]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - drivers/infiniband/hw/cxgb4/device.c
RDMA/cxgb4: Fix LE hash collision bug for passive open connection
[~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         return 0;
284 }
285
286 static int stats_open(struct inode *inode, struct file *file)
287 {
288         return single_open(file, stats_show, inode->i_private);
289 }
290
291 static ssize_t stats_clear(struct file *file, const char __user *buf,
292                 size_t count, loff_t *pos)
293 {
294         struct c4iw_dev *dev = ((struct seq_file *)file->private_data)->private;
295
296         mutex_lock(&dev->rdev.stats.lock);
297         dev->rdev.stats.pd.max = 0;
298         dev->rdev.stats.pd.fail = 0;
299         dev->rdev.stats.qid.max = 0;
300         dev->rdev.stats.qid.fail = 0;
301         dev->rdev.stats.stag.max = 0;
302         dev->rdev.stats.stag.fail = 0;
303         dev->rdev.stats.pbl.max = 0;
304         dev->rdev.stats.pbl.fail = 0;
305         dev->rdev.stats.rqt.max = 0;
306         dev->rdev.stats.rqt.fail = 0;
307         dev->rdev.stats.ocqp.max = 0;
308         dev->rdev.stats.ocqp.fail = 0;
309         dev->rdev.stats.db_full = 0;
310         dev->rdev.stats.db_empty = 0;
311         dev->rdev.stats.db_drop = 0;
312         dev->rdev.stats.db_state_transitions = 0;
313         mutex_unlock(&dev->rdev.stats.lock);
314         return count;
315 }
316
317 static const struct file_operations stats_debugfs_fops = {
318         .owner   = THIS_MODULE,
319         .open    = stats_open,
320         .release = single_release,
321         .read    = seq_read,
322         .llseek  = seq_lseek,
323         .write   = stats_clear,
324 };
325
326 static int setup_debugfs(struct c4iw_dev *devp)
327 {
328         struct dentry *de;
329
330         if (!devp->debugfs_root)
331                 return -1;
332
333         de = debugfs_create_file("qps", S_IWUSR, devp->debugfs_root,
334                                  (void *)devp, &qp_debugfs_fops);
335         if (de && de->d_inode)
336                 de->d_inode->i_size = 4096;
337
338         de = debugfs_create_file("stags", S_IWUSR, devp->debugfs_root,
339                                  (void *)devp, &stag_debugfs_fops);
340         if (de && de->d_inode)
341                 de->d_inode->i_size = 4096;
342
343         de = debugfs_create_file("stats", S_IWUSR, devp->debugfs_root,
344                         (void *)devp, &stats_debugfs_fops);
345         if (de && de->d_inode)
346                 de->d_inode->i_size = 4096;
347
348         return 0;
349 }
350
351 void c4iw_release_dev_ucontext(struct c4iw_rdev *rdev,
352                                struct c4iw_dev_ucontext *uctx)
353 {
354         struct list_head *pos, *nxt;
355         struct c4iw_qid_list *entry;
356
357         mutex_lock(&uctx->lock);
358         list_for_each_safe(pos, nxt, &uctx->qpids) {
359                 entry = list_entry(pos, struct c4iw_qid_list, entry);
360                 list_del_init(&entry->entry);
361                 if (!(entry->qid & rdev->qpmask)) {
362                         c4iw_put_resource(&rdev->resource.qid_table,
363                                           entry->qid);
364                         mutex_lock(&rdev->stats.lock);
365                         rdev->stats.qid.cur -= rdev->qpmask + 1;
366                         mutex_unlock(&rdev->stats.lock);
367                 }
368                 kfree(entry);
369         }
370
371         list_for_each_safe(pos, nxt, &uctx->qpids) {
372                 entry = list_entry(pos, struct c4iw_qid_list, entry);
373                 list_del_init(&entry->entry);
374                 kfree(entry);
375         }
376         mutex_unlock(&uctx->lock);
377 }
378
379 void c4iw_init_dev_ucontext(struct c4iw_rdev *rdev,
380                             struct c4iw_dev_ucontext *uctx)
381 {
382         INIT_LIST_HEAD(&uctx->qpids);
383         INIT_LIST_HEAD(&uctx->cqids);
384         mutex_init(&uctx->lock);
385 }
386
387 /* Caller takes care of locking if needed */
388 static int c4iw_rdev_open(struct c4iw_rdev *rdev)
389 {
390         int err;
391
392         c4iw_init_dev_ucontext(rdev, &rdev->uctx);
393
394         /*
395          * qpshift is the number of bits to shift the qpid left in order
396          * to get the correct address of the doorbell for that qp.
