *sge = *ss->sg_list++;
230                 } else if (sge->length == 0 && sge->mr->lkey) {
231                         if (++sge->n >= QIB_SEGSZ) {
232                                 if (++sge->m >= sge->mr->mapsz)
233                                         break;
234                                 sge->n = 0;
235                         }
236                         sge->vaddr =
237                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
238                         sge->length =
239                                 sge->mr->map[sge->m]->segs[sge->n].length;
240                 }
241                 length -= len;
242         }
243 }
244
245 /*
246  * Count the number of DMA descriptors needed to send length bytes of data.
247  * Don't modify the qib_sge_state to get the count.
248  * Return zero if any of the segments is not aligned.
249  */
250 static u32 qib_count_sge(struct qib_sge_state *ss, u32 length)
251 {
252         struct qib_sge *sg_list = ss->sg_list;
253         struct qib_sge sge = ss->sge;
254         u8 num_sge = ss->num_sge;
255         u32 ndesc = 1;  /* count the header */
256
257         while (length) {
258                 u32 len = sge.length;
259
260                 if (len > length)
261                         len = length;
262                 if (len > sge.sge_length)
263                         len = sge.sge_length;
264                 BUG_ON(len == 0);
265                 if (((long) sge.vaddr & (sizeof(u32) - 1)) ||
266                     (len != length && (len & (sizeof(u32) - 1)))) {
267                         ndesc = 0;
268                         break;
269                 }
270                 ndesc++;
271                 sge.vaddr += len;
272                 sge.length -= len;
273                 sge.sge_length -= len;
274                 if (sge.sge_length == 0) {
275                         if (--num_sge)
276                                 sge = *sg_list++;
277                 } else if (sge.length == 0 && sge.mr->lkey) {
278                         if (++sge.n >= QIB_SEGSZ) {
279                                 if (++sge.m >= sge.mr->mapsz)
280                                         break;
281                                 sge.n = 0;
282                         }
283                         sge.vaddr =
284                                 sge.mr->map[sge.m]->segs[sge.n].vaddr;
285                         sge.length =
286                                 sge.mr->map[sge.m]->segs[sge.n].length;
287                 }
288                 length -= len;
289         }
290         return ndesc;
291 }
292
293 /*
294  * Copy from the SGEs to the data buffer.
295  */
296 static void qib_copy_from_sge(void *data, struct qib_sge_state *ss, u32 length)
297 {
298         struct qib_sge *sge = &ss->sge;
299
300         while (length) {
301                 u32 len = sge->length;
302
303                 if (len > length)
304                         len = length;
305                 if (len > sge->sge_length)
306                         len = sge->sge_length;
307                 BUG_ON(len == 0);
308                 memcpy(data, sge->vaddr, len);
309                 sge->vaddr += len;
310                 sge->length -= len;
311                 sge->sge_length -= len;
312                 if (sge->sge_length == 0) {
313                         if (--ss->num_sge)
314                                 *sge = *ss->sg_list++;
315                 } else if (sge->length == 0 && sge->mr->lkey) {
316                         if (++sge->n >= QIB_SEGSZ) {
317                                 if (++sge->m >= sge->mr->mapsz)
318                                         break;
319                                 sge->n = 0;
320                         }
321                         sge->vaddr =
322                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
323                         sge->length =
324                                 sge->mr->map[sge->m]->segs[sge->n].length;
325                 }
326                 data += len;
327                 length -= len;
328         }
329 }
330
331 /**
332  * qib_post_one_send - post one RC, UC, or UD send work request
333  * @qp: the QP to post on
334  * @wr: the work request to send
335  */
336 static int qib_post_one_send(struct qib_qp *qp, struct ib_send_wr *wr)
337 {
338         struct qib_swqe *wqe;
339         u32 next;
340         int i;
341         int j;
342         int acc;
343         int ret;
344         unsigned long flags;
345         struct qib_lkey_table *rkt;
346         struct qib_pd *pd;
347
348         spin_lock_irqsave(&qp->s_lock, flags);
349
350         /* Check that state is OK to post send. */
351         if (unlikely(!(ib_qib_state_ops[qp->state] & QIB_POST_SEND_OK)))
352                 goto bail_inval;
353
354         /* IB spec says that num_sge == 0 is OK. */
355         if (wr->num_sge > qp->s_max_sge)
356                 goto bail_inval;
357
358         /*
359          * Don't allow RDMA reads or atomic operations on UC or
360          * undefined operations.
361          * Make sure buffer is large enough to hold the result for atomics.
362          */
363         if (wr->opcode == IB_WR_FAST_REG_MR) {
364                 if (qib_fast_reg_mr(qp, wr))
365                         goto bail_inval;
366         } else 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_RC) {
370                 /* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_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         rkt = &to_idev(qp->ibqp.device)->lk_table;
396         pd = to_ipd(qp->ibqp.pd);
397         wqe = get_swqe_ptr(qp, qp->s_head);
398         wqe->wr = *wr;
399         wqe->length = 0;
400         j = 0;
401         if (wr->num_sge) {
402                 acc = wr->opcode >= IB_WR_RDMA_READ ?
403                         IB_ACCESS_LOCAL_WRITE : 0;
404                 for (i = 0; i < wr->num_sge; i++) {
405                         u32 length = wr->sg_list[i].length;
406                         int ok;
407
408                         if (length == 0)
409                                 continue;
410                         ok = qib_lkey_ok(rkt, pd, &wqe->sg_list[j],
411                                          &wr->sg_list[i], acc);
412                         if (!ok)
413                                 goto bail_inval_free;
414                         wqe->length += length;
415                         j++;
416                 }
417                 wqe->wr.num_sge = j;
418         }
419         if (qp->ibqp.qp_type == IB_QPT_UC ||
420             qp->ibqp.qp_type == IB_QPT_RC) {
421                 if (wqe->length > 0x80000000U)
422                         goto bail_inval_free;
423         } else if (wqe->length > (dd_from_ibdev(qp->ibqp.device)->pport +
424                                   qp->port_num - 1)->ibmtu)
425                 goto bail_inval_free;
426         else
427                 atomic_inc(&to_iah(wr->wr.ud.ah)->refcount);
428         wqe->ssn = qp->s_ssn++;
429         qp->s_head = next;
430
431         ret = 0;
432         goto bail;
433
434 bail_inval_free:
435         while (j) {
436                 struct qib_sge *sge = &wqe->sg_list[--j];
437
438                 qib_put_mr(sge->mr);
439         }
440 bail_inval:
441         ret = -EINVAL;
442 bail:
443         spin_unlock_irqrestore(&qp->s_lock, flags);
444         return ret;
445 }
446
447 /**
448  * qib_post_send - post a send on a QP
449  * @ibqp: the QP to post the send on
450  * @wr: the list of work requests to post
451  * @bad_wr: the first bad WR is put here
452  *
453  * This may be called from interrupt context.
454  */
455 static int qib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
456                          struct ib_send_wr **bad_wr)
457 {
458         struct qib_qp *qp = to_iqp(ibqp);
459         int err = 0;
460
461         for (; wr; wr = wr->next) {
462                 err = qib_post_one_send(qp, wr);
463                 if (err) {
464                         *bad_wr = wr;
465                         goto bail;
466                 }
467         }
468
469         /* Try to do the send work in the caller's context. */
470         qib_do_send(&qp->s_work);
471
472 bail:
473         return err;
474 }
475
476 /**
477  * qib_post_receive - post a receive on a QP
478  * @ibqp: the QP to post the receive on
479  * @wr: the WR to post
480  * @bad_wr: the first bad WR is put here
481  *
482  * This may be called from interrupt context.
483  */
484 static int qib_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
485                             struct ib_recv_wr **bad_wr)
486 {
487         struct qib_qp *qp = to_iqp(ibqp);
488         struct qib_rwq *wq = qp->r_rq.wq;
489         unsigned long flags;
490         int ret;
491
492         /* Check that state is OK to post receive. */
493         if (!(ib_qib_state_ops[qp->state] & QIB_POST_RECV_OK) || !wq) {
494                 *bad_wr = wr;
495                 ret = -EINVAL;
496                 goto bail;
497         }
498
499         for (; wr; wr = wr->next) {
500                 struct qib_rwqe *wqe;
501                 u32 next;
502                 int i;
503
504                 if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
505                         *bad_wr = wr;
506                         ret = -EINVAL;
507                         goto bail;
508                 }
509
510                 spin_lock_irqsave(&qp->r_rq.lock, flags);
511                 next = wq->head + 1;
512                 if (next >= qp->r_rq.size)
513                         next = 0;
514                 if (next == wq->tail) {
515                         spin_unlock_irqrestore(&qp->r_rq.lock, flags);
516                         *bad_wr = wr;
517                         ret = -ENOMEM;
518                         goto bail;
519                 }
520
521                 wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
522                 wqe->wr_id = wr->wr_id;
523                 wqe->num_sge = wr->num_sge;
524                 for (i = 0; i < wr->num_sge; i++)
525                         wqe->sg_list[i] = wr->sg_list[i];
526                 /* Make sure queue entry is written before the head index. */
527                 smp_wmb();
528                 wq->head = next;
529                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
530         }
531         ret = 0;
532
533 bail:
534         return ret;
535 }
536
537 /**
538  * qib_qp_rcv - processing an incoming packet on a QP
539  * @rcd: the context pointer
540  * @hdr: the packet header
541  * @has_grh: true if the packet has a GRH
542  * @data: the packet data
543  * @tlen: the packet length
544  * @qp: the QP the packet came on
545  *
546  * This is called from qib_ib_rcv() to process an incoming packet
547  * for the given QP.
548  * Called at interrupt level.
