2 * Copyright(c) 2015, 2016 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
48 #include <linux/spinlock.h>
49 #include <linux/pci.h>
51 #include <linux/delay.h>
52 #include <linux/netdevice.h>
53 #include <linux/vmalloc.h>
54 #include <linux/module.h>
55 #include <linux/prefetch.h>
56 #include <rdma/ib_verbs.h>
64 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
67 * The size has to be longer than this string, so we can append
68 * board/chip information to it in the initialization code.
70 const char ib_hfi1_version[] = HFI1_DRIVER_VERSION "\n";
72 DEFINE_SPINLOCK(hfi1_devs_lock);
73 LIST_HEAD(hfi1_dev_list);
74 DEFINE_MUTEX(hfi1_mutex); /* general driver use */
76 unsigned int hfi1_max_mtu = HFI1_DEFAULT_MAX_MTU;
77 module_param_named(max_mtu, hfi1_max_mtu, uint, S_IRUGO);
78 MODULE_PARM_DESC(max_mtu, "Set max MTU bytes, default is " __stringify(
79 HFI1_DEFAULT_MAX_MTU));
81 unsigned int hfi1_cu = 1;
82 module_param_named(cu, hfi1_cu, uint, S_IRUGO);
83 MODULE_PARM_DESC(cu, "Credit return units");
85 unsigned long hfi1_cap_mask = HFI1_CAP_MASK_DEFAULT;
86 static int hfi1_caps_set(const char *, const struct kernel_param *);
87 static int hfi1_caps_get(char *, const struct kernel_param *);
88 static const struct kernel_param_ops cap_ops = {
92 module_param_cb(cap_mask, &cap_ops, &hfi1_cap_mask, S_IWUSR | S_IRUGO);
93 MODULE_PARM_DESC(cap_mask, "Bit mask of enabled/disabled HW features");
95 MODULE_LICENSE("Dual BSD/GPL");
96 MODULE_DESCRIPTION("Intel Omni-Path Architecture driver");
97 MODULE_VERSION(HFI1_DRIVER_VERSION);
100 * MAX_PKT_RCV is the max # if packets processed per receive interrupt.
102 #define MAX_PKT_RECV 64
103 #define EGR_HEAD_UPDATE_THRESHOLD 16
105 struct hfi1_ib_stats hfi1_stats;
107 static int hfi1_caps_set(const char *val, const struct kernel_param *kp)
110 unsigned long *cap_mask_ptr = (unsigned long *)kp->arg,
111 cap_mask = *cap_mask_ptr, value, diff,
112 write_mask = ((HFI1_CAP_WRITABLE_MASK << HFI1_CAP_USER_SHIFT) |
113 HFI1_CAP_WRITABLE_MASK);
115 ret = kstrtoul(val, 0, &value);
117 pr_warn("Invalid module parameter value for 'cap_mask'\n");
120 /* Get the changed bits (except the locked bit) */
121 diff = value ^ (cap_mask & ~HFI1_CAP_LOCKED_SMASK);
123 /* Remove any bits that are not allowed to change after driver load */
124 if (HFI1_CAP_LOCKED() && (diff & ~write_mask)) {
125 pr_warn("Ignoring non-writable capability bits %#lx\n",
130 /* Mask off any reserved bits */
131 diff &= ~HFI1_CAP_RESERVED_MASK;
132 /* Clear any previously set and changing bits */
134 /* Update the bits with the new capability */
135 cap_mask |= (value & diff);
136 /* Check for any kernel/user restrictions */
137 diff = (cap_mask & (HFI1_CAP_MUST_HAVE_KERN << HFI1_CAP_USER_SHIFT)) ^
138 ((cap_mask & HFI1_CAP_MUST_HAVE_KERN) << HFI1_CAP_USER_SHIFT);
140 /* Set the bitmask to the final set */
141 *cap_mask_ptr = cap_mask;
146 static int hfi1_caps_get(char *buffer, const struct kernel_param *kp)
148 unsigned long cap_mask = *(unsigned long *)kp->arg;
150 cap_mask &= ~HFI1_CAP_LOCKED_SMASK;
151 cap_mask |= ((cap_mask & HFI1_CAP_K2U) << HFI1_CAP_USER_SHIFT);
153 return scnprintf(buffer, PAGE_SIZE, "0x%lx", cap_mask);
156 const char *get_unit_name(int unit)
158 static char iname[16];
160 snprintf(iname, sizeof(iname), DRIVER_NAME "_%u", unit);
164 const char *get_card_name(struct rvt_dev_info *rdi)
166 struct hfi1_ibdev *ibdev = container_of(rdi, struct hfi1_ibdev, rdi);
167 struct hfi1_devdata *dd = container_of(ibdev,
168 struct hfi1_devdata, verbs_dev);
169 return get_unit_name(dd->unit);
172 struct pci_dev *get_pci_dev(struct rvt_dev_info *rdi)
174 struct hfi1_ibdev *ibdev = container_of(rdi, struct hfi1_ibdev, rdi);
175 struct hfi1_devdata *dd = container_of(ibdev,
176 struct hfi1_devdata, verbs_dev);
181 * Return count of units with at least one port ACTIVE.
