xref: /linux/drivers/infiniband/hw/qib/qib_driver.c (revision 93d4e8bb3f137e8037a65ea96f175f81c25c50e5)
1 /*
2  * Copyright (c) 2021 Cornelis Networks. All rights reserved.
3  * Copyright (c) 2013 Intel Corporation. All rights reserved.
4  * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved.
5  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  */
35 
36 #include <linux/spinlock.h>
37 #include <linux/pci.h>
38 #include <linux/io.h>
39 #include <linux/delay.h>
40 #include <linux/netdevice.h>
41 #include <linux/vmalloc.h>
42 #include <linux/module.h>
43 #include <linux/prefetch.h>
44 
45 #include "qib.h"
46 
47 DEFINE_MUTEX(qib_mutex);	/* general driver use */
48 
49 unsigned qib_ibmtu;
50 module_param_named(ibmtu, qib_ibmtu, uint, S_IRUGO);
51 MODULE_PARM_DESC(ibmtu, "Set max IB MTU (0=2KB, 1=256, 2=512, ... 5=4096");
52 
53 unsigned qib_compat_ddr_negotiate = 1;
54 module_param_named(compat_ddr_negotiate, qib_compat_ddr_negotiate, uint,
55 		   S_IWUSR | S_IRUGO);
56 MODULE_PARM_DESC(compat_ddr_negotiate,
57 		 "Attempt pre-IBTA 1.2 DDR speed negotiation");
58 
59 MODULE_LICENSE("Dual BSD/GPL");
60 MODULE_AUTHOR("Cornelis <support@cornelisnetworks.com>");
61 MODULE_DESCRIPTION("Cornelis IB driver");
62 
63 /*
64  * QIB_PIO_MAXIBHDR is the max IB header size allowed for in our
65  * PIO send buffers.  This is well beyond anything currently
66  * defined in the InfiniBand spec.
67  */
68 #define QIB_PIO_MAXIBHDR 128
69 
70 /*
71  * QIB_MAX_PKT_RCV is the max # if packets processed per receive interrupt.
72  */
73 #define QIB_MAX_PKT_RECV 64
74 
75 struct qlogic_ib_stats qib_stats;
76 
77 struct pci_dev *qib_get_pci_dev(struct rvt_dev_info *rdi)
78 {
79 	struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
80 	struct qib_devdata *dd = container_of(ibdev,
81 					      struct qib_devdata, verbs_dev);
82 	return dd->pcidev;
83 }
84 
85 /*
86  * Return count of units with at least one port ACTIVE.
87  */
88 int qib_count_active_units(void)
89 {
90 	struct qib_devdata *dd;
91 	struct qib_pportdata *ppd;
92 	unsigned long index, flags;
93 	int pidx, nunits_active = 0;
94 
95 	xa_lock_irqsave(&qib_dev_table, flags);
96 	xa_for_each(&qib_dev_table, index, dd) {
97 		if (!(dd->flags & QIB_PRESENT) || !dd->kregbase)
98 			continue;
99 		for (pidx = 0; pidx < dd->num_pports; ++pidx) {
100 			ppd = dd->pport + pidx;
101 			if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
102 					 QIBL_LINKARMED | QIBL_LINKACTIVE))) {
103 				nunits_active++;
104 				break;
105 			}
106 		}
107 	}
108 	xa_unlock_irqrestore(&qib_dev_table, flags);
109 	return nunits_active;
110 }
111 
112 /*
113  * Return count of all units, optionally return in arguments
114  * the number of usable (present) units, and the number of
115  * ports that are up.