549  */
550 static void qib_qp_rcv(struct qib_ctxtdata *rcd, struct qib_ib_header *hdr,
551                        int has_grh, void *data, u32 tlen, struct qib_qp *qp)
552 {
553         struct qib_ibport *ibp = &rcd->ppd->ibport_data;
554
555         spin_lock(&qp->r_lock);
556
557         /* Check for valid receive state. */
558         if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK)) {
559                 ibp->n_pkt_drops++;
560                 goto unlock;
561         }
562
563         switch (qp->ibqp.qp_type) {
564         case IB_QPT_SMI:
565         case IB_QPT_GSI:
566                 if (ib_qib_disable_sma)
567                         break;
568                 /* FALLTHROUGH */
569         case IB_QPT_UD:
570                 qib_ud_rcv(ibp, hdr, has_grh, data, tlen, qp);
571                 break;
572
573         case IB_QPT_RC:
574                 qib_rc_rcv(rcd, hdr, has_grh, data, tlen, qp);
575                 break;
576
577         case IB_QPT_UC:
578                 qib_uc_rcv(ibp, hdr, has_grh, data, tlen, qp);
579                 break;
580
581         default:
582                 break;
583         }
584
585 unlock:
586         spin_unlock(&qp->r_lock);
587 }
588
589 /**
590  * qib_ib_rcv - process an incoming packet
591  * @rcd: the context pointer
592  * @rhdr: the header of the packet
593  * @data: the packet payload
594  * @tlen: the packet length
595  *
596  * This is called from qib_kreceive() to process an incoming packet at
597  * interrupt level. Tlen is the length of the header + data + CRC in bytes.
598  */
599 void qib_ib_rcv(struct qib_ctxtdata *rcd, void *rhdr, void *data, u32 tlen)
600 {
601         struct qib_pportdata *ppd = rcd->ppd;
602         struct qib_ibport *ibp = &ppd->ibport_data;
603         struct qib_ib_header *hdr = rhdr;
604         struct qib_other_headers *ohdr;
605         struct qib_qp *qp;
606         u32 qp_num;
607         int lnh;
608         u8 opcode;
609         u16 lid;
610
611         /* 24 == LRH+BTH+CRC */
612         if (unlikely(tlen < 24))
613                 goto drop;
614
615         /* Check for a valid destination LID (see ch. 7.11.1). */
616         lid = be16_to_cpu(hdr->lrh[1]);
617         if (lid < QIB_MULTICAST_LID_BASE) {
618                 lid &= ~((1 << ppd->lmc) - 1);
619                 if (unlikely(lid != ppd->lid))
620                         goto drop;
621         }
622
623         /* Check for GRH */
624         lnh = be16_to_cpu(hdr->lrh[0]) & 3;
625         if (lnh == QIB_LRH_BTH)
626                 ohdr = &hdr->u.oth;
627         else if (lnh == QIB_LRH_GRH) {
628                 u32 vtf;
629
630                 ohdr = &hdr->u.l.oth;
631                 if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
632                         goto drop;
633                 vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
634                 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
635                         goto drop;
636         } else
637                 goto drop;
638
639         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
640         ibp->opstats[opcode & 0x7f].n_bytes += tlen;
641         ibp->opstats[opcode & 0x7f].n_packets++;
642
643         /* Get the destination QP number. */
644         qp_num = be32_to_cpu(ohdr->bth[1]) & QIB_QPN_MASK;
645         if (qp_num == QIB_MULTICAST_QPN) {
646                 struct qib_mcast *mcast;
647                 struct qib_mcast_qp *p;
648
649                 if (lnh != QIB_LRH_GRH)
650                         goto drop;
651                 mcast = qib_mcast_find(ibp, &hdr->u.l.grh.dgid);
652                 if (mcast == NULL)
653                         goto drop;
654                 ibp->n_multicast_rcv++;
655                 list_for_each_entry_rcu(p, &mcast->qp_list, list)
656                         qib_qp_rcv(rcd, hdr, 1, data, tlen, p->qp);
657                 /*
658                  * Notify qib_multicast_detach() if it is waiting for us
659                  * to finish.
660                  */
661                 if (atomic_dec_return(&mcast->refcount) <= 1)
662                         wake_up(&mcast->wait);
663         } else {
664                 if (rcd->lookaside_qp) {
665                         if (rcd->lookaside_qpn != qp_num) {
666                                 if (atomic_dec_and_test(
667                                         &rcd->lookaside_qp->refcount))
668                                         wake_up(
669                                          &rcd->lookaside_qp->wait);
670                                         rcd->lookaside_qp = NULL;
671                                 }
672                 }
673                 if (!rcd->lookaside_qp) {
674                         qp = qib_lookup_qpn(ibp, qp_num);
675                         if (!qp)
676                                 goto drop;
677                         rcd->lookaside_qp = qp;
678                         rcd->lookaside_qpn = qp_num;
679                 } else
680                         qp = rcd->lookaside_qp;
681                 ibp->n_unicast_rcv++;
682                 qib_qp_rcv(rcd, hdr, lnh == QIB_LRH_GRH, data, tlen, qp);
683         }
684         return;
685
686 drop:
687         ibp->n_pkt_drops++;
688 }
689
690 /*
691  * This is called from a timer to check for QPs
692  * which need kernel memory in order to send a packet.
693  */
694 static void mem_timer(unsigned long data)
695 {
696         struct qib_ibdev *dev = (struct qib_ibdev *) data;
697         struct list_head *list = &dev->memwait;
698         struct qib_qp *qp = NULL;
699         unsigned long flags;
700
701         spin_lock_irqsave(&dev->pending_lock, flags);
702         if (!list_empty(list)) {
703                 qp = list_entry(list->next, struct qib_qp, iowait);
704                 list_del_init(&qp->iowait);
705                 atomic_inc(&qp->refcount);
706                 if (!list_empty(list))
707                         mod_timer(&dev->mem_timer, jiffies + 1);
708         }
709         spin_unlock_irqrestore(&dev->pending_lock, flags);
710
711         if (qp) {
712                 spin_lock_irqsave(&qp->s_lock, flags);
713                 if (qp->s_flags & QIB_S_WAIT_KMEM) {
714                         qp->s_flags &= ~QIB_S_WAIT_KMEM;
715                         qib_schedule_send(qp);
716                 }
717                 spin_unlock_irqrestore(&qp->s_lock, flags);
718                 if (atomic_dec_and_test(&qp->refcount))
719                         wake_up(&qp->wait);
720         }
721 }
722
723 static void update_sge(struct qib_sge_state *ss, u32 length)
724 {
725         struct qib_sge *sge = &ss->sge;
726
727         sge->vaddr += length;
728         sge->length -= length;
729         sge->sge_length -= length;
730         if (sge->sge_length == 0) {
731                 if (--ss->num_sge)
732                         *sge = *ss->sg_list++;
733         } else if (sge->length == 0 && sge->mr->lkey) {
734                 if (++sge->n >= QIB_SEGSZ) {
735                         if (++sge->m >= sge->mr->mapsz)
736                                 return;
737                         sge->n = 0;
738                 }
739                 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
740                 sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
741         }
742 }
743
744 #ifdef __LITTLE_ENDIAN
745 static inline u32 get_upper_bits(u32 data, u32 shift)
746 {
747         return data >> shift;
748 }
749
750 static inline u32 set_upper_bits(u32 data, u32 shift)
751 {
752         return data << shift;
753 }
754
755 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
756 {
757         data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
758         data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
759         return data;
760 }
761 #else
762 static inline u32 get_upper_bits(u32 data, u32 shift)
763 {
764         return data << shift;
765 }
766
767 static inline u32 set_upper_bits(u32 data, u32 shift)
768 {
769         return data >> shift;
770 }
771
772 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
773 {
774         data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
775         data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
776         return data;
777 }
778 #endif
779
780 static void copy_io(u32 __iomem *piobuf, struct qib_sge_state *ss,
781                     u32 length, unsigned flush_wc)
782 {
783         u32 extra = 0;
784         u32 data = 0;
785         u32 last;
786
787         while (1) {
788                 u32 len = ss->sge.length;
789                 u32 off;
790
791                 if (len > length)
792                         len = length;
793                 if (len > ss->sge.sge_length)
794                         len = ss->sge.sge_length;
795                 BUG_ON(len == 0);
796                 /* If the source address is not aligned, try to align it. */
797                 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
798                 if (off) {
799                         u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
800                                             ~(sizeof(u32) - 1));
801                         u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
802                         u32 y;
803
804                         y = sizeof(u32) - off;
805                         if (len > y)
806                                 len = y;
807                         if (len + extra >= sizeof(u32)) {
808                                 data |= set_upper_bits(v, extra *
809                                                        BITS_PER_BYTE);
810                                 len = sizeof(u32) - extra;
811                                 if (len == length) {
812                                         last = data;
813                                         break;
814                                 }
815                                 __raw_writel(data, piobuf);
816                                 piobuf++;
817                                 extra = 0;
818                                 data = 0;
819                         } else {
820                                 /* Clear unused upper bytes */
821                                 data |= clear_upper_bytes(v, len, extra);
822                                 if (len == length) {
823                                         last = data;
824                                         break;
825                                 }
826                                 extra += len;
827                         }
828                 } else if (extra) {
829                         /* Source address is aligned. */
830                         u32 *addr = (u32 *) ss->sge.vaddr;
831                         int shift = extra * BITS_PER_BYTE;
832                         int ushift = 32 - shift;
833                         u32 l = len;
834
835                         while (l >= sizeof(u32)) {
836                                 u32 v = *addr;
837
838                                 data |= set_upper_bits(v, shift);
839                                 __raw_writel(data, piobuf);
840                                 data = get_upper_bits(v, ushift);
841                                 piobuf++;
842                                 addr++;
843                                 l -= sizeof(u32);
844                         }
845                         /*
846                          * We still have 'extra' number of bytes leftover.