183 int hfi1_count_active_units(void)
185 struct hfi1_devdata *dd;
186 struct hfi1_pportdata *ppd;
188 int pidx, nunits_active = 0;
190 spin_lock_irqsave(&hfi1_devs_lock, flags);
191 list_for_each_entry(dd, &hfi1_dev_list, list) {
192 if (!(dd->flags & HFI1_PRESENT) || !dd->kregbase)
194 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
195 ppd = dd->pport + pidx;
196 if (ppd->lid && ppd->linkup) {
202 spin_unlock_irqrestore(&hfi1_devs_lock, flags);
203 return nunits_active;
207 * Return count of all units, optionally return in arguments
208 * the number of usable (present) units, and the number of
211 int hfi1_count_units(int *npresentp, int *nupp)
213 int nunits = 0, npresent = 0, nup = 0;
214 struct hfi1_devdata *dd;
217 struct hfi1_pportdata *ppd;
219 spin_lock_irqsave(&hfi1_devs_lock, flags);
221 list_for_each_entry(dd, &hfi1_dev_list, list) {
223 if ((dd->flags & HFI1_PRESENT) && dd->kregbase)
225 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
226 ppd = dd->pport + pidx;
227 if (ppd->lid && ppd->linkup)
232 spin_unlock_irqrestore(&hfi1_devs_lock, flags);
235 *npresentp = npresent;
243 * Get address of eager buffer from it's index (allocated in chunks, not
246 static inline void *get_egrbuf(const struct hfi1_ctxtdata *rcd, u64 rhf,
249 u32 idx = rhf_egr_index(rhf), offset = rhf_egr_buf_offset(rhf);
251 *update |= !(idx & (rcd->egrbufs.threshold - 1)) && !offset;
252 return (void *)(((u64)(rcd->egrbufs.rcvtids[idx].addr)) +
253 (offset * RCV_BUF_BLOCK_SIZE));
257 * Validate and encode the a given RcvArray Buffer size.
258 * The function will check whether the given size falls within
259 * allowed size ranges for the respective type and, optionally,
260 * return the proper encoding.
262 inline int hfi1_rcvbuf_validate(u32 size, u8 type, u16 *encoded)
264 if (unlikely(!PAGE_ALIGNED(size)))
266 if (unlikely(size < MIN_EAGER_BUFFER))
269 (type == PT_EAGER ? MAX_EAGER_BUFFER : MAX_EXPECTED_BUFFER))
272 *encoded = ilog2(size / PAGE_SIZE) + 1;
276 static void rcv_hdrerr(struct hfi1_ctxtdata *rcd, struct hfi1_pportdata *ppd,
277 struct hfi1_packet *packet)
279 struct hfi1_message_header *rhdr = packet->hdr;
280 u32 rte = rhf_rcv_type_err(packet->rhf);
281 int lnh = be16_to_cpu(rhdr->lrh[0]) & 3;
282 struct hfi1_ibport *ibp = &ppd->ibport_data;
283 struct hfi1_devdata *dd = ppd->dd;
284 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
286 if (packet->rhf & (RHF_VCRC_ERR | RHF_ICRC_ERR))
289 if (packet->rhf & RHF_TID_ERR) {
290 /* For TIDERR and RC QPs preemptively schedule a NAK */
291 struct hfi1_ib_header *hdr = (struct hfi1_ib_header *)rhdr;
292 struct hfi1_other_headers *ohdr = NULL;
293 u32 tlen = rhf_pkt_len(packet->rhf); /* in bytes */
294 u16 lid = be16_to_cpu(hdr->lrh[1]);
298 /* Sanity check packet */
303 if (lnh == HFI1_LRH_BTH) {
305 } else if (lnh == HFI1_LRH_GRH) {
308 ohdr = &hdr->u.l.oth;
309 if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
311 vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
312 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
314 rcv_flags |= HFI1_HAS_GRH;
318 /* Get the destination QP number. */
319 qp_num = be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK;
320 if (lid < be16_to_cpu(IB_MULTICAST_LID_BASE)) {
325 qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
332 * Handle only RC QPs - for other QP types drop error
335 spin_lock_irqsave(&qp->r_lock, flags);
337 /* Check for valid receive state. */
338 if (!(ib_rvt_state_ops[qp->state] &
339 RVT_PROCESS_RECV_OK)) {
340 ibp->rvp.n_pkt_drops++;
343 switch (qp->ibqp.qp_type) {
352 /* For now don't handle any other QP types */
356 spin_unlock_irqrestore(&qp->r_lock, flags);
359 } /* Valid packet with TIDErr */
361 /* handle "RcvTypeErr" flags */
363 case RHF_RTE_ERROR_OP_CODE_ERR:
369 if (rhf_use_egr_bfr(packet->rhf))
373 goto drop; /* this should never happen */
375 if (lnh == HFI1_LRH_BTH)
376 bth = (__be32 *)ebuf;
377 else if (lnh == HFI1_LRH_GRH)
378 bth = (__be32 *)((char *)ebuf + sizeof(struct ib_grh));
382 opcode = be32_to_cpu(bth[0]) >> 24;
385 if (opcode == IB_OPCODE_CNP) {
387 * Only in pre-B0 h/w is the CNP_OPCODE handled
388 * via this code path.
