]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - drivers/infiniband/hw/cxgb4/device.c
Merge branches 'cxgb4', 'ipoib' and 'mlx4' into for-next
[~shefty/rdma-dev.git] / drivers / infiniband / hw / cxgb4 / device.c
1 /*
2  * Copyright (c) 2009-2010 Chelsio, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 #include <linux/module.h>
33 #include <linux/moduleparam.h>
34 #include <linux/debugfs.h>
35 #include <linux/vmalloc.h>
36
37 #include <rdma/ib_verbs.h>
38
39 #include "iw_cxgb4.h"
40
41 #define DRV_VERSION "0.1"
42
43 MODULE_AUTHOR("Steve Wise");
44 MODULE_DESCRIPTION("Chelsio T4 RDMA Driver");
45 MODULE_LICENSE("Dual BSD/GPL");
46 MODULE_VERSION(DRV_VERSION);
47
48 struct uld_ctx {
49         struct list_head entry;
50         struct cxgb4_lld_info lldi;
51         struct c4iw_dev *dev;
52 };
53
54 static LIST_HEAD(uld_ctx_list);
55 static DEFINE_MUTEX(dev_mutex);
56
57 static struct dentry *c4iw_debugfs_root;
58
59 struct c4iw_debugfs_data {
60         struct c4iw_dev *devp;
61         char *buf;
62         int bufsize;
63         int pos;
64 };
65
66 static int count_idrs(int id, void *p, void *data)
67 {
68         int *countp = data;
69
70         *countp = *countp + 1;
71         return 0;
72 }
73
74 static ssize_t debugfs_read(struct file *file, char __user *buf, size_t count,
75                             loff_t *ppos)
76 {
77         struct c4iw_debugfs_data *d = file->private_data;
78
79         return simple_read_from_buffer(buf, count, ppos, d->buf, d->pos);
80 }
81
82 static int dump_qp(int id, void *p, void *data)
83 {
84         struct c4iw_qp *qp = p;
85         struct c4iw_debugfs_data *qpd = data;
86         int space;
87         int cc;
88
89         if (id != qp->wq.sq.qid)
90                 return 0;
91
92         space = qpd->bufsize - qpd->pos - 1;
93         if (space == 0)
94                 return 1;
95
96         if (qp->ep)
97                 cc = snprintf(qpd->buf + qpd->pos, space,
98                              "qp sq id %u rq id %u state %u onchip %u "
99                              "ep tid %u state %u %pI4:%u->%pI4:%u\n",
100                              qp->wq.sq.qid, qp->wq.rq.qid, (int)qp->attr.state,
101                              qp->wq.sq.flags & T4_SQ_ONCHIP,
102                              qp->ep->hwtid, (int)qp->ep->com.state,
103                              &qp->ep->com.local_addr.sin_addr.s_addr,
104                              ntohs(qp->ep->com.local_addr.sin_port),
105                              &qp->ep->com.remote_addr.sin_addr.s_addr,
106                              ntohs(qp->ep->com.remote_addr.sin_port));
107         else
108                 cc = snprintf(qpd->buf + qpd->pos, space,
109                              "qp sq id %u rq id %u state %u onchip %u\n",
110                               qp->wq.sq.qid, qp->wq.rq.qid,
111                               (int)qp->attr.state,
112                               qp->wq.sq.flags & T4_SQ_ONCHIP);
113         if (cc < space)
114                 qpd->pos += cc;
115         return 0;
116 }
117
118 static int qp_release(struct inode *inode, struct file *file)
119 {
120         struct c4iw_debugfs_data *qpd = file->private_data;
121         if (!qpd) {
122                 printk(KERN_INFO "%s null qpd?\n", __func__);
123                 return 0;
124         }
125         vfree(qpd->buf);
126         kfree(qpd);
127         return 0;
128 }
129
130 static int qp_open(struct inode *inode, struct file *file)
131 {
132         struct c4iw_debugfs_data *qpd;
133         int ret = 0;
134         int count = 1;
135
136         qpd = kmalloc(sizeof *qpd, GFP_KERNEL);
137         if (!qpd) {
138                 ret = -ENOMEM;
139                 goto out;
140         }
141         qpd->devp = inode->i_private;
142         qpd->pos = 0;
143
144         spin_lock_irq(&qpd->devp->lock);
145         idr_for_each(&qpd->devp->qpidr, count_idrs, &count);
146         spin_unlock_irq(&qpd->devp->lock);
147
148         qpd->bufsize = count * 128;
149         qpd->buf = vmalloc(qpd->bufsize);
150         if (!qpd->buf) {
151                 ret = -ENOMEM;
152                 goto err1;
153         }
154
155         spin_lock_irq(&qpd->devp->lock);
156         idr_for_each(&qpd->devp->qpidr, dump_qp, qpd);
157         spin_unlock_irq(&qpd->devp->lock);
158
159         qpd->buf[qpd->pos++] = 0;
160         file->private_data = qpd;
161         goto out;
162 err1:
163         kfree(qpd);
164 out:
165         return ret;
166 }
167
168 static const struct file_operations qp_debugfs_fops = {
169         .owner   = THIS_MODULE,
170         .open    = qp_open,
171         .release = qp_release,
172         .read    = debugfs_read,
173         .llseek  = default_llseek,
174 };
175
176 static int dump_stag(int id, void *p, void *data)
177 {
178         struct c4iw_debugfs_data *stagd = data;
179         int space;
180         int cc;
181
182         space = stagd->bufsize - stagd->pos - 1;
183         if (space == 0)
184                 return 1;
185
186         cc = snprintf(stagd->buf + stagd->pos, space, "0x%x\n", id<<8);
187         if (cc < space)
