xref: /linux/drivers/infiniband/hw/cxgb4/device.c (revision c1aac62f36c1e37ee81c9e09ee9ee733eef05dcb)
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 #include <linux/math64.h>
37 
38 #include <rdma/ib_verbs.h>
39 
40 #include "iw_cxgb4.h"
41 
42 #define DRV_VERSION "0.1"
43 
44 MODULE_AUTHOR("Steve Wise");
45 MODULE_DESCRIPTION("Chelsio T4/T5 RDMA Driver");
46 MODULE_LICENSE("Dual BSD/GPL");
47 MODULE_VERSION(DRV_VERSION);
48 
49 static int allow_db_fc_on_t5;
50 module_param(allow_db_fc_on_t5, int, 0644);
51 MODULE_PARM_DESC(allow_db_fc_on_t5,
52 		 "Allow DB Flow Control on T5 (default = 0)");
53 
54 static int allow_db_coalescing_on_t5;
55 module_param(allow_db_coalescing_on_t5, int, 0644);
56 MODULE_PARM_DESC(allow_db_coalescing_on_t5,
57 		 "Allow DB Coalescing on T5 (default = 0)");
58 
59 int c4iw_wr_log = 0;
60 module_param(c4iw_wr_log, int, 0444);
61 MODULE_PARM_DESC(c4iw_wr_log, "Enables logging of work request timing data.");
62 
63 static int c4iw_wr_log_size_order = 12;
64 module_param(c4iw_wr_log_size_order, int, 0444);
65 MODULE_PARM_DESC(c4iw_wr_log_size_order,
66 		 "Number of entries (log2) in the work request timing log.");
67 
68 struct uld_ctx {
69 	struct list_head entry;
70 	struct cxgb4_lld_info lldi;
71 	struct c4iw_dev *dev;
72 };
73 
74 static LIST_HEAD(uld_ctx_list);
75 static DEFINE_MUTEX(dev_mutex);
76 
77 #define DB_FC_RESUME_SIZE 64
78 #define DB_FC_RESUME_DELAY 1
79 #define DB_FC_DRAIN_THRESH 0
80 
81 static struct dentry *c4iw_debugfs_root;
82 
83 struct c4iw_debugfs_data {
84 	struct c4iw_dev *devp;
85 	char *buf;
86 	int bufsize;
87 	int pos;
88 };
89 
90 static int count_idrs(int id, void *p, void *data)
91 {
92 	int *countp = data;
93 
94 	*countp = *countp + 1;
95 	return 0;
96 }
97 
98 static ssize_t debugfs_read(struct file *file, char __user *buf, size_t count,
99 			    loff_t *ppos)
100 {
101 	struct c4iw_debugfs_data *d = file->private_data;
102 
103 	return simple_read_from_buffer(buf, count, ppos, d->buf, d->pos);
104 }
105 
106 void c4iw_log_wr_stats(struct t4_wq *wq, struct t4_cqe *cqe)
107 {
108 	struct wr_log_entry le;
109 	int idx;
110 
111 	if (!wq->rdev->wr_log)
112 		return;
113 
114 	idx = (atomic_inc_return(&wq->rdev->wr_log_idx) - 1) &
115 		(wq->rdev->wr_log_size - 1);
116 	le.poll_sge_ts = cxgb4_read_sge_timestamp(wq->rdev->lldi.ports[0]);
117 	getnstimeofday(&le.poll_host_ts);
118 	le.valid = 1;
119 	le.cqe_sge_ts = CQE_TS(cqe);
120 	if (SQ_TYPE(cqe)) {
121 		le.qid = wq->sq.qid;
122 		le.opcode = CQE_OPCODE(cqe);
123 		le.post_host_ts = wq->sq.sw_sq[wq->sq.cidx].host_ts;
124 		le.post_sge_ts = wq->sq.sw_sq[wq->sq.cidx].sge_ts;
125 		le.wr_id = CQE_WRID_SQ_IDX(cqe);
126 	} else {
127 		le.qid = wq->rq.qid;
128 		le.opcode = FW_RI_RECEIVE;
129 		le.post_host_ts = wq->rq.sw_rq[wq->rq.cidx].host_ts;
130 		le.post_sge_ts = wq->rq.sw_rq[wq->rq.cidx].sge_ts;
131 		le.wr_id = CQE_WRID_MSN(cqe);
132 	}
133 	wq->rdev->wr_log[idx] = le;
134 }
135 
136 static int wr_log_show(struct seq_file *seq, void *v)
137 {
138 	struct c4iw_dev *dev = seq->private;
139 	struct timespec prev_ts = {0, 0};
140 	struct wr_log_entry *lep;
141 	int prev_ts_set = 0;
142 	int idx, end;
143 
144 #define ts2ns(ts) div64_u64((ts) * dev->rdev.lldi.cclk_ps, 1000)
145 
146 	idx = atomic_read(&dev->rdev.wr_log_idx) &
147 		(dev->rdev.wr_log_size - 1);
148 	end = idx - 1;
149 	if (end < 0)
150 		end = dev->rdev.wr_log_size - 1;
151 	lep = &dev->rdev.wr_log[idx];
152 	while (idx != end) {
153 		if (lep->valid) {
154 			if (!prev_ts_set) {
155 				prev_ts_set = 1;
156 				prev_ts = lep->poll_host_ts;
157 			}
158 			seq_printf(seq, "%04u: sec %lu nsec %lu qid %u opcode "
159 				   "%u %s 0x%x host_wr_delta sec %lu nsec %lu "
160 				   "post_sge_ts 0x%llx cqe_sge_ts 0x%llx "
161 				   "poll_sge_ts 0x%llx post_poll_delta_ns %llu "
162 				   "cqe_poll_delta_ns %llu\n",
163 				   idx,
164 				   timespec_sub(lep->poll_host_ts,
165 						prev_ts).tv_sec,
166 				   timespec_sub(lep->poll_host_ts,
167 						prev_ts).tv_nsec,
168 				   lep->qid, lep->opcode,
169 				   lep->opcode == FW_RI_RECEIVE ?
170 							"msn" : "wrid",
171 				   lep->wr_id,
172 				   timespec_sub(lep->poll_host_ts,
173 						lep->post_host_ts).tv_sec,
174 				   timespec_sub(lep->poll_host_ts,
175 						lep->post_host_ts).tv_nsec,
176 				   lep->post_sge_ts, lep->cqe_sge_ts,
177 				   lep->poll_sge_ts,
178 				   ts2ns(lep->poll_sge_ts - lep->post_sge_ts),
179 				   ts2ns(lep->poll_sge_ts - lep->cqe_sge_ts));
180 			prev_ts = lep->poll_host_ts;
181 		}
182 		idx++;
183 		if (idx > (dev->rdev.wr_log_size - 1))
184 			idx = 0;
185 		lep = &dev->rdev.wr_log[idx];
186 	}
187 #undef ts2ns
188 	return 0;
189 }
190 
191 static int wr_log_open(struct inode *inode, struct file *file)
192 {
193 	return single_open(file, wr_log_show, inode->i_private);
194 }
195 
196 static ssize_t wr_log_clear(struct file *file, const char __user *buf,
197 			    size_t count, loff_t *pos)
198 {
199 	struct c4iw_dev *dev = ((struct seq_file *)file->private_data)->private;
200 	int i;
201 
202 	if (dev->rdev.wr_log)
203 		for (i = 0; i < dev->rdev.wr_log_size; i++)
204 			dev->rdev.wr_log[i].valid = 0;
205 	return count;
206 }
207 
208 static const struct file_operations wr_log_debugfs_fops = {
209 	.owner   = THIS_MODULE,
210 	.open    = wr_log_open,
211 	.release = single_release,
212 	.read    = seq_read,
213 	.llseek  = seq_lseek,
214 	.write   = wr_log_clear,
215 };
216 
217 static int dump_qp(int id, void *p, void *data)
218 {
219 	struct c4iw_qp *qp = p;
220 	struct c4iw_debugfs_data *qpd = data;
221 	int space;
222 	int cc;
223 
224 	if (id != qp->wq.sq.qid)
225 		return 0;
226 
227 	space = qpd->bufsize - qpd->pos - 1;
228 	if (space == 0)
229 		return 1;
230 
231 	if (qp->ep) {
232 		if (qp->ep->com.local_addr.ss_family == AF_INET) {
233 			struct sockaddr_in *lsin = (struct sockaddr_in *)