116  */
117 int qib_count_units(int *npresentp, int *nupp)
118 {
119 	int nunits = 0, npresent = 0, nup = 0;
120 	struct qib_devdata *dd;
121 	unsigned long index, flags;
122 	int pidx;
123 	struct qib_pportdata *ppd;
124 
125 	xa_lock_irqsave(&qib_dev_table, flags);
126 	xa_for_each(&qib_dev_table, index, dd) {
127 		nunits++;
128 		if ((dd->flags & QIB_PRESENT) && dd->kregbase)
129 			npresent++;
130 		for (pidx = 0; pidx < dd->num_pports; ++pidx) {
131 			ppd = dd->pport + pidx;
132 			if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
133 					 QIBL_LINKARMED | QIBL_LINKACTIVE)))
134 				nup++;
135 		}
136 	}
137 	xa_unlock_irqrestore(&qib_dev_table, flags);
138 
139 	if (npresentp)
140 		*npresentp = npresent;
141 	if (nupp)
142 		*nupp = nup;
143 
144 	return nunits;
145 }
146 
147 /**
148  * qib_wait_linkstate - wait for an IB link state change to occur
149  * @ppd: the qlogic_ib device
150  * @state: the state to wait for
151  * @msecs: the number of milliseconds to wait
152  *
153  * wait up to msecs milliseconds for IB link state change to occur for
154  * now, take the easy polling route.  Currently used only by
155  * qib_set_linkstate.  Returns 0 if state reached, otherwise
156  * -ETIMEDOUT state can have multiple states set, for any of several
157  * transitions.
158  */
159 int qib_wait_linkstate(struct qib_pportdata *ppd, u32 state, int msecs)
160 {
161 	int ret;
162 	unsigned long flags;
163 
164 	spin_lock_irqsave(&ppd->lflags_lock, flags);
165 	if (ppd->state_wanted) {
166 		spin_unlock_irqrestore(&ppd->lflags_lock, flags);
167 		ret = -EBUSY;
168 		goto bail;
169 	}
170 	ppd->state_wanted = state;
171 	spin_unlock_irqrestore(&ppd->lflags_lock, flags);
172 	wait_event_interruptible_timeout(ppd->state_wait,
173 					 (ppd->lflags & state),
174 					 msecs_to_jiffies(msecs));
175 	spin_lock_irqsave(&ppd->lflags_lock, flags);
176 	ppd->state_wanted = 0;
177 	spin_unlock_irqrestore(&ppd->lflags_lock, flags);
178 
179 	if (!(ppd->lflags & state))
180 		ret = -ETIMEDOUT;
181 	else
182 		ret = 0;
183 bail:
184 	return ret;
185 }
186 
187 int qib_set_linkstate(struct qib_pportdata *ppd, u8 newstate)
188 {
189 	u32 lstate;
190 	int ret;
191 	struct qib_devdata *dd = ppd->dd;
192 	unsigned long flags;
193 
194 	switch (newstate) {
195 	case QIB_IB_LINKDOWN_ONLY:
196 		dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
197 				 IB_LINKCMD_DOWN | IB_LINKINITCMD_NOP);
198 		/* don't wait */
199 		ret = 0;
200 		goto bail;
201 
202 	case QIB_IB_LINKDOWN:
203 		dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
204 				 IB_LINKCMD_DOWN | IB_LINKINITCMD_POLL);
205 		/* don't wait */
206 		ret = 0;
207 		goto bail;
208 
209 	case QIB_IB_LINKDOWN_SLEEP:
210 		dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
211 				 IB_LINKCMD_DOWN | IB_LINKINITCMD_SLEEP);
212 		/* don't wait */
213 		ret = 0;
214 		goto bail;
215 
216 	case QIB_IB_LINKDOWN_DISABLE:
217 		dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
218 				 IB_LINKCMD_DOWN | IB_LINKINITCMD_DISABLE);
219 		/* don't wait */
220 		ret = 0;
221 		goto bail;
222 
223 	case QIB_IB_LINKARM:
224 		if (ppd->lflags & QIBL_LINKARMED) {
225 			ret = 0;
226 			goto bail;
227 		}
228 		if (!(ppd->lflags & (QIBL_LINKINIT | QIBL_LINKACTIVE))) {
229 			ret = -EINVAL;
230 			goto bail;
231 		}
232 		/*
233 		 * Since the port can be ACTIVE when we ask for ARMED,
234 		 * clear QIBL_LINKV so we can wait for a transition.
235 		 * If the link isn't ARMED, then something else happened
236 		 * and there is no point waiting for ARMED.