188                 stagd->pos += cc;
189         return 0;
190 }
191
192 static int stag_release(struct inode *inode, struct file *file)
193 {
194         struct c4iw_debugfs_data *stagd = file->private_data;
195         if (!stagd) {
196                 printk(KERN_INFO "%s null stagd?\n", __func__);
197                 return 0;
198         }
199         kfree(stagd->buf);
200         kfree(stagd);
201         return 0;
202 }
203
204 static int stag_open(struct inode *inode, struct file *file)
205 {
206         struct c4iw_debugfs_data *stagd;
207         int ret = 0;
208         int count = 1;
209
210         stagd = kmalloc(sizeof *stagd, GFP_KERNEL);
211         if (!stagd) {
212                 ret = -ENOMEM;
213                 goto out;
214         }
215         stagd->devp = inode->i_private;
216         stagd->pos = 0;
217
218         spin_lock_irq(&stagd->devp->lock);
219         idr_for_each(&stagd->devp->mmidr, count_idrs, &count);
220         spin_unlock_irq(&stagd->devp->lock);
221
222         stagd->bufsize = count * sizeof("0x12345678\n");
223         stagd->buf = kmalloc(stagd->bufsize, GFP_KERNEL);
224         if (!stagd->buf) {
225                 ret = -ENOMEM;
226                 goto err1;
227         }
228
229         spin_lock_irq(&stagd->devp->lock);
230         idr_for_each(&stagd->devp->mmidr, dump_stag, stagd);
231         spin_unlock_irq(&stagd->devp->lock);
232
233         stagd->buf[stagd->pos++] = 0;
234         file->private_data = stagd;
235         goto out;
236 err1:
237         kfree(stagd);
238 out:
239         return ret;
240 }
241
242 static const struct file_operations stag_debugfs_fops = {
243         .owner   = THIS_MODULE,
244         .open    = stag_open,
245         .release = stag_release,
246         .read    = debugfs_read,
247         .llseek  = default_llseek,
248 };
249
250 static char *db_state_str[] = {"NORMAL", "FLOW_CONTROL", "RECOVERY"};
251
252 static int stats_show(struct seq_file *seq, void *v)
253 {
254         struct c4iw_dev *dev = seq->private;
255
256         seq_printf(seq, "   Object: %10s %10s %10s %10s\n", "Total", "Current",
257                    "Max", "Fail");
258         seq_printf(seq, "     PDID: %10llu %10llu %10llu %10llu\n",
259                         dev->rdev.stats.pd.total, dev->rdev.stats.pd.cur,
260                         dev->rdev.stats.pd.max, dev->rdev.stats.pd.fail);
261         seq_printf(seq, "      QID: %10llu %10llu %10llu %10llu\n",
262                         dev->rdev.stats.qid.total, dev->rdev.stats.qid.cur,
263                         dev->rdev.stats.qid.max, dev->rdev.stats.qid.fail);
264         seq_printf(seq, "   TPTMEM: %10llu %10llu %10llu %10llu\n",
265                         dev->rdev.stats.stag.total, dev->rdev.stats.stag.cur,
266                         dev->rdev.stats.stag.max, dev->rdev.stats.stag.fail);
267         seq_printf(seq, "   PBLMEM: %10llu %10llu %10llu %10llu\n",
268                         dev->rdev.stats.pbl.total, dev->rdev.stats.pbl.cur,
269                         dev->rdev.stats.pbl.max, dev->rdev.stats.pbl.fail);
270         seq_printf(seq, "   RQTMEM: %10llu %10llu %10llu %10llu\n",
271                         dev->rdev.stats.rqt.total, dev->rdev.stats.rqt.cur,
272                         dev->rdev.stats.rqt.max, dev->rdev.stats.rqt.fail);
273         seq_printf(seq, "  OCQPMEM: %10llu %10llu %10llu %10llu\n",
274                         dev->rdev.stats.ocqp.total, dev->rdev.stats.ocqp.cur,
275                         dev->rdev.stats.ocqp.max, dev->rdev.stats.ocqp.fail);
276         seq_printf(seq, "  DB FULL: %10llu\n", dev->rdev.stats.db_full);
277         seq_printf(seq, " DB EMPTY: %10llu\n", dev->rdev.stats.db_empty);
278         seq_printf(seq, "  DB DROP: %10llu\n", dev->rdev.stats.db_drop);
279         seq_printf(seq, " DB State: %s Transitions %llu\n",
280                    db_state_str[dev->db_state],
281                    dev->rdev.stats.db_state_transitions);
282         seq_printf(seq, "TCAM_FULL: %10llu\n", dev->rdev.stats.tcam_full);
283         seq_printf(seq, "ACT_OFLD_CONN_FAILS: %10llu\n",
284                    dev->rdev.stats.act_ofld_conn_fails);
285         seq_printf(seq, "PAS_OFLD_CONN_FAILS: %10llu\n",
286                    dev->rdev.stats.pas_ofld_conn_fails);
287         return 0;
288 }
289
290 static int stats_open(struct inode *inode, struct file *file)
291 {
292         return single_open(file, stats_show, inode->i_private);
293 }
294
295 static ssize_t stats_clear(struct file *file, const char __user *buf,
296                 size_t count, loff_t *pos)
297 {
298         struct c4iw_dev *dev = ((struct seq_file *)file->private_data)->private;
299
300         mutex_lock(&dev->rdev.stats.lock);
301         dev->rdev.stats.pd.max = 0;
302         dev->rdev.stats.pd.fail = 0;
303         dev->rdev.stats.qid.max = 0;
304         dev->rdev.stats.qid.fail = 0;
305         dev->rdev.stats.stag.max = 0;
306         dev->rdev.stats.stag.fail = 0;
307         dev->rdev.stats.pbl.max = 0;
308         dev->rdev.stats.pbl.fail = 0;
309         dev->rdev.stats.rqt.max = 0;
