Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
[firefly-linux-kernel-4.4.55.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/T5 RDMA Driver");
45 MODULE_LICENSE("Dual BSD/GPL");
46 MODULE_VERSION(DRV_VERSION);
47
48 static int allow_db_fc_on_t5;
49 module_param(allow_db_fc_on_t5, int, 0644);
50 MODULE_PARM_DESC(allow_db_fc_on_t5,
51                  "Allow DB Flow Control on T5 (default = 0)");
52
53 static int allow_db_coalescing_on_t5;
54 module_param(allow_db_coalescing_on_t5, int, 0644);
55 MODULE_PARM_DESC(allow_db_coalescing_on_t5,
56                  "Allow DB Coalescing on T5 (default = 0)");
57
58 struct uld_ctx {
59         struct list_head entry;
60         struct cxgb4_lld_info lldi;
61         struct c4iw_dev *dev;
62 };
63
64 static LIST_HEAD(uld_ctx_list);
65 static DEFINE_MUTEX(dev_mutex);
66
67 static struct dentry *c4iw_debugfs_root;
68
69 struct c4iw_debugfs_data {
70         struct c4iw_dev *devp;
71         char *buf;
72         int bufsize;
73         int pos;
74 };
75
76 static int count_idrs(int id, void *p, void *data)
77 {
78         int *countp = data;
79
80         *countp = *countp + 1;
81         return 0;
82 }
83
84 static ssize_t debugfs_read(struct file *file, char __user *buf, size_t count,
85                             loff_t *ppos)
86 {
87         struct c4iw_debugfs_data *d = file->private_data;
88
89         return simple_read_from_buffer(buf, count, ppos, d->buf, d->pos);
90 }
91
92 static int dump_qp(int id, void *p, void *data)
93 {
94         struct c4iw_qp *qp = p;
95         struct c4iw_debugfs_data *qpd = data;
96         int space;
97         int cc;
98
99         if (id != qp->wq.sq.qid)
100                 return 0;
101
102         space = qpd->bufsize - qpd->pos - 1;
103         if (space == 0)
104                 return 1;
105
106         if (qp->ep)
107                 cc = snprintf(qpd->buf + qpd->pos, space,
108                              "qp sq id %u rq id %u state %u onchip %u "
109                              "ep tid %u state %u %pI4:%u->%pI4:%u\n",
110                              qp->wq.sq.qid, qp->wq.rq.qid, (int)qp->attr.state,
111                              qp->wq.sq.flags & T4_SQ_ONCHIP,
112                              qp->ep->hwtid, (int)qp->ep->com.state,
113                              &qp->ep->com.local_addr.sin_addr.s_addr,
114                              ntohs(qp->ep->com.local_addr.sin_port),
115                              &qp->ep->com.remote_addr.sin_addr.s_addr,
116                              ntohs(qp->ep->com.remote_addr.sin_port));
117         else
118                 cc = snprintf(qpd->buf + qpd->pos, space,
119                              "qp sq id %u rq id %u state %u onchip %u\n",
120                               qp->wq.sq.qid, qp->wq.rq.qid,
121                               (int)qp->attr.state,
122                               qp->wq.sq.flags & T4_SQ_ONCHIP);
123         if (cc < space)
124                 qpd->pos += cc;
125         return 0;
126 }
127
128 static int qp_release(struct inode *inode, struct file *file)
129 {
130         struct c4iw_debugfs_data *qpd = file->private_data;
131         if (!qpd) {
132                 printk(KERN_INFO "%s null qpd?\n", __func__);
133                 return 0;
134         }
135         vfree(qpd->buf);
136         kfree(qpd);
137         return 0;
138 }
139
140 static int qp_open(struct inode *inode, struct file *file)
141 {
142         struct c4iw_debugfs_data *qpd;
143         int ret = 0;
144         int count = 1;
145
146         qpd = kmalloc(sizeof *qpd, GFP_KERNEL);
147         if (!qpd) {
148                 ret = -ENOMEM;
149                 goto out;
150         }
151         qpd->devp = inode->i_private;
152         qpd->pos = 0;
153
154         spin_lock_irq(&qpd->devp->lock);
155         idr_for_each(&qpd->devp->qpidr, count_idrs, &count);
156         spin_unlock_irq(&qpd->devp->lock);
157
158         qpd->bufsize = count * 128;
159         qpd->buf = vmalloc(qpd->bufsize);
160         if (!qpd->buf) {
161                 ret = -ENOMEM;
162                 goto err1;
163         }
164
165         spin_lock_irq(&qpd->devp->lock);
166         idr_for_each(&qpd->devp->qpidr, dump_qp, qpd);
167         spin_unlock_irq(&qpd->devp->lock);
168
169         qpd->buf[qpd->pos++] = 0;
170         file->private_data = qpd;
171         goto out;
172 err1:
173         kfree(qpd);
174 out:
175         return ret;
176 }
177
178 static const struct file_operations qp_debugfs_fops = {
179         .owner   = THIS_MODULE,
180         .open    = qp_open,
181         .release = qp_release,
182         .read    = debugfs_read,
183         .llseek  = default_llseek,
184 };
185
186 static int dump_stag(int id, void *p, void *data)
187 {
188         struct c4iw_debugfs_data *stagd = data;
189         int space;
190         int cc;
191
192         space = stagd->bufsize - stagd->pos - 1;
193         if (space == 0)
194                 return 1;
195
196         cc = snprintf(stagd->buf + stagd->pos, space, "0x%x\n", id<<8);
197         if (cc < space)
198                 stagd->pos += cc;
199         return 0;
200 }
201
202 static int stag_release(struct inode *inode, struct file *file)
203 {
