2 * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/spinlock.h>
35 #include <linux/pci.h>
37 #include <linux/delay.h>
38 #include <linux/netdevice.h>
39 #include <linux/vmalloc.h>
44 * The size has to be longer than this string, so we can append
45 * board/chip information to it in the init code.
47 const char ib_qib_version[] = QIB_IDSTR "\n";
49 DEFINE_SPINLOCK(qib_devs_lock);
50 LIST_HEAD(qib_dev_list);
51 DEFINE_MUTEX(qib_mutex); /* general driver use */
54 module_param_named(ibmtu, qib_ibmtu, uint, S_IRUGO);
55 MODULE_PARM_DESC(ibmtu, "Set max IB MTU (0=2KB, 1=256, 2=512, ... 5=4096");
57 unsigned qib_compat_ddr_negotiate = 1;
58 module_param_named(compat_ddr_negotiate, qib_compat_ddr_negotiate, uint,
60 MODULE_PARM_DESC(compat_ddr_negotiate,
61 "Attempt pre-IBTA 1.2 DDR speed negotiation");
63 MODULE_LICENSE("Dual BSD/GPL");
64 MODULE_AUTHOR("QLogic <support@qlogic.com>");
65 MODULE_DESCRIPTION("QLogic IB driver");
68 * QIB_PIO_MAXIBHDR is the max IB header size allowed for in our
69 * PIO send buffers. This is well beyond anything currently
70 * defined in the InfiniBand spec.
72 #define QIB_PIO_MAXIBHDR 128
74 struct qlogic_ib_stats qib_stats;
76 const char *qib_get_unit_name(int unit)
78 static char iname[16];
80 snprintf(iname, sizeof iname, "infinipath%u", unit);
85 * Return count of units with at least one port ACTIVE.
87 int qib_count_active_units(void)
89 struct qib_devdata *dd;
90 struct qib_pportdata *ppd;
92 int pidx, nunits_active = 0;
94 spin_lock_irqsave(&qib_devs_lock, flags);
95 list_for_each_entry(dd, &qib_dev_list, list) {
96 if (!(dd->flags & QIB_PRESENT) || !dd->kregbase)
98 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
99 ppd = dd->pport + pidx;
100 if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
101 QIBL_LINKARMED | QIBL_LINKACTIVE))) {
107 spin_unlock_irqrestore(&qib_devs_lock, flags);
108 return nunits_active;
112 * Return count of all units, optionally return in arguments
113 * the number of usable (present) units, and the number of
116 int qib_count_units(int *npresentp, int *nupp)
118 int nunits = 0, npresent = 0, nup = 0;
119 struct qib_devdata *dd;
122 struct qib_pportdata *ppd;
124 spin_lock_irqsave(&qib_devs_lock, flags);
126 list_for_each_entry(dd, &qib_dev_list, list) {
128 if ((dd->flags & QIB_PRESENT) && dd->kregbase)
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)))
138 spin_unlock_irqrestore(&qib_devs_lock, flags);
141 *npresentp = npresent;
149 * qib_wait_linkstate - wait for an IB link state change to occur
150 * @dd: the qlogic_ib device
151 * @state: the state to wait for
152 * @msecs: the number of milliseconds to wait
154 * wait up to msecs milliseconds for IB link state change to occur for
155 * now, take the easy polling route. Currently used only by
156 * qib_set_linkstate. Returns 0 if state reached, otherwise
157 * -ETIMEDOUT state can have multiple states set, for any of several
160 int qib_wait_linkstate(struct qib_pportdata *ppd, u32 state, int msecs)
165 spin_lock_irqsave(&ppd->lflags_lock, flags);
166 if (ppd->state_wanted) {
167 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
171 ppd->state_wanted = state;
172 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
173 wait_event_interruptible_timeout(ppd->state_wait,
174 (ppd->lflags & state),
175 msecs_to_jiffies(msecs));
176 spin_lock_irqsave(&ppd->lflags_lock, flags);
177 ppd->state_wanted = 0;
178 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
180 if (!(ppd->lflags & state))
188 int qib_set_linkstate(struct qib_pportdata *ppd, u8 newstate)
192 struct qib_devdata *dd = ppd->dd;
196 case QIB_IB_LINKDOWN_ONLY:
197 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
198 IB_LINKCMD_DOWN | IB_LINKINITCMD_NOP);
203 case QIB_IB_LINKDOWN:
204 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
205 IB_LINKCMD_DOWN | IB_LINKINITCMD_POLL);
210 case QIB_IB_LINKDOWN_SLEEP:
211 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
212 IB_LINKCMD_DOWN | IB_LINKINITCMD_SLEEP);
217 case QIB_IB_LINKDOWN_DISABLE:
218 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
219 IB_LINKCMD_DOWN | IB_LINKINITCMD_DISABLE);
225 if (ppd->lflags & QIBL_LINKARMED) {
229 if (!(ppd->lflags & (QIBL_LINKINIT | QIBL_LINKACTIVE))) {
234 * Since the port can be ACTIVE when we ask for ARMED,
235 * clear QIBL_LINKV so we can wait for a transition.
