]> git.karo-electronics.de Git - karo-tx-linux.git/blob - drivers/infiniband/hw/qib/qib_qp.c
dm, dax: Make sure dm_dax_flush() is called if device supports it
[karo-tx-linux.git] / drivers / infiniband / hw / qib / qib_qp.c
1 /*
2  * Copyright (c) 2012, 2013 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 <linux/err.h>
36 #include <linux/vmalloc.h>
37 #include <rdma/rdma_vt.h>
38 #ifdef CONFIG_DEBUG_FS
39 #include <linux/seq_file.h>
40 #endif
41
42 #include "qib.h"
43
44 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
45                               struct rvt_qpn_map *map, unsigned off)
46 {
47         return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
48 }
49
50 static inline unsigned find_next_offset(struct rvt_qpn_table *qpt,
51                                         struct rvt_qpn_map *map, unsigned off,
52                                         unsigned n, u16 qpt_mask)
53 {
54         if (qpt_mask) {
55                 off++;
56                 if (((off & qpt_mask) >> 1) >= n)
57                         off = (off | qpt_mask) + 2;
58         } else {
59                 off = find_next_zero_bit(map->page, RVT_BITS_PER_PAGE, off);
60         }
61         return off;
62 }
63
64 const struct rvt_operation_params qib_post_parms[RVT_OPERATION_MAX] = {
65 [IB_WR_RDMA_WRITE] = {
66         .length = sizeof(struct ib_rdma_wr),
67         .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
68 },
69
70 [IB_WR_RDMA_READ] = {
71         .length = sizeof(struct ib_rdma_wr),
72         .qpt_support = BIT(IB_QPT_RC),
73         .flags = RVT_OPERATION_ATOMIC,
74 },
75
76 [IB_WR_ATOMIC_CMP_AND_SWP] = {
77         .length = sizeof(struct ib_atomic_wr),
78         .qpt_support = BIT(IB_QPT_RC),
79         .flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
80 },
81
82 [IB_WR_ATOMIC_FETCH_AND_ADD] = {
83         .length = sizeof(struct ib_atomic_wr),
84         .qpt_support = BIT(IB_QPT_RC),
85         .flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
86 },
87
88 [IB_WR_RDMA_WRITE_WITH_IMM] = {
89         .length = sizeof(struct ib_rdma_wr),
90         .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
91 },
92
93 [IB_WR_SEND] = {
94         .length = sizeof(struct ib_send_wr),
95         .qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
96                        BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
97 },
98
99 [IB_WR_SEND_WITH_IMM] = {
100         .length = sizeof(struct ib_send_wr),
101         .qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
102                        BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
103 },
104
105 };
106
107 static void get_map_page(struct rvt_qpn_table *qpt, struct rvt_qpn_map *map,
108                          gfp_t gfp)
109 {
110         unsigned long page = get_zeroed_page(gfp);
111
112         /*
113          * Free the page if someone raced with us installing it.
114          */
115
116         spin_lock(&qpt->lock);
117         if (map->page)
118                 free_page(page);
119         else
120                 map->page = (void *)page;
121         spin_unlock(&qpt->lock);
122 }
123
124 /*
125  * Allocate the next available QPN or
126  * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI.