234 				&qp->ep->com.cm_id->local_addr;
235 			struct sockaddr_in *rsin = (struct sockaddr_in *)
236 				&qp->ep->com.cm_id->remote_addr;
237 			struct sockaddr_in *mapped_lsin = (struct sockaddr_in *)
238 				&qp->ep->com.cm_id->m_local_addr;
239 			struct sockaddr_in *mapped_rsin = (struct sockaddr_in *)
240 				&qp->ep->com.cm_id->m_remote_addr;
241 
242 			cc = snprintf(qpd->buf + qpd->pos, space,
243 				      "rc qp sq id %u rq id %u state %u "
244 				      "onchip %u ep tid %u state %u "
245 				      "%pI4:%u/%u->%pI4:%u/%u\n",
246 				      qp->wq.sq.qid, qp->wq.rq.qid,
247 				      (int)qp->attr.state,
248 				      qp->wq.sq.flags & T4_SQ_ONCHIP,
249 				      qp->ep->hwtid, (int)qp->ep->com.state,
250 				      &lsin->sin_addr, ntohs(lsin->sin_port),
251 				      ntohs(mapped_lsin->sin_port),
252 				      &rsin->sin_addr, ntohs(rsin->sin_port),
253 				      ntohs(mapped_rsin->sin_port));
254 		} else {
255 			struct sockaddr_in6 *lsin6 = (struct sockaddr_in6 *)
256 				&qp->ep->com.cm_id->local_addr;
257 			struct sockaddr_in6 *rsin6 = (struct sockaddr_in6 *)
258 				&qp->ep->com.cm_id->remote_addr;
259 			struct sockaddr_in6 *mapped_lsin6 =
260 				(struct sockaddr_in6 *)
261 				&qp->ep->com.cm_id->m_local_addr;
262 			struct sockaddr_in6 *mapped_rsin6 =
263 				(struct sockaddr_in6 *)
264 				&qp->ep->com.cm_id->m_remote_addr;
265 
266 			cc = snprintf(qpd->buf + qpd->pos, space,
267 				      "rc qp sq id %u rq id %u state %u "
268 				      "onchip %u ep tid %u state %u "
269 				      "%pI6:%u/%u->%pI6:%u/%u\n",
270 				      qp->wq.sq.qid, qp->wq.rq.qid,
271 				      (int)qp->attr.state,
272 				      qp->wq.sq.flags & T4_SQ_ONCHIP,
273 				      qp->ep->hwtid, (int)qp->ep->com.state,
274 				      &lsin6->sin6_addr,
275 				      ntohs(lsin6->sin6_port),
276 				      ntohs(mapped_lsin6->sin6_port),
277 				      &rsin6->sin6_addr,
278 				      ntohs(rsin6->sin6_port),
279 				      ntohs(mapped_rsin6->sin6_port));
280 		}
281 	} else
282 		cc = snprintf(qpd->buf + qpd->pos, space,
283 			     "qp sq id %u rq id %u state %u onchip %u\n",
284 			      qp->wq.sq.qid, qp->wq.rq.qid,
285 			      (int)qp->attr.state,
286 			      qp->wq.sq.flags & T4_SQ_ONCHIP);
287 	if (cc < space)
288 		qpd->pos += cc;
289 	return 0;
290 }
291 
292 static int qp_release(struct inode *inode, struct file *file)
293 {
294 	struct c4iw_debugfs_data *qpd = file->private_data;
295 	if (!qpd) {
296 		printk(KERN_INFO "%s null qpd?\n", __func__);
297 		return 0;
298 	}
299 	vfree(qpd->buf);
300 	kfree(qpd);
301 	return 0;
302 }
303 
304 static int qp_open(struct inode *inode, struct file *file)
305 {
306 	struct c4iw_debugfs_data *qpd;
307 	int count = 1;
308 
309 	qpd = kmalloc(sizeof *qpd, GFP_KERNEL);
310 	if (!qpd)
311 		return -ENOMEM;
312 
313 	qpd->devp = inode->i_private;
314 	qpd->pos = 0;
315 
316 	spin_lock_irq(&qpd->devp->lock);
317 	idr_for_each(&qpd->devp->qpidr, count_idrs, &count);
318 	spin_unlock_irq(&qpd->devp->lock);
319 
320 	qpd->bufsize = count * 180;
321 	qpd->buf = vmalloc(qpd->bufsize);
322 	if (!qpd->buf) {
323 		kfree(qpd);
324 		return -ENOMEM;
325 	}
326 
327 	spin_lock_irq(&qpd->devp->lock);
328 	idr_for_each(&qpd->devp->qpidr, dump_qp, qpd);
329 	spin_unlock_irq(&qpd->devp->lock);
330 
331 	qpd->buf[qpd->pos++] = 0;
332 	file->private_data = qpd;
333 	return 0;
334 }
335 
336 static const struct file_operations qp_debugfs_fops = {
337 	.owner   = THIS_MODULE,
338 	.open    = qp_open,
339 	.release = qp_release,
340 	.read    = debugfs_read,
341 	.llseek  = default_llseek,
342 };
343 
344 static int dump_stag(int id, void *p, void *data)
345 {
346 	struct c4iw_debugfs_data *stagd = data;
347 	int space;
348 	int cc;
349 	struct fw_ri_tpte tpte;
350 	int ret;
351 
352 	space = stagd->bufsize - stagd->pos - 1;
353 	if (space == 0)
354 		return 1;
355 
356 	ret = cxgb4_read_tpte(stagd->devp->rdev.lldi.ports[0], (u32)id<<8,
357 			      (__be32 *)&tpte);
358 	if (ret) {
359 		dev_err(&stagd->devp->rdev.lldi.pdev->dev,
360 			"%s cxgb4_read_tpte err %d\n", __func__, ret);
361 		return ret;
362 	}
363 	cc = snprintf(stagd->buf + stagd->pos, space,
364 		      "stag: idx 0x%x valid %d key 0x%x state %d pdid %d "
365 		      "perm 0x%x ps %d len 0x%llx va 0x%llx\n",
366 		      (u32)id<<8,
367 		      FW_RI_TPTE_VALID_G(ntohl(tpte.valid_to_pdid)),
368 		      FW_RI_TPTE_STAGKEY_G(ntohl(tpte.valid_to_pdid)),
369 		      FW_RI_TPTE_STAGSTATE_G(ntohl(tpte.valid_to_pdid)),
370 		      FW_RI_TPTE_PDID_G(ntohl(tpte.valid_to_pdid)),
371 		      FW_RI_TPTE_PERM_G(ntohl(tpte.locread_to_qpid)),
372 		      FW_RI_TPTE_PS_G(ntohl(tpte.locread_to_qpid)),
373 		      ((u64)ntohl(tpte.len_hi) << 32) | ntohl(tpte.len_lo),
374 		      ((u64)ntohl(tpte.va_hi) << 32) | ntohl(tpte.va_lo_fbo));
375 	if (cc < space)
376 		stagd->pos += cc;
377 	return 0;
378 }
379 
380 static int stag_release(struct inode *inode, struct file *file)
381 {
382 	struct c4iw_debugfs_data *stagd = file->private_data;
383 	if (!stagd) {
384 		printk(KERN_INFO "%s null stagd?\n", __func__);
385 		return 0;
386 	}
387 	vfree(stagd->buf);
388 	kfree(stagd);
389 	return 0;
390 }
391 
392 static int stag_open(struct inode *inode, struct file *file)
393 {
394 	struct c4iw_debugfs_data *stagd;
395 	int ret = 0;
396 	int count = 1;
397 
398 	stagd = kmalloc(sizeof *stagd, GFP_KERNEL);
399 	if (!stagd) {
400 		ret = -ENOMEM;
401 		goto out;
402 	}
403 	stagd->devp = inode->i_private;
404 	stagd->pos = 0;
405 
406 	spin_lock_irq(&stagd->devp->lock);
407 	idr_for_each(&stagd->devp->mmidr, count_idrs, &count);
408 	spin_unlock_irq(&stagd->devp->lock);
409 
410 	stagd->bufsize = count * 256;
411 	stagd->buf = vmalloc(stagd->bufsize);
412 	if (!stagd->buf) {
413 		ret = -ENOMEM;
414 		goto err1;
415 	}
416 
417 	spin_lock_irq(&stagd->devp->lock);
418 	idr_for_each(&stagd->devp->mmidr, dump_stag, stagd);
419 	spin_unlock_irq(&stagd->devp->lock);