397          */
398         rdev->qpshift = PAGE_SHIFT - ilog2(rdev->lldi.udb_density);
399         rdev->qpmask = rdev->lldi.udb_density - 1;
400         rdev->cqshift = PAGE_SHIFT - ilog2(rdev->lldi.ucq_density);
401         rdev->cqmask = rdev->lldi.ucq_density - 1;
402         PDBG("%s dev %s stag start 0x%0x size 0x%0x num stags %d "
403              "pbl start 0x%0x size 0x%0x rq start 0x%0x size 0x%0x "
404              "qp qid start %u size %u cq qid start %u size %u\n",
405              __func__, pci_name(rdev->lldi.pdev), rdev->lldi.vr->stag.start,
406              rdev->lldi.vr->stag.size, c4iw_num_stags(rdev),
407              rdev->lldi.vr->pbl.start,
408              rdev->lldi.vr->pbl.size, rdev->lldi.vr->rq.start,
409              rdev->lldi.vr->rq.size,
410              rdev->lldi.vr->qp.start,
411              rdev->lldi.vr->qp.size,
412              rdev->lldi.vr->cq.start,
413              rdev->lldi.vr->cq.size);
414         PDBG("udb len 0x%x udb base %p db_reg %p gts_reg %p qpshift %lu "
415              "qpmask 0x%x cqshift %lu cqmask 0x%x\n",
416              (unsigned)pci_resource_len(rdev->lldi.pdev, 2),
417              (void *)pci_resource_start(rdev->lldi.pdev, 2),
418              rdev->lldi.db_reg,
419              rdev->lldi.gts_reg,
420              rdev->qpshift, rdev->qpmask,
421              rdev->cqshift, rdev->cqmask);
422
423         if (c4iw_num_stags(rdev) == 0) {
424                 err = -EINVAL;
425                 goto err1;
426         }
427
428         rdev->stats.pd.total = T4_MAX_NUM_PD;
429         rdev->stats.stag.total = rdev->lldi.vr->stag.size;
430         rdev->stats.pbl.total = rdev->lldi.vr->pbl.size;
431         rdev->stats.rqt.total = rdev->lldi.vr->rq.size;
432         rdev->stats.ocqp.total = rdev->lldi.vr->ocq.size;
433         rdev->stats.qid.total = rdev->lldi.vr->qp.size;
434
435         err = c4iw_init_resource(rdev, c4iw_num_stags(rdev), T4_MAX_NUM_PD);
436         if (err) {
437                 printk(KERN_ERR MOD "error %d initializing resources\n", err);
438                 goto err1;
439         }
440         err = c4iw_pblpool_create(rdev);
441         if (err) {
442                 printk(KERN_ERR MOD "error %d initializing pbl pool\n", err);
443                 goto err2;
444         }
445         err = c4iw_rqtpool_create(rdev);
446         if (err) {
447                 printk(KERN_ERR MOD "error %d initializing rqt pool\n", err);
448                 goto err3;
449         }
450         err = c4iw_ocqp_pool_create(rdev);
451         if (err) {
452                 printk(KERN_ERR MOD "error %d initializing ocqp pool\n", err);
453                 goto err4;
454         }
455         return 0;
456 err4:
457         c4iw_rqtpool_destroy(rdev);
458 err3:
459         c4iw_pblpool_destroy(rdev);
460 err2:
461         c4iw_destroy_resource(&rdev->resource);
462 err1:
463         return err;
464 }
465
466 static void c4iw_rdev_close(struct c4iw_rdev *rdev)
467 {
468         c4iw_pblpool_destroy(rdev);
469         c4iw_rqtpool_destroy(rdev);
470         c4iw_destroy_resource(&rdev->resource);
471 }
472
473 static void c4iw_dealloc(struct uld_ctx *ctx)
474 {
475         c4iw_rdev_close(&ctx->dev->rdev);
476         idr_destroy(&ctx->dev->cqidr);
477         idr_destroy(&ctx->dev->qpidr);
478         idr_destroy(&ctx->dev->mmidr);
479         iounmap(ctx->dev->rdev.oc_mw_kva);
480         ib_dealloc_device(&ctx->dev->ibdev);
481         ctx->dev = NULL;
482 }
483
484 static void c4iw_remove(struct uld_ctx *ctx)
485 {
486         PDBG("%s c4iw_dev %p\n", __func__,  ctx->dev);
487         c4iw_unregister_device(ctx->dev);
488         c4iw_dealloc(ctx);
489 }
490
491 static int rdma_supported(const struct cxgb4_lld_info *infop)
492 {
493         return infop->vr->stag.size > 0 && infop->vr->pbl.size > 0 &&