204         struct c4iw_debugfs_data *stagd = file->private_data;
205         if (!stagd) {
206                 printk(KERN_INFO "%s null stagd?\n", __func__);
207                 return 0;
208         }
209         kfree(stagd->buf);
210         kfree(stagd);
211         return 0;
212 }
213
214 static int stag_open(struct inode *inode, struct file *file)
215 {
216         struct c4iw_debugfs_data *stagd;
217         int ret = 0;
218         int count = 1;
219
220         stagd = kmalloc(sizeof *stagd, GFP_KERNEL);
221         if (!stagd) {
222                 ret = -ENOMEM;
223                 goto out;
224         }
225         stagd->devp = inode->i_private;
226         stagd->pos = 0;
227
228         spin_lock_irq(&stagd->devp->lock);
229         idr_for_each(&stagd->devp->mmidr, count_idrs, &count);
230         spin_unlock_irq(&stagd->devp->lock);
231
232         stagd->bufsize = count * sizeof("0x12345678\n");
233         stagd->buf = kmalloc(stagd->bufsize, GFP_KERNEL);
234         if (!stagd->buf) {
235                 ret = -ENOMEM;
236                 goto err1;
237         }
238
239         spin_lock_irq(&stagd->devp->lock);
240         idr_for_each(&stagd->devp->mmidr, dump_stag, stagd);
241         spin_unlock_irq(&stagd->devp->lock);
242
243         stagd->buf[stagd->pos++] = 0;
244         file->private_data = stagd;
245         goto out;
246 err1:
247         kfree(stagd);
248 out:
249         return ret;
250 }
251
252 static const struct file_operations stag_debugfs_fops = {
253         .owner   = THIS_MODULE,
254         .open    = stag_open,
255         .release = stag_release,
256         .read    = debugfs_read,
257         .llseek  = default_llseek,
258 };
259
260 static char *db_state_str[] = {"NORMAL", "FLOW_CONTROL", "RECOVERY"};
261
262 static int stats_show(struct seq_file *seq, void *v)
263 {
264         struct c4iw_dev *dev = seq->private;
265
266         seq_printf(seq, "   Object: %10s %10s %10s %10s\n", "Total", "Current",
267                    "Max", "Fail");
268         seq_printf(seq, "     PDID: %10llu %10llu %10llu %10llu\n",
269                         dev->rdev.stats.pd.total, dev->rdev.stats.pd.cur,
270                         dev->rdev.stats.pd.max, dev->rdev.stats.pd.fail);
271         seq_printf(seq, "      QID: %10llu %10llu %10llu %10llu\n",
272                         dev->rdev.stats.qid.total, dev->rdev.stats.qid.cur,
273                         dev->rdev.stats.qid.max, dev->rdev.stats.qid.fail);
274         seq_printf(seq, "   TPTMEM: %10llu %10llu %10llu %10llu\n",
275                         dev->rdev.stats.stag.total, dev->rdev.stats.stag.cur,
276                         dev->rdev.stats.stag.max, dev->rdev.stats.stag.fail);
277         seq_printf(seq, "   PBLMEM: %10llu %10llu %10llu %10llu\n",
278                         dev->rdev.stats.pbl.total, dev->rdev.stats.pbl.cur,
279                         dev->rdev.stats.pbl.max, dev->rdev.stats.pbl.fail);
280         seq_printf(seq, "   RQTMEM: %10llu %10llu %10llu %10llu\n",
281                         dev->rdev.stats.rqt.total, dev->rdev.stats.rqt.cur,
282                         dev->rdev.stats.rqt.max, dev->rdev.stats.rqt.fail);
283         seq_printf(seq, "  OCQPMEM: %10llu %10llu %10llu %10llu\n",
284                         dev->rdev.stats.ocqp.total, dev->rdev.stats.ocqp.cur,
285                         dev->rdev.stats.ocqp.max, dev->rdev.stats.ocqp.fail);
286         seq_printf(seq, "  DB FULL: %10llu\n", dev->rdev.stats.db_full);
287         seq_printf(seq, " DB EMPTY: %10llu\n", dev->rdev.stats.db_empty);
288         seq_printf(seq, "  DB DROP: %10llu\n", dev->rdev.stats.db_drop);
289         seq_printf(seq, " DB State: %s Transitions %llu\n",
290                    db_state_str[dev->db_state],
291                    dev->rdev.stats.db_state_transitions);
292         seq_printf(seq, "TCAM_FULL: %10llu\n", dev->rdev.stats.tcam_full);
293         seq_printf(seq, "ACT_OFLD_CONN_FAILS: %10llu\n",
294                    dev->rdev.stats.act_ofld_conn_fails);
295         seq_printf(seq, "PAS_OFLD_CONN_FAILS: %10llu\n",
296                    dev->rdev.stats.pas_ofld_conn_fails);
297         return 0;
298 }
299
300 static int stats_open(struct inode *inode, struct file *file)
301 {
302         return single_open(file, stats_show, inode->i_private);
303 }
304
305 static ssize_t stats_clear(struct file *file, const char __user *buf,
306                 size_t count, loff_t *pos)
307 {
308         struct c4iw_dev *dev = ((struct seq_file *)file->private_data)->private;
309
310         mutex_lock(&dev->rdev.stats.lock);
311         dev->rdev.stats.pd.max = 0;
312         dev->rdev.stats.pd.fail = 0;
313         dev->rdev.stats.qid.max = 0;
314         dev->rdev.stats.qid.fail = 0;
315         dev->rdev.stats.stag.max = 0;
316         dev->rdev.stats.stag.fail = 0;
317         dev->rdev.stats.pbl.max = 0;
318         dev->rdev.stats.pbl.fail = 0;
319         dev->rdev.stats.rqt.max = 0;
320         dev->rdev.stats.rqt.fail = 0;
321         dev->rdev.stats.ocqp.max = 0;
322         dev->rdev.stats.ocqp.fail = 0;
323         dev->rdev.stats.db_full = 0;
324         dev->rdev.stats.db_empty = 0;
325         dev->rdev.stats.db_drop = 0;
