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>
40 #include <linux/module.h>
45 * The size has to be longer than this string, so we can append
46 * board/chip information to it in the init code.
48 const char ib_qib_version[] = QIB_IDSTR "\n";
50 DEFINE_SPINLOCK(qib_devs_lock);
51 LIST_HEAD(qib_dev_list);
52 DEFINE_MUTEX(qib_mutex); /* general driver use */
55 module_param_named(ibmtu, qib_ibmtu, uint, S_IRUGO);
56 MODULE_PARM_DESC(ibmtu, "Set max IB MTU (0=2KB, 1=256, 2=512, ... 5=4096");
58 unsigned qib_compat_ddr_negotiate = 1;
59 module_param_named(compat_ddr_negotiate, qib_compat_ddr_negotiate, uint,
61 MODULE_PARM_DESC(compat_ddr_negotiate,
62 "Attempt pre-IBTA 1.2 DDR speed negotiation");
64 MODULE_LICENSE("Dual BSD/GPL");
65 MODULE_AUTHOR("QLogic <support@qlogic.com>");
66 MODULE_DESCRIPTION("QLogic IB driver");
69 * QIB_PIO_MAXIBHDR is the max IB header size allowed for in our
70 * PIO send buffers. This is well beyond anything currently
71 * defined in the InfiniBand spec.
73 #define QIB_PIO_MAXIBHDR 128
76 * QIB_MAX_PKT_RCV is the max # if packets processed per receive interrupt.
78 #define QIB_MAX_PKT_RECV 64
80 struct qlogic_ib_stats qib_stats;
82 const char *qib_get_unit_name(int unit)
84 static char iname[16];
86 snprintf(iname, sizeof iname, "infinipath%u", unit);
91 * Return count of units with at least one port ACTIVE.
93 int qib_count_active_units(void)
95 struct qib_devdata *dd;
96 struct qib_pportdata *ppd;
98 int pidx, nunits_active = 0;
100 spin_lock_irqsave(&qib_devs_lock, flags);
101 list_for_each_entry(dd, &qib_dev_list, list) {
102 if (!(dd->flags & QIB_PRESENT) || !dd->kregbase)
104 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
105 ppd = dd->pport + pidx;
106 if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
107 QIBL_LINKARMED | QIBL_LINKACTIVE))) {
113 spin_unlock_irqrestore(&qib_devs_lock, flags);
114 return nunits_active;
118 * Return count of all units, optionally return in arguments
119 * the number of usable (present) units, and the number of
122 int qib_count_units(int *npresentp, int *nupp)
124 int nunits = 0, npresent = 0, nup = 0;
125 struct qib_devdata *dd;
128 struct qib_pportdata *ppd;
130 spin_lock_irqsave(&qib_devs_lock, flags);
132 list_for_each_entry(dd, &qib_dev_list, list) {
134 if ((dd->flags & QIB_PRESENT) && dd->kregbase)
136 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
137 ppd = dd->pport + pidx;
138 if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
139 QIBL_LINKARMED | QIBL_LINKACTIVE)))
144 spin_unlock_irqrestore(&qib_devs_lock, flags);
147 *npresentp = npresent;
155 * qib_wait_linkstate - wait for an IB link state change to occur
156 * @dd: the qlogic_ib device
157 * @state: the state to wait for
158 * @msecs: the number of milliseconds to wait
160 * wait up to msecs milliseconds for IB link state change to occur for
161 * now, take the easy polling route. Currently used only by
162 * qib_set_linkstate. Returns 0 if state reached, otherwise
163 * -ETIMEDOUT state can have multiple states set, for any of several
166 int qib_wait_linkstate(struct qib_pportdata *ppd, u32 state, int msecs)
171 spin_lock_irqsave(&ppd->lflags_lock, flags);
172 if (ppd->state_wanted) {
173 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
177 ppd->state_wanted = state;
178 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
179 wait_event_interruptible_timeout(ppd->state_wait,
180 (ppd->lflags & state),
181 msecs_to_jiffies(msecs));
182 spin_lock_irqsave(&ppd->lflags_lock, flags);
183 ppd->state_wanted = 0;
184 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
186 if (!(ppd->lflags & state))
194 int qib_set_linkstate(struct qib_pportdata *ppd, u8 newstate)
198 struct qib_devdata *dd = ppd->dd;
202 case QIB_IB_LINKDOWN_ONLY:
203 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
204 IB_LINKCMD_DOWN | IB_LINKINITCMD_NOP);
209 case QIB_IB_LINKDOWN:
210 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
211 IB_LINKCMD_DOWN | IB_LINKINITCMD_POLL);
216 case QIB_IB_LINKDOWN_SLEEP:
217 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
218 IB_LINKCMD_DOWN | IB_LINKINITCMD_SLEEP);
223 case QIB_IB_LINKDOWN_DISABLE:
224 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
225 IB_LINKCMD_DOWN | IB_LINKINITCMD_DISABLE);
231 if (ppd->lflags & QIBL_LINKARMED) {
235 if (!(ppd->lflags & (QIBL_LINKINIT | QIBL_LINKACTIVE))) {
240 * Since the port can be ACTIVE when we ask for ARMED,
241 * clear QIBL_LINKV so we can wait for a transition.