494                infop->vr->rq.size > 0 && infop->vr->qp.size > 0 &&
495                infop->vr->cq.size > 0 && infop->vr->ocq.size > 0;
496 }
497
498 static struct c4iw_dev *c4iw_alloc(const struct cxgb4_lld_info *infop)
499 {
500         struct c4iw_dev *devp;
501         int ret;
502
503         if (!rdma_supported(infop)) {
504                 printk(KERN_INFO MOD "%s: RDMA not supported on this device.\n",
505                        pci_name(infop->pdev));
506                 return ERR_PTR(-ENOSYS);
507         }
508         devp = (struct c4iw_dev *)ib_alloc_device(sizeof(*devp));
509         if (!devp) {
510                 printk(KERN_ERR MOD "Cannot allocate ib device\n");
511                 return ERR_PTR(-ENOMEM);
512         }
513         devp->rdev.lldi = *infop;
514
515         devp->rdev.oc_mw_pa = pci_resource_start(devp->rdev.lldi.pdev, 2) +
516                 (pci_resource_len(devp->rdev.lldi.pdev, 2) -
517                  roundup_pow_of_two(devp->rdev.lldi.vr->ocq.size));
518         devp->rdev.oc_mw_kva = ioremap_wc(devp->rdev.oc_mw_pa,
519                                                devp->rdev.lldi.vr->ocq.size);
520
521         PDBG(KERN_INFO MOD "ocq memory: "
522                "hw_start 0x%x size %u mw_pa 0x%lx mw_kva %p\n",
523                devp->rdev.lldi.vr->ocq.start, devp->rdev.lldi.vr->ocq.size,
524                devp->rdev.oc_mw_pa, devp->rdev.oc_mw_kva);
525
526         ret = c4iw_rdev_open(&devp->rdev);
527         if (ret) {
528                 printk(KERN_ERR MOD "Unable to open CXIO rdev err %d\n", ret);
529                 ib_dealloc_device(&devp->ibdev);
530                 return ERR_PTR(ret);
531         }
532
533         idr_init(&devp->cqidr);
534         idr_init(&devp->qpidr);
535         idr_init(&devp->mmidr);
536         spin_lock_init(&devp->lock);
537         mutex_init(&devp->rdev.stats.lock);
538         mutex_init(&devp->db_mutex);
539
540         if (c4iw_debugfs_root) {
541                 devp->debugfs_root = debugfs_create_dir(
542                                         pci_name(devp->rdev.lldi.pdev),
543                                         c4iw_debugfs_root);
544                 setup_debugfs(devp);
545         }
546         return devp;
547 }
548
549 static void *c4iw_uld_add(const struct cxgb4_lld_info *infop)
550 {
551         struct uld_ctx *ctx;
552         static int vers_printed;
553         int i;
554
555         if (!vers_printed++)
556                 printk(KERN_INFO MOD "Chelsio T4 RDMA Driver - version %s\n",
557                        DRV_VERSION);
558
559         ctx = kzalloc(sizeof *ctx, GFP_KERNEL);
560         if (!ctx) {
561                 ctx = ERR_PTR(-ENOMEM);
562                 goto out;
563         }
564         ctx->lldi = *infop;
565
566         PDBG("%s found device %s nchan %u nrxq %u ntxq %u nports %u\n",
567              __func__, pci_name(ctx->lldi.pdev),
568              ctx->lldi.nchan, ctx->lldi.nrxq,
569              ctx->lldi.ntxq, ctx->lldi.nports);
570
571         mutex_lock(&dev_mutex);
572         list_add_tail(&ctx->entry, &uld_ctx_list);
573         mutex_unlock(&dev_mutex);
574
575         for (i = 0; i < ctx->lldi.nrxq; i++)
576                 PDBG("rxqid[%u] %u\n", i, ctx->lldi.rxq_ids[i]);
577 out:
578         return ctx;
579 }
580
581 static inline struct sk_buff *copy_gl_to_skb_pkt(const struct pkt_gl *gl,
582                                                  const __be64 *rsp,
583                                                  u32 pktshift)
584 {
585         struct sk_buff *skb;
586
587         /*
588          * Allocate space for cpl_pass_accept_req which will be synthesized by
589          * driver. Once the driver synthesizes the request the skb will go
590          * through the regular cpl_pass_accept_req processing.
