2 * Copyright(c) 2015-2018 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>
67 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
70 * The size has to be longer than this string, so we can append
71 * board/chip information to it in the initialization code.
73 const char ib_hfi1_version
[] = HFI1_DRIVER_VERSION
"\n";
75 DEFINE_MUTEX(hfi1_mutex
); /* general driver use */
77 unsigned int hfi1_max_mtu
= HFI1_DEFAULT_MAX_MTU
;
78 module_param_named(max_mtu
, hfi1_max_mtu
, uint
, S_IRUGO
);
79 MODULE_PARM_DESC(max_mtu
, "Set max MTU bytes, default is " __stringify(
80 HFI1_DEFAULT_MAX_MTU
));
82 unsigned int hfi1_cu
= 1;
83 module_param_named(cu
, hfi1_cu
, uint
, S_IRUGO
);
84 MODULE_PARM_DESC(cu
, "Credit return units");
86 unsigned long hfi1_cap_mask
= HFI1_CAP_MASK_DEFAULT
;
87 static int hfi1_caps_set(const char *val
, const struct kernel_param
*kp
);
88 static int hfi1_caps_get(char *buffer
, const struct kernel_param
*kp
);
89 static const struct kernel_param_ops cap_ops
= {
93 module_param_cb(cap_mask
, &cap_ops
, &hfi1_cap_mask
, S_IWUSR
| S_IRUGO
);
94 MODULE_PARM_DESC(cap_mask
, "Bit mask of enabled/disabled HW features");
96 MODULE_LICENSE("Dual BSD/GPL");
97 MODULE_DESCRIPTION("Intel Omni-Path Architecture driver");
100 * MAX_PKT_RCV is the max # if packets processed per receive interrupt.
102 #define MAX_PKT_RECV 64
104 * MAX_PKT_THREAD_RCV is the max # of packets processed before
105 * the qp_wait_list queue is flushed.
107 #define MAX_PKT_RECV_THREAD (MAX_PKT_RECV * 4)
108 #define EGR_HEAD_UPDATE_THRESHOLD 16
110 struct hfi1_ib_stats hfi1_stats
;
112 static int hfi1_caps_set(const char *val
, const struct kernel_param
*kp
)
115 unsigned long *cap_mask_ptr
= (unsigned long *)kp
->arg
,
116 cap_mask
= *cap_mask_ptr
, value
, diff
,
117 write_mask
= ((HFI1_CAP_WRITABLE_MASK
<< HFI1_CAP_USER_SHIFT
) |
118 HFI1_CAP_WRITABLE_MASK
);
120 ret
= kstrtoul(val
, 0, &value
);
122 pr_warn("Invalid module parameter value for 'cap_mask'\n");
125 /* Get the changed bits (except the locked bit) */
126 diff
= value
^ (cap_mask
& ~HFI1_CAP_LOCKED_SMASK
);
128 /* Remove any bits that are not allowed to change after driver load */
129 if (HFI1_CAP_LOCKED() && (diff
& ~write_mask
)) {
130 pr_warn("Ignoring non-writable capability bits %#lx\n",
135 /* Mask off any reserved bits */
136 diff
&= ~HFI1_CAP_RESERVED_MASK
;
137 /* Clear any previously set and changing bits */
139 /* Update the bits with the new capability */
140 cap_mask
|= (value
& diff
);
141 /* Check for any kernel/user restrictions */
142 diff
= (cap_mask
& (HFI1_CAP_MUST_HAVE_KERN
<< HFI1_CAP_USER_SHIFT
)) ^
143 ((cap_mask
& HFI1_CAP_MUST_HAVE_KERN
) << HFI1_CAP_USER_SHIFT
);
145 /* Set the bitmask to the final set */
146 *cap_mask_ptr
= cap_mask
;
151 static int hfi1_caps_get(char *buffer
, const struct kernel_param
*kp
)
153 unsigned long cap_mask
= *(unsigned long *)kp
->arg
;
155 cap_mask
&= ~HFI1_CAP_LOCKED_SMASK
;
156 cap_mask
|= ((cap_mask
& HFI1_CAP_K2U
) << HFI1_CAP_USER_SHIFT
);
158 return scnprintf(buffer
, PAGE_SIZE
, "0x%lx", cap_mask
);
161 struct pci_dev
*get_pci_dev(struct rvt_dev_info
*rdi
)
163 struct hfi1_ibdev
*ibdev
= container_of(rdi
, struct hfi1_ibdev
, rdi
);
164 struct hfi1_devdata
*dd
= container_of(ibdev
,
165 struct hfi1_devdata
, verbs_dev
);
170 * Return count of units with at least one port ACTIVE.
172 int hfi1_count_active_units(void)
174 struct hfi1_devdata
*dd
;
175 struct hfi1_pportdata
*ppd
;
176 unsigned long index
, flags
;
177 int pidx
, nunits_active
= 0;
179 xa_lock_irqsave(&hfi1_dev_table
, flags
);
180 xa_for_each(&hfi1_dev_table
, index
, dd
) {
181 if (!(dd
->flags
& HFI1_PRESENT
) || !dd
->kregbase1
)
183 for (pidx
= 0; pidx
< dd
->num_pports
; ++pidx
) {
184 ppd
= dd
->pport
+ pidx
;
185 if (ppd
->lid
&& ppd
->linkup
) {
191 xa_unlock_irqrestore(&hfi1_dev_table
, flags
);
192 return nunits_active
;
196 * Get address of eager buffer from it's index (allocated in chunks, not
199 static inline void *get_egrbuf(const struct hfi1_ctxtdata
*rcd
, u64 rhf
,
202 u32 idx
= rhf_egr_index(rhf
), offset
= rhf_egr_buf_offset(rhf
);
204 *update
|= !(idx
& (rcd
->egrbufs
.threshold
- 1)) && !offset
;
205 return (void *)(((u64
)(rcd
->egrbufs
.rcvtids
[idx
].addr
)) +
206 (offset
* RCV_BUF_BLOCK_SIZE
));
209 static inline void *hfi1_get_header(struct hfi1_ctxtdata
*rcd
,
212 u32 offset
= rhf_hdrq_offset(rhf_to_cpu(rhf_addr
));
214 return (void *)(rhf_addr
- rcd
->rhf_offset
+ offset
);
217 static inline struct ib_header
*hfi1_get_msgheader(struct hfi1_ctxtdata
*rcd
,
220 return (struct ib_header
*)hfi1_get_header(rcd
, rhf_addr
);
223 static inline struct hfi1_16b_header
224 *hfi1_get_16B_header(struct hfi1_ctxtdata
*rcd
,
227 return (struct hfi1_16b_header
*)hfi1_get_header(rcd
, rhf_addr
);
231 * Validate and encode the a given RcvArray Buffer size.
232 * The function will check whether the given size falls within
233 * allowed size ranges for the respective type and, optionally,
234 * return the proper encoding.
236 int hfi1_rcvbuf_validate(u32 size
, u8 type
, u16
*encoded
)
238 if (unlikely(!PAGE_ALIGNED(size
)))
240 if (unlikely(size
< MIN_EAGER_BUFFER
))
243 (type
== PT_EAGER
? MAX_EAGER_BUFFER
: MAX_EXPECTED_BUFFER
))
246 *encoded
= ilog2(size
/ PAGE_SIZE
) + 1;
250 static void rcv_hdrerr(struct hfi1_ctxtdata
*rcd
, struct hfi1_pportdata
*ppd
,
251 struct hfi1_packet
*packet
)
253 struct ib_header
*rhdr
= packet
->hdr
;
254 u32 rte
= rhf_rcv_type_err(packet
->rhf
);
256 struct hfi1_ibport
*ibp
= rcd_to_iport(rcd
);
257 struct hfi1_devdata
*dd
= ppd
->dd
;
258 struct hfi1_ibdev
*verbs_dev
= &dd
->verbs_dev
;
259 struct rvt_dev_info
*rdi
= &verbs_dev
->rdi
;
261 if ((packet
->rhf
& RHF_DC_ERR
) &&
262 hfi1_dbg_fault_suppress_err(verbs_dev
))
265 if (packet
->rhf
& RHF_ICRC_ERR
)
268 if (packet
->etype
== RHF_RCV_TYPE_BYPASS
) {
271 u8 lnh
= ib_get_lnh(rhdr
);
273 mlid_base
= be16_to_cpu(IB_MULTICAST_LID_BASE
);
274 if (lnh
== HFI1_LRH_BTH
) {
275 packet
->ohdr
= &rhdr
->u
.oth
;
276 } else if (lnh
== HFI1_LRH_GRH
) {
277 packet
->ohdr
= &rhdr
->u
.l
.oth
;
278 packet
->grh
= &rhdr
->u
.l
.grh
;
284 if (packet
->rhf
& RHF_TID_ERR
) {
285 /* For TIDERR and RC QPs preemptively schedule a NAK */
286 u32 tlen
= rhf_pkt_len(packet
->rhf
); /* in bytes */
287 u32 dlid
= ib_get_dlid(rhdr
);
290 /* Sanity check packet */
297 struct ib_grh
*grh
= packet
->grh
;
299 if (grh
->next_hdr
!= IB_GRH_NEXT_HDR
)
301 vtf
= be32_to_cpu(grh
->version_tclass_flow
);
302 if ((vtf
>> IB_GRH_VERSION_SHIFT
) != IB_GRH_VERSION
)
306 /* Get the destination QP number. */
307 qp_num
= ib_bth_get_qpn(packet
->ohdr
);
308 if (dlid
< mlid_base
) {
313 qp
= rvt_lookup_qpn(rdi
, &ibp
->rvp
, qp_num
);
320 * Handle only RC QPs - for other QP types drop error
323 spin_lock_irqsave(&qp
->r_lock
, flags
);
325 /* Check for valid receive state. */
326 if (!(ib_rvt_state_ops
[qp
->state
] &
327 RVT_PROCESS_RECV_OK
)) {
328 ibp
->rvp
.n_pkt_drops
++;
331 switch (qp
->ibqp
.qp_type
) {
333 hfi1_rc_hdrerr(rcd
, packet
, qp
);
336 /* For now don't handle any other QP types */
340 spin_unlock_irqrestore(&qp
->r_lock
, flags
);
343 } /* Valid packet with TIDErr */
345 /* handle "RcvTypeErr" flags */
347 case RHF_RTE_ERROR_OP_CODE_ERR
:
352 if (rhf_use_egr_bfr(packet
->rhf
))
356 goto drop
; /* this should never happen */
358 opcode
= ib_bth_get_opcode(packet
->ohdr
);
359 if (opcode
== IB_OPCODE_CNP
) {
361 * Only in pre-B0 h/w is the CNP_OPCODE handled
362 * via this code path.
364 struct rvt_qp
*qp
= NULL
;
367 u8 svc_type
, sl
, sc5
;
369 sc5
= hfi1_9B_get_sc5(rhdr
, packet
->rhf
);
370 sl
= ibp
->sc_to_sl
[sc5
];
372 lqpn
= ib_bth_get_qpn(packet
->ohdr
);
374 qp
= rvt_lookup_qpn(rdi
, &ibp
->rvp
, lqpn
);
380 switch (qp
->ibqp
.qp_type
) {
384 svc_type
= IB_CC_SVCTYPE_UD
;
387 rlid
= ib_get_slid(rhdr
);
388 rqpn
= qp
->remote_qpn
;
389 svc_type
= IB_CC_SVCTYPE_UC
;
396 process_becn(ppd
, sl
, rlid
, lqpn
, rqpn
, svc_type
);
400 packet
->rhf
&= ~RHF_RCV_TYPE_ERR_SMASK
;
411 static inline void init_packet(struct hfi1_ctxtdata
*rcd
,
412 struct hfi1_packet
*packet
)
414 packet
->rsize
= rcd
->rcvhdrqentsize
; /* words */
415 packet
->maxcnt
= rcd
->rcvhdrq_cnt
* packet
->rsize
; /* words */
419 packet
->rhf_addr
= get_rhf_addr(rcd
);
420 packet
->rhf
= rhf_to_cpu(packet
->rhf_addr
);
421 packet
->rhqoff
= rcd
->head
;
425 /* We support only two types - 9B and 16B for now */
426 static const hfi1_handle_cnp hfi1_handle_cnp_tbl
[2] = {
427 [HFI1_PKT_TYPE_9B
] = &return_cnp
,
428 [HFI1_PKT_TYPE_16B
] = &return_cnp_16B
432 * hfi1_process_ecn_slowpath - Process FECN or BECN bits
433 * @qp: The packet's destination QP
434 * @pkt: The packet itself.
435 * @prescan: Is the caller the RXQ prescan
437 * Process the packet's FECN or BECN bits. By now, the packet
438 * has already been evaluated whether processing of those bit should
440 * The significance of the @prescan argument is that if the caller
441 * is the RXQ prescan, a CNP will be send out instead of waiting for the
442 * normal packet processing to send an ACK with BECN set (or a CNP).
444 bool hfi1_process_ecn_slowpath(struct rvt_qp
*qp
, struct hfi1_packet
*pkt
,
447 struct hfi1_ibport
*ibp
= to_iport(qp
->ibqp
.device
, qp
->port_num
);
448 struct hfi1_pportdata
*ppd
= ppd_from_ibp(ibp
);
449 struct ib_other_headers
*ohdr
= pkt
->ohdr
;
450 struct ib_grh
*grh
= pkt
->grh
;
453 u32 rlid
, slid
, dlid
= 0;
454 u8 hdr_type
, sc
, svc_type
, opcode
;
455 bool is_mcast
= false, ignore_fecn
= false, do_cnp
= false,
458 /* can be called from prescan */
459 if (pkt
->etype
== RHF_RCV_TYPE_BYPASS
) {
460 pkey
= hfi1_16B_get_pkey(pkt
->hdr
);
461 sc
= hfi1_16B_get_sc(pkt
->hdr
);
462 dlid
= hfi1_16B_get_dlid(pkt
->hdr
);
463 slid
= hfi1_16B_get_slid(pkt
->hdr
);
464 is_mcast
= hfi1_is_16B_mcast(dlid
);
465 opcode
= ib_bth_get_opcode(ohdr
);
466 hdr_type
= HFI1_PKT_TYPE_16B
;
467 fecn
= hfi1_16B_get_fecn(pkt
->hdr
);
468 becn
= hfi1_16B_get_becn(pkt
->hdr
);
470 pkey
= ib_bth_get_pkey(ohdr
);
471 sc
= hfi1_9B_get_sc5(pkt
->hdr
, pkt
->rhf
);
472 dlid
= qp
->ibqp
.qp_type
!= IB_QPT_UD
? ib_get_dlid(pkt
->hdr
) :
474 slid
= ib_get_slid(pkt
->hdr
);
475 is_mcast
= (dlid
> be16_to_cpu(IB_MULTICAST_LID_BASE
)) &&
476 (dlid
!= be16_to_cpu(IB_LID_PERMISSIVE
));
477 opcode
= ib_bth_get_opcode(ohdr
);
478 hdr_type
= HFI1_PKT_TYPE_9B
;
479 fecn
= ib_bth_get_fecn(ohdr
);
480 becn
= ib_bth_get_becn(ohdr
);
483 switch (qp
->ibqp
.qp_type
) {
486 rqpn
= ib_get_sqpn(pkt
->ohdr
);
487 svc_type
= IB_CC_SVCTYPE_UD
;
492 rqpn
= ib_get_sqpn(pkt
->ohdr
);
493 svc_type
= IB_CC_SVCTYPE_UD
;
496 rlid
= rdma_ah_get_dlid(&qp
->remote_ah_attr
);
497 rqpn
= qp
->remote_qpn
;
498 svc_type
= IB_CC_SVCTYPE_UC
;
501 rlid
= rdma_ah_get_dlid(&qp
->remote_ah_attr
);
502 rqpn
= qp
->remote_qpn
;
503 svc_type
= IB_CC_SVCTYPE_RC
;
509 ignore_fecn
= is_mcast
|| (opcode
== IB_OPCODE_CNP
) ||
510 (opcode
== IB_OPCODE_RC_ACKNOWLEDGE
);
512 * ACKNOWLEDGE packets do not get a CNP but this will be
513 * guarded by ignore_fecn above.
516 (opcode
>= IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST
&&
517 opcode
<= IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE
) ||
518 opcode
== TID_OP(READ_RESP
) ||
519 opcode
== TID_OP(ACK
);
521 /* Call appropriate CNP handler */
522 if (!ignore_fecn
&& do_cnp
&& fecn
)
523 hfi1_handle_cnp_tbl
[hdr_type
](ibp
, qp
, rqpn
, pkey
,
524 dlid
, rlid
, sc
, grh
);
527 u32 lqpn
= be32_to_cpu(ohdr
->bth
[1]) & RVT_QPN_MASK
;
528 u8 sl
= ibp
->sc_to_sl
[sc
];
530 process_becn(ppd
, sl
, rlid
, lqpn
, rqpn
, svc_type
);
532 return !ignore_fecn
&& fecn
;
536 struct hfi1_ctxtdata
*rcd
;
544 static inline void init_ps_mdata(struct ps_mdata
*mdata
,
545 struct hfi1_packet
*packet
)
547 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
550 mdata
->rsize
= packet
->rsize
;
551 mdata
->maxcnt
= packet
->maxcnt
;
552 mdata
->ps_head
= packet
->rhqoff
;
554 if (HFI1_CAP_KGET_MASK(rcd
->flags
, DMA_RTAIL
)) {
555 mdata
->ps_tail
= get_rcvhdrtail(rcd
);
556 if (rcd
->ctxt
== HFI1_CTRL_CTXT
)
557 mdata
->ps_seq
= rcd
->seq_cnt
;
559 mdata
->ps_seq
= 0; /* not used with DMA_RTAIL */
561 mdata
->ps_tail
= 0; /* used only with DMA_RTAIL*/
562 mdata
->ps_seq
= rcd
->seq_cnt
;
566 static inline int ps_done(struct ps_mdata
*mdata
, u64 rhf
,
567 struct hfi1_ctxtdata
*rcd
)
569 if (HFI1_CAP_KGET_MASK(rcd
->flags
, DMA_RTAIL
))
570 return mdata
->ps_head
== mdata
->ps_tail
;
571 return mdata
->ps_seq
!= rhf_rcv_seq(rhf
);
574 static inline int ps_skip(struct ps_mdata
*mdata
, u64 rhf
,
575 struct hfi1_ctxtdata
*rcd
)
578 * Control context can potentially receive an invalid rhf.
581 if ((rcd
->ctxt
== HFI1_CTRL_CTXT
) && (mdata
->ps_head
!= mdata
->ps_tail
))
582 return mdata
->ps_seq
!= rhf_rcv_seq(rhf
);
587 static inline void update_ps_mdata(struct ps_mdata
*mdata
,
588 struct hfi1_ctxtdata
*rcd
)
590 mdata
->ps_head
+= mdata
->rsize
;
591 if (mdata
->ps_head
>= mdata
->maxcnt
)
594 /* Control context must do seq counting */
595 if (!HFI1_CAP_KGET_MASK(rcd
->flags
, DMA_RTAIL
) ||
596 (rcd
->ctxt
== HFI1_CTRL_CTXT
)) {
597 if (++mdata
->ps_seq
> 13)
603 * prescan_rxq - search through the receive queue looking for packets
604 * containing Excplicit Congestion Notifications (FECNs, or BECNs).
605 * When an ECN is found, process the Congestion Notification, and toggle
607 * This is declared as a macro to allow quick checking of the port to avoid
608 * the overhead of a function call if not enabled.
610 #define prescan_rxq(rcd, packet) \
612 if (rcd->ppd->cc_prescan) \
613 __prescan_rxq(packet); \
615 static void __prescan_rxq(struct hfi1_packet
*packet
)
617 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
618 struct ps_mdata mdata
;
620 init_ps_mdata(&mdata
, packet
);
623 struct hfi1_ibport
*ibp
= rcd_to_iport(rcd
);
624 __le32
*rhf_addr
= (__le32
*)rcd
->rcvhdrq
+ mdata
.ps_head
+
625 packet
->rcd
->rhf_offset
;
627 struct ib_header
*hdr
;
628 struct rvt_dev_info
*rdi
= &rcd
->dd
->verbs_dev
.rdi
;
629 u64 rhf
= rhf_to_cpu(rhf_addr
);
630 u32 etype
= rhf_rcv_type(rhf
), qpn
, bth1
;
633 if (ps_done(&mdata
, rhf
, rcd
))
636 if (ps_skip(&mdata
, rhf
, rcd
))
639 if (etype
!= RHF_RCV_TYPE_IB
)
642 packet
->hdr
= hfi1_get_msgheader(packet
->rcd
, rhf_addr
);
644 lnh
= ib_get_lnh(hdr
);
646 if (lnh
== HFI1_LRH_BTH
) {
647 packet
->ohdr
= &hdr
->u
.oth
;
649 } else if (lnh
== HFI1_LRH_GRH
) {
650 packet
->ohdr
= &hdr
->u
.l
.oth
;
651 packet
->grh
= &hdr
->u
.l
.grh
;
653 goto next
; /* just in case */
656 if (!hfi1_may_ecn(packet
))
659 bth1
= be32_to_cpu(packet
->ohdr
->bth
[1]);
660 qpn
= bth1
& RVT_QPN_MASK
;
662 qp
= rvt_lookup_qpn(rdi
, &ibp
->rvp
, qpn
);
669 hfi1_process_ecn_slowpath(qp
, packet
, true);
672 /* turn off BECN, FECN */
673 bth1
&= ~(IB_FECN_SMASK
| IB_BECN_SMASK
);
674 packet
->ohdr
->bth
[1] = cpu_to_be32(bth1
);
676 update_ps_mdata(&mdata
, rcd
);
680 static void process_rcv_qp_work(struct hfi1_packet
*packet
)
682 struct rvt_qp
*qp
, *nqp
;
683 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
686 * Iterate over all QPs waiting to respond.
687 * The list won't change since the IRQ is only run on one CPU.
689 list_for_each_entry_safe(qp
, nqp
, &rcd
->qp_wait_list
, rspwait
) {
690 list_del_init(&qp
->rspwait
);
691 if (qp
->r_flags
& RVT_R_RSP_NAK
) {
692 qp
->r_flags
&= ~RVT_R_RSP_NAK
;
694 hfi1_send_rc_ack(packet
, 0);
696 if (qp
->r_flags
& RVT_R_RSP_SEND
) {
699 qp
->r_flags
&= ~RVT_R_RSP_SEND
;
700 spin_lock_irqsave(&qp
->s_lock
, flags
);
701 if (ib_rvt_state_ops
[qp
->state
] &
702 RVT_PROCESS_OR_FLUSH_SEND
)
703 hfi1_schedule_send(qp
);
704 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
710 static noinline
int max_packet_exceeded(struct hfi1_packet
*packet
, int thread
)
713 if ((packet
->numpkt
& (MAX_PKT_RECV_THREAD
- 1)) == 0)
714 /* allow defered processing */
715 process_rcv_qp_work(packet
);
719 this_cpu_inc(*packet
->rcd
->dd
->rcv_limit
);
720 return RCV_PKT_LIMIT
;
724 static inline int check_max_packet(struct hfi1_packet
*packet
, int thread
)
726 int ret
= RCV_PKT_OK
;
728 if (unlikely((packet
->numpkt
& (MAX_PKT_RECV
- 1)) == 0))
729 ret
= max_packet_exceeded(packet
, thread
);
733 static noinline
int skip_rcv_packet(struct hfi1_packet
*packet
, int thread
)
737 /* Set up for the next packet */
738 packet
->rhqoff
+= packet
->rsize
;
739 if (packet
->rhqoff
>= packet
->maxcnt
)
743 ret
= check_max_packet(packet
, thread
);
745 packet
->rhf_addr
= (__le32
*)packet
->rcd
->rcvhdrq
+ packet
->rhqoff
+
746 packet
->rcd
->rhf_offset
;
747 packet
->rhf
= rhf_to_cpu(packet
->rhf_addr
);
752 static inline int process_rcv_packet(struct hfi1_packet
*packet
, int thread
)
756 packet
->etype
= rhf_rcv_type(packet
->rhf
);
759 packet
->tlen
= rhf_pkt_len(packet
->rhf
); /* in bytes */
760 /* retrieve eager buffer details */
762 if (rhf_use_egr_bfr(packet
->rhf
)) {
763 packet
->etail
= rhf_egr_index(packet
->rhf
);
764 packet
->ebuf
= get_egrbuf(packet
->rcd
, packet
->rhf
,
767 * Prefetch the contents of the eager buffer. It is
768 * OK to send a negative length to prefetch_range().
769 * The +2 is the size of the RHF.
771 prefetch_range(packet
->ebuf
,
772 packet
->tlen
- ((packet
->rcd
->rcvhdrqentsize
-
773 (rhf_hdrq_offset(packet
->rhf
)
778 * Call a type specific handler for the packet. We
779 * should be able to trust that etype won't be beyond
780 * the range of valid indexes. If so something is really
781 * wrong and we can probably just let things come
782 * crashing down. There is no need to eat another
783 * comparison in this performance critical code.
785 packet
->rcd
->rhf_rcv_function_map
[packet
->etype
](packet
);
788 /* Set up for the next packet */
789 packet
->rhqoff
+= packet
->rsize
;
790 if (packet
->rhqoff
>= packet
->maxcnt
)
793 ret
= check_max_packet(packet
, thread
);
795 packet
->rhf_addr
= (__le32
*)packet
->rcd
->rcvhdrq
+ packet
->rhqoff
+
796 packet
->rcd
->rhf_offset
;
797 packet
->rhf
= rhf_to_cpu(packet
->rhf_addr
);
802 static inline void process_rcv_update(int last
, struct hfi1_packet
*packet
)
805 * Update head regs etc., every 16 packets, if not last pkt,
806 * to help prevent rcvhdrq overflows, when many packets
807 * are processed and queue is nearly full.
808 * Don't request an interrupt for intermediate updates.
810 if (!last
&& !(packet
->numpkt
& 0xf)) {
811 update_usrhead(packet
->rcd
, packet
->rhqoff
, packet
->updegr
,
812 packet
->etail
, 0, 0);
818 static inline void finish_packet(struct hfi1_packet
*packet
)
821 * Nothing we need to free for the packet.
823 * The only thing we need to do is a final update and call for an
826 update_usrhead(packet
->rcd
, packet
->rcd
->head
, packet
->updegr
,
827 packet
->etail
, rcv_intr_dynamic
, packet
->numpkt
);
831 * Handle receive interrupts when using the no dma rtail option.
833 int handle_receive_interrupt_nodma_rtail(struct hfi1_ctxtdata
*rcd
, int thread
)
836 int last
= RCV_PKT_OK
;
837 struct hfi1_packet packet
;
839 init_packet(rcd
, &packet
);
840 seq
= rhf_rcv_seq(packet
.rhf
);
841 if (seq
!= rcd
->seq_cnt
) {
846 prescan_rxq(rcd
, &packet
);
848 while (last
== RCV_PKT_OK
) {
849 last
= process_rcv_packet(&packet
, thread
);
850 seq
= rhf_rcv_seq(packet
.rhf
);
851 if (++rcd
->seq_cnt
> 13)
853 if (seq
!= rcd
->seq_cnt
)
855 process_rcv_update(last
, &packet
);
857 process_rcv_qp_work(&packet
);
858 rcd
->head
= packet
.rhqoff
;
860 finish_packet(&packet
);
864 int handle_receive_interrupt_dma_rtail(struct hfi1_ctxtdata
*rcd
, int thread
)
867 int last
= RCV_PKT_OK
;
868 struct hfi1_packet packet
;
870 init_packet(rcd
, &packet
);
871 hdrqtail
= get_rcvhdrtail(rcd
);
872 if (packet
.rhqoff
== hdrqtail
) {
876 smp_rmb(); /* prevent speculative reads of dma'ed hdrq */
878 prescan_rxq(rcd
, &packet
);
880 while (last
== RCV_PKT_OK
) {
881 last
= process_rcv_packet(&packet
, thread
);
882 if (packet
.rhqoff
== hdrqtail
)
884 process_rcv_update(last
, &packet
);
886 process_rcv_qp_work(&packet
);
887 rcd
->head
= packet
.rhqoff
;
889 finish_packet(&packet
);
893 static inline void set_nodma_rtail(struct hfi1_devdata
*dd
, u16 ctxt
)
895 struct hfi1_ctxtdata
*rcd
;
899 * For dynamically allocated kernel contexts (like vnic) switch
900 * interrupt handler only for that context. Otherwise, switch
901 * interrupt handler for all statically allocated kernel contexts.
903 if (ctxt
>= dd
->first_dyn_alloc_ctxt
) {
904 rcd
= hfi1_rcd_get_by_index_safe(dd
, ctxt
);
907 &handle_receive_interrupt_nodma_rtail
;
913 for (i
= HFI1_CTRL_CTXT
+ 1; i
< dd
->first_dyn_alloc_ctxt
; i
++) {
914 rcd
= hfi1_rcd_get_by_index(dd
, i
);
917 &handle_receive_interrupt_nodma_rtail
;
922 static inline void set_dma_rtail(struct hfi1_devdata
*dd
, u16 ctxt
)
924 struct hfi1_ctxtdata
*rcd
;
928 * For dynamically allocated kernel contexts (like vnic) switch
929 * interrupt handler only for that context. Otherwise, switch
930 * interrupt handler for all statically allocated kernel contexts.
932 if (ctxt
>= dd
->first_dyn_alloc_ctxt
) {
933 rcd
= hfi1_rcd_get_by_index_safe(dd
, ctxt
);
936 &handle_receive_interrupt_dma_rtail
;
942 for (i
= HFI1_CTRL_CTXT
+ 1; i
< dd
->first_dyn_alloc_ctxt
; i
++) {
943 rcd
= hfi1_rcd_get_by_index(dd
, i
);
946 &handle_receive_interrupt_dma_rtail
;
951 void set_all_slowpath(struct hfi1_devdata
*dd
)
953 struct hfi1_ctxtdata
*rcd
;
956 /* HFI1_CTRL_CTXT must always use the slow path interrupt handler */
957 for (i
= HFI1_CTRL_CTXT
+ 1; i
< dd
->num_rcv_contexts
; i
++) {
958 rcd
= hfi1_rcd_get_by_index(dd
, i
);
961 if (i
< dd
->first_dyn_alloc_ctxt
|| rcd
->is_vnic
)
962 rcd
->do_interrupt
= &handle_receive_interrupt
;
968 static inline int set_armed_to_active(struct hfi1_ctxtdata
*rcd
,
969 struct hfi1_packet
*packet
,
970 struct hfi1_devdata
*dd
)
972 struct work_struct
*lsaw
= &rcd
->ppd
->linkstate_active_work
;
973 u8 etype
= rhf_rcv_type(packet
->rhf
);
976 if (etype
== RHF_RCV_TYPE_IB
) {
977 struct ib_header
*hdr
= hfi1_get_msgheader(packet
->rcd
,
979 sc
= hfi1_9B_get_sc5(hdr
, packet
->rhf
);
980 } else if (etype
== RHF_RCV_TYPE_BYPASS
) {
981 struct hfi1_16b_header
*hdr
= hfi1_get_16B_header(
984 sc
= hfi1_16B_get_sc(hdr
);
986 if (sc
!= SC15_PACKET
) {
987 int hwstate
= driver_lstate(rcd
->ppd
);
989 if (hwstate
!= IB_PORT_ACTIVE
) {
991 "Unexpected link state %s\n",
992 opa_lstate_name(hwstate
));
996 queue_work(rcd
->ppd
->link_wq
, lsaw
);
1003 * handle_receive_interrupt - receive a packet
1006 * Called from interrupt handler for errors or receive interrupt.
1007 * This is the slow path interrupt handler.
1009 int handle_receive_interrupt(struct hfi1_ctxtdata
*rcd
, int thread
)
1011 struct hfi1_devdata
*dd
= rcd
->dd
;
1013 int needset
, last
= RCV_PKT_OK
;
1014 struct hfi1_packet packet
;
1017 /* Control context will always use the slow path interrupt handler */
1018 needset
= (rcd
->ctxt
== HFI1_CTRL_CTXT
) ? 0 : 1;
1020 init_packet(rcd
, &packet
);
1022 if (!HFI1_CAP_KGET_MASK(rcd
->flags
, DMA_RTAIL
)) {
1023 u32 seq
= rhf_rcv_seq(packet
.rhf
);
1025 if (seq
!= rcd
->seq_cnt
) {
1026 last
= RCV_PKT_DONE
;
1031 hdrqtail
= get_rcvhdrtail(rcd
);
1032 if (packet
.rhqoff
== hdrqtail
) {
1033 last
= RCV_PKT_DONE
;
1036 smp_rmb(); /* prevent speculative reads of dma'ed hdrq */
1039 * Control context can potentially receive an invalid
1040 * rhf. Drop such packets.
1042 if (rcd
->ctxt
== HFI1_CTRL_CTXT
) {
1043 u32 seq
= rhf_rcv_seq(packet
.rhf
);
1045 if (seq
!= rcd
->seq_cnt
)
1050 prescan_rxq(rcd
, &packet
);
1052 while (last
== RCV_PKT_OK
) {
1053 if (unlikely(dd
->do_drop
&&
1054 atomic_xchg(&dd
->drop_packet
, DROP_PACKET_OFF
) ==
1058 /* On to the next packet */
1059 packet
.rhqoff
+= packet
.rsize
;
1060 packet
.rhf_addr
= (__le32
*)rcd
->rcvhdrq
+
1063 packet
.rhf
= rhf_to_cpu(packet
.rhf_addr
);
1065 } else if (skip_pkt
) {
1066 last
= skip_rcv_packet(&packet
, thread
);
1069 /* Auto activate link on non-SC15 packet receive */
1070 if (unlikely(rcd
->ppd
->host_link_state
==
1072 set_armed_to_active(rcd
, &packet
, dd
))
1074 last
= process_rcv_packet(&packet
, thread
);
1077 if (!HFI1_CAP_KGET_MASK(rcd
->flags
, DMA_RTAIL
)) {
1078 u32 seq
= rhf_rcv_seq(packet
.rhf
);
1080 if (++rcd
->seq_cnt
> 13)
1082 if (seq
!= rcd
->seq_cnt
)
1083 last
= RCV_PKT_DONE
;
1085 dd_dev_info(dd
, "Switching to NO_DMA_RTAIL\n");
1086 set_nodma_rtail(dd
, rcd
->ctxt
);
1090 if (packet
.rhqoff
== hdrqtail
)
1091 last
= RCV_PKT_DONE
;
1093 * Control context can potentially receive an invalid
1094 * rhf. Drop such packets.
1096 if (rcd
->ctxt
== HFI1_CTRL_CTXT
) {
1097 u32 seq
= rhf_rcv_seq(packet
.rhf
);
1099 if (++rcd
->seq_cnt
> 13)
1101 if (!last
&& (seq
!= rcd
->seq_cnt
))
1107 "Switching to DMA_RTAIL\n");
1108 set_dma_rtail(dd
, rcd
->ctxt
);
1113 process_rcv_update(last
, &packet
);
1116 process_rcv_qp_work(&packet
);
1117 rcd
->head
= packet
.rhqoff
;
1121 * Always write head at end, and setup rcv interrupt, even
1122 * if no packets were processed.
1124 finish_packet(&packet
);
1129 * We may discover in the interrupt that the hardware link state has
1130 * changed from ARMED to ACTIVE (due to the arrival of a non-SC15 packet),
1131 * and we need to update the driver's notion of the link state. We cannot
1132 * run set_link_state from interrupt context, so we queue this function on
1135 * We delay the regular interrupt processing until after the state changes
1136 * so that the link will be in the correct state by the time any application
1137 * we wake up attempts to send a reply to any message it received.
1138 * (Subsequent receive interrupts may possibly force the wakeup before we
1139 * update the link state.)
1141 * The rcd is freed in hfi1_free_ctxtdata after hfi1_postinit_cleanup invokes
1142 * dd->f_cleanup(dd) to disable the interrupt handler and flush workqueues,
1143 * so we're safe from use-after-free of the rcd.
1145 void receive_interrupt_work(struct work_struct
*work
)
1147 struct hfi1_pportdata
*ppd
= container_of(work
, struct hfi1_pportdata
,
1148 linkstate_active_work
);
1149 struct hfi1_devdata
*dd
= ppd
->dd
;
1150 struct hfi1_ctxtdata
*rcd
;
1153 /* Received non-SC15 packet implies neighbor_normal */
1154 ppd
->neighbor_normal
= 1;
1155 set_link_state(ppd
, HLS_UP_ACTIVE
);
1158 * Interrupt all statically allocated kernel contexts that could
1159 * have had an interrupt during auto activation.
1161 for (i
= HFI1_CTRL_CTXT
; i
< dd
->first_dyn_alloc_ctxt
; i
++) {
1162 rcd
= hfi1_rcd_get_by_index(dd
, i
);
1164 force_recv_intr(rcd
);
1170 * Convert a given MTU size to the on-wire MAD packet enumeration.
1171 * Return -1 if the size is invalid.
1173 int mtu_to_enum(u32 mtu
, int default_if_bad
)
1176 case 0: return OPA_MTU_0
;
1177 case 256: return OPA_MTU_256
;
1178 case 512: return OPA_MTU_512
;
1179 case 1024: return OPA_MTU_1024
;
1180 case 2048: return OPA_MTU_2048
;
1181 case 4096: return OPA_MTU_4096
;
1182 case 8192: return OPA_MTU_8192
;
1183 case 10240: return OPA_MTU_10240
;
1185 return default_if_bad
;
1188 u16
enum_to_mtu(int mtu
)
1191 case OPA_MTU_0
: return 0;
1192 case OPA_MTU_256
: return 256;
1193 case OPA_MTU_512
: return 512;
1194 case OPA_MTU_1024
: return 1024;
1195 case OPA_MTU_2048
: return 2048;
1196 case OPA_MTU_4096
: return 4096;
1197 case OPA_MTU_8192
: return 8192;
1198 case OPA_MTU_10240
: return 10240;
1199 default: return 0xffff;
1204 * set_mtu - set the MTU
1205 * @ppd: the per port data
1207 * We can handle "any" incoming size, the issue here is whether we
1208 * need to restrict our outgoing size. We do not deal with what happens
1209 * to programs that are already running when the size changes.
1211 int set_mtu(struct hfi1_pportdata
*ppd
)
1213 struct hfi1_devdata
*dd
= ppd
->dd
;
1214 int i
, drain
, ret
= 0, is_up
= 0;
1217 for (i
= 0; i
< ppd
->vls_supported
; i
++)
1218 if (ppd
->ibmtu
< dd
->vld
[i
].mtu
)
1219 ppd
->ibmtu
= dd
->vld
[i
].mtu
;
1220 ppd
->ibmaxlen
= ppd
->ibmtu
+ lrh_max_header_bytes(ppd
->dd
);
1222 mutex_lock(&ppd
->hls_lock
);
1223 if (ppd
->host_link_state
== HLS_UP_INIT
||
1224 ppd
->host_link_state
== HLS_UP_ARMED
||
1225 ppd
->host_link_state
== HLS_UP_ACTIVE
)
1228 drain
= !is_ax(dd
) && is_up
;
1232 * MTU is specified per-VL. To ensure that no packet gets
1233 * stuck (due, e.g., to the MTU for the packet's VL being
1234 * reduced), empty the per-VL FIFOs before adjusting MTU.
1236 ret
= stop_drain_data_vls(dd
);
1239 dd_dev_err(dd
, "%s: cannot stop/drain VLs - refusing to change per-VL MTUs\n",
1244 hfi1_set_ib_cfg(ppd
, HFI1_IB_CFG_MTU
, 0);
1247 open_fill_data_vls(dd
); /* reopen all VLs */
1250 mutex_unlock(&ppd
->hls_lock
);
1255 int hfi1_set_lid(struct hfi1_pportdata
*ppd
, u32 lid
, u8 lmc
)
1257 struct hfi1_devdata
*dd
= ppd
->dd
;
1261 hfi1_set_ib_cfg(ppd
, HFI1_IB_CFG_LIDLMC
, 0);
1263 dd_dev_info(dd
, "port %u: got a lid: 0x%x\n", ppd
->port
, lid
);
1268 void shutdown_led_override(struct hfi1_pportdata
*ppd
)
1270 struct hfi1_devdata
*dd
= ppd
->dd
;
1273 * This pairs with the memory barrier in hfi1_start_led_override to
1274 * ensure that we read the correct state of LED beaconing represented
1275 * by led_override_timer_active
1278 if (atomic_read(&ppd
->led_override_timer_active
)) {
1279 del_timer_sync(&ppd
->led_override_timer
);
1280 atomic_set(&ppd
->led_override_timer_active
, 0);
1281 /* Ensure the atomic_set is visible to all CPUs */
1285 /* Hand control of the LED to the DC for normal operation */
1286 write_csr(dd
, DCC_CFG_LED_CNTRL
, 0);
1289 static void run_led_override(struct timer_list
*t
)
1291 struct hfi1_pportdata
*ppd
= from_timer(ppd
, t
, led_override_timer
);
1292 struct hfi1_devdata
*dd
= ppd
->dd
;
1293 unsigned long timeout
;
1296 if (!(dd
->flags
& HFI1_INITTED
))
1299 phase_idx
= ppd
->led_override_phase
& 1;
1301 setextled(dd
, phase_idx
);
1303 timeout
= ppd
->led_override_vals
[phase_idx
];
1305 /* Set up for next phase */
1306 ppd
->led_override_phase
= !ppd
->led_override_phase
;
1308 mod_timer(&ppd
->led_override_timer
, jiffies
+ timeout
);
1312 * To have the LED blink in a particular pattern, provide timeon and timeoff
1314 * To turn off custom blinking and return to normal operation, use
1315 * shutdown_led_override()
1317 void hfi1_start_led_override(struct hfi1_pportdata
*ppd
, unsigned int timeon
,
1318 unsigned int timeoff
)
1320 if (!(ppd
->dd
->flags
& HFI1_INITTED
))
1323 /* Convert to jiffies for direct use in timer */
1324 ppd
->led_override_vals
[0] = msecs_to_jiffies(timeoff
);
1325 ppd
->led_override_vals
[1] = msecs_to_jiffies(timeon
);
1327 /* Arbitrarily start from LED on phase */
1328 ppd
->led_override_phase
= 1;
1331 * If the timer has not already been started, do so. Use a "quick"
1332 * timeout so the handler will be called soon to look at our request.
1334 if (!timer_pending(&ppd
->led_override_timer
)) {
1335 timer_setup(&ppd
->led_override_timer
, run_led_override
, 0);
1336 ppd
->led_override_timer
.expires
= jiffies
+ 1;
1337 add_timer(&ppd
->led_override_timer
);
1338 atomic_set(&ppd
->led_override_timer_active
, 1);
1339 /* Ensure the atomic_set is visible to all CPUs */
1345 * hfi1_reset_device - reset the chip if possible
1346 * @unit: the device to reset
1348 * Whether or not reset is successful, we attempt to re-initialize the chip
1349 * (that is, much like a driver unload/reload). We clear the INITTED flag
1350 * so that the various entry points will fail until we reinitialize. For
1351 * now, we only allow this if no user contexts are open that use chip resources
1353 int hfi1_reset_device(int unit
)
1356 struct hfi1_devdata
*dd
= hfi1_lookup(unit
);
1357 struct hfi1_pportdata
*ppd
;
1365 dd_dev_info(dd
, "Reset on unit %u requested\n", unit
);
1367 if (!dd
->kregbase1
|| !(dd
->flags
& HFI1_PRESENT
)) {
1369 "Invalid unit number %u or not initialized or not present\n",
1375 /* If there are any user/vnic contexts, we cannot reset */
1376 mutex_lock(&hfi1_mutex
);
1378 if (hfi1_stats
.sps_ctxts
) {
1379 mutex_unlock(&hfi1_mutex
);
1383 mutex_unlock(&hfi1_mutex
);
1385 for (pidx
= 0; pidx
< dd
->num_pports
; ++pidx
) {
1386 ppd
= dd
->pport
+ pidx
;
1388 shutdown_led_override(ppd
);
1390 if (dd
->flags
& HFI1_HAS_SEND_DMA
)
1393 hfi1_reset_cpu_counters(dd
);
1395 ret
= hfi1_init(dd
, 1);
1399 "Reinitialize unit %u after reset failed with %d\n",
1402 dd_dev_info(dd
, "Reinitialized unit %u after resetting\n",
1409 static inline void hfi1_setup_ib_header(struct hfi1_packet
*packet
)
1411 packet
->hdr
= (struct hfi1_ib_message_header
*)
1412 hfi1_get_msgheader(packet
->rcd
,
1414 packet
->hlen
= (u8
*)packet
->rhf_addr
- (u8
*)packet
->hdr
;
1417 static int hfi1_bypass_ingress_pkt_check(struct hfi1_packet
*packet
)
1419 struct hfi1_pportdata
*ppd
= packet
->rcd
->ppd
;
1421 /* slid and dlid cannot be 0 */
1422 if ((!packet
->slid
) || (!packet
->dlid
))
1425 /* Compare port lid with incoming packet dlid */
1426 if ((!(hfi1_is_16B_mcast(packet
->dlid
))) &&
1428 opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE
), 16B
))) {
1429 if ((packet
->dlid
& ~((1 << ppd
->lmc
) - 1)) != ppd
->lid
)
1433 /* No multicast packets with SC15 */
1434 if ((hfi1_is_16B_mcast(packet
->dlid
)) && (packet
->sc
== 0xF))
1437 /* Packets with permissive DLID always on SC15 */
1438 if ((packet
->dlid
== opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE
),
1440 (packet
->sc
!= 0xF))
1446 static int hfi1_setup_9B_packet(struct hfi1_packet
*packet
)
1448 struct hfi1_ibport
*ibp
= rcd_to_iport(packet
->rcd
);
1449 struct ib_header
*hdr
;
1452 hfi1_setup_ib_header(packet
);
1455 lnh
= ib_get_lnh(hdr
);
1456 if (lnh
== HFI1_LRH_BTH
) {
1457 packet
->ohdr
= &hdr
->u
.oth
;
1459 } else if (lnh
== HFI1_LRH_GRH
) {
1462 packet
->ohdr
= &hdr
->u
.l
.oth
;
1463 packet
->grh
= &hdr
->u
.l
.grh
;
1464 if (packet
->grh
->next_hdr
!= IB_GRH_NEXT_HDR
)
1466 vtf
= be32_to_cpu(packet
->grh
->version_tclass_flow
);
1467 if ((vtf
>> IB_GRH_VERSION_SHIFT
) != IB_GRH_VERSION
)
1473 /* Query commonly used fields from packet header */
1474 packet
->payload
= packet
->ebuf
;
1475 packet
->opcode
= ib_bth_get_opcode(packet
->ohdr
);
1476 packet
->slid
= ib_get_slid(hdr
);
1477 packet
->dlid
= ib_get_dlid(hdr
);
1478 if (unlikely((packet
->dlid
>= be16_to_cpu(IB_MULTICAST_LID_BASE
)) &&
1479 (packet
->dlid
!= be16_to_cpu(IB_LID_PERMISSIVE
))))
1480 packet
->dlid
+= opa_get_mcast_base(OPA_MCAST_NR
) -
1481 be16_to_cpu(IB_MULTICAST_LID_BASE
);
1482 packet
->sl
= ib_get_sl(hdr
);
1483 packet
->sc
= hfi1_9B_get_sc5(hdr
, packet
->rhf
);
1484 packet
->pad
= ib_bth_get_pad(packet
->ohdr
);
1485 packet
->extra_byte
= 0;
1486 packet
->pkey
= ib_bth_get_pkey(packet
->ohdr
);
1487 packet
->migrated
= ib_bth_is_migration(packet
->ohdr
);
1491 ibp
->rvp
.n_pkt_drops
++;
1495 static int hfi1_setup_bypass_packet(struct hfi1_packet
*packet
)
1498 * Bypass packets have a different header/payload split
1499 * compared to an IB packet.
1500 * Current split is set such that 16 bytes of the actual
1501 * header is in the header buffer and the remining is in
1502 * the eager buffer. We chose 16 since hfi1 driver only
1503 * supports 16B bypass packets and we will be able to
1504 * receive the entire LRH with such a split.
1507 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
1508 struct hfi1_pportdata
*ppd
= rcd
->ppd
;
1509 struct hfi1_ibport
*ibp
= &ppd
->ibport_data
;
1512 packet
->hdr
= (struct hfi1_16b_header
*)
1513 hfi1_get_16B_header(packet
->rcd
,
1515 l4
= hfi1_16B_get_l4(packet
->hdr
);
1516 if (l4
== OPA_16B_L4_IB_LOCAL
) {
1517 packet
->ohdr
= packet
->ebuf
;
1519 packet
->opcode
= ib_bth_get_opcode(packet
->ohdr
);
1520 packet
->pad
= hfi1_16B_bth_get_pad(packet
->ohdr
);
1521 /* hdr_len_by_opcode already has an IB LRH factored in */
1522 packet
->hlen
= hdr_len_by_opcode
[packet
->opcode
] +
1523 (LRH_16B_BYTES
- LRH_9B_BYTES
);
1524 packet
->migrated
= opa_bth_is_migration(packet
->ohdr
);
1525 } else if (l4
== OPA_16B_L4_IB_GLOBAL
) {
1527 u8 grh_len
= sizeof(struct ib_grh
);
1529 packet
->ohdr
= packet
->ebuf
+ grh_len
;
1530 packet
->grh
= packet
->ebuf
;
1531 packet
->opcode
= ib_bth_get_opcode(packet
->ohdr
);
1532 packet
->pad
= hfi1_16B_bth_get_pad(packet
->ohdr
);
1533 /* hdr_len_by_opcode already has an IB LRH factored in */
1534 packet
->hlen
= hdr_len_by_opcode
[packet
->opcode
] +
1535 (LRH_16B_BYTES
- LRH_9B_BYTES
) + grh_len
;
1536 packet
->migrated
= opa_bth_is_migration(packet
->ohdr
);
1538 if (packet
->grh
->next_hdr
!= IB_GRH_NEXT_HDR
)
1540 vtf
= be32_to_cpu(packet
->grh
->version_tclass_flow
);
1541 if ((vtf
>> IB_GRH_VERSION_SHIFT
) != IB_GRH_VERSION
)
1543 } else if (l4
== OPA_16B_L4_FM
) {
1544 packet
->mgmt
= packet
->ebuf
;
1545 packet
->ohdr
= NULL
;
1547 packet
->opcode
= IB_OPCODE_UD_SEND_ONLY
;
1548 packet
->pad
= OPA_16B_L4_FM_PAD
;
1549 packet
->hlen
= OPA_16B_L4_FM_HLEN
;
1550 packet
->migrated
= false;
1555 /* Query commonly used fields from packet header */
1556 packet
->payload
= packet
->ebuf
+ packet
->hlen
- LRH_16B_BYTES
;
1557 packet
->slid
= hfi1_16B_get_slid(packet
->hdr
);
1558 packet
->dlid
= hfi1_16B_get_dlid(packet
->hdr
);
1559 if (unlikely(hfi1_is_16B_mcast(packet
->dlid
)))
1560 packet
->dlid
+= opa_get_mcast_base(OPA_MCAST_NR
) -
1561 opa_get_lid(opa_get_mcast_base(OPA_MCAST_NR
),
1563 packet
->sc
= hfi1_16B_get_sc(packet
->hdr
);
1564 packet
->sl
= ibp
->sc_to_sl
[packet
->sc
];
1565 packet
->extra_byte
= SIZE_OF_LT
;
1566 packet
->pkey
= hfi1_16B_get_pkey(packet
->hdr
);
1568 if (hfi1_bypass_ingress_pkt_check(packet
))
1573 hfi1_cdbg(PKT
, "%s: packet dropped\n", __func__
);
1574 ibp
->rvp
.n_pkt_drops
++;
1578 static void show_eflags_errs(struct hfi1_packet
*packet
)
1580 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
1581 u32 rte
= rhf_rcv_type_err(packet
->rhf
);
1584 "receive context %d: rhf 0x%016llx, errs [ %s%s%s%s%s%s%s] rte 0x%x\n",
1585 rcd
->ctxt
, packet
->rhf
,
1586 packet
->rhf
& RHF_K_HDR_LEN_ERR
? "k_hdr_len " : "",
1587 packet
->rhf
& RHF_DC_UNC_ERR
? "dc_unc " : "",
1588 packet
->rhf
& RHF_DC_ERR
? "dc " : "",
1589 packet
->rhf
& RHF_TID_ERR
? "tid " : "",
1590 packet
->rhf
& RHF_LEN_ERR
? "len " : "",
1591 packet
->rhf
& RHF_ECC_ERR
? "ecc " : "",
1592 packet
->rhf
& RHF_ICRC_ERR
? "icrc " : "",
1596 void handle_eflags(struct hfi1_packet
*packet
)
1598 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
1600 rcv_hdrerr(rcd
, rcd
->ppd
, packet
);
1601 if (rhf_err_flags(packet
->rhf
))
1602 show_eflags_errs(packet
);
1606 * The following functions are called by the interrupt handler. They are type
1607 * specific handlers for each packet type.
1609 static int process_receive_ib(struct hfi1_packet
*packet
)
1611 if (hfi1_setup_9B_packet(packet
))
1612 return RHF_RCV_CONTINUE
;
1614 if (unlikely(hfi1_dbg_should_fault_rx(packet
)))
1615 return RHF_RCV_CONTINUE
;
1617 trace_hfi1_rcvhdr(packet
);
1619 if (unlikely(rhf_err_flags(packet
->rhf
))) {
1620 handle_eflags(packet
);
1621 return RHF_RCV_CONTINUE
;
1624 hfi1_ib_rcv(packet
);
1625 return RHF_RCV_CONTINUE
;
1628 static inline bool hfi1_is_vnic_packet(struct hfi1_packet
*packet
)
1630 /* Packet received in VNIC context via RSM */
1631 if (packet
->rcd
->is_vnic
)
1634 if ((hfi1_16B_get_l2(packet
->ebuf
) == OPA_16B_L2_TYPE
) &&
1635 (hfi1_16B_get_l4(packet
->ebuf
) == OPA_16B_L4_ETHR
))
1641 static int process_receive_bypass(struct hfi1_packet
*packet
)
1643 struct hfi1_devdata
*dd
= packet
->rcd
->dd
;
1645 if (hfi1_is_vnic_packet(packet
)) {
1646 hfi1_vnic_bypass_rcv(packet
);
1647 return RHF_RCV_CONTINUE
;
1650 if (hfi1_setup_bypass_packet(packet
))
1651 return RHF_RCV_CONTINUE
;
1653 trace_hfi1_rcvhdr(packet
);
1655 if (unlikely(rhf_err_flags(packet
->rhf
))) {
1656 handle_eflags(packet
);
1657 return RHF_RCV_CONTINUE
;
1660 if (hfi1_16B_get_l2(packet
->hdr
) == 0x2) {
1661 hfi1_16B_rcv(packet
);
1664 "Bypass packets other than 16B are not supported in normal operation. Dropping\n");
1665 incr_cntr64(&dd
->sw_rcv_bypass_packet_errors
);
1666 if (!(dd
->err_info_rcvport
.status_and_code
&
1667 OPA_EI_STATUS_SMASK
)) {
1668 u64
*flits
= packet
->ebuf
;
1670 if (flits
&& !(packet
->rhf
& RHF_LEN_ERR
)) {
1671 dd
->err_info_rcvport
.packet_flit1
= flits
[0];
1672 dd
->err_info_rcvport
.packet_flit2
=
1673 packet
->tlen
> sizeof(flits
[0]) ?
1676 dd
->err_info_rcvport
.status_and_code
|=
1677 (OPA_EI_STATUS_SMASK
| BAD_L2_ERR
);
1680 return RHF_RCV_CONTINUE
;
1683 static int process_receive_error(struct hfi1_packet
*packet
)
1685 /* KHdrHCRCErr -- KDETH packet with a bad HCRC */
1687 hfi1_dbg_fault_suppress_err(&packet
->rcd
->dd
->verbs_dev
) &&
1688 (rhf_rcv_type_err(packet
->rhf
) == RHF_RCV_TYPE_ERROR
||
1689 packet
->rhf
& RHF_DC_ERR
)))
1690 return RHF_RCV_CONTINUE
;
1692 hfi1_setup_ib_header(packet
);
1693 handle_eflags(packet
);
1695 if (unlikely(rhf_err_flags(packet
->rhf
)))
1696 dd_dev_err(packet
->rcd
->dd
,
1697 "Unhandled error packet received. Dropping.\n");
1699 return RHF_RCV_CONTINUE
;
1702 static int kdeth_process_expected(struct hfi1_packet
*packet
)
1704 hfi1_setup_9B_packet(packet
);
1705 if (unlikely(hfi1_dbg_should_fault_rx(packet
)))
1706 return RHF_RCV_CONTINUE
;
1708 if (unlikely(rhf_err_flags(packet
->rhf
))) {
1709 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
1711 if (hfi1_handle_kdeth_eflags(rcd
, rcd
->ppd
, packet
))
1712 return RHF_RCV_CONTINUE
;
1715 hfi1_kdeth_expected_rcv(packet
);
1716 return RHF_RCV_CONTINUE
;
1719 static int kdeth_process_eager(struct hfi1_packet
*packet
)
1721 hfi1_setup_9B_packet(packet
);
1722 if (unlikely(hfi1_dbg_should_fault_rx(packet
)))
1723 return RHF_RCV_CONTINUE
;
1725 trace_hfi1_rcvhdr(packet
);
1726 if (unlikely(rhf_err_flags(packet
->rhf
))) {
1727 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
1729 show_eflags_errs(packet
);
1730 if (hfi1_handle_kdeth_eflags(rcd
, rcd
->ppd
, packet
))
1731 return RHF_RCV_CONTINUE
;
1734 hfi1_kdeth_eager_rcv(packet
);
1735 return RHF_RCV_CONTINUE
;
1738 static int process_receive_invalid(struct hfi1_packet
*packet
)
1740 dd_dev_err(packet
->rcd
->dd
, "Invalid packet type %d. Dropping\n",
1741 rhf_rcv_type(packet
->rhf
));
1742 return RHF_RCV_CONTINUE
;
1745 void seqfile_dump_rcd(struct seq_file
*s
, struct hfi1_ctxtdata
*rcd
)
1747 struct hfi1_packet packet
;
1748 struct ps_mdata mdata
;
1750 seq_printf(s
, "Rcd %u: RcvHdr cnt %u entsize %u %s head %llu tail %llu\n",
1751 rcd
->ctxt
, rcd
->rcvhdrq_cnt
, rcd
->rcvhdrqentsize
,
1752 HFI1_CAP_KGET_MASK(rcd
->flags
, DMA_RTAIL
) ?
1753 "dma_rtail" : "nodma_rtail",
1754 read_uctxt_csr(rcd
->dd
, rcd
->ctxt
, RCV_HDR_HEAD
) &
1755 RCV_HDR_HEAD_HEAD_MASK
,
1756 read_uctxt_csr(rcd
->dd
, rcd
->ctxt
, RCV_HDR_TAIL
));
1758 init_packet(rcd
, &packet
);
1759 init_ps_mdata(&mdata
, &packet
);
1762 __le32
*rhf_addr
= (__le32
*)rcd
->rcvhdrq
+ mdata
.ps_head
+
1764 struct ib_header
*hdr
;
1765 u64 rhf
= rhf_to_cpu(rhf_addr
);
1766 u32 etype
= rhf_rcv_type(rhf
), qpn
;
1771 if (ps_done(&mdata
, rhf
, rcd
))
1774 if (ps_skip(&mdata
, rhf
, rcd
))
1777 if (etype
> RHF_RCV_TYPE_IB
)
1780 packet
.hdr
= hfi1_get_msgheader(rcd
, rhf_addr
);
1783 lnh
= be16_to_cpu(hdr
->lrh
[0]) & 3;
1785 if (lnh
== HFI1_LRH_BTH
)
1786 packet
.ohdr
= &hdr
->u
.oth
;
1787 else if (lnh
== HFI1_LRH_GRH
)
1788 packet
.ohdr
= &hdr
->u
.l
.oth
;
1790 goto next
; /* just in case */
1792 opcode
= (be32_to_cpu(packet
.ohdr
->bth
[0]) >> 24);
1793 qpn
= be32_to_cpu(packet
.ohdr
->bth
[1]) & RVT_QPN_MASK
;
1794 psn
= mask_psn(be32_to_cpu(packet
.ohdr
->bth
[2]));
1796 seq_printf(s
, "\tEnt %u: opcode 0x%x, qpn 0x%x, psn 0x%x\n",
1797 mdata
.ps_head
, opcode
, qpn
, psn
);
1799 update_ps_mdata(&mdata
, rcd
);
1803 const rhf_rcv_function_ptr normal_rhf_rcv_functions
[] = {
1804 [RHF_RCV_TYPE_EXPECTED
] = kdeth_process_expected
,
1805 [RHF_RCV_TYPE_EAGER
] = kdeth_process_eager
,
1806 [RHF_RCV_TYPE_IB
] = process_receive_ib
,
1807 [RHF_RCV_TYPE_ERROR
] = process_receive_error
,
1808 [RHF_RCV_TYPE_BYPASS
] = process_receive_bypass
,
1809 [RHF_RCV_TYPE_INVALID5
] = process_receive_invalid
,
1810 [RHF_RCV_TYPE_INVALID6
] = process_receive_invalid
,
1811 [RHF_RCV_TYPE_INVALID7
] = process_receive_invalid
,