127  */
128 int qib_alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
129                   enum ib_qp_type type, u8 port, gfp_t gfp)
130 {
131         u32 i, offset, max_scan, qpn;
132         struct rvt_qpn_map *map;
133         u32 ret;
134         struct qib_ibdev *verbs_dev = container_of(rdi, struct qib_ibdev, rdi);
135         struct qib_devdata *dd = container_of(verbs_dev, struct qib_devdata,
136                                               verbs_dev);
137         u16 qpt_mask = dd->qpn_mask;
138
139         if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
140                 unsigned n;
141
142                 ret = type == IB_QPT_GSI;
143                 n = 1 << (ret + 2 * (port - 1));
144                 spin_lock(&qpt->lock);
145                 if (qpt->flags & n)
146                         ret = -EINVAL;
147                 else
148                         qpt->flags |= n;
149                 spin_unlock(&qpt->lock);
150                 goto bail;
151         }
152
153         qpn = qpt->last + 2;
154         if (qpn >= RVT_QPN_MAX)
155                 qpn = 2;
156         if (qpt_mask && ((qpn & qpt_mask) >> 1) >= dd->n_krcv_queues)
157                 qpn = (qpn | qpt_mask) + 2;
158         offset = qpn & RVT_BITS_PER_PAGE_MASK;
159         map = &qpt->map[qpn / RVT_BITS_PER_PAGE];
160         max_scan = qpt->nmaps - !offset;
161         for (i = 0;;) {
162                 if (unlikely(!map->page)) {
163                         get_map_page(qpt, map, gfp);
164                         if (unlikely(!map->page))
165                                 break;
166                 }
167                 do {
168                         if (!test_and_set_bit(offset, map->page)) {
169                                 qpt->last = qpn;
170                                 ret = qpn;
171                                 goto bail;
172                         }
173                         offset = find_next_offset(qpt, map, offset,
174                                 dd->n_krcv_queues, qpt_mask);
175                         qpn = mk_qpn(qpt, map, offset);
176                         /*
177                          * This test differs from alloc_pidmap().
178                          * If find_next_offset() does find a zero
179                          * bit, we don't need to check for QPN
180                          * wrapping around past our starting QPN.
181                          * We just need to be sure we don't loop
182                          * forever.
183                          */
184                 } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX);
185                 /*
186                  * In order to keep the number of pages allocated to a
187                  * minimum, we scan the all existing pages before increasing
188                  * the size of the bitmap table.
189                  */
190                 if (++i > max_scan) {
191                         if (qpt->nmaps == RVT_QPNMAP_ENTRIES)
192                                 break;
193                         map = &qpt->map[qpt->nmaps++];
194                         offset = 0;
195                 } else if (map < &qpt->map[qpt->nmaps]) {
196                         ++map;
197                         offset = 0;
198                 } else {
199                         map = &qpt->map[0];
200                         offset = 2;
201                 }
202                 qpn = mk_qpn(qpt, map, offset);
203         }
204
205         ret = -ENOMEM;
206
207 bail:
208         return ret;
209 }
210
211 /**
212  * qib_free_all_qps - check for QPs still in use
213  */
214 unsigned qib_free_all_qps(struct rvt_dev_info *rdi)
215 {
216         struct qib_ibdev *verbs_dev = container_of(rdi, struct qib_ibdev, rdi);
217         struct qib_devdata *dd = container_of(verbs_dev, struct qib_devdata,
218                                               verbs_dev);
219         unsigned n, qp_inuse = 0;
220
221         for (n = 0; n < dd->num_pports; n++) {
222                 struct qib_ibport *ibp = &dd->pport[n].ibport_data;
223
224                 rcu_read_lock();
225                 if (rcu_dereference(ibp->rvp.qp[0]))
226                         qp_inuse++;
227                 if (rcu_dereference(ibp->rvp.qp[1]))
228                         qp_inuse++;
229                 rcu_read_unlock();
230         }
231         return qp_inuse;
232 }
233
234 void qib_notify_qp_reset(struct rvt_qp *qp)
235 {
236         struct qib_qp_priv *priv = qp->priv;
237