237 		 */
238 		spin_lock_irqsave(&ppd->lflags_lock, flags);
239 		ppd->lflags &= ~QIBL_LINKV;
240 		spin_unlock_irqrestore(&ppd->lflags_lock, flags);
241 		dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
242 				 IB_LINKCMD_ARMED | IB_LINKINITCMD_NOP);
243 		lstate = QIBL_LINKV;
244 		break;
245 
246 	case QIB_IB_LINKACTIVE:
247 		if (ppd->lflags & QIBL_LINKACTIVE) {
248 			ret = 0;
249 			goto bail;
250 		}
251 		if (!(ppd->lflags & QIBL_LINKARMED)) {
252 			ret = -EINVAL;
253 			goto bail;
254 		}
255 		dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
256 				 IB_LINKCMD_ACTIVE | IB_LINKINITCMD_NOP);
257 		lstate = QIBL_LINKACTIVE;
258 		break;
259 
260 	default:
261 		ret = -EINVAL;
262 		goto bail;
263 	}
264 	ret = qib_wait_linkstate(ppd, lstate, 10);
265 
266 bail:
267 	return ret;
268 }
269 
270 /*
271  * Get address of eager buffer from it's index (allocated in chunks, not
272  * contiguous).
273  */
274 static inline void *qib_get_egrbuf(const struct qib_ctxtdata *rcd, u32 etail)
275 {
276 	const u32 chunk = etail >> rcd->rcvegrbufs_perchunk_shift;
277 	const u32 idx =  etail & ((u32)rcd->rcvegrbufs_perchunk - 1);
278 
279 	return rcd->rcvegrbuf[chunk] + (idx << rcd->dd->rcvegrbufsize_shift);
280 }
281 
282 /*
283  * Returns 1 if error was a CRC, else 0.
284  * Needed for some chip's synthesized error counters.
285  */
286 static u32 qib_rcv_hdrerr(struct qib_ctxtdata *rcd, struct qib_pportdata *ppd,
287 			  u32 ctxt, u32 eflags, u32 l, u32 etail,
288 			  __le32 *rhf_addr, struct qib_message_header *rhdr)
289 {
290 	u32 ret = 0;
291 
292 	if (eflags & (QLOGIC_IB_RHF_H_ICRCERR | QLOGIC_IB_RHF_H_VCRCERR))
293 		ret = 1;
294 	else if (eflags == QLOGIC_IB_RHF_H_TIDERR) {
295 		/* For TIDERR and RC QPs premptively schedule a NAK */
296 		struct ib_header *hdr = (struct ib_header *)rhdr;
297 		struct ib_other_headers *ohdr = NULL;
298 		struct qib_ibport *ibp = &ppd->ibport_data;
299 		struct qib_devdata *dd = ppd->dd;
300 		struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
301 		struct rvt_qp *qp = NULL;
302 		u32 tlen = qib_hdrget_length_in_bytes(rhf_addr);
303 		u16 lid  = be16_to_cpu(hdr->lrh[1]);
304 		int lnh = be16_to_cpu(hdr->lrh[0]) & 3;
305 		u32 qp_num;
306 		u32 opcode;
307 		u32 psn;
308 		int diff;
309 
310 		/* Sanity check packet */
311 		if (tlen < 24)
312 			goto drop;
313 
314 		if (lid < be16_to_cpu(IB_MULTICAST_LID_BASE)) {
315 			lid &= ~((1 << ppd->lmc) - 1);
316 			if (unlikely(lid != ppd->lid))
317 				goto drop;
318 		}
319 
320 		/* Check for GRH */
321 		if (lnh == QIB_LRH_BTH)
322 			ohdr = &hdr->u.oth;
323 		else if (lnh == QIB_LRH_GRH) {
324 			u32 vtf;
325 
326 			ohdr = &hdr->u.l.oth;
327 			if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
328 				goto drop;
329 			vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
330 			if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
331 				goto drop;
332 		} else
333 			goto drop;
334 
335 		/* Get opcode and PSN from packet */
336 		opcode = be32_to_cpu(ohdr->bth[0]);
337 		opcode >>= 24;
338 		psn = be32_to_cpu(ohdr->bth[2]);
339 
340 		/* Get the destination QP number. */
341 		qp_num = be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK;
342 		if (qp_num != QIB_MULTICAST_QPN) {
343 			int ruc_res;
344 
345 			rcu_read_lock();
346 			qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
347 			if (!qp) {
348 				rcu_read_unlock();
349 				goto drop;
350 			}
351 
352 			/*
353 			 * Handle only RC QPs - for other QP types drop error
354 			 * packet.