326         dev->rdev.stats.db_state_transitions = 0;
327         dev->rdev.stats.tcam_full = 0;
328         dev->rdev.stats.act_ofld_conn_fails = 0;
329         dev->rdev.stats.pas_ofld_conn_fails = 0;
330         mutex_unlock(&dev->rdev.stats.lock);
331         return count;
332 }
333
334 static const struct file_operations stats_debugfs_fops = {
335         .owner   = THIS_MODULE,
336         .open    = stats_open,
337         .release = single_release,
338         .read    = seq_read,
339         .llseek  = seq_lseek,
340         .write   = stats_clear,
341 };
342
343 static int dump_ep(int id, void *p, void *data)
344 {
345         struct c4iw_ep *ep = p;
346         struct c4iw_debugfs_data *epd = data;
347         int space;
348         int cc;
349
350         space = epd->bufsize - epd->pos - 1;
351         if (space == 0)
352                 return 1;
353
354         cc = snprintf(epd->buf + epd->pos, space,
355                         "ep %p cm_id %p qp %p state %d flags 0x%lx history 0x%lx "
356                         "hwtid %d atid %d %pI4:%d <-> %pI4:%d\n",
357                         ep, ep->com.cm_id, ep->com.qp, (int)ep->com.state,
358                         ep->com.flags, ep->com.history, ep->hwtid, ep->atid,
359                         &ep->com.local_addr.sin_addr.s_addr,
360                         ntohs(ep->com.local_addr.sin_port),
361                         &ep->com.remote_addr.sin_addr.s_addr,
362                         ntohs(ep->com.remote_addr.sin_port));
363         if (cc < space)
364                 epd->pos += cc;
365         return 0;
366 }
367
368 static int dump_listen_ep(int id, void *p, void *data)
369 {
370         struct c4iw_listen_ep *ep = p;
371         struct c4iw_debugfs_data *epd = data;
372         int space;
373         int cc;
374
375         space = epd->bufsize - epd->pos - 1;
376         if (space == 0)
377                 return 1;
378
379         cc = snprintf(epd->buf + epd->pos, space,
380                         "ep %p cm_id %p state %d flags 0x%lx stid %d backlog %d "
381                         "%pI4:%d\n", ep, ep->com.cm_id, (int)ep->com.state,
382                         ep->com.flags, ep->stid, ep->backlog,
383                         &ep->com.local_addr.sin_addr.s_addr,
384                         ntohs(ep->com.local_addr.sin_port));
385         if (cc < space)
386                 epd->pos += cc;
387         return 0;
388 }
389
390 static int ep_release(struct inode *inode, struct file *file)
391 {
392         struct c4iw_debugfs_data *epd = file->private_data;
393         if (!epd) {
394                 pr_info("%s null qpd?\n", __func__);
395                 return 0;
396         }
397         vfree(epd->buf);
398         kfree(epd);
399         return 0;
400 }
401
402 static int ep_open(struct inode *inode, struct file *file)
403 {
404         struct c4iw_debugfs_data *epd;
405         int ret = 0;
406         int count = 1;
407
408         epd = kmalloc(sizeof(*epd), GFP_KERNEL);
409         if (!epd) {
410                 ret = -ENOMEM;
411                 goto out;
412         }
413         epd->devp = inode->i_private;
414         epd->pos = 0;
415
416         spin_lock_irq(&epd->devp->lock);
417         idr_for_each(&epd->devp->hwtid_idr, count_idrs, &count);
418         idr_for_each(&epd->devp->atid_idr, count_idrs, &count);
419         idr_for_each(&epd->devp->stid_idr, count_idrs, &count);
420         spin_unlock_irq(&epd->devp->lock);
421
422         epd->bufsize = count * 160;
423         epd->buf = vmalloc(epd->bufsize);
424         if (!epd->buf) {
425                 ret = -ENOMEM;
426                 goto err1;
427         }
428
429         spin_lock_irq(&epd->devp->lock);
430         idr_for_each(&epd->devp->hwtid_idr, dump_ep, epd);
431         idr_for_each(&epd->devp->atid_idr, dump_ep, epd);
432         idr_for_each(&epd->devp->stid_idr, dump_listen_ep, epd);
433         spin_unlock_irq(&epd->devp->lock);
434
435         file->private_data = epd;
436         goto out;
437 err1:
438         kfree(epd);
439 out:
440         return ret;
441 }
442
443 static const struct file_operations ep_debugfs_fops = {
444         .owner   = THIS_MODULE,
445         .open    = ep_open,
446         .release = ep_release,
447         .read    = debugfs_read,
448 };
449
450 static int setup_debugfs(struct c4iw_dev *devp)
451 {
452         struct dentry *de;
453
454         if (!devp->debugfs_root)
455                 return -1;
456
457         de = debugfs_create_file("qps", S_IWUSR, devp->debugfs_root,
458                                  (void *)devp, &qp_debugfs_fops);
459         if (de && de->d_inode)
460                 de->d_inode->i_size = 4096;
461
462         de = debugfs_create_file("stags", S_IWUSR, devp->debugfs_root,
463                                  (void *)devp, &stag_debugfs_fops);
464         if (de && de->d_inode)
465                 de->d_inode->i_size = 4096;
466
467         de = debugfs_create_file("stats", S_IWUSR, devp->debugfs_root,
468                         (void *)devp, &stats_debugfs_fops);
469         if (de && de->d_inode)
470                 de->d_inode->i_size = 4096;
471
472         de = debugfs_create_file("eps", S_IWUSR, devp->debugfs_root,