591          * The math here assumes sizeof cpl_pass_accept_req >= sizeof
592          * cpl_rx_pkt.
593          */
594         skb = alloc_skb(gl->tot_len + sizeof(struct cpl_pass_accept_req) +
595                         sizeof(struct rss_header) - pktshift, GFP_ATOMIC);
596         if (unlikely(!skb))
597                 return NULL;
598
599          __skb_put(skb, gl->tot_len + sizeof(struct cpl_pass_accept_req) +
600                    sizeof(struct rss_header) - pktshift);
601
602         /*
603          * This skb will contain:
604          *   rss_header from the rspq descriptor (1 flit)
605          *   cpl_rx_pkt struct from the rspq descriptor (2 flits)
606          *   space for the difference between the size of an
607          *      rx_pkt and pass_accept_req cpl (1 flit)
608          *   the packet data from the gl
609          */
610         skb_copy_to_linear_data(skb, rsp, sizeof(struct cpl_pass_accept_req) +
611                                 sizeof(struct rss_header));
612         skb_copy_to_linear_data_offset(skb, sizeof(struct rss_header) +
613                                        sizeof(struct cpl_pass_accept_req),
614                                        gl->va + pktshift,
615                                        gl->tot_len - pktshift);
616         return skb;
617 }
618
619 static inline int recv_rx_pkt(struct c4iw_dev *dev, const struct pkt_gl *gl,
620                            const __be64 *rsp)
621 {
622         unsigned int opcode = *(u8 *)rsp;
623         struct sk_buff *skb;
624
625         if (opcode != CPL_RX_PKT)
626                 goto out;
627
628         skb = copy_gl_to_skb_pkt(gl , rsp, dev->rdev.lldi.sge_pktshift);
629         if (skb == NULL)
630                 goto out;
631
632         if (c4iw_handlers[opcode] == NULL) {
633                 pr_info("%s no handler opcode 0x%x...\n", __func__,
634                        opcode);
635                 kfree_skb(skb);
636                 goto out;
637         }
638         c4iw_handlers[opcode](dev, skb);
639         return 1;
640 out:
641         return 0;
642 }
643
644 static int c4iw_uld_rx_handler(void *handle, const __be64 *rsp,
645                         const struct pkt_gl *gl)
646 {
647         struct uld_ctx *ctx = handle;
648         struct c4iw_dev *dev = ctx->dev;
649         struct sk_buff *skb;
650         u8 opcode;
651
652         if (gl == NULL) {
653                 /* omit RSS and rsp_ctrl at end of descriptor */
654                 unsigned int len = 64 - sizeof(struct rsp_ctrl) - 8;
655
656                 skb = alloc_skb(256, GFP_ATOMIC);
657                 if (!skb)
658                         goto nomem;
659                 __skb_put(skb, len);
660                 skb_copy_to_linear_data(skb, &rsp[1], len);
661         } else if (gl == CXGB4_MSG_AN) {
662                 const struct rsp_ctrl *rc = (void *)rsp;
663
664                 u32 qid = be32_to_cpu(rc->pldbuflen_qid);
665                 c4iw_ev_handler(dev, qid);
666                 return 0;
667         } else if (unlikely(*(u8 *)rsp != *(u8 *)gl->va)) {
668                 if (recv_rx_pkt(dev, gl, rsp))
669                         return 0;
670
671                 pr_info("%s: unexpected FL contents at %p, " \
672                        "RSS %#llx, FL %#llx, len %u\n",
673                        pci_name(ctx->lldi.pdev), gl->va,
674                        (unsigned long long)be64_to_cpu(*rsp),
675                        (unsigned long long)be64_to_cpu(*(u64 *)gl->va),
676                        gl->tot_len);
677
678                 return 0;
679         } else {
680                 skb = cxgb4_pktgl_to_skb(gl, 128, 128);
681                 if (unlikely(!skb))
682                         goto nomem;
683         }
684
685         opcode = *(u8 *)rsp;
686         if (c4iw_handlers[opcode])
687                 c4iw_handlers[opcode](dev, skb);
688         else
689                 pr_info("%s no handler opcode 0x%x...\n", __func__,
690                        opcode);
691
692         return 0;
693 nomem:
694         return -1;
695 }
696
697 static int c4iw_uld_state_change(void *handle, enum cxgb4_state new_state)
698 {
699         struct uld_ctx *ctx = handle;
700
701         PDBG("%s new_state %u\n", __func__, new_state);
702         switch (new_state) {
703         case CXGB4_STATE_UP:
704                 printk(KERN_INFO MOD "%s: Up\n", pci_name(ctx->lldi.pdev));
705                 if (!ctx->dev) {
706                         int ret;
707
708                         ctx->dev = c4iw_alloc(&ctx->lldi);
709                         if (IS_ERR(ctx->dev)) {
710                                 printk(KERN_ERR MOD
711                                        "%s: initialization failed: %ld\n",
712                                        pci_name(ctx->lldi.pdev),
713                                        PTR_ERR(ctx->dev));
714                                 ctx->dev = NULL;
715                                 break;
716                         }
717                         ret = c4iw_register_device(ctx->dev);
718                         if (ret) {
719                                 printk(KERN_ERR MOD
720                                        "%s: RDMA registration failed: %d\n",
721                                        pci_name(ctx->lldi.pdev), ret);
722                                 c4iw_dealloc(ctx);
723                         }
724                 }
725                 break;
726         case CXGB4_STATE_DOWN:
727                 printk(KERN_INFO MOD "%s: Down\n",
728                        pci_name(ctx->lldi.pdev));
729                 if (ctx->dev)
730                         c4iw_remove(ctx);
731                 break;
732         case CXGB4_STATE_START_RECOVERY:
733                 printk(KERN_INFO MOD "%s: Fatal Error\n",
734                        pci_name(ctx->lldi.pdev));
735                 if (ctx->dev) {
736                         struct ib_event event;
737
738                         ctx->dev->rdev.flags |= T4_FATAL_ERROR;
739                         memset(&event, 0, sizeof event);
740                         event.event  = IB_EVENT_DEVICE_FATAL;
741                         event.device = &ctx->dev->ibdev;
742                         ib_dispatch_event(&event);
743                         c4iw_remove(ctx);
744                 }
745                 break;
746         case CXGB4_STATE_DETACH:
747                 printk(KERN_INFO MOD "%s: Detach\n",
748                        pci_name(ctx->lldi.pdev));
749                 if (ctx->dev)
750                         c4iw_remove(ctx);
751                 break;
752         }
753         return 0;
754 }
755
756 static int disable_qp_db(int id, void *p, void *data)
757 {
758         struct c4iw_qp *qp = p;
759
760         t4_disable_wq_db(&qp->wq);
761         return 0;
762 }
763
764 static void stop_queues(struct uld_ctx *ctx)
765 {
766         spin_lock_irq(&ctx->dev->lock);
767         if (ctx->dev->db_state == NORMAL) {
768                 ctx->dev->rdev.stats.db_state_transitions++;
769                 ctx->dev->db_state = FLOW_CONTROL;
770                 idr_for_each(&ctx->dev->qpidr, disable_qp_db, NULL);
771         }
772         spin_unlock_irq(&ctx->dev->lock);
773 }
774
775 static int enable_qp_db(int id, void *p, void *data)
776 {
777         struct c4iw_qp *qp = p;
778
779         t4_enable_wq_db(&qp->wq);
780         return 0;
781 }
782
783 static void resume_queues(struct uld_ctx *ctx)
784 {
785         spin_lock_irq(&ctx->dev->lock);
786         if (ctx->dev->qpcnt <= db_fc_threshold &&
787             ctx->dev->db_state == FLOW_CONTROL) {
788                 ctx->dev->db_state = NORMAL;
789                 ctx->dev->rdev.stats.db_state_transitions++;
790                 idr_for_each(&ctx->dev->qpidr, enable_qp_db, NULL);
791         }
792         spin_unlock_irq(&ctx->dev->lock);
793 }
794
795 struct qp_list {
796         unsigned idx;
797         struct c4iw_qp **qps;
798 };
799
800 static int add_and_ref_qp(int id, void *p, void *data)
801 {
802         struct qp_list *qp_listp = data;
803         struct c4iw_qp *qp = p;