355 			 */
356 			spin_lock(&qp->r_lock);
357 
358 			/* Check for valid receive state. */
359 			if (!(ib_rvt_state_ops[qp->state] &
360 			      RVT_PROCESS_RECV_OK)) {
361 				ibp->rvp.n_pkt_drops++;
362 				goto unlock;
363 			}
364 
365 			switch (qp->ibqp.qp_type) {
366 			case IB_QPT_RC:
367 				ruc_res =
368 					qib_ruc_check_hdr(
369 						ibp, hdr,
370 						lnh == QIB_LRH_GRH,
371 						qp,
372 						be32_to_cpu(ohdr->bth[0]));
373 				if (ruc_res)
374 					goto unlock;
375 
376 				/* Only deal with RDMA Writes for now */
377 				if (opcode <
378 				    IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
379 					diff = qib_cmp24(psn, qp->r_psn);
380 					if (!qp->r_nak_state && diff >= 0) {
381 						ibp->rvp.n_rc_seqnak++;
382 						qp->r_nak_state =
383 							IB_NAK_PSN_ERROR;
384 						/* Use the expected PSN. */
385 						qp->r_ack_psn = qp->r_psn;
386 						/*
387 						 * Wait to send the sequence
388 						 * NAK until all packets
389 						 * in the receive queue have
390 						 * been processed.
391 						 * Otherwise, we end up
392 						 * propagating congestion.
393 						 */
394 						if (list_empty(&qp->rspwait)) {
395 							qp->r_flags |=
396 								RVT_R_RSP_NAK;
397 							rvt_get_qp(qp);
398 							list_add_tail(
399 							 &qp->rspwait,
400 							 &rcd->qp_wait_list);
401 						}
402 					} /* Out of sequence NAK */
403 				} /* QP Request NAKs */
404 				break;
405 			case IB_QPT_SMI:
406 			case IB_QPT_GSI:
407 			case IB_QPT_UD:
408 			case IB_QPT_UC:
409 			default:
410 				/* For now don't handle any other QP types */
411 				break;
412 			}
413 
414 unlock:
415 			spin_unlock(&qp->r_lock);
416 			rcu_read_unlock();
417 		} /* Unicast QP */
418 	} /* Valid packet with TIDErr */
419 
420 drop:
421 	return ret;
422 }
423 
424 /*
425  * qib_kreceive - receive a packet
426  * @rcd: the qlogic_ib context
427  * @llic: gets count of good packets needed to clear lli,
428  *          (used with chips that need need to track crcs for lli)
429  *
430  * called from interrupt handler for errors or receive interrupt
431  * Returns number of CRC error packets, needed by some chips for
432  * local link integrity tracking.   crcs are adjusted down by following
433  * good packets, if any, and count of good packets is also tracked.
434  */
435 u32 qib_kreceive(struct qib_ctxtdata *rcd, u32 *llic, u32 *npkts)
436 {
437 	struct qib_devdata *dd = rcd->dd;
438 	struct qib_pportdata *ppd = rcd->ppd;
439 	__le32 *rhf_addr;
440 	void *ebuf;
441 	const u32 rsize = dd->rcvhdrentsize;        /* words */
442 	const u32 maxcnt = dd->rcvhdrcnt * rsize;   /* words */
443 	u32 etail = -1, l, hdrqtail;
444 	struct qib_message_header *hdr;
445 	u32 eflags, etype, tlen, i = 0, updegr = 0, crcs = 0;
446 	int last;
447 	u64 lval;
448 	struct rvt_qp *qp, *nqp;
449 
450 	l = rcd->head;
451 	rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
452 	if (dd->flags & QIB_NODMA_RTAIL) {
453 		u32 seq = qib_hdrget_seq(rhf_addr);
454 
455 		if (seq != rcd->seq_cnt)
456 			goto bail;
457 		hdrqtail = 0;
458 	} else {
459 		hdrqtail = qib_get_rcvhdrtail(rcd);
460 		if (l == hdrqtail)
461 			goto bail;
462 		smp_rmb();  /* prevent speculative reads of dma'ed hdrq */
463 	}
464 
465 	for (last = 0, i = 1; !last; i += !last) {
466 		hdr = dd->f_get_msgheader(dd, rhf_addr);
467 		eflags = qib_hdrget_err_flags(rhf_addr);
468 		etype = qib_hdrget_rcv_type(rhf_addr);
469 		/* total length */
470 		tlen = qib_hdrget_length_in_bytes(rhf_addr);
471 		ebuf = NULL;
472 		if ((dd->flags & QIB_NODMA_RTAIL) ?