390 struct rvt_qp *qp = NULL;
393 u8 svc_type, sl, sc5;
395 sc5 = (be16_to_cpu(rhdr->lrh[0]) >> 12) & 0xf;
396 if (rhf_dc_info(packet->rhf))
398 sl = ibp->sc_to_sl[sc5];
400 lqpn = be32_to_cpu(bth[1]) & RVT_QPN_MASK;
402 qp = rvt_lookup_qpn(rdi, &ibp->rvp, lqpn);
408 switch (qp->ibqp.qp_type) {
412 svc_type = IB_CC_SVCTYPE_UD;
415 rlid = be16_to_cpu(rhdr->lrh[3]);
416 rqpn = qp->remote_qpn;
417 svc_type = IB_CC_SVCTYPE_UC;
423 process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
427 packet->rhf &= ~RHF_RCV_TYPE_ERR_SMASK;
438 static inline void init_packet(struct hfi1_ctxtdata *rcd,
439 struct hfi1_packet *packet)
441 packet->rsize = rcd->rcvhdrqentsize; /* words */
442 packet->maxcnt = rcd->rcvhdrq_cnt * packet->rsize; /* words */
446 packet->rhf_addr = get_rhf_addr(rcd);
447 packet->rhf = rhf_to_cpu(packet->rhf_addr);
448 packet->rhqoff = rcd->head;
450 packet->rcv_flags = 0;
453 static void process_ecn(struct rvt_qp *qp, struct hfi1_ib_header *hdr,
454 struct hfi1_other_headers *ohdr,
455 u64 rhf, u32 bth1, struct ib_grh *grh)
457 struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
462 switch (qp->ibqp.qp_type) {
466 rlid = be16_to_cpu(hdr->lrh[3]);
467 rqpn = be32_to_cpu(ohdr->u.ud.deth[1]) & RVT_QPN_MASK;
468 svc_type = IB_CC_SVCTYPE_UD;
471 rlid = qp->remote_ah_attr.dlid;
472 rqpn = qp->remote_qpn;
473 svc_type = IB_CC_SVCTYPE_UC;
476 rlid = qp->remote_ah_attr.dlid;
477 rqpn = qp->remote_qpn;
478 svc_type = IB_CC_SVCTYPE_RC;
484 sc5 = (be16_to_cpu(hdr->lrh[0]) >> 12) & 0xf;
485 if (rhf_dc_info(rhf))
488 if (bth1 & HFI1_FECN_SMASK) {
489 u16 pkey = (u16)be32_to_cpu(ohdr->bth[0]);
490 u16 dlid = be16_to_cpu(hdr->lrh[1]);
492 return_cnp(ibp, qp, rqpn, pkey, dlid, rlid, sc5, grh);
495 if (bth1 & HFI1_BECN_SMASK) {
496 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
497 u32 lqpn = bth1 & RVT_QPN_MASK;
498 u8 sl = ibp->sc_to_sl[sc5];
500 process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
505 struct hfi1_ctxtdata *rcd;
513 static inline void init_ps_mdata(struct ps_mdata *mdata,
514 struct hfi1_packet *packet)
516 struct hfi1_ctxtdata *rcd = packet->rcd;
519 mdata->rsize = packet->rsize;
520 mdata->maxcnt = packet->maxcnt;
521 mdata->ps_head = packet->rhqoff;
523 if (HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
524 mdata->ps_tail = get_rcvhdrtail(rcd);
525 if (rcd->ctxt == HFI1_CTRL_CTXT)
526 mdata->ps_seq = rcd->seq_cnt;
528 mdata->ps_seq = 0; /* not used with DMA_RTAIL */
530 mdata->ps_tail = 0; /* used only with DMA_RTAIL*/
531 mdata->ps_seq = rcd->seq_cnt;
535 static inline int ps_done(struct ps_mdata *mdata, u64 rhf,
536 struct hfi1_ctxtdata *rcd)
538 if (HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL))
539 return mdata->ps_head == mdata->ps_tail;
540 return mdata->ps_seq != rhf_rcv_seq(rhf);
543 static inline int ps_skip(struct ps_mdata *mdata, u64 rhf,
544 struct hfi1_ctxtdata *rcd)
547 * Control context can potentially receive an invalid rhf.
550 if ((rcd->ctxt == HFI1_CTRL_CTXT) && (mdata->ps_head != mdata->ps_tail))
551 return mdata->ps_seq != rhf_rcv_seq(rhf);
556 static inline void update_ps_mdata(struct ps_mdata *mdata,
557 struct hfi1_ctxtdata *rcd)
559 mdata->ps_head += mdata->rsize;
560 if (mdata->ps_head >= mdata->maxcnt)
563 /* Control context must do seq counting */
564 if (!HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL) ||
565 (rcd->ctxt == HFI1_CTRL_CTXT)) {
566 if (++mdata->ps_seq > 13)
572 * prescan_rxq - search through the receive queue looking for packets
573 * containing Excplicit Congestion Notifications (FECNs, or BECNs).
