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
= hdr2sc(rhdr
, packet
->rhf
);
396 sl
= ibp
->sc_to_sl
[sc5
];
398 lqpn
= be32_to_cpu(bth
[1]) & RVT_QPN_MASK
;
400 qp
= rvt_lookup_qpn(rdi
, &ibp
->rvp
, lqpn
);
406 switch (qp
->ibqp
.qp_type
) {
410 svc_type
= IB_CC_SVCTYPE_UD
;
413 rlid
= be16_to_cpu(rhdr
->lrh
[3]);
414 rqpn
= qp
->remote_qpn
;
415 svc_type
= IB_CC_SVCTYPE_UC
;
421 process_becn(ppd
, sl
, rlid
, lqpn
, rqpn
, svc_type
);
425 packet
->rhf
&= ~RHF_RCV_TYPE_ERR_SMASK
;
436 static inline void init_packet(struct hfi1_ctxtdata
*rcd
,
437 struct hfi1_packet
*packet
)
439 packet
->rsize
= rcd
->rcvhdrqentsize
; /* words */
440 packet
->maxcnt
= rcd
->rcvhdrq_cnt
* packet
->rsize
; /* words */
444 packet
->rhf_addr
= get_rhf_addr(rcd
);
445 packet
->rhf
= rhf_to_cpu(packet
->rhf_addr
);
446 packet
->rhqoff
= rcd
->head
;
448 packet
->rcv_flags
= 0;
451 void hfi1_process_ecn_slowpath(struct rvt_qp
*qp
, struct hfi1_packet
*pkt
,
454 struct hfi1_ibport
*ibp
= to_iport(qp
->ibqp
.device
, qp
->port_num
);
455 struct hfi1_ib_header
*hdr
= pkt
->hdr
;
456 struct hfi1_other_headers
*ohdr
= pkt
->ohdr
;
457 struct ib_grh
*grh
= NULL
;
459 u16 rlid
, dlid
= be16_to_cpu(hdr
->lrh
[1]);
461 bool is_mcast
= false;
463 if (pkt
->rcv_flags
& HFI1_HAS_GRH
)
466 switch (qp
->ibqp
.qp_type
) {
470 rlid
= be16_to_cpu(hdr
->lrh
[3]);
471 rqpn
= be32_to_cpu(ohdr
->u
.ud
.deth
[1]) & RVT_QPN_MASK
;
472 svc_type
= IB_CC_SVCTYPE_UD
;
473 is_mcast
= (dlid
> be16_to_cpu(IB_MULTICAST_LID_BASE
)) &&
474 (dlid
!= be16_to_cpu(IB_LID_PERMISSIVE
));
477 rlid
= qp
->remote_ah_attr
.dlid
;
478 rqpn
= qp
->remote_qpn
;
479 svc_type
= IB_CC_SVCTYPE_UC
;
482 rlid
= qp
->remote_ah_attr
.dlid
;
483 rqpn
= qp
->remote_qpn
;
484 svc_type
= IB_CC_SVCTYPE_RC
;
490 sc
= hdr2sc((struct hfi1_message_header
*)hdr
, pkt
->rhf
);
492 bth1
= be32_to_cpu(ohdr
->bth
[1]);
493 if (do_cnp
&& (bth1
& HFI1_FECN_SMASK
)) {
494 u16 pkey
= (u16
)be32_to_cpu(ohdr
->bth
[0]);
496 return_cnp(ibp
, qp
, rqpn
, pkey
, dlid
, rlid
, sc
, grh
);
499 if (!is_mcast
&& (bth1
& HFI1_BECN_SMASK
)) {
500 struct hfi1_pportdata
*ppd
= ppd_from_ibp(ibp
);
501 u32 lqpn
= bth1
& RVT_QPN_MASK
;
502 u8 sl
= ibp
->sc_to_sl
[sc
];
504 process_becn(ppd
, sl
, rlid
, lqpn
, rqpn
, svc_type
);
510 struct hfi1_ctxtdata
*rcd
;
518 static inline void init_ps_mdata(struct ps_mdata
*mdata
,
519 struct hfi1_packet
*packet
)
521 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
524 mdata
->rsize
= packet
->rsize
;
525 mdata
->maxcnt
= packet
->maxcnt
;
526 mdata
->ps_head
= packet
->rhqoff
;
528 if (HFI1_CAP_KGET_MASK(rcd
->flags
, DMA_RTAIL
)) {
529 mdata
->ps_tail
= get_rcvhdrtail(rcd
);
530 if (rcd
->ctxt
== HFI1_CTRL_CTXT
)
531 mdata
->ps_seq
= rcd
->seq_cnt
;
533 mdata
->ps_seq
= 0; /* not used with DMA_RTAIL */
535 mdata
->ps_tail
= 0; /* used only with DMA_RTAIL*/
536 mdata
->ps_seq
= rcd
->seq_cnt
;
540 static inline int ps_done(struct ps_mdata
*mdata
, u64 rhf
,
541 struct hfi1_ctxtdata
*rcd
)
543 if (HFI1_CAP_KGET_MASK(rcd
->flags
, DMA_RTAIL
))
544 return mdata
->ps_head
== mdata
->ps_tail
;
545 return mdata
->ps_seq
!= rhf_rcv_seq(rhf
);
548 static inline int ps_skip(struct ps_mdata
*mdata
, u64 rhf
,
549 struct hfi1_ctxtdata
*rcd
)
552 * Control context can potentially receive an invalid rhf.
555 if ((rcd
->ctxt
== HFI1_CTRL_CTXT
) && (mdata
->ps_head
!= mdata
->ps_tail
))
556 return mdata
->ps_seq
!= rhf_rcv_seq(rhf
);
561 static inline void update_ps_mdata(struct ps_mdata
*mdata
,
562 struct hfi1_ctxtdata
*rcd
)
564 mdata
->ps_head
+= mdata
->rsize
;
565 if (mdata
->ps_head
>= mdata
->maxcnt
)
568 /* Control context must do seq counting */
569 if (!HFI1_CAP_KGET_MASK(rcd
->flags
, DMA_RTAIL
) ||
570 (rcd
->ctxt
== HFI1_CTRL_CTXT
)) {
571 if (++mdata
->ps_seq
> 13)
577 * prescan_rxq - search through the receive queue looking for packets
578 * containing Excplicit Congestion Notifications (FECNs, or BECNs).
579 * When an ECN is found, process the Congestion Notification, and toggle
581 * This is declared as a macro to allow quick checking of the port to avoid
582 * the overhead of a function call if not enabled.
584 #define prescan_rxq(rcd, packet) \
586 if (rcd->ppd->cc_prescan) \
587 __prescan_rxq(packet); \
589 static void __prescan_rxq(struct hfi1_packet
*packet
)
591 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
592 struct ps_mdata mdata
;
594 init_ps_mdata(&mdata
, packet
);
597 struct hfi1_devdata
*dd
= rcd
->dd
;
598 struct hfi1_ibport
*ibp
= &rcd
->ppd
->ibport_data
;
599 __le32
*rhf_addr
= (__le32
*)rcd
->rcvhdrq
+ mdata
.ps_head
+
602 struct hfi1_ib_header
*hdr
;
603 struct hfi1_other_headers
*ohdr
;
604 struct rvt_dev_info
*rdi
= &dd
->verbs_dev
.rdi
;
605 u64 rhf
= rhf_to_cpu(rhf_addr
);
606 u32 etype
= rhf_rcv_type(rhf
), qpn
, bth1
;
610 if (ps_done(&mdata
, rhf
, rcd
))
613 if (ps_skip(&mdata
, rhf
, rcd
))
616 if (etype
!= RHF_RCV_TYPE_IB
)
619 hdr
= (struct hfi1_ib_header
*)
620 hfi1_get_msgheader(dd
, rhf_addr
);
621 lnh
= be16_to_cpu(hdr
->lrh
[0]) & 3;
623 if (lnh
== HFI1_LRH_BTH
)
625 else if (lnh
== HFI1_LRH_GRH
)
626 ohdr
= &hdr
->u
.l
.oth
;
628 goto next
; /* just in case */
630 bth1
= be32_to_cpu(ohdr
->bth
[1]);
631 is_ecn
= !!(bth1
& (HFI1_FECN_SMASK
| HFI1_BECN_SMASK
));
636 qpn
= bth1
& RVT_QPN_MASK
;
638 qp
= rvt_lookup_qpn(rdi
, &ibp
->rvp
, qpn
);
645 process_ecn(qp
, packet
, true);
648 /* turn off BECN, FECN */
649 bth1
&= ~(HFI1_FECN_SMASK
| HFI1_BECN_SMASK
);
650 ohdr
->bth
[1] = cpu_to_be32(bth1
);
652 update_ps_mdata(&mdata
, rcd
);
656 static inline int skip_rcv_packet(struct hfi1_packet
*packet
, int thread
)
658 int ret
= RCV_PKT_OK
;
660 /* Set up for the next packet */
661 packet
->rhqoff
+= packet
->rsize
;
662 if (packet
->rhqoff
>= packet
->maxcnt
)
666 if (unlikely((packet
->numpkt
& (MAX_PKT_RECV
- 1)) == 0)) {
671 this_cpu_inc(*packet
->rcd
->dd
->rcv_limit
);
675 packet
->rhf_addr
= (__le32
*)packet
->rcd
->rcvhdrq
+ packet
->rhqoff
+
676 packet
->rcd
->dd
->rhf_offset
;
677 packet
->rhf
= rhf_to_cpu(packet
->rhf_addr
);
682 static inline int process_rcv_packet(struct hfi1_packet
*packet
, int thread
)
684 int ret
= RCV_PKT_OK
;
686 packet
->hdr
= hfi1_get_msgheader(packet
->rcd
->dd
,
688 packet
->hlen
= (u8
*)packet
->rhf_addr
- (u8
*)packet
->hdr
;
689 packet
->etype
= rhf_rcv_type(packet
->rhf
);
691 packet
->tlen
= rhf_pkt_len(packet
->rhf
); /* in bytes */
692 /* retrieve eager buffer details */
694 if (rhf_use_egr_bfr(packet
->rhf
)) {
695 packet
->etail
= rhf_egr_index(packet
->rhf
);
696 packet
->ebuf
= get_egrbuf(packet
->rcd
, packet
->rhf
,
699 * Prefetch the contents of the eager buffer. It is
700 * OK to send a negative length to prefetch_range().
701 * The +2 is the size of the RHF.
703 prefetch_range(packet
->ebuf
,
704 packet
->tlen
- ((packet
->rcd
->rcvhdrqentsize
-
705 (rhf_hdrq_offset(packet
->rhf
)
710 * Call a type specific handler for the packet. We
711 * should be able to trust that etype won't be beyond
712 * the range of valid indexes. If so something is really
713 * wrong and we can probably just let things come
714 * crashing down. There is no need to eat another
715 * comparison in this performance critical code.
717 packet
->rcd
->dd
->rhf_rcv_function_map
[packet
->etype
](packet
);
720 /* Set up for the next packet */
721 packet
->rhqoff
+= packet
->rsize
;
722 if (packet
->rhqoff
>= packet
->maxcnt
)
725 if (unlikely((packet
->numpkt
& (MAX_PKT_RECV
- 1)) == 0)) {
730 this_cpu_inc(*packet
->rcd
->dd
->rcv_limit
);
734 packet
->rhf_addr
= (__le32
*)packet
->rcd
->rcvhdrq
+ packet
->rhqoff
+
735 packet
->rcd
->dd
->rhf_offset
;
736 packet
->rhf
= rhf_to_cpu(packet
->rhf_addr
);
741 static inline void process_rcv_update(int last
, struct hfi1_packet
*packet
)
744 * Update head regs etc., every 16 packets, if not last pkt,
745 * to help prevent rcvhdrq overflows, when many packets
746 * are processed and queue is nearly full.
747 * Don't request an interrupt for intermediate updates.
749 if (!last
&& !(packet
->numpkt
& 0xf)) {
750 update_usrhead(packet
->rcd
, packet
->rhqoff
, packet
->updegr
,
751 packet
->etail
, 0, 0);
754 packet
->rcv_flags
= 0;
757 static inline void finish_packet(struct hfi1_packet
*packet
)
760 * Nothing we need to free for the packet.
762 * The only thing we need to do is a final update and call for an
765 update_usrhead(packet
->rcd
, packet
->rcd
->head
, packet
->updegr
,
766 packet
->etail
, rcv_intr_dynamic
, packet
->numpkt
);
769 static inline void process_rcv_qp_work(struct hfi1_packet
*packet
)
771 struct hfi1_ctxtdata
*rcd
;
772 struct rvt_qp
*qp
, *nqp
;
775 rcd
->head
= packet
->rhqoff
;
778 * Iterate over all QPs waiting to respond.
779 * The list won't change since the IRQ is only run on one CPU.
781 list_for_each_entry_safe(qp
, nqp
, &rcd
->qp_wait_list
, rspwait
) {
782 list_del_init(&qp
->rspwait
);
783 if (qp
->r_flags
& RVT_R_RSP_NAK
) {
784 qp
->r_flags
&= ~RVT_R_RSP_NAK
;
785 hfi1_send_rc_ack(rcd
, qp
, 0);
787 if (qp
->r_flags
& RVT_R_RSP_SEND
) {
790 qp
->r_flags
&= ~RVT_R_RSP_SEND
;
791 spin_lock_irqsave(&qp
->s_lock
, flags
);
792 if (ib_rvt_state_ops
[qp
->state
] &
793 RVT_PROCESS_OR_FLUSH_SEND
)
794 hfi1_schedule_send(qp
);
795 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
797 if (atomic_dec_and_test(&qp
->refcount
))
803 * Handle receive interrupts when using the no dma rtail option.
805 int handle_receive_interrupt_nodma_rtail(struct hfi1_ctxtdata
*rcd
, int thread
)
808 int last
= RCV_PKT_OK
;
809 struct hfi1_packet packet
;
811 init_packet(rcd
, &packet
);
812 seq
= rhf_rcv_seq(packet
.rhf
);
813 if (seq
!= rcd
->seq_cnt
) {
818 prescan_rxq(rcd
, &packet
);
820 while (last
== RCV_PKT_OK
) {
821 last
= process_rcv_packet(&packet
, thread
);
822 seq
= rhf_rcv_seq(packet
.rhf
);
823 if (++rcd
->seq_cnt
> 13)
825 if (seq
!= rcd
->seq_cnt
)
827 process_rcv_update(last
, &packet
);
829 process_rcv_qp_work(&packet
);
831 finish_packet(&packet
);
835 int handle_receive_interrupt_dma_rtail(struct hfi1_ctxtdata
*rcd
, int thread
)
838 int last
= RCV_PKT_OK
;
839 struct hfi1_packet packet
;
841 init_packet(rcd
, &packet
);
842 hdrqtail
= get_rcvhdrtail(rcd
);
843 if (packet
.rhqoff
== hdrqtail
) {
847 smp_rmb(); /* prevent speculative reads of dma'ed hdrq */
849 prescan_rxq(rcd
, &packet
);
851 while (last
== RCV_PKT_OK
) {
852 last
= process_rcv_packet(&packet
, thread
);
853 if (packet
.rhqoff
== hdrqtail
)
855 process_rcv_update(last
, &packet
);
857 process_rcv_qp_work(&packet
);
859 finish_packet(&packet
);
863 static inline void set_all_nodma_rtail(struct hfi1_devdata
*dd
)
867 for (i
= HFI1_CTRL_CTXT
+ 1; i
< dd
->first_user_ctxt
; i
++)
868 dd
->rcd
[i
]->do_interrupt
=
869 &handle_receive_interrupt_nodma_rtail
;
872 static inline void set_all_dma_rtail(struct hfi1_devdata
*dd
)
876 for (i
= HFI1_CTRL_CTXT
+ 1; i
< dd
->first_user_ctxt
; i
++)
877 dd
->rcd
[i
]->do_interrupt
=
878 &handle_receive_interrupt_dma_rtail
;
881 void set_all_slowpath(struct hfi1_devdata
*dd
)
885 /* HFI1_CTRL_CTXT must always use the slow path interrupt handler */
886 for (i
= HFI1_CTRL_CTXT
+ 1; i
< dd
->first_user_ctxt
; i
++)
887 dd
->rcd
[i
]->do_interrupt
= &handle_receive_interrupt
;
890 static inline int set_armed_to_active(struct hfi1_ctxtdata
*rcd
,
891 struct hfi1_packet packet
,
892 struct hfi1_devdata
*dd
)
894 struct work_struct
*lsaw
= &rcd
->ppd
->linkstate_active_work
;
895 struct hfi1_message_header
*hdr
= hfi1_get_msgheader(packet
.rcd
->dd
,
898 if (hdr2sc(hdr
, packet
.rhf
) != 0xf) {
899 int hwstate
= read_logical_state(dd
);
901 if (hwstate
!= LSTATE_ACTIVE
) {
902 dd_dev_info(dd
, "Unexpected link state %d\n", hwstate
);
906 queue_work(rcd
->ppd
->hfi1_wq
, lsaw
);
913 * handle_receive_interrupt - receive a packet
916 * Called from interrupt handler for errors or receive interrupt.
917 * This is the slow path interrupt handler.
919 int handle_receive_interrupt(struct hfi1_ctxtdata
*rcd
, int thread
)
921 struct hfi1_devdata
*dd
= rcd
->dd
;
923 int needset
, last
= RCV_PKT_OK
;
924 struct hfi1_packet packet
;
927 /* Control context will always use the slow path interrupt handler */
928 needset
= (rcd
->ctxt
== HFI1_CTRL_CTXT
) ? 0 : 1;
930 init_packet(rcd
, &packet
);
932 if (!HFI1_CAP_KGET_MASK(rcd
->flags
, DMA_RTAIL
)) {
933 u32 seq
= rhf_rcv_seq(packet
.rhf
);
935 if (seq
!= rcd
->seq_cnt
) {
941 hdrqtail
= get_rcvhdrtail(rcd
);
942 if (packet
.rhqoff
== hdrqtail
) {
946 smp_rmb(); /* prevent speculative reads of dma'ed hdrq */
949 * Control context can potentially receive an invalid
950 * rhf. Drop such packets.
952 if (rcd
->ctxt
== HFI1_CTRL_CTXT
) {
953 u32 seq
= rhf_rcv_seq(packet
.rhf
);
955 if (seq
!= rcd
->seq_cnt
)
960 prescan_rxq(rcd
, &packet
);
962 while (last
== RCV_PKT_OK
) {
963 if (unlikely(dd
->do_drop
&&
964 atomic_xchg(&dd
->drop_packet
, DROP_PACKET_OFF
) ==
968 /* On to the next packet */
969 packet
.rhqoff
+= packet
.rsize
;
970 packet
.rhf_addr
= (__le32
*)rcd
->rcvhdrq
+
973 packet
.rhf
= rhf_to_cpu(packet
.rhf_addr
);
975 } else if (skip_pkt
) {
976 last
= skip_rcv_packet(&packet
, thread
);
979 /* Auto activate link on non-SC15 packet receive */
980 if (unlikely(rcd
->ppd
->host_link_state
==
982 set_armed_to_active(rcd
, packet
, dd
))
984 last
= process_rcv_packet(&packet
, thread
);
987 if (!HFI1_CAP_KGET_MASK(rcd
->flags
, DMA_RTAIL
)) {
988 u32 seq
= rhf_rcv_seq(packet
.rhf
);
990 if (++rcd
->seq_cnt
> 13)
992 if (seq
!= rcd
->seq_cnt
)
995 dd_dev_info(dd
, "Switching to NO_DMA_RTAIL\n");
996 set_all_nodma_rtail(dd
);
1000 if (packet
.rhqoff
== hdrqtail
)
1001 last
= RCV_PKT_DONE
;
1003 * Control context can potentially receive an invalid
1004 * rhf. Drop such packets.
1006 if (rcd
->ctxt
== HFI1_CTRL_CTXT
) {
1007 u32 seq
= rhf_rcv_seq(packet
.rhf
);
1009 if (++rcd
->seq_cnt
> 13)
1011 if (!last
&& (seq
!= rcd
->seq_cnt
))
1017 "Switching to DMA_RTAIL\n");
1018 set_all_dma_rtail(dd
);
1023 process_rcv_update(last
, &packet
);
1026 process_rcv_qp_work(&packet
);
1030 * Always write head at end, and setup rcv interrupt, even
1031 * if no packets were processed.
1033 finish_packet(&packet
);
1038 * We may discover in the interrupt that the hardware link state has
1039 * changed from ARMED to ACTIVE (due to the arrival of a non-SC15 packet),
1040 * and we need to update the driver's notion of the link state. We cannot
1041 * run set_link_state from interrupt context, so we queue this function on
1044 * We delay the regular interrupt processing until after the state changes
1045 * so that the link will be in the correct state by the time any application
1046 * we wake up attempts to send a reply to any message it received.
1047 * (Subsequent receive interrupts may possibly force the wakeup before we
1048 * update the link state.)
1050 * The rcd is freed in hfi1_free_ctxtdata after hfi1_postinit_cleanup invokes
1051 * dd->f_cleanup(dd) to disable the interrupt handler and flush workqueues,
1052 * so we're safe from use-after-free of the rcd.
1054 void receive_interrupt_work(struct work_struct
*work
)
1056 struct hfi1_pportdata
*ppd
= container_of(work
, struct hfi1_pportdata
,
1057 linkstate_active_work
);
1058 struct hfi1_devdata
*dd
= ppd
->dd
;
1061 /* Received non-SC15 packet implies neighbor_normal */
1062 ppd
->neighbor_normal
= 1;
1063 set_link_state(ppd
, HLS_UP_ACTIVE
);
1066 * Interrupt all kernel contexts that could have had an
1067 * interrupt during auto activation.
1069 for (i
= HFI1_CTRL_CTXT
; i
< dd
->first_user_ctxt
; i
++)
1070 force_recv_intr(dd
->rcd
[i
]);
1074 * Convert a given MTU size to the on-wire MAD packet enumeration.
1075 * Return -1 if the size is invalid.
1077 int mtu_to_enum(u32 mtu
, int default_if_bad
)
1080 case 0: return OPA_MTU_0
;
1081 case 256: return OPA_MTU_256
;
1082 case 512: return OPA_MTU_512
;
1083 case 1024: return OPA_MTU_1024
;
1084 case 2048: return OPA_MTU_2048
;
1085 case 4096: return OPA_MTU_4096
;
1086 case 8192: return OPA_MTU_8192
;
1087 case 10240: return OPA_MTU_10240
;
1089 return default_if_bad
;
1092 u16
enum_to_mtu(int mtu
)
1095 case OPA_MTU_0
: return 0;
1096 case OPA_MTU_256
: return 256;
1097 case OPA_MTU_512
: return 512;
1098 case OPA_MTU_1024
: return 1024;
1099 case OPA_MTU_2048
: return 2048;
1100 case OPA_MTU_4096
: return 4096;
1101 case OPA_MTU_8192
: return 8192;
1102 case OPA_MTU_10240
: return 10240;
1103 default: return 0xffff;
1108 * set_mtu - set the MTU
1109 * @ppd: the per port data
1111 * We can handle "any" incoming size, the issue here is whether we
1112 * need to restrict our outgoing size. We do not deal with what happens
1113 * to programs that are already running when the size changes.
1115 int set_mtu(struct hfi1_pportdata
*ppd
)
1117 struct hfi1_devdata
*dd
= ppd
->dd
;
1118 int i
, drain
, ret
= 0, is_up
= 0;
1121 for (i
= 0; i
< ppd
->vls_supported
; i
++)
1122 if (ppd
->ibmtu
< dd
->vld
[i
].mtu
)
1123 ppd
->ibmtu
= dd
->vld
[i
].mtu
;
1124 ppd
->ibmaxlen
= ppd
->ibmtu
+ lrh_max_header_bytes(ppd
->dd
);
1126 mutex_lock(&ppd
->hls_lock
);
1127 if (ppd
->host_link_state
== HLS_UP_INIT
||
1128 ppd
->host_link_state
== HLS_UP_ARMED
||
1129 ppd
->host_link_state
== HLS_UP_ACTIVE
)
1132 drain
= !is_ax(dd
) && is_up
;
1136 * MTU is specified per-VL. To ensure that no packet gets
1137 * stuck (due, e.g., to the MTU for the packet's VL being
1138 * reduced), empty the per-VL FIFOs before adjusting MTU.
1140 ret
= stop_drain_data_vls(dd
);
1143 dd_dev_err(dd
, "%s: cannot stop/drain VLs - refusing to change per-VL MTUs\n",
1148 hfi1_set_ib_cfg(ppd
, HFI1_IB_CFG_MTU
, 0);
1151 open_fill_data_vls(dd
); /* reopen all VLs */
1154 mutex_unlock(&ppd
->hls_lock
);
1159 int hfi1_set_lid(struct hfi1_pportdata
*ppd
, u32 lid
, u8 lmc
)
1161 struct hfi1_devdata
*dd
= ppd
->dd
;
1165 hfi1_set_ib_cfg(ppd
, HFI1_IB_CFG_LIDLMC
, 0);
1167 dd_dev_info(dd
, "port %u: got a lid: 0x%x\n", ppd
->port
, lid
);
1172 void shutdown_led_override(struct hfi1_pportdata
*ppd
)
1174 struct hfi1_devdata
*dd
= ppd
->dd
;
1177 * This pairs with the memory barrier in hfi1_start_led_override to
1178 * ensure that we read the correct state of LED beaconing represented
1179 * by led_override_timer_active
1182 if (atomic_read(&ppd
->led_override_timer_active
)) {
1183 del_timer_sync(&ppd
->led_override_timer
);
1184 atomic_set(&ppd
->led_override_timer_active
, 0);
1185 /* Ensure the atomic_set is visible to all CPUs */
1189 /* Hand control of the LED to the DC for normal operation */
1190 write_csr(dd
, DCC_CFG_LED_CNTRL
, 0);
1193 static void run_led_override(unsigned long opaque
)
1195 struct hfi1_pportdata
*ppd
= (struct hfi1_pportdata
*)opaque
;
1196 struct hfi1_devdata
*dd
= ppd
->dd
;
1197 unsigned long timeout
;
1200 if (!(dd
->flags
& HFI1_INITTED
))
1203 phase_idx
= ppd
->led_override_phase
& 1;
1205 setextled(dd
, phase_idx
);
1207 timeout
= ppd
->led_override_vals
[phase_idx
];
1209 /* Set up for next phase */
1210 ppd
->led_override_phase
= !ppd
->led_override_phase
;
1212 mod_timer(&ppd
->led_override_timer
, jiffies
+ timeout
);
1216 * To have the LED blink in a particular pattern, provide timeon and timeoff
1218 * To turn off custom blinking and return to normal operation, use
1219 * shutdown_led_override()
1221 void hfi1_start_led_override(struct hfi1_pportdata
*ppd
, unsigned int timeon
,
1222 unsigned int timeoff
)
1224 if (!(ppd
->dd
->flags
& HFI1_INITTED
))
1227 /* Convert to jiffies for direct use in timer */
1228 ppd
->led_override_vals
[0] = msecs_to_jiffies(timeoff
);
1229 ppd
->led_override_vals
[1] = msecs_to_jiffies(timeon
);
1231 /* Arbitrarily start from LED on phase */
1232 ppd
->led_override_phase
= 1;
1235 * If the timer has not already been started, do so. Use a "quick"
1236 * timeout so the handler will be called soon to look at our request.
1238 if (!timer_pending(&ppd
->led_override_timer
)) {
1239 setup_timer(&ppd
->led_override_timer
, run_led_override
,
1240 (unsigned long)ppd
);
1241 ppd
->led_override_timer
.expires
= jiffies
+ 1;
1242 add_timer(&ppd
->led_override_timer
);
1243 atomic_set(&ppd
->led_override_timer_active
, 1);
1244 /* Ensure the atomic_set is visible to all CPUs */
1250 * hfi1_reset_device - reset the chip if possible
1251 * @unit: the device to reset
1253 * Whether or not reset is successful, we attempt to re-initialize the chip
1254 * (that is, much like a driver unload/reload). We clear the INITTED flag
1255 * so that the various entry points will fail until we reinitialize. For
1256 * now, we only allow this if no user contexts are open that use chip resources
1258 int hfi1_reset_device(int unit
)
1261 struct hfi1_devdata
*dd
= hfi1_lookup(unit
);
1262 struct hfi1_pportdata
*ppd
;
1263 unsigned long flags
;
1271 dd_dev_info(dd
, "Reset on unit %u requested\n", unit
);
1273 if (!dd
->kregbase
|| !(dd
->flags
& HFI1_PRESENT
)) {
1275 "Invalid unit number %u or not initialized or not present\n",
1281 spin_lock_irqsave(&dd
->uctxt_lock
, flags
);
1283 for (i
= dd
->first_user_ctxt
; i
< dd
->num_rcv_contexts
; i
++) {
1284 if (!dd
->rcd
[i
] || !dd
->rcd
[i
]->cnt
)
1286 spin_unlock_irqrestore(&dd
->uctxt_lock
, flags
);
1290 spin_unlock_irqrestore(&dd
->uctxt_lock
, flags
);
1292 for (pidx
= 0; pidx
< dd
->num_pports
; ++pidx
) {
1293 ppd
= dd
->pport
+ pidx
;
1295 shutdown_led_override(ppd
);
1297 if (dd
->flags
& HFI1_HAS_SEND_DMA
)
1300 hfi1_reset_cpu_counters(dd
);
1302 ret
= hfi1_init(dd
, 1);
1306 "Reinitialize unit %u after reset failed with %d\n",
1309 dd_dev_info(dd
, "Reinitialized unit %u after resetting\n",
1316 void handle_eflags(struct hfi1_packet
*packet
)
1318 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
1319 u32 rte
= rhf_rcv_type_err(packet
->rhf
);
1321 rcv_hdrerr(rcd
, rcd
->ppd
, packet
);
1322 if (rhf_err_flags(packet
->rhf
))
1324 "receive context %d: rhf 0x%016llx, errs [ %s%s%s%s%s%s%s%s] rte 0x%x\n",
1325 rcd
->ctxt
, packet
->rhf
,
1326 packet
->rhf
& RHF_K_HDR_LEN_ERR
? "k_hdr_len " : "",
1327 packet
->rhf
& RHF_DC_UNC_ERR
? "dc_unc " : "",
1328 packet
->rhf
& RHF_DC_ERR
? "dc " : "",
1329 packet
->rhf
& RHF_TID_ERR
? "tid " : "",
1330 packet
->rhf
& RHF_LEN_ERR
? "len " : "",
1331 packet
->rhf
& RHF_ECC_ERR
? "ecc " : "",
1332 packet
->rhf
& RHF_VCRC_ERR
? "vcrc " : "",
1333 packet
->rhf
& RHF_ICRC_ERR
? "icrc " : "",
1338 * The following functions are called by the interrupt handler. They are type
1339 * specific handlers for each packet type.
1341 int process_receive_ib(struct hfi1_packet
*packet
)
1343 trace_hfi1_rcvhdr(packet
->rcd
->ppd
->dd
,
1345 rhf_err_flags(packet
->rhf
),
1350 rhf_egr_index(packet
->rhf
));
1352 if (unlikely(rhf_err_flags(packet
->rhf
))) {
1353 handle_eflags(packet
);
1354 return RHF_RCV_CONTINUE
;
1357 hfi1_ib_rcv(packet
);
1358 return RHF_RCV_CONTINUE
;
1361 int process_receive_bypass(struct hfi1_packet
*packet
)
1363 if (unlikely(rhf_err_flags(packet
->rhf
)))
1364 handle_eflags(packet
);
1366 dd_dev_err(packet
->rcd
->dd
,
1367 "Bypass packets are not supported in normal operation. Dropping\n");
1368 incr_cntr64(&packet
->rcd
->dd
->sw_rcv_bypass_packet_errors
);
1369 return RHF_RCV_CONTINUE
;
1372 int process_receive_error(struct hfi1_packet
*packet
)
1374 handle_eflags(packet
);
1376 if (unlikely(rhf_err_flags(packet
->rhf
)))
1377 dd_dev_err(packet
->rcd
->dd
,
1378 "Unhandled error packet received. Dropping.\n");
1380 return RHF_RCV_CONTINUE
;
1383 int kdeth_process_expected(struct hfi1_packet
*packet
)
1385 if (unlikely(rhf_err_flags(packet
->rhf
)))
1386 handle_eflags(packet
);
1388 dd_dev_err(packet
->rcd
->dd
,
1389 "Unhandled expected packet received. Dropping.\n");
1390 return RHF_RCV_CONTINUE
;
1393 int kdeth_process_eager(struct hfi1_packet
*packet
)
1395 if (unlikely(rhf_err_flags(packet
->rhf
)))
1396 handle_eflags(packet
);
1398 dd_dev_err(packet
->rcd
->dd
,
1399 "Unhandled eager packet received. Dropping.\n");
1400 return RHF_RCV_CONTINUE
;
1403 int process_receive_invalid(struct hfi1_packet
*packet
)
1405 dd_dev_err(packet
->rcd
->dd
, "Invalid packet type %d. Dropping\n",
1406 rhf_rcv_type(packet
->rhf
));
1407 return RHF_RCV_CONTINUE
;