473 		    qib_hdrget_use_egr_buf(rhf_addr) :
474 		    (etype != RCVHQ_RCV_TYPE_EXPECTED)) {
475 			etail = qib_hdrget_index(rhf_addr);
476 			updegr = 1;
477 			if (tlen > sizeof(*hdr) ||
478 			    etype >= RCVHQ_RCV_TYPE_NON_KD) {
479 				ebuf = qib_get_egrbuf(rcd, etail);
480 				prefetch_range(ebuf, tlen - sizeof(*hdr));
481 			}
482 		}
483 		if (!eflags) {
484 			u16 lrh_len = be16_to_cpu(hdr->lrh[2]) << 2;
485 
486 			if (lrh_len != tlen) {
487 				qib_stats.sps_lenerrs++;
488 				goto move_along;
489 			}
490 		}
491 		if (etype == RCVHQ_RCV_TYPE_NON_KD && !eflags &&
492 		    ebuf == NULL &&
493 		    tlen > (dd->rcvhdrentsize - 2 + 1 -
494 				qib_hdrget_offset(rhf_addr)) << 2) {
495 			goto move_along;
496 		}
497 
498 		/*
499 		 * Both tiderr and qibhdrerr are set for all plain IB
500 		 * packets; only qibhdrerr should be set.
501 		 */
502 		if (unlikely(eflags))
503 			crcs += qib_rcv_hdrerr(rcd, ppd, rcd->ctxt, eflags, l,
504 					       etail, rhf_addr, hdr);
505 		else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
506 			qib_ib_rcv(rcd, hdr, ebuf, tlen);
507 			if (crcs)
508 				crcs--;
509 			else if (llic && *llic)
510 				--*llic;
511 		}
512 move_along:
513 		l += rsize;
514 		if (l >= maxcnt)
515 			l = 0;
516 		if (i == QIB_MAX_PKT_RECV)
517 			last = 1;
518 
519 		rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
520 		if (dd->flags & QIB_NODMA_RTAIL) {
521 			u32 seq = qib_hdrget_seq(rhf_addr);
522 
523 			if (++rcd->seq_cnt > 13)
524 				rcd->seq_cnt = 1;
525 			if (seq != rcd->seq_cnt)
526 				last = 1;
527 		} else if (l == hdrqtail)
528 			last = 1;
529 		/*
530 		 * Update head regs etc., every 16 packets, if not last pkt,
531 		 * to help prevent rcvhdrq overflows, when many packets
532 		 * are processed and queue is nearly full.
533 		 * Don't request an interrupt for intermediate updates.
534 		 */
535 		lval = l;
536 		if (!last && !(i & 0xf)) {
537 			dd->f_update_usrhead(rcd, lval, updegr, etail, i);
538 			updegr = 0;
539 		}
540 	}
541 
542 	rcd->head = l;
543 
544 	/*
545 	 * Iterate over all QPs waiting to respond.
546 	 * The list won't change since the IRQ is only run on one CPU.
547 	 */
548 	list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
549 		list_del_init(&qp->rspwait);
550 		if (qp->r_flags & RVT_R_RSP_NAK) {
551 			qp->r_flags &= ~RVT_R_RSP_NAK;
552 			qib_send_rc_ack(qp);
553 		}
554 		if (qp->r_flags & RVT_R_RSP_SEND) {
555 			unsigned long flags;
556 
557 			qp->r_flags &= ~RVT_R_RSP_SEND;
558 			spin_lock_irqsave(&qp->s_lock, flags);
559 			if (ib_rvt_state_ops[qp->state] &
560 					RVT_PROCESS_OR_FLUSH_SEND)
561 				qib_schedule_send(qp);
562 			spin_unlock_irqrestore(&qp->s_lock, flags);
563 		}
564 		rvt_put_qp(qp);
565 	}
566 
567 bail:
568 	/* Report number of packets consumed */
569 	if (npkts)
570 		*npkts = i;
571 
572 	/*
573 	 * Always write head at end, and setup rcv interrupt, even
574 	 * if no packets were processed.