473                         (void *)devp, &ep_debugfs_fops);
474         if (de && de->d_inode)
475                 de->d_inode->i_size = 4096;
476
477         return 0;
478 }
479
480 void c4iw_release_dev_ucontext(struct c4iw_rdev *rdev,
481                                struct c4iw_dev_ucontext *uctx)
482 {
483         struct list_head *pos, *nxt;
484         struct c4iw_qid_list *entry;
485
486         mutex_lock(&uctx->lock);
487         list_for_each_safe(pos, nxt, &uctx->qpids) {
488                 entry = list_entry(pos, struct c4iw_qid_list, entry);
489                 list_del_init(&entry->entry);
490                 if (!(entry->qid & rdev->qpmask)) {
491                         c4iw_put_resource(&rdev->resource.qid_table,
492                                           entry->qid);
493                         mutex_lock(&rdev->stats.lock);
494                         rdev->stats.qid.cur -= rdev->qpmask + 1;
495                         mutex_unlock(&rdev->stats.lock);
496                 }
497                 kfree(entry);
498         }
499
500         list_for_each_safe(pos, nxt, &uctx->qpids) {
501                 entry = list_entry(pos, struct c4iw_qid_list, entry);
502                 list_del_init(&entry->entry);
503                 kfree(entry);
504         }
505         mutex_unlock(&uctx->lock);
506 }
507
508 void c4iw_init_dev_ucontext(struct c4iw_rdev *rdev,
509                             struct c4iw_dev_ucontext *uctx)
510 {
511         INIT_LIST_HEAD(&uctx->qpids);
512         INIT_LIST_HEAD(&uctx->cqids);
513         mutex_init(&uctx->lock);
514 }
515
516 /* Caller takes care of locking if needed */
517 static int c4iw_rdev_open(struct c4iw_rdev *rdev)
518 {
519         int err;
520
521         c4iw_init_dev_ucontext(rdev, &rdev->uctx);
522
523         /*
524          * qpshift is the number of bits to shift the qpid left in order
525          * to get the correct address of the doorbell for that qp.
526          */
527         rdev->qpshift = PAGE_SHIFT - ilog2(rdev->lldi.udb_density);
528         rdev->qpmask = rdev->lldi.udb_density - 1;
529         rdev->cqshift = PAGE_SHIFT - ilog2(rdev->lldi.ucq_density);
530         rdev->cqmask = rdev->lldi.ucq_density - 1;
531         PDBG("%s dev %s stag start 0x%0x size 0x%0x num stags %d "
532              "pbl start 0x%0x size 0x%0x rq start 0x%0x size 0x%0x "
533              "qp qid start %u size %u cq qid start %u size %u\n",
534              __func__, pci_name(rdev->lldi.pdev), rdev->lldi.vr->stag.start,
535              rdev->lldi.vr->stag.size, c4iw_num_stags(rdev),
536              rdev->lldi.vr->pbl.start,
537              rdev->lldi.vr->pbl.size, rdev->lldi.vr->rq.start,
538              rdev->lldi.vr->rq.size,
539              rdev->lldi.vr->qp.start,
540              rdev->lldi.vr->qp.size,
541              rdev->lldi.vr->cq.start,
542              rdev->lldi.vr->cq.size);
543         PDBG("udb len 0x%x udb base %p db_reg %p gts_reg %p qpshift %lu "
544              "qpmask 0x%x cqshift %lu cqmask 0x%x\n",
545              (unsigned)pci_resource_len(rdev->lldi.pdev, 2),
546              (void *)(unsigned long)pci_resource_start(rdev->lldi.pdev, 2),
547              rdev->lldi.db_reg,
548              rdev->lldi.gts_reg,
549              rdev->qpshift, rdev->qpmask,
550              rdev->cqshift, rdev->cqmask);
551
552         if (c4iw_num_stags(rdev) == 0) {
553                 err = -EINVAL;
554                 goto err1;
555         }
556
557         rdev->stats.pd.total = T4_MAX_NUM_PD;
558         rdev->stats.stag.total = rdev->lldi.vr->stag.size;
559         rdev->stats.pbl.total = rdev->lldi.vr->pbl.size;
560         rdev->stats.rqt.total = rdev->lldi.vr->rq.size;
561         rdev->stats.ocqp.total = rdev->lldi.vr->ocq.size;
562         rdev->stats.qid.total = rdev->lldi.vr->qp.size;
563
564         err = c4iw_init_resource(rdev, c4iw_num_stags(rdev), T4_MAX_NUM_PD);
565         if (err) {
566                 printk(KERN_ERR MOD "error %d initializing resources\n", err);
567                 goto err1;
568         }
569         err = c4iw_pblpool_create(rdev);
570         if (err) {
571                 printk(KERN_ERR MOD "error %d initializing pbl pool\n", err);
572                 goto err2;
573         }
574         err = c4iw_rqtpool_create(rdev);
575         if (err) {
576                 printk(KERN_ERR MOD "error %d initializing rqt pool\n", err);
577                 goto err3;
578         }
579         err = c4iw_ocqp_pool_create(rdev);
580         if (err) {
581                 printk(KERN_ERR MOD "error %d initializing ocqp pool\n", err);
582                 goto err4;
583         }
584         return 0;
585 err4:
586         c4iw_rqtpool_destroy(rdev);
587 err3:
588         c4iw_pblpool_destroy(rdev);
589 err2:
590         c4iw_destroy_resource(&rdev->resource);
591 err1:
592         return err;
593 }
594
595 static void c4iw_rdev_close(struct c4iw_rdev *rdev)
596 {
597         c4iw_pblpool_destroy(rdev);
598         c4iw_rqtpool_destroy(rdev);
599         c4iw_destroy_resource(&rdev->resource);
600 }
601
602 static void c4iw_dealloc(struct uld_ctx *ctx)
603 {
604         c4iw_rdev_close(&ctx->dev->rdev);
605         idr_destroy(&ctx->dev->cqidr);