420 
421 	stagd->buf[stagd->pos++] = 0;
422 	file->private_data = stagd;
423 	goto out;
424 err1:
425 	kfree(stagd);
426 out:
427 	return ret;
428 }
429 
430 static const struct file_operations stag_debugfs_fops = {
431 	.owner   = THIS_MODULE,
432 	.open    = stag_open,
433 	.release = stag_release,
434 	.read    = debugfs_read,
435 	.llseek  = default_llseek,
436 };
437 
438 static char *db_state_str[] = {"NORMAL", "FLOW_CONTROL", "RECOVERY", "STOPPED"};
439 
440 static int stats_show(struct seq_file *seq, void *v)
441 {
442 	struct c4iw_dev *dev = seq->private;
443 
444 	seq_printf(seq, "   Object: %10s %10s %10s %10s\n", "Total", "Current",
445 		   "Max", "Fail");
446 	seq_printf(seq, "     PDID: %10llu %10llu %10llu %10llu\n",
447 			dev->rdev.stats.pd.total, dev->rdev.stats.pd.cur,
448 			dev->rdev.stats.pd.max, dev->rdev.stats.pd.fail);
449 	seq_printf(seq, "      QID: %10llu %10llu %10llu %10llu\n",
450 			dev->rdev.stats.qid.total, dev->rdev.stats.qid.cur,
451 			dev->rdev.stats.qid.max, dev->rdev.stats.qid.fail);
452 	seq_printf(seq, "   TPTMEM: %10llu %10llu %10llu %10llu\n",
453 			dev->rdev.stats.stag.total, dev->rdev.stats.stag.cur,
454 			dev->rdev.stats.stag.max, dev->rdev.stats.stag.fail);
455 	seq_printf(seq, "   PBLMEM: %10llu %10llu %10llu %10llu\n",
456 			dev->rdev.stats.pbl.total, dev->rdev.stats.pbl.cur,
457 			dev->rdev.stats.pbl.max, dev->rdev.stats.pbl.fail);
458 	seq_printf(seq, "   RQTMEM: %10llu %10llu %10llu %10llu\n",
459 			dev->rdev.stats.rqt.total, dev->rdev.stats.rqt.cur,
460 			dev->rdev.stats.rqt.max, dev->rdev.stats.rqt.fail);
461 	seq_printf(seq, "  OCQPMEM: %10llu %10llu %10llu %10llu\n",
462 			dev->rdev.stats.ocqp.total, dev->rdev.stats.ocqp.cur,
463 			dev->rdev.stats.ocqp.max, dev->rdev.stats.ocqp.fail);
464 	seq_printf(seq, "  DB FULL: %10llu\n", dev->rdev.stats.db_full);
465 	seq_printf(seq, " DB EMPTY: %10llu\n", dev->rdev.stats.db_empty);
466 	seq_printf(seq, "  DB DROP: %10llu\n", dev->rdev.stats.db_drop);
467 	seq_printf(seq, " DB State: %s Transitions %llu FC Interruptions %llu\n",
468 		   db_state_str[dev->db_state],
469 		   dev->rdev.stats.db_state_transitions,
470 		   dev->rdev.stats.db_fc_interruptions);
471 	seq_printf(seq, "TCAM_FULL: %10llu\n", dev->rdev.stats.tcam_full);
472 	seq_printf(seq, "ACT_OFLD_CONN_FAILS: %10llu\n",
473 		   dev->rdev.stats.act_ofld_conn_fails);
474 	seq_printf(seq, "PAS_OFLD_CONN_FAILS: %10llu\n",
475 		   dev->rdev.stats.pas_ofld_conn_fails);
476 	seq_printf(seq, "NEG_ADV_RCVD: %10llu\n", dev->rdev.stats.neg_adv);
477 	seq_printf(seq, "AVAILABLE IRD: %10u\n", dev->avail_ird);
478 	return 0;
479 }
480 
481 static int stats_open(struct inode *inode, struct file *file)
482 {
483 	return single_open(file, stats_show, inode->i_private);
484 }
485 
486 static ssize_t stats_clear(struct file *file, const char __user *buf,
487 		size_t count, loff_t *pos)
488 {
489 	struct c4iw_dev *dev = ((struct seq_file *)file->private_data)->private;
490 
491 	mutex_lock(&dev->rdev.stats.lock);
492 	dev->rdev.stats.pd.max = 0;
493 	dev->rdev.stats.pd.fail = 0;
494 	dev->rdev.stats.qid.max = 0;
495 	dev->rdev.stats.qid.fail = 0;
496 	dev->rdev.stats.stag.max = 0;
497 	dev->rdev.stats.stag.fail = 0;
498 	dev->rdev.stats.pbl.max = 0;
499 	dev->rdev.stats.pbl.fail = 0;
500 	dev->rdev.stats.rqt.max = 0;
501 	dev->rdev.stats.rqt.fail = 0;
502 	dev->rdev.stats.ocqp.max = 0;
503 	dev->rdev.stats.ocqp.fail = 0;
504 	dev->rdev.stats.db_full = 0;
505 	dev->rdev.stats.db_empty = 0;
506 	dev->rdev.stats.db_drop = 0;
507 	dev->rdev.stats.db_state_transitions = 0;
508 	dev->rdev.stats.tcam_full = 0;
509 	dev->rdev.stats.act_ofld_conn_fails = 0;
510 	dev->rdev.stats.pas_ofld_conn_fails = 0;
511 	mutex_unlock(&dev->rdev.stats.lock);
512 	return count;
513 }
514 
515 static const struct file_operations stats_debugfs_fops = {
516 	.owner   = THIS_MODULE,
517 	.open    = stats_open,
518 	.release = single_release,
519 	.read    = seq_read,
520 	.llseek  = seq_lseek,
521 	.write   = stats_clear,
522 };
523 
524 static int dump_ep(int id, void *p, void *data)
525 {
526 	struct c4iw_ep *ep = p;
527 	struct c4iw_debugfs_data *epd = data;
528 	int space;
529 	int cc;
530 
531 	space = epd->bufsize - epd->pos - 1;
532 	if (space == 0)
533 		return 1;
534 
535 	if (ep->com.local_addr.ss_family == AF_INET) {
536 		struct sockaddr_in *lsin = (struct sockaddr_in *)
537 			&ep->com.cm_id->local_addr;
538 		struct sockaddr_in *rsin = (struct sockaddr_in *)
539 			&ep->com.cm_id->remote_addr;
540 		struct sockaddr_in *mapped_lsin = (struct sockaddr_in *)
541 			&ep->com.cm_id->m_local_addr;
542 		struct sockaddr_in *mapped_rsin = (struct sockaddr_in *)
543 			&ep->com.cm_id->m_remote_addr;
544 
545 		cc = snprintf(epd->buf + epd->pos, space,
546 			      "ep %p cm_id %p qp %p state %d flags 0x%lx "
547 			      "history 0x%lx hwtid %d atid %d "
548 			      "conn_na %u abort_na %u "
549 			      "%pI4:%d/%d <-> %pI4:%d/%d\n",
550 			      ep, ep->com.cm_id, ep->com.qp,
551 			      (int)ep->com.state, ep->com.flags,
552 			      ep->com.history, ep->hwtid, ep->atid,
553 			      ep->stats.connect_neg_adv,
554 			      ep->stats.abort_neg_adv,
555 			      &lsin->sin_addr, ntohs(lsin->sin_port),
556 			      ntohs(mapped_lsin->sin_port),
557 			      &rsin->sin_addr, ntohs(rsin->sin_port),
558 			      ntohs(mapped_rsin->sin_port));
559 	} else {
560 		struct sockaddr_in6 *lsin6 = (struct sockaddr_in6 *)
561 			&ep->com.cm_id->local_addr;
562 		struct sockaddr_in6 *rsin6 = (struct sockaddr_in6 *)
563 			&ep->com.cm_id->remote_addr;
564 		struct sockaddr_in6 *mapped_lsin6 = (struct sockaddr_in6 *)
565 			&ep->com.cm_id->m_local_addr;
566 		struct sockaddr_in6 *mapped_rsin6 = (struct sockaddr_in6 *)
567 			&ep->com.cm_id->m_remote_addr;
568 
569 		cc = snprintf(epd->buf + epd->pos, space,
570 			      "ep %p cm_id %p qp %p state %d flags 0x%lx "