242 * If the link isn't ARMED, then something else happened
243 * and there is no point waiting for ARMED.
245 spin_lock_irqsave(&ppd->lflags_lock, flags);
246 ppd->lflags &= ~QIBL_LINKV;
247 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
248 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
249 IB_LINKCMD_ARMED | IB_LINKINITCMD_NOP);
253 case QIB_IB_LINKACTIVE:
254 if (ppd->lflags & QIBL_LINKACTIVE) {
258 if (!(ppd->lflags & QIBL_LINKARMED)) {
262 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
263 IB_LINKCMD_ACTIVE | IB_LINKINITCMD_NOP);
264 lstate = QIBL_LINKACTIVE;
271 ret = qib_wait_linkstate(ppd, lstate, 10);
278 * Get address of eager buffer from it's index (allocated in chunks, not
281 static inline void *qib_get_egrbuf(const struct qib_ctxtdata *rcd, u32 etail)
283 const u32 chunk = etail / rcd->rcvegrbufs_perchunk;
284 const u32 idx = etail % rcd->rcvegrbufs_perchunk;
286 return rcd->rcvegrbuf[chunk] + idx * rcd->dd->rcvegrbufsize;
290 * Returns 1 if error was a CRC, else 0.
291 * Needed for some chip's synthesized error counters.
293 static u32 qib_rcv_hdrerr(struct qib_ctxtdata *rcd, struct qib_pportdata *ppd,
294 u32 ctxt, u32 eflags, u32 l, u32 etail,
295 __le32 *rhf_addr, struct qib_message_header *rhdr)
299 if (eflags & (QLOGIC_IB_RHF_H_ICRCERR | QLOGIC_IB_RHF_H_VCRCERR))
301 else if (eflags == QLOGIC_IB_RHF_H_TIDERR) {
302 /* For TIDERR and RC QPs premptively schedule a NAK */
303 struct qib_ib_header *hdr = (struct qib_ib_header *) rhdr;
304 struct qib_other_headers *ohdr = NULL;
305 struct qib_ibport *ibp = &ppd->ibport_data;
306 struct qib_qp *qp = NULL;
307 u32 tlen = qib_hdrget_length_in_bytes(rhf_addr);
308 u16 lid = be16_to_cpu(hdr->lrh[1]);
309 int lnh = be16_to_cpu(hdr->lrh[0]) & 3;
316 /* Sanity check packet */
320 if (lid < QIB_MULTICAST_LID_BASE) {
321 lid &= ~((1 << ppd->lmc) - 1);
322 if (unlikely(lid != ppd->lid))
327 if (lnh == QIB_LRH_BTH)
329 else if (lnh == QIB_LRH_GRH) {
332 ohdr = &hdr->u.l.oth;
333 if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
335 vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
336 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
341 /* Get opcode and PSN from packet */
342 opcode = be32_to_cpu(ohdr->bth[0]);
344 psn = be32_to_cpu(ohdr->bth[2]);
346 /* Get the destination QP number. */
347 qp_num = be32_to_cpu(ohdr->bth[1]) & QIB_QPN_MASK;
348 if (qp_num != QIB_MULTICAST_QPN) {
350 qp = qib_lookup_qpn(ibp, qp_num);
355 * Handle only RC QPs - for other QP types drop error
358 spin_lock(&qp->r_lock);
360 /* Check for valid receive state. */
361 if (!(ib_qib_state_ops[qp->state] &
362 QIB_PROCESS_RECV_OK)) {
367 switch (qp->ibqp.qp_type) {
369 spin_lock_irqsave(&qp->s_lock, flags);
375 be32_to_cpu(ohdr->bth[0]));
377 spin_unlock_irqrestore(&qp->s_lock,
381 spin_unlock_irqrestore(&qp->s_lock, flags);
383 /* Only deal with RDMA Writes for now */
385 IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
386 diff = qib_cmp24(psn, qp->r_psn);
387 if (!qp->r_nak_state && diff >= 0) {
391 /* Use the expected PSN. */
392 qp->r_ack_psn = qp->r_psn;
394 * Wait to send the sequence
395 * NAK until all packets
396 * in the receive queue have
398 * Otherwise, we end up
399 * propagating congestion.