606         idr_destroy(&ctx->dev->qpidr);
607         idr_destroy(&ctx->dev->mmidr);
608         idr_destroy(&ctx->dev->hwtid_idr);
609         idr_destroy(&ctx->dev->stid_idr);
610         idr_destroy(&ctx->dev->atid_idr);
611         iounmap(ctx->dev->rdev.oc_mw_kva);
612         ib_dealloc_device(&ctx->dev->ibdev);
613         ctx->dev = NULL;
614 }
615
616 static void c4iw_remove(struct uld_ctx *ctx)
617 {
618         PDBG("%s c4iw_dev %p\n", __func__,  ctx->dev);
619         c4iw_unregister_device(ctx->dev);
620         c4iw_dealloc(ctx);
621 }
622
623 static int rdma_supported(const struct cxgb4_lld_info *infop)
624 {
625         return infop->vr->stag.size > 0 && infop->vr->pbl.size > 0 &&
626                infop->vr->rq.size > 0 && infop->vr->qp.size > 0 &&
627                infop->vr->cq.size > 0;
628 }
629
630 static struct c4iw_dev *c4iw_alloc(const struct cxgb4_lld_info *infop)
631 {
632         struct c4iw_dev *devp;
633         int ret;
634
635         if (!rdma_supported(infop)) {
636                 printk(KERN_INFO MOD "%s: RDMA not supported on this device.\n",
637                        pci_name(infop->pdev));
638                 return ERR_PTR(-ENOSYS);
639         }
640         if (!ocqp_supported(infop))
641                 pr_info("%s: On-Chip Queues not supported on this device.\n",
642                         pci_name(infop->pdev));
643
644         if (!is_t4(infop->adapter_type)) {
645                 if (!allow_db_fc_on_t5) {
646                         db_fc_threshold = 100000;
647                         pr_info("DB Flow Control Disabled.\n");
648                 }
649
650                 if (!allow_db_coalescing_on_t5) {
651                         db_coalescing_threshold = -1;
652                         pr_info("DB Coalescing Disabled.\n");
653                 }
654         }
655
656         devp = (struct c4iw_dev *)ib_alloc_device(sizeof(*devp));
657         if (!devp) {
658                 printk(KERN_ERR MOD "Cannot allocate ib device\n");
659                 return ERR_PTR(-ENOMEM);
660         }
661         devp->rdev.lldi = *infop;
662
663         devp->rdev.oc_mw_pa = pci_resource_start(devp->rdev.lldi.pdev, 2) +
664                 (pci_resource_len(devp->rdev.lldi.pdev, 2) -
665                  roundup_pow_of_two(devp->rdev.lldi.vr->ocq.size));
666         devp->rdev.oc_mw_kva = ioremap_wc(devp->rdev.oc_mw_pa,
667                                                devp->rdev.lldi.vr->ocq.size);
668
669         PDBG(KERN_INFO MOD "ocq memory: "
670                "hw_start 0x%x size %u mw_pa 0x%lx mw_kva %p\n",
671                devp->rdev.lldi.vr->ocq.start, devp->rdev.lldi.vr->ocq.size,
672                devp->rdev.oc_mw_pa, devp->rdev.oc_mw_kva);
673
674         ret = c4iw_rdev_open(&devp->rdev);
675         if (ret) {
676                 printk(KERN_ERR MOD "Unable to open CXIO rdev err %d\n", ret);
677                 ib_dealloc_device(&devp->ibdev);
678                 return ERR_PTR(ret);
679         }
680
681         idr_init(&devp->cqidr);
682         idr_init(&devp->qpidr);
683         idr_init(&devp->mmidr);
684         idr_init(&devp->hwtid_idr);
685         idr_init(&devp->stid_idr);
686         idr_init(&devp->atid_idr);
687         spin_lock_init(&devp->lock);
688         mutex_init(&devp->rdev.stats.lock);
689         mutex_init(&devp->db_mutex);
690
691         if (c4iw_debugfs_root) {
692                 devp->debugfs_root = debugfs_create_dir(
693                                         pci_name(devp->rdev.lldi.pdev),
694                                         c4iw_debugfs_root);
695                 setup_debugfs(devp);
696         }
697         return devp;
698 }
699
700 static void *c4iw_uld_add(const struct cxgb4_lld_info *infop)
701 {
702         struct uld_ctx *ctx;
703         static int vers_printed;
704         int i;
705
706         if (!vers_printed++)
707                 pr_info("Chelsio T4/T5 RDMA Driver - version %s\n",
708                         DRV_VERSION);
709
710         ctx = kzalloc(sizeof *ctx, GFP_KERNEL);
711         if (!ctx) {
712                 ctx = ERR_PTR(-ENOMEM);
713                 goto out;
714         }
715         ctx->lldi = *infop;
716
717         PDBG("%s found device %s nchan %u nrxq %u ntxq %u nports %u\n",
718              __func__, pci_name(ctx->lldi.pdev),
719              ctx->lldi.nchan, ctx->lldi.nrxq,
720              ctx->lldi.ntxq, ctx->lldi.nports);
721
722         mutex_lock(&dev_mutex);
723         list_add_tail(&ctx->entry, &uld_ctx_list);
724         mutex_unlock(&dev_mutex);
725
726         for (i = 0; i < ctx->lldi.nrxq; i++)
727                 PDBG("rxqid[%u] %u\n", i, ctx->lldi.rxq_ids[i]);
728 out:
729         return ctx;
730 }
731
732 static inline struct sk_buff *copy_gl_to_skb_pkt(const struct pkt_gl *gl,
733                                                  const __be64 *rsp,
734                                                  u32 pktshift)
735 {
736         struct sk_buff *skb;
737
738         /*
739          * Allocate space for cpl_pass_accept_req which will be synthesized by
740          * driver. Once the driver synthesizes the request the skb will go
741          * through the regular cpl_pass_accept_req processing.