571 			      "history 0x%lx hwtid %d atid %d "
572 			      "conn_na %u abort_na %u "
573 			      "%pI6:%d/%d <-> %pI6:%d/%d\n",
574 			      ep, ep->com.cm_id, ep->com.qp,
575 			      (int)ep->com.state, ep->com.flags,
576 			      ep->com.history, ep->hwtid, ep->atid,
577 			      ep->stats.connect_neg_adv,
578 			      ep->stats.abort_neg_adv,
579 			      &lsin6->sin6_addr, ntohs(lsin6->sin6_port),
580 			      ntohs(mapped_lsin6->sin6_port),
581 			      &rsin6->sin6_addr, ntohs(rsin6->sin6_port),
582 			      ntohs(mapped_rsin6->sin6_port));
583 	}
584 	if (cc < space)
585 		epd->pos += cc;
586 	return 0;
587 }
588 
589 static int dump_listen_ep(int id, void *p, void *data)
590 {
591 	struct c4iw_listen_ep *ep = p;
592 	struct c4iw_debugfs_data *epd = data;
593 	int space;
594 	int cc;
595 
596 	space = epd->bufsize - epd->pos - 1;
597 	if (space == 0)
598 		return 1;
599 
600 	if (ep->com.local_addr.ss_family == AF_INET) {
601 		struct sockaddr_in *lsin = (struct sockaddr_in *)
602 			&ep->com.cm_id->local_addr;
603 		struct sockaddr_in *mapped_lsin = (struct sockaddr_in *)
604 			&ep->com.cm_id->m_local_addr;
605 
606 		cc = snprintf(epd->buf + epd->pos, space,
607 			      "ep %p cm_id %p state %d flags 0x%lx stid %d "
608 			      "backlog %d %pI4:%d/%d\n",
609 			      ep, ep->com.cm_id, (int)ep->com.state,
610 			      ep->com.flags, ep->stid, ep->backlog,
611 			      &lsin->sin_addr, ntohs(lsin->sin_port),
612 			      ntohs(mapped_lsin->sin_port));
613 	} else {
614 		struct sockaddr_in6 *lsin6 = (struct sockaddr_in6 *)
615 			&ep->com.cm_id->local_addr;
616 		struct sockaddr_in6 *mapped_lsin6 = (struct sockaddr_in6 *)
617 			&ep->com.cm_id->m_local_addr;
618 
619 		cc = snprintf(epd->buf + epd->pos, space,
620 			      "ep %p cm_id %p state %d flags 0x%lx stid %d "
621 			      "backlog %d %pI6:%d/%d\n",
622 			      ep, ep->com.cm_id, (int)ep->com.state,
623 			      ep->com.flags, ep->stid, ep->backlog,
624 			      &lsin6->sin6_addr, ntohs(lsin6->sin6_port),
625 			      ntohs(mapped_lsin6->sin6_port));
626 	}
627 	if (cc < space)
628 		epd->pos += cc;
629 	return 0;
630 }
631 
632 static int ep_release(struct inode *inode, struct file *file)
633 {
634 	struct c4iw_debugfs_data *epd = file->private_data;
635 	if (!epd) {
636 		pr_info("%s null qpd?\n", __func__);
637 		return 0;
638 	}
639 	vfree(epd->buf);
640 	kfree(epd);
641 	return 0;
642 }
643 
644 static int ep_open(struct inode *inode, struct file *file)
645 {
646 	struct c4iw_debugfs_data *epd;
647 	int ret = 0;
648 	int count = 1;
649 
650 	epd = kmalloc(sizeof(*epd), GFP_KERNEL);
651 	if (!epd) {
652 		ret = -ENOMEM;
653 		goto out;
654 	}
655 	epd->devp = inode->i_private;
656 	epd->pos = 0;
657 
658 	spin_lock_irq(&epd->devp->lock);
659 	idr_for_each(&epd->devp->hwtid_idr, count_idrs, &count);
660 	idr_for_each(&epd->devp->atid_idr, count_idrs, &count);
661 	idr_for_each(&epd->devp->stid_idr, count_idrs, &count);
662 	spin_unlock_irq(&epd->devp->lock);
663 
664 	epd->bufsize = count * 240;
665 	epd->buf = vmalloc(epd->bufsize);
666 	if (!epd->buf) {
667 		ret = -ENOMEM;
668 		goto err1;
669 	}
670 
671 	spin_lock_irq(&epd->devp->lock);
672 	idr_for_each(&epd->devp->hwtid_idr, dump_ep, epd);
673 	idr_for_each(&epd->devp->atid_idr, dump_ep, epd);
674 	idr_for_each(&epd->devp->stid_idr, dump_listen_ep, epd);
675 	spin_unlock_irq(&epd->devp->lock);
676 
677 	file->private_data = epd;
678 	goto out;
679 err1:
680 	kfree(epd);
681 out:
682 	return ret;
683 }
684 
685 static const struct file_operations ep_debugfs_fops = {
686 	.owner   = THIS_MODULE,
687 	.open    = ep_open,
688 	.release = ep_release,
689 	.read    = debugfs_read,
690 };
691 
692 static int setup_debugfs(struct c4iw_dev *devp)
693 {
694 	if (!devp->debugfs_root)
695 		return -1;
696 
697 	debugfs_create_file_size("qps", S_IWUSR, devp->debugfs_root,
698 				 (void *)devp, &qp_debugfs_fops, 4096);
699 
700 	debugfs_create_file_size("stags", S_IWUSR, devp->debugfs_root,
701 				 (void *)devp, &stag_debugfs_fops, 4096);
702 
703 	debugfs_create_file_size("stats", S_IWUSR, devp->debugfs_root,
704 				 (void *)devp, &stats_debugfs_fops, 4096);
705 
706 	debugfs_create_file_size("eps", S_IWUSR, devp->debugfs_root,
707 				 (void *)devp, &ep_debugfs_fops, 4096);
708 
709 	if (c4iw_wr_log)
710 		debugfs_create_file_size("wr_log", S_IWUSR, devp->debugfs_root,
711 					 (void *)devp, &wr_log_debugfs_fops, 4096);
712 	return 0;
713 }
714 
715 void c4iw_release_dev_ucontext(struct c4iw_rdev *rdev,
716 			       struct c4iw_dev_ucontext *uctx)
717 {
718 	struct list_head *pos, *nxt;
719 	struct c4iw_qid_list *entry;
720 
721 	mutex_lock(&uctx->lock);
722 	list_for_each_safe(pos, nxt, &uctx->qpids) {
723 		entry = list_entry(pos, struct c4iw_qid_list, entry);
724 		list_del_init(&entry->entry);
725 		if (!(entry->qid & rdev->qpmask)) {
726 			c4iw_put_resource(&rdev->resource.qid_table,
727 					  entry->qid);
728 			mutex_lock(&rdev->stats.lock);
729 			rdev->stats.qid.cur -= rdev->qpmask + 1;
730 			mutex_unlock(&rdev->stats.lock);
731 		}
732 		kfree(entry);
733 	}
734 
735 	list_for_each_safe(pos, nxt, &uctx->qpids) {
736 		entry = list_entry(pos, struct c4iw_qid_list, entry);
737 		list_del_init(&entry->entry);
738 		kfree(entry);
739 	}
740 	mutex_unlock(&uctx->lock);
741 }
742 
743 void c4iw_init_dev_ucontext(struct c4iw_rdev *rdev,
744 			    struct c4iw_dev_ucontext *uctx)
745 {
746 	INIT_LIST_HEAD(&uctx->qpids);
747 	INIT_LIST_HEAD(&uctx->cqids);
748 	mutex_init(&uctx->lock);
749 }
750 
751 /* Caller takes care of locking if needed */
752 static int c4iw_rdev_open(struct c4iw_rdev *rdev)
753 {
754 	int err;
755 
756 	c4iw_init_dev_ucontext(rdev, &rdev->uctx);
757 
758 	/*
759 	 * This implementation assumes udb_density == ucq_density!  Eventually
760 	 * we might need to support this but for now fail the open. Also the
761 	 * cqid and qpid range must match for now.