804
805         c4iw_qp_add_ref(&qp->ibqp);
806         qp_listp->qps[qp_listp->idx++] = qp;
807         return 0;
808 }
809
810 static int count_qps(int id, void *p, void *data)
811 {
812         unsigned *countp = data;
813         (*countp)++;
814         return 0;
815 }
816
817 static void deref_qps(struct qp_list qp_list)
818 {
819         int idx;
820
821         for (idx = 0; idx < qp_list.idx; idx++)
822                 c4iw_qp_rem_ref(&qp_list.qps[idx]->ibqp);
823 }
824
825 static void recover_lost_dbs(struct uld_ctx *ctx, struct qp_list *qp_list)
826 {
827         int idx;
828         int ret;
829
830         for (idx = 0; idx < qp_list->idx; idx++) {
831                 struct c4iw_qp *qp = qp_list->qps[idx];
832
833                 ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
834                                           qp->wq.sq.qid,
835                                           t4_sq_host_wq_pidx(&qp->wq),
836                                           t4_sq_wq_size(&qp->wq));
837                 if (ret) {
838                         printk(KERN_ERR MOD "%s: Fatal error - "
839                                "DB overflow recovery failed - "
840                                "error syncing SQ qid %u\n",
841                                pci_name(ctx->lldi.pdev), qp->wq.sq.qid);
842                         return;
843                 }
844
845                 ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
846                                           qp->wq.rq.qid,
847                                           t4_rq_host_wq_pidx(&qp->wq),
848                                           t4_rq_wq_size(&qp->wq));
849
850                 if (ret) {
851                         printk(KERN_ERR MOD "%s: Fatal error - "
852                                "DB overflow recovery failed - "
853                                "error syncing RQ qid %u\n",
854                                pci_name(ctx->lldi.pdev), qp->wq.rq.qid);
855                         return;
856                 }
857
858                 /* Wait for the dbfifo to drain */
859                 while (cxgb4_dbfifo_count(qp->rhp->rdev.lldi.ports[0], 1) > 0) {
860                         set_current_state(TASK_UNINTERRUPTIBLE);
861                         schedule_timeout(usecs_to_jiffies(10));
862                 }
863         }
864 }
865
866 static void recover_queues(struct uld_ctx *ctx)
867 {
868         int count = 0;
869         struct qp_list qp_list;
870         int ret;
871
872         /* lock out kernel db ringers */
873         mutex_lock(&ctx->dev->db_mutex);
874
875         /* put all queues in to recovery mode */
876         spin_lock_irq(&ctx->dev->lock);
877         ctx->dev->db_state = RECOVERY;
878         ctx->dev->rdev.stats.db_state_transitions++;
879         idr_for_each(&ctx->dev->qpidr, disable_qp_db, NULL);
880         spin_unlock_irq(&ctx->dev->lock);
881
882         /* slow everybody down */
883         set_current_state(TASK_UNINTERRUPTIBLE);
884         schedule_timeout(usecs_to_jiffies(1000));
885
886         /* Wait for the dbfifo to completely drain. */
887         while (cxgb4_dbfifo_count(ctx->dev->rdev.lldi.ports[0], 1) > 0) {
888                 set_current_state(TASK_UNINTERRUPTIBLE);
889                 schedule_timeout(usecs_to_jiffies(10));
890         }
891
892         /* flush the SGE contexts */
893         ret = cxgb4_flush_eq_cache(ctx->dev->rdev.lldi.ports[0]);
894         if (ret) {
895                 printk(KERN_ERR MOD "%s: Fatal error - DB overflow recovery failed\n",
896                        pci_name(ctx->lldi.pdev));
897                 goto out;
898         }
899
900         /* Count active queues so we can build a list of queues to recover */
901         spin_lock_irq(&ctx->dev->lock);
902         idr_for_each(&ctx->dev->qpidr, count_qps, &count);
903
904         qp_list.qps = kzalloc(count * sizeof *qp_list.qps, GFP_ATOMIC);
905         if (!qp_list.qps) {
906                 printk(KERN_ERR MOD "%s: Fatal error - DB overflow recovery failed\n",