236 * If the link isn't ARMED, then something else happened
237 * and there is no point waiting for ARMED.
239 spin_lock_irqsave(&ppd->lflags_lock, flags);
240 ppd->lflags &= ~QIBL_LINKV;
241 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
242 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
243 IB_LINKCMD_ARMED | IB_LINKINITCMD_NOP);
247 case QIB_IB_LINKACTIVE:
248 if (ppd->lflags & QIBL_LINKACTIVE) {
252 if (!(ppd->lflags & QIBL_LINKARMED)) {
256 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
257 IB_LINKCMD_ACTIVE | IB_LINKINITCMD_NOP);
258 lstate = QIBL_LINKACTIVE;
265 ret = qib_wait_linkstate(ppd, lstate, 10);
272 * Get address of eager buffer from it's index (allocated in chunks, not
275 static inline void *qib_get_egrbuf(const struct qib_ctxtdata *rcd, u32 etail)
277 const u32 chunk = etail / rcd->rcvegrbufs_perchunk;
278 const u32 idx = etail % rcd->rcvegrbufs_perchunk;
280 return rcd->rcvegrbuf[chunk] + idx * rcd->dd->rcvegrbufsize;
284 * Returns 1 if error was a CRC, else 0.
285 * Needed for some chip's synthesized error counters.
287 static u32 qib_rcv_hdrerr(struct qib_pportdata *ppd, u32 ctxt,
288 u32 eflags, u32 l, u32 etail, __le32 *rhf_addr,
289 struct qib_message_header *hdr)
293 if (eflags & (QLOGIC_IB_RHF_H_ICRCERR | QLOGIC_IB_RHF_H_VCRCERR))
299 * qib_kreceive - receive a packet
300 * @rcd: the qlogic_ib context
301 * @llic: gets count of good packets needed to clear lli,
302 * (used with chips that need need to track crcs for lli)
304 * called from interrupt handler for errors or receive interrupt
305 * Returns number of CRC error packets, needed by some chips for
306 * local link integrity tracking. crcs are adjusted down by following
307 * good packets, if any, and count of good packets is also tracked.
309 u32 qib_kreceive(struct qib_ctxtdata *rcd, u32 *llic, u32 *npkts)
311 struct qib_devdata *dd = rcd->dd;
312 struct qib_pportdata *ppd = rcd->ppd;
315 const u32 rsize = dd->rcvhdrentsize; /* words */
316 const u32 maxcnt = dd->rcvhdrcnt * rsize; /* words */
317 u32 etail = -1, l, hdrqtail;
318 struct qib_message_header *hdr;
319 u32 eflags, etype, tlen, i = 0, updegr = 0, crcs = 0;
322 struct qib_qp *qp, *nqp;
325 rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
326 if (dd->flags & QIB_NODMA_RTAIL) {
327 u32 seq = qib_hdrget_seq(rhf_addr);
328 if (seq != rcd->seq_cnt)
332 hdrqtail = qib_get_rcvhdrtail(rcd);
335 smp_rmb(); /* prevent speculative reads of dma'ed hdrq */
338 for (last = 0, i = 1; !last; i += !last) {
339 hdr = dd->f_get_msgheader(dd, rhf_addr);
340 eflags = qib_hdrget_err_flags(rhf_addr);
341 etype = qib_hdrget_rcv_type(rhf_addr);
343 tlen = qib_hdrget_length_in_bytes(rhf_addr);
345 if ((dd->flags & QIB_NODMA_RTAIL) ?