847                          */
848                         if (l) {
849                                 u32 v = *addr;
850
851                                 if (l + extra >= sizeof(u32)) {
852                                         data |= set_upper_bits(v, shift);
853                                         len -= l + extra - sizeof(u32);
854                                         if (len == length) {
855                                                 last = data;
856                                                 break;
857                                         }
858                                         __raw_writel(data, piobuf);
859                                         piobuf++;
860                                         extra = 0;
861                                         data = 0;
862                                 } else {
863                                         /* Clear unused upper bytes */
864                                         data |= clear_upper_bytes(v, l, extra);
865                                         if (len == length) {
866                                                 last = data;
867                                                 break;
868                                         }
869                                         extra += l;
870                                 }
871                         } else if (len == length) {
872                                 last = data;
873                                 break;
874                         }
875                 } else if (len == length) {
876                         u32 w;
877
878                         /*
879                          * Need to round up for the last dword in the
880                          * packet.
881                          */
882                         w = (len + 3) >> 2;
883                         qib_pio_copy(piobuf, ss->sge.vaddr, w - 1);
884                         piobuf += w - 1;
885                         last = ((u32 *) ss->sge.vaddr)[w - 1];
886                         break;
887                 } else {
888                         u32 w = len >> 2;
889
890                         qib_pio_copy(piobuf, ss->sge.vaddr, w);
891                         piobuf += w;
892
893                         extra = len & (sizeof(u32) - 1);
894                         if (extra) {
895                                 u32 v = ((u32 *) ss->sge.vaddr)[w];
896
897                                 /* Clear unused upper bytes */
898                                 data = clear_upper_bytes(v, extra, 0);
899                         }
900                 }
901                 update_sge(ss, len);
902                 length -= len;
903         }
904         /* Update address before sending packet. */
905         update_sge(ss, length);
906         if (flush_wc) {
907                 /* must flush early everything before trigger word */
908                 qib_flush_wc();
909                 __raw_writel(last, piobuf);
910                 /* be sure trigger word is written */
911                 qib_flush_wc();
912         } else
913                 __raw_writel(last, piobuf);
914 }
915
916 static noinline struct qib_verbs_txreq *__get_txreq(struct qib_ibdev *dev,
917                                            struct qib_qp *qp)
918 {
919         struct qib_verbs_txreq *tx;
920         unsigned long flags;
921
922         spin_lock_irqsave(&qp->s_lock, flags);
923         spin_lock(&dev->pending_lock);
924
925         if (!list_empty(&dev->txreq_free)) {
926                 struct list_head *l = dev->txreq_free.next;
927
928                 list_del(l);
929                 spin_unlock(&dev->pending_lock);
930                 spin_unlock_irqrestore(&qp->s_lock, flags);
931                 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
932         } else {
933                 if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK &&
934                     list_empty(&qp->iowait)) {
935                         dev->n_txwait++;
936                         qp->s_flags |= QIB_S_WAIT_TX;
937                         list_add_tail(&qp->iowait, &dev->txwait);
938                 }
939                 qp->s_flags &= ~QIB_S_BUSY;
940                 spin_unlock(&dev->pending_lock);
941                 spin_unlock_irqrestore(&qp->s_lock, flags);
942                 tx = ERR_PTR(-EBUSY);
943         }
944         return tx;
945 }
946
947 static inline struct qib_verbs_txreq *get_txreq(struct qib_ibdev *dev,
948                                          struct qib_qp *qp)
949 {
950         struct qib_verbs_txreq *tx;
951         unsigned long flags;
952
953         spin_lock_irqsave(&dev->pending_lock, flags);
954         /* assume the list non empty */
955         if (likely(!list_empty(&dev->txreq_free))) {
956                 struct list_head *l = dev->txreq_free.next;
957
958                 list_del(l);
959                 spin_unlock_irqrestore(&dev->pending_lock, flags);
960                 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
961         } else {
962                 /* call slow path to get the extra lock */
963                 spin_unlock_irqrestore(&dev->pending_lock, flags);
964                 tx =  __get_txreq(dev, qp);
965         }
966         return tx;
967 }
968
969 void qib_put_txreq(struct qib_verbs_txreq *tx)
970 {
971         struct qib_ibdev *dev;
972         struct qib_qp *qp;
973         unsigned long flags;
974
975         qp = tx->qp;
976         dev = to_idev(qp->ibqp.device);
977
978         if (atomic_dec_and_test(&qp->refcount))
979                 wake_up(&qp->wait);
980         if (tx->mr) {
981                 qib_put_mr(tx->mr);
982                 tx->mr = NULL;
983         }
984         if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) {
985                 tx->txreq.flags &= ~QIB_SDMA_TXREQ_F_FREEBUF;
986                 dma_unmap_single(&dd_from_dev(dev)->pcidev->dev,
987                                  tx->txreq.addr, tx->hdr_dwords << 2,
988                                  DMA_TO_DEVICE);
989                 kfree(tx->align_buf);
990         }
991
992         spin_lock_irqsave(&dev->pending_lock, flags);
993
994         /* Put struct back on free list */
995         list_add(&tx->txreq.list, &dev->txreq_free);
996
997         if (!list_empty(&dev->txwait)) {
998                 /* Wake up first QP wanting a free struct */
999                 qp = list_entry(dev->txwait.next, struct qib_qp, iowait);
1000                 list_del_init(&qp->iowait);
1001                 atomic_inc(&qp->refcount);
1002                 spin_unlock_irqrestore(&dev->pending_lock, flags);
1003
1004                 spin_lock_irqsave(&qp->s_lock, flags);
1005                 if (qp->s_flags & QIB_S_WAIT_TX) {
1006                         qp->s_flags &= ~QIB_S_WAIT_TX;
1007                         qib_schedule_send(qp);
1008                 }
1009                 spin_unlock_irqrestore(&qp->s_lock, flags);
1010
1011                 if (atomic_dec_and_test(&qp->refcount))
1012                         wake_up(&qp->wait);
1013         } else
1014                 spin_unlock_irqrestore(&dev->pending_lock, flags);
1015 }
1016
1017 /*
1018  * This is called when there are send DMA descriptors that might be
1019  * available.
1020  *
1021  * This is called with ppd->sdma_lock held.
1022  */
1023 void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail)
1024 {
1025         struct qib_qp *qp, *nqp;
1026         struct qib_qp *qps[20];
1027         struct qib_ibdev *dev;
1028         unsigned i, n;
1029
1030         n = 0;
1031         dev = &ppd->dd->verbs_dev;
1032         spin_lock(&dev->pending_lock);
1033
1034         /* Search wait list for first QP wanting DMA descriptors. */
1035         list_for_each_entry_safe(qp, nqp, &dev->dmawait, iowait) {
1036                 if (qp->port_num != ppd->port)
1037                         continue;
1038                 if (n == ARRAY_SIZE(qps))
1039                         break;
1040                 if (qp->s_tx->txreq.sg_count > avail)
1041                         break;
1042                 avail -= qp->s_tx->txreq.sg_count;
1043                 list_del_init(&qp->iowait);
1044                 atomic_inc(&qp->refcount);
1045                 qps[n++] = qp;
1046         }
1047
1048         spin_unlock(&dev->pending_lock);
1049
1050         for (i = 0; i < n; i++) {
1051                 qp = qps[i];
1052                 spin_lock(&qp->s_lock);
1053                 if (qp->s_flags & QIB_S_WAIT_DMA_DESC) {
1054                         qp->s_flags &= ~QIB_S_WAIT_DMA_DESC;
1055                         qib_schedule_send(qp);
1056                 }
1057                 spin_unlock(&qp->s_lock);
1058                 if (atomic_dec_and_test(&qp->refcount))
1059                         wake_up(&qp->wait);
1060         }
1061 }
1062
1063 /*
1064  * This is called with ppd->sdma_lock held.