762 	 */
763 	if (rdev->lldi.udb_density != rdev->lldi.ucq_density) {
764 		pr_err(MOD "%s: unsupported udb/ucq densities %u/%u\n",
765 		       pci_name(rdev->lldi.pdev), rdev->lldi.udb_density,
766 		       rdev->lldi.ucq_density);
767 		return -EINVAL;
768 	}
769 	if (rdev->lldi.vr->qp.start != rdev->lldi.vr->cq.start ||
770 	    rdev->lldi.vr->qp.size != rdev->lldi.vr->cq.size) {
771 		pr_err(MOD "%s: unsupported qp and cq id ranges "
772 		       "qp start %u size %u cq start %u size %u\n",
773 		       pci_name(rdev->lldi.pdev), rdev->lldi.vr->qp.start,
774 		       rdev->lldi.vr->qp.size, rdev->lldi.vr->cq.size,
775 		       rdev->lldi.vr->cq.size);
776 		return -EINVAL;
777 	}
778 
779 	rdev->qpmask = rdev->lldi.udb_density - 1;
780 	rdev->cqmask = rdev->lldi.ucq_density - 1;
781 	PDBG("%s dev %s stag start 0x%0x size 0x%0x num stags %d "
782 	     "pbl start 0x%0x size 0x%0x rq start 0x%0x size 0x%0x "
783 	     "qp qid start %u size %u cq qid start %u size %u\n",
784 	     __func__, pci_name(rdev->lldi.pdev), rdev->lldi.vr->stag.start,
785 	     rdev->lldi.vr->stag.size, c4iw_num_stags(rdev),
786 	     rdev->lldi.vr->pbl.start,
787 	     rdev->lldi.vr->pbl.size, rdev->lldi.vr->rq.start,
788 	     rdev->lldi.vr->rq.size,
789 	     rdev->lldi.vr->qp.start,
790 	     rdev->lldi.vr->qp.size,
791 	     rdev->lldi.vr->cq.start,
792 	     rdev->lldi.vr->cq.size);
793 	PDBG("udb %pR db_reg %p gts_reg %p "
794 	     "qpmask 0x%x cqmask 0x%x\n",
795 		&rdev->lldi.pdev->resource[2],
796 	     rdev->lldi.db_reg, rdev->lldi.gts_reg,
797 	     rdev->qpmask, rdev->cqmask);
798 
799 	if (c4iw_num_stags(rdev) == 0)
800 		return -EINVAL;
801 
802 	rdev->stats.pd.total = T4_MAX_NUM_PD;
803 	rdev->stats.stag.total = rdev->lldi.vr->stag.size;
804 	rdev->stats.pbl.total = rdev->lldi.vr->pbl.size;
805 	rdev->stats.rqt.total = rdev->lldi.vr->rq.size;
806 	rdev->stats.ocqp.total = rdev->lldi.vr->ocq.size;
807 	rdev->stats.qid.total = rdev->lldi.vr->qp.size;
808 
809 	err = c4iw_init_resource(rdev, c4iw_num_stags(rdev), T4_MAX_NUM_PD);
810 	if (err) {
811 		printk(KERN_ERR MOD "error %d initializing resources\n", err);
812 		return err;
813 	}
814 	err = c4iw_pblpool_create(rdev);
815 	if (err) {
816 		printk(KERN_ERR MOD "error %d initializing pbl pool\n", err);
817 		goto destroy_resource;
818 	}
819 	err = c4iw_rqtpool_create(rdev);
820 	if (err) {
821 		printk(KERN_ERR MOD "error %d initializing rqt pool\n", err);
822 		goto destroy_pblpool;
823 	}
824 	err = c4iw_ocqp_pool_create(rdev);
825 	if (err) {
826 		printk(KERN_ERR MOD "error %d initializing ocqp pool\n", err);
827 		goto destroy_rqtpool;
828 	}
829 	rdev->status_page = (struct t4_dev_status_page *)
830 			    __get_free_page(GFP_KERNEL);
831 	if (!rdev->status_page) {
832 		err = -ENOMEM;
833 		goto destroy_ocqp_pool;
834 	}
835 	rdev->status_page->qp_start = rdev->lldi.vr->qp.start;
836 	rdev->status_page->qp_size = rdev->lldi.vr->qp.size;
837 	rdev->status_page->cq_start = rdev->lldi.vr->cq.start;
838 	rdev->status_page->cq_size = rdev->lldi.vr->cq.size;
839 
840 	if (c4iw_wr_log) {
841 		rdev->wr_log = kzalloc((1 << c4iw_wr_log_size_order) *
842 				       sizeof(*rdev->wr_log), GFP_KERNEL);
843 		if (rdev->wr_log) {
844 			rdev->wr_log_size = 1 << c4iw_wr_log_size_order;
845 			atomic_set(&rdev->wr_log_idx, 0);
846 		}
847 	}
848 
849 	rdev->free_workq = create_singlethread_workqueue("iw_cxgb4_free");
850 	if (!rdev->free_workq) {
851 		err = -ENOMEM;
852 		goto err_free_status_page;
853 	}
854 
855 	rdev->status_page->db_off = 0;
856 
857 	return 0;
858 err_free_status_page:
859 	free_page((unsigned long)rdev->status_page);
860 destroy_ocqp_pool:
861 	c4iw_ocqp_pool_destroy(rdev);
862 destroy_rqtpool:
863 	c4iw_rqtpool_destroy(rdev);
864 destroy_pblpool:
865 	c4iw_pblpool_destroy(rdev);
866 destroy_resource:
867 	c4iw_destroy_resource(&rdev->resource);
868 	return err;
869 }
870 
871 static void c4iw_rdev_close(struct c4iw_rdev *rdev)
872 {
873 	destroy_workqueue(rdev->free_workq);
874 	kfree(rdev->wr_log);
875 	free_page((unsigned long)rdev->status_page);
876 	c4iw_pblpool_destroy(rdev);
877 	c4iw_rqtpool_destroy(rdev);
878 	c4iw_destroy_resource(&rdev->resource);
879 }
880 
881 static void c4iw_dealloc(struct uld_ctx *ctx)
882 {
883 	c4iw_rdev_close(&ctx->dev->rdev);
884 	WARN_ON_ONCE(!idr_is_empty(&ctx->dev->cqidr));
885 	idr_destroy(&ctx->dev->cqidr);
886 	WARN_ON_ONCE(!idr_is_empty(&ctx->dev->qpidr));
887 	idr_destroy(&ctx->dev->qpidr);
888 	WARN_ON_ONCE(!idr_is_empty(&ctx->dev->mmidr));
889 	idr_destroy(&ctx->dev->mmidr);
890 	wait_event(ctx->dev->wait, idr_is_empty(&ctx->dev->hwtid_idr));
891 	idr_destroy(&ctx->dev->hwtid_idr);
892 	idr_destroy(&ctx->dev->stid_idr);
893 	idr_destroy(&ctx->dev->atid_idr);
894 	if (ctx->dev->rdev.bar2_kva)
895 		iounmap(ctx->dev->rdev.bar2_kva);
896 	if (ctx->dev->rdev.oc_mw_kva)
897 		iounmap(ctx->dev->rdev.oc_mw_kva);
898 	ib_dealloc_device(&ctx->dev->ibdev);
899 	ctx->dev = NULL;
900 }
901 
902 static void c4iw_remove(struct uld_ctx *ctx)
903 {
904 	PDBG("%s c4iw_dev %p\n", __func__,  ctx->dev);
905 	c4iw_unregister_device(ctx->dev);
906 	c4iw_dealloc(ctx);
907 }
908 
909 static int rdma_supported(const struct cxgb4_lld_info *infop)
910 {
911 	return infop->vr->stag.size > 0 && infop->vr->pbl.size > 0 &&
912 	       infop->vr->rq.size > 0 && infop->vr->qp.size > 0 &&
913 	       infop->vr->cq.size > 0;
914 }
915 
916 static struct c4iw_dev *c4iw_alloc(const struct cxgb4_lld_info *infop)
917 {
918 	struct c4iw_dev *devp;
919 	int ret;
920 
921 	if (!rdma_supported(infop)) {
922 		printk(KERN_INFO MOD "%s: RDMA not supported on this device.\n",
923 		       pci_name(infop->pdev));
924 		return ERR_PTR(-ENOSYS);
925 	}
926 	if (!ocqp_supported(infop))
927 		pr_info("%s: On-Chip Queues not supported on this device.\n",
928 			pci_name(infop->pdev));
929 
930 	devp = (struct c4iw_dev *)ib_alloc_device(sizeof(*devp));
931 	if (!devp) {
932 		printk(KERN_ERR MOD "Cannot allocate ib device\n");
933 		return ERR_PTR(-ENOMEM);
934 	}
935 	devp->rdev.lldi = *infop;
936 
937 	/* init various hw-queue params based on lld info */
938 	PDBG("%s: Ing. padding boundary is %d, egrsstatuspagesize = %d\n",
939 	     __func__, devp->rdev.lldi.sge_ingpadboundary,
940 	     devp->rdev.lldi.sge_egrstatuspagesize);
941 
942 	devp->rdev.hw_queue.t4_eq_status_entries =
943 		devp->rdev.lldi.sge_ingpadboundary > 64 ? 2 : 1;
944 	devp->rdev.hw_queue.t4_max_eq_size = 65520;
945 	devp->rdev.hw_queue.t4_max_iq_size = 65520;
946 	devp->rdev.hw_queue.t4_max_rq_size = 8192 -
947 		devp->rdev.hw_queue.t4_eq_status_entries - 1;
948 	devp->rdev.hw_queue.t4_max_sq_size =
949 		devp->rdev.hw_queue.t4_max_eq_size -
950 		devp->rdev.hw_queue.t4_eq_status_entries - 1;
951 	devp->rdev.hw_queue.t4_max_qp_depth =
952 		devp->rdev.hw_queue.t4_max_rq_size;
953 	devp->rdev.hw_queue.t4_max_cq_depth =
954 		devp->rdev.hw_queue.t4_max_iq_size - 2;
955 	devp->rdev.hw_queue.t4_stat_len =
956 		devp->rdev.lldi.sge_egrstatuspagesize;
957 
958 	/*
959 	 * For T5/T6 devices, we map all of BAR2 with WC.