574 * When an ECN is found, process the Congestion Notification, and toggle
576 * This is declared as a macro to allow quick checking of the port to avoid
577 * the overhead of a function call if not enabled.
579 #define prescan_rxq(rcd, packet) \
581 if (rcd->ppd->cc_prescan) \
582 __prescan_rxq(packet); \
584 static void __prescan_rxq(struct hfi1_packet *packet)
586 struct hfi1_ctxtdata *rcd = packet->rcd;
587 struct ps_mdata mdata;
589 init_ps_mdata(&mdata, packet);
592 struct hfi1_devdata *dd = rcd->dd;
593 struct hfi1_ibport *ibp = &rcd->ppd->ibport_data;
594 __le32 *rhf_addr = (__le32 *)rcd->rcvhdrq + mdata.ps_head +
597 struct hfi1_ib_header *hdr;
598 struct hfi1_other_headers *ohdr;
599 struct ib_grh *grh = NULL;
600 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
601 u64 rhf = rhf_to_cpu(rhf_addr);
602 u32 etype = rhf_rcv_type(rhf), qpn, bth1;
606 if (ps_done(&mdata, rhf, rcd))
609 if (ps_skip(&mdata, rhf, rcd))
612 if (etype != RHF_RCV_TYPE_IB)
615 hdr = (struct hfi1_ib_header *)
616 hfi1_get_msgheader(dd, rhf_addr);
617 lnh = be16_to_cpu(hdr->lrh[0]) & 3;
619 if (lnh == HFI1_LRH_BTH) {
621 } else if (lnh == HFI1_LRH_GRH) {
622 ohdr = &hdr->u.l.oth;
625 goto next; /* just in case */
627 bth1 = be32_to_cpu(ohdr->bth[1]);
628 is_ecn = !!(bth1 & (HFI1_FECN_SMASK | HFI1_BECN_SMASK));
633 qpn = bth1 & RVT_QPN_MASK;
635 qp = rvt_lookup_qpn(rdi, &ibp->rvp, qpn);
642 process_ecn(qp, hdr, ohdr, rhf, bth1, grh);
645 /* turn off BECN, FECN */
646 bth1 &= ~(HFI1_FECN_SMASK | HFI1_BECN_SMASK);
647 ohdr->bth[1] = cpu_to_be32(bth1);
649 update_ps_mdata(&mdata, rcd);
653 static inline int skip_rcv_packet(struct hfi1_packet *packet, int thread)
655 int ret = RCV_PKT_OK;
657 /* Set up for the next packet */
658 packet->rhqoff += packet->rsize;
659 if (packet->rhqoff >= packet->maxcnt)
663 if (unlikely((packet->numpkt & (MAX_PKT_RECV - 1)) == 0)) {
668 this_cpu_inc(*packet->rcd->dd->rcv_limit);
672 packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff +
673 packet->rcd->dd->rhf_offset;
674 packet->rhf = rhf_to_cpu(packet->rhf_addr);
679 static inline int process_rcv_packet(struct hfi1_packet *packet, int thread)
681 int ret = RCV_PKT_OK;
683 packet->hdr = hfi1_get_msgheader(packet->rcd->dd,
685 packet->hlen = (u8 *)packet->rhf_addr - (u8 *)packet->hdr;
686 packet->etype = rhf_rcv_type(packet->rhf);
688 packet->tlen = rhf_pkt_len(packet->rhf); /* in bytes */
689 /* retrieve eager buffer details */
691 if (rhf_use_egr_bfr(packet->rhf)) {
692 packet->etail = rhf_egr_index(packet->rhf);
693 packet->ebuf = get_egrbuf(packet->rcd, packet->rhf,
696 * Prefetch the contents of the eager buffer. It is
697 * OK to send a negative length to prefetch_range().
698 * The +2 is the size of the RHF.
700 prefetch_range(packet->ebuf,
701 packet->tlen - ((packet->rcd->rcvhdrqentsize -
702 (rhf_hdrq_offset(packet->rhf)
707 * Call a type specific handler for the packet. We
708 * should be able to trust that etype won't be beyond
709 * the range of valid indexes. If so something is really
710 * wrong and we can probably just let things come
711 * crashing down. There is no need to eat another
712 * comparison in this performance critical code.
714 packet->rcd->dd->rhf_rcv_function_map[packet->etype](packet);
717 /* Set up for the next packet */
718 packet->rhqoff += packet->rsize;
719 if (packet->rhqoff >= packet->maxcnt)
722 if (unlikely((packet->numpkt & (MAX_PKT_RECV - 1)) == 0)) {
727 this_cpu_inc(*packet->rcd->dd->rcv_limit);
731 packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff +
732 packet->rcd->dd->rhf_offset;
733 packet->rhf = rhf_to_cpu(packet->rhf_addr);
738 static inline void process_rcv_update(int last, struct hfi1_packet *packet)
741 * Update head regs etc., every 16 packets, if not last pkt,
742 * to help prevent rcvhdrq overflows, when many packets
743 * are processed and queue is nearly full.