1065  */
1066 static void sdma_complete(struct qib_sdma_txreq *cookie, int status)
1067 {
1068         struct qib_verbs_txreq *tx =
1069                 container_of(cookie, struct qib_verbs_txreq, txreq);
1070         struct qib_qp *qp = tx->qp;
1071
1072         spin_lock(&qp->s_lock);
1073         if (tx->wqe)
1074                 qib_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
1075         else if (qp->ibqp.qp_type == IB_QPT_RC) {
1076                 struct qib_ib_header *hdr;
1077
1078                 if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF)
1079                         hdr = &tx->align_buf->hdr;
1080                 else {
1081                         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
1082
1083                         hdr = &dev->pio_hdrs[tx->hdr_inx].hdr;
1084                 }
1085                 qib_rc_send_complete(qp, hdr);
1086         }
1087         if (atomic_dec_and_test(&qp->s_dma_busy)) {
1088                 if (qp->state == IB_QPS_RESET)
1089                         wake_up(&qp->wait_dma);
1090                 else if (qp->s_flags & QIB_S_WAIT_DMA) {
1091                         qp->s_flags &= ~QIB_S_WAIT_DMA;
1092                         qib_schedule_send(qp);
1093                 }
1094         }
1095         spin_unlock(&qp->s_lock);
1096
1097         qib_put_txreq(tx);
1098 }
1099
1100 static int wait_kmem(struct qib_ibdev *dev, struct qib_qp *qp)
1101 {
1102         unsigned long flags;
1103         int ret = 0;
1104
1105         spin_lock_irqsave(&qp->s_lock, flags);
1106         if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) {
1107                 spin_lock(&dev->pending_lock);
1108                 if (list_empty(&qp->iowait)) {
1109                         if (list_empty(&dev->memwait))
1110                                 mod_timer(&dev->mem_timer, jiffies + 1);
1111                         qp->s_flags |= QIB_S_WAIT_KMEM;
1112                         list_add_tail(&qp->iowait, &dev->memwait);
1113                 }
1114                 spin_unlock(&dev->pending_lock);
1115                 qp->s_flags &= ~QIB_S_BUSY;
1116                 ret = -EBUSY;
1117         }
1118         spin_unlock_irqrestore(&qp->s_lock, flags);
1119
1120         return ret;
1121 }
1122
1123 static int qib_verbs_send_dma(struct qib_qp *qp, struct qib_ib_header *hdr,
1124                               u32 hdrwords, struct qib_sge_state *ss, u32 len,
1125                               u32 plen, u32 dwords)
1126 {
1127         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
1128         struct qib_devdata *dd = dd_from_dev(dev);
1129         struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1130         struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1131         struct qib_verbs_txreq *tx;
1132         struct qib_pio_header *phdr;
1133         u32 control;
1134         u32 ndesc;
1135         int ret;
1136
1137         tx = qp->s_tx;
1138         if (tx) {
1139                 qp->s_tx = NULL;
1140                 /* resend previously constructed packet */
1141                 ret = qib_sdma_verbs_send(ppd, tx->ss, tx->dwords, tx);
1142                 goto bail;
1143         }
1144
1145         tx = get_txreq(dev, qp);
1146         if (IS_ERR(tx))
1147                 goto bail_tx;
1148
1149         control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
1150                                        be16_to_cpu(hdr->lrh[0]) >> 12);
1151         tx->qp = qp;
1152         atomic_inc(&qp->refcount);
1153         tx->wqe = qp->s_wqe;
1154         tx->mr = qp->s_rdma_mr;
1155         if (qp->s_rdma_mr)
1156                 qp->s_rdma_mr = NULL;
1157         tx->txreq.callback = sdma_complete;
1158         if (dd->flags & QIB_HAS_SDMA_TIMEOUT)
1159                 tx->txreq.flags = QIB_SDMA_TXREQ_F_HEADTOHOST;
1160         else
1161                 tx->txreq.flags = QIB_SDMA_TXREQ_F_INTREQ;
1162         if (plen + 1 > dd->piosize2kmax_dwords)
1163                 tx->txreq.flags |= QIB_SDMA_TXREQ_F_USELARGEBUF;
1164
1165         if (len) {
1166                 /*
1167                  * Don't try to DMA if it takes more descriptors than
1168                  * the queue holds.
1169                  */
1170                 ndesc = qib_count_sge(ss, len);
1171                 if (ndesc >= ppd->sdma_descq_cnt)
1172                         ndesc = 0;
1173         } else
1174                 ndesc = 1;
1175         if (ndesc) {
1176                 phdr = &dev->pio_hdrs[tx->hdr_inx];
1177                 phdr->pbc[0] = cpu_to_le32(plen);
1178                 phdr->pbc[1] = cpu_to_le32(control);
1179                 memcpy(&phdr->hdr, hdr, hdrwords << 2);
1180                 tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEDESC;
1181                 tx->txreq.sg_count = ndesc;
1182                 tx->txreq.addr = dev->pio_hdrs_phys +
1183                         tx->hdr_inx * sizeof(struct qib_pio_header);
1184                 tx->hdr_dwords = hdrwords + 2; /* add PBC length */
1185                 ret = qib_sdma_verbs_send(ppd, ss, dwords, tx);
1186                 goto bail;
1187         }
1188
1189         /* Allocate a buffer and copy the header and payload to it. */
1190         tx->hdr_dwords = plen + 1;
1191         phdr = kmalloc(tx->hdr_dwords << 2, GFP_ATOMIC);
1192         if (!phdr)
1193                 goto err_tx;
1194         phdr->pbc[0] = cpu_to_le32(plen);
1195         phdr->pbc[1] = cpu_to_le32(control);
1196         memcpy(&phdr->hdr, hdr, hdrwords << 2);
1197         qib_copy_from_sge((u32 *) &phdr->hdr + hdrwords, ss, len);
1198
1199         tx->txreq.addr = dma_map_single(&dd->pcidev->dev, phdr,
1200                                         tx->hdr_dwords << 2, DMA_TO_DEVICE);
1201         if (dma_mapping_error(&dd->pcidev->dev, tx->txreq.addr))
1202                 goto map_err;
1203         tx->align_buf = phdr;
1204         tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEBUF;
1205         tx->txreq.sg_count = 1;
1206         ret = qib_sdma_verbs_send(ppd, NULL, 0, tx);
1207         goto unaligned;
1208
1209 map_err:
1210         kfree(phdr);
1211 err_tx:
1212         qib_put_txreq(tx);
1213         ret = wait_kmem(dev, qp);
1214 unaligned:
1215         ibp->n_unaligned++;
1216 bail:
1217         return ret;
1218 bail_tx:
1219         ret = PTR_ERR(tx);
1220         goto bail;
1221 }
1222
1223 /*
1224  * If we are now in the error state, return zero to flush the
1225  * send work request.
1226  */
1227 static int no_bufs_available(struct qib_qp *qp)
1228 {
1229         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
1230         struct qib_devdata *dd;
1231         unsigned long flags;
1232         int ret = 0;
1233
1234         /*
1235          * Note that as soon as want_buffer() is called and
1236          * possibly before it returns, qib_ib_piobufavail()
1237          * could be called. Therefore, put QP on the I/O wait list before
1238          * enabling the PIO avail interrupt.
1239          */
1240         spin_lock_irqsave(&qp->s_lock, flags);
1241         if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) {
1242                 spin_lock(&dev->pending_lock);
1243                 if (list_empty(&qp->iowait)) {
1244                         dev->n_piowait++;
1245                         qp->s_flags |= QIB_S_WAIT_PIO;
1246                         list_add_tail(&qp->iowait, &dev->piowait);
1247                         dd = dd_from_dev(dev);
1248                         dd->f_wantpiobuf_intr(dd, 1);
1249                 }
1250                 spin_unlock(&dev->pending_lock);
1251                 qp->s_flags &= ~QIB_S_BUSY;
1252                 ret = -EBUSY;
1253         }
1254         spin_unlock_irqrestore(&qp->s_lock, flags);
1255         return ret;
1256 }
1257
1258 static int qib_verbs_send_pio(struct qib_qp *qp, struct qib_ib_header *ibhdr,
1259                               u32 hdrwords, struct qib_sge_state *ss, u32 len,
1260                               u32 plen, u32 dwords)
1261 {
1262         struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1263         struct qib_pportdata *ppd = dd->pport + qp->port_num - 1;
1264         u32 *hdr = (u32 *) ibhdr;
1265         u32 __iomem *piobuf_orig;
1266         u32 __iomem *piobuf;
1267         u64 pbc;
1268         unsigned long flags;
1269         unsigned flush_wc;
1270         u32 control;
1271         u32 pbufn;
1272
1273         control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
1274                 be16_to_cpu(ibhdr->lrh[0]) >> 12);
1275         pbc = ((u64) control << 32) | plen;
1276         piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
1277         if (unlikely(piobuf == NULL))
1278                 return no_bufs_available(qp);
1279
1280         /*
1281          * Write the pbc.
1282          * We have to flush after the PBC for correctness on some cpus
1283          * or WC buffer can be written out of order.
1284          */
1285         writeq(pbc, piobuf);
1286         piobuf_orig = piobuf;
1287         piobuf += 2;
1288
1289         flush_wc = dd->flags & QIB_PIO_FLUSH_WC;
1290         if (len == 0) {
1291                 /*
1292                  * If there is just the header portion, must flush before
1293                  * writing last word of header for correctness, and after
1294                  * the last header word (trigger word).
1295                  */
1296                 if (flush_wc) {
1297                         qib_flush_wc();
1298                         qib_pio_copy(piobuf, hdr, hdrwords - 1);
1299                         qib_flush_wc();
1300                         __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
1301                         qib_flush_wc();
1302                 } else
1303                         qib_pio_copy(piobuf, hdr, hdrwords);
1304                 goto done;
1305         }
1306
1307         if (flush_wc)
1308                 qib_flush_wc();
1309         qib_pio_copy(piobuf, hdr, hdrwords);
1310         piobuf += hdrwords;
1311
1312         /* The common case is aligned and contained in one segment. */
1313         if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
1314                    !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
1315                 u32 *addr = (u32 *) ss->sge.vaddr;
1316
1317                 /* Update address before sending packet. */
1318                 update_sge(ss, len);
1319                 if (flush_wc) {
1320                         qib_pio_copy(piobuf, addr, dwords - 1);
1321                         /* must flush early everything before trigger word */
1322                         qib_flush_wc();
1323                         __raw_writel(addr[dwords - 1], piobuf + dwords - 1);
1324                         /* be sure trigger word is written */
1325                         qib_flush_wc();
1326                 } else
1327                         qib_pio_copy(piobuf, addr, dwords);
1328                 goto done;
1329         }
1330         copy_io(piobuf, ss, len, flush_wc);
1331 done:
1332         if (dd->flags & QIB_USE_SPCL_TRIG) {
1333                 u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
1334                 qib_flush_wc();
1335                 __raw_writel(0xaebecede, piobuf_orig + spcl_off);
1336         }
1337         qib_sendbuf_done(dd, pbufn);
1338         if (qp->s_rdma_mr) {
1339                 qib_put_mr(qp->s_rdma_mr);
1340                 qp->s_rdma_mr = NULL;
1341         }
1342         if (qp->s_wqe) {
1343                 spin_lock_irqsave(&qp->s_lock, flags);
1344                 qib_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1345                 spin_unlock_irqrestore(&qp->s_lock, flags);
1346         } else if (qp->ibqp.qp_type == IB_QPT_RC) {
1347                 spin_lock_irqsave(&qp->s_lock, flags);
1348                 qib_rc_send_complete(qp, ibhdr);
1349                 spin_unlock_irqrestore(&qp->s_lock, flags);
1350         }
1351         return 0;
1352 }
1353
1354 /**
1355  * qib_verbs_send - send a packet
1356  * @qp: the QP to send on
1357  * @hdr: the packet header
1358  * @hdrwords: the number of 32-bit words in the header
1359  * @ss: the SGE to send
1360  * @len: the length of the packet in bytes
1361  *
1362  * Return zero if packet is sent or queued OK.