310         dev->rdev.stats.rqt.fail = 0;
311         dev->rdev.stats.ocqp.max = 0;
312         dev->rdev.stats.ocqp.fail = 0;
313         dev->rdev.stats.db_full = 0;
314         dev->rdev.stats.db_empty = 0;
315         dev->rdev.stats.db_drop = 0;
316         dev->rdev.stats.db_state_transitions = 0;
317         dev->rdev.stats.tcam_full = 0;
318         dev->rdev.stats.act_ofld_conn_fails = 0;
319         dev->rdev.stats.pas_ofld_conn_fails = 0;
320         mutex_unlock(&dev->rdev.stats.lock);
321         return count;
322 }
323
324 static const struct file_operations stats_debugfs_fops = {
325         .owner   = THIS_MODULE,
326         .open    = stats_open,
327         .release = single_release,
328         .read    = seq_read,
329         .llseek  = seq_lseek,
330         .write   = stats_clear,
331 };
332
333 static int dump_ep(int id, void *p, void *data)
334 {
335         struct c4iw_ep *ep = p;
336         struct c4iw_debugfs_data *epd = data;
337         int space;
338         int cc;
339
340         space = epd->bufsize - epd->pos - 1;
341         if (space == 0)
342                 return 1;
343
344         cc = snprintf(epd->buf + epd->pos, space,
345                         "ep %p cm_id %p qp %p state %d flags 0x%lx history 0x%lx "
346                         "hwtid %d atid %d %pI4:%d <-> %pI4:%d\n",
347                         ep, ep->com.cm_id, ep->com.qp, (int)ep->com.state,
348                         ep->com.flags, ep->com.history, ep->hwtid, ep->atid,
349                         &ep->com.local_addr.sin_addr.s_addr,
350                         ntohs(ep->com.local_addr.sin_port),
351                         &ep->com.remote_addr.sin_addr.s_addr,
352                         ntohs(ep->com.remote_addr.sin_port));
353         if (cc < space)
354                 epd->pos += cc;
355         return 0;
356 }
357
358 static int dump_listen_ep(int id, void *p, void *data)
359 {
360         struct c4iw_listen_ep *ep = p;
361         struct c4iw_debugfs_data *epd = data;
362         int space;
363         int cc;
364
365         space = epd->bufsize - epd->pos - 1;
366         if (space == 0)
367                 return 1;
368
369         cc = snprintf(epd->buf + epd->pos, space,
370                         "ep %p cm_id %p state %d flags 0x%lx stid %d backlog %d "
371                         "%pI4:%d\n", ep, ep->com.cm_id, (int)ep->com.state,
372                         ep->com.flags, ep->stid, ep->backlog,
373                         &ep->com.local_addr.sin_addr.s_addr,
374                         ntohs(ep->com.local_addr.sin_port));
375         if (cc < space)
376                 epd->pos += cc;
377         return 0;
378 }
379
380 static int ep_release(struct inode *inode, struct file *file)
381 {
382         struct c4iw_debugfs_data *epd = file->private_data;
383         if (!epd) {
384                 pr_info("%s null qpd?\n", __func__);
385                 return 0;
386         }
387         vfree(epd->buf);
388         kfree(epd);
389         return 0;
390 }
391
392 static int ep_open(struct inode *inode, struct file *file)
393 {
394         struct c4iw_debugfs_data *epd;
395         int ret = 0;
396         int count = 1;
397
398         epd = kmalloc(sizeof(*epd), GFP_KERNEL);
399         if (!epd) {
400                 ret = -ENOMEM;
401                 goto out;
402         }
403         epd->devp = inode->i_private;
404         epd->pos = 0;
405
406         spin_lock_irq(&epd->devp->lock);
407         idr_for_each(&epd->devp->hwtid_idr, count_idrs, &count);
408         idr_for_each(&epd->devp->atid_idr, count_idrs, &count);
409         idr_for_each(&epd->devp->stid_idr, count_idrs, &count);
410         spin_unlock_irq(&epd->devp->lock);
411
412         epd->bufsize = count * 160;
413         epd->buf = vmalloc(epd->bufsize);
414         if (!epd->buf) {
415                 ret = -ENOMEM;
416                 goto err1;
417         }
418
419         spin_lock_irq(&epd->devp->lock);
420         idr_for_each(&epd->devp->hwtid_idr, dump_ep, epd);
421         idr_for_each(&epd->devp->atid_idr, dump_ep, epd);
422         idr_for_each(&epd->devp->stid_idr, dump_listen_ep, epd);
423         spin_unlock_irq(&epd->devp->lock);
424
425         file->private_data = epd;
426         goto out;
427 err1:
428         kfree(epd);
429 out:
430         return ret;
431 }
432
433 static const struct file_operations ep_debugfs_fops = {
434         .owner   = THIS_MODULE,
435         .open    = ep_open,
436         .release = ep_release,
437         .read    = debugfs_read,
438 };
439
440 static int setup_debugfs(struct c4iw_dev *devp)
441 {
442         struct dentry *de;
443
444         if (!devp->debugfs_root)
445                 return -1;
446
447         de = debugfs_create_file("qps", S_IWUSR, devp->debugfs_root,
448                                  (void *)devp, &qp_debugfs_fops);
449         if (de && de->d_inode)
450                 de->d_inode->i_size = 4096;
451
452         de = debugfs_create_file("stags", S_IWUSR, devp->debugfs_root,
453                                  (void *)devp, &stag_debugfs_fops);