744 * Don't request an interrupt for intermediate updates.
746 if (!last && !(packet->numpkt & 0xf)) {
747 update_usrhead(packet->rcd, packet->rhqoff, packet->updegr,
748 packet->etail, 0, 0);
751 packet->rcv_flags = 0;
754 static inline void finish_packet(struct hfi1_packet *packet)
757 * Nothing we need to free for the packet.
759 * The only thing we need to do is a final update and call for an
762 update_usrhead(packet->rcd, packet->rcd->head, packet->updegr,
763 packet->etail, rcv_intr_dynamic, packet->numpkt);
766 static inline void process_rcv_qp_work(struct hfi1_packet *packet)
768 struct hfi1_ctxtdata *rcd;
769 struct rvt_qp *qp, *nqp;
772 rcd->head = packet->rhqoff;
775 * Iterate over all QPs waiting to respond.
776 * The list won't change since the IRQ is only run on one CPU.
778 list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
779 list_del_init(&qp->rspwait);
780 if (qp->r_flags & RVT_R_RSP_NAK) {
781 qp->r_flags &= ~RVT_R_RSP_NAK;
782 hfi1_send_rc_ack(rcd, qp, 0);
784 if (qp->r_flags & RVT_R_RSP_SEND) {
787 qp->r_flags &= ~RVT_R_RSP_SEND;
788 spin_lock_irqsave(&qp->s_lock, flags);
789 if (ib_rvt_state_ops[qp->state] &
790 RVT_PROCESS_OR_FLUSH_SEND)
791 hfi1_schedule_send(qp);
792 spin_unlock_irqrestore(&qp->s_lock, flags);
794 if (atomic_dec_and_test(&qp->refcount))
800 * Handle receive interrupts when using the no dma rtail option.
802 int handle_receive_interrupt_nodma_rtail(struct hfi1_ctxtdata *rcd, int thread)
805 int last = RCV_PKT_OK;
806 struct hfi1_packet packet;
808 init_packet(rcd, &packet);
809 seq = rhf_rcv_seq(packet.rhf);
810 if (seq != rcd->seq_cnt) {
815 prescan_rxq(rcd, &packet);
817 while (last == RCV_PKT_OK) {
818 last = process_rcv_packet(&packet, thread);
819 seq = rhf_rcv_seq(packet.rhf);
820 if (++rcd->seq_cnt > 13)
822 if (seq != rcd->seq_cnt)
824 process_rcv_update(last, &packet);
826 process_rcv_qp_work(&packet);
828 finish_packet(&packet);
832 int handle_receive_interrupt_dma_rtail(struct hfi1_ctxtdata *rcd, int thread)
835 int last = RCV_PKT_OK;
836 struct hfi1_packet packet;
838 init_packet(rcd, &packet);
839 hdrqtail = get_rcvhdrtail(rcd);
840 if (packet.rhqoff == hdrqtail) {
844 smp_rmb(); /* prevent speculative reads of dma'ed hdrq */
846 prescan_rxq(rcd, &packet);
848 while (last == RCV_PKT_OK) {
849 last = process_rcv_packet(&packet, thread);
850 if (packet.rhqoff == hdrqtail)
852 process_rcv_update(last, &packet);
854 process_rcv_qp_work(&packet);
856 finish_packet(&packet);
860 static inline void set_all_nodma_rtail(struct hfi1_devdata *dd)
864 for (i = HFI1_CTRL_CTXT + 1; i < dd->first_user_ctxt; i++)
865 dd->rcd[i]->do_interrupt =
866 &handle_receive_interrupt_nodma_rtail;
869 static inline void set_all_dma_rtail(struct hfi1_devdata *dd)
873 for (i = HFI1_CTRL_CTXT + 1; i < dd->first_user_ctxt; i++)
874 dd->rcd[i]->do_interrupt =
875 &handle_receive_interrupt_dma_rtail;
878 void set_all_slowpath(struct hfi1_devdata *dd)
882 /* HFI1_CTRL_CTXT must always use the slow path interrupt handler */
883 for (i = HFI1_CTRL_CTXT + 1; i < dd->first_user_ctxt; i++)
884 dd->rcd[i]->do_interrupt = &handle_receive_interrupt;
887 static inline int set_armed_to_active(struct hfi1_ctxtdata *rcd,
888 struct hfi1_packet packet,
889 struct hfi1_devdata *dd)
891 struct work_struct *lsaw = &rcd->ppd->linkstate_active_work;
892 struct hfi1_message_header *hdr = hfi1_get_msgheader(packet.rcd->dd,
895 if (hdr2sc(hdr, packet.rhf) != 0xf) {
896 int hwstate = read_logical_state(dd);
898 if (hwstate != LSTATE_ACTIVE) {
899 dd_dev_info(dd, "Unexpected link state %d\n", hwstate);
903 queue_work(rcd->ppd->hfi1_wq, lsaw);
910 * handle_receive_interrupt - receive a packet
913 * Called from interrupt handler for errors or receive interrupt.