401 if (list_empty(&qp->rspwait)) {
410 } /* Out of sequence NAK */
411 } /* QP Request NAKs */
418 /* For now don't handle any other QP types */
423 spin_unlock(&qp->r_lock);
425 * Notify qib_destroy_qp() if it is waiting
428 if (atomic_dec_and_test(&qp->refcount))
431 } /* Valid packet with TIDErr */
438 * qib_kreceive - receive a packet
439 * @rcd: the qlogic_ib context
440 * @llic: gets count of good packets needed to clear lli,
441 * (used with chips that need need to track crcs for lli)
443 * called from interrupt handler for errors or receive interrupt
444 * Returns number of CRC error packets, needed by some chips for
445 * local link integrity tracking. crcs are adjusted down by following
446 * good packets, if any, and count of good packets is also tracked.
448 u32 qib_kreceive(struct qib_ctxtdata *rcd, u32 *llic, u32 *npkts)
450 struct qib_devdata *dd = rcd->dd;
451 struct qib_pportdata *ppd = rcd->ppd;
454 const u32 rsize = dd->rcvhdrentsize; /* words */
455 const u32 maxcnt = dd->rcvhdrcnt * rsize; /* words */
456 u32 etail = -1, l, hdrqtail;
457 struct qib_message_header *hdr;
458 u32 eflags, etype, tlen, i = 0, updegr = 0, crcs = 0;
461 struct qib_qp *qp, *nqp;
464 rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
465 if (dd->flags & QIB_NODMA_RTAIL) {
466 u32 seq = qib_hdrget_seq(rhf_addr);
467 if (seq != rcd->seq_cnt)
471 hdrqtail = qib_get_rcvhdrtail(rcd);
474 smp_rmb(); /* prevent speculative reads of dma'ed hdrq */
477 for (last = 0, i = 1; !last; i += !last) {
478 hdr = dd->f_get_msgheader(dd, rhf_addr);
479 eflags = qib_hdrget_err_flags(rhf_addr);
480 etype = qib_hdrget_rcv_type(rhf_addr);
482 tlen = qib_hdrget_length_in_bytes(rhf_addr);
484 if ((dd->flags & QIB_NODMA_RTAIL) ?
485 qib_hdrget_use_egr_buf(rhf_addr) :
486 (etype != RCVHQ_RCV_TYPE_EXPECTED)) {
487 etail = qib_hdrget_index(rhf_addr);
489 if (tlen > sizeof(*hdr) ||
490 etype >= RCVHQ_RCV_TYPE_NON_KD)
491 ebuf = qib_get_egrbuf(rcd, etail);
494 u16 lrh_len = be16_to_cpu(hdr->lrh[2]) << 2;
496 if (lrh_len != tlen) {
497 qib_stats.sps_lenerrs++;
501 if (etype == RCVHQ_RCV_TYPE_NON_KD && !eflags &&
503 tlen > (dd->rcvhdrentsize - 2 + 1 -
504 qib_hdrget_offset(rhf_addr)) << 2) {
509 * Both tiderr and qibhdrerr are set for all plain IB
510 * packets; only qibhdrerr should be set.