238         atomic_set(&priv->s_dma_busy, 0);
239 }
240
241 void qib_notify_error_qp(struct rvt_qp *qp)
242 {
243         struct qib_qp_priv *priv = qp->priv;
244         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
245
246         spin_lock(&dev->rdi.pending_lock);
247         if (!list_empty(&priv->iowait) && !(qp->s_flags & RVT_S_BUSY)) {
248                 qp->s_flags &= ~RVT_S_ANY_WAIT_IO;
249                 list_del_init(&priv->iowait);
250         }
251         spin_unlock(&dev->rdi.pending_lock);
252
253         if (!(qp->s_flags & RVT_S_BUSY)) {
254                 qp->s_hdrwords = 0;
255                 if (qp->s_rdma_mr) {
256                         rvt_put_mr(qp->s_rdma_mr);
257                         qp->s_rdma_mr = NULL;
258                 }
259                 if (priv->s_tx) {
260                         qib_put_txreq(priv->s_tx);
261                         priv->s_tx = NULL;
262                 }
263         }
264 }
265
266 static int mtu_to_enum(u32 mtu)
267 {
268         int enum_mtu;
269
270         switch (mtu) {
271         case 4096:
272                 enum_mtu = IB_MTU_4096;
273                 break;
274         case 2048:
275                 enum_mtu = IB_MTU_2048;
276                 break;
277         case 1024:
278                 enum_mtu = IB_MTU_1024;
279                 break;
280         case 512:
281                 enum_mtu = IB_MTU_512;
282                 break;
283         case 256:
284                 enum_mtu = IB_MTU_256;
285                 break;
286         default:
287                 enum_mtu = IB_MTU_2048;
288         }
289         return enum_mtu;
290 }
291
292 int qib_get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
293                            struct ib_qp_attr *attr)
294 {
295         int mtu, pmtu, pidx = qp->port_num - 1;
296         struct qib_ibdev *verbs_dev = container_of(rdi, struct qib_ibdev, rdi);
297         struct qib_devdata *dd = container_of(verbs_dev, struct qib_devdata,
298                                               verbs_dev);
299         mtu = ib_mtu_enum_to_int(attr->path_mtu);
300         if (mtu == -1)
301                 return -EINVAL;
302
303         if (mtu > dd->pport[pidx].ibmtu)
304                 pmtu = mtu_to_enum(dd->pport[pidx].ibmtu);
305         else
306                 pmtu = attr->path_mtu;
307         return pmtu;
308 }
309
310 int qib_mtu_to_path_mtu(u32 mtu)
311 {
312         return mtu_to_enum(mtu);
313 }
314
315 u32 qib_mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
316 {
317         return ib_mtu_enum_to_int(pmtu);
318 }
319
320 void *qib_qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp, gfp_t gfp)
321 {
322         struct qib_qp_priv *priv;
323
324         priv = kzalloc(sizeof(*priv), gfp);
325         if (!priv)
326                 return ERR_PTR(-ENOMEM);
327         priv->owner = qp;
328
329         priv->s_hdr = kzalloc(sizeof(*priv->s_hdr), gfp);
330         if (!priv->s_hdr) {
331                 kfree(priv);
332                 return ERR_PTR(-ENOMEM);
333         }
334         init_waitqueue_head(&priv->wait_dma);
335         INIT_WORK(&priv->s_work, _qib_do_send);
336         INIT_LIST_HEAD(&priv->iowait);
337
338         return priv;
339 }
340
341 void qib_qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
342 {
343         struct qib_qp_priv *priv = qp->priv;
344
345         kfree(priv->s_hdr);
346         kfree(priv);
347 }
348
349 void qib_stop_send_queue(struct rvt_qp *qp)
350 {
351         struct qib_qp_priv *priv = qp->priv;
352
353         cancel_work_sync(&priv->s_work);
354 }
355
356 void qib_quiesce_qp(struct rvt_qp *qp)
357 {
358         struct qib_qp_priv *priv = qp->priv;
359
360         wait_event(priv->wait_dma, !atomic_read(&priv->s_dma_busy));
361         if (priv->s_tx) {
362                 qib_put_txreq(priv->s_tx);
363                 priv->s_tx = NULL;
364         }
365 }
366
367 void qib_flush_qp_waiters(struct rvt_qp *qp)
368 {
369         struct qib_qp_priv *priv = qp->priv;
370         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
371
372         spin_lock(&dev->rdi.pending_lock);
373         if (!list_empty(&priv->iowait))
374                 list_del_init(&priv->iowait);
375         spin_unlock(&dev->rdi.pending_lock);
376 }
377
378 /**
379  * qib_check_send_wqe - validate wr/wqe
380  * @qp - The qp
381  * @wqe - The built wqe
382  *
383  * validate wr/wqe.  This is called
384  * prior to inserting the wqe into
385  * the ring but after the wqe has been
386  * setup.