742          * The math here assumes sizeof cpl_pass_accept_req >= sizeof
743          * cpl_rx_pkt.
744          */
745         skb = alloc_skb(gl->tot_len + sizeof(struct cpl_pass_accept_req) +
746                         sizeof(struct rss_header) - pktshift, GFP_ATOMIC);
747         if (unlikely(!skb))
748                 return NULL;
749
750          __skb_put(skb, gl->tot_len + sizeof(struct cpl_pass_accept_req) +
751                    sizeof(struct rss_header) - pktshift);
752
753         /*
754          * This skb will contain:
755          *   rss_header from the rspq descriptor (1 flit)
756          *   cpl_rx_pkt struct from the rspq descriptor (2 flits)
757          *   space for the difference between the size of an
758          *      rx_pkt and pass_accept_req cpl (1 flit)
759          *   the packet data from the gl
760          */
761         skb_copy_to_linear_data(skb, rsp, sizeof(struct cpl_pass_accept_req) +
762                                 sizeof(struct rss_header));
763         skb_copy_to_linear_data_offset(skb, sizeof(struct rss_header) +
764                                        sizeof(struct cpl_pass_accept_req),
765                                        gl->va + pktshift,
766                                        gl->tot_len - pktshift);
767         return skb;
768 }
769
770 static inline int recv_rx_pkt(struct c4iw_dev *dev, const struct pkt_gl *gl,
771                            const __be64 *rsp)
772 {
773         unsigned int opcode = *(u8 *)rsp;
774         struct sk_buff *skb;
775
776         if (opcode != CPL_RX_PKT)
777                 goto out;
778
779         skb = copy_gl_to_skb_pkt(gl , rsp, dev->rdev.lldi.sge_pktshift);
780         if (skb == NULL)
781                 goto out;
782
783         if (c4iw_handlers[opcode] == NULL) {
784                 pr_info("%s no handler opcode 0x%x...\n", __func__,
785                        opcode);
786                 kfree_skb(skb);
787                 goto out;
788         }
789         c4iw_handlers[opcode](dev, skb);
790         return 1;
791 out:
792         return 0;
793 }
794
795 static int c4iw_uld_rx_handler(void *handle, const __be64 *rsp,
796                         const struct pkt_gl *gl)
797 {
798         struct uld_ctx *ctx = handle;
799         struct c4iw_dev *dev = ctx->dev;
800         struct sk_buff *skb;
801         u8 opcode;
802
803         if (gl == NULL) {
804                 /* omit RSS and rsp_ctrl at end of descriptor */
805                 unsigned int len = 64 - sizeof(struct rsp_ctrl) - 8;
806
807                 skb = alloc_skb(256, GFP_ATOMIC);
808                 if (!skb)
809                         goto nomem;
810                 __skb_put(skb, len);
811                 skb_copy_to_linear_data(skb, &rsp[1], len);
812         } else if (gl == CXGB4_MSG_AN) {
813                 const struct rsp_ctrl *rc = (void *)rsp;
814
815                 u32 qid = be32_to_cpu(rc->pldbuflen_qid);
816                 c4iw_ev_handler(dev, qid);
817                 return 0;
818         } else if (unlikely(*(u8 *)rsp != *(u8 *)gl->va)) {
819                 if (recv_rx_pkt(dev, gl, rsp))
820                         return 0;
821
822                 pr_info("%s: unexpected FL contents at %p, " \
823                        "RSS %#llx, FL %#llx, len %u\n",
824                        pci_name(ctx->lldi.pdev), gl->va,
825                        (unsigned long long)be64_to_cpu(*rsp),
826                        (unsigned long long)be64_to_cpu(
827                        *(__force __be64 *)gl->va),
828                        gl->tot_len);
829
830                 return 0;
831         } else {
832                 skb = cxgb4_pktgl_to_skb(gl, 128, 128);
833                 if (unlikely(!skb))
834                         goto nomem;
835         }
836
837         opcode = *(u8 *)rsp;
838         if (c4iw_handlers[opcode])
839                 c4iw_handlers[opcode](dev, skb);
840         else
841                 pr_info("%s no handler opcode 0x%x...\n", __func__,
842                        opcode);
843
844         return 0;
845 nomem:
846         return -1;
847 }
848
849 static int c4iw_uld_state_change(void *handle, enum cxgb4_state new_state)
850 {
851         struct uld_ctx *ctx = handle;
852
853         PDBG("%s new_state %u\n", __func__, new_state);
854         switch (new_state) {
855         case CXGB4_STATE_UP:
856                 printk(KERN_INFO MOD "%s: Up\n", pci_name(ctx->lldi.pdev));
857                 if (!ctx->dev) {
858                         int ret;
859
860                         ctx->dev = c4iw_alloc(&ctx->lldi);
861                         if (IS_ERR(ctx->dev)) {
862                                 printk(KERN_ERR MOD
863                                        "%s: initialization failed: %ld\n",
864                                        pci_name(ctx->lldi.pdev),
865                                        PTR_ERR(ctx->dev));
866                                 ctx->dev = NULL;
867                                 break;
868                         }
869                         ret = c4iw_register_device(ctx->dev);
870                         if (ret) {
871                                 printk(KERN_ERR MOD
872                                        "%s: RDMA registration failed: %d\n",
873                                        pci_name(ctx->lldi.pdev), ret);
874                                 c4iw_dealloc(ctx);
875                         }
876                 }