1363  * Return non-zero and clear qp->s_flags QIB_S_BUSY otherwise.
1364  */
1365 int qib_verbs_send(struct qib_qp *qp, struct qib_ib_header *hdr,
1366                    u32 hdrwords, struct qib_sge_state *ss, u32 len)
1367 {
1368         struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1369         u32 plen;
1370         int ret;
1371         u32 dwords = (len + 3) >> 2;
1372
1373         /*
1374          * Calculate the send buffer trigger address.
1375          * The +1 counts for the pbc control dword following the pbc length.
1376          */
1377         plen = hdrwords + dwords + 1;
1378
1379         /*
1380          * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1381          * can defer SDMA restart until link goes ACTIVE without
1382          * worrying about just how we got there.
1383          */
1384         if (qp->ibqp.qp_type == IB_QPT_SMI ||
1385             !(dd->flags & QIB_HAS_SEND_DMA))
1386                 ret = qib_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1387                                          plen, dwords);
1388         else
1389                 ret = qib_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1390                                          plen, dwords);
1391
1392         return ret;
1393 }
1394
1395 int qib_snapshot_counters(struct qib_pportdata *ppd, u64 *swords,
1396                           u64 *rwords, u64 *spkts, u64 *rpkts,
1397                           u64 *xmit_wait)
1398 {
1399         int ret;
1400         struct qib_devdata *dd = ppd->dd;
1401
1402         if (!(dd->flags & QIB_PRESENT)) {
1403                 /* no hardware, freeze, etc. */
1404                 ret = -EINVAL;
1405                 goto bail;
1406         }
1407         *swords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDSEND);
1408         *rwords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDRCV);
1409         *spkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTSEND);
1410         *rpkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTRCV);
1411         *xmit_wait = dd->f_portcntr(ppd, QIBPORTCNTR_SENDSTALL);
1412
1413         ret = 0;
1414
1415 bail:
1416         return ret;
1417 }
1418
1419 /**
1420  * qib_get_counters - get various chip counters
1421  * @dd: the qlogic_ib device
1422  * @cntrs: counters are placed here
1423  *
1424  * Return the counters needed by recv_pma_get_portcounters().
1425  */
1426 int qib_get_counters(struct qib_pportdata *ppd,
1427                      struct qib_verbs_counters *cntrs)
1428 {
1429         int ret;
1430
1431         if (!(ppd->dd->flags & QIB_PRESENT)) {
1432                 /* no hardware, freeze, etc. */
1433                 ret = -EINVAL;
1434                 goto bail;
1435         }
1436         cntrs->symbol_error_counter =
1437                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBSYMBOLERR);
1438         cntrs->link_error_recovery_counter =
1439                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKERRRECOV);
1440         /*
1441          * The link downed counter counts when the other side downs the
1442          * connection.  We add in the number of times we downed the link
1443          * due to local link integrity errors to compensate.
1444          */
1445         cntrs->link_downed_counter =
1446                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKDOWN);
1447         cntrs->port_rcv_errors =
1448                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXDROPPKT) +
1449                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVOVFL) +
1450                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERR_RLEN) +
1451                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_INVALIDRLEN) +
1452                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLINK) +
1453                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRICRC) +
1454                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRVCRC) +
1455                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLPCRC) +
1456                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_BADFORMAT);
1457         cntrs->port_rcv_errors +=
1458                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXLOCALPHYERR);
1459         cntrs->port_rcv_errors +=
1460                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXVLERR);
1461         cntrs->port_rcv_remphys_errors =
1462                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVEBP);
1463         cntrs->port_xmit_discards =
1464                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_UNSUPVL);
1465         cntrs->port_xmit_data = ppd->dd->f_portcntr(ppd,
1466                         QIBPORTCNTR_WORDSEND);
1467         cntrs->port_rcv_data = ppd->dd->f_portcntr(ppd,
1468                         QIBPORTCNTR_WORDRCV);
1469         cntrs->port_xmit_packets = ppd->dd->f_portcntr(ppd,
1470                         QIBPORTCNTR_PKTSEND);
1471         cntrs->port_rcv_packets = ppd->dd->f_portcntr(ppd,
1472                         QIBPORTCNTR_PKTRCV);
1473         cntrs->local_link_integrity_errors =
1474                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_LLI);
1475         cntrs->excessive_buffer_overrun_errors =
1476                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_EXCESSBUFOVFL);
1477         cntrs->vl15_dropped =
1478                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_VL15PKTDROP);
1479
1480         ret = 0;
1481
1482 bail:
1483         return ret;
1484 }
1485
1486 /**
1487  * qib_ib_piobufavail - callback when a PIO buffer is available
1488  * @dd: the device pointer
1489  *
1490  * This is called from qib_intr() at interrupt level when a PIO buffer is
1491  * available after qib_verbs_send() returned an error that no buffers were
1492  * available. Disable the interrupt if there are no more QPs waiting.
1493  */
1494 void qib_ib_piobufavail(struct qib_devdata *dd)
1495 {
1496         struct qib_ibdev *dev = &dd->verbs_dev;
1497         struct list_head *list;
1498         struct qib_qp *qps[5];
1499         struct qib_qp *qp;
1500         unsigned long flags;
1501         unsigned i, n;
1502
1503         list = &dev->piowait;
1504         n = 0;
1505
1506         /*
1507          * Note: checking that the piowait list is empty and clearing
1508          * the buffer available interrupt needs to be atomic or we
1509          * could end up with QPs on the wait list with the interrupt
1510          * disabled.
1511          */
1512         spin_lock_irqsave(&dev->pending_lock, flags);
1513         while (!list_empty(list)) {
1514                 if (n == ARRAY_SIZE(qps))
1515                         goto full;
1516                 qp = list_entry(list->next, struct qib_qp, iowait);
1517                 list_del_init(&qp->iowait);
1518                 atomic_inc(&qp->refcount);
1519                 qps[n++] = qp;
1520         }
1521         dd->f_wantpiobuf_intr(dd, 0);
1522 full:
1523         spin_unlock_irqrestore(&dev->pending_lock, flags);
1524
1525         for (i = 0; i < n; i++) {
1526                 qp = qps[i];
1527
1528                 spin_lock_irqsave(&qp->s_lock, flags);
1529                 if (qp->s_flags & QIB_S_WAIT_PIO) {
1530                         qp->s_flags &= ~QIB_S_WAIT_PIO;
1531                         qib_schedule_send(qp);
1532                 }
1533                 spin_unlock_irqrestore(&qp->s_lock, flags);
1534
1535                 /* Notify qib_destroy_qp() if it is waiting. */
1536                 if (atomic_dec_and_test(&qp->refcount))
1537                         wake_up(&qp->wait);
1538         }
1539 }
1540
1541 static int qib_query_device(struct ib_device *ibdev,
1542                             struct ib_device_attr *props)
1543 {
1544         struct qib_devdata *dd = dd_from_ibdev(ibdev);
1545         struct qib_ibdev *dev = to_idev(ibdev);
1546
1547         memset(props, 0, sizeof(*props));
1548
1549         props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1550                 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1551                 IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1552                 IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1553         props->page_size_cap = PAGE_SIZE;
1554         props->vendor_id =
1555                 QIB_SRC_OUI_1 << 16 | QIB_SRC_OUI_2 << 8 | QIB_SRC_OUI_3;
1556         props->vendor_part_id = dd->deviceid;
1557         props->hw_ver = dd->minrev;
1558         props->sys_image_guid = ib_qib_sys_image_guid;
1559         props->max_mr_size = ~0ULL;
1560         props->max_qp = ib_qib_max_qps;
1561         props->max_qp_wr = ib_qib_max_qp_wrs;
1562         props->max_sge = ib_qib_max_sges;
1563         props->max_cq = ib_qib_max_cqs;
1564         props->max_ah = ib_qib_max_ahs;
1565         props->max_cqe = ib_qib_max_cqes;
1566         props->max_mr = dev->lk_table.max;
1567         props->max_fmr = dev->lk_table.max;
1568         props->max_map_per_fmr = 32767;
1569         props->max_pd = ib_qib_max_pds;
1570         props->max_qp_rd_atom = QIB_MAX_RDMA_ATOMIC;
1571         props->max_qp_init_rd_atom = 255;
1572         /* props->max_res_rd_atom */
1573         props->max_srq = ib_qib_max_srqs;
1574         props->max_srq_wr = ib_qib_max_srq_wrs;
1575         props->max_srq_sge = ib_qib_max_srq_sges;
1576         /* props->local_ca_ack_delay */
1577         props->atomic_cap = IB_ATOMIC_GLOB;
1578         props->max_pkeys = qib_get_npkeys(dd);
1579         props->max_mcast_grp = ib_qib_max_mcast_grps;
1580         props->max_mcast_qp_attach = ib_qib_max_mcast_qp_attached;
1581         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1582                 props->max_mcast_grp;
1583
1584         return 0;
1585 }
1586
1587 static int qib_query_port(struct ib_device *ibdev, u8 port,
1588                           struct ib_port_attr *props)
1589 {
1590         struct qib_devdata *dd = dd_from_ibdev(ibdev);