454         if (de && de->d_inode)
455                 de->d_inode->i_size = 4096;
456
457         de = debugfs_create_file("stats", S_IWUSR, devp->debugfs_root,
458                         (void *)devp, &stats_debugfs_fops);
459         if (de && de->d_inode)
460                 de->d_inode->i_size = 4096;
461
462         de = debugfs_create_file("eps", S_IWUSR, devp->debugfs_root,
463                         (void *)devp, &ep_debugfs_fops);
464         if (de && de->d_inode)
465                 de->d_inode->i_size = 4096;
466
467         return 0;
468 }
469
470 void c4iw_release_dev_ucontext(struct c4iw_rdev *rdev,
471                                struct c4iw_dev_ucontext *uctx)
472 {
473         struct list_head *pos, *nxt;
474         struct c4iw_qid_list *entry;
475
476         mutex_lock(&uctx->lock);
477         list_for_each_safe(pos, nxt, &uctx->qpids) {
478                 entry = list_entry(pos, struct c4iw_qid_list, entry);
479                 list_del_init(&entry->entry);
480                 if (!(entry->qid & rdev->qpmask)) {
481                         c4iw_put_resource(&rdev->resource.qid_table,
482                                           entry->qid);
483                         mutex_lock(&rdev->stats.lock);
484                         rdev->stats.qid.cur -= rdev->qpmask + 1;
485                         mutex_unlock(&rdev->stats.lock);
486                 }
487                 kfree(entry);
488         }
489
490         list_for_each_safe(pos, nxt, &uctx->qpids) {
491                 entry = list_entry(pos, struct c4iw_qid_list, entry);
492                 list_del_init(&entry->entry);
493                 kfree(entry);
494         }
495         mutex_unlock(&uctx->lock);
496 }
497
498 void c4iw_init_dev_ucontext(struct c4iw_rdev *rdev,
499                             struct c4iw_dev_ucontext *uctx)
500 {
501         INIT_LIST_HEAD(&uctx->qpids);
502         INIT_LIST_HEAD(&uctx->cqids);
503         mutex_init(&uctx->lock);
504 }
505
506 /* Caller takes care of locking if needed */
507 static int c4iw_rdev_open(struct c4iw_rdev *rdev)
508 {
509         int err;
510
511         c4iw_init_dev_ucontext(rdev, &rdev->uctx);
512
513         /*
514          * qpshift is the number of bits to shift the qpid left in order
515          * to get the correct address of the doorbell for that qp.
516          */
517         rdev->qpshift = PAGE_SHIFT - ilog2(rdev->lldi.udb_density);
518         rdev->qpmask = rdev->lldi.udb_density - 1;
519         rdev->cqshift = PAGE_SHIFT - ilog2(rdev->lldi.ucq_density);
520         rdev->cqmask = rdev->lldi.ucq_density - 1;
521         PDBG("%s dev %s stag start 0x%0x size 0x%0x num stags %d "
522              "pbl start 0x%0x size 0x%0x rq start 0x%0x size 0x%0x "
523              "qp qid start %u size %u cq qid start %u size %u\n",
524              __func__, pci_name(rdev->lldi.pdev), rdev->lldi.vr->stag.start,
525              rdev->lldi.vr->stag.size, c4iw_num_stags(rdev),
526              rdev->lldi.vr->pbl.start,
527              rdev->lldi.vr->pbl.size, rdev->lldi.vr->rq.start,
528              rdev->lldi.vr->rq.size,
529              rdev->lldi.vr->qp.start,
530              rdev->lldi.vr->qp.size,
531              rdev->lldi.vr->cq.start,
532              rdev->lldi.vr->cq.size);
533         PDBG("udb len 0x%x udb base %p db_reg %p gts_reg %p qpshift %lu "
534              "qpmask 0x%x cqshift %lu cqmask 0x%x\n",
535              (unsigned)pci_resource_len(rdev->lldi.pdev, 2),
536              (void *)pci_resource_start(rdev->lldi.pdev, 2),
537              rdev->lldi.db_reg,
538              rdev->lldi.gts_reg,
539              rdev->qpshift, rdev->qpmask,
540              rdev->cqshift, rdev->cqmask);
541
542         if (c4iw_num_stags(rdev) == 0) {
543                 err = -EINVAL;
544                 goto err1;
545         }
546
547         rdev->stats.pd.total = T4_MAX_NUM_PD;
548         rdev->stats.stag.total = rdev->lldi.vr->stag.size;
549         rdev->stats.pbl.total = rdev->lldi.vr->pbl.size;
550         rdev->stats.rqt.total = rdev->lldi.vr->rq.size;
551         rdev->stats.ocqp.total = rdev->lldi.vr->ocq.size;
552         rdev->stats.qid.total = rdev->lldi.vr->qp.size;
553
554         err = c4iw_init_resource(rdev, c4iw_num_stags(rdev), T4_MAX_NUM_PD);
555         if (err) {
556                 printk(KERN_ERR MOD "error %d initializing resources\n", err);
557                 goto err1;
558         }
559         err = c4iw_pblpool_create(rdev);
560         if (err) {
561                 printk(KERN_ERR MOD "error %d initializing pbl pool\n", err);
562                 goto err2;
563         }
564         err = c4iw_rqtpool_create(rdev);
565         if (err) {
566                 printk(KERN_ERR MOD "error %d initializing rqt pool\n", err);
567                 goto err3;
568         }
569         err = c4iw_ocqp_pool_create(rdev);
570         if (err) {
571                 printk(KERN_ERR MOD "error %d initializing ocqp pool\n", err);
572                 goto err4;
573         }
574         return 0;
575 err4:
576         c4iw_rqtpool_destroy(rdev);
577 err3:
578         c4iw_pblpool_destroy(rdev);
579 err2:
580         c4iw_destroy_resource(&rdev->resource);
581 err1:
582         return err;
583 }
584
585 static void c4iw_rdev_close(struct c4iw_rdev *rdev)
586 {
587         c4iw_pblpool_destroy(rdev);
588         c4iw_rqtpool_destroy(rdev);
589         c4iw_destroy_resource(&rdev->resource);
590 }
591
592 static void c4iw_dealloc(struct uld_ctx *ctx)
593 {
594         c4iw_rdev_close(&ctx->dev->rdev);
595         idr_destroy(&ctx->dev->cqidr);
596         idr_destroy(&ctx->dev->qpidr);
597         idr_destroy(&ctx->dev->mmidr);
598         idr_destroy(&ctx->dev->hwtid_idr);
599         idr_destroy(&ctx->dev->stid_idr);
600         idr_destroy(&ctx->dev->atid_idr);
601         iounmap(ctx->dev->rdev.oc_mw_kva);
602         ib_dealloc_device(&ctx->dev->ibdev);
603         ctx->dev = NULL;
604 }
605
606 static void c4iw_remove(struct uld_ctx *ctx)
607 {
608         PDBG("%s c4iw_dev %p\n", __func__,  ctx->dev);
609         c4iw_unregister_device(ctx->dev);
610         c4iw_dealloc(ctx);
611 }
612
613 static int rdma_supported(const struct cxgb4_lld_info *infop)
614 {
615         return infop->vr->stag.size > 0 && infop->vr->pbl.size > 0 &&
616                infop->vr->rq.size > 0 && infop->vr->qp.size > 0 &&
617                infop->vr->cq.size > 0 && infop->vr->ocq.size > 0;
618 }
619
620 static struct c4iw_dev *c4iw_alloc(const struct cxgb4_lld_info *infop)
621 {
622         struct c4iw_dev *devp;
623         int ret;
624
625         if (!rdma_supported(infop)) {
626                 printk(KERN_INFO MOD "%s: RDMA not supported on this device.\n",
627                        pci_name(infop->pdev));
628                 return ERR_PTR(-ENOSYS);
629         }
630         devp = (struct c4iw_dev *)ib_alloc_device(sizeof(*devp));
631         if (!devp) {
632                 printk(KERN_ERR MOD "Cannot allocate ib device\n");
633                 return ERR_PTR(-ENOMEM);
634         }
635         devp->rdev.lldi = *infop;
636
637         devp->rdev.oc_mw_pa = pci_resource_start(devp->rdev.lldi.pdev, 2) +
638                 (pci_resource_len(devp->rdev.lldi.pdev, 2) -
639                  roundup_pow_of_two(devp->rdev.lldi.vr->ocq.size));
640         devp->rdev.oc_mw_kva = ioremap_wc(devp->rdev.oc_mw_pa,
641                                                devp->rdev.lldi.vr->ocq.size);
642
643         PDBG(KERN_INFO MOD "ocq memory: "
644                "hw_start 0x%x size %u mw_pa 0x%lx mw_kva %p\n",
645                devp->rdev.lldi.vr->ocq.start, devp->rdev.lldi.vr->ocq.size,
646                devp->rdev.oc_mw_pa, devp->rdev.oc_mw_kva);
647
648         ret = c4iw_rdev_open(&devp->rdev);
649         if (ret) {
650                 printk(KERN_ERR MOD "Unable to open CXIO rdev err %d\n", ret);
651                 ib_dealloc_device(&devp->ibdev);
652                 return ERR_PTR(ret);
653         }
654
655         idr_init(&devp->cqidr);
656         idr_init(&devp->qpidr);
657         idr_init(&devp->mmidr);
658         idr_init(&devp->hwtid_idr);
659         idr_init(&devp->stid_idr);
660         idr_init(&devp->atid_idr);
661         spin_lock_init(&devp->lock);
662         mutex_init(&devp->rdev.stats.lock);
663         mutex_init(&devp->db_mutex);
664
665         if (c4iw_debugfs_root) {
666                 devp->debugfs_root = debugfs_create_dir(
667                                         pci_name(devp->rdev.lldi.pdev),
668                                         c4iw_debugfs_root);
669                 setup_debugfs(devp);
670         }
671         return devp;
672 }
673
674 static void *c4iw_uld_add(const struct cxgb4_lld_info *infop)
675 {
676         struct uld_ctx *ctx;
677         static int vers_printed;
678         int i;
679
680         if (!vers_printed++)
681                 printk(KERN_INFO MOD "Chelsio T4 RDMA Driver - version %s\n",
682                        DRV_VERSION);
683
684         ctx = kzalloc(sizeof *ctx, GFP_KERNEL);
685         if (!ctx) {
686                 ctx = ERR_PTR(-ENOMEM);
687                 goto out;
688         }
689         ctx->lldi = *infop;
690
691         PDBG("%s found device %s nchan %u nrxq %u ntxq %u nports %u\n",
692              __func__, pci_name(ctx->lldi.pdev),
693              ctx->lldi.nchan, ctx->lldi.nrxq,
694              ctx->lldi.ntxq, ctx->lldi.nports);
695
696         mutex_lock(&dev_mutex);
697         list_add_tail(&ctx->entry, &uld_ctx_list);
698         mutex_unlock(&dev_mutex);
699
700         for (i = 0; i < ctx->lldi.nrxq; i++)
701                 PDBG("rxqid[%u] %u\n", i, ctx->lldi.rxq_ids[i]);
702 out:
703         return ctx;
704 }
705
706 static inline struct sk_buff *copy_gl_to_skb_pkt(const struct pkt_gl *gl,
707                                                  const __be64 *rsp,
708                                                  u32 pktshift)
709 {
710         struct sk_buff *skb;
711
712         /*
713          * Allocate space for cpl_pass_accept_req which will be synthesized by
714          * driver. Once the driver synthesizes the request the skb will go
715          * through the regular cpl_pass_accept_req processing.