877                 break;
878         case CXGB4_STATE_DOWN:
879                 printk(KERN_INFO MOD "%s: Down\n",
880                        pci_name(ctx->lldi.pdev));
881                 if (ctx->dev)
882                         c4iw_remove(ctx);
883                 break;
884         case CXGB4_STATE_START_RECOVERY:
885                 printk(KERN_INFO MOD "%s: Fatal Error\n",
886                        pci_name(ctx->lldi.pdev));
887                 if (ctx->dev) {
888                         struct ib_event event;
889
890                         ctx->dev->rdev.flags |= T4_FATAL_ERROR;
891                         memset(&event, 0, sizeof event);
892                         event.event  = IB_EVENT_DEVICE_FATAL;
893                         event.device = &ctx->dev->ibdev;
894                         ib_dispatch_event(&event);
895                         c4iw_remove(ctx);
896                 }
897                 break;
898         case CXGB4_STATE_DETACH:
899                 printk(KERN_INFO MOD "%s: Detach\n",
900                        pci_name(ctx->lldi.pdev));
901                 if (ctx->dev)
902                         c4iw_remove(ctx);
903                 break;
904         }
905         return 0;
906 }
907
908 static int disable_qp_db(int id, void *p, void *data)
909 {
910         struct c4iw_qp *qp = p;
911
912         t4_disable_wq_db(&qp->wq);
913         return 0;
914 }
915
916 static void stop_queues(struct uld_ctx *ctx)
917 {
918         spin_lock_irq(&ctx->dev->lock);
919         if (ctx->dev->db_state == NORMAL) {
920                 ctx->dev->rdev.stats.db_state_transitions++;
921                 ctx->dev->db_state = FLOW_CONTROL;
922                 idr_for_each(&ctx->dev->qpidr, disable_qp_db, NULL);
923         }
924         spin_unlock_irq(&ctx->dev->lock);
925 }
926
927 static int enable_qp_db(int id, void *p, void *data)
928 {
929         struct c4iw_qp *qp = p;
930
931         t4_enable_wq_db(&qp->wq);
932         return 0;
933 }
934
935 static void resume_queues(struct uld_ctx *ctx)
936 {
937         spin_lock_irq(&ctx->dev->lock);
938         if (ctx->dev->qpcnt <= db_fc_threshold &&
939             ctx->dev->db_state == FLOW_CONTROL) {
940                 ctx->dev->db_state = NORMAL;
941                 ctx->dev->rdev.stats.db_state_transitions++;
942                 idr_for_each(&ctx->dev->qpidr, enable_qp_db, NULL);
943         }
944         spin_unlock_irq(&ctx->dev->lock);
945 }
946
947 struct qp_list {
948         unsigned idx;
949         struct c4iw_qp **qps;
950 };
951
952 static int add_and_ref_qp(int id, void *p, void *data)
953 {
954         struct qp_list *qp_listp = data;
955         struct c4iw_qp *qp = p;
956
957         c4iw_qp_add_ref(&qp->ibqp);
958         qp_listp->qps[qp_listp->idx++] = qp;
959         return 0;
960 }
961
962 static int count_qps(int id, void *p, void *data)
963 {
964         unsigned *countp = data;
965         (*countp)++;
966         return 0;
967 }
968
969 static void deref_qps(struct qp_list qp_list)
970 {
971         int idx;
972
973         for (idx = 0; idx < qp_list.idx; idx++)
974                 c4iw_qp_rem_ref(&qp_list.qps[idx]->ibqp);
975 }
976
977 static void recover_lost_dbs(struct uld_ctx *ctx, struct qp_list *qp_list)
978 {
979         int idx;
980         int ret;
981
982         for (idx = 0; idx < qp_list->idx; idx++) {
983                 struct c4iw_qp *qp = qp_list->qps[idx];
984
985                 ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
986                                           qp->wq.sq.qid,
987                                           t4_sq_host_wq_pidx(&qp->wq),
988                                           t4_sq_wq_size(&qp->wq));
989                 if (ret) {
990                         printk(KERN_ERR MOD "%s: Fatal error - "
991                                "DB overflow recovery failed - "
992                                "error syncing SQ qid %u\n",
993                                pci_name(ctx->lldi.pdev), qp->wq.sq.qid);
994                         return;
995                 }
996
997                 ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
998                                           qp->wq.rq.qid,
999                                           t4_rq_host_wq_pidx(&qp->wq),
1000                                           t4_rq_wq_size(&qp->wq));
1001
1002                 if (ret) {
1003                         printk(KERN_ERR MOD "%s: Fatal error - "
1004                                "DB overflow recovery failed - "
1005                                "error syncing RQ qid %u\n",
1006                                pci_name(ctx->lldi.pdev), qp->wq.rq.qid);
1007                         return;
1008                 }
1009
1010                 /* Wait for the dbfifo to drain */
1011                 while (cxgb4_dbfifo_count(qp->rhp->rdev.lldi.ports[0], 1) > 0) {
1012                         set_current_state(TASK_UNINTERRUPTIBLE);
1013                         schedule_timeout(usecs_to_jiffies(10));
1014                 }
1015         }
1016 }
1017
1018 static void recover_queues(struct uld_ctx *ctx)
1019 {
1020         int count = 0;
1021         struct qp_list qp_list;
1022         int ret;
1023
1024         /* lock out kernel db ringers */
1025         mutex_lock(&ctx->dev->db_mutex);
1026
1027         /* put all queues in to recovery mode */
1028         spin_lock_irq(&ctx->dev->lock);
1029         ctx->dev->db_state = RECOVERY;