1591         struct qib_ibport *ibp = to_iport(ibdev, port);
1592         struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1593         enum ib_mtu mtu;
1594         u16 lid = ppd->lid;
1595
1596         memset(props, 0, sizeof(*props));
1597         props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1598         props->lmc = ppd->lmc;
1599         props->sm_lid = ibp->sm_lid;
1600         props->sm_sl = ibp->sm_sl;
1601         props->state = dd->f_iblink_state(ppd->lastibcstat);
1602         props->phys_state = dd->f_ibphys_portstate(ppd->lastibcstat);
1603         props->port_cap_flags = ibp->port_cap_flags;
1604         props->gid_tbl_len = QIB_GUIDS_PER_PORT;
1605         props->max_msg_sz = 0x80000000;
1606         props->pkey_tbl_len = qib_get_npkeys(dd);
1607         props->bad_pkey_cntr = ibp->pkey_violations;
1608         props->qkey_viol_cntr = ibp->qkey_violations;
1609         props->active_width = ppd->link_width_active;
1610         /* See rate_show() */
1611         props->active_speed = ppd->link_speed_active;
1612         props->max_vl_num = qib_num_vls(ppd->vls_supported);
1613         props->init_type_reply = 0;
1614
1615         props->max_mtu = qib_ibmtu ? qib_ibmtu : IB_MTU_4096;
1616         switch (ppd->ibmtu) {
1617         case 4096:
1618                 mtu = IB_MTU_4096;
1619                 break;
1620         case 2048:
1621                 mtu = IB_MTU_2048;
1622                 break;
1623         case 1024:
1624                 mtu = IB_MTU_1024;
1625                 break;
1626         case 512:
1627                 mtu = IB_MTU_512;
1628                 break;
1629         case 256:
1630                 mtu = IB_MTU_256;
1631                 break;
1632         default:
1633                 mtu = IB_MTU_2048;
1634         }
1635         props->active_mtu = mtu;
1636         props->subnet_timeout = ibp->subnet_timeout;
1637
1638         return 0;
1639 }
1640
1641 static int qib_modify_device(struct ib_device *device,
1642                              int device_modify_mask,
1643                              struct ib_device_modify *device_modify)
1644 {
1645         struct qib_devdata *dd = dd_from_ibdev(device);
1646         unsigned i;
1647         int ret;
1648
1649         if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1650                                    IB_DEVICE_MODIFY_NODE_DESC)) {
1651                 ret = -EOPNOTSUPP;
1652                 goto bail;
1653         }
1654
1655         if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1656                 memcpy(device->node_desc, device_modify->node_desc, 64);
1657                 for (i = 0; i < dd->num_pports; i++) {
1658                         struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1659
1660                         qib_node_desc_chg(ibp);
1661                 }
1662         }
1663
1664         if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1665                 ib_qib_sys_image_guid =
1666                         cpu_to_be64(device_modify->sys_image_guid);
1667                 for (i = 0; i < dd->num_pports; i++) {
1668                         struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1669
1670                         qib_sys_guid_chg(ibp);
1671                 }
1672         }
1673
1674         ret = 0;
1675
1676 bail:
1677         return ret;
1678 }
1679
1680 static int qib_modify_port(struct ib_device *ibdev, u8 port,
1681                            int port_modify_mask, struct ib_port_modify *props)
1682 {
1683         struct qib_ibport *ibp = to_iport(ibdev, port);
1684         struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1685
1686         ibp->port_cap_flags |= props->set_port_cap_mask;
1687         ibp->port_cap_flags &= ~props->clr_port_cap_mask;
1688         if (props->set_port_cap_mask || props->clr_port_cap_mask)
1689                 qib_cap_mask_chg(ibp);
1690         if (port_modify_mask & IB_PORT_SHUTDOWN)
1691                 qib_set_linkstate(ppd, QIB_IB_LINKDOWN);
1692         if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1693                 ibp->qkey_violations = 0;
1694         return 0;
1695 }
1696
1697 static int qib_query_gid(struct ib_device *ibdev, u8 port,
1698                          int index, union ib_gid *gid)
1699 {
1700         struct qib_devdata *dd = dd_from_ibdev(ibdev);
1701         int ret = 0;
1702
1703         if (!port || port > dd->num_pports)
1704                 ret = -EINVAL;
1705         else {
1706                 struct qib_ibport *ibp = to_iport(ibdev, port);
1707                 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1708
1709                 gid->global.subnet_prefix = ibp->gid_prefix;
1710                 if (index == 0)
1711                         gid->global.interface_id = ppd->guid;
1712                 else if (index < QIB_GUIDS_PER_PORT)
1713                         gid->global.interface_id = ibp->guids[index - 1];
1714                 else
1715                         ret = -EINVAL;
1716         }
1717
1718         return ret;
1719 }
1720
1721 static struct ib_pd *qib_alloc_pd(struct ib_device *ibdev,
1722                                   struct ib_ucontext *context,
1723                                   struct ib_udata *udata)
1724 {
1725         struct qib_ibdev *dev = to_idev(ibdev);
1726         struct qib_pd *pd;
1727         struct ib_pd *ret;
1728
1729         /*
1730          * This is actually totally arbitrary.  Some correctness tests
1731          * assume there's a maximum number of PDs that can be allocated.
1732          * We don't actually have this limit, but we fail the test if
1733          * we allow allocations of more than we report for this value.
1734          */
1735
1736         pd = kmalloc(sizeof *pd, GFP_KERNEL);
1737         if (!pd) {
1738                 ret = ERR_PTR(-ENOMEM);
1739                 goto bail;
1740         }
1741
1742         spin_lock(&dev->n_pds_lock);
1743         if (dev->n_pds_allocated == ib_qib_max_pds) {
1744                 spin_unlock(&dev->n_pds_lock);
1745                 kfree(pd);
1746                 ret = ERR_PTR(-ENOMEM);
1747                 goto bail;
1748         }
1749
1750         dev->n_pds_allocated++;
1751         spin_unlock(&dev->n_pds_lock);
1752
1753         /* ib_alloc_pd() will initialize pd->ibpd. */
1754         pd->user = udata != NULL;
1755
1756         ret = &pd->ibpd;
1757
1758 bail:
1759         return ret;
1760 }
1761
1762 static int qib_dealloc_pd(struct ib_pd *ibpd)
1763 {
1764         struct qib_pd *pd = to_ipd(ibpd);
1765         struct qib_ibdev *dev = to_idev(ibpd->device);
1766
1767         spin_lock(&dev->n_pds_lock);
1768         dev->n_pds_allocated--;
1769         spin_unlock(&dev->n_pds_lock);
1770
1771         kfree(pd);
1772
1773         return 0;
1774 }
1775
1776 int qib_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr)
1777 {
1778         /* A multicast address requires a GRH (see ch. 8.4.1). */
1779         if (ah_attr->dlid >= QIB_MULTICAST_LID_BASE &&
1780             ah_attr->dlid != QIB_PERMISSIVE_LID &&
1781             !(ah_attr->ah_flags & IB_AH_GRH))
1782                 goto bail;
1783         if ((ah_attr->ah_flags & IB_AH_GRH) &&
1784             ah_attr->grh.sgid_index >= QIB_GUIDS_PER_PORT)
1785                 goto bail;
1786         if (ah_attr->dlid == 0)
1787                 goto bail;
1788         if (ah_attr->port_num < 1 ||
1789             ah_attr->port_num > ibdev->phys_port_cnt)
1790                 goto bail;
1791         if (ah_attr->static_rate != IB_RATE_PORT_CURRENT &&
1792             ib_rate_to_mult(ah_attr->static_rate) < 0)
1793                 goto bail;
1794         if (ah_attr->sl > 15)
1795                 goto bail;
1796         return 0;
1797 bail:
1798         return -EINVAL;
1799 }
1800
1801 /**
1802  * qib_create_ah - create an address handle
1803  * @pd: the protection domain
1804  * @ah_attr: the attributes of the AH
1805  *
1806  * This may be called from interrupt context.
1807  */
1808 static struct ib_ah *qib_create_ah(struct ib_pd *pd,
1809                                    struct ib_ah_attr *ah_attr)
1810 {
1811         struct qib_ah *ah;
1812         struct ib_ah *ret;
1813         struct qib_ibdev *dev = to_idev(pd->device);
1814         unsigned long flags;
1815
1816         if (qib_check_ah(pd->device, ah_attr)) {
1817                 ret = ERR_PTR(-EINVAL);
1818                 goto bail;
1819         }
1820
1821         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
1822         if (!ah) {
1823                 ret = ERR_PTR(-ENOMEM);
1824                 goto bail;
1825         }
1826
1827         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1828         if (dev->n_ahs_allocated == ib_qib_max_ahs) {
1829                 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1830                 kfree(ah);
1831                 ret = ERR_PTR(-ENOMEM);
1832                 goto bail;
1833         }
1834
1835         dev->n_ahs_allocated++;
1836         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1837
1838         /* ib_create_ah() will initialize ah->ibah. */
1839         ah->attr = *ah_attr;
1840         atomic_set(&ah->refcount, 0);
1841
1842         ret = &ah->ibah;
1843
1844 bail:
1845         return ret;
1846 }
1847
1848 struct ib_ah *qib_create_qp0_ah(struct qib_ibport *ibp, u16 dlid)
1849 {
1850         struct ib_ah_attr attr;
1851         struct ib_ah *ah = ERR_PTR(-EINVAL);
1852         struct qib_qp *qp0;
1853
1854         memset(&attr, 0, sizeof attr);
1855         attr.dlid = dlid;
1856         attr.port_num = ppd_from_ibp(ibp)->port;
1857         rcu_read_lock();
1858         qp0 = rcu_dereference(ibp->qp0);
1859         if (qp0)
1860                 ah = ib_create_ah(qp0->ibqp.pd, &attr);
1861         rcu_read_unlock();
1862         return ah;
1863 }
1864
1865 /**
1866  * qib_destroy_ah - destroy an address handle
1867  * @ibah: the AH to destroy
1868  *
1869  * This may be called from interrupt context.