960 	 * For T4 devices with onchip qp mem, we map only that part
961 	 * of BAR2 with WC.
962 	 */
963 	devp->rdev.bar2_pa = pci_resource_start(devp->rdev.lldi.pdev, 2);
964 	if (!is_t4(devp->rdev.lldi.adapter_type)) {
965 		devp->rdev.bar2_kva = ioremap_wc(devp->rdev.bar2_pa,
966 			pci_resource_len(devp->rdev.lldi.pdev, 2));
967 		if (!devp->rdev.bar2_kva) {
968 			pr_err(MOD "Unable to ioremap BAR2\n");
969 			ib_dealloc_device(&devp->ibdev);
970 			return ERR_PTR(-EINVAL);
971 		}
972 	} else if (ocqp_supported(infop)) {
973 		devp->rdev.oc_mw_pa =
974 			pci_resource_start(devp->rdev.lldi.pdev, 2) +
975 			pci_resource_len(devp->rdev.lldi.pdev, 2) -
976 			roundup_pow_of_two(devp->rdev.lldi.vr->ocq.size);
977 		devp->rdev.oc_mw_kva = ioremap_wc(devp->rdev.oc_mw_pa,
978 			devp->rdev.lldi.vr->ocq.size);
979 		if (!devp->rdev.oc_mw_kva) {
980 			pr_err(MOD "Unable to ioremap onchip mem\n");
981 			ib_dealloc_device(&devp->ibdev);
982 			return ERR_PTR(-EINVAL);
983 		}
984 	}
985 
986 	PDBG(KERN_INFO MOD "ocq memory: "
987 	       "hw_start 0x%x size %u mw_pa 0x%lx mw_kva %p\n",
988 	       devp->rdev.lldi.vr->ocq.start, devp->rdev.lldi.vr->ocq.size,
989 	       devp->rdev.oc_mw_pa, devp->rdev.oc_mw_kva);
990 
991 	ret = c4iw_rdev_open(&devp->rdev);
992 	if (ret) {
993 		printk(KERN_ERR MOD "Unable to open CXIO rdev err %d\n", ret);
994 		ib_dealloc_device(&devp->ibdev);
995 		return ERR_PTR(ret);
996 	}
997 
998 	idr_init(&devp->cqidr);
999 	idr_init(&devp->qpidr);
1000 	idr_init(&devp->mmidr);
1001 	idr_init(&devp->hwtid_idr);
1002 	idr_init(&devp->stid_idr);
1003 	idr_init(&devp->atid_idr);
1004 	spin_lock_init(&devp->lock);
1005 	mutex_init(&devp->rdev.stats.lock);
1006 	mutex_init(&devp->db_mutex);
1007 	INIT_LIST_HEAD(&devp->db_fc_list);
1008 	init_waitqueue_head(&devp->wait);
1009 	devp->avail_ird = devp->rdev.lldi.max_ird_adapter;
1010 
1011 	if (c4iw_debugfs_root) {
1012 		devp->debugfs_root = debugfs_create_dir(
1013 					pci_name(devp->rdev.lldi.pdev),
1014 					c4iw_debugfs_root);
1015 		setup_debugfs(devp);
1016 	}
1017 
1018 
1019 	return devp;
1020 }
1021 
1022 static void *c4iw_uld_add(const struct cxgb4_lld_info *infop)
1023 {
1024 	struct uld_ctx *ctx;
1025 	static int vers_printed;
1026 	int i;
1027 
1028 	if (!vers_printed++)
1029 		pr_info("Chelsio T4/T5 RDMA Driver - version %s\n",
1030 			DRV_VERSION);
1031 
1032 	ctx = kzalloc(sizeof *ctx, GFP_KERNEL);
1033 	if (!ctx) {
1034 		ctx = ERR_PTR(-ENOMEM);
1035 		goto out;
1036 	}
1037 	ctx->lldi = *infop;
1038 
1039 	PDBG("%s found device %s nchan %u nrxq %u ntxq %u nports %u\n",
1040 	     __func__, pci_name(ctx->lldi.pdev),
1041 	     ctx->lldi.nchan, ctx->lldi.nrxq,
1042 	     ctx->lldi.ntxq, ctx->lldi.nports);
1043 
1044 	mutex_lock(&dev_mutex);
1045 	list_add_tail(&ctx->entry, &uld_ctx_list);
1046 	mutex_unlock(&dev_mutex);
1047 
1048 	for (i = 0; i < ctx->lldi.nrxq; i++)
1049 		PDBG("rxqid[%u] %u\n", i, ctx->lldi.rxq_ids[i]);
1050 out:
1051 	return ctx;
1052 }
1053 
1054 static inline struct sk_buff *copy_gl_to_skb_pkt(const struct pkt_gl *gl,
1055 						 const __be64 *rsp,
1056 						 u32 pktshift)
1057 {
1058 	struct sk_buff *skb;
1059 
1060 	/*
1061 	 * Allocate space for cpl_pass_accept_req which will be synthesized by
1062 	 * driver. Once the driver synthesizes the request the skb will go
1063 	 * through the regular cpl_pass_accept_req processing.
1064 	 * The math here assumes sizeof cpl_pass_accept_req >= sizeof
1065 	 * cpl_rx_pkt.
1066 	 */
1067 	skb = alloc_skb(gl->tot_len + sizeof(struct cpl_pass_accept_req) +
1068 			sizeof(struct rss_header) - pktshift, GFP_ATOMIC);
1069 	if (unlikely(!skb))
1070 		return NULL;
1071 
1072 	 __skb_put(skb, gl->tot_len + sizeof(struct cpl_pass_accept_req) +
1073 		   sizeof(struct rss_header) - pktshift);
1074 
1075 	/*
1076 	 * This skb will contain:
1077 	 *   rss_header from the rspq descriptor (1 flit)
1078 	 *   cpl_rx_pkt struct from the rspq descriptor (2 flits)
1079 	 *   space for the difference between the size of an
1080 	 *      rx_pkt and pass_accept_req cpl (1 flit)
1081 	 *   the packet data from the gl
1082 	 */
1083 	skb_copy_to_linear_data(skb, rsp, sizeof(struct cpl_pass_accept_req) +
1084 				sizeof(struct rss_header));
1085 	skb_copy_to_linear_data_offset(skb, sizeof(struct rss_header) +
1086 				       sizeof(struct cpl_pass_accept_req),
1087 				       gl->va + pktshift,
1088 				       gl->tot_len - pktshift);
1089 	return skb;
1090 }
1091 
1092 static inline int recv_rx_pkt(struct c4iw_dev *dev, const struct pkt_gl *gl,
1093 			   const __be64 *rsp)
1094 {
1095 	unsigned int opcode = *(u8 *)rsp;
1096 	struct sk_buff *skb;
1097 
1098 	if (opcode != CPL_RX_PKT)
1099 		goto out;
1100 
1101 	skb = copy_gl_to_skb_pkt(gl , rsp, dev->rdev.lldi.sge_pktshift);
1102 	if (skb == NULL)
1103 		goto out;
1104 
1105 	if (c4iw_handlers[opcode] == NULL) {
1106 		pr_info("%s no handler opcode 0x%x...\n", __func__,
1107 		       opcode);
1108 		kfree_skb(skb);
1109 		goto out;
1110 	}
1111 	c4iw_handlers[opcode](dev, skb);
1112 	return 1;
1113 out:
1114 	return 0;
1115 }
1116 
1117 static int c4iw_uld_rx_handler(void *handle, const __be64 *rsp,
1118 			const struct pkt_gl *gl)
1119 {
1120 	struct uld_ctx *ctx = handle;
1121 	struct c4iw_dev *dev = ctx->dev;
1122 	struct sk_buff *skb;
1123 	u8 opcode;
1124 
1125 	if (gl == NULL) {
1126 		/* omit RSS and rsp_ctrl at end of descriptor */
1127 		unsigned int len = 64 - sizeof(struct rsp_ctrl) - 8;
1128 
1129 		skb = alloc_skb(256, GFP_ATOMIC);
1130 		if (!skb)
1131 			goto nomem;
1132 		__skb_put(skb, len);
1133 		skb_copy_to_linear_data(skb, &rsp[1], len);
1134 	} else if (gl == CXGB4_MSG_AN) {
1135 		const struct rsp_ctrl *rc = (void *)rsp;
1136 
1137 		u32 qid = be32_to_cpu(rc->pldbuflen_qid);