716          * The math here assumes sizeof cpl_pass_accept_req >= sizeof
717          * cpl_rx_pkt.
718          */
719         skb = alloc_skb(gl->tot_len + sizeof(struct cpl_pass_accept_req) +
720                         sizeof(struct rss_header) - pktshift, GFP_ATOMIC);
721         if (unlikely(!skb))
722                 return NULL;
723
724          __skb_put(skb, gl->tot_len + sizeof(struct cpl_pass_accept_req) +
725                    sizeof(struct rss_header) - pktshift);
726
727         /*
728          * This skb will contain:
729          *   rss_header from the rspq descriptor (1 flit)
730          *   cpl_rx_pkt struct from the rspq descriptor (2 flits)
731          *   space for the difference between the size of an
732          *      rx_pkt and pass_accept_req cpl (1 flit)
733          *   the packet data from the gl
734          */
735         skb_copy_to_linear_data(skb, rsp, sizeof(struct cpl_pass_accept_req) +
736                                 sizeof(struct rss_header));
737         skb_copy_to_linear_data_offset(skb, sizeof(struct rss_header) +
738                                        sizeof(struct cpl_pass_accept_req),
739                                        gl->va + pktshift,
740                                        gl->tot_len - pktshift);
741         return skb;
742 }
743
744 static inline int recv_rx_pkt(struct c4iw_dev *dev, const struct pkt_gl *gl,
745                            const __be64 *rsp)
746 {
747         unsigned int opcode = *(u8 *)rsp;
748         struct sk_buff *skb;
749
750         if (opcode != CPL_RX_PKT)
751                 goto out;
752
753         skb = copy_gl_to_skb_pkt(gl , rsp, dev->rdev.lldi.sge_pktshift);
754         if (skb == NULL)
755                 goto out;
756
757         if (c4iw_handlers[opcode] == NULL) {
758                 pr_info("%s no handler opcode 0x%x...\n", __func__,
759                        opcode);
760                 kfree_skb(skb);
761                 goto out;
762         }
763         c4iw_handlers[opcode](dev, skb);
764         return 1;
765 out:
766         return 0;
767 }
768
769 static int c4iw_uld_rx_handler(void *handle, const __be64 *rsp,
770                         const struct pkt_gl *gl)
771 {
772         struct uld_ctx *ctx = handle;
773         struct c4iw_dev *dev = ctx->dev;
774         struct sk_buff *skb;
775         u8 opcode;
776
777         if (gl == NULL) {
778                 /* omit RSS and rsp_ctrl at end of descriptor */
779                 unsigned int len = 64 - sizeof(struct rsp_ctrl) - 8;
780
781                 skb = alloc_skb(256, GFP_ATOMIC);
782                 if (!skb)
783                         goto nomem;
784                 __skb_put(skb, len);
785                 skb_copy_to_linear_data(skb, &rsp[1], len);
786         } else if (gl == CXGB4_MSG_AN) {
787                 const struct rsp_ctrl *rc = (void *)rsp;
788
789                 u32 qid = be32_to_cpu(rc->pldbuflen_qid);
790                 c4iw_ev_handler(dev, qid);
791                 return 0;
792         } else if (unlikely(*(u8 *)rsp != *(u8 *)gl->va)) {
793                 if (recv_rx_pkt(dev, gl, rsp))
794                         return 0;
795
796                 pr_info("%s: unexpected FL contents at %p, " \
797                        "RSS %#llx, FL %#llx, len %u\n",
798                        pci_name(ctx->lldi.pdev), gl->va,
799                        (unsigned long long)be64_to_cpu(*rsp),
800                        (unsigned long long)be64_to_cpu(*(u64 *)gl->va),
801                        gl->tot_len);
802
803                 return 0;
804         } else {
805                 skb = cxgb4_pktgl_to_skb(gl, 128, 128);
806                 if (unlikely(!skb))
807                         goto nomem;
808         }
809
810         opcode = *(u8 *)rsp;
811         if (c4iw_handlers[opcode])
812                 c4iw_handlers[opcode](dev, skb);
813         else
814                 pr_info("%s no handler opcode 0x%x...\n", __func__,
815                        opcode);
816
817         return 0;
818 nomem:
819         return -1;
820 }
821
822 static int c4iw_uld_state_change(void *handle, enum cxgb4_state new_state)
823 {
824         struct uld_ctx *ctx = handle;
825
826         PDBG("%s new_state %u\n", __func__, new_state);
827         switch (new_state) {
828         case CXGB4_STATE_UP:
829                 printk(KERN_INFO MOD "%s: Up\n", pci_name(ctx->lldi.pdev));
830                 if (!ctx->dev) {
831                         int ret;
832
833                         ctx->dev = c4iw_alloc(&ctx->lldi);
834                         if (IS_ERR(ctx->dev)) {
835                                 printk(KERN_ERR MOD
836                                        "%s: initialization failed: %ld\n",
837                                        pci_name(ctx->lldi.pdev),
838                                        PTR_ERR(ctx->dev));
839                                 ctx->dev = NULL;
840                                 break;
841                         }
842                         ret = c4iw_register_device(ctx->dev);
843                         if (ret) {
844                                 printk(KERN_ERR MOD
845                                        "%s: RDMA registration failed: %d\n",
846                                        pci_name(ctx->lldi.pdev), ret);
847                                 c4iw_dealloc(ctx);
848                         }
849                 }
850                 break;