914 * This is the slow path interrupt handler.
916 int handle_receive_interrupt(struct hfi1_ctxtdata *rcd, int thread)
918 struct hfi1_devdata *dd = rcd->dd;
920 int needset, last = RCV_PKT_OK;
921 struct hfi1_packet packet;
924 /* Control context will always use the slow path interrupt handler */
925 needset = (rcd->ctxt == HFI1_CTRL_CTXT) ? 0 : 1;
927 init_packet(rcd, &packet);
929 if (!HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
930 u32 seq = rhf_rcv_seq(packet.rhf);
932 if (seq != rcd->seq_cnt) {
938 hdrqtail = get_rcvhdrtail(rcd);
939 if (packet.rhqoff == hdrqtail) {
943 smp_rmb(); /* prevent speculative reads of dma'ed hdrq */
946 * Control context can potentially receive an invalid
947 * rhf. Drop such packets.
949 if (rcd->ctxt == HFI1_CTRL_CTXT) {
950 u32 seq = rhf_rcv_seq(packet.rhf);
952 if (seq != rcd->seq_cnt)
957 prescan_rxq(rcd, &packet);
959 while (last == RCV_PKT_OK) {
960 if (unlikely(dd->do_drop &&
961 atomic_xchg(&dd->drop_packet, DROP_PACKET_OFF) ==
965 /* On to the next packet */
966 packet.rhqoff += packet.rsize;
967 packet.rhf_addr = (__le32 *)rcd->rcvhdrq +
970 packet.rhf = rhf_to_cpu(packet.rhf_addr);
972 } else if (skip_pkt) {
973 last = skip_rcv_packet(&packet, thread);
976 /* Auto activate link on non-SC15 packet receive */
977 if (unlikely(rcd->ppd->host_link_state ==
979 set_armed_to_active(rcd, packet, dd))
981 last = process_rcv_packet(&packet, thread);
984 if (!HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
985 u32 seq = rhf_rcv_seq(packet.rhf);
987 if (++rcd->seq_cnt > 13)
989 if (seq != rcd->seq_cnt)
992 dd_dev_info(dd, "Switching to NO_DMA_RTAIL\n");
993 set_all_nodma_rtail(dd);
997 if (packet.rhqoff == hdrqtail)
1000 * Control context can potentially receive an invalid
1001 * rhf. Drop such packets.
1003 if (rcd->ctxt == HFI1_CTRL_CTXT) {
1004 u32 seq = rhf_rcv_seq(packet.rhf);
1006 if (++rcd->seq_cnt > 13)
1008 if (!last && (seq != rcd->seq_cnt))
1014 "Switching to DMA_RTAIL\n");
1015 set_all_dma_rtail(dd);
1020 process_rcv_update(last, &packet);
1023 process_rcv_qp_work(&packet);
1027 * Always write head at end, and setup rcv interrupt, even
1028 * if no packets were processed.
1030 finish_packet(&packet);
1035 * We may discover in the interrupt that the hardware link state has
1036 * changed from ARMED to ACTIVE (due to the arrival of a non-SC15 packet),
1037 * and we need to update the driver's notion of the link state. We cannot
1038 * run set_link_state from interrupt context, so we queue this function on
1041 * We delay the regular interrupt processing until after the state changes
1042 * so that the link will be in the correct state by the time any application
1043 * we wake up attempts to send a reply to any message it received.
1044 * (Subsequent receive interrupts may possibly force the wakeup before we
1045 * update the link state.)
1047 * The rcd is freed in hfi1_free_ctxtdata after hfi1_postinit_cleanup invokes
1048 * dd->f_cleanup(dd) to disable the interrupt handler and flush workqueues,
1049 * so we're safe from use-after-free of the rcd.
1051 void receive_interrupt_work(struct work_struct *work)
1053 struct hfi1_pportdata *ppd = container_of(work, struct hfi1_pportdata,
1054 linkstate_active_work);
1055 struct hfi1_devdata *dd = ppd->dd;
1058 /* Received non-SC15 packet implies neighbor_normal */
1059 ppd->neighbor_normal = 1;
1060 set_link_state(ppd, HLS_UP_ACTIVE);
1063 * Interrupt all kernel contexts that could have had an
1064 * interrupt during auto activation.
1066 for (i = HFI1_CTRL_CTXT; i < dd->first_user_ctxt; i++)
1067 force_recv_intr(dd->rcd[i]);
1071 * Convert a given MTU size to the on-wire MAD packet enumeration.
1072 * Return -1 if the size is invalid.