512 if (unlikely(eflags))
513 crcs += qib_rcv_hdrerr(rcd, ppd, rcd->ctxt, eflags, l,
514 etail, rhf_addr, hdr);
515 else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
516 qib_ib_rcv(rcd, hdr, ebuf, tlen);
519 else if (llic && *llic)
526 if (i == QIB_MAX_PKT_RECV)
529 rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
530 if (dd->flags & QIB_NODMA_RTAIL) {
531 u32 seq = qib_hdrget_seq(rhf_addr);
533 if (++rcd->seq_cnt > 13)
535 if (seq != rcd->seq_cnt)
537 } else if (l == hdrqtail)
540 * Update head regs etc., every 16 packets, if not last pkt,
541 * to help prevent rcvhdrq overflows, when many packets
542 * are processed and queue is nearly full.
543 * Don't request an interrupt for intermediate updates.
546 if (!last && !(i & 0xf)) {
547 dd->f_update_usrhead(rcd, lval, updegr, etail, i);
556 * Iterate over all QPs waiting to respond.
557 * The list won't change since the IRQ is only run on one CPU.
559 list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
560 list_del_init(&qp->rspwait);
561 if (qp->r_flags & QIB_R_RSP_NAK) {
562 qp->r_flags &= ~QIB_R_RSP_NAK;
565 if (qp->r_flags & QIB_R_RSP_SEND) {
568 qp->r_flags &= ~QIB_R_RSP_SEND;
569 spin_lock_irqsave(&qp->s_lock, flags);
570 if (ib_qib_state_ops[qp->state] &
571 QIB_PROCESS_OR_FLUSH_SEND)
572 qib_schedule_send(qp);
573 spin_unlock_irqrestore(&qp->s_lock, flags);
575 if (atomic_dec_and_test(&qp->refcount))
580 /* Report number of packets consumed */
585 * Always write head at end, and setup rcv interrupt, even
586 * if no packets were processed.
588 lval = (u64)rcd->head | dd->rhdrhead_intr_off;
589 dd->f_update_usrhead(rcd, lval, updegr, etail, i);
594 * qib_set_mtu - set the MTU
595 * @ppd: the perport data
598 * We can handle "any" incoming size, the issue here is whether we
599 * need to restrict our outgoing size. For now, we don't do any
600 * sanity checking on this, and we don't deal with what happens to
601 * programs that are already running when the size changes.
602 * NOTE: changing the MTU will usually cause the IBC to go back to
605 int qib_set_mtu(struct qib_pportdata *ppd, u16 arg)
610 if (arg != 256 && arg != 512 && arg != 1024 && arg != 2048 &&
615 chk = ib_mtu_enum_to_int(qib_ibmtu);
616 if (chk > 0 && arg > chk) {
621 piosize = ppd->ibmaxlen;
624 if (arg >= (piosize - QIB_PIO_MAXIBHDR)) {
625 /* Only if it's not the initial value (or reset to it) */
626 if (piosize != ppd->init_ibmaxlen) {
627 if (arg > piosize && arg <= ppd->init_ibmaxlen)
628 piosize = ppd->init_ibmaxlen - 2 * sizeof(u32);
629 ppd->ibmaxlen = piosize;
631 } else if ((arg + QIB_PIO_MAXIBHDR) != ppd->ibmaxlen) {
632 piosize = arg + QIB_PIO_MAXIBHDR - 2 * sizeof(u32);
633 ppd->ibmaxlen = piosize;
636 ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_MTU, 0);
644 int qib_set_lid(struct qib_pportdata *ppd, u32 lid, u8 lmc)
646 struct qib_devdata *dd = ppd->dd;
650 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LIDLMC,
651 lid | (~((1U << lmc) - 1)) << 16);
653 qib_devinfo(dd->pcidev, "IB%u:%u got a lid: 0x%x\n",
654 dd->unit, ppd->port, lid);
660 * Following deal with the "obviously simple" task of overriding the state
661 * of the LEDS, which normally indicate link physical and logical status.
662 * The complications arise in dealing with different hardware mappings
663 * and the board-dependent routine being called from interrupts.
664 * and then there's the requirement to _flash_ them.