1870  */
1871 static int qib_destroy_ah(struct ib_ah *ibah)
1872 {
1873         struct qib_ibdev *dev = to_idev(ibah->device);
1874         struct qib_ah *ah = to_iah(ibah);
1875         unsigned long flags;
1876
1877         if (atomic_read(&ah->refcount) != 0)
1878                 return -EBUSY;
1879
1880         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1881         dev->n_ahs_allocated--;
1882         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1883
1884         kfree(ah);
1885
1886         return 0;
1887 }
1888
1889 static int qib_modify_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1890 {
1891         struct qib_ah *ah = to_iah(ibah);
1892
1893         if (qib_check_ah(ibah->device, ah_attr))
1894                 return -EINVAL;
1895
1896         ah->attr = *ah_attr;
1897
1898         return 0;
1899 }
1900
1901 static int qib_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1902 {
1903         struct qib_ah *ah = to_iah(ibah);
1904
1905         *ah_attr = ah->attr;
1906
1907         return 0;
1908 }
1909
1910 /**
1911  * qib_get_npkeys - return the size of the PKEY table for context 0
1912  * @dd: the qlogic_ib device
1913  */
1914 unsigned qib_get_npkeys(struct qib_devdata *dd)
1915 {
1916         return ARRAY_SIZE(dd->rcd[0]->pkeys);
1917 }
1918
1919 /*
1920  * Return the indexed PKEY from the port PKEY table.
1921  * No need to validate rcd[ctxt]; the port is setup if we are here.
1922  */
1923 unsigned qib_get_pkey(struct qib_ibport *ibp, unsigned index)
1924 {
1925         struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1926         struct qib_devdata *dd = ppd->dd;
1927         unsigned ctxt = ppd->hw_pidx;
1928         unsigned ret;
1929
1930         /* dd->rcd null if mini_init or some init failures */
1931         if (!dd->rcd || index >= ARRAY_SIZE(dd->rcd[ctxt]->pkeys))
1932                 ret = 0;
1933         else
1934                 ret = dd->rcd[ctxt]->pkeys[index];
1935
1936         return ret;
1937 }
1938
1939 static int qib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1940                           u16 *pkey)
1941 {
1942         struct qib_devdata *dd = dd_from_ibdev(ibdev);
1943         int ret;
1944
1945         if (index >= qib_get_npkeys(dd)) {
1946                 ret = -EINVAL;
1947                 goto bail;
1948         }
1949
1950         *pkey = qib_get_pkey(to_iport(ibdev, port), index);
1951         ret = 0;
1952
1953 bail:
1954         return ret;
1955 }
1956
1957 /**
1958  * qib_alloc_ucontext - allocate a ucontest
1959  * @ibdev: the infiniband device
1960  * @udata: not used by the QLogic_IB driver
1961  */
1962
1963 static struct ib_ucontext *qib_alloc_ucontext(struct ib_device *ibdev,
1964                                               struct ib_udata *udata)
1965 {
1966         struct qib_ucontext *context;
1967         struct ib_ucontext *ret;
1968
1969         context = kmalloc(sizeof *context, GFP_KERNEL);
1970         if (!context) {
1971                 ret = ERR_PTR(-ENOMEM);
1972                 goto bail;
1973         }
1974
1975         ret = &context->ibucontext;
1976
1977 bail:
1978         return ret;
1979 }
1980
1981 static int qib_dealloc_ucontext(struct ib_ucontext *context)
1982 {
1983         kfree(to_iucontext(context));
1984         return 0;
1985 }
1986
1987 static void init_ibport(struct qib_pportdata *ppd)
1988 {
1989         struct qib_verbs_counters cntrs;
1990         struct qib_ibport *ibp = &ppd->ibport_data;
1991
1992         spin_lock_init(&ibp->lock);
1993         /* Set the prefix to the default value (see ch. 4.1.1) */
1994         ibp->gid_prefix = IB_DEFAULT_GID_PREFIX;
1995         ibp->sm_lid = be16_to_cpu(IB_LID_PERMISSIVE);
1996         ibp->port_cap_flags = IB_PORT_SYS_IMAGE_GUID_SUP |
1997                 IB_PORT_CLIENT_REG_SUP | IB_PORT_SL_MAP_SUP |
1998                 IB_PORT_TRAP_SUP | IB_PORT_AUTO_MIGR_SUP |
1999                 IB_PORT_DR_NOTICE_SUP | IB_PORT_CAP_MASK_NOTICE_SUP |
2000                 IB_PORT_OTHER_LOCAL_CHANGES_SUP;
2001         if (ppd->dd->flags & QIB_HAS_LINK_LATENCY)
2002                 ibp->port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
2003         ibp->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
2004         ibp->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
2005         ibp->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
2006         ibp->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
2007         ibp->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
2008
2009         /* Snapshot current HW counters to "clear" them. */
2010         qib_get_counters(ppd, &cntrs);
2011         ibp->z_symbol_error_counter = cntrs.symbol_error_counter;
2012         ibp->z_link_error_recovery_counter =
2013                 cntrs.link_error_recovery_counter;
2014         ibp->z_link_downed_counter = cntrs.link_downed_counter;
2015         ibp->z_port_rcv_errors = cntrs.port_rcv_errors;
2016         ibp->z_port_rcv_remphys_errors = cntrs.port_rcv_remphys_errors;
2017         ibp->z_port_xmit_discards = cntrs.port_xmit_discards;
2018         ibp->z_port_xmit_data = cntrs.port_xmit_data;
2019         ibp->z_port_rcv_data = cntrs.port_rcv_data;
2020         ibp->z_port_xmit_packets = cntrs.port_xmit_packets;
2021         ibp->z_port_rcv_packets = cntrs.port_rcv_packets;
2022         ibp->z_local_link_integrity_errors =
2023                 cntrs.local_link_integrity_errors;
2024         ibp->z_excessive_buffer_overrun_errors =
2025                 cntrs.excessive_buffer_overrun_errors;
2026         ibp->z_vl15_dropped = cntrs.vl15_dropped;
2027         RCU_INIT_POINTER(ibp->qp0, NULL);
2028         RCU_INIT_POINTER(ibp->qp1, NULL);
2029 }
2030
2031 /**
2032  * qib_register_ib_device - register our device with the infiniband core
2033  * @dd: the device data structure
2034  * Return the allocated qib_ibdev pointer or NULL on error.
2035  */
2036 int qib_register_ib_device(struct qib_devdata *dd)
2037 {
2038         struct qib_ibdev *dev = &dd->verbs_dev;
2039         struct ib_device *ibdev = &dev->ibdev;
2040         struct qib_pportdata *ppd = dd->pport;
2041         unsigned i, lk_tab_size;
2042         int ret;
2043
2044         dev->qp_table_size = ib_qib_qp_table_size;
2045         get_random_bytes(&dev->qp_rnd, sizeof(dev->qp_rnd));
2046         dev->qp_table = kmalloc(dev->qp_table_size * sizeof *dev->qp_table,
2047                                 GFP_KERNEL);
2048         if (!dev->qp_table) {
2049                 ret = -ENOMEM;
2050                 goto err_qpt;
2051         }
2052         for (i = 0; i < dev->qp_table_size; i++)
2053                 RCU_INIT_POINTER(dev->qp_table[i], NULL);
2054
2055         for (i = 0; i < dd->num_pports; i++)
2056                 init_ibport(ppd + i);
2057
2058         /* Only need to initialize non-zero fields. */
2059         spin_lock_init(&dev->qpt_lock);
2060         spin_lock_init(&dev->n_pds_lock);
2061         spin_lock_init(&dev->n_ahs_lock);
2062         spin_lock_init(&dev->n_cqs_lock);
2063         spin_lock_init(&dev->n_qps_lock);
2064         spin_lock_init(&dev->n_srqs_lock);
2065         spin_lock_init(&dev->n_mcast_grps_lock);
2066         init_timer(&dev->mem_timer);
2067         dev->mem_timer.function = mem_timer;
2068         dev->mem_timer.data = (unsigned long) dev;
2069
2070         qib_init_qpn_table(dd, &dev->qpn_table);
2071
2072         /*
2073          * The top ib_qib_lkey_table_size bits are used to index the
2074          * table.  The lower 8 bits can be owned by the user (copied from
2075          * the LKEY).  The remaining bits act as a generation number or tag.