1074 int mtu_to_enum(u32 mtu, int default_if_bad)
1077 case 0: return OPA_MTU_0;
1078 case 256: return OPA_MTU_256;
1079 case 512: return OPA_MTU_512;
1080 case 1024: return OPA_MTU_1024;
1081 case 2048: return OPA_MTU_2048;
1082 case 4096: return OPA_MTU_4096;
1083 case 8192: return OPA_MTU_8192;
1084 case 10240: return OPA_MTU_10240;
1086 return default_if_bad;
1089 u16 enum_to_mtu(int mtu)
1092 case OPA_MTU_0: return 0;
1093 case OPA_MTU_256: return 256;
1094 case OPA_MTU_512: return 512;
1095 case OPA_MTU_1024: return 1024;
1096 case OPA_MTU_2048: return 2048;
1097 case OPA_MTU_4096: return 4096;
1098 case OPA_MTU_8192: return 8192;
1099 case OPA_MTU_10240: return 10240;
1100 default: return 0xffff;
1105 * set_mtu - set the MTU
1106 * @ppd: the per port data
1108 * We can handle "any" incoming size, the issue here is whether we
1109 * need to restrict our outgoing size. We do not deal with what happens
1110 * to programs that are already running when the size changes.
1112 int set_mtu(struct hfi1_pportdata *ppd)
1114 struct hfi1_devdata *dd = ppd->dd;
1115 int i, drain, ret = 0, is_up = 0;
1118 for (i = 0; i < ppd->vls_supported; i++)
1119 if (ppd->ibmtu < dd->vld[i].mtu)
1120 ppd->ibmtu = dd->vld[i].mtu;
1121 ppd->ibmaxlen = ppd->ibmtu + lrh_max_header_bytes(ppd->dd);
1123 mutex_lock(&ppd->hls_lock);
1124 if (ppd->host_link_state == HLS_UP_INIT ||
1125 ppd->host_link_state == HLS_UP_ARMED ||
1126 ppd->host_link_state == HLS_UP_ACTIVE)
1129 drain = !is_ax(dd) && is_up;
1133 * MTU is specified per-VL. To ensure that no packet gets
1134 * stuck (due, e.g., to the MTU for the packet's VL being
1135 * reduced), empty the per-VL FIFOs before adjusting MTU.
1137 ret = stop_drain_data_vls(dd);
1140 dd_dev_err(dd, "%s: cannot stop/drain VLs - refusing to change per-VL MTUs\n",
1145 hfi1_set_ib_cfg(ppd, HFI1_IB_CFG_MTU, 0);
1148 open_fill_data_vls(dd); /* reopen all VLs */
1151 mutex_unlock(&ppd->hls_lock);
1156 int hfi1_set_lid(struct hfi1_pportdata *ppd, u32 lid, u8 lmc)
1158 struct hfi1_devdata *dd = ppd->dd;
1162 hfi1_set_ib_cfg(ppd, HFI1_IB_CFG_LIDLMC, 0);
1164 dd_dev_info(dd, "port %u: got a lid: 0x%x\n", ppd->port, lid);
1169 void shutdown_led_override(struct hfi1_pportdata *ppd)
1171 struct hfi1_devdata *dd = ppd->dd;
1174 * This pairs with the memory barrier in hfi1_start_led_override to
1175 * ensure that we read the correct state of LED beaconing represented
1176 * by led_override_timer_active
1179 if (atomic_read(&ppd->led_override_timer_active)) {
1180 del_timer_sync(&ppd->led_override_timer);
1181 atomic_set(&ppd->led_override_timer_active, 0);
1182 /* Ensure the atomic_set is visible to all CPUs */
1186 /* Hand control of the LED to the DC for normal operation */
1187 write_csr(dd, DCC_CFG_LED_CNTRL, 0);
1190 static void run_led_override(unsigned long opaque)
1192 struct hfi1_pportdata *ppd = (struct hfi1_pportdata *)opaque;
1193 struct hfi1_devdata *dd = ppd->dd;
1194 unsigned long timeout;
1197 if (!(dd->flags & HFI1_INITTED))
1200 phase_idx = ppd->led_override_phase & 1;
1202 setextled(dd, phase_idx);
1204 timeout = ppd->led_override_vals[phase_idx];
1206 /* Set up for next phase */
1207 ppd->led_override_phase = !ppd->led_override_phase;
1209 mod_timer(&ppd->led_override_timer, jiffies + timeout);
1213 * To have the LED blink in a particular pattern, provide timeon and timeoff
1215 * To turn off custom blinking and return to normal operation, use
1216 * shutdown_led_override()
1218 void hfi1_start_led_override(struct hfi1_pportdata *ppd, unsigned int timeon,
1219 unsigned int timeoff)
1221 if (!(ppd->dd->flags & HFI1_INITTED))
1224 /* Convert to jiffies for direct use in timer */
1225 ppd->led_override_vals[0] = msecs_to_jiffies(timeoff);
1226 ppd->led_override_vals[1] = msecs_to_jiffies(timeon);
1228 /* Arbitrarily start from LED on phase */
1229 ppd->led_override_phase = 1;
1232 * If the timer has not already been started, do so. Use a "quick"
1233 * timeout so the handler will be called soon to look at our request.