666 #define LED_OVER_FREQ_SHIFT 8
667 #define LED_OVER_FREQ_MASK (0xFF<<LED_OVER_FREQ_SHIFT)
668 /* Below is "non-zero" to force override, but both actual LEDs are off */
669 #define LED_OVER_BOTH_OFF (8)
671 static void qib_run_led_override(unsigned long opaque)
673 struct qib_pportdata *ppd = (struct qib_pportdata *)opaque;
674 struct qib_devdata *dd = ppd->dd;
678 if (!(dd->flags & QIB_INITTED))
681 ph_idx = ppd->led_override_phase++ & 1;
682 ppd->led_override = ppd->led_override_vals[ph_idx];
683 timeoff = ppd->led_override_timeoff;
685 dd->f_setextled(ppd, 1);
687 * don't re-fire the timer if user asked for it to be off; we let
688 * it fire one more time after they turn it off to simplify
690 if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
691 mod_timer(&ppd->led_override_timer, jiffies + timeoff);
694 void qib_set_led_override(struct qib_pportdata *ppd, unsigned int val)
696 struct qib_devdata *dd = ppd->dd;
699 if (!(dd->flags & QIB_INITTED))
702 /* First check if we are blinking. If not, use 1HZ polling */
704 freq = (val & LED_OVER_FREQ_MASK) >> LED_OVER_FREQ_SHIFT;
707 /* For blink, set each phase from one nybble of val */
708 ppd->led_override_vals[0] = val & 0xF;
709 ppd->led_override_vals[1] = (val >> 4) & 0xF;
710 timeoff = (HZ << 4)/freq;
712 /* Non-blink set both phases the same. */
713 ppd->led_override_vals[0] = val & 0xF;
714 ppd->led_override_vals[1] = val & 0xF;
716 ppd->led_override_timeoff = timeoff;
719 * If the timer has not already been started, do so. Use a "quick"
720 * timeout so the function will be called soon, to look at our request.
722 if (atomic_inc_return(&ppd->led_override_timer_active) == 1) {
723 /* Need to start timer */
724 init_timer(&ppd->led_override_timer);
725 ppd->led_override_timer.function = qib_run_led_override;
726 ppd->led_override_timer.data = (unsigned long) ppd;
727 ppd->led_override_timer.expires = jiffies + 1;
728 add_timer(&ppd->led_override_timer);
730 if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
731 mod_timer(&ppd->led_override_timer, jiffies + 1);
732 atomic_dec(&ppd->led_override_timer_active);
737 * qib_reset_device - reset the chip if possible
738 * @unit: the device to reset
740 * Whether or not reset is successful, we attempt to re-initialize the chip
741 * (that is, much like a driver unload/reload). We clear the INITTED flag
742 * so that the various entry points will fail until we reinitialize. For
743 * now, we only allow this if no user contexts are open that use chip resources
745 int qib_reset_device(int unit)
748 struct qib_devdata *dd = qib_lookup(unit);
749 struct qib_pportdata *ppd;
758 qib_devinfo(dd->pcidev, "Reset on unit %u requested\n", unit);
760 if (!dd->kregbase || !(dd->flags & QIB_PRESENT)) {
761 qib_devinfo(dd->pcidev, "Invalid unit number %u or "
762 "not initialized or not present\n", unit);
767 spin_lock_irqsave(&dd->uctxt_lock, flags);
769 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
770 if (!dd->rcd[i] || !dd->rcd[i]->cnt)
772 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
776 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
778 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
779 ppd = dd->pport + pidx;
780 if (atomic_read(&ppd->led_override_timer_active)) {
781 /* Need to stop LED timer, _then_ shut off LEDs */
782 del_timer_sync(&ppd->led_override_timer);
783 atomic_set(&ppd->led_override_timer_active, 0);
786 /* Shut off LEDs after we are sure timer is not running */
787 ppd->led_override = LED_OVER_BOTH_OFF;
788 dd->f_setextled(ppd, 0);
789 if (dd->flags & QIB_HAS_SEND_DMA)
790 qib_teardown_sdma(ppd);
793 ret = dd->f_reset(dd);
795 ret = qib_init(dd, 1);
799 qib_dev_err(dd, "Reinitialize unit %u after "
800 "reset failed with %d\n", unit, ret);
802 qib_devinfo(dd->pcidev, "Reinitialized unit %u after "
803 "resetting\n", unit);