387  *
388  * Returns 1 to force direct progress, 0 otherwise, -EINVAL on failure
389  */
390 int qib_check_send_wqe(struct rvt_qp *qp,
391                        struct rvt_swqe *wqe)
392 {
393         struct rvt_ah *ah;
394         int ret = 0;
395
396         switch (qp->ibqp.qp_type) {
397         case IB_QPT_RC:
398         case IB_QPT_UC:
399                 if (wqe->length > 0x80000000U)
400                         return -EINVAL;
401                 break;
402         case IB_QPT_SMI:
403         case IB_QPT_GSI:
404         case IB_QPT_UD:
405                 ah = ibah_to_rvtah(wqe->ud_wr.ah);
406                 if (wqe->length > (1 << ah->log_pmtu))
407                         return -EINVAL;
408                 /* progress hint */
409                 ret = 1;
410                 break;
411         default:
412                 break;
413         }
414         return ret;
415 }
416
417 #ifdef CONFIG_DEBUG_FS
418
419 struct qib_qp_iter {
420         struct qib_ibdev *dev;
421         struct rvt_qp *qp;
422         int n;
423 };
424
425 struct qib_qp_iter *qib_qp_iter_init(struct qib_ibdev *dev)
426 {
427         struct qib_qp_iter *iter;
428
429         iter = kzalloc(sizeof(*iter), GFP_KERNEL);
430         if (!iter)
431                 return NULL;
432
433         iter->dev = dev;
434
435         return iter;
436 }
437
438 int qib_qp_iter_next(struct qib_qp_iter *iter)
439 {
440         struct qib_ibdev *dev = iter->dev;
441         int n = iter->n;
442         int ret = 1;
443         struct rvt_qp *pqp = iter->qp;
444         struct rvt_qp *qp;
445
446         for (; n < dev->rdi.qp_dev->qp_table_size; n++) {
447                 if (pqp)
448                         qp = rcu_dereference(pqp->next);
449                 else
450                         qp = rcu_dereference(dev->rdi.qp_dev->qp_table[n]);
451                 pqp = qp;
452                 if (qp) {
453                         iter->qp = qp;
454                         iter->n = n;
455                         return 0;
456                 }
457         }
458         return ret;
459 }
460
461 static const char * const qp_type_str[] = {
462         "SMI", "GSI", "RC", "UC", "UD",
463 };
464
465 void qib_qp_iter_print(struct seq_file *s, struct qib_qp_iter *iter)
466 {
467         struct rvt_swqe *wqe;
468         struct rvt_qp *qp = iter->qp;
469         struct qib_qp_priv *priv = qp->priv;
470
471         wqe = rvt_get_swqe_ptr(qp, qp->s_last);
472         seq_printf(s,
473                    "N %d QP%u %s %u %u %u f=%x %u %u %u %u %u PSN %x %x %x %x %x (%u %u %u %u %u %u) QP%u LID %x\n",
474                    iter->n,
475                    qp->ibqp.qp_num,
476                    qp_type_str[qp->ibqp.qp_type],
477                    qp->state,
478                    wqe->wr.opcode,
479                    qp->s_hdrwords,
480                    qp->s_flags,
481                    atomic_read(&priv->s_dma_busy),
482                    !list_empty(&priv->iowait),
483                    qp->timeout,
484                    wqe->ssn,
485                    qp->s_lsn,
486                    qp->s_last_psn,
487                    qp->s_psn, qp->s_next_psn,
488                    qp->s_sending_psn, qp->s_sending_hpsn,
489                    qp->s_last, qp->s_acked, qp->s_cur,
490                    qp->s_tail, qp->s_head, qp->s_size,
491                    qp->remote_qpn,
492                    rdma_ah_get_dlid(&qp->remote_ah_attr));
493 }
494
495 #endif