2076          */
2077         spin_lock_init(&dev->lk_table.lock);
2078         dev->lk_table.max = 1 << ib_qib_lkey_table_size;
2079         lk_tab_size = dev->lk_table.max * sizeof(*dev->lk_table.table);
2080         dev->lk_table.table = (struct qib_mregion __rcu **)
2081                 __get_free_pages(GFP_KERNEL, get_order(lk_tab_size));
2082         if (dev->lk_table.table == NULL) {
2083                 ret = -ENOMEM;
2084                 goto err_lk;
2085         }
2086         RCU_INIT_POINTER(dev->dma_mr, NULL);
2087         for (i = 0; i < dev->lk_table.max; i++)
2088                 RCU_INIT_POINTER(dev->lk_table.table[i], NULL);
2089         INIT_LIST_HEAD(&dev->pending_mmaps);
2090         spin_lock_init(&dev->pending_lock);
2091         dev->mmap_offset = PAGE_SIZE;
2092         spin_lock_init(&dev->mmap_offset_lock);
2093         INIT_LIST_HEAD(&dev->piowait);
2094         INIT_LIST_HEAD(&dev->dmawait);
2095         INIT_LIST_HEAD(&dev->txwait);
2096         INIT_LIST_HEAD(&dev->memwait);
2097         INIT_LIST_HEAD(&dev->txreq_free);
2098
2099         if (ppd->sdma_descq_cnt) {
2100                 dev->pio_hdrs = dma_alloc_coherent(&dd->pcidev->dev,
2101                                                 ppd->sdma_descq_cnt *
2102                                                 sizeof(struct qib_pio_header),
2103                                                 &dev->pio_hdrs_phys,
2104                                                 GFP_KERNEL);
2105                 if (!dev->pio_hdrs) {
2106                         ret = -ENOMEM;
2107                         goto err_hdrs;
2108                 }
2109         }
2110
2111         for (i = 0; i < ppd->sdma_descq_cnt; i++) {
2112                 struct qib_verbs_txreq *tx;
2113
2114                 tx = kzalloc(sizeof *tx, GFP_KERNEL);
2115                 if (!tx) {
2116                         ret = -ENOMEM;
2117                         goto err_tx;
2118                 }
2119                 tx->hdr_inx = i;
2120                 list_add(&tx->txreq.list, &dev->txreq_free);
2121         }
2122
2123         /*
2124          * The system image GUID is supposed to be the same for all
2125          * IB HCAs in a single system but since there can be other
2126          * device types in the system, we can't be sure this is unique.
2127          */
2128         if (!ib_qib_sys_image_guid)
2129                 ib_qib_sys_image_guid = ppd->guid;
2130
2131         strlcpy(ibdev->name, "qib%d", IB_DEVICE_NAME_MAX);
2132         ibdev->owner = THIS_MODULE;
2133         ibdev->node_guid = ppd->guid;
2134         ibdev->uverbs_abi_ver = QIB_UVERBS_ABI_VERSION;
2135         ibdev->uverbs_cmd_mask =
2136                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
2137                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
2138                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
2139                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
2140                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
2141                 (1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
2142                 (1ull << IB_USER_VERBS_CMD_MODIFY_AH)           |
2143                 (1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
2144                 (1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
2145                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
2146                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
2147                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2148                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
2149                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
2150                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
2151                 (1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
2152                 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
2153                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
2154                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
2155                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
2156                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
2157                 (1ull << IB_USER_VERBS_CMD_POST_SEND)           |
2158                 (1ull << IB_USER_VERBS_CMD_POST_RECV)           |
2159                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
2160                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
2161                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
2162                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
2163                 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
2164                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
2165                 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
2166         ibdev->node_type = RDMA_NODE_IB_CA;
2167         ibdev->phys_port_cnt = dd->num_pports;
2168         ibdev->num_comp_vectors = 1;
2169         ibdev->dma_device = &dd->pcidev->dev;
2170         ibdev->query_device = qib_query_device;
2171         ibdev->modify_device = qib_modify_device;
2172         ibdev->query_port = qib_query_port;
2173         ibdev->modify_port = qib_modify_port;
2174         ibdev->query_pkey = qib_query_pkey;
2175         ibdev->query_gid = qib_query_gid;
2176         ibdev->alloc_ucontext = qib_alloc_ucontext;
2177         ibdev->dealloc_ucontext = qib_dealloc_ucontext;
2178         ibdev->alloc_pd = qib_alloc_pd;
2179         ibdev->dealloc_pd = qib_dealloc_pd;
2180         ibdev->create_ah = qib_create_ah;
2181         ibdev->destroy_ah = qib_destroy_ah;
2182         ibdev->modify_ah = qib_modify_ah;
2183         ibdev->query_ah = qib_query_ah;
2184         ibdev->create_srq = qib_create_srq;
2185         ibdev->modify_srq = qib_modify_srq;
2186         ibdev->query_srq = qib_query_srq;
2187         ibdev->destroy_srq = qib_destroy_srq;
2188         ibdev->create_qp = qib_create_qp;
2189         ibdev->modify_qp = qib_modify_qp;
2190         ibdev->query_qp = qib_query_qp;
2191         ibdev->destroy_qp = qib_destroy_qp;
2192         ibdev->post_send = qib_post_send;
2193         ibdev->post_recv = qib_post_receive;
2194         ibdev->post_srq_recv = qib_post_srq_receive;
2195         ibdev->create_cq = qib_create_cq;
2196         ibdev->destroy_cq = qib_destroy_cq;
2197         ibdev->resize_cq = qib_resize_cq;
2198         ibdev->poll_cq = qib_poll_cq;
2199         ibdev->req_notify_cq = qib_req_notify_cq;
2200         ibdev->get_dma_mr = qib_get_dma_mr;
2201         ibdev->reg_phys_mr = qib_reg_phys_mr;
2202         ibdev->reg_user_mr = qib_reg_user_mr;
2203         ibdev->dereg_mr = qib_dereg_mr;
2204         ibdev->alloc_fast_reg_mr = qib_alloc_fast_reg_mr;
2205         ibdev->alloc_fast_reg_page_list = qib_alloc_fast_reg_page_list;
2206         ibdev->free_fast_reg_page_list = qib_free_fast_reg_page_list;
2207         ibdev->alloc_fmr = qib_alloc_fmr;
2208         ibdev->map_phys_fmr = qib_map_phys_fmr;
2209         ibdev->unmap_fmr = qib_unmap_fmr;
2210         ibdev->dealloc_fmr = qib_dealloc_fmr;
2211         ibdev->attach_mcast = qib_multicast_attach;
2212         ibdev->detach_mcast = qib_multicast_detach;
2213         ibdev->process_mad = qib_process_mad;
2214         ibdev->mmap = qib_mmap;
2215         ibdev->dma_ops = &qib_dma_mapping_ops;
2216
2217         snprintf(ibdev->node_desc, sizeof(ibdev->node_desc),
2218                  QIB_IDSTR " %s", init_utsname()->nodename);
2219
2220         ret = ib_register_device(ibdev, qib_create_port_files);
2221         if (ret)
2222                 goto err_reg;
2223
2224         ret = qib_create_agents(dev);
2225         if (ret)
2226                 goto err_agents;
2227
2228         if (qib_verbs_register_sysfs(dd))
2229                 goto err_class;
2230
2231         goto bail;
2232
2233 err_class:
2234         qib_free_agents(dev);
2235 err_agents:
2236         ib_unregister_device(ibdev);
2237 err_reg:
2238 err_tx:
2239         while (!list_empty(&dev->txreq_free)) {
2240                 struct list_head *l = dev->txreq_free.next;
2241                 struct qib_verbs_txreq *tx;
2242
2243                 list_del(l);
2244                 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
2245                 kfree(tx);
2246         }
2247         if (ppd->sdma_descq_cnt)
2248                 dma_free_coherent(&dd->pcidev->dev,
2249                                   ppd->sdma_descq_cnt *
2250                                         sizeof(struct qib_pio_header),
2251                                   dev->pio_hdrs, dev->pio_hdrs_phys);
2252 err_hdrs:
2253         free_pages((unsigned long) dev->lk_table.table, get_order(lk_tab_size));
2254 err_lk:
2255         kfree(dev->qp_table);
2256 err_qpt:
2257         qib_dev_err(dd, "cannot register verbs: %d!\n", -ret);
2258 bail:
2259         return ret;
2260 }
2261
2262 void qib_unregister_ib_device(struct qib_devdata *dd)
2263 {
2264         struct qib_ibdev *dev = &dd->verbs_dev;
2265         struct ib_device *ibdev = &dev->ibdev;
2266         u32 qps_inuse;
2267         unsigned lk_tab_size;
2268
2269         qib_verbs_unregister_sysfs(dd);
2270
2271         qib_free_agents(dev);
2272
2273         ib_unregister_device(ibdev);
2274
2275         if (!list_empty(&dev->piowait))
2276                 qib_dev_err(dd, "piowait list not empty!\n");
2277         if (!list_empty(&dev->dmawait))
2278                 qib_dev_err(dd, "dmawait list not empty!\n");
2279         if (!list_empty(&dev->txwait))
2280                 qib_dev_err(dd, "txwait list not empty!\n");
2281         if (!list_empty(&dev->memwait))
2282                 qib_dev_err(dd, "memwait list not empty!\n");
2283         if (dev->dma_mr)
2284                 qib_dev_err(dd, "DMA MR not NULL!\n");
2285
2286         qps_inuse = qib_free_all_qps(dd);
2287         if (qps_inuse)
2288                 qib_dev_err(dd, "QP memory leak! %u still in use\n",
2289                             qps_inuse);
2290
2291         del_timer_sync(&dev->mem_timer);
2292         qib_free_qpn_table(&dev->qpn_table);
2293         while (!list_empty(&dev->txreq_free)) {
2294                 struct list_head *l = dev->txreq_free.next;
2295                 struct qib_verbs_txreq *tx;
2296
2297                 list_del(l);
2298                 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
2299                 kfree(tx);
2300         }
2301         if (dd->pport->sdma_descq_cnt)
2302                 dma_free_coherent(&dd->pcidev->dev,
2303                                   dd->pport->sdma_descq_cnt *
2304                                         sizeof(struct qib_pio_header),
2305                                   dev->pio_hdrs, dev->pio_hdrs_phys);
2306         lk_tab_size = dev->lk_table.max * sizeof(*dev->lk_table.table);
2307         free_pages((unsigned long) dev->lk_table.table,
2308                    get_order(lk_tab_size));
2309         kfree(dev->qp_table);
2310 }