1235 if (!timer_pending(&ppd->led_override_timer)) {
1236 setup_timer(&ppd->led_override_timer, run_led_override,
1237 (unsigned long)ppd);
1238 ppd->led_override_timer.expires = jiffies + 1;
1239 add_timer(&ppd->led_override_timer);
1240 atomic_set(&ppd->led_override_timer_active, 1);
1241 /* Ensure the atomic_set is visible to all CPUs */
1247 * hfi1_reset_device - reset the chip if possible
1248 * @unit: the device to reset
1250 * Whether or not reset is successful, we attempt to re-initialize the chip
1251 * (that is, much like a driver unload/reload). We clear the INITTED flag
1252 * so that the various entry points will fail until we reinitialize. For
1253 * now, we only allow this if no user contexts are open that use chip resources
1255 int hfi1_reset_device(int unit)
1258 struct hfi1_devdata *dd = hfi1_lookup(unit);
1259 struct hfi1_pportdata *ppd;
1260 unsigned long flags;
1268 dd_dev_info(dd, "Reset on unit %u requested\n", unit);
1270 if (!dd->kregbase || !(dd->flags & HFI1_PRESENT)) {
1272 "Invalid unit number %u or not initialized or not present\n",
1278 spin_lock_irqsave(&dd->uctxt_lock, flags);
1280 for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
1281 if (!dd->rcd[i] || !dd->rcd[i]->cnt)
1283 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1287 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1289 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1290 ppd = dd->pport + pidx;
1292 shutdown_led_override(ppd);
1294 if (dd->flags & HFI1_HAS_SEND_DMA)
1297 hfi1_reset_cpu_counters(dd);
1299 ret = hfi1_init(dd, 1);
1303 "Reinitialize unit %u after reset failed with %d\n",
1306 dd_dev_info(dd, "Reinitialized unit %u after resetting\n",
1313 void handle_eflags(struct hfi1_packet *packet)
1315 struct hfi1_ctxtdata *rcd = packet->rcd;
1316 u32 rte = rhf_rcv_type_err(packet->rhf);
1318 rcv_hdrerr(rcd, rcd->ppd, packet);
1319 if (rhf_err_flags(packet->rhf))
1321 "receive context %d: rhf 0x%016llx, errs [ %s%s%s%s%s%s%s%s] rte 0x%x\n",
1322 rcd->ctxt, packet->rhf,
1323 packet->rhf & RHF_K_HDR_LEN_ERR ? "k_hdr_len " : "",
1324 packet->rhf & RHF_DC_UNC_ERR ? "dc_unc " : "",
1325 packet->rhf & RHF_DC_ERR ? "dc " : "",
1326 packet->rhf & RHF_TID_ERR ? "tid " : "",
1327 packet->rhf & RHF_LEN_ERR ? "len " : "",
1328 packet->rhf & RHF_ECC_ERR ? "ecc " : "",
1329 packet->rhf & RHF_VCRC_ERR ? "vcrc " : "",
1330 packet->rhf & RHF_ICRC_ERR ? "icrc " : "",
1335 * The following functions are called by the interrupt handler. They are type
1336 * specific handlers for each packet type.
1338 int process_receive_ib(struct hfi1_packet *packet)
1340 trace_hfi1_rcvhdr(packet->rcd->ppd->dd,
1342 rhf_err_flags(packet->rhf),
1347 rhf_egr_index(packet->rhf));
1349 if (unlikely(rhf_err_flags(packet->rhf))) {
1350 handle_eflags(packet);
1351 return RHF_RCV_CONTINUE;
1354 hfi1_ib_rcv(packet);
1355 return RHF_RCV_CONTINUE;
1358 int process_receive_bypass(struct hfi1_packet *packet)
1360 if (unlikely(rhf_err_flags(packet->rhf)))
1361 handle_eflags(packet);
1363 dd_dev_err(packet->rcd->dd,
1364 "Bypass packets are not supported in normal operation. Dropping\n");
1365 return RHF_RCV_CONTINUE;
1368 int process_receive_error(struct hfi1_packet *packet)
1370 handle_eflags(packet);
1372 if (unlikely(rhf_err_flags(packet->rhf)))
1373 dd_dev_err(packet->rcd->dd,
1374 "Unhandled error packet received. Dropping.\n");
1376 return RHF_RCV_CONTINUE;
1379 int kdeth_process_expected(struct hfi1_packet *packet)
1381 if (unlikely(rhf_err_flags(packet->rhf)))
1382 handle_eflags(packet);
1384 dd_dev_err(packet->rcd->dd,
1385 "Unhandled expected packet received. Dropping.\n");
1386 return RHF_RCV_CONTINUE;
1389 int kdeth_process_eager(struct hfi1_packet *packet)
1391 if (unlikely(rhf_err_flags(packet->rhf)))
1392 handle_eflags(packet);
1394 dd_dev_err(packet->rcd->dd,
1395 "Unhandled eager packet received. Dropping.\n");
1396 return RHF_RCV_CONTINUE;
1399 int process_receive_invalid(struct hfi1_packet *packet)
1401 dd_dev_err(packet->rcd->dd, "Invalid packet type %d. Dropping\n",
1402 rhf_rcv_type(packet->rhf));
1403 return RHF_RCV_CONTINUE;