346 qib_hdrget_use_egr_buf(rhf_addr) :
347 (etype != RCVHQ_RCV_TYPE_EXPECTED)) {
348 etail = qib_hdrget_index(rhf_addr);
350 if (tlen > sizeof(*hdr) ||
351 etype >= RCVHQ_RCV_TYPE_NON_KD)
352 ebuf = qib_get_egrbuf(rcd, etail);
355 u16 lrh_len = be16_to_cpu(hdr->lrh[2]) << 2;
357 if (lrh_len != tlen) {
358 qib_stats.sps_lenerrs++;
362 if (etype == RCVHQ_RCV_TYPE_NON_KD && !eflags &&
364 tlen > (dd->rcvhdrentsize - 2 + 1 -
365 qib_hdrget_offset(rhf_addr)) << 2) {
370 * Both tiderr and qibhdrerr are set for all plain IB
371 * packets; only qibhdrerr should be set.
373 if (unlikely(eflags))
374 crcs += qib_rcv_hdrerr(ppd, rcd->ctxt, eflags, l,
375 etail, rhf_addr, hdr);
376 else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
377 qib_ib_rcv(rcd, hdr, ebuf, tlen);
380 else if (llic && *llic)
387 rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
388 if (dd->flags & QIB_NODMA_RTAIL) {
389 u32 seq = qib_hdrget_seq(rhf_addr);
391 if (++rcd->seq_cnt > 13)
393 if (seq != rcd->seq_cnt)
395 } else if (l == hdrqtail)
398 * Update head regs etc., every 16 packets, if not last pkt,
399 * to help prevent rcvhdrq overflows, when many packets
400 * are processed and queue is nearly full.
401 * Don't request an interrupt for intermediate updates.
404 if (!last && !(i & 0xf)) {
405 dd->f_update_usrhead(rcd, lval, updegr, etail);
414 * Iterate over all QPs waiting to respond.
415 * The list won't change since the IRQ is only run on one CPU.
417 list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
418 list_del_init(&qp->rspwait);
419 if (qp->r_flags & QIB_R_RSP_NAK) {
420 qp->r_flags &= ~QIB_R_RSP_NAK;
423 if (qp->r_flags & QIB_R_RSP_SEND) {
426 qp->r_flags &= ~QIB_R_RSP_SEND;
427 spin_lock_irqsave(&qp->s_lock, flags);
428 if (ib_qib_state_ops[qp->state] &
429 QIB_PROCESS_OR_FLUSH_SEND)
430 qib_schedule_send(qp);
431 spin_unlock_irqrestore(&qp->s_lock, flags);
433 if (atomic_dec_and_test(&qp->refcount))
438 /* Report number of packets consumed */
443 * Always write head at end, and setup rcv interrupt, even
444 * if no packets were processed.
446 lval = (u64)rcd->head | dd->rhdrhead_intr_off;
447 dd->f_update_usrhead(rcd, lval, updegr, etail);
452 * qib_set_mtu - set the MTU
453 * @ppd: the perport data
456 * We can handle "any" incoming size, the issue here is whether we
457 * need to restrict our outgoing size. For now, we don't do any
458 * sanity checking on this, and we don't deal with what happens to
459 * programs that are already running when the size changes.
460 * NOTE: changing the MTU will usually cause the IBC to go back to
463 int qib_set_mtu(struct qib_pportdata *ppd, u16 arg)
468 if (arg != 256 && arg != 512 && arg != 1024 && arg != 2048 &&
473 chk = ib_mtu_enum_to_int(qib_ibmtu);
474 if (chk > 0 && arg > chk) {
479 piosize = ppd->ibmaxlen;
482 if (arg >= (piosize - QIB_PIO_MAXIBHDR)) {
483 /* Only if it's not the initial value (or reset to it) */
484 if (piosize != ppd->init_ibmaxlen) {
485 if (arg > piosize && arg <= ppd->init_ibmaxlen)
486 piosize = ppd->init_ibmaxlen - 2 * sizeof(u32);
487 ppd->ibmaxlen = piosize;
489 } else if ((arg + QIB_PIO_MAXIBHDR) != ppd->ibmaxlen) {
490 piosize = arg + QIB_PIO_MAXIBHDR - 2 * sizeof(u32);
491 ppd->ibmaxlen = piosize;
494 ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_MTU, 0);
502 int qib_set_lid(struct qib_pportdata *ppd, u32 lid, u8 lmc)
504 struct qib_devdata *dd = ppd->dd;
508 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LIDLMC,
509 lid | (~((1U << lmc) - 1)) << 16);
511 qib_devinfo(dd->pcidev, "IB%u:%u got a lid: 0x%x\n",
512 dd->unit, ppd->port, lid);
518 * Following deal with the "obviously simple" task of overriding the state
519 * of the LEDS, which normally indicate link physical and logical status.