1138 		c4iw_ev_handler(dev, qid);
1139 		return 0;
1140 	} else if (unlikely(*(u8 *)rsp != *(u8 *)gl->va)) {
1141 		if (recv_rx_pkt(dev, gl, rsp))
1142 			return 0;
1143 
1144 		pr_info("%s: unexpected FL contents at %p, " \
1145 		       "RSS %#llx, FL %#llx, len %u\n",
1146 		       pci_name(ctx->lldi.pdev), gl->va,
1147 		       (unsigned long long)be64_to_cpu(*rsp),
1148 		       (unsigned long long)be64_to_cpu(
1149 		       *(__force __be64 *)gl->va),
1150 		       gl->tot_len);
1151 
1152 		return 0;
1153 	} else {
1154 		skb = cxgb4_pktgl_to_skb(gl, 128, 128);
1155 		if (unlikely(!skb))
1156 			goto nomem;
1157 	}
1158 
1159 	opcode = *(u8 *)rsp;
1160 	if (c4iw_handlers[opcode]) {
1161 		c4iw_handlers[opcode](dev, skb);
1162 	} else {
1163 		pr_info("%s no handler opcode 0x%x...\n", __func__,
1164 		       opcode);
1165 		kfree_skb(skb);
1166 	}
1167 
1168 	return 0;
1169 nomem:
1170 	return -1;
1171 }
1172 
1173 static int c4iw_uld_state_change(void *handle, enum cxgb4_state new_state)
1174 {
1175 	struct uld_ctx *ctx = handle;
1176 
1177 	PDBG("%s new_state %u\n", __func__, new_state);
1178 	switch (new_state) {
1179 	case CXGB4_STATE_UP:
1180 		printk(KERN_INFO MOD "%s: Up\n", pci_name(ctx->lldi.pdev));
1181 		if (!ctx->dev) {
1182 			int ret;
1183 
1184 			ctx->dev = c4iw_alloc(&ctx->lldi);
1185 			if (IS_ERR(ctx->dev)) {
1186 				printk(KERN_ERR MOD
1187 				       "%s: initialization failed: %ld\n",
1188 				       pci_name(ctx->lldi.pdev),
1189 				       PTR_ERR(ctx->dev));
1190 				ctx->dev = NULL;
1191 				break;
1192 			}
1193 			ret = c4iw_register_device(ctx->dev);
1194 			if (ret) {
1195 				printk(KERN_ERR MOD
1196 				       "%s: RDMA registration failed: %d\n",
1197 				       pci_name(ctx->lldi.pdev), ret);
1198 				c4iw_dealloc(ctx);
1199 			}
1200 		}
1201 		break;
1202 	case CXGB4_STATE_DOWN:
1203 		printk(KERN_INFO MOD "%s: Down\n",
1204 		       pci_name(ctx->lldi.pdev));
1205 		if (ctx->dev)
1206 			c4iw_remove(ctx);
1207 		break;
1208 	case CXGB4_STATE_START_RECOVERY:
1209 		printk(KERN_INFO MOD "%s: Fatal Error\n",
1210 		       pci_name(ctx->lldi.pdev));
1211 		if (ctx->dev) {
1212 			struct ib_event event;
1213 
1214 			ctx->dev->rdev.flags |= T4_FATAL_ERROR;
1215 			memset(&event, 0, sizeof event);
1216 			event.event  = IB_EVENT_DEVICE_FATAL;
1217 			event.device = &ctx->dev->ibdev;
1218 			ib_dispatch_event(&event);
1219 			c4iw_remove(ctx);
1220 		}
1221 		break;
1222 	case CXGB4_STATE_DETACH:
1223 		printk(KERN_INFO MOD "%s: Detach\n",
1224 		       pci_name(ctx->lldi.pdev));
1225 		if (ctx->dev)
1226 			c4iw_remove(ctx);
1227 		break;
1228 	}
1229 	return 0;
1230 }
1231 
1232 static int disable_qp_db(int id, void *p, void *data)
1233 {
1234 	struct c4iw_qp *qp = p;
1235 
1236 	t4_disable_wq_db(&qp->wq);
1237 	return 0;
1238 }
1239 
1240 static void stop_queues(struct uld_ctx *ctx)
1241 {
1242 	unsigned long flags;
1243 
1244 	spin_lock_irqsave(&ctx->dev->lock, flags);
1245 	ctx->dev->rdev.stats.db_state_transitions++;
1246 	ctx->dev->db_state = STOPPED;
1247 	if (ctx->dev->rdev.flags & T4_STATUS_PAGE_DISABLED)
1248 		idr_for_each(&ctx->dev->qpidr, disable_qp_db, NULL);
1249 	else
1250 		ctx->dev->rdev.status_page->db_off = 1;
1251 	spin_unlock_irqrestore(&ctx->dev->lock, flags);
1252 }
1253 
1254 static int enable_qp_db(int id, void *p, void *data)
1255 {
1256 	struct c4iw_qp *qp = p;
1257 
1258 	t4_enable_wq_db(&qp->wq);
1259 	return 0;
1260 }
1261 
1262 static void resume_rc_qp(struct c4iw_qp *qp)
1263 {
1264 	spin_lock(&qp->lock);
1265 	t4_ring_sq_db(&qp->wq, qp->wq.sq.wq_pidx_inc, NULL);
1266 	qp->wq.sq.wq_pidx_inc = 0;
1267 	t4_ring_rq_db(&qp->wq, qp->wq.rq.wq_pidx_inc, NULL);
1268 	qp->wq.rq.wq_pidx_inc = 0;
1269 	spin_unlock(&qp->lock);
1270 }
1271 
1272 static void resume_a_chunk(struct uld_ctx *ctx)
1273 {
1274 	int i;
1275 	struct c4iw_qp *qp;
1276 
1277 	for (i = 0; i < DB_FC_RESUME_SIZE; i++) {
1278 		qp = list_first_entry(&ctx->dev->db_fc_list, struct c4iw_qp,
1279 				      db_fc_entry);
1280 		list_del_init(&qp->db_fc_entry);
1281 		resume_rc_qp(qp);
1282 		if (list_empty(&ctx->dev->db_fc_list))
1283 			break;
1284 	}
1285 }
1286 
1287 static void resume_queues(struct uld_ctx *ctx)
1288 {
1289 	spin_lock_irq(&ctx->dev->lock);
1290 	if (ctx->dev->db_state != STOPPED)
1291 		goto out;
1292 	ctx->dev->db_state = FLOW_CONTROL;
1293 	while (1) {
1294 		if (list_empty(&ctx->dev->db_fc_list)) {
1295 			WARN_ON(ctx->dev->db_state != FLOW_CONTROL);
1296 			ctx->dev->db_state = NORMAL;
1297 			ctx->dev->rdev.stats.db_state_transitions++;
1298 			if (ctx->dev->rdev.flags & T4_STATUS_PAGE_DISABLED) {
1299 				idr_for_each(&ctx->dev->qpidr, enable_qp_db,
1300 					     NULL);
1301 			} else {
1302 				ctx->dev->rdev.status_page->db_off = 0;
1303 			}
1304 			break;
1305 		} else {
1306 			if (cxgb4_dbfifo_count(ctx->dev->rdev.lldi.ports[0], 1)
1307 			    < (ctx->dev->rdev.lldi.dbfifo_int_thresh <<
1308 			       DB_FC_DRAIN_THRESH)) {
1309 				resume_a_chunk(ctx);
1310 			}
1311 			if (!list_empty(&ctx->dev->db_fc_list)) {
1312 				spin_unlock_irq(&ctx->dev->lock);
1313 				if (DB_FC_RESUME_DELAY) {
1314 					set_current_state(TASK_UNINTERRUPTIBLE);
1315 					schedule_timeout(DB_FC_RESUME_DELAY);
1316 				}
1317 				spin_lock_irq(&ctx->dev->lock);
1318 				if (ctx->dev->db_state != FLOW_CONTROL)
1319 					break;
1320 			}
1321 		}
1322 	}
1323 out:
1324 	if (ctx->dev->db_state != NORMAL)
1325 		ctx->dev->rdev.stats.db_fc_interruptions++;
1326 	spin_unlock_irq(&ctx->dev->lock);
1327 }
1328 
1329 struct qp_list {
1330 	unsigned idx;
1331 	struct c4iw_qp **qps;
1332 };
1333 
1334 static int add_and_ref_qp(int id, void *p, void *data)
1335 {
1336 	struct qp_list *qp_listp = data;
1337 	struct c4iw_qp *qp = p;
1338 