851         case CXGB4_STATE_DOWN:
852                 printk(KERN_INFO MOD "%s: Down\n",
853                        pci_name(ctx->lldi.pdev));
854                 if (ctx->dev)
855                         c4iw_remove(ctx);
856                 break;
857         case CXGB4_STATE_START_RECOVERY:
858                 printk(KERN_INFO MOD "%s: Fatal Error\n",
859                        pci_name(ctx->lldi.pdev));
860                 if (ctx->dev) {
861                         struct ib_event event;
862
863                         ctx->dev->rdev.flags |= T4_FATAL_ERROR;
864                         memset(&event, 0, sizeof event);
865                         event.event  = IB_EVENT_DEVICE_FATAL;
866                         event.device = &ctx->dev->ibdev;
867                         ib_dispatch_event(&event);
868                         c4iw_remove(ctx);
869                 }
870                 break;
871         case CXGB4_STATE_DETACH:
872                 printk(KERN_INFO MOD "%s: Detach\n",
873                        pci_name(ctx->lldi.pdev));
874                 if (ctx->dev)
875                         c4iw_remove(ctx);
876                 break;
877         }
878         return 0;
879 }
880
881 static int disable_qp_db(int id, void *p, void *data)
882 {
883         struct c4iw_qp *qp = p;
884
885         t4_disable_wq_db(&qp->wq);
886         return 0;
887 }
888
889 static void stop_queues(struct uld_ctx *ctx)
890 {
891         spin_lock_irq(&ctx->dev->lock);
892         if (ctx->dev->db_state == NORMAL) {
893                 ctx->dev->rdev.stats.db_state_transitions++;
894                 ctx->dev->db_state = FLOW_CONTROL;
895                 idr_for_each(&ctx->dev->qpidr, disable_qp_db, NULL);
896         }
897         spin_unlock_irq(&ctx->dev->lock);
898 }
899
900 static int enable_qp_db(int id, void *p, void *data)
901 {
902         struct c4iw_qp *qp = p;
903
904         t4_enable_wq_db(&qp->wq);
905         return 0;
906 }
907
908 static void resume_queues(struct uld_ctx *ctx)
909 {
910         spin_lock_irq(&ctx->dev->lock);
911         if (ctx->dev->qpcnt <= db_fc_threshold &&
912             ctx->dev->db_state == FLOW_CONTROL) {
913                 ctx->dev->db_state = NORMAL;
914                 ctx->dev->rdev.stats.db_state_transitions++;
915                 idr_for_each(&ctx->dev->qpidr, enable_qp_db, NULL);
916         }
917         spin_unlock_irq(&ctx->dev->lock);
918 }
919
920 struct qp_list {
921         unsigned idx;
922         struct c4iw_qp **qps;
923 };
924
925 static int add_and_ref_qp(int id, void *p, void *data)
926 {
927         struct qp_list *qp_listp = data;
928         struct c4iw_qp *qp = p;
929
930         c4iw_qp_add_ref(&qp->ibqp);
931         qp_listp->qps[qp_listp->idx++] = qp;
932         return 0;
933 }
934
935 static int count_qps(int id, void *p, void *data)
936 {
937         unsigned *countp = data;
938         (*countp)++;
939         return 0;
940 }
941
942 static void deref_qps(struct qp_list qp_list)
943 {
944         int idx;
945
946         for (idx = 0; idx < qp_list.idx; idx++)
947                 c4iw_qp_rem_ref(&qp_list.qps[idx]->ibqp);
948 }
949
950 static void recover_lost_dbs(struct uld_ctx *ctx, struct qp_list *qp_list)
951 {
952         int idx;
953         int ret;
954
955         for (idx = 0; idx < qp_list->idx; idx++) {
956                 struct c4iw_qp *qp = qp_list->qps[idx];
957
958                 ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
959                                           qp->wq.sq.qid,
960                                           t4_sq_host_wq_pidx(&qp->wq),
961                                           t4_sq_wq_size(&qp->wq));
962                 if (ret) {
963                         printk(KERN_ERR MOD "%s: Fatal error - "
964                                "DB overflow recovery failed - "
965                                "error syncing SQ qid %u\n",
966                                pci_name(ctx->lldi.pdev), qp->wq.sq.qid);
967                         return;
968                 }
969
970                 ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
971                                           qp->wq.rq.qid,
972                                           t4_rq_host_wq_pidx(&qp->wq),
973                                           t4_rq_wq_size(&qp->wq));
974
975                 if (ret) {
976                         printk(KERN_ERR MOD "%s: Fatal error - "
977                                "DB overflow recovery failed - "
978                                "error syncing RQ qid %u\n",
979                                pci_name(ctx->lldi.pdev), qp->wq.rq.qid);
980                         return;
981                 }
982
983                 /* Wait for the dbfifo to drain */
984                 while (cxgb4_dbfifo_count(qp->rhp->rdev.lldi.ports[0], 1) > 0) {
985                         set_current_state(TASK_UNINTERRUPTIBLE);
986                         schedule_timeout(usecs_to_jiffies(10));
987                 }
988         }
989 }
990
991 static void recover_queues(struct uld_ctx *ctx)
992 {
993         int count = 0;
994         struct qp_list qp_list;
995         int ret;
996
997         /* lock out kernel db ringers */
998         mutex_lock(&ctx->dev->db_mutex);
999
1000         /* put all queues in to recovery mode */
1001         spin_lock_irq(&ctx->dev->lock);
1002         ctx->dev->db_state = RECOVERY;
1003         ctx->dev->rdev.stats.db_state_transitions++;