520 * The complications arise in dealing with different hardware mappings
521 * and the board-dependent routine being called from interrupts.
522 * and then there's the requirement to _flash_ them.
524 #define LED_OVER_FREQ_SHIFT 8
525 #define LED_OVER_FREQ_MASK (0xFF<<LED_OVER_FREQ_SHIFT)
526 /* Below is "non-zero" to force override, but both actual LEDs are off */
527 #define LED_OVER_BOTH_OFF (8)
529 static void qib_run_led_override(unsigned long opaque)
531 struct qib_pportdata *ppd = (struct qib_pportdata *)opaque;
532 struct qib_devdata *dd = ppd->dd;
536 if (!(dd->flags & QIB_INITTED))
539 ph_idx = ppd->led_override_phase++ & 1;
540 ppd->led_override = ppd->led_override_vals[ph_idx];
541 timeoff = ppd->led_override_timeoff;
543 dd->f_setextled(ppd, 1);
545 * don't re-fire the timer if user asked for it to be off; we let
546 * it fire one more time after they turn it off to simplify
548 if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
549 mod_timer(&ppd->led_override_timer, jiffies + timeoff);
552 void qib_set_led_override(struct qib_pportdata *ppd, unsigned int val)
554 struct qib_devdata *dd = ppd->dd;
557 if (!(dd->flags & QIB_INITTED))
560 /* First check if we are blinking. If not, use 1HZ polling */
562 freq = (val & LED_OVER_FREQ_MASK) >> LED_OVER_FREQ_SHIFT;
565 /* For blink, set each phase from one nybble of val */
566 ppd->led_override_vals[0] = val & 0xF;
567 ppd->led_override_vals[1] = (val >> 4) & 0xF;
568 timeoff = (HZ << 4)/freq;
570 /* Non-blink set both phases the same. */
571 ppd->led_override_vals[0] = val & 0xF;
572 ppd->led_override_vals[1] = val & 0xF;
574 ppd->led_override_timeoff = timeoff;
577 * If the timer has not already been started, do so. Use a "quick"
578 * timeout so the function will be called soon, to look at our request.
580 if (atomic_inc_return(&ppd->led_override_timer_active) == 1) {
581 /* Need to start timer */
582 init_timer(&ppd->led_override_timer);
583 ppd->led_override_timer.function = qib_run_led_override;
584 ppd->led_override_timer.data = (unsigned long) ppd;
585 ppd->led_override_timer.expires = jiffies + 1;
586 add_timer(&ppd->led_override_timer);
588 if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
589 mod_timer(&ppd->led_override_timer, jiffies + 1);
590 atomic_dec(&ppd->led_override_timer_active);
595 * qib_reset_device - reset the chip if possible
596 * @unit: the device to reset
598 * Whether or not reset is successful, we attempt to re-initialize the chip
599 * (that is, much like a driver unload/reload). We clear the INITTED flag
600 * so that the various entry points will fail until we reinitialize. For
601 * now, we only allow this if no user contexts are open that use chip resources
603 int qib_reset_device(int unit)
606 struct qib_devdata *dd = qib_lookup(unit);
607 struct qib_pportdata *ppd;
616 qib_devinfo(dd->pcidev, "Reset on unit %u requested\n", unit);
618 if (!dd->kregbase || !(dd->flags & QIB_PRESENT)) {
619 qib_devinfo(dd->pcidev, "Invalid unit number %u or "
620 "not initialized or not present\n", unit);
625 spin_lock_irqsave(&dd->uctxt_lock, flags);
627 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
628 if (!dd->rcd[i] || !dd->rcd[i]->cnt)
630 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
634 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
636 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
637 ppd = dd->pport + pidx;
638 if (atomic_read(&ppd->led_override_timer_active)) {
639 /* Need to stop LED timer, _then_ shut off LEDs */
640 del_timer_sync(&ppd->led_override_timer);
641 atomic_set(&ppd->led_override_timer_active, 0);
644 /* Shut off LEDs after we are sure timer is not running */
645 ppd->led_override = LED_OVER_BOTH_OFF;
646 dd->f_setextled(ppd, 0);
647 if (dd->flags & QIB_HAS_SEND_DMA)
648 qib_teardown_sdma(ppd);
651 ret = dd->f_reset(dd);
653 ret = qib_init(dd, 1);
657 qib_dev_err(dd, "Reinitialize unit %u after "
658 "reset failed with %d\n", unit, ret);
660 qib_devinfo(dd->pcidev, "Reinitialized unit %u after "
661 "resetting\n", unit);