907                        pci_name(ctx->lldi.pdev));
908                 spin_unlock_irq(&ctx->dev->lock);
909                 goto out;
910         }
911         qp_list.idx = 0;
912
913         /* add and ref each qp so it doesn't get freed */
914         idr_for_each(&ctx->dev->qpidr, add_and_ref_qp, &qp_list);
915
916         spin_unlock_irq(&ctx->dev->lock);
917
918         /* now traverse the list in a safe context to recover the db state*/
919         recover_lost_dbs(ctx, &qp_list);
920
921         /* we're almost done!  deref the qps and clean up */
922         deref_qps(qp_list);
923         kfree(qp_list.qps);
924
925         /* Wait for the dbfifo to completely drain again */
926         while (cxgb4_dbfifo_count(ctx->dev->rdev.lldi.ports[0], 1) > 0) {
927                 set_current_state(TASK_UNINTERRUPTIBLE);
928                 schedule_timeout(usecs_to_jiffies(10));
929         }
930
931         /* resume the queues */
932         spin_lock_irq(&ctx->dev->lock);
933         if (ctx->dev->qpcnt > db_fc_threshold)
934                 ctx->dev->db_state = FLOW_CONTROL;
935         else {
936                 ctx->dev->db_state = NORMAL;
937                 idr_for_each(&ctx->dev->qpidr, enable_qp_db, NULL);
938         }
939         ctx->dev->rdev.stats.db_state_transitions++;
940         spin_unlock_irq(&ctx->dev->lock);
941
942 out:
943         /* start up kernel db ringers again */
944         mutex_unlock(&ctx->dev->db_mutex);
945 }
946
947 static int c4iw_uld_control(void *handle, enum cxgb4_control control, ...)
948 {
949         struct uld_ctx *ctx = handle;
950
951         switch (control) {
952         case CXGB4_CONTROL_DB_FULL:
953                 stop_queues(ctx);
954                 mutex_lock(&ctx->dev->rdev.stats.lock);
955                 ctx->dev->rdev.stats.db_full++;
956                 mutex_unlock(&ctx->dev->rdev.stats.lock);
957                 break;
958         case CXGB4_CONTROL_DB_EMPTY:
959                 resume_queues(ctx);
960                 mutex_lock(&ctx->dev->rdev.stats.lock);
961                 ctx->dev->rdev.stats.db_empty++;
962                 mutex_unlock(&ctx->dev->rdev.stats.lock);
963                 break;
964         case CXGB4_CONTROL_DB_DROP:
965                 recover_queues(ctx);
966                 mutex_lock(&ctx->dev->rdev.stats.lock);
967                 ctx->dev->rdev.stats.db_drop++;
968                 mutex_unlock(&ctx->dev->rdev.stats.lock);
969                 break;
970         default:
971                 printk(KERN_WARNING MOD "%s: unknown control cmd %u\n",
972                        pci_name(ctx->lldi.pdev), control);
973                 break;
974         }
975         return 0;
976 }
977
978 static struct cxgb4_uld_info c4iw_uld_info = {
979         .name = DRV_NAME,
980         .add = c4iw_uld_add,
981         .rx_handler = c4iw_uld_rx_handler,
982         .state_change = c4iw_uld_state_change,
983         .control = c4iw_uld_control,
984 };
985
986 static int __init c4iw_init_module(void)
987 {
988         int err;
989
990         err = c4iw_cm_init();
991         if (err)
992                 return err;
993
994         c4iw_debugfs_root = debugfs_create_dir(DRV_NAME, NULL);
995         if (!c4iw_debugfs_root)
996                 printk(KERN_WARNING MOD
997                        "could not create debugfs entry, continuing\n");
998
999         cxgb4_register_uld(CXGB4_ULD_RDMA, &c4iw_uld_info);
1000
1001         return 0;
1002 }
1003
1004 static void __exit c4iw_exit_module(void)
1005 {
1006         struct uld_ctx *ctx, *tmp;
1007
1008         mutex_lock(&dev_mutex);
1009         list_for_each_entry_safe(ctx, tmp, &uld_ctx_list, entry) {
1010                 if (ctx->dev)
1011                         c4iw_remove(ctx);
1012                 kfree(ctx);
1013         }
1014         mutex_unlock(&dev_mutex);
1015         cxgb4_unregister_uld(CXGB4_ULD_RDMA);
1016         c4iw_cm_term();
1017         debugfs_remove_recursive(c4iw_debugfs_root);
1018 }
1019
1020 module_init(c4iw_init_module);
1021 module_exit(c4iw_exit_module);