1004         idr_for_each(&ctx->dev->qpidr, disable_qp_db, NULL);
1005         spin_unlock_irq(&ctx->dev->lock);
1006
1007         /* slow everybody down */
1008         set_current_state(TASK_UNINTERRUPTIBLE);
1009         schedule_timeout(usecs_to_jiffies(1000));
1010
1011         /* Wait for the dbfifo to completely drain. */
1012         while (cxgb4_dbfifo_count(ctx->dev->rdev.lldi.ports[0], 1) > 0) {
1013                 set_current_state(TASK_UNINTERRUPTIBLE);
1014                 schedule_timeout(usecs_to_jiffies(10));
1015         }
1016
1017         /* flush the SGE contexts */
1018         ret = cxgb4_flush_eq_cache(ctx->dev->rdev.lldi.ports[0]);
1019         if (ret) {
1020                 printk(KERN_ERR MOD "%s: Fatal error - DB overflow recovery failed\n",
1021                        pci_name(ctx->lldi.pdev));
1022                 goto out;
1023         }
1024
1025         /* Count active queues so we can build a list of queues to recover */
1026         spin_lock_irq(&ctx->dev->lock);
1027         idr_for_each(&ctx->dev->qpidr, count_qps, &count);
1028
1029         qp_list.qps = kzalloc(count * sizeof *qp_list.qps, GFP_ATOMIC);
1030         if (!qp_list.qps) {
1031                 printk(KERN_ERR MOD "%s: Fatal error - DB overflow recovery failed\n",
1032                        pci_name(ctx->lldi.pdev));
1033                 spin_unlock_irq(&ctx->dev->lock);
1034                 goto out;
1035         }
1036         qp_list.idx = 0;
1037
1038         /* add and ref each qp so it doesn't get freed */
1039         idr_for_each(&ctx->dev->qpidr, add_and_ref_qp, &qp_list);
1040
1041         spin_unlock_irq(&ctx->dev->lock);
1042
1043         /* now traverse the list in a safe context to recover the db state*/
1044         recover_lost_dbs(ctx, &qp_list);
1045
1046         /* we're almost done!  deref the qps and clean up */
1047         deref_qps(qp_list);
1048         kfree(qp_list.qps);
1049
1050         /* Wait for the dbfifo to completely drain again */
1051         while (cxgb4_dbfifo_count(ctx->dev->rdev.lldi.ports[0], 1) > 0) {
1052                 set_current_state(TASK_UNINTERRUPTIBLE);
1053                 schedule_timeout(usecs_to_jiffies(10));
1054         }
1055
1056         /* resume the queues */
1057         spin_lock_irq(&ctx->dev->lock);
1058         if (ctx->dev->qpcnt > db_fc_threshold)
1059                 ctx->dev->db_state = FLOW_CONTROL;
1060         else {
1061                 ctx->dev->db_state = NORMAL;
1062                 idr_for_each(&ctx->dev->qpidr, enable_qp_db, NULL);
1063         }
1064         ctx->dev->rdev.stats.db_state_transitions++;
1065         spin_unlock_irq(&ctx->dev->lock);
1066
1067 out:
1068         /* start up kernel db ringers again */
1069         mutex_unlock(&ctx->dev->db_mutex);
1070 }
1071
1072 static int c4iw_uld_control(void *handle, enum cxgb4_control control, ...)
1073 {
1074         struct uld_ctx *ctx = handle;
1075
1076         switch (control) {
1077         case CXGB4_CONTROL_DB_FULL:
1078                 stop_queues(ctx);
1079                 mutex_lock(&ctx->dev->rdev.stats.lock);
1080                 ctx->dev->rdev.stats.db_full++;
1081                 mutex_unlock(&ctx->dev->rdev.stats.lock);
1082                 break;
1083         case CXGB4_CONTROL_DB_EMPTY:
1084                 resume_queues(ctx);
1085                 mutex_lock(&ctx->dev->rdev.stats.lock);
1086                 ctx->dev->rdev.stats.db_empty++;
1087                 mutex_unlock(&ctx->dev->rdev.stats.lock);
1088                 break;
1089         case CXGB4_CONTROL_DB_DROP:
1090                 recover_queues(ctx);
1091                 mutex_lock(&ctx->dev->rdev.stats.lock);
1092                 ctx->dev->rdev.stats.db_drop++;
1093                 mutex_unlock(&ctx->dev->rdev.stats.lock);
1094                 break;
1095         default:
1096                 printk(KERN_WARNING MOD "%s: unknown control cmd %u\n",
1097                        pci_name(ctx->lldi.pdev), control);
1098                 break;
1099         }
1100         return 0;
1101 }
1102
1103 static struct cxgb4_uld_info c4iw_uld_info = {
1104         .name = DRV_NAME,
1105         .add = c4iw_uld_add,
1106         .rx_handler = c4iw_uld_rx_handler,
1107         .state_change = c4iw_uld_state_change,
1108         .control = c4iw_uld_control,
1109 };
1110
1111 static int __init c4iw_init_module(void)
1112 {
1113         int err;
1114
1115         err = c4iw_cm_init();
1116         if (err)
1117                 return err;
1118
1119         c4iw_debugfs_root = debugfs_create_dir(DRV_NAME, NULL);
1120         if (!c4iw_debugfs_root)
1121                 printk(KERN_WARNING MOD
1122                        "could not create debugfs entry, continuing\n");
1123
1124         cxgb4_register_uld(CXGB4_ULD_RDMA, &c4iw_uld_info);
1125
1126         return 0;
1127 }
1128
1129 static void __exit c4iw_exit_module(void)
1130 {
1131         struct uld_ctx *ctx, *tmp;
1132
1133         mutex_lock(&dev_mutex);
1134         list_for_each_entry_safe(ctx, tmp, &uld_ctx_list, entry) {
1135                 if (ctx->dev)
1136                         c4iw_remove(ctx);
1137                 kfree(ctx);
1138         }
1139         mutex_unlock(&dev_mutex);
1140         cxgb4_unregister_uld(CXGB4_ULD_RDMA);
1141         c4iw_cm_term();
1142         debugfs_remove_recursive(c4iw_debugfs_root);
1143 }
1144
1145 module_init(c4iw_init_module);
1146 module_exit(c4iw_exit_module);