1339 	c4iw_qp_add_ref(&qp->ibqp);
1340 	qp_listp->qps[qp_listp->idx++] = qp;
1341 	return 0;
1342 }
1343 
1344 static int count_qps(int id, void *p, void *data)
1345 {
1346 	unsigned *countp = data;
1347 	(*countp)++;
1348 	return 0;
1349 }
1350 
1351 static void deref_qps(struct qp_list *qp_list)
1352 {
1353 	int idx;
1354 
1355 	for (idx = 0; idx < qp_list->idx; idx++)
1356 		c4iw_qp_rem_ref(&qp_list->qps[idx]->ibqp);
1357 }
1358 
1359 static void recover_lost_dbs(struct uld_ctx *ctx, struct qp_list *qp_list)
1360 {
1361 	int idx;
1362 	int ret;
1363 
1364 	for (idx = 0; idx < qp_list->idx; idx++) {
1365 		struct c4iw_qp *qp = qp_list->qps[idx];
1366 
1367 		spin_lock_irq(&qp->rhp->lock);
1368 		spin_lock(&qp->lock);
1369 		ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
1370 					  qp->wq.sq.qid,
1371 					  t4_sq_host_wq_pidx(&qp->wq),
1372 					  t4_sq_wq_size(&qp->wq));
1373 		if (ret) {
1374 			pr_err(MOD "%s: Fatal error - "
1375 			       "DB overflow recovery failed - "
1376 			       "error syncing SQ qid %u\n",
1377 			       pci_name(ctx->lldi.pdev), qp->wq.sq.qid);
1378 			spin_unlock(&qp->lock);
1379 			spin_unlock_irq(&qp->rhp->lock);
1380 			return;
1381 		}
1382 		qp->wq.sq.wq_pidx_inc = 0;
1383 
1384 		ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
1385 					  qp->wq.rq.qid,
1386 					  t4_rq_host_wq_pidx(&qp->wq),
1387 					  t4_rq_wq_size(&qp->wq));
1388 
1389 		if (ret) {
1390 			pr_err(MOD "%s: Fatal error - "
1391 			       "DB overflow recovery failed - "
1392 			       "error syncing RQ qid %u\n",
1393 			       pci_name(ctx->lldi.pdev), qp->wq.rq.qid);
1394 			spin_unlock(&qp->lock);
1395 			spin_unlock_irq(&qp->rhp->lock);
1396 			return;
1397 		}
1398 		qp->wq.rq.wq_pidx_inc = 0;
1399 		spin_unlock(&qp->lock);
1400 		spin_unlock_irq(&qp->rhp->lock);
1401 
1402 		/* Wait for the dbfifo to drain */
1403 		while (cxgb4_dbfifo_count(qp->rhp->rdev.lldi.ports[0], 1) > 0) {
1404 			set_current_state(TASK_UNINTERRUPTIBLE);
1405 			schedule_timeout(usecs_to_jiffies(10));
1406 		}
1407 	}
1408 }
1409 
1410 static void recover_queues(struct uld_ctx *ctx)
1411 {
1412 	int count = 0;
1413 	struct qp_list qp_list;
1414 	int ret;
1415 
1416 	/* slow everybody down */
1417 	set_current_state(TASK_UNINTERRUPTIBLE);
1418 	schedule_timeout(usecs_to_jiffies(1000));
1419 
1420 	/* flush the SGE contexts */
1421 	ret = cxgb4_flush_eq_cache(ctx->dev->rdev.lldi.ports[0]);
1422 	if (ret) {
1423 		printk(KERN_ERR MOD "%s: Fatal error - DB overflow recovery failed\n",
1424 		       pci_name(ctx->lldi.pdev));
1425 		return;
1426 	}
1427 
1428 	/* Count active queues so we can build a list of queues to recover */
1429 	spin_lock_irq(&ctx->dev->lock);
1430 	WARN_ON(ctx->dev->db_state != STOPPED);
1431 	ctx->dev->db_state = RECOVERY;
1432 	idr_for_each(&ctx->dev->qpidr, count_qps, &count);
1433 
1434 	qp_list.qps = kzalloc(count * sizeof *qp_list.qps, GFP_ATOMIC);
1435 	if (!qp_list.qps) {
1436 		spin_unlock_irq(&ctx->dev->lock);
1437 		return;
1438 	}
1439 	qp_list.idx = 0;
1440 
1441 	/* add and ref each qp so it doesn't get freed */
1442 	idr_for_each(&ctx->dev->qpidr, add_and_ref_qp, &qp_list);
1443 
1444 	spin_unlock_irq(&ctx->dev->lock);
1445 
1446 	/* now traverse the list in a safe context to recover the db state*/
1447 	recover_lost_dbs(ctx, &qp_list);
1448 
1449 	/* we're almost done!  deref the qps and clean up */
1450 	deref_qps(&qp_list);
1451 	kfree(qp_list.qps);
1452 
1453 	spin_lock_irq(&ctx->dev->lock);
1454 	WARN_ON(ctx->dev->db_state != RECOVERY);
1455 	ctx->dev->db_state = STOPPED;
1456 	spin_unlock_irq(&ctx->dev->lock);
1457 }
1458 
1459 static int c4iw_uld_control(void *handle, enum cxgb4_control control, ...)
1460 {
1461 	struct uld_ctx *ctx = handle;
1462 
1463 	switch (control) {
1464 	case CXGB4_CONTROL_DB_FULL:
1465 		stop_queues(ctx);
1466 		ctx->dev->rdev.stats.db_full++;
1467 		break;
1468 	case CXGB4_CONTROL_DB_EMPTY:
1469 		resume_queues(ctx);
1470 		mutex_lock(&ctx->dev->rdev.stats.lock);
1471 		ctx->dev->rdev.stats.db_empty++;
1472 		mutex_unlock(&ctx->dev->rdev.stats.lock);
1473 		break;
1474 	case CXGB4_CONTROL_DB_DROP:
1475 		recover_queues(ctx);
1476 		mutex_lock(&ctx->dev->rdev.stats.lock);
1477 		ctx->dev->rdev.stats.db_drop++;
1478 		mutex_unlock(&ctx->dev->rdev.stats.lock);
1479 		break;
1480 	default:
1481 		printk(KERN_WARNING MOD "%s: unknown control cmd %u\n",
1482 		       pci_name(ctx->lldi.pdev), control);
1483 		break;
1484 	}
1485 	return 0;
1486 }
1487 
1488 static struct cxgb4_uld_info c4iw_uld_info = {
1489 	.name = DRV_NAME,
1490 	.nrxq = MAX_ULD_QSETS,
1491 	.ntxq = MAX_ULD_QSETS,
1492 	.rxq_size = 511,
1493 	.ciq = true,
1494 	.lro = false,
1495 	.add = c4iw_uld_add,
1496 	.rx_handler = c4iw_uld_rx_handler,
1497 	.state_change = c4iw_uld_state_change,
1498 	.control = c4iw_uld_control,
1499 };
1500 
1501 static int __init c4iw_init_module(void)
1502 {
1503 	int err;
1504 
1505 	err = c4iw_cm_init();
1506 	if (err)
1507 		return err;
1508 
1509 	c4iw_debugfs_root = debugfs_create_dir(DRV_NAME, NULL);
1510 	if (!c4iw_debugfs_root)
1511 		printk(KERN_WARNING MOD
1512 		       "could not create debugfs entry, continuing\n");
1513 
1514 	cxgb4_register_uld(CXGB4_ULD_RDMA, &c4iw_uld_info);
1515 
1516 	return 0;
1517 }
1518 
1519 static void __exit c4iw_exit_module(void)
1520 {
1521 	struct uld_ctx *ctx, *tmp;
1522 
1523 	mutex_lock(&dev_mutex);
1524 	list_for_each_entry_safe(ctx, tmp, &uld_ctx_list, entry) {
1525 		if (ctx->dev)
1526 			c4iw_remove(ctx);
1527 		kfree(ctx);
1528 	}
1529 	mutex_unlock(&dev_mutex);
1530 	cxgb4_unregister_uld(CXGB4_ULD_RDMA);
1531 	c4iw_cm_term();
1532 	debugfs_remove_recursive(c4iw_debugfs_root);
1533 }
1534 
1535 module_init(c4iw_init_module);
1536 module_exit(c4iw_exit_module);
1537