1030         ctx->dev->rdev.stats.db_state_transitions++;
1031         idr_for_each(&ctx->dev->qpidr, disable_qp_db, NULL);
1032         spin_unlock_irq(&ctx->dev->lock);
1033
1034         /* slow everybody down */
1035         set_current_state(TASK_UNINTERRUPTIBLE);
1036         schedule_timeout(usecs_to_jiffies(1000));
1037
1038         /* Wait for the dbfifo to completely drain. */
1039         while (cxgb4_dbfifo_count(ctx->dev->rdev.lldi.ports[0], 1) > 0) {
1040                 set_current_state(TASK_UNINTERRUPTIBLE);
1041                 schedule_timeout(usecs_to_jiffies(10));
1042         }
1043
1044         /* flush the SGE contexts */
1045         ret = cxgb4_flush_eq_cache(ctx->dev->rdev.lldi.ports[0]);
1046         if (ret) {
1047                 printk(KERN_ERR MOD "%s: Fatal error - DB overflow recovery failed\n",
1048                        pci_name(ctx->lldi.pdev));
1049                 goto out;
1050         }
1051
1052         /* Count active queues so we can build a list of queues to recover */
1053         spin_lock_irq(&ctx->dev->lock);
1054         idr_for_each(&ctx->dev->qpidr, count_qps, &count);
1055
1056         qp_list.qps = kzalloc(count * sizeof *qp_list.qps, GFP_ATOMIC);
1057         if (!qp_list.qps) {
1058                 printk(KERN_ERR MOD "%s: Fatal error - DB overflow recovery failed\n",
1059                        pci_name(ctx->lldi.pdev));
1060                 spin_unlock_irq(&ctx->dev->lock);
1061                 goto out;
1062         }
1063         qp_list.idx = 0;
1064
1065         /* add and ref each qp so it doesn't get freed */
1066         idr_for_each(&ctx->dev->qpidr, add_and_ref_qp, &qp_list);
1067
1068         spin_unlock_irq(&ctx->dev->lock);
1069
1070         /* now traverse the list in a safe context to recover the db state*/
1071         recover_lost_dbs(ctx, &qp_list);
1072
1073         /* we're almost done!  deref the qps and clean up */
1074         deref_qps(qp_list);
1075         kfree(qp_list.qps);
1076
1077         /* Wait for the dbfifo to completely drain again */
1078         while (cxgb4_dbfifo_count(ctx->dev->rdev.lldi.ports[0], 1) > 0) {
1079                 set_current_state(TASK_UNINTERRUPTIBLE);
1080                 schedule_timeout(usecs_to_jiffies(10));
1081         }
1082
1083         /* resume the queues */
1084         spin_lock_irq(&ctx->dev->lock);
1085         if (ctx->dev->qpcnt > db_fc_threshold)
1086                 ctx->dev->db_state = FLOW_CONTROL;
1087         else {
1088                 ctx->dev->db_state = NORMAL;
1089                 idr_for_each(&ctx->dev->qpidr, enable_qp_db, NULL);
1090         }
1091         ctx->dev->rdev.stats.db_state_transitions++;
1092         spin_unlock_irq(&ctx->dev->lock);
1093
1094 out:
1095         /* start up kernel db ringers again */
1096         mutex_unlock(&ctx->dev->db_mutex);
1097 }
1098
1099 static int c4iw_uld_control(void *handle, enum cxgb4_control control, ...)
1100 {
1101         struct uld_ctx *ctx = handle;
1102
1103         switch (control) {
1104         case CXGB4_CONTROL_DB_FULL:
1105                 stop_queues(ctx);
1106                 mutex_lock(&ctx->dev->rdev.stats.lock);
1107                 ctx->dev->rdev.stats.db_full++;
1108                 mutex_unlock(&ctx->dev->rdev.stats.lock);
1109                 break;
1110         case CXGB4_CONTROL_DB_EMPTY:
1111                 resume_queues(ctx);
1112                 mutex_lock(&ctx->dev->rdev.stats.lock);
1113                 ctx->dev->rdev.stats.db_empty++;
1114                 mutex_unlock(&ctx->dev->rdev.stats.lock);
1115                 break;
1116         case CXGB4_CONTROL_DB_DROP:
1117                 recover_queues(ctx);
1118                 mutex_lock(&ctx->dev->rdev.stats.lock);
1119                 ctx->dev->rdev.stats.db_drop++;
1120                 mutex_unlock(&ctx->dev->rdev.stats.lock);
1121                 break;
1122         default:
1123                 printk(KERN_WARNING MOD "%s: unknown control cmd %u\n",
1124                        pci_name(ctx->lldi.pdev), control);
1125                 break;
1126         }
1127         return 0;
1128 }
1129
1130 static struct cxgb4_uld_info c4iw_uld_info = {
1131         .name = DRV_NAME,
1132         .add = c4iw_uld_add,
1133         .rx_handler = c4iw_uld_rx_handler,
1134         .state_change = c4iw_uld_state_change,
1135         .control = c4iw_uld_control,
1136 };
1137
1138 static int __init c4iw_init_module(void)
1139 {
1140         int err;
1141
1142         err = c4iw_cm_init();
1143         if (err)
1144                 return err;
1145
1146         c4iw_debugfs_root = debugfs_create_dir(DRV_NAME, NULL);
1147         if (!c4iw_debugfs_root)
1148                 printk(KERN_WARNING MOD
1149                        "could not create debugfs entry, continuing\n");
1150
1151         cxgb4_register_uld(CXGB4_ULD_RDMA, &c4iw_uld_info);
1152
1153         return 0;
1154 }
1155
1156 static void __exit c4iw_exit_module(void)
1157 {
1158         struct uld_ctx *ctx, *tmp;
1159
1160         mutex_lock(&dev_mutex);
1161         list_for_each_entry_safe(ctx, tmp, &uld_ctx_list, entry) {
1162                 if (ctx->dev)
1163                         c4iw_remove(ctx);
1164                 kfree(ctx);
1165         }
1166         mutex_unlock(&dev_mutex);
1167         cxgb4_unregister_uld(CXGB4_ULD_RDMA);
1168         c4iw_cm_term();
1169         debugfs_remove_recursive(c4iw_debugfs_root);
1170 }
1171
1172 module_init(c4iw_init_module);
1173 module_exit(c4iw_exit_module);