575 	 */
576 	lval = (u64)rcd->head | dd->rhdrhead_intr_off;
577 	dd->f_update_usrhead(rcd, lval, updegr, etail, i);
578 	return crcs;
579 }
580 
581 /**
582  * qib_set_mtu - set the MTU
583  * @ppd: the perport data
584  * @arg: the new MTU
585  *
586  * We can handle "any" incoming size, the issue here is whether we
587  * need to restrict our outgoing size.   For now, we don't do any
588  * sanity checking on this, and we don't deal with what happens to
589  * programs that are already running when the size changes.
590  * NOTE: changing the MTU will usually cause the IBC to go back to
591  * link INIT state...
592  */
593 int qib_set_mtu(struct qib_pportdata *ppd, u16 arg)
594 {
595 	u32 piosize;
596 	int ret, chk;
597 
598 	if (arg != 256 && arg != 512 && arg != 1024 && arg != 2048 &&
599 	    arg != 4096) {
600 		ret = -EINVAL;
601 		goto bail;
602 	}
603 	chk = ib_mtu_enum_to_int(qib_ibmtu);
604 	if (chk > 0 && arg > chk) {
605 		ret = -EINVAL;
606 		goto bail;
607 	}
608 
609 	piosize = ppd->ibmaxlen;
610 	ppd->ibmtu = arg;
611 
612 	if (arg >= (piosize - QIB_PIO_MAXIBHDR)) {
613 		/* Only if it's not the initial value (or reset to it) */
614 		if (piosize != ppd->init_ibmaxlen) {
615 			if (arg > piosize && arg <= ppd->init_ibmaxlen)
616 				piosize = ppd->init_ibmaxlen - 2 * sizeof(u32);
617 			ppd->ibmaxlen = piosize;
618 		}
619 	} else if ((arg + QIB_PIO_MAXIBHDR) != ppd->ibmaxlen) {
620 		piosize = arg + QIB_PIO_MAXIBHDR - 2 * sizeof(u32);
621 		ppd->ibmaxlen = piosize;
622 	}
623 
624 	ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_MTU, 0);
625 
626 	ret = 0;
627 
628 bail:
629 	return ret;
630 }
631 
632 int qib_set_lid(struct qib_pportdata *ppd, u32 lid, u8 lmc)
633 {
634 	struct qib_devdata *dd = ppd->dd;
635 
636 	ppd->lid = lid;
637 	ppd->lmc = lmc;
638 
639 	dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LIDLMC,
640 			 lid | (~((1U << lmc) - 1)) << 16);
641 
642 	qib_devinfo(dd->pcidev, "IB%u:%u got a lid: 0x%x\n",
643 		    dd->unit, ppd->port, lid);
644 
645 	return 0;
646 }
647 
648 /*
649  * Following deal with the "obviously simple" task of overriding the state
650  * of the LEDS, which normally indicate link physical and logical status.
651  * The complications arise in dealing with different hardware mappings
652  * and the board-dependent routine being called from interrupts.
653  * and then there's the requirement to _flash_ them.
654  */
655 #define LED_OVER_FREQ_SHIFT 8
656 #define LED_OVER_FREQ_MASK (0xFF<<LED_OVER_FREQ_SHIFT)
657 /* Below is "non-zero" to force override, but both actual LEDs are off */
658 #define LED_OVER_BOTH_OFF (8)
659 
660 static void qib_run_led_override(struct timer_list *t)
661 {
662 	struct qib_pportdata *ppd = from_timer(ppd, t,
663 						    led_override_timer);
664 	struct qib_devdata *dd = ppd->dd;
665 	int timeoff;
666 	int ph_idx;
667 
668 	if (!(dd->flags & QIB_INITTED))
669 		return;
670 
671 	ph_idx = ppd->led_override_phase++ & 1;
672 	ppd->led_override = ppd->led_override_vals[ph_idx];
673 	timeoff = ppd->led_override_timeoff;
674 
675 	dd->f_setextled(ppd, 1);
676 	/*
677 	 * don't re-fire the timer if user asked for it to be off; we let
678 	 * it fire one more time after they turn it off to simplify
679 	 */
680 	if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
681 		mod_timer(&ppd->led_override_timer, jiffies + timeoff);
682 }
683 
684 void qib_set_led_override(struct qib_pportdata *ppd, unsigned int val)
685 {
686 	struct qib_devdata *dd = ppd->dd;
687 	int timeoff, freq;
688 
689 	if (!(dd->flags & QIB_INITTED))
690 		return;
691 
692 	/* First check if we are blinking. If not, use 1HZ polling */
693 	timeoff = HZ;
694 	freq = (val & LED_OVER_FREQ_MASK) >> LED_OVER_FREQ_SHIFT;
695 
696 	if (freq) {
697 		/* For blink, set each phase from one nybble of val */
698 		ppd->led_override_vals[0] = val & 0xF;
699 		ppd->led_override_vals[1] = (val >> 4) & 0xF;
700 		timeoff = (HZ << 4)/freq;
701 	} else {
702 		/* Non-blink set both phases the same. */
703 		ppd->led_override_vals[0] = val & 0xF;
704 		ppd->led_override_vals[1] = val & 0xF;
705 	}
706 	ppd->led_override_timeoff = timeoff;
707 
708 	/*
709 	 * If the timer has not already been started, do so. Use a "quick"
710 	 * timeout so the function will be called soon, to look at our request.
711 	 */
712 	if (atomic_inc_return(&ppd->led_override_timer_active) == 1) {
713 		/* Need to start timer */
714 		timer_setup(&ppd->led_override_timer, qib_run_led_override, 0);
715 		ppd->led_override_timer.expires = jiffies + 1;
716 		add_timer(&ppd->led_override_timer);
717 	} else {
718 		if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
719 			mod_timer(&ppd->led_override_timer, jiffies + 1);
720 		atomic_dec(&ppd->led_override_timer_active);
721 	}
722 }
723 
724 /**
725  * qib_reset_device - reset the chip if possible
726  * @unit: the device to reset
727  *
728  * Whether or not reset is successful, we attempt to re-initialize the chip
729  * (that is, much like a driver unload/reload).  We clear the INITTED flag
730  * so that the various entry points will fail until we reinitialize.  For
731  * now, we only allow this if no user contexts are open that use chip resources
732  */
733 int qib_reset_device(int unit)
734 {
735 	int ret, i;
736 	struct qib_devdata *dd = qib_lookup(unit);
737 	struct qib_pportdata *ppd;
738 	unsigned long flags;
739 	int pidx;
740 
741 	if (!dd) {
742 		ret = -ENODEV;
743 		goto bail;
744 	}
745 
746 	qib_devinfo(dd->pcidev, "Reset on unit %u requested\n", unit);
747 
748 	if (!dd->kregbase || !(dd->flags & QIB_PRESENT)) {
749 		qib_devinfo(dd->pcidev,
750 			"Invalid unit number %u or not initialized or not present\n",
751 			unit);
752 		ret = -ENXIO;
753 		goto bail;
754 	}
755 
756 	spin_lock_irqsave(&dd->uctxt_lock, flags);
757 	if (dd->rcd)
758 		for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
759 			if (!dd->rcd[i] || !dd->rcd[i]->cnt)
760 				continue;
761 			spin_unlock_irqrestore(&dd->uctxt_lock, flags);
762 			ret = -EBUSY;
763 			goto bail;
764 		}
765 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
766 
767 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
768 		ppd = dd->pport + pidx;
769 		if (atomic_read(&ppd->led_override_timer_active)) {
770 			/* Need to stop LED timer, _then_ shut off LEDs */
771 			del_timer_sync(&ppd->led_override_timer);
772 			atomic_set(&ppd->led_override_timer_active, 0);
773 		}
774 
775 		/* Shut off LEDs after we are sure timer is not running */
776 		ppd->led_override = LED_OVER_BOTH_OFF;
777 		dd->f_setextled(ppd, 0);
778 		if (dd->flags & QIB_HAS_SEND_DMA)
779 			qib_teardown_sdma(ppd);
780 	}
781 
782 	ret = dd->f_reset(dd);
783 	if (ret == 1)
784 		ret = qib_init(dd, 1);
785 	else
786 		ret = -EAGAIN;
787 	if (ret)
788 		qib_dev_err(dd,
789 			"Reinitialize unit %u after reset failed with %d\n",
790 			unit, ret);
791 	else
792 		qib_devinfo(dd->pcidev,
793 			"Reinitialized unit %u after resetting\n",
794 			unit);
795 
796 bail:
797 